Don't you need to update SCTLR_EL1 to reflect introduced variant(s)?

@murzinv Do you mean litmus should explicitly set those bits when the variants are used? I introduced a check whether the bits are already enabled, otherwise throw an error.

@murzinv Do you mean litmus should explicitly set those bits when the variants are used? I introduced a check whether the bits are already enabled, otherwise throw an error.

Well, in current form litmus would just accept new variants, but they have no effect. IIUC, these variants corresponds to FEAT_ExS and allow to control aspects of contexts synchronization. Thus absence of any setting in litmus7 when these variants are present look odd. I do not know what are defaults for those bits in SCTLR_EL1 (architecture says the are RES1) so I'd assume litmus7 should toggle/clear them. Also it'd be handy if there is check for FEAT_ExS so user would be aware if toggling those bits make sense in the first place.

I tested this on hardware using the following test

AArch64 SVC-ifetch.exs1eis0
variant=self,exs
CacheType=DIC
{
  0:X0=instr:"NOP"; 0:X1=P0:L2;
}
 P0          | P0.F       ;
 STR W0,[X1] | ERET       ;
 DSB ISH     |;
 SVC #0      |;
L2:          |;
 B L1        |;
 MOV W9,#1   |;
L1:          |;

exists(0:X9=0)

I get the following result:

Test SVC-ifetch.exs1eis0 Allowed
Histogram (2 states)
39999 *>0:X9=0;
1     :>0:X9=1;
Ok

Witnesses
Positive: 39999, Negative: 1
Condition exists (0:X9=0) is validated
Hash=30df2009caaef631aff2dc9d9ac6a587
variant=self,exs
CacheType=DIC
Observation SVC-ifetch.exs1eis0 Sometimes 39999 1
Topology 39999 :> part=0 [[0]]
Time SVC-ifetch.exs1eis0 0.01

I opted for if variant eis then the relaxation provided by SCTLR_EL1.EIS=0 is in place. I find this a bit mind boggling, let me know if you can think of a better solution. Given this, my understanding is that without variant eis it should be guaranteed that X9=1. Since the SVC is a context synchronising event, should that not guarantee that P0:L2 is populated with a NOP? Am I missing something?

I just glaceed over herd side of things and it seems that it is done with assumption that SCTLR_EL1.{EIS,EOS}={0,0} as default. My impression is that default SCTLR_EL1.{EIS,EOS}={1,1} provides stronger guaranties, I think that inverted polarity is the reason of

mind boggling

Obviously, herd and litmus need to be aligned.