1 |
709 |
jeremybenn |
;; ARM 1026EJ-S Pipeline Description
|
2 |
|
|
;; Copyright (C) 2003, 2007 Free Software Foundation, Inc.
|
3 |
|
|
;; Written by CodeSourcery, LLC.
|
4 |
|
|
;;
|
5 |
|
|
;; This file is part of GCC.
|
6 |
|
|
;;
|
7 |
|
|
;; GCC is free software; you can redistribute it and/or modify it
|
8 |
|
|
;; under the terms of the GNU General Public License as published by
|
9 |
|
|
;; the Free Software Foundation; either version 3, or (at your option)
|
10 |
|
|
;; any later version.
|
11 |
|
|
;;
|
12 |
|
|
;; GCC is distributed in the hope that it will be useful, but
|
13 |
|
|
;; WITHOUT ANY WARRANTY; without even the implied warranty of
|
14 |
|
|
;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
15 |
|
|
;; General Public License for more details.
|
16 |
|
|
;;
|
17 |
|
|
;; You should have received a copy of the GNU General Public License
|
18 |
|
|
;; along with GCC; see the file COPYING3. If not see
|
19 |
|
|
;; . */
|
20 |
|
|
|
21 |
|
|
;; These descriptions are based on the information contained in the
|
22 |
|
|
;; ARM1026EJ-S Technical Reference Manual, Copyright (c) 2003 ARM
|
23 |
|
|
;; Limited.
|
24 |
|
|
;;
|
25 |
|
|
|
26 |
|
|
;; This automaton provides a pipeline description for the ARM
|
27 |
|
|
;; 1026EJ-S core.
|
28 |
|
|
;;
|
29 |
|
|
;; The model given here assumes that the condition for all conditional
|
30 |
|
|
;; instructions is "true", i.e., that all of the instructions are
|
31 |
|
|
;; actually executed.
|
32 |
|
|
|
33 |
|
|
(define_automaton "arm1026ejs")
|
34 |
|
|
|
35 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
36 |
|
|
;; Pipelines
|
37 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
38 |
|
|
|
39 |
|
|
;; There are two pipelines:
|
40 |
|
|
;;
|
41 |
|
|
;; - An Arithmetic Logic Unit (ALU) pipeline.
|
42 |
|
|
;;
|
43 |
|
|
;; The ALU pipeline has fetch, issue, decode, execute, memory, and
|
44 |
|
|
;; write stages. We only need to model the execute, memory and write
|
45 |
|
|
;; stages.
|
46 |
|
|
;;
|
47 |
|
|
;; - A Load-Store Unit (LSU) pipeline.
|
48 |
|
|
;;
|
49 |
|
|
;; The LSU pipeline has decode, execute, memory, and write stages.
|
50 |
|
|
;; We only model the execute, memory and write stages.
|
51 |
|
|
|
52 |
|
|
(define_cpu_unit "a_e,a_m,a_w" "arm1026ejs")
|
53 |
|
|
(define_cpu_unit "l_e,l_m,l_w" "arm1026ejs")
|
54 |
|
|
|
55 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
56 |
|
|
;; ALU Instructions
|
57 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
58 |
|
|
|
59 |
|
|
;; ALU instructions require three cycles to execute, and use the ALU
|
60 |
|
|
;; pipeline in each of the three stages. The results are available
|
61 |
|
|
;; after the execute stage stage has finished.
|
62 |
|
|
;;
|
63 |
|
|
;; If the destination register is the PC, the pipelines are stalled
|
64 |
|
|
;; for several cycles. That case is not modeled here.
|
65 |
|
|
|
66 |
|
|
;; ALU operations with no shifted operand
|
67 |
|
|
(define_insn_reservation "alu_op" 1
|
68 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
69 |
|
|
(eq_attr "type" "alu"))
|
70 |
|
|
"a_e,a_m,a_w")
|
71 |
|
|
|
72 |
|
|
;; ALU operations with a shift-by-constant operand
|
73 |
|
|
(define_insn_reservation "alu_shift_op" 1
|
74 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
75 |
|
|
(eq_attr "type" "alu_shift"))
|
76 |
|
|
"a_e,a_m,a_w")
|
77 |
|
|
|
78 |
|
|
;; ALU operations with a shift-by-register operand
|
79 |
|
|
;; These really stall in the decoder, in order to read
|
80 |
|
|
;; the shift value in a second cycle. Pretend we take two cycles in
|
81 |
|
|
;; the execute stage.
|
82 |
|
|
(define_insn_reservation "alu_shift_reg_op" 2
|
83 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
84 |
|
|
(eq_attr "type" "alu_shift_reg"))
|
85 |
|
|
"a_e*2,a_m,a_w")
|
86 |
|
|
|
87 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
88 |
|
|
;; Multiplication Instructions
|
89 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
90 |
|
|
|
91 |
|
|
;; Multiplication instructions loop in the execute stage until the
|
92 |
|
|
;; instruction has been passed through the multiplier array enough
|
93 |
|
|
;; times.
|
94 |
|
|
|
95 |
|
|
;; The result of the "smul" and "smulw" instructions is not available
|
96 |
|
|
;; until after the memory stage.
|
97 |
|
|
(define_insn_reservation "mult1" 2
|
98 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
99 |
|
|
(eq_attr "insn" "smulxy,smulwy"))
|
100 |
|
|
"a_e,a_m,a_w")
|
101 |
|
|
|
102 |
|
|
;; The "smlaxy" and "smlawx" instructions require two iterations through
|
103 |
|
|
;; the execute stage; the result is available immediately following
|
104 |
|
|
;; the execute stage.
|
105 |
|
|
(define_insn_reservation "mult2" 2
|
106 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
107 |
|
|
(eq_attr "insn" "smlaxy,smlalxy,smlawx"))
|
108 |
|
|
"a_e*2,a_m,a_w")
|
109 |
|
|
|
110 |
|
|
;; The "smlalxy", "mul", and "mla" instructions require two iterations
|
111 |
|
|
;; through the execute stage; the result is not available until after
|
112 |
|
|
;; the memory stage.
|
113 |
|
|
(define_insn_reservation "mult3" 3
|
114 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
115 |
|
|
(eq_attr "insn" "smlalxy,mul,mla"))
|
116 |
|
|
"a_e*2,a_m,a_w")
|
117 |
|
|
|
118 |
|
|
;; The "muls" and "mlas" instructions loop in the execute stage for
|
119 |
|
|
;; four iterations in order to set the flags. The value result is
|
120 |
|
|
;; available after three iterations.
|
121 |
|
|
(define_insn_reservation "mult4" 3
|
122 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
123 |
|
|
(eq_attr "insn" "muls,mlas"))
|
124 |
|
|
"a_e*4,a_m,a_w")
|
125 |
|
|
|
126 |
|
|
;; Long multiply instructions that produce two registers of
|
127 |
|
|
;; output (such as umull) make their results available in two cycles;
|
128 |
|
|
;; the least significant word is available before the most significant
|
129 |
|
|
;; word. That fact is not modeled; instead, the instructions are
|
130 |
|
|
;; described as if the entire result was available at the end of the
|
131 |
|
|
;; cycle in which both words are available.
|
132 |
|
|
|
133 |
|
|
;; The "umull", "umlal", "smull", and "smlal" instructions all take
|
134 |
|
|
;; three iterations through the execute cycle, and make their results
|
135 |
|
|
;; available after the memory cycle.
|
136 |
|
|
(define_insn_reservation "mult5" 4
|
137 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
138 |
|
|
(eq_attr "insn" "umull,umlal,smull,smlal"))
|
139 |
|
|
"a_e*3,a_m,a_w")
|
140 |
|
|
|
141 |
|
|
;; The "umulls", "umlals", "smulls", and "smlals" instructions loop in
|
142 |
|
|
;; the execute stage for five iterations in order to set the flags.
|
143 |
|
|
;; The value result is available after four iterations.
|
144 |
|
|
(define_insn_reservation "mult6" 4
|
145 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
146 |
|
|
(eq_attr "insn" "umulls,umlals,smulls,smlals"))
|
147 |
|
|
"a_e*5,a_m,a_w")
|
148 |
|
|
|
149 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
150 |
|
|
;; Load/Store Instructions
|
151 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
152 |
|
|
|
153 |
|
|
;; The models for load/store instructions do not accurately describe
|
154 |
|
|
;; the difference between operations with a base register writeback
|
155 |
|
|
;; (such as "ldm!"). These models assume that all memory references
|
156 |
|
|
;; hit in dcache.
|
157 |
|
|
|
158 |
|
|
;; LSU instructions require six cycles to execute. They use the ALU
|
159 |
|
|
;; pipeline in all but the 5th cycle, and the LSU pipeline in cycles
|
160 |
|
|
;; three through six.
|
161 |
|
|
;; Loads and stores which use a scaled register offset or scaled
|
162 |
|
|
;; register pre-indexed addressing mode take three cycles EXCEPT for
|
163 |
|
|
;; those that are base + offset with LSL of 0 or 2, or base - offset
|
164 |
|
|
;; with LSL of zero. The remainder take 1 cycle to execute.
|
165 |
|
|
;; For 4byte loads there is a bypass from the load stage
|
166 |
|
|
|
167 |
|
|
(define_insn_reservation "load1_op" 2
|
168 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
169 |
|
|
(eq_attr "type" "load_byte,load1"))
|
170 |
|
|
"a_e+l_e,l_m,a_w+l_w")
|
171 |
|
|
|
172 |
|
|
(define_insn_reservation "store1_op" 0
|
173 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
174 |
|
|
(eq_attr "type" "store1"))
|
175 |
|
|
"a_e+l_e,l_m,a_w+l_w")
|
176 |
|
|
|
177 |
|
|
;; A load's result can be stored by an immediately following store
|
178 |
|
|
(define_bypass 1 "load1_op" "store1_op" "arm_no_early_store_addr_dep")
|
179 |
|
|
|
180 |
|
|
;; On a LDM/STM operation, the LSU pipeline iterates until all of the
|
181 |
|
|
;; registers have been processed.
|
182 |
|
|
;;
|
183 |
|
|
;; The time it takes to load the data depends on whether or not the
|
184 |
|
|
;; base address is 64-bit aligned; if it is not, an additional cycle
|
185 |
|
|
;; is required. This model assumes that the address is always 64-bit
|
186 |
|
|
;; aligned. Because the processor can load two registers per cycle,
|
187 |
|
|
;; that assumption means that we use the same instruction reservations
|
188 |
|
|
;; for loading 2k and 2k - 1 registers.
|
189 |
|
|
;;
|
190 |
|
|
;; The ALU pipeline is stalled until the completion of the last memory
|
191 |
|
|
;; stage in the LSU pipeline. That is modeled by keeping the ALU
|
192 |
|
|
;; execute stage busy until that point.
|
193 |
|
|
;;
|
194 |
|
|
;; As with ALU operations, if one of the destination registers is the
|
195 |
|
|
;; PC, there are additional stalls; that is not modeled.
|
196 |
|
|
|
197 |
|
|
(define_insn_reservation "load2_op" 2
|
198 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
199 |
|
|
(eq_attr "type" "load2"))
|
200 |
|
|
"a_e+l_e,l_m,a_w+l_w")
|
201 |
|
|
|
202 |
|
|
(define_insn_reservation "store2_op" 0
|
203 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
204 |
|
|
(eq_attr "type" "store2"))
|
205 |
|
|
"a_e+l_e,l_m,a_w+l_w")
|
206 |
|
|
|
207 |
|
|
(define_insn_reservation "load34_op" 3
|
208 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
209 |
|
|
(eq_attr "type" "load3,load4"))
|
210 |
|
|
"a_e+l_e,a_e+l_e+l_m,a_e+l_m,a_w+l_w")
|
211 |
|
|
|
212 |
|
|
(define_insn_reservation "store34_op" 0
|
213 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
214 |
|
|
(eq_attr "type" "store3,store4"))
|
215 |
|
|
"a_e+l_e,a_e+l_e+l_m,a_e+l_m,a_w+l_w")
|
216 |
|
|
|
217 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
218 |
|
|
;; Branch and Call Instructions
|
219 |
|
|
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
|
220 |
|
|
|
221 |
|
|
;; Branch instructions are difficult to model accurately. The ARM
|
222 |
|
|
;; core can predict most branches. If the branch is predicted
|
223 |
|
|
;; correctly, and predicted early enough, the branch can be completely
|
224 |
|
|
;; eliminated from the instruction stream. Some branches can
|
225 |
|
|
;; therefore appear to require zero cycles to execute. We assume that
|
226 |
|
|
;; all branches are predicted correctly, and that the latency is
|
227 |
|
|
;; therefore the minimum value.
|
228 |
|
|
|
229 |
|
|
(define_insn_reservation "branch_op" 0
|
230 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
231 |
|
|
(eq_attr "type" "branch"))
|
232 |
|
|
"nothing")
|
233 |
|
|
|
234 |
|
|
;; The latency for a call is not predictable. Therefore, we use 32 as
|
235 |
|
|
;; roughly equivalent to positive infinity.
|
236 |
|
|
|
237 |
|
|
(define_insn_reservation "call_op" 32
|
238 |
|
|
(and (eq_attr "tune" "arm1026ejs")
|
239 |
|
|
(eq_attr "type" "call"))
|
240 |
|
|
"nothing")
|