[Ping] [patch 0/3] New macro PREFERRED_RENAME_CLASS
Eric Botcazou
ebotcazou@adacore.com
Mon Dec 13 13:10:00 GMT 2010
> Bootstrap and regression tested on x86_64-linux. Committed. Thanks a
> lot for your patient review.
You're welcome. However, I've pondered some more on this and I now think that
we went a little too far. The problem is the sorting of DU chains. Although
this isn't explicit in the pass (lack of general comment at the top of the
file among other things), it uses a round-robin algorithm to rename the
registers (the 'tick' business); sorting the chains damages the algorithm.
So I think that we cannot sort the chains, at least not so abruptly. The
attached patch removes the sort, but in exchange forces the renaming of
registers that don't belong to the preferred class to registers that do, even
though they were used recently. It also adds the missing general comment.
Could you give it a whirl on your favorite set of tests and see how it fares?
--
Eric Botcazou
-------------- next part --------------
A non-text attachment was scrubbed...
Name: regrename.diff
Type: text/x-diff
Size: 11968 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20101213/9af1827e/attachment.bin>
More information about the Gcc-patches
mailing list