DebugInfo: Hash DW_OP_convert in loclists when using Split DWARF

This code was incorrectly emitting extra bytes into arbitrary parts of
the object file when it was meant to be hashing them to compute the DWO
ID.

Follow-up patch(es) will refactor this API somewhat to make such bugs
harder to introduce, hopefully.

diff --git a/llvm/lib/CodeGen/AsmPrinter/ByteStreamer.h b/llvm/lib/CodeGen/AsmPrinter/ByteStreamer.h
index 09f7496..b60f68c 100644 (file)
--- a/llvm/lib/CodeGen/AsmPrinter/ByteStreamer.h
+++ b/llvm/lib/CodeGen/AsmPrinter/ByteStreamer.h
@@ -50,7 +50,7 @@ public:
   }
   void EmitULEB128(uint64_t DWord, const Twine &Comment, unsigned PadTo) override {
     AP.OutStreamer->AddComment(Comment);
-    AP.EmitULEB128(DWord);
+    AP.EmitULEB128(DWord, nullptr, PadTo);
   }
 };
 

diff --git a/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp b/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
index 0923394..447d3fa 100644 (file)
--- a/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
+++ b/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
@@ -2241,7 +2241,7 @@ void DwarfDebug::emitDebugLocEntry(ByteStreamer &Streamer,
         uint64_t Offset =
             CU->ExprRefedBaseTypes[Op.getRawOperand(I)].Die->getOffset();
         assert(Offset < (1ULL << (ULEB128PadSize * 7)) && "Offset wont fit");
-        Asm->EmitULEB128(Offset, nullptr, ULEB128PadSize);
+        Streamer.EmitULEB128(Offset, "", ULEB128PadSize);
         // Make sure comments stay aligned.
         for (unsigned J = 0; J < ULEB128PadSize; ++J)
           if (Comment != End)

diff --git a/llvm/test/DebugInfo/X86/convert-loclist.ll b/llvm/test/DebugInfo/X86/convert-loclist.ll
index 1d5b6f9..e510cd6 100644 (file)
--- a/llvm/test/DebugInfo/X86/convert-loclist.ll
+++ b/llvm/test/DebugInfo/X86/convert-loclist.ll
@@ -1,8 +1,26 @@
-; RUN: llc -dwarf-version=5 -filetype=obj -O0 < %s | llvm-dwarfdump -debug-info -debug-loclists - | FileCheck %s
-; RUN: llc -dwarf-version=5 -split-dwarf-file=foo.dwo -filetype=obj -O0 < %s | llvm-dwarfdump -debug-info -debug-loclists - | FileCheck %s
+; RUN: llc -dwarf-version=5 -filetype=obj -O0 < %s \
+; RUN:   | llvm-dwarfdump -debug-info -debug-loclists - | FileCheck %s
+; RUN: llc -dwarf-version=5 -split-dwarf-file=foo.dwo -filetype=obj -O0 < %s \
+; RUN:   | llvm-dwarfdump -debug-info -debug-loclists - | FileCheck --check-prefix=CHECK --check-prefix=SPLIT %s
+; RUN: llc -dwarf-version=5 -split-dwarf-file=foo.dwo -filetype=asm -O0 < %s | FileCheck --check-prefix=ASM %s
+
+; A bit of a brittle test - this is testing the specific DWO_id. The
+; alternative would be to test two files with different DW_OP_convert values &
+; ensuring the DWO IDs differ when the DW_OP_convert parameter differs.
+
+; So if this test ends up being a brittle pain to maintain, updating the DWO ID
+; often - add another IR file with a different DW_OP_convert that's otherwise
+; identical and demonstrate that they have different DWO IDs.
+
+; SPLIT: 0x00000000: Compile Unit: {{.*}} DWO_id = 0xafd73565c68bc661
+
+; Regression testing a fairly quirky bug where instead of hashing (see above),
+; extra bytes would be emitted into the output assembly in no
+; particular/intentional section - so let's check they don't show up at all:
+; ASM-NOT: .asciz  "\200\200\200"
 
 ; CHECK: 0x{{0*}}[[TYPE:.*]]: DW_TAG_base_type
-; CHECK-NEXT:                DW_AT_name ("DW_ATE_unsigned_32")
+; CHECK-NEXT:                   DW_AT_name ("DW_ATE_unsigned_32")
 
 ; CHECK: DW_LLE_offset_pair ({{.*}}): DW_OP_consts +7, DW_OP_convert 0x[[TYPE]], DW_OP_stack_value