diff mbox

Fix twolf -funroll-loops -O3 miscompilation (a semi-latent web.c bug)

Message ID CABu31nPUNnpNFUv=g-HSSsy7eU2S0DrSLywoCSF7NBCNSWC_1A@mail.gmail.com
State New
Headers show

Commit Message

Steven Bosscher Nov. 27, 2012, 1:04 p.m. UTC 
On Tue, Nov 27, 2012 at 1:28 PM, Steven Bosscher wrote:
> Dominique, could you give this a try and see if it helps?
> (But as I said up-thread: I'm not sure this is a proper fix, or just
> another band-aid...)

And more band-aid, but this time I'm not even sure where things go
wrong. In any case, we end up with a REG_EQUAL note referencing a
SUBREG of a dead register. This leaked because df_remove_dead_eq_notes
uses loc_mentioned_in_p not on the note but on the first operand of
the note.

Comments

Dominique d'Humières Nov. 27, 2012, 2:25 p.m. UTC | #1 
> And more band-aid, ...

The gcc_assert triggers at bootstrap when compiling gcc/ada/ali.adb:

+===========================GNAT BUG DETECTED==============================+
| 4.8.0 20121127 (experimental) [trunk revision 193848p10] (x86_64-apple-darwin10.8.0) GCC error:|
| in df_remove_dead_eq_notes, at df-problems.c:2917                        |
| Error detected around ../../work/gcc/ada/ali.adb:2682:8                  |
| Please submit a bug report; see http://gcc.gnu.org/bugs.html.            |
| Use a subject line meaningful to you and us to track the bug.            |
| Include the entire contents of this bug box in the report.               |
| Include the exact gcc or gnatmake command that you entered.              |
| Also include sources listed below in gnatchop format                     |
| (concatenated together with no headers between files).                   |
+==========================================================================+

Dominique
Steven Bosscher Nov. 27, 2012, 2:48 p.m. UTC | #2 
On Tue, Nov 27, 2012 at 3:25 PM, Dominique Dhumieres <dominiq@lps.ens.fr> wrote:
>> And more band-aid, ...
>
> The gcc_assert triggers at bootstrap when compiling gcc/ada/ali.adb:
>
> +===========================GNAT BUG DETECTED==============================+
> | 4.8.0 20121127 (experimental) [trunk revision 193848p10] (x86_64-apple-darwin10.8.0) GCC error:|
> | in df_remove_dead_eq_notes, at df-problems.c:2917                        |
> | Error detected around ../../work/gcc/ada/ali.adb:2682:8                  |
> | Please submit a bug report; see http://gcc.gnu.org/bugs.html.            |
> | Use a subject line meaningful to you and us to track the bug.            |
> | Include the entire contents of this bug box in the report.               |
> | Include the exact gcc or gnatmake command that you entered.              |
> | Also include sources listed below in gnatchop format                     |
> | (concatenated together with no headers between files).                   |
> +==========================================================================+

Yes, I found that one already, too.

And, oh joy, we have pseudos in REG_EQUAL notes after LRA! (Probably
also after reload, btw.).  In the ICE I got, a pseudo's live range got
split and an inheritance move is injected, but REG_EQUAL notes were
not updated or removed. Finding and removing the notes is hard in IRA
and LRA because they don't use the DF caches.

Ciao!
Steven
diff mbox

Patch

Index: df-problems.c
===================================================================
--- df-problems.c       (revision 193394)
+++ df-problems.c       (working copy)
@@ -2907,9 +2907,10 @@  df_remove_dead_eq_notes (rtx insn, bitma
                if (DF_REF_REGNO (use) > FIRST_PSEUDO_REGISTER
                    && DF_REF_LOC (use)
                    && (DF_REF_FLAGS (use) & DF_REF_IN_NOTE)
-                   && ! bitmap_bit_p (live, DF_REF_REGNO (use))
-                   && loc_mentioned_in_p (DF_REF_LOC (use), XEXP (link, 0)))
+                   && ! bitmap_bit_p (live, DF_REF_REGNO (use)))
                  {
+                   /* Make sure that DF_SCAN is up-to-date.  */
+                   gcc_assert (loc_mentioned_in_p (DF_REF_LOC (use), link));
                    deleted = true;
                    break;
                  }