[llvm-commits] [LLVM, llc] TypeLegalization, DAGCombining, vectors loading: Obvious fixes.
Duncan Sands
baldrick at free.fr
Thu Dec 15 00:41:50 PST 2011
Hi Stepan,
> For 'load' I fixed the next:
>
> Next string will always wrong for non round types. For 2 x i7 it returns 1 byte!
> unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
>
> I just replaced it with this one:
>
> unsigned Stride = SrcVT.getScalarType().getStoreSize();
>
> This code just calculates proper size in bytes.
I think it is pointless trying to fix things for <2 x i7> and friends until it
is decided how they should be represented in memory.
Ciao, Duncan.
>
> -Stepan
>
> 15.12.2011, 04:48, "Dan Gohman"<gohman at apple.com>:
>> On Dec 14, 2011, at 12:57 PM, Stepan Dyatkovskiy wrote:
>>
>>> Hi all. Please find some obvious fixes for<n x i1> ..<n x i7> vectors.
>>> I fixed 'load' expansion, since there was improper stride calculation.
>>> I also fixed some assertion for trunc store. Assertion dropped for trunc stores with non simple MemoryVT. MemoryVT<n x i1> will never became simple in ToT. I inserted the store expansion in this case.
>>
>> At least the load part of this patch is wrong. PR1784 specifically says that
>> vector elements are to be bit-packed.
>>
>> Dan
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
More information about the llvm-commits
mailing list