[llvm-commits] [llvm] r158937 - in /llvm/trunk: include/llvm/Analysis/MemoryBuiltins.h lib/Analysis/MemoryBuiltins.cpp lib/Transforms/InstCombine/InstCombineCalls.cpp lib/Transforms/InstCombine/InstructionCombining.cpp test/Transforms/InstCombine

[Duncan Sands]

Hi Nuno,

>> Note that instcombine has "setPreservesCFG" set, which means that it really
>> shouldn't be touching any terminator instructions. It currently does to
>> swap things like "br i1 (xor b, true), label %l1, label %l2" and arguably
>> that doesn't really change the CFG (though I think that in an ideal world
>> instcombine should not be doing that either). This patch certainly does
>> though, and instcombine shouldn't be doing it. You could put it in
>> SimplifyCFG maybe?
>
> Right, just noticed that setPreservesCFG bit as well..
> First I was thinking to create an artificial edge with 'br i1 true,
> normalDest, unwindDest', but LLVM's verifier rejects that.
> This trick is actually used in another place in instcombine, which
> means that the following code crashes instcombine:
>     %call = invoke noalias i8* undef()
>             to label %invoke.cont unwind label %lpad
>
> So now we have 2 places in instcombine that are trying to remove
> invoke instructions..
> A simple solution could be to replace 'invoke @malloc' with 'invoke
> undef', and then move the logic to remove these invokes to
> SimplifyCFG.  What do you think?

if you are changing an invoke into a call, aren't you in essence setting
nounwind on @malloc?  You could just do that and then the invoke will be
dissolved into a call automatically by some other pass.  However, if
that is what you are doing (making malloc as nounwind), shouldn't it be done
in SimplifyLibCalls?

Ciao, Duncan.