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;
}
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;
}
jit_pointer_t compile(const std::vector<Oper>& ops, jit_word_t *memory, const bool flush = true)
{
jit_prolog();
jit_movi(JIT_V0, reinterpret_cast<jit_word_t>(memory));
jit_movi(JIT_V1, 0);
std::stack<Loop> loops;
jit_node_t* start = jit_note(__FILE__, __LINE__);
for ( size_t n=0; n<ops.size(); ++n ) {
switch ( ops[n].code ) {
case '<':
jit_str(JIT_V0, JIT_V1);
jit_subi(JIT_V0, JIT_V0, ops[n].count * sizeof(jit_word_t));
jit_ldr(JIT_V1, JIT_V0);
break;
case '>':
jit_str(JIT_V0, JIT_V1);
jit_addi(JIT_V0, JIT_V0, ops[n].count * sizeof(jit_word_t));
jit_ldr(JIT_V1, JIT_V0);
break;
case 'z':
jit_movi(JIT_V1, 0);
break;
case '+':
jit_addi(JIT_V1, JIT_V1, ops[n].count);
break;
case '-':
jit_subi(JIT_V1, JIT_V1, ops[n].count);
break;
case '.':
jit_prepare();
jit_pushargr(JIT_V1);
jit_finishi(reinterpret_cast<jit_pointer_t>(putchar));
if ( flush ) {
jit_prepare();
jit_pushargi(reinterpret_cast<jit_word_t>(stdout));
jit_finishi(reinterpret_cast<jit_pointer_t>(fflush));
}
break;
case ',':
jit_prepare();
jit_finishi(reinterpret_cast<jit_pointer_t>(getchar));
jit_retval(JIT_V1);
break;
case '[': {
Loop loop;
loop.end = jit_forward();
jit_node_t *j = jit_beqi(JIT_V1, 0);
jit_patch_at(j, loop.end);
loop.body = jit_label();
loops.push(loop);
} break;
case ']': {
Loop loop = loops.top();
jit_node_t *j = jit_bnei(JIT_V1, 0);
jit_patch_at(j, loop.body);
jit_link(loop.end);
loops.pop();
break;
}
default:
break;
}
}
jit_node_t* stop = jit_note(__FILE__, __LINE__);
jit_ret();
jit_epilog();
jit_pointer_t r = jit_emit();
fprintf(stderr, "compiled to %zu bytes\n", (char*)jit_address(stop) -
(char*)jit_address(start));
return r;
}
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();
}
int scheme_inline_alloc(mz_jit_state *jitter, int amt, Scheme_Type ty, int flags,
int keep_r0_r1, int keep_fpr1, int inline_retry,
int keep_extfpr1)
/* Puts allocated result at JIT_V1; first word is GC tag.
Uses JIT_R2 as temporary. The allocated memory is "dirty" (i.e., not 0ed).
Save FP0 when FP ops are enabled. */
{
GC_CAN_IGNORE jit_insn *ref, *reffail;
#ifdef MZ_GC_STRESS_TESTING
GC_CAN_IGNORE jit_insn *refstress;
#endif
intptr_t a_word, sz, algn;
sz = GC_compute_alloc_size(amt);
algn = GC_alloc_alignment();
__START_TINY_JUMPS__(1);
#ifdef MZ_GC_STRESS_TESTING
refstress = jit_jmpi(jit_forward());
#endif
reffail = jit_get_ip();
mz_tl_ldi_p(JIT_V1, tl_GC_gen0_alloc_page_ptr);
jit_subi_l(JIT_R2, JIT_V1, 1);
jit_andi_l(JIT_R2, JIT_R2, (algn - 1));
ref = jit_blti_l(jit_forward(), JIT_R2, (algn - sz));
#ifdef MZ_GC_STRESS_TESTING
mz_patch_ucbranch(refstress);
#endif
CHECK_LIMIT();
__END_TINY_JUMPS__(1);
/* Failure handling */
if (inline_retry) {
int mode;
if (keep_r0_r1)
mode = 1;
else if (keep_fpr1)
mode = 2;
else if (keep_extfpr1)
mode = 3;
else
mode = 0;
scheme_generate_alloc_retry(jitter, mode);
CHECK_LIMIT();
} else {
if (keep_r0_r1) {
(void)jit_calli(sjc.retry_alloc_code_keep_r0_r1);
} else if (keep_fpr1) {
(void)jit_calli(sjc.retry_alloc_code_keep_fpr1);
#ifdef MZ_LONG_DOUBLE
} else if (keep_extfpr1) {
(void)jit_calli(sjc.retry_alloc_code_keep_extfpr1);
#endif
} else {
(void)jit_calli(sjc.retry_alloc_code);
}
}
__START_TINY_JUMPS__(1);
(void)jit_jmpi(reffail);
__END_SHORT_JUMPS__(1);
__START_TINY_JUMPS__(1);
mz_patch_branch(ref);
jit_addi_ul(JIT_R2, JIT_V1, sz);
(void)mz_tl_sti_l(tl_GC_gen0_alloc_page_ptr, JIT_R2, JIT_R0);
/* GC header: */
if (ty >= 0) {
if ((ty == scheme_pair_type)
|| (ty == scheme_mutable_pair_type)
|| (ty == scheme_raw_pair_type))
a_word = GC_pair_initial_word(amt);
else
a_word = GC_initial_word(amt);
jit_stir_l(JIT_V1, a_word);
/* Scheme_Object header: */
a_word = initial_tag_word(ty, flags);
jit_stixi_l(sizeof(intptr_t), JIT_V1, a_word);
} else {
/* an array of pointers */
a_word = GC_array_initial_word(amt);
jit_stir_l(JIT_V1, a_word);
}
CHECK_LIMIT();
__END_TINY_JUMPS__(1);
return 1;
}
int scheme_inline_alloc(mz_jit_state *jitter, int amt, Scheme_Type ty, int immut,
int keep_r0_r1, int keep_fpr1, int inline_retry)
/* Puts allocated result at JIT_V1; first word is GC tag.
Uses JIT_R2 as temporary. The allocated memory is "dirty" (i.e., not 0ed).
Save FP0 when FP ops are enabled. */
{
GC_CAN_IGNORE jit_insn *ref, *reffail;
intptr_t a_word, sz, algn;
sz = GC_compute_alloc_size(amt);
algn = GC_alloc_alignment();
__START_TINY_JUMPS__(1);
reffail = _jit.x.pc;
mz_tl_ldi_p(JIT_V1, tl_GC_gen0_alloc_page_ptr);
jit_subi_l(JIT_R2, JIT_V1, 1);
jit_andi_l(JIT_R2, JIT_R2, (algn - 1));
ref = jit_blti_l(jit_forward(), JIT_R2, (algn - sz));
CHECK_LIMIT();
__END_TINY_JUMPS__(1);
/* Failure handling */
if (keep_r0_r1) {
if (inline_retry) {
scheme_generate_alloc_retry(jitter, 1);
CHECK_LIMIT();
} else {
(void)jit_calli(sjc.retry_alloc_code_keep_r0_r1);
}
} else if (keep_fpr1) {
(void)jit_calli(sjc.retry_alloc_code_keep_fpr1);
} else {
(void)jit_calli(sjc.retry_alloc_code);
}
__START_TINY_JUMPS__(1);
(void)jit_jmpi(reffail);
__END_SHORT_JUMPS__(1);
__START_TINY_JUMPS__(1);
mz_patch_branch(ref);
jit_addi_ul(JIT_R2, JIT_V1, sz);
(void)mz_tl_sti_l(tl_GC_gen0_alloc_page_ptr, JIT_R2, JIT_R0);
/* GC header: */
if (ty >= 0) {
a_word = GC_initial_word(amt);
jit_movi_l(JIT_R2, a_word);
jit_str_l(JIT_V1, JIT_R2);
/* Scheme_Object header: */
a_word = initial_tag_word(ty, immut);
jit_movi_l(JIT_R2, a_word);
jit_stxi_l(sizeof(intptr_t), JIT_V1, JIT_R2);
} else {
/* an array of pointers */
a_word = GC_array_initial_word(amt);
jit_movi_l(JIT_R2, a_word);
jit_str_l(JIT_V1, JIT_R2);
}
CHECK_LIMIT();
__END_TINY_JUMPS__(1);
return 1;
}
