+APInt APInt::rotl(uint32_t rotateAmt) const {
+ if (rotateAmt == 0)
+ return *this;
+ // Don't get too fancy, just use existing shift/or facilities
+ APInt hi(*this);
+ APInt lo(*this);
+ hi.shl(rotateAmt);
+ lo.lshr(BitWidth - rotateAmt);
+ return hi | lo;
+}
+
+APInt APInt::rotr(uint32_t rotateAmt) const {
+ if (rotateAmt == 0)
+ return *this;
+ // Don't get too fancy, just use existing shift/or facilities
+ APInt hi(*this);
+ APInt lo(*this);
+ lo.lshr(rotateAmt);
+ hi.shl(BitWidth - rotateAmt);
+ return hi | lo;
+}
+
+// Square Root - this method computes and returns the square root of "this".
+// Three mechanisms are used for computation. For small values (<= 5 bits),
+// a table lookup is done. This gets some performance for common cases. For
+// values using less than 52 bits, the value is converted to double and then
+// the libc sqrt function is called. The result is rounded and then converted
+// back to a uint64_t which is then used to construct the result. Finally,
+// the Babylonian method for computing square roots is used.
+APInt APInt::sqrt() const {
+
+ // Determine the magnitude of the value.
+ uint32_t magnitude = getActiveBits();
+
+ // Use a fast table for some small values. This also gets rid of some
+ // rounding errors in libc sqrt for small values.
+ if (magnitude <= 5) {
+ static const uint8_t results[32] = {
+ /* 0 */ 0,
+ /* 1- 2 */ 1, 1,
+ /* 3- 6 */ 2, 2, 2, 2,
+ /* 7-12 */ 3, 3, 3, 3, 3, 3,
+ /* 13-20 */ 4, 4, 4, 4, 4, 4, 4, 4,
+ /* 21-30 */ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+ /* 31 */ 6
+ };
+ return APInt(BitWidth, results[ (isSingleWord() ? VAL : pVal[0]) ]);
+ }
+
+ // If the magnitude of the value fits in less than 52 bits (the precision of
+ // an IEEE double precision floating point value), then we can use the
+ // libc sqrt function which will probably use a hardware sqrt computation.
+ // This should be faster than the algorithm below.
+ if (magnitude < 52) {
+#ifdef _MSC_VER
+ // Amazingly, VC++ doesn't have round().
+ return APInt(BitWidth,
+ uint64_t(::sqrt(double(isSingleWord()?VAL:pVal[0]))) + 0.5);
+#else
+ return APInt(BitWidth,
+ uint64_t(::round(::sqrt(double(isSingleWord()?VAL:pVal[0])))));
+#endif
+ }
+
+ // Okay, all the short cuts are exhausted. We must compute it. The following
+ // is a classical Babylonian method for computing the square root. This code
+ // was adapted to APINt from a wikipedia article on such computations.
+ // See http://www.wikipedia.org/ and go to the page named
+ // Calculate_an_integer_square_root.
+ uint32_t nbits = BitWidth, i = 4;
+ APInt testy(BitWidth, 16);
+ APInt x_old(BitWidth, 1);
+ APInt x_new(BitWidth, 0);
+ APInt two(BitWidth, 2);
+
+ // Select a good starting value using binary logarithms.
+ for (;; i += 2, testy = testy.shl(2))
+ if (i >= nbits || this->ule(testy)) {
+ x_old = x_old.shl(i / 2);
+ break;
+ }
+
+ // Use the Babylonian method to arrive at the integer square root:
+ for (;;) {
+ x_new = (this->udiv(x_old) + x_old).udiv(two);
+ if (x_old.ule(x_new))
+ break;
+ x_old = x_new;
+ }
+
+ // Make sure we return the closest approximation
+ // NOTE: The rounding calculation below is correct. It will produce an
+ // off-by-one discrepancy with results from pari/gp. That discrepancy has been
+ // determined to be a rounding issue with pari/gp as it begins to use a
+ // floating point representation after 192 bits. There are no discrepancies
+ // between this algorithm and pari/gp for bit widths < 192 bits.
+ APInt square(x_old * x_old);
+ APInt nextSquare((x_old + 1) * (x_old +1));
+ if (this->ult(square))
+ return x_old;
+ else if (this->ule(nextSquare)) {
+ APInt midpoint((nextSquare - square).udiv(two));
+ APInt offset(*this - square);
+ if (offset.ult(midpoint))
+ return x_old;
+ else
+ return x_old + 1;
+ } else
+ assert(0 && "Error in APInt::sqrt computation");
+ return x_old + 1;
+}
+