This is the mail archive of the
mailing list for the GCC project.
More stats on new reg allocator
- To: <gcc at gcc dot gnu dot org>
- Subject: More stats on new reg allocator
- From: Daniel Berlin <dan at www dot cgsoftware dot com>
- Date: Wed, 24 Jan 2001 17:09:55 -0500 (EST)
So i found some random code that generate spills with the old register
allocator (rsne.c in libI77), and wanted to see how the new allocator
Well, the stats are very encouraging.
(When i say spill, i mean reload couldn't avoid spilling it either, if
you count the ones reload puts back into registers, you have a lot
While we spill at least once (usually at least 3-4 times) per procedure
old allocator, we don't spill at all with the new one.
In fact, in the procedure with the most regs, x_rsne (72 regs need
allocating with -O2), we spill with the old allocator about 30 times, and
use 25 regs, with the new allocator, we don't spill, and use 14 regs.
(For those wondering how i testd to make sure it spills when
necessary, and spills the right stuff, i told the allocator it only
had 3 registers available)