self compile problems with SSA simplifier

Matthew Fluet Matthew Fluet <fluet@CS.Cornell.EDU>
Wed, 14 Nov 2001 21:46:06 -0500 (EST)


> It is interesting that the omission of globalizing constants didn't
> hurt any other benchmarks.

Yes, that is somewhat counter-intuitive.

> > Check out the diff for once.fun that I just checked in.  See if actually
> > globalizing constants helps. ;)
> 
> BTW, the slight slowness in SSA optimization was with a MLton built by
> an old stable version, so was not impacted by the once bug.

O.k.  I couldn't tell from your e-mail which G? you were using.
But, still, the input program to each of the SSA passes would be different
considering that constant propagation was acting funky.

> My guess is that we just need some tuning.  I'd say we're in quite
> good shape for a first whack.

Yes.  I'm sure I left plenty of places where Vector and List operations on
the function and block sets could be different since we don't care about
keeping the order of the functions.  (I tried making the output of a pass
look the same as the input, just so that diffs between pre and post were
useful.  But, as we need less debugging, that's less of an issue.)