[Mlir-commits] [mlir] [MLIR][OpenMP] Preserve the debug location for the reduction barrier. (PR #220605)

[Abid Qadeer]

https://github.com/abidh created https://github.com/llvm/llvm-project/pull/220605

`createReductionsAndCleanup` passed the insert point returned by `createReductions` straight into `createBarrier`. That selects` LocationDescription(const InsertPointTy &)`, which leaves the debug location default-constructed, so the `__kmpc_barrier` and the `__kmpc_global_thread_num` generated by `createBarrier` are emitted without a !dbg.

That breaks the device link of a target region compiled with -g once a device runtime built with debug info is present. The verifier fails with the following error:

  inlinable function call in a function with debug info must have a !dbg location
    %31 = call i32 @__kmpc_global_thread_num
  inlinable function call in a function with debug info must have a !dbg location
    call void @__kmpc_barrier
  LLVM ERROR: Broken module found, compilation aborted!

This is the same class as the sites converted in #211254. Fix it by passing a valid debug location.

>From 28a2bd28e5e7dd96d3188d8d042509f73c512e4b Mon Sep 17 00:00:00 2001
From: Abid Qadeer <haqadeer at amd.com>
Date: Fri, 21 Aug 2026 20:03:55 +0100
Subject: [PATCH] [MLIR][OpenMP] Preserve the debug location for the reduction
 barrier.

createReductionsAndCleanup passed the insert point returned by
createReductions straight into createBarrier. That selects
LocationDescription(const InsertPointTy &), which leaves the debug
location default-constructed, so the __kmpc_barrier and the
__kmpc_global_thread_num generated by createBarrier are emitted without
a !dbg.

That breaks the device link of a target region compiled with -g once a
device runtime built with debug info is present. The verifier fails with
the following error:

  inlinable function call in a function with debug info must have a !dbg location
    %31 = call i32 @__kmpc_global_thread_num
  inlinable function call in a function with debug info must have a !dbg location
    call void @__kmpc_barrier
  LLVM ERROR: Broken module found, compilation aborted!

This is the same class as the sites converted in #211254. Fix it by
passing a valid debug location.

Co-authored-by: Cursor <cursoragent at cursor.com>
---
 .../OpenMP/OpenMPToLLVMIRTranslation.cpp      |  4 +-
 .../openmp-reduction-barrier-debug-loc.mlir   | 51 +++++++++++++++++++
 2 files changed, 54 insertions(+), 1 deletion(-)
 create mode 100644 mlir/test/Target/LLVMIR/openmp-reduction-barrier-debug-loc.mlir

diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index e66a699f7071e..cedc7b95e577a 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -1719,6 +1719,7 @@ static LogicalResult createReductionsAndCleanup(
   // and remove it later.
   llvm::UnreachableInst *tempTerminator = builder.CreateUnreachable();
   builder.SetInsertPoint(tempTerminator);
+  llvm::DebugLoc reductionLoc = builder.getCurrentDebugLocation();
   llvm::OpenMPIRBuilder::InsertPointOrErrorTy contInsertPoint =
       ompBuilder->createReductions(builder, allocaIP, reductionInfos, isByRef,
                                    isNowait, isTeamsReduction);
@@ -1732,7 +1733,8 @@ static LogicalResult createReductionsAndCleanup(
   llvm::OpenMPIRBuilder::InsertPointTy afterIP = *contInsertPoint;
   if (!isTeamsReduction) {
     llvm::OpenMPIRBuilder::InsertPointOrErrorTy barrierIP =
-        ompBuilder->createBarrier(*contInsertPoint, llvm::omp::OMPD_for);
+        ompBuilder->createBarrier({*contInsertPoint, reductionLoc},
+                                  llvm::omp::OMPD_for);
 
     if (failed(handleError(barrierIP, *op)))
       return failure();
diff --git a/mlir/test/Target/LLVMIR/openmp-reduction-barrier-debug-loc.mlir b/mlir/test/Target/LLVMIR/openmp-reduction-barrier-debug-loc.mlir
new file mode 100644
index 0000000000000..dffbfc3e4a9a3
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-reduction-barrier-debug-loc.mlir
@@ -0,0 +1,51 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+// subroutine reduce(s)
+//   integer :: i, s
+//   !$omp do reduction(+:s)
+//   do i = 1, 10
+//      s = s + i
+//   end do
+// end subroutine
+
+#file = #llvm.di_file<"reduce.f90" in "">
+#void = #llvm.di_null_type
+#cu = #llvm.di_compile_unit<id = distinct[0]<>, sourceLanguage = DW_LANG_Fortran95, file = #file, producer = "flang", isOptimized = false, emissionKind = Full>
+#sub_type = #llvm.di_subroutine_type<callingConvention = DW_CC_normal, types = #void>
+#sp = #llvm.di_subprogram<id = distinct[1]<>, compileUnit = #cu, scope = #file, name = "reduce", linkageName = "reduce_", file = #file, line = 1, scopeLine = 1, subprogramFlags = "Definition", type = #sub_type>
+
+#loc1 = loc("reduce.f90":1:1)
+#loc2 = loc("reduce.f90":3:9)
+#loc_fn = loc(fused<#sp>[#loc1])
+
+module attributes {omp.version = #omp.version<version = 31>} {
+  omp.declare_reduction @add_i32 : i32 init {
+  ^bb0(%arg0: i32):
+    %0 = llvm.mlir.constant(0 : i32) : i32 loc(#loc2)
+    omp.yield(%0 : i32) loc(#loc2)
+  } combiner {
+  ^bb0(%arg0: i32, %arg1: i32):
+    %0 = llvm.add %arg0, %arg1 : i32 loc(#loc2)
+    omp.yield(%0 : i32) loc(#loc2)
+  } loc(#loc2)
+
+  llvm.func @reduce_(%arg0: !llvm.ptr) {
+    %c1 = llvm.mlir.constant(1 : i32) : i32 loc(#loc2)
+    %cn = llvm.mlir.constant(10 : i32) : i32 loc(#loc2)
+    omp.wsloop reduction(@add_i32 %arg0 -> %prv : !llvm.ptr) {
+      omp.loop_nest (%iv) : i32 = (%c1) to (%cn) inclusive step (%c1) {
+        %0 = llvm.load %prv : !llvm.ptr -> i32 loc(#loc2)
+        %1 = llvm.add %0, %iv : i32 loc(#loc2)
+        llvm.store %1, %prv : i32, !llvm.ptr loc(#loc2)
+        omp.yield loc(#loc2)
+      } loc(#loc2)
+    } loc(#loc2)
+    llvm.return loc(#loc_fn)
+  } loc(#loc_fn)
+}
+
+// CHECK-LABEL: define void @reduce_(
+// CHECK-SAME:      ptr %{{.*}}) !dbg
+// CHECK:       reduce.finalize:
+// CHECK-NEXT:    %{{.*}} = call i32 @__kmpc_global_thread_num({{.*}}), !dbg
+// CHECK-NEXT:    call void @__kmpc_barrier({{.*}}), !dbg