math_error(7) still describes glibc 2.8

The math_error(7) man page on this Ubuntu 24.04 box has a NOTES section that begins:

The math_errhandling identifier specified by C99 and POSIX.1 is not supported by glibc. ... The current (glibc 2.8) situation under glibc is messy.

glibc 2.8 shipped in 2008. Michael Kerrisk added that paragraph on 2008-07-30, and the page it sits on is dated 2023-10-31 (Linux man-pages 6.7). Nobody has touched the text in the eighteen years between.

The binary disagrees. On glibc 2.39, math_errhandling is defined and expands to both mechanisms:

printf("math_errhandling = %d\n", math_errhandling);  /* 3 */

I wrote a small program that zeroes errno, clears the floating-point exception flags, calls a math function, and prints what changed. glibc 2.39, gcc 13:

log      x=-1        r=-nan         errno=33  exc=FE_INVALID
sqrt     x=-1        r=-nan         errno=33  exc=FE_INVALID
log      x=0         r=-inf         errno=34  exc=FE_DIVBYZERO
exp      x=1000      r=inf          errno=34  exc=FE_OVERFLOW
exp      x=-1000     r=0            errno=34  exc=FE_UNDERFLOW

errno 33 is EDOM, 34 is ERANGE. Domain error, pole error, overflow, underflow: each one sets errno and raises the matching exception. Messy was a fair word for 2.8. It is not a fair word for this.

glibc's own header agrees. math.h in 2.39 says "By default all math functions support both errno and exception handling," and defines math_errhandling as MATH_ERRNO | MATH_ERREXCEPT. The macro landed in glibc 2.11 (commit 7b943af6, August 2009), a year after the man page paragraph.

musl does something different. It doesn't compute math_errhandling; it hardcodes it:

#define math_errhandling 2

That is MATH_ERREXCEPT, always. Under -ffast-math, glibc changes the macro to 0 (its own comment calls that nonconforming) while musl leaves it at 2. And musl never sets errno in a math function: none of the 380 .c files under src/math/ in the 1.2.4 tarball mention errno. I built musl 1.2.4 and ran the same program:

log      x=-1        r=-nan         errno=0   exc=FE_INVALID
sqrt     x=-1        r=-nan         errno=0   exc=FE_INVALID
log      x=0         r=-inf         errno=0   exc=FE_DIVBYZERO
exp      x=1000      r=inf          errno=0   exc=FE_OVERFLOW
exp      x=-1000     r=0            errno=0   exc=FE_UNDERFLOW

Same exceptions, errno untouched. The recipe the man page hands you near the top, set errno to zero, call feclearexcept, treat either being nonzero as an error, works on glibc and has a dead branch on musl. That errno check can never fire there.

The 2008 commit is sharper than ordinary bit-rot. The same diff that added the paragraph deleted a comment: "I've tested glibc 2.3.3 and glibc 2.8, and both seem to support both mechanisms. A quick look at the glibc source code suggests that support goes back to glibc 2.1 at least." The rewrite turned "glibc supports both" into "glibc doesn't support the identifier and its behavior is messy." glibc added the identifier the next year.

The portable answer is the one the page reaches a few paragraphs later: check the inputs before the call and the outputs after. NaN, zero and out-of-domain arguments cover domain and pole errors; HUGE_VAL covers overflow; a zero result where the function cannot legitimately produce one covers underflow. That is more code, and it needs a per-function idea of what a bad answer looks like. It also doesn't depend on a macro two libcs disagree about.

sebsite's note is where I found the man page paragraph. The builds above are mine.