This PR changes `toLCNF` to stop caching translations of expressions
upon seeing an expression marked `never_extract`. This is more
coarse-grained than it needs to be, but it is difficult to do any
better, as the new compiler's `Expr` cache is based on structural
identity (rather than the pointer identity of the old compiler).
The newly added `tests/compiler/never_extract.lean` is also converted
into a `run` tests, because during development I found the order of the
output to `stderr` to be a bit finicky. The reason for making it a
`compiler` test in the first place is that closed term decls work
slightly differently between native code and the interpreter, and it
would be good to test both, but we already have separate tests for
`never_extract` and closed term extraction.
Fixes#8944.
This PR fixes the handling of the `never_extract` attribute in the
compiler's CSE pass. There is an interesting debate to be had about
exactly how hard the compiler should try to avoid duplicating anything
that transitively uses `never_extract`, but this is the simplest form
and roughly matches the check in the old compiler (although due to
different handling of local function decls in the two compilers, the
consequences might be slightly different).
This gets half of the way to #8944.
This PR adds explanations for a few errors concerning noncomputability,
redundant match alternatives, and invalid inductive declarations.
These adopt a lower-case error naming style, which is also applied to
existing error explanation tests.
The `.impure` LCNF `Phase` is not currently used, but was intended for a
potential future where the current `IR` passes (which operate on a
highly impure representation) were rewritten to operate on LCNF instead.
For several reasons, I don't think this is very likely to happen, and
instead we are more likely to remove some of the unnecessary differences
between LCNF and IR while keeping them distinct.
This PR improves IR generation for constructors of inductive types that
are represented by scalars. Surprisingly, this isn't required for
correctness, because the boxing pass will fix it up. The extra `unbox`
operation it inserts shouldn't matter when compiling to native code,
because it's trivial for a C compiler to optimize, but it does matter
for the interpreter.
This PR adds a cache for constructor info in toIR. This is called for
all constructors, projections, and cases alternatives, so it makes sense
to cache.
This PR adds constant folding for Char.ofNat in LCNF simp. This
implicitly relies on the representation of `Char` as `UInt32` rather
than making a separate `.char` literal type, which seems reasonable as
`Char` is erased by the trivial structure optimization in `toMono`.
This PR fixes a bug in `floatLetIn` where if one decl (e.g. a join
point) is floated into a case arm and it uses another decl (e.g. another
join point) that does not have any other existing uses in that arm, then
the second decl does not get floated in despite this being perfectly
legal. This was causing artificial array linearity issues in
`Lean.Elab.Tactic.BVDecide.LRAT.trim.useAnalysis`.
This PR adds the `nondep` field of `Expr.letE` to the C++ data model.
Previously this field has been unused, and in followup PRs the
elaborator will use it to encode `have` expressions (non-dependent
`let`s). The kernel does not verify that `nondep` is correctly applied
during typechecking. The `letE` delaborator now prints `have`s when
`nondep` is true, though `have` still elaborates as `letFun` for now.
Breaking change: `Expr.updateLet!` is renamed to `Expr.updateLetE!`.
This PR also fixes a bug in `Expr.letFun?` and `Expr.letFunAppArgs?`
when the body is not a lambda. In any case, these functions will be
removed once the `Expr.letE (nondep := true)` encoding of `have`
expressions is complete.
This PR changes the LCNF pass pipeline so checks are no longer run by
default after every pass, only after `init`, `saveBase`, `toMono` and
`saveMono`. This is a compile time improvement, and the utility of these
checks is decreased a bit after the decision to no longer attempt to
preserve types throughout compilation. They have not been a significant
way to discover issues during development of the new compiler.
This PR adds caching for the `hasTrivialStructure?` function for LCNF
types. This is one of the hottest small functions in the new compiler,
so adding a cache makes a lot of sense.
This PR changes the implementation of computed fields in the new
compiler, which should enable more optimizations (and remove a
questionable hack in `toLCNF` that was only suitable for bringup). We
convert `casesOn` to `cases` like we do for other inductive types, all
constructors get replaced by their real implementations late in the base
phase, and then the `cases` expression is rewritten to use the real
constructors in `toMono`.
In the future, it might be better to move to a model where the `cases`
expression gets rewritten earlier or the constructors get replaced
later, so that both are done at the same time.
This PR fixes an issue where the `extendJoinPointContext` pass can lift
join points containing projections to the top level, as siblings of
`cases` constructs matching on other projections of the same base value.
This prevents the `structProjCases` pass from projecting both at once,
extending the lifetime of the parent value and breaking linearity at
runtime.
This would theoretically be possible to fix in `structProjCases`, but it
would require some better infrastructure for handling join points. It's
also likely that the IR passes dealing with reference counting would
have similar bugs that pessimize the code. For this reason, the simplest
thing is to just perform the `structProjCases` pass earlier, which
prevents `extendJoinPointContext` from lifting these join points.
This PR changes LCNF's `FVarSubst` to use `Arg` rather than `Expr`. This
enforces the requirements on substitutions, which match the requirements
on `Arg`.
This PR adds documentation to builtin attributes like `@[refl]` or
`@[implemented_by]`.
Closes#8432
---------
Co-authored-by: David Thrane Christiansen <david@davidchristiansen.dk>
Co-authored-by: David Thrane Christiansen <david@lean-fro.org>
This PR uses the fvar substitution mechanism to replace erased code.
This isn't entirely satisfactory, since LCNF's `.return` doesn't support
a general `Arg` (which has a `.erased` constructor), it only supports an
`FVarId`. This is in contrast to the IR `.ret`, which does support a
general `Arg`.
This PR makes any type application of an erased term to be erased. This
comes up a bit more than one would expect in the implementation of Lean
itself.
This PR optimizes let decls of an erased type to an erased value.
Specialization can create local functions that produce a Prop, and
there's no point in keeping them around.
This PR handles constants with erased types in `toMonoType`. It is much
harder to write a test case for this than you would think, because most
references to such types get replaced with `lcErased` earlier.
This PR adds an optimization to the LCNF simp pass where the
discriminant of a `cases` construct will only be mark used if it has a
non-default alternative.
This PR increases the precision of the new compiler's non computable
check, particularly around irrelevant uses of `noncomputable` defs in
applications.
There are no tests included because they don't pass with the old
compiler. They are on the new compiler's branch and they will be merged
when it is enabled.
This PR improves the precision of the new compiler's `noncomputable`
check for projections. There is no test included because while this was
reduced from Mathlib, the old compiler does not correctly handle the
reduced test case. It's not entirely clear to me if the check is passing
with the old compiler for correct reasons. A test will be added to the
new compiler's branch.
This PR adds support for the `compiler.extract_closed` option to the new
compiler, since this is used by the definition of `unsafeBaseIO`. We'll
revisit this once we switch to the new compiler and rethink its
relationship with IO.
This PR removes incorrect optimizations for strictOr/strictAnd from the
old compiler, along with deleting an incorrect test. In order to do
these optimizations correctly, nontermination analysis is required.
Arguably, the correct way to express these optimizations is by exposing
the implementation of strictOr/strictAnd to a nontermination-aware phase
of the compiler, and then having them follow from more general
transformations.
