http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54717
--- Comment #16 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-11-15
10:52:13 UTC ---
OK, 4.7 vectorize two loops in the function in cptrf2
loop at ../a.f90:3538
if (nxtr < 4) then
kerr = 1
do ixtr = 1, nxtr - 1
ixtrt (ixtr) = ixtr + 1
enddo
goto 9000
endif
and
loop at ../a.f90:3530
ixtrt = 0
The second loop is recognized as memset by mainline, so it remains to figure
out what is wrong with the first loop. It is unrolled:
Analyzing # of iterations of loop 9
exit condition [1, + , 1](no_overflow) != ival2_27 + -1
bounds on difference of bases: 0 ... 1
result:
# of iterations (unsigned int) ival2_27 + 4294967294, bounded by 1
Loop 9 iterates at most 1 times.
Estimating sizes for loop 9
BB: 8, after_exit: 0
size: 0 _38 = (integer(kind=8)) ixtr_12;
Induction variable computation will be folded away.
size: 1 _39 = _38 + -1;
Induction variable computation will be folded away.
size: 1 ixtr_40 = ixtr_12 + 1;
Induction variable computation will be folded away.
size: 1 *ixtrt_33(D)[_39] = ixtr_40;
size: 2 if (ixtr_12 == _37)
Exit condition will be eliminated in last copy.
BB: 79, after_exit: 1
size: 5-2, last_iteration: 5-4
Loop size: 5
Estimated size after unrolling: 2
Unrolled loop 9 completely (duplicated 1 times).
I do not quite see why it iterates at most once, but if seems to work. So I
would say that it is good idea to unroll rather than vectorize.
Is the slowdown still reproducing with my patch?