fenv_access 

  • [アーティクル]

Disables (ON) or enables (OFF) optimizations that could change flag tests and mode changes.

#pragma fenv_access [ON | OFF]

Remarks

By default, fenv_access is OFF.

For more information on floating-point behavior, see /fp (Specify Floating-Point Behavior).

The kinds of optimizations that are subject to fenv_access are:

  • Global common subexpression elimination

  • Code motion

  • Constant folding

Other floating-point pragmas include:

Example

// pragma_directive_fenv_access_x86.cpp
// compile with: /O2
// processor: x86
#include <stdio.h>
#include <float.h> 
#include <errno.h>
#pragma fenv_access (on)

int main() {
   double z, b = 0.1, t = 0.1;
   unsigned int currentControl;
   errno_t err;

   err = _controlfp_s(&currentControl, _PC_24, _MCW_PC);
   if (err != 0) {
      printf_s("The function _controlfp_s failed!\n");
      return -1;
   }
   z = b * t;
   printf_s ("out=%.15e\n",z);
}

Output

out=9.999999776482582e-003

The following sample is for compiler producing output files for Itanium processors. /fp:precise keeps the intermediate results in extended precision where values greater than FLT_MAX (3.402823466e+38F) can be calculated and as a result of that sum will have 1.0 result, as it should if manually calculated. /fp:strict keeps intermediate results in their source precision (float) so the first addition produces infinity, which is kept throughout the expression.

// pragma_directive_fenv_access_IPF.cpp
// compile with: /O2 /fp:precise
// processor: IPF
// compiling with /fp:precise prints 1.0F
// compile with /fp:strict to print infinity

#include <stdio.h>
float arr[5] = {3.402823465e+38F, 
               3.402823462e+38F,
               3.402823464e+38F,
               3.402823463e+38F,
               1.0F};

int main() {
   float sum = 0;
   sum = arr[0] + arr[1] - arr[2] - arr[3] + arr[4];
   printf_s("%f\n", sum);
}

Output

1.000000

When commenting out #pragma fenv_access (on) from the previous sample, note that the output is different because the compiler does compile-time evaluation, which does not use the control mode.

// pragma_directive_fenv_access_2.cpp
// compile with: /O2
#include <stdio.h>
#include <float.h> 

int main() {
   double z, b = 0.1, t = 0.1;
   unsigned int currentControl;
   errno_t err;

   err = _controlfp_s(&currentControl, _PC_24, _MCW_PC);
   if (err != 0) {
      printf_s("The function _controlfp_s failed!\n");
      return -1;
   }
   z = b * t;
   printf_s ("out=%.15e\n",z);
}

Output

out=1.000000000000000e-002

See Also

Reference

Pragma Directives