radlink: performance, memory, IFC/GDK fixes (refreshed on dev@47de6bf3)

[honkstar1]

September 8 refresh

This block supersedes older base-revision and validation notes below.

• Head: ad4d21695e504237004061b2b3ba0d35560dd0c0.
• Updated dependency/base: upstream dev 47de6bf3. The PR still targets upstream dev.
• Incremental review: 47de6bf3...ad4d2169.
• Local validation: Clang/LLD 20.1.8, Release -O2; 178 passed, 0 failed/crashed, 1 excluded in linker/* torture/* base/* eval2/*. Exclusion: ms_link_icf_section_flag_eligibility (the previously noted Microsoft-linker comparison). These counts were revalidated after the portability correction; no farm/runtime-soak result is claimed.
• Linux follow-up: CI exposed missing internal linkage on lnk_obj_dist_weight_is_before after the C99 portability refresh. This is corrected in radlink: performance, memory, IFC/GDK fixes (refreshed on dev@47de6bf3) #842 and inherited by all dependent layers. Local Linux debug builds with Clang 18 and GCC 13 succeeded, and the GitHub matrix has been rerun for these heads.
• Updated with merge commits and fast-forward pushes only; no force rewrite. Existing draft status is unchanged.

Conflict resolution retains the regression-proven archive-symbol/function-override rescans. Upstream ICF offsets, the pending-input guard, single-task fast path and portability/build changes are integrated. Migrated tests preserve the upstream ICF, DEFAULTLIB and MSF coverage.

Older text is retained below for historical context. Already-integrated upstream changes are behind the refreshed merge base rather than new standalone fixes in this refresh.

---

Radlink performance, memory, and correctness series

Refreshed onto dev at b4c52780 on 2026-09-02. Head: 6fa35bcc.
This PR contains 105 ordered commits, 52 changed files, +8,258 / -1,383 lines.
Compressed OBJ support is separate in #892; it is not included here. The shared-pool work formerly proposed in #847 is consolidated here.

Reconciliation with upstream

Of the previously published 95 commits, 94 are retained: 92 are patch-equivalent, the subsection-accessor change is adapted to the new test harness, and the library-rescan regression is ported from C to a declarative .tst. git range-diff represents that test migration as delete/add, not as a retained patch.

The remaining commit, the C11/BLAKE3 CPU-feature-load build workaround, is removed because upstream b4c52780 supersedes it. The four implementations already removed in the August cleanup stay removed: the old unresolved-symbol line mapper, writable raw-symbol copies, leaf-reference node arenas, and final-layout side arrays. Upstream's new declarative tests and BLAKE3 assembly implementation are preserved. Clang builds now use llvm-ml for the assembly rather than invoking ml64.

Review areas

• Correctness and diagnostics: header-unit IFC debug-record resolution, GDK import variants, library-search symbol-state rescans, invalid-library indexing, entry-inference state, and missing DLL-import dependency diagnostics.
• Debug fidelity: folded-function records/source identity and function-scoped statics; /OPT:GCTYPES remains opt-in and disabled by default.
• Input and memory lifetime: read-only input mappings, private patch metadata, reusable debug-section windows/journals, bounded fault-heavy work, earlier input/buffer release, and arena reaping.
• Type merge and PDB scaling: unique-count sizing, compact type-index descriptors, giant-stream checkpoints, array-backed GSI/PSI buckets, streamed payloads, direct module copies, and overlapped decommit.
• REF/ICF throughput: cached relocation resolution, parallel dead-section removal, and compact refinement/associative-graph storage.
• Shared workers: opt-in cross-process fair-share governor; non-shared operation retains the barrier path. Stranded-grant/semaphore failure paths do not perform potentially blocking logging.

Newly included, individually reviewable fixes

• f555a9dd: emit BigObj for oversized generated import objects, with writer and 22,000-import regressions. This entire commit applies cleanly directly to upstream b4c52780, including both tests; it is independent of the rest of this series and of compressed OBJ support.
• 2aa63c01: index function-override aliases instead of repeatedly scanning ordinary aliases; declarative regression included.
• c71da326: park the governor once demand is covered; repeated queued/tiny/barrier-cohort regression.
• 1076b06c: correct PDB summary timer brackets and check the reported module interval against the enclosing invocation.
• a721072c: correct the test sort-callback signature for Clang.
• 34587780, 9f9a097d, c7f33b3e, 4353099d: consolidate the existing general diagnostic/shared-pool hardening from the compressed-object stack into this PR.

Fresh validation of this revision

• Windows x64, Clang/LLVM 20.1.8: release and debug/assert-enabled builds succeed.
• Release suite: 158 passed, 0 failed, 0 crashed, 12 skipped.
• Debug suite: 158 passed, 0 failed, 0 crashed, 12 skipped.
• The 12 exclusions are the MSVC-invoking cases: generic_build_previous_exit_condition, ms_link_icf_section_flag_eligibility, determ_test, icf_cpp_identical_functions_fold, icf_cpp_multihop_functions_do_not_fold, icf_llvm_addrsig_external_symbol, icf_llvm_addrsig, import_export, infer_asan, ms_link_icfs_identical_comdats, generic_build_default_output, and generic_build_output_with_spaces. No local MSVC compiler/linker was used for this refresh; GitHub runs its existing compiler matrix automatically.
• Raw-input links from this head participate in a 24-run, same-path EXE/PDB byte-parity matrix against the previous integration binary and the ordinary/Oodle radlink: portable compressed OBJ support (restacked on #842) #892 builds. Every linked fixture executes successfully. These are small correctness fixtures, not a production benchmark.

Outstanding validation and scope limits

Large-input ICF nondeterminism remains unresolved. A September 2 frozen raw-OBJ reproduction produces differing image layout/PDB bytes with ICF enabled; a rebuilt pre-merge b93924f9 control also reproduces it, while its ICF-disabled pair is byte-identical. It is not yet established whether the root cause is upstream ICF or retained PR work. Successful small tests do not close this issue.

The exact refreshed heads have not been rerun against full Common/Engine inputs, launched in the editor, or validated across a Linux/Wine/UBA farm. Earlier large-link measurements belong to older revisions and are not fresh performance or determinism claims for this head. Production process-exit measurements must include the post-Process/End kernel teardown tail; printed phase totals are not that measurement.

The recent relocation-decode-window experiment, independent /RAD_UNMAP_WORKERS tuning, and sharded compressed-cache backing are deliberately not included in either refreshed PR. They remain separate pending memory, performance, and production-path validation.

GitHub CI follow-up

The automatic matrix is not green. All 26 build-only jobs have passed; at the latest inspection three debug run jobs failed and two release run jobs were still running.

• Windows Clang debug reproduces upstream's GetOpenFileNameW debugger link failure and ASan _ITERATOR_DEBUG_LEVEL mismatch in the mule fixture. Windows MSVC debug reaches the existing ms_link_icf_section_flag_eligibility failure. Those failures also occur in the upstream b4c52780 run.
• Linux now gets past upstream's incompatible sort-callback compile error, fixed here. It stops in unchanged delay_import.tst because libcmt.lib is unavailable, followed by DLL entry-point inference failure. The upstream Linux run failed earlier at compilation, so no passing upstream Linux runtime control is claimed.
• The local supported-subset results above remain separate from the full CI result. No failing test has been disabled or weakened in these PRs to hide the CI failures.