llvm dev - Jul 2019 - Replacing ptrtoint+inttoptr with addrspacecast

Given a ptrtoint followed by inttoptr:

%tmp = ptrtoint i16 addrspace(4)* %a to i32
%b = inttoptr i32 %tmp to i8*

InstCombine will normally collapse them to a bitcast (or no-op, if the src and
dst pointer types are the same).  But if the pointers are in different address
spaces, as shown above, it won't, due to the following code in
Instructions.cpp:

      // Cannot simplify if address spaces are different!
      if (SrcTy->getPointerAddressSpace() !=
DstTy->getPointerAddressSpace())
        return 0;

Is this correct?  Given that the instructions above are effectively casting the
pointer between two address spaces, can/should InstCombine replace the code
above with an addrspacecast?  Is there something in the LLVM language spec that
makes it illegal or undesirable?

Thanks,

David
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190715/586664c2/attachment.html>

> On Jul 15, 2019, at 14:34, Cashman, David via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> Given a ptrtoint followed by inttoptr:
>  
> %tmp = ptrtoint i16 addrspace(4)* %a to i32
> %b = inttoptr i32 %tmp to i8*
>  
> InstCombine will normally collapse them to a bitcast (or no-op, if the src
and dst pointer types are the same).  But if the pointers are in different
address spaces, as shown above, it won’t, due to the following code in
Instructions.cpp:
> 
>       // Cannot simplify if address spaces are different!
>       if (SrcTy->getPointerAddressSpace() !=
DstTy->getPointerAddressSpace())
>         return 0;
>  
> Is this correct?  Given that the instructions above are effectively casting
the pointer between two address spaces, can/should InstCombine replace the code
above with an addrspacecast?  Is there something in the LLVM language spec that
makes it illegal or undesirable?
>  
> Thanks,
> 
> David
> 
This would be illegal. This is something resembling a no-op addrspacecast, and
would only be valid if the target defined representations of both address spaces
is known identical. Addrspacecast can change the representation of the pointer
in complex ways which would not happen with the ptrtoint/inttoptr pair.

-Matt
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190715/865b7d33/attachment.html>

People have argued that eliminating a ptr -> int -> ptr cast sequence is 
wrong even if all pointers are in the same address space. Therefore the 
optimization currently performed by the optimizer is unsafe. See 
https://www.youtube.com/watch?v=CPokKMqAVdY

Additionally there might be additional reasons why replacing ptr -> int 
-> ptr cast sequence with addrspacecast is wrong. It depends on what the 
exact semantics of the addrspacecast is.

On 2019-07-15 20:34, Cashman, David via llvm-dev wrote:
> Given a ptrtoint followed by inttoptr:
> 
> %tmp = ptrtoint i16 addrspace(4)* %a to i32
> %b = inttoptr i32 %tmp to i8*
> 
> InstCombine will normally collapse them to a bitcast (or no-op, if the
> src and dst pointer types are the same).  But if the pointers are in
> different address spaces, as shown above, it won't, due to the
> following code in Instructions.cpp:
> 
>       // Cannot simplify if address spaces are different!
>       if (SrcTy->getPointerAddressSpace() != 
> DstTy->getPointerAddressSpace())
>         return 0;
> 
> Is this correct?  Given that the instructions above are effectively
> casting the pointer between two address spaces, can/should InstCombine
> replace the code above with an addrspacecast?  Is there something in
> the LLVM language spec that makes it illegal or undesirable?
> 
> Thanks,
> 
> David
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev