背景(Background)

我们是一个游戏公司，基于Adobe Flash Player开发3D游戏。 在一个项目中我们需要使用Adobe的Crossbridge编译器编译physx代码（C++ ），使之能运行在Flash上。

Flash Player里内嵌了一个虚拟机，称之为AVM(ActionScript Virtual Machine)。 Crossbridge能把C++代码成编译成avm机器码ABC Code，从而能在Flash上执行。

Crossbridge的编译器是在llvm-2.9,llvm-gcc-4.2基础上修改得到的

We are a game developmenting company, who has developed several 3D games based on Adobe Flash Player. In one of our projects, we are trying to use Adobe Crossbridge to compile the physx library (C++ based) to make it runnable on Flash.

Flash player embedded a virtual machine called AVM(ActionScript Virtual Machine), and Crossbridge can turn C++ code into AVM machine code(named the ABC code) thus make it runnable on flash.

Crossbridge is based on llvm-2.9 and llvm-gcc-4.2, and have made some changes to them.

问题现象 (Problem Summary)

我们碰到的问题是，同样的输入，CrossBridge编译出来的程序会死循环，而原生的使用原生llvm-2.9、llvm-gcc-4.2在Windows上编译得到的程序不会。

死循环发生在一个名为_gjkRelativeRayCast的函数。经过一些测试，我们发现两个现象：

The problem we have encounterd is that, fed with same input the program generated by crossbridge will produce different results with the program generated with original llvm-2.9,llvm-gcc-4.2 for windows. The former will run into dead loop sometimes.

The dead loop happens in a function named _gjkRelativeRayCast. After some testing, we have two observations:

现象一(Observation 1)　###

CrossBridge编译后的程序，变量sDist在循环的迭代N中的值，在进入下一个迭代N+1后，会更改。

If the following source code is compiled by CrossBridge, variable sDist will be changed unexpectedly at the end of one loop iteration.

while(BAllEq(bNotTerminated, bTrue))
	{
		// 在迭代N+1开始时，sDist 突然变为 0.0506684184074401855468750
		// when iternal N+1 begins, sDist becomes 0.0506684184074401855468750
		printf("Start Loop: sDist.x is %.60f\n", sDist.x); 
		minDist = sDist;
		
		... // 省略的代码
		
		
		sDist = V3Dot(closest, closest);

		... // 省略的代码

		// 在迭代N结束时， sDist值为 0.050668418176030627364525571465492248535156250
		// when iteraion N ends, sDist equals 0.050668418176030627364525571465492248535156250
		printf("End Loop: sDist.x is %.60f\n", sDist.x); 
		
	}

现象二 (Observation 2)　###

两段有相同语义的代码会有不同的执行结果。　

• 在代码Ａ中，minDist在两个地方被赋值(赋值1,赋值2）。
• 在代码Ｂ中，我们确定minDist有没有发生赋值2。如果没有发生赋值2，再执行赋值1，因为赋值1最终反正会被赋值2覆盖。

注：minDist和sDist没有在省略的代码中被引用。

Two sematicly equal blocks of source code, will have different result:

• In Code A, minDist will be assined at two locations(assignment 1, assignment 2)
• In code B, we determine if assignment 2 to minDist happens. If not we then run assignment 1, else do nothing. That's becuase assignment 1 will be override by assignment 2 anyway.

Note: minDist And sDist is not referenced in the omitted code.

代码A (code A)

... // 省略的代码(Intentionally omitted)
if(FAllGrtr(vw, zero))
{
	minDist = sDist;　// 赋值1(assignment 1)
	if(FAllGrtrOrEq(vr, zero))
	{
		return false;
	}
	else
	{
		const FloatV _oldLambda = _lambda;
		_lambda = FSub(_lambda, FDiv(vw, vr));
		if(FAllGrtr(_lambda, _oldLambda))
		{
			... //　省略的代码(Intentionally omitted)
			minDist = maxDist; // 赋值2(assignment 2)
			nor = closest;
		}
	}
}
... // 省略的代码(Intentionally omitted)

代码B (code B)

... // 省略的代码(Intentionally omitted)
bool modified = false;
if(FAllGrtr(vw, zero))
{
	if(FAllGrtrOrEq(vr, zero))
	{
		return false;
	}
	else
	{
		const FloatV _oldLambda = _lambda;
		_lambda = FSub(_lambda, FDiv(vw, vr));
		if(FAllGrtr(_lambda, _oldLambda))
		{
			... // 省略的代码(Intentionally omitted)
			minDist = maxDist;　// 赋值2(assignment 2)
			modified = true;
			nor = closest;
			
		}
	}
}
if(modified == false) {
	minDist = sDist;　// 赋值1(assignment 1)
}
... // 省略的代码(Intentionally omitted)

Result

The problem is that Crossbridge does not implement FP_ROUND (Rounding from a larger floating point type down to the precision of the destination), what it does is instead simply an assignment.