Re: Will tree-ssa be GCC 3.5?

[law at redhat dot com]

In message <10312020203.AA01836@vlsi1.ultra.nyu.edu>, Richard Kenner writes:
 >There are conflicting goals here.  In order to do the sorts of
 >optimizations best handled at higher level (e.g., Fortran loop
 >optimizers), you need to keep the tree at the highest level possible
 >for as long as possible during the optimization process.  But that
 >means the RTL generation will produce more things for RTL optimizers
 >to optimize.  I admittedly don't follow the tree-ssa project extremely
 >closely, but from what I see, it is far from clear how this dilemma
 >will be resolved and I suspect it will be an ongoing process for quite
 >some time.  So I think any prediction of removing a significant number
 >(and certainly *most*) of the RTL optimizers is premature.  Indeed, I
 >think that viewing that as a goal may get in the way of making the
 >best effort at the important high-level optimizations.
Well, we already have cases where the tree-ssa creates initial RTL which is
*better* than the RTL you would have after running the first cse pass.

We're certainly not there for all cases, but that's definitely where we're
going.

Some of this happens because the algorithms we can use on the SSA form are
simply more powerful than those we can use at the RTL level.  Some of this
happens because we can make transformations which are exceedingly difficult
to do at the RTL level because we have more information about the underlying
objects and types.

Please don't make statements about the status or goals of the tree-ssa
work if you aren't following it closely.

jeff