https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116776
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2024-09-19
Ever confirmed|0 |1
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's different, only the outermost condition load is hoisted (the inner ones
could trap!) and unswitching would only iteratively hoist (it unswitches
on the x->dims[1] condition), it does not consider unswitching on a
whole sub-CFG representing a not short-circuited && condition.
An intermediate step would be to transform this into
if (shouldthischange(x, 0, 0) && shouldthischange(y, 0, 0) &&
shouldthischange(z, 0, 0))
flag = 1;
else
flag = 0;
for (...)
if (flag)
...
loop invariant motion can hoist some conditions but in reality it looks
at invariant PHIs, hoisting them as ?: and rely on CFG cleanup to elide
empty ifs. In principle LIM can compute all invariant info but of course
it doesn't implement hoisting a sub-CFG or re-materializing a sub-CFG
to compute a condition argument.
