2004-11-08 Ben Maurer <bmaurer@ximian.com>

[mono-project.git] / docs / mini-porting.txt

                       Mono JIT porting guide.
                   Paolo Molaro (lupus@ximian.com)

* Introduction

        This documents describes the process of porting the mono JIT
        to a new CPU architecture. The new mono JIT has been designed
        to make porting easier though at the same time enable the port
        to take full advantage from the new architecture features and
        instructions. Knowledge of the mini architecture (described in
        the mini-doc.txt file) is a requirement for understanding this
        guide, as well as an earlier document about porting the mono
        interpreter (available on the web site).
        
        There are six main areas that a port needs to implement to
        have a fully-functional JIT for a given architecture:
        
                1) instruction selection
                2) native code emission
                3) call conventions and register allocation
                4) method trampolines
                5) exception handling
                6) minor helper methods
        
        To take advantage of some not-so-common processor features
        (for example conditional execution of instructions as may be
        found on ARM or ia64), it may be needed to develop an
        high-level optimization, but doing so is not a requirement for
        getting the JIT to work.
        
        We'll see in more details each of the steps required, note,
        though, that a new port may just as well start from a
        cut&paste of an existing port to a similar architecture (for
        example from x86 to amd64, or from powerpc to sparc).
        
        The architecture specific code is split from the rest of the
        JIT, for example the x86 specific code and data is all
        included in the following files in the distribution:
        
                mini-x86.h mini-x86.c
                inssel-x86.brg
                cpu-pentium.md
                tramp-x86.c 
                exceptions-x86.c 
        
        I suggest a similar split for other architectures as well.
        
        Note that this document is still incomplete: some sections are
        only sketched and some are missing, but the important info to
        get a port going is already described.


* Architecture-specific instructions and instruction selection.

        The JIT already provides a set of instructions that can be
        easily mapped to a great variety of different processor
        instructions.  Sometimes it may be necessary or advisable to
        add a new instruction that represent more closely an
        instruction in the architecture.  Note that a mini instruction
        can be used to represent also a short sequence of CPU
        low-level instructions, but note that each instruction
        represents the minimum amount of code the instruction
        scheduler will handle (i.e., the scheduler won't schedule the
        instructions that compose the low-level sequence as individual
        instructions, but just the whole sequence, as an indivisible
        block).

        New instructions are created by adding a line in the
        mini-ops.h file, assigning an opcode and a name. To specify
        the input and output for the instruction, there are two
        different places, depending on the context in which the
        instruction gets used.

        If the instruction is used in the tree representation, the
        input and output types are defined by the BURG rules in the
        *.brg files (the usual non-terminals are 'reg' to represent a
        normal register, 'lreg' to represent a register or two that
        hold a 64 bit value, freg for a floating point register).

        If an instruction is used as a low-level CPU instruction, the
        info is specified in a machine description file. The
        description file is processed by the genmdesc program to
        provide a data structure that can be easily used from C code
        to query the needed info about the instruction.

        As an example, let's consider the add instruction for both x86
        and ppc:
        
        x86 version:
                add: dest:i src1:i src2:i len:2 clob:1
        ppc version:
                add: dest:i src1:i src2:i len:4
        
        Note that the instruction takes two input integer registers on
        both CPU, but on x86 the first source register is clobbered
        (clob:1) and the length in bytes of the instruction differs.

        Note that integer adds and floating point adds use different
        opcodes, unlike the IL language (64 bit add is done with two
        instructions on 32 bit architectures, using a add that sets
        the carry and an add with carry).

        A specific CPU port may assign any meaning to the clob field
        for an instruction since the value will be processed in an
        arch-specific file anyway.

        See the top of the existing cpu-pentium.md file for more info
        on other fields: the info may or may not be applicable to a
        different CPU, in this latter case the info can be ignored.

        The code in mini.c together with the BURG rules in inssel.brg,
        inssel-float.brg and inssel-long32.brg provides general
        purpose mappings from the tree representation to a set of
        instructions that should be easily implemented in any
        architecture.  To allow for additional arch-specific
        functionality, an arch-specific BURG file can be used: in this
        file arch-specific instructions can be selected that provide
        better performance than the general instructions or that
        provide functionality that is needed by the JIT but that
        cannot be expressed in a general enough way.
        
        As an example, x86 has the special instruction "push" to make
        it easier to implement the default call convention (passing
        arguments on the stack): almost all the other architectures
        don't have such an instruction (and don't need it anyway), so
        we added a special rule in the inssel-x86.brg file for it.
        
        So, one of the first things needed in a port is to write a
        cpu-$(arch).md machine description file and fill it with the
        needed info. As a start, only a few instructions can be
        specified, like the ones required to do simple integer
        operations. The default rules of the instruction selector will
        emit the common instructions and so we're ready to go for the
        next step in porting the JIT.
        

*) Native code emission

        Since the first step in porting mono to a new CPU is to port
        the interpreter, there should be already a file that allows
        the emission of binary native code in a buffer for the
        architecture. This file should be placed in the

                mono/arch/$(arch)/

        directory.

        The bulk of the code emission happens in the mini-$(arch).c
        file, in a function called mono_arch_output_basic_block
        (). This function takes a basic block, walks the list of
        instructions in the block and emits the binary code for each.
        Optionally a peephole optimization pass is done on the basic
        block, but this can be left for later, when the port actually
        works.

        This function is very simple, there is just a big switch on
        the instruction opcode and in the corresponding case the
        functions or macros to emit the binary native code are
        used. Note that in this function the lengths of the
        instructions are used to determine if the buffer for the code
        needs enlarging.
        
        To complete the code emission for a method, a few other
        functions need implementing as well:
        
                mono_arch_emit_prolog ()
                mono_arch_emit_epilog ()
                mono_arch_patch_code ()
        
        mono_arch_emit_prolog () will emit the code to setup the stack
        frame for a method, optionally call the callbacks used in
        profiling and tracing, and move the arguments to their home
        location (in a caller-save register if the variable was
        allocated to one, or in a stack location if the argument was
        passed in a volatile register and wasn't allocated a
        non-volatile one). caller-save registers used by the function
        are saved in the prolog as well.
        
        mono_arch_emit_epilog () will emit the code needed to return
        from the function, optionally calling the profiling or tracing
        callbacks. At this point the basic blocks or the code that was
        moved out of the normal flow for the function can be emitted
        as well (this is usually done to provide better info for the
        static branch predictor).  In the epilog, caller-save
        registers are restored if they were used.

        Note that, to help exception handling and stack unwinding,
        when there is a transition from managed to unmanaged code,
        some special processing needs to be done (basically, saving
        all the registers and setting up the links in the Last Managed
        Frame structure).
        
        When the epilog has been emitted, the upper level code
        arranges for the buffer of memory that contains the native
        code to be copied in an area of executable memory and at this
        point, instructions that use relative addressing need to be
        patched to have the right offsets: this work is done by
        mono_arch_patch_code ().


* Call conventions and register allocation

        To account for the differences in the call conventions, a few functions need to
        be implemented.
        
        mono_arch_allocate_vars () assigns to both arguments and local
        variables the offset relative to the frame register where they
        are stored, dead variables are simply discarded. The total
        amount of stack needed is calculated.
        
        mono_arch_call_opcode () is the function that more closely
        deals with the call convention on a given system. For each
        argument to a function call, an instruction is created that
        actually puts the argument where needed, be it the stack or a
        specific register. This function can also re-arrange th order
        of evaluation when multiple arguments are involved if needed
        (like, on x86 arguments are pushed on the stack in reverse
        order). The function needs to carefully take into accounts
        platform specific issues, like how structures are returned as
        well as the differences in size and/or alignment of managed
        and corresponding unmanaged structures.
        
        The other chunk of code that needs to deal with the call
        convention and other specifics of a CPU, is the local register
        allocator, implemented in a function named
        mono_arch_local_regalloc (). The local allocator deals with a
        basic block at a time and basically just allocates registers
        for temporary values during expression evaluation, spilling
        and unspilling as necessary.

        The local allocator needs to take into account clobbering
        information, both during simple instructions and during
        function calls and it needs to deal with other
        architecture-specific weirdnesses, like instructions that take
        inputs only in specific registers or output only is some.

        Some effort will be put later in moving most of the local
        register allocator to a common file so that the code can be
        shared more for similar, risc-like CPUs.  The register
        allocator does a first pass on the instructions in a block,
        collecting liveness information and in a backward pass on the
        same list performs the actual register allocation, inserting
        the instructions needed to spill values, if necessary.
        
        When this part of code is implemented, some testing can be
        done with the generated code for the new architecture. Most
        helpful is the use of the --regression command line switch to
        run the regression tests (basic.cs, for example).

        Note that the JIT will try to initialize the runtime, but it
        may not be able yet to compile and execute complex code:
        commenting most of the code in the mini_init() function in
        mini.c is needed to let the JIT just compile the regression
        tests.  Also, using multiple -v switches on the command line
        makes the JIT dump an increasing amount of information during
        compilation.
        
        
* Method trampolines

        To get better startup performance, the JIT actually compiles a
        method only when needed. To achieve this, when a call to a
        method is compiled, we actually emit a call to a magic
        trampoline. The magic trampoline is a function written in
        assembly that invokes the compiler to compile the given method
        and jumps to the newly compiled code, ensuring the arguments
        it received are passed correctly to the actual method.

        Before jumping to the new code, though, the magic trampoline
        takes care of patching the call site so that next time the
        call will go directly to the method instead of the
        trampoline. How does this all work?

        mono_arch_create_jit_trampoline () creates a small function
        that just preserves the arguments passed to it and adds an
        additional argument (the method to compile) before calling the
        generic trampoline. This small function is called the specific
        trampoline, because it is method-specific (the method to
        compile is hard-code in the instruction stream).

        The generic trampoline saves all the arguments that could get
        clobbered and calls a C function that will do two things:
        
        *) actually call the JIT to compile the method
        *) identify the calling code so that it can be patched to call directly
        the actual method
        
        If the 'this' argument to a method is a boxed valuetype that
        is passed to a method that expects just a pointer to the data,
        an additional unboxing trampoline will need to be inserted as
        well.
        

* Exception handling

        Exception handling is likely the most difficult part of the
        port, as it needs to deal with unwinding (both managed and
        unmanaged code) and calling catch and filter blocks. It also
        needs to deal with signals, because mono takes advantage of
        the MMU in the CPU and of the operation system to handle
        dereferences of the NULL pointer. Some of the function needed
        to implement the mechanisms are:
        
        mono_arch_get_throw_exception () returns a function that takes
        an exception object and invokes an arch-specific function that
        will enter the exception processing.  To do so, all the
        relevant registers need to be saved and passed on.
        
        mono_arch_handle_exception () this function takes the
        exception thrown and a context that describes the state of the
        CPU at the time the exception was thrown. The function needs
        to implement the exception handling mechanism, so it makes a
        search for an handler for the exception and if none is found,
        it follows the unhandled exception path (that can print a
        trace and exit or just abort the current thread). The
        difficulty here is to unwind the stack correctly, by restoring
        the register state at each call site in the call chain,
        calling finally, filters and handler blocks while doing so.
        
        As part of exception handling a couple of internal calls need
        to be implemented as well.

        ves_icall_get_frame_info () returns info about a specific
        frame.

        mono_jit_walk_stack () walks the stack and calls a callback with info for
        each frame found.

        ves_icall_get_trace () return an array of StackFrame objects.
        
** Code generation for filter/finally handlers

        Filter and finally handlers are called from 2 different locations:
        
               1.) from within the method containing the exception clauses
               2.) from the stack unwinding code
        
        To make this possible we implement them like subroutines,
        ending with a "return" statement. The subroutine does not save
        the base pointer, because we need access to the local
        variables of the enclosing method. Its is possible that
        instructions inside those handlers modify the stack pointer,
        thus we save the stack pointer at the start of the handler,
        and restore it at the end. We have to use a "call" instruction
        to execute such finally handlers.
        
        The MIR code for filter and finally handlers looks like:
        
            OP_START_HANDLER
            ...
            OP_END_FINALLY | OP_ENDFILTER(reg)
        
        OP_START_HANDLER: should save the stack pointer somewhere
        OP_END_FINALLY: restores the stack pointers and returns.
        OP_ENDFILTER (reg): restores the stack pointers and returns the value in "reg".
        
** Calling finally/filter handlers 

        There is a special opcode to call those handler, its called
        OP_CALL_HANDLER. It simple emits a call instruction.
        
        Its a bit more complex to call handler from outside (in the
        stack unwinding code), because we have to restore the whole
        context of the method first. After that we simply emit a call
        instruction to invoke the handler. Its usually possible to use
        the same code to call filter and finally handlers (see
        arch_get_call_filter).
        
** Calling catch handlers

        Catch handlers are always called from the stack unwinding
        code. Unlike finally clauses or filters, catch handler never
        return. Instead we simply restore the whole context, and
        restart execution at the catch handler.
        
** Passing Exception objects to catch handlers and filters.

        We use a local variable to store exception objects. The stack
        unwinding code must store the exception object into this
        variable before calling catch handler or filter.
        
* Minor helper methods

        A few minor helper methods are referenced from the arch-independent code.
        Some of them are:
        
        *) mono_arch_cpu_optimizations ()
                This function returns a mask of optimizations that
                should be enabled for the current CPU and a mask of
                optimizations that should be excluded, instead.
        
        *) mono_arch_regname ()
                Returns the name for a numeric register.
        
        *) mono_arch_get_allocatable_int_vars ()
                Returns a list of variables that can be allocated to
                the integer registers in the current architecture.
        
        *) mono_arch_get_global_int_regs ()
                Returns a list of caller-save registers that can be
                used to allocate variables in the current method.
        
        *) mono_arch_instrument_mem_needs ()
        *) mono_arch_instrument_prolog ()
        *) mono_arch_instrument_epilog ()
                Functions needed to implement the profiling interface.
        
        
* Writing regression tests

        Regression tests for the JIT should be written for any bug
        found in the JIT in one of the *.cs files in the mini
        directory. Eventually all the operations of the JIT should be
        tested (including the ones that get selected only when some
        specific optimization is enabled).
        

* Platform specific optimizations

        An example of a platform-specific optimization is the peephole
        optimization: we look at a small window of code at a time and
        we replace one or more instructions with others that perform
        better for the given architecture or CPU.
        
* 64 bit support tips, by Zoltan Varga (vargaz@gmail.com)

        For a 64-bit port of the Mono runtime, you will typically do
        the following:

                * need to use inssel-long.brg instead of
                  inssel-long32.brg.

                * need to implement lots of new opcodes:
                       OP_I<OP> is 32 bit op
                       OP_L<OP> and CEE_<OP> are 64 bit ops


        The 64 bit version of an existing port might share the code
        with the 32 bit port (for example SPARC/SPARV9), or it might
        be separate (x86/AMD64).  

        That will depend on the similarities of the two instructions
        sets/ABIs etc.

        The runtime and most parts of the JIT are 64 bit clean
        at this point, so the only parts which require changing are
        the arch dependent files.