compiler: Fix stale reads and lost writes in buffering

[ggorman]

Stacked on #3062 (which is on #3061); the base switches as they merge.

Bugs predating #3050, found against opt='noop'. Every test program has at most one firing write per step, slot and point, except the last one, which says so.

1. ir: Let GuardBoundNext be negated. BaseGuardBoundNext._new did not accept evaluate, so Guard.negated raised a TypeError for bound guards, and so did any And over guards containing one (SymPy negates the relational arguments of an And to detect contradictions). Latent on main; commit 4 builds such an And for every buffered Function, which meets the bound guards of prefetches. Test test_ir.py::TestGuards::test_guard_bound_next_negated.
2. Read-before-write and read-ahead slots were stale. g = f; f = T read the slot left by the previous iteration: with f.data[:] = 10 and eqs = [Eq(T, time), Eq(g, f), Eq(f, T)], main gives g = [10, 0, 1], noop [10, 10, 10]. Such slots are now loaded from the Function and the leading written slot is copied back. Tests test_read_before_write, test_read_ahead.
3. Non-consecutive Clusters with equal guards formed distinct groups (itertools.groupby); the buffer was sized from the last group only. Test test_non_consecutive_guards.
4. Reads after writes under other guards now share the writers' buffer, loaded (slots behind the written index included) before the first access. A load keeps only the guard conjuncts that read nothing written in the loop; an access under a guard reading a Function does not count as having loaded or written the slot; no sharing at a non-constant slot distance or across different spatial Dimensions/SubDomains (separate buffers instead of a CompilationError). Main-wrong examples: test_mutable_read_guards[repeated] gives h = [0, 2, 20, 30] vs noop [1, 2, 20, 30]; test_guard_changes_between_accesses[write-read|read-read] (with buf-async-degree=2 main reads uninitialised memory). Tests test_read_write_different_guards, test_read_write_shifted_slot, test_read_behind_guarded_write, test_read_write_different_subdomains, test_guard_changes_between_accesses, test_mutable_read_guards.
5. Guards mutated between writes are refused with a CompilationError. Test test_mutable_guards[stable|mutable|indirect].
6. Several written indices per iteration (f[time+1] and f[time]) are refused with a CompilationError. Test test_write_multi_index.
7. A copy-back is not skipped if the slot is read in between. The _overwritten docstring example, a subsampled usave save plus a last-step save with g = usave under the last-step guard between them, reads the slot from storage (commit 4), which the skipped copy-back left unwritten: g = 0 instead of 6. This slot has two firing writes at that step; the test asserts the read between them and the final value, which both follow program order. Test test_slot_read_between_writes.

Each commit's tests fail with only its production change reverted to its parent. The buffering, dimension, iet, operator and ir suites pass (679 passed, 1 xfail).

Depends on #3062: with its production change reverted on this head, test_mutable_read_guards[after] gives f = [0, 2, 20, 30] vs noop [1, 2, 20, 30].

Validation beyond the unit tests: a differential review of 1,521 generated programs (each under noop, buffering and buf-async-degree=2) found no violation and no case where this branch is worse than main. Two follow-ups are left out: 8 programs with two syntactic writers under a guard reading a Function written between them are refused although a baseline computes them, and 80 programs are wrong on main as well. On an H200 node the downstream GPU test suites (serial and 2 MPI ranks) give failure sets identical to main.

🤖 Generated with Claude Code