1 ;; Copyright (C) 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
3 ;; This file is part of GCC.
5 ;; GCC is free software; you can redistribute it and/or modify
6 ;; it under the terms of the GNU General Public License as published by
7 ;; the Free Software Foundation; either version 3, or (at your option)
10 ;; GCC is distributed in the hope that it will be useful,
11 ;; but WITHOUT ANY WARRANTY; without even the implied warranty of
12 ;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 ;; GNU General Public License for more details.
15 ;; You should have received a copy of the GNU General Public License
16 ;; along with GCC; see the file COPYING3. If not see
17 ;; <http://www.gnu.org/licenses/>.
19 ;; DFA-based pipeline description for Broadcom SB-1
22 ;; The Broadcom SB-1 core is 4-way superscalar, in-order. It has 2 load/store
23 ;; pipes (one of which can support some ALU operations), 2 alu pipes, 2 FP
24 ;; pipes, and 1 MDMX pipes. It can issue 2 ls insns and 2 exe/fpu/mdmx insns
27 ;; We model the 4-way issue by ordering unit choices. The possible choices are
28 ;; {ex1,fp1}|{ex0,fp0}|ls1|ls0. Instructions issue to the first eligible unit
29 ;; in the list in most cases. Non-indexed load/stores issue to ls0 first.
30 ;; simple alu operations issue to ls1 if it is still available, and their
31 ;; operands are ready (no co-issue with loads), otherwise to the first
34 ;; When exceptions are enabled, can only issue FP insns to fp1. This is
35 ;; to ensure that instructions complete in order. The -mfp-exceptions option
36 ;; can be used to specify whether the system has FP exceptions enabled or not.
38 ;; In 32-bit mode, dependent FP can't co-issue with load, and only one FP exe
39 ;; insn can issue per cycle (fp1).
41 ;; The A1 MDMX pipe is separate from the FP pipes, but uses the same register
42 ;; file. As a result, once an MDMX insn is issued, no FP insns can be issued
43 ;; for 3 cycles. When an FP insn is issued, no MDMX insn can be issued for
44 ;; 5 cycles. This is currently not handled because there is no MDMX insn
48 ;; We use two automata. sb1_cpu_div is for the integer divides, which are
49 ;; not pipelined. sb1_cpu is for everything else.
51 (define_automaton "sb1_cpu, sb1_cpu_div")
53 ;; Load/store function units.
54 (define_cpu_unit "sb1_ls0" "sb1_cpu")
55 (define_cpu_unit "sb1_ls1" "sb1_cpu")
57 ;; CPU function units.
58 (define_cpu_unit "sb1_ex0" "sb1_cpu")
59 (define_cpu_unit "sb1_ex1" "sb1_cpu")
61 ;; The divide unit is not pipelined, and blocks hi/lo reads and writes.
62 (define_cpu_unit "sb1_div" "sb1_cpu_div")
63 ;; DMULT block any multiply from issuing in the next cycle.
64 (define_cpu_unit "sb1_mul" "sb1_cpu")
66 ;; Floating-point units.
67 (define_cpu_unit "sb1_fp0" "sb1_cpu")
68 (define_cpu_unit "sb1_fp1" "sb1_cpu")
70 ;; Can only issue to one of the ex and fp pipes at a time.
71 (exclusion_set "sb1_ex0" "sb1_fp0")
72 (exclusion_set "sb1_ex1" "sb1_fp1")
74 ;; Define an SB-1 specific attribute to simplify some FP descriptions.
75 ;; We can use 2 FP pipes only if we have 64-bit FP code, and exceptions are
78 (define_attr "sb1_fp_pipes" "one,two"
79 (cond [(and (match_test "TARGET_FLOAT64")
80 (not (match_test "TARGET_FP_EXCEPTIONS")))
82 (const_string "one")))
84 ;; Define reservations for common combinations.
86 ;; For long cycle operations, the FPU has a 4 cycle pipeline that repeats,
87 ;; effectively re-issuing the operation every 4 cycles. This means that we
88 ;; can have at most 4 long-cycle operations per pipe.
90 ;; ??? The fdiv operations should be e.g.
91 ;; sb1_fp1_4cycles*7" | "sb1_fp0_4cycle*7
92 ;; but the DFA is too large when we do that. Perhaps have to use scheduler
95 ;; ??? Try limiting scheduler to 2 long latency operations, and see if this
96 ;; results in a usable DFA, and whether it helps code performance.
98 ;;(define_reservation "sb1_fp0_4cycles" "sb1_fp0, nothing*3")
99 ;;(define_reservation "sb1_fp1_4cycles" "sb1_fp1, nothing*3")
102 ;; The ordering of the instruction-execution-path/resource-usage
103 ;; descriptions (also known as reservation RTL) is roughly ordered
104 ;; based on the define attribute RTL for the "type" classification.
105 ;; When modifying, remember that the first test that matches is the
109 (define_insn_reservation "ir_sb1_unknown" 1
110 (and (eq_attr "cpu" "sb1,sb1a")
111 (eq_attr "type" "unknown,multi"))
112 "sb1_ls0+sb1_ls1+sb1_ex0+sb1_ex1+sb1_fp0+sb1_fp1")
114 ;; predicted taken branch causes 2 cycle ifetch bubble. predicted not
115 ;; taken branch causes 0 cycle ifetch bubble. mispredicted branch causes 8
116 ;; cycle ifetch bubble. We assume all branches predicted not taken.
118 ;; ??? This assumption that branches are predicated not taken should be
119 ;; investigated. Maybe using 2 here will give better results.
121 (define_insn_reservation "ir_sb1_branch" 0
122 (and (eq_attr "cpu" "sb1,sb1a")
123 (eq_attr "type" "branch,jump,call"))
126 ;; ??? This is 1 cycle for ldl/ldr to ldl/ldr when they use the same data
127 ;; register as destination.
129 ;; ??? SB-1 can co-issue a load with a dependent arith insn if it executes on
130 ;; an EX unit. Can not co-issue if the dependent insn executes on an LS unit.
131 ;; SB-1A can always co-issue here.
133 ;; A load normally has a latency of zero cycles. In some cases, dependent
134 ;; insns can be issued in the same cycle. However, a value of 1 gives
135 ;; better performance in empirical testing.
137 (define_insn_reservation "ir_sb1_load" 1
138 (and (eq_attr "cpu" "sb1")
139 (eq_attr "type" "load,prefetch"))
142 (define_insn_reservation "ir_sb1a_load" 0
143 (and (eq_attr "cpu" "sb1a")
144 (eq_attr "type" "load,prefetch"))
147 ;; Can not co-issue fpload with fp exe when in 32-bit mode.
149 (define_insn_reservation "ir_sb1_fpload" 0
150 (and (eq_attr "cpu" "sb1,sb1a")
151 (and (eq_attr "type" "fpload")
152 (match_test "TARGET_FLOAT64")))
155 (define_insn_reservation "ir_sb1_fpload_32bitfp" 1
156 (and (eq_attr "cpu" "sb1,sb1a")
157 (and (eq_attr "type" "fpload")
158 (not (match_test "TARGET_FLOAT64"))))
161 ;; Indexed loads can only execute on LS1 pipe.
163 (define_insn_reservation "ir_sb1_fpidxload" 0
164 (and (eq_attr "cpu" "sb1,sb1a")
165 (and (eq_attr "type" "fpidxload")
166 (match_test "TARGET_FLOAT64")))
169 (define_insn_reservation "ir_sb1_fpidxload_32bitfp" 1
170 (and (eq_attr "cpu" "sb1,sb1a")
171 (and (eq_attr "type" "fpidxload")
172 (not (match_test "TARGET_FLOAT64"))))
175 ;; prefx can only execute on the ls1 pipe.
177 (define_insn_reservation "ir_sb1_prefetchx" 0
178 (and (eq_attr "cpu" "sb1,sb1a")
179 (eq_attr "type" "prefetchx"))
182 ;; ??? There is a 4.5 cycle latency if a store is followed by a load, and
183 ;; there is a RAW dependency.
185 (define_insn_reservation "ir_sb1_store" 1
186 (and (eq_attr "cpu" "sb1,sb1a")
187 (eq_attr "type" "store"))
188 "sb1_ls0+sb1_ex1 | sb1_ls0+sb1_ex0 | sb1_ls1+sb1_ex1 | sb1_ls1+sb1_ex0")
190 (define_insn_reservation "ir_sb1_fpstore" 1
191 (and (eq_attr "cpu" "sb1,sb1a")
192 (eq_attr "type" "fpstore"))
193 "sb1_ls0+sb1_fp1 | sb1_ls0+sb1_fp0 | sb1_ls1+sb1_fp1 | sb1_ls1+sb1_fp0")
195 ;; Indexed stores can only execute on LS1 pipe.
197 (define_insn_reservation "ir_sb1_fpidxstore" 1
198 (and (eq_attr "cpu" "sb1,sb1a")
199 (eq_attr "type" "fpidxstore"))
200 "sb1_ls1+sb1_fp1 | sb1_ls1+sb1_fp0")
202 ;; Load latencies are 3 cycles for one load to another load or store (address
203 ;; only). This is 0 cycles for one load to a store using it as the data
206 ;; This assumes that if a load is dependent on a previous insn, then it must
207 ;; be an address dependence.
210 "ir_sb1_load,ir_sb1a_load,ir_sb1_fpload,ir_sb1_fpload_32bitfp,
211 ir_sb1_fpidxload,ir_sb1_fpidxload_32bitfp"
212 "ir_sb1_load,ir_sb1a_load,ir_sb1_fpload,ir_sb1_fpload_32bitfp,
213 ir_sb1_fpidxload,ir_sb1_fpidxload_32bitfp,ir_sb1_prefetchx")
216 "ir_sb1_load,ir_sb1a_load,ir_sb1_fpload,ir_sb1_fpload_32bitfp,
217 ir_sb1_fpidxload,ir_sb1_fpidxload_32bitfp"
218 "ir_sb1_store,ir_sb1_fpstore,ir_sb1_fpidxstore"
219 "mips_store_data_bypass_p")
221 ;; On SB-1, simple alu instructions can execute on the LS1 unit.
223 ;; ??? A simple alu insn issued on an LS unit has 0 cycle latency to an EX
224 ;; insn, to a store (for data), and to an xfer insn. It has 1 cycle latency to
225 ;; another LS insn (excluding store data). A simple alu insn issued on an EX
226 ;; unit has a latency of 5 cycles when the results goes to a LS unit (excluding
227 ;; store data), otherwise a latency of 1 cycle.
229 ;; ??? We cannot handle latencies properly for simple alu instructions
230 ;; within the DFA pipeline model. Latencies can be defined only from one
231 ;; insn reservation to another. We can't make them depend on which function
232 ;; unit was used. This isn't a DFA flaw. There is a conflict here, as we
233 ;; need to know the latency before we can determine which unit will be
234 ;; available, but we need to know which unit it is issued to before we can
235 ;; compute the latency. Perhaps this can be handled via scheduler hooks.
236 ;; This needs to be investigated.
238 ;; ??? Optimal scheduling taking the LS units into account seems to require
239 ;; a pre-scheduling pass. We need to determine which instructions feed results
240 ;; into store/load addresses, and thus benefit most from being issued to the
241 ;; LS unit. Also, we need to prune the list to ensure we don't overschedule
242 ;; insns to the LS unit, and that we don't conflict with insns that need LS1
243 ;; such as indexed loads. We then need to emit nops to ensure that simple
244 ;; alu instructions that are not supposed to be scheduled to LS1 don't
245 ;; accidentally end up there because LS1 is free when they are issued. This
246 ;; will be a lot of work, and it isn't clear how useful it will be.
248 ;; Empirical testing shows that 2 gives the best result.
250 (define_insn_reservation "ir_sb1_simple_alu" 2
251 (and (eq_attr "cpu" "sb1")
252 (eq_attr "type" "const,arith,logical,move,signext"))
253 "sb1_ls1 | sb1_ex1 | sb1_ex0")
255 ;; On SB-1A, simple alu instructions can not execute on the LS1 unit, and we
256 ;; have none of the above problems.
258 (define_insn_reservation "ir_sb1a_simple_alu" 1
259 (and (eq_attr "cpu" "sb1a")
260 (eq_attr "type" "const,arith,logical,move,signext"))
263 ;; ??? condmove also includes some FP instructions that execute on the FP
264 ;; units. This needs to be clarified.
266 (define_insn_reservation "ir_sb1_alu" 1
267 (and (eq_attr "cpu" "sb1,sb1a")
268 (eq_attr "type" "condmove,nop,shift"))
271 ;; These are type arith/darith that only execute on the EX0 unit.
273 (define_insn_reservation "ir_sb1_alu_0" 1
274 (and (eq_attr "cpu" "sb1,sb1a")
275 (eq_attr "type" "slt,clz,trap"))
278 ;; An alu insn issued on an EX unit has a latency of 5 cycles when the
279 ;; result goes to a LS unit (excluding store data).
281 ;; This assumes that if a load is dependent on a previous insn, then it must
282 ;; be an address dependence.
285 "ir_sb1a_simple_alu,ir_sb1_alu,ir_sb1_alu_0,ir_sb1_mfhi,ir_sb1_mflo"
286 "ir_sb1_load,ir_sb1a_load,ir_sb1_fpload,ir_sb1_fpload_32bitfp,
287 ir_sb1_fpidxload,ir_sb1_fpidxload_32bitfp,ir_sb1_prefetchx")
290 "ir_sb1a_simple_alu,ir_sb1_alu,ir_sb1_alu_0,ir_sb1_mfhi,ir_sb1_mflo"
291 "ir_sb1_store,ir_sb1_fpstore,ir_sb1_fpidxstore"
292 "mips_store_data_bypass_p")
294 ;; mf{hi,lo} is 1 cycle.
296 (define_insn_reservation "ir_sb1_mfhi" 1
297 (and (eq_attr "cpu" "sb1,sb1a")
298 (and (eq_attr "type" "mfhilo")
299 (not (match_operand 1 "lo_operand"))))
302 (define_insn_reservation "ir_sb1_mflo" 1
303 (and (eq_attr "cpu" "sb1,sb1a")
304 (and (eq_attr "type" "mfhilo")
305 (match_operand 1 "lo_operand")))
308 ;; mt{hi,lo} to mul/div is 4 cycles.
310 (define_insn_reservation "ir_sb1_mthilo" 4
311 (and (eq_attr "cpu" "sb1,sb1a")
312 (eq_attr "type" "mthilo"))
315 ;; mt{hi,lo} to mf{hi,lo} is 3 cycles.
317 (define_bypass 3 "ir_sb1_mthilo" "ir_sb1_mfhi,ir_sb1_mflo")
319 ;; multiply latency to an EX operation is 3 cycles.
321 ;; ??? Should check whether we need to make multiply conflict with moves
322 ;; to/from hilo registers.
324 (define_insn_reservation "ir_sb1_mulsi" 3
325 (and (eq_attr "cpu" "sb1,sb1a")
326 (and (eq_attr "type" "imul,imul3,imadd")
327 (eq_attr "mode" "SI")))
330 ;; muldi to mfhi is 4 cycles.
331 ;; Blocks any other multiply insn issue for 1 cycle.
333 (define_insn_reservation "ir_sb1_muldi" 4
334 (and (eq_attr "cpu" "sb1,sb1a")
335 (and (eq_attr "type" "imul,imul3")
336 (eq_attr "mode" "DI")))
337 "sb1_ex1+sb1_mul, sb1_mul")
339 ;; muldi to mflo is 3 cycles.
341 (define_bypass 3 "ir_sb1_muldi" "ir_sb1_mflo")
343 ;; mul latency is 7 cycles if the result is used by any LS insn.
345 ;; This assumes that if a load is dependent on a previous insn, then it must
346 ;; be an address dependence.
349 "ir_sb1_mulsi,ir_sb1_muldi"
350 "ir_sb1_load,ir_sb1a_load,ir_sb1_fpload,ir_sb1_fpload_32bitfp,
351 ir_sb1_fpidxload,ir_sb1_fpidxload_32bitfp,ir_sb1_prefetchx")
354 "ir_sb1_mulsi,ir_sb1_muldi"
355 "ir_sb1_store,ir_sb1_fpstore,ir_sb1_fpidxstore"
356 "mips_store_data_bypass_p")
358 ;; The divide unit is not pipelined. Divide busy is asserted in the 4th
359 ;; cycle, and then deasserted on the latency cycle. So only one divide at
360 ;; a time, but the first/last 4 cycles can overlap.
362 ;; ??? All divides block writes to hi/lo regs. hi/lo regs are written 4 cycles
363 ;; after the latency cycle for divides (e.g. 40/72). dmult writes lo in
364 ;; cycle 7, and hi in cycle 8. All other insns write hi/lo regs in cycle 7.
365 ;; Default for output dependencies is the difference in latencies, which is
366 ;; only 1 cycle off here, e.g. div to mtlo stalls for 32 cycles, but should
367 ;; stall for 33 cycles. This does not seem significant enough to worry about.
369 (define_insn_reservation "ir_sb1_divsi" 36
370 (and (eq_attr "cpu" "sb1,sb1a")
371 (and (eq_attr "type" "idiv")
372 (eq_attr "mode" "SI")))
373 "sb1_ex1, nothing*3, sb1_div*32")
375 (define_insn_reservation "ir_sb1_divdi" 68
376 (and (eq_attr "cpu" "sb1,sb1a")
377 (and (eq_attr "type" "idiv")
378 (eq_attr "mode" "DI")))
379 "sb1_ex1, nothing*3, sb1_div*64")
381 (define_insn_reservation "ir_sb1_fpu_2pipes" 4
382 (and (eq_attr "cpu" "sb1,sb1a")
383 (and (eq_attr "type" "fmove,fadd,fmul,fabs,fneg,fcvt,frdiv1,frsqrt1")
384 (eq_attr "sb1_fp_pipes" "two")))
387 (define_insn_reservation "ir_sb1_fpu_1pipe" 4
388 (and (eq_attr "cpu" "sb1,sb1a")
389 (and (eq_attr "type" "fmove,fadd,fmul,fabs,fneg,fcvt,frdiv1,frsqrt1")
390 (eq_attr "sb1_fp_pipes" "one")))
393 (define_insn_reservation "ir_sb1_fpu_step2_2pipes" 8
394 (and (eq_attr "cpu" "sb1,sb1a")
395 (and (eq_attr "type" "frdiv2,frsqrt2")
396 (eq_attr "sb1_fp_pipes" "two")))
399 (define_insn_reservation "ir_sb1_fpu_step2_1pipe" 8
400 (and (eq_attr "cpu" "sb1,sb1a")
401 (and (eq_attr "type" "frdiv2,frsqrt2")
402 (eq_attr "sb1_fp_pipes" "one")))
405 ;; ??? madd/msub 4-cycle latency to itself (same fr?), but 8 cycle latency
408 ;; ??? Blocks issue of another non-madd/msub after 4 cycles.
410 (define_insn_reservation "ir_sb1_fmadd_2pipes" 8
411 (and (eq_attr "cpu" "sb1,sb1a")
412 (and (eq_attr "type" "fmadd")
413 (eq_attr "sb1_fp_pipes" "two")))
416 (define_insn_reservation "ir_sb1_fmadd_1pipe" 8
417 (and (eq_attr "cpu" "sb1,sb1a")
418 (and (eq_attr "type" "fmadd")
419 (eq_attr "sb1_fp_pipes" "one")))
422 (define_insn_reservation "ir_sb1_fcmp" 4
423 (and (eq_attr "cpu" "sb1,sb1a")
424 (eq_attr "type" "fcmp"))
427 ;; mtc1 latency 5 cycles.
429 (define_insn_reservation "ir_sb1_mtxfer" 5
430 (and (eq_attr "cpu" "sb1,sb1a")
431 (eq_attr "type" "mtc"))
434 ;; mfc1 latency 1 cycle.
436 (define_insn_reservation "ir_sb1_mfxfer" 1
437 (and (eq_attr "cpu" "sb1,sb1a")
438 (eq_attr "type" "mfc"))
441 ;; ??? Can deliver at most 1 result per every 6 cycles because of issue
444 (define_insn_reservation "ir_sb1_divsf_2pipes" 24
445 (and (eq_attr "cpu" "sb1,sb1a")
446 (and (eq_attr "type" "fdiv")
447 (and (eq_attr "mode" "SF")
448 (eq_attr "sb1_fp_pipes" "two"))))
451 (define_insn_reservation "ir_sb1_divsf_1pipe" 24
452 (and (eq_attr "cpu" "sb1,sb1a")
453 (and (eq_attr "type" "fdiv")
454 (and (eq_attr "mode" "SF")
455 (eq_attr "sb1_fp_pipes" "one"))))
458 ;; ??? Can deliver at most 1 result per every 8 cycles because of issue
461 (define_insn_reservation "ir_sb1_divdf_2pipes" 32
462 (and (eq_attr "cpu" "sb1,sb1a")
463 (and (eq_attr "type" "fdiv")
464 (and (eq_attr "mode" "DF")
465 (eq_attr "sb1_fp_pipes" "two"))))
468 (define_insn_reservation "ir_sb1_divdf_1pipe" 32
469 (and (eq_attr "cpu" "sb1,sb1a")
470 (and (eq_attr "type" "fdiv")
471 (and (eq_attr "mode" "DF")
472 (eq_attr "sb1_fp_pipes" "one"))))
475 ;; ??? Can deliver at most 1 result per every 3 cycles because of issue
478 (define_insn_reservation "ir_sb1_recipsf_2pipes" 12
479 (and (eq_attr "cpu" "sb1,sb1a")
480 (and (eq_attr "type" "frdiv")
481 (and (eq_attr "mode" "SF")
482 (eq_attr "sb1_fp_pipes" "two"))))
485 (define_insn_reservation "ir_sb1_recipsf_1pipe" 12
486 (and (eq_attr "cpu" "sb1,sb1a")
487 (and (eq_attr "type" "frdiv")
488 (and (eq_attr "mode" "SF")
489 (eq_attr "sb1_fp_pipes" "one"))))
492 ;; ??? Can deliver at most 1 result per every 5 cycles because of issue
495 (define_insn_reservation "ir_sb1_recipdf_2pipes" 20
496 (and (eq_attr "cpu" "sb1,sb1a")
497 (and (eq_attr "type" "frdiv")
498 (and (eq_attr "mode" "DF")
499 (eq_attr "sb1_fp_pipes" "two"))))
502 (define_insn_reservation "ir_sb1_recipdf_1pipe" 20
503 (and (eq_attr "cpu" "sb1,sb1a")
504 (and (eq_attr "type" "frdiv")
505 (and (eq_attr "mode" "DF")
506 (eq_attr "sb1_fp_pipes" "one"))))
509 ;; ??? Can deliver at most 1 result per every 7 cycles because of issue
512 (define_insn_reservation "ir_sb1_sqrtsf_2pipes" 28
513 (and (eq_attr "cpu" "sb1,sb1a")
514 (and (eq_attr "type" "fsqrt")
515 (and (eq_attr "mode" "SF")
516 (eq_attr "sb1_fp_pipes" "two"))))
519 (define_insn_reservation "ir_sb1_sqrtsf_1pipe" 28
520 (and (eq_attr "cpu" "sb1,sb1a")
521 (and (eq_attr "type" "fsqrt")
522 (and (eq_attr "mode" "SF")
523 (eq_attr "sb1_fp_pipes" "one"))))
526 ;; ??? Can deliver at most 1 result per every 10 cycles because of issue
529 (define_insn_reservation "ir_sb1_sqrtdf_2pipes" 40
530 (and (eq_attr "cpu" "sb1,sb1a")
531 (and (eq_attr "type" "fsqrt")
532 (and (eq_attr "mode" "DF")
533 (eq_attr "sb1_fp_pipes" "two"))))
536 (define_insn_reservation "ir_sb1_sqrtdf_1pipe" 40
537 (and (eq_attr "cpu" "sb1,sb1a")
538 (and (eq_attr "type" "fsqrt")
539 (and (eq_attr "mode" "DF")
540 (eq_attr "sb1_fp_pipes" "one"))))
543 ;; ??? Can deliver at most 1 result per every 4 cycles because of issue
546 (define_insn_reservation "ir_sb1_rsqrtsf_2pipes" 16
547 (and (eq_attr "cpu" "sb1,sb1a")
548 (and (eq_attr "type" "frsqrt")
549 (and (eq_attr "mode" "SF")
550 (eq_attr "sb1_fp_pipes" "two"))))
553 (define_insn_reservation "ir_sb1_rsqrtsf_1pipe" 16
554 (and (eq_attr "cpu" "sb1,sb1a")
555 (and (eq_attr "type" "frsqrt")
556 (and (eq_attr "mode" "SF")
557 (eq_attr "sb1_fp_pipes" "one"))))
560 ;; ??? Can deliver at most 1 result per every 7 cycles because of issue
563 (define_insn_reservation "ir_sb1_rsqrtdf_2pipes" 28
564 (and (eq_attr "cpu" "sb1,sb1a")
565 (and (eq_attr "type" "frsqrt")
566 (and (eq_attr "mode" "DF")
567 (eq_attr "sb1_fp_pipes" "two"))))
570 (define_insn_reservation "ir_sb1_rsqrtdf_1pipe" 28
571 (and (eq_attr "cpu" "sb1,sb1a")
572 (and (eq_attr "type" "frsqrt")
573 (and (eq_attr "mode" "DF")
574 (eq_attr "sb1_fp_pipes" "one"))))