From: Dave Cheney <dave@cheney.net>
Date: Mon, 5 Aug 2013 21:51:37 +0000 (+1000)
Subject: runtime: tune append crossover on amd64 and 386
X-Git-Tag: go1.2rc2~780
X-Git-Url: http://www.git.cypherpunks.su/?a=commitdiff_plain;h=8ce8adbe7a194cba2819b533b603d92df40fe799;p=gostls13.git

runtime: tune append crossover on amd64 and 386

Fixes #4963.

Sets the append crossover to 0 on intel platforms.

Results for linux/amd64 Core i5 SNB

benchmark                     old ns/op    new ns/op    delta
BenchmarkAppend                     102          104   +1.96%
BenchmarkAppend1Byte                 10           11   +0.92%
BenchmarkAppend4Bytes                15           11  -28.10%
BenchmarkAppend7Bytes                17           12  -32.58%
BenchmarkAppend8Bytes                18           12  -36.17%
BenchmarkAppend15Bytes               24           11  -55.02%
BenchmarkAppend16Bytes               25           11  -56.03%
BenchmarkAppend32Bytes               11           12   +4.31%
BenchmarkAppendStr1Byte               8            9  +13.99%
BenchmarkAppendStr4Bytes             11            9  -17.52%
BenchmarkAppendStr8Bytes             14            9  -35.70%
BenchmarkAppendStr16Bytes            21            9  -55.19%
BenchmarkAppendStr32Bytes            10           10   -5.66%
BenchmarkAppendSpecialCase           49           52   +7.96%

Results for linux/386 Atom(TM) CPU 330 @ 1.60GHz

benchmark                     old ns/op    new ns/op    delta
BenchmarkAppend                     219          218   -0.46%
BenchmarkAppend1Byte                 75           72   -3.44%
BenchmarkAppend4Bytes                92           73  -19.87%
BenchmarkAppend7Bytes               108           74  -31.20%
BenchmarkAppend8Bytes               116           74  -35.95%
BenchmarkAppend15Bytes              162           77  -52.22%
BenchmarkAppend16Bytes              169           77  -54.20%
BenchmarkAppend32Bytes               88           86   -2.38%
BenchmarkAppendStr1Byte              57           59   +3.32%
BenchmarkAppendStr4Bytes             72           59  -17.40%
BenchmarkAppendStr8Bytes             92           60  -34.70%
BenchmarkAppendStr16Bytes           141           63  -54.89%
BenchmarkAppendStr32Bytes            75           73   -2.64%
BenchmarkAppendSpecialCase          270          270   +0.00%

R=golang-dev, r
CC=golang-dev
https://golang.org/cl/12440044
---

diff --git a/src/pkg/runtime/append_test.go b/src/pkg/runtime/append_test.go
index 36390181e8..8a4e4a383d 100644
--- a/src/pkg/runtime/append_test.go
+++ b/src/pkg/runtime/append_test.go
@@ -38,10 +38,18 @@ func BenchmarkAppend4Bytes(b *testing.B) {
 	benchmarkAppendBytes(b, 4)
 }
 
+func BenchmarkAppend7Bytes(b *testing.B) {
+	benchmarkAppendBytes(b, 7)
+}
+
 func BenchmarkAppend8Bytes(b *testing.B) {
 	benchmarkAppendBytes(b, 8)
 }
 
+func BenchmarkAppend15Bytes(b *testing.B) {
+	benchmarkAppendBytes(b, 15)
+}
+
 func BenchmarkAppend16Bytes(b *testing.B) {
 	benchmarkAppendBytes(b, 16)
 }
diff --git a/src/pkg/runtime/arch_386.h b/src/pkg/runtime/arch_386.h
index 7e74d8f923..6c8550d61d 100644
--- a/src/pkg/runtime/arch_386.h
+++ b/src/pkg/runtime/arch_386.h
@@ -6,6 +6,6 @@ enum {
 	thechar = '8',
 	BigEndian = 0,
 	CacheLineSize = 64,
-	appendCrossover = 16,
+	appendCrossover = 0,
 	PCQuantum = 1
 };
diff --git a/src/pkg/runtime/arch_amd64.h b/src/pkg/runtime/arch_amd64.h
index 21144111b2..761183a9d3 100644
--- a/src/pkg/runtime/arch_amd64.h
+++ b/src/pkg/runtime/arch_amd64.h
@@ -6,6 +6,6 @@ enum {
 	thechar = '6',
 	BigEndian = 0,
 	CacheLineSize = 64,
-	appendCrossover = 16,
+	appendCrossover = 0,
 	PCQuantum = 1
 };
diff --git a/src/pkg/runtime/slice.c b/src/pkg/runtime/slice.c
index 3adbe51837..58086b61c7 100644
--- a/src/pkg/runtime/slice.c
+++ b/src/pkg/runtime/slice.c
@@ -95,7 +95,7 @@ runtime·appendslice(SliceType *t, Slice x, Slice y, Slice ret)
 	p = ret.array+ret.len*w;
 	q = y.array;
 	w *= y.len;
-	if(w <= appendCrossover) {
+	if(appendCrossover > 0 && w <= appendCrossover) {
 		if(p <= q || w <= p-q) // No overlap.
 			while(w-- > 0)
 				*p++ = *q++;
@@ -148,7 +148,7 @@ runtime·appendstr(SliceType *t, Slice x, String y, Slice ret)
 	w = y.len;
 	p = ret.array+ret.len;
 	q = y.str;
-	if(w <= appendCrossover) {
+	if(appendCrossover > 0 && w <= appendCrossover) {
 		while(w-- > 0)
 			*p++ = *q++;
 	} else {