LEAQ  256(inp), inp
	LEAQ  256(oup), oup
	SUBQ  $256, inl
	JMP   openSSEMainLoop

openSSEMainLoopDone:
	// Handle the various tail sizes efficiently
	TESTQ inl, inl
	JE    openSSEFinalize
	CMPQ  inl, $64
	JBE   openSSETail64
	CMPQ  inl, $128
	JBE   openSSETail128
	CMPQ  inl, $192
	JBE   openSSETail192
	JMP   openSSETail256

openSSEFinalize:

		// TESTx: compare (arg0 & arg1) to 0
		{name: "TESTQ", argLength: 2, reg: gp2flags, commutative: true, asm: "TESTQ", typ: "Flags"},
		{name: "TESTL", argLength: 2, reg: gp2flags, commutative: true, asm: "TESTL", typ: "Flags"},

  func.return %0 : tensor<1x80x128x!quant.uniform<i8:f32, 0.60070550441741943:-128>>
}

// -----

// CHECK-LABEL: testSqrt
func.func @testSqrt(tensor<? x f32>) -> tensor<? x f32> {
^bb0(%arg0: tensor<? x f32>):
  // CHECK: "tfl.sqrt"(%arg0)
  %0 = "tfl.sqrt"(%arg0): (tensor<? x f32>) -> tensor<? x f32>
  func.return %0 : tensor<? x f32>
}

	if x.pos >= len(x.last) {
		x.last = append(x.last, choice{x.r.Intn(max), 0, max})
	}
	c := &x.last[x.pos]
	x.pos++
	if c.max != max {
		panic("inconsistent use of exhaustive tester")
	}
	return (c.n + c.off) % max
}

func (x *exhaustive) Maybe() bool {
	return x.Choose(2) == 1
}

func GCFunc(args []Value) []Value {
	runtime.GC()
	return []Value{}
}