1757406516, 1757862468, 1757862472

[0xjunha]

It seems these traces have some PVM invocations with Panic or PageFault exit reasons for Ψ1 due to invalid memory read/write.
According to accumulate_gas_used field in statistics state of the traces, it seems those cases charge gas for the failing instructions (which attempted invalid memory accesses), whereas GP states they should not mutate VM state at all, immediately exiting with Panic or PageFault.

GP reference: https://graypaper.fluffylabs.dev/#/38c4e62/251901251901?v=0.7.0

If this was ambiguous in v0.7.0, it is clarified in v0.7.2:
https://graypaper.fluffylabs.dev/#/ab2cdbd/256c01256c01?v=0.7.2

[clearloop]

If I recall correctly, the basic PVM tests have a lot of this case as well

[davxy]

which one?

[clearloop]

most of the *_nok tests are broken at memory operation and charges the gas

[davxy]

Sorry but I don't understand. What/where are the *_nok cases? Can you point me to one of these tests? Are in the jam-test-vectors repo?

[0xjunha]

I guess @clearloop referred to https://github.com/koute/jamtestvectors/blob/master_pvm_initial/pvm/TESTCASES.md

[davxy]

@0xjunha indeed according to A.8 the gas is left untouched.
However, I think that makes more sense to charge for the instruction that triggered the fault.

Begin of section A.1

Assuming the program blob is valid (which can be validated statically), some gas is always charged whenever execution
is attempted. This is the case even if no instruction is effectively executed and machine state is unchanged (i.e. the result
state is equal to the parameter).

Otherwise no gas is charged if the very first instruction is one that triggers a trap for memory access.

[0xjunha]

@0xjunha indeed according to A.8 the gas is left untouched. However, I think that makes more sense to charge for the instruction that triggered the fault.

Begin of section A.1

Assuming the program blob is valid (which can be validated statically), some gas is always charged whenever execution
is attempted. This is the case even if no instruction is effectively executed and machine state is unchanged (i.e. the result
state is equal to the parameter).

Otherwise no gas is charged if the very first instruction is one that triggers a trap for memory access.

Makes sense. Thank you for the input and the PR!

[davxy]

gavofyork/graypaper#497