Moved away from using lazy_static for this one since every access to a
lazy_static reference adds a runtime check, Currently using a build.rs script to generate a const table instead.
Worked out the ARM decoding issues, seems to run fine now.
Though the dynamic decoding for ARM is broken now since I had to change things in Data Processing and MSR instructions
TODO use `const fn` when it becomes stable
Former-commit-id: ba09748ff74a403c7016adcbe0ca553b591f6855
Lookup tables are generally faster than matching in most architectures.
There is some trouble in decoding some arm data processing instructions
so this is non-default feature for now
* Crashes on armwrestler but many games seem to work
Former-commit-id: 3c06ea344ae0097947d8cd5bd05b1f882c7f743a
Previously I did not account for gamepak 32-bit accesses being two 16bit (NSEQ+SEQ or SEQ+SEQ) accesses,
This resulted in less cycles being accounted for, resulting in more wasted CPU cycles per frame and was a huge performance hit.
I re-implemented this with look-up-tables also to speed up the cycle
count routing `SysBus::get_cycles` and also account for 32bit gamepak memory access.
Former-commit-id: fe6a9a570c843d40e38971a2a36e6511df1b8894
Profiling with 'perf' I discovered that too much time was spent inside
Timer::update when the prescalar was 0 (frequency of 1 cycle).
This was due to the while loop I used. I optimized this loop out and
there is an appearent performance increase.
Former-commit-id: d228450625fad003f054e3161f6eeee888245cac
Also, get rid of unnecessary derive Copy.
Since the ArmInstruction (and Thumb) derived from Copy, the compiler
always saves them to the stack when they are passed by value to other
functions, thus making some unwanted performance overhead.
I removed the derive Copy, and also pass them as references. This
project has a lot of "derive Copy" everywhere, and I should take note if
this is happenning elsewhere aswell.
Former-commit-id: 2f94c6050fa26c5b777244bd26706d4e6e2f0dc9
- Wrap static globals with a mutex.
- Implement InputInterface
Testing this on a standalone Java Console Application (without
rendering) shows sufficiant performanse for now (AVG fps of 180 in the bios)
Former-commit-id: cfbb9abb7e91b04258c41fc20e8a22c16d797386
I should really pay more attention to the docs..
The previous fix (bb843ea5) was not enough as it caused vblank,hblank
and vcount-match flags to be overwritten when they really should have
just been ignored.
This caused Dragon Ball - Advance Adevnture to fail booting
Former-commit-id: 1810e32692a0abd79f2df385c7c822debf186c79
Initially I began implementing the gdb protocol on my own, but I then found a
nice work-in-progress crate on https://github.com/daniel5151/gdbstub
that suited my needs.
Former-commit-id: f77557cbbd8652c2ed05ac439efc1956d8e99729
For some reason, probably typing mistkae, I `or`ed the value being written to DISPSTAT,
causing bits to never actually turn off, affectivly making HBLANK/VBLANK
interrupts linger forever.
This probably has caused some nasty glitches that I don't even know of, but two confirm issues solved
by this are Mariokart and Pacman.
This commit fixes#14 and #13
Former-commit-id: 72ad6efcb9e8f6b582d9279c5daab2603d5a408e
Ran cargo-fix to automatically fix most of the build warnings,
Cleaned up dead code, and fix the rest manually
Former-commit-id: f35faba46b40eaf9c047efb8ab1e77ffa24d41b6
Move reload_pipeline calls inside instructions.
This commit yeilds yet another 5% performance improvment.
The next step is to move `advance_pc` into the instructions themselves
and save the `match result` per executed instruction
Former-commit-id: 42193ffc48fda9943665e6a74e873186627a0b4a
In preperation for later optimization in the CPU pipeline
implementation, this commit refactors the arm/thumb exec functions to return a
CpuAction (Whether to advance the program counter or to flush the
pipeline)
Currently, a lot of host cycles are wasted in the arm7tdmi pipeline
Refill1 & Refill2 states. Optimizing these steps out would make the CPU
a bit faster.
Former-commit-id: 9be7966eaad22cceeb443fcc5823bbd945284027
This commit removes the error handling (CpuResult<>) in order to reduce
overhead in the cpu implementation.
Also, some cleanup of warning messages.
Notice: this commit breaks '--feature debugger' for now
Former-commit-id: d4484047c3f5d509eff89cef7090aa88b07a8d17
