Re: [PATCH v2] aarch64: thunderx2 memcpy branches reordering

[Wilco Dijkstra <Wilco dot Dijkstra at arm dot com>]

Hi Anton,

> Rewrote the branches in load and merge chunk
> so that the order is more in line with the 
> most probable case.

Why not make the code more efficient as well? The loop has 2 redundant
moves and the code for each alignment is twice as large as it could be.
It's not a large amount of effort to do better, see below.

@@ -557,17 +558,9 @@ L(ext_size_ ## shft):;\
 	ext     A_v.16b, C_v.16b, D_v.16b, 16-shft;\
 	ext     B_v.16b, D_v.16b, E_v.16b, 16-shft;\

These instructions are already part of the main loop, so why not directly
branch into it rather than repeat them?

 	subs    count, count, 32;\
-	b.ge    2f;\
+	b.lt    2f;\

This branch is completely redundant since count is always larger than 32.

 1:;\
 	stp     A_q, B_q, [dst], #32;\
-	ext     H_v.16b, E_v.16b, F_v.16b, 16-shft;\
-	ext     I_v.16b, F_v.16b, G_v.16b, 16-shft;\
-	stp     H_q, I_q, [dst], #16;\
-	add     dst, dst, tmp1;\
-	str     G_q, [dst], #16;\
-	b       L(copy_long_check32);\
-2:;\
-	stp     A_q, B_q, [dst], #32;\
 	prfm    pldl1strm, [src, MEMCPY_PREFETCH_LDR];\
 	ldp     D_q, J_q, [src], #32;\
 	ext     H_v.16b, E_v.16b, F_v.16b, 16-shft;\
@@ -579,8 +572,15 @@ L(ext_size_ ## shft):;\
 	ext     B_v.16b, D_v.16b, J_v.16b, 16-shft;\
 	mov     E_v.16b, J_v.16b;\

Redundant move in loop (2x). You might as well execute the next ext!

 	subs    count, count, 64;\
-	b.ge    2b;\
-	b	1b;\
+	b.ge    1b;\
+2:;\
+	stp     A_q, B_q, [dst], #32;\
+	ext     H_v.16b, E_v.16b, F_v.16b, 16-shft;\
+	ext     I_v.16b, F_v.16b, G_v.16b, 16-shft;\

These instructions appear in the loop already, why not break out of
the loop after executing them?

+	stp     H_q, I_q, [dst], #16;\
+	add     dst, dst, tmp1;\
+	str     G_q, [dst], #16;\
+	b       L(copy_long_check32);\
 
These instructions are invariant with respect to alignment, so why repeat
these many times?

Wilco