4 files changed, 47 insertions, 84 deletions

diff --git a/src/lj_asm.c b/src/lj_asm.c
index 215dfb13..27ccfe9a 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -2283,55 +2283,12 @@ static void asm_tdup(ASMState *as, IRIns *ir)
 }
 
 #if LJ_HASFFI
-static RegSet asm_cnew_init(ASMState *as, IRRef ref, int32_t ofs, RegSet allow)
-{
-  IRIns *ir = IR(ref);
-  if (irref_isk(ref)) {
-#if LJ_64
-    if (ir->o == IR_KNUM || ir->o == IR_KINT64) {
-      uint64_t k = ir_k64(ir)->u64;
-      if (checki32((int64_t)k)) {
-        emit_i32(as, (int32_t)k);
-        emit_rmro(as, XO_MOVmi, REX_64, RID_RET, ofs);
-      } else {
-        emit_movtomro(as, RID_ECX|REX_64, RID_RET, ofs);
-        emit_loadu64(as, RID_ECX, k);
-      }
-    } else {
-      emit_movmroi(as, RID_RET, ofs, ir->i);
-    }
-#else
-    if (ir->o == IR_KNUM) {
-      emit_rmro(as, XO_MOVSDto, RID_XMM0, RID_RET, ofs);
-      emit_loadn(as, RID_XMM0, ir_k64(ir));
-    } else if (ir->o == IR_KINT64) {
-      uint64_t k = ir_k64(ir)->u64;
-      emit_movmroi(as, RID_RET, ofs, (int32_t)k);
-      emit_movmroi(as, RID_RET, ofs+4, (int32_t)(k >> 32));
-    } else {
-      emit_movmroi(as, RID_RET, ofs, ir->i);
-    }
-#endif
-  } else {
-    Reg r;
-    if (irt_isnum(ir->t)) {
-      r = ra_alloc1(as, ref, (RSET_FPR & allow));
-      emit_rmro(as, XO_MOVSDto, r, RID_RET, ofs);
-    } else {
-      r = ra_alloc1(as, ref, (RSET_GPR & allow));
-      emit_movtomro(as, REX_64IR(ir, r), RID_RET, ofs);
-    }
-    rset_clear(allow, r);
-  }
-  return allow;
-}
-
 static void asm_cnew(ASMState *as, IRIns *ir)
 {
   CTState *cts = ctype_ctsG(J2G(as->J));
-  CTypeID typeid = (CTypeID)IR(ir->op2)->i;
-  CTSize sz = (ir->o == IR_CNEWI || ir->op1 == REF_NIL) ?
-              lj_ctype_size(cts, typeid) : (CTSize)IR(ir->op1)->i;
+  CTypeID typeid = (CTypeID)IR(ir->op1)->i;
+  CTSize sz = (ir->o == IR_CNEWP || 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];
   lua_assert(sz != CTSIZE_INVALID);
@@ -2339,18 +2296,35 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   args[0] = ASMREF_L;     /* lua_State *L */
   args[1] = ASMREF_TMP1;  /* MSize size   */
   as->gcsteps++;
-  asm_setupresult(as, ir, ci);  /* GCobj * */
+  asm_setupresult(as, ir, ci);  /* GCcdata * */
 
-  /* Initialize immutable cdata object. */
-  if (ir->o == IR_CNEWI) {
-    RegSet allow = ~RSET_SCRATCH;
-    IRRef ref = ir->op1;
-    if (IR(ref)->o == IR_CARG) {  /* 2nd initializer. */
-      IRIns *ira = IR(ref);
-      allow = asm_cnew_init(as, ira->op2, sizeof(GCcdata) + (sz>>1), allow);
-      ref = ira->op1;
+  /* Initialize pointer cdata object. */
+  if (ir->o == IR_CNEWP) {
+    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);
+        }
+      } else {
+#endif
+        lua_assert(sz == 4);
+        emit_movmroi(as, RID_RET, sizeof(GCcdata), irk->i);
+#if LJ_64
+      }
+#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));
     }
-    asm_cnew_init(as, ref, sizeof(GCcdata), allow);  /* 1st initializer. */
   }
 
   /* Combine initialization of marked, gct and typeid. */
@@ -3675,7 +3649,7 @@ 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_CNEWI: asm_cnew(as, ir); break;
+  case IR_CNEW: case IR_CNEWP: asm_cnew(as, ir); break;
 
   /* Write barriers. */
   case IR_TBAR: asm_tbar(as, ir); break;
@@ -3793,7 +3767,7 @@ 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_CNEWI: case IR_TOSTR:
+    case IR_TNEW: case IR_TDUP: case IR_CNEW: case IR_CNEWP: 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 2f7e0f34..1df6c70c 100644
--- a/src/lj_crecord.c
+++ b/src/lj_crecord.c
@@ -260,7 +260,7 @@ static TRef crec_tv_ct(jit_State *J, CType *s, CTypeID sid, TRef sp)
     ptrdiff_t esz = (ptrdiff_t)(s->size >> 1);
     TRef ptr, tr1, tr2, dp;
     if (t == IRT_CDATA) goto err_nyi;  /* NYI: float IRType. */
-    dp = emitir(IRTG(IR_CNEW, IRT_CDATA), TREF_NIL, lj_ir_kint(J, sid));
+    dp = emitir(IRTG(IR_CNEW, IRT_CDATA), lj_ir_kint(J, sid), TREF_NIL);
     tr1 = emitir(IRT(IR_XLOAD, t), sp, 0);
     ptr = emitir(IRT(IR_ADD, IRT_PTR), sp, lj_ir_kintp(J, esz));
     tr2 = emitir(IRT(IR_XLOAD, t), ptr, 0);
@@ -274,7 +274,7 @@ static TRef crec_tv_ct(jit_State *J, CType *s, CTypeID sid, TRef sp)
   err_nyi:
     lj_trace_err(J, LJ_TRERR_NYICONV);
   }
-  return emitir(IRTG(IR_CNEWI, IRT_CDATA), sp, lj_ir_kint(J, sid));
+  return emitir(IRTG(IR_CNEWP, IRT_CDATA), lj_ir_kint(J, sid), sp);
 }
 
 /* -- Convert TValue to C type (store) ------------------------------------ */
@@ -300,7 +300,7 @@ static void crec_ct_tv(jit_State *J, CType *d, TRef dp, TRef sp, TValue *sval)
     s = ctype_raw(cts, sid);
     if (ctype_isptr(s->info)) {
       IRType t = (LJ_64 && s->size == 8) ? IRT_P64 : IRT_P32;
-      sp = emitir(IRT(IR_FLOAD, t), sp, IRFL_CDATA_INIT1);
+      sp = emitir(IRT(IR_FLOAD, t), sp, IRFL_CDATA_PTR);
       if (ctype_isref(s->info))
         s = ctype_rawchild(cts, s);
       else
@@ -356,7 +356,7 @@ void LJ_FASTCALL recff_cdata_index(jit_State *J, RecordFFData *rd)
   if (ctype_isptr(ct->info)) {
     IRType t = (LJ_64 && ct->size == 8) ? IRT_P64 : IRT_P32;
     if (ctype_isref(ct->info)) ct = ctype_rawchild(cts, ct);
-    ptr = emitir(IRT(IR_FLOAD, t), ptr, IRFL_CDATA_INIT1);
+    ptr = emitir(IRT(IR_FLOAD, t), ptr, IRFL_CDATA_PTR);
     ofs = 0;
     ptr = crec_reassoc_ofs(J, ptr, &ofs, 1);
   }
diff --git a/src/lj_ir.h b/src/lj_ir.h
index 6495a780..33edf76c 100644
--- a/src/lj_ir.h
+++ b/src/lj_ir.h
@@ -111,7 +111,7 @@
   _(TNEW,       AW, lit, lit) \
   _(TDUP,       AW, ref, ___) \
   _(CNEW,       AW, ref, ref) \
-  _(CNEWI,      NW, ref, ref)  /* CSE is ok, not marked as A. */ \
+  _(CNEWP,      NW, ref, ref)  /* CSE is ok, not marked as A. */ \
   \
   /* Write barriers. */ \
   _(TBAR,       S , ref, ___) \
@@ -188,9 +188,7 @@ IRFPMDEF(FPMENUM)
   _(UDATA_UDTYPE, offsetof(GCudata, udtype)) \
   _(UDATA_FILE, sizeof(GCudata)) \
   _(CDATA_TYPEID, offsetof(GCcdata, typeid)) \
-  _(CDATA_INIT1, sizeof(GCcdata)) \
-  _(CDATA_INIT2_4, sizeof(GCcdata)+4) \
-  _(CDATA_INIT2_8, sizeof(GCcdata)+8)
+  _(CDATA_PTR, sizeof(GCcdata))
 
 typedef enum {
 #define FLENUM(name, ofs)       IRFL_##name,
diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index 74b15f32..d2c20546 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -153,7 +153,7 @@ 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_CNEWI] || J->chain[IR_TOSTR]))
+    J->chain[IR_CNEW] || J->chain[IR_CNEWP] || J->chain[IR_TOSTR]))
 
 /* -- Constant folding for FP numbers ------------------------------------- */
 
@@ -1587,35 +1587,26 @@ LJFOLDF(fload_str_len_snew)
 
 /* The C type ID of cdata objects is immutable. */
 LJFOLD(FLOAD CNEW IRFL_CDATA_TYPEID)
-LJFOLD(FLOAD CNEWI IRFL_CDATA_TYPEID)
+LJFOLD(FLOAD CNEWP IRFL_CDATA_TYPEID)
 LJFOLDF(fload_cdata_typeid_cnewi)
 {
   if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD))
-    return fleft->op2;  /* No PHI barrier needed. CNEW/CNEWI op2 is const. */
+    return fleft->op1;  /* No PHI barrier needed. CNEW/CNEWP op1 is const. */
   return NEXTFOLD;
 }
 
-/* Fixed initializers in cdata objects are immutable. */
-LJFOLD(FLOAD CNEWI IRFL_CDATA_INIT1)
-LJFOLD(FLOAD CNEWI IRFL_CDATA_INIT2_4)
-LJFOLD(FLOAD CNEWI IRFL_CDATA_INIT2_8)
-LJFOLDF(fload_cdata_init_cnew)
+/* Pointer cdata objects are immutable. */
+LJFOLD(FLOAD CNEWP IRFL_CDATA_PTR)
+LJFOLDF(fload_cdata_ptr_cnewi)
 {
-  if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD)) {
-    IRIns *ir = fleft;
-    /* Fold even across PHI to avoid expensive allocations. */
-    lua_assert(ir->op1 != REF_NIL);
-    if (IR(ir->op1)->o == IR_CARG) ir = IR(ir->op1);
-    return fins->op2 == IRFL_CDATA_INIT1 ? ir->op1 : ir->op2;
-  }
+  if (LJ_LIKELY(J->flags & JIT_F_OPT_FOLD))
+    return fleft->op2;  /* Fold even across PHI to avoid allocations. */
   return NEXTFOLD;
 }
 
 LJFOLD(FLOAD any IRFL_STR_LEN)
 LJFOLD(FLOAD any IRFL_CDATA_TYPEID)
-LJFOLD(FLOAD any IRFL_CDATA_INIT1)
-LJFOLD(FLOAD any IRFL_CDATA_INIT2_4)
-LJFOLD(FLOAD any IRFL_CDATA_INIT2_8)
+LJFOLD(FLOAD any IRFL_CDATA_PTR)
 LJFOLD(VLOAD any any)  /* Vararg loads have no corresponding stores. */
 LJFOLDX(lj_opt_cse)