[Perf] Add qd.field_array(N, dtype) for indexed @qd.dataclass fields

[hughperkins]

Adds a new qd.field_array(N, dtype) annotation for @qd.dataclass that exposes a logical N-element array as obj.r[i] while storing N individually- named synthetic scalar fields (_r0..._r{N-1}) under the hood. For python-int indices (including qd.static(range(N))-unrolled loop variables), the AST transformer rewrites obj.r[i] directly to obj._r{i}, so the generated LLVM IR / PTX is byte-identical to a hand-rolled named-field struct.

Motivation: today's idiomatic r: qd.types.vector(N, dtype) group field leaves an alloca that LLVM SROA can't decompose once register pressure crosses a threshold (e.g. two concurrent tiles in a Cholesky+TRSM kernel), causing runtime regressions via local-memory spills. The named-field cascade pattern avoids the spill but balloons source size (32-way if k == N: self.rN = val write cascades duplicated at every callsite). field_array collapses those cascades to one AST node per callsite while preserving the named-field IR.

Changes:

• lang/struct.py: FieldArray type wrapper, field_array(count, dtype) constructor, expansion in StructType.__init__ (synthetic field names plus _field_groups metadata), propagation in StructType.__call__, _FieldArrayRef transient proxy.
• lang/impl.py: preserve _qd_field_groups across the Struct rewrap in expr_init.
• lang/ast/ast_transformer.py: build_Attribute returns a _FieldArrayRef for group access; build_Subscript resolves it to a direct field reference for python-int indices.
• tests/python/test_field_array.py: 5 tests covering construction, static python-int index, qd.static loop-var index, runtime-index rejection (clear error), and static-index OOB rejection.

Runtime-int indexing is intentionally rejected with a friendly error pointing at qd.static; existing cascade helpers continue to handle the runtime case by spelling out the _rN fields directly. Adding runtime-int support is a small follow-up.

Verified on a field_array port of genesis _tile32.py: PTX byte-identical to the named-field S1 baseline (modulo the per-session-nonce comment) on both chol_kernel and chol_trsm_kernel; zero local-memory spills (S1: 0/0, FA: 0/0, F4-A vector-field variant: 42/97); 25% compile-time reduction on the single-tile harness (5.60s -> 4.19s, 3-run mean). Source dropped from 1068 to 515 lines (-52%). Full writeup in perso_hugh/doc/qd_field_array_2026may23.md.

All 201 tests in test_py_dataclass.py + test_complex_struct.py + test_struct.py continue to pass; the 5 new tests pass in 1.76s total.

Issue: #

Brief Summary

copilot:summary

Walkthrough

copilot:walkthrough

[hughperkins]

Need some user-facing doc.

[hughperkins]

(Also, might want to revisit the name)

[github-actions]

Total: 4 file(s) changed, +203 -3 code lines.

[github-actions]

Diff coverage: 70% · 200 lines, 61 missing