[HLSL][DXIL] Implement asdouble intrinsic (llvm#114847)

- define intrinsic as builtin in Builtins.td
- link intrinsic in hlsl_intrinsics.h
- add semantic analysis to SemaHLSL.cpp
- lower to `llvm` or a `dx` intrinsic when applicable in CGBuiltin.cpp
- define DXIL intrinsic in IntrinsicsDirectX.td
- add DXIL op and mapping in DXIL.td
- enable scalarization of intrinsic

- add basic sema checking to asdouble-errors.hlsl
    
 Resolves llvm#99081

Author: inbelic

clang/include/clang/Basic/Builtins.td

@@ -4750,6 +4750,12 @@ def HLSLAny : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "bool(...)";
 }
 
+def HLSLAsDouble : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_asdouble"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 def HLSLWaveActiveAnyTrue : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_wave_active_any_true"];
   let Attributes = [NoThrow, Const];

clang/lib/CodeGen/CGBuiltin.cpp

@@ -209,6 +209,41 @@ static Value *handleHlslSplitdouble(const CallExpr *E, CodeGenFunction *CGF) {
   return LastInst;
 }
 
+Value *handleAsDoubleBuiltin(CodeGenFunction &CGF, const CallExpr *E) {
+  assert((E->getArg(0)->getType()->hasUnsignedIntegerRepresentation() &&
+          E->getArg(1)->getType()->hasUnsignedIntegerRepresentation()) &&
+         "asdouble operands types mismatch");
+  Value *OpLowBits = CGF.EmitScalarExpr(E->getArg(0));
+  Value *OpHighBits = CGF.EmitScalarExpr(E->getArg(1));
+
+  llvm::Type *ResultType = CGF.DoubleTy;
+  int N = 1;
+  if (auto *VTy = E->getArg(0)->getType()->getAs<clang::VectorType>()) {
+    N = VTy->getNumElements();
+    ResultType = llvm::FixedVectorType::get(CGF.DoubleTy, N);
+  }
+
+  if (CGF.CGM.getTarget().getTriple().isDXIL())
+    return CGF.Builder.CreateIntrinsic(
+        /*ReturnType=*/ResultType, Intrinsic::dx_asdouble,
+        ArrayRef<Value *>{OpLowBits, OpHighBits}, nullptr, "hlsl.asdouble");
+
+  if (!E->getArg(0)->getType()->isVectorType()) {
+    OpLowBits = CGF.Builder.CreateVectorSplat(1, OpLowBits);
+    OpHighBits = CGF.Builder.CreateVectorSplat(1, OpHighBits);
+  }
+
+  llvm::SmallVector<int> Mask;
+  for (int i = 0; i < N; i++) {
+    Mask.push_back(i);
+    Mask.push_back(i + N);
+  }
+
+  Value *BitVec = CGF.Builder.CreateShuffleVector(OpLowBits, OpHighBits, Mask);
+
+  return CGF.Builder.CreateBitCast(BitVec, ResultType);
+}
+
 /// getBuiltinLibFunction - Given a builtin id for a function like
 /// "__builtin_fabsf", return a Function* for "fabsf".
 llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
@@ -19023,6 +19058,8 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
         CGM.getHLSLRuntime().getAnyIntrinsic(), ArrayRef<Value *>{Op0}, nullptr,
         "hlsl.any");
   }
+  case Builtin::BI__builtin_hlsl_asdouble:
+    return handleAsDoubleBuiltin(*this, E);
   case Builtin::BI__builtin_hlsl_elementwise_clamp: {
     Value *OpX = EmitScalarExpr(E->getArg(0));
     Value *OpMin = EmitScalarExpr(E->getArg(1));

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -361,6 +361,24 @@ bool any(double3);
 _HLSL_BUILTIN_ALIAS(__builtin_hlsl_any)
 bool any(double4);
 
+//===----------------------------------------------------------------------===//
+// asdouble builtins
+//===----------------------------------------------------------------------===//
+
+/// \fn double asdouble(uint LowBits, uint HighBits)
+/// \brief Reinterprets a cast value (two 32-bit values) into a double.
+/// \param LowBits The low 32-bit pattern of the input value.
+/// \param HighBits The high 32-bit pattern of the input value.
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_asdouble)
+double asdouble(uint, uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_asdouble)
+double2 asdouble(uint2, uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_asdouble)
+double3 asdouble(uint3, uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_asdouble)
+double4 asdouble(uint4, uint4);
+
 //===----------------------------------------------------------------------===//
 // asfloat builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaHLSL.cpp

@@ -1888,6 +1888,15 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
+  case Builtin::BI__builtin_hlsl_asdouble: {
+    if (SemaRef.checkArgCount(TheCall, 2))
+      return true;
+    if (CheckUnsignedIntRepresentation(&SemaRef, TheCall))
+      return true;
+
+    SetElementTypeAsReturnType(&SemaRef, TheCall, getASTContext().DoubleTy);
+    break;
+  }
   case Builtin::BI__builtin_hlsl_elementwise_clamp: {
     if (SemaRef.checkArgCount(TheCall, 3))
       return true;

clang/test/CodeGenHLSL/builtins/asdouble.hlsl

@@ -0,0 +1,37 @@
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.3-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-DXIL
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN:   spirv-pc-vulkan-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-SPV
+
+// Test lowering of asdouble expansion to shuffle/bitcast and splat when required
+
+// CHECK-LABEL: test_uint
+double test_uint(uint low, uint high) {
+  // CHECK-SPV: %[[LOW_INSERT:.*]] = insertelement <1 x i32>
+  // CHECK-SPV: %[[LOW_SHUFFLE:.*]] = shufflevector <1 x i32> %[[LOW_INSERT]], {{.*}} zeroinitializer
+  // CHECK-SPV: %[[HIGH_INSERT:.*]] = insertelement <1 x i32>
+  // CHECK-SPV: %[[HIGH_SHUFFLE:.*]] = shufflevector <1 x i32> %[[HIGH_INSERT]], {{.*}} zeroinitializer
+
+  // CHECK-SPV:      %[[SHUFFLE0:.*]] = shufflevector <1 x i32> %[[LOW_SHUFFLE]], <1 x i32> %[[HIGH_SHUFFLE]],
+  // CHECK-SPV-SAME: {{.*}} <i32 0, i32 1>
+  // CHECK-SPV:      bitcast <2 x i32> %[[SHUFFLE0]] to double
+
+  // CHECK-DXIL: call double @llvm.dx.asdouble.i32
+  return asdouble(low, high);
+}
+
+// CHECK-DXIL: declare double @llvm.dx.asdouble.i32
+
+// CHECK-LABEL: test_vuint
+double3 test_vuint(uint3 low, uint3 high) {
+  // CHECK-SPV:      %[[SHUFFLE1:.*]] = shufflevector
+  // CHECK-SPV-SAME: {{.*}} <i32 0, i32 3, i32 1, i32 4, i32 2, i32 5>
+  // CHECK-SPV:      bitcast <6 x i32> %[[SHUFFLE1]] to <3 x double>
+
+  // CHECK-DXIL: call <3 x double> @llvm.dx.asdouble.v3i32
+  return asdouble(low, high);
+}
+
+// CHECK-DXIL: declare <3 x double> @llvm.dx.asdouble.v3i32

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

@@ -0,0 +1,16 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -emit-llvm-only -disable-llvm-passes -verify
+
+double test_too_few_arg() {
+  return __builtin_hlsl_asdouble();
+  // expected-error@-1 {{too few arguments to function call, expected 2, have 0}}
+}
+
+double test_too_few_arg_1(uint p0) {
+  return __builtin_hlsl_asdouble(p0);
+  // expected-error@-1 {{too few arguments to function call, expected 2, have 1}}
+}
+
+double test_too_many_arg(uint p0) {
+  return __builtin_hlsl_asdouble(p0, p0, p0);
+  // expected-error@-1 {{too many arguments to function call, expected 2, have 3}}
+}

llvm/include/llvm/IR/IntrinsicsDirectX.td

@@ -46,6 +46,7 @@ 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_asdouble : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [llvm_anyint_ty, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_uclamp : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_sclamp : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_nclamp : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;

llvm/lib/Target/DirectX/DXIL.td

@@ -818,6 +818,15 @@ def FlattenedThreadIdInGroup :  DXILOp<96, flattenedThreadIdInGroup> {
   let attributes = [Attributes<DXIL1_0, [ReadNone]>];
 }
 
+def MakeDouble :  DXILOp<101, makeDouble> {
+  let Doc = "creates a double value";
+  let LLVMIntrinsic = int_dx_asdouble;
+  let arguments = [Int32Ty, Int32Ty];
+  let result = DoubleTy;
+  let stages = [Stages<DXIL1_0, [all_stages]>];
+  let attributes = [Attributes<DXIL1_0, [ReadNone]>];
+}
+
 def SplitDouble :  DXILOp<102, splitDouble> {
   let Doc = "Splits a double into 2 uints";
   let arguments = [OverloadTy];

llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp

@@ -28,6 +28,8 @@ bool DirectXTTIImpl::isTargetIntrinsicWithScalarOpAtArg(Intrinsic::ID ID,
 bool DirectXTTIImpl::isVectorIntrinsicWithOverloadTypeAtArg(Intrinsic::ID ID,
                                                             int ScalarOpdIdx) {
   switch (ID) {
+  case Intrinsic::dx_asdouble:
+    return ScalarOpdIdx == 0;
   default:
     return ScalarOpdIdx == -1;
   }
@@ -39,6 +41,7 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable(
   case Intrinsic::dx_frac:
   case Intrinsic::dx_rsqrt:
   case Intrinsic::dx_wave_readlane:
+  case Intrinsic::dx_asdouble:
   case Intrinsic::dx_splitdouble:
   case Intrinsic::dx_firstbituhigh:
   case Intrinsic::dx_firstbitshigh:

llvm/test/CodeGen/DirectX/asdouble.ll

@@ -0,0 +1,22 @@
+; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+
+; Test that for scalar and vector inputs, asdouble maps down to the makeDouble
+; DirectX op
+
+define noundef double @asdouble_scalar(i32 noundef %low, i32 noundef %high) {
+; CHECK: call double @dx.op.makeDouble(i32 101, i32 %low, i32 %high)
+  %ret = call double @llvm.dx.asdouble.i32(i32 %low, i32 %high)
+  ret double %ret
+}
+
+declare double @llvm.dx.asdouble.i32(i32, i32)
+
+define noundef <3 x double> @asdouble_vec(<3 x i32> noundef %low, <3 x i32> noundef %high) {
+; CHECK: call double @dx.op.makeDouble(i32 101, i32 %low.i0, i32 %high.i0)
+; CHECK: call double @dx.op.makeDouble(i32 101, i32 %low.i1, i32 %high.i1)
+; CHECK: call double @dx.op.makeDouble(i32 101, i32 %low.i2, i32 %high.i2)
+  %ret = call <3 x double> @llvm.dx.asdouble.v3i32(<3 x i32> %low, <3 x i32> %high)
+  ret <3 x double> %ret
+}
+
+declare <3 x double> @llvm.dx.asdouble.v3i32(<3 x i32>, <3 x i32>)