JIT: Transform floating SELECT(a < b, a, b) to NativeMin(a, b)

[BoyBaykiller]

Example:

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

;; ------ BASE
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      

;; ------ DIFF
G_M21498_IG02: 
       vminss   xmm0, xmm0, xmm1

G_M21498_IG03:
       ret      

Handles all these cases. I compared the codegen of every single one against it to be sure its correct:
https://godbolt.org/z/389jvnfW4

Fix #116802

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR extends CoreCLR JIT if-conversion select optimization to recognize floating-point conditional selects that compute min/max (e.g., a < b ? a : b) and replace them with a native SIMD min/max node on xarch, enabling codegen like vminss/vmaxss instead of compare + branch.

Changes:

• Allow GT_STORE_LCL_VAR / GT_RETURN if-conversion when the value type is floating-point (in addition to integral/I).
• Add TrySelectToMinMax to rewrite eligible floating GT_SELECT nodes into gtNewSimdMinMaxNativeNode (xarch).
• Add an early bailout to avoid producing floating-point GT_SELECT nodes unless they optimize to a non-SELECT form.

[BoyBaykiller]

we call gtNewConditionalNode before which counts as change

[BoyBaykiller]

@MihuBot -nuget

[tannergooding]

As per Discord, I'm not in favor of this specific optimization.

The a < b ? a : b pattern is fairly specific to xarch and is likely a deoptimization on other platforms since they do not have a single instruction variant. Such code is rarely dependent on the "return b for NaN or if both are zero" behavior and so switching to use MinNative(a, b) is almost always preferred, as it accelerates to minss on xarch and to an appropriate single instruction on other platforms.

In general I'd like to push users towards a path of success and accelerating a "bad practice" doesn't help achieve that.