The code pipeline now supports lazy compilation, which can be used to only compile certain FunctionNodes and leave others be, saving startup time. When these uncompiled nodes are hit, a trampoline will force them to be recompiled. This can also be used to specialize compilation fixing parameter types and return types to a callsite specific compilation. This will give performance.
Reviewed-by: attila, sundar
This commit streamlines the bytecode function signatures, prologue, local variable use, scope creation, and invocation. It started out quite innocently when we noticed that we always emit __callee__ parameters for all functions even when they are not needed, but it turned out to be quite a deep rabbit hole. In the end, I identified exact conditions when functions need to have a callee parameter, when they need to receive parent scope, when they need to create their own scope, when they need to have variable arity signature, and when they need to have an "arguments" object, and made sure that callee parameters in signatures only show up when they are needed, that parent function's scope is only passed to a child function when it is needed, that the function only creates its own scope when it is needed. In crypto.js, the number of scopes dropped from 446 to 244, and the number of callees dropped from 315 to 145.
Reviewed-by: jlaskey, lagergren
Lower suffered from being a "God class" trying to do everything at once. As Nashorn code generation has grown, so has Lower. It does several post processing passes, tries to do several things at once even though all type information isn't in place, adjusting state afterwards and so on. It also performs control flow analysis, type attribution and constant folding, and everything else code generation related before byte code emission. I have now separated the compilation process into Lower (create low level nodes from high level ones, copy code such as finally block inlining etc), Attr (assign types and symbols to all nodes - freeze slot and scope information) and FinalizeTypes (insert explicit casts, specialize invoke dynamic types for scope accesses). I've removed the kludgy AccessSpecializer, as this now integrates naturally with typing. Everything is now much easier to read and each module performs only one thing. I have added separate loggers for the separate tiers. In the process I have also fixed: (1) problems with type coercion (see test/script/basic/typecoercion.js, basically our coercion was too late and our symbol inference was erroneous. This only manifested itself in very rare occasions where toNumber coercion has side effects, such as for example when valueOf is overridden) (2) copying literal nodes (literal copy did not use the superclass copy, which made all the Node specific fields not to be copied (3) erroneous literal tokenization (literals shouldn't always just inherit token information from whatever node that creates them) (4) splitter weighnodes - unary nodes were considered weightless (4) removed the hateful and kludgy "VarNode.shouldAppend", which really isn't needed when we have an attribution phase that determines self reference symbols (the only thing it was used for) (5) duplicate line number issues in the parser (6) convert bug in CodeGenerator for intermediate results of scope accesses (see test/script/basic/access-specializer.js) ... Several of these things just stopped being problems with the new architecture "can't happen anymore" and are not bug fixes per se. All tests run. No performance regressions exist that I've been able to measure. Some increases in performance were measured, but in the statistical margin of error (which is very wide as HotSpot currently has warmup issues with LambdaForms/invoke dynamic). Compile speed has not measurably increased.
Reviewed-by: jlaskey, attila
