1 ;; .........................
3 ;; DFA-based pipeline description for Sandcraft SR3 (MIPS64 based)
5 ;; The SR3 is describeds as:
6 ;; - nine-stage pipeline, insn buffering with out-of-order issue to
7 ;; multiple function units, with an average dispatch rate of 2
8 ;; insn.s per cycle (max 6 insns: 2 fpu, 4 cpu).
10 ;; The details on this are scant except for a diagram in
11 ;; Chap. 6 of Rev. 1.0 SR3 Spec.
13 ;; The model employed below is designed to closely approximate the
14 ;; published latencies. Emulation of out-of-order issue and the insn
15 ;; buffering is done via a VLIW dispatch style (with a packing of 6 insns);
16 ;; the function unit reservations restrictions (define_*_set) are
17 ;; contrived to support published timings.
20 ;; "SR3 Microporocessor Specification, System development information,"
21 ;; Revision 1.0, 13 December 2000.
24 ;; Reservation model is based on:
25 ;; 1) Figure 6-1, from the 1.0 specicification.
26 ;; 2) Chapter 19, from the 1.0 specificication.
27 ;; 3) following questions(Red Hat)/answers(Sandcraft):
28 ;; RH> From Section 19.1
29 ;; RH> 1) In terms of figure 6-1, are all the instructions in
30 ;; RH> table 19-1 restricted
31 ;; RH> to ALUx? When ALUx is not in use for an instruction in table;; RH> 19-1 is
32 ;; RH> it fully compatible with all insns that issue to ALUy?
34 ;; Yes, all the instructions in Table 19-1 only go to ALUX, and all the
35 ;; instructions that can be issued to ALUY can also be issued to ALUX.
38 ;; RH> From Section 19.2
39 ;; RH> 2) Explain conditional moves execution path (in terms of
42 ;; Conditional move of integer registers (based on floating point condition
43 ;; codes or integer register value) go to ALUX or ALUY.
45 ;; RH> 3) Explain floating point store execution path (in terms of
48 ;; Floating point stores go to Ld/St and go to MOV in the floating point
51 ;; Floating point loads go to Ld/St and go to LOAD in the floating point
54 ;; RH> 4) Explain branch on floating condition (in terms of figure 6-1);;
55 ;; Branch on floating condition go to BRU.
57 ;; RH> 5) Is the column for single RECIP instruction latency correct?
58 ;; RH> What about for RSQRT single and double?
60 ;; The latency/repeat for RECIP and RSQRT are correct.
64 ;; Use four automata to isolate long latency operations, and to
65 ;; reduce the complexity of cpu+fpu, reducing space.
67 (define_automaton "sr71_cpu, sr71_cpu1, sr71_cp1, sr71_cp2, sr71_fextra, sr71_imacc")
69 ;; feeders for CPU function units and feeders for fpu (CP1 interface)
70 (define_cpu_unit "sr_iss0,sr_iss1,sr_iss2,sr_iss3,sr_iss4,sr_iss5" "sr71_cpu")
73 (define_cpu_unit "ipu_bru" "sr71_cpu1")
74 (define_cpu_unit "ipu_alux" "sr71_cpu1")
75 (define_cpu_unit "ipu_aluy" "sr71_cpu1")
76 (define_cpu_unit "ipu_ldst" "sr71_cpu1")
77 (define_cpu_unit "ipu_macc_iter" "sr71_imacc")
80 ;; Floating-point unit (Co-processor interface 1).
81 (define_cpu_unit "fpu_mov" "sr71_cp1")
82 (define_cpu_unit "fpu_load" "sr71_cp1")
83 (define_cpu_unit "fpu_fpu" "sr71_cp2")
85 ;; fictitous unit to track long float insns with separate automaton
86 (define_cpu_unit "fpu_iter" "sr71_fextra")
90 ;; Define common execution path (reservation) combinations
94 (define_reservation "cpu_iss" "sr_iss0|sr_iss1|sr_iss2|sr_iss3")
96 ;; two cycles are used for instruction using the fpu as it runs
97 ;; at half the clock speed of the cpu. By adding an extra cycle
98 ;; to the issue units, the default/minimum "repeat" dispatch delay is
99 ;; accounted for all insn.s
100 (define_reservation "cp1_iss" "(sr_iss4*2)|(sr_iss5*2)")
102 (define_reservation "serial_dispatch" "sr_iss0+sr_iss1+sr_iss2+sr_iss3+sr_iss4+sr_iss5")
104 ;; Simulate a 6 insn VLIW dispatch, 1 cycle in dispatch followed by
105 ;; reservation of function unit.
106 (define_reservation "ri_insns" "cpu_iss,(ipu_alux|ipu_aluy)")
107 (define_reservation "ri_mem" "cpu_iss,ipu_ldst")
108 (define_reservation "ri_alux" "cpu_iss,ipu_alux")
109 (define_reservation "ri_branch" "cpu_iss,ipu_bru")
111 (define_reservation "rf_insn" "cp1_iss,fpu_fpu")
112 (define_reservation "rf_ldmem" "cp1_iss,fpu_load")
114 ; simultaneous reservation of pseudo-unit keeps cp1 fpu tied
115 ; up until long cycle insn is finished...
116 (define_reservation "rf_multi1" "rf_insn+fpu_iter")
119 ;; The ordering of the instruction-execution-path/resource-usage
120 ;; descriptions (also known as reservation RTL) is roughly ordered
121 ;; based on the define attribute RTL for the "type" classification.
122 ;; When modifying, remember that the first test that matches is the
127 (define_insn_reservation "ir_sr70_unknown"
129 (and (eq_attr "cpu" "sr71000")
130 (eq_attr "type" "unknown"))
134 ;; Assume prediction fails.
135 (define_insn_reservation "ir_sr70_branch"
137 (and (eq_attr "cpu" "sr71000")
138 (eq_attr "type" "branch,jump,call"))
141 (define_insn_reservation "ir_sr70_load"
143 (and (eq_attr "cpu" "sr71000")
144 (and (eq_attr "type" "load")
145 (eq_attr "mode" "!SF,DF,FPSW")))
148 (define_insn_reservation "ir_sr70_store"
150 (and (eq_attr "cpu" "sr71000")
151 (and (eq_attr "type" "store")
152 (eq_attr "mode" "!SF,DF,FPSW")))
157 ;; float loads/stores flow through both cpu and cp1...
159 (define_insn_reservation "ir_sr70_fload"
161 (and (eq_attr "cpu" "sr71000")
162 (and (eq_attr "type" "load")
163 (eq_attr "mode" "SF,DF")))
164 "(cpu_iss+cp1_iss),(ri_mem+rf_ldmem)")
166 (define_insn_reservation "ir_sr70_fstore"
168 (and (eq_attr "cpu" "sr71000")
169 (and (eq_attr "type" "store")
170 (eq_attr "mode" "SF,DF")))
171 "(cpu_iss+cp1_iss),(fpu_mov+ri_mem)")
174 ;; This reservation is for conditional move based on integer
175 ;; or floating point CC. This could probably use some refinement
176 ;; as "move" type attr seems to be overloaded in rtl.
177 (define_insn_reservation "ir_sr70_move"
179 (and (eq_attr "cpu" "sr71000")
180 (eq_attr "type" "move"))
183 ;; Try to discriminate move-from-cp1 versus move-to-cp1 as latencies
184 ;; are different. Like float load/store, these insns use multiple
185 ;; resources simultaneously
186 (define_insn_reservation "ir_sr70_xfer_from"
188 (and (eq_attr "cpu" "sr71000")
189 (and (eq_attr "type" "xfer")
190 (eq_attr "mode" "!SF,DF,FPSW")))
191 "(cpu_iss+cp1_iss),(fpu_mov+ri_mem)")
193 (define_insn_reservation "ir_sr70_xfer_to"
195 (and (eq_attr "cpu" "sr71000")
196 (and (eq_attr "type" "xfer")
197 (eq_attr "mode" "SF,DF")))
198 "(cpu_iss+cp1_iss),(ri_mem+rf_ldmem)")
200 (define_insn_reservation "ir_sr70_hilo"
202 (and (eq_attr "cpu" "sr71000")
203 (eq_attr "type" "hilo"))
206 (define_insn_reservation "ir_sr70_arith"
208 (and (eq_attr "cpu" "sr71000")
209 (eq_attr "type" "arith,darith,const"))
212 ;; emulate repeat (dispatch stall) by spending extra cycle(s) in
214 (define_insn_reservation "ir_sr70_imul_si"
216 (and (eq_attr "cpu" "sr71000")
217 (and (eq_attr "type" "imul,imadd")
218 (eq_attr "mode" "SI")))
219 "ri_alux,ipu_alux,ipu_macc_iter")
221 (define_insn_reservation "ir_sr70_imul_di"
223 (and (eq_attr "cpu" "sr71000")
224 (and (eq_attr "type" "imul,imadd")
225 (eq_attr "mode" "DI")))
226 "ri_alux,ipu_alux,(ipu_macc_iter*3)")
228 ;; Divide algorithm is early out with best latency of 7 pcycles.
229 ;; Use worst case for scheduling purposes.
230 (define_insn_reservation "ir_sr70_idiv_si"
232 (and (eq_attr "cpu" "sr71000")
233 (and (eq_attr "type" "idiv")
234 (eq_attr "mode" "SI")))
235 "ri_alux,ipu_alux,(ipu_macc_iter*38)")
237 (define_insn_reservation "ir_sr70_idiv_di"
239 (and (eq_attr "cpu" "sr71000")
240 (and (eq_attr "type" "idiv")
241 (eq_attr "mode" "DI")))
242 "ri_alux,ipu_alux,(ipu_macc_iter*70)")
244 (define_insn_reservation "ir_sr70_icmp"
246 (and (eq_attr "cpu" "sr71000")
247 (eq_attr "type" "icmp"))
250 ;; extra reservations of fpu_fpu are for repeat latency
251 (define_insn_reservation "ir_sr70_fadd_sf"
253 (and (eq_attr "cpu" "sr71000")
254 (and (eq_attr "type" "fadd")
255 (eq_attr "mode" "SF")))
258 (define_insn_reservation "ir_sr70_fadd_df"
260 (and (eq_attr "cpu" "sr71000")
261 (and (eq_attr "type" "fadd")
262 (eq_attr "mode" "DF")))
265 ;; Latencies for MADD,MSUB, NMADD, NMSUB assume the Multiply is fused
266 ;; with the sub or add.
267 (define_insn_reservation "ir_sr70_fmul_sf"
269 (and (eq_attr "cpu" "sr71000")
270 (and (eq_attr "type" "fmul,fmadd")
271 (eq_attr "mode" "SF")))
274 ;; tie up the fpu unit to emulate the balance for the "repeat
275 ;; rate" of 8 (2 are spent in the iss unit)
276 (define_insn_reservation "ir_sr70_fmul_df"
278 (and (eq_attr "cpu" "sr71000")
279 (and (eq_attr "type" "fmul,fmadd")
280 (eq_attr "mode" "DF")))
284 ;; RECIP insn uses same type attr as div, and for SR3, has same
285 ;; timings for double. However, single RECIP has a latency of
286 ;; 28 -- only way to fix this is to introduce new insn attrs.
287 ;; cycles spent in iter unit are designed to satisfy balance
288 ;; of "repeat" latency after insn uses up rf_multi1 reservation
289 (define_insn_reservation "ir_sr70_fdiv_sf"
291 (and (eq_attr "cpu" "sr71000")
292 (and (eq_attr "type" "fdiv")
293 (eq_attr "mode" "SF")))
294 "rf_multi1+(fpu_iter*51)")
296 (define_insn_reservation "ir_sr70_fdiv_df"
298 (and (eq_attr "cpu" "sr71000")
299 (and (eq_attr "type" "fdiv")
300 (eq_attr "mode" "DF")))
301 "rf_multi1+(fpu_iter*109)")
303 (define_insn_reservation "ir_sr70_fabs"
305 (and (eq_attr "cpu" "sr71000")
306 (eq_attr "type" "fabs,fneg"))
309 (define_insn_reservation "ir_sr70_fcmp"
311 (and (eq_attr "cpu" "sr71000")
312 (eq_attr "type" "fcmp"))
315 ;; "fcvt" type attribute covers a number of diff insns, most have the same
316 ;; latency descriptions, a few vary. We use the
317 ;; most common timing (which is also worst case).
318 (define_insn_reservation "ir_sr70_fcvt"
320 (and (eq_attr "cpu" "sr71000")
321 (eq_attr "type" "fcvt"))
324 (define_insn_reservation "ir_sr70_fsqrt_sf"
326 (and (eq_attr "cpu" "sr71000")
327 (and (eq_attr "type" "fsqrt")
328 (eq_attr "mode" "SF")))
329 "rf_multi1+(fpu_iter*53)")
331 (define_insn_reservation "ir_sr70_fsqrt_df"
333 (and (eq_attr "cpu" "sr71000")
334 (and (eq_attr "type" "fsqrt")
335 (eq_attr "mode" "DF")))
336 "rf_multi1+(fpu_iter*111)")
338 (define_insn_reservation "ir_sr70_frsqrt_sf"
340 (and (eq_attr "cpu" "sr71000")
341 (and (eq_attr "type" "frsqrt")
342 (eq_attr "mode" "SF")))
343 "rf_multi1+(fpu_iter*39)")
345 (define_insn_reservation "ir_sr70_frsqrt_df"
347 (and (eq_attr "cpu" "sr71000")
348 (and (eq_attr "type" "frsqrt")
349 (eq_attr "mode" "DF")))
350 "rf_multi1+(fpu_iter*229)")
352 (define_insn_reservation "ir_sr70_multi"
354 (and (eq_attr "cpu" "sr71000")
355 (eq_attr "type" "multi"))
358 (define_insn_reservation "ir_sr70_nop"
360 (and (eq_attr "cpu" "sr71000")
361 (eq_attr "type" "nop"))