This doesn't happen on just amd64, I've seen this bug on all targets 
when I was playing with a patch to blow away libcall notes.  On ia64 
it lead to a link failure because of missing symbols, so fixing this 
problem will probably uncover a few latent bugs elsewhere. 
 

When I looked at this, the libcall was eliminated in the call to 
delete_trivially_dead_insns, somewhere around jump1. 
 

*** Bug 25418 has been marked as a duplicate of this bug. ***

*** Bug 27261 has been marked as a duplicate of this bug. ***

There are a few issues, first we use emit_libcall_block to emit the trapping
PLUS
which sets a REG_EQUAL note with a non-trapping PLUS.

Second we analyze iaddv as not possibly throwing so we remove the call from
main()
during tree optimization as the result is unused.

Third, we widen the integer addition to DImode and dispatch to libgcc2
__addvdi3
which looks like

Dump of assembler code for function __addvdi3:
0x0000000000400580 <__addvdi3+0>:       sub    $0x8,%rsp
0x0000000000400584 <__addvdi3+4>:       test   %rsi,%rsi
0x0000000000400587 <__addvdi3+7>:       lea    (%rdi,%rsi,1),%rax
0x000000000040058b <__addvdi3+11>:      js     0x4005a0 <__addvdi3+32>
0x000000000040058d <__addvdi3+13>:      cmp    %rax,%rdi

so it tests for 64bit overflow instead of 32bit one.  Obviously allowind
LIBCALL_WIDEN is wrong for the trapping optabs, too.

*** Bug 32153 has been marked as a duplicate of this bug. ***

With trunk as of today, the test case of comment #0 does trap if optimization
is disabled. At -O, the libcall is optimized away, but the call to iaddv() from
main() is also optimized away because iaddv is found to be pure.  You must use
the result of iaddv() to avoid this:

#include <limits.h>

int __attribute__((noinline))
iaddv (int a, int b)
{
  return a + b;
}

int main(void)
{
  return iaddv (INT_MAX, 1);
}

On the expand side it is just wrong (and asking for trouble) to go the usual
expand_binop way for expanding trapping arithmetic.  At least we would need to
pass down a flag.

If you remove libcall notes around all binops then the following testcase
passes in dg-torture:

/* { dg-do run { target hppa*-*-hpux* *-*-linux* powerpc*-*-darwin*
*-*-darwin[912]* } } */
/* { dg-options "-ftrapv" } */

#include <stdlib.h>
#include <limits.h>
#include <signal.h>

int ok;

void handle_abort (int sig)
{
  if (sig == SIGABRT)
    {
      ok = 1;
      exit (0);
    }
}

int __attribute__((noinline))
iaddv (int a, int b)
{
  return a + b;
}

volatile int x;

int main(void)
{
  struct sigaction sa;
  sigemptyset (&sa.sa_mask);
  sa.sa_handler = handle_abort;
  sa.sa_flags = SA_RESETHAND;
  if (sigaction (SIGABRT, &sa, NULL) == -1)
    return 0;
  /* ???  iaddv is analyzed to be pure, so use the result.  */
  x = iaddv (INT_MAX, 1);
  sa.sa_handler = SIG_DFL;
  if (sigaction (SIGABRT, &sa, NULL) == -1)
    return 0;
  if (!ok)
    abort ();
  return 0;
}

but I get a build failure of libjava then:

gnu/java/nio/natVMSelector.cc: In static member function 'static jint
gnu::java::nio::VMSelector::select(JArray<__java_int>*, JArray<__java_int>*,
JArray<__java_int>*, jlong)':
gnu/java/nio/natVMSelector.cc:127: error: unable to find a register to spill in
class 'AD_REGS'
gnu/java/nio/natVMSelector.cc:127: error: this is the insn:
(insn 58 57 59 6 gnu/java/nio/natVMSelector.cc:82 (parallel [
            (set (reg:DI 4 si [165])
                (mult:DI (zero_extend:DI (reg:SI 0 ax))
                    (zero_extend:DI (reg:SI 2 cx [166]))))
            (clobber (reg:CC 17 flags))
        ]) 304 {*umulsidi3_insn} (expr_list:REG_DEAD (reg:SI 2 cx [166])
        (expr_list:REG_DEAD (reg:SI 0 ax)
            (expr_list:REG_UNUSED (reg:CC 17 flags)
                (expr_list:REG_EQUAL (mult:DI (zero_extend:DI (reg:SI 0 ax))
                        (const_int 1000 [0x3e8]))
                    (nil))))))
gnu/java/nio/natVMSelector.cc:127: internal compiler error: in spill_failure,
at reload1.c:2000
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.


with the following patch

Index: optabs.c
===================================================================
--- optabs.c    (revision 132790)
+++ optabs.c    (working copy)
@@ -2148,12 +2148,9 @@ expand_binop (enum machine_mode mode, op

       insns = get_insns ();
       end_sequence ();
+      emit_insn (insns);

-      target = gen_reg_rtx (mode);
-      emit_libcall_block (insns, target, value,
-              gen_rtx_fmt_ee (binoptab->code, mode, op0, op1));
-
-      return target;
+      return value;
     }

   delete_insns_since (last);


I'll test forcing value to a register.

Doesn't help.

Quoting your insn once more:

(insn 58 57 59 6 gnu/java/nio/natVMSelector.cc:82 (parallel [
            (set (reg:DI 4 si [165])
                (mult:DI (zero_extend:DI (reg:SI 0 ax))
                    (zero_extend:DI (reg:SI 2 cx [166]))))
            (clobber (reg:CC 17 flags))
        ]) 304 {*umulsidi3_insn} (expr_list:REG_DEAD (reg:SI 2 cx [166])
        (expr_list:REG_DEAD (reg:SI 0 ax)
            (expr_list:REG_UNUSED (reg:CC 17 flags)
                (expr_list:REG_EQUAL (mult:DI (zero_extend:DI (reg:SI 0 ax))
                        (const_int 1000 [0x3e8]))
                    (nil))))))

The register allocator failed to make your insn satisfy its constraints.
Operand 0 is (reg:DI 4 si) but the constraint for it is "=A", i.e. the ax
register.

The funny thing is that reload wants to make operand 0 be (reg:DI 0 ax), but
somehow it can't, even though (reg:SI 0 ax) dies in this insn. So apparently
the high part of (reg:DI 0 ax), i.e. (reg:SI 1 dx), lives across the insn.

This looks to me like a case of PR35404.  What do you think, Ian?

*** Bug 36398 has been marked as a duplicate of this bug. ***

*** Bug 36868 has been marked as a duplicate of this bug. ***

libcalls are gone in 4.4 and later.

I'm not sure if this bug is fixed, though.  -ftrapv is still broken afaik.

> I'm not sure if this bug is fixed, though.  -ftrapv is still broken afaik.

Probably, but let's ditch antiquated stuff and start afresh.