add note about decision tree multiplication

diff --git a/foo.S b/foo.S
index bde793cb03c0f8ed7c3c9009e21255e51bcb4fbb..d89dec548a0ac1953fefeeb49f69a4422090554b 100644 (file)
--- a/foo.S
+++ b/foo.S
@@ -125,14 +125,20 @@ g: ; g(i, t) -> t
        #undef zero
        #undef tmp
 
-       ;TODO: check correctness!
        LDI     Xlo, lo8(notes)
        ADD     Xlo, t ; NOTE: can't overflow, since RAMEND == 0x5F
        LD      t, X
 
        CLR     a2
 
-       /* decision tree multiplication saves cycles and (hopefully) reduces code size
+       /* decision tree multiplication:
+          there is only a limited number of coefficients, so we heavily
+          optimize for those only, and only compute the bits we
+          actually need. this reduces cycle count from 38 for the
+          (optimized) classic approach to 31. instruction count
+          increases from 38 to 100. in the end it turned out that we
+          would've had enough cycles to spare to just use the standard
+          algorithm.
                             _xxx?
                         /           \
                   _xx?0                _xx1?
@@ -145,7 +151,8 @@ g: ; g(i, t) -> t
           |     |       |     |       |     |     |
          ...   ...     ...   ...     ...   ...   ...
           |     |       |     |       |     |     |
-          B0    58     84    8C      69     75    9D   */
+          B0    58     84    8C      69     75    9D
+         27cy  28cy   26cy  28cy    26cy   31cy  30cy  */
                test    0, m____1
        m____0: shift16
                never   1