[Bug math/22902] float128 test failures with GCC 8

[joseph at codesourcery dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=22902

--- Comment #3 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
On Tue, 27 Feb 2018, hjl.tools at gmail dot com wrote:

> https://sourceware.org/bugzilla/show_bug.cgi?id=22902
> 
> --- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> ---
> (In reply to joseph@codesourcery.com from comment #1)
> > I suspect a GCC bug.
> > 
> > These functions rely on knowing that libgcc's soft-fp for float128 uses 
> > the x87 rounding mode for 32-bit but the SSE rounding mode for 64-bit - 
> 
> My GCC is configured with -with-fpmath=sse.

I don't think that should make any difference as regards rounding modes.

fenv_private.h redefines libc_feholdexcept_setroundf128 and 
libc_feupdateenv_testf128 to use the _sse functions in the __x86_64__ 
case.  When defined, float128_private.h redefines 
libc_feholdexcept_setroundl and libc_feupdateenv_testl to use the *f128 
versions.  If not defined, the *l names won't be redefined - but those 
names for i386 also use x87 unconditionally.  And libgcc's 
32/sfp-machine.h always uses fnstcw to get the rounding mode, i.e. always 
uses the x87 rounding mode for float128, which is what glibc expects for 
32-bit for float128.

Maybe there's an issue with exceptions rather than rounding modes?  If so, 
the 32-bit, __SSE_MATH__ case in glibc would need to have a version of 
those two functions that clears SSE exceptions as well as 387 exceptions 
in the first function, and tests/restores both in the second function.

-- 
You are receiving this mail because:
You are on the CC list for the bug.