[Bug target/116629] New: Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[Bug target/116629] New: Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

            Bug ID: 116629
           Summary: Building openvino with -flto ICEs in
                    aarch64_sve::gimple_folder::redirect_pred_x
           Product: gcc
           Version: 14.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

When building openvino with LTO for openSUSE we ICE on aarch64:

[ 4390s] during IPA pass: inline
[ 4390s]
../src/plugins/intel_cpu/thirdparty/ComputeLibrary/src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp:
In function ‘elementwise_sve_op’:
[ 4390s]
../src/plugins/intel_cpu/thirdparty/ComputeLibrary/src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp:75:
internal compiler error: Segmentation fault
[ 4390s]    75 | void elementwise_sve_op(const ITensor *in, ITensor *out, const
Window &window, ElementWiseUnary op)
[ 4390s] 0xd5dcb7 internal_error(char const*, ...)
[ 4390s]        ???:0
[ 4390s] 0x93d4d8 aarch64_sve::gimple_folder::redirect_pred_x()
[ 4390s]        ???:0
[ 4390s] 0x93d5ef aarch64_sve::gimple_folder::fold()
[ 4390s]        ???:0
[ 4390s] 0x93d69f aarch64_sve::gimple_fold_builtin(unsigned int,
gimple_stmt_iterator*, gcall*)
[ 4390s]        ???:0
[ 4390s] 0x11a7f0f optimize_inline_calls(tree_node*)
[ 4390s]        ???:0
[ 4390s] 0x11a6303 inline_transform(cgraph_node*)
[ 4390s]        ???:0
[ 4390s] 0x11a590b execute_all_ipa_transforms(bool)
[ 4390s]        ???:0
[ 4390s] 0x100e697 cgraph_node::expand()
[ 4390s]        ???:0
[ 4390s] 0x17224f3 symbol_table::compile()
[ 4390s]        ???:0
[ 4390s] 0x171a1b7 lto_main()
[ 4390s]        ???:0

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.3
            Summary|Building openvino with      |[14/15 Regression] Building
                   |-flto ICEs in               |openvino with -flto ICEs in
                   |aarch64_sve::gimple_folder: |aarch64_sve::gimple_folder:
                   |:redirect_pred_x            |:redirect_pred_x

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
See https://build.opensuse.org/package/show/openSUSE:Factory:ARM/openvino for
getting to sources/flags and a more complete build log.  Sorry for not
providing a useful testcase at this point.

It worked fine with GCC 13.

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=106326

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
gimple_folder::redirect_pred_x was added with r14-5881-ge09007308c96a0

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> See https://build.opensuse.org/package/show/openSUSE:Factory:ARM/openvino for
> getting to sources/flags and a more complete build log.  Sorry for not
> providing a useful testcase at this point.

This code depends on so many things. It is hard to build from scratch.
I wonder if it could be reproduced by compiling ComputeLibrary and its examples
with LTO.

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #3)
> I wonder if it could be reproduced by compiling ComputeLibrary and its
> examples with LTO.

Yes just building upstream version of ComputeLibrary (with some include fixes)
causes the same ICE:
scons extra_cxx_flags="-O3 -g -flto" Werror=1 -j8 debug=0 asserts=1 neon=1 
os=linux arch=armv8.2-a-sve
...
src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp: In function
‘elementwise_arithmetic_op’:
src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp:38: internal compiler
error: tree check: expected function_type or method_type, have integer_type in
function_args_iter_init, at tree.h:5458
   38 | void elementwise_arithmetic_op(
0x20ce39e internal_error(char const*, ...)
        ../../gcc/diagnostic-global-context.cc:492
0x8644b2 tree_check_failed(tree_node const*, char const*, int, char const*,
...)
        ../../gcc/tree.cc:9003
0x8651f2 tree_check2(tree_node const*, char const*, int, char const*,
tree_code, tree_code)
        ../../gcc/tree.h:3940
0x8651f2 function_args_iter_init(function_args_iterator*, tree_node const*)
        ../../gcc/tree.h:5458
0x8651f2 type_argument_type(tree_node const*, unsigned int)
        ../../gcc/tree.cc:6369
0x144ca7a aarch64_sve::gimple_folder::redirect_pred_x()
        ../../gcc/config/aarch64/aarch64-sve-builtins.cc:3545
0x144cbbc aarch64_sve::gimple_folder::redirect_pred_x()
        ../../gcc/config/aarch64/aarch64-sve-builtins.cc:3541
0x144cbbc aarch64_sve::gimple_folder::fold()
        ../../gcc/config/aarch64/aarch64-sve-builtins.cc:3656
0x144cc87 aarch64_sve::gimple_fold_builtin(unsigned int, gimple_stmt_iterator*,
gcall*)
        ../../gcc/config/aarch64/aarch64-sve-builtins.cc:4723
0x1356aaa aarch64_gimple_fold_builtin
        ../../gcc/config/aarch64/aarch64.cc:15645
0xbd7c98 gimple_fold_call
        ../../gcc/gimple-fold.cc:5747
0xbda464 fold_stmt_1
        ../../gcc/gimple-fold.cc:6538
0xfd6917 fold_marked_statements
        ../../gcc/tree-inline.cc:5439
0xfe0e2e optimize_inline_calls(tree_node*)
        ../../gcc/tree-inline.cc:5587
0xcaf26b inline_transform(cgraph_node*)
        ../../gcc/ipa-inline-transform.cc:792
0xe2c5e6 execute_one_ipa_transform_pass
        ../../gcc/passes.cc:2346
0xe2c5e6 execute_all_ipa_transforms(bool)
        ../../gcc/passes.cc:2409
0xa435fd cgraph_node::expand()
        ../../gcc/cgraphunit.cc:1838
0xa435fd cgraph_node::expand()
        ../../gcc/cgraphunit.cc:1798
0xa44a21 expand_all_functions
        ../../gcc/cgraphunit.cc:2028
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
lto-wrapper: fatal error: aarch64-linux-gnu-g++ returned 1 exit status
compilation terminated.

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
  tree arg0_type = type_argument_type (TREE_TYPE (fndecl), 1);

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Trying to get a reduced testcase somehow ...

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Reduced down to one object file:
```
aarch64-linux-gnu-g++  build/src/cpu/kernels/activation/generic/sve/fp16.os 
-shared
```

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 59067
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59067&action=edit
Preprocessed testcase

aarch64-linux-gnu-g++ -shared t.ii -O3 -flto  -march=armv8.2-a+sve+fp16+dotprod

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #9 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 59068
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59068&action=edit
slightly reduced

`-shared -O3  -march=armv9-a -flto`

Will reduce it further in a few minutes.

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 59069
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59069&action=edit
Reduced all the way I can make it

Still LTO only: `-shared  -O3 -march=armv9-a -flto`

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #59068|0                           |1
        is obsolete|                            |
  Attachment #59069|0                           |1
        is obsolete|                            |

--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 59070
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59070&action=edit
Actually all the way reduced

Still need lto

[Bug target/116629] [14/15 Regression] Building openvino with -flto ICEs in aarch64_sve::gimple_folder::redirect_pred_x

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-09-07

--- Comment #12 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.
You need to compile the testcase using the C++ front-end and with lto to hit
the ICE.

Even the testcase that was added for PR 106326 fails when compiling with the
C++ front-end and with LTO.