[Bug target/109004] New: wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[Bug target/109004] New: wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

            Bug ID: 109004
           Summary: wrong code for -O2 (any above -O0) with g++ 11.3 for
                    POWER9 (cross-compiler on x86_64 host)
           Product: gcc
           Version: 11.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: bugreporter66 at gmail dot com
  Target Milestone: ---

Created attachment 54574
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54574&action=edit
archive with all the necessary source and output files

The code sample is attached with all the necessary output files. It's a
regression from g++ 9.x series, where it worked properly.

The line that fails the test is a ternary operator working on two memory
floating point operands: farr[n] < farr[m] ? farr[n] : farr[m];

Two potential workarounds for this are turning optimizations off (-O0) or
re-targeting to POWER8 instead (-O2 -mcpu=power8).

The test case is taken from https://github.com/VectorChief/UniSIMD-assembler,
where that test fails (v1.1.0c) for POWER9 target on Ubuntu MATE 22.04.2

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |9.5.0
            Summary|wrong code for -O2 (any     |[10/11/12/13 Regression]
                   |above -O0) with g++ 11.3    |wrong code for -O2 (any
                   |for POWER9 (cross-compiler  |above -O0) with g++ 11.3
                   |on x86_64 host)             |for POWER9 (cross-compiler
                   |                            |on x86_64 host)
           Keywords|                            |wrong-code
   Target Milestone|---                         |10.5
             Target|                            |powerpc64le

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #1 from bugreporter66 at gmail dot com ---
Checked g++ 10.4 today, it works as it should.
11.3 and 12.1 were tested to show the issue so far.

The command line for building:
powerpc64le-linux-gnu-g++ -O2 -static test.cpp -o test_p64

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

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

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
You haven't told what you think is a bug and what you think is correct
behavior.
I've tried to compile using 12.2.1 20221121
g++ -o test_p64-test{0,.ii} -mcpu=power9 -O0 -fpie -pie
g++ -o test_p64-test{2,.ii} -mcpu=power9 -O2 -fpie -pie
./test_p64-test0 > 0
./test_p64-test2 > 2
diff -up 0 2
g++ -o test_p64-test{0,.ii} -mcpu=power8 -O0 -fpie -pie
g++ -o test_p64-test{2,.ii} -fno-strict-aliasing -fwrapv
-fno-aggressive-loop-optimizations -mcpu=power9 -O2 -fpie -pie
./test_p64-test0 > 0
./test_p64-test2 > 2
diff -up 0 2
And neither shows any differences.
Attaching a screenshot is a very bad idea, one can't copy from it.  Only
preprocessed source should be attached, the rest should go in text here,
describe what to run and what is correct and what is wrong (best testcases just
abort if they don't behave as expected and exit with 0 otherwise).

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #3 from bugreporter66 at gmail dot com ---
Created attachment 54680
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54680&action=edit
binaries and output files as text

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #4 from bugreporter66 at gmail dot com ---
The test prints two array values (farr[11], farr[3]) and then calculates and
prints their respective minimum and maximum values. The first line
(unoptimized) is a correct behavior (only showing the first few lines of the
diff here).

ubuntu-mate@ubuntu-mate:~/Downloads/bug$ diff -up 0 2
--- 0   2023-03-16 10:20:33.592616511 +0000
+++ 2   2023-03-16 10:20:46.624839476 +0000
@@ -1,24 +1,24 @@
 farr[11] = 4.318749e+04, farr[3] = 1.468766e+00
-C MIN(farr[11],farr[3]) = 1.468766e+00, MAX(farr[11],farr[3]) = 4.318749e+04
+C MIN(farr[11],farr[3]) = 4.318749e+04, MAX(farr[11],farr[3]) = 4.318749e+04

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #5 from bugreporter66 at gmail dot com ---
The exact command line used to produce the new binaries (added -static):

ubuntu-mate@ubuntu-mate:~/Downloads/bug$ powerpc64le-linux-gnu-g++  -o
test_p64-test{0,.ii} -mcpu=power9 -O0 -fpie -pie -static
ubuntu-mate@ubuntu-mate:~/Downloads/bug$ 
ubuntu-mate@ubuntu-mate:~/Downloads/bug$ powerpc64le-linux-gnu-g++  -o
test_p64-test{2,.ii} -mcpu=power9 -O2 -fpie -pie -static

Also tried different QEMU versions (6.2.0, 5.1.0, 4.2.0) to eliminate the
possibility of emulation issue. The source and the compiler are the same as
with the original report.

ubuntu-mate@ubuntu-mate:~/Downloads/bug$ powerpc64le-linux-gnu-g++  --version
powerpc64le-linux-gnu-g++ (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |WORKSFORME
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Looks like qemu bug to me.  I get the exact same output even with your binaries
on native power9.

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #7 from bugreporter66 at gmail dot com ---
Good to know. I will file a QEMU bug then. Thank you for your efforts.

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The program uses xsmincdp and xsmaxcdp instructions and it feels as if qemu
implements the latter exactly the same as the former.
So, I guess simpler testcase would be compile with -O2 -mcpu=power9 -ffast-math
__attribute__((noipa)) float
mymin (float x, float y)
{
  return __builtin_fminf (x, y);
}

__attribute__((noipa)) float
mymax (float x, float y)
{
  return __builtin_fmaxf (x, y);
}

int
main ()
{
  if (mymin (1.0f, 2.0f) != 1.0f)
    __builtin_abort ();
  if (mymax (1.0f, 2.0f) != 2.0f)
    __builtin_abort ();
}
and see if it aborts under qemu or not.

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #9 from bugreporter66 at gmail dot com ---
Created attachment 54683
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54683&action=edit
sample_p64

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #10 from bugreporter66 at gmail dot com ---
I checked the simple version of the test with QEMU 6.2.0 and 7.0.0:

ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ powerpc64le-linux-gnu-g++ -O2
-mcpu=power9 -ffast-math -static sample.cpp -o sample_p64
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ qemu-ppc64le --version
qemu-ppc64le version 6.2.0
Copyright (c) 2003-2021 Fabrice Bellard and the QEMU Project developers
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ qemu-ppc64le -cpu POWER9
sample_p64 
qemu: uncaught target signal 6 (Aborted) - core dumped
Aborted (core dumped)
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 

...

ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ qemu-ppc64le --version
qemu-ppc64le version 7.0.0
Copyright (c) 2003-2022 Fabrice Bellard and the QEMU Project developers
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ qemu-ppc64le -cpu POWER9
sample_p64 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 
ubuntu-mate@ubuntu-mate:~/Downloads/test_p64$ 

So it works after 7.0.0

[Bug target/109004] [10/11/12/13 Regression] wrong code for -O2 (any above -O0) with g++ 11.3 for POWER9 (cross-compiler on x86_64 host)

[bugreporter66 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109004

--- Comment #11 from bugreporter66 at gmail dot com ---
Created a QEMU bug here:
https://gitlab.com/qemu-project/qemu/-/issues/1547