[Bug target/110960] New: TestSatWidenMulPairwiseAdd in the Google Highway Test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[Bug target/110960] New: TestSatWidenMulPairwiseAdd in the Google Highway Test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

            Bug ID: 110960
           Summary: TestSatWidenMulPairwiseAdd in the Google Highway Test
                    suite fails when compiled with GCC 12 or later with
                    the -mcpu=power9 option
           Product: gcc
           Version: 12.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: john_platts at hotmail dot com
  Target Milestone: ---

Here are the steps to reproduce the problem:
1. Clone the Google Highway git repository from
https://github.com/google/highway.git.
2. Create a build directory called hwy_ppc9_gcc12_build in the highway
directory.
3. In the hwy_ppc9_gcc12_build directory, execute the following command:
CC=powerpc64le-linux-gnu-gcc-12 CXX=powerpc64le-linux-g++-12 cmake ..
-DCMAKE_C_COMPILER_TARGET="powerpc64le-linux-gnu"
-DCMAKE_CXX_COMPILER_TARGET="powerpc64le-linux-gnu"
-DCMAKE_C_FLAGS='-mcpu=power9' -DCMAKE_CXX_FLAGS='-mcpu=power9
-DHWY_DISABLED_TARGETS=6918232715082858496 -DHWY_BROKEN_TARGETS=0'
-DHWY_ENABLE_CONTRIB=OFF
4. After the cmake command is executed, execute the "make mul_test" command in
the hwy_ppc9_gcc12_build directory.
5. Execute the ./tests/mul_test program in the hwy_ppc9_gcc12_build directory.

Here is the expected output of the mul_test program:
[==========] Running 9 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 9 tests from HwyMulTestGroup/HwyMulTest
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC9 (21 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC9 (2 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC9 (45 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC9 (3 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC9 (22 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC9 (3 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllSatWidenMulPairwiseAdd/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllSatWidenMulPairwiseAdd/PPC9 (5
ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllReorderWidenMulAccumulate/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllReorderWidenMulAccumulate/PPC9
(5 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllRearrangeToOddPlusEven/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllRearrangeToOddPlusEven/PPC9 (0
ms)
[----------] 9 tests from HwyMulTestGroup/HwyMulTest (112 ms total)

[----------] Global test environment tear-down
[==========] 9 tests from 1 test suite ran. (117 ms total)
[  PASSED  ] 9 tests.

Here is the actual output of the mul_test program when compiled with GCC 12 and
the -mcpu=power9 options:
[==========] Running 9 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 9 tests from HwyMulTestGroup/HwyMulTest
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC9 (11 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC9 (3 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC9 (46 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC9 (1 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC9 (15 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC9
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC9 (4 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllSatWidenMulPairwiseAdd/PPC9


i16x8 expect [0+ ->]:
  0xEB00,0xEA00,0xE900,0xE800,0xE700,0xE600,0xE500,
i16x8 actual [0+ ->]:
  0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,
Abort at
/home/some_user/hwy_ppc9_build_080923/highway/hwy/tests/mul_test.cc:585: PPC9,
i16x8 lane 0 mismatch: expected '0xEB00', got '0x8000'.

Aborted (core dumped)

The TestAllSatWidenMulPairwiseAdd test failure does not occur when
hwy/tests/mul_test.cc is compiled with GCC 11 or Clang 16 with the
"-mcpu=power9 -DHWY_DISABLED_TARGETS=6918232715082858496
-DHWY_BROKEN_TARGETS=0" options.

I have written some test programs to try to reproduce the problem (one test
program that only uses the Altivec intrinsics and another test program that
uses the Google Highway library), and I was not encountering the same failures
with those test programs.

Attached is another test program to try to reproduce the problem, and the
attached test program runs successfully when compiled with GCC 12 with the
-mcpu=power9 option.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway Test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #1 from John Platts <john_platts at hotmail dot com> ---
Created attachment 55710
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55710&action=edit
Test program to reproduce SatWidenMulPairwiseAdd compilation bug

The attached ppc9_sat_widen_mul_pairwise_add_test_080823_4.cpp program needs to
be compiled with the "-mcpu=power9 -std=c++17" options, but has no dependencies
on Google Highway or CMake.

The attached ppc9_sat_widen_mul_pairwise_add_test_080823_4 program runs
successfully if compiled with the "-mcpu=power9 -std=c++17" option with GCC 9.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #2 from John Platts <john_platts at hotmail dot com> ---
Created attachment 55711
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55711&action=edit
Test program to reproduce SatWidenMulPairwiseAdd compilation bug (requires
CMake and Google Highway)

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #3 from John Platts <john_platts at hotmail dot com> ---
Here is the output of running the "./tests/mul_test" program in the Google
Highway test suite when compiled with the "-mcpu=power8
-DHWY_DISABLED_TARGETS=6917951240106147840" options when compiled with GCC 12:
[==========] Running 9 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 9 tests from HwyMulTestGroup/HwyMulTest
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMul/PPC8 (11 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulHigh/PPC8 (4 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulFixedPoint15/PPC8 (67 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulEven/PPC8 (2 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllMulAdd/PPC8 (20 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllWidenMulPairwiseAdd/PPC8 (2 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllSatWidenMulPairwiseAdd/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllSatWidenMulPairwiseAdd/PPC8 (6
ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllReorderWidenMulAccumulate/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllReorderWidenMulAccumulate/PPC8
(4 ms)
[ RUN      ] HwyMulTestGroup/HwyMulTest.TestAllRearrangeToOddPlusEven/PPC8
[       OK ] HwyMulTestGroup/HwyMulTest.TestAllRearrangeToOddPlusEven/PPC8 (0
ms)
[----------] 9 tests from HwyMulTestGroup/HwyMulTest (122 ms total)

[----------] Global test environment tear-down
[==========] 9 tests from 1 test suite ran. (127 ms total)
[  PASSED  ] 9 tests.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #4 from John Platts <john_platts at hotmail dot com> ---
I had made some changes to TestSatWidenMulPairwiseAdd in hwy/tests/mul_test.cc
that would get TestSatWidenMulPairwiseAdd to pass successfully on POWER9 when
compiled with GCC 12 with the "-mcpu=power9
-DHWY_DISABLED_TARGETS=6918232715082858496 -DHWY_BROKEN_TARGETS=0" options.

The version of Google Highway with the TestSatWidenMulPairwiseAdd changes to
get TestSatWidenMulPairwiseAdd to pass successfully on POWER9 with the
"-mcpu=power9 -DHWY_DISABLED_TARGETS=6918232715082858496
-DHWY_BROKEN_TARGETS=0" options can be found in the
jep_hwy_ppc9_multest_diff_080923 repository at
https://github.com/johnplatts/jep_google_highway.git.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #5 from John Platts <john_platts at hotmail dot com> ---
The version of Google Highway with the TestSatWidenMulPairwiseAdd changes to
get TestSatWidenMulPairwiseAdd to pass successfully on POWER9 with the
"-mcpu=power9 -DHWY_DISABLED_TARGETS=6918232715082858496
-DHWY_BROKEN_TARGETS=0" options can be found in the
jep_hwy_ppc9_multest_diff_080923 branch in the
https://github.com/johnplatts/jep_google_highway.git repository.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #6 from John Platts <john_platts at hotmail dot com> ---
Need to use revision ff1ad85a96c0bc8483b582d6dbceb8bc07edd226 of Google Highway
to reproduce the PPC9 codegen bug with GCC 12 as the TestSatWidenMulPairwiseAdd
will now pass on PPC9 due to a recent update to TestSatWidenMulPairwiseAdd in
the master branch of the Google Highway Git repository.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[malat at debian dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #7 from Mathieu Malaterre <malat at debian dot org> ---
@rguenth You added `needs-bisection` keyword, but the example is quite small:
154 lines of code.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Mathieu Malaterre from comment #7)
> @rguenth You added `needs-bisection` keyword, but the example is quite
> small: 154 lines of code.

needs-bisection means to figure out which commit caused the regression.
That is different from needs-reduction keyword which says we need a reduction
still ...

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #9 from John Platts <john_platts at hotmail dot com> ---
Created attachment 55867
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55867&action=edit
Test program to reproduce SatWidenMulPairwiseAdd compilation bug

The attached ppc9_test_sat_widen_pairwise_add_090923_2.cpp test program
generates the following error if compiled with the -O2 -mcpu=power9 options:
Mismatch in lane 0 (ppc9_test_sat_widen_pairwise_add_090923_2.cpp:341):
Expected: -5376, -5632, -5888, -6144, -6400, -6656, -6912, -7168
Actual: -32768, -32768, -32768, -32768, -32768, -32768, -32768, -32768
Aborted (core dumped)

Here is the expected output of the
ppc9_test_sat_widen_pairwise_add_090923_2.cpp test program:
Test completed successfully

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #10 from John Platts <john_platts at hotmail dot com> ---
Created attachment 55868
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55868&action=edit
Test program to reproduce SatWidenMulPairwiseAdd compilation bug

The ppc9_test_sat_widen_pairwise_add_090923_2b.cpp compiles successfully with
GCC 12, even if compiled with the -O2 -mcpu=power9 options.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[john_platts at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

--- Comment #11 from John Platts <john_platts at hotmail dot com> ---
Created attachment 55869
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55869&action=edit
Test program to reproduce GCC 12 compilation bug

Here is the expected output of the ppc9_test_sat_add_090923.cpp test program:
Test completed successfully

The ppc9_test_sat_add_090923.cpp test program results in the following output
when compiled with GCC 12 with -O2 -mcpu=power9 -std=c++14:
Expected vector not equal to actual
Aborted (core dumped)

The ppc9_test_sat_add_090923.cpp test program does generate the expected
results when compiled with GCC 11 with the -O2 -mcpu=power9 -std=c++14 options.

[Bug target/110960] TestSatWidenMulPairwiseAdd in the Google Highway test suite fails when compiled with GCC 12 or later with the -mcpu=power9 option

[meissner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110960

Michael Meissner <meissner at gcc dot gnu.org> changed:

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

--- Comment #12 from Michael Meissner <meissner at gcc dot gnu.org> ---
The test case actually shows on power8 GCC was generating incorrect code, and
power9 is actually doing the right thing.  But the test case was written
assuming the previous behavior was correct.

TL;DNR answer power8 generated STVX instead of STXVD2VX.  Power9 generates
STXV.

To explain what the issue is, we need to go back in history.

PowerPC processors (and Power before it) were originally designed for big
endian environments.  The Altivec instruction set had limited vector save and
load instructions (STVX and LVX) which ignored the bottom 4 bits of the
address.  STVX and LVX did the correct byte swapping if the PowerPC was running
in little endian mode.

When power7 came out with the VSX instruction set, the vector save and load
instructions (STXVD2X, STXV4X, LXVD2X, and LXV4X) were added.  These
instructions allowed saving and loading all 64 VSX registers (32 registers that
overlapped with floating point registers, and 32 registers that overlapped with
traditional Altivec registers).  However, these instructions only store and
load values using big endian ordering.

After the power8 came out, the PowerPC Linux systems were moved from being big
endian to little endian.  This meant that after doing a vector load
instruction, we had to do explicit byte swapping, and before a vector save we
had to do the byte swapping of the value before doing the save.

We added an optimization to GCC that in the special case of storing/loading
temporaries on the stack, we would use the Altivec instructions STVX and LVX
and elimiante the byte swapping instructions since we could insure that all
temporaries were correctly aligned.  But we couldn't use STVX and LVX in
general due to these instructions ignoring the bottom 4 bits of the address and
they restricted the vector registers to just the VSX registers that overlap
with the Altivec registers.

When power9 came out, we added new vector store and load instructions (STXV,
STXVX, LXV, and LXVX) that did the correct byte swapping on little endian
systems.  GCC now generates these instructions and eliminates the special code
to use the Altivec STVX and LVX instructions.

In the test case, VerifyVecEqToActual takes 2 vector arguments, and creates 2
16 byte arrays, and stores each vector into the array.  It uses
reinterpret_cast to convert this into a store instruction.

However, since the temporary is on the stack, on power8 this uses the Altivec
STVX instruction and it gets byte swapped.