lightning.git - Portable just-in-time compiler library

diff --git a/lib/jit_arm-cpu.c b/lib/jit_arm-cpu.c
index 948f260..1922a6c 100644
--- a/lib/jit_arm-cpu.c
+++ b/lib/jit_arm-cpu.c

@@ -409,7 +409,7 @@ static void _torl(jit_state_t*,int,int,int) maybe_unused;

# define CC_MVN(cc,rd,rm) corrr(cc,ARM_MVN,0,rd,rm)

# define MVN(rd,rm) CC_MVN(ARM_CC_AL,rd,rm)

# define T1_MVN(rd,rm) is(THUMB_MVN|(_u3(rm)<<3)|_u3(rd))

-# define T2_MVN(rd,rm) torrr(THUMB2_MVN,rd,_R15_REGNO,rm)

+# define T2_MVN(rd,rm) torrr(THUMB2_MVN,_R15_REGNO,rd,rm)

# define CC_MVNI(cc,rd,im) corri(cc,ARM_MVN|ARM_I,0,rd,im)

# define MVNI(rd,im) CC_MVNI(ARM_CC_AL,rd,im)

# define T2_MVNI(rd,im) torri(THUMB2_MVNI,_R15_REGNO,rd,im)

@@ -816,10 +816,10 @@ static void _torl(jit_state_t*,int,int,int) maybe_unused;

# define T2_POP(im) tpp(THUMB2_POP,im)

# define jit_get_reg_args() \

do { \

- (void)jit_get_reg(_R0|jit_class_gpr); \

- (void)jit_get_reg(_R1|jit_class_gpr); \

- (void)jit_get_reg(_R2|jit_class_gpr); \

- (void)jit_get_reg(_R3|jit_class_gpr); \

+ (void)jit_get_reg(_R0|jit_class_named|jit_class_gpr); \

+ (void)jit_get_reg(_R1|jit_class_named|jit_class_gpr); \

+ (void)jit_get_reg(_R2|jit_class_named|jit_class_gpr); \

+ (void)jit_get_reg(_R3|jit_class_named|jit_class_gpr); \

} while (0)

# define jit_unget_reg_args() \

do { \

diff --git a/lib/jit_arm-swf.c b/lib/jit_arm-swf.c
index cc311df..ad0d12d 100644
--- a/lib/jit_arm-swf.c
+++ b/lib/jit_arm-swf.c

@@ -156,8 +156,8 @@ static void _swf_negr_d(jit_state_t*,jit_int32_t,jit_int32_t);

# define swf_muli_d(r0,r1,i0) swf_ddd_(__aeabi_dmul,r0,r1,i0)

# define swf_divr_f(r0,r1,r2) swf_fff(__aeabi_fdiv,r0,r1,r2)

# define swf_divi_f(r0,r1,i0) swf_fff_(__aeabi_fdiv,r0,r1,i0)

-# define swf_divr_d(r0,r1,r2) swf_ddd(__aeabi_dsub,r0,r1,r2)

-# define swf_divi_d(r0,r1,i0) swf_ddd_(__aeabi_dsub,r0,r1,i0)

+# define swf_divr_d(r0,r1,r2) swf_ddd(__aeabi_ddiv,r0,r1,r2)

+# define swf_divi_d(r0,r1,i0) swf_ddd_(__aeabi_ddiv,r0,r1,i0)

# define swf_ltr_f(r0,r1,r2) swf_iff(__aeabi_fcmplt,r0,r1,r2)

# define swf_lti_f(r0,r1,i0) swf_iff_(__aeabi_fcmplt,r0,r1,i0)

# define swf_ltr_d(r0,r1,r2) swf_idd(__aeabi_dcmplt,r0,r1,r2)

diff --git a/lib/jit_arm.c b/lib/jit_arm.c
index 319703e..b706440 100644
--- a/lib/jit_arm.c
+++ b/lib/jit_arm.c

@@ -37,10 +37,6 @@

#define jit_exchange_p() 1

/* FIXME is it really required to not touch _R10? */

-#define jit_reg_free_p(regno) \

- (!jit_regset_tstbit(_jit->reglive, regno) && \

- !jit_regset_tstbit(_jit->regarg, regno) && \

- !jit_regset_tstbit(_jit->regsav, regno))

/*

* Types

@@ -290,28 +286,28 @@ _jit_reti(jit_state_t *_jit, jit_word_t u)

void

_jit_retr_f(jit_state_t *_jit, jit_int32_t u)

{

- jit_movr_f(JIT_RET, u);

+ jit_movr_f(JIT_FRET, u);

jit_ret();

}

void

_jit_reti_f(jit_state_t *_jit, jit_float32_t u)

{

- jit_movi_f(JIT_RET, u);

+ jit_movi_f(JIT_FRET, u);

jit_ret();

}

void

_jit_retr_d(jit_state_t *_jit, jit_int32_t u)

{

- jit_movr_d(JIT_RET, u);

+ jit_movr_d(JIT_FRET, u);

jit_ret();

}

void

_jit_reti_d(jit_state_t *_jit, jit_float64_t u)

{

- jit_movi_d(JIT_RET, u);

+ jit_movi_d(JIT_FRET, u);

jit_ret();

}

diff --git a/lib/jit_x86-cpu.c b/lib/jit_x86-cpu.c
index bef8d4c..ed02008 100644
--- a/lib/jit_x86-cpu.c
+++ b/lib/jit_x86-cpu.c

@@ -40,7 +40,7 @@

# define stxi(u, v, w) stxi_l(u, v, w)

# define can_sign_extend_int_p(im) \

(((im) >= 0 && (long)(im) <= 0x7fffffffL) || \

- ((im) < 0 && (long)(im) >= -0x80000000L))

+ ((im) < 0 && (long)(im) > -0x80000000L))

# define can_zero_extend_int_p(im) \

((im) >= 0 && (im) < 0x80000000L)

# define fits_uint32_p(im) ((im & 0xffffffff00000000L) == 0)

@@ -975,8 +975,8 @@ _subi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)

lea(-i0, r1, _NOREG, _SCL1, r0);

}

else if (r0 != r1) {

- movi(r0, i0);

- isubr(r0, r1);

+ movi(r0, -i0);

+ iaddr(r0, r1);

}

else {

reg = jit_get_reg(jit_class_gpr);

@@ -1064,7 +1064,7 @@ _imuli(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)

{

jit_int32_t reg;

if (can_sign_extend_int_p(i0)) {

- rex(0, 1, r1, _NOREG, r0);

+ rex(0, 1, r0, _NOREG, r1);

if ((jit_int8_t)i0 == i0) {

ic(0x6b);

mrm(0x03, r7(r0), r7(r1));

@@ -1147,16 +1147,25 @@ _divremr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2,

{

jit_int32_t div;

jit_int32_t reg;

-

- if (r0 != _RDX_REGNO)

- (void)jit_get_reg(_RDX|jit_class_gpr);

- if (r0 != _RAX_REGNO)

- (void)jit_get_reg(_RAX|jit_class_gpr);

+ jit_int32_t set;

+ jit_int32_t use;

+

+ set = use = 0;

+ if (r0 != _RDX_REGNO && r1 != _RDX_REGNO && r2 != _RDX_REGNO)

+ set |= 1 << _RDX_REGNO;

+ if (r0 != _RAX_REGNO && r1 != _RAX_REGNO && r2 != _RAX_REGNO)

+ set |= 1 << _RAX_REGNO;

+ if (set & (1 <<_RDX_REGNO))

+ (void)jit_get_reg(_RDX|jit_class_gpr|jit_class_named);

+ if (set & (1 << _RAX_REGNO))

+ (void)jit_get_reg(_RAX|jit_class_gpr|jit_class_named);

if (r2 == _RAX_REGNO) {

if (r0 == _RAX_REGNO || r0 == _RDX_REGNO) {

if ((reg = jit_get_reg(jit_class_gpr|jit_class_chk)) == JIT_NOREG)

- reg = jit_get_reg(r1 == _RCX_REGNO ? _RBX : _RCX);

+ reg = jit_get_reg((r1 == _RCX_REGNO ? _RBX : _RCX) |

+ jit_class_gpr|jit_class_named);

+ use = 1;

div = rn(reg);

movr(div, _RAX_REGNO);

if (r1 != _RAX_REGNO)

@@ -1172,13 +1181,14 @@ _divremr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2,

movr(_RAX_REGNO, r1);

}

div = r0;

- reg = 0;

}

}

else if (r2 == _RDX_REGNO) {

if (r0 == _RAX_REGNO || r0 == _RDX_REGNO) {

if ((reg = jit_get_reg(jit_class_gpr|jit_class_chk)) == JIT_NOREG)

- reg = jit_get_reg(r1 == _RCX_REGNO ? _RBX : _RCX);

+ reg = jit_get_reg((r1 == _RCX_REGNO ? _RBX : _RCX) |

+ jit_class_gpr|jit_class_named);

+ use = 1;

div = rn(reg);

movr(div, _RDX_REGNO);

if (r1 != _RAX_REGNO)

@@ -1189,14 +1199,12 @@ _divremr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2,

movr(_RAX_REGNO, r1);

movr(r0, _RDX_REGNO);

div = r0;

- reg = 0;

}

}

else {

if (r1 != _RAX_REGNO)

movr(_RAX_REGNO, r1);

div = r2;

- reg = 0;

}

if (sign) {

@@ -1208,19 +1216,21 @@ _divremr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2,

idivr_u(div);

}

- if (reg)

+ if (use)

jit_unget_reg(reg);

if (r0 != _RAX_REGNO) {

if (divide)

movr(r0, _RAX_REGNO);

- jit_unget_reg(_RAX);

}

if (r0 != _RDX_REGNO) {

if (!divide)

movr(r0, _RDX_REGNO);

- jit_unget_reg(_RDX);

}

+ if (set & (1 <<_RDX_REGNO))

+ jit_unget_reg(_RDX);

+ if (set & (1 << _RAX_REGNO))

+ jit_unget_reg(_RAX);

}

static void

@@ -1229,6 +1239,8 @@ _divremi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0,

{

jit_int32_t reg;

jit_int32_t div;

+ jit_int32_t set;

+ jit_int32_t use;

if (divide) {

switch (i0) {

@@ -1275,23 +1287,28 @@ _divremi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0,

return;

}

- if (r0 != _RDX_REGNO)

- (void)jit_get_reg(_RDX|jit_class_gpr);

- if (r0 != _RAX_REGNO)

- (void)jit_get_reg(_RAX|jit_class_gpr);

+ set = use = 0;

+ if (r0 != _RDX_REGNO && r1 != _RDX_REGNO)

+ set |= 1 << _RDX_REGNO;

+ if (r0 != _RAX_REGNO && r1 != _RAX_REGNO)

+ set |= 1 << _RAX_REGNO;

+ if (set & (1 <<_RDX_REGNO))

+ (void)jit_get_reg(_RDX|jit_class_gpr|jit_class_named);

+ if (set & (1 << _RAX_REGNO))

+ (void)jit_get_reg(_RAX|jit_class_gpr|jit_class_named);

- if (r0 == _RAX_REGNO || r0 == _RDX_REGNO) {

+ if (r0 == _RAX_REGNO || r0 == _RDX_REGNO || r0 == r1) {

if ((reg = jit_get_reg(jit_class_gpr|jit_class_chk)) == JIT_NOREG)

- reg = jit_get_reg(_RCX);

+ reg = jit_get_reg((r1 == _RCX_REGNO ? _RBX : _RCX) |

+ jit_class_gpr|jit_class_named);

+ use = 1;

div = rn(reg);

}

- else {

- reg = 0;

+ else

div = r0;

- }

movi(div, i0);

- movr(_RAX, r1);

+ movr(_RAX_REGNO, r1);

if (sign) {

sign_extend_rdx_rax();

@@ -1302,19 +1319,21 @@ _divremi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0,

idivr_u(div);

}

- if (reg)

+ if (use)

jit_unget_reg(reg);

if (r0 != _RAX_REGNO) {

if (divide)

movr(r0, _RAX_REGNO);

- jit_unget_reg(_RAX);

}

if (r0 != _RDX_REGNO) {

if (!divide)

movr(r0, _RDX_REGNO);

- jit_unget_reg(_RDX);

}

+ if (set & (1 <<_RDX_REGNO))

+ jit_unget_reg(_RDX);

+ if (set & (1 << _RAX_REGNO))

+ jit_unget_reg(_RAX);

}

static void

@@ -1386,7 +1405,7 @@ _ori(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)

}

else if (r0 != r1) {

movi(r0, i0);

- ixorr(r0, r1);

+ iorr(r0, r1);

}

else {

reg = jit_get_reg(jit_class_gpr);

@@ -1448,6 +1467,7 @@ _rotshr(jit_state_t *_jit, jit_int32_t code,

jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)

{

jit_int32_t reg;

+ jit_int32_t use;

if (r0 == _RCX_REGNO) {

reg = jit_get_reg(jit_class_gpr);

@@ -1459,13 +1479,30 @@ _rotshr(jit_state_t *_jit, jit_int32_t code,

jit_unget_reg(reg);

}

else if (r2 != _RCX_REGNO) {

- reg = jit_get_reg(jit_class_gpr);

- movr(rn(reg), _RCX_REGNO);

- movr(_RCX_REGNO, r2);

- movr(r0, r1);

+ use = !jit_reg_free_p(_RCX);

+ if (use) {

+ reg = jit_get_reg(jit_class_gpr);

+ movr(rn(reg), _RCX_REGNO);

+ }

+ else

+ reg = 0;

+ if (r1 == _RCX_REGNO) {

+ if (r0 == r2)

+ xchgr(r0, _RCX_REGNO);

+ else {

+ movr(r0, r1);

+ movr(_RCX_REGNO, r2);

+ }

+ }

+ else {

+ movr(_RCX_REGNO, r2);

+ movr(r0, r1);

+ }

irotshr(code, r0);

- movr(_RCX_REGNO, rn(reg));

- jit_unget_reg(reg);

+ if (use) {

+ movr(_RCX_REGNO, rn(reg));

+ jit_unget_reg(reg);

+ }

}

else {

movr(r0, r1);

diff --git a/lib/lightning.c b/lib/lightning.c
index ac06ffe..69abeeb 100644
--- a/lib/lightning.c
+++ b/lib/lightning.c

@@ -148,17 +148,18 @@ _jit_get_reg(jit_state_t *_jit, jit_int32_t regspec)

jit_int32_t spec;

jit_int32_t regno;

- /* if asking for an explicit register value, assume it will

- * properly handle the case of the register also being an

- * argument for the instruction, or the register value

- * being live */

spec = regspec & ~(jit_class_chk|jit_class_nospill);

- if ((regno = jit_regno(spec))) {

+ if (spec & jit_class_named) {

+ regno = jit_regno(spec);

if (jit_regset_tstbit(_jit->regsav, regno))

/* fail if register is spilled */

goto fail;

- if (jit_regset_tstbit(_jit->regarg, regno)) {

+ if (jit_regset_tstbit(_jit->regarg, regno))

+ /* fail if register is an argument to current instruction */

+ goto fail;

+ if (jit_regset_tstbit(_jit->reglive, regno)) {

if (regspec & jit_class_nospill)

+ /* fail if register is live and should not spill/reload */

goto fail;

goto spill;

}

@@ -1453,6 +1454,12 @@ _thread_jumps(jit_state_t *_jit)

case jit_code_callr: case jit_code_calli:

/* non optimizable jump like code */

break;

+ case jit_code_beqr_f: case jit_code_beqi_f:

+ case jit_code_beqr_d: case jit_code_beqi_d:

+ case jit_code_bltgtr_f: case jit_code_bltgti_f:

+ case jit_code_bltgtr_d: case jit_code_bltgti_d:

+ /* non optimizable jump code */

+ break;

default:

mask = jit_classify(node->code);

if (mask & jit_cc_a0_jmp) {

@@ -1633,8 +1640,10 @@ reverse_jump_code(jit_code_t code)

case jit_code_blti_f: return (jit_code_bungei_f);

case jit_code_bler_f: return (jit_code_bungtr_f);

case jit_code_blei_f: return (jit_code_bungti_f);

+#if 0

case jit_code_beqr_f: return (jit_code_bltgtr_f);

case jit_code_beqi_f: return (jit_code_bltgti_f);

+#endif

case jit_code_bger_f: return (jit_code_bunltr_f);

case jit_code_bgei_f: return (jit_code_bunlti_f);

case jit_code_bgtr_f: return (jit_code_bunler_f);

@@ -1651,8 +1660,10 @@ reverse_jump_code(jit_code_t code)

case jit_code_bungei_f: return (jit_code_blti_f);

case jit_code_bungtr_f: return (jit_code_bler_f);

case jit_code_bungti_f: return (jit_code_blei_f);

+#if 0

case jit_code_bltgtr_f: return (jit_code_beqr_f);

case jit_code_bltgti_f: return (jit_code_beqi_f);

+#endif

case jit_code_bordr_f: return (jit_code_bunordr_f);

case jit_code_bordi_f: return (jit_code_bunordi_f);

case jit_code_bunordr_f:return (jit_code_bordr_f);