[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - allow insertps intrinsic to be used in constexp (llvm#165513)

ahmednoursphinx · web-flow · commit cf85cf452368 · 2025-10-30T19:07:00.000Z
Resolves llvm#165161
diff --git a/clang/include/clang/Basic/BuiltinsX86.td b/clang/include/clang/Basic/BuiltinsX86.td
@@ -328,7 +328,6 @@ let Features = "ssse3", Attributes = [NoThrow, Const, Constexpr, RequiredVectorW
 }
 
 let Features = "sse4.1", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
-  def insertps128 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<4, float>, _Constant char)">;
   def roundps : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Constant int)">;
   def roundss : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<4, float>, _Constant int)">;
   def roundsd : X86Builtin<"_Vector<2, double>(_Vector<2, double>, _Vector<2, double>, _Constant int)">;
@@ -342,6 +341,8 @@ let Features = "sse4.1", Attributes = [NoThrow, Const, RequiredVectorWidth<128>]
 
 let Features = "sse4.1",
     Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in {
+  def insertps128 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, "
+                               "_Vector<4, float>, _Constant char)">;
   def ptestz128
       : X86Builtin<"int(_Vector<2, long long int>, _Vector<2, long long int>)">;
   def ptestc128
diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
@@ -3411,7 +3411,7 @@ static bool interp__builtin_x86_byteshift(
 
 static bool interp__builtin_ia32_shuffle_generic(
     InterpState &S, CodePtr OpPC, const CallExpr *Call,
-    llvm::function_ref<std::pair<unsigned, unsigned>(unsigned, unsigned)>
+    llvm::function_ref<std::pair<unsigned, int>(unsigned, unsigned)>
         GetSourceIndex) {
 
   assert(Call->getNumArgs() == 3);
@@ -3428,8 +3428,19 @@ static bool interp__builtin_ia32_shuffle_generic(
 
   for (unsigned DstIdx = 0; DstIdx != NumElems; ++DstIdx) {
     auto [SrcVecIdx, SrcIdx] = GetSourceIndex(DstIdx, ShuffleMask);
-    const Pointer &Src = (SrcVecIdx == 0) ? A : B;
-    TYPE_SWITCH(ElemT, { Dst.elem<T>(DstIdx) = Src.elem<T>(SrcIdx); });
+
+    if (SrcIdx < 0) {
+      // Zero out this element
+      if (ElemT == PT_Float) {
+        Dst.elem<Floating>(DstIdx) = Floating(
+            S.getASTContext().getFloatTypeSemantics(VecT->getElementType()));
+      } else {
+        INT_TYPE_SWITCH_NO_BOOL(ElemT, { Dst.elem<T>(DstIdx) = T::from(0); });
+      }
+    } else {
+      const Pointer &Src = (SrcVecIdx == 0) ? A : B;
+      TYPE_SWITCH(ElemT, { Dst.elem<T>(DstIdx) = Src.elem<T>(SrcIdx); });
+    }
   }
   Dst.initializeAllElements();
 
@@ -4382,7 +4393,8 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
           unsigned SrcIdx = ElemInLane >= NumSelectableElems ? 1 : 0;
           unsigned BitIndex = (DstIdx * BitsPerElem) % MaskBits;
           unsigned Index = (ShuffleMask >> BitIndex) & IndexMask;
-          return std::pair<unsigned, unsigned>{SrcIdx, LaneOffset + Index};
+          return std::pair<unsigned, int>{SrcIdx,
+                                          static_cast<int>(LaneOffset + Index)};
         });
   case X86::BI__builtin_ia32_shufpd:
   case X86::BI__builtin_ia32_shufpd256:
@@ -4400,7 +4412,27 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
           unsigned SrcIdx = ElemInLane >= NumSelectableElems ? 1 : 0;
           unsigned BitIndex = (DstIdx * BitsPerElem) % MaskBits;
           unsigned Index = (ShuffleMask >> BitIndex) & IndexMask;
-          return std::pair<unsigned, unsigned>{SrcIdx, LaneOffset + Index};
+          return std::pair<unsigned, int>{SrcIdx,
+                                          static_cast<int>(LaneOffset + Index)};
+        });
+  case X86::BI__builtin_ia32_insertps128:
+    return interp__builtin_ia32_shuffle_generic(
+        S, OpPC, Call, [](unsigned DstIdx, unsigned Mask) {
+          // Bits [3:0]: zero mask - if bit is set, zero this element
+          if ((Mask & (1 << DstIdx)) != 0) {
+            return std::pair<unsigned, int>{0, -1};
+          }
+          // Bits [7:6]: select element from source vector Y (0-3)
+          // Bits [5:4]: select destination position (0-3)
+          unsigned SrcElem = (Mask >> 6) & 0x3;
+          unsigned DstElem = (Mask >> 4) & 0x3;
+          if (DstIdx == DstElem) {
+            // Insert element from source vector (B) at this position
+            return std::pair<unsigned, int>{1, static_cast<int>(SrcElem)};
+          } else {
+            // Copy from destination vector (A)
+            return std::pair<unsigned, int>{0, static_cast<int>(DstIdx)};
+          }
         });
   case X86::BI__builtin_ia32_pshufb128:
   case X86::BI__builtin_ia32_pshufb256:
diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp
@@ -11621,7 +11621,7 @@ static bool evalPackBuiltin(const CallExpr *E, EvalInfo &Info, APValue &Result,
 
 static bool evalShuffleGeneric(
     EvalInfo &Info, const CallExpr *Call, APValue &Out,
-    llvm::function_ref<std::pair<unsigned, unsigned>(unsigned, unsigned)>
+    llvm::function_ref<std::pair<unsigned, int>(unsigned, unsigned)>
         GetSourceIndex) {
 
   const auto *VT = Call->getType()->getAs<VectorType>();
@@ -11644,8 +11644,16 @@ static bool evalShuffleGeneric(
 
   for (unsigned DstIdx = 0; DstIdx != NumElts; ++DstIdx) {
     auto [SrcVecIdx, SrcIdx] = GetSourceIndex(DstIdx, ShuffleMask);
-    const APValue &Src = (SrcVecIdx == 0) ? A : B;
-    ResultElements.push_back(Src.getVectorElt(SrcIdx));
+
+    if (SrcIdx < 0) {
+      // Zero out this element
+      QualType ElemTy = VT->getElementType();
+      ResultElements.push_back(
+          APValue(APFloat::getZero(Info.Ctx.getFloatTypeSemantics(ElemTy))));
+    } else {
+      const APValue &Src = (SrcVecIdx == 0) ? A : B;
+      ResultElements.push_back(Src.getVectorElt(SrcIdx));
+    }
   }
 
   Out = APValue(ResultElements.data(), ResultElements.size());
@@ -12438,7 +12446,7 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) {
     if (!evalShuffleGeneric(
             Info, E, R,
             [](unsigned DstIdx,
-               unsigned ShuffleMask) -> std::pair<unsigned, unsigned> {
+               unsigned ShuffleMask) -> std::pair<unsigned, int> {
               constexpr unsigned LaneBits = 128u;
               unsigned NumElemPerLane = LaneBits / 32;
               unsigned NumSelectableElems = NumElemPerLane / 2;
@@ -12451,7 +12459,7 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) {
               unsigned BitIndex = (DstIdx * BitsPerElem) % MaskBits;
               unsigned SrcIdx = (ElemInLane < NumSelectableElems) ? 0 : 1;
               unsigned Index = (ShuffleMask >> BitIndex) & IndexMask;
-              return {SrcIdx, LaneOffset + Index};
+              return {SrcIdx, static_cast<int>(LaneOffset + Index)};
             }))
       return false;
     return Success(R, E);
@@ -12463,7 +12471,7 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) {
     if (!evalShuffleGeneric(
             Info, E, R,
             [](unsigned DstIdx,
-               unsigned ShuffleMask) -> std::pair<unsigned, unsigned> {
+               unsigned ShuffleMask) -> std::pair<unsigned, int> {
               constexpr unsigned LaneBits = 128u;
               unsigned NumElemPerLane = LaneBits / 64;
               unsigned NumSelectableElems = NumElemPerLane / 2;
@@ -12476,7 +12484,31 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) {
               unsigned BitIndex = (DstIdx * BitsPerElem) % MaskBits;
               unsigned SrcIdx = (ElemInLane < NumSelectableElems) ? 0 : 1;
               unsigned Index = (ShuffleMask >> BitIndex) & IndexMask;
-              return {SrcIdx, LaneOffset + Index};
+              return {SrcIdx, static_cast<int>(LaneOffset + Index)};
+            }))
+      return false;
+    return Success(R, E);
+  }
+  case X86::BI__builtin_ia32_insertps128: {
+    APValue R;
+    if (!evalShuffleGeneric(
+            Info, E, R,
+            [](unsigned DstIdx, unsigned Mask) -> std::pair<unsigned, int> {
+              // Bits [3:0]: zero mask - if bit is set, zero this element
+              if ((Mask & (1 << DstIdx)) != 0) {
+                return {0, -1};
+              }
+              // Bits [7:6]: select element from source vector Y (0-3)
+              // Bits [5:4]: select destination position (0-3)
+              unsigned SrcElem = (Mask >> 6) & 0x3;
+              unsigned DstElem = (Mask >> 4) & 0x3;
+              if (DstIdx == DstElem) {
+                // Insert element from source vector (B) at this position
+                return {1, static_cast<int>(SrcElem)};
+              } else {
+                // Copy from destination vector (A)
+                return {0, static_cast<int>(DstIdx)};
+              }
             }))
       return false;
     return Success(R, E);
diff --git a/clang/test/CodeGen/X86/sse41-builtins.c b/clang/test/CodeGen/X86/sse41-builtins.c
@@ -307,6 +307,16 @@ __m128 test_mm_insert_ps(__m128 x, __m128 y) {
   return _mm_insert_ps(x, y, 4);
 }
 
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x10), 1.0f, 10.0f, 3.0f, 4.0f))); // Insert Y[0] into X[1]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x00), 10.0f, 2.0f, 3.0f, 4.0f))); // Insert Y[0] into X[0]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x20), 1.0f, 2.0f, 10.0f, 4.0f))); // Insert Y[0] into X[2]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x30), 1.0f, 2.0f, 3.0f, 10.0f))); // Insert Y[0] into X[3]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x80), 30.0f, 2.0f, 3.0f, 4.0f))); // Insert Y[2] into X[0]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x01), 0.0f, 2.0f, 3.0f, 4.0f))); // Insert Y[0] into X[0], zero X[0]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x0A), 10.0f, 0.0f, 3.0f, 0.0f))); // Insert Y[0] into X[0], zero X[1] and X[3]
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0x0F), 0.0f, 0.0f, 0.0f, 0.0f))); // Insert Y[0] into X[0], zero all
+TEST_CONSTEXPR((match_m128(_mm_insert_ps(((__m128)(__v4sf){1.0f, 2.0f, 3.0f, 4.0f}), ((__m128)(__v4sf){10.0f, 20.0f, 30.0f, 40.0f}), 0xCF), 0.0f, 0.0f, 0.0f, 0.0f))); // Insert Y[3] into X[0], zero all
+
 __m128i test_mm_max_epi8(__m128i x, __m128i y) {
   // CHECK-LABEL: test_mm_max_epi8
   // CHECK: call <16 x i8> @llvm.smax.v16i8(<16 x i8> %{{.*}}, <16 x i8> %{{.*}})
