Move cblock, lblock and fwdrefs out of frontend, making it a part of IRState

[ibuclaw]

As requested in #4569

@yebblies - this has O(n) search complexity, it should really be O(1).
@WalterBright - useless refactoring that only benefits GDC+LDC. But at least the crossing between C++ <-> D boundaries have been reduced.

I noticed during writing this that Blocks fwdrefs will only ever have an array length of 0 or 1 - this seemed odd until I discovered that DMD only errors at the label location, and not the location of the original goto. This seems a bit backwards, considering (contrived example):

    if (foo > 42)
        goto Lbar;
    try {
        if (foo < -42)
            goto Lbar;
        throw new Exception("unhandled");
      Lbar:          // compiler error location is here, but which goto caused it?
        /* ... */
    }
    catch {
        /* ... */
    }

So I fixed it up so it errors at the location of the goto (mimicking what GDC does), for good or ill...

If all is good, I'll add cblock to this list too, as I've made it so that getLabel is compatible.

[ibuclaw]

Tests were passing, so push in commit for cblock too.

[ibuclaw]

I've now removed uses of cblock, lblock, fwdrefs and ctype in my 2.067 branch of GDC, and will land in master shortly.

DMD has some catching up to do.

[yebblies]

Oh ok, I see what's going on. How about using aav instead of a linear lookup? The number of labels per function can get big when auto-generated switch-cases are involved.

[yebblies]

It would also allow removing some of the types from headers, since aav uses void* and casting.

[ibuclaw]

I've now completely remove IRState from GDC. :o)

[ibuclaw]

@yebblies - the fact that it's a pointer doesn't seem to sit too well (unlike the stack allocated array). If the AA is initialized deep in a nested block, the parents won't see the initialization.

statement = 0x7fbaeae05fd0, label = 0x7fbae80eeae0, irs = 0x7ffc1f4bd6d0
statement = 0x7fbaeae05fd0, label = 0x7fbae80eeaf0, irs = 0x7ffc1f4bd1a0

I'll see if there is a destructor that allows copying back up.

[ibuclaw]

Ooh, actually it's a bit worse. There's a slight reorder that needs to be done.

[ibuclaw]

@yebblies - I think it should be good... There's now a pointer to the stack which should be shared across all IRState levels in the function.