[llvm] e0b24d9 - [llvm-dis] Fix non-deterministic disassembly across multiple inputs (#110988)

[via llvm-commits]

Author: Peter Waller
Date: 2024-10-07T11:23:30+01:00
New Revision: e0b24d947570385c5b6c3afea46fea09fc75ddab

URL: https://github.com/llvm/llvm-project/commit/e0b24d947570385c5b6c3afea46fea09fc75ddab
DIFF: https://github.com/llvm/llvm-project/commit/e0b24d947570385c5b6c3afea46fea09fc75ddab.diff

LOG: [llvm-dis] Fix non-deterministic disassembly across multiple inputs (#110988)

Prior to this patch, the LLVMContext was shared across inputs to
llvm-dis.

Consequently, NamedStructTypes was shared across inputs, which impacts
StructType::setName - if a name was reused across inputs, it would get
renamed during construction of the struct type, leading to tricky to
diagnose confusion.

Added: 
    llvm/test/tools/llvm-dis/multiple-files-equivalent.ll

Modified: 
    llvm/tools/llvm-dis/llvm-dis.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/test/tools/llvm-dis/multiple-files-equivalent.ll b/llvm/test/tools/llvm-dis/multiple-files-equivalent.ll
new file mode 100644
index 00000000000000..21f7abb67b2866
--- /dev/null
+++ b/llvm/test/tools/llvm-dis/multiple-files-equivalent.ll
@@ -0,0 +1,15 @@
+; RUN: llvm-as -o %t0 %s
+; RUN: cp %t0 %t1
+; RUN: llvm-dis %t0 %t1
+; RUN: FileCheck %s < %t0.ll
+; RUN: FileCheck %s < %t1.ll
+
+; Test that if we disassemble the same bitcode twice, the type names are
+; unchanged between the two. This protects against a bug whereby state was
+; preserved across inputs and the types ended up with 
diff erent names.
+
+; CHECK: %Foo = type { ptr }
+%Foo = type { ptr }
+
+; CHECK: @foo = global %Foo zeroinitializer
+ at foo = global %Foo zeroinitializer

diff  --git a/llvm/tools/llvm-dis/llvm-dis.cpp b/llvm/tools/llvm-dis/llvm-dis.cpp
index 744201289b4a42..49acc9cd456ff4 100644
--- a/llvm/tools/llvm-dis/llvm-dis.cpp
+++ b/llvm/tools/llvm-dis/llvm-dis.cpp
@@ -191,10 +191,6 @@ int main(int argc, char **argv) {
   if (LoadBitcodeIntoNewDbgInfoFormat == cl::boolOrDefault::BOU_UNSET)
     LoadBitcodeIntoNewDbgInfoFormat = cl::boolOrDefault::BOU_TRUE;
 
-  LLVMContext Context;
-  Context.setDiagnosticHandler(
-      std::make_unique<LLVMDisDiagnosticHandler>(argv[0]));
-
   if (InputFilenames.size() < 1) {
     InputFilenames.push_back("-");
   } else if (InputFilenames.size() > 1 && !OutputFilename.empty()) {
@@ -204,6 +200,12 @@ int main(int argc, char **argv) {
   }
 
   for (const auto &InputFilename : InputFilenames) {
+    // Use a fresh context for each input to avoid state
+    // cross-contamination across inputs (e.g. type name collisions).
+    LLVMContext Context;
+    Context.setDiagnosticHandler(
+        std::make_unique<LLVMDisDiagnosticHandler>(argv[0]));
+
     ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
         MemoryBuffer::getFileOrSTDIN(InputFilename);
     if (std::error_code EC = BufferOrErr.getError()) {