ARM64 branches have fairly small relative distances they can encode.
These can be +-1MB, or even +-32KB. The largest relative branch is
+-128MB, which we already set as an upper limit of our block JIT cache
size.
We have for a long time just compiled these without checking with the
expectation that things just happen to work. We didn't hit the asserts
so it was relatively low priority. Apparently now with Steam and a
MaxInst limit of 5000, we are now hitting an assert where we are
encoding too large of a range.
Implement support for long jumping from anywhere in the JIT for when a
long jump tries to be encoded and fails, allowing us to restart the JIT
at any moment. This is implemented as a long jump when this singular
feature could have gotten away with some sort of invasive check and
early exit path for two reasons. For one, that would be even more
invasive, effectively doing try-catch logic manually. And two, the next
step is supporting JIT buffer overflow for when our block size heuristic
fails.
This next step will mandate longjump on SIGSEGV (with cooperative
interaction with the frontend) from effectively /anywhere/ in the JIT.
One of the design goals of the CodeEmitter is that every code emission
function doesn't do a size remaining check to allow the compiler to do
some very effective optimization of emitting code blocks to memory (and
it works!).
But we lose the ability to sanely size check. When writing the emitter I
knew we were going to need to write this cooperative guard page handler,
and we're finally at a point where it needs to be done. This will be in
the next PR although.
Shared code buffer support introduced the concept of having a single
GuestToHostMaps shared across many threads. In the common case all
threads will share one however if e.g. a resize recently occured and
specific thread is yet to compile any code with the new codebuffer it
will still use the old GuestToHostMap. The current invalidation
approach handles this by repeatedly calling erase for every single
thread's GuestToHostMap, even if it is repeated. An accumulator is used
to ensure when two threads share a map, the L1/L2 cache entries in the
second thread will still be invalidated even if the the iteration for
the first thread removed them from the map.
Unfortunately this is incredibly slow in cases with many threads, as
a significant number of redundant map lookups and L1/L2 cache erasures
on threads that never even observed a given block can occur. Solve this
by introducing a two-pass model:
- First, all active codebuffers (and their associated GuestToHostMaps)
have their entries invalidated for the given range, these codebuffers
are tracked internally within FEXCore. It is at this point that delinking
callbacks are ran.
- Second, each thread will have its caches invalidated. But rather than
naively invalidating the L1/L2 caches for every invalidated block for
every thread, threads now track on their own what specific entries
have been potentially fetched into their L1/L2 caches. This is
aided by GuestToHostMap now tracking the pages each block touches. (an
inverse CodePages so to speak).
We currently rely on the frontend to keep track of threads and then
iterate over all threads to perform per-codebuffer operations. However
as codebuffers are shared between many threads (the common case is a
single code buffer across all) this ends up being inefficient. Introduce
a list of codebuffers to solve that (new codebuffers are very rare, so a
vector is plenty fine here for erasing invalid weak refs).
This was missed before, where the non-repeating strings instructions
were still using TSO even when the memcpy/set config option was
disabled. Make sure it listens to the config option and disable TSO in
those instances.
Noticed this while profiling Dishonored, and WINE's `sse2_memmove`
function was showing up as a high amount of CPU time. This is due to
them using non-repeating string operations on the header and tail of
their memmove to align to 16-byte.
With this fixed, it causes the game to go from ~62FPS to ~67FPS,
becoming bottlenecked by x87 emulation instead of memmove. Doing about
23 million soft-float operations per second, because it needs full
precision to remove some flickering artifacts.
- Cache miss counts
- Useful for determining if L2 cache or dynamic cache could help
- Cache read/write lock contention times
- Useful to see if threads are blocking each other on contention
- Read lock is the case where a read-lock is beneficial, even if we
currently use a write lock.
- JIT count
- Useful to see if any new JIT blocks are generating
On top of #4951 because it fiddles with the cache stuff.
With our flags being optimized, this does even less than when it was
introduced. It's a hack, people are tinkering with it thinking it'll do
something. Get rid of it.
L1 cache residency can get quite large. Solution, start out small and
scale quickly on L1 cache misses but L2/L3 cache hits.
Some stats on L1 cache residency change:
- Teardown: 40MB -> 16MB (40%)
- Ender Lilies: 79MB -> 32MB (40.5%)
- Death Stranding: 186MB -> 93MB (50%)
- Steam: 75MB -> 7MB (9.3%)
The cost of this option is effectively free in our JIT. It changes a
single LDR to be a single LDP, which on Cortex CPUs cost the same. We do
this by moving the L1 pointer mask in to the CPUState object, making it
dynamic so it lives next to the L1 pointer. We then use that directly
rather than having the hardcoded value.
The lookup cache does a little bit of additional tracking and heuristics
to determine when the current L1 cache should increase or decrease in
size. From 128KB to 16MB per thread, allocating the full VA range as
previously.
Once the heuristic determines that L1 should be increased, it simply
changes the max and the L1 pointer size to compensate, the kernel will
fault in whichever pages are necessary.
Decreasing the size is a little bit more complex, as we want to madvise
the resulting L1 range to ensure we don't have that memory as resident
anymore. Same heuristic but going in the opposite direction otherwise.
Tends to be the case that L1 cache increases a bit on loading screens
then backs down once in-game.
These heuristic values are exposed for increasing and decreasing because
while I think I've picked reasonable values, we will likely need some
more fine tuning over time. Kind of expert user toggles at that point.
Based on #4940 as a base which needs to be merged first.
Full tracked stats from steam as an example of where we are:
```
Total (1000 millisecond sample period):
JIT Time: 0.486630 ms/second (0.00 percent)
Signal Time: 0.065880 ms/second (0.00 percent)
SIGBUS Cnt: 38 (38.160780 per second)
SMC Cnt: 0
Softfloat Cnt: 0
FEX JIT Load: 0.004585 (cycles: 552510)
Total FEX Anon memory resident: 368 mB
JIT resident: 95 mB
OpDispatcher resident: 38 mB
Frontend resident: 8 mB
CPUBackend resident: 624 kB
Lookup cache resident: 0 (null)
Lookup L1 cache resident: 7 mB
ThreadStates resident: 460 kB
Unaccounted resident: 217 mB
```
This mode has been broken for a long time because it's mostly untested.
Barriers, and backpatching while slow have proven that they work.
Maintain the one TSO path, at least until all ARM hardware gains support for
x86-TSO memory model mode.
We want to take advantage of 16-byte single-copy atomicity. Which I am
relying on, but didn't codify it the first time.
Additionally add comments to explain that new members should be added to
the end to allow tools time to gain support gradually. This will allow
me to add new members without fully breaking mangohud, they'll just not
display the new information until support is added.
We're not guaranteeing backwards compatibility, just an attempt not to
constantly churn the format unless necessary. This way if we do break
compatibility, the tool will have an upper bound on supported versions
before needing to rewrite code.
Currently FEX doesn't properly support partial decoded instructions,
which behave slightly differently than full noexec or invalid
instruction decodings. Before this commit we didn't even have a way to
detect the difference.
Primary difference is that the faulting RIP is the beginning of
instruction decode, while the fault address is the first byte that
couldn't be fetched due to memory permissions. This shows up as a
difference between the RIP in mcontext and si_addr in siginfo in the
Linux signal handler.
Right now just change the log so we can determine if we need to support
this edge case.
This saves a whole bunch of memory. Cutting `Just Cause 2`'s title
screen from 1132MB anonymous FEX memory down to 438MB. 629MB in L2
alone.
L2 is primarily a means to reduce overhead in map queries, so it's all
about performance. But because it consumes a lot of people it's kind of
hard.
One idea is that the L2 lookups can be moved to shared data structures,
since we already pull the shared lock when doing an L2 lookup this is
already halfway there.
Side note, we're using unique locks even with read-only code paths
which we can't use the shared lock because this terrible recursive
mutex!
Instead of outright changing L2 behaviour and potentially wrecking
havoc, add a config option for now so testing can happen over time.
before:
```
Total FEX Anon memory resident: 1132 mB
JIT resident: 60 mB
OpDispatcher resident: 97 mB
Frontend resident: 37 mB
CPUBackend resident: 500 kB
Lookup cache resident: 629 mB
Lookup L1 cache resident: 108 mB
ThreadStates resident: 436 kB
```
after:
```
Total FEX Anon memory resident: 438 mB
JIT resident: 62 mB
OpDispatcher resident: 56 mB
Frontend resident: 22 mB
CPUBackend resident: 496 kB
Lookup cache resident: 0 (null)
Lookup L1 cache resident: 109 mB
ThreadStates resident: 436 kB
```
Every time I see this recursive mutex I glare at it. Remove the last one
so that we no longer need to deal with it.
The only reason why this recursive mutex still existed today was because
it is fairly intertwined with the ContextImpl and tracing it all was a
pain.
Peel back the layers and follow the idiom to have ContextImpl pull the
write mutex when requiredand pass it through by reference to ensure it stays alive.
This allows us to entirely give rid of the recursive nature of the
mutex, which means that `FindBlock` can eventually be switched over to a
read-lock to improve multiple threads reading the caches at the same
time.
I didn't do that exercise since that can be followed up in a subsequent
PR.
This captures the remaining FEX allocations that /aren't/ coming from
JEMalloc, allowing us to separate our mapped regions versus just
jemalloc allocations.
With some additional naming in jemalloc (which I'm not adding here) this
gets us interesting results:
```
Misc resident: 54 MiB
JEMalloc resident: 208 MiB
```
So 208MB of active jemalloc allocations in this particular case. These will be able to be tracked in heaptrack-like applications if careful.
This should let us target down whatever live allocations we're keeping
large amounts of data around if possible.
The frontend did a quirky widening check which was accidentally working
in this case, but it is supposed to be for the couple of GPR handling
AVX instructions.
Correct the implementation to use the correct register size for FMA.
Billy Laws