From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-bugs-return-442395-listarch-gcc-bugs=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 1529 invoked by alias); 3 Feb 2014 13:51:06 -0000
Mailing-List: contact gcc-bugs-help@gcc.gnu.org; run by ezmlm
Precedence: bulk
List-Id: <gcc-bugs.gcc.gnu.org>
List-Archive: <http://gcc.gnu.org/ml/gcc-bugs/>
List-Post: <mailto:gcc-bugs@gcc.gnu.org>
List-Help: <mailto:gcc-bugs-help@gcc.gnu.org>
Sender: gcc-bugs-owner@gcc.gnu.org
Received: (qmail 1490 invoked by uid 48); 3 Feb 2014 13:51:01 -0000
From: "rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug tree-optimization/60042] vectorizer still does too many dependence tests for himeno:jacobi
Date: Mon, 03 Feb 2014 13:51:00 -0000
X-Bugzilla-Reason: CC
X-Bugzilla-Type: changed
X-Bugzilla-Watch-Reason: None
X-Bugzilla-Product: gcc
X-Bugzilla-Component: tree-optimization
X-Bugzilla-Version: 4.9.0
X-Bugzilla-Keywords: missed-optimization
X-Bugzilla-Severity: normal
X-Bugzilla-Who: rguenth at gcc dot gnu.org
X-Bugzilla-Status: UNCONFIRMED
X-Bugzilla-Priority: P3
X-Bugzilla-Assigned-To: unassigned at gcc dot gnu.org
X-Bugzilla-Target-Milestone: ---
X-Bugzilla-Flags:
X-Bugzilla-Changed-Fields:
Message-ID: <bug-60042-4-9dPAoJaa7G@http.gcc.gnu.org/bugzilla/>
In-Reply-To: <bug-60042-4@http.gcc.gnu.org/bugzilla/>
References: <bug-60042-4@http.gcc.gnu.org/bugzilla/>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 7bit
X-Bugzilla-URL: http://gcc.gnu.org/bugzilla/
Auto-Submitted: auto-generated
MIME-Version: 1.0
X-SW-Source: 2014-02/txt/msg00152.txt.bz2

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60042

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
With some more dumping I seee

himenobmtxpa.c:296:9: note: === vect_prune_runtime_alias_test_list ===
himenobmtxpa.c:296:9: note: merging ranges for *_205, *_324 and *_49, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_205, *_324 and *_192, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_168, *_324 and *_69, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_168, *_324 and *_154, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_265, *_324 and *_296, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_265, *_324 and *_89, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_174, *_324 and *_248, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_174, *_324 and *_161, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_211, *_324 and *_231, *_324
himenobmtxpa.c:296:9: note: merging ranges for *_211, *_324 and *_199, *_324
himenobmtxpa.c:296:9: note: improved number of alias checks from 31 to 21

and

Creating dr for *_205
analyze_innermost: success.
        base_address: pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1009
* 4)
        offset from base address: 0
        constant offset from base address: 0
        step: 4
        aligned to: 128
        base_object: *pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1009
* 4)
        Access function 0: {0B, +, 4}_7
Creating dr for *_168
analyze_innermost: success.
        base_address: pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1023
* 4)
        offset from base address: 0
        constant offset from base address: 0
        step: 4
        aligned to: 128
        base_object: *pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1023
* 4)
        Access function 0: {0B, +, 4}_7
Creating dr for *_265
analyze_innermost: success.
        base_address: pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1034
* 4)
        offset from base address: 0
        constant offset from base address: 0
        step: 4
        aligned to: 128
        base_object: *pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1034
* 4)
        Access function 0: {0B, +, 4}_7
Creating dr for *_174
analyze_innermost: success.
        base_address: pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1063
* 4)
        offset from base address: 0
        constant offset from base address: 0
        step: 4
        aligned to: 128
        base_object: *pretmp_1004 + (sizetype) ((long unsigned int) pretmp_1063
* 4)
        Access function 0: {0B, +, 4}_7
...

so the remaining DDRs against *_324 all look related.

  pretmp_1062 = pretmp_1020 + pretmp_1047;
  pretmp_1063 = _25 * pretmp_1062;

  pretmp_1033 = j_380 + pretmp_1020;
  pretmp_1034 = _25 * pretmp_1033;

  pretmp_1022 = pretmp_1020 + pretmp_1021;
  pretmp_1023 = _25 * pretmp_1022;

but SCEV doesn't expand stmts before the loop and thus doesn't see this.
It's obviously far from trivial to merge segments with symbolic start
addresses ... these are multi-dimensional accesses:

        for(k=1 ; k<kmax ; k++){
          s0= MR(a,0,i,j,k)*MR(p,0,i+1,j,  k)
            + MR(a,1,i,j,k)*MR(p,0,i,  j+1,k)
            + MR(a,2,i,j,k)*MR(p,0,i,  j,  k+1)
            + MR(b,0,i,j,k)
             *( MR(p,0,i+1,j+1,k) - MR(p,0,i+1,j-1,k)
              - MR(p,0,i-1,j+1,k) + MR(p,0,i-1,j-1,k) )
            + MR(b,1,i,j,k)
             *( MR(p,0,i,j+1,k+1) - MR(p,0,i,j-1,k+1)
              - MR(p,0,i,j+1,k-1) + MR(p,0,i,j-1,k-1) )
            + MR(b,2,i,j,k)
             *( MR(p,0,i+1,j,k+1) - MR(p,0,i-1,j,k+1)
              - MR(p,0,i+1,j,k-1) + MR(p,0,i-1,j,k-1) )
            + MR(c,0,i,j,k) * MR(p,0,i-1,j,  k)
            + MR(c,1,i,j,k) * MR(p,0,i,  j-1,k)
            + MR(c,2,i,j,k) * MR(p,0,i,  j,  k-1)
            + MR(wrk1,0,i,j,k);

          ss= (s0*MR(a,3,i,j,k) - MR(p,0,i,j,k))*MR(bnd,0,i,j,k);

          gosa+= ss*ss;
          MR(wrk2,0,i,j,k)= MR(p,0,i,j,k) + omega*ss;
        }

and we manage to merge the fastest varying dimension +-1 ones AFAIK,
but not for example the ones for MR(p,0,i+1,j+1,k) and MR(p,0,i+1,j-1,k).
Ideally we would be able to derive a single check for each array
(which would require analyzing the DRs in the outer loops as well to
gather info about the other dimensions).