The JIT was doing a bunch of additional work where it was saving and
restoring registers and then juggling the arguments back in to a stack
frame. All of this is nonsensical without the optimization where we
could call syscalls inline without a stack frame.
Instead remove this optimization entirely and behave like a "generic"
syscall path always. The Linux syscall handler now pulls the arguments
out of the CPU context directly and stores the result back in to RAX
directly as well.
This has knock-on effects where technically syscalls are
going to be slightly faster because no stack frame setup for the
arguments, but additionally we are going to be able to have syscalls be
proper serialization points where we can interrupt the syscall and
long-jump out without problems.
Bumps the DiskCache version again because it causes codegen to change.
Two added failure modes here. If the disk cache version has changed then
the json files must be updated to the new version to ensure correct
tracking.
Additional failure mode is that codegen actually changed but the disk
cache version hasn't. This is the expected common failure mode and we
need to do additional work before updating json results. This just means
incrementing the disk cache version before updating the instcountci
results. Have a fairly lengthy error message to showcase how much of an
impact this might have.
We actually never use this anymore, we instead always pass zero for
both, and then rely on the thread inheritance model or setting the
values manually. Now that we expose visibility of the
InternalThreadState to the frontend they just access it directly.
Just a smidge of cleanup, NFC.
Prevents invalidations being missed under the following circumstances:
Thread A JITs block A into the global codebuffer, adding the guest to host
mapping to its CodePages, thread A is then killed.
Thread B then performs SMC on block A. An exception will be triggered but
as CodePages was stored per-thread, and thread A is now killed when all
threads are iterated over by the frontend to perform invalidations it
will be missed.
The accumulator is introduced to handle the case where multiple threads
have the same code entry in their local caches but share the same codebuffer.
Consider a thread C in the above example that also has block A in its cache,
without an accumulator, when invalidating thread B the entrypoint of A is erased
from the shared guest to host map. So when C is invalidated, the local cache entry
for A is not removed since it was removed from CodePages when invalidating B.
Apparently I completely missed that cpuid, xgetbv, syscall,
l{u,}{div,rem} were failing to hit their optimized cases for inlining
and avoiding 128-bit software divide.
The divisions are a clear performance regression for 64-bit applications
since that is the only real way to do a 64-bit division on x86, I added
those specifically because it sped up games.
CPUID depends heavily on the game, since some games use that as a
serialization instruction fairly heavily.
XGETBV is trivial since it matches behaviour of CPUID (and is basically
an extension of it).
Syscall inlining can save a decent amount of time, again heavily depends
on game.
Adds multi-inst tests for all of these situations so that once it gets
fixed (Apparently broken once RCLSE got stripped out), we can see that
they keep working. Obviously tests couldn't have existed in instcountCI
before since we didn't support multi-instruction tests.
Due to how jit block tail padding is working, there's no real good way
to determine the true "implementation size" of an instruction without
the backend being aware of wanting to investigate it.
Trying to inject another instruction, or another IR operation actually
subtly changes codegen in a way that gives invalid results. The only
real way to get around this is to inject a known token in to the
instruction stream as we `ExitFunction`.
So inject a `udf #0x420f`, and change the scanning behaviour to find the
first one and cut everything else off afterwards.
This already scoops out some code in some game blocks that were
accidentally landing ExitFunction code in the json.
This also has been tested to work with #4528 with its InstCountCI
specific changes reverted.
This means we don't need to play subtle padding tricks in the JIT to get
the information we want in InstcountCI.
This changes the instcountCI code to consistently load test data in to
RIP 0x1'0000 so we don't have any spurious changes due to json changes.
This has been a minor annoyance where if a test was added, it had the
potential to shift the rest of the data in the tests. This now ensures
it is consistent.
Instead of keeping the vlaue as a string array in the MetaLayer, convert
the value to its final type once.
Improves performance in some hotpaths that were doing config based
string conversion in a relatively high frequency.
No functional change here.
- CoreRunningMode enum and variable wasn't used anymore.
- Code was moved to the frontend
- CustomCPUFactory wasn't used anymore
- All special signal handling and various features were moved to
TestHarnessRunner
- We also don't want to support actual custom CPU cores.
- TestHarnessRunner just runs as a host runner if compiled on an
x86-64 device if vixl sim isn't enabled now.
- Removes the Core config option entirely.
- Moves VDSOPointers struct to the frontend
- Every use of this lives in the Linux frontend instead now
The class constructor for ContextImpl::CPUID requires HostFeatures to be
available at construction time. Pass the host features struct directly
through during construction time instead, which cleans up the interface
slightly and fixes that issue.
This moves the CPU feature querying to the frontend. The primary purpose
here is for the wow64 frontend to not require linux-isms for querying
these features. This is required since non-Linux environments don't have
the "CPUID" feature for reading EL1 MSRs in EL0.
Wiring up the remaining wow64 registry querying is left for a future
exercise.
This also technically removes an xbyak requirement from FEXCore for when
building the x86 Test harness runner, but that doesn't really matter for
regular use cases.
Depending on where the assembly was getting loaded in to memory it was
causing slight code generation differences.
Map the entire file to the same fixed offset as our ASM tests to ensure
consistency and removing flakes in CI.
This is what we'll actually ship (I hope), so that's the config we want to
track long-term. It's also a lot more managable resulting asm.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
When code invalidation is happening we currently have the issue that a
thread can acquire the code invalidation mutex in the middle of
invalidation. This is due to us acquiring and releasing the mutex
between each thread's code invalidation.
We need to hold the mutex for the entire duration for all thread's code
invalidation.
This fixes a rare hang on proton startup and resolves a consistent hang
on Proton application shutdown.
This now puts us on par with FEX-2312.1 with hanging.
This does not fix a relatively rare hang on fork (which also existed with FEX-2312.1).
This also does not fix the issue that the intersection of our mutexes
between frontend and backend are very convoluted. In part of the work
that is going to fix the rare fork mutex hang will change more of this.
This is used for instcountci to ensure instruction counts don't change
when a compiler supports this feature or not. Always runtime disable
when running in instcountci.
CMake option from #3394 can still be useful so leaving that in place.
Lots going on here.
This moves OS thread object lifetime management and internal thread
state lifetime management to the frontend. This causes a bunch of thread
handling to move from the FEXCore Context to the frontend.
Looking at `FEXCore/include/FEXCore/Core/Context.h` really shows how
much of the API has moved to the frontend that FEXCore no longer needs
to manage. Primarily this makes FEXCore itself no longer need to care
about most of the management of the emulation state.
A large amount of the behaviour moved wholesale from Core.cpp to
LinuxEmulation's ThreadManager.cpp. Which this manages the lifetimes of
both the OS threads and the FEXCore thread state objects.
One feature lost was the instruction capability, but this was already
buggy and is going to be rewritten/fixed when gdbserver work continues.
Now that all of this management is moved to the frontend, the gdbserver
can start improving since it can start managing all thread state
directly.
Similar to #3284 but works around some of the bugs that one introduced.
This is the minimal amount of changes to move the ownership from FEXCore
to the frontend. Since the frontends don't yet have a full thread state
tracking, there is an opaque pointer that needs to be managed.
In the followup commits this will be changed to have the syscall handler
to be the thread object manager.
Instruction count CI has transformed the way we work on FEX… I love the system
and want to make it better. there’s one part of instruction count CI that isn’t
so lovable: the problematic “optimal” flag on instructions.
There are several issues with this flag, both philosophical and practical.
– it is tedious to update the optimal flag when making an implementation
optimal. The effect of that is discouraging people from making instructions,
optimal, or encouraging people to fail to update the flag, and dilute the value
of it. Either way, since we care far more about optimal implementations, then we
do about updating the flag, clearly we should prioritize the implementation and
not the flag. This issue was not obvious at the outset, when instruction count,
CI was introduced, and still quite small. The problem magnified when we started
duplicating instructions in bulk for different combinations of CPU features
(flagm, AFP, etc.) that intern multiplies the manual work required to update the
flags by the corresponding constant factor. if it comes down to a choice between
removing this extra coverage and removing the flag, I think we all agree that
removing the flag is the lesser evil.
– The definition of “optimal” is fundamentally problematic. I have often
improved the instruction count of an instruction that was already “optimal”.
This is all kinds of silly, and calls into question whether there’s any value
whatsoever in the existing classifications of the flag. Furthermore, it is often
unknowable, whether an implementation really is optimal. Is it possible to
implement BZHI (with flag calculations) in fewer than eight instructions? We
don’t know, and it’s silly to pretend that we do.
– as a consequence of the problematic definitions , there are so many errors in
both directions that I don’t think there’s much value in preserving the existing
classification at the expense of +progress. Being able to say “32% of
instructions are translated optimally” is neat, but it really doesn’t tell us
anything whatsoever when you dig a little deeper.
So, as the flag is misleading at best and perhaps harmful at worst, let’s remove
it and make the instruction count CI, more useful overall. let’s let the
expected count and the assembly speak for themselves, and cut away the chaff. if
we want a meaningless number to report to management, we can instead calculate
the average blowup factor ;-)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
There are some cases where we want to test multiple instructions where
we can do optimizations that would overwise be hard to see.
eg:
```asm
; Can be optimized to a single stp
push eax
push ebx
; Can remove half of the copy since we know the direction
cld
rep movsb
; Can remove a redundant insert
addss xmm0, xmm1
addss xmm0, xmm2
```
This lets us have arbitrary sized code in instruction count CI, with the
original json key becoming only a label if the instruction array is
provided.
There are still some major limitations to this, instructions that
generate side-effects might have "garbage" after the end of the block
that isn't correctly accounted for. So care must be taken.
Example in the json
```json
"push ax, bx": {
"ExpectedInstructionCount": 4,
"Optimal": "No",
"Comment": "0x50",
"x86Insts": [
"push ax",
"push bx"
],
"ExpectedArm64ASM": [
"uxth w20, w4",
"strh w20, [x8, #-2]!",
"uxth w20, w7",
"strh w20, [x8, #-2]!"
]
}
```
This allows us to use reciprocal instructions which matches precision of
what x86 expects rather than converting everything to float divides.
Currently no hardware supports this, and even the upcoming X4/A720/A520
won't support it, but it was trivial to implement so wire it up.
Enables in InstCountCI so Pi users can run InstCountCI can run the tests
without breaking on crypto operations.
When crypto is enabled or disabled just wholesale change AES, CRC32, and
PMULL 128-bit in one step. We don't really care about partial support
here.
Ryan Houdek