//
// 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.
//
//===----------------------------------------------------------------------===//
//
#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
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");
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() {
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.
return Count;
}
+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();
+ 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 Count;
}
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");
uint32_t wordsBefore = 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];
// 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 = wordBits == 0 ? 0 : ~0ULL << wordBits;
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");
uint32_t wordsBefore = 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);
+
+ // 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 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();
}
+
+// 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());
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())
*this += apdigit;
}
// If its negative, put it in two's complement form
- if (isNegative) {
+ if (isNeg) {
+ (*this)--;
this->flip();
- (*this)++;
}
}
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;
else for (unsigned i = getNumWords(); i > 0; i--) {
cerr << pVal[i-1] << " ";
}
- cerr << " (" << this->toString(10, false) << ")\n" << std::setbase(10);
+ cerr << " U(" << this->toString(10) << ") S(" << this->toStringSigned(10)
+ << ")\n" << std::setbase(10);
}
#endif
projects / oota-llvm.git / blobdiff