[HLSL] Implementation of dot intrinsic

This change implements #70073

HLSL has a dot intrinsic defined here:
https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-dot

The intrinsic itself is defined as a HLSL_LANG LangBuiltin in BuiltinsHLSL.td.
This is used to associate all the dot product typdef defined hlsl_intrinsics.h
with a single intrinsic check in CGBuiltin.cpp & SemaChecking.cpp.

As a side note adding the dot product intrinsic to BuiltinsHLSL.td had a
significant impact on re-compile time speeds. I recommend we move the
other hlsl functions here. Further it lets us tap into the existing
target specifc code organizations that exist in Sema and CodeGen.

In IntrinsicsDirectX.td we define the llvmIR for the dot product.
A few goals were in mind for this IR. First it should operate on only
vectors. Second the return type should be the vector element type. Third
the second parameter vector should be of the same size as the first
parameter. Finally `a dot b` should be the same as `b dot a`.

In CGBuiltin.cpp hlsl has built on top of existing clang intrinsics via EmitBuiltinExpr. Dot
product though is a target specific intrinsic and so needed to establish
a pattern for Target builtins via EmitDXILBuiltinExpr.
The call chain looks like this now: EmitBuiltinExpr -> EmitTargetBuiltinExpr -> EmitTargetArchBuiltinExpr -> EmitDXILBuiltinExp

EmitDXILBuiltinExp dot product intrinsics makes a destinction
between vectors and scalars. This is because HLSL supports dot product on scalars which simplifies down to multiply.

Sema.h & SemaChecking.cpp saw the addition of CheckHLSLBuiltinFunctionCall, a target specific semantic validation that can be expanded for other hlsl specific intrinsics.

Author: farzonl

clang/include/clang/Basic/BuiltinsHLSL.td

@@ -0,0 +1,15 @@
+//===--- BuiltinsHLSL.td - HLSL Builtin function database ---------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+include "clang/Basic/BuiltinsBase.td"
+
+def HLSLDotProduct : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_dot"];
+  let Attributes = [NoThrow, Const, CustomTypeChecking];
+  let Prototype = "void(...)";
+}

clang/include/clang/Basic/CMakeLists.txt

@@ -65,6 +65,10 @@ clang_tablegen(BuiltinsBPF.inc -gen-clang-builtins
   SOURCE BuiltinsBPF.td
   TARGET ClangBuiltinsBPF)
 
+  clang_tablegen(BuiltinsHLSL.inc -gen-clang-builtins
+  SOURCE BuiltinsHLSL.td
+  TARGET ClangBuiltinsHLSL)
+
 # ARM NEON and MVE
 clang_tablegen(arm_neon.inc -gen-arm-neon-sema
   SOURCE arm_neon.td

clang/include/clang/Basic/TargetBuiltins.h

@@ -89,6 +89,16 @@ namespace clang {
   };
   }
 
+  /// HLSL builtins
+  namespace hlsl {
+  enum {
+    LastTIBuiltin = clang::Builtin::FirstTSBuiltin - 1,
+#define BUILTIN(ID, TYPE, ATTRS) BI##ID,
+#include "clang/Basic/BuiltinsHLSL.inc"
+    LastTSBuiltin
+  };
+  } // namespace hlsl
+
   /// PPC builtins
   namespace PPC {
     enum {

clang/include/clang/Sema/Sema.h

@@ -14017,6 +14017,7 @@ class Sema final {
   bool CheckPPCBuiltinFunctionCall(const TargetInfo &TI, unsigned BuiltinID,
                                    CallExpr *TheCall);
   bool CheckAMDGCNBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall);
+  bool CheckHLSLBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall);
   bool CheckRISCVLMUL(CallExpr *TheCall, unsigned ArgNum);
   bool CheckRISCVBuiltinFunctionCall(const TargetInfo &TI, unsigned BuiltinID,
                                      CallExpr *TheCall);

clang/lib/Basic/Targets/DirectX.cpp

@@ -12,11 +12,24 @@
 
 #include "DirectX.h"
 #include "Targets.h"
+#include "clang/Basic/MacroBuilder.h"
+#include "clang/Basic/TargetBuiltins.h"
 
 using namespace clang;
 using namespace clang::targets;
 
+static constexpr Builtin::Info BuiltinInfo[] = {
+#define BUILTIN(ID, TYPE, ATTRS)                                               \
+  {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES},
+#include "clang/Basic/BuiltinsHLSL.inc"
+};
+
 void DirectXTargetInfo::getTargetDefines(const LangOptions &Opts,
                                          MacroBuilder &Builder) const {
   DefineStd(Builder, "DIRECTX", Opts);
 }
+
+ArrayRef<Builtin::Info> DirectXTargetInfo::getTargetBuiltins() const {
+  return llvm::ArrayRef(BuiltinInfo,
+                        clang::hlsl::LastTSBuiltin - Builtin::FirstTSBuiltin);
+}

clang/lib/Basic/Targets/DirectX.h

@@ -73,9 +73,7 @@ class LLVM_LIBRARY_VISIBILITY DirectXTargetInfo : public TargetInfo {
     return Feature == "directx";
   }
 
-  ArrayRef<Builtin::Info> getTargetBuiltins() const override {
-    return std::nullopt;
-  }
+  ArrayRef<Builtin::Info> getTargetBuiltins() const override;
 
   std::string_view getClobbers() const override { return ""; }
 

clang/lib/CodeGen/CGBuiltin.cpp

@@ -44,6 +44,7 @@
 #include "llvm/IR/IntrinsicsAMDGPU.h"
 #include "llvm/IR/IntrinsicsARM.h"
 #include "llvm/IR/IntrinsicsBPF.h"
+#include "llvm/IR/IntrinsicsDirectX.h"
 #include "llvm/IR/IntrinsicsHexagon.h"
 #include "llvm/IR/IntrinsicsNVPTX.h"
 #include "llvm/IR/IntrinsicsPowerPC.h"
@@ -6018,6 +6019,8 @@ static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF,
   case llvm::Triple::bpfeb:
   case llvm::Triple::bpfel:
     return CGF->EmitBPFBuiltinExpr(BuiltinID, E);
+  case llvm::Triple::dxil:
+    return CGF->EmitDXILBuiltinExpr(BuiltinID, E);
   case llvm::Triple::x86:
   case llvm::Triple::x86_64:
     return CGF->EmitX86BuiltinExpr(BuiltinID, E);
@@ -17895,6 +17898,44 @@ llvm::Value *CodeGenFunction::EmitScalarOrConstFoldImmArg(unsigned ICEArguments,
   return Arg;
 }
 
+Value *CodeGenFunction::EmitDXILBuiltinExpr(unsigned BuiltinID,
+                                            const CallExpr *E) {
+  switch (BuiltinID) {
+  case hlsl::BI__builtin_hlsl_dot: {
+    Value *Op0 = EmitScalarExpr(E->getArg(0));
+    Value *Op1 = EmitScalarExpr(E->getArg(1));
+    llvm::Type *T0 = Op0->getType();
+    llvm::Type *T1 = Op1->getType();
+    if (!T0->isVectorTy() && !T1->isVectorTy()) {
+      if (T0->isFloatingPointTy()) {
+        return Builder.CreateFMul(Op0, Op1, "dx.dot");
+      }
+
+      if (T0->isIntegerTy()) {
+        return Builder.CreateMul(Op0, Op1, "dx.dot");
+      }
+      assert(
+          false &&
+          "Dot product on a scalar is only supported on integers and floats.");
+    }
+    assert(T0->isVectorTy() && T1->isVectorTy() &&
+           "Dot product of vector and scalar is not supported.");
+    assert(T0->getScalarType() == T1->getScalarType() &&
+           "Dot product of vectors need the same element types.");
+
+    auto *VecTy0 = E->getArg(0)->getType()->getAs<VectorType>();
+    auto *VecTy1 = E->getArg(1)->getType()->getAs<VectorType>();
+    assert(VecTy0->getNumElements() == VecTy1->getNumElements() &&
+           "Dot product requires vectors to be of the same size.");
+
+    return Builder.CreateIntrinsic(
+        /*ReturnType*/ T0->getScalarType(), Intrinsic::dx_dot,
+        ArrayRef<Value *>{Op0, Op1}, nullptr, "dx.dot");
+  } break;
+  }
+  return nullptr;
+}
+
 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
                                               const CallExpr *E) {
   llvm::AtomicOrdering AO = llvm::AtomicOrdering::SequentiallyConsistent;

clang/lib/CodeGen/CodeGenFunction.h

@@ -4392,6 +4392,7 @@ class CodeGenFunction : public CodeGenTypeCache {
   llvm::Value *EmitX86BuiltinExpr(unsigned BuiltinID, const CallExpr *E);
   llvm::Value *EmitPPCBuiltinExpr(unsigned BuiltinID, const CallExpr *E);
   llvm::Value *EmitAMDGPUBuiltinExpr(unsigned BuiltinID, const CallExpr *E);
+  llvm::Value *EmitDXILBuiltinExpr(unsigned BuiltinID, const CallExpr *E);
   llvm::Value *EmitScalarOrConstFoldImmArg(unsigned ICEArguments, unsigned Idx,
                                            const CallExpr *E);
   llvm::Value *EmitSystemZBuiltinExpr(unsigned BuiltinID, const CallExpr *E);

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -144,6 +144,74 @@ double3 cos(double3);
 _HLSL_BUILTIN_ALIAS(__builtin_elementwise_cos)
 double4 cos(double4);
 
+//===----------------------------------------------------------------------===//
+// dot product builtins
+//===----------------------------------------------------------------------===//
+#ifdef __HLSL_ENABLE_16_BIT
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+half dot(half, half);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+half dot(half2, half2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+half dot(half3, half3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+half dot(half4, half4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int16_t dot(int16_t, int16_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int16_t dot(int16_t2, int16_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int16_t dot(int16_t3, int16_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int16_t dot(int16_t4, int16_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint16_t dot(uint16_t, uint16_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint16_t dot(uint16_t2, uint16_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint16_t dot(uint16_t3, uint16_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint16_t dot(uint16_t4, uint16_t4);
+#endif
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+float dot(float, float);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+float dot(float2, float2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+float dot(float3, float3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+float dot(float4, float4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+double dot(double, double);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int dot(int, int);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int dot(int2, int2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int dot(int3, int3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int dot(int4, int4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint dot(uint, uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint dot(uint2, uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint dot(uint3, uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint dot(uint4, uint4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+int64_t dot(int64_t, int64_t);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot)
+uint64_t dot(uint64_t, uint64_t);
+
 //===----------------------------------------------------------------------===//
 // floor builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaChecking.cpp

@@ -2084,6 +2084,8 @@ bool Sema::CheckTSBuiltinFunctionCall(const TargetInfo &TI, unsigned BuiltinID,
     return CheckBPFBuiltinFunctionCall(BuiltinID, TheCall);
   case llvm::Triple::hexagon:
     return CheckHexagonBuiltinFunctionCall(BuiltinID, TheCall);
+  case llvm::Triple::dxil:
+    return CheckHLSLBuiltinFunctionCall(BuiltinID, TheCall);
   case llvm::Triple::mips:
   case llvm::Triple::mipsel:
   case llvm::Triple::mips64:
@@ -2120,10 +2122,11 @@ bool Sema::CheckTSBuiltinFunctionCall(const TargetInfo &TI, unsigned BuiltinID,
 // not a valid type, emit an error message and return true. Otherwise return
 // false.
 static bool checkMathBuiltinElementType(Sema &S, SourceLocation Loc,
-                                        QualType Ty) {
-  if (!Ty->getAs<VectorType>() && !ConstantMatrixType::isValidElementType(Ty)) {
+                                        QualType ArgTy, int ArgIndex) {
+  if (!ArgTy->getAs<VectorType>() &&
+      !ConstantMatrixType::isValidElementType(ArgTy)) {
     return S.Diag(Loc, diag::err_builtin_invalid_arg_type)
-           << 1 << /* vector, integer or float ty*/ 0 << Ty;
+           << ArgIndex << /* vector, integer or float ty*/ 0 << ArgTy;
   }
 
   return false;
@@ -5158,6 +5161,64 @@ bool Sema::CheckPPCMMAType(QualType Type, SourceLocation TypeLoc) {
   return false;
 }
 
+// Note: returning true in this case results in CheckBuiltinFunctionCall
+// returning an ExprError
+bool Sema::CheckHLSLBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
+  switch (BuiltinID) {
+  case hlsl::BI__builtin_hlsl_dot: {
+    if (checkArgCount(*this, TheCall, 2)) {
+      return true;
+    }
+    Expr *Arg0 = TheCall->getArg(0);
+    QualType ArgTy0 = Arg0->getType();
+
+    Expr *Arg1 = TheCall->getArg(1);
+    QualType ArgTy1 = Arg1->getType();
+
+    auto *VecTy0 = ArgTy0->getAs<VectorType>();
+    auto *VecTy1 = ArgTy1->getAs<VectorType>();
+    SourceLocation BuiltinLoc = TheCall->getBeginLoc();
+
+    // if arg0 is bool then call Diag with err_builtin_invalid_arg_type
+    if (checkMathBuiltinElementType(*this, Arg0->getBeginLoc(), ArgTy0, 1)) {
+      return true;
+    }
+
+    // if arg1 is bool then call Diag with err_builtin_invalid_arg_type
+    if (checkMathBuiltinElementType(*this, Arg1->getBeginLoc(), ArgTy1, 2)) {
+      return true;
+    }
+
+    if (VecTy0 == nullptr && VecTy1 == nullptr) {
+      if (ArgTy0 != ArgTy1) {
+        return true;
+      } else {
+        return false;
+      }
+    }
+
+    if ((VecTy0 == nullptr && VecTy1 != nullptr) ||
+        (VecTy0 != nullptr && VecTy1 == nullptr)) {
+
+      Diag(BuiltinLoc, diag::err_vec_builtin_non_vector)
+          << TheCall->getDirectCallee()
+          << SourceRange(TheCall->getArg(0)->getBeginLoc(),
+                         TheCall->getArg(1)->getEndLoc());
+      return true;
+    }
+
+    if (VecTy0->getElementType() != VecTy1->getElementType()) {
+      return true;
+    }
+    if (VecTy0->getNumElements() != VecTy1->getNumElements()) {
+      return true;
+    }
+    break;
+  }
+  }
+  return false;
+}
+
 bool Sema::CheckAMDGCNBuiltinFunctionCall(unsigned BuiltinID,
                                           CallExpr *TheCall) {
   // position of memory order and scope arguments in the builtin
@@ -19577,7 +19638,7 @@ bool Sema::PrepareBuiltinElementwiseMathOneArgCall(CallExpr *TheCall) {
   TheCall->setArg(0, A.get());
   QualType TyA = A.get()->getType();
 
-  if (checkMathBuiltinElementType(*this, A.get()->getBeginLoc(), TyA))
+  if (checkMathBuiltinElementType(*this, A.get()->getBeginLoc(), TyA, 1))
     return true;
 
   TheCall->setType(TyA);
@@ -19605,7 +19666,7 @@ bool Sema::SemaBuiltinElementwiseMath(CallExpr *TheCall) {
                 diag::err_typecheck_call_different_arg_types)
            << TyA << TyB;
 
-  if (checkMathBuiltinElementType(*this, A.get()->getBeginLoc(), TyA))
+  if (checkMathBuiltinElementType(*this, A.get()->getBeginLoc(), TyA, 1))
     return true;
 
   TheCall->setArg(0, A.get());