[HLSL] Implement WaveReadLaneAt intrinsic

    - create a clang built-in in Builtins.td
    - add semantic checking in SemaHLSL.cpp
    - link the WaveReadLaneAt api in hlsl_intrinsics.h
    - add lowering to spirv backend op GroupNonUniformShuffle
      with Scope = 2 (Group) in SPIRVInstructionSelector.cpp

    - add tests for HLSL intrinsic lowering to spirv intrinsic in
      WaveReadLaneAt.hlsl
    - add tests for sema checks in WaveReadLaneAt-errors.hlsl
    - add spir-v backend tests in WaveReadLaneAt.ll

Author: inbelic

clang/include/clang/Basic/Builtins.td

@@ -4703,6 +4703,12 @@ 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

@@ -18835,6 +18835,23 @@ 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, {}, false, true),
+        ArrayRef{OpExpr, OpIndex}, "hlsl.wave.read.lane.at");
+  }
   case Builtin::BI__builtin_hlsl_elementwise_sign: {
     Value *Op0 = EmitScalarExpr(E->getArg(0));
     llvm::Type *Xty = Op0->getType();

clang/lib/CodeGen/CGHLSLRuntime.h

@@ -87,6 +87,7 @@ 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_read_lane_at)
 
   //===----------------------------------------------------------------------===//
   // End of reserved area for HLSL intrinsic getters.

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -2015,6 +2015,13 @@ _HLSL_AVAILABILITY(shadermodel, 6.0)
 _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_is_first_lane)
 __attribute__((convergent)) bool WaveIsFirstLane();
 
+// \brief Returns the value of the expression for the given lane index within
+// the specified wave.
+template <typename T>
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
+    __attribute__((convergent)) T WaveReadLaneAt(T, int32_t);
+
 //===----------------------------------------------------------------------===//
 // sign builtins
 //===----------------------------------------------------------------------===//

clang/lib/Sema/SemaHLSL.cpp

@@ -1956,6 +1956,26 @@ 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->hasIntegerRepresentation()) {
+      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_elementwise_acos:
   case Builtin::BI__builtin_elementwise_asin:
   case Builtin::BI__builtin_elementwise_atan:

clang/test/CodeGenHLSL/builtins/WaveReadLaneAt.hlsl

@@ -0,0 +1,40 @@
+// 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.read.lane.at.i32([[TY]] %[[#]], i32 %[[#]])
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.read.lane.at.i32([[TY]] %[[#]], i32 %[[#]])
+
+  // CHECK:  ret [[TY]] %[[RET]]
+  return WaveReadLaneAt(expr, idx);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.read.lane.at.i32([[TY]], i32) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.read.lane.at.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.read.lane.at.v4f32([[TY1]] %[[#]], i32 %[[#]])
+  // CHECK-DXIL:  %[[RET1:.*]] = call [[TY1:.*]] @llvm.dx.wave.read.lane.at.v4f32([[TY1]] %[[#]], i32 %[[#]])
+
+  // CHECK:  ret [[TY1]] %[[RET1]]
+  return WaveReadLaneAt(expr, idx);
+}
+
+// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.read.lane.at.v4f32([[TY1]], i32) #[[#attr]]
+// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.read.lane.at.v4f32([[TY1]], i32) #[[#attr]]
+
+// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}}

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

@@ -0,0 +1,21 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -fnative-half-type -emit-llvm-only -disable-llvm-passes -verify -verify-ignore-unexpected
+
+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'}}
+}

llvm/include/llvm/IR/IntrinsicsSPIRV.td

@@ -82,5 +82,6 @@ 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_read_lane_at : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent]>;
   def int_spv_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty]>;
 }

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -2653,6 +2653,21 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
         .addUse(GR.getSPIRVTypeID(ResType))
         .addUse(GR.getOrCreateConstInt(3, I, IntTy, TII));
   }
+  case Intrinsic::spv_wave_read_lane_at: {
+    assert(I.getNumOperands() == 4);
+    assert(I.getOperand(2).isReg());
+    assert(I.getOperand(3).isReg());
+
+    // Defines the execution scope currently 2 for group, see scope table
+    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(2, I, IntTy, TII));
+  }
   case Intrinsic::spv_step:
     return selectStep(ResVReg, ResType, I);
   // Discard intrinsics which we do not expect to actually represent code after

llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveReadLaneAt.ll

@@ -0,0 +1,28 @@
+; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv32v1.3-vulkan-unknown %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-vulkan-unknown %s -o - -filetype=obj | spirv-val %}
+
+; Test lowering to spir-v backend
+
+; CHECK-DAG:   %[[#uint:]] = OpTypeInt 32 0
+; CHECK-DAG:   %[[#scope:]] = OpConstant %[[#uint]] 2
+; CHECK-DAG:   %[[#f32:]] = OpTypeFloat 32
+; CHECK-DAG:   %[[#expr:]] = OpFunctionParameter %[[#f32]]
+; CHECK-DAG:   %[[#idx:]] = OpFunctionParameter %[[#uint]]
+
+define spir_func void @test_1(float %expr, i32 %idx) #0 {
+entry:
+  %0 = call token @llvm.experimental.convergence.entry()
+; CHECK:   %[[#ret:]] = OpGroupNonUniformShuffle %[[#f32]] %[[#expr]] %[[#idx]] %[[#scope]]
+  %1 = call float @llvm.spv.wave.read.lane.at(float %expr, i32 %idx) [ "convergencectrl"(token %0) ]
+  ret void
+}
+
+declare i32 @__hlsl_wave_get_lane_index() #1
+
+attributes #0 = { convergent norecurse "hlsl.numthreads"="1,1,1" "hlsl.shader"="compute" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
+attributes #1 = { convergent }
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 4, !"dx.disable_optimizations", i32 1}