possible bug in COFFObjectFile::getSymbolType()

I’m working on adding data detection to llvm-objdump, so it prints out raw data for symbols with type SymbolRef::ST_Data instead of disassembling them. This causes llvm/test/MC/ARM/Windows/mov32t-range.s to fail, because the symbol “truncation” comes back as SymbolRef::ST_Data, and the code gets dumped as bytes instead of disassembled.

I traced the problem back to COFFObjectFile::getSymbolType() in llvm/lib/Object/COFFObjectFile.cpp :

if (Characteristics & COFF::IMAGE_SCN_MEM_READ &&

~Characteristics & COFF::IMAGE_SCN_MEM_WRITE) // Read only.

Result = SymbolRef::ST_Data;

I think it should also check for ~Characteristics & COFF::IMAGE_SCN_MEM_EXECUTE. The symbol, in this case, is READ, EXECUTE, !WRITE.

if (Characteristics & COFF::IMAGE_SCN_MEM_READ &&

~Characteristics & COFF::IMAGE_SCN_MEM_WRITE &&

~Characteristics & COFF::IMAGE_SCN_MEM_EXECUTE) // Read only, not execute.

Result = SymbolRef::ST_Data;

I’m not very familiar with COFF; what ramifications would this change have?

Ted

COFF sections also contain stuff like IMAGE_SCN_CNT_CODE and IMAGE_SCN_CNT_INITIALIZED_DATA. Are either of those set on the section in question?

It may make more sense to try to key off of those section characteristics as well.

The section is .text, so I assume it’s got IMAGE_SCM_CNT_CODE set. The problem is the symbol is marked (read & execute), but the test is (read & !write), so the symbol gets marked as ST_Data instead of ST_Other.

I believe this is fixed in r220952.

Hi David,

I apologize for taking this long to reply. I’ve been on vacation.

We’ve tested r220952, and it fixes the issue we saw with COFF symbol types. Thanks!

Ted