Can't see range text in FCS tests

[auduchinok]

It's difficult to analyze the tree node sources in the tests:

A possible fix and other options suggestions have been discussed in #13403.

[vzarytovskii]

Let's see if copilot can figure it out

[majocha]

There are only a few places in the codebase where the source text and associated filename meet.
If we identify one that occurs early enough that we can hope it covers most of our use cases, for example here:

fsharp/src/Compiler/SyntaxTree/LexHelpers.fs

Line 116 in 087ffa2

let usingLexbufForParsing (lexbuf: Lexbuf, fileName) f =

we can just cache the contents keyed by filename. Make that cache exist only in debug mode and visible from Range.DebugCode.

[Martin521]

we can just cache the contents keyed by filename

Would we want to add another layer of complexity?
I'd rather want to remove all file system access from the compiler core and have string-based sources only. But that's of course a bigger task.

[majocha]

we can just cache the contents keyed by filename

Would we want to add another layer of complexity? I'd rather want to remove all file system access from the compiler core and have string-based sources only. But that's of course a bigger task.

Absolutely agreed. But that would be orthogonal to the issue. In fact, this way we could at least remove the current debug mechanism of reading from filesystem in Range.DebugCode.

[auduchinok]

Just to chime in, we use the File System shim heavily in Rider for the file sources, since we have to use the IntelliJ virtual file system instead of the normal one.

[T-Gro]

I think that fits well.

The good idea really is to eliminate direct FS access from the compiler itself, and only work with ISourceText and other abstraction (some of which might not yet exist). And the abstraction might as well be coming with a default implementation being an FileSystemProvider or whatever.

Is the complexity/benefit ration good here?

[Martin521]

Yes, I think so. At some level (like fsc and FSharpChecker) of course we need to deal with file system IO. But not deep down in range or CompilerDiagnostics.

I believe we would need a type Source which contains a name (unique in the project, typically the source file path) and the source text (or ISourceText). The FileIndexTable would then map FileIndex and Source (rather than FileIndex and FileName). We then have source access through range and no longer need, deep in the compiler, any re-reads from the file system.