These IR operations are required to support AFP's NEP mode which does
vector insert in to the destination register. Additionally it gives us
tracking information to allow optimizing out redundant inserts on
devices that don't support AFP natively.
In order to match x86 semantics we need to support binary and unary
scalar operations that do a final insert in to a vector. With optional
zeroing of the top 128-bits for AVX variants.
A tricky thing is that in binary operations this means that the
destination and first source have an intrinsically linked property
depending on if it is SSE or AVX.
SSE example:
- addss xmm0, xmm1
- xmm0 is both the destination and the first source.
- This means xmm0[31:0] = xmm0[31:0] + xmm1[31:0]
- Bits [127:32] are UNMODIFIED.
FEX's JIT jumps through some hoops so that if the destination register
equals the first source register, then it hits the optimal path the
AFP.NEP will insert in to the result. AVX throws a small wrench in to
this due to changed behaviour
AVX example:
- vaddss xmm0, xmm1, xmm2
- xmm0 is ONLY the destination, xmm1 and xmm2 are the sources
- This operation copies the bits above the scalar result from the
first source (xmm1).
- Additionally this will zero bits above the original 128-bit xmm
register.
- xmm0[31:0] = xmm1[31:0] + xmm2[31:0]
- xmm0[127:32] = xmm1[127:32]
- ymm0[255:127] = 0
This causes these instructions to support a fairly large table depending
on if the instruction is an SSE or AVX instruction, plus if the host CPU
supports AFP or not.
So while fairly complex, it's handling all the edge cases and gives us
optimization opportunities as we move forward. Currently on non-AFP
supporting devices this has a minor benefit that these IR operations
remove one temporary register, lowering the Register Allocation
overhead.
In the coming weeks I am likely to introduce an optimization pass that
removes redundant inserts because FEX currently does /really/ badly with
scalar code loops.
Needs #3184 merged first.
Suggested by Alyssa. Adding an IR operation can be a little tedious
since you need to add the definition to JIT.cpp for the dispatch switch,
JITClass.h for the function declared, and then actually defining the
implementation in the correct file.
Instead support the common case where an IR operation just gets
dispatched through to the regular handler. This lets the developer just
put the function definition in to the json and the relevent cpp file and
it just gets picked up.
Some minor things:
- Needs to support dynamic dispatch for {Load,Store}Register and
{Load,Store}Mem
- This is just a bool in the json
- It needs to not output JIT dispatch for some IR operations
- SSE4.2 string instructions and x87 operations
- These go down the "Unhandled" path
- Needs to support a Dispatcher function override
- This is just for handling NoOp IR operations that get used for
other reasons.
- Finally removes VSMul and VUMul, consolidating to VMul
- Unlike V{U,S}Mull, signed or unsigned doesn't change behaviour here
- Fixed a couple random handler names not matching the IR operation
name.
This is the cause of a bunch of redundant moves that shows up in
InstCountCI. Fixing this aliasing and pre-colouring issue causes a ton
of 256-bit operations to become optimal.
Mostly copypaste of Orlshl... we really should deduplicate this mess somehow.
Maybe a shift enum on the core Or op?
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
If we const-prop the required functions and leafs then we can directly
encode the CPUID information rather than jumping out of the JIT.
In testing almost all CPUID executions const-prop which function is
getting called. Worst case that I found was only 85% const-prop rate.
This isn't quite 100% optimal since we need to call the RCLSE and
Constprop passes after we optimize these, which would remove some
redundant moves.
Sadly there seems to be a bug in the constprop pass that starts crashing
applications if that is done.
Easily enough tested by running Half-Life 2 and it immediately hitting
SIGILL.
Even without this optimization, this is stil a significant savings since
we aren't jumping out of the JIT anymore for these optimized CPUIDs.
The bug that was causing crashes with this was due to inline syscalls.
Now that this is fixed we can re-enable store->load operations.
This allows constant propagation to work significantly better, which
means inline syscalls start working again. This can significantly
improve syscall performance in some cases.
This is most likely to improve performance in dxsetup and vc_redist but
hard to get a real profile.
Additionally this will let us inline cpuid results in the future which
is pretty nice.
The range was slightly incorrect which mostly wouldn't have caused
issues.
The lowest byte would have just generated slightly less optimal code.
The upper byte could have generated broken code, which our CI couldn't
catch since TSO instructions only get enabled when multiple threads are
in-flight.
Easy enough to fix.
Add new synthetic condition codes that do an AND as their relational operator,
testing the result. This is 1 IR op for things like
(A & B) == 0 ? C : D
This can translate to
tst A, B
csel A, B, eq
In the future, if A is the NZCV register and B is a supported immediate, eg
(NZCV & 0x80000000) == 0 ? C : D
this will be able to translate to a single instruction with the appropriate
condition
csel A, B, pl
but that needs RA support.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This is blocking performance improvements. This backend is almost
unilaterally unused except for when I'm testing if games run on Radeon
video drivers.
Hopefully AmpereOne and Orin/Grace can fulfill this role when they
launch next year.
While this interface is usually pretty fast because it is a write and
forget operation, this has issues when there are multiple threads
hitting the perf map file at the same time. In particular this interface
becomes a bottleneck due to a locking mutex on writes in the kernel.
The situations when this bottleneck occurs is when a bunch of threads
get spawned and they are all jitting code as quickly as possible. In
particular Geekbench's clang benchmark hits this hard where each CPU
thread spends ~40% CPU time on all eight CPU threads because they are
stalled waiting for this mutex to unlock.
To work around this issue, buffer the writes a small amount. Either up
to a page-ish of data or 100ms of time. This completely eliminates
threads waiting on the kernel mutex.
- Around a page of buffer space was chosen by profiling Geekbench's
clang benchmark and seeing how frequently it was still writing.
- 1024 bytes was still fairly aggressive, 4096 seemed fine.
- 100ms was chosen to ensure we don't wait /too/ long to write JIT
symbols.
- In most cases 100ms is enough that you won't notice the blip in
perf.
One thing of note is that with profiling enabled and checking the time
on every JIT block still ends up with 2-3% CPUtime in vdso
clock_gettime. We can improve this by using the cyclecounter directly
since that is still guaranteed to be monotonic. Maybe we'll come back to
that if it is actually an issue here.
If we are going to throw away the updated value of CF anyway there is no point
wasting an instruction to invert CF. Add an IR toggle for that so the arm64 JIT
can make better choices.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Changes the helper which all the source uses to still calculate the size
implicitly. This is going to take a while to convert all implicit uses
over to the explicit operation.
Get us started by at least having the IR operation itself be explicit.
Skips implementing it for the x86 JIT because that's a bit of a
nightmare to think about.
The ARM64 implementation requires sequential registers which means if
the incoming sources aren't sequential then we need to move the sources
in to the two vector temporaries. This is fine since we have zero-cycle
vector renames and the alternative is slower.
Currently FEX will always jump out of the JIT any time FCW was getting
written to, ensuring that the softfloat state is setup to rounding at
the time of FCW getting written.
This has the unintended side-effect that even in "x87 reduced precision"
mode we were jumping out of the JIT.
This hit a real world use case of an installer reloading FCW after every
x87 operation and generating a block with 2297 instructions.
Instead when jumping out of the JIT for handling x87 operations, load
FCW and pass it as the first argument of the handler. Setting the
softfloat state at that point.
This helps the installer's hottest block by cutting it down to 1477
instructions. 64.3% of the original size. The code block is still
burning 90% of the CPU time of the installer but the performance is
significantly better while it is doing its decompression.
In order to optimize this installer's block of code more then we will
likely need to optimize out x87 stack usage.
This is taking steps to start fixing RCLSE which was started by #2700.
Same situation as that PR, since #2170 when we converted
{Load,Store}Context in to {Load,Store}Register we broke this pass
entirely. It hasn't been doing anything for redundant GPRs and FPRs
since at least November of last year.
Technically it was potentially still optimizing redundant MMX
accesses, but it is so broken that it doesn't matter.
Instead of going all in like #2700 did, tear down the pass and start
again. We are now /only/ optimizing redundant context/register loads.
This fixes an issue that comes up commonly where the same register used
as sources was getting loaded twice, causing redundant moves.
`packsswb xmm0, xmm0` for example was generating a four instruction
sequence instead of three instructions because we weren't eliminating
the redundant load.
Going to take reimplementing all the optimizations that this pass does
in steps. This way we can track any regression in the independent steps
unlike what happened in #2700.
Confirmed that Proton/Sonic Mania still works after this.
It turns out that pure SSA isn't a great choice for the sort of emulation we do.
On one hand, it discards information from the guest binary's register allocation
that would let us skip stuff. On the other hand, it doesn't have nearly as many
benefits in this setting as in a traditional compiler... We really *don't* want
to do global RA or really any global optimization. We assume the guest optimizer
did its job for x86, we just need to clean up the mess left from going x86 ->
arm. So we just need enough SSA to peephole optimize.
My concrete IR proposals are that:
* SSA values must be killed in the same block that they are defined.
* Explicit LoadGPR/StoreGPR instructions can be used for global persistence.
* LoadGPR/StoreGPR are eliminated in favour of SSA within a block.
This has a lot of nice properties for our setting:
* Except for some internal REP instruction emulation (etc), we already have
registers for everything that escapes block boundaries, so this form is very
easy to go into -- straightforward local value numbering, not a full into
SSA pass.
* Spilling is entirely local (if it happens at all), since everything is in
registers at block boundaries. This is excellent, because Belady's algorithm
lets us spill nearly optimally in linear-time for individual blocks. (And
the global version of Belady's algorithm is massively more complicated...)
A nice fit for a JIT.
Relatedly, it turns out allowing spilling is probably a decent decision,
since the same spiller code can be used to rematerialize constants in a
straightforward way. This is an issue with the current RA.
* Register assignment is entirely local. For the same reason, we can assign
registers "optimally" in linear time & memory (e.g. with linear scan). And
the impl is massively simpler than a full blown SSA-based tree scan RA. For
example, we don't have to worry about parallel copies or coalescing phis or
anything. Massively nicer algorithm to deal with.
* SSA value names can be block local which makes the validation implicit :~)
It also has remarkably few drawbacks, because we didn't want to do CFG global
optimization anyway given our time budget and the diminishng returns. The few
global optimizations we might want (flag escape analysis?) don't necessarily
benefit from pure SSA anyway.
Anyway, we explicitly don't want phi nodes in any of this. They're currently
unused. Let's just remove them so nobody gets the bright idea of changing that.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Now that PF calculation is deferred, the cost of calculating PF correctly should
be tolerable. Remove the speed hack to skip PF. It's fundamentally broken, and
there are enough broken things in FEX as it is that we don't need to maintain
this one ;-)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
FEX has a problem with large blocks that uses a ton of constants spread
throughout the block. Once a block gets large enough with enough
constants that have large live ranges, FEX slows down to unusable speeds
due to the register allocator spending more time calculating node
interferences than anything else in the program.
This adds a little heuristic to ensure that constants aren't reused if
the previous value is past a certain distance threshold. This threshold
works well enough that XeSS's pedantic initialization code doesn't have
issues now. See https://github.com/FEX-Emu/FEX/issues/2688 for more
information about that.
FEX itself should work to remove bad constant usages to make this pass
less necessary anyway. In most cases we are materializing duplicated 0,
1, and masks which could be done without a constant entirely.
Maybe once we've improve that enough we could remove this constant
pooling entirely.
To note, this doesn't fix the issue that XeSS causes our register
allocator, this is purely a heuristic workaround.
With Or, Orlshl, Bfe, and Bfi there were some assumptions made that i8
and i16 operations made sense. Which required us to disable the IR
validation for these operations when it was just added.
This removes the final assumptions about these IR operations supporting
these small operating sizes allowing us to enable the IR validation.
Also a very minor optimization by moving a couple extracts from source
before trying to BFI from it, making RA more optimal.
The FXSAVE and FSAVE tag words are written out in different formats,
with FXSAVE using an abridged version that lacks the zero/special/valid
distinction. Switch to using this abridged version internally for
simplicity, and to allow the calculation of zero/special/valid
distinction to be deferred until an fxsave instruction (in the future,
currently the distinction is ignored and only valid/empty states are
possible).
AddNZCV is a new op to return the NZCV for an addition directly, which lets us
skip software flag calculation in some cases. In the future it would be nice to
fuse this into the Add itself as a second destination to avoid repeating the
addition, but that's a very involved change and right now I'm building FEX on an
old Chromebook because my M1 kernel is FUBAR.
Similarly, SubNZCV returns flags for Sub. This has the extra twist of needing to
invert the carry bit due to the inverted definition between arm64 and x86_64.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
A bunch of the AES operations take a zero register upfront and we
currently materialize it for each instruction.
Considering that most AES operations are used back to back, we can
eliminate these materializations by caching it between instructions.
Additionally removes a move in the optimal case when destination matches
the state register, which is exactly what the SSE operation ends up
doing.
AESKeyGenAssist has an edge case that if the destination RA overlaps the
zero register then we still need to eat a move, hopefully doesn't happen
too frequently in practice. This is also the lesser used instruction so
it isn't a big deal. RA constraints could solve that still.
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.
Ryan Houdek