On Tue, Sep 27, 2016 at 09:29:10PM +0200, Florian Weimer wrote:
> Not sure if I read this code correctly, but if we fall through from
> V32HImode, and we have TARGET_SSE2 set, we execute this code:
> 
>       tmp = "p<logic>";
>       ssesuffix = TARGET_AVX512VL ? "q" : "";
> 
> And not gcc_unreachable (), as is probably intended.

It really doesn't matter.
The instruction uses
(define_mode_iterator VI12_AVX_AVX512F
  [ (V64QI "TARGET_AVX512F") (V32QI "TARGET_AVX") V16QI
    (V32HI "TARGET_AVX512F") (V16HI "TARGET_AVX") V8HI])
iterator (and, after all, ix86_hard_regno_mode_ok should ensure the same),
which means V64QI/V32HI will only show up for TARGET_AVX512F, V32QI/V16HI
only for TARGET_AVX (which implies TARGET_SSE2), and the slightly
nonsensical
gcc_assert (TARGET_SSE2 || TARGET_AVX512VL);
before the switch (the  || TARGET_AVX512VL is pointless, because
TARGET_AVX512VL implies TARGET_SSE2 as well as TARGET_AVX2).
So, I'd go perhaps for (untested) following patch, first diff -up, followed
by diff -upb:

	Jakub