1
0
mirror of https://github.com/golang/go synced 2024-11-23 17:00:07 -07:00

cmd/compile: prefer to evict a rematerializable register

This resolves a long-standing regalloc TODO:
If you must evict a register, choose to evict a register
containing a rematerializable value, since that value
won't need to be spilled.

Provides very minor performance and size improvements.

name                     old time/op    new time/op    delta
BinaryTree17-8              2.20s ± 3%     2.18s ± 2%  -0.77%  (p=0.000 n=45+49)
Fannkuch11-8                2.14s ± 2%     2.15s ± 2%  +0.73%  (p=0.000 n=43+44)
FmtFprintfEmpty-8          30.6ns ± 4%    30.2ns ± 3%  -1.14%  (p=0.000 n=50+48)
FmtFprintfString-8         54.5ns ± 6%    53.6ns ± 5%  -1.64%  (p=0.001 n=50+48)
FmtFprintfInt-8            58.0ns ± 7%    57.6ns ± 4%    ~     (p=0.220 n=50+50)
FmtFprintfIntInt-8         85.3ns ± 2%    84.8ns ± 3%  -0.62%  (p=0.001 n=44+47)
FmtFprintfPrefixedInt-8    93.9ns ± 6%    93.6ns ± 5%    ~     (p=0.706 n=50+48)
FmtFprintfFloat-8           178ns ± 4%     177ns ± 4%    ~     (p=0.107 n=49+50)
FmtManyArgs-8               376ns ± 4%     374ns ± 3%  -0.58%  (p=0.013 n=45+50)
GobDecode-8                4.77ms ± 2%    4.76ms ± 3%    ~     (p=0.059 n=47+46)
GobEncode-8                4.04ms ± 2%    3.99ms ± 3%  -1.13%  (p=0.000 n=49+49)
Gzip-8                      177ms ± 2%     180ms ± 3%  +1.43%  (p=0.000 n=48+48)
Gunzip-8                   28.5ms ± 6%    28.3ms ± 5%    ~     (p=0.104 n=50+49)
HTTPClientServer-8         72.1µs ± 1%    72.0µs ± 1%  -0.15%  (p=0.042 n=48+42)
JSONEncode-8               9.81ms ± 5%   10.03ms ± 6%  +2.29%  (p=0.000 n=50+49)
JSONDecode-8               39.2ms ± 3%    39.3ms ± 2%    ~     (p=0.095 n=49+49)
Mandelbrot200-8            3.48ms ± 2%    3.46ms ± 2%  -0.80%  (p=0.000 n=47+48)
GoParse-8                  2.54ms ± 3%    2.51ms ± 3%  -1.35%  (p=0.000 n=49+49)
RegexpMatchEasy0_32-8      66.0ns ± 7%    65.7ns ± 8%    ~     (p=0.331 n=50+50)
RegexpMatchEasy0_1K-8       155ns ± 4%     154ns ± 4%    ~     (p=0.986 n=49+50)
RegexpMatchEasy1_32-8      62.6ns ± 8%    62.2ns ± 5%    ~     (p=0.395 n=50+49)
RegexpMatchEasy1_1K-8       260ns ± 5%     255ns ± 3%  -1.92%  (p=0.000 n=49+49)
RegexpMatchMedium_32-8     92.9ns ± 2%    91.8ns ± 2%  -1.25%  (p=0.000 n=46+48)
RegexpMatchMedium_1K-8     27.7µs ± 3%    27.0µs ± 2%  -2.59%  (p=0.000 n=49+49)
RegexpMatchHard_32-8       1.23µs ± 4%    1.21µs ± 2%  -2.16%  (p=0.000 n=49+44)
RegexpMatchHard_1K-8       36.4µs ± 2%    35.7µs ± 2%  -1.87%  (p=0.000 n=48+49)
Revcomp-8                   274ms ± 2%     276ms ± 3%  +0.70%  (p=0.034 n=45+48)
Template-8                 45.1ms ± 8%    45.1ms ± 8%    ~     (p=0.643 n=50+50)
TimeParse-8                 223ns ± 2%     223ns ± 2%    ~     (p=0.401 n=47+47)
TimeFormat-8                245ns ± 2%     246ns ± 3%    ~     (p=0.758 n=49+50)
[Geo mean]                 36.5µs         36.3µs       -0.54%


name        old object-bytes  new object-bytes  delta
Template          480kB ± 0%        480kB ± 0%    ~     (all equal)
Unicode           214kB ± 0%        214kB ± 0%    ~     (all equal)
GoTypes          1.54MB ± 0%       1.54MB ± 0%  -0.03%  (p=0.008 n=5+5)
Compiler         5.75MB ± 0%       5.75MB ± 0%    ~     (all equal)
SSA              14.6MB ± 0%       14.6MB ± 0%  -0.01%  (p=0.008 n=5+5)
Flate             300kB ± 0%        300kB ± 0%  -0.01%  (p=0.008 n=5+5)
GoParser          366kB ± 0%        366kB ± 0%    ~     (all equal)
Reflect          1.20MB ± 0%       1.20MB ± 0%    ~     (all equal)
Tar               413kB ± 0%        413kB ± 0%    ~     (all equal)
XML               529kB ± 0%        528kB ± 0%  -0.13%  (p=0.008 n=5+5)
[Geo mean]        909kB             909kB       -0.02%


Change-Id: I46d37a55197683a98913f35801dc2b0d609653c8
Reviewed-on: https://go-review.googlesource.com/103240
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
This commit is contained in:
Josh Bleecher Snyder 2018-03-16 07:15:59 -07:00
parent 5d8d3d52b0
commit 014a9048d4

View File

@ -1749,6 +1749,7 @@ type edgeState struct {
usedRegs regMask // registers currently holding something
uniqueRegs regMask // registers holding the only copy of a value
finalRegs regMask // registers holding final target
rematerializeableRegs regMask // registers that hold rematerializeable values
}
type contentRecord struct {
@ -1782,6 +1783,7 @@ func (e *edgeState) setup(idx int, srcReg []endReg, dstReg []startReg, stacklive
e.usedRegs = 0
e.uniqueRegs = 0
e.finalRegs = 0
e.rematerializeableRegs = 0
// Live registers can be sources.
for _, x := range srcReg {
@ -2041,6 +2043,9 @@ func (e *edgeState) set(loc Location, vid ID, c *Value, final bool, pos src.XPos
e.uniqueRegs &^= regMask(1) << uint(t.num)
}
}
if e.s.values[vid].rematerializeable {
e.rematerializeableRegs |= regMask(1) << uint(r.num)
}
}
if e.s.f.pass.debug > regDebug {
fmt.Printf("%s\n", c.LongString())
@ -2082,6 +2087,7 @@ func (e *edgeState) erase(loc Location) {
if cr.final {
e.finalRegs &^= regMask(1) << uint(r.num)
}
e.rematerializeableRegs &^= regMask(1) << uint(r.num)
}
if len(a) == 1 {
if r, ok := e.s.f.getHome(a[0].ID).(*Register); ok {
@ -2105,7 +2111,7 @@ func (e *edgeState) findRegFor(typ *types.Type) Location {
// 1) an unused register
// 2) a non-unique register not holding a final value
// 3) a non-unique register
// 4) TODO: a register holding a rematerializeable value
// 4) a register holding a rematerializeable value
x := m &^ e.usedRegs
if x != 0 {
return &e.s.registers[pickReg(x)]
@ -2118,6 +2124,10 @@ func (e *edgeState) findRegFor(typ *types.Type) Location {
if x != 0 {
return &e.s.registers[pickReg(x)]
}
x = m & e.rematerializeableRegs
if x != 0 {
return &e.s.registers[pickReg(x)]
}
// No register is available.
// Pick a register to spill.
@ -2145,7 +2155,7 @@ func (e *edgeState) findRegFor(typ *types.Type) Location {
}
}
fmt.Printf("m:%d unique:%d final:%d\n", m, e.uniqueRegs, e.finalRegs)
fmt.Printf("m:%d unique:%d final:%d rematerializable:%d\n", m, e.uniqueRegs, e.finalRegs, e.rematerializeableRegs)
for _, vid := range e.cachedVals {
a := e.cache[vid]
for _, c := range a {