2 * McKinley-optimized version of copy_page().
4 * Copyright (C) 2002 Hewlett-Packard Co
5 * David Mosberger <davidm@hpl.hp.com>
8 * in0: address of target page
9 * in1: address of source page
14 * - use regular loads and stores to prefetch data to avoid consuming M-slot just for
15 * lfetches => good for in-cache performance
16 * - avoid l2 bank-conflicts by not storing into the same 16-byte bank within a single
19 * Principle of operation:
20 * First, note that L1 has a line-size of 64 bytes and L2 a line-size of 128 bytes.
21 * To avoid secondary misses in L2, we prefetch both source and destination with a line-size
22 * of 128 bytes. When both of these lines are in the L2 and the first half of the
23 * source line is in L1, we start copying the remaining words. The second half of the
24 * source line is prefetched in an earlier iteration, so that by the time we start
25 * accessing it, it's also present in the L1.
27 * We use a software-pipelined loop to control the overall operation. The pipeline
28 * has 2*PREFETCH_DIST+K stages. The first PREFETCH_DIST stages are used for prefetching
29 * source cache-lines. The second PREFETCH_DIST stages are used for prefetching destination
30 * cache-lines, the last K stages are used to copy the cache-line words not copied by
31 * the prefetches. The four relevant points in the pipelined are called A, B, C, D:
32 * p[A] is TRUE if a source-line should be prefetched, p[B] is TRUE if a destination-line
33 * should be prefetched, p[C] is TRUE if the second half of an L2 line should be brought
34 * into L1D and p[D] is TRUE if a cacheline needs to be copied.
36 * This all sounds very complicated, but thanks to the modulo-scheduled loop support,
37 * the resulting code is very regular and quite easy to follow (once you get the idea).
39 * As a secondary optimization, the first 2*PREFETCH_DIST iterations are implemented
40 * as the separate .prefetch_loop. Logically, this loop performs exactly like the
41 * main-loop (.line_copy), but has all known-to-be-predicated-off instructions removed,
42 * so that each loop iteration is faster (again, good for cached case).
44 * When reading the code, it helps to keep the following picture in mind:
51 * | t6 | t7 | | 128 bytes
52 * | n[y] | t9 | | (L2 cache line)
58 * Here, v[x] is copied by the (memory) prefetch. n[y] is loaded at p[C]
59 * to fetch the second-half of the L2 cache line into L1, and the tX words are copied in
60 * an order that avoids bank conflicts.
62 #include <asm/asmmacro.h>
65 #define PREFETCH_DIST 8 // McKinley sustains 16 outstanding L2 misses (8 ld, 8 st)
71 #define src_pre_mem r11
72 #define dst_pre_mem r14
73 #define src_pre_l2 r15
74 #define dst_pre_l2 r16
79 #define t5 t1 // alias!
80 #define t6 t2 // alias!
81 #define t7 t3 // alias!
82 #define t9 t5 // alias!
83 #define t10 t4 // alias!
84 #define t11 t7 // alias!
85 #define t12 t6 // alias!
86 #define t14 t10 // alias!
94 #define B (PREFETCH_DIST)
95 #define C (B + PREFETCH_DIST)
98 #define Nrot ((N + 7) & ~7)
100 GLOBAL_ENTRY(copy_page)
102 alloc r8 = ar.pfs, 2, Nrot-2, 0, Nrot
104 .rotr v[2*PREFETCH_DIST], n[D-C+1]
107 .save ar.lc, saved_lc
113 mov src_pre_mem = in1
115 mov ar.ec = 1 // special unrolled loop
117 mov dst_pre_mem = in0
118 mov ar.lc = 2*PREFETCH_DIST - 1
120 add src_pre_l2 = 8*8, in1
121 add dst_pre_l2 = 8*8, in0
122 add src0 = 8, in1 // first t1 src
123 add src1 = 3*8, in1 // first t3 src
124 add dst0 = 8, in0 // first t1 dst
125 add dst1 = 3*8, in0 // first t3 dst
126 mov t1 = (PAGE_SIZE/128) - (2*PREFETCH_DIST) - 1
130 // same as .line_copy loop, but with all predicated-off instructions removed:
132 (p[A]) ld8 v[A] = [src_pre_mem], 128 // M0
133 (p[B]) st8 [dst_pre_mem] = v[B], 128 // M2
134 br.ctop.sptk .prefetch_loop
136 cmp.eq p16, p0 = r0, r0 // reset p16 to 1 (br.ctop cleared it to zero)
137 mov ar.lc = t1 // with 64KB pages, t1 is too big to fit in 8 bits!
138 mov ar.ec = N // # of stages in pipeline
141 (p[D]) ld8 t2 = [src0], 3*8 // M0
142 (p[D]) ld8 t4 = [src1], 3*8 // M1
143 (p[B]) st8 [dst_pre_mem] = v[B], 128 // M2 prefetch dst from memory
144 (p[D]) st8 [dst_pre_l2] = n[D-C], 128 // M3 prefetch dst from L2
146 (p[A]) ld8 v[A] = [src_pre_mem], 128 // M0 prefetch src from memory
147 (p[C]) ld8 n[0] = [src_pre_l2], 128 // M1 prefetch src from L2
148 (p[D]) st8 [dst0] = t1, 8 // M2
149 (p[D]) st8 [dst1] = t3, 8 // M3
151 (p[D]) ld8 t5 = [src0], 8
152 (p[D]) ld8 t7 = [src1], 3*8
153 (p[D]) st8 [dst0] = t2, 3*8
154 (p[D]) st8 [dst1] = t4, 3*8
156 (p[D]) ld8 t6 = [src0], 3*8
157 (p[D]) ld8 t10 = [src1], 8
158 (p[D]) st8 [dst0] = t5, 8
159 (p[D]) st8 [dst1] = t7, 3*8
161 (p[D]) ld8 t9 = [src0], 3*8
162 (p[D]) ld8 t11 = [src1], 3*8
163 (p[D]) st8 [dst0] = t6, 3*8
164 (p[D]) st8 [dst1] = t10, 8
166 (p[D]) ld8 t12 = [src0], 8
167 (p[D]) ld8 t14 = [src1], 8
168 (p[D]) st8 [dst0] = t9, 3*8
169 (p[D]) st8 [dst1] = t11, 3*8
171 (p[D]) ld8 t13 = [src0], 4*8
172 (p[D]) ld8 t15 = [src1], 4*8
173 (p[D]) st8 [dst0] = t12, 8
174 (p[D]) st8 [dst1] = t14, 8
176 (p[D-1])ld8 t1 = [src0], 8
177 (p[D-1])ld8 t3 = [src1], 8
178 (p[D]) st8 [dst0] = t13, 4*8
179 (p[D]) st8 [dst1] = t15, 4*8
180 br.ctop.sptk .line_copy
183 mov pr = saved_pr, -1