support multi-labels on AMD64

.github/workflows/test.yml

@@ -20,7 +20,10 @@ jobs:
       - name: Install GOAT
         run: go install .
       - name: Run GOAT
-        run: goat src/avx_mul_to.c -O3 -mavx -mfma
+        run: |
+          for file in src/*.c; do
+            goat $file -O3 -mavx -mfma
+          done
         working-directory: example
       - name: Run tests
         run: go test -v ./...

example/avx_dot_test.go

@@ -0,0 +1,24 @@
+package example
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func AVXDot(a, b []float32) float32 {
+	if len(a) != len(b) {
+		panic("floats: slice lengths do not match")
+	}
+	var c float32
+	avx_dot(unsafe.Pointer(&a[0]), unsafe.Pointer(&b[0]), unsafe.Pointer(uintptr(len(a))), unsafe.Pointer(&c))
+	return c
+}
+
+func TestDot(t *testing.T) {
+	a := []float32{1, 2, 3, 4}
+	b := []float32{5, 6, 7, 8}
+	c := AVXDot(a, b)
+	if c != 70 {
+		t.Errorf("AVXDot(%v, %v) = %v, want %v", a, b, c, 70)
+	}
+}

example/src/avx_dot.c

@@ -0,0 +1,39 @@
+#include <immintrin.h>
+#include <stdint.h>
+
+void avx_dot(float *a, float *b, int64_t n, float *ret)
+{
+    int epoch = n / 8;
+    int remain = n % 8;
+    __m256 s;
+    if (epoch > 0)
+    {
+        __m256 v1 = _mm256_loadu_ps(a);
+        __m256 v2 = _mm256_loadu_ps(b);
+        s = _mm256_mul_ps(v1, v2);
+        a += 8;
+        b += 8;
+    }
+    for (int i = 1; i < epoch; i++)
+    {
+        __m256 v1 = _mm256_loadu_ps(a);
+        __m256 v2 = _mm256_loadu_ps(b);
+        s = _mm256_add_ps(_mm256_mul_ps(v1, v2), s);
+        a += 8;
+        b += 8;
+    }
+    __m128 s7_6_5_4 = _mm256_extractf128_ps(s, 1);
+    __m128 s3_2_1_0 = _mm256_castps256_ps128(s);
+    __m128 s37_26_15_04 = _mm_add_ps(s7_6_5_4, s3_2_1_0);
+    __m128 sxx_15_04 = s37_26_15_04;
+    __m128 sxx_37_26 = _mm_movehl_ps(s37_26_15_04, s37_26_15_04);
+    const __m128 sxx_1357_0246 = _mm_add_ps(sxx_15_04, sxx_37_26);
+    const __m128 sxxx_0246 = sxx_1357_0246;
+    const __m128 sxxx_1357 = _mm_shuffle_ps(sxx_1357_0246, sxx_1357_0246, 0x1);
+    __m128 sxxx_01234567 = _mm_add_ss(sxxx_0246, sxxx_1357);
+    *ret = _mm_cvtss_f32(sxxx_01234567);
+    for (int i = 0; i < remain; i++)
+    {
+        *ret += a[i] * b[i];
+    }
+}

parser_amd64.go

@@ -39,15 +39,15 @@ var (
 )
 
 type Line struct {
-	Label    string
+	Labels   []string
 	Assembly string
 	Binary   []string
 }
 
 func (line *Line) String() string {
 	var builder strings.Builder
-	if len(line.Label) > 0 {
-		builder.WriteString(line.Label)
+	for _, label := range line.Labels {
+		builder.WriteString(label)
 		builder.WriteString(":\n")
 	}
 	builder.WriteString("\t")
@@ -115,15 +115,21 @@ func parseAssembly(path string) (map[string][]Line, error) {
 		} else if labelLine.MatchString(line) {
 			labelName = strings.Split(line, ":")[0]
 			labelName = labelName[1:]
-			functions[functionName] = append(functions[functionName], Line{Label: labelName})
+			lines := functions[functionName]
+			if len(lines) > 0 && lines[len(lines)-1].Assembly == "" {
+				// If the last line is a label, append the label to the last line.
+				lines[len(lines)-1].Labels = append(lines[len(lines)-1].Labels, labelName)
+			} else {
+				functions[functionName] = append(functions[functionName], Line{Labels: []string{labelName}})
+			}
 		} else if codeLine.MatchString(line) {
 			asm := sanitizeAsm(line)
 			if labelName == "" {
 				functions[functionName] = append(functions[functionName], Line{Assembly: asm})
 			} else {
 				lines := functions[functionName]
 				if len(lines) == 0 {
-					functions[functionName] = append(functions[functionName], Line{Label: labelName})
+					functions[functionName] = append(functions[functionName], Line{Labels: []string{labelName}})
 					lines = functions[functionName]
 				}