Re: Tru64, egcs, and optimization

[Brian Gough <bjg at network-theory dot co dot uk>]

That is good to know. GSL might make a good compiler test, if nothing
else ;-)

I would be interested to hear from anyone on the list who is familiar
with the Vegas algorithm and can explain why monte/vegas.c does not
integrate the function f2() properly with double-precision rounding,
but does work in extended-precision.

Steve ROBBINS writes:
 > Hullo,
 > 
 > A while back, I reported some `make check' failures in gsl/siman under
 > `Tru64' unix on an alpha machine.  Later, I reported that foregoing
 > optimization generated code that passed the tests.  But that was using an
 > install of egcs from a year back.  I speculated that a newer GCC might
 > work right.  Indeed, it does. 
 > 
 > I have just compiled GSL using GCC 2.95.2, WITH -O2 optimization,
 > and it PASSES all `siman' tests that used to fail.  In fact, it
 > passes all tests except 
 > 
 >     FAIL: vegas(f2), dim=9, err=0.0003, chisq=0.6418 
 >         (0.49968100184661118 observed vs 1 expected) 
 > 
 > (which was already known).
 > 
 > 
 > In summary:
 > 
 > On my "Digital UNIX V4.0F (Rev. 1229)" machine, with an alpha processor,
 > compiling with default (-O2) optimization using egcs-2.91.66 produced code
 > that fails siman/test.  Omitting the optimization flag completely produces
 > code that passes the test.  I did not try dropping back to `-O1'.
 > 
 > On the same platform, gcc 2.95.2 produced code that passes siman/test,
 > using the default `-O2' optimization.
 > 
 > -smr
 > 
 >