On Thu, Oct 16, 2014 at 1:42 PM, Martin Liška <mli...@suse.cz> wrote:
> On 10/16/2014 01:31 PM, Richard Biener wrote:
>>
>> On Wed, Oct 15, 2014 at 6:26 PM, Martin Liška <mli...@suse.cz> wrote:
>>>
>>> Hello.
>>>
>>> Following patch introduces a new class called callgraph_annotation. Idea
>>> behind the patch is to provide a generic interface one can use to
>>> register
>>> custom info related to a cgraph_node. As you know, symbol_table provides
>>> hooks for creation, deletion and duplication of a cgraph_node. If you
>>> have a
>>> pass, you need to handle all these hooks and store custom data in your
>>> data
>>> structure.
>>>
>>> As an example, after discussion with Martin, I chose usage in ipa-prop.h:
>>>
>>> data structure:
>>> vec<ipa_node_params> ipa_node_params_vector
>>>
>>> if the pass handles an event, following chunk is executed:
>>> if (ipa_node_params_vector.length () <= (unsigned)
>>> symtab->cgraph_max_uid)
>>>      ipa_node_params_vector.safe_grow_cleared (symtab->cgraph_max_uid +
>>> 1);
>>>
>>> The problem is that you can have sparse UIDs of cgraph_nodes and every
>>> time
>>> you have to allocate a vector of size equal to cgraph_max_uid.
>>>
>>> As a replacement, I implemented first version of cgraph_annotation that
>>> internally uses hash_map<cgraph_unique_identifier, T>.
>>> Every time a node is deleted, we remove corresponding data associated to
>>> the
>>> node.
>>>
>>> What do you think about it?
>>
>>
>> I don't like "generic annotation" facilities at all.  Would it be possible
>> to make cgraph UIDs not sparse?  (keep a free-list of cgraph nodes
>> with UID < cgraph_max_uid, only really free nodes at the end)
>> Using a different data structure than a vector indexed by cgraph UID
>> should also be easily possible (a map from UID to data, hash_map <int,
>> T>).
>
>
> Hello.
>
> If I recall correctly, we recycle cgraph_nodes and it's possible that an UID
> is given to different nodes:
> symbol_table::allocate_cgraph_symbol (void). Such uid is problematic from
> perspective that it cannot be used as a index to a vector.
>
> It was also Honza's note that one can choose inner implementation of such
> annotation class. We can implement both sparse (hash_map) and consecutive
> vector data structure.
>
> According to first numbers I was given, Inkscape allocates about ~64k
> cgraph_nodes in WPA. After function merging is processed, it shrinks to
> about a half. So that, our free list contains the half of nodes. If we use
> consecutive vector, our memory impact is bigger thank necessary.

I don't think there is anything that forces us to retain the original
UID allocation after WPA merging?  So why not compact it?

Richard.

> Martin
>
>
>>
>> Richard.
>>
>>> Thank you,
>>> Martin
>
>

Reply via email to