[Bug rtl-optimization/52203] New: ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[Bug rtl-optimization/52203] New: ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

             Bug #: 52203
           Summary: ICE: in reset_sched_cycles_in_current_ebb, at
                    sel-sched.c:7136 with -fsel-sched-pipelining
                    -fselective-scheduling2 and other custom flags
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zsojka@seznam.cz
              Host: x86_64-pc-linux-gnu
            Target: x86_64-pc-linux-gnu


Created attachment 26639
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26639
reduced testcase

After a long time, there comes one sel-sched ICE.

Compiler output:
$ gcc -O -fPIC -fno-dce -fgcse -fschedule-insns2 -fsel-sched-pipelining
-fselective-scheduling2 -ftree-vectorize -funroll-loops -fno-web
-mno-avx256-split-unaligned-store -mfma --param=max-pending-list-length=0
--param=selsched-insns-to-rename=70 testcase.c 
testcase.c: In function 'foo':
testcase.c:9:1: internal compiler error: in reset_sched_cycles_in_current_ebb,
at sel-sched.c:7136
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

(gdb) bt
#0  fancy_abort (file=0x129c408 "/mnt/svn/gcc-trunk/gcc/sel-sched.c",
line=7136, function=0x129dce0 "reset_sched_cycles_in_current_ebb")
    at /mnt/svn/gcc-trunk/gcc/diagnostic.c:898
#1  0x00000000009a919d in reset_sched_cycles_in_current_ebb () at
/mnt/svn/gcc-trunk/gcc/sel-sched.c:7136
#2  0x00000000009bea53 in sel_region_target_finish (reset_sched_cycles_p=true)
at /mnt/svn/gcc-trunk/gcc/sel-sched.c:7209
#3  sel_region_finish (reset_sched_cycles_p=<optimized out>) at
/mnt/svn/gcc-trunk/gcc/sel-sched.c:7266
#4  sel_sched_region (rgn=<optimized out>) at
/mnt/svn/gcc-trunk/gcc/sel-sched.c:7622
#5  0x00000000009bf23a in run_selective_scheduling () at
/mnt/svn/gcc-trunk/gcc/sel-sched.c:7693
#6  0x000000000098c076 in rest_of_handle_sched2 () at
/mnt/svn/gcc-trunk/gcc/sched-rgn.c:3515
#7  0x00000000008fbd85 in execute_one_pass (pass=0x17ca8e0) at
/mnt/svn/gcc-trunk/gcc/passes.c:2081
#8  0x00000000008fc145 in execute_pass_list (pass=0x17ca8e0) at
/mnt/svn/gcc-trunk/gcc/passes.c:2136
#9  0x00000000008fc157 in execute_pass_list (pass=0x17c9ee0) at
/mnt/svn/gcc-trunk/gcc/passes.c:2137
#10 0x00000000008fc157 in execute_pass_list (pass=0x17c9f40) at
/mnt/svn/gcc-trunk/gcc/passes.c:2137
#11 0x0000000000a5e3ee in tree_rest_of_compilation (fndecl=0x7ffff5bd2700) at
/mnt/svn/gcc-trunk/gcc/tree-optimize.c:422
#12 0x00000000006b006a in cgraph_expand_function (node=0x7ffff5a976c0) at
/mnt/svn/gcc-trunk/gcc/cgraphunit.c:1829
#13 0x00000000006b1f5c in cgraph_expand_all_functions () at
/mnt/svn/gcc-trunk/gcc/cgraphunit.c:1896
#14 cgraph_optimize () at /mnt/svn/gcc-trunk/gcc/cgraphunit.c:2210
#15 0x00000000006b25aa in cgraph_finalize_compilation_unit () at
/mnt/svn/gcc-trunk/gcc/cgraphunit.c:1338
#16 0x000000000058ecb0 in c_write_global_declarations () at
/mnt/svn/gcc-trunk/gcc/c-decl.c:10030
#17 0x00000000009f0b9c in compile_file () at
/mnt/svn/gcc-trunk/gcc/toplev.c:573
#18 do_compile () at /mnt/svn/gcc-trunk/gcc/toplev.c:1938
#19 toplev_main (argc=28, argv=0x7fffffffd658) at
/mnt/svn/gcc-trunk/gcc/toplev.c:2014
#20 0x00007ffff61cc09d in __libc_start_main () from /lib64/libc.so.6
#21 0x00000000005723b1 in _start ()

Tested revisions:
r184068 - crash
4.6 r180325 - OK

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-02-10
                 CC|                            |abel at gcc dot gnu.org
     Ever Confirmed|0                           |1

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

Andrey Belevantsev <abel at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |abel at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #1 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-02-11 12:52:38 UTC ---
Thanks, Zdenek and Steven, I'll look at this on Monday.  I bet this is caused
by yet another insn without a reservation.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #2 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-02-13 08:48:35 UTC ---
Well, in this case the insn is vcvtps2pd (from avx_cvtps2pd256 and
*avx_cvtps2pd256_2 patterns) but not having a reservation looks correct as the
manual says "decode type -- microcode".  Uros, is that right?  In this case,
the below fixes the test for me (and I was stupid enough not to do this in the
first place when this bug stream just started...  However, we did fix some real
bugs.)

iff --git a/gcc/sel-sched.c b/gcc/sel-sched.c
index 2af01ae..2829f60 100644
--- a/gcc/sel-sched.c
+++ b/gcc/sel-sched.c
@@ -4265,9 +4265,10 @@ invoke_aftermath_hooks (fence_t fence, rtx best_insn,
int issue_more)
   return issue_more;
 }

-/* Estimate the cost of issuing INSN on DFA state STATE.  */
+/* Estimate the cost of issuing INSN on DFA state STATE.  Write to PEMPTY
+   true when INSN does not change the processor state.  */
 static int
-estimate_insn_cost (rtx insn, state_t state)
+estimate_insn_cost (rtx insn, state_t state, bool *pempty)
 {
   static state_t temp = NULL;
   int cost;
@@ -4277,6 +4278,8 @@ estimate_insn_cost (rtx insn, state_t state)

   memcpy (temp, state, dfa_state_size);
   cost = state_transition (temp, insn);
+  if (pempty)
+    *pempty = (memcmp (temp, state, dfa_state_size) == 0);

   if (cost < 0)
     return 0;
@@ -4307,7 +4310,7 @@ get_expr_cost (expr_t expr, fence_t fence)
        return 0;
     }
   else
-    return estimate_insn_cost (insn, FENCE_STATE (fence));
+    return estimate_insn_cost (insn, FENCE_STATE (fence), NULL);
 }

 /* Find the best insn for scheduling, either via max_issue or just take
@@ -7020,7 +7023,7 @@ reset_sched_cycles_in_current_ebb (void)
     {
       int cost, haifa_cost;
       int sort_p;
-      bool asm_p, real_insn, after_stall, all_issued;
+      bool asm_p, real_insn, after_stall, all_issued, empty;
       int clock;

       if (!INSN_P (insn))
@@ -7047,7 +7050,7 @@ reset_sched_cycles_in_current_ebb (void)
            haifa_cost = 0;
        }
       else
-        haifa_cost = estimate_insn_cost (insn, curr_state);
+        haifa_cost = estimate_insn_cost (insn, curr_state, &empty);

       /* Stall for whatever cycles we've stalled before.  */
       after_stall = 0;
@@ -7081,7 +7084,7 @@ reset_sched_cycles_in_current_ebb (void)
               if (!after_stall
                   && real_insn
                   && haifa_cost > 0
-                  && estimate_insn_cost (insn, curr_state) == 0)
+                  && estimate_insn_cost (insn, curr_state, NULL) == 0)
                 break;

               /* When the data dependency stall is longer than the DFA stall,
@@ -7093,7 +7096,7 @@ reset_sched_cycles_in_current_ebb (void)
               if ((after_stall || all_issued)
                   && real_insn
                   && haifa_cost == 0)
-                haifa_cost = estimate_insn_cost (insn, curr_state);
+                haifa_cost = estimate_insn_cost (insn, curr_state, NULL);
             }

          haifa_clock += i;
@@ -7125,7 +7128,8 @@ reset_sched_cycles_in_current_ebb (void)
       if (real_insn)
        {
          cost = state_transition (curr_state, insn);
-         issued_insns++;
+         if (!empty)
+           issued_insns++;

           if (sched_verbose >= 2)
            {

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

Andrey Belevantsev <abel at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ubizjak at gmail dot com

--- Comment #3 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-02-13 08:49:38 UTC ---
Even better with Uros in CC.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2012-02-13 15:54:28 UTC ---
(In reply to comment #2)
> Well, in this case the insn is vcvtps2pd (from avx_cvtps2pd256 and
> *avx_cvtps2pd256_2 patterns) but not having a reservation looks correct as the
> manual says "decode type -- microcode".  Uros, is that right?  In this case,
> the below fixes the test for me (and I was stupid enough not to do this in the
> first place when this bug stream just started...  However, we did fix some real
> bugs.)

There is nothing wrong with insn definition, the problem is only that
attributes are missing.  Attributes are considered a tool to fine-tune the
scheduler and in any case, gcc should not ICE on an insn without reservation.
The insn doesn't reserve anything in this case and scheduler should live with
this.

The bugs you are referring to were real bugs, where wrong or missing operand
was accessed from attribute calculation machinery.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #5 from Uros Bizjak <ubizjak at gmail dot com> 2012-02-13 16:00:27 UTC ---
(In reply to comment #4)

> The bugs you are referring to were real bugs, where wrong or missing operand
> was accessed from attribute calculation machinery.

... and PR 49014, where indeed insn without reservation crashed scheduler in
the same way. But there was a clear inconsistency in the model, so we got away
with the adjustment.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #6 from Uros Bizjak <ubizjak at gmail dot com> 2012-02-13 16:01:40 UTC ---
(In reply to comment #5)
> (In reply to comment #4)
> 
> > The bugs you are referring to were real bugs, where wrong or missing operand
> > was accessed from attribute calculation machinery.
> 
> ... and PR 49014, where indeed insn without reservation crashed scheduler in
> the same way. But there was a clear inconsistency in the model, so we got away
> with the adjustment.

Maybe a warning in the logs could be emitted for insns without reservations?

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |steven at gcc dot gnu.org

--- Comment #7 from Steven Bosscher <steven at gcc dot gnu.org> 2012-02-13 21:14:35 UTC ---
A warning in the scheduler log for insns without a reservation would be nice.
Doing it from genautomata would be quite difficult, but the scheduler can
detect the situation easily and issue a diagnostic.

How does the Haifa scheduler handle such insns?

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #8 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-02-14 06:56:10 UTC ---
Sorry, I didn't explain clearly from the start.  Regarding the backend, I just
wanted to double check that for the given insn not having a reservation is
correct.  Now that Uros confirmed this, I can proceed with the scheduler patch.

Uros and Steven, regarding the big picture, you can read the thread on PR
49014, starting from http://gcc.gnu.org/ml/gcc-patches/2011-05/msg01801.html
(our first discussions with Bernd) and initial patch with the remainder of the
thread at http://gcc.gnu.org/ml/gcc-patches/2011-07/msg00045.html.  Basically,
we allow insns that recog as >=0 but do not change the DFA state, such insns do
not get counted against issue rate.  I was wondering whether we could mark such
insns specially in the MD files and then assert that any other real insn should
affect the DFA.  The reason for this was that at least three real backend bugs
(missing DFA attributes) were fixed due to the more strict checking the
selective scheduler makes wrt the Haifa scheduler.  

Suprisingly, I could bootstrap x86-64 with just some hours of work with the
patch from http://gcc.gnu.org/ml/gcc-patches/2011-07/msg00045.html (the patch
marks the offending insns with the new attribute and adds an assert; Uros, if
you can spot any insns there that indeed should have a reservation, I can make
a patch fixing those).  However, at the time I gave in pursuing this work
further, as nobody was very fascinated with the idea and I didn't have much
free time.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #9 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-03-07 12:00:42 UTC ---
Author: abel
Date: Wed Mar  7 12:00:37 2012
New Revision: 185036

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185036
Log:
        PR rtl-optimization/52203
        * sel-sched.c (estimate_insn_cost): New parameter pempty.  Adjust
        all callers to pass NULL except ...
        (reset_sched_cycles_in_current_ebb): ... here, save the value
        in new variable 'empty'.  Increase issue_rate only for
        non-empty insns.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/sel-sched.c

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

Andrey Belevantsev <abel at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.8.0

--- Comment #10 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-03-07 12:02:28 UTC ---
Fixed on trunk, will be ported for 4.7.1.

[Bug rtl-optimization/52203] ICE: in reset_sched_cycles_in_current_ebb, at sel-sched.c:7136 with -fsel-sched-pipelining -fselective-scheduling2 and other custom flags

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52203

--- Comment #11 from Andrey Belevantsev <abel at gcc dot gnu.org> 2012-04-13 09:36:46 UTC ---
Author: abel
Date: Fri Apr 13 09:36:42 2012
New Revision: 186410

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=186410
Log:
        PR rtl-optimization/52203
        PR rtl-optimization/52715

        Revert the 2012-03-07 fix for PR 52203.
        * sel-sched.c (reset_sched_cycles_in_current_ebb): Check that
        the insn does not modify DFA right before issuing, adjust
        issue_rate accordingly.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/sel-sched.c