suzanne.soy/racket
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

jit tweaks

Browse Source 
svn: r11224
This commit is contained in:
Matthew Flatt 2008-08-13 19:04:53 +00:00
parent caf7bdd363
commit eadbc25953
6 changed files with 71 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/mzscheme/src/eval.c
View File

@ -248,6 +248,7 @@ typedef void (*DW_PrePost_Proc)(void *);
#ifdef USE_STACK_BOUNDARY_VAR
unsigned long scheme_stack_boundary;
unsigned long volatile scheme_jit_stack_boundary;
#endif
#ifdef MZ_PRECISE_GC
@ -678,6 +679,10 @@ void scheme_init_stack_check()
# endif
}
#endif
#ifdef USE_STACK_BOUNDARY_VAR
scheme_jit_stack_boundary = scheme_stack_boundary;
#endif
}

57
src/mzscheme/src/jit.c
View File

@ -997,9 +997,9 @@ static void _jit_prolog_again(mz_jit_state *jitter, int n, int ret_addr_reg)
# endif
# ifdef X86_ALIGN_STACK
/* Maintain 4-byte stack alignment. */
# define mz_prolog(x) (SUBQir(STACK_ALIGN_WORDS * JIT_WORD_SIZE, JIT_SP))
# define mz_epilog_without_jmp() ADDQir((STACK_ALIGN_WORDS + 1) * JIT_WORD_SIZE, JIT_SP)
# define mz_epilog(x) (ADDQir(STACK_ALIGN_WORDS * JIT_WORD_SIZE, JIT_SP), RET_())
# define mz_prolog(x) (ADDQiBr(-(STACK_ALIGN_WORDS * JIT_WORD_SIZE), JIT_SP))
# define mz_epilog_without_jmp() ADDQiBr((STACK_ALIGN_WORDS + 1) * JIT_WORD_SIZE, JIT_SP)
# define mz_epilog(x) (ADDQiBr(STACK_ALIGN_WORDS * JIT_WORD_SIZE, JIT_SP), RET_())
# define LOCAL_FRAME_SIZE 3
# define JIT_LOCAL3 -(JIT_WORD_SIZE * 6)
# else
@ -1036,9 +1036,13 @@ static void _jit_prolog_again(mz_jit_state *jitter, int n, int ret_addr_reg)
/* A tiny jump has to be between -128 and 127 bytes. */
# define __START_TINY_JUMPS__(cond) if (cond) { _jitl.tiny_jumps = 1; }
# define __END_TINY_JUMPS__(cond) if (cond) { _jitl.tiny_jumps = 0; }
# define __START_INNER_TINY__(cond) __END_SHORT_JUMPS__(cond); __START_TINY_JUMPS__(1);
# define __END_INNER_TINY__(cond) __END_TINY_JUMPS__(1); __START_SHORT_JUMPS__(cond);
#else
# define __START_TINY_JUMPS__(cond) __START_SHORT_JUMPS__(cond)
# define __END_TINY_JUMPS__(cond) __END_SHORT_JUMPS__(cond)
# define __START_INNER_TINY__(cond) /* empty */
# define __END_INNER_TINY__(cond) /* empty */
#endif
#define __START_TINY_OR_SHORT_JUMPS__(tcond, cond) if (tcond) { __START_TINY_JUMPS__(1); } else { __START_SHORT_JUMPS__(cond); }
@ -1862,7 +1866,10 @@ static int generate_non_tail_call(mz_jit_state *jitter, int num_rands, int direc
If num_rands < 0, then argc is in R0, and need to pop runstack before returning.
If num_rands == -1, skip prolog. */
GC_CAN_IGNORE jit_insn *ref, *ref2, *ref4, *ref5, *ref6, *ref7, *ref8, *ref9;
GC_CAN_IGNORE jit_insn *ref10, *ref11, *reftop = NULL;
GC_CAN_IGNORE jit_insn *ref10, *reftop = NULL;
#ifndef FUEL_AUTODECEREMENTS
GC_CAN_IGNORE jit_insn *ref11;
#endif
__START_SHORT_JUMPS__(1);
@ -1893,16 +1900,16 @@ static int generate_non_tail_call(mz_jit_state *jitter, int num_rands, int direc
CHECK_LIMIT();
/* Before inlined native, check stack depth: */
(void)jit_movi_p(JIT_R1, &scheme_stack_boundary);
(void)jit_movi_p(JIT_R1, &scheme_jit_stack_boundary); /* assumes USE_STACK_BOUNDARY_VAR */
jit_ldr_l(JIT_R1, JIT_R1);
ref9 = jit_bltr_ul(jit_forward(), JIT_STACK, JIT_R1);
ref9 = jit_bltr_ul(jit_forward(), JIT_STACK, JIT_R1); /* assumes down-growing stack */
CHECK_LIMIT();
#ifndef FUEL_AUTODECEREMENTS
/* Finally, check for thread swap: */
(void)jit_movi_p(JIT_R1, &scheme_fuel_counter);
jit_ldr_i(JIT_R2, JIT_R1);
ref11 = jit_blei_i(jit_forward(), JIT_R2, 0);
#ifndef FUEL_AUTODECEREMENTS
jit_subi_p(JIT_R2, JIT_R2, 0x1);
jit_str_i(JIT_R1, JIT_R2);
#endif
@ -1952,6 +1959,27 @@ static int generate_non_tail_call(mz_jit_state *jitter, int num_rands, int direc
CHECK_LIMIT();
jit_retval(JIT_R0);
VALIDATE_RESULT(JIT_R0);
/* Fast common-case return */
if (pop_and_jump) {
jit_insn *refc;
__START_INNER_TINY__(1);
refc = jit_blei_p(jit_forward(), JIT_R0, SCHEME_MULTIPLE_VALUES);
__END_INNER_TINY__(1);
if (num_rands < 0) {
/* At this point, argc must be in V1 */
jit_lshi_l(JIT_R1, JIT_V1, JIT_LOG_WORD_SIZE);
jit_addr_p(JIT_RUNSTACK, JIT_RUNSTACK, JIT_R1);
}
if (pop_and_jump) {
mz_epilog(JIT_V1);
}
__START_INNER_TINY__(1);
mz_patch_branch(refc);
__END_INNER_TINY__(1);
CHECK_LIMIT();
}
if (!multi_ok) {
jit_insn *refm;
__END_SHORT_JUMPS__(1);
@ -2037,7 +2065,9 @@ static int generate_non_tail_call(mz_jit_state *jitter, int num_rands, int direc
}
mz_patch_branch(ref4);
mz_patch_branch(ref9);
#ifndef FUEL_AUTODECEREMENTS
mz_patch_branch(ref11);
#endif
if (need_set_rs) {
JIT_UPDATE_THREAD_RSPTR();
}
@ -2066,6 +2096,7 @@ static int generate_non_tail_call(mz_jit_state *jitter, int num_rands, int direc
if (!direct_native) {
mz_patch_branch(ref10);
}
/* Note: same return code is above for faster common-case return */
if (num_rands < 0) {
/* At this point, argc must be in V1 */
jit_lshi_l(JIT_R1, JIT_V1, JIT_LOG_WORD_SIZE);
@ -3142,15 +3173,19 @@ static int generate_arith(mz_jit_state *jitter, Scheme_Object *rator, Scheme_Obj
} else if (arith == 9) {
/* min */
jit_insn *refc;
__START_INNER_TINY__(branch_short);
refc = jit_bltr_l(jit_forward(), JIT_R0, JIT_R1);
jit_movr_l(JIT_R0, JIT_R1);
mz_patch_branch(refc);
__END_INNER_TINY__(branch_short);
} else if (arith == 10) {
/* max */
jit_insn *refc;
__START_INNER_TINY__(branch_short);
refc = jit_bgtr_l(jit_forward(), JIT_R0, JIT_R1);
jit_movr_l(JIT_R0, JIT_R1);
mz_patch_branch(refc);
__END_INNER_TINY__(branch_short);
}
} else {
/* Non-constant arg is in JIT_R0 */
@ -3224,19 +3259,25 @@ static int generate_arith(mz_jit_state *jitter, Scheme_Object *rator, Scheme_Obj
} else if (arith == 9) {
/* min */
jit_insn *refc;
__START_INNER_TINY__(branch_short);
refc = jit_blti_l(jit_forward(), JIT_R0, (long)scheme_make_integer(v));
jit_movi_l(JIT_R0, (long)scheme_make_integer(v));
mz_patch_branch(refc);
__END_INNER_TINY__(branch_short);
} else if (arith == 10) {
/* max */
jit_insn *refc;
__START_INNER_TINY__(branch_short);
refc = jit_bgti_l(jit_forward(), JIT_R0, (long)scheme_make_integer(v));
jit_movi_l(JIT_R0, (long)scheme_make_integer(v));
mz_patch_branch(refc);
__END_INNER_TINY__(branch_short);
} else if (arith == 11) {
/* abs */
jit_insn *refc;
__START_INNER_TINY__(branch_short);
refc = jit_bgei_l(jit_forward(), JIT_R0, (long)scheme_make_integer(0));
__END_INNER_TINY__(branch_short);
/* watch out for most negative fixnum! */
(void)jit_beqi_p(refslow, JIT_R0, (void *)(((long)1 << ((8 * JIT_WORD_SIZE) - 1)) | 0x1));
jit_rshi_l(JIT_R0, JIT_R0, 1);
@ -3244,7 +3285,9 @@ static int generate_arith(mz_jit_state *jitter, Scheme_Object *rator, Scheme_Obj
jit_subr_l(JIT_R0, JIT_R1, JIT_R0);
jit_lshi_l(JIT_R0, JIT_R0, 1);
jit_ori_l(JIT_R0, JIT_R0, 0x1);
__START_INNER_TINY__(branch_short);
mz_patch_branch(refc);
__END_INNER_TINY__(branch_short);
CHECK_LIMIT();
}
}

2
src/mzscheme/src/lightning/i386/asm.h
View File

@ -325,6 +325,8 @@ typedef _uc jit_insn;
#define ADDQrr(RS, RD) _qO_Mrm (0x01 ,_b11,_r8(RS),_r8(RD) )
#define ADDQir(IM, RD) _qOs_Mrm_sL (0x81 ,_b11,_b000 ,_r8(RD) ,IM )
#define ADDQiBr(IM, RD) _qO_Mrm_B (0x83 ,_b11,_b000 ,_r1(RD) ,_su8(IM))
#define ANDBrr(RS, RD) _O_Mrm (0x20 ,_b11,_r1(RS),_r1(RD) )
#define ANDBmr(MD, MB, MI, MS, RD) _O_r_X (0x22 ,_r1(RD) ,MD,MB,MI,MS )
#define ANDBrm(RS, MD, MB, MI, MS) _O_r_X (0x20 ,_r1(RS) ,MD,MB,MI,MS )

4
src/mzscheme/src/port.c
View File

@ -8098,6 +8098,7 @@ static long ITimer(void)
while (1) {
if (WaitForSingleObject(itimer_semaphore, itimer_delay / 1000) == WAIT_TIMEOUT) {
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
WaitForSingleObject(itimer_semaphore, INFINITE);
}
}
@ -8140,7 +8141,8 @@ static void *run_itimer(void *p)
while (1) {
usleep(itimer_delay);
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
pthread_mutex_lock(&itimer_mutex);
if (itimer_continue) {
itimer_continue = 0;

3
src/mzscheme/src/schpriv.h
View File

@ -1172,6 +1172,9 @@ typedef struct Scheme_Overflow {
|| defined(PALM_FIND_STACK_BOUNDS)
# define USE_STACK_BOUNDARY_VAR
extern unsigned long scheme_stack_boundary;
/* Same as scheme_stack_boundary, but set to an extreme value when feul auto-expires,
so that JIT-generated code can check just one variable: */
extern unsigned long volatile scheme_jit_stack_boundary;
#endif
typedef struct Scheme_Meta_Continuation {

10
src/mzscheme/src/thread.c
View File

@ -1870,6 +1870,7 @@ void scheme_schedule_custodian_close(Scheme_Custodian *c)
scheduled_kills = scheme_make_pair((Scheme_Object *)c, scheduled_kills);
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
}
static void check_scheduled_kills()
@ -3528,6 +3529,7 @@ START_XFORM_SKIP;
static void timer_expired(int ignored)
{
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
# ifdef SIGSET_NEEDS_REINSTALL
MZ_SIGSET(SIGPROF, timer_expired);
# endif
@ -3787,8 +3789,10 @@ void scheme_break_thread(Scheme_Thread *p)
p->external_break = 1;
if (p == scheme_current_thread) {
if (scheme_can_break(p))
if (scheme_can_break(p)) {
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
}
}
scheme_weak_resume_thread(p);
# if defined(WINDOWS_PROCESSES) || defined(WINDOWS_FILE_HANDLES)
@ -4100,7 +4104,8 @@ void scheme_thread_block(float sleep_time)
if (do_atomic)
missed_context_switch = 1;
MZTHREADELEM(p, fuel_counter) = p->engine_weight;
scheme_fuel_counter = p->engine_weight;
scheme_jit_stack_boundary = scheme_stack_boundary;
#ifdef USE_ITIMER
{
@ -7162,6 +7167,7 @@ static void get_ready_for_GC()
#endif
scheme_fuel_counter = 0;
scheme_jit_stack_boundary = (unsigned long)-1;
#ifdef WINDOWS_PROCESSES
scheme_suspend_remembered_threads();