Re: Eliminate write-only variables

[Richard Biener <richard dot guenther at gmail dot com>]

On Tue, Jun 3, 2014 at 6:19 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>> On Mon, Jun 2, 2014 at 8:59 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>> >>
>> >> Yeah, I discussed this with martin today on irc.  For aliasing we'd like to know whether a decl possibly has its address taken. Currently we only trust TREE_ADDRESSABLE for statics - and lto might change those to hidden visibility public :(
>> >>
>> >> So we want deck_referred_to_by_single_function
>> >
>> > OK, I will implement this one in IPA mini-pass.  It is easy.
>> > This property changes by clonning and inlining. What Martin wants to use it for?
>> > (I.e. my plan would be to compute it as last IPA pass making it useless for
>> > IPA analysis&propagation)
>>
>> He doesn't want to use it but we talked and he said he'd have a look.
>>
>> Note that it's important the decls are not refered to by global initializers
>> either.
>>
>> Why is a new pass necessary - can't the IPA reference maintaining code
>> update some flag in the varpool magically?
>
> If the code to add/remove reference was updating the flag, it would became out of
> date as we remove callgraph during the late compilation (we do not keep references for
> functions we already compiled).
> I do not want the flags to be computed before end of IPA queue so they won't become
> out of date as we clone/inline.
>
> We basically need to walk references and check that they are all functions &
> either one given function F or a clones inlined to it.  Assuming that function
> does not have unbounded number of references to given variale, this is
> basically constant time test.
>>
>> >> and deck_may_have_address_taken.
>> >
>> > We currently clear TREE_ADDRESSABLE for statics that have no address taken at
>> > WPA time and that ought to keep the flag cleared at ltrans (I think I even made
>> > testcase for this)
>> >
>> > What I think we could improve is to not consider string functions as ADDR operations
>> > for this analysis, i.e. it is common to only memset to 0.
>> >
>> > How may_have_address_taken would differ here?
>>
>> I want tree.h:may_be_aliased to return false if a decl doesn't have its address
>> taken.  We don't trust TREE_ADDRESSABLE for TREE_PUBLIC/DECL_EXTERNAL
>> decls because other TUs may take the address.  I want the check to be replaced
>> with sth more symtab aware, that is, bring in benefits from LTO (and at the same
>> time be not confused by statics brought public with hidden visibility).
>
>
> I see, are you sure you need to ignore TREE_ADDRESSABLE for PUBLIC/EXTERNAL?
> I belive frontends are resposible to set them for all data that may have address
> taken (even externally) and IPA code maintains it - we clear the flagonly for
> variables that are static.

Not sure - we've always not trusted this flag on public/externals.  Probably
because there could be aliases to them that have their address taken?
(well, that's true for locals as well ...)

That said, having a varpool way of querying whether a decl may be aliased
by a pointer would be nice (with bonus points of handling the alias case).

> Or we can just set the flag to true for externally visible vars ealry in IPA
> queue if this is false.

Well, let me do a simple check (remove the restriction and bootstrap/test).

Richard.

> Honza
>>
>> Richard.
>>
>> > Honza