Skip to content

Commit 88f423e

Browse files
josharianjosharian
committed
cmd/internal/obj/x86: improve static branch prediction for wrapper prologue
Static branch prediction assumes that forward branches are not taken. The existing wrapper prologue almost always takes the first forward branch. Move the rare case to the end of the function. This CL is amd64 only. Other architectures will be done in separate CLs. Updates #19042. Package sort benchmarks: SearchWrappers-8 104ns ± 2% 104ns ± 0% -0.41% (p=0.006 n=30+41) SortString1K-8 128µs ± 1% 128µs ± 1% -0.25% (p=0.045 n=30+56) SortString1K_Slice-8 117µs ± 1% 117µs ± 1% ~ (p=0.855 n=30+59) StableString1K-8 18.6µs ± 1% 18.6µs ± 1% ~ (p=0.599 n=29+60) SortInt1K-8 61.0µs ± 1% 56.5µs ± 1% -7.36% (p=0.000 n=29+58) StableInt1K-8 74.6µs ± 1% 70.4µs ± 3% -5.54% (p=0.000 n=28+60) StableInt1K_Slice-8 59.9µs ± 1% 58.3µs ± 4% -2.64% (p=0.000 n=29+60) SortInt64K-8 6.02ms ± 2% 5.98ms ± 2% -0.60% (p=0.000 n=29+59) SortInt64K_Slice-8 5.07ms ± 2% 5.05ms ± 2% -0.38% (p=0.006 n=30+58) StableInt64K-8 6.41ms ± 1% 6.22ms ± 1% -3.00% (p=0.000 n=27+58) Sort1e2-8 37.4µs ± 1% 37.1µs ± 1% -0.91% (p=0.000 n=30+57) Stable1e2-8 74.8µs ± 1% 75.2µs ± 1% +0.52% (p=0.000 n=30+57) Sort1e4-8 8.11ms ± 1% 8.01ms ± 1% -1.20% (p=0.000 n=30+59) Stable1e4-8 24.3ms ± 1% 24.3ms ± 1% ~ (p=0.157 n=30+60) Sort1e6-8 1.25s ± 1% 1.23s ± 1% -1.43% (p=0.000 n=29+58) Stable1e6-8 4.93s ± 1% 4.90s ± 1% -0.56% (p=0.000 n=29+59) [Geo mean] 720µs 709µs -1.52% Assembly for sort.(*intPairs).Swap: Before: "".(*intPairs).Swap t=1 size=147 args=0x18 locals=0x8 0x0000 00000 (<autogenerated>:1) TEXT "".(*intPairs).Swap(SB), $8-24 0x0000 00000 (<autogenerated>:1) MOVQ (TLS), CX 0x0009 00009 (<autogenerated>:1) SUBQ $8, SP 0x000d 00013 (<autogenerated>:1) MOVQ BP, (SP) 0x0011 00017 (<autogenerated>:1) LEAQ (SP), BP 0x0015 00021 (<autogenerated>:1) MOVQ 32(CX), BX 0x0019 00025 (<autogenerated>:1) TESTQ BX, BX 0x001c 00028 (<autogenerated>:1) JEQ 43 0x001e 00030 (<autogenerated>:1) LEAQ 16(SP), DI 0x0023 00035 (<autogenerated>:1) CMPQ (BX), DI 0x0026 00038 (<autogenerated>:1) JNE 43 0x0028 00040 (<autogenerated>:1) MOVQ SP, (BX) 0x002b 00043 (<autogenerated>:1) NOP 0x002b 00043 (<autogenerated>:1) FUNCDATA $0, gclocals·e6397a44f8e1b6e77d0f200b4fba5269(SB) 0x002b 00043 (<autogenerated>:1) FUNCDATA $1, gclocals·69c1753bd5f81501d95132d08af04464(SB) 0x002b 00043 (<autogenerated>:1) MOVQ ""..this+16(FP), AX 0x0030 00048 (<autogenerated>:1) TESTQ AX, AX 0x0033 00051 (<autogenerated>:1) JEQ $0, 140 0x0035 00053 (<autogenerated>:1) MOVQ (AX), CX 0x0038 00056 (<autogenerated>:1) MOVQ 8(AX), AX 0x003c 00060 (<autogenerated>:1) MOVQ "".i+24(FP), DX 0x0041 00065 (<autogenerated>:1) CMPQ DX, AX 0x0044 00068 (<autogenerated>:1) JCC $0, 133 0x0046 00070 (<autogenerated>:1) SHLQ $4, DX 0x004a 00074 (<autogenerated>:1) MOVQ 8(CX)(DX*1), BX 0x004f 00079 (<autogenerated>:1) MOVQ (CX)(DX*1), SI 0x0053 00083 (<autogenerated>:1) MOVQ "".j+32(FP), DI 0x0058 00088 (<autogenerated>:1) CMPQ DI, AX 0x005b 00091 (<autogenerated>:1) JCC $0, 133 0x005d 00093 (<autogenerated>:1) SHLQ $4, DI 0x0061 00097 (<autogenerated>:1) MOVQ 8(CX)(DI*1), AX 0x0066 00102 (<autogenerated>:1) MOVQ (CX)(DI*1), R8 0x006a 00106 (<autogenerated>:1) MOVQ R8, (CX)(DX*1) 0x006e 00110 (<autogenerated>:1) MOVQ AX, 8(CX)(DX*1) 0x0073 00115 (<autogenerated>:1) MOVQ SI, (CX)(DI*1) 0x0077 00119 (<autogenerated>:1) MOVQ BX, 8(CX)(DI*1) 0x007c 00124 (<autogenerated>:1) MOVQ (SP), BP 0x0080 00128 (<autogenerated>:1) ADDQ $8, SP 0x0084 00132 (<autogenerated>:1) RET 0x0085 00133 (<autogenerated>:1) PCDATA $0, $1 0x0085 00133 (<autogenerated>:1) CALL runtime.panicindex(SB) 0x008a 00138 (<autogenerated>:1) UNDEF 0x008c 00140 (<autogenerated>:1) PCDATA $0, $1 0x008c 00140 (<autogenerated>:1) CALL runtime.panicwrap(SB) 0x0091 00145 (<autogenerated>:1) UNDEF After: "".(*intPairs).Swap t=1 size=149 args=0x18 locals=0x8 0x0000 00000 (<autogenerated>:1) TEXT "".(*intPairs).Swap(SB), $8-24 0x0000 00000 (<autogenerated>:1) MOVQ (TLS), CX 0x0009 00009 (<autogenerated>:1) SUBQ $8, SP 0x000d 00013 (<autogenerated>:1) MOVQ BP, (SP) 0x0011 00017 (<autogenerated>:1) LEAQ (SP), BP 0x0015 00021 (<autogenerated>:1) MOVQ 32(CX), BX 0x0019 00025 (<autogenerated>:1) TESTQ BX, BX 0x001c 00028 (<autogenerated>:1) JNE 134 0x001e 00030 (<autogenerated>:1) NOP 0x001e 00030 (<autogenerated>:1) FUNCDATA $0, gclocals·e6397a44f8e1b6e77d0f200b4fba5269(SB) 0x001e 00030 (<autogenerated>:1) FUNCDATA $1, gclocals·69c1753bd5f81501d95132d08af04464(SB) 0x001e 00030 (<autogenerated>:1) MOVQ ""..this+16(FP), AX 0x0023 00035 (<autogenerated>:1) TESTQ AX, AX 0x0026 00038 (<autogenerated>:1) JEQ $0, 127 0x0028 00040 (<autogenerated>:1) MOVQ (AX), CX 0x002b 00043 (<autogenerated>:1) MOVQ 8(AX), AX 0x002f 00047 (<autogenerated>:1) MOVQ "".i+24(FP), DX 0x0034 00052 (<autogenerated>:1) CMPQ DX, AX 0x0037 00055 (<autogenerated>:1) JCC $0, 120 0x0039 00057 (<autogenerated>:1) SHLQ $4, DX 0x003d 00061 (<autogenerated>:1) MOVQ 8(CX)(DX*1), BX 0x0042 00066 (<autogenerated>:1) MOVQ (CX)(DX*1), SI 0x0046 00070 (<autogenerated>:1) MOVQ "".j+32(FP), DI 0x004b 00075 (<autogenerated>:1) CMPQ DI, AX 0x004e 00078 (<autogenerated>:1) JCC $0, 120 0x0050 00080 (<autogenerated>:1) SHLQ $4, DI 0x0054 00084 (<autogenerated>:1) MOVQ 8(CX)(DI*1), AX 0x0059 00089 (<autogenerated>:1) MOVQ (CX)(DI*1), R8 0x005d 00093 (<autogenerated>:1) MOVQ R8, (CX)(DX*1) 0x0061 00097 (<autogenerated>:1) MOVQ AX, 8(CX)(DX*1) 0x0066 00102 (<autogenerated>:1) MOVQ SI, (CX)(DI*1) 0x006a 00106 (<autogenerated>:1) MOVQ BX, 8(CX)(DI*1) 0x006f 00111 (<autogenerated>:1) MOVQ (SP), BP 0x0073 00115 (<autogenerated>:1) ADDQ $8, SP 0x0077 00119 (<autogenerated>:1) RET 0x0078 00120 (<autogenerated>:1) PCDATA $0, $1 0x0078 00120 (<autogenerated>:1) CALL runtime.panicindex(SB) 0x007d 00125 (<autogenerated>:1) UNDEF 0x007f 00127 (<autogenerated>:1) PCDATA $0, $1 0x007f 00127 (<autogenerated>:1) CALL runtime.panicwrap(SB) 0x0084 00132 (<autogenerated>:1) UNDEF 0x0086 00134 (<autogenerated>:1) LEAQ 16(SP), DI 0x008b 00139 (<autogenerated>:1) CMPQ (BX), DI 0x008e 00142 (<autogenerated>:1) JNE 30 0x0090 00144 (<autogenerated>:1) MOVQ SP, (BX) 0x0093 00147 (<autogenerated>:1) JMP 30 Change-Id: Ie8c37f384bba10fbacaa754bb0a6b0a7e520ef01 Reviewed-on: https://go-review.googlesource.com/36893 Reviewed-by: Keith Randall <[email protected]>
1 parent f7f3514 commit 88f423e

File tree

1 file changed

+35
-16
lines changed

1 file changed

+35
-16
lines changed

‎src/cmd/internal/obj/x86/obj6.go

+35-16
Original file line numberDiff line numberDiff line change
@@ -748,16 +748,22 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
748748
//
749749
// MOVQ g_panic(CX), BX
750750
// TESTQ BX, BX
751-
// JEQ end
751+
// JNE checkargp
752+
// end:
753+
// NOP
754+
// ... rest of function ...
755+
// checkargp:
752756
// LEAQ (autoffset+8)(SP), DI
753757
// CMPQ panic_argp(BX), DI
754758
// JNE end
755-
// MOVQ SP, panic_argp(BX)
756-
// end:
757-
// NOP
759+
// MOVQ SP, panic_argp(BX)
760+
// JMP end
758761
//
759762
// The NOP is needed to give the jumps somewhere to land.
760763
// It is a liblink NOP, not an x86 NOP: it encodes to 0 instruction bytes.
764+
//
765+
// The layout is chosen to help static branch prediction:
766+
// Both conditional jumps are unlikely, so they are arranged to be forward jumps.
761767

762768
// MOVQ g_panic(CX), BX
763769
p = obj.Appendp(ctxt, p)
@@ -789,14 +795,23 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
789795
p.As = ATESTL
790796
}
791797

792-
// JEQ end
793-
p = obj.Appendp(ctxt, p)
794-
p.As = AJEQ
795-
p.To.Type = obj.TYPE_BRANCH
796-
p1 := p
798+
// JNE checkargp (checkargp to be resolved later)
799+
jne := obj.Appendp(ctxt, p)
800+
jne.As = AJNE
801+
jne.To.Type = obj.TYPE_BRANCH
802+
803+
// end:
804+
// NOP
805+
end := obj.Appendp(ctxt, jne)
806+
end.As = obj.ANOP
807+
808+
// Fast forward to end of function.
809+
var last *obj.Prog
810+
for last = end; last.Link != nil; last = last.Link {
811+
}
797812

798813
// LEAQ (autoffset+8)(SP), DI
799-
p = obj.Appendp(ctxt, p)
814+
p = obj.Appendp(ctxt, last)
800815
p.As = ALEAQ
801816
p.From.Type = obj.TYPE_MEM
802817
p.From.Reg = REG_SP
@@ -807,6 +822,9 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
807822
p.As = ALEAL
808823
}
809824

825+
// Set jne branch target.
826+
jne.Pcond = p
827+
810828
// CMPQ panic_argp(BX), DI
811829
p = obj.Appendp(ctxt, p)
812830
p.As = ACMPQ
@@ -830,7 +848,7 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
830848
p = obj.Appendp(ctxt, p)
831849
p.As = AJNE
832850
p.To.Type = obj.TYPE_BRANCH
833-
p2 := p
851+
p.Pcond = end
834852

835853
// MOVQ SP, panic_argp(BX)
836854
p = obj.Appendp(ctxt, p)
@@ -851,13 +869,14 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
851869
p.As = AMOVL
852870
}
853871

854-
// NOP
872+
// JMP end
855873
p = obj.Appendp(ctxt, p)
856-
p.As = obj.ANOP
874+
p.As = obj.AJMP
875+
p.To.Type = obj.TYPE_BRANCH
876+
p.Pcond = end
857877

858-
// Set targets for jumps above to the NOP
859-
p1.Pcond = p
860-
p2.Pcond = p
878+
// Reset p for following code.
879+
p = end
861880
}
862881

863882
for ; p != nil; p = p.Link {

0 commit comments

Comments
 (0)