[LLVMdev] how to correctly erase a bunch of instructions

[Daniel Berlin]

On Sat, Jun 20, 2015 at 8:03 AM, Frank Winter <fwinter at jlab.org> wrote:
> I have a basic block BB and a SetVector<Value*> for_erasure. The set vector
> contains the instructions which should be erased from BB but they are in no
> particular order. Thus something like this
>
> for ( Value* v: for_erasure ) {
>     if (Instruction *Inst = dyn_cast<Instruction>(v)) {
>       Inst->dropAllReferences();
>       Inst->eraseFromParent();
>     }
>   }
>
> results in
>
> While deleting: float* %
> Use still stuck around after Def is destroyed: <badref> = getelementptr
>

You need to replace all the uses with something. It's not enough to
drop all references, dropping all refs just makes badrefs, as you've
seen :)

If the remaining uses are in, say, unreachable blocks, or whatever,
the right thing to do is to replace the remaining uses with undef.

> I was thinking of reverse iterating through the basic block and testing each
> instruction whether it is contained in the set vector. I have however no
> success with it (segfaults). I think it's because erasing the instruction
> invalidates the (reverse) iterator.


I fixed this in trunk. The eraseFromParent()'s now return an
instruction you can use to create an iterator.
IE, the following now works:

  for (auto I = BB.rbegin(), E = BB.rend(); I
!= E;  ++I) {

    Instruction& inst = *I;
     I = inst.eraseFromParent();
}


> Can I somehow declare instructions as dead and then happily erase them (so
> that the above error's not happening)?
>

The above error is simply because you haven't replaced all the uses.
So replace them :)

> What's the standard way of erasing instructions in the given case? Some
> pseudo-codewould be appreciated.


The following should fix the above, if you insert it before the eraseFromParent

if (!Inst->use_empty())
Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));


Obviously, you should examine why it still has uses hanging around,
and whether you care about them, before you apply this :)
>
> Thanks
> Frank
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev