From e1a453ce38d77f80c171b2752bedd1be54fd0eb4 Mon Sep 17 00:00:00 2001
From: Christian Starkjohann <cs+github@obdev.at>
Date: Fri, 22 Nov 2013 15:48:48 +0100
Subject: Optimized fast CRC routine to be shorter and faster.

---
 usbdrv/usbdrvasm.S | 37 ++++++++++++++++---------------------
 1 file changed, 16 insertions(+), 21 deletions(-)

diff --git a/usbdrv/usbdrvasm.S b/usbdrv/usbdrvasm.S
index 32ce8ef..6fa575d 100644
--- a/usbdrv/usbdrvasm.S
+++ b/usbdrv/usbdrvasm.S
@@ -145,7 +145,8 @@ RTMODEL "__rt_version", "3"
 #if USB_USE_FAST_CRC
 
 ; This implementation is faster, but has bigger code size
-; Thanks to Slawomir Fras (BoskiDialer) for this code!
+; Thanks to Slawomir Fras (BoskiDialer) for this code and to Shay Green for
+; even further optimizations!
 ; It implements the following C pseudo-code:
 ; unsigned table(unsigned char x)
 ; {
@@ -175,35 +176,29 @@ RTMODEL "__rt_version", "3"
 ;   resCrc  r24+r25 / r16+r17
 ;   ptr     X / Z
 usbCrc16:
-    mov     ptrL, argPtrL
-    mov     ptrH, argPtrH
+    movw    ptrL, argPtrL
     ldi     resCrcL, 0xFF
     ldi     resCrcH, 0xFF
+    clr     bitCnt          ; zero reg
     rjmp    usbCrc16LoopTest
 usbCrc16ByteLoop:
     ld      byte, ptr+
-    eor     resCrcL, byte   ; resCrcL is now 'x' in table()
-    mov     byte, resCrcL   ; compute parity of 'x'
+    eor     byte, resCrcL   ; scratch is now 'x' in table()
+    mov     scratch, byte   ; compute parity of 'x'
     swap    byte
-    eor     byte, resCrcL
-    mov     scratch, byte
+    eor     byte, scratch
+    mov     resCrcL, byte
     lsr     byte
     lsr     byte
-    eor     byte, scratch
+    eor     byte, resCrcL
     inc     byte
-    lsr     byte
-    andi    byte, 1         ; byte is now parity(x)
-    mov     scratch, resCrcL
-    mov     resCrcL, resCrcH
-    eor     resCrcL, byte   ; low byte of if(parity(x)) value ^= 0xc001;
-    neg     byte
-    andi    byte, 0xc0
-    mov     resCrcH, byte   ; high byte of if(parity(x)) value ^= 0xc001;
-    clr     byte
-    lsr     scratch
-    ror     byte
-    eor     resCrcH, scratch
-    eor     resCrcL, byte
+    andi    byte, 2        ; byte is now parity(x) << 1
+    cp      bitCnt, byte   ; c = (byte != 0), then put in high bit
+    ror     scratch        ; so that after xoring, shifting, and xoring, it gives
+    ror     byte           ; the desired 0xC0 with resCrcH
+    mov     resCrcL, byte
+    eor     resCrcL, resCrcH
+    mov     resCrcH, scratch
     lsr     scratch
     ror     byte
     eor     resCrcH, scratch
-- 
cgit v1.2.3