On Mon, May 07, 2012 at 09:36:38AM +0200, Uros Bizjak wrote:
> On Mon, May 7, 2012 at 7:43 AM, Jakub Jelinek <ja...@redhat.com> wrote:
>
> >> Index: i386.md
> >> ===================================================================
> >> --- i386.md (revision 187217)
> >> +++ i386.md (working copy)
> >> @@ -12112,9 +12112,22 @@
> >> (set (match_operand:SWI48 0 "register_operand" "=r")
> >> (ctz:SWI48 (match_dup 1)))]
> >> ""
> >> - "bsf{<imodesuffix>}\t{%1, %0|%0, %1}"
> >> +{
> >> + if (optimize_function_for_size_p (cfun))
> >> + return "bsf{<imodesuffix>}\t{%1, %0|%0, %1}";
> >> + else if (TARGET_BMI)
> >> + return "tzcnt{<imodesuffix>}\t{%1, %0|%0, %1}";
> >> + else
> >> + /* tzcnt expands to rep;bsf and we can use it even if !TARGET_BMI. */
> >> + return "rep; bsf{<imodesuffix>}\t{%1, %0|%0, %1}";
> >> +}
> >
> > Shouldn't that be done only for generic tuning? If somebody uses
> > -mtune=native, then emitting rep; bsf is overkill, the code is intended
> > to be run on a CPU without (or with TARGET_BMI with) tzcnt insn support.
>
> Yes, this is a good idea.
>
> Something like attached patch?
Yeah, but probably with the TARGET_BMI and optimize_function_for_size_p
lines swapped.
Because i386.h defines
CTZ_DEFINED_VALUE_AT_ZERO to 1 for TARGET_BMI, therefore it isn't
undefined for TARGET_BMI, so even for -Os you can't just use bsf.
Jakub
