Re: [patch] gcse.c: Fix the header ordering.

[Graham Stott <graham dot stott at btinternet dot com>]

Kazu Hirata wrote:
> Hi Graham,
>
>
> > >      rtx dest = SET_DEST (sset);
> > >      if ((REG_P (dest) || CC0_P (dest))
> > > 	  && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
> > >
> > > Since rtl.h is included before insn-config.h, which contains HAVE_cc0,
> > > CC0_P is defined to be 0, so the above "if" statement doesn't let us
> > > call cprop_jump at all.
> > Actually it will if dest is a REG rtx :-)
>
> Right. :-)
>
>
> > > A quick grep shows that combine.c also uses CC0_P but includes
> > > insn-config.h after rtl.h.  I haven't seen what happens if I correct
> > > the order.
> > >
> > > I am going to run the testsuite with the patch below, but the problem
> > > seems to be a little deeper.  Should rtl.h include insn-config.h?
> > > Should rtl.h give an error if insn-config.h is not included yet?
> > I think rtl.h should include insn-config.h
It would also need be guarded by ifndef GENERATOR_FILE and also added to
RTL_BASE_H list in Makefile.in

> > Note in the not to distant past such nested included were discouraged but
> > it would appear that that position has changed because grepping for
> > "#include" in *.h files within the gcc directory gets quite a few hits.
>
> Discouraged for simplicity?  Or for a little more specific reason?
I'm not sure on the specifics I think it was just convention.

> I just put insn-config.h before rtl.h in combine.c, the generated code
> seems to get slightly better.  (That is, CC0_P in combine.c is now
> triggered.)
>
> Anyway, grepping in *.c arena is scary.  A lot of files include
> insn-config.h after rtl.h...
For the majority of files it currently makes no difference.

> Kazu Hirata
Graham