Reconstruct disconnected owned conditional blocks

Failure domain and proof

Let V be the blocks owned by a FunctionDefinition and R its entry-rooted executable successor traversal. The dominator-tree domain D equals R; declarative merge/body/update links do not add executable edges. In general R is a subset of V. CFG canonicalization can disconnect an old block without releasing it. That block remains part of the public XIR module contract.

The verifier and final residual-branch check quantify over V, but the remaining-divergent conditional index previously covered only R. Thus a raw conditional in V minus R could never enter the rewrite worklist and would always fail the final check. More fixed-point iterations cannot repair this domain mismatch. Indexed-branch reconstruction already covers V.

The reduced example has four blocks and one boolean argument:

entry: return
dead_header: branch condition, dead_true, dead_false
dead_true: return
dead_false: return

Both transactional and in-place restructuring reject it before the repair, reporting one remaining conditional and no invalid or irreducible construct. No renderer, GPU, floating-point computation or coroutine ABI is required.

Minimal repair

Index every owned block’s structured roles and conditional candidates. Keep the old traversal order for R, then append V minus R in ownership order, so live-CFG rewrite ordering is unchanged. The index and its fresh-analysis oracle now use the same domain as the final contract.

A header outside D has no executable lexical dominance context. As in indexed-branch reconstruction, it receives a fresh synthetic unreachable merge instead of adopting a global post-dominator that might belong to live code. Its condition, payload and executable branch targets are unchanged. No block is deleted and no new entry-reachable edge is introduced.

Each rewrite still consumes exactly one original raw conditional and adds only an If plus branch/unreachable blocks. No new raw conditional or loop role is created, so the existing finite drain and exact dominance/post- dominance overlay contracts remain applicable. Structural validation is not weakened, and no application-specific compiler setting is introduced.

Permanent tests

test_xir_pass_restructure_cfg_owned_blocks is registered with CMake and xmake. It checks both public mutation modes, the four-block red case, idempotence, retained block/instruction identities and strict output verification. A second fixture has seventeen disconnected selections whose nontrivial store arms converge on live entry code. It checks that no live merge is claimed, no store is dropped and no executable edge is retargeted. Both fixtures enable the fresh remaining-divergent index/post-dominator oracle; 310 assertions pass, also with intermediate verification enabled.

cmake --build build-codex-xir --parallel 32
ctest --test-dir build-codex-xir --parallel 32 --output-on-failure \
  -L '^(unit_xir|unit_coro)$'
env LUISA_XIR_VERIFY_INTERMEDIATE=1 \
  ./build-codex-xir/bin/test_xir_pass_restructure_cfg_owned_blocks

The all-target build succeeds and the XIR/coroutine host selection is 70/70. Do not use an unanchored unit_coro label as a host-only selection: it also matches the separately labeled unit_coro_runtime device test.

An exploratory all-unit run was 149/151. The remaining assertions are in test_spirv_xir_dialect (old special-register enum endpoint) and test_spirv_target_feature_codegen (a required optimized instruction spelling), not the restructuring tests. A separate test-only correction extends both enum scans through the appended raster depth/builtin tags. The constant fixture’s actual SPIR-V retains all four signed 16-bit literals and selects between them, so it now checks those exact payloads plus the existing positive/negative storage-feature requirements instead of insisting on OpConstantComposite. No code-generation or numerical policy changes. The subsequent all-unit run passes 151/151 (luisa-unit-all-v2.log); spirv-constant-red-dump.log retains the original emitted selection chain.

Original complete application

The triggering Psycles change restores the original Cycles post-lamp INTERSECT_CLOSEST stage, including a fresh BVH traversal and the last lamp’s ray-self identity. Its small production-pipeline HIP regression initially passed, but full Barbershop failed during continuation 5 restructuring: one residual conditional at owned block 242, with targets 263 and 264. All three blocks are absent from the final entry-rooted structural dump.

After this generic repair, the unchanged full original application compiles and renders Barbershop at 2048x858, 64 spp, with the index oracle enabled. Its six-stage frame remains 93 fields / 416 bytes. Actual closest visits are 444,263,321 versus original Cycles 444,271,208. This is a correctness and whole-module compilation gate, not a renderer-speedup claim. The same small production routing regression also passes HIP and fallback with SoA/AoS, capacity refills and repeated dispatches.

Local evidence is /var/tmp/psycles-lamp-routing-sCLxKs: psycles.log, restructure-trace.log, restructure-dump-v2.log, xir-owned-red.log, xir-owned-expanded-green.log, xir-coro-all-v2.log, luisa-unit-all.log, xir-full-build.log, psycles-fixed.log, and final-focused-{hip,fallback}.log. Large-function dump instrumentation used to diagnose the failure has been removed. No diagnostic switch or application workaround is part of the fix.