On Thu, Jan 7, 2021 at 10:14 AM Uros Bizjak <ubiz...@gmail.com> wrote:
>
> On Thu, Jan 7, 2021 at 9:56 AM Jakub Jelinek <ja...@redhat.com> wrote:
> >
> > Hi!
> >
> > The BLSI instruction sets SF and ZF based on the result and clears OF.
> > CF is set to something unrelated.
> >
> > The following patch optimizes BLSI followed by comparison, so we don't need
> > to emit a TEST insn in between.
> >
> > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> >
> > 2021-01-07 Jakub Jelinek <ja...@redhat.com>
> >
> > PR target/98567
> > * config/i386/i386.md (*bmi_blsi_<mode>_cmp): New define_insn.
> >
> > * gcc.target/i386/pr98567-1.c: New test.
> > * gcc.target/i386/pr98567-2.c: New test.
> >
> > --- gcc/config/i386/i386.md.jj 2021-01-04 10:25:45.072163178 +0100
> > +++ gcc/config/i386/i386.md 2021-01-06 17:49:13.251966127 +0100
> > @@ -14568,6 +14568,21 @@ (define_insn "*bmi_blsi_<mode>"
> > (set_attr "btver2_decode" "double")
> > (set_attr "mode" "<MODE>")])
> >
> > +(define_insn "*bmi_blsi_<mode>_cmp"
> > + [(set (reg FLAGS_REG)
> > + (compare
> > + (and:SWI48
> > + (neg:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "rm"))
> > + (match_dup 1))
> > + (const_int 0)))
> > + (set (match_operand:SWI48 0 "register_operand" "=r")
> > + (and:SWI48 (neg:SWI48 (match_dup 1)) (match_dup 1)))]
> > + "TARGET_BMI && ix86_match_ccmode (insn, CCNOmode)"
> > + "blsi\t{%1, %0|%0, %1}"
> > + [(set_attr "type" "bitmanip")
> > + (set_attr "btver2_decode" "double")
> > + (set_attr "mode" "<MODE>")])
>
> I wonder if we should also add _cc variant where scratch is used:
Er, _ccno in this particular case.
> (define_insn "*bmi_blsi_<mode>_cc"
> [(set (reg FLAGS_REG)
> (compare
> (and:SWI48
> (neg:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "rm"))
> (match_dup 1))
> (const_int 0)))
> (clobber (match_scratch:SWI48 0 "=r"))]
> "TARGET_BMI && ix86_match_ccmode (insn, CCNOmode)"
> "blsi\t{%1, %0|%0, %1}"
> [(set_attr "type" "bitmanip")
> (set_attr "btver2_decode" "double")
> (set_attr "mode" "<MODE>")])
>
> The output is unused in the testcases, so there may be no difference
> in the generated code, but it looks to me that additional pattern
> gives the compiler more freedom. Also note, that all other CC setting
> insns come in three variants.
If I'd have to choose between _cmp and _ccno variants, I'd prefer
__ccno one (with the scratch), since output and flags are seldom used
together (and when they do, a new pack of optimization problems
arises, e.g. recently mentioned global RTX CSE issue).
Uros.
