Reduce generated code (#185)

* Reduce generated code

Use a define (with hacks)

.github/workflows/go.yml

@@ -87,4 +87,5 @@ jobs:
       run: go test -no-avx512 -no-avx2 -no-ssse3 -short -race .
 
     - name: Test Microarch v4
+      shell: bash {0}
       run: go run testlevel.go 4;if [ $? -eq 0 ]; then GOAMD64=v4 go test -no-avx512 ./...; else true; fi

_gen/cleanup.go

@@ -21,7 +21,19 @@ func main() {
 		if err != nil {
 			log.Fatalln(err)
 		}
-		data = bytes.Replace(data, []byte("\t// #"), []byte("#"), -1)
+		data = bytes.ReplaceAll(data, []byte("\t// #"), []byte("#"))
+		data = bytes.ReplaceAll(data, []byte("\t// @"), []byte(""))
+		data = bytes.ReplaceAll(data, []byte("VPTERNLOGQ"), []byte("XOR3WAY("))
+		split := bytes.Split(data, []byte("\n"))
+		// Add closing ')'
+		want := []byte("\tXOR3WAY(")
+		for i, b := range split {
+			if bytes.Contains(b, want) {
+				b = []byte(string(b) + ")")
+				split[i] = b
+			}
+		}
+		data = bytes.Join(split, []byte("\n"))
 		data, err = asmfmt.Format(bytes.NewBuffer(data))
 		if err != nil {
 			log.Fatalln(err)

_gen/gen.go

@@ -38,6 +38,17 @@ func main() {
 	Constraint(buildtags.Not("nogen").ToConstraint())
 	Constraint(buildtags.Term("gc").ToConstraint())
 
+	TEXT("_dummy_", 0, "func()")
+	Comment("#ifdef GOAMD64_v4")
+	Comment("#define XOR3WAY(ignore, a, b, dst)\\")
+	Comment("@\tVPTERNLOGD $0x96, a, b, dst")
+	Comment("#else")
+	Comment("#define XOR3WAY(ignore, a, b, dst)\\")
+	Comment("@\tVPXOR a, dst, dst\\")
+	Comment("@\tVPXOR b, dst, dst")
+	Comment("#endif")
+	RET()
+
 	const perLoopBits = 6
 	const perLoop = 1 << perLoopBits
 
@@ -123,13 +134,8 @@ func galMulSlicesAvx2Xor(matrix []byte, in, out [][]byte, start, stop int) int {
 
 // VPXOR3way will 3-way xor a and b and dst.
 func VPXOR3way(a, b, dst reg.VecVirtual) {
-	Comment("#ifdef GOAMD64_v4")
-	// AVX512F and AVX512VL required
-	VPTERNLOGD(U8(0x96), a, b, dst)
-	Comment("#else")
-	VPXOR(a, dst, dst) // dst = a^dst
-	VPXOR(b, dst, dst) // dst = (a^dst)^b
-	Comment("#endif")
+	// VPTERNLOGQ is replaced by XOR3WAY - we just use an equivalent operation
+	VPTERNLOGQ(U8(0), a, b, dst)
 }
 
 func genMulAvx2(name string, inputs int, outputs int, xor bool) {

galois_gen_amd64.go

@@ -5,6 +5,8 @@
 
 package reedsolomon
 
+func _dummy_()
+
 // mulAvxTwo_1x1 takes 1 inputs and produces 1 outputs.
 // The output is initialized to 0.
 //go:noescape