//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Sheng Zhou and Reid Spencer and is distributed
-// under the // University of Illinois Open Source License. See LICENSE.TXT
-// for details.
+// This file was developed by Sheng Zhou and is distributed under the
+// University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#define DEBUG_TYPE "apint"
#include "llvm/ADT/APInt.h"
-#include "llvm/DerivedTypes.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
+#include <math.h>
+#include <limits>
#include <cstring>
#include <cstdlib>
-#ifndef NDEBUG
#include <iomanip>
-#endif
using namespace llvm;
+
+/// This enumeration just provides for internal constants used in this
+/// translation unit.
+enum {
+ MIN_INT_BITS = 1, ///< Minimum number of bits that can be specified
+ ///< Note that this must remain synchronized with IntegerType::MIN_INT_BITS
+ MAX_INT_BITS = (1<<23)-1 ///< Maximum number of bits that can be specified
+ ///< Note that this must remain synchronized with IntegerType::MAX_INT_BITS
+};
+
/// A utility function for allocating memory, checking for allocation failures,
/// and ensuring the contents are zeroed.
inline static uint64_t* getClearedMemory(uint32_t numWords) {
return result;
}
-APInt::APInt(uint32_t numBits, uint64_t val)
+APInt::APInt(uint32_t numBits, uint64_t val, bool isSigned)
: BitWidth(numBits), VAL(0) {
- assert(BitWidth >= IntegerType::MIN_INT_BITS && "bitwidth too small");
- assert(BitWidth <= IntegerType::MAX_INT_BITS && "bitwidth too large");
+ assert(BitWidth >= MIN_INT_BITS && "bitwidth too small");
+ assert(BitWidth <= MAX_INT_BITS && "bitwidth too large");
if (isSingleWord())
VAL = val;
else {
pVal = getClearedMemory(getNumWords());
pVal[0] = val;
+ if (isSigned && int64_t(val) < 0)
+ for (unsigned i = 1; i < getNumWords(); ++i)
+ pVal[i] = -1ULL;
}
clearUnusedBits();
}
-APInt::APInt(uint32_t numBits, uint32_t numWords, uint64_t bigVal[])
+APInt::APInt(uint32_t numBits, uint32_t numWords, const uint64_t bigVal[])
: BitWidth(numBits), VAL(0) {
- assert(BitWidth >= IntegerType::MIN_INT_BITS && "bitwidth too small");
- assert(BitWidth <= IntegerType::MAX_INT_BITS && "bitwidth too large");
+ assert(BitWidth >= MIN_INT_BITS && "bitwidth too small");
+ assert(BitWidth <= MAX_INT_BITS && "bitwidth too large");
assert(bigVal && "Null pointer detected!");
if (isSingleWord())
VAL = bigVal[0];
APInt::APInt(uint32_t numbits, const char StrStart[], uint32_t slen,
uint8_t radix)
: BitWidth(numbits), VAL(0) {
+ assert(BitWidth >= MIN_INT_BITS && "bitwidth too small");
+ assert(BitWidth <= MAX_INT_BITS && "bitwidth too large");
fromString(numbits, StrStart, slen, radix);
}
APInt::APInt(uint32_t numbits, const std::string& Val, uint8_t radix)
: BitWidth(numbits), VAL(0) {
+ assert(BitWidth >= MIN_INT_BITS && "bitwidth too small");
+ assert(BitWidth <= MAX_INT_BITS && "bitwidth too large");
assert(!Val.empty() && "String empty?");
fromString(numbits, Val.c_str(), Val.size(), radix);
}
APInt::APInt(const APInt& that)
: BitWidth(that.BitWidth), VAL(0) {
+ assert(BitWidth >= MIN_INT_BITS && "bitwidth too small");
+ assert(BitWidth <= MAX_INT_BITS && "bitwidth too large");
if (isSingleWord())
VAL = that.VAL;
else {
APInt::~APInt() {
if (!isSingleWord() && pVal)
- delete[] pVal;
+ delete [] pVal;
}
APInt& APInt::operator=(const APInt& RHS) {
- assert(BitWidth == RHS.BitWidth && "Bit widths must be the same");
- if (isSingleWord())
+ // Don't do anything for X = X
+ if (this == &RHS)
+ return *this;
+
+ // If the bitwidths are the same, we can avoid mucking with memory
+ if (BitWidth == RHS.getBitWidth()) {
+ if (isSingleWord())
+ VAL = RHS.VAL;
+ else
+ memcpy(pVal, RHS.pVal, getNumWords() * APINT_WORD_SIZE);
+ return *this;
+ }
+
+ if (isSingleWord())
+ if (RHS.isSingleWord())
+ VAL = RHS.VAL;
+ else {
+ VAL = 0;
+ pVal = getMemory(RHS.getNumWords());
+ memcpy(pVal, RHS.pVal, RHS.getNumWords() * APINT_WORD_SIZE);
+ }
+ else if (getNumWords() == RHS.getNumWords())
+ memcpy(pVal, RHS.pVal, RHS.getNumWords() * APINT_WORD_SIZE);
+ else if (RHS.isSingleWord()) {
+ delete [] pVal;
VAL = RHS.VAL;
- else
- memcpy(pVal, RHS.pVal, getNumWords() * APINT_WORD_SIZE);
- return *this;
+ } else {
+ delete [] pVal;
+ pVal = getMemory(RHS.getNumWords());
+ memcpy(pVal, RHS.pVal, RHS.getNumWords() * APINT_WORD_SIZE);
+ }
+ BitWidth = RHS.BitWidth;
+ return clearUnusedBits();
}
APInt& APInt::operator=(uint64_t RHS) {
pVal[0] = RHS;
memset(pVal+1, 0, (getNumWords() - 1) * APINT_WORD_SIZE);
}
- return *this;
+ return clearUnusedBits();
}
/// add_1 - This function adds a single "digit" integer, y, to the multiple
APInt APInt::operator^(const APInt& RHS) const {
assert(BitWidth == RHS.BitWidth && "Bit widths must be the same");
if (isSingleWord())
- return APInt(BitWidth, VAL ^ RHS.VAL).clearUnusedBits();
+ return APInt(BitWidth, VAL ^ RHS.VAL);
uint32_t numWords = getNumWords();
uint64_t *val = getMemory(numWords);
APInt APInt::operator*(const APInt& RHS) const {
assert(BitWidth == RHS.BitWidth && "Bit widths must be the same");
if (isSingleWord())
- return APInt(BitWidth, VAL * RHS.VAL).clearUnusedBits();
+ return APInt(BitWidth, VAL * RHS.VAL);
APInt Result(*this);
Result *= RHS;
return Result.clearUnusedBits();
APInt APInt::operator+(const APInt& RHS) const {
assert(BitWidth == RHS.BitWidth && "Bit widths must be the same");
if (isSingleWord())
- return APInt(BitWidth, VAL + RHS.VAL).clearUnusedBits();
+ return APInt(BitWidth, VAL + RHS.VAL);
APInt Result(BitWidth, 0);
add(Result.pVal, this->pVal, RHS.pVal, getNumWords());
return Result.clearUnusedBits();
APInt APInt::operator-(const APInt& RHS) const {
assert(BitWidth == RHS.BitWidth && "Bit widths must be the same");
if (isSingleWord())
- return APInt(BitWidth, VAL - RHS.VAL).clearUnusedBits();
+ return APInt(BitWidth, VAL - RHS.VAL);
APInt Result(BitWidth, 0);
sub(Result.pVal, this->pVal, RHS.pVal, getNumWords());
return Result.clearUnusedBits();
}
bool APInt::operator==(const APInt& RHS) const {
+ assert(BitWidth == RHS.BitWidth && "Comparison requires equal bit widths");
if (isSingleWord())
return VAL == RHS.VAL;
return pVal[0] < RHS.pVal[0];
// Otherwise, compare all words
- for (int i = whichWord(n1 - 1); i >= 0; --i) {
+ uint32_t topWord = whichWord(std::max(n1,n2)-1);
+ for (int i = topWord; i >= 0; --i) {
if (pVal[i] > RHS.pVal[i])
return false;
if (pVal[i] < RHS.pVal[i])
}
APInt lhs(*this);
- APInt rhs(*this);
- bool lhsNegative = false;
- bool rhsNegative = false;
- if (lhs[BitWidth-1]) {
- // Sign bit is set so make a note of it and perform two's complement
- lhsNegative = true;
+ APInt rhs(RHS);
+ bool lhsNeg = isNegative();
+ bool rhsNeg = rhs.isNegative();
+ if (lhsNeg) {
+ // Sign bit is set so perform two's complement to make it positive
lhs.flip();
lhs++;
}
- if (rhs[BitWidth-1]) {
- // Sign bit is set so make a note of it and perform two's complement
- rhsNegative = true;
+ if (rhsNeg) {
+ // Sign bit is set so perform two's complement to make it positive
rhs.flip();
rhs++;
}
// Now we have unsigned values to compare so do the comparison if necessary
// based on the negativeness of the values.
- if (lhsNegative)
- if (rhsNegative)
- return !lhs.ult(rhs);
+ if (lhsNeg)
+ if (rhsNeg)
+ return lhs.ugt(rhs);
else
return true;
- else if (rhsNegative)
+ else if (rhsNeg)
return false;
else
return lhs.ult(rhs);
}
// Set all the bits in all the words.
- for (uint32_t i = 0; i < getNumWords() - 1; ++i)
+ for (uint32_t i = 0; i < getNumWords(); ++i)
pVal[i] = -1ULL;
// Clear the unused ones
return clearUnusedBits();
/// @brief Bitwise NOT operator. Performs a bitwise logical NOT operation on
/// this APInt.
APInt APInt::operator~() const {
- APInt API(*this);
- API.flip();
- return API;
+ APInt Result(*this);
+ Result.flip();
+ return Result;
}
/// @brief Toggle every bit to its opposite value.
APInt& APInt::flip() {
if (isSingleWord()) {
- VAL = ~VAL;
+ VAL ^= -1ULL;
return clearUnusedBits();
}
for (uint32_t i = 0; i < getNumWords(); ++i)
- pVal[i] = ~pVal[i];
+ pVal[i] ^= -1ULL;
return clearUnusedBits();
}
return *this;
}
-/// getMaxValue - This function returns the largest value
-/// for an APInt of the specified bit-width and if isSign == true,
-/// it should be largest signed value, otherwise unsigned value.
-APInt APInt::getMaxValue(uint32_t numBits, bool isSign) {
- APInt Result(numBits, 0);
- Result.set();
- if (isSign)
- Result.clear(numBits - 1);
- return Result;
-}
+uint32_t APInt::getBitsNeeded(const char* str, uint32_t slen, uint8_t radix) {
+ assert(str != 0 && "Invalid value string");
+ assert(slen > 0 && "Invalid string length");
-/// getMinValue - This function returns the smallest value for
-/// an APInt of the given bit-width and if isSign == true,
-/// it should be smallest signed value, otherwise zero.
-APInt APInt::getMinValue(uint32_t numBits, bool isSign) {
- APInt Result(numBits, 0);
- if (isSign)
- Result.set(numBits - 1);
- return Result;
-}
+ // Each computation below needs to know if its negative
+ uint32_t isNegative = str[0] == '-';
+ if (isNegative) {
+ slen--;
+ str++;
+ }
+ // For radixes of power-of-two values, the bits required is accurately and
+ // easily computed
+ if (radix == 2)
+ return slen + isNegative;
+ if (radix == 8)
+ return slen * 3 + isNegative;
+ if (radix == 16)
+ return slen * 4 + isNegative;
+
+ // Otherwise it must be radix == 10, the hard case
+ assert(radix == 10 && "Invalid radix");
+
+ // This is grossly inefficient but accurate. We could probably do something
+ // with a computation of roughly slen*64/20 and then adjust by the value of
+ // the first few digits. But, I'm not sure how accurate that could be.
-/// getAllOnesValue - This function returns an all-ones value for
-/// an APInt of the specified bit-width.
-APInt APInt::getAllOnesValue(uint32_t numBits) {
- return getMaxValue(numBits, false);
+ // Compute a sufficient number of bits that is always large enough but might
+ // be too large. This avoids the assertion in the constructor.
+ uint32_t sufficient = slen*64/18;
+
+ // Convert to the actual binary value.
+ APInt tmp(sufficient, str, slen, radix);
+
+ // Compute how many bits are required.
+ return isNegative + tmp.logBase2() + 1;
}
-/// getNullValue - This function creates an '0' value for an
-/// APInt of the specified bit-width.
-APInt APInt::getNullValue(uint32_t numBits) {
- return getMinValue(numBits, false);
+uint64_t APInt::getHashValue() const {
+ // Put the bit width into the low order bits.
+ uint64_t hash = BitWidth;
+
+ // Add the sum of the words to the hash.
+ if (isSingleWord())
+ hash += VAL << 6; // clear separation of up to 64 bits
+ else
+ for (uint32_t i = 0; i < getNumWords(); ++i)
+ hash += pVal[i] << 6; // clear sepration of up to 64 bits
+ return hash;
}
/// HiBits - This function returns the high "numBits" bits of this APInt.
uint32_t remainder = BitWidth % APINT_BITS_PER_WORD;
if (remainder)
Count -= APINT_BITS_PER_WORD - remainder;
+ return std::min(Count, BitWidth);
+}
+
+static uint32_t countLeadingOnes_64(uint64_t V, uint32_t skip) {
+ uint32_t Count = 0;
+ if (skip)
+ V <<= skip;
+ while (V && (V & (1ULL << 63))) {
+ Count++;
+ V <<= 1;
+ }
+ return Count;
+}
+
+uint32_t APInt::countLeadingOnes() const {
+ if (isSingleWord())
+ return countLeadingOnes_64(VAL, APINT_BITS_PER_WORD - BitWidth);
+
+ uint32_t highWordBits = BitWidth % APINT_BITS_PER_WORD;
+ uint32_t shift = (highWordBits == 0 ? 0 : APINT_BITS_PER_WORD - highWordBits);
+ int i = getNumWords() - 1;
+ uint32_t Count = countLeadingOnes_64(pVal[i], shift);
+ if (Count == highWordBits) {
+ for (i--; i >= 0; --i) {
+ if (pVal[i] == -1ULL)
+ Count += APINT_BITS_PER_WORD;
+ else {
+ Count += countLeadingOnes_64(pVal[i], 0);
+ break;
+ }
+ }
+ }
return Count;
}
uint32_t APInt::countTrailingZeros() const {
if (isSingleWord())
- return CountTrailingZeros_64(VAL);
- APInt Tmp( ~(*this) & ((*this) - APInt(BitWidth,1)) );
- return getNumWords() * APINT_BITS_PER_WORD - Tmp.countLeadingZeros();
+ return std::min(CountTrailingZeros_64(VAL), BitWidth);
+ uint32_t Count = 0;
+ uint32_t i = 0;
+ for (; i < getNumWords() && pVal[i] == 0; ++i)
+ Count += APINT_BITS_PER_WORD;
+ if (i < getNumWords())
+ Count += CountTrailingZeros_64(pVal[i]);
+ return std::min(Count, BitWidth);
}
uint32_t APInt::countPopulation() const {
APInt APInt::byteSwap() const {
assert(BitWidth >= 16 && BitWidth % 16 == 0 && "Cannot byteswap!");
if (BitWidth == 16)
- return APInt(BitWidth, ByteSwap_16(VAL));
+ return APInt(BitWidth, ByteSwap_16(uint16_t(VAL)));
else if (BitWidth == 32)
- return APInt(BitWidth, ByteSwap_32(VAL));
+ return APInt(BitWidth, ByteSwap_32(uint32_t(VAL)));
else if (BitWidth == 48) {
- uint64_t Tmp1 = ((VAL >> 32) << 16) | (VAL & 0xFFFF);
+ uint32_t Tmp1 = uint32_t(VAL >> 16);
Tmp1 = ByteSwap_32(Tmp1);
- uint64_t Tmp2 = (VAL >> 16) & 0xFFFF;
+ uint16_t Tmp2 = uint16_t(VAL);
Tmp2 = ByteSwap_16(Tmp2);
- return
- APInt(BitWidth,
- (Tmp1 & 0xff) | ((Tmp1<<16) & 0xffff00000000ULL) | (Tmp2 << 16));
+ return APInt(BitWidth, (uint64_t(Tmp2) << 32) | Tmp1);
} else if (BitWidth == 64)
return APInt(BitWidth, ByteSwap_64(VAL));
else {
return A;
}
-APInt llvm::APIntOps::RoundDoubleToAPInt(double Double) {
+APInt llvm::APIntOps::RoundDoubleToAPInt(double Double, uint32_t width) {
union {
double D;
uint64_t I;
} T;
T.D = Double;
+
+ // Get the sign bit from the highest order bit
bool isNeg = T.I >> 63;
+
+ // Get the 11-bit exponent and adjust for the 1023 bit bias
int64_t exp = ((T.I >> 52) & 0x7ff) - 1023;
+
+ // If the exponent is negative, the value is < 0 so just return 0.
if (exp < 0)
- return APInt(64ull, 0u);
- uint64_t mantissa = ((T.I << 12) >> 12) | (1ULL << 52);
+ return APInt(width, 0u);
+
+ // Extract the mantissa by clearing the top 12 bits (sign + exponent).
+ uint64_t mantissa = (T.I & (~0ULL >> 12)) | 1ULL << 52;
+
+ // If the exponent doesn't shift all bits out of the mantissa
if (exp < 52)
- return isNeg ? -APInt(64u, mantissa >> (52 - exp)) :
- APInt(64u, mantissa >> (52 - exp));
- APInt Tmp(exp + 1, mantissa);
+ return isNeg ? -APInt(width, mantissa >> (52 - exp)) :
+ APInt(width, mantissa >> (52 - exp));
+
+ // If the client didn't provide enough bits for us to shift the mantissa into
+ // then the result is undefined, just return 0
+ if (width <= exp - 52)
+ return APInt(width, 0);
+
+ // Otherwise, we have to shift the mantissa bits up to the right location
+ APInt Tmp(width, mantissa);
Tmp = Tmp.shl(exp - 52);
return isNeg ? -Tmp : Tmp;
}
// Return infinity for exponent overflow
if (exp > 1023) {
if (!isSigned || !isNeg)
- return double(1.0E300 * 1.0E300); // positive infinity
+ return std::numeric_limits<double>::infinity();
else
- return double(-1.0E300 * 1.0E300); // negative infinity
+ return -std::numeric_limits<double>::infinity();
}
exp += 1023; // Increment for 1023 bias
}
// Truncate to new width.
-void APInt::trunc(uint32_t width) {
+APInt &APInt::trunc(uint32_t width) {
assert(width < BitWidth && "Invalid APInt Truncate request");
- assert(width >= IntegerType::MIN_INT_BITS && "Can't truncate to 0 bits");
+ assert(width >= MIN_INT_BITS && "Can't truncate to 0 bits");
uint32_t wordsBefore = getNumWords();
BitWidth = width;
uint32_t wordsAfter = getNumWords();
if (wordsAfter == 1) {
uint64_t *tmp = pVal;
VAL = pVal[0];
- delete tmp;
+ delete [] tmp;
} else {
uint64_t *newVal = getClearedMemory(wordsAfter);
for (uint32_t i = 0; i < wordsAfter; ++i)
newVal[i] = pVal[i];
- delete pVal;
+ delete [] pVal;
pVal = newVal;
}
}
- clearUnusedBits();
+ return clearUnusedBits();
}
// Sign extend to a new width.
-void APInt::sext(uint32_t width) {
+APInt &APInt::sext(uint32_t width) {
assert(width > BitWidth && "Invalid APInt SignExtend request");
- assert(width <= IntegerType::MAX_INT_BITS && "Too many bits");
- bool isNegative = (*this)[BitWidth-1];
+ assert(width <= MAX_INT_BITS && "Too many bits");
// If the sign bit isn't set, this is the same as zext.
- if (!isNegative) {
+ if (!isNegative()) {
zext(width);
- return;
+ return *this;
}
// The sign bit is set. First, get some facts
if (wordsBefore == wordsAfter) {
uint32_t newWordBits = width % APINT_BITS_PER_WORD;
// The extension is contained to the wordsBefore-1th word.
- uint64_t mask = (~0ULL >> (APINT_BITS_PER_WORD - newWordBits)) << wordBits;
+ uint64_t mask = ~0ULL;
+ if (newWordBits)
+ mask >>= APINT_BITS_PER_WORD - newWordBits;
+ mask <<= wordBits;
if (wordsBefore == 1)
VAL |= mask;
else
pVal[wordsBefore-1] |= mask;
- clearUnusedBits();
- return;
+ return clearUnusedBits();
}
- uint64_t mask = ~0ULL << wordBits;
+ uint64_t mask = wordBits == 0 ? 0 : ~0ULL << wordBits;
uint64_t *newVal = getMemory(wordsAfter);
if (wordsBefore == 1)
newVal[0] = VAL | mask;
for (uint32_t i = wordsBefore; i < wordsAfter; i++)
newVal[i] = -1ULL;
if (wordsBefore != 1)
- delete pVal;
+ delete [] pVal;
pVal = newVal;
- clearUnusedBits();
+ return clearUnusedBits();
}
// Zero extend to a new width.
-void APInt::zext(uint32_t width) {
+APInt &APInt::zext(uint32_t width) {
assert(width > BitWidth && "Invalid APInt ZeroExtend request");
- assert(width <= IntegerType::MAX_INT_BITS && "Too many bits");
+ assert(width <= MAX_INT_BITS && "Too many bits");
uint32_t wordsBefore = getNumWords();
BitWidth = width;
uint32_t wordsAfter = getNumWords();
for (uint32_t i = 0; i < wordsBefore; ++i)
newVal[i] = pVal[i];
if (wordsBefore != 1)
- delete pVal;
+ delete [] pVal;
pVal = newVal;
}
+ return *this;
+}
+
+APInt &APInt::zextOrTrunc(uint32_t width) {
+ if (BitWidth < width)
+ return zext(width);
+ if (BitWidth > width)
+ return trunc(width);
+ return *this;
+}
+
+APInt &APInt::sextOrTrunc(uint32_t width) {
+ if (BitWidth < width)
+ return sext(width);
+ if (BitWidth > width)
+ return trunc(width);
+ return *this;
}
/// Arithmetic right-shift this APInt by shiftAmt.
/// @brief Arithmetic right-shift function.
APInt APInt::ashr(uint32_t shiftAmt) const {
+ assert(shiftAmt <= BitWidth && "Invalid shift amount");
+ // Handle a degenerate case
+ if (shiftAmt == 0)
+ return *this;
+
+ // Handle single word shifts with built-in ashr
if (isSingleWord()) {
if (shiftAmt == BitWidth)
+ return APInt(BitWidth, 0); // undefined
+ else {
+ uint32_t SignBit = APINT_BITS_PER_WORD - BitWidth;
+ return APInt(BitWidth,
+ (((int64_t(VAL) << SignBit) >> SignBit) >> shiftAmt));
+ }
+ }
+
+ // If all the bits were shifted out, the result is, technically, undefined.
+ // We return -1 if it was negative, 0 otherwise. We check this early to avoid
+ // issues in the algorithm below.
+ if (shiftAmt == BitWidth) {
+ if (isNegative())
return APInt(BitWidth, -1ULL);
else
- return APInt(BitWidth,
- (((int64_t(VAL) << (APINT_BITS_PER_WORD - BitWidth)) >>
- (APINT_BITS_PER_WORD - BitWidth)) >> shiftAmt)).clearUnusedBits();
+ return APInt(BitWidth, 0);
}
- APInt Result(*this);
- if (shiftAmt >= BitWidth) {
- memset(Result.pVal, Result[BitWidth-1] ? 1 : 0,
- (getNumWords()-1) * APINT_WORD_SIZE);
- return Result.clearUnusedBits();
- }
+ // Create some space for the result.
+ uint64_t * val = new uint64_t[getNumWords()];
- // FIXME: bit-at-a-time shift is really slow.
- uint32_t i = 0;
- for (; i < BitWidth - shiftAmt; ++i)
- if (Result[i+shiftAmt])
- Result.set(i);
- else
- Result.clear(i);
- for (; i < BitWidth; ++i)
- if (Result[BitWidth-1])
- Result.set(i);
- else
- Result.clear(i);
- return Result;
+ // Compute some values needed by the following shift algorithms
+ uint32_t wordShift = shiftAmt % APINT_BITS_PER_WORD; // bits to shift per word
+ uint32_t offset = shiftAmt / APINT_BITS_PER_WORD; // word offset for shift
+ uint32_t breakWord = getNumWords() - 1 - offset; // last word affected
+ uint32_t bitsInWord = whichBit(BitWidth); // how many bits in last word?
+ if (bitsInWord == 0)
+ bitsInWord = APINT_BITS_PER_WORD;
+
+ // If we are shifting whole words, just move whole words
+ if (wordShift == 0) {
+ // Move the words containing significant bits
+ for (uint32_t i = 0; i <= breakWord; ++i)
+ val[i] = pVal[i+offset]; // move whole word
+
+ // Adjust the top significant word for sign bit fill, if negative
+ if (isNegative())
+ if (bitsInWord < APINT_BITS_PER_WORD)
+ val[breakWord] |= ~0ULL << bitsInWord; // set high bits
+ } else {
+ // Shift the low order words
+ for (uint32_t i = 0; i < breakWord; ++i) {
+ // This combines the shifted corresponding word with the low bits from
+ // the next word (shifted into this word's high bits).
+ val[i] = (pVal[i+offset] >> wordShift) |
+ (pVal[i+offset+1] << (APINT_BITS_PER_WORD - wordShift));
+ }
+
+ // Shift the break word. In this case there are no bits from the next word
+ // to include in this word.
+ val[breakWord] = pVal[breakWord+offset] >> wordShift;
+
+ // Deal with sign extenstion in the break word, and possibly the word before
+ // it.
+ if (isNegative()) {
+ if (wordShift > bitsInWord) {
+ if (breakWord > 0)
+ val[breakWord-1] |=
+ ~0ULL << (APINT_BITS_PER_WORD - (wordShift - bitsInWord));
+ val[breakWord] |= ~0ULL;
+ } else
+ val[breakWord] |= (~0ULL << (bitsInWord - wordShift));
+ }
+ }
+
+ // Remaining words are 0 or -1, just assign them.
+ uint64_t fillValue = (isNegative() ? -1ULL : 0);
+ for (uint32_t i = breakWord+1; i < getNumWords(); ++i)
+ val[i] = fillValue;
+ return APInt(val, BitWidth).clearUnusedBits();
}
/// Logical right-shift this APInt by shiftAmt.
/// @brief Logical right-shift function.
APInt APInt::lshr(uint32_t shiftAmt) const {
- if (isSingleWord())
+ if (isSingleWord()) {
if (shiftAmt == BitWidth)
return APInt(BitWidth, 0);
else
return APInt(BitWidth, this->VAL >> shiftAmt);
+ }
- APInt Result(*this);
- if (shiftAmt >= BitWidth) {
- Result.clear();
- return Result;
+ // If all the bits were shifted out, the result is 0. This avoids issues
+ // with shifting by the size of the integer type, which produces undefined
+ // results. We define these "undefined results" to always be 0.
+ if (shiftAmt == BitWidth)
+ return APInt(BitWidth, 0);
+
+ // If none of the bits are shifted out, the result is *this. This avoids
+ // issues with shifting byt he size of the integer type, which produces
+ // undefined results in the code below. This is also an optimization.
+ if (shiftAmt == 0)
+ return *this;
+
+ // Create some space for the result.
+ uint64_t * val = new uint64_t[getNumWords()];
+
+ // If we are shifting less than a word, compute the shift with a simple carry
+ if (shiftAmt < APINT_BITS_PER_WORD) {
+ uint64_t carry = 0;
+ for (int i = getNumWords()-1; i >= 0; --i) {
+ val[i] = (pVal[i] >> shiftAmt) | carry;
+ carry = pVal[i] << (APINT_BITS_PER_WORD - shiftAmt);
+ }
+ return APInt(val, BitWidth).clearUnusedBits();
}
- // FIXME: bit at a time shift is really slow
- uint32_t i = 0;
- for (i = 0; i < Result.BitWidth - shiftAmt; ++i)
- if (Result[i+shiftAmt])
- Result.set(i);
- else
- Result.clear(i);
- for (; i < Result.BitWidth; ++i)
- Result.clear(i);
- return Result;
+ // Compute some values needed by the remaining shift algorithms
+ uint32_t wordShift = shiftAmt % APINT_BITS_PER_WORD;
+ uint32_t offset = shiftAmt / APINT_BITS_PER_WORD;
+
+ // If we are shifting whole words, just move whole words
+ if (wordShift == 0) {
+ for (uint32_t i = 0; i < getNumWords() - offset; ++i)
+ val[i] = pVal[i+offset];
+ for (uint32_t i = getNumWords()-offset; i < getNumWords(); i++)
+ val[i] = 0;
+ return APInt(val,BitWidth).clearUnusedBits();
+ }
+
+ // Shift the low order words
+ uint32_t breakWord = getNumWords() - offset -1;
+ for (uint32_t i = 0; i < breakWord; ++i)
+ val[i] = (pVal[i+offset] >> wordShift) |
+ (pVal[i+offset+1] << (APINT_BITS_PER_WORD - wordShift));
+ // Shift the break word.
+ val[breakWord] = pVal[breakWord+offset] >> wordShift;
+
+ // Remaining words are 0
+ for (uint32_t i = breakWord+1; i < getNumWords(); ++i)
+ val[i] = 0;
+ return APInt(val, BitWidth).clearUnusedBits();
}
/// Left-shift this APInt by shiftAmt.
if (isSingleWord()) {
if (shiftAmt == BitWidth)
return APInt(BitWidth, 0); // avoid undefined shift results
- return APInt(BitWidth, VAL << shiftAmt).clearUnusedBits();
+ return APInt(BitWidth, VAL << shiftAmt);
}
// If all the bits were shifted out, the result is 0. This avoids issues
if (shiftAmt == BitWidth)
return APInt(BitWidth, 0);
+ // If none of the bits are shifted out, the result is *this. This avoids a
+ // lshr by the words size in the loop below which can produce incorrect
+ // results. It also avoids the expensive computation below for a common case.
+ if (shiftAmt == 0)
+ return *this;
+
// Create some space for the result.
uint64_t * val = new uint64_t[getNumWords()];
// If we are shifting less than a word, do it the easy way
if (shiftAmt < APINT_BITS_PER_WORD) {
uint64_t carry = 0;
- shiftAmt %= APINT_BITS_PER_WORD;
for (uint32_t i = 0; i < getNumWords(); i++) {
val[i] = pVal[i] << shiftAmt | carry;
carry = pVal[i] >> (APINT_BITS_PER_WORD - shiftAmt);
return APInt(val, BitWidth).clearUnusedBits();
}
+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;
+}
+
/// Implementation of Knuth's Algorithm D (Division of nonnegative integers)
/// from "Art of Computer Programming, Volume 2", section 4.3.1, p. 272. The
/// variables here have the same names as in the algorithm. Comments explain
// (u[j+n]u[j+n-1]..u[j]) - qp * (v[n-1]...v[1]v[0]). This computation
// consists of a simple multiplication by a one-place number, combined with
// a subtraction.
- bool isNegative = false;
+ bool isNeg = false;
for (uint32_t i = 0; i < n; ++i) {
uint64_t u_tmp = uint64_t(u[j+i]) | (uint64_t(u[j+i+1]) << 32);
uint64_t subtrahend = uint64_t(qp) * uint64_t(v[i]);
u[k]--;
k++;
}
- isNegative |= borrow;
+ isNeg |= borrow;
DEBUG(cerr << "KnuthDiv: u[j+i] == " << u[j+i] << ", u[j+i+1] == " <<
u[j+i+1] << '\n');
}
// true value plus b**(n+1), namely as the b's complement of
// the true value, and a "borrow" to the left should be remembered.
//
- if (isNegative) {
+ if (isNeg) {
bool carry = true; // true because b's complement is "complement + 1"
for (uint32_t i = 0; i <= m+n; ++i) {
u[i] = ~u[i] + carry; // b's complement
// D5. [Test remainder.] Set q[j] = qp. If the result of step D4 was
// negative, go to step D6; otherwise go on to step D7.
q[j] = qp;
- if (isNegative) {
+ if (isNeg) {
// D6. [Add back]. The probability that this step is necessary is very
// small, on the order of only 2/b. Make sure that test data accounts for
// this possibility. Decrease q[j] by 1
if (Quotient->isSingleWord())
Quotient->VAL = 0;
else
- delete Quotient->pVal;
+ delete [] Quotient->pVal;
Quotient->BitWidth = LHS.BitWidth;
if (!Quotient->isSingleWord())
Quotient->pVal = getClearedMemory(Quotient->getNumWords());
if (Remainder->isSingleWord())
Remainder->VAL = 0;
else
- delete Remainder->pVal;
+ delete [] Remainder->pVal;
Remainder->BitWidth = RHS.BitWidth;
if (!Remainder->isSingleWord())
Remainder->pVal = getClearedMemory(Remainder->getNumWords());
return APInt(BitWidth, pVal[0] % RHS.pVal[0]);
}
- // We have to compute it the hard way. Invoke the Knute divide algorithm.
+ // We have to compute it the hard way. Invoke the Knuth divide algorithm.
APInt Remainder(1,0);
divide(*this, lhsWords, RHS, rhsWords, 0, &Remainder);
return Remainder;
}
+void APInt::udivrem(const APInt &LHS, const APInt &RHS,
+ APInt &Quotient, APInt &Remainder) {
+ // Get some size facts about the dividend and divisor
+ uint32_t lhsBits = LHS.getActiveBits();
+ uint32_t lhsWords = !lhsBits ? 0 : (APInt::whichWord(lhsBits - 1) + 1);
+ uint32_t rhsBits = RHS.getActiveBits();
+ uint32_t rhsWords = !rhsBits ? 0 : (APInt::whichWord(rhsBits - 1) + 1);
+
+ // Check the degenerate cases
+ if (lhsWords == 0) {
+ Quotient = 0; // 0 / Y ===> 0
+ Remainder = 0; // 0 % Y ===> 0
+ return;
+ }
+
+ if (lhsWords < rhsWords || LHS.ult(RHS)) {
+ Quotient = 0; // X / Y ===> 0, iff X < Y
+ Remainder = LHS; // X % Y ===> X, iff X < Y
+ return;
+ }
+
+ if (LHS == RHS) {
+ Quotient = 1; // X / X ===> 1
+ Remainder = 0; // X % X ===> 0;
+ return;
+ }
+
+ if (lhsWords == 1 && rhsWords == 1) {
+ // There is only one word to consider so use the native versions.
+ if (LHS.isSingleWord()) {
+ Quotient = APInt(LHS.getBitWidth(), LHS.VAL / RHS.VAL);
+ Remainder = APInt(LHS.getBitWidth(), LHS.VAL % RHS.VAL);
+ } else {
+ Quotient = APInt(LHS.getBitWidth(), LHS.pVal[0] / RHS.pVal[0]);
+ Remainder = APInt(LHS.getBitWidth(), LHS.pVal[0] % RHS.pVal[0]);
+ }
+ return;
+ }
+
+ // Okay, lets do it the long way
+ divide(LHS, lhsWords, RHS, rhsWords, &Quotient, &Remainder);
+}
+
void APInt::fromString(uint32_t numbits, const char *str, uint32_t slen,
uint8_t radix) {
// Check our assumptions here
assert((radix == 10 || radix == 8 || radix == 16 || radix == 2) &&
"Radix should be 2, 8, 10, or 16!");
assert(str && "String is null?");
- bool isNegative = str[0] == '-';
- if (isNegative)
+ bool isNeg = str[0] == '-';
+ if (isNeg)
str++, slen--;
- assert(slen <= numbits || radix != 2 && "Insufficient bit width");
- assert(slen*3 <= numbits || radix != 8 && "Insufficient bit width");
- assert(slen*4 <= numbits || radix != 16 && "Insufficient bit width");
- assert((slen*64)/20 <= numbits || radix != 10 && "Insufficient bit width");
+ assert((slen <= numbits || radix != 2) && "Insufficient bit width");
+ assert((slen*3 <= numbits || radix != 8) && "Insufficient bit width");
+ assert((slen*4 <= numbits || radix != 16) && "Insufficient bit width");
+ assert(((slen*64)/22 <= numbits || radix != 10) && "Insufficient bit width");
// Allocate memory
if (!isSingleWord())
// Get a digit
uint32_t digit = 0;
char cdigit = str[i];
- if (isdigit(cdigit))
- digit = cdigit - '0';
- else if (isxdigit(cdigit))
- if (cdigit >= 'a')
+ if (radix == 16) {
+ if (!isxdigit(cdigit))
+ assert(0 && "Invalid hex digit in string");
+ if (isdigit(cdigit))
+ digit = cdigit - '0';
+ else if (cdigit >= 'a')
digit = cdigit - 'a' + 10;
else if (cdigit >= 'A')
digit = cdigit - 'A' + 10;
else
- assert(0 && "huh?");
- else
+ assert(0 && "huh? we shouldn't get here");
+ } else if (isdigit(cdigit)) {
+ digit = cdigit - '0';
+ } else {
assert(0 && "Invalid character in digit string");
+ }
- // Shift or multiple the value by the radix
+ // Shift or multiply the value by the radix
if (shift)
- this->shl(shift);
+ *this <<= shift;
else
*this *= apradix;
*this += apdigit;
}
// If its negative, put it in two's complement form
- if (isNegative) {
+ if (isNeg) {
+ (*this)--;
this->flip();
- (*this)++;
}
}
}
if (radix != 10) {
- uint64_t mask = radix - 1;
- uint32_t shift = (radix == 16 ? 4 : radix == 8 ? 3 : 1);
- uint32_t nibbles = APINT_BITS_PER_WORD / shift;
- for (uint32_t i = 0; i < getNumWords(); ++i) {
- uint64_t value = pVal[i];
- for (uint32_t j = 0; j < nibbles; ++j) {
- result.insert(0, digits[ value & mask ]);
- value >>= shift;
+ // For the 2, 8 and 16 bit cases, we can just shift instead of divide
+ // because the number of bits per digit (1,3 and 4 respectively) divides
+ // equaly. We just shift until there value is zero.
+
+ // First, check for a zero value and just short circuit the logic below.
+ if (*this == 0)
+ result = "0";
+ else {
+ APInt tmp(*this);
+ size_t insert_at = 0;
+ if (wantSigned && this->isNegative()) {
+ // They want to print the signed version and it is a negative value
+ // Flip the bits and add one to turn it into the equivalent positive
+ // value and put a '-' in the result.
+ tmp.flip();
+ tmp++;
+ result = "-";
+ insert_at = 1;
+ }
+ // Just shift tmp right for each digit width until it becomes zero
+ uint32_t shift = (radix == 16 ? 4 : (radix == 8 ? 3 : 1));
+ uint64_t mask = radix - 1;
+ APInt zero(tmp.getBitWidth(), 0);
+ while (tmp.ne(zero)) {
+ unsigned digit = (tmp.isSingleWord() ? tmp.VAL : tmp.pVal[0]) & mask;
+ result.insert(insert_at, digits[digit]);
+ tmp = tmp.lshr(shift);
}
}
return result;
APInt tmp2(tmp.getBitWidth(), 0);
divide(tmp, tmp.getNumWords(), divisor, divisor.getNumWords(), &tmp2,
&APdigit);
- uint32_t digit = APdigit.getValue();
+ uint32_t digit = APdigit.getZExtValue();
assert(digit < radix && "divide failed");
result.insert(insert_at,digits[digit]);
tmp = tmp2;
return result;
}
-#ifndef NDEBUG
void APInt::dump() const
{
cerr << "APInt(" << BitWidth << ")=" << std::setbase(16);
else for (unsigned i = getNumWords(); i > 0; i--) {
cerr << pVal[i-1] << " ";
}
- cerr << " (" << this->toString(10, false) << ")\n" << std::setbase(10);
+ cerr << " U(" << this->toStringUnsigned(10) << ") S("
+ << this->toStringSigned(10) << ")" << std::setbase(10);
+}
+
+// This implements a variety of operations on a representation of
+// arbitrary precision, two's-complement, bignum integer values.
+
+/* Assumed by lowHalf, highHalf, partMSB and partLSB. A fairly safe
+ and unrestricting assumption. */
+COMPILE_TIME_ASSERT(integerPartWidth % 2 == 0);
+
+/* Some handy functions local to this file. */
+namespace {
+
+ /* Returns the integer part with the least significant BITS set.
+ BITS cannot be zero. */
+ inline integerPart
+ lowBitMask(unsigned int bits)
+ {
+ assert (bits != 0 && bits <= integerPartWidth);
+
+ return ~(integerPart) 0 >> (integerPartWidth - bits);
+ }
+
+ /* Returns the value of the lower half of PART. */
+ inline integerPart
+ lowHalf(integerPart part)
+ {
+ return part & lowBitMask(integerPartWidth / 2);
+ }
+
+ /* Returns the value of the upper half of PART. */
+ inline integerPart
+ highHalf(integerPart part)
+ {
+ return part >> (integerPartWidth / 2);
+ }
+
+ /* Returns the bit number of the most significant set bit of a part.
+ If the input number has no bits set -1U is returned. */
+ unsigned int
+ partMSB(integerPart value)
+ {
+ unsigned int n, msb;
+
+ if (value == 0)
+ return -1U;
+
+ n = integerPartWidth / 2;
+
+ msb = 0;
+ do {
+ if (value >> n) {
+ value >>= n;
+ msb += n;
+ }
+
+ n >>= 1;
+ } while (n);
+
+ return msb;
+ }
+
+ /* Returns the bit number of the least significant set bit of a
+ part. If the input number has no bits set -1U is returned. */
+ unsigned int
+ partLSB(integerPart value)
+ {
+ unsigned int n, lsb;
+
+ if (value == 0)
+ return -1U;
+
+ lsb = integerPartWidth - 1;
+ n = integerPartWidth / 2;
+
+ do {
+ if (value << n) {
+ value <<= n;
+ lsb -= n;
+ }
+
+ n >>= 1;
+ } while (n);
+
+ return lsb;
+ }
+}
+
+/* Sets the least significant part of a bignum to the input value, and
+ zeroes out higher parts. */
+void
+APInt::tcSet(integerPart *dst, integerPart part, unsigned int parts)
+{
+ unsigned int i;
+
+ assert (parts > 0);
+
+ dst[0] = part;
+ for(i = 1; i < parts; i++)
+ dst[i] = 0;
+}
+
+/* Assign one bignum to another. */
+void
+APInt::tcAssign(integerPart *dst, const integerPart *src, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ dst[i] = src[i];
+}
+
+/* Returns true if a bignum is zero, false otherwise. */
+bool
+APInt::tcIsZero(const integerPart *src, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ if (src[i])
+ return false;
+
+ return true;
+}
+
+/* Extract the given bit of a bignum; returns 0 or 1. */
+int
+APInt::tcExtractBit(const integerPart *parts, unsigned int bit)
+{
+ return(parts[bit / integerPartWidth]
+ & ((integerPart) 1 << bit % integerPartWidth)) != 0;
+}
+
+/* Set the given bit of a bignum. */
+void
+APInt::tcSetBit(integerPart *parts, unsigned int bit)
+{
+ parts[bit / integerPartWidth] |= (integerPart) 1 << (bit % integerPartWidth);
+}
+
+/* Returns the bit number of the least significant set bit of a
+ number. If the input number has no bits set -1U is returned. */
+unsigned int
+APInt::tcLSB(const integerPart *parts, unsigned int n)
+{
+ unsigned int i, lsb;
+
+ for(i = 0; i < n; i++) {
+ if (parts[i] != 0) {
+ lsb = partLSB(parts[i]);
+
+ return lsb + i * integerPartWidth;
+ }
+ }
+
+ return -1U;
+}
+
+/* Returns the bit number of the most significant set bit of a number.
+ If the input number has no bits set -1U is returned. */
+unsigned int
+APInt::tcMSB(const integerPart *parts, unsigned int n)
+{
+ unsigned int msb;
+
+ do {
+ --n;
+
+ if (parts[n] != 0) {
+ msb = partMSB(parts[n]);
+
+ return msb + n * integerPartWidth;
+ }
+ } while (n);
+
+ return -1U;
+}
+
+/* Copy the bit vector of width srcBITS from SRC, starting at bit
+ srcLSB, to DST, of dstCOUNT parts, such that the bit srcLSB becomes
+ the least significant bit of DST. All high bits above srcBITS in
+ DST are zero-filled. */
+void
+APInt::tcExtract(integerPart *dst, unsigned int dstCount, const integerPart *src,
+ unsigned int srcBits, unsigned int srcLSB)
+{
+ unsigned int firstSrcPart, dstParts, shift, n;
+
+ dstParts = (srcBits + integerPartWidth - 1) / integerPartWidth;
+ assert (dstParts <= dstCount);
+
+ firstSrcPart = srcLSB / integerPartWidth;
+ tcAssign (dst, src + firstSrcPart, dstParts);
+
+ shift = srcLSB % integerPartWidth;
+ tcShiftRight (dst, dstParts, shift);
+
+ /* We now have (dstParts * integerPartWidth - shift) bits from SRC
+ in DST. If this is less that srcBits, append the rest, else
+ clear the high bits. */
+ n = dstParts * integerPartWidth - shift;
+ if (n < srcBits) {
+ integerPart mask = lowBitMask (srcBits - n);
+ dst[dstParts - 1] |= ((src[firstSrcPart + dstParts] & mask)
+ << n % integerPartWidth);
+ } else if (n > srcBits) {
+ if (srcBits % integerPartWidth)
+ dst[dstParts - 1] &= lowBitMask (srcBits % integerPartWidth);
+ }
+
+ /* Clear high parts. */
+ while (dstParts < dstCount)
+ dst[dstParts++] = 0;
+}
+
+/* DST += RHS + C where C is zero or one. Returns the carry flag. */
+integerPart
+APInt::tcAdd(integerPart *dst, const integerPart *rhs,
+ integerPart c, unsigned int parts)
+{
+ unsigned int i;
+
+ assert(c <= 1);
+
+ for(i = 0; i < parts; i++) {
+ integerPart l;
+
+ l = dst[i];
+ if (c) {
+ dst[i] += rhs[i] + 1;
+ c = (dst[i] <= l);
+ } else {
+ dst[i] += rhs[i];
+ c = (dst[i] < l);
+ }
+ }
+
+ return c;
+}
+
+/* DST -= RHS + C where C is zero or one. Returns the carry flag. */
+integerPart
+APInt::tcSubtract(integerPart *dst, const integerPart *rhs,
+ integerPart c, unsigned int parts)
+{
+ unsigned int i;
+
+ assert(c <= 1);
+
+ for(i = 0; i < parts; i++) {
+ integerPart l;
+
+ l = dst[i];
+ if (c) {
+ dst[i] -= rhs[i] + 1;
+ c = (dst[i] >= l);
+ } else {
+ dst[i] -= rhs[i];
+ c = (dst[i] > l);
+ }
+ }
+
+ return c;
+}
+
+/* Negate a bignum in-place. */
+void
+APInt::tcNegate(integerPart *dst, unsigned int parts)
+{
+ tcComplement(dst, parts);
+ tcIncrement(dst, parts);
+}
+
+/* DST += SRC * MULTIPLIER + CARRY if add is true
+ DST = SRC * MULTIPLIER + CARRY if add is false
+
+ Requires 0 <= DSTPARTS <= SRCPARTS + 1. If DST overlaps SRC
+ they must start at the same point, i.e. DST == SRC.
+
+ If DSTPARTS == SRCPARTS + 1 no overflow occurs and zero is
+ returned. Otherwise DST is filled with the least significant
+ DSTPARTS parts of the result, and if all of the omitted higher
+ parts were zero return zero, otherwise overflow occurred and
+ return one. */
+int
+APInt::tcMultiplyPart(integerPart *dst, const integerPart *src,
+ integerPart multiplier, integerPart carry,
+ unsigned int srcParts, unsigned int dstParts,
+ bool add)
+{
+ unsigned int i, n;
+
+ /* Otherwise our writes of DST kill our later reads of SRC. */
+ assert(dst <= src || dst >= src + srcParts);
+ assert(dstParts <= srcParts + 1);
+
+ /* N loops; minimum of dstParts and srcParts. */
+ n = dstParts < srcParts ? dstParts: srcParts;
+
+ for(i = 0; i < n; i++) {
+ integerPart low, mid, high, srcPart;
+
+ /* [ LOW, HIGH ] = MULTIPLIER * SRC[i] + DST[i] + CARRY.
+
+ This cannot overflow, because
+
+ (n - 1) * (n - 1) + 2 (n - 1) = (n - 1) * (n + 1)
+
+ which is less than n^2. */
+
+ srcPart = src[i];
+
+ if (multiplier == 0 || srcPart == 0) {
+ low = carry;
+ high = 0;
+ } else {
+ low = lowHalf(srcPart) * lowHalf(multiplier);
+ high = highHalf(srcPart) * highHalf(multiplier);
+
+ mid = lowHalf(srcPart) * highHalf(multiplier);
+ high += highHalf(mid);
+ mid <<= integerPartWidth / 2;
+ if (low + mid < low)
+ high++;
+ low += mid;
+
+ mid = highHalf(srcPart) * lowHalf(multiplier);
+ high += highHalf(mid);
+ mid <<= integerPartWidth / 2;
+ if (low + mid < low)
+ high++;
+ low += mid;
+
+ /* Now add carry. */
+ if (low + carry < low)
+ high++;
+ low += carry;
+ }
+
+ if (add) {
+ /* And now DST[i], and store the new low part there. */
+ if (low + dst[i] < low)
+ high++;
+ dst[i] += low;
+ } else
+ dst[i] = low;
+
+ carry = high;
+ }
+
+ if (i < dstParts) {
+ /* Full multiplication, there is no overflow. */
+ assert(i + 1 == dstParts);
+ dst[i] = carry;
+ return 0;
+ } else {
+ /* We overflowed if there is carry. */
+ if (carry)
+ return 1;
+
+ /* We would overflow if any significant unwritten parts would be
+ non-zero. This is true if any remaining src parts are non-zero
+ and the multiplier is non-zero. */
+ if (multiplier)
+ for(; i < srcParts; i++)
+ if (src[i])
+ return 1;
+
+ /* We fitted in the narrow destination. */
+ return 0;
+ }
+}
+
+/* DST = LHS * RHS, where DST has the same width as the operands and
+ is filled with the least significant parts of the result. Returns
+ one if overflow occurred, otherwise zero. DST must be disjoint
+ from both operands. */
+int
+APInt::tcMultiply(integerPart *dst, const integerPart *lhs,
+ const integerPart *rhs, unsigned int parts)
+{
+ unsigned int i;
+ int overflow;
+
+ assert(dst != lhs && dst != rhs);
+
+ overflow = 0;
+ tcSet(dst, 0, parts);
+
+ for(i = 0; i < parts; i++)
+ overflow |= tcMultiplyPart(&dst[i], lhs, rhs[i], 0, parts,
+ parts - i, true);
+
+ return overflow;
+}
+
+/* DST = LHS * RHS, where DST has width the sum of the widths of the
+ operands. No overflow occurs. DST must be disjoint from both
+ operands. Returns the number of parts required to hold the
+ result. */
+unsigned int
+APInt::tcFullMultiply(integerPart *dst, const integerPart *lhs,
+ const integerPart *rhs, unsigned int lhsParts,
+ unsigned int rhsParts)
+{
+ /* Put the narrower number on the LHS for less loops below. */
+ if (lhsParts > rhsParts) {
+ return tcFullMultiply (dst, rhs, lhs, rhsParts, lhsParts);
+ } else {
+ unsigned int n;
+
+ assert(dst != lhs && dst != rhs);
+
+ tcSet(dst, 0, rhsParts);
+
+ for(n = 0; n < lhsParts; n++)
+ tcMultiplyPart(&dst[n], rhs, lhs[n], 0, rhsParts, rhsParts + 1, true);
+
+ n = lhsParts + rhsParts;
+
+ return n - (dst[n - 1] == 0);
+ }
+}
+
+/* If RHS is zero LHS and REMAINDER are left unchanged, return one.
+ Otherwise set LHS to LHS / RHS with the fractional part discarded,
+ set REMAINDER to the remainder, return zero. i.e.
+
+ OLD_LHS = RHS * LHS + REMAINDER
+
+ SCRATCH is a bignum of the same size as the operands and result for
+ use by the routine; its contents need not be initialized and are
+ destroyed. LHS, REMAINDER and SCRATCH must be distinct.
+*/
+int
+APInt::tcDivide(integerPart *lhs, const integerPart *rhs,
+ integerPart *remainder, integerPart *srhs,
+ unsigned int parts)
+{
+ unsigned int n, shiftCount;
+ integerPart mask;
+
+ assert(lhs != remainder && lhs != srhs && remainder != srhs);
+
+ shiftCount = tcMSB(rhs, parts) + 1;
+ if (shiftCount == 0)
+ return true;
+
+ shiftCount = parts * integerPartWidth - shiftCount;
+ n = shiftCount / integerPartWidth;
+ mask = (integerPart) 1 << (shiftCount % integerPartWidth);
+
+ tcAssign(srhs, rhs, parts);
+ tcShiftLeft(srhs, parts, shiftCount);
+ tcAssign(remainder, lhs, parts);
+ tcSet(lhs, 0, parts);
+
+ /* Loop, subtracting SRHS if REMAINDER is greater and adding that to
+ the total. */
+ for(;;) {
+ int compare;
+
+ compare = tcCompare(remainder, srhs, parts);
+ if (compare >= 0) {
+ tcSubtract(remainder, srhs, 0, parts);
+ lhs[n] |= mask;
+ }
+
+ if (shiftCount == 0)
+ break;
+ shiftCount--;
+ tcShiftRight(srhs, parts, 1);
+ if ((mask >>= 1) == 0)
+ mask = (integerPart) 1 << (integerPartWidth - 1), n--;
+ }
+
+ return false;
+}
+
+/* Shift a bignum left COUNT bits in-place. Shifted in bits are zero.
+ There are no restrictions on COUNT. */
+void
+APInt::tcShiftLeft(integerPart *dst, unsigned int parts, unsigned int count)
+{
+ if (count) {
+ unsigned int jump, shift;
+
+ /* Jump is the inter-part jump; shift is is intra-part shift. */
+ jump = count / integerPartWidth;
+ shift = count % integerPartWidth;
+
+ while (parts > jump) {
+ integerPart part;
+
+ parts--;
+
+ /* dst[i] comes from the two parts src[i - jump] and, if we have
+ an intra-part shift, src[i - jump - 1]. */
+ part = dst[parts - jump];
+ if (shift) {
+ part <<= shift;
+ if (parts >= jump + 1)
+ part |= dst[parts - jump - 1] >> (integerPartWidth - shift);
+ }
+
+ dst[parts] = part;
+ }
+
+ while (parts > 0)
+ dst[--parts] = 0;
+ }
+}
+
+/* Shift a bignum right COUNT bits in-place. Shifted in bits are
+ zero. There are no restrictions on COUNT. */
+void
+APInt::tcShiftRight(integerPart *dst, unsigned int parts, unsigned int count)
+{
+ if (count) {
+ unsigned int i, jump, shift;
+
+ /* Jump is the inter-part jump; shift is is intra-part shift. */
+ jump = count / integerPartWidth;
+ shift = count % integerPartWidth;
+
+ /* Perform the shift. This leaves the most significant COUNT bits
+ of the result at zero. */
+ for(i = 0; i < parts; i++) {
+ integerPart part;
+
+ if (i + jump >= parts) {
+ part = 0;
+ } else {
+ part = dst[i + jump];
+ if (shift) {
+ part >>= shift;
+ if (i + jump + 1 < parts)
+ part |= dst[i + jump + 1] << (integerPartWidth - shift);
+ }
+ }
+
+ dst[i] = part;
+ }
+ }
+}
+
+/* Bitwise and of two bignums. */
+void
+APInt::tcAnd(integerPart *dst, const integerPart *rhs, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ dst[i] &= rhs[i];
+}
+
+/* Bitwise inclusive or of two bignums. */
+void
+APInt::tcOr(integerPart *dst, const integerPart *rhs, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ dst[i] |= rhs[i];
+}
+
+/* Bitwise exclusive or of two bignums. */
+void
+APInt::tcXor(integerPart *dst, const integerPart *rhs, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ dst[i] ^= rhs[i];
+}
+
+/* Complement a bignum in-place. */
+void
+APInt::tcComplement(integerPart *dst, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ dst[i] = ~dst[i];
+}
+
+/* Comparison (unsigned) of two bignums. */
+int
+APInt::tcCompare(const integerPart *lhs, const integerPart *rhs,
+ unsigned int parts)
+{
+ while (parts) {
+ parts--;
+ if (lhs[parts] == rhs[parts])
+ continue;
+
+ if (lhs[parts] > rhs[parts])
+ return 1;
+ else
+ return -1;
+ }
+
+ return 0;
+}
+
+/* Increment a bignum in-place, return the carry flag. */
+integerPart
+APInt::tcIncrement(integerPart *dst, unsigned int parts)
+{
+ unsigned int i;
+
+ for(i = 0; i < parts; i++)
+ if (++dst[i] != 0)
+ break;
+
+ return i == parts;
+}
+
+/* Set the least significant BITS bits of a bignum, clear the
+ rest. */
+void
+APInt::tcSetLeastSignificantBits(integerPart *dst, unsigned int parts,
+ unsigned int bits)
+{
+ unsigned int i;
+
+ i = 0;
+ while (bits > integerPartWidth) {
+ dst[i++] = ~(integerPart) 0;
+ bits -= integerPartWidth;
+ }
+
+ if (bits)
+ dst[i++] = ~(integerPart) 0 >> (integerPartWidth - bits);
+
+ while (i < parts)
+ dst[i++] = 0;
}
-#endif