Re: We're out of tree codes; now what?

[Dirk Mueller <dmueller at suse dot de>]

On Monday, 12. March 2007, Steven Bosscher wrote:

> with a larger-than-necessary tree structure... :-(  (plus, it's not 32
> bits but 64 bits extra on 64 bits hosts...)

Actually, for 64bit hosts the tree code extension would be cost-free due to 
the pointer that has been clutched after the tree code (which causes 4 bytes 
extra unused padding already now). 

An extension to 9 bit whould be carefully benchmarked, I would assume that gcc 
is really inefficient in reading bitfields, and reading 9 bits cannot be 
broken down into byte or short word reading. I can hardly believe though that 
we cannot get rid of some of the tree codes. 255 ought to be enough for 
everyone. 

Actually, is there an easy way to get a profile of which tree codes are used 
how often for a certain testcase?

It seems to me that annotations are only used on gimple stmts, var decls or 
function decls. moving the annotate pointer out of the tree_node seems easy 
(and I'm mostly done with it except that I couldn't find the place which 
needs patching in order to be able to allocate a gimple_stmt node for 
statements). The question if its worthwile depends however on how the usual 
allocation pattern is. 


Dirk