I spent several hours yesterday and today hunting bugs in the IR that somehow went unnoticed until now. Notably, the constant propagation algorithm would produce infinite loops over programs containing infinite loops (I had to reread the IBM paper on the topic to figure out what was wrong). This is now fixed, and I have improved the implementation a little bit to deal with more cases.

I have also decided to change the implementation of IIR constants, which you all seemed puzzled and displeased with. I implemented small primitive functions that will dynamically cast boxed integers into the desired types (eg: pint, i32, u16, i16). The constant propagation algorithm is able to convert those to constants at compile time when the input to the functions is constant. This deals with negative constant without hacks, without the need for constant propagation over the AST, and I also took the opportunity to lighten the syntax.

Before the change: iir.cst special form, essentially serves as a macro to generate typed constants. May be confusing as it looks like a function call or an iir instruction, but it is neither.

for (var i = iir.cst(IRType.pint, 0); i < STR_TBL_INIT_SIZE; i += iir.cst(IRType.pint, 1)) { ... }

Now: call the function "pint" to get the pint value of an integer. This call becomes a pint constant after constant propagation. Similarly, i16, u16 and i32 are defined for those other types.

for (var i = pint(0); i < STR_TBL_INIT_SIZE; i += pint(1)) { ... }

- Maxime