Re: [PATCH] PR35371 GCSE loses track of REG_POINTER attribute

[Jeff Law <law@redhat.com>]

Peter Bergner wrote:
> On Mon, 2008-03-03 at 12:15 -0700, Jeff Law wrote:
>>> Do you mean fix it up and then call it from more than just CSE?
>>> Currently, the only call to reg_scan() isn't in a location that
>>> will help me.
>> No.  I mean make it smarter.  If you read the code it's amazingly
>> simplistic and punts propagation of REG_POINTER for any pseudo
>> that is set more than once.
>>
>> It shouldn't be terribly difficult to build a simple propagation
>> engine that handles multiple sets.
> 
> Sorry, making it "smarter" is what I meant by "fix it up".
> My problem with it, as I mentioned in my previous note, is that
> the only location it is currently called doesn't help me.
> I guess what I was asking was there shouldn't be a problem
> with me calling it from another location, correct?
Ah.  A misunderstanding on my part.


I do recall some problems with passes substituting a pseudo
without REG_POINTER set for a register with REG_POINTER set,
so if you run it early, you might lose some REG_POINTER
attributes.

That's not a fatal problem -- the PA backend already knows
to cope with this issue.  And if the pass is safe, you ought
to be able to just run it twice.


Ideally propagation of REG_POINTER would use both the
assignment to the pseudo and the use of the pseudo to
try and determine if it's a pointer.


Jeff