I ran the test suite using 32-bit PHP. There were 2 places where changes
were needed due to 32-bit timestamps.
Reader\\Xml.php was using strtotime as an intermediate step in converting
a string timestamp to an Excel timestamp. The XML file type stores pure timestamps
(i.e. no date portion) as, e.g., 1899-12-31T02:30:00.000, and that value
causes an error using strtotime on a 32-bit system. However, it is sufficient
to use that value in a DateTime constructor, and that will work for 32- and 64-bit.
There was no test for that particular cell, so I added one to the XML read test.
And that's when I discovered the getFormattedValue bug. The cell's format
is `hh":"mm":"ss`. The quotes around the colons are disrupting the formatting.
PhpSpreadsheet formats the cell by converting the Excel format
to a Php Date format, in this case `H\:m\:s`.
That's a problem,
since Excel thinks 'm' means *minutes*, but PHP thinks it means *months*.
This is not a problem when the colon is not quoted; there are ample tests for that.
I added my best guess as to how to recognize this situation,
changing `\:m` to `:i`. The XML read test
now succeeds, and no other tests were broken by this change.
Test Shared\\DateTest had one test where the expected result of converting to a
Unix timestamp exceeds 2**32. Since a Unix timestamp is strictly an int,
that test fails on a 32-bit system. In the discussion regarding recently merged
PR #1870, it was felt that the user base might still be using the functions
that convert to and from a timestamp. So, we should not drop this test, but,
since it cannot succeed on a 32-bit system, I changed it to be skipped
whenever the expected result exceeded PHP_INT_MAX. There are 3 "toTimestamp"
functions within that test. Only one of these had been affected, but I thought
it was a good idea to add additional tests to the others to demonstrate this
condition.
In the course of testing, I also discovered some 32-bit problems with
bitwise and base-conversion functions. I am preparing separate PRs to
deal with those.
* 100% Coverage for Calculation/DateTime
The code in DateTime is now completely covered.
Along the way, some errors were discovered and corrected.
- The tests which have had to be changed at the start of every year are
replaced by more robust equivalents which do not require annual changes.
- Several places in the code where Gnumeric and OpenOffice were thought to differ
from Excel do not appear to have had any justification.
I have left a comment where such code has been removed.
- Use DateTime when possible rather than date, time, or strftime functions to avoid
potential Y2038 problems.
- Some impossible code has been removed, replaced by an explanatory comment.
- NETWORKDAYS had a bug when the start date was Sunday. There had been no tests
of this condition.
- Some functions allow boolean and null arguments where a number is expected.
This is more complicated than the equivalent situations in MathTrig because
the initial date for these calculations can be Day 1 rather than Day 0.
- More testing for dates from 1900-01-01 through the fictitious
everywhere-but-Excel 1900-01-29.
- This showed that there is an additional Excel bug - Excel evaluates
WEEKNUM(emptycell) as 0, which is not a valid result for
WEEKNUM without a second argument.
PhpSpreadsheet now duplicates this bug.
- There is a similar and even worse bug for 1904-01-01 in 1904 calculations.
Weeknum returns 0 for this,
but returns the correct value for arguments of 0 or null.
- DATEVALUE should accept 1900-02-29 (sigh) and relatives.
PhpSpreadsheet now duplicates this bug.
- Testing bootstrap sets default timezone. This appears to be a relic from
the releases of PHP where the unwise decision, subsequenly reversed,
was made to issue messages for
"no default timezone is set" rather than just use a sensible default.
This was a disruptive setting for some of the tests I added.
There is only one test in the entire suite which is default-timezone-dependent.
Setting and resetting of default timezone is moved to that test
(Reader/ODS/ODSTest), and out of bootstrap.
- There had been no testing of NOW() function.
- DATEVALUE test had no tests for 1904 calendar and needs some.
- DATE test changed 1900/1904 calendar in use without restoring it.
- WEEKDAY test had no tests for 1904 calendar and needs some.
- Which revealed a bug in Shared/Date (excelToDateTimeObject was not
recognizing 1904-01-01 as valid when 1904 calendar is in use).
- And an additional bug in that legal 1904-calendar values in the 0.0-1.0
range yielded the same "wrong" answers as 1900-calendar (see "One note" below).
Also the comment for one of the calendar-1904 tests was wrong in attempting
to identify what time of day the fraction represented.
I had wanted to break this up into a set of smaller modules, a process already
started for Engineering and MathTrig.
However the number of source code changes was sufficient that I wanted
a clean delta for this request.
If it is merged, I will work on breaking it up afterwards.
One note - Shared/Date/excelToDateTimeObject, when calendar-1900 is in use,
returns an unexpected result if its argument is between 0 and 1,
which is nominally invalid for that calendar.
It uses a base-1970 calendar in that instance. That check is not justifiable
for calendar-1904, where values in that range are legal,
so I made the check specific to calendar-1900,
and adjusted 3 1904 unit test results accordingly. However, I have to admit that
I don't understand why that check should be made even for calendar-1900.
It certainly doesn't match anything that Excel does.
I would recommend scrapping that code altogether.
If agreed, I would do this as part of the break-up into smaller modules.
Another note -
more controversially, it is clear that PhpSpreadsheet needs to support
the Excel and PHP date formats. Although it requires further study,
I am not convinced that it needs to support Unix timestamp format.
Since that is a potential source of Y2038 problems on 32-bit systems,
I would like to open a PR to deprecate the use of that format.
Please let me know if you are aware of a valid reason to continue to support it.
oleibman
Adrien Crivelli