[Bug target/126049] RISC-V auto-vectorization misoptimizes fmax()/fmin() regarding NaN

[uwu at icenowy dot me]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126049

--- Comment #2 from Icenowy Zheng <uwu at icenowy dot me> ---
Created attachment 64901
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=64901&action=edit
A proposed patch changing the initial value to NaN

It seems that the initial value could just be NaN because vfredmin/vfredmax are
implementing IEEE754 minimumNumber/maximumNumber instead of minimum/maximum.