1 |
709 |
jeremybenn |
;; Pipeline description for the AppliedMicro Titan core.
|
2 |
|
|
;; Copyright (C) 2010 Free Software Foundation, Inc.
|
3 |
|
|
;; Contributed by Theobroma Systems Design und Consulting GmbH
|
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
|
9 |
|
|
;; by the Free Software Foundation; either version 3, or (at your
|
10 |
|
|
;; option) any later version.
|
11 |
|
|
;;
|
12 |
|
|
;; GCC is distributed in the hope that it will be useful, but WITHOUT
|
13 |
|
|
;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
14 |
|
|
;; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
|
15 |
|
|
;; 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 |
|
|
;; AppliedMicro Titan core complex
|
22 |
|
|
|
23 |
|
|
(define_automaton "titan_core,titan_fpu,titan_fxu,titan_bpu,titan_lsu")
|
24 |
|
|
(define_cpu_unit "titan_issue_0,titan_issue_1" "titan_core")
|
25 |
|
|
|
26 |
|
|
;; Some useful abbreviations.
|
27 |
|
|
(define_reservation "titan_issue" "titan_issue_0|titan_issue_1")
|
28 |
|
|
|
29 |
|
|
;; === FXU scheduling ===
|
30 |
|
|
|
31 |
|
|
(define_cpu_unit "titan_fxu_sh,titan_fxu_wb" "titan_fxu")
|
32 |
|
|
|
33 |
|
|
;; The 1-cycle adder executes add, addi, subf, neg, compare and trap
|
34 |
|
|
;; instructions. It provides its own, dedicated result-bus, so we
|
35 |
|
|
;; don't need the titan_fxu_wb reservation to complete.
|
36 |
|
|
(define_insn_reservation "titan_fxu_adder" 1
|
37 |
|
|
(and (eq_attr "type" "cmp,fast_compare,trap")
|
38 |
|
|
(eq_attr "cpu" "titan"))
|
39 |
|
|
"titan_issue,titan_fxu_sh")
|
40 |
|
|
|
41 |
|
|
;; Keep the titan_imul and titan_mulhw (half-word) rules in order, to
|
42 |
|
|
;; ensure the proper match: the half-word instructions are tagged as
|
43 |
|
|
;; imul3 only, whereas regular multiplys will always carry a imul tag.
|
44 |
|
|
|
45 |
|
|
(define_insn_reservation "titan_imul" 5
|
46 |
|
|
(and (eq_attr "type" "imul,imul2,imul_compare")
|
47 |
|
|
(eq_attr "cpu" "titan"))
|
48 |
|
|
"titan_issue,titan_fxu_sh,nothing*5,titan_fxu_wb")
|
49 |
|
|
|
50 |
|
|
(define_insn_reservation "titan_mulhw" 4
|
51 |
|
|
(and (eq_attr "type" "imul3")
|
52 |
|
|
(eq_attr "cpu" "titan"))
|
53 |
|
|
"titan_issue,titan_fxu_sh,nothing*4,titan_fxu_wb")
|
54 |
|
|
|
55 |
|
|
(define_bypass 2 "titan_mulhw" "titan_mulhw")
|
56 |
|
|
|
57 |
|
|
(define_insn_reservation "titan_fxu_shift_and_rotate" 2
|
58 |
|
|
(and (eq_attr "type" "insert_word,shift,var_shift_rotate,cntlz")
|
59 |
|
|
(eq_attr "cpu" "titan"))
|
60 |
|
|
"titan_issue,titan_fxu_sh,nothing*2,titan_fxu_wb")
|
61 |
|
|
|
62 |
|
|
;; We model the divider for the worst-case (i.e. a full 32-bit
|
63 |
|
|
;; divide). To model the bypass for byte-wise completion, a
|
64 |
|
|
;; define_bypass with a guard-function could be used... however, this
|
65 |
|
|
;; would be an optimization of doubtful value, as a large number of
|
66 |
|
|
;; divides will operate on 32-bit variables.
|
67 |
|
|
|
68 |
|
|
;; To avoid an unmanagably large automata (generating the automata
|
69 |
|
|
;; would require well over 2GB in memory), we don't model the shared
|
70 |
|
|
;; result bus on this one. The divider-pipeline is thus modeled
|
71 |
|
|
;; through its latency and initial disptach bottlenecks (i.e. issue
|
72 |
|
|
;; slots and fxu scheduler availability)
|
73 |
|
|
(define_insn_reservation "titan_fxu_div" 34
|
74 |
|
|
(and (eq_attr "type" "idiv")
|
75 |
|
|
(eq_attr "cpu" "titan"))
|
76 |
|
|
"titan_issue,titan_fxu_sh")
|
77 |
|
|
|
78 |
|
|
(define_insn_reservation "titan_fxu_alu" 1
|
79 |
|
|
(and (eq_attr "type" "integer,exts")
|
80 |
|
|
(eq_attr "cpu" "titan"))
|
81 |
|
|
"titan_issue,titan_fxu_sh,nothing,titan_fxu_wb")
|
82 |
|
|
|
83 |
|
|
;; === BPU scheduling ===
|
84 |
|
|
|
85 |
|
|
(define_cpu_unit "titan_bpu_sh" "titan_bpu")
|
86 |
|
|
|
87 |
|
|
(define_insn_reservation "titan_bpu" 2
|
88 |
|
|
(and (eq_attr "type" "branch,jmpreg,cr_logical,delayed_cr")
|
89 |
|
|
(eq_attr "cpu" "titan"))
|
90 |
|
|
"titan_issue,titan_bpu_sh")
|
91 |
|
|
|
92 |
|
|
;; === LSU scheduling ===
|
93 |
|
|
|
94 |
|
|
(define_cpu_unit "titan_lsu_sh" "titan_lsu")
|
95 |
|
|
|
96 |
|
|
;; Loads.
|
97 |
|
|
(define_insn_reservation "titan_lsu_load" 3
|
98 |
|
|
(and (eq_attr "type" "load,load_ext,load_ext_u,load_ext_ux,load_ux,load_u,\
|
99 |
|
|
load_l,sync")
|
100 |
|
|
(eq_attr "cpu" "titan"))
|
101 |
|
|
"titan_issue,titan_lsu_sh")
|
102 |
|
|
|
103 |
|
|
(define_insn_reservation "titan_lsu_fpload" 12
|
104 |
|
|
(and (eq_attr "type" "fpload,fpload_ux,fpload_u")
|
105 |
|
|
(eq_attr "cpu" "titan"))
|
106 |
|
|
"titan_issue,titan_lsu_sh")
|
107 |
|
|
|
108 |
|
|
;; Note that the isync is not clearly placed within any execution
|
109 |
|
|
;; unit. We've made the assumption that it will be running out of the
|
110 |
|
|
;; LSU, as msync is also executed within the LSU.
|
111 |
|
|
(define_insn_reservation "titan_lsu_sync" 20
|
112 |
|
|
(and (eq_attr "type" "sync")
|
113 |
|
|
(eq_attr "cpu" "titan"))
|
114 |
|
|
"titan_issue,titan_lsu_sh*20")
|
115 |
|
|
|
116 |
|
|
;; Stores.
|
117 |
|
|
(define_insn_reservation "titan_lsu_store" 12
|
118 |
|
|
(and (eq_attr "type" "store,store_ux,store_u,store_c")
|
119 |
|
|
(eq_attr "cpu" "titan"))
|
120 |
|
|
"titan_issue,titan_lsu_sh")
|
121 |
|
|
|
122 |
|
|
(define_insn_reservation "titan_lsu_fpstore" 12
|
123 |
|
|
(and (eq_attr "type" "fpstore,fpstore_ux,fpstore_u")
|
124 |
|
|
(eq_attr "cpu" "titan"))
|
125 |
|
|
"titan_issue,titan_lsu_sh")
|
126 |
|
|
|
127 |
|
|
;; === FPU scheduling ===
|
128 |
|
|
|
129 |
|
|
;; In order to keep the automaton for the Titan FPU efficient and
|
130 |
|
|
;; maintainable, we've kept in as concise as possible and created a
|
131 |
|
|
;; mapping for the main "choke points" only instead of modelling the
|
132 |
|
|
;; overall flow of instructions through the FP-pipeline(s).
|
133 |
|
|
|
134 |
|
|
;; The key elements modelled are:
|
135 |
|
|
;; * each FP-instruction takes up one of the two issue slots
|
136 |
|
|
;; * the FPU runs at half the core frequency
|
137 |
|
|
;; * divides are not pipelined (but execute in a separate unit)
|
138 |
|
|
;; * the FPU has a shared result bus for all its units
|
139 |
|
|
|
140 |
|
|
(define_cpu_unit "titan_fp0,titan_fpdiv,titan_fpwb" "titan_fpu")
|
141 |
|
|
|
142 |
|
|
(define_insn_reservation "titan_fp_div_double" 72
|
143 |
|
|
(and (eq_attr "type" "ddiv")
|
144 |
|
|
(eq_attr "cpu" "titan"))
|
145 |
|
|
"titan_issue,titan_fpdiv*72,titan_fpwb")
|
146 |
|
|
|
147 |
|
|
(define_insn_reservation "titan_fp_div_single" 46
|
148 |
|
|
(and (eq_attr "type" "sdiv")
|
149 |
|
|
(eq_attr "cpu" "titan"))
|
150 |
|
|
"titan_issue,titan_fpdiv*46,titan_fpwb")
|
151 |
|
|
|
152 |
|
|
(define_insn_reservation "titan_fp_single" 12
|
153 |
|
|
(and (eq_attr "fp_type" "fp_addsub_s,fp_mul_s,fp_maddsub_s")
|
154 |
|
|
(eq_attr "cpu" "titan"))
|
155 |
|
|
"titan_issue,titan_fp0*2,nothing*10,titan_fpwb")
|
156 |
|
|
|
157 |
|
|
;; Make sure the "titan_fp" rule stays last, as it's a catch all for
|
158 |
|
|
;; double-precision and unclassified (e.g. fsel) FP-instructions
|
159 |
|
|
(define_insn_reservation "titan_fp" 10
|
160 |
|
|
(and (eq_attr "type" "fpcompare,fp,dmul")
|
161 |
|
|
(eq_attr "cpu" "titan"))
|
162 |
|
|
"titan_issue,titan_fp0*2,nothing*8,titan_fpwb")
|
163 |
|
|
|
164 |
|
|
;; Please note, that the non-pipelined FP-instructions "mcrfs",
|
165 |
|
|
;; "mtfsb0[.]", "mtfsb1[.]", "mtfsf[.]", "mtfsfi[.]" are not
|
166 |
|
|
;; accessible from regular language constructs (i.e. they are not used
|
167 |
|
|
;; by the code generator, except for special purpose sequences defined
|
168 |
|
|
;; in rs6000.md), no special provisions are made for these.
|
169 |
|
|
|