[Bug c/106755] New: Incorrect code gen for altivec intrinsics with constant inputs

[Bug c/106755] New: Incorrect code gen for altivec intrinsics with constant inputs

[munroesj at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

            Bug ID: 106755
           Summary: Incorrect code gen for altivec intrinsics with
                    constant inputs
           Product: gcc
           Version: 12.2.1
            Status: UNCONFIRMED
          Severity: blocker
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: munroesj at gcc dot gnu.org
  Target Milestone: ---

Created attachment 53514
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53514&action=edit
Reducted test case for vec_muludq() with make

Compiling the PVECLIB project V1.0.4-4 fails unit test (make check) when
compiled with GCC 12 on Fedora 36/37.

Two units test for:
Vector Multiply Unsigned Double Quadword. vec_muludq()
and
Vector Multiply-Add Unsigned Quadword. vec_madduq()

The tests that fail are passing local vector constants to in-lined instants of
these functions.

Current status; the PVECLIB package is blocked for Fedora 37 because it will
not compile with the default GCC-12 compiler.

[Bug c/106755] Incorrect code gen for altivec intrinsics with constant inputs

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

Peter Bergner <bergner at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2022-08-26
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
                 CC|                            |bergner at gcc dot gnu.org,
                   |                            |linkw at gcc dot gnu.org,
                   |                            |meissner at gcc dot gnu.org,
                   |                            |segher at gcc dot gnu.org

--- Comment #1 from Peter Bergner <bergner at gcc dot gnu.org> ---
I can confirm I see the pveclib test errors.  I'm having a look.

[Bug c/106755] Incorrect code gen for altivec intrinsics with constant inputs

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

--- Comment #2 from Peter Bergner <bergner at gcc dot gnu.org> ---
So the tests (I've removed all static inline usage and always use -fno-inline)
pass with -O1 and fail with -O2 and -O3.  Looking at all of the optimizations
enabled by -O2 that are not in -O1 and using -fno-* for them, the only option
that allows the tests to pass with -O2 is -fno-strict-aliasing.  That said,
-Wall and -Wstrict-aliasing do not flag any warnings with the code.  I suppose
they could miss some issues in the test case code???

In addition, if I move the vec_muludq() function to its own source file, then
the tests pass with -O2 and -O3.

[Bug target/106755] Incorrect code gen for altivec intrinsics with constant inputs

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

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

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

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Peter Bergner from comment #2)
> That said, -Wall and -Wstrict-aliasing do not flag
> any warnings with the code.  I suppose they could miss some issues in the
> test case code???

typedef __vector unsigned int vui32_t;
...
typedef __vector unsigned __int128 vui128_t;
...

static inline vui128_t
vec_muludq (vui128_t *mulu, vui128_t a, vui128_t b)
{
...
  *mulu = (vui128_t) t;
...

int
test_muludq (void)
{
  vui32_t i, j, k, l /*, m*/;
....
  k = (vui32_t) test_vec_muludq((vui128_t* )&l, (vui128_t)i, (vui128_t)j);
....
  print_vint128_prod ("2**96-1 * 2**96-1 ", k, i, j, l);


So yes there is aliasing violation as you do the store as vui128_t aka
"__vector unsigned __int128" (which has the same aliasing set) as "__int128"
but then do the load from it as vui32_t aka "__vector unsigned int vui32_t"
(which is the same aliasing set as int). 

-Wstrict-aliasing=3 might warn about this case I can't remember exactly of the
levels of the options but the -Wstrict-aliasing warnings are always not going
to happen, sometimes it is easier to read the code.

[Bug target/106755] Incorrect code gen for altivec intrinsics with constant inputs

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

--- Comment #4 from Peter Bergner <bergner at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #3)
> So yes there is aliasing violation as you do the store as vui128_t aka
> "__vector unsigned __int128" (which has the same aliasing set) as "__int128"
> but then do the load from it as vui32_t aka "__vector unsigned int vui32_t"
> (which is the same aliasing set as int). 

Yes, I came to the same conclusion and sent Steve a patch offline that fixes
the issue.

> -Wstrict-aliasing=3 might warn about this case I can't remember exactly of
> the levels of the options but the -Wstrict-aliasing warnings are always not
> going to happen, sometimes it is easier to read the code.

Ah, I didn't realize we could increase the warning level.  Cool!  It seems
-Wstrict-aliasing=2 is enough to catch the problem:

gcc -O3 -Wall -c -fmessage-length=0 -mcpu=power8 -mtune=power8
-Wstrict-aliasing=2  -m64 ../src/vec_bug6.c
../src/vec_bug6.c: In function ‘test_muludq’:
../src/vec_bug6.c:969:45: warning: dereferencing type-punned pointer will break
strict-aliasing rules [-Wstrict-aliasing]
  969 |   k = (vui32_t) test_vec_muludq((vui128_t* )&l, (vui128_t)i,
(vui128_t)j);
...

[Bug target/106755] Incorrect code gen for altivec intrinsics with constant inputs

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106755

--- Comment #5 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to Peter Bergner from comment #2)
> So the tests (I've removed all static inline usage and always use
> -fno-inline) pass with -O1 and fail with -O2 and -O3.  Looking at all of the
> optimizations enabled by -O2 that are not in -O1 and using -fno-* for them,
> the only option that allows the tests to pass with -O2 is
> -fno-strict-aliasing.  That said, -Wall and -Wstrict-aliasing do not flag
> any warnings with the code.  I suppose they could miss some issues in the
> test case code???

There are 10 cases (in just the .cc files) that use "optimize > 1", and
another 5 that do "optimize >= 2".

The -O options do more than just being shorthand for some -f option selections.