[PATCH] D20789: Consecutive memory access in Loop Vectorizer - fixed and simplified

[Hal Finkel via llvm-commits]

hfinkel added a comment.

In http://reviews.llvm.org/D20789#449385, @Ayal wrote:

> (1) Indeed *p++ should be treated just as consecutively as p[i++] by all means.
>
> (2) Could we simply use getPtrStride(), like so:
>
>   int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides);
>   if (Stride == 1 || Stride == -1)
>     return Stride;
>   return 0;
>
> making it work with specialization for unit-stride (vectorizing the load in version-mem-access.ll)?
>
> (3) If you test for
>
>   *to++ == *from--
>
> you'll also catch reverse consecutive, and won't compete with -loop-idiom's folding to memcpy; albeit cost will be higher.


You can also do `*to++ = 1 + *from++;` which will avoid the memcpy idiom issue.


Repository:
  rL LLVM

http://reviews.llvm.org/D20789