The various places that were using the CodeBuffer object were using
internal implementation details that are changing as we move over to a
bitmap allocator.
Preempt this by hiding some of the implementation details early without
changing behaviour. `GetBufferBase` is still technically leaking some of
the internal details, but it needs changes around how relocations are
being handled and how the disk cache validation works in order to handle
that right now.
Should be no functional change.
NFC
- Renames CodeBufferManager to SharedCodeBufferManager to be more
explicit about it being shared between threads
- Renames `CodeBuffers` to `SharedCodeBuffers` to make it more explicit
about sharing these buffers between threads.
- Separates the Manager to its own file so it is distinct from the rest
of the CPUBackend code
Makes it easier to parse ownership and lifetime semantics of these
buffers.
This both take a void*, so the casting is unnecessary to begin with,
since this would occur anyway without it. We can also avoid a
duplication to reduce line noise.
using <sys/prctl.h> and <linux/prctl.h> simultaneously causes clang to
fail:
```
In file included from FEX/FEXCore/Source/Utils/AllocatorHooks.cpp:6:
/usr/include/sys/prctl.h:88:8: error: redefinition of 'prctl_mm_map'
88 | struct prctl_mm_map {
| ^
/usr/include/linux/prctl.h:134:8: note: previous definition is here
134 | struct prctl_mm_map {
| ^
1 error generated.
```
prefer <sys/prctl.h> and do not include <linux/prctl.h>
fix: #5454
Signed-off-by: Pepper Gray <hello@peppergray.xyz>
Disables THP on some key locations that are fairly sparse
- rpmalloc
- This is the big one as this allocates some heavy sparse buffers.
- CallRet stacks
- These get in the hundreds of megabytes, while not being sparse they
trend towards only using a handful of pages and ballooning to 2MB
per thread is quite heavy.
- Lookup cache
- L1 specifically gets hit here which adds a decent chunk of overhead
due to sparsity.
Win32 for all of these also aren't handled, but that will need to be a
followup.
If/When this works, the amount of iTLB misses drop dramatically, which
reduces L2 TLB pressure, which just improves performance for our CPU
cores that have itty-bitty L1 iTLB entry counts.
We can't use `mmap(MAP_HUGETLB)` directly for terrible reasons, so we
are required to lean on madvise instead.
The current CodeBuffer regrowth code discards any existing contents. This is
undesirable with code caching, since those contents can't be re-fetched from
the disk cache and instead need to be recompiled at runtime.
Additionally, FEXOfflineCompiler obviously should never discard compiled code.
Using a large enough CodeBuffer right away reduces the likelihood that FEX
runs into such scenarios.
There's no longer a distinction between AArch64 and x86 and everything
effectively falls under "Common" now. This means flattening the entire
structure just cleans it up.
NFC. (Although instcountCI will update because of a couple pointer
offsets changing)
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 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.
Will allow external tools to track how much memory FEX allocates.
Necessary since we can't use traditional memory usage tools to track FEX
memory allocation independently of guest allocations. Plus most tools
like heaptrack hook allocation symbols, which break under jemalloc.
Using this information I can see with Steam loaded with my library that
FEX consumes ~825MB. Total process resident memory is 1204M, accounting
for around 379MB being used by steam itself. This is /relatively/ close
to my desktop running steam at around 261MB. There's a bit of variance
due to what Steam chooses to do at startup.
This tracking will be the first step towards seeing where our memory
usage is going.
This is changes the interface of CodeBuffer to that of a partially persistent
data structure based on reference counting:
- Exactly one CodeBuffer is now designated as "active", which means data can
be *appended* to it
- Lossy modifications to the active CodeBuffer will not invalidate any data
in use by other threads, which enables save sharing across threads
- Instead, such lossy modifications trigger a new "version" of the data in
the modifying thread. Old versions of the CodeBuffer persist as read-only
data for use by the other threads.
- The other threads can update their version of the CodeBuffer. This will
decrease the reference count and eventually trigger deallocation of the
old version
protects last page of codebuffer. This should cause a
SIGSEGV if we try to access it. Until now it was possible to go over
and access out of bounds.
In addition, there a couple of clang-tidy fixes which should be NFC.
Telemetry value address generation was forcing an indirection at all
times which was unnecessary. These values live in the BSS, zero
initialized at process start and is unnecessary.
Instead change the wrapper defines to directly operate on the enum
passed in which saves an indirection on all of these telemetry
operations (except for the ones in the JIT which are required to be PIC
compliant).
This also fixes an annoying warning about
`FEXCORE_TELEMETRY_STATIC_INIT` causing initialization and destruction
order being unspecified, so two wins.
```diff
"sha1rnds4 xmm0, xmm1, 10b": {
- "ExpectedInstructionCount": 55,
+ "ExpectedInstructionCount": 10,
```
So I spent a few hours glaring at this instruction. Then spent a few
more glaring in to the sunset and then found the optimization.
It's faster to load the f80 sign mask from our named vector constants
than synthesizing the values. Changes a 4 instruction sequence to
synthesize to be 1 load.
This is no longer necessary to be part of the public API. Moves the
header internally.
Needed to pass through `IsAddressInCodeBuffer` from CPUBackend through
the Context object, but otherwise no functional change.
I was looking at some other JIT overheads and this cropped up as some
overhead. Instead of materializing a constant using mov+movk+movk+movk,
load it from the named vector constant array.
In a micro-benchmark this improved performance by 34%.
In bytemark this improved on subbench by 0.82%
This will result in FEX not being able to allocate executable memory.
We can use shared memory in the future to work around this but for now
we don't support that as a fix.
Requires #3238 to be merged first since this uses the tbx IR operation.
Worst case is now a three instruction sequence of ldr+ldr+tbx.
Some operations are special-cased, which definitely doesn't cover all
possible cases we could use without tbx, but as a worst case improvement
this is a significant improvement.
A bunch of blendps swizzles weren't optimal. This optimizes all swizzles
to be optimal.
Two instructions can be more optimal without a tbx but the rest required
tbx to be optimal since they don't match ARM's swizzle mechanics.
The motivation towards just having a pointer array in CpuState was that
initialization was fairly cheap and that we have limited space inside
the encoding depending on what we want to do.
Initialization cost is still a concern but doing a memcpy of 128-bytes
isn't that big of a deal.
Limited space in CpuState, while a concern isn't a significant one.
- Needs to currently be less than 1 page in size
- Needs to be under the architectural offset limitations of loadstore
scaled offsets. Which is 65KB for 128-bit vectors
Still keeps the pointer array around for cases when we would need
synthesize an address offset and it's just easier to load the
process-wide table.
The performance improvement here is removing the dependency in the
ldr+ldr chain. In microbenchmarks this has shown to have an improvement
of ~4% by removing this dependency chain on Cortex-X1C.
In the case that source registers are sequential then this turns in to a
load of the vector constant (2 instructions) and the single tbl
instruction.
If the registers aren't sequential then the tbl turns in to 2 moves and
then the single tbl, which with zero-cycle rename isn't too bad.
Since this is a worst case option this is significantly better than the
previous implementation doing a bunch of inserts which was always 9
instructions.
We should still strive to implement faster versions without the use of
TBL2 if possible but this makes it less of a concern.
We can load the swizzle table from our constant pool now. This removes
the only usage of VTMP3 from our Arm64 JIT.
I would say the this is now optimal for the version without RCON set.
With RCON we could technically make some of the move of the constant
more optimal.
This now improves the instruction implementation from 17 instructions
down to 5 or 6 depending on if the host supports SVE.
I would say this is now optimal.
Use a named constant for loading the sign inversion, then EOR the second
source and just FAdd it all.
In a vacuum it isn't a significant improvement, but as soon as more than
one instruction is in a block it will eventually get optimized with
named constant caching and be a significant win.
Thanks to @rygorous for the idea!
VRev32 matches Arm64 semantics directly.
LoadNamedVectorConstant allows FEX to quickly load "named constants".
This will allow us to have specific hardcoded vector constant values
that we can load with a ldr(State)+ldr(Value) and will be more abused in
the future.
This also allows us to do a very simple optimization in the future where
we can optimize away redundant loads of these loads if they are used
multiple times in the same block. (Not implemented here).
Due to Intel dropping support for legacy segment registers[1] there is a
concern that this will break legacy 32-bit software that is doing some
magic segment register handling.
Adds some simple telemetry for 32-bit applications that when they
encounter an instruction that sets the segment register or uses a
segment register that the JIT will do a /relatively/ quick four
instruction check to see if it is not a null segment.
It's not enough to just check if the segment index is 0 or not, 32-bit
Linux software starts with non-zero segment register indexes but the LDT
for each segment index is a null-descriptor.
Once the segment address is loaded, the IR operation will do a quick
check against zero and if it /isn't/ zero then set the telemetry value.
A very minor optimization that segment registers only get checked once
per block to ensure overhead stays low.
[1] https://www.intel.com/content/www/us/en/developer/articles/technical/envisioning-future-simplified-architecture.html
- 3.6 - Restricted Subset of Segmentation
- `Bases are supported for FS, GS, GDT, IDT, LDT, and TSS
registers; the base for CS, DS, ES, and SS is ignored for 32-bit
mode, same as 64-bit mode (treated as zero).`
- 4.2.17 - MOV to Segment Register
- Will fault if SS is written (Breaking anything that writes to
SS).
- Will not fault if CS, DS, ES are written (Thus it sets the
segment but gets ignored due to 3.6).
It is not an external component, and it makes paths needlessly long.
Ryan seemed amenable to this when we discussed on IRC earlier.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Ryan Houdek