https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112746
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
Summary|Missed optimization for |Missed optimization for
|redundancy computation |exact division with
|elimination (fre1(tree) for |multi-use addition chain
|global variable) |
Last reconfirmed| |2023-11-28
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue with test2 is that with value-numbering we have
(n.3_1 * 2 + n.3_1) / n.3_1
and that does not simplify. That is, we do not simplify 2 * n + n to
3 * n as in general that wouldn't be profitable if 2 * n is live after
the use is elided (and it is live since it's stored to 'b').
Which means we ask for (n.3_1 * 2 + n.3_1) / n.3_1 which we currently
cannot simplify to a constant. Handling cases like this in match.pd
feels wrong.
Note that with -fwrapv the optimization wouldn't be valid.
Other passes face the same issue, 'b' keeps n*2 live so an add
looks cheaper than to compute n*3 for the division. We're not
anticipating the division here.
But:
_3 = n.3_1 + _2;
_4 = _3 / n.3_1;
could be simplified to
_5 = _2 / n.3_1;
_4 = _5 + 1;
if we know _2 is a multiple of n.3_1 which then could be simplified as well.
Note that all passes doing analyses on addition chains also stop at the
multi-use, so we'd need to improve at that point somehow.