This PR slightly improves the types involved in creating boxed
declarations. Previously the type of
the vdecl used for the return was always `tobj` when returning a boxed
scalar. This is not the most
precise annotation we can give.
This PR sorts the declarations fed into ElimDeadBranches in increasing
size. This can improve performance when we are dealing with a lot of
iterations.
The motivation for this change is as follows. Currently the algorithm
for doing one step of abstract interpretation is:
```
for decl in scc do
interpDecl
if summaryChanged decl then
return true
return false
```
whenever we return true we run another step. Now suppose we are in a
situation where we have an SCC with one big decl in the front and then
`n` small ones afterwards. For each time that the small ones change
their summary, we will re-run analysis of the big one in the front.
Currently the ordering is basically at "random" based on how other
compilers inject things into the SCC. This change ensures the behavior
is consistent and at least somewhat intelligent. By putting the small
declarations first, whenever we trigger a rerun of the loop we bias
analyzing the small declarations first, thus decreasing run time.
Note that this change does not have much effect on the current pipeline
because: We usually construct the SCCs in a way such that small ones
happen to be in front anyways. However, with upcomping changes on
specialization this is about to change.
This PR adds “non-branching case statements”: For each inductive
constructor `T.con` this adds a function `T.con.with` that is similar
`T.casesOn`, but has only one arm (the one for `con`), and an additional
`t.toCtorIdx = 12` assumption.
For example:
```lean
inductive Vec (α : Type) : Nat → Type where
| nil : Vec α 0
| cons {n} : α → Vec α n → Vec α (n + 1)
/--
info: @[reducible] protected def Vec.cons.elim.{u} : {α : Type} →
{motive : (a : Nat) → Vec α a → Sort u} →
{a : Nat} →
(t : Vec α a) →
t.ctorIdx = 1 → ({n : Nat} → (a : α) → (a_1 : Vec α n) → motive (n + 1) (Vec.cons a a_1)) → motive a t
-/
#guard_msgs in
#print sig Vec.cons.elim
```
This is a building block for non-quadratic implementations of `BEq` and
`DecidableEq` etc.
Builds on top of #9951.
The compiled code for a these functions could presumably, without
branching on the inductive value, directly access the fields. Achieving
this optimization (and achieving it without a quadratic compilation
cost) is not in scope for this PR.
This PR creates the deprecated `.toCtorIdx` alias only for enumeration
types, which are the types that used to have this function. No need
generating an alias for types that never had it. Should reduce the
number of symbols in the standard library.
This PR generates `.ctorIdx` functions for all inductive types, not just
enumeration types. This can be a building block for other constructions
(`BEq`, `noConfusion`) that are size-efficient even for large
inductives.
It also renames it from `.toCtorIdx` to `.ctorIdx`, which is the more
idiomatic naming.
The old name exists as an alias, with a deprecation attribute to be
added after the next
stage0 update.
These functions can arguably compiled down to a rather efficient tag
lookup, rather than a `case` statement. This is future work (but
hopefully near future).
For a fair number of basic types the compiler is not able to compile a
function using `casesOn` until further definitions have been defined.
This therefore (ab)uses the `genInjectivity` flag and
`gen_injective_theorems%` command to also control the generation of this
construct.
For (slightly) more efficient kernel reduction one could use `.rec`
rather than `.casesOn`. I did not do that yet, also because it
complicates compilation.
This PR fixes an issue when running Mathlib's `FintypeCat` as code,
where an erased type former is passed to a polymorphic function. We were
lowering the arrow type to`object`, which conflicts with the runtime
representation of an erased value as a tagged scalar.
This PR changes `IRType.boxed` to map `erased` to `tobject` rather than
`object`, since `erased` has a representation of a boxed scalar 0 when
we are forced to represent it at runtime. This case does not occur at
all in the Lean codebase.
This PR changes ToIR to call `lowerEnumToScalarType?` with
`ConstructorVal.induct` rather than the name of the constructor itself.
This was an oversight in some refactoring of code in the new compiler
before landing it. It should not affect runtime of compiled code (due to
the extra tagging/untagging being optimized by LLVM), but it does make
IR for the interpreter slightly more efficient.
This PR enables code generation to proceed in parallel to further
elaboration.
It does not aim to make further refinements such as generating code for
different declarations in parallel or removing the dependency on kernel
checking.
