C99 Pragmas

Support has been added for the three C99 standard pragmas:

     #pragma STDC FENV_ACCESS         (defaults to OFF)
     #pragma STDC FP_CONTRACT         (defaults to OFF)
     #pragma STDC CX_LIMITED_RANGE    (defaults to OFF)

Only FENV_ACCESS currently has any effect; with it off, the compiler can assume that floating-point operations have no side-effects, and ignores any possibility of exceptions or traps. With it on, care is taken to ensure that all arithmetic occurs (as if) at run-time. For example

         double fred(void)
         {
             return 1.0 / 3.0;
         }

will compile to

         fred
                 LDFD    f0, =0.3333333333333333
                 MOV     pc, lr

but

         #include 
         
         double fred(void)
         {
             #pragma STDC FENV_ACCESS ON
             return 1.0 / 3.0;
         }

will compile to

         fred
                 MVFD    f0, #1
                 DVFD    f0, f0, #3
                 MOV     pc, lr

to ensure that the inexact flag gets set (and with a view to rounding mode altering the result). The compiler will still fold expressions that can be determined to raise no exceptions (eg 1.0 / 2.0). Compile-time evaluation can also be forced thus:

         double fred(void)
         {
             #pragma STDC FENV_ACCESS ON
             static const double one_third = 1.0 / 3.0;
             return one_third;
         }

Other optimisations, particularly CSE, are effectively inhibited on floating-point arithmetic if FENV_ACCESS is on.

In C90 mode, the compiler functions as before, and only considers overflow, invalid operation and divide by zero exceptions as significant. Thus 1.0 / 0.0 would be performed at run-time, but 1.0 / 3.0 would not. For other optimisations, behaviour is somewhere in-between the two C99 states.