Finish M extension #434
Finish M extension #434
Conversation
feat: m extension support in tracer
| // All instructions in virtual sequence are mapped from the same | ||
| // ELF address. Thus if an instruction is virtual (and not the last one | ||
| // in its sequence), then we should *not* update the PC. | ||
| flags[11] = match self.virtual_sequence_remaining { | ||
| Some(i) => i != 0, | ||
| None => false | ||
| }; |
There was a problem hiding this comment.
We need this flag (and so we had to reverse the virtual sequence indexing) for the new non-uniform constraint, which ensures that virtual sequences are executed in full and in order
| @@ -26,6 +26,23 @@ impl JoltField for ark_bn254::Fr { | |||
| } | |||
| } | |||
|
|
|||
| fn to_u64(&self) -> Option<u64> { | |||
| @@ -190,7 +201,26 @@ impl<const WORD_SIZE: usize> VirtualInstructionSequence for DIVInstruction<WORD_ | |||
| advice_value: None, | |||
| }); | |||
|
|
|||
| virtual_sequence | |||
| virtual_trace.push(RVTraceRow { | |||
There was a problem hiding this comment.
Previously, we had elided the MOVE instruction as described in the Jolt paper as an optimization, instead writing the advice directly to the destination register
But as it turns out, we actually did need the separate virtual register + MOVE for the edge case where rd is equal to rs1 or rs2
There was a problem hiding this comment.
This was implemented as a signed ≤, but it should be an unsigned ≤
| SubtableIndices::from(C - 1), | ||
| ), | ||
| ( | ||
| // Not used for lookup, but this implicitly range-checks |
There was a problem hiding this comment.
Let's make a utility function for this so it's self documenting and use across all the memory ops that need range-checking.
fn RangeCheck(C) -> (Box<Subtable<...>>, SubtableIndices)
Finishes wiring up the M extension to Jolt: