akirilov-arm commented on Issue #2369:
Out of curiosity, is there a performance difference between
fmov d0, d1
andmov v0.8b, v1.8b
?According to the Cortex-A55 optimization guide, the
MOV
latency is 2, while theFMOV
one is 1. Strangely, theORR
(which is whatMOV
aliases to in this case) latency is also 1, but I have decided to be conservative and have made the change. That's the only difference I have found.Personally, I have been trying to use the term
Fpu
in VCode names to refer to SIMD & FP scalars, whileVec
(e.g.Inst::VecMisc
) - to vectors (and yes, there are 128-bit "scalars", though that subtlety pertains mostly to loads, and then to big-endian environments). With that in mindFMOV
is better, but ultimately it is the same operation.
cfallin commented on Issue #2369:
@akirilov-arm looks like there's a merge conflict on some tests; if you don't mind rebasing, I'm happy to merge this!
akirilov-arm commented on Issue #2369:
@cfallin Done.
Last updated: Dec 23 2024 at 12:05 UTC