[Bug middle-end/85090] [8 Regression] wrong code with -O2 -fno-tree-dominator-opts -mavx512f -fira-algorithm=priority

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85090

--- Comment #3 from ktkachov at gcc dot gnu.org ---
Hmm, I don't have access to AVX512F hardware so I can't reproduce the runtime
failure.
The vector simplifications that my patch introduces look correct to me from
looking at the dumps.

I'm not very familiar with i386.md but the *movdi_internal pattern that
produces the vmovq that zeroes out the top of the z register doesn't seem to
represent that in RTL.

So GCC ends up loading a DImode register xmm20 but then stores it as a
V32HImode register, the zero-extending effects of the DImode load are not
represented at RTL.

Any ideas?