Performance: Consider using `float.{Min/Max}Native` instead of comparison+branch logic

[BoyBaykiller]

Hi, just wanted to let you know about places where float.MinNative/float.MaxNative could perhaps be used over the branchful
a < b ? a : b (or similar) logic to maybe get some perf:

https://gist.github.com/BoyBaykiller/78738c2e26770f19e954e2361bd49198

---

To give some context, I was prototyping an optimization in RyuJIT for patterns like:

float Min(float a, float b)
{
    return a < b ? a : b;
}

where we currently emit a compare+branch:

G_M000_IG02: 
       vucomiss xmm1, xmm0
       jbe      SHORT G_M000_IG04
 
G_M000_IG03:
       ret      
 
G_M000_IG04:
       vmovaps  xmm0, xmm1
 
G_M000_IG05: 
       ret      

to instead use the native x64-specific vminss/vmaxss instruction:

G_M21498_IG02: 
       vminss   xmm0, xmm0, xmm1

G_M21498_IG03:
       ret   

And the link above shows the automated assembly differences it produced in methods from this project.

Note that my prototyped optimization won't make it into the JIT.
float.{Min/Max}Native is the way to get good perf accross all platforms, assuming potentially different edge-case handling can be tolerated.
On x64 it will use the same vminss/vmaxss

[RossNordby]

Thanks for the heads up!