/* Array and hash load forwarding. */
static TRef fwd_ahload(jit_State *J, IRRef xref)
{
IRIns *xr = IR(xref);
IRRef lim = xref; /* Search limit. */
IRRef ref;
/* Search for conflicting stores. */
ref = J->chain[fins->o+IRDELTA_L2S];
while (ref > xref) {
IRIns *store = IR(ref);
switch (aa_ahref(J, xr, IR(store->op1))) {
case ALIAS_NO: break; /* Continue searching. */
case ALIAS_MAY: lim = ref; goto cselim; /* Limit search for load. */
case ALIAS_MUST: return store->op2; /* Store forwarding. */
}
ref = store->prev;
}
/* No conflicting store (yet): const-fold loads from allocations. */
{
IRIns *ir = (xr->o == IR_HREFK || xr->o == IR_AREF) ? IR(xr->op1) : xr;
IRRef tab = ir->op1;
ir = IR(tab);
if (ir->o == IR_TNEW || (ir->o == IR_TDUP && irref_isk(xr->op2))) {
/* A NEWREF with a number key may end up pointing to the array part.
** But it's referenced from HSTORE and not found in the ASTORE chain.
** For now simply consider this a conflict without forwarding anything.
*/
if (xr->o == IR_AREF) {
IRRef ref2 = J->chain[IR_NEWREF];
while (ref2 > tab) {
IRIns *newref = IR(ref2);
if (irt_isnum(IR(newref->op2)->t))
goto cselim;
ref2 = newref->prev;
}
}
/* NEWREF inhibits CSE for HREF, and dependent FLOADs from HREFK/AREF.
** But the above search for conflicting stores was limited by xref.
** So continue searching, limited by the TNEW/TDUP. Store forwarding
** is ok, too. A conflict does NOT limit the search for a matching load.
*/
while (ref > tab) {
IRIns *store = IR(ref);
switch (aa_ahref(J, xr, IR(store->op1))) {
case ALIAS_NO: break; /* Continue searching. */
case ALIAS_MAY: goto cselim; /* Conflicting store. */
case ALIAS_MUST: return store->op2; /* Store forwarding. */
}
ref = store->prev;
}
lua_assert(ir->o != IR_TNEW || irt_isnil(fins->t));
if (irt_ispri(fins->t)) {
return TREF_PRI(irt_type(fins->t));
} else if (irt_isnum(fins->t) || irt_isstr(fins->t)) {
TValue keyv;
cTValue *tv;
IRIns *key = IR(xr->op2);
if (key->o == IR_KSLOT) key = IR(key->op1);
lj_ir_kvalue(J->L, &keyv, key);
tv = lj_tab_get(J->L, ir_ktab(IR(ir->op1)), &keyv);
lua_assert(itype2irt(tv) == irt_type(fins->t));
if (irt_isnum(fins->t))
return lj_ir_knum_u64(J, tv->u64);
else
return lj_ir_kstr(J, strV(tv));
}
/* Othwerwise: don't intern as a constant. */
}
}
cselim:
/* Try to find a matching load. Below the conflicting store, if any. */
ref = J->chain[fins->o];
while (ref > lim) {
IRIns *load = IR(ref);
if (load->op1 == xref)
return ref; /* Load forwarding. */
ref = load->prev;
}
return 0; /* Conflict or no match. */
}
/* 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. */
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);
}
}
/* Replay snapshot state to setup side trace. */
void lj_snap_replay(jit_State *J, GCtrace *T)
{
SnapShot *snap = &T->snap[J->exitno];
SnapEntry *map = &T->snapmap[snap->mapofs];
MSize n, nent = snap->nent;
BloomFilter seen = 0;
int pass23 = 0;
J->framedepth = 0;
/* Emit IR for slots inherited from parent snapshot. */
for (n = 0; n < nent; n++) {
SnapEntry sn = map[n];
BCReg s = snap_slot(sn);
IRRef ref = snap_ref(sn);
IRIns *ir = &T->ir[ref];
TRef tr;
/* The bloom filter avoids O(nent^2) overhead for de-duping slots. */
if (bloomtest(seen, ref) && (tr = snap_dedup(J, map, n, ref)) != 0)
goto setslot;
bloomset(seen, ref);
if (irref_isk(ref)) {
tr = snap_replay_const(J, ir);
} else if (!regsp_used(ir->prev)) {
pass23 = 1;
lua_assert(s != 0);
tr = s;
} else {
IRType t = irt_type(ir->t);
uint32_t mode = IRSLOAD_INHERIT|IRSLOAD_PARENT;
if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM)) t = IRT_NUM;
if (ir->o == IR_SLOAD) mode |= (ir->op2 & IRSLOAD_READONLY);
tr = emitir_raw(IRT(IR_SLOAD, t), s, mode);
}
setslot:
J->slot[s] = tr | (sn&(SNAP_CONT|SNAP_FRAME)); /* Same as TREF_* flags. */
J->framedepth += ((sn & (SNAP_CONT|SNAP_FRAME)) && s);
if ((sn & SNAP_FRAME))
J->baseslot = s+1;
}
if (pass23) {
IRIns *irlast = &T->ir[snap->ref];
pass23 = 0;
/* Emit dependent PVALs. */
for (n = 0; n < nent; n++) {
SnapEntry sn = map[n];
IRRef refp = snap_ref(sn);
IRIns *ir = &T->ir[refp];
if (regsp_reg(ir->r) == RID_SUNK) {
if (J->slot[snap_slot(sn)] != snap_slot(sn)) continue;
pass23 = 1;
lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP ||
ir->o == IR_CNEW || ir->o == IR_CNEWI);
if (ir->op1 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op1);
if (ir->op2 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op2);
if (LJ_HASFFI && ir->o == IR_CNEWI) {
if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP)
snap_pref(J, T, map, nent, seen, (ir+1)->op2);
} else {
IRIns *irs;
for (irs = ir+1; irs < irlast; irs++)
if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
if (snap_pref(J, T, map, nent, seen, irs->op2) == 0)
snap_pref(J, T, map, nent, seen, T->ir[irs->op2].op1);
else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
irs+1 < irlast && (irs+1)->o == IR_HIOP)
snap_pref(J, T, map, nent, seen, (irs+1)->op2);
}
}
} else if (!irref_isk(refp) && !regsp_used(ir->prev)) {
lua_assert(ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
J->slot[snap_slot(sn)] = snap_pref(J, T, map, nent, seen, ir->op1);
}
}
/* Replay sunk instructions. */
for (n = 0; pass23 && n < nent; n++) {
SnapEntry sn = map[n];
IRRef refp = snap_ref(sn);
IRIns *ir = &T->ir[refp];
if (regsp_reg(ir->r) == RID_SUNK) {
TRef op1, op2;
if (J->slot[snap_slot(sn)] != snap_slot(sn)) { /* De-dup allocs. */
J->slot[snap_slot(sn)] = J->slot[J->slot[snap_slot(sn)]];
continue;
}
op1 = ir->op1;
if (op1 >= T->nk) op1 = snap_pref(J, T, map, nent, seen, op1);
op2 = ir->op2;
if (op2 >= T->nk) op2 = snap_pref(J, T, map, nent, seen, op2);
if (LJ_HASFFI && ir->o == IR_CNEWI) {
if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP) {
lj_needsplit(J); /* Emit joining HIOP. */
op2 = emitir_raw(IRT(IR_HIOP, IRT_I64), op2,
snap_pref(J, T, map, nent, seen, (ir+1)->op2));
}
J->slot[snap_slot(sn)] = emitir(ir->ot, op1, op2);
} else {
IRIns *irs;
TRef tr = emitir(ir->ot, op1, op2);
J->slot[snap_slot(sn)] = tr;
for (irs = ir+1; irs < irlast; irs++)
if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
IRIns *irr = &T->ir[irs->op1];
TRef val, key = irr->op2, tmp = tr;
if (irr->o != IR_FREF) {
IRIns *irk = &T->ir[key];
if (irr->o == IR_HREFK)
key = lj_ir_kslot(J, snap_replay_const(J, &T->ir[irk->op1]),
irk->op2);
else
key = snap_replay_const(J, irk);
if (irr->o == IR_HREFK || irr->o == IR_AREF) {
IRIns *irf = &T->ir[irr->op1];
tmp = emitir(irf->ot, tmp, irf->op2);
}
}
tmp = emitir(irr->ot, tmp, key);
val = snap_pref(J, T, map, nent, seen, irs->op2);
if (val == 0) {
IRIns *irc = &T->ir[irs->op2];
lua_assert(irc->o == IR_CONV && irc->op2 == IRCONV_NUM_INT);
val = snap_pref(J, T, map, nent, seen, irc->op1);
val = emitir(IRTN(IR_CONV), val, IRCONV_NUM_INT);
} else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
irs+1 < irlast && (irs+1)->o == IR_HIOP) {
IRType t = IRT_I64;
if (LJ_SOFTFP && irt_type((irs+1)->t) == IRT_SOFTFP)
t = IRT_NUM;
lj_needsplit(J);
if (irref_isk(irs->op2) && irref_isk((irs+1)->op2)) {
uint64_t k = (uint32_t)T->ir[irs->op2].i +
((uint64_t)T->ir[(irs+1)->op2].i << 32);
val = lj_ir_k64(J, t == IRT_I64 ? IR_KINT64 : IR_KNUM,
lj_ir_k64_find(J, k));
} else {
val = emitir_raw(IRT(IR_HIOP, t), val,
snap_pref(J, T, map, nent, seen, (irs+1)->op2));
}
tmp = emitir(IRT(irs->o, t), tmp, val);
continue;
}
tmp = emitir(irs->ot, tmp, val);
} else if (LJ_HASFFI && irs->o == IR_XBAR && ir->o == IR_CNEW) {
emitir(IRT(IR_XBAR, IRT_NIL), 0, 0);
}
}
}
}
}
J->base = J->slot + J->baseslot;
J->maxslot = snap->nslots - J->baseslot;
lj_snap_add(J);
if (pass23) /* Need explicit GC step _after_ initial snapshot. */
emitir_raw(IRTG(IR_GCSTEP, IRT_NIL), 0, 0);
}
/* Emit or eliminate collected PHIs. */
static void loop_emit_phi(jit_State *J, IRRef1 *subst, IRRef1 *phi, IRRef nphi,
SnapNo onsnap)
{
int passx = 0;
IRRef i, j, nslots;
IRRef invar = J->chain[IR_LOOP];
/* Pass #1: mark redundant and potentially redundant PHIs. */
for (i = 0, j = 0; i < nphi; i++) {
IRRef lref = phi[i];
IRRef rref = subst[lref];
if (lref == rref || rref == REF_DROP) { /* Invariants are redundant. */
irt_clearphi(IR(lref)->t);
} else {
phi[j++] = (IRRef1)lref;
if (!(IR(rref)->op1 == lref || IR(rref)->op2 == lref)) {
/* Quick check for simple recurrences failed, need pass2. */
irt_setmark(IR(lref)->t);
passx = 1;
}
}
}
nphi = j;
/* Pass #2: traverse variant part and clear marks of non-redundant PHIs. */
if (passx) {
SnapNo s;
for (i = J->cur.nins-1; i > invar; i--) {
IRIns *ir = IR(i);
if (!irref_isk(ir->op2)) irt_clearmark(IR(ir->op2)->t);
if (!irref_isk(ir->op1)) {
irt_clearmark(IR(ir->op1)->t);
if (ir->op1 < invar &&
ir->o >= IR_CALLN && ir->o <= IR_CARG) { /* ORDER IR */
ir = IR(ir->op1);
while (ir->o == IR_CARG) {
if (!irref_isk(ir->op2)) irt_clearmark(IR(ir->op2)->t);
if (irref_isk(ir->op1)) break;
ir = IR(ir->op1);
irt_clearmark(ir->t);
}
}
}
}
for (s = J->cur.nsnap-1; s >= onsnap; s--) {
SnapShot *snap = &J->cur.snap[s];
SnapEntry *map = &J->cur.snapmap[snap->mapofs];
MSize n, nent = snap->nent;
for (n = 0; n < nent; n++) {
IRRef ref = snap_ref(map[n]);
if (!irref_isk(ref)) irt_clearmark(IR(ref)->t);
}
}
}
/* Pass #3: add PHIs for variant slots without a corresponding SLOAD. */
nslots = J->baseslot+J->maxslot;
for (i = 1; i < nslots; i++) {
IRRef ref = tref_ref(J->slot[i]);
while (!irref_isk(ref) && ref != subst[ref]) {
IRIns *ir = IR(ref);
irt_clearmark(ir->t); /* Unmark potential uses, too. */
if (irt_isphi(ir->t) || irt_ispri(ir->t))
break;
irt_setphi(ir->t);
if (nphi >= LJ_MAX_PHI)
lj_trace_err(J, LJ_TRERR_PHIOV);
phi[nphi++] = (IRRef1)ref;
ref = subst[ref];
if (ref > invar)
break;
}
}
/* Pass #4: propagate non-redundant PHIs. */
while (passx) {
passx = 0;
for (i = 0; i < nphi; i++) {
IRRef lref = phi[i];
IRIns *ir = IR(lref);
if (!irt_ismarked(ir->t)) { /* Propagate only from unmarked PHIs. */
IRIns *irr = IR(subst[lref]);
if (irt_ismarked(irr->t)) { /* Right ref points to other PHI? */
irt_clearmark(irr->t); /* Mark that PHI as non-redundant. */
passx = 1; /* Retry. */
}
}
}
}
/* Pass #5: emit PHI instructions or eliminate PHIs. */
for (i = 0; i < nphi; i++) {
IRRef lref = phi[i];
IRIns *ir = IR(lref);
if (!irt_ismarked(ir->t)) { /* Emit PHI if not marked. */
IRRef rref = subst[lref];
if (rref > invar)
irt_setphi(IR(rref)->t);
emitir_raw(IRT(IR_PHI, irt_type(ir->t)), lref, rref);
} else { /* Otherwise eliminate PHI. */
irt_clearmark(ir->t);
irt_clearphi(ir->t);
}
}
}
/* 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 && !irt_isint(IR(store->op2)->t)) {
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;
/* 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);
}
}
}
}