https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85186
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-04-04
CC| |law at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
Blocks| |53947
Summary|ifcvt can peel iterations |jump threading can rotate
|affecting code form, and |loops affecting loop form,
|causing vectorization not |and causing vectorization
|to happen |not to happen
Ever confirmed|0 |1
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's not if-conversion but jump threading doing this. If you specify
-fno-tree-dominator-opts you get back the vectorization. Likely a regression
from the
time DOM gained the ability to CSE A[0] = 36 across unrelated stores.
It's also not affecting code form but the number of iterations which on x86
is then too small to be vectorized.
Not sure how we can identify good/bad jump threadings... in this case it
rotates the loop into non-canonical (not loop header copied) form and
the loop header copying pass before the vectorizer completes the peeling
then to give it a chance to be vectorized.
I guess DOM needs a cost model to determine when loop rotation is profitable.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
