[9fans] more on gcc bug qed [plan or side effect]

[9fans] more on gcc bug qed [plan or side effect]

[ozan s. yigit]

at the expense of wasting more eyeball time on this topic, here are more
details. i don't know nor care where the bug is. the program used to find
the bug has been around for a long time, and is intended to test floating
point implementations. it is available from netlib or search for paranoia
and kahan. [btw, kahan paper on java's floating point is a masterpiece,
and should be required reading for anyone interested in either topic] the
failure of the gcc optimizer was established on a freebsd system with
freshly installed 3.0.4 compiler. similar errors should pop up on any
intel platform eg. a fresh redhat installation with gcc 2.96:

gcc -o paranoia paranoia.c

[paranoia output elided]

The number of  DEFECTs  discovered =         1.
The number of  FLAWs  discovered =           1.

gcc -O2 -o paranoia paranoia.c

[paranoia output elided]

The number of  FAILUREs  encountered =       4.
The number of  SERIOUS DEFECTs  discovered = 4.
The number of  DEFECTs  discovered =         2.
The number of  FLAWs  discovered =           2.

the bug does not exist for the sparc compiler[s]. i have
not tried other architectures.

oz
---
careful - we don't want to learn from this! -- calvin