🚀 go-pugleaf

In article <1992Feb23.213917.13705@murdoch.acc.Virginia.EDU>, gsh7w@fermi.clas.Virginia.EDU (Greg Hennessy) writes:
> 
> Now that GCC 2.0 is available, I was wondering if anyone could report
> some benchmarks of the new version of GCC verses the versions that are
> shipped by the vendor. I am especially interested in how GCC-2.0 fares
> against both the bundled and unbundled compilers for Sparcstations.

Hi. 
I tried compiling a large FORTRAN numerical package, using Sun Fortran 1.4, Sun
CC 1.1 (unbundled) Sun CC and GCC 2.0, computer SStation IPX.

The compile times seemed quite long for Sun unbundled and Sun Fortran.

Fortran source, tminl2.f r-minl2.f, 85 and 5670 lines respectively

C source (from f2c), tminl2.c r-minl2.c, 179 and 8045 lines resp.

`tminl2.?' is the main program which sets up arguments and calls the
numbercrunching routine in `r-minl2.?', with a 1000 times repetition
around the setup and call of the numbercrunching routine.

Compiler command lines

1)	Sun bundled	/bin/cc -O4 -c ..
2)	Sun unbundled	/usr/lang/cc -O4 -c ..
3)	F77		f77 -O4 -c ..
4)	GCC 2		gcc -O2 -fstrength-reduce -funroll-loops -c ..

sizes 	r-minl2.o/text 	tminl2/text 	tminl2/data 	tminl2/bss

1)	72104		131072  	16384   	12608 
2)	88784		147456		16384		12616	
3)	93784		253952		24576		7272	
4)	53392		106496		16384		13320	

Execution time (user time field of /bin/time output, average of 2)

1)	6.2s
2)	6.6s
3)	5.4s (large variance, 5.3 -- 5.6)
4)	6.1s

All runs produced the same results.



Conclusion 
==========

GCC certainly generates very compact code, and Sun Fortran must know something
that f2c combined with even a very good C compiler doesn't know.

-- 
Mr Jens Tingleff, M.Sc.EE. PhD student at 
     Imperial College, Dept of EE,  Exhibition Road, London SW7 2BT, England
jensting@titan.ee.ic.ac.uk or jensting@ee.ic...  (used to be jensting@diku.dk) 
'You got a light, Mac?'  'No, but I got a dark brown overcoat' Bonzo Dog Band