[llvm-commits] [llvm] r160589 - in /llvm/trunk: include/llvm/Transforms/Utils/Local.h lib/Analysis/MemoryBuiltins.cpp test/Instrumentation/BoundsChecking/simple.ll

[Duncan Sands]

Hi Nuno,

>>> baby steps toward fixing some problems with inbound GEPs that
>>> overflow, as discussed 2 months ago or so.
>>> Make sure we do not emit index computations with NSW flags so that
>>> we dont get an undef value if the GEP overflows
>>
>> does this really help?  I mean, the optimizers may have exploited
>> nsw, undef and
>> inbounds before this point to produce really weird bitcode, and may
>> exploit them
>> afterwards, eg because after inlining they see that your computation
>> is fed by
>> undef.  I'm worried that this appears to fix something, but doesn't
>> really.  I
>> don't think this kind of issue can be fixed without some kind of compiler
>> barrier like the llvm.identity intrinsic I mentioned a while ago.
>
> It surely helps.
> When we reach this point, the GEP may be undef (or become undef
> later), yes, but in that case the load/store was probably removed
> already. Theoretically this may not be truth, but it happens in
> practice.

it depends on what guarantees your pass is providing.  If it is OK that it
doesn't catch all out-of-bounds accesses, just most of them, then this patch
is fine.  If that is not OK then this patch doesn't make any sense to me.

Ciao, Duncan.