https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126049
--- Comment #19 from Icenowy Zheng <uwu at icenowy dot me> --- On GCC 15 the generated code isn't well optimized -- a scalar fallback codepath exists, and is used for data not able to fill a single vector register (for float, array elements less than vlenb/4). To investigate the problem on GCC 15, the count of NaNs passed needs to be increased to be more than vlenb/4 (I tried 10 NaNs on qemu-user (which by default has vlenb = 16) and it worked).
