I have a nyquist script using snd-fft api which have a random beaviour on my computer
where floats are 32 bits and long 64 bits.
Each time I execute the script, result is different, when it should be about equal.

I 'd like to know if same random issue occur on 32 bits systems.
-------------8<--- Here after extracts of the debian source 
-------------8<--- ./lib-src/libnyquist/nyquist/nyqsrc/fft.c
110          * in the block. Here, there are 4 more longs after the size, and
111          * then room for 3*len floats (assumes that floats and longs take
112          * equal space).

130         window = (float *) &(s->extra[OFFSET + 2 * len]);

141     samples = (float *) &(s->extra[OFFSET]);
142     temp_fft = samples + len;
143     window = temp_fft + len;

// Note that as s->extra seams to be long pointer while samples and temp_fft seams to
// be float pointers, I am not confident that windows from line 130 and 143 might overlap.

-------------8<--- Here after the bits processor detector

// long_and_float_size.c
void main ()
{
  printf ("sizeof ( long ) = %d\n", sizeof(long) );
  printf ("sizeof ( float ) = %d\n", sizeof(float) );
  return sizeof(float) == sizeof(long);
}
// :~$ ./a.out
// sizeof ( long ) = 8
// sizeof ( float ) = 4

-------------8<--- Here after the nyquist script

;nyquist plug-in
;version 1
;type analyze
;categories "http://audacityteam.org/namespace#OnsetDetector"
;name "FFT ..."
;action "FFT tests..."
;info "Released under terms of the GNU General Public License version 2"
;control freq_basses2 "Filtre basses Hz" int "" 40 5 200

(setf s1 (if (arrayp s) (snd-add (aref s 0) (aref s 1)) s))
; ajout des canaux gaucHe et droite dans un canal unifie, s1

(defun signal_basses () (force-srate 1000 (lp s1 60.0 ) ) )
; Filtre des frequences inferieures à 60 Hz

(setq l NIL )
(setq result-fft NIL)
(setq result-fft (snd-fft (snd-copy (signal_basses) ) 128 0 NIL ) )
(do ( (n 0) (X 1.0) )
    ( (> n 60) NIL )
  (print (list n (aref result-fft (* 2 n ) ) ) )
  (setq n ( + 1 n ) )
)
(setq label "toto")
(setq l (cons (list 0.0 label) l))
l

To reproduce the bug.

Use a 64 amd processor.

Install the nyquist script in ~/.audacity-files/plug-ins/fft.ny

If you cannot load a sound with bass such as bachata,
use the audacity menu: Générer/Sons/ Sinuosiadal 40Hz Amplitue 0,8
or nicer and more audible : Générer/Risset Drum/50 Hz Decay 2s

Launch the FFT script from Analyse/FFT ...
Choose to Validate or Debug

And look the 60 values displayed in the console.
Values about offset 40 or 50 are assumed to be greater
than others for a signal frequency around 40 or 50 Hz..

Here is a patch that probably only work on amd 64 bits machines and compatibles as
it assumes sizeof long equals sizeof double.
--- ./lib-src/libnyquist/nyquist/nyqsrc/fft.c.original	2011-03-10 21:44:23.000000000 +0100
+++ ./lib-src/libnyquist/nyquist/nyqsrc/fft.c	2011-03-10 21:58:01.000000000 +0100
@@ -31,7 +31,7 @@
  * extra[2] -> INDEX (current sample index in current block)
  * extra[3] -> FILLCNT (how many samples in buffer)
  * extra[4] -> TERMCNT (how many samples until termination)
- * extra[5 .. 5+len-1] -> samples (stored as floats)
+ * extra[5 .. 5+len-1] -> samples (stored as doubles)
  * extra[5+len .. 5+2*len-1] -> array of samples to fft
  * extra[5+2*len ... 5+3*len-1] -> window coefficients
  *
@@ -56,7 +56,7 @@
 #define SAMPLES list->block->samples

 /* DEBUGGING PRINT FUNCTION:
-    void printfloats(char *caption, float *data, int len)
+    void printdoubles(char *caption, double *data, int len)
     {
         int i;
         printf("%s: ", caption);
@@ -67,7 +67,7 @@
     }
 */