Problem:
The crash happened in FORMAT(double, dec>=31, 'de_DE').
The patch for MDEV-23118 (commit 0041dacc1b)
did not take into account that String::set_real() has a limit of 31
(FLOATING_POINT_DECIMALS) fractional digits. So for the range of 31..38
digits, set_real() switches to use:
- my_fcvt() - decimal point notation, e.g. 1.9999999999
- my_gcvt() - scientific notation, e.g. 1e22
my_gcvt() returned a shorter string than Item_func_format::val_str_ascii()
expected to get after the my_fcvt() call, so it crashed on assert.
Solution:
We cannot extend set_real() to use the my_fcvt() mode for the range of
31..38 fractional digits, because set_real() is used in a lot of places
and such a change will break everything.
Introducing String::set_fcvt() which always prints using my_fcvt()
for the whole range of decimals 0..38, supported by the FORMAT() function.
FORMAT() can print more integer digits (than the argument has)
if rounding happens:
FORMAT(9.9,0) -> '10'
The old code did not take this into account.
Fix:
1. One extra digit is needed in case of rounding
- If args[1] is a not-NULL constant, then reserve space for one extra integer
digit if the requested number of decimals is less than args[0]->decimals.
- Otherwise, reserve space for one extra integer digit if
args[0]->decimals is not 0, because rounding can potentially happen
(depending on the exact data in arguments).
2. One extra digit is also needed if the argument has no integer digits,
e.g. in a data type like DECIMAL(38,38).
The conditions 1 and 2 are ORed.
3. Fixing FORMAT_MAX_DECIMALS from 30 to 38. This was forgotten in 10.2.1
(when the limit for the number of fractional digits in DECIMAL was extended).
The fix consists of three commits backported from 10.3:
1) Cleanup isnan() portability checks
(cherry picked from commit 7ffd7fe962)
2) Cleanup isinf() portability checks
Original problem reported by Wlad: re-compilation of 10.3 on top of 10.2
build would cache undefined HAVE_ISINF from 10.2, whereas it is expected
to be 1 in 10.3.
std::isinf() seem to be available on all supported platforms.
(cherry picked from commit bc469a0bdf)
3) Use std::isfinite in C++ code
This is addition to parent revision fixing build failures.
(cherry picked from commit 54999f4e75)
This change takes into account a column's GENERATED ALWAYS AS
expression dependcy on sql_mode's PAD_CHAR_TO_FULL_LENGTH and
NO_UNSIGNED_SUBTRACTION flags.
Indexed virtual columns as well as persistent generated columns are
now not allowed to have such dependencies to avoid inconsistent data
or index files on sql_mode changes.
So an error is now returned in cases like this:
CREATE OR REPLACE TABLE t1
(
a CHAR(5),
v VARCHAR(5) AS (a) PERSISTENT -- CHAR->VARCHAR or CHAR->TEXT = ERROR
);
Functions RPAD() and RTRIM() can now remove dependency on
PAD_CHAR_TO_FULL_LENGTH. So this can be used instead:
CREATE OR REPLACE TABLE t1
(
a CHAR(5),
v VARCHAR(5) AS (RTRIM(a)) PERSISTENT
);
Note, unlike CHAR->VARCHAR and CHAR->TEXT this still works,
not RPAD(a) is needed:
CREATE OR REPLACE TABLE t1
(
a CHAR(5),
v CHAR(5) AS (a) PERSISTENT -- CHAR->CHAR is OK
);
More sql_mode flags may affect values of generated columns.
They will be addressed separately.
See comments in sql_mode.h for implementation details.
cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Debug
Maintainer mode makes all warnings errors. This patch fix warnings. Mostly about
deprecated `register` keyword.
Too much warnings came from Mroonga and I gave up on it.
The error message modified.
Then the TABLE_SHARE::error_table_name() implementation taken from 10.3,
to be used as a name of the table in this message.
The problem happened because {{Field_xxx::store(longlong nr, bool unsigned_val)}} erroneously passed {{unsigned_flag}} to the {{usec}} parameter of this constructor:
{code:cpp}
Datetime(int *warn, longlong sec, ulong usec, date_conv_mode_t flags)
{code}
1. Changing Time and Datetime constructors to accept data as Sec6 rather than as
longlong/double/my_decimal, so it's not possible to do such mistakes
in the future. Additional good effect of these changes:
- This reduced some amount of similar code (minus ~35 lines).
- The code now does not rely on the fact that "unsigned_flag" is
not important inside Datetime().
The constructor always gets all three parts: sign, integer part,
fractional part. The simple the better.
2. Fixing Field_xxx::store() to use the new Datetime constructor format.
This change actually fixes the problem.
3. Adding "explicit" keyword to all Sec6 constructors,
to avoid automatic hidden conversion from double/my_decimal to Sec6,
as well as from longlong/ulonglong through double to Sec6.
4. Change#1 caused (as a dependency) changes in a few places
with code like this:
bool neg= nr < 0 && !unsigned_val;
ulonglong value= m_neg ? (ulonglong) -nr : (ulonglong) nr;
These fragments relied on a non-standard behavior with
the operator "minus" applied to the lowest possible negative
signed long long value. This can lead to different results
depending on the platform and compilation flags.
We have fixed such bugs a few times already.
So instead of modifying the old wrong code to a new wrong code,
replacing all such fragments to use Longlong_hybrid,
which correctly handles this special case with -LONGLONG_MIN
in its method abs().
This also reduced the amount of similar code
(1 or 2 new lines instead 3 old lines in all 6 such fragments).
5. Removing ErrConvInteger(longlong nr, bool unsigned_flag= false)
and adding ErrConvInteger(Longlong_hybrid) instead, to encourage
use of safe Longlong_hybrid instead of unsafe pairs nr+neg.
6. Removing unused ErrConvInteger from Item_cache_temporal::get_date()
This warning come from a copy() operation of type:
memcpy(ptr, ptr+A, B), which is safe but produces a warning
when run with valgrind.
To avoid the warning, I added copy_or_move() method which uses
memmove() instead of memcpy().
In 10.3 the change in item_strfunc::Item_func_concat() has to be mirroed
in Item_func_concat_oracle() to avoid future valgrind warnings.
After the MDEV-13118 fix there's no code in the server that
wants caseup/casedn to change the argument in place for simple
charsets. Let's remove this logic and always return the result in a
new string for all charsets, both simple and complex.
1. Removing the optimization that *some* character sets used in casedn()
and caseup(), which allowed (and required) to change the case in-place,
overwriting the string passed as the "src" argument.
Now all CHARSET_INFO's work in the same way:
non of them change the source string in-place, all of them now convert
case from the source string to the destination string, leaving
the source string untouched.
2. Adding "const" qualifier to the "char *src" parameter
to caseup() and casedn().
3. Removing duplicate implementations in ctype-mb.c.
Now both caseup() and casedn() implementations for all CJK character sets
use internally the same function my_casefold_mb()
(the former my_casefold_mb_varlen()).
4. Removing the "unused" attribute from parameters of some my_case{up|dn}_xxx()
implementations, as the affected parameters are now *used* in the code.
Previously these parameters were used only in DBUG_ASSERT().
