diff --git a/artifacts/gust_dma_model.yaml b/artifacts/gust_dma_model.yaml index 1f63cb1..f6289a5 100644 --- a/artifacts/gust_dma_model.yaml +++ b/artifacts/gust_dma_model.yaml @@ -285,6 +285,20 @@ artifacts: (https://github.com/cpetig/wasm-shm-test/blob/main/DESIGN.md). cpetig is aware; this is recorded as the forward direction, not committed scope. + The framing question upstream is Luke Wagner's talk "Does the Component Model + Require Extra Copying?" (https://youtu.be/YZG7waS2XZg). That question is not + academic on an 8 KB part: a copy at the seam is a copy we cannot afford, which + is why own is specified as a TRANSFER (attach/detach carry it) and + not as a hand-off plus memcpy. Our contribution to that question is an answer + that is verified rather than argued -- the Kani-proven ownership FSM, on + silicon, in 220 B. + + Also relevant to timing: the Bytecode Alliance's Component Model 1.0 roadmap + (Eric Gregory, 2026-06-08) places shared-everything threads AFTER 1.0 + ("these advancements also pave the road for shared-everything threads later"). + The general mechanism is future work; our specific verified instance exists + now, which is the window in which contributing it is worth something. + Alignment: his "host may answer ZERO to the allocation request and use a FIXED valid address … MPU not MMU" is exactly gale's embedded reality (meld merged-memory + synth --native-pointer-abi fixed base, F100/M3 MPU, BYO-OS); @@ -301,7 +315,7 @@ artifacts: (3) gale's differentiated contribution = the VERIFIED ownership-transfer discipline (Kani-proven attach/detach FSM + barrier-pairing) + the MPU/no-MMU embedded profile as a working reference implementation. - tags: [gust, dma, shared-memory, component-model, bytecode-alliance, canonical-abi, next-step] + tags: [gust, dma, shared-memory, component-model, bytecode-alliance, canonical-abi, next-step, upstream-reference] links: - type: related-to target: SAC-DMA-SEAM