mirror of
https://github.com/golang/go.git
synced 2025-12-08 06:10:04 +00:00
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
This commit is contained in:
Josh Bleecher Snyder
parent
974c201f74
commit
6b33b0e98e
2 changed files with 23 additions and 13 deletions
|
|
@ -37,6 +37,12 @@ func makeslice(t *slicetype, len64, cap64 int64) slice {
|
|||
// It is passed the slice type, the old slice, and the desired new minimum capacity,
|
||||
// and it returns a new slice with at least that capacity, with the old data
|
||||
// copied into it.
|
||||
// The new slice's length is set to the old slice's length,
|
||||
// NOT to the new requested capacity.
|
||||
// This is for codegen convenience. The old slice's length is used immediately
|
||||
// to calculate where to write new values during an append.
|
||||
// TODO: When the old backend is gone, reconsider this decision.
|
||||
// The SSA backend might prefer the new length or to return only ptr/cap and save stack space.
|
||||
func growslice(t *slicetype, old slice, cap int) slice {
|
||||
if raceenabled {
|
||||
callerpc := getcallerpc(unsafe.Pointer(&t))
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue