0
|
1 ;; DFA scheduling description for SH4.
|
145
|
2 ;; Copyright (C) 2004-2020 Free Software Foundation, Inc.
|
0
|
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 3, 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 COPYING3. If not see
|
|
18 ;; <http://www.gnu.org/licenses/>.
|
|
19
|
|
20 ;; Load and store instructions save a cycle if they are aligned on a
|
|
21 ;; four byte boundary. Using a function unit for stores encourages
|
|
22 ;; gcc to separate load and store instructions by one instruction,
|
|
23 ;; which makes it more likely that the linker will be able to word
|
|
24 ;; align them when relaxing.
|
|
25
|
|
26 ;; The following description models the SH4 pipeline using the DFA based
|
|
27 ;; scheduler. The DFA based description is better way to model a
|
|
28 ;; superscalar pipeline as compared to function unit reservation model.
|
|
29 ;; 1. The function unit based model is oriented to describe at most one
|
|
30 ;; unit reservation by each insn. It is difficult to model unit reservations
|
|
31 ;; in multiple pipeline units by same insn. This can be done using DFA
|
|
32 ;; based description.
|
|
33 ;; 2. The execution performance of DFA based scheduler does not depend on
|
|
34 ;; processor complexity.
|
|
35 ;; 3. Writing all unit reservations for an instruction class is a more natural
|
|
36 ;; description of the pipeline and makes the interface to the hazard
|
|
37 ;; recognizer simpler than the old function unit based model.
|
|
38 ;; 4. The DFA model is richer and is a part of greater overall framework
|
|
39 ;; of RCSP.
|
|
40
|
|
41
|
|
42 ;; Two automata are defined to reduce number of states
|
|
43 ;; which a single large automaton will have. (Factoring)
|
|
44 (define_automaton "inst_pipeline,fpu_pipe")
|
|
45
|
|
46 ;; This unit is basically the decode unit of the processor.
|
|
47 ;; Since SH4 is a dual issue machine,it is as if there are two
|
|
48 ;; units so that any insn can be processed by either one
|
|
49 ;; of the decoding unit.
|
|
50 (define_cpu_unit "pipe_01,pipe_02" "inst_pipeline")
|
|
51
|
|
52
|
|
53 ;; The fixed point arithmetic calculator(?? EX Unit).
|
|
54 (define_cpu_unit "int" "inst_pipeline")
|
|
55
|
|
56 ;; f1_1 and f1_2 are floating point units.Actually there is
|
|
57 ;; a f1 unit which can overlap with other f1 unit but
|
|
58 ;; not another F1 unit.It is as though there were two
|
|
59 ;; f1 units.
|
|
60 (define_cpu_unit "f1_1,f1_2" "fpu_pipe")
|
|
61
|
|
62 ;; The floating point units (except FS - F2 always precedes it.)
|
|
63 (define_cpu_unit "F0,F1,F2,F3" "fpu_pipe")
|
|
64
|
|
65 ;; This is basically the MA unit of SH4
|
|
66 ;; used in LOAD/STORE pipeline.
|
|
67 (define_cpu_unit "memory" "inst_pipeline")
|
|
68
|
|
69 ;; However, there are LS group insns that don't use it, even ones that
|
|
70 ;; complete in 0 cycles. So we use an extra unit for the issue of LS insns.
|
|
71 (define_cpu_unit "load_store" "inst_pipeline")
|
|
72
|
|
73 ;; The address calculator used for branch instructions.
|
|
74 ;; This will be reserved after "issue" of branch instructions
|
|
75 ;; and this is to make sure that no two branch instructions
|
|
76 ;; can be issued in parallel.
|
|
77
|
|
78 (define_cpu_unit "pcr_addrcalc" "inst_pipeline")
|
|
79
|
|
80 ;; ----------------------------------------------------
|
|
81 ;; This reservation is to simplify the dual issue description.
|
|
82 (define_reservation "issue" "pipe_01|pipe_02")
|
|
83
|
|
84 ;; This is to express the locking of D stage.
|
|
85 ;; Note that the issue of a CO group insn also effectively locks the D stage.
|
|
86 (define_reservation "d_lock" "pipe_01+pipe_02")
|
|
87
|
|
88 ;; Every FE instruction but fipr / ftrv starts with issue and this.
|
|
89 (define_reservation "F01" "F0+F1")
|
|
90
|
|
91 ;; This is to simplify description where F1,F2,FS
|
|
92 ;; are used simultaneously.
|
|
93 (define_reservation "fpu" "F1+F2")
|
|
94
|
|
95 ;; This is to highlight the fact that f1
|
|
96 ;; cannot overlap with F1.
|
|
97 (exclusion_set "f1_1,f1_2" "F1")
|
|
98
|
|
99 (define_insn_reservation "nil" 0 (eq_attr "type" "nil") "nothing")
|
|
100
|
|
101 ;; Although reg moves have a latency of zero
|
|
102 ;; we need to highlight that they use D stage
|
|
103 ;; for one cycle.
|
|
104
|
|
105 ;; Group: MT
|
|
106 (define_insn_reservation "reg_mov" 0
|
|
107 (and (eq_attr "pipe_model" "sh4")
|
|
108 (eq_attr "type" "move"))
|
|
109 "issue")
|
|
110
|
|
111 ;; Group: LS
|
|
112 (define_insn_reservation "freg_mov" 0
|
|
113 (and (eq_attr "pipe_model" "sh4")
|
|
114 (eq_attr "type" "fmove"))
|
|
115 "issue+load_store")
|
|
116
|
|
117 ;; We don't model all pipeline stages; we model the issue ('D') stage
|
|
118 ;; inasmuch as we allow only two instructions to issue simultaneously,
|
|
119 ;; and CO instructions prevent any simultaneous issue of another instruction.
|
|
120 ;; (This uses pipe_01 and pipe_02).
|
|
121 ;; Double issue of EX insns is prevented by using the int unit in the EX stage.
|
|
122 ;; Double issue of EX / BR insns is prevented by using the int unit /
|
|
123 ;; pcr_addrcalc unit in the EX stage.
|
|
124 ;; Double issue of BR / LS instructions is prevented by using the
|
|
125 ;; pcr_addrcalc / load_store unit in the issue cycle.
|
|
126 ;; Double issue of FE instructions is prevented by using F0 in the first
|
|
127 ;; pipeline stage after the first D stage.
|
|
128 ;; There is no need to describe the [ES]X / [MN]A / S stages after a D stage
|
|
129 ;; (except in the cases outlined above), nor to describe the FS stage after
|
|
130 ;; the F2 stage.
|
|
131
|
|
132 ;; Other MT group instructions(1 step operations)
|
|
133 ;; Group: MT
|
|
134 ;; Latency: 1
|
|
135 ;; Issue Rate: 1
|
|
136 (define_insn_reservation "mt" 1
|
|
137 (and (eq_attr "pipe_model" "sh4")
|
|
138 (eq_attr "type" "mt_group"))
|
|
139 "issue")
|
|
140
|
|
141 ;; Fixed Point Arithmetic Instructions(1 step operations)
|
|
142 ;; Group: EX
|
|
143 ;; Latency: 1
|
|
144 ;; Issue Rate: 1
|
|
145 (define_insn_reservation "sh4_simple_arith" 1
|
|
146 (and (eq_attr "pipe_model" "sh4")
|
|
147 (eq_attr "insn_class" "ex_group"))
|
|
148 "issue,int")
|
|
149
|
|
150 ;; Load and store instructions have no alignment peculiarities for the SH4,
|
|
151 ;; but they use the load-store unit, which they share with the fmove type
|
|
152 ;; insns (fldi[01]; fmov frn,frm; flds; fsts; fabs; fneg) .
|
|
153 ;; Loads have a latency of two.
|
|
154 ;; However, call insns can only paired with a preceding insn, and have
|
|
155 ;; a delay slot, so that we want two more insns to be scheduled between the
|
|
156 ;; load of the function address and the call. This is equivalent to a
|
|
157 ;; latency of three.
|
|
158 ;; ADJUST_COST can only properly handle reductions of the cost, so we
|
|
159 ;; use a latency of three here, which gets multiplied by 10 to yield 30.
|
|
160 ;; We only do this for SImode loads of general registers, to make the work
|
|
161 ;; for ADJUST_COST easier.
|
|
162
|
|
163 ;; Load Store instructions. (MOV.[BWL]@(d,GBR)
|
|
164 ;; Group: LS
|
|
165 ;; Latency: 2
|
|
166 ;; Issue Rate: 1
|
|
167 (define_insn_reservation "sh4_load" 2
|
|
168 (and (eq_attr "pipe_model" "sh4")
|
|
169 (eq_attr "type" "load,pcload"))
|
|
170 "issue+load_store,nothing,memory")
|
|
171
|
|
172 ;; calls / sfuncs need an extra instruction for their delay slot.
|
|
173 ;; Moreover, estimating the latency for SImode loads as 3 will also allow
|
|
174 ;; adjust_cost to meaningfully bump it back up to 3 if they load the shift
|
|
175 ;; count of a dynamic shift.
|
|
176 (define_insn_reservation "sh4_load_si" 3
|
|
177 (and (eq_attr "pipe_model" "sh4")
|
|
178 (eq_attr "type" "load_si,pcload_si"))
|
|
179 "issue+load_store,nothing,memory")
|
|
180
|
|
181 ;; (define_bypass 2 "sh4_load_si" "!sh4_call")
|
|
182
|
|
183 ;; The load latency is upped to three higher if the dependent insn does
|
|
184 ;; double precision computation. We want the 'default' latency to reflect
|
|
185 ;; that increased latency because otherwise the insn priorities won't
|
|
186 ;; allow proper scheduling.
|
|
187 (define_insn_reservation "sh4_fload" 3
|
|
188 (and (eq_attr "pipe_model" "sh4")
|
|
189 (eq_attr "type" "fload,pcfload"))
|
|
190 "issue+load_store,nothing,memory")
|
|
191
|
|
192 ;; (define_bypass 2 "sh4_fload" "!")
|
|
193
|
|
194 (define_insn_reservation "sh4_store" 1
|
|
195 (and (eq_attr "pipe_model" "sh4")
|
|
196 (eq_attr "type" "store,fstore"))
|
|
197 "issue+load_store,nothing,memory")
|
|
198
|
|
199 (define_insn_reservation "mac_mem" 1
|
|
200 (and (eq_attr "pipe_model" "sh4")
|
|
201 (eq_attr "type" "mac_mem"))
|
|
202 "d_lock,nothing,memory")
|
|
203
|
|
204 ;; Load Store instructions.
|
|
205 ;; Group: LS
|
|
206 ;; Latency: 1
|
|
207 ;; Issue Rate: 1
|
|
208 (define_insn_reservation "sh4_gp_fpul" 1
|
|
209 (and (eq_attr "pipe_model" "sh4")
|
|
210 (eq_attr "type" "gp_fpul"))
|
|
211 "issue+load_store")
|
|
212
|
|
213 ;; Load Store instructions.
|
|
214 ;; Group: LS
|
|
215 ;; Latency: 3
|
|
216 ;; Issue Rate: 1
|
|
217 (define_insn_reservation "sh4_fpul_gp" 3
|
|
218 (and (eq_attr "pipe_model" "sh4")
|
|
219 (eq_attr "type" "fpul_gp"))
|
|
220 "issue+load_store")
|
|
221
|
|
222 ;; Branch (BF,BF/S,BT,BT/S,BRA)
|
|
223 ;; Group: BR
|
|
224 ;; Latency when taken: 2 (or 1)
|
|
225 ;; Issue Rate: 1
|
|
226 ;; The latency is 1 when displacement is 0.
|
|
227 ;; We can't really do much with the latency, even if we could express it,
|
|
228 ;; but the pairing restrictions are useful to take into account.
|
|
229 ;; ??? If the branch is likely, we might want to fill the delay slot;
|
|
230 ;; if the branch is likely, but not very likely, should we pretend to use
|
|
231 ;; a resource that CO instructions use, to get a pairable delay slot insn?
|
|
232 (define_insn_reservation "sh4_branch" 1
|
|
233 (and (eq_attr "pipe_model" "sh4")
|
|
234 (eq_attr "type" "cbranch,jump"))
|
|
235 "issue+pcr_addrcalc")
|
|
236
|
|
237 ;; Branch Far (JMP,RTS,BRAF)
|
|
238 ;; Group: CO
|
|
239 ;; Latency: 3
|
|
240 ;; Issue Rate: 2
|
|
241 ;; ??? Scheduling happens before branch shortening, and hence jmp and braf
|
|
242 ;; can't be distinguished from bra for the "jump" pattern.
|
|
243 (define_insn_reservation "sh4_return" 3
|
|
244 (and (eq_attr "pipe_model" "sh4")
|
|
245 (eq_attr "type" "return,jump_ind"))
|
|
246 "d_lock*2")
|
|
247
|
|
248 ;; RTE
|
|
249 ;; Group: CO
|
|
250 ;; Latency: 5
|
|
251 ;; Issue Rate: 5
|
|
252 ;; this instruction can be executed in any of the pipelines
|
|
253 ;; and blocks the pipeline for next 4 stages.
|
|
254 (define_insn_reservation "sh4_return_from_exp" 5
|
|
255 (and (eq_attr "pipe_model" "sh4")
|
|
256 (eq_attr "type" "rte"))
|
|
257 "d_lock*5")
|
|
258
|
|
259 ;; OCBP, OCBWB
|
|
260 ;; Group: CO
|
|
261 ;; Latency: 1-5
|
|
262 ;; Issue Rate: 1
|
111
|
263 ;; cwb is used for the sequence
|
|
264 ;; ocbwb @%0
|
|
265 ;; extu.w %0,%2
|
|
266 ;; or %1,%2
|
|
267 ;; mov.l %0,@%2
|
0
|
268 ;; ocbwb on its own would be "d_lock,nothing,memory*5"
|
|
269 (define_insn_reservation "ocbwb" 6
|
|
270 (and (eq_attr "pipe_model" "sh4")
|
|
271 (eq_attr "type" "cwb"))
|
|
272 "d_lock*2,(d_lock+memory)*3,issue+load_store+memory,memory*2")
|
|
273
|
|
274 ;; LDS to PR,JSR
|
|
275 ;; Group: CO
|
|
276 ;; Latency: 3
|
|
277 ;; Issue Rate: 2
|
|
278 ;; The SX stage is blocked for last 2 cycles.
|
|
279 ;; OTOH, the only time that has an effect for insns generated by the compiler
|
|
280 ;; is when lds to PR is followed by sts from PR - and that is highly unlikely -
|
|
281 ;; or when we are doing a function call - and we don't do inter-function
|
|
282 ;; scheduling. For the function call case, it's really best that we end with
|
|
283 ;; something that models an rts.
|
|
284 (define_insn_reservation "sh4_lds_to_pr" 3
|
|
285 (and (eq_attr "pipe_model" "sh4")
|
|
286 (eq_attr "type" "prset") )
|
|
287 "d_lock*2")
|
|
288
|
|
289 ;; calls introduce a longisch delay that is likely to flush the pipelines
|
|
290 ;; of the caller's instructions. Ordinary functions tend to end with a
|
|
291 ;; load to restore a register (in the delay slot of rts), while sfuncs
|
|
292 ;; tend to end with an EX or MT insn. But that is not actually relevant,
|
|
293 ;; since there are no instructions that contend for memory access early.
|
|
294 ;; We could, of course, provide exact scheduling information for specific
|
|
295 ;; sfuncs, if that should prove useful.
|
|
296 (define_insn_reservation "sh4_call" 16
|
|
297 (and (eq_attr "pipe_model" "sh4")
|
|
298 (eq_attr "type" "call,sfunc"))
|
|
299 "d_lock*16")
|
|
300
|
|
301 ;; LDS.L to PR
|
|
302 ;; Group: CO
|
|
303 ;; Latency: 3
|
|
304 ;; Issue Rate: 2
|
|
305 ;; The SX unit is blocked for last 2 cycles.
|
|
306 (define_insn_reservation "ldsmem_to_pr" 3
|
|
307 (and (eq_attr "pipe_model" "sh4")
|
|
308 (eq_attr "type" "pload"))
|
|
309 "d_lock*2")
|
|
310
|
|
311 ;; STS from PR
|
|
312 ;; Group: CO
|
|
313 ;; Latency: 2
|
|
314 ;; Issue Rate: 2
|
|
315 ;; The SX unit in second and third cycles.
|
|
316 (define_insn_reservation "sts_from_pr" 2
|
|
317 (and (eq_attr "pipe_model" "sh4")
|
|
318 (eq_attr "type" "prget"))
|
|
319 "d_lock*2")
|
|
320
|
|
321 ;; STS.L from PR
|
|
322 ;; Group: CO
|
|
323 ;; Latency: 2
|
|
324 ;; Issue Rate: 2
|
|
325 (define_insn_reservation "sh4_prstore_mem" 2
|
|
326 (and (eq_attr "pipe_model" "sh4")
|
|
327 (eq_attr "type" "pstore"))
|
|
328 "d_lock*2,nothing,memory")
|
|
329
|
|
330 ;; LDS to FPSCR
|
|
331 ;; Group: CO
|
|
332 ;; Latency: 4
|
|
333 ;; Issue Rate: 1
|
|
334 ;; F1 is blocked for last three cycles.
|
|
335 (define_insn_reservation "fpscr_load" 4
|
|
336 (and (eq_attr "pipe_model" "sh4")
|
|
337 (eq_attr "type" "gp_fpscr"))
|
|
338 "d_lock,nothing,F1*3")
|
|
339
|
|
340 ;; LDS.L to FPSCR
|
|
341 ;; Group: CO
|
|
342 ;; Latency: 1 / 4
|
|
343 ;; Latency to update Rn is 1 and latency to update FPSCR is 4
|
|
344 ;; Issue Rate: 1
|
|
345 ;; F1 is blocked for last three cycles.
|
|
346 (define_insn_reservation "fpscr_load_mem" 4
|
|
347 (and (eq_attr "pipe_model" "sh4")
|
|
348 (eq_attr "type" "mem_fpscr"))
|
|
349 "d_lock,nothing,(F1+memory),F1*2")
|
|
350
|
|
351
|
|
352 ;; Fixed point multiplication (DMULS.L DMULU.L MUL.L MULS.W,MULU.W)
|
|
353 ;; Group: CO
|
|
354 ;; Latency: 4 / 4
|
|
355 ;; Issue Rate: 2
|
|
356 (define_insn_reservation "multi" 4
|
|
357 (and (eq_attr "pipe_model" "sh4")
|
|
358 (eq_attr "type" "smpy,dmpy"))
|
|
359 "d_lock,(d_lock+f1_1),(f1_1|f1_2)*3,F2")
|
|
360
|
|
361 ;; Fixed STS from, and LDS to MACL / MACH
|
|
362 ;; Group: CO
|
|
363 ;; Latency: 3
|
|
364 ;; Issue Rate: 1
|
|
365 (define_insn_reservation "sh4_mac_gp" 3
|
|
366 (and (eq_attr "pipe_model" "sh4")
|
|
367 (eq_attr "type" "mac_gp,gp_mac,mem_mac"))
|
|
368 "d_lock")
|
|
369
|
|
370
|
|
371 ;; Single precision floating point computation FCMP/EQ,
|
|
372 ;; FCMP/GT, FADD, FLOAT, FMAC, FMUL, FSUB, FTRC, FRCHG, FSCHG
|
|
373 ;; Group: FE
|
|
374 ;; Latency: 3/4
|
|
375 ;; Issue Rate: 1
|
|
376 (define_insn_reservation "fp_arith" 3
|
|
377 (and (eq_attr "pipe_model" "sh4")
|
|
378 (eq_attr "type" "fp,fp_cmp"))
|
|
379 "issue,F01,F2")
|
|
380
|
|
381 ;; We don't model the resource usage of this exactly because that would
|
|
382 ;; introduce a bogus latency.
|
|
383 (define_insn_reservation "sh4_fpscr_toggle" 1
|
|
384 (and (eq_attr "pipe_model" "sh4")
|
|
385 (eq_attr "type" "fpscr_toggle"))
|
|
386 "issue")
|
|
387
|
|
388 (define_insn_reservation "fp_arith_ftrc" 3
|
|
389 (and (eq_attr "pipe_model" "sh4")
|
|
390 (eq_attr "type" "ftrc_s"))
|
|
391 "issue,F01,F2")
|
|
392
|
|
393 (define_bypass 1 "fp_arith_ftrc" "sh4_fpul_gp")
|
|
394
|
|
395 ;; Single Precision FDIV/SQRT
|
|
396 ;; Group: FE
|
|
397 ;; Latency: 12/13 (FDIV); 11/12 (FSQRT)
|
|
398 ;; Issue Rate: 1
|
|
399 ;; We describe fdiv here; fsqrt is actually one cycle faster.
|
|
400 (define_insn_reservation "fp_div" 12
|
|
401 (and (eq_attr "pipe_model" "sh4")
|
|
402 (eq_attr "type" "fdiv"))
|
|
403 "issue,F01+F3,F2+F3,F3*7,F1+F3,F2")
|
|
404
|
|
405 ;; Double Precision floating point computation
|
|
406 ;; (FCNVDS, FCNVSD, FLOAT, FTRC)
|
|
407 ;; Group: FE
|
|
408 ;; Latency: (3,4)/5
|
|
409 ;; Issue Rate: 1
|
|
410 (define_insn_reservation "dp_float" 4
|
|
411 (and (eq_attr "pipe_model" "sh4")
|
|
412 (eq_attr "type" "dfp_conv"))
|
|
413 "issue,F01,F1+F2,F2")
|
|
414
|
|
415 ;; Double-precision floating-point (FADD,FMUL,FSUB)
|
|
416 ;; Group: FE
|
|
417 ;; Latency: (7,8)/9
|
|
418 ;; Issue Rate: 1
|
|
419 (define_insn_reservation "fp_double_arith" 8
|
|
420 (and (eq_attr "pipe_model" "sh4")
|
|
421 (eq_attr "type" "dfp_arith,dfp_mul"))
|
|
422 "issue,F01,F1+F2,fpu*4,F2")
|
|
423
|
|
424 ;; Double-precision FCMP (FCMP/EQ,FCMP/GT)
|
|
425 ;; Group: CO
|
|
426 ;; Latency: 3/5
|
|
427 ;; Issue Rate: 2
|
|
428 (define_insn_reservation "fp_double_cmp" 3
|
|
429 (and (eq_attr "pipe_model" "sh4")
|
|
430 (eq_attr "type" "dfp_cmp"))
|
|
431 "d_lock,(d_lock+F01),F1+F2,F2")
|
|
432
|
|
433 ;; Double precision FDIV/SQRT
|
|
434 ;; Group: FE
|
|
435 ;; Latency: (24,25)/26
|
|
436 ;; Issue Rate: 1
|
|
437 (define_insn_reservation "dp_div" 25
|
|
438 (and (eq_attr "pipe_model" "sh4")
|
|
439 (eq_attr "type" "dfdiv"))
|
|
440 "issue,F01+F3,F1+F2+F3,F2+F3,F3*16,F1+F3,(fpu+F3)*2,F2")
|
|
441
|
|
442
|
|
443 ;; Use the branch-not-taken case to model arith3 insns. For the branch taken
|
|
444 ;; case, we'd get a d_lock instead of issue at the end.
|
|
445 (define_insn_reservation "arith3" 3
|
|
446 (and (eq_attr "pipe_model" "sh4")
|
|
447 (eq_attr "type" "arith3"))
|
|
448 "issue,d_lock+pcr_addrcalc,issue")
|
|
449
|
|
450 ;; arith3b insns schedule the same no matter if the branch is taken or not.
|
|
451 (define_insn_reservation "arith3b" 2
|
|
452 (and (eq_attr "pipe_model" "sh4")
|
|
453 (eq_attr "type" "arith3"))
|
|
454 "issue,d_lock+pcr_addrcalc")
|