armfortas Public

Notifications Fork 0 Star 0

Commits

trunk
Switch branches/tags
✓ trunk default
All users
mfwolffe
View commits for all users
Until Apr 9, 2026
April 2026
‹ ›
Su Mo Tu We Th Fr Sa
29 30 31 1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 1 2
3 4 5 6 7 8 9
Clear Today

Commits on April 9, 2026

  1. Fix contained dummy array lowering
    mfwolffe committed
    3df6882
  2. Fix chained mem2reg promotion
    mfwolffe committed
    9b07e0f
  3. Add GVN audit regression
    mfwolffe committed
    71d2ac0
  4. Restore GVN in O3 pipeline
    mfwolffe committed
    89b9a8f
  5. Wire array bounds checks end to end
    mfwolffe committed
    697a46b
  6. Re-enable safe global load forwarding
    mfwolffe committed
    74a593b
  7. Honor opt level in capture stages
    mfwolffe committed
    33a0d10
  8. add transfer report: GlobalLsf/BCE status, testing methodology, .refs findings
    mfwolffe committed
    958483e
  9. enable SROA + GVN in Os and O3/Ofast pipelines
    mfwolffe committed
    7a9c3c2
  10. re-enable SROA after Mem2Reg+ConstFold (GCC SRA pattern): 117 e2e pass 
    SROA runs after SSA construction and constant folding, matching
GCC's pass ordering. Second Mem2Reg promotes SROA-created scalar
allocas. Placed before inlining so contained functions also benefit.

Active O2 pipeline now has SROA + GVN enabled (21 passes).
GlobalLsf remains disabled (path-sensitivity bug).
    mfwolffe committed
    81da431
  11. re-enable GVN (after loop passes), fix GlobalLsf load-block clobber check, keep GlobalLsf disabled 
    GVN placed after all loop passes avoids unswitching SSA conflicts.
117 e2e programs pass with GVN enabled.

GlobalLsf: added load-block clobber check (calls before load prevent
cross-block forwarding). Still has path-sensitivity bug: forwards
past calls on non-dominator paths (fibonacci loop bound mismatch).
Needs proper MemorySSA-style walk (every path from store to load
must be clobber-free). Disabled pending proper implementation.
    mfwolffe committed
    f2c89ea
  12. document 29.8 lessons: SROA/GVN/GlobalLsf interaction bugs, study needed before re-enabling
    mfwolffe committed
    aa23cd5
  13. disable 29.8 passes from pipeline: SROA/GVN/GlobalLsf have interaction bugs 
    SROA + second Mem2Reg triggers unswitching SSA errors on nested loops.
GlobalLsf incorrectly forwards stores across calls that clobber memory.
GVN alone works but combined pipeline is unstable.

The pass implementations exist and their unit tests pass, but they
cannot safely run in the O2+ pipeline yet. Fix: SROA byte-GEP
rejection added (ptr<i8> GEPs from array constructors).

Deferred to 29.8.5: study GCC/LLVM .refs for proper SROA placement
and GlobalLsf clobber analysis. The code is correct in isolation;
the pipeline ordering and interaction with existing passes needs work.
    mfwolffe committed
    5782d4a
  14. add bounds check elimination framework: CheckBounds RuntimeFunc + BCE pass 
    Adds RuntimeFunc::CheckBounds for array bounds checking. BCE pass
removes CheckBounds calls when the index is provably in-bounds
(constant index in [lower, upper], or loop IV within loop bounds).

Bounds check INSERTION (adding CheckBounds at array access sites
in the lowerer) is deferred — this sprint delivers the elimination
framework. Unit tests verify constant-index elimination works.
    mfwolffe committed
    1df1e75
  15. add cross-block load-store forwarding using dominance + alias analysis
    mfwolffe committed
    6661da3
  16. add GVN pass: cross-block redundancy elimination via dominator-tree hash-consing
    mfwolffe committed
    884d943
  17. add Fortran-specific alias analysis: distinct alloca/global→NoAlias, GEP offset disambiguation
    mfwolffe committed
    f1d228b
  18. add SROA pass: decompose small aggregate allocas into individual scalars 
    Detects alloca [T x N] where N≤8 with constant-index GEPs and no
address escape. Creates N individual alloca T, rewrites GEP users.
Combined with second Mem2Reg pass, enables full scalar promotion
of small arrays and complex numbers.
    mfwolffe committed
    b3370ef
  19. document inline performance concern in noted_issues: one-at-a-time is correct, re-scan optimization deferred
    mfwolffe committed
    b109c6c
  20. fix critical multi-site inline SSA bug: revert to one-at-a-time with re-scan 
    Two calls to the same function in one block broke SSA: the second
inline split a block already modified by the first, leaving stale
ValueId references. Reverted to safe one-at-a-time approach with
pass manager fixpoint re-invocation.

Also: removed unused preds variable and dead || true in SimplifyCFG.
Added inline_multisite.f90 regression test (3 calls to helper in
one block).
    mfwolffe committed
    b088c5f
  21. fix audit findings: inlining() includes O1, rename bottom_up_order, add PURE annotation to IR printer, loop cost multiplier
    mfwolffe committed
    6ad65a6
  22. add loop cost multiplier to inline cost model: 4x per loop in callee
    mfwolffe committed
    a30dfab
  23. inline all eligible call sites in one pass invocation (was one-at-a-time)
    mfwolffe committed
    4d068f6
  24. add dead function elimination: remove module functions with zero callers after inlining
    mfwolffe committed
    63a5269
  25. add SimplifyCFG pass: thread jumps through empty blocks, merge single-pred chains
    mfwolffe committed
    dff853d
  26. add function inlining pass with call graph, cost model, and Internal ref codegen fix 
    Inlining clones callee blocks into caller, maps params to args,
replaces Return with Branch to post-call block. Cost model: O1 ≤20
insts (PURE only), O2 ≤100, O3 ≤200. Non-recursive only.

Also fixed codegen to resolve FuncRef::Internal(idx) to actual
function names via select_module name table (was emitting _func_N
placeholder). All 866 unit tests + all test programs O0==O2.
    mfwolffe committed
    b93533e
  27. add call graph with recursion detection and bottom-up ordering
    mfwolffe committed
    f104b8a
  28. add is_pure/is_elemental metadata to IR Function, propagate from AST prefix
    mfwolffe committed
    264f8a9
  29. add call resolution pass: rewrite External→Internal for module-local functions
    mfwolffe committed
    d625cae
  30. rewrite fission/fusion CFG surgery, re-enable in pipeline 
    Fission: clone loop, remove store from each copy (LLVM-inspired).
Fusion: latch-redirect pattern — clone gap block constants into B's
body, redirect A's body exit → B's body → B's latch → A's latch.
Remap iv_b → iv_a. Exit block content copying.

Guards: skip loops with inner loops, skip fission-produced clones.
Dep analysis: fusion_legal now remaps IVs so same-iteration RAW deps
are correctly recognized as fusion-legal.

All 12 stress/loop tests pass O0==O2. 859 unit tests pass.
    mfwolffe committed
    f9aca25