chorus: keep the tap read position inside the delay line

Message

A read position a few float ulps below zero rounded to exactly bufsize after the wrap, indexing one past the buffer. Found by review; an exhaustive index probe hit it 2,425 times in 25M (position, delay) pairs and 0 after the fix. Also document that comp applies makeup at ratio 1 and treats a ratio below 1 as bypass.

Changed

 src/c/fx4.c | 3 +++
 src/c/fx4.h | 3 ++-
 2 files changed, 5 insertions(+), 1 deletion(-)

Diff

src/c/fx4.cmodified

@@ -44,6 +44,9 @@ static inline SPFLOAT chorus_tap(const sp_chorus *p, SPFLOAT delay)

if (delay > maxd) delay = maxd;

SPFLOAT rpos = (SPFLOAT)p->pos - delay;

while (rpos < 0.0f) rpos += (SPFLOAT)p->bufsize;

+47

/* A tiny negative rpos rounds to exactly bufsize after the add

+48

* (float ulp at ~1768), which would index one past the buffer. */

+49

if (rpos >= (SPFLOAT)p->bufsize) rpos -= (SPFLOAT)p->bufsize;

int i0 = (int)rpos;

SPFLOAT frac = rpos - (SPFLOAT)i0;

int i1 = i0 + 1;

src/c/fx4.hmodified

@@ -11,7 +11,8 @@

* rate reduction. compute(in, drive, bits, rate_div)

* sp_comp feed-forward peak compressor with a one-pole attack/release

* envelope. compute(in, threshold_db, ratio); attack, release

−14

* and makeup gain are init-time.

+14

* and makeup gain are init-time; makeup is applied even at

+15

* ratio 1, and a ratio below 1 is a bypass, not expansion.

*

* Deterministic: no randomness, no libm on the per-sample path except

* tanhf in the crush drive stage and the dB conversions in the compressor