]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/config/i386/pentium.md
Update copyright years.
[thirdparty/gcc.git] / gcc / config / i386 / pentium.md
CommitLineData
af2728a4 1;; Pentium Scheduling
a945c346 2;; Copyright (C) 2002-2024 Free Software Foundation, Inc.
af2728a4 3;;
a805d35f 4;; This file is part of GCC.
af2728a4 5;;
a805d35f 6;; GCC is free software; you can redistribute it and/or modify
af2728a4 7;; it under the terms of the GNU General Public License as published by
2f83c7d6 8;; the Free Software Foundation; either version 3, or (at your option)
af2728a4
JL
9;; any later version.
10;;
a805d35f 11;; GCC is distributed in the hope that it will be useful,
af2728a4
JL
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
2f83c7d6
NC
17;; along with GCC; see the file COPYING3. If not see
18;; <http://www.gnu.org/licenses/>. */
af2728a4
JL
19;;
20;; The Pentium is an in-order core with two integer pipelines.
21
22;; True for insns that behave like prefixed insns on the Pentium.
23(define_attr "pent_prefix" "false,true"
24 (if_then_else (ior (eq_attr "prefix_0f" "1")
25 (ior (eq_attr "prefix_data16" "1")
26 (eq_attr "prefix_rep" "1")))
27 (const_string "true")
28 (const_string "false")))
29
30;; Categorize how an instruction slots.
31
32;; The non-MMX Pentium slots an instruction with prefixes on U pipe only,
33;; while MMX Pentium can slot it on either U or V. Model non-MMX Pentium
34;; rules, because it results in noticeably better code on non-MMX Pentium
35;; and doesn't hurt much on MMX. (Prefixed instructions are not very
d1f87653 36;; common, so the scheduler usually has a non-prefixed insn to pair).
af2728a4
JL
37
38(define_attr "pent_pair" "uv,pu,pv,np"
39 (cond [(eq_attr "imm_disp" "true")
40 (const_string "np")
41 (ior (eq_attr "type" "alu1,alu,imov,icmp,test,lea,incdec")
42 (and (eq_attr "type" "pop,push")
43 (eq_attr "memory" "!both")))
44 (if_then_else (eq_attr "pent_prefix" "true")
45 (const_string "pu")
46 (const_string "uv"))
47 (eq_attr "type" "ibr")
48 (const_string "pv")
49 (and (eq_attr "type" "ishift")
82e86dc6 50 (match_operand 2 "const_int_operand"))
af2728a4
JL
51 (const_string "pu")
52 (and (eq_attr "type" "rotate")
82e86dc6 53 (match_operand 2 "const1_operand"))
af2728a4 54 (const_string "pu")
1b245ade 55 (and (eq_attr "type" "ishift1")
82e86dc6 56 (match_operand 1 "const_int_operand"))
1b245ade
JH
57 (const_string "pu")
58 (and (eq_attr "type" "rotate1")
82e86dc6 59 (match_operand 1 "const1_operand"))
1b245ade 60 (const_string "pu")
af2728a4 61 (and (eq_attr "type" "call")
82e86dc6 62 (match_operand 0 "constant_call_address_operand"))
af2728a4
JL
63 (const_string "pv")
64 (and (eq_attr "type" "callv")
82e86dc6 65 (match_operand 1 "constant_call_address_operand"))
af2728a4
JL
66 (const_string "pv")
67 ]
68 (const_string "np")))
69
70(define_automaton "pentium,pentium_fpu")
71
72;; Pentium do have U and V pipes. Instruction to both pipes
d1f87653 73;; are always issued together, much like on VLIW.
af2728a4
JL
74;;
75;; predecode
76;; / \
77;; decodeu decodev
78;; / | |
79;; fpu executeu executev
80;; | | |
81;; fpu retire retire
82;; |
83;; fpu
84;; We add dummy "port" pipes allocated only first cycle of
9cd10576 85;; instruction to specify this behavior.
af2728a4
JL
86
87(define_cpu_unit "pentium-portu,pentium-portv" "pentium")
88(define_cpu_unit "pentium-u,pentium-v" "pentium")
89(absence_set "pentium-portu" "pentium-u,pentium-v")
90(presence_set "pentium-portv" "pentium-portu")
91
92;; Floating point instructions can overlap with new issue of integer
93;; instructions. We model only first cycle of FP pipeline, as it is
94;; fully pipelined.
95(define_cpu_unit "pentium-fp" "pentium_fpu")
96
97;; There is non-pipelined multiplier unit used for complex operations.
98(define_cpu_unit "pentium-fmul" "pentium_fpu")
99
100;; Pentium preserves memory ordering, so when load-execute-store
101;; instruction is executed together with other instruction loading
102;; data, the execution of the other instruction is delayed to very
103;; last cycle of first instruction, when data are bypassed.
104;; We model this by allocating "memory" unit when store is pending
105;; and using conflicting load units together.
106
107(define_cpu_unit "pentium-memory" "pentium")
108(define_cpu_unit "pentium-load0" "pentium")
109(define_cpu_unit "pentium-load1" "pentium")
110(absence_set "pentium-load0,pentium-load1" "pentium-memory")
111
112(define_reservation "pentium-load" "(pentium-load0 | pentium-load1)")
113(define_reservation "pentium-np" "(pentium-u + pentium-v)")
114(define_reservation "pentium-uv" "(pentium-u | pentium-v)")
115(define_reservation "pentium-portuv" "(pentium-portu | pentium-portv)")
116(define_reservation "pentium-firstu" "(pentium-u + pentium-portu)")
117(define_reservation "pentium-firstv" "(pentium-v + pentium-portuv)")
118(define_reservation "pentium-firstuv" "(pentium-uv + pentium-portuv)")
119(define_reservation "pentium-firstuload" "(pentium-load + pentium-firstu)")
120(define_reservation "pentium-firstvload" "(pentium-load + pentium-firstv)")
121(define_reservation "pentium-firstuvload" "(pentium-load + pentium-firstuv)
122 | (pentium-firstv,pentium-v,
123 (pentium-load+pentium-firstv))")
124(define_reservation "pentium-firstuboth" "(pentium-load + pentium-firstu
125 + pentium-memory)")
942579db 126(define_reservation "pentium-firstvboth" "(pentium-load + pentium-firstv
af2728a4
JL
127 + pentium-memory)")
128(define_reservation "pentium-firstuvboth" "(pentium-load + pentium-firstuv
129 + pentium-memory)
130 | (pentium-firstv,pentium-v,
131 (pentium-load+pentium-firstv))")
132
133;; Few common long latency instructions
134(define_insn_reservation "pent_mul" 11
135 (and (eq_attr "cpu" "pentium")
136 (eq_attr "type" "imul"))
137 "pentium-np*11")
138
139(define_insn_reservation "pent_str" 12
140 (and (eq_attr "cpu" "pentium")
141 (eq_attr "type" "str"))
142 "pentium-np*12")
143
144;; Integer division and some other long latency instruction block all
145;; units, including the FP pipe. There is no value in modeling the
146;; latency of these instructions and not modeling the latency
147;; decreases the size of the DFA.
148(define_insn_reservation "pent_block" 1
149 (and (eq_attr "cpu" "pentium")
150 (eq_attr "type" "idiv"))
151 "pentium-np+pentium-fp")
152
af2728a4
JL
153;; Moves usually have one cycle penalty, but there are exceptions.
154(define_insn_reservation "pent_fmov" 1
155 (and (eq_attr "cpu" "pentium")
156 (and (eq_attr "type" "fmov")
157 (eq_attr "memory" "none,load")))
158 "(pentium-fp+pentium-np)")
159
160(define_insn_reservation "pent_fpmovxf" 3
161 (and (eq_attr "cpu" "pentium")
162 (and (eq_attr "type" "fmov")
163 (and (eq_attr "memory" "load,store")
164 (eq_attr "mode" "XF"))))
165 "(pentium-fp+pentium-np)*3")
166
167(define_insn_reservation "pent_fpstore" 2
168 (and (eq_attr "cpu" "pentium")
169 (and (eq_attr "type" "fmov")
82e86dc6 170 (ior (match_operand 1 "immediate_operand")
af2728a4
JL
171 (eq_attr "memory" "store"))))
172 "(pentium-fp+pentium-np)*2")
173
174(define_insn_reservation "pent_imov" 1
175 (and (eq_attr "cpu" "pentium")
176 (eq_attr "type" "imov"))
177 "pentium-firstuv")
178
179;; Push and pop instructions have 1 cycle latency and special
180;; hardware bypass allows them to be paired with other push,pop
181;; and call instructions.
182(define_bypass 0 "pent_push,pent_pop" "pent_push,pent_pop,pent_call")
183(define_insn_reservation "pent_push" 1
184 (and (eq_attr "cpu" "pentium")
185 (and (eq_attr "type" "push")
186 (eq_attr "memory" "store")))
187 "pentium-firstuv")
188
189(define_insn_reservation "pent_pop" 1
190 (and (eq_attr "cpu" "pentium")
4977bab6 191 (eq_attr "type" "pop,leave"))
af2728a4
JL
192 "pentium-firstuv")
193
194;; Call and branch instruction can execute in either pipe, but
195;; they are only pairable when in the v pipe.
196(define_insn_reservation "pent_call" 10
197 (and (eq_attr "cpu" "pentium")
198 (eq_attr "type" "call,callv"))
199 "pentium-firstv,pentium-v*9")
200
201(define_insn_reservation "pent_branch" 1
202 (and (eq_attr "cpu" "pentium")
203 (eq_attr "type" "ibr"))
204 "pentium-firstv")
205
206;; Floating point instruction dispatch in U pipe, but continue
2a43945f 207;; in FP pipeline allowing other instructions to be executed.
af2728a4
JL
208(define_insn_reservation "pent_fp" 3
209 (and (eq_attr "cpu" "pentium")
210 (eq_attr "type" "fop,fistp"))
211 "(pentium-firstu+pentium-fp),nothing,nothing")
212
213;; First two cycles of fmul are not pipelined.
214(define_insn_reservation "pent_fmul" 3
215 (and (eq_attr "cpu" "pentium")
216 (eq_attr "type" "fmul"))
217 "(pentium-firstuv+pentium-fp+pentium-fmul),pentium-fmul,nothing")
218
219;; Long latency FP instructions overlap with integer instructions,
220;; but only last 2 cycles with FP ones.
221(define_insn_reservation "pent_fdiv" 39
222 (and (eq_attr "cpu" "pentium")
223 (eq_attr "type" "fdiv"))
224 "(pentium-np+pentium-fp+pentium-fmul),
225 (pentium-fp+pentium-fmul)*36,pentium-fmul*2")
226
227(define_insn_reservation "pent_fpspc" 70
228 (and (eq_attr "cpu" "pentium")
229 (eq_attr "type" "fpspc"))
230 "(pentium-np+pentium-fp+pentium-fmul),
231 (pentium-fp+pentium-fmul)*67,pentium-fmul*2")
232
233;; Integer instructions. Load/execute/store takes 3 cycles,
234;; load/execute 2 cycles and execute only one cycle.
235(define_insn_reservation "pent_uv_both" 3
236 (and (eq_attr "cpu" "pentium")
237 (and (eq_attr "pent_pair" "uv")
238 (eq_attr "memory" "both")))
239 "pentium-firstuvboth,pentium-uv+pentium-memory,pentium-uv")
240
241(define_insn_reservation "pent_u_both" 3
242 (and (eq_attr "cpu" "pentium")
243 (and (eq_attr "pent_pair" "pu")
244 (eq_attr "memory" "both")))
245 "pentium-firstuboth,pentium-u+pentium-memory,pentium-u")
246
247(define_insn_reservation "pent_v_both" 3
248 (and (eq_attr "cpu" "pentium")
249 (and (eq_attr "pent_pair" "pv")
250 (eq_attr "memory" "both")))
251 "pentium-firstvboth,pentium-v+pentium-memory,pentium-v")
252
253(define_insn_reservation "pent_np_both" 3
254 (and (eq_attr "cpu" "pentium")
255 (and (eq_attr "pent_pair" "np")
256 (eq_attr "memory" "both")))
257 "pentium-np,pentium-np,pentium-np")
258
259(define_insn_reservation "pent_uv_load" 2
260 (and (eq_attr "cpu" "pentium")
261 (and (eq_attr "pent_pair" "uv")
262 (eq_attr "memory" "load")))
263 "pentium-firstuvload,pentium-uv")
264
265(define_insn_reservation "pent_u_load" 2
266 (and (eq_attr "cpu" "pentium")
267 (and (eq_attr "pent_pair" "pu")
268 (eq_attr "memory" "load")))
269 "pentium-firstuload,pentium-u")
270
271(define_insn_reservation "pent_v_load" 2
272 (and (eq_attr "cpu" "pentium")
273 (and (eq_attr "pent_pair" "pv")
274 (eq_attr "memory" "load")))
275 "pentium-firstvload,pentium-v")
276
277(define_insn_reservation "pent_np_load" 2
278 (and (eq_attr "cpu" "pentium")
279 (and (eq_attr "pent_pair" "np")
280 (eq_attr "memory" "load")))
281 "pentium-np,pentium-np")
282
283(define_insn_reservation "pent_uv" 1
284 (and (eq_attr "cpu" "pentium")
285 (and (eq_attr "pent_pair" "uv")
286 (eq_attr "memory" "none")))
287 "pentium-firstuv")
288
289(define_insn_reservation "pent_u" 1
290 (and (eq_attr "cpu" "pentium")
291 (and (eq_attr "pent_pair" "pu")
292 (eq_attr "memory" "none")))
293 "pentium-firstu")
294
295(define_insn_reservation "pent_v" 1
296 (and (eq_attr "cpu" "pentium")
297 (and (eq_attr "pent_pair" "pv")
298 (eq_attr "memory" "none")))
299 "pentium-firstv")
300
301(define_insn_reservation "pent_np" 1
302 (and (eq_attr "cpu" "pentium")
303 (and (eq_attr "pent_pair" "np")
304 (eq_attr "memory" "none")))
305 "pentium-np")
306