Re: [new-regalloc-branch] patch: subreg things 2/2

[Daniel Berlin <dan at cgsoftware dot com>]

Michael Matz <matzmich@cs.tu-berlin.de> writes:

> Hi,
> 
> On 21 Jun 2001, Daniel Berlin wrote:
> 
>> Michael,
>> Just FYI, I plan on getting rid of the conservative coalescing part.
>> Until optimistic coalescing is fully implemented, this will mainly
>> consist of commenting out the call to the conservative test, and
>> freezing of moves.
> 
> Note also, that some code in ra.c canīt correctly handle spilled nodes, if
> they result from coalescing (i.e. something which is not possible right
> now, which explains this fact).

Right.
I know this. I worked on it to 
> 
> Iīve also read a paper about optimistic coalescing (and then added the
> TODO item ;-)), but this involves some things and bookkeeping (e.g.
> breaking coalesced webs, if that coalescing results in spilling
> afterwards). 

Well, the undo of coalescing is more to prevent spilling than anything
else. It's not strictly necessary.

WE also almost already have the book keeping.
We have it for one of the two nodes that are coalesced together, since
we just follow alias pointers.

Adding a compressed bitmap representing the original edges for each
web shouldn't take up that much memory.
Of course, it *shouldn't*, but it probably will.


> I would like to have this (switching off conservative
> coalescing) optional, may be based on a -fra-xxx flag, so performance
> impact of this is easier measurable.

I've been doing testing by just commenting out the conservative test,
and freeze_moves (which means things get put on the freeze worklist,
then just moved to the ismplify worklist, of course. Not good for
speed measurements of the allocator, but okay for measuring the
performance of generated code).

> 
>> This has also been shown in the paper on optimistic register
>> coalescing, among other papers.
> 
> If this is the same paper I was refering above (I donīt have the title
> handy here at university), then this (the betterness as iterated
> coalescing) was reached by the man above, the reconsidering of coalscing
> if that rsulted in spilling.  This is also, what I want to implement one
> of those days.
> 
> 
> Ciao,
> Michael.

-- 
"One time a cop pulled me over for running a stop sign.  He said,
"Didn't you see the stop sign?"  I said, "Yeah, but I don't
believe everything I read."
"-Steven Wright