Richard Biener wrote:
Apart from Jeffs comment - the usual fix for the undesired
vectorization is to put
a __asm__ volatile (""); in the loop.
In vect-strided-a-u16-i4.c, narrowing the scope of the declaration seemed to preserve the original intent. I've been able to drop the other testsuite changes.
+ /* If any block in the loop has an exit edge, and code after it, it is + not a do-while loop. */ + basic_block *body = get_loop_body (loop); + for (unsigned i = 0; i < loop->num_nodes; i++) wouldn't it be easier to verify that the predecessor of the loop latch contains the (only) loop exit?
It's not guaranteed that the loop latch has only one predecessor. The testsuite contains quite a few examples, e.g. gcc.c-torture/compile/20011114.c (at -O3). However, I've found a simpler (and equivalent) test, as we have the unique exit edge and it's source already.
Note that single_exit () only works when the loop state has LOOPS_HAVE_RECORDED_EXITS
Hah, thanks - didn't realize that. So using single_exit_p did make pass_ch behave differently from pass_ch_vect. I've restored the original code for the original pass_ch...
> I think pass_ch_vect > should be only executed if flag_tree_loop_vectorize is enabled....agreed; and handling loop->force_vectorize and loop->dont_vectorize properly required splitting the two phases up more anyway, so I've used clearly-different predicates in each.
Coding-style wise, can you please move the "common" pass_ch_vect::execute out of the pass_ch_vect class?
Yes, I've done some reorg, introducing a third base class with the common execute bits calling a virtual method returning bool.
loop_optimizer_init (LOOPS_NORMAL - | LOOPS_HAVE_RECORDED_EXITS); + | LOOPS_HAVE_RECORDED_EXITS + | LOOPS_HAVE_PREHEADERS + | LOOPS_HAVE_SIMPLE_LATCHES); already included in LOOPS_NORMAL.
So it is. Thanks!TYVM for the review - I've posted a v2 at https://gcc.gnu.org/ml/gcc-patches/2015-06/msg01355.html .
Cheers, Alan
