Verify that just-emitted module interfaces parse and typecheck

[beccadax]

This PR makes Swift verify module interfaces and private module interfaces after emitting them. To do this, it adds a new frontend mode, -typecheck-module-from-interface, and modifies the driver to optionally schedule a typechecking job for each emitted module interface after merge-modules.

This is currently done unconditionally, but I'll probably add some kind of flag later on to turn it on only when requested. This extra verification step is enabled by default in assert compilers, but can be enabled in any compiler by passing the -verify-emitted-module-interface flag to the driver.

Fixes rdar://66710563.

[beccadax]

@swift-ci please test

[beccadax]

I still need to figure out how this should interact with the prebuild script's -check-only flag. Should that mode use -typecheck-from-module-interface instead of -compile-from-module-interface?

Edit: No. -check-only still builds modules from the interfaces (and needs to do so); it just skips work that's unnecessary if we know the prebuilt cache will be thrown away quickly.

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - d602bca81e4e03262bdd117ebdf4eb2c62b1b00d

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - d602bca81e4e03262bdd117ebdf4eb2c62b1b00d

[nkcsgexi]

Great addition overall! Why not we just reuse CompileModuleFromInterface action instead of defining a new one? We could potentially add another flag to indicate stopping after typechecking for -compile-module-from-interface. Maintaining an additional frontend action doesn't seem appealing to me.

[beccadax]

-typecheck and -compile-module-from-interface are both compiler modes (ModeOpts), so they're mutually exclusive. Also note that their behavior is pretty different—if you take a look at the FrontendOptions.cpp diff, ActionType::TypecheckModuleFromInterface's behavior tends to match ActionType::Typecheck, not ActionType::CompileModuleFromInterface.

There's something to be said for the idea of removing -compile-module-from-interface and instead having people use -emit-module to do that; either there would be a flag to restart the compilation in the module interface builder, or this would just happen automatically when you passed a module interface. This would have a couple of nice benefits, like allowing other flags such as -dump-ast or -emit-sil. But that would be a much bigger change and I'm not sure it can be fully justified.

(Ultimately, I think the new driver probably ought to treat .swiftinterface inputs as implicitly being requests to either compile or typecheck the interface—not sure which. But that's a whole different story, and I think it's better handled in the new driver instead of the old one.)

[beccadax]

@swift-ci please test

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - d602bca81e4e03262bdd117ebdf4eb2c62b1b00d

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - d602bca81e4e03262bdd117ebdf4eb2c62b1b00d

[beccadax]

SwiftPM test is failing:

18:14:35 Test Case '-[CommandsTests.BuildToolTests testParseableInterfaces]' started.
18:14:35 /Users/buildnode/jenkins/workspace/swift-PR-osx/branch-master/swiftpm/Tests/CommandsTests/BuildToolTests.swift:205: error: -[CommandsTests.BuildToolTests testParseableInterfaces] : failed - error: failed parsing the Swift compiler output: unexpected JSON message: : keyNotFound(CodingKeys(stringValue: "outputs", intValue: nil), Swift.DecodingError.Context(codingPath: [], debugDescription: "No value associated with key CodingKeys(stringValue: \"outputs\", intValue: nil) (\"outputs\").", underlyingError: nil))

Seems totally plausible that something might need to be updated there.

[beccadax]

With swiftlang/swift-package-manager#2854

@swift-ci please test

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - 5488473c17bf85906223f8b79b4943844c725012

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - 5488473c17bf85906223f8b79b4943844c725012

[beccadax]

@swift-ci please test