[X86][Clang] Support constexpr evaluation of cvtpd2ps intrinsics

clang/include/clang/Basic/BuiltinsX86.td

@@ -166,14 +166,20 @@ let Features = "sse2", Attributes = [NoThrow] in {
   def movnti : X86Builtin<"void(int *, int)">;
 }
 
+let Features = "sse2", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in {
+  def cvtpd2ps : X86Builtin<"_Vector<4, float>(_Vector<2, double>)">;
+  def cvtsd2ss : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<2, double>)">;
+}
+let Features = "avx512f", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in {
+  def cvtsd2ss_round_mask : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<2, double>, _Vector<4, float>, unsigned char, _Constant int)">;
+}
+
 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 cvtpd2ps : X86Builtin<"_Vector<4, float>(_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>)">;
-  def cvtsd2ss : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<2, double>)">;
   def cvtps2dq : X86Builtin<"_Vector<4, int>(_Vector<4, float>)">;
   def cvttps2dq : X86Builtin<"_Vector<4, int>(_Vector<4, float>)">;
 }
@@ -462,19 +468,21 @@ let Features = "avx", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWid
   def vpermilvarps256 : X86Builtin<"_Vector<8, float>(_Vector<8, float>, _Vector<8, int>)">;
 }
 
+let Features = "avx", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<256>] in {
+  def cvtpd2ps256 : X86Builtin<"_Vector<4, float>(_Vector<4, double>)">;
+}
+
 let Features = "avx", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
   def dpps256 : X86Builtin<"_Vector<8, float>(_Vector<8, float>, _Vector<8, float>, _Constant char)">;
   def cmppd256 : X86Builtin<"_Vector<4, double>(_Vector<4, double>, _Vector<4, double>, _Constant char)">;
   def cmpps256 : X86Builtin<"_Vector<8, float>(_Vector<8, float>, _Vector<8, float>, _Constant char)">;
-  def cvtpd2ps256 : X86Builtin<"_Vector<4, float>(_Vector<4, double>)">;
   def cvtps2dq256 : X86Builtin<"_Vector<8, int>(_Vector<8, float>)">;
   def cvttpd2dq256 : X86Builtin<"_Vector<4, int>(_Vector<4, double>)">;
   def cvtpd2dq256 : X86Builtin<"_Vector<4, int>(_Vector<4, double>)">;
   def cvttps2dq256 : X86Builtin<"_Vector<8, int>(_Vector<8, float>)">;
   def vperm2f128_pd256 : X86Builtin<"_Vector<4, double>(_Vector<4, double>, _Vector<4, double>, _Constant int)">;
   def vperm2f128_ps256 : X86Builtin<"_Vector<8, float>(_Vector<8, float>, _Vector<8, float>, _Constant int)">;
   def vperm2f128_si256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Constant int)">;
-
   foreach Op = ["max", "min"] in {
     def Op#pd256 : X86Builtin<"_Vector<4, double>(_Vector<4, double>, _Vector<4, double>)">;
     def Op#ps256 : X86Builtin<"_Vector<8, float>(_Vector<8, float>, _Vector<8, float>)">;
@@ -1004,6 +1012,10 @@ let Features = "avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<128
   def cmppd128_mask : X86Builtin<"unsigned char(_Vector<2, double>, _Vector<2, double>, _Constant int, unsigned char)">;
 }
 
+let Features = "avx512f", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<512>] in {
+  def cvtpd2ps512_mask : X86Builtin<"_Vector<8, float>(_Vector<8, double>, _Vector<8, float>, unsigned char, _Constant int)">;
+}
+
 let Features = "avx512f", Attributes = [NoThrow, Const, RequiredVectorWidth<512>] in {
   def rndscaleps_mask : X86Builtin<"_Vector<16, float>(_Vector<16, float>, _Constant int, _Vector<16, float>, unsigned short, _Constant int)">;
   def rndscalepd_mask : X86Builtin<"_Vector<8, double>(_Vector<8, double>, _Constant int, _Vector<8, double>, unsigned char, _Constant int)">;
@@ -1017,7 +1029,6 @@ let Features = "avx512f", Attributes = [NoThrow, Const, RequiredVectorWidth<512>
   def maxpd512 : X86Builtin<"_Vector<8, double>(_Vector<8, double>, _Vector<8, double>, _Constant int)">;
   def cvtdq2ps512_mask : X86Builtin<"_Vector<16, float>(_Vector<16, int>, _Vector<16, float>, unsigned short, _Constant int)">;
   def cvtudq2ps512_mask : X86Builtin<"_Vector<16, float>(_Vector<16, int>, _Vector<16, float>, unsigned short, _Constant int)">;
-  def cvtpd2ps512_mask : X86Builtin<"_Vector<8, float>(_Vector<8, double>, _Vector<8, float>, unsigned char, _Constant int)">;
   def vcvtps2ph512_mask : X86Builtin<"_Vector<16, short>(_Vector<16, float>, _Constant int, _Vector<16, short>, unsigned short)">;
   def vcvtph2ps512_mask : X86Builtin<"_Vector<16, float>(_Vector<16, short>, _Vector<16, float>, unsigned short, _Constant int)">;
 }
@@ -1452,9 +1463,12 @@ let Features = "avx512vl", Attributes = [NoThrow, RequiredVectorWidth<256>] in {
   def compressstoresi256_mask : X86Builtin<"void(_Vector<8, int *>, _Vector<8, int>, unsigned char)">;
 }
 
+let Features = "avx512vl", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in {
+  def cvtpd2ps_mask : X86Builtin<"_Vector<4, float>(_Vector<2, double>, _Vector<4, float>, unsigned char)">;
+}
+
 let Features = "avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
   def cvtpd2dq128_mask : X86Builtin<"_Vector<4, int>(_Vector<2, double>, _Vector<4, int>, unsigned char)">;
-  def cvtpd2ps_mask : X86Builtin<"_Vector<4, float>(_Vector<2, double>, _Vector<4, float>, unsigned char)">;
   def cvtpd2udq128_mask : X86Builtin<"_Vector<4, int>(_Vector<2, double>, _Vector<4, int>, unsigned char)">;
 }
 
@@ -3288,7 +3302,6 @@ let Features = "avx512bw,avx512vl",
 }
 
 let Features = "avx512f", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
-  def cvtsd2ss_round_mask : X86Builtin<"_Vector<4, float>(_Vector<4, float>, _Vector<2, double>, _Vector<4, float>, unsigned char, _Constant int)">;
   def cvtsi2ss32 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, int, _Constant int)">;
   def cvtss2sd_round_mask : X86Builtin<"_Vector<2, double>(_Vector<2, double>, _Vector<4, float>, _Vector<2, double>, unsigned char, _Constant int)">;
   def cvtusi2ss32 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, unsigned int, _Constant int)">;

clang/lib/AST/ByteCode/InterpBuiltin.cpp

@@ -167,6 +167,38 @@ static llvm::APSInt convertBoolVectorToInt(const Pointer &Val) {
   return Result;
 }
 
+// Strict double -> float conversion used for X86 PD2PS/cvtsd2ss intrinsics.
+// Reject NaN/Inf/Subnormal inputs and any lossy/inexact conversions.
+static bool convertDoubleToFloatStrict(APFloat Src, Floating &Dst,
+                                       InterpState &S, const Expr *DiagExpr) {
+  if (Src.isInfinity()) {
+    if (S.diagnosing())
+      S.CCEDiag(DiagExpr, diag::note_constexpr_float_arithmetic) << 0;
+    return false;
+  }
+  if (Src.isNaN()) {
+    if (S.diagnosing())
+      S.CCEDiag(DiagExpr, diag::note_constexpr_float_arithmetic) << 1;
+    return false;
+  }
+  APFloat Val = Src;
+  bool LosesInfo = false;
+  APFloat::opStatus Status = Val.convert(
+      APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, &LosesInfo);
+  if (LosesInfo || Val.isDenormal()) {
+    if (S.diagnosing())
+      S.CCEDiag(DiagExpr, diag::note_constexpr_float_arithmetic_strict);
+    return false;
+  }
+  if (Status != APFloat::opOK) {
+    if (S.diagnosing())
+      S.CCEDiag(DiagExpr, diag::note_invalid_subexpr_in_const_expr);
+    return false;
+  }
+  Dst.copy(Val);
+  return true;
+}
+
 static bool interp__builtin_is_constant_evaluated(InterpState &S, CodePtr OpPC,
                                                   const InterpFrame *Frame,
                                                   const CallExpr *Call) {
@@ -3359,6 +3391,146 @@ static bool interp__builtin_ia32_cvt_vec2mask(InterpState &S, CodePtr OpPC,
   pushInteger(S, RetMask, Call->getType());
   return true;
 }
+static bool interp__builtin_ia32_cvtsd2ss(InterpState &S, CodePtr OpPC,
+                                          const CallExpr *Call) {
+  assert(Call->getNumArgs() == 2);
+
+  const Pointer &B = S.Stk.pop<Pointer>();
+  const Pointer &A = S.Stk.pop<Pointer>();
+  if (!CheckLoad(S, OpPC, A) || !CheckLoad(S, OpPC, B))
+    return false;
+
+  const auto *DstVTy = Call->getType()->castAs<VectorType>();
+  unsigned NumElems = DstVTy->getNumElements();
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  // Copy all elements except lane 0 (overwritten below) from A to Dst.
+  for (unsigned I = 1; I != NumElems; ++I)
+    Dst.elem<Floating>(I) = A.elem<Floating>(I);
+
+  // Convert element 0 from double to float.
+  Floating Conv = S.allocFloat(
+      S.getASTContext().getFloatTypeSemantics(DstVTy->getElementType()));
+  APFloat SrcVal = B.elem<Floating>(0).getAPFloat();
+  if (!convertDoubleToFloatStrict(SrcVal, Conv, S, Call))
+    return false;
+  Dst.elem<Floating>(0) = Conv;
+
+  Dst.initializeAllElements();
+  return true;
+}
+
+static bool interp__builtin_ia32_cvtsd2ss_round_mask(InterpState &S,
+                                                     CodePtr OpPC,
+                                                     const CallExpr *Call) {
+  assert(Call->getNumArgs() == 5);
+
+  APSInt Rounding = popToAPSInt(S, Call->getArg(4));
+  APSInt MaskInt = popToAPSInt(S, Call->getArg(3));
+  const Pointer &Src = S.Stk.pop<Pointer>();
+  const Pointer &B = S.Stk.pop<Pointer>();
+  const Pointer &A = S.Stk.pop<Pointer>();
+  if (!CheckLoad(S, OpPC, A) || !CheckLoad(S, OpPC, B) ||
+      !CheckLoad(S, OpPC, Src))
+    return false;
+
+  const auto *DstVTy = Call->getType()->castAs<VectorType>();
+  unsigned NumElems = DstVTy->getNumElements();
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  // Copy all elements except lane 0 (overwritten below) from A to Dst.
+  for (unsigned I = 1; I != NumElems; ++I)
+    Dst.elem<Floating>(I) = A.elem<Floating>(I);
+
+  // If mask bit 0 is set, convert element 0 from double to float; otherwise use
+  // Src.
+  if (MaskInt.getZExtValue() & 0x1) {
+
+    assert(S.getASTContext().FloatTy == DstVTy->getElementType() &&
+           "cvtsd2ss requires float element type in destination vector");
+
+    Floating Conv = S.allocFloat(
+        S.getASTContext().getFloatTypeSemantics(DstVTy->getElementType()));
+    APFloat SrcVal = B.elem<Floating>(0).getAPFloat();
+    if (!convertDoubleToFloatStrict(SrcVal, Conv, S, Call))
+      return false;
+    Dst.elem<Floating>(0) = Conv;
+  } else {
+    Dst.elem<Floating>(0) = Src.elem<Floating>(0);
+  }
+
+  Dst.initializeAllElements();
+  return true;
+}
+
+static bool interp__builtin_ia32_cvtpd2ps(InterpState &S, CodePtr OpPC,
+                                          const CallExpr *Call,
+                                          unsigned BuiltinID) {
+  bool IsMasked = (BuiltinID == X86::BI__builtin_ia32_cvtpd2ps_mask ||
+                   BuiltinID == X86::BI__builtin_ia32_cvtpd2ps512_mask);
+  bool HasRounding = (BuiltinID == X86::BI__builtin_ia32_cvtpd2ps512_mask);
+
+  APSInt MaskVal;
+  Pointer PassThrough;
+  Pointer Src;
+  APSInt Rounding;
+
+  if (IsMasked) {
+    // Pop in reverse order.
+    if (HasRounding) {
+      Rounding = popToAPSInt(S, Call->getArg(3));
+      MaskVal = popToAPSInt(S, Call->getArg(2));
+      PassThrough = S.Stk.pop<Pointer>();
+      Src = S.Stk.pop<Pointer>();
+    } else {
+      MaskVal = popToAPSInt(S, Call->getArg(2));
+      PassThrough = S.Stk.pop<Pointer>();
+      Src = S.Stk.pop<Pointer>();
+    }
+
+    if (!CheckLoad(S, OpPC, PassThrough))
+      return false;
+  } else {
+    // Pop source only.
+    Src = S.Stk.pop<Pointer>();
+  }
+
+  if (!CheckLoad(S, OpPC, Src))
+    return false;
+
+  const auto *RetVTy = Call->getType()->castAs<VectorType>();
+  unsigned RetElems = RetVTy->getNumElements();
+  unsigned SrcElems = Src.getNumElems();
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  // Initialize destination with passthrough or zeros.
+  for (unsigned I = 0; I != RetElems; ++I)
+    if (IsMasked)
+      Dst.elem<Floating>(I) = PassThrough.elem<Floating>(I);
+    else
+      Dst.elem<Floating>(I) = Floating(APFloat(0.0f));
+
+  assert(S.getASTContext().FloatTy == RetVTy->getElementType() &&
+         "cvtpd2ps requires float element type in return vector");
+
+  // Convert double to float for enabled elements (only process source elements
+  // that exist).
+  for (unsigned I = 0; I != SrcElems; ++I) {
+    if (IsMasked && !MaskVal[I])
+      continue;
+
+    APFloat SrcVal = Src.elem<Floating>(I).getAPFloat();
+
+    Floating Conv = S.allocFloat(
+        S.getASTContext().getFloatTypeSemantics(RetVTy->getElementType()));
+    if (!convertDoubleToFloatStrict(SrcVal, Conv, S, Call))
+      return false;
+    Dst.elem<Floating>(I) = Conv;
+  }
+
+  Dst.initializeAllElements();
+  return true;
+}
 
 static bool interp__builtin_ia32_shuffle_generic(
     InterpState &S, CodePtr OpPC, const CallExpr *Call,
@@ -5169,6 +5341,18 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
   case X86::BI__builtin_ia32_cvtq2mask512:
     return interp__builtin_ia32_cvt_vec2mask(S, OpPC, Call, BuiltinID);
 
+  case X86::BI__builtin_ia32_cvtsd2ss:
+    return interp__builtin_ia32_cvtsd2ss(S, OpPC, Call);
+
+  case X86::BI__builtin_ia32_cvtsd2ss_round_mask:
+    return interp__builtin_ia32_cvtsd2ss_round_mask(S, OpPC, Call);
+
+  case X86::BI__builtin_ia32_cvtpd2ps:
+  case X86::BI__builtin_ia32_cvtpd2ps256:
+  case X86::BI__builtin_ia32_cvtpd2ps_mask:
+  case X86::BI__builtin_ia32_cvtpd2ps512_mask:
+    return interp__builtin_ia32_cvtpd2ps(S, OpPC, Call, BuiltinID);
+
   case X86::BI__builtin_ia32_cmpb128_mask:
   case X86::BI__builtin_ia32_cmpw128_mask:
   case X86::BI__builtin_ia32_cmpd128_mask: