[HLSL] implement the rcp intrinsic

clang/include/clang/Basic/Builtins.td

@@ -4584,6 +4584,12 @@ def HLSLMad : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void(...)";
 }
 
+def HLSLRcp : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_elementwise_rcp"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 // Builtins for XRay.
 def XRayCustomEvent : Builtin {
   let Spellings = ["__xray_customevent"];

clang/lib/CodeGen/CGBuiltin.cpp

@@ -17966,7 +17966,7 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
   case Builtin::BI__builtin_hlsl_elementwise_any: {
     Value *Op0 = EmitScalarExpr(E->getArg(0));
     return Builder.CreateIntrinsic(
-        /*ReturnType*/ llvm::Type::getInt1Ty(getLLVMContext()),
+        /*ReturnType=*/llvm::Type::getInt1Ty(getLLVMContext()),
         Intrinsic::dx_any, ArrayRef<Value *>{Op0}, nullptr, "dx.any");
   }
   case Builtin::BI__builtin_hlsl_dot: {
@@ -18002,7 +18002,7 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
            "Dot product requires vectors to be of the same size.");
 
     return Builder.CreateIntrinsic(
-        /*ReturnType*/ T0->getScalarType(), Intrinsic::dx_dot,
+        /*ReturnType=*/T0->getScalarType(), Intrinsic::dx_dot,
         ArrayRef<Value *>{Op0, Op1}, nullptr, "dx.dot");
   } break;
   case Builtin::BI__builtin_hlsl_lerp: {
@@ -18039,15 +18039,15 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
            XVecTy->getElementType() == SVecTy->getElementType() &&
            "Lerp requires float vectors to be of the same type.");
     return Builder.CreateIntrinsic(
-        /*ReturnType*/ Xty, Intrinsic::dx_lerp, ArrayRef<Value *>{X, Y, S},
+        /*ReturnType=*/Xty, Intrinsic::dx_lerp, ArrayRef<Value *>{X, Y, S},
         nullptr, "dx.lerp");
   }
   case Builtin::BI__builtin_hlsl_elementwise_frac: {
     Value *Op0 = EmitScalarExpr(E->getArg(0));
     if (!E->getArg(0)->getType()->hasFloatingRepresentation())
       llvm_unreachable("frac operand must have a float representation");
     return Builder.CreateIntrinsic(
-        /*ReturnType*/ Op0->getType(), Intrinsic::dx_frac,
+        /*ReturnType=*/Op0->getType(), Intrinsic::dx_frac,
         ArrayRef<Value *>{Op0}, nullptr, "dx.frac");
   }
   case Builtin::BI__builtin_hlsl_mad: {
@@ -18066,9 +18066,17 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
     }
     assert(E->getArg(0)->getType()->hasUnsignedIntegerRepresentation());
     return Builder.CreateIntrinsic(
-        /*ReturnType*/ M->getType(), Intrinsic::dx_umad,
+        /*ReturnType=*/M->getType(), Intrinsic::dx_umad,
         ArrayRef<Value *>{M, A, B}, nullptr, "dx.umad");
   }
+  case Builtin::BI__builtin_hlsl_elementwise_rcp: {
+    Value *Op0 = EmitScalarExpr(E->getArg(0));
+    if (!E->getArg(0)->getType()->hasFloatingRepresentation())
+      llvm_unreachable("rcp operand must have a float representation");
+    return Builder.CreateIntrinsic(
+        /*ReturnType=*/Op0->getType(), Intrinsic::dx_rcp,
+        ArrayRef<Value *>{Op0}, nullptr, "dx.rcp");
+  }
   }
   return nullptr;
 }

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -1112,6 +1112,47 @@ uint64_t3 reversebits(uint64_t3);
 _HLSL_BUILTIN_ALIAS(__builtin_elementwise_bitreverse)
 uint64_t4 reversebits(uint64_t4);
 
+//===----------------------------------------------------------------------===//
+// rcp builtins
+//===----------------------------------------------------------------------===//
+
+/// \fn T rcp(T x)
+/// \brief Calculates a fast, approximate, per-component reciprocal ie 1 / \a x.
+/// \param x The specified input value.
+///
+/// The return value is the reciprocal of the \a x parameter.
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+half rcp(half);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+half2 rcp(half2);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+half3 rcp(half3);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+half4 rcp(half4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+float rcp(float);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+float2 rcp(float2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+float3 rcp(float3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+float4 rcp(float4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+double rcp(double);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+double2 rcp(double2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+double3 rcp(double3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_rcp)
+double4 rcp(double4);
+
 //===----------------------------------------------------------------------===//
 // round builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaChecking.cpp

@@ -5285,6 +5285,7 @@ bool Sema::CheckHLSLBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
+  case Builtin::BI__builtin_hlsl_elementwise_rcp:
   case Builtin::BI__builtin_hlsl_elementwise_frac: {
     if (PrepareBuiltinElementwiseMathOneArgCall(TheCall))
       return true;

clang/test/CodeGenHLSL/builtins/rcp.hlsl

@@ -0,0 +1,53 @@
+// 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: %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
+
+// NATIVE_HALF: define noundef half @
+// NATIVE_HALF: %dx.rcp = call half @llvm.dx.rcp.f16(
+// NATIVE_HALF: ret half %dx.rcp
+// NO_HALF: define noundef float @"?test_rcp_half@@YA$halff@$halff@@Z"(
+// NO_HALF: %dx.rcp = call float @llvm.dx.rcp.f32(
+// NO_HALF: ret float %dx.rcp
+half test_rcp_half(half p0) { return rcp(p0); }
+// NATIVE_HALF: define noundef <2 x half> @
+// NATIVE_HALF: %dx.rcp = call <2 x half> @llvm.dx.rcp.v2f16
+// NATIVE_HALF: ret <2 x half> %dx.rcp
+// NO_HALF: define noundef <2 x float> @
+// NO_HALF: %dx.rcp = call <2 x float> @llvm.dx.rcp.v2f32(
+// NO_HALF: ret <2 x float> %dx.rcp
+half2 test_rcp_half2(half2 p0) { return rcp(p0); }
+// NATIVE_HALF: define noundef <3 x half> @
+// NATIVE_HALF: %dx.rcp = call <3 x half> @llvm.dx.rcp.v3f16
+// NATIVE_HALF: ret <3 x half> %dx.rcp
+// NO_HALF: define noundef <3 x float> @
+// NO_HALF: %dx.rcp = call <3 x float> @llvm.dx.rcp.v3f32(
+// NO_HALF: ret <3 x float> %dx.rcp
+half3 test_rcp_half3(half3 p0) { return rcp(p0); }
+// NATIVE_HALF: define noundef <4 x half> @
+// NATIVE_HALF: %dx.rcp = call <4 x half> @llvm.dx.rcp.v4f16
+// NATIVE_HALF: ret <4 x half> %dx.rcp
+// NO_HALF: define noundef <4 x float> @
+// NO_HALF: %dx.rcp = call <4 x float> @llvm.dx.rcp.v4f32(
+// NO_HALF: ret <4 x float> %dx.rcp
+half4 test_rcp_half4(half4 p0) { return rcp(p0); }
+
+// CHECK: define noundef float @
+// CHECK: %dx.rcp = call float @llvm.dx.rcp.f32(
+// CHECK: ret float %dx.rcp
+float test_rcp_float(float p0) { return rcp(p0); }
+// CHECK: define noundef <2 x float> @
+// CHECK: %dx.rcp = call <2 x float> @llvm.dx.rcp.v2f32
+// CHECK: ret <2 x float> %dx.rcp
+float2 test_rcp_float2(float2 p0) { return rcp(p0); }
+// CHECK: define noundef <3 x float> @
+// CHECK: %dx.rcp = call <3 x float> @llvm.dx.rcp.v3f32
+// CHECK: ret <3 x float> %dx.rcp
+float3 test_rcp_float3(float3 p0) { return rcp(p0); }
+// CHECK: define noundef <4 x float> @
+// CHECK: %dx.rcp = call <4 x float> @llvm.dx.rcp.v4f32
+// CHECK: ret <4 x float> %dx.rcp
+float4 test_rcp_float4(float4 p0) { return rcp(p0); }

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

@@ -0,0 +1,27 @@
+
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -fnative-half-type -emit-llvm -disable-llvm-passes -verify -verify-ignore-unexpected
+
+float test_too_few_arg() {
+  return __builtin_hlsl_elementwise_rcp();
+  // expected-error@-1 {{too few arguments to function call, expected 1, have 0}}
+}
+
+float2 test_too_many_arg(float2 p0) {
+  return __builtin_hlsl_elementwise_rcp(p0, p0);
+  // expected-error@-1 {{too many arguments to function call, expected 1, have 2}}
+}
+
+float builtin_bool_to_float_type_promotion(bool p1) {
+  return __builtin_hlsl_elementwise_rcp(p1);
+  // expected-error@-1 {{1st argument must be a vector, integer or floating point type (was 'bool')}}
+}
+
+float builtin_rcp_int_to_float_promotion(int p1) {
+  return __builtin_hlsl_elementwise_rcp(p1);
+  // expected-error@-1 {{passing 'int' to parameter of incompatible type 'float'}}
+}
+
+float2 builtin_rcp_int2_to_float2_promotion(int2 p1) {
+  return __builtin_hlsl_elementwise_rcp(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)}}
+}

llvm/include/llvm/IR/IntrinsicsDirectX.td

@@ -34,6 +34,7 @@ def int_dx_lerp :
     [llvm_anyvector_ty, LLVMScalarOrSameVectorWidth<0, LLVMVectorElementType<0>>,LLVMScalarOrSameVectorWidth<0, LLVMVectorElementType<0>>],
     [IntrNoMem, IntrWillReturn] >;
 
-def int_dx_imad  : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>]>;
-def int_dx_umad  : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>]>;
+def int_dx_imad : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>]>;
+def int_dx_umad : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>]>;
+def int_dx_rcp  : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
 }