YJIT: Use starting context for status === CantCompile

[jimmyhmiller]

No description provided.

[jimmyhmiller]

I looked into the failure. It is unrelated to my change. I can actually recreate it locally on master. I will make an issue about it.

[maximecb]

Thanks for fixing this and adding new tests.

[k0kubun]

When we do stack temp register allocation, just looking at a stack operand (even without push/pop) can result in a register spill, i.e. a destructive ctx change. So I'm planning to remove all starting_ctx patterns, including this PR's change.

I'll work on it myself tomorrow, but if you already have insight as to which place should be fixed (is it just the ctx.stack_pop(1) for block arg?), your input would be appreciated.

edit: I changed my mind. The following idea seems easier to maintain.

[k0kubun]

Another idea: If we store asm.insns.len() before compiling each YARV insn, when CantCompile is returned, maybe we could rewind asm to the point before the code for the cant-compile YARV insn is generated. Then we could safely use starting_ctx even with stack temp register allocation.

[maximecb]

Another idea: If we store asm.insns.len() before compiling each YARV insn, when CantCompile is returned, maybe we could rewind asm to the point before the code for the cant-compile YARV insn is generated. Then we could safely use starting_ctx even with stack temp register allocation.

I'm a bit confused.

Normally, we return CantCompile before any instructions are generated? Is this not the case here?

[k0kubun]

Normally, we return CantCompile before any instructions are generated?

This doesn't seem like the case for method calls. invokesuper and invokeblock generating guards before calling gen_send_iseq (which may return CantCompile) are fine-ish because it doesn't touch the stack. However, for OPTIMIZED_METHOD_TYPE_SEND (send(:foo, bar)), you inevitably need to generate a guard that looks at the stack first and then go through the normal method call sequence (which may return CantCompile) again. It doesn't seem trivial to implement the assumption without rewinding an Assembler.

[k0kubun]

I can still look at simply not using starting_ctx for CantCompile though. It's probably too late to avoid generating code before CantCompile, but avoiding stack_pop before CantCompile might be easier than that.