Re: Bug in gcc-20020114: m68k doesn't use tstsf/tstdf patterns

[Peter Barada <pbarada at mail dot wm dot sps dot mot dot com>]

>Also notice that the ftst is redundant as we could use condition code
>values set by the fadd.s.  That would create a loop which looks like:
>
>.L8:
>        fsglmul.s %d0,%fp0
>        fadd.x %fp1,%fp0
>        fbnlt .L4
>        move.w #1,%a0
>.L4:
>        subq.l #1,%d1
>        jbpl .L8

Hmmm, I'll assume that I should look in notice_update_cc to get the
compiler to generate this sequence.

>I did something slightly different.  0.0 is really only cheaper in a
>compare insn because we can use a tst insn instead (which we might later
>manage to remove).  Thus my change restricts the reduction in cost to
>cases where OUTER_CODE is a COMPARE.
>
>Here's the patch I've checked into the development sources (not the
>GCC 3.1 branch):

Sounds better.  You have the advantage of the experience of looking at
the *larger* picture, something I haven't gotten yet :-)

Thanx for updating the tree.  If you don't mind my asking, why didnt'
it go into the 3.1 branch as well?

-- 
Peter Barada                                   Peter.Barada@motorola.com
Wizard                                         781-852-2768 (direct)
WaveMark Solutions(wholly owned by Motorola)   781-270-0193 (fax)