[X86] Use __builtin_ia32_paddq and __builtin_ia32_psubq to implement a couple...

[X86] Use __builtin_ia32_paddq and __builtin_ia32_psubq to implement a couple intrinsics that were supposed to operate on MMX registers. Otherwise we end up operating on GPRs. Throw in a test for _mm_mul_su32 while I was there.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@252711 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Headers/emmintrin.h

@@ -647,7 +647,7 @@ _mm_add_epi32(__m128i __a, __m128i __b)
 static __inline__ __m64 __DEFAULT_FN_ATTRS
 _mm_add_si64(__m64 __a, __m64 __b)
 {
-  return __a + __b;
+  return (__m64)__builtin_ia32_paddq(__a, __b);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
@@ -779,7 +779,7 @@ _mm_sub_epi32(__m128i __a, __m128i __b)
 static __inline__ __m64 __DEFAULT_FN_ATTRS
 _mm_sub_si64(__m64 __a, __m64 __b)
 {
-  return __a - __b;
+  return (__m64)__builtin_ia32_psubq(__a, __b);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS

test/CodeGen/sse-builtins.c

@@ -495,3 +495,21 @@ __m128i test_mm_undefined_si128() {
   // CHECK: ret <2 x i64> undef
   return _mm_undefined_si128();
 }
+
+__m64 test_mm_add_si64(__m64 __a, __m64 __b) {
+  // CHECK-LABEL: @test_mm_add_si64
+  // CHECK @llvm.x86.mmx.padd.q(x86_mmx %{{.*}}, x86_mmx %{{.*}})
+  return _mm_add_si64(__a, __b);
+}
+
+__m64 test_mm_sub_si64(__m64 __a, __m64 __b) {
+  // CHECK-LABEL: @test_mm_sub_si64
+  // CHECK @llvm.x86.mmx.psub.q(x86_mmx %{{.*}}, x86_mmx %{{.*}})
+  return _mm_sub_si64(__a, __b);
+}
+
+__m64 test_mm_mul_su32(__m64 __a, __m64 __b) {
+  // CHECK-LABEL: @test_mm_mul_su32
+  // CHECK @llvm.x86.mmx.pmulu.dq(x86_mmx %{{.*}}, x86_mmx %{{.*}})
+  return _mm_mul_su32(__a, __b);
+}