cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills. Storing newlen too early caused uses of append in the runtime itself to receive an inconsistent view of a slice, leading to corruption. This CL makes the generate code much more similar to the old backend. It spills more than before, but those spills have been contained to the grow path. It recalculates newlen unnecessarily on the fast path, but that's measurably cheaper than spilling it. CL 21891 caused runtime failures in 6 of 2000 runs of net/http and crypto/x509 in my test setup. This CL has gone 6000 runs without a failure. Benchmarks going from master to this CL: name old time/op new time/op delta AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27) AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29) AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30) AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26) AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30) AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29) AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30) AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30) AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29) AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27) However, for the large no-grow cases, there is still more work to be done. Going from this CL to the non-SSA backend: name old time/op new time/op delta AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29) AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26) AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29) AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28) AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28) AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29) AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30) AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29) AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30) AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26) New generated code: var x []byte func a() { x = append(x, 1) } "".a t=1 size=208 args=0x0 locals=0x48 0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0 0x0000 00000 (a.go:5) MOVQ (TLS), CX 0x0009 00009 (a.go:5) CMPQ SP, 16(CX) 0x000d 00013 (a.go:5) JLS 190 0x0013 00019 (a.go:5) SUBQ $72, SP 0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB) 0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB) 0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX 0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX 0x0025 00037 (a.go:6) MOVQ "".x(SB), BX 0x002c 00044 (a.go:6) LEAQ 1(DX), BP 0x0030 00048 (a.go:6) CMPQ BP, CX 0x0033 00051 (a.go:6) JGT $0, 73 0x0035 00053 (a.go:6) LEAQ 1(DX), AX 0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB) 0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1) 0x0044 00068 (a.go:7) ADDQ $72, SP 0x0048 00072 (a.go:7) RET 0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX 0x0050 00080 (a.go:6) MOVQ AX, (SP) 0x0054 00084 (a.go:6) MOVQ BX, 8(SP) 0x0059 00089 (a.go:6) MOVQ DX, 16(SP) 0x005e 00094 (a.go:6) MOVQ CX, 24(SP) 0x0063 00099 (a.go:6) MOVQ BP, 32(SP) 0x0068 00104 (a.go:6) PCDATA $0, $0 0x0068 00104 (a.go:6) CALL runtime.growslice(SB) 0x006d 00109 (a.go:6) MOVQ 40(SP), CX 0x0072 00114 (a.go:6) MOVQ 48(SP), DX 0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP) 0x007c 00124 (a.go:6) MOVQ 56(SP), BX 0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB) 0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX 0x008e 00142 (a.go:6) TESTB AL, AL 0x0090 00144 (a.go:6) JNE $0, 162 0x0092 00146 (a.go:6) MOVQ CX, "".x(SB) 0x0099 00153 (a.go:6) MOVQ "".x(SB), BX 0x00a0 00160 (a.go:6) JMP 53 0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX 0x00a9 00169 (a.go:6) MOVQ BX, (SP) 0x00ad 00173 (a.go:6) MOVQ CX, 8(SP) 0x00b2 00178 (a.go:6) PCDATA $0, $0 0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB) 0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX 0x00bc 00188 (a.go:6) JMP 153 0x00be 00190 (a.go:6) NOP 0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB) 0x00c3 00195 (a.go:5) JMP 0 Fixes #14969 again Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d Reviewed-on: https://go-review.googlesource.com/22197 Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
Showing
Please register or sign in to comment