go/src/cmd/compile/internal/gc/reflect.go

1542 lines
36 KiB
Go
Raw Normal View History

// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package gc
import (
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
"cmd/internal/gcprog"
"cmd/internal/obj"
"fmt"
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
"os"
"sort"
"strings"
)
// runtime interface and reflection data structures
var signatlist *NodeList
// byMethodNameAndPackagePath sorts method signatures by name, then package path.
type byMethodNameAndPackagePath []*Sig
func (x byMethodNameAndPackagePath) Len() int { return len(x) }
func (x byMethodNameAndPackagePath) Swap(i, j int) { x[i], x[j] = x[j], x[i] }
func (x byMethodNameAndPackagePath) Less(i, j int) bool {
return siglt(x[i], x[j])
}
// siglt reports whether a < b
func siglt(a, b *Sig) bool {
if a.name != b.name {
return a.name < b.name
}
if a.pkg == b.pkg {
return false
}
if a.pkg == nil {
return true
}
if b.pkg == nil {
return false
}
return a.pkg.Path < b.pkg.Path
}
// Builds a type representing a Bucket structure for
// the given map type. This type is not visible to users -
// we include only enough information to generate a correct GC
// program for it.
// Make sure this stays in sync with ../../../../runtime/hashmap.go!
const (
BUCKETSIZE = 8
MAXKEYSIZE = 128
MAXVALSIZE = 128
)
func makefield(name string, t *Type) *Type {
f := typ(TFIELD)
f.Type = t
f.Sym = nopkg.Lookup(name)
return f
}
func mapbucket(t *Type) *Type {
if t.Bucket != nil {
return t.Bucket
}
bucket := typ(TSTRUCT)
keytype := t.Down
valtype := t.Type
dowidth(keytype)
dowidth(valtype)
if keytype.Width > MAXKEYSIZE {
keytype = Ptrto(keytype)
}
if valtype.Width > MAXVALSIZE {
valtype = Ptrto(valtype)
}
// The first field is: uint8 topbits[BUCKETSIZE].
arr := typ(TARRAY)
arr.Type = Types[TUINT8]
arr.Bound = BUCKETSIZE
cmd/compile, runtime: fix placement of map bucket overflow pointer on nacl On most systems, a pointer is the worst case alignment, so adding a pointer field at the end of a struct guarantees there will be no padding added after that field (to satisfy overall struct alignment due to some more-aligned field also present). In the runtime, the map implementation needs a quick way to get to the overflow pointer, which is last in the bucket struct, so it uses size - sizeof(pointer) as the offset. NaCl/amd64p32 is the exception, as always. The worst case alignment is 64 bits but pointers are 32 bits. There's a long history that is not worth going into, but when we moved the overflow pointer to the end of the struct, we didn't get the padding computation right. The compiler computed the regular struct size and then on amd64p32 added another 32-bit field. And the runtime assumed it could step back two 32-bit fields (one 64-bit register size) to get to the overflow pointer. But in fact if the struct needed 64-bit alignment, the computation of the regular struct size would have added a 32-bit pad already, and then the code unconditionally added a second 32-bit pad. This placed the overflow pointer three words from the end, not two. The last two were padding, and since the runtime was consistent about using the second-to-last word as the overflow pointer, no harm done in the sense of overwriting useful memory. But writing the overflow pointer to a non-pointer word of memory means that the GC can't see the overflow blocks, so it will collect them prematurely. Then bad things happen. Correct all this in a few steps: 1. Add an explicit check at the end of the bucket layout in the compiler that the overflow field is last in the struct, never followed by padding. 2. When padding is needed on nacl (not always, just when needed), insert it before the overflow pointer, to preserve the "last in the struct" property. 3. Let the compiler have the final word on the width of the struct, by inserting an explicit padding field instead of overwriting the results of the width computation it does. 4. For the same reason (tell the truth to the compiler), set the type of the overflow field when we're trying to pretend its not a pointer (in this case the runtime maintains a list of the overflow blocks elsewhere). 5. Make the runtime use "last in the struct" as its location algorithm. This fixes TestTraceStress on nacl/amd64p32. The 'bad map state' and 'invalid free list' failures no longer occur. Fixes #11838. Change-Id: If918887f8f252d988db0a35159944d2b36512f92 Reviewed-on: https://go-review.googlesource.com/12971 Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Austin Clements <austin@google.com>
2015-07-30 22:05:51 -04:00
field := make([]*Type, 0, 5)
field = append(field, makefield("topbits", arr))
arr = typ(TARRAY)
arr.Type = keytype
arr.Bound = BUCKETSIZE
cmd/compile, runtime: fix placement of map bucket overflow pointer on nacl On most systems, a pointer is the worst case alignment, so adding a pointer field at the end of a struct guarantees there will be no padding added after that field (to satisfy overall struct alignment due to some more-aligned field also present). In the runtime, the map implementation needs a quick way to get to the overflow pointer, which is last in the bucket struct, so it uses size - sizeof(pointer) as the offset. NaCl/amd64p32 is the exception, as always. The worst case alignment is 64 bits but pointers are 32 bits. There's a long history that is not worth going into, but when we moved the overflow pointer to the end of the struct, we didn't get the padding computation right. The compiler computed the regular struct size and then on amd64p32 added another 32-bit field. And the runtime assumed it could step back two 32-bit fields (one 64-bit register size) to get to the overflow pointer. But in fact if the struct needed 64-bit alignment, the computation of the regular struct size would have added a 32-bit pad already, and then the code unconditionally added a second 32-bit pad. This placed the overflow pointer three words from the end, not two. The last two were padding, and since the runtime was consistent about using the second-to-last word as the overflow pointer, no harm done in the sense of overwriting useful memory. But writing the overflow pointer to a non-pointer word of memory means that the GC can't see the overflow blocks, so it will collect them prematurely. Then bad things happen. Correct all this in a few steps: 1. Add an explicit check at the end of the bucket layout in the compiler that the overflow field is last in the struct, never followed by padding. 2. When padding is needed on nacl (not always, just when needed), insert it before the overflow pointer, to preserve the "last in the struct" property. 3. Let the compiler have the final word on the width of the struct, by inserting an explicit padding field instead of overwriting the results of the width computation it does. 4. For the same reason (tell the truth to the compiler), set the type of the overflow field when we're trying to pretend its not a pointer (in this case the runtime maintains a list of the overflow blocks elsewhere). 5. Make the runtime use "last in the struct" as its location algorithm. This fixes TestTraceStress on nacl/amd64p32. The 'bad map state' and 'invalid free list' failures no longer occur. Fixes #11838. Change-Id: If918887f8f252d988db0a35159944d2b36512f92 Reviewed-on: https://go-review.googlesource.com/12971 Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Austin Clements <austin@google.com>
2015-07-30 22:05:51 -04:00
field = append(field, makefield("keys", arr))
arr = typ(TARRAY)
arr.Type = valtype
arr.Bound = BUCKETSIZE
cmd/compile, runtime: fix placement of map bucket overflow pointer on nacl On most systems, a pointer is the worst case alignment, so adding a pointer field at the end of a struct guarantees there will be no padding added after that field (to satisfy overall struct alignment due to some more-aligned field also present). In the runtime, the map implementation needs a quick way to get to the overflow pointer, which is last in the bucket struct, so it uses size - sizeof(pointer) as the offset. NaCl/amd64p32 is the exception, as always. The worst case alignment is 64 bits but pointers are 32 bits. There's a long history that is not worth going into, but when we moved the overflow pointer to the end of the struct, we didn't get the padding computation right. The compiler computed the regular struct size and then on amd64p32 added another 32-bit field. And the runtime assumed it could step back two 32-bit fields (one 64-bit register size) to get to the overflow pointer. But in fact if the struct needed 64-bit alignment, the computation of the regular struct size would have added a 32-bit pad already, and then the code unconditionally added a second 32-bit pad. This placed the overflow pointer three words from the end, not two. The last two were padding, and since the runtime was consistent about using the second-to-last word as the overflow pointer, no harm done in the sense of overwriting useful memory. But writing the overflow pointer to a non-pointer word of memory means that the GC can't see the overflow blocks, so it will collect them prematurely. Then bad things happen. Correct all this in a few steps: 1. Add an explicit check at the end of the bucket layout in the compiler that the overflow field is last in the struct, never followed by padding. 2. When padding is needed on nacl (not always, just when needed), insert it before the overflow pointer, to preserve the "last in the struct" property. 3. Let the compiler have the final word on the width of the struct, by inserting an explicit padding field instead of overwriting the results of the width computation it does. 4. For the same reason (tell the truth to the compiler), set the type of the overflow field when we're trying to pretend its not a pointer (in this case the runtime maintains a list of the overflow blocks elsewhere). 5. Make the runtime use "last in the struct" as its location algorithm. This fixes TestTraceStress on nacl/amd64p32. The 'bad map state' and 'invalid free list' failures no longer occur. Fixes #11838. Change-Id: If918887f8f252d988db0a35159944d2b36512f92 Reviewed-on: https://go-review.googlesource.com/12971 Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Austin Clements <austin@google.com>
2015-07-30 22:05:51 -04:00
field = append(field, makefield("values", arr))
// Make sure the overflow pointer is the last memory in the struct,
// because the runtime assumes it can use size-ptrSize as the
// offset of the overflow pointer. We double-check that property
// below once the offsets and size are computed.
//
// BUCKETSIZE is 8, so the struct is aligned to 64 bits to this point.
// On 32-bit systems, the max alignment is 32-bit, and the
// overflow pointer will add another 32-bit field, and the struct
// will end with no padding.
// On 64-bit systems, the max alignment is 64-bit, and the
// overflow pointer will add another 64-bit field, and the struct
// will end with no padding.
// On nacl/amd64p32, however, the max alignment is 64-bit,
// but the overflow pointer will add only a 32-bit field,
// so if the struct needs 64-bit padding (because a key or value does)
// then it would end with an extra 32-bit padding field.
// Preempt that by emitting the padding here.
if int(t.Type.Align) > Widthptr || int(t.Down.Align) > Widthptr {
field = append(field, makefield("pad", Types[TUINTPTR]))
}
// If keys and values have no pointers, the map implementation
// can keep a list of overflow pointers on the side so that
// buckets can be marked as having no pointers.
// Arrange for the bucket to have no pointers by changing
// the type of the overflow field to uintptr in this case.
// See comment on hmap.overflow in ../../../../runtime/hashmap.go.
otyp := Ptrto(bucket)
if !haspointers(t.Type) && !haspointers(t.Down) && t.Type.Width <= MAXKEYSIZE && t.Down.Width <= MAXVALSIZE {
otyp = Types[TUINTPTR]
}
ovf := makefield("overflow", otyp)
field = append(field, ovf)
// link up fields
bucket.Noalg = true
bucket.Local = t.Local
bucket.Type = field[0]
for n := int32(0); n < int32(len(field)-1); n++ {
field[n].Down = field[n+1]
}
field[len(field)-1].Down = nil
dowidth(bucket)
cmd/compile, runtime: fix placement of map bucket overflow pointer on nacl On most systems, a pointer is the worst case alignment, so adding a pointer field at the end of a struct guarantees there will be no padding added after that field (to satisfy overall struct alignment due to some more-aligned field also present). In the runtime, the map implementation needs a quick way to get to the overflow pointer, which is last in the bucket struct, so it uses size - sizeof(pointer) as the offset. NaCl/amd64p32 is the exception, as always. The worst case alignment is 64 bits but pointers are 32 bits. There's a long history that is not worth going into, but when we moved the overflow pointer to the end of the struct, we didn't get the padding computation right. The compiler computed the regular struct size and then on amd64p32 added another 32-bit field. And the runtime assumed it could step back two 32-bit fields (one 64-bit register size) to get to the overflow pointer. But in fact if the struct needed 64-bit alignment, the computation of the regular struct size would have added a 32-bit pad already, and then the code unconditionally added a second 32-bit pad. This placed the overflow pointer three words from the end, not two. The last two were padding, and since the runtime was consistent about using the second-to-last word as the overflow pointer, no harm done in the sense of overwriting useful memory. But writing the overflow pointer to a non-pointer word of memory means that the GC can't see the overflow blocks, so it will collect them prematurely. Then bad things happen. Correct all this in a few steps: 1. Add an explicit check at the end of the bucket layout in the compiler that the overflow field is last in the struct, never followed by padding. 2. When padding is needed on nacl (not always, just when needed), insert it before the overflow pointer, to preserve the "last in the struct" property. 3. Let the compiler have the final word on the width of the struct, by inserting an explicit padding field instead of overwriting the results of the width computation it does. 4. For the same reason (tell the truth to the compiler), set the type of the overflow field when we're trying to pretend its not a pointer (in this case the runtime maintains a list of the overflow blocks elsewhere). 5. Make the runtime use "last in the struct" as its location algorithm. This fixes TestTraceStress on nacl/amd64p32. The 'bad map state' and 'invalid free list' failures no longer occur. Fixes #11838. Change-Id: If918887f8f252d988db0a35159944d2b36512f92 Reviewed-on: https://go-review.googlesource.com/12971 Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Austin Clements <austin@google.com>
2015-07-30 22:05:51 -04:00
// Double-check that overflow field is final memory in struct,
// with no padding at end. See comment above.
if ovf.Width != bucket.Width-int64(Widthptr) {
Yyerror("bad math in mapbucket for %v", t)
}
t.Bucket = bucket
bucket.Map = t
return bucket
}
// Builds a type representing a Hmap structure for the given map type.
// Make sure this stays in sync with ../../../../runtime/hashmap.go!
func hmap(t *Type) *Type {
if t.Hmap != nil {
return t.Hmap
}
bucket := mapbucket(t)
var field [8]*Type
field[0] = makefield("count", Types[TINT])
field[1] = makefield("flags", Types[TUINT8])
field[2] = makefield("B", Types[TUINT8])
field[3] = makefield("hash0", Types[TUINT32])
field[4] = makefield("buckets", Ptrto(bucket))
field[5] = makefield("oldbuckets", Ptrto(bucket))
field[6] = makefield("nevacuate", Types[TUINTPTR])
field[7] = makefield("overflow", Types[TUNSAFEPTR])
h := typ(TSTRUCT)
h.Noalg = true
h.Local = t.Local
h.Type = field[0]
for n := int32(0); n < int32(len(field)-1); n++ {
field[n].Down = field[n+1]
}
field[len(field)-1].Down = nil
dowidth(h)
t.Hmap = h
h.Map = t
return h
}
func hiter(t *Type) *Type {
if t.Hiter != nil {
return t.Hiter
}
// build a struct:
// hiter {
// key *Key
// val *Value
// t *MapType
// h *Hmap
// buckets *Bucket
// bptr *Bucket
// overflow0 unsafe.Pointer
// overflow1 unsafe.Pointer
// startBucket uintptr
// stuff uintptr
// bucket uintptr
// checkBucket uintptr
// }
// must match ../../../../runtime/hashmap.go:hiter.
var field [12]*Type
field[0] = makefield("key", Ptrto(t.Down))
field[1] = makefield("val", Ptrto(t.Type))
field[2] = makefield("t", Ptrto(Types[TUINT8]))
field[3] = makefield("h", Ptrto(hmap(t)))
field[4] = makefield("buckets", Ptrto(mapbucket(t)))
field[5] = makefield("bptr", Ptrto(mapbucket(t)))
field[6] = makefield("overflow0", Types[TUNSAFEPTR])
field[7] = makefield("overflow1", Types[TUNSAFEPTR])
field[8] = makefield("startBucket", Types[TUINTPTR])
field[9] = makefield("stuff", Types[TUINTPTR]) // offset+wrapped+B+I
field[10] = makefield("bucket", Types[TUINTPTR])
field[11] = makefield("checkBucket", Types[TUINTPTR])
// build iterator struct holding the above fields
i := typ(TSTRUCT)
i.Noalg = true
i.Type = field[0]
for n := int32(0); n < int32(len(field)-1); n++ {
field[n].Down = field[n+1]
}
field[len(field)-1].Down = nil
dowidth(i)
if i.Width != int64(12*Widthptr) {
Yyerror("hash_iter size not correct %d %d", i.Width, 12*Widthptr)
}
t.Hiter = i
i.Map = t
return i
}
// f is method type, with receiver.
// return function type, receiver as first argument (or not).
func methodfunc(f *Type, receiver *Type) *Type {
var in []*Node
if receiver != nil {
d := Nod(ODCLFIELD, nil, nil)
d.Type = receiver
in = append(in, d)
}
var d *Node
for t := getinargx(f).Type; t != nil; t = t.Down {
d = Nod(ODCLFIELD, nil, nil)
d.Type = t.Type
d.Isddd = t.Isddd
in = append(in, d)
}
var out []*Node
for t := getoutargx(f).Type; t != nil; t = t.Down {
d = Nod(ODCLFIELD, nil, nil)
d.Type = t.Type
out = append(out, d)
}
t := functype(nil, in, out)
if f.Nname != nil {
// Link to name of original method function.
t.Nname = f.Nname
}
return t
}
// methods returns the methods of the non-interface type t, sorted by name.
// Generates stub functions as needed.
func methods(t *Type) []*Sig {
// method type
mt := methtype(t, 0)
if mt == nil {
return nil
}
expandmeth(mt)
// type stored in interface word
it := t
if !isdirectiface(it) {
it = Ptrto(t)
}
// make list of methods for t,
// generating code if necessary.
var ms []*Sig
for f := mt.Xmethod; f != nil; f = f.Down {
if f.Etype != TFIELD {
Fatalf("methods: not field %v", f)
}
if f.Type.Etype != TFUNC || f.Type.Thistuple == 0 {
Fatalf("non-method on %v method %v %v\n", mt, f.Sym, f)
}
if getthisx(f.Type).Type == nil {
Fatalf("receiver with no type on %v method %v %v\n", mt, f.Sym, f)
}
if f.Nointerface {
continue
}
method := f.Sym
if method == nil {
continue
}
// get receiver type for this particular method.
// if pointer receiver but non-pointer t and
// this is not an embedded pointer inside a struct,
// method does not apply.
this := getthisx(f.Type).Type.Type
if Isptr[this.Etype] && this.Type == t {
continue
}
if Isptr[this.Etype] && !Isptr[t.Etype] && f.Embedded != 2 && !isifacemethod(f.Type) {
continue
}
var sig Sig
ms = append(ms, &sig)
sig.name = method.Name
if !exportname(method.Name) {
if method.Pkg == nil {
Fatalf("methods: missing package")
}
sig.pkg = method.Pkg
}
sig.isym = methodsym(method, it, 1)
sig.tsym = methodsym(method, t, 0)
sig.type_ = methodfunc(f.Type, t)
sig.mtype = methodfunc(f.Type, nil)
if sig.isym.Flags&SymSiggen == 0 {
sig.isym.Flags |= SymSiggen
if !Eqtype(this, it) || this.Width < Types[Tptr].Width {
compiling_wrappers = 1
genwrapper(it, f, sig.isym, 1)
compiling_wrappers = 0
}
}
if sig.tsym.Flags&SymSiggen == 0 {
sig.tsym.Flags |= SymSiggen
if !Eqtype(this, t) {
compiling_wrappers = 1
genwrapper(t, f, sig.tsym, 0)
compiling_wrappers = 0
}
}
}
sort.Sort(byMethodNameAndPackagePath(ms))
return ms
}
// imethods returns the methods of the interface type t, sorted by name.
func imethods(t *Type) []*Sig {
var methods []*Sig
for f := t.Type; f != nil; f = f.Down {
if f.Etype != TFIELD {
Fatalf("imethods: not field")
}
if f.Type.Etype != TFUNC || f.Sym == nil {
continue
}
method := f.Sym
var sig = Sig{
name: method.Name,
}
if !exportname(method.Name) {
if method.Pkg == nil {
Fatalf("imethods: missing package")
}
sig.pkg = method.Pkg
}
sig.mtype = f.Type
sig.offset = 0
sig.type_ = methodfunc(f.Type, nil)
if n := len(methods); n > 0 {
last := methods[n-1]
if !(siglt(last, &sig)) {
Fatalf("sigcmp vs sortinter %s %s", last.name, sig.name)
}
}
methods = append(methods, &sig)
// Compiler can only refer to wrappers for non-blank methods.
if isblanksym(method) {
continue
}
// NOTE(rsc): Perhaps an oversight that
// IfaceType.Method is not in the reflect data.
// Generate the method body, so that compiled
// code can refer to it.
isym := methodsym(method, t, 0)
if isym.Flags&SymSiggen == 0 {
isym.Flags |= SymSiggen
genwrapper(t, f, isym, 0)
}
}
return methods
}
var dimportpath_gopkg *Pkg
func dimportpath(p *Pkg) {
if p.Pathsym != nil {
return
}
// If we are compiling the runtime package, there are two runtime packages around
// -- localpkg and Runtimepkg. We don't want to produce import path symbols for
// both of them, so just produce one for localpkg.
if myimportpath == "runtime" && p == Runtimepkg {
return
}
if dimportpath_gopkg == nil {
dimportpath_gopkg = mkpkg("go")
dimportpath_gopkg.Name = "go"
}
nam := "importpath." + p.Prefix + "."
n := Nod(ONAME, nil, nil)
n.Sym = Pkglookup(nam, dimportpath_gopkg)
n.Class = PEXTERN
n.Xoffset = 0
p.Pathsym = n.Sym
if p == localpkg {
// Note: myimportpath != "", or else dgopkgpath won't call dimportpath.
gdatastring(n, myimportpath)
} else {
gdatastring(n, p.Path)
}
ggloblsym(n.Sym, int32(Types[TSTRING].Width), obj.DUPOK|obj.RODATA)
}
func dgopkgpath(s *Sym, ot int, pkg *Pkg) int {
if pkg == nil {
return dgostringptr(s, ot, "")
}
if pkg == localpkg && myimportpath == "" {
// If we don't know the full path of the package being compiled (i.e. -p
// was not passed on the compiler command line), emit reference to
// go.importpath.""., which 6l will rewrite using the correct import path.
// Every package that imports this one directly defines the symbol.
var ns *Sym
if ns == nil {
ns = Pkglookup("importpath.\"\".", mkpkg("go"))
}
return dsymptr(s, ot, ns, 0)
}
dimportpath(pkg)
return dsymptr(s, ot, pkg.Pathsym, 0)
}
// uncommonType
// ../../../../runtime/type.go:/uncommonType
func dextratype(sym *Sym, off int, t *Type, ptroff int) int {
m := methods(t)
if t.Sym == nil && len(m) == 0 {
return off
}
// fill in *extraType pointer in header
off = int(Rnd(int64(off), int64(Widthptr)))
dsymptr(sym, ptroff, sym, off)
for _, a := range m {
dtypesym(a.type_)
}
ot := off
s := sym
if t.Sym != nil && t != Types[t.Etype] && t != errortype {
ot = dgopkgpath(s, ot, t.Sym.Pkg)
} else {
ot = dgostringptr(s, ot, "")
}
// slice header
ot = dsymptr(s, ot, s, ot+Widthptr+2*Widthint)
n := len(m)
ot = duintxx(s, ot, uint64(n), Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
// methods
for _, a := range m {
// method
// ../../../../runtime/type.go:/method
ot = dgostringptr(s, ot, a.name)
ot = dgopkgpath(s, ot, a.pkg)
ot = dsymptr(s, ot, dtypesym(a.mtype), 0)
ot = dsymptr(s, ot, dtypesym(a.type_), 0)
if a.isym != nil {
ot = dsymptr(s, ot, a.isym, 0)
} else {
ot = duintptr(s, ot, 0)
}
if a.tsym != nil {
ot = dsymptr(s, ot, a.tsym, 0)
} else {
ot = duintptr(s, ot, 0)
}
}
return ot
}
var kinds = []int{
TINT: obj.KindInt,
TUINT: obj.KindUint,
TINT8: obj.KindInt8,
TUINT8: obj.KindUint8,
TINT16: obj.KindInt16,
TUINT16: obj.KindUint16,
TINT32: obj.KindInt32,
TUINT32: obj.KindUint32,
TINT64: obj.KindInt64,
TUINT64: obj.KindUint64,
TUINTPTR: obj.KindUintptr,
TFLOAT32: obj.KindFloat32,
TFLOAT64: obj.KindFloat64,
TBOOL: obj.KindBool,
TSTRING: obj.KindString,
TPTR32: obj.KindPtr,
TPTR64: obj.KindPtr,
TSTRUCT: obj.KindStruct,
TINTER: obj.KindInterface,
TCHAN: obj.KindChan,
TMAP: obj.KindMap,
TARRAY: obj.KindArray,
TFUNC: obj.KindFunc,
TCOMPLEX64: obj.KindComplex64,
TCOMPLEX128: obj.KindComplex128,
TUNSAFEPTR: obj.KindUnsafePointer,
}
func haspointers(t *Type) bool {
if t.Haspointers != 0 {
return t.Haspointers-1 != 0
}
var ret bool
switch t.Etype {
case TINT,
TUINT,
TINT8,
TUINT8,
TINT16,
TUINT16,
TINT32,
TUINT32,
TINT64,
TUINT64,
TUINTPTR,
TFLOAT32,
TFLOAT64,
TCOMPLEX64,
TCOMPLEX128,
TBOOL:
ret = false
case TARRAY:
if t.Bound < 0 { // slice
ret = true
break
}
if t.Bound == 0 { // empty array
ret = false
break
}
ret = haspointers(t.Type)
case TSTRUCT:
ret = false
for t1 := t.Type; t1 != nil; t1 = t1.Down {
if haspointers(t1.Type) {
ret = true
break
}
}
case TSTRING,
TPTR32,
TPTR64,
TUNSAFEPTR,
TINTER,
TCHAN,
TMAP,
TFUNC:
fallthrough
default:
ret = true
case TFIELD:
Fatalf("haspointers: unexpected type, %v", t)
}
t.Haspointers = 1 + uint8(obj.Bool2int(ret))
return ret
}
// typeptrdata returns the length in bytes of the prefix of t
// containing pointer data. Anything after this offset is scalar data.
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
func typeptrdata(t *Type) int64 {
if !haspointers(t) {
return 0
}
switch t.Etype {
case TPTR32,
TPTR64,
TUNSAFEPTR,
TFUNC,
TCHAN,
TMAP:
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return int64(Widthptr)
case TSTRING:
// struct { byte *str; intgo len; }
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return int64(Widthptr)
case TINTER:
// struct { Itab *tab; void *data; } or
// struct { Type *type; void *data; }
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return 2 * int64(Widthptr)
case TARRAY:
if Isslice(t) {
// struct { byte *array; uintgo len; uintgo cap; }
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return int64(Widthptr)
}
// haspointers already eliminated t.Bound == 0.
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return (t.Bound-1)*t.Type.Width + typeptrdata(t.Type)
case TSTRUCT:
// Find the last field that has pointers.
var lastPtrField *Type
for t1 := t.Type; t1 != nil; t1 = t1.Down {
if haspointers(t1.Type) {
lastPtrField = t1
}
}
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
return lastPtrField.Width + typeptrdata(lastPtrField.Type)
default:
Fatalf("typeptrdata: unexpected type, %v", t)
return 0
}
}
// commonType
// ../../runtime/type.go:/commonType
var dcommontype_algarray *Sym
func dcommontype(s *Sym, ot int, t *Type) int {
if ot != 0 {
Fatalf("dcommontype %d", ot)
}
sizeofAlg := 2 * Widthptr
if dcommontype_algarray == nil {
dcommontype_algarray = Pkglookup("algarray", Runtimepkg)
}
dowidth(t)
alg := algtype(t)
var algsym *Sym
if alg < 0 || alg == AMEM {
algsym = dalgsym(t)
}
tptr := Ptrto(t)
if !Isptr[t.Etype] && (t.Sym != nil || methods(tptr) != nil) {
sptr := dtypesym(tptr)
r := obj.Addrel(Linksym(s))
r.Off = 0
r.Siz = 0
r.Sym = sptr.Lsym
r.Type = obj.R_USETYPE
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
gcsym, useGCProg, ptrdata := dgcsym(t)
// ../../../../reflect/type.go:/^type.rtype
// actual type structure
// type rtype struct {
// size uintptr
// ptrdata uintptr
// hash uint32
// _ uint8
// align uint8
// fieldAlign uint8
// kind uint8
// alg *typeAlg
// gcdata *byte
// string *string
// *uncommonType
// }
ot = duintptr(s, ot, uint64(t.Width))
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
ot = duintptr(s, ot, uint64(ptrdata))
ot = duint32(s, ot, typehash(t))
ot = duint8(s, ot, 0) // unused
// runtime (and common sense) expects alignment to be a power of two.
i := int(t.Align)
if i == 0 {
i = 1
}
if i&(i-1) != 0 {
Fatalf("invalid alignment %d for %v", t.Align, t)
}
ot = duint8(s, ot, t.Align) // align
ot = duint8(s, ot, t.Align) // fieldAlign
i = kinds[t.Etype]
if t.Etype == TARRAY && t.Bound < 0 {
i = obj.KindSlice
}
if !haspointers(t) {
i |= obj.KindNoPointers
}
if isdirectiface(t) {
i |= obj.KindDirectIface
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
if useGCProg {
i |= obj.KindGCProg
}
ot = duint8(s, ot, uint8(i)) // kind
if algsym == nil {
ot = dsymptr(s, ot, dcommontype_algarray, alg*sizeofAlg)
} else {
ot = dsymptr(s, ot, algsym, 0)
}
ot = dsymptr(s, ot, gcsym, 0) // gcdata
p := Tconv(t, obj.FmtLeft|obj.FmtUnsigned)
// If we're writing out type T,
// we are very likely to write out type *T as well.
// Use the string "*T"[1:] for "T", so that the two
// share storage. This is a cheap way to reduce the
// amount of space taken up by reflect strings.
prefix := 0
if !strings.HasPrefix(p, "*") {
p = "*"+p
prefix = 1
}
_, symdata := stringsym(p) // string
ot = dsymptr(s, ot, symdata, prefix)
ot = duintxx(s, ot, uint64(len(p)-prefix), Widthint)
//fmt.Printf("dcommontype: %s\n", p)
// skip pointer to extraType,
// which follows the rest of this type structure.
// caller will fill in if needed.
// otherwise linker will assume 0.
ot += Widthptr
return ot
}
func typesym(t *Type) *Sym {
return Pkglookup(Tconv(t, obj.FmtLeft), typepkg)
}
func tracksym(t *Type) *Sym {
return Pkglookup(Tconv(t.Outer, obj.FmtLeft)+"."+t.Sym.Name, trackpkg)
}
func typelinksym(t *Type) *Sym {
// %-uT is what the generated Type's string field says.
// It uses (ambiguous) package names instead of import paths.
// %-T is the complete, unambiguous type name.
// We want the types to end up sorted by string field,
// so use that first in the name, and then add :%-T to
// disambiguate. We use a tab character as the separator to
// ensure the types appear sorted by their string field. The
// names are a little long but they are discarded by the linker
// and do not end up in the symbol table of the final binary.
p := Tconv(t, obj.FmtLeft|obj.FmtUnsigned) + "\t" + Tconv(t, obj.FmtLeft)
s := Pkglookup(p, typelinkpkg)
//print("typelinksym: %s -> %+S\n", p, s);
return s
}
func typesymprefix(prefix string, t *Type) *Sym {
p := prefix + "." + Tconv(t, obj.FmtLeft)
s := Pkglookup(p, typepkg)
//print("algsym: %s -> %+S\n", p, s);
return s
}
func typenamesym(t *Type) *Sym {
if t == nil || (Isptr[t.Etype] && t.Type == nil) || isideal(t) {
Fatalf("typename %v", t)
}
s := typesym(t)
if s.Def == nil {
n := Nod(ONAME, nil, nil)
n.Sym = s
n.Type = Types[TUINT8]
n.Addable = true
n.Ullman = 1
n.Class = PEXTERN
n.Xoffset = 0
n.Typecheck = 1
s.Def = n
signatlist = list(signatlist, typenod(t))
}
return s.Def.Sym
}
func typename(t *Type) *Node {
s := typenamesym(t)
n := Nod(OADDR, s.Def, nil)
n.Type = Ptrto(s.Def.Type)
n.Addable = true
n.Ullman = 2
n.Typecheck = 1
return n
}
// isreflexive reports whether t has a reflexive equality operator.
// That is, if x==x for all x of type t.
func isreflexive(t *Type) bool {
switch t.Etype {
case TBOOL,
TINT,
TUINT,
TINT8,
TUINT8,
TINT16,
TUINT16,
TINT32,
TUINT32,
TINT64,
TUINT64,
TUINTPTR,
TPTR32,
TPTR64,
TUNSAFEPTR,
TSTRING,
TCHAN:
return true
case TFLOAT32,
TFLOAT64,
TCOMPLEX64,
TCOMPLEX128,
TINTER:
return false
case TARRAY:
if Isslice(t) {
Fatalf("slice can't be a map key: %v", t)
}
return isreflexive(t.Type)
case TSTRUCT:
for t1 := t.Type; t1 != nil; t1 = t1.Down {
if !isreflexive(t1.Type) {
return false
}
}
return true
default:
Fatalf("bad type for map key: %v", t)
return false
}
}
// needkeyupdate reports whether map updates with t as a key
// need the key to be updated.
func needkeyupdate(t *Type) bool {
switch t.Etype {
case TBOOL,
TINT,
TUINT,
TINT8,
TUINT8,
TINT16,
TUINT16,
TINT32,
TUINT32,
TINT64,
TUINT64,
TUINTPTR,
TPTR32,
TPTR64,
TUNSAFEPTR,
TCHAN:
return false
case TFLOAT32, // floats can be +0/-0
TFLOAT64,
TCOMPLEX64,
TCOMPLEX128,
TINTER,
TSTRING: // strings might have smaller backing stores
return true
case TARRAY:
if Isslice(t) {
Fatalf("slice can't be a map key: %v", t)
}
return needkeyupdate(t.Type)
case TSTRUCT:
for t1 := t.Type; t1 != nil; t1 = t1.Down {
if needkeyupdate(t1.Type) {
return true
}
}
return false
default:
Fatalf("bad type for map key: %v", t)
return true
}
}
func dtypesym(t *Type) *Sym {
// Replace byte, rune aliases with real type.
// They've been separate internally to make error messages
// better, but we have to merge them in the reflect tables.
if t == bytetype || t == runetype {
t = Types[t.Etype]
}
if isideal(t) {
Fatalf("dtypesym %v", t)
}
s := typesym(t)
if s.Flags&SymSiggen != 0 {
return s
}
s.Flags |= SymSiggen
// special case (look for runtime below):
// when compiling package runtime,
// emit the type structures for int, float, etc.
tbase := t
if Isptr[t.Etype] && t.Sym == nil && t.Type.Sym != nil {
tbase = t.Type
}
dupok := 0
if tbase.Sym == nil {
dupok = obj.DUPOK
}
if myimportpath == "runtime" && (tbase == Types[tbase.Etype] || tbase == bytetype || tbase == runetype || tbase == errortype) { // int, float, etc
goto ok
}
// named types from other files are defined only by those files
if tbase.Sym != nil && !tbase.Local {
return s
}
if isforw[tbase.Etype] {
return s
}
ok:
ot := 0
xt := 0
switch t.Etype {
default:
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
case TARRAY:
if t.Bound >= 0 {
// ../../../../runtime/type.go:/arrayType
s1 := dtypesym(t.Type)
t2 := typ(TARRAY)
t2.Type = t.Type
t2.Bound = -1 // slice
s2 := dtypesym(t2)
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s1, 0)
ot = dsymptr(s, ot, s2, 0)
ot = duintptr(s, ot, uint64(t.Bound))
} else {
// ../../../../runtime/type.go:/sliceType
s1 := dtypesym(t.Type)
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s1, 0)
}
// ../../../../runtime/type.go:/chanType
case TCHAN:
s1 := dtypesym(t.Type)
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s1, 0)
ot = duintptr(s, ot, uint64(t.Chan))
case TFUNC:
for t1 := getthisx(t).Type; t1 != nil; t1 = t1.Down {
dtypesym(t1.Type)
}
isddd := false
for t1 := getinargx(t).Type; t1 != nil; t1 = t1.Down {
isddd = t1.Isddd
dtypesym(t1.Type)
}
for t1 := getoutargx(t).Type; t1 != nil; t1 = t1.Down {
dtypesym(t1.Type)
}
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = duint8(s, ot, uint8(obj.Bool2int(isddd)))
// two slice headers: in and out.
ot = int(Rnd(int64(ot), int64(Widthptr)))
ot = dsymptr(s, ot, s, ot+2*(Widthptr+2*Widthint))
n := t.Thistuple + t.Intuple
ot = duintxx(s, ot, uint64(n), Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
ot = dsymptr(s, ot, s, ot+1*(Widthptr+2*Widthint)+n*Widthptr)
ot = duintxx(s, ot, uint64(t.Outtuple), Widthint)
ot = duintxx(s, ot, uint64(t.Outtuple), Widthint)
// slice data
for t1 := getthisx(t).Type; t1 != nil; t1 = t1.Down {
ot = dsymptr(s, ot, dtypesym(t1.Type), 0)
n++
}
for t1 := getinargx(t).Type; t1 != nil; t1 = t1.Down {
ot = dsymptr(s, ot, dtypesym(t1.Type), 0)
n++
}
for t1 := getoutargx(t).Type; t1 != nil; t1 = t1.Down {
ot = dsymptr(s, ot, dtypesym(t1.Type), 0)
n++
}
case TINTER:
m := imethods(t)
n := len(m)
for _, a := range m {
dtypesym(a.type_)
}
// ../../../../runtime/type.go:/interfaceType
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s, ot+Widthptr+2*Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
for _, a := range m {
// ../../../../runtime/type.go:/imethod
ot = dgostringptr(s, ot, a.name)
ot = dgopkgpath(s, ot, a.pkg)
ot = dsymptr(s, ot, dtypesym(a.type_), 0)
}
// ../../../../runtime/type.go:/mapType
case TMAP:
s1 := dtypesym(t.Down)
s2 := dtypesym(t.Type)
s3 := dtypesym(mapbucket(t))
s4 := dtypesym(hmap(t))
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s1, 0)
ot = dsymptr(s, ot, s2, 0)
ot = dsymptr(s, ot, s3, 0)
ot = dsymptr(s, ot, s4, 0)
if t.Down.Width > MAXKEYSIZE {
ot = duint8(s, ot, uint8(Widthptr))
ot = duint8(s, ot, 1) // indirect
} else {
ot = duint8(s, ot, uint8(t.Down.Width))
ot = duint8(s, ot, 0) // not indirect
}
if t.Type.Width > MAXVALSIZE {
ot = duint8(s, ot, uint8(Widthptr))
ot = duint8(s, ot, 1) // indirect
} else {
ot = duint8(s, ot, uint8(t.Type.Width))
ot = duint8(s, ot, 0) // not indirect
}
ot = duint16(s, ot, uint16(mapbucket(t).Width))
ot = duint8(s, ot, uint8(obj.Bool2int(isreflexive(t.Down))))
ot = duint8(s, ot, uint8(obj.Bool2int(needkeyupdate(t.Down))))
case TPTR32, TPTR64:
if t.Type.Etype == TANY {
// ../../../../runtime/type.go:/UnsafePointerType
ot = dcommontype(s, ot, t)
break
}
// ../../../../runtime/type.go:/ptrType
s1 := dtypesym(t.Type)
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s1, 0)
// ../../../../runtime/type.go:/structType
// for security, only the exported fields.
case TSTRUCT:
n := 0
for t1 := t.Type; t1 != nil; t1 = t1.Down {
dtypesym(t1.Type)
n++
}
ot = dcommontype(s, ot, t)
xt = ot - 1*Widthptr
ot = dsymptr(s, ot, s, ot+Widthptr+2*Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
ot = duintxx(s, ot, uint64(n), Widthint)
for t1 := t.Type; t1 != nil; t1 = t1.Down {
// ../../../../runtime/type.go:/structField
if t1.Sym != nil && t1.Embedded == 0 {
ot = dgostringptr(s, ot, t1.Sym.Name)
if exportname(t1.Sym.Name) {
ot = dgostringptr(s, ot, "")
} else {
ot = dgopkgpath(s, ot, t1.Sym.Pkg)
}
} else {
ot = dgostringptr(s, ot, "")
if t1.Type.Sym != nil &&
(t1.Type.Sym.Pkg == builtinpkg || !exportname(t1.Type.Sym.Name)) {
ot = dgopkgpath(s, ot, localpkg)
} else {
ot = dgostringptr(s, ot, "")
}
}
ot = dsymptr(s, ot, dtypesym(t1.Type), 0)
ot = dgostrlitptr(s, ot, t1.Note)
ot = duintptr(s, ot, uint64(t1.Width)) // field offset
}
}
ot = dextratype(s, ot, t, xt)
ggloblsym(s, int32(ot), int16(dupok|obj.RODATA))
// generate typelink.foo pointing at s = type.foo.
// The linker will leave a table of all the typelinks for
// types in the binary, so reflect can find them.
// We only need the link for unnamed composites that
// we want be able to find.
if t.Sym == nil {
switch t.Etype {
case TPTR32, TPTR64, TARRAY, TCHAN, TFUNC, TMAP:
slink := typelinksym(t)
dsymptr(slink, 0, s, 0)
ggloblsym(slink, int32(Widthptr), int16(dupok|obj.RODATA))
}
}
return s
}
func dumptypestructs() {
var n *Node
// copy types from externdcl list to signatlist
for _, n := range externdcl {
if n.Op != OTYPE {
continue
}
signatlist = list(signatlist, n)
}
// process signatlist
var t *Type
for l := signatlist; l != nil; l = l.Next {
n = l.N
if n.Op != OTYPE {
continue
}
t = n.Type
dtypesym(t)
if t.Sym != nil {
dtypesym(Ptrto(t))
}
}
// generate import strings for imported packages
for _, p := range pkgs {
if p.Direct {
dimportpath(p)
}
}
// do basic types if compiling package runtime.
// they have to be in at least one package,
// and runtime is always loaded implicitly,
// so this is as good as any.
// another possible choice would be package main,
// but using runtime means fewer copies in .6 files.
if myimportpath == "runtime" {
for i := EType(1); i <= TBOOL; i++ {
dtypesym(Ptrto(Types[i]))
}
dtypesym(Ptrto(Types[TSTRING]))
dtypesym(Ptrto(Types[TUNSAFEPTR]))
// emit type structs for error and func(error) string.
// The latter is the type of an auto-generated wrapper.
dtypesym(Ptrto(errortype))
dtypesym(functype(nil, []*Node{Nod(ODCLFIELD, nil, typenod(errortype))}, []*Node{Nod(ODCLFIELD, nil, typenod(Types[TSTRING]))}))
// add paths for runtime and main, which 6l imports implicitly.
dimportpath(Runtimepkg)
if flag_race != 0 {
dimportpath(racepkg)
}
if flag_msan != 0 {
dimportpath(msanpkg)
}
dimportpath(mkpkg("main"))
}
}
func dalgsym(t *Type) *Sym {
var s *Sym
var hashfunc *Sym
var eqfunc *Sym
// dalgsym is only called for a type that needs an algorithm table,
// which implies that the type is comparable (or else it would use ANOEQ).
if algtype(t) == AMEM {
// we use one algorithm table for all AMEM types of a given size
p := fmt.Sprintf(".alg%d", t.Width)
s = Pkglookup(p, typepkg)
if s.Flags&SymAlgGen != 0 {
return s
}
s.Flags |= SymAlgGen
// make hash closure
p = fmt.Sprintf(".hashfunc%d", t.Width)
hashfunc = Pkglookup(p, typepkg)
ot := 0
ot = dsymptr(hashfunc, ot, Pkglookup("memhash_varlen", Runtimepkg), 0)
ot = duintxx(hashfunc, ot, uint64(t.Width), Widthptr) // size encoded in closure
ggloblsym(hashfunc, int32(ot), obj.DUPOK|obj.RODATA)
// make equality closure
p = fmt.Sprintf(".eqfunc%d", t.Width)
eqfunc = Pkglookup(p, typepkg)
ot = 0
ot = dsymptr(eqfunc, ot, Pkglookup("memequal_varlen", Runtimepkg), 0)
ot = duintxx(eqfunc, ot, uint64(t.Width), Widthptr)
ggloblsym(eqfunc, int32(ot), obj.DUPOK|obj.RODATA)
} else {
// generate an alg table specific to this type
s = typesymprefix(".alg", t)
hash := typesymprefix(".hash", t)
eq := typesymprefix(".eq", t)
hashfunc = typesymprefix(".hashfunc", t)
eqfunc = typesymprefix(".eqfunc", t)
genhash(hash, t)
geneq(eq, t)
// make Go funcs (closures) for calling hash and equal from Go
dsymptr(hashfunc, 0, hash, 0)
ggloblsym(hashfunc, int32(Widthptr), obj.DUPOK|obj.RODATA)
dsymptr(eqfunc, 0, eq, 0)
ggloblsym(eqfunc, int32(Widthptr), obj.DUPOK|obj.RODATA)
}
// ../../../../runtime/alg.go:/typeAlg
ot := 0
ot = dsymptr(s, ot, hashfunc, 0)
ot = dsymptr(s, ot, eqfunc, 0)
ggloblsym(s, int32(ot), obj.DUPOK|obj.RODATA)
return s
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// maxPtrmaskBytes is the maximum length of a GC ptrmask bitmap,
// which holds 1-bit entries describing where pointers are in a given type.
// 16 bytes is enough to describe 128 pointer-sized words, 512 or 1024 bytes
// depending on the system. Above this length, the GC information is
// recorded as a GC program, which can express repetition compactly.
// In either form, the information is used by the runtime to initialize the
// heap bitmap, and for large types (like 128 or more words), they are
// roughly the same speed. GC programs are never much larger and often
// more compact. (If large arrays are involved, they can be arbitrarily more
// compact.)
//
// The cutoff must be large enough that any allocation large enough to
// use a GC program is large enough that it does not share heap bitmap
// bytes with any other objects, allowing the GC program execution to
// assume an aligned start and not use atomic operations. In the current
// runtime, this means all malloc size classes larger than the cutoff must
// be multiples of four words. On 32-bit systems that's 16 bytes, and
// all size classes >= 16 bytes are 16-byte aligned, so no real constraint.
// On 64-bit systems, that's 32 bytes, and 32-byte alignment is guaranteed
// for size classes >= 256 bytes. On a 64-bit system, 256 bytes allocated
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// is 32 pointers, the bits for which fit in 4 bytes. So maxPtrmaskBytes
// must be >= 4.
//
// We used to use 16 because the GC programs do have some constant overhead
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// to get started, and processing 128 pointers seems to be enough to
// amortize that overhead well.
//
// To make sure that the runtime's chansend can call typeBitsBulkBarrier,
// we raised the limit to 2048, so that even 32-bit systems are guaranteed to
// use bitmaps for objects up to 64 kB in size.
//
// Also known to reflect/type.go.
//
const maxPtrmaskBytes = 2048
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// dgcsym emits and returns a data symbol containing GC information for type t,
// along with a boolean reporting whether the UseGCProg bit should be set in
// the type kind, and the ptrdata field to record in the reflect type information.
func dgcsym(t *Type) (sym *Sym, useGCProg bool, ptrdata int64) {
ptrdata = typeptrdata(t)
if ptrdata/int64(Widthptr) <= maxPtrmaskBytes*8 {
sym = dgcptrmask(t)
return
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
useGCProg = true
sym, ptrdata = dgcprog(t)
return
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// dgcptrmask emits and returns the symbol containing a pointer mask for type t.
func dgcptrmask(t *Type) *Sym {
ptrmask := make([]byte, (typeptrdata(t)/int64(Widthptr)+7)/8)
fillptrmask(t, ptrmask)
p := fmt.Sprintf("gcbits.%x", ptrmask)
sym := Pkglookup(p, Runtimepkg)
if sym.Flags&SymUniq == 0 {
sym.Flags |= SymUniq
for i, x := range ptrmask {
duint8(sym, i, x)
}
ggloblsym(sym, int32(len(ptrmask)), obj.DUPOK|obj.RODATA|obj.LOCAL)
}
return sym
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// fillptrmask fills in ptrmask with 1s corresponding to the
// word offsets in t that hold pointers.
// ptrmask is assumed to fit at least typeptrdata(t)/Widthptr bits.
func fillptrmask(t *Type, ptrmask []byte) {
for i := range ptrmask {
ptrmask[i] = 0
}
if !haspointers(t) {
return
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
vec := bvalloc(8 * int32(len(ptrmask)))
xoffset := int64(0)
cmd/internal/gc, runtime: use 1-bit bitmap for stack frames, data, bss The bitmaps were 2 bits per pointer because we needed to distinguish scalar, pointer, multiword, and we used the leftover value to distinguish uninitialized from scalar, even though the garbage collector (GC) didn't care. Now that there are no multiword structures from the GC's point of view, cut the bitmaps down to 1 bit per pointer, recording just live pointer vs not. The GC assumes the same layout for stack frames and for the maps describing the global data and bss sections, so change them all in one CL. The code still refers to 4-bit heap bitmaps and 2-bit "type bitmaps", since the 2-bit representation lives (at least for now) in some of the reflect data. Because these stack frame bitmaps are stored directly in the rodata in the binary, this CL reduces the size of the 6g binary by about 1.1%. Performance change is basically a wash, but using less memory, and smaller binaries, and enables other bitmap reductions. name old mean new mean delta BenchmarkBinaryTree17 13.2s × (0.97,1.03) 13.0s × (0.99,1.01) -0.93% (p=0.005) BenchmarkBinaryTree17-2 9.69s × (0.96,1.05) 9.51s × (0.96,1.03) -1.86% (p=0.001) BenchmarkBinaryTree17-4 10.1s × (0.97,1.05) 10.0s × (0.96,1.05) ~ (p=0.141) BenchmarkFannkuch11 4.35s × (0.99,1.01) 4.43s × (0.98,1.04) +1.75% (p=0.001) BenchmarkFannkuch11-2 4.31s × (0.99,1.03) 4.32s × (1.00,1.00) ~ (p=0.095) BenchmarkFannkuch11-4 4.32s × (0.99,1.02) 4.38s × (0.98,1.04) +1.38% (p=0.008) BenchmarkFmtFprintfEmpty 83.5ns × (0.97,1.10) 87.3ns × (0.92,1.11) +4.55% (p=0.014) BenchmarkFmtFprintfEmpty-2 81.8ns × (0.98,1.04) 82.5ns × (0.97,1.08) ~ (p=0.364) BenchmarkFmtFprintfEmpty-4 80.9ns × (0.99,1.01) 82.6ns × (0.97,1.08) +2.12% (p=0.010) BenchmarkFmtFprintfString 320ns × (0.95,1.04) 322ns × (0.97,1.05) ~ (p=0.368) BenchmarkFmtFprintfString-2 303ns × (0.97,1.04) 304ns × (0.97,1.04) ~ (p=0.484) BenchmarkFmtFprintfString-4 305ns × (0.97,1.05) 306ns × (0.98,1.05) ~ (p=0.543) BenchmarkFmtFprintfInt 311ns × (0.98,1.03) 319ns × (0.97,1.03) +2.63% (p=0.000) BenchmarkFmtFprintfInt-2 297ns × (0.98,1.04) 301ns × (0.97,1.04) +1.19% (p=0.023) BenchmarkFmtFprintfInt-4 302ns × (0.98,1.02) 304ns × (0.97,1.03) ~ (p=0.126) BenchmarkFmtFprintfIntInt 554ns × (0.96,1.05) 554ns × (0.97,1.03) ~ (p=0.975) BenchmarkFmtFprintfIntInt-2 520ns × (0.98,1.03) 517ns × (0.98,1.02) ~ (p=0.153) BenchmarkFmtFprintfIntInt-4 524ns × (0.98,1.02) 525ns × (0.98,1.03) ~ (p=0.597) BenchmarkFmtFprintfPrefixedInt 433ns × (0.97,1.06) 434ns × (0.97,1.06) ~ (p=0.804) BenchmarkFmtFprintfPrefixedInt-2 413ns × (0.98,1.04) 413ns × (0.98,1.03) ~ (p=0.881) BenchmarkFmtFprintfPrefixedInt-4 420ns × (0.97,1.03) 421ns × (0.97,1.03) ~ (p=0.561) BenchmarkFmtFprintfFloat 620ns × (0.99,1.03) 636ns × (0.97,1.03) +2.57% (p=0.000) BenchmarkFmtFprintfFloat-2 601ns × (0.98,1.02) 617ns × (0.98,1.03) +2.58% (p=0.000) BenchmarkFmtFprintfFloat-4 613ns × (0.98,1.03) 626ns × (0.98,1.02) +2.15% (p=0.000) BenchmarkFmtManyArgs 2.19µs × (0.96,1.04) 2.23µs × (0.97,1.02) +1.65% (p=0.000) BenchmarkFmtManyArgs-2 2.08µs × (0.98,1.03) 2.10µs × (0.99,1.02) +0.79% (p=0.019) BenchmarkFmtManyArgs-4 2.10µs × (0.98,1.02) 2.13µs × (0.98,1.02) +1.72% (p=0.000) BenchmarkGobDecode 21.3ms × (0.97,1.05) 21.1ms × (0.97,1.04) -1.36% (p=0.025) BenchmarkGobDecode-2 20.0ms × (0.97,1.03) 19.2ms × (0.97,1.03) -4.00% (p=0.000) BenchmarkGobDecode-4 19.5ms × (0.99,1.02) 19.0ms × (0.99,1.01) -2.39% (p=0.000) BenchmarkGobEncode 18.3ms × (0.95,1.07) 18.1ms × (0.96,1.08) ~ (p=0.305) BenchmarkGobEncode-2 16.8ms × (0.97,1.02) 16.4ms × (0.98,1.02) -2.79% (p=0.000) BenchmarkGobEncode-4 15.4ms × (0.98,1.02) 15.4ms × (0.98,1.02) ~ (p=0.465) BenchmarkGzip 650ms × (0.98,1.03) 655ms × (0.97,1.04) ~ (p=0.075) BenchmarkGzip-2 652ms × (0.98,1.03) 655ms × (0.98,1.02) ~ (p=0.337) BenchmarkGzip-4 656ms × (0.98,1.04) 653ms × (0.98,1.03) ~ (p=0.291) BenchmarkGunzip 143ms × (1.00,1.01) 143ms × (1.00,1.01) ~ (p=0.507) BenchmarkGunzip-2 143ms × (1.00,1.01) 143ms × (1.00,1.01) ~ (p=0.313) BenchmarkGunzip-4 143ms × (1.00,1.01) 143ms × (1.00,1.01) ~ (p=0.312) BenchmarkHTTPClientServer 110µs × (0.98,1.03) 109µs × (0.99,1.02) -1.40% (p=0.000) BenchmarkHTTPClientServer-2 154µs × (0.90,1.08) 149µs × (0.90,1.08) -3.43% (p=0.007) BenchmarkHTTPClientServer-4 138µs × (0.97,1.04) 138µs × (0.96,1.04) ~ (p=0.670) BenchmarkJSONEncode 40.2ms × (0.98,1.02) 40.2ms × (0.98,1.05) ~ (p=0.828) BenchmarkJSONEncode-2 35.1ms × (0.99,1.02) 35.2ms × (0.98,1.03) ~ (p=0.392) BenchmarkJSONEncode-4 35.3ms × (0.98,1.03) 35.3ms × (0.98,1.02) ~ (p=0.813) BenchmarkJSONDecode 119ms × (0.97,1.02) 117ms × (0.98,1.02) -1.80% (p=0.000) BenchmarkJSONDecode-2 115ms × (0.99,1.02) 114ms × (0.98,1.02) -1.18% (p=0.000) BenchmarkJSONDecode-4 116ms × (0.98,1.02) 114ms × (0.98,1.02) -1.43% (p=0.000) BenchmarkMandelbrot200 6.03ms × (1.00,1.01) 6.03ms × (1.00,1.01) ~ (p=0.985) BenchmarkMandelbrot200-2 6.03ms × (1.00,1.01) 6.02ms × (1.00,1.01) ~ (p=0.320) BenchmarkMandelbrot200-4 6.03ms × (1.00,1.01) 6.03ms × (1.00,1.01) ~ (p=0.799) BenchmarkGoParse 8.63ms × (0.89,1.10) 8.58ms × (0.93,1.09) ~ (p=0.667) BenchmarkGoParse-2 8.20ms × (0.97,1.04) 8.37ms × (0.97,1.04) +1.96% (p=0.001) BenchmarkGoParse-4 8.00ms × (0.98,1.02) 8.14ms × (0.99,1.02) +1.75% (p=0.000) BenchmarkRegexpMatchEasy0_32 162ns × (1.00,1.01) 164ns × (0.98,1.04) +1.35% (p=0.011) BenchmarkRegexpMatchEasy0_32-2 161ns × (1.00,1.01) 161ns × (1.00,1.00) ~ (p=0.185) BenchmarkRegexpMatchEasy0_32-4 161ns × (1.00,1.00) 161ns × (1.00,1.00) -0.19% (p=0.001) BenchmarkRegexpMatchEasy0_1K 540ns × (0.99,1.02) 566ns × (0.98,1.04) +4.98% (p=0.000) BenchmarkRegexpMatchEasy0_1K-2 540ns × (0.99,1.01) 557ns × (0.99,1.01) +3.21% (p=0.000) BenchmarkRegexpMatchEasy0_1K-4 541ns × (0.99,1.01) 559ns × (0.99,1.01) +3.26% (p=0.000) BenchmarkRegexpMatchEasy1_32 139ns × (0.98,1.04) 139ns × (0.99,1.03) ~ (p=0.979) BenchmarkRegexpMatchEasy1_32-2 139ns × (0.99,1.04) 139ns × (0.99,1.02) ~ (p=0.777) BenchmarkRegexpMatchEasy1_32-4 139ns × (0.98,1.04) 139ns × (0.99,1.04) ~ (p=0.771) BenchmarkRegexpMatchEasy1_1K 890ns × (0.99,1.03) 885ns × (1.00,1.01) -0.50% (p=0.004) BenchmarkRegexpMatchEasy1_1K-2 888ns × (0.99,1.01) 885ns × (0.99,1.01) -0.37% (p=0.004) BenchmarkRegexpMatchEasy1_1K-4 890ns × (0.99,1.02) 884ns × (1.00,1.00) -0.70% (p=0.000) BenchmarkRegexpMatchMedium_32 252ns × (0.99,1.01) 251ns × (0.99,1.01) ~ (p=0.081) BenchmarkRegexpMatchMedium_32-2 254ns × (0.99,1.04) 252ns × (0.99,1.01) -0.78% (p=0.027) BenchmarkRegexpMatchMedium_32-4 253ns × (0.99,1.04) 252ns × (0.99,1.01) -0.70% (p=0.022) BenchmarkRegexpMatchMedium_1K 72.9µs × (0.99,1.01) 72.7µs × (1.00,1.00) ~ (p=0.064) BenchmarkRegexpMatchMedium_1K-2 74.1µs × (0.98,1.05) 72.9µs × (1.00,1.01) -1.61% (p=0.001) BenchmarkRegexpMatchMedium_1K-4 73.6µs × (0.99,1.05) 72.8µs × (1.00,1.00) -1.13% (p=0.007) BenchmarkRegexpMatchHard_32 3.88µs × (0.99,1.03) 3.92µs × (0.98,1.05) ~ (p=0.143) BenchmarkRegexpMatchHard_32-2 3.89µs × (0.99,1.03) 3.93µs × (0.98,1.09) ~ (p=0.278) BenchmarkRegexpMatchHard_32-4 3.90µs × (0.99,1.05) 3.93µs × (0.98,1.05) ~ (p=0.252) BenchmarkRegexpMatchHard_1K 118µs × (0.99,1.01) 117µs × (0.99,1.02) -0.54% (p=0.003) BenchmarkRegexpMatchHard_1K-2 118µs × (0.99,1.01) 118µs × (0.99,1.03) ~ (p=0.581) BenchmarkRegexpMatchHard_1K-4 118µs × (0.99,1.02) 117µs × (0.99,1.01) -0.54% (p=0.002) BenchmarkRevcomp 991ms × (0.95,1.10) 989ms × (0.94,1.08) ~ (p=0.879) BenchmarkRevcomp-2 978ms × (0.95,1.11) 962ms × (0.96,1.08) ~ (p=0.257) BenchmarkRevcomp-4 979ms × (0.96,1.07) 974ms × (0.96,1.11) ~ (p=0.678) BenchmarkTemplate 141ms × (0.99,1.02) 145ms × (0.99,1.02) +2.75% (p=0.000) BenchmarkTemplate-2 135ms × (0.98,1.02) 138ms × (0.99,1.02) +2.34% (p=0.000) BenchmarkTemplate-4 136ms × (0.98,1.02) 140ms × (0.99,1.02) +2.71% (p=0.000) BenchmarkTimeParse 640ns × (0.99,1.01) 622ns × (0.99,1.01) -2.88% (p=0.000) BenchmarkTimeParse-2 640ns × (0.99,1.01) 622ns × (1.00,1.00) -2.81% (p=0.000) BenchmarkTimeParse-4 640ns × (1.00,1.01) 622ns × (0.99,1.01) -2.82% (p=0.000) BenchmarkTimeFormat 730ns × (0.98,1.02) 731ns × (0.98,1.03) ~ (p=0.767) BenchmarkTimeFormat-2 709ns × (0.99,1.02) 707ns × (0.99,1.02) ~ (p=0.347) BenchmarkTimeFormat-4 717ns × (0.98,1.01) 718ns × (0.98,1.02) ~ (p=0.793) Change-Id: Ie779c47e912bf80eb918bafa13638bd8dfd6c2d9 Reviewed-on: https://go-review.googlesource.com/9406 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-04-27 22:45:57 -04:00
onebitwalktype1(t, &xoffset, vec)
runtime: reintroduce ``dead'' space during GC scan Reintroduce an optimization discarded during the initial conversion from 4-bit heap bitmaps to 2-bit heap bitmaps: when we reach the place in the bitmap where there are no more pointers, mark that position for the GC so that it can avoid scanning past that place. During heapBitsSetType we can also avoid initializing heap bitmap beyond that location, which gives a bit of a win compared to Go 1.4. This particular optimization (not initializing the heap bitmap) may not last: we might change typedmemmove to use the heap bitmap, in which case it would all need to be initialized. The early stop in the GC scan will stay no matter what. Compared to Go 1.4 (github.com/rsc/go, branch go14bench): name old mean new mean delta SetTypeNode64 80.7ns × (1.00,1.01) 57.4ns × (1.00,1.01) -28.83% (p=0.000) SetTypeNode64Dead 80.5ns × (1.00,1.01) 13.1ns × (0.99,1.02) -83.77% (p=0.000) SetTypeNode64Slice 2.16µs × (1.00,1.01) 1.54µs × (1.00,1.01) -28.75% (p=0.000) SetTypeNode64DeadSlice 2.16µs × (1.00,1.01) 1.52µs × (1.00,1.00) -29.74% (p=0.000) Compared to previous CL: name old mean new mean delta SetTypeNode64 56.7ns × (1.00,1.00) 57.4ns × (1.00,1.01) +1.19% (p=0.000) SetTypeNode64Dead 57.2ns × (1.00,1.00) 13.1ns × (0.99,1.02) -77.15% (p=0.000) SetTypeNode64Slice 1.56µs × (1.00,1.01) 1.54µs × (1.00,1.01) -0.89% (p=0.000) SetTypeNode64DeadSlice 1.55µs × (1.00,1.01) 1.52µs × (1.00,1.00) -2.23% (p=0.000) This is the last CL in the sequence converting from the 4-bit heap to the 2-bit heap, with all the same optimizations reenabled. Compared to before that process began (compared to CL 9701 patch set 1): name old mean new mean delta BinaryTree17 5.87s × (0.94,1.09) 5.91s × (0.96,1.06) ~ (p=0.578) Fannkuch11 4.32s × (1.00,1.00) 4.32s × (1.00,1.00) ~ (p=0.474) FmtFprintfEmpty 89.1ns × (0.95,1.16) 89.0ns × (0.93,1.10) ~ (p=0.942) FmtFprintfString 283ns × (0.98,1.02) 298ns × (0.98,1.06) +5.33% (p=0.000) FmtFprintfInt 284ns × (0.98,1.04) 286ns × (0.98,1.03) ~ (p=0.208) FmtFprintfIntInt 486ns × (0.98,1.03) 498ns × (0.97,1.06) +2.48% (p=0.000) FmtFprintfPrefixedInt 400ns × (0.99,1.02) 408ns × (0.98,1.02) +2.23% (p=0.000) FmtFprintfFloat 566ns × (0.99,1.01) 587ns × (0.98,1.01) +3.69% (p=0.000) FmtManyArgs 1.91µs × (0.99,1.02) 1.94µs × (0.99,1.02) +1.81% (p=0.000) GobDecode 15.5ms × (0.98,1.05) 15.8ms × (0.98,1.03) +1.94% (p=0.002) GobEncode 11.9ms × (0.97,1.03) 12.0ms × (0.96,1.09) ~ (p=0.263) Gzip 648ms × (0.99,1.01) 648ms × (0.99,1.01) ~ (p=0.992) Gunzip 143ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.585) HTTPClientServer 89.2µs × (0.99,1.02) 90.3µs × (0.98,1.01) +1.24% (p=0.000) JSONEncode 32.3ms × (0.97,1.06) 31.6ms × (0.99,1.01) -2.29% (p=0.000) JSONDecode 106ms × (0.99,1.01) 107ms × (1.00,1.01) +0.62% (p=0.000) Mandelbrot200 6.02ms × (1.00,1.00) 6.03ms × (1.00,1.01) ~ (p=0.250) GoParse 6.57ms × (0.97,1.06) 6.53ms × (0.99,1.03) ~ (p=0.243) RegexpMatchEasy0_32 162ns × (1.00,1.00) 161ns × (1.00,1.01) -0.80% (p=0.000) RegexpMatchEasy0_1K 561ns × (0.99,1.02) 541ns × (0.99,1.01) -3.67% (p=0.000) RegexpMatchEasy1_32 145ns × (0.95,1.04) 138ns × (1.00,1.00) -5.04% (p=0.000) RegexpMatchEasy1_1K 864ns × (0.99,1.04) 887ns × (0.99,1.01) +2.57% (p=0.000) RegexpMatchMedium_32 255ns × (0.99,1.04) 253ns × (0.99,1.01) -1.05% (p=0.012) RegexpMatchMedium_1K 73.9µs × (0.98,1.04) 72.8µs × (1.00,1.00) -1.51% (p=0.005) RegexpMatchHard_32 3.92µs × (0.98,1.04) 3.85µs × (1.00,1.01) -1.88% (p=0.002) RegexpMatchHard_1K 120µs × (0.98,1.04) 117µs × (1.00,1.01) -2.02% (p=0.001) Revcomp 936ms × (0.95,1.08) 922ms × (0.97,1.08) ~ (p=0.234) Template 130ms × (0.98,1.04) 126ms × (0.99,1.01) -2.99% (p=0.000) TimeParse 638ns × (0.98,1.05) 628ns × (0.99,1.01) -1.54% (p=0.004) TimeFormat 674ns × (0.99,1.01) 668ns × (0.99,1.01) -0.80% (p=0.001) The slowdown of the first few benchmarks seems to be due to the new atomic operations for certain small size allocations. But the larger benchmarks mostly improve, probably due to the decreased memory pressure from having half as much heap bitmap. CL 9706, which removes the (never used anymore) wbshadow mode, gets back what is lost in the early microbenchmarks. Change-Id: I37423a209e8ec2a2e92538b45cac5422a6acd32d Reviewed-on: https://go-review.googlesource.com/9705 Reviewed-by: Rick Hudson <rlh@golang.org>
2015-05-04 22:53:54 -04:00
nptr := typeptrdata(t) / int64(Widthptr)
for i := int64(0); i < nptr; i++ {
if bvget(vec, int32(i)) == 1 {
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
ptrmask[i/8] |= 1 << (uint(i) % 8)
}
}
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
// dgcprog emits and returns the symbol containing a GC program for type t
// along with the size of the data described by the program (in the range [typeptrdata(t), t.Width]).
// In practice, the size is typeptrdata(t) except for non-trivial arrays.
// For non-trivial arrays, the program describes the full t.Width size.
func dgcprog(t *Type) (*Sym, int64) {
dowidth(t)
if t.Width == BADWIDTH {
Fatalf("dgcprog: %v badwidth", t)
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
}
sym := typesymprefix(".gcprog", t)
var p GCProg
p.init(sym)
p.emit(t, 0)
offset := p.w.BitIndex() * int64(Widthptr)
p.end()
if ptrdata := typeptrdata(t); offset < ptrdata || offset > t.Width {
Fatalf("dgcprog: %v: offset=%d but ptrdata=%d size=%d", t, offset, ptrdata, t.Width)
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
}
return sym, offset
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
type GCProg struct {
sym *Sym
symoff int
w gcprog.Writer
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
var Debug_gcprog int // set by -d gcprog
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
func (p *GCProg) init(sym *Sym) {
p.sym = sym
p.symoff = 4 // first 4 bytes hold program length
p.w.Init(p.writeByte)
if Debug_gcprog > 0 {
fmt.Fprintf(os.Stderr, "compile: start GCProg for %v\n", sym)
p.w.Debug(os.Stderr)
}
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
func (p *GCProg) writeByte(x byte) {
p.symoff = duint8(p.sym, p.symoff, x)
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
func (p *GCProg) end() {
p.w.End()
duint32(p.sym, 0, uint32(p.symoff-4))
ggloblsym(p.sym, int32(p.symoff), obj.DUPOK|obj.RODATA|obj.LOCAL)
if Debug_gcprog > 0 {
fmt.Fprintf(os.Stderr, "compile: end GCProg for %v\n", p.sym)
}
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
func (p *GCProg) emit(t *Type, offset int64) {
dowidth(t)
if !haspointers(t) {
return
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
if t.Width == int64(Widthptr) {
p.w.Ptr(offset / int64(Widthptr))
return
}
switch t.Etype {
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
default:
Fatalf("GCProg.emit: unexpected type %v", t)
case TSTRING:
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
p.w.Ptr(offset / int64(Widthptr))
case TINTER:
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
p.w.Ptr(offset / int64(Widthptr))
p.w.Ptr(offset/int64(Widthptr) + 1)
case TARRAY:
if Isslice(t) {
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
p.w.Ptr(offset / int64(Widthptr))
return
}
if t.Bound == 0 {
// should have been handled by haspointers check above
Fatalf("GCProg.emit: empty array")
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
}
// Flatten array-of-array-of-array to just a big array by multiplying counts.
count := t.Bound
elem := t.Type
for Isfixedarray(elem) {
count *= elem.Bound
elem = elem.Type
}
if !p.w.ShouldRepeat(elem.Width/int64(Widthptr), count) {
// Cheaper to just emit the bits.
for i := int64(0); i < count; i++ {
p.emit(elem, offset+i*elem.Width)
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
return
}
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
p.emit(elem, offset)
p.w.ZeroUntil((offset + elem.Width) / int64(Widthptr))
p.w.Repeat(elem.Width/int64(Widthptr), count-1)
case TSTRUCT:
for t1 := t.Type; t1 != nil; t1 = t1.Down {
runtime: replace GC programs with simpler encoding, faster decoder Small types record the location of pointers in their memory layout by using a simple bitmap. In Go 1.4 the bitmap held 4-bit entries, and in Go 1.5 the bitmap holds 1-bit entries, but in both cases using a bitmap for a large type containing arrays does not make sense: if someone refers to the type [1<<28]*byte in a program in such a way that the type information makes it into the binary, it would be a waste of space to write a 128 MB (for 4-bit entries) or even 32 MB (for 1-bit entries) bitmap full of 1s into the binary or even to keep one in memory during the execution of the program. For large types containing arrays, it is much more compact to describe the locations of pointers using a notation that can express repetition than to lay out a bitmap of pointers. Go 1.4 included such a notation, called ``GC programs'' but it was complex, required recursion during decoding, and was generally slow. Dmitriy measured the execution of these programs writing directly to the heap bitmap as being 7x slower than copying from a preunrolled 4-bit mask (and frankly that code was not terribly fast either). For some tests, unrollgcprog1 was seen costing as much as 3x more than the rest of malloc combined. This CL introduces a different form for the GC programs. They use a simple Lempel-Ziv-style encoding of the 1-bit pointer information, in which the only operations are (1) emit the following n bits and (2) repeat the last n bits c more times. This encoding can be generated directly from the Go type information (using repetition only for arrays or large runs of non-pointer data) and it can be decoded very efficiently. In particular the decoding requires little state and no recursion, so that the entire decoding can run without any memory accesses other than the reads of the encoding and the writes of the decoded form to the heap bitmap. For recursive types like arrays of arrays of arrays, the inner instructions are only executed once, not n times, so that large repetitions run at full speed. (In contrast, large repetitions in the old programs repeated the individual bit-level layout of the inner data over and over.) The result is as much as 25x faster decoding compared to the old form. Because the old decoder was so slow, Go 1.4 had three (or so) cases for how to set the heap bitmap bits for an allocation of a given type: (1) If the type had an even number of words up to 32 words, then the 4-bit pointer mask for the type fit in no more than 16 bytes; store the 4-bit pointer mask directly in the binary and copy from it. (1b) If the type had an odd number of words up to 15 words, then the 4-bit pointer mask for the type, doubled to end on a byte boundary, fit in no more than 16 bytes; store that doubled mask directly in the binary and copy from it. (2) If the type had an even number of words up to 128 words, or an odd number of words up to 63 words (again due to doubling), then the 4-bit pointer mask would fit in a 64-byte unrolled mask. Store a GC program in the binary, but leave space in the BSS for the unrolled mask. Execute the GC program to construct the mask the first time it is needed, and thereafter copy from the mask. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. (This is the case that was 7x slower than the other two.) Because the new pointer masks store 1-bit entries instead of 4-bit entries and because using the decoder no longer carries a significant overhead, after this CL (that is, for Go 1.5) there are only two cases: (1) If the type is 128 words or less (no condition about odd or even), store the 1-bit pointer mask directly in the binary and use it to initialize the heap bitmap during malloc. (Implemented in CL 9702.) (2) There is no case 2 anymore. (3) Otherwise, store a GC program and execute it to write directly to the heap bitmap each time an object of that type is allocated. Executing the GC program directly into the heap bitmap (case (3) above) was disabled for the Go 1.5 dev cycle, both to avoid needing to use GC programs for typedmemmove and to avoid updating that code as the heap bitmap format changed. Typedmemmove no longer uses this type information; as of CL 9886 it uses the heap bitmap directly. Now that the heap bitmap format is stable, we reintroduce GC programs and their space savings. Benchmarks for heapBitsSetType, before this CL vs this CL: name old mean new mean delta SetTypePtr 7.59ns × (0.99,1.02) 5.16ns × (1.00,1.00) -32.05% (p=0.000) SetTypePtr8 21.0ns × (0.98,1.05) 21.4ns × (1.00,1.00) ~ (p=0.179) SetTypePtr16 24.1ns × (0.99,1.01) 24.6ns × (1.00,1.00) +2.41% (p=0.001) SetTypePtr32 31.2ns × (0.99,1.01) 32.4ns × (0.99,1.02) +3.72% (p=0.001) SetTypePtr64 45.2ns × (1.00,1.00) 47.2ns × (1.00,1.00) +4.42% (p=0.000) SetTypePtr126 75.8ns × (0.99,1.01) 79.1ns × (1.00,1.00) +4.25% (p=0.000) SetTypePtr128 74.3ns × (0.99,1.01) 77.6ns × (1.00,1.01) +4.55% (p=0.000) SetTypePtrSlice 726ns × (1.00,1.01) 712ns × (1.00,1.00) -1.95% (p=0.001) SetTypeNode1 20.0ns × (0.99,1.01) 20.7ns × (1.00,1.00) +3.71% (p=0.000) SetTypeNode1Slice 112ns × (1.00,1.00) 113ns × (0.99,1.00) ~ (p=0.070) SetTypeNode8 23.9ns × (1.00,1.00) 24.7ns × (1.00,1.01) +3.18% (p=0.000) SetTypeNode8Slice 294ns × (0.99,1.02) 287ns × (0.99,1.01) -2.38% (p=0.015) SetTypeNode64 52.8ns × (0.99,1.03) 51.8ns × (0.99,1.01) ~ (p=0.069) SetTypeNode64Slice 1.13µs × (0.99,1.05) 1.14µs × (0.99,1.00) ~ (p=0.767) SetTypeNode64Dead 36.0ns × (1.00,1.01) 32.5ns × (0.99,1.00) -9.67% (p=0.000) SetTypeNode64DeadSlice 1.43µs × (0.99,1.01) 1.40µs × (1.00,1.00) -2.39% (p=0.001) SetTypeNode124 75.7ns × (1.00,1.01) 79.0ns × (1.00,1.00) +4.44% (p=0.000) SetTypeNode124Slice 1.94µs × (1.00,1.01) 2.04µs × (0.99,1.01) +4.98% (p=0.000) SetTypeNode126 75.4ns × (1.00,1.01) 77.7ns × (0.99,1.01) +3.11% (p=0.000) SetTypeNode126Slice 1.95µs × (0.99,1.01) 2.03µs × (1.00,1.00) +3.74% (p=0.000) SetTypeNode128 85.4ns × (0.99,1.01) 122.0ns × (1.00,1.00) +42.89% (p=0.000) SetTypeNode128Slice 2.20µs × (1.00,1.01) 2.36µs × (0.98,1.02) +7.48% (p=0.001) SetTypeNode130 83.3ns × (1.00,1.00) 123.0ns × (1.00,1.00) +47.61% (p=0.000) SetTypeNode130Slice 2.30µs × (0.99,1.01) 2.40µs × (0.98,1.01) +4.37% (p=0.000) SetTypeNode1024 498ns × (1.00,1.00) 537ns × (1.00,1.00) +7.96% (p=0.000) SetTypeNode1024Slice 15.5µs × (0.99,1.01) 17.8µs × (1.00,1.00) +15.27% (p=0.000) The above compares always using a cached pointer mask (and the corresponding waste of memory) against using the programs directly. Some slowdown is expected, in exchange for having a better general algorithm. The GC programs kick in for SetTypeNode128, SetTypeNode130, SetTypeNode1024, along with the slice variants of those. It is possible that the cutoff of 128 words (bits) should be raised in a followup CL, but even with this low cutoff the GC programs are faster than Go 1.4's "fast path" non-GC program case. Benchmarks for heapBitsSetType, Go 1.4 vs this CL: name old mean new mean delta SetTypePtr 6.89ns × (1.00,1.00) 5.17ns × (1.00,1.00) -25.02% (p=0.000) SetTypePtr8 25.8ns × (0.97,1.05) 21.5ns × (1.00,1.00) -16.70% (p=0.000) SetTypePtr16 39.8ns × (0.97,1.02) 24.7ns × (0.99,1.01) -37.81% (p=0.000) SetTypePtr32 68.8ns × (0.98,1.01) 32.2ns × (1.00,1.01) -53.18% (p=0.000) SetTypePtr64 130ns × (1.00,1.00) 47ns × (1.00,1.00) -63.67% (p=0.000) SetTypePtr126 241ns × (0.99,1.01) 79ns × (1.00,1.01) -67.25% (p=0.000) SetTypePtr128 2.07µs × (1.00,1.00) 0.08µs × (1.00,1.00) -96.27% (p=0.000) SetTypePtrSlice 1.05µs × (0.99,1.01) 0.72µs × (0.99,1.02) -31.70% (p=0.000) SetTypeNode1 16.0ns × (0.99,1.01) 20.8ns × (0.99,1.03) +29.91% (p=0.000) SetTypeNode1Slice 184ns × (0.99,1.01) 112ns × (0.99,1.01) -39.26% (p=0.000) SetTypeNode8 29.5ns × (0.97,1.02) 24.6ns × (1.00,1.00) -16.50% (p=0.000) SetTypeNode8Slice 624ns × (0.98,1.02) 285ns × (1.00,1.00) -54.31% (p=0.000) SetTypeNode64 135ns × (0.96,1.08) 52ns × (0.99,1.02) -61.32% (p=0.000) SetTypeNode64Slice 3.83µs × (1.00,1.00) 1.14µs × (0.99,1.01) -70.16% (p=0.000) SetTypeNode64Dead 134ns × (0.99,1.01) 32ns × (1.00,1.01) -75.74% (p=0.000) SetTypeNode64DeadSlice 3.83µs × (0.99,1.00) 1.40µs × (1.00,1.01) -63.42% (p=0.000) SetTypeNode124 240ns × (0.99,1.01) 79ns × (1.00,1.01) -67.05% (p=0.000) SetTypeNode124Slice 7.27µs × (1.00,1.00) 2.04µs × (1.00,1.00) -71.95% (p=0.000) SetTypeNode126 2.06µs × (0.99,1.01) 0.08µs × (0.99,1.01) -96.23% (p=0.000) SetTypeNode126Slice 64.4µs × (1.00,1.00) 2.0µs × (1.00,1.00) -96.85% (p=0.000) SetTypeNode128 2.09µs × (1.00,1.01) 0.12µs × (1.00,1.00) -94.15% (p=0.000) SetTypeNode128Slice 65.4µs × (1.00,1.00) 2.4µs × (0.99,1.03) -96.39% (p=0.000) SetTypeNode130 2.11µs × (1.00,1.00) 0.12µs × (1.00,1.00) -94.18% (p=0.000) SetTypeNode130Slice 66.3µs × (1.00,1.00) 2.4µs × (0.97,1.08) -96.34% (p=0.000) SetTypeNode1024 16.0µs × (1.00,1.01) 0.5µs × (1.00,1.00) -96.65% (p=0.000) SetTypeNode1024Slice 512µs × (1.00,1.00) 18µs × (0.98,1.04) -96.45% (p=0.000) SetTypeNode124 uses a 124 data + 2 ptr = 126-word allocation. Both Go 1.4 and this CL are using pointer bitmaps for this case, so that's an overall 3x speedup for using pointer bitmaps. SetTypeNode128 uses a 128 data + 2 ptr = 130-word allocation. Both Go 1.4 and this CL are running the GC program for this case, so that's an overall 17x speedup when using GC programs (and I've seen >20x on other systems). Comparing Go 1.4's SetTypeNode124 (pointer bitmap) against this CL's SetTypeNode128 (GC program), the slow path in the code in this CL is 2x faster than the fast path in Go 1.4. The Go 1 benchmarks are basically unaffected compared to just before this CL. Go 1 benchmarks, before this CL vs this CL: name old mean new mean delta BinaryTree17 5.87s × (0.97,1.04) 5.91s × (0.96,1.04) ~ (p=0.306) Fannkuch11 4.38s × (1.00,1.00) 4.37s × (1.00,1.01) -0.22% (p=0.006) FmtFprintfEmpty 90.7ns × (0.97,1.10) 89.3ns × (0.96,1.09) ~ (p=0.280) FmtFprintfString 282ns × (0.98,1.04) 287ns × (0.98,1.07) +1.72% (p=0.039) FmtFprintfInt 269ns × (0.99,1.03) 282ns × (0.97,1.04) +4.87% (p=0.000) FmtFprintfIntInt 478ns × (0.99,1.02) 481ns × (0.99,1.02) +0.61% (p=0.048) FmtFprintfPrefixedInt 399ns × (0.98,1.03) 400ns × (0.98,1.05) ~ (p=0.533) FmtFprintfFloat 563ns × (0.99,1.01) 570ns × (1.00,1.01) +1.37% (p=0.000) FmtManyArgs 1.89µs × (0.99,1.01) 1.92µs × (0.99,1.02) +1.88% (p=0.000) GobDecode 15.2ms × (0.99,1.01) 15.2ms × (0.98,1.05) ~ (p=0.609) GobEncode 11.6ms × (0.98,1.03) 11.9ms × (0.98,1.04) +2.17% (p=0.000) Gzip 648ms × (0.99,1.01) 648ms × (1.00,1.01) ~ (p=0.835) Gunzip 142ms × (1.00,1.00) 143ms × (1.00,1.01) ~ (p=0.169) HTTPClientServer 90.5µs × (0.98,1.03) 91.5µs × (0.98,1.04) +1.04% (p=0.045) JSONEncode 31.5ms × (0.98,1.03) 31.4ms × (0.98,1.03) ~ (p=0.549) JSONDecode 111ms × (0.99,1.01) 107ms × (0.99,1.01) -3.21% (p=0.000) Mandelbrot200 6.01ms × (1.00,1.00) 6.01ms × (1.00,1.00) ~ (p=0.878) GoParse 6.54ms × (0.99,1.02) 6.61ms × (0.99,1.03) +1.08% (p=0.004) RegexpMatchEasy0_32 160ns × (1.00,1.01) 161ns × (1.00,1.00) +0.40% (p=0.000) RegexpMatchEasy0_1K 560ns × (0.99,1.01) 559ns × (0.99,1.01) ~ (p=0.088) RegexpMatchEasy1_32 138ns × (0.99,1.01) 138ns × (1.00,1.00) ~ (p=0.380) RegexpMatchEasy1_1K 877ns × (1.00,1.00) 878ns × (1.00,1.00) ~ (p=0.157) RegexpMatchMedium_32 251ns × (0.99,1.00) 251ns × (1.00,1.01) +0.28% (p=0.021) RegexpMatchMedium_1K 72.6µs × (1.00,1.00) 72.6µs × (1.00,1.00) ~ (p=0.539) RegexpMatchHard_32 3.84µs × (1.00,1.00) 3.84µs × (1.00,1.00) ~ (p=0.378) RegexpMatchHard_1K 117µs × (1.00,1.00) 117µs × (1.00,1.00) ~ (p=0.067) Revcomp 904ms × (0.99,1.02) 904ms × (0.99,1.01) ~ (p=0.943) Template 125ms × (0.99,1.02) 127ms × (0.99,1.01) +1.79% (p=0.000) TimeParse 627ns × (0.99,1.01) 622ns × (0.99,1.01) -0.88% (p=0.000) TimeFormat 655ns × (0.99,1.02) 655ns × (0.99,1.02) ~ (p=0.976) For the record, Go 1 benchmarks, Go 1.4 vs this CL: name old mean new mean delta BinaryTree17 4.61s × (0.97,1.05) 5.91s × (0.98,1.03) +28.35% (p=0.000) Fannkuch11 4.40s × (0.99,1.03) 4.41s × (0.99,1.01) ~ (p=0.212) FmtFprintfEmpty 102ns × (0.99,1.01) 84ns × (0.99,1.02) -18.38% (p=0.000) FmtFprintfString 302ns × (0.98,1.01) 303ns × (0.99,1.02) ~ (p=0.203) FmtFprintfInt 313ns × (0.97,1.05) 270ns × (0.99,1.01) -13.69% (p=0.000) FmtFprintfIntInt 524ns × (0.98,1.02) 477ns × (0.99,1.00) -8.87% (p=0.000) FmtFprintfPrefixedInt 424ns × (0.98,1.02) 386ns × (0.99,1.01) -8.96% (p=0.000) FmtFprintfFloat 652ns × (0.98,1.02) 594ns × (0.97,1.05) -8.97% (p=0.000) FmtManyArgs 2.13µs × (0.99,1.02) 1.94µs × (0.99,1.01) -8.92% (p=0.000) GobDecode 17.1ms × (0.99,1.02) 14.9ms × (0.98,1.03) -13.07% (p=0.000) GobEncode 13.5ms × (0.98,1.03) 11.5ms × (0.98,1.03) -15.25% (p=0.000) Gzip 656ms × (0.99,1.02) 647ms × (0.99,1.01) -1.29% (p=0.000) Gunzip 143ms × (0.99,1.02) 144ms × (0.99,1.01) ~ (p=0.204) HTTPClientServer 88.2µs × (0.98,1.02) 90.8µs × (0.98,1.01) +2.93% (p=0.000) JSONEncode 32.2ms × (0.98,1.02) 30.9ms × (0.97,1.04) -4.06% (p=0.001) JSONDecode 121ms × (0.98,1.02) 110ms × (0.98,1.05) -8.95% (p=0.000) Mandelbrot200 6.06ms × (0.99,1.01) 6.11ms × (0.98,1.04) ~ (p=0.184) GoParse 6.76ms × (0.97,1.04) 6.58ms × (0.98,1.05) -2.63% (p=0.003) RegexpMatchEasy0_32 195ns × (1.00,1.01) 155ns × (0.99,1.01) -20.43% (p=0.000) RegexpMatchEasy0_1K 479ns × (0.98,1.03) 535ns × (0.99,1.02) +11.59% (p=0.000) RegexpMatchEasy1_32 169ns × (0.99,1.02) 131ns × (0.99,1.03) -22.44% (p=0.000) RegexpMatchEasy1_1K 1.53µs × (0.99,1.01) 0.87µs × (0.99,1.02) -43.07% (p=0.000) RegexpMatchMedium_32 334ns × (0.99,1.01) 242ns × (0.99,1.01) -27.53% (p=0.000) RegexpMatchMedium_1K 125µs × (1.00,1.01) 72µs × (0.99,1.03) -42.53% (p=0.000) RegexpMatchHard_32 6.03µs × (0.99,1.01) 3.79µs × (0.99,1.01) -37.12% (p=0.000) RegexpMatchHard_1K 189µs × (0.99,1.02) 115µs × (0.99,1.01) -39.20% (p=0.000) Revcomp 935ms × (0.96,1.03) 926ms × (0.98,1.02) ~ (p=0.083) Template 146ms × (0.97,1.05) 119ms × (0.99,1.01) -18.37% (p=0.000) TimeParse 660ns × (0.99,1.01) 624ns × (0.99,1.02) -5.43% (p=0.000) TimeFormat 670ns × (0.98,1.02) 710ns × (1.00,1.01) +5.97% (p=0.000) This CL is a bit larger than I would like, but the compiler, linker, runtime, and package reflect all need to be in sync about the format of these programs, so there is no easy way to split this into independent changes (at least while keeping the build working at each change). Fixes #9625. Fixes #10524. Change-Id: I9e3e20d6097099d0f8532d1cb5b1af528804989a Reviewed-on: https://go-review.googlesource.com/9888 Reviewed-by: Austin Clements <austin@google.com> Run-TryBot: Russ Cox <rsc@golang.org>
2015-05-08 01:43:18 -04:00
p.emit(t1.Type, offset+t1.Width)
}
}
}