On Mon, Oct 20, 2014 at 05:30:36PM +0400, Kirill Yukhin wrote:
> > Unfortunately this caused PR63600. The problem is that VI_AVX2
> > mode iterator includes V2DI and for AVX2 also V4DI, but for pre-ssse3
> > ix86_expand_sse2_abs doesn't handle V2DI (and can't easily, we don't have
> > PSRAQ instruction), for ssse3 there is no vpabsq instruction, and for
> > avx2 neither.
> > We can handle V2DI/V4DI only for TARGET_AVX512VL, and V8DI for
> > TARGET_AVX512F.
> > Thus, IMHO the mode iterator on at least
> > (define_insn "*abs<mode>2"
> > and on
> > (define_expand "abs<mode>2"
> > is wrong, should not include V2DI/V4DI unless TARGET_AVX512VL
> > (so new (or ressurrected, was that VI124_AVX2_48_AVX512F?)
> > specialized mode iterator?).
>
>
> This patch removes absq insn patterns for non-AVX-512 targets.
>
>
> gcc/
> * config/i386/sse.md (define_mode_iterator VI_AVX2): Restore to 128-,
> 256- bit integer modes only.
> (define_mode_iterator VI_AVX2_AVX512): New.
> (define_expand "neg<mode>2"): Use VI_AVX2_AVX512 mode iterator.
> (define_expand "<plusminus_insn><mode>3"): Ditto.
> (define_insn "*<plusminus_insn><mode>3"): Ditto.
> (define_expand "<sse2_avx2>_andnot<mode>3"): Ditto.
> (define_mode_iterator VI1248_AVX512VL_AVX512BW): New.
> (define_insn "abs<VI1248_AVX512VL_AVX512BW:mode>2"): Ditto.
>
> Bootstrap in progress. AVX-512 tests pass.
>
> Is it ok for trunk?
I'll certainly leave the review to Uros, whatever he prefers.
That said, I was expecting you'd keep VI_AVX2 as is (because from the patch
clearly that is what is used most commonly, the V?DI modes are for most
insns normal integral vector modes, VI* uses the same modes and VI_AVX2
used to be just like VI, just with TARGET_AVX conditions replaced with
TARGET_AVX2), and just add a new mode iterator for the two abs patterns
(*abs<mode>2 and abs<mode>2), it can be specialized mode iterator just
for the abs with ABS in names or something.
Jakub
