Draft neon support

Author: jeschkies

pkg/search/asm_neon.go

@@ -0,0 +1,42 @@
+//go:build ignore
+
+package main
+
+import (
+	. "github.com/mmcloughlin/avo/build"
+	. "github.com/mmcloughlin/avo/operand"
+	"github.com/mmcloughlin/avo/reg"
+)
+
+// main generates assembly code for NEON on ARM aarch64.
+func main() {
+	TEXT("findInChunk", NOSPLIT, "func(needle []byte, haystack []byte) int64")
+	Doc("findInChunk is only generated for testing.")
+	//hptr := Load(Param("haystack").Base(), GP64())
+	needleLen := Load(Param("needle").Len(), GP64()); DECQ(needleLen)
+	needle := Load(Param("needle").Base(), GP64())
+	_, _ = inlineSplat(needle, needleLen)
+
+	//offset := inlineFindInChunk("test", f, l, hptr, needle, needleLen)
+	offset := GP64()
+
+	Store(offset, ReturnIndex(0))
+	RET()
+}
+
+// inlineSplat fills one 128bit register with repeated first neelde char and
+// another with repeated last needle char.
+func inlineSplat(needle0, needleLen reg.Register) (reg.VecVirtual, reg.VecVirtual) {
+	Comment("create vector filled with first and last character")
+	f := YMM()
+	l := YMM()
+
+	VDUP
+
+	needle1 := GP64(); MOVQ(needle0, needle1);
+	ADDQ(needleLen, needle1)
+	VPBROADCASTB(Mem{Base: needle0}, f)
+	VPBROADCASTB(Mem{Base: needle1}, l)
+
+	return f, l
+}

pkg/search/bytes_neon_arm64.go

@@ -4,6 +4,9 @@ import (
 	"bytes"
 )
 
+// findInChunk is only generated for testing.
+func findInChunk(needle []byte, haystack []byte) int64
+
 // indexNeon returns the first position the needle is in the haystack.
 func indexNeon (haystack []byte, needle []byte) int64 {
 	// TODO: port to ARM64 Neon. This file will be generated.

pkg/search/bytes_test.go

@@ -9,10 +9,10 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/require"
-	"golang.org/x/sys/cpu"
 )
 
 //go:generate go run asm_avx2.go -out bytes_avx2_amd64.s -stubs bytes_avx2_amd64.go
+//go:generate go run asm_neon.go -out bytes_neon_arm64.s -stubs bytes_neon_arm64.go
 
 func TestSimpleIndex(t *testing.T) {
 	for _, tt := range []struct {
@@ -129,10 +129,6 @@ func TestMask(t *testing.T) {
 	}
 }
 
-func TestArm(t * testing.T) {
-	require.True(t, cpu.ARM64.HasASIMD)
-}
-
 func BenchmarkIndexSmall(b *testing.B) {
 	needle := []byte("goldner7875")
 	haystack, err := loadHaystack("small.log")