Recent changes to the stdlib resulted in some expressions involving
literals and operators that have both generic and non-generic overloads
to become ambiguous. The ambiguity is due to an old performance hack in
the solver which unfortunately needs to remain in place at the moment to
avoid regressing expression type checker performance.
This commit changes the performance hack a bit in that instead of
stopping visiting the options in a disjunction as soon as we have a
solution involving non-generic operators and come across a constraint
involving generic operators, we instead continue visiting the elements
in the disjunction, but just skip over the generic operators.
Fixes rdar://problem/31695865 and rdar://problem/31698831.
We previously used a simple heuristic of visiting the disjunction with
the fewest number of elements in it.
Instead, this commit introduces a new way to select the best disjunction
to explore based on attempting disjunctions based on either how much
information we have about the associated type variables (e.g. how many
of the arguments of a function call in the case of bind overload
disjunctions) or by other criteria that help constraint the search or
split the connected components (e.g. coercions and calls to
initializers).
A key part of the improvement here is allowing the type checker to
attempt bindings of types to type variables when there are still
argument conversion constraints between type variables in the
system. The insight here is if there are no other constraints blocking
an attempt to bind types and we have types to try, we will be able to
test those types against active conformance constraints and potentially
fail much earlier while solving.
Visiting disjunctions in a different order exposed some other problems
with the type checker including a couple cases where map is used that
are now considered ambiguous due to problems in how we are ranking
solutions.
I measured an 11-13% reduction in type checking time for the standard
library using a release build of the compiler.
When determining whether our inference of an optional type should add
a layer of optionality, look through lvalue types.
Fixes rdar://problem/31779785.
We have various hacks in the constraint solver to improve the
compile-time performance characteristics of operators. Some of those
didn't respect availability annotations, causing us to incorrectly
choose an unavailable operator when available options exist.
Fixes rdar://problem/31592529.
There were various problems with layout constraints either
being ignored or handled incorrectly. Now that I've exercised
this support with an upcoming patch, there are some fixes
here.
Also, introduce a new ExistentialLayout::getLayoutConstriant()
which returns a value for existentials which are class-constrained
but don't have a superclass or any class-constrained protocols;
an example would be AnyObject, or AnyObject & P for some
non-class protocol P.
NFC for now, since these layout-constrained existentials cannot
be constructed yet.
This is an initial implementation of the constraint propagation pass,
disabled by default at the moment as it does not currently pass all
tests.
I have other changes that I'll hold off on for now because they are
incomplete. Without those changes, this pass doesn't really achieve
much, so this is really just a checkpoint on some progress and not
something really worth trying out at this point.
As solveRec() already does, exit immediately if we have a failed
constraint from constraint generation. Not doing so means that we can
create a SolverScope in constraint propagation, and on destruction of
that SolverScope we clear out the failedConstraint field, obscuring the
fact that we had a failure.
The inactive list may contain other disjunctions associated with bound
type variables. For now, make sure we recover the orphan directly to
fix the crash in SR-4056 / rdar://problem/30686926. Later, we can
treat these as orphans, too.
The constraint graph models type variables (as the nodes) and
constraints (as the multi-edges connecting nodes). The connected
components within this (multi-)graph are independent subproblems that
are solved separately; the results from each subproblem are then
combined. The approach helps curtail exponential behavior, because
(e.g.) the disjunctions/type variables in one component won't ever be
explored while solving for another component
This approach assumes that all of the constraints that cannot be
immediately solved are associated with one or more type
variables. This is almost entirely true---constraints that don't
involve type variables are immediately simplified.
Except for disjunctions. A disjunction involving no type variables
would not appear *at all* in the constraint graph. Worse, it's
independence from other constraints could not be established, so the
constraint solver would go exponential for every one of these
constraints. This has always been an issue, but it got worse with the
separation of type checking of "as" into the "coercion" case and the
"bridging" case, which introduced more of these disjunctions. This led
to counterintuitive behavior where adding "as Foo" would cause the
type checking to take *more* time than leaving it off, if both sides
of the "as" were known to be concrete. rdar://problem/30545483
captures a case (now in the new test case) where we saw such
exponential blow-ups.
Teach the constraint graph to keep track of "orphaned" constraints
that don't reference any type variables, and treat each "orphaned"
constraint as a separate connected component. That way, they're solved
independently.
Fixes rdar://problem/30545483 and will likely curtain other
exponential behavior we're seeing in the solver.
When trying to solve for the test case we attempt to simplify a value
member constraint and it fails because we've bound the LHS type
variable to an optional as a result of other constraints involving
other type variables in the equivalence class of this type
variable.
We don't have enough information to directly deduce the non-optional
type directly from other constraints involving this type variable.
This change results in one interesting type checking anomoly. In Swift
3 mode, we now successfully typecheck an expression that we previously
did not. Although the type checking technically makes sense given the
type checking rules we have in place, it can be a bit surprising to
users. Fortunately we emit a warning that calls out the surprising
behavior of considering an expression unused.
Fixes rdar://problem/30271695.
Shrinking such expressions will immediately restrict the solution set
to the default literal types (Int or Double), causing later solutions
to fail. Fixes rdar://problem/30220565.
While shrinking constraint system, avoid candidates which contain
closures inside, because closure expressions require special handling
and allow no rollback.
Since `ConstraintSystem::shrink` is going to attempt to type-check
sub-expressions separately it's essential to clean-up AST if constraint
generation or solving of the such expressions fails, otherwise if
such solving resulted in creation of implicit expression type variables
might leak to the outside.
Since `ConstraintSystem::shrink` is going to attempt to type-check
sub-expressions separately it's essential to clean-up AST if constraint
generation or solving of the such expressions fails, otherwise if
such solving resulted in creation of implicit expression type variables
might leak to the outside.
This commit introduces new kind of requirements: layout requirements.
This kind of requirements allows to expose that a type should satisfy certain layout properties, e.g. it should be a trivial type, have a given size and alignment, etc.
Rather than waiting until we've used a huge amount of memory, attempt to
make the choice to bail out based on the number of type bindings /
disjunction choices we visit.
I expect this will generally fail faster than the Swift 3 metric, but
will still only fail when we've got clearly exponential type checking
behvior.
Since we have multiple sources of exponential behavior today, I don't
want to make the bounds too tight. Once we fix some/most of that
behavior we can look at further tightening up the metric.
We were only reporting it in cases where we found no solutions before
deciding the expression was too complex. Had we found some solution, we
would go ahead and use that solution before bailing out due to
complexity, which means e.g. we might have been allowing otherwise
ambiguous expression to type check.
Checked casts are dependent on run-time queries; we should not attempt
to infer type variable bindings from them, because doing so produces
unreasonable bindings. Fixes rdar://problem/29894174.
