On 13-06-19 1:23 AM, Wei Mi wrote:
Ping.
On Wed, Jun 12, 2013 at 2:44 PM, Wei Mi <w...@google.com> wrote:
Hi,
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57518
pr57518 happened because update_equiv_regs in IRA marked a reg
equivalent with a mem, lowered its mem_cost in scan_one_insn, set
NO_REGS to its rclass, but didn't consider the reg was used in
paradoxical subreg which prevented the reg from being replaced by mem
in LRA phase.
This patch is to check whether a reg is used in a paradoxical subreg
in update_equiv_regs before reg is set as equivalent to a mem.
bootstrap and regression test on x86_64-linux-gnu ok. Is it ok for
trunk and gcc-4.8 branch?
Thanks for working on this PR, Wei, and sorry for the delay with the
answer (I was on vacation).
In general, the PR analysis and the proposed solution looks ok. I only
worry that you are adding additional full scan of all RTL code. It
might add 0.5% to GCC compilation time if data cache is rewritten (which
will happen for moderate size or big functions). It would be nice to do
it on some other existing RTL traversing. Unfortunately, this info is
calculated later (reg_max_width in reload or biggest_mode in LRA). I am
in doubt that other solutions I see now are better:
o calculate this info in regstat_... function and store it in reg_info_p
o calculate it with update_equiv_regs and use it for invalidation the
equiv info later
The first one increases reg_info_p footprint and calculation is done
many times although it is used once.
The second one results in complicated code.
So I think the current patch is ok to commit.
Thanks, again.
