Patch: don't generate dead bytecode

[Per Bothner]

Tom Tromey wrote:

>Today I ran `make class-check' on libgcj.  One of the problems is a
>bug in the front end (previously reported).  All the other problems
>turned out to be dead byte code which was generated.  It turns out we
>generate code for a statement like this:
>
>    if (false) { do something }
>
>This patch removes all that dead bytecode.  It does so by simply not
>generating code along a dead branch.
>
Wouldn't it be a lot simpler and cleaner to optimize this at the tree level,
sort of like constant folding?  For example in patch_if_else_statement.

One possible complication is that CAN_COMPLETE_NORMALLY
needs to be set on the resulting node even if only the optimized-away
branch can complete normally.  I'll try a patch that does this.