[lldb] Prefer triple from module in precise invocations

[kastiglione]

When using precise compiler invocations, prefer the triple of the dylib/framework (aka image/module), and not the target binary. This is particularly vital for explicitly built modules, where the pcms for the dylib target can differ from the main target. The difference in pcms can cause expression evaluation to fail.

rdar://130284825

[kastiglione]

@swift-ci test

[kastiglione]

@swift-ci test

[fredriss]

Should playgrounds just turn off precise compiler invocations rather than threading this as a different dimension in the API?

[kastiglione]

@fredriss here's the new commit, which disables precise compiler invocation for playgrounds https://github.com/apple/llvm-project/pull/8941/commits/9ec444b043e3017abbbb0a670ad22e96dfc2ec24

The tests in TestSwiftPlaygrounds.py expect the semantics that the target's triple is used to evaluate Swift code, not the current module.

[kastiglione]

Should playgrounds just turn off precise compiler invocations rather than threading this as a different dimension in the API?

@fredriss let me look into that.

[kastiglione]

@fredriss it turns out that the existing SymbolContext parameter itself controls whether a precise compiler invocation is used. I've changed the code to pass nullptr for the symbol context, whenever the expression is run from a playground (or repl). This avoids risks that could come with unprotected temporary mutation of a global (the swift-precise-compiler-invocation setting).

Thoughts?

[kastiglione]

@swift-ci test

[fredriss]

LGTM

[kastiglione]

@swift-ci test linux

[kastiglione]

The linux test failures need investigation.

[kastiglione]

@swift-ci test linux

[kastiglione]

@swift-ci test linux

[kastiglione]

@swift-ci test linux

[kastiglione]

@swift-ci test linux

[kastiglione]

@swift-ci test linux