Praxis3D/Dependencies/include/Lua - Copy/lj_asm_arm64.h

/*
** ARM64 IR assembler (SSA IR -> machine code).
** Copyright (C) 2005-2023 Mike Pall. See Copyright Notice in luajit.h
**
** Contributed by Djordje Kovacevic and Stefan Pejic from RT-RK.com.
** Sponsored by Cisco Systems, Inc.
*/

/* -- Register allocator extensions --------------------------------------- */

/* Allocate a register with a hint. */
static Reg ra_hintalloc(ASMState *as, IRRef ref, Reg hint, RegSet allow)
{
  Reg r = IR(ref)->r;
  if (ra_noreg(r)) {
    if (!ra_hashint(r) && !iscrossref(as, ref))
      ra_sethint(IR(ref)->r, hint);  /* Propagate register hint. */
    r = ra_allocref(as, ref, allow);
  }
  ra_noweak(as, r);
  return r;
}

/* Allocate two source registers for three-operand instructions. */
static Reg ra_alloc2(ASMState *as, IRIns *ir, RegSet allow)
{
  IRIns *irl = IR(ir->op1), *irr = IR(ir->op2);
  Reg left = irl->r, right = irr->r;
  if (ra_hasreg(left)) {
    ra_noweak(as, left);
    if (ra_noreg(right))
      right = ra_allocref(as, ir->op2, rset_exclude(allow, left));
    else
      ra_noweak(as, right);
  } else if (ra_hasreg(right)) {
    ra_noweak(as, right);
    left = ra_allocref(as, ir->op1, rset_exclude(allow, right));
  } else if (ra_hashint(right)) {
    right = ra_allocref(as, ir->op2, allow);
    left = ra_alloc1(as, ir->op1, rset_exclude(allow, right));
  } else {
    left = ra_allocref(as, ir->op1, allow);
    right = ra_alloc1(as, ir->op2, rset_exclude(allow, left));
  }
  return left | (right << 8);
}

/* -- Guard handling ------------------------------------------------------ */

/* Setup all needed exit stubs. */
static void asm_exitstub_setup(ASMState *as, ExitNo nexits)
{
  ExitNo i;
  MCode *mxp = as->mctop;
  if (mxp - (nexits + 3 + MCLIM_REDZONE) < as->mclim)
    asm_mclimit(as);
  /* 1: str lr,[sp]; bl ->vm_exit_handler; movz w0,traceno; bl <1; bl <1; ... */
  for (i = nexits-1; (int32_t)i >= 0; i--)
    *--mxp = A64I_LE(A64I_BL | A64F_S26(-3-i));
  *--mxp = A64I_LE(A64I_MOVZw | A64F_U16(as->T->traceno));
  mxp--;
  *mxp = A64I_LE(A64I_BL | A64F_S26(((MCode *)(void *)lj_vm_exit_handler-mxp)));
  *--mxp = A64I_LE(A64I_STRx | A64F_D(RID_LR) | A64F_N(RID_SP));
  as->mctop = mxp;
}

static MCode *asm_exitstub_addr(ASMState *as, ExitNo exitno)
{
  /* Keep this in-sync with exitstub_trace_addr(). */
  return as->mctop + exitno + 3;
}

/* Emit conditional branch to exit for guard. */
static void asm_guardcc(ASMState *as, A64CC cc)
{
  MCode *target = asm_exitstub_addr(as, as->snapno);
  MCode *p = as->mcp;
  if (LJ_UNLIKELY(p == as->invmcp)) {
    as->loopinv = 1;
    *p = A64I_B | A64F_S26(target-p);
    emit_cond_branch(as, cc^1, p-1);
    return;
  }
  emit_cond_branch(as, cc, target);
}

/* Emit test and branch instruction to exit for guard, if in range. */
static int asm_guardtnb(ASMState *as, A64Ins ai, Reg r, uint32_t bit)
{
  MCode *target = asm_exitstub_addr(as, as->snapno);
  MCode *p = as->mcp;
  ptrdiff_t delta = target - p;
  if (LJ_UNLIKELY(p == as->invmcp)) {
    if (as->orignins > 1023) return 0;  /* Delta might end up too large. */
    as->loopinv = 1;
    *p = A64I_B | A64F_S26(delta);
    ai ^= 0x01000000u;
    target = p-1;
  } else if (LJ_UNLIKELY(delta >= 0x1fff)) {
    return 0;
  }
  emit_tnb(as, ai, r, bit, target);
  return 1;
}

/* Emit compare and branch instruction to exit for guard. */
static void asm_guardcnb(ASMState *as, A64Ins ai, Reg r)
{
  MCode *target = asm_exitstub_addr(as, as->snapno);
  MCode *p = as->mcp;
  if (LJ_UNLIKELY(p == as->invmcp)) {
    as->loopinv = 1;
    *p = A64I_B | A64F_S26(target-p);
    emit_cnb(as, ai^0x01000000u, r, p-1);
    return;
  }
  emit_cnb(as, ai, r, target);
}

/* -- Operand fusion ------------------------------------------------------ */

/* Limit linear search to this distance. Avoids O(n^2) behavior. */
#define CONFLICT_SEARCH_LIM	31

static int asm_isk32(ASMState *as, IRRef ref, int32_t *k)
{
  if (irref_isk(ref)) {
    IRIns *ir = IR(ref);
    if (ir->o == IR_KNULL || !irt_is64(ir->t)) {
      *k = ir->i;
      return 1;
    } else if (checki32((int64_t)ir_k64(ir)->u64)) {
      *k = (int32_t)ir_k64(ir)->u64;
      return 1;
    }
  }
  return 0;
}

/* Check if there's no conflicting instruction between curins and ref. */
static int noconflict(ASMState *as, IRRef ref, IROp conflict)
{
  IRIns *ir = as->ir;
  IRRef i = as->curins;
  if (i > ref + CONFLICT_SEARCH_LIM)
    return 0;  /* Give up, ref is too far away. */
  while (--i > ref)
    if (ir[i].o == conflict)
      return 0;  /* Conflict found. */
  return 1;  /* Ok, no conflict. */
}

/* Fuse the array base of colocated arrays. */
static int32_t asm_fuseabase(ASMState *as, IRRef ref)
{
  IRIns *ir = IR(ref);
  if (ir->o == IR_TNEW && ir->op1 <= LJ_MAX_COLOSIZE &&
      !neverfuse(as) && noconflict(as, ref, IR_NEWREF))
    return (int32_t)sizeof(GCtab);
  return 0;
}

#define FUSE_REG	0x40000000

/* Fuse array/hash/upvalue reference into register+offset operand. */
static Reg asm_fuseahuref(ASMState *as, IRRef ref, int32_t *ofsp, RegSet allow,
			  A64Ins ins)
{
  IRIns *ir = IR(ref);
  if (ra_noreg(ir->r)) {
    if (ir->o == IR_AREF) {
      if (mayfuse(as, ref)) {
	if (irref_isk(ir->op2)) {
	  IRRef tab = IR(ir->op1)->op1;
	  int32_t ofs = asm_fuseabase(as, tab);
	  IRRef refa = ofs ? tab : ir->op1;
	  ofs += 8*IR(ir->op2)->i;
	  if (emit_checkofs(ins, ofs)) {
	    *ofsp = ofs;
	    return ra_alloc1(as, refa, allow);
	  }
	} else {
	  Reg base = ra_alloc1(as, ir->op1, allow);
	  *ofsp = FUSE_REG|ra_alloc1(as, ir->op2, rset_exclude(allow, base));
	  return base;
	}
      }
    } else if (ir->o == IR_HREFK) {
      if (mayfuse(as, ref)) {
	int32_t ofs = (int32_t)(IR(ir->op2)->op2 * sizeof(Node));
	if (emit_checkofs(ins, ofs)) {
	  *ofsp = ofs;
	  return ra_alloc1(as, ir->op1, allow);
	}
      }
    } else if (ir->o == IR_UREFC) {
      if (irref_isk(ir->op1)) {
	GCfunc *fn = ir_kfunc(IR(ir->op1));
	GCupval *uv = &gcref(fn->l.uvptr[(ir->op2 >> 8)])->uv;
	int64_t ofs = glofs(as, &uv->tv);
	if (emit_checkofs(ins, ofs)) {
	  *ofsp = (int32_t)ofs;
	  return RID_GL;
	}
      }
    } else if (ir->o == IR_TMPREF) {
      *ofsp = (int32_t)glofs(as, &J2G(as->J)->tmptv);
      return RID_GL;
    }
  }
  *ofsp = 0;
  return ra_alloc1(as, ref, allow);
}

/* Fuse m operand into arithmetic/logic instructions. */
static uint32_t asm_fuseopm(ASMState *as, A64Ins ai, IRRef ref, RegSet allow)
{
  IRIns *ir = IR(ref);
  int logical = (ai & 0x1f000000) == 0x0a000000;
  if (ra_hasreg(ir->r)) {
    ra_noweak(as, ir->r);
    return A64F_M(ir->r);
  } else if (irref_isk(ref)) {
    int64_t k = get_k64val(as, ref);
    uint32_t m = logical ? emit_isk13(k, irt_is64(ir->t)) :
			   emit_isk12(irt_is64(ir->t) ? k : (int32_t)k);
    if (m)
      return m;
  } else if (mayfuse(as, ref)) {
    if ((ir->o >= IR_BSHL && ir->o <= IR_BSAR && irref_isk(ir->op2)) ||
	(ir->o == IR_ADD && ir->op1 == ir->op2)) {
      A64Shift sh = ir->o == IR_BSHR ? A64SH_LSR :
		    ir->o == IR_BSAR ? A64SH_ASR : A64SH_LSL;
      int shift = ir->o == IR_ADD ? 1 :
		    (IR(ir->op2)->i & (irt_is64(ir->t) ? 63 : 31));
      IRIns *irl = IR(ir->op1);
      if (sh == A64SH_LSL &&
	  irl->o == IR_CONV && !logical &&
	  irl->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT) &&
	  shift <= 4 &&
	  canfuse(as, irl)) {
	Reg m = ra_alloc1(as, irl->op1, allow);
	return A64F_M(m) | A64F_EXSH(A64EX_SXTW, shift);
      } else {
	Reg m = ra_alloc1(as, ir->op1, allow);
	return A64F_M(m) | A64F_SH(sh, shift);
      }
    } else if (ir->o == IR_BROR && logical && irref_isk(ir->op2)) {
      Reg m = ra_alloc1(as, ir->op1, allow);
      int shift = (IR(ir->op2)->i & (irt_is64(ir->t) ? 63 : 31));
      return A64F_M(m) | A64F_SH(A64SH_ROR, shift);
    } else if (ir->o == IR_CONV && !logical &&
	       ir->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT)) {
      Reg m = ra_alloc1(as, ir->op1, allow);
      return A64F_M(m) | A64F_EX(A64EX_SXTW);
    }
  }
  return A64F_M(ra_allocref(as, ref, allow));
}

/* Fuse XLOAD/XSTORE reference into load/store operand. */
static void asm_fusexref(ASMState *as, A64Ins ai, Reg rd, IRRef ref,
			 RegSet allow)
{
  IRIns *ir = IR(ref);
  Reg base;
  int32_t ofs = 0;
  if (ra_noreg(ir->r) && canfuse(as, ir)) {
    if (ir->o == IR_ADD) {
      if (asm_isk32(as, ir->op2, &ofs) && emit_checkofs(ai, ofs)) {
	ref = ir->op1;
      } else {
	Reg rn, rm;
	IRRef lref = ir->op1, rref = ir->op2;
	IRIns *irl = IR(lref);
	if (mayfuse(as, irl->op1)) {
	  unsigned int shift = 4;
	  if (irl->o == IR_BSHL && irref_isk(irl->op2)) {
	    shift = (IR(irl->op2)->i & 63);
	  } else if (irl->o == IR_ADD && irl->op1 == irl->op2) {
	    shift = 1;
	  }
	  if ((ai >> 30) == shift) {
	    lref = irl->op1;
	    irl = IR(lref);
	    ai |= A64I_LS_SH;
	  }
	}
	if (irl->o == IR_CONV &&
	    irl->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT) &&
	    canfuse(as, irl)) {
	  lref = irl->op1;
	  ai |= A64I_LS_SXTWx;
	} else {
	  ai |= A64I_LS_LSLx;
	}
	rm = ra_alloc1(as, lref, allow);
	rn = ra_alloc1(as, rref, rset_exclude(allow, rm));
	emit_dnm(as, (ai^A64I_LS_R), (rd & 31), rn, rm);
	return;
      }
    } else if (ir->o == IR_STRREF) {
      if (asm_isk32(as, ir->op2, &ofs)) {
	ref = ir->op1;
      } else if (asm_isk32(as, ir->op1, &ofs)) {
	ref = ir->op2;
      } else {
	Reg refk = irref_isk(ir->op1) ? ir->op1 : ir->op2;
	Reg refv = irref_isk(ir->op1) ? ir->op2 : ir->op1;
	Reg rn = ra_alloc1(as, refv, allow);
	IRIns *irr = IR(refk);
	uint32_t m;
	if (irr+1 == ir && !ra_used(irr) &&
	    irr->o == IR_ADD && irref_isk(irr->op2)) {
	  ofs = sizeof(GCstr) + IR(irr->op2)->i;
	  if (emit_checkofs(ai, ofs)) {
	    Reg rm = ra_alloc1(as, irr->op1, rset_exclude(allow, rn));
	    m = A64F_M(rm) | A64F_EX(A64EX_SXTW);
	    goto skipopm;
	  }
	}
	m = asm_fuseopm(as, 0, refk, rset_exclude(allow, rn));
	ofs = sizeof(GCstr);
      skipopm:
	emit_lso(as, ai, rd, rd, ofs);
	emit_dn(as, A64I_ADDx^m, rd, rn);
	return;
      }
      ofs += sizeof(GCstr);
      if (!emit_checkofs(ai, ofs)) {
	Reg rn = ra_alloc1(as, ref, allow);
	Reg rm = ra_allock(as, ofs, rset_exclude(allow, rn));
	emit_dnm(as, (ai^A64I_LS_R)|A64I_LS_UXTWx, rd, rn, rm);
	return;
      }
    }
  }
  base = ra_alloc1(as, ref, allow);
  emit_lso(as, ai, (rd & 31), base, ofs);
}

/* Fuse FP multiply-add/sub. */
static int asm_fusemadd(ASMState *as, IRIns *ir, A64Ins ai, A64Ins air)
{
  IRRef lref = ir->op1, rref = ir->op2;
  IRIns *irm;
  if ((as->flags & JIT_F_OPT_FMA) &&
      lref != rref &&
      ((mayfuse(as, lref) && (irm = IR(lref), irm->o == IR_MUL) &&
       ra_noreg(irm->r)) ||
       (mayfuse(as, rref) && (irm = IR(rref), irm->o == IR_MUL) &&
       (rref = lref, ai = air, ra_noreg(irm->r))))) {
    Reg dest = ra_dest(as, ir, RSET_FPR);
    Reg add = ra_hintalloc(as, rref, dest, RSET_FPR);
    Reg left = ra_alloc2(as, irm,
			 rset_exclude(rset_exclude(RSET_FPR, dest), add));
    Reg right = (left >> 8); left &= 255;
    emit_dnma(as, ai, (dest & 31), (left & 31), (right & 31), (add & 31));
    return 1;
  }
  return 0;
}

/* Fuse BAND + BSHL/BSHR into UBFM. */
static int asm_fuseandshift(ASMState *as, IRIns *ir)
{
  IRIns *irl = IR(ir->op1);
  lj_assertA(ir->o == IR_BAND, "bad usage");
  if (canfuse(as, irl) && irref_isk(ir->op2)) {
    uint64_t mask = get_k64val(as, ir->op2);
    if (irref_isk(irl->op2) && (irl->o == IR_BSHR || irl->o == IR_BSHL)) {
      int32_t shmask = irt_is64(irl->t) ? 63 : 31;
      int32_t shift = (IR(irl->op2)->i & shmask);
      int32_t imms = shift;
      if (irl->o == IR_BSHL) {
	mask >>= shift;
	shift = (shmask-shift+1) & shmask;
	imms = 0;
      }
      if (mask && !((mask+1) & mask)) {  /* Contiguous 1-bits at the bottom. */
	Reg dest = ra_dest(as, ir, RSET_GPR);
	Reg left = ra_alloc1(as, irl->op1, RSET_GPR);
	A64Ins ai = shmask == 63 ? A64I_UBFMx : A64I_UBFMw;
	imms += 63 - emit_clz64(mask);
	if (imms > shmask) imms = shmask;
	emit_dn(as, ai | A64F_IMMS(imms) | A64F_IMMR(shift), dest, left);
	return 1;
      }
    }
  }
  return 0;
}

/* Fuse BOR(BSHL, BSHR) into EXTR/ROR. */
static int asm_fuseorshift(ASMState *as, IRIns *ir)
{
  IRIns *irl = IR(ir->op1), *irr = IR(ir->op2);
  lj_assertA(ir->o == IR_BOR, "bad usage");
  if (canfuse(as, irl) && canfuse(as, irr) &&
      ((irl->o == IR_BSHR && irr->o == IR_BSHL) ||
       (irl->o == IR_BSHL && irr->o == IR_BSHR))) {
    if (irref_isk(irl->op2) && irref_isk(irr->op2)) {
      IRRef lref = irl->op1, rref = irr->op1;
      uint32_t lshift = IR(irl->op2)->i, rshift = IR(irr->op2)->i;
      if (irl->o == IR_BSHR) {  /* BSHR needs to be the right operand. */
	uint32_t tmp2;
	IRRef tmp1 = lref; lref = rref; rref = tmp1;
	tmp2 = lshift; lshift = rshift; rshift = tmp2;
      }
      if (rshift + lshift == (irt_is64(ir->t) ? 64 : 32)) {
	A64Ins ai = irt_is64(ir->t) ? A64I_EXTRx : A64I_EXTRw;
	Reg dest = ra_dest(as, ir, RSET_GPR);
	Reg left = ra_alloc1(as, lref, RSET_GPR);
	Reg right = ra_alloc1(as, rref, rset_exclude(RSET_GPR, left));
	emit_dnm(as, ai | A64F_IMMS(rshift), dest, left, right);
	return 1;
      }
    }
  }
  return 0;
}

/* -- Calls --------------------------------------------------------------- */

/* Generate a call to a C function. */
static void asm_gencall(ASMState *as, const CCallInfo *ci, IRRef *args)
{
  uint32_t n, nargs = CCI_XNARGS(ci);
  int32_t spofs = 0, spalign = LJ_HASFFI && LJ_TARGET_OSX ? 0 : 7;
  Reg gpr, fpr = REGARG_FIRSTFPR;
  if (ci->func)
    emit_call(as, ci->func);
  for (gpr = REGARG_FIRSTGPR; gpr <= REGARG_LASTGPR; gpr++)
    as->cost[gpr] = REGCOST(~0u, ASMREF_L);
  gpr = REGARG_FIRSTGPR;
#if LJ_HASFFI && LJ_ABI_WIN
  if ((ci->flags & CCI_VARARG)) {
    fpr = REGARG_LASTFPR+1;
  }
#endif
  for (n = 0; n < nargs; n++) { /* Setup args. */
    IRRef ref = args[n];
    IRIns *ir = IR(ref);
    if (ref) {
      if (irt_isfp(ir->t)) {
	if (fpr <= REGARG_LASTFPR) {
	  lj_assertA(rset_test(as->freeset, fpr),
		     "reg %d not free", fpr);  /* Must have been evicted. */
	  ra_leftov(as, fpr, ref);
	  fpr++;
#if LJ_HASFFI && LJ_ABI_WIN
	} else if ((ci->flags & CCI_VARARG) && (gpr <= REGARG_LASTGPR)) {
	  Reg rf = ra_alloc1(as, ref, RSET_FPR);
	  emit_dn(as, A64I_FMOV_R_D, gpr++, rf & 31);
#endif
	} else {
	  Reg r = ra_alloc1(as, ref, RSET_FPR);
	  int32_t al = spalign;
#if LJ_HASFFI && LJ_TARGET_OSX
	  al |= irt_isnum(ir->t) ? 7 : 3;
#endif
	  spofs = (spofs + al) & ~al;
	  if (LJ_BE && al >= 7 && !irt_isnum(ir->t)) spofs += 4, al -= 4;
	  emit_spstore(as, ir, r, spofs);
	  spofs += al + 1;
	}
      } else {
	if (gpr <= REGARG_LASTGPR) {
	  lj_assertA(rset_test(as->freeset, gpr),
		     "reg %d not free", gpr);  /* Must have been evicted. */
	  ra_leftov(as, gpr, ref);
	  gpr++;
	} else {
	  Reg r = ra_alloc1(as, ref, RSET_GPR);
	  int32_t al = spalign;
#if LJ_HASFFI && LJ_TARGET_OSX
	  al |= irt_size(ir->t) - 1;
#endif
	  spofs = (spofs + al) & ~al;
	  if (al >= 3) {
	    if (LJ_BE && al >= 7 && !irt_is64(ir->t)) spofs += 4, al -= 4;
	    emit_spstore(as, ir, r, spofs);
	  } else {
	    lj_assertA(al == 0 || al == 1, "size %d unexpected", al + 1);
	    emit_lso(as, al ? A64I_STRH : A64I_STRB, r, RID_SP, spofs);
	  }
	  spofs += al + 1;
	}
      }
#if LJ_HASFFI && LJ_TARGET_OSX
    } else {  /* Marker for start of varargs. */
      gpr = REGARG_LASTGPR+1;
      fpr = REGARG_LASTFPR+1;
      spalign = 7;
#endif
    }
  }
}

/* Setup result reg/sp for call. Evict scratch regs. */
static void asm_setupresult(ASMState *as, IRIns *ir, const CCallInfo *ci)
{
  RegSet drop = RSET_SCRATCH;
  int hiop = ((ir+1)->o == IR_HIOP && !irt_isnil((ir+1)->t));
  if (ra_hasreg(ir->r))
    rset_clear(drop, ir->r); /* Dest reg handled below. */
  if (hiop && ra_hasreg((ir+1)->r))
    rset_clear(drop, (ir+1)->r);  /* Dest reg handled below. */
  ra_evictset(as, drop); /* Evictions must be performed first. */
  if (ra_used(ir)) {
    lj_assertA(!irt_ispri(ir->t), "PRI dest");
    if (irt_isfp(ir->t)) {
      if (ci->flags & CCI_CASTU64) {
	Reg dest = ra_dest(as, ir, RSET_FPR) & 31;
	emit_dn(as, irt_isnum(ir->t) ? A64I_FMOV_D_R : A64I_FMOV_S_R,
		dest, RID_RET);
      } else {
	ra_destreg(as, ir, RID_FPRET);
      }
    } else if (hiop) {
      ra_destpair(as, ir);
    } else {
      ra_destreg(as, ir, RID_RET);
    }
  }
  UNUSED(ci);
}

static void asm_callx(ASMState *as, IRIns *ir)
{
  IRRef args[CCI_NARGS_MAX*2];
  CCallInfo ci;
  IRRef func;
  IRIns *irf;
  ci.flags = asm_callx_flags(as, ir);
  asm_collectargs(as, ir, &ci, args);
  asm_setupresult(as, ir, &ci);
  func = ir->op2; irf = IR(func);
  if (irf->o == IR_CARG) { func = irf->op1; irf = IR(func); }
  if (irref_isk(func)) {  /* Call to constant address. */
    ci.func = (ASMFunction)(ir_k64(irf)->u64);
  } else {  /* Need a non-argument register for indirect calls. */
    Reg freg = ra_alloc1(as, func, RSET_RANGE(RID_X8, RID_MAX_GPR)-RSET_FIXED);
    emit_n(as, A64I_BLR_AUTH, freg);
    ci.func = (ASMFunction)(void *)0;
  }
  asm_gencall(as, &ci, args);
}

/* -- Returns ------------------------------------------------------------- */

/* Return to lower frame. Guard that it goes to the right spot. */
static void asm_retf(ASMState *as, IRIns *ir)
{
  Reg base = ra_alloc1(as, REF_BASE, RSET_GPR);
  void *pc = ir_kptr(IR(ir->op2));
  int32_t delta = 1+LJ_FR2+bc_a(*((const BCIns *)pc - 1));
  as->topslot -= (BCReg)delta;
  if ((int32_t)as->topslot < 0) as->topslot = 0;
  irt_setmark(IR(REF_BASE)->t);  /* Children must not coalesce with BASE reg. */
  emit_setgl(as, base, jit_base);
  emit_addptr(as, base, -8*delta);
  asm_guardcc(as, CC_NE);
  emit_nm(as, A64I_CMPx, RID_TMP,
	  ra_allock(as, i64ptr(pc), rset_exclude(RSET_GPR, base)));
  emit_lso(as, A64I_LDRx, RID_TMP, base, -8);
}

/* -- Buffer operations --------------------------------------------------- */

#if LJ_HASBUFFER
static void asm_bufhdr_write(ASMState *as, Reg sb)
{
  Reg tmp = ra_scratch(as, rset_exclude(RSET_GPR, sb));
  IRIns irgc;
  irgc.ot = IRT(0, IRT_PGC);  /* GC type. */
  emit_storeofs(as, &irgc, RID_TMP, sb, offsetof(SBuf, L));
  emit_dn(as, A64I_BFMx | A64F_IMMS(lj_fls(SBUF_MASK_FLAG)) | A64F_IMMR(0), RID_TMP, tmp);
  emit_getgl(as, RID_TMP, cur_L);
  emit_loadofs(as, &irgc, tmp, sb, offsetof(SBuf, L));
}
#endif

/* -- Type conversions ---------------------------------------------------- */

static void asm_tointg(ASMState *as, IRIns *ir, Reg left)
{
  Reg tmp = ra_scratch(as, rset_exclude(RSET_FPR, left));
  Reg dest = ra_dest(as, ir, RSET_GPR);
  asm_guardcc(as, CC_NE);
  emit_nm(as, A64I_FCMPd, (tmp & 31), (left & 31));
  emit_dn(as, A64I_FCVT_F64_S32, (tmp & 31), dest);
  emit_dn(as, A64I_FCVT_S32_F64, dest, (left & 31));
}

static void asm_tobit(ASMState *as, IRIns *ir)
{
  RegSet allow = RSET_FPR;
  Reg left = ra_alloc1(as, ir->op1, allow);
  Reg right = ra_alloc1(as, ir->op2, rset_clear(allow, left));
  Reg tmp = ra_scratch(as, rset_clear(allow, right));
  Reg dest = ra_dest(as, ir, RSET_GPR);
  emit_dn(as, A64I_FMOV_R_S, dest, (tmp & 31));
  emit_dnm(as, A64I_FADDd, (tmp & 31), (left & 31), (right & 31));
}

static void asm_conv(ASMState *as, IRIns *ir)
{
  IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
  int st64 = (st == IRT_I64 || st == IRT_U64 || st == IRT_P64);
  int stfp = (st == IRT_NUM || st == IRT_FLOAT);
  IRRef lref = ir->op1;
  lj_assertA(irt_type(ir->t) != st, "inconsistent types for CONV");
  if (irt_isfp(ir->t)) {
    Reg dest = ra_dest(as, ir, RSET_FPR);
    if (stfp) {  /* FP to FP conversion. */
      emit_dn(as, st == IRT_NUM ? A64I_FCVT_F32_F64 : A64I_FCVT_F64_F32,
	      (dest & 31), (ra_alloc1(as, lref, RSET_FPR) & 31));
    } else {  /* Integer to FP conversion. */
      Reg left = ra_alloc1(as, lref, RSET_GPR);
      A64Ins ai = irt_isfloat(ir->t) ?
	(((IRT_IS64 >> st) & 1) ?
	 (st == IRT_I64 ? A64I_FCVT_F32_S64 : A64I_FCVT_F32_U64) :
	 (st == IRT_INT ? A64I_FCVT_F32_S32 : A64I_FCVT_F32_U32)) :
	(((IRT_IS64 >> st) & 1) ?
	 (st == IRT_I64 ? A64I_FCVT_F64_S64 : A64I_FCVT_F64_U64) :
	 (st == IRT_INT ? A64I_FCVT_F64_S32 : A64I_FCVT_F64_U32));
      emit_dn(as, ai, (dest & 31), left);
    }
  } else if (stfp) {  /* FP to integer conversion. */
    if (irt_isguard(ir->t)) {
      /* Checked conversions are only supported from number to int. */
      lj_assertA(irt_isint(ir->t) && st == IRT_NUM,
		 "bad type for checked CONV");
      asm_tointg(as, ir, ra_alloc1(as, lref, RSET_FPR));
    } else {
      Reg left = ra_alloc1(as, lref, RSET_FPR);
      Reg dest = ra_dest(as, ir, RSET_GPR);
      A64Ins ai = irt_is64(ir->t) ?
	(st == IRT_NUM ?
	 (irt_isi64(ir->t) ? A64I_FCVT_S64_F64 : A64I_FCVT_U64_F64) :
	 (irt_isi64(ir->t) ? A64I_FCVT_S64_F32 : A64I_FCVT_U64_F32)) :
	(st == IRT_NUM ?
	 (irt_isint(ir->t) ? A64I_FCVT_S32_F64 : A64I_FCVT_U32_F64) :
	 (irt_isint(ir->t) ? A64I_FCVT_S32_F32 : A64I_FCVT_U32_F32));
      emit_dn(as, ai, dest, (left & 31));
    }
  } else if (st >= IRT_I8 && st <= IRT_U16) { /* Extend to 32 bit integer. */
    Reg dest = ra_dest(as, ir, RSET_GPR);
    Reg left = ra_alloc1(as, lref, RSET_GPR);
    A64Ins ai = st == IRT_I8 ? A64I_SXTBw :
		st == IRT_U8 ? A64I_UXTBw :
		st == IRT_I16 ? A64I_SXTHw : A64I_UXTHw;
    lj_assertA(irt_isint(ir->t) || irt_isu32(ir->t), "bad type for CONV EXT");
    emit_dn(as, ai, dest, left);
  } else {
    Reg dest = ra_dest(as, ir, RSET_GPR);
    if (irt_is64(ir->t)) {
      if (st64 || !(ir->op2 & IRCONV_SEXT)) {
	/* 64/64 bit no-op (cast) or 32 to 64 bit zero extension. */
	ra_leftov(as, dest, lref);  /* Do nothing, but may need to move regs. */
      } else {  /* 32 to 64 bit sign extension. */
	Reg left = ra_alloc1(as, lref, RSET_GPR);
	emit_dn(as, A64I_SXTW, dest, left);
      }
    } else {
      if (st64 && !(ir->op2 & IRCONV_NONE)) {
	/* This is either a 32 bit reg/reg mov which zeroes the hiword
	** or a load of the loword from a 64 bit address.
	*/
	Reg left = ra_alloc1(as, lref, RSET_GPR);
	emit_dm(as, A64I_MOVw, dest, left);
      } else {  /* 32/32 bit no-op (cast). */
	ra_leftov(as, dest, lref);  /* Do nothing, but may need to move regs. */
      }
    }
  }
}

static void asm_strto(ASMState *as, IRIns *ir)
{
  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_strscan_num];
  IRRef args[2];
  Reg tmp;
  int32_t ofs = 0;
  ra_evictset(as, RSET_SCRATCH);
  if (ra_used(ir)) {
    if (ra_hasspill(ir->s)) {
      ofs = sps_scale(ir->s);
      if (ra_hasreg(ir->r)) {
	ra_free(as, ir->r);
	ra_modified(as, ir->r);
	emit_spload(as, ir, ir->r, ofs);
      }
    } else {
      Reg dest = ra_dest(as, ir, RSET_FPR);
      emit_lso(as, A64I_LDRd, (dest & 31), RID_SP, 0);
    }
  }
  asm_guardcnb(as, A64I_CBZ, RID_RET);
  args[0] = ir->op1; /* GCstr *str */
  args[1] = ASMREF_TMP1; /* TValue *n  */
  asm_gencall(as, ci, args);
  tmp = ra_releasetmp(as, ASMREF_TMP1);
  emit_opk(as, A64I_ADDx, tmp, RID_SP, ofs, RSET_GPR);
}

/* -- Memory references --------------------------------------------------- */

/* Store tagged value for ref at base+ofs. */
static void asm_tvstore64(ASMState *as, Reg base, int32_t ofs, IRRef ref)
{
  RegSet allow = rset_exclude(RSET_GPR, base);
  IRIns *ir = IR(ref);
  lj_assertA(irt_ispri(ir->t) || irt_isaddr(ir->t) || irt_isinteger(ir->t),
	     "store of IR type %d", irt_type(ir->t));
  if (irref_isk(ref)) {
    TValue k;
    lj_ir_kvalue(as->J->L, &k, ir);
    emit_lso(as, A64I_STRx, ra_allock(as, k.u64, allow), base, ofs);
  } else {
    Reg src = ra_alloc1(as, ref, allow);
    rset_clear(allow, src);
    if (irt_isinteger(ir->t)) {
      Reg type = ra_allock(as, (int64_t)irt_toitype(ir->t) << 47, allow);
      emit_lso(as, A64I_STRx, RID_TMP, base, ofs);
      emit_dnm(as, A64I_ADDx | A64F_EX(A64EX_UXTW), RID_TMP, type, src);
    } else {
      Reg type = ra_allock(as, (int32_t)irt_toitype(ir->t), allow);
      emit_lso(as, A64I_STRx, RID_TMP, base, ofs);
      emit_dnm(as, A64I_ADDx | A64F_SH(A64SH_LSL, 47), RID_TMP, src, type);
    }
  }
}

/* Get pointer to TValue. */
static void asm_tvptr(ASMState *as, Reg dest, IRRef ref, MSize mode)
{
  if ((mode & IRTMPREF_IN1)) {
    IRIns *ir = IR(ref);
    if (irt_isnum(ir->t)) {
      if (irref_isk(ref) && !(mode & IRTMPREF_OUT1)) {
	/* Use the number constant itself as a TValue. */
	ra_allockreg(as, i64ptr(ir_knum(ir)), dest);
	return;
      }
      emit_lso(as, A64I_STRd, (ra_alloc1(as, ref, RSET_FPR) & 31), dest, 0);
    } else {
      asm_tvstore64(as, dest, 0, ref);
    }
  }
  /* g->tmptv holds the TValue(s). */
  emit_dn(as, A64I_ADDx^emit_isk12(glofs(as, &J2G(as->J)->tmptv)), dest, RID_GL);
}

static void asm_aref(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg idx, base;
  if (irref_isk(ir->op2)) {
    IRRef tab = IR(ir->op1)->op1;
    int32_t ofs = asm_fuseabase(as, tab);
    IRRef refa = ofs ? tab : ir->op1;
    uint32_t k = emit_isk12(ofs + 8*IR(ir->op2)->i);
    if (k) {
      base = ra_alloc1(as, refa, RSET_GPR);
      emit_dn(as, A64I_ADDx^k, dest, base);
      return;
    }
  }
  base = ra_alloc1(as, ir->op1, RSET_GPR);
  idx = ra_alloc1(as, ir->op2, rset_exclude(RSET_GPR, base));
  emit_dnm(as, A64I_ADDx | A64F_EXSH(A64EX_UXTW, 3), dest, base, idx);
}

/* Inlined hash lookup. Specialized for key type and for const keys.
** The equivalent C code is:
**   Node *n = hashkey(t, key);
**   do {
**     if (lj_obj_equal(&n->key, key)) return &n->val;
**   } while ((n = nextnode(n)));
**   return niltv(L);
*/
static void asm_href(ASMState *as, IRIns *ir, IROp merge)
{
  RegSet allow = RSET_GPR;
  int destused = ra_used(ir);
  Reg dest = ra_dest(as, ir, allow);
  Reg tab = ra_alloc1(as, ir->op1, rset_clear(allow, dest));
  Reg tmp = RID_TMP, type = RID_NONE, key, tkey;
  IRRef refkey = ir->op2;
  IRIns *irkey = IR(refkey);
  int isk = irref_isk(refkey);
  IRType1 kt = irkey->t;
  uint32_t k = 0;
  uint32_t khash;
  MCLabel l_end, l_loop;
  rset_clear(allow, tab);

  /* Allocate register for tkey outside of the loop. */
  if (isk) {
    int64_t kk;
    if (irt_isaddr(kt)) {
      kk = ((int64_t)irt_toitype(kt) << 47) | irkey[1].tv.u64;
    } else if (irt_isnum(kt)) {
      kk = (int64_t)ir_knum(irkey)->u64;
      /* Assumes -0.0 is already canonicalized to +0.0. */
    } else {
      lj_assertA(irt_ispri(kt) && !irt_isnil(kt), "bad HREF key type");
      kk = ~((int64_t)~irt_toitype(kt) << 47);
    }
    k = emit_isk12(kk);
    tkey = k ? 0 : ra_allock(as, kk, allow);
  } else {
    tkey = ra_scratch(as, allow);
  }

  /* Key not found in chain: jump to exit (if merged) or load niltv. */
  l_end = emit_label(as);
  as->invmcp = NULL;
  if (merge == IR_NE) {
    asm_guardcc(as, CC_AL);
  } else if (destused) {
    uint32_t k12 = emit_isk12(offsetof(global_State, nilnode.val));
    lj_assertA(k12 != 0, "Cannot k12 encode niltv(L)");
    emit_dn(as, A64I_ADDx^k12, dest, RID_GL);
  }

  /* Follow hash chain until the end. */
  l_loop = --as->mcp;
  if (destused)
    emit_lso(as, A64I_LDRx, dest, dest, offsetof(Node, next));

  /* Type and value comparison. */
  if (merge == IR_EQ)
    asm_guardcc(as, CC_EQ);
  else
    emit_cond_branch(as, CC_EQ, l_end);
  emit_nm(as, A64I_CMPx^k, tmp, tkey);
  if (!destused)
    emit_lso(as, A64I_LDRx, dest, dest, offsetof(Node, next));
  emit_lso(as, A64I_LDRx, tmp, dest, offsetof(Node, key));
  *l_loop = A64I_X | A64I_CBNZ | A64F_S19(as->mcp - l_loop) | dest;

  /* Construct tkey as canonicalized or tagged key. */
  if (!isk) {
    if (irt_isnum(kt)) {
      key = ra_alloc1(as, refkey, RSET_FPR);
      emit_dnm(as, A64I_CSELx | A64F_CC(CC_EQ), tkey, RID_ZERO, tkey);
      /* A64I_FMOV_R_D from key to tkey done below. */
    } else {
      lj_assertA(irt_isaddr(kt), "bad HREF key type");
      key = ra_alloc1(as, refkey, allow);
      type = ra_allock(as, irt_toitype(kt) << 15, rset_clear(allow, key));
      emit_dnm(as, A64I_ADDx | A64F_SH(A64SH_LSL, 32), tkey, key, type);
    }
  }

  /* Load main position relative to tab->node into dest. */
  khash = isk ? ir_khash(as, irkey) : 1;
  if (khash == 0) {
    emit_lso(as, A64I_LDRx, dest, tab, offsetof(GCtab, node));
  } else {
    emit_dnm(as, A64I_ADDx | A64F_SH(A64SH_LSL, 3), dest, tmp, dest);
    emit_dnm(as, A64I_ADDx | A64F_SH(A64SH_LSL, 1), dest, dest, dest);
    emit_lso(as, A64I_LDRx, tmp, tab, offsetof(GCtab, node));
    if (isk) {
      Reg tmphash = ra_allock(as, khash, allow);
      emit_dnm(as, A64I_ANDw, dest, dest, tmphash);
      emit_lso(as, A64I_LDRw, dest, tab, offsetof(GCtab, hmask));
    } else if (irt_isstr(kt)) {
      emit_dnm(as, A64I_ANDw, dest, dest, tmp);
      emit_lso(as, A64I_LDRw, tmp, key, offsetof(GCstr, sid));
      emit_lso(as, A64I_LDRw, dest, tab, offsetof(GCtab, hmask));
    } else {  /* Must match with hash*() in lj_tab.c. */
      emit_dnm(as, A64I_ANDw, dest, dest, tmp);
      emit_lso(as, A64I_LDRw, tmp, tab, offsetof(GCtab, hmask));
      emit_dnm(as, A64I_SUBw, dest, dest, tmp);
      emit_dnm(as, A64I_EXTRw | (A64F_IMMS(32-HASH_ROT3)), tmp, tmp, tmp);
      emit_dnm(as, A64I_EORw | A64F_SH(A64SH_ROR, 32-HASH_ROT2), dest, tmp, dest);
      emit_dnm(as, A64I_SUBw, tmp, tmp, dest);
      emit_dnm(as, A64I_EXTRw | (A64F_IMMS(32-HASH_ROT1)), dest, dest, dest);
      if (irt_isnum(kt)) {
	emit_dnm(as, A64I_EORw, tmp, tkey, dest);
	emit_dnm(as, A64I_ADDw, dest, dest, dest);
	emit_dn(as, A64I_LSRx | A64F_IMMR(32)|A64F_IMMS(32), dest, tkey);
	emit_nm(as, A64I_FCMPZd, (key & 31), 0);
	emit_dn(as, A64I_FMOV_R_D, tkey, (key & 31));
      } else {
	emit_dnm(as, A64I_EORw, tmp, key, dest);
	emit_dnm(as, A64I_EORx | A64F_SH(A64SH_LSR, 32), dest, type, key);
      }
    }
  }
}

static void asm_hrefk(ASMState *as, IRIns *ir)
{
  IRIns *kslot = IR(ir->op2);
  IRIns *irkey = IR(kslot->op1);
  int32_t ofs = (int32_t)(kslot->op2 * sizeof(Node));
  int32_t kofs = ofs + (int32_t)offsetof(Node, key);
  int bigofs = !emit_checkofs(A64I_LDRx, kofs);
  Reg dest = (ra_used(ir) || bigofs) ? ra_dest(as, ir, RSET_GPR) : RID_NONE;
  Reg node = ra_alloc1(as, ir->op1, RSET_GPR);
  Reg idx = node;
  RegSet allow = rset_exclude(RSET_GPR, node);
  uint64_t k;
  lj_assertA(ofs % sizeof(Node) == 0, "unaligned HREFK slot");
  if (bigofs) {
    idx = dest;
    rset_clear(allow, dest);
    kofs = (int32_t)offsetof(Node, key);
  } else if (ra_hasreg(dest)) {
    emit_opk(as, A64I_ADDx, dest, node, ofs, allow);
  }
  asm_guardcc(as, CC_NE);
  if (irt_ispri(irkey->t)) {
    k = ~((int64_t)~irt_toitype(irkey->t) << 47);
  } else if (irt_isnum(irkey->t)) {
    k = ir_knum(irkey)->u64;
  } else {
    k = ((uint64_t)irt_toitype(irkey->t) << 47) | (uint64_t)ir_kgc(irkey);
  }
  emit_nm(as, A64I_CMPx, RID_TMP, ra_allock(as, k, allow));
  emit_lso(as, A64I_LDRx, RID_TMP, idx, kofs);
  if (bigofs)
    emit_opk(as, A64I_ADDx, dest, node, ofs, rset_exclude(RSET_GPR, node));
}

static void asm_uref(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  int guarded = (irt_t(ir->t) & (IRT_GUARD|IRT_TYPE)) == (IRT_GUARD|IRT_PGC);
  if (irref_isk(ir->op1) && !guarded) {
    GCfunc *fn = ir_kfunc(IR(ir->op1));
    MRef *v = &gcref(fn->l.uvptr[(ir->op2 >> 8)])->uv.v;
    emit_lsptr(as, A64I_LDRx, dest, v);
  } else {
    if (guarded)
      asm_guardcnb(as, ir->o == IR_UREFC ? A64I_CBZ : A64I_CBNZ, RID_TMP);
    if (ir->o == IR_UREFC)
      emit_opk(as, A64I_ADDx, dest, dest,
	       (int32_t)offsetof(GCupval, tv), RSET_GPR);
    else
      emit_lso(as, A64I_LDRx, dest, dest, (int32_t)offsetof(GCupval, v));
    if (guarded)
      emit_lso(as, A64I_LDRB, RID_TMP, dest,
	       (int32_t)offsetof(GCupval, closed));
    if (irref_isk(ir->op1)) {
      GCfunc *fn = ir_kfunc(IR(ir->op1));
      uint64_t k = gcrefu(fn->l.uvptr[(ir->op2 >> 8)]);
      emit_loadu64(as, dest, k);
    } else {
      emit_lso(as, A64I_LDRx, dest, ra_alloc1(as, ir->op1, RSET_GPR),
	       (int32_t)offsetof(GCfuncL, uvptr) + 8*(int32_t)(ir->op2 >> 8));
    }
  }
}

static void asm_fref(ASMState *as, IRIns *ir)
{
  UNUSED(as); UNUSED(ir);
  lj_assertA(!ra_used(ir), "unfused FREF");
}

static void asm_strref(ASMState *as, IRIns *ir)
{
  RegSet allow = RSET_GPR;
  Reg dest = ra_dest(as, ir, allow);
  Reg base = ra_alloc1(as, ir->op1, allow);
  IRIns *irr = IR(ir->op2);
  int32_t ofs = sizeof(GCstr);
  uint32_t m;
  rset_clear(allow, base);
  if (irref_isk(ir->op2) && (m = emit_isk12(ofs + irr->i))) {
    emit_dn(as, A64I_ADDx^m, dest, base);
  } else {
    emit_dn(as, (A64I_ADDx^A64I_K12) | A64F_U12(ofs), dest, dest);
    emit_dnm(as, A64I_ADDx, dest, base, ra_alloc1(as, ir->op2, allow));
  }
}

/* -- Loads and stores ---------------------------------------------------- */

static A64Ins asm_fxloadins(IRIns *ir)
{
  switch (irt_type(ir->t)) {
  case IRT_I8: return A64I_LDRB ^ A64I_LS_S;
  case IRT_U8: return A64I_LDRB;
  case IRT_I16: return A64I_LDRH ^ A64I_LS_S;
  case IRT_U16: return A64I_LDRH;
  case IRT_NUM: return A64I_LDRd;
  case IRT_FLOAT: return A64I_LDRs;
  default: return irt_is64(ir->t) ? A64I_LDRx : A64I_LDRw;
  }
}

static A64Ins asm_fxstoreins(IRIns *ir)
{
  switch (irt_type(ir->t)) {
  case IRT_I8: case IRT_U8: return A64I_STRB;
  case IRT_I16: case IRT_U16: return A64I_STRH;
  case IRT_NUM: return A64I_STRd;
  case IRT_FLOAT: return A64I_STRs;
  default: return irt_is64(ir->t) ? A64I_STRx : A64I_STRw;
  }
}

static void asm_fload(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg idx;
  A64Ins ai = asm_fxloadins(ir);
  int32_t ofs;
  if (ir->op1 == REF_NIL) {  /* FLOAD from GG_State with offset. */
    idx = RID_GL;
    ofs = (ir->op2 << 2) - GG_OFS(g);
  } else {
    idx = ra_alloc1(as, ir->op1, RSET_GPR);
    if (ir->op2 == IRFL_TAB_ARRAY) {
      ofs = asm_fuseabase(as, ir->op1);
      if (ofs) {  /* Turn the t->array load into an add for colocated arrays. */
	emit_dn(as, (A64I_ADDx^A64I_K12) | A64F_U12(ofs), dest, idx);
	return;
      }
    }
    ofs = field_ofs[ir->op2];
  }
  emit_lso(as, ai, (dest & 31), idx, ofs);
}

static void asm_fstore(ASMState *as, IRIns *ir)
{
  if (ir->r != RID_SINK) {
    Reg src = ra_alloc1(as, ir->op2, RSET_GPR);
    IRIns *irf = IR(ir->op1);
    Reg idx = ra_alloc1(as, irf->op1, rset_exclude(RSET_GPR, src));
    int32_t ofs = field_ofs[irf->op2];
    emit_lso(as, asm_fxstoreins(ir), (src & 31), idx, ofs);
  }
}

static void asm_xload(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, irt_isfp(ir->t) ? RSET_FPR : RSET_GPR);
  lj_assertA(!(ir->op2 & IRXLOAD_UNALIGNED), "unaligned XLOAD");
  asm_fusexref(as, asm_fxloadins(ir), dest, ir->op1, RSET_GPR);
}

static void asm_xstore(ASMState *as, IRIns *ir)
{
  if (ir->r != RID_SINK) {
    Reg src = ra_alloc1(as, ir->op2, irt_isfp(ir->t) ? RSET_FPR : RSET_GPR);
    asm_fusexref(as, asm_fxstoreins(ir), src, ir->op1,
		 rset_exclude(RSET_GPR, src));
  }
}

static void asm_ahuvload(ASMState *as, IRIns *ir)
{
  Reg idx, tmp;
  int32_t ofs = 0;
  RegSet gpr = RSET_GPR, allow = irt_isnum(ir->t) ? RSET_FPR : RSET_GPR;
  lj_assertA(irt_isnum(ir->t) || irt_ispri(ir->t) || irt_isaddr(ir->t) ||
	     irt_isint(ir->t),
	     "bad load type %d", irt_type(ir->t));
  if (ra_used(ir)) {
    Reg dest = ra_dest(as, ir, allow);
    tmp = irt_isnum(ir->t) ? ra_scratch(as, rset_clear(gpr, dest)) : dest;
    if (irt_isaddr(ir->t)) {
      emit_dn(as, A64I_ANDx^emit_isk13(LJ_GCVMASK, 1), dest, dest);
    } else if (irt_isnum(ir->t)) {
      emit_dn(as, A64I_FMOV_D_R, (dest & 31), tmp);
    } else if (irt_isint(ir->t)) {
      emit_dm(as, A64I_MOVw, dest, dest);
    }
  } else {
    tmp = ra_scratch(as, gpr);
  }
  idx = asm_fuseahuref(as, ir->op1, &ofs, rset_clear(gpr, tmp), A64I_LDRx);
  rset_clear(gpr, idx);
  if (ofs & FUSE_REG) rset_clear(gpr, ofs & 31);
  if (ir->o == IR_VLOAD) ofs += 8 * ir->op2;
  /* Always do the type check, even if the load result is unused. */
  asm_guardcc(as, irt_isnum(ir->t) ? CC_LS : CC_NE);
  if (irt_type(ir->t) >= IRT_NUM) {
    lj_assertA(irt_isinteger(ir->t) || irt_isnum(ir->t),
	       "bad load type %d", irt_type(ir->t));
    emit_nm(as, A64I_CMPx | A64F_SH(A64SH_LSR, 32),
	    ra_allock(as, LJ_TISNUM << 15, gpr), tmp);
  } else if (irt_isaddr(ir->t)) {
    emit_n(as, (A64I_CMNx^A64I_K12) | A64F_U12(-irt_toitype(ir->t)), RID_TMP);
    emit_dn(as, A64I_ASRx | A64F_IMMR(47), RID_TMP, tmp);
  } else if (irt_isnil(ir->t)) {
    emit_n(as, (A64I_CMNx^A64I_K12) | A64F_U12(1), tmp);
  } else {
    emit_nm(as, A64I_CMPx | A64F_SH(A64SH_LSR, 32),
	    ra_allock(as, (irt_toitype(ir->t) << 15) | 0x7fff, gpr), tmp);
  }
  if (ofs & FUSE_REG)
    emit_dnm(as, (A64I_LDRx^A64I_LS_R)|A64I_LS_UXTWx|A64I_LS_SH, tmp, idx, (ofs & 31));
  else
    emit_lso(as, A64I_LDRx, tmp, idx, ofs);
}

static void asm_ahustore(ASMState *as, IRIns *ir)
{
  if (ir->r != RID_SINK) {
    RegSet allow = RSET_GPR;
    Reg idx, src = RID_NONE, tmp = RID_TMP, type = RID_NONE;
    int32_t ofs = 0;
    if (irt_isnum(ir->t)) {
      src = ra_alloc1(as, ir->op2, RSET_FPR);
      idx = asm_fuseahuref(as, ir->op1, &ofs, allow, A64I_STRd);
      if (ofs & FUSE_REG)
	emit_dnm(as, (A64I_STRd^A64I_LS_R)|A64I_LS_UXTWx|A64I_LS_SH, (src & 31), idx, (ofs &31));
      else
	emit_lso(as, A64I_STRd, (src & 31), idx, ofs);
    } else {
      if (!irt_ispri(ir->t)) {
	src = ra_alloc1(as, ir->op2, allow);
	rset_clear(allow, src);
	if (irt_isinteger(ir->t))
	  type = ra_allock(as, (uint64_t)(int32_t)LJ_TISNUM << 47, allow);
	else
	  type = ra_allock(as, irt_toitype(ir->t), allow);
      } else {
	tmp = type = ra_allock(as, ~((int64_t)~irt_toitype(ir->t)<<47), allow);
      }
      idx = asm_fuseahuref(as, ir->op1, &ofs, rset_exclude(allow, type),
			   A64I_STRx);
      if (ofs & FUSE_REG)
	emit_dnm(as, (A64I_STRx^A64I_LS_R)|A64I_LS_UXTWx|A64I_LS_SH, tmp, idx, (ofs & 31));
      else
	emit_lso(as, A64I_STRx, tmp, idx, ofs);
      if (ra_hasreg(src)) {
	if (irt_isinteger(ir->t)) {
	  emit_dnm(as, A64I_ADDx | A64F_EX(A64EX_UXTW), tmp, type, src);
	} else {
	  emit_dnm(as, A64I_ADDx | A64F_SH(A64SH_LSL, 47), tmp, src, type);
	}
      }
    }
  }
}

static void asm_sload(ASMState *as, IRIns *ir)
{
  int32_t ofs = 8*((int32_t)ir->op1-2);
  IRType1 t = ir->t;
  Reg dest = RID_NONE, base;
  RegSet allow = RSET_GPR;
  lj_assertA(!(ir->op2 & IRSLOAD_PARENT),
	     "bad parent SLOAD");  /* Handled by asm_head_side(). */
  lj_assertA(irt_isguard(t) || !(ir->op2 & IRSLOAD_TYPECHECK),
	     "inconsistent SLOAD variant");
  if ((ir->op2 & IRSLOAD_CONVERT) && irt_isguard(t) && irt_isint(t)) {
    dest = ra_scratch(as, RSET_FPR);
    asm_tointg(as, ir, dest);
    t.irt = IRT_NUM;  /* Continue with a regular number type check. */
  } else if (ra_used(ir)) {
    Reg tmp = RID_NONE;
    if ((ir->op2 & IRSLOAD_CONVERT))
      tmp = ra_scratch(as, irt_isint(t) ? RSET_FPR : RSET_GPR);
    lj_assertA((irt_isnum(t)) || irt_isint(t) || irt_isaddr(t),
	       "bad SLOAD type %d", irt_type(t));
    dest = ra_dest(as, ir, irt_isnum(t) ? RSET_FPR : allow);
    base = ra_alloc1(as, REF_BASE, rset_clear(allow, dest));
    if (irt_isaddr(t)) {
      emit_dn(as, A64I_ANDx^emit_isk13(LJ_GCVMASK, 1), dest, dest);
    } else if ((ir->op2 & IRSLOAD_CONVERT)) {
      if (irt_isint(t)) {
	emit_dn(as, A64I_FCVT_S32_F64, dest, (tmp & 31));
	/* If value is already loaded for type check, move it to FPR. */
	if ((ir->op2 & IRSLOAD_TYPECHECK))
	  emit_dn(as, A64I_FMOV_D_R, (tmp & 31), dest);
	else
	  dest = tmp;
	t.irt = IRT_NUM;  /* Check for original type. */
      } else {
	emit_dn(as, A64I_FCVT_F64_S32, (dest & 31), tmp);
	dest = tmp;
	t.irt = IRT_INT;  /* Check for original type. */
      }
    } else if (irt_isint(t) && (ir->op2 & IRSLOAD_TYPECHECK)) {
      emit_dm(as, A64I_MOVw, dest, dest);
    }
    goto dotypecheck;
  }
  base = ra_alloc1(as, REF_BASE, allow);
dotypecheck:
  rset_clear(allow, base);
  if ((ir->op2 & IRSLOAD_TYPECHECK)) {
    Reg tmp;
    if (ra_hasreg(dest) && rset_test(RSET_GPR, dest)) {
      tmp = dest;
    } else {
      tmp = ra_scratch(as, allow);
      rset_clear(allow, tmp);
    }
    if (ra_hasreg(dest) && tmp != dest)
      emit_dn(as, A64I_FMOV_D_R, (dest & 31), tmp);
    /* Need type check, even if the load result is unused. */
    asm_guardcc(as, irt_isnum(t) ? CC_LS : CC_NE);
    if (irt_type(t) >= IRT_NUM) {
      lj_assertA(irt_isinteger(t) || irt_isnum(t),
		 "bad SLOAD type %d", irt_type(t));
      emit_nm(as, A64I_CMPx | A64F_SH(A64SH_LSR, 32),
	      ra_allock(as, (ir->op2 & IRSLOAD_KEYINDEX) ? LJ_KEYINDEX : (LJ_TISNUM << 15), allow), tmp);
    } else if (irt_isnil(t)) {
      emit_n(as, (A64I_CMNx^A64I_K12) | A64F_U12(1), tmp);
    } else if (irt_ispri(t)) {
      emit_nm(as, A64I_CMPx,
	      ra_allock(as, ~((int64_t)~irt_toitype(t) << 47) , allow), tmp);
    } else {
      emit_n(as, (A64I_CMNx^A64I_K12) | A64F_U12(-irt_toitype(t)), RID_TMP);
      emit_dn(as, A64I_ASRx | A64F_IMMR(47), RID_TMP, tmp);
    }
    emit_lso(as, A64I_LDRx, tmp, base, ofs);
    return;
  }
  if (ra_hasreg(dest)) {
    emit_lso(as, irt_isnum(t) ? A64I_LDRd :
	     (irt_isint(t) ? A64I_LDRw : A64I_LDRx), (dest & 31), base,
	     ofs ^ ((LJ_BE && irt_isint(t) ? 4 : 0)));
  }
}

/* -- Allocations --------------------------------------------------------- */

#if LJ_HASFFI
static void asm_cnew(ASMState *as, IRIns *ir)
{
  CTState *cts = ctype_ctsG(J2G(as->J));
  CTypeID id = (CTypeID)IR(ir->op1)->i;
  CTSize sz;
  CTInfo info = lj_ctype_info(cts, id, &sz);
  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_mem_newgco];
  IRRef args[4];
  RegSet allow = (RSET_GPR & ~RSET_SCRATCH);
  lj_assertA(sz != CTSIZE_INVALID || (ir->o == IR_CNEW && ir->op2 != REF_NIL),
	     "bad CNEW/CNEWI operands");

  as->gcsteps++;
  asm_setupresult(as, ir, ci);  /* GCcdata * */
  /* Initialize immutable cdata object. */
  if (ir->o == IR_CNEWI) {
    int32_t ofs = sizeof(GCcdata);
    Reg r = ra_alloc1(as, ir->op2, allow);
    lj_assertA(sz == 4 || sz == 8, "bad CNEWI size %d", sz);
    emit_lso(as, sz == 8 ? A64I_STRx : A64I_STRw, r, RID_RET, ofs);
  } else if (ir->op2 != REF_NIL) {  /* Create VLA/VLS/aligned cdata. */
    ci = &lj_ir_callinfo[IRCALL_lj_cdata_newv];
    args[0] = ASMREF_L;     /* lua_State *L */
    args[1] = ir->op1;      /* CTypeID id   */
    args[2] = ir->op2;      /* CTSize sz    */
    args[3] = ASMREF_TMP1;  /* CTSize align */
    asm_gencall(as, ci, args);
    emit_loadi(as, ra_releasetmp(as, ASMREF_TMP1), (int32_t)ctype_align(info));
    return;
  }

  /* Initialize gct and ctypeid. lj_mem_newgco() already sets marked. */
  {
    Reg r = (id < 65536) ? RID_X1 : ra_allock(as, id, allow);
    emit_lso(as, A64I_STRB, RID_TMP, RID_RET, offsetof(GCcdata, gct));
    emit_lso(as, A64I_STRH, r, RID_RET, offsetof(GCcdata, ctypeid));
    emit_d(as, A64I_MOVZw | A64F_U16(~LJ_TCDATA), RID_TMP);
    if (id < 65536) emit_d(as, A64I_MOVZw | A64F_U16(id), RID_X1);
  }
  args[0] = ASMREF_L;     /* lua_State *L */
  args[1] = ASMREF_TMP1;  /* MSize size   */
  asm_gencall(as, ci, args);
  ra_allockreg(as, (int32_t)(sz+sizeof(GCcdata)),
	       ra_releasetmp(as, ASMREF_TMP1));
}
#endif

/* -- Write barriers ------------------------------------------------------ */

static void asm_tbar(ASMState *as, IRIns *ir)
{
  Reg tab = ra_alloc1(as, ir->op1, RSET_GPR);
  Reg link = ra_scratch(as, rset_exclude(RSET_GPR, tab));
  Reg mark = RID_TMP;
  MCLabel l_end = emit_label(as);
  emit_lso(as, A64I_STRB, mark, tab, (int32_t)offsetof(GCtab, marked));
  /* Keep STRx in the middle to avoid LDP/STP fusion with surrounding code. */
  emit_lso(as, A64I_STRx, link, tab, (int32_t)offsetof(GCtab, gclist));
  emit_setgl(as, tab, gc.grayagain);
  emit_dn(as, A64I_ANDw^emit_isk13(~LJ_GC_BLACK, 0), mark, mark);
  emit_getgl(as, link, gc.grayagain);
  emit_cond_branch(as, CC_EQ, l_end);
  emit_n(as, A64I_TSTw^emit_isk13(LJ_GC_BLACK, 0), mark);
  emit_lso(as, A64I_LDRB, mark, tab, (int32_t)offsetof(GCtab, marked));
}

static void asm_obar(ASMState *as, IRIns *ir)
{
  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_gc_barrieruv];
  IRRef args[2];
  MCLabel l_end;
  Reg obj, val, tmp;
  /* No need for other object barriers (yet). */
  lj_assertA(IR(ir->op1)->o == IR_UREFC, "bad OBAR type");
  ra_evictset(as, RSET_SCRATCH);
  l_end = emit_label(as);
  args[0] = ASMREF_TMP1;  /* global_State *g */
  args[1] = ir->op1;      /* TValue *tv      */
  asm_gencall(as, ci, args);
  emit_dm(as, A64I_MOVx, ra_releasetmp(as, ASMREF_TMP1), RID_GL);
  obj = IR(ir->op1)->r;
  tmp = ra_scratch(as, rset_exclude(RSET_GPR, obj));
  emit_tnb(as, A64I_TBZ, tmp, lj_ffs(LJ_GC_BLACK), l_end);
  emit_cond_branch(as, CC_EQ, l_end);
  emit_n(as, A64I_TSTw^emit_isk13(LJ_GC_WHITES, 0), RID_TMP);
  val = ra_alloc1(as, ir->op2, rset_exclude(RSET_GPR, obj));
  emit_lso(as, A64I_LDRB, tmp, obj,
	   (int32_t)offsetof(GCupval, marked)-(int32_t)offsetof(GCupval, tv));
  emit_lso(as, A64I_LDRB, RID_TMP, val, (int32_t)offsetof(GChead, marked));
}

/* -- Arithmetic and logic operations ------------------------------------- */

static void asm_fparith(ASMState *as, IRIns *ir, A64Ins ai)
{
  Reg dest = ra_dest(as, ir, RSET_FPR);
  Reg right, left = ra_alloc2(as, ir, RSET_FPR);
  right = (left >> 8); left &= 255;
  emit_dnm(as, ai, (dest & 31), (left & 31), (right & 31));
}

static void asm_fpunary(ASMState *as, IRIns *ir, A64Ins ai)
{
  Reg dest = ra_dest(as, ir, RSET_FPR);
  Reg left = ra_hintalloc(as, ir->op1, dest, RSET_FPR);
  emit_dn(as, ai, (dest & 31), (left & 31));
}

static void asm_fpmath(ASMState *as, IRIns *ir)
{
  IRFPMathOp fpm = (IRFPMathOp)ir->op2;
  if (fpm == IRFPM_SQRT) {
    asm_fpunary(as, ir, A64I_FSQRTd);
  } else if (fpm <= IRFPM_TRUNC) {
    asm_fpunary(as, ir, fpm == IRFPM_FLOOR ? A64I_FRINTMd :
			fpm == IRFPM_CEIL ? A64I_FRINTPd : A64I_FRINTZd);
  } else {
    asm_callid(as, ir, IRCALL_lj_vm_floor + fpm);
  }
}

static int asm_swapops(ASMState *as, IRRef lref, IRRef rref)
{
  IRIns *ir;
  if (irref_isk(rref))
    return 0;  /* Don't swap constants to the left. */
  if (irref_isk(lref))
    return 1;  /* But swap constants to the right. */
  ir = IR(rref);
  if ((ir->o >= IR_BSHL && ir->o <= IR_BROR) ||
      (ir->o == IR_ADD && ir->op1 == ir->op2) ||
      (ir->o == IR_CONV && ir->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT)))
    return 0;  /* Don't swap fusable operands to the left. */
  ir = IR(lref);
  if ((ir->o >= IR_BSHL && ir->o <= IR_BROR) ||
      (ir->o == IR_ADD && ir->op1 == ir->op2) ||
      (ir->o == IR_CONV && ir->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT)))
    return 1;  /* But swap fusable operands to the right. */
  return 0;  /* Otherwise don't swap. */
}

static void asm_intop(ASMState *as, IRIns *ir, A64Ins ai)
{
  IRRef lref = ir->op1, rref = ir->op2;
  Reg left, dest = ra_dest(as, ir, RSET_GPR);
  uint32_t m;
  if ((ai & ~A64I_S) != A64I_SUBw && asm_swapops(as, lref, rref)) {
    IRRef tmp = lref; lref = rref; rref = tmp;
  }
  left = ra_hintalloc(as, lref, dest, RSET_GPR);
  if (irt_is64(ir->t)) ai |= A64I_X;
  m = asm_fuseopm(as, ai, rref, rset_exclude(RSET_GPR, left));
  if (irt_isguard(ir->t)) {  /* For IR_ADDOV etc. */
    asm_guardcc(as, CC_VS);
    ai |= A64I_S;
  }
  emit_dn(as, ai^m, dest, left);
}

static void asm_intop_s(ASMState *as, IRIns *ir, A64Ins ai)
{
  if (as->flagmcp == as->mcp) {  /* Drop cmp r, #0. */
    as->flagmcp = NULL;
    as->mcp++;
    ai |= A64I_S;
  }
  asm_intop(as, ir, ai);
}

static void asm_intneg(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg left = ra_hintalloc(as, ir->op1, dest, RSET_GPR);
  emit_dm(as, irt_is64(ir->t) ? A64I_NEGx : A64I_NEGw, dest, left);
}

/* NYI: use add/shift for MUL(OV) with constants. FOLD only does 2^k. */
static void asm_intmul(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
  Reg right = ra_alloc1(as, ir->op2, rset_exclude(RSET_GPR, left));
  if (irt_isguard(ir->t)) {  /* IR_MULOV */
    asm_guardcc(as, CC_NE);
    emit_dm(as, A64I_MOVw, dest, dest);  /* Zero-extend. */
    emit_nm(as, A64I_CMPx | A64F_EX(A64EX_SXTW), dest, dest);
    emit_dnm(as, A64I_SMULL, dest, right, left);
  } else {
    emit_dnm(as, irt_is64(ir->t) ? A64I_MULx : A64I_MULw, dest, left, right);
  }
}

static void asm_add(ASMState *as, IRIns *ir)
{
  if (irt_isnum(ir->t)) {
    if (!asm_fusemadd(as, ir, A64I_FMADDd, A64I_FMADDd))
      asm_fparith(as, ir, A64I_FADDd);
    return;
  }
  asm_intop_s(as, ir, A64I_ADDw);
}

static void asm_sub(ASMState *as, IRIns *ir)
{
  if (irt_isnum(ir->t)) {
    if (!asm_fusemadd(as, ir, A64I_FNMSUBd, A64I_FMSUBd))
      asm_fparith(as, ir, A64I_FSUBd);
    return;
  }
  asm_intop_s(as, ir, A64I_SUBw);
}

static void asm_mul(ASMState *as, IRIns *ir)
{
  if (irt_isnum(ir->t)) {
    asm_fparith(as, ir, A64I_FMULd);
    return;
  }
  asm_intmul(as, ir);
}

#define asm_addov(as, ir)	asm_add(as, ir)
#define asm_subov(as, ir)	asm_sub(as, ir)
#define asm_mulov(as, ir)	asm_mul(as, ir)

#define asm_fpdiv(as, ir)	asm_fparith(as, ir, A64I_FDIVd)
#define asm_abs(as, ir)		asm_fpunary(as, ir, A64I_FABS)

static void asm_neg(ASMState *as, IRIns *ir)
{
  if (irt_isnum(ir->t)) {
    asm_fpunary(as, ir, A64I_FNEGd);
    return;
  }
  asm_intneg(as, ir);
}

static void asm_band(ASMState *as, IRIns *ir)
{
  A64Ins ai = A64I_ANDw;
  if (asm_fuseandshift(as, ir))
    return;
  if (as->flagmcp == as->mcp) {
    /* Try to drop cmp r, #0. */
    as->flagmcp = NULL;
    as->mcp++;
    ai = A64I_ANDSw;
  }
  asm_intop(as, ir, ai);
}

static void asm_borbxor(ASMState *as, IRIns *ir, A64Ins ai)
{
  IRRef lref = ir->op1, rref = ir->op2;
  IRIns *irl = IR(lref), *irr = IR(rref);
  if ((canfuse(as, irl) && irl->o == IR_BNOT && !irref_isk(rref)) ||
      (canfuse(as, irr) && irr->o == IR_BNOT && !irref_isk(lref))) {
    Reg left, dest = ra_dest(as, ir, RSET_GPR);
    uint32_t m;
    if (irl->o == IR_BNOT) {
      IRRef tmp = lref; lref = rref; rref = tmp;
    }
    left = ra_alloc1(as, lref, RSET_GPR);
    ai |= A64I_ON;
    if (irt_is64(ir->t)) ai |= A64I_X;
    m = asm_fuseopm(as, ai, IR(rref)->op1, rset_exclude(RSET_GPR, left));
    emit_dn(as, ai^m, dest, left);
  } else {
    asm_intop(as, ir, ai);
  }
}

static void asm_bor(ASMState *as, IRIns *ir)
{
  if (asm_fuseorshift(as, ir))
    return;
  asm_borbxor(as, ir, A64I_ORRw);
}

#define asm_bxor(as, ir)	asm_borbxor(as, ir, A64I_EORw)

static void asm_bnot(ASMState *as, IRIns *ir)
{
  A64Ins ai = A64I_MVNw;
  Reg dest = ra_dest(as, ir, RSET_GPR);
  uint32_t m = asm_fuseopm(as, ai, ir->op1, RSET_GPR);
  if (irt_is64(ir->t)) ai |= A64I_X;
  emit_d(as, ai^m, dest);
}

static void asm_bswap(ASMState *as, IRIns *ir)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
  emit_dn(as, irt_is64(ir->t) ? A64I_REVx : A64I_REVw, dest, left);
}

static void asm_bitshift(ASMState *as, IRIns *ir, A64Ins ai, A64Shift sh)
{
  int32_t shmask = irt_is64(ir->t) ? 63 : 31;
  if (irref_isk(ir->op2)) {  /* Constant shifts. */
    Reg left, dest = ra_dest(as, ir, RSET_GPR);
    int32_t shift = (IR(ir->op2)->i & shmask);
    IRIns *irl = IR(ir->op1);
    if (shmask == 63) ai += A64I_UBFMx - A64I_UBFMw;

    /* Fuse BSHL + BSHR/BSAR into UBFM/SBFM aka UBFX/SBFX/UBFIZ/SBFIZ. */
    if ((sh == A64SH_LSR || sh == A64SH_ASR) && canfuse(as, irl)) {
      if (irl->o == IR_BSHL && irref_isk(irl->op2)) {
	int32_t shift2 = (IR(irl->op2)->i & shmask);
	shift = ((shift - shift2) & shmask);
	shmask -= shift2;
	ir = irl;
      }
    }

    left = ra_alloc1(as, ir->op1, RSET_GPR);
    switch (sh) {
    case A64SH_LSL:
      emit_dn(as, ai | A64F_IMMS(shmask-shift) |
		  A64F_IMMR((shmask-shift+1)&shmask), dest, left);
      break;
    case A64SH_LSR: case A64SH_ASR:
      emit_dn(as, ai | A64F_IMMS(shmask) | A64F_IMMR(shift), dest, left);
      break;
    case A64SH_ROR:
      emit_dnm(as, ai | A64F_IMMS(shift), dest, left, left);
      break;
    }
  } else {  /* Variable-length shifts. */
    Reg dest = ra_dest(as, ir, RSET_GPR);
    Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
    Reg right = ra_alloc1(as, ir->op2, rset_exclude(RSET_GPR, left));
    emit_dnm(as, (shmask == 63 ? A64I_SHRx : A64I_SHRw) | A64F_BSH(sh), dest, left, right);
  }
}

#define asm_bshl(as, ir)	asm_bitshift(as, ir, A64I_UBFMw, A64SH_LSL)
#define asm_bshr(as, ir)	asm_bitshift(as, ir, A64I_UBFMw, A64SH_LSR)
#define asm_bsar(as, ir)	asm_bitshift(as, ir, A64I_SBFMw, A64SH_ASR)
#define asm_bror(as, ir)	asm_bitshift(as, ir, A64I_EXTRw, A64SH_ROR)
#define asm_brol(as, ir)	lj_assertA(0, "unexpected BROL")

static void asm_intmin_max(ASMState *as, IRIns *ir, A64CC cc)
{
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg left = ra_hintalloc(as, ir->op1, dest, RSET_GPR);
  Reg right = ra_alloc1(as, ir->op2, rset_exclude(RSET_GPR, left));
  emit_dnm(as, A64I_CSELw|A64F_CC(cc), dest, left, right);
  emit_nm(as, A64I_CMPw, left, right);
}

static void asm_fpmin_max(ASMState *as, IRIns *ir, A64CC fcc)
{
  Reg dest = (ra_dest(as, ir, RSET_FPR) & 31);
  Reg right, left = ra_alloc2(as, ir, RSET_FPR);
  right = ((left >> 8) & 31); left &= 31;
  emit_dnm(as, A64I_FCSELd | A64F_CC(fcc), dest, right, left);
  emit_nm(as, A64I_FCMPd, left, right);
}

static void asm_min_max(ASMState *as, IRIns *ir, A64CC cc, A64CC fcc)
{
  if (irt_isnum(ir->t))
    asm_fpmin_max(as, ir, fcc);
  else
    asm_intmin_max(as, ir, cc);
}

#define asm_min(as, ir)		asm_min_max(as, ir, CC_LT, CC_PL)
#define asm_max(as, ir)		asm_min_max(as, ir, CC_GT, CC_LE)

/* -- Comparisons --------------------------------------------------------- */

/* Map of comparisons to flags. ORDER IR. */
static const uint8_t asm_compmap[IR_ABC+1] = {
  /* op  FP swp  int cc   FP cc */
  /* LT       */ CC_GE + (CC_HS << 4),
  /* GE    x  */ CC_LT + (CC_HI << 4),
  /* LE       */ CC_GT + (CC_HI << 4),
  /* GT    x  */ CC_LE + (CC_HS << 4),
  /* ULT   x  */ CC_HS + (CC_LS << 4),
  /* UGE      */ CC_LO + (CC_LO << 4),
  /* ULE   x  */ CC_HI + (CC_LO << 4),
  /* UGT      */ CC_LS + (CC_LS << 4),
  /* EQ       */ CC_NE + (CC_NE << 4),
  /* NE       */ CC_EQ + (CC_EQ << 4),
  /* ABC      */ CC_LS + (CC_LS << 4)  /* Same as UGT. */
};

/* FP comparisons. */
static void asm_fpcomp(ASMState *as, IRIns *ir)
{
  Reg left, right;
  A64Ins ai;
  int swp = ((ir->o ^ (ir->o >> 2)) & ~(ir->o >> 3) & 1);
  if (!swp && irref_isk(ir->op2) && ir_knum(IR(ir->op2))->u64 == 0) {
    left = (ra_alloc1(as, ir->op1, RSET_FPR) & 31);
    right = 0;
    ai = A64I_FCMPZd;
  } else {
    left = ra_alloc2(as, ir, RSET_FPR);
    if (swp) {
      right = (left & 31); left = ((left >> 8) & 31);
    } else {
      right = ((left >> 8) & 31); left &= 31;
    }
    ai = A64I_FCMPd;
  }
  asm_guardcc(as, (asm_compmap[ir->o] >> 4));
  emit_nm(as, ai, left, right);
}

/* Integer comparisons. */
static void asm_intcomp(ASMState *as, IRIns *ir)
{
  A64CC oldcc, cc = (asm_compmap[ir->o] & 15);
  A64Ins ai = irt_is64(ir->t) ? A64I_CMPx : A64I_CMPw;
  IRRef lref = ir->op1, rref = ir->op2;
  Reg left;
  uint32_t m;
  int cmpprev0 = 0;
  lj_assertA(irt_is64(ir->t) || irt_isint(ir->t) ||
	     irt_isu32(ir->t) || irt_isaddr(ir->t) || irt_isu8(ir->t),
	     "bad comparison data type %d", irt_type(ir->t));
  if (asm_swapops(as, lref, rref)) {
    IRRef tmp = lref; lref = rref; rref = tmp;
    if (cc >= CC_GE) cc ^= 7;  /* LT <-> GT, LE <-> GE */
    else if (cc > CC_NE) cc ^= 11;  /* LO <-> HI, LS <-> HS */
  }
  oldcc = cc;
  if (irref_isk(rref) && get_k64val(as, rref) == 0) {
    IRIns *irl = IR(lref);
    if (cc == CC_GE) cc = CC_PL;
    else if (cc == CC_LT) cc = CC_MI;
    else if (cc > CC_NE) goto nocombine;  /* Other conds don't work with tst. */
    cmpprev0 = (irl+1 == ir);
    /* Combine and-cmp-bcc into tbz/tbnz or and-cmp into tst. */
    if (cmpprev0 && irl->o == IR_BAND && !ra_used(irl)) {
      IRRef blref = irl->op1, brref = irl->op2;
      uint32_t m2 = 0;
      Reg bleft;
      if (asm_swapops(as, blref, brref)) {
	Reg tmp = blref; blref = brref; brref = tmp;
      }
      bleft = ra_alloc1(as, blref, RSET_GPR);
      if (irref_isk(brref)) {
	uint64_t k = get_k64val(as, brref);
	if (k && !(k & (k-1)) && (cc == CC_EQ || cc == CC_NE) &&
	    asm_guardtnb(as, cc == CC_EQ ? A64I_TBZ : A64I_TBNZ, bleft,
			 emit_ctz64(k)))
	  return;
	m2 = emit_isk13(k, irt_is64(irl->t));
      }
      ai = (irt_is64(irl->t) ? A64I_TSTx : A64I_TSTw);
      if (!m2)
	m2 = asm_fuseopm(as, ai, brref, rset_exclude(RSET_GPR, bleft));
      asm_guardcc(as, cc);
      emit_n(as, ai^m2, bleft);
      return;
    }
    if (cc == CC_EQ || cc == CC_NE) {
      /* Combine cmp-bcc into cbz/cbnz. */
      ai = cc == CC_EQ ? A64I_CBZ : A64I_CBNZ;
      if (irt_is64(ir->t)) ai |= A64I_X;
      asm_guardcnb(as, ai, ra_alloc1(as, lref, RSET_GPR));
      return;
    }
  }
nocombine:
  left = ra_alloc1(as, lref, RSET_GPR);
  m = asm_fuseopm(as, ai, rref, rset_exclude(RSET_GPR, left));
  asm_guardcc(as, cc);
  emit_n(as, ai^m, left);
  /* Signed comparison with zero and referencing previous ins? */
  if (cmpprev0 && (oldcc <= CC_NE || oldcc >= CC_GE))
    as->flagmcp = as->mcp;  /* Allow elimination of the compare. */
}

static void asm_comp(ASMState *as, IRIns *ir)
{
  if (irt_isnum(ir->t))
    asm_fpcomp(as, ir);
  else
    asm_intcomp(as, ir);
}

#define asm_equal(as, ir)	asm_comp(as, ir)

/* -- Split register ops -------------------------------------------------- */

/* Hiword op of a split 64/64 bit op. Previous op is the loword op. */
static void asm_hiop(ASMState *as, IRIns *ir)
{
  /* HIOP is marked as a store because it needs its own DCE logic. */
  int uselo = ra_used(ir-1), usehi = ra_used(ir);  /* Loword/hiword used? */
  if (LJ_UNLIKELY(!(as->flags & JIT_F_OPT_DCE))) uselo = usehi = 1;
  if (!usehi) return;  /* Skip unused hiword op for all remaining ops. */
  switch ((ir-1)->o) {
  case IR_CALLN:
  case IR_CALLL:
  case IR_CALLS:
  case IR_CALLXS:
    if (!uselo)
      ra_allocref(as, ir->op1, RID2RSET(RID_RETLO));  /* Mark lo op as used. */
    break;
  default: lj_assertA(0, "bad HIOP for op %d", (ir-1)->o); break;
  }
}

/* -- Profiling ----------------------------------------------------------- */

static void asm_prof(ASMState *as, IRIns *ir)
{
  uint32_t k = emit_isk13(HOOK_PROFILE, 0);
  lj_assertA(k != 0, "HOOK_PROFILE does not fit in K13");
  UNUSED(ir);
  asm_guardcc(as, CC_NE);
  emit_n(as, A64I_TSTw^k, RID_TMP);
  emit_lsptr(as, A64I_LDRB, RID_TMP, (void *)&J2G(as->J)->hookmask);
}

/* -- Stack handling ------------------------------------------------------ */

/* Check Lua stack size for overflow. Use exit handler as fallback. */
static void asm_stack_check(ASMState *as, BCReg topslot,
			    IRIns *irp, RegSet allow, ExitNo exitno)
{
  uint32_t k;
  Reg pbase = RID_BASE;
  if (irp) {
    pbase = irp->r;
    if (!ra_hasreg(pbase))
      pbase = allow ? (0x40 | rset_pickbot(allow)) : (0xC0 | RID_RET);
  }
  emit_cond_branch(as, CC_LS, asm_exitstub_addr(as, exitno));
  if (pbase & 0x80)  /* Restore temp. register. */
    emit_lso(as, A64I_LDRx, (pbase & 31), RID_SP, 0);
  k = emit_isk12((8*topslot));
  lj_assertA(k, "slot offset %d does not fit in K12", 8*topslot);
  emit_n(as, A64I_CMPx^k, RID_TMP);
  emit_dnm(as, A64I_SUBx, RID_TMP, RID_TMP, (pbase & 31));
  emit_lso(as, A64I_LDRx, RID_TMP, RID_TMP,
	   (int32_t)offsetof(lua_State, maxstack));
  if (pbase & 0x40) {
    emit_getgl(as, (pbase & 31), jit_base);
    if (pbase & 0x80)  /* Save temp register. */
      emit_lso(as, A64I_STRx, (pbase & 31), RID_SP, 0);
  }
  emit_getgl(as, RID_TMP, cur_L);
}

/* Restore Lua stack from on-trace state. */
static void asm_stack_restore(ASMState *as, SnapShot *snap)
{
  SnapEntry *map = &as->T->snapmap[snap->mapofs];
#ifdef LUA_USE_ASSERT
  SnapEntry *flinks = &as->T->snapmap[snap_nextofs(as->T, snap)-1-LJ_FR2];
#endif
  MSize n, nent = snap->nent;
  /* Store the value of all modified slots to the Lua stack. */
  for (n = 0; n < nent; n++) {
    SnapEntry sn = map[n];
    BCReg s = snap_slot(sn);
    int32_t ofs = 8*((int32_t)s-1-LJ_FR2);
    IRRef ref = snap_ref(sn);
    IRIns *ir = IR(ref);
    if ((sn & SNAP_NORESTORE))
      continue;
    if ((sn & SNAP_KEYINDEX)) {
      RegSet allow = rset_exclude(RSET_GPR, RID_BASE);
      Reg r = irref_isk(ref) ? ra_allock(as, ir->i, allow) :
			       ra_alloc1(as, ref, allow);
      rset_clear(allow, r);
      emit_lso(as, A64I_STRw, r, RID_BASE, ofs);
      emit_lso(as, A64I_STRw, ra_allock(as, LJ_KEYINDEX, allow), RID_BASE, ofs+4);
    } else if (irt_isnum(ir->t)) {
      Reg src = ra_alloc1(as, ref, RSET_FPR);
      emit_lso(as, A64I_STRd, (src & 31), RID_BASE, ofs);
    } else {
      asm_tvstore64(as, RID_BASE, ofs, ref);
    }
    checkmclim(as);
  }
  lj_assertA(map + nent == flinks, "inconsistent frames in snapshot");
}

/* -- GC handling --------------------------------------------------------- */

/* Marker to prevent patching the GC check exit. */
#define ARM64_NOPATCH_GC_CHECK \
  (A64I_ORRx|A64F_D(RID_ZERO)|A64F_M(RID_ZERO)|A64F_N(RID_ZERO))

/* Check GC threshold and do one or more GC steps. */
static void asm_gc_check(ASMState *as)
{
  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_gc_step_jit];
  IRRef args[2];
  MCLabel l_end;
  Reg tmp2;
  ra_evictset(as, RSET_SCRATCH);
  l_end = emit_label(as);
  /* Exit trace if in GCSatomic or GCSfinalize. Avoids syncing GC objects. */
  asm_guardcnb(as, A64I_CBNZ, RID_RET); /* Assumes asm_snap_prep() is done. */
  *--as->mcp = ARM64_NOPATCH_GC_CHECK;
  args[0] = ASMREF_TMP1;  /* global_State *g */
  args[1] = ASMREF_TMP2;  /* MSize steps     */
  asm_gencall(as, ci, args);
  emit_dm(as, A64I_MOVx, ra_releasetmp(as, ASMREF_TMP1), RID_GL);
  tmp2 = ra_releasetmp(as, ASMREF_TMP2);
  emit_loadi(as, tmp2, as->gcsteps);
  /* Jump around GC step if GC total < GC threshold. */
  emit_cond_branch(as, CC_LS, l_end);
  emit_nm(as, A64I_CMPx, RID_TMP, tmp2);
  emit_getgl(as, tmp2, gc.threshold);
  emit_getgl(as, RID_TMP, gc.total);
  as->gcsteps = 0;
  checkmclim(as);
}

/* -- Loop handling ------------------------------------------------------- */

/* Fixup the loop branch. */
static void asm_loop_fixup(ASMState *as)
{
  MCode *p = as->mctop;
  MCode *target = as->mcp;
  if (as->loopinv) {  /* Inverted loop branch? */
    uint32_t mask = (p[-2] & 0x7e000000) == 0x36000000 ? 0x3fffu : 0x7ffffu;
    ptrdiff_t delta = target - (p - 2);
    /* asm_guard* already inverted the bcc/tnb/cnb and patched the final b. */
    p[-2] |= ((uint32_t)delta & mask) << 5;
  } else {
    ptrdiff_t delta = target - (p - 1);
    p[-1] = A64I_B | A64F_S26(delta);
  }
}

/* Fixup the tail of the loop. */
static void asm_loop_tail_fixup(ASMState *as)
{
  UNUSED(as);  /* Nothing to do. */
}

/* -- Head of trace ------------------------------------------------------- */

/* Coalesce BASE register for a root trace. */
static void asm_head_root_base(ASMState *as)
{
  IRIns *ir = IR(REF_BASE);
  Reg r = ir->r;
  if (ra_hasreg(r)) {
    ra_free(as, r);
    if (rset_test(as->modset, r) || irt_ismarked(ir->t))
      ir->r = RID_INIT;  /* No inheritance for modified BASE register. */
    if (r != RID_BASE)
      emit_movrr(as, ir, r, RID_BASE);
  }
}

/* Coalesce BASE register for a side trace. */
static Reg asm_head_side_base(ASMState *as, IRIns *irp)
{
  IRIns *ir = IR(REF_BASE);
  Reg r = ir->r;
  if (ra_hasreg(r)) {
    ra_free(as, r);
    if (rset_test(as->modset, r) || irt_ismarked(ir->t))
      ir->r = RID_INIT;  /* No inheritance for modified BASE register. */
    if (irp->r == r) {
      return r;  /* Same BASE register already coalesced. */
    } else if (ra_hasreg(irp->r) && rset_test(as->freeset, irp->r)) {
      /* Move from coalesced parent reg. */
      emit_movrr(as, ir, r, irp->r);
      return irp->r;
    } else {
      emit_getgl(as, r, jit_base);  /* Otherwise reload BASE. */
    }
  }
  return RID_NONE;
}

/* -- Tail of trace ------------------------------------------------------- */

/* Fixup the tail code. */
static void asm_tail_fixup(ASMState *as, TraceNo lnk)
{
  MCode *p = as->mctop;
  MCode *target;
  /* Undo the sp adjustment in BC_JLOOP when exiting to the interpreter. */
  int32_t spadj = as->T->spadjust + (lnk ? 0 : sps_scale(SPS_FIXED));
  if (spadj == 0) {
    *--p = A64I_LE(A64I_NOP);
    as->mctop = p;
  } else {
    /* Patch stack adjustment. */
    uint32_t k = emit_isk12(spadj);
    lj_assertA(k, "stack adjustment %d does not fit in K12", spadj);
    p[-2] = (A64I_ADDx^k) | A64F_D(RID_SP) | A64F_N(RID_SP);
  }
  /* Patch exit branch. */
  target = lnk ? traceref(as->J, lnk)->mcode : (MCode *)lj_vm_exit_interp;
  p[-1] = A64I_B | A64F_S26((target-p)+1);
}

/* Prepare tail of code. */
static void asm_tail_prep(ASMState *as)
{
  MCode *p = as->mctop - 1;  /* Leave room for exit branch. */
  if (as->loopref) {
    as->invmcp = as->mcp = p;
  } else {
    as->mcp = p-1;  /* Leave room for stack pointer adjustment. */
    as->invmcp = NULL;
  }
  *p = 0;  /* Prevent load/store merging. */
}

/* -- Trace setup --------------------------------------------------------- */

/* Ensure there are enough stack slots for call arguments. */
static Reg asm_setup_call_slots(ASMState *as, IRIns *ir, const CCallInfo *ci)
{
#if LJ_HASFFI
  uint32_t i, nargs = CCI_XNARGS(ci);
  if (nargs > (REGARG_NUMGPR < REGARG_NUMFPR ? REGARG_NUMGPR : REGARG_NUMFPR) ||
      (LJ_TARGET_OSX && (ci->flags & CCI_VARARG))) {
    IRRef args[CCI_NARGS_MAX*2];
    int ngpr = REGARG_NUMGPR, nfpr = REGARG_NUMFPR;
    int spofs = 0, spalign = LJ_TARGET_OSX ? 0 : 7, nslots;
    asm_collectargs(as, ir, ci, args);
#if LJ_ABI_WIN
    if ((ci->flags & CCI_VARARG)) nfpr = 0;
#endif
    for (i = 0; i < nargs; i++) {
      int al = spalign;
      if (!args[i]) {
#if LJ_TARGET_OSX
	/* Marker for start of varaargs. */
	nfpr = 0;
	ngpr = 0;
	spalign = 7;
#endif
      } else if (irt_isfp(IR(args[i])->t)) {
	if (nfpr > 0) { nfpr--; continue; }
#if LJ_ABI_WIN
	if ((ci->flags & CCI_VARARG) && ngpr > 0) { ngpr--; continue; }
#elif LJ_TARGET_OSX
	al |= irt_isnum(IR(args[i])->t) ? 7 : 3;
#endif
      } else {
	if (ngpr > 0) { ngpr--; continue; }
#if LJ_TARGET_OSX
	al |= irt_size(IR(args[i])->t) - 1;
#endif
      }
      spofs = (spofs + 2*al+1) & ~al;  /* Align and bump stack pointer. */
    }
    nslots = (spofs + 3) >> 2;
    if (nslots > as->evenspill)  /* Leave room for args in stack slots. */
      as->evenspill = nslots;
  }
#endif
  return REGSP_HINT(irt_isfp(ir->t) ? RID_FPRET : RID_RET);
}

static void asm_setup_target(ASMState *as)
{
  /* May need extra exit for asm_stack_check on side traces. */
  asm_exitstub_setup(as, as->T->nsnap + (as->parent ? 1 : 0));
}

#if LJ_BE
/* ARM64 instructions are always little-endian. Swap for ARM64BE. */
static void asm_mcode_fixup(MCode *mcode, MSize size)
{
  MCode *pe = (MCode *)((char *)mcode + size);
  while (mcode < pe) {
    MCode ins = *mcode;
    *mcode++ = lj_bswap(ins);
  }
}
#define LJ_TARGET_MCODE_FIXUP	1
#endif

/* -- Trace patching ------------------------------------------------------ */

/* Patch exit jumps of existing machine code to a new target. */
void lj_asm_patchexit(jit_State *J, GCtrace *T, ExitNo exitno, MCode *target)
{
  MCode *p = T->mcode;
  MCode *pe = (MCode *)((char *)p + T->szmcode);
  MCode *cstart = NULL;
  MCode *mcarea = lj_mcode_patch(J, p, 0);
  MCode *px = exitstub_trace_addr(T, exitno);
  int patchlong = 1;
  /* Note: this assumes a trace exit is only ever patched once. */
  for (; p < pe; p++) {
    /* Look for exitstub branch, replace with branch to target. */
    ptrdiff_t delta = target - p;
    MCode ins = A64I_LE(*p);
    if ((ins & 0xff000000u) == 0x54000000u &&
	((ins ^ ((px-p)<<5)) & 0x00ffffe0u) == 0) {
      /* Patch bcc, if within range. */
      if (A64F_S_OK(delta, 19)) {
	*p = A64I_LE((ins & 0xff00001fu) | A64F_S19(delta));
	if (!cstart) cstart = p;
      }
    } else if ((ins & 0xfc000000u) == 0x14000000u &&
	       ((ins ^ (px-p)) & 0x03ffffffu) == 0) {
      /* Patch b. */
      lj_assertJ(A64F_S_OK(delta, 26), "branch target out of range");
      *p = A64I_LE((ins & 0xfc000000u) | A64F_S26(delta));
      if (!cstart) cstart = p;
    } else if ((ins & 0x7e000000u) == 0x34000000u &&
	       ((ins ^ ((px-p)<<5)) & 0x00ffffe0u) == 0) {
      /* Patch cbz/cbnz, if within range. */
      if (p[-1] == ARM64_NOPATCH_GC_CHECK) {
	patchlong = 0;
      } else if (A64F_S_OK(delta, 19)) {
	*p = A64I_LE((ins & 0xff00001fu) | A64F_S19(delta));
	if (!cstart) cstart = p;
      }
    } else if ((ins & 0x7e000000u) == 0x36000000u &&
	       ((ins ^ ((px-p)<<5)) & 0x0007ffe0u) == 0) {
      /* Patch tbz/tbnz, if within range. */
      if (A64F_S_OK(delta, 14)) {
	*p = A64I_LE((ins & 0xfff8001fu) | A64F_S14(delta));
	if (!cstart) cstart = p;
      }
    }
  }
  /* Always patch long-range branch in exit stub itself. Except, if we can't. */
  if (patchlong) {
    ptrdiff_t delta = target - px;
    lj_assertJ(A64F_S_OK(delta, 26), "branch target out of range");
    *px = A64I_B | A64F_S26(delta);
    if (!cstart) cstart = px;
  }
  if (cstart) lj_mcode_sync(cstart, px+1);
  lj_mcode_patch(J, mcarea, 1);
}