LAPACKE test suite via the extended-API source generator

[ACSimon33]

Summary

LAPACKE currently has no systematic test suite. Instead of copying the Fortran LAPACK test suite and hand-renaming symbols (the CBLAS-testing approach), reuse the extended-API source generator (CMAKE/Generate64BitSuffixedSource.cmake) to rewrite the existing TESTING/LIN and TESTING/EIG sources so that calls to LAPACKE-covered routines become calls to _TEST-suffixed symbols (e.g. DGELS → DGELS_TEST). Those symbols are provided by a library of generated C wrappers that forward to LAPACKE_*_work, in both column-major and row-major flavors. The unmodified numerical verification and error-exit machinery of the Fortran suite then tests LAPACKE.

This is the same pattern already used to build the _64 test executables (xlintstd_64, xeigtstd_64, ...), so the generator is proven on exactly these sources.

Design

TESTING/LIN/*.f --(generator: SUFFIX=_TEST, allowlist, NO_STRING_REPLACEMENTS)--> renamed sources
                                                                                  (calls DGELS_TEST; DLATMS/ILAENV/harness stay native)
xlintstd_lapacke_{cm,rm} = renamed test objects (compiled once, OBJECT library)
                         + liblapacke_test_wrappers_{cm,rm} (generated C, one source compiled twice)
                         + lapacke + tmglib + lapack + blas

Each generated wrapper has the Fortran-mangled name (dgels_test_), the exact Fortran calling convention (hidden strlen args declared and ignored, reusing the lapack.h macros), and:

• CM flavor: faithful passthrough to LAPACKE_dgels_work(LAPACK_COL_MAJOR, ...) — preserves workspace queries, minimal-workspace driver loops, and LWORK error exits.
• RM flavor: allocates row-major shadow copies of matrix arguments, converts with LAPACKE's own LAPACKE_?{ge,tr,sy,po,gb,tb,tp,pp,...}_trans helpers, calls the _work function with LAPACK_ROW_MAJOR, converts back.
• Info mapping: *info = ret < 0 ? ret + 1 : ret (undo the layout-argument shift); on any negative return, additionally call the interposed xerbla_ with the Fortran argument number (benign duplicate for Fortran-detected errors, the only notification for LAPACKE-detected ones).

Wrappers target the _work (middle-level) API, not the high-level API: high-level functions ignore caller WORK/LWORK, which breaks LWORK error-exit tests and workspace-stress loops. High-level/NaN-check coverage is a later opt-in flavor.

Row-major runs must disable error-exit tests (TSTERR = F input-file variants, generated at configure time): the xERR* routines encode column-major LD* semantics, and a transposing wrapper never forwards the caller's deliberately-bogus LDA. LAPACKE's RM-specific ld checks get a small dedicated C unit test instead.

Tasks

Phase 0 — Spike (validate all risky assumptions end to end)

• Parameterize the generator: SUFFIX= (default _64) and SYMBOL_ALLOWLIST_FILE= options in Generate64BitSuffixedSource.cmake / ExtendedAPIHelpers.cmake, back-compat preserved
• Hand-write ~10 double-precision wrappers (dgetrf/s, dpotrf, dgels, dgeqrf, dorgqr, dormqr, dlange)
• Build xlintstd_lapacke_cm with a restricted allowlist; run dtest.in (GE/PO/QR paths) green, error-exit tests included

Phase 1 — Wrapper generator + full column-major LIN

• Python generator: parse lapack.h (LAPACK_x_base) + lapacke.h (LAPACKE_x_work) prototypes; self-validating 1:1 argument zip; manual-override table for nonuniform interfaces; one generated file per routine, committed (LAPACKE precedent)
• Handle: value-returning functions (?lange), no-INFO routines (?lacpy), callbacks (?gees/?gges SELECT), output char* args (EQUED)
• Allowlist derivation at configure time (LAPACKE _work base names minus exclusions/local redefinitions)
• LAPACKE_TESTING CMake option (requires LAPACKE=ON; CMake-only, like _64)
• xlintstd_lapacke_cm green on full dtest.in; then s/c/z
• Post-build nm audit: report allowlist symbols referenced but not renamed (missing EXTERNAL declarations)
• Extend add_lapack_test + teach lapack_testing.py the new output names

Phase 2 — Row-major flavor

• RM shape-lifting from lapacke_*_work.c RM branches; CM-fallback + gap report for unparseable routines
• Generate TSTERR = F input-file variants at configure time
• xlintst{s,d,c,z}_lapacke_rm runs; triage failures (expect genuine LAPACKE RM bugs — that is the payoff)
• Dedicated C unit test for LAPACKE RM ld error checks

Phase 3 — Remaining suites

• EIG (xeigtst*_lapacke_{cm,rm})
• RFP (xlintstrf*), mixed precision (xlintstds/xlintstzc), DMD

Phase 4 — Future extensions

• High-level-API wrapper flavor (NaN-check + auto-allocation coverage; LWORK error tests expected-failed; LAPACKE_set_nancheck control)
• Combined _64 + LAPACKE variant (LAPACKE_64 already exists via API_SUFFIX; generator parameters compose)

I'll split this into many smaller PRs and extend the test suite bit by bit.