From: Richard Sandiford <richard.sandiford@linaro.org>
To: Richard Biener <richard.guenther@gmail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: Use nonzero bits to refine range in split_constant_offset (PR 81635)
Date: Thu, 08 Feb 2018 12:09:00 -0000 [thread overview]
Message-ID: <87d11fy971.fsf@linaro.org> (raw)
In-Reply-To: <CAFiYyc3BB7BhTSoUNPaFQXBk45=9ndJDkqKfwevHHnYJXmdqLg@mail.gmail.com> (Richard Biener's message of "Thu, 8 Feb 2018 12:45:39 +0100")
Richard Biener <richard.guenther@gmail.com> writes:
> On Fri, Feb 2, 2018 at 3:12 PM, Richard Sandiford
> <richard.sandiford@linaro.org> wrote:
>> Index: gcc/tree-data-ref.c
>> ===================================================================
>> --- gcc/tree-data-ref.c 2018-02-02 14:03:53.964530009 +0000
>> +++ gcc/tree-data-ref.c 2018-02-02 14:03:54.184521826 +0000
>> @@ -721,7 +721,13 @@ split_constant_offset_1 (tree type, tree
>> if (TREE_CODE (tmp_var) != SSA_NAME)
>> return false;
>> wide_int var_min, var_max;
>> - if (get_range_info (tmp_var, &var_min, &var_max) != VR_RANGE)
>> + value_range_type vr_type = get_range_info (tmp_var, &var_min,
>> + &var_max);
>> + wide_int var_nonzero = get_nonzero_bits (tmp_var);
>> + signop sgn = TYPE_SIGN (itype);
>> + if (intersect_range_with_nonzero_bits (vr_type, &var_min,
>> + &var_max, var_nonzero,
>> + sgn) != VR_RANGE)
>
> Above it looks like we could go from VR_RANGE to VR_UNDEFINED.
> I'm not sure if the original range-info might be useful in this case -
> if it may be
> can we simply use only the range info if it was VR_RANGE?
I think we only drop to VR_UNDEFINED if we have contradictory
information: nonzero bits says some bits must be clear, but the range
only contains values for which the bits are set. In that case I think
we should either be conservative and not use the information, or be
aggressive and say that we have undefined behaviour, so overflow is OK.
It seems a bit of a fudge to go back to the old range when we know it's
false, and use it to allow the split some times and not others.
Thanks,
Richard
>
> Ok otherwise.
> Thanks,
> Richard.
>
>> return false;
>>
>> /* See whether the range of OP0 (i.e. TMP_VAR + TMP_OFF)
>> @@ -729,7 +735,6 @@ split_constant_offset_1 (tree type, tree
>> operations done in ITYPE. The addition must overflow
>> at both ends of the range or at neither. */
>> bool overflow[2];
>> - signop sgn = TYPE_SIGN (itype);
>> unsigned int prec = TYPE_PRECISION (itype);
>> wide_int woff = wi::to_wide (tmp_off, prec);
>> wide_int op0_min = wi::add (var_min, woff, sgn, &overflow[0]);
>> Index: gcc/testsuite/gcc.dg/vect/bb-slp-pr81635-3.c
>> ===================================================================
>> --- /dev/null 2018-02-02 09:03:36.168354735 +0000
>> +++ gcc/testsuite/gcc.dg/vect/bb-slp-pr81635-3.c 2018-02-02 14:03:54.183521863 +0000
>> @@ -0,0 +1,62 @@
>> +/* { dg-do compile } */
>> +/* { dg-additional-options "-fno-tree-loop-vectorize" } */
>> +/* { dg-require-effective-target vect_double } */
>> +/* { dg-require-effective-target lp64 } */
>> +
>> +void
>> +f1 (double *p, double *q, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = 0; i < n; i += 4)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +void
>> +f2 (double *p, double *q, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = 0; i < n; i += 2)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +void
>> +f3 (double *p, double *q, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = 0; i < n; i += 6)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +void
>> +f4 (double *p, double *q, unsigned int start, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = start & -2; i < n; i += 2)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times "basic block vectorized" 4 "slp1" } } */
>> Index: gcc/testsuite/gcc.dg/vect/bb-slp-pr81635-4.c
>> ===================================================================
>> --- /dev/null 2018-02-02 09:03:36.168354735 +0000
>> +++ gcc/testsuite/gcc.dg/vect/bb-slp-pr81635-4.c 2018-02-02 14:03:54.183521863 +0000
>> @@ -0,0 +1,47 @@
>> +/* { dg-do compile } */
>> +/* { dg-additional-options "-fno-tree-loop-vectorize" } */
>> +/* { dg-require-effective-target lp64 } */
>> +
>> +void
>> +f1 (double *p, double *q, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = 0; i < n; i += 1)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +void
>> +f2 (double *p, double *q, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = 0; i < n; i += 3)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +void
>> +f3 (double *p, double *q, unsigned int start, unsigned int n)
>> +{
>> + p = (double *) __builtin_assume_aligned (p, sizeof (double) * 2);
>> + q = (double *) __builtin_assume_aligned (q, sizeof (double) * 2);
>> + for (unsigned int i = start; i < n; i += 2)
>> + {
>> + double a = q[i] + p[i];
>> + double b = q[i + 1] + p[i + 1];
>> + q[i] = a;
>> + q[i + 1] = b;
>> + }
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-not "basic block vectorized" "slp1" } } */
next prev parent reply other threads:[~2018-02-08 12:09 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-02-02 14:12 Richard Sandiford
2018-02-08 11:45 ` Richard Biener
2018-02-08 12:09 ` Richard Sandiford [this message]
2018-02-08 15:06 ` Richard Biener
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=87d11fy971.fsf@linaro.org \
--to=richard.sandiford@linaro.org \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.guenther@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).