Re: [PATCH][4.4] Add a global DECL_UID -> tree mapping

[Richard Guenther <rguenther@suse.de>]

On Fri, 9 Nov 2007, Richard Guenther wrote:

> This patch adds a global map to look up a decl from its uid.  At the
> moment all places doing this use the referenced_vars () hashtables which
> exist per function and only during a part of the compilation.
> 
> This patch enables making these hashtables bitmaps which allows to
> do unused vars removal more efficient and makes walking referenced vars
> order independed of the hashtable size (the walk is now sorted after
> UID).  See the followup patch.

I am going to go forward with this (and the followup) early next week.

http://gcc.gnu.org/ml/gcc-patches/2007-11/msg00524.html

This will make referenced_vars a bitmap which allows stable traversal
and fixes most of the bootstrap-debug fails if we re-enable
remove_unused_vars before final inlining.

> Bootstrapped and tested on x86_64-unknown-linux-gnu.
>
> For tramp3d the global hashtable peaks at 1M entries which makes its
> cost 8MB on a x86_64 architecture.
> 
> FYI, I was just sitting on this patch for a while now.
> 
> You can also see the single problem:
> 
> +   /* We should never try to re-insert a decl with the same uid.
> +      ???  The C++ frontend breaks this invariant.  Hopefully in a
> +      non-fatal way, so just overwrite the slot in this case.  */
> + #if 0
> +   gcc_assert (!*slot);
> + #endif
> 
> so the C++ frontend does something "interesting" here, which still
> needs investigation (and fixing).

I will not do this though, but file a PR instead and let the C++ FE
maintainers sort this out.

Constructive complaints welcome;  if this is in I am going to
tackle the unexpanded_vars list next, which if made a bitmap will
allow us getting rid of TREE_CHAIN of (hopefully all) decls.

If we do the same for BLOCK_VARS we can get rid of remove_unused_vars
pass and let GC do its work instead.  (remove_unused_vars takes quite
some compile-time on tramp3d, which of course overall pays back)

Thanks,
Richard.