Commit Graph
5 Commits
Author SHA1 Message Date
Valiantsin DzerakhandClaude Opus 5.5 5c48f62fca The FP fast paths of the lifted code check FPCR.RMode inside the result compare and pass vector slow-path operands through the Cpu
The result check of every scalar and vector FP op compares |r| against a per-thread threshold in the Cpu (fp_th_s,
fp_th_d): the smallest normal in round-to-nearest, NaN under a directed RMode, so a directed mode sends every result
to the redo path and the separate load and test of c->fpcr per op goes. rt_fpcr_threshold sets it; rt_fpcr_sync
calls it on every host, and the Cpu inits and the fiber switch's fpcr copy call it too.

The vector slow path FN(vredo) takes its operands from c->fp_arg instead of three stack copies per call site, which
takes ___chkstk_ms out of the large lifted functions (275 call sites in main to 5). It reruns only the lanes that
failed the check. A NaN result in round-to-nearest goes from redo straight to fix with the FZ-flushed operands. The
static FP helpers use preserve_most, so their cold call sites do not spill the caller-saved registers.

Results are bit-identical to the old headers (equivalence harness of build/v143-genko/static2, 200M cases).
On the 42 hottest 1.4.3 functions: instructions -8.2%, bytes -8.8%, hot-path instructions -4.3%.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-15 00:31:00 +03:00
Valiantsin DzerakhandClaude Opus 5.5 922edb177d Cheaper FPCR checks in the FP helpers
FCMP and the vector compares no longer flush denormal operands under FPCR.FZ: rt_fpcr_sync keeps MXCSR.DAZ equal to
FPCR.FZ, and x86 compares under DAZ read a denormal as a zero of its sign, as ARM does under FZ. The directed RMode
check of FN(op) and FN(vop) moved into the result check's branch, and min/max skip it, since they do not round.

Hot lifted functions shrink by about 10% of their x86 instructions; the town camera-pan bench (1080p, two runs each)
goes from 10.25 to 9.78 ms per frame. simd.py (60000 encodings), lift.py and func.py report 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-08 16:03:14 +03:00
deanzx06 ec9029def4 runtime: vectorized AVX round-to-odd FMA emulation for non-FMA/AVX-only CPUs 2026-11-02 05:17:58 +07:00
Valiantsin DzerakhandClaude Opus 5.5 2fe6a6d4c1 Keep MXCSR in line with FPCR where FPCR changes instead of checking it in every FP op
The FP helpers compared c->fpcr with a cached copy before every operation to resync MXCSR lazily. MXCSR now
follows FPCR.FZ at once: MSR FPCR calls rt_fpcr_sync, and so do rt_call (thread, fiber and host entry into guest
code), fiber resume (FIBER_FLAG_FLOAT_SWITCH fibers keep their own MXCSR) and the differential test harness per
case. op and vop only test FPCR.RMode for the slow path, and their redo check is one ordered-not-equal compare of
|r| with the smallest normal, the same NaN-or-smallest-normal set as before.

Hot guest code shrinks by about a quarter (132 functions holding half the guest cycles, 1.4.3); fight frame time
moves 1-3%, inside the run-to-run spread. lift, simd, func, atomic and flags_check pass on 1.4.3; simd.py fails
without the harness sync.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-28 12:25:25 +03:00
Valiantsin DzerakhandClaude Opus 5.5 19c9b648d1 Add the recompiler, Windows runtime, tools and docs
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 22:27:54 +03:00