fix: mismatch parsed and output length on amd64 (#2)

.github/workflows/test.yml

@@ -2,9 +2,9 @@ name: test
 
 on:
   push:
-    branches: [ main ]
+    branches: [main]
   pull_request:
-    branches: [ main ]
+    branches: [main]
 
 jobs:
   example:
@@ -43,3 +43,8 @@ jobs:
           for file in tests/*.c; do
             goat $file -O3 -mavx -mfma
           done
+      - name: Run amd64-only tests
+        run: |
+          for file in tests/amd64/*.c; do
+            goat $file -O3 -mavx -mfma
+          done

parser_amd64.go

@@ -111,17 +111,22 @@ func parseAssembly(path string) (map[string][]Line, error) {
 		} else if nameLine.MatchString(line) {
 			functionName = strings.Split(line, ":")[0]
 			functions[functionName] = make([]Line, 0)
+			labelName = ""
 		} else if labelLine.MatchString(line) {
 			labelName = strings.Split(line, ":")[0]
 			labelName = labelName[1:]
 			functions[functionName] = append(functions[functionName], Line{Label: labelName})
 		} else if codeLine.MatchString(line) {
-			asm := strings.Split(line, "#")[0]
-			asm = strings.TrimSpace(asm)
+			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})
+					lines = functions[functionName]
+				}
+
 				lines[len(lines)-1].Assembly = asm
 				labelName = ""
 			}
@@ -134,6 +139,14 @@ func parseAssembly(path string) (map[string][]Line, error) {
 	return functions, nil
 }
 
+func sanitizeAsm(asm string) string {
+	asm = strings.TrimSpace(asm)
+	asm = strings.Split(asm, "//")[0]
+	asm = strings.TrimSpace(asm)
+
+	return asm
+}
+
 func parseObjectDump(dump string, functions map[string][]Line) error {
 	var (
 		functionName string
@@ -161,6 +174,12 @@ func parseObjectDump(dump string, functions map[string][]Line) error {
 				}
 				binary = append(binary, s)
 			}
+
+			assembly = sanitizeAsm(assembly)
+			if strings.Contains(assembly, "nop") {
+				continue
+			}
+
 			if assembly == "" {
 				return fmt.Errorf("try to increase --insn-width of objdump")
 			} else if strings.HasPrefix(assembly, "nop") ||

parser_arm64.go

@@ -99,7 +99,9 @@ func parseAssembly(path string) (map[string][]Line, error) {
 				functions[functionName] = append(functions[functionName], Line{Assembly: asm})
 			} else {
 				lines := functions[functionName]
-				lines[len(lines)-1].Assembly = asm
+				if len(lines) > 0 {
+					lines[len(lines)-1].Assembly = asm
+				}
 				labelName = ""
 			}
 		}

tests/amd64/empty_line.c

@@ -0,0 +1,39 @@
+#include <immintrin.h>
+
+void l2(float *a, float *b, float *res, long *len)
+{
+    int n = *len;
+    float sum = 0;
+
+    __m256 acc[4];
+    acc[0] = _mm256_setzero_ps();
+    acc[1] = _mm256_setzero_ps();
+
+    while (n)
+    {
+        __m256 a_vec0 = _mm256_loadu_ps(a);
+        __m256 b_vec0 = _mm256_loadu_ps(b);
+
+        __m256 diff0 = _mm256_sub_ps(a_vec0, b_vec0);
+
+        acc[0] = _mm256_fmadd_ps(diff0, diff0, acc[0]);
+
+        n--;
+        a++;
+        b++;
+    }
+
+    acc[0] = _mm256_add_ps(acc[1], acc[0]);
+    if (*len >= 32)
+    {
+        acc[2] = _mm256_add_ps(acc[3], acc[2]);
+        acc[0] = _mm256_add_ps(acc[2], acc[0]);
+    }
+
+    __m256 t1 = _mm256_hadd_ps(acc[0], acc[0]);
+    __m256 t2 = _mm256_hadd_ps(t1, t1);
+    __m128 t3 = _mm256_extractf128_ps(t2, 1);
+    __m128 t4 = _mm_add_ps(_mm256_castps256_ps128(t2), t3);
+    sum += _mm_cvtss_f32(t4);
+    *res = sum;
+}