1;; Cavium ThunderX pipeline description
2;; Copyright (C) 2014-2022 Free Software Foundation, Inc.
3;;
4;; Written by Andrew Pinski  <apinski@cavium.com>
5
6;; This file is part of GCC.
7
8;; GCC is free software; you can redistribute it and/or modify
9;; it under the terms of the GNU General Public License as published by
10;; the Free Software Foundation; either version 3, or (at your option)
11;; any later version.
12
13;; GCC is distributed in the hope that it will be useful,
14;; but WITHOUT ANY WARRANTY; without even the implied warranty of
15;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16;; GNU General Public License for more details.
17
18;; You should have received a copy of the GNU General Public License
19;; along with GCC; see the file COPYING3.  If not see
20;; <http://www.gnu.org/licenses/>.
21
22
23;; Thunder is a dual-issue processor that can issue all instructions on
24;; pipe0 and a subset on pipe1.
25
26
27(define_automaton "thunderx_main, thunderx_mult, thunderx_divide, thunderx_simd")
28
29(define_cpu_unit "thunderx_pipe0" "thunderx_main")
30(define_cpu_unit "thunderx_pipe1" "thunderx_main")
31(define_cpu_unit "thunderx_mult" "thunderx_mult")
32(define_cpu_unit "thunderx_divide" "thunderx_divide")
33(define_cpu_unit "thunderx_simd" "thunderx_simd")
34
35(define_insn_reservation "thunderx_add" 1
36  (and (eq_attr "tune" "thunderx")
37       (eq_attr "type" "adc_imm,adc_reg,adr,alu_imm,alu_sreg,alus_imm,alus_sreg,extend,logic_imm,logic_reg,logics_imm,logics_reg,mov_imm,mov_reg"))
38  "thunderx_pipe0 | thunderx_pipe1")
39
40(define_insn_reservation "thunderx_shift" 1
41  (and (eq_attr "tune" "thunderx")
42       (eq_attr "type" "bfm,bfx,extend,rotate_imm,shift_imm,shift_reg,rbit,rev"))
43  "thunderx_pipe0 | thunderx_pipe1")
44
45
46;; Arthimentic instructions with an extra shift or extend is two cycles.
47;; FIXME: This needs more attributes on aarch64 than what is currently there;
48;;    this is conserative for now.
49;; Except this is not correct as this is only for !(LSL && shift by 0/1/2/3)
50;; Except this is not correct as this is only for !(zero extend)
51
52(define_insn_reservation "thunderx_arith_shift" 2
53  (and (eq_attr "tune" "thunderx")
54       (eq_attr "type" "alu_ext,alu_shift_imm_lsl_1to4,alu_shift_imm_other,alu_shift_reg,alus_ext,logic_shift_imm,logic_shift_reg,logics_shift_imm,logics_shift_reg,alus_shift_imm"))
55  "thunderx_pipe0 | thunderx_pipe1")
56
57(define_insn_reservation "thunderx_csel" 2
58  (and (eq_attr "tune" "thunderx")
59       (eq_attr "type" "csel"))
60  "thunderx_pipe0 | thunderx_pipe1")
61
62;; Multiply and mulitply accumulate and count leading zeros can only happen on pipe 1
63
64(define_insn_reservation "thunderx_mul" 4
65  (and (eq_attr "tune" "thunderx")
66       (eq_attr "type" "mul,muls,mla,mlas,clz,smull,umull,smlal,umlal"))
67  "thunderx_pipe1 + thunderx_mult")
68
69;; crcb,crch,crcw is 4 cycles and can only happen on pipe 1
70
71(define_insn_reservation "thunderx_crc32" 4
72  (and (eq_attr "tune" "thunderx")
73       (eq_attr "type" "crc"))
74  "thunderx_pipe1 + thunderx_mult")
75
76;; crcx is 5 cycles and only happen on pipe 1
77;(define_insn_reservation "thunderx_crc64" 5
78;  (and (eq_attr "tune" "thunderx")
79;       (eq_attr "type" "crc")
80;       (eq_attr "mode" "DI"))
81;  "thunderx_pipe1 + thunderx_mult")
82
83(define_insn_reservation "thunderx_div32" 22
84  (and (eq_attr "tune" "thunderx")
85       (eq_attr "type" "udiv,sdiv"))
86  "thunderx_pipe1 + thunderx_divide, thunderx_divide * 21")
87
88;(define_insn_reservation "thunderx_div64" 38
89;  (and (eq_attr "tune" "thunderx")
90;       (eq_attr "type" "udiv,sdiv")
91;       (eq_attr "mode" "DI"))
92;  "thunderx_pipe1 + thunderx_divide, thunderx_divide * 34")
93
94;; Stores take one cycle in pipe 0
95(define_insn_reservation "thunderx_store" 1
96  (and (eq_attr "tune" "thunderx")
97       (eq_attr "type" "store_4"))
98  "thunderx_pipe0")
99
100;; Store pair are single issued
101(define_insn_reservation "thunderx_storepair" 1
102  (and (eq_attr "tune" "thunderx")
103       (eq_attr "type" "store_8,store_16"))
104  "thunderx_pipe0 + thunderx_pipe1")
105
106;; Prefetch are single issued
107;(define_insn_reservation "thunderx_prefetch" 1
108;  (and (eq_attr "tune" "thunderx")
109;       (eq_attr "type" "prefetch"))
110;  "thunderx_pipe0 + thunderx_pipe1")
111
112;; loads (and load pairs) from L1 take 3 cycles in pipe 0
113(define_insn_reservation "thunderx_load" 3
114  (and (eq_attr "tune" "thunderx")
115       (eq_attr "type" "load_4, load_8, load_16"))
116  "thunderx_pipe0")
117
118(define_insn_reservation "thunderx_brj" 1
119  (and (eq_attr "tune" "thunderx")
120       (eq_attr "type" "branch,trap,call"))
121  "thunderx_pipe1")
122
123;; FPU
124
125(define_insn_reservation "thunderx_fadd" 4
126  (and (eq_attr "tune" "thunderx")
127       (eq_attr "type" "faddd,fadds"))
128  "thunderx_pipe1")
129
130(define_insn_reservation "thunderx_fconst" 1
131  (and (eq_attr "tune" "thunderx")
132       (eq_attr "type" "fconsts,fconstd"))
133  "thunderx_pipe1")
134
135;; Moves between fp are 2 cycles including min/max
136(define_insn_reservation "thunderx_fmov" 2
137  (and (eq_attr "tune" "thunderx")
138       (eq_attr "type" "fmov,f_minmaxs,f_minmaxd"))
139  "thunderx_pipe1")
140
141;; ABS, and NEG are 1 cycle
142(define_insn_reservation "thunderx_fabs" 1
143  (and (eq_attr "tune" "thunderx")
144       (eq_attr "type" "ffariths,ffarithd"))
145  "thunderx_pipe1")
146
147(define_insn_reservation "thunderx_fcsel" 3
148  (and (eq_attr "tune" "thunderx")
149       (eq_attr "type" "fcsel"))
150  "thunderx_pipe1")
151
152(define_insn_reservation "thunderx_fmovgpr" 2
153  (and (eq_attr "tune" "thunderx")
154       (eq_attr "type" "f_mrc, f_mcr"))
155  "thunderx_pipe1")
156
157(define_insn_reservation "thunderx_fcmp" 3
158  (and (eq_attr "tune" "thunderx")
159       (eq_attr "type" "fcmps,fcmpd,fccmps,fccmpd"))
160  "thunderx_pipe1")
161
162(define_insn_reservation "thunderx_fmul" 6
163  (and (eq_attr "tune" "thunderx")
164       (eq_attr "type" "fmacs,fmacd,fmuls,fmuld"))
165  "thunderx_pipe1")
166
167(define_insn_reservation "thunderx_fdivs" 12
168  (and (eq_attr "tune" "thunderx")
169       (eq_attr "type" "fdivs"))
170  "thunderx_pipe1 + thunderx_divide, thunderx_divide*8")
171
172(define_insn_reservation "thunderx_fdivd" 22
173  (and (eq_attr "tune" "thunderx")
174       (eq_attr "type" "fdivd"))
175  "thunderx_pipe1 + thunderx_divide, thunderx_divide*18")
176
177(define_insn_reservation "thunderx_fsqrts" 17
178  (and (eq_attr "tune" "thunderx")
179       (eq_attr "type" "fsqrts"))
180  "thunderx_pipe1 + thunderx_divide, thunderx_divide*13")
181
182(define_insn_reservation "thunderx_fsqrtd" 31
183  (and (eq_attr "tune" "thunderx")
184       (eq_attr "type" "fsqrtd"))
185  "thunderx_pipe1 + thunderx_divide, thunderx_divide*27")
186
187;; The rounding conversion inside fp is 4 cycles
188(define_insn_reservation "thunderx_frint" 4
189  (and (eq_attr "tune" "thunderx")
190       (eq_attr "type" "f_cvt,f_rints,f_rintd"))
191  "thunderx_pipe1")
192
193;; Float to integer with a move from int to/from float is 6 cycles
194(define_insn_reservation "thunderx_f_cvt" 6
195  (and (eq_attr "tune" "thunderx")
196       (eq_attr "type" "f_cvtf2i,f_cvti2f"))
197  "thunderx_pipe1")
198
199;; FP/SIMD load/stores happen in pipe 0
200;; 64bit Loads register/pairs are 4 cycles from L1
201(define_insn_reservation "thunderx_64simd_fp_load" 4
202  (and (eq_attr "tune" "thunderx")
203       (eq_attr "type" "f_loadd,f_loads,neon_load1_1reg,\
204			neon_load1_1reg_q,neon_load1_2reg"))
205  "thunderx_pipe0")
206
207;; 128bit load pair is singled issue and 4 cycles from L1
208(define_insn_reservation "thunderx_128simd_pair_load" 4
209  (and (eq_attr "tune" "thunderx")
210       (eq_attr "type" "neon_load1_2reg_q"))
211  "thunderx_pipe0+thunderx_pipe1")
212
213;; FP/SIMD Stores takes one cycle in pipe 0
214;; ST1 with one registers either multiple structures or single structure is
215;;    also one cycle.
216(define_insn_reservation "thunderx_simd_fp_store" 1
217  (and (eq_attr "tune" "thunderx")
218       (eq_attr "type" "f_stored,f_stores,neon_store1_1reg,neon_store1_1reg_q, \
219			neon_store1_one_lane, neon_store1_one_lane_q"))
220  "thunderx_pipe0")
221
222;; 64bit neon store pairs are single issue for one cycle
223(define_insn_reservation "thunderx_64neon_storepair" 1
224  (and (eq_attr "tune" "thunderx")
225       (eq_attr "type" "neon_store1_2reg"))
226  "thunderx_pipe0 + thunderx_pipe1")
227
228;; 128bit neon store pair are single issued for two cycles
229(define_insn_reservation "thunderx_128neon_storepair" 2
230  (and (eq_attr "tune" "thunderx")
231       (eq_attr "type" "neon_store1_2reg_q"))
232  "(thunderx_pipe0 + thunderx_pipe1)*2")
233
234;; LD1R/LD1 (with a single struct) takes 6 cycles and issued in pipe0
235(define_insn_reservation "thunderx_neon_ld1" 6
236  (and (eq_attr "tune" "thunderx")
237       (eq_attr "type" "neon_load1_all_lanes"))
238  "thunderx_pipe0")
239
240;; SIMD/NEON (q forms take an extra cycle)
241;; SIMD For ThunderX is 64bit wide,
242
243;; ThunderX simd move instruction types - 2/3 cycles
244;; ThunderX dup, ins is the same
245;; ThunderX SIMD fabs/fneg instruction types
246(define_insn_reservation "thunderx_neon_move" 2
247  (and (eq_attr "tune" "thunderx")
248       (eq_attr "type" "neon_logic, neon_bsl, neon_fp_compare_s, \
249			neon_fp_compare_d, neon_move, neon_dup, \
250			neon_ins, neon_from_gp, neon_to_gp, \
251			neon_abs, neon_neg, \
252			neon_fp_neg_s, neon_fp_abs_s"))
253  "thunderx_pipe1 + thunderx_simd")
254
255(define_insn_reservation "thunderx_neon_move_q" 3
256  (and (eq_attr "tune" "thunderx")
257       (eq_attr "type" "neon_logic_q, neon_bsl_q, neon_fp_compare_s_q, \
258			neon_fp_compare_d_q, neon_move_q, neon_dup_q, \
259			neon_ins_q, neon_from_gp_q, neon_to_gp_q, \
260			neon_abs_q, neon_neg_q, \
261			neon_fp_neg_s_q, neon_fp_neg_d_q, \
262			neon_fp_abs_s_q, neon_fp_abs_d_q"))
263  "thunderx_pipe1 + thunderx_simd, thunderx_simd")
264
265;; ThunderX simd simple/add instruction types - 4/5 cycles
266
267(define_insn_reservation "thunderx_neon_add" 4
268  (and (eq_attr "tune" "thunderx")
269       (eq_attr "type" "neon_reduc_add, neon_reduc_minmax, neon_fp_reduc_add_s, \
270			neon_fp_reduc_add_d, neon_fp_to_int_s, neon_fp_to_int_d, \
271			neon_add_halve, neon_sub_halve, neon_qadd, neon_compare, \
272			neon_compare_zero, neon_minmax, neon_abd, neon_add, neon_sub, \
273			neon_fp_minmax_s, neon_fp_minmax_d, neon_reduc_add, neon_cls, \
274			neon_qabs, neon_qneg, neon_fp_addsub_s, neon_fp_addsub_d, \
275			neon_arith_acc, neon_rev, neon_fp_abd_s, neon_fp_abd_d, \
276			neon_fp_reduc_minmax_s"))
277  "thunderx_pipe1 + thunderx_simd")
278
279;; BIG NOTE: neon_add_long/neon_sub_long don't have a q form which is incorrect
280
281(define_insn_reservation "thunderx_neon_add_q" 5
282  (and (eq_attr "tune" "thunderx")
283       (eq_attr "type" "neon_reduc_add_q, neon_reduc_minmax_q, neon_fp_reduc_add_s_q, \
284			neon_fp_reduc_add_d_q, neon_fp_to_int_s_q, neon_fp_to_int_d_q, \
285			neon_add_halve_q, neon_sub_halve_q, neon_qadd_q, neon_compare_q, \
286			neon_compare_zero_q, neon_minmax_q, neon_abd_q, neon_add_q, neon_sub_q, \
287			neon_fp_minmax_s_q, neon_fp_minmax_d_q, neon_reduc_add_q, neon_cls_q, \
288			neon_qabs_q, neon_qneg_q, neon_fp_addsub_s_q, neon_fp_addsub_d_q, \
289			neon_add_long, neon_sub_long, neon_fp_abd_s_q, neon_fp_abd_d_q, \
290			neon_arith_acc_q, neon_rev_q, \
291			neon_fp_reduc_minmax_s_q, neon_fp_reduc_minmax_d_q"))
292  "thunderx_pipe1 + thunderx_simd, thunderx_simd")
293
294;; Multiplies (float and integer) and shifts and permutes (except for TBL) and float conversions
295;; are 6/7 cycles
296(define_insn_reservation "thunderx_neon_mult" 6
297  (and (eq_attr "tune" "thunderx")
298       (eq_attr "type" "neon_fp_mul_s, neon_fp_mul_d, neon_fp_mla_s, neon_fp_mla_d, \
299			neon_mla_b, neon_mla_h, neon_mla_s, \
300			neon_mla_h_scalar, neon_mla_s_scalar, \
301			neon_ext, neon_shift_imm, neon_permute, \
302			neon_int_to_fp_s, neon_int_to_fp_d, neon_shift_reg, \
303			neon_sat_shift_reg, neon_shift_acc, \
304			neon_mul_b, neon_mul_h, neon_mul_s, \
305			neon_mul_h_scalar, neon_mul_s_scalar, \
306			neon_fp_mul_s_scalar, \
307			neon_fp_mla_s_scalar"))
308  "thunderx_pipe1 + thunderx_simd")
309
310(define_insn_reservation "thunderx_neon_mult_q" 7
311  (and (eq_attr "tune" "thunderx")
312       (eq_attr "type" "neon_fp_mul_s_q, neon_fp_mul_d_q, neon_fp_mla_s_q, neon_fp_mla_d_q, \
313			neon_mla_b_q, neon_mla_h_q, neon_mla_s_q, \
314			neon_mla_h_scalar_q, neon_mla_s_scalar_q, \
315			neon_ext_q, neon_shift_imm_q, neon_permute_q, \
316			neon_int_to_fp_s_q, neon_int_to_fp_d_q, neon_shift_reg_q, \
317			neon_sat_shift_reg_q, neon_shift_acc_q, \
318			neon_shift_imm_long, \
319			neon_mul_b_q, neon_mul_h_q, neon_mul_s_q, \
320			neon_mul_h_scalar_q, neon_mul_s_scalar_q, \
321			neon_fp_mul_s_scalar_q, neon_fp_mul_d_scalar_q, \
322			neon_mul_b_long, neon_mul_h_long, neon_mul_s_long, \
323			neon_shift_imm_narrow_q, neon_fp_cvt_widen_s, neon_fp_cvt_narrow_d_q, \
324			neon_fp_mla_s_scalar_q, neon_fp_mla_d_scalar_q"))
325  "thunderx_pipe1 + thunderx_simd, thunderx_simd")
326
327
328;; AES[ED] is 5 cycles
329(define_insn_reservation "thunderx_crypto_aese" 5
330  (and (eq_attr "tune" "thunderx")
331       (eq_attr "type" "crypto_aese"))
332  "thunderx_pipe1 + thunderx_simd, thunderx_simd")
333
334;; AES{,I}MC is 3 cycles
335(define_insn_reservation "thunderx_crypto_aesmc" 3
336  (and (eq_attr "tune" "thunderx")
337       (eq_attr "type" "crypto_aesmc"))
338  "thunderx_pipe1 + thunderx_simd, thunderx_simd")
339
340
341;; Thunder 128bit SIMD reads the upper halve in cycle 2 and writes upper halve in the last cycle
342(define_bypass 2 "thunderx_neon_move_q" "thunderx_neon_move_q, thunderx_neon_add_q, thunderx_neon_mult_q")
343(define_bypass 4 "thunderx_neon_add_q" "thunderx_neon_move_q, thunderx_neon_add_q, thunderx_neon_mult_q")
344(define_bypass 6 "thunderx_neon_mult_q" "thunderx_neon_move_q, thunderx_neon_add_q, thunderx_neon_mult_q")
345
346;; 64bit TBL is emulated and takes 160 cycles
347(define_insn_reservation "thunderx_tbl" 160
348  (and (eq_attr "tune" "thunderx")
349       (eq_attr "type" "neon_tbl1"))
350  "(thunderx_pipe1+thunderx_pipe0)*160")
351
352;; 128bit TBL is emulated and takes 320 cycles
353(define_insn_reservation "thunderx_tblq" 320
354  (and (eq_attr "tune" "thunderx")
355       (eq_attr "type" "neon_tbl1_q"))
356  "(thunderx_pipe1+thunderx_pipe0)*320")
357
358;; Assume both pipes are needed for unknown and multiple-instruction
359;; patterns.
360
361(define_insn_reservation "thunderx_unknown" 1
362  (and (eq_attr "tune" "thunderx")
363       (eq_attr "type" "untyped,multiple"))
364  "thunderx_pipe0 + thunderx_pipe1")
365
366
367