file .\dtoa.c

The assertion failure was correct because the 'width' argument
of my_gcvt() has the signed integer type, whereas the unsigned
value UINT_MAX32 was being passed by the caller
(Field_double::val_str()) leading to a negative width in
my_gcvt().

The following chain of problems was found by further analysis:

1. The display width for a floating point number is calculated
in Field_double::val_str() as either field_length or the
maximum possible length of string representation of a floating
point number, whichever is greater. Since in the bug's test
case field_length is UINT_MAX32, we get the same value as the
display width. This does not make any sense because for numeric
values field_length only matters for ZEROFILL columns,
otherwise it does not make sense to allocate that much memory
just to print a number. Field_float::val_str() has a similar
problem.

2. Even if the above wasn't the case, we would still get a
crash on a slightly different test case when trying to allocate
UINT_MAX32 bytes with String::alloc() because the latter does
not handle such large input values correctly due to alignment
overflows.

3. Even when String::alloc() is fixed to return an error when
an alignment overflow occurs, there is still a problem because
almost no callers check its return value, and
Field_double::val_str() is not an exception (same for
Field_float::val_str()).

4. Even if all of the above wasn't the case, creating a
Field_double object with UINT_MAX32 as its field_length does
not make much sense either, since the .frm code limits it to
MAX_FIELD_CHARLENGTH (255) bytes. Such a beast can only be
created by create_tmp_field_from_item() from an Item with
REAL_RESULT as its result_type() and UINT_MAX32 as its
max_length.

5. For the bug's test case, the above condition (REAL_RESULT
Item with max_length = UINT_MAX32) was a result of
Item_func_if::fix_length_and_dec() "shortcutting" aggregation
of argument types when one of the arguments was a constant
NULL. In this case, the attributes of the aggregated type were
simply copied from the other, non-NULL argument, but max_length
was still calculated as per the general, non-shortcut case, by
choosing the greatest of argument's max_length, which is
obviously not correct.

The patch addresses all of the above problems, even though
fixing the assertion failure for the particular test case would
require only a subset of the above problems to be solved.