runtime: port performance-critical functions to regabi

This CL ports a few performance-critical runtime assembly functions to
use register arguments directly. While using the faster ABI is nice,
the real win here is that we avoid ABI wrappers: since these are
"builtin" functions in the compiler, it can generate calls to them
without knowing that their native implementation is ABI0. Hence, it
generates ABIInternal calls that go through ABI wrappers. By porting
them to use ABIInternal natively, we avoid the overhead of the ABI
wrapper.

This significantly improves performance on several benchmarks,
comparing regabiwrappers before and after this change:

name                                old time/op  new time/op  delta
BiogoIgor                            15.7s ± 2%   15.7s ± 2%    ~     (p=0.617 n=25+25)
BiogoKrishna                         18.5s ± 5%   17.7s ± 2%  -4.61%  (p=0.000 n=25+25)
BleveIndexBatch100                   5.91s ± 3%   5.82s ± 3%  -1.60%  (p=0.000 n=25+25)
BleveQuery                           6.76s ± 0%   6.60s ± 1%  -2.31%  (p=0.000 n=22+25)
CompileTemplate                      248ms ± 5%   245ms ± 1%    ~     (p=0.643 n=25+20)
CompileUnicode                      94.4ms ± 3%  93.9ms ± 2%    ~     (p=0.152 n=24+23)
CompileGoTypes                       1.60s ± 2%   1.59s ± 2%    ~     (p=0.059 n=24+24)
CompileCompiler                      104ms ± 3%   103ms ± 1%    ~     (p=0.056 n=25+22)
CompileSSA                           10.9s ± 1%   10.9s ± 1%    ~     (p=0.052 n=25+25)
CompileFlate                         156ms ± 8%   152ms ± 1%  -2.49%  (p=0.008 n=25+21)
CompileGoParser                      248ms ± 1%   249ms ± 2%    ~     (p=0.058 n=21+20)
CompileReflect                       595ms ± 3%   601ms ± 4%    ~     (p=0.182 n=25+25)
CompileTar                           211ms ± 2%   211ms ± 1%    ~     (p=0.663 n=23+23)
CompileXML                           282ms ± 2%   284ms ± 5%    ~     (p=0.456 n=21+23)
CompileStdCmd                        13.6s ± 2%   13.5s ± 2%    ~     (p=0.112 n=25+24)
FoglemanFauxGLRenderRotateBoat       8.69s ± 2%   8.67s ± 0%    ~     (p=0.094 n=22+25)
FoglemanPathTraceRenderGopherIter1   20.2s ± 2%   20.7s ± 3%  +2.53%  (p=0.000 n=24+24)
GopherLuaKNucleotide                 31.4s ± 1%   31.0s ± 1%  -1.28%  (p=0.000 n=25+24)
MarkdownRenderXHTML                  246ms ± 1%   244ms ± 1%  -0.79%  (p=0.000 n=20+21)
Tile38WithinCircle100kmRequest       843µs ± 4%   818µs ± 4%  -2.93%  (p=0.000 n=25+25)
Tile38IntersectsCircle100kmRequest  1.06ms ± 5%  1.05ms ± 3%  -1.19%  (p=0.021 n=24+25)
Tile38KNearestLimit100Request       1.01ms ± 1%  1.01ms ± 2%    ~     (p=0.335 n=22+25)
[Geo mean]                           596ms        592ms       -0.71%

(https://perf.golang.org/search?q=upload:20210411.5)

It also significantly reduces the performance penalty of enabling
regabiwrappers, though it doesn't yet fully close the gap on all
benchmarks:

name                                old time/op  new time/op  delta
BiogoIgor                            15.7s ± 1%   15.7s ± 2%    ~     (p=0.366 n=24+25)
BiogoKrishna                         17.7s ± 2%   17.7s ± 2%    ~     (p=0.315 n=23+25)
BleveIndexBatch100                   5.86s ± 4%   5.82s ± 3%    ~     (p=0.137 n=24+25)
BleveQuery                           6.55s ± 0%   6.60s ± 1%  +0.83%  (p=0.000 n=24+25)
CompileTemplate                      244ms ± 1%   245ms ± 1%    ~     (p=0.208 n=21+20)
CompileUnicode                      94.0ms ± 4%  93.9ms ± 2%    ~     (p=0.666 n=24+23)
CompileGoTypes                       1.60s ± 2%   1.59s ± 2%    ~     (p=0.154 n=25+24)
CompileCompiler                      103ms ± 1%   103ms ± 1%    ~     (p=0.905 n=24+22)
CompileSSA                           10.9s ± 2%   10.9s ± 1%    ~     (p=0.803 n=25+25)
CompileFlate                         153ms ± 1%   152ms ± 1%    ~     (p=0.182 n=23+21)
CompileGoParser                      250ms ± 2%   249ms ± 2%    ~     (p=0.843 n=24+20)
CompileReflect                       595ms ± 4%   601ms ± 4%    ~     (p=0.141 n=25+25)
CompileTar                           212ms ± 3%   211ms ± 1%    ~     (p=0.499 n=23+23)
CompileXML                           282ms ± 1%   284ms ± 5%    ~     (p=0.129 n=20+23)
CompileStdCmd                        13.5s ± 2%   13.5s ± 2%    ~     (p=0.480 n=24+24)
FoglemanFauxGLRenderRotateBoat       8.66s ± 1%   8.67s ± 0%    ~     (p=0.325 n=25+25)
FoglemanPathTraceRenderGopherIter1   20.6s ± 3%   20.7s ± 3%    ~     (p=0.137 n=25+24)
GopherLuaKNucleotide                 30.5s ± 2%   31.0s ± 1%  +1.68%  (p=0.000 n=23+24)
MarkdownRenderXHTML                  243ms ± 1%   244ms ± 1%  +0.51%  (p=0.000 n=23+21)
Tile38WithinCircle100kmRequest       801µs ± 2%   818µs ± 4%  +2.11%  (p=0.000 n=25+25)
Tile38IntersectsCircle100kmRequest  1.01ms ± 2%  1.05ms ± 3%  +4.34%  (p=0.000 n=24+25)
Tile38KNearestLimit100Request       1.00ms ± 1%  1.01ms ± 2%  +0.81%  (p=0.008 n=21+25)
[Geo mean]                           589ms        592ms       +0.50%

(https://perf.golang.org/search?q=upload:20210411.6)

Change-Id: I8f77f010b0abc658064df569a27a9c7a7b1c7bf9
Reviewed-on: https://go-review.googlesource.com/c/go/+/308931
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

1 files changed, 57 insertions, 48 deletions

diff --git a/src/runtime/memclr_amd64.s b/src/runtime/memclr_amd64.s
index 37fe9745b1..b4bc9988ec 100644
--- a/src/runtime/memclr_amd64.s
+++ b/src/runtime/memclr_amd64.s
@@ -12,9 +12,16 @@
 // See memclrNoHeapPointers Go doc for important implementation constraints.
 
 // func memclrNoHeapPointers(ptr unsafe.Pointer, n uintptr)
-TEXT runtime·memclrNoHeapPointers(SB), NOSPLIT, $0-16
+// ABIInternal for performance.
+TEXT runtime·memclrNoHeapPointers<ABIInternal>(SB), NOSPLIT, $0-16
+#ifdef GOEXPERIMENT_regabiargs
+	// AX = ptr
+	// BX = n
+	MOVQ	AX, DI	// DI = ptr
+#else
 	MOVQ	ptr+0(FP), DI
 	MOVQ	n+8(FP), BX
+#endif
 	XORQ	AX, AX
 
 	// MOVOU seems always faster than REP STOSQ.
@@ -31,7 +38,9 @@ tail:
 	JE	_8
 	CMPQ	BX, $16
 	JBE	_9through16
-	PXOR	X0, X0
+#ifndef GOEXPERIMENT_regabig
+	PXOR	X15, X15
+#endif
 	CMPQ	BX, $32
 	JBE	_17through32
 	CMPQ	BX, $64
@@ -45,22 +54,22 @@ tail:
 	// TODO: for really big clears, use MOVNTDQ, even without AVX2.
 
 loop:
-	MOVOU	X0, 0(DI)
-	MOVOU	X0, 16(DI)
-	MOVOU	X0, 32(DI)
-	MOVOU	X0, 48(DI)
-	MOVOU	X0, 64(DI)
-	MOVOU	X0, 80(DI)
-	MOVOU	X0, 96(DI)
-	MOVOU	X0, 112(DI)
-	MOVOU	X0, 128(DI)
-	MOVOU	X0, 144(DI)
-	MOVOU	X0, 160(DI)
-	MOVOU	X0, 176(DI)
-	MOVOU	X0, 192(DI)
-	MOVOU	X0, 208(DI)
-	MOVOU	X0, 224(DI)
-	MOVOU	X0, 240(DI)
+	MOVOU	X15, 0(DI)
+	MOVOU	X15, 16(DI)
+	MOVOU	X15, 32(DI)
+	MOVOU	X15, 48(DI)
+	MOVOU	X15, 64(DI)
+	MOVOU	X15, 80(DI)
+	MOVOU	X15, 96(DI)
+	MOVOU	X15, 112(DI)
+	MOVOU	X15, 128(DI)
+	MOVOU	X15, 144(DI)
+	MOVOU	X15, 160(DI)
+	MOVOU	X15, 176(DI)
+	MOVOU	X15, 192(DI)
+	MOVOU	X15, 208(DI)
+	MOVOU	X15, 224(DI)
+	MOVOU	X15, 240(DI)
 	SUBQ	$256, BX
 	ADDQ	$256, DI
 	CMPQ	BX, $256
@@ -141,40 +150,40 @@ _9through16:
 	MOVQ	AX, -8(DI)(BX*1)
 	RET
 _17through32:
-	MOVOU	X0, (DI)
-	MOVOU	X0, -16(DI)(BX*1)
+	MOVOU	X15, (DI)
+	MOVOU	X15, -16(DI)(BX*1)
 	RET
 _33through64:
-	MOVOU	X0, (DI)
-	MOVOU	X0, 16(DI)
-	MOVOU	X0, -32(DI)(BX*1)
-	MOVOU	X0, -16(DI)(BX*1)
+	MOVOU	X15, (DI)
+	MOVOU	X15, 16(DI)
+	MOVOU	X15, -32(DI)(BX*1)
+	MOVOU	X15, -16(DI)(BX*1)
 	RET
 _65through128:
-	MOVOU	X0, (DI)
-	MOVOU	X0, 16(DI)
-	MOVOU	X0, 32(DI)
-	MOVOU	X0, 48(DI)
-	MOVOU	X0, -64(DI)(BX*1)
-	MOVOU	X0, -48(DI)(BX*1)
-	MOVOU	X0, -32(DI)(BX*1)
-	MOVOU	X0, -16(DI)(BX*1)
+	MOVOU	X15, (DI)
+	MOVOU	X15, 16(DI)
+	MOVOU	X15, 32(DI)
+	MOVOU	X15, 48(DI)
+	MOVOU	X15, -64(DI)(BX*1)
+	MOVOU	X15, -48(DI)(BX*1)
+	MOVOU	X15, -32(DI)(BX*1)
+	MOVOU	X15, -16(DI)(BX*1)
 	RET
 _129through256:
-	MOVOU	X0, (DI)
-	MOVOU	X0, 16(DI)
-	MOVOU	X0, 32(DI)
-	MOVOU	X0, 48(DI)
-	MOVOU	X0, 64(DI)
-	MOVOU	X0, 80(DI)
-	MOVOU	X0, 96(DI)
-	MOVOU	X0, 112(DI)
-	MOVOU	X0, -128(DI)(BX*1)
-	MOVOU	X0, -112(DI)(BX*1)
-	MOVOU	X0, -96(DI)(BX*1)
-	MOVOU	X0, -80(DI)(BX*1)
-	MOVOU	X0, -64(DI)(BX*1)
-	MOVOU	X0, -48(DI)(BX*1)
-	MOVOU	X0, -32(DI)(BX*1)
-	MOVOU	X0, -16(DI)(BX*1)
+	MOVOU	X15, (DI)
+	MOVOU	X15, 16(DI)
+	MOVOU	X15, 32(DI)
+	MOVOU	X15, 48(DI)
+	MOVOU	X15, 64(DI)
+	MOVOU	X15, 80(DI)
+	MOVOU	X15, 96(DI)
+	MOVOU	X15, 112(DI)
+	MOVOU	X15, -128(DI)(BX*1)
+	MOVOU	X15, -112(DI)(BX*1)
+	MOVOU	X15, -96(DI)(BX*1)
+	MOVOU	X15, -80(DI)(BX*1)
+	MOVOU	X15, -64(DI)(BX*1)
+	MOVOU	X15, -48(DI)(BX*1)
+	MOVOU	X15, -32(DI)(BX*1)
+	MOVOU	X15, -16(DI)(BX*1)
 	RET