absolute worst case 2 : new variant 2 cycles faster
best case i've found: 30 cycles faster
all my tests showed identical results to the existing function

time to sleep....

did not work because of missing return - for num_channels<=1, "left channel" was still running..
(1  means pass trough)

In my testsketch with 2 x Chorus with n_chrous=4 & #define CHORUS_DELAY_LENGTH (24*AUDIO_BLOCK_SAMPLES),
Cpu-usage decreases from 13.8% to 9.8%

Access Q-bit from c

You mentioned in synth_dc.h "// if only C language provided a way to test Q status bit...."
Here it is :) . That return (out & 0x8000000)>>27; translates to only one more instruction : ubfx    r0, r0, #27, #1 , i decided to use the c -variant because it is better to readable.

I added a function to clear that bit ( clr_q_psr() ), too. Unfortunately, three instructions are needed (read-modify-write)

(PlatformIO) Use tagged source code from repo

Len noticed a bug in binary velocity output

Merge changes from LenShustek/miditones

https://forum.pjrc.com/threads/32580-Audio-System-Design-Tool-Typo

Move miditones to "extras"-folder to prevent compiling

https://hackaday.io/project/5912-teensy-super-audio-board

http://forum.arduino.cc/index.php?topic=372960.msg2574826#msg2574826

Bug fix so that downward tone sweeps will work.

Fixed problem with tone_hi - tone_lo computation being done
with unsigned arithmetic, and thus always giving a positive result.

don't save r12

indeed..you're right, saving r12 seems to be not necessary.
i did some tests, even with gcc 5, all worked. but  i'm still only 99.9% sure .-)

Just learned that saving r0-r3 is superflues

http://stackoverflow.com/questions/261419/arm-to-c-calling-convention-registers-to-save

update