diff --git a/Makefile b/Makefile
index f3e9c3e..3fe8240 100644
--- a/Makefile
+++ b/Makefile
@@ -29,6 +29,10 @@ dist: tarball
 
 upload: dist
 	ncftpput upload.sourceforge.net incoming $(ZIPFILE) $(TARBALL)
-	
-asm:
+
+asm: kiss_fft.s
+
+kiss_fft.s: kiss_fft.c kiss_fft.h _kiss_fft_guts.h
+	[ -e kiss_fft.s ] && mv kiss_fft.s kiss_fft.s~ || true
 	gcc -S kiss_fft.c -O3 -march=pentiumpro -ffast-math -fomit-frame-pointer -dA -fverbose-asm 
+	[ -e kiss_fft.s~ ] && diff kiss_fft.s~ kiss_fft.s || true
diff --git a/_kiss_fft_guts.h b/_kiss_fft_guts.h
index 9071c06..eefc17f 100644
--- a/_kiss_fft_guts.h
+++ b/_kiss_fft_guts.h
@@ -84,15 +84,6 @@ kiss_fft_cpx kf_cexp(double phase) /* returns e ** (j*phase)   */
     return x;
 }
 
-void kf_work(
-        kiss_fft_cpx * Fout,
-        const kiss_fft_cpx * f,
-        int fstride,
-        int in_skip,
-        int * factors,
-        const kiss_fft_state * st
-        );
-
 /* a debugging function */
 #define pcpx(c)\
     fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) )
diff --git a/kiss_fft.c b/kiss_fft.c
index 7674024..7e73d41 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -26,7 +26,7 @@ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 
 static void kf_bfly2(
         kiss_fft_cpx * Fout,
-        int fstride,
+        const int fstride,
         const kiss_fft_state * st,
         int m
         )
@@ -49,7 +49,7 @@ static void kf_bfly2(
 
 static void kf_bfly4(
         kiss_fft_cpx * Fout,
-        int fstride,
+        const int fstride,
         const kiss_fft_state * st,
         int m
         )
@@ -97,7 +97,7 @@ static void kf_bfly4(
 
 static void kf_bfly3(
          kiss_fft_cpx * Fout,
-         int fstride,
+         const int fstride,
          const kiss_fft_state * st,
          int m
          )
@@ -143,7 +143,7 @@ static void kf_bfly3(
 
 static void kf_bfly5(
         kiss_fft_cpx * Fout,
-        int fstride,
+        const int fstride,
         const kiss_fft_state * st,
         int m
         )
@@ -205,7 +205,7 @@ static void kf_bfly5(
 /* perform the butterfly for one stage of a mixed radix FFT */
 static void kf_bfly_generic(
         kiss_fft_cpx * Fout,
-        int fstride,
+        const int fstride,
         const kiss_fft_state * st,
         int m,
         int p
@@ -244,7 +244,7 @@ static
 void kf_work(
         kiss_fft_cpx * Fout,
         const kiss_fft_cpx * f,
-        int fstride,
+        const int fstride,
         int in_stride,
         int * factors,
         const kiss_fft_state * st