Various parts of the LLVM code generator assume that the address
argument of a dbg.declare is not a ptrtoint-of-alloca. ASan breaks
this assumption, and this results in local variables sometimes being
unavailable at -O0.
GlobalISel, SelectionDAG, and FastISel all do not appear to expect
dbg.declares to have a ptrtoint as an operand. This means that they do
not place entry block allocas in the usual side table reserved for local
variables available in the whole function scope. This isn't always a
problem, as LLVM can try to lower the dbg.declare to a DBG_VALUE, but
those DBG_VALUEs can get dropped for all the usual reasons DBG_VALUEs
get dropped. In the ObjC test case I'm looking at, the cause happens to
be that replaceDbgDeclare has hoisted dbg.declares into the entry
block, causing LiveDebugValues to "kill" the DBG_VALUEs because the
lexical dominance check fails.
To address this, I propose:
- Have ASan (always) pass an alloca to dbg.declares (this patch). This is a narrow bugfix for -O0 debugging.
- Make replaceDbgDeclare not move dbg.declares around. This should be a generic improvement for optimized debug info, as it would prevent the lexical dominance check in LiveDebugValues from killing as many variables.
This means reverting llvm/r227544, which fixed an assertion failure (llvm.org/PR22386) but no longer seems to be necessary. I was able to complete a stage2 build with the revert in place.