rllback: improve compression by allowing two-byte run lengths

With the command line size being set to 2K, it's common to need longer
runs.  Allow runs up to 8K.

diff --git a/rllpack.inc b/rllpack.inc
index e5d1d5e..a556e00 100644 (file)
--- a/rllpack.inc
+++ b/rllpack.inc
@@ -18,7 +18,8 @@
 ;
 ; Format of leading byte
 ; 1-128                = x verbatim bytes follow
-; 129-255      = (x-126) times subsequent byte
+; 129-223      = (x-126) times subsequent byte
+; 224-255      = (x-224)*256+(next byte) times the following byte
 ; 0            = end of data
 ;
 ; These structures are stored *in reverse order* in high memory.
@@ -46,8 +47,8 @@ rllpack:
                xor ebx,ebx             ; Run length zero
                dec edi
                mov edx,edi             ; Pointer to header byte
-               mov [edi],al
-               jcxz .done_null
+               mov [edi],al            ; Create header byte
+               jcxz .done              ; If done, this was the terminator
 .stdbyte:
                lodsb
                dec edi
@@ -61,7 +62,7 @@ rllpack:
 .plainbyte:
                inc bx
                inc byte [edx]
-               jcxz .done
+               jcxz .startseq
                jns .stdbyte
                jmp .startseq
 .same:
@@ -74,32 +75,35 @@ rllpack:
                                        ; drop start byte
 .normal:
                inc bx
-               add edi,ebx
-               mov al,bl
-               add al,126
-               dec edi
-               mov edx,edi
-               mov [edi],al
-               dec edi
-               mov [edi],ah
+               add edi,ebx             ; Remove the stored run bytes
 .getrun:
-               jcxz .done
-               cmp bl,255-126
-               jae .startseq
+               jcxz .nomatch
                lodsb
                cmp al,ah
                jne .nomatch
+               cmp bx,(256-224)*256-1  ; Maximum run size
+               jae .nomatch
                inc bx
-               inc byte [edx]
                dec cx
                jmp .getrun
 .nomatch:
-               dec si
+               cmp bx,224-126
+               jae .twobyte
+.onebyte:
+               add bl,126
+               dec edi
+               mov [edi],bl
+               jmp .storebyte
+.twobyte:
+               add bh,224
+               sub edi,2
+               mov [edi],bx
+.storebyte:
+               dec edi
+               mov [edi],ah
+               dec si                  ; Reload subsequent byte
                jmp .startseq
 .done:
-               dec edi
-               mov [edi],cl            ; CX = 0 here
-.done_null:
                pop edx
                pop ebx
                pop cx
@@ -119,13 +123,11 @@ rllunpack:
                xor cx,cx
 .header:
                dec esi
-               mov al,[esi]
-               and al,al
-               jz .done
-               cmp al,129
+               mov cl,[esi]
+               jcxz .done
+               cmp cl,129
                jae .isrun
                ; Not a run
-               mov cl,al
 .copy:
                dec esi
                mov al,[esi]
@@ -133,12 +135,20 @@ rllunpack:
                loop .copy
                jmp .header
 .isrun:
-               sub al,126
-               mov cl,al
+               cmp cl,224
+               jae .longrun
+               sub cl,126
+.dorun:
                dec esi
                mov al,[esi]
                rep stosb
                jmp .header
+.longrun:
+               sub cl,224
+               mov ch,cl
+               dec esi
+               mov cl,[esi]
+               jmp .dorun
 .done:
                pop cx
                sub cx,di