cmd/compile: performance slowdown with zero-length field added

[dsnet]

Using go1.10.1

Consider the following benchmark:

type Scalar struct {
	incomparable [0]func()
	A, B *int
	C, D uint64
}

func Foo(v int64) Scalar {
	return Scalar{C: uint64(v), D: 5}
}

var sink Scalar

func BenchmarkA(b *testing.B) {
	for i := 0; i < b.N; i++ {
		sink = Scalar{C: uint64(i), D: 5}
	}
}
func BenchmarkB(b *testing.B) {
	for i := 0; i < b.N; i++ {
		sink = Foo(int64(i))
	}
}

One my machine, this takes:

BenchmarkA-8   	200000000	         6.15 ns/op
BenchmarkB-8   	200000000	         8.69 ns/op

which is an abnormally high amount of time to copy 32 bytes.

If you comment out the incomparable field, the benchmarks dramatically improves:

BenchmarkA-8   	2000000000	         0.84 ns/op
BenchmarkB-8   	2000000000	         0.84 ns/op

I don't quite see how a zero-length field could make such a big impact of performance.

The assembly for BenchmarkB (with incomparable):

0x0000 00000	TEXT	"".BenchmarkB(SB), $104-8
0x0000 00000	MOVQ	(TLS), CX
0x0009 00009	CMPQ	SP, 16(CX)
0x000d 00013	JLS	241
0x0013 00019	SUBQ	$104, SP
0x0017 00023	MOVQ	BP, 96(SP)
0x001c 00028	LEAQ	96(SP), BP
0x0021 00033	FUNCDATA	$0, gclocals·a36216b97439c93dafebe03e7f0808b5(SB)
0x0021 00033	FUNCDATA	$1, gclocals·1d0ed49f611d7e40a62328b5976a2ede(SB)
0x0021 00033	XORL	AX, AX
0x0023 00035	JMP	40
0x0025 00037	INCQ	AX
0x0028 00040	MOVQ	"".b+112(SP), CX
0x002d 00045	MOVQ	240(CX), DX
0x0034 00052	CMPQ	AX, DX
0x0037 00055	JGE	231
0x003d 00061	XORPS	X0, X0
0x0040 00064	MOVUPS	X0, "".~r1+32(SP)
0x0045 00069	MOVUPS	X0, "".~r1+48(SP)
0x004a 00074	MOVUPS	X0, ""..autotmp_6+64(SP)
0x004f 00079	MOVUPS	X0, ""..autotmp_6+80(SP)
0x0054 00084	MOVQ	AX, ""..autotmp_6+80(SP)
0x0059 00089	MOVQ	$5, ""..autotmp_6+88(SP)
0x0062 00098	LEAQ	"".~r1+32(SP), DI
0x0067 00103	LEAQ	""..autotmp_6+64(SP), SI
0x006c 00108	DUFFCOPY	$868
0x007f 00127	MOVL	runtime.writeBarrier(SB), DX
0x0085 00133	TESTL	DX, DX
0x0087 00135	JNE	173
0x0089 00137	LEAQ	"".sink(SB), DI
0x0090 00144	LEAQ	"".~r1+32(SP), SI
0x0095 00149	DUFFCOPY	$868
0x00a8 00168	JMP	37
0x00ad 00173	MOVQ	AX, ""..autotmp_7+24(SP)
0x00b2 00178	LEAQ	type."".Scalar(SB), AX
0x00b9 00185	MOVQ	AX, (SP)
0x00bd 00189	LEAQ	"".sink(SB), AX
0x00c4 00196	MOVQ	AX, 8(SP)
0x00c9 00201	LEAQ	"".~r1+32(SP), CX
0x00ce 00206	MOVQ	CX, 16(SP)
0x00d3 00211	PCDATA	$0, $0
0x00d3 00211	CALL	runtime.typedmemmove(SB)
0x00d8 00216	MOVQ	""..autotmp_7+24(SP), AX
0x00dd 00221	MOVQ	"".b+112(SP), CX
0x00e2 00226	JMP	37
0x00e7 00231	MOVQ	96(SP), BP
0x00ec 00236	ADDQ	$104, SP
0x00f0 00240	RET
0x00f1 00241	NOP
0x00f1 00241	PCDATA	$0, $-1
0x00f1 00241	CALL	runtime.morestack_noctxt(SB)
0x00f6 00246	JMP	0

The assembly for BenchmarkA (with incomparable):

0x0000 00000	TEXT	"".BenchmarkA(SB), $72-8
0x0000 00000	MOVQ	(TLS), CX
0x0009 00009	CMPQ	SP, 16(CX)
0x000d 00013	JLS	199
0x0013 00019	SUBQ	$72, SP
0x0017 00023	MOVQ	BP, 64(SP)
0x001c 00028	LEAQ	64(SP), BP
0x0021 00033	FUNCDATA	$0, gclocals·a36216b97439c93dafebe03e7f0808b5(SB)
0x0021 00033	FUNCDATA	$1, gclocals·ff19ed39bdde8a01a800918ac3ef0ec7(SB)
0x0021 00033	XORL	AX, AX
0x0023 00035	JMP	40
0x0025 00037	INCQ	AX
0x0028 00040	MOVQ	"".b+80(SP), CX
0x002d 00045	MOVQ	240(CX), DX
0x0034 00052	CMPQ	AX, DX
0x0037 00055	JGE	189
0x003d 00061	XORPS	X0, X0
0x0040 00064	MOVUPS	X0, ""..autotmp_3+32(SP)
0x0045 00069	MOVUPS	X0, ""..autotmp_3+48(SP)
0x004a 00074	MOVQ	AX, ""..autotmp_3+48(SP)
0x004f 00079	MOVQ	$5, ""..autotmp_3+56(SP)
0x0058 00088	MOVL	runtime.writeBarrier(SB), DX
0x005e 00094	TESTL	DX, DX
0x0060 00096	JNE	131
0x0062 00098	LEAQ	"".sink(SB), DI
0x0069 00105	LEAQ	""..autotmp_3+32(SP), SI
0x006e 00110	DUFFCOPY	$868
0x0081 00129	JMP	37
0x0083 00131	MOVQ	AX, "".i+24(SP)
0x0088 00136	LEAQ	type."".Scalar(SB), AX
0x008f 00143	MOVQ	AX, (SP)
0x0093 00147	LEAQ	"".sink(SB), AX
0x009a 00154	MOVQ	AX, 8(SP)
0x009f 00159	LEAQ	""..autotmp_3+32(SP), CX
0x00a4 00164	MOVQ	CX, 16(SP)
0x00a9 00169	PCDATA	$0, $0
0x00a9 00169	CALL	runtime.typedmemmove(SB)
0x00ae 00174	MOVQ	"".i+24(SP), AX
0x00b3 00179	MOVQ	"".b+80(SP), CX
0x00b8 00184	JMP	37
0x00bd 00189	MOVQ	64(SP), BP
0x00c2 00194	ADDQ	$72, SP
0x00c6 00198	RET
0x00c7 00199	NOP
0x00c7 00199	PCDATA	$0, $-1
0x00c7 00199	CALL	runtime.morestack_noctxt(SB)
0x00cc 00204	JMP	0

The assembly for both BenchmarkA and BenchmarkB (without incomparable):

0x0000 00000	TEXT	"".BenchmarkA(SB), $8-8
0x0000 00000	MOVQ	(TLS), CX
0x0009 00009	CMPQ	SP, 16(CX)
0x000d 00013	JLS	134
0x000f 00015	SUBQ	$8, SP
0x0013 00019	MOVQ	BP, (SP)
0x0017 00023	LEAQ	(SP), BP
0x001b 00027	FUNCDATA	$0, gclocals·a36216b97439c93dafebe03e7f0808b5(SB)
0x001b 00027	FUNCDATA	$1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x001b 00027	MOVQ	"".b+16(SP), CX
0x0020 00032	XORL	AX, AX
0x0022 00034	JMP	39
0x0024 00036	INCQ	AX
0x0027 00039	MOVQ	240(CX), DX
0x002e 00046	CMPQ	AX, DX
0x0031 00049	JGE	125
0x0033 00051	MOVQ	AX, "".sink+16(SB)
0x003a 00058	MOVQ	$5, "".sink+24(SB)
0x0045 00069	MOVL	runtime.writeBarrier(SB), DX
0x004b 00075	TESTL	DX, DX
0x004d 00077	JNE	91
0x004f 00079	XORPS	X0, X0
0x0052 00082	MOVUPS	X0, "".sink(SB)
0x0059 00089	JMP	36
0x005b 00091	LEAQ	"".sink(SB), DI
0x0062 00098	MOVQ	AX, DX
0x0065 00101	XORL	AX, AX
0x0067 00103	CALL	runtime.gcWriteBarrier(SB)
0x006c 00108	LEAQ	"".sink+8(SB), DI
0x0073 00115	CALL	runtime.gcWriteBarrier(SB)
0x0078 00120	MOVQ	DX, AX
0x007b 00123	JMP	36
0x007d 00125	MOVQ	(SP), BP
0x0081 00129	ADDQ	$8, SP
0x0085 00133	RET
0x0086 00134	NOP
0x0086 00134	PCDATA	$0, $-1
0x0086 00134	CALL	runtime.morestack_noctxt(SB)
0x008b 00139	JMP	0

The presence of the incomparable field seems to trigger the use of runtime.duffcopy, which is slow on short copies.

[ALTree]

I don't think this is specifically caused by the zero-length field. It's just that structures with more than 4 fields are not ssa-able. If you change scalar to

type Scalar struct {
	XX   int
	A, B *int
	C, D uint64
}

you'll get the same slowdown.

The general tracking bug for this is: #24416.

[dsnet]

Indeed, a duplicate of #24416, which explicitly states the 4 field limit.