Repository navigation
Affine.access carries neither scalar gatedness nor local effects, so every between-two-points hazard question re-walks the raw code #1016
Description
Grounded in lukstafi#750 (footprint-scoped materialization, #616). Its eligibility query asks one question — "between the producer's last write and the reader, and inside the producer's own span, what else runs?" — and the review surfaced its answer one member per round, because Affine.access (the routine's access relations, Low_level.affine_accesses) cannot say: round 1 tensor writes between producer and reader (fixed by placing the prologue at the producer), round 2 a sibling write inside the producer's statement, round 3 a write between two accumulating components, round 4 a Set_local between components, round 5 a Set_local inside the producer's statement and dead loops counted as writers, round 7 a read under a Where arm (a_guarded is the If statement only; scalar gates are invisible), round 10 a local written by the reader's statement. The query now carries three private walks over the raw code beside the access list — effectful / local_effects (statements with any effect, and with a local effect) and gated_readers (an Access_fold pass with gated_operands = Visit) — and so do the fan-in guard's reads_of_proc, computation_reads_merge and template_leaves, each a hand-rolled variant of the same read.
What a fix would touch: give Affine.access an a_gated : bool (the Access_fold context already computes it) and either emit local writes and declarations as effect rows of the same list (a Set_local is an access of a scope-local "node") or expose a sibling statement_effects view from the same walk in affine_accesses; then the eligibility query's three walks, and the guard's/merge scan's/leaves' hand-rolled reads, collapse into queries over one relation. The reviewer's genre — "an effect the access relations do not carry" — becomes unrepresentable rather than one-more-arm.
Activity
lukstafi commentedon Sep 19, 2026
Another concrete instance landed in staging#758, resolving #1001:
Online_softmax.has_opaquemust classifyTile_mmaitself as a code-motion barrier. Descending only into its fallback missed that hazard. The fix adds the explicit classification, but leaves another raw-statement walk beside the write census. This supports the shared effect/hazard representation proposed here.lukstafi commentedon Sep 25, 2026
Wave coordinator (2026-09-25 fleet wave), decision gate: proceeding on the sequencing plan's recommendation unless vetoed here before the worker's PR opens — a sibling
statement_effectsview built from the same walk, witha_gatedadded onAffine.access, rather than making local writes and declarations effect rows ofaccess. ASet_localis not an access of a tensor node, and the four hand-rolled reads want "what else runs between two points", which a per-statement effect row answers directly. Evidence: the four reads collapse onto the view, with cc and Metal parity on the footprint tests.lukstafi commentedon Sep 25, 2026
Implementation PR: lukstafi#786. It follows the decision comment:
a_gatedonAffine.access, and a siblingAffine.statement_effectview from the same walk (Low_level.affine_relations). The footprint eligibility query,template_leaves,computation_reads_mergeandOnline_softmax.has_opaquenow query that pair.One departure from the comment: the fan-in guard's
reads_of_procstays a walk. It is a cost count that charges the widerWherearm, the same per-arm maximumtrace_node_factsuses, and a boolean gate cannot say which arm a read belongs to. The PR gives the full rationale and the executed parity on cc and Metal.Merge pull request #786 from lukstafi/claude/issue-1016-affine-effect…
lukstafi commentedon Sep 25, 2026
Landed in lukstafi#786, merge commit
b1e12c8ca. Codex approved it in round 1 with no findings. Every PR-matrix check on head7963ca6cwas green.What landed
Affine.access.a_gated: a read under a scalar gate (aWherearm or a gated operand), following theAccess_foldconvention. ALocal_scopebody inside a gate is hoisted, so its reads are not gated.Affine.statement_effectrows, built by the same walk (Low_level.affine_relations;affine_accessesandstatement_effectsare its two halves). Kinds:Local_write,Local_declare,Scope_body,Barrier,Staged,Mma,Merge_read. Helpers:Affine.loops_live,in_scope_body,within_statement.footprint_eligibility_query(its three private walks),template_leaves,computation_reads_mergeandOnline_softmax.has_opaque. Each classifies effect kinds with an exhaustive match, so adding a kind forces every consumer to handle it.docs/agent-notes/lowering-and-analysis.md. The footprint paragraph invirtualization-and-inlining.mdnow namesa_gatedand the effect rows.Parity and controls. Executed parity passed on cc and Metal for footprint_materialization, footprint_diagonal_einsum, online_softmax (plus fast_math) and virtual_chain_fanin. Negative controls: breaking any one of the four readers makes its own test rows fail.
Deliberately not done
reads_of_procstays a walk. It is a cost count that charges the widerWherearm, the same per-arm maximumtrace_node_factsuses, and a boolean gate cannot say which arm a read belongs to.writes_of_stmt/reads_of_body) stays hand-rolled, soScope_bodystays opaque there.Pre-existing gap found (follow-up proposed to the wave coordinator): footprint_materialization's
inherited-localrow does not isolate the round-10 "reader statement writes a local" check. The row still passes with the check disabled, because a later retraction to inlining covers the case.Merge pull request #1016 from lukstafi/claude/w1005-1218-scalar-host-…