On Fri, Jul 8, 2011 at 03:32, Richard Guenther <rguent...@suse.de> wrote:
> On Thu, 7 Jul 2011, Sebastian Pop wrote:
>
>> Hi,
>>
>> First there are two cleanup patches independent of the fix:
>>
>> Start counting nesting level from 0.
>> Do not compute twice type, lb, and ub.
>>
>> Then the patch that fixes PR47654:
>>
>> Fix PR47654: Compute LB and UB of a CLAST expression.
>>
>> One of the reasons we cannot determine the IV type only from the
>> polyhedral representation is that as in the testcase of PR47654, we
>> are asked to generate an induction variable going from 0 to 127. That
>> could be represented with a "char". However the upper bound
>> expression of the loop generated by CLOOG is min (127, 51*scat_1 + 50)
>> and that would overflow if we use a "char" type. To evaluate a type
>> in which the expression 51*scat_1 + 50 does not overflow, we have to
>> compute an upper and lower bound for the expression.
>>
>> To fix the problem exposed by Tobias:
>>
>> > for (i = 0 ; i < 2; i++)
>> > for (j = i ; j < i + 1; j++)
>> > for (k = j ; k < j + 1; k++)
>> > for (m = k ; m < k + 1; m++)
>> > for (n = m ; n < m + 1; n++)
>> > A[0] += A[n];
>> >
>> > I am a little bit afraid that we will increase the type size by an
>> > order of magnitude (or at least one bit) for each nesting level.
>>
>> instead of computing the lb and ub of scat_1 in "51*scat_1 + 50" based
>> on the type of scat_1 (that we already code generated when building
>> the outer loop), we use the polyhedral representation to get an
>> accurate lb and ub for scat_1.
>>
>> When translating the substitutions of a user statement using this
>> precise method, like for example S5 in vect-pr43423.c:
>>
>> for (scat_1=0;scat_1<=min(T_3-1,T_4-1);scat_1++) {
>> S5(scat_1);
>>
>> we get a type that is too precise: based on the interval [0,99] we get
>> the type "unsigned char" when the type of scat_1 is "int", misleading
>> the vectorizer due to the insertion of spurious casts:
>>
>> # Access function 0: (int) {(<unnamed-unsigned:8>) graphite_IV.7_56, +,
>> 1}_3;
>> #)
>> affine dependence test not usable: access function not affine or constant.
>>
>> So we have to keep around the previous code gcc_type_for_clast_* that
>> computes the type of an expression as the max precision of the
>> components of that expression, and use that when computing the types
>> of substitution expressions.
>>
>> The patches passed together a full bootstrap and test on amd64-linux.
>> Ok for trunk?
>
> The idea sounds good to me and the middle-end-like looking pieces
> look good. I'd appreciate a 2nd look from Tobias.
>
Tobias, could you please have a look at these patches as well?
Thanks,
Sebastian
