[Mlir-commits] [mlir] e959a28 - [mlir][openacc] Do not erase acc routine information early (#210813)

[llvmlistbot at llvm.org]

Author: Valentin Clement (バレンタイン クレメン)
Date: 2026-07-21T09:47:51-07:00
New Revision: e959a280d5ac91d219abb802913c4e5cadb45f2a

URL: https://github.com/llvm/llvm-project/commit/e959a280d5ac91d219abb802913c4e5cadb45f2a
DIFF: https://github.com/llvm/llvm-project/commit/e959a280d5ac91d219abb802913c4e5cadb45f2a.diff

LOG: [mlir][openacc] Do not erase acc routine information early (#210813)

acc routine information is used in ACCCGToGPU to compute parallelism. If
the info is missing, predicated region might be wrong and lead to
deadlock.

Just keep them in ACCRountineToGPUFunc and they will be cleaned up later
in the pipeline.

Added: 
    mlir/test/Dialect/OpenACC/acc-cg-to-gpu-bound-routine-call.mlir

Modified: 
    mlir/lib/Dialect/OpenACC/Transforms/ACCRoutineToGPUFunc.cpp

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/OpenACC/Transforms/ACCRoutineToGPUFunc.cpp b/mlir/lib/Dialect/OpenACC/Transforms/ACCRoutineToGPUFunc.cpp
index 27a34695decb2..6bbb4b53cb5bb 100644
--- a/mlir/lib/Dialect/OpenACC/Transforms/ACCRoutineToGPUFunc.cpp
+++ b/mlir/lib/Dialect/OpenACC/Transforms/ACCRoutineToGPUFunc.cpp
@@ -320,8 +320,6 @@ class ACCRoutineToGPUFunc
       return signalPassFailure();
 
     cleanupHostModule(funcsToClone);
-    for (RoutineOp bindOp : bindAccRoutines)
-      bindOp.erase();
   }
 };
 

diff  --git a/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-bound-routine-call.mlir b/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-bound-routine-call.mlir
new file mode 100644
index 0000000000000..c55f87b73d181
--- /dev/null
+++ b/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-bound-routine-call.mlir
@@ -0,0 +1,28 @@
+// RUN: mlir-opt %s --pass-pipeline="builtin.module(func.func(acc-bind-routine),acc-routine-to-gpu-func,func.func(acc-cg-to-gpu))" | FileCheck %s
+
+// Bound acc.routine metadata must remain available after ACCBindRoutine and
+// ACCRoutineToGPUFunc so ACCCGToGPU can classify the bound callee as vector.
+// Otherwise the call inside acc.predicate_region is incorrectly predicated to
+// thread_id.x == 0.
+
+// CHECK-LABEL: func.func @bound_vector_call
+// CHECK: gpu.launch
+// CHECK-NOT: scf.if
+// CHECK: func.call @bound_vector
+func.func @bound_vector_call(%arg0: memref<4xf32>) {
+  %c32 = arith.constant 32 : index
+  %thread_x = acc.par_width %c32 {par_dim = #acc.par_dim<thread_x>}
+  acc.compute_region launch(%tx = %thread_x) ins(%arg10 = %arg0) : (memref<4xf32>) {
+    acc.predicate_region {
+      func.call @wrapped_vector(%arg10) : (memref<4xf32>) -> ()
+    }
+    acc.yield
+  } {origin = "acc.parallel"}
+  return
+}
+
+func.func private @wrapped_vector(memref<4xf32>)
+    attributes {acc.routine_info = #acc.routine_info<[@acc_routine_bound_vector]>}
+acc.routine @acc_routine_bound_vector func(@wrapped_vector) vector bind("bound_vector")
+func.func private @bound_vector(memref<4xf32>)
+    attributes {acc.routine_info = #acc.routine_info<[@acc_routine_bound_vector]>}