[llvm] c985f28 - [OMPIRBuilder] Hoist alloca's to entry blocks of compiler-emitted GPU reduction functions (#181359)

[via llvm-commits]

Author: Kareem Ergawy
Date: 2026-02-17T09:08:17+01:00
New Revision: c985f285e4911affa46ea79e21263cbaf4c8322f

URL: https://github.com/llvm/llvm-project/commit/c985f285e4911affa46ea79e21263cbaf4c8322f
DIFF: https://github.com/llvm/llvm-project/commit/c985f285e4911affa46ea79e21263cbaf4c8322f.diff

LOG: [OMPIRBuilder] Hoist alloca's to entry blocks of compiler-emitted GPU reduction functions (#181359)

Fixes a bug in GPU reductions when `-O0` was used to compile GPU
reductions. There were invalid memory accesses at runtime for the
following example:

```fortran
program test_array_reduction()
  integer :: red_array(1)
  integer :: i

  red_array = 0

  !$omp target teams distribute parallel do reduction(+:red_array)
  do i = 1, 100
    red_array(1) = red_array(1) + 4422
  end do
  !$omp end target teams distribute parallel do

  print *, red_array
end program test_array_reduction
```

The issue was caused by alloca's for some temp values in the combiner
region of the reduction op being inlined beyond the entry blocks of the
GPU reduction functions emitted by the compiler.

This PR fixes the issue by hoisting all alloca's to the entry block
after the reduction functions are completely emitted by the compiler.

Added: 
    

Modified: 
    llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
    mlir/test/Target/LLVMIR/allocatable_gpu_reduction.mlir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index 464ec5b5a2ece..1947323ef85fd 100644
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -787,6 +787,13 @@ static void hoistNonEntryAllocasToEntryBlock(llvm::BasicBlock &Block) {
     AllocaInst->moveBefore(InsertPoint);
 }
 
+static void hoistNonEntryAllocasToEntryBlock(llvm::Function *Func) {
+  PostDominatorTree PostDomTree(*Func);
+  for (llvm::BasicBlock &BB : *Func)
+    if (PostDomTree.properlyDominates(&BB, &Func->getEntryBlock()))
+      hoistNonEntryAllocasToEntryBlock(BB);
+}
+
 void OpenMPIRBuilder::finalize(Function *Fn) {
   SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
   SmallVector<BasicBlock *, 32> Blocks;
@@ -893,12 +900,8 @@ void OpenMPIRBuilder::finalize(Function *Fn) {
     if (OI.PostOutlineCB)
       OI.PostOutlineCB(*OutlinedFn);
 
-    if (OI.FixUpNonEntryAllocas) {
-      PostDominatorTree PostDomTree(*OutlinedFn);
-      for (llvm::BasicBlock &BB : *OutlinedFn)
-        if (PostDomTree.properlyDominates(&BB, &OutlinedFn->getEntryBlock()))
-          hoistNonEntryAllocasToEntryBlock(BB);
-    }
+    if (OI.FixUpNonEntryAllocas)
+      hoistNonEntryAllocasToEntryBlock(OutlinedFn);
   }
 
   // Remove work items that have been completed.
@@ -4234,6 +4237,13 @@ Expected<Function *> OpenMPIRBuilder::createReductionFunction(
     }
 
   Builder.CreateRetVoid();
+  // Compiling with `-O0`, `alloca`s emitted in non-entry blocks are not hoisted
+  // to the entry block (this is dones for higher opt levels by later passes in
+  // the pipeline). This has caused issues because non-entry `alloca`s force the
+  // function to use dynamic stack allocations and we might run out of scratch
+  // memory.
+  hoistNonEntryAllocasToEntryBlock(ReductionFunc);
+
   return ReductionFunc;
 }
 

diff  --git a/mlir/test/Target/LLVMIR/allocatable_gpu_reduction.mlir b/mlir/test/Target/LLVMIR/allocatable_gpu_reduction.mlir
index 95d12f304aca0..c9ff6de8cc951 100644
--- a/mlir/test/Target/LLVMIR/allocatable_gpu_reduction.mlir
+++ b/mlir/test/Target/LLVMIR/allocatable_gpu_reduction.mlir
@@ -24,6 +24,8 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<"dlti.alloca_memory_space" = 5 :
     "llvm.intr.memcpy"(%5, %arg0, %6) <{isVolatile = false}> : (!llvm.ptr, !llvm.ptr, i32) -> ()
     %7 = llvm.mlir.constant(24 : i32) : i32
     "llvm.intr.memcpy"(%2, %arg1, %7) <{isVolatile = false}> : (!llvm.ptr, !llvm.ptr, i32) -> ()
+    llvm.br ^bb1
+  ^bb1:
     %8 = llvm.getelementptr %5[0, 0] : (!llvm.ptr) -> !llvm.ptr, !llvm.struct<(ptr, i64, i32, i8, i8, i8, i8)>
     %9 = llvm.load %8 : !llvm.ptr -> !llvm.ptr
     %10 = llvm.getelementptr %2[0, 0] : (!llvm.ptr) -> !llvm.ptr, !llvm.struct<(ptr, i64, i32, i8, i8, i8, i8)>
@@ -63,6 +65,21 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<"dlti.alloca_memory_space" = 5 :
   }
 }
 
+// CHECK: define internal void @"{{.*}}$reduction$reduction_func"(ptr noundef %[[ARG_0:.*]], ptr noundef %[[ARG_1:.*]]) {{.*}} {
+// CHECK: entry:
+// CHECK:   %[[TEMP_1:.*]] = alloca { ptr, i64, i32, i8, i8, i8, i8 }
+// CHECK:   %[[TEMP_2:.*]] = alloca { ptr, i64, i32, i8, i8, i8, i8 }
+// CHECK:   br label %[[RED_BODY:omp.reduction.nonatomic.body]]
+// Verify that no allocas are emitted beyond the entry block.
+// CHECK-NOT: alloca
+
+// CHECK: [[RED_BODY]]:
+// CHECK:   %[[TEMP_1_ACAST:.*]] = addrspacecast ptr addrspace(5) %[[TEMP_1]] to ptr
+// CHECK:   %[[TEMP_2_ACAST:.*]] = addrspacecast ptr addrspace(5) %[[TEMP_2]] to ptr
+// CHECK:   call void @llvm.memcpy.p0.p0.i32(ptr %[[TEMP_2_ACAST]], ptr %{{.*}}, i32 24, i1 false)
+// CHECK:   call void @llvm.memcpy.p0.p0.i32(ptr %[[TEMP_1_ACAST]], ptr %{{.*}}, i32 24, i1 false)
+// CHECK: }
+
 // CHECK: define {{.*}} @_omp_reduction_shuffle_and_reduce_func({{.*}}) {{.*}} {
 // CHECK:   %[[REMOTE_RED_LIST:.omp.reduction.remote_reduce_list]] = alloca [1 x ptr], align 8, addrspace(5)
 // CHECK:   %[[RED_ELEM:.omp.reduction.element]] = alloca { ptr, i64, i32, i8, i8, i8, i8 }, align 8, addrspace(5)