// instructions, based on their operands.
	ABEQZ
	ABGEZ
	ABGT
	ABGTU
	ABGTZ
	ABLE
	ABLEU
	ABLEZ
	ABLTZ
	ABNEZ
	AFABSD
	AFABSS
	AFNEGD
	AFNEGS
	AFNED
	AFNES
	AMOV
	AMOVB
	AMOVBU
	AMOVF
	AMOVD
	AMOVH
	AMOVHU
	AMOVW
	AMOVWU
	ANEG
	ANEGW
	ANOT
	ASEQZ
	ASNEZ

	// End marker
	ALAST
)

	CALL (AX)(PC*1)                 // ERROR "invalid instruction"
	CALL (AX)(SB*1)                 // ERROR "invalid instruction"
	CALL (AX)(FP*1)                 // ERROR "invalid instruction"
	// Forbid memory operands for MOV CR/DR. See #24981.
	MOVQ CR0, (AX)                  // ERROR "invalid instruction"
	MOVQ CR2, (AX)                  // ERROR "invalid instruction"
	MOVQ CR3, (AX)                  // ERROR "invalid instruction"

//     its least significant bit [Yes|No] inverted, e.g.
//       arg_cond_AllowALNV_Normal
//       arg_cond_NotAllowALNV_Invert
//
// - arg_immediate_OptLSL_amount_16_0_48:
//     An immediate for MOV[KNZ] instruction encoded in imm16[20:5] with an optional
//     left shift of 16 in the range [0, 48] encoded in hw[22, 21]
//

	CVTTSS2SI: true,
	DEC:       true,
	DIV:       true,
	FLDENV:    true,
	FRSTOR:    true,
	IDIV:      true,
	IMUL:      true,
	IN:        true,
	INC:       true,
	LEA:       true,
	MOV:       true,
	MOVNTI:    true,
	MUL:       true,
	NEG:       true,
	NOP:       true,
	NOT:       true,
	OR:        true,
	OUT:       true,
	POP:       true,
	POPA:      true,
	POPCNT:    true,

		// SYNC
		// MOV(B|W)	(Rarg0), Rout
		// SYNC
		{name: "LoweredAtomicLoad8", argLength: 2, reg: gpload, faultOnNilArg0: true},
		{name: "LoweredAtomicLoad32", argLength: 2, reg: gpload, faultOnNilArg0: true},

		// store arg1 to arg0. arg2=mem. returns memory.
		// SYNC
		// MOV(B|W)	Rarg1, (Rarg0)
		// SYNC

	case LIS:
		return "ADDIS $0," + args[1] + "," + args[0]
	// store instructions with index registers
	case STBX, STBUX, STHX, STHUX, STWX, STWUX, STDX, STDUX,
		STHBRX, STWBRX, STDBRX, STSWX, STFIWX:
		return "MOV" + op[2:len(op)-1] + " " + args[0] + ",(" + args[2] + ")(" + args[1] + ")"

	case STDCXCC, STWCXCC, STHCXCC, STBCXCC:
		return op + " " + args[0] + ",(" + args[2] + ")(" + args[1] + ")"

	AVBIT
	AVBSL
	AVCMEQ
	AVCMTST
	AVCNT
	AVDUP
	AVEOR
	AVEOR3
	AVEXT
	AVFMLA
	AVFMLS
	AVLD1
	AVLD1R
	AVLD2
	AVLD2R
	AVLD3
	AVLD3R
	AVLD4
	AVLD4R
	AVMOV
	AVMOVD
	AVMOVI
	AVMOVQ
	AVMOVS
	AVORR
	AVPMULL
	AVPMULL2
	AVRAX1
	AVRBIT
	AVREV16
	AVREV32
	AVREV64
	AVSHL
	AVSLI
	AVSRI
	AVST1
	AVST2
	AVST3
	AVST4

		thunkfunc.SetType(sym.STEXT)
		ldr.SetAttrLocal(thunkfunc.Sym(), true)
		o := func(op ...uint8) {
			for _, op1 := range op {
				thunkfunc.AddUint8(op1)
			}
		}
		// 8b 04 24	mov    (%esp),%eax
		// Destination register is in bits 3-5 of the middle byte, so add that in.
		o(0x8b, 0x04+r.num<<3, 0x24)
		// c3		ret
		o(0xc3)

		thunks = append(thunks, thunkfunc.Sym())
	}

		{name: "CMOVWCS", argLength: 3, reg: gp21, asm: "CMOVWCS", resultInArg0: true},

		// CMOV with floating point instructions. We need separate pseudo-op to handle
		// InvertFlags correctly, and to generate special code that handles NaN (unordered flag).
		// NOTE: the fact that CMOV*EQF here is marked to generate CMOV*NE is not a bug. See
		// code generation in amd64/ssa.go.

				writeableData[r.Off()-2] = 0x8d
				su.SetRelocType(rIdx, objabi.R_PCREL)
				su.SetRelocAdd(rIdx, r.Add()+4)
				return true
			}
		}

		// fall back to using GOT and hope for the best (CMOV*)
		// TODO: just needs relocation, no need to put in .dynsym
		ld.AddGotSym(target, ldr, syms, targ, uint32(elf.R_X86_64_GLOB_DAT))

		su.SetRelocType(rIdx, objabi.R_PCREL)
		su.SetRelocSym(rIdx, syms.GOT)