[HLSL] Implement WaveReadLaneAt intrinsic

clang/include/clang/Basic/Builtins.td

@@ -4733,6 +4733,12 @@ def HLSLWaveIsFirstLane : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "bool()";
 }
 
+def HLSLWaveReadLaneAt : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_wave_read_lane_at"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 def HLSLClamp : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_elementwise_clamp"];
   let Attributes = [NoThrow, Const];

clang/lib/CodeGen/CGBuiltin.cpp

@@ -18887,6 +18887,24 @@ case Builtin::BI__builtin_hlsl_elementwise_isinf: {
     Intrinsic::ID ID = CGM.getHLSLRuntime().getWaveIsFirstLaneIntrinsic();
     return EmitRuntimeCall(Intrinsic::getDeclaration(&CGM.getModule(), ID));
   }
+  case Builtin::BI__builtin_hlsl_wave_read_lane_at: {
+    // Due to the use of variadic arguments we must explicitly retreive them and
+    // create our function type.
+    Value *OpExpr = EmitScalarExpr(E->getArg(0));
+    Value *OpIndex = EmitScalarExpr(E->getArg(1));
+    llvm::FunctionType *FT = llvm::FunctionType::get(
+        OpExpr->getType(), ArrayRef{OpExpr->getType(), OpIndex->getType()},
+        false);
+
+    // Get overloaded name
+    std::string Name =
+        Intrinsic::getName(CGM.getHLSLRuntime().getWaveReadLaneAtIntrinsic(),
+                           ArrayRef{OpExpr->getType()}, &CGM.getModule());
+    return EmitRuntimeCall(CGM.CreateRuntimeFunction(FT, Name, {},
+                                                     /*Local=*/false,
+                                                     /*AssumeConvergent=*/true),
+                           ArrayRef{OpExpr, OpIndex}, "hlsl.wave.readlaneAt");
+  }
   case Builtin::BI__builtin_hlsl_elementwise_sign: {
     auto *Arg0 = E->getArg(0);
     Value *Op0 = EmitScalarExpr(Arg0);

clang/lib/CodeGen/CGHLSLRuntime.h

@@ -89,6 +89,7 @@ class CGHLSLRuntime {
   GENERATE_HLSL_INTRINSIC_FUNCTION(SDot, sdot)
   GENERATE_HLSL_INTRINSIC_FUNCTION(UDot, udot)
   GENERATE_HLSL_INTRINSIC_FUNCTION(WaveIsFirstLane, wave_is_first_lane)
+  GENERATE_HLSL_INTRINSIC_FUNCTION(WaveReadLaneAt, wave_readlaneat)
 
   //===----------------------------------------------------------------------===//
   // End of reserved area for HLSL intrinsic getters.

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -2067,6 +2067,86 @@ _HLSL_AVAILABILITY(shadermodel, 6.0)
 _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_is_first_lane)
 __attribute__((convergent)) bool WaveIsFirstLane();
 
+//===----------------------------------------------------------------------===//
+// WaveReadLaneAt builtins
+//===----------------------------------------------------------------------===//
+
+// \brief Returns the value of the expression for the given lane index within
+// the specified wave.
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) bool WaveReadLaneAt(bool, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) bool2 WaveReadLaneAt(bool2, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) bool3 WaveReadLaneAt(bool3, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) bool4 WaveReadLaneAt(bool4, int32_t);
+
+#ifdef __HLSL_ENABLE_16_BIT
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int16_t WaveReadLaneAt(int16_t, int32_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int16_t2 WaveReadLaneAt(int16_t2, int32_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int16_t3 WaveReadLaneAt(int16_t3, int32_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int16_t4 WaveReadLaneAt(int16_t4, int32_t);
+#endif
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) half WaveReadLaneAt(half, int32_t);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) half2 WaveReadLaneAt(half2, int32_t);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) half3 WaveReadLaneAt(half3, int32_t);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) half4 WaveReadLaneAt(half4, int32_t);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int WaveReadLaneAt(int, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int2 WaveReadLaneAt(int2, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int3 WaveReadLaneAt(int3, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int4 WaveReadLaneAt(int4, int32_t);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) float WaveReadLaneAt(float, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) float2 WaveReadLaneAt(float2, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) float3 WaveReadLaneAt(float3, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) float4 WaveReadLaneAt(float4, int32_t);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int64_t WaveReadLaneAt(int64_t, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int64_t2 WaveReadLaneAt(int64_t2, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int64_t3 WaveReadLaneAt(int64_t3, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) int64_t4 WaveReadLaneAt(int64_t4, int32_t);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) double WaveReadLaneAt(double, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) double2 WaveReadLaneAt(double2, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) double3 WaveReadLaneAt(double3, int32_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+__attribute__((convergent)) double4 WaveReadLaneAt(double4, int32_t);
+
 //===----------------------------------------------------------------------===//
 // sign builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaHLSL.cpp

@@ -1992,6 +1992,26 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
+  case Builtin::BI__builtin_hlsl_wave_read_lane_at: {
+    if (SemaRef.checkArgCount(TheCall, 2))
+      return true;
+
+    // Ensure index parameter type can be interpreted as a uint
+    ExprResult Index = TheCall->getArg(1);
+    QualType ArgTyIndex = Index.get()->getType();
+    if (!ArgTyIndex->isIntegerType()) {
+      SemaRef.Diag(TheCall->getArg(1)->getBeginLoc(),
+                   diag::err_typecheck_convert_incompatible)
+          << ArgTyIndex << SemaRef.Context.UnsignedIntTy << 1 << 0 << 0;
+      return true;
+    }
+
+    // Ensure return type is the same as the input expr type
+    ExprResult Expr = TheCall->getArg(0);
+    QualType ArgTyExpr = Expr.get()->getType();
+    TheCall->setType(ArgTyExpr);
+    break;
+  }
   case Builtin::BI__builtin_hlsl_wave_get_lane_index: {
     if (SemaRef.checkArgCount(TheCall, 0))
       return true;

clang/test/CodeGenHLSL/builtins/WaveReadLaneAt.hlsl

@@ -0,0 +1,36 @@
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -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 -std=hlsl2021 -finclude-default-header -x hlsl -triple   \
+// RUN:   spirv-pc-vulkan-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-SPIRV
+
+// Test basic lowering to runtime function call.
+
+// CHECK-LABEL: test_int
+int test_int(int expr, uint idx) {
+  // CHECK-SPIRV: %[[#entry_tok:]] = call token @llvm.experimental.convergence.entry()
+  // CHECK-SPIRV:  %[[RET:.*]] = call [[TY:.*]] @llvm.spv.wave.readlaneat.i32([[TY]] %[[#]], i32 %[[#]]) [ "convergencectrl"(token %[[#entry_tok]]) ]
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.readlaneat.i32([[TY]] %[[#]], i32 %[[#]])
+  // CHECK:  ret [[TY]] %[[RET]]
+  return WaveReadLaneAt(expr, idx);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.readlaneat.i32([[TY]], i32) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.readlaneat.i32([[TY]], i32) #[[#attr:]]
+
+// Test basic lowering to runtime function call with array and float value.
+
+// CHECK-LABEL: test_floatv4
+float4 test_floatv4(float4 expr, uint idx) {
+  // CHECK-SPIRV: %[[#entry_tok1:]] = call token @llvm.experimental.convergence.entry()
+  // CHECK-SPIRV:  %[[RET1:.*]] = call [[TY1:.*]] @llvm.spv.wave.readlaneat.v4f32([[TY1]] %[[#]], i32 %[[#]]) [ "convergencectrl"(token %[[#entry_tok1]]) ]
+  // CHECK-DXIL:  %[[RET1:.*]] = call [[TY1:.*]] @llvm.dx.wave.readlaneat.v4f32([[TY1]] %[[#]], i32 %[[#]])
+  // CHECK:  ret [[TY1]] %[[RET1]]
+  return WaveReadLaneAt(expr, idx);
+}
+
+// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.readlaneat.v4f32([[TY1]], i32) #[[#attr]]
+// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.readlaneat.v4f32([[TY1]], i32) #[[#attr]]
+
+// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}}

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

@@ -0,0 +1,26 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -emit-llvm-only -disable-llvm-passes -verify
+
+bool test_too_few_arg() {
+  return __builtin_hlsl_wave_read_lane_at();
+  // expected-error@-1 {{too few arguments to function call, expected 2, have 0}}
+}
+
+float2 test_too_few_arg_1(float2 p0) {
+  return __builtin_hlsl_wave_read_lane_at(p0);
+  // expected-error@-1 {{too few arguments to function call, expected 2, have 1}}
+}
+
+float2 test_too_many_arg(float2 p0) {
+  return __builtin_hlsl_wave_read_lane_at(p0, p0, p0);
+  // expected-error@-1 {{too many arguments to function call, expected 2, have 3}}
+}
+
+float3 test_index_type_check(float3 p0, double idx) {
+  return __builtin_hlsl_wave_read_lane_at(p0, idx);
+  // expected-error@-1 {{passing 'double' to parameter of incompatible type 'unsigned int'}}
+}
+
+float3 test_index_type_check(float3 p0, int3 idxs) {
+  return __builtin_hlsl_wave_read_lane_at(p0, idxs);
+  // expected-error@-1 {{passing 'int3' (aka 'vector<int, 3>') to parameter of incompatible type 'unsigned int'}}
+}

llvm/include/llvm/IR/IntrinsicsDirectX.td

@@ -85,6 +85,7 @@ def int_dx_normalize : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_
 def int_dx_rsqrt  : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_wave_getlaneindex : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent, IntrNoMem]>;
 def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrConvergent]>;
+def int_dx_wave_readlaneat : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>;
 def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_step : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;

llvm/include/llvm/IR/IntrinsicsSPIRV.td

@@ -83,6 +83,7 @@ let TargetPrefix = "spv" in {
     [llvm_anyint_ty, LLVMScalarOrSameVectorWidth<0, LLVMVectorElementType<0>>],
     [IntrNoMem, Commutative] >;
   def int_spv_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrConvergent]>;
+  def int_spv_wave_readlaneat : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>;
   def int_spv_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>;
   def int_spv_radians : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty], [IntrNoMem]>;
 

llvm/lib/Target/DirectX/DXIL.td

@@ -802,6 +802,16 @@ def WaveIsFirstLane :  DXILOp<110, waveIsFirstLane> {
   let attributes = [Attributes<DXIL1_0, [ReadNone]>];
 }
 
+def WaveReadLaneAt:  DXILOp<117, waveReadLaneAt> {
+  let Doc = "returns the value from the specified lane";
+  let LLVMIntrinsic = int_dx_wave_readlaneat;
+  let arguments = [OverloadTy, Int32Ty];
+  let result = OverloadTy;
+  let overloads = [Overloads<DXIL1_0, [HalfTy, FloatTy, DoubleTy, Int1Ty, Int16Ty, Int32Ty]>];
+    let stages = [Stages<DXIL1_0, [all_stages]>];
+  let attributes = [Attributes<DXIL1_0, [ReadNone]>];
+}
+
 def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
   let Doc = "returns the index of the current lane in the wave";
   let LLVMIntrinsic = int_dx_wave_getlaneindex;

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -229,6 +229,9 @@ class SPIRVInstructionSelector : public InstructionSelector {
 
   bool selectSpvThreadId(Register ResVReg, const SPIRVType *ResType,
                          MachineInstr &I) const;
+
+  bool selectWaveReadLaneAt(Register ResVReg, const SPIRVType *ResType,
+                            MachineInstr &I) const;
 
   bool selectUnmergeValues(MachineInstr &I) const;
 
@@ -417,6 +420,7 @@ bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
 
   case TargetOpcode::G_INTRINSIC:
   case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
+  case TargetOpcode::G_INTRINSIC_CONVERGENT:
   case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS:
     return selectIntrinsic(ResVReg, ResType, I);
   case TargetOpcode::G_BITREVERSE:
@@ -1758,6 +1762,26 @@ bool SPIRVInstructionSelector::selectSign(Register ResVReg,
   return Result;
 }
 
+bool SPIRVInstructionSelector::selectWaveReadLaneAt(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();
+
+  // IntTy is used to define the execution scope, set to 3 to denote a
+  // cross-lane interaction equivalent to a SPIR-V subgroup.
+  SPIRVType *IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
+  return BuildMI(BB, I, I.getDebugLoc(),
+                 TII.get(SPIRV::OpGroupNonUniformShuffle))
+      .addDef(ResVReg)
+      .addUse(GR.getSPIRVTypeID(ResType))
+      .addUse(I.getOperand(2).getReg())
+      .addUse(I.getOperand(3).getReg())
+      .addUse(GR.getOrCreateConstInt(3, I, IntTy, TII));
+}
+
 bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg,
                                                 const SPIRVType *ResType,
                                                 MachineInstr &I) const {
@@ -2541,6 +2565,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
         .addUse(GR.getSPIRVTypeID(ResType))
         .addUse(GR.getOrCreateConstInt(3, I, IntTy, TII));
   }
+  case Intrinsic::spv_wave_readlaneat:
+    return selectWaveReadLaneAt(ResVReg, ResType, I);
   case Intrinsic::spv_step:
     return selectExtInst(ResVReg, ResType, I, CL::step, GL::Step);
   case Intrinsic::spv_radians:

llvm/test/CodeGen/DirectX/WaveReadLaneAt.ll

@@ -0,0 +1,53 @@
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-compute %s | FileCheck %s
+
+; Test that for scalar values, WaveReadLaneAt maps down to the DirectX op
+
+define noundef half @wave_rla_half(half noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call half @dx.op.waveReadLaneAt.f16(i32 117, half %expr, i32 %idx)
+  %ret = call half @llvm.dx.wave.readlaneat.f16(half %expr, i32 %idx)
+  ret half %ret
+}
+
+define noundef float @wave_rla_float(float noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call float @dx.op.waveReadLaneAt.f32(i32 117, float %expr, i32 %idx)
+  %ret = call float @llvm.dx.wave.readlaneat(float %expr, i32 %idx)
+  ret float %ret
+}
+
+define noundef double @wave_rla_double(double noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call double @dx.op.waveReadLaneAt.f64(i32 117, double %expr, i32 %idx)
+  %ret = call double @llvm.dx.wave.readlaneat(double %expr, i32 %idx)
+  ret double %ret
+}
+
+define noundef i1 @wave_rla_i1(i1 noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call i1 @dx.op.waveReadLaneAt.i1(i32 117, i1 %expr, i32 %idx)
+  %ret = call i1 @llvm.dx.wave.readlaneat.i1(i1 %expr, i32 %idx)
+  ret i1 %ret
+}
+
+define noundef i16 @wave_rla_i16(i16 noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call i16 @dx.op.waveReadLaneAt.i16(i32 117, i16 %expr, i32 %idx)
+  %ret = call i16 @llvm.dx.wave.readlaneat.i16(i16 %expr, i32 %idx)
+  ret i16 %ret
+}
+
+define noundef i32 @wave_rla_i32(i32 noundef %expr, i32 noundef %idx) {
+entry:
+; CHECK: call i32 @dx.op.waveReadLaneAt.i32(i32 117, i32 %expr, i32 %idx)
+  %ret = call i32 @llvm.dx.wave.readlaneat.i32(i32 %expr, i32 %idx)
+  ret i32 %ret
+}
+
+declare half @llvm.dx.wave.readlaneat.f16(half, i32)
+declare float @llvm.dx.wave.readlaneat.f32(float, i32)
+declare double @llvm.dx.wave.readlaneat.f64(double, i32)
+
+declare i1 @llvm.dx.wave.readlaneat.i1(i1, i32)
+declare i16 @llvm.dx.wave.readlaneat.i16(i16, i32)
+declare i32 @llvm.dx.wave.readlaneat.i32(i32, i32)