[MLIR][LLVM] Translate Debug EmissionKind

[zyx-billy]

Translate debug emission kind into LLVM (the importer already supports this).

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-llvm

Author: Billy Zhu (zyx-billy)

Changes

Translate debug emission kind into LLVM (the importer already supports this).

---

Full diff: https://github.com/llvm/llvm-project/pull/74376.diff

2 Files Affected:

• (modified) mlir/lib/Target/LLVMIR/DebugTranslation.cpp (+3-1)
• (modified) mlir/test/Target/LLVMIR/llvmir-debug.mlir (+21)

diff --git a/mlir/lib/Target/LLVMIR/DebugTranslation.cpp b/mlir/lib/Target/LLVMIR/DebugTranslation.cpp
index 6d845e27ffa24..26950d7287304 100644
--- a/mlir/lib/Target/LLVMIR/DebugTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/DebugTranslation.cpp
@@ -102,7 +102,9 @@ llvm::DICompileUnit *DebugTranslation::translateImpl(DICompileUnitAttr attr) {
       attr.getSourceLanguage(), translate(attr.getFile()),
       attr.getProducer() ? attr.getProducer().getValue() : "",
       attr.getIsOptimized(),
-      /*Flags=*/"", /*RV=*/0);
+      /*Flags=*/"", /*RV=*/0, /*SplitName=*/{},
+      static_cast<llvm::DICompileUnit::DebugEmissionKind>(
+          attr.getEmissionKind()));
 }
 
 /// Returns a new `DINodeT` that is either distinct or not, depending on
diff --git a/mlir/test/Target/LLVMIR/llvmir-debug.mlir b/mlir/test/Target/LLVMIR/llvmir-debug.mlir
index e62ee35f6c007..3cd995d4c294f 100644
--- a/mlir/test/Target/LLVMIR/llvmir-debug.mlir
+++ b/mlir/test/Target/LLVMIR/llvmir-debug.mlir
@@ -286,3 +286,24 @@ llvm.func @dbg_intrinsics_with_no_location(%arg0: i32) -> (i32) {
 #di_basic_type_2 = #llvm.di_basic_type<tag = DW_TAG_base_type, name = "uint64_t", sizeInBits = 64, encoding = DW_ATE_unsigned>
 llvm.mlir.global external @global_with_expr_1() {addr_space = 0 : i32, dbg_expr = #llvm.di_global_variable_expression<var = <scope = #di_compile_unit_2, name = "global_with_expr_1", linkageName = "global_with_expr_1", file = #di_file_2, line = 370, type = #di_basic_type_2>, expr = <>>} : i64
 llvm.mlir.global external @global_with_expr_2() {addr_space = 0 : i32, dbg_expr = #llvm.di_global_variable_expression<var = <scope = #di_compile_unit_2, name = "global_with_expr_2", linkageName = "global_with_expr_2", file = #di_file_2, line = 371, type = #di_basic_type_2, isLocalToUnit = true, isDefined = true, alignInBits = 8>, expr = <>>} : i64
+
+// -----
+
+// CHECK-DAG: ![[FILE1:.*]] = !DIFile(filename: "foo1.mlir", directory: "/test/")
+// CHECK-DAG: ![[FILE2:.*]] = !DIFile(filename: "foo2.mlir", directory: "/test/")
+#di_file_1 = #llvm.di_file<"foo1.mlir" in "/test/">
+#di_file_2 = #llvm.di_file<"foo2.mlir" in "/test/">
+// CHECK-DAG: ![[SCOPE1:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE1]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly)
+// CHECK-DAG: ![[SCOPE2:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE2]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: DebugDirectivesOnly)
+#di_compile_unit_1 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_1, producer = "MLIR", isOptimized = true, emissionKind = LineTablesOnly>
+#di_compile_unit_2 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_2, producer = "MLIR", isOptimized = true, emissionKind = DebugDirectivesOnly>
+#di_subprogram_1 = #llvm.di_subprogram<compileUnit = #di_compile_unit_1, scope = #di_file_1, name = "func1", file = #di_file_1, subprogramFlags = "Definition|Optimized">
+#di_subprogram_2 = #llvm.di_subprogram<compileUnit = #di_compile_unit_2, scope = #di_file_2, name = "func2", file = #di_file_2, subprogramFlags = "Definition|Optimized">
+
+llvm.func @func_line_tables() {
+  llvm.return
+} loc(fused<#di_subprogram_1>["foo1.mlir":0:0])
+
+llvm.func @func_debug_directives() {
+  llvm.return
+} loc(fused<#di_subprogram_2>["foo2.mlir":0:0])

[gysit]

Thanks for fixing!

LGTM

[gysit]

Suggested change

	// CHECK-DAG: ![[FILE1:.*]] = !DIFile(filename: "foo1.mlir", directory: "/test/")
	// CHECK-DAG: ![[FILE2:.*]] = !DIFile(filename: "foo2.mlir", directory: "/test/")
	#di_file_1 = #llvm.di_file<"foo1.mlir" in "/test/">
	#di_file_2 = #llvm.di_file<"foo2.mlir" in "/test/">
	// CHECK-DAG: ![[SCOPE1:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE1]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly)
	// CHECK-DAG: ![[SCOPE2:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE2]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: DebugDirectivesOnly)
	#di_compile_unit_1 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_1, producer = "MLIR", isOptimized = true, emissionKind = LineTablesOnly>
	#di_compile_unit_2 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_2, producer = "MLIR", isOptimized = true, emissionKind = DebugDirectivesOnly>
	// CHECK-DAG: ![[FILE1:.*]] = !DIFile(filename: "foo1.mlir", directory: "/test/")
	#di_file_1 = #llvm.di_file<"foo1.mlir" in "/test/">
	// CHECK-DAG: ![[FILE2:.*]] = !DIFile(filename: "foo2.mlir", directory: "/test/")
	#di_file_2 = #llvm.di_file<"foo2.mlir" in "/test/">
	// CHECK-DAG: ![[SCOPE1:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE1]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly)
	#di_compile_unit_1 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_1, producer = "MLIR", isOptimized = true, emissionKind = LineTablesOnly>
	// CHECK-DAG: ![[SCOPE2:.*]] = distinct !DICompileUnit(language: DW_LANG_C, file: ![[FILE2]], producer: "MLIR", isOptimized: true, runtimeVersion: 0, emissionKind: DebugDirectivesOnly)
	#di_compile_unit_2 = #llvm.di_compile_unit<sourceLanguage = DW_LANG_C, file = #di_file_2, producer = "MLIR", isOptimized = true, emissionKind = DebugDirectivesOnly>

ultra nit: I would try to put the check line right above the line you are matching.

[gysit]

I wonder why we did not choose an LLVM enum here... That way we would have a guarantee that the two enum values match. However, this is orthogonal to your revision.

[zyx-billy]

Yeah there are other enums like DISubprogramFlags that are also just redefined and rely on static_cast. Might be a good idea to reuse llvm enums if possible, or set up explicit mappings?

[gysit]

Yeah I guess I would prefer if we use LLVM enums since there are no large switch statements needed (I assume that is what you refer to by explicit mappings)? But this is unrelated to your PR. I just realized that things are a bit inconsistent in that regard.

[zyx-billy]

Yeah that's what I meant. That makes sense. We can leave it at this then for this PR. Can you help me click merge once it's ready? Thanks! 🙏

[gysit]

sure!