You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
LexLeanRuntime.Laws loops forever when Lean 4.32.1 emits C for it (breaks Prism app builds, PP5007) #454
LexLeanRuntime.Laws` loops forever when Lean 4.32.1 emits C for it (breaks Prism app builds, PP5007)
Compiling the runtime module LexLeanRuntime.Laws (language 1.2) with C output never finishes. Lean 4.32.1's compiler hits two assertions and then panics in a loop:
PANIC at Lean.Compiler.LCNF.getImpureDeclIndices Lean.Compiler.LCNF.PhaseExt:238:4: assertion violation: i != 0
PANIC at Lean.Compiler.LCNF.getImpureDeclIndices Lean.Compiler.LCNF.PhaseExt:240:2: assertion violation: map.size == targets.size
PANIC at Std.DHashMap.Internal.AssocList.get! Std.Data.DHashMap.Internal.AssocList.Basic:163:11: key is not present in hash table
(the last line repeats until the process is killed: 1,219 times in 30 seconds, 4 MB of output)
The same module compiles in about 3 seconds without -c.
Why it matters
LexLean imports LexLeanRuntime.Laws into any module with a linear_arithmetic or law proof. PrismPM's application-lean stage runs lake build, which emits C for every module, so a Prism text application whose modules reach such a proof never builds. It fails with PP5007: application-lean timed out after 600 seconds and gives no hint of the cause.
lexlean build the runtime library at language/runtime-1.2/library. I used LexLean at the commit before the 0.4.0 bump, which matches the 0.3.0 SDK.
Put the generated LexLeanRuntime/{Portable,Collections,Reasoning,Laws}.lean in a lake project with roots = ["LexLeanRuntime.Portable","LexLeanRuntime.Collections","LexLeanRuntime.Reasoning","LexLeanRuntime.Laws"].
lake build. It hangs on LexLeanRuntime.Laws.
Directly: lake env lean LexLeanRuntime/Laws.lean -o out.olean -c out.c loops, while lake env lean LexLeanRuntime/Laws.lean -o out.olean exits 0 at once.
The runtime from LexLean 0.4.0 (main) behaves the same way.
Workaround we use
We keep every theorem in separate …Laws modules and build the Prism app from a second LexLean project whose entry points reach only the definition modules. With that change the app's Lean build takes 4 seconds. (Kappa Exchange: kx-model/lexlean-app.toml.)
Possible fixes
Mark the Laws declarations that the compiler need not touch as noncomputable, or move its JSON/Name helpers (fail, field, nameOf, …) out of the proof-only module, so the compiler never sees them.
Or have the runtime import Laws only at proof time, so app builds never include it.
The underlying LCNF assertion is likely a Lean bug worth reporting to leanprover/lean4 once LexLean has a minimal reproduction.
Read-only triage confirms that linear_arithmetic/law modules can import LexLeanRuntime.Laws, and our existing verifier invokes Lean for .olean output without -c. Those checks therefore do not establish that the downstream C-emission path succeeds.
Could you attach the exact four generated Lean inputs (Portable, Collections, Reasoning and Laws), the Lake configuration, and the exact LexLean commit used? The reported Lean identity is already specific. Reusing those inputs would let us isolate the -o versus -o -c difference without regenerating the runtime or repeating the full application build.
The suggested decoder helpers are emitted by the core backend; they are not authored in Laws.lex.tex. Laws also contains executable sum definitions, so blanket noncomputable marking or removing the import needs a semantic review before adoption. No reproduction or repair is claimed yet, and this is separate from the existing #433 toolchain upgrade and #437 formatting work.
Activity
afflom commentedon Nov 10, 2026
Read-only triage confirms that
linear_arithmetic/lawmodules can importLexLeanRuntime.Laws, and our existing verifier invokes Lean for.oleanoutput without-c. Those checks therefore do not establish that the downstream C-emission path succeeds.Could you attach the exact four generated Lean inputs (Portable, Collections, Reasoning and Laws), the Lake configuration, and the exact LexLean commit used? The reported Lean identity is already specific. Reusing those inputs would let us isolate the
-oversus-o -cdifference without regenerating the runtime or repeating the full application build.The suggested decoder helpers are emitted by the core backend; they are not authored in Laws.lex.tex. Laws also contains executable sum definitions, so blanket
noncomputablemarking or removing the import needs a semantic review before adoption. No reproduction or repair is claimed yet, and this is separate from the existing #433 toolchain upgrade and #437 formatting work.