This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: PATCH: remove bb_ann_d's num_preds field


On Wed, 2004-09-29 at 14:29, Jeffrey A Law wrote:
> On Wed, 2004-09-29 at 12:20, Andrew MacLeod wrote:
> 
> > Immediate_uses will give you direct access to the stmt from the operand,
> > so that gets you the result of a PHI node pretty easily too.
> > 
> > Its just a matter of how much memory I can chew up :-).  Its a continual
> > trade off of memory for speed, and the balance is delicate. sigh.  I
> > trying yet another angle now.
> Well, I was thinking about this in terms of the only representation
> for PHIs -- ie, we'd drop PHI nodes as we know them today and
> replace them with a new scheme.

They will have to be associated with the PHI results somehow, so either
the vector of today, or the list of tomorrow :-)   Believe me, I have
seriously considered trying to implement the PHI arguments as a linked
list.  It makes keeping the immediate-use information quicker, as well
as resolves any PHI node resizing issues, but again, more memory :-(.

We do look through all the arguments of a PHI frequently I think, so we
need to be able to traverse the arguments as well.  I have envisioned a
*9* word PHI argument, and it makes me shiver a bit. It would be very
good to work with, but seems like a lot of memory.

*shiver*

Andrew


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]