Merge pull request #75 from howjmay/undefined

feat: Add _mm_undefined_[pd|ps|si128]
pattonkan · Jan 27, 2024 · 036a6e6 · 036a6e6
2 parents 3216dc6 + b73b9cf
commit 036a6e6
Show file tree

Hide file tree

Showing 2 changed files with 60 additions and 21 deletions.
diff --git a/sse2rvv.h b/sse2rvv.h
@@ -3476,11 +3476,50 @@ FORCE_INLINE int _mm_ucomineq_ss(__m128 a, __m128 b) {
   return _mm_comineq_ss(a, b);
 }
 
-// FORCE_INLINE _mm_undefined_pd (void) {}
+FORCE_INLINE __m128d _mm_undefined_pd(void) {
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuninitialized"
+#endif
+  __m128d a;
+#if defined(_MSC_VER)
+  a = _mm_setzero_pd();
+#endif
+  return a;
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+}
 
-// FORCE_INLINE __m128 _mm_undefined_ps (void) {}
+FORCE_INLINE __m128 _mm_undefined_ps(void) {
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuninitialized"
+#endif
+  __m128 a;
+#if defined(_MSC_VER)
+  a = _mm_setzero_ps();
+#endif
+  return a;
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+}
 
-// FORCE_INLINE __m128i _mm_undefined_si128 (void) {}
+FORCE_INLINE __m128i _mm_undefined_si128(void) {
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuninitialized"
+#endif
+  __m128i a;
+#if defined(_MSC_VER)
+  a = _mm_setzero_si128();
+#endif
+  return a;
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+}
 
 FORCE_INLINE __m128i _mm_unpackhi_epi16(__m128i a, __m128i b) {
   vuint16m2_t _a = __riscv_vlmul_ext_v_u16m1_u16m2(vreinterpretq_m128i_u16(a));

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -3649,13 +3649,13 @@ result_t test_mm_ucomineq_ss(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
 }
 
 result_t test_mm_undefined_ps(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   __m128 a = _mm_undefined_ps();
-  //   a = _mm_xor_ps(a, a);
-  //   return validate_float(a, 0, 0, 0, 0);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  __m128 a = _mm_undefined_ps();
+  a = _mm_xor_ps(a, a);
+  return validate_float(a, 0, 0, 0, 0);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_unpackhi_ps(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
@@ -8048,23 +8048,23 @@ result_t test_mm_ucomineq_sd(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
 }
 
 result_t test_mm_undefined_pd(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   __m128d a = _mm_undefined_pd();
-  //   a = _mm_xor_pd(a, a);
-  //   return validate_double(a, 0, 0);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  __m128d a = _mm_undefined_pd();
+  a = _mm_xor_pd(a, a);
+  return validate_double(a, 0, 0);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_undefined_si128(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   __m128i a = _mm_undefined_si128();
-  //   a = _mm_xor_si128(a, a);
-  //   return validate_int64(a, 0, 0);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  __m128i a = _mm_undefined_si128();
+  a = _mm_xor_si128(a, a);
+  return validate_int64(a, 0, 0);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_unpackhi_epi16(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {