/* Record N-ary bit.band, bit.bor, bit.bxor. */
static void LJ_FASTCALL recff_bit_nary(jit_State *J, RecordFFData *rd)
{
TRef tr = lj_opt_narrow_tobit(J, J->base[0]);
uint32_t op = rd->data;
BCReg i;
for (i = 1; J->base[i] != 0; i++)
tr = emitir(IRTI(op), tr, lj_opt_narrow_tobit(J, J->base[i]));
J->base[0] = tr;
}
static void LJ_FASTCALL recff_rawlen(jit_State *J, RecordFFData *rd)
{
TRef tr = J->base[0];
if (tref_isstr(tr))
J->base[0] = emitir(IRTI(IR_FLOAD), tr, IRFL_STR_LEN);
else if (tref_istab(tr))
J->base[0] = lj_ir_call(J, IRCALL_lj_tab_len, tr);
/* else: Interpreter will throw. */
UNUSED(rd);
}
/* Get a pointer to the other 32 bit word (LE: hiword, BE: loword). */
static IRRef split_ptr(jit_State *J, IRRef ref)
{
IRIns *ir = IR(ref);
int32_t ofs = 4;
if (ir->o == IR_ADD && irref_isk(ir->op2)) { /* Reassociate address. */
ofs += IR(ir->op2)->i;
ref = ir->op1;
if (ofs == 0) return ref;
}
return split_emit(J, IRTI(IR_ADD), ref, lj_ir_kint(J, ofs));
}
/* Get a pointer to the other 32 bit word (LE: hiword, BE: loword). */
static IRRef split_ptr(jit_State *J, IRIns *oir, IRRef ref)
{
IRRef nref = oir[ref].prev;
IRIns *ir = IR(nref);
int32_t ofs = 4;
if (ir->o == IR_ADD && irref_isk(ir->op2) && !irt_isphi(oir[ref].t)) {
/* Reassociate address. */
ofs += IR(ir->op2)->i;
nref = ir->op1;
if (ofs == 0) return nref;
}
return split_emit(J, IRTI(IR_ADD), nref, lj_ir_kint(J, ofs));
}
/* Emit a CALLN with one split 64 bit argument. */
static IRRef split_call_l(jit_State *J, IRRef1 *hisubst, IRIns *oir,
IRIns *ir, IRCallID id)
{
IRRef tmp, op1 = ir->op1;
J->cur.nins--;
#if LJ_LE
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), oir[op1].prev, hisubst[op1]);
#else
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), hisubst[op1], oir[op1].prev);
#endif
ir->prev = tmp = split_emit(J, IRTI(IR_CALLN), tmp, id);
return split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), tmp, tmp);
}
static void LJ_FASTCALL recff_ipairs_aux(jit_State *J, RecordFFData *rd)
{
RecordIndex ix;
ix.tab = J->base[0];
if (tref_istab(ix.tab)) {
if (!tvisnumber(&rd->argv[1])) /* No support for string coercion. */
lj_trace_err(J, LJ_TRERR_BADTYPE);
setintV(&ix.keyv, numberVint(&rd->argv[1])+1);
settabV(J->L, &ix.tabv, tabV(&rd->argv[0]));
ix.val = 0; ix.idxchain = 0;
ix.key = lj_opt_narrow_toint(J, J->base[1]);
J->base[0] = ix.key = emitir(IRTI(IR_ADD), ix.key, lj_ir_kint(J, 1));
J->base[1] = lj_record_idx(J, &ix);
rd->nres = tref_isnil(J->base[1]) ? 0 : 2;
} /* else: Interpreter will throw. */
}
/* Narrowing of modulo operator. */
TRef lj_opt_narrow_mod(jit_State *J, TRef rb, TRef rc)
{
TRef tmp;
if ((J->flags & JIT_F_OPT_NARROW) &&
tref_isk(rc) && tref_isint(rc)) { /* Optimize x % k. */
int32_t k = IR(tref_ref(rc))->i;
if (k > 0 && (k & (k-1)) == 0) { /* i % 2^k ==> band(i, 2^k-1) */
if (tref_isinteger(rb))
return emitir(IRTI(IR_BAND), rb, lj_ir_kint(J, k-1));
}
}
/* b % c ==> b - floor(b/c)*c */
rb = lj_ir_tonum(J, rb);
rc = lj_ir_tonum(J, rc);
tmp = emitir(IRTN(IR_DIV), rb, rc);
tmp = emitir(IRTN(IR_FPMATH), tmp, IRFPM_FLOOR);
tmp = emitir(IRTN(IR_MUL), tmp, rc);
return emitir(IRTN(IR_SUB), rb, tmp);
}
/* Emit a CALLN with two split 64 bit arguments. */
static IRRef split_call_ll(jit_State *J, IRRef1 *hisubst, IRIns *oir,
IRIns *ir, IRCallID id)
{
IRRef tmp, op1 = ir->op1, op2 = ir->op2;
J->cur.nins--;
#if LJ_LE
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), oir[op1].prev, hisubst[op1]);
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, oir[op2].prev);
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, hisubst[op2]);
#else
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), hisubst[op1], oir[op1].prev);
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, hisubst[op2]);
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, oir[op2].prev);
#endif
ir->prev = tmp = split_emit(J, IRTI(IR_CALLN), tmp, id);
return split_emit(J,
IRT(IR_HIOP, (LJ_SOFTFP && irt_isnum(ir->t)) ? IRT_SOFTFP : IRT_INT),
tmp, tmp);
}
static void LJ_FASTCALL recff_table_insert(jit_State *J, RecordFFData *rd)
{
RecordIndex ix;
ix.tab = J->base[0];
ix.val = J->base[1];
rd->nres = 0;
if (tref_istab(ix.tab) && ix.val) {
if (!J->base[2]) { /* Simple push: t[#t+1] = v */
TRef trlen = lj_ir_call(J, IRCALL_lj_tab_len, ix.tab);
GCtab *t = tabV(&rd->argv[0]);
ix.key = emitir(IRTI(IR_ADD), trlen, lj_ir_kint(J, 1));
settabV(J->L, &ix.tabv, t);
setintV(&ix.keyv, lj_tab_len(t) + 1);
ix.idxchain = 0;
lj_record_idx(J, &ix); /* Set new value. */
} else { /* Complex case: insert in the middle. */
recff_nyiu(J);
}
} /* else: Interpreter will throw. */
}
static void LJ_FASTCALL recff_io_write(jit_State *J, RecordFFData *rd)
{
TRef fp = recff_io_fp(J, rd->data);
TRef zero = lj_ir_kint(J, 0);
TRef one = lj_ir_kint(J, 1);
ptrdiff_t i = rd->data == 0 ? 1 : 0;
for (; J->base[i]; i++) {
TRef str = lj_ir_tostr(J, J->base[i]);
TRef buf = emitir(IRT(IR_STRREF, IRT_P32), str, zero);
TRef len = emitir(IRTI(IR_FLOAD), str, IRFL_STR_LEN);
if (tref_isk(len) && IR(tref_ref(len))->i == 1) {
TRef tr = emitir(IRT(IR_XLOAD, IRT_U8), buf, IRXLOAD_READONLY);
tr = lj_ir_call(J, IRCALL_fputc, tr, fp);
if (results_wanted(J) != 0) /* Check result only if not ignored. */
emitir(IRTGI(IR_NE), tr, lj_ir_kint(J, -1));
} else {
TRef tr = lj_ir_call(J, IRCALL_fwrite, buf, one, len, fp);
if (results_wanted(J) != 0) /* Check result only if not ignored. */
emitir(IRTGI(IR_EQ), tr, len);
}
}
J->base[0] = TREF_TRUE;
}
static void LJ_FASTCALL recff_string_len(jit_State *J, RecordFFData *rd)
{
J->base[0] = emitir(IRTI(IR_FLOAD), lj_ir_tostr(J, J->base[0]), IRFL_STR_LEN);
UNUSED(rd);
}
/* Handle string.byte (rd->data = 0) and string.sub (rd->data = 1). */
static void LJ_FASTCALL recff_string_range(jit_State *J, RecordFFData *rd)
{
TRef trstr = lj_ir_tostr(J, J->base[0]);
TRef trlen = emitir(IRTI(IR_FLOAD), trstr, IRFL_STR_LEN);
TRef tr0 = lj_ir_kint(J, 0);
TRef trstart, trend;
GCstr *str = argv2str(J, &rd->argv[0]);
int32_t start, end;
if (rd->data) { /* string.sub(str, start [,end]) */
start = argv2int(J, &rd->argv[1]);
trstart = lj_opt_narrow_toint(J, J->base[1]);
trend = J->base[2];
if (tref_isnil(trend)) {
trend = lj_ir_kint(J, -1);
end = -1;
} else {
trend = lj_opt_narrow_toint(J, trend);
end = argv2int(J, &rd->argv[2]);
}
} else { /* string.byte(str, [,start [,end]]) */
if (J->base[1]) {
start = argv2int(J, &rd->argv[1]);
trstart = lj_opt_narrow_toint(J, J->base[1]);
trend = J->base[2];
if (tref_isnil(trend)) {
trend = trstart;
end = start;
} else {
trend = lj_opt_narrow_toint(J, trend);
end = argv2int(J, &rd->argv[2]);
}
} else {
trend = trstart = lj_ir_kint(J, 1);
end = start = 1;
}
}
if (end < 0) {
emitir(IRTGI(IR_LT), trend, tr0);
trend = emitir(IRTI(IR_ADD), emitir(IRTI(IR_ADD), trlen, trend),
lj_ir_kint(J, 1));
end = end+(int32_t)str->len+1;
} else if ((MSize)end <= str->len) {
emitir(IRTGI(IR_ULE), trend, trlen);
} else {
emitir(IRTGI(IR_GT), trend, trlen);
end = (int32_t)str->len;
trend = trlen;
}
if (start < 0) {
emitir(IRTGI(IR_LT), trstart, tr0);
trstart = emitir(IRTI(IR_ADD), trlen, trstart);
start = start+(int32_t)str->len;
emitir(start < 0 ? IRTGI(IR_LT) : IRTGI(IR_GE), trstart, tr0);
if (start < 0) {
trstart = tr0;
start = 0;
}
} else {
if (start == 0) {
emitir(IRTGI(IR_EQ), trstart, tr0);
trstart = tr0;
} else {
trstart = emitir(IRTI(IR_ADD), trstart, lj_ir_kint(J, -1));
emitir(IRTGI(IR_GE), trstart, tr0);
start--;
}
}
if (rd->data) { /* Return string.sub result. */
if (end - start >= 0) {
/* Also handle empty range here, to avoid extra traces. */
TRef trptr, trslen = emitir(IRTI(IR_SUB), trend, trstart);
emitir(IRTGI(IR_GE), trslen, tr0);
trptr = emitir(IRT(IR_STRREF, IRT_P32), trstr, trstart);
J->base[0] = emitir(IRT(IR_SNEW, IRT_STR), trptr, trslen);
} else { /* Range underflow: return empty string. */
emitir(IRTGI(IR_LT), trend, trstart);
J->base[0] = lj_ir_kstr(J, lj_str_new(J->L, strdata(str), 0));
}
} else { /* Return string.byte result(s). */
ptrdiff_t i, len = end - start;
if (len > 0) {
TRef trslen = emitir(IRTI(IR_SUB), trend, trstart);
emitir(IRTGI(IR_EQ), trslen, lj_ir_kint(J, (int32_t)len));
if (J->baseslot + len > LJ_MAX_JSLOTS)
lj_trace_err_info(J, LJ_TRERR_STACKOV);
rd->nres = len;
for (i = 0; i < len; i++) {
TRef tmp = emitir(IRTI(IR_ADD), trstart, lj_ir_kint(J, (int32_t)i));
tmp = emitir(IRT(IR_STRREF, IRT_P32), trstr, tmp);
J->base[i] = emitir(IRT(IR_XLOAD, IRT_U8), tmp, IRXLOAD_READONLY);
}
} else { /* Empty range or range underflow: return no results. */
emitir(IRTGI(IR_LE), trend, trstart);
rd->nres = 0;
}
}
}
/* Record unary bit.tobit, bit.bnot, bit.bswap. */
static void LJ_FASTCALL recff_bit_unary(jit_State *J, RecordFFData *rd)
{
TRef tr = lj_opt_narrow_tobit(J, J->base[0]);
J->base[0] = (rd->data == IR_TOBIT) ? tr : emitir(IRTI(rd->data), tr, 0);
}
/* Backpropagate narrowing conversion. Return number of needed conversions. */
static int narrow_conv_backprop(NarrowConv *nc, IRRef ref, int depth)
{
jit_State *J = nc->J;
IRIns *ir = IR(ref);
IRRef cref;
/* Check the easy cases first. */
if (ir->o == IR_TONUM) { /* Undo inverse conversion. */
*nc->sp++ = NARROWINS(NARROW_REF, ir->op1);
return 0;
} else if (ir->o == IR_KNUM) { /* Narrow FP constant. */
lua_Number n = ir_knum(ir)->n;
if (nc->mode == IRTOINT_TOBIT) { /* Allows a wider range of constants. */
int64_t k64 = (int64_t)n;
if (n == cast_num(k64)) { /* Only if constant doesn't lose precision. */
*nc->sp++ = NARROWINS(NARROW_INT, 0);
*nc->sp++ = (NarrowIns)k64; /* But always truncate to 32 bits. */
return 0;
}
} else {
int32_t k = lj_num2int(n);
if (n == cast_num(k)) { /* Only if constant is really an integer. */
*nc->sp++ = NARROWINS(NARROW_INT, 0);
*nc->sp++ = (NarrowIns)k;
return 0;
}
}
return 10; /* Never narrow other FP constants (this is rare). */
}
/* Try to CSE the conversion. Stronger checks are ok, too. */
for (cref = J->chain[fins->o]; cref > ref; cref = IR(cref)->prev)
if (IR(cref)->op1 == ref &&
irt_isguard(IR(cref)->t) >= irt_isguard(fins->t)) {
*nc->sp++ = NARROWINS(NARROW_REF, cref);
return 0; /* Already there, no additional conversion needed. */
}
/* Backpropagate across ADD/SUB. */
if (ir->o == IR_ADD || ir->o == IR_SUB) {
/* Try cache lookup first. */
IRRef bpref, mode = nc->mode;
if (mode == IRTOINT_INDEX && depth > 0)
mode = IRTOINT_CHECK; /* Inner conversions need a stronger check. */
bpref = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
if (bpref) {
*nc->sp++ = NARROWINS(NARROW_REF, bpref);
return 0;
}
if (++depth < NARROW_MAX_BACKPROP && nc->sp < nc->maxsp) {
NarrowIns *savesp = nc->sp;
int count = narrow_conv_backprop(nc, ir->op1, depth);
count += narrow_conv_backprop(nc, ir->op2, depth);
if (count <= nc->lim) { /* Limit total number of conversions. */
*nc->sp++ = NARROWINS(IRTI(ir->o), ref);
return count;
}
nc->sp = savesp; /* Too many conversions, need to backtrack. */
}
}
/* Otherwise add a conversion. */
*nc->sp++ = NARROWINS(NARROW_CONV, ref);
return 1;
}
/* XLOAD forwarding. */
TRef LJ_FASTCALL lj_opt_fwd_xload(jit_State *J)
{
IRRef xref = fins->op1;
IRIns *xr = IR(xref);
IRRef lim = xref; /* Search limit. */
IRRef ref;
if ((fins->op2 & IRXLOAD_READONLY))
goto cselim;
if ((fins->op2 & IRXLOAD_VOLATILE))
goto doemit;
/* Search for conflicting stores. */
ref = J->chain[IR_XSTORE];
retry:
if (J->chain[IR_CALLXS] > lim) lim = J->chain[IR_CALLXS];
if (J->chain[IR_XBAR] > lim) lim = J->chain[IR_XBAR];
while (ref > lim) {
IRIns *store = IR(ref);
switch (aa_xref(J, xr, fins, store)) {
case ALIAS_NO: break; /* Continue searching. */
case ALIAS_MAY: lim = ref; goto cselim; /* Limit search for load. */
case ALIAS_MUST:
/* Emit conversion if the loaded type doesn't match the forwarded type. */
if (!irt_sametype(fins->t, IR(store->op2)->t)) {
IRType st = irt_type(fins->t);
if (st == IRT_I8 || st == IRT_I16) { /* Trunc + sign-extend. */
st |= IRCONV_SEXT;
} else if (st == IRT_U8 || st == IRT_U16) { /* Trunc + zero-extend. */
} else if (st == IRT_INT) {
st = irt_type(IR(store->op2)->t); /* Needs dummy CONV.int.*. */
} else { /* I64/U64 are boxed, U32 is hidden behind a CONV.num.u32. */
goto store_fwd;
}
fins->ot = IRTI(IR_CONV);
fins->op1 = store->op2;
fins->op2 = (IRT_INT<<5)|st;
return RETRYFOLD;
}
store_fwd:
return store->op2; /* Store forwarding. */
}
ref = store->prev;
}
cselim:
/* Try to find a matching load. Below the conflicting store, if any. */
ref = J->chain[IR_XLOAD];
while (ref > lim) {
/* CSE for XLOAD depends on the type, but not on the IRXLOAD_* flags. */
if (IR(ref)->op1 == xref && irt_sametype(IR(ref)->t, fins->t))
return ref;
ref = IR(ref)->prev;
}
/* Reassociate XLOAD across PHIs to handle a[i-1] forwarding case. */
if (!(fins->op2 & IRXLOAD_READONLY) && J->chain[IR_LOOP] &&
xref == fins->op1 && (xref = reassoc_xref(J, xr)) != 0) {
ref = J->chain[IR_XSTORE];
while (ref > lim) /* Skip stores that have already been checked. */
ref = IR(ref)->prev;
lim = xref;
xr = IR(xref);
goto retry; /* Retry with the reassociated reference. */
}
doemit:
return EMITFOLD;
}
/* Transform the old IR to the new IR. */
static void split_ir(jit_State *J)
{
IRRef nins = J->cur.nins, nk = J->cur.nk;
MSize irlen = nins - nk;
MSize need = (irlen+1)*(sizeof(IRIns) + sizeof(IRRef1));
IRIns *oir = (IRIns *)lj_str_needbuf(J->L, &G(J->L)->tmpbuf, need);
IRRef1 *hisubst;
IRRef ref, snref;
SnapShot *snap;
/* Copy old IR to buffer. */
memcpy(oir, IR(nk), irlen*sizeof(IRIns));
/* Bias hiword substitution table and old IR. Loword kept in field prev. */
hisubst = (IRRef1 *)&oir[irlen] - nk;
oir -= nk;
/* Remove all IR instructions, but retain IR constants. */
J->cur.nins = REF_FIRST;
J->loopref = 0;
/* Process constants and fixed references. */
for (ref = nk; ref <= REF_BASE; ref++) {
IRIns *ir = &oir[ref];
if ((LJ_SOFTFP && ir->o == IR_KNUM) || ir->o == IR_KINT64) {
/* Split up 64 bit constant. */
TValue tv = *ir_k64(ir);
ir->prev = lj_ir_kint(J, (int32_t)tv.u32.lo);
hisubst[ref] = lj_ir_kint(J, (int32_t)tv.u32.hi);
} else {
ir->prev = ref; /* Identity substitution for loword. */
hisubst[ref] = 0;
}
}
/* Process old IR instructions. */
snap = J->cur.snap;
snref = snap->ref;
for (ref = REF_FIRST; ref < nins; ref++) {
IRIns *ir = &oir[ref];
IRRef nref = lj_ir_nextins(J);
IRIns *nir = IR(nref);
IRRef hi = 0;
if (ref >= snref) {
snap->ref = nref;
split_subst_snap(J, snap++, oir);
snref = snap < &J->cur.snap[J->cur.nsnap] ? snap->ref : ~(IRRef)0;
}
/* Copy-substitute old instruction to new instruction. */
nir->op1 = ir->op1 < nk ? ir->op1 : oir[ir->op1].prev;
nir->op2 = ir->op2 < nk ? ir->op2 : oir[ir->op2].prev;
ir->prev = nref; /* Loword substitution. */
nir->o = ir->o;
nir->t.irt = ir->t.irt & ~(IRT_MARK|IRT_ISPHI);
hisubst[ref] = 0;
/* Split 64 bit instructions. */
#if LJ_SOFTFP
if (irt_isnum(ir->t)) {
nir->t.irt = IRT_INT | (nir->t.irt & IRT_GUARD); /* Turn into INT op. */
/* Note: hi ref = lo ref + 1! Required for SNAP_SOFTFPNUM logic. */
switch (ir->o) {
case IR_ADD:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_softfp_add);
break;
case IR_SUB:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_softfp_sub);
break;
case IR_MUL:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_softfp_mul);
break;
case IR_DIV:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_softfp_div);
break;
case IR_POW:
hi = split_call_li(J, hisubst, oir, ir, IRCALL_lj_vm_powi);
break;
case IR_FPMATH:
/* Try to rejoin pow from EXP2, MUL and LOG2. */
if (nir->op2 == IRFPM_EXP2 && nir->op1 > J->loopref) {
IRIns *irp = IR(nir->op1);
if (irp->o == IR_CALLN && irp->op2 == IRCALL_softfp_mul) {
IRIns *irm4 = IR(irp->op1);
IRIns *irm3 = IR(irm4->op1);
IRIns *irm12 = IR(irm3->op1);
IRIns *irl1 = IR(irm12->op1);
if (irm12->op1 > J->loopref && irl1->o == IR_CALLN &&
irl1->op2 == IRCALL_lj_vm_log2) {
IRRef tmp = irl1->op1; /* Recycle first two args from LOG2. */
IRRef arg3 = irm3->op2, arg4 = irm4->op2;
J->cur.nins--;
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, arg3);
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, arg4);
ir->prev = tmp = split_emit(J, IRTI(IR_CALLN), tmp, IRCALL_pow);
hi = split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), tmp, tmp);
break;
}
}
}
hi = split_call_l(J, hisubst, oir, ir, IRCALL_lj_vm_floor + ir->op2);
break;
case IR_ATAN2:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_atan2);
break;
case IR_LDEXP:
hi = split_call_li(J, hisubst, oir, ir, IRCALL_ldexp);
break;
case IR_NEG: case IR_ABS:
nir->o = IR_CONV; /* Pass through loword. */
nir->op2 = (IRT_INT << 5) | IRT_INT;
hi = split_emit(J, IRT(ir->o == IR_NEG ? IR_BXOR : IR_BAND, IRT_SOFTFP),
hisubst[ir->op1], hisubst[ir->op2]);
break;
case IR_SLOAD:
if ((nir->op2 & IRSLOAD_CONVERT)) { /* Convert from int to number. */
nir->op2 &= ~IRSLOAD_CONVERT;
ir->prev = nref = split_emit(J, IRTI(IR_CALLN), nref,
IRCALL_softfp_i2d);
hi = split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), nref, nref);
break;
}
/* fallthrough */
case IR_ALOAD: case IR_HLOAD: case IR_ULOAD: case IR_VLOAD:
case IR_STRTO:
hi = split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), nref, nref);
break;
case IR_XLOAD: {
IRIns inslo = *nir; /* Save/undo the emit of the lo XLOAD. */
J->cur.nins--;
hi = split_ptr(J, oir, ir->op1); /* Insert the hiref ADD. */
nref = lj_ir_nextins(J);
nir = IR(nref);
*nir = inslo; /* Re-emit lo XLOAD immediately before hi XLOAD. */
hi = split_emit(J, IRT(IR_XLOAD, IRT_SOFTFP), hi, ir->op2);
#if LJ_LE
ir->prev = nref;
#else
ir->prev = hi; hi = nref;
#endif
break;
}
case IR_ASTORE: case IR_HSTORE: case IR_USTORE: case IR_XSTORE:
split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), nir->op1, hisubst[ir->op2]);
break;
case IR_CONV: { /* Conversion to number. Others handled below. */
IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
UNUSED(st);
#if LJ_32 && LJ_HASFFI
if (st == IRT_I64 || st == IRT_U64) {
hi = split_call_l(J, hisubst, oir, ir,
st == IRT_I64 ? IRCALL_fp64_l2d : IRCALL_fp64_ul2d);
break;
}
#endif
lua_assert(st == IRT_INT ||
(LJ_32 && LJ_HASFFI && (st == IRT_U32 || st == IRT_FLOAT)));
nir->o = IR_CALLN;
#if LJ_32 && LJ_HASFFI
nir->op2 = st == IRT_INT ? IRCALL_softfp_i2d :
st == IRT_FLOAT ? IRCALL_softfp_f2d :
IRCALL_softfp_ui2d;
#else
nir->op2 = IRCALL_softfp_i2d;
#endif
hi = split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), nref, nref);
break;
}
case IR_CALLN:
case IR_CALLL:
case IR_CALLS:
case IR_CALLXS:
goto split_call;
case IR_PHI:
if (nir->op1 == nir->op2)
J->cur.nins--; /* Drop useless PHIs. */
if (hisubst[ir->op1] != hisubst[ir->op2])
split_emit(J, IRT(IR_PHI, IRT_SOFTFP),
hisubst[ir->op1], hisubst[ir->op2]);
break;
case IR_HIOP:
J->cur.nins--; /* Drop joining HIOP. */
ir->prev = nir->op1;
hi = nir->op2;
break;
default:
lua_assert(ir->o <= IR_NE || ir->o == IR_MIN || ir->o == IR_MAX);
hi = split_emit(J, IRTG(IR_HIOP, IRT_SOFTFP),
hisubst[ir->op1], hisubst[ir->op2]);
break;
}
} else
#endif
#if LJ_32 && LJ_HASFFI
if (irt_isint64(ir->t)) {
IRRef hiref = hisubst[ir->op1];
nir->t.irt = IRT_INT | (nir->t.irt & IRT_GUARD); /* Turn into INT op. */
switch (ir->o) {
case IR_ADD:
case IR_SUB:
/* Use plain op for hiword if loword cannot produce a carry/borrow. */
if (irref_isk(nir->op2) && IR(nir->op2)->i == 0) {
ir->prev = nir->op1; /* Pass through loword. */
nir->op1 = hiref; nir->op2 = hisubst[ir->op2];
hi = nref;
break;
}
/* fallthrough */
case IR_NEG:
hi = split_emit(J, IRTI(IR_HIOP), hiref, hisubst[ir->op2]);
break;
case IR_MUL:
hi = split_call_ll(J, hisubst, oir, ir, IRCALL_lj_carith_mul64);
break;
case IR_DIV:
hi = split_call_ll(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
IRCALL_lj_carith_divu64);
break;
case IR_MOD:
hi = split_call_ll(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
IRCALL_lj_carith_modu64);
break;
case IR_POW:
hi = split_call_ll(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
IRCALL_lj_carith_powu64);
break;
case IR_FLOAD:
lua_assert(ir->op2 == IRFL_CDATA_INT64);
hi = split_emit(J, IRTI(IR_FLOAD), nir->op1, IRFL_CDATA_INT64_4);
#if LJ_BE
ir->prev = hi; hi = nref;
#endif
break;
case IR_XLOAD:
hi = split_emit(J, IRTI(IR_XLOAD), split_ptr(J, oir, ir->op1), ir->op2);
#if LJ_BE
ir->prev = hi; hi = nref;
#endif
break;
case IR_XSTORE:
split_emit(J, IRTI(IR_HIOP), nir->op1, hisubst[ir->op2]);
break;
case IR_CONV: { /* Conversion to 64 bit integer. Others handled below. */
IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
#if LJ_SOFTFP
if (st == IRT_NUM) { /* NUM to 64 bit int conv. */
hi = split_call_l(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_fp64_d2l : IRCALL_fp64_d2ul);
} else if (st == IRT_FLOAT) { /* FLOAT to 64 bit int conv. */
nir->o = IR_CALLN;
nir->op2 = irt_isi64(ir->t) ? IRCALL_fp64_f2l : IRCALL_fp64_f2ul;
hi = split_emit(J, IRTI(IR_HIOP), nref, nref);
}
#else
if (st == IRT_NUM || st == IRT_FLOAT) { /* FP to 64 bit int conv. */
hi = split_emit(J, IRTI(IR_HIOP), nir->op1, nref);
}
#endif
else if (st == IRT_I64 || st == IRT_U64) { /* 64/64 bit cast. */
/* Drop cast, since assembler doesn't care. But fwd both parts. */
hi = hiref;
goto fwdlo;
} else if ((ir->op2 & IRCONV_SEXT)) { /* Sign-extend to 64 bit. */
IRRef k31 = lj_ir_kint(J, 31);
nir = IR(nref); /* May have been reallocated. */
ir->prev = nir->op1; /* Pass through loword. */
nir->o = IR_BSAR; /* hi = bsar(lo, 31). */
nir->op2 = k31;
hi = nref;
} else { /* Zero-extend to 64 bit. */
hi = lj_ir_kint(J, 0);
goto fwdlo;
}
break;
}
case IR_CALLXS:
goto split_call;
case IR_PHI: {
IRRef hiref2;
if ((irref_isk(nir->op1) && irref_isk(nir->op2)) ||
nir->op1 == nir->op2)
J->cur.nins--; /* Drop useless PHIs. */
hiref2 = hisubst[ir->op2];
if (!((irref_isk(hiref) && irref_isk(hiref2)) || hiref == hiref2))
split_emit(J, IRTI(IR_PHI), hiref, hiref2);
break;
}
case IR_HIOP:
J->cur.nins--; /* Drop joining HIOP. */
ir->prev = nir->op1;
hi = nir->op2;
break;
default:
lua_assert(ir->o <= IR_NE); /* Comparisons. */
split_emit(J, IRTGI(IR_HIOP), hiref, hisubst[ir->op2]);
break;
}
} else
#endif
#if LJ_SOFTFP
if (ir->o == IR_SLOAD) {
if ((nir->op2 & IRSLOAD_CONVERT)) { /* Convert from number to int. */
nir->op2 &= ~IRSLOAD_CONVERT;
if (!(nir->op2 & IRSLOAD_TYPECHECK))
nir->t.irt = IRT_INT; /* Drop guard. */
split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), nref, nref);
ir->prev = split_num2int(J, nref, nref+1, irt_isguard(ir->t));
}
} else if (ir->o == IR_TOBIT) {
IRRef tmp, op1 = ir->op1;
J->cur.nins--;
#if LJ_LE
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), oir[op1].prev, hisubst[op1]);
#else
tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), hisubst[op1], oir[op1].prev);
#endif
ir->prev = split_emit(J, IRTI(IR_CALLN), tmp, IRCALL_lj_vm_tobit);
} else if (ir->o == IR_TOSTR) {
if (hisubst[ir->op1]) {
if (irref_isk(ir->op1))
nir->op1 = ir->op1;
else
split_emit(J, IRT(IR_HIOP, IRT_NIL), hisubst[ir->op1], nref);
}
} else if (ir->o == IR_HREF || ir->o == IR_NEWREF) {
if (irref_isk(ir->op2) && hisubst[ir->op2])
nir->op2 = ir->op2;
} else
#endif
if (ir->o == IR_CONV) { /* See above, too. */
IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
#if LJ_32 && LJ_HASFFI
if (st == IRT_I64 || st == IRT_U64) { /* Conversion from 64 bit int. */
#if LJ_SOFTFP
if (irt_isfloat(ir->t)) {
split_call_l(J, hisubst, oir, ir,
st == IRT_I64 ? IRCALL_fp64_l2f : IRCALL_fp64_ul2f);
J->cur.nins--; /* Drop unused HIOP. */
}
#else
if (irt_isfp(ir->t)) { /* 64 bit integer to FP conversion. */
ir->prev = split_emit(J, IRT(IR_HIOP, irt_type(ir->t)),
hisubst[ir->op1], nref);
}
#endif
else { /* Truncate to lower 32 bits. */
fwdlo:
ir->prev = nir->op1; /* Forward loword. */
/* Replace with NOP to avoid messing up the snapshot logic. */
nir->ot = IRT(IR_NOP, IRT_NIL);
nir->op1 = nir->op2 = 0;
}
}
#endif
#if LJ_SOFTFP && LJ_32 && LJ_HASFFI
else if (irt_isfloat(ir->t)) {
if (st == IRT_NUM) {
split_call_l(J, hisubst, oir, ir, IRCALL_softfp_d2f);
J->cur.nins--; /* Drop unused HIOP. */
} else {
nir->o = IR_CALLN;
nir->op2 = st == IRT_INT ? IRCALL_softfp_i2f : IRCALL_softfp_ui2f;
}
} else if (st == IRT_FLOAT) {
nir->o = IR_CALLN;
nir->op2 = irt_isint(ir->t) ? IRCALL_softfp_f2i : IRCALL_softfp_f2ui;
} else
#endif
#if LJ_SOFTFP
if (st == IRT_NUM || (LJ_32 && LJ_HASFFI && st == IRT_FLOAT)) {
if (irt_isguard(ir->t)) {
lua_assert(st == IRT_NUM && irt_isint(ir->t));
J->cur.nins--;
ir->prev = split_num2int(J, nir->op1, hisubst[ir->op1], 1);
} else {
split_call_l(J, hisubst, oir, ir,
#if LJ_32 && LJ_HASFFI
st == IRT_NUM ?
(irt_isint(ir->t) ? IRCALL_softfp_d2i : IRCALL_softfp_d2ui) :
(irt_isint(ir->t) ? IRCALL_softfp_f2i : IRCALL_softfp_f2ui)
#else
IRCALL_softfp_d2i
#endif
);
J->cur.nins--; /* Drop unused HIOP. */
}
}
#endif
} else if (ir->o == IR_CALLXS) {
IRRef hiref;
split_call:
hiref = hisubst[ir->op1];
if (hiref) {
IROpT ot = nir->ot;
IRRef op2 = nir->op2;
nir->ot = IRT(IR_CARG, IRT_NIL);
#if LJ_LE
nir->op2 = hiref;
#else
nir->op2 = nir->op1; nir->op1 = hiref;
#endif
ir->prev = nref = split_emit(J, ot, nref, op2);
}
if (LJ_SOFTFP ? irt_is64(ir->t) : irt_isint64(ir->t))
hi = split_emit(J,
IRT(IR_HIOP, (LJ_SOFTFP && irt_isnum(ir->t)) ? IRT_SOFTFP : IRT_INT),
nref, nref);
} else if (ir->o == IR_CARG) {
IRRef hiref = hisubst[ir->op1];
if (hiref) {
IRRef op2 = nir->op2;
#if LJ_LE
nir->op2 = hiref;
#else
nir->op2 = nir->op1; nir->op1 = hiref;
#endif
ir->prev = nref = split_emit(J, IRT(IR_CARG, IRT_NIL), nref, op2);
nir = IR(nref);
}
hiref = hisubst[ir->op2];
if (hiref) {
#if !LJ_TARGET_X86
int carg = 0;
IRIns *cir;
for (cir = IR(nir->op1); cir->o == IR_CARG; cir = IR(cir->op1))
carg++;
if ((carg & 1) == 0) { /* Align 64 bit arguments. */
IRRef op2 = nir->op2;
nir->op2 = REF_NIL;
nref = split_emit(J, IRT(IR_CARG, IRT_NIL), nref, op2);
nir = IR(nref);
}
#endif
#if LJ_BE
{ IRRef tmp = nir->op2; nir->op2 = hiref; hiref = tmp; }
#endif
ir->prev = split_emit(J, IRT(IR_CARG, IRT_NIL), nref, hiref);
}
} else if (ir->o == IR_CNEWI) {
if (hisubst[ir->op2])
split_emit(J, IRT(IR_HIOP, IRT_NIL), nref, hisubst[ir->op2]);
} else if (ir->o == IR_LOOP) {
J->loopref = nref; /* Needed by assembler. */
}
hisubst[ref] = hi; /* Store hiword substitution. */
}
if (snref == nins) { /* Substitution for last snapshot. */
snap->ref = J->cur.nins;
split_subst_snap(J, snap, oir);
}
/* Add PHI marks. */
for (ref = J->cur.nins-1; ref >= REF_FIRST; ref--) {
IRIns *ir = IR(ref);
if (ir->o != IR_PHI) break;
if (!irref_isk(ir->op1)) irt_setphi(IR(ir->op1)->t);
if (ir->op2 > J->loopref) irt_setphi(IR(ir->op2)->t);
}
}
/* Transform the old IR to the new IR. */
static void split_ir(jit_State *J)
{
IRRef nins = J->cur.nins, nk = J->cur.nk;
MSize irlen = nins - nk;
MSize need = (irlen+1)*(sizeof(IRIns) + sizeof(IRRef1));
IRIns *oir = (IRIns *)lj_str_needbuf(J->L, &G(J->L)->tmpbuf, need);
IRRef1 *hisubst;
IRRef ref;
/* Copy old IR to buffer. */
memcpy(oir, IR(nk), irlen*sizeof(IRIns));
/* Bias hiword substitution table and old IR. Loword kept in field prev. */
hisubst = (IRRef1 *)&oir[irlen] - nk;
oir -= nk;
/* Remove all IR instructions, but retain IR constants. */
J->cur.nins = REF_FIRST;
/* Process constants and fixed references. */
for (ref = nk; ref <= REF_BASE; ref++) {
IRIns *ir = &oir[ref];
if (ir->o == IR_KINT64) { /* Split up 64 bit constant. */
TValue tv = *ir_k64(ir);
ir->prev = lj_ir_kint(J, (int32_t)tv.u32.lo);
hisubst[ref] = lj_ir_kint(J, (int32_t)tv.u32.hi);
} else {
ir->prev = ref; /* Identity substitution for loword. */
hisubst[ref] = 0;
}
}
/* Process old IR instructions. */
for (ref = REF_FIRST; ref < nins; ref++) {
IRIns *ir = &oir[ref];
IRRef nref = lj_ir_nextins(J);
IRIns *nir = IR(nref);
IRRef hi = 0;
/* Copy-substitute old instruction to new instruction. */
nir->op1 = ir->op1 < nk ? ir->op1 : oir[ir->op1].prev;
nir->op2 = ir->op2 < nk ? ir->op2 : oir[ir->op2].prev;
ir->prev = nref; /* Loword substitution. */
nir->o = ir->o;
nir->t.irt = ir->t.irt & ~(IRT_MARK|IRT_ISPHI);
hisubst[ref] = 0;
/* Split 64 bit instructions. */
if (irt_isint64(ir->t)) {
IRRef hiref = hisubst[ir->op1];
nir->t.irt = IRT_INT | (nir->t.irt & IRT_GUARD); /* Turn into INT op. */
switch (ir->o) {
case IR_ADD:
case IR_SUB:
/* Use plain op for hiword if loword cannot produce a carry/borrow. */
if (irref_isk(nir->op2) && IR(nir->op2)->i == 0) {
ir->prev = nir->op1; /* Pass through loword. */
nir->op1 = hiref; nir->op2 = hisubst[ir->op2];
hi = nref;
break;
}
/* fallthrough */
case IR_NEG:
hi = split_emit(J, IRTI(IR_HIOP), hiref, hisubst[ir->op2]);
break;
case IR_MUL:
hi = split_call64(J, hisubst, oir, ir, IRCALL_lj_carith_mul64);
break;
case IR_DIV:
hi = split_call64(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
IRCALL_lj_carith_divu64);
break;
case IR_MOD:
hi = split_call64(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
IRCALL_lj_carith_modu64);
break;
case IR_POW:
hi = split_call64(J, hisubst, oir, ir,
irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
IRCALL_lj_carith_powu64);
break;
case IR_FLOAD:
lua_assert(ir->op2 == IRFL_CDATA_INT64);
hi = split_emit(J, IRTI(IR_FLOAD), nir->op1, IRFL_CDATA_INT64HI);
#if LJ_BE
ir->prev = hi; hi = nref;
#endif
break;
case IR_XLOAD:
hi = split_emit(J, IRTI(IR_XLOAD), split_ptr(J, nir->op1), ir->op2);
#if LJ_BE
ir->prev = hi; hi = nref;
#endif
break;
case IR_XSTORE:
#if LJ_LE
hiref = hisubst[ir->op2];
#else
hiref = nir->op2; nir->op2 = hisubst[ir->op2];
#endif
split_emit(J, IRTI(IR_XSTORE), split_ptr(J, nir->op1), hiref);
break;
case IR_CONV: { /* Conversion to 64 bit integer. Others handled below. */
IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
if (st == IRT_NUM || st == IRT_FLOAT) { /* FP to 64 bit int conv. */
hi = split_emit(J, IRTI(IR_HIOP), nir->op1, nref);
} else if (st == IRT_I64 || st == IRT_U64) { /* 64/64 bit cast. */
/* Drop cast, since assembler doesn't care. */
goto fwdlo;
} else if ((ir->op2 & IRCONV_SEXT)) { /* Sign-extend to 64 bit. */
IRRef k31 = lj_ir_kint(J, 31);
nir = IR(nref); /* May have been reallocated. */
ir->prev = nir->op1; /* Pass through loword. */
nir->o = IR_BSAR; /* hi = bsar(lo, 31). */
nir->op2 = k31;
hi = nref;
} else { /* Zero-extend to 64 bit. */
hi = lj_ir_kint(J, 0);
goto fwdlo;
}
break;
}
case IR_PHI: {
IRRef hiref2;
if ((irref_isk(nir->op1) && irref_isk(nir->op2)) ||
nir->op1 == nir->op2)
J->cur.nins--; /* Drop useless PHIs. */
hiref2 = hisubst[ir->op2];
if (!((irref_isk(hiref) && irref_isk(hiref2)) || hiref == hiref2))
split_emit(J, IRTI(IR_PHI), hiref, hiref2);
break;
}
default:
lua_assert(ir->o <= IR_NE); /* Comparisons. */
split_emit(J, IRTGI(IR_HIOP), hiref, hisubst[ir->op2]);
break;
}
} else if (ir->o == IR_CONV) { /* See above, too. */
IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
if (st == IRT_I64 || st == IRT_U64) { /* Conversion from 64 bit int. */
if (irt_isfp(ir->t)) { /* 64 bit integer to FP conversion. */
ir->prev = split_emit(J, IRT(IR_HIOP, irt_type(ir->t)),
hisubst[ir->op1], nref);
} else { /* Truncate to lower 32 bits. */
fwdlo:
ir->prev = nir->op1; /* Forward loword. */
/* Replace with NOP to avoid messing up the snapshot logic. */
nir->ot = IRT(IR_NOP, IRT_NIL);
nir->op1 = nir->op2 = 0;
}
}
} else if (ir->o == IR_CNEWI) {
if (hisubst[ir->op2])
split_emit(J, IRT(IR_HIOP, IRT_NIL), nref, hisubst[ir->op2]);
} else if (ir->o == IR_LOOP) {
J->loopref = nref; /* Needed by assembler. */
}
hisubst[ref] = hi; /* Store hiword substitution. */
}
/* Add PHI marks. */
for (ref = J->cur.nins-1; ref >= REF_FIRST; ref--) {
IRIns *ir = IR(ref);
if (ir->o != IR_PHI) break;
if (!irref_isk(ir->op1)) irt_setphi(IR(ir->op1)->t);
if (ir->op2 > J->loopref) irt_setphi(IR(ir->op2)->t);
}
/* Substitute snapshot maps. */
oir[nins].prev = J->cur.nins; /* Substitution for last snapshot. */
{
SnapNo i, nsnap = J->cur.nsnap;
for (i = 0; i < nsnap; i++) {
SnapShot *snap = &J->cur.snap[i];
SnapEntry *map = &J->cur.snapmap[snap->mapofs];
MSize n, nent = snap->nent;
snap->ref = oir[snap->ref].prev;
for (n = 0; n < nent; n++) {
SnapEntry sn = map[n];
map[n] = ((sn & 0xffff0000) | oir[snap_ref(sn)].prev);
}
}
}
}
