/* USTORE elimination. */
TRef LJ_FASTCALL lj_opt_dse_ustore(jit_State *J)
{
IRRef xref = fins->op1; /* xREF reference. */
IRRef val = fins->op2; /* Stored value reference. */
IRIns *xr = IR(xref);
IRRef1 *refp = &J->chain[IR_USTORE];
IRRef ref = *refp;
while (ref > xref) { /* Search for redundant or conflicting stores. */
IRIns *store = IR(ref);
switch (aa_uref(xr, IR(store->op1))) {
case ALIAS_NO:
break; /* Continue searching. */
case ALIAS_MAY: /* Store to MAYBE the same location. */
if (store->op2 != val) /* Conflict if the value is different. */
goto doemit;
break; /* Otherwise continue searching. */
case ALIAS_MUST: /* Store to the same location. */
if (store->op2 == val) /* Same value: drop the new store. */
return DROPFOLD;
/* Different value: try to eliminate the redundant store. */
if (ref > J->chain[IR_LOOP]) { /* Quick check to avoid crossing LOOP. */
IRIns *ir;
/* Check for any intervening guards (includes conflicting loads). */
for (ir = IR(J->cur.nins-1); ir > store; ir--)
if (irt_isguard(ir->t))
goto doemit; /* No elimination possible. */
/* Remove redundant store from chain and replace with NOP. */
*refp = store->prev;
store->o = IR_NOP; /* Unchained NOP -- does anybody care? */
store->t.irt = IRT_NIL;
store->op1 = store->op2 = 0;
store->prev = 0;
/* Now emit the new store instead. */
}
goto doemit;
}
ref = *(refp = &store->prev);
}
doemit:
return EMITFOLD; /* Otherwise we have a conflict or simply no match. */
}
/* Add or merge a snapshot. */
void lj_snap_add(jit_State *J)
{
MSize nsnap = J->cur.nsnap;
MSize nsnapmap = J->cur.nsnapmap;
/* Merge if no ins. inbetween or if requested and no guard inbetween. */
if ((nsnap > 0 && J->cur.snap[nsnap-1].ref == J->cur.nins) ||
(J->mergesnap && !irt_isguard(J->guardemit))) {
if (nsnap == 1) { /* But preserve snap #0 PC. */
emitir_raw(IRT(IR_NOP, IRT_NIL), 0, 0);
goto nomerge;
}
nsnapmap = J->cur.snap[--nsnap].mapofs;
} else {
nomerge:
lj_snap_grow_buf(J, nsnap+1);
J->cur.nsnap = (uint16_t)(nsnap+1);
}
J->mergesnap = 0;
J->guardemit.irt = 0;
snapshot_stack(J, &J->cur.snap[nsnap], nsnapmap);
}
/* Emit the conversions collected during backpropagation. */
static IRRef narrow_conv_emit(jit_State *J, NarrowConv *nc)
{
/* The fins fields must be saved now -- emitir() overwrites them. */
IROpT guardot = irt_isguard(fins->t) ? IRTG(IR_ADDOV-IR_ADD, 0) : 0;
IROpT convot = fins->ot;
IRRef1 convop2 = fins->op2;
NarrowIns *next = nc->stack; /* List of instructions from backpropagation. */
NarrowIns *last = nc->sp;
NarrowIns *sp = nc->stack; /* Recycle the stack to store operands. */
while (next < last) { /* Simple stack machine to process the ins. list. */
NarrowIns ref = *next++;
IROpT op = narrow_op(ref);
if (op == NARROW_REF) {
*sp++ = ref;
} else if (op == NARROW_CONV) {
*sp++ = emitir_raw(convot, ref, convop2); /* Raw emit avoids a loop. */
} else if (op == NARROW_INT) {
lua_assert(next < last);
*sp++ = lj_ir_kint(J, *next++);
} else { /* Regular IROpT. Pops two operands and pushes one result. */
IRRef mode = nc->mode;
lua_assert(sp >= nc->stack+2);
sp--;
/* Omit some overflow checks for array indexing. See comments above. */
if (mode == IRTOINT_INDEX) {
if (next == last && irref_isk(narrow_ref(sp[0])) &&
(uint32_t)IR(narrow_ref(sp[0]))->i + 0x40000000 < 0x80000000)
guardot = 0;
else
mode = IRTOINT_CHECK; /* Otherwise cache a stronger check. */
}
sp[-1] = emitir(op+guardot, sp[-1], sp[0]);
/* Add to cache. */
narrow_bpc_set(J, narrow_ref(ref), narrow_ref(sp[-1]), mode);
}
}
lua_assert(sp == nc->stack+1);
return nc->stack[0];
}
/* Add or merge a snapshot. */
void lj_snap_add(jit_State *J)
{
MSize nsnap = J->cur.nsnap;
MSize nsnapmap = J->cur.nsnapmap;
/* Merge if no ins. inbetween or if requested and no guard inbetween. */
if (J->mergesnap ? !irt_isguard(J->guardemit) :
(nsnap > 0 && J->cur.snap[nsnap-1].ref == J->cur.nins)) {
nsnapmap = J->cur.snap[--nsnap].mapofs;
} else {
/* Need to grow snapshot buffer? */
if (LJ_UNLIKELY(nsnap >= J->sizesnap)) {
MSize maxsnap = (MSize)J->param[JIT_P_maxsnap];
if (nsnap >= maxsnap)
lj_trace_err(J, LJ_TRERR_SNAPOV);
lj_mem_growvec(J->L, J->snapbuf, J->sizesnap, maxsnap, SnapShot);
J->cur.snap = J->snapbuf;
}
J->cur.nsnap = (uint16_t)(nsnap+1);
}
J->mergesnap = 0;
J->guardemit.irt = 0;
snapshot_stack(J, &J->cur.snap[nsnap], nsnapmap);
}
/* 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_CONV && (ir->op2 & IRCONV_SRCMASK) == IRT_INT) {
if ((nc->mode & IRCONV_CONVMASK) <= IRCONV_ANY)
narrow_stripov_backprop(nc, ir->op1, depth+1);
else
*nc->sp++ = NARROWINS(NARROW_REF, ir->op1); /* Undo conversion. */
if (nc->t == IRT_I64)
*nc->sp++ = NARROWINS(NARROW_SEXT, 0); /* Sign-extend integer. */
return 0;
} else if (ir->o == IR_KNUM) { /* Narrow FP constant. */
lua_Number n = ir_knum(ir)->n;
if ((nc->mode & IRCONV_CONVMASK) == IRCONV_TOBIT) {
/* Allows a wider range of constants. */
int64_t k64 = (int64_t)n;
if (n == (lua_Number)k64) { /* Only if const 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);
/* Only if constant is a small integer. */
if (checki16(k) && n == (lua_Number)k) {
*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. */
cref = J->chain[fins->o];
while (cref > ref) {
IRIns *cr = IR(cref);
if (cr->op1 == ref &&
(fins->o == IR_TOBIT ||
((cr->op2 & IRCONV_MODEMASK) == (nc->mode & IRCONV_MODEMASK) &&
irt_isguard(cr->t) >= irt_isguard(fins->t)))) {
*nc->sp++ = NARROWINS(NARROW_REF, cref);
return 0; /* Already there, no additional conversion needed. */
}
cref = cr->prev;
}
/* Backpropagate across ADD/SUB. */
if (ir->o == IR_ADD || ir->o == IR_SUB) {
/* Try cache lookup first. */
IRRef mode = nc->mode;
BPropEntry *bp;
/* Inner conversions need a stronger check. */
if ((mode & IRCONV_CONVMASK) == IRCONV_INDEX && depth > 0)
mode += IRCONV_CHECK-IRCONV_INDEX;
bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
if (bp) {
*nc->sp++ = NARROWINS(NARROW_REF, bp->val);
return 0;
} else if (nc->t == IRT_I64) {
/* Try sign-extending from an existing (checked) conversion to int. */
mode = (IRT_INT<<5)|IRT_NUM|IRCONV_INDEX;
bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
if (bp) {
*nc->sp++ = NARROWINS(NARROW_REF, bp->val);
*nc->sp++ = NARROWINS(NARROW_SEXT, 0);
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(IRT(ir->o, nc->t), ref);
return count;
}
nc->sp = savesp; /* Too many conversions, need to backtrack. */
}
}
/* Otherwise add a conversion. */
*nc->sp++ = NARROWINS(NARROW_CONV, ref);
return 1;
}
/* Unroll loop. */
static void loop_unroll(jit_State *J)
{
IRRef1 phi[LJ_MAX_PHI];
uint32_t nphi = 0;
IRRef1 *subst;
SnapNo onsnap;
SnapShot *osnap, *loopsnap;
SnapEntry *loopmap, *psentinel;
IRRef ins, invar;
/* Use temp buffer for substitution table.
** Only non-constant refs in [REF_BIAS,invar) are valid indexes.
** Caveat: don't call into the VM or run the GC or the buffer may be gone.
*/
invar = J->cur.nins;
subst = (IRRef1 *)lj_str_needbuf(J->L, &G(J->L)->tmpbuf,
(invar-REF_BIAS)*sizeof(IRRef1)) - REF_BIAS;
subst[REF_BASE] = REF_BASE;
/* LOOP separates the pre-roll from the loop body. */
emitir_raw(IRTG(IR_LOOP, IRT_NIL), 0, 0);
/* Grow snapshot buffer and map for copy-substituted snapshots.
** Need up to twice the number of snapshots minus #0 and loop snapshot.
** Need up to twice the number of entries plus fallback substitutions
** from the loop snapshot entries for each new snapshot.
** Caveat: both calls may reallocate J->cur.snap and J->cur.snapmap!
*/
onsnap = J->cur.nsnap;
lj_snap_grow_buf(J, 2*onsnap-2);
lj_snap_grow_map(J, J->cur.nsnapmap*2+(onsnap-2)*J->cur.snap[onsnap-1].nent);
/* The loop snapshot is used for fallback substitutions. */
loopsnap = &J->cur.snap[onsnap-1];
loopmap = &J->cur.snapmap[loopsnap->mapofs];
/* The PC of snapshot #0 and the loop snapshot must match. */
psentinel = &loopmap[loopsnap->nent];
lua_assert(*psentinel == J->cur.snapmap[J->cur.snap[0].nent]);
*psentinel = SNAP(255, 0, 0); /* Replace PC with temporary sentinel. */
/* Start substitution with snapshot #1 (#0 is empty for root traces). */
osnap = &J->cur.snap[1];
/* Copy and substitute all recorded instructions and snapshots. */
for (ins = REF_FIRST; ins < invar; ins++) {
IRIns *ir;
IRRef op1, op2;
if (ins >= osnap->ref) /* Instruction belongs to next snapshot? */
loop_subst_snap(J, osnap++, loopmap, subst); /* Copy-substitute it. */
/* Substitute instruction operands. */
ir = IR(ins);
op1 = ir->op1;
if (!irref_isk(op1)) op1 = subst[op1];
op2 = ir->op2;
if (!irref_isk(op2)) op2 = subst[op2];
if (irm_kind(lj_ir_mode[ir->o]) == IRM_N &&
op1 == ir->op1 && op2 == ir->op2) { /* Regular invariant ins? */
subst[ins] = (IRRef1)ins; /* Shortcut. */
} else {
/* Re-emit substituted instruction to the FOLD/CSE/etc. pipeline. */
IRType1 t = ir->t; /* Get this first, since emitir may invalidate ir. */
IRRef ref = tref_ref(emitir(ir->ot & ~IRT_ISPHI, op1, op2));
subst[ins] = (IRRef1)ref;
if (ref != ins) {
IRIns *irr = IR(ref);
if (ref < invar) { /* Loop-carried dependency? */
/* Potential PHI? */
if (!irref_isk(ref) && !irt_isphi(irr->t) && !irt_ispri(irr->t)) {
irt_setphi(irr->t);
if (nphi >= LJ_MAX_PHI)
lj_trace_err(J, LJ_TRERR_PHIOV);
phi[nphi++] = (IRRef1)ref;
}
/* Check all loop-carried dependencies for type instability. */
if (!irt_sametype(t, irr->t)) {
if (irt_isinteger(t) && irt_isinteger(irr->t))
continue;
else if (irt_isnum(t) && irt_isinteger(irr->t)) /* Fix int->num. */
ref = tref_ref(emitir(IRTN(IR_CONV), ref, IRCONV_NUM_INT));
else if (irt_isnum(irr->t) && irt_isinteger(t)) /* Fix num->int. */
ref = tref_ref(emitir(IRTGI(IR_CONV), ref,
IRCONV_INT_NUM|IRCONV_CHECK));
else
lj_trace_err(J, LJ_TRERR_TYPEINS);
subst[ins] = (IRRef1)ref;
irr = IR(ref);
goto phiconv;
}
} else if (ref != REF_DROP && irr->o == IR_CONV &&
ref > invar && irr->op1 < invar) {
/* May need an extra PHI for a CONV. */
ref = irr->op1;
irr = IR(ref);
phiconv:
if (ref < invar && !irref_isk(ref) && !irt_isphi(irr->t)) {
irt_setphi(irr->t);
if (nphi >= LJ_MAX_PHI)
lj_trace_err(J, LJ_TRERR_PHIOV);
phi[nphi++] = (IRRef1)ref;
}
}
}
}
}
if (!irt_isguard(J->guardemit)) /* Drop redundant snapshot. */
J->cur.nsnapmap = (uint16_t)J->cur.snap[--J->cur.nsnap].mapofs;
lua_assert(J->cur.nsnapmap <= J->sizesnapmap);
*psentinel = J->cur.snapmap[J->cur.snap[0].nent]; /* Restore PC. */
loop_emit_phi(J, subst, phi, nphi, onsnap);
}
/* 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);
}
}
/* 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;
}
