Replace the x64 STVL/STVR lowering with explicit byte stores that preserve the
guest-visible vector byte order.
The previous vector shuffle/blend path could write incorrect bytes for partial
vector stores, which broke optimized guest memcpy implementations using
stvlx/stvrx for unaligned heads and tails. In 4156081C this corrupted skinned
model render command data and eventually caused render-thread crashes.
Add regression coverage for StoreVectorLeft/StoreVectorRight across common
unaligned offsets, plus a memcpy-head style case matching the affected guest
instruction pattern.
- DIV_I32/I64 and MUL_HI_I64: use source registers directly in
sdiv/udiv/umulh/smulh instead of copying to scratch first
- SWIZZLE: fast paths for identity (no-op), broadcast (dup), and
pair-swap (rev64) before falling back to general TBL
- LOAD_VECTOR_SHL/SHR: build base index vector with fmov+ins
instead of stp+ldr stack round-trip
Replace per-op save/restore of FPCR in EmitWithVmxFpcr with tracked
lazy switching via ChangeFpcrMode. Consecutive VMX float ops in the
same basic block now emit a single FPCR switch instead of one per op.
ChangeFpcrMode loads pre-computed values from the backend context
(fpcr_fpu/fpcr_vmx) instead of MRS + read-modify-write, eliminating
an expensive system register read per switch.
FPU mode is restored at block boundaries and calls via ForgetFpcrMode.
Scalar FP sequences guard with ChangeFpcrMode(Fpu) which is a no-op
when already in FPU mode.
The LSE paths in RESERVED_STORE_I32/I64 returned early from Emit,
leaving forward-referenced no_reserve/done labels unbound. Xbyak
encodes these as offset-zero branches (branch-to-self), causing
infinite loops when the reservation check fails on LSE hardware.
Replace early return with branch-to-done inside an if/else so
labels are always bound regardless of the LSE path taken.
Adds tests for RESERVED_STORE I64 success path and I32/I64
no-reservation failure path with timeout-based hang detection.
* Add unit tests
* Use `dc zva` to zero out entire cache-lines worth of data, when possible
* Special instruction that zaps an entire cache-line into being zero-values, typically 64-bytes at a time
* Statically read and initialize the `DCZID_EL0` register value since reading registers with `mrs` can be kinda slow and isn't worth doing redundantly at JIT-time when it never changes.
* Update store-instructions to use inline post-increments to avoid an additional `add` instruction and keep encoded instruction immediates small
Pretty much every arm processor has a cache line size of 64 bytes, so a typical `dcbz128` to clear 128-bytes of data now results in:
```
dc zva, x0
add x0, x0, 64
dc zva, x0
```
rather than a sequence of 8 `stp xzr, xzr, ...` instructions
Most non-Windows code paths use standard POSIX APIs (sockets, signals,
dlopen, threading) that work on any POSIX platform. Change Linux-specific
guards to !WIN32 or #else where the code is portable. Linux-specific
APIs (SIGRTMIN, Vulkan/X11, fontconfig/GTK) remain Linux-guarded.
Optimize `LoadV128Const` to detect even more opportunities
to use `movi`, `mvni`, and `fmov` and such for 16-bit, 32-bit
and 64-bit splats.
Update some `mov`+`dup` idioms into `LoadV128Const` as well
to catch more opportunities to optimize these const-loads when
possible.
Adds `platform_arm64` similar to `platform_arm64` to populate a 64-bit feature-flag value and creates a `a64_extension_mask` configuration-mask for the detection and utilization of certain arm ISA features. The first feature being `FEAT_LSE` which maps to `XBYAK_AARCH64_HWCAP_ATOMIC`.
Just like the x64 backend, implements `IsFeatureEnabled` to the emitter for checking if the host processor supports a certain subset of feature flag(s) at once.
Optimizes `OPCODE_ATOMIC_EXCHANGE`, `OPCODE_ATOMIC_COMPARE_EXCHANGE`,
and `OPCODE_RESERVED_STORE` with `FEAT_LSE` instructions.
Replace CallNativeSafe round-trips with fcvtl/fcvtl2 for half-to-float
conversion. A rev64 fixup corrects the PPC word-pair ordering. FLOAT16_2
uses ext to rotate the two source halfs into position and sets the
constant {0.0, 1.0} tail lanes. PACK_FLOAT16 kept as C helpers since
fcvtn doesn't match Xenos overflow saturation semantics.
Replace CallNativeSafe round-trips with inline TBL shuffles. LVL uses
a single-register TBL where out-of-range indices naturally produce
zero. LVR uses a 2-register TBL over {zeros, mem} so indices 0-15 map
to zeros and 16-31 map to memory bytes.
Add sentinel bits after the left-shift so clz stops at the correct
count (8 or 16) instead of counting through all 32 bits of the
widened register. Unlikely to hit in practice since PPC only has
cntlzw/cntlzd, but corrects existing sub-word variants.
Expands the vector register allocator from 20 to 28 registers by
including v8-v15. The HostToGuestThunk now saves/restores full Q8-Q15
(128-bit) instead of just D8-D15, and the GuestToHostThunk adds Q8-Q15
to its save/restore set since AAPCS64 only preserves their lower 64
bits across C calls.
Reduces register pressure for VMX-heavy guest code at the cost of
slightly larger thunk stack frames.
Reserves x19 to permanently hold the A64BackendContext pointer
(x20 - sizeof(A64BackendContext)), set once in the HostToGuestThunk.
Since x19 is AAPCS64 callee-saved, it survives all host calls for free.
Eliminates repeated 2-instruction mov+sub sequences in stackpoint
push/pop, stack synchronization, FPCR management, and reserved
load/store paths.
Reduces allocatable GPRs from 8 to 7 (like on x64).
Use the provided scratch register-index, defaults to W0/X0, and insert 64-bit elements directly into the vector-register rather than writing/reading to scratch memory.
Should allow certain splats to opportunistically emit the best assembly when possible.
Also use `std::bit_cast` rather than an inline `union` type to directly
cast a `float` into a `uint32_t`.
The `BRK` instruction code needs to be specified correctly so that a debugger can handle it properly(and continue from it). `brk 0xF000` implements a debug-breakpoint that can be continued from, and `brk 0xF001` is implemented an assertion failing. Any other codes like `0` and `F1` will not be able to be continued from or handled correctly by debuggers.
These seem to be standard brk codes:
```
Breakpoint 0xF000
Assert 0xF001
Debug Service 0xF002
Fastfail 0xF003
Divide by 0 0xF004
```
Rather than using a `mov` and vector-`dup` for vector-splats, just use the `movi` instruction where applicable(8-bit element-splats).
Eventually we might want to detect opportunities to use `FMOV` and the 64-bit form of `MOVI` over in `LoadV128Const`.
Track PPC scalar FP state in the backend context and restore it at
host/guest boundaries so translated FP code has a stable rounding mode.
Convert remaining VMX sequences to scoped EmitWithVmxFpcr. Remove
unnecessary ChangeFpcrMode from sequences that only call native helpers.
Co-Authored-By: Reality <reality@xenios.jp>
Save and restore FPCR around each VMX FP operation so vector code
doesn't leak FPCR state into scalar paths. Drop the DN bit so NaN
propagation fixup can inspect original payloads.
The A64 sequence table is populated by registration statics spread across
multiple translation units. Keeping the backing map as a global object makes
registration depend on cross-TU initialization order and can hit the table
before it has been constructed.
Move the sequence table behind a function-local static so sequence
registration is safe regardless of which unit initializes first.
Based entirely off existing xbyak x86 implementation and available
tests. Still needs a lot of optimization and testing on non-Windows
platforms.
So far passes all tests and boots at least some games on Windows.
Copy src1 into dest before the masked vrangeps so that unmasked lanes
(normal/inf/NaN values) are preserved via merge-masking. Previously,
when the register allocator assigned different physical registers for
dest and src1, unmasked lanes would retain stale values from dest.
Tested with Intel SDE and AVX-512 cpu to verify.
Also shift unused FpRangeSign constants into bits [3:2] of the vrangeps
immediate, matching the Intel encoding. Only one used in the code was 0
so it was already correct by accident but using the others in the future
would produce incorrect results.
Use the `vptestnmd`-instruction to quickly test if the exponent bits of
each element is zero and put these results into a mask register. This
mask register can then be used to write `+0.0` or `-0.0` values
depending on the original value's sign-bit.
A masked `vrangeps` instruction is used to move zero-values into the effected
elements while preserving the sign-bit without having to touch memory.
