[Bug c++/106041] New: Long/infinite compile time for Arm SIMD -O1 or -O2 but not -O0

[Bug c++/106041] New: Long/infinite compile time for Arm SIMD -O1 or -O2 but not -O0

[jan.wassenberg at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

            Bug ID: 106041
           Summary: Long/infinite compile time for Arm SIMD -O1 or -O2 but
                    not -O0
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jan.wassenberg at gmail dot com
  Target Milestone: ---

Created attachment 53179
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53179&action=edit
Zipped preprocessed source

Long (overnight) compile time for Arm target, vs 28sec for Clang.

Repro: time aarch64-linux-gnu-g++-12 -O2 memory_test.cc.ii

Output: none, not even with -ftime-report. Commenting out the contents of
TestLoadStoreInterleaved3 and TestLoadStoreInterleaved4 avoids the problem.

-v output:
Using built-in specs.
COLLECT_GCC=aarch64-linux-gnu-g++-12
COLLECT_LTO_WRAPPER=/usr/lib/gcc-cross/aarch64-linux-gnu/12/lto-wrapper
Target: aarch64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Debian 12-20220302-1'
--with-bugurl=file:///usr/share/doc/gcc-12/README.Bugs
--enable-languages=c,ada,c++,go,d,fortran,objc,obj-c++,m2 --prefix=/usr
--with-gcc-major-version-only --program-suffix=-12 --enable-shared
--enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext
--enable-threads=posix --libdir=/usr/lib --enable-nls --with-sysroot=/
--enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes
--with-default-libstdcxx-abi=new --enable-gnu-unique-object
--disable-libquadmath --disable-libquadmath-support --enable-plugin
--enable-default-pie --with-system-zlib --enable-libphobos-checking=release
--without-target-system-zlib --enable-multiarch --enable-fix-cortex-a53-843419
--disable-werror --enable-checking=release --build=x86_64-linux-gnu
--host=x86_64-linux-gnu --target=aarch64-linux-gnu
--program-prefix=aarch64-linux-gnu- --includedir=/usr/aarch64-linux-gnu/include
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 12.0.1 20220302 (experimental) [master r12-7448-g58394373a70]
(Debian 12-20220302-1)

[Bug rtl-optimization/106041] Long/infinite compile time for Arm SIMD -O1 or -O2 but not -O0

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |rtl-optimization

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
#0  0x0000000000bd9e74 in df_worklist_propagate_backward (age=<optimized out>,
last_change_age=<synthetic pointer>..., considered=0x3f4de80,
pending=0x44d0f20, bbindex_to_postorder=0x3ecf3d0, bb_index=<optimized out>,
dataflow=0x3ecd500) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-core.cc:971
#1  df_worklist_dataflow_doublequeue (n_blocks=<optimized out>,
bbindex_to_postorder=0x3ecf3d0, blocks_in_postorder=0x3bc1200,
considered=0x3f4de80, pending=0x44d0f20, dataflow=0x3ecd500) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-core.cc:1044
#2  df_worklist_dataflow (dataflow=0x3ecd500, blocks_to_consider=<optimized
out>, blocks_in_postorder=0x3bc1200, n_blocks=<optimized out>) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-core.cc:1121
#3  0x0000000000bda0c8 in df_analyze_problem (dflow=0x3ecd500,
blocks_to_consider=blocks_to_consider@entry=0x3b990b0,
postorder=postorder@entry=0x3bc1200, n_blocks=n_blocks@entry=38) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-core.cc:1175
#4  0x0000000001c3dfa8 in fast_dce (word_level=word_level@entry=false) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/dce.cc:1169
#5  0x0000000001c3e1e8 in rest_of_handle_fast_dce () at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/dce.cc:1194
#6  run_fast_df_dce () at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/dce.cc:1242
#7  0x0000000000be0ef8 in df_lr_finalize (all_blocks=0x44d0f40) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-problems.cc:1065
#8  0x0000000000bda0dc in df_analyze_problem (dflow=0x3ecd500,
blocks_to_consider=0x44d0f40, postorder=0x3bc1200, n_blocks=38) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/df-core.cc:1180
#9  0x0000000000bda1cc in df_analyze_1 () at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/vec.h:1472
#10 0x0000000001c27180 in execute_compare_elim_after_reload () at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/compare-elim.cc:910
#11 (anonymous namespace)::pass_compare_elim_after_reload::execute
(this=<optimized out>) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/compare-elim.cc:967
#12 0x0000000000fb2420 in execute_one_pass (pass=pass@entry=0x2981b60) at
/home/ubuntu/src/upstream-gcc-aarch64/gcc/gcc/passes.cc:2637

I need to double check this but it seems to be stuck here.

Also this is with "void hwy::N_NEON::TestLoadStoreInterleaved2::operator()(T,
D) [with T = double; D = hwy::N_NEON::Simd<double, 1, 0>]" function.

[Bug rtl-optimization/106041] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Long/infinite compile time  |infinite loop in fast_dce
                   |for Arm SIMD -O1 or -O2 but |at -O1 with aarch64
                   |not -O0                     |

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> I need to double check this but it seems to be stuck here.

Yes it is stuck in fast_dce. The basic blocks don't look that big either ...

[Bug rtl-optimization/106041] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-reduction

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
If I get some time next weekend I might reduce this further to understand it.
There are only 38 BBs in this function too and they are not large either.

[Bug rtl-optimization/106041] infinite loop in fast_dce at -O1 with aarch64

[jan.wassenberg at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #4 from Jan Wassenberg <jan.wassenberg at gmail dot com> ---
Thanks for having a look! BTW forgot to mention: version 11.0 does not have
this issue.

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|12.0                        |12.1.0
   Target Milestone|---                         |12.2
            Summary|infinite loop in fast_dce   |[12/13 Regression] infinite
                   |at -O1 with aarch64         |loop in fast_dce at -O1
                   |                            |with aarch64

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #5 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Jan Wassenberg from comment #4)
> Thanks for having a look! BTW forgot to mention: version 11.0 does not have
> this issue.

Can you please share pre-processed source from GCC 11? And can you verify that
taking it and using it with GCC 12 is stuck?

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[jan.wassenberg at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #6 from Jan Wassenberg <jan.wassenberg at gmail dot com> ---
Created attachment 53181
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53181&action=edit
GCC11 zipped preprocessed source

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[jan.wassenberg at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #7 from Jan Wassenberg <jan.wassenberg at gmail dot com> ---
Sure, added GCC11 .ii. It does indeed still freeze GCC 12.

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[malat at debian dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #9 from Mathieu Malaterre <malat at debian dot org> ---
(In reply to Andrew Pinski from comment #3)
> If I get some time next weekend I might reduce this further to understand it.
> There are only 38 BBs in this function too and they are not large either.

Seems confirmed at least.

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed, it never finishes word-level iteration.

static void
fast_dce (bool word_level)

{
...
  while (global_changed)
    {
...
    }

IMHO that "fast" DCE should hard-limit the iteration count anyway.  I'm trying
to reduce the testcase based on that.

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 54503
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54503&action=edit
reduced testcase

Reduced testcase (on the GCC 12 branch).  Reduced with -O1 and

diff --git a/gcc/dce.cc b/gcc/dce.cc
index 6676cbcd429..8af54a96394 100644
--- a/gcc/dce.cc
+++ b/gcc/dce.cc
@@ -36,6 +36,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "tree-pass.h"
 #include "dbgcnt.h"
 #include "rtl-iter.h"
+#include "diagnostic-core.h"


 /* -------------------------------------------------------------------------
@@ -1103,9 +1104,13 @@ fast_dce (bool word_level)

   dead_debug_global_init (&global_debug, NULL);

+  unsigned iters = 0;
   while (global_changed)
     {
       global_changed = false;
+      ++iters;
+      if (iters > 50)
+       internal_error ("too many iterations");

       for (i = 0; i < n_blocks; i++)
        {

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|needs-reduction             |
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2023-02-21

--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
The reduced testcase doesn't fail on the trunk (but I lack preprocessed source
for trunk to re-reduce it).  The unreduced testcase doesn't build on trunk due
to standard library header portability issues.

The reduced testcase is very small - two BBs, one BB with an infinite cycle and
no path to exit.

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #13 from Andrew Pinski <pinskia at gcc dot gnu.org> ---

(insn 42 41 43 3 (clobber (reg/v:V4x1DF 40 v8 [orig:96 tup ] [96])) -1
     (nil))

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=108681

--- Comment #14 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #12)
> The reduced testcase doesn't fail on the trunk (but I lack preprocessed
> source for trunk to re-reduce it).  The unreduced testcase doesn't build on
> trunk due to standard library header portability issues.
> 
> The reduced testcase is very small - two BBs, one BB with an infinite cycle
> and no path to exit.

I suspect this is a dup of bug 108681 ....

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #15 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #12)
> The reduced testcase doesn't fail on the trunk 

Most likely because it was fixed when PR 108681 was fixed .

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

--- Comment #16 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #11)
> Created attachment 54503 [details]
> reduced testcase

This reduced testcase looks almost the same as the reduced testcase in bug
108681 comment #4 (when comparing the gimple at optimized stage).

[Bug rtl-optimization/106041] [12/13 Regression] infinite loop in fast_dce at -O1 with aarch64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106041

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |DUPLICATE

--- Comment #17 from Richard Biener <rguenth at gcc dot gnu.org> ---
Duplicate then.

*** This bug has been marked as a duplicate of bug 108681 ***