diff options
author | W. J. van der Laan <laanwj@protonmail.com> | 2021-05-26 09:41:49 +0200 |
---|---|---|
committer | W. J. van der Laan <laanwj@protonmail.com> | 2021-05-26 10:16:41 +0200 |
commit | 707ba8692b0013f4824dc3c2ea6554ccad5d20c2 (patch) | |
tree | 345a8c04ba50a587f49f6c20c02a395e57ed0227 /src | |
parent | 7aa41fc58196843bc021ab06a7d3f37189b4a871 (diff) | |
parent | 66545da2008cd9e806e41b74522ded259cd64f86 (diff) |
Merge bitcoin/bitcoin#21966: Remove double serialization; use software encoder for fee estimation
66545da2008cd9e806e41b74522ded259cd64f86 Remove support for double serialization (Pieter Wuille)
fff1cae43af959a601cf2558cb3c77f3c2b1aa80 Convert uses of double-serialization to {En,De}codeDouble (Pieter Wuille)
afd964d70b6f7583ecf89c380f80db07f5b66a60 Convert existing float encoding tests (Pieter Wuille)
bda33f98e2f32f2411fb0a8f5fb4f0a32abdf7d4 Add unit tests for serfloat module (Pieter Wuille)
2be4cd94f4c7d92a4287971233a20d68db81c9c9 Add platform-independent float encoder/decoder (Pieter Wuille)
e40224d0c77674348bf0a518365208bc118f39a4 Remove unused float serialization (MarcoFalke)
Pull request description:
Based on #21981.
This adds a software-based platform-independent float/double encoder/decoder (platform independent in the sense that it only uses arithmetic and library calls, but never inspects the binary representation). This should strengthen our guarantee that encoded float/double values are portable across platforms. It then removes the functionality to serialize doubles from serialize.h, and replaces its only (non-test) use for fee estimation data serialization with the software encoder.
At least on x86/ARM, the only difference should be how certain NaN values are encoded/decoded (but not *whether* they are NaN or not).
It comes with tests that verify on is_iec559 platforms (which are the only ones we support, at least for now) that the serialized bytes exactly match the binary representation of floats in memory (for non-NaN).
ACKs for top commit:
laanwj:
Code review re-ACK 66545da2008cd9e806e41b74522ded259cd64f86
practicalswift:
cr re-ACK 66545da2008cd9e806e41b74522ded259cd64f86
Tree-SHA512: 62ad9adc26e28707b2eb12a919feefd4fd10cf9032652dbb1ca1cc97638ac21de89e240858e80d293d5112685c623e58affa3d316a9783ff0e6d291977a141f5
Diffstat (limited to 'src')
-rw-r--r-- | src/Makefile.am | 2 | ||||
-rw-r--r-- | src/Makefile.test.include | 1 | ||||
-rw-r--r-- | src/compat/assumptions.h | 5 | ||||
-rw-r--r-- | src/policy/fees.cpp | 44 | ||||
-rw-r--r-- | src/serialize.h | 32 | ||||
-rw-r--r-- | src/test/fuzz/float.cpp | 35 | ||||
-rw-r--r-- | src/test/fuzz/util.h | 4 | ||||
-rw-r--r-- | src/test/serfloat_tests.cpp | 129 | ||||
-rw-r--r-- | src/test/serialize_tests.cpp | 86 | ||||
-rw-r--r-- | src/util/serfloat.cpp | 64 | ||||
-rw-r--r-- | src/util/serfloat.h | 16 |
11 files changed, 258 insertions, 160 deletions
diff --git a/src/Makefile.am b/src/Makefile.am index 770bb76226..2c25f04d08 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -253,6 +253,7 @@ BITCOIN_CORE_H = \ util/moneystr.h \ util/rbf.h \ util/readwritefile.h \ + util/serfloat.h \ util/settings.h \ util/sock.h \ util/spanparsing.h \ @@ -594,6 +595,7 @@ libbitcoin_util_a_SOURCES = \ util/settings.cpp \ util/thread.cpp \ util/threadnames.cpp \ + util/serfloat.cpp \ util/spanparsing.cpp \ util/strencodings.cpp \ util/string.cpp \ diff --git a/src/Makefile.test.include b/src/Makefile.test.include index dc79ea3125..e5f9c4cb52 100644 --- a/src/Makefile.test.include +++ b/src/Makefile.test.include @@ -121,6 +121,7 @@ BITCOIN_TESTS =\ test/script_tests.cpp \ test/script_standard_tests.cpp \ test/scriptnum_tests.cpp \ + test/serfloat_tests.cpp \ test/serialize_tests.cpp \ test/settings_tests.cpp \ test/sighash_tests.cpp \ diff --git a/src/compat/assumptions.h b/src/compat/assumptions.h index 5f50cde3ff..7a254c3b67 100644 --- a/src/compat/assumptions.h +++ b/src/compat/assumptions.h @@ -36,11 +36,6 @@ static_assert(std::numeric_limits<double>::is_iec559, "IEEE 754 double assumed") // Example(s): Everywhere :-) static_assert(std::numeric_limits<unsigned char>::digits == 8, "8-bit byte assumed"); -// Assumption: We assume floating-point widths. -// Example(s): Type punning in serialization code (ser_{float,double}_to_uint{32,64}). -static_assert(sizeof(float) == 4, "32-bit float assumed"); -static_assert(sizeof(double) == 8, "64-bit double assumed"); - // Assumption: We assume integer widths. // Example(s): GetSizeOfCompactSize and WriteCompactSize in the serialization // code. diff --git a/src/policy/fees.cpp b/src/policy/fees.cpp index 52c3362166..2ae5798ebe 100644 --- a/src/policy/fees.cpp +++ b/src/policy/fees.cpp @@ -10,6 +10,7 @@ #include <logging.h> #include <streams.h> #include <txmempool.h> +#include <util/serfloat.h> #include <util/system.h> static const char* FEE_ESTIMATES_FILENAME = "fee_estimates.dat"; @@ -26,6 +27,25 @@ std::string StringForFeeEstimateHorizon(FeeEstimateHorizon horizon) assert(false); } +namespace { + +struct EncodedDoubleFormatter +{ + template<typename Stream> void Ser(Stream &s, double v) + { + s << EncodeDouble(v); + } + + template<typename Stream> void Unser(Stream& s, double& v) + { + uint64_t encoded; + s >> encoded; + v = DecodeDouble(encoded); + } +}; + +} // namespace + /** * We will instantiate an instance of this class to track transactions that were * included in a block. We will lump transactions into a bucket according to their @@ -356,12 +376,12 @@ double TxConfirmStats::EstimateMedianVal(int confTarget, double sufficientTxVal, void TxConfirmStats::Write(CAutoFile& fileout) const { - fileout << decay; + fileout << Using<EncodedDoubleFormatter>(decay); fileout << scale; - fileout << m_feerate_avg; - fileout << txCtAvg; - fileout << confAvg; - fileout << failAvg; + fileout << Using<VectorFormatter<EncodedDoubleFormatter>>(m_feerate_avg); + fileout << Using<VectorFormatter<EncodedDoubleFormatter>>(txCtAvg); + fileout << Using<VectorFormatter<VectorFormatter<EncodedDoubleFormatter>>>(confAvg); + fileout << Using<VectorFormatter<VectorFormatter<EncodedDoubleFormatter>>>(failAvg); } void TxConfirmStats::Read(CAutoFile& filein, int nFileVersion, size_t numBuckets) @@ -372,7 +392,7 @@ void TxConfirmStats::Read(CAutoFile& filein, int nFileVersion, size_t numBuckets size_t maxConfirms, maxPeriods; // The current version will store the decay with each individual TxConfirmStats and also keep a scale factor - filein >> decay; + filein >> Using<EncodedDoubleFormatter>(decay); if (decay <= 0 || decay >= 1) { throw std::runtime_error("Corrupt estimates file. Decay must be between 0 and 1 (non-inclusive)"); } @@ -381,15 +401,15 @@ void TxConfirmStats::Read(CAutoFile& filein, int nFileVersion, size_t numBuckets throw std::runtime_error("Corrupt estimates file. Scale must be non-zero"); } - filein >> m_feerate_avg; + filein >> Using<VectorFormatter<EncodedDoubleFormatter>>(m_feerate_avg); if (m_feerate_avg.size() != numBuckets) { throw std::runtime_error("Corrupt estimates file. Mismatch in feerate average bucket count"); } - filein >> txCtAvg; + filein >> Using<VectorFormatter<EncodedDoubleFormatter>>(txCtAvg); if (txCtAvg.size() != numBuckets) { throw std::runtime_error("Corrupt estimates file. Mismatch in tx count bucket count"); } - filein >> confAvg; + filein >> Using<VectorFormatter<VectorFormatter<EncodedDoubleFormatter>>>(confAvg); maxPeriods = confAvg.size(); maxConfirms = scale * maxPeriods; @@ -402,7 +422,7 @@ void TxConfirmStats::Read(CAutoFile& filein, int nFileVersion, size_t numBuckets } } - filein >> failAvg; + filein >> Using<VectorFormatter<VectorFormatter<EncodedDoubleFormatter>>>(failAvg); if (maxPeriods != failAvg.size()) { throw std::runtime_error("Corrupt estimates file. Mismatch in confirms tracked for failures"); } @@ -884,7 +904,7 @@ bool CBlockPolicyEstimator::Write(CAutoFile& fileout) const else { fileout << historicalFirst << historicalBest; } - fileout << buckets; + fileout << Using<VectorFormatter<EncodedDoubleFormatter>>(buckets); feeStats->Write(fileout); shortStats->Write(fileout); longStats->Write(fileout); @@ -920,7 +940,7 @@ bool CBlockPolicyEstimator::Read(CAutoFile& filein) throw std::runtime_error("Corrupt estimates file. Historical block range for estimates is invalid"); } std::vector<double> fileBuckets; - filein >> fileBuckets; + filein >> Using<VectorFormatter<EncodedDoubleFormatter>>(fileBuckets); size_t numBuckets = fileBuckets.size(); if (numBuckets <= 1 || numBuckets > 1000) { throw std::runtime_error("Corrupt estimates file. Must have between 2 and 1000 feerate buckets"); diff --git a/src/serialize.h b/src/serialize.h index d9ca984f9c..5ef846b9e9 100644 --- a/src/serialize.h +++ b/src/serialize.h @@ -122,34 +122,6 @@ template<typename Stream> inline uint64_t ser_readdata64(Stream &s) s.read((char*)&obj, 8); return le64toh(obj); } -inline uint64_t ser_double_to_uint64(double x) -{ - uint64_t tmp; - std::memcpy(&tmp, &x, sizeof(x)); - static_assert(sizeof(tmp) == sizeof(x), "double and uint64_t assumed to have the same size"); - return tmp; -} -inline uint32_t ser_float_to_uint32(float x) -{ - uint32_t tmp; - std::memcpy(&tmp, &x, sizeof(x)); - static_assert(sizeof(tmp) == sizeof(x), "float and uint32_t assumed to have the same size"); - return tmp; -} -inline double ser_uint64_to_double(uint64_t y) -{ - double tmp; - std::memcpy(&tmp, &y, sizeof(y)); - static_assert(sizeof(tmp) == sizeof(y), "double and uint64_t assumed to have the same size"); - return tmp; -} -inline float ser_uint32_to_float(uint32_t y) -{ - float tmp; - std::memcpy(&tmp, &y, sizeof(y)); - static_assert(sizeof(tmp) == sizeof(y), "float and uint32_t assumed to have the same size"); - return tmp; -} ///////////////////////////////////////////////////////////////// @@ -234,8 +206,6 @@ template<typename Stream> inline void Serialize(Stream& s, int32_t a ) { ser_wri template<typename Stream> inline void Serialize(Stream& s, uint32_t a) { ser_writedata32(s, a); } template<typename Stream> inline void Serialize(Stream& s, int64_t a ) { ser_writedata64(s, a); } template<typename Stream> inline void Serialize(Stream& s, uint64_t a) { ser_writedata64(s, a); } -template<typename Stream> inline void Serialize(Stream& s, float a ) { ser_writedata32(s, ser_float_to_uint32(a)); } -template<typename Stream> inline void Serialize(Stream& s, double a ) { ser_writedata64(s, ser_double_to_uint64(a)); } template<typename Stream, int N> inline void Serialize(Stream& s, const char (&a)[N]) { s.write(a, N); } template<typename Stream, int N> inline void Serialize(Stream& s, const unsigned char (&a)[N]) { s.write(CharCast(a), N); } template<typename Stream> inline void Serialize(Stream& s, const Span<const unsigned char>& span) { s.write(CharCast(span.data()), span.size()); } @@ -252,8 +222,6 @@ template<typename Stream> inline void Unserialize(Stream& s, int32_t& a ) { a = template<typename Stream> inline void Unserialize(Stream& s, uint32_t& a) { a = ser_readdata32(s); } template<typename Stream> inline void Unserialize(Stream& s, int64_t& a ) { a = ser_readdata64(s); } template<typename Stream> inline void Unserialize(Stream& s, uint64_t& a) { a = ser_readdata64(s); } -template<typename Stream> inline void Unserialize(Stream& s, float& a ) { a = ser_uint32_to_float(ser_readdata32(s)); } -template<typename Stream> inline void Unserialize(Stream& s, double& a ) { a = ser_uint64_to_double(ser_readdata64(s)); } template<typename Stream, int N> inline void Unserialize(Stream& s, char (&a)[N]) { s.read(a, N); } template<typename Stream, int N> inline void Unserialize(Stream& s, unsigned char (&a)[N]) { s.read(CharCast(a), N); } template<typename Stream> inline void Unserialize(Stream& s, Span<unsigned char>& span) { s.read(CharCast(span.data()), span.size()); } diff --git a/src/test/fuzz/float.cpp b/src/test/fuzz/float.cpp index d18a87d177..adef66a3ee 100644 --- a/src/test/fuzz/float.cpp +++ b/src/test/fuzz/float.cpp @@ -3,14 +3,14 @@ // file COPYING or http://www.opensource.org/licenses/mit-license.php. #include <memusage.h> -#include <serialize.h> -#include <streams.h> #include <test/fuzz/FuzzedDataProvider.h> #include <test/fuzz/fuzz.h> +#include <util/serfloat.h> #include <version.h> #include <cassert> -#include <cstdint> +#include <cmath> +#include <limits> FUZZ_TARGET(float) { @@ -19,24 +19,17 @@ FUZZ_TARGET(float) { const double d = fuzzed_data_provider.ConsumeFloatingPoint<double>(); (void)memusage::DynamicUsage(d); - assert(ser_uint64_to_double(ser_double_to_uint64(d)) == d); - CDataStream stream(SER_NETWORK, INIT_PROTO_VERSION); - stream << d; - double d_deserialized; - stream >> d_deserialized; - assert(d == d_deserialized); - } - - { - const float f = fuzzed_data_provider.ConsumeFloatingPoint<float>(); - (void)memusage::DynamicUsage(f); - assert(ser_uint32_to_float(ser_float_to_uint32(f)) == f); - - CDataStream stream(SER_NETWORK, INIT_PROTO_VERSION); - stream << f; - float f_deserialized; - stream >> f_deserialized; - assert(f == f_deserialized); + uint64_t encoded = EncodeDouble(d); + if constexpr (std::numeric_limits<double>::is_iec559) { + if (!std::isnan(d)) { + uint64_t encoded_in_memory; + std::copy((const unsigned char*)&d, (const unsigned char*)(&d + 1), (unsigned char*)&encoded_in_memory); + assert(encoded_in_memory == encoded); + } + } + double d_deserialized = DecodeDouble(encoded); + assert(std::isnan(d) == std::isnan(d_deserialized)); + assert(std::isnan(d) || d == d_deserialized); } } diff --git a/src/test/fuzz/util.h b/src/test/fuzz/util.h index 86b203c6b5..36b1d5035c 100644 --- a/src/test/fuzz/util.h +++ b/src/test/fuzz/util.h @@ -513,8 +513,6 @@ void WriteToStream(FuzzedDataProvider& fuzzed_data_provider, Stream& stream) noe WRITE_TO_STREAM_CASE(uint32_t, fuzzed_data_provider.ConsumeIntegral<uint32_t>()), WRITE_TO_STREAM_CASE(int64_t, fuzzed_data_provider.ConsumeIntegral<int64_t>()), WRITE_TO_STREAM_CASE(uint64_t, fuzzed_data_provider.ConsumeIntegral<uint64_t>()), - WRITE_TO_STREAM_CASE(float, fuzzed_data_provider.ConsumeFloatingPoint<float>()), - WRITE_TO_STREAM_CASE(double, fuzzed_data_provider.ConsumeFloatingPoint<double>()), WRITE_TO_STREAM_CASE(std::string, fuzzed_data_provider.ConsumeRandomLengthString(32)), WRITE_TO_STREAM_CASE(std::vector<char>, ConsumeRandomLengthIntegralVector<char>(fuzzed_data_provider))); } catch (const std::ios_base::failure&) { @@ -545,8 +543,6 @@ void ReadFromStream(FuzzedDataProvider& fuzzed_data_provider, Stream& stream) no READ_FROM_STREAM_CASE(uint32_t), READ_FROM_STREAM_CASE(int64_t), READ_FROM_STREAM_CASE(uint64_t), - READ_FROM_STREAM_CASE(float), - READ_FROM_STREAM_CASE(double), READ_FROM_STREAM_CASE(std::string), READ_FROM_STREAM_CASE(std::vector<char>)); } catch (const std::ios_base::failure&) { diff --git a/src/test/serfloat_tests.cpp b/src/test/serfloat_tests.cpp new file mode 100644 index 0000000000..54e07b0f61 --- /dev/null +++ b/src/test/serfloat_tests.cpp @@ -0,0 +1,129 @@ +// Copyright (c) 2014-2020 The Bitcoin Core developers +// Distributed under the MIT software license, see the accompanying +// file COPYING or http://www.opensource.org/licenses/mit-license.php. + +#include <hash.h> +#include <test/util/setup_common.h> +#include <util/serfloat.h> +#include <serialize.h> +#include <streams.h> + +#include <boost/test/unit_test.hpp> + +#include <cmath> +#include <limits> + +BOOST_FIXTURE_TEST_SUITE(serfloat_tests, BasicTestingSetup) + +namespace { + +uint64_t TestDouble(double f) { + uint64_t i = EncodeDouble(f); + double f2 = DecodeDouble(i); + if (std::isnan(f)) { + // NaN is not guaranteed to round-trip exactly. + BOOST_CHECK(std::isnan(f2)); + } else { + // Everything else is. + BOOST_CHECK(!std::isnan(f2)); + uint64_t i2 = EncodeDouble(f2); + BOOST_CHECK_EQUAL(f, f2); + BOOST_CHECK_EQUAL(i, i2); + } + return i; +} + +} // namespace + +BOOST_AUTO_TEST_CASE(double_serfloat_tests) { + BOOST_CHECK_EQUAL(TestDouble(0.0), 0); + BOOST_CHECK_EQUAL(TestDouble(-0.0), 0x8000000000000000); + BOOST_CHECK_EQUAL(TestDouble(std::numeric_limits<double>::infinity()), 0x7ff0000000000000); + BOOST_CHECK_EQUAL(TestDouble(-std::numeric_limits<double>::infinity()), 0xfff0000000000000); + BOOST_CHECK_EQUAL(TestDouble(0.5), 0x3fe0000000000000ULL); + BOOST_CHECK_EQUAL(TestDouble(1.0), 0x3ff0000000000000ULL); + BOOST_CHECK_EQUAL(TestDouble(2.0), 0x4000000000000000ULL); + BOOST_CHECK_EQUAL(TestDouble(4.0), 0x4010000000000000ULL); + BOOST_CHECK_EQUAL(TestDouble(785.066650390625), 0x4088888880000000ULL); + + // Roundtrip test on IEC559-compatible systems + if (std::numeric_limits<double>::is_iec559) { + BOOST_CHECK_EQUAL(sizeof(double), 8); + BOOST_CHECK_EQUAL(sizeof(uint64_t), 8); + // Test extreme values + TestDouble(std::numeric_limits<double>::min()); + TestDouble(-std::numeric_limits<double>::min()); + TestDouble(std::numeric_limits<double>::max()); + TestDouble(-std::numeric_limits<double>::max()); + TestDouble(std::numeric_limits<double>::lowest()); + TestDouble(-std::numeric_limits<double>::lowest()); + TestDouble(std::numeric_limits<double>::quiet_NaN()); + TestDouble(-std::numeric_limits<double>::quiet_NaN()); + TestDouble(std::numeric_limits<double>::signaling_NaN()); + TestDouble(-std::numeric_limits<double>::signaling_NaN()); + TestDouble(std::numeric_limits<double>::denorm_min()); + TestDouble(-std::numeric_limits<double>::denorm_min()); + // Test exact encoding: on currently supported platforms, EncodeDouble + // should produce exactly the same as the in-memory representation for non-NaN. + for (int j = 0; j < 1000; ++j) { + // Iterate over 9 specific bits exhaustively; the others are chosen randomly. + // These specific bits are the sign bit, and the 2 top and bottom bits of + // exponent and mantissa in the IEEE754 binary64 format. + for (int x = 0; x < 512; ++x) { + uint64_t v = InsecureRandBits(64); + v &= ~(uint64_t{1} << 0); + if (x & 1) v |= (uint64_t{1} << 0); + v &= ~(uint64_t{1} << 1); + if (x & 2) v |= (uint64_t{1} << 1); + v &= ~(uint64_t{1} << 50); + if (x & 4) v |= (uint64_t{1} << 50); + v &= ~(uint64_t{1} << 51); + if (x & 8) v |= (uint64_t{1} << 51); + v &= ~(uint64_t{1} << 52); + if (x & 16) v |= (uint64_t{1} << 52); + v &= ~(uint64_t{1} << 53); + if (x & 32) v |= (uint64_t{1} << 53); + v &= ~(uint64_t{1} << 61); + if (x & 64) v |= (uint64_t{1} << 61); + v &= ~(uint64_t{1} << 62); + if (x & 128) v |= (uint64_t{1} << 62); + v &= ~(uint64_t{1} << 63); + if (x & 256) v |= (uint64_t{1} << 63); + double f; + memcpy(&f, &v, 8); + uint64_t v2 = TestDouble(f); + if (!std::isnan(f)) BOOST_CHECK_EQUAL(v, v2); + } + } + } +} + +/* +Python code to generate the below hashes: + + def reversed_hex(x): + return binascii.hexlify(''.join(reversed(x))) + def dsha256(x): + return hashlib.sha256(hashlib.sha256(x).digest()).digest() + + reversed_hex(dsha256(''.join(struct.pack('<d', x) for x in range(0,1000)))) == '43d0c82591953c4eafe114590d392676a01585d25b25d433557f0d7878b23f96' +*/ +BOOST_AUTO_TEST_CASE(doubles) +{ + CDataStream ss(SER_DISK, 0); + // encode + for (int i = 0; i < 1000; i++) { + ss << EncodeDouble(i); + } + BOOST_CHECK(Hash(ss) == uint256S("43d0c82591953c4eafe114590d392676a01585d25b25d433557f0d7878b23f96")); + + // decode + for (int i = 0; i < 1000; i++) { + uint64_t val; + ss >> val; + double j = DecodeDouble(val); + BOOST_CHECK_MESSAGE(i == j, "decoded:" << j << " expected:" << i); + } +} + +BOOST_AUTO_TEST_SUITE_END() diff --git a/src/test/serialize_tests.cpp b/src/test/serialize_tests.cpp index f77cda7ba2..58709178a4 100644 --- a/src/test/serialize_tests.cpp +++ b/src/test/serialize_tests.cpp @@ -70,8 +70,6 @@ BOOST_AUTO_TEST_CASE(sizes) BOOST_CHECK_EQUAL(sizeof(uint32_t), GetSerializeSize(uint32_t(0), 0)); BOOST_CHECK_EQUAL(sizeof(int64_t), GetSerializeSize(int64_t(0), 0)); BOOST_CHECK_EQUAL(sizeof(uint64_t), GetSerializeSize(uint64_t(0), 0)); - BOOST_CHECK_EQUAL(sizeof(float), GetSerializeSize(float(0), 0)); - BOOST_CHECK_EQUAL(sizeof(double), GetSerializeSize(double(0), 0)); // Bool is serialized as char BOOST_CHECK_EQUAL(sizeof(char), GetSerializeSize(bool(0), 0)); @@ -85,93 +83,9 @@ BOOST_AUTO_TEST_CASE(sizes) BOOST_CHECK_EQUAL(GetSerializeSize(uint32_t(0), 0), 4U); BOOST_CHECK_EQUAL(GetSerializeSize(int64_t(0), 0), 8U); BOOST_CHECK_EQUAL(GetSerializeSize(uint64_t(0), 0), 8U); - BOOST_CHECK_EQUAL(GetSerializeSize(float(0), 0), 4U); - BOOST_CHECK_EQUAL(GetSerializeSize(double(0), 0), 8U); BOOST_CHECK_EQUAL(GetSerializeSize(bool(0), 0), 1U); } -BOOST_AUTO_TEST_CASE(floats_conversion) -{ - // Choose values that map unambiguously to binary floating point to avoid - // rounding issues at the compiler side. - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x00000000), 0.0F); - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x3f000000), 0.5F); - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x3f800000), 1.0F); - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x40000000), 2.0F); - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x40800000), 4.0F); - BOOST_CHECK_EQUAL(ser_uint32_to_float(0x44444444), 785.066650390625F); - - BOOST_CHECK_EQUAL(ser_float_to_uint32(0.0F), 0x00000000U); - BOOST_CHECK_EQUAL(ser_float_to_uint32(0.5F), 0x3f000000U); - BOOST_CHECK_EQUAL(ser_float_to_uint32(1.0F), 0x3f800000U); - BOOST_CHECK_EQUAL(ser_float_to_uint32(2.0F), 0x40000000U); - BOOST_CHECK_EQUAL(ser_float_to_uint32(4.0F), 0x40800000U); - BOOST_CHECK_EQUAL(ser_float_to_uint32(785.066650390625F), 0x44444444U); -} - -BOOST_AUTO_TEST_CASE(doubles_conversion) -{ - // Choose values that map unambiguously to binary floating point to avoid - // rounding issues at the compiler side. - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x0000000000000000ULL), 0.0); - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x3fe0000000000000ULL), 0.5); - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x3ff0000000000000ULL), 1.0); - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x4000000000000000ULL), 2.0); - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x4010000000000000ULL), 4.0); - BOOST_CHECK_EQUAL(ser_uint64_to_double(0x4088888880000000ULL), 785.066650390625); - - BOOST_CHECK_EQUAL(ser_double_to_uint64(0.0), 0x0000000000000000ULL); - BOOST_CHECK_EQUAL(ser_double_to_uint64(0.5), 0x3fe0000000000000ULL); - BOOST_CHECK_EQUAL(ser_double_to_uint64(1.0), 0x3ff0000000000000ULL); - BOOST_CHECK_EQUAL(ser_double_to_uint64(2.0), 0x4000000000000000ULL); - BOOST_CHECK_EQUAL(ser_double_to_uint64(4.0), 0x4010000000000000ULL); - BOOST_CHECK_EQUAL(ser_double_to_uint64(785.066650390625), 0x4088888880000000ULL); -} -/* -Python code to generate the below hashes: - - def reversed_hex(x): - return binascii.hexlify(''.join(reversed(x))) - def dsha256(x): - return hashlib.sha256(hashlib.sha256(x).digest()).digest() - - reversed_hex(dsha256(''.join(struct.pack('<f', x) for x in range(0,1000)))) == '8e8b4cf3e4df8b332057e3e23af42ebc663b61e0495d5e7e32d85099d7f3fe0c' - reversed_hex(dsha256(''.join(struct.pack('<d', x) for x in range(0,1000)))) == '43d0c82591953c4eafe114590d392676a01585d25b25d433557f0d7878b23f96' -*/ -BOOST_AUTO_TEST_CASE(floats) -{ - CDataStream ss(SER_DISK, 0); - // encode - for (int i = 0; i < 1000; i++) { - ss << float(i); - } - BOOST_CHECK(Hash(ss) == uint256S("8e8b4cf3e4df8b332057e3e23af42ebc663b61e0495d5e7e32d85099d7f3fe0c")); - - // decode - for (int i = 0; i < 1000; i++) { - float j; - ss >> j; - BOOST_CHECK_MESSAGE(i == j, "decoded:" << j << " expected:" << i); - } -} - -BOOST_AUTO_TEST_CASE(doubles) -{ - CDataStream ss(SER_DISK, 0); - // encode - for (int i = 0; i < 1000; i++) { - ss << double(i); - } - BOOST_CHECK(Hash(ss) == uint256S("43d0c82591953c4eafe114590d392676a01585d25b25d433557f0d7878b23f96")); - - // decode - for (int i = 0; i < 1000; i++) { - double j; - ss >> j; - BOOST_CHECK_MESSAGE(i == j, "decoded:" << j << " expected:" << i); - } -} - BOOST_AUTO_TEST_CASE(varints) { // encode diff --git a/src/util/serfloat.cpp b/src/util/serfloat.cpp new file mode 100644 index 0000000000..8edca924cd --- /dev/null +++ b/src/util/serfloat.cpp @@ -0,0 +1,64 @@ +// Copyright (c) 2021 The Bitcoin Core developers +// Distributed under the MIT software license, see the accompanying +// file COPYING or http://www.opensource.org/licenses/mit-license.php. + +#include <util/serfloat.h> + +#include <cmath> +#include <limits> + +double DecodeDouble(uint64_t v) noexcept { + static constexpr double NANVAL = std::numeric_limits<double>::quiet_NaN(); + static constexpr double INFVAL = std::numeric_limits<double>::infinity(); + double sign = 1.0; + if (v & 0x8000000000000000) { + sign = -1.0; + v ^= 0x8000000000000000; + } + // Zero + if (v == 0) return copysign(0.0, sign); + // Infinity + if (v == 0x7ff0000000000000) return copysign(INFVAL, sign); + // Other numbers + int exp = (v & 0x7FF0000000000000) >> 52; + uint64_t man = v & 0xFFFFFFFFFFFFF; + if (exp == 2047) { + // NaN + return NANVAL; + } else if (exp == 0) { + // Subnormal + return copysign(ldexp((double)man, -1074), sign); + } else { + // Normal + return copysign(ldexp((double)(man + 0x10000000000000), -1075 + exp), sign); + } +} + +uint64_t EncodeDouble(double f) noexcept { + int cls = std::fpclassify(f); + uint64_t sign = 0; + if (copysign(1.0, f) == -1.0) { + f = -f; + sign = 0x8000000000000000; + } + // Zero + if (cls == FP_ZERO) return sign; + // Infinity + if (cls == FP_INFINITE) return sign | 0x7ff0000000000000; + // NaN + if (cls == FP_NAN) return 0x7ff8000000000000; + // Other numbers + int exp; + uint64_t man = std::round(std::frexp(f, &exp) * 9007199254740992.0); + if (exp < -1021) { + // Too small to represent, encode 0 + if (exp < -1084) return sign; + // Subnormal numbers + return sign | (man >> (-1021 - exp)); + } else { + // Too big to represent, encode infinity + if (exp > 1024) return sign | 0x7ff0000000000000; + // Normal numbers + return sign | (((uint64_t)(1022 + exp)) << 52) | (man & 0xFFFFFFFFFFFFF); + } +} diff --git a/src/util/serfloat.h b/src/util/serfloat.h new file mode 100644 index 0000000000..4d912b0176 --- /dev/null +++ b/src/util/serfloat.h @@ -0,0 +1,16 @@ +// Copyright (c) 2021 The Bitcoin Core developers +// Distributed under the MIT software license, see the accompanying +// file COPYING or http://www.opensource.org/licenses/mit-license.php. + +#ifndef BITCOIN_UTIL_SERFLOAT_H +#define BITCOIN_UTIL_SERFLOAT_H + +#include <stdint.h> + +/* Encode a double using the IEEE 754 binary64 format. All NaNs are encoded as x86/ARM's + * positive quiet NaN with payload 0. */ +uint64_t EncodeDouble(double f) noexcept; +/* Reverse operation of DecodeDouble. DecodeDouble(EncodeDouble(f))==f unless isnan(f). */ +double DecodeDouble(uint64_t v) noexcept; + +#endif // BITCOIN_UTIL_SERFLOAT_H |