[llvm] [SPIRV] Add support for the SPV_KHR_subgroup_rotate extension (PR #82374)
Vyacheslav Levytskyy via llvm-commits
llvm-commits at lists.llvm.org
Tue Feb 20 07:57:22 PST 2024
https://github.com/VyacheslavLevytskyy created https://github.com/llvm/llvm-project/pull/82374
This PR adds support for the SPV_KHR_subgroup_rotate extension that enables rotating values across invocations within a subgroup:
* https://github.com/KhronosGroup/SPIRV-Registry/blob/main/extensions/KHR/SPV_KHR_subgroup_rotate.asciidoc
>From 7877b1b0f15855dba9a158d3f13d95e7cbe932b8 Mon Sep 17 00:00:00 2001
From: "Levytskyy, Vyacheslav" <vyacheslav.levytskyy at intel.com>
Date: Tue, 20 Feb 2024 07:53:53 -0800
Subject: [PATCH] add support of the SPV_KHR_subgroup_rotate extension
---
llvm/lib/Target/SPIRV/SPIRVBuiltins.td | 7 +-
llvm/lib/Target/SPIRV/SPIRVInstrInfo.td | 5 +
llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp | 9 +
llvm/lib/Target/SPIRV/SPIRVSubtarget.cpp | 4 +
.../lib/Target/SPIRV/SPIRVSymbolicOperands.td | 1 +
.../subgroup-rotate.ll | 357 ++++++++++++++++++
6 files changed, 382 insertions(+), 1 deletion(-)
create mode 100644 llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_subgroup_rotate/subgroup-rotate.ll
diff --git a/llvm/lib/Target/SPIRV/SPIRVBuiltins.td b/llvm/lib/Target/SPIRV/SPIRVBuiltins.td
index 571cfcfd6e7e5c..bcd27c82439f96 100644
--- a/llvm/lib/Target/SPIRV/SPIRVBuiltins.td
+++ b/llvm/lib/Target/SPIRV/SPIRVBuiltins.td
@@ -619,7 +619,8 @@ class GroupBuiltin<string name, Op operation> {
!eq(operation, OpGroupNonUniformShuffleDown),
!eq(operation, OpGroupBroadcast),
!eq(operation, OpGroupNonUniformBroadcast),
- !eq(operation, OpGroupNonUniformBroadcastFirst));
+ !eq(operation, OpGroupNonUniformBroadcastFirst),
+ !eq(operation, OpGroupNonUniformRotateKHR));
bit HasBoolArg = !or(!and(IsAllOrAny, !eq(IsAllEqual, false)), IsBallot, IsLogical);
}
@@ -877,6 +878,10 @@ defm : DemangledGroupBuiltin<"group_non_uniform_scan_inclusive_logical_xors", Wo
defm : DemangledGroupBuiltin<"group_non_uniform_scan_exclusive_logical_xors", WorkOrSub, OpGroupNonUniformLogicalXor>;
defm : DemangledGroupBuiltin<"group_clustered_reduce_logical_xor", WorkOrSub, OpGroupNonUniformLogicalXor>;
+// cl_khr_subgroup_rotate / SPV_KHR_subgroup_rotate
+defm : DemangledGroupBuiltin<"group_rotate", OnlySub, OpGroupNonUniformRotateKHR>;
+defm : DemangledGroupBuiltin<"group_clustered_rotate", OnlySub, OpGroupNonUniformRotateKHR>;
+
// cl_khr_work_group_uniform_arithmetic / SPV_KHR_uniform_group_instructions
defm : DemangledGroupBuiltin<"group_reduce_imul", OnlyWork, OpGroupIMulKHR>;
defm : DemangledGroupBuiltin<"group_reduce_mulu", OnlyWork, OpGroupIMulKHR>;
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td b/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
index 0f11bc34d176f7..86f65b6320d530 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
+++ b/llvm/lib/Target/SPIRV/SPIRVInstrInfo.td
@@ -765,6 +765,11 @@ def OpGroupNonUniformLogicalAnd: OpGroupNUGroup<"LogicalAnd", 362>;
def OpGroupNonUniformLogicalOr: OpGroupNUGroup<"LogicalOr", 363>;
def OpGroupNonUniformLogicalXor: OpGroupNUGroup<"LogicalXor", 364>;
+// SPV_KHR_subgroup_rotate
+def OpGroupNonUniformRotateKHR: Op<4431, (outs ID:$res),
+ (ins TYPE:$type, ID:$scope, ID:$value, ID:$delta, variable_ops),
+ "$res = OpGroupNonUniformRotateKHR $type $scope $value $delta">;
+
// 3.49.7, Constant-Creation Instructions
// - SPV_INTEL_function_pointers
diff --git a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
index dbda2871e153de..9b9575b9879948 100644
--- a/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVModuleAnalysis.cpp
@@ -1069,6 +1069,15 @@ void addInstrRequirements(const MachineInstr &MI,
Reqs.addCapability(SPIRV::Capability::FunctionPointersINTEL);
}
break;
+ case SPIRV::OpGroupNonUniformRotateKHR:
+ if (!ST.canUseExtension(SPIRV::Extension::SPV_KHR_subgroup_rotate))
+ report_fatal_error("OpGroupNonUniformRotateKHR instruction requires the "
+ "following SPIR-V extension: SPV_KHR_subgroup_rotate",
+ false);
+ Reqs.addExtension(SPIRV::Extension::SPV_KHR_subgroup_rotate);
+ Reqs.addCapability(SPIRV::Capability::GroupNonUniformRotateKHR);
+ Reqs.addCapability(SPIRV::Capability::GroupNonUniform);
+ break;
case SPIRV::OpGroupIMulKHR:
case SPIRV::OpGroupFMulKHR:
case SPIRV::OpGroupBitwiseAndKHR:
diff --git a/llvm/lib/Target/SPIRV/SPIRVSubtarget.cpp b/llvm/lib/Target/SPIRV/SPIRVSubtarget.cpp
index e186154aa408bd..4694363614ef60 100644
--- a/llvm/lib/Target/SPIRV/SPIRVSubtarget.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVSubtarget.cpp
@@ -75,6 +75,10 @@ cl::list<SPIRV::Extension::Extension> Extensions(
"Allows to use the LinkOnceODR linkage type that is to let "
"a function or global variable to be merged with other functions "
"or global variables of the same name when linkage occurs."),
+ clEnumValN(SPIRV::Extension::SPV_KHR_subgroup_rotate,
+ "SPV_KHR_subgroup_rotate",
+ "Adds a new instruction that enables rotating values across "
+ "invocations within a subgroup."),
clEnumValN(SPIRV::Extension::SPV_INTEL_function_pointers,
"SPV_INTEL_function_pointers",
"Allows translation of function pointers.")));
diff --git a/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td b/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
index 4e5ac0d531b2d5..6c36087baa85ed 100644
--- a/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
+++ b/llvm/lib/Target/SPIRV/SPIRVSymbolicOperands.td
@@ -455,6 +455,7 @@ defm BitInstructions : CapabilityOperand<6025, 0, 0, [SPV_KHR_bit_instructions],
defm ExpectAssumeKHR : CapabilityOperand<5629, 0, 0, [SPV_KHR_expect_assume], []>;
defm FunctionPointersINTEL : CapabilityOperand<5603, 0, 0, [SPV_INTEL_function_pointers], []>;
defm IndirectReferencesINTEL : CapabilityOperand<5604, 0, 0, [SPV_INTEL_function_pointers], []>;
+defm GroupNonUniformRotateKHR : CapabilityOperand<6026, 0, 0, [SPV_KHR_subgroup_rotate], [GroupNonUniform]>;
defm AtomicFloat32AddEXT : CapabilityOperand<6033, 0, 0, [SPV_EXT_shader_atomic_float_add], []>;
defm AtomicFloat64AddEXT : CapabilityOperand<6034, 0, 0, [SPV_EXT_shader_atomic_float_add], []>;
defm AtomicFloat16AddEXT : CapabilityOperand<6095, 0, 0, [SPV_EXT_shader_atomic_float16_add], []>;
diff --git a/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_subgroup_rotate/subgroup-rotate.ll b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_subgroup_rotate/subgroup-rotate.ll
new file mode 100644
index 00000000000000..f4af0773084d3e
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/extensions/SPV_KHR_subgroup_rotate/subgroup-rotate.ll
@@ -0,0 +1,357 @@
+; RUN: not llc -O0 -mtriple=spirv32-unknown-unknown %s -o %t.spvt 2>&1 | FileCheck %s --check-prefix=CHECK-ERROR
+; RUN: llc -O0 -mtriple=spirv32-unknown-unknown --spirv-extensions=SPV_KHR_subgroup_rotate %s -o - | FileCheck %s
+; TODO: %if spirv-tools %{ llc -O0 -mtriple=spirv64-unknown-unknown --spirv-extensions=SPV_KHR_subgroup_rotate %s -o - -filetype=obj | spirv-val %}
+
+; CHECK-ERROR: LLVM ERROR: OpGroupNonUniformRotateKHR instruction requires the following SPIR-V extension: SPV_KHR_subgroup_rotate
+
+; CHECK: OpCapability GroupNonUniformRotateKHR
+; CHECK: OpExtension "SPV_KHR_subgroup_rotate"
+
+; CHECK-DAG: %[[TyInt8:.*]] = OpTypeInt 8 0
+; CHECK-DAG: %[[TyInt16:.*]] = OpTypeInt 16 0
+; CHECK-DAG: %[[TyInt32:.*]] = OpTypeInt 32 0
+; CHECK-DAG: %[[TyInt64:.*]] = OpTypeInt 64 0
+; CHECK-DAG: %[[TyFloat:.*]] = OpTypeFloat 32
+; CHECK-DAG: %[[TyHalf:.*]] = OpTypeFloat 16
+; CHECK-DAG: %[[TyDouble:.*]] = OpTypeFloat 64
+; CHECK-DAG: %[[ScopeSubgroup:.*]] = OpConstant %[[TyInt32]] 3
+; CHECK-DAG: %[[ConstInt2:.*]] = OpConstant %[[TyInt32]] 2
+; CHECK-DAG: %[[ConstInt4:.*]] = OpConstant %[[TyInt32]] 4
+
+target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir"
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateChar(ptr addrspace(1) noundef align 1 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !5 !kernel_arg_base_type !5 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i8, align 1
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i8 0, ptr %v, align 1
+ %0 = load i8, ptr %v, align 1
+; CHECK: OpGroupNonUniformRotateKHR %[[TyInt8]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func signext i8 @_Z16sub_group_rotateci(i8 noundef signext %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i8, ptr addrspace(1) %1, i32 0
+ store i8 %call, ptr addrspace(1) %arrayidx, align 1
+ %2 = load i8, ptr %v, align 1
+; CHECK: OpGroupNonUniformRotateKHR %[[TyInt8]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func signext i8 @_Z26sub_group_clustered_rotatecij(i8 noundef signext %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i8, ptr addrspace(1) %3, i32 1
+ store i8 %call1, ptr addrspace(1) %arrayidx2, align 1
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func signext i8 @_Z16sub_group_rotateci(i8 noundef signext, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func signext i8 @_Z26sub_group_clustered_rotatecij(i8 noundef signext, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateUChar(ptr addrspace(1) noundef align 1 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !7 !kernel_arg_base_type !7 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i8, align 1
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i8 0, ptr %v, align 1
+ %0 = load i8, ptr %v, align 1
+; CHECK: OpGroupNonUniformRotateKHR %[[TyInt8]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func zeroext i8 @_Z16sub_group_rotatehi(i8 noundef zeroext %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i8, ptr addrspace(1) %1, i32 0
+ store i8 %call, ptr addrspace(1) %arrayidx, align 1
+ %2 = load i8, ptr %v, align 1
+; CHECK: OpGroupNonUniformRotateKHR %[[TyInt8]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func zeroext i8 @_Z26sub_group_clustered_rotatehij(i8 noundef zeroext %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i8, ptr addrspace(1) %3, i32 1
+ store i8 %call1, ptr addrspace(1) %arrayidx2, align 1
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func zeroext i8 @_Z16sub_group_rotatehi(i8 noundef zeroext, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func zeroext i8 @_Z26sub_group_clustered_rotatehij(i8 noundef zeroext, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateShort(ptr addrspace(1) noundef align 2 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !8 !kernel_arg_base_type !8 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i16, align 2
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i16 0, ptr %v, align 2
+ %0 = load i16, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt16]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func signext i16 @_Z16sub_group_rotatesi(i16 noundef signext %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i16, ptr addrspace(1) %1, i32 0
+ store i16 %call, ptr addrspace(1) %arrayidx, align 2
+ %2 = load i16, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt16]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func signext i16 @_Z26sub_group_clustered_rotatesij(i16 noundef signext %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i16, ptr addrspace(1) %3, i32 1
+ store i16 %call1, ptr addrspace(1) %arrayidx2, align 2
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func signext i16 @_Z16sub_group_rotatesi(i16 noundef signext, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func signext i16 @_Z26sub_group_clustered_rotatesij(i16 noundef signext, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateUShort(ptr addrspace(1) noundef align 2 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !9 !kernel_arg_base_type !9 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i16, align 2
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i16 0, ptr %v, align 2
+ %0 = load i16, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt16]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func zeroext i16 @_Z16sub_group_rotateti(i16 noundef zeroext %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i16, ptr addrspace(1) %1, i32 0
+ store i16 %call, ptr addrspace(1) %arrayidx, align 2
+ %2 = load i16, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt16]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func zeroext i16 @_Z26sub_group_clustered_rotatetij(i16 noundef zeroext %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i16, ptr addrspace(1) %3, i32 1
+ store i16 %call1, ptr addrspace(1) %arrayidx2, align 2
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func zeroext i16 @_Z16sub_group_rotateti(i16 noundef zeroext, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func zeroext i16 @_Z26sub_group_clustered_rotatetij(i16 noundef zeroext, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateInt(ptr addrspace(1) noundef align 4 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !10 !kernel_arg_base_type !10 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i32, align 4
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i32 0, ptr %v, align 4
+ %0 = load i32, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt32]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func i32 @_Z16sub_group_rotateii(i32 noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i32, ptr addrspace(1) %1, i32 0
+ store i32 %call, ptr addrspace(1) %arrayidx, align 4
+ %2 = load i32, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt32]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func i32 @_Z26sub_group_clustered_rotateiij(i32 noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %3, i32 1
+ store i32 %call1, ptr addrspace(1) %arrayidx2, align 4
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func i32 @_Z16sub_group_rotateii(i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func i32 @_Z26sub_group_clustered_rotateiij(i32 noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateUInt(ptr addrspace(1) noundef align 4 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !11 !kernel_arg_base_type !11 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i32, align 4
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i32 0, ptr %v, align 4
+ %0 = load i32, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt32]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func i32 @_Z16sub_group_rotateji(i32 noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i32, ptr addrspace(1) %1, i32 0
+ store i32 %call, ptr addrspace(1) %arrayidx, align 4
+ %2 = load i32, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt32]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func i32 @_Z26sub_group_clustered_rotatejij(i32 noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %3, i32 1
+ store i32 %call1, ptr addrspace(1) %arrayidx2, align 4
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func i32 @_Z16sub_group_rotateji(i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func i32 @_Z26sub_group_clustered_rotatejij(i32 noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateLong(ptr addrspace(1) noundef align 8 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !12 !kernel_arg_base_type !12 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i64, align 8
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i64 0, ptr %v, align 8
+ %0 = load i64, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt64]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func i64 @_Z16sub_group_rotateli(i64 noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i64, ptr addrspace(1) %1, i32 0
+ store i64 %call, ptr addrspace(1) %arrayidx, align 8
+ %2 = load i64, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt64]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func i64 @_Z26sub_group_clustered_rotatelij(i64 noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i64, ptr addrspace(1) %3, i32 1
+ store i64 %call1, ptr addrspace(1) %arrayidx2, align 8
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func i64 @_Z16sub_group_rotateli(i64 noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func i64 @_Z26sub_group_clustered_rotatelij(i64 noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateULong(ptr addrspace(1) noundef align 8 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !13 !kernel_arg_base_type !13 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca i64, align 8
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store i64 0, ptr %v, align 8
+ %0 = load i64, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt64]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func i64 @_Z16sub_group_rotatemi(i64 noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds i64, ptr addrspace(1) %1, i32 0
+ store i64 %call, ptr addrspace(1) %arrayidx, align 8
+ %2 = load i64, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyInt64]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func i64 @_Z26sub_group_clustered_rotatemij(i64 noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds i64, ptr addrspace(1) %3, i32 1
+ store i64 %call1, ptr addrspace(1) %arrayidx2, align 8
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func i64 @_Z16sub_group_rotatemi(i64 noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func i64 @_Z26sub_group_clustered_rotatemij(i64 noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateFloat(ptr addrspace(1) noundef align 4 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !14 !kernel_arg_base_type !14 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca float, align 4
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store float 0.000000e+00, ptr %v, align 4
+ %0 = load float, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyFloat]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func float @_Z16sub_group_rotatefi(float noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds float, ptr addrspace(1) %1, i32 0
+ store float %call, ptr addrspace(1) %arrayidx, align 4
+ %2 = load float, ptr %v, align 4
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyFloat]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func float @_Z26sub_group_clustered_rotatefij(float noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds float, ptr addrspace(1) %3, i32 1
+ store float %call1, ptr addrspace(1) %arrayidx2, align 4
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func float @_Z16sub_group_rotatefi(float noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func float @_Z26sub_group_clustered_rotatefij(float noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateHalf(ptr addrspace(1) noundef align 2 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !15 !kernel_arg_base_type !15 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca half, align 2
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store half 0xH0000, ptr %v, align 2
+ %0 = load half, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyHalf]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func half @_Z16sub_group_rotateDhi(half noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds half, ptr addrspace(1) %1, i32 0
+ store half %call, ptr addrspace(1) %arrayidx, align 2
+ %2 = load half, ptr %v, align 2
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyHalf]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func half @_Z26sub_group_clustered_rotateDhij(half noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds half, ptr addrspace(1) %3, i32 1
+ store half %call1, ptr addrspace(1) %arrayidx2, align 2
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func half @_Z16sub_group_rotateDhi(half noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func half @_Z26sub_group_clustered_rotateDhij(half noundef, i32 noundef, i32 noundef) #1
+
+; Function Attrs: convergent noinline norecurse nounwind optnone
+define dso_local spir_kernel void @testRotateDouble(ptr addrspace(1) noundef align 8 %dst) #0 !kernel_arg_addr_space !3 !kernel_arg_access_qual !4 !kernel_arg_type !16 !kernel_arg_base_type !16 !kernel_arg_type_qual !6 {
+entry:
+ %dst.addr = alloca ptr addrspace(1), align 4
+ %v = alloca double, align 8
+ store ptr addrspace(1) %dst, ptr %dst.addr, align 4
+ store double 0.000000e+00, ptr %v, align 8
+ %0 = load double, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyDouble]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]]
+ %call = call spir_func double @_Z16sub_group_rotatedi(double noundef %0, i32 noundef 2) #2
+ %1 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx = getelementptr inbounds double, ptr addrspace(1) %1, i32 0
+ store double %call, ptr addrspace(1) %arrayidx, align 8
+ %2 = load double, ptr %v, align 8
+ ; CHECK: OpGroupNonUniformRotateKHR %[[TyDouble]] %[[ScopeSubgroup]] %[[#]] %[[ConstInt2]] %[[ConstInt4]]
+ %call1 = call spir_func double @_Z26sub_group_clustered_rotatedij(double noundef %2, i32 noundef 2, i32 noundef 4) #2
+ %3 = load ptr addrspace(1), ptr %dst.addr, align 4
+ %arrayidx2 = getelementptr inbounds double, ptr addrspace(1) %3, i32 1
+ store double %call1, ptr addrspace(1) %arrayidx2, align 8
+ ret void
+}
+
+; Function Attrs: convergent nounwind
+declare spir_func double @_Z16sub_group_rotatedi(double noundef, i32 noundef) #1
+
+; Function Attrs: convergent nounwind
+declare spir_func double @_Z26sub_group_clustered_rotatedij(double noundef, i32 noundef, i32 noundef) #1
+
+attributes #0 = { convergent noinline norecurse nounwind optnone "no-trapping-math"="true" "stack-protector-buffer-size"="8" "uniform-work-group-size"="false" }
+attributes #1 = { convergent nounwind "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
+attributes #2 = { convergent nounwind }
+
+!llvm.module.flags = !{!0}
+!opencl.ocl.version = !{!1}
+!opencl.spir.version = !{!1}
+!llvm.ident = !{!2}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 2, i32 0}
+!2 = !{!"clang version 19.0.0"}
+!3 = !{i32 1}
+!4 = !{!"none"}
+!5 = !{!"char*"}
+!6 = !{!""}
+!7 = !{!"uchar*"}
+!8 = !{!"short*"}
+!9 = !{!"ushort*"}
+!10 = !{!"int*"}
+!11 = !{!"uint*"}
+!12 = !{!"long*"}
+!13 = !{!"ulong*"}
+!14 = !{!"float*"}
+!15 = !{!"half*"}
+!16 = !{!"double*"}
More information about the llvm-commits
mailing list