Audit accesses to the source_map

[cjgillot]

The exact source code text in the SourceMap is not tracked for incremental compilation. Accessing it and changing behaviour depending on this text is bound to cause issues with incremental compilation. Accesses to the source_map should therefore be audited, to remove access to such untracked information.

As the main use for these is diagnostics (accessing the exact source code, adjusting a span...). As a consequence, we don't expect ICEs or miscompilations. The worse that could happen is weird diagnostics.

Are fine:

• accesses that happen before incremental compilation kicks in: parsing, macro expansion, AST passes, lowering, name resolution...
• accesses that are hashed in impl HashStable for Span: filename, line number, column number...
• accesses in rustdoc, clippy, rustfmt, as they don't use rustc's incremental compilation.

Steps:

1. create a query source_map: () -> &'tcx SourceMap, marked eval_always and no_hash;
2. for each access to tcx.sess.source_map() (there are many), determine if it looks at the source code text;
   2.2 if it does, see below;
   2.1. if it does not: leave it as it, wrap this access in a method on TyCtxt in rustc_middle::ty::context and document why this access is fine.

When an untracked access to the source code is found, there are three options, in order:

• try to replace the logic with one based on information available in the HIR;
• add the required information to the HIR and use it, it's meant for this sort of things;
• replace tcx.sess.source_map() by tcx.source_map(()), which will force the query system to recompute everything properly.

Note: there are many of such accesses. A quick grep counted ~300.
We do not expect a single contributor to audit all of them at once.

Do not hesitate to contact @cjgillot on zulip if necessary.

[mdx97]

@rustbot claim

[cjgillot]

In addition: the compiler uses guess_head_span in many places to obtain a span which corresponds to the first line of an item. For instance the impl Trait for Foo in impl Trait for Foo { ... }.

We could add a head_span field to hir::Item, hir::TraitItem, hir::ImplItem and hir::ForeignItem which is set to sess.source_map().guess_head_span(item.span) (lowering happens before incremental compilation starts, so this is fine), or something more elaborate when available. The function tcx.hir().opt_span could then be modified to return this field. This would alleviate the need for many calls to guess_head_span in the codebase.