[llvm-commits] [llvm] r69574 - /llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp

[Sanjiv Gupta]

Dan Gohman wrote:
> On Apr 20, 2009, at 9:16 AM, Sanjiv Gupta wrote:
>
>   
>> Dan Gohman wrote:
>>     
>>> Hello,
>>>
>>> Are you planning to resubmit r58505?  It was reverted in r58547 with
>>> a message indicating that it would be resubmitted.  That patch would
>>> appear to make this r69574 unnecessary.
>>>
>>> Dan
>>>
>>>
>>>       
>> Right. Even after that this change will still hold good just in case  
>> if
>> a pointer size doesn't make valid index.
>>     
>
> If Instcombine isn't using a valid index type for this optimization
> for some target, it really should be fixed. Silently falling back to
> not doing this may just be hiding problems.
>
>   
I agree.
>> Chris had suggested that 58505 alone is not enough as the bc encoding
>> uses 1 bit to encode i32/i64 for index types.
>> He had suggested that we should be able to write i16 type indexes and
>> should be able to get back them by the command seq below.
>>  $ llvm-as < test.ll | llvm-dis | llvm-as | llvm-dis
>>
>> When I tried this with 58505, it worked.
>> So the only additional thing is that we  probably need to do the  
>> changes
>> in the documentation.
>>
>> Let me know if this understanding is incorrect. Better you can get  
>> me a
>> test case.
>>     
>
> I'm not very familiar with the bitcode encoding, but at a quick
> glance it looks like a full type is stored with each operand.
> If this just needs a documentation update, please do that and
> re-apply 58505.
>
> Thanks,
>
> D
>   
I too had similar observation last time I debugged that area.
- Sanjiv
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>