// CHECK-DAG:  %[[CONSTANT1:.*]] = arith.constant dense<0.000000e+00> : tensor<4xf32>
//              variance + epsilon
// CHECK:  %[[ADD1:.*]] = "tf.Add"(%[[ARG4:.*]], %[[CONSTANT]])
//              rsqrt(variance + epsilon)
// CHECK:  %[[RSQRT:.*]] = "tf.Rsqrt"(%[[ADD1]])
//              scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL1:.*]] = "tf.Mul"(%[[ARG1:.*]], %[[RSQRT]])

	MOVD $0, R0                     // 38000000

	ADD $1, R3                      // 38630001
	ADD $1, R3, R4                  // 38830001
	ADD $-1, R4                     // 3884ffff
	ADD $-1, R4, R5                 // 38a4ffff
	ADD $65535, R5                  // 601fffff7cbf2a14 or 0600000038a5ffff
	ADD $65535, R5, R6              // 601fffff7cdf2a14 or 0600000038c5ffff
	ADD $65536, R6                  // 3cc60001

// floating point negative abs
(FNEG (F(ABS|NABS) x)) => (F(NABS|ABS) x)

// floating-point fused multiply-add/sub
(F(ADD|SUB) (FMUL x y) z) && x.Block.Func.useFMA(v) => (FM(ADD|SUB) x y z)
(F(ADDS|SUBS) (FMULS x y) z) && x.Block.Func.useFMA(v) => (FM(ADDS|SUBS) x y z)

// Arch-specific inlining for small or disjoint runtime.memmove

    %add:2 = tf_executor.island wraps "tf.Add"(%arg1, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %stack_push2:2 = tf_executor.island(%stack_push#1) wraps "tf.StackPushV2"(%stack_handle#0, %add#0) : (!tf_res, tensor<f32>) -> tensor<f32>
    %one:2 = tf_executor.island wraps "tf.Const"() { value = dense<1> : tensor<i32> } : () -> tensor<i32>

		n = min(n, maxObletBytes)
		tp = s.typePointersOfUnchecked(s.base())
		tp = tp.fastForward(b-tp.addr, b+n)
	} else {
		tp = s.typePointersOfUnchecked(b)
	}

	var scanSize uintptr
	for {
		var addr uintptr
		if tp, addr = tp.nextFast(); addr == 0 {
			if tp, addr = tp.next(b + n); addr == 0 {
				break
			}
		}

		// Keep track of farthest pointer we found, so we can

	ADD	R0->28, R1           // 401e81e0
	ADD	R0@>28, R1           // 601e81e0
	ADD.S	R0<<28, R1           // 001e91e0
	ADD.S	R0>>28, R1           // 201e91e0
	ADD.S	R0->28, R1           // 401e91e0
	ADD.S	R0@>28, R1           // 601e91e0
	ADD	R0<<R1, R2, R3       // 103182e0
	ADD	R0>>R1, R2, R3       // 303182e0
	ADD	R0->R1, R2, R3       // 503182e0
	ADD	R0@>R1, R2, R3       // 703182e0

		// in the span since some of them might be scavenged.
		sysUsed(unsafe.Pointer(base), nbytes, scav)
		gcController.heapReleased.add(-int64(scav))
	}
	// Update stats.
	gcController.heapFree.add(-int64(nbytes - scav))
	if typ == spanAllocHeap {
		gcController.heapInUse.add(int64(nbytes))
	}
	// Update consistent stats.
	stats := memstats.heapStats.acquire()
	atomic.Xaddint64(&stats.committed, int64(scav))