https://github.com/kcloudy0717 created 
https://github.com/llvm/llvm-project/pull/179423

This PR adds WavePrefixProduct intrinsic support in HLSL with codegen for both 
DirectX and SPIRV backends

>From 9c51ae0c966b1e331d0303ae9fb3ffdeaabc09f6 Mon Sep 17 00:00:00 2001
From: Kai Huang <[email protected]>
Date: Sat, 6 Dec 2025 22:28:19 +0800
Subject: [PATCH] [HLSL][DXIL][SPIRV] WavePrefixProduct intrinsic support

This PR adds WavePrefixProduct intrinsic support in HLSL with codegen for both 
DirectX and SPIRV backends
---
 clang/include/clang/Basic/Builtins.td         |   6 +
 clang/lib/CodeGen/CGHLSLBuiltins.cpp          |  27 ++++
 .../lib/Headers/hlsl/hlsl_alias_intrinsics.h  |  99 ++++++++++++
 clang/lib/Sema/SemaHLSL.cpp                   |   3 +-
 .../builtins/WavePrefixProduct.hlsl           |  46 ++++++
 .../BuiltIns/WavePrefixProduct-errors.hlsl    |  28 ++++
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |   2 +
 llvm/include/llvm/IR/IntrinsicsSPIRV.td       |   1 +
 llvm/lib/Target/DirectX/DXIL.td               |  10 ++
 llvm/lib/Target/DirectX/DXILShaderFlags.cpp   |   2 +
 .../DirectX/DirectXTargetTransformInfo.cpp    |   2 +
 .../Target/SPIRV/SPIRVInstructionSelector.cpp |  18 +++
 .../CodeGen/DirectX/ShaderFlags/wave-ops.ll   |  14 ++
 .../test/CodeGen/DirectX/WavePrefixProduct.ll | 143 ++++++++++++++++++
 .../hlsl-intrinsics/WavePrefixProduct.ll      |  41 +++++
 15 files changed, 441 insertions(+), 1 deletion(-)
 create mode 100644 clang/test/CodeGenHLSL/builtins/WavePrefixProduct.hlsl
 create mode 100644 clang/test/SemaHLSL/BuiltIns/WavePrefixProduct-errors.hlsl
 create mode 100644 llvm/test/CodeGen/DirectX/WavePrefixProduct.ll
 create mode 100644 llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixProduct.ll

diff --git a/clang/include/clang/Basic/Builtins.td 
b/clang/include/clang/Basic/Builtins.td
index 69699d1f74463..d06338b048bb5 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -5156,6 +5156,12 @@ def HLSLWavePrefixSum : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void(...)";
 }
 
+def HLSLWavePrefixProduct : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_wave_prefix_product"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 def HLSLClamp : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_elementwise_clamp"];
   let Attributes = [NoThrow, Const, CustomTypeChecking];
diff --git a/clang/lib/CodeGen/CGHLSLBuiltins.cpp 
b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
index 02e4852a2a83e..a18436648f2be 100644
--- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp
+++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
@@ -376,6 +376,24 @@ static Intrinsic::ID 
getWavePrefixSumIntrinsic(llvm::Triple::ArchType Arch,
   }
 }
 
+// Return wave prefix product that corresponds to the QT scalar type
+static Intrinsic::ID getWavePrefixProductIntrinsic(llvm::Triple::ArchType Arch,
+                                                   CGHLSLRuntime &RT,
+                                                   QualType QT) {
+  switch (Arch) {
+  case llvm::Triple::spirv:
+    return Intrinsic::spv_wave_prefix_product;
+  case llvm::Triple::dxil: {
+    if (QT->isUnsignedIntegerType())
+      return Intrinsic::dx_wave_prefix_uproduct;
+    return Intrinsic::dx_wave_prefix_product;
+  }
+  default:
+    llvm_unreachable("Intrinsic WavePrefixProduct"
+                     " not supported by target architecture");
+  }
+}
+
 // Returns the mangled name for a builtin function that the SPIR-V backend
 // will expand into a spec Constant.
 static std::string getSpecConstantFunctionName(clang::QualType 
SpecConstantType,
@@ -1064,6 +1082,15 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned 
BuiltinID,
                                &CGM.getModule(), IID, {OpExpr->getType()}),
                            ArrayRef{OpExpr}, "hlsl.wave.prefix.sum");
   }
+  case Builtin::BI__builtin_hlsl_wave_prefix_product: {
+    Value *OpExpr = EmitScalarExpr(E->getArg(0));
+    Intrinsic::ID IID = getWavePrefixProductIntrinsic(
+        getTarget().getTriple().getArch(), CGM.getHLSLRuntime(),
+        E->getArg(0)->getType());
+    return EmitRuntimeCall(Intrinsic::getOrInsertDeclaration(
+                               &CGM.getModule(), IID, {OpExpr->getType()}),
+                           ArrayRef{OpExpr}, "hlsl.wave.prefix.product");
+  }
   case Builtin::BI__builtin_hlsl_elementwise_sign: {
     auto *Arg0 = E->getArg(0);
     Value *Op0 = EmitScalarExpr(Arg0);
diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h 
b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
index aaad8f94e23af..2543401bdfbf9 100644
--- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
@@ -3007,6 +3007,105 @@ __attribute__((convergent)) double3 
WavePrefixSum(double3);
 _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum)
 __attribute__((convergent)) double4 WavePrefixSum(double4);
 
+//===----------------------------------------------------------------------===//
+// WavePrefixProduct builtins
+//===----------------------------------------------------------------------===//
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) half WavePrefixProduct(half);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) half2 WavePrefixProduct(half2);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) half3 WavePrefixProduct(half3);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) half4 WavePrefixProduct(half4);
+
+#ifdef __HLSL_ENABLE_16_BIT
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int16_t WavePrefixProduct(int16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int16_t2 WavePrefixProduct(int16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int16_t3 WavePrefixProduct(int16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int16_t4 WavePrefixProduct(int16_t4);
+
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint16_t WavePrefixProduct(uint16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint16_t2 WavePrefixProduct(uint16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint16_t3 WavePrefixProduct(uint16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint16_t4 WavePrefixProduct(uint16_t4);
+#endif
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int WavePrefixProduct(int);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int2 WavePrefixProduct(int2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int3 WavePrefixProduct(int3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int4 WavePrefixProduct(int4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint WavePrefixProduct(uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint2 WavePrefixProduct(uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint3 WavePrefixProduct(uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint4 WavePrefixProduct(uint4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int64_t WavePrefixProduct(int64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int64_t2 WavePrefixProduct(int64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int64_t3 WavePrefixProduct(int64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) int64_t4 WavePrefixProduct(int64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint64_t WavePrefixProduct(uint64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint64_t2 WavePrefixProduct(uint64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint64_t3 WavePrefixProduct(uint64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) uint64_t4 WavePrefixProduct(uint64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) float WavePrefixProduct(float);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) float2 WavePrefixProduct(float2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) float3 WavePrefixProduct(float3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) float4 WavePrefixProduct(float4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) double WavePrefixProduct(double);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) double2 WavePrefixProduct(double2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) double3 WavePrefixProduct(double3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
+__attribute__((convergent)) double4 WavePrefixProduct(double4);
+
 
//===----------------------------------------------------------------------===//
 // sign builtins
 
//===----------------------------------------------------------------------===//
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 45f62b78e71ca..18cca52aad20e 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -3756,7 +3756,8 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned 
BuiltinID, CallExpr *TheCall) {
       return true;
     break;
   }
-  case Builtin::BI__builtin_hlsl_wave_prefix_sum: {
+  case Builtin::BI__builtin_hlsl_wave_prefix_sum:
+  case Builtin::BI__builtin_hlsl_wave_prefix_product: {
     if (SemaRef.checkArgCount(TheCall, 1))
       return true;
 
diff --git a/clang/test/CodeGenHLSL/builtins/WavePrefixProduct.hlsl 
b/clang/test/CodeGenHLSL/builtins/WavePrefixProduct.hlsl
new file mode 100644
index 0000000000000..a45cbf29b87f2
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/WavePrefixProduct.hlsl
@@ -0,0 +1,46 @@
+// RUN: %clang_cc1 -std=hlsl2021 -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 -std=hlsl2021 -finclude-default-header -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) {
+  // CHECK-SPIRV:  %[[RET:.*]] = call spir_func [[TY:.*]] 
@llvm.spv.wave.prefix.product.i32([[TY]] %[[#]])
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY:.*]] 
@llvm.dx.wave.prefix.product.i32([[TY]] %[[#]])
+  // CHECK:  ret [[TY]] %[[RET]]
+  return WavePrefixProduct(expr);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.prefix.product.i32([[TY]]) 
#[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.prefix.product.i32([[TY]]) 
#[[#attr:]]
+
+// CHECK-LABEL: test_uint64_t
+uint64_t test_uint64_t(uint64_t expr) {
+  // CHECK-SPIRV:  %[[RET:.*]] = call spir_func [[TY:.*]] 
@llvm.spv.wave.prefix.product.i64([[TY]] %[[#]])
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY:.*]] 
@llvm.dx.wave.prefix.uproduct.i64([[TY]] %[[#]])
+  // CHECK:  ret [[TY]] %[[RET]]
+  return WavePrefixProduct(expr);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.prefix.uproduct.i64([[TY]]) 
#[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.prefix.product.i64([[TY]]) 
#[[#attr:]]
+
+// Test basic lowering to runtime function call with array and float value.
+
+// CHECK-LABEL: test_floatv4
+float4 test_floatv4(float4 expr) {
+  // CHECK-SPIRV:  %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn 
spir_func [[TY1:.*]] @llvm.spv.wave.prefix.product.v4f32([[TY1]] %[[#]]
+  // CHECK-DXIL:  %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn 
[[TY1:.*]] @llvm.dx.wave.prefix.product.v4f32([[TY1]] %[[#]])
+  // CHECK:  ret [[TY1]] %[[RET1]]
+  return WavePrefixProduct(expr);
+}
+
+// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.prefix.product.v4f32([[TY1]]) 
#[[#attr]]
+// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.prefix.product.v4f32([[TY1]]) 
#[[#attr]]
+
+// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}}
+
diff --git a/clang/test/SemaHLSL/BuiltIns/WavePrefixProduct-errors.hlsl 
b/clang/test/SemaHLSL/BuiltIns/WavePrefixProduct-errors.hlsl
new file mode 100644
index 0000000000000..7fd44b2705b5a
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/WavePrefixProduct-errors.hlsl
@@ -0,0 +1,28 @@
+// RUN: %clang_cc1 -finclude-default-header -triple 
dxil-pc-shadermodel6.6-library %s -emit-llvm-only -disable-llvm-passes -verify
+
+int test_too_few_arg() {
+  return __builtin_hlsl_wave_prefix_product();
+  // expected-error@-1 {{too few arguments to function call, expected 1, have 
0}}
+}
+
+float2 test_too_many_arg(float2 p0) {
+  return __builtin_hlsl_wave_prefix_product(p0, p0);
+  // expected-error@-1 {{too many arguments to function call, expected 1, have 
2}}
+}
+
+bool test_expr_bool_type_check(bool p0) {
+  return __builtin_hlsl_wave_prefix_product(p0);
+  // expected-error@-1 {{invalid operand of type 'bool'}}
+}
+
+bool2 test_expr_bool_vec_type_check(bool2 p0) {
+  return __builtin_hlsl_wave_prefix_product(p0);
+  // expected-error@-1 {{invalid operand of type 'bool2' (aka 'vector<bool, 
2>')}}
+}
+
+struct S { float f; };
+
+S test_expr_struct_type_check(S p0) {
+  return __builtin_hlsl_wave_prefix_product(p0);
+  // expected-error@-1 {{invalid operand of type 'S' where a scalar or vector 
is required}}
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td 
b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 88732bfa5a892..09b15e3994aab 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -182,6 +182,8 @@ def int_dx_wave_get_lane_count
     : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent]>;
 def int_dx_wave_prefix_sum : DefaultAttrsIntrinsic<[llvm_any_ty], 
[LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
 def int_dx_wave_prefix_usum : DefaultAttrsIntrinsic<[llvm_anyint_ty], 
[LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
+def int_dx_wave_prefix_product : DefaultAttrsIntrinsic<[llvm_any_ty], 
[LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
+def int_dx_wave_prefix_uproduct : DefaultAttrsIntrinsic<[llvm_anyint_ty], 
[LLVMMatchType<0>], [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_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, 
LLVMMatchType<0>],
diff --git a/llvm/include/llvm/IR/IntrinsicsSPIRV.td 
b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
index 41526b8c9afc3..c4f09e2a2109b 100644
--- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td
+++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
@@ -134,6 +134,7 @@ def int_spv_rsqrt : 
DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty]
   def int_spv_wave_get_lane_count
       : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent]>;
   def int_spv_wave_prefix_sum : DefaultAttrsIntrinsic<[llvm_any_ty], 
[LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
+  def int_spv_wave_prefix_product : DefaultAttrsIntrinsic<[llvm_any_ty], 
[LLVMMatchType<0>], [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]>;
   def int_spv_group_memory_barrier_with_group_sync : 
ClangBuiltin<"__builtin_spirv_group_barrier">,
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index 8751484496395..59a5b7fe4d508 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -1137,6 +1137,16 @@ def WavePrefixOp : DXILOp<121, wavePrefixOp> {
                    IntrinArgIndex<0>, IntrinArgI8<WaveOpKind_Sum>,
                    IntrinArgI8<SignedOpKind_Unsigned>
                  ]>,
+    IntrinSelect<int_dx_wave_prefix_product,
+                 [
+                   IntrinArgIndex<0>, IntrinArgI8<WaveOpKind_Product>,
+                   IntrinArgI8<SignedOpKind_Signed>
+                 ]>,
+    IntrinSelect<int_dx_wave_prefix_uproduct,
+                 [
+                   IntrinArgIndex<0>, IntrinArgI8<WaveOpKind_Product>,
+                   IntrinArgI8<SignedOpKind_Unsigned>
+                 ]>,
   ];
 
   let arguments = [OverloadTy, Int8Ty, Int8Ty];
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp 
b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index 54cf56f92277b..52993ee1c1220 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -100,6 +100,8 @@ static bool checkWaveOps(Intrinsic::ID IID) {
     // Wave Prefix Op Variants
   case Intrinsic::dx_wave_prefix_sum:
   case Intrinsic::dx_wave_prefix_usum:
+  case Intrinsic::dx_wave_prefix_product:
+  case Intrinsic::dx_wave_prefix_uproduct:
     return true;
   }
 }
diff --git a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp 
b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
index 4d6f6d6472b25..8018b09c9f248 100644
--- a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
+++ b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
@@ -62,10 +62,12 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable(
   case Intrinsic::dx_wave_reduce_min:
   case Intrinsic::dx_wave_reduce_sum:
   case Intrinsic::dx_wave_prefix_sum:
+  case Intrinsic::dx_wave_prefix_product:
   case Intrinsic::dx_wave_reduce_umax:
   case Intrinsic::dx_wave_reduce_umin:
   case Intrinsic::dx_wave_reduce_usum:
   case Intrinsic::dx_wave_prefix_usum:
+  case Intrinsic::dx_wave_prefix_uproduct:
   case Intrinsic::dx_imad:
   case Intrinsic::dx_umad:
   case Intrinsic::dx_ddx_coarse:
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp 
b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index 765fa6277d4ca..eefbab657fe5d 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -249,6 +249,10 @@ class SPIRVInstructionSelector : public 
InstructionSelector {
   bool selectWaveExclusiveScanSum(Register ResVReg, const SPIRVType *ResType,
                                   MachineInstr &I) const;
 
+  bool selectWaveExclusiveScanProduct(Register ResVReg,
+                                      const SPIRVType *ResType,
+                                      MachineInstr &I) const;
+
   bool selectConst(Register ResVReg, const SPIRVType *ResType,
                    MachineInstr &I) const;
 
@@ -2872,6 +2876,18 @@ bool 
SPIRVInstructionSelector::selectWaveExclusiveScanSum(
                                  });
 }
 
+bool SPIRVInstructionSelector::selectWaveExclusiveScanProduct(
+    Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const {
+  return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false,
+                                 [&](Register InputRegister, bool IsUnsigned) {
+                                   bool IsFloatTy = GR.isScalarOrVectorOfType(
+                                       InputRegister, SPIRV::OpTypeFloat);
+                                   return IsFloatTy
+                                              ? SPIRV::OpGroupNonUniformFMul
+                                              : SPIRV::OpGroupNonUniformIMul;
+                                 });
+}
+
 template <typename PickOpcodeFn>
 bool SPIRVInstructionSelector::selectWaveExclusiveScan(
     Register ResVReg, const SPIRVType *ResType, MachineInstr &I,
@@ -3992,6 +4008,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register 
ResVReg,
                             SPIRV::OpGroupNonUniformShuffle);
   case Intrinsic::spv_wave_prefix_sum:
     return selectWaveExclusiveScanSum(ResVReg, ResType, I);
+  case Intrinsic::spv_wave_prefix_product:
+    return selectWaveExclusiveScanProduct(ResVReg, ResType, I);
   case Intrinsic::spv_step:
     return selectExtInst(ResVReg, ResType, I, CL::step, GL::Step);
   case Intrinsic::spv_radians:
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll 
b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
index 12dca7a78cc27..be53d19aca8f2 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
@@ -125,3 +125,17 @@ entry:
   %ret = call i32 @llvm.dx.wave.prefix.usum.i32(i32 %x)
   ret i32 %ret
 }
+
+define noundef i32 @wave_prefix_product(i32 noundef %x) {
+entry:
+  ; CHECK: Function wave_prefix_product : [[WAVE_FLAG]]
+  %ret = call i32 @llvm.dx.wave.prefix.product.i32(i32 %x)
+  ret i32 %ret
+}
+
+define noundef i32 @wave_prefix_uproduct(i32 noundef %x) {
+entry:
+  ; CHECK: Function wave_prefix_uproduct : [[WAVE_FLAG]]
+  %ret = call i32 @llvm.dx.wave.prefix.uproduct.i32(i32 %x)
+  ret i32 %ret
+}
diff --git a/llvm/test/CodeGen/DirectX/WavePrefixProduct.ll 
b/llvm/test/CodeGen/DirectX/WavePrefixProduct.ll
new file mode 100644
index 0000000000000..cb1cffd169d92
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/WavePrefixProduct.ll
@@ -0,0 +1,143 @@
+; RUN: opt -S -scalarizer -dxil-op-lower 
-mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s
+
+; Test that for scalar values, WavePrefixProduct maps down to the DirectX op
+
+define noundef half @wave_prefix_product_half(half noundef %expr) {
+entry:
+; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr, i8 1, i8 0)
+  %ret = call half @llvm.dx.wave.prefix.product.f16(half %expr)
+  ret half %ret
+}
+
+define noundef float @wave_prefix_product_float(float noundef %expr) {
+entry:
+; CHECK: call float @dx.op.wavePrefixOp.f32(i32 121, float %expr, i8 1, i8 0)
+  %ret = call float @llvm.dx.wave.prefix.product.f32(float %expr)
+  ret float %ret
+}
+
+define noundef double @wave_prefix_product_double(double noundef %expr) {
+entry:
+; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr, i8 1, i8 0)
+  %ret = call double @llvm.dx.wave.prefix.product.f64(double %expr)
+  ret double %ret
+}
+
+define noundef i16 @wave_prefix_product_i16(i16 noundef %expr) {
+entry:
+; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr, i8 1, i8 0)
+  %ret = call i16 @llvm.dx.wave.prefix.product.i16(i16 %expr)
+  ret i16 %ret
+}
+
+define noundef i32 @wave_prefix_product_i32(i32 noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr, i8 1, i8 0)
+  %ret = call i32 @llvm.dx.wave.prefix.product.i32(i32 %expr)
+  ret i32 %ret
+}
+
+define noundef i64 @wave_prefix_product_i64(i64 noundef %expr) {
+entry:
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr, i8 1, i8 0)
+  %ret = call i64 @llvm.dx.wave.prefix.product.i64(i64 %expr)
+  ret i64 %ret
+}
+
+define noundef i16 @wave_prefix_uproduct_i16(i16 noundef %expr) {
+entry:
+; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr, i8 1, i8 1)
+  %ret = call i16 @llvm.dx.wave.prefix.uproduct.i16(i16 %expr)
+  ret i16 %ret
+}
+
+define noundef i32 @wave_prefix_uproduct_i32(i32 noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr, i8 1, i8 1)
+  %ret = call i32 @llvm.dx.wave.prefix.uproduct.i32(i32 %expr)
+  ret i32 %ret
+}
+
+define noundef i64 @wave_prefix_uproduct_i64(i64 noundef %expr) {
+entry:
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr, i8 1, i8 1)
+  %ret = call i64 @llvm.dx.wave.prefix.uproduct.i64(i64 %expr)
+  ret i64 %ret
+}
+
+declare half @llvm.dx.wave.prefix.product.f16(half)
+declare float @llvm.dx.wave.prefix.product.f32(float)
+declare double @llvm.dx.wave.prefix.product.f64(double)
+
+declare i16 @llvm.dx.wave.prefix.product.i16(i16)
+declare i32 @llvm.dx.wave.prefix.product.i32(i32)
+declare i64 @llvm.dx.wave.prefix.product.i64(i64)
+
+declare i16 @llvm.dx.wave.prefix.uproduct.i16(i16)
+declare i32 @llvm.dx.wave.prefix.uproduct.i32(i32)
+declare i64 @llvm.dx.wave.prefix.uproduct.i64(i64)
+
+; Test that for vector values, WavePrefixProduct scalarizes and maps down to 
the
+; DirectX op
+
+define noundef <2 x half> @wave_prefix_product_v2half(<2 x half> noundef 
%expr) {
+entry:
+; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr.i0, i8 1, i8 0)
+; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr.i1, i8 1, i8 0)
+  %ret = call <2 x half> @llvm.dx.wave.prefix.product.v2f16(<2 x half> %expr)
+  ret <2 x half> %ret
+}
+
+define noundef <3 x i32> @wave_prefix_product_v3i32(<3 x i32> noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i0, i8 1, i8 0)
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i1, i8 1, i8 0)
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i2, i8 1, i8 0)
+  %ret = call <3 x i32> @llvm.dx.wave.prefix.product.v3i32(<3 x i32> %expr)
+  ret <3 x i32> %ret
+}
+
+define noundef <4 x double> @wave_prefix_product_v4f64(<4 x double> noundef 
%expr) {
+entry:
+; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i0, i8 1, 
i8 0)
+; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i1, i8 1, 
i8 0)
+; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i2, i8 1, 
i8 0)
+; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i3, i8 1, 
i8 0)
+  %ret = call <4 x double> @llvm.dx.wave.prefix.product.v464(<4 x double> 
%expr)
+  ret <4 x double> %ret
+}
+
+declare <2 x half> @llvm.dx.wave.prefix.product.v2f16(<2 x half>)
+declare <3 x i32> @llvm.dx.wave.prefix.product.v3i32(<3 x i32>)
+declare <4 x double> @llvm.dx.wave.prefix.product.v4f64(<4 x double>)
+
+define noundef <2 x i16> @wave_prefix_uproduct_v2i16(<2 x i16> noundef %expr) {
+entry:
+; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr.i0, i8 1, i8 1)
+; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr.i1, i8 1, i8 1)
+  %ret = call <2 x i16> @llvm.dx.wave.prefix.uproduct.v2f16(<2 x i16> %expr)
+  ret <2 x i16> %ret
+}
+
+define noundef <3 x i32> @wave_prefix_uproduct_v3i32(<3 x i32> noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i0, i8 1, i8 1)
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i1, i8 1, i8 1)
+; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i2, i8 1, i8 1)
+  %ret = call <3 x i32> @llvm.dx.wave.prefix.uproduct.v3i32(<3 x i32> %expr)
+  ret <3 x i32> %ret
+}
+
+define noundef <4 x i64> @wave_prefix_uproduct_v4f64(<4 x i64> noundef %expr) {
+entry:
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i0, i8 1, i8 1)
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i1, i8 1, i8 1)
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i2, i8 1, i8 1)
+; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i3, i8 1, i8 1)
+  %ret = call <4 x i64> @llvm.dx.wave.prefix.uproduct.v464(<4 x i64> %expr)
+  ret <4 x i64> %ret
+}
+
+declare <2 x i16> @llvm.dx.wave.prefix.uproduct.v2f16(<2 x i16>)
+declare <3 x i32> @llvm.dx.wave.prefix.uproduct.v3i32(<3 x i32>)
+declare <4 x i64> @llvm.dx.wave.prefix.uproduct.v4f64(<4 x i64>)
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixProduct.ll 
b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixProduct.ll
new file mode 100644
index 0000000000000..8c4b6050ae330
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixProduct.ll
@@ -0,0 +1,41 @@
+; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv-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 for various types and scalar/vector
+
+; CHECK-DAG:   %[[#f16:]] = OpTypeFloat 16
+; CHECK-DAG:   %[[#f32:]] = OpTypeFloat 32
+; CHECK-DAG:   %[[#uint:]] = OpTypeInt 32 0
+; CHECK-DAG:   %[[#v4_half:]] = OpTypeVector %[[#f16]] 4
+; CHECK-DAG:   %[[#scope:]] = OpConstant %[[#uint]] 3
+
+; CHECK-LABEL: Begin function test_float
+; CHECK:   %[[#fexpr:]] = OpFunctionParameter %[[#f32]]
+define float @test_float(float %fexpr) {
+entry:
+; CHECK:   %[[#fret:]] = OpGroupNonUniformFMul %[[#f32]] %[[#scope]] 
ExclusiveScan %[[#fexpr]]
+  %0 = call float @llvm.spv.wave.prefix.product.f32(float %fexpr)
+  ret float %0
+}
+
+; CHECK-LABEL: Begin function test_int
+; CHECK:   %[[#iexpr:]] = OpFunctionParameter %[[#uint]]
+define i32 @test_int(i32 %iexpr) {
+entry:
+; CHECK:   %[[#iret:]] = OpGroupNonUniformIMul %[[#uint]] %[[#scope]] 
ExclusiveScan %[[#iexpr]]
+  %0 = call i32 @llvm.spv.wave.prefix.product.i32(i32 %iexpr)
+  ret i32 %0
+}
+
+; CHECK-LABEL: Begin function test_vhalf
+; CHECK:   %[[#vbexpr:]] = OpFunctionParameter %[[#v4_half]]
+define <4 x half> @test_vhalf(<4 x half> %vbexpr) {
+entry:
+; CHECK:   %[[#vhalfret:]] = OpGroupNonUniformFMul %[[#v4_half]] %[[#scope]] 
ExclusiveScan %[[#vbexpr]]
+  %0 = call <4 x half> @llvm.spv.wave.prefix.product.v4half(<4 x half> %vbexpr)
+  ret <4 x half> %0
+}
+
+declare float @llvm.spv.wave.prefix.product.f32(float)
+declare i32 @llvm.spv.wave.prefix.product.i32(i32)
+declare <4 x half> @llvm.spv.wave.prefix.product.v4half(<4 x half>)

_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to