[Bug tree-optimization/48295] Incorrect code generated with dynamic floating point rounding mode switches

["joseph at codesourcery dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48295

--- Comment #4 from joseph at codesourcery dot com <joseph at codesourcery dot com> 2011-03-28 10:50:44 UTC ---
On Mon, 28 Mar 2011, rguenth at gcc dot gnu.org wrote:

> Btw, your testcase would be kindof invalid as you are not using the
> documented standard way of accessing fenv but using inline-asm (and
> we don't have a special clobber that tells GCC you touched the FP
> control word).

You mark the asm as reading and clobbering fpcr - that register name 
already exists in GCC, and it's a bug that glibc's fpu_control.h doesn't 
use it, though I suppose you'd need to add the "mxcsr" name as well.  
When -frounding-math is actually properly implemented it ought to be 
taught about target-specific registers representing rounding modes (and 
likewise exception status) so it knows how asms interact with the 
floating-point state.  (Non-const, non-pure function calls will also need 
to be presumed to interact with the state in unknown ways since they may 
end up calling fenv.h functions.)