floatFunc
makeCallFloatFunc (floatFunc theFunc)
{
floatFunc retVal;
int dbl1, dbl2;
retVal = (floatFunc) jit_get_ip ().iptr;
jit_prolog (2);
dbl1 = jit_arg_f ();
dbl2 = jit_arg_f ();
jit_prepare_f (2);
jit_getarg_f (JIT_FPR0, dbl1);
jit_getarg_f (JIT_FPR1, dbl2);
jit_mulr_f (JIT_FPR1, JIT_FPR1, JIT_FPR0);
jit_pusharg_f (JIT_FPR1);
jit_pusharg_f (JIT_FPR0);
jit_finish ((void *) theFunc);
jit_retval_f (JIT_FPRET);
jit_ret ();
jit_flush_code ((char *) retVal, jit_get_ip ().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble (stderr, (char *) retVal, jit_get_ip ().ptr);
#endif
return retVal;
}
static int_return_int_t
generate_function_proxy (int_return_int_t func)
{
static jit_insn buffer[1024];
static jit_state _jit;
int_return_int_t result;
int arg;
result = (int_return_int_t)(jit_set_ip (buffer).ptr);
jit_prolog (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
/* Reset `JIT_RET'. */
jit_movi_i (JIT_RET, -1);
/* Invoke a FUNC. */
jit_prepare (1);
jit_pusharg_i (JIT_R1);
(void)jit_finish (func);
/* Copy the result of FUNC from `JIT_RET' into our own result register. */
jit_retval_i (JIT_RET);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
int main()
{
pifi nfibs;
int in; /* offset of the argument */
jit_insn *ref; /* to patch the forward reference */
jit_insn *mref; /* ref of move to backpatch */
jit_insn *tp; /* location to patch */
codeBuffer = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANON, -1, 0);
if (codeBuffer == MAP_FAILED) {
perror("mmap");
exit(0);
}
nfibs = (pifi) (jit_set_ip(codeBuffer).iptr);
jit_prolog (1);
in = jit_arg_ui ();
jit_getarg_ui(JIT_V0, in); /* V0 = n */
mref= jit_movi_p(JIT_V2,jit_forward ()); /* Generate a dumb movi */
jit_jmpr(JIT_V2);
/* generate some dump filler that will never be executed!*/
jit_addi_ui(JIT_V0,JIT_V0,1);
jit_addi_ui(JIT_V0,JIT_V0,1);
jit_addi_ui(JIT_V0,JIT_V0,1);
jit_addi_ui(JIT_V0,JIT_V0,1);
tp = jit_get_label ();
ref = jit_blti_ui (jit_forward(), JIT_V0, 2);
jit_subi_ui (JIT_V1, JIT_V0, 1); /* V1 = n-1 */
jit_subi_ui (JIT_V2, JIT_V0, 2); /* V2 = n-2 */
jit_prepare (1);
jit_pusharg_ui(JIT_V1);
jit_finish(nfibs);
jit_retval(JIT_V1); /* V1 = nfibs(n-1) */
jit_prepare(1);
jit_pusharg_ui(JIT_V2);
jit_finish(nfibs);
jit_retval(JIT_V2); /* V2 = nfibs(n-2) */
jit_addi_ui(JIT_V1, JIT_V1, 1);
jit_addr_ui(JIT_RET, JIT_V1, JIT_V2); /* RET = V1 + V2 + 1 */
jit_ret();
jit_patch(ref); /* patch jump */
jit_movi_i(JIT_RET, 1); /* RET = 1 */
jit_ret();
jit_patch_movi(mref,tp); /* Ok. Do the back-patching */
/* call the generated code, passing 32 as an argument */
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
printf("nfibs(%d) = %d\n", 32, nfibs(32));
#endif
return 0;
}
static divider_t
generate_divider (int operand, unsigned int *size)
{
divider_t result;
int arg;
buffer = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANON, -1, 0);
if (buffer == MAP_FAILED) {
perror("mmap");
exit(0);
}
result = (divider_t)(jit_set_ip (buffer).iptr);
jit_leaf (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
jit_divi_i (JIT_R2, JIT_R1, operand);
jit_movr_i (JIT_RET, JIT_R2);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
*size = (jit_insn *)jit_get_ip ().ptr - buffer;
return result;
}
mod_t
generate_modi (int operand)
{
mod_t result;
int arg;
buffer = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANON, -1, 0);
if (buffer == MAP_FAILED) {
perror("mmap");
exit(0);
}
result = (mod_t)(jit_set_ip (buffer).iptr);
jit_leaf (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
jit_modi_i (JIT_R2, JIT_R1, operand);
jit_movr_i (JIT_RET, JIT_R2);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
int main()
{
static jit_insn codeBuffer[1024];
static jit_state _jit;
pvfi myFunction; /* ptr to generated code */
char *start, *end; /* a couple of labels */
int ofs; /* to get the argument */
myFunction = (pvfi) (jit_set_ip(codeBuffer).vptr);
start = jit_get_ip().ptr;
jit_prolog(1);
ofs = jit_arg_i();
jit_movi_p(JIT_R0, "looks like %d bytes sufficed\n");
jit_getarg_i(JIT_R1, ofs);
jit_prepare_i(2);
jit_pusharg_i(JIT_R1); /* push in reverse order */
jit_pusharg_p(JIT_R0);
jit_finish(display_message);
jit_ret();
end = jit_get_ip().ptr;
jit_flush_code(codeBuffer, end);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, codeBuffer, end);
#endif
#ifndef LIGHTNING_CROSS
/* call the generated code, passing its size as argument */
myFunction(sizeof(codeBuffer));
#endif
return 0;
}
static stakumilo_t
generate_push_pop (void)
{
static const char msg[] = "we got %i\n";
stakumilo_t result;
int arg;
int retval;
buffer = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANON, -1, 0);
if (buffer == MAP_FAILED) {
perror("mmap");
exit(0);
}
result = (stakumilo_t)(jit_set_ip (buffer).ptr);
jit_prolog (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
/* Save R1 on the stack. */
jit_pushr_i (JIT_R1);
/* Save two other registers just for the sake of using the stack. */
jit_movi_i (JIT_R0, -1);
jit_movi_i (JIT_R2, -1);
jit_pushr_i (JIT_R0);
jit_pushr_i (JIT_R2);
/* most likely need stack aligned at 16 bytes, dummy push to force align */
jit_pushr_i (JIT_R2);
jit_movr_i (JIT_R0, JIT_R1);
jit_movi_p (JIT_R1, msg);
/* Invoke a function that may modify R1. */
jit_prepare (2);
jit_pusharg_i (JIT_R0);
jit_pusharg_p (JIT_R1);
(void)jit_finish (display_message);
/* dummy pop for the sake of calling function with 16 byte aligned stack */
jit_popr_i (JIT_R2);
/* Restore the dummy registers. */
jit_popr_i (JIT_R2);
jit_popr_i (JIT_R0);
/* Restore R1. */
jit_popr_i (JIT_R1);
jit_movr_i (JIT_RET, JIT_R1);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
static stakumilo_t
generate_push_pop (void)
{
static const char msg[] = "we got %i\n";
static char buffer[1024];
stakumilo_t result;
int arg;
result = (stakumilo_t)(jit_set_ip (buffer).ptr);
jit_prolog (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
/* Save R1 on the stack. */
jit_pushr_i (JIT_R1);
/* Save two other registers just for the sake of using the stack. */
jit_movi_i (JIT_R0, -1);
jit_movi_i (JIT_R2, -1);
jit_pushr_i (JIT_R0);
jit_pushr_i (JIT_R2);
jit_movr_i (JIT_R0, JIT_R1);
jit_movi_p (JIT_R1, msg);
/* Invoke a function that may modify R1. */
jit_prepare (2);
jit_pusharg_i (JIT_R0);
jit_pusharg_p (JIT_R1);
(void)jit_finish (display_message);
/* Restore the dummy registers. */
jit_popr_i (JIT_R2);
jit_popr_i (JIT_R0);
/* Restore R1. */
jit_popr_i (JIT_R1);
jit_movr_i (JIT_RET, JIT_R1);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
static mover_t
generate_movi (const void *operand)
{
static char buffer[1024];
mover_t result;
/* printf ("si?=%i ui?=%i\n", _siP (16, operand), _uiP (16, operand)); */
result = (mover_t)(jit_set_ip (buffer).iptr);
jit_leaf (1);
jit_movi_p (JIT_R0, operand);
jit_movr_p (JIT_RET, JIT_R0);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
int JIT_evaluate(ASTNode *root)
{
#ifdef __APPLE__
// Mac OS X hack
codeBuffer = mmap (NULL, 4096, PROT_EXEC | PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANON, -1, 0);
#endif
pifi function = (pifi) (jit_set_ip(codeBuffer).iptr);
jit_leaf(1);
int regs[MAX_REG] = {0};
int ret = alloc_reg(regs);
JIT_visit(root, regs, ret);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
return function();
}
mod_t
generate_modi (int operand)
{
static char buffer[1024];
mod_t result;
int arg;
result = (mod_t)(jit_set_ip (buffer).iptr);
jit_leaf (1);
arg = jit_arg_i ();
jit_getarg_i (JIT_R1, arg);
jit_modi_i (JIT_R2, JIT_R1, operand);
jit_movr_i (JIT_RET, JIT_R2);
jit_ret ();
jit_flush_code (buffer, jit_get_ip ().ptr);
return result;
}
int main()
{
pifi myFunction= (pifi) (jit_set_ip(codeBuffer).iptr);
int ofs; /* offset of the argument */
jit_leaf(1);
ofs = jit_arg_i();
jit_getarg_i(JIT_R0, ofs);
jit_addi_i(JIT_RET, JIT_R0, 1);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
/* call the generated code, passing its size as argument */
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
printf("%d + 1 = %d\n", 5, myFunction(5));
#endif
return 0;
}
/* Generate a function that computes and returns the sum of
its two double arguments (return an int)
i.e., double foo(double x,double y) { return x + y;} */
floatFunc
makeFloatFunc ()
{
floatFunc retVal;
int dbl1, dbl2;
retVal = (floatFunc) jit_get_ip ().iptr;
jit_prolog (2);
dbl1 = jit_arg_f ();
dbl2 = jit_arg_f ();
jit_getarg_f (JIT_FPR0, dbl1);
jit_getarg_f (JIT_FPR1, dbl2);
jit_addr_f (JIT_FPR0, JIT_FPR0, JIT_FPR1);
jit_movr_f (JIT_FPRET, JIT_FPR0);
jit_ret ();
jit_flush_code ((char *) retVal, jit_get_ip ().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble (stderr, (char *) retVal, jit_get_ip ().ptr);
#endif
return retVal;
}
int main()
{
pifi nfibs = (pifi) (jit_set_ip(codeBuffer).iptr);
int in; /* offset of the argument */
jit_insn *ref; /* to patch the forward reference */
jit_insn *loop; /* start of the loop */
jit_prolog (1);
in = jit_arg_ui ();
jit_getarg_ui(JIT_R2, in); /* V0 = n */
jit_movi_ui (JIT_R1, 1);
ref = jit_blti_ui (jit_forward(), JIT_R2, 2);
jit_subi_ui (JIT_R2, JIT_R2, 1);
jit_movi_ui (JIT_R0, 1);
loop= jit_get_label();
jit_subi_ui (JIT_R2, JIT_R2, 1); /* we'll calculate one more */
jit_addr_ui (JIT_V0, JIT_R0, JIT_R1); /* V0 = R0 + R1 */
jit_movr_ui (JIT_R0, JIT_R1); /* R0 = R1 */
jit_addi_ui (JIT_R1, JIT_V0, 1); /* R1 = V0 + 1 */
jit_bnei_ui (loop, JIT_R2, 0); /* if (R2) goto loop; */
jit_patch(ref); /* patch forward jump */
jit_movr_ui (JIT_RET, JIT_R1); /* RET = R1 */
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *) codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
/* call the generated code, passing 36 as an argument */
printf("nfibs(%d) = %d\n", 36, nfibs(36));
#endif
return 0;
}
int main()
{
pvfi myFunction; /* ptr to generated code */
char *start, *end; /* a couple of labels */
int ofs; /* to get the argument */
codeBuffer = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANON, -1, 0);
if (codeBuffer == MAP_FAILED) {
perror("mmap");
exit(0);
}
myFunction = (pvfi) (jit_set_ip(codeBuffer).vptr);
start = jit_get_ip().ptr;
jit_prolog(1);
ofs = jit_arg_i();
jit_movi_p(JIT_R0, "looks like %d bytes sufficed\n");
jit_getarg_i(JIT_R1, ofs);
jit_prepare_i(2);
jit_pusharg_i(JIT_R1); /* push in reverse order */
jit_pusharg_p(JIT_R0);
jit_finish(display_message);
jit_ret();
end = jit_get_ip().ptr;
jit_flush_code(codeBuffer, end);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, codeBuffer, end);
#endif
#ifndef LIGHTNING_CROSS
/* call the generated code, with a dummy argument */
myFunction(1024);
#endif
return 0;
}
/* This function does all of lexing, parsing, and picking a good
order of evaluation... Needless to say, this is not the best
possible design, but it avoids cluttering everything with globals. */
pifi
compile_rpn (char *expr)
{
struct stack_element stack[32];
int sp = 0;
int curr_tos = -1; /* stack element currently in R0 */
int spill_base, spill_sp;
pifi fn;
int ofs;
fn = (pifi) (jit_get_ip ().iptr);
jit_leaf (1);
ofs = jit_arg_i ();
spill_sp = spill_base = jit_allocai (32 * sizeof (int));
while (*expr)
{
int with_imm;
int imm;
int tok;
int src1, src2;
/* This is the lexer. */
switch (*expr)
{
case ' ': case '\t':
expr++;
continue;
case '0': case '1': case '2': case '3': case '4':
case '5': case '6': case '7': case '8': case '9':
stack[sp].kind = IMM;
stack[sp++].imm = strtol (expr, &expr, 0);
continue;
case 'x':
expr++;
stack[sp++].kind = ARG;
continue;
case '~':
/* NOT. Implemented as a XOR with -1. */
stack[sp].kind = IMM;
stack[sp++].imm = ~0;
tok = '^';
break;
case '_':
/* Unary minus. Transform to 0 - X and go on.
Also used to enter negative constants (32_ = -32). */
expr++;
stack[sp] = stack[sp - 1];
/* Ensure CURR_TOS is correct. */
if (curr_tos == sp - 1)
curr_tos = sp;
stack[sp - 1].kind = IMM;
stack[sp - 1].imm = 0;
sp++;
tok = '-';
break;
case '+':
case '-':
case '*':
case '/':
case '%':
case '&':
case '|':
case '^':
case '=':
tok = *expr++;
break;
case '!':
/* Get != */
expr++;
assert (*expr == '=');
tok = NE;
break;
case '<':
/* Get <, <<, <= */
if (expr[1] == '=')
expr += 2, tok = LE;
else if (expr[1] == '<')
expr += 2, tok = LSH;
else
expr++, tok = '<';
break;
case '>':
/* Get >, >>, >>>, >= */
if (expr[1] == '=')
expr += 2, tok = GE;
else if (expr[1] == '>' && expr[2] == '>')
expr += 3, tok = RSHU;
else if (expr[1] == '>')
expr += 2, tok = RSH;
else
expr++, tok = '>';
break;
default:
abort ();
}
assert (sp >= 2);
/* Constant folding. */
if (stack[sp - 1].kind == IMM && stack[sp - 2].kind == IMM)
{
stack[sp - 2].imm =
fold (stack[sp - 2].imm, stack[sp - 1].imm, tok);
sp--;
continue;
}
/* If possible, ensure that the constant is the RHS, possibly
by changing TOK (if it is a comparison). */
if (stack[sp - 2].kind == IMM)
{
int swapped_operation = swap_op (tok);
if (swapped_operation)
{
tok = swapped_operation;
stack[sp - 2].kind = stack[sp - 1].kind;
stack[sp - 1].kind = IMM;
stack[sp - 1].imm = stack[sp - 2].imm;
/* Ensure CURR_TOS is correct. */
if (curr_tos == sp - 1)
curr_tos = sp - 2;
}
}
/* Get the second argument into a register, if not an immediate.
Also decide which argument will be prepared into JIT_R0 and
which will be prepared into JIT_V0. */
with_imm = 0;
src1 = JIT_R0;
src2 = JIT_V0;
switch (stack[sp - 1].kind)
{
case IMM:
/* RHS is an immediate, use an immediate instruction. */
with_imm = 1;
imm = stack[sp - 1].imm;
break;
case EXPR:
/* RHS is an expression, check if it is already in JIT_R0. */
if (curr_tos == sp - 1)
{
/* Invert the two sources. */
src1 = JIT_V0;
src2 = JIT_R0;
}
else
popr (JIT_V0, &spill_sp);
curr_tos = -1;
break;
case ARG:
jit_getarg_i (JIT_V0, ofs);
break;
}
/* Get the first argument into a register indicated by SRC1. */
switch (stack[sp - 2].kind)
{
case IMM:
/* LHS is an immediate, check if we must spill the top of stack. */
if (curr_tos != -1)
{
pushr (JIT_R0, &spill_sp);
curr_tos = -1;
}
jit_movi_i (src1, stack[sp - 2].imm);
break;
case EXPR:
/* LHS is an expression, check if it is already in JIT_R0. */
if (curr_tos != sp - 2)
{
popr (src1, &spill_sp);
curr_tos = -1;
}
else
assert (src1 == JIT_R0);
break;
case ARG:
if (curr_tos != -1)
{
pushr (JIT_R0, &spill_sp);
curr_tos = -1;
}
jit_getarg_i (src1, ofs);
break;
}
/* Set up the new stack entry, which is cached in R0. */
sp -= 2;
curr_tos = sp;
stack[sp++].kind = EXPR;
/* Perform the computation. */
if (with_imm)
gen_reg_imm (src1, imm, tok);
else
gen_reg_reg (src1, src2, tok);
}
assert (sp == 1);
switch (stack[0].kind)
{
case IMM:
jit_movi_i (JIT_RET, stack[0].imm);
break;
case EXPR:
assert (curr_tos == 0);
jit_movr_i (JIT_RET, JIT_R0);
break;
case ARG:
jit_getarg_i (JIT_V0, ofs);
break;
}
jit_ret ();
jit_flush_code ((char *) fn, jit_get_ip ().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble (stderr, (char *) fn, jit_get_ip ().ptr);
#endif
return fn;
}
int
main()
{
jit_code code;
volatile double x = 0.0;
code.ptr = (char *) codeBuffer;
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_movi_d(JIT_FPR1, 0.0);
jit_gtr_d(JIT_R0, JIT_FPR0, JIT_FPR1);
jit_ltr_d(JIT_R1, JIT_FPR0, JIT_FPR1);
jit_subr_i(JIT_RET, JIT_R0, JIT_R1); /* [greater] - [less] = -1/0/1 */
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("compare", code, -2.6, -2.4, 0, 2.4, 2.6);
#endif
#ifdef __GNUC__
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_movi_d(JIT_FPR1, 0.0);
jit_eqr_d(JIT_R0, JIT_FPR0, JIT_FPR1);
jit_ltgtr_d(JIT_R1, JIT_FPR0, JIT_FPR1);
jit_lshi_i(JIT_R1, JIT_R1, 1);
jit_orr_i(JIT_RET, JIT_R0, JIT_R1);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("nans", code, x / x, 1 / (a - a), -1 / (a - a), 0.0, -2.0);
#endif
#else
printf ("nans\t\t1 3 3 0 3\n");
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_truncr_d_i(JIT_RET, JIT_FPR0);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("trunc", code, -2.6, -2.4, 0, 2.4, 2.6);
int_test("trunc", code, -3, -2, 0, 2, 3);
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_ceilr_d_i(JIT_RET, JIT_FPR0);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("ceil", code, -2.6, -2.4, 0, 2.4, 2.6);
int_test("ceil", code, -3, -2, 0, 2, 3);
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_floorr_d_i(JIT_RET, JIT_FPR0);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("floor", code, -2.6, -2.4, 0, 2.4, 2.6);
int_test("floor", code, -3, -2, 0, 2, 3);
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jit_ldi_d(JIT_FPR0, &a);
jit_roundr_d_i(JIT_RET, JIT_FPR0);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
int_test("round", code, -2.6, -2.4, 0, 2.4, 2.6);
int_test("round", code, -3, -2, 0, 2, 3);
#endif
#if 0 && defined JIT_TRANSCENDENTAL
jit_set_ip(codeBuffer);
jit_leaf(0);
jitfp_sti_d(&a,
jitfp_log(
jitfp_exp(jitfp_imm(1.0))
)
);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
code.vptr();
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
printf("log e = \t%f\n", a);
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jitfp_sti_d(&a,
jitfp_atn(
jitfp_imm(1.732050807657)
)
);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
code.vptr();
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
printf("pi = \t%f\n", a*3);
#endif
jit_set_ip(codeBuffer);
jit_leaf(0);
jitfp_sti_d(&a,
jitfp_tan(
jitfp_ldi_d(&a)
)
);
jit_ret();
jit_flush_code(codeBuffer, jit_get_ip().ptr);
code.vptr();
#ifdef LIGHTNING_DISASSEMBLE
disassemble(stderr, (char *)codeBuffer, jit_get_ip().ptr);
#endif
#ifndef LIGHTNING_CROSS
printf("tan^2 pi/3 = \t%f\n", a*a);
#endif
#endif /* JIT_TRANSCEDENTAL */
return (0);
}
void
run_gnulightning(void)
{
struct bfi * n = bfprog;
int maxstack = 0, stackptr = 0;
char *strbuf = 0;
size_t maxstrlen = 0;
#ifdef GNULIGHTv1
jit_insn** loopstack = 0;
jit_insn *codeBuffer;
void * startptr;
int argp;
#endif
#ifdef GNULIGHTv2
jit_node_t *start, *end; /* For size of code */
jit_node_t** loopstack = 0;
jit_node_t *argp;
#endif
if (cell_size == 8) tape_step = 1; else
tape_step = sizeof(int);
#ifdef GNULIGHTv1
/* TODO: Use mmap for allocating memory, the x86 execute protection
* bit is on the segment so Linux has to say thay everything below
* a specific address is executable. If you ask mmap for executable
* memory it can put it below the current value. The mprotect()
* function can't do this.
*/
if (total_nodes < 4096)
codeBuffer = malloc(65536);
else
codeBuffer = malloc(16 * total_nodes);
save_ptr_for_free(codeBuffer);
codeptr = (codeptr_t) jit_set_ip(codeBuffer).vptr;
startptr = jit_get_ip().ptr;
/* Function call prolog */
jit_prolog(1);
/* Get the data area pointer */
argp = jit_arg_p();
jit_getarg_p(REG_P, argp);
#endif
#ifdef GNULIGHTv2
init_jit(NULL); // argv[0]);
_jit = jit_new_state();
start = jit_note(__FILE__, __LINE__);
jit_prolog();
/* Get the data area pointer */
argp = jit_arg();
jit_getarg(REG_P, argp);
#endif
while(n)
{
switch(n->type)
{
case T_MOV:
if (acc_loaded)
acc_offset -= n->count;
jit_addi(REG_P, REG_P, n->count * tape_step);
break;
case T_ADD:
load_acc_offset(n->offset);
set_acc_offset(n->offset);
jit_addi(REG_ACC, REG_ACC, n->count);
break;
case T_SET:
set_acc_offset(n->offset);
if (acc_const && acc_const_val == n->count) {
;
} else if (acc_const && acc_const_val+1 == n->count) {
jit_addi(REG_ACC, REG_ACC, 1);
} else if (acc_const && acc_const_val-1 == n->count) {
jit_addi(REG_ACC, REG_ACC, -1);
} else {
jit_movi(REG_ACC, n->count);
}
acc_const = 1;
acc_const_val = n->count;
break;
case T_CALC:
if (n->offset == n->offset2 && n->count2 == 1) {
load_acc_offset(n->offset);
set_acc_offset(n->offset);
if (n->count)
jit_addi(REG_ACC, REG_ACC, n->count);
} else if (n->count2 != 0) {
load_acc_offset(n->offset2);
set_acc_offset(n->offset);
if (n->count2 == -1)
jit_negr(REG_ACC, REG_ACC);
else if (n->count2 != 1)
jit_muli(REG_ACC, REG_ACC, n->count2);
if (n->count)
jit_addi(REG_ACC, REG_ACC, n->count);
} else {
clean_acc();
set_acc_offset(n->offset);
jit_movi(REG_ACC, n->count);
if (n->count2 != 0) {
if (tape_step > 1)
jit_ldxi_i(REG_A1, REG_P, n->offset2 * tape_step);
else
jit_ldxi_uc(REG_A1, REG_P, n->offset2);
if (n->count2 == -1)
jit_negr(REG_A1, REG_A1);
else if (n->count2 != 1)
jit_muli(REG_A1, REG_A1, n->count2);
jit_addr(REG_ACC, REG_ACC, REG_A1);
}
}
if (n->count3 != 0) {
if (tape_step > 1)
jit_ldxi_i(REG_A1, REG_P, n->offset3 * tape_step);
else
jit_ldxi_uc(REG_A1, REG_P, n->offset3);
if (n->count3 == -1)
jit_negr(REG_A1, REG_A1);
else if (n->count3 != 1)
jit_muli(REG_A1, REG_A1, n->count3);
jit_addr(REG_ACC, REG_ACC, REG_A1);
}
break;
case T_IF: case T_MULT: case T_CMULT:
case T_WHL:
load_acc_offset(n->offset);
clean_acc();
acc_const = acc_loaded = 0;
if (stackptr >= maxstack) {
loopstack = realloc(loopstack,
((maxstack+=32)+2)*sizeof(*loopstack));
if (loopstack == 0) {
perror("loop stack realloc failure");
exit(1);
}
}
if (cell_mask > 0 && acc_hi_dirty) {
if (cell_mask == 0xFF)
jit_extr_uc(REG_ACC,REG_ACC);
else
jit_andi(REG_ACC, REG_ACC, cell_mask);
}
#ifdef GNULIGHTv1
loopstack[stackptr] = jit_beqi_i(jit_forward(), REG_ACC, 0);
loopstack[stackptr+1] = jit_get_label();
#endif
#ifdef GNULIGHTv2
loopstack[stackptr] = jit_beqi(REG_ACC, 0);
loopstack[stackptr+1] = jit_label();
#endif
stackptr += 2;
break;
case T_END:
load_acc_offset(n->offset);
clean_acc();
stackptr -= 2;
if (stackptr < 0) {
fprintf(stderr, "Code gen failure: Stack pointer negative.\n");
exit(1);
}
if (cell_mask > 0 && acc_hi_dirty) {
if (cell_mask == 0xFF)
jit_extr_uc(REG_ACC,REG_ACC);
else
jit_andi(REG_ACC, REG_ACC, cell_mask);
}
#ifdef GNULIGHTv1
jit_bnei_i(loopstack[stackptr+1], REG_ACC, 0);
jit_patch(loopstack[stackptr]);
#endif
#ifdef GNULIGHTv2
{
jit_node_t *ref;
ref = jit_bnei(REG_ACC, 0);
jit_patch_at(ref, loopstack[stackptr+1]);
jit_patch(loopstack[stackptr]);
}
#endif
break;
case T_ENDIF:
clean_acc();
acc_const = acc_loaded = 0;
stackptr -= 2;
if (stackptr < 0) {
fprintf(stderr, "Code gen failure: Stack pointer negative.\n");
exit(1);
}
jit_patch(loopstack[stackptr]);
break;
case T_PRT:
clean_acc();
load_acc_offset(n->offset);
acc_loaded = 0;
#ifdef GNULIGHTv1
jit_prepare_i(1);
jit_pusharg_i(REG_ACC);
jit_finish(putch);
#endif
#ifdef GNULIGHTv2
jit_prepare();
jit_pushargr(REG_ACC);
jit_finishi(putch);
#endif
break;
case T_CHR:
clean_acc();
acc_const = acc_loaded = 0;
if (n->count <= 0 || (n->count >= 127 && iostyle == 1) ||
!n->next || n->next->type != T_CHR) {
jit_movi(REG_ACC, n->count);
#ifdef GNULIGHTv1
jit_prepare_i(1);
jit_pusharg_i(REG_ACC);
jit_finish(putch);
#endif
#ifdef GNULIGHTv2
jit_prepare();
jit_pushargr(REG_ACC);
jit_finishi(putch);
#endif
} else {
unsigned i = 0;
struct bfi * v = n;
char *s;
while(v->next && v->next->type == T_CHR &&
v->next->count > 0 &&
(v->next->count < 127 || iostyle != 1)) {
if (i+2 > maxstrlen) {
if (maxstrlen) maxstrlen *= 2; else maxstrlen = 4096;
strbuf = realloc(strbuf, maxstrlen);
if (!strbuf) {
fprintf(stderr, "Reallocate of string buffer failed\n");
exit(42);
}
}
strbuf[i++] = (char) /*GCC -Wconversion*/ v->count;
n = v;
v = v->next;
}
strbuf[i] = 0;
s = strdup(strbuf);
if (!s) {
fprintf(stderr, "Save of string failed\n");
exit(43);
}
save_ptr_for_free(s);
#ifdef GNULIGHTv1
jit_movi_p(REG_ACC, s);
jit_prepare_i(1);
jit_pusharg_i(REG_ACC);
jit_finish(puts_without_nl);
#endif
#ifdef GNULIGHTv2
jit_prepare();
jit_pushargi((jit_word_t) s);
jit_finishi(puts_without_nl);
#endif
}
break;
case T_INP:
load_acc_offset(n->offset);
set_acc_offset(n->offset);
#ifdef GNULIGHTv1
jit_prepare_i(1);
jit_pusharg_i(REG_ACC);
jit_finish(getch);
jit_retval_i(REG_ACC);
#endif
#ifdef GNULIGHTv2
jit_prepare();
jit_pushargr(REG_ACC);
jit_finishi(getch);
jit_retval(REG_ACC);
#endif
break;
case T_STOP:
#ifdef GNULIGHTv1
jit_prepare_i(0);
jit_finish(failout);
#endif
#ifdef GNULIGHTv2
jit_prepare();
jit_finishi(failout);
#endif
break;
case T_NOP:
case T_DUMP:
fprintf(stderr, "Warning on code generation: "
"%s node: ptr+%d, cnt=%d, @(%d,%d).\n",
tokennames[n->type],
n->offset, n->count, n->line, n->col);
break;
default:
fprintf(stderr, "Code gen error: "
"%s\t"
"%d:%d, %d:%d, %d:%d\n",
tokennames[n->type],
n->offset, n->count,
n->offset2, n->count2,
n->offset3, n->count3);
exit(1);
}
n=n->next;
#if 0 /*def GNULIGHTv2 */
if(n && enable_trace) {
char *p, buf[250];
clean_acc();
acc_loaded = 0;
sprintf(buf, "@(%d,%d)\n", n->line, n->col);
p = strdup(buf);
save_ptr_for_free(p);
jit_prepare();
jit_pushargi((jit_word_t) p);
jit_finishi(puts_without_nl);
}
#endif
#ifdef GNULIGHTv1
/* TODO -- Check for codeBuffer overflow (add jmp to new) */
#endif
}
jit_ret();
if (strbuf) { maxstrlen = 0; free(strbuf); strbuf = 0; }
delete_tree();
#ifdef GNULIGHTv1
jit_flush_code(startptr, jit_get_ip().ptr);
if (verbose)
fprintf(stderr, "Generated %d bytes of V1 GNU Lightning code, running\n",
(int)(jit_get_ip().ptr - (char*)startptr));
start_runclock();
codeptr(map_hugeram());
finish_runclock(&run_time, &io_time);
#endif
#ifdef GNULIGHTv2
jit_epilog();
end = jit_note(__FILE__, __LINE__);
codeptr = jit_emit();
if (verbose)
fprintf(stderr, "Generated %d bytes of V2 GNU Lightning code, running\n",
(int)((char*)jit_address(end) - (char*)jit_address(start)));
jit_clear_state();
// jit_disassemble();
start_runclock();
codeptr(map_hugeram());
finish_runclock(&run_time, &io_time);
#endif
#if 0
/* This code writes the generated instructions to a file
* so we can disassemble it using: ndisasm -b 32/64 code.bin */
{
#ifdef GNULIGHTv1
char *p = startptr;
int s = jit_get_ip().ptr - p;
#endif
#ifdef GNULIGHTv2
char *p = (char*)jit_address(start);
int s = (char*)jit_address(end) - p;
#endif
FILE *fp = fopen("code.bin", "w");
int i;
for (i = 0; i < s; ++i) {
fputc(p[i], fp);
}
fclose(fp);
}
#endif
#ifdef GNULIGHTv2
jit_destroy_state();
finish_jit();
#endif
codeptr = 0;
if (loopstack) { free(loopstack); loopstack = 0; }
free_saved_memory();
}
pdfd
compile_rpn (char *expr)
{
pdfd fn;
int ofs, sp = 1;
fn = (pdfd) (jit_get_ip ().dptr);
jit_leaf (1);
ofs = jit_arg_d ();
jit_getarg_d (regs[0], ofs);
while (*expr)
{
char buf[32];
int n;
/* This scanner is much less advanced than the one in rpn.c. */
if (sscanf (expr, "%[0-9]%n", buf, &n))
{
double d = strtod (buf, NULL);
expr += n - 1;
jit_movi_d (regs[sp], d);
sp++;
}
else if (*expr == '+')
{
jit_addr_d (regs[sp - 2], regs[sp - 2], regs[sp - 1]);
sp--;
}
else if (*expr == '-')
{
jit_subr_d (regs[sp - 2], regs[sp - 2], regs[sp - 1]);
sp--;
}
else if (*expr == '*')
{
jit_mulr_d (regs[sp - 2], regs[sp - 2], regs[sp - 1]);
sp--;
}
else if (*expr == '/')
{
jit_divr_d (regs[sp - 2], regs[sp - 2], regs[sp - 1]);
sp--;
}
else
{
fprintf (stderr, "cannot compile: %s\n", expr);
abort ();
}
++expr;
}
jit_movr_d (JIT_FPRET, regs[0]);
jit_ret ();
jit_flush_code ((char *) fn, jit_get_ip ().ptr);
#ifdef LIGHTNING_DISASSEMBLE
disassemble (stderr, (char *) fn, jit_get_ip ().ptr);
#endif
return fn;
}
