This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH i386 AVX512] [17/n] Split VI48_AVX512F into VI4_AVX512VL and VI248_AVX512, extend vcvtps2udq,vpbroadcastmb2d.
- From: Uros Bizjak <ubizjak at gmail dot com>
- To: Kirill Yukhin <kirill dot yukhin at gmail dot com>
- Cc: Jakub Jelinek <jakub at redhat dot com>, Richard Henderson <rth at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Mon, 18 Aug 2014 21:05:30 +0200
- Subject: Re: [PATCH i386 AVX512] [17/n] Split VI48_AVX512F into VI4_AVX512VL and VI248_AVX512, extend vcvtps2udq,vpbroadcastmb2d.
- Authentication-results: sourceware.org; auth=none
- References: <20140815114256 dot GC35144 at msticlxl57 dot ims dot intel dot com> <CAFULd4Z_rbyrQuLCXBmYUY1xmOYYTOQoURVW_gccr-ouDb_26Q at mail dot gmail dot com> <20140818182757 dot GA28898 at msticlxl57 dot ims dot intel dot com>
On Mon, Aug 18, 2014 at 8:30 PM, Kirill Yukhin <firstname.lastname@example.org> wrote:
>> > (define_insn "<avx2_avx512bw>_ashrv<mode><mask_name>"): New.
>> It looks to me that the macroization is somehow wrong for ashrv. I'd
>> split the mode iterator to:
>> > +(define_mode_iterator VI248_AVX512
>> > + [(V16SI "TARGET_AVX512F") (V8SI "TARGET_AVX2") (V4SI "TARGET_AVX2")
>> > + (V32HI "TARGET_AVX512BW")
>> > + (V16HI "TARGET_AVX512BW && TARGET_AVX512VL")
>> > + (V8HI "TARGET_AVX512BW && TARGET_AVX512VL")
>> > + (V8DI "TARGET_AVX512F") (V4DI "TARGET_AVX512VL") (V2DI "TARGET_AVX512VL")])
>> > +
>> V4SI, V8SI, V16SI (AVX512F), V8DI (AVX512F), V4DI (AVX512VL), V2DI
>> (AVX512VL), with AVX2 as the baseline
>> V32HI, V16HI (AVX512VL), V8HI (AVX512VL), with AVX512BW as the baseline
> Agreed. After doing `make mddump' I can see that I've enabled
> vpsavw for AVX2.
> I (hopefully) implemented the proposal. Thanks for review!
> Is it ok now?
Yes, this looks much better!
I hope that some simple logic behind AVX512VL and AVX512BW (etc)
combinations can be found. At least by using simple insn constraint
and simple mode constraints, one can find which modes are supported by
the pattern in a much easier way.
Please also use the same approach in your future patches (using simple
insn and mode constraints). We can always merge patterns later, if any
additional macroization possibilities would be found.
The patch is OK.
BTW: Taking these new findings into account, is it possible to split
V_AVX512VL mode iterator in the same way to avoid compound conditions
in the mode iterator?