All but one TGI drivers didn't use IRQs. Especially when the TGI driver kernel was the only .interruptor this meant quite some unnecessary overhead because it pulled in the whole IRQ infrastructure.
The one driver using IRQs (the graphics driver for the 160x102x16 mode on the Lynx) now uses a library reference to set up a JMP to its IRQ handler.
All but one joystick drivers didn't use IRQs. Espsecially when the joystick driver kernel was the only .interruptor this meant quite some unnecessary overhead because it pulled in the whole IRQ infrastructure.
I was told that the one driver using IRQs (the DXS/HIT-4 Player joystick driver for the C64) can be reworked to not do it. Until this is done that driver is defunct.
As discussed in https://github.com/cc65/cc65/pull/452 after my premature merge the two functions in question don't work as expected.
Additionally I adjusted several style deviations in the pull request in question.
So far the joy_masks array allowed several joystick drivers for a single target to each have different joy_read return values. However this meant that every call to joy_read implied an additional joy_masks lookup to post-process the return value.
Given that almost all targets only come with a single joystick driver this seems an inappropriate overhead. Therefore now the target header files contain constants matching the return value of the joy_read of the joystick driver(s) on that target.
If there indeed are several joystick drivers for a single target they must agree on a common return value for joy_read. In some cases this was alredy the case as there's a "natural" return value for joy_read. However a few joystick drivers need to be adjusted. This may cause some overhead inside the driver. But that is for sure smaller than the overhead introduced by the joy_masks lookup before.
!!! ToDo !!!
The following three joystick drivers become broken with this commit and need to be adjusted:
- atrmj8.s
- c64-numpad.s
- vic20-stdjoy.s
About all CONIO functions offering a <...>xy variant call
popa
_gotoxy
By providing an internal gotoxy variant that starts with a popa all those CONIO function can be shortened by 3 bytes. As soon as program calls more than one CONIO function this means an overall code size reduction.
The BSS segment and the ONCE segment share the same start address. So they need to be placed in two different memory areas.
So far BSS was placed in the MAIN memory area and ONCE was placed in an additional memory area. Both memory areas were written to the output file. They just "happened" to be loadable and runnable at a stretch.
Now ONCE is placed in the MAIN memory area and BSS is placed in an additional memory area. Only MAIN is written to the output file. It becomes more obvious that BSS is "just" defined to share memory with ONCE.
The way we want to use the INITBSS segment - and especially the fact that it won't have the type bss on all ROM based targets - means that the name INITBSS is misleading. After all INIT is the best name from my perspective as it serves several purposes and therefore needs a rather generic name.
Unfortunately this means that the current INIT segment needs to be renamed too. Looking for a short (ideally 4 letter) name I came up with ONCE as it contains all code (and data) accessed only once during initialization.
Conceptually the INITBSS segment is not initialized in any way. Therefore it makes sense to not load it from disk. However the INIT segment has to be loaded from disk and therefore moved to its run location above the INITBSS segment. The necessary move routine increases runtime RAM usage :-(
Therefore we now "unnecessarily" load the INITBSS segment from disk too meaning that the INIT segment is loaded at its run location. Therefore there's no need for the move routine anymore.
After all we trade disk space for (runtime) RAM space - an easy decision ;-)
Notes:
- The code allowing to re-run a program without re-load present so far could not have worked as far as I can see as it only avoided to re-run the move routine but still tried to re-run the code in the INIT segment that was clobbered by zeroing the BSS. Therefore I removed the code in question altogether. I'm personally not into this "dirty re-run" but if someone wants to add an actually working solution I won't block that.
- INITBSS is intentionally not just merged with the DATA segment as ROM-based targets can't reuse the INIT segment for the BSS and therefore have no reason to place the INIT segment above INITBSS.
- Because ROM-based targets don't copy INITBSS from the ROM (like it is done with the DATA segment) all users of INITBSS _MUST_NOT_ presume INITBSS to be initialized with zeros!
A call to $FDA3 cannot be used because it re-enables the BASIC ROM. If a large program (such as Contiki's webbrowser80) has destructor code or data "behind" that ROM, then the program might crash when it tries to quit gracefully. Changing that code to set CIA2_PRA works well enough.
So far the INIT segment was run from the later heap+stack. Now the INIT segment is run from the later BSS. The background is that so far the INIT segment was pretty small (from $80 to $180 bytes). But upcoming changes will increase the INIT segment in certain scenarios up to ~ $1000 bytes. So programs with very limited heap+stack might just not been able to move the INIT segment to its run location. But moving the INIT segment to the later BSS allows it to occupy the later BSS+heap+stack.
In order to allow that the constructors are _NOT_ allowed anymore to access the BSS. Rather they must use the DATA segment or the new INITBSS segment. The latter isn't cleared at any point so the constructors may use it to expose values to the main program. However they must make sure to always write the values as they are not pre-initialized.
Made other changes that were recommended by Oliver.
* Changed its name from move_init to moveinit.
* Used self-modifying code in the subroutine.
* The INIT segment doesn't need to be optional (it's used by the start-up file).
When a program starts running, INIT is moved from one place to another place. Then, INIT's code is executed; and, the first place is re-used for variables. After the INIT code has finished, the second place can be re-used by the heap and the C stack. That means that initiation code and data won't waste any RAM space after they stop being needed.
Up to now static drivers were created via co65 from dynamic drivers. However there was an issue with that approach:
The dynamic drivers are "o65 simple files" which obligates that they start with the 'code' segment. However dynamic drivers need to start with the module header - which is written to. For dynamic drivers this isn't more than a conceptual issue because they are always contain a 'data' segment and may therefore only be loaded into writable memory.
However when dynamic drivers are converted to static drivers using co65 then that issue becomes a real problem as then the 'code' segment may end up in non-writable memory - and thus writing to the module header fails.
Instead of changing the way dynamic drivers work I opted to rather make static driver creation totally independent from dynamic drivers. This allows to place the module header in the 'data' segment (see 'module.mac').
This change was suppsed to fix the issue that the former JUMPTABLE is merked as 'ro' while it is actually written to in several scenarios. When drivers are converted using co65 and then compiled into ROMs the JUMPTABLE isn't copied to RAM and therefore the write operations in question fail.
However unfortunately I didn't succeed in changing that :-( Just setting the former JUMPTABLE to 'rw' broke the drivers. So I placed the DATA segment directly after the former JUMPTABLE segment. This made the drivers converted with co65 work again - obviously after changing libsrc/Makefile:235 from '--code-label' to '--data-label'. But the actual dynamic drivers still didn't work as the former JUMPTABLE wasn't placed as the beginning of the loaded file anymore. That effect could be changed by exchanging src/ld65/o65.c:1391 with src/ld65/o65.c:1394 but doing so broke the drivers again :-((
The Apple2 doesn't have sprites so the Apple2 mouse callbacks place a special character on the text screen to indicate the mouse position. In order to support the necessary character removing and redrawing the Apple2 mouse driver called the Apple2 mouse callbacks in an "unusual way". So far so (sort of) good.
However the upcoming Atari mouse driver aims to support both "sprite-type" mouse callbacks as well as "text-char-type" mouse callbacks. Therefore the interface between mouse drivers and callbacks needs to be extended to allow the mouse callbacks to hide their different types from the mouse driver.
The nature of this change can be seen best by looking at the Apple2 file modifications. The CBM drivers and callbacks (at least the current ones) don't benefit from this change.
