[Mlir-commits] [mlir] [mlir] Use non thread-local allocator for DistinctAttr when threading is disabled (PR #128566)

[Artemiy Bulavin]

================
@@ -0,0 +1,19 @@
+// Test that the enable-debug-info-scope-on-llvm-func pass can create its
+// DI attributes when running in the crash reproducer thread,
+
+// RUN: mlir-opt --mlir-disable-threading --mlir-pass-pipeline-crash-reproducer=. \
+// RUN:          --pass-pipeline="builtin.module(ensure-debug-info-scope-on-llvm-func)" \
+// RUN:          --mlir-print-debuginfo %s | FileCheck %s
+
+module {
+  llvm.func @func_no_debug() {
+    llvm.return loc(unknown)
+  } loc(unknown)
+} loc(unknown)
+
+// CHECK-LABEL: llvm.func @func_no_debug()
+// CHECK: llvm.return loc(#loc
+// CHECK: loc(#loc[[LOC:[0-9]+]])
+// CHECK: #di_file = #llvm.di_file<"<unknown>" in "">
+// CHECK: #di_subprogram = #llvm.di_subprogram<id = distinct[{{.*}}]<>, compileUnit = #di_compile_unit, scope = #di_file, name = "func_no_debug", linkageName = "func_no_debug", file = #di_file, line = 1, scopeLine = 1, subprogramFlags = "Definition|Optimized", type = #di_subroutine_type>
----------------
abulavin wrote:

I wanted to add this as a regression test because I initially discovered this issue as a result of running crash reproduction with this debuginfo pass. I think that the crash reproduction in MLIR lacks comprehensive testing so a test that verifies that this pass runs successfully with a crash reproducer is better than nothing. I will trim the checks lines to only target the distinct attributes.

https://github.com/llvm/llvm-project/pull/128566