path: root/src/vm_arm64.dasc

|// Low-level VM code for ARM64 CPUs.
|// Bytecode interpreter, fast functions and helper functions.
|// Copyright (C) 2005-2022 Mike Pall. See Copyright Notice in luajit.h
|
|.arch arm64
|.section code_op, code_sub
|
|.actionlist build_actionlist
|.globals GLOB_
|.globalnames globnames
|.externnames extnames
|
|// Note: The ragged indentation of the instructions is intentional.
|//       The starting columns indicate data dependencies.
|
|//-----------------------------------------------------------------------
|
|// ARM64 registers and the AAPCS64 ABI 1.0 at a glance:
|//
|// x0-x17 temp, x19-x28 callee-saved, x29 fp, x30 lr
|// x18 is reserved on most platforms. Don't use it, save it or restore it.
|// x31 doesn't exist. Register number 31 either means xzr/wzr (zero) or sp,
|// depending on the instruction.
|// v0-v7 temp, v8-v15 callee-saved (only d8-d15 preserved), v16-v31 temp
|//
|// x0-x7/v0-v7 hold parameters and results.
|
|// Fixed register assignments for the interpreter.
|
|// The following must be C callee-save.
|.define BASE,		x19	// Base of current Lua stack frame.
|.define KBASE,		x20	// Constants of current Lua function.
|.define PC,		x21	// Next PC.
|.define GLREG,		x22	// Global state.
|.define LREG,		x23	// Register holding lua_State (also in SAVE_L).
|.define TISNUM,	x24	// Constant LJ_TISNUM << 47.
|.define TISNUMhi,	x25	// Constant LJ_TISNUM << 15.
|.define TISNIL,	x26	// Constant -1LL.
|.define fp,		x29	// Yes, we have to maintain a frame pointer.
|
|.define ST_INTERP,	w26	// Constant -1.
|
|// The following temporaries are not saved across C calls, except for RA/RC.
|.define RA,		x27
|.define RC,		x28
|.define RB,		x17
|.define RAw,		w27
|.define RCw,		w28
|.define RBw,		w17
|.define INS,		x16
|.define INSw,		w16
|.define ITYPE,		x15
|.define TMP0,		x8
|.define TMP1,		x9
|.define TMP2,		x10
|.define TMP3,		x11
|.define TMP0w,		w8
|.define TMP1w,		w9
|.define TMP2w,		w10
|.define TMP3w,		w11
|
|// Calling conventions. Also used as temporaries.
|.define CARG1,		x0
|.define CARG2,		x1
|.define CARG3,		x2
|.define CARG4,		x3
|.define CARG5,		x4
|.define CARG1w,	w0
|.define CARG2w,	w1
|.define CARG3w,	w2
|.define CARG4w,	w3
|.define CARG5w,	w4
|
|.define FARG1,		d0
|.define FARG2,		d1
|
|.define CRET1,		x0
|.define CRET1w,	w0
|
|// Stack layout while in interpreter. Must match with lj_frame.h.
|
|.define CFRAME_SPACE,	208
|//----- 16 byte aligned, <-- sp entering interpreter
|.define SAVE_FP_LR_,	192
|.define SAVE_GPR_,	112		// 112+10*8: 64 bit GPR saves
|.define SAVE_FPR_,	48		// 48+8*8: 64 bit FPR saves
|// Unused		[sp, #44]	// 32 bit values
|.define SAVE_NRES,	[sp, #40]
|.define SAVE_ERRF,	[sp, #36]
|.define SAVE_MULTRES,	[sp, #32]
|.define TMPD,		[sp, #24]	// 64 bit values
|.define SAVE_L,	[sp, #16]
|.define SAVE_PC,	[sp, #8]
|.define SAVE_CFRAME,	[sp, #0]
|//----- 16 byte aligned, <-- sp while in interpreter.
|
|.define TMPDofs,	#24
|
|.macro save_, gpr1, gpr2, fpr1, fpr2
|  stp d..fpr2, d..fpr1, [sp, # SAVE_FPR_+(14-fpr1)*8]
|  stp x..gpr2, x..gpr1, [sp, # SAVE_GPR_+(27-gpr1)*8]
|.endmacro
|.macro rest_, gpr1, gpr2, fpr1, fpr2
|  ldp d..fpr2, d..fpr1, [sp, # SAVE_FPR_+(14-fpr1)*8]
|  ldp x..gpr2, x..gpr1, [sp, # SAVE_GPR_+(27-gpr1)*8]
|.endmacro
|
|.macro saveregs
|  sub sp, sp, # CFRAME_SPACE
|  stp fp, lr, [sp, # SAVE_FP_LR_]
|  add fp, sp, # SAVE_FP_LR_
|  stp x20, x19, [sp, # SAVE_GPR_+(27-19)*8]
|  save_ 21, 22, 8, 9
|  save_ 23, 24, 10, 11
|  save_ 25, 26, 12, 13
|  save_ 27, 28, 14, 15
|.endmacro
|.macro restoreregs
|  ldp x20, x19, [sp, # SAVE_GPR_+(27-19)*8]
|  rest_ 21, 22, 8, 9
|  rest_ 23, 24, 10, 11
|  rest_ 25, 26, 12, 13
|  rest_ 27, 28, 14, 15
|  ldp fp, lr, [sp, # SAVE_FP_LR_]
|  add sp, sp, # CFRAME_SPACE
|.endmacro
|
|// Type definitions. Some of these are only used for documentation.
|.type L,		lua_State,	LREG
|.type GL,		global_State,	GLREG
|.type TVALUE,		TValue
|.type GCOBJ,		GCobj
|.type STR,		GCstr
|.type TAB,		GCtab
|.type LFUNC,		GCfuncL
|.type CFUNC,		GCfuncC
|.type PROTO,		GCproto
|.type UPVAL,		GCupval
|.type NODE,		Node
|.type NARGS8,		int
|.type TRACE,		GCtrace
|.type SBUF,		SBuf
|
|//-----------------------------------------------------------------------
|
|// Trap for not-yet-implemented parts.
|.macro NYI; brk; .endmacro
|
|//-----------------------------------------------------------------------
|
|// Access to frame relative to BASE.
|.define FRAME_FUNC,	#-16
|.define FRAME_PC,	#-8
|
|// Endian-specific defines.
|.if ENDIAN_LE
|.define LO,		0
|.define OFS_RD,	2
|.define OFS_RB,	3
|.define OFS_RA,	1
|.define OFS_OP,	0
|.else
|.define LO,		4
|.define OFS_RD,	0
|.define OFS_RB,	0
|.define OFS_RA,	2
|.define OFS_OP,	3
|.endif
|
|.macro decode_RA, dst, ins; ubfx dst, ins, #8, #8; .endmacro
|.macro decode_RB, dst, ins; ubfx dst, ins, #24, #8; .endmacro
|.macro decode_RC, dst, ins; ubfx dst, ins, #16, #8; .endmacro
|.macro decode_RD, dst, ins; ubfx dst, ins, #16, #16; .endmacro
|.macro decode_RC8RD, dst, src; ubfiz dst, src, #3, #8; .endmacro
|
|// Instruction decode+dispatch.
|.macro ins_NEXT
|  ldr INSw, [PC], #4
|  add TMP1, GL, INS, uxtb #3
|   decode_RA RA, INS
|  ldr TMP0, [TMP1, #GG_G2DISP]
|   decode_RD RC, INS
|  br TMP0
|.endmacro
|
|// Instruction footer.
|.if 1
|  // Replicated dispatch. Less unpredictable branches, but higher I-Cache use.
|  .define ins_next, ins_NEXT
|  .define ins_next_, ins_NEXT
|.else
|  // Common dispatch. Lower I-Cache use, only one (very) unpredictable branch.
|  // Affects only certain kinds of benchmarks (and only with -j off).
|  .macro ins_next
|    b ->ins_next
|  .endmacro
|  .macro ins_next_
|  ->ins_next:
|    ins_NEXT
|  .endmacro
|.endif
|
|// Call decode and dispatch.
|.macro ins_callt
|  // BASE = new base, CARG3 = LFUNC/CFUNC, RC = nargs*8, FRAME_PC(BASE) = PC
|  ldr PC, LFUNC:CARG3->pc
|  ldr INSw, [PC], #4
|  add TMP1, GL, INS, uxtb #3
|   decode_RA RA, INS
|  ldr TMP0, [TMP1, #GG_G2DISP]
|   add RA, BASE, RA, lsl #3
|  br TMP0
|.endmacro
|
|.macro ins_call
|  // BASE = new base, CARG3 = LFUNC/CFUNC, RC = nargs*8, PC = caller PC
|  str PC, [BASE, FRAME_PC]
|  ins_callt
|.endmacro
|
|//-----------------------------------------------------------------------
|
|// Macros to check the TValue type and extract the GCobj. Branch on failure.
|.macro checktp, reg, tp, target
|  asr ITYPE, reg, #47
|  cmn ITYPE, #-tp
|   and reg, reg, #LJ_GCVMASK
|  bne target
|.endmacro
|.macro checktp, dst, reg, tp, target
|  asr ITYPE, reg, #47
|  cmn ITYPE, #-tp
|   and dst, reg, #LJ_GCVMASK
|  bne target
|.endmacro
|.macro checkstr, reg, target; checktp reg, LJ_TSTR, target; .endmacro
|.macro checktab, reg, target; checktp reg, LJ_TTAB, target; .endmacro
|.macro checkfunc, reg, target; checktp reg, LJ_TFUNC, target; .endmacro
|.macro checkint, reg, target
|  cmp TISNUMhi, reg, lsr #32
|  bne target
|.endmacro
|.macro checknum, reg, target
|  cmp TISNUMhi, reg, lsr #32
|  bls target
|.endmacro
|.macro checknumber, reg, target
|  cmp TISNUMhi, reg, lsr #32
|  blo target
|.endmacro
|
|.macro mov_false, reg; movn reg, #0x8000, lsl #32; .endmacro
|.macro mov_true, reg; movn reg, #0x0001, lsl #48; .endmacro
|
#define GL_J(field)	(GG_G2J + (int)offsetof(jit_State, field))
|
#define PC2PROTO(field)  ((int)offsetof(GCproto, field)-(int)sizeof(GCproto))
|
|.macro hotcheck, delta
|  lsr CARG1, PC, #1
|  and CARG1, CARG1, #126
|  add CARG1, CARG1, #GG_G2DISP+GG_DISP2HOT
|  ldrh CARG2w, [GL, CARG1]
|  subs CARG2, CARG2, #delta
|  strh CARG2w, [GL, CARG1]
|.endmacro
|
|.macro hotloop
|  hotcheck HOTCOUNT_LOOP
|  blo ->vm_hotloop
|.endmacro
|
|.macro hotcall
|  hotcheck HOTCOUNT_CALL
|  blo ->vm_hotcall
|.endmacro
|
|// Set current VM state.
|.macro mv_vmstate, reg, st; movn reg, #LJ_VMST_..st; .endmacro
|.macro st_vmstate, reg; str reg, GL->vmstate; .endmacro
|
|// Move table write barrier back. Overwrites mark and tmp.
|.macro barrierback, tab, mark, tmp
|  ldr tmp, GL->gc.grayagain
|   and mark, mark, #~LJ_GC_BLACK	// black2gray(tab)
|  str tab, GL->gc.grayagain
|   strb mark, tab->marked
|  str tmp, tab->gclist
|.endmacro
|
|//-----------------------------------------------------------------------

#if !LJ_DUALNUM
#error "Only dual-number mode supported for ARM64 target"
#endif

/* Generate subroutines used by opcodes and other parts of the VM. */
/* The .code_sub section should be last to help static branch prediction. */
static void build_subroutines(BuildCtx *ctx)
{
  |.code_sub
  |
  |//-----------------------------------------------------------------------
  |//-- Return handling ----------------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |->vm_returnp:
  |  // See vm_return. Also: RB = previous base.
  |  tbz PC, #2, ->cont_dispatch	// (PC & FRAME_P) == 0?
  |
  |  // Return from pcall or xpcall fast func.
  |  ldr PC, [RB, FRAME_PC]		// Fetch PC of previous frame.
  |   mov_true TMP0
  |  mov BASE, RB
  |  // Prepending may overwrite the pcall frame, so do it at the end.
  |   str TMP0, [RA, #-8]!		// Prepend true to results.
  |
  |->vm_returnc:
  |  adds RC, RC, #8			// RC = (nresults+1)*8.
  |  mov CRET1, #LUA_YIELD
  |  beq ->vm_unwind_c_eh
  |  str RCw, SAVE_MULTRES
  |  ands CARG1, PC, #FRAME_TYPE
  |  beq ->BC_RET_Z			// Handle regular return to Lua.
  |
  |->vm_return:
  |  // BASE = base, RA = resultptr, RC/MULTRES = (nresults+1)*8, PC = return
  |  // CARG1 = PC & FRAME_TYPE
  |  and RB, PC, #~FRAME_TYPEP
  |   cmp CARG1, #FRAME_C
  |  sub RB, BASE, RB			// RB = previous base.
  |   bne ->vm_returnp
  |
  |  str RB, L->base
  |   ldrsw CARG2, SAVE_NRES		// CARG2 = nresults+1.
  |    mv_vmstate TMP0w, C
  |   sub BASE, BASE, #16
  |  subs TMP2, RC, #8
  |    st_vmstate TMP0w
  |  beq >2
  |1:
  |  subs TMP2, TMP2, #8
  |   ldr TMP0, [RA], #8
  |   str TMP0, [BASE], #8
  |  bne <1
  |2:
  |  cmp RC, CARG2, lsl #3		// More/less results wanted?
  |  bne >6
  |3:
  |  str BASE, L->top			// Store new top.
  |
  |->vm_leave_cp:
  |  ldr RC, SAVE_CFRAME		// Restore previous C frame.
  |   mov CRET1, #0			// Ok return status for vm_pcall.
  |  str RC, L->cframe
  |
  |->vm_leave_unw:
  |  restoreregs
  |  ret
  |
  |6:
  |  bgt >7				// Less results wanted?
  |  // More results wanted. Check stack size and fill up results with nil.
  |  ldr CARG3, L->maxstack
  |  cmp BASE, CARG3
  |  bhs >8
  |   str TISNIL, [BASE], #8
  |  add RC, RC, #8
  |  b <2
  |
  |7:  // Less results wanted.
  |  cbz CARG2, <3			// LUA_MULTRET+1 case?
  |  sub CARG1, RC, CARG2, lsl #3
  |  sub BASE, BASE, CARG1		// Shrink top.
  |  b <3
  |
  |8:  // Corner case: need to grow stack for filling up results.
  |  // This can happen if:
  |  // - A C function grows the stack (a lot).
  |  // - The GC shrinks the stack in between.
  |  // - A return back from a lua_call() with (high) nresults adjustment.
  |  str BASE, L->top			// Save current top held in BASE (yes).
  |  mov CARG1, L
  |  bl extern lj_state_growstack	// (lua_State *L, int n)
  |  ldr BASE, L->top			// Need the (realloced) L->top in BASE.
  |  ldrsw CARG2, SAVE_NRES
  |  b <2
  |
  |->vm_unwind_c:			// Unwind C stack, return from vm_pcall.
  |  // (void *cframe, int errcode)
  |  mov sp, CARG1
  |  mov CRET1, CARG2
  |->vm_unwind_c_eh:			// Landing pad for external unwinder.
  |  ldr L, SAVE_L
  |   mv_vmstate TMP0w, C
  |  ldr GL, L->glref
  |   st_vmstate TMP0w
  |  b ->vm_leave_unw
  |
  |->vm_unwind_ff:			// Unwind C stack, return from ff pcall.
  |  // (void *cframe)
  |  and sp, CARG1, #CFRAME_RAWMASK
  |->vm_unwind_ff_eh:			// Landing pad for external unwinder.
  |  ldr L, SAVE_L
  |    movz TISNUM, #(LJ_TISNUM>>1)&0xffff, lsl #48
  |    movz TISNUMhi, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |    movn TISNIL, #0
  |    mov RC, #16			// 2 results: false + error message.
  |  ldr BASE, L->base
  |   ldr GL, L->glref			// Setup pointer to global state.
  |    mov_false TMP0
  |  sub RA, BASE, #8			// Results start at BASE-8.
  |  ldr PC, [BASE, FRAME_PC]		// Fetch PC of previous frame.
  |    str TMP0, [BASE, #-8]		// Prepend false to error message.
  |   st_vmstate ST_INTERP
  |  b ->vm_returnc
  |
  |//-----------------------------------------------------------------------
  |//-- Grow stack for calls -----------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |->vm_growstack_c:			// Grow stack for C function.
  |  // CARG1 = L
  |  mov CARG2, #LUA_MINSTACK
  |  b >2
  |
  |->vm_growstack_l:			// Grow stack for Lua function.
  |  // BASE = new base, RA = BASE+framesize*8, RC = nargs*8, PC = first PC
  |  add RC, BASE, RC
  |   sub RA, RA, BASE
  |    mov CARG1, L
  |  stp BASE, RC, L->base
  |   add PC, PC, #4			// Must point after first instruction.
  |   lsr CARG2, RA, #3
  |2:
  |  // L->base = new base, L->top = top
  |  str PC, SAVE_PC
  |  bl extern lj_state_growstack	// (lua_State *L, int n)
  |  ldp BASE, RC, L->base
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]
  |   sub NARGS8:RC, RC, BASE
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  // BASE = new base, RB = LFUNC/CFUNC, RC = nargs*8, FRAME_PC(BASE) = PC
  |  ins_callt				// Just retry the call.
  |
  |//-----------------------------------------------------------------------
  |//-- Entry points into the assembler VM ---------------------------------
  |//-----------------------------------------------------------------------
  |
  |->vm_resume:				// Setup C frame and resume thread.
  |  // (lua_State *L, TValue *base, int nres1 = 0, ptrdiff_t ef = 0)
  |  saveregs
  |  mov L, CARG1
  |    ldr GL, L->glref			// Setup pointer to global state.
  |  mov BASE, CARG2
  |   str L, SAVE_L
  |  mov PC, #FRAME_CP
  |   str wzr, SAVE_NRES
  |    add TMP0, sp, #CFRAME_RESUME
  |  ldrb TMP1w, L->status
  |   str wzr, SAVE_ERRF
  |   str L, SAVE_PC			// Any value outside of bytecode is ok.
  |   str xzr, SAVE_CFRAME
  |    str TMP0, L->cframe
  |  cbz TMP1w, >3
  |
  |  // Resume after yield (like a return).
  |  str L, GL->cur_L
  |  mov RA, BASE
  |   ldp BASE, CARG1, L->base
  |    movz TISNUM, #(LJ_TISNUM>>1)&0xffff, lsl #48
  |    movz TISNUMhi, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |  ldr PC, [BASE, FRAME_PC]
  |     strb wzr, L->status
  |    movn TISNIL, #0
  |   sub RC, CARG1, BASE
  |  ands CARG1, PC, #FRAME_TYPE
  |   add RC, RC, #8
  |     st_vmstate ST_INTERP
  |   str RCw, SAVE_MULTRES
  |  beq ->BC_RET_Z
  |  b ->vm_return
  |
  |->vm_pcall:				// Setup protected C frame and enter VM.
  |  // (lua_State *L, TValue *base, int nres1, ptrdiff_t ef)
  |  saveregs
  |  mov PC, #FRAME_CP
  |  str CARG4w, SAVE_ERRF
  |  b >1
  |
  |->vm_call:				// Setup C frame and enter VM.
  |  // (lua_State *L, TValue *base, int nres1)
  |  saveregs
  |  mov PC, #FRAME_C
  |
  |1:  // Entry point for vm_pcall above (PC = ftype).
  |  ldr RC, L:CARG1->cframe
  |   str CARG3w, SAVE_NRES
  |    mov L, CARG1
  |   str CARG1, SAVE_L
  |    ldr GL, L->glref			// Setup pointer to global state.
  |     mov BASE, CARG2
  |   str CARG1, SAVE_PC		// Any value outside of bytecode is ok.
  |  add TMP0, sp, #0
  |   str RC, SAVE_CFRAME
  |  str TMP0, L->cframe		// Add our C frame to cframe chain.
  |
  |3:  // Entry point for vm_cpcall/vm_resume (BASE = base, PC = ftype).
  |  str L, GL->cur_L
  |  ldp RB, CARG1, L->base		// RB = old base (for vmeta_call).
  |    movz TISNUM, #(LJ_TISNUM>>1)&0xffff, lsl #48
  |    movz TISNUMhi, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |  add PC, PC, BASE
  |    movn TISNIL, #0
  |  sub PC, PC, RB			// PC = frame delta + frame type
  |   sub NARGS8:RC, CARG1, BASE
  |    st_vmstate ST_INTERP
  |
  |->vm_call_dispatch:
  |  // RB = old base, BASE = new base, RC = nargs*8, PC = caller PC
  |  ldr CARG3, [BASE, FRAME_FUNC]
  |  checkfunc CARG3, ->vmeta_call
  |
  |->vm_call_dispatch_f:
  |  ins_call
  |  // BASE = new base, CARG3 = func, RC = nargs*8, PC = caller PC
  |
  |->vm_cpcall:				// Setup protected C frame, call C.
  |  // (lua_State *L, lua_CFunction func, void *ud, lua_CPFunction cp)
  |  saveregs
  |  mov L, CARG1
  |   ldr RA, L:CARG1->stack
  |  str CARG1, SAVE_L
  |    ldr GL, L->glref			// Setup pointer to global state.
  |   ldr RB, L->top
  |  str CARG1, SAVE_PC			// Any value outside of bytecode is ok.
  |  ldr RC, L->cframe
  |   sub RA, RA, RB			// Compute -savestack(L, L->top).
  |   str RAw, SAVE_NRES		// Neg. delta means cframe w/o frame.
  |  str wzr, SAVE_ERRF			// No error function.
  |  add TMP0, sp, #0
  |   str RC, SAVE_CFRAME
  |  str TMP0, L->cframe		// Add our C frame to cframe chain.
  |    str L, GL->cur_L
  |  blr CARG4			// (lua_State *L, lua_CFunction func, void *ud)
  |  mov BASE, CRET1
  |   mov PC, #FRAME_CP
  |  cbnz BASE, <3			// Else continue with the call.
  |  b ->vm_leave_cp			// No base? Just remove C frame.
  |
  |//-----------------------------------------------------------------------
  |//-- Metamethod handling ------------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |//-- Continuation dispatch ----------------------------------------------
  |
  |->cont_dispatch:
  |  // BASE = meta base, RA = resultptr, RC = (nresults+1)*8
  |  ldr LFUNC:CARG3, [RB, FRAME_FUNC]
  |    ldr CARG1, [BASE, #-32]		// Get continuation.
  |   mov CARG4, BASE
  |   mov BASE, RB			// Restore caller BASE.
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |.if FFI
  |    cmp CARG1, #1
  |.endif
  |   ldr PC, [CARG4, #-24]		// Restore PC from [cont|PC].
  |    add TMP0, RA, RC
  |    str TISNIL, [TMP0, #-8]		// Ensure one valid arg.
  |.if FFI
  |    bls >1
  |.endif
  |  ldr CARG3, LFUNC:CARG3->pc
  |  ldr KBASE, [CARG3, #PC2PROTO(k)]
  |  // BASE = base, RA = resultptr, CARG4 = meta base
  |    br CARG1
  |
  |.if FFI
  |1:
  |  beq ->cont_ffi_callback		// cont = 1: return from FFI callback.
  |  // cont = 0: tailcall from C function.
  |   sub CARG4, CARG4, #32
  |   sub RC, CARG4, BASE
  |  b ->vm_call_tail
  |.endif
  |
  |->cont_cat:				// RA = resultptr, CARG4 = meta base
  |  ldr INSw, [PC, #-4]
  |   sub CARG2, CARG4, #32
  |   ldr TMP0, [RA]
  |     str BASE, L->base
  |  decode_RB RB, INS
  |   decode_RA RA, INS
  |  add TMP1, BASE, RB, lsl #3
  |  subs TMP1, CARG2, TMP1
  |  beq >1
  |   str TMP0, [CARG2]
  |  lsr CARG3, TMP1, #3
  |  b ->BC_CAT_Z
  |
  |1:
  |   str TMP0, [BASE, RA, lsl #3]
  |  b ->cont_nop
  |
  |//-- Table indexing metamethods -----------------------------------------
  |
  |->vmeta_tgets1:
  |  movn CARG4, #~LJ_TSTR
  |   add CARG2, BASE, RB, lsl #3
  |  add CARG4, STR:RC, CARG4, lsl #47
  |  b >2
  |
  |->vmeta_tgets:
  |  movk CARG2, #(LJ_TTAB>>1)&0xffff, lsl #48
  |  str CARG2, GL->tmptv
  |  add CARG2, GL, #offsetof(global_State, tmptv)
  |2:
  |   add CARG3, sp, TMPDofs
  |  str CARG4, TMPD
  |  b >1
  |
  |->vmeta_tgetb:			// RB = table, RC = index
  |  add RC, RC, TISNUM
  |   add CARG2, BASE, RB, lsl #3
  |   add CARG3, sp, TMPDofs
  |  str RC, TMPD
  |  b >1
  |
  |->vmeta_tgetv:			// RB = table, RC = key
  |  add CARG2, BASE, RB, lsl #3
  |   add CARG3, BASE, RC, lsl #3
  |1:
  |   str BASE, L->base
  |  mov CARG1, L
  |   str PC, SAVE_PC
  |  bl extern lj_meta_tget		// (lua_State *L, TValue *o, TValue *k)
  |  // Returns TValue * (finished) or NULL (metamethod).
  |  cbz CRET1, >3
  |  ldr TMP0, [CRET1]
  |  str TMP0, [BASE, RA, lsl #3]
  |  ins_next
  |
  |3:  // Call __index metamethod.
  |  // BASE = base, L->top = new base, stack = cont/func/t/k
  |   sub TMP1, BASE, #FRAME_CONT
  |  ldr BASE, L->top
  |    mov NARGS8:RC, #16		// 2 args for func(t, k).
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]  // Guaranteed to be a function here.
  |    str PC, [BASE, #-24]		// [cont|PC]
  |   sub PC, BASE, TMP1
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  b ->vm_call_dispatch_f
  |
  |->vmeta_tgetr:
  |  sxtw CARG2, TMP1w
  |  bl extern lj_tab_getinth		// (GCtab *t, int32_t key)
  |  // Returns cTValue * or NULL.
  |  mov TMP0, TISNIL
  |  cbz CRET1, ->BC_TGETR_Z
  |  ldr TMP0, [CRET1]
  |  b ->BC_TGETR_Z
  |
  |//-----------------------------------------------------------------------
  |
  |->vmeta_tsets1:
  |  movn CARG4, #~LJ_TSTR
  |   add CARG2, BASE, RB, lsl #3
  |  add CARG4, STR:RC, CARG4, lsl #47
  |  b >2
  |
  |->vmeta_tsets:
  |  movk CARG2, #(LJ_TTAB>>1)&0xffff, lsl #48
  |  str CARG2, GL->tmptv
  |  add CARG2, GL, #offsetof(global_State, tmptv)
  |2:
  |   add CARG3, sp, TMPDofs
  |  str CARG4, TMPD
  |  b >1
  |
  |->vmeta_tsetb:			// RB = table, RC = index
  |  add RC, RC, TISNUM
  |   add CARG2, BASE, RB, lsl #3
  |   add CARG3, sp, TMPDofs
  |  str RC, TMPD
  |  b >1
  |
  |->vmeta_tsetv:
  |  add CARG2, BASE, RB, lsl #3
  |   add CARG3, BASE, RC, lsl #3
  |1:
  |   str BASE, L->base
  |  mov CARG1, L
  |   str PC, SAVE_PC
  |  bl extern lj_meta_tset		// (lua_State *L, TValue *o, TValue *k)
  |  // Returns TValue * (finished) or NULL (metamethod).
  |   ldr TMP0, [BASE, RA, lsl #3]
  |  cbz CRET1, >3
  |  // NOBARRIER: lj_meta_tset ensures the table is not black.
  |   str TMP0, [CRET1]
  |  ins_next
  |
  |3:  // Call __newindex metamethod.
  |  // BASE = base, L->top = new base, stack = cont/func/t/k/(v)
  |   sub TMP1, BASE, #FRAME_CONT
  |  ldr BASE, L->top
  |    mov NARGS8:RC, #24		// 3 args for func(t, k, v).
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]  // Guaranteed to be a function here.
  |   str TMP0, [BASE, #16]		// Copy value to third argument.
  |    str PC, [BASE, #-24]		// [cont|PC]
  |   sub PC, BASE, TMP1
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  b ->vm_call_dispatch_f
  |
  |->vmeta_tsetr:
  |  sxtw CARG3, TMP1w
  |  str BASE, L->base
  |  mov CARG1, L
  |  str PC, SAVE_PC
  |  bl extern lj_tab_setinth  // (lua_State *L, GCtab *t, int32_t key)
  |  // Returns TValue *.
  |  b ->BC_TSETR_Z
  |
  |//-- Comparison metamethods ---------------------------------------------
  |
  |->vmeta_comp:
  |  add CARG2, BASE, RA, lsl #3
  |   sub PC, PC, #4
  |  add CARG3, BASE, RC, lsl #3
  |   str BASE, L->base
  |  mov CARG1, L
  |   str PC, SAVE_PC
  |  uxtb CARG4w, INSw
  |  bl extern lj_meta_comp  // (lua_State *L, TValue *o1, *o2, int op)
  |  // Returns 0/1 or TValue * (metamethod).
  |3:
  |  cmp CRET1, #1
  |  bhi ->vmeta_binop
  |4:
  |   ldrh RBw, [PC, # OFS_RD]
  |    add PC, PC, #4
  |   add RB, PC, RB, lsl #2
  |   sub RB, RB, #0x20000
  |  csel PC, PC, RB, lo
  |->cont_nop:
  |  ins_next
  |
  |->cont_ra:				// RA = resultptr
  |  ldr INSw, [PC, #-4]
  |   ldr TMP0, [RA]
  |  decode_RA TMP1, INS
  |   str TMP0, [BASE, TMP1, lsl #3]
  |  b ->cont_nop
  |
  |->cont_condt:			// RA = resultptr
  |  ldr TMP0, [RA]
  |   mov_true TMP1
  |  cmp TMP1, TMP0			// Branch if result is true.
  |  b <4
  |
  |->cont_condf:			// RA = resultptr
  |  ldr TMP0, [RA]
  |   mov_false TMP1
  |  cmp TMP0, TMP1			// Branch if result is false.
  |  b <4
  |
  |->vmeta_equal:
  |  // CARG2, CARG3, CARG4 are already set by BC_ISEQV/BC_ISNEV.
  |  and TAB:CARG3, CARG3, #LJ_GCVMASK
  |  sub PC, PC, #4
  |   str BASE, L->base
  |   mov CARG1, L
  |  str PC, SAVE_PC
  |  bl extern lj_meta_equal  // (lua_State *L, GCobj *o1, *o2, int ne)
  |  // Returns 0/1 or TValue * (metamethod).
  |  b <3
  |
  |->vmeta_equal_cd:
  |.if FFI
  |  sub PC, PC, #4
  |   str BASE, L->base
  |   mov CARG1, L
  |   mov CARG2, INS
  |  str PC, SAVE_PC
  |  bl extern lj_meta_equal_cd		// (lua_State *L, BCIns op)
  |  // Returns 0/1 or TValue * (metamethod).
  |  b <3
  |.endif
  |
  |->vmeta_istype:
  |  sub PC, PC, #4
  |   str BASE, L->base
  |   mov CARG1, L
  |   mov CARG2, RA
  |   mov CARG3, RC
  |  str PC, SAVE_PC
  |  bl extern lj_meta_istype  // (lua_State *L, BCReg ra, BCReg tp)
  |  b ->cont_nop
  |
  |//-- Arithmetic metamethods ---------------------------------------------
  |
  |->vmeta_arith_vn:
  |  add CARG3, BASE, RB, lsl #3
  |   add CARG4, KBASE, RC, lsl #3
  |  b >1
  |
  |->vmeta_arith_nv:
  |  add CARG4, BASE, RB, lsl #3
  |   add CARG3, KBASE, RC, lsl #3
  |  b >1
  |
  |->vmeta_unm:
  |  add CARG3, BASE, RC, lsl #3
  |  mov CARG4, CARG3
  |  b >1
  |
  |->vmeta_arith_vv:
  |  add CARG3, BASE, RB, lsl #3
  |   add CARG4, BASE, RC, lsl #3
  |1:
  |  uxtb CARG5w, INSw
  |   add CARG2, BASE, RA, lsl #3
  |    str BASE, L->base
  |   mov CARG1, L
  |    str PC, SAVE_PC
  |  bl extern lj_meta_arith  // (lua_State *L, TValue *ra,*rb,*rc, BCReg op)
  |  // Returns NULL (finished) or TValue * (metamethod).
  |  cbz CRET1, ->cont_nop
  |
  |  // Call metamethod for binary op.
  |->vmeta_binop:
  |  // BASE = old base, CRET1 = new base, stack = cont/func/o1/o2
  |  sub TMP1, CRET1, BASE
  |   str PC, [CRET1, #-24]		// [cont|PC]
  |  add PC, TMP1, #FRAME_CONT
  |  mov BASE, CRET1
  |   mov NARGS8:RC, #16		// 2 args for func(o1, o2).
  |  b ->vm_call_dispatch
  |
  |->vmeta_len:
  |  add CARG2, BASE, RC, lsl #3
#if LJ_52
  |  mov TAB:RC, TAB:CARG1		// Save table (ignored for other types).
#endif
  |   str BASE, L->base
  |  mov CARG1, L
  |   str PC, SAVE_PC
  |  bl extern lj_meta_len		// (lua_State *L, TValue *o)
  |  // Returns NULL (retry) or TValue * (metamethod base).
#if LJ_52
  |  cbnz CRET1, ->vmeta_binop		// Binop call for compatibility.
  |  mov TAB:CARG1, TAB:RC
  |  b ->BC_LEN_Z
#else
  |  b ->vmeta_binop			// Binop call for compatibility.
#endif
  |
  |//-- Call metamethod ----------------------------------------------------
  |
  |->vmeta_call:			// Resolve and call __call metamethod.
  |  // RB = old base, BASE = new base, RC = nargs*8
  |  mov CARG1, L
  |   str RB, L->base			// This is the callers base!
  |  sub CARG2, BASE, #16
  |   str PC, SAVE_PC
  |  add CARG3, BASE, NARGS8:RC
  |  bl extern lj_meta_call	// (lua_State *L, TValue *func, TValue *top)
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]  // Guaranteed to be a function here.
  |   add NARGS8:RC, NARGS8:RC, #8	// Got one more argument now.
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  ins_call
  |
  |->vmeta_callt:			// Resolve __call for BC_CALLT.
  |  // BASE = old base, RA = new base, RC = nargs*8
  |  mov CARG1, L
  |   str BASE, L->base
  |  sub CARG2, RA, #16
  |   str PC, SAVE_PC
  |  add CARG3, RA, NARGS8:RC
  |  bl extern lj_meta_call	// (lua_State *L, TValue *func, TValue *top)
  |  ldr TMP1, [RA, FRAME_FUNC]		// Guaranteed to be a function here.
  |   ldr PC, [BASE, FRAME_PC]
  |   add NARGS8:RC, NARGS8:RC, #8	// Got one more argument now.
  |  and LFUNC:CARG3, TMP1, #LJ_GCVMASK
  |  b ->BC_CALLT2_Z
  |
  |//-- Argument coercion for 'for' statement ------------------------------
  |
  |->vmeta_for:
  |  mov CARG1, L
  |   str BASE, L->base
  |  mov CARG2, RA
  |   str PC, SAVE_PC
  |  bl extern lj_meta_for	// (lua_State *L, TValue *base)
  |  ldr INSw, [PC, #-4]
  |.if JIT
  |   uxtb TMP0w, INSw
  |.endif
  |  decode_RA RA, INS
  |  decode_RD RC, INS
  |.if JIT
  |   cmp TMP0, #BC_JFORI
  |   beq =>BC_JFORI
  |.endif
  |  b =>BC_FORI
  |
  |//-----------------------------------------------------------------------
  |//-- Fast functions -----------------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |.macro .ffunc, name
  |->ff_ .. name:
  |.endmacro
  |
  |.macro .ffunc_1, name
  |->ff_ .. name:
  |  ldr CARG1, [BASE]
  |   cmp NARGS8:RC, #8
  |   blo ->fff_fallback
  |.endmacro
  |
  |.macro .ffunc_2, name
  |->ff_ .. name:
  |  ldp CARG1, CARG2, [BASE]
  |   cmp NARGS8:RC, #16
  |   blo ->fff_fallback
  |.endmacro
  |
  |.macro .ffunc_n, name
  |  .ffunc name
  |  ldr CARG1, [BASE]
  |   cmp NARGS8:RC, #8
  |  ldr FARG1, [BASE]
  |   blo ->fff_fallback
  |  checknum CARG1, ->fff_fallback
  |.endmacro
  |
  |.macro .ffunc_nn, name
  |  .ffunc name
  |  ldp CARG1, CARG2, [BASE]
  |   cmp NARGS8:RC, #16
  |  ldp FARG1, FARG2, [BASE]
  |   blo ->fff_fallback
  |  checknum CARG1, ->fff_fallback
  |  checknum CARG2, ->fff_fallback
  |.endmacro
  |
  |// Inlined GC threshold check. Caveat: uses CARG1 and CARG2.
  |.macro ffgccheck
  |  ldp CARG1, CARG2, GL->gc.total	// Assumes threshold follows total.
  |  cmp CARG1, CARG2
  |  blt >1
  |  bl ->fff_gcstep
  |1:
  |.endmacro
  |
  |//-- Base library: checks -----------------------------------------------
  |
  |.ffunc_1 assert
  |   ldr PC, [BASE, FRAME_PC]
  |  mov_false TMP1
  |  cmp CARG1, TMP1
  |  bhs ->fff_fallback
  |  str CARG1, [BASE, #-16]
  |  sub RB, BASE, #8
  |  subs RA, NARGS8:RC, #8
  |   add RC, NARGS8:RC, #8		// Compute (nresults+1)*8.
  |  cbz RA, ->fff_res			// Done if exactly 1 argument.
  |1:
  |   ldr CARG1, [RB, #16]
  |  sub RA, RA, #8
  |   str CARG1, [RB], #8
  |  cbnz RA, <1
  |  b ->fff_res
  |
  |.ffunc_1 type
  |  mov TMP0, #~LJ_TISNUM
  |  asr ITYPE, CARG1, #47
  |  cmn ITYPE, #~LJ_TISNUM
  |  csinv TMP1, TMP0, ITYPE, lo
  |  add TMP1, TMP1, #offsetof(GCfuncC, upvalue)/8
  |  ldr CARG1, [CFUNC:CARG3, TMP1, lsl #3]
  |  b ->fff_restv
  |
  |//-- Base library: getters and setters ---------------------------------
  |
  |.ffunc_1 getmetatable
  |  asr ITYPE, CARG1, #47
  |  cmn ITYPE, #-LJ_TTAB
  |  ccmn ITYPE, #-LJ_TUDATA, #4, ne
  |   and TAB:CARG1, CARG1, #LJ_GCVMASK
  |  bne >6
  |1:  // Field metatable must be at same offset for GCtab and GCudata!
  |  ldr TAB:RB, TAB:CARG1->metatable
  |2:
  |   mov CARG1, TISNIL
  |   ldr STR:RC, GL->gcroot[GCROOT_MMNAME+MM_metatable]
  |  cbz TAB:RB, ->fff_restv
  |  ldr TMP1w, TAB:RB->hmask
  |   ldr TMP2w, STR:RC->sid
  |    ldr NODE:CARG3, TAB:RB->node
  |  and TMP1w, TMP1w, TMP2w		// idx = str->sid & tab->hmask
  |  add TMP1, TMP1, TMP1, lsl #1
  |  movn CARG4, #~LJ_TSTR
  |    add NODE:CARG3, NODE:CARG3, TMP1, lsl #3  // node = tab->node + idx*3*8
  |  add CARG4, STR:RC, CARG4, lsl #47	// Tagged key to look for.
  |3:  // Rearranged logic, because we expect _not_ to find the key.
  |  ldp CARG1, TMP0, NODE:CARG3->val
  |   ldr NODE:CARG3, NODE:CARG3->next
  |  cmp TMP0, CARG4
  |  beq >5
  |  cbnz NODE:CARG3, <3
  |4:
  |  mov CARG1, RB			// Use metatable as default result.
  |  movk CARG1, #(LJ_TTAB>>1)&0xffff, lsl #48
  |  b ->fff_restv
  |5:
  |  cmp TMP0, TISNIL
  |  bne ->fff_restv
  |  b <4
  |
  |6:
  |  movn TMP0, #~LJ_TISNUM
  |  cmp ITYPE, TMP0
  |  csel ITYPE, ITYPE, TMP0, hs
  |  sub TMP1, GL, ITYPE, lsl #3
  |  ldr TAB:RB, [TMP1, #offsetof(global_State, gcroot[GCROOT_BASEMT])-8]
  |  b <2
  |
  |.ffunc_2 setmetatable
  |  // Fast path: no mt for table yet and not clearing the mt.
  |  checktp TMP1, CARG1, LJ_TTAB, ->fff_fallback
  |   ldr TAB:TMP0, TAB:TMP1->metatable
  |  asr ITYPE, CARG2, #47
  |   ldrb TMP2w, TAB:TMP1->marked
  |  cmn ITYPE, #-LJ_TTAB
  |    and TAB:CARG2, CARG2, #LJ_GCVMASK
  |  ccmp TAB:TMP0, #0, #0, eq
  |  bne ->fff_fallback
  |    str TAB:CARG2, TAB:TMP1->metatable
  |   tbz TMP2w, #2, ->fff_restv	// isblack(table)
  |  barrierback TAB:TMP1, TMP2w, TMP0
  |  b ->fff_restv
  |
  |.ffunc rawget
  |  ldr CARG2, [BASE]
  |   cmp NARGS8:RC, #16
  |   blo ->fff_fallback
  |  checktab CARG2, ->fff_fallback
  |   mov CARG1, L
  |   add CARG3, BASE, #8
  |  bl extern lj_tab_get  // (lua_State *L, GCtab *t, cTValue *key)
  |  // Returns cTValue *.
  |  ldr CARG1, [CRET1]
  |  b ->fff_restv
  |
  |//-- Base library: conversions ------------------------------------------
  |
  |.ffunc tonumber
  |  // Only handles the number case inline (without a base argument).
  |  ldr CARG1, [BASE]
  |   cmp NARGS8:RC, #8
  |   bne ->fff_fallback
  |  checknumber CARG1, ->fff_fallback
  |  b ->fff_restv
  |
  |.ffunc_1 tostring
  |  // Only handles the string or number case inline.
  |  asr ITYPE, CARG1, #47
  |  cmn ITYPE, #-LJ_TSTR
  |  // A __tostring method in the string base metatable is ignored.
  |  beq ->fff_restv
  |  // Handle numbers inline, unless a number base metatable is present.
  |  ldr TMP1, GL->gcroot[GCROOT_BASEMT_NUM]
  |   str BASE, L->base
  |  cmn ITYPE, #-LJ_TISNUM
  |  ccmp TMP1, #0, #0, ls
  |   str PC, SAVE_PC			// Redundant (but a defined value).
  |  bne ->fff_fallback
  |  ffgccheck
  |  mov CARG1, L
  |  mov CARG2, BASE
  |  bl extern lj_strfmt_number		// (lua_State *L, cTValue *o)
  |  // Returns GCstr *.
  |   movn TMP1, #~LJ_TSTR
  |  ldr BASE, L->base
  |   add CARG1, CARG1, TMP1, lsl #47
  |  b ->fff_restv
  |
  |//-- Base library: iterators -------------------------------------------
  |
  |.ffunc_1 next
  |  checktp CARG1, LJ_TTAB, ->fff_fallback
  |  str TISNIL, [BASE, NARGS8:RC]	// Set missing 2nd arg to nil.
  |  ldr PC, [BASE, FRAME_PC]
  |  add CARG2, BASE, #8
  |  sub CARG3, BASE, #16
  |  bl extern lj_tab_next		// (GCtab *t, cTValue *key, TValue *o)
  |  // Returns 1=found, 0=end, -1=error.
  |   mov RC, #(2+1)*8
  |  tbnz CRET1w, #31, ->fff_fallback	// Invalid key.
  |  cbnz CRET1, ->fff_res		// Found key/value.
  |  // End of traversal: return nil.
  |  str TISNIL, [BASE, #-16]
  |  b ->fff_res1
  |
  |.ffunc_1 pairs
  |  checktp TMP1, CARG1, LJ_TTAB, ->fff_fallback
#if LJ_52
  |  ldr TAB:CARG2, TAB:TMP1->metatable
#endif
  |   ldr CFUNC:CARG4, CFUNC:CARG3->upvalue[0]
  |    ldr PC, [BASE, FRAME_PC]
#if LJ_52
  |  cbnz TAB:CARG2, ->fff_fallback
#endif
  |  mov RC, #(3+1)*8
  |  stp CARG1, TISNIL, [BASE, #-8]
  |   str CFUNC:CARG4, [BASE, #-16]
  |  b ->fff_res
  |
  |.ffunc_2 ipairs_aux
  |  checktab CARG1, ->fff_fallback
  |   checkint CARG2, ->fff_fallback
  |  ldr TMP1w, TAB:CARG1->asize
  |   ldr CARG3, TAB:CARG1->array
  |    ldr TMP0w, TAB:CARG1->hmask
  |  add CARG2w, CARG2w, #1
  |  cmp CARG2w, TMP1w
  |    ldr PC, [BASE, FRAME_PC]
  |     add TMP2, CARG2, TISNUM
  |   mov RC, #(0+1)*8
  |     str TMP2, [BASE, #-16]
  |  bhs >2				// Not in array part?
  |  ldr TMP0, [CARG3, CARG2, lsl #3]
  |1:
  |   mov TMP1, #(2+1)*8
  |   cmp TMP0, TISNIL
  |  str TMP0, [BASE, #-8]
  |   csel RC, RC, TMP1, eq
  |  b ->fff_res
  |2:  // Check for empty hash part first. Otherwise call C function.
  |  cbz TMP0w, ->fff_res
  |  bl extern lj_tab_getinth		// (GCtab *t, int32_t key)
  |  // Returns cTValue * or NULL.
  |  cbz CRET1, ->fff_res
  |  ldr TMP0, [CRET1]
  |  b <1
  |
  |.ffunc_1 ipairs
  |  checktp TMP1, CARG1, LJ_TTAB, ->fff_fallback
#if LJ_52
  |  ldr TAB:CARG2, TAB:TMP1->metatable
#endif
  |   ldr CFUNC:CARG4, CFUNC:CARG3->upvalue[0]
  |    ldr PC, [BASE, FRAME_PC]
#if LJ_52
  |  cbnz TAB:CARG2, ->fff_fallback
#endif
  |  mov RC, #(3+1)*8
  |  stp CARG1, TISNUM, [BASE, #-8]
  |   str CFUNC:CARG4, [BASE, #-16]
  |  b ->fff_res
  |
  |//-- Base library: catch errors ----------------------------------------
  |
  |.ffunc pcall
  |   cmp NARGS8:RC, #8
  |  ldrb TMP0w, GL->hookmask
  |   blo ->fff_fallback
  |   sub NARGS8:RC, NARGS8:RC, #8
  |    mov RB, BASE
  |    add BASE, BASE, #16
  |  ubfx TMP0w, TMP0w, #HOOK_ACTIVE_SHIFT, #1
  |  add PC, TMP0, #16+FRAME_PCALL
  |   beq ->vm_call_dispatch
  |1:
  |   add TMP2, BASE, NARGS8:RC
  |2:
  |   ldr TMP0, [TMP2, #-16]
  |   str TMP0, [TMP2, #-8]!
  |  cmp TMP2, BASE
  |  bne <2
  |  b ->vm_call_dispatch
  |
  |.ffunc xpcall
  |     ldp CARG1, CARG2, [BASE]
  |  ldrb TMP0w, GL->hookmask
  |   subs NARGS8:TMP1, NARGS8:RC, #16
  |   blo ->fff_fallback
  |    mov RB, BASE
  |     asr ITYPE, CARG2, #47
  |  ubfx TMP0w, TMP0w, #HOOK_ACTIVE_SHIFT, #1
  |     cmn ITYPE, #-LJ_TFUNC
  |  add PC, TMP0, #24+FRAME_PCALL
  |     bne ->fff_fallback		// Traceback must be a function.
  |   mov NARGS8:RC, NARGS8:TMP1
  |    add BASE, BASE, #24
  |     stp CARG2, CARG1, [RB]		// Swap function and traceback.
  |   cbz NARGS8:RC, ->vm_call_dispatch
  |  b <1
  |
  |//-- Coroutine library --------------------------------------------------
  |
  |.macro coroutine_resume_wrap, resume
  |.if resume
  |.ffunc_1 coroutine_resume
  |  checktp CARG1, LJ_TTHREAD, ->fff_fallback
  |.else
  |.ffunc coroutine_wrap_aux
  |  ldr L:CARG1, CFUNC:CARG3->upvalue[0].gcr
  |  and L:CARG1, CARG1, #LJ_GCVMASK
  |.endif
  |   ldr PC, [BASE, FRAME_PC]
  |     str BASE, L->base
  |  ldp RB, CARG2, L:CARG1->base
  |   ldrb TMP1w, L:CARG1->status
  |  add TMP0, CARG2, TMP1
  |   str PC, SAVE_PC
  |  cmp TMP0, RB
  |  beq ->fff_fallback
  |   cmp TMP1, #LUA_YIELD
  |    add TMP0, CARG2, #8
  |   csel CARG2, CARG2, TMP0, hs
  |   ldr CARG4, L:CARG1->maxstack
  |   add CARG3, CARG2, NARGS8:RC
  |    ldr RB, L:CARG1->cframe
  |   ccmp CARG3, CARG4, #2, ls
  |    ccmp RB, #0, #2, ls
  |    bhi ->fff_fallback
  |.if resume
  |  sub CARG3, CARG3, #8		// Keep resumed thread in stack for GC.
  |  add BASE, BASE, #8
  |  sub NARGS8:RC, NARGS8:RC, #8
  |.endif
  |  str CARG3, L:CARG1->top
  |  str BASE, L->top
  |  cbz NARGS8:RC, >3
  |2:  // Move args to coroutine.
  |   ldr TMP0, [BASE, RB]
  |  cmp RB, NARGS8:RC
  |   str TMP0, [CARG2, RB]
  |   add RB, RB, #8
  |  bne <2
  |3:
  |  mov CARG3, #0
  |   mov L:RA, L:CARG1
  |  mov CARG4, #0
  |  bl ->vm_resume			// (lua_State *L, TValue *base, 0, 0)
  |  // Returns thread status.
  |4:
  |  ldp CARG3, CARG4, L:RA->base
  |   cmp CRET1, #LUA_YIELD
  |  ldr BASE, L->base
  |    str L, GL->cur_L
  |    st_vmstate ST_INTERP
  |   bhi >8
  |  sub RC, CARG4, CARG3
  |   ldr CARG1, L->maxstack
  |   add CARG2, BASE, RC
  |  cbz RC, >6				// No results?
  |  cmp CARG2, CARG1
  |   mov RB, #0
  |  bhi >9				// Need to grow stack?
  |
  |  sub CARG4, RC, #8
  |   str CARG3, L:RA->top		// Clear coroutine stack.
  |5:  // Move results from coroutine.
  |   ldr TMP0, [CARG3, RB]
  |  cmp RB, CARG4
  |   str TMP0, [BASE, RB]
  |   add RB, RB, #8
  |  bne <5
  |6:
  |.if resume
  |  mov_true TMP1
  |   add RC, RC, #16
  |7:
  |  str TMP1, [BASE, #-8]		// Prepend true/false to results.
  |   sub RA, BASE, #8
  |.else
  |   mov RA, BASE
  |   add RC, RC, #8
  |.endif
  |  ands CARG1, PC, #FRAME_TYPE
  |   str PC, SAVE_PC
  |   str RCw, SAVE_MULTRES
  |  beq ->BC_RET_Z
  |  b ->vm_return
  |
  |8:  // Coroutine returned with error (at co->top-1).
  |.if resume
  |  ldr TMP0, [CARG4, #-8]!
  |   mov_false TMP1
  |    mov RC, #(2+1)*8
  |  str CARG4, L:RA->top		// Remove error from coroutine stack.
  |  str TMP0, [BASE]			// Copy error message.
  |  b <7
  |.else
  |  mov CARG1, L
  |  mov CARG2, L:RA
  |  bl extern lj_ffh_coroutine_wrap_err  // (lua_State *L, lua_State *co)
  |  // Never returns.
  |.endif
  |
  |9:  // Handle stack expansion on return from yield.
  |  mov CARG1, L
  |  lsr CARG2, RC, #3
  |  bl extern lj_state_growstack	// (lua_State *L, int n)
  |  mov CRET1, #0
  |  b <4
  |.endmacro
  |
  |  coroutine_resume_wrap 1		// coroutine.resume
  |  coroutine_resume_wrap 0		// coroutine.wrap
  |
  |.ffunc coroutine_yield
  |  ldr TMP0, L->cframe
  |   add TMP1, BASE, NARGS8:RC
  |    mov CRET1, #LUA_YIELD
  |   stp BASE, TMP1, L->base
  |  tbz TMP0, #0, ->fff_fallback
  |   str xzr, L->cframe
  |    strb CRET1w, L->status
  |  b ->vm_leave_unw
  |
  |//-- Math library -------------------------------------------------------
  |
  |.macro math_round, func, round
  |  .ffunc math_ .. func
  |  ldr CARG1, [BASE]
  |   cmp NARGS8:RC, #8
  |  ldr d0, [BASE]
  |   blo ->fff_fallback
  |  cmp TISNUMhi, CARG1, lsr #32
  |  beq ->fff_restv
  |  blo ->fff_fallback
  |  round d0, d0
  |  b ->fff_resn
  |.endmacro
  |
  |  math_round floor, frintm
  |  math_round ceil, frintp
  |
  |.ffunc_1 math_abs
  |  checknumber CARG1, ->fff_fallback
  |  and CARG1, CARG1, #U64x(7fffffff,ffffffff)
  |  bne ->fff_restv
  |  eor CARG2w, CARG1w, CARG1w, asr #31
  |   movz CARG3, #0x41e0, lsl #48	// 2^31.
  |  subs CARG1w, CARG2w, CARG1w, asr #31
  |   add CARG1, CARG1, TISNUM
  |  csel CARG1, CARG1, CARG3, pl
  |  // Fallthrough.
  |
  |->fff_restv:
  |  // CARG1 = TValue result.
  |  ldr PC, [BASE, FRAME_PC]
  |  str CARG1, [BASE, #-16]
  |->fff_res1:
  |  // PC = return.
  |  mov RC, #(1+1)*8
  |->fff_res:
  |  // RC = (nresults+1)*8, PC = return.
  |  ands CARG1, PC, #FRAME_TYPE
  |   str RCw, SAVE_MULTRES
  |   sub RA, BASE, #16
  |  bne ->vm_return
  |  ldr INSw, [PC, #-4]
  |  decode_RB RB, INS
  |5:
  |  cmp RC, RB, lsl #3			// More results expected?
  |  blo >6
  |  decode_RA TMP1, INS
  |  // Adjust BASE. KBASE is assumed to be set for the calling frame.
  |  sub BASE, RA, TMP1, lsl #3
  |  ins_next
  |
  |6:  // Fill up results with nil.
  |  add TMP1, RA, RC
  |   add RC, RC, #8
  |  str TISNIL, [TMP1, #-8]
  |  b <5
  |
  |.macro math_extern, func
  |  .ffunc_n math_ .. func
  |  bl extern func
  |  b ->fff_resn
  |.endmacro
  |
  |.macro math_extern2, func
  |  .ffunc_nn math_ .. func
  |  bl extern func
  |  b ->fff_resn
  |.endmacro
  |
  |.ffunc_n math_sqrt
  |  fsqrt d0, d0
  |->fff_resn:
  |  ldr PC, [BASE, FRAME_PC]
  |  str d0, [BASE, #-16]
  |  b ->fff_res1
  |
  |.ffunc math_log
  |  ldr CARG1, [BASE]
  |   cmp NARGS8:RC, #8
  |  ldr FARG1, [BASE]
  |   bne ->fff_fallback			// Need exactly 1 argument.
  |  checknum CARG1, ->fff_fallback
  |  bl extern log
  |  b ->fff_resn
  |
  |  math_extern log10
  |  math_extern exp
  |  math_extern sin
  |  math_extern cos
  |  math_extern tan
  |  math_extern asin
  |  math_extern acos
  |  math_extern atan
  |  math_extern sinh
  |  math_extern cosh
  |  math_extern tanh
  |  math_extern2 pow
  |  math_extern2 atan2
  |  math_extern2 fmod
  |
  |.ffunc_2 math_ldexp
  |  ldr FARG1, [BASE]
  |  checknum CARG1, ->fff_fallback
  |  checkint CARG2, ->fff_fallback
  |  sxtw CARG1, CARG2w
  |  bl extern ldexp			// (double x, int exp)
  |  b ->fff_resn
  |
  |.ffunc_n math_frexp
  |  add CARG1, sp, TMPDofs
  |  bl extern frexp
  |   ldr CARG2w, TMPD
  |    ldr PC, [BASE, FRAME_PC]
  |  str d0, [BASE, #-16]
  |    mov RC, #(2+1)*8
  |   add CARG2, CARG2, TISNUM
  |   str CARG2, [BASE, #-8]
  |  b ->fff_res
  |
  |.ffunc_n math_modf
  |  sub CARG1, BASE, #16
  |   ldr PC, [BASE, FRAME_PC]
  |  bl extern modf
  |   mov RC, #(2+1)*8
  |  str d0, [BASE, #-8]
  |  b ->fff_res
  |
  |.macro math_minmax, name, cond, fcond
  |  .ffunc_1 name
  |   add RB, BASE, RC
  |   add RA, BASE, #8
  |  checkint CARG1, >4
  |1:  // Handle integers.
  |  ldr CARG2, [RA]
  |   cmp RA, RB
  |   bhs ->fff_restv
  |  checkint CARG2, >3
  |  cmp CARG1w, CARG2w
  |   add RA, RA, #8
  |  csel CARG1, CARG2, CARG1, cond
  |  b <1
  |3:  // Convert intermediate result to number and continue below.
  |  scvtf d0, CARG1w
  |  blo ->fff_fallback
  |  ldr d1, [RA]
  |  b >6
  |
  |4:
  |  ldr d0, [BASE]
  |  blo ->fff_fallback
  |5:  // Handle numbers.
  |  ldr CARG2, [RA]
  |  ldr d1, [RA]
  |   cmp RA, RB
  |   bhs ->fff_resn
  |  checknum CARG2, >7
  |6:
  |  fcmp d0, d1
  |   add RA, RA, #8
  |  fcsel d0, d1, d0, fcond
  |  b <5
  |7:  // Convert integer to number and continue above.
  |  scvtf d1, CARG2w
  |  blo ->fff_fallback
  |  b <6
  |.endmacro
  |
  |  math_minmax math_min, gt, pl
  |  math_minmax math_max, lt, le
  |
  |//-- String library -----------------------------------------------------
  |
  |.ffunc string_byte			// Only handle the 1-arg case here.
  |  ldp PC, CARG1, [BASE, FRAME_PC]
  |   cmp NARGS8:RC, #8
  |  asr ITYPE, CARG1, #47
  |  ccmn ITYPE, #-LJ_TSTR, #0, eq
  |   and STR:CARG1, CARG1, #LJ_GCVMASK
  |  bne ->fff_fallback
  |  ldrb TMP0w, STR:CARG1[1]		// Access is always ok (NUL at end).
  |   ldr CARG3w, STR:CARG1->len
  |  add TMP0, TMP0, TISNUM
  |  str TMP0, [BASE, #-16]
  |  mov RC, #(0+1)*8
  |   cbz CARG3, ->fff_res
  |  b ->fff_res1
  |
  |.ffunc string_char			// Only handle the 1-arg case here.
  |  ffgccheck
  |  ldp PC, CARG1, [BASE, FRAME_PC]
  |  cmp CARG1w, #255
  |   ccmp NARGS8:RC, #8, #0, ls		// Need exactly 1 argument.
  |  bne ->fff_fallback
  |  checkint CARG1, ->fff_fallback
  |  mov CARG3, #1
  |  // Point to the char inside the integer in the stack slot.
  |.if ENDIAN_LE
  |  mov CARG2, BASE
  |.else
  |  add CARG2, BASE, #7
  |.endif
  |->fff_newstr:
  |  // CARG2 = str, CARG3 = len.
  |   str BASE, L->base
  |  mov CARG1, L
  |   str PC, SAVE_PC
  |  bl extern lj_str_new		// (lua_State *L, char *str, size_t l)
  |->fff_resstr:
  |  // Returns GCstr *.
  |  ldr BASE, L->base
  |   movn TMP1, #~LJ_TSTR
  |  add CARG1, CARG1, TMP1, lsl #47
  |  b ->fff_restv
  |
  |.ffunc string_sub
  |  ffgccheck
  |  ldr CARG1, [BASE]
  |    ldr CARG3, [BASE, #16]
  |   cmp NARGS8:RC, #16
  |    movn RB, #0
  |   beq >1
  |   blo ->fff_fallback
  |    checkint CARG3, ->fff_fallback
  |    sxtw RB, CARG3w
  |1:
  |  ldr CARG2, [BASE, #8]
  |  checkstr CARG1, ->fff_fallback
  |   ldr TMP1w, STR:CARG1->len
  |  checkint CARG2, ->fff_fallback
  |  sxtw CARG2, CARG2w
  |  // CARG1 = str, TMP1 = str->len, CARG2 = start, RB = end
  |   add TMP2, RB, TMP1
  |   cmp RB, #0
  |  add TMP0, CARG2, TMP1
  |   csinc RB, RB, TMP2, ge		// if (end < 0) end += len+1
  |  cmp CARG2, #0
  |  csinc CARG2, CARG2, TMP0, ge	// if (start < 0) start += len+1
  |   cmp RB, #0
  |   csel RB, RB, xzr, ge		// if (end < 0) end = 0
  |  cmp CARG2, #1
  |  csinc CARG2, CARG2, xzr, ge	// if (start < 1) start = 1
  |   cmp RB, TMP1
  |   csel RB, RB, TMP1, le		// if (end > len) end = len
  |  add CARG1, STR:CARG1, #sizeof(GCstr)-1
  |   subs CARG3, RB, CARG2		// len = end - start
  |  add CARG2, CARG1, CARG2
  |   add CARG3, CARG3, #1		// len += 1
  |   bge ->fff_newstr
  |  add STR:CARG1, GL, #offsetof(global_State, strempty)
  |   movn TMP1, #~LJ_TSTR
  |  add CARG1, CARG1, TMP1, lsl #47
  |  b ->fff_restv
  |
  |.macro ffstring_op, name
  |  .ffunc string_ .. name
  |  ffgccheck
  |  ldr CARG2, [BASE]
  |   cmp NARGS8:RC, #8
  |  asr ITYPE, CARG2, #47
  |  ccmn ITYPE, #-LJ_TSTR, #0, hs
  |   and STR:CARG2, CARG2, #LJ_GCVMASK
  |  bne ->fff_fallback
  |  ldr TMP0, GL->tmpbuf.b
  |   add SBUF:CARG1, GL, #offsetof(global_State, tmpbuf)
  |   str BASE, L->base
  |   str PC, SAVE_PC
  |   str L, GL->tmpbuf.L
  |  str TMP0, GL->tmpbuf.w
  |  bl extern lj_buf_putstr_ .. name
  |  bl extern lj_buf_tostr
  |  b ->fff_resstr
  |.endmacro
  |
  |ffstring_op reverse
  |ffstring_op lower
  |ffstring_op upper
  |
  |//-- Bit library --------------------------------------------------------
  |
  |// FP number to bit conversion for soft-float. Clobbers CARG1-CARG3
  |->vm_tobit_fb:
  |  bls ->fff_fallback
  |  add CARG2, CARG1, CARG1
  |  mov CARG3, #1076
  |  sub CARG3, CARG3, CARG2, lsr #53
  |  cmp CARG3, #53
  |  bhi >1
  |  and CARG2, CARG2, #U64x(001fffff,ffffffff)
  |  orr CARG2, CARG2, #U64x(00200000,00000000)
  |   cmp CARG1, #0
  |  lsr CARG2, CARG2, CARG3
  |   cneg CARG1w, CARG2w, mi
  |  br lr
  |1:
  |  mov CARG1w, #0
  |  br lr
  |
  |.macro .ffunc_bit, name
  |  .ffunc_1 bit_..name
  |  adr lr, >1
  |  checkint CARG1, ->vm_tobit_fb
  |1:
  |.endmacro
  |
  |.macro .ffunc_bit_op, name, ins
  |  .ffunc_bit name
  |  mov RA, #8
  |  mov TMP0w, CARG1w
  |  adr lr, >2
  |1:
  |  ldr CARG1, [BASE, RA]
  |   cmp RA, NARGS8:RC
  |    add RA, RA, #8
  |   bge >9
  |  checkint CARG1, ->vm_tobit_fb
  |2:
  |  ins TMP0w, TMP0w, CARG1w
  |  b <1
  |.endmacro
  |
  |.ffunc_bit_op band, and
  |.ffunc_bit_op bor, orr
  |.ffunc_bit_op bxor, eor
  |
  |.ffunc_bit tobit
  |  mov TMP0w, CARG1w
  |9:  // Label reused by .ffunc_bit_op users.
  |  add CARG1, TMP0, TISNUM
  |  b ->fff_restv
  |
  |.ffunc_bit bswap
  |  rev TMP0w, CARG1w
  |  add CARG1, TMP0, TISNUM
  |  b ->fff_restv
  |
  |.ffunc_bit bnot
  |  mvn TMP0w, CARG1w
  |  add CARG1, TMP0, TISNUM
  |  b ->fff_restv
  |
  |.macro .ffunc_bit_sh, name, ins, shmod
  |  .ffunc bit_..name
  |  ldp TMP0, CARG1, [BASE]
  |   cmp NARGS8:RC, #16
  |   blo ->fff_fallback
  |  adr lr, >1
  |  checkint CARG1, ->vm_tobit_fb
  |1:
  |.if shmod == 0
  |  mov TMP1, CARG1
  |.else
  |  neg TMP1, CARG1
  |.endif
  |  mov CARG1, TMP0
  |  adr lr, >2
  |  checkint CARG1, ->vm_tobit_fb
  |2:
  |  ins TMP0w, CARG1w, TMP1w
  |  add CARG1, TMP0, TISNUM
  |  b ->fff_restv
  |.endmacro
  |
  |.ffunc_bit_sh lshift, lsl, 0
  |.ffunc_bit_sh rshift, lsr, 0
  |.ffunc_bit_sh arshift, asr, 0
  |.ffunc_bit_sh rol, ror, 1
  |.ffunc_bit_sh ror, ror, 0
  |
  |//-----------------------------------------------------------------------
  |
  |->fff_fallback:			// Call fast function fallback handler.
  |  // BASE = new base, RC = nargs*8
  |   ldp CFUNC:CARG3, PC, [BASE, FRAME_FUNC]	// Fallback may overwrite PC.
  |  ldr TMP2, L->maxstack
  |  add TMP1, BASE, NARGS8:RC
  |  stp BASE, TMP1, L->base
  |   and CFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  add TMP1, TMP1, #8*LUA_MINSTACK
  |   ldr CARG3, CFUNC:CARG3->f
  |    str PC, SAVE_PC			// Redundant (but a defined value).
  |  cmp TMP1, TMP2
  |   mov CARG1, L
  |  bhi >5				// Need to grow stack.
  |   blr CARG3				// (lua_State *L)
  |  // Either throws an error, or recovers and returns -1, 0 or nresults+1.
  |   ldr BASE, L->base
  |  cmp CRET1w, #0
  |   lsl RC, CRET1, #3
  |   sub RA, BASE, #16
  |  bgt ->fff_res			// Returned nresults+1?
  |1:  // Returned 0 or -1: retry fast path.
  |   ldr CARG1, L->top
  |    ldr CFUNC:CARG3, [BASE, FRAME_FUNC]
  |   sub NARGS8:RC, CARG1, BASE
  |  bne ->vm_call_tail			// Returned -1?
  |    and CFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  ins_callt				// Returned 0: retry fast path.
  |
  |// Reconstruct previous base for vmeta_call during tailcall.
  |->vm_call_tail:
  |  ands TMP0, PC, #FRAME_TYPE
  |   and TMP1, PC, #~FRAME_TYPEP
  |  bne >3
  |  ldrb RAw, [PC, #-4+OFS_RA]
  |  lsl RA, RA, #3
  |  add TMP1, RA, #16
  |3:
  |  sub RB, BASE, TMP1
  |  b ->vm_call_dispatch		// Resolve again for tailcall.
  |
  |5:  // Grow stack for fallback handler.
  |  mov CARG2, #LUA_MINSTACK
  |  bl extern lj_state_growstack	// (lua_State *L, int n)
  |  ldr BASE, L->base
  |  cmp CARG1, CARG1			// Set zero-flag to force retry.
  |  b <1
  |
  |->fff_gcstep:			// Call GC step function.
  |  // BASE = new base, RC = nargs*8
  |   add CARG2, BASE, NARGS8:RC	// Calculate L->top.
  |  mov RA, lr
  |   stp BASE, CARG2, L->base
  |   str PC, SAVE_PC			// Redundant (but a defined value).
  |  mov CARG1, L
  |  bl extern lj_gc_step		// (lua_State *L)
  |  ldp BASE, CARG2, L->base
  |   ldr CFUNC:CARG3, [BASE, FRAME_FUNC]
  |  mov lr, RA				// Help return address predictor.
  |  sub NARGS8:RC, CARG2, BASE		// Calculate nargs*8.
  |   and CFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  ret
  |
  |//-----------------------------------------------------------------------
  |//-- Special dispatch targets -------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |->vm_record:				// Dispatch target for recording phase.
  |.if JIT
  |  ldrb CARG1w, GL->hookmask
  |  tst CARG1, #HOOK_VMEVENT		// No recording while in vmevent.
  |  bne >5
  |  // Decrement the hookcount for consistency, but always do the call.
  |   ldr CARG2w, GL->hookcount
  |  tst CARG1, #HOOK_ACTIVE
  |  bne >1
  |   sub CARG2w, CARG2w, #1
  |  tst CARG1, #LUA_MASKLINE|LUA_MASKCOUNT
  |  beq >1
  |   str CARG2w, GL->hookcount
  |  b >1
  |.endif
  |
  |->vm_rethook:			// Dispatch target for return hooks.
  |  ldrb TMP2w, GL->hookmask
  |  tbz TMP2w, #HOOK_ACTIVE_SHIFT, >1	// Hook already active?
  |5:  // Re-dispatch to static ins.
  |  ldr TMP0, [TMP1, #GG_G2DISP+GG_DISP2STATIC]
  |  br TMP0
  |
  |->vm_inshook:			// Dispatch target for instr/line hooks.
  |  ldrb TMP2w, GL->hookmask
  |   ldr TMP3w, GL->hookcount
  |  tbnz TMP2w, #HOOK_ACTIVE_SHIFT, <5	// Hook already active?
  |  tst TMP2w, #LUA_MASKLINE|LUA_MASKCOUNT
  |  beq <5
  |   sub TMP3w, TMP3w, #1
  |   str TMP3w, GL->hookcount
  |   cbz TMP3w, >1
  |  tbz TMP2w, #LUA_HOOKLINE, <5
  |1:
  |  mov CARG1, L
  |   str BASE, L->base
  |  mov CARG2, PC
  |  // SAVE_PC must hold the _previous_ PC. The callee updates it with PC.
  |  bl extern lj_dispatch_ins		// (lua_State *L, const BCIns *pc)
  |3:
  |  ldr BASE, L->base
  |4:  // Re-dispatch to static ins.
  |  ldr INSw, [PC, #-4]
  |  add TMP1, GL, INS, uxtb #3
  |   decode_RA RA, INS
  |  ldr TMP0, [TMP1, #GG_G2DISP+GG_DISP2STATIC]
  |   decode_RD RC, INS
  |  br TMP0
  |
  |->cont_hook:				// Continue from hook yield.
  |  ldr CARG1, [CARG4, #-40]
  |   add PC, PC, #4
  |  str CARG1w, SAVE_MULTRES		// Restore MULTRES for *M ins.
  |  b <4
  |
  |->vm_hotloop:			// Hot loop counter underflow.
  |.if JIT
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]  // Same as curr_topL(L).
  |   add CARG1, GL, #GG_G2DISP+GG_DISP2J
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |   str PC, SAVE_PC
  |  ldr CARG3, LFUNC:CARG3->pc
  |   mov CARG2, PC
  |   str L, [GL, #GL_J(L)]
  |  ldrb CARG3w, [CARG3, #PC2PROTO(framesize)]
  |   str BASE, L->base
  |  add CARG3, BASE, CARG3, lsl #3
  |  str CARG3, L->top
  |  bl extern lj_trace_hot		// (jit_State *J, const BCIns *pc)
  |  b <3
  |.endif
  |
  |->vm_callhook:			// Dispatch target for call hooks.
  |  mov CARG2, PC
  |.if JIT
  |  b >1
  |.endif
  |
  |->vm_hotcall:			// Hot call counter underflow.
  |.if JIT
  |  orr CARG2, PC, #1
  |1:
  |.endif
  |  add TMP1, BASE, NARGS8:RC
  |   str PC, SAVE_PC
  |   mov CARG1, L
  |   sub RA, RA, BASE
  |  stp BASE, TMP1, L->base
  |  bl extern lj_dispatch_call		// (lua_State *L, const BCIns *pc)
  |  // Returns ASMFunction.
  |  ldp BASE, TMP1, L->base
  |   str xzr, SAVE_PC			// Invalidate for subsequent line hook.
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]
  |  add RA, BASE, RA
  |  sub NARGS8:RC, TMP1, BASE
  |   ldr INSw, [PC, #-4]
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  br CRET1
  |
  |->cont_stitch:			// Trace stitching.
  |.if JIT
  |  // RA = resultptr, CARG4 = meta base
  |   ldr RBw, SAVE_MULTRES
  |  ldr INSw, [PC, #-4]
  |    ldr TRACE:CARG3, [CARG4, #-40]	// Save previous trace.
  |   subs RB, RB, #8
  |  decode_RA RC, INS			// Call base.
  |    and CARG3, CARG3, #LJ_GCVMASK
  |   beq >2
  |1:  // Move results down.
  |  ldr CARG1, [RA]
  |    add RA, RA, #8
  |   subs RB, RB, #8
  |  str CARG1, [BASE, RC, lsl #3]
  |    add RC, RC, #1
  |   bne <1
  |2:
  |   decode_RA RA, INS
  |   decode_RB RB, INS
  |   add RA, RA, RB
  |3:
  |   cmp RA, RC
  |   bhi >9				// More results wanted?
  |
  |  ldrh RAw, TRACE:CARG3->traceno
  |  ldrh RCw, TRACE:CARG3->link
  |  cmp RCw, RAw
  |  beq ->cont_nop			// Blacklisted.
  |  cmp RCw, #0
  |  bne =>BC_JLOOP			// Jump to stitched trace.
  |
  |  // Stitch a new trace to the previous trace.
  |  mov CARG1, #GL_J(exitno)
  |  str RAw, [GL, CARG1]
  |  mov CARG1, #GL_J(L)
  |  str L, [GL, CARG1]
  |  str BASE, L->base
  |  add CARG1, GL, #GG_G2J
  |  mov CARG2, PC
  |  bl extern lj_dispatch_stitch	// (jit_State *J, const BCIns *pc)
  |  ldr BASE, L->base
  |  b ->cont_nop
  |
  |9:  // Fill up results with nil.
  |  str TISNIL, [BASE, RC, lsl #3]
  |  add RC, RC, #1
  |  b <3
  |.endif
  |
  |->vm_profhook:			// Dispatch target for profiler hook.
#if LJ_HASPROFILE
  |  mov CARG1, L
  |   str BASE, L->base
  |  mov CARG2, PC
  |  bl extern lj_dispatch_profile	// (lua_State *L, const BCIns *pc)
  |  // HOOK_PROFILE is off again, so re-dispatch to dynamic instruction.
  |  ldr BASE, L->base
  |  sub PC, PC, #4
  |  b ->cont_nop
#endif
  |
  |//-----------------------------------------------------------------------
  |//-- Trace exit handler -------------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |.macro savex_, a, b
  |  stp d..a, d..b, [sp, #a*8]
  |  stp x..a, x..b, [sp, #32*8+a*8]
  |.endmacro
  |
  |->vm_exit_handler:
  |.if JIT
  |  sub     sp, sp, #(64*8)
  |  savex_, 0, 1
  |  savex_, 2, 3
  |  savex_, 4, 5
  |  savex_, 6, 7
  |  savex_, 8, 9
  |  savex_, 10, 11
  |  savex_, 12, 13
  |  savex_, 14, 15
  |  savex_, 16, 17
  |  savex_, 18, 19
  |  savex_, 20, 21
  |  savex_, 22, 23
  |  savex_, 24, 25
  |  savex_, 26, 27
  |  savex_, 28, 29
  |  stp d30, d31, [sp, #30*8]
  |  ldr CARG1, [sp, #64*8]	// Load original value of lr.
  |   add CARG3, sp, #64*8	// Recompute original value of sp.
  |    mv_vmstate CARG4w, EXIT
  |   stp xzr, CARG3, [sp, #62*8]	// Store 0/sp in RID_LR/RID_SP.
  |  sub CARG1, CARG1, lr
  |   ldr L, GL->cur_L
  |  lsr CARG1, CARG1, #2
  |   ldr BASE, GL->jit_base
  |  sub CARG1, CARG1, #2
  |   ldr CARG2w, [lr]		// Load trace number.
  |    st_vmstate CARG4w
  |.if ENDIAN_BE
  |   rev32 CARG2, CARG2
  |.endif
  |   str BASE, L->base
  |  ubfx CARG2w, CARG2w, #5, #16
  |  str CARG1w, [GL, #GL_J(exitno)]
  |   str CARG2w, [GL, #GL_J(parent)]
  |   str L, [GL, #GL_J(L)]
  |  str xzr, GL->jit_base
  |  add CARG1, GL, #GG_G2J
  |  mov CARG2, sp
  |  bl extern lj_trace_exit		// (jit_State *J, ExitState *ex)
  |  // Returns MULTRES (unscaled) or negated error code.
  |  ldr CARG2, L->cframe
  |   ldr BASE, L->base
  |  and sp, CARG2, #CFRAME_RAWMASK
  |   ldr PC, SAVE_PC			// Get SAVE_PC.
  |  str L, SAVE_L			// Set SAVE_L (on-trace resume/yield).
  |  b >1
  |.endif
  |
  |->vm_exit_interp:
  |  // CARG1 = MULTRES or negated error code, BASE, PC and GL set.
  |.if JIT
  |  ldr L, SAVE_L
  |1:
  |  cmp CARG1w, #0
  |  blt >9				// Check for error from exit.
  |   lsl RC, CARG1, #3
  |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
  |    movz TISNUM, #(LJ_TISNUM>>1)&0xffff, lsl #48
  |    movz TISNUMhi, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |    movn TISNIL, #0
  |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
  |   str RCw, SAVE_MULTRES
  |   str BASE, L->base
  |  ldr CARG2, LFUNC:CARG2->pc
  |   str xzr, GL->jit_base
  |    mv_vmstate CARG4w, INTERP
  |  ldr KBASE, [CARG2, #PC2PROTO(k)]
  |  // Modified copy of ins_next which handles function header dispatch, too.
  |  ldrb RBw, [PC, # OFS_OP]
  |   ldr INSw, [PC], #4
  |    st_vmstate CARG4w
  |  cmp RBw, #BC_FUNCC+2		// Fast function?
  |   add TMP1, GL, INS, uxtb #3
  |  bhs >4
  |2:
  |  cmp RBw, #BC_FUNCF			// Function header?
  |  add TMP0, GL, RB, uxtb #3
  |  ldr RB, [TMP0, #GG_G2DISP]
  |   decode_RA RA, INS
  |   lsr TMP0, INS, #16
  |   csel RC, TMP0, RC, lo
  |   blo >5
  |   ldr CARG3, [BASE, FRAME_FUNC]
  |   sub RC, RC, #8
  |   add RA, BASE, RA, lsl #3	// Yes: RA = BASE+framesize*8, RC = nargs*8
  |   and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |5:
  |  br RB
  |
  |4:  // Check frame below fast function.
  |  ldr CARG1, [BASE, FRAME_PC]
  |  ands CARG2, CARG1, #FRAME_TYPE
  |  bne <2			// Trace stitching continuation?
  |  // Otherwise set KBASE for Lua function below fast function.
  |  ldr CARG3w, [CARG1, #-4]
  |  decode_RA CARG1, CARG3
  |  sub CARG2, BASE, CARG1, lsl #3
  |  ldr LFUNC:CARG3, [CARG2, #-32]
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  ldr CARG3, LFUNC:CARG3->pc
  |  ldr KBASE, [CARG3, #PC2PROTO(k)]
  |  b <2
  |
  |9:  // Rethrow error from the right C frame.
  |  neg CARG2w, CARG1w
  |  mov CARG1, L
  |  bl extern lj_err_trace		// (lua_State *L, int errcode)
  |.endif
  |
  |//-----------------------------------------------------------------------
  |//-- Math helper functions ----------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |  // int lj_vm_modi(int dividend, int divisor);
  |->vm_modi:
  |    eor CARG4w, CARG1w, CARG2w
  |    cmp CARG4w, #0
  |  eor CARG3w, CARG1w, CARG1w, asr #31
  |   eor CARG4w, CARG2w, CARG2w, asr #31
  |  sub CARG3w, CARG3w, CARG1w, asr #31
  |   sub CARG4w, CARG4w, CARG2w, asr #31
  |  udiv CARG1w, CARG3w, CARG4w
  |  msub CARG1w, CARG1w, CARG4w, CARG3w
  |    ccmp CARG1w, #0, #4, mi
  |    sub CARG3w, CARG1w, CARG4w
  |    csel CARG1w, CARG1w, CARG3w, eq
  |  eor CARG3w, CARG1w, CARG2w
  |  cmp CARG3w, #0
  |  cneg CARG1w, CARG1w, mi
  |  ret
  |
  |//-----------------------------------------------------------------------
  |//-- Miscellaneous functions --------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |.define NEXT_TAB,		TAB:CARG1
  |.define NEXT_RES,		CARG1
  |.define NEXT_IDX,		CARG2w
  |.define NEXT_LIM,		CARG3w
  |.define NEXT_TMP0,		TMP0
  |.define NEXT_TMP0w,		TMP0w
  |.define NEXT_TMP1,		TMP1
  |.define NEXT_TMP1w,		TMP1w
  |.define NEXT_RES_PTR,	sp
  |.define NEXT_RES_VAL,	[sp]
  |.define NEXT_RES_KEY,	[sp, #8]
  |
  |// TValue *lj_vm_next(GCtab *t, uint32_t idx)
  |// Next idx returned in CRET2w.
  |->vm_next:
  |.if JIT
  |  ldr NEXT_LIM, NEXT_TAB->asize
  |   ldr NEXT_TMP1, NEXT_TAB->array
  |1:  // Traverse array part.
  |  subs NEXT_TMP0w, NEXT_IDX, NEXT_LIM
  |  bhs >5				// Index points after array part?
  |  ldr NEXT_TMP0, [NEXT_TMP1, NEXT_IDX, uxtw #3]
  |  cmn NEXT_TMP0, #-LJ_TNIL
  |   cinc NEXT_IDX, NEXT_IDX, eq
  |  beq <1				// Skip holes in array part.
  |  str NEXT_TMP0, NEXT_RES_VAL
  |   movz NEXT_TMP0w, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |   stp NEXT_IDX, NEXT_TMP0w, NEXT_RES_KEY
  |  add NEXT_IDX, NEXT_IDX, #1
  |  mov NEXT_RES, NEXT_RES_PTR
  |4:
  |  ret
  |
  |5:  // Traverse hash part.
  |  ldr NEXT_TMP1w, NEXT_TAB->hmask
  |   ldr NODE:NEXT_RES, NEXT_TAB->node
  |   add NEXT_TMP0w, NEXT_TMP0w, NEXT_TMP0w, lsl #1
  |  add NEXT_LIM, NEXT_LIM, NEXT_TMP1w
  |   add NODE:NEXT_RES, NODE:NEXT_RES, NEXT_TMP0w, uxtw #3
  |6:
  |  cmp NEXT_IDX, NEXT_LIM
  |  bhi >9
  |  ldr NEXT_TMP0, NODE:NEXT_RES->val
  |  cmn NEXT_TMP0, #-LJ_TNIL
  |   add NEXT_IDX, NEXT_IDX, #1
  |  bne <4
  |  // Skip holes in hash part.
  |  add NODE:NEXT_RES, NODE:NEXT_RES, #sizeof(Node)
  |  b <6
  |
  |9:  // End of iteration. Set the key to nil (not the value).
  |  movn NEXT_TMP0, #0
  |  str NEXT_TMP0, NEXT_RES_KEY
  |  mov NEXT_RES, NEXT_RES_PTR
  |  ret
  |.endif
  |
  |//-----------------------------------------------------------------------
  |//-- FFI helper functions -----------------------------------------------
  |//-----------------------------------------------------------------------
  |
  |// Handler for callback functions.
  |// Saveregs already performed. Callback slot number in [sp], g in r12.
  |->vm_ffi_callback:
  |.if FFI
  |.type CTSTATE, CTState, PC
  |  saveregs
  |  ldr CTSTATE, GL:x10->ctype_state
  |  mov GL, x10
  |    add x10, sp, # CFRAME_SPACE
  |  str w9, CTSTATE->cb.slot
  |  stp x0, x1, CTSTATE->cb.gpr[0]
  |   stp d0, d1, CTSTATE->cb.fpr[0]
  |  stp x2, x3, CTSTATE->cb.gpr[2]
  |   stp d2, d3, CTSTATE->cb.fpr[2]
  |  stp x4, x5, CTSTATE->cb.gpr[4]
  |   stp d4, d5, CTSTATE->cb.fpr[4]
  |  stp x6, x7, CTSTATE->cb.gpr[6]
  |   stp d6, d7, CTSTATE->cb.fpr[6]
  |    str x10, CTSTATE->cb.stack
  |  mov CARG1, CTSTATE
  |   str CTSTATE, SAVE_PC		// Any value outside of bytecode is ok.
  |  mov CARG2, sp
  |  bl extern lj_ccallback_enter	// (CTState *cts, void *cf)
  |  // Returns lua_State *.
  |  ldp BASE, RC, L:CRET1->base
  |   movz TISNUM, #(LJ_TISNUM>>1)&0xffff, lsl #48
  |   movz TISNUMhi, #(LJ_TISNUM>>1)&0xffff, lsl #16
  |   movn TISNIL, #0
  |   mov L, CRET1
  |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]
  |  sub RC, RC, BASE
  |   st_vmstate ST_INTERP
  |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
  |  ins_callt
  |.endif
  |
  |->cont_ffi_callback:			// Return from FFI callback.
  |.if FFI
  |  ldr CTSTATE, GL->ctype_state
  |   stp BASE, CARG4, L->base
  |  str L, CTSTATE->L
  |  mov CARG1, CTSTATE
  |  mov CARG2, RA
  |  bl extern lj_ccallback_leave       // (CTState *cts, TValue *o)
  |  ldp x0, x1, CTSTATE->cb.gpr[0]
  |   ldp d0, d1, CTSTATE->cb.fpr[0]
  |  b ->vm_leave_unw
  |.endif
  |
  |->vm_ffi_call:			// Call C function via FFI.
  |  // Caveat: needs special frame unwinding, see below.
  |.if FFI
  |  .type CCSTATE, CCallState, x19
  |  stp x20, CCSTATE, [sp, #-32]!
  |  stp fp, lr, [sp, #16]
  |  add fp, sp, #16
  |  mov CCSTATE, x0
  |  ldr TMP0w, CCSTATE:x0->spadj
  |   ldrb TMP1w, CCSTATE->nsp
  |    add TMP2, CCSTATE, #offsetof(CCallState, stack)
  |   subs TMP1, TMP1, #1
  |    ldr TMP3, CCSTATE->func
  |  sub sp, sp, TMP0
  |   bmi >2
  |1:  // Copy stack slots
  |  ldr TMP0, [TMP2, TMP1, lsl #3]
  |  str TMP0, [sp, TMP1, lsl #3]
  |  subs TMP1, TMP1, #1
  |  bpl <1
  |2:
  |  ldp x0, x1, CCSTATE->gpr[0]
  |   ldp d0, d1, CCSTATE->fpr[0]
  |  ldp x2, x3, CCSTATE->gpr[2]
  |   ldp d2, d3, CCSTATE->fpr[2]
  |  ldp x4, x5, CCSTATE->gpr[4]
  |   ldp d4, d5, CCSTATE->fpr[4]
  |  ldp x6, x7, CCSTATE->gpr[6]
  |   ldp d6, d7, CCSTATE->fpr[6]
  |  ldr x8, CCSTATE->retp
  |  blr TMP3
  |  sub sp, fp, #16
  |  stp x0, x1, CCSTATE->gpr[0]
  |   stp d0, d1, CCSTATE->fpr[0]
  |   stp d2, d3, CCSTATE->fpr[2]
  |  ldp fp, lr, [sp, #16]
  |  ldp x20, CCSTATE, [sp], #32
  |  ret
  |.endif
  |// Note: vm_ffi_call must be the last function in this object file!
  |
  |//-----------------------------------------------------------------------
}

/* Generate the code for a single instruction. */
static void build_ins(BuildCtx *ctx, BCOp op, int defop)
{
  int vk = 0;
  |=>defop:

  switch (op) {

  /* -- Comparison ops ---------------------------------------------------- */

  /* Remember: all ops branch for a true comparison, fall through otherwise. */

  case BC_ISLT: case BC_ISGE: case BC_ISLE: case BC_ISGT:
    |  // RA = src1, RC = src2, JMP with RC = target
    |  ldr CARG1, [BASE, RA, lsl #3]
    |    ldrh RBw, [PC, # OFS_RD]
    |   ldr CARG2, [BASE, RC, lsl #3]
    |    add PC, PC, #4
    |    add RB, PC, RB, lsl #2
    |    sub RB, RB, #0x20000
    |  checkint CARG1, >3
    |   checkint CARG2, >4
    |  cmp CARG1w, CARG2w
    if (op == BC_ISLT) {
      |  csel PC, RB, PC, lt
    } else if (op == BC_ISGE) {
      |  csel PC, RB, PC, ge
    } else if (op == BC_ISLE) {
      |  csel PC, RB, PC, le
    } else {
      |  csel PC, RB, PC, gt
    }
    |1:
    |  ins_next
    |
    |3:  // RA not int.
    |    ldr FARG1, [BASE, RA, lsl #3]
    |  blo ->vmeta_comp
    |    ldr FARG2, [BASE, RC, lsl #3]
    |   cmp TISNUMhi, CARG2, lsr #32
    |   bhi >5
    |   bne ->vmeta_comp
    |  // RA number, RC int.
    |  scvtf FARG2, CARG2w
    |  b >5
    |
    |4:  // RA int, RC not int
    |    ldr FARG2, [BASE, RC, lsl #3]
    |   blo ->vmeta_comp
    |  // RA int, RC number.
    |  scvtf FARG1, CARG1w
    |
    |5:  // RA number, RC number
    |  fcmp FARG1, FARG2
    |  // To preserve NaN semantics GE/GT branch on unordered, but LT/LE don't.
    if (op == BC_ISLT) {
      |  csel PC, RB, PC, lo
    } else if (op == BC_ISGE) {
      |  csel PC, RB, PC, hs
    } else if (op == BC_ISLE) {
      |  csel PC, RB, PC, ls
    } else {
      |  csel PC, RB, PC, hi
    }
    |  b <1
    break;

  case BC_ISEQV: case BC_ISNEV:
    vk = op == BC_ISEQV;
    |  // RA = src1, RC = src2, JMP with RC = target
    |  ldr CARG1, [BASE, RA, lsl #3]
    |   add RC, BASE, RC, lsl #3
    |    ldrh RBw, [PC, # OFS_RD]
    |   ldr CARG3, [RC]
    |    add PC, PC, #4
    |    add RB, PC, RB, lsl #2
    |    sub RB, RB, #0x20000
    |  asr ITYPE, CARG3, #47
    |  cmn ITYPE, #-LJ_TISNUM
    if (vk) {
      |  bls ->BC_ISEQN_Z
    } else {
      |  bls ->BC_ISNEN_Z
    }
    |  // RC is not a number.
    |   asr TMP0, CARG1, #47
    |.if FFI
    |  // Check if RC or RA is a cdata.
    |  cmn ITYPE, #-LJ_TCDATA
    |   ccmn TMP0, #-LJ_TCDATA, #4, ne
    |  beq ->vmeta_equal_cd
    |.endif
    |  cmp CARG1, CARG3
    |  bne >2
    |  // Tag and value are equal.
    if (vk) {
      |->BC_ISEQV_Z:
      |  mov PC, RB			// Perform branch.
    }
    |1:
    |  ins_next
    |
    |2:  // Check if the tags are the same and it's a table or userdata.
    |  cmp ITYPE, TMP0
    |  ccmn ITYPE, #-LJ_TISTABUD, #2, eq
    if (vk) {
      |  bhi <1
    } else {
      |  bhi ->BC_ISEQV_Z		// Reuse code from opposite instruction.
    }
    |  // Different tables or userdatas. Need to check __eq metamethod.
    |  // Field metatable must be at same offset for GCtab and GCudata!
    |  and TAB:CARG2, CARG1, #LJ_GCVMASK
    |  ldr TAB:TMP2, TAB:CARG2->metatable
    if (vk) {
      |  cbz TAB:TMP2, <1		// No metatable?
      |  ldrb TMP1w, TAB:TMP2->nomm
      |   mov CARG4, #0			// ne = 0
      |  tbnz TMP1w, #MM_eq, <1		// 'no __eq' flag set: done.
    } else {
      |  cbz TAB:TMP2, ->BC_ISEQV_Z	// No metatable?
      |  ldrb TMP1w, TAB:TMP2->nomm
      |   mov CARG4, #1			// ne = 1.
      |  tbnz TMP1w, #MM_eq, ->BC_ISEQV_Z	// 'no __eq' flag set: done.
    }
    |  b ->vmeta_equal
    break;

  case BC_ISEQS: case BC_ISNES:
    vk = op == BC_ISEQS;
    |  // RA = src, RC = str_const (~), JMP with RC = target
    |  ldr CARG1, [BASE, RA, lsl #3]
    |   mvn RC, RC
    |    ldrh RBw, [PC, # OFS_RD]
    |   ldr CARG2, [KBASE, RC, lsl #3]
    |    add PC, PC, #4
    |   movn TMP0, #~LJ_TSTR
    |.if FFI
    |  asr ITYPE, CARG1, #47
    |.endif
    |    add RB, PC, RB, lsl #2
    |   add CARG2, CARG2, TMP0, lsl #47
    |    sub RB, RB, #0x20000
    |.if FFI
    |  cmn ITYPE, #-LJ_TCDATA
    |  beq ->vmeta_equal_cd
    |.endif
    |  cmp CARG1, CARG2
    if (vk) {
      |  csel PC, RB, PC, eq
    } else {
      |  csel PC, RB, PC, ne
    }
    |  ins_next
    break;

  case BC_ISEQN: case BC_ISNEN:
    vk = op == BC_ISEQN;
    |  // RA = src, RC = num_const (~), JMP with RC = target
    |  ldr CARG1, [BASE, RA, lsl #3]
    |   add RC, KBASE, RC, lsl #3
    |    ldrh RBw, [PC, # OFS_RD]
    |   ldr CARG3, [RC]
    |    add PC, PC, #4
    |    add RB, PC, RB, lsl #2
    |    sub RB, RB, #0x20000
    if (vk) {
      |->BC_ISEQN_Z:
    } else {
      |->BC_ISNEN_Z:
    }
    |  checkint CARG1, >4
    |   checkint CARG3, >6
    |  cmp CARG1w, CARG3w
    |1:
    if (vk) {
      |  csel PC, RB, PC, eq
      |2:
    } else {
      |2:
      |  csel PC, RB, PC, ne
    }
    |3:
    |  ins_next
    |
    |4:  // RA not int.
    |.if FFI
    |  blo >7
    |.else
    |  blo <2
    |.endif
    |    ldr FARG1, [BASE, RA, lsl #3]
    |    ldr FARG2, [RC]
    |   cmp TISNUMhi, CARG3, lsr #32
    |   bne >5
    |  // RA number, RC int.
    |  scvtf FARG2, CARG3w
    |5:
    |  // RA number, RC number.
    |  fcmp FARG1, FARG2
    |  b <1
    |
    |6:  // RA int, RC number
    |  ldr FARG2, [RC]
    |  scvtf FARG1, CARG1w
    |  fcmp FARG1, FARG2
    |  b <1
    |
    |.if FFI
    |7:
    |  asr ITYPE, CARG1, #47
    |  cmn ITYPE, #-LJ_TCDATA
    |  bne <2
    |  b ->vmeta_equal_cd
    |.endif
    break;

  case BC_ISEQP: case BC_ISNEP:
    vk = op == BC_ISEQP;
    |  // RA = src, RC = primitive_type (~), JMP with RC = target
    |  ldr TMP0, [BASE, RA, lsl #3]
    |   ldrh RBw, [PC, # OFS_RD]
    |   add PC, PC, #4
    |  add RC, RC, #1
    |   add RB, PC, RB, lsl #2
    |.if FFI
    |  asr ITYPE, TMP0, #47
    |  cmn ITYPE, #-LJ_TCDATA
    |  beq ->vmeta_equal_cd
    |  cmn RC, ITYPE
    |.else
    |  cmn RC, TMP0, asr #47
    |.endif
    |   sub RB, RB, #0x20000
    if (vk) {
      |  csel PC, RB, PC, eq
    } else {
      |  csel PC, RB, PC, ne
    }
    |  ins_next
    break;

  /* -- Unary test and copy ops ------------------------------------------- */

  case BC_ISTC: case BC_ISFC: case BC_IST: case BC_ISF:
    |  // RA = dst or unused, RC = src, JMP with RC = target
    |   ldrh RBw, [PC, # OFS_RD]
    |  ldr TMP0, [BASE, RC, lsl #3]
    |   add PC, PC, #4
    |  mov_false TMP1
    |   add RB, PC, RB, lsl #2
    |  cmp TMP0, TMP1
    |   sub RB, RB, #0x20000
    if (op == BC_ISTC || op == BC_IST) {
      if (op == BC_ISTC) {
	|  csel RA, RA, RC, lo
      }
      |  csel PC, RB, PC, lo
    } else {
      if (op == BC_ISFC) {
	|  csel RA, RA, RC, hs
      }
      |  csel PC, RB, PC, hs
    }
    if (op == BC_ISTC || op == BC_ISFC) {
      |  str TMP0, [BASE, RA, lsl #3]
    }
    |  ins_next
    break;

  case BC_ISTYPE:
    |  // RA = src, RC = -type
    |  ldr TMP0, [BASE, RA, lsl #3]
    |  cmn RC, TMP0, asr #47
    |  bne ->vmeta_istype
    |  ins_next
    break;
  case BC_ISNUM:
    |  // RA = src, RC = -(TISNUM-1)
    |  ldr TMP0, [BASE, RA]
    |  checknum TMP0, ->vmeta_istype
    |  ins_next
    break;

  /* -- Unary ops --------------------------------------------------------- */

  case BC_MOV:
    |  // RA = dst, RC = src
    |  ldr TMP0, [BASE, RC, lsl #3]
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_NOT:
    |  // RA = dst, RC = src
    |  ldr TMP0, [BASE, RC, lsl #3]
    |   mov_false TMP1
    |   mov_true TMP2
    |  cmp TMP0, TMP1
    |  csel TMP0, TMP1, TMP2, lo
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_UNM:
    |  // RA = dst, RC = src
    |  ldr TMP0, [BASE, RC, lsl #3]
    |  asr ITYPE, TMP0, #47
    |  cmn ITYPE, #-LJ_TISNUM
    |  bhi ->vmeta_unm
    |  eor TMP0, TMP0, #U64x(80000000,00000000)
    |  bne >5
    |  negs TMP0w, TMP0w
    |   movz CARG3, #0x41e0, lsl #48	// 2^31.
    |   add TMP0, TMP0, TISNUM
    |  csel TMP0, TMP0, CARG3, vc
    |5:
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_LEN:
    |  // RA = dst, RC = src
    |  ldr CARG1, [BASE, RC, lsl #3]
    |  asr ITYPE, CARG1, #47
    |  cmn ITYPE, #-LJ_TSTR
    |   and CARG1, CARG1, #LJ_GCVMASK
    |  bne >2
    |  ldr CARG1w, STR:CARG1->len
    |1:
    |  add CARG1, CARG1, TISNUM
    |  str CARG1, [BASE, RA, lsl #3]
    |  ins_next
    |
    |2:
    |  cmn ITYPE, #-LJ_TTAB
    |  bne ->vmeta_len
#if LJ_52
    |  ldr TAB:CARG2, TAB:CARG1->metatable
    |  cbnz TAB:CARG2, >9
    |3:
#endif
    |->BC_LEN_Z:
    |  bl extern lj_tab_len		// (GCtab *t)
    |  // Returns uint32_t (but less than 2^31).
    |  b <1
    |
#if LJ_52
    |9:
    |  ldrb TMP1w, TAB:CARG2->nomm
    |  tbnz TMP1w, #MM_len, <3		// 'no __len' flag set: done.
    |  b ->vmeta_len
#endif
    break;

  /* -- Binary ops -------------------------------------------------------- */

    |.macro ins_arithcheck_int, target
    |  checkint CARG1, target
    |  checkint CARG2, target
    |.endmacro
    |
    |.macro ins_arithcheck_num, target
    |  checknum CARG1, target
    |  checknum CARG2, target
    |.endmacro
    |
    |.macro ins_arithcheck_nzdiv, target
    |  cbz CARG2w, target
    |.endmacro
    |
    |.macro ins_arithhead
    ||vk = ((int)op - BC_ADDVN) / (BC_ADDNV-BC_ADDVN);
    ||if (vk == 1) {
    |   and RC, RC, #255
    |    decode_RB RB, INS
    ||} else {
    |   decode_RB RB, INS
    |    and RC, RC, #255
    ||}
    |.endmacro
    |
    |.macro ins_arithload, reg1, reg2
    |  // RA = dst, RB = src1, RC = src2 | num_const
    ||switch (vk) {
    ||case 0:
    |   ldr reg1, [BASE, RB, lsl #3]
    |    ldr reg2, [KBASE, RC, lsl #3]
    ||  break;
    ||case 1:
    |   ldr reg1, [KBASE, RC, lsl #3]
    |    ldr reg2, [BASE, RB, lsl #3]
    ||  break;
    ||default:
    |   ldr reg1, [BASE, RB, lsl #3]
    |    ldr reg2, [BASE, RC, lsl #3]
    ||  break;
    ||}
    |.endmacro
    |
    |.macro ins_arithfallback, ins
    ||switch (vk) {
    ||case 0:
    |   ins ->vmeta_arith_vn
    ||  break;
    ||case 1:
    |   ins ->vmeta_arith_nv
    ||  break;
    ||default:
    |   ins ->vmeta_arith_vv
    ||  break;
    ||}
    |.endmacro
    |
    |.macro ins_arithmod, res, reg1, reg2
    |  fdiv d2, reg1, reg2
    |  frintm d2, d2
    |  fmsub res, d2, reg2, reg1
    |.endmacro
    |
    |.macro ins_arithdn, intins, fpins
    |  ins_arithhead
    |  ins_arithload CARG1, CARG2
    |  ins_arithcheck_int >5
    |.if "intins" == "smull"
    |  smull CARG1, CARG1w, CARG2w
    |  cmp CARG1, CARG1, sxtw
    |   mov CARG1w, CARG1w
    |  ins_arithfallback bne
    |.elif "intins" == "ins_arithmodi"
    |  ins_arithfallback ins_arithcheck_nzdiv
    |  bl ->vm_modi
    |.else
    |  intins CARG1w, CARG1w, CARG2w
    |  ins_arithfallback bvs
    |.endif
    |  add CARG1, CARG1, TISNUM
    |  str CARG1, [BASE, RA, lsl #3]
    |4:
    |  ins_next
    |
    |5:  // FP variant.
    |  ins_arithload FARG1, FARG2
    |  ins_arithfallback ins_arithcheck_num
    |  fpins FARG1, FARG1, FARG2
    |  str FARG1, [BASE, RA, lsl #3]
    |  b <4
    |.endmacro
    |
    |.macro ins_arithfp, fpins
    |  ins_arithhead
    |  ins_arithload CARG1, CARG2
    |  ins_arithload FARG1, FARG2
    |  ins_arithfallback ins_arithcheck_num
    |.if "fpins" == "fpow"
    |  bl extern pow
    |.else
    |  fpins FARG1, FARG1, FARG2
    |.endif
    |  str FARG1, [BASE, RA, lsl #3]
    |  ins_next
    |.endmacro

  case BC_ADDVN: case BC_ADDNV: case BC_ADDVV:
    |  ins_arithdn adds, fadd
    break;
  case BC_SUBVN: case BC_SUBNV: case BC_SUBVV:
    |  ins_arithdn subs, fsub
    break;
  case BC_MULVN: case BC_MULNV: case BC_MULVV:
    |  ins_arithdn smull, fmul
    break;
  case BC_DIVVN: case BC_DIVNV: case BC_DIVVV:
    |  ins_arithfp fdiv
    break;
  case BC_MODVN: case BC_MODNV: case BC_MODVV:
    |  ins_arithdn ins_arithmodi, ins_arithmod
    break;
  case BC_POW:
    |  // NYI: (partial) integer arithmetic.
    |  ins_arithfp fpow
    break;

  case BC_CAT:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = src_start, RC = src_end
    |   str BASE, L->base
    |  sub CARG3, RC, RB
    |  add CARG2, BASE, RC, lsl #3
    |->BC_CAT_Z:
    |  // RA = dst, CARG2 = top-1, CARG3 = left
    |  mov CARG1, L
    |   str PC, SAVE_PC
    |  bl extern lj_meta_cat		// (lua_State *L, TValue *top, int left)
    |  // Returns NULL (finished) or TValue * (metamethod).
    |  ldrb RBw, [PC, #-4+OFS_RB]
    |   ldr BASE, L->base
    |   cbnz CRET1, ->vmeta_binop
    |  ldr TMP0, [BASE, RB, lsl #3]
    |  str TMP0, [BASE, RA, lsl #3]	// Copy result to RA.
    |  ins_next
    break;

  /* -- Constant ops ------------------------------------------------------ */

  case BC_KSTR:
    |  // RA = dst, RC = str_const (~)
    |  mvn RC, RC
    |  ldr TMP0, [KBASE, RC, lsl #3]
    |   movn TMP1, #~LJ_TSTR
    |  add TMP0, TMP0, TMP1, lsl #47
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_KCDATA:
    |.if FFI
    |  // RA = dst, RC = cdata_const (~)
    |  mvn RC, RC
    |  ldr TMP0, [KBASE, RC, lsl #3]
    |   movn TMP1, #~LJ_TCDATA
    |  add TMP0, TMP0, TMP1, lsl #47
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    |.endif
    break;
  case BC_KSHORT:
    |  // RA = dst, RC = int16_literal
    |  sxth RCw, RCw
    |  add TMP0, RC, TISNUM
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_KNUM:
    |  // RA = dst, RC = num_const
    |  ldr TMP0, [KBASE, RC, lsl #3]
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_KPRI:
    |  // RA = dst, RC = primitive_type (~)
    |  mvn TMP0, RC, lsl #47
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_KNIL:
    |  // RA = base, RC = end
    |  add RA, BASE, RA, lsl #3
    |   add RC, BASE, RC, lsl #3
    |  str TISNIL, [RA], #8
    |1:
    |   cmp RA, RC
    |  str TISNIL, [RA], #8
    |   blt <1
    |  ins_next_
    break;

  /* -- Upvalue and function ops ------------------------------------------ */

  case BC_UGET:
    |  // RA = dst, RC = uvnum
    |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
    |   add RC, RC, #offsetof(GCfuncL, uvptr)/8
    |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
    |  ldr UPVAL:CARG2, [LFUNC:CARG2, RC, lsl #3]
    |  ldr CARG2, UPVAL:CARG2->v
    |  ldr TMP0, [CARG2]
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;
  case BC_USETV:
    |  // RA = uvnum, RC = src
    |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
    |   add RA, RA, #offsetof(GCfuncL, uvptr)/8
    |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
    |  ldr UPVAL:CARG1, [LFUNC:CARG2, RA, lsl #3]
    |   ldr CARG3, [BASE, RC, lsl #3]
    |    ldr CARG2, UPVAL:CARG1->v
    |  ldrb TMP2w, UPVAL:CARG1->marked
    |  ldrb TMP0w, UPVAL:CARG1->closed
    |    asr ITYPE, CARG3, #47
    |   str CARG3, [CARG2]
    |    add ITYPE, ITYPE, #-LJ_TISGCV
    |  tst TMP2w, #LJ_GC_BLACK		// isblack(uv)
    |  ccmp TMP0w, #0, #4, ne		// && uv->closed
    |    ccmn ITYPE, #-(LJ_TNUMX - LJ_TISGCV), #0, ne	// && tvisgcv(v)
    |  bhi >2
    |1:
    |  ins_next
    |
    |2:  // Check if new value is white.
    |  and GCOBJ:CARG3, CARG3, #LJ_GCVMASK
    |  ldrb TMP1w, GCOBJ:CARG3->gch.marked
    |  tst TMP1w, #LJ_GC_WHITES		// iswhite(str)
    |  beq <1
    |  // Crossed a write barrier. Move the barrier forward.
    |  mov CARG1, GL
    |  bl extern lj_gc_barrieruv	// (global_State *g, TValue *tv)
    |  b <1
    break;
  case BC_USETS:
    |  // RA = uvnum, RC = str_const (~)
    |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
    |   add RA, RA, #offsetof(GCfuncL, uvptr)/8
    |    mvn RC, RC
    |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
    |  ldr UPVAL:CARG1, [LFUNC:CARG2, RA, lsl #3]
    |   ldr STR:CARG3, [KBASE, RC, lsl #3]
    |   movn TMP0, #~LJ_TSTR
    |    ldr CARG2, UPVAL:CARG1->v
    |  ldrb TMP2w, UPVAL:CARG1->marked
    |   add TMP0, STR:CARG3, TMP0, lsl #47
    |    ldrb TMP1w, STR:CARG3->marked
    |   str TMP0, [CARG2]
    |  tbnz TMP2w, #2, >2		// isblack(uv)
    |1:
    |  ins_next
    |
    |2:  // Check if string is white and ensure upvalue is closed.
    |  ldrb TMP0w, UPVAL:CARG1->closed
    |    tst TMP1w, #LJ_GC_WHITES	// iswhite(str)
    |  ccmp TMP0w, #0, #4, ne
    |  beq <1
    |  // Crossed a write barrier. Move the barrier forward.
    |  mov CARG1, GL
    |  bl extern lj_gc_barrieruv	// (global_State *g, TValue *tv)
    |  b <1
    break;
  case BC_USETN:
    |  // RA = uvnum, RC = num_const
    |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
    |   add RA, RA, #offsetof(GCfuncL, uvptr)/8
    |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
    |  ldr UPVAL:CARG2, [LFUNC:CARG2, RA, lsl #3]
    |   ldr TMP0, [KBASE, RC, lsl #3]
    |  ldr CARG2, UPVAL:CARG2->v
    |   str TMP0, [CARG2]
    |  ins_next
    break;
  case BC_USETP:
    |  // RA = uvnum, RC = primitive_type (~)
    |  ldr LFUNC:CARG2, [BASE, FRAME_FUNC]
    |   add RA, RA, #offsetof(GCfuncL, uvptr)/8
    |  and LFUNC:CARG2, CARG2, #LJ_GCVMASK
    |  ldr UPVAL:CARG2, [LFUNC:CARG2, RA, lsl #3]
    |   mvn TMP0, RC, lsl #47
    |  ldr CARG2, UPVAL:CARG2->v
    |   str TMP0, [CARG2]
    |  ins_next
    break;

  case BC_UCLO:
    |  // RA = level, RC = target
    |  ldr CARG3, L->openupval
    |   add RC, PC, RC, lsl #2
    |    str BASE, L->base
    |   sub PC, RC, #0x20000
    |  cbz CARG3, >1
    |  mov CARG1, L
    |  add CARG2, BASE, RA, lsl #3
    |  bl extern lj_func_closeuv	// (lua_State *L, TValue *level)
    |  ldr BASE, L->base
    |1:
    |  ins_next
    break;

  case BC_FNEW:
    |  // RA = dst, RC = proto_const (~) (holding function prototype)
    |  mvn RC, RC
    |   str BASE, L->base
    |  ldr LFUNC:CARG3, [BASE, FRAME_FUNC]
    |    str PC, SAVE_PC
    |   ldr CARG2, [KBASE, RC, lsl #3]
    |    mov CARG1, L
    |  and LFUNC:CARG3, CARG3, #LJ_GCVMASK
    |  // (lua_State *L, GCproto *pt, GCfuncL *parent)
    |  bl extern lj_func_newL_gc
    |  // Returns GCfuncL *.
    |  ldr BASE, L->base
    |   movn TMP0, #~LJ_TFUNC
    |   add CRET1, CRET1, TMP0, lsl #47
    |  str CRET1, [BASE, RA, lsl #3]
    |  ins_next
    break;

  /* -- Table ops --------------------------------------------------------- */

  case BC_TNEW:
  case BC_TDUP:
    |  // RA = dst, RC = (hbits|asize) | tab_const (~)
    |  ldp CARG3, CARG4, GL->gc.total	// Assumes threshold follows total.
    |   str BASE, L->base
    |   str PC, SAVE_PC
    |   mov CARG1, L
    |  cmp CARG3, CARG4
    |  bhs >5
    |1:
    if (op == BC_TNEW) {
      |  and CARG2, RC, #0x7ff
      |   lsr CARG3, RC, #11
      |  cmp CARG2, #0x7ff
      |  mov TMP0, #0x801
      |  csel CARG2, CARG2, TMP0, ne
      |  bl extern lj_tab_new  // (lua_State *L, int32_t asize, uint32_t hbits)
      |  // Returns GCtab *.
    } else {
      |  mvn RC, RC
      |  ldr CARG2, [KBASE, RC, lsl #3]
      |  bl extern lj_tab_dup  // (lua_State *L, Table *kt)
      |  // Returns GCtab *.
    }
    |  ldr BASE, L->base
    |   movk CRET1, #(LJ_TTAB>>1)&0xffff, lsl #48
    |  str CRET1, [BASE, RA, lsl #3]
    |  ins_next
    |
    |5:
    |  bl extern lj_gc_step_fixtop  // (lua_State *L)
    |  mov CARG1, L
    |  b <1
    break;

  case BC_GGET:
    |  // RA = dst, RC = str_const (~)
  case BC_GSET:
    |  // RA = src, RC = str_const (~)
    |  ldr LFUNC:CARG1, [BASE, FRAME_FUNC]
    |   mvn RC, RC
    |  and LFUNC:CARG1, CARG1, #LJ_GCVMASK
    |  ldr TAB:CARG2, LFUNC:CARG1->env
    |   ldr STR:RC, [KBASE, RC, lsl #3]
    if (op == BC_GGET) {
      |  b ->BC_TGETS_Z
    } else {
      |  b ->BC_TSETS_Z
    }
    break;

  case BC_TGETV:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = table, RC = key
    |  ldr CARG2, [BASE, RB, lsl #3]
    |   ldr TMP1, [BASE, RC, lsl #3]
    |  checktab CARG2, ->vmeta_tgetv
    |  checkint TMP1, >9		// Integer key?
    |  ldr CARG3, TAB:CARG2->array
    |   ldr CARG1w, TAB:CARG2->asize
    |  add CARG3, CARG3, TMP1, uxtw #3
    |   cmp TMP1w, CARG1w		// In array part?
    |   bhs ->vmeta_tgetv
    |  ldr TMP0, [CARG3]
    |  cmp TMP0, TISNIL
    |  beq >5
    |1:
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    |
    |5:  // Check for __index if table value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <1		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_index, <1	// 'no __index' flag set: done.
    |  b ->vmeta_tgetv
    |
    |9:
    |  asr ITYPE, TMP1, #47
    |  cmn ITYPE, #-LJ_TSTR		// String key?
    |  bne ->vmeta_tgetv
    |   and STR:RC, TMP1, #LJ_GCVMASK
    |  b ->BC_TGETS_Z
    break;
  case BC_TGETS:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = table, RC = str_const (~)
    |  ldr CARG2, [BASE, RB, lsl #3]
    |   mvn RC, RC
    |   ldr STR:RC, [KBASE, RC, lsl #3]
    |  checktab CARG2, ->vmeta_tgets1
    |->BC_TGETS_Z:
    |  // TAB:CARG2 = GCtab *, STR:RC = GCstr *, RA = dst
    |  ldr TMP1w, TAB:CARG2->hmask
    |   ldr TMP2w, STR:RC->sid
    |    ldr NODE:CARG3, TAB:CARG2->node
    |  and TMP1w, TMP1w, TMP2w		// idx = str->sid & tab->hmask
    |  add TMP1, TMP1, TMP1, lsl #1
    |  movn CARG4, #~LJ_TSTR
    |    add NODE:CARG3, NODE:CARG3, TMP1, lsl #3  // node = tab->node + idx*3*8
    |  add CARG4, STR:RC, CARG4, lsl #47	// Tagged key to look for.
    |1:
    |  ldp TMP0, CARG1, NODE:CARG3->val
    |   ldr NODE:CARG3, NODE:CARG3->next
    |  cmp CARG1, CARG4
    |  bne >4
    |  cmp TMP0, TISNIL
    |  beq >5
    |3:
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    |
    |4:  // Follow hash chain.
    |  cbnz NODE:CARG3, <1
    |  // End of hash chain: key not found, nil result.
    |   mov TMP0, TISNIL
    |
    |5:  // Check for __index if table value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <3		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_index, <3	// 'no __index' flag set: done.
    |  b ->vmeta_tgets
    break;
  case BC_TGETB:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = table, RC = index
    |  ldr CARG2, [BASE, RB, lsl #3]
    |  checktab CARG2, ->vmeta_tgetb
    |  ldr CARG3, TAB:CARG2->array
    |   ldr CARG1w, TAB:CARG2->asize
    |  add CARG3, CARG3, RC, lsl #3
    |   cmp RCw, CARG1w			// In array part?
    |   bhs ->vmeta_tgetb
    |  ldr TMP0, [CARG3]
    |  cmp TMP0, TISNIL
    |  beq >5
    |1:
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    |
    |5:  // Check for __index if table value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <1		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_index, <1	// 'no __index' flag set: done.
    |  b ->vmeta_tgetb
    break;
  case BC_TGETR:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = table, RC = key
    |  ldr CARG1, [BASE, RB, lsl #3]
    |   ldr TMP1, [BASE, RC, lsl #3]
    |  and TAB:CARG1, CARG1, #LJ_GCVMASK
    |  ldr CARG3, TAB:CARG1->array
    |   ldr TMP2w, TAB:CARG1->asize
    |  add CARG3, CARG3, TMP1w, uxtw #3
    |   cmp TMP1w, TMP2w		// In array part?
    |   bhs ->vmeta_tgetr
    |  ldr TMP0, [CARG3]
    |->BC_TGETR_Z:
    |  str TMP0, [BASE, RA, lsl #3]
    |  ins_next
    break;

  case BC_TSETV:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = src, RB = table, RC = key
    |  ldr CARG2, [BASE, RB, lsl #3]
    |   ldr TMP1, [BASE, RC, lsl #3]
    |  checktab CARG2, ->vmeta_tsetv
    |  checkint TMP1, >9		// Integer key?
    |  ldr CARG3, TAB:CARG2->array
    |   ldr CARG1w, TAB:CARG2->asize
    |  add CARG3, CARG3, TMP1, uxtw #3
    |   cmp TMP1w, CARG1w		// In array part?
    |   bhs ->vmeta_tsetv
    |  ldr TMP1, [CARG3]
    |   ldr TMP0, [BASE, RA, lsl #3]
    |    ldrb TMP2w, TAB:CARG2->marked
    |  cmp TMP1, TISNIL			// Previous value is nil?
    |  beq >5
    |1:
    |   str TMP0, [CARG3]
    |    tbnz TMP2w, #2, >7		// isblack(table)
    |2:
    |   ins_next
    |
    |5:  // Check for __newindex if previous value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <1		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_newindex, <1	// 'no __newindex' flag set: done.
    |  b ->vmeta_tsetv
    |
    |7:  // Possible table write barrier for the value. Skip valiswhite check.
    |  barrierback TAB:CARG2, TMP2w, TMP1
    |  b <2
    |
    |9:
    |  asr ITYPE, TMP1, #47
    |  cmn ITYPE, #-LJ_TSTR		// String key?
    |  bne ->vmeta_tsetv
    |   and STR:RC, TMP1, #LJ_GCVMASK
    |  b ->BC_TSETS_Z
    break;
  case BC_TSETS:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = dst, RB = table, RC = str_const (~)
    |  ldr CARG2, [BASE, RB, lsl #3]
    |   mvn RC, RC
    |   ldr STR:RC, [KBASE, RC, lsl #3]
    |  checktab CARG2, ->vmeta_tsets1
    |->BC_TSETS_Z:
    |  // TAB:CARG2 = GCtab *, STR:RC = GCstr *, RA = src
    |  ldr TMP1w, TAB:CARG2->hmask
    |   ldr TMP2w, STR:RC->sid
    |    ldr NODE:CARG3, TAB:CARG2->node
    |  and TMP1w, TMP1w, TMP2w		// idx = str->sid & tab->hmask
    |  add TMP1, TMP1, TMP1, lsl #1
    |  movn CARG4, #~LJ_TSTR
    |    add NODE:CARG3, NODE:CARG3, TMP1, lsl #3  // node = tab->node + idx*3*8
    |  add CARG4, STR:RC, CARG4, lsl #47	// Tagged key to look for.
    |   strb wzr, TAB:CARG2->nomm	// Clear metamethod cache.
    |1:
    |  ldp TMP1, CARG1, NODE:CARG3->val
    |   ldr NODE:TMP3, NODE:CARG3->next
    |    ldrb TMP2w, TAB:CARG2->marked
    |  cmp CARG1, CARG4
    |  bne >5
    |   ldr TMP0, [BASE, RA, lsl #3]
    |  cmp TMP1, TISNIL			// Previous value is nil?
    |  beq >4
    |2:
    |   str TMP0, NODE:CARG3->val
    |    tbnz TMP2w, #2, >7		// isblack(table)
    |3:
    |  ins_next
    |
    |4:  // Check for __newindex if previous value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <2		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_newindex, <2	// 'no __newindex' flag set: done.
    |  b ->vmeta_tsets
    |
    |5:  // Follow hash chain.
    |  mov NODE:CARG3, NODE:TMP3
    |  cbnz NODE:TMP3, <1
    |  // End of hash chain: key not found, add a new one.
    |
    |  // But check for __newindex first.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, >6		// No metatable: continue.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  // 'no __newindex' flag NOT set: check.
    |  tbz TMP1w, #MM_newindex, ->vmeta_tsets
    |6:
    |  movn TMP1, #~LJ_TSTR
    |   str PC, SAVE_PC
    |  add TMP0, STR:RC, TMP1, lsl #47
    |   str BASE, L->base
    |   mov CARG1, L
    |  str TMP0, TMPD
    |   add CARG3, sp, TMPDofs
    |  bl extern lj_tab_newkey		// (lua_State *L, GCtab *t, TValue *k)
    |  // Returns TValue *.
    |  ldr BASE, L->base
    |  ldr TMP0, [BASE, RA, lsl #3]
    |  str TMP0, [CRET1]
    |  b <3				// No 2nd write barrier needed.
    |
    |7:  // Possible table write barrier for the value. Skip valiswhite check.
    |  barrierback TAB:CARG2, TMP2w, TMP1
    |  b <3
    break;
  case BC_TSETB:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = src, RB = table, RC = index
    |  ldr CARG2, [BASE, RB, lsl #3]
    |  checktab CARG2, ->vmeta_tsetb
    |  ldr CARG3, TAB:CARG2->array
    |   ldr CARG1w, TAB:CARG2->asize
    |  add CARG3, CARG3, RC, lsl #3
    |   cmp RCw, CARG1w			// In array part?
    |   bhs ->vmeta_tsetb
    |  ldr TMP1, [CARG3]
    |   ldr TMP0, [BASE, RA, lsl #3]
    |    ldrb TMP2w, TAB:CARG2->marked
    |  cmp TMP1, TISNIL			// Previous value is nil?
    |  beq >5
    |1:
    |   str TMP0, [CARG3]
    |    tbnz TMP2w, #2, >7		// isblack(table)
    |2:
    |   ins_next
    |
    |5:  // Check for __newindex if previous value is nil.
    |  ldr TAB:CARG1, TAB:CARG2->metatable
    |  cbz TAB:CARG1, <1		// No metatable: done.
    |  ldrb TMP1w, TAB:CARG1->nomm
    |  tbnz TMP1w, #MM_newindex, <1	// 'no __newindex' flag set: done.
    |  b ->vmeta_tsetb
    |
    |7:  // Possible table write barrier for the value. Skip valiswhite check.
    |  barrierback TAB:CARG2, TMP2w, TMP1
    |  b <2
    break;
  case BC_TSETR:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = src, RB = table, RC = key
    |  ldr CARG2, [BASE, RB, lsl #3]
    |   ldr TMP1, [BASE, RC, lsl #3]
    |  and TAB:CARG2, CARG2, #LJ_GCVMASK
    |  ldr CARG1, TAB:CARG2->array
    |    ldrb TMP2w, TAB:CARG2->marked
    |   ldr CARG4w, TAB:CARG2->asize
    |  add CARG1, CARG1, TMP1, uxtw #3
    |    tbnz TMP2w, #2, >7		// isblack(table)
    |2:
    |   cmp TMP1w, CARG4w		// In array part?
    |   bhs ->vmeta_tsetr
    |->BC_TSETR_Z:
    |   ldr TMP0, [BASE, RA, lsl #3]
    |   str TMP0, [CARG1]
    |   ins_next
    |
    |7:  // Possible table write barrier for the value. Skip valiswhite check.
    |  barrierback TAB:CARG2, TMP2w, TMP0
    |  b <2
    break;

  case BC_TSETM:
    |  // RA = base (table at base-1), RC = num_const (start index)
    |  add RA, BASE, RA, lsl #3
    |1:
    |   ldr RBw, SAVE_MULTRES
    |  ldr TAB:CARG2, [RA, #-8]		// Guaranteed to be a table.
    |   ldr TMP1, [KBASE, RC, lsl #3]	// Integer constant is in lo-word.
    |    sub RB, RB, #8
    |    cbz RB, >4			// Nothing to copy?
    |  and TAB:CARG2, CARG2, #LJ_GCVMASK
    |  ldr CARG1w, TAB:CARG2->asize
    |   add CARG3w, TMP1w, RBw, lsr #3
    |   ldr CARG4, TAB:CARG2->array
    |  cmp CARG3, CARG1
    |    add RB, RA, RB
    |  bhi >5
    |   add TMP1, CARG4, TMP1w, uxtw #3
    |    ldrb TMP2w, TAB:CARG2->marked
    |3:  // Copy result slots to table.
    |   ldr TMP0, [RA], #8
    |   str TMP0, [TMP1], #8
    |  cmp RA, RB
    |  blo <3
    |    tbnz TMP2w, #2, >7		// isblack(table)
    |4:
    |  ins_next
    |
    |5:  // Need to resize array part.
    |   str BASE, L->base
    |  mov CARG1, L
    |   str PC, SAVE_PC
    |  bl extern lj_tab_reasize		// (lua_State *L, GCtab *t, int nasize)
    |  // Must not reallocate the stack.
    |  b <1
    |
    |7:  // Possible table write barrier for any value. Skip valiswhite check.
    |  barrierback TAB:CARG2, TMP2w, TMP1
    |  b <4
    break;

  /* -- Calls and vararg handling ----------------------------------------- */

  case BC_CALLM:
    |  // RA = base, (RB = nresults+1,) RC = extra_nargs
    |  ldr TMP0w, SAVE_MULTRES
    |  decode_RC8RD NARGS8:RC, RC
    |  add NARGS8:RC, NARGS8:RC, TMP0
    |  b ->BC_CALL_Z
    break;
  case BC_CALL:
    |  decode_RC8RD NARGS8:RC, RC
    |  // RA = base, (RB = nresults+1,) RC = (nargs+1)*8
    |->BC_CALL_Z:
    |  mov RB, BASE			// Save old BASE for vmeta_call.
    |  add BASE, BASE, RA, lsl #3
    |  ldr CARG3, [BASE]
    |   sub NARGS8:RC, NARGS8:RC, #8
    |   add BASE, BASE, #16
    |  checkfunc CARG3, ->vmeta_call
    |  ins_call
    break;

  case BC_CALLMT:
    |  // RA = base, (RB = 0,) RC = extra_nargs
    |  ldr TMP0w, SAVE_MULTRES
    |  add NARGS8:RC, TMP0, RC, lsl #3
    |  b ->BC_CALLT1_Z
    break;
  case BC_CALLT:
    |  lsl NARGS8:RC, RC, #3
    |  // RA = base, (RB = 0,) RC = (nargs+1)*8
    |->BC_CALLT1_Z:
    |  add RA, BASE, RA, lsl #3
    |  ldr TMP1, [RA]
    |   sub NARGS8:RC, NARGS8:RC, #8
    |   add RA, RA, #16
    |  checktp CARG3, TMP1, LJ_TFUNC, ->vmeta_callt
    |  ldr PC, [BASE, FRAME_PC]
    |->BC_CALLT2_Z:
    |   mov RB, #0
    |   ldrb TMP2w, LFUNC:CARG3->ffid
    |  tst PC, #FRAME_TYPE
    |  bne >7
    |1:
    |  str TMP1, [BASE, FRAME_FUNC]	// Copy function down, but keep PC.
    |  cbz NARGS8:RC, >3
    |2:
    |  ldr TMP0, [RA, RB]
    |   add TMP1, RB, #8
    |   cmp TMP1, NARGS8:RC
    |  str TMP0, [BASE, RB]
    |    mov RB, TMP1
    |   bne <2
    |3:
    |  cmp TMP2, #1			// (> FF_C) Calling a fast function?
    |  bhi >5
    |4:
    |  ins_callt
    |
    |5:  // Tailcall to a fast function with a Lua frame below.
    |  ldrb RAw, [PC, #-4+OFS_RA]
    |  sub CARG1, BASE, RA, lsl #3
    |  ldr LFUNC:CARG1, [CARG1, #-32]
    |  and LFUNC:CARG1, CARG1, #LJ_GCVMASK
    |  ldr CARG1, LFUNC:CARG1->pc
    |  ldr KBASE, [CARG1, #PC2PROTO(k)]
    |  b <4
    |
    |7:  // Tailcall from a vararg function.
    |  eor PC, PC, #FRAME_VARG
    |  tst PC, #FRAME_TYPEP		// Vararg frame below?
    |  csel TMP2, RB, TMP2, ne		// Clear ffid if no Lua function below.
    |  bne <1
    |  sub BASE, BASE, PC
    |  ldr PC, [BASE, FRAME_PC]
    |  tst PC, #FRAME_TYPE
    |  csel TMP2, RB, TMP2, ne		// Clear ffid if no Lua function below.
    |  b <1
    break;

  case BC_ITERC:
    |  // RA = base, (RB = nresults+1, RC = nargs+1 (2+1))
    |  add RA, BASE, RA, lsl #3
    |  ldr CARG3, [RA, #-24]
    |    mov RB, BASE			// Save old BASE for vmeta_call.
    |   ldp CARG1, CARG2, [RA, #-16]
    |    add BASE, RA, #16
    |    mov NARGS8:RC, #16		// Iterators get 2 arguments.
    |  str CARG3, [RA]			// Copy callable.
    |   stp CARG1, CARG2, [RA, #16]	// Copy state and control var.
    |  checkfunc CARG3, ->vmeta_call
    |  ins_call
    break;

  case BC_ITERN:
    |.if JIT
    |  hotloop
    |.endif
    |->vm_IITERN:
    |  // RA = base, (RB = nresults+1, RC = nargs+1 (2+1))
    |  add RA, BASE, RA, lsl #3
    |  ldr TAB:RB, [RA, #-16]
    |    ldrh TMP3w, [PC, # OFS_RD]
    |  ldr CARG1w, [RA, #-8+LO]		// Get index from control var.
    |    add PC, PC, #4
    |    add TMP3, PC, TMP3, lsl #2
    |  and TAB:RB, RB, #LJ_GCVMASK
    |    sub TMP3, TMP3, #0x20000
    |  ldr TMP1w, TAB:RB->asize
    |   ldr CARG2, TAB:RB->array
    |1:  // Traverse array part.
    |  subs RC, CARG1, TMP1
    |   add CARG3, CARG2, CARG1, lsl #3
    |  bhs >5				// Index points after array part?
    |   ldr TMP0, [CARG3]
    |   cmp TMP0, TISNIL
    |   cinc CARG1, CARG1, eq		// Skip holes in array part.
    |   beq <1
    |   add CARG1, CARG1, TISNUM
    |   stp CARG1, TMP0, [RA]
    |    add CARG1, CARG1, #1
    |3:
    |    str CARG1w, [RA, #-8+LO]	// Update control var.
    |  mov PC, TMP3
    |4:
    |  ins_next
    |
    |5:  // Traverse hash part.
    |  ldr TMP2w, TAB:RB->hmask
    |   ldr NODE:RB, TAB:RB->node
    |6:
    |   add CARG1, RC, RC, lsl #1
    |  cmp RC, TMP2			// End of iteration? Branch to ITERN+1.
    |   add NODE:CARG3, NODE:RB, CARG1, lsl #3  // node = tab->node + idx*3*8
    |  bhi <4
    |  ldp TMP0, CARG1, NODE:CARG3->val
    |  cmp TMP0, TISNIL
    |   add RC, RC, #1
    |  beq <6				// Skip holes in hash part.
    |  stp CARG1, TMP0, [RA]
    |  add CARG1, RC, TMP1
    |  b <3
    break;

  case BC_ISNEXT:
    |  // RA = base, RC = target (points to ITERN)
    |  add RA, BASE, RA, lsl #3
    |  ldr CFUNC:CARG1, [RA, #-24]
    |     add RC, PC, RC, lsl #2
    |   ldp TAB:CARG3, CARG4, [RA, #-16]
    |     sub RC, RC, #0x20000
    |  checkfunc CFUNC:CARG1, >5
    |   asr TMP0, TAB:CARG3, #47
    |  ldrb TMP1w, CFUNC:CARG1->ffid
    |   cmn TMP0, #-LJ_TTAB
    |   ccmp CARG4, TISNIL, #0, eq
    |  ccmp TMP1w, #FF_next_N, #0, eq
    |  bne >5
    |  mov TMP0w, #0xfffe7fff		// LJ_KEYINDEX
    |  lsl TMP0, TMP0, #32
    |  str TMP0, [RA, #-8]		// Initialize control var.
    |1:
    |     mov PC, RC
    |  ins_next
    |
    |5:  // Despecialize bytecode if any of the checks fail.
    |.if JIT
    |  ldrb TMP2w, [RC, # OFS_OP]
    |.endif
    |  mov TMP0, #BC_JMP
    |   mov TMP1, #BC_ITERC
    |  strb TMP0w, [PC, #-4+OFS_OP]
    |.if JIT
    |  cmp TMP2w, #BC_ITERN
    |  bne >6
    |.endif
    |   strb TMP1w, [RC, # OFS_OP]
    |  b <1
    |.if JIT
    |6:  // Unpatch JLOOP.
    |  ldr RA, [GL, #GL_J(trace)]
    |  ldrh TMP2w, [RC, # OFS_RD]
    |  ldr TRACE:RA, [RA, TMP2, lsl #3]
    |  ldr TMP2w, TRACE:RA->startins
    |  bfxil TMP2w, TMP1w, #0, #8
    |  str TMP2w, [RC]
    |  b <1
    |.endif
    break;

  case BC_VARG:
    |  decode_RB RB, INS
    |   and RC, RC, #255
    |  // RA = base, RB = (nresults+1), RC = numparams
    |  ldr TMP1, [BASE, FRAME_PC]
    |  add RC, BASE, RC, lsl #3
    |   add RA, BASE, RA, lsl #3
    |  add RC, RC, #FRAME_VARG
    |   add TMP2, RA, RB, lsl #3
    |  sub RC, RC, TMP1			// RC = vbase
    |  // Note: RC may now be even _above_ BASE if nargs was < numparams.
    |   sub TMP3, BASE, #16		// TMP3 = vtop
    |  cbz RB, >5
    |   sub TMP2, TMP2, #16
    |1:  // Copy vararg slots to destination slots.
    |  cmp RC, TMP3
    |  ldr TMP0, [RC], #8
    |  csel TMP0, TMP0, TISNIL, lo
    |   cmp RA, TMP2
    |  str TMP0, [RA], #8
    |   blo <1
    |2:
    |  ins_next
    |
    |5:  // Copy all varargs.
    |  ldr TMP0, L->maxstack
    |   subs TMP2, TMP3, RC
    |   csel RB, xzr, TMP2, le		// MULTRES = (max(vtop-vbase,0)+1)*8
    |   add RB, RB, #8
    |  add TMP1, RA, TMP2
    |   str RBw, SAVE_MULTRES
    |   ble <2				// Nothing to copy.
    |  cmp TMP1, TMP0
    |  bhi >7
    |6:
    |  ldr TMP0, [RC], #8
    |  str TMP0, [RA], #8
    |  cmp RC, TMP3
    |  blo <6
    |  b <2
    |
    |7:  // Grow stack for varargs.
    |  lsr CARG2, TMP2, #3
    |   stp BASE, RA, L->base
    |  mov CARG1, L
    |  sub RC, RC, BASE			// Need delta, because BASE may change.
    |   str PC, SAVE_PC
    |  bl extern lj_state_growstack	// (lua_State *L, int n)
    |  ldp BASE, RA, L->base
    |  add RC, BASE, RC
    |  sub TMP3, BASE, #16
    |  b <6
    break;

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

  case BC_RETM:
    |  // RA = results, RC = extra results
    |  ldr TMP0w, SAVE_MULTRES
    |   ldr PC, [BASE, FRAME_PC]
    |    add RA, BASE, RA, lsl #3
    |  add RC, TMP0, RC, lsl #3
    |  b ->BC_RETM_Z
    break;

  case BC_RET:
    |  // RA = results, RC = nresults+1
    |  ldr PC, [BASE, FRAME_PC]
    |   lsl RC, RC, #3
    |    add RA, BASE, RA, lsl #3
    |->BC_RETM_Z:
    |   str RCw, SAVE_MULTRES
    |1:
    |  ands CARG1, PC, #FRAME_TYPE
    |   eor CARG2, PC, #FRAME_VARG
    |  bne ->BC_RETV2_Z
    |
    |->BC_RET_Z:
    |  // BASE = base, RA = resultptr, RC = (nresults+1)*8, PC = return
    |  ldr INSw, [PC, #-4]
    |  subs TMP1, RC, #8
    |   sub CARG3, BASE, #16
    |  beq >3
    |2:
    |  ldr TMP0, [RA], #8
    |   add BASE, BASE, #8
    |   sub TMP1, TMP1, #8
    |  str TMP0, [BASE, #-24]
    |   cbnz TMP1, <2
    |3:
    |  decode_RA RA, INS
    |  sub CARG4, CARG3, RA, lsl #3
    |   decode_RB RB, INS
    |  ldr LFUNC:CARG1, [CARG4, FRAME_FUNC]
    |5:
    |  cmp RC, RB, lsl #3		// More results expected?
    |  blo >6
    |  and LFUNC:CARG1, CARG1, #LJ_GCVMASK
    |  mov BASE, CARG4
    |  ldr CARG2, LFUNC:CARG1->pc
    |  ldr KBASE, [CARG2, #PC2PROTO(k)]
    |   ins_next
    |
    |6:  // Fill up results with nil.
    |  add BASE, BASE, #8
    |   add RC, RC, #8
    |  str TISNIL, [BASE, #-24]
    |  b <5
    |
    |->BC_RETV1_Z:  // Non-standard return case.
    |  add RA, BASE, RA, lsl #3
    |->BC_RETV2_Z:
    |  tst CARG2, #FRAME_TYPEP
    |  bne ->vm_return
    |  // Return from vararg function: relocate BASE down.
    |  sub BASE, BASE, CARG2
    |  ldr PC, [BASE, FRAME_PC]
    |  b <1
    break;

  case BC_RET0: case BC_RET1:
    |  // RA = results, RC = nresults+1
    |  ldr PC, [BASE, FRAME_PC]
    |   lsl RC, RC, #3
    |   str RCw, SAVE_MULTRES
    |  ands CARG1, PC, #FRAME_TYPE
    |   eor CARG2, PC, #FRAME_VARG
    |  bne ->BC_RETV1_Z
    |   ldr INSw, [PC, #-4]
    if (op == BC_RET1) {
      |  ldr TMP0, [BASE, RA, lsl #3]
    }
    |  sub CARG4, BASE, #16
    |   decode_RA RA, INS
    |  sub BASE, CARG4, RA, lsl #3
    if (op == BC_RET1) {
      |  str TMP0, [CARG4], #8
    }
    |   decode_RB RB, INS
    |  ldr LFUNC:CARG1, [BASE, FRAME_FUNC]
    |5:
    |  cmp RC, RB, lsl #3
    |  blo >6
    |  and LFUNC:CARG1, CARG1, #LJ_GCVMASK
    |  ldr CARG2, LFUNC:CARG1->pc
    |  ldr KBASE, [CARG2, #PC2PROTO(k)]
    |  ins_next
    |
    |6:  // Fill up results with nil.
    |  add RC, RC, #8
    |  str TISNIL, [CARG4], #8
    |  b <5
    break;

  /* -- Loops and branches ------------------------------------------------ */

  |.define FOR_IDX,  [RA];      .define FOR_TIDX,  [RA, #4]
  |.define FOR_STOP, [RA, #8];  .define FOR_TSTOP, [RA, #12]
  |.define FOR_STEP, [RA, #16]; .define FOR_TSTEP, [RA, #20]
  |.define FOR_EXT,  [RA, #24]; .define FOR_TEXT,  [RA, #28]

  case BC_FORL:
    |.if JIT
    |  hotloop
    |.endif
    |  // Fall through. Assumes BC_IFORL follows.
    break;

  case BC_JFORI:
  case BC_JFORL:
#if !LJ_HASJIT
    break;
#endif
  case BC_FORI:
  case BC_IFORL:
    |  // RA = base, RC = target (after end of loop or start of loop)
    vk = (op == BC_IFORL || op == BC_JFORL);
    |  add RA, BASE, RA, lsl #3
    |  ldp CARG1, CARG2, FOR_IDX		// CARG1 = IDX, CARG2 = STOP
    |   ldr CARG3, FOR_STEP			// CARG3 = STEP
    if (op != BC_JFORL) {
      |   add RC, PC, RC, lsl #2
      |   sub RC, RC, #0x20000
    }
    |  checkint CARG1, >5
    if (!vk) {
      |  checkint CARG2, ->vmeta_for
      |   checkint CARG3, ->vmeta_for
      |  tbnz CARG3w, #31, >4
      |  cmp CARG1w, CARG2w
    } else {
      |  adds CARG1w, CARG1w, CARG3w
      |  bvs >2
      |   add TMP0, CARG1, TISNUM
      |  tbnz CARG3w, #31, >4
      |  cmp CARG1w, CARG2w
    }
    |1:
    if (op == BC_FORI) {
      |  csel PC, RC, PC, gt
    } else if (op == BC_JFORI) {
      |  mov PC, RC
      |  ldrh RCw, [RC, #-4+OFS_RD]
    } else if (op == BC_IFORL) {
      |  csel PC, RC, PC, le
    }
    if (vk) {
      |   str TMP0, FOR_IDX
      |   str TMP0, FOR_EXT
    } else {
      |  str CARG1, FOR_EXT
    }
    if (op == BC_JFORI || op == BC_JFORL) {
      |  ble =>BC_JLOOP
    }
    |2:
    |   ins_next
    |
    |4:  // Invert check for negative step.
    |  cmp CARG2w, CARG1w
    |  b <1
    |
    |5:  // FP loop.
    |  ldp d0, d1, FOR_IDX
    |  blo ->vmeta_for
    if (!vk) {
      |  checknum CARG2, ->vmeta_for
      |   checknum CARG3, ->vmeta_for
      |  str d0, FOR_EXT
    } else {
      |  ldr d2, FOR_STEP
      |  fadd d0, d0, d2
    }
    |  tbnz CARG3, #63, >7
    |  fcmp d0, d1
    |6:
    if (vk) {
      |  str d0, FOR_IDX
      |  str d0, FOR_EXT
    }
    if (op == BC_FORI) {
      |  csel PC, RC, PC, hi
    } else if (op == BC_JFORI) {
      |  ldrh RCw, [RC, #-4+OFS_RD]
      |  bls =>BC_JLOOP
    } else if (op == BC_IFORL) {
      |  csel PC, RC, PC, ls
    } else {
      |  bls =>BC_JLOOP
    }
    |  b <2
    |
    |7:  // Invert check for negative step.
    |  fcmp d1, d0
    |  b <6
    break;

  case BC_ITERL:
    |.if JIT
    |  hotloop
    |.endif
    |  // Fall through. Assumes BC_IITERL follows.
    break;

  case BC_JITERL:
#if !LJ_HASJIT
    break;
#endif
  case BC_IITERL:
    |  // RA = base, RC = target
    |  ldr CARG1, [BASE, RA, lsl #3]
    |   add TMP1, BASE, RA, lsl #3
    |  cmp CARG1, TISNIL
    |  beq >1				// Stop if iterator returned nil.
    if (op == BC_JITERL) {
      |  str CARG1, [TMP1, #-8]
      |  b =>BC_JLOOP
    } else {
      |  add TMP0, PC, RC, lsl #2	// Otherwise save control var + branch.
      |  sub PC, TMP0, #0x20000
      |  str CARG1, [TMP1, #-8]
    }
    |1:
    |  ins_next
    break;

  case BC_LOOP:
    |  // RA = base, RC = target (loop extent)
    |  // Note: RA/RC is only used by trace recorder to determine scope/extent
    |  // This opcode does NOT jump, it's only purpose is to detect a hot loop.
    |.if JIT
    |  hotloop
    |.endif
    |  // Fall through. Assumes BC_ILOOP follows.
    break;

  case BC_ILOOP:
    |  // RA = base, RC = target (loop extent)
    |  ins_next
    break;

  case BC_JLOOP:
    |.if JIT
    |  // RA = base (ignored), RC = traceno
    |  ldr CARG1, [GL, #GL_J(trace)]
    |   mov CARG2w, #0  // Traces on ARM64 don't store the trace #, so use 0.
    |  ldr TRACE:RC, [CARG1, RC, lsl #3]
    |   st_vmstate CARG2w
    |  ldr RA, TRACE:RC->mcode
    |   str BASE, GL->jit_base
    |   str L, GL->tmpbuf.L
    |  sub sp, sp, #16	// See SPS_FIXED. Avoids sp adjust in every root trace.
    |  br RA
    |.endif
    break;

  case BC_JMP:
    |  // RA = base (only used by trace recorder), RC = target
    |  add RC, PC, RC, lsl #2
    |  sub PC, RC, #0x20000
    |  ins_next
    break;

  /* -- Function headers -------------------------------------------------- */

  case BC_FUNCF:
    |.if JIT
    |  hotcall
    |.endif
  case BC_FUNCV:  /* NYI: compiled vararg functions. */
    |  // Fall through. Assumes BC_IFUNCF/BC_IFUNCV follow.
    break;

  case BC_JFUNCF:
#if !LJ_HASJIT
    break;
#endif
  case BC_IFUNCF:
    |  // BASE = new base, RA = BASE+framesize*8, CARG3 = LFUNC, RC = nargs*8
    |  ldr CARG1, L->maxstack
    |   ldrb TMP1w, [PC, #-4+PC2PROTO(numparams)]
    |    ldr KBASE, [PC, #-4+PC2PROTO(k)]
    |  cmp RA, CARG1
    |  bhi ->vm_growstack_l
    |2:
    |  cmp NARGS8:RC, TMP1, lsl #3	// Check for missing parameters.
    |  blo >3
    if (op == BC_JFUNCF) {
      |  decode_RD RC, INS
      |  b =>BC_JLOOP
    } else {
      |  ins_next
    }
    |
    |3:  // Clear missing parameters.
    |  str TISNIL, [BASE, NARGS8:RC]
    |  add NARGS8:RC, NARGS8:RC, #8
    |  b <2
    break;

  case BC_JFUNCV:
#if !LJ_HASJIT
    break;
#endif
    |  NYI  // NYI: compiled vararg functions
    break;  /* NYI: compiled vararg functions. */

  case BC_IFUNCV:
    |  // BASE = new base, RA = BASE+framesize*8, CARG3 = LFUNC, RC = nargs*8
    |  ldr CARG1, L->maxstack
    |   movn TMP0, #~LJ_TFUNC
    |   add TMP2, BASE, RC
    |   add LFUNC:CARG3, CARG3, TMP0, lsl #47
    |  add RA, RA, RC
    |   add TMP0, RC, #16+FRAME_VARG
    |   str LFUNC:CARG3, [TMP2], #8	// Store (tagged) copy of LFUNC.
    |    ldr KBASE, [PC, #-4+PC2PROTO(k)]
    |  cmp RA, CARG1
    |   str TMP0, [TMP2], #8		// Store delta + FRAME_VARG.
    |  bhs ->vm_growstack_l
    |   sub RC, TMP2, #16
    |  ldrb TMP1w, [PC, #-4+PC2PROTO(numparams)]
    |   mov RA, BASE
    |   mov BASE, TMP2
    |  cbz TMP1, >2
    |1:
    |  cmp RA, RC			// Less args than parameters?
    |  bhs >3
    |   ldr TMP0, [RA]
    |  sub TMP1, TMP1, #1
    |    str TISNIL, [RA], #8		// Clear old fixarg slot (help the GC).
    |   str TMP0, [TMP2], #8
    |  cbnz TMP1, <1
    |2:
    |  ins_next
    |
    |3:
    |  sub TMP1, TMP1, #1
    |   str TISNIL, [TMP2], #8
    |  cbz TMP1, <2
    |  b <3
    break;

  case BC_FUNCC:
  case BC_FUNCCW:
    |  // BASE = new base, RA = BASE+framesize*8, CARG3 = CFUNC, RC = nargs*8
    if (op == BC_FUNCC) {
      |  ldr CARG4, CFUNC:CARG3->f
    } else {
      |  ldr CARG4, GL->wrapf
    }
    |   add CARG2, RA, NARGS8:RC
    |   ldr CARG1, L->maxstack
    |  add RC, BASE, NARGS8:RC
    |   cmp CARG2, CARG1
    |  stp BASE, RC, L->base
    if (op == BC_FUNCCW) {
      |  ldr CARG2, CFUNC:CARG3->f
    }
    |    mv_vmstate TMP0w, C
    |  mov CARG1, L
    |   bhi ->vm_growstack_c		// Need to grow stack.
    |    st_vmstate TMP0w
    |  blr CARG4			// (lua_State *L [, lua_CFunction f])
    |  // Returns nresults.
    |  ldp BASE, TMP1, L->base
    |    str L, GL->cur_L
    |   sbfiz RC, CRET1, #3, #32
    |    st_vmstate ST_INTERP
    |  ldr PC, [BASE, FRAME_PC]
    |   sub RA, TMP1, RC		// RA = L->top - nresults*8
    |  b ->vm_returnc
    break;

  /* ---------------------------------------------------------------------- */

  default:
    fprintf(stderr, "Error: undefined opcode BC_%s\n", bc_names[op]);
    exit(2);
    break;
  }
}

static int build_backend(BuildCtx *ctx)
{
  int op;

  dasm_growpc(Dst, BC__MAX);

  build_subroutines(ctx);

  |.code_op
  for (op = 0; op < BC__MAX; op++)
    build_ins(ctx, (BCOp)op, op);

  return BC__MAX;
}

/* Emit pseudo frame-info for all assembler functions. */
static void emit_asm_debug(BuildCtx *ctx)
{
  int fcofs = (int)((uint8_t *)ctx->glob[GLOB_vm_ffi_call] - ctx->code);
  int i;
  switch (ctx->mode) {
  case BUILD_elfasm:
    fprintf(ctx->fp, "\t.section .debug_frame,\"\",%%progbits\n");
    fprintf(ctx->fp,
	".Lframe0:\n"
	"\t.long .LECIE0-.LSCIE0\n"
	".LSCIE0:\n"
	"\t.long 0xffffffff\n"
	"\t.byte 0x1\n"
	"\t.string \"\"\n"
	"\t.uleb128 0x1\n"
	"\t.sleb128 -8\n"
	"\t.byte 30\n"				/* Return address is in lr. */
	"\t.byte 0xc\n\t.uleb128 29\n\t.uleb128 16\n"	/* def_cfa fp 16 */
	"\t.align 3\n"
	".LECIE0:\n\n");
    fprintf(ctx->fp,
	".LSFDE0:\n"
	"\t.long .LEFDE0-.LASFDE0\n"
	".LASFDE0:\n"
	"\t.long .Lframe0\n"
	"\t.quad .Lbegin\n"
	"\t.quad %d\n"
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n",		/* offset fp */
	fcofs);
    for (i = 19; i <= 28; i++)  /* offset x19-x28 */
      fprintf(ctx->fp, "\t.byte 0x%x\n\t.uleb128 %d\n", 0x80+i, i+(3-19));
    for (i = 8; i <= 15; i++)  /* offset d8-d15 */
      fprintf(ctx->fp, "\t.byte 5\n\t.uleb128 0x%x\n\t.uleb128 %d\n",
	      64+i, i+(3+(28-19+1)-8));
    fprintf(ctx->fp,
	"\t.align 3\n"
	".LEFDE0:\n\n");
#if LJ_HASFFI
    fprintf(ctx->fp,
	".LSFDE1:\n"
	"\t.long .LEFDE1-.LASFDE1\n"
	".LASFDE1:\n"
	"\t.long .Lframe0\n"
	"\t.quad lj_vm_ffi_call\n"
	"\t.quad %d\n"
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n"		/* offset fp */
	"\t.byte 0x93\n\t.uleb128 3\n"		/* offset x19 */
	"\t.byte 0x94\n\t.uleb128 4\n"		/* offset x20 */
	"\t.align 3\n"
	".LEFDE1:\n\n", (int)ctx->codesz - fcofs);
#endif
#if !LJ_NO_UNWIND
    fprintf(ctx->fp, "\t.section .eh_frame,\"a\",%%progbits\n");
    fprintf(ctx->fp,
	".Lframe1:\n"
	"\t.long .LECIE1-.LSCIE1\n"
	".LSCIE1:\n"
	"\t.long 0\n"
	"\t.byte 0x1\n"
	"\t.string \"zPR\"\n"
	"\t.uleb128 0x1\n"
	"\t.sleb128 -8\n"
	"\t.byte 30\n"				/* Return address is in lr. */
	"\t.uleb128 6\n"			/* augmentation length */
	"\t.byte 0x1b\n"			/* pcrel|sdata4 */
	"\t.long lj_err_unwind_dwarf-.\n"
	"\t.byte 0x1b\n"			/* pcrel|sdata4 */
	"\t.byte 0xc\n\t.uleb128 29\n\t.uleb128 16\n"	/* def_cfa fp 16 */
	"\t.align 3\n"
	".LECIE1:\n\n");
    fprintf(ctx->fp,
	".LSFDE2:\n"
	"\t.long .LEFDE2-.LASFDE2\n"
	".LASFDE2:\n"
	"\t.long .LASFDE2-.Lframe1\n"
	"\t.long .Lbegin-.\n"
	"\t.long %d\n"
	"\t.uleb128 0\n"			/* augmentation length */
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n",		/* offset fp */
	fcofs);
    for (i = 19; i <= 28; i++)  /* offset x19-x28 */
      fprintf(ctx->fp, "\t.byte 0x%x\n\t.uleb128 %d\n", 0x80+i, i+(3-19));
    for (i = 8; i <= 15; i++)  /* offset d8-d15 */
      fprintf(ctx->fp, "\t.byte 5\n\t.uleb128 0x%x\n\t.uleb128 %d\n",
	      64+i, i+(3+(28-19+1)-8));
    fprintf(ctx->fp,
	"\t.align 3\n"
	".LEFDE2:\n\n");
#if LJ_HASFFI
    fprintf(ctx->fp,
	".Lframe2:\n"
	"\t.long .LECIE2-.LSCIE2\n"
	".LSCIE2:\n"
	"\t.long 0\n"
	"\t.byte 0x1\n"
	"\t.string \"zR\"\n"
	"\t.uleb128 0x1\n"
	"\t.sleb128 -8\n"
	"\t.byte 30\n"				/* Return address is in lr. */
	"\t.uleb128 1\n"			/* augmentation length */
	"\t.byte 0x1b\n"			/* pcrel|sdata4 */
	"\t.byte 0xc\n\t.uleb128 29\n\t.uleb128 16\n"	/* def_cfa fp 16 */
	"\t.align 3\n"
	".LECIE2:\n\n");
    fprintf(ctx->fp,
	".LSFDE3:\n"
	"\t.long .LEFDE3-.LASFDE3\n"
	".LASFDE3:\n"
	"\t.long .LASFDE3-.Lframe2\n"
	"\t.long lj_vm_ffi_call-.\n"
	"\t.long %d\n"
	"\t.uleb128 0\n"			/* augmentation length */
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n"		/* offset fp */
	"\t.byte 0x93\n\t.uleb128 3\n"		/* offset x19 */
	"\t.byte 0x94\n\t.uleb128 4\n"		/* offset x20 */
	"\t.align 3\n"
	".LEFDE3:\n\n", (int)ctx->codesz - fcofs);
#endif
#endif
    break;
#if !LJ_NO_UNWIND
  case BUILD_machasm: {
#if LJ_HASFFI
    int fcsize = 0;
#endif
    int j;
    fprintf(ctx->fp, "\t.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support\n");
    fprintf(ctx->fp,
	"EH_frame1:\n"
	"\t.set L$set$x,LECIEX-LSCIEX\n"
	"\t.long L$set$x\n"
	"LSCIEX:\n"
	"\t.long 0\n"
	"\t.byte 0x1\n"
	"\t.ascii \"zPR\\0\"\n"
	"\t.uleb128 0x1\n"
	"\t.sleb128 -8\n"
	"\t.byte 30\n"				/* Return address is in lr. */
	"\t.uleb128 6\n"			/* augmentation length */
	"\t.byte 0x9b\n"			/* indirect|pcrel|sdata4 */
	"\t.long _lj_err_unwind_dwarf@GOT-.\n"
	"\t.byte 0x1b\n"			/* pcrel|sdata4 */
	"\t.byte 0xc\n\t.uleb128 29\n\t.uleb128 16\n"	/* def_cfa fp 16 */
	"\t.align 3\n"
	"LECIEX:\n\n");
    for (j = 0; j < ctx->nsym; j++) {
      const char *name = ctx->sym[j].name;
      int32_t size = ctx->sym[j+1].ofs - ctx->sym[j].ofs;
      if (size == 0) continue;
#if LJ_HASFFI
      if (!strcmp(name, "_lj_vm_ffi_call")) { fcsize = size; continue; }
#endif
      fprintf(ctx->fp,
	"LSFDE%d:\n"
	"\t.set L$set$%d,LEFDE%d-LASFDE%d\n"
	"\t.long L$set$%d\n"
	"LASFDE%d:\n"
	"\t.long LASFDE%d-EH_frame1\n"
	"\t.long %s-.\n"
	"\t.long %d\n"
	"\t.uleb128 0\n"			/* augmentation length */
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n",		/* offset fp */
	j, j, j, j, j, j, j, name, size);
      for (i = 19; i <= 28; i++)  /* offset x19-x28 */
	fprintf(ctx->fp, "\t.byte 0x%x\n\t.uleb128 %d\n", 0x80+i, i+(3-19));
      for (i = 8; i <= 15; i++)  /* offset d8-d15 */
	fprintf(ctx->fp, "\t.byte 5\n\t.uleb128 0x%x\n\t.uleb128 %d\n",
		64+i, i+(3+(28-19+1)-8));
      fprintf(ctx->fp,
	"\t.align 3\n"
	"LEFDE%d:\n\n", j);
    }
#if LJ_HASFFI
    if (fcsize) {
      fprintf(ctx->fp,
	"EH_frame2:\n"
	"\t.set L$set$y,LECIEY-LSCIEY\n"
	"\t.long L$set$y\n"
	"LSCIEY:\n"
	"\t.long 0\n"
	"\t.byte 0x1\n"
	"\t.ascii \"zR\\0\"\n"
	"\t.uleb128 0x1\n"
	"\t.sleb128 -8\n"
	"\t.byte 30\n"				/* Return address is in lr. */
	"\t.uleb128 1\n"			/* augmentation length */
	"\t.byte 0x1b\n"			/* pcrel|sdata4 */
	"\t.byte 0xc\n\t.uleb128 29\n\t.uleb128 16\n"	/* def_cfa fp 16 */
	"\t.align 3\n"
	"LECIEY:\n\n");
      fprintf(ctx->fp,
	"LSFDEY:\n"
	"\t.set L$set$yy,LEFDEY-LASFDEY\n"
	"\t.long L$set$yy\n"
	"LASFDEY:\n"
	"\t.long LASFDEY-EH_frame2\n"
	"\t.long _lj_vm_ffi_call-.\n"
	"\t.long %d\n"
	"\t.uleb128 0\n"			/* augmentation length */
	"\t.byte 0x9e\n\t.uleb128 1\n"		/* offset lr */
	"\t.byte 0x9d\n\t.uleb128 2\n"		/* offset fp */
	"\t.byte 0x93\n\t.uleb128 3\n"		/* offset x19 */
	"\t.byte 0x94\n\t.uleb128 4\n"		/* offset x20 */
	"\t.align 3\n"
	"LEFDEY:\n\n", fcsize);
    }
#endif
    fprintf(ctx->fp, ".subsections_via_symbols\n");
    }
    break;
#endif
  default:
    break;
  }
}