`ad548bb`

Burn down final audit xfails

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 1 month ago

SHA: ad548bbc108457cf2b8e8b4c559fdf383cc06df8
Parents: 204993d
Tree: b0049e2

4 changed files

Status	File	+	-
M	`.docs/sprints/sprint29_11.md`	19	10
M	`noted_items.md`	0	15
M	`test_programs/realworld_assumed_shape_size.f90`	0	1
M	`test_programs/realworld_suffix_scan.f90`	1	5

.docs/sprints/sprint29_11.mdmodified

  Kickoff items:
  - source-level audit for helper-before-program entry lowering / dead-function root handling
 -- living XFAIL for module procedure host-association over module globals
 +- source-level audit for module procedure host-association over module globals
  - real-world stdlib-style kernel:
    - tridiagonal sparse matvec (`realworld_tridiag_spmv.f90`)
  - real-world BLAS-style kernel:
    real-world default-kind array-query assignments could produce invalid asm like
    `mov w21, x20`; the new runtime-shape audit keeps that truncation surface
    honest
 +- fixed: fpm-style suffix classification in `realworld_suffix_scan.f90` first
 +  needed fixed-length character arrays to lower as real element-addressable
 +  arrays instead of scalar strings, then needed scalar character dummy intrinsics
 +  like `INDEX(name, suffix)` to carry a real runtime length story through
 +  `afs_c_strlen`, and finally exposed two optimizer-side bugs: mixed-width GEP
 +  offsets were being compared without pointee-size scaling in alias analysis,
 +  and SROA was scalarizing aggregates even when a GEP address escaped through
 +  the synthesized descriptor. The program is now a passing audit probe with
 +  cross-opt equality plus IR/asm/object/run reproducibility
 +- fixed: dummy-array `SIZE(...)` queries in `realworld_assumed_shape_size.f90`
 +  were not actually using assumed-shape descriptors because bare `(:)` dummy
 +  declarations lowered through the `Deferred` array-spec path, and then the
 +  synthesized descriptor lost `rank` and `flags` at `-O2+` because mixed-width
 +  GEP offsets in alias analysis made descriptor field stores look like they
 +  overlapped. Assumed-shape dummies now classify as descriptor-backed, query
 +  results are kept in default-kind scalars, and the real-world canary now
 +  passes across optimization levels with deterministic artifacts
  - proven: LICM hoists invariant scalar dummy loads out of a real-world affine
    update loop in `realworld_affine_shift.f90` once BCE clears the loop body
  - proven: GVN reduces duplicated branch-join PURE helper calls in
    (`realworld_ipo_chain.f90`), small-loop DO CONCURRENT exploitation
    (`realworld_doconc_square.f90`), and explicit-DO vectorization onto the bulk
    runtime kernels (`realworld_vector_stage.f90`)
 -- deferred with living XFAIL: `FPM-SUFFIX-1` fpm-style source suffix scan still
 -  fails after parsing with an `i32`/`i64` IR store mismatch
 -- deferred with living XFAIL: `ASHAPE-SIZE-1` dummy-array `SIZE(...)` lowering
 -  still routes dummy arrays into the descriptor runtime path even though ordinary
 -  dummy arrays are carried as base pointers today. The default-integer query
 -  typing bug is now fixed, so the remaining failure is the real descriptor
 -  mismatch: `realworld_assumed_shape_size.f90` reaches `afs_array_size` with
 -  bogus dummy-array metadata and panics in the runtime
  - separately deferred parser gap: typed character array constructors using an
    explicit type-spec inside `[]`
  - `172` programs with `CHECK`
  - `30` programs with `IR_CHECK`
  - `6` programs with `IR_NOT`
 -- `2` living `XFAIL`s
 +- `0` living `XFAIL`s
  ## Brutal Audit Priorities

noted_items.mdmodified

    The full-sprint audit tripped this while building the fpm-inspired
    `realworld_suffix_scan.f90` reproducer, which is currently written in a more
    conservative source form instead of the typed constructor spelling.
 -- Audit FPM-SUFFIX-1: the conservative `realworld_suffix_scan.f90` variant now
 -  gets past parsing but still fails IR verification with an `i32`/`i64` store
 -  type mismatch.
 -  This is now a living XFAIL in the real-world audit corpus, so the Sprint 29
 -  closeout has a standing canary for the remaining source-scanner gap.
 -- Audit ASHAPE-SIZE-1: dummy-array `SIZE(...)` queries still flow into the
 -  descriptor runtime path even though ordinary dummy arrays are currently
 -  lowered as base pointers rather than descriptors.
 -  The mixed-width query-result bug that originally masked this is now fixed, so
 -  the remaining failure is the real one: `afs_array_size` receives bogus
 -  dummy-array metadata and panics in the runtime.
 -  The 29.11 claims audit surfaced this first through the original
 -  `realworld_ipo_chain.f90` helper loop, and
 -  `test_programs/realworld_assumed_shape_size.f90` now captures it as a living
 -  XFAIL canary instead of letting the finding disappear.
  - Revisit ambitious array/vectorization-style `integer(16)` rewrites only after the
    scalar/runtime ABI surface is fully closed and audited.

test_programs/realworld_assumed_shape_size.f90modified

  ! Real-world-style assumed-shape SIZE query canary.
 -! XFAIL: audit ASHAPE-SIZE-1 (assumed-shape dummy arrays still flow into afs_array_size without a descriptor)
  ! CHECK: 6
  ! REPRO_CHECK: asm
  ! REPRO_CHECK: obj

test_programs/realworld_suffix_scan.f90modified

  ! fpm-inspired source suffix classifier shape.
  ! Conservatively written because the typed constructor spelling
  ! `[character(len=20) :: ...]` is still a separate parser-gap note.
 -! XFAIL: audit FPM-SUFFIX-1 (fpm-style suffix scan hits i32/i64 IR store mismatch)
 -! CHECK: 3
 -! CHECK: 2
 -! CHECK: 2
 -! CHECK: 322
 +! CHECK: 3 2 2 322
  ! IR_CHECK: call @classify_sources(
  ! REPRO_CHECK: asm
  ! REPRO_CHECK: obj