I discussed this with JoeP, and he said that 100 additions is plenty to verify that we don't go exponential in this case. 1000 additions causes SILGen to exhaust the stack in release builds, but we don't have the time to refactor it to avoid the exhaustion right now. <rdar://problem/17898364>
Swift SVN r21035
This addresses or improves several existing TC perf bugs (rdar://problem/15933674, rdar://problem/17110608, rdar://problem/17240816, rdar://problem/17244974), and seems to speed up our unit test runs by ~5%. (On my MacBook Pro, total average execution time is reduced from 557.28s to 531.89s.)
Swift SVN r19939
expression applications
(rdar://problem/15933674, rdar://problem/17365394 and many, many dupes.)
When solving for the type of a binOp expression, factor the operand expression
types into account when collating overloads for the operator being applied.
This allows the type checker to now infer types for some binary operations with
hundreds of nested components, whereas previously we could only handle a handful.
(E.g., "1+2+3+4+5+6" previously sent the compiler into a tailspin.)
Specifically, if one of the operands is a literal, favor operator overloads
whose operand, result or contextual types are the default type of the literal
convertible conformance of the the argument literal type.
By doing so we can prevent exponential behavior in the solver and massively
reduce the complexity of many commonly found constraint systems. At the same
time, we'll still defer to "better" overloads if the default one cannot be
applied. (When adding an Int8 to an Int, for example.)
This obviously doesn't solve all of our performance problems (there are more
changes coming), but there are couple of nice side-effects:
- By tracking literal/convertible protocol conformance info within type
variables, I can potentially eliminate many instances of "$T0" and the
like from our diagnostics.
- Favored constraints are placed at the front of the overload resolution
disjunction, so if a system fails to produce a solution they'll be the
first to be mined for a cause. This helps preserve user intent, and leads
to better diagnostics being produced in some cases.
Swift SVN r19848
