[CodeCompletion] Parse #if block containing CC token as active

[rintaro]

Treat the region containing the completion location as "active" during parsing.

1. When #if is found, see if the completion location is between #if and #endif (using backtracking.) If not, parse the entire #if ... #endif normally
2. For each region, see if the completion location is in the region (using backtracking.) If yes, parse it as "active" block, otherwise treat it "inactive"

This is not so efficient because the Lexer runs up to 3 times for each IfConfig region. But Lexer is pretty fast, and doesn't consume much memory. So it shouldn't be a problem.

This also guarantees that the completion never happens inside "inactive" blocks. Since Sema usually don't expect type checking inside inactive blocks, this change might improve the overall stability of code-completion inside #if ... #endif.

The second commit is removing includeInactiveIfConfigClauses from ASTScope. This flag was prepared for code-completion, but had never been enabled. Now that, we guarantee that name lookup never happen inside inactive blocks, we don't need this flag anymore. cc: @davidungar

rdar://problem/67027408

[rintaro]

@swift-ci Please smoke test

[rintaro]

tok::code_complete was here to prevent excessive skipping beyond the completion point. But we sometimes want to skip code_complete token, for example canParse*** kind of functions. This doesn't cause any visible regression because skipUntilDeclRBrace() family skipping functions still stop at tok::code_complete.

As for tok::pound_else and tok::pound_elseif, they should be treated as the same as tok::pound_endif in terms of "skipping" tokens.

[rintaro]

@swift-ci Please smoke test

[benlangmuir]

This approach only applies to the #if immediately surrounding the completion, so it can lead to non-sensical completions

#if FOO
let x: Foo
#else
let x: NotFoo
#endif 
...
#if FOO
x.<here>
#endif

If we moved #if condition evaluation after parsing we could look at the #if condition for the completion location and choose a config that is consistent with that for all other #ifs. If there are || branches we might choose something invalid, but seems overall reasonable. What do you think?

---

What impact does this change have on AST reuse if the #if is outside of a function body? Will it still work? Will it cause the interface hash to change if you complete in the inactive block?

---

This is not so efficient because the Lexer runs up to 3 times for each IfConfig region

What if we do a single extra pass and track the #else locations along the way so we can know directly which branch contains the completion?

[benlangmuir]

We could also check that this is not after the completion loc already.

[rintaro]

If we moved #if condition evaluation after parsing we could look at the #if condition for the completion location and choose a config that is consistent with that for all other #ifs.

It is a long standing desire to separate #if conditions evaluation from parsing. We want to do it eventually. But, for code-completion inside inactive blocks, I'm not sure we'd gain enough benefit for the effort.
For example:

#if os(macOS)
import AppKit
#else
import UIKit
#endif 

#if os(macOS)
let color = NSColor.black
#else
let color = UIColor.#HERE#
#endif

This fails if the user is working on macOS target because macOS SDK doesn't have UIKit anyway.

[rintaro]

What impact does this change have on AST reuse if the #if is outside of a function body? Will it still work? Will it cause the interface hash to change if you complete in the inactive block?

My original patch had a bug where interface hash continued to be recorded while backtracking. After fixing that, it works as expected. Switching between active and inactive blocks changes the interface hash.

[rintaro]

What if we do a single extra pass and track the #else locations along the way so we can know directly which branch contains the completion?

Great idea! Implemented.

[benlangmuir]

Typo: locaiton -> location

[rintaro]

🤦

[rintaro]

@swift-ci Please smoke test

[rintaro]

@swift-ci Please smoke test macOS