[clang] [llvm] [HLSL] Add QuadAll Emulation support (PR #240645)
Joshua Batista via cfe-commits
cfe-commits at lists.llvm.org
Fri Nov 13 17:06:10 PST 2026
https://github.com/bob80905 updated https://github.com/llvm/llvm-project/pull/240645
>From 76027fba1b5e048f1f64bca3ec9c890e37d21948 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Mon, 9 Nov 2026 16:45:24 -0800
Subject: [PATCH 1/9] first attempt
Co-authored-by: Copilot <223556219+Copilot at users.noreply.github.com>
Copilot-Session: 365638e8-2291-4925-8f1e-1b5c23220818
---
clang/include/clang/Basic/HLSLIntrinsics.td | 2 +-
clang/test/CodeGenHLSL/builtins/QuadAll.hlsl | 3 ++
.../BuiltIns/QuadAll-availability.hlsl | 10 +++---
llvm/docs/SPIRVUsage.md | 2 ++
.../Target/DirectX/DXILIntrinsicExpansion.cpp | 24 +++++++++++++
.../Target/SPIRV/SPIRVInstructionSelector.cpp | 32 +++++++++++++----
.../test/CodeGen/DirectX/QuadAll-expansion.ll | 36 +++++++++++++++++++
llvm/test/CodeGen/DirectX/QuadAll.ll | 21 +++++++++++
.../quad-all.ll | 22 +++++++++---
.../CodeGen/SPIRV/hlsl-intrinsics/QuadAll.ll | 35 ++++++++++++++++--
10 files changed, 168 insertions(+), 19 deletions(-)
create mode 100644 llvm/test/CodeGen/DirectX/QuadAll-expansion.ll
diff --git a/clang/include/clang/Basic/HLSLIntrinsics.td b/clang/include/clang/Basic/HLSLIntrinsics.td
index d0500838a81f7..dcabb5a6c05d0 100644
--- a/clang/include/clang/Basic/HLSLIntrinsics.td
+++ b/clang/include/clang/Basic/HLSLIntrinsics.td
@@ -1405,7 +1405,7 @@ def hlsl_quad_all : HLSLBuiltin<"QuadAll", "__builtin_hlsl_quad_all"> {
let Args = [BoolTy];
let ReturnType = BoolTy;
let IsConvergent = 1;
- let Availability = SM6_7;
+ let Availability = SM6_0;
}
// Reads the value from the lane across the X axis of the quad.
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll.hlsl
index 986178e8d9ab1..864c2ededec3c 100644
--- a/clang/test/CodeGenHLSL/builtins/QuadAll.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll.hlsl
@@ -2,6 +2,9 @@
// RUN: dxil-pc-shadermodel6.7-library %s -emit-llvm \
// RUN: -disable-llvm-passes -o - | FileCheck %s --check-prefixes=CHECK,DXIL
// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.0-library %s -emit-llvm \
+// RUN: -disable-llvm-passes -o - | FileCheck %s --check-prefixes=CHECK,DXIL
+// RUN: %clang_cc1 -finclude-default-header -triple \
// RUN: spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - | \
// RUN: FileCheck %s --check-prefixes=CHECK,SPIRV
diff --git a/clang/test/SemaHLSL/BuiltIns/QuadAll-availability.hlsl b/clang/test/SemaHLSL/BuiltIns/QuadAll-availability.hlsl
index c6b3bea64e760..6617fecf40280 100644
--- a/clang/test/SemaHLSL/BuiltIns/QuadAll-availability.hlsl
+++ b/clang/test/SemaHLSL/BuiltIns/QuadAll-availability.hlsl
@@ -1,13 +1,15 @@
// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.0-library %s -fhlsl-strict-availability \
+// RUN: -fsyntax-only -verify
+// RUN: %clang_cc1 -finclude-default-header -triple \
// RUN: dxil-pc-shadermodel6.6-library %s -fhlsl-strict-availability \
-// RUN: -fsyntax-only -verify -verify-ignore-unexpected=note
+// RUN: -fsyntax-only -verify
// RUN: %clang_cc1 -finclude-default-header -triple \
// RUN: dxil-pc-shadermodel6.7-library %s -fhlsl-strict-availability \
-// RUN: -fsyntax-only -verify=sm67
+// RUN: -fsyntax-only -verify
-// sm67-no-diagnostics
+// expected-no-diagnostics
bool test_availability(bool cond) {
- // expected-error at +1 {{'QuadAll' is only available on Shader Model 6.7}}
return QuadAll(cond);
}
diff --git a/llvm/docs/SPIRVUsage.md b/llvm/docs/SPIRVUsage.md
index 2cfe73bf398d9..16e4cb48865f7 100644
--- a/llvm/docs/SPIRVUsage.md
+++ b/llvm/docs/SPIRVUsage.md
@@ -242,6 +242,8 @@ Below is a list of supported SPIR-V extensions, sorted alphabetically by their e
- Adds the quad vote instruction for HLSL `QuadAll` and the
`QuadDerivativesKHR` and `RequireFullQuadsKHR` execution modes.
These features also require `SPV_KHR_maximal_reconvergence`.
+ Without `SPV_KHR_quad_control`, `QuadAll` uses quad-read emulation.
+ This path requires SPIR-V 1.3 but not maximal reconvergence.
* - ``SPV_ALTERA_blocking_pipes``
- Adds new pipe read and write functions that have blocking semantics instead of the non-blocking semantics of the existing pipe read/write functions.
* - ``SPV_ALTERA_arbitrary_precision_fixed_point``
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 25bdd17b7f7bc..44892aead9161 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -242,6 +242,9 @@ static bool isIntrinsicExpansion(Function &F) {
case Intrinsic::vector_reduce_or:
return F.getParent()->getTargetTriple().getDXILVersion() <
VersionTuple(1, 9);
+ case Intrinsic::dx_quad_all:
+ return F.getParent()->getTargetTriple().getDXILVersion() <
+ VersionTuple(1, 7);
case Intrinsic::dx_fdot:
return shouldExpandFloatDotIntrinsic(F);
case Intrinsic::dx_resource_load_rawbuffer:
@@ -1261,6 +1264,24 @@ static Value *expandLoadInput(CallInst *Orig) {
return Vec;
}
+static Value *expandQuadAll(CallInst *Orig) {
+ IRBuilder<> Builder(Orig);
+ SmallVector<OperandBundleDef, 1> Bundles;
+ Orig->getOperandBundlesAsDefs(Bundles);
+
+ Value *Cond =
+ Builder.CreateSExt(Orig->getArgOperand(0), Builder.getInt32Ty());
+ Value *Result = Cond;
+ for (Intrinsic::ID ID :
+ {Intrinsic::dx_quad_read_across_x, Intrinsic::dx_quad_read_across_y,
+ Intrinsic::dx_quad_read_across_diagonal}) {
+ Value *QuadValue = Builder.CreateIntrinsic(ID, {Builder.getInt32Ty()}, Cond,
+ nullptr, "", Bundles);
+ Result = Builder.CreateAnd(Result, QuadValue);
+ }
+ return Builder.CreateTrunc(Result, Builder.getInt1Ty());
+}
+
static bool expandIntrinsic(Function &F, CallInst *Orig) {
Value *Result = nullptr;
Intrinsic::ID IntrinsicId = F.getIntrinsicID();
@@ -1327,6 +1348,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
case Intrinsic::dx_sign:
Result = expandSignIntrinsic(Orig);
break;
+ case Intrinsic::dx_quad_all:
+ Result = expandQuadAll(Orig);
+ break;
case Intrinsic::dx_load_input:
Result = expandLoadInput(Orig);
break;
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index 301bd5a486eb3..d109733892e7c 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -329,6 +329,9 @@ class SPIRVInstructionSelector : public InstructionSelector {
bool selectQuadSwap(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
unsigned Direction) const;
+ bool selectQuadAll(Register ResVReg, SPIRVTypeInst ResType,
+ MachineInstr &I) const;
+
bool selectConst(Register ResVReg, SPIRVTypeInst ResType,
MachineInstr &I) const;
@@ -4110,6 +4113,28 @@ bool SPIRVInstructionSelector::selectQuadSwap(Register ResVReg,
return true;
}
+bool SPIRVInstructionSelector::selectQuadAll(Register ResVReg,
+ SPIRVTypeInst ResType,
+ MachineInstr &I) const {
+ if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_quad_control))
+ return selectOpWithSrcs(ResVReg, ResType, I, {I.getOperand(2).getReg()},
+ SPIRV::OpGroupNonUniformQuadAllKHR);
+
+ Register Result = I.getOperand(2).getReg();
+ for (unsigned Direction = 0; Direction != 3; ++Direction) {
+ Register QuadValue = createVirtualRegister(ResType, &GR, MRI, MRI->getMF());
+ Register Next =
+ Direction == 2 ? ResVReg
+ : createVirtualRegister(ResType, &GR, MRI, MRI->getMF());
+ if (!selectQuadSwap(QuadValue, ResType, I, Direction) ||
+ !selectOpWithSrcs(Next, ResType, I, {Result, QuadValue},
+ SPIRV::OpLogicalAnd))
+ return false;
+ Result = Next;
+ }
+ return true;
+}
+
bool SPIRVInstructionSelector::selectBitreverseViaI32(Register ResVReg,
SPIRVTypeInst ResType,
MachineInstr &I,
@@ -5895,12 +5920,7 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
case Intrinsic::spv_wave_prefix_product:
return selectWaveExclusiveScanProduct(ResVReg, ResType, I);
case Intrinsic::spv_quad_all:
- BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGroupNonUniformQuadAllKHR))
- .addDef(ResVReg)
- .addUse(GR.getSPIRVTypeID(ResType))
- .addUse(I.getOperand(2).getReg())
- .constrainAllUses(TII, TRI, RBI);
- return true;
+ return selectQuadAll(ResVReg, ResType, I);
case Intrinsic::spv_quad_read_across_x: {
return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 0);
}
diff --git a/llvm/test/CodeGen/DirectX/QuadAll-expansion.ll b/llvm/test/CodeGen/DirectX/QuadAll-expansion.ll
new file mode 100644
index 0000000000000..75a9fe825de83
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/QuadAll-expansion.ll
@@ -0,0 +1,36 @@
+; RUN: opt -S -dxil-intrinsic-expansion \
+; RUN: -mtriple=dxil-pc-shadermodel6.0-library %s \
+; RUN: | FileCheck %s --check-prefix=EMULATED \
+; RUN: --implicit-check-not=llvm.dx.quad.all
+; RUN: opt -S -dxil-intrinsic-expansion \
+; RUN: -mtriple=dxil-pc-shadermodel6.6-library %s \
+; RUN: | FileCheck %s --check-prefix=EMULATED \
+; RUN: --implicit-check-not=llvm.dx.quad.all
+; RUN: opt -S -dxil-intrinsic-expansion \
+; RUN: -mtriple=dxil-pc-shadermodel6.7-library %s \
+; RUN: | FileCheck %s --check-prefix=NATIVE \
+; RUN: --implicit-check-not=llvm.dx.quad.read
+
+; EMULATED-LABEL: define i1 @quad_all(
+; EMULATED-NEXT: %[[TOKEN:.*]] = call token @llvm.experimental.convergence.entry()
+; EMULATED-NEXT: %[[COND:.*]] = sext i1 %cond to i32
+; EMULATED-NEXT: %[[X:.*]] = call i32 @llvm.dx.quad.read.across.x.i32(i32 %[[COND]]) [ "convergencectrl"(token %[[TOKEN]]) ]
+; EMULATED-NEXT: %[[ANDX:.*]] = and i32 %[[COND]], %[[X]]
+; EMULATED-NEXT: %[[Y:.*]] = call i32 @llvm.dx.quad.read.across.y.i32(i32 %[[COND]]) [ "convergencectrl"(token %[[TOKEN]]) ]
+; EMULATED-NEXT: %[[ANDY:.*]] = and i32 %[[ANDX]], %[[Y]]
+; EMULATED-NEXT: %[[D:.*]] = call i32 @llvm.dx.quad.read.across.diagonal.i32(i32 %[[COND]]) [ "convergencectrl"(token %[[TOKEN]]) ]
+; EMULATED-NEXT: %[[ANDD:.*]] = and i32 %[[ANDY]], %[[D]]
+; EMULATED-NEXT: %[[RET:.*]] = trunc i32 %[[ANDD]] to i1
+; EMULATED-NEXT: ret i1 %[[RET]]
+
+; NATIVE-LABEL: define i1 @quad_all(
+; NATIVE-NEXT: %[[TOKEN:.*]] = call token @llvm.experimental.convergence.entry()
+; NATIVE-NEXT: %ret = call i1 @llvm.dx.quad.all(i1 %cond) [ "convergencectrl"(token %[[TOKEN]]) ]
+; NATIVE-NEXT: ret i1 %ret
+define i1 @quad_all(i1 %cond) convergent {
+ %tok = call token @llvm.experimental.convergence.entry()
+ %ret = call i1 @llvm.dx.quad.all(i1 %cond) [ "convergencectrl"(token %tok) ]
+ ret i1 %ret
+}
+
+declare i1 @llvm.dx.quad.all(i1) convergent
diff --git a/llvm/test/CodeGen/DirectX/QuadAll.ll b/llvm/test/CodeGen/DirectX/QuadAll.ll
index 383de1135adab..a3da4f6919dae 100644
--- a/llvm/test/CodeGen/DirectX/QuadAll.ll
+++ b/llvm/test/CodeGen/DirectX/QuadAll.ll
@@ -8,10 +8,31 @@
; RUN: | FileCheck %s
; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.7-mesh %s \
; RUN: | FileCheck %s
+; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower \
+; RUN: -mtriple=dxil-pc-shadermodel6.0-library %s \
+; RUN: | FileCheck %s --check-prefix=EMULATED --implicit-check-not=quadVote
+; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower \
+; RUN: -mtriple=dxil-pc-shadermodel6.6-compute %s \
+; RUN: | FileCheck %s --check-prefix=EMULATED --implicit-check-not=quadVote
+; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower \
+; RUN: -mtriple=dxil-pc-shadermodel6.6-pixel %s \
+; RUN: | FileCheck %s --check-prefix=EMULATED --implicit-check-not=quadVote
+; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower \
+; RUN: -mtriple=dxil-pc-shadermodel6.7-library %s | FileCheck %s
; CHECK-LABEL: define i1 @quad_all(
; CHECK: %[[RET:.*]] = call i1 @dx.op.quadVote.i1(i32 222, i1 %cond, i8 1)
; CHECK-NEXT: ret i1 %[[RET]]
+; EMULATED-LABEL: define i1 @quad_all(
+; EMULATED-NEXT: %[[COND:.*]] = sext i1 %cond to i32
+; EMULATED-NEXT: %[[X:.*]] = call i32 @dx.op.quadOp.i32(i32 123, i32 %[[COND]], i8 0)
+; EMULATED-NEXT: %[[ANDX:.*]] = and i32 %[[COND]], %[[X]]
+; EMULATED-NEXT: %[[Y:.*]] = call i32 @dx.op.quadOp.i32(i32 123, i32 %[[COND]], i8 1)
+; EMULATED-NEXT: %[[ANDY:.*]] = and i32 %[[ANDX]], %[[Y]]
+; EMULATED-NEXT: %[[D:.*]] = call i32 @dx.op.quadOp.i32(i32 123, i32 %[[COND]], i8 2)
+; EMULATED-NEXT: %[[ANDD:.*]] = and i32 %[[ANDY]], %[[D]]
+; EMULATED-NEXT: %[[RET:.*]] = trunc i32 %[[ANDD]] to i1
+; EMULATED-NEXT: ret i1 %[[RET]]
define i1 @quad_all(i1 %cond) {
%ret = call i1 @llvm.dx.quad.all(i1 %cond)
ret i1 %ret
diff --git a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/quad-all.ll b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/quad-all.ll
index e62c3d2af0755..549c4ea0ac126 100644
--- a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/quad-all.ll
+++ b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/quad-all.ll
@@ -4,17 +4,27 @@
; RUN: llc -O0 -verify-machineinstrs \
; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
; RUN: --spirv-ext=%{deriv-ext},%{quad-exts} \
-; RUN: %s -o - | FileCheck %s
+; RUN: %s -o - | FileCheck %s --check-prefixes=CHECK,NATIVE
+; RUN: llc -O0 -verify-machineinstrs \
+; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
+; RUN: --spirv-ext=%{deriv-ext},-SPV_KHR_quad_control \
+; RUN: %s -o - | FileCheck %s --check-prefixes=CHECK,EMULATED \
+; RUN: --implicit-check-not=SPV_KHR_quad_control \
+; RUN: --implicit-check-not=SPV_KHR_maximal_reconvergence
; RUN: %if spirv-tools %{ llc -O0 \
; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
; RUN: --spirv-ext=%{deriv-ext},%{quad-exts} \
; RUN: %s -o - -filetype=obj | spirv-val --target-env vulkan1.2 %}
+; RUN: %if spirv-tools %{ llc -O0 \
+; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
+; RUN: --spirv-ext=%{deriv-ext},-SPV_KHR_quad_control \
+; RUN: %s -o - -filetype=obj | spirv-val --target-env vulkan1.2 %}
; CHECK-DAG: OpCapability ComputeDerivativeGroupQuadsKHR
-; CHECK-DAG: OpCapability QuadControlKHR
+; NATIVE-DAG: OpCapability QuadControlKHR
; CHECK-DAG: OpExtension "SPV_KHR_compute_shader_derivatives"
-; CHECK-DAG: OpExtension "SPV_KHR_quad_control"
-; CHECK-DAG: OpExtension "SPV_KHR_maximal_reconvergence"
+; NATIVE-DAG: OpExtension "SPV_KHR_quad_control"
+; NATIVE-DAG: OpExtension "SPV_KHR_maximal_reconvergence"
; CHECK: OpEntryPoint GLCompute %[[#MAIN:]] "main"
; CHECK-DAG: OpExecutionMode %[[#MAIN]] DerivativeGroupQuadsKHR
; CHECK-DAG: OpExecutionMode %[[#MAIN]] LocalSize 2 2 1
@@ -22,7 +32,9 @@
; CHECK-DAG: %[[#TRUE:]] = OpConstantTrue %[[#BOOL]]
; CHECK: %[[#MAIN]] = OpFunction
-; CHECK: %[[#ALL:]] = OpGroupNonUniformQuadAllKHR %[[#BOOL]] %[[#TRUE]]
+; NATIVE: %[[#ALL:]] = OpGroupNonUniformQuadAllKHR %[[#BOOL]] %[[#TRUE]]
+; EMULATED-COUNT-3: OpGroupNonUniformQuadSwap %[[#BOOL]] %{{[0-9]+}} %[[#TRUE]]
+; EMULATED-NEXT: %[[#ALL:]] = OpLogicalAnd %[[#BOOL]]
; CHECK: %[[#VALUE:]] = OpSelect %{{[0-9]+}} %[[#ALL]]
; CHECK: OpStore %{{[0-9]+}} %[[#VALUE]] Volatile
define void @main() #0 {
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadAll.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadAll.ll
index 10e8faad8cb0f..58f2a33c8bf94 100644
--- a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadAll.ll
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/QuadAll.ll
@@ -1,8 +1,17 @@
; RUN: llc -O0 -verify-machineinstrs -mtriple=spirv1.5-unknown-vulkan-library \
; RUN: --spirv-ext=+SPV_KHR_quad_control,+SPV_KHR_maximal_reconvergence \
; RUN: %s -o - | FileCheck %s
-; RUN: not llc -O0 -mtriple=spirv1.5-unknown-vulkan-library \
-; RUN: %s -o - 2>&1 | FileCheck %s --check-prefix=NO-EXT
+; RUN: llc -O0 -verify-machineinstrs \
+; RUN: -mtriple=spirv32v1.3-unknown-vulkan-library %s -o - \
+; RUN: | FileCheck %s --check-prefix=EMULATED \
+; RUN: --implicit-check-not=OpGroupNonUniformQuadAllKHR \
+; RUN: --implicit-check-not=OpExtension
+; RUN: llc -O0 -verify-machineinstrs \
+; RUN: -mtriple=spirv1.5-unknown-vulkan-library \
+; RUN: --spirv-ext=-SPV_KHR_quad_control,-SPV_KHR_maximal_reconvergence \
+; RUN: %s -o - | FileCheck %s --check-prefix=EMULATED \
+; RUN: --implicit-check-not=OpGroupNonUniformQuadAllKHR \
+; RUN: --implicit-check-not=OpExtension
; RUN: not llc -O0 -mtriple=spirv1.5-unknown-vulkan-library \
; RUN: --spirv-ext=+SPV_KHR_quad_control,-SPV_KHR_maximal_reconvergence \
; RUN: %s -o - 2>&1 | FileCheck %s --check-prefix=NO-RECONVERGENCE
@@ -13,8 +22,10 @@
; RUN: -mtriple=spirv1.5-unknown-vulkan-library \
; RUN: --spirv-ext=+SPV_KHR_quad_control,+SPV_KHR_maximal_reconvergence \
; RUN: %s -o - -filetype=obj | spirv-val %}
+; RUN: %if spirv-tools %{ llc -O0 \
+; RUN: -mtriple=spirv32v1.3-unknown-vulkan-library \
+; RUN: %s -o - -filetype=obj | spirv-val %}
-; NO-EXT: OpGroupNonUniformQuadAllKHR requires the following SPIR-V extension: SPV_KHR_quad_control
; NO-RECONVERGENCE: SPV_KHR_quad_control requires the following SPIR-V extension: SPV_KHR_maximal_reconvergence
; VERSION: OpGroupNonUniformQuadAllKHR requires SPIR-V 1.3
@@ -28,6 +39,24 @@
; CHECK: %[[#COND:]] = OpFunctionParameter %[[#BOOL]]
; CHECK: %[[#RET:]] = OpGroupNonUniformQuadAllKHR %[[#BOOL]] %[[#COND]]
; CHECK-NEXT: OpReturnValue %[[#RET]]
+
+; EMULATED-DAG: OpCapability GroupNonUniformQuad
+; EMULATED-NOT: OpExecutionMode
+; EMULATED-DAG: %[[#BOOL:]] = OpTypeBool
+; EMULATED-DAG: %[[#INT:]] = OpTypeInt 32 0
+; EMULATED-DAG: %[[#SCOPE:]] = OpConstant %[[#INT]] 3
+; EMULATED-DAG: %[[#XDIR:]] = OpConstant %[[#INT]] 0
+; EMULATED-DAG: %[[#YDIR:]] = OpConstant %[[#INT]] 1
+; EMULATED-DAG: %[[#DDIR:]] = OpConstant %[[#INT]] 2
+; EMULATED-LABEL: Begin function quad_all
+; EMULATED: %[[#COND:]] = OpFunctionParameter %[[#BOOL]]
+; EMULATED: %[[#X:]] = OpGroupNonUniformQuadSwap %[[#BOOL]] %[[#SCOPE]] %[[#COND]] %[[#XDIR]]
+; EMULATED-NEXT: %[[#ANDX:]] = OpLogicalAnd %[[#BOOL]] %[[#COND]] %[[#X]]
+; EMULATED-NEXT: %[[#Y:]] = OpGroupNonUniformQuadSwap %[[#BOOL]] %[[#SCOPE]] %[[#COND]] %[[#YDIR]]
+; EMULATED-NEXT: %[[#ANDY:]] = OpLogicalAnd %[[#BOOL]] %[[#ANDX]] %[[#Y]]
+; EMULATED-NEXT: %[[#D:]] = OpGroupNonUniformQuadSwap %[[#BOOL]] %[[#SCOPE]] %[[#COND]] %[[#DDIR]]
+; EMULATED-NEXT: %[[#RET:]] = OpLogicalAnd %[[#BOOL]] %[[#ANDY]] %[[#D]]
+; EMULATED-NEXT: OpReturnValue %[[#RET]]
define i1 @quad_all(i1 %cond) convergent {
%tok = call token @llvm.experimental.convergence.entry()
%ret = call i1 @llvm.spv.quad.all(i1 %cond) [ "convergencectrl"(token %tok) ]
>From 8af9df36b31e630e4df82d4dd66a8bdf40fb47f0 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Tue, 10 Nov 2026 16:41:37 -0800
Subject: [PATCH 2/9] address Deric, disallow quadall except for mesh / amp /
compute / pixel
---
.../clang/Basic/DiagnosticSemaKinds.td | 3 +
clang/lib/Sema/SemaHLSL.cpp | 21 +++++
.../BuiltIns/QuadAll-stages.errors.hlsl | 86 +++++++++++++++++++
3 files changed, 110 insertions(+)
create mode 100644 clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 3e56980e22902..1bd9316499262 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -13986,6 +13986,9 @@ def err_hlsl_atomic_arg_addr_space : Error<
def err_hlsl_builtin_requires_sm : Error<
"%0 requires shader model %1 or newer">;
+def err_hlsl_builtin_unsupported_in_stage : Error<
+ "%0 is not supported in '%1' shaders">;
+
def err_hlsl_export_not_on_function : Error<
"export declaration can only be used on functions">;
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 506ac8e13fe30..53f57970c2599 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -3191,6 +3191,7 @@ class DiagnoseHLSLAvailability : public DynamicRecursiveASTVisitor {
SourceRange Range);
const AvailabilityAttr *FindAvailabilityAttr(const Decl *D);
bool HasMatchingEnvironmentOrNone(const AvailabilityAttr *AA);
+ void DiagnoseQuadAllCall(CallExpr *CE);
void DiagnoseBarrierCall(CallExpr *CE);
uint64_t DiagnoseBarrierGroupMemory(Expr *MemoryArg, uint64_t MemoryFlags,
bool HasVisibleGroup, bool IsAllMemory);
@@ -3226,11 +3227,31 @@ class DiagnoseHLSLAvailability : public DynamicRecursiveASTVisitor {
}
bool VisitCallExpr(CallExpr *CE) override {
+ DiagnoseQuadAllCall(CE);
DiagnoseBarrierCall(CE);
return true;
}
};
+void DiagnoseHLSLAvailability::DiagnoseQuadAllCall(CallExpr *CE) {
+ const FunctionDecl *FD = CE->getDirectCallee();
+ if (!FD || FD->getBuiltinID() != Builtin::BI__builtin_hlsl_quad_all)
+ return;
+
+ switch (GetCurrentShaderEnvironment()) {
+ case llvm::Triple::UnknownEnvironment:
+ case llvm::Triple::Pixel:
+ case llvm::Triple::Compute:
+ case llvm::Triple::Mesh:
+ case llvm::Triple::Amplification:
+ return;
+ default:
+ SemaRef.Diag(CE->getExprLoc(), diag::err_hlsl_builtin_unsupported_in_stage)
+ << FD
+ << llvm::Triple::getEnvironmentTypeName(GetCurrentShaderEnvironment());
+ }
+}
+
uint64_t DiagnoseHLSLAvailability::DiagnoseBarrierGroupMemory(
Expr *MemoryArg, uint64_t MemoryFlags, bool HasVisibleGroup,
bool IsAllMemory) {
diff --git a/clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl b/clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl
new file mode 100644
index 0000000000000..1832f01eaaf09
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl
@@ -0,0 +1,86 @@
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.0-library %s -fsyntax-only -verify
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.6-library %s -fsyntax-only -verify \
+// RUN: -fhlsl-strict-availability
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.7-library %s -fsyntax-only -verify
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: spirv-pc-vulkan-library %s -fsyntax-only -verify
+
+bool test_quad_all(bool cond) {
+ return QuadAll(cond); // #quad
+}
+
+export bool exported_quad_all(bool cond) {
+ return test_quad_all(cond);
+}
+
+[shader("pixel")]
+void pixel_main() {
+ test_quad_all(true);
+ (void)__builtin_hlsl_quad_all(true);
+}
+
+[shader("compute")]
+[numthreads(2, 2, 1)]
+void compute_main() {
+ test_quad_all(true);
+ (void)__builtin_hlsl_quad_all(true);
+}
+
+[shader("mesh")]
+[numthreads(2, 2, 1)]
+void mesh_main() {
+ test_quad_all(true);
+}
+
+[shader("amplification")]
+[numthreads(2, 2, 1)]
+void amplification_main() {
+ test_quad_all(true);
+}
+
+// expected-error@#quad {{'QuadAll' is not supported in 'vertex' shaders}}
+[shader("vertex")]
+void vertex_main() {
+ test_quad_all(true);
+ // expected-error at +1 {{'__builtin_hlsl_quad_all' is not supported}}
+ (void)__builtin_hlsl_quad_all(true);
+}
+
+// expected-error@#quad {{'QuadAll' is not supported in 'geometry' shaders}}
+[shader("geometry")]
+void geometry_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'hull' shaders}}
+[shader("hull")]
+void hull_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'domain' shaders}}
+[shader("domain")]
+void domain_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'raygeneration'}}
+[shader("raygeneration")]
+void raygeneration_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'intersection' shaders}}
+[shader("intersection")]
+void intersection_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'anyhit' shaders}}
+[shader("anyhit")]
+void anyhit_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'closesthit' shaders}}
+[shader("closesthit")]
+void closesthit_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'miss' shaders}}
+[shader("miss")]
+void miss_main() { test_quad_all(true); }
+
+// expected-error@#quad {{'QuadAll' is not supported in 'callable' shaders}}
+[shader("callable")]
+void callable_main() { test_quad_all(true); }
>From 047c68196d7e49144a8197ca17f05477719caa7b Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Wed, 11 Nov 2026 14:38:43 -0800
Subject: [PATCH 3/9] add derivative mode if applicable
Co-authored-by: Copilot <223556219+Copilot at users.noreply.github.com>
Copilot-Session: 365638e8-2291-4925-8f1e-1b5c23220818
---
clang/lib/CodeGen/CGHLSLRuntime.cpp | 34 +++++++++
.../builtins/QuadAll-derivative-group.hlsl | 71 +++++++++++++++++++
llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp | 11 +--
.../lib/Target/SPIRV/SPIRVSymbolicOperands.td | 10 ++-
4 files changed, 119 insertions(+), 7 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.cpp b/clang/lib/CodeGen/CGHLSLRuntime.cpp
index e65ada1ee8100..8eccf28f20656 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.cpp
+++ b/clang/lib/CodeGen/CGHLSLRuntime.cpp
@@ -35,6 +35,7 @@
#include "llvm/ADT/Enum.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/ScopeExit.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Frontend/HLSL/HLSLResource.h"
@@ -1000,6 +1001,37 @@ void CGHLSLRuntime::addHLSLBufferLayoutType(const RecordType *StructType,
LayoutTypes[StructType] = LayoutTy;
}
+static void addSPIRVQuadAllExecutionModes(llvm::Module &M) {
+ Function *QuadAll =
+ Intrinsic::getDeclarationIfExists(&M, Intrinsic::spv_quad_all);
+ if (!QuadAll)
+ return;
+
+ SmallVector<Function *, 8> Worklist{QuadAll};
+ SmallPtrSet<Function *, 16> Visited;
+ IRBuilder<> B(M);
+ while (!Worklist.empty()) {
+ Function *F = Worklist.pop_back_val();
+ if (!Visited.insert(F).second)
+ continue;
+
+ if (F->hasFnAttribute("hlsl.numthreads")) {
+ StringRef NumThreads =
+ F->getFnAttribute("hlsl.numthreads").getValueAsString();
+ // DerivativeGroupLinearKHR is 5290; DerivativeGroupQuadsKHR is 5289.
+ unsigned Mode = NumThreads.ends_with(",1,1") ? 5290 : 5289;
+ MDNode *MD = MDNode::get(
+ M.getContext(),
+ {ValueAsMetadata::get(F), ConstantAsMetadata::get(B.getInt32(Mode))});
+ M.getOrInsertNamedMetadata("spirv.ExecutionMode")->addOperand(MD);
+ }
+
+ for (User *U : F->users())
+ if (auto *Call = dyn_cast<CallBase>(U))
+ Worklist.push_back(Call->getFunction());
+ }
+}
+
void CGHLSLRuntime::finishCodeGen() {
auto &TargetOpts = CGM.getTarget().getTargetOpts();
auto &CodeGenOpts = CGM.getCodeGenOpts();
@@ -1044,6 +1076,8 @@ void CGHLSLRuntime::finishCodeGen() {
}
generateGlobalCtorDtorCalls();
+ if (T.isSPIRV())
+ addSPIRVQuadAllExecutionModes(M);
}
void clang::CodeGen::CGHLSLRuntime::setHLSLEntryAttributes(
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
new file mode 100644
index 0000000000000..5fa6afefac66c
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
@@ -0,0 +1,71 @@
+// DEFINE: %{deriv-ext} = +SPV_KHR_compute_shader_derivatives
+// DEFINE: %{quad-exts} = +SPV_KHR_quad_control,+SPV_KHR_maximal_reconvergence
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - \
+// RUN: | FileCheck %s --check-prefix=IR
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.7-library %s -emit-llvm \
+// RUN: -disable-llvm-passes -o - | FileCheck %s --check-prefix=DXIL
+// RUN: %if spirv-registered-target %{ \
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: spirv1.6-unknown-vulkan1.3-library -O3 -S %s -o - \
+// RUN: -mllvm -spirv-ext=%{deriv-ext},%{quad-exts} \
+// RUN: | FileCheck %s --check-prefixes=SPIRV,NATIVE %}
+// RUN: %if spirv-registered-target %{ \
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: spirv1.6-unknown-vulkan1.3-library -O3 -S %s -o - \
+// RUN: -mllvm -spirv-ext=%{deriv-ext},-SPV_KHR_quad_control \
+// RUN: | FileCheck %s --check-prefixes=SPIRV,EMULATED \
+// RUN: --implicit-check-not=OpGroupNonUniformQuadAllKHR %}
+
+RWStructuredBuffer<uint> Out;
+
+bool vote(bool cond) {
+ return QuadAll(cond);
+}
+
+bool helper(bool cond) {
+ return vote(cond);
+}
+
+export bool exported_vote(bool cond) {
+ return helper(cond);
+}
+
+[shader("compute")]
+[numthreads(4, 2, 1)]
+void compute_2d(uint index : SV_GroupIndex) {
+ Out[index] = helper(index != 0) && helper(index < 8);
+}
+
+[shader("compute")]
+[numthreads(8, 1, 1)]
+void compute_1d(uint index : SV_GroupIndex) {
+ Out[index] = __builtin_hlsl_quad_all(index != 0);
+}
+
+[shader("pixel")]
+void pixel_main() {
+ Out[0] = helper(Out[0] != 0);
+}
+
+[shader("compute")]
+[numthreads(4, 2, 1)]
+void no_vote(uint index : SV_GroupIndex) {
+ Out[index] = index;
+}
+
+// IR: !spirv.ExecutionMode = !{!{{[0-9]+}}, !{{[0-9]+}}}
+// IR-DAG: !{{[0-9]+}} = !{ptr @compute_2d, i32 5289}
+// IR-DAG: !{{[0-9]+}} = !{ptr @compute_1d, i32 5290}
+// DXIL-NOT: !spirv.ExecutionMode
+
+// SPIRV-DAG: OpCapability ComputeDerivativeGroupQuadsKHR
+// SPIRV-DAG: OpCapability ComputeDerivativeGroupLinearKHR
+// SPIRV-DAG: OpExtension "SPV_KHR_compute_shader_derivatives"
+// SPIRV-DAG: OpEntryPoint GLCompute %[[#QUADS:]] "compute_2d"
+// SPIRV-DAG: OpEntryPoint GLCompute %[[#LINEAR:]] "compute_1d"
+// SPIRV-DAG: OpExecutionMode %[[#QUADS]] DerivativeGroupQuadsKHR
+// SPIRV-DAG: OpExecutionMode %[[#LINEAR]] DerivativeGroupLinearKHR
+// NATIVE: OpGroupNonUniformQuadAllKHR
+// EMULATED: OpGroupNonUniformQuadSwap
diff --git a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
index c3766765bc4bc..59df96461dd5f 100644
--- a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
@@ -872,12 +872,13 @@ void SPIRV::RequirementHandler::getAndAddRequirements(
addQuadControlRequirements(*this, ST,
getSymbolicOperandMnemonic(Category, i));
if (Category == SPIRV::OperandCategory::ExecutionModeOperand &&
- i == SPIRV::ExecutionMode::DerivativeGroupQuadsKHR &&
+ (i == SPIRV::ExecutionMode::DerivativeGroupQuadsKHR ||
+ i == SPIRV::ExecutionMode::DerivativeGroupLinearKHR) &&
!ST.canUseExtension(SPIRV::Extension::SPV_KHR_compute_shader_derivatives))
- report_fatal_error(
- "DerivativeGroupQuadsKHR requires the following SPIR-V extension: "
- "SPV_KHR_compute_shader_derivatives",
- false);
+ report_fatal_error(Twine(getSymbolicOperandMnemonic(Category, i)) +
+ " requires the following SPIR-V extension: "
+ "SPV_KHR_compute_shader_derivatives",
+ false);
addRequirements(getSymbolicOperandRequirements(Category, i, ST, *this));
}
diff --git a/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td b/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
index 9c8c6d1051f86..e8a9edad64db2 100644
--- a/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
+++ b/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
@@ -577,7 +577,11 @@ defm ComputeDerivativeGroupQuadsKHR
def ComputeDerivativeGroupQuadsNV
: Capability<"ComputeDerivativeGroupQuadsNV", 5288>;
defm DemoteToHelperInvocation : CapabilityOperand<5379, 0x10600, 0, [SPV_EXT_demote_to_helper_invocation], []>;
-defm ComputeDerivativeGroupLinearNV : CapabilityOperand<5350, 0, 0, [], []>;
+defm ComputeDerivativeGroupLinearKHR
+ : CapabilityOperand<5350, 0, 0, [SPV_KHR_compute_shader_derivatives],
+ [Shader]>;
+def ComputeDerivativeGroupLinearNV
+ : Capability<"ComputeDerivativeGroupLinearNV", 5350>;
defm FragmentDensityEXT : CapabilityOperand<5291, 0, 0, [], [Shader]>;
defm PhysicalStorageBufferAddressesEXT : CapabilityOperand<5347, 0, 0, [], [Shader]>;
defm CooperativeMatrixNV : CapabilityOperand<5357, 0, 0, [], [Shader]>;
@@ -864,7 +868,9 @@ defm OutputLinesNV : ExecutionModeOperand<5269, [MeshShadingNV]>;
defm DerivativeGroupQuadsKHR
: ExecutionModeOperand<5289, [ComputeDerivativeGroupQuadsKHR]>;
def DerivativeGroupQuadsNV : ExecutionMode<"DerivativeGroupQuadsNV", 5289>;
-defm DerivativeGroupLinearNV : ExecutionModeOperand<5290, [ComputeDerivativeGroupLinearNV]>;
+defm DerivativeGroupLinearKHR
+ : ExecutionModeOperand<5290, [ComputeDerivativeGroupLinearKHR]>;
+def DerivativeGroupLinearNV : ExecutionMode<"DerivativeGroupLinearNV", 5290>;
defm OutputTrianglesNV : ExecutionModeOperand<5298, [MeshShadingNV]>;
defm RoundingModeRTPINTEL : ExecutionModeOperand<5620, [RoundToInfinityINTEL]>;
defm RoundingModeRTNINTEL : ExecutionModeOperand<5621, [RoundToInfinityINTEL]>;
>From cd33fd3ffc87385d5fc3537799a2cd6a7b6d7e6e Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Thu, 12 Nov 2026 11:34:03 -0800
Subject: [PATCH 4/9] address Deric, fix diag msg
---
.../clang/Basic/DiagnosticSemaKinds.td | 3 +-
clang/lib/Sema/SemaHLSL.cpp | 27 +++++++++------
....errors.hlsl => QuadAll-stage-errors.hlsl} | 34 +++++++++++++------
3 files changed, 41 insertions(+), 23 deletions(-)
rename clang/test/SemaHLSL/BuiltIns/{QuadAll-stages.errors.hlsl => QuadAll-stage-errors.hlsl} (71%)
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 1bd9316499262..b0356b82f0260 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -13987,7 +13987,8 @@ def err_hlsl_builtin_requires_sm : Error<
"%0 requires shader model %1 or newer">;
def err_hlsl_builtin_unsupported_in_stage : Error<
- "%0 is not supported in '%1' shaders">;
+ "%0 is not supported in '%1' shaders, requires "
+ "%select{|one of the following: }2%3">;
def err_hlsl_export_not_on_function : Error<
"export declaration can only be used on functions">;
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 53f57970c2599..d3b88bbdfb614 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -3238,18 +3238,23 @@ void DiagnoseHLSLAvailability::DiagnoseQuadAllCall(CallExpr *CE) {
if (!FD || FD->getBuiltinID() != Builtin::BI__builtin_hlsl_quad_all)
return;
- switch (GetCurrentShaderEnvironment()) {
- case llvm::Triple::UnknownEnvironment:
- case llvm::Triple::Pixel:
- case llvm::Triple::Compute:
- case llvm::Triple::Mesh:
- case llvm::Triple::Amplification:
+ const llvm::Triple::EnvironmentType AllowedStages[] = {
+ llvm::Triple::Pixel, llvm::Triple::Compute, llvm::Triple::Mesh,
+ llvm::Triple::Amplification};
+ llvm::Triple::EnvironmentType Stage = GetCurrentShaderEnvironment();
+ if (Stage == llvm::Triple::UnknownEnvironment ||
+ llvm::is_contained(AllowedStages, Stage))
return;
- default:
- SemaRef.Diag(CE->getExprLoc(), diag::err_hlsl_builtin_unsupported_in_stage)
- << FD
- << llvm::Triple::getEnvironmentTypeName(GetCurrentShaderEnvironment());
- }
+
+ SmallVector<StringRef, 4> StageStrings;
+ llvm::transform(AllowedStages, std::back_inserter(StageStrings),
+ [](llvm::Triple::EnvironmentType ST) {
+ return StringRef(
+ HLSLShaderAttr::ConvertEnvironmentTypeToStr(ST));
+ });
+ SemaRef.Diag(CE->getExprLoc(), diag::err_hlsl_builtin_unsupported_in_stage)
+ << FD << llvm::Triple::getEnvironmentTypeName(Stage)
+ << (std::size(AllowedStages) != 1) << join(StageStrings, ", ");
}
uint64_t DiagnoseHLSLAvailability::DiagnoseBarrierGroupMemory(
diff --git a/clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl b/clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
similarity index 71%
rename from clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl
rename to clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
index 1832f01eaaf09..1c375034d5554 100644
--- a/clang/test/SemaHLSL/BuiltIns/QuadAll-stages.errors.hlsl
+++ b/clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
@@ -41,46 +41,58 @@ void amplification_main() {
test_quad_all(true);
}
-// expected-error@#quad {{'QuadAll' is not supported in 'vertex' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'vertex' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("vertex")]
void vertex_main() {
test_quad_all(true);
- // expected-error at +1 {{'__builtin_hlsl_quad_all' is not supported}}
+ // expected-error at +3 {{'__builtin_hlsl_quad_all' is not supported in \
+'vertex' shaders, requires one of the following: \
+pixel, compute, mesh, amplification}}
(void)__builtin_hlsl_quad_all(true);
}
-// expected-error@#quad {{'QuadAll' is not supported in 'geometry' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'geometry' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("geometry")]
void geometry_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'hull' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'hull' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("hull")]
void hull_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'domain' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'domain' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("domain")]
void domain_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'raygeneration'}}
+// expected-error@#quad {{'QuadAll' is not supported in 'raygeneration' \
+shaders, requires one of the following: pixel, compute, mesh, amplification}}
[shader("raygeneration")]
void raygeneration_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'intersection' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'intersection' \
+shaders, requires one of the following: pixel, compute, mesh, amplification}}
[shader("intersection")]
void intersection_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'anyhit' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'anyhit' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("anyhit")]
void anyhit_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'closesthit' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'closesthit' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("closesthit")]
void closesthit_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'miss' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'miss' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("miss")]
void miss_main() { test_quad_all(true); }
-// expected-error@#quad {{'QuadAll' is not supported in 'callable' shaders}}
+// expected-error@#quad {{'QuadAll' is not supported in 'callable' shaders, \
+requires one of the following: pixel, compute, mesh, amplification}}
[shader("callable")]
void callable_main() { test_quad_all(true); }
>From beef838c078c5433170e6f2282ce39e465cd545c Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 13 Nov 2026 14:18:32 -0800
Subject: [PATCH 5/9] address Helena
---
.../clang/Basic/DiagnosticFrontendKinds.td | 4 +
.../clang/Basic/DiagnosticSemaKinds.td | 4 -
clang/lib/CodeGen/CGHLSLRuntime.cpp | 33 ++++++-
clang/lib/Sema/SemaHLSL.cpp | 26 -----
.../QuadAll-derivative-group-errors.hlsl | 57 +++++++++++
.../builtins/QuadAll-derivative-group.hlsl | 23 -----
.../BuiltIns/QuadAll-stage-errors.hlsl | 98 -------------------
.../derivative-group-linear.ll | 30 ++++++
8 files changed, 119 insertions(+), 156 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
delete mode 100644 clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
create mode 100644 llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/derivative-group-linear.ll
diff --git a/clang/include/clang/Basic/DiagnosticFrontendKinds.td b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
index b51aa3ff0a0cb..be82686552733 100644
--- a/clang/include/clang/Basic/DiagnosticFrontendKinds.td
+++ b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
@@ -504,6 +504,10 @@ def warn_hlsl_langstd_minimal :
InGroup<HLSLDXCCompat>;
def err_hlsl_signature_packing : Error<
"failed to pack %select{input|output}0 signature: %1">;
+def err_hlsl_spirv_quad_all_numthreads : Error<
+ "entry point '%0' uses QuadAll with invalid numthreads: "
+ "%select{DerivativeGroupQuadsKHR requires X and Y to be multiples of 2|"
+ "DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4}1">;
def warn_hlsl_signature_packing_ignored : Warning<
"semantic signature packing mode '%0' is ignored as "
"%select{vertex shader inputs|pixel shader outputs}1 always use "
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index b0356b82f0260..3e56980e22902 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -13986,10 +13986,6 @@ def err_hlsl_atomic_arg_addr_space : Error<
def err_hlsl_builtin_requires_sm : Error<
"%0 requires shader model %1 or newer">;
-def err_hlsl_builtin_unsupported_in_stage : Error<
- "%0 is not supported in '%1' shaders, requires "
- "%select{|one of the following: }2%3">;
-
def err_hlsl_export_not_on_function : Error<
"export declaration can only be used on functions">;
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.cpp b/clang/lib/CodeGen/CGHLSLRuntime.cpp
index 8eccf28f20656..3f9e01716ffb6 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.cpp
+++ b/clang/lib/CodeGen/CGHLSLRuntime.cpp
@@ -68,6 +68,11 @@ using llvm::hlsl::CBufferRowSizeInBytes;
namespace {
+enum class SPIRVExecutionMode : uint32_t {
+ DerivativeGroupQuadsKHR = 5289,
+ DerivativeGroupLinearKHR = 5290
+};
+
void addDxilValVersion(StringRef ValVersionStr, llvm::Module &M) {
// The validation of ValVersionStr is done at HLSLToolChain::TranslateArgs.
// Assume ValVersionStr is legal here.
@@ -1001,7 +1006,8 @@ void CGHLSLRuntime::addHLSLBufferLayoutType(const RecordType *StructType,
LayoutTypes[StructType] = LayoutTy;
}
-static void addSPIRVQuadAllExecutionModes(llvm::Module &M) {
+static void addSPIRVQuadAllExecutionModes(CodeGenModule &CGM) {
+ llvm::Module &M = CGM.getModule();
Function *QuadAll =
Intrinsic::getDeclarationIfExists(&M, Intrinsic::spv_quad_all);
if (!QuadAll)
@@ -1018,11 +1024,28 @@ static void addSPIRVQuadAllExecutionModes(llvm::Module &M) {
if (F->hasFnAttribute("hlsl.numthreads")) {
StringRef NumThreads =
F->getFnAttribute("hlsl.numthreads").getValueAsString();
- // DerivativeGroupLinearKHR is 5290; DerivativeGroupQuadsKHR is 5289.
- unsigned Mode = NumThreads.ends_with(",1,1") ? 5290 : 5289;
+ SmallVector<StringRef, 3> Dimensions;
+ NumThreads.split(Dimensions, ',');
+ assert(Dimensions.size() == 3 && "expected three numthreads dimensions");
+ std::array<uint32_t, 3> Threads;
+ for (auto [Dimension, Thread] : llvm::zip(Dimensions, Threads)) {
+ [[maybe_unused]] bool Invalid = Dimension.getAsInteger(10, Thread);
+ assert(!Invalid && "expected an integer numthreads dimension");
+ }
+ bool IsLinear = Threads[1] == 1 && Threads[2] == 1;
+ if (IsLinear ? (uint64_t(Threads[0]) * Threads[1] * Threads[2]) % 4 != 0
+ : Threads[0] % 2 != 0 || Threads[1] % 2 != 0) {
+ CGM.getDiags().Report(diag::err_hlsl_spirv_quad_all_numthreads)
+ << F->getName() << IsLinear;
+ continue;
+ }
+ SPIRVExecutionMode Mode =
+ IsLinear ? SPIRVExecutionMode::DerivativeGroupLinearKHR
+ : SPIRVExecutionMode::DerivativeGroupQuadsKHR;
MDNode *MD = MDNode::get(
M.getContext(),
- {ValueAsMetadata::get(F), ConstantAsMetadata::get(B.getInt32(Mode))});
+ {ValueAsMetadata::get(F),
+ ConstantAsMetadata::get(B.getInt32(llvm::to_underlying(Mode)))});
M.getOrInsertNamedMetadata("spirv.ExecutionMode")->addOperand(MD);
}
@@ -1077,7 +1100,7 @@ void CGHLSLRuntime::finishCodeGen() {
generateGlobalCtorDtorCalls();
if (T.isSPIRV())
- addSPIRVQuadAllExecutionModes(M);
+ addSPIRVQuadAllExecutionModes(CGM);
}
void clang::CodeGen::CGHLSLRuntime::setHLSLEntryAttributes(
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index d3b88bbdfb614..506ac8e13fe30 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -3191,7 +3191,6 @@ class DiagnoseHLSLAvailability : public DynamicRecursiveASTVisitor {
SourceRange Range);
const AvailabilityAttr *FindAvailabilityAttr(const Decl *D);
bool HasMatchingEnvironmentOrNone(const AvailabilityAttr *AA);
- void DiagnoseQuadAllCall(CallExpr *CE);
void DiagnoseBarrierCall(CallExpr *CE);
uint64_t DiagnoseBarrierGroupMemory(Expr *MemoryArg, uint64_t MemoryFlags,
bool HasVisibleGroup, bool IsAllMemory);
@@ -3227,36 +3226,11 @@ class DiagnoseHLSLAvailability : public DynamicRecursiveASTVisitor {
}
bool VisitCallExpr(CallExpr *CE) override {
- DiagnoseQuadAllCall(CE);
DiagnoseBarrierCall(CE);
return true;
}
};
-void DiagnoseHLSLAvailability::DiagnoseQuadAllCall(CallExpr *CE) {
- const FunctionDecl *FD = CE->getDirectCallee();
- if (!FD || FD->getBuiltinID() != Builtin::BI__builtin_hlsl_quad_all)
- return;
-
- const llvm::Triple::EnvironmentType AllowedStages[] = {
- llvm::Triple::Pixel, llvm::Triple::Compute, llvm::Triple::Mesh,
- llvm::Triple::Amplification};
- llvm::Triple::EnvironmentType Stage = GetCurrentShaderEnvironment();
- if (Stage == llvm::Triple::UnknownEnvironment ||
- llvm::is_contained(AllowedStages, Stage))
- return;
-
- SmallVector<StringRef, 4> StageStrings;
- llvm::transform(AllowedStages, std::back_inserter(StageStrings),
- [](llvm::Triple::EnvironmentType ST) {
- return StringRef(
- HLSLShaderAttr::ConvertEnvironmentTypeToStr(ST));
- });
- SemaRef.Diag(CE->getExprLoc(), diag::err_hlsl_builtin_unsupported_in_stage)
- << FD << llvm::Triple::getEnvironmentTypeName(Stage)
- << (std::size(AllowedStages) != 1) << join(StageStrings, ", ");
-}
-
uint64_t DiagnoseHLSLAvailability::DiagnoseBarrierGroupMemory(
Expr *MemoryArg, uint64_t MemoryFlags, bool HasVisibleGroup,
bool IsAllMemory) {
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
new file mode 100644
index 0000000000000..29574af763267
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
@@ -0,0 +1,57 @@
+// RUN: not %clang_cc1 -finclude-default-header -triple \
+// RUN: spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes \
+// RUN: -o %t 2>&1 | FileCheck %s \
+// RUN: --implicit-check-not="entry point 'valid" \
+// RUN: --implicit-check-not="entry point 'no_vote'"
+// RUN: %clang_cc1 -finclude-default-header -triple \
+// RUN: dxil-pc-shadermodel6.7-library %s -emit-llvm \
+// RUN: -disable-llvm-passes -o %t
+
+bool helper(bool cond) {
+ return QuadAll(cond);
+}
+
+[shader("compute")]
+[numthreads(4, 1, 1)]
+void valid_linear() {
+ helper(true);
+}
+
+[shader("compute")]
+[numthreads(2, 2, 3)]
+void valid_quads() {
+ helper(true);
+}
+
+[shader("compute")]
+[numthreads(6, 1, 1)]
+void invalid_linear() {
+ (void)__builtin_hlsl_quad_all(true);
+}
+
+[shader("compute")]
+[numthreads(3, 4, 1)]
+void invalid_quad_x() {
+ helper(true);
+}
+
+[shader("compute")]
+[numthreads(4, 3, 1)]
+void invalid_quad_y() {
+ helper(true);
+}
+
+[shader("compute")]
+[numthreads(2, 1, 2)]
+void invalid_quad_3d() {
+ helper(true);
+}
+
+[shader("compute")]
+[numthreads(3, 1, 1)]
+void no_vote() {}
+
+// CHECK-DAG: error: entry point 'invalid_linear' uses QuadAll with invalid numthreads: DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4
+// CHECK-DAG: error: entry point 'invalid_quad_x' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: error: entry point 'invalid_quad_y' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: error: entry point 'invalid_quad_3d' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
index 5fa6afefac66c..52f33e20ca09d 100644
--- a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group.hlsl
@@ -1,22 +1,9 @@
-// DEFINE: %{deriv-ext} = +SPV_KHR_compute_shader_derivatives
-// DEFINE: %{quad-exts} = +SPV_KHR_quad_control,+SPV_KHR_maximal_reconvergence
// RUN: %clang_cc1 -finclude-default-header -triple \
// RUN: spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - \
// RUN: | FileCheck %s --check-prefix=IR
// RUN: %clang_cc1 -finclude-default-header -triple \
// RUN: dxil-pc-shadermodel6.7-library %s -emit-llvm \
// RUN: -disable-llvm-passes -o - | FileCheck %s --check-prefix=DXIL
-// RUN: %if spirv-registered-target %{ \
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: spirv1.6-unknown-vulkan1.3-library -O3 -S %s -o - \
-// RUN: -mllvm -spirv-ext=%{deriv-ext},%{quad-exts} \
-// RUN: | FileCheck %s --check-prefixes=SPIRV,NATIVE %}
-// RUN: %if spirv-registered-target %{ \
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: spirv1.6-unknown-vulkan1.3-library -O3 -S %s -o - \
-// RUN: -mllvm -spirv-ext=%{deriv-ext},-SPV_KHR_quad_control \
-// RUN: | FileCheck %s --check-prefixes=SPIRV,EMULATED \
-// RUN: --implicit-check-not=OpGroupNonUniformQuadAllKHR %}
RWStructuredBuffer<uint> Out;
@@ -59,13 +46,3 @@ void no_vote(uint index : SV_GroupIndex) {
// IR-DAG: !{{[0-9]+}} = !{ptr @compute_2d, i32 5289}
// IR-DAG: !{{[0-9]+}} = !{ptr @compute_1d, i32 5290}
// DXIL-NOT: !spirv.ExecutionMode
-
-// SPIRV-DAG: OpCapability ComputeDerivativeGroupQuadsKHR
-// SPIRV-DAG: OpCapability ComputeDerivativeGroupLinearKHR
-// SPIRV-DAG: OpExtension "SPV_KHR_compute_shader_derivatives"
-// SPIRV-DAG: OpEntryPoint GLCompute %[[#QUADS:]] "compute_2d"
-// SPIRV-DAG: OpEntryPoint GLCompute %[[#LINEAR:]] "compute_1d"
-// SPIRV-DAG: OpExecutionMode %[[#QUADS]] DerivativeGroupQuadsKHR
-// SPIRV-DAG: OpExecutionMode %[[#LINEAR]] DerivativeGroupLinearKHR
-// NATIVE: OpGroupNonUniformQuadAllKHR
-// EMULATED: OpGroupNonUniformQuadSwap
diff --git a/clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
deleted file mode 100644
index 1c375034d5554..0000000000000
--- a/clang/test/SemaHLSL/BuiltIns/QuadAll-stage-errors.hlsl
+++ /dev/null
@@ -1,98 +0,0 @@
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: dxil-pc-shadermodel6.0-library %s -fsyntax-only -verify
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: dxil-pc-shadermodel6.6-library %s -fsyntax-only -verify \
-// RUN: -fhlsl-strict-availability
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: dxil-pc-shadermodel6.7-library %s -fsyntax-only -verify
-// RUN: %clang_cc1 -finclude-default-header -triple \
-// RUN: spirv-pc-vulkan-library %s -fsyntax-only -verify
-
-bool test_quad_all(bool cond) {
- return QuadAll(cond); // #quad
-}
-
-export bool exported_quad_all(bool cond) {
- return test_quad_all(cond);
-}
-
-[shader("pixel")]
-void pixel_main() {
- test_quad_all(true);
- (void)__builtin_hlsl_quad_all(true);
-}
-
-[shader("compute")]
-[numthreads(2, 2, 1)]
-void compute_main() {
- test_quad_all(true);
- (void)__builtin_hlsl_quad_all(true);
-}
-
-[shader("mesh")]
-[numthreads(2, 2, 1)]
-void mesh_main() {
- test_quad_all(true);
-}
-
-[shader("amplification")]
-[numthreads(2, 2, 1)]
-void amplification_main() {
- test_quad_all(true);
-}
-
-// expected-error@#quad {{'QuadAll' is not supported in 'vertex' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("vertex")]
-void vertex_main() {
- test_quad_all(true);
- // expected-error at +3 {{'__builtin_hlsl_quad_all' is not supported in \
-'vertex' shaders, requires one of the following: \
-pixel, compute, mesh, amplification}}
- (void)__builtin_hlsl_quad_all(true);
-}
-
-// expected-error@#quad {{'QuadAll' is not supported in 'geometry' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("geometry")]
-void geometry_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'hull' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("hull")]
-void hull_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'domain' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("domain")]
-void domain_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'raygeneration' \
-shaders, requires one of the following: pixel, compute, mesh, amplification}}
-[shader("raygeneration")]
-void raygeneration_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'intersection' \
-shaders, requires one of the following: pixel, compute, mesh, amplification}}
-[shader("intersection")]
-void intersection_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'anyhit' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("anyhit")]
-void anyhit_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'closesthit' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("closesthit")]
-void closesthit_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'miss' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("miss")]
-void miss_main() { test_quad_all(true); }
-
-// expected-error@#quad {{'QuadAll' is not supported in 'callable' shaders, \
-requires one of the following: pixel, compute, mesh, amplification}}
-[shader("callable")]
-void callable_main() { test_quad_all(true); }
diff --git a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/derivative-group-linear.ll b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/derivative-group-linear.ll
new file mode 100644
index 0000000000000..b4e53e32c82aa
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_compute_shader_derivatives/derivative-group-linear.ll
@@ -0,0 +1,30 @@
+; RUN: llc -O0 -verify-machineinstrs \
+; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
+; RUN: --spirv-ext=+SPV_KHR_compute_shader_derivatives \
+; RUN: %s -o - | FileCheck %s
+; RUN: not llc -O0 -mtriple=spirv1.5-unknown-vulkan-compute \
+; RUN: --spirv-ext=-SPV_KHR_compute_shader_derivatives \
+; RUN: %s -o - 2>&1 | FileCheck %s --check-prefix=NO-EXT
+; RUN: %if spirv-tools %{ llc -O0 \
+; RUN: -mtriple=spirv1.5-unknown-vulkan-compute \
+; RUN: --spirv-ext=+SPV_KHR_compute_shader_derivatives \
+; RUN: %s -o - -filetype=obj | spirv-val --target-env vulkan1.2 %}
+
+; NO-EXT: DerivativeGroupLinearKHR requires the following SPIR-V extension: SPV_KHR_compute_shader_derivatives
+
+; CHECK-DAG: OpCapability Shader
+; CHECK-DAG: OpCapability ComputeDerivativeGroupLinearKHR
+; CHECK: OpExtension "SPV_KHR_compute_shader_derivatives"
+; CHECK-NOT: OpExtension
+; CHECK: OpEntryPoint GLCompute %[[#MAIN:]] "main"
+; CHECK: OpExecutionMode %[[#MAIN]] DerivativeGroupLinearKHR
+; CHECK: OpExecutionMode %[[#MAIN]] LocalSize 8 1 1
+
+define void @main() #0 {
+ ret void
+}
+
+attributes #0 = { "hlsl.numthreads"="8,1,1" "hlsl.shader"="compute" }
+
+!spirv.ExecutionMode = !{!0}
+!0 = !{ptr @main, i32 5290}
>From dc26ed1ba8a2e07acaa6b4432553b9210fe3a89a Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 13 Nov 2026 15:02:22 -0800
Subject: [PATCH 6/9] address deric, extract helper function for identifying
derivative mode
---
clang/lib/CodeGen/CGHLSLRuntime.cpp | 46 ++++++++++++++++-------------
1 file changed, 26 insertions(+), 20 deletions(-)
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.cpp b/clang/lib/CodeGen/CGHLSLRuntime.cpp
index 3f9e01716ffb6..1958761ffec1b 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.cpp
+++ b/clang/lib/CodeGen/CGHLSLRuntime.cpp
@@ -1006,6 +1006,28 @@ void CGHLSLRuntime::addHLSLBufferLayoutType(const RecordType *StructType,
LayoutTypes[StructType] = LayoutTy;
}
+static std::optional<SPIRVExecutionMode>
+getSPIRVDerivativeGroupMode(CodeGenModule &CGM, const Function &F) {
+ StringRef NumThreads = F.getFnAttribute("hlsl.numthreads").getValueAsString();
+ SmallVector<StringRef, 3> Dimensions;
+ NumThreads.split(Dimensions, ',');
+ assert(Dimensions.size() == 3 && "expected three numthreads dimensions");
+ std::array<uint32_t, 3> Threads;
+ for (auto [Dimension, Thread] : llvm::zip(Dimensions, Threads)) {
+ [[maybe_unused]] bool Invalid = Dimension.getAsInteger(10, Thread);
+ assert(!Invalid && "expected an integer numthreads dimension");
+ }
+ bool IsLinear = Threads[1] == 1 && Threads[2] == 1;
+ if (IsLinear ? (uint64_t(Threads[0]) * Threads[1] * Threads[2]) % 4 != 0
+ : Threads[0] % 2 != 0 || Threads[1] % 2 != 0) {
+ CGM.getDiags().Report(diag::err_hlsl_spirv_quad_all_numthreads)
+ << F.getName() << IsLinear;
+ return std::nullopt;
+ }
+ return IsLinear ? SPIRVExecutionMode::DerivativeGroupLinearKHR
+ : SPIRVExecutionMode::DerivativeGroupQuadsKHR;
+}
+
static void addSPIRVQuadAllExecutionModes(CodeGenModule &CGM) {
llvm::Module &M = CGM.getModule();
Function *QuadAll =
@@ -1022,30 +1044,14 @@ static void addSPIRVQuadAllExecutionModes(CodeGenModule &CGM) {
continue;
if (F->hasFnAttribute("hlsl.numthreads")) {
- StringRef NumThreads =
- F->getFnAttribute("hlsl.numthreads").getValueAsString();
- SmallVector<StringRef, 3> Dimensions;
- NumThreads.split(Dimensions, ',');
- assert(Dimensions.size() == 3 && "expected three numthreads dimensions");
- std::array<uint32_t, 3> Threads;
- for (auto [Dimension, Thread] : llvm::zip(Dimensions, Threads)) {
- [[maybe_unused]] bool Invalid = Dimension.getAsInteger(10, Thread);
- assert(!Invalid && "expected an integer numthreads dimension");
- }
- bool IsLinear = Threads[1] == 1 && Threads[2] == 1;
- if (IsLinear ? (uint64_t(Threads[0]) * Threads[1] * Threads[2]) % 4 != 0
- : Threads[0] % 2 != 0 || Threads[1] % 2 != 0) {
- CGM.getDiags().Report(diag::err_hlsl_spirv_quad_all_numthreads)
- << F->getName() << IsLinear;
+ std::optional<SPIRVExecutionMode> Mode =
+ getSPIRVDerivativeGroupMode(CGM, *F);
+ if (!Mode)
continue;
- }
- SPIRVExecutionMode Mode =
- IsLinear ? SPIRVExecutionMode::DerivativeGroupLinearKHR
- : SPIRVExecutionMode::DerivativeGroupQuadsKHR;
MDNode *MD = MDNode::get(
M.getContext(),
{ValueAsMetadata::get(F),
- ConstantAsMetadata::get(B.getInt32(llvm::to_underlying(Mode)))});
+ ConstantAsMetadata::get(B.getInt32(llvm::to_underlying(*Mode)))});
M.getOrInsertNamedMetadata("spirv.ExecutionMode")->addOperand(MD);
}
>From 3bad2efe250ddfd6be925294ae618fd690c2e5ef Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 13 Nov 2026 15:12:05 -0800
Subject: [PATCH 7/9] generalize diagnostic message and name
---
clang/include/clang/Basic/DiagnosticFrontendKinds.td | 4 ++--
clang/lib/CodeGen/CGHLSLRuntime.cpp | 2 +-
.../builtins/QuadAll-derivative-group-errors.hlsl | 8 ++++----
3 files changed, 7 insertions(+), 7 deletions(-)
diff --git a/clang/include/clang/Basic/DiagnosticFrontendKinds.td b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
index be82686552733..cff8808f7d7b1 100644
--- a/clang/include/clang/Basic/DiagnosticFrontendKinds.td
+++ b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
@@ -504,8 +504,8 @@ def warn_hlsl_langstd_minimal :
InGroup<HLSLDXCCompat>;
def err_hlsl_signature_packing : Error<
"failed to pack %select{input|output}0 signature: %1">;
-def err_hlsl_spirv_quad_all_numthreads : Error<
- "entry point '%0' uses QuadAll with invalid numthreads: "
+def err_hlsl_spirv_derivative_group_numthreads : Error<
+ "entry point '%0' has invalid numthreads: "
"%select{DerivativeGroupQuadsKHR requires X and Y to be multiples of 2|"
"DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4}1">;
def warn_hlsl_signature_packing_ignored : Warning<
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.cpp b/clang/lib/CodeGen/CGHLSLRuntime.cpp
index 1958761ffec1b..b4c3a5278d601 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.cpp
+++ b/clang/lib/CodeGen/CGHLSLRuntime.cpp
@@ -1020,7 +1020,7 @@ getSPIRVDerivativeGroupMode(CodeGenModule &CGM, const Function &F) {
bool IsLinear = Threads[1] == 1 && Threads[2] == 1;
if (IsLinear ? (uint64_t(Threads[0]) * Threads[1] * Threads[2]) % 4 != 0
: Threads[0] % 2 != 0 || Threads[1] % 2 != 0) {
- CGM.getDiags().Report(diag::err_hlsl_spirv_quad_all_numthreads)
+ CGM.getDiags().Report(diag::err_hlsl_spirv_derivative_group_numthreads)
<< F.getName() << IsLinear;
return std::nullopt;
}
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
index 29574af763267..cbc18c4e06e97 100644
--- a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
@@ -51,7 +51,7 @@ void invalid_quad_3d() {
[numthreads(3, 1, 1)]
void no_vote() {}
-// CHECK-DAG: error: entry point 'invalid_linear' uses QuadAll with invalid numthreads: DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4
-// CHECK-DAG: error: entry point 'invalid_quad_x' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
-// CHECK-DAG: error: entry point 'invalid_quad_y' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
-// CHECK-DAG: error: entry point 'invalid_quad_3d' uses QuadAll with invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: error: entry point 'invalid_linear' has invalid numthreads: DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4
+// CHECK-DAG: error: entry point 'invalid_quad_x' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: error: entry point 'invalid_quad_y' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: error: entry point 'invalid_quad_3d' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
>From a5e659b74819d3549df7f60a074021b04c21a519 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 13 Nov 2026 16:25:33 -0800
Subject: [PATCH 8/9] add context that quadall is causing the error
---
clang/include/clang/Basic/DiagnosticFrontendKinds.td | 3 +++
clang/lib/CodeGen/CGHLSLRuntime.cpp | 6 +++++-
.../builtins/QuadAll-derivative-group-errors.hlsl | 4 ++++
3 files changed, 12 insertions(+), 1 deletion(-)
diff --git a/clang/include/clang/Basic/DiagnosticFrontendKinds.td b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
index cff8808f7d7b1..5ec3d1bb1114e 100644
--- a/clang/include/clang/Basic/DiagnosticFrontendKinds.td
+++ b/clang/include/clang/Basic/DiagnosticFrontendKinds.td
@@ -508,6 +508,9 @@ def err_hlsl_spirv_derivative_group_numthreads : Error<
"entry point '%0' has invalid numthreads: "
"%select{DerivativeGroupQuadsKHR requires X and Y to be multiples of 2|"
"DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4}1">;
+def note_hlsl_spirv_derivative_group_requirement : Note<
+ "'%0' implicitly requires a derivative group execution mode "
+ "for SPIR-V entry point '%1'">;
def warn_hlsl_signature_packing_ignored : Warning<
"semantic signature packing mode '%0' is ignored as "
"%select{vertex shader inputs|pixel shader outputs}1 always use "
diff --git a/clang/lib/CodeGen/CGHLSLRuntime.cpp b/clang/lib/CodeGen/CGHLSLRuntime.cpp
index b4c3a5278d601..3be39712b5cab 100644
--- a/clang/lib/CodeGen/CGHLSLRuntime.cpp
+++ b/clang/lib/CodeGen/CGHLSLRuntime.cpp
@@ -1046,8 +1046,12 @@ static void addSPIRVQuadAllExecutionModes(CodeGenModule &CGM) {
if (F->hasFnAttribute("hlsl.numthreads")) {
std::optional<SPIRVExecutionMode> Mode =
getSPIRVDerivativeGroupMode(CGM, *F);
- if (!Mode)
+ if (!Mode) {
+ CGM.getDiags().Report(
+ diag::note_hlsl_spirv_derivative_group_requirement)
+ << "QuadAll" << F->getName();
continue;
+ }
MDNode *MD = MDNode::get(
M.getContext(),
{ValueAsMetadata::get(F),
diff --git a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
index cbc18c4e06e97..03e7962faf061 100644
--- a/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/QuadAll-derivative-group-errors.hlsl
@@ -52,6 +52,10 @@ void invalid_quad_3d() {
void no_vote() {}
// CHECK-DAG: error: entry point 'invalid_linear' has invalid numthreads: DerivativeGroupLinearKHR requires X * Y * Z to be a multiple of 4
+// CHECK-DAG: note: 'QuadAll' implicitly requires a derivative group execution mode for SPIR-V entry point 'invalid_linear'
// CHECK-DAG: error: entry point 'invalid_quad_x' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: note: 'QuadAll' implicitly requires a derivative group execution mode for SPIR-V entry point 'invalid_quad_x'
// CHECK-DAG: error: entry point 'invalid_quad_y' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: note: 'QuadAll' implicitly requires a derivative group execution mode for SPIR-V entry point 'invalid_quad_y'
// CHECK-DAG: error: entry point 'invalid_quad_3d' has invalid numthreads: DerivativeGroupQuadsKHR requires X and Y to be multiples of 2
+// CHECK-DAG: note: 'QuadAll' implicitly requires a derivative group execution mode for SPIR-V entry point 'invalid_quad_3d'
>From 7ea001b2e59883846da41c5132688536e21f413b Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Fri, 13 Nov 2026 17:05:48 -0800
Subject: [PATCH 9/9] address recent validator updates
Co-authored-by: Copilot <223556219+Copilot at users.noreply.github.com>
Copilot-Session: 365638e8-2291-4925-8f1e-1b5c23220818
---
.../cooperative_matrix.ll | 34 +++++++++----------
.../untyped-ptr-coopmatrix.ll | 10 +++---
2 files changed, 22 insertions(+), 22 deletions(-)
diff --git a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_cooperative_matrix/cooperative_matrix.ll b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_cooperative_matrix/cooperative_matrix.ll
index ad9425afe6394..2502883ec35ba 100644
--- a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_cooperative_matrix/cooperative_matrix.ll
+++ b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_cooperative_matrix/cooperative_matrix.ll
@@ -37,17 +37,17 @@
; CHECK: OpCooperativeMatrixMulAddKHR %[[#MatTy6]] %[[#MatA]] %[[#MatB]] %[[#MatC]] MatrixASignedComponentsKHR|MatrixBSignedComponentsKHR
; CHECK: OpCooperativeMatrixStoreKHR
-define spir_kernel void @matr_mult(ptr addrspace(1) align 1 %_arg_accA, ptr addrspace(1) align 1 %_arg_accB, ptr addrspace(1) align 4 %_arg_accC, i64 %_arg_N, i64 %_arg_K) {
+define spir_kernel void @matr_mult(ptr addrspace(1) align 1 %_arg_accA, ptr addrspace(1) align 1 %_arg_accB, ptr addrspace(1) align 4 %_arg_accC, i32 %_arg_N, i32 %_arg_K) {
entry:
%addr1 = alloca target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), align 8
%res = alloca target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), align 8
%m1 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z26__spirv_CompositeConstruct(i32 0)
store target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m1, ptr %addr1, align 8
%accA3 = addrspacecast ptr addrspace(1) %_arg_accA to ptr addrspace(3)
- %m2 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) @_Z32__spirv_CooperativeMatrixLoadKHR_1(ptr addrspace(3) %accA3, i32 0, i64 %_arg_K, i32 1)
+ %m2 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) @_Z32__spirv_CooperativeMatrixLoadKHR_1(ptr addrspace(3) %accA3, i32 0, i32 %_arg_K, i32 1)
%len = tail call spir_func i32 @_Z34__spirv_CooperativeMatrixLengthKHR(target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) %m2)
%accB3 = addrspacecast ptr addrspace(1) %_arg_accB to ptr addrspace(3)
- %m3 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1) @_Z32__spirv_CooperativeMatrixLoadKHR_2(ptr addrspace(3) %accB3, i32 0, i64 0)
+ %m3 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1) @_Z32__spirv_CooperativeMatrixLoadKHR_2(ptr addrspace(3) %accB3, i32 0, i32 0)
%m4 = load target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), ptr %addr1, align 8
%m5 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z34__spirv_CooperativeMatrixMulAddKHR(target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) %m2, target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1) %m3, target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m4, i32 12)
store target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m5, ptr %res, align 8
@@ -55,34 +55,34 @@ entry:
store i64 %r, ptr %addr1, align 8
%accC3 = addrspacecast ptr addrspace(1) %_arg_accC to ptr addrspace(3)
%m6 = load target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), ptr %addr1, align 8
- tail call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(3) %accC3, target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m6, i32 0, i64 %_arg_N, i32 1)
+ tail call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(3) %accC3, target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m6, i32 0, i32 %_arg_N, i32 1)
ret void
}
-define spir_kernel void @matr_mult_with_complex_mangling(ptr addrspace(1) %_arg_A, ptr addrspace(1) %_arg_B, i64 %_arg_M, i64 %_arg_N) {
+define spir_kernel void @matr_mult_with_complex_mangling(ptr addrspace(1) %_arg_A, ptr addrspace(1) %_arg_B, i32 %_arg_M, i32 %_arg_N) {
entry:
- ; __spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)0>* __spirv_CooperativeMatrixLoadKHR<short AS1, short, 8ul, 16ul, (__spv::MatrixUse)0, (__spv::MatrixLayout)1, (__spv::Scope::Flag)3>(short AS1*, __spv::MatrixLayout, unsigned long, int)
- %call = call spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm8ELm16ELN5__spv9MatrixUseE0ELNS1_12MatrixLayoutE1ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) %_arg_A, i32 1, i64 %_arg_M, i32 0)
- ; __spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 16ul, 16ul, (__spv::MatrixUse)1>* __spirv_CooperativeMatrixLoadKHR<short AS1, short, 16ul, 16ul, (__spv::MatrixUse)1, (__spv::MatrixLayout)0, (__spv::Scope::Flag)3>(short AS1*, __spv::MatrixLayout, unsigned long, int)
- %call1 = call spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm16ELm16ELN5__spv9MatrixUseE1ELNS1_12MatrixLayoutE0ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) %_arg_B, i32 0, i64 %_arg_N, i32 0)
+ ; __spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)0>* __spirv_CooperativeMatrixLoadKHR<short AS1, short, 8ul, 16ul, (__spv::MatrixUse)0, (__spv::MatrixLayout)1, (__spv::Scope::Flag)3>(short AS1*, __spv::MatrixLayout, unsigned int, int)
+ %call = call spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm8ELm16ELN5__spv9MatrixUseE0ELNS1_12MatrixLayoutE1ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_ji(ptr addrspace(1) %_arg_A, i32 1, i32 %_arg_M, i32 0)
+ ; __spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 16ul, 16ul, (__spv::MatrixUse)1>* __spirv_CooperativeMatrixLoadKHR<short AS1, short, 16ul, 16ul, (__spv::MatrixUse)1, (__spv::MatrixLayout)0, (__spv::Scope::Flag)3>(short AS1*, __spv::MatrixLayout, unsigned int, int)
+ %call1 = call spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm16ELm16ELN5__spv9MatrixUseE1ELNS1_12MatrixLayoutE0ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_ji(ptr addrspace(1) %_arg_B, i32 0, i32 %_arg_N, i32 0)
; __spv::__spirv_CooperativeMatrixKHR<int, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)2>* __spirv_CompositeConstruct<int, int, 8ul, 16ul, (__spv::MatrixUse)2, (__spv::MatrixLayout)3, (__spv::Scope::Flag)3>(int)
%call2 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) @_Z26__spirv_CompositeConstructIiiLm8ELm16ELN5__spv9MatrixUseE2ELNS0_12MatrixLayoutE3ELNS0_5Scope4FlagE3EEPNS0_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEET_(i32 1)
; __spv::__spirv_CooperativeMatrixKHR<int, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)2>* __spirv_CooperativeMatrixMulAddKHR<short, short, int, int, 8ul, 16ul, 16ul, (__spv::MatrixUse)0, (__spv::MatrixUse)1, (__spv::MatrixUse)2, (__spv::MatrixLayout)0, (__spv::MatrixLayout)0, (__spv::MatrixLayout)0, (__spv::Scope::Flag)3>(__spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)0>*, __spv::__spirv_CooperativeMatrixKHR<short, (__spv::Scope::Flag)3, 16ul, 16ul, (__spv::MatrixUse)1>*, __spv::__spirv_CooperativeMatrixKHR<int, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)2>*, unsigned long)
%call3 = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) @_Z34__spirv_CooperativeMatrixMulAddKHRIssiiLm8ELm16ELm16ELN5__spv9MatrixUseE0ELS1_1ELS1_2ELNS0_12MatrixLayoutE0ELS2_0ELS2_0ELNS0_5Scope4FlagE3EEPNS0_28__spirv_CooperativeMatrixKHRIT2_XT12_EXT3_EXT5_EXT8_EEEPNS5_IT_XT12_EXT3_EXT4_EXT6_EEEPNS5_IT0_XT12_EXT4_EXT5_EXT7_EEEPNS5_IT1_XT12_EXT3_EXT5_EXT8_EEEm(target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0) %call, target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1) %call1, target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) %call2, i64 3)
- ; void __spirv_CooperativeMatrixStoreKHR<int AS1, int, 8ul, 16ul, (__spv::MatrixUse)2, (__spv::MatrixLayout)3, (__spv::Scope::Flag)3>(int AS1*, __spv::__spirv_CooperativeMatrixKHR<int, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)2>*, __spv::MatrixLayout, unsigned long, int)
- call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRIU3AS1iiLm8ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEvPT_PNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEES3_mi(ptr addrspace(1) %_arg_A, target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) %call2, i32 1, i64 %_arg_M, i32 0)
+ ; void __spirv_CooperativeMatrixStoreKHR<int AS1, int, 8ul, 16ul, (__spv::MatrixUse)2, (__spv::MatrixLayout)3, (__spv::Scope::Flag)3>(int AS1*, __spv::__spirv_CooperativeMatrixKHR<int, (__spv::Scope::Flag)3, 8ul, 16ul, (__spv::MatrixUse)2>*, __spv::MatrixLayout, unsigned int, int)
+ call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRIU3AS1iiLm8ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEvPT_PNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEES3_ji(ptr addrspace(1) %_arg_A, target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) %call2, i32 1, i32 %_arg_M, i32 0)
ret void
}
declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z26__spirv_CompositeConstruct(i32)
declare dso_local spir_func i32 @_Z34__spirv_CooperativeMatrixLengthKHR(target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0))
-declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) @_Z32__spirv_CooperativeMatrixLoadKHR_1(ptr addrspace(3), i32, i64, i32)
-declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1) @_Z32__spirv_CooperativeMatrixLoadKHR_2(ptr addrspace(3), i32, i64)
+declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0) @_Z32__spirv_CooperativeMatrixLoadKHR_1(ptr addrspace(3), i32, i32, i32)
+declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1) @_Z32__spirv_CooperativeMatrixLoadKHR_2(ptr addrspace(3), i32, i32)
declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z34__spirv_CooperativeMatrixMulAddKHR(target("spirv.CooperativeMatrixKHR", i32, 3, 12, 48, 0), target("spirv.CooperativeMatrixKHR", i32, 3, 48, 12, 1), target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), i32)
-declare dso_local spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(3), target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), i32, i64, i32)
+declare dso_local spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(3), target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), i32, i32, i32)
declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) @_Z26__spirv_CompositeConstructIiiLm8ELm16ELN5__spv9MatrixUseE2ELNS0_12MatrixLayoutE3ELNS0_5Scope4FlagE3EEPNS0_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEET_(i32)
-declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm8ELm16ELN5__spv9MatrixUseE0ELNS1_12MatrixLayoutE1ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1), i32, i64, i32)
-declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm16ELm16ELN5__spv9MatrixUseE1ELNS1_12MatrixLayoutE0ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1), i32, i64, i32)
+declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm8ELm16ELN5__spv9MatrixUseE0ELNS1_12MatrixLayoutE1ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_ji(ptr addrspace(1), i32, i32, i32)
+declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ssLm16ELm16ELN5__spv9MatrixUseE1ELNS1_12MatrixLayoutE0ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_ji(ptr addrspace(1), i32, i32, i32)
declare dso_local spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2) @_Z34__spirv_CooperativeMatrixMulAddKHRIssiiLm8ELm16ELm16ELN5__spv9MatrixUseE0ELS1_1ELS1_2ELNS0_12MatrixLayoutE0ELS2_0ELS2_0ELNS0_5Scope4FlagE3EEPNS0_28__spirv_CooperativeMatrixKHRIT2_XT12_EXT3_EXT5_EXT8_EEEPNS5_IT_XT12_EXT3_EXT4_EXT6_EEEPNS5_IT0_XT12_EXT4_EXT5_EXT7_EEEPNS5_IT1_XT12_EXT3_EXT5_EXT8_EEEm(target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0), target("spirv.CooperativeMatrixKHR", i16, 3, 16, 16, 1), target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2), i64)
-declare dso_local spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRIU3AS1iiLm8ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEvPT_PNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEES3_mi(ptr addrspace(1), target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2), i32, i64, i32)
+declare dso_local spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRIU3AS1iiLm8ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEvPT_PNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEES3_ji(ptr addrspace(1), target("spirv.CooperativeMatrixKHR", i32, 3, 8, 16, 2), i32, i32, i32)
diff --git a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_untyped_pointers/untyped-ptr-coopmatrix.ll b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_untyped_pointers/untyped-ptr-coopmatrix.ll
index 2b77049a95aaa..bca423233be98 100644
--- a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_untyped_pointers/untyped-ptr-coopmatrix.ll
+++ b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_untyped_pointers/untyped-ptr-coopmatrix.ll
@@ -6,10 +6,10 @@
; CHECK-DAG: %[[#CROSS:]] = OpTypeUntypedPointerKHR CrossWorkgroup
; CHECK: OpCooperativeMatrixLoadKHR
; CHECK: OpCooperativeMatrixStoreKHR
-define spir_kernel void @t(ptr addrspace(1) %src, ptr addrspace(1) %dst, i64 %stride) {
- %m = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z32__spirv_CooperativeMatrixLoadKHR(ptr addrspace(1) %src, i32 0, i64 %stride, i32 1)
- tail call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(1) %dst, target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m, i32 0, i64 %stride, i32 1)
+define spir_kernel void @t(ptr addrspace(1) %src, ptr addrspace(1) %dst, i32 %stride) {
+ %m = tail call spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z32__spirv_CooperativeMatrixLoadKHR(ptr addrspace(1) %src, i32 0, i32 %stride, i32 1)
+ tail call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(1) %dst, target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) %m, i32 0, i32 %stride, i32 1)
ret void
}
-declare spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z32__spirv_CooperativeMatrixLoadKHR(ptr addrspace(1), i32, i64, i32)
-declare spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(1), target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), i32, i64, i32)
+declare spir_func target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2) @_Z32__spirv_CooperativeMatrixLoadKHR(ptr addrspace(1), i32, i32, i32)
+declare spir_func void @_Z33__spirv_CooperativeMatrixStoreKHR(ptr addrspace(1), target("spirv.CooperativeMatrixKHR", i32, 3, 12, 12, 2), i32, i32, i32)
More information about the cfe-commits
mailing list