S390: Optimize __memcpy_z196.

This patch introduces an extra loop without pfd instructions
as it turned out that the pfd instructions are usefull
for copies >=64KB but are counterproductive for smaller copies.

diff --git a/sysdeps/s390/memcpy-z900.S b/sysdeps/s390/memcpy-z900.S
index f2e9aae..dc2f491 100644 (file)
--- a/sysdeps/s390/memcpy-z900.S
+++ b/sysdeps/s390/memcpy-z900.S
@@ -184,25 +184,34 @@ ENTRY(MEMCPY_Z196)
        je      .L_Z196_4
 .L_Z196_start2:
        aghi    %r4,-1
-       srlg    %r5,%r4,8
-       ltgr    %r5,%r5
+       risbg   %r5,%r4,8,128+63,56 # r0 = r5 / 256
        jne     .L_Z196_5
 .L_Z196_3:
        exrl    %r4,.L_Z196_14
 .L_Z196_4:
        br      %r14
 .L_Z196_5:
-       cgfi    %r5,262144      # Switch to mvcle for copies >64MB
-       jh      __memcpy_mvcle
+       cgfi    %r5,255         # Switch to loop with pfd for copies >=64kB
+       jh      .L_Z196_6
 .L_Z196_2:
-       pfd     1,768(%r3)
-       pfd     2,768(%r1)
        mvc     0(256,%r1),0(%r3)
        aghi    %r5,-1
        la      %r1,256(%r1)
        la      %r3,256(%r3)
        jne     .L_Z196_2
        j       .L_Z196_3
+.L_Z196_6:
+       cgfi    %r5,262144      # Switch to mvcle for copies >64MB
+       jh      __memcpy_mvcle
+.L_Z196_7:
+       pfd     1,1024(%r3)
+       pfd     2,1024(%r1)
+       mvc     0(256,%r1),0(%r3)
+       aghi    %r5,-1
+       la      %r1,256(%r1)
+       la      %r3,256(%r3)
+       jne     .L_Z196_7
+       j       .L_Z196_3
 .L_Z196_14:
        mvc     0(1,%r1),0(%r3)
 END(MEMCPY_Z196)