Loop oddity

[John David Anglin]

Here is a loop oddity that I noticed.  The loop 

  for (i = 1000000; i; i--) {};

yields the following assembly code when compiled at -O2 on the PA:

	ldil L'999424,%r19
        ldo 576(%r19),%r19
	ldo -625(%r19),%r19
L$0109
	comib,<>,n 0,%r19,L$0109
	ldo -625(%r19),%r19

You get similar results on other architectures.  I was mystified by
the transformation of the decrement from 1 to 625.  Looking
at the output of the loop pass, I see that it has unrolled the loop:

Loop from 12 to 26: 2 real insns.
Continue at insn 18.
Biv 94: insn 20 const (-1)
Biv 94: verified
Biv 94: initialized at insn 11: initial value (1000000)
Cannot eliminate nonneg biv 94.
Sorted combine statistics:

Loop unrolling: 1000000 iterations.
Unrolling loop 25 times.
Biv 94 safe to split.


Loop from 12 to 26: 2 real insns.
Continue at insn 102.
Biv 94: insn 103 const (-25)
Biv 94: verified
Biv 94: initialized at insn 11: initial value (1000000)
Cannot eliminate nonneg biv 94.
Sorted combine statistics:

Loop unrolling: 40000 iterations.
Unrolling loop 25 times.

The documentation states that loops are not unrolled at -O2.
-fno-unroll-loops has no effect.  Of course, if there is something
in the body of the loop, this doesn't happen.  I noticed this
trying to write a little program to check insn timing.  Thus,
I didn't want the loop transformed.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)