Every runtime module load sets up the import stubs of all loaded modules
again, and each setup built a fresh trampoline and rewrote the stub's 16 bytes
-- under guest threads that may be executing exactly those bytes.
A trampoline only encodes its import index, so the trampolines are now cached
by index and reused, and a stub that already jumps to the trampoline it would
be given is left untouched, with no unprotect/write/restore cycle at all. A
reload of a module whose stubs are already patched therefore writes no guest
code.
The trampoline ABI test builds its backend with GetUninitializedObject, which
skips field initializers, so it has to supply the new cache; both fields now
go through one helper that names a field it cannot find.
The export returns the context memory size through one size_t. For heap
output it also wrote a second qword (an alignment) after it, and titles keep
other locals there: Octopath Traveler II's allocator reads the dword at +12 of
that block as a divisor, so it faulted with #DE during boot. Write only the
size, wherever the output lives.
A suspend point inserts GPU commands that drain the graphics pipe. The call
that inserts it does not wait for those commands, and the hardware allows one
suspend point to be in flight, but the queue implemented it by blocking the
calling thread until every accepted submission had been processed.
The frame boundary is now a queued submission of its own, ordered behind the
graphics work it ends, and processing it flushes what the frame recorded. The
caller returns immediately for the first boundary and only waits when a second
one arrives while the first is still pending. Shutdown still drains, because no
slice will run a queued boundary once the queue stops accepting work, and a
cancelled queue drops pending boundaries along with the rest.
Measured on Octopath Traveler II: the call went from 223 ms to 72 ms, and the
guest render thread spent 38 s instead of 107 s of its 170 s inside it.
PredicationZPassWaitOp is 0 for kWaitForQueryResults and 1 for
kDoNotPredicateIfQueryResultsNotReady, so the wait bit was being read
inverted. Every SET_PREDICATION packet that asked the command processor not to
stall drained the whole host queue instead, and the packets that did ask to
wait read the predication label without waiting for the results at all, so
they could act on a stale value.
The drain now happens only for kWaitForQueryResults, and only on the branch
that actually reads the label. A packet tagged
kDoNotPredicateIfQueryResultsNotReady leaves its work unpredicated, which is
what the hardware does while the results are still outstanding, and clearing
predication reads no memory at all so it never drains.
A resource descriptor the host materializes is built by replaying the shader's
scalar operations on the host. The validator that decides whether a value can
be replayed listed only part of what the scalar evaluator implements, so a
descriptor computed through any of the others was rejected even though the
host can compute it exactly.
The list now covers every pure scalar function the evaluator supports: the
remaining integer maximum, absolute value and high multiplies, the signed and
unsigned comparisons, 64-bit equality, the bit-counting and bit-scanning
operations, the quad mask, the integer/float conversions, and the float
arithmetic and comparisons. Anything with side effects or a lane-varying
result is still rejected.
S_CODE_END pads the space after a shader's last instruction. When the code
after the final S_ENDPGM ends in a backward branch, the linear decoder walks
straight into that padding and failed with an unknown SOPP opcode 0x1F, which
stopped the session.
The decoder now names 0x1F and ends the program there, unless a forward branch
still targets code beyond it -- in which case the padding really would be
inside the program and the decode still fails.
Two things were wrong in the SOPK forms, in both the SPIR-V and the MSL
backend.
S_CMPK_*_U32 compares against SIMM16 zero-extended; only the signed compares
sign-extend it. Both translators sign-extended it for every compare, so
"s_cmpk_le_u32 s4, 0xffff" was true for any value at all. NGG geometry
programs use exactly that test on the packed vertex and primitive count, so
they took their no-primitive path and exported a single null primitive, which
left Unreal's volume passes empty. The register S_CMPK names is also only
read, yet user-data liveness recorded it as a definition, so a compare on a
user-data register left that register unloaded; it now counts as a use, as
does the register of the SOPK accumulations.
S_ADDK_I32 is defined as D.i = D.i + signext(SIMM16) with SCC set on signed
overflow, the same SCC rule as S_ADD_I32, which both backends already
implement. Neither wrote SCC for the SOPK form, so a following
S_CBRANCH_SCC0/1 or S_CSELECT read whatever the previous instruction had left
there. S_MOVK_I32 and S_MULK_I32 correctly leave SCC alone.
v_min_f32 and v_max_f32 return the other operand when one of them is NaN, and
v_cvt_u32_f32 and v_cvt_i32_f32 saturate an out-of-range value to the target
type's limit and turn NaN into zero. The GLSL FMin and FMax instructions and
SPIR-V's ConvertFToU and ConvertFToS all leave those cases undefined, so a
host driver is free to hand back the NaN or an arbitrary integer, and Metal's
fast math does return the NaN.
The minimum and maximum now select the other operand when either side is NaN,
and the two conversions clamp at the type's limits and return zero for NaN.
NaN is tested on the value's bits rather than with an ordered comparison, so
fast math cannot assume the test away.
Device tests cover this and the wave-wide maximum that Unreal's motion blur
computes with DPP row_shr, v_permlanex16_b32 and v_readlane_b32 across a
64-lane wave.
Patching a guest code site unprotects the page it lives on, writes, and
restores the old protection. Two threads patching sites that share a page
therefore race: one restores the page to read-execute while the other is
still writing into it, and the second write faults or is lost. Both happen in
practice -- a runtime module load sets up import stubs while guest threads are
running, and guest threads patch their own lazy TLS access sites as they first
reach them, several of which land on one page.
Every import stub patch and every lazy TLS site patch now runs under a single
gate, so the unprotect/write/restore sequence is never interleaved.
A CPU access to a page the GPU memory tracker protects faults on every first
touch after an upload or a GPU write, thousands of times a second in a frame.
The POSIX bridge converted each one into a full Win64 CONTEXT (general and
vector registers), then walked the interrupt, image-tracker, lazy-commit and
allocator-hole handlers before the tracker ever saw it.
The tracker needs only the fault address and the access kind, so it now runs
right after the image tracker, straight off the raw signal. An env var
restores the old handler order.
The POSIX bridge installed handlers for SIGSEGV, SIGBUS, SIGILL, SIGTRAP and
SIGABRT, but not SIGFPE. A guest integer divide by zero therefore took the
runtime default and killed the process with no exception record and no guest
context, so the faulting instruction was invisible.
SIGFPE now routes to the same handler and is reported as an integer divide by
zero. Like SIGILL, it skips the image write tracker, which only ever resolves
data-access faults.
CaptureFlip created a 1920x1080 RGBA16F VkImage with its own VkDeviceMemory
for every flip and destroyed both when the flip retired: around twenty
vkCreateImage plus vkAllocateMemory pairs a second for the same extent, and on
MoltenVK as many Metal texture allocations.
Retired snapshots are now kept in a small per-format, per-extent pool and
handed back instead. A snapshot is fully overwritten by the copy that follows
it and its barrier takes it from Undefined, so neither its previous contents
nor its layout matter. The pool holds at most eight images, is guarded because
the command-stream thread takes snapshots while the render thread retires
them, and is drained at teardown. SHARPEMU_FLIP_SNAPSHOT_POOL=0 turns it off.
CaptureFlip recorded the display-surface copy into the guest batch, flushed
it, and then read the shared _submitTimeline field for the presentation's
RequiredTick. The presenter thread writes that same field from its own
submission, so the flip could be tagged with an older tick that was already
complete. The render loop then blitted the freshly created (never written)
flip snapshot, which presents as fully zeroed RGBA - a black frame with an
alpha of 0. Take the tick that this flush returned instead.
A float guest flip is blitted through an sRGB image before it is copied into
the UNORM swapchain image, and one such intermediate served every frame. Each
frame discards its contents by transitioning it Undefined -> TransferDst
before blitting, so with two frames in flight the next frame's blit raced the
previous frame's copy-out and the presented image came out fully zeroed. The
guest's own display buffers held the correct picture the whole time, so the
window simply went black for long stretches.
Each swapchain image now owns its own intermediate, created and destroyed with
the swapchain, so a frame can only ever discard the intermediate belonging to
the image it is about to present.
On Superliminal the same run went from 60 of 84 presented-swapchain readbacks
black to 10 of 187, all of them from before the menu appears, and the menu no
longer turns black after a couple of minutes.
Guest faults are delivered on the guest stack, and the managed handler ran
there. The runtime only knows the thread's own stack; Windows swaps the TEB
stack bounds while guest code runs, but POSIX has no equivalent. A garbage
collection during a guest fault then misread the thread's frames and
reported garbage object references, and the process failed with "attempted
to call a UnmanagedCallersOnly method from managed code". A small native
trampoline now moves to the host stack below the frames that entered guest
code before it calls the managed handler.
The two remaining reserved pixel system inputs were also left reading zero.
SAMPLE_COVERAGE now reads the first element of the SampleMask built-in, and
POS_FIXED_PT packs the integer pixel position taken from FragCoord, x in bits
[15:0] and y in bits [31:16], so no enabled pixel system input is left unset.
A pixel shader that enables FRONT_FACE or ANCILLARY reads those VGPRs, but the
translator only reserved their compact slots and left them reading zero.
Unreal Engine bakes its colour grading LUT one volume slice per instance and
has the pixel shader read the slice back out of ANCILLARY[28:16], the render
target array index. With that register zero every slice was baked from the
blue = 0 slice, so the tonemapped image lost its blue channel.
ANCILLARY is now the Layer built-in placed at bit 16, and FRONT_FACE the
FrontFacing built-in mapped to 1.0f for a front face and 0 for a back face,
which satisfies both the float greater-than and the integer not-equal tests
compilers emit for it.
Guest memcpy went through the managed gateway on every call: two write
tracking notifications, the memory reader lock and two region lookups, about
0.8 us per call under Rosetta even for a copy of a few bytes. Dreaming Sarah
makes roughly 45k of those calls a second.
Guest memory is identity-mapped, and every fault a copy can take -- write
tracking, lazy commit -- is resolved from the fault address alone, so the copy
can run as a forward rep movsb in a fast-path stub. A destination that sits
inside the source, where a forward copy would read bytes it has already
written, and the write-watch diagnostic both keep the managed export.
The managed import gateway costs a guest-to-host stack switch, a full guest
register pack, a GCHandle dereference and the whole dispatch bookkeeping
before the HLE export even runs. For the two hottest trivial imports that is
three orders of magnitude more work than the answer itself: one UE5 title's
load measured 12.9M scePthreadGetspecific calls (38 s of wall time) and 4.3M
scePthreadSelf calls, at roughly 2.9 us each.
The few values those two imports need are now published into a per-host-thread
native block whose address lives in a native TLS slot, rebound whenever a
thread enters or is restored as a guest thread, and the two imports are
patched to emitted x86-64 stubs that read the block directly. pthread-specific
values live in native per-thread tables so a stub and the managed export read
exactly the same storage and cannot drift apart. The managed path stays the
source of truth for every case a stub declines -- no block published, no guest
thread bound, no value table, a key outside the table -- which it reaches by
jumping to the import's normal trampoline with every register intact,
including the incoming RAX variadic vector count.
The value-table slot follows whichever handle the current thread reports, the
bound guest thread or the host thread's own synthetic handle, so
pthread_getspecific is answered on unbound host threads as well; the primary
execution thread never enters through the guest-thread path and is a major
caller in UE5 titles. The self-handle slot stays guest-only, because
pthread_self on an unbound thread registers that thread's CPU context with the
guest scheduler and must keep reaching the managed export. Each stub also
bumps a per-thread hit counter, reported next to the existing HLE cost
histogram, so a call that misses the stub is visible rather than invisible.
The stubs are gated to macOS x64, where the emitted TLS read matches the
platform's pthread key array, and switched off whenever import or pthread
tracing needs every call to pass through the gateway.
Two sources of instructions the driver has to compile but can never need.
A read of a planned slot in the host-flattened resource table is always
inside the table the host binds, because the table holds every planned read
and written range. Those loads nevertheless carried a runtime array-length
compare and two selects each; a large pixel shader made about 376 of them.
A slot below the planned reserved word count now loads directly, and only a
slot computed at run time keeps the check.
ds_swizzle and ds_bpermute exchange data between lanes and GDS instructions
address the global data share, so none of them touches LDS. A graphics stage
that uses only those was still given the 8 KiB zero-initialized private LDS
array, which Metal pays for in compile time and in private memory per
invocation. The array is now declared only for instructions that really
address LDS.
The translator materializes a good deal of translation-time knowledge as
SPIR-V instructions: a specialized descriptor's buffer format, swizzles,
strides, and the conditions that guard host-guaranteed ranges all end up as
arithmetic, comparisons and selections over operands the builder itself just
emitted as constants. That reaches the driver as runtime work, and on Metal
every one of those instructions costs pipeline-creation time.
The module builder now remembers the value of each 32-bit integer and boolean
constant it emits and folds pure instructions over them while it builds:
integer arithmetic, bit operations, shifts, bitfield extraction, the integer
and logical comparisons, the logical connectives, and a selection on a known
condition. Folding happens at emission, so a folded result feeds further
folding without a separate pass. Anything whose operands are not all known
constants is emitted exactly as before.
The buffer format decoder uses this directly: a specialized descriptor makes
the unified format a constant, so it decodes on the host instead of emitting
a module-level lookup table and an indexed read.
The test that covered that path asserted the table's presence and now asserts
its absence.
SIGTERM (launchd, process managers, kill) used the runtime default and
exited with GPU work in flight. It now takes the same host-interrupt path
as Ctrl+C: the presenter closes and tears Vulkan down before exit.
Submit's backpressure wait is what paces every audio grain on the SDL
output, and it polled the queued byte count with Thread.Sleep(1), costing
the audio thread several wakeups per grain.
It now sleeps for the time the device needs to drain back under the queue
cap, still bounded by the existing deadline and the maximum wait.
The depth attachment layout followed each draw's depth/stencil write mask, so
a draw sequence that toggles depth writes — opaque geometry, then decals and
translucency — changed the layout every few draws, and every layout change
ended the rendering scope.
The read-only attachment layouts are only required when the draw also samples
the attachment. Every other draw now uses the writable layout for the
aspects the format has, and the write mask alone selects the access, so the
layout stays constant and the scope survives. The presenter answers whether a
texture bound for the draw being prepared views the depth attachment's
subresources, defaulting to yes for hosts that do not track it.
sceAudioOut2ContextAdvance and sceAudioOut2ContextPush each submitted the
context's pending PCM and, whenever nothing reached a host device, each slept
a whole grain in the software pacing fallback. A title that calls Advance then
Push once per grain therefore spent two grain periods per grain, so its audio
thread ran at half real time and held up every thread waiting on it.
Advance now claims the next grain on the context's clock and returns without
blocking (or, when its own submit already blocked in the backend, marks the
grain served), and Push waits for the grain Advance claimed; a Push with no
preceding Advance claims one itself. Neither call skips pending PCM. Because
the output holds QueueDepth grains, a grain may be queued as soon as the grain
QueueDepth-1 ahead of it starts playing, so a context can run that far ahead
of real time before a push blocks.
sceAudioOut2ContextGetQueueLevel reported an empty queue with the full depth
available on every call. Titles size their rendering by those two values --
Wwise renders exactly puiAvailableQueues grains per update and pushes them
non-blocking -- so a queue that never filled made it render a full queue every
update and then block inside Push for a grain per push while holding its audio
lock, which starved the game thread. The level now comes from the context's
grain clock, or, once a grain has reached a host device, from that device's
queued milliseconds beyond a short cushion.
In wave32 a lane mask is its low register alone: EXEC_HI and VCC_HI are
ordinary SGPRs. Writing one of them therefore must not recompute the wave
lane state, and the lane state has to be tested against the low dword rather
than against a 64-bit mask composed from the low and high registers.
The lane-bit test that guards this covered the wave64 constant width only, so
it now accepts a mask of either width.
Tile-based GPUs pay for every rendering scope: ending one stores the
attachments out of tile memory and beginning the next one loads them back.
The presenter ended the scope for three things that do not actually need it.
An image barrier that repeats an attachment access in the same layout only
matters to work that runs after the scope, because accesses to an attachment
inside one scope are already ordered by rasterization order. Such barriers
are now queued and recorded as the scope ends instead of ending it.
A guest cache flush (a global memory barrier) inside a scope whose draws only
wrote attachments is deferred the same way, and every scope now opens with
that global barrier so anything recorded before the scope stays ordered
before its draws. A draw whose vertex or pixel stage stores to buffers,
storage images or device addresses ends a scope with a deferred flush first,
so a store can never move ahead of it. Finally, the DMA fill and copy entry
points no longer end the scope up front: the buffer cache completes most
transfers in guest memory, and the paths that do record GPU work end the
scope themselves.
SHARPEMU_DEFER_GLOBAL_BARRIERS=0 restores ending the scope at every flush.
kuba
ParantezTech