[RemoveDIs] Don't convert debug-info in bitcode-loading just now (#80865)

We've been building and testing this no-debug-intrinsic work inside of
the pass manager for a while, so that optimisation passes get exercised
and tested when we turn it on. However, by converting to the
non-intrinsic form in the bitcode loader, we accidentally caused all
parts of LLVM to potentially see non-intrinsic debug-info.

Seeing how we're trying to turn things on incrementally, it was a
mistake to go this far this fast: we can instead just focus on enabling
during optimisations for the moment, then all the other parts of LLVM
later.

Author: jmorse

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -6632,9 +6632,6 @@ Error BitcodeReader::materialize(GlobalValue *GV) {
   if (Error Err = materializeMetadata())
     return Err;
 
-  bool NewDebugInfoRequested = F->IsNewDbgInfoFormat;
-  F->IsNewDbgInfoFormat = false;
-
   // Move the bit stream to the saved position of the deferred function body.
   if (Error JumpFailed = Stream.JumpToBit(DFII->second))
     return JumpFailed;
@@ -6710,14 +6707,6 @@ Error BitcodeReader::materialize(GlobalValue *GV) {
   // Look for functions that rely on old function attribute behavior.
   UpgradeFunctionAttributes(*F);
 
-  // If we've materialized a function set up in "new" debug-info mode, the
-  // contents just loaded will still be in dbg.value mode. Switch to the new
-  // mode now. NB: we can add more complicated logic here in the future to
-  // correctly identify when we do and don't need to autoupgrade.
-  if (NewDebugInfoRequested) {
-    F->convertToNewDbgValues();
-  }
-
   // Bring in any functions that this function forward-referenced via
   // blockaddresses.
   return materializeForwardReferencedFunctions();
@@ -8042,14 +8031,6 @@ BitcodeModule::getModuleImpl(LLVMContext &Context, bool MaterializeAll,
       return std::move(Err);
   }
 
-  // If we are operating in a "new debug-info" context, upgrade the debug-info
-  // in the loaded module. This is a transitional approach as we enable "new"
-  // debug-info in LLVM, which will eventually be pushed down into the
-  // autoupgrade path once the bitcode-encoding is finalised. Non-materialised
-  // functions will be upgraded in the materialize method.
-  if (UseNewDbgInfoFormat && !M->IsNewDbgInfoFormat)
-    M->convertToNewDbgValues();
-
   return std::move(M);
 }
 

llvm/lib/Linker/IRMover.cpp

@@ -1775,6 +1775,8 @@ IRMover::IRMover(Module &M) : Composite(M) {
 Error IRMover::move(std::unique_ptr<Module> Src,
                     ArrayRef<GlobalValue *> ValuesToLink,
                     LazyCallback AddLazyFor, bool IsPerformingImport) {
+  if (getModule().IsNewDbgInfoFormat)
+    Src->convertToNewDbgValues();
   IRLinker TheIRLinker(Composite, SharedMDs, IdentifiedStructTypes,
                        std::move(Src), ValuesToLink, std::move(AddLazyFor),
                        IsPerformingImport);

llvm/test/ThinLTO/X86/crash_debuginfo.ll

@@ -4,6 +4,9 @@
 ; RUN: opt -passes=function-import,inline -summary-file %t-index.thinlto.bc %t-dst.bc -o %t.out
 ; RUN: llvm-nm -U %t.out | FileCheck %s --implicit-check-not=_bar
 
+; Ensure we can do the same thing in RemoveDIs mode.
+; RUN: opt -passes=function-import,inline -summary-file %t-index.thinlto.bc %t-dst.bc -o %t.out --try-experimental-debuginfo-iterators
+
 ; Verify that we import bar and inline it. It use to crash importing due to ODR type uniquing
 ; CHECK: _foo