Skip to content

Commit

Permalink
cmd/internal/obj/x86: make function prologue more predictable
Browse files Browse the repository at this point in the history 
Static branch prediction guesses that forward branches aren't taken.
Since stacks are rarely grown, make the forward branch mean grow.

Sample disassembly for

func f() {
	_ = [128]byte{}
}

Before:

TEXT main.f(SB) x.go
	x.go:3	0x2000	65488b0c25a0080000	GS MOVQ GS:0x8a0, CX
	x.go:3	0x2009	483b6110		CMPQ 0x10(CX), SP
	x.go:3	0x200d	7707			JA 0x2016
	x.go:3	0x200f	e88c410400		CALL runtime.morestack_noctxt(SB)
	x.go:3	0x2014	ebea			JMP main.f(SB)
	x.go:3	0x2016	4881ec80000000		SUBQ $0x80, SP
	x.go:4	0x201d	488d3c24		LEAQ 0(SP), DI
	x.go:4	0x2021	31c0			XORL AX, AX
	x.go:4	0x2023	e8cc640400		CALL 0x484f4
	x.go:5	0x2028	4881c480000000		ADDQ $0x80, SP
	x.go:5	0x202f	c3			RET

After:

TEXT main.f(SB) x.go
	x.go:3	0x2000	65488b0c25a0080000	GS MOVQ GS:0x8a0, CX
	x.go:3	0x2009	483b6110		CMPQ 0x10(CX), SP
	x.go:3	0x200d	761a			JBE 0x2029
	x.go:3	0x200f	4881ec80000000		SUBQ $0x80, SP
	x.go:4	0x2016	488d3c24		LEAQ 0(SP), DI
	x.go:4	0x201a	31c0			XORL AX, AX
	x.go:4	0x201c	e813740400		CALL 0x49434
	x.go:5	0x2021	4881c480000000		ADDQ $0x80, SP
	x.go:5	0x2028	c3			RET
	x.go:3	0x2029	e8224f0400		CALL runtime.morestack_noctxt(SB)
	x.go:3	0x202e	ebd0			JMP main.f(SB)

Updates golang#10587.

Sample benchmarks on a 2.8 GHz Intel Core i7:

package sort

name            old mean              new mean              delta
SearchWrappers   134ns × (0.99,1.01)   132ns × (0.99,1.01)  -1.73% (p=0.000 n=15+14)
SortString1K     215µs × (0.99,1.01)   213µs × (0.99,1.01)  -0.61% (p=0.020 n=14+15)
StableString1K   311µs × (0.99,1.02)   309µs × (0.99,1.02)    ~    (p=0.077 n=14+15)
SortInt1K        103µs × (0.99,1.02)   100µs × (0.98,1.01)  -3.34% (p=0.000 n=15+15)
StableInt1K      102µs × (0.99,1.01)    98µs × (0.97,1.04)  -3.53% (p=0.000 n=15+15)
SortInt64K      10.1ms × (0.98,1.02)   9.7ms × (0.99,1.01)  -3.86% (p=0.000 n=14+15)
StableInt64K    8.70ms × (0.99,1.01)  8.44ms × (0.99,1.03)  -2.93% (p=0.000 n=14+15)
Sort1e2         51.2µs × (1.00,1.01)  48.9µs × (0.99,1.02)  -4.48% (p=0.000 n=13+15)
Stable1e2        100µs × (0.99,1.02)    99µs × (0.99,1.01)  -1.15% (p=0.000 n=14+13)
Sort1e4         11.1ms × (0.99,1.02)  10.4ms × (0.99,1.01)  -6.02% (p=0.000 n=15+14)
Stable1e4       30.6ms × (0.99,1.01)  30.3ms × (0.99,1.02)  -1.02% (p=0.001 n=15+14)
Sort1e6          1.75s × (0.99,1.02)   1.66s × (0.98,1.03)  -4.95% (p=0.000 n=14+15)
Stable1e6        6.31s × (0.99,1.01)   6.26s × (0.99,1.01)  -0.79% (p=0.002 n=15+15)

package regexp

name                          old mean              new mean              delta
Literal                        131ns × (0.99,1.01)   130ns × (0.99,1.03)  -1.07% (p=0.004 n=14+15)
NotLiteral                    2.13µs × (0.99,1.01)  2.01µs × (0.99,1.03)  -5.71% (p=0.000 n=14+14)
MatchClass                    3.15µs × (0.99,1.01)  3.04µs × (0.99,1.02)  -3.40% (p=0.000 n=15+15)
MatchClass_InRange            2.92µs × (0.99,1.01)  2.77µs × (0.99,1.02)  -5.05% (p=0.000 n=13+15)
ReplaceAll                    2.17µs × (0.99,1.02)  2.06µs × (0.99,1.01)  -5.19% (p=0.000 n=15+13)
AnchoredLiteralShortNonMatch   116ns × (0.99,1.02)   113ns × (0.99,1.01)  -2.75% (p=0.000 n=15+14)
AnchoredLiteralLongNonMatch    125ns × (0.99,1.01)   127ns × (0.98,1.02)  +1.49% (p=0.000 n=15+15)
AnchoredShortMatch             178ns × (0.99,1.02)   175ns × (0.99,1.01)  -1.62% (p=0.000 n=15+13)
AnchoredLongMatch              328ns × (0.99,1.00)   341ns × (0.99,1.01)  +3.73% (p=0.000 n=12+15)
OnePassShortA                  773ns × (0.99,1.02)   752ns × (0.99,1.01)  -2.78% (p=0.000 n=15+13)
NotOnePassShortA               794ns × (0.99,1.03)   780ns × (0.99,1.02)  -1.75% (p=0.001 n=15+15)
OnePassShortB                  608ns × (0.99,1.01)   591ns × (0.99,1.02)  -2.86% (p=0.000 n=15+14)
NotOnePassShortB               576ns × (0.99,1.01)   571ns × (0.99,1.02)  -0.74% (p=0.035 n=15+15)
OnePassLongPrefix              131ns × (0.99,1.02)   130ns × (0.99,1.02)  -1.32% (p=0.003 n=15+15)
OnePassLongNotPrefix           503ns × (0.99,1.02)   481ns × (0.99,1.01)  -4.34% (p=0.000 n=15+13)
MatchEasy0_32                  102ns × (0.98,1.01)   101ns × (0.99,1.02)    ~    (p=0.907 n=15+14)
MatchEasy0_1K                  617ns × (0.99,1.02)   634ns × (0.98,1.02)  +2.77% (p=0.000 n=15+15)
MatchEasy0_32K                10.9µs × (0.99,1.01)  11.1µs × (0.99,1.01)  +1.59% (p=0.000 n=15+15)
MatchEasy0_1M                  406µs × (0.99,1.02)   410µs × (0.99,1.02)  +1.01% (p=0.000 n=14+15)
MatchEasy0_32M                13.4ms × (0.99,1.01)  13.7ms × (0.99,1.02)  +1.64% (p=0.000 n=12+15)
MatchEasy1_32                 83.7ns × (0.98,1.02)  83.0ns × (0.98,1.02)    ~    (p=0.190 n=15+15)
MatchEasy1_1K                 1.46µs × (0.99,1.02)  1.39µs × (0.99,1.02)  -4.83% (p=0.000 n=15+15)
MatchEasy1_32K                49.4µs × (0.99,1.01)  49.4µs × (0.99,1.01)    ~    (p=0.205 n=15+15)
MatchEasy1_1M                 1.72ms × (0.99,1.02)  1.75ms × (0.99,1.01)  +1.34% (p=0.000 n=15+15)
MatchEasy1_32M                55.5ms × (0.99,1.01)  56.1ms × (0.99,1.02)  +1.10% (p=0.002 n=15+15)
MatchMedium_32                1.37µs × (0.99,1.04)  1.33µs × (0.99,1.01)  -2.87% (p=0.000 n=15+15)
MatchMedium_1K                41.1µs × (0.99,1.02)  40.4µs × (0.99,1.02)  -1.59% (p=0.000 n=15+15)
MatchMedium_32K               1.71ms × (0.99,1.01)  1.75ms × (0.99,1.02)  +2.36% (p=0.000 n=14+15)
MatchMedium_1M                54.5ms × (0.99,1.01)  56.1ms × (0.99,1.01)  +2.94% (p=0.000 n=13+15)
MatchMedium_32M                1.75s × (0.99,1.01)   1.80s × (0.99,1.01)  +2.77% (p=0.000 n=15+15)
MatchHard_32                  2.12µs × (0.99,1.02)  2.06µs × (0.99,1.01)  -2.60% (p=0.000 n=15+14)
MatchHard_1K                  64.4µs × (0.98,1.02)  62.2µs × (0.99,1.01)  -3.33% (p=0.000 n=15+15)
MatchHard_32K                 2.74ms × (0.99,1.01)  2.75ms × (0.99,1.01)    ~    (p=0.310 n=15+14)
MatchHard_1M                  87.1ms × (0.99,1.02)  88.2ms × (0.99,1.01)  +1.36% (p=0.000 n=14+15)
MatchHard_32M                  2.79s × (0.99,1.02)   2.83s × (0.99,1.02)  +1.26% (p=0.004 n=15+14)

go1 benchmarks

name                   old time/op    new time/op    delta
BinaryTree17              3.34s ± 3%     3.28s ± 2%  -1.86%  (p=0.000 n=67+66)
Fannkuch11                2.50s ± 1%     2.51s ± 1%  +0.24%  (p=0.016 n=63+66)
FmtFprintfEmpty          50.3ns ± 1%    50.2ns ± 2%  -0.30%  (p=0.001 n=62+67)
FmtFprintfString          178ns ± 1%     166ns ± 1%  -7.10%  (p=0.000 n=62+59)
FmtFprintfInt             168ns ± 1%     161ns ± 2%  -4.41%  (p=0.000 n=66+64)
FmtFprintfIntInt          292ns ± 1%     282ns ± 2%  -3.55%  (p=0.000 n=62+60)
FmtFprintfPrefixedInt     245ns ± 2%     239ns ± 2%  -2.24%  (p=0.000 n=66+65)
FmtFprintfFloat           338ns ± 2%     326ns ± 1%  -3.42%  (p=0.000 n=64+59)
FmtManyArgs              1.14µs ± 1%    1.10µs ± 2%  -3.55%  (p=0.000 n=62+62)
GobDecode                8.88ms ± 2%    8.74ms ± 1%  -1.55%  (p=0.000 n=66+62)
GobEncode                6.84ms ± 2%    6.61ms ± 2%  -3.32%  (p=0.000 n=61+67)
Gzip                      356ms ± 2%     352ms ± 2%  -1.07%  (p=0.000 n=67+66)
Gunzip                   90.6ms ± 2%    89.8ms ± 1%  -0.83%  (p=0.000 n=65+64)
HTTPClientServer         82.6µs ± 2%    82.5µs ± 2%    ~     (p=0.832 n=65+63)
JSONEncode               17.5ms ± 2%    16.8ms ± 2%  -3.77%  (p=0.000 n=63+63)
JSONDecode               63.3ms ± 2%    59.0ms ± 2%  -6.85%  (p=0.000 n=64+63)
Mandelbrot200            3.85ms ± 1%    3.85ms ± 1%    ~     (p=0.127 n=65+62)
GoParse                  3.75ms ± 2%    3.66ms ± 2%  -2.39%  (p=0.000 n=66+64)
RegexpMatchEasy0_32       100ns ± 2%     100ns ± 1%  -0.65%  (p=0.000 n=62+64)
RegexpMatchEasy0_1K       342ns ± 1%     341ns ± 1%  -0.43%  (p=0.000 n=65+64)
RegexpMatchEasy1_32      82.8ns ± 2%    82.8ns ± 2%    ~     (p=0.977 n=63+64)
RegexpMatchEasy1_1K       511ns ± 2%     506ns ± 2%  -1.01%  (p=0.000 n=63+64)
RegexpMatchMedium_32      139ns ± 1%     134ns ± 3%  -3.27%  (p=0.000 n=59+60)
RegexpMatchMedium_1K     41.8µs ± 2%    40.5µs ± 2%  -3.05%  (p=0.000 n=62+64)
RegexpMatchHard_32       2.13µs ± 1%    2.09µs ± 1%  -2.22%  (p=0.000 n=60+65)
RegexpMatchHard_1K       64.4µs ± 3%    62.8µs ± 2%  -2.58%  (p=0.000 n=65+59)
Revcomp                   531ms ± 2%     529ms ± 1%  -0.28%  (p=0.022 n=61+61)
Template                 73.2ms ± 1%    73.1ms ± 1%    ~     (p=0.794 n=66+63)
TimeParse                 369ns ± 1%     352ns ± 1%  -4.68%  (p=0.000 n=65+66)
TimeFormat                374ns ± 2%     348ns ± 2%  -7.01%  (p=0.000 n=66+64)

Change-Id: Ib190b5bb48a3e9087711d9e3383621d3103dd342
Reviewed-on: https://go-review.googlesource.com/10367
Reviewed-by: Russ Cox <[email protected]>
  • Loading branch information
@josharian
josharian committed Jun 4, 2015
1 parent 2db587c commit f4b48de
Showing 1 changed file with 29 additions and 33 deletions.
62 changes: 29 additions & 33 deletions src/cmd/internal/obj/x86/obj6.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -497,9 +497,8 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
p = load_g_cx(ctxt, p) // load g into CX
}

var q *obj.Prog
if cursym.Text.From3Offset()&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autoffset, int32(textarg), &q) // emit split check
p = stacksplit(ctxt, p, autoffset, int32(textarg)) // emit split check
}

if autoffset != 0 {
Expand All @@ -524,9 +523,6 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
p.Spadj = int32(ctxt.Arch.Ptrsize)
}

if q != nil {
q.Pcond = p
}
deltasp := autoffset

if bpsize > 0 {
Expand Down Expand Up @@ -856,9 +852,7 @@ func load_g_cx(ctxt *obj.Link, p *obj.Prog) *obj.Prog {
// Appends to (does not overwrite) p.
// Assumes g is in CX.
// Returns last new instruction.
// On return, *jmpok is the instruction that should jump
// to the stack frame allocation if no split is needed.
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32, textarg int32, jmpok **obj.Prog) *obj.Prog {
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32, textarg int32) *obj.Prog {
cmp := ACMPQ
lea := ALEAQ
mov := AMOVQ
Expand Down Expand Up @@ -973,37 +967,39 @@ func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32, textarg int32, jmp
}

// common
p = obj.Appendp(ctxt, p)

p.As = AJHI
p.To.Type = obj.TYPE_BRANCH
q := p

p = obj.Appendp(ctxt, p)
p.As = obj.ACALL
p.To.Type = obj.TYPE_BRANCH
if ctxt.Cursym.Cfunc != 0 {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestackc", 0)
} else if ctxt.Cursym.Text.From3Offset()&obj.NEEDCTXT == 0 {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
} else {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack", 0)
}
jls := obj.Appendp(ctxt, p)
jls.As = AJLS
jls.To.Type = obj.TYPE_BRANCH

p = obj.Appendp(ctxt, p)
p.As = obj.AJMP
p.To.Type = obj.TYPE_BRANCH
p.Pcond = ctxt.Cursym.Text.Link
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}

if q != nil {
q.Pcond = p.Link
call := obj.Appendp(ctxt, last)
call.Lineno = ctxt.Cursym.Text.Lineno
call.Mode = ctxt.Cursym.Text.Mode
call.As = obj.ACALL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc != 0:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3Offset()&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)

jmp := obj.Appendp(ctxt, call)
jmp.As = obj.AJMP
jmp.To.Type = obj.TYPE_BRANCH
jmp.Pcond = ctxt.Cursym.Text.Link

jls.Pcond = call
if q1 != nil {
q1.Pcond = q.Link
q1.Pcond = call
}

*jmpok = q
return p
return jls
}

func follow(ctxt *obj.Link, s *obj.LSym) {
Expand Down

0 comments on commit f4b48de

Please sign in to comment.