[clang] [llvm] [HLSL][DXIL][SPIRV] QuadReadAcrossX intrinsic support (PR #184360)
via cfe-commits
cfe-commits at lists.llvm.org
Thu Mar 12 19:34:32 PDT 2026
https://github.com/kcloudy0717 updated https://github.com/llvm/llvm-project/pull/184360
>From ef9b6551b006c10191b13fead9f79c1bf5550868 Mon Sep 17 00:00:00 2001
From: Kai Huang <kcloudy0717 at gmail.com>
Date: Tue, 3 Mar 2026 19:28:09 +0800
Subject: [PATCH 1/3] [HLSL][DXIL][SPIRV] QuadReadAcrossX intrinsic support
---
clang/include/clang/Basic/Builtins.td | 6 ++
clang/lib/CodeGen/CGHLSLBuiltins.cpp | 7 ++
clang/lib/CodeGen/CGHLSLRuntime.h | 1 +
.../lib/Headers/hlsl/hlsl_alias_intrinsics.h | 99 +++++++++++++++++++
clang/lib/Sema/SemaHLSL.cpp | 13 +++
.../CodeGenHLSL/builtins/QuadReadAcrossX.hlsl | 46 +++++++++
.../BuiltIns/QuadReadAcrossX-errors.hlsl | 28 ++++++
llvm/include/llvm/IR/IntrinsicsDirectX.td | 1 +
llvm/include/llvm/IR/IntrinsicsSPIRV.td | 1 +
llvm/lib/Target/DirectX/DXIL.td | 22 +++++
llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 3 +-
.../DirectX/DirectXTargetTransformInfo.cpp | 1 +
llvm/lib/Target/SPIRV/SPIRVInstrInfo.td | 1 +
.../Target/SPIRV/SPIRVInstructionSelector.cpp | 30 ++++++
llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp | 6 +-
llvm/test/CodeGen/DirectX/QuadReadAcrossX.ll | 87 ++++++++++++++++
.../CodeGen/DirectX/ShaderFlags/wave-ops.ll | 7 ++
.../SPIRV/hlsl-intrinsics/QuadReadAcrossX.ll | 44 +++++++++
18 files changed, 401 insertions(+), 2 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/builtins/QuadReadAcrossX.hlsl
create mode 100644 clang/test/SemaHLSL/BuiltIns/QuadReadAcrossX-errors.hlsl
create mode 100644 llvm/test/CodeGen/DirectX/QuadReadAcrossX.ll
create mode 100644 llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadReadAcrossX.ll
diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index 4805c93eef2bf..8c6c08dd67893 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -5264,6 +5264,12 @@ def HLSLWavePrefixProduct : LangBuiltin<"HLSL_LANG"> {
let Prototype = "void(...)";
}
+def HLSLQuadReadAcrossX : LangBuiltin<"HLSL_LANG"> {
+ let Spellings = ["__builtin_hlsl_quad_read_across_x"];
+ 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 677f33eb938a7..d990ddf683194 100644
--- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp
+++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
@@ -1377,6 +1377,13 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
&CGM.getModule(), IID, {OpExpr->getType()}),
ArrayRef{OpExpr}, "hlsl.wave.prefix.product");
}
+ case Builtin::BI__builtin_hlsl_quad_read_across_x: {
+ Value *OpExpr = EmitScalarExpr(E->getArg(0));
+ Intrinsic::ID ID = CGM.getHLSLRuntime().getQuadReadAcrossXIntrinsic();
+ return EmitRuntimeCall(Intrinsic::getOrInsertDeclaration(
+ &CGM.getModule(), ID, {OpExpr->getType()}),
+ ArrayRef{OpExpr}, "hlsl.quad.read.across.x");
+ }
case Builtin::BI__builtin_hlsl_elementwise_sign: {
auto *Arg0 = E->getArg(0);
Value *Op0 = EmitScalarExpr(Arg0);
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.h b/clang/lib/CodeGen/CGHLSLRuntime.h
index 10b72f450dd2c..21500aad190a5 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.h
+++ b/clang/lib/CodeGen/CGHLSLRuntime.h
@@ -156,6 +156,7 @@ class CGHLSLRuntime {
GENERATE_HLSL_INTRINSIC_FUNCTION(WaveIsFirstLane, wave_is_first_lane)
GENERATE_HLSL_INTRINSIC_FUNCTION(WaveGetLaneCount, wave_get_lane_count)
GENERATE_HLSL_INTRINSIC_FUNCTION(WaveReadLaneAt, wave_readlane)
+ GENERATE_HLSL_INTRINSIC_FUNCTION(QuadReadAcrossX, quad_read_across_x)
GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitUHigh, firstbituhigh)
GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitSHigh, firstbitshigh)
GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitLow, firstbitlow)
diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
index 440bb5533c278..ac10ba731bb54 100644
--- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
@@ -3451,6 +3451,105 @@ __attribute__((convergent)) double3 WavePrefixProduct(double3);
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_product)
__attribute__((convergent)) double4 WavePrefixProduct(double4);
+//===----------------------------------------------------------------------===//
+// QuadReadAcrossX builtins
+//===----------------------------------------------------------------------===//
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) half QuadReadAcrossX(half);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) half2 QuadReadAcrossX(half2);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) half3 QuadReadAcrossX(half3);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) half4 QuadReadAcrossX(half4);
+
+#ifdef __HLSL_ENABLE_16_BIT
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int16_t QuadReadAcrossX(int16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int16_t2 QuadReadAcrossX(int16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int16_t3 QuadReadAcrossX(int16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int16_t4 QuadReadAcrossX(int16_t4);
+
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint16_t QuadReadAcrossX(uint16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint16_t2 QuadReadAcrossX(uint16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint16_t3 QuadReadAcrossX(uint16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint16_t4 QuadReadAcrossX(uint16_t4);
+#endif
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int QuadReadAcrossX(int);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int2 QuadReadAcrossX(int2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int3 QuadReadAcrossX(int3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int4 QuadReadAcrossX(int4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint QuadReadAcrossX(uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint2 QuadReadAcrossX(uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint3 QuadReadAcrossX(uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint4 QuadReadAcrossX(uint4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int64_t QuadReadAcrossX(int64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int64_t2 QuadReadAcrossX(int64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int64_t3 QuadReadAcrossX(int64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) int64_t4 QuadReadAcrossX(int64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint64_t QuadReadAcrossX(uint64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint64_t2 QuadReadAcrossX(uint64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint64_t3 QuadReadAcrossX(uint64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) uint64_t4 QuadReadAcrossX(uint64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) float QuadReadAcrossX(float);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) float2 QuadReadAcrossX(float2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) float3 QuadReadAcrossX(float3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) float4 QuadReadAcrossX(float4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) double QuadReadAcrossX(double);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) double2 QuadReadAcrossX(double2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) double3 QuadReadAcrossX(double3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
+__attribute__((convergent)) double4 QuadReadAcrossX(double4);
+
//===----------------------------------------------------------------------===//
// sign builtins
//===----------------------------------------------------------------------===//
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index e200c595ab2c2..a4c7a37821a35 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -4208,6 +4208,19 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
TheCall->setType(ArgTyExpr);
break;
}
+ case Builtin::BI__builtin_hlsl_quad_read_across_x: {
+ if (SemaRef.checkArgCount(TheCall, 1))
+ return true;
+
+ if (CheckAnyScalarOrVector(&SemaRef, TheCall, 0))
+ return true;
+ if (CheckNotBoolScalarOrVector(&SemaRef, TheCall, 0))
+ return true;
+ ExprResult Expr = TheCall->getArg(0);
+ QualType ArgTyExpr = Expr.get()->getType();
+ TheCall->setType(ArgTyExpr);
+ break;
+ }
case Builtin::BI__builtin_hlsl_elementwise_splitdouble: {
if (SemaRef.checkArgCount(TheCall, 3))
return true;
diff --git a/clang/test/CodeGenHLSL/builtins/QuadReadAcrossX.hlsl b/clang/test/CodeGenHLSL/builtins/QuadReadAcrossX.hlsl
new file mode 100644
index 0000000000000..e0f14cd8be9b5
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/QuadReadAcrossX.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.quad.read.across.x.i32([[TY]] %[[#]])
+ // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.quad.read.across.x.i32([[TY]] %[[#]])
+ // CHECK: ret [[TY]] %[[RET]]
+ return QuadReadAcrossX(expr);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.quad.read.across.x.i32([[TY]]) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.quad.read.across.x.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.quad.read.across.x.i64([[TY]] %[[#]])
+ // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.quad.read.across.x.i64([[TY]] %[[#]])
+ // CHECK: ret [[TY]] %[[RET]]
+ return QuadReadAcrossX(expr);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.quad.read.across.x.i64([[TY]]) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.quad.read.across.x.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.quad.read.across.x.v4f32([[TY1]] %[[#]]
+ // CHECK-DXIL: %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn [[TY1:.*]] @llvm.dx.quad.read.across.x.v4f32([[TY1]] %[[#]])
+ // CHECK: ret [[TY1]] %[[RET1]]
+ return QuadReadAcrossX(expr);
+}
+
+// CHECK-DXIL: declare [[TY1]] @llvm.dx.quad.read.across.x.v4f32([[TY1]]) #[[#attr]]
+// CHECK-SPIRV: declare [[TY1]] @llvm.spv.quad.read.across.x.v4f32([[TY1]]) #[[#attr]]
+
+// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}}
+
diff --git a/clang/test/SemaHLSL/BuiltIns/QuadReadAcrossX-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/QuadReadAcrossX-errors.hlsl
new file mode 100644
index 0000000000000..a9dcc162bbbb5
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/QuadReadAcrossX-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_quad_read_across_x();
+ // expected-error at -1 {{too few arguments to function call, expected 1, have 0}}
+}
+
+float2 test_too_many_arg(float2 p0) {
+ return __builtin_hlsl_quad_read_across_x(p0, p0);
+ // expected-error at -1 {{too many arguments to function call, expected 1, have 2}}
+}
+
+bool test_expr_bool_type_check(bool p0) {
+ return __builtin_hlsl_quad_read_across_x(p0);
+ // expected-error at -1 {{invalid operand of type 'bool'}}
+}
+
+bool2 test_expr_bool_vec_type_check(bool2 p0) {
+ return __builtin_hlsl_quad_read_across_x(p0);
+ // expected-error at -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_quad_read_across_x(p0);
+ // expected-error at -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 52a305888f699..72727f4c98c15 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -253,6 +253,7 @@ def int_dx_wave_prefix_sum : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType
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_quad_read_across_x : DefaultAttrsIntrinsic<[llvm_any_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 5bc93ec87ef66..8e24dce0cdb2a 100644
--- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td
+++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
@@ -146,6 +146,7 @@ def int_spv_rsqrt : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty]
: 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_quad_read_across_x : 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 1dc3f34e152e2..8c7074fddc009 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -324,6 +324,10 @@ defvar WaveBitOpKind_Xor = 2;
defvar SignedOpKind_Signed = 0;
defvar SignedOpKind_Unsigned = 1;
+defvar QuadOpKind_ReadAcrossX = 0;
+defvar QuadOpKind_ReadAcrossY = 1;
+defvar QuadOpKind_ReadAcrossDiagonal = 2;
+
// Intrinsic arg selection
class IntrinArgSelectType;
def IntrinArgSelect_Index : IntrinArgSelectType;
@@ -1199,6 +1203,24 @@ def WavePrefixOp : DXILOp<121, wavePrefixOp> {
let attributes = [Attributes<DXIL1_0, []>];
}
+def QuadOp : DXILOp<123, quadOp> {
+ let Doc = "returns the value from another lane within the quad by swapping values in a direction";
+ let intrinsics = [
+ IntrinSelect<int_dx_quad_read_across_x,
+ [
+ IntrinArgIndex<0>, IntrinArgI8<QuadOpKind_ReadAcrossX>
+ ]>,
+ ];
+
+ let arguments = [OverloadTy, Int8Ty];
+ let result = OverloadTy;
+ let overloads = [
+ Overloads<DXIL1_0, [HalfTy, FloatTy, DoubleTy, Int16Ty, Int32Ty, Int64Ty]>
+ ];
+ let stages = [Stages<DXIL1_0, [all_stages]>];
+ let attributes = [Attributes<DXIL1_0, []>];
+}
+
def WavePrefixBitCount : DXILOp<136, wavePrefixOp> {
let Doc = "returns the count of bits of Expr set to 1 on prior lanes";
let intrinsics = [IntrinSelect<int_dx_wave_prefix_bit_count>];
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index b9541daeead31..d42ad6a6ec967 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -70,7 +70,6 @@ static bool checkWaveOps(Intrinsic::ID IID) {
// case Intrinsic::dx_wave_reduce.xor:
// case Intrinsic::dx_wave_prefixop:
// case Intrinsic::dx_quad.readat:
- // case Intrinsic::dx_quad.readacrossx:
// case Intrinsic::dx_quad.readacrossy:
// case Intrinsic::dx_quad.readacrossdiagonal:
// case Intrinsic::dx_wave_prefixballot:
@@ -105,6 +104,8 @@ static bool checkWaveOps(Intrinsic::ID IID) {
case Intrinsic::dx_wave_prefix_usum:
case Intrinsic::dx_wave_prefix_product:
case Intrinsic::dx_wave_prefix_uproduct:
+ // Quad Op Variants
+ case Intrinsic::dx_quad_read_across_x:
return true;
}
}
diff --git a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
index a36ca97dba2be..f12500bcfbc7f 100644
--- a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
+++ b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
@@ -75,6 +75,7 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable(
case Intrinsic::dx_wave_uproduct:
case Intrinsic::dx_wave_prefix_usum:
case Intrinsic::dx_wave_prefix_uproduct:
+ case Intrinsic::dx_quad_read_across_x:
case Intrinsic::dx_imad:
case Intrinsic::dx_umad:
case Intrinsic::dx_ddx_coarse:
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td b/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
index 819cdd6107d0d..84fbfce31cb78 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
+++ b/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
@@ -849,6 +849,7 @@ def OpGroupNonUniformBitwiseXor: OpGroupNUGroup<"BitwiseXor", 361>;
def OpGroupNonUniformLogicalAnd: OpGroupNUGroup<"LogicalAnd", 362>;
def OpGroupNonUniformLogicalOr: OpGroupNUGroup<"LogicalOr", 363>;
def OpGroupNonUniformLogicalXor: OpGroupNUGroup<"LogicalXor", 364>;
+def OpGroupNonUniformQuadSwap: OpGroupNU4<"QuadSwap", 366>;
// SPV_KHR_subgroup_rotate
def OpGroupNonUniformRotateKHR: Op<4431, (outs ID:$res),
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index 3e50e4a0e8c80..d18763c1707ac 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -274,6 +274,9 @@ class SPIRVInstructionSelector : public InstructionSelector {
bool selectWaveExclusiveScanProduct(Register ResVReg, SPIRVTypeInst ResType,
MachineInstr &I) const;
+ bool selectQuadSwap(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
+ unsigned Direction) const;
+
bool selectConst(Register ResVReg, SPIRVTypeInst ResType,
MachineInstr &I) const;
@@ -3188,6 +3191,30 @@ bool SPIRVInstructionSelector::selectWaveExclusiveScan(
return true;
}
+bool SPIRVInstructionSelector::selectQuadSwap(Register ResVReg,
+ SPIRVTypeInst ResType,
+ MachineInstr &I,
+ unsigned Direction) const {
+ assert(I.getNumOperands() == 3);
+ assert(I.getOperand(2).isReg());
+ MachineBasicBlock &BB = *I.getParent();
+ Register InputRegister = I.getOperand(2).getReg();
+
+ SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
+ bool ZeroAsNull = !STI.isShader();
+ Register DirectionReg =
+ GR.getOrCreateConstInt(Direction, I, IntTy, TII, ZeroAsNull);
+ BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGroupNonUniformQuadSwap))
+ .addDef(ResVReg)
+ .addUse(GR.getSPIRVTypeID(ResType))
+ .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
+ ZeroAsNull))
+ .addUse(InputRegister)
+ .addUse(DirectionReg)
+ .constrainAllUses(TII, TRI, RBI);
+ return true;
+}
+
bool SPIRVInstructionSelector::selectBitreverse16(Register ResVReg,
SPIRVTypeInst ResType,
MachineInstr &I,
@@ -4421,6 +4448,9 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
return selectWaveExclusiveScanSum(ResVReg, ResType, I);
case Intrinsic::spv_wave_prefix_product:
return selectWaveExclusiveScanProduct(ResVReg, ResType, I);
+ case Intrinsic::spv_quad_read_across_x: {
+ return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 0);
+ }
case Intrinsic::spv_step:
return selectExtInst(ResVReg, ResType, I, CL::step, GL::Step);
case Intrinsic::spv_radians:
diff --git a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
index 38bc72ff154fa..2cd80d5564d23 100644
--- a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
@@ -869,7 +869,8 @@ void RequirementHandler::initAvailableCapabilities(const SPIRVSubtarget &ST) {
Capability::GroupNonUniformBallot,
Capability::GroupNonUniformClustered,
Capability::GroupNonUniformShuffle,
- Capability::GroupNonUniformShuffleRelative});
+ Capability::GroupNonUniformShuffleRelative,
+ Capability::GroupNonUniformQuad});
if (ST.isAtLeastSPIRVVer(VersionTuple(1, 6)))
addAvailableCaps({Capability::DotProduct, Capability::DotProductInputAll,
@@ -1673,6 +1674,9 @@ void addInstrRequirements(const MachineInstr &MI,
}
break;
}
+ case SPIRV::OpGroupNonUniformQuadSwap:
+ Reqs.addCapability(SPIRV::Capability::GroupNonUniformQuad);
+ break;
case SPIRV::OpImageQueryFormat: {
Register ResultReg = MI.getOperand(0).getReg();
const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
diff --git a/llvm/test/CodeGen/DirectX/QuadReadAcrossX.ll b/llvm/test/CodeGen/DirectX/QuadReadAcrossX.ll
new file mode 100644
index 0000000000000..fd4fd165a52c1
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/QuadReadAcrossX.ll
@@ -0,0 +1,87 @@
+; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s
+
+; Test that for scalar values, QuadReadAcrossX maps down to the DirectX op
+
+define noundef half @quad_read_across_x_half(half noundef %expr) {
+entry:
+; CHECK: call half @dx.op.quadOp.f16(i32 123, half %expr, i8 0)
+ %ret = call half @llvm.dx.quad.read.across.x.f16(half %expr)
+ ret half %ret
+}
+
+define noundef float @quad_read_across_x_float(float noundef %expr) {
+entry:
+; CHECK: call float @dx.op.quadOp.f32(i32 123, float %expr, i8 0)
+ %ret = call float @llvm.dx.quad.read.across.x.f32(float %expr)
+ ret float %ret
+}
+
+define noundef double @quad_read_across_x_double(double noundef %expr) {
+entry:
+; CHECK: call double @dx.op.quadOp.f64(i32 123, double %expr, i8 0)
+ %ret = call double @llvm.dx.quad.read.across.x.f64(double %expr)
+ ret double %ret
+}
+
+define noundef i16 @quad_read_across_x_i16(i16 noundef %expr) {
+entry:
+; CHECK: call i16 @dx.op.quadOp.i16(i32 123, i16 %expr, i8 0)
+ %ret = call i16 @llvm.dx.quad.read.across.x.i16(i16 %expr)
+ ret i16 %ret
+}
+
+define noundef i32 @quad_read_across_x_i32(i32 noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.quadOp.i32(i32 123, i32 %expr, i8 0)
+ %ret = call i32 @llvm.dx.quad.read.across.x.i32(i32 %expr)
+ ret i32 %ret
+}
+
+define noundef i64 @quad_read_across_x_i64(i64 noundef %expr) {
+entry:
+; CHECK: call i64 @dx.op.quadOp.i64(i32 123, i64 %expr, i8 0)
+ %ret = call i64 @llvm.dx.quad.read.across.x.i64(i64 %expr)
+ ret i64 %ret
+}
+
+declare half @llvm.dx.quad.read.across.x.f16(half)
+declare float @llvm.dx.quad.read.across.x.f32(float)
+declare double @llvm.dx.quad.read.across.x.f64(double)
+
+declare i16 @llvm.dx.quad.read.across.x.i16(i16)
+declare i32 @llvm.dx.quad.read.across.x.i32(i32)
+declare i64 @llvm.dx.quad.read.across.x.i64(i64)
+
+; Test that for vector values, QuadReadAcrossX scalarizes and maps down to the
+; DirectX op
+
+define noundef <2 x half> @quad_read_across_x_v2half(<2 x half> noundef %expr) {
+entry:
+; CHECK: call half @dx.op.quadOp.f16(i32 123, half %expr.i0, i8 0)
+; CHECK: call half @dx.op.quadOp.f16(i32 123, half %expr.i1, i8 0)
+ %ret = call <2 x half> @llvm.dx.quad.read.across.x.v2f16(<2 x half> %expr)
+ ret <2 x half> %ret
+}
+
+define noundef <3 x i32> @quad_read_across_x_v3i32(<3 x i32> noundef %expr) {
+entry:
+; CHECK: call i32 @dx.op.quadOp.i32(i32 123, i32 %expr.i0, i8 0)
+; CHECK: call i32 @dx.op.quadOp.i32(i32 123, i32 %expr.i1, i8 0)
+; CHECK: call i32 @dx.op.quadOp.i32(i32 123, i32 %expr.i2, i8 0)
+ %ret = call <3 x i32> @llvm.dx.quad.read.across.x.v3i32(<3 x i32> %expr)
+ ret <3 x i32> %ret
+}
+
+define noundef <4 x double> @quad_read_across_x_v4f64(<4 x double> noundef %expr) {
+entry:
+; CHECK: call double @dx.op.quadOp.f64(i32 123, double %expr.i0, i8 0)
+; CHECK: call double @dx.op.quadOp.f64(i32 123, double %expr.i1, i8 0)
+; CHECK: call double @dx.op.quadOp.f64(i32 123, double %expr.i2, i8 0)
+; CHECK: call double @dx.op.quadOp.f64(i32 123, double %expr.i3, i8 0)
+ %ret = call <4 x double> @llvm.dx.quad.read.across.x.v464(<4 x double> %expr)
+ ret <4 x double> %ret
+}
+
+declare <2 x half> @llvm.dx.quad.read.across.x.v2f16(<2 x half>)
+declare <3 x i32> @llvm.dx.quad.read.across.x.v3i32(<3 x i32>)
+declare <4 x double> @llvm.dx.quad.read.across.x.v4f64(<4 x double>)
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
index c42bf2d5fdf17..19a39b464b7bb 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll
@@ -167,3 +167,10 @@ entry:
%ret = call i32 @llvm.dx.wave.prefix.uproduct.i32(i32 %x)
ret i32 %ret
}
+
+define noundef i32 @quad_read_across_x_i32(i32 noundef %expr) {
+entry:
+ ; CHECK: Function quad_read_across_x_i32 : [[WAVE_FLAG]]
+ %ret = call i32 @llvm.dx.quad.read.across.x.i32(i32 %expr)
+ ret i32 %ret
+}
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadReadAcrossX.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadReadAcrossX.ll
new file mode 100644
index 0000000000000..688ee1bce4ef4
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadReadAcrossX.ll
@@ -0,0 +1,44 @@
+; 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: OpCapability GroupNonUniformQuad
+
+; 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-DAG: %[[#direction:]] = OpConstant %[[#uint]] 0
+
+; CHECK-LABEL: Begin function test_float
+; CHECK: %[[#fexpr:]] = OpFunctionParameter %[[#f32]]
+define float @test_float(float %fexpr) {
+entry:
+; CHECK: %[[#fret:]] = OpGroupNonUniformQuadSwap %[[#f32]] %[[#scope]] %[[#fexpr]] %[[#direction]]
+ %0 = call float @llvm.spv.quad.read.across.x.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:]] = OpGroupNonUniformQuadSwap %[[#uint]] %[[#scope]] %[[#iexpr]] %[[#direction]]
+ %0 = call i32 @llvm.spv.quad.read.across.x.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:]] = OpGroupNonUniformQuadSwap %[[#v4_half]] %[[#scope]] %[[#vbexpr]] %[[#direction]]
+ %0 = call <4 x half> @llvm.spv.quad.read.across.x.v4half(<4 x half> %vbexpr)
+ ret <4 x half> %0
+}
+
+declare float @llvm.spv.quad.read.across.x.f32(float)
+declare i32 @llvm.spv.quad.read.across.x.i32(i32)
+declare <4 x half> @llvm.spv.quad.read.across.x.v4half(<4 x half>)
>From c0dd576d49660b2e84b866cb18607c72e57451c5 Mon Sep 17 00:00:00 2001
From: Kai Huang <kcloudy0717 at gmail.com>
Date: Fri, 13 Mar 2026 02:21:48 +0800
Subject: [PATCH 2/3] Fix shader model target for _HLSL_16BIT_AVAILABILITY
macro
---
clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
index ac10ba731bb54..be0703c4c5e53 100644
--- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
@@ -3455,16 +3455,16 @@ __attribute__((convergent)) double4 WavePrefixProduct(double4);
// QuadReadAcrossX builtins
//===----------------------------------------------------------------------===//
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) half QuadReadAcrossX(half);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) half2 QuadReadAcrossX(half2);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) half3 QuadReadAcrossX(half3);
-_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) half4 QuadReadAcrossX(half4);
>From 0f2109539af6040508fdb5f839f0162f0462ea29 Mon Sep 17 00:00:00 2001
From: Kai Huang <kcloudy0717 at gmail.com>
Date: Fri, 13 Mar 2026 10:34:13 +0800
Subject: [PATCH 3/3] Fixed shader model target for intrinsics within
__HLSL_ENABLE_16_BIT macro too
---
clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
index be0703c4c5e53..02ccbd3132f3b 100644
--- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
@@ -3469,29 +3469,29 @@ _HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) half4 QuadReadAcrossX(half4);
#ifdef __HLSL_ENABLE_16_BIT
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) int16_t QuadReadAcrossX(int16_t);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) int16_t2 QuadReadAcrossX(int16_t2);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) int16_t3 QuadReadAcrossX(int16_t3);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) int16_t4 QuadReadAcrossX(int16_t4);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) uint16_t QuadReadAcrossX(uint16_t);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) uint16_t2 QuadReadAcrossX(uint16_t2);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) uint16_t3 QuadReadAcrossX(uint16_t3);
-_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_AVAILABILITY(shadermodel, 6.2)
_HLSL_BUILTIN_ALIAS(__builtin_hlsl_quad_read_across_x)
__attribute__((convergent)) uint16_t4 QuadReadAcrossX(uint16_t4);
#endif
More information about the cfe-commits
mailing list