[Bug target/70021] [6 Regression] Test miscompiled with -O3 option for -march=core-avx2.

[ienkovich at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70021

--- Comment #7 from Ilya Enkovich <ienkovich at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #6)
> Created attachment 37834 [details]
> gcc6-pr70021-wip.patch
> 
> But that would mean we don't vectorize the loop.
> I see 2 cases here, one where we look up the def_stmt and look through it,
> that is the same case as in the other PR and we shouldn't do that, otherwise
> we reference something that is not the final SSA_NAME of a pattern sequence
> outside of that sequence.
> The other case is where we don't look through that or just avoid that, in
> that case we add a cast into the current pattern sequence, and refer to the
> final SSA_NAME of the other pattern sequence.  I really don't see why that
> shouldn't work right.  Except as the patch shows, it doesn't for some reason.

We didn't vectorize this loop before r230098 so it wouldn't be a regression. 
But surely be less restrictive is always good!

You'd probably want to check loop is vectorized in your patch then?

BTW this test shows another problem.  Conversion statement replaced with a
pattern is still used for vectorization factor computation causing VF=32 with
no vectors with more than 8 elements used.