NAME ^

docs/jit.pod - Parrot JIT Subsystem

ABSTRACT ^

This PDD describes the Parrot Just In Time compilation subsystem.

DESCRIPTION ^

The Just In Time, or JIT, subsystem converts a bytecode file to native machine code instructions and executes the generated instruction sequence directly.

IMPLEMENTATION ^

Currently works on ALPHA, Arm, Intel x86, PPC, and SPARC version 8 processor systems, on most operating systems. Currently only 32-bit INTVALs are supported.

The initial step in generating native code is to invoke Parrot_jit_begin, which generally provides architecture specific preamble code. For each parrot opcode in the bytecode, either a generic or opcode specific sequence of native code is generated. The .jit files provide functions that generate native code for specific opcode functions, for a given instruction set architecture. If a function is not provided for a specific opcode, a generic sequence of native code is output which calls the interpreter C function that implements the opcode. Such opcode are handled by Parrot_jit_normal_op.

If the opcode can cause a control flow change, as in the case of a branch or call opcode, an extended or modified version of this generic code is used that tracks changes in the bytecode program counter with changes in the hardware program counter. This type of opcode is handled by Parrot_jit_cpcf_op.

While generating native code, certain offsets and absolute addresses may not be available. This occurs with forward opcode branches, as the native code corresponding to the branch target has not yet been generated. On some platforms, function calls are performed using program-counter relative addresses. Since the location of the buffer holding the native code may move as code is generated (due to growing of the buffer), these relative addresses may only be calculated once the buffer is guaranteed to no longer move. To handle these instances, the JIT subsystem uses "fixups", which record locations in native code where adjustments to the native code are required.

FILES ^

jit/${jitcpuarch}/jit_emit.h

This file defines Parrot_jit_begin, Parrot_jit_dofixup, Parrot_jit_normal_op, Parrot_jit_cpcf_op, Parrot_jit_restart_op and optionally Parrot_jit_vtable*_op. In addition, this file defines the macros and static functions used in .jit files to produce binary representations of native instructions.

For moving registers from processor to parrot and vice versa, the Parrot_jit_emit_mov* functions have to be implemented.

jit/${jitcpuarch}/core.jit

The functions to generate native code for core parrot opcodes are specified here. To simplify the maintenance of these functions, they are specified in a format that is pre-processed by jit2c.pl to produce a valid C source file, jit_cpu.c. See "Format of .jit Files" below.

include/parrot/jit.h

This file contains definitions of generic structures used by the JIT subsystem.

The op_jit array of jit_fn_info_t structures, provides for each opcode, a pointer to the function that generates native code for the opcode, whether the generic Parrot_jit_normal_op or Parrot_jit_cpcf_op functions or an opcode specific function. Parrot_jit_restart_op is like Parrot_jit_cpcf_op with the addition to check for a zero program counter. The Parrot_jit_vtable*_op functions are defined as Parrot_jit_normal_op or Parrot_jit_cpcf_op and may be implemented to do native vtable calls (s. jit/i386/jit_emit.h for an example).

The Parrot_jit_fixup structure records the offset in native code where a fixup must be applied, the type of fixup required and the specific information needed to perform the parameters of the fixup. Currently, a fixup parameter is either an opcode_t value or a function pointer.

The Parrot_jit_info structure holds data used while producing and executing native code. An important piece of data in this structure is the op_map array, which maps from opcode addresses to native code addresses.

jit.c

parrot_build_asm() is the main routine of the code generator, which loops over the parrot bytecode, calling the code generating routines for each opcode while filling in the op_map array. This array is used by the JIT subsystem to perform certain types of fixups on native code, as well as by the native code itself to convert bytecode program counters values (opcode_t *'s) to hardware program counter values.

The bytecode is considered an array of opcode_t sized elements, with parallel entries in op_map. op_map is initially populated with the offsets into the native code corresponding to the opcodes in the bytecode. Once code generation is complete and fixups have been applied, the native code offsets are converted to absolute addresses. This trades the low up-front cost of converting all offsets once, for the unknown cost of repeatedly converting these offsets while executing native code.

See src/jit/skeleton/jit_emit.h for details.

jit2c.pl

Preprocesses the .jit files to produce jit_cpu.c.

Defines in jit_emit.h ^

The architecture specific jit_emit.h file communicates some defines and tables with jit.c and languages/imcc/imc.c. The structure of the file and the defines must therefore follow a specific syntax.

Overall structure ^

    #if JIT_EMIT

    ... emit code

    #else

    ... defines
    static const jit_arch_info arch_info = {
       ... initialization of maps
       ... and possibly private static functions
    }

    #endif

See jit/skeleton/jit_emit.h for a more detailed explanation.

Defines ^

XXX most are moved into jit_arch_info now.

INT_REGISTERS_TO_MAP

This is the amount of integer registers to be mapped to processor registers. The corresponding intval_map[] has to have exactly this amount of register numbers. A register with the value of zero can not be in the list.

FLOAT_REGISTERS_TO_MAP

When this is defined, it works like above for floating point registers.

PRESERVED_INT_REGS

When this is defined, it's the amount of integer registers, that are preserved over function calls. These preserved registers have to be first in intval_map. When this is not defined, it is assumed that all registers are preserved over function calls.

PRESERVED_FLOAT_REGS

Same for floating point registers.

jit_emit_noop(pc)

JUMP_ALIGN

If these are defined, JUMP_ALIGN should be a small number stating the desired alignment of jump targets is 1 << JUMP_ALIGN. The jit_emit_noop gets called with the unaligned pc repeatedly, until the pc has the desired alignment. So the function can either emit a one byte noop instruction, or a noop like instruction (sequence) with the desired size, to achieve the necessary padding. The emitted code must not have any side effects.

ALLOCATE_REGISTERS_PER_SECTION

Normally jit.c does register allocation per section, but there is a somewhat experimental feature, to allocate registers per basic block.

MAP

Jit code generated by the imcc JIT optimizer used negative numbers for mapped registers and positive numbers for non mapped parrot registers. To use this feature, the definition of mapped registers can be redefined like so:

    #define MAP(i) OMAP(i)
    #undef MAP
    #define MAP(i) (i) >= 0 ? 0 : OMAP(i)
Parrot_jit_emit_get_base_reg_no(pc)

This macro should return the register number of the register base pointer.

S. jit/i386/jit_emit.h for actual usage of these defines.

Format of .jit Files ^

Jit files are interpreted as follows:

op-name { \n body \n }

Where op-name is the name of the Parrot opcode, and body consists of C syntax code which may contain any of the identifiers listed in the following section.

The closing curly brace has to be in the first column.

Comment lines

Comments are marked with a ; in the first column. These and empty lines are ignored.

Identifiers

In general, prefixing an identifier with & yields an address. The * prefix specifies a value. Since Parrot register values vary during code execution, their values can not be obtained through identifier substitution alone, therefore offsets are used for accessing registers.

To obtain register offsets, a set of macros exists, that have OFFS in their names:

REG_OFFS_INT(reg_no) ...

ROFFS_INT(n) ...

INT_CONST[n]

Gets replaced by the INTVAL constant specified in the nth argument.

NUM_CONST[n]

Gets replaced by the FLOATVAL constant specified in the nth argument.

MAP[n]

The nth integer or floating processor register, mapped in this section.

Note: The register with the physical number zero can not be mapped.

NATIVECODE

Gets replaced by the current native program counter.

*CUR_OPCODE[n]

Gets replaced by the address of the current opcode in the Parrot bytecode.

ISRn FSRn

The nth integer or floating point scratch register.

TEMPLATE template-name { \n body \n }

Defines a template for similar functions, e.g. all the binary ops taking three variable parameters.

template-name perl-subst ...

Take a template and do all substitutions to generate the implementation for this jit function.

Example:

    TEMPLATE Parrot_set_x_ic {
    if (MAP[1]) {
        jit_emit_mov_ri<_N>(NATIVECODE, MAP[1], <typ>_CONST[2]);
    }
    else {
        jit_emit_mov_mi<_N>(NATIVECODE, &INT_REG[1], <typ>_CONST[2]);
    }
    }

    Parrot_set_i_ic {
    Parrot_set_x_ic s/<_N>/_i/ s/<typ>/*INT/
    }

    Parrot_set_n_ic {
    Parrot_set_x_ic s/<_N>/_ni/ s/<typ>/&INT/ s/INT_R/NUM_R/
    }
The jit function Parrot_set_i_ic is based on the template Parrot_set_x_ic, the s/x/y/ are substitutions on the template body, to generate the actual function body. These substitutions are done before the other substitutions.

s. jit/i386/core.jit for more.

Naming convention for jit_emit functions ^

To make it easier to share core.jit files between machines of similar architecture, the jit_emit functions should follow this syntax:

jit_emit_<op>_<args>_<type>

<op>

This is the operation like mov, add or bxor. In normal cases this is the PASM name of the op.

<args>

args specify the arguments of the function in the PASM sequence dest, source ... The args consist of one letter per argument:

r

A mapped processor register.

m

A memory operand, the address of the parrot register.

i

An immediate operand, i.e. an integer constant.

<type>

Specifies if this operation works on integer or floating point arguments. If all arguments are of the same type, only one type specifier is needed.

i

An integer argument

n

A float argument.

Examples:

jit_emit_sub_rm_i

Subtract integer at memory from integer processor register.

jit_emit_mov_ri_ni

Move integer constant (immediate) to floating point register.

ALPHA Notes ^

The access to Parrot registers is done relative to $6, all other memory access is done relative to $27, to access float constants relative to $7 so you must preside the instruction with ldah $7,0($27).

i386 Notes ^

Only 32 bit INTVALs are supported. Long double FLOATVALs are ok.

There are four mapped integer registers %edi, %esi, %ecx, and %edx. The first 2 of these are callee saved, they preserve their value around extern function calls.

Four floating point operations the registers ST1 ... ST4 are mapped and considered as preserved over function calls.

The register %ebx holds the register frame pointer.

EXAMPLE ^

Let's see how this works:

Parrot Assembly:

 set I0,8
 set I2,I0
 print I2
 end

Parrot Bytecode: (only the bytecode segment is shown)

 +--------------------------------------+
 | 73 | 0 | 8 | 72 | 2 | 0 | 21 | 2 | 0 |
 +-|------------|------------|--------|-+
   |            |            |        |
   |            |            |        +----------- end (no arguments)
   |            |            +-------------------- print_i (1 argument)
   |            +--------------------------------- set_i_i (2 arguments)
   +---------------------------------------------- set_i_ic (2 arguments)

Please note that the opcode numbers used might have already changed. Also generated assembly code might be different.

Intel x86 assembly version of the Parrot ops:

Parrot_jit_begin

    0x817ddd0 <jit_func>:   push   %ebp
    0x817ddd1 <jit_func+1>: mov    %esp,%ebp
    0x817ddd3 <jit_func+3>: push   %ebx
    0x817ddd4 <jit_func+4>: push   %esi
    0x817ddd5 <jit_func+5>: push   %edi

  normal function header till here, now push interpreter

    0x817ddd6 <jit_func+6>: push   $0x8164420

  get jit function table to %ebp and
  jump to first instruction

    0x817dddb <jit_func+11>:    mov    0xc(%ebp),%eax
    0x817ddde <jit_func+14>:    mov    $0x81773f0,%ebp
    0x817dde3 <jit_func+19>:    sub    $0x81774a8,%eax
    0x817dde9 <jit_func+25>:    jmp    *%ds:0x0(%ebp,%eax,1)

set_i_ic

    0x817ddee <jit_func+30>:    mov    $0x8,%edi

set_i_i

    0x817ddf3 <jit_func+35>:    mov    %edi,%ebx

Parrot_jit_save_registers

    0x817ddf5 <jit_func+37>:    mov    %edi,0x8164420
    0x817ddfb <jit_func+43>:    mov    %ebx,0x8164428

Parrot_jit_normal_op

    0x817de01 <jit_func+49>:    push   $0x81774c0
    0x817de06 <jit_func+54>:    call   0x804be00 <Parrot_print_i>
    0x817de0b <jit_func+59>:    add    $0x4,%esp

Parrot_jit_end

    0x817de0e <jit_func+62>:    add    $0x4,%esp
    0x817de14 <jit_func+68>:    pop    %edi
    0x817de16 <jit_func+70>:    pop    %ebx
    0x817de18 <jit_func+72>:    pop    %esi
    0x817de1a <jit_func+74>:    pop    %ebp
    0x817de1c <jit_func+76>:    ret

Please note the reverse argument direction. PASM and JIT notations use dest,src,src, while gdb and the internal macros in jit_emit.h have src,dest.

Debugging ^

Above listing was generated by gdb, the GNU debugger, with a little help from Parrot_jit_debug, which generates a symbol file in stabs format, s. info stabs for more (or less :-()

The following script calls ddd (the graphic debugger fronted) and attaches the symbol file, after it got built in parrot_build_asm.

    # dddp
    # run ddd parrot with given file
    # gdb confirmations should be off
    parrot -o $1.pbc -d1 $1.pasm
    echo "b runops_jit
    r -D4 -j $1.pbc
    n
    add-symbol-file $1.o 0
    s
    " > .ddd

    ddd --command .ddd parrot &

Run this with e.g. dddp t/op/jit_2, then turn on the register status, step or nexti through the source, or set break points as with any other language.

You can examine parrot registers via the debugger or even set them and you can always step into external opcode and look at *interpreter.

The tests t/op/jit*.t have some test cases for testing register allocation. These tests are written for a mapping of 4 processor registers. If your processor architecture has more mapped registers, reduce them to 4 and run these tests.

Example for a debug session ^

  $ cat j.pasm
        set I0, 10
        set N1, 1.1
        set S2, "abc"
        print "\n"
        end
  $ dddp j

(ddd shows above source code and assembly (startup code snipped):

    0x815de46 <jit_func+30>:    mov    $0xa,%ebx
    0x815de4b <jit_func+35>:    fldl   0x81584c0
    0x815de51 <jit_func+41>:    fstp   %st(2)
    0x815de53 <jit_func+43>:    mov    %ebx,0x8158098
    0x815de59 <jit_func+49>:    fld    %st(1)
    0x815de5b <jit_func+51>:    fstpl  0x8158120
    0x815de61 <jit_func+57>:    push   $0x815cd90
    0x815de66 <jit_func+62>:    call   0x804db90 <Parrot_set_s_sc>
    0x815de6b <jit_func+67>:    add    $0x4,%esp
    0x815de6e <jit_func+70>:    push   $0x815cd9c
    0x815de73 <jit_func+75>:    call   0x804bcd0 <Parrot_print_sc>
    0x815de78 <jit_func+80>:    add    $0x4,%esp
    0x815de7b <jit_func+83>:    add    $0x4,%esp
    0x815de81 <jit_func+89>:    pop    %edi
    0x815de83 <jit_func+91>:    pop    %ebx
    0x815de85 <jit_func+93>:    pop    %esi
    0x815de87 <jit_func+95>:    pop    %ebp
    0x815de89 <jit_func+97>:    ret
  (gdb) n
  (gdb) n
  (gdb) n
  (gdb) p I0
  $1 = 10
  (gdb) p N1
  $2 = 1.1000000000000001
  (gdb) p *S2
  $3 = {bufstart = 0x815ad30, buflen = 15, flags = 336128, bufused =
  3, strstart = 0x815ad30 "abc"}
  (gdb) p &I0
  $4 = (INTVAL *) 0x8158098

XXX (p)rinting register contents like shown above is currently not supported.

SEE ALSO ^

docs/dev/jit_i386.pod, jit/skeleton/jit_emit.h


parrot