This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][dataflow] Fix PR 32339, bad DF_REF_READ_WRITE marking


On 6/15/07, Seongbae Park (박성배, 朴成培) <seongbae.park@gmail.com> wrote:
On 6/14/07, Seongbae Park (박성배, 朴成培) <seongbae.park@gmail.com> wrote:
> "flags" was modified, leaving later calls to df_uses_record
> to mark uses as DF_REF_READ_WRITE.
> This lead to missing REG_DEAD note for such a use,
> causing different views of live ranges
> between forward-scanning (global, using REG_DEAD)
> and backward-scanning (REG_N_CALLS_CROSSED, which is computed by
> regstat_bb_compute_ri).
> This inconsistent view lead to a call-clobbered hard register
> to not have a save area, but later was attempted to be spilled
> because a pseudo register that doesn't cross a live range got allocated to it
> and the forward scanning in global thought that the pseudo reg
> lives across the call.
>
> Bootstrapped and regtested on ia64.
> OK fo mainline ?
>
> ChangeLog:
>
> 2007-06-14  Seongbae Park  <seongbae.park@gmail.com>
>
>         PR rtl-optimization/32339
>         * df-scan.c (df_uses_record): Don't modify flags but just add to it for
>         df_ref_record.
> --
> #pragma ident "Seongbae Park, compiler, http://seongbae.blogspot.com";
>
>
> Index: gcc/df-scan.c
> ===================================================================
> --- gcc/df-scan.c       (revision 125735)
> +++ gcc/df-scan.c       (working copy)
> @@ -2982,9 +2982,9 @@ df_uses_record (struct df_collection_rec
>      case PRE_MODIFY:
>      case POST_MODIFY:
>        /* Catch the def of the register being modified.  */
> -      flags |= DF_REF_READ_WRITE | DF_REF_PRE_POST_MODIFY;
>        df_ref_record (collection_rec, XEXP (x, 0), &XEXP (x, 0), bb, insn,
> -                    DF_REF_REG_DEF, flags);
> +                    DF_REF_REG_DEF,
> +                     flags | DF_REF_READ_WRITE | DF_REF_PRE_POST_MODIFY);
>
>        /* ... Fall through to handle uses ...  */

Andrew Pinski pointed out that I didn't add the testcase to our testsuite.
Here's the patch for adding the test.
--
#pragma ident "Seongbae Park, compiler, http://seongbae.blogspot.com";

ChangeLog:

2007-06-15 Seongbae Park <seongbae.park@gmail.com>

        PR rtl-optimization/32339
        * gcc.c-torture/compile/pr32339.c: New test.


Index: testsuite/gcc.c-torture/compile/pr32339.c =================================================================== --- testsuite/gcc.c-torture/compile/pr32339.c (revision 0) +++ testsuite/gcc.c-torture/compile/pr32339.c (revision 0) @@ -0,0 +1,17 @@ +/* We used to ICE in insert_save at caller-save.c, + due to missing REG_DEAD for register use in post-decrement on ia64. */ +struct city_dialog +{ + struct city *pcity; + char change_list_names[200][200]; + int change_list_ids[200]; +}; +change_callback (void) +{ + struct city_dialog *pdialog; + int n; + int i; + get_city_dialog_production_full (pdialog->change_list_names[n], + pdialog->pcity); + pdialog->change_list_ids[n++] = i; +}


*PING* I've updated the attribution in the changelog entry. No change in the test itself.

ChangeLog:

2007-06-15 Martin Michlmayr <tbm@cyrius.com>

      PR rtl-optimization/32339
      * gcc.c-torture/compile/pr32339.c: New test.
--
#pragma ident "Seongbae Park, compiler, http://seongbae.blogspot.com";

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]