notes

proposals/p3797.md

@@ -22,8 +22,8 @@ SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 -   [Rationale](#rationale)
 -   [Alternatives considered](#alternatives-considered)
     -   [Other raw identifier syntaxes](#other-raw-identifier-syntaxes)
-    -   [Restrict `r#` to current and future keywords](#restrict-r-to-current-and-future-keywords)
-    -   [Don't require `r#` for references to raw identifiers](#dont-require-r-for-references-to-raw-identifiers)
+    -   [Restrict raw identifier syntax to current and future keywords](#restrict-raw-identifier-syntax-to-current-and-future-keywords)
+    -   [Don't require syntax for references to raw identifiers](#dont-require-syntax-for-references-to-raw-identifiers)
     -   [Don't provide raw identifier syntax](#dont-provide-raw-identifier-syntax)
 
 <!-- tocstop -->
@@ -129,40 +129,72 @@ Advantages to `r#` are:
     literals. The `r#` syntax offers consistency with this, and will hopefully
     be recognizable to users.
 -   Consistency with Rust.
--   Avoids reserving an otherwise unused character for a syntax that should have
-    narrow usage.
 
-Disadvantages are that any `r`-prefixed identifier parses substantially slower,
-as noted in [PR #3044](https://github.com/carbon-language/carbon-lang/pull/3344)
-which implemented `r#` syntax.
-
-Other syntaxes we considered are:
-
--   `#` without `r`.
+A disadvantage is that any `r`-prefixed identifier parses substantially slower,
+as noted by the benchmarks in
+[PR #3044](https://github.com/carbon-language/carbon-lang/pull/3344) which
+implemented `r#` syntax. A 2% benchmark slowdown indicates around 2x because `r`
+is about 1-in-55 identifiers. This may be reduced if we enable tail calls and
+other optimizations.
+
+Various other prefixes have been discussed, mostly using a special character
+prefix in order to restrict the lexing impact. In particular:
+
+-   `\` prefix.
+    -   Similar to `\` escaping in strings.
+    -   More intuitive "escaping" semantic for some developers versus `r#`.
+    -   Creates a different meaning for `\n` as an identifier versus `\n` as a
+        character escape.
+        -   Some of this could be addressed by restricting `\` raw identifiers
+            to only keywords in the language, meaning `\n` would only be a
+            character escape. The alternative
+            [Restrict raw identifier syntax to current and future keywords](#restrict-raw-identifier-syntax-to-current-and-future-keywords)
+            applies to this solution.
+-   `#` prefix without `r`.
     -   Would be more consistent with string literals, and avoid the lexing
         overhead.
     -   We are considering using a `#` prefix for metaprogramming, so the `r`
         offers a way to keep the `#` prefix available for other purposes.
     -   `#if` may look to C++ developers like a compiler directive, rather than
-        a raw identifier.
--   Backticks, as in Swift.
+        a raw identifier for `if`.
+-   Backticks, consistent with Swift.
     -   We prefer not to use backticks for Carbon syntax so that it is easy to
         write in Markdown, which uses backticks for inline code.
--   `@` prefix, as in C#.
+-   `@` prefix, consistent with C#.
     -   We've also discussed using `@` for attributes, similar to Python.
--   Other currently unused characters, such as `~` or `%`.
-    -   Reserves a character for a feature with limited usage.
-    -   Misses an opportunity to provide cross-language consistency.
-
-### Restrict `r#` to current and future keywords
+-   Other currently unused characters, such as `~`, `$`, or `%`.
+    -   We expect raw identifiers to be relatively rare. There may be future
+        uses for these characters that allow us to serve a broader use-case.
+    -   While we could change raw string literal syntax to use the same
+        character, it would be helpful if raw string literal syntax had some
+        degree of cross-language syntactic consistency in order to reduce
+        learning curves.
+
+Raw identifier syntax is expected to be an edge case of the language. As a
+consequence, it should probably be expected that developers reading it will be
+more likely to rely on their understanding of the syntax either from other parts
+of Carbon, or from other languages. This means it's helpful if the syntax can be
+understood on its own, but if it's confusable with C++ syntax, the relative
+rarity could exacerbate understandability issues.
+
+### Restrict raw identifier syntax to current and future keywords
 
 We had discussed maintaining a list of current and future keywords, and only
-allowing `r#` in those cases. We aren't doing that because it creates an
-additional burden for software evolution, that the language must release a
-version that "declares" future keywords without turning them into actual
-keywords.
-
-### Don't require `r#` for references to raw identifiers
+allowing raw identifier syntax in those cases. If this were done as part of the
+toolchain, releases would need to push versions that "declare" future keywords
+without turning them into actual keywords. For a library that used those
+identifiers, it would initially be compatible with compiler versions up to and
+including the "future" keyword version; upon using raw identifier syntax, that
+would become the minimum compiler version. This creates a compiler versioning
+dependency that it might be helpful to avoid.
+
+As an alternative approach, Carbon could provide a command line option which
+libraries could use to specify future keywords that are used in the program.
+While some systems such as `bazel` allow libraries to indicate options they need
+for compilation, other build systems such as `cmake` might require library users
+to update their dependencies as well.
+
+### Don't require syntax for references to raw identifiers
 
 We could say that, in a scope where a raw identifier has been declared, the
 token without `r#` now refers to the identifier instead of the keyword. If the