Re: ifcvt fix

[Franz Sirl <Franz dot Sirl-kernel at lauterbach dot com>]

On Saturday 30 March 2002 13:05, Jan Hubicka wrote:
> > On Thursday 28 March 2002 20:07, Richard Henderson wrote:
> > > On Thu, Mar 28, 2002 at 07:21:52PM +0100, Jan Hubicka wrote:
> > > > 	* ifcvt.c (if_convert): Clear aux_for_blocks early enought.
> > >
> > > Ok everywhere.
> >
> > Jan, please don't apply it on the 3_1-branch without further testing, I
> > get a _lot_ of testsuite regressions (it doesn't even bootstrap) on
> > powerpc-linux-gnu with it, for example gcc.c-torture/compile/900116-1.c:
> >
> > (gdb) r -fpreprocessed 900116-1.i -quiet -dumpbase 900116-1.c -O1
> > -version -o 900116-1.s
> > Starting program: /cvsx/rawhide/BUILD/obj-gcc31-ppc/gcc/./cc1
> > -fpreprocessed 900116-1.i -quiet -dumpbase 900116-1.c -O1 -version -o
> > 900116-1.s Breakpoint 2 at 0xfeec090
> > Breakpoint 3 at 0xfeead08
> > GNU CPP version 3.1 20020329 (prerelease) (cpplib) (PowerPC GNU/Linux)
> > GNU C version 3.1 20020329 (prerelease) (ppc-linux)
> >         compiled by GNU C version 3.1 20020329 (prerelease).
> >
> > Program received signal SIGSEGV, Segmentation fault.
> > calculate_global_regs_live (blocks_in=0x41, blocks_out=0x10429b80,
> > flags=16) at /cvsx/rawhide/BUILD/gcc-3.1/gcc/flow.c:1118
> > 1118          bb->aux = NULL;
> > (gdb) l
> > 1113          edge e;
> > 1114
> > 1115          bb = *qhead++;
> > 1116          if (qhead == qend)
> > 1117            qhead = queue;
> > 1118          bb->aux = NULL;
> > 1119
> > 1120          /* Begin by propagating live_at_start from the successor
> > blocks.  */
> > 1121          CLEAR_REG_SET (new_live_at_end);
> > 1122
> > (gdb) bt
>
> Uhm, perhaps it is because the 3.1 version of patch needs to be somewhat
> different. I am just going to test:
>
> Index: ifcvt.c
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/ifcvt.c,v
> retrieving revision 1.78.2.1
> diff -c -3 -p -r1.78.2.1 ifcvt.c
> *** ifcvt.c	2002/03/18 00:18:41	1.78.2.1
> --- ifcvt.c	2002/03/30 12:04:48
> *************** if_convert (x_life_data_ok)
> *** 2756,2761 ****
> --- 2756,2762 ----
>   	if (UPDATE_LIFE (BASIC_BLOCK (block_num)))
>   	  SET_BIT (update_life_blocks, block_num);
>
> +       clear_aux_for_blocks ();
>         count_or_remove_death_notes (update_life_blocks, 1);
>         /* ??? See about adding a mode that verifies that the initial
>   	set of blocks don't let registers come live.  */
> *************** if_convert (x_life_data_ok)
> *** 2765,2771 ****
>
>         sbitmap_free (update_life_blocks);
>       }
> !   clear_aux_for_blocks ();
>
>     /* Write the final stats.  */
>     if (rtl_dump_file && num_possible_if_blocks > 0)
> --- 2766,2773 ----
>
>         sbitmap_free (update_life_blocks);
>       }
> !   else
> !     clear_aux_for_blocks ();
>
>     /* Write the final stats.  */
>     if (rtl_dump_file && num_possible_if_blocks > 0)

Yes, this patch works better on a quick testsuite check, I will run a full 
bootstrap tomorrow.

Thanks,
Franz.