/* Narrowing of power operator or math.pow. */
TRef lj_opt_narrow_pow(jit_State *J, TRef rb, TRef rc, TValue *vc)
{
lua_Number n;
if (tvisstr(vc) && !lj_str_tonum(strV(vc), vc))
lj_trace_err(J, LJ_TRERR_BADTYPE);
n = numV(vc);
/* Limit narrowing for pow to small exponents (or for two constants). */
if ((tref_isk(rc) && tref_isint(rc) && tref_isk(rb)) ||
((J->flags & JIT_F_OPT_NARROW) &&
(numisint(n) && n >= -65536.0 && n <= 65536.0))) {
TRef tmp;
if (!tref_isinteger(rc)) {
if (tref_isstr(rc))
rc = emitir(IRTG(IR_STRTO, IRT_NUM), rc, 0);
rc = emitir(IRTGI(IR_TOINT), rc, IRTOINT_CHECK); /* Guarded TOINT! */
}
if (!tref_isk(rc)) { /* Range guard: -65536 <= i <= 65536 */
tmp = emitir(IRTI(IR_ADD), rc, lj_ir_kint(J, 65536-2147483647-1));
emitir(IRTGI(IR_LE), tmp, lj_ir_kint(J, 2*65536-2147483647-1));
}
return emitir(IRTN(IR_POWI), rb, rc);
}
/* FOLD covers most cases, but some are easier to do here. */
if (tref_isk(rb) && tvispone(ir_knum(IR(tref_ref(rb)))))
return rb; /* 1 ^ x ==> 1 */
rc = lj_ir_tonum(J, rc);
if (tref_isk(rc) && ir_knum(IR(tref_ref(rc)))->n == 0.5)
return emitir(IRTN(IR_FPMATH), rb, IRFPM_SQRT); /* x ^ 0.5 ==> sqrt(x) */
/* Split up b^c into exp2(c*log2(b)). Assembler may rejoin later. */
rb = emitir(IRTN(IR_FPMATH), rb, IRFPM_LOG2);
rc = emitir(IRTN(IR_MUL), rb, rc);
return emitir(IRTN(IR_FPMATH), rc, IRFPM_EXP2);
}
/* Replay constant from parent trace. */
static TRef snap_replay_const(jit_State *J, IRIns *ir)
{
/* Only have to deal with constants that can occur in stack slots. */
switch ((IROp)ir->o) {
case IR_KPRI: return TREF_PRI(irt_type(ir->t));
case IR_KINT: return lj_ir_kint(J, ir->i);
case IR_KGC: return lj_ir_kgc(J, ir_kgc(ir), irt_t(ir->t));
case IR_KNUM: return lj_ir_k64(J, IR_KNUM, ir_knum(ir));
case IR_KINT64: return lj_ir_k64(J, IR_KINT64, ir_kint64(ir));
case IR_KPTR: return lj_ir_kptr(J, ir_kptr(ir)); /* Continuation. */
default: lua_assert(0); return TREF_NIL; break;
}
}
/* 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;
}
/* 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;
}
