RFA: Fix rtl-optimization/57425

[Joern Rennecke]

Quoting Michael Matz <matz@suse.de>:

> So, first the potentially clobbering write, then the read.  And indeed in
> check_dependence d->exp is the write and x the read that is potentially
> clobbered.

Oops, you are right.  I got confused because what is X in cse.c:invalidate
ends up as d->exp in cse.c:check_dependence, and what is &p->canon_exp in
cse.c:invalidate ends up as X in cse.c:check_dependence.

I was a bit puzzled why we could have a STRICT_LOW_PART for MEM.  Now it
makes more sense.

I also see now that both the read expression from the hash table and the
write address are canonicalized, so if we have a canon_* interface, we
want to use both.

> Generally I would prefer simple interfaces to the query functions,
> dependence problems are hard enough to think about without functions
> needing four arguments.  Does it really save much to not canonicalize the
> mem address for some calls?

I haven't measured it myself, but I suppose it must be there for a  
good reason.
Of course a simpler interface would be nice, however...
When cse encounters a write, it goes through the entire hash table of
expressions and checks dependencies.  So for (extended?) basic blocks with
lots of expressions & writes, we got O(n^2) dependency checks.