All ram pages containing code are write protected. In that case, no need
for block checks. Memory reads in the same block(s) can also be executed
at compile time and the results propagated as constants.
When a write is detected in a protected area, the corresponding blocks
are discarded and recompiled using traditional (slow) block checks.
Backported the blkmap code finding change from upstream.
Use smart pointers for block management to avoid reference issues.
Added WriteAfterWrite ssa pass
Fixed crash in ssa ConstProp pass when op list is modified
native implementation of negc and xtrct for x64 and arm64
rec-arm64: pass exception pc in w27 instead of sh4 context
inline mmu_intruction_translation() and clean up
Tested: Both with and without the feature, works only for x64 CPUs for
now, but supported in both windows and linux (see vmem implementation
for it, using mem-mapped files).
Linked them both toghether since you can't really define one and
not the other (plus Linux honors one windows the other in some
cases).
More refactoring on this area to follow.
detect frequent SMC check failures and use a specific code cache area
for these blocks.
flush the temp area when full but keep the main code cache area
generates direct vmem read & write accesses
trap sigsegv and rewrite using slow path
add w29 to allocatable registers
get rid of literals and use pc-rel branching
minor optimizations and cleanup
This adds support for separate config and data dirs.
On Linux, these will be compliant XDG Basedir Specification, i.e.
XDG_CONFIG_HOME and XDG_CONFIG_DIRS (or XDG_DATA_HOME and XDG_DATA_DIRS
respectively). On all other platforms, there currently just set to the
homedir path (so no previous behaviour has been changed).
If reicast wants to read and write a data file, it just calls
get_data_path("/samplefile.txt"). If it does not need to write to
that file, it just uses get_data_path("/samplefile.txt", false). That
way, we can also use system-wide dirs (like /usr/share/reicast on
linux), that the user usually doesn't have write access to.
The same applies for config file, where you use get_config_path(args)
respectively.
This consolidates some of the work done for TARGET_NO_NVMEM and
feat/no-direct-memmap. If nvmem is disabled at compile time or alloc
fails _nvmem_enabled() will return false. Various other fixes
and cleanups all around.
This works, but is extremelly hacky. Must be started without attached debugger, lldb doesn't want to let go of EXC_BAD_ADDRESS, but reicast really depends on it getting delivered as SIGSEGV/SIGBUS. Also xcode has a really bad day upon seeing the jit code. Oh well.
There's some dynarec bug that causes color corruption on bios logo/boot triagles, TA crash on ikaruga and infinitive loop on crazy taxi. I'd guess some fp-memory-write thingy, abi, or smth. Too bad.
- Force code to compile in arm mode (arm jit -> thumb mem functions is complicated)
- SIGILL, SIGBUS. Works w/o Mach exceptions and EXC_BAD_ADDRESS
- Code buffers move to __TEXT, munmapped && memmapped to actually work
- Primitive input. Button + start, or left (works to get out of bios date screen)
- Fixup emitter for thumb2/interworking (didn't work though, reverted to arm cc)
- Block Manager: Disable mem saving / page fault alloc-on-demand logic
- Move cycle counter to r11, r9 is not clean on iOS. Remove r11 from reg alloc list
- Cache flushes for iOS
- log to log.txt
- load game.chd
Flyinghead