[llvm] [AMDGPU] Fix isKnownNeverNaN for FMIN_LEGACY/FMAX_LEGACY (PR #216338)

[Matt Arsenault via llvm-commits]

================
@@ -9248,6 +9248,407 @@ define double @v_test_nnan_input_fmed3_r_i_i_f64_maximum_minimum(double %a) {
   ret double %med
 }
 
+; FMIN_LEGACY passes an sNaN operand through bit-for-bit, so the canonicalize
+; must not be folded away.
+define amdgpu_kernel void @canonicalize_fmin_legacy(ptr addrspace(1) %out, float %a, float %b) #3 {
----------------
arsenm wrote:

Avoid using kernel + load/store

https://github.com/llvm/llvm-project/pull/216338