For certain large values of x, expm1() returns a bogus FE_INVALID exception and returns a NaN instead of positive invinity.
Created attachment 2885 [details] test program
$ /tmp/mt_expm1 1e5 errno == 0 fetestexcept() says: FE_INVALID FE_OVERFLOW FE_INEXACT expm1(1.00000000000000000e+05)=nan 0 FE_OVERFLOW nan $ /tmp/mt_expm1 1.00199970127e5 errno == 0 fetestexcept() says: FE_INVALID FE_OVERFLOW FE_INEXACT expm1(1.00199970126999993e+05)=nan 0 FE_OVERFLOW nan $ /tmp/mt_expm1 1.00199970128e5 errno == 0 fetestexcept() says: FE_OVERFLOW FE_INEXACT expm1(1.00199970128000001e+05)=inf 0 FE_OVERFLOW +inf
Comment #2 should have started with thw words: "The first two runs below demonstrate the problem:"
This appears to have been fixed, probably by various changes in 2009 (there is a test in the testsuite of expm1 overflow, which will verify there is no such inappropriate exception, and that test was added in 2009). However, the OVERFLOW exception for expm1 seems to have disappeared on x86, so I've filed bug 13787 for that problem.
Fix documented for man-pages-5.07. commit fdd3fe51793b75b03ab7e48797e11dc46c935f3b Author: Michael Kerrisk <mtk.manpages@gmail.com> Date: Sat May 23 11:56:12 2020 +0200 expm1.3: The bogus invalid floating-point exception bug has been fixed Sometime around 2009 (approximately 2.11) according to https://www.sourceware.org/bugzilla/show_bug.cgi?id=6814. Signed-off-by: Michael Kerrisk <mtk.manpages@gmail.com> diff --git a/man3/expm1.3 b/man3/expm1.3 index 460068a6d..8b60c3453 100644 --- a/man3/expm1.3 +++ b/man3/expm1.3 @@ -153,16 +153,16 @@ for some large negative .I x values (where the function result approaches \-1), .PP -For some large positive -.I x -values, -.BR expm1 () -raises a bogus invalid floating-point exception in addition to the expected -overflow exception, and returns a NaN instead of positive infinity. -.\" FIXME . -.\" Bug raised: http://sources.redhat.com/bugzilla/show_bug.cgi?id=6814 +Before approximately glibc version 2.11, +.\" http://sources.redhat.com/bugzilla/show_bug.cgi?id=6814 .\" e.g., expm1(1e5) through expm1(1.00199970127e5), .\" but not expm1(1.00199970128e5) and beyond. +.BR expm1 () +raised a bogus invalid floating-point exception in addition to the expected +overflow exception, and returned a NaN instead of positive infinity. +for some large positive +.I x +values, .PP Before version 2.11, .\" It looks like the fix was in 2.11, or possibly 2.12.