30672136635

Committed 31 Jul 2026 11:10PM UTC coverage: 85.23% (-0.02%) from 85.249%

Build # 30672136635

Build Type

push

github

Committed by

web-flow

Commit Message

fix(posit): clear -Wsign-conversion in posit_impl convert_ (#1281) (#1286)

Sub-issue of Epic #1279. Four sites in the posit encode/rounding path
(convert_), all value-preserving:

- `exponent = e % (1ull << es)`: the signed scale `e` was already promoted
  to unsigned long long by the modulo (the other operand is 1ull); make it
  explicit. This is correct for a power-of-two modulus -- 2^64 mod 2^es is
  0, so the unsigned reinterpretation of a two's-complement negative e
  yields exactly the low es bits, which is the posit exponent field (and
  matches the arithmetic-shift regime split e >> es).
- `regime <<= es + nrFbits + 1u`, `exponent <<= nrFbits + 1u`,
  `pt_bits <<= pt_len - len`: unsigned shift-amount / left-align distance
  flowing into operator<<=(int); each is >= 1 and small by construction.
  Cast to int explicitly.

Proven behavior-neutral: optimized assembly (g++ -O2) for a TU
instantiating convert_ across posit<8,0>/<16,1>/<32,2>/<64,3> is
byte-for-byte identical before and after -- the casts are explicit forms
of conversions the compiler already performed implicitly.

The dead `convert_to_bb` (inside `#ifdef TBD`, never called) has the same
idiom but does not compile, so it is intentionally left untouched.

Verified posit_impl -Wsign-conversion AND -Wconversion clean on gcc AND
clang; posit conversion/assignment/multiplication regressions PASS on
both. (The posit<64,2..4> division failures are pre-existing on main --
identical counts with the fix reverted -- and out of scope here.)

Relates to #1279
Resolves #1281

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Coverage Stats

4 of 4 new or added lines in 1 file covered. (100.0%)

36 existing lines in 4 files now uncovered.

41708 of 48936 relevant lines covered (85.23%)

7271799.74 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

94.69

/include/sw/universal/internal/blockbinary/blockbinary.hpp

#pragma once
// blockbinary.hpp: parameterized blocked binary number system representing a 2's complement binary number
//
// Copyright (C) 2017 Stillwater Supercomputing, Inc.
// SPDX-License-Identifier: MIT
//
// This file is part of the universal numbers project, which is released under an MIT Open Source license.
#include <cstdint>
#include <iostream>
#include <iomanip>
#include <string>
#include <sstream>
#include <type_traits>
#include <universal/number/shared/specific_value_encoding.hpp>
#include <universal/internal/blocktype/carry.hpp>

#include <universal/internal/bit_manipulation.hpp>

namespace sw { namespace universal {

enum class BinaryNumberType {
        Signed   = 0, // { ...,-3,-2,-1,0,1,2,3,... }    // 2's complement encoding
        Unsigned = 1  // {              0,1,2,3,... }    // binary encoding
};

// forward references
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType> class blockbinary;
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType> constexpr blockbinary<nbits, BlockType, NumberType> twosComplement(const blockbinary<nbits, BlockType, NumberType>&);
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType> struct quorem;
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType> constexpr quorem<nbits, BlockType, NumberType> longdivision(const blockbinary<nbits, BlockType, NumberType>&, const blockbinary<nbits, BlockType, NumberType>&);

// idiv_t for blockbinary<nbits> to capture quotient and remainder during long division
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
struct quorem {
        int exceptionId;
        blockbinary<nbits, BlockType, NumberType> quo; // quotient
        blockbinary<nbits, BlockType, NumberType> rem;  // remainder
};

// maximum positive 2's complement number: b01111...1111
template<unsigned nbits, typename BlockType = uint8_t, BinaryNumberType NumberType>
constexpr blockbinary<nbits, BlockType, NumberType>& maxpos(blockbinary<nbits, BlockType, NumberType>& a) {
        a.clear();
        a.flip();
        if constexpr (NumberType == BinaryNumberType::Signed) {
                a.setbit(nbits - 1, false);
        }
        return a;
}

// maximum negative 2's complement number: b1000...0000
template<unsigned nbits, typename BlockType = uint8_t, BinaryNumberType NumberType>
constexpr blockbinary<nbits, BlockType, NumberType>& maxneg(blockbinary<nbits, BlockType, NumberType>& a) {
        a.clear();
        if constexpr (NumberType == BinaryNumberType::Signed) {
                a.setbit(nbits - 1);
        }
        return a;
}

// generate the 2's complement of the block binary number
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
constexpr blockbinary<nbits, BlockType, NumberType> twosComplement(const blockbinary<nbits, BlockType, NumberType>& orig) {
        blockbinary<nbits, BlockType, NumberType> twosC(orig);
        blockbinary<nbits, BlockType, NumberType> plusOne(1);
        twosC.flip();
        twosC += plusOne;
        return twosC;
}

// Truncate a bigger posit to fit in a smaller
template<unsigned srcbits, unsigned tgtbits, typename bt, BinaryNumberType nt>
constexpr void truncate(const blockbinary<srcbits, bt, nt>& src, blockbinary<tgtbits, bt, nt>& tgt) {
        static_assert(tgtbits < srcbits, "truncate requires source posit to be bigger than target posit");
        constexpr unsigned diff = srcbits - tgtbits;
        for (unsigned i = 0; i < tgtbits; ++i) { // TODO: optimize for limbs
                tgt.setbit(i, src.test(i + diff));
        }
}

/*
NOTES

For block arithmetic, we need to manage a carry bit.
For uint8_t, uint16_t, and uint32_t limbs, we cast up to uint64_t to detect overflow.
For uint64_t limbs, there is no larger native type, so we use platform-specific
intrinsics (see carry.hpp) for carry propagation: compiler builtins on MSVC,
unsigned __int128 on GCC/Clang, or a portable comparison-based fallback.
*/

// a block-based binary number configurable to be signed or unsigned. When signed it uses 2's complement encoding
template<unsigned _nbits, typename bt = uint8_t, BinaryNumberType _NumberType = BinaryNumberType::Signed>
class blockbinary {
public:
        static constexpr unsigned nbits = _nbits;
        typedef bt BlockType;
        static constexpr BinaryNumberType NumberType = _NumberType;

        static constexpr unsigned bitsInByte = 8;
        static constexpr unsigned bitsInBlock = sizeof(bt) * bitsInByte;
        static constexpr unsigned nrBlocks = (0 == nbits ? 1 : (1ull + ((nbits - 1ull) / bitsInBlock)));
        static constexpr uint64_t storageMask = (0xFFFFFFFFFFFFFFFFull >> (64 - bitsInBlock));
        static constexpr bt       maxBlockValue = bt(-1);

        static constexpr unsigned MSU = nrBlocks - 1; // MSU == Most Significant Unit
        static constexpr bt       ALL_ONES = bt(~0);
        static constexpr unsigned maxShift = (0 == nbits ? 0 : (nrBlocks* bitsInBlock - nbits)); // protect the shift that is >= sizeof(bt)
        static constexpr bt       MSU_MASK = (0 == nbits ? bt(0) : (ALL_ONES >> maxShift));      // the other side of this protection
        static constexpr bt       SIGN_BIT_MASK = (0 == nbits ? bt(0) : (bt(bt(1) << ((nbits - 1ull) % bitsInBlock))));

        // uint64_t multi-block arithmetic is supported via carry-detection intrinsics (see carry.hpp)
        static_assert(bitsInBlock <= 64, "storage unit for block arithmetic needs to be one of [uint8_t | uint16_t | uint32_t | uint64_t]");

        /// trivial constructor
        blockbinary() = default;

        /// construct a blockbinary from another: bt must be the same
        template<unsigned nnbits>
        constexpr blockbinary(const blockbinary<nnbits, BlockType, NumberType>& rhs) : _block{} { this->assign(rhs); }

        // initializer for long long
        constexpr blockbinary(long long initial_value) noexcept : _block{} { *this = initial_value; }

        // specific value constructors
        constexpr blockbinary(const std::string& s) noexcept : _block{} {  }  // TODO
        constexpr blockbinary(const SpecificValue code) : _block{} {
                switch (code) {
                case SpecificValue::infpos:
                case SpecificValue::maxpos:
                        maxpos();
                        break;
                case SpecificValue::minpos:
                        minpos();
                        break;
                case SpecificValue::qnan:
                case SpecificValue::snan:
                case SpecificValue::nar:
                case SpecificValue::zero:
                default:
                        zero();
                        break;
                case SpecificValue::minneg:
                        minneg();
                        break;
                case SpecificValue::infneg:
                case SpecificValue::maxneg:
                        maxneg();
                        break;
                }
        }

        constexpr blockbinary& operator=(long long rhs) noexcept {
                if constexpr (1 < nrBlocks) {
                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                _block[i] = static_cast<uint64_t>(rhs) & storageMask;
                                if constexpr (bitsInBlock < 64) {
                                        rhs >>= bitsInBlock;
                                }
                                else {
                                        rhs = (rhs < 0) ? -1LL : 0LL; // sign-extend for uint64_t limbs
                                }
                        }
                        // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                        _block[MSU] &= MSU_MASK;
                }
                else if constexpr (1 == nrBlocks) {
                        _block[0] = static_cast<uint64_t>(rhs) & storageMask;
                        // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                        _block[MSU] &= MSU_MASK;
                }
                return *this;
        }

        // conversion operators
        constexpr explicit operator int() const                { return int(to_sll()); }
        constexpr explicit operator long() const               { return long(to_sll()); }
        constexpr explicit operator long long() const          { return to_sll(); }
        constexpr explicit operator unsigned int() const       { return unsigned(to_ull()); }
        constexpr explicit operator unsigned long() const      { return (unsigned long)to_ull(); }
        constexpr explicit operator unsigned long long() const { return to_ull(); }
        // TODO: these need proper implementations that can convert very large integers to the proper scale afforded by the floating-point formats
        explicit operator float() const              { return to_native<float>(); }
        explicit operator double() const             { return to_native<double>(); }

#if LONG_DOUBLE_SUPPORT
        explicit operator long double() const        { return to_native<long double>(); }
#endif

        // limb access operators
//        constexpr BlockType& operator[](unsigned index) { return _block[index]; }
        constexpr BlockType operator[](unsigned index) const { return _block[index]; }

        // prefix operators
        constexpr blockbinary operator-() const {
                blockbinary negated(*this);
                blockbinary plusOne(1);
                negated.flip();
                negated += plusOne;
                return negated;
        }
        // one's complement
        constexpr blockbinary operator~() const {
                blockbinary complement(*this);
                complement.flip();
                return complement;
        }
        // increment/decrement
        constexpr blockbinary operator++(int) {
                blockbinary tmp(*this);
                operator++();
                return tmp;
        }
        constexpr blockbinary& operator++() {
                blockbinary increment;
                increment.setbits(0x1);
                *this += increment;
                return *this;
        }
        constexpr blockbinary operator--(int) {
                blockbinary tmp(*this);
                operator--();
                return tmp;
        }
        constexpr blockbinary& operator--() {
                blockbinary decrement;
                decrement.setbits(0x1);
                return *this -= decrement;
        }
        // arithmetic operators
        constexpr blockbinary& operator+=(const blockbinary& rhs) {
                if constexpr (nrBlocks == 1) {
                        _block[0] = static_cast<bt>(_block[0] + rhs.block(0));
                        // null any leading bits that fall outside of nbits
                        _block[MSU] = static_cast<bt>(MSU_MASK & _block[MSU]);
                }
                else if constexpr (bitsInBlock == 64) {
                        // uint64_t limbs: addcarry uses platform intrinsics (not constexpr on MSVC).
                        // In a constant-evaluated context, fall back to portable carry detection.
                        blockbinary sum;
                        if (std::is_constant_evaluated()) {
                                uint64_t carry = 0;
                                for (unsigned i = 0; i < nrBlocks; ++i) {
                                        uint64_t a = _block[i];
                                        uint64_t b = rhs._block[i];
                                        uint64_t s1 = a + carry;
                                        uint64_t c1 = (s1 < a) ? 1ull : 0ull;
                                        uint64_t r  = s1 + b;
                                        uint64_t c2 = (r < s1) ? 1ull : 0ull;
                                        sum._block[i] = r;
                                        carry = c1 + c2;
                                }
                        }
                        else {
                                uint64_t carry = 0;
                                for (unsigned i = 0; i < nrBlocks; ++i) {
                                        sum._block[i] = addcarry(_block[i], rhs._block[i], carry, carry);
                                }
                        }
                        // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                        sum._block[MSU] = static_cast<bt>(MSU_MASK & sum._block[MSU]);
                        *this = sum;
                }
                else {
                        // Multi-limb path for bt = uint8_t/uint16_t/uint32_t.
                        // Index-based loop (constexpr-friendly).
                        blockbinary sum;
                        std::uint64_t carry = 0;
                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                carry += static_cast<std::uint64_t>(_block[i]) + static_cast<std::uint64_t>(rhs._block[i]);
                                sum._block[i] = static_cast<bt>(carry);
                                carry >>= bitsInBlock;
                        }
                        // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                        sum._block[MSU] = static_cast<bt>(MSU_MASK & sum._block[MSU]);
                        *this = sum;
                }
                return *this;
        }
        constexpr blockbinary& operator-=(const blockbinary& rhs) {
                return operator+=(sw::universal::twosComplement(rhs));
        }
#define BLOCKBINARY_FAST_MUL
#ifdef BLOCKBINARY_FAST_MUL
        constexpr blockbinary& operator*=(const blockbinary& rhs) {
                if constexpr (NumberType == BinaryNumberType::Signed) {
                        if constexpr (nrBlocks == 1) {
                                _block[0] = static_cast<bt>(static_cast<std::uint64_t>(_block[0]) * static_cast<std::uint64_t>(rhs.block(0)));
                        }
                        else if constexpr (bitsInBlock == 64) {
                                // uint64_t limbs: mul128/addcarry use platform intrinsics that are not
                                // constexpr on MSVC. In a constant-evaluated context, fall back to a
                                // portable __uint128_t-based carry-propagation loop (gcc/clang only).
                                blockbinary<nbits + 1, BlockType, NumberType> base(*this);
                                blockbinary<nbits + 1, BlockType, NumberType> multiplicant(rhs);
                                bool resultIsNeg = (base.isneg() ^ multiplicant.isneg());
                                if (base.isneg()) {
                                        base.twosComplement();
                                }
                                if (multiplicant.isneg()) {
                                        multiplicant.twosComplement();
                                }
                                clear();
                                if (std::is_constant_evaluated()) {
#if defined(__SIZEOF_INT128__)
                                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                                uint128_t carry = 0;
                                                for (unsigned j = 0; j < nrBlocks; ++j) {
                                                        if (i + j < nrBlocks) {
                                                                uint128_t product = static_cast<uint128_t>(base.block(i)) * multiplicant.block(j);
                                                                uint128_t sum = product + _block[i + j] + carry;
                                                                _block[i + j] = static_cast<bt>(sum);
                                                                carry = sum >> 64;
                                                        }
                                                }
                                        }
#else
                                        // Constexpr eval of uint64-limb mul without __int128 is unsupported
                                        // (MSVC). Forcing a throw in the constant-evaluated branch turns this
                                        // into a compile error rather than a silent zero-result. The branch
                                        // is unreachable at runtime because is_constant_evaluated() is false
                                        // there; the runtime intrinsic path below handles all execution.
                                        // Users on MSVC needing compile-time eval should use uint32_t limbs.
                                        throw "blockbinary<N, uint64_t> constexpr multiply requires __int128 support (gcc/clang); on MSVC use uint32_t limbs for compile-time evaluation";
#endif
                                }
                                else {
                                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                                uint64_t carry = 0;
                                                for (unsigned j = 0; j < nrBlocks; ++j) {
                                                        if (i + j < nrBlocks) {
                                                                uint64_t lo, hi;
                                                                mul128(base.block(i), multiplicant.block(j), lo, hi);
                                                                // accumulate: _block[i+j] += lo + carry
                                                                // use two separate additions to avoid passing multi-bit carry
                                                                // to addcarry (MSVC _addcarry_u64 only accepts 0/1 carry_in)
                                                                uint64_t c1 = 0;
                                                                uint64_t sum = addcarry(_block[i + j], lo, 0, c1);
                                                                uint64_t c2 = 0;
                                                                sum = addcarry(sum, carry, 0, c2);
                                                                _block[i + j] = sum;
                                                                carry = hi + c1 + c2;
                                                        }
                                                }
                                        }
                                }
                                if (resultIsNeg) twosComplement();
                        }
                        else {
                                // is there a better way than upconverting to deal with maxneg in a 2's complement encoding?
                                blockbinary<nbits + 1, BlockType, NumberType> base(*this);
                                blockbinary<nbits + 1, BlockType, NumberType> multiplicant(rhs);
                                bool resultIsNeg = (base.isneg() ^ multiplicant.isneg());
                                if (base.isneg()) {
                                        base.twosComplement();
                                }
                                if (multiplicant.isneg()) {
                                        multiplicant.twosComplement();
                                }
                                clear();
                                for (unsigned i = 0; i < static_cast<unsigned>(nrBlocks); ++i) {
                                        std::uint64_t segment(0);
                                        for (unsigned j = 0; j < static_cast<unsigned>(nrBlocks); ++j) {
                                                segment += static_cast<std::uint64_t>(base.block(i)) * static_cast<std::uint64_t>(multiplicant.block(j));

                                                if (i + j < static_cast<unsigned>(nrBlocks)) {
                                                        segment += _block[i + j];
                                                        _block[i + j] = static_cast<bt>(segment);
                                                        segment >>= bitsInBlock;
                                                }
                                        }
                                }
                                if (resultIsNeg) twosComplement();
                        }
                }
                else {  // unsigned
                        if constexpr (nrBlocks == 1) {
                                _block[0] = static_cast<bt>(static_cast<std::uint64_t>(_block[0]) * static_cast<std::uint64_t>(rhs.block(0)));
                        }
                        else if constexpr (bitsInBlock == 64) {
                                // uint64_t limbs: same is_constant_evaluated dispatch as the signed path.
                                blockbinary base(*this);
                                blockbinary multiplicant(rhs);
                                clear();
                                if (std::is_constant_evaluated()) {
#if defined(__SIZEOF_INT128__)
                                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                                uint128_t carry = 0;
                                                for (unsigned j = 0; j < nrBlocks; ++j) {
                                                        if (i + j < nrBlocks) {
                                                                uint128_t product = static_cast<uint128_t>(base.block(i)) * multiplicant.block(j);
                                                                uint128_t sum = product + _block[i + j] + carry;
                                                                _block[i + j] = static_cast<bt>(sum);
                                                                carry = sum >> 64;
                                                        }
                                                }
                                        }
#else
                                        // See the signed branch: forcing a throw turns silent zero-result
                                        // into a compile error on platforms without __int128 (MSVC).
                                        throw "blockbinary<N, uint64_t> constexpr multiply requires __int128 support (gcc/clang); on MSVC use uint32_t limbs for compile-time evaluation";
#endif
                                }
                                else {
                                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                                uint64_t carry = 0;
                                                for (unsigned j = 0; j < nrBlocks; ++j) {
                                                        if (i + j < nrBlocks) {
                                                                uint64_t lo, hi;
                                                                mul128(base.block(i), multiplicant.block(j), lo, hi);
                                                                // accumulate: _block[i+j] += lo + carry
                                                                // use two separate additions to avoid passing multi-bit carry
                                                                // to addcarry (MSVC _addcarry_u64 only accepts 0/1 carry_in)
                                                                uint64_t c1 = 0;
                                                                uint64_t sum = addcarry(_block[i + j], lo, 0, c1);
                                                                uint64_t c2 = 0;
                                                                sum = addcarry(sum, carry, 0, c2);
                                                                _block[i + j] = sum;
                                                                carry = hi + c1 + c2;
                                                        }
                                                }
                                        }
                                }
                        }
                        else {
                                blockbinary base(*this);
                                blockbinary multiplicant(rhs);
                                clear();
                                for (unsigned i = 0; i < static_cast<unsigned>(nrBlocks); ++i) {
                                        std::uint64_t segment(0);
                                        for (unsigned j = 0; j < static_cast<unsigned>(nrBlocks); ++j) {
                                                segment += static_cast<std::uint64_t>(base.block(i)) * static_cast<std::uint64_t>(multiplicant.block(j));

                                                if (i + j < static_cast<unsigned>(nrBlocks)) {
                                                        segment += _block[i + j];
                                                        _block[i + j] = static_cast<bt>(segment);
                                                        segment >>= bitsInBlock;
                                                }
                                        }
                                }
                        }
                }
                // null any leading bits that fall outside of nbits
                _block[MSU] = static_cast<bt>(MSU_MASK & _block[MSU]);
                return *this;
        }
#else
        constexpr blockbinary& operator*=(const blockbinary& rhs) { // modulo in-place
                blockbinary base(*this);
                blockbinary multiplicant(rhs);
                clear();
                for (unsigned i = 0; i < nbits; ++i) {
                        if (base.at(i)) {
                                operator+=(multiplicant);
                        }
                        multiplicant <<= 1;
                }
                // since we used operator+=, which enforces the nulling of leading bits
                // we don't need to null here
                return *this;
        }
#endif
        constexpr blockbinary& operator/=(const blockbinary& rhs) {
                if constexpr (nbits == (sizeof(BlockType) * 8)) {
                        if (rhs.iszero()) {
                                *this = 0;
                                return *this;
                        }
                        if constexpr (sizeof(BlockType) == 1) {
                                _block[0] = static_cast<bt>(std::int8_t(_block[0]) / std::int8_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 2) {
                                _block[0] = static_cast<bt>(std::int16_t(_block[0]) / std::int16_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 4) {
                                _block[0] = static_cast<bt>(std::int32_t(_block[0]) / std::int32_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 8) {
                                _block[0] = static_cast<bt>(std::int64_t(_block[0]) / std::int64_t(rhs._block[0]));
                        }
                        _block[0] = static_cast<bt>(MSU_MASK & _block[0]);
                }
                else {
                        quorem<nbits, BlockType, NumberType> result = longdivision(*this, rhs);
                        *this = result.quo;
                }
                return *this;
        }
        constexpr blockbinary& operator%=(const blockbinary& rhs) {
                if constexpr (nbits == (sizeof(BlockType) * 8)) {
                        if (rhs.iszero()) {
                                *this = 0;
                                return *this;
                        }
                        if constexpr (sizeof(BlockType) == 1) {
                                _block[0] = static_cast<bt>(std::int8_t(_block[0]) % std::int8_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 2) {
                                _block[0] = static_cast<bt>(std::int16_t(_block[0]) % std::int16_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 4) {
                                _block[0] = static_cast<bt>(std::int32_t(_block[0]) % std::int32_t(rhs._block[0]));
                        }
                        else if constexpr (sizeof(BlockType) == 8) {
                                _block[0] = static_cast<bt>(std::int64_t(_block[0]) % std::int64_t(rhs._block[0]));
                        }
                        _block[0] = static_cast<bt>(MSU_MASK & _block[0]);
                }
                else {
                        quorem<nbits, BlockType, NumberType> result = longdivision(*this, rhs);
                        *this = result.rem;
                }
                return *this;
        }
        
        ///////////////////////////////////////////////////////////////////
        ///              logic operators

        constexpr blockbinary& operator|=(const blockbinary& rhs) noexcept {
                for (unsigned i = 0; i < nrBlocks; ++i) {
                        _block[i] |= rhs._block[i];
                }
                _block[MSU] &= MSU_MASK; // assert precondition of properly nulled leading non-bits
                return *this;
        }
        constexpr blockbinary& operator&=(const blockbinary& rhs) noexcept {
                for (unsigned i = 0; i < nrBlocks; ++i) {
                        _block[i] &= rhs._block[i];
                }
                _block[MSU] &= MSU_MASK; // assert precondition of properly nulled leading non-bits
                return *this;
        }
        constexpr blockbinary& operator^=(const blockbinary& rhs) noexcept {
                for (unsigned i = 0; i < nrBlocks; ++i) {
                        _block[i] ^= rhs._block[i];
                }
                _block[MSU] &= MSU_MASK; // assert precondition of properly nulled leading non-bits
                return *this;
        }

        // shift left operator
        constexpr blockbinary& operator<<=(int bitsToShift) {
                if (bitsToShift == 0) return *this;
                if (bitsToShift < 0) return operator>>=(-bitsToShift);
                if (bitsToShift > static_cast<int>(nbits)) {
                        setzero();
                        return *this;
                }
                if (bitsToShift >= static_cast<int>(bitsInBlock)) {
                        int blockShift = bitsToShift / static_cast<int>(bitsInBlock);
                        for (int i = static_cast<int>(MSU); i >= blockShift; --i) {
                                _block[i] = _block[i - blockShift];
                        }
                        for (int i = blockShift - 1; i >= 0; --i) {
                                _block[i] = bt(0);
                        }
                        // adjust the shift
                        bitsToShift -= static_cast<int>(static_cast<unsigned>(blockShift) * bitsInBlock);
                        // enforce the invariant that bits outside nbits are nulled before the early return
                        if (bitsToShift == 0) { _block[MSU] &= MSU_MASK; return *this; }
                }
                if constexpr (MSU > 0) {
                        // construct the mask for the upper bits in the block that needs to move to the higher word
                        bt mask = bit_high_mask<bt>(static_cast<unsigned>(bitsToShift), bitsInBlock);
                        for (unsigned i = MSU; i > 0; --i) {
                                _block[i] <<= bitsToShift;
                                // mix in the bits from the right
                                bt bits = bt(mask & _block[i - 1]);
                                _block[i] |= (bits >> (bitsInBlock - static_cast<unsigned>(bitsToShift)));
                        }
                }
                _block[0] <<= bitsToShift;
                // left shifts can push valid bits into the unused MSU storage bits; null them
                // so iszero()/operator== (which compare raw blocks) stay correct
                _block[MSU] &= MSU_MASK;
                return *this;
        }
        // arithmetic shift right operator
        constexpr blockbinary& operator>>=(int bitsToShift) {
                if (bitsToShift == 0) return *this;
                if (bitsToShift < 0) return operator<<=(-bitsToShift);
                if (bitsToShift >= static_cast<int>(nbits)) {
                        setzero();
                        return *this;
                }
                bool signext = sign();
                unsigned blockShift = 0;
                if (bitsToShift >= static_cast<int>(bitsInBlock)) {
                        blockShift = static_cast<unsigned>(bitsToShift) / bitsInBlock;
                        if (MSU >= blockShift) {
                                // shift by blocks
                                for (unsigned i = 0; i <= MSU - blockShift; ++i) {
                                        _block[i] = _block[i + blockShift];
                                }
                        }
                        // adjust the shift
                        bitsToShift -= static_cast<int>(blockShift * bitsInBlock);
                        if (bitsToShift == 0) {
                                // fix up the leading zeros if we have a negative number
                                if (signext) {
                                        // bitsToShift is guaranteed to be less than nbits
                                        bitsToShift += static_cast<int>(blockShift * bitsInBlock);
                                        for (unsigned i = nbits - static_cast<unsigned>(bitsToShift); i < nbits; ++i) {
                                                this->setbit(i);
                                        }
                                }
                                else {
                                        // clean up the blocks we have shifted clean
                                        bitsToShift += static_cast<int>(blockShift * bitsInBlock);
                                        for (unsigned i = nbits - static_cast<unsigned>(bitsToShift); i < nbits; ++i) {
                                                this->setbit(i, false);
                                        }
                                }
                                return *this;
                        }
                }
                if constexpr (MSU > 0) {
                        bt mask = ALL_ONES;
                        // mask for the lower bits in the block that need to move to the lower word
                        mask >>= (bitsInBlock - static_cast<unsigned>(bitsToShift));
                        for (unsigned i = 0; i < MSU; ++i) {  // TODO: can this be improved? we should not have to work on the upper blocks in case we block shifted
                                _block[i] >>= bitsToShift;
                                // mix in the bits from the left
                                bt bits = bt(mask & _block[i + 1]);
                                _block[i] |= (bits << (bitsInBlock - static_cast<unsigned>(bitsToShift)));
                        }
                }
                _block[MSU] >>= bitsToShift;

                // fix up the leading zeros if we have a negative number
                if (signext) {
                        // bitsToShift is guaranteed to be less than nbits
                        bitsToShift += static_cast<int>(blockShift * bitsInBlock);
                        for (unsigned i = nbits - static_cast<unsigned>(bitsToShift); i < nbits; ++i) {
                                this->setbit(i);
                        }
                }
                else {
                        // clean up the blocks we have shifted clean
                        bitsToShift += static_cast<int>(blockShift * bitsInBlock);
                        for (unsigned i = nbits - static_cast<unsigned>(bitsToShift); i < nbits; ++i) {
                                this->setbit(i, false);
                        }
                }

                // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                _block[MSU] &= MSU_MASK;
                return *this;
        }


        ///////////////////////////////////////////////////////////////////
        ///                  modifiers

        constexpr void clear() noexcept {
                for (unsigned i = 0; i < nrBlocks; ++i) {
                        _block[i] = bt(0ull);
                }
        }
        constexpr void setzero() noexcept { clear(); }
        constexpr void set() noexcept { // set all bits to 1
                if constexpr (nrBlocks > 1) {
                        for (unsigned i = 0; i < nrBlocks - 1; ++i) {
                                _block[i] = ALL_ONES;
                        }
                }
                _block[MSU] = ALL_ONES & MSU_MASK;
        }
        constexpr void reset() noexcept { clear(); } // set all bits to 0
        constexpr void set(unsigned i) noexcept {        setbit(i, true); }
        constexpr void reset(unsigned i) noexcept { setbit(i, false); }
        constexpr void setbit(unsigned i, bool v = true) noexcept {
                if (i >= nbits) return;
                unsigned blockIndex = i / bitsInBlock;
                if (blockIndex < nrBlocks) {
                        // GCC -O3 produces false-positive -Warray-bounds when inlining
                        // setbit across blockbinary instantiations with different nbits
                        // (e.g., blockbinary<129> calling into blockbinary<128> context).
                        // The bounds checks above are correct but GCC's interprocedural
                        // analysis loses track of which instantiation's _block is accessed.
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Warray-bounds"
#if __GNUC__ >= 12
#pragma GCC diagnostic ignored "-Wstringop-overflow"
#endif
#endif
                        bt blockBits = _block[blockIndex];
                        bt null = bit_clear_mask<bt>(i, bitsInBlock);
                        bt bit = bt(v ? 1 : 0);
                        bt mask = bt(bit << (i % bitsInBlock));
                        _block[blockIndex] = bt((blockBits & null) | mask);
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic pop
#endif
                }
        }
        constexpr void setbits(uint64_t value) noexcept {
                if constexpr (1 == nrBlocks) {
                        _block[0] = value & storageMask;
                }
                else if constexpr (1 < nrBlocks) {
                        for (unsigned i = 0; i < nrBlocks; ++i) {
                                _block[i] = value & storageMask;
                                if constexpr (bitsInBlock < 64) {
                                        value >>= bitsInBlock;
                                }
                                else {
                                        value = 0;  // avoid shift overflow when bitsInBlock >= 64
                                }
                        }
                }
                _block[MSU] &= MSU_MASK; // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
        }
        constexpr void setblock(unsigned b, const bt& blockBits) noexcept {
                if (b < nrBlocks) _block[b] = blockBits; // nop if b is out of range
        }        
        constexpr blockbinary& flip() noexcept { // in-place one's complement
                for (unsigned i = 0; i < nrBlocks; ++i) {
                        _block[i] = bt(~_block[i]);
                }                
                _block[MSU] &= MSU_MASK; // assert precondition of properly nulled leading non-bits
                return *this;
        }
        /// <summary>
        /// in-place 2's complement
        /// </summary>
        /// <returns>2's complement of original</returns>
        constexpr blockbinary& twosComplement() noexcept {
                blockbinary plusOne(1);
                if constexpr (NumberType == BinaryNumberType::Signed) {
                        flip();
                }
                else {
                        static_assert(NumberType == BinaryNumberType::Signed, "calling in-place 2's complement on an unsigned blockbinary"); // should this be allowed?
                }
                return *this += plusOne;
        }

        // minimum positive value of the blockbinary configuration
        constexpr blockbinary& minpos() noexcept {
                // minpos = 0000....00001
                clear();
                setbit(0, true);
                return *this;
        }
        // maximum positive value of the blockbinary configuration
        constexpr blockbinary& maxpos() noexcept {
                if constexpr (NumberType == BinaryNumberType::Signed) {
                        // maxpos = 01111....1111
                        clear();
                        flip();
                        setbit(nbits - 1, false);
                }
                else {
                        // maxpos = 11111....1111
                        clear();
                        flip();
                }
                return *this;
        }
        // zero
        constexpr blockbinary& zero() noexcept {
                clear();
                return *this;
        }
        // minimum negative value of the blockbinary configuration
        constexpr blockbinary& minneg() noexcept {
                if constexpr (NumberType == BinaryNumberType::Signed) {
                        // minneg = 11111....11111
                        clear();
                        flip();
                }
                else {
                        // minneg = 00000....00000
                        clear();
                }
                return *this;
        }
        // maximum negative value of the blockbinary configuration
        constexpr blockbinary& maxneg() noexcept {
                if constexpr (NumberType == BinaryNumberType::Signed) {
                        // maxneg = 10000....0000
                        clear();
                        setbit(nbits - 1);
                }
                else {
                        // maxneg = 00000....00000
                        clear();
                }
                                
                return *this;
        }

        // selectors
        constexpr bool sign() const noexcept { return _block[MSU] & SIGN_BIT_MASK; }
        constexpr bool ispos() const noexcept { return !sign(); }
        constexpr bool isneg() const noexcept { return sign(); }
        constexpr bool iszero() const noexcept {
                for (unsigned i = 0; i < nrBlocks; ++i) if (_block[i] != 0) return false;
                return true;
        }
        constexpr bool isodd() const noexcept { return _block[0] & 0x1;        }
        constexpr bool iseven() const noexcept { return !isodd(); }

        constexpr bool all() const noexcept {
                if constexpr (nrBlocks > 1) for (unsigned i = 0; i < nrBlocks - 1; ++i) if (_block[i] != ALL_ONES) return false;
                if (_block[MSU] != MSU_MASK) return false;
                return true;
        }
        constexpr bool any() const noexcept {
                if constexpr (nrBlocks > 1) for (unsigned i = 0; i < nrBlocks - 1; ++i) if (_block[i] != 0) return true;
                if (_block[MSU] & MSU_MASK) return true;
                return false;
        }
        constexpr bool anyAfter(unsigned bitIndex) const noexcept {  // TODO: optimize for limbs
                unsigned limit = (bitIndex < nbits) ? bitIndex : nbits;
                for (unsigned i = 0; i < limit; ++i) if (test(i)) return true;
                return false;
        }

        constexpr bool none() const noexcept {
                if constexpr (nrBlocks > 1) for (unsigned i = 0; i < nrBlocks - 1; ++i) if (_block[i] != 0) return false;
                if (_block[MSU] & MSU_MASK) return false;
                return true;
        }
        constexpr unsigned count() const noexcept { // TODO: optimize for limbs
                unsigned nrOnes = 0;
                for (unsigned i = 0; i < nbits; ++i) {
                        if (test(i)) ++nrOnes;
                }
                return nrOnes;
        }
        constexpr bool test(unsigned bitIndex) const noexcept { return at(bitIndex); }
        constexpr bool at(unsigned bitIndex) const noexcept {
                if (bitIndex >= nbits) return false; // fail silently as no-op
                unsigned blockIndex = bitIndex / bitsInBlock;
                bt limb = _block[blockIndex];
                bt mask = bt(1ull << (bitIndex % bitsInBlock));
                return (limb & mask);
        }
        constexpr uint8_t nibble(unsigned n) const noexcept {
                uint8_t retval{ 0 };
                if (n < (1 + ((nbits - 1) >> 2))) {
                        bt word = _block[(n * 4) / bitsInBlock];
                        unsigned nibbleIndexInWord = n % (bitsInBlock >> 2);
                        bt mask = static_cast<bt>(bt(0x0F) << (nibbleIndexInWord*4));
                        bt nibblebits = static_cast<bt>(mask & word);
                        retval = static_cast<uint8_t>(nibblebits >> (nibbleIndexInWord*4));
                }
                else { // nop when nibble index out of bounds
                        retval = 0;
                }
                return retval;
        }
        constexpr bt block(unsigned b) const noexcept {
                if (b < nrBlocks) return _block[b]; 
                return bt(0); // return 0 when block index out of bounds
        }

        // copy a value over from one blockbinary to this blockbinary
        // blockbinary is a 2's complement encoding when Signed, so we sign-extend
        // by default for Signed; for Unsigned, the high bit is just data and must
        // be zero-extended on widening (clear() above already zeroed the storage).
        template<unsigned srcbits>
        constexpr blockbinary<nbits, bt, NumberType>& assign(const blockbinary<srcbits, bt, NumberType>& rhs) {
                clear();
                // since bt is the same, we can simply copy the blocks in
                unsigned minNrBlocks = (this->nrBlocks < rhs.nrBlocks) ? this->nrBlocks : rhs.nrBlocks;
                for (unsigned i = 0; i < minNrBlocks; ++i) {
                        _block[i] = rhs.block(i);
                }
                if constexpr (nbits > srcbits && NumberType == BinaryNumberType::Signed) {
                        if (rhs.sign()) {
                                for (unsigned i = srcbits; i < nbits; ++i) { // TODO: replace bit-oriented sequence with block
                                        setbit(i);
                                }
                        }
                }
                // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                _block[MSU] &= MSU_MASK;
                return *this;
        }

        // copy a value over from one blockbinary to this without sign-extending the value
        // blockbinary is a 2's complement encoding, so we sign-extend by default
        // for fraction/significent encodings, we need to turn off sign-extending.
        template<unsigned srcbits>
        constexpr blockbinary<nbits, bt, NumberType>& assignWithoutSignExtend(const blockbinary<srcbits, bt, NumberType>& rhs) {
                clear();
                // since bt is the same, we can simply copy the blocks in
                unsigned minNrBlocks = (this->nrBlocks < rhs.nrBlocks) ? this->nrBlocks : rhs.nrBlocks;
                for (unsigned i = 0; i < minNrBlocks; ++i) {
                        _block[i] = rhs.block(i);
                }
                // enforce precondition for fast comparison by properly nulling bits that are outside of nbits
                _block[MSU] &= MSU_MASK;
                return *this;
        }

        // return the position of the most significant bit, -1 if v == 0
        constexpr int msb() const noexcept {
                for (int i = int(MSU); i >= 0; --i) {
                        if (_block[i] != 0) {
                                bt mask = (bt(1u) << (bitsInBlock-1));
                                for (int j = bitsInBlock - 1; j >= 0; --j) {
                                        if (_block[i] & mask) {
                                                return i * static_cast<int>(bitsInBlock) + j;
                                        }
                                        mask >>= 1;
                                }
                        }
                }
                return -1; // no significant bit found, all bits are zero
        }
        // conversion to native types
        constexpr int64_t to_sll() const {
                constexpr unsigned sizeoflonglong = 8 * sizeof(long long);
                int64_t ll{ 0 };
                int64_t mask{ 1 };
                unsigned upper = (nbits < sizeoflonglong ? nbits : sizeoflonglong);
                for (unsigned i = 0; i < upper; ++i) {
                        ll |= at(i) ? mask : 0;
                        mask <<= 1;
                }
                if (sign() && upper < sizeoflonglong) { // sign extend
                        for (unsigned i = upper; i < sizeoflonglong; ++i) {
                                ll |= mask;
                                mask <<= 1;
                        }
                }
                return ll;
        }
        constexpr uint64_t to_ull() const {
                uint64_t ull{ 0 };
                uint64_t mask{ 1 };
                uint32_t msb = nbits < 64 ? nbits : 64;
                for (uint32_t i = 0; i < msb; ++i) {
                        ull |= at(i) ? mask : 0;
                        mask <<= 1;
                }
                return ull;
        }
        template<typename Real,
                typename = typename std::enable_if< std::is_floating_point<Real>::value, Real >::type>
        Real to_native() const {
                blockbinary tmp(*this);
                if (isneg()) tmp.twosComplement();
                Real v{ 0.0 }, base{ 1.0 };
                for (unsigned i = 0; i < nbits; ++i) {
                        if (tmp.test(i)) v += base;
                        base *= 2.0;
                }
                return (isneg() ? -v : v);
        }
        // determine the rounding mode: result needs to be rounded up if true
        constexpr bool roundingMode(unsigned targetLsb) const {
                bool lsb = at(targetLsb);
                bool guard = (targetLsb == 0 ? false : at(targetLsb - 1));
                bool round = (targetLsb > 1 ? at(targetLsb - 2) : false);
                bool sticky =(targetLsb < 3 ? false : any(targetLsb - 3));
                bool tie = guard && !round && !sticky;
                return (lsb && tie) || (guard && !tie);
        }
        constexpr bool any(unsigned msb) const {
                msb = (msb > nbits - 1 ? nbits - 1 : msb);
                unsigned topBlock = msb / bitsInBlock;
                bt mask = bt(ALL_ONES >> (bitsInBlock - 1 - (msb % bitsInBlock)));
                for (unsigned i = 0; i < topBlock; ++i) {
                        if (_block[i] > 0) return true;
                }
                // process the partial block
                if (_block[topBlock] & mask) return true;
                return false;
        }

protected:
        // HELPER methods
        // none

private:
        bt _block[nrBlocks];

        //////////////////////////////////////////////////////////////////////////////
        // friend functions

        // integer - integer logic comparisons
        template<unsigned N, typename B, BinaryNumberType T>
        friend constexpr bool operator==(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs);
        template<unsigned N, typename B, BinaryNumberType T>
        friend constexpr bool operator!=(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs);
        // the other logic operators are defined in terms of arithmetic terms

        template<unsigned N, typename B, BinaryNumberType T>
        friend std::ostream& operator<<(std::ostream& ostr, const blockbinary<N, B, T>& v);
};

// Generate a type tag for blockbinary
template<unsigned N, typename B, BinaryNumberType T>
std::string type_tag(const blockbinary<N, B, T>& = {}) {
        std::stringstream str;
        str << "blockbinary<"
                << std::setw(4) << N << ", "
                << typeid(B).name() << ", "
                << typeid(T).name() << '>';
        return str.str();
}

//////////////////////////////////////////////////////////////////////////////////
// logic operators

template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator==(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        for (unsigned i = 0; i < lhs.nrBlocks; ++i) {
                if (lhs._block[i] != rhs._block[i]) {
                        return false;
                }
        }
        return true;
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator!=(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        return !operator==(lhs, rhs);
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator<(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        if (lhs.ispos() && rhs.isneg()) return false; // need to filter out possible overflow conditions
        if (lhs.isneg() && rhs.ispos()) return true;  // need to filter out possible underflow conditions
        if (lhs == rhs) return false; // so the maxneg logic works
        blockbinary<N, B, T> mneg; maxneg<N, B>(mneg);
        if (rhs == mneg) return false; // special case: nothing is smaller than maximum negative
        blockbinary<N, B, T> diff = lhs - rhs;
        return diff.isneg();
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator<=(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        return (lhs < rhs || lhs == rhs);
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator>(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        return !(lhs <= rhs);
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr bool operator>=(const blockbinary<N, B, T>& lhs, const blockbinary<N, B, T>& rhs) {
        return !(lhs < rhs);
}
///////////////////////////////////////////////////////////////////////////////
// binary operators

template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<N, B, T> operator+(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N, B, T> c(a);
        return c += b;
}
template<unsigned N, typename B, BinaryNumberType T >
constexpr blockbinary<N, B, T> operator-(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N, B, T> c(a);
        return c -= b;
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr inline blockbinary<N, B, T> operator*(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N, B, T> c(a);
        return c *= b;
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr inline blockbinary<N, B, T> operator/(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N, B, T> c(a);
        return c /= b;
}
template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<N, B, T> operator%(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N, B, T> c(a);
        return c %= b;
}

template<unsigned N, typename B, BinaryNumberType T>
blockbinary<N, B, T> operator<<(const blockbinary<N, B, T>& a, long b) {
        blockbinary<N, B, T> c(a);
        return c <<= b;
}
template<unsigned N, typename B, BinaryNumberType T>
blockbinary<N, B, T> operator>>(const blockbinary<N, B, T>& a, long b) {
        blockbinary<N, B, T> c(a);
        return c >>= b;
}

// divide a by b and return both quotient and remainder
template<unsigned N, typename B, BinaryNumberType T>
constexpr quorem<N, B, T> longdivision(const blockbinary<N, B, T>& dividend, const blockbinary<N, B, T>& divisor) {
        static_assert(T == BinaryNumberType::Signed, "longdivision requires signed blockbinary types");
        using BlockBinary = blockbinary<N + 1, B, T>;
        quorem<N, B, T> result = { 0, 0, 0 };
        if (divisor.iszero()) {
                result.exceptionId = 1; // division by zero
                return result;
        }
        // generate the absolute values to do long division 
        // 2's complement special case -max requires an signed int that is 1 bit bigger to represent abs()
        bool a_sign = dividend.sign();
        bool b_sign = divisor.sign();
        bool result_negative = (a_sign ^ b_sign);
        // normalize both arguments to positive, which requires expansion by 1-bit to deal with maxneg
        BlockBinary a(dividend);
        BlockBinary b(divisor);
        if (a_sign) a.twosComplement();
        if (b_sign) b.twosComplement();

        if (a < b) { // optimization for integer numbers
                result.rem = dividend; // a % b = a when a / b = 0
                return result;         // a / b = 0 when b > a
        }
        // initialize the long division
        BlockBinary accumulator = a;
        // prepare the subtractand
        BlockBinary subtractand = b;
        int msb_b = b.msb();
        int msb_a = a.msb();
        int shift = msb_a - msb_b;
        subtractand <<= shift;
        // long division
        for (int i = shift; i >= 0; --i) {
                if (subtractand <= accumulator) {
                        accumulator -= subtractand;
                        result.quo.setbit(static_cast<unsigned>(i));
                }
                else {
                        result.quo.setbit(static_cast<unsigned>(i), false);
                }
                subtractand >>= 1;
        }
        if (result_negative) {  // take 2's complement
                result.quo.flip();
                result.quo += 1;
        }
        if (a_sign) {
                result.rem = -accumulator;
        }
        else {
                result.rem = accumulator;
        }
        return result;
}

///////////////////////////////////////////////////////////////////////////////
// specialty binary operators

// unrounded addition, returns a blockbinary that is of size nbits+1
template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<N + 1, B, T> uradd(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N + 1, B, T> result(a);
        return result += blockbinary<N + 1, B, T>(b);
}

// unrounded subtraction, returns a blockbinary that is of size nbits+1
template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<N + 1, B, T> ursub(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<N + 1, B, T> result(a);
        return result -= blockbinary<N + 1, B, T>(b);
}

#define TRACE_URMUL 0
// unrounded multiplication, returns a blockbinary that is of size 2*nbits
// using brute-force sign-extending of operands to yield correct sign-extended result for 2*nbits 2's complement.
template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<2*N, B, T> urmul(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        using BlockBinary = blockbinary<2 * N, B, T>;
        BlockBinary result(0);
        if (a.iszero() || b.iszero()) return result;

        // compute the result
        BlockBinary signextended_a(a);
        BlockBinary multiplicant(b);
#if TRACE_URMUL
        std::cout << "    " << to_binary(a) << " * " << to_binary(b) << std::endl;
        std::cout << std::setw(3) << 0 << ' ' << to_binary(multiplicant) << ' ' << to_binary(result) << std::endl;
#endif
        for (unsigned i = 0; i < 2*N; ++i) {
                if (signextended_a.at(i)) {
                        result += multiplicant;
                }
                multiplicant <<= 1;
#if TRACE_URMUL
                std::cout << std::setw(3) << i << ' ' << to_binary(multiplicant) << ' ' << to_binary(result) << std::endl;
#endif

        }
#if TRACE_URMUL
        std::cout << "fnl " << to_binary(result) << std::endl;
#endif
        //blockbinary<2 * nbits, bt> clipped(result);
        // since we used operator+=, which enforces the nulling of leading bits
        // we don't need to null here
        return result;
}

// unrounded multiplication, returns a blockbinary that is of size 2*nbits
// using nbits modulo arithmetic with final sign
template<unsigned N, typename B, BinaryNumberType T>
constexpr blockbinary<2 * N, B, T> urmul2(const blockbinary<N, B, T>& a, const blockbinary<N, B, T>& b) {
        blockbinary<2 * N, B, T> result(0);
        if (a.iszero() || b.iszero()) return result;

        // compute the result
        bool result_sign = a.sign() ^ b.sign();
        // normalize both arguments to positive in new size
        blockbinary<N + 1, B, T> a_new(a); // TODO optimize: now create a, create _a.bb, copy, destroy _a.bb_copy
        blockbinary<N + 1, B, T> b_new(b);
        if (a.sign()) a_new.twosComplement();
        if (b.sign()) b_new.twosComplement();
        blockbinary<2*N, B, T> multiplicant(b_new);

#if TRACE_URMUL
        std::cout << "    " << a_new << " * " << b_new << std::endl;
        std::cout << std::setw(3) << 0 << ' ' << multiplicant << ' ' << result << std::endl;
#endif
        for (unsigned i = 0; i < (N+1); ++i) {
                if (a_new.at(i)) {
                        result += multiplicant;  // if multiplicant is not the same size as result, the assignment will get sign-extended if the MSB is true, this is not correct because we are assuming unsigned binaries in this loop
                }
                multiplicant <<= 1;
#if TRACE_URMUL
                std::cout << std::setw(3) << i << ' ' << multiplicant << ' ' << result << std::endl;
#endif
        }
        if (result_sign) result.twosComplement();
#if TRACE_URMUL
        std::cout << "fnl " << result << std::endl;
#endif
        return result;
}

#define TRACE_DIV 0
// unrounded division, returns a blockbinary that is of size 2*nbits
template<unsigned nbits, unsigned roundingBits, typename B, BinaryNumberType T>
blockbinary<2 * nbits + roundingBits, B, T> urdiv(const blockbinary<nbits, B, T>& a, const blockbinary<nbits, B, T>& b) {
        if (b.iszero()) {
                // division by zero
                throw "urdiv divide by zero";
        }
        // generate the absolute values to do long division 
        // 2's complement special case -max requires an signed int that is 1 bit bigger to represent abs()
        bool a_sign = a.sign();
        bool b_sign = b.sign();
        bool result_negative = (a_sign ^ b_sign);

        // normalize both arguments to positive, which requires expansion by 1-bit to deal with maxneg
        blockbinary<nbits + 1, B, T> a_new(a); // TODO optimize: now create a, create _a.bb, copy, destroy _a.bb_copy
        blockbinary<nbits + 1, B, T> b_new(b);
#if TRACE_DIV
        std::cout << "a " << to_binary(a_new) << '\n';
        std::cout << "b " << to_binary(b_new) << '\n';
#endif
        if (a_sign) a_new.twosComplement();
        if (b_sign) b_new.twosComplement();
#if TRACE_DIV
        std::cout << "a " << to_binary(a_new) << '\n';
        std::cout << "b " << to_binary(b_new) << '\n';
#endif

        // initialize the long division
        blockbinary<2 * nbits + roundingBits + 1, B, T> decimator(a_new);
        blockbinary<2 * nbits + roundingBits + 1, B, T> subtractand(b_new); // prepare the subtractand
        blockbinary<2 * nbits + roundingBits + 1, B, T> result;

        constexpr unsigned msp = nbits + roundingBits; // msp = most significant position
        decimator <<= msp; // scale the decimator to the largest possible positive value

        int msb_b = subtractand.msb();
        int msb_a = decimator.msb();
        int shift = msb_a - msb_b;
        subtractand <<= shift;
        int offset = msb_a - static_cast<int>(msp);  // msb of the result
        int scale  = shift - static_cast<int>(msp);  // scale of the result quotient

#if TRACE_DIV
        std::cout << "  " << to_binary(decimator, true)   << " msp  : " << msp << '\n';
        std::cout << "- " << to_binary(subtractand, true) << " shift: " << shift << '\n';
#endif
        // long division
        for (int i = msb_a; i >= 0; --i) {

                if (subtractand <= decimator) {
                        decimator -= subtractand;
                        result.setbit(static_cast<unsigned>(i));
                }
                else {
                        result.setbit(static_cast<unsigned>(i), false);
                }
                subtractand >>= 1;

#if TRACE_DIV
                std::cout << "  " << to_binary(decimator, true) << "  current quotient: " << to_binary(result, true) << '\n';
                std::cout << "- " << to_binary(subtractand, true) << '\n';
#endif
        }
        result <<= (scale - offset);
#if TRACE_DIV
        std::cout << "  " << "scaled result: " << to_binary(result, true) << " scale : " << scale << " offset : " << offset << '\n';
#endif
        if (result_negative) result.twosComplement();
        return result;
}

//////////////////////////////////////////////////////////////////////////////
// conversions to string representations

// create a binary representation of the storage
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
std::string to_binary(const blockbinary<nbits, BlockType, NumberType>& number, bool nibbleMarker = false) {
        std::stringstream s;
        s << "0b";
        for (unsigned i = 0; i < nbits; ++i) {
                unsigned bitIndex = nbits - 1ull - i;
                s << (number.at(bitIndex) ? '1' : '0');
                if (bitIndex > 0 && (bitIndex % 4) == 0 && nibbleMarker) s << '\'';
        }
        return s.str();
}

// local helper to display the contents of a byte array
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
std::string to_hex(const blockbinary<nbits, BlockType, NumberType>& number, bool nibbleMarker = true) {
        static constexpr unsigned bitsInByte = 8;
        static constexpr unsigned bitsInBlock = sizeof(BlockType) * bitsInByte;
        char hexChar[16] = {
                '0', '1', '2', '3', '4', '5', '6', '7',
                '8', '9', 'A', 'B', 'C', 'D', 'E', 'F',
        };
        std::stringstream ss;
        ss << "0x" << std::hex;
        int nrNibbles = int(1 + ((nbits - 1) >> 2));
        for (int n = nrNibbles - 1; n >= 0; --n) {
                uint8_t nibble = number.nibble(static_cast<unsigned>(n));
                ss << hexChar[nibble];
                if (nibbleMarker && n > 0 && ((n * 4ll) % bitsInBlock) == 0) ss << '\'';
        }
        return ss.str();
}

// decimal string conversion
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
std::string to_decimal(const blockbinary<nbits, BlockType, NumberType>& number) {
        if (number.iszero()) return "0";

        std::string result;
        blockbinary<nbits, BlockType, NumberType> dividend(number);
        bool isNegative = false;

        // Handle negative numbers for signed types
        if constexpr (NumberType == BinaryNumberType::Signed) {
                if (dividend.isneg()) {
                        isNegative = true;
                        dividend.twosComplement(); // Convert to positive
                }
        }

        // Repeatedly divide by 10 and collect remainders
        blockbinary<nbits, BlockType, NumberType> ten(10);
        while (!dividend.iszero()) {
                if constexpr (nbits <= 64) {
                        // For smaller sizes, use native division to avoid complexity
                        uint64_t temp = dividend.to_ull();
                        uint64_t remainder = temp % 10;
                        result = char('0' + remainder) + result;
                        dividend = temp / 10;
                } else {
                        // For larger sizes, use blockbinary division operators
                        blockbinary<nbits, BlockType, NumberType> remainder = dividend % ten;
                        uint64_t digit = remainder.to_ull();
                        result = char('0' + digit) + result;
                        dividend /= ten;
                }
        }

        if (isNegative) {
                result = "-" + result;
        }

        return result;
}

// ostream operator
template<unsigned nbits, typename BlockType, BinaryNumberType NumberType>
std::ostream& operator<<(std::ostream& ostr, const blockbinary<nbits, BlockType, NumberType>& number) {
        ostr << to_decimal(number);
        return ostr;
}


}} // namespace sw::universal

stillwater-sc / universal / 30672136635

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous