// framesize is large enough that SP-framesize may
		// underflow, causing a direct comparison with the
		// stack guard to incorrectly succeed. We explicitly
		// guard against underflow.
		//
		//	// Try subtracting from SP and check for underflow.
		//	// If this underflows, it sets C to 0.
		//	SUB.S $(framesize-StackSmall), SP, R2
		//	// If C is 1 (unsigned >=), compare with guard.
		//	CMP.HS stackguard, R2

    static final Comparator<Entry<?>> INSTANCE = new DecreasingCount();

    @Override
    public int compare(Entry<?> entry1, Entry<?> entry2) {
      return entry2.getCount() - entry1.getCount(); // subtracting two nonnegative integers
    }
  }

  /**
   * An {@link AbstractMultiset} with additional default implementations, some of them linear-time

    static final Comparator<Entry<?>> INSTANCE = new DecreasingCount();

    @Override
    public int compare(Entry<?> entry1, Entry<?> entry2) {
      return entry2.getCount() - entry1.getCount(); // subtracting two nonnegative integers
    }
  }

  /**
   * An {@link AbstractMultiset} with additional default implementations, some of them linear-time

        break;
      case HALF_EVEN:
      case HALF_DOWN:
      case HALF_UP:
        long absRem = abs(rem);
        long cmpRemToHalfDivisor = absRem - (abs(q) - absRem);
        // subtracting two nonnegative longs can't overflow
        // cmpRemToHalfDivisor has the same sign as compare(abs(rem), abs(q) / 2).
        if (cmpRemToHalfDivisor == 0) { // exactly on the half mark

        break;
      case HALF_EVEN:
      case HALF_DOWN:
      case HALF_UP:
        long absRem = abs(rem);
        long cmpRemToHalfDivisor = absRem - (abs(q) - absRem);
        // subtracting two nonnegative longs can't overflow
        // cmpRemToHalfDivisor has the same sign as compare(abs(rem), abs(q) / 2).
        if (cmpRemToHalfDivisor == 0) { // exactly on the half mark

		// Because divRecursiveStep overwrote the top part of uu with
		// the remainder r̂, the full uu already contains the equivalent
		// of r̂·B + uₙ₋₂ from the “Refining Guesses” discussion.
		// Subtracting q̂·vₙ₋₂ from it will compute the full-length remainder.
		// If that subtraction underflows, q̂·v > u, which we fix up
		// by decrementing q̂ and adding v back, same as in long division.

			t.Errorf("Expected %v, actual %v", testValue, v2.String())
		}
	}
}

// TestQuantitySubZeroPreservesSuffix verifies that a suffix is preserved
// independent of the order of operations when subtracting a zero and non-zero val
func TestQuantitySubZeroPreservesSuffix(t *testing.T) {
	testValues := []string{"100m", "1Gi"}
	zero := MustParse("0")
	for _, testValue := range testValues {
		value := MustParse(testValue)

    // numerical stability, subtract each batch with their max element so that
    // the maximum input value is zero. It can be shown that softmax computed
    // after adding or subtracting all inputs in a batch using a common value
    // gives mathematically equivalent result.
    auto max_logits =
        rewriter.create<TF::MaxOp>(loc, logits, reduce_dim,

	return (x + to - 1) &^ (to - 1)
}

// splitImm24uScaled splits an immediate into a scaled 12 bit unsigned lo value
// and an unscaled shifted 12 bit unsigned hi value. These are typically used
// by adding or subtracting the hi value and using the lo value as the offset
// for a load or store.
func splitImm24uScaled(v int32, shift int) (int32, int32, error) {
	if v < 0 {

      // the gradient. Note that we save one `PadOp` and some unnecessary kernel
      // computations, compared to the `xla::AvgPoolGrad` implementation, by
      // subtracting the original exterior padding before `ReduceWindowOp`
      // instead of trimming the result of `ReduceWindowOp` (the final result is
      // the same because all strides are 1).
      low_padding[dim] =