[flang-commits] [flang] [flang][OpenMP] Fix data race with LINEAR clause on composite DO SIMD (PR #195634)
via flang-commits
flang-commits at lists.llvm.org
Mon May 4 04:20:08 PDT 2026
https://github.com/Ritanya-B-Bharadwaj created https://github.com/llvm/llvm-project/pull/195634
Fixes https://github.com/llvm/llvm-project/issues/180093
LINEAR variables on composite DO SIMD were being lowered onto omp.simd, which writes back unconditionally causing a race inside PARALLEL. Move them to omp.wsloop instead, which already has correct last-iter write-back with a barrier.
>From ca887764b7d4b0912ec35432c84357bbb395600f Mon Sep 17 00:00:00 2001
From: Ritanya B Bharadwaj <ritanya.b.bharadwaj at gmail.com>
Date: Mon, 4 May 2026 06:14:29 -0500
Subject: [PATCH] [flang][OpenMP] Fix data race with LINEAR clause on composite
DO SIMD
---
flang/lib/Lower/OpenMP/OpenMP.cpp | 24 +++++++++++
.../Lower/OpenMP/composite_simd_linear.f90 | 41 +++++++++++--------
flang/test/Lower/OpenMP/linear_modifier.f90 | 11 ++---
3 files changed, 53 insertions(+), 23 deletions(-)
diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index 88d28cf94b045..ed36df5ddcc46 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -3527,6 +3527,17 @@ static mlir::omp::DistributeOp genCompositeDistributeParallelDoSimd(
genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps,
simdReductionSyms, &reductionVarCache);
+ // Same as genCompositeDoSimd.
+ if (!simdClauseOps.linearVars.empty()) {
+ wsloopClauseOps.linearVars = std::move(simdClauseOps.linearVars);
+ wsloopClauseOps.linearStepVars = std::move(simdClauseOps.linearStepVars);
+ wsloopClauseOps.linearVarTypes = simdClauseOps.linearVarTypes;
+ wsloopClauseOps.linearModifiers = simdClauseOps.linearModifiers;
+ simdClauseOps.linearVars.clear();
+ simdClauseOps.linearStepVars.clear();
+ simdClauseOps.linearVarTypes = nullptr;
+ simdClauseOps.linearModifiers = nullptr;
+ }
DataSharingProcessor simdItemDSP(converter, semaCtx, simdItem->clauses, eval,
/*shouldCollectPreDeterminedSymbols=*/true,
/*useDelayedPrivatization=*/true, symTable);
@@ -3662,6 +3673,19 @@ static mlir::omp::WsloopOp genCompositeDoSimd(
genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps,
simdReductionSyms, &reductionVarCache);
+ // omp.simd writes back linear vars unconditionally, causing a race when
+ // inside a parallel region. Move them to wsloop which has proper last-iter
+ // write-back guarded by a barrier.
+ if (!simdClauseOps.linearVars.empty()) {
+ wsloopClauseOps.linearVars = std::move(simdClauseOps.linearVars);
+ wsloopClauseOps.linearStepVars = std::move(simdClauseOps.linearStepVars);
+ wsloopClauseOps.linearVarTypes = simdClauseOps.linearVarTypes;
+ wsloopClauseOps.linearModifiers = simdClauseOps.linearModifiers;
+ simdClauseOps.linearVars.clear();
+ simdClauseOps.linearStepVars.clear();
+ simdClauseOps.linearVarTypes = nullptr;
+ simdClauseOps.linearModifiers = nullptr;
+ }
DataSharingProcessor wsloopItemDSP(
converter, semaCtx, doItem->clauses, eval,
/*shouldCollectPreDeterminedSymbols=*/false,
diff --git a/flang/test/Lower/OpenMP/composite_simd_linear.f90 b/flang/test/Lower/OpenMP/composite_simd_linear.f90
index 38ef80292326d..dbe9961cafbdc 100644
--- a/flang/test/Lower/OpenMP/composite_simd_linear.f90
+++ b/flang/test/Lower/OpenMP/composite_simd_linear.f90
@@ -7,13 +7,13 @@ subroutine do_simd
!CHECK: %[[X:.*]]:2 = hlfir.declare %{{.*}} {uniq_name = "_QFdo_simdEx"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
!CHECK: %[[CONST:.*]] = arith.constant 1 : i32
!CHECK: %{{.*}} = arith.constant 1 : i32
-!CHECK: %[[IV_STEP:.*]] = arith.constant 1 : i32
-!CHECK: omp.wsloop {
-!DEFAULT: omp.simd linear(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32, %[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32) private(@_QFdo_simdEi_private_i32 {{.*}} -> %arg0 : !fir.ref<i32>) {
-!OPENMP52: omp.simd linear(val(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32), val(%[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32)) private(@_QFdo_simdEi_private_i32 {{.*}} -> %arg0 : !fir.ref<i32>) {
+!CHECK: %{{.*}} = arith.constant 1 : i32
+!DEFAULT: omp.wsloop linear(%[[X]]#0 : !fir.ref<i32> = {{.*}}) {
+!OPENMP52: omp.wsloop linear(val(%[[X]]#0 : !fir.ref<i32> = {{.*}})) {
+!CHECK: omp.simd linear({{.*}}) private(@_QFdo_simdEi_private_i32 {{.*}} -> %arg0 : !fir.ref<i32>) {
!CHECK: }
-!CHECK: } {linear_var_types = [i32, i32], omp.composite}
-!CHECK: } {omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
integer :: x
!$omp do simd linear(x:1)
do i = 1, N
@@ -45,15 +45,18 @@ subroutine distribute_parallel_do
!CHECK: omp.teams {
!CHECK: omp.parallel {
!CHECK: %[[CONST]] = arith.constant 1 : i32
+!CHECK: %{{.*}} = arith.constant 1 : i32
+!CHECK: %[[IV_STEP:.*]] = arith.constant 1 : i32
!CHECK: omp.distribute {
-!CHECK: omp.wsloop {
-!DEFAULT: omp.simd linear(%[[I]]#0 : !fir.ref<i32> = %[[CONST]] : i32) private(@_QFdistribute_parallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
-!OPENMP52: omp.simd linear(val(%[[I]]#0 : !fir.ref<i32> = %[[CONST]] : i32)) private(@_QFdistribute_parallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
+!DEFAULT: omp.wsloop linear(%[[I]]#0 : !fir.ref<i32> = {{.*}}) {
+!OPENMP52: omp.wsloop linear(val(%[[I]]#0 : !fir.ref<i32> = {{.*}})) {
+!CHECK: omp.simd linear({{.*}}) private(@_QFdistribute_parallel_doEi_private_i32 {{.*}}) {
!$omp teams
!$omp distribute parallel do simd linear(i:1)
do i = 1, N
end do
!$omp end distribute parallel do simd
+!CHECK: } {linear_var_types = [i32], omp.composite}
!CHECK: } {linear_var_types = [i32], omp.composite}
!$omp end teams
end subroutine distribute_parallel_do
@@ -64,16 +67,17 @@ subroutine parallel_do
!CHECK: omp.parallel {
!CHECK: %[[LINEAR_STEP:.*]] = arith.constant 2 : i32
!CHECK: %{{.*}} = arith.constant 1 : i32
-!CHECK: %[[IV_STEP:.*]] = arith.constant 1 : i32
-!CHECK: omp.wsloop {
-!DEFAULT: omp.simd linear(%[[X]]#0 : !fir.ref<i32> = %[[LINEAR_STEP]] : i32, %[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32) private(@_QFparallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
-!OPENMP52: omp.simd linear(val(%[[X]]#0 : !fir.ref<i32> = %[[LINEAR_STEP]] : i32), val(%[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32)) private(@_QFparallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
+!CHECK: %{{.*}} = arith.constant 1 : i32
+!DEFAULT: omp.wsloop linear(%[[X]]#0 : !fir.ref<i32> = {{.*}}) {
+!OPENMP52: omp.wsloop linear(val(%[[X]]#0 : !fir.ref<i32> = {{.*}})) {
+!CHECK: omp.simd linear({{.*}}) private(@_QFparallel_doEi_private_i32 {{.*}}) {
integer :: x
!$omp parallel do simd linear(x:2)
do i = 1, N
end do
!$omp end parallel do simd
-!CHECK: } {linear_var_types = [i32, i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
end subroutine parallel_do
subroutine teams_distribute
@@ -103,13 +107,14 @@ subroutine teams_distribute_parallel_do
!CHECK: %{{.*}} = arith.constant 1 : i32
!CHECK: %[[IV_STEP:.*]] = arith.constant 1 : i32
!CHECK: omp.distribute {
-!CHECK: omp.wsloop {
-!DEFAULT: omp.simd linear(%[[X]]#0 : !fir.ref<i32> = %c1_i32 : i32, %[[I]]#0 : !fir.ref<i32> = %c1_i32_1 : i32) private(@_QFteams_distribute_parallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
-!OPENMP52: omp.simd linear(val(%[[X]]#0 : !fir.ref<i32> = %c1_i32 : i32), val(%[[I]]#0 : !fir.ref<i32> = %c1_i32_1 : i32)) private(@_QFteams_distribute_parallel_doEi_private_i32 %[[I]]#0 -> %arg0 : !fir.ref<i32>) {
+!DEFAULT: omp.wsloop linear(%[[X]]#0 : !fir.ref<i32> = {{.*}}) {
+!OPENMP52: omp.wsloop linear(val(%[[X]]#0 : !fir.ref<i32> = {{.*}})) {
+!CHECK: omp.simd linear({{.*}}) private(@_QFteams_distribute_parallel_doEi_private_i32 {{.*}}) {
integer :: x
!$omp teams distribute parallel do simd linear(x)
do i = 1, N
end do
!$omp end teams distribute parallel do simd
-!CHECK: } {linear_var_types = [i32, i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
end subroutine teams_distribute_parallel_do
diff --git a/flang/test/Lower/OpenMP/linear_modifier.f90 b/flang/test/Lower/OpenMP/linear_modifier.f90
index 8364e5d698f06..cce565bec86d3 100644
--- a/flang/test/Lower/OpenMP/linear_modifier.f90
+++ b/flang/test/Lower/OpenMP/linear_modifier.f90
@@ -41,14 +41,15 @@ subroutine do_simd_linear
!CHECK: %[[CONST:.*]] = arith.constant 1 : i32
!CHECK: %{{.*}} = arith.constant 1 : i32
!CHECK: %[[IV_STEP:.*]] = arith.constant 1 : i32
-!CHECK: omp.wsloop {
-!OPENMP52: omp.simd linear(val(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32), val(%[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32)) private({{.*}}) {
-!OPENMP45: omp.simd linear(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32, %[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32) private({{.*}}) {
+!OPENMP52: omp.wsloop linear(val(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32)) {
+!OPENMP45: omp.wsloop linear(%[[X]]#0 : !fir.ref<i32> = %[[CONST]] : i32) {
+!OPENMP52: omp.simd linear(val(%[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32)) private({{.*}}) {
+!OPENMP45: omp.simd linear(%[[I]]#0 : !fir.ref<i32> = %[[IV_STEP]] : i32) private({{.*}}) {
integer :: x
!$omp do simd linear(x:1)
do i = 1, 10
end do
!$omp end do simd
-!CHECK: } {linear_var_types = [i32, i32], omp.composite}
-!CHECK: } {omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
+!CHECK: } {linear_var_types = [i32], omp.composite}
end subroutine do_simd_linear
More information about the flang-commits
mailing list