x86 sse4.1 mm_testz_si128: fix backwards short circuit logic

Co-authored-by: Florent Hivert <[email protected]>
simd-everywhere · Oct 27, 2023 · f132275 · f132275
1 parent d5d6d10
commit f132275
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 5 deletions.
diff --git a/simde/x86/sse4.1.h b/simde/x86/sse4.1.h
@@ -2340,14 +2340,19 @@ simde_mm_testz_si128 (simde__m128i a, simde__m128i b) {
     #elif defined(SIMDE_WASM_SIMD128_NATIVE)
       v128_t m = wasm_v128_and(a_.wasm_v128, b_.wasm_v128);
       return (wasm_i64x2_extract_lane(m, 0) | wasm_i64x2_extract_lane(m, 1)) == 0;
+    #elif defined(SIMDE_HAVE_INT128_)
+      if ((a_.u128[0] & b_.u128[0]) == 0) {
+        return 1;
+      }
+      return 0;
     #else
       for (size_t i = 0 ; i < (sizeof(a_.u64) / sizeof(a_.u64[0])) ; i++) {
-        if ((a_.u64[i] & b_.u64[i]) == 0)
-          return 1;
+        if ((a_.u64[i] & b_.u64[i]) > 0)
+          return 0;
       }
     #endif
 
-    return 0;
+    return 1;
   #endif
 }
 #if defined(SIMDE_X86_SSE4_1_ENABLE_NATIVE_ALIASES)

diff --git a/test/x86/sse4.1.c b/test/x86/sse4.1.c
@@ -3552,7 +3552,7 @@ test_simde_mm_testz_si128(SIMDE_MUNIT_TEST_ARGS) {
     simde__m128i a;
     simde__m128i b;
     int r;
-  } test_vec[8] = {
+  } test_vec[] = {
     { simde_x_mm_set_epu32(UINT32_C(3990889899), UINT32_C(4230789057), UINT32_C(3808461042), UINT32_C(3371427410)),
       simde_x_mm_set_epu32(UINT32_C(    331792), UINT32_C(  63127558), UINT32_C( 483990789), UINT32_C(    530828)),
       1 },
@@ -3576,7 +3576,10 @@ test_simde_mm_testz_si128(SIMDE_MUNIT_TEST_ARGS) {
       1 },
     { simde_x_mm_set_epu32(UINT32_C(3782150825), UINT32_C(3325635017), UINT32_C(1617333560), UINT32_C(3634437083)),
       simde_x_mm_set_epu32(UINT32_C( 167825730), UINT32_C( 294047748), UINT32_C( 126906945), UINT32_C(  17884164)),
-      1 }
+      1 },
+    { simde_x_mm_set_epu8(UINT8_C(0), UINT8_C(1), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0)),
+      simde_x_mm_set_epu8(UINT8_C(0), UINT8_C(1), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0), UINT8_C(0)),
+      0 }
   };
 
   for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {