http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60138
--- Comment #3 from Oleg Endo <olegendo at gcc dot gnu.org> --- (In reply to Rich Felker from comment #2) > > No need to move it to TLS, since the FPSCR register is usually already part > > of > > an execution context (i.e. thread) and will be saved and restored > > accordingly. > > This does not help. > The compiler constantly clobbers the FPSCR with the > values read from __fpscr_values[]. If you want to affect a change to the fpu > behavior, you'd have to modify the values in this array, and they are > global, not thread-local, which is wrong. > > If use of __fpscr_values is eliminated, of course this will not matter. > > Right now this is a blocking issue for musl libc having properly a working > sh port (floating point tests fail because fenv is not honored). We could > treat sh as a no-fenv port, but I'd rather not do that since the hardware > does have working fenv and it's just gcc messing it up; also we don't really > support no-fenv targets in a conforming way (we don't provide errno for math > functions) so I'd much rather have fenv. It's not a 5 minute job. I'll see what I can do for the next GCC version (after 4.9). However, it will most likely not be backported to 4.8 or 4.9.
