https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113734
--- Comment #19 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
Ok, removing all the noise shows that this is the same issue as I saw before.
The code out of the vectorizer is correct, but cunroll does a dodgee unrolling.
-fdisable-tree-cunroll confirms it's the unrolling.
cunroll claims:
Loop 4 iterates at most 16 times.
Loop 4 likely iterates at most 16 times.
Analyzing # of
iterations of loop 4
exit condition [1, + , 1](no_overflow) < bnd.157_285
...
;; Guessed iterations of loop 4 is 7.347979. New upper bound 16.
Making edge 23->36 impossible by redistributing
probability to other edges. Original probability: 93.8% (guessed)
misc.c:147:31:
optimized: loop with 16 iterations completely unrolled (header execution count
2859449)
Last iteration exit edge was proved true.
Not peeling: number of iterations
is not estimated
but the i is bounded by i < 306. and VF=8. so binding loop iteration to 16
means it just cut off half the loop.
so it looks like the upper bounds is wrong. Checking what we write out during
loop vect.
