On Thu, Oct 21, 2021 at 8:04 AM Andrew MacLeod <[email protected]> wrote:
>
> On 10/19/21 7:13 PM, Andrew Pinski wrote:
> > On Tue, Oct 19, 2021 at 3:32 PM Andrew MacLeod <[email protected]> wrote:
> >> On 10/19/21 5:13 PM, Andrew Pinski wrote:
> >>> On Tue, Oct 19, 2021 at 1:29 PM Andrew MacLeod via Gcc-patches
> >>> <[email protected]> wrote:
> >>>> using testcase ifcvt-4.c:
> >>>>
> >>>>
> >>>> typedef int word __attribute__((mode(word)));
> >>>>
> >>>> word
> >>>> foo (word x, word y, word a)
> >>>> {
> >>>> word i = x;
> >>>> word j = y;
> >>>> /* Try to make taking the branch likely. */
> >>>> __builtin_expect (x > y, 1);
> >>>> if (x > y)
> >>>> {
> >>>> i = a;
> >>>> j = i;
> >>>> }
> >>>> return i * j;
> >>>>
> >
> > The testcase is broken anyways.
> > The builtin_expect should be inside the if to have any effect. Look
> > at the estimated values:
> > if (x_3(D) > y_4(D))
> > goto <bb 4>; [50.00%] <<-- has been reversed.
> > else
> > goto <bb 3>; [50.00%]
> > ;; succ: 4 [50.0% (guessed)] count:536870912 (estimated
> > locally) (TRUE_VALUE,EXECUTABLE)
> > ;; 3 [50.0% (guessed)] count:536870912 (estimated
> > locally) (FALSE_VALUE,EXECUTABLE)
> >
> > See how it is 50/50?
> > The testcase is not even testing what it says it is testing. Just
> > happened to work previously does not mean anything. Move the
> > builtin_expect inside the if and try again. I am shocked it took this
> > long to find the testcase issue really.
> >
> > Thanks,
> > Andrew Pinski
> >
> Moving the expect around doesn't change anything, in fact, it makes it
> worse since fre and evrp immediately eliminate it as true if it is in
> the THEN block.
I think you misunderstood the change I was saying to do.
Try this:
typedef int word __attribute__((mode(word)));
word
foo (word x, word y, word a)
{
word i = x;
word j = y;
/* Try to make taking the branch likely. */
if (__builtin_expect (x > y, 1))
{
i = a;
j = i;
}
return i * j;
}
/* { dg-final { scan-rtl-dump "2 true changes made" "ce1" } } */
This should fix the "estimated values" to be more correct.
Thanks,
Andrew Pinski
>
> It looks like it is eliminated by the CDDCE pass:
>
> cddce1 sees:
>
> _1 = x_5(D) > y_7(D);
> # RANGE [0, 1] NONZERO 1
> _2 = (long int) _1;
> __builtin_expect (_2, 1);
> if (x_5(D) > y_7(D))
> goto <bb 3>; [INV]
> else
> goto <bb 4>; [INV]
>
> and proceeds:
>
> Marking useful stmt: if (x_5(D) > y_7(D))
> processing: if (x_5(D) > y_7(D))
> processing: i_3 = PHI <x_5(D)(2), a_9(D)(3)>
>
> Eliminating unnecessary statements:
> Deleting : __builtin_expect (_2, 1);
> Deleting : _2 = (long int) _1;
> Deleting : _1 = x_5(D) > y_7(D);
>
> IF we are suppose to reverse the If, it is not obvious to me who is
> suppose to.. You seem to be right that its a crap shot that VRP2 does
> it because there isnt enough info to dictate it.. unless somewhere it
> detects that a THEN targets an empty block which fallthrus to the ELSE
> block should be swapped. Or maybe you are right and that it flukeily
> happens due to the ASSERTS being added and removed.
>
> IF i turn of DCE, then this all works like it si ssupopse to.. so maybe
> DCE isnt supopse to remove this?
>
> Andrew
>
