Created attachment 587376 [details]
Simple test code for pow()

Description of problem:
Some functions from libm library (in glibc) return unpredictable result in non default rounding mode. Functions are: exp, pow etc. 


Version-Release number of selected component (if applicable): In RHEL 5.x 6.x, Fedora 16


How reproducible:
Always


Steps to Reproduce:
1.Compile simple code 

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <fenv.h>

static int modes[4] = { FE_TONEAREST, FE_TOWARDZERO, FE_DOWNWARD, FE_UPWARD };

int main (int argc, char *argv[])
{
	int i;
	char *end;
	double x = strtod (argv[1], &end);
	if (*end != '\0')
		exit (EXIT_FAILURE);

	for (i = 0; i < 4; i++)
        {
		double y;
		fesetround (modes[i]);
		y = pow (x,-0.34);
		printf ("pow(%.17g,-0.34) = %.17g\n", x, y);
        }
	return 0;
}

with gcc: "gcc -lm -o round_mode_test round_mode_test.c"
2. Run test code with some input number
$ ./round_mode_test 2
pow(2,-0.34) = 0.79004131186337712
pow(2,-0.34) = 0.50032986487121911
pow(2,-0.34) = 0.50032986487121911
pow(2,-0.34) = 0.79004131186337723
$ ./round_mode_test 0.9
pow(0.90000000000000002,-0.34) = 1.0364719344786322
pow(0.90000000000000002,-0.34) = 1.036471934478632
pow(0.90000000000000002,-0.34) = 1.036471934478632
pow(0.90000000000000002,-0.34) = 0.0061468530371330907

  
Actual results:
Different result in different rounding mode

Expected results:
Same results

Additional info:
This is glibc bug http://sources.redhat.com/bugzilla/show_bug.cgi?format=multiple&id=3976 with bugfix available.