[Bug target/55433] New: [4.8 Regression] ICE on excessive reloads

["PHHargrove at lbl dot gov" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55433

             Bug #: 55433
           Summary: [4.8 Regression] ICE on excessive reloads
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: PHHargrove@lbl.gov
                CC: gary@intrepid.com, nenad@intrepid.com


The 9-line stand-alone test case at the end of this report fails as show below
on MacOS (at least versions 10.6 and 10.8) on x86-64 machines when targeting
the ILP32 ABI.  It notably does NOT fail on Linux/x86-64.

This was seen in SVN r193679 (on MacOSX 10.6) and r193709 (on MacOSX 10.8).

--cut--
$ gcc -m32 -O1 -fpreprocessed -c testcase.c
testcase.c: In function 'test':
testcase.c:10:1: internal compiler error: Max. number of generated reload insns
per insn is achieved (90)

 }
 ^

testcase.c:9:1: internal compiler error: Abort trap
gcc: internal compiler error: Abort trap (program cc1)
Abort trap
--cut--

Note that while the error message is the same as bug 55247, I have confirmed
that the test case attached to that bug report has NOT regressed.  Therefore, I
believe this is a different bug (though possibly related).

Some additional data:
+  -m32 is required (no bug seen w/ -m64)
+  Linux host is fine (w/ -m32 and -m64)
+  -O1 is sufficient, but -O2 and -O3 fail in the same manner
+  removing "unsigned" from the typedef eliminates the problem
+  adding -mno-sse to the command line eliminates the problem


The testcase:

typedef unsigned long long tick_t; // NOTE: no bug if "unsigned" is removed
extern int foo(void);
extern tick_t tick(void);
double test(void) {
  struct { tick_t ticks; } st;
  st.ticks = tick();
  foo(); // NOTE: no bug if this call is removed
  return (double)st.ticks;
}