Is there a recommended calling convention for stacking the SIMD/PIE registers across function and subroutine calls?
There is no calling convention, and accordingly no compiler support. In theory,
every function call may clobber any and all PIE registers/state.
OTOH you can implement your own "convention", or just avoid calls to "3rd-party" functions in the middle of PIE instruction sequences. I find this not hard to do because in the high-performance sections I don't want to do "slow" function calls anyway.
In C++ I also found it useful to use template functions, where the Q registers to use by a function are template parameters - so defined explicitly by the
caller.
Lastly, saving or restoring a Q register (except QACC) to/from an (aligned) location on the stack can be as little as 1 clock cycle each, so it's actually pretty cheap to do.
That said, neither the compiler, nor libc(++), nor the IDF use any PIE instructions or registers; but libraries like esp-dsp or esp-dl do.
In practice I never had the need to save/restore any PIE state, because the PIE code sections tend to be small and self-contained.
Which registers are preserved by the RTOS?
FreeRTOS saves/restores the full PIE "coproc" state upon context switch if needed, so nothing to worry about there.
Are there registers I should preserve before using?
Nope.
(On a note, some of the SIMD instructions (VMUL...) use the value from the Xtensa core "shift amount register" ("special register SAR") which is also used by the compiler, so may be "clobbered" at any time "outside" of assembly code.)
