Hi Segher,

On 28/07/15 19:04, Segher Boessenkool wrote:

On Tue, Jul 28, 2015 at 12:19:54PM +0200, Thomas Huth wrote:

quoted

 : invert-region ( addr len -- )
-   0 ?DO dup dup rb@ -1 xor swap rb! 1+ LOOP drop
-;
-
-: invert-region-x ( addr len -- )
-   /x / 0 ?DO dup dup rx@ -1 xor swap rx! xa1+ LOOP drop
+   2dup or 7 and CASE
+      0 OF 3 rshift 0 ?DO dup dup rx@ -1 xor swap rx! xa1+ LOOP ENDOF
+      2 OF 1 rshift 0 ?DO dup dup rw@ -1 xor swap rw! wa1+ LOOP ENDOF
+      4 OF 2 rshift 0 ?DO dup dup rl@ -1 xor swap rl! la1+ LOOP ENDOF
+      6 OF 1 rshift 0 ?DO dup dup rw@ -1 xor swap rw! wa1+ LOOP ENDOF
+      dup OF 0 ?DO dup dup rb@ -1 xor swap rb! 1+ LOOP ENDOF
+   ENDCASE
+   drop
 ;

Can you access device memory as 64 bits for all supported devices?

Yes, should be fine since 64 bit access was already used in the original
code, see fb8-invert-screen in 
https://github.com/aik/SLOF/commit/99c534ecc7a8566bd9ca6346915d9ac1bfacae1e

You can get a bigger speedup by writing some of the core blitting
functions in C, btw.

Well, the above code is for js2x only ... so this is likely not worth
the effort anymore. The code for qemu-spapr calls into a hypercall
already, so this is already accelerated.

A small simplification:

   2dup or 7 and CASE
      0 OF 3 rshift 0 ?DO dup dup rx@ -1 xor swap rx! xa1+ LOOP ENDOF
      4 OF 2 rshift 0 ?DO dup dup rl@ -1 xor swap rl! la1+ LOOP ENDOF
      3 and
      2 OF 1 rshift 0 ?DO dup dup rw@ -1 xor swap rw! wa1+ LOOP ENDOF
      dup OF 0 ?DO dup dup rb@ -1 xor swap rb! 1+ LOOP ENDOF
   ENDCASE

Ok, nice idea, makes sense! I'll include it in v2 (after waiting a little
bit to see if there's other feedback)
 

If this code is often called unaligned, it makes more sense to special-
case the begin and end probably.

It's only used for drawing the cursor, so it always should be aligned.

 Thomas