[Bug c++/109724] New: [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp

[Bug c++/109724] New: [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

            Bug ID: 109724
           Summary: [10 regression] Huge memory usage when building
                    qtwebengine's SkOpAngle.cpp
           Product: gcc
           Version: 10.4.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: sjames at gcc dot gnu.org
                CC: rguenth at gcc dot gnu.org
  Target Milestone: ---

Created attachment 54988
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54988&action=edit
SkOpAngle.ii.xz

Originally reported downstream in Gentoo at https://bugs.gentoo.org/905634. 

Matthew Ogilvie reports that 10.4.1 20230119
(18bc39a4c83f8d59175061a58637de9f4b2a54d4) is OK, but 10.4.1 20230426
(db0b8927d663cf54ad13dc2629ec195e3883d0d0) is not.

releases/gcc-10 is bad too.

---


```
$ g++ -O2 /tmp/SkOpAngle.ii -c
[consumes a huge amount of memory quite quickly and takes a long time to
compile, OOMs before I can complete compilation (it may not even terminate?)]

```

----

Bisect says

051f78a5c1d6994c10ee7c35453ff0ccee94e5c6 is the first bad commit
commit 051f78a5c1d6994c10ee7c35453ff0ccee94e5c6
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Sep 2 13:36:13 2022 +0200

    tree-optimization/106809 - compile time hog in VN

i.e. r10-11201-g051f78a5c1d699.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

Sam James <sjames at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |compile-time-hog,
                   |                            |memory-hog

--- Comment #1 from Sam James <sjames at gcc dot gnu.org> ---
Reverting it on the branch makes compile succeed within about a second.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #2 from Sam James <sjames at gcc dot gnu.org> ---
I'm going to try reduce this with a timeout of ~3s or so and pray, but please
tell me if you have a better idea to help speed it up.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #3 from Sam James <sjames at gcc dot gnu.org> ---
Created attachment 54989
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54989&action=edit
reduced.ii

This reduced version seems to exhibit the same behaviour but please verify.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-bisection
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
   Target Milestone|---                         |10.5
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-05-04

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Huh, interesting.

Did you by chance verify if GCC 11, 12 or 13 are affected similarly?

I wonder if we can bisect what fixed it if not.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

Sam James <sjames at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |marxin at gcc dot gnu.org

--- Comment #5 from Sam James <sjames at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #4)
> Did you by chance verify if GCC 11, 12 or 13 are affected similarly?

releases/{11,12,13} are fine.

> I wonder if we can bisect what fixed it if not.

The 11.1 and 12.1 tags are fine too.

I'm worried that it didn't ever exist on trunk and is a combination of
backports manifesting badly.

I don't remember it happening on any of the branches either, and given how
common a package qtwebengine is, I'm confident we would've noticed if it
manifested on any of the release/* branches for an extended period at least.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
I checked 11 and 12 and they are not affected.  We run in circles through
valueize_refs_1 in vn_reference_maybe_forwprop_address, repeatedly picking up

_24 = &MEM <const struct SkDCubic> [(const struct SkDPoint
&)iftmp.77_74].fPts[idx2_62].fX;

where the available definition for the value if iftmp.77_74 is _24.

We have

<bb 70> [local count: 224314031]:
_24 = &MEM <const struct SkDCubic> [(const struct SkDPoint
&)iftmp.77_74].fPts[idx2_62].fX;
_259 = minX;
_260 = MEM <const struct SkDCubic> [(const double
&)iftmp.77_74].fPts[idx2_62].fX;

and the def of _74 is

iftmp.77_74 = PHI <iftmp.77_178(64), iftmp.77_177(65)>

_177 is

iftmp.77_177 = &this_93(D)->fPart.fCurve;

abd _62 is value-numbered to zero.  But since on the 10 branch we have
get_addr_base_and_unit_offset without valueization we likely do not
notice the address doesn't actually change.

Backporting r11-199-g2b42509f8b7bdf fixes this issue (well, it avoids it).
I'm not 100% sure it doesn't just make the issue latent.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
The real issue is probably that when value-numbering

Value numbering stmt = iftmp.77_74 = PHI <iftmp.77_178(64), iftmp.77_177(65)>
Setting value number of iftmp.77_74 to iftmp.77_177 (changed)
iftmp.77_177 is available for iftmp.77_177

we have _177 available but for

Value numbering stmt = _24 = &MEM <const struct SkDCubic> [(const struct
SkDPoint &)iftmp.77_74].fPts[idx2_62].fX;
Setting value number of _24 to iftmp.77_177 (changed)
Making available beyond BB70 _24 for value iftmp.77_177

we appearantly have not, despite the _74 def dominating this stmt.  Of course
that's an inherent limitation of the dominated_by_p_w_unex tool that was
patched with the offending revision.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|needs-bisection             |

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'm reducing a testcase and will revert the backport that caused this.  That
should be the least risky.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Reduced testcase.  I've reduced with

diff --git a/gcc/tree-ssa-sccvn.c b/gcc/tree-ssa-sccvn.c
index 666d7061aa8..31879fb761b 100644
--- a/gcc/tree-ssa-sccvn.c
+++ b/gcc/tree-ssa-sccvn.c
@@ -1558,6 +1558,7 @@ valueize_refs_1 (vec<vn_reference_op_s> *orig, bool
*value
ized_anything,
 {
   vn_reference_op_t vro;
   unsigned int i;
+  auto_bitmap visited;

   *valueized_anything = false;

@@ -1609,8 +1610,12 @@ valueize_refs_1 (vec<vn_reference_op_s> *orig, bool
*valueized_anything,
               && vro->opcode == SSA_NAME
               && (*orig)[i - 1].opcode == MEM_REF)
        {
+         tree name = vro->op0;
          if (vn_reference_maybe_forwprop_address (orig, &i))
-           *valueized_anything = true;
+           {
+             *valueized_anything = true;
+             gcc_assert (bitmap_set_bit (visited, SSA_NAME_VERSION (name)));
+           }
        }
       /* If it transforms a non-constant ARRAY_REF into a constant
         one, adjust the constant offset.  */

--

double &SkTMax(double &a, double &b) { return a ?: b; }
struct SkDPoint {
  double fX;
};
struct SkDCubic {
  SkDPoint &operator[](int n) { return fPts[n]; }
  SkDPoint fPts[4];
};
struct SkDCurve {
  SkDCubic fCubic;
  SkDPoint &operator[](int n) { return fCubic[n]; }
};
struct SkDCurveSweep {
  SkDCurve fCurve;
};
int endsIntersect_rPts;
double endsIntersect_maxX;
struct SkOpAngle {
  void endsIntersect();
  SkDCurveSweep fPart;
} endsIntersect_rh;
void SkOpAngle::endsIntersect() {
  for (int index = 0;; ++index) {
    SkDCurve &curve = index ? endsIntersect_rh.fPart.fCurve : fPart.fCurve;
    for (int idx2 = 0; endsIntersect_rPts; ++idx2) {
      SkDPoint &__trans_tmp_1 = curve[idx2];
      endsIntersect_maxX = SkTMax(endsIntersect_maxX, __trans_tmp_1.fX);
    }
  }
}

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:ee99aaae4aeecd55f1d945a959652cf07e3b2e9e

commit r14-470-gee99aaae4aeecd55f1d945a959652cf07e3b2e9e
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 4 09:41:33 2023 +0200

    tree-optimization/109724 - new testcase

    The following adds a testcase for PR109724 which was caused by
    backporting r13-2375-gbe1b42de9c151d and fixed by r11-199-g2b42509f8b7bdf.

            PR tree-optimization/109724
            * g++.dg/torture/pr109724.C: New testcase.

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-13 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:de45793d0a102fc71785b90ce9ce75e57878da5d

commit r13-7293-gde45793d0a102fc71785b90ce9ce75e57878da5d
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 4 09:41:33 2023 +0200

    tree-optimization/109724 - new testcase

    The following adds a testcase for PR109724 which was caused by
    backporting r13-2375-gbe1b42de9c151d and fixed by r11-199-g2b42509f8b7bdf.

            PR tree-optimization/109724
            * g++.dg/torture/pr109724.C: New testcase.

    (cherry picked from commit ee99aaae4aeecd55f1d945a959652cf07e3b2e9e)

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-12 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:9945e5f0e705588eada3ae26bb2168124c88461a

commit r12-9510-g9945e5f0e705588eada3ae26bb2168124c88461a
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 4 09:41:33 2023 +0200

    tree-optimization/109724 - new testcase

    The following adds a testcase for PR109724 which was caused by
    backporting r13-2375-gbe1b42de9c151d and fixed by r11-199-g2b42509f8b7bdf.

            PR tree-optimization/109724
            * g++.dg/torture/pr109724.C: New testcase.

    (cherry picked from commit ee99aaae4aeecd55f1d945a959652cf07e3b2e9e)

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #13 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-11 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:7d7e2d67bfd316125d7344bc713396ecab2ac3df

commit r11-10744-g7d7e2d67bfd316125d7344bc713396ecab2ac3df
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 4 09:41:33 2023 +0200

    tree-optimization/109724 - new testcase

    The following adds a testcase for PR109724 which was caused by
    backporting r13-2375-gbe1b42de9c151d and fixed by r11-199-g2b42509f8b7bdf.

            PR tree-optimization/109724
            * g++.dg/torture/pr109724.C: New testcase.

    (cherry picked from commit ee99aaae4aeecd55f1d945a959652cf07e3b2e9e)

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

--- Comment #14 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:946dc2eb7aeb1f3bf0085975a7325941acfb978f

commit r10-11392-g946dc2eb7aeb1f3bf0085975a7325941acfb978f
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 4 09:41:33 2023 +0200

    tree-optimization/109724 - new testcase

    The following adds a testcase for PR109724 which was caused by
    backporting r13-2375-gbe1b42de9c151d and fixed by r11-199-g2b42509f8b7bdf.

            PR tree-optimization/109724
            * g++.dg/torture/pr109724.C: New testcase.

    (cherry picked from commit ee99aaae4aeecd55f1d945a959652cf07e3b2e9e)

[Bug c++/109724] [10 regression] Huge memory usage when building qtwebengine's SkOpAngle.cpp since r10-11201-g051f78a5c1d699

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109724

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #15 from Richard Biener <rguenth at gcc dot gnu.org> ---
I have reverted the offending commit.  Many thanks for the report.