[patch, fortran] Fix PR 42131, improvement in do loops

[Daniel Kraft]

Thomas Koenig wrote:
> Hello world,
> 
> this fixes PR 42131 by using an alternative method of calculating
> the count for non-obvious DO loops.  This means that loops with strides
> unknown at compile-time can be unrolled.  Just in time for the stage4
> deadline :-)
> 
> Regression-tested on x86_64-unknown-linux-gnu.  OK for trunk?

Ok (after clearing my mind with regard to the question below), thanks 
for working on this!

I'm just curious about this:

+      tmp = fold_build2 (RSHIFT_EXPR, type, step,
+			 build_int_cst (type,
+					TYPE_PRECISION (type) - 1));
+
+      tmp = fold_build2 (MULT_EXPR, type, tmp,
+			 build_int_cst (type, 2));
+
+      step_sign = fold_build2 (PLUS_EXPR, type, tmp,
+			       fold_convert (type, integer_one_node));

How does this calculate the sign of step?  To me this looks like:

1) Shift step right so that afterwards only the highest bit (which 
contains its sign) is present any longer; thus step will be 0 or 1.

2) Multiply it by two, gives 0 or 2.  (Why no left-shift?)

3) Add 1, gives 1 or 3 (?).

What am I missing here?  (I'm no trans* expert...)

Yours,
Daniel

-- 
Done:  Arc-Bar-Cav-Ran-Rog-Sam-Tou-Val-Wiz
To go: Hea-Kni-Mon-Pri