[trunk] patch to remove REG_NO_CONFLICT_BLOCKS
Vladimir Makarov
vmakarov@redhat.com
Wed Apr 23 21:11:00 GMT 2008
Steven Bosscher wrote:
> On Wed, Apr 23, 2008 at 5:34 PM, Vladimir Makarov <vmakarov@redhat.com> wrote:
>
>> It would be nice to know you current position what are you going to do with
>> the local allocator. As I wrote personally I don't care about the old RA
>> anymore but I see other people do really care about local-alloc.
>>
>
> Does that mean you will also try to make IRA work at -O0? Otherwise
> we will still end up with (at least remnants of) the old allocator and
> IRA side-by-side...
>
>
IRA works for -O0 but has the same problem as the global. I'll have
some ideas how to make more compact conflict info. Currently IRA uses
bit vector or vector of conflict allocnos (during IR flattening vector
of allocnos may have even duplications) whatever is more profitable.
The problem is now that there is a short intermediate step when conflict
bit matrix exists. To solve the problem I should somehow eliminate the
intermediate step.
I am going to start this work on next week. Now I am working on
spill/restore placement optimizations (as I wrote earlier the code was
originally supposed to be temporary) and on teaching the reload how to
use saved hard registers for the reloads in holes between the
spill/restore code. Currently I am stuck in the reload problems.
More information about the Gcc-patches
mailing list