llvm-project

86779ac9 - [AMDGPU][GlobalISel] Fix i64-to-i16 clamp combine on inverted bounds (#216306)

[AMDGPU][GlobalISel] Fix i64-to-i16 clamp combine on inverted bounds (#216306) `abs(Cmp2 - Cmp1)` check couldn't tell inverted bounds `(Lo > Hi)` from valid ones, letting the combine build a med3 that clamped to the wrong range instead of the constant smin/smax result