1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
|
Copyright 2000-2005 Free Software Foundation, Inc.
This file is part of the GNU MP Library.
The GNU MP Library is free software; you can redistribute it and/or modify
it under the terms of either:
* the GNU Lesser General Public License as published by the Free
Software Foundation; either version 3 of the License, or (at your
option) any later version.
or
* the GNU General Public License as published by the Free Software
Foundation; either version 2 of the License, or (at your option) any
later version.
or both in parallel, as here.
The GNU MP Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received copies of the GNU General Public License and the
GNU Lesser General Public License along with the GNU MP Library. If not,
see https://www.gnu.org/licenses/.
IA-64 MPN SUBROUTINES
This directory contains mpn functions for the IA-64 architecture.
CODE ORGANIZATION
mpn/ia64 itanium-2, and generic ia64
The code here has been optimized primarily for Itanium 2. Very few Itanium 1
chips were ever sold, and Itanium 2 is more powerful, so the latter is what
we concentrate on.
CHIP NOTES
The IA-64 ISA keeps instructions three and three in 128 bit bundles.
Programmers/compilers need to put explicit breaks `;;' when there are WAW or
RAW dependencies, with some notable exceptions. Such "breaks" are typically
at the end of a bundle, but can be put between operations within some bundle
types too.
The Itanium 1 and Itanium 2 implementations can under ideal conditions
execute two bundles per cycle. The Itanium 1 allows 4 of these instructions
to do integer operations, while the Itanium 2 allows all 6 to be integer
operations.
Taken cloop branches seem to insert a bubble into the pipeline most of the
time on Itanium 1.
Loads to the fp registers bypass the L1 cache and thus get extremely long
latencies, 9 cycles on the Itanium 1 and 6 cycles on the Itanium 2.
The software pipeline stuff using br.ctop instruction causes delays, since
many issue slots are taken up by instructions with zero predicates, and
since many extra instructions are needed to set things up. These features
are clearly designed for code density, not speed.
Misc pipeline limitations (Itanium 1):
* The getf.sig instruction can only execute in M0.
* At most four integer instructions/cycle.
* Nops take up resources like any plain instructions.
Misc pipeline limitations (Itanium 2):
* The getf.sig instruction can only execute in M0.
* Nops take up resources like any plain instructions.
ASSEMBLY SYNTAX
.align pads with nops in a text segment, but gas 2.14 and earlier
incorrectly byte-swaps its nop bundle in big endian mode (eg. hpux), making
it come out as break instructions. We use the ALIGN() macro in
mpn/ia64/ia64-defs.m4 when it might be executed across. That macro
suppresses any .align if the problem is detected by configure. Lack of
alignment might hurt performance but will at least be correct.
foo:: to create a global symbol is not accepted by gas. Use separate
".global foo" and "foo:" instead.
.global is the standard global directive. gas accepts .globl, but hpux "as"
doesn't.
.proc / .endp generates the appropriate .type and .size information for ELF,
so the latter directives don't need to be given explicitly.
.pred.rel "mutex"... is standard for annotating predicate register
relationships. gas also accepts .pred.rel.mutex, but hpux "as" doesn't.
.pred directives can't be put on a line with a label, like
".Lfoo: .pred ...", the HP assembler on HP-UX 11.23 rejects that.
gas is happy with it, and past versions of HP had seemed ok.
// is the standard comment sequence, but we prefer "C" since it inhibits m4
macro expansion. See comments in ia64-defs.m4.
REGISTER USAGE
Special:
r0: constant 0
r1: global pointer (gp)
r8: return value
r12: stack pointer (sp)
r13: thread pointer (tp)
Caller-saves: r8-r11 r14-r31 f6-f15 f32-f127
Caller-saves but rotating: r32-
================================================================
mpn_add_n, mpn_sub_n:
The current code runs at 1.25 c/l on Itanium 2.
================================================================
mpn_mul_1:
The current code runs at 2 c/l on Itanium 2.
Using a blocked approach, working off of 4 separate places in the operands,
one could make use of the xma accumulation, and approach 1 c/l.
ldf8 [up]
xma.l
xma.hu
stf8 [wrp]
================================================================
mpn_addmul_1:
The current code runs at 2 c/l on Itanium 2.
It seems possible to use a blocked approach, as with mpn_mul_1. We should
read rp[] to integer registers, allowing for just one getf.sig per cycle.
ld8 [rp]
ldf8 [up]
xma.l
xma.hu
getf.sig
add+add+cmp+cmp
st8 [wrp]
These 10 instructions can be scheduled to approach 1.667 cycles, and with
the 4 cycle latency of xma, this means we need at least 3 blocks. Using
ldfp8 we could approach 1.583 c/l.
================================================================
mpn_submul_1:
The current code runs at 2.25 c/l on Itanium 2. Getting to 2 c/l requires
ldfp8 with all alignment headache that implies.
================================================================
mpn_addmul_N
For best speed, we need to give up using mpn_addmul_2 as the main multiply
building block, and instead take multiple v limbs per loop. For the Itanium
1, we need to take about 8 limbs at a time for full speed. For the Itanium
2, something like mpn_addmul_4 should be enough.
The add+cmp+cmp+add we use on the other codes is optimal for shortening
recurrencies (1 cycle) but the sequence takes up 4 execution slots. When
recurrency depth is not critical, a more standard 3-cycle add+cmp+add is
better.
/* First load the 8 values from v */
ldfp8 v0, v1 = [r35], 16;;
ldfp8 v2, v3 = [r35], 16;;
ldfp8 v4, v5 = [r35], 16;;
ldfp8 v6, v7 = [r35], 16;;
/* In the inner loop, get a new U limb and store a result limb. */
mov lc = un
Loop: ldf8 u0 = [r33], 8
ld8 r0 = [r32]
xma.l lp0 = v0, u0, hp0
xma.hu hp0 = v0, u0, hp0
xma.l lp1 = v1, u0, hp1
xma.hu hp1 = v1, u0, hp1
xma.l lp2 = v2, u0, hp2
xma.hu hp2 = v2, u0, hp2
xma.l lp3 = v3, u0, hp3
xma.hu hp3 = v3, u0, hp3
xma.l lp4 = v4, u0, hp4
xma.hu hp4 = v4, u0, hp4
xma.l lp5 = v5, u0, hp5
xma.hu hp5 = v5, u0, hp5
xma.l lp6 = v6, u0, hp6
xma.hu hp6 = v6, u0, hp6
xma.l lp7 = v7, u0, hp7
xma.hu hp7 = v7, u0, hp7
getf.sig l0 = lp0
getf.sig l1 = lp1
getf.sig l2 = lp2
getf.sig l3 = lp3
getf.sig l4 = lp4
getf.sig l5 = lp5
getf.sig l6 = lp6
add+cmp+add xx, l0, r0
add+cmp+add acc0, acc1, l1
add+cmp+add acc1, acc2, l2
add+cmp+add acc2, acc3, l3
add+cmp+add acc3, acc4, l4
add+cmp+add acc4, acc5, l5
add+cmp+add acc5, acc6, l6
getf.sig acc6 = lp7
st8 [r32] = xx, 8
br.cloop Loop
49 insn at max 6 insn/cycle: 8.167 cycles/limb8
11 memops at max 2 memops/cycle: 5.5 cycles/limb8
16 fpops at max 2 fpops/cycle: 8 cycles/limb8
21 intops at max 4 intops/cycle: 5.25 cycles/limb8
11+21 memops+intops at max 4/cycle 8 cycles/limb8
================================================================
mpn_lshift, mpn_rshift
The current code runs at 1 cycle/limb on Itanium 2.
Using 63 separate loops, we could use the double-word shrp instruction.
That instruction has a plain single-cycle latency. We need 63 loops since
this instruction only accept immediate count. That would lead to a somewhat
silly code size, but the speed would be 0.75 c/l on Itanium 2 (by using shrp
each cycle plus shl/shr going down I1 for a further limb every second
cycle).
================================================================
mpn_copyi, mpn_copyd
The current code runs at 0.5 c/l on Itanium 2. But that is just for L1
cache hit. The 4-way unrolled loop takes just 2 cycles, and thus load-use
scheduling isn't great. It might be best to actually use modulo scheduled
loops, since that will allow us to do better load-use scheduling without too
much unrolling.
Depending on size or operand alignment, we get 1 c/l or 0.5 c/l on Itanium
2, according to tune/speed. Cache bank conflicts?
REFERENCES
Intel Itanium Architecture Software Developer's Manual, volumes 1 to 3,
Intel document 245317-004, 245318-004, 245319-004 October 2002. Volume 1
includes an Itanium optimization guide.
Intel Itanium Processor-specific Application Binary Interface (ABI), Intel
document 245370-003, May 2001. Describes C type sizes, dynamic linking,
etc.
Intel Itanium Architecture Assembly Language Reference Guide, Intel document
248801-004, 2000-2002. Describes assembly instruction syntax and other
directives.
Itanium Software Conventions and Runtime Architecture Guide, Intel document
245358-003, May 2001. Describes calling conventions, including stack
unwinding requirements.
Intel Itanium Processor Reference Manual for Software Optimization, Intel
document 245473-003, November 2001.
Intel Itanium-2 Processor Reference Manual for Software Development and
Optimization, Intel document 251110-003, May 2004.
All the above documents can be found online at
http://developer.intel.com/design/itanium/manuals.htm
|