Reconstruct do-loops

[Simn]

Currently, the analyzer does not reconstruct do-loops so the generators never really see them. This PR changes this and surfaced some problems:

• C++

This somehow generates a label after closing the loop, which throws off native compilation: while((cur_match < 10)) _hx_goto_1:;

• Lua

Fails with /home/travis/build/Simn/haxe/std/lua/_std/EReg.hx:142: characters 5-10 : This expression cannot be compiled to Lua

• HL

Just hangs.

---

I'm assigning the related target maintainers. I don't want to merge this for 4.0 anyway, but we should look into supporting this properly again.

[hughsando]

Are you thinking about reconstructing for loops? IIRC, I had a few optimizations in there, which may have got a bit rusty.

[hughsando]

I put a fix in for hxcpp:
c025565
There is also a fix in the hxcpp repo for cppia jit case.

[Simn]

I have added support for do-while loops to the JVM target. It's not perfect but good enough for now.

@jdonaldson Lua still fails: std/lua/_std/EReg.hx:156: characters 5-10 : This expression cannot be compiled to Lua

@ncannasse HL still hangs

[jdonaldson]

I'll check into this, sorry for the delay.

[Simn]

Updated the branch and added a checkbox for Justin. :)

[jdonaldson]

[jdonaldson]

/// CHECKBOX CLICKING INTENSIFIES ///

[Simn]

I want to check this for 4.3 and see where we're at. Maybe I'm even brave enough to fix HL.

[Simn]

Shockingly, the HL debugger worked and pointed me to EReg.map. The problem can be reproduced with that:

function main() {
	var a = ~/a+/g.map("aaabacx", function(r) return "[" + r.matched(0).substr(1) + "]");
	trace(a);
}

According to the debugger, offset is 0 on every iteration, which means there's never any break.

[Simn]

final global = true;

function main() {
	var offset = 0;
	do {
		trace(offset);
		if (offset >= 3)
			break;
		offset = 3;
	} while (global);
}

This infinitely prints 0. Dump looks fine, so I think it's the codegen.

[Simn]

Version without trace:

final global = true;

@:pure(false)
function dontTrace(d:Int) {}

function getOffset() {
	return 0;
}

function main() {
	var offset = getOffset();
	do {
		dontTrace(offset);
		if (offset >= 3)
			break;
		offset = 3;
	} while (global);
}

This is some problem with hlopt. Diff with/without -D hl_no_opt:

The original compares via registers 11 -> 1 -> 8, the optimized version stores in register 4 but never does anything with it. At the very least this would require a mov 0,4 before jumping back.

This might be a bit above my paygrade because it likely goes into the register liveness analysis. I'll probably just disable do-reconstruction for HL, but at some point this should definitely be looked into.