Inefficient ref counting in else if expressions

[brson]

The way else if is currently translated the following code

auto x = if (a) {
  b
} else if (c) {
  d
} else {
  e
};

translates the same as

auto x = if (a) {
  b
} else {
  if (c) {
    d
  } else {
    e
  }
};

This creates a new sub-scope for each else if branch.

The mechanism for returning results from blocks entails copying the result into a new alloca and dropping any refs for it in the parent scope. These two things combined mean that an else if chain with n else if branches will generate allocas for n results and n take/drop pairs.

I suppose it could be possible to eliminate them in an optimization pass that eliminated redundant take/drops, but this also might be a good argument for reverting to the previous, non-recursive definition of else if.

[marijnh]

This problem also occurs for other types of nested blocks, so I think we should optimize the general case rather than change the representation of if/else again.

[marijnh]

This was mostly handled by 7d68cbd . There's still something to be gained by returning block results with move semantics, but that's not really what this bug is about.