Macro Exterminator #15601
Macro Exterminator #15601
Conversation
|
One interesting implication of this that I didn't think about before is that --pretty expanded source will not export macros that the unexpanded source would. I don't know how much we care about the semantics of the expanded source, though, as there are already other issues with e.g. hygiene. |
|
On Jul 10, 2014, at 11:30 PM, Steven Fackler [email protected] wrote:
I thought about that a bit. As you point out, though, there are much larger problems with pretty —expanded. plus, all the tests pass!
|
|
@sfackler Personally I would love to see the FWIW, the only feature I know of offhand that should be impossible to implement in |
|
@kballard your point is well taken, but I'm reluctant to spend time on fixing pretty --expanded for two reasons:
|
| fn visit_mac(&mut self, macro: &ast::Mac, _:()) { | ||
| self.sess.span_diagnostic.span_err(macro.span, | ||
| "macro exterminator: expected AST \ | ||
| with no macro invocations"); |
There was a problem hiding this comment.
Should this be span_bug instead of span_err?
There was a problem hiding this comment.
yes! right you are.
There were two things named MacroExpander, which was confusing. I renamed one of them TTMacroExpander. [breaking change]
Per discussion with @sfackler, refactored the expander to change the way that exported macros are collected. Specifically, a crate now contains a side table of spans that exported macros go into. This has two benefits. First, the encoder doesn't need to scan through the expanded crate in order to discover exported macros. Second, the expander can drop all expanded macros from the crate, with the pleasant result that a fully expanded crate contains no macro invocations (which include macro definitions).
macros can expand into arbitrary items, exprs, etc. This means that using a default walker or folder on an AST before macro expansion is complete will miss things (the things that the macros expand into). As a partial fence against this, this commit moves the default traversal of macros into a separate procedure, and makes the default trait implementation signal an error. This means that Folders and Visitors can traverse macros if they want to, but they need to explicitly add an impl that calls the walk_mac or fold_mac procedure This should prevent problems down the road.
the Macro Exterminator ensures that there are no macro invocations in an AST. This should help make later passes confident that there aren't hidden items, methods, expressions, etc.
…, r=alexcrichton Our AST definition can include macro invocations, which can expand into all kinds of things. Macro invocations are expanded away during expansion time, and the rest of the compiler doesn't have to deal with them. However, we have no way of enforcing this. This patch adds two protective mechanisms. First, it adds a (quick) explicit check that ensures there are no macro invocations remaining in the AST after expansion. Second, it updates the visit and fold mechanisms so that by default, they will not traverse macro invocations. It's easy enough to add this, if desired (it's documented in the source, and examples appear, e.g. in the IdentFinder. Along the way, I also consulted with @sfackler to refactor the macro export mechanism so that it stores macro text spans in a side table, rather than leaving them in the AST.
fix: Temporarily skip decl check in derive expansions "Fixes rust-lang/rust-analyzer#15344"
Our AST definition can include macro invocations, which can expand into all kinds of things. Macro invocations are expanded away during expansion time, and the rest of the compiler doesn't have to deal with them. However, we have no way of enforcing this.
This patch adds two protective mechanisms.
First, it adds a (quick) explicit check that ensures there are no macro invocations remaining in the AST after expansion. Second, it updates the visit and fold mechanisms so that by default, they will not traverse macro invocations. It's easy enough to add this, if desired (it's documented in the source, and examples appear, e.g. in the IdentFinder.
Along the way, I also consulted with @sfackler to refactor the macro export mechanism so that it stores macro text spans in a side table, rather than leaving them in the AST.