This reverts commit b0c4e6fd2f.
After investigation, this is related to builtin functions. Using
-fno-builtin is a less invasive way to solve the problem. However this
appears to be a bug [1]; in theory fat LTO objects should solve that. I
will handle this matter at the fxSDK level by adding -fno-builtin right
and experimenting with the bugfix in binutils 2.43.
[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114337
As of now (fxSDK 2.11), this flag is set globally by the fxSDK, so this
change doesn't have an immediate effect. However, I've been
experimenting with removing -ffreestanding at the global level to allow
applications to build with the hosted C++ library (which isn't fully
supported by has many supported features, like containers, that aren't
available in free-standing mode).
Without -ffreestanding, LTO makes weird decisions and ends up pruning
way too many symbols from libraries, leading to undefined references for
symbols provided by the standard library. Here is a minimal example.
---
% cat abort.c
void abort(void)
{
__builtin_unreachable();
}
% cat main.c
extern void abort(void);
int start(void)
{
abort();
return 0;
}
% cat sh-min.x
ENTRY(_start)
OUTPUT_FORMAT(elf32-sh)
SECTIONS {
. = 0x1000;
.text : { *(.text) }
.data : { *(.data) }
.bss : { *(.bss) }
}
% sh-elf-gcc -flto -nostdlib -ffreestanding -c abort.c -o abort.o
% sh-elf-gcc-ar rcs abort.a abort.o
% sh-elf-gcc -flto -nostdlib -c main.c -o main.o
% sh-elf-gcc -flto -nostdlib -save-temps main.o -o main -T ./sh-min.x abort.a -lgcc
ld: ./main.ltrans0.ltrans.o: in function `_start':
<artificial>:(.text+0xc): undefined reference to `_abort'
---
To solve the bug in this example, add -ffreestanding when making main.o.
I haven't been able to sufficiently dump/introspect intermediate files
to understand what's happening yet. It's also unclear whether the fix is
clean since LTO normally requires all files to be built with the same
settings, so adding -ffreestanding to gint but not the add-in seems
suspicious. There is, however, an inherent incompatibility in the
conjunction of (1) building kernels with -ffreestanding, (2) building
add-ins that use the C++ library thus require -fhosted, and (3) only
linking objects files built with the same options.
* Set -flto
* Use gcc-ar to build the archive so the LTO plugin is loaded and LTO
sections aren't stripped off the object files
* Slightly conservative KEEP()s in the linker script (not required but
I'm paranoid)
* Make the drivers __attribute__((externall_visible)) so they are
generated by the link-time back-end and not left in the archive
* Remove the unused isappli/optnum parameters of main, which have been
unused for years, are irrelevant on fx-CG and lead to a link-time
warning with LTO. I'll add APIs to access them later.
Features will come in slowly while I restructure for gint 3. With this
third big target for gint, the legacy aspects of gint 2's structure and
API are getting felt, so a major revision will be in order.
See the TODO file at this commit for info on what works and not.
- Define a draft of the video interface
- Implement that dragt for CG for a single mode
* Includes stub of brightness setting from disassembling 3.60
- Use the video interface to show visual feedback on GDB on CG
Using the video interface avoids directly linking into a driver, which
will serve modularity in gint 3.
Nothing interesting in this commit, just moving files around, making
sure every target compiles every file, and controlling with macros
instead.
Basic support for fxg3a target in CMakeLists.txt, cmake/FindGint.cmake,
giteapc.make and include/gint/config.h.in. The rest is forgettable.
* Honor dwindow settings immediately (avoids useless dline() calls)
* Bound to ymin/ymax instead of doing many useless cut computations
* Remove the need for floating-point operations and division
* Move logic around tracking transfers to asyncio.c.
* Add a "short buffer" holding 0-3 bytes between writes, so that the
driver performs only 4-byte writes in the FIFO and a short write in
the commit, if needed.
- This is partially due to me thinking at some point that degrading
writing size was impossible, but it might actually be possible by
writing to FIFO/FIFO+2 or FIFO/FIFO+1/FIFO+2/FIFO+3.
- In any case I think this new approach wins on performance.
* Get rid of unit_size since we now always use 4 bytes.
* Add a waiting function which is used in usb_close() (and once tested
should be used in world switches too).
* Eliminate some of the special cases for the DCP, though not all (in
particular I can't get the commit to rely on the BEMP interrupt yet,
nor can I properly clear PID to NAK when unbinding).
* Create a heap arena over the OS stack, large enough to hold two VRAMs
as was previously done, unless GINT_NO_OS_STACK is set at compile
time. (This replaces GINT_USER_VRAM.)
* Allocate a single VRAM in the heap at startup.
* Use double buffering by default as triple buffering is almost entirely
useless. dudpate() waits if both VRAMs are identical to prevent
corruption, but this can be bypassed with R61524 functions as usual.
This adds about 180 kB of heap data to any add-in using default
settings.
* Stop trying to be smart and generate repeats on the fly; this breaks
time consistency. Also if repeats are not handled in time this causes
infinite loops.
* Move rarely-used functions to external files, simplify stuff, get rid
of internal driver events; saves ~1 kB per add-in overall.
