[PATCH] [InstCombine] Don't fold bitcast into store if it would need addrspacecast

[Richard Osborne]

On 19/03/14 17:40, Matt Arsenault wrote:
>
> Index: lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
> ===================================================================
> --- lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
> +++ lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
> @@ -517,27 +517,35 @@
>         IC.getDataLayout()->getTypeSizeInBits(DestPTy))
>       return 0;
>   
> +  // Don't introduce casts between pointers to different address spaces. We
> +  // can't use the addrspacecast instruction in this case since, depending on
> +  // the target, addrspacecast may not be a no-op cast.
> +  if (SrcPTy->isPointerTy() && DestPTy->isPointerTy() &&
> +      cast<PointerType>(SrcPTy)->getAddressSpace() !=
> +      cast<PointerType>(DestPTy)->getAddressSpace())
> +    return 0;
> +
>
> Right above here is a check leftover from when bitcast was allowed if 
> the pointers were the same size. You should update that and replace it 
> with this.

Are you talking about this code?

       // If the pointers point into different address spaces or if they
    point to
       // values with different sizes, we can't do the transformation.
       if (!IC.getDataLayout() ||
           SrcTy->getAddressSpace() !=
    cast<PointerType>(CI->getType())->getAddressSpace() ||
           IC.getDataLayout()->getTypeSizeInBits(SrcPTy) !=
           IC.getDataLayout()->getTypeSizeInBits(DestPTy))
         return 0;

I think the address space check here is still required. It might be 
cleaner to bailout earlier on if the cast instruction is an address 
space cast instead of a bitcast. Is that what you had in mind?

-- 
Richard Osborne | XMOS
http://www.xmos.com

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140319/e2d663ee/attachment.html>