Skip to content

Instantly share code, notes, and snippets.

@bdw

bdw/fail.p6 Secret

Created Nov 1, 2015
Embed
What would you like to do?
What is wrong with the JIT
class C::C { };
EVAL 'sub foo() { return C::C.new(a => 1, b => 2) }; foo()' for ^500
What is currently going wrong with the JIT?
First of all, we have the mystery of the cleared jit_entry_label.
This one applies to master. It should be debuggable easily enough.
It manifests as a failed lookup of a dynamic (contextual) variable, in the compiler.
So that is the first problem here; this is a 'deep' problem.
Attached file demonstrates it. What happens is that the frame that defines the $*IN_RETURN variable is compiled after some 105 times of running it.
Then for the first 5 lookups or so, everything is fine. The 6th lookup in the JIT compiled frame, the jit_entry_label is NULL, and so the variable cannot be found.
This is very confusing because within a call chain the JIT entry label should never be NULL; that is, unless the code has reverted to the non-JIT case.
I cannot recall finding evidence of that.... but I should check again.
The second problem is that of even-moar-jit.
Roughly speaking, too many pieces are compiling in too many different ways for me to get a useful idea of what works and what doesn't.
I should probably try to remove templates temporarily so as to check which pieces do and which do not work.
I should also, but I dread doing it, refactor the tiler to
a): linearize it's output, so that compilation is no longer a tree-walk but a list-traversal
b): make register allocation offline, so that it's results can be logged and validated and understood
c): handle both values and parameter args, i.e. in acquiring values
d): generates the minimum-cost table correctly to take into account the 'implied costs' of certain rules.
(E.g. (load reg) is not cheaper than (load (addr reg)) when they can co-occur, because they can only co-occur when the reg is (addr) anyway. )
So that is a very large amount of work, and what makes it even more difficult is that each part must be correct to start with.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment