Re: Clear some of pointers in SSA_NAME

[Jeffrey A Law <law at redhat dot com>]

On Sun, 2004-09-19 at 09:04, Jan Hubicka wrote:
> Hi,
> release_ssa_name don't clear the released SSA name keeping dead
> statements alive from GGC point of view.  This patch fixes everything
> except for SSA_NAME_VAR that is still accessed for released sse name in
> tree-into-ssa.  See my other mail.
> 
> This patch saves about 30% of memory usage on rt-3.4.ii testcase I got
> somewhere from testsuite.  This is kind of cheat as the testcase merely
> use hords of templates to construct mostly empty fucntion and dce1 pass
> kills most of the statements before we happen to consume the memory
> later.  But it shows that keeping dead pointers around is not the
> coolest idea on the planet.
> 
> Honza
> bootstrapped/regtested ppc-linux, OK?
> 
> 2004-09-19  Jan Hubicka  <jh@suse.cz>
> 	* tree-ssanames.c (release_ssa_name): Zero out SSA_NAME datastructure.
I've been looking at similar stuff.  Though my tests haven't shown
this to be a problem at all.  In fact, a similar change I tried had
absolutely _no_ effect on Gerald's testcase.

I'd be very curious to know which of those pointers is keeping data
live -- mostly because it may point to a pass that is not cleaning
up things as well as it should.

On a style note, I'd probably suggest using memset to wipe the
entire node, then put back the 2-3 fields we actually need, ie
the tree code, the ssa version # associated with the node and
(for a little while longer) the SSA_NAME_VAR associated with
the node.

[ Also note that once you wipe the node en-masse in release_ssa_name,
  you no longer have to do so in the code which recycles them. ]


jeff