[PATCH] x86: Correct EVEX to 128-bit EVEX optimization

[H.J. Lu]

On Mon, Mar 18, 2019 at 7:31 PM Jan Beulich <JBeulich@suse.com> wrote:
>
> >>> On 16.03.19 at 23:48, <hjl.tools@gmail.com> wrote:
> > We can optimize 512-bit EVEX to 128-bit EVEX encoding for upper 16
> > vector registers only when AVX512VL is enabled.  We can't optimize
> > EVEX to 128-bit VEX encoding when AVX isn't enabled.
>
> I don't understand the last sentence: AVX is a prereq to anything
> that's EVEX-encoded, at least as of now. "-march=+noavx" should
> really result in all of AVX512 to also get disabled.
>
> > --- a/gas/config/tc-i386.c
> > +++ b/gas/config/tc-i386.c
> > @@ -3975,10 +3975,13 @@ optimize_encoding (void)
> >                  && !i.rounding
> >                  && is_evex_encoding (&i.tm)
> >                  && (i.vec_encoding != vex_encoding_evex
> > +                    || cpu_arch_flags.bitfield.cpuavx
> > +                    || cpu_arch_isa_flags.bitfield.cpuavx
> > +                    || cpu_arch_flags.bitfield.cpuavx512vl
> > +                    || cpu_arch_isa_flags.bitfield.cpuavx512vl
>
> cpu_arch_flags starts out with (almost) all bits set. It was for that
> reason that ...
>
> >                      || i.tm.cpu_flags.bitfield.cpuavx512vl
> >                      || (i.tm.operand_types[2].bitfield.zmmword
> > -                        && i.types[2].bitfield.ymmword)
> > -                    || cpu_arch_isa_flags.bitfield.cpuavx512vl)))
> > +                        && i.types[2].bitfield.ymmword))))
>
> ... originally only cpu_arch_isa_flags got checked here.
>

Here is the patch.

-- 
H.J.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-x86-Correct-EVEX-to-128-bit-EVEX-optimization.patch
Type: text/x-patch
Size: 20925 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/binutils/attachments/20190319/fc8da79c/attachment.bin>