[clang][x86] Fix the return type of the cvtpd2dq builtin

[jmgorius]

The CVTPD2DQ instruction converts packed 64-bit floating-point values to packed 32-bit signed integer values. This patch fixes the return type of the corresponding builtin, which previously returned a vector of two 64-bit signed integers. The new behavior is in line with the return type of the CVTTPD2DQ builtin.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-backend-x86

Author: Jean-Michel Gorius (jmgorius)

Changes

The CVTPD2DQ instruction converts packed 64-bit floating-point values to packed 32-bit signed integer values. This patch fixes the return type of the corresponding builtin, which previously returned a vector of two 64-bit signed integers. The new behavior is in line with the return type of the CVTTPD2DQ builtin.

---

Full diff: https://github.com/llvm/llvm-project/pull/189254.diff

3 Files Affected:

• (modified) clang/include/clang/Basic/BuiltinsX86.td (+1-1)
• (modified) clang/lib/Headers/emmintrin.h (+1-1)
• (modified) clang/test/CodeGen/builtins-x86.c (+1-1)

diff --git a/clang/include/clang/Basic/BuiltinsX86.td b/clang/include/clang/Basic/BuiltinsX86.td
index f47532a63de04..0cab8c77d465d 100644
--- a/clang/include/clang/Basic/BuiltinsX86.td
+++ b/clang/include/clang/Basic/BuiltinsX86.td
@@ -180,7 +180,7 @@ let Features = "avx512f", Attributes = [NoThrow, Const, Constexpr, RequiredVecto
 
 let Features = "sse2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
   def psadbw128 : X86Builtin<"_Vector<2, long long int>(_Vector<16, char>, _Vector<16, char>)">;
-  def cvtpd2dq : X86Builtin<"_Vector<2, long long int>(_Vector<2, double>)">;
+  def cvtpd2dq : X86Builtin<"_Vector<4, int>(_Vector<2, double>)">;
   def cvttpd2dq : X86Builtin<"_Vector<4, int>(_Vector<2, double>)">;
   def cvtsd2si : X86Builtin<"int(_Vector<2, double>)">;
   def cvttsd2si : X86Builtin<"int(_Vector<2, double>)">;
diff --git a/clang/lib/Headers/emmintrin.h b/clang/lib/Headers/emmintrin.h
index bbf366133c68a..43c93263f015a 100644
--- a/clang/lib/Headers/emmintrin.h
+++ b/clang/lib/Headers/emmintrin.h
@@ -1343,7 +1343,7 @@ _mm_cvtepi32_pd(__m128i __a) {
 /// \returns A 128-bit vector of [4 x i32] whose lower 64 bits contain the
 ///    converted values. The upper 64 bits are set to zero.
 static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvtpd_epi32(__m128d __a) {
-  return __builtin_ia32_cvtpd2dq((__v2df)__a);
+  return (__m128i)__builtin_ia32_cvtpd2dq((__v2df)__a);
 }
 
 /// Converts the low-order element of a 128-bit vector of [2 x double]
diff --git a/clang/test/CodeGen/builtins-x86.c b/clang/test/CodeGen/builtins-x86.c
index a1e63d59e88e1..0f66d8c4e3120 100644
--- a/clang/test/CodeGen/builtins-x86.c
+++ b/clang/test/CodeGen/builtins-x86.c
@@ -290,7 +290,7 @@ void f0(void) {
   (void) __builtin_ia32_movnti64(tmp_LLip, tmp_LLi);
 #endif
   tmp_V2LLi = __builtin_ia32_psadbw128(tmp_V16c, tmp_V16c);
-  tmp_V2LLi = __builtin_ia32_cvtpd2dq(tmp_V2d);
+  tmp_V4i = __builtin_ia32_cvtpd2dq(tmp_V2d);
   tmp_V4f = __builtin_ia32_cvtpd2ps(tmp_V2d);
   tmp_V4i = __builtin_ia32_cvttpd2dq(tmp_V2d);
   tmp_i = __builtin_ia32_cvtsd2si(tmp_V2d);

[RKSimon]

LGTM - will correct match int_x86_sse2_cvtpd2dq as well

[jmgorius]

Thanks @RKSimon !

Would it make sense to open more PRs if I come across inconsistent types in the x86 builtins? I found a few that, e.g., should semantically operate on unsigned values but take vectors of signed values instead (packuswb128 comes to mind).

[RKSimon]

Sure - keep each patch focused on a particular instruction type though and expect some issues with constexpr evaluation which often uses the signedness from the types in the APSInt values