[clang] [llvm] Add cross builtins and cross HLSL function to DirectX and SPIR-V backend (PR #109180)

Joshua Batista via cfe-commits cfe-commits at lists.llvm.org
Fri Sep 20 15:32:12 PDT 2024


https://github.com/bob80905 updated https://github.com/llvm/llvm-project/pull/109180

>From 3c58861f3e8c2f1333d0b36c6f5b7ba7f3d9e187 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Wed, 18 Sep 2024 12:20:19 -0700
Subject: [PATCH 1/3] add cross hlsl function

---
 clang/include/clang/Basic/Builtins.td         |  6 ++
 clang/lib/CodeGen/CGBuiltin.cpp               | 15 +++++
 clang/lib/CodeGen/CGHLSLRuntime.h             |  1 +
 clang/lib/Headers/hlsl/hlsl_intrinsics.h      | 22 +++++++
 clang/lib/Sema/SemaHLSL.cpp                   | 14 +++++
 clang/test/CodeGenHLSL/builtins/cross.hlsl    | 36 ++++++++++++
 .../test/SemaHLSL/BuiltIns/cross-errors.hlsl  | 31 ++++++++++
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  1 +
 llvm/include/llvm/IR/IntrinsicsSPIRV.td       |  1 +
 .../Target/DirectX/DXILIntrinsicExpansion.cpp | 40 +++++++++++++
 .../Target/SPIRV/SPIRVInstructionSelector.cpp | 24 +++++++-
 llvm/test/CodeGen/DirectX/cross.ll            | 57 +++++++++++++++++++
 .../CodeGen/SPIRV/hlsl-intrinsics/cross.ll    | 33 +++++++++++
 13 files changed, 280 insertions(+), 1 deletion(-)
 create mode 100644 clang/test/CodeGenHLSL/builtins/cross.hlsl
 create mode 100644 clang/test/SemaHLSL/BuiltIns/cross-errors.hlsl
 create mode 100644 llvm/test/CodeGen/DirectX/cross.ll
 create mode 100644 llvm/test/CodeGen/SPIRV/hlsl-intrinsics/cross.ll

diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index 8c5d7ad763bf97..c4735b59dfb0be 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -4709,6 +4709,12 @@ def HLSLCreateHandle : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void*(unsigned char)";
 }
 
+def HLSLCross: LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_cross"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 def HLSLDotProduct : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_dot"];
   let Attributes = [NoThrow, Const];
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 7e18aafcdd4b8a..0883ad17ce0ffd 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -18639,6 +18639,21 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
         IsUnsigned ? Intrinsic::dx_uclamp : Intrinsic::dx_clamp,
         ArrayRef<Value *>{OpX, OpMin, OpMax}, nullptr, "dx.clamp");
   }
+  case Builtin::BI__builtin_hlsl_cross: {
+    Value *Op0 = EmitScalarExpr(E->getArg(0));
+    Value *Op1 = EmitScalarExpr(E->getArg(1));
+    assert(E->getArg(0)->getType()->hasFloatingRepresentation() &&
+           E->getArg(1)->getType()->hasFloatingRepresentation() &&
+           "step operands must have a float representation");
+    // make sure each vector has exactly 3 elements
+    auto *XVecTy1 = E->getArg(0)->getType()->getAs<VectorType>();
+    auto *XVecTy2 = E->getArg(1)->getType()->getAs<VectorType>();
+    assert(XVecTy1->getNumElements() == 3 && XVecTy2->getNumElements() == 3 &&
+           "input vectors must have 3 elements each");
+    return Builder.CreateIntrinsic(
+        /*ReturnType=*/Op0->getType(), CGM.getHLSLRuntime().getCrossIntrinsic(),
+        ArrayRef<Value *>{Op0, Op1}, nullptr, "hlsl.cross");
+  }
   case Builtin::BI__builtin_hlsl_dot: {
     Value *Op0 = EmitScalarExpr(E->getArg(0));
     Value *Op1 = EmitScalarExpr(E->getArg(1));
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.h b/clang/lib/CodeGen/CGHLSLRuntime.h
index a8aabca7348ffb..6722d2c7c50a2b 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.h
+++ b/clang/lib/CodeGen/CGHLSLRuntime.h
@@ -74,6 +74,7 @@ class CGHLSLRuntime {
 
   GENERATE_HLSL_INTRINSIC_FUNCTION(All, all)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Any, any)
+  GENERATE_HLSL_INTRINSIC_FUNCTION(Cross, cross)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Frac, frac)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Length, length)
   GENERATE_HLSL_INTRINSIC_FUNCTION(Lerp, lerp)
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
index 6cd6a2caf19994..e4b40978dea6b9 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
@@ -1563,6 +1563,28 @@ uint64_t3 reversebits(uint64_t3);
 _HLSL_BUILTIN_ALIAS(__builtin_elementwise_bitreverse)
 uint64_t4 reversebits(uint64_t4);
 
+//===----------------------------------------------------------------------===//
+// cross builtins
+//===----------------------------------------------------------------------===//
+
+/// \fn T cross(T x, T y)
+/// \brief Returns the cross product of two floating-point, 3D vectors.
+/// \param x [in] The first floating-point, 3D vector.
+/// \param y [in] The second floating-point, 3D vector.
+///
+/// Result is the cross product of x and y, i.e., the resulting
+/// components are, in order :
+/// x[1] * y[2] - y[1] * x[2]
+/// x[2] * y[0] - y[2] * x[0]
+/// x[0] * y[1] - y[0] * x[1]
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_cross)
+half3 cross(half3, half3);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_cross)
+float3 cross(float3, float3);
+
 //===----------------------------------------------------------------------===//
 // rcp builtins
 //===----------------------------------------------------------------------===//
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index a303f211501348..e7e7a3c259e27a 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -1704,6 +1704,20 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
+  case Builtin::BI__builtin_hlsl_cross: {
+    if (SemaRef.checkArgCount(TheCall, 2))
+      return true;
+    if (CheckVectorElementCallArgs(&SemaRef, TheCall))
+      return true;
+    if (CheckFloatOrHalfRepresentations(&SemaRef, TheCall))
+      return true;
+
+    ExprResult A = TheCall->getArg(0);
+    QualType ArgTyA = A.get()->getType();
+    // return type is the same as the input type
+    TheCall->setType(ArgTyA);
+    break;
+  }
   case Builtin::BI__builtin_hlsl_dot: {
     if (SemaRef.checkArgCount(TheCall, 2))
       return true;
diff --git a/clang/test/CodeGenHLSL/builtins/cross.hlsl b/clang/test/CodeGenHLSL/builtins/cross.hlsl
new file mode 100644
index 00000000000000..047e7fef7136fe
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/cross.hlsl
@@ -0,0 +1,36 @@
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -fnative-half-type \
+// RUN:   -emit-llvm -disable-llvm-passes -o - | FileCheck %s \
+// RUN:   --check-prefixes=CHECK,NATIVE_HALF \
+// RUN:   -DFNATTRS=noundef -DTARGET=dx
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN:   dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes \
+// RUN:   -o - | FileCheck %s --check-prefixes=CHECK,NO_HALF \
+// RUN:   -DFNATTRS=noundef -DTARGET=dx
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN:   spirv-unknown-vulkan-compute %s -fnative-half-type \
+// RUN:   -emit-llvm -disable-llvm-passes -o - | FileCheck %s \
+// RUN:   --check-prefixes=CHECK,NATIVE_HALF \
+// RUN:   -DFNATTRS="spir_func noundef" -DTARGET=spv
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN:   spirv-unknown-vulkan-compute %s -emit-llvm -disable-llvm-passes \
+// RUN:   -o - | FileCheck %s --check-prefixes=CHECK,NO_HALF \
+// RUN:   -DFNATTRS="spir_func noundef" -DTARGET=spv
+
+// NATIVE_HALF: define [[FNATTRS]] <3 x half> @
+// NATIVE_HALF: call <3 x half> @llvm.[[TARGET]].cross.v3f16(<3 x half>
+// NO_HALF: call <3 x float> @llvm.[[TARGET]].cross.v3f32(<3 x float>
+// NATIVE_HALF: ret <3 x half> %hlsl.cross
+// NO_HALF: ret <3 x float> %hlsl.cross
+half3 test_cross_half3(half3 p0, half3 p1)
+{
+    return cross(p0, p1);
+}
+
+// CHECK: define [[FNATTRS]] <3 x float> @
+// CHECK: %hlsl.cross = call <3 x float> @llvm.[[TARGET]].cross.v3f32(
+// CHECK: ret <3 x float> %hlsl.cross
+float3 test_cross_float3(float3 p0, float3 p1)
+{
+    return cross(p0, p1);
+}
diff --git a/clang/test/SemaHLSL/BuiltIns/cross-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/cross-errors.hlsl
new file mode 100644
index 00000000000000..40ab4b533a495e
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/cross-errors.hlsl
@@ -0,0 +1,31 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -fnative-half-type -disable-llvm-passes -verify -verify-ignore-unexpected
+
+void test_too_few_arg()
+{
+  return __builtin_hlsl_cross();
+  // expected-error at -1 {{too few arguments to function call, expected 2, have 0}}
+}
+
+void test_too_many_arg(float3 p0)
+{
+  return __builtin_hlsl_cross(p0, p0, p0);
+  // expected-error at -1 {{too many arguments to function call, expected 2, have 3}}
+}
+
+bool builtin_bool_to_float_type_promotion(bool p1)
+{
+  return __builtin_hlsl_cross(p1, p1);
+  // expected-error at -1 {passing 'bool' to parameter of incompatible type 'float'}}
+}
+
+bool builtin_cross_int_to_float_promotion(int p1)
+{
+  return __builtin_hlsl_cross(p1, p1);
+  // expected-error at -1 {{passing 'int' to parameter of incompatible type 'float'}}
+}
+
+bool2 builtin_cross_int2_to_float2_promotion(int2 p1)
+{
+  return __builtin_hlsl_cross(p1, p1);
+  // expected-error at -1 {{passing 'int2' (aka 'vector<int, 2>') to parameter of incompatible type '__attribute__((__vector_size__(2 * sizeof(float)))) float' (vector of 2 'float' values)}}
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 3ce7b8b987ef86..f4242772bab20c 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -44,6 +44,7 @@ def int_dx_cast_handle : Intrinsic<[llvm_any_ty], [llvm_any_ty]>;
 def int_dx_all : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_any : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_clamp : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+def int_dx_cross : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_uclamp : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_saturate : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 
diff --git a/llvm/include/llvm/IR/IntrinsicsSPIRV.td b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
index a4c01952927175..480b391bd54fdf 100644
--- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td
+++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
@@ -60,6 +60,7 @@ let TargetPrefix = "spv" in {
       Intrinsic<[ llvm_ptr_ty ], [llvm_i8_ty], [IntrWillReturn]>;
   def int_spv_all : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty]>;
   def int_spv_any : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty]>;
+  def int_spv_cross : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
   def int_spv_frac : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty]>;
   def int_spv_lerp : Intrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>,LLVMMatchType<0>],
     [IntrNoMem, IntrWillReturn] >;
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index dd73b895b14d37..e921dffede38f8 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -43,6 +43,7 @@ static bool isIntrinsicExpansion(Function &F) {
   case Intrinsic::dx_all:
   case Intrinsic::dx_any:
   case Intrinsic::dx_clamp:
+  case Intrinsic::dx_cross:
   case Intrinsic::dx_uclamp:
   case Intrinsic::dx_lerp:
   case Intrinsic::dx_length:
@@ -73,6 +74,42 @@ static Value *expandAbs(CallInst *Orig) {
                                  "dx.max");
 }
 
+static Value *expandCrossIntrinsic(CallInst *Orig) {
+
+  VectorType *VT = cast<VectorType>(Orig->getType());
+  if (cast<FixedVectorType>(VT)->getNumElements() != 3)
+    report_fatal_error(Twine("return vector must have exactly 3 elements"),
+                       /* gen_crash_diag=*/false);
+
+  Value *op0 = Orig->getOperand(0);
+  Value *op1 = Orig->getOperand(1);
+  IRBuilder<> Builder(Orig);
+
+  Value *op0_x = Builder.CreateExtractElement(op0, (uint64_t)0);
+  Value *op0_y = Builder.CreateExtractElement(op0, 1);
+  Value *op0_z = Builder.CreateExtractElement(op0, 2);
+
+  Value *op1_x = Builder.CreateExtractElement(op1, (uint64_t)0);
+  Value *op1_y = Builder.CreateExtractElement(op1, 1);
+  Value *op1_z = Builder.CreateExtractElement(op1, 2);
+
+  auto MulSub = [&](Value *x0, Value *y0, Value *x1, Value *y1) -> Value * {
+    Value *xy = Builder.CreateFMul(x0, y1);
+    Value *yx = Builder.CreateFMul(y0, x1);
+    return Builder.CreateFSub(xy, yx);
+  };
+
+  Value *yz_zy = MulSub(op0_y, op0_z, op1_y, op1_z);
+  Value *zx_xz = MulSub(op0_z, op0_x, op1_z, op1_x);
+  Value *xy_yx = MulSub(op0_x, op0_y, op1_x, op1_y);
+
+  Value *cross = UndefValue::get(VT);
+  cross = Builder.CreateInsertElement(cross, yz_zy, (uint64_t)0);
+  cross = Builder.CreateInsertElement(cross, zx_xz, 1);
+  cross = Builder.CreateInsertElement(cross, xy_yx, 2);
+  return cross;
+}
+
 // Create appropriate DXIL float dot intrinsic for the given A and B operands
 // The appropriate opcode will be determined by the size of the operands
 // The dot product is placed in the position indicated by Orig
@@ -434,6 +471,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
   case Intrinsic::dx_any:
     Result = expandAnyOrAllIntrinsic(Orig, IntrinsicId);
     break;
+  case Intrinsic::dx_cross:
+    Result = expandCrossIntrinsic(Orig);
+    break;
   case Intrinsic::dx_uclamp:
   case Intrinsic::dx_clamp:
     Result = expandClampIntrinsic(Orig, IntrinsicId);
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index b526c9f29f1e6a..677119840709aa 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -167,7 +167,8 @@ class SPIRVInstructionSelector : public InstructionSelector {
 
   bool selectCmp(Register ResVReg, const SPIRVType *ResType,
                  unsigned comparisonOpcode, MachineInstr &I) const;
-
+  bool selectCross(Register ResVReg, const SPIRVType *ResType,
+                   MachineInstr &I) const;
   bool selectICmp(Register ResVReg, const SPIRVType *ResType,
                   MachineInstr &I) const;
   bool selectFCmp(Register ResVReg, const SPIRVType *ResType,
@@ -1465,6 +1466,25 @@ bool SPIRVInstructionSelector::selectAny(Register ResVReg,
   return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAny);
 }
 
+bool SPIRVInstructionSelector::selectCross(Register ResVReg,
+                                           const SPIRVType *ResType,
+                                           MachineInstr &I) const {
+
+  assert(I.getNumOperands() == 4);
+  assert(I.getOperand(2).isReg());
+  assert(I.getOperand(3).isReg());
+  MachineBasicBlock &BB = *I.getParent();
+
+  return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
+      .addDef(ResVReg)
+      .addUse(GR.getSPIRVTypeID(ResType))
+      .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
+      .addImm(GL::Cross)
+      .addUse(I.getOperand(2).getReg())
+      .addUse(I.getOperand(3).getReg())
+      .constrainAllUses(TII, TRI, RBI);
+}
+
 bool SPIRVInstructionSelector::selectFmix(Register ResVReg,
                                           const SPIRVType *ResType,
                                           MachineInstr &I) const {
@@ -2458,6 +2478,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
     return selectAll(ResVReg, ResType, I);
   case Intrinsic::spv_any:
     return selectAny(ResVReg, ResType, I);
+  case Intrinsic::spv_cross:
+    return selectCross(ResVReg, ResType, I);
   case Intrinsic::spv_lerp:
     return selectFmix(ResVReg, ResType, I);
   case Intrinsic::spv_length:
diff --git a/llvm/test/CodeGen/DirectX/cross.ll b/llvm/test/CodeGen/DirectX/cross.ll
new file mode 100644
index 00000000000000..90847ac635dbba
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/cross.ll
@@ -0,0 +1,57 @@
+; RUN: opt -S  -dxil-intrinsic-expansion  < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+
+; Make sure dxil operation function calls for cross are generated for half/float.
+
+declare <3 x half> @llvm.dx.cross.v3f16(<3 x half>, <3 x half>)
+declare <3 x float> @llvm.dx.cross.v3f32(<3 x float>, <3 x float>)
+
+define noundef <3 x half> @test_cross_half3(<3 x half> noundef %p0, <3 x half> noundef %p1) {
+entry:
+  ; CHECK: %0 = extractelement <3 x half> %p0, i64 0
+  ; CHECK: %1 = extractelement <3 x half> %p0, i64 1
+  ; CHECK: %2 = extractelement <3 x half> %p0, i64 2
+  ; CHECK: %3 = extractelement <3 x half> %p1, i64 0
+  ; CHECK: %4 = extractelement <3 x half> %p1, i64 1
+  ; CHECK: %5 = extractelement <3 x half> %p1, i64 2
+  ; CHECK: %6 = fmul half %1, %5
+  ; CHECK: %7 = fmul half %2, %4
+  ; CHECK: %8 = fsub half %6, %7
+  ; CHECK: %9 = fmul half %2, %3
+  ; CHECK: %10 = fmul half %0, %5
+  ; CHECK: %11 = fsub half %9, %10
+  ; CHECK: %12 = fmul half %0, %4
+  ; CHECK: %13 = fmul half %1, %3
+  ; CHECK: %14 = fsub half %12, %13
+  ; CHECK: %15 = insertelement <3 x half> undef, half %8, i64 0
+  ; CHECK: %16 = insertelement <3 x half> %15, half %11, i64 1
+  ; CHECK: %17 = insertelement <3 x half> %16, half %14, i64 2
+  ; CHECK: ret <3 x half> %17
+  %hlsl.cross = call <3 x half> @llvm.dx.cross.v3f16(<3 x half> %p0, <3 x half> %p1)
+  ret <3 x half> %hlsl.cross
+}
+
+define noundef <3 x float> @test_cross_float3(<3 x float> noundef %p0, <3 x float> noundef %p1) {
+entry:
+  ; CHECK: %0 = extractelement <3 x float> %p0, i64 0
+  ; CHECK: %1 = extractelement <3 x float> %p0, i64 1
+  ; CHECK: %2 = extractelement <3 x float> %p0, i64 2
+  ; CHECK: %3 = extractelement <3 x float> %p1, i64 0
+  ; CHECK: %4 = extractelement <3 x float> %p1, i64 1
+  ; CHECK: %5 = extractelement <3 x float> %p1, i64 2
+  ; CHECK: %6 = fmul float %1, %5
+  ; CHECK: %7 = fmul float %2, %4
+  ; CHECK: %8 = fsub float %6, %7
+  ; CHECK: %9 = fmul float %2, %3
+  ; CHECK: %10 = fmul float %0, %5
+  ; CHECK: %11 = fsub float %9, %10
+  ; CHECK: %12 = fmul float %0, %4
+  ; CHECK: %13 = fmul float %1, %3
+  ; CHECK: %14 = fsub float %12, %13
+  ; CHECK: %15 = insertelement <3 x float> undef, float %8, i64 0
+  ; CHECK: %16 = insertelement <3 x float> %15, float %11, i64 1
+  ; CHECK: %17 = insertelement <3 x float> %16, float %14, i64 2
+  ; CHECK: ret <3 x float> %17
+  %hlsl.cross = call <3 x float> @llvm.dx.cross.v3f32(<3 x float> %p0, <3 x float> %p1)
+  ret <3 x float> %hlsl.cross
+}
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/cross.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/cross.ll
new file mode 100644
index 00000000000000..2e0eb8c429ac27
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/cross.ll
@@ -0,0 +1,33 @@
+; RUN: llc -O0 -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-unknown-unknown %s -o - -filetype=obj | spirv-val %}
+
+; Make sure SPIRV operation function calls for cross are lowered correctly.
+
+; CHECK-DAG: %[[#op_ext_glsl:]] = OpExtInstImport "GLSL.std.450"
+; CHECK-DAG: %[[#float_32:]] = OpTypeFloat 32
+; CHECK-DAG: %[[#float_16:]] = OpTypeFloat 16
+; CHECK-DAG: %[[#vec3_float_16:]] = OpTypeVector %[[#float_16]] 3
+; CHECK-DAG: %[[#vec3_float_32:]] = OpTypeVector %[[#float_32]] 3
+
+define noundef <3 x half> @cross_half4(<3 x half> noundef %a, <3 x half> noundef %b) {
+entry:
+  ; CHECK: %[[#]] = OpFunction %[[#vec3_float_16]] None %[[#]]
+  ; CHECK: %[[#arg0:]] = OpFunctionParameter %[[#vec3_float_16]]
+  ; CHECK: %[[#arg1:]] = OpFunctionParameter %[[#vec3_float_16]]
+  ; CHECK: %[[#]] = OpExtInst %[[#vec3_float_16]] %[[#op_ext_glsl]] Cross %[[#arg0]] %[[#arg1]]
+  %hlsl.cross = call <3 x half> @llvm.spv.cross.v4f16(<3 x half> %a, <3 x half> %b)
+  ret <3 x half> %hlsl.cross
+}
+
+define noundef <3 x float> @cross_float4(<3 x float> noundef %a, <3 x float> noundef %b) {
+entry:
+  ; CHECK: %[[#]] = OpFunction %[[#vec3_float_32]] None %[[#]]
+  ; CHECK: %[[#arg0:]] = OpFunctionParameter %[[#vec3_float_32]]
+  ; CHECK: %[[#arg1:]] = OpFunctionParameter %[[#vec3_float_32]]
+  ; CHECK: %[[#]] = OpExtInst %[[#vec3_float_32]] %[[#op_ext_glsl]] Cross %[[#arg0]] %[[#arg1]]
+  %hlsl.cross = call <3 x float> @llvm.spv.cross.v4f32(<3 x float> %a, <3 x float> %b)
+  ret <3 x float> %hlsl.cross
+}
+
+declare <3 x half> @llvm.spv.cross.v4f16(<3 x half>, <3 x half>)
+declare <3 x float> @llvm.spv.cross.v4f32(<3 x float>, <3 x float>)

>From 9109b48786888ffc708e163441d5f6c7fb8977c3 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 20 Sep 2024 15:08:16 -0700
Subject: [PATCH 2/3] address Farzon

---
 clang/lib/CodeGen/CGBuiltin.cpp               |  2 +-
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  2 +-
 .../Target/DirectX/DXILIntrinsicExpansion.cpp | 12 ++--
 llvm/test/CodeGen/DirectX/cross.ll            | 72 +++++++++----------
 4 files changed, 44 insertions(+), 44 deletions(-)

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 0883ad17ce0ffd..47fff98e99a4ba 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -18644,7 +18644,7 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
     Value *Op1 = EmitScalarExpr(E->getArg(1));
     assert(E->getArg(0)->getType()->hasFloatingRepresentation() &&
            E->getArg(1)->getType()->hasFloatingRepresentation() &&
-           "step operands must have a float representation");
+           "cross operands must have a float representation");
     // make sure each vector has exactly 3 elements
     auto *XVecTy1 = E->getArg(0)->getType()->getAs<VectorType>();
     auto *XVecTy2 = E->getArg(1)->getType()->getAs<VectorType>();
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index f4242772bab20c..c3ae6c9168ef4c 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -44,8 +44,8 @@ def int_dx_cast_handle : Intrinsic<[llvm_any_ty], [llvm_any_ty]>;
 def int_dx_all : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_any : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_clamp : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
-def int_dx_cross : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_uclamp : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+def int_dx_cross : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_saturate : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 
 def int_dx_dot2 :
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index e921dffede38f8..445da7b30139f1 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -85,13 +85,13 @@ static Value *expandCrossIntrinsic(CallInst *Orig) {
   Value *op1 = Orig->getOperand(1);
   IRBuilder<> Builder(Orig);
 
-  Value *op0_x = Builder.CreateExtractElement(op0, (uint64_t)0);
-  Value *op0_y = Builder.CreateExtractElement(op0, 1);
-  Value *op0_z = Builder.CreateExtractElement(op0, 2);
+  Value *op0_x = Builder.CreateExtractElement(op0, (uint64_t)0, "x0");
+  Value *op0_y = Builder.CreateExtractElement(op0, 1, "y0");
+  Value *op0_z = Builder.CreateExtractElement(op0, 2, "z0");
 
-  Value *op1_x = Builder.CreateExtractElement(op1, (uint64_t)0);
-  Value *op1_y = Builder.CreateExtractElement(op1, 1);
-  Value *op1_z = Builder.CreateExtractElement(op1, 2);
+  Value *op1_x = Builder.CreateExtractElement(op1, (uint64_t)0, "x1");
+  Value *op1_y = Builder.CreateExtractElement(op1, 1, "y1");
+  Value *op1_z = Builder.CreateExtractElement(op1, 2, "z1");
 
   auto MulSub = [&](Value *x0, Value *y0, Value *x1, Value *y1) -> Value * {
     Value *xy = Builder.CreateFMul(x0, y1);
diff --git a/llvm/test/CodeGen/DirectX/cross.ll b/llvm/test/CodeGen/DirectX/cross.ll
index 90847ac635dbba..bc665244ed20c6 100644
--- a/llvm/test/CodeGen/DirectX/cross.ll
+++ b/llvm/test/CodeGen/DirectX/cross.ll
@@ -8,50 +8,50 @@ declare <3 x float> @llvm.dx.cross.v3f32(<3 x float>, <3 x float>)
 
 define noundef <3 x half> @test_cross_half3(<3 x half> noundef %p0, <3 x half> noundef %p1) {
 entry:
-  ; CHECK: %0 = extractelement <3 x half> %p0, i64 0
-  ; CHECK: %1 = extractelement <3 x half> %p0, i64 1
-  ; CHECK: %2 = extractelement <3 x half> %p0, i64 2
-  ; CHECK: %3 = extractelement <3 x half> %p1, i64 0
-  ; CHECK: %4 = extractelement <3 x half> %p1, i64 1
-  ; CHECK: %5 = extractelement <3 x half> %p1, i64 2
-  ; CHECK: %6 = fmul half %1, %5
-  ; CHECK: %7 = fmul half %2, %4
+  ; CHECK: %x0 = extractelement <3 x half> %p0, i64 0
+  ; CHECK: %y0 = extractelement <3 x half> %p0, i64 1
+  ; CHECK: %z0 = extractelement <3 x half> %p0, i64 2
+  ; CHECK: %x1 = extractelement <3 x half> %p1, i64 0
+  ; CHECK: %y1 = extractelement <3 x half> %p1, i64 1
+  ; CHECK: %z1 = extractelement <3 x half> %p1, i64 2
+  ; CHECK: %0 = fmul half %y0, %z1
+  ; CHECK: %1 = fmul half %z0, %y1
+  ; CHECK: %2 = fsub half %0, %1
+  ; CHECK: %3 = fmul half %z0, %x1
+  ; CHECK: %4 = fmul half %x0, %z1
+  ; CHECK: %5 = fsub half %3, %4
+  ; CHECK: %6 = fmul half %x0, %y1
+  ; CHECK: %7 = fmul half %y0, %x1
   ; CHECK: %8 = fsub half %6, %7
-  ; CHECK: %9 = fmul half %2, %3
-  ; CHECK: %10 = fmul half %0, %5
-  ; CHECK: %11 = fsub half %9, %10
-  ; CHECK: %12 = fmul half %0, %4
-  ; CHECK: %13 = fmul half %1, %3
-  ; CHECK: %14 = fsub half %12, %13
-  ; CHECK: %15 = insertelement <3 x half> undef, half %8, i64 0
-  ; CHECK: %16 = insertelement <3 x half> %15, half %11, i64 1
-  ; CHECK: %17 = insertelement <3 x half> %16, half %14, i64 2
-  ; CHECK: ret <3 x half> %17
+  ; CHECK: %9 = insertelement <3 x half> undef, half %2, i64 0
+  ; CHECK: %10 = insertelement <3 x half> %9, half %5, i64 1
+  ; CHECK: %11 = insertelement <3 x half> %10, half %8, i64 2
+  ; CHECK: ret <3 x half> %11
   %hlsl.cross = call <3 x half> @llvm.dx.cross.v3f16(<3 x half> %p0, <3 x half> %p1)
   ret <3 x half> %hlsl.cross
 }
 
 define noundef <3 x float> @test_cross_float3(<3 x float> noundef %p0, <3 x float> noundef %p1) {
 entry:
-  ; CHECK: %0 = extractelement <3 x float> %p0, i64 0
-  ; CHECK: %1 = extractelement <3 x float> %p0, i64 1
-  ; CHECK: %2 = extractelement <3 x float> %p0, i64 2
-  ; CHECK: %3 = extractelement <3 x float> %p1, i64 0
-  ; CHECK: %4 = extractelement <3 x float> %p1, i64 1
-  ; CHECK: %5 = extractelement <3 x float> %p1, i64 2
-  ; CHECK: %6 = fmul float %1, %5
-  ; CHECK: %7 = fmul float %2, %4
+  ; CHECK: %x0 = extractelement <3 x float> %p0, i64 0
+  ; CHECK: %y0 = extractelement <3 x float> %p0, i64 1
+  ; CHECK: %z0 = extractelement <3 x float> %p0, i64 2
+  ; CHECK: %x1 = extractelement <3 x float> %p1, i64 0
+  ; CHECK: %y1 = extractelement <3 x float> %p1, i64 1
+  ; CHECK: %z1 = extractelement <3 x float> %p1, i64 2
+  ; CHECK: %0 = fmul float %y0, %z1
+  ; CHECK: %1 = fmul float %z0, %y1
+  ; CHECK: %2 = fsub float %0, %1
+  ; CHECK: %3 = fmul float %z0, %x1
+  ; CHECK: %4 = fmul float %x0, %z1
+  ; CHECK: %5 = fsub float %3, %4
+  ; CHECK: %6 = fmul float %x0, %y1
+  ; CHECK: %7 = fmul float %y0, %x1
   ; CHECK: %8 = fsub float %6, %7
-  ; CHECK: %9 = fmul float %2, %3
-  ; CHECK: %10 = fmul float %0, %5
-  ; CHECK: %11 = fsub float %9, %10
-  ; CHECK: %12 = fmul float %0, %4
-  ; CHECK: %13 = fmul float %1, %3
-  ; CHECK: %14 = fsub float %12, %13
-  ; CHECK: %15 = insertelement <3 x float> undef, float %8, i64 0
-  ; CHECK: %16 = insertelement <3 x float> %15, float %11, i64 1
-  ; CHECK: %17 = insertelement <3 x float> %16, float %14, i64 2
-  ; CHECK: ret <3 x float> %17
+  ; CHECK: %9 = insertelement <3 x float> undef, float %2, i64 0
+  ; CHECK: %10 = insertelement <3 x float> %9, float %5, i64 1
+  ; CHECK: %11 = insertelement <3 x float> %10, float %8, i64 2
+  ; CHECK: ret <3 x float> %11
   %hlsl.cross = call <3 x float> @llvm.dx.cross.v3f32(<3 x float> %p0, <3 x float> %p1)
   ret <3 x float> %hlsl.cross
 }

>From b5dc3e74bb9e2a347eb637f89b5c850af3d0ece4 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 20 Sep 2024 15:31:53 -0700
Subject: [PATCH 3/3] use correct naming convention for vec elements

---
 .../Target/DirectX/DXILIntrinsicExpansion.cpp |  8 ++---
 llvm/test/CodeGen/DirectX/cross.ll            | 36 +++++++++----------
 2 files changed, 22 insertions(+), 22 deletions(-)

diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 445da7b30139f1..c6b5e8ddb13bc2 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -86,12 +86,12 @@ static Value *expandCrossIntrinsic(CallInst *Orig) {
   IRBuilder<> Builder(Orig);
 
   Value *op0_x = Builder.CreateExtractElement(op0, (uint64_t)0, "x0");
-  Value *op0_y = Builder.CreateExtractElement(op0, 1, "y0");
-  Value *op0_z = Builder.CreateExtractElement(op0, 2, "z0");
+  Value *op0_y = Builder.CreateExtractElement(op0, 1, "x1");
+  Value *op0_z = Builder.CreateExtractElement(op0, 2, "x2");
 
-  Value *op1_x = Builder.CreateExtractElement(op1, (uint64_t)0, "x1");
+  Value *op1_x = Builder.CreateExtractElement(op1, (uint64_t)0, "y0");
   Value *op1_y = Builder.CreateExtractElement(op1, 1, "y1");
-  Value *op1_z = Builder.CreateExtractElement(op1, 2, "z1");
+  Value *op1_z = Builder.CreateExtractElement(op1, 2, "y2");
 
   auto MulSub = [&](Value *x0, Value *y0, Value *x1, Value *y1) -> Value * {
     Value *xy = Builder.CreateFMul(x0, y1);
diff --git a/llvm/test/CodeGen/DirectX/cross.ll b/llvm/test/CodeGen/DirectX/cross.ll
index bc665244ed20c6..74934a628988c3 100644
--- a/llvm/test/CodeGen/DirectX/cross.ll
+++ b/llvm/test/CodeGen/DirectX/cross.ll
@@ -9,19 +9,19 @@ declare <3 x float> @llvm.dx.cross.v3f32(<3 x float>, <3 x float>)
 define noundef <3 x half> @test_cross_half3(<3 x half> noundef %p0, <3 x half> noundef %p1) {
 entry:
   ; CHECK: %x0 = extractelement <3 x half> %p0, i64 0
-  ; CHECK: %y0 = extractelement <3 x half> %p0, i64 1
-  ; CHECK: %z0 = extractelement <3 x half> %p0, i64 2
-  ; CHECK: %x1 = extractelement <3 x half> %p1, i64 0
+  ; CHECK: %x1 = extractelement <3 x half> %p0, i64 1
+  ; CHECK: %x2 = extractelement <3 x half> %p0, i64 2
+  ; CHECK: %y0 = extractelement <3 x half> %p1, i64 0
   ; CHECK: %y1 = extractelement <3 x half> %p1, i64 1
-  ; CHECK: %z1 = extractelement <3 x half> %p1, i64 2
-  ; CHECK: %0 = fmul half %y0, %z1
-  ; CHECK: %1 = fmul half %z0, %y1
+  ; CHECK: %y2 = extractelement <3 x half> %p1, i64 2
+  ; CHECK: %0 = fmul half %x1, %y2
+  ; CHECK: %1 = fmul half %x2, %y1
   ; CHECK: %2 = fsub half %0, %1
-  ; CHECK: %3 = fmul half %z0, %x1
-  ; CHECK: %4 = fmul half %x0, %z1
+  ; CHECK: %3 = fmul half %x2, %y0
+  ; CHECK: %4 = fmul half %x0, %y2
   ; CHECK: %5 = fsub half %3, %4
   ; CHECK: %6 = fmul half %x0, %y1
-  ; CHECK: %7 = fmul half %y0, %x1
+  ; CHECK: %7 = fmul half %x1, %y0
   ; CHECK: %8 = fsub half %6, %7
   ; CHECK: %9 = insertelement <3 x half> undef, half %2, i64 0
   ; CHECK: %10 = insertelement <3 x half> %9, half %5, i64 1
@@ -34,19 +34,19 @@ entry:
 define noundef <3 x float> @test_cross_float3(<3 x float> noundef %p0, <3 x float> noundef %p1) {
 entry:
   ; CHECK: %x0 = extractelement <3 x float> %p0, i64 0
-  ; CHECK: %y0 = extractelement <3 x float> %p0, i64 1
-  ; CHECK: %z0 = extractelement <3 x float> %p0, i64 2
-  ; CHECK: %x1 = extractelement <3 x float> %p1, i64 0
+  ; CHECK: %x1 = extractelement <3 x float> %p0, i64 1
+  ; CHECK: %x2 = extractelement <3 x float> %p0, i64 2
+  ; CHECK: %y0 = extractelement <3 x float> %p1, i64 0
   ; CHECK: %y1 = extractelement <3 x float> %p1, i64 1
-  ; CHECK: %z1 = extractelement <3 x float> %p1, i64 2
-  ; CHECK: %0 = fmul float %y0, %z1
-  ; CHECK: %1 = fmul float %z0, %y1
+  ; CHECK: %y2 = extractelement <3 x float> %p1, i64 2
+  ; CHECK: %0 = fmul float %x1, %y2
+  ; CHECK: %1 = fmul float %x2, %y1
   ; CHECK: %2 = fsub float %0, %1
-  ; CHECK: %3 = fmul float %z0, %x1
-  ; CHECK: %4 = fmul float %x0, %z1
+  ; CHECK: %3 = fmul float %x2, %y0
+  ; CHECK: %4 = fmul float %x0, %y2
   ; CHECK: %5 = fsub float %3, %4
   ; CHECK: %6 = fmul float %x0, %y1
-  ; CHECK: %7 = fmul float %y0, %x1
+  ; CHECK: %7 = fmul float %x1, %y0
   ; CHECK: %8 = fsub float %6, %7
   ; CHECK: %9 = insertelement <3 x float> undef, float %2, i64 0
   ; CHECK: %10 = insertelement <3 x float> %9, float %5, i64 1



More information about the cfe-commits mailing list