riastradh, 15 days ago

@jspath55 FYI, the long double tests in t_fpclassify are enabled now on all architectures, whether or not long double is larger than double.

The issue with exp2f may be that the rpi0w CPU doesn't support gradual underflow or is configured in nonstandard flush-to-zero mode -- in IEEE 754 binary32, 2^{-127} is subnormal. Whatever the underlying cause, this should be handled more gracefully. Can you file a PR for that one?

jspath55, 14 days ago

@riastradh Great. I ran the newer t_fpclassify and saw 3 failures where there were 2 before on the 0W. I will write it up.

jspath55, 14 days ago

@riastradh Next question. I reviewed prior lest logs and see skips on the 0W, such as:

tp-start: 1715245975.744968, lib/libc/gen/t_floatunditf, 1
tc-start: 1715245975.745004, floatunditf
tc-end: 1715245975.765321, floatunditf, skipped, Requires long double support
tp-end: 1715245975.770136, lib/libc/gen/t_floatunditf

and

tc-end: 1711963192.337442, scalbnl_zero_pos, skipped, Requires long double support

(this is on #NetBSD 10.0)

Suggestions?

riastradh, 14 days ago

@jspath55 I made those tests run too in HEAD. I think I got all the spuriously skipped long double tests now but let me know if you see any more.

jspath55, 14 days ago

@riastradh I am re-running tests from #57419 where I reported the t_fpclassify + related failures on the 0W and the same cases on a 02W.