Re: [RFC PATCH, i386]: Enable post-reload compare elimination pass

[Jakub Jelinek <jakub@redhat.com>]

On Fri, May 12, 2017 at 05:42:59AM -0400, Hans-Peter Nilsson wrote:
> (To-list pruned, my correction doesn't need attention.)
> 
> On Thu, 11 May 2017, Hans-Peter Nilsson wrote:
> > On Wed, 10 May 2017, Jakub Jelinek wrote:
> >
> > > On Wed, May 10, 2017 at 09:57:56PM +0200, Uros Bizjak wrote:
> > > > BTW: This patch now catches 417 cases (instead of 200+) in linux
> > > > build, including e.g.:
> > > >
> > > > (parallel [
> > > >         (set (reg:CCZ 17 flags)
> > > >             (compare:CCZ (lshiftrt:SI (reg:SI 4 si [orig:93 _10 ] [93])
> > > >                     (const_int 1 [0x1]))
> > > >                 (const_int 0 [0])))
> > > >         (set (reg:DI 4 si)
> > > >             (zero_extend:DI (lshiftrt:SI (reg:SI 4 si [orig:93 _10 ] [93])
> > > >                     (const_int 1 [0x1]))))
> > > >     ])
> 
> > Anyway, people seem to drift towards the ccreg-last variant
> 
> JFTR, I miswrote that; I meant "towards the variant with
> ccreg-first" as in Uros' example kept above and as opposed to my
> example.

If I skim the current primary/secondary targets, then i386/x86_64, rs6000,
aarch64, arm, sparc, s390* all use the ccreg-first order, in mips I couldn't
find either of the orders.
Looking at other targets, in alpha, avr, bfin, c6x, cr16, cris, fr30, ft32,
h8300, ia64, iq2000, lm32, m32c, m32r, m68k, mcore, microblaze, mmix, moxie,
nds32, nios2, nvptx, pa, pdp11, riscv, rl78, sh, spu, stormy16, tilegx,
tilepro, v850, vax, xtensa I can't find any order,
arc, epiphany are ccreg-first,
frv is some weird mixture of ccreg-first (e.g. *combo_intop_compare2) and
ccreg-last (e.g. adddi3_lower),
mn10300, rx and visium are ccreg-last that we want to switch over now to
ccreg-first.
It was all from quickly skimming (mostly) the main config/<cpu>/<cpu>.md
looking for \(compare and/or \(reg.*CC and looking if there are patterns
that have the same arithmetics inside a compare or something similar vs.
the operation repeated on a set, so it is possible I've missed something.

But if the above is roughly true, then it is of course preferable to change
3 less used targets than 6 primary/secondary ones + 2 further ones.

	Jakub