This PR removes a FIXME in Sym-based mvcgen' concerning a hardcoded step
limit for grind simplification. I tested that this is no longer
necessary even at highest setting for the GetThrowSetGrind benchmark.
This PR replaces eager let-expression zeta-reduction in the sym-based
mvcgen with on-demand unfolding that mirrors the production mvcgen's
behavior.
Previously, all let-expressions in the program head were immediately
zeta-reduced. Now, let-expressions are hoisted to the top of the goal
target, and the value is only inlined if it is duplicable (literals,
fvars, consts, `OfNat.ofNat`). Complex values are introduced into the
local context via `introsSimp`, preserving SymM's maximal sharing
invariants, and unfolded on demand when the fvar later appears as the
program head.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR switches the mvcgen worklist from BFS (queue) to DFS (stack)
ordering for subgoal processing.
With the new do elaborator, `if`-without-`else` generates asymmetric
bind depth between branches (`pure () >>= cont` is optimized to just
`cont` in the else branch). This caused BFS-based VC numbering to depend
on elaborator internals, swapping vc10/vc11 in test cases. DFS ordering
follows the syntactic program structure more naturally and is robust to
such bind-depth asymmetries.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR adds `simplifying_assumptions [Nat.add_assoc]` to the
`vcgen_get_throw_set_grind` benchmark, recovering hypothesis
simplification lost in a54eafb ("refactor: decouple solve from grind").
That refactor introduced `PreTac.processHypotheses` which wraps
`simpNewHyps`, but the call sites in `work` and `main` call
`Grind.processHypotheses` directly, leaving `simpNewHyps` as dead code.
Without it, long `s + 1 + … + 1` chains are never collapsed, causing an
asymptotic slowdown visible by a factor of 2 at n=150 (largest radar
input size).
Benchmark results (VCGen time in ms):
| n | Before | After | Speedup |
|---|--------|-------|---------|
| 50 | 222 | 186 | 1.2× |
| 100 | 391 | 251 | 1.6× |
| 150 | 647 | 329 | 2.0× |
| 200 | 995 | 415 | 2.4× |
| 300 | 1894 | 589 | 3.2× |
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR replaces the manual `simpForallDomains` / `implies_congr`
machinery in `introsSimp` with `Sym.Simp.simpTelescope` as the
pre-combinator, which already handles simplifying forall telescope
domains.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR adds a `simplifying_assumptions` clause to the `mvcgen'` tactic
that allows users to specify Sym.simp rewrite theorems for simplifying
hypotheses during VC generation. The syntax is `mvcgen'
simplifying_assumptions [thm₁, thm₂, ...]`. This replaces the previous
approach of hardcoding `reassocNatAdd` in `mvcgen' with grind` mode,
making hypothesis simplification user-extensible and independent of
grind.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR refactors the sym-based VCGen (`tests/bench/mvcgen/sym`) to
separate concerns between
goal decomposition and VC discharge, following the architecture of
loom2's `mvcgen'`.
- `solve` now operates on plain `MVarId` with no knowledge of grind,
returning `List MVarId`
in `SolveResult.goals`.
- `work` handles grind E-graph internalization: after `solve` returns
multiple subgoals, it
calls `processHypotheses` on the parent goal to share context before
forking.
- `emitVC` dispatches on a new `PreTac` enum (`.none`, `.grind`,
`.tactic`) to try solving
each VC, replacing the previous inline grind logic and post-hoc tactic
loop in the elaborator.
- The redundant `WorkItem` wrapper (which duplicated `Grind.Goal`'s
`mvarId`) is removed; the
worklist operates directly on `Grind.Goal`.
- `GrindContext` is replaced by `PreTac` + `hypSimpMethods` fields in
`VCGen.Context`, cleanly
separating hypothesis simplification from the discharge strategy.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR fixes the sym mvcgen benchmark script and tunes input sizes.
**run_bench.sh**: Replace `| tee` with the `capture` helper from
`util.sh`.
Without `pipefail`, piping through `tee` masks non-zero exit codes from
`lake build`, so build failures (OOM, stack overflow) go unnoticed.
**Benchmark sizes**: Scale down inputs for benchmarks that exceeded the
2s
budget so each benchmark completes in 1-2s across its 3 linearly
increasing
inputs.
**Metric collision**: Copy `GetThrowSet.Goal` into a `GetThrowSetGrind`
namespace so the grind variant reports as `GetThrowSetGrind(n)` instead
of
colliding with `GetThrowSet(n)` in `measurements.jsonl`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR adds iota reduction to the sym-based `mvcgen'` tactic by calling
`reduceRecMatcher?` before falling back to the match split backward
rule.
When a matcher/recursor has a concrete discriminant, it is reduced
directly
instead of constructing and applying a splitting backward rule, which is
significantly faster for benchmarks like `MatchIota` (previously
`MatchSplit`)
where `loop n` unrolls into `n` nested matches with known `Nat`
discriminants.
The old `MatchSplit` test case (concrete discriminants) is renamed to
`MatchIota`
and a new `MatchSplit` test case with symbolic discriminants (matching
on state)
is added to keep exercising the split backward rule code path.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR extends the sym-based `mvcgen'` tactic with two new modes:
1. `mvcgen' with <tac>`: run VCGen, then apply `<tac>` to each remaining
VC.
2. `mvcgen' with grind`: integrate grind into the VCGen loop for
incremental context internalization. Each VC inherits the parent's
E-graph state, so hypothesis processing is shared across sibling VCs,
avoiding O(n) re-internalization per VC.
The grind mode accepts the full grind configuration syntax (`mvcgen'
with grind (config := { ... }) [params]`).
A persistent `Sym.Simp` cache with a `reassocNatAdd` simproc normalizes
hypothesis types (e.g., `s + 1 + 1 + 1` → `s + 3`) before grind
internalization, achieving O(1) amortized simplification per VC.
Benchmark results for GetThrowSet (`mvcgen' with grind`):
- n=100: 400ms total, 180ms kernel
- n=250: 855ms total, 1.8s kernel
- n=500: 1.9s total, 11.8s kernel
Kernel checking time grows superlinearly and is the dominant cost at
larger sizes. This is a separate issue from VCGen performance.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The ultimate goal of this work is to turn production of `.ir` files into
separate build step so that it does not block non-`meta` imports and can
be skipped entirely when not needed. This PR implements the main logic
of this new `leanir` compiler executable and runs it after `lean` inside
the same Lake build step but leaves its use disabled behind a
`compiler.postponeCompile` flag until further Lake adjustments move it
to a separate facet so that its use can be actually beneficial.
---------
Co-authored-by: Joscha <joscha@plugh.de>
This PR adds per-result `contextDependent` tracking to `Sym.Simp.Result`
and splits the simplifier cache into persistent (context-independent)
and transient (context-dependent, cleared on binder entry). This
replaces the coarse `wellBehavedMethods` flag.
Key changes:
- Add `contextDependent : Bool := false` to `Result.rfl` and
`Result.step`
- Split `State.cache` into `persistentCache` and `transientCache`
- Remove `wellBehavedMethods` from `Methods`
- Replace `withoutModifyingCacheIfNotWellBehaved` with
`withFreshTransientCache`
- Change `DischargeResult` to an inductive (`.failed`/`.solved`)
- Add `dischargeAssumption` (context-dependent discharger for testing)
- Add `sym.simp.debug.cache` trace class
- Propagate `contextDependent` through all combinators (congruence,
transitivity, control flow, arrows, rewriting)
- Add `mkRflResult`/`mkRflResultCD` to avoid dynamic allocation of rfl
results
- Fix `isRfl` to ignore `contextDependent` (was silently broken by the
extra field)
Propagation invariant: when combining sub-results, `cd` is the
disjunction of ALL sub-results' flags — including `.rfl` results. If
`simp` returned `.rfl (contextDependent := true)`, it means `simp` might
take a completely different code path in another local context, so all
downstream results must be marked context-dependent.
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The tests need to run with certain environment variables set that only
cmake really knows and that differ between stages. Cmake could just set
the variables directly when running the tests and benchmarks, but that
would leave no good way to manually run a single benchmark. So cmake
generates some stage-specific scripts instead that set the required
environment variables.
Previously, those scripts were sourced directly by the individual
`run_*` scripts, so the env scripts of different stages would overwrite
each other. This PR changes the setup so they can instead be generated
next to each other. This also simplifies the `run_*` scripts themselves
a bit, and makes `tests/bench/build` less of a hack.
This PR generalizes the sym MVCGen's match splitting from `ite`-only to
`ite`, `dite`, and arbitrary matchers. Previously, only `ite` was
supported; `dite` and match expressions were rejected with an error.
`mkBackwardRuleForSplit` uses `SplitInfo.splitWith` to build the
splitting proof. Hypothesis types are discovered via `rwIfOrMatcher`
inside the splitter telescope, and `TransformAltFVars.all` provides the
proper fvars for `mkForallFVars`. Subgoal type metavariables use
`mkFreshExprSyntheticOpaqueMVar` so that `rwIfOrMatcher`'s internal
`assumption` tactic cannot assign them.
Adds `DiteSplit`, `MatchSplit`, and `MatchSplitState` test cases and a
`vcgen_match_split` benchmark.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR adds support for simp/equational spec theorems in the SymM-based
`mvcgen'` tactic,
catching up with a feature that the original `mvcgen` has supported for
a long time.
Users can write `@[spec] theorem : get (m := StateT σ m) = fun s => pure
(s, s) := rfl`
instead of manually specifying equivalent Hoare triples. The equational
form is more
concise and natural for specs that simply unfold definitions.
The universe level normalization (`normalizeLevelsExpr`) applied in
`work` and the backward
rule constructors is a workaround; ideally this should be integrated
into
`preprocessMVar`/`preprocessExpr` in the SymM framework so all users
benefit.
Changes:
- Add `SpecTheoremKind` to distinguish triple vs simp specs in
`SpecTheoremNew`
- Add `mkSpecTheoremNewFromSimpDecl?` to create spec entries from
equational lemmas, filtering no-op equations
- Add `mkBackwardRuleFromSimpSpec` to build backward rules via
`Eq.mpr`/`congrArg`, with instance synthesis, projection reduction, and
`unfoldReducible` on the RHS
- Migrate simp theorems from `SimpTheorems` database during
`migrateSpecTheoremsDatabase`
- Normalize universe levels so structural matching in
`BackwardRule.apply` succeeds when `max u v` vs `max v u` arise from
different code paths
- Simplify `mkSpecContext` by removing the mock `simp` context
construction
- Use `mkBackwardRuleFromExpr` instead of `mkAuxLemma` for triple specs,
since the proof may contain free variables from the goal context
- Add `AddSubCancelSimp` benchmark case and test exercising the simp
spec code path
- Change `AddSubCancel` spec proofs from `mvcgen` to `mvcgen'`
(dogfooding)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR extracts the example programs from the sym mvcgen benchmarks
into
shared `Cases.*` modules so that both benchmarks and a new fast test
suite
can reuse them. It also renames `vcgen_deep_add_sub_cancel` to
`vcgen_add_sub_cancel_deep` for consistency.
The test suite (`test_vcgen.lean`) runs all cases at n=10, completing in
~2s vs minutes for the full benchmarks. It is wired up as a `lake test`
driver and integrated with the lean4 test/bench infrastructure via
`run_test`/`run_bench` scripts registered in `CMakeLists.txt`.
Benchmark output now uses aligned `CaseName(n):` labels. The `run_bench`
script extracts per-case vcgen and kernel timings into
`measurements.jsonl`.
Benchmarks run single-threaded (`LEAN_NUM_THREADS=1`) for
reproducibility.
`vcgen_get_throw_set` is excluded from benchmarks due to pathological
`instantiateMVars` behavior.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR implements a merge sort algorithm on arrays. It has been
measured to be about twice as fast as `List.mergeSort` for large arrays
with random elements, but for small or almost sorted ones, the list
implementation is faster. Compared to `Array.qsort`, it is stable and
has O(n log n) worst-case cost. Note: There is still a lot of potential
for optimization. The current implementation allocates O(n log n)
arrays, one per recursive call.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This PR deprecates `levelZero` in favor of `Level.zero` and `levelOne`
in favor of the new `Level.one`, and updates all usages throughout the
codebase. The `levelZero` alias was previously required for computed
field `data` to work, but this is no longer needed.
🤖 Prepared with Claude Code
This PR changes the spec lookup procedure in Sym-based mvcgen so that
1. Spec candidates are sorted first before being filtered
2. Instead of filtering the whole set of candidates using
`spec.pattern.match?`, we take the first match with the highest
priority.
The second point means we will do a lot fewer matches when the highest
priority spec matches immediately. In this case, the one match is still
partially redundant with the final application of the backward rule
application. It would be great if could somehow specialize the backward
rule after it has been created. Still, this yields some welcome
speedups. Before and after for each.
```
vcgen_add_sub_cancel:
goal_1000: 865 ms, 1 VCs by grind: 228 ms, kernel: 435 ms
goal_1000: 540 ms, 1 VCs by grind: 229 ms, kernel: 426 ms
vcgen_ping_pong:
goal_1000: 458 ms, 0 VCs, kernel: 431 ms
goal_1000: 454 ms, 0 VCs, kernel: 443 ms (unchanged, because there is only ever one candidate spec)
vcgen_deep_add_sub_cancel:
goal_1000: 986 ms, 1 VCs by grind: 234 ms, kernel: 735 ms
goal_1000: 728 ms, 1 VCs by grind: 231 ms, kernel: 708 ms
vcgen_reader_state:
goal_1000: 746 ms, 1 VCs by sorry: 1 ms, kernel: 803 ms
goal_1000: 525 ms, 1 VCs by sorry: 1 ms, kernel: 840 ms
```
This PR adds support for generating and discharging postcondition VCs in
Sym-based `mvcgen`. It also adds a new benchmark case
`vcgen_ping_pong.lean` that tests this functionality. This benchmark
required a more diligent approach to maintain maximal sharing in goal
preprocessing. Goal preprocessing was subsequently merged into the main
VC generation function.
This PR sets up the new integrated test/bench suite. It then migrates
all benchmarks and some related tests to the new suite. There's also
some documentation and some linting.
For now, a lot of the old tests are left alone so this PR doesn't become
even larger than it already is. Eventually, all tests should be migrated
to the new suite though so there isn't a confusing mix of two systems.
This PR changes all `lean-toolchain` to use relative toolchain paths
instead of `lean4` and `lean4-stage0` identifiers, which removes the
need for manually linking toolchains via Elan.
After this PR, at least Elan 4.2.0 and 0.0.224 of the Lean VS Code
extension will be needed to edit core.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This PR adds the benchmark vcgen_reader_state that is a variant of
vcgen_add_sub_cancel that takes the value to subtract from a `ReaderT`
layer. Measurements:
```
goal_100: 201 ms, 1 VCs by sorry: 0 ms, kernel: 52 ms
goal_500: 382 ms, 1 VCs by sorry: 0 ms, kernel: 327 ms
goal_1000: 674 ms, 1 VCs by sorry: 1 ms, kernel: 741 ms
```
Which suggests it scales linearly. The generated VC triggers superlinear
behavior in `grind`, though, hence it is discharged by `sorry`.
This PR adds a System F formalization as a `cbv` tactic benchmark. It is
a translation of the Rocq case study from:
*Definitional Proof Irrelevance Made Accessible* by Thiago Felicissimo,
Yann Leray, Loïc Pujet, Nicolas Tabareau, Éric Tanter, Théo Winterhalter
The authors have given permission to use their development.
The benchmark includes:
- A full System F formalization (substitution lemmas, confluence of
λ-calculus, strong normalization)
- A `pow2DoubleEq` benchmark that verifies 2^(n+1) = 2^n + 2^n via
normalization in System F, measuring both `cbv` tactic time and kernel
checking time for n = 0..6
Co-Authored-By: @david-christiansen
Co-authored-by: David Thrane Christiansen <david@davidchristiansen.dk>
This PR improves the Sym VCGen such that we can use Sym.simp to unfold
definitions in the benchmark driver. To do so, it adds support for
zeta-reduction in the VCGen and ensures that proof terms are maximally
shared before being sent to the kernel.
This PR removes the unnecessary and potentially broken handling of
`let`s by zeta-reduction in Sym-based `mvcgen`.
It turns out to be unnecessary for the benchmarks so far, so there is a
lack of motivation to publicize `betaRevS` which would be needed to fix
it.
This PR shares the driver code from the Sym-based mvcgen benchmarks. It
also moves the `simp only [loop, step]` call out of the measured
section, so that we measure purely the overhead of VC generation.
The new benchmark results are as follows. All measurements for n=1000:
```
baseline_add_sub_cancel: 719.318425 ms, kernel: 382.708178 ms
vcgen_add_sub_cancel: 306.883079 ms, kernel: 455.050825 ms
vcgen_deep_add_sub_cancel: 543.350543 ms, kernel: 896.926298 ms
vcgen_get_throw_set: 669.566541 ms, kernel: 60754.202714 ms
```
Note that `vcgen_add_sub_cancel` sped up by 100% because we no longer
measure unfolding `loop` and `step`. The baseline didn't speed up as
much because it unfolded in the same `Sym.simp` call that also does
other rewrites, so there was no `simp` pass that could be eliminated.
This PR adds two more benchmarks for the Sym-based mvcgen prototype in
the style of `add_sub_cancel`.
The first is `deep_add_sub_cancel`, which is like `add_sub_cancel` but
with a much deeper monad stack:
```lean
abbrev M := ExceptT String <| ReaderT String <| ExceptT Nat <| StateT Nat <| ExceptT Unit <| StateM Unit
```
By specializing the specs for `get` and `set`, we get competitive
performance:
```
goal_100: 180.365086 ms, kernel: 79.634989 ms
goal_200: 313.465611 ms, kernel: 187.808631 ms
goal_300: 478.278585 ms, kernel: 270.210634 ms
goal_400: 638.884320 ms, kernel: 380.381127 ms
goal_500: 759.802772 ms, kernel: 472.662882 ms
goal_600: 933.575180 ms, kernel: 649.040746 ms
goal_700: 1174.367200 ms, kernel: 759.470010 ms
goal_800: 1298.866482 ms, kernel: 864.420171 ms
goal_900: 1475.315552 ms, kernel: 1008.662783 ms
goal_1000: 1627.957444 ms, kernel: 1078.627830 ms
```
Recall that `add_sub_cancel` had `goal_1000: 824.476962 ms, kernel:
477.069045 ms`, but that doesn't need to repeatedly unwrap 3 layers of
the monad.
The second benchmark is `get_throw_set`. Its kernel is
```lean
def step (lim : Nat) : ExceptT String (StateM Nat) Unit := do
let s ← get
if s > lim then
throw "s is too large"
set (s + 1)
def loop (n : Nat) : ExceptT String (StateM Nat) Unit := do
match n with
| 0 => pure ()
| n+1 => loop n; step n
def Goal (n : Nat) : Prop := ⦃fun s => ⌜s = 0⌝⦄ loop n ⦃⇓_ s => ⌜s = n⌝⦄
```
It will generate `n+1` VCs. We get `n` VCs of the form
```
s✝ : Nat
_ : ¬0 < s✝
...
_ : n < s✝ + 1 ...<n times>... + 1
⊢ ⌜s✝ = 0⌝ ⊢ₛ ⌜False⌝ (s✝ + ...<n times>...)
```
and one VC of the form
```
⌜s✝ = 0⌝ ⊢ₛ ⌜s✝ + 1 + <n times> ... + 1 = n⌝
```
which can be discharged by `grind`, but presently are discharged with
`sorry`.
Statistics:
```
goal_100: 209.435869 ms, kernel: 128.768919 ms
goal_200: 386.639441 ms, kernel: 482.244717 ms
goal_300: 559.795137 ms, kernel: 1251.777405 ms
goal_400: 753.243978 ms, kernel: 3020.878177 ms
goal_500: 1014.939522 ms, kernel: 5182.120327 ms
goal_600: 1229.173622 ms, kernel: 9296.551442 ms
goal_700: 1410.024180 ms, kernel: 16655.954682 ms
goal_800: 1684.059305 ms, kernel: 32065.951705 ms
goal_900: 1905.602401 ms, kernel: 55299.942894 ms
goal_1000: 2172.823244 ms, kernel: 84082.492485 ms
```
Need to look at kernel times here, but tactic time looks about alright.
Using `grind` to discharge just `n=100` goals took 8s.
This PR refactors the main loop of the `cbv` tactic. Rather than using
multiple simprocs, a central pre simproc is introduced. Moreover, let
expressions are no longer immediately zeta-reduced due to performance on
one of the benchmarks (`leroy.lean`).
Stacked on top of #12416
This PR adds two benchmarks (sieve of Eratosthenes, removing duplicates
from the list) and one test (a function with sublinear complexity
defined via well-founded recursion evaluated on large naturals with up
to `60` digits).
The tests have been suggested by @b-mehta.
