[FMV][AArch64] Fix UPASS when "SSBS not fully self-synchronizing".

[labrinea]

When running try_ssbs2() on hardware which is affected by the "SSBS not fully self-synchronizing" errata, the linux kernel mutes the detection of ssbs2 via hardware caps. As a result the default version ends up running the ssbs2 code which was expected to trap originally.

To work around this UPASS I am passing an additional macro parameter to indicate whether the default version is exempt from diagnostics.

This fixes llvm/llvm-project#109304.

[nikic]

LGTM

[pawosm-arm]

@labrinea I'm seeing UPASS on more features when running this test on neoverse-v1 and neoverse-v2 running RHEL8 in a docker image:

• on neoverse-v1 these are: dpb2, bf16, i8mm, rng (/proc/cpuinfo features: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs paca pacg dcpodp svei8mm svebf16 i8mm bf16 dgh rng)

• on neoverse-v2 these are: dpb2, bf16, i8mm, rng, bti, frintts, sve2, sve2-aes, sve2-bitperm, sve2-sha3 (/proc/cpuinfo features: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti)

Of course changing the parameter value from false to true for each of these features make this test pass, but I suppose it may not be the right thing to do.

[labrinea]

@pawosm-arm without looking at any logs I can only guess what's going on. We have a builder in place with similar hardware to yours https://lab.llvm.org/buildbot/#/builders/198 which makes me think the problem may relate to your operating system or the command line used. This issue may relate to yours llvm/llvm-project#123383.

[pawosm-arm]

@pawosm-arm without looking at any logs I can only guess what's going on.

There are no logs. I don't need to run entire suite for this to reproduce it:

$ clang -DNDEBUG  -O3 -fmath-errno    -O3 -DNDEBUG --rtlib=compiler-rt -o acle-fmv-features acle-fmv-features.c

$ ./acle-fmv-features
flagm
flagm2
        UPASS
dotprod
sha3
rdm
lse
sha2
aes
rcpc
rcpc2
fcma
jscvt
dpb
dpb2
        UPASS
bf16
        UPASS
i8mm
        UPASS
dit
fp16
ssbs
bti
        UPASS
simd
fp
crc
sme
sme2
f32mm
f64mm
fp16fml
frintts
        UPASS
rcpc3
rng
        UPASS
sve
sve2
        UPASS
sve2-aes
        UPASS
sve2-bitperm
        UPASS
sve2-sha3
        UPASS
sve2-sm4
wfxt
sb
sm4
sme-f64f64
sme-i16i64
mops
memtag
cssc

[labrinea]

What happens if you specify -mcpu=neoverse-v2 (or v1 depending on which machine you're running this on)? Any better?

[pawosm-arm]

What happens if you specify -mcpu=neoverse-v2 (or v1 depending on which machine you're running this on)? Any better?

Tried this on neoverse-v2. Doesn't seem to be any different:

$ clang -DNDEBUG  -O3 -fmath-errno    -O3 -DNDEBUG --rtlib=compiler-rt -o acle-fmv-features acle-fmv-features.c -mcpu=neoverse-v2
clang: warning: math errno enabled by '-fmath-errno' after it was implicitly disabled by '-fveclib=ArmPL', this may limit the utilization of the vector library [-Wmath-errno-enabled-with-veclib]
acle-fmv-features.c:195:55: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
  195 |         asm volatile("mrs %0, ID_AA64PFR1_EL1" : "=r"(val));
      |                                                       ^
acle-fmv-features.c:195:27: note: use constraint modifier "w"
  195 |         asm volatile("mrs %0, ID_AA64PFR1_EL1" : "=r"(val));
      |                           ^~
      |                           %w0
acle-fmv-features.c:22:13: note: expanded from macro 'CHECK'
   22 |             BODY \
      |             ^
acle-fmv-features.c:382:13: warning: omitting the parameter name in a function definition is a C23 extension [-Wc23-extensions]
  382 | int main(int, const char **) {
      |             ^
acle-fmv-features.c:382:28: warning: omitting the parameter name in a function definition is a C23 extension [-Wc23-extensions]
  382 | int main(int, const char **) {
      |                            ^
3 warnings generated.

$ ./acle-fmv-features
flagm
flagm2
	UPASS
dotprod
sha3
rdm
lse
sha2
aes
rcpc
rcpc2
fcma
jscvt
dpb
dpb2
	UPASS
bf16
i8mm
dit
fp16
ssbs
bti
simd
fp
crc
sme
sme2
f32mm
f64mm
fp16fml
frintts
	UPASS
rcpc3
rng
sve
sve2
sve2-aes
	UPASS
sve2-bitperm
	UPASS
sve2-sha3
	UPASS
sve2-sm4
wfxt
sb
sm4
sme-f64f64
sme-i16i64
mops
memtag
cssc

[pawosm-arm]

Happens also at -O1:

$ clang -DNDEBUG  -O1 -DNDEBUG --rtlib=compiler-rt -o acle-fmv-features acle-fmv-features.c
acle-fmv-features.c:195:55: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
  195 |         asm volatile("mrs %0, ID_AA64PFR1_EL1" : "=r"(val));
      |                                                       ^
acle-fmv-features.c:195:27: note: use constraint modifier "w"
  195 |         asm volatile("mrs %0, ID_AA64PFR1_EL1" : "=r"(val));
      |                           ^~
      |                           %w0
acle-fmv-features.c:22:13: note: expanded from macro 'CHECK'
   22 |             BODY \
      |             ^
acle-fmv-features.c:382:13: warning: omitting the parameter name in a function definition is a C23 extension [-Wc23-extensions]
  382 | int main(int, const char **) {
      |             ^
acle-fmv-features.c:382:28: warning: omitting the parameter name in a function definition is a C23 extension [-Wc23-extensions]
  382 | int main(int, const char **) {
      |                            ^
3 warnings generated.

$ ./acle-fmv-features
flagm
flagm2
        UPASS
dotprod
sha3
rdm
lse
sha2
aes
rcpc
rcpc2
fcma
jscvt
dpb
dpb2
        UPASS
bf16
        UPASS
i8mm
        UPASS
dit
fp16
ssbs
bti
        UPASS
simd
fp
crc
sme
sme2
f32mm
f64mm
fp16fml
frintts
        UPASS
rcpc3
rng
        UPASS
sve
sve2
        UPASS
sve2-aes
        UPASS
sve2-bitperm
        UPASS
sve2-sha3
        UPASS
sve2-sm4
wfxt
sb
sm4
sme-f64f64
sme-i16i64
mops
memtag
cssc

[pawosm-arm]

...and at -O0 too!

[pawosm-arm]

Is there anything that can be done to this issue? Our RHEL-8 CI is red for more than a month now...