Re: Question about RTL for bitwise AND

[Ian Lance Taylor <ian@airs.com>]

"Matt Lee" <reachmatt.lee@gmail.com> writes:

> Thanks again. I do agree that my costs calculation do not recurse into
> the sub-expressions, but I was just mimicing other ports in this
> regard (see rs6000 for instance) . They do not recurse into
> sub-expressions as well.

In the current mainline, the costs calculation in rs6000.c returns
false for AND and friends.  I haven't looked at the 3.4 sources (I
don't have them checked out).

> Also, I think I traced the problem to fold-const.c::fold_single_bit_test()
> 
>       /* If we have (A & C) != 0 or (A & C) == 0 and C is a power of
>          2, then fold the expression into shifts and logical operations.  */
>       tem = fold_single_bit_test (code, arg0, arg1, type);
>       if (tem)
>         return tem;

You're right, it looks like this is fixed in 4.1 and current mainline
with the patch in PR 14846.

> On a related note, it looks the if-conversion phase as well, chooses
> to use a bunch of shifts and other logical operations to convert a
> branch into linear code. This is fine, except that the resultant
> operations prove more expensive than a branch. I will take a closer
> look at the IFCVT_* macros and see if one of them is useful for
> catching these conditions.

If-conversion also uses costs to make these determinations.  The thing
to tweak is BRANCH_COST, and perhaps MAX_CONDITIONAL_EXECUTE.

Hope this helps.

Ian