[Bug rtl-optimization/57067] New: Missing control flow edges for setjmp/longjmp

[Bug rtl-optimization/57067] New: Missing control flow edges for setjmp/longjmp

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

             Bug #: 57067
           Summary: Missing control flow edges for setjmp/longjmp
    Classification: Unclassified
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: krebbel@gcc.gnu.org


The fix for PR56982 adds abnormal control flow edges from function calls to a
setjmp call in the same function. Unfortunately these edges do not survive
until RTL so that the RTL passes might still do the wrong thing. 

The edges are removed in gimple_expand_cfg:

      /* At the moment not all abnormal edges match the RTL
         representation.  It is safe to remove them here as
         find_many_sub_basic_blocks will rediscover them.
         In the future we should get this fixed properly.  */
      if ((e->flags & EDGE_ABNORMAL)
          && !(e->flags & EDGE_SIBCALL))
        remove_edge (e);
      else
        ei_next (&ei);


find_many_sub_basic_blocks needs a fix to add them back as well.

I don't have a testcase for GCC head.  The testcase I have fails only with GCC
4.4: http://gcc.gnu.org/ml/gcc/2013-04/msg00237.html

In this case the RTL scheduler pass generates broken code due to the missing
control flow info.  I think this could potentially happen with GCC head as
well.

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-04-25
                 CC|                            |matz at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-25 09:17:03 UTC ---
I suppose more selectively removing edges would be best.  Eventually this is
done to cater for expansions of builtin calls to non-calls?  Then maybe
those edges should be removed later, _after_ builtin expansion and
find_many_sub_basic_blocks?

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #2 from Andreas Krebbel <krebbel at gcc dot gnu.org> 2013-04-25 14:46:09 UTC ---
(In reply to comment #1)
> I suppose more selectively removing edges would be best.  Eventually this is
> done to cater for expansions of builtin calls to non-calls?  Then maybe
> those edges should be removed later, _after_ builtin expansion and
> find_many_sub_basic_blocks?

Wouldn't it be better to keep all the edges and fix the fallout from the RTL
passes?

Do you remember any example where this currently fails?

I've verified with GCC 4.4 that dropping the edge removal code from cfgexpand.c
together with a backport of the PR56982 patch fixes the actual problem.

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> 2013-04-25 15:11:00 UTC ---
On Thu, 25 Apr 2013, krebbel at gcc dot gnu.org wrote:

> 
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067
> 
> --- Comment #2 from Andreas Krebbel <krebbel at gcc dot gnu.org> 2013-04-25 14:46:09 UTC ---
> (In reply to comment #1)
> > I suppose more selectively removing edges would be best.  Eventually this is
> > done to cater for expansions of builtin calls to non-calls?  Then maybe
> > those edges should be removed later, _after_ builtin expansion and
> > find_many_sub_basic_blocks?
> 
> Wouldn't it be better to keep all the edges and fix the fallout from the RTL
> passes?

Yes, we should be able to prune them when expansion sees that is 
necessary.

> Do you remember any example where this currently fails?

Not sure.  Run bootstrap & testing?

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[gretay at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #4 from gretay at gcc dot gnu.org 2013-04-29 16:58:34 UTC ---
Created attachment 29974
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29974
testcase

The attached test case fails for arm-none-eabi on trunk, caused by r198096 (the
fix for PR56982). It seems related to this PR because the control flow edge
after testing the return value of malloc disappears and it is sensitive to the
setjmp() call placement.

Compile with:
/work/apr-builds/r198096/install/bin/arm-none-eabi-gcc -O1 -mcpu=cortex-a15
pr57067.c -o bad.elf
The test case should print PASS, but it prints FAIL.

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
The RTL except.c:can_nonlocal_goto () function does not consider setjmp.
Does

Index: gcc/except.c
===================================================================
--- gcc/except.c        (revision 198867)
+++ gcc/except.c        (working copy)
@@ -1936,7 +1936,9 @@ insn_nothrow_p (const_rtx insn)
 bool
 can_nonlocal_goto (const_rtx insn)
 {
-  if (nonlocal_goto_handler_labels && CALL_P (insn))
+  if ((nonlocal_goto_handler_labels
+       || cfun->calls_stjmp)
+      && CALL_P (insn))
     {
       rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
       if (!note || INTVAL (XEXP (note, 0)) != INT_MIN)

fix it?

For retaining edges we have to teach find_many_sub_basic_blocks to treat
abnormal outgoing edges similar to EH outgoing edges - they have to be
re-directed from the block ending in the actual call (and made
EDGE_ABNORMAL_CALL which doesn't exist on the GIMPLE CFG).

Also

          if ((e->flags & EDGE_ABNORMAL)
              && !(e->flags & EDGE_SIBCALL))
            remove_edge (e);

will happily remove an EDGE_FALLTHRU|EDGE_ABNORMAL edge.

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[gretay at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

gretay at gcc dot gnu.org changed:

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

--- Comment #6 from gretay at gcc dot gnu.org ---
I tried the proposed patch (after fixing a typo: calls_setjmp / calls_stjmp).
It causes an ICE on the attached testcase (and some regressions):

$ /work/apr-builds/pr57067/install/bin/arm-none-eabi-gcc -O1 -mcpu=cortex-a15
pr57067.c -o pr57067.elf 
pr57067.c: In function 'main':
pr57067.c:18:7: warning: incompatible implicit declaration of built-in function
'malloc' [enabled by default]
   p = malloc (0x1000);
       ^
pr57067.c:22:7: warning: incompatible implicit declaration of built-in function
'printf' [enabled by default]
       printf ("\nFAIL\n");
       ^
pr57067.c:23:7: warning: incompatible implicit declaration of built-in function
'exit' [enabled by default]
       exit (1);
       ^
pr57067.c:32:3: warning: incompatible implicit declaration of built-in function
'printf' [enabled by default]
   printf ("\nPASS\n");
   ^
pr57067.c:34:1: error: in basic block 2:
 }
 ^
pr57067.c:34:1: error: flow control insn inside a basic block
(call_insn 8 7 9 2 (parallel [
            (set (reg:SI 0 r0)
                (call (mem:SI (symbol_ref:SI ("malloc") [flags 0x41]
<function_decl 0x7f0cc7362900 malloc>) [0 __builtin_malloc S4 A32])
                    (const_int 0 [0])))
            (use (const_int 0 [0]))
            (clobber (reg:SI 14 lr))
        ]) pr57067.c:18 -1
     (expr_list:REG_EH_REGION (const_int 0 [0])
        (nil))
    (expr_list:REG_CFA_WINDOW_SAVE (use (reg:SI 0 r0))
        (nil)))
pr57067.c:34:1: internal compiler error: in rtl_verify_flow_info_1, at
cfgrtl.c:2321
0x7cc383 _fatal_insn(char const*, rtx_def const*, char const*, int, char
const*)
    /work/local-checkouts/gcc-fsf/gcc/rtl-error.c:109
0x52ed65 rtl_verify_flow_info_1
    /work/local-checkouts/gcc-fsf/gcc/cfgrtl.c:2321
0x52efb2 rtl_verify_flow_info
    /work/local-checkouts/gcc-fsf/gcc/cfgrtl.c:2345
0x51ef49 verify_flow_info()
    /work/local-checkouts/gcc-fsf/gcc/cfghooks.c:258
0xc02464 try_optimize_cfg
    /work/local-checkouts/gcc-fsf/gcc/cfgcleanup.c:2812
0xc02464 cleanup_cfg(int)
    /work/local-checkouts/gcc-fsf/gcc/cfgcleanup.c:2974
0x51c7dd gimple_expand_cfg
    /work/local-checkouts/gcc-fsf/gcc/cfgexpand.c:4781

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067
Bug 57067 depends on bug 56982, which changed state.

Bug 56982 Summary: [4.8 Regression] Bad optimization with setjmp()
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56982

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

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

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 108132 has been marked as a duplicate of this bug. ***

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |fxue at os dot amperecomputing.com

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 108117 has been marked as a duplicate of this bug. ***

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #9 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
*** Bug 108117 has been marked as a duplicate of this bug. ***

[Bug rtl-optimization/57067] Missing control flow edges for setjmp/longjmp

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57067

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 108132 has been marked as a duplicate of this bug. ***