NAME
docs/dev/jit_i386.pod - Parrot JIT (i386/gcc)
ABSTRACT
This PDD describes the i386 gcc JIT implementation.
DESCRIPTION
JIT i386/gcc is a combination of unrolled assembly instructions and the Computed Goto Predereferenced (CGP) run loop. For branch instructions the function implementation in the standard core is called.
Another difference of JIT/i386 is that most vtable functions are JITed instructions which use register mappings.
For a better understanding of the control flow between these basically 3 run loop cores, an example shows the gory details.
EXAMPLE
Given the following PASM program, the righthand three columns show where each opcode gets executed:
PASM JIT ops Normal CGP ops
set I0, 10 # set_i_ic print I0 # (call) print_i print "\n" # print_sc bsr inc # (call) bsr_ic cpu_ret end # (jmp) HALT end (ret) # end (ret) inc: inc I0 # inc_i new P0, 'String' # new_p_sc set P0, I0 # set_p_i print P0 # (call) print_p print "\n" # print_sc ret # (call) ret cpu_ret
Startup sequence
In runops_jit a prederefed copy of the opcode stream is built by init_prederef. Then build_asm generates the assembler code sequence as usual. This generated code (shown as runops_jit in ddd) is then executed.
Generate minimal stack frame, save %ebx
0x812c510 <jit_func>: push %ebp
0x812c511 <jit_func+1>: mov %esp,%ebp
0x812c513 <jit_func+3>: push %ebx
Get the program counter to %ebx
0x812c514 <jit_func+4>: mov 0xc(%ebp),%ebx
Push interpreter and (opcode_t*) 1 and call cgp_core
0x812c517 <jit_func+7>: push $0x8113db8
0x812c51c <jit_func+12>: push $0x1
0x812c521 <jit_func+17>: mov $0x1,%eax
0x812c526 <jit_func+22>: call 0x80b5830 <cgp_core>
In cgp_core all callee saved registers are saved.
0x80b5830 <cgp_core>: push %ebp
0x80b5831 <cgp_core+1>: mov %esp,%ebp
0x80b5833 <cgp_core+3>: sub $0xdc,%esp
0x80b5839 <cgp_core+9>: lea 0x8(%ebp),%eax
0x80b583c <cgp_core+12>: push %edi
0x80b583d <cgp_core+13>: push %esi
0x80b583e <cgp_core+14>: push %ebx
In %eax the init flag is set to -1
0x80b583f <cgp_core+15>: mov %eax,0xfffffff
The parameter *cur_op (the program counter) is put into %esi and ...
0x80b5842 <cgp_core+18>: mov 0x8(%ebp),%esi
0x80b5845 <cgp_core+21>: test %esi,%esi
0x80b5847 <cgp_core+23>: jne 0x80b5853 <cgp_core+35>
0x80b5849 <cgp_core+25>: mov $0x810ca60,%eax
0x80b584e <cgp_core+30>: jmp 0x80bb470 <cgp_core+23616>
... compared to 1
0x80b5853 <cgp_core+35>: cmp $0x1,%esi
0x80b5856 <cgp_core+38>: jne 0x80b5860 <cgp_core+48>
If true, the program jumps to the return address of above function call, i.e. it jumps back again to JIT code.
0x80b5858 <cgp_core+40>: jmp *0x4(%ebp)
Back again in JIT code, the init flag is checked
0x812c52b <jit_func+27>: test %eax,%eax
0x812c52d <jit_func+29>: jne 0x812c536 <jit_func+38>
... and if zero, the function would be left.
[ 0x812c52f <jit_func+31>: pop %ebx ] [ 0x812c531 <jit_func+33>: mov %ebp,%esp ] [ 0x812c533 <jit_func+35>: pop %ebp ] [ 0x812c535 <jit_func+37>: ret ]
When coming from the init sequence, program flow continues by checking the resume_offset and jumping to the desired instruction
0x812c536 <jit_func+38>: mov %ebx,%eax
0x812c538 <jit_func+40>: sub $0x400140c0,%eax
0x812c53e <jit_func+46>: mov $0x812c4a8,%edx
0x812c543 <jit_func+51>: jmp *(%edx,%eax,1)
set I0, 10 and save_registers
0x812c546 <jit_func+54>: mov $0xa,%ebx
0x812c54b <jit_func+59>: mov %ebx,0x8113db8
Now non-JITed code follows -- get the address from the prederefed op_func_table and call it:
0x812c551 <jit_func+65>: mov $0x812ac0c,%esi
0x812c556 <jit_func+70>: call *(%esi)
inline op print(in INT) {
printf(INTVAL_FMT, (INTVAL)$1);
goto NEXT();
}
where the goto NEXT() is a simple:
0x80b5b49 <cgp_core+793>: jmp *(%esi)
op print(in STR) {
...
goto NEXT();
}
As the last instruction of the non-JITed code sequence is a branch, this is not executed in CGP, but the opcode:
inline op cpu_ret() {
#ifdef __GNUC__
# ifdef I386
asm("ret")
is executed. This opcode is patched into the prederefed code stream by Parrot_jit_normal_op at the end of a non-JITed code sequence. This returns to JIT code again, where the next instruction gets called as a function in the standard core ...
0x812c558 <jit_func+72>: push $0x8113db8
0x812c55d <jit_func+77>: push $0x400140dc
0x812c562 <jit_func+82>: call 0x805be60 <Parrot_bsr_ic>
0x812c567 <jit_func+87>: add $0x8,%esp
... and from the return result in %eax, the new code position in JIT is calculated and gets jumped to:
0x812c56a <jit_func+90>: sub $0x400140c0,%eax
0x812c570 <jit_func+96>: mov $0x812c4a8,%edx
0x812c575 <jit_func+101>: jmp *(%edx,%eax,1)
Now in the subroutine inc:
0x812c580 <jit_func+112>: mov 0x8113db8,%ebx
0x812c586 <jit_func+118>: inc %ebx
Save register and arguments and call pmc_new_noinit:
0x812c587 <jit_func+119>: push %edx
0x812c588 <jit_func+120>: push $0x11
0x812c58d <jit_func+125>: push $0x8113db8
0x812c592 <jit_func+130>: call 0x806fc60 <pmc_new_noinit>
put the PMC* into Parrot's register:
0x812c597 <jit_func+135>: mov %eax,0x8113fb8
and prepare arguments for a VTABLE call:
0x812c59d <jit_func+141>: push %eax
0x812c59e <jit_func+142>: push $0x8113db8
0x812c5a3 <jit_func+147>: mov 0x10(%eax),%eax
0x812c5a6 <jit_func+150>: call *0x18(%eax)
0x812c5a9 <jit_func+153>: add $0x10,%esp
0x812c5ac <jit_func+156>: pop %edx
and another one:
0x812c5ae <jit_func+158>: push %edx
Here, with the mapped register in %ebx, push I0, the PMC and the interpreter:
0x812c5af <jit_func+159>: push %ebx
0x812c5b0 <jit_func+160>: mov 0x8113fb8,%eax
0x812c5b6 <jit_func+166>: push %eax
0x812c5b7 <jit_func+167>: push $0x8113db8
and call the vtable:
0x812c5bc <jit_func+172>: mov 0x10(%eax),%eax
0x812c5bf <jit_func+175>: call *0xdc(%eax)
0x812c5c5 <jit_func+181>: add $0xc,%esp
0x812c5c8 <jit_func+184>: pop %edx
As this ends the JITed section, used registers are saved back to Parrot's register:
0x812c5ca <jit_func+186>: mov %ebx,0x8113db8
and again the code in cgp_core gets called:
0x812c5d0 <jit_func+192>: mov $0x812ac48,%esi
0x812c5d5 <jit_func+197>: call *(%esi)
which after executing the print returns back here in JIT, where the ret is called:
0x812c5d7 <jit_func+199>: push $0x8113db8
0x812c5dc <jit_func+204>: push $0x40014118
0x812c5e1 <jit_func+209>: call 0x805d5e0 <Parrot_ret>
0x812c5e6 <jit_func+214>: add $0x8,%esp
From the returned PC a JIT address is calculated, which gets executed:
0x812c5e9 <jit_func+217>: sub $0x400140c0,%eax
0x812c5ef <jit_func+223>: mov $0x812c4a8,%edx
0x812c5f4 <jit_func+228>: jmp *(%edx,%eax,1)
Now at the end opcode, the CGP code for HALT() gets jumped to:
0x812c578 <jit_func+104>: mov $0x80b5877,%esi
0x812c57d <jit_func+109>: jmp *%esi
which is:
inline op end() {
HALT();
}
or, set return result:
0x80b8b6f <cgp_core+13119>: xor %eax,%eax
...
and clean up stack frame and ret:
0x80bb470 <cgp_core+23616>: lea 0xffffff18(%ebp),%esp
0x80bb476 <cgp_core+23622>: pop %ebx
0x80bb477 <cgp_core+23623>: pop %esi
0x80bb478 <cgp_core+23624>: pop %edi
0x80bb479 <cgp_core+23625>: mov %ebp,%esp
0x80bb47b <cgp_core+23627>: pop %ebp
0x80bb47c <cgp_core+23628>: ret
This returns after the position where cgp_core was called during the init sequence, but now the return value %eax is zero and the..
0x812c52b <jit_func+27>: test %eax,%eax
0x812c52d <jit_func+29>: jne 0x812c536 <jit_func+38>
0x812c52f <jit_func+31>: pop %ebx
0x812c531 <jit_func+33>: mov %ebp,%esp
0x812c533 <jit_func+35>: pop %ebp
0x812c535 <jit_func+37>: ret
... whole story ends here, we are back again in runops_jit.
So this is rather simple once it gets going.
BUGS
The floating point registers do not get saved to Parrot before vtable calls. This assumes that external routines preserve the FP stack pointer and don't use more the 4 floating point registers at once.
AUTHOR
Leopold Toetsch lt@toetsch.at
VERSION
CURRENT
14.02.2003 by Leopold Toetsch