https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77445
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
CC| |hubicka at gcc dot gnu.org
Target Milestone|--- |7.0
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think r239219 looks bogus in that it considers the outgoing edge frequency
but not the frequency of the path to it. That is, we usually have
if ()
/ \
B A
\ /
X
if ()
\
taken_edge
and we should duplicate 'X' if the path leading to taken_edge (either
through A or B) is to be optimized for speed (or ! for size). The
frequency of the edge out of the condition isn't the only relevant
one (in fact whether we should consider it hot or cold needs to be
adjusted by the incoming path frequency given the now static prediction).
Given r239219 uses ninsns > 1 in the else path it IMHO should use
!optimize_edge_for_size_p or retain a > 1 PARAM use.
