On 12-09-28 1:48 PM, Andi Kleen wrote:
Steven Bosscher <stevenb....@gmail.com> writes:
On Fri, Sep 28, 2012 at 12:56 AM, Vladimir Makarov <vmaka...@redhat.com> wrote:
Any comments and proposals are appreciated. Even if GCC community
decides that it is too late to submit it to gcc4.8, the earlier reviews
are always useful.
I would like to see some benchmark numbers, both for code quality and
compile time impact for the most notorious compile time hog PRs for
large routines where IRA performs poorly (e.g. PR54146, PR26854).
I would be interested in some numbers how much the new XMM spilling
helps on x86 and how it affects code size.
I have some results which I got after implementation of spilling into
SSE regs:
Average code size change: Corei7 Bulldozer
SPECInt 32-bit -0.15% -0.14%
SPECFP 32-bit -0.36% -0.24%
SPECInt 64-bit -0.03% -0.07%
SPECFP 64-bit -0.11% -0.09%
Rate change: Corei7 Bulldozer
SPECInt 32-bit +0.6% -1.2%
SPECFP 32-bit +0.3% 0%
SPECInt 64-bit 0% 0%
SPECFP 64-bit 0% 0%
I used -O3 -mtune=corei7 -march=corei7 for Corei7 and -O3
-mtune=bdver1 -march=bdver1 for Bulldozer processor. Additionally I
enabled inter unit moves for Bulldozer when the optimization works
because without this spilling general regs into SSE regs is not
possible.
