1 files changed, 838 insertions, 183 deletions

diff --git a/src/lj_asm.c b/src/lj_asm.c
index 9ff9215f..7abafbf4 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -11,6 +11,7 @@
 #if LJ_HASJIT
 
 #include "lj_gc.h"
+#include "lj_buf.h"
 #include "lj_str.h"
 #include "lj_tab.h"
 #include "lj_frame.h"
@@ -71,6 +72,7 @@ typedef struct ASMState {
   IRRef snaprename;     /* Rename highwater mark for snapshot check. */
   SnapNo snapno;        /* Current snapshot number. */
   SnapNo loopsnapno;    /* Loop snapshot number. */
+  int snapalloc;        /* Current snapshot needs allocation. */
   BloomFilter snapfilt1, snapfilt2;     /* Filled with snapshot refs. */
 
   IRRef fuseref;        /* Fusion limit (loopref, 0 or FUSE_DISABLED). */
@@ -85,18 +87,25 @@ typedef struct ASMState {
 
   MCode *mcbot;         /* Bottom of reserved MCode. */
   MCode *mctop;         /* Top of generated MCode. */
+  MCode *mctoporig;     /* Original top of generated MCode. */
   MCode *mcloop;        /* Pointer to loop MCode (or NULL). */
   MCode *invmcp;        /* Points to invertible loop branch (or NULL). */
   MCode *flagmcp;       /* Pending opportunity to merge flag setting ins. */
   MCode *realign;       /* Realign loop if not NULL. */
 
 #ifdef RID_NUM_KREF
-  int32_t krefk[RID_NUM_KREF];
+  intptr_t krefk[RID_NUM_KREF];
 #endif
   IRRef1 phireg[RID_MAX];  /* PHI register references. */
   uint16_t parentmap[LJ_MAX_JSLOTS];  /* Parent instruction to RegSP map. */
 } ASMState;
 
+#ifdef LUA_USE_ASSERT
+#define lj_assertA(c, ...)      lj_assertG_(J2G(as->J), (c), __VA_ARGS__)
+#else
+#define lj_assertA(c, ...)      ((void)as)
+#endif
+
 #define IR(ref)                 (&as->ir[(ref)])
 
 #define ASMREF_TMP1             REF_TRUE        /* Temp. register. */
@@ -128,9 +137,8 @@ static LJ_AINLINE void checkmclim(ASMState *as)
 #ifdef LUA_USE_ASSERT
   if (as->mcp + MCLIM_REDZONE < as->mcp_prev) {
     IRIns *ir = IR(as->curins+1);
-    fprintf(stderr, "RED ZONE OVERFLOW: %p IR %04d  %02d %04d %04d\n", as->mcp,
-            as->curins+1-REF_BIAS, ir->o, ir->op1-REF_BIAS, ir->op2-REF_BIAS);
-    lua_assert(0);
+    lj_assertA(0, "red zone overflow: %p IR %04d  %02d %04d %04d\n", as->mcp,
+      as->curins+1-REF_BIAS, ir->o, ir->op1-REF_BIAS, ir->op2-REF_BIAS);
   }
 #endif
   if (LJ_UNLIKELY(as->mcp < as->mclim)) asm_mclimit(as);
@@ -144,7 +152,7 @@ static LJ_AINLINE void checkmclim(ASMState *as)
 #define ra_krefreg(ref)         ((Reg)(RID_MIN_KREF + (Reg)(ref)))
 #define ra_krefk(as, ref)       (as->krefk[(ref)])
 
-static LJ_AINLINE void ra_setkref(ASMState *as, Reg r, int32_t k)
+static LJ_AINLINE void ra_setkref(ASMState *as, Reg r, intptr_t k)
 {
   IRRef ref = (IRRef)(r - RID_MIN_KREF);
   as->krefk[ref] = k;
@@ -171,6 +179,8 @@ IRFLDEF(FLOFS)
 #include "lj_emit_x86.h"
 #elif LJ_TARGET_ARM
 #include "lj_emit_arm.h"
+#elif LJ_TARGET_ARM64
+#include "lj_emit_arm64.h"
 #elif LJ_TARGET_PPC
 #include "lj_emit_ppc.h"
 #elif LJ_TARGET_MIPS
@@ -179,6 +189,12 @@ IRFLDEF(FLOFS)
 #error "Missing instruction emitter for target CPU"
 #endif
 
+/* Generic load/store of register from/to stack slot. */
+#define emit_spload(as, ir, r, ofs) \
+  emit_loadofs(as, ir, (r), RID_SP, (ofs))
+#define emit_spstore(as, ir, r, ofs) \
+  emit_storeofs(as, ir, (r), RID_SP, (ofs))
+
 /* -- Register allocator debugging ---------------------------------------- */
 
 /* #define LUAJIT_DEBUG_RA */
@@ -236,7 +252,7 @@ static void ra_dprintf(ASMState *as, const char *fmt, ...)
           *p++ = *q >= 'A' && *q <= 'Z' ? *q + 0x20 : *q;
       } else {
         *p++ = '?';
-        lua_assert(0);
+        lj_assertA(0, "bad register %d for debug format \"%s\"", r, fmt);
       }
     } else if (e[1] == 'f' || e[1] == 'i') {
       IRRef ref;
@@ -254,7 +270,7 @@ static void ra_dprintf(ASMState *as, const char *fmt, ...)
     } else if (e[1] == 'x') {
       p += sprintf(p, "%08x", va_arg(argp, int32_t));
     } else {
-      lua_assert(0);
+      lj_assertA(0, "bad debug format code");
     }
     fmt = e+2;
   }
@@ -313,37 +329,51 @@ static Reg ra_rematk(ASMState *as, IRRef ref)
   Reg r;
   if (ra_iskref(ref)) {
     r = ra_krefreg(ref);
-    lua_assert(!rset_test(as->freeset, r));
+    lj_assertA(!rset_test(as->freeset, r), "rematk of free reg %d", r);
     ra_free(as, r);
     ra_modified(as, r);
+#if LJ_64
+    emit_loadu64(as, r, ra_krefk(as, ref));
+#else
     emit_loadi(as, r, ra_krefk(as, ref));
+#endif
     return r;
   }
   ir = IR(ref);
   r = ir->r;
-  lua_assert(ra_hasreg(r) && !ra_hasspill(ir->s));
+  lj_assertA(ra_hasreg(r), "rematk of K%03d has no reg", REF_BIAS - ref);
+  lj_assertA(!ra_hasspill(ir->s),
+             "rematk of K%03d has spill slot [%x]", REF_BIAS - ref, ir->s);
   ra_free(as, r);
   ra_modified(as, r);
   ir->r = RID_INIT;  /* Do not keep any hint. */
   RA_DBGX((as, "remat     $i $r", ir, r));
-#if !LJ_SOFTFP
+#if !LJ_SOFTFP32
   if (ir->o == IR_KNUM) {
-    emit_loadn(as, r, ir_knum(ir));
+    emit_loadk64(as, r, ir);
   } else
 #endif
   if (emit_canremat(REF_BASE) && ir->o == IR_BASE) {
     ra_sethint(ir->r, RID_BASE);  /* Restore BASE register hint. */
     emit_getgl(as, r, jit_base);
   } else if (emit_canremat(ASMREF_L) && ir->o == IR_KPRI) {
-    lua_assert(irt_isnil(ir->t));  /* REF_NIL stores ASMREF_L register. */
-    emit_getgl(as, r, jit_L);
+    /* REF_NIL stores ASMREF_L register. */
+    lj_assertA(irt_isnil(ir->t), "rematk of bad ASMREF_L");
+    emit_getgl(as, r, cur_L);
 #if LJ_64
   } else if (ir->o == IR_KINT64) {
     emit_loadu64(as, r, ir_kint64(ir)->u64);
+#if LJ_GC64
+  } else if (ir->o == IR_KGC) {
+    emit_loadu64(as, r, (uintptr_t)ir_kgc(ir));
+  } else if (ir->o == IR_KPTR || ir->o == IR_KKPTR) {
+    emit_loadu64(as, r, (uintptr_t)ir_kptr(ir));
+#endif
 #endif
   } else {
-    lua_assert(ir->o == IR_KINT || ir->o == IR_KGC ||
-               ir->o == IR_KPTR || ir->o == IR_KKPTR || ir->o == IR_KNULL);
+    lj_assertA(ir->o == IR_KINT || ir->o == IR_KGC ||
+               ir->o == IR_KPTR || ir->o == IR_KKPTR || ir->o == IR_KNULL,
+               "rematk of bad IR op %d", ir->o);
     emit_loadi(as, r, ir->i);
   }
   return r;
@@ -353,7 +383,8 @@ static Reg ra_rematk(ASMState *as, IRRef ref)
 static int32_t ra_spill(ASMState *as, IRIns *ir)
 {
   int32_t slot = ir->s;
-  lua_assert(ir >= as->ir + REF_TRUE);
+  lj_assertA(ir >= as->ir + REF_TRUE,
+             "spill of K%03d", REF_BIAS - (int)(ir - as->ir));
   if (!ra_hasspill(slot)) {
     if (irt_is64(ir->t)) {
       slot = as->evenspill;
@@ -378,7 +409,9 @@ static Reg ra_releasetmp(ASMState *as, IRRef ref)
 {
   IRIns *ir = IR(ref);
   Reg r = ir->r;
-  lua_assert(ra_hasreg(r) && !ra_hasspill(ir->s));
+  lj_assertA(ra_hasreg(r), "release of TMP%d has no reg", ref-ASMREF_TMP1+1);
+  lj_assertA(!ra_hasspill(ir->s),
+             "release of TMP%d has spill slot [%x]", ref-ASMREF_TMP1+1, ir->s);
   ra_free(as, r);
   ra_modified(as, r);
   ir->r = RID_INIT;
@@ -394,7 +427,7 @@ static Reg ra_restore(ASMState *as, IRRef ref)
     IRIns *ir = IR(ref);
     int32_t ofs = ra_spill(as, ir);  /* Force a spill slot. */
     Reg r = ir->r;
-    lua_assert(ra_hasreg(r));
+    lj_assertA(ra_hasreg(r), "restore of IR %04d has no reg", ref - REF_BIAS);
     ra_sethint(ir->r, r);  /* Keep hint. */
     ra_free(as, r);
     if (!rset_test(as->weakset, r)) {  /* Only restore non-weak references. */
@@ -423,14 +456,15 @@ static Reg ra_evict(ASMState *as, RegSet allow)
 {
   IRRef ref;
   RegCost cost = ~(RegCost)0;
-  lua_assert(allow != RSET_EMPTY);
+  lj_assertA(allow != RSET_EMPTY, "evict from empty set");
   if (RID_NUM_FPR == 0 || allow < RID2RSET(RID_MAX_GPR)) {
     GPRDEF(MINCOST)
   } else {
     FPRDEF(MINCOST)
   }
   ref = regcost_ref(cost);
-  lua_assert(ra_iskref(ref) || (ref >= as->T->nk && ref < as->T->nins));
+  lj_assertA(ra_iskref(ref) || (ref >= as->T->nk && ref < as->T->nins),
+             "evict of out-of-range IR %04d", ref - REF_BIAS);
   /* Preferably pick any weak ref instead of a non-weak, non-const ref. */
   if (!irref_isk(ref) && (as->weakset & allow)) {
     IRIns *ir = IR(ref);
@@ -512,7 +546,7 @@ static void ra_evictk(ASMState *as)
 
 #ifdef RID_NUM_KREF
 /* Allocate a register for a constant. */
-static Reg ra_allock(ASMState *as, int32_t k, RegSet allow)
+static Reg ra_allock(ASMState *as, intptr_t k, RegSet allow)
 {
   /* First try to find a register which already holds the same constant. */
   RegSet pick, work = ~as->freeset & RSET_GPR;
@@ -521,9 +555,31 @@ static Reg ra_allock(ASMState *as, int32_t k, RegSet allow)
     IRRef ref;
     r = rset_pickbot(work);
     ref = regcost_ref(as->cost[r]);
+#if LJ_64
+    if (ref < ASMREF_L) {
+      if (ra_iskref(ref)) {
+        if (k == ra_krefk(as, ref))
+          return r;
+      } else {
+        IRIns *ir = IR(ref);
+        if ((ir->o == IR_KINT64 && k == (int64_t)ir_kint64(ir)->u64) ||
+#if LJ_GC64
+            (ir->o == IR_KINT && k == ir->i) ||
+            (ir->o == IR_KGC && k == (intptr_t)ir_kgc(ir)) ||
+            ((ir->o == IR_KPTR || ir->o == IR_KKPTR) &&
+             k == (intptr_t)ir_kptr(ir))
+#else
+            (ir->o != IR_KINT64 && k == ir->i)
+#endif
+           )
+          return r;
+      }
+    }
+#else
     if (ref < ASMREF_L &&
         k == (ra_iskref(ref) ? ra_krefk(as, ref) : IR(ref)->i))
       return r;
+#endif
     rset_clear(work, r);
   }
   pick = as->freeset & allow;
@@ -543,7 +599,7 @@ static Reg ra_allock(ASMState *as, int32_t k, RegSet allow)
 }
 
 /* Allocate a specific register for a constant. */
-static void ra_allockreg(ASMState *as, int32_t k, Reg r)
+static void ra_allockreg(ASMState *as, intptr_t k, Reg r)
 {
   Reg kr = ra_allock(as, k, RID2RSET(r));
   if (kr != r) {
@@ -566,7 +622,8 @@ static Reg ra_allocref(ASMState *as, IRRef ref, RegSet allow)
   IRIns *ir = IR(ref);
   RegSet pick = as->freeset & allow;
   Reg r;
-  lua_assert(ra_noreg(ir->r));
+  lj_assertA(ra_noreg(ir->r),
+             "IR %04d already has reg %d", ref - REF_BIAS, ir->r);
   if (pick) {
     /* First check register hint from propagation or PHI. */
     if (ra_hashint(ir->r)) {
@@ -613,15 +670,27 @@ static Reg ra_alloc1(ASMState *as, IRRef ref, RegSet allow)
   return r;
 }
 
+/* Add a register rename to the IR. */
+static void ra_addrename(ASMState *as, Reg down, IRRef ref, SnapNo snapno)
+{
+  IRRef ren;
+  lj_ir_set(as->J, IRT(IR_RENAME, IRT_NIL), ref, snapno);
+  ren = tref_ref(lj_ir_emit(as->J));
+  as->J->cur.ir[ren].r = (uint8_t)down;
+  as->J->cur.ir[ren].s = SPS_NONE;
+}
+
 /* Rename register allocation and emit move. */
 static void ra_rename(ASMState *as, Reg down, Reg up)
 {
-  IRRef ren, ref = regcost_ref(as->cost[up] = as->cost[down]);
+  IRRef ref = regcost_ref(as->cost[up] = as->cost[down]);
   IRIns *ir = IR(ref);
   ir->r = (uint8_t)up;
   as->cost[down] = 0;
-  lua_assert((down < RID_MAX_GPR) == (up < RID_MAX_GPR));
-  lua_assert(!rset_test(as->freeset, down) && rset_test(as->freeset, up));
+  lj_assertA((down < RID_MAX_GPR) == (up < RID_MAX_GPR),
+             "rename between GPR/FPR %d and %d", down, up);
+  lj_assertA(!rset_test(as->freeset, down), "rename from free reg %d", down);
+  lj_assertA(rset_test(as->freeset, up), "rename to non-free reg %d", up);
   ra_free(as, down);  /* 'down' is free ... */
   ra_modified(as, down);
   rset_clear(as->freeset, up);  /* ... and 'up' is now allocated. */
@@ -629,11 +698,14 @@ static void ra_rename(ASMState *as, Reg down, Reg up)
   RA_DBGX((as, "rename    $f $r $r", regcost_ref(as->cost[up]), down, up));
   emit_movrr(as, ir, down, up);  /* Backwards codegen needs inverse move. */
   if (!ra_hasspill(IR(ref)->s)) {  /* Add the rename to the IR. */
-    lj_ir_set(as->J, IRT(IR_RENAME, IRT_NIL), ref, as->snapno);
-    ren = tref_ref(lj_ir_emit(as->J));
-    as->ir = as->T->ir;  /* The IR may have been reallocated. */
-    IR(ren)->r = (uint8_t)down;
-    IR(ren)->s = SPS_NONE;
+    /*
+    ** The rename is effective at the subsequent (already emitted) exit
+    ** branch. This is for the current snapshot (as->snapno). Except if we
+    ** haven't yet allocated any refs for the snapshot (as->snapalloc == 1),
+    ** then it belongs to the next snapshot.
+    ** See also the discussion at asm_snap_checkrename().
+    */
+    ra_addrename(as, down, ref, as->snapno + as->snapalloc);
   }
 }
 
@@ -666,7 +738,7 @@ static void ra_destreg(ASMState *as, IRIns *ir, Reg r)
 {
   Reg dest = ra_dest(as, ir, RID2RSET(r));
   if (dest != r) {
-    lua_assert(rset_test(as->freeset, r));
+    lj_assertA(rset_test(as->freeset, r), "dest reg %d is not free", r);
     ra_modified(as, r);
     emit_movrr(as, ir, dest, r);
   }
@@ -683,20 +755,25 @@ static void ra_left(ASMState *as, Reg dest, IRRef lref)
   if (ra_noreg(left)) {
     if (irref_isk(lref)) {
       if (ir->o == IR_KNUM) {
-        cTValue *tv = ir_knum(ir);
         /* FP remat needs a load except for +0. Still better than eviction. */
-        if (tvispzero(tv) || !(as->freeset & RSET_FPR)) {
-          emit_loadn(as, dest, tv);
+        if (tvispzero(ir_knum(ir)) || !(as->freeset & RSET_FPR)) {
+          emit_loadk64(as, dest, ir);
           return;
         }
 #if LJ_64
       } else if (ir->o == IR_KINT64) {
-        emit_loadu64(as, dest, ir_kint64(ir)->u64);
+        emit_loadk64(as, dest, ir);
+        return;
+#if LJ_GC64
+      } else if (ir->o == IR_KGC || ir->o == IR_KPTR || ir->o == IR_KKPTR) {
+        emit_loadk64(as, dest, ir);
         return;
 #endif
-      } else {
-        lua_assert(ir->o == IR_KINT || ir->o == IR_KGC ||
-                   ir->o == IR_KPTR || ir->o == IR_KKPTR || ir->o == IR_KNULL);
+#endif
+      } else if (ir->o != IR_KPRI) {
+        lj_assertA(ir->o == IR_KINT || ir->o == IR_KGC ||
+                   ir->o == IR_KPTR || ir->o == IR_KKPTR || ir->o == IR_KNULL,
+                   "K%03d has bad IR op %d", REF_BIAS - lref, ir->o);
         emit_loadi(as, dest, ir->i);
         return;
       }
@@ -741,11 +818,11 @@ static void ra_leftov(ASMState *as, Reg dest, IRRef lref)
 }
 #endif
 
-#if !LJ_64
 /* Force a RID_RETLO/RID_RETHI destination register pair (marked as free). */
 static void ra_destpair(ASMState *as, IRIns *ir)
 {
   Reg destlo = ir->r, desthi = (ir+1)->r;
+  IRIns *irx = (LJ_64 && !irt_is64(ir->t)) ? ir+1 : ir;
   /* First spill unrelated refs blocking the destination registers. */
   if (!rset_test(as->freeset, RID_RETLO) &&
       destlo != RID_RETLO && desthi != RID_RETLO)
@@ -769,29 +846,28 @@ static void ra_destpair(ASMState *as, IRIns *ir)
   /* Check for conflicts and shuffle the registers as needed. */
   if (destlo == RID_RETHI) {
     if (desthi == RID_RETLO) {
-#if LJ_TARGET_X86
-      *--as->mcp = XI_XCHGa + RID_RETHI;
+#if LJ_TARGET_X86ORX64
+      *--as->mcp = REX_64IR(irx, XI_XCHGa + RID_RETHI);
 #else
-      emit_movrr(as, ir, RID_RETHI, RID_TMP);
-      emit_movrr(as, ir, RID_RETLO, RID_RETHI);
-      emit_movrr(as, ir, RID_TMP, RID_RETLO);
+      emit_movrr(as, irx, RID_RETHI, RID_TMP);
+      emit_movrr(as, irx, RID_RETLO, RID_RETHI);
+      emit_movrr(as, irx, RID_TMP, RID_RETLO);
 #endif
     } else {
-      emit_movrr(as, ir, RID_RETHI, RID_RETLO);
-      if (desthi != RID_RETHI) emit_movrr(as, ir, desthi, RID_RETHI);
+      emit_movrr(as, irx, RID_RETHI, RID_RETLO);
+      if (desthi != RID_RETHI) emit_movrr(as, irx, desthi, RID_RETHI);
     }
   } else if (desthi == RID_RETLO) {
-    emit_movrr(as, ir, RID_RETLO, RID_RETHI);
-    if (destlo != RID_RETLO) emit_movrr(as, ir, destlo, RID_RETLO);
+    emit_movrr(as, irx, RID_RETLO, RID_RETHI);
+    if (destlo != RID_RETLO) emit_movrr(as, irx, destlo, RID_RETLO);
   } else {
-    if (desthi != RID_RETHI) emit_movrr(as, ir, desthi, RID_RETHI);
-    if (destlo != RID_RETLO) emit_movrr(as, ir, destlo, RID_RETLO);
+    if (desthi != RID_RETHI) emit_movrr(as, irx, desthi, RID_RETHI);
+    if (destlo != RID_RETLO) emit_movrr(as, irx, destlo, RID_RETLO);
   }
   /* Restore spill slots (if any). */
   if (ra_hasspill((ir+1)->s)) ra_save(as, ir+1, RID_RETHI);
   if (ra_hasspill(ir->s)) ra_save(as, ir, RID_RETLO);
 }
-#endif
 
 /* -- Snapshot handling --------- ----------------------------------------- */
 
@@ -841,11 +917,14 @@ static void asm_snap_alloc1(ASMState *as, IRRef ref)
 #endif
       {  /* Allocate stored values for TNEW, TDUP and CNEW. */
         IRIns *irs;
-        lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP || ir->o == IR_CNEW);
+        lj_assertA(ir->o == IR_TNEW || ir->o == IR_TDUP || ir->o == IR_CNEW,
+                   "sink of IR %04d has bad op %d", ref - REF_BIAS, ir->o);
         for (irs = IR(as->snapref-1); irs > ir; irs--)
           if (irs->r == RID_SINK && asm_sunk_store(as, ir, irs)) {
-            lua_assert(irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
-                       irs->o == IR_FSTORE || irs->o == IR_XSTORE);
+            lj_assertA(irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
+                       irs->o == IR_FSTORE || irs->o == IR_XSTORE,
+                       "sunk store IR %04d has bad op %d",
+                       (int)(irs - as->ir) - REF_BIAS, irs->o);
             asm_snap_alloc1(as, irs->op2);
             if (LJ_32 && (irs+1)->o == IR_HIOP)
               asm_snap_alloc1(as, (irs+1)->op2);
@@ -881,9 +960,9 @@ static void asm_snap_alloc1(ASMState *as, IRRef ref)
 }
 
 /* Allocate refs escaping to a snapshot. */
-static void asm_snap_alloc(ASMState *as)
+static void asm_snap_alloc(ASMState *as, int snapno)
 {
-  SnapShot *snap = &as->T->snap[as->snapno];
+  SnapShot *snap = &as->T->snap[snapno];
   SnapEntry *map = &as->T->snapmap[snap->mapofs];
   MSize n, nent = snap->nent;
   as->snapfilt1 = as->snapfilt2 = 0;
@@ -893,7 +972,9 @@ static void asm_snap_alloc(ASMState *as)
     if (!irref_isk(ref)) {
       asm_snap_alloc1(as, ref);
       if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM)) {
-        lua_assert(irt_type(IR(ref+1)->t) == IRT_SOFTFP);
+        lj_assertA(irt_type(IR(ref+1)->t) == IRT_SOFTFP,
+                   "snap %d[%d] points to bad SOFTFP IR %04d",
+                   snapno, n, ref - REF_BIAS);
         asm_snap_alloc1(as, ref+1);
       }
     }
@@ -919,67 +1000,55 @@ static int asm_snap_checkrename(ASMState *as, IRRef ren)
   return 0;  /* Not found. */
 }
 
-/* Prepare snapshot for next guard instruction. */
+/* Prepare snapshot for next guard or throwing instruction. */
 static void asm_snap_prep(ASMState *as)
 {
-  if (as->curins < as->snapref) {
-    do {
-      if (as->snapno == 0) return;  /* Called by sunk stores before snap #0. */
-      as->snapno--;
-      as->snapref = as->T->snap[as->snapno].ref;
-    } while (as->curins < as->snapref);
-    asm_snap_alloc(as);
+  if (as->snapalloc) {
+    /* Alloc on first invocation for each snapshot. */
+    as->snapalloc = 0;
+    asm_snap_alloc(as, as->snapno);
     as->snaprename = as->T->nins;
   } else {
-    /* Process any renames above the highwater mark. */
+    /* Check any renames above the highwater mark. */
     for (; as->snaprename < as->T->nins; as->snaprename++) {
-      IRIns *ir = IR(as->snaprename);
+      IRIns *ir = &as->T->ir[as->snaprename];
       if (asm_snap_checkrename(as, ir->op1))
         ir->op2 = REF_BIAS-1;  /* Kill rename. */
     }
   }
 }
 
-/* -- Miscellaneous helpers ----------------------------------------------- */
-
-/* Collect arguments from CALL* and CARG instructions. */
-static void asm_collectargs(ASMState *as, IRIns *ir,
-                            const CCallInfo *ci, IRRef *args)
+/* Move to previous snapshot when we cross the current snapshot ref. */
+static void asm_snap_prev(ASMState *as)
 {
-  uint32_t n = CCI_NARGS(ci);
-  lua_assert(n <= CCI_NARGS_MAX*2);  /* Account for split args. */
-  if ((ci->flags & CCI_L)) { *args++ = ASMREF_L; n--; }
-  while (n-- > 1) {
-    ir = IR(ir->op1);
-    lua_assert(ir->o == IR_CARG);
-    args[n] = ir->op2 == REF_NIL ? 0 : ir->op2;
+  if (as->curins < as->snapref) {
+    uintptr_t ofs = (uintptr_t)(as->mctoporig - as->mcp);
+    if (ofs >= 0x10000) lj_trace_err(as->J, LJ_TRERR_MCODEOV);
+    do {
+      if (as->snapno == 0) return;
+      as->snapno--;
+      as->snapref = as->T->snap[as->snapno].ref;
+      as->T->snap[as->snapno].mcofs = (uint16_t)ofs;  /* Remember mcode ofs. */
+    } while (as->curins < as->snapref);  /* May have no ins inbetween. */
+    as->snapalloc = 1;
   }
-  args[0] = ir->op1 == REF_NIL ? 0 : ir->op1;
-  lua_assert(IR(ir->op1)->o != IR_CARG);
 }
 
-/* Reconstruct CCallInfo flags for CALLX*. */
-static uint32_t asm_callx_flags(ASMState *as, IRIns *ir)
+/* Fixup snapshot mcode offsetst. */
+static void asm_snap_fixup_mcofs(ASMState *as)
 {
-  uint32_t nargs = 0;
-  if (ir->op1 != REF_NIL) {  /* Count number of arguments first. */
-    IRIns *ira = IR(ir->op1);
-    nargs++;
-    while (ira->o == IR_CARG) { nargs++; ira = IR(ira->op1); }
+  uint32_t sz = (uint32_t)(as->mctoporig - as->mcp);
+  SnapShot *snap = as->T->snap;
+  SnapNo i;
+  for (i = as->T->nsnap-1; i > 0; i--) {
+    /* Compute offset from mcode start and store in correct snapshot. */
+    snap[i].mcofs = (uint16_t)(sz - snap[i-1].mcofs);
   }
-#if LJ_HASFFI
-  if (IR(ir->op2)->o == IR_CARG) {  /* Copy calling convention info. */
-    CTypeID id = (CTypeID)IR(IR(ir->op2)->op2)->i;
-    CType *ct = ctype_get(ctype_ctsG(J2G(as->J)), id);
-    nargs |= ((ct->info & CTF_VARARG) ? CCI_VARARG : 0);
-#if LJ_TARGET_X86
-    nargs |= (ctype_cconv(ct->info) << CCI_CC_SHIFT);
-#endif
-  }
-#endif
-  return (nargs | (ir->t.irt << CCI_OTSHIFT));
+  snap[0].mcofs = 0;
 }
 
+/* -- Miscellaneous helpers ----------------------------------------------- */
+
 /* Calculate stack adjustment. */
 static int32_t asm_stack_adjust(ASMState *as)
 {
@@ -989,21 +1058,26 @@ static int32_t asm_stack_adjust(ASMState *as)
 }
 
 /* Must match with hash*() in lj_tab.c. */
-static uint32_t ir_khash(IRIns *ir)
+static uint32_t ir_khash(ASMState *as, IRIns *ir)
 {
   uint32_t lo, hi;
+  UNUSED(as);
   if (irt_isstr(ir->t)) {
-    return ir_kstr(ir)->hash;
+    return ir_kstr(ir)->sid;
   } else if (irt_isnum(ir->t)) {
     lo = ir_knum(ir)->u32.lo;
     hi = ir_knum(ir)->u32.hi << 1;
   } else if (irt_ispri(ir->t)) {
-    lua_assert(!irt_isnil(ir->t));
+    lj_assertA(!irt_isnil(ir->t), "hash of nil key");
     return irt_type(ir->t)-IRT_FALSE;
   } else {
-    lua_assert(irt_isgcv(ir->t));
+    lj_assertA(irt_isgcv(ir->t), "hash of bad IR type %d", irt_type(ir->t));
     lo = u32ptr(ir_kgc(ir));
+#if LJ_GC64
+    hi = (uint32_t)(u64ptr(ir_kgc(ir)) >> 32) | (irt_toitype(ir->t) << 15);
+#else
     hi = lo + HASH_BIAS;
+#endif
   }
   return hashrot(lo, hi);
 }
@@ -1017,6 +1091,7 @@ static void asm_snew(ASMState *as, IRIns *ir)
 {
   const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_str_new];
   IRRef args[3];
+  asm_snap_prep(as);
   args[0] = ASMREF_L;  /* lua_State *L    */
   args[1] = ir->op1;   /* const char *str */
   args[2] = ir->op2;   /* size_t len      */
@@ -1029,6 +1104,7 @@ static void asm_tnew(ASMState *as, IRIns *ir)
 {
   const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_tab_new1];
   IRRef args[2];
+  asm_snap_prep(as);
   args[0] = ASMREF_L;     /* lua_State *L    */
   args[1] = ASMREF_TMP1;  /* uint32_t ahsize */
   as->gcsteps++;
@@ -1041,6 +1117,7 @@ static void asm_tdup(ASMState *as, IRIns *ir)
 {
   const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_tab_dup];
   IRRef args[2];
+  asm_snap_prep(as);
   args[0] = ASMREF_L;  /* lua_State *L    */
   args[1] = ir->op1;   /* const GCtab *kt */
   as->gcsteps++;
@@ -1064,6 +1141,260 @@ static void asm_gcstep(ASMState *as, IRIns *ir)
   as->gcsteps = 0x80000000;  /* Prevent implicit GC check further up. */
 }
 
+/* -- Buffer operations --------------------------------------------------- */
+
+static void asm_tvptr(ASMState *as, Reg dest, IRRef ref, MSize mode);
+#if LJ_HASBUFFER
+static void asm_bufhdr_write(ASMState *as, Reg sb);
+#endif
+
+static void asm_bufhdr(ASMState *as, IRIns *ir)
+{
+  Reg sb = ra_dest(as, ir, RSET_GPR);
+  switch (ir->op2) {
+  case IRBUFHDR_RESET: {
+    Reg tmp = ra_scratch(as, rset_exclude(RSET_GPR, sb));
+    IRIns irbp;
+    irbp.ot = IRT(0, IRT_PTR);  /* Buffer data pointer type. */
+    emit_storeofs(as, &irbp, tmp, sb, offsetof(SBuf, w));
+    emit_loadofs(as, &irbp, tmp, sb, offsetof(SBuf, b));
+    break;
+    }
+  case IRBUFHDR_APPEND: {
+    /* Rematerialize const buffer pointer instead of likely spill. */
+    IRIns *irp = IR(ir->op1);
+    if (!(ra_hasreg(irp->r) || irp == ir-1 ||
+          (irp == ir-2 && !ra_used(ir-1)))) {
+      while (!(irp->o == IR_BUFHDR && irp->op2 == IRBUFHDR_RESET))
+        irp = IR(irp->op1);
+      if (irref_isk(irp->op1)) {
+        ra_weak(as, ra_allocref(as, ir->op1, RSET_GPR));
+        ir = irp;
+      }
+    }
+    break;
+    }
+#if LJ_HASBUFFER
+  case IRBUFHDR_WRITE:
+    asm_bufhdr_write(as, sb);
+    break;
+#endif
+  default: lj_assertA(0, "bad BUFHDR op2 %d", ir->op2); break;
+  }
+#if LJ_TARGET_X86ORX64
+  ra_left(as, sb, ir->op1);
+#else
+  ra_leftov(as, sb, ir->op1);
+#endif
+}
+
+static void asm_bufput(ASMState *as, IRIns *ir)
+{
+  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_buf_putstr];
+  IRRef args[3];
+  IRIns *irs;
+  int kchar = -129;
+  args[0] = ir->op1;  /* SBuf * */
+  args[1] = ir->op2;  /* GCstr * */
+  irs = IR(ir->op2);
+  lj_assertA(irt_isstr(irs->t),
+             "BUFPUT of non-string IR %04d", ir->op2 - REF_BIAS);
+  if (irs->o == IR_KGC) {
+    GCstr *s = ir_kstr(irs);
+    if (s->len == 1) {  /* Optimize put of single-char string constant. */
+      kchar = (int8_t)strdata(s)[0];  /* Signed! */
+      args[1] = ASMREF_TMP1;  /* int, truncated to char */
+      ci = &lj_ir_callinfo[IRCALL_lj_buf_putchar];
+    }
+  } else if (mayfuse(as, ir->op2) && ra_noreg(irs->r)) {
+    if (irs->o == IR_TOSTR) {  /* Fuse number to string conversions. */
+      if (irs->op2 == IRTOSTR_NUM) {
+        args[1] = ASMREF_TMP1;  /* TValue * */
+        ci = &lj_ir_callinfo[IRCALL_lj_strfmt_putnum];
+      } else {
+        lj_assertA(irt_isinteger(IR(irs->op1)->t),
+                   "TOSTR of non-numeric IR %04d", irs->op1);
+        args[1] = irs->op1;  /* int */
+        if (irs->op2 == IRTOSTR_INT)
+          ci = &lj_ir_callinfo[IRCALL_lj_strfmt_putint];
+        else
+          ci = &lj_ir_callinfo[IRCALL_lj_buf_putchar];
+      }
+    } else if (irs->o == IR_SNEW) {  /* Fuse string allocation. */
+      args[1] = irs->op1;  /* const void * */
+      args[2] = irs->op2;  /* MSize */
+      ci = &lj_ir_callinfo[IRCALL_lj_buf_putmem];
+    }
+  }
+  asm_setupresult(as, ir, ci);  /* SBuf * */
+  asm_gencall(as, ci, args);
+  if (args[1] == ASMREF_TMP1) {
+    Reg tmp = ra_releasetmp(as, ASMREF_TMP1);
+    if (kchar == -129)
+      asm_tvptr(as, tmp, irs->op1, IRTMPREF_IN1);
+    else
+      ra_allockreg(as, kchar, tmp);
+  }
+}
+
+static void asm_bufstr(ASMState *as, IRIns *ir)
+{
+  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_buf_tostr];
+  IRRef args[1];
+  args[0] = ir->op1;  /* SBuf *sb */
+  as->gcsteps++;
+  asm_setupresult(as, ir, ci);  /* GCstr * */
+  asm_gencall(as, ci, args);
+}
+
+/* -- Type conversions ---------------------------------------------------- */
+
+static void asm_tostr(ASMState *as, IRIns *ir)
+{
+  const CCallInfo *ci;
+  IRRef args[2];
+  asm_snap_prep(as);
+  args[0] = ASMREF_L;
+  as->gcsteps++;
+  if (ir->op2 == IRTOSTR_NUM) {
+    args[1] = ASMREF_TMP1;  /* cTValue * */
+    ci = &lj_ir_callinfo[IRCALL_lj_strfmt_num];
+  } else {
+    args[1] = ir->op1;  /* int32_t k */
+    if (ir->op2 == IRTOSTR_INT)
+      ci = &lj_ir_callinfo[IRCALL_lj_strfmt_int];
+    else
+      ci = &lj_ir_callinfo[IRCALL_lj_strfmt_char];
+  }
+  asm_setupresult(as, ir, ci);  /* GCstr * */
+  asm_gencall(as, ci, args);
+  if (ir->op2 == IRTOSTR_NUM)
+    asm_tvptr(as, ra_releasetmp(as, ASMREF_TMP1), ir->op1, IRTMPREF_IN1);
+}
+
+#if LJ_32 && LJ_HASFFI && !LJ_SOFTFP && !LJ_TARGET_X86
+static void asm_conv64(ASMState *as, IRIns *ir)
+{
+  IRType st = (IRType)((ir-1)->op2 & IRCONV_SRCMASK);
+  IRType dt = (((ir-1)->op2 & IRCONV_DSTMASK) >> IRCONV_DSH);
+  IRCallID id;
+  IRRef args[2];
+  lj_assertA((ir-1)->o == IR_CONV && ir->o == IR_HIOP,
+             "not a CONV/HIOP pair at IR %04d", (int)(ir - as->ir) - REF_BIAS);
+  args[LJ_BE] = (ir-1)->op1;
+  args[LJ_LE] = ir->op1;
+  if (st == IRT_NUM || st == IRT_FLOAT) {
+    id = IRCALL_fp64_d2l + ((st == IRT_FLOAT) ? 2 : 0) + (dt - IRT_I64);
+    ir--;
+  } else {
+    id = IRCALL_fp64_l2d + ((dt == IRT_FLOAT) ? 2 : 0) + (st - IRT_I64);
+  }
+  {
+#if LJ_TARGET_ARM && !LJ_ABI_SOFTFP
+    CCallInfo cim = lj_ir_callinfo[id], *ci = &cim;
+    cim.flags |= CCI_VARARG;  /* These calls don't use the hard-float ABI! */
+#else
+    const CCallInfo *ci = &lj_ir_callinfo[id];
+#endif
+    asm_setupresult(as, ir, ci);
+    asm_gencall(as, ci, args);
+  }
+}
+#endif
+
+/* -- Memory references --------------------------------------------------- */
+
+static void asm_newref(ASMState *as, IRIns *ir)
+{
+  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_tab_newkey];
+  IRRef args[3];
+  if (ir->r == RID_SINK)
+    return;
+  asm_snap_prep(as);
+  args[0] = ASMREF_L;     /* lua_State *L */
+  args[1] = ir->op1;      /* GCtab *t     */
+  args[2] = ASMREF_TMP1;  /* cTValue *key */
+  asm_setupresult(as, ir, ci);  /* TValue * */
+  asm_gencall(as, ci, args);
+  asm_tvptr(as, ra_releasetmp(as, ASMREF_TMP1), ir->op2, IRTMPREF_IN1);
+}
+
+static void asm_tmpref(ASMState *as, IRIns *ir)
+{
+  Reg r = ra_dest(as, ir, RSET_GPR);
+  asm_tvptr(as, r, ir->op1, ir->op2);
+}
+
+static void asm_lref(ASMState *as, IRIns *ir)
+{
+  Reg r = ra_dest(as, ir, RSET_GPR);
+#if LJ_TARGET_X86ORX64
+  ra_left(as, r, ASMREF_L);
+#else
+  ra_leftov(as, r, ASMREF_L);
+#endif
+}
+
+/* -- Calls --------------------------------------------------------------- */
+
+/* Collect arguments from CALL* and CARG instructions. */
+static void asm_collectargs(ASMState *as, IRIns *ir,
+                            const CCallInfo *ci, IRRef *args)
+{
+  uint32_t n = CCI_XNARGS(ci);
+  /* Account for split args. */
+  lj_assertA(n <= CCI_NARGS_MAX*2, "too many args %d to collect", n);
+  if ((ci->flags & CCI_L)) { *args++ = ASMREF_L; n--; }
+  while (n-- > 1) {
+    ir = IR(ir->op1);
+    lj_assertA(ir->o == IR_CARG, "malformed CALL arg tree");
+    args[n] = ir->op2 == REF_NIL ? 0 : ir->op2;
+  }
+  args[0] = ir->op1 == REF_NIL ? 0 : ir->op1;
+  lj_assertA(IR(ir->op1)->o != IR_CARG, "malformed CALL arg tree");
+}
+
+/* Reconstruct CCallInfo flags for CALLX*. */
+static uint32_t asm_callx_flags(ASMState *as, IRIns *ir)
+{
+  uint32_t nargs = 0;
+  if (ir->op1 != REF_NIL) {  /* Count number of arguments first. */
+    IRIns *ira = IR(ir->op1);
+    nargs++;
+    while (ira->o == IR_CARG) { nargs++; ira = IR(ira->op1); }
+  }
+#if LJ_HASFFI
+  if (IR(ir->op2)->o == IR_CARG) {  /* Copy calling convention info. */
+    CTypeID id = (CTypeID)IR(IR(ir->op2)->op2)->i;
+    CType *ct = ctype_get(ctype_ctsG(J2G(as->J)), id);
+    nargs |= ((ct->info & CTF_VARARG) ? CCI_VARARG : 0);
+#if LJ_TARGET_X86
+    nargs |= (ctype_cconv(ct->info) << CCI_CC_SHIFT);
+#endif
+  }
+#endif
+  return (nargs | (ir->t.irt << CCI_OTSHIFT));
+}
+
+static void asm_callid(ASMState *as, IRIns *ir, IRCallID id)
+{
+  const CCallInfo *ci = &lj_ir_callinfo[id];
+  IRRef args[2];
+  args[0] = ir->op1;
+  args[1] = ir->op2;
+  asm_setupresult(as, ir, ci);
+  asm_gencall(as, ci, args);
+}
+
+static void asm_call(ASMState *as, IRIns *ir)
+{
+  IRRef args[CCI_NARGS_MAX];
+  const CCallInfo *ci = &lj_ir_callinfo[ir->op2];
+  asm_collectargs(as, ir, ci, args);
+  asm_setupresult(as, ir, ci);
+  asm_gencall(as, ci, args);
+}
+
 /* -- PHI and loop handling ----------------------------------------------- */
 
 /* Break a PHI cycle by renaming to a free register (evict if needed). */
@@ -1249,12 +1580,7 @@ static void asm_phi_fixup(ASMState *as)
       irt_clearmark(ir->t);
       /* Left PHI gained a spill slot before the loop? */
       if (ra_hasspill(ir->s)) {
-        IRRef ren;
-        lj_ir_set(as->J, IRT(IR_RENAME, IRT_NIL), lref, as->loopsnapno);
-        ren = tref_ref(lj_ir_emit(as->J));
-        as->ir = as->T->ir;  /* The IR may have been reallocated. */
-        IR(ren)->r = (uint8_t)r;
-        IR(ren)->s = SPS_NONE;
+        ra_addrename(as, r, lref, as->loopsnapno);
       }
     }
     rset_clear(work, r);
@@ -1329,6 +1655,8 @@ static void asm_loop(ASMState *as)
 #include "lj_asm_x86.h"
 #elif LJ_TARGET_ARM
 #include "lj_asm_arm.h"
+#elif LJ_TARGET_ARM64
+#include "lj_asm_arm64.h"
 #elif LJ_TARGET_PPC
 #include "lj_asm_ppc.h"
 #elif LJ_TARGET_MIPS
@@ -1337,6 +1665,204 @@ static void asm_loop(ASMState *as)
 #error "Missing assembler for target CPU"
 #endif
 
+/* -- Common instruction helpers ------------------------------------------ */
+
+#if !LJ_SOFTFP32
+#if !LJ_TARGET_X86ORX64
+#define asm_ldexp(as, ir)       asm_callid(as, ir, IRCALL_ldexp)
+#define asm_fppowi(as, ir)      asm_callid(as, ir, IRCALL_lj_vm_powi)
+#endif
+
+static void asm_pow(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isnum(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
+                                          IRCALL_lj_carith_powu64);
+  else
+#endif
+  if (irt_isnum(IR(ir->op2)->t))
+    asm_callid(as, ir, IRCALL_pow);
+  else
+    asm_fppowi(as, ir);
+}
+
+static void asm_div(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isnum(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
+                                          IRCALL_lj_carith_divu64);
+  else
+#endif
+    asm_fpdiv(as, ir);
+}
+#endif
+
+static void asm_mod(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isint(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
+                                          IRCALL_lj_carith_modu64);
+  else
+#endif
+    asm_callid(as, ir, IRCALL_lj_vm_modi);
+}
+
+static void asm_fuseequal(ASMState *as, IRIns *ir)
+{
+  /* Fuse HREF + EQ/NE. */
+  if ((ir-1)->o == IR_HREF && ir->op1 == as->curins-1) {
+    as->curins--;
+    asm_href(as, ir-1, (IROp)ir->o);
+  } else {
+    asm_equal(as, ir);
+  }
+}
+
+static void asm_alen(ASMState *as, IRIns *ir)
+{
+  asm_callid(as, ir, ir->op2 == REF_NIL ? IRCALL_lj_tab_len :
+                                          IRCALL_lj_tab_len_hint);
+}
+
+/* -- Instruction dispatch ------------------------------------------------ */
+
+/* Assemble a single instruction. */
+static void asm_ir(ASMState *as, IRIns *ir)
+{
+  switch ((IROp)ir->o) {
+  /* Miscellaneous ops. */
+  case IR_LOOP: asm_loop(as); break;
+  case IR_NOP: case IR_XBAR:
+    lj_assertA(!ra_used(ir),
+               "IR %04d not unused", (int)(ir - as->ir) - REF_BIAS);
+    break;
+  case IR_USE:
+    ra_alloc1(as, ir->op1, irt_isfp(ir->t) ? RSET_FPR : RSET_GPR); break;
+  case IR_PHI: asm_phi(as, ir); break;
+  case IR_HIOP: asm_hiop(as, ir); break;
+  case IR_GCSTEP: asm_gcstep(as, ir); break;
+  case IR_PROF: asm_prof(as, ir); break;
+
+  /* Guarded assertions. */
+  case IR_LT: case IR_GE: case IR_LE: case IR_GT:
+  case IR_ULT: case IR_UGE: case IR_ULE: case IR_UGT:
+  case IR_ABC:
+    asm_comp(as, ir);
+    break;
+  case IR_EQ: case IR_NE: asm_fuseequal(as, ir); break;
+
+  case IR_RETF: asm_retf(as, ir); break;
+
+  /* Bit ops. */
+  case IR_BNOT: asm_bnot(as, ir); break;
+  case IR_BSWAP: asm_bswap(as, ir); break;
+  case IR_BAND: asm_band(as, ir); break;
+  case IR_BOR: asm_bor(as, ir); break;
+  case IR_BXOR: asm_bxor(as, ir); break;
+  case IR_BSHL: asm_bshl(as, ir); break;
+  case IR_BSHR: asm_bshr(as, ir); break;
+  case IR_BSAR: asm_bsar(as, ir); break;
+  case IR_BROL: asm_brol(as, ir); break;
+  case IR_BROR: asm_bror(as, ir); break;
+
+  /* Arithmetic ops. */
+  case IR_ADD: asm_add(as, ir); break;
+  case IR_SUB: asm_sub(as, ir); break;
+  case IR_MUL: asm_mul(as, ir); break;
+  case IR_MOD: asm_mod(as, ir); break;
+  case IR_NEG: asm_neg(as, ir); break;
+#if LJ_SOFTFP32
+  case IR_DIV: case IR_POW: case IR_ABS:
+  case IR_LDEXP: case IR_FPMATH: case IR_TOBIT:
+    /* Unused for LJ_SOFTFP32. */
+    lj_assertA(0, "IR %04d with unused op %d",
+                  (int)(ir - as->ir) - REF_BIAS, ir->o);
+    break;
+#else
+  case IR_DIV: asm_div(as, ir); break;
+  case IR_POW: asm_pow(as, ir); break;
+  case IR_ABS: asm_abs(as, ir); break;
+  case IR_LDEXP: asm_ldexp(as, ir); break;
+  case IR_FPMATH: asm_fpmath(as, ir); break;
+  case IR_TOBIT: asm_tobit(as, ir); break;
+#endif
+  case IR_MIN: asm_min(as, ir); break;
+  case IR_MAX: asm_max(as, ir); break;
+
+  /* Overflow-checking arithmetic ops. */
+  case IR_ADDOV: asm_addov(as, ir); break;
+  case IR_SUBOV: asm_subov(as, ir); break;
+  case IR_MULOV: asm_mulov(as, ir); break;
+
+  /* Memory references. */
+  case IR_AREF: asm_aref(as, ir); break;
+  case IR_HREF: asm_href(as, ir, 0); break;
+  case IR_HREFK: asm_hrefk(as, ir); break;
+  case IR_NEWREF: asm_newref(as, ir); break;
+  case IR_UREFO: case IR_UREFC: asm_uref(as, ir); break;
+  case IR_FREF: asm_fref(as, ir); break;
+  case IR_TMPREF: asm_tmpref(as, ir); break;
+  case IR_STRREF: asm_strref(as, ir); break;
+  case IR_LREF: asm_lref(as, ir); break;
+
+  /* Loads and stores. */
+  case IR_ALOAD: case IR_HLOAD: case IR_ULOAD: case IR_VLOAD:
+    asm_ahuvload(as, ir);
+    break;
+  case IR_FLOAD: asm_fload(as, ir); break;
+  case IR_XLOAD: asm_xload(as, ir); break;
+  case IR_SLOAD: asm_sload(as, ir); break;
+  case IR_ALEN: asm_alen(as, ir); break;
+
+  case IR_ASTORE: case IR_HSTORE: case IR_USTORE: asm_ahustore(as, ir); break;
+  case IR_FSTORE: asm_fstore(as, ir); break;
+  case IR_XSTORE: asm_xstore(as, ir); break;
+
+  /* Allocations. */
+  case IR_SNEW: case IR_XSNEW: 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:
+#if LJ_HASFFI
+    asm_cnew(as, ir);
+#else
+    lj_assertA(0, "IR %04d with unused op %d",
+                  (int)(ir - as->ir) - REF_BIAS, ir->o);
+#endif
+    break;
+
+  /* Buffer operations. */
+  case IR_BUFHDR: asm_bufhdr(as, ir); break;
+  case IR_BUFPUT: asm_bufput(as, ir); break;
+  case IR_BUFSTR: asm_bufstr(as, ir); break;
+
+  /* Write barriers. */
+  case IR_TBAR: asm_tbar(as, ir); break;
+  case IR_OBAR: asm_obar(as, ir); break;
+
+  /* Type conversions. */
+  case IR_CONV: asm_conv(as, ir); break;
+  case IR_TOSTR: asm_tostr(as, ir); break;
+  case IR_STRTO: asm_strto(as, ir); break;
+
+  /* Calls. */
+  case IR_CALLA:
+    as->gcsteps++;
+    /* fallthrough */
+  case IR_CALLN: case IR_CALLL: case IR_CALLS: asm_call(as, ir); break;
+  case IR_CALLXS: asm_callx(as, ir); break;
+  case IR_CARG: break;
+
+  default:
+    setintV(&as->J->errinfo, ir->o);
+    lj_trace_err_info(as->J, LJ_TRERR_NYIIR);
+    break;
+  }
+}
+
 /* -- Head of trace ------------------------------------------------------- */
 
 /* Head of a root trace. */
@@ -1373,8 +1899,7 @@ static void asm_head_side(ASMState *as)
 
   if (as->snapno && as->topslot > as->parent->topslot) {
     /* Force snap #0 alloc to prevent register overwrite in stack check. */
-    as->snapno = 0;
-    asm_snap_alloc(as);
+    asm_snap_alloc(as, 0);
   }
   allow = asm_head_side_base(as, irp, allow);
 
@@ -1382,8 +1907,10 @@ static void asm_head_side(ASMState *as)
   for (i = as->stopins; i > REF_BASE; i--) {
     IRIns *ir = IR(i);
     RegSP rs;
-    lua_assert((ir->o == IR_SLOAD && (ir->op2 & IRSLOAD_PARENT)) ||
-               (LJ_SOFTFP && ir->o == IR_HIOP) || ir->o == IR_PVAL);
+    lj_assertA((ir->o == IR_SLOAD && (ir->op2 & IRSLOAD_PARENT)) ||
+               (LJ_SOFTFP && ir->o == IR_HIOP) || ir->o == IR_PVAL,
+               "IR %04d has bad parent op %d",
+               (int)(ir - as->ir) - REF_BIAS, ir->o);
     rs = as->parentmap[i - REF_FIRST];
     if (ra_hasreg(ir->r)) {
       rset_clear(allow, ir->r);
@@ -1535,7 +2062,7 @@ static BCReg asm_baseslot(ASMState *as, SnapShot *snap, int *gotframe)
     SnapEntry sn = map[n-1];
     if ((sn & SNAP_FRAME)) {
       *gotframe = 1;
-      return snap_slot(sn);
+      return snap_slot(sn) - LJ_FR2;
     }
   }
   return 0;
@@ -1555,19 +2082,23 @@ static void asm_tail_link(ASMState *as)
 
   if (as->T->link == 0) {
     /* Setup fixed registers for exit to interpreter. */
-    const BCIns *pc = snap_pc(as->T->snapmap[snap->mapofs + snap->nent]);
+    const BCIns *pc = snap_pc(&as->T->snapmap[snap->mapofs + snap->nent]);
     int32_t mres;
     if (bc_op(*pc) == BC_JLOOP) {  /* NYI: find a better way to do this. */
       BCIns *retpc = &traceref(as->J, bc_d(*pc))->startins;
       if (bc_isret(bc_op(*retpc)))
         pc = retpc;
     }
+#if LJ_GC64
+    emit_loadu64(as, RID_LPC, u64ptr(pc));
+#else
     ra_allockreg(as, i32ptr(J2GG(as->J)->dispatch), RID_DISPATCH);
     ra_allockreg(as, i32ptr(pc), RID_LPC);
-    mres = (int32_t)(snap->nslots - baseslot);
+#endif
+    mres = (int32_t)(snap->nslots - baseslot - LJ_FR2);
     switch (bc_op(*pc)) {
     case BC_CALLM: case BC_CALLMT:
-      mres -= (int32_t)(1 + bc_a(*pc) + bc_c(*pc)); break;
+      mres -= (int32_t)(1 + LJ_FR2 + bc_a(*pc) + bc_c(*pc)); break;
     case BC_RETM: mres -= (int32_t)(bc_a(*pc) + bc_d(*pc)); break;
     case BC_TSETM: mres -= (int32_t)bc_a(*pc); break;
     default: if (bc_op(*pc) < BC_FUNCF) mres = 0; break;
@@ -1579,6 +2110,11 @@ static void asm_tail_link(ASMState *as)
   }
   emit_addptr(as, RID_BASE, 8*(int32_t)baseslot);
 
+  if (as->J->ktrace) {  /* Patch ktrace slot with the final GCtrace pointer. */
+    setgcref(IR(as->J->ktrace)[LJ_GC64].gcr, obj2gco(as->J->curfinal));
+    IR(as->J->ktrace)->o = IR_KGC;
+  }
+
   /* Sync the interpreter state with the on-trace state. */
   asm_stack_restore(as, snap);
 
@@ -1602,22 +2138,32 @@ static void asm_setup_regsp(ASMState *as)
 #endif
 
   ra_setup(as);
+#if LJ_TARGET_ARM64
+  ra_setkref(as, RID_GL, (intptr_t)J2G(as->J));
+#endif
 
   /* Clear reg/sp for constants. */
-  for (ir = IR(T->nk), lastir = IR(REF_BASE); ir < lastir; ir++)
+  for (ir = IR(T->nk), lastir = IR(REF_BASE); ir < lastir; ir++) {
     ir->prev = REGSP_INIT;
+    if (irt_is64(ir->t) && ir->o != IR_KNULL) {
+#if LJ_GC64
+      /* The false-positive of irt_is64() for ASMREF_L (REF_NIL) is OK here. */
+      ir->i = 0;  /* Will become non-zero only for RIP-relative addresses. */
+#else
+      /* Make life easier for backends by putting address of constant in i. */
+      ir->i = (int32_t)(intptr_t)(ir+1);
+#endif
+      ir++;
+    }
+  }
 
   /* REF_BASE is used for implicit references to the BASE register. */
   lastir->prev = REGSP_HINT(RID_BASE);
 
-  ir = IR(nins-1);
-  if (ir->o == IR_RENAME) {
-    do { ir--; nins--; } while (ir->o == IR_RENAME);
-    T->nins = nins;  /* Remove any renames left over from ASM restart. */
-  }
   as->snaprename = nins;
   as->snapref = nins;
   as->snapno = T->nsnap;
+  as->snapalloc = 0;
 
   as->stopins = REF_BASE;
   as->orignins = nins;
@@ -1627,7 +2173,7 @@ static void asm_setup_regsp(ASMState *as)
   ir = IR(REF_FIRST);
   if (as->parent) {
     uint16_t *p;
-    lastir = lj_snap_regspmap(as->parent, as->J->exitno, ir);
+    lastir = lj_snap_regspmap(as->J, as->parent, as->J->exitno, ir);
     if (lastir - ir > LJ_MAX_JSLOTS)
       lj_trace_err(as->J, LJ_TRERR_NYICOAL);
     as->stopins = (IRRef)((lastir-1) - as->ir);
@@ -1666,6 +2212,10 @@ static void asm_setup_regsp(ASMState *as)
       ir->prev = (uint16_t)REGSP_HINT((rload & 15));
       rload = lj_ror(rload, 4);
       continue;
+    case IR_TMPREF:
+      if ((ir->op2 & IRTMPREF_OUT2) && as->evenspill < 4)
+        as->evenspill = 4;  /* TMPREF OUT2 needs two TValues on the stack. */
+      break;
 #endif
     case IR_CALLXS: {
       CCallInfo ci;
@@ -1675,7 +2225,17 @@ static void asm_setup_regsp(ASMState *as)
         as->modset |= RSET_SCRATCH;
       continue;
       }
-    case IR_CALLN: case IR_CALLL: case IR_CALLS: {
+    case IR_CALLL:
+      /* lj_vm_next needs two TValues on the stack. */
+#if LJ_TARGET_X64 && LJ_ABI_WIN
+      if (ir->op2 == IRCALL_lj_vm_next && as->evenspill < SPS_FIRST + 4)
+        as->evenspill = SPS_FIRST + 4;
+#else
+      if (SPS_FIRST < 4 && ir->op2 == IRCALL_lj_vm_next && as->evenspill < 4)
+        as->evenspill = 4;
+#endif
+      /* fallthrough */
+    case IR_CALLN: case IR_CALLA: case IR_CALLS: {
       const CCallInfo *ci = &lj_ir_callinfo[ir->op2];
       ir->prev = asm_setup_call_slots(as, ir, ci);
       if (inloop)
@@ -1683,7 +2243,6 @@ static void asm_setup_regsp(ASMState *as)
                       (RSET_SCRATCH & ~RSET_FPR) : RSET_SCRATCH;
       continue;
       }
-#if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
     case IR_HIOP:
       switch ((ir-1)->o) {
 #if LJ_SOFTFP && LJ_TARGET_ARM
@@ -1694,15 +2253,15 @@ static void asm_setup_regsp(ASMState *as)
         }
         break;
 #endif
-#if !LJ_SOFTFP && LJ_NEED_FP64
+#if !LJ_SOFTFP && LJ_NEED_FP64 && LJ_32 && LJ_HASFFI
       case IR_CONV:
         if (irt_isfp((ir-1)->t)) {
           ir->prev = REGSP_HINT(RID_FPRET);
           continue;
         }
-        /* fallthrough */
 #endif
-      case IR_CALLN: case IR_CALLXS:
+      /* fallthrough */
+      case IR_CALLN: case IR_CALLL: case IR_CALLS: case IR_CALLXS:
 #if LJ_SOFTFP
       case IR_MIN: case IR_MAX:
 #endif
@@ -1713,18 +2272,29 @@ static void asm_setup_regsp(ASMState *as)
         break;
       }
       break;
-#endif
 #if LJ_SOFTFP
     case IR_MIN: case IR_MAX:
       if ((ir+1)->o != IR_HIOP) break;
 #endif
     /* fallthrough */
     /* C calls evict all scratch regs and return results in RID_RET. */
-    case IR_SNEW: case IR_XSNEW: case IR_NEWREF:
+    case IR_SNEW: case IR_XSNEW: case IR_NEWREF: case IR_BUFPUT:
       if (REGARG_NUMGPR < 3 && as->evenspill < 3)
         as->evenspill = 3;  /* lj_str_new and lj_tab_newkey need 3 args. */
+#if LJ_TARGET_X86 && LJ_HASFFI
+      if (0) {
+    case IR_CNEW:
+        if (ir->op2 != REF_NIL && as->evenspill < 4)
+          as->evenspill = 4;  /* lj_cdata_newv needs 4 args. */
+      }
       /* fallthrough */
-    case IR_TNEW: case IR_TDUP: case IR_CNEW: case IR_CNEWI: case IR_TOSTR:
+#else
+      /* fallthrough */
+    case IR_CNEW:
+#endif
+      /* fallthrough */
+    case IR_TNEW: case IR_TDUP: case IR_CNEWI: case IR_TOSTR:
+    case IR_BUFSTR:
       ir->prev = REGSP_HINT(RID_RET);
       if (inloop)
         as->modset = RSET_SCRATCH;
@@ -1733,58 +2303,73 @@ static void asm_setup_regsp(ASMState *as)
       if (inloop)
         as->modset = RSET_SCRATCH;
       break;
-#if !LJ_TARGET_X86ORX64 && !LJ_SOFTFP
-    case IR_ATAN2: case IR_LDEXP:
+#if !LJ_SOFTFP
+#if !LJ_TARGET_X86ORX64
+    case IR_LDEXP:
+#endif
 #endif
+      /* fallthrough */
     case IR_POW:
       if (!LJ_SOFTFP && irt_isnum(ir->t)) {
-#if LJ_TARGET_X86ORX64
-        ir->prev = REGSP_HINT(RID_XMM0);
         if (inloop)
-          as->modset |= RSET_RANGE(RID_XMM0, RID_XMM1+1)|RID2RSET(RID_EAX);
+          as->modset |= RSET_SCRATCH;
+#if LJ_TARGET_X86
+        if (irt_isnum(IR(ir->op2)->t)) {
+          if (as->evenspill < 4)  /* Leave room to call pow(). */
+            as->evenspill = 4;
+        }
+        break;
 #else
         ir->prev = REGSP_HINT(RID_FPRET);
-        if (inloop)
-          as->modset |= RSET_SCRATCH;
-#endif
         continue;
+#endif
       }
       /* fallthrough */ /* for integer POW */
     case IR_DIV: case IR_MOD:
-      if (!irt_isnum(ir->t)) {
+      if ((LJ_64 && LJ_SOFTFP) || !irt_isnum(ir->t)) {
         ir->prev = REGSP_HINT(RID_RET);
         if (inloop)
           as->modset |= (RSET_SCRATCH & RSET_GPR);
         continue;
       }
       break;
-    case IR_FPMATH:
-#if LJ_TARGET_X86ORX64
-      if (ir->op2 == IRFPM_EXP2) {  /* May be joined to lj_vm_pow_sse. */
-        ir->prev = REGSP_HINT(RID_XMM0);
-#if !LJ_64
-        if (as->evenspill < 4)  /* Leave room for 16 byte scratch area. */
-          as->evenspill = 4;
-#endif
-        if (inloop)
-          as->modset |= RSET_RANGE(RID_XMM0, RID_XMM2+1)|RID2RSET(RID_EAX);
-        continue;
-      } else if (ir->op2 <= IRFPM_TRUNC && !(as->flags & JIT_F_SSE4_1)) {
-        ir->prev = REGSP_HINT(RID_XMM0);
+#if LJ_64 && LJ_SOFTFP
+    case IR_ADD: case IR_SUB: case IR_MUL:
+      if (irt_isnum(ir->t)) {
+        ir->prev = REGSP_HINT(RID_RET);
         if (inloop)
-          as->modset |= RSET_RANGE(RID_XMM0, RID_XMM3+1)|RID2RSET(RID_EAX);
+          as->modset |= (RSET_SCRATCH & RSET_GPR);
         continue;
       }
       break;
-#else
-      ir->prev = REGSP_HINT(RID_FPRET);
+#endif
+    case IR_FPMATH:
+#if LJ_TARGET_X86ORX64
+      if (ir->op2 <= IRFPM_TRUNC) {
+        if (!(as->flags & JIT_F_SSE4_1)) {
+          ir->prev = REGSP_HINT(RID_XMM0);
+          if (inloop)
+            as->modset |= RSET_RANGE(RID_XMM0, RID_XMM3+1)|RID2RSET(RID_EAX);
+          continue;
+        }
+        break;
+      }
+#endif
       if (inloop)
         as->modset |= RSET_SCRATCH;
+#if LJ_TARGET_X86
+      break;
+#else
+      ir->prev = REGSP_HINT(RID_FPRET);
       continue;
 #endif
 #if LJ_TARGET_X86ORX64
     /* Non-constant shift counts need to be in RID_ECX on x86/x64. */
-    case IR_BSHL: case IR_BSHR: case IR_BSAR: case IR_BROL: case IR_BROR:
+    case IR_BSHL: case IR_BSHR: case IR_BSAR:
+      if ((as->flags & JIT_F_BMI2))  /* Except if BMI2 is available. */
+        break;
+      /* fallthrough */
+    case IR_BROL: case IR_BROR:
       if (!irref_isk(ir->op2) && !ra_hashint(IR(ir->op2)->r)) {
         IR(ir->op2)->r = REGSP_HINT(RID_ECX);
         if (inloop)
@@ -1828,16 +2413,26 @@ void lj_asm_trace(jit_State *J, GCtrace *T)
 {
   ASMState as_;
   ASMState *as = &as_;
-  MCode *origtop;
+
+  /* Remove nops/renames left over from ASM restart due to LJ_TRERR_MCODELM. */
+  {
+    IRRef nins = T->nins;
+    IRIns *ir = &T->ir[nins-1];
+    if (ir->o == IR_NOP || ir->o == IR_RENAME) {
+      do { ir--; nins--; } while (ir->o == IR_NOP || ir->o == IR_RENAME);
+      T->nins = nins;
+    }
+  }
 
   /* Ensure an initialized instruction beyond the last one for HIOP checks. */
-  J->cur.nins = lj_ir_nextins(J);
-  lj_ir_nop(&J->cur.ir[J->cur.nins]);
+  /* This also allows one RENAME to be added without reallocating curfinal. */
+  as->orignins = lj_ir_nextins(J);
+  lj_ir_nop(&J->cur.ir[as->orignins]);
 
   /* Setup initial state. Copy some fields to reduce indirections. */
   as->J = J;
   as->T = T;
-  as->ir = T->ir;
+  J->curfinal = lj_trace_alloc(J->L, T);  /* This copies the IR, too. */
   as->flags = J->flags;
   as->loopref = J->loopref;
   as->realign = NULL;
@@ -1845,17 +2440,46 @@ void lj_asm_trace(jit_State *J, GCtrace *T)
   as->parent = J->parent ? traceref(J, J->parent) : NULL;
 
   /* Reserve MCode memory. */
-  as->mctop = origtop = lj_mcode_reserve(J, &as->mcbot);
+  as->mctop = as->mctoporig = lj_mcode_reserve(J, &as->mcbot);
   as->mcp = as->mctop;
   as->mclim = as->mcbot + MCLIM_REDZONE;
   asm_setup_target(as);
 
-  do {
+  /*
+  ** This is a loop, because the MCode may have to be (re-)assembled
+  ** multiple times:
+  **
+  ** 1. as->realign is set (and the assembly aborted), if the arch-specific
+  **    backend wants the MCode to be aligned differently.
+  **
+  **    This is currently only the case on x86/x64, where small loops get
+  **    an aligned loop body plus a short branch. Not much effort is wasted,
+  **    because the abort happens very quickly and only once.
+  **
+  ** 2. The IR is immovable, since the MCode embeds pointers to various
+  **    constants inside the IR. But RENAMEs may need to be added to the IR
+  **    during assembly, which might grow and reallocate the IR. We check
+  **    at the end if the IR (in J->cur.ir) has actually grown, resize the
+  **    copy (in J->curfinal.ir) and try again.
+  **
+  **    95% of all traces have zero RENAMEs, 3% have one RENAME, 1.5% have
+  **    2 RENAMEs and only 0.5% have more than that. That's why we opt to
+  **    always have one spare slot in the IR (see above), which means we
+  **    have to redo the assembly for only ~2% of all traces.
+  **
+  **    Very, very rarely, this needs to be done repeatedly, since the
+  **    location of constants inside the IR (actually, reachability from
+  **    a global pointer) may affect register allocation and thus the
+  **    number of RENAMEs.
+  */
+  for (;;) {
     as->mcp = as->mctop;
 #ifdef LUA_USE_ASSERT
     as->mcp_prev = as->mcp;
 #endif
-    as->curins = T->nins;
+    as->ir = J->curfinal->ir;  /* Use the copied IR. */
+    as->curins = J->cur.nins = as->orignins;
+
     RA_DBG_START();
     RA_DBGX((as, "===== STOP ====="));
 
@@ -1874,7 +2498,11 @@ void lj_asm_trace(jit_State *J, GCtrace *T)
     /* Assemble a trace in linear backwards order. */
     for (as->curins--; as->curins > as->stopins; as->curins--) {
       IRIns *ir = IR(as->curins);
-      lua_assert(!(LJ_32 && irt_isint64(ir->t)));  /* Handled by SPLIT. */
+      /* 64 bit types handled by SPLIT for 32 bit archs. */
+      lj_assertA(!(LJ_32 && irt_isint64(ir->t)),
+                 "IR %04d has unsplit 64 bit type",
+                 (int)(ir - as->ir) - REF_BIAS);
+      asm_snap_prev(as);
       if (!ra_used(ir) && !ir_sideeff(ir) && (as->flags & JIT_F_OPT_DCE))
         continue;  /* Dead-code elimination can be soooo easy. */
       if (irt_isguard(ir->t))
@@ -1883,22 +2511,43 @@ void lj_asm_trace(jit_State *J, GCtrace *T)
       checkmclim(as);
       asm_ir(as, ir);
     }
-  } while (as->realign);  /* Retry in case the MCode needs to be realigned. */
 
-  /* Emit head of trace. */
-  RA_DBG_REF();
-  checkmclim(as);
-  if (as->gcsteps > 0) {
-    as->curins = as->T->snap[0].ref;
-    asm_snap_prep(as);  /* The GC check is a guard. */
-    asm_gc_check(as);
+    if (as->realign && J->curfinal->nins >= T->nins)
+      continue;  /* Retry in case only the MCode needs to be realigned. */
+
+    /* Emit head of trace. */
+    RA_DBG_REF();
+    checkmclim(as);
+    if (as->gcsteps > 0) {
+      as->curins = as->T->snap[0].ref;
+      asm_snap_prep(as);  /* The GC check is a guard. */
+      asm_gc_check(as);
+      as->curins = as->stopins;
+    }
+    ra_evictk(as);
+    if (as->parent)
+      asm_head_side(as);
+    else
+      asm_head_root(as);
+    asm_phi_fixup(as);
+
+    if (J->curfinal->nins >= T->nins) {  /* IR didn't grow? */
+      lj_assertA(J->curfinal->nk == T->nk, "unexpected IR constant growth");
+      memcpy(J->curfinal->ir + as->orignins, T->ir + as->orignins,
+             (T->nins - as->orignins) * sizeof(IRIns));  /* Copy RENAMEs. */
+      T->nins = J->curfinal->nins;
+      /* Fill mcofs of any unprocessed snapshots. */
+      as->curins = REF_FIRST;
+      asm_snap_prev(as);
+      break;  /* Done. */
+    }
+
+    /* Otherwise try again with a bigger IR. */
+    lj_trace_free(J2G(J), J->curfinal);
+    J->curfinal = NULL;  /* In case lj_trace_alloc() OOMs. */
+    J->curfinal = lj_trace_alloc(J->L, T);
+    as->realign = NULL;
   }
-  ra_evictk(as);
-  if (as->parent)
-    asm_head_side(as);
-  else
-    asm_head_root(as);
-  asm_phi_fixup(as);
 
   RA_DBGX((as, "===== START ===="));
   RA_DBG_FLUSH();
@@ -1908,10 +2557,16 @@ void lj_asm_trace(jit_State *J, GCtrace *T)
   /* Set trace entry point before fixing up tail to allow link to self. */
   T->mcode = as->mcp;
   T->mcloop = as->mcloop ? (MSize)((char *)as->mcloop - (char *)as->mcp) : 0;
-  if (!as->loopref)
+  if (as->loopref)
+    asm_loop_tail_fixup(as);
+  else
     asm_tail_fixup(as, T->link);  /* Note: this may change as->mctop! */
   T->szmcode = (MSize)((char *)as->mctop - (char *)as->mcp);
-  lj_mcode_sync(T->mcode, origtop);
+  asm_snap_fixup_mcofs(as);
+#if LJ_TARGET_MCODE_FIXUP
+  asm_mcode_fixup(T->mcode, T->szmcode);
+#endif
+  lj_mcode_sync(T->mcode, as->mctoporig);
 }
 
 #undef IR