From Bugzilla Helper:
User-Agent: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.7.3) Gecko/20040913

Description of problem:
On AMD64: after a fesetround(FE_UPWARD) (ISO C99 standard function for changing IEEE arithmetics rounding mode to round-to-plus-infinity, i.e. round to the nearest greater or equal representable value), pow() on certain operands will result in a segmentation fault, and exp() on some operands will yield negative results (the exponential of a real number is never negative).

Version-Release number of selected component (if applicable):
2.3.3

How reproducible:
Always

Steps to Reproduce:
1) x=79.252413924118542354335659183561801910400390625, y=0.333333299999999999041477849459624849259853363037109375
2) fesetround(FE_UPWARD);
3a) pow(x, y)
3b) exp(log(x)*y)
(see attachment)
  

Actual Results:  3a) Segmentation fault.
3b) result = -0.246008

Expected Results:  Both programs should have given something around 4.295405, as it happens on IA32.

Additional info:

The glibc code (e_pow.c) that is used for some computations states, in a comment:
"Assumption: Machine arithmetic operations are performed in round to nearest
mode of IEEE 754 standard."

The bug has been filed in glibc bugzilla (http://sources.redhat.com/bugzilla/show_bug.cgi?id=763).

This bug has resulted in annoying crashes in scientific code doing interval arithmetic.