On Tue, 28 Jun 2005, Mark Mitchell wrote:
Joe Buck wrote:
I don't think we should give the user any such promise, and if we do
give such a promise, we will never catch icc. The main problem is that
we will no longer be able to optimize many loops.
It's entirely possible that I was naive in assuming that this wouldn't have a
big optimization impact. Reiterating my response to Daniel, if it is in fact
the case that this is a major loss for optimization, then I would have to
retract my claim.
I've got some notes in to various XLC people to get info if they have
actual numbers.
However, in the meanwhile, the manual actually says it causes very severe
performance
degradation, and that you are better off doing -O2 and
-qnostrict_induction than -O3 + qstrict_induction, if it makes your code
work.
(which means it says you are better off turning off high level loop opts
like interchange/fusion/etc and still letting lower level stuff make the
assumption, than you are turning off the assumption and turning on high
level loop opts).
I imagine it very badly degrades your ability to move things out of loops,
perform any kind of cache transformations, etc.
and it's just so common to write signed loops without caring about
overflow that i'd guess you lose 90% of your opportunities to do anything
useful or cool with the loop.
