maandag 20 juli 2015

Of Values


In the interest of the common interest in my little project, I think it's time for me to blog again. Today marks the end of the fifth week of my project period, since my project was planned to take 10 weeks, that means I'm now halfway. So what have I achieved, and what have I learned?

Well, I promised to deliver the following things:
  1. An implementation of the code generation algorithm described below, including instruction selection tables for the x64 architecture
  2. A runtime representation of machine operations ('expressions') that will form the input to the code generator and is suitable for targeting to different architectures
  3. A patched version of DynASM capable of addressing the extended registers of x64
  4. Conversion of large parts of the current JIT to the new algorithm
  5. An extension API for REPR ops to insert (inline) expressions into the JIT in place of some operations
  6. A set of automated tests known to trigger JIT compilation to either the NQP or Rakudo Perl 6 test suite.
  7. Reports and documentation explaining the JIT and it's API
I have delivered 2, and 3, and the main reason I haven't done 4 yet is that 1 is not completely done. (The consequence of converting everything to the expression tree format would be that testing the soundness of compilation algorithms would become much more difficult). I have delivered tooling (i.e. a preprocessor) to elegantly and efficiently transform MoarVM bytecode segments to the expression tree format.

I think my almost-weekly blog reports do something for 7, but real documentation is still lacking. In the case of 6 (the test suite), it turns out that practically any NQP program - including bootstrapping NQP itself - already exercises the JIT quite a bit, including the new expression tree pseudocompiler. Thus, during development it has not yet been necessary to develop an explicit test suite, but I expect it will become more useful when the core compiler has stabilized. So in short, although I am not quite near the finish line, I think I am well underway to delivering a usable and useful compiler.

What have I learned that I think will help me go forward?
  1. A lot of things that look like simple expressions in MoarVM are quite complex underneath. Some things include conditional evaluation, some include evaluation lists. Many things have common subexpressions. Many other things are really statements.
  2. A MoarVM basic block is a much larger unit than a machine basic block, and the former may include many of the latter. A basic block in the expression tree is also quite conceptually difficult, given that
  3. Lazy evaluation is not compatible with single evaluation in case of conditional evaluation.
  4. The issues of evaluation order, value management, register management and instruction selection are distinct but closely related. Each greatly influences the order. For instance,
    1. A register can hold multiple values (values can be aliased to the same register).
    2. Values may be classified as intermediaries (single use not representing a final variable), temporaries (multiple uses not representing a final variable) and locals (that do represent a final variable).
    3. Value uniqueness falls directly out of the expression tree format.
    4. Instruction selection influences the amount of registers required and should precede register management.
    5. Register selection benefits from a consistent way to get a 'free register', and either a heap or a stack are decent ways to provide this; more importantly, it benefits from a way to efficiently subset the register set.
  5. It's nearly impossible to compile decent code using only a single pass traversal, because you don't know where values will end up, and to which of the 3 classes above it belongs.
  6. The expression tree is really a Directed Acyclic Graph, and traversal and compilation can be significantly more complex for a DAG than they are for a tree.
Accordingly, I've spent most of my last week learning these things, in various degrees of hard and easy ways to learn them. This is why, as far as features are concerned, I don't have so much news to report this week. I hope next week I'll have more exciting news to report. See you then!

maandag 13 juli 2015

A Progress Report

Another week, another moment for reporting JIT compiler progress. I don't really have an interesting story to tell, so I'll keep it to a list of goals achieved.

I implemented a macro facility in the expression tree template builder, and changed the parser to use s-expressions throughout, making it much simpler.  I've used it to implement some opcode templates, learning much about what is required of the expression tree.

I've introduced a macro-based dynamic array implementation and refactored the JIT graph builder and expression tree builder to use it. This is necessary to allow the expression tree builder to use JIT graph labels. (For the record, the graph is the structure representing the whole routine or frame, and the expression tree represents a small part of interconnected expressions or statements. Expression tree is a misnomer for the data type I'm manipulating, because it is a DAG rather than a tree, and it holds statements rather than expressions. But the name is there, and I haven't really got a replacement ready).

I've implemented a 'generic' tree-walking mechanism on the expression tree, and designed some algorithms for it to help compiling, such as live-range calculations and common subexpression elimination (CSE). CSE is not just a useful optimization, but as a result of it, all sorts of useful information can be calculated, informing register allocation and/or spill decisions. Another useful optimization, and not a very difficult one, is constant folding.

I've added and changed and removed a bunch of expression tree node types and macro's. There are some interesting language-design details there; for instance that all and any can stand in for boolean or and and when these are used for binary computations, as is the case for machine code.

I've started writing a 'pseudocompiler', that is to say, a routine that logs the machine code statements that would be produced by the expression tree compiler to the JIT log, allowing me to inspect the logs to find bugs rather than deep down in GDB. Predictably, there were many bugs, most of which I think I've now fixed.

I've implemented the worlds most naive register allocator, based on a ring of usable registers and spilling to stack. This was more complex than I had assumed, so doing so was another learning experience. I noticed that without use information, there is no way to insert spills

I've also encountered some difficulties. The notion of a basic block - an uninterrupted sequence of operations - differs between the JIT compiler and spesh, because many MoarVM-level instructions are implemented as function calls. Function calls imply spills (because registers are not persisted between calls); but because the call may be conditional, there is potentially a path with and without spills; implying the load will be garbage. Or in other words, spills should precede conditionals, because conditionals break up the basic block. I think the SSA information from spesh could be useful here, but I have so far not figured out how to combine this information with the expression tree.

Some things (pure operations without side effects) can potentially be recalculated rather than spilled. Address calculations, which can be done inline (for the most part) in x64 instructions, are a prime example of this. (The current pseudocompiler computes these values into real registers, because the current pseudocompiler is dumb).

That is most of it, I guess. See you next week, when it's milestone time.

vrijdag 3 juli 2015

Intermediate Progress on an Intermediate Representation

In which I write about my mad-scientists approach to building an expression tree and muse about the recursive nature of compilation.

Last week I reported some success in hacking register addressing for amd64 extended register into the DynASM assembler and preprocessor. This week I thought I could do better, and implement something very much like DynASM myself. Well, that wasn't part of the original plan at all, so I think that justifies an explanation.

Maybe you'll recall that the original aim for my proposal was to make the JIT compiler generate better code. In this case, we're lucky enough to know what better means: smaller and with less memory traffic. The old JIT has two principal limitations to prevent it from achieving this goal: it couldn't address registers and it had no way to decouple computations from memory access. The first of these problems involves the aforementioned DynASM hackery. The second of these problems is the topic for the rest of this post.

To generate good code, a compiler needs to know how values are used and how memory is used. For instance, it is useless to commit a temporary variable to memory, especially if it is used directly after, and it is especially wasteful to load a value from memory if it already exists in a register. It is my opinion that a tree structure that explicitly represents the memory access and value usage in a code segment (a basic block in compiler jargon) is the best way to discover opportunities for generating efficient code. I call this tree structure the 'expression tree'. This is especially relevant as the x86 architecture, being a CISC architecture, has many ways of encoding the same computation, so that finding the optimal way is not obvious.  In a way, the same thing that make it easy for a human to program x86 makes it more difficult for a compiler.

As a sidenote: programming x86, especially the amd64 dialect, really is easy, and I would suggest that learning it is an excellent investment of time. There are literally hundreds of guides, most of them quite reasonable (although few have been updated for amd64).

It follows that if one wants to generate code from an expression tree one must first acquire or build such a tree from some input. The input for the JIT compiler is a spesh graph, which is a graphical-and-linear representation of MoarVM bytecode. It is very suitable for analysis and manipulation, but it is not so well suited for low-level code generation (in my opinion), because all memory access is implicit, as are relations between values. (Actually, they are encoded using SSA form, but it takes explicit analysis to find the relations). To summarise, before we can compile an expression tree to machine code, we should first compile the MoarVM bytecode to an expression tree.

I think a good way to do that is to use templates for the expression tree that correspond to particular MoarVM instructions, which are then filled in with information from the specific instruction. Using a relatively simple algorithm, computed values from earlier instructions are then associated with their use in later instructions, forming a tree structure. (Actually, a DAG structure, because these computed values can be used by multiple computations). Whenever a value is first loaded from memory - or we know the register values to have been invalidated somehow - an explicit load node is inserted. Similarily an 'immediate' value node is inserted whenever an instruction has a constant value operand. This ensures that the use of a value is always linked to the most recent computation of it.

Another aside: the use of 'template filling' is made significantly easier by the use of a linear tree representation. Rather than use pointers, I use indices into the array to refer to child nodes. This has several advantages, realloc safety for one, and trivial linking of templates into the tree for another. I use 64 bit integers for holding each tree node, which is immensely wasteful for the tree nodes, but very handy for holding immediate values. Finally, generating the tree in this manner into a linear array implies that the array can be used directly for code generation - because code using an operand is always preceded by code defining it.

If you agree with me that template filling is a good method for generating the low-level IR - considering the most obvious alternative is coding the tree by hand - then maybe you'll also agree that a lookup table is the most obvious way to map MoarVM instructions to templates. And maybe you'll agree that hand-writing a linear tree representation can be a huge pain, because it requires you to exactly match nodes to indices. Moreover, because in C one cannot declare the template array inline to a struct declaration - although one can declare a string inline - these trees would either have to be stored in nearly a thousand separate variables, or in a single giant array. For the purpose of not polluting the namespace unnecessarily, the last solution is preferable.

I'm not sure I can expect my reader to follow me this deep into the rabbit hole. But my narrative isn't done yet. It was clear to me now that I had to use some form of preprocessor to generate the templates (as well as the lookup tables and some runtime usage instructions). (Of course, the language of this preprocessor had to be perl). The last question then was how to represent the template trees. Since these templates could have a tree structure themselves, using the linear array format would've been rather annoying. A lot of people today would probably choose JSON. (That would've been a fine choice, to be honest). Not me, I pick s-expressions. S-expressions are not only trivial to parse (current implementation costs 23 lines), they are also compact and represent trees without any ambiguity. Using just the tiniest bit of syntactic sugar, I've added macro facilities and let statements. This preprocessor is now complete, but I still need  to implement the template filling algorithm, define all the node types for the IR, and of course hook it into the current JIT. With so much still left to do, I'm hoping (but reasonably confident) that this detour of writing an expression template generator will eventually be worth the time. (For one thing, I expect it to make creating the extension API a bit easier).

Next week I plan to finish the IR tree generation and write a simple code generator for it. That code generator will not produce optimal code just yet, but it will demonstrate that the tree structure works, and it will serve to probe difficulties in implementing a more advanced tree-walking code generator. See you then!

dinsdag 23 juni 2015

Adventures in Instruction Encoding

In which I update you of my progress in patching DynASM to Do What I Mean.

Some of you may have noticed but I've officially started working on the MoarVM JIT compiler a little over a week ago. I've been spending that time catching up on reading, thinking about the proper way to represent the low-level IR, tiles, registers, and other bits and pieces. At the advice of Jonathan, I also confronted the problem of dynamic register addressing head-on, which was an interesting experience.

As I have mentioned in my earlier posts, we use DynASM for generating x86-64 machine code, something which it does very well. (Link goes to the unofficial rather than the official documentation, since the former is much more useful than the latter). The advantage of DynASM is that it allows you to write snippets of assembly code just as you would for a regular assembler, and at runtime these are then assembled into real machine code. As such, it hides the user from the gory details of instruction encoding. I think it's safe to say using DynASM made developing the JIT dramatically simpler.

However, DynASM as we used it has an important limitation. The x86-64 instruction set architecture specifies 16 general-purpose registers, but the dynamic addressing feature of DynASM (which allows you to specify at runtime which registers are the operands of a instruction) was limited to using only the 8 registers already present in x86. This is an effect of the way instructions are encoded in x86 - namely, using 3 bits (in octal). 3 bits are not enough to specify 16 registers, so how are the extra registers dealt with?

The answer is: using a special bit in a special prefix byte (REX byte). This byte signifies the use of 64 bit operands or the use of the extended registers. To make matters more difficult, x86 instructions can use up to three registers for 2 instructions, so it is kind of important to know which bit to set, and which not to set. Furthermore, at instruction encoding time you will need to know where that REX byte is, because any number of instruction parameters might have come between the byte that holds the register address and the REX byte. (You might notice I've been talking about bytes, a lot, by now. Instruction encoding is a byte business).

I finally implemented this by specifically marking REX bytes whenever they are followed by a dynamic register declaration, and then adding in the required bits at address encoding time. This required some coordination between the lua part and the C part of DynASM, but ultimately it worked out. Patches are here. In due course I plan to backport this branch to LuaJIT This patch is not entirely complete, though, because the REX byte is not always present when using dynamic registers, only if we use 64 bit operands. Thus, it needs to be conditionally added when using extended registers in the case of 32 bit operands. We don't really expect to use that, though, since MoarVM uses 64 bit almost exclusively, especially on a 64 bit platform.

The importance of this all is that it unblocks the development of a tiler, register selection, register allocation, and in general all the nice stuff of compiler development. Next weeks, I'll start by developing a prototype tiler, which I can then integrate into the MoarVM JIT. There are still plenty of issues to deal with before that is done, and so I'll just try to keep you up to date.

Finally, if you're going to YAPC::EU, note that I'll be presenting on the topic of JIT compilers (especially in the context of MoarVM and perl6). If this interests you, be sure to check it out.

zondag 7 juni 2015

Studying

Odds are that if you read my blog, you also read either perl6 weekly or the perl foundation blog, in which case you already know that my grant application has been accepted. Yay! I should really have blogged somewhat earlier about that, but I've been very busy the last few weeks. But for clarity, that means I start working on the MoarVM JIT compiler ('expression compiler') on the 14th of June, and I hope to have reached a first milestone 5 weeks later, and a number of 'inchstones' before that.

In the meantime, I have just merged a branch that timotimo and I worked on to JIT-compile a larger number of frames containing exceptions. That caused some problems because, as it turns out, there is more than one way to throw and catch exceptions in MoarVM. To be specific, catching an exception sometimes means invoking a handler routine and sometimes means jumping to a specific point within a frame. To make matters more confusing, sometimes that means we descend in the stack (call from our current frame) and sometimes that means we ascend, and sometimes we just jump around in our current frame. And to top it of, perl6 (as one of the few languages I know of) allows you to resume an exception, like so:


sub foo() {
    try {
        say "TRY";
        die "DIE";
        say "RESUMED";
        CATCH {
            say "CATCH";
            default {
                $_.resume;
            }
        }
    }
}

loop (my int $i = 0; $i < 500; $i++) {
    foo();
}
say "FINISHED";

There is no question to me that this is super-cool, of course, but it can be a bit tricky to implement. For the JIT, it meant storing a pointer to a label *just* after the current 'throwish' operation in the exception body. (A 'throwish' op is a VM-level operation that throws exception and thus causes flow control to jump around relatively unpredictably. This is also why (some) C++ programmers dislike exceptions, by the way). This way when an exception is resumed the JIT trampolining mechanisms will ensure that control is resumed where we left off. Unless of course we have jumped to a handler in the same frame, in which case we jump there directly. And because timotimo has implemented the auxiliary ops that come with exception handling we can now JIT quite a few more frames.

Anyway, there are two reasons this post has it's name. The first is of course that I'm still busy finishing this study year for a final week, which is stressful in itself. And the second reason is that I've started reading up on articles concerning JIT compilation and code generation. A short list of these include:
Finally, I've submitted a talk proposal to YAPC::EU 2015 to discuss all the interesting bits of JIT compilation. I hope it will be accepted because (I think) there is no shortage of interesting stuff to to talk about.

zondag 19 april 2015

Grant Application

Hi everybody, I just wanted to point you to my Grant Application at the Perl Foundation to develop a better JIT compiler. Feedback is highly appreciated.

dinsdag 24 maart 2015

Advancing the JIT compiler

It is customary to reintroduce oneself after a long blogging hiatus like the one I displayed here for some 6 months. So here I am, ready to talk JIT compilers again. Or compilers in general, if you wish it. Today I want to discuss possible ways to improve the MoarVM JIT compiler for great performance wins! Or so I hope.

It is no secret (at least not to me) that the MoarVM JIT compiler isn't as good as, say, any of {v8, luajit, HotSpot, Spidermonkey, PyPy}. There are many reasons why and it is helpful to consider the structure of the Rakudo-MoarVM 'stack'.
So to help you do that I've drawn this picture:
A picture may be worth more than a thousand words, but this is just a diagram, and I can summarize the important points as:
  1. MoarVM is the 'end of the pipe'. To your program many optimizations may be applied before your code ever reaches this layer. Many more optimizations cannot be proven safe due to perl6 semantics, however, which is why spesh is a speculative optimizer.
  2. The JIT only kicks in after spesh has applied several optimizations and in the current system only after spesh has generated optimized bytecode. (This is a flaw that we are looking to correct). This is an important point that is often missed, possibly because in many other systems the optimization and JIT compilation steps are not so clearly delineated.
  3. Within MoarVM, 6model plays a central role. 6model is the name for Rakudo's implementation of the perl6 (meta-) object system. Many operations such as array indexes and hash lookups are implemented as 6model operations. In general, these ultimately resolve to 6model 'representation' objects, which are implemented in C. Many steps in the execution of a perl6 program are really (virtually) dispatched to these 'REPR' objects.
So what benefit can an improved JIT compiler bring? The JIT has an advantage compared to all other components in that it has on one hand all information of the layers above and on the other hand it has the full flexibility of the underlying machine. Many tricks are easy and possible at the level of machine code which are awkward (and compiler-specific) to express at higher levels. Generating code at runtime is more-or-less a superpower, which is probably what LISP fans will tell you too.

The current JIT cannot really take advantage of this power because it is really very simple. All operations in a given piece of code are compiled independently and stitched together to form the entire subroutine. You can think of it as a 'cut-and-paste' compiler. Because all pieces are independent, each piece has load and store values directly to memory, causing unnecessary memory traffic. There is also no way to reorder operations or coalesce operations into a smaller amount of instructions. Indeed most forms of machine-level optimization are virtually impossible.

I propose to replace - step-by-step - our current 'cut and paste JIT' with a more advanced 'expression JIT'. The idea is very simple - in order to generate good code for a machine it is first necessary to reify the operations on that machine. Compilation is then the process of ordering these operations in a (directed acyclic) graph and selecting instructions to perform those operations. There is nothing new about that idea, 'real' compilers have been written that way for ages, and there is very little preventing us from doing the same. (Currently, DynASM doesn't support dynamic selection of registers other than those present in x86, which are rather limited. This is something that can be fixed, though).


To be sure, I think we should leave large parts of the current JIT infrastructure intact, with maybe an improvement here or there. And I think this 'expression JIT' can be introduced piecemeal, with the current 'cut-and-paste' code snippets serving as a template. The expression JIT will function as another node type for the current JIT, although probably most code will eventually be converted to it. When this is done it opens up a lot of possibilities:
  • Load and store elision, and in general 'optimal' code generation by register selection and allocation. We are aided in a way by the fact that x86-64 has become more 'RISC-y' over the years, which makes instruction selection simpler.
  • JIT compilation of REPR object methods. The 'expression tree' on which the JIT operates is architecture-independent, so it's feasible to have REPR objects generate a tree for specific operations, effectively 'inlining' such operations into the compilation frame. In real terms, this may translate a high-level array access into a simple pointer reference.
  • NativeCall calls may be more easily converted into JIT-level C calls, which is ultimately the same thing, just much more efficient.
  • Many optimizations that become much simpler even if they were possible before. For example 'small bigint' optimization which lets us execute small integer operations with big integer semantics, courtesy of cheap overflow checking at the machine level. Or possibly transforming tight operation loops into equivalent SIMD instructions, although that one is in fact rather more involved.
To conclude, I think the current MoarVM JIT compiler can be radically improved. I also think that this is nothing groundbreaking intellectually speaking and that there are few barriers to implementation. And finally, that with these improvements MoarVM will have a strong base for future development. I'd love to hear your thoughts.