[980221]: Performance regression on i586-linux-gnulibc1?

[Manfred Hollstein <manfred at s-direktnet dot de>]

I just ran the good old dhrystone from libg++ with the following compilers:

  1. gcc-2.7.2.3
  2. gcc-2.8.1-980218
  3. egcs-1.0.2 prerelease
  4. egcs-980221

It looks like we've got a significant drop with egcs-980221:

  Compilation flags were:
    -O9 -fomit-frame-pointer -malign-loops=0 -malign-jumps=0 -malign-functions=0
  and:
    -m486				for 2.7.2.3
    -march=pentium -mcpu=pentium	for all else


	-funroll-all-loops	-funroll-loops
1.	243902			192307
2.	243902			205479
3.	252100			218978
4.	211267			209790

Obviously, 1.0.2 performs at its best; 980221 is still quite good with only
`-funroll-loops', but `-funroll-all-loops' seems to be completely down.

Looks like we've got now what's in gcc's info pages since years:

`-funroll-all-loops'
     Perform the optimization of loop unrolling.  This is done for all
     loops and usually makes programs run more slowly.

Manfred