FFI: Rename IR_CNEWP to IR_CNEWI and use it to box 64 bit integers.

Generates smaller IR and DCE eliminates many intermediate boxes.
Needs allocation sinking to eliminate the boxes kept alive by PHIs.

5 files changed, 140 insertions, 82 deletions

diff --git a/src/lj_asm.c b/src/lj_asm.c
index 8864c9a3..77b55f0c 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -2518,7 +2518,7 @@ static void asm_cnew(ASMState *as, IRIns *ir)
 {
   CTState *cts = ctype_ctsG(J2G(as->J));
   CTypeID typeid = (CTypeID)IR(ir->op1)->i;
-  CTSize sz = (ir->o == IR_CNEWP || ir->op2 == REF_NIL) ?
+  CTSize sz = (ir->o == IR_CNEWI || ir->op2 == REF_NIL) ?
               lj_ctype_size(cts, typeid) : (CTSize)IR(ir->op2)->i;
   const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_mem_newgco];
   IRRef args[2];
@@ -2529,33 +2529,45 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   as->gcsteps++;
   asm_setupresult(as, ir, ci);  /* GCcdata * */
 
-  /* Initialize pointer cdata object. */
-  if (ir->o == IR_CNEWP) {
+  /* Initialize immutable cdata object. */
+  if (ir->o == IR_CNEWI) {
+    RegSet allow = (RSET_GPR & ~RSET_SCRATCH);
+#if LJ_64
+    Reg r64 = sz == 8 ? REX_64 : 0;
     if (irref_isk(ir->op2)) {
       IRIns *irk = IR(ir->op2);
-#if LJ_64
-      if (irk->o == IR_KINT64) {
-        uint64_t k = ir_k64(irk)->u64;
-        lua_assert(sz == 8);
-        if (checki32((int64_t)k)) {
-          emit_i32(as, (int32_t)k);
-          emit_rmro(as, XO_MOVmi, REX_64, RID_RET, sizeof(GCcdata));
-        } else {
-          emit_movtomro(as, RID_ECX|REX_64, RID_RET, sizeof(GCcdata));
-          emit_loadu64(as, RID_ECX, k);
-        }
+      uint64_t k = irk->o == IR_KINT64 ? ir_k64(irk)->u64 :
+                                         (uint64_t)(uint32_t)irk->i;
+      if (sz == 4 || checki32((int64_t)k)) {
+        emit_i32(as, (int32_t)k);
+        emit_rmro(as, XO_MOVmi, r64, RID_RET, sizeof(GCcdata));
       } else {
-#endif
-        lua_assert(sz == 4);
-        emit_movmroi(as, RID_RET, sizeof(GCcdata), irk->i);
-#if LJ_64
+        emit_movtomro(as, RID_ECX + r64, RID_RET, sizeof(GCcdata));
+        emit_loadu64(as, RID_ECX, k);
       }
-#endif
     } else {
-      Reg r = ra_alloc1(as, ir->op2, (RSET_GPR & ~RSET_SCRATCH));
-      emit_movtomro(as, r + ((LJ_64 && sz == 8) ? REX_64 : 0),
-                    RID_RET, sizeof(GCcdata));
+      Reg r = ra_alloc1(as, ir->op2, allow);
+      emit_movtomro(as, r + r64, RID_RET, sizeof(GCcdata));
+    }
+#else
+    int32_t ofs = sizeof(GCcdata);
+    if (LJ_HASFFI && sz == 8) {
+      ofs += 4; ir++;
+      lua_assert(ir->o == IR_HIOP);
     }
+    do {
+      if (irref_isk(ir->op2)) {
+        emit_movmroi(as, RID_RET, ofs, IR(ir->op2)->i);
+      } else {
+        Reg r = ra_alloc1(as, ir->op2, allow);
+        emit_movtomro(as, r, RID_RET, ofs);
+        rset_clear(allow, r);
+      }
+      if (!LJ_HASFFI || ofs == sizeof(GCcdata)) break;
+      ofs -= 4; ir--;
+    } while (1);
+#endif
+    lua_assert(sz == 4 || (sz == 8 && (LJ_64 || LJ_HASFFI)));
   }
 
   /* Combine initialization of marked, gct and typeid. */
@@ -3289,6 +3301,9 @@ static void asm_hiop(ASMState *as, IRIns *ir)
     if (!uselo)
       ra_allocref(as, ir->op1, RID2RSET(RID_RET));  /* Mark call as used. */
     break;
+  case IR_CNEWI:
+    /* Nothing to do here. Handled by CNEWI itself. */
+    break;
   default: lua_assert(0); break;
   }
 #else
@@ -4057,7 +4072,7 @@ static void asm_ir(ASMState *as, IRIns *ir)
   case IR_SNEW: asm_snew(as, ir); break;
   case IR_TNEW: asm_tnew(as, ir); break;
   case IR_TDUP: asm_tdup(as, ir); break;
-  case IR_CNEW: case IR_CNEWP: asm_cnew(as, ir); break;
+  case IR_CNEW: case IR_CNEWI: asm_cnew(as, ir); break;
 
   /* Write barriers. */
   case IR_TBAR: asm_tbar(as, ir); break;
@@ -4164,8 +4179,10 @@ static void asm_setup_regsp(ASMState *as, GCtrace *T)
       }
 #if LJ_32 && LJ_HASFFI
     case IR_HIOP:
-      if ((ir-1)->o == IR_CALLN)
+      if ((ir-1)->o == IR_CALLN) {
         ir->prev = REGSP_HINT(RID_RETHI);
+        continue;
+      }
       break;
 #endif
     /* C calls evict all scratch regs and return results in RID_RET. */
@@ -4174,7 +4191,7 @@ static void asm_setup_regsp(ASMState *as, GCtrace *T)
       if (as->evenspill < 3)  /* lj_str_new and lj_tab_newkey need 3 args. */
         as->evenspill = 3;
 #endif
-    case IR_TNEW: case IR_TDUP: case IR_CNEW: case IR_CNEWP: case IR_TOSTR:
+    case IR_TNEW: case IR_TDUP: case IR_CNEW: case IR_CNEWI: case IR_TOSTR:
       ir->prev = REGSP_HINT(RID_RET);
       if (inloop)
         as->modset = RSET_SCRATCH;
diff --git a/src/lj_crecord.c b/src/lj_crecord.c
index cd5c7d49..1ba98ae8 100644
--- a/src/lj_crecord.c
+++ b/src/lj_crecord.c
@@ -137,7 +137,7 @@ static int crec_isnonzero(CType *s, void *p)
   }
 }
 
-static void crec_ct_ct(jit_State *J, CType *d, CType *s, TRef dp, TRef sp,
+static TRef crec_ct_ct(jit_State *J, CType *d, CType *s, TRef dp, TRef sp,
                        void *svisnz)
 {
   CTSize dsize = d->size, ssize = s->size;
@@ -190,6 +190,7 @@ static void crec_ct_ct(jit_State *J, CType *d, CType *s, TRef dp, TRef sp,
 #endif
   xstore:
     if (dt == IRT_I64 || dt == IRT_U64) lj_needsplit(J);
+    if (dp == 0) return sp;
     emitir(IRT(IR_XSTORE, dt), dp, sp);
     break;
   case CCX(I, C):
@@ -290,6 +291,7 @@ static void crec_ct_ct(jit_State *J, CType *d, CType *s, TRef dp, TRef sp,
     lj_trace_err(J, LJ_TRERR_NYICONV);
     break;
   }
+  return 0;
 }
 
 /* -- Convert C type to TValue (load) ------------------------------------- */
@@ -306,21 +308,18 @@ static TRef crec_tv_ct(jit_State *J, CType *s, CTypeID sid, TRef sp)
       goto err_nyi;  /* NYI: copyval of >64 bit integers. */
     tr = emitir(IRT(IR_XLOAD, t), sp, 0);
     if (t == IRT_FLOAT || t == IRT_U32) {  /* Keep uint32_t/float as numbers. */
-      tr = emitconv(tr, IRT_NUM, t, 0);
+      return emitconv(tr, IRT_NUM, t, 0);
     } else if (t == IRT_I64 || t == IRT_U64) {  /* Box 64 bit integer. */
-      TRef dp = emitir(IRTG(IR_CNEW, IRT_CDATA), lj_ir_kint(J, sid), TREF_NIL);
-      TRef ptr = emitir(IRT(IR_ADD, IRT_PTR), dp,
-                        lj_ir_kintp(J, sizeof(GCcdata)));
-      emitir(IRT(IR_XSTORE, t), ptr, tr);
+      sp = tr;
       lj_needsplit(J);
-      return dp;
     } else if ((sinfo & CTF_BOOL)) {
       /* Assume not equal to zero. Fixup and emit pending guard later. */
       lj_ir_set(J, IRTGI(IR_NE), tr, lj_ir_kint(J, 0));
       J->postproc = LJ_POST_FIXGUARD;
-      tr = TREF_TRUE;
+      return TREF_TRUE;
+    } else {
+      return tr;
     }
-    return tr;
   } else if (ctype_isptr(sinfo)) {
     IRType t = (LJ_64 && s->size == 8) ? IRT_P64 : IRT_P32;
     sp = emitir(IRT(IR_XLOAD, t), sp, 0);
@@ -345,13 +344,13 @@ static TRef crec_tv_ct(jit_State *J, CType *s, CTypeID sid, TRef sp)
   err_nyi:
     lj_trace_err(J, LJ_TRERR_NYICONV);
   }
-  /* Box pointer or ref. */
-  return emitir(IRTG(IR_CNEWP, IRT_CDATA), lj_ir_kint(J, sid), sp);
+  /* Box pointer, ref or 64 bit integer. */
+  return emitir(IRTG(IR_CNEWI, IRT_CDATA), lj_ir_kint(J, sid), sp);
 }
 
 /* -- Convert TValue to C type (store) ------------------------------------ */
 
-static void crec_ct_tv(jit_State *J, CType *d, TRef dp, TRef sp, TValue *sval)
+static TRef crec_ct_tv(jit_State *J, CType *d, TRef dp, TRef sp, TValue *sval)
 {
   CTState *cts = ctype_ctsG(J2G(J));
   CTypeID sid = CTID_P_VOID;
@@ -402,6 +401,12 @@ static void crec_ct_tv(jit_State *J, CType *d, TRef dp, TRef sp, TValue *sval)
       } else {
         goto doconv;  /* The pointer value was loaded, don't load number. */
       }
+
+    } else if (ctype_isnum(s->info) && s->size == 8) {
+      IRType t = (s->info & CTF_UNSIGNED) ? IRT_U64 : IRT_I64;
+      sp = emitir(IRT(IR_FLOAD, t), sp, IRFL_CDATA_INT64);
+      lj_needsplit(J);
+      goto doconv;
     } else {
       sp = emitir(IRT(IR_ADD, IRT_PTR), sp, lj_ir_kintp(J, sizeof(GCcdata)));
     }
@@ -418,7 +423,7 @@ static void crec_ct_tv(jit_State *J, CType *d, TRef dp, TRef sp, TValue *sval)
   s = ctype_get(cts, sid);
 doconv:
   if (ctype_isenum(d->info)) d = ctype_child(cts, d);
-  crec_ct_ct(J, d, s, dp, sp, svisnz);
+  return crec_ct_ct(J, d, s, dp, sp, svisnz);
 }
 
 /* -- C data metamethods -------------------------------------------------- */
@@ -578,15 +583,18 @@ static void crec_alloc(jit_State *J, RecordFFData *rd, CTypeID id)
   CTState *cts = ctype_ctsG(J2G(J));
   CTSize sz;
   CTInfo info = lj_ctype_info(cts, id, &sz);
+  CType *d = ctype_raw(cts, id);
   TRef trid;
   if (sz == 0 || sz > 64 || (info & CTF_VLA) || ctype_align(info) > CT_MEMALIGN)
     lj_trace_err(J, LJ_TRERR_NYICONV);  /* NYI: large/special allocations. */
   trid = lj_ir_kint(J, id);
-  if (ctype_isptr(info)) {
-    TRef sp = J->base[1] ? J->base[1] : lj_ir_kptr(J, NULL);
-    J->base[0] = emitir(IRTG(IR_CNEWP, IRT_CDATA), trid, sp);
+  /* Use special instruction to box pointer or 64 bit integer. */
+  if (ctype_isptr(info) || (ctype_isnum(info) && sz == 8)) {
+    TRef sp = J->base[1] ? crec_ct_tv(J, d, 0, J->base[1], &rd->argv[1]) :
+              ctype_isptr(info) ? lj_ir_kptr(J, NULL) :
+              (lj_needsplit(J), lj_ir_kint64(J, 0));
+    J->base[0] = emitir(IRTG(IR_CNEWI, IRT_CDATA), trid, sp);
   } else {
-    CType *d = ctype_raw(cts, id);
     TRef trcd = emitir(IRTG(IR_CNEW, IRT_CDATA), trid, TREF_NIL);
     J->base[0] = trcd;
     if (J->base[1] && !J->base[2] && !lj_cconv_multi_init(d, &rd->argv[1])) {
@@ -598,7 +606,7 @@ static void crec_alloc(jit_State *J, RecordFFData *rd, CTypeID id)
       TValue tv;
       TValue *sval = &tv;
       MSize i;
-      setnumV(&tv, 0);
+      tv.u64 = 0;
       if (!(ctype_isnum(dc->info) || ctype_isptr(dc->info)))
         lj_trace_err(J, LJ_TRERR_NYICONV);  /* NYI: init array of aggregates. */
       for (i = 1, ofs = 0; ofs < sz; ofs += esize) {
@@ -645,11 +653,16 @@ static void crec_alloc(jit_State *J, RecordFFData *rd, CTypeID id)
         }
       }
     } else {
-      TRef sp, dp;
+      TRef dp;
     single_init:
-      sp = J->base[1] ? J->base[1] : lj_ir_kint(J, 0);
       dp = emitir(IRT(IR_ADD, IRT_PTR), trcd, lj_ir_kintp(J, sizeof(GCcdata)));
-      crec_ct_tv(J, d, dp, sp, &rd->argv[1]);
+      if (J->base[1]) {
+        crec_ct_tv(J, d, dp, J->base[1], &rd->argv[1]);
+      } else {
+        TValue tv;
+        tv.u64 = 0;
+        crec_ct_tv(J, d, dp, lj_ir_kint(J, 0), &tv);
+      }
     }
   }
 }
@@ -669,7 +682,7 @@ static TRef crec_arith_int64(jit_State *J, TRef *sp, CType **s, MMS mm)
   if (ctype_isnum(s[0]->info) && ctype_isnum(s[1]->info)) {
     IRType dt;
     CTypeID id;
-    TRef tr, dp, ptr;
+    TRef tr;
     MSize i;
     lj_needsplit(J);
     if (((s[0]->info & CTF_UNSIGNED) && s[0]->size == 8) ||
@@ -702,10 +715,7 @@ static TRef crec_arith_int64(jit_State *J, TRef *sp, CType **s, MMS mm)
     } else {
       tr = emitir(IRT(mm+(int)IR_ADD-(int)MM_add, dt), sp[0], sp[1]);
     }
-    dp = emitir(IRTG(IR_CNEW, IRT_CDATA), lj_ir_kint(J, id), TREF_NIL);
-    ptr = emitir(IRT(IR_ADD, IRT_PTR), dp, lj_ir_kintp(J, sizeof(GCcdata)));
-    emitir(IRT(IR_XSTORE, dt), ptr, tr);
-    return dp;
+    return emitir(IRTG(IR_CNEWI, IRT_CDATA), lj_ir_kint(J, id), tr);
   }
   return 0;
 }
@@ -767,7 +777,7 @@ static TRef crec_arith_ptr(jit_State *J, TRef *sp, CType **s, MMS mm)
     tr = emitir(IRT(IR_ADD, IRT_PTR), sp[0], tr);
     id = lj_ctype_intern(cts, CTINFO(CT_PTR, CTALIGN_PTR|ctype_cid(ctp->info)),
                          CTSIZE_PTR);
-    return emitir(IRTG(IR_CNEWP, IRT_CDATA), lj_ir_kint(J, id), tr);
+    return emitir(IRTG(IR_CNEWI, IRT_CDATA), lj_ir_kint(J, id), tr);
   }
 }
 
@@ -787,6 +797,11 @@ void LJ_FASTCALL recff_cdata_arith(jit_State *J, RecordFFData *rd)
         IRType t = (LJ_64 && ct->size == 8) ? IRT_P64 : IRT_P32;
         if (ctype_isref(ct->info)) ct = ctype_rawchild(cts, ct);
         tr = emitir(IRT(IR_FLOAD, t), tr, IRFL_CDATA_PTR);
+      } else if (ctype_isnum(ct->info) && ct->size == 8) {
+        IRType t = (ct->info & CTF_UNSIGNED) ? IRT_U64 : IRT_I64;
+        tr = emitir(IRT(IR_FLOAD, t), tr, IRFL_CDATA_INT64);
+        lj_needsplit(J);
+        goto ok;
       } else {
         tr = emitir(IRT(IR_ADD, IRT_PTR), tr, lj_ir_kintp(J, sizeof(GCcdata)));
       }
@@ -807,6 +822,7 @@ void LJ_FASTCALL recff_cdata_arith(jit_State *J, RecordFFData *rd)
     } else if (!tref_isnum(tr)) {
       goto err_type;
     }
+  ok:
     s[i] = ct;
     sp[i] = tr;
   }
diff --git a/src/lj_ir.h b/src/lj_ir.h
index dfafc5db..bde0ac04 100644
--- a/src/lj_ir.h
+++ b/src/lj_ir.h
@@ -113,7 +113,7 @@
   _(TNEW,       AW, lit, lit) \
   _(TDUP,       AW, ref, ___) \
   _(CNEW,       AW, ref, ref) \
-  _(CNEWP,      NW, ref, ref)  /* CSE is ok, not marked as A. */ \
+  _(CNEWI,      NW, ref, ref)  /* CSE is ok, not marked as A. */ \
   \
   /* Write barriers. */ \
   _(TBAR,       S , ref, ___) \
@@ -188,7 +188,9 @@ IRFPMDEF(FPMENUM)
   _(UDATA_UDTYPE, offsetof(GCudata, udtype)) \
   _(UDATA_FILE, sizeof(GCudata)) \
   _(CDATA_TYPEID, offsetof(GCcdata, typeid)) \
-  _(CDATA_PTR, sizeof(GCcdata))
+  _(CDATA_PTR,  sizeof(GCcdata)) \
+  _(CDATA_INT64, sizeof(GCcdata)) \
+  _(CDATA_INT64HI, sizeof(GCcdata) + 4)
 
 typedef enum {
 #define FLENUM(name, ofs)       IRFL_##name,
@@ -588,12 +590,12 @@ typedef union IRIns {
 #define ir_kptr(ir) \
   check_exp((ir)->o == IR_KPTR || (ir)->o == IR_KKPTR, mref((ir)->ptr, void))
 
+LJ_STATIC_ASSERT((int)IRT_GUARD == (int)IRM_W);
+
 /* A store or any other op with a non-weak guard has a side-effect. */
 static LJ_AINLINE int ir_sideeff(IRIns *ir)
 {
   return (((ir->t.irt | ~IRT_GUARD) & lj_ir_mode[ir->o]) >= IRM_S);
 }
 
-LJ_STATIC_ASSERT((int)IRT_GUARD == (int)IRM_W);
-
 #endif
diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index c3b0a082..28758013 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -154,7 +154,7 @@ typedef IRRef (LJ_FASTCALL *FoldFunc)(jit_State *J);
 #define gcstep_barrier(J, ref) \
   ((ref) < J->chain[IR_LOOP] && \
    (J->chain[IR_SNEW] || J->chain[IR_TNEW] || J->chain[IR_TDUP] || \
-    J->chain[IR_CNEW] || J->chain[IR_CNEWP] || J->chain[IR_TOSTR]))
+    J->chain[IR_CNEW] || J->chain[IR_CNEWI] || J->chain[IR_TOSTR]))
 
 /* -- Constant folding for FP numbers ------------------------------------- */
 
@@ -307,7 +307,7 @@ static uint64_t kfold_int64arith(uint64_t k1, uint64_t k2, IROp op)
   case IR_BOR: k1 |= k2; break;
   case IR_BXOR: k1 ^= k2; break;
 #endif
-  default: lua_assert(0); break;
+  default: UNUSED(k2); lua_assert(0); break;
   }
   return k1;
 }
@@ -1765,18 +1765,28 @@ LJFOLDF(fload_cdata_typeid_kgc)
   return NEXTFOLD;
 }
 
+/* The content of int64 cdata objects is immutable. */
+LJFOLD(FLOAD KGC IRFL_CDATA_INT64)
+LJFOLDF(fload_cdata_int64_kgc)
+{
+  if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD))
+    return INT64FOLD(*(uint64_t *)cdataptr(ir_kcdata(fleft)));
+  return NEXTFOLD;
+}
+
 LJFOLD(FLOAD CNEW IRFL_CDATA_TYPEID)
-LJFOLD(FLOAD CNEWP IRFL_CDATA_TYPEID)
+LJFOLD(FLOAD CNEWI IRFL_CDATA_TYPEID)
 LJFOLDF(fload_cdata_typeid_cnew)
 {
   if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD))
-    return fleft->op1;  /* No PHI barrier needed. CNEW/CNEWP op1 is const. */
+    return fleft->op1;  /* No PHI barrier needed. CNEW/CNEWI op1 is const. */
   return NEXTFOLD;
 }
 
-/* Pointer cdata objects are immutable. */
-LJFOLD(FLOAD CNEWP IRFL_CDATA_PTR)
-LJFOLDF(fload_cdata_ptr_cnew)
+/* Pointer and int64 cdata objects are immutable. */
+LJFOLD(FLOAD CNEWI IRFL_CDATA_PTR)
+LJFOLD(FLOAD CNEWI IRFL_CDATA_INT64)
+LJFOLDF(fload_cdata_ptr_int64_cnew)
 {
   if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD))
     return fleft->op2;  /* Fold even across PHI to avoid allocations. */
@@ -1786,6 +1796,7 @@ LJFOLDF(fload_cdata_ptr_cnew)
 LJFOLD(FLOAD any IRFL_STR_LEN)
 LJFOLD(FLOAD any IRFL_CDATA_TYPEID)
 LJFOLD(FLOAD any IRFL_CDATA_PTR)
+LJFOLD(FLOAD any IRFL_CDATA_INT64)
 LJFOLD(VLOAD any any)  /* Vararg loads have no corresponding stores. */
 LJFOLDX(lj_opt_cse)
 
diff --git a/src/lj_opt_split.c b/src/lj_opt_split.c
index f53616b3..90b2b49c 100644
--- a/src/lj_opt_split.c
+++ b/src/lj_opt_split.c
@@ -159,7 +159,8 @@ static void split_ir(jit_State *J)
       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 = (IRRef1)ref;  /* Identity substitution for loword. */
+      ir->prev = ref;  /* Identity substitution for loword. */
+      hisubst[ref] = 0;
     }
   }
 
@@ -168,6 +169,7 @@ static void split_ir(jit_State *J)
     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;
@@ -175,10 +177,11 @@ static void split_ir(jit_State *J)
     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 hi = hisubst[ir->op1];
+      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:
@@ -186,13 +189,13 @@ static void split_ir(jit_State *J)
         /* 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 = hi; nir->op2 = hisubst[ir->op2];
+          nir->op1 = hiref; nir->op2 = hisubst[ir->op2];
           hi = nref;
           break;
         }
         /* fallthrough */
       case IR_NEG:
-        hi = split_emit(J, IRTI(IR_HIOP), hi, hisubst[ir->op2]);
+        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);
@@ -212,6 +215,13 @@ static void split_ir(jit_State *J)
                           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
@@ -220,19 +230,18 @@ static void split_ir(jit_State *J)
         break;
       case IR_XSTORE:
 #if LJ_LE
-        hi = hisubst[ir->op2];
+        hiref = hisubst[ir->op2];
 #else
-        hi = nir->op2; nir->op2 = hisubst[ir->op2];
+        hiref = nir->op2; nir->op2 = hisubst[ir->op2];
 #endif
-        split_emit(J, IRTI(IR_XSTORE), split_ptr(J, nir->op1), hi);
-        continue;
+        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. */
-          hisubst[ref] = hi;
           goto fwdlo;
         } else if ((ir->op2 & IRCONV_SEXT)) {  /* Sign-extend to 64 bit. */
           IRRef k31 = lj_ir_kint(J, 31);
@@ -242,27 +251,26 @@ static void split_ir(jit_State *J)
           nir->op2 = k31;
           hi = nref;
         } else {  /* Zero-extend to 64 bit. */
-          hisubst[ref] = lj_ir_kint(J, 0);
+          hi = lj_ir_kint(J, 0);
           goto fwdlo;
         }
         break;
         }
       case IR_PHI: {
-        IRRef hi2;
+        IRRef hiref2;
         if ((irref_isk(nir->op1) && irref_isk(nir->op2)) ||
             nir->op1 == nir->op2)
           J->cur.nins--;  /* Drop useless PHIs. */
-        hi2 = hisubst[ir->op2];
-        if (!((irref_isk(hi) && irref_isk(hi2)) || hi == hi2))
-          split_emit(J, IRTI(IR_PHI), hi, hi2);
-        continue;
+        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);
-        split_emit(J, IRTGI(IR_HIOP), hi, hisubst[ir->op2]);  /* Comparisons. */
-        continue;
+        lua_assert(ir->o <= IR_NE);  /* Comparisons. */
+        split_emit(J, IRTGI(IR_HIOP), hiref, hisubst[ir->op2]);
+        break;
       }
-      hisubst[ref] = hi;  /* Store hiword substitution. */
     } 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. */
@@ -277,9 +285,13 @@ static void split_ir(jit_State *J)
           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. */