Add description for LUB Coercion

Author: ldm0

src/type-coercions.md

@@ -184,3 +184,47 @@ unsized coercion to `Foo<U>`.
 
 [`Unsize`]: ../std/marker/trait.Unsize.html
 [`CoerceUnsized`]: ../std/ops/trait.CoerceUnsized.html
+
+## LUB Coercion
+
+This is not in the RFC 401, but it's the behavior of current compiler. In LUB
+Coercion, we tried to find the least upper bound of given types. It can be
+observed in array collection, if-else/match arms.
+
+For example:
+
+```rust
+let emm = if true {
+    A
+} else if false {
+    B
+} else {
+    C
+};
+
+let foo = match 42 {
+    0 => A,
+    1 => B,
+    _ => C,
+};
+
+let bar = [A, B, C];
+```
+In this example, both `emm` and `foo` has type
+`LubCoerce(typeof(A), typeof(B), typeof(C))` and `bar` has type
+`[LubCoerce(typeof(A), typeof(B), typeof(C)); 3]`.
+
+LUB Coercion has some features:
+1. Order independent: e.g. `LubCoerce(ty0, ty1, ty2, ty3)` equals to
+    `LubCoerce(ty1, ty0, ty4, ty3)`.
+2. `LubCoerce(ty0, ty1, ty2, ...)` equals to
+    `LubCoerce(LubCoerce(ty0, ty1), ty2, ...)`
+4. `LubCoerce(ty0, ty1) == ty2` means both `ty0` and `ty1` can be coerced to
+    `ty2`.
+
+Notice the feature No.3, it uses the word "means" rather than "if and only if".
+That's because currently if `ty0` and `ty1` can be coerced to `ty2` and `ty2`
+equals to neither `ty0` nor `ty1`, there are only two special cases we can get
+`LubCoerce(ty0, ty1) == ty2`:
+1. `LubCoerce(FnDef, FnDef) == FnPtr`,
+2. `LubCoerce(Closure, Closure) == FnPtr`(where Closure is non-capturing).