[nicole-graus]

Motivation

main has no accelerator for memcpy, memmove or memset, and proves the write syscall with a dedicated COMMIT chip. This adds the three memory operations and folds write into the same chip, following the merged COPY spec.

Description

One COPY table (41 columns, 33 bus interactions, 39 constraints, degree 3) serves memcpy/memmove, memset and write. The COMMIT chip is deleted.

Which operation a row runs is decoded, never chosen by the caller: the ECALL number is a linear function of is_set and is_write, so the three accepted numbers are exactly 64, -30 and -31 and nothing else; both flags ride inside the chain tuple, so a sequence cannot change operation half way through.

• memset is a propagating copy. The stub seeds eight bytes and calls with dst = src + 8; the chip inverts the read/write timestamps so each row observes the previous row's write. The gap is pinned in-circuit: without it a dst == src call would leave the copied value unconstrained.
• write reads its destination from x254, the running commit index, writes dst + count back, checks the file descriptor is 1 and returns count in x10.
• A chain ends on the row that moves the last bytes; there is no terminal row. end is pinned by restrict_end against the borrow out of count - step.
• The width is a free per-row choice upward, and pinned exactly downward: a wide row with fewer than eight bytes left borrows, and the borrow demands count == 0.

ECALL renumbering: memset moves to -31 to match the spec, and hint to -32. x254 becomes a two-word register (addresses 508-509), so the register file grows by one word address. Every guest ELF must be rebuilt: in-tree artifacts are gitignored and rebuild automatically, but build caches and the bench box will not.

[github-actions]

Benchmark Results for unmodified programs 🚀

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base binary_search	38.7 ± 0.6	37.6	39.5	1.00
head binary_search	39.7 ± 0.9	38.1	41.3	1.03 ± 0.03

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base bitwise_ops	39.0 ± 0.6	38.0	39.9	1.00
head bitwise_ops	39.6 ± 1.1	38.5	41.4	1.02 ± 0.03

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base fibonacci_26	40.9 ± 0.6	40.0	41.8	1.00
head fibonacci_26	41.3 ± 0.5	40.7	42.0	1.01 ± 0.02

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base matrix_multiply	41.3 ± 0.9	39.8	42.6	1.01 ± 0.02
head matrix_multiply	40.8 ± 0.4	40.2	41.5	1.00

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base modular_exp	38.9 ± 0.5	38.3	39.8	1.00
head modular_exp	41.6 ± 4.7	38.5	52.6	1.07 ± 0.12

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base quicksort	41.8 ± 0.5	40.4	42.2	1.00
head quicksort	42.0 ± 0.7	40.8	42.9	1.01 ± 0.02

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base sieve	42.0 ± 0.5	41.2	42.9	1.00
head sieve	43.1 ± 0.8	41.7	43.8	1.03 ± 0.02

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
base sum_array	49.4 ± 0.6	48.5	50.3	1.00
head sum_array	49.9 ± 0.6	48.6	50.8	1.01 ± 0.02

[nicole-graus]

/bench

[nicole-graus]

/bench-verify

[github-actions]

⏳ Benchmark started on the bench server. Two verifier arms (monolithic + continuations over an ethrex 20-tx block), then the recursion-guest cycle comparison, which adds guest builds on top — longer on a cold runner. The bench server is occupied until it finishes.