[RFC/RFT] Should we hoist FP constants on x87?

[Grigory Zagorodnev]

Roger Sayle wrote:
> Wow!  Hopefully your POV-ray results are impressive enough to prompt
> someone with SPEC to evaluate it on IA-32.  For the curious the original

Hi!
My cpu2000 run of gcc 4.2.0 20060216 showed no significant impact of 
this patch on geomeaned SPEC marks. However few individual benchmarks 
have got the real performance gain:

     191.fma3d	1.5%-7.2%
     168.wupwise	average is 1%
     171.swim	average is 1%
     300.twolf	gain is shady: +4.5% on Prescott, -3.4% on Yonah

Below is the full listing of performance impact on ia32 (Prescott, 
Yonah) and x86_64 (Nocona) at -O2 optimization level:

            Prescott   Nocona   Yonah
164.gzip     0.1%      1.2%    -0.6%
175.vpr      0.7%     -0.3%     0.3%
176.gcc      0.6%     -0.5%    -0.3%
181.mcf      0.6%      1.3%    -0.5%
186.crafty  -0.1%      0.1%     0.6%
197.parser  -0.1%     -0.2%    -1.2%
252.eon      0.0%      1.7%     0.3%
253.perlbmk  0.3%      1.3%     0.9%
254.gap     -0.4%      0.1%     0.7%
255.vortex  -0.1%      0.3%     0.1%
256.bzip2    0.1%      0.1%     1.0%
300.twolf    4.5%     -0.5%    -3.4%
CINT2000     0.5%      0.4%    -0.2%
			
168.wupwise  0.5%      0.1%     1.8%
171.swim     1.6%      1.3%    -0.2%
172.mgrid   -0.2%     -0.2%    -0.3%
173.applu   -1.6%      0.5%     0.0%
177.mesa     0.5%      1.6%     0.5%
178.galgel  -0.6%     -2.5%     0.9%
179.art     -1.1%      0.1%    -0.4%
183.equake   0.4%      1.8%    -0.8%
187.facerec  0.9%     -0.7%    -1.5%
188.ammp    -0.3%      0.1%     0.0%
189.lucas    0.0%      0.7%    -0.6%
191.fma3d    1.5%      7.2%     3.9%
200.sixtrack 1.5%      0.0%    -0.4%
301.apsi    -0.7%      0.2%    -0.5%
CFP2000      0.2%      0.7%     0.2%

- Grigory