Re: [tree-ssa] live analysis on local static functions

[Jan Hubicka <jh at suse dot cz>]

> 
> 
> In message <20031017212058.GE26035@kam.mff.cuni.cz>, Jan Hubicka writes:
>  >> In message <20031017190337.GF15904@atrey.karlin.mff.cuni.cz>, Jan Hubicka w
>  >rite
>  >> s:
>  >>  >> On Fri, Oct 17, 2003 at 06:57:47PM +0200, Jan Hubicka wrote:
>  >>  >> > Also can I ask about the status of your changes in tree-ssa?  (it loo
>  >ks
>  >>  >> > like they are responsible for the new misscompilations I am seeing wi
>  >th
>  >>  >> > unit-at-a-time enabled.  I will look into these in case you don't hav
>  >e
>  >>  >> > something ready already)
>  >>  >> 
>  >>  >> It's looking generally good.  I'm sure there are bugs, but they're
>  >>  >> not overwhelming.
>  >>  >I actually was referring to Jeff's changes from this week,
>  >> Which specifically? I need specifics to be able to address any problems.
>  >
>  >Except for the uninitialized variables, I am seeing misscopmilation of
>  >read_rtx with unit-at-a-time enabled.  This is really funny one.
>  >read_rtx does call read_name that gets inlined as called one.
> You're missing my point.  I need testcases!   This is especially true
> since none of my changes this month should be effecting code generation!

If I had one, I would send it.
The failure I am looking into is reproducible only using stage2 compiler
so it is in reality misscompilation of reload when unit-at-a-time is
enabled as in attached patch.
So far I tracked it to fact that find_reg choose ebp as best reg that is
obivously wrong.  find_reg itself gets inlined into find_regs that get
inlined into reload() that gets gigantic and debug info completelly
messed up so I am not really sure what is going on here.  I am still
investigating.

I am bootrapping i686-linux.  I will try to produce testcase for the
missed optimization in read_rtx once I figure out what is happening here
(in fact this week changes apparently just made the bug to happen
earlier since last week I was also hitting reload misscomplation just
the symptomps were even more funny)

Honza

Index: opts.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/opts.c,v
retrieving revision 1.31.2.12
diff -c -3 -p -r1.31.2.12 opts.c
*** opts.c	17 Oct 2003 06:43:58 -0000	1.31.2.12
--- opts.c	17 Oct 2003 21:58:44 -0000
*************** decode_options (unsigned int argc, const
*** 562,568 ****
        flag_delete_null_pointer_checks = 1;
        flag_reorder_blocks = 1;
        flag_reorder_functions = 1;
!       /* flag_unit_at_a_time = 1; */
      }
  
    if (optimize >= 3)
--- 562,568 ----
        flag_delete_null_pointer_checks = 1;
        flag_reorder_blocks = 1;
        flag_reorder_functions = 1;
!       flag_unit_at_a_time = 1;
      }
  
    if (optimize >= 3)