linkage property destroyed by BitcodeReader::Dematerialize()?

[Petar Jovanovic]

Hi everyone,

In NativeClient for MIPS, we have come across an issue [1] in which
NaClBitcodeReader, which mimics BitcodeReader, calls Dematerialize()
which calls F->deleteBody() eventually.
The problem is that deleteBody() also converts the linkage to external
and thus destroys original linkage type value. Lack of correct linkage
type causes wrong relocations to be emitted later.

For NaClBitcodeReader, a proposal is to change Dematerialize() to
call dropAllReferences() instead of deleteBody() like this:

Index: lib/Bitcode/NaCl/Reader/NaClBitcodeReader.cpp
diff --git a/lib/Bitcode/NaCl/Reader/NaClBitcodeReader.cpp b/lib/Bitcode/NaCl/Reader/NaClBitcodeReader.cpp
index d613d2d84836142db874b9a78b94c197ebf0f30f..d4465d17639f7f74114a08fc14b5ffc6a45b529c 100644
--- a/lib/Bitcode/NaCl/Reader/NaClBitcodeReader.cpp
+++ b/lib/Bitcode/NaCl/Reader/NaClBitcodeReader.cpp
@@ -1651,7 +1651,7 @@ void NaClBitcodeReader::Dematerialize(GlobalValue *GV) {
   assert(DeferredFunctionInfo.count(F) && "No info to read function later?");
 
   // Just forget the function body, we can remat it later.
-  F->deleteBody();
+  F->dropAllReferences();
 }
 

Still, there is a question if this change is applicable for
BitcodeReader::Dematerialize() as well. Any opinions?

Thanks in advance.

Kind regards,
Petar

[1] Linkage property gets destroyed by NaClBitcodeReader::Dematerialize(),
https://code.google.com/p/nativeclient/issues/detail?id=3943