The createpatch() function was using a fixed-size buffer. I've changed it
to be dynamically allocated. I saw that the RPN format used in patches is
slightly different from the one used internally in the assembler, so I
added a new member to the Expression struct to track the patch size.
I've also limited the RPN expression length to 1MB. I realized that the
patch RPN expression could potentially be longer than the internal RPN
expression, so the internal expression would need a limit smaller than
UINT32_MAX. I thought 1MB would be a reasonable limit.
When a macro arg appears in a symbol name, the contents are appended.
However, the contents of the macro arg were not being validated.
Any character, regardless of whether it was allowed in a symbol name,
would be appended. With this change, the contents of the macro arg
are now validated character by character. The symbol name is considered
to end at the last valid character. The remainder of the macro arg is
treated as though it followed the symbol name in the asm source code.
It seemed that the consensus in our discussions of signed integer
overflow, which invokes undefined behavior in C, was that integer
arithmetic should be two's complement and there should be no warning for
overflows. I have implemented that by converting values to unsigned types
when appropriate. These changes will mostly preserve existing behavior,
except for a few cases that were being handled incorrectly before.
The case of dividing INT_MIN by -1 previously resulted in a CPU
exception and program termination. Now, that case is detected and results
in a warning and a value of INT_MIN.
Similarly, INT_MIN % -1 would have resulted in a CPU exception. Since this
is a mathematically valid operation with a result of 0, it now simply
gives that result without a warning.
I noticed that in rpn.c, there were attempts in certain operation handlers
to validate the nVal members of the source expressions even when the
expressions may have been relocatable expressions with meaningless numbers
for the nVal member. This could have caused spurious errors/warnings, so I
made those handlers confirm that isReloc is false before validating nVal.
Also, integer constants that are too large now result in a warning. The
post-conversion values have not been changed, in order to preserve
backward compatibility.
The linker script now allows you to assign a section with the same
attributes as in the source.
To do this, I've removed a check from AssignSectionAddressAndBankByName
that would never be triggered, due to that condition being checked
before. Shouldn't this and IsSectionSameTypeBankAndAttrs be condensed
into a single function?
Currently, all symbols are assigned a filename and line when they're
first encountered and added to the internal hash table. This is often
not expected and leads to erroneous error messages.
This ensures that build breaks to any of the test projects don’t immediately cause rgbds tests to fail.
On clone, I’ve set it up to pull the commits since the day before the desired commit. Sadly, this will clone more recent commits that we’re not testing, but at least it ensures that the desired commit can be checked out. This is hopefully a good enough replacement for —depth=1.
Signed-off-by: Ben10do <Ben10do@users.noreply.github.com>
We no longer assume that the test repos don’t exist when we run run-tests.sh. This allows developers to choose to keep them, to allow them to run the tests more quickly.
- Add the test repos to the .gitignore.
- Check if the directory for each repo already exists, before trying to clone it.
- Do a `git pull` for each repo, to ensure that existing copies of repos are up-to-date.
Signed-off-by: Ben10do <Ben10do@users.noreply.github.com>
The tests are not exhaustive, there are some conditions that aren't
checked. The tests are based in the C standard rules about undefined
behaviour.
This is a compatibility break but, hopefully, all projects are using
sane values. If not, there is no guarantee that the projects will build
in any platform where RGBDS can be compiled, so it would be better to
fix them.
Even though, technically, the left shift of a negative value is always
undefined, some projects rely on its current behaviour. This is the
reason why this doesn't cause a fatal error.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
Small tests like the ones included in this repository are good to test
individual features, but it is also a good idea to test some real
projects.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
NULL error messages have been given a description.
Messages that weren't descriptive enough now also print the name of the
function that has failed.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
This ensures that the test scripts are correctly run with the Bourne shell, regardless of the (potentially more exotic) shell that is used to invoke the script.
rgblink option -w has been restored to its previous behaviour: make WRAM
a continous section instead of spliting it into WRAM0 and WRAMX.
To enable DMG mode, option -d has to be used instead. This option
automatically enables -w.
Update tests.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
The error message shouldn't specify the name of the binary, that's
supposed to be known by the caller.
Update test reference outputs.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
It compares the results of the operators with the expected result if
doing the same thing manually.
Signed-off-by: Antonio Niño Díaz <antonio_nd@outlook.com>
Check all sections when testing BANK().
Add scripts to verify the tests and update the reference if needed.
Signed-off-by: AntonioND <antonio_nd@outlook.com>
Due to recent changes, lots of tests generated a slightly different
error output.
- bank-noexist : This test doesn't generate any error output now as
unknown labels are left for the linker to resolve.
- null-in-macro : This test used to crash. Now, the parser verifies that
a MACRO ends in ENDM, generating an error message if not.
Signed-off-by: AntonioND <antonio_nd@outlook.com>
