1 |
12 |
jlechner |
;; Scheduling description for UltraSPARC-I/II.
|
2 |
|
|
;; Copyright (C) 2002, 2004 Free Software Foundation, Inc.
|
3 |
|
|
;;
|
4 |
|
|
;; This file is part of GCC.
|
5 |
|
|
;;
|
6 |
|
|
;; GCC is free software; you can redistribute it and/or modify
|
7 |
|
|
;; it under the terms of the GNU General Public License as published by
|
8 |
|
|
;; the Free Software Foundation; either version 2, or (at your option)
|
9 |
|
|
;; any later version.
|
10 |
|
|
;;
|
11 |
|
|
;; GCC is distributed in the hope that it will be useful,
|
12 |
|
|
;; but WITHOUT ANY WARRANTY; without even the implied warranty of
|
13 |
|
|
;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
14 |
|
|
;; GNU General Public License for more details.
|
15 |
|
|
;;
|
16 |
|
|
;; You should have received a copy of the GNU General Public License
|
17 |
|
|
;; along with GCC; see the file COPYING. If not, write to
|
18 |
|
|
;; the Free Software Foundation, 51 Franklin Street, Fifth Floor,
|
19 |
|
|
;; Boston, MA 02110-1301, USA.
|
20 |
|
|
|
21 |
|
|
;; UltraSPARC-I and II are quad-issue processors. Interesting features
|
22 |
|
|
;; to note:
|
23 |
|
|
;;
|
24 |
|
|
;; - Buffered loads, they can queue waiting for the actual data until
|
25 |
|
|
;; an instruction actually tries to reference the destination register
|
26 |
|
|
;; as an input
|
27 |
|
|
;; - Two integer units. Only one of them can do shifts, and the other
|
28 |
|
|
;; is the only one which may do condition code setting instructions.
|
29 |
|
|
;; Complicating things further, a shift may go only into the first
|
30 |
|
|
;; slot in a dispatched group. And if you have a non-condition code
|
31 |
|
|
;; setting instruction and one that does set the condition codes. The
|
32 |
|
|
;; former must be issued first in order for both of them to issue.
|
33 |
|
|
;; - Stores can issue before the value being stored is available. As long
|
34 |
|
|
;; as the input data becomes ready before the store is to move out of the
|
35 |
|
|
;; store buffer, it will not cause a stall.
|
36 |
|
|
;; - Branches may issue in the same cycle as an instruction setting the
|
37 |
|
|
;; condition codes being tested by that branch. This does not apply
|
38 |
|
|
;; to floating point, only integer.
|
39 |
|
|
|
40 |
|
|
(define_automaton "ultrasparc_0,ultrasparc_1")
|
41 |
|
|
|
42 |
|
|
(define_cpu_unit "us1_fdivider,us1_fpm" "ultrasparc_0");
|
43 |
|
|
(define_cpu_unit "us1_fpa,us1_load_writeback" "ultrasparc_1")
|
44 |
|
|
(define_cpu_unit "us1_fps_0,us1_fps_1,us1_fpd_0,us1_fpd_1" "ultrasparc_1")
|
45 |
|
|
(define_cpu_unit "us1_slot0,us1_slot1,us1_slot2,us1_slot3" "ultrasparc_1")
|
46 |
|
|
(define_cpu_unit "us1_ieu0,us1_ieu1,us1_cti,us1_lsu" "ultrasparc_1")
|
47 |
|
|
|
48 |
|
|
(define_reservation "us1_slot012" "(us1_slot0 | us1_slot1 | us1_slot2)")
|
49 |
|
|
(define_reservation "us1_slotany" "(us1_slot0 | us1_slot1 | us1_slot2 | us1_slot3)")
|
50 |
|
|
(define_reservation "us1_single_issue" "us1_slot0 + us1_slot1 + us1_slot2 + us1_slot3")
|
51 |
|
|
|
52 |
|
|
(define_reservation "us1_fp_single" "(us1_fps_0 | us1_fps_1)")
|
53 |
|
|
(define_reservation "us1_fp_double" "(us1_fpd_0 | us1_fpd_1)")
|
54 |
|
|
|
55 |
|
|
;; This is a simplified representation of the issue at hand.
|
56 |
|
|
;; For most cases, going from one FP precision type insn to another
|
57 |
|
|
;; just breaks up the insn group. However for some cases, such
|
58 |
|
|
;; a situation causes the second insn to stall 2 more cycles.
|
59 |
|
|
(exclusion_set "us1_fps_0,us1_fps_1" "us1_fpd_0,us1_fpd_1")
|
60 |
|
|
|
61 |
|
|
;; If we have to schedule an ieu1 specific instruction and we want
|
62 |
|
|
;; to reserve the ieu0 unit as well, we must reserve it first. So for
|
63 |
|
|
;; example we could not schedule this sequence:
|
64 |
|
|
;; COMPARE IEU1
|
65 |
|
|
;; IALU IEU0
|
66 |
|
|
;; but we could schedule them together like this:
|
67 |
|
|
;; IALU IEU0
|
68 |
|
|
;; COMPARE IEU1
|
69 |
|
|
;; This basically requires that ieu0 is reserved before ieu1 when
|
70 |
|
|
;; it is required that both be reserved.
|
71 |
|
|
(absence_set "us1_ieu0" "us1_ieu1")
|
72 |
|
|
|
73 |
|
|
;; This defines the slotting order. Most IEU instructions can only
|
74 |
|
|
;; execute in the first three slots, FPU and branches can go into
|
75 |
|
|
;; any slot. We represent instructions which "break the group"
|
76 |
|
|
;; as requiring reservation of us1_slot0.
|
77 |
|
|
(absence_set "us1_slot0" "us1_slot1,us1_slot2,us1_slot3")
|
78 |
|
|
(absence_set "us1_slot1" "us1_slot2,us1_slot3")
|
79 |
|
|
(absence_set "us1_slot2" "us1_slot3")
|
80 |
|
|
|
81 |
|
|
(define_insn_reservation "us1_single" 1
|
82 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
83 |
|
|
(eq_attr "type" "multi,savew,flushw,iflush,trap"))
|
84 |
|
|
"us1_single_issue")
|
85 |
|
|
|
86 |
|
|
(define_insn_reservation "us1_simple_ieuN" 1
|
87 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
88 |
|
|
(eq_attr "type" "ialu"))
|
89 |
|
|
"(us1_ieu0 | us1_ieu1) + us1_slot012")
|
90 |
|
|
|
91 |
|
|
(define_insn_reservation "us1_simple_ieu0" 1
|
92 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
93 |
|
|
(eq_attr "type" "shift"))
|
94 |
|
|
"us1_ieu0 + us1_slot012")
|
95 |
|
|
|
96 |
|
|
(define_insn_reservation "us1_simple_ieu1" 1
|
97 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
98 |
|
|
(eq_attr "type" "compare"))
|
99 |
|
|
"us1_ieu1 + us1_slot012")
|
100 |
|
|
|
101 |
|
|
(define_insn_reservation "us1_ialuX" 1
|
102 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
103 |
|
|
(eq_attr "type" "ialuX"))
|
104 |
|
|
"us1_single_issue")
|
105 |
|
|
|
106 |
|
|
(define_insn_reservation "us1_cmove" 2
|
107 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
108 |
|
|
(eq_attr "type" "cmove"))
|
109 |
|
|
"us1_single_issue, nothing")
|
110 |
|
|
|
111 |
|
|
(define_insn_reservation "us1_imul" 1
|
112 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
113 |
|
|
(eq_attr "type" "imul"))
|
114 |
|
|
"us1_single_issue")
|
115 |
|
|
|
116 |
|
|
(define_insn_reservation "us1_idiv" 1
|
117 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
118 |
|
|
(eq_attr "type" "idiv"))
|
119 |
|
|
"us1_single_issue")
|
120 |
|
|
|
121 |
|
|
;; For loads, the "delayed return mode" behavior of the chip
|
122 |
|
|
;; is represented using the us1_load_writeback resource.
|
123 |
|
|
(define_insn_reservation "us1_load" 2
|
124 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
125 |
|
|
(eq_attr "type" "load,fpload"))
|
126 |
|
|
"us1_lsu + us1_slot012, us1_load_writeback")
|
127 |
|
|
|
128 |
|
|
(define_insn_reservation "us1_load_signed" 3
|
129 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
130 |
|
|
(eq_attr "type" "sload"))
|
131 |
|
|
"us1_lsu + us1_slot012, nothing, us1_load_writeback")
|
132 |
|
|
|
133 |
|
|
(define_insn_reservation "us1_store" 1
|
134 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
135 |
|
|
(eq_attr "type" "store,fpstore"))
|
136 |
|
|
"us1_lsu + us1_slot012")
|
137 |
|
|
|
138 |
|
|
(define_insn_reservation "us1_branch" 1
|
139 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
140 |
|
|
(eq_attr "type" "branch"))
|
141 |
|
|
"us1_cti + us1_slotany")
|
142 |
|
|
|
143 |
|
|
(define_insn_reservation "us1_call_jmpl" 1
|
144 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
145 |
|
|
(eq_attr "type" "call,sibcall,call_no_delay_slot,uncond_branch"))
|
146 |
|
|
"us1_cti + us1_ieu1 + us1_slot0")
|
147 |
|
|
|
148 |
|
|
(define_insn_reservation "us1_fmov_single" 1
|
149 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
150 |
|
|
(eq_attr "type" "fpmove"))
|
151 |
|
|
(eq_attr "fptype" "single"))
|
152 |
|
|
"us1_fpa + us1_fp_single + us1_slotany")
|
153 |
|
|
|
154 |
|
|
(define_insn_reservation "us1_fmov_double" 1
|
155 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
156 |
|
|
(eq_attr "type" "fpmove"))
|
157 |
|
|
(eq_attr "fptype" "double"))
|
158 |
|
|
"us1_fpa + us1_fp_double + us1_slotany")
|
159 |
|
|
|
160 |
|
|
(define_insn_reservation "us1_fcmov_single" 2
|
161 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
162 |
|
|
(eq_attr "type" "fpcmove,fpcrmove"))
|
163 |
|
|
(eq_attr "fptype" "single"))
|
164 |
|
|
"us1_fpa + us1_fp_single + us1_slotany, nothing")
|
165 |
|
|
|
166 |
|
|
(define_insn_reservation "us1_fcmov_double" 2
|
167 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
168 |
|
|
(eq_attr "type" "fpcmove,fpcrmove"))
|
169 |
|
|
(eq_attr "fptype" "double"))
|
170 |
|
|
"us1_fpa + us1_fp_double + us1_slotany, nothing")
|
171 |
|
|
|
172 |
|
|
(define_insn_reservation "us1_faddsub_single" 4
|
173 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
174 |
|
|
(eq_attr "type" "fp"))
|
175 |
|
|
(eq_attr "fptype" "single"))
|
176 |
|
|
"us1_fpa + us1_fp_single + us1_slotany, nothing*3")
|
177 |
|
|
|
178 |
|
|
(define_insn_reservation "us1_faddsub_double" 4
|
179 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
180 |
|
|
(eq_attr "type" "fp"))
|
181 |
|
|
(eq_attr "fptype" "double"))
|
182 |
|
|
"us1_fpa + us1_fp_double + us1_slotany, nothing*3")
|
183 |
|
|
|
184 |
|
|
(define_insn_reservation "us1_fpcmp_single" 1
|
185 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
186 |
|
|
(eq_attr "type" "fpcmp"))
|
187 |
|
|
(eq_attr "fptype" "single"))
|
188 |
|
|
"us1_fpa + us1_fp_single + us1_slotany")
|
189 |
|
|
|
190 |
|
|
(define_insn_reservation "us1_fpcmp_double" 1
|
191 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
192 |
|
|
(eq_attr "type" "fpcmp"))
|
193 |
|
|
(eq_attr "fptype" "double"))
|
194 |
|
|
"us1_fpa + us1_fp_double + us1_slotany")
|
195 |
|
|
|
196 |
|
|
(define_insn_reservation "us1_fmult_single" 4
|
197 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
198 |
|
|
(eq_attr "type" "fpmul"))
|
199 |
|
|
(eq_attr "fptype" "single"))
|
200 |
|
|
"us1_fpm + us1_fp_single + us1_slotany, nothing*3")
|
201 |
|
|
|
202 |
|
|
(define_insn_reservation "us1_fmult_double" 4
|
203 |
|
|
(and (and (eq_attr "cpu" "ultrasparc")
|
204 |
|
|
(eq_attr "type" "fpmul"))
|
205 |
|
|
(eq_attr "fptype" "double"))
|
206 |
|
|
"us1_fpm + us1_fp_double + us1_slotany, nothing*3")
|
207 |
|
|
|
208 |
|
|
;; This is actually in theory dangerous, because it is possible
|
209 |
|
|
;; for the chip to prematurely dispatch the dependent instruction
|
210 |
|
|
;; in the G stage, resulting in a 9 cycle stall. However I have never
|
211 |
|
|
;; been able to trigger this case myself even with hand written code,
|
212 |
|
|
;; so it must require some rare complicated pipeline state.
|
213 |
|
|
(define_bypass 3
|
214 |
|
|
"us1_faddsub_single,us1_faddsub_double,us1_fmult_single,us1_fmult_double"
|
215 |
|
|
"us1_faddsub_single,us1_faddsub_double,us1_fmult_single,us1_fmult_double")
|
216 |
|
|
|
217 |
|
|
;; Floating point divide and square root use the multiplier unit
|
218 |
|
|
;; for final rounding 3 cycles before the divide/sqrt is complete.
|
219 |
|
|
|
220 |
|
|
(define_insn_reservation "us1_fdivs"
|
221 |
|
|
13
|
222 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
223 |
|
|
(eq_attr "type" "fpdivs,fpsqrts"))
|
224 |
|
|
"(us1_fpm + us1_fdivider + us1_slot0), us1_fdivider*8, (us1_fpm + us1_fdivider), us1_fdivider*2"
|
225 |
|
|
)
|
226 |
|
|
|
227 |
|
|
(define_bypass
|
228 |
|
|
12
|
229 |
|
|
"us1_fdivs"
|
230 |
|
|
"us1_faddsub_single,us1_faddsub_double,us1_fmult_single,us1_fmult_double")
|
231 |
|
|
|
232 |
|
|
(define_insn_reservation "us1_fdivd"
|
233 |
|
|
23
|
234 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
235 |
|
|
(eq_attr "type" "fpdivd,fpsqrtd"))
|
236 |
|
|
"(us1_fpm + us1_fdivider + us1_slot0), us1_fdivider*18, (us1_fpm + us1_fdivider), us1_fdivider*2"
|
237 |
|
|
)
|
238 |
|
|
(define_bypass
|
239 |
|
|
22
|
240 |
|
|
"us1_fdivd"
|
241 |
|
|
"us1_faddsub_single,us1_faddsub_double,us1_fmult_single,us1_fmult_double")
|
242 |
|
|
|
243 |
|
|
;; Any store may multi issue with the insn creating the source
|
244 |
|
|
;; data as long as that creating insn is not an FPU div/sqrt.
|
245 |
|
|
;; We need a special guard function because this bypass does
|
246 |
|
|
;; not apply to the address inputs of the store.
|
247 |
|
|
(define_bypass 0 "us1_simple_ieuN,us1_simple_ieu1,us1_simple_ieu0,us1_faddsub_single,us1_faddsub_double,us1_fmov_single,us1_fmov_double,us1_fcmov_single,us1_fcmov_double,us1_fmult_single,us1_fmult_double" "us1_store"
|
248 |
|
|
"store_data_bypass_p")
|
249 |
|
|
|
250 |
|
|
;; An integer branch may execute in the same cycle as the compare
|
251 |
|
|
;; creating the condition codes.
|
252 |
|
|
(define_bypass 0 "us1_simple_ieu1" "us1_branch")
|
253 |
|
|
|
254 |
|
|
;; VIS scheduling
|
255 |
|
|
(define_insn_reservation "us1_fga_single"
|
256 |
|
|
2
|
257 |
|
|
(and (and
|
258 |
|
|
(eq_attr "cpu" "ultrasparc")
|
259 |
|
|
(eq_attr "type" "fga"))
|
260 |
|
|
(eq_attr "fptype" "single"))
|
261 |
|
|
"us1_fpa + us1_fp_single + us1_slotany, nothing")
|
262 |
|
|
|
263 |
|
|
(define_bypass 1 "us1_fga_single" "us1_fga_single")
|
264 |
|
|
|
265 |
|
|
(define_insn_reservation "us1_fga_double"
|
266 |
|
|
2
|
267 |
|
|
(and (and
|
268 |
|
|
(eq_attr "cpu" "ultrasparc")
|
269 |
|
|
(eq_attr "type" "fga"))
|
270 |
|
|
(eq_attr "fptype" "double"))
|
271 |
|
|
"us1_fpa + us1_fp_double + us1_slotany, nothing")
|
272 |
|
|
|
273 |
|
|
(define_bypass 1 "us1_fga_double" "us1_fga_double")
|
274 |
|
|
|
275 |
|
|
(define_insn_reservation "us1_fgm_single"
|
276 |
|
|
4
|
277 |
|
|
(and (and
|
278 |
|
|
(eq_attr "cpu" "ultrasparc")
|
279 |
|
|
(eq_attr "type" "fgm_pack,fgm_mul,fgm_cmp"))
|
280 |
|
|
(eq_attr "fptype" "single"))
|
281 |
|
|
"us1_fpm + us1_fp_single + us1_slotany, nothing*3")
|
282 |
|
|
|
283 |
|
|
(define_bypass 3 "us1_fgm_single" "us1_fga_single")
|
284 |
|
|
|
285 |
|
|
(define_insn_reservation "us1_fgm_double"
|
286 |
|
|
4
|
287 |
|
|
(and (and
|
288 |
|
|
(eq_attr "cpu" "ultrasparc")
|
289 |
|
|
(eq_attr "type" "fgm_pack,fgm_mul,fgm_cmp"))
|
290 |
|
|
(eq_attr "fptype" "double"))
|
291 |
|
|
"us1_fpm + us1_fp_double + us1_slotany, nothing*3")
|
292 |
|
|
|
293 |
|
|
(define_bypass 3 "us1_fgm_double" "us1_fga_double")
|
294 |
|
|
|
295 |
|
|
(define_insn_reservation "us1_pdist"
|
296 |
|
|
4
|
297 |
|
|
(and (eq_attr "cpu" "ultrasparc")
|
298 |
|
|
(eq_attr "type" "fgm_pdist"))
|
299 |
|
|
"us1_fpm + us1_fp_double + us1_slotany, nothing*3")
|
300 |
|
|
|
301 |
|
|
(define_bypass 3 "us1_pdist" "us1_fga_double,us1_fga_single")
|
302 |
|
|
(define_bypass 1 "us1_pdist" "us1_pdist")
|