New IR ops do them four lanes at a time where the backend has them, the
ARM64 JIT and the IR interpreter so far. vmin and vmax compare by
vfpu_minmax_key() in NEON, which took about twelve instructions a lane
through the integer registers before; vslt and vsge are a compare and an
AND with 1.0.
vi2f becomes one SCVTF with fractional bits.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A new IR op, Vec4VfpuMultiDot, does four vfpu_dot()s against one vector,
the other operands being the rows or Vec4s of a 4x4 block of registers.
The frontend emits it for vmmul and vtfm/vhtfm, and for runs of vdot that
fit it, like a matrix times a vector written out.
On ARM64 it calls a routine that has the dots in lanes rather than the
terms, so nothing is horizontal and nothing goes through the integer
registers. The block comes in with LD1, or LD4 for the transposed form,
when none of it is in host registers. The IR interpreter runs it through
vfpu_dot(); the other backends don't get it yet.
The CPU fuzzer gets runs of vdot shaped for it, and finds no difference
between the cores on them.
There's a routine per number of terms, so vdot.t runs don't pay for a
fourth. The vector's significands are worked out once and multiplied by
element, and the rounding at the end is UCVTF's in round-to-nearest,
which is what games run in; the integer version stays for the other
modes. The routine saves nothing, and the call site keeps only those
registers it clobbers that hold guest state.
JitDotBench, ns per dot: vtfm4.q 3.40, vmmul.q 3.12, a run of vdot.t
2.83, against 4.6 for single dots.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The frontend lowers vdot, vhdp, vfad, vavg, vdet, vcrsp/vqmul, vtfm/vhtfm and
vmmul to it then, with the operands built from the same lanes, constants and
negations the interpreter feeds vfpu_dot(). BuildDotOperands() puts them in
Vec4 regs, using the free temps where they aren't one already, and the ops
fall back to the interpreter when there aren't enough. The IR interpreter runs
it through vfpu_dot(), and the native backends through their generic fallback
for now. vmmul takes the same path for the MoreAccurateVMMUL compat flag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Blocks ending in a branch dispatch a conditional exit and then the
fallthrough ExitToConst. One op now returns either target, reading the
second from the ExitToConst, which stays behind unexecuted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- An lwl/lwr pair was combined into one load even when the first half loads
into the base register, which changes the address of the second half.
- ApplyMemoryValidation shared one sp check across the block even past an
Interpret or CallReplacement, which may change sp.
- Drop a duplicate FSqrt meta entry, and name Load8Ext correctly.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A new FSinCos op writes both from one argument reduction. arm64 and x64 get
both back from a single call, packed in one double; RISC-V and LoongArch make
the two calls.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
vh2f always went to the interpreter in the IR. A new FHalfToFloat op
converts the lower or upper half of a word, and the native backends
call vfpu_h2f for it like FSin. The legacy arm64 JIT now makes the same
call instead of computing the conversion inline; vh2f is rare, and the
call is much less code.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Vec4ClampToZero and Vec2ClampToZero only ever fed Vec4Pack31To8 and
Vec2Pack31To16, for vi2uc and vi2us. The packs now clamp negative lanes
to zero themselves, which saves an op and a vector temp, and lets x64
clamp with PACKUSWB's saturation after an arithmetic shift.
While at it, RISC-V compiles Vec2Unpack16To31, Vec2Pack31To16 and
Vec4Pack32To8, and LoongArch Vec2Unpack16To31, Vec2Pack31To16 and the
non-LSX Vec4Pack32To8, all of which went to the IR interpreter.
LoongArch's Vec2Pack32To16 and Vec2Unpack16To32 now take their scalar
path with LSX too, instead of falling back.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
These were always interpreted. They're now IR ops that the native
backends compile to calls to vfpu_exp2 and vfpu_log2, like FSin and
FAsin. vrexp2 is FNeg followed by FExp2, which is how vfpu_rexp2
computes it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
These went through the host's sqrt and division everywhere except the
interpreter's vrcp and vnrcp (vsqrt and vrsq there only behind
USE_VFPU_SQRT, now gone). They now always give the PSP's bits: the IR
gets FVSqrt (FSqrt stays the FPU's IEEE sqrt.s), and FRSqrt and FRecip,
which only the VFPU emits, become vfpu_rsqrt and vfpu_rcp; the IR
interpreter and the x64, arm64, RISC-V and LoongArch backends call them.
The old JITs call them directly, the ARM ones keeping the lanes in
callee-saved registers across the calls. cpu/vfpu/exact now passes on every core.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
