https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53513
--- Comment #22 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #21)
> (In reply to Oleg Endo from comment #17)
> >
> > In the 'addsf3_i' pattern, I've tried replacing the
> >
> > (use (match_operand:PSI 3 "fpscr_operand" "c"))
> >
> > with
> >
> > (set (match_operand:PSI 3 "fpscr_operand" "=&c")
> > (unspec:PSI [(match_dup 3)] UNSPEC_FPSCR_SET))]
> >
> >
> > I haven't checked all the other side effects it could have, but at least the
> > FMA combine patterns still seem work after that change.
>
> With those changes above applied to all the other FP insns, the FMA tests in
> gcc.target/sh fail.
As well as some of the FSCA tests. E.g. 'return sinf (x) + cosf (x)' would
expand into two sincos patterns and normally combine would fold them into one.
However, it doesn't understand the unspec:PSI stuff and tries to nest it such
as:
Failed to match this instruction:
(parallel [
(set (reg:SI 174)
(fix:SI (reg:SF 167)))
(set (reg/v:PSI 151 )
(unspec:PSI [
(unspec:PSI [
(reg/v:PSI 151 )
] UNSPEC_FPSCR)
] UNSPEC_FPSCR))
])
I've tried adding a predicate that recursively matches those nested unspec:PSI
things, which makes combine happy, but then reload would bail out for some
reason. Moreover, fp insns that do this
(set (match_operand:PSI 3 "fpscr_operand" "=&c")
(unspec:PSI [(match_dup 3)] UNSPEC_FPSCR_SET))]
won't get eliminated even if their results are unused/dead.
I'm trying to come up with some alternative approach.
