[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[mkuvyrkov at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

--- Comment #5 from Maxim Kuvyrkov  ---
Looking at this problem more, I think the issue is due to ARM target trying
hard to avoid UNSUPPORTED tests, instead of embracing them.

For the vectorization NEON check we have ...
===
proc check_effective_target_arm_neon_ok_nocache { } {
global et_arm_neon_flags
set et_arm_neon_flags ""
if { [check_effective_target_arm32] } {
foreach flags {"" "-mfloat-abi=softfp" "-mfpu=neon" "-mfpu=neon
-mfloat-abi=softfp" "-mfpu=neon -mfloat-abi=softfp -march=armv7-a"
"-mfloat-abi=hard" "-mfpu=neon -mfloat-abi=hard" "-mfpu=neon -mfloat-abi=hard
-march=armv7-a"} {
if { [check_no_compiler_messages_nocache arm_neon_ok object {
#include 
...
===
... where target tries to find a set of flags compatible with _any_ of the
built multilibs to run the testsuite.

I think this is excessive, since each multilib should be tested on its own
merits, and if armv7-m does not support vectorization, there should be no
effort to try and switch to armv7-a or armv8-m+mve multilib in order to run
vectorization tests.  In other words, vectorization tests should be marked
UNSUPPORTED in armv7-m, and PASS/FAIL in armv7-a and/or armv8-m+mve.

In practical terms, my proposed solution to this problem is to remove all
"foreach flags" options except for the default "".

ARM maintainers, what am I missing?

[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[clyon at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

--- Comment #4 from Christophe Lyon  ---
I'm wondering whether you missed check_effective_target_arm_arch_FUNC_link and
friends?

Maybe check_effective_target_arm_arch_v7a_neon_link would work here, but it
does not use the exact same flags.

[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[mkuvyrkov at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

--- Comment #3 from Maxim Kuvyrkov  ---
Changing from compile-only to link test is as simple as changing "object" to
"executable" in
[check_no_compiler_messages_nocache arm_neon_ok object ...]
.

However, ... this pattern of checking for ARM architectural features is shared
by 20+ check_effective_target_arm_* routines.  IMO, we should either update all
of these to be link tests (unless there is a good reason to keep them as
compile-only that we can document in the comments).  Or just accept this
vectorization test failure on ARM targets that don't support vectorization.

[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[clyon at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

--- Comment #2 from Christophe Lyon  ---
I think the last -march option overrides the previous one(s).

I'd say the test should use an effective-target which checks that linking is
actually OK rather than just a compile OK test. Not sure if an adequate one
already exists, but there are already plenty :-)

[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[mkuvyrkov at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

Maxim Kuvyrkov  changed:

   What|Removed |Added

 CC||clyon at gcc dot gnu.org,
   ||rearnsha at gcc dot gnu.org

--- Comment #1 from Maxim Kuvyrkov  ---
The test now fails with linker error:
.../arm-eabi/bin/ld: error: /tmp/cc2Q27GE.o: conflicting architecture profiles
A/M

This is due to command line having
-mthumb -march=armv7-m -mtune=cortex-m3 -mfloat-abi=softfp -mfpu=auto ...
-mfpu=neon -mfloat-abi=softfp -march=armv7-a

The first part comes from toolchain configuration settings, and the second part
(-mfpu=neon -mfloat-abi=softfp -march=armv7-a) comes from
check_effective_target_arm_neon_ok_nocache().

Surprisingly (to me), GCC accepts such mixed options, which makes
check_effective_target_arm_neon_ok_nocache() succeed, since it's only doing a
compilation test.  The linker, though, fails.

Richard E., it is expected that GCC accepts conflicting -march= options?

[Bug testsuite/114568] [14 regression] g++.dg/vect/pr84556.cc fails to produce executable since r14-9706-gb8e7aaaf350a45

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114568

Richard Biener  changed:

   What|Removed |Added

   Target Milestone|--- |14.0
 Target||arm