cmd/compile: optimize regalloc for phi value

When allocating registers for phi value, only the primary predecessor is considered.
Taking into account the allocation status of other predecessors can help reduce
unnecessary copy or spill operations. Many such cases can be found in the standard
library, such as runtime.wirep, moveByType, etc. The test results from benchstat
also show that this change helps reduce the file size.

name                      old time/op       new time/op       delta
Template                        328ms ± 5%        326ms ± 4%    ~     (p=0.254 n=50+47)
Unicode                         156ms ± 7%        158ms ±10%    ~     (p=0.412 n=49+49)
GoTypes                         1.07s ± 3%        1.07s ± 2%    ~     (p=0.664 n=48+49)
Compiler                        4.43s ± 3%        4.44s ± 3%    ~     (p=0.758 n=48+50)
SSA                             10.3s ± 2%        10.4s ± 2%  +0.43%  (p=0.017 n=50+46)
Flate                           208ms ± 9%        209ms ± 7%    ~     (p=0.920 n=49+46)
GoParser                        260ms ± 5%        262ms ± 4%    ~     (p=0.063 n=50+48)
Reflect                         687ms ± 3%        685ms ± 2%    ~     (p=0.459 n=50+48)
Tar                             293ms ± 4%        293ms ± 5%    ~     (p=0.695 n=49+48)
XML                             391ms ± 4%        389ms ± 3%    ~     (p=0.109 n=49+46)
LinkCompiler                    570ms ± 5%        563ms ± 5%  -1.10%  (p=0.006 n=46+47)
ExternalLinkCompiler            1.57s ± 3%        1.56s ± 3%    ~     (p=0.118 n=47+46)
LinkWithoutDebugCompiler        349ms ± 6%        349ms ± 5%    ~     (p=0.726 n=49+47)
[Geo mean]                      645ms             645ms       -0.05%

name                      old user-time/op  new user-time/op  delta
Template                        507ms ±14%        513ms ±14%    ~     (p=0.398 n=48+49)
Unicode                         345ms ±29%        345ms ±38%    ~     (p=0.521 n=47+49)
GoTypes                         1.95s ±16%        1.94s ±19%    ~     (p=0.324 n=50+50)
Compiler                        8.26s ±16%        8.22s ±14%    ~     (p=0.834 n=50+50)
SSA                             19.6s ± 8%        19.2s ±15%    ~     (p=0.056 n=50+50)
Flate                           293ms ± 9%        299ms ±12%    ~     (p=0.057 n=47+50)
GoParser                        388ms ± 9%        387ms ±14%    ~     (p=0.660 n=46+50)
Reflect                         1.15s ±28%        1.12s ±18%    ~     (p=0.648 n=49+48)
Tar                             456ms ±10%        476ms ±15%  +4.48%  (p=0.001 n=46+48)
XML                             648ms ±27%        634ms ±16%    ~     (p=0.685 n=50+46)
LinkCompiler                    1.00s ± 8%        1.00s ± 8%    ~     (p=0.638 n=50+50)
ExternalLinkCompiler            1.96s ± 5%        1.96s ± 5%    ~     (p=0.792 n=50+50)
LinkWithoutDebugCompiler        443ms ±10%        442ms ±11%    ~     (p=0.813 n=50+50)
[Geo mean]                      1.05s             1.05s       -0.09%

name                      old alloc/op      new alloc/op      delta
Template                       36.0MB ± 0%       36.0MB ± 0%    ~     (p=0.599 n=49+50)
Unicode                        29.8MB ± 0%       29.8MB ± 0%    ~     (p=0.739 n=50+50)
GoTypes                         118MB ± 0%        118MB ± 0%    ~     (p=0.436 n=50+50)
Compiler                        562MB ± 0%        562MB ± 0%    ~     (p=0.693 n=50+50)
SSA                            1.42GB ± 0%       1.42GB ± 0%  -0.10%  (p=0.000 n=50+49)
Flate                          22.5MB ± 0%       22.5MB ± 0%    ~     (p=0.429 n=48+49)
GoParser                       27.7MB ± 0%       27.7MB ± 0%    ~     (p=0.705 n=49+48)
Reflect                        77.7MB ± 0%       77.7MB ± 0%  -0.01%  (p=0.043 n=50+50)
Tar                            33.8MB ± 0%       33.8MB ± 0%    ~     (p=0.241 n=49+50)
XML                            42.8MB ± 0%       42.8MB ± 0%    ~     (p=0.677 n=47+49)
LinkCompiler                   98.3MB ± 0%       98.3MB ± 0%    ~     (p=0.157 n=50+50)
ExternalLinkCompiler           89.4MB ± 0%       89.4MB ± 0%    ~     (p=0.683 n=50+50)
LinkWithoutDebugCompiler       56.7MB ± 0%       56.7MB ± 0%    ~     (p=0.155 n=49+49)
[Geo mean]                     77.3MB            77.3MB       -0.01%

name                      old allocs/op     new allocs/op     delta
Template                         367k ± 0%         367k ± 0%    ~     (p=0.863 n=50+50)
Unicode                          345k ± 0%         345k ± 0%    ~     (p=0.744 n=49+49)
GoTypes                         1.28M ± 0%        1.28M ± 0%    ~     (p=0.957 n=48+50)
Compiler                        5.39M ± 0%        5.39M ± 0%  +0.00%  (p=0.012 n=50+49)
SSA                             13.9M ± 0%        13.9M ± 0%  +0.02%  (p=0.000 n=47+49)
Flate                            230k ± 0%         230k ± 0%  -0.01%  (p=0.007 n=47+49)
GoParser                         292k ± 0%         292k ± 0%    ~     (p=0.891 n=50+49)
Reflect                          977k ± 0%         977k ± 0%    ~     (p=0.274 n=50+50)
Tar                              343k ± 0%         343k ± 0%    ~     (p=0.942 n=50+50)
XML                              418k ± 0%         418k ± 0%    ~     (p=0.374 n=50+49)
LinkCompiler                     516k ± 0%         516k ± 0%    ~     (p=0.205 n=49+47)
ExternalLinkCompiler             570k ± 0%         570k ± 0%    ~     (p=0.783 n=49+47)
LinkWithoutDebugCompiler         169k ± 0%         169k ± 0%    ~     (p=0.233 n=50+46)
[Geo mean]                       672k              672k       +0.00%

name                      old maxRSS/op     new maxRSS/op     delta
Template                        34.5M ± 3%        34.4M ± 3%    ~     (p=0.566 n=49+48)
Unicode                         36.0M ± 6%        35.9M ± 6%    ~     (p=0.736 n=50+50)
GoTypes                         75.7M ± 7%        75.4M ± 5%    ~     (p=0.412 n=50+50)
Compiler                         314M ±10%         313M ± 8%    ~     (p=0.708 n=50+50)
SSA                              730M ± 6%         735M ± 6%    ~     (p=0.324 n=50+50)
Flate                           25.8M ± 5%        25.6M ± 6%    ~     (p=0.415 n=49+50)
GoParser                        28.5M ± 3%        28.5M ± 4%    ~     (p=0.977 n=46+50)
Reflect                         57.4M ± 4%        57.2M ± 3%    ~     (p=0.173 n=50+50)
Tar                             33.3M ± 3%        33.2M ± 4%    ~     (p=0.621 n=48+50)
XML                             39.6M ± 5%        39.6M ± 4%    ~     (p=0.997 n=50+50)
LinkCompiler                     168M ± 2%         167M ± 1%    ~     (p=0.072 n=49+45)
ExternalLinkCompiler             179M ± 1%         179M ± 1%    ~     (p=0.147 n=48+50)
LinkWithoutDebugCompiler         136M ± 1%         136M ± 1%    ~     (p=0.789 n=47+49)
[Geo mean]                      79.2M             79.1M       -0.12%

name                      old text-bytes    new text-bytes    delta
HelloSize                       812kB ± 0%        811kB ± 0%  -0.06%  (p=0.000 n=50+50)

name                      old data-bytes    new data-bytes    delta
HelloSize                      13.3kB ± 0%       13.3kB ± 0%    ~     (all equal)

name                      old bss-bytes     new bss-bytes     delta
HelloSize                       206kB ± 0%        206kB ± 0%    ~     (all equal)

name                      old exe-bytes     new exe-bytes     delta
HelloSize                      1.21MB ± 0%       1.21MB ± 0%  -0.03%  (p=0.000 n=50+50)

file      before    after     Δ       %
addr2line 4057421   4056237   -1184   -0.029%
api       4952451   4946715   -5736   -0.116%
asm       4888993   4888185   -808    -0.017%
buildid   2617705   2616441   -1264   -0.048%
cgo       4521849   4520681   -1168   -0.026%
compile   19143451  19141243  -2208   -0.012%
cover     4847391   4837151   -10240  -0.211%
dist      3473877   3472565   -1312   -0.038%
doc       3821496   3820432   -1064   -0.028%
fix       3220587   3220659   +72     +0.002%
link      6587504   6582576   -4928   -0.075%
nm        4000154   3998690   -1464   -0.037%
objdump   4409449   4407625   -1824   -0.041%
pack      2398086   2393110   -4976   -0.207%
pprof     13599060  13606111  +7051   +0.052%
test2json 2645148   2645692   +544    +0.021%
trace     10355281  10355862  +581    +0.006%
vet       6780026   6779666   -360    -0.005%
total     106319929 106289641 -30288  -0.028%

Change-Id: Ia5399286958c187c8664c769bbddf7bc4c1cae99
Reviewed-on: https://go-review.googlesource.com/c/go/+/263600
Run-TryBot: eric fang <eric.fang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: eric fang <eric.fang@arm.com>
Reviewed-by: Keith Randall <khr@golang.org>

diff --git a/src/cmd/compile/internal/ssa/regalloc.go b/src/cmd/compile/internal/ssa/regalloc.go
index 691530ec0b468d9776f4fb931377da51d92b5742..7b97c8e097b109c421e919c6dad5f90f9bf6f834 100644 (file)
--- a/src/cmd/compile/internal/ssa/regalloc.go
+++ b/src/cmd/compile/internal/ssa/regalloc.go
@@ -1012,8 +1012,8 @@ func (s *regAllocState) regalloc(f *Func) {
                        // Copy phi ops into new schedule.
                        b.Values = append(b.Values, phis...)
 
-                       // Third pass - pick registers for phis whose inputs
-                       // were not in a register.
+                       // Third pass - pick registers for phis whose input
+                       // was not in a register in the primary predecessor.
                        for i, v := range phis {
                                if !s.values[v.ID].needReg {
                                        continue
@@ -1022,6 +1022,24 @@ func (s *regAllocState) regalloc(f *Func) {
                                        continue
                                }
                                m := s.compatRegs(v.Type) &^ phiUsed &^ s.used
+                               // If one of the other inputs of v is in a register, and the register is available,
+                               // select this register, which can save some unnecessary copies.
+                               for i, pe := range b.Preds {
+                                       if int32(i) == idx {
+                                               continue
+                                       }
+                                       ri := noRegister
+                                       for _, er := range s.endRegs[pe.b.ID] {
+                                               if er.v == s.orig[v.Args[i].ID] {
+                                                       ri = er.r
+                                                       break
+                                               }
+                                       }
+                                       if ri != noRegister && m>>ri&1 != 0 {
+                                               m = regMask(1) << ri
+                                               break
+                                       }
+                               }
                                if m != 0 {
                                        r := pickReg(m)
                                        phiRegs[i] = r
@@ -1119,7 +1137,19 @@ func (s *regAllocState) regalloc(f *Func) {
                                }
                                rp, ok := s.f.getHome(v.ID).(*Register)
                                if !ok {
-                                       continue
+                                       // If v is not assigned a register, pick a register assigned to one of v's inputs.
+                                       // Hopefully v will get assigned that register later.
+                                       // If the inputs have allocated register information, add it to desired,
+                                       // which may reduce spill or copy operations when the register is available.
+                                       for _, a := range v.Args {
+                                               rp, ok = s.f.getHome(a.ID).(*Register)
+                                               if ok {
+                                                       break
+                                               }
+                                       }
+                                       if !ok {
+                                               continue
+                                       }
                                }
                                desired.add(v.Args[pidx].ID, register(rp.num))
                        }