intel/compiler: remove branch weight heuristic
Current logic is too simple and helps only a small number of cases, harming more common ones.
As a result of this patch, the compiler chooses SIMD32 shaders more frequently.
Jason Ekstrand described the problem here: Does subgroup/wave size matter?
In Intel perf CI it improves FPS in:
| gen9 | gen11 | |
|---|---|---|
| gfxbench5 alu2 | 21.92% | 23.7% |
| synmark OglShMapVsm | 3.26% | 4.52% |
| gfxbench5 car chase | 1.34% | 1.32% |
No observed regressions.
In my testing, it also improves FPS in:
| gen9 | gen11 | |
|---|---|---|
| The Talos Principle | 2.9% | (not tested) |
The other 16 games I tested had very minor changes in performance (2/3 positive, but either not significant enough to list here or too noisy).
cc @currojerez