RFC for removing integer inference fallback

active/0000-rm-integer-fallback.md

@@ -0,0 +1,73 @@
+- Start Date: 2014-06-11
+- RFC PR #: (leave this empty)
+- Rust Issue #: 6023
+
+# Summary
+
+Currently we use inference to find the current type of
+otherwise-unannotated integer literals, and when that fails the type
+defaults to `int`. This is often felt to be potentially error-prone
+behavior.
+
+This proposal removes the integer inference fallback and strengthens
+the types required for several language features that interact with
+integer inference.
+
+# Motivation
+
+With the integer fallback, small changes to code can change the
+inferred type in unexpected ways. It's not clear how big a problem
+this is, but previous experiments[1] indicate that removing
+the fallback has a relatively small impact on existing code,
+so it's reasonable to back off of this feature in favor of more
+strict typing.
+
+See also https://github.com/mozilla/rust/issues/6023.
+
+[1]: https://gist.github.com/nikomatsakis/11179747
+
+# Detailed design
+
+The primary change here is that, when integer type inference fails,
+the compiler will emit an error instead of assigning the value the
+type `int`.
+
+This change alone will cause a fair bit of existing code to be
+unable to type check because of lack of constraints. To add more
+constraints and increase likelihood of unification, we 'tighten'
+up what kinds of integers are required in some situations:
+
+* Array repeat counts must be uint (`[expr, .. count]`)
+* << and >> require uint when shifting integral types
+
+Finally, inference for `as` will be modified to track the types
+a value is being cast *to* for cases where the value being cast
+is unconstrained, like `0 as u8`.
+
+# Drawbacks
+
+This will force users to cast somewhat more often. In particular,
+ranges of unsigned ints may need to be type-hinted:
+
+```
+for _ in range(0u, 10) { }
+```
+
+# Alternatives
+
+Do none of this.
+
+# Unresolved questions
+
+* If we're putting new restrictions on shift operators, should we
+  change the traits, or just make the primitives special?
+
+There is some question about how to treat enum discriminants:
+
+```
+enum Color { Red = 0, Green = 1, Blue = 2 }
+```
+
+Currently these default to `int`, but we need to change the
+behavior. Niko suggests just making discriminants always `int`, but
+how does that interact with `repr`?