lp_test_blend could be faster
Or at least could probably be chunked up. For this s390x job it took 222 seconds, of which 106 was lp_test_blend; granted that's in parallel with the rest of the tests, but the next slowest took 29 seconds so that's still 70+ seconds waiting on one test.
It would probably be sufficient to split lp_test_blend into five pieces, one for each blendfunc, and arrange for meson to invoke them as five different tests.