[Mlir-commits] [mlir] [mlir][OpenMP] Fix update of linear iteration variables (PR #183800)
Leandro Lupori
llvmlistbot at llvm.org
Fri Jul 24 07:59:45 PDT 2026
https://github.com/luporl updated https://github.com/llvm/llvm-project/pull/183800
>From 49d7a2054c0b123a6783a49383c71066b7e1d1d9 Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Fri, 27 Feb 2026 15:32:58 -0300
Subject: [PATCH 1/6] [mlir][OpenMP] Fix update of linear iteration variables
The final value of a linear iteration variable must be the loop
limit_value + step. Before this patch it was limit_value.
This fixes the second issue reported in #170784.
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 129 ++++++++++++------
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 15 +-
2 files changed, 103 insertions(+), 41 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index deefa29c157f3..993cd89f74300 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -150,6 +150,37 @@ class LinearClauseProcessor {
llvm::BasicBlock *linearFinalizationBB;
llvm::BasicBlock *linearExitBB;
llvm::BasicBlock *linearLastIterExitBB;
+ Value linearLoopIV;
+ Value linearLoopIVStart;
+
+ void updateLinearVar(llvm::IRBuilderBase &builder, llvm::Type *varType,
+ llvm::Value *var, llvm::Value *varStart,
+ llvm::Value *step, llvm::Value *iv) {
+ if (!iv->getType()->isIntegerTy())
+ llvm_unreachable("OpenMP loop induction variable must be an integer "
+ "type");
+
+ if (varType->isIntegerTy()) {
+ // Integer path: normalize all arithmetic to linearVarType
+ iv = builder.CreateSExtOrTrunc(iv, varType);
+ step = builder.CreateSExtOrTrunc(step, varType);
+
+ llvm::Value *mulInst = builder.CreateMul(iv, step);
+ llvm::Value *addInst = builder.CreateAdd(varStart, mulInst);
+ builder.CreateStore(addInst, var);
+ } else if (varType->isFloatingPointTy()) {
+ // Float path: perform multiply in integer, then convert to float
+ step = builder.CreateSExtOrTrunc(step, iv->getType());
+
+ llvm::Value *mulInst = builder.CreateMul(iv, step);
+ llvm::Value *mulFp = builder.CreateSIToFP(mulInst, varType);
+ llvm::Value *addInst = builder.CreateFAdd(varStart, mulFp);
+ builder.CreateStore(addInst, var);
+ } else {
+ llvm_unreachable(
+ "Linear variable must be of integer or floating-point type");
+ }
+ }
public:
// Register type for the linear variables
@@ -189,46 +220,63 @@ class LinearClauseProcessor {
}
}
+ // Find linear iteration variable and save it for later updates
+ void initLinearIV(omp::SimdOp simdOp) {
+ auto loopOp = cast<omp::LoopNestOp>(simdOp.getWrappedLoop());
+ // NOTE iteration variables can only be linear in non-nested loops.
+ if (loopOp.getIVs().size() != 1)
+ return;
+ // The linear IV is the loop IV's store address.
+ BlockArgument arg = loopOp.getIVs().front();
+ for (const Operation *user : arg.getUsers()) {
+ if (auto storeOp = dyn_cast<LLVM::StoreOp>(user)) {
+ for (Value linearVar : simdOp.getLinearVars()) {
+ if (linearVar == storeOp.getAddr()) {
+ linearLoopIV = linearVar;
+ linearLoopIVStart = loopOp.getLoopLowerBounds().front();
+ break;
+ }
+ }
+ }
+ }
+ }
+
// Emit IR for updating Linear variables
- void updateLinearVar(llvm::IRBuilderBase &builder, llvm::BasicBlock *loopBody,
- llvm::Value *loopInductionVar) {
+ void updateLinearVars(llvm::IRBuilderBase &builder,
+ llvm::BasicBlock *loopBody,
+ llvm::Value *loopInductionVar) {
builder.SetInsertPoint(loopBody->getTerminator());
for (size_t index = 0; index < linearPreconditionVars.size(); index++) {
- llvm::Type *linearVarType = linearVarTypes[index];
- llvm::Value *iv = loopInductionVar;
- llvm::Value *step = linearSteps[index];
-
- if (!iv->getType()->isIntegerTy())
- llvm_unreachable("OpenMP loop induction variable must be an integer "
- "type");
-
- if (linearVarType->isIntegerTy()) {
- // Integer path: normalize all arithmetic to linearVarType
- iv = builder.CreateSExtOrTrunc(iv, linearVarType);
- step = builder.CreateSExtOrTrunc(step, linearVarType);
-
- llvm::LoadInst *linearVarStart =
- builder.CreateLoad(linearVarType, linearPreconditionVars[index]);
- llvm::Value *mulInst = builder.CreateMul(iv, step);
- llvm::Value *addInst = builder.CreateAdd(linearVarStart, mulInst);
- builder.CreateStore(addInst, linearLoopBodyTemps[index]);
- } else if (linearVarType->isFloatingPointTy()) {
- // Float path: perform multiply in integer, then convert to float
- step = builder.CreateSExtOrTrunc(step, iv->getType());
- llvm::Value *mulInst = builder.CreateMul(iv, step);
-
- llvm::LoadInst *linearVarStart =
- builder.CreateLoad(linearVarType, linearPreconditionVars[index]);
- llvm::Value *mulFp = builder.CreateSIToFP(mulInst, linearVarType);
- llvm::Value *addInst = builder.CreateFAdd(linearVarStart, mulFp);
- builder.CreateStore(addInst, linearLoopBodyTemps[index]);
- } else {
- llvm_unreachable(
- "Linear variable must be of integer or floating-point type");
- }
+ llvm::LoadInst *linearVarStart = builder.CreateLoad(
+ linearVarTypes[index], linearPreconditionVars[index]);
+ updateLinearVar(builder, linearVarTypes[index],
+ linearLoopBodyTemps[index], linearVarStart,
+ linearSteps[index], loopInductionVar);
}
}
+ // Emit IR for updating linear iteration variables on loop exit
+ void updateLinearIV(llvm::IRBuilderBase &builder,
+ LLVM::ModuleTranslation &moduleTranslation,
+ llvm::Value *loopIV) {
+ if (!linearLoopIV)
+ return;
+ llvm::Value *linearIV = moduleTranslation.lookupValue(linearLoopIV);
+ llvm::Value *linearIVStart =
+ moduleTranslation.lookupValue(linearLoopIVStart);
+
+ // Find linearIV's index
+ size_t index;
+ for (index = 0; index < linearOrigVal.size(); index++)
+ if (linearIV == linearOrigVal[index])
+ break;
+ if (index == linearOrigVal.size())
+ return;
+
+ updateLinearVar(builder, linearVarTypes[index], linearLoopBodyTemps[index],
+ linearIVStart, linearSteps[index], loopIV);
+ }
+
// Linear variable finalization is conditional on the last logical iteration.
// Create BB splits to manage the same.
void splitLinearFiniBB(llvm::IRBuilderBase &builder,
@@ -3852,8 +3900,8 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
if (failed(handleError(afterBarrierIP, *loopOp)))
return failure();
builder.restoreIP(*afterBarrierIP);
- linearClauseProcessor.updateLinearVar(builder, loopInfo->getBody(),
- loopInfo->getIndVar());
+ linearClauseProcessor.updateLinearVars(builder, loopInfo->getBody(),
+ loopInfo->getIndVar());
linearClauseProcessor.splitLinearFiniBB(builder, loopInfo->getExit());
}
@@ -4153,6 +4201,8 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
// Initialize linear variables and linear step
LinearClauseProcessor linearClauseProcessor;
+ linearClauseProcessor.initLinearIV(simdOp);
+
if (!simdOp.getLinearVars().empty()) {
auto linearVarTypes = simdOp.getLinearVarTypes().value();
for (mlir::Attribute linearVarType : linearVarTypes)
@@ -4250,8 +4300,8 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
linearClauseProcessor.initLinearVar(builder, moduleTranslation,
loopInfo->getPreheader());
- linearClauseProcessor.updateLinearVar(builder, loopInfo->getBody(),
- loopInfo->getIndVar());
+ linearClauseProcessor.updateLinearVars(builder, loopInfo->getBody(),
+ loopInfo->getIndVar());
}
builder.SetInsertPoint(*regionBlock, (*regionBlock)->begin());
@@ -4261,6 +4311,9 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
: nullptr,
order, simdlen, safelen);
+ linearClauseProcessor.updateLinearIV(builder, moduleTranslation,
+ loopInfo->getIndVar());
+
linearClauseProcessor.emitStoresForLinearVar(builder);
// Check if this SIMD loop contains ordered regions
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index 38f10320bd3ae..45eb6ac1ef957 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -743,6 +743,7 @@ llvm.func @simd_simple(%lb : i64, %ub : i64, %step : i64, %arg0: !llvm.ptr) {
llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// CHECK-LABEL: @simd_linear
+// CHECK-SAME: (i32 %[[LB:.*]], i32 %{{.*}}, i32 %[[STEP:.*]], ptr %[[X:.*]])
// CHECK: %[[LINEAR_VAR:.*]] = alloca i32, align 4
// CHECK: %[[LINEAR_RESULT:.*]] = alloca i32, align 4
@@ -758,8 +759,16 @@ llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// CHECK: %[[MUL:.*]] = mul i32 %omp_loop.iv, {{.*}}
// CHECK: %[[ADD:.*]] = add i32 %[[LOAD]], %[[MUL]]
// CHECK: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]], align 4, !llvm.access.group !1
+
+// CHECK: omp.region.cont:
+// CHECK: %[[MUL:.*]] = mul i32 %omp_loop.iv, %[[STEP]]
+// CHECK: %[[ADD:.*]] = add i32 %[[LB]], %[[MUL]]
+// CHECK: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]], align 4
+// CHECK: %[[LOAD:.*]] = load i32, ptr %[[LINEAR_RESULT]], align 4
+// CHECK: store i32 %[[LOAD]], ptr %[[X]], align 4
omp.simd linear(%x : !llvm.ptr = %step : i32) {
omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
+ llvm.store %iv, %x : i32, !llvm.ptr
omp.yield
}
} {linear_var_types = [i32]}
@@ -784,8 +793,8 @@ llvm.func @simd_linear_i64_var_i32_step(%lb : i32, %ub : i32, %x : !llvm.ptr) {
// CHECK: omp_loop.body:
// Verify type conversions: iv (i32) is extended to i64 before multiplication
-// CHECK: %[[IV_I64:.*]] = sext i32 %omp_loop.iv to i64
// CHECK: %[[LOAD:.*]] = load i64, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
+// CHECK: %[[IV_I64:.*]] = sext i32 %omp_loop.iv to i64
// Verify multiplication and addition use consistent i64 types
// CHECK: %[[MUL:.*]] = mul i64 %[[IV_I64]], {{.*}}
// CHECK: %[[ADD:.*]] = add i64 %[[LOAD]], %[[MUL]]
@@ -817,8 +826,8 @@ llvm.func @simd_linear_f64_var_i32_step(%lb : i32, %ub : i32, %x : !llvm.ptr) {
// CHECK: omp_loop.body:
// Verify integer multiplication, load, and conversion to float
// CHECK: mul i32 %omp_loop.iv
-// CHECK: %[[MUL_INT:.*]] = mul i32 %omp_loop.iv, {{.*}}
-// CHECK-NEXT: %[[LOAD:.*]] = load double, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
+// CHECK: %[[LOAD:.*]] = load double, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
+// CHECK-NEXT: %[[MUL_INT:.*]] = mul i32 %omp_loop.iv, {{.*}}
// CHECK-NEXT: %[[MUL_FP:.*]] = sitofp i32 %[[MUL_INT]] to double
// CHECK-NEXT: %[[ADD:.*]] = fadd double %[[LOAD]], %[[MUL_FP]]
// CHECK-NEXT: store double %[[ADD]], ptr %[[LINEAR_RESULT]], {{.*}}!llvm.access.group
>From 20ac4aef52312cd93944ad17f0fef0b126211102 Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Tue, 19 May 2026 12:06:23 -0300
Subject: [PATCH 2/6] Fix "instruction does not dominate all uses" issue
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 104 +++++++++---------
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 48 ++++++--
2 files changed, 90 insertions(+), 62 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 993cd89f74300..aace5283e4362 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -151,36 +151,6 @@ class LinearClauseProcessor {
llvm::BasicBlock *linearExitBB;
llvm::BasicBlock *linearLastIterExitBB;
Value linearLoopIV;
- Value linearLoopIVStart;
-
- void updateLinearVar(llvm::IRBuilderBase &builder, llvm::Type *varType,
- llvm::Value *var, llvm::Value *varStart,
- llvm::Value *step, llvm::Value *iv) {
- if (!iv->getType()->isIntegerTy())
- llvm_unreachable("OpenMP loop induction variable must be an integer "
- "type");
-
- if (varType->isIntegerTy()) {
- // Integer path: normalize all arithmetic to linearVarType
- iv = builder.CreateSExtOrTrunc(iv, varType);
- step = builder.CreateSExtOrTrunc(step, varType);
-
- llvm::Value *mulInst = builder.CreateMul(iv, step);
- llvm::Value *addInst = builder.CreateAdd(varStart, mulInst);
- builder.CreateStore(addInst, var);
- } else if (varType->isFloatingPointTy()) {
- // Float path: perform multiply in integer, then convert to float
- step = builder.CreateSExtOrTrunc(step, iv->getType());
-
- llvm::Value *mulInst = builder.CreateMul(iv, step);
- llvm::Value *mulFp = builder.CreateSIToFP(mulInst, varType);
- llvm::Value *addInst = builder.CreateFAdd(varStart, mulFp);
- builder.CreateStore(addInst, var);
- } else {
- llvm_unreachable(
- "Linear variable must be of integer or floating-point type");
- }
- }
public:
// Register type for the linear variables
@@ -233,7 +203,6 @@ class LinearClauseProcessor {
for (Value linearVar : simdOp.getLinearVars()) {
if (linearVar == storeOp.getAddr()) {
linearLoopIV = linearVar;
- linearLoopIVStart = loopOp.getLoopLowerBounds().front();
break;
}
}
@@ -242,28 +211,51 @@ class LinearClauseProcessor {
}
// Emit IR for updating Linear variables
- void updateLinearVars(llvm::IRBuilderBase &builder,
- llvm::BasicBlock *loopBody,
- llvm::Value *loopInductionVar) {
+ void updateLinearVar(llvm::IRBuilderBase &builder, llvm::BasicBlock *loopBody,
+ llvm::Value *loopInductionVar) {
builder.SetInsertPoint(loopBody->getTerminator());
for (size_t index = 0; index < linearPreconditionVars.size(); index++) {
- llvm::LoadInst *linearVarStart = builder.CreateLoad(
- linearVarTypes[index], linearPreconditionVars[index]);
- updateLinearVar(builder, linearVarTypes[index],
- linearLoopBodyTemps[index], linearVarStart,
- linearSteps[index], loopInductionVar);
+ llvm::Type *linearVarType = linearVarTypes[index];
+ llvm::Value *iv = loopInductionVar;
+ llvm::Value *step = linearSteps[index];
+
+ if (!iv->getType()->isIntegerTy())
+ llvm_unreachable("OpenMP loop induction variable must be an integer "
+ "type");
+
+ if (linearVarType->isIntegerTy()) {
+ // Integer path: normalize all arithmetic to linearVarType
+ iv = builder.CreateSExtOrTrunc(iv, linearVarType);
+ step = builder.CreateSExtOrTrunc(step, linearVarType);
+
+ llvm::LoadInst *linearVarStart =
+ builder.CreateLoad(linearVarType, linearPreconditionVars[index]);
+ llvm::Value *mulInst = builder.CreateMul(iv, step);
+ llvm::Value *addInst = builder.CreateAdd(linearVarStart, mulInst);
+ builder.CreateStore(addInst, linearLoopBodyTemps[index]);
+ } else if (linearVarType->isFloatingPointTy()) {
+ // Float path: perform multiply in integer, then convert to float
+ step = builder.CreateSExtOrTrunc(step, iv->getType());
+ llvm::Value *mulInst = builder.CreateMul(iv, step);
+
+ llvm::LoadInst *linearVarStart =
+ builder.CreateLoad(linearVarType, linearPreconditionVars[index]);
+ llvm::Value *mulFp = builder.CreateSIToFP(mulInst, linearVarType);
+ llvm::Value *addInst = builder.CreateFAdd(linearVarStart, mulFp);
+ builder.CreateStore(addInst, linearLoopBodyTemps[index]);
+ } else {
+ llvm_unreachable(
+ "Linear variable must be of integer or floating-point type");
+ }
}
}
// Emit IR for updating linear iteration variables on loop exit
void updateLinearIV(llvm::IRBuilderBase &builder,
- LLVM::ModuleTranslation &moduleTranslation,
- llvm::Value *loopIV) {
+ LLVM::ModuleTranslation &moduleTranslation) {
if (!linearLoopIV)
return;
llvm::Value *linearIV = moduleTranslation.lookupValue(linearLoopIV);
- llvm::Value *linearIVStart =
- moduleTranslation.lookupValue(linearLoopIVStart);
// Find linearIV's index
size_t index;
@@ -273,8 +265,17 @@ class LinearClauseProcessor {
if (index == linearOrigVal.size())
return;
- updateLinearVar(builder, linearVarTypes[index], linearLoopBodyTemps[index],
- linearIVStart, linearSteps[index], loopIV);
+ // Add one more step to the linear iteration variable
+ llvm::Type *varType = linearVarTypes[index];
+ llvm::Value *var = linearLoopBodyTemps[index];
+ llvm::Value *step = linearSteps[index];
+ if (!varType->isIntegerTy())
+ llvm_unreachable("Linear iteration variable must be of integer type");
+
+ step = builder.CreateSExtOrTrunc(step, varType);
+ llvm::Value *val = builder.CreateLoad(varType, var);
+ llvm::Value *addInst = builder.CreateAdd(val, step);
+ builder.CreateStore(addInst, var);
}
// Linear variable finalization is conditional on the last logical iteration.
@@ -3900,8 +3901,8 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
if (failed(handleError(afterBarrierIP, *loopOp)))
return failure();
builder.restoreIP(*afterBarrierIP);
- linearClauseProcessor.updateLinearVars(builder, loopInfo->getBody(),
- loopInfo->getIndVar());
+ linearClauseProcessor.updateLinearVar(builder, loopInfo->getBody(),
+ loopInfo->getIndVar());
linearClauseProcessor.splitLinearFiniBB(builder, loopInfo->getExit());
}
@@ -4200,7 +4201,6 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
// Initialize linear variables and linear step
LinearClauseProcessor linearClauseProcessor;
-
linearClauseProcessor.initLinearIV(simdOp);
if (!simdOp.getLinearVars().empty()) {
@@ -4300,8 +4300,8 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
linearClauseProcessor.initLinearVar(builder, moduleTranslation,
loopInfo->getPreheader());
- linearClauseProcessor.updateLinearVars(builder, loopInfo->getBody(),
- loopInfo->getIndVar());
+ linearClauseProcessor.updateLinearVar(builder, loopInfo->getBody(),
+ loopInfo->getIndVar());
}
builder.SetInsertPoint(*regionBlock, (*regionBlock)->begin());
@@ -4311,9 +4311,7 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
: nullptr,
order, simdlen, safelen);
- linearClauseProcessor.updateLinearIV(builder, moduleTranslation,
- loopInfo->getIndVar());
-
+ linearClauseProcessor.updateLinearIV(builder, moduleTranslation);
linearClauseProcessor.emitStoresForLinearVar(builder);
// Check if this SIMD loop contains ordered regions
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index 45eb6ac1ef957..51a12e23a0e5e 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -743,7 +743,7 @@ llvm.func @simd_simple(%lb : i64, %ub : i64, %step : i64, %arg0: !llvm.ptr) {
llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// CHECK-LABEL: @simd_linear
-// CHECK-SAME: (i32 %[[LB:.*]], i32 %{{.*}}, i32 %[[STEP:.*]], ptr %[[X:.*]])
+// CHECK-SAME: (i32 %{{.*}}, i32 %{{.*}}, i32 %[[STEP:.*]], ptr %[[X:.*]])
// CHECK: %[[LINEAR_VAR:.*]] = alloca i32, align 4
// CHECK: %[[LINEAR_RESULT:.*]] = alloca i32, align 4
@@ -761,11 +761,11 @@ llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// CHECK: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]], align 4, !llvm.access.group !1
// CHECK: omp.region.cont:
-// CHECK: %[[MUL:.*]] = mul i32 %omp_loop.iv, %[[STEP]]
-// CHECK: %[[ADD:.*]] = add i32 %[[LB]], %[[MUL]]
-// CHECK: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]], align 4
-// CHECK: %[[LOAD:.*]] = load i32, ptr %[[LINEAR_RESULT]], align 4
-// CHECK: store i32 %[[LOAD]], ptr %[[X]], align 4
+// CHECK: %[[VAL:.*]] = load i32, ptr %[[LINEAR_RESULT]]
+// CHECK-NEXT: %[[ADD:.*]] = add i32 %[[VAL]], %[[STEP]]
+// CHECK-NEXT: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]]
+// CHECK-NEXT: %[[LOAD:.*]] = load i32, ptr %[[LINEAR_RESULT]]
+// CHECK-NEXT: store i32 %[[LOAD]], ptr %[[X]], align 4
omp.simd linear(%x : !llvm.ptr = %step : i32) {
omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
llvm.store %iv, %x : i32, !llvm.ptr
@@ -793,8 +793,8 @@ llvm.func @simd_linear_i64_var_i32_step(%lb : i32, %ub : i32, %x : !llvm.ptr) {
// CHECK: omp_loop.body:
// Verify type conversions: iv (i32) is extended to i64 before multiplication
-// CHECK: %[[LOAD:.*]] = load i64, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
// CHECK: %[[IV_I64:.*]] = sext i32 %omp_loop.iv to i64
+// CHECK: %[[LOAD:.*]] = load i64, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
// Verify multiplication and addition use consistent i64 types
// CHECK: %[[MUL:.*]] = mul i64 %[[IV_I64]], {{.*}}
// CHECK: %[[ADD:.*]] = add i64 %[[LOAD]], %[[MUL]]
@@ -826,8 +826,8 @@ llvm.func @simd_linear_f64_var_i32_step(%lb : i32, %ub : i32, %x : !llvm.ptr) {
// CHECK: omp_loop.body:
// Verify integer multiplication, load, and conversion to float
// CHECK: mul i32 %omp_loop.iv
-// CHECK: %[[LOAD:.*]] = load double, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
-// CHECK-NEXT: %[[MUL_INT:.*]] = mul i32 %omp_loop.iv, {{.*}}
+// CHECK: %[[MUL_INT:.*]] = mul i32 %omp_loop.iv, {{.*}}
+// CHECK-NEXT: %[[LOAD:.*]] = load double, ptr %[[LINEAR_VAR]], {{.*}}!llvm.access.group
// CHECK-NEXT: %[[MUL_FP:.*]] = sitofp i32 %[[MUL_INT]] to double
// CHECK-NEXT: %[[ADD:.*]] = fadd double %[[LOAD]], %[[MUL_FP]]
// CHECK-NEXT: store double %[[ADD]], ptr %[[LINEAR_RESULT]], {{.*}}!llvm.access.group
@@ -841,6 +841,36 @@ llvm.func @simd_linear_f64_var_i32_step(%lb : i32, %ub : i32, %x : !llvm.ptr) {
// -----
+// Test the update of omp.simd linear iteration variables, when nested inside
+// omp.wsloop.
+llvm.func @wsloop_simd_linear(%x : !llvm.ptr) {
+
+// CHECK-LABEL: @wsloop_simd_linear
+
+// CHECK: omp.wsloop.region:
+// CHECK: %[[LINEAR_VAR:.*]] = alloca i32
+// CHECK: %[[LINEAR_RESULT:.*]] = alloca i32
+
+// CHECK: omp.region.cont2:
+// CHECK: %[[VAL:.*]] = load i32, ptr %[[LINEAR_RESULT]]
+// CHECK-NEXT: %[[ADD:.*]] = add i32 %[[VAL]], 25
+// CHECK-NEXT: store i32 %[[ADD]], ptr %[[LINEAR_RESULT]]
+ %lb = llvm.mlir.constant(1 : i32) : i32
+ %ub = llvm.mlir.constant(100 : i32) : i32
+ %step = llvm.mlir.constant(25 : i32) : i32
+ omp.wsloop {
+ omp.simd linear(%x : !llvm.ptr = %step : i32) {
+ omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
+ llvm.store %iv, %x : i32, !llvm.ptr
+ omp.yield
+ }
+ } {linear_var_types = [i32], omp.composite}
+ } {omp.composite}
+ llvm.return
+}
+
+// -----
+
// CHECK-LABEL: @simd_simple_multiple
llvm.func @simd_simple_multiple(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64, %ub2 : i64, %step2 : i64, %arg0: !llvm.ptr, %arg1: !llvm.ptr) {
omp.simd {
>From 030a957a941fb50d15c99f9d0031e26019b53992 Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Tue, 16 Jun 2026 09:58:04 -0300
Subject: [PATCH 3/6] Verify that linear variables don't appear in other DSA
clauses
---
mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp | 10 ++++++++++
mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir | 14 ++++++++++++++
mlir/test/Target/LLVMIR/openmp-simd-ordered.mlir | 12 +++++-------
3 files changed, 29 insertions(+), 7 deletions(-)
diff --git a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
index db5fd8f2e3230..b328e15313538 100644
--- a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
+++ b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
@@ -3363,6 +3363,16 @@ LogicalResult SimdOp::verify() {
if (getLinearVars().size() &&
getLinearVarTypes().value().size() != getLinearVars().size())
return emitError() << "Ill-formed type attributes for linear variables";
+
+ llvm::DenseSet<Value> privateVars(llvm::from_range, getPrivateVars());
+ llvm::DenseSet<Value> reductionVars(llvm::from_range, getReductionVars());
+ // TODO Check lastprivate vars when their support is added to SimdOp.
+ for (Value var : getLinearVars()) {
+ if (privateVars.contains(var) || reductionVars.contains(var))
+ return emitOpError()
+ << "linear variables cannot appear in other data-sharing clauses";
+ }
+
return success();
}
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir b/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
index d9459ace54422..7e2e0061f80be 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
@@ -116,6 +116,20 @@ llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// -----
+omp.private {type = private} @i_private_i32 : i32
+llvm.func @simd_linear_private(%lb : i32, %ub : i32, %step : i32, %i : !llvm.ptr) {
+ // expected-error @below {{linear variables cannot appear in other data-sharing clauses}}
+ omp.simd linear(%i : !llvm.ptr = %step : i32)
+ private(@i_private_i32 %i -> %priv_i : !llvm.ptr) {
+ omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
+ omp.yield
+ }
+ } {linear_var_types = [i32]}
+ llvm.return
+}
+
+// -----
+
module attributes {llvm.target_triple = "amdgcn-amd-amdhsa", omp.is_target_device = true} {
llvm.func @host_op_in_device(%arg0 : !llvm.ptr) {
// expected-error @below {{unsupported host op found in device}}
diff --git a/mlir/test/Target/LLVMIR/openmp-simd-ordered.mlir b/mlir/test/Target/LLVMIR/openmp-simd-ordered.mlir
index 8da1fab1f48af..a84c51f08daa7 100644
--- a/mlir/test/Target/LLVMIR/openmp-simd-ordered.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-simd-ordered.mlir
@@ -17,8 +17,6 @@
// !$omp end simd
module {
- omp.private {type = private} @i_private_i32 : i32
-
// CHECK-LABEL: define void @simd_ordered_linear
llvm.func @simd_ordered_linear() {
%c0_i64 = llvm.mlir.constant(0 : i64) : i64
@@ -38,14 +36,14 @@ module {
// CHECK: %.linear_var = alloca i32
// CHECK: %.linear_result = alloca i32
- omp.simd linear(%i : !llvm.ptr = %c1_i32 : i32) private(@i_private_i32 %i -> %arg0 : !llvm.ptr) {
+ omp.simd linear(%i : !llvm.ptr = %c1_i32 : i32) {
omp.loop_nest (%iv) : i32 = (%c1_i32) to (%c10_i32) inclusive step (%c1_i32) {
// CHECK: omp.loop_nest.region:
// CHECK: load i32, ptr %.linear_result
- llvm.store %iv, %arg0 : i32, !llvm.ptr
+ llvm.store %iv, %i : i32, !llvm.ptr
// Compute a[i] = b[i] * 10
- %i_val = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_val = llvm.load %i : !llvm.ptr -> i32
%i_idx = llvm.sext %i_val : i32 to i64
%i_off = llvm.sub %i_idx, %c1_i64 : i64
%b_ptr = llvm.getelementptr %b[%i_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
@@ -58,7 +56,7 @@ module {
omp.ordered.region par_level_simd {
// CHECK: omp.ordered.region:
// CHECK: load i32, ptr %.linear_result
- %i_ord = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_ord = llvm.load %i : !llvm.ptr -> i32
%i_ord_idx = llvm.sext %i_ord : i32 to i64
%i_ord_off = llvm.sub %i_ord_idx, %c1_i64 : i64
%a_ord_ptr = llvm.getelementptr %a[%i_ord_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
@@ -69,7 +67,7 @@ module {
// Compute c[i] = a[i] * 2 (code after ordered region)
// CHECK: omp_region.finalize:
// CHECK: load i32, ptr %.linear_result
- %i_post = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_post = llvm.load %i : !llvm.ptr -> i32
%i_post_idx = llvm.sext %i_post : i32 to i64
%i_post_off = llvm.sub %i_post_idx, %c1_i64 : i64
%a_post_ptr = llvm.getelementptr %a[%i_post_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
>From 2f8b6c3b6f7ca958505c6e6d7cc41ec53cb0fe0d Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Tue, 16 Jun 2026 11:10:29 -0300
Subject: [PATCH 4/6] Update mlir tests
---
mlir/test/Target/LLVMIR/openmp-simd-guided.mlir | 3 +--
.../Target/LLVMIR/openmp-wsloop-simd-ordered.mlir | 12 +++++-------
2 files changed, 6 insertions(+), 9 deletions(-)
diff --git a/mlir/test/Target/LLVMIR/openmp-simd-guided.mlir b/mlir/test/Target/LLVMIR/openmp-simd-guided.mlir
index b1f18b9f85417..496669c5b90d5 100644
--- a/mlir/test/Target/LLVMIR/openmp-simd-guided.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-simd-guided.mlir
@@ -1,7 +1,6 @@
// Ensure that schedule can be used with the guided kind-type and simd construct
// RUN: mlir-translate -mlir-to-llvmir -split-input-file %s | FileCheck %s
-omp.private {type = private} @_QFEi_private_i32 : i32
llvm.func @test_simd_guided() {
%0 = llvm.mlir.constant (1 : i64) : i64
%c1_i32 = llvm.mlir.constant (1 : i32) : i32
@@ -10,7 +9,7 @@ llvm.func @test_simd_guided() {
%c4_i32 = llvm.mlir.constant (4 : i32) : i32
%1 = llvm.alloca %0 x i32 {bindc_name = "i"} : (i64) -> !llvm.ptr
omp.wsloop schedule(guided = %c4_i32 : i32, simd) {
- omp.simd linear(%1 : !llvm.ptr = %c1_i32 : i32) private(@_QFEi_private_i32 %1 -> %arg0 : !llvm.ptr) {
+ omp.simd linear(%1 : !llvm.ptr = %c1_i32 : i32) {
omp.loop_nest (%arg1) : i32 = (%c0_i32) to (%c64_i32) inclusive step (%c1_i32) {
omp.yield
}
diff --git a/mlir/test/Target/LLVMIR/openmp-wsloop-simd-ordered.mlir b/mlir/test/Target/LLVMIR/openmp-wsloop-simd-ordered.mlir
index beb300c23651f..8d5b89ecbd00a 100644
--- a/mlir/test/Target/LLVMIR/openmp-wsloop-simd-ordered.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-wsloop-simd-ordered.mlir
@@ -17,8 +17,6 @@
// !$omp end do simd
module {
- omp.private {type = private} @i_private_i32 : i32
-
// CHECK-LABEL: define void @wsloop_simd_ordered_linear
llvm.func @wsloop_simd_ordered_linear() {
%c0_i64 = llvm.mlir.constant(0 : i64) : i64
@@ -39,14 +37,14 @@ module {
// CHECK: %.linear_result = alloca i32
omp.wsloop ordered(0) {
- omp.simd linear(%i : !llvm.ptr = %c1_i32 : i32) private(@i_private_i32 %i -> %arg0 : !llvm.ptr) {
+ omp.simd linear(%i : !llvm.ptr = %c1_i32 : i32) {
omp.loop_nest (%iv) : i32 = (%c1_i32) to (%c100_i32) inclusive step (%c1_i32) {
// CHECK: omp.loop_nest.region:
// CHECK: load i32, ptr %.linear_result
- llvm.store %iv, %arg0 : i32, !llvm.ptr
+ llvm.store %iv, %i : i32, !llvm.ptr
// Compute a[i] = b[i] * 10
- %i_val = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_val = llvm.load %i : !llvm.ptr -> i32
%i_idx = llvm.sext %i_val : i32 to i64
%i_off = llvm.sub %i_idx, %c1_i64 : i64
%b_ptr = llvm.getelementptr %b[%i_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
@@ -59,7 +57,7 @@ module {
omp.ordered.region par_level_simd {
// CHECK: omp.ordered.region:
// CHECK: load i32, ptr %.linear_result
- %i_ord = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_ord = llvm.load %i : !llvm.ptr -> i32
%i_ord_idx = llvm.sext %i_ord : i32 to i64
%i_ord_off = llvm.sub %i_ord_idx, %c1_i64 : i64
%a_ord_ptr = llvm.getelementptr %a[%i_ord_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
@@ -71,7 +69,7 @@ module {
// Compute c[i] = a[i] * 2 (code after ordered region)
// CHECK: omp_region.finalize:
// CHECK: load i32, ptr %.linear_result
- %i_post = llvm.load %arg0 : !llvm.ptr -> i32
+ %i_post = llvm.load %i : !llvm.ptr -> i32
%i_post_idx = llvm.sext %i_post : i32 to i64
%i_post_off = llvm.sub %i_post_idx, %c1_i64 : i64
%a_post_ptr = llvm.getelementptr %a[%i_post_off] : (!llvm.ptr, i64) -> !llvm.ptr, i32
>From b671d6cf6efa8413abd061ac101b1ea14137ff5f Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Thu, 23 Jul 2026 16:48:58 -0300
Subject: [PATCH 5/6] Report an error if the linear variable can't be
determined unambiguously
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 20 ++++++++++++++-----
.../Target/LLVMIR/openmp-llvm-invalid.mlir | 15 ++++++++++++++
2 files changed, 30 insertions(+), 5 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 2a246ebf98b3c..b938572fe3720 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -191,23 +191,32 @@ class LinearClauseProcessor {
}
// Find linear iteration variable and save it for later updates
- void initLinearIV(omp::SimdOp simdOp) {
+ LogicalResult initLinearIV(omp::SimdOp simdOp) {
auto loopOp = cast<omp::LoopNestOp>(simdOp.getWrappedLoop());
// NOTE iteration variables can only be linear in non-nested loops.
if (loopOp.getIVs().size() != 1)
- return;
- // The linear IV is the loop IV's store address.
+ return success();
+ // Currently, frontends using `omp.simd` always generate a store from the
+ // `omp.loop_nest`'s IV to the corresponding iteration variable.
+ // We leverage this to find the linear iteration variable.
+ //
+ // TODO Add an attribute to `omp.loop_nest` that explicitly lists the
+ // variables that correspond to the loop induction variables.
BlockArgument arg = loopOp.getIVs().front();
for (const Operation *user : arg.getUsers()) {
if (auto storeOp = dyn_cast<LLVM::StoreOp>(user)) {
for (Value linearVar : simdOp.getLinearVars()) {
if (linearVar == storeOp.getAddr()) {
+ if (linearLoopIV)
+ return simdOp.emitError(
+ "Could not determine the linear variable associated with the "
+ "loop nest induction variable");
linearLoopIV = linearVar;
- break;
}
}
}
}
+ return success();
}
// Emit IR for updating Linear variables
@@ -4858,7 +4867,8 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
// Initialize linear variables and linear step
LinearClauseProcessor linearClauseProcessor;
- linearClauseProcessor.initLinearIV(simdOp);
+ if (linearClauseProcessor.initLinearIV(simdOp).failed())
+ return failure();
if (!simdOp.getLinearVars().empty()) {
auto linearVarTypes = simdOp.getLinearVarTypes().value();
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir b/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
index 6eff5accca138..234f404ff5690 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm-invalid.mlir
@@ -130,6 +130,21 @@ llvm.func @simd_linear_private(%lb : i32, %ub : i32, %step : i32, %i : !llvm.ptr
// -----
+llvm.func @simd_linear_ambiguous_iv(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr, %y : !llvm.ptr) {
+ // expected-error @below {{Could not determine the linear variable associated with the loop nest induction variable}}
+ // expected-error @below {{LLVM Translation failed for operation: omp.simd}}
+ omp.simd linear(%x : !llvm.ptr = %step : i32, %y : !llvm.ptr = %step : i32) {
+ omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
+ llvm.store %iv, %x : i32, !llvm.ptr
+ llvm.store %iv, %y : i32, !llvm.ptr
+ omp.yield
+ }
+ } {linear_var_types = [i32, i32]}
+ llvm.return
+}
+
+// -----
+
module attributes {llvm.target_triple = "amdgcn-amd-amdhsa", omp.is_target_device = true} {
llvm.func @host_op_in_device(%arg0 : !llvm.ptr) {
// expected-error @below {{unsupported host op found in device}}
>From 7d19047506073f01cb6135659cb64561812a3a88 Mon Sep 17 00:00:00 2001
From: Leandro Lupori <leandro.lupori at linaro.org>
Date: Fri, 24 Jul 2026 11:57:10 -0300
Subject: [PATCH 6/6] Allow repeated stores to the same location
---
.../Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp | 2 +-
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 1 +
2 files changed, 2 insertions(+), 1 deletion(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index b938572fe3720..ca320ea0b1453 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -207,7 +207,7 @@ class LinearClauseProcessor {
if (auto storeOp = dyn_cast<LLVM::StoreOp>(user)) {
for (Value linearVar : simdOp.getLinearVars()) {
if (linearVar == storeOp.getAddr()) {
- if (linearLoopIV)
+ if (linearLoopIV && linearLoopIV != linearVar)
return simdOp.emitError(
"Could not determine the linear variable associated with the "
"loop nest induction variable");
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index 01fe0ae690955..de51468efdec1 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -768,6 +768,7 @@ llvm.func @simd_linear(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// CHECK-NEXT: store i32 %[[LOAD]], ptr %[[X]], align 4
omp.simd linear(%x : !llvm.ptr = %step : i32) {
omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
+ llvm.store %iv, %x : i32, !llvm.ptr
llvm.store %iv, %x : i32, !llvm.ptr
omp.yield
}
More information about the Mlir-commits
mailing list