mirrors/scylladb
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cql3: castas_fcts: do not rely on boost casting large multiprecision integers to floats behavior

Browse Source 
In [1] a bug casting large multiprecision integers to floats is documented (note that it
received two fixes, the most recent and relevant is [2]). Even with the fix, boost now
returns NaN instead of ±∞ as it did before [3].

Since we cannot rely on boost, detect the conditions that trigger the bug and return
the expected result.

The unit test is extended to cover large negative numbers.

Boost version behavior:
 - 1.78 - returns ±∞
 - 1.79 - terminates
 - 1.79 + fix - returns NaN

Fixes https://github.com/scylladb/scylladb/issues/18508

[1] https://github.com/boostorg/multiprecision/issues/553
[2] ea786494db
[3] https://github.com/boostorg/math/issues/1132

Closes scylladb/scylladb#18532
This commit is contained in:
Avi Kivity
2024-05-06 12:17:14 +03:00
committed by Nadav Har'El
parent 4639ca1bf5
commit 51d09e6a2a
2 changed files with 13 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
cql3/functions/castas_fcts.cc
View File

@@ -69,6 +69,16 @@ using bytes_opt = std::optional<bytes>;
template<typename ToType, typename FromType>
static data_value castas_fctn_simple(data_value from) {
auto val_from = value_cast<FromType>(from);
// Workaround for https://github.com/boostorg/multiprecision/issues/553 (the additional bug discovered post-closing)
if constexpr (std::is_floating_point_v<ToType> && std::is_same_v<FromType, utils::multiprecision_int>) {
static auto min = utils::multiprecision_int(std::numeric_limits<ToType>::lowest());
static auto max = utils::multiprecision_int(std::numeric_limits<ToType>::max());
if (val_from < min) {
return -std::numeric_limits<ToType>::infinity();
} else if (val_from > max) {
return std::numeric_limits<ToType>::infinity();
}
}
return static_cast<ToType>(val_from);
}

3
test/cql-pytest/test_cast_data.py
View File

@@ -55,8 +55,10 @@ def signed(number, bits):
# numbers.
def test_cast_from_large_varint(cql, table1):
p = unique_key_int()
p_negative = unique_key_int()
v = 32767456456456456456545678943512357658768763546575675
cql.execute(f'INSERT INTO {table1} (p, cVarint) VALUES ({p}, {v})')
cql.execute(f'INSERT INTO {table1} (p, cVarint) VALUES ({p_negative}, {-v})')
# We can read back the original number without a cast, or with a cast
# to the same type. The "decimal" type can also hold a varint and return
# the same number.
@@ -71,6 +73,7 @@ def test_cast_from_large_varint(cql, table1):
# Casting to a 32-bit floating point, which only supports numbers up
# to 1e38, results in infinity
assert [(math.inf,)] == list(cql.execute(f"SELECT CAST(cVarint AS float) FROM {table1} WHERE p={p}"))
assert [(-math.inf,)] == list(cql.execute(f"SELECT CAST(cVarint AS float) FROM {table1} WHERE p={p_negative}"))
# Casting to a 64-bit floating point, which supports the range of our
# given number (though not its full precision!) is allowed, and some
# precision is lost. Confusingly, Python's 64-bit floating point is