IMUL3Q $0x13, AX, AX
	MULQ   24(BX)
	ADDQ   AX, R9
	ADCQ   DX, R8

	// r1 += 19×a4×b2
	MOVQ   32(CX), AX
	IMUL3Q $0x13, AX, AX
	MULQ   16(BX)
	ADDQ   AX, R9
	ADCQ   DX, R8

	// r2 = a0×b2
	MOVQ (CX), AX
	MULQ 16(BX)
	MOVQ AX, R11
	MOVQ DX, R10

	// r2 += a1×b1
	MOVQ 8(CX), AX
	MULQ 8(BX)
	ADDQ AX, R11
	ADCQ DX, R10

	// r2 += a2×b0
	MOVQ 16(CX), AX
	MULQ (BX)
	ADDQ AX, R11

	MOVQ (8*0)(x_ptr), t0

	MOVQ (8*1)(x_ptr), AX
	MULQ t0
	MOVQ AX, acc1
	MOVQ DX, acc2

	MOVQ (8*2)(x_ptr), AX
	MULQ t0
	ADDQ AX, acc2
	ADCQ $0, DX
	MOVQ DX, acc3

	MOVQ (8*3)(x_ptr), AX
	MULQ t0
	ADDQ AX, acc3
	ADCQ $0, DX
	MOVQ DX, acc4
	// y[2:] * y[1]
	MOVQ (8*1)(x_ptr), t0

	MOVQ (8*2)(x_ptr), AX
	MULQ t0
	ADDQ AX, acc3
	ADCQ $0, DX
	MOVQ DX, t1

	ADCQ $1, h2;      \
	LEAQ 16(msg), msg

#define POLY1305_MUL(h0, h1, h2, r0, r1, t0, t1, t2, t3) \
	MOVQ  r0, AX;                  \
	MULQ  h0;                      \
	MOVQ  AX, t0;                  \
	MOVQ  DX, t1;                  \
	MOVQ  r0, AX;                  \
	MULQ  h1;                      \
	ADDQ  AX, t1;                  \
	ADCQ  $0, DX;                  \
	MOVQ  r0, t2;                  \

	MOVQ (0*8)(R8)(BX*8), AX
	MULQ R9
	ADDQ CX, AX
	ADCQ $0, DX
	MOVQ AX, (0*8)(R10)(BX*8)
	MOVQ DX, CX
	MOVQ (1*8)(R8)(BX*8), AX
	MULQ R9
	ADDQ CX, AX
	ADCQ $0, DX
	MOVQ AX, (1*8)(R10)(BX*8)
	MOVQ DX, CX
	MOVQ (2*8)(R8)(BX*8), AX
	MULQ R9
	ADDQ CX, AX
	ADCQ $0, DX
	MOVQ AX, (2*8)(R10)(BX*8)
	MOVQ DX, CX
	MOVQ (3*8)(R8)(BX*8), AX
	MULQ R9
	ADDQ CX, AX
	ADCQ $0, DX

Regexps are always matched from the start of the instructions line.
This means, for example, that the "MULQ" regexp is equivalent to
"^MULQ" (^ representing the start of the line), and it will NOT match
the following assembly line:

  IMULQ	$99, AX

To force a match at any point of the line, ".*MULQ" should be used.

For the same reason, a negative regexp like -"memmove" is not enough

	SUBQ	DI, AX
	IMULQ	$100, AX

	// generated code for
	//	func f(x uint64) (uint64, uint64) { return x/1000000000, x%1000000000 }
	// adapted to reduce duplication
	MOVQ	AX, CX
	MOVQ	$1360296554856532783, AX
	MULQ	CX
	ADDQ	CX, DX
	RCRQ	$1, DX
	SHRQ	$29, DX
	MOVQ	DX, sec+0(FP)
	IMULQ	$1000000000, DX
	SUBQ	DX, CX
	MOVL	CX, nsec+8(FP)

		Load(aX, RAX)
	case 2:
		Comment(fmt.Sprintf("%s = 2×%s×%s", r, aX, bX))
		Load(aX, RAX)
		SHLQ(Imm(1), RAX)
	default:
		panic("unsupported i value")
	}
	MULQ(mustAddr(bX)) // RDX, RAX = RAX * bX
	MOVQ(RAX, r.lo)
	MOVQ(RDX, r.hi)
}

// addMul64 sets r to r + i * aX * bX.
func addMul64(r uint128, i uint64, aX, bX namedComponent) {
	switch i {
	case 1:

	// ppc64x:"SRAD"
	b := n2 / 64 // signed

	return a, b
}

// Check that constant divisions get turned into MULs
func ConstDivs(n1 uint, n2 int) (uint, int) {
	// amd64:"MOVQ\t[$]-1085102592571150095","MULQ",-"DIVQ"
	// 386:"MOVL\t[$]-252645135","MULL",-"DIVL"
	// arm64:`MOVD`,`UMULH`,-`DIV`
	// arm:`MOVW`,`MUL`,-`.*udiv`
	a := n1 / 17 // unsigned

	// amd64:"MOVQ\t[$]-1085102592571150095","IMULQ",-"IDIVQ"

//    bits.Mul*    //
// --------------- //

func Mul(x, y uint) (hi, lo uint) {
	// amd64:"MULQ"
	// arm64:"UMULH","MUL"
	// ppc64x:"MULHDU","MULLD"
	// s390x:"MLGR"
	// mips64: "MULVU"
	// riscv64:"MULHU","MUL"
	return bits.Mul(x, y)
}

func Mul64(x, y uint64) (hi, lo uint64) {
	// amd64:"MULQ"
	// arm64:"UMULH","MUL"
	// ppc64x:"MULHDU","MULLD"
	// s390x:"MLGR"
	// mips64: "MULVU"

		{name: "SUBLconst", argLength: 1, reg: gp11, asm: "SUBL", aux: "Int32", resultInArg0: true, clobberFlags: true},

		{name: "MULQ", argLength: 2, reg: gp21, asm: "IMULQ", commutative: true, resultInArg0: true, clobberFlags: true},
		{name: "MULL", argLength: 2, reg: gp21, asm: "IMULL", commutative: true, resultInArg0: true, clobberFlags: true},