a bit faster, a bit smaller

asm/8088/decompress_small_v1.S

@@ -54,11 +54,12 @@ lzsa1_decompress:
    jne .mid_literals
 
    lodsw                   ; grab 16-bit extra length
-   jmp short .got_literals
+   .byte 81H               ; mask inc ah/lodsb
+                           ; (*like jmp short .got_literals but faster)
 
 .mid_literals:
-   lodsb                   ; grab single extra length byte
    inc ah                  ; add 256
+   lodsb                   ; grab single extra length byte
 
 .got_literals:
    xchg cx,ax
@@ -67,16 +68,17 @@ lzsa1_decompress:
    test dl,dl              ; check match offset size in token (O bit)
    js .get_long_offset
 
-   xchg ax,cx              ; clear ah - cx is zero from the rep movsb above
+   dec cx
+   xchg cx,ax              ; ah to 0xff - cx was zero from the rep movsb above
    lodsb
-   dec ah
-   jmp short .get_match_length
+   .byte 3CH               ; mask lodsw
+                           ; (*like jmp short .get_match_length but faster)
 
 .get_long_offset:
    lodsw                   ; Get 2-byte match offset
 
 .get_match_length:
-   xchg ax,dx              ; dx: match offset  ax: original token
+   xchg dx,ax              ; dx: match offset  ax: original token
    and al,0FH              ; isolate match length in token (MMMM)
    add al,3                ; add MIN_MATCH_SIZE
 
@@ -90,8 +92,13 @@ lzsa1_decompress:
 
    lodsw                   ; grab 16-bit length
    test ax,ax              ; bail if we hit EOD
-   je short .done_decompressing 
-   jmp short .got_matchlen
+   jne short .got_matchlen
+
+.done_decompressing:
+   pop ax                  ; retrieve the original decompression offset
+   xchg ax,di              ; compute decompressed size
+   sub ax,di
+   ret                     ; done
 
 .mid_matchlen:
    lodsb                   ; grab single extra length byte
@@ -109,9 +116,3 @@ lzsa1_decompress:
    xchg si,ax              ; restore ds:si
    pop ds
    jmp short .decode_token ; go decode another token
-
-.done_decompressing:
-   pop ax                  ; retrieve the original decompression offset
-   xchg ax,di              ; compute decompressed size
-   sub ax,di
-   ret                     ; done

asm/8088/decompress_small_v2.S

@@ -72,7 +72,7 @@ lzsa2_decompress:
    jnb .offset_9_bit
 
                            ; 5 bit offset
-   xchg ax,cx              ; clear ah - cx is zero from the rep movsb above
+   xchg cx,ax              ; clear ah - cx is zero from the rep movsb above
    mov al,020H             ; shift Z (offset bit 4) in place
    and al,dl
    shl al,1
@@ -85,7 +85,7 @@ lzsa2_decompress:
    jmp short .get_match_length
 
 .offset_9_bit:             ; 9 bit offset
-   xchg ax,cx              ; clear ah - cx is zero from the rep movsb above
+   xchg cx,ax              ; clear ah - cx is zero from the rep movsb above
    lodsb                   ; get 8 bit offset from stream in A
    dec ah                  ; set offset bits 15-8 to 1
    test dl,020H            ; test bit Z (offset bit 8)
@@ -120,9 +120,9 @@ lzsa2_decompress:
    xchg ah,al
 
 .get_match_length:
-   mov bp,ax               ; bp: offset
+   xchg bp,ax              ; bp: offset
 .repeat_match:
-   mov ax,dx               ; ax: original token
+   xchg ax,dx              ; ax: original token
    and al,07H              ; isolate match length in token (MMM)
    add al,2                ; add MIN_MATCH_SIZE_V2
 
@@ -156,7 +156,7 @@ lzsa2_decompress:
 
 .done_decompressing:
    pop ax                  ; retrieve the original decompression offset
-   xchg ax,di              ; compute decompressed size
+   xchg di,ax              ; compute decompressed size
    sub ax,di
    ret                     ; done
 
@@ -166,7 +166,7 @@ lzsa2_decompress:
    
    xchg bx,ax
    lodsb                   ; load two nibbles
-   xchg ax,bx
+   xchg bx,ax
 
 .has_nibble:
    mov cl,4                ; swap 4 high and low bits of nibble