}
	return x // ERROR "converted OpPhi to OrB$"
}

//go:noinline
func f5and(a int, b bool) bool {
	var x bool
	if a == 0 {
		x = b
	} else {
		x = false
	}
	return x // ERROR "converted OpPhi to AndB$"
}

//go:noinline
func f6or(a int, b bool) bool {
	x := b
	if a == 0 {
		// f6or has side effects so the OpPhi should not be converted.
		x = f6or(a, b)
	}
	return x
}

//go:noinline

(MOVWZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVWZreg ((OR|XOR|AND) <t> x y))
(MOVHZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVHZreg ((OR|XOR|AND) <t> x y))
(MOVHZreg ((OR|XOR|AND) <t> x (MOVHZreg y))) => (MOVHZreg ((OR|XOR|AND) <t> x y))
(MOVBZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVBZreg ((OR|XOR|AND) <t> x y))
(MOVBZreg ((OR|XOR|AND) <t> x (MOVHZreg y))) => (MOVBZreg ((OR|XOR|AND) <t> x y))

	RORL	$2, AX; \
	MOVL	(a*4)(DI), DX; \
	ANDL	(b*4)(DI), BX; \
	RORL	$13, DX; \
	MOVL	(a*4)(DI), CX; \
	ANDL	(c*4)(DI), CX; \
	XORL	DX, AX; \
	XORL	CX, BX; \
	MOVL	(a*4)(DI), DX; \
	MOVL	(b*4)(DI), CX; \
	RORL	$22, DX; \
	ANDL	(a*4)(DI), CX; \
	XORL	CX, BX; \
	XORL	DX, AX; \
	ADDL	AX, BX

// Calculate T1 and T2, then e = d + T1 and a = T1 + T2.

	LEAL	const(a)(DI*1), a; \
	ANDL	b, BP; \
	XORL d, BP; \
	MOVL (index*4)(SI), DI; \
	ADDL BP, a; \
	ROLL $shift, a; \
	MOVL c, BP; \
	ADDL b, a

#define ROUND2(a, b, c, d, index, const, shift) \
	LEAL	const(a)(DI*1),a; \
	MOVL	d,		DI; \
	ANDL	b,		DI; \
	MOVL	d,		BP; \
	NOTL	BP; \
	ANDL	c,		BP; \
	ORL	DI,		BP; \
	MOVL	(index*4)(SI),DI; \
	ADDL	BP,		a; \
	ROLL	$shift,	a; \
	ADDL	b,		a

// Uses https://github.com/animetosho/md5-optimisation#dependency-shortcut-in-g-function

#define ROUND2(a, b, c, d, index, const, shift) \
	XORL	R11, R9; \
	ADDL	$const,	a; \
	ADDL	R8,	a; \
	ANDL	b,		R10; \
	ANDL	c,		R9; \
	MOVL	(index*4)(SI),R8; \
	ADDL	R9,	a; \
	ADDL	R10,	a; \
	MOVL	c,		R9; \
	MOVL	c,		R10; \
	ROLL	$shift,	a; \
	ADDL	b,		a

	ROUND2(AX,BX,CX,DX, 6,0xf61e2562, 5);

	// amd64: `SUBL\t8\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*4\), [A-Z]+[0-9]*`
	s -= x[i+2]
	// 386: `IMULL\t12\([A-Z]+\)\([A-Z]+\*4\), [A-Z]+`
	s *= x[i+3]
	// 386: `ANDL\t16\([A-Z]+\)\([A-Z]+\*4\), [A-Z]+`
	// amd64: `ANDL\t16\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*4\), [A-Z]+[0-9]*`
	s &= x[i+4]
	// 386: `ORL\t20\([A-Z]+\)\([A-Z]+\*4\), [A-Z]+`
	// amd64: `ORL\t20\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*4\), [A-Z]+[0-9]*`

func BenchmarkAnd8(b *testing.B) {
	var x [512]uint8 // give byte its own cache line
	sink = &x
	for i := 0; i < b.N; i++ {
		atomic.And8(&x[255], uint8(i))
	}
}

func BenchmarkAnd(b *testing.B) {
	var x [128]uint32 // give x its own cache line
	sink = &x
	for i := 0; i < b.N; i++ {
		atomic.And(&x[63], uint32(i))
	}
}

func BenchmarkAnd8Parallel(b *testing.B) {
	var x [512]uint8 // give byte its own cache line

//go:noescape
func LoadAcq64(ptr *uint64) uint64

//go:noescape
func LoadAcquintptr(ptr *uintptr) uintptr

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

// NOTE: Do not add atomicxor8 (XOR is not idempotent).

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape
func And32(ptr *uint32, val uint32) uint32

}

//go:nosplit
//go:noinline
func StoreReluintptr(ptr *uintptr, val uintptr) {
	*ptr = val
}

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

// NOTE: Do not add atomicxor8 (XOR is not idempotent).

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape

	POPCNT	R3, R4                // b9e10043

	AND	R1, R2                // b9800021
	AND	R1, R2, R3            // b9e42031
	AND	$-2, R1               // a517fffe
	AND	$-65536, R1           // c01bffff0000
	AND	$1, R1                // c0a100000001b980001a
	ANDW	R1, R2                // 1421
	ANDW	R1, R2, R3            // b9f42031
	ANDW	$1, R1                // c01b00000001
	ANDW	$131071, R1           // a5160001