3 * Optimized version of the standard copy_page() function
6 * in0: address of target page
7 * in1: address of source page
11 * Copyright (C) 1999, 2001 Hewlett-Packard Co
12 * Stephane Eranian <eranian@hpl.hp.com>
13 * David Mosberger <davidm@hpl.hp.com>
15 * 4/06/01 davidm Tuned to make it perform well both for cached and uncached copies.
17 #include <asm/asmmacro.h>
19 #include <asm/export.h>
22 #define EPI p[PIPE_DEPTH-1]
36 #define Nrot ((8*PIPE_DEPTH+7)&~7)
38 GLOBAL_ENTRY(copy_page)
40 .save ar.pfs, saved_pfs
41 alloc saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
43 .rotr t1[PIPE_DEPTH], t2[PIPE_DEPTH], t3[PIPE_DEPTH], t4[PIPE_DEPTH], \
44 t5[PIPE_DEPTH], t6[PIPE_DEPTH], t7[PIPE_DEPTH], t8[PIPE_DEPTH]
51 mov lcount=PAGE_SIZE/64-1
60 mov tgt_last = PAGE_SIZE
67 add tgt_last = tgt_last, in0
70 (p[0]) ld8 t1[0]=[src1],16
71 (EPI) st8 [tgt1]=t1[PIPE_DEPTH-1],16
72 (p[0]) ld8 t2[0]=[src2],16
73 (EPI) st8 [tgt2]=t2[PIPE_DEPTH-1],16
74 cmp.ltu p6,p0 = tgtf, tgt_last
76 (p[0]) ld8 t3[0]=[src1],16
77 (EPI) st8 [tgt1]=t3[PIPE_DEPTH-1],16
78 (p[0]) ld8 t4[0]=[src2],16
79 (EPI) st8 [tgt2]=t4[PIPE_DEPTH-1],16
81 (p[0]) ld8 t5[0]=[src1],16
82 (EPI) st8 [tgt1]=t5[PIPE_DEPTH-1],16
83 (p[0]) ld8 t6[0]=[src2],16
84 (EPI) st8 [tgt2]=t6[PIPE_DEPTH-1],16
86 (p[0]) ld8 t7[0]=[src1],16
87 (EPI) st8 [tgt1]=t7[PIPE_DEPTH-1],16
88 (p[0]) ld8 t8[0]=[src2],16
89 (EPI) st8 [tgt2]=t8[PIPE_DEPTH-1],16
91 (p6) lfetch [srcf], 64
92 (p6) lfetch [tgtf], 64
95 mov pr=saved_pr,0xffffffffffff0000 // restore predicates
100 EXPORT_SYMBOL(copy_page)