On Mon, Nov 19, 2012 at 12:24 PM, Eric Botcazou <[email protected]> wrote:
>> Yes, I'll be looking into this soon.
>
> We have a related regression on SPARC:
>
> FAIL: gfortran.dg/minmaxloc_5.f90 -O3 -fomit-frame-pointer -funroll-loops
> execution test
> FAIL: gfortran.dg/minmaxloc_5.f90 -O3 -fomit-frame-pointer -funroll-all-loops
> -finline-functions execution test
> FAIL: gfortran.dg/minmaxloc_6.f90 -O3 -fomit-frame-pointer -funroll-loops
> execution test
> FAIL: gfortran.dg/minmaxloc_6.f90 -O3 -fomit-frame-pointer -funroll-all-loops
> -finline-functions execution test
>
> whose failure mode is exactly the same as for Honza's testcase. I even have a
> more reduced testcase (6 lines, attached) in case you'd prefer working on it.
>
> Reproducible with a cross to sparc-linux with -mcpu=v8 -O3 -funroll-loops and
> grepping for mem:SF (const_int 0 [0]) in the RTL dumps.
Thanks for this reduced test case, that's saving me a lot of work!
Can you please try and see if the following C test case also fails?
It looks like the memcpy is expanded to a loop that is unrolled, and
then mis-optimized. I haven't found out yet why...
Ciao!
Steven
extern void abort (void) __attribute__ ((__noreturn__));
extern void *memcpy (void *__restrict, const void *__restrict, __SIZE_TYPE__);
static void __attribute__ ((__noinline__, __noclone__)) ga4076 (void)
{
int D833;
float D834;
int D837;
int D840;
float D841;
int D844;
float dda[100];
int ids;
static float A0[100] =
{ 10e+0, 20e+0, 30e+0, 40e+0, 50e+0, 60e+0, 70e+0, 80e+0, 90e+0,
10e+1, 11e+1, 12e+1, 13e+1, 14e+1, 15e+1, 16e+1, 17e+1, 18e+1,
19e+1, 20e+1, 21e+1, 22e+1, 23e+1, 24e+1, 25e+1, 26e+1, 27e+1,
28e+1, 29e+1, 30e+1, 31e+1, 32e+1, 33e+1, 34e+1, 35e+1, 36e+1,
37e+1, 38e+1, 39e+1, 40e+1, 41e+1, 42e+1, 43e+1, 44e+1, 45e+1,
46e+1, 47e+1, 48e+1, 49e+1, 50e+1, 51e+1, 52e+1, 53e+1, 54e+1,
55e+1, 56e+1, 57e+1, 58e+1, 59e+1, 60e+1, 61e+1, 62e+1, 63e+1,
64e+1, 65e+1, 66e+1, 67e+1, 68e+1, 69e+1, 70e+1, 71e+1, 72e+1,
73e+1, 74e+1, 75e+1, 76e+1, 77e+1, 78e+1, 79e+1, 80e+1, 81e+1,
82e+1, 83e+1, 84e+1, 85e+1, 86e+1, 87e+1, 88e+1, 89e+1, 90e+1,
91e+1, 92e+1, 93e+1, 94e+1, 95e+1, 96e+1, 97e+1, 98e+1, 99e+1,
10e+2
};
memcpy (dda, A0, sizeof (dda));
float limit3;
int offset2;
int pos1;
int S4;
limit3 = -1. * __builtin_inf();
pos1 = 0;
offset2 = 0;
S4 = 1;
D831:
if (S4 > 100) goto L3; else goto D832;
D832:
D833 = S4 - 1;
D834 = dda[D833];
if (D834 >= limit3) goto D835; else goto D836;
D835:
D837 = S4 - 1;
limit3 = dda[D837];
pos1 = S4 + offset2;
goto L1;
D836:
S4 = S4 + 1;
goto D831;
L3:
pos1 = 1;
goto L2;
L1:
if (S4 > 100) goto L2; else goto D839;
D839:
D840 = S4 - 1;
D841 = dda[D840];
if (D841 > limit3) goto D842; else goto D843;
D842:
D844 = S4 - 1;
limit3 = dda[D844];
pos1 = S4 + offset2;
D843:
S4 = S4 + 1;
goto L1;
L2:
ids = pos1;
if (ids != 100)
abort ();
}
int
main (void)
{
ga4076 ();
return 0;
}
