Confirmed, let me bisect.

Note you may need -march=x86-64 to reproduce.

Started with r228231:

commit 3ef16338b7f4970607d1d0b9c8228bfe3aeeca94
Author: ienkovich <ienkovich@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Tue Sep 29 09:32:40 2015 +0000

    gcc/
    
        PR target/65105
        * config/i386/i386.c: Include dbgcnt.h.
...

Looks like a RA issue to me.
At least, on -m32 -O3 -march=x86-64 compiled:
short a;
int b = 1, f;
char c, e = 1;
long long d;

static short
foo ()
{
  unsigned g, h = 0;
  int i = 0 || d * (b | e);
  char j = a << i, l = a;
  short k;
  int m = -b;
  if (m < b)
    {
      k = m = b;
      g = (k || l) / (b / e);
      if (b)
	__builtin_printf ("foo=%lld\n", (long long) a);
    }
  a = b = m;
  if (j || e)
    {
      h = g;
      i = m;
      g = j * k / (i - d);
      if (m)
	b = j && b;
      e = b * (h & d) || g;
    }
  b = i;
  char n = e || h | d;
  e = i < d & k / n;
  return f;
}

int
main ()
{
  if (foo ())
    if (c)
     lab:
      goto lab;
  return 0;
}

we in *.ira still have correct:
(insn 30 28 31 2 (parallel [
            (set (reg/v:SI 103 [ m ])
                (neg:SI (reg:SI 91 [ b.2_7 ])))
            (clobber (reg:CC 17 flags))
        ]) pr68990-4.c:13 463 {*negsi2_1}
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))
(insn 31 30 32 2 (set (reg:CCGC 17 flags)
        (compare:CCGC (reg:SI 91 [ b.2_7 ])
            (reg/v:SI 103 [ m ]))) pr68990-4.c:14 7 {*cmpsi_1}
     (nil))
(jump_insn 32 31 33 2 (set (pc)
        (if_then_else (le (reg:CCGC 17 flags)
                (const_int 0 [0]))
            (label_ref 56)
            (pc))) pr68990-4.c:14 631 {*jcc_1}
     (expr_list:REG_DEAD (reg:CCGC 17 flags)
        (int_list:REG_BR_PROB 5000 (nil)))
and
31:r91  l0   mem
13:r103 l0     3
but LRA turns that into:
(insn 256 207 208 2 (set (reg:SI 1 dx [orig:91 b.2_7 ] [91])
        (mem/c:SI (plus:SI (reg/f:SI 6 bp)
                (const_int -28 [0xffffffffffffffe4])) [5 %sfp+-4 S4 A32])) pr68990-4.c:13 86 {*movsi_internal}
     (nil))
(insn 208 256 30 2 (set (reg/v:SI 3 bx [orig:103 m ] [103])
        (reg:SI 1 dx [orig:91 b.2_7 ] [91])) pr68990-4.c:13 86 {*movsi_internal}
     (nil))
(insn 30 208 286 2 (parallel [
            (set (reg/v:SI 3 bx [orig:103 m ] [103])
                (neg:SI (reg/v:SI 3 bx [orig:103 m ] [103])))
            (clobber (reg:CC 17 flags))
        ]) pr68990-4.c:13 463 {*negsi2_1}
     (nil))
(insn 286 30 210 2 (set (mem/c:SI (plus:SI (reg/f:SI 6 bp)
                (const_int -28 [0xffffffffffffffe4])) [5 %sfp+-4 S4 A32])
        (reg:SI 3 bx [orig:91 b.2_7 ] [91])) pr68990-4.c:13 86 {*movsi_internal}
     (nil))
(insn 210 286 31 2 (set (reg:SI 3 bx [orig:91 b.2_7 ] [91])
        (reg:SI 1 dx [orig:91 b.2_7 ] [91])) pr68990-4.c:14 86 {*movsi_internal}
     (nil))
(insn 31 210 32 2 (set (reg:CCGC 17 flags)
        (compare:CCGC (reg:SI 3 bx [orig:91 b.2_7 ] [91])
            (reg:SI 3 bx [orig:91 b.2_7 ] [91]))) pr68990-4.c:14 7 {*cmpsi_1}
     (nil))
(jump_insn 32 31 33 2 (set (pc)
        (if_then_else (le (reg:CCGC 17 flags)
                (const_int 0 [0]))
            (label_ref 56)
            (pc))) pr68990-4.c:14 631 {*jcc_1}
     (int_list:REG_BR_PROB 5000 (nil))
 -> 56)
which is just wrong, instead of comparing m < b it suddenly compares b < b.

I think the problem is:
********** Pseudos coalescing #1: **********

  Coalescing move 5:r91(91)-r103(103) (freq=1)
         Removing move 5 (freq=1)
deleting insn with uid = 5.
         Make unique value for inheritance r256
         Make unique value for inheritance r257
         Make unique value for inheritance r258
         Make unique value for inheritance r259
         Make unique value for inheritance r260
         Make unique value for inheritance r262
         Make unique value for inheritance r264
Coalesced Moves = 1

in the other spot where insn 5 is it is perhaps possible to coalesce the two pseudos, but not at the
  int m = -b;
  if (m < b)
comparison, where it breaks it.

BTW, with -mno-stv -O3 the testcase passes, and while the *.stv pass actually doesn't do anything at all on this testcase, that switch already changes expansion in a couple of places so no wonder that the RA makes different decisions.

And, I haven't managed to create a testcase that would abort vs. return 0 from main for good vs. bad, even other fail defining a stdarg function like printf didn't help.  Do, unless somebody comes up with one, guess we'll need a dg-output directive and try to match what is printed (without the newline in there).

(In reply to Jakub Jelinek from comment #4)
> I think the problem is:
> ********** Pseudos coalescing #1: **********
> 
>   Coalescing move 5:r91(91)-r103(103) (freq=1)
>          Removing move 5 (freq=1)
> deleting insn with uid = 5.
>          Make unique value for inheritance r256
>          Make unique value for inheritance r257
>          Make unique value for inheritance r258
>          Make unique value for inheritance r259
>          Make unique value for inheritance r260
>          Make unique value for inheritance r262
>          Make unique value for inheritance r264
> Coalesced Moves = 1
> 
> in the other spot where insn 5 is it is perhaps possible to coalesce the two
> pseudos, but not at the
>   int m = -b;
>   if (m < b)
> comparison, where it breaks it.

I've reproduced the bug and checked the coalescing.   The coalescing seems the right transformation as p91 becomes dead right before insn#30 and actually its value before insn#30 through an inheritance pseudo is used in insn #31.  The problem actually occurs later in pseudo assignment sub-pass.

I suspect that something is wrong with pseudo values.  Changes affecting pseudo values much affect the performance much.  So when I have the patch, I need to check the performance impact too.  I think if it is everything ok, the patch will be ready at the end of week.

Author: vmakarov
Date: Thu Jan 21 16:01:22 2016
New Revision: 232679

URL: https://gcc.gnu.org/viewcvs?rev=232679&root=gcc&view=rev
Log:
2016-01-21  Vladimir Makarov  <vmakarov@redhat.com>

	PR rtl-optimization/68990
	* lra-coalesce.c (lra_coalesce): Invalidate value for the result
	pseudo instead of inheritance ones.

2016-01-21  Vladimir Makarov  <vmakarov@redhat.com>

	PR rtl-optimization/68990
	* gcc.target/i386/pr68990: New.



Added:
    trunk/gcc/testsuite/gcc.target/i386/pr68990.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/lra-coalesce.c
    trunk/gcc/testsuite/ChangeLog

Fixed by Vlad's commit on the trunk.