Re: Starting a JIT backend for Ravi yet again

[Alexander Nasonov <alnsn@...>]

Doug Currie wrote:
> Have you considered SLJIT?
> 
> https://sourceforge.net/projects/sljit/

C API is very verbose. To see what I mean, take a look at my bpfjit code [1].

To address this issue, I started Lua bindings:

https://github.com/alnsn/luaSljit

but I never finished the project.

Emit calls are much shorter in Lua and you can easily emit one asm
instruction per line (unlike C which needs several lines).

For instance, this code will generate fast_divide32 function

	local sljit = require 'sljit'
	local mul, sh1, sh2 = ...
	return sljit.create_compiler()
	    :emit_enter{args=1, saveds=1, scratches=1}
	    :emit_op2('MUL', 'R0', 'S0', sljit.imm(mul))
	    :emit_op2('LSHR', 'R0', 'R0', sljit.imm(32))
	    :emit_op2('ISUB', 'S0', 'S0', 'R0')
	    :emit_op2('ILSHR', 'S0', 'S0', sljit.imm(sh1))
	    :emit_op2('IADD', 'R0', 'R0', 'S0')
	    :emit_op2('ILSHR', 'R0', 'R0', sljit.imm(sh2))
	    :emit_return('MOV_UI', 'R0', 0)
	    :generate_code()

--
Alex

[1] https://github.com/alnsn/bpfjit/tree/master/src/bpfjit.c

It the function below, it's about 6 line per a single asm instruction.

/*
 * Emit code for BPF_LD+BPF_H+BPF_ABS    A <- P[k:2].
 */
static int
emit_read16(struct sljit_compiler *compiler, sljit_si src, uint32_t k)
{
	int status;

	BJ_ASSERT(k <= UINT32_MAX - 1);

	/* A = buf[k]; */
	status = sljit_emit_op1(compiler,
	    SLJIT_MOV_UB,
	    BJ_AREG, 0,
	    SLJIT_MEM1(src), k);
	if (status != SLJIT_SUCCESS)
		return status;

	/* tmp1 = buf[k+1]; */
	status = sljit_emit_op1(compiler,
	    SLJIT_MOV_UB,
	    BJ_TMP1REG, 0,
	    SLJIT_MEM1(src), k+1);
	if (status != SLJIT_SUCCESS)
		return status;

	/* A = A << 8; */
	status = sljit_emit_op2(compiler,
	    SLJIT_SHL,
	    BJ_AREG, 0,
	    BJ_AREG, 0,
	    SLJIT_IMM, 8);
	if (status != SLJIT_SUCCESS)
		return status;

	/* A = A + tmp1; */
	status = sljit_emit_op2(compiler,
	    SLJIT_ADD,
	    BJ_AREG, 0,
	    BJ_AREG, 0,
	    BJ_TMP1REG, 0);
	return status;
}