https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77975
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 40936
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40936&action=edit
gcc7-pr77975.patch
Actually I think your patch is fully correct. If next[j] (the phi argument
from the latch edge) is constant in addition to val[j] (the phi argument from
preheader edge), then in the val[j] = get_val_for (next[j], val[j]); we want to
set val[j] = next[j];, the processing of the statements in the loop is then
done in the other get_val_for call - the aval[j] = get_val_for (op[j], val[j]);
So, the first time we evaluate the statements for base of val[j] and second
time for base of next[j] if both are constants. The only thing I've added in
addition to some comment/formatting changes and testcase is a compile time
optimization, MAX_ITERATIONS_TO_TRACK can be very large (I'm surprised we don't
limit also the number of statements between op and the header phi node, there
could be hundreds of thousands of them), but if the phi contains two same
constants, then it is enough to do one iteration of the loop, if two different
constants, then two, all further work is wasted.
