go/src/cmd/compile/internal/ssa/lca.go

124 lines
3.7 KiB
Go
Raw Normal View History

cmd/compile: improve tighten pass Move a value to the block which is the lowest common ancestor in the dominator tree of all of its uses. Make sure not to move a value into a loop. Makes the tighten pass on average (across go1 benchmarks) 40% slower. Still not a big contributor to overall compile time. Binary size is just a tad smaller. name old time/op new time/op delta BinaryTree17-12 2.77s ± 9% 2.76s ± 9% ~ (p=0.878 n=8+8) Fannkuch11-12 2.75s ± 1% 2.74s ± 1% ~ (p=0.232 n=8+7) FmtFprintfEmpty-12 48.9ns ± 9% 47.7ns ± 0% ~ (p=0.431 n=8+8) FmtFprintfString-12 143ns ± 8% 142ns ± 1% ~ (p=0.257 n=8+7) FmtFprintfInt-12 123ns ± 1% 122ns ± 1% -1.04% (p=0.026 n=7+8) FmtFprintfIntInt-12 195ns ± 7% 185ns ± 0% -5.32% (p=0.000 n=8+8) FmtFprintfPrefixedInt-12 194ns ± 4% 195ns ± 0% +0.81% (p=0.015 n=7+7) FmtFprintfFloat-12 267ns ± 0% 268ns ± 0% +0.37% (p=0.001 n=7+6) FmtManyArgs-12 800ns ± 0% 762ns ± 1% -4.78% (p=0.000 n=8+8) GobDecode-12 7.67ms ± 2% 7.60ms ± 2% ~ (p=0.234 n=8+8) GobEncode-12 6.55ms ± 0% 6.57ms ± 1% ~ (p=0.336 n=7+8) Gzip-12 237ms ± 0% 238ms ± 0% +0.40% (p=0.017 n=7+7) Gunzip-12 40.8ms ± 0% 40.2ms ± 0% -1.52% (p=0.000 n=7+8) HTTPClientServer-12 208µs ± 3% 209µs ± 3% ~ (p=0.955 n=8+7) JSONEncode-12 16.2ms ± 1% 17.2ms ±11% +5.80% (p=0.001 n=7+8) JSONDecode-12 57.3ms ±12% 55.5ms ± 3% ~ (p=0.867 n=8+7) Mandelbrot200-12 4.68ms ± 6% 4.46ms ± 1% ~ (p=0.442 n=8+8) GoParse-12 4.27ms ±44% 3.42ms ± 1% -19.95% (p=0.005 n=8+8) RegexpMatchEasy0_32-12 75.1ns ± 0% 75.8ns ± 1% +0.99% (p=0.002 n=7+7) RegexpMatchEasy0_1K-12 963ns ± 0% 1021ns ± 6% +5.98% (p=0.001 n=7+7) RegexpMatchEasy1_32-12 72.4ns ±11% 70.8ns ± 1% ~ (p=0.368 n=8+8) RegexpMatchEasy1_1K-12 394ns ± 1% 399ns ± 0% +1.23% (p=0.000 n=8+7) RegexpMatchMedium_32-12 114ns ± 0% 115ns ± 1% +0.63% (p=0.021 n=7+7) RegexpMatchMedium_1K-12 35.9µs ± 0% 37.6µs ± 1% +4.72% (p=0.000 n=7+8) RegexpMatchHard_32-12 1.93µs ± 2% 1.91µs ± 0% -0.91% (p=0.001 n=7+7) RegexpMatchHard_1K-12 60.2µs ± 3% 61.2µs ±10% ~ (p=0.442 n=8+8) Revcomp-12 404ms ± 1% 406ms ± 1% ~ (p=0.054 n=8+7) Template-12 64.6ms ± 1% 63.5ms ± 1% -1.66% (p=0.000 n=8+8) TimeParse-12 347ns ± 8% 309ns ± 0% -11.13% (p=0.000 n=8+7) TimeFormat-12 343ns ± 4% 331ns ± 0% -3.34% (p=0.000 n=8+7) Change-Id: Id6da1239ddd4d0cb074ff29cffb06302d1c6d08f Reviewed-on: https://go-review.googlesource.com/28712 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-09-07 14:04:31 -07:00
// Copyright 2016 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 ssa
// Code to compute lowest common ancestors in the dominator tree.
// https://en.wikipedia.org/wiki/Lowest_common_ancestor
// https://en.wikipedia.org/wiki/Range_minimum_query#Solution_using_constant_time_and_linearithmic_space
// lcaRange is a data structure that can compute lowest common ancestor queries
// in O(n lg n) precomputed space and O(1) time per query.
type lcaRange struct {
// Additional information about each block (indexed by block ID).
blocks []lcaRangeBlock
// Data structure for range minimum queries.
// rangeMin[k][i] contains the ID of the minimum depth block
// in the Euler tour from positions i to i+1<<k-1, inclusive.
rangeMin [][]ID
}
type lcaRangeBlock struct {
b *Block
parent ID // parent in dominator tree. 0 = no parent (entry or unreachable)
firstChild ID // first child in dominator tree
sibling ID // next child of parent
pos int32 // an index in the Euler tour where this block appears (any one of its occurrences)
depth int32 // depth in dominator tree (root=0, its children=1, etc.)
}
func makeLCArange(f *Func) *lcaRange {
dom := f.Idom()
cmd/compile: improve tighten pass Move a value to the block which is the lowest common ancestor in the dominator tree of all of its uses. Make sure not to move a value into a loop. Makes the tighten pass on average (across go1 benchmarks) 40% slower. Still not a big contributor to overall compile time. Binary size is just a tad smaller. name old time/op new time/op delta BinaryTree17-12 2.77s ± 9% 2.76s ± 9% ~ (p=0.878 n=8+8) Fannkuch11-12 2.75s ± 1% 2.74s ± 1% ~ (p=0.232 n=8+7) FmtFprintfEmpty-12 48.9ns ± 9% 47.7ns ± 0% ~ (p=0.431 n=8+8) FmtFprintfString-12 143ns ± 8% 142ns ± 1% ~ (p=0.257 n=8+7) FmtFprintfInt-12 123ns ± 1% 122ns ± 1% -1.04% (p=0.026 n=7+8) FmtFprintfIntInt-12 195ns ± 7% 185ns ± 0% -5.32% (p=0.000 n=8+8) FmtFprintfPrefixedInt-12 194ns ± 4% 195ns ± 0% +0.81% (p=0.015 n=7+7) FmtFprintfFloat-12 267ns ± 0% 268ns ± 0% +0.37% (p=0.001 n=7+6) FmtManyArgs-12 800ns ± 0% 762ns ± 1% -4.78% (p=0.000 n=8+8) GobDecode-12 7.67ms ± 2% 7.60ms ± 2% ~ (p=0.234 n=8+8) GobEncode-12 6.55ms ± 0% 6.57ms ± 1% ~ (p=0.336 n=7+8) Gzip-12 237ms ± 0% 238ms ± 0% +0.40% (p=0.017 n=7+7) Gunzip-12 40.8ms ± 0% 40.2ms ± 0% -1.52% (p=0.000 n=7+8) HTTPClientServer-12 208µs ± 3% 209µs ± 3% ~ (p=0.955 n=8+7) JSONEncode-12 16.2ms ± 1% 17.2ms ±11% +5.80% (p=0.001 n=7+8) JSONDecode-12 57.3ms ±12% 55.5ms ± 3% ~ (p=0.867 n=8+7) Mandelbrot200-12 4.68ms ± 6% 4.46ms ± 1% ~ (p=0.442 n=8+8) GoParse-12 4.27ms ±44% 3.42ms ± 1% -19.95% (p=0.005 n=8+8) RegexpMatchEasy0_32-12 75.1ns ± 0% 75.8ns ± 1% +0.99% (p=0.002 n=7+7) RegexpMatchEasy0_1K-12 963ns ± 0% 1021ns ± 6% +5.98% (p=0.001 n=7+7) RegexpMatchEasy1_32-12 72.4ns ±11% 70.8ns ± 1% ~ (p=0.368 n=8+8) RegexpMatchEasy1_1K-12 394ns ± 1% 399ns ± 0% +1.23% (p=0.000 n=8+7) RegexpMatchMedium_32-12 114ns ± 0% 115ns ± 1% +0.63% (p=0.021 n=7+7) RegexpMatchMedium_1K-12 35.9µs ± 0% 37.6µs ± 1% +4.72% (p=0.000 n=7+8) RegexpMatchHard_32-12 1.93µs ± 2% 1.91µs ± 0% -0.91% (p=0.001 n=7+7) RegexpMatchHard_1K-12 60.2µs ± 3% 61.2µs ±10% ~ (p=0.442 n=8+8) Revcomp-12 404ms ± 1% 406ms ± 1% ~ (p=0.054 n=8+7) Template-12 64.6ms ± 1% 63.5ms ± 1% -1.66% (p=0.000 n=8+8) TimeParse-12 347ns ± 8% 309ns ± 0% -11.13% (p=0.000 n=8+7) TimeFormat-12 343ns ± 4% 331ns ± 0% -3.34% (p=0.000 n=8+7) Change-Id: Id6da1239ddd4d0cb074ff29cffb06302d1c6d08f Reviewed-on: https://go-review.googlesource.com/28712 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-09-07 14:04:31 -07:00
// Build tree
blocks := make([]lcaRangeBlock, f.NumBlocks())
for _, b := range f.Blocks {
blocks[b.ID].b = b
if dom[b.ID] == nil {
continue // entry or unreachable
}
parent := dom[b.ID].ID
blocks[b.ID].parent = parent
blocks[b.ID].sibling = blocks[parent].firstChild
blocks[parent].firstChild = b.ID
}
// Compute euler tour ordering.
// Each reachable block will appear #children+1 times in the tour.
tour := make([]ID, 0, f.NumBlocks()*2-1)
type queueEntry struct {
bid ID // block to work on
cid ID // child we're already working on (0 = haven't started yet)
}
q := []queueEntry{{f.Entry.ID, 0}}
for len(q) > 0 {
n := len(q) - 1
bid := q[n].bid
cid := q[n].cid
q = q[:n]
// Add block to tour.
blocks[bid].pos = int32(len(tour))
tour = append(tour, bid)
// Proceed down next child edge (if any).
if cid == 0 {
// This is our first visit to b. Set its depth.
blocks[bid].depth = blocks[blocks[bid].parent].depth + 1
// Then explore its first child.
cid = blocks[bid].firstChild
} else {
// We've seen b before. Explore the next child.
cid = blocks[cid].sibling
}
if cid != 0 {
q = append(q, queueEntry{bid, cid}, queueEntry{cid, 0})
}
}
// Compute fast range-minimum query data structure
var rangeMin [][]ID
rangeMin = append(rangeMin, tour) // 1-size windows are just the tour itself.
for logS, s := 1, 2; s < len(tour); logS, s = logS+1, s*2 {
r := make([]ID, len(tour)-s+1)
for i := 0; i < len(tour)-s+1; i++ {
bid := rangeMin[logS-1][i]
bid2 := rangeMin[logS-1][i+s/2]
if blocks[bid2].depth < blocks[bid].depth {
bid = bid2
}
r[i] = bid
}
rangeMin = append(rangeMin, r)
}
return &lcaRange{blocks: blocks, rangeMin: rangeMin}
}
// find returns the lowest common ancestor of a and b.
func (lca *lcaRange) find(a, b *Block) *Block {
if a == b {
return a
}
// Find the positions of a and bin the Euler tour.
p1 := lca.blocks[a.ID].pos
p2 := lca.blocks[b.ID].pos
if p1 > p2 {
p1, p2 = p2, p1
}
// The lowest common ancestor is the minimum depth block
// on the tour from p1 to p2. We've precomputed minimum
// depth blocks for powers-of-two subsequences of the tour.
// Combine the right two precomputed values to get the answer.
logS := uint(log64(int64(p2 - p1)))
cmd/compile: improve tighten pass Move a value to the block which is the lowest common ancestor in the dominator tree of all of its uses. Make sure not to move a value into a loop. Makes the tighten pass on average (across go1 benchmarks) 40% slower. Still not a big contributor to overall compile time. Binary size is just a tad smaller. name old time/op new time/op delta BinaryTree17-12 2.77s ± 9% 2.76s ± 9% ~ (p=0.878 n=8+8) Fannkuch11-12 2.75s ± 1% 2.74s ± 1% ~ (p=0.232 n=8+7) FmtFprintfEmpty-12 48.9ns ± 9% 47.7ns ± 0% ~ (p=0.431 n=8+8) FmtFprintfString-12 143ns ± 8% 142ns ± 1% ~ (p=0.257 n=8+7) FmtFprintfInt-12 123ns ± 1% 122ns ± 1% -1.04% (p=0.026 n=7+8) FmtFprintfIntInt-12 195ns ± 7% 185ns ± 0% -5.32% (p=0.000 n=8+8) FmtFprintfPrefixedInt-12 194ns ± 4% 195ns ± 0% +0.81% (p=0.015 n=7+7) FmtFprintfFloat-12 267ns ± 0% 268ns ± 0% +0.37% (p=0.001 n=7+6) FmtManyArgs-12 800ns ± 0% 762ns ± 1% -4.78% (p=0.000 n=8+8) GobDecode-12 7.67ms ± 2% 7.60ms ± 2% ~ (p=0.234 n=8+8) GobEncode-12 6.55ms ± 0% 6.57ms ± 1% ~ (p=0.336 n=7+8) Gzip-12 237ms ± 0% 238ms ± 0% +0.40% (p=0.017 n=7+7) Gunzip-12 40.8ms ± 0% 40.2ms ± 0% -1.52% (p=0.000 n=7+8) HTTPClientServer-12 208µs ± 3% 209µs ± 3% ~ (p=0.955 n=8+7) JSONEncode-12 16.2ms ± 1% 17.2ms ±11% +5.80% (p=0.001 n=7+8) JSONDecode-12 57.3ms ±12% 55.5ms ± 3% ~ (p=0.867 n=8+7) Mandelbrot200-12 4.68ms ± 6% 4.46ms ± 1% ~ (p=0.442 n=8+8) GoParse-12 4.27ms ±44% 3.42ms ± 1% -19.95% (p=0.005 n=8+8) RegexpMatchEasy0_32-12 75.1ns ± 0% 75.8ns ± 1% +0.99% (p=0.002 n=7+7) RegexpMatchEasy0_1K-12 963ns ± 0% 1021ns ± 6% +5.98% (p=0.001 n=7+7) RegexpMatchEasy1_32-12 72.4ns ±11% 70.8ns ± 1% ~ (p=0.368 n=8+8) RegexpMatchEasy1_1K-12 394ns ± 1% 399ns ± 0% +1.23% (p=0.000 n=8+7) RegexpMatchMedium_32-12 114ns ± 0% 115ns ± 1% +0.63% (p=0.021 n=7+7) RegexpMatchMedium_1K-12 35.9µs ± 0% 37.6µs ± 1% +4.72% (p=0.000 n=7+8) RegexpMatchHard_32-12 1.93µs ± 2% 1.91µs ± 0% -0.91% (p=0.001 n=7+7) RegexpMatchHard_1K-12 60.2µs ± 3% 61.2µs ±10% ~ (p=0.442 n=8+8) Revcomp-12 404ms ± 1% 406ms ± 1% ~ (p=0.054 n=8+7) Template-12 64.6ms ± 1% 63.5ms ± 1% -1.66% (p=0.000 n=8+8) TimeParse-12 347ns ± 8% 309ns ± 0% -11.13% (p=0.000 n=8+7) TimeFormat-12 343ns ± 4% 331ns ± 0% -3.34% (p=0.000 n=8+7) Change-Id: Id6da1239ddd4d0cb074ff29cffb06302d1c6d08f Reviewed-on: https://go-review.googlesource.com/28712 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-09-07 14:04:31 -07:00
bid1 := lca.rangeMin[logS][p1]
bid2 := lca.rangeMin[logS][p2-1<<logS+1]
if lca.blocks[bid1].depth < lca.blocks[bid2].depth {
return lca.blocks[bid1].b
}
return lca.blocks[bid2].b
}