This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Question about vectorization limit
- From: Xinliang David Li <davidxl at google dot com>
- To: Jakub Jelinek <jakub at redhat dot com>
- Cc: Toon Moene <toon at moene dot org>, Richard Biener <richard dot guenther at gmail dot com>, Dehao Chen <dehao at google dot com>, GCC Development <gcc at gcc dot gnu dot org>
- Date: Fri, 31 May 2013 12:01:50 -0700
- Subject: Re: Question about vectorization limit
- References: <CAO2gOZX7_-08m_+AEybF0RwG=8Y_qPG_+wjmgsq6ymVWTr3=Vw at mail dot gmail dot com> <CAFiYyc3ehiZeyrXUb+wgj_tBi7WqmeHNQdFK9vDinnMWYHYswA at mail dot gmail dot com> <51A8A3EF dot 3010508 at moene dot org> <20130531134131 dot GT1493 at tucnak dot redhat dot com> <51A8AA4B dot 6080001 at moene dot org> <20130531135414 dot GU1493 at tucnak dot redhat dot com>
On Fri, May 31, 2013 at 6:54 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Fri, May 31, 2013 at 03:48:59PM +0200, Toon Moene wrote:
>> >If you rewrite the above into:
>> >SUBROUTINE XYZ(A, B, N)
>> >DIMENSION A(N), B(N)
>> >DO I = 1, N
>> > C = B(I)
>> > IF (A(I)> 0.0) THEN
>> > A(I) = C / A(I)
>> > ELSE
>> > A(I) = C
>> > ENDIF
>> >ENDDO
>> >END
>> >
>> >then it is vectorized just fine.
>>
>> But this "inner loop" has at least 3 basic blocks - so what does the
>> "loop->num_nodes != 2" test exactly codify ?
>
> With the above testcase it has just 2.
> Before ifcvt pass it still has 4:
> <bb 4>:
> # i_1 = PHI <1(3), i_18(7)>
> _8 = (integer(kind=8)) i_1;
> _9 = _8 + -1;
> c_11 = *b_10(D)[_9];
> _13 = *a_12(D)[_9];
> if (_13 > 0.0)
> goto <bb 5>;
> else
> goto <bb 6>;
>
> <bb 5>:
> _14 = c_11 / _13;
>
> <bb 6>:
> # cstore_17 = PHI <_14(5), c_11(4)>
> *a_12(D)[_9] = cstore_17;
> i_18 = i_1 + 1;
> if (i_1 == _7)
> goto <bb 8>;
> else
> goto <bb 7>;
>
> <bb 7>:
> goto <bb 4>;
> but ifcvt transforms that into:
> <bb 4>:
> # i_1 = PHI <1(3), i_18(5)>
> _8 = (integer(kind=8)) i_1;
> _9 = _8 + -1;
> c_11 = *b_10(D)[_9];
> _13 = *a_12(D)[_9];
> _14 = c_11 / _13;
> cstore_17 = _13 > 0.0 ? _14 : c_11;
> *a_12(D)[_9] = cstore_17;
> i_18 = i_1 + 1;
> if (i_1 == _7)
> goto <bb 6>;
> else
> goto <bb 5>;
>
> <bb 5>:
> goto <bb 4>;
> which is already generally vectorizable. Guess ifcvt can be certainly
> taught if it finds a possibly trapping statement to check if the same
> statement isn't present in all possible branches, though the question is if
> this is best done in ifcvt or some other pass.
This is code hoisting, which should be done earlier. Similarly
missing in GCC is speculative PRE for loads and stores that are safe
to control speculate, which include global scalars, global array
access within bounds. Fancier ones require some analysis. For
instance, a field reference is fully available or fully anticipated,
then it is safe to control speculate a mem ref to a different field:
... p[i].a ....
if (cond0)
{
// frequent
t = p[i].b;
..
}
if (cond1)
{
// frequent
.. = p[i].b;
}
Should be PREed into:
... p[i].a ....
if (cond0)
{
// frequent
t = p[i].b;
..
}
else
{
t = p[i].b;
}
if (cond1)
{
// frequent
.. = t;
}
David
>
> Jakub