Add cross builtins and cross HLSL function to DirectX and SPIR-V back…

…end (llvm#109180)

This PR adds the step intrinsic and an HLSL function that uses it.
The SPIRV backend is also implemented.

Used llvm#106471 as a reference.
Fixes llvm#99095

clang/include/clang/Basic/Builtins.td

@@ -4739,6 +4739,12 @@ def HLSLClamp : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void(...)";
 }
 
+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];

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -10510,8 +10510,8 @@ def err_first_argument_to_cwsc_pdtor_call : Error<
 def err_second_argument_to_cwsc_not_pointer : Error<
   "second argument to __builtin_call_with_static_chain must be of pointer type">;
 
-def err_vector_incorrect_num_initializers : Error<
-  "%select{too many|too few}0 elements in vector initialization (expected %1 elements, have %2)">;
+def err_vector_incorrect_num_elements : Error<
+  "%select{too many|too few}0 elements in vector %select{initialization|operand}3 (expected %1 elements, have %2)">;
 def err_altivec_empty_initializer : Error<"expected initializer">;
 
 def err_invalid_neon_type_code : Error<

clang/lib/CodeGen/CGBuiltin.cpp

@@ -18667,6 +18667,21 @@ 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() &&
+           "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>();
+    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));

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)

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -1642,6 +1642,28 @@ 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
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaHLSL.cpp

@@ -1828,6 +1828,41 @@ 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;
+    // ensure both args have 3 elements
+    int NumElementsArg1 =
+        TheCall->getArg(0)->getType()->getAs<VectorType>()->getNumElements();
+    int NumElementsArg2 =
+        TheCall->getArg(1)->getType()->getAs<VectorType>()->getNumElements();
+
+    if (NumElementsArg1 != 3) {
+      int LessOrMore = NumElementsArg1 > 3 ? 1 : 0;
+      SemaRef.Diag(TheCall->getBeginLoc(),
+                   diag::err_vector_incorrect_num_elements)
+          << LessOrMore << 3 << NumElementsArg1 << /*operand*/ 1;
+      return true;
+    }
+    if (NumElementsArg2 != 3) {
+      int LessOrMore = NumElementsArg2 > 3 ? 1 : 0;
+
+      SemaRef.Diag(TheCall->getBeginLoc(),
+                   diag::err_vector_incorrect_num_elements)
+          << LessOrMore << 3 << NumElementsArg2 << /*operand*/ 1;
+      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;

clang/lib/Sema/SemaInit.cpp

@@ -1976,8 +1976,9 @@ void InitListChecker::CheckVectorType(const InitializedEntity &Entity,
   if (numEltsInit != maxElements) {
     if (!VerifyOnly)
       SemaRef.Diag(IList->getBeginLoc(),
-                   diag::err_vector_incorrect_num_initializers)
-          << (numEltsInit < maxElements) << maxElements << numEltsInit;
+                   diag::err_vector_incorrect_num_elements)
+          << (numEltsInit < maxElements) << maxElements << numEltsInit
+          << /*initialization*/ 0;
     hadError = true;
   }
 }

clang/test/CodeGenHLSL/builtins/cross.hlsl

@@ -0,0 +1,37 @@
+// 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>
+// NATIVE_HALF: ret <3 x half> %hlsl.cross
+// NO_HALF: define [[FNATTRS]] <3 x float> @
+// NO_HALF: call <3 x float> @llvm.[[TARGET]].cross.v3f32(<3 x float>
+// 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);
+}

clang/test/SemaHLSL/BuiltIns/cross-errors.hlsl

@@ -0,0 +1,43 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -fnative-half-type -disable-llvm-passes -verify
+
+void test_too_few_arg()
+{
+  return __builtin_hlsl_cross();
+  // expected-error@-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@-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@-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@-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@-1 {{passing 'int2' (aka 'vector<int, 2>') to parameter of incompatible type '__attribute__((__vector_size__(2 * sizeof(float)))) float' (vector of 2 'float' values)}}
+}
+
+float2 builtin_cross_float2(float2 p1, float2 p2)
+{
+  return __builtin_hlsl_cross(p1, p2);
+  // expected-error@-1 {{too many elements in vector operand (expected 3 elements, have 2)}}
+}
+
+float3  builtin_cross_float3_int3(float3 p1, int3 p2)
+{
+  return __builtin_hlsl_cross(p1, p2);
+  // expected-error@-1 {{all arguments to '__builtin_hlsl_cross' must have the same type}}
+}

llvm/include/llvm/IR/IntrinsicsDirectX.td

@@ -42,6 +42,7 @@ 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_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 :

llvm/include/llvm/IR/IntrinsicsSPIRV.td

@@ -61,6 +61,7 @@ let TargetPrefix = "spv" in {
   def int_spv_thread_id : Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem, IntrWillReturn]>;
   def int_spv_all : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
   def int_spv_any : DefaultAttrsIntrinsic<[llvm_i1_ty], [llvm_any_ty], [IntrNoMem]>;
+  def int_spv_cross : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
   def int_spv_frac : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty], [IntrNoMem]>;
   def int_spv_lerp : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>,LLVMMatchType<0>],
     [IntrNoMem] >;

llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp

@@ -54,6 +54,7 @@ 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:
@@ -84,6 +85,42 @@ 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, "x0");
+  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, "y0");
+  Value *op1_y = Builder.CreateExtractElement(op1, 1, "y1");
+  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);
+    Value *yx = Builder.CreateFMul(y0, x1);
+    return Builder.CreateFSub(xy, yx, Orig->getName());
+  };
+
+  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
@@ -496,6 +533,9 @@ 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);

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -145,7 +145,8 @@ 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,
@@ -2500,6 +2501,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
     return selectAll(ResVReg, ResType, I);
   case Intrinsic::spv_any:
     return selectAny(ResVReg, ResType, I);
+  case Intrinsic::spv_cross:
+    return selectExtInst(ResVReg, ResType, I, CL::cross, GL::Cross);
   case Intrinsic::spv_lerp:
     return selectExtInst(ResVReg, ResType, I, CL::mix, GL::FMix);
   case Intrinsic::spv_length:

llvm/test/CodeGen/DirectX/cross.ll

@@ -0,0 +1,56 @@
+; RUN: opt -S  -dxil-intrinsic-expansion  < %s | FileCheck %s
+
+; 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: %x0 = extractelement <3 x half> %p0, 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: %y2 = extractelement <3 x half> %p1, i64 2
+  ; CHECK: %0 = fmul half %x1, %y2
+  ; CHECK: %1 = fmul half %x2, %y1
+  ; CHECK: %hlsl.cross1 = fsub half %0, %1
+  ; CHECK: %2 = fmul half %x2, %y0
+  ; CHECK: %3 = fmul half %x0, %y2
+  ; CHECK: %hlsl.cross2 = fsub half %2, %3
+  ; CHECK: %4 = fmul half %x0, %y1
+  ; CHECK: %5 = fmul half %x1, %y0
+  ; CHECK: %hlsl.cross3 = fsub half %4, %5
+  ; CHECK: %6 = insertelement <3 x half> undef, half %hlsl.cross1, i64 0
+  ; CHECK: %7 = insertelement <3 x half> %6, half %hlsl.cross2, i64 1
+  ; CHECK: %8 = insertelement <3 x half> %7, half %hlsl.cross3, i64 2
+  ; CHECK: ret <3 x half> %8
+  %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: %x0 = extractelement <3 x float> %p0, 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: %y2 = extractelement <3 x float> %p1, i64 2
+  ; CHECK: %0 = fmul float %x1, %y2
+  ; CHECK: %1 = fmul float %x2, %y1
+  ; CHECK: %hlsl.cross1 = fsub float %0, %1
+  ; CHECK: %2 = fmul float %x2, %y0
+  ; CHECK: %3 = fmul float %x0, %y2
+  ; CHECK: %hlsl.cross2 = fsub float %2, %3
+  ; CHECK: %4 = fmul float %x0, %y1
+  ; CHECK: %5 = fmul float %x1, %y0
+  ; CHECK: %hlsl.cross3 = fsub float %4, %5
+  ; CHECK: %6 = insertelement <3 x float> undef, float %hlsl.cross1, i64 0
+  ; CHECK: %7 = insertelement <3 x float> %6, float %hlsl.cross2, i64 1
+  ; CHECK: %8 = insertelement <3 x float> %7, float %hlsl.cross3, i64 2
+  ; CHECK: ret <3 x float> %8
+  %hlsl.cross = call <3 x float> @llvm.dx.cross.v3f32(<3 x float> %p0, <3 x float> %p1)
+  ret <3 x float> %hlsl.cross
+}