]>
Commit | Line | Data |
---|---|---|
c3427c7d | 1 | /* Subroutines for insn-output.c for SPARC. |
4592bdcb | 2 | Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998, |
4900aa61 | 3 | 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
62de55f7 | 4 | Free Software Foundation, Inc. |
ab835497 | 5 | Contributed by Michael Tiemann (tiemann@cygnus.com) |
c3427c7d | 6 | 64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans, |
61a55e8b | 7 | at Cygnus Support. |
ab835497 | 8 | |
4d8af13f | 9 | This file is part of GCC. |
ab835497 | 10 | |
4d8af13f | 11 | GCC is free software; you can redistribute it and/or modify |
ab835497 | 12 | it under the terms of the GNU General Public License as published by |
2f83c7d6 | 13 | the Free Software Foundation; either version 3, or (at your option) |
ab835497 RK |
14 | any later version. |
15 | ||
4d8af13f | 16 | GCC is distributed in the hope that it will be useful, |
ab835497 RK |
17 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
18 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
19 | GNU General Public License for more details. | |
20 | ||
21 | You should have received a copy of the GNU General Public License | |
2f83c7d6 NC |
22 | along with GCC; see the file COPYING3. If not see |
23 | <http://www.gnu.org/licenses/>. */ | |
ab835497 | 24 | |
ab835497 | 25 | #include "config.h" |
284d86e9 | 26 | #include "system.h" |
4977bab6 ZW |
27 | #include "coretypes.h" |
28 | #include "tm.h" | |
210aa14a | 29 | #include "tree.h" |
ab835497 RK |
30 | #include "rtl.h" |
31 | #include "regs.h" | |
32 | #include "hard-reg-set.h" | |
33 | #include "real.h" | |
34 | #include "insn-config.h" | |
16f59241 | 35 | #include "insn-codes.h" |
ab835497 | 36 | #include "conditions.h" |
ab835497 RK |
37 | #include "output.h" |
38 | #include "insn-attr.h" | |
39 | #include "flags.h" | |
49ad7cfa | 40 | #include "function.h" |
ab835497 | 41 | #include "expr.h" |
e78d8e51 | 42 | #include "optabs.h" |
ab835497 | 43 | #include "recog.h" |
487a6e06 | 44 | #include "toplev.h" |
d07d525a | 45 | #include "ggc.h" |
b1474bb7 | 46 | #include "tm_p.h" |
e78d8e51 | 47 | #include "debug.h" |
672a6f42 NB |
48 | #include "target.h" |
49 | #include "target-def.h" | |
117dca74 | 50 | #include "cfglayout.h" |
726a989a | 51 | #include "gimple.h" |
16f59241 | 52 | #include "langhooks.h" |
47eb5b32 | 53 | #include "params.h" |
6fb5fa3c | 54 | #include "df.h" |
9fd52c6c | 55 | #include "dwarf2out.h" |
ab835497 | 56 | |
18df6de9 DM |
57 | /* Processor costs */ |
58 | static const | |
59 | struct processor_costs cypress_costs = { | |
ffe384ff DM |
60 | COSTS_N_INSNS (2), /* int load */ |
61 | COSTS_N_INSNS (2), /* int signed load */ | |
62 | COSTS_N_INSNS (2), /* int zeroed load */ | |
63 | COSTS_N_INSNS (2), /* float load */ | |
64 | COSTS_N_INSNS (5), /* fmov, fneg, fabs */ | |
65 | COSTS_N_INSNS (5), /* fadd, fsub */ | |
66 | COSTS_N_INSNS (1), /* fcmp */ | |
67 | COSTS_N_INSNS (1), /* fmov, fmovr */ | |
68 | COSTS_N_INSNS (7), /* fmul */ | |
69 | COSTS_N_INSNS (37), /* fdivs */ | |
70 | COSTS_N_INSNS (37), /* fdivd */ | |
71 | COSTS_N_INSNS (63), /* fsqrts */ | |
72 | COSTS_N_INSNS (63), /* fsqrtd */ | |
73 | COSTS_N_INSNS (1), /* imul */ | |
74 | COSTS_N_INSNS (1), /* imulX */ | |
18df6de9 | 75 | 0, /* imul bit factor */ |
ffe384ff DM |
76 | COSTS_N_INSNS (1), /* idiv */ |
77 | COSTS_N_INSNS (1), /* idivX */ | |
78 | COSTS_N_INSNS (1), /* movcc/movr */ | |
18df6de9 DM |
79 | 0, /* shift penalty */ |
80 | }; | |
81 | ||
82 | static const | |
83 | struct processor_costs supersparc_costs = { | |
ffe384ff DM |
84 | COSTS_N_INSNS (1), /* int load */ |
85 | COSTS_N_INSNS (1), /* int signed load */ | |
86 | COSTS_N_INSNS (1), /* int zeroed load */ | |
87 | COSTS_N_INSNS (0), /* float load */ | |
88 | COSTS_N_INSNS (3), /* fmov, fneg, fabs */ | |
89 | COSTS_N_INSNS (3), /* fadd, fsub */ | |
90 | COSTS_N_INSNS (3), /* fcmp */ | |
91 | COSTS_N_INSNS (1), /* fmov, fmovr */ | |
92 | COSTS_N_INSNS (3), /* fmul */ | |
93 | COSTS_N_INSNS (6), /* fdivs */ | |
94 | COSTS_N_INSNS (9), /* fdivd */ | |
95 | COSTS_N_INSNS (12), /* fsqrts */ | |
96 | COSTS_N_INSNS (12), /* fsqrtd */ | |
97 | COSTS_N_INSNS (4), /* imul */ | |
98 | COSTS_N_INSNS (4), /* imulX */ | |
18df6de9 | 99 | 0, /* imul bit factor */ |
ffe384ff DM |
100 | COSTS_N_INSNS (4), /* idiv */ |
101 | COSTS_N_INSNS (4), /* idivX */ | |
102 | COSTS_N_INSNS (1), /* movcc/movr */ | |
18df6de9 DM |
103 | 1, /* shift penalty */ |
104 | }; | |
105 | ||
106 | static const | |
107 | struct processor_costs hypersparc_costs = { | |
ffe384ff DM |
108 | COSTS_N_INSNS (1), /* int load */ |
109 | COSTS_N_INSNS (1), /* int signed load */ | |
110 | COSTS_N_INSNS (1), /* int zeroed load */ | |
111 | COSTS_N_INSNS (1), /* float load */ | |
112 | COSTS_N_INSNS (1), /* fmov, fneg, fabs */ | |
113 | COSTS_N_INSNS (1), /* fadd, fsub */ | |
114 | COSTS_N_INSNS (1), /* fcmp */ | |
115 | COSTS_N_INSNS (1), /* fmov, fmovr */ | |
116 | COSTS_N_INSNS (1), /* fmul */ | |
117 | COSTS_N_INSNS (8), /* fdivs */ | |
118 | COSTS_N_INSNS (12), /* fdivd */ | |
119 | COSTS_N_INSNS (17), /* fsqrts */ | |
120 | COSTS_N_INSNS (17), /* fsqrtd */ | |
121 | COSTS_N_INSNS (17), /* imul */ | |
122 | COSTS_N_INSNS (17), /* imulX */ | |
18df6de9 | 123 | 0, /* imul bit factor */ |
ffe384ff DM |
124 | COSTS_N_INSNS (17), /* idiv */ |
125 | COSTS_N_INSNS (17), /* idivX */ | |
126 | COSTS_N_INSNS (1), /* movcc/movr */ | |
18df6de9 DM |
127 | 0, /* shift penalty */ |
128 | }; | |
129 | ||
130 | static const | |
131 | struct processor_costs sparclet_costs = { | |
ffe384ff DM |
132 | COSTS_N_INSNS (3), /* int load */ |
133 | COSTS_N_INSNS (3), /* int signed load */ | |
134 | COSTS_N_INSNS (1), /* int zeroed load */ | |
135 | COSTS_N_INSNS (1), /* float load */ | |
136 | COSTS_N_INSNS (1), /* fmov, fneg, fabs */ | |
137 | COSTS_N_INSNS (1), /* fadd, fsub */ | |
138 | COSTS_N_INSNS (1), /* fcmp */ | |
139 | COSTS_N_INSNS (1), /* fmov, fmovr */ | |
140 | COSTS_N_INSNS (1), /* fmul */ | |
141 | COSTS_N_INSNS (1), /* fdivs */ | |
142 | COSTS_N_INSNS (1), /* fdivd */ | |
143 | COSTS_N_INSNS (1), /* fsqrts */ | |
144 | COSTS_N_INSNS (1), /* fsqrtd */ | |
145 | COSTS_N_INSNS (5), /* imul */ | |
146 | COSTS_N_INSNS (5), /* imulX */ | |
18df6de9 | 147 | 0, /* imul bit factor */ |
ffe384ff DM |
148 | COSTS_N_INSNS (5), /* idiv */ |
149 | COSTS_N_INSNS (5), /* idivX */ | |
150 | COSTS_N_INSNS (1), /* movcc/movr */ | |
18df6de9 DM |
151 | 0, /* shift penalty */ |
152 | }; | |
153 | ||
154 | static const | |
155 | struct processor_costs ultrasparc_costs = { | |
ffe384ff DM |
156 | COSTS_N_INSNS (2), /* int load */ |
157 | COSTS_N_INSNS (3), /* int signed load */ | |
158 | COSTS_N_INSNS (2), /* int zeroed load */ | |
159 | COSTS_N_INSNS (2), /* float load */ | |
160 | COSTS_N_INSNS (1), /* fmov, fneg, fabs */ | |
161 | COSTS_N_INSNS (4), /* fadd, fsub */ | |
162 | COSTS_N_INSNS (1), /* fcmp */ | |
163 | COSTS_N_INSNS (2), /* fmov, fmovr */ | |
164 | COSTS_N_INSNS (4), /* fmul */ | |
165 | COSTS_N_INSNS (13), /* fdivs */ | |
166 | COSTS_N_INSNS (23), /* fdivd */ | |
167 | COSTS_N_INSNS (13), /* fsqrts */ | |
168 | COSTS_N_INSNS (23), /* fsqrtd */ | |
169 | COSTS_N_INSNS (4), /* imul */ | |
170 | COSTS_N_INSNS (4), /* imulX */ | |
18df6de9 | 171 | 2, /* imul bit factor */ |
ffe384ff DM |
172 | COSTS_N_INSNS (37), /* idiv */ |
173 | COSTS_N_INSNS (68), /* idivX */ | |
174 | COSTS_N_INSNS (2), /* movcc/movr */ | |
18df6de9 DM |
175 | 2, /* shift penalty */ |
176 | }; | |
177 | ||
178 | static const | |
179 | struct processor_costs ultrasparc3_costs = { | |
ffe384ff DM |
180 | COSTS_N_INSNS (2), /* int load */ |
181 | COSTS_N_INSNS (3), /* int signed load */ | |
182 | COSTS_N_INSNS (3), /* int zeroed load */ | |
183 | COSTS_N_INSNS (2), /* float load */ | |
184 | COSTS_N_INSNS (3), /* fmov, fneg, fabs */ | |
185 | COSTS_N_INSNS (4), /* fadd, fsub */ | |
186 | COSTS_N_INSNS (5), /* fcmp */ | |
187 | COSTS_N_INSNS (3), /* fmov, fmovr */ | |
188 | COSTS_N_INSNS (4), /* fmul */ | |
189 | COSTS_N_INSNS (17), /* fdivs */ | |
190 | COSTS_N_INSNS (20), /* fdivd */ | |
191 | COSTS_N_INSNS (20), /* fsqrts */ | |
192 | COSTS_N_INSNS (29), /* fsqrtd */ | |
193 | COSTS_N_INSNS (6), /* imul */ | |
194 | COSTS_N_INSNS (6), /* imulX */ | |
18df6de9 | 195 | 0, /* imul bit factor */ |
ffe384ff DM |
196 | COSTS_N_INSNS (40), /* idiv */ |
197 | COSTS_N_INSNS (71), /* idivX */ | |
198 | COSTS_N_INSNS (2), /* movcc/movr */ | |
18df6de9 DM |
199 | 0, /* shift penalty */ |
200 | }; | |
201 | ||
4c837a1e DM |
202 | static const |
203 | struct processor_costs niagara_costs = { | |
204 | COSTS_N_INSNS (3), /* int load */ | |
205 | COSTS_N_INSNS (3), /* int signed load */ | |
206 | COSTS_N_INSNS (3), /* int zeroed load */ | |
207 | COSTS_N_INSNS (9), /* float load */ | |
208 | COSTS_N_INSNS (8), /* fmov, fneg, fabs */ | |
209 | COSTS_N_INSNS (8), /* fadd, fsub */ | |
210 | COSTS_N_INSNS (26), /* fcmp */ | |
211 | COSTS_N_INSNS (8), /* fmov, fmovr */ | |
212 | COSTS_N_INSNS (29), /* fmul */ | |
213 | COSTS_N_INSNS (54), /* fdivs */ | |
214 | COSTS_N_INSNS (83), /* fdivd */ | |
215 | COSTS_N_INSNS (100), /* fsqrts - not implemented in hardware */ | |
216 | COSTS_N_INSNS (100), /* fsqrtd - not implemented in hardware */ | |
217 | COSTS_N_INSNS (11), /* imul */ | |
218 | COSTS_N_INSNS (11), /* imulX */ | |
219 | 0, /* imul bit factor */ | |
220 | COSTS_N_INSNS (72), /* idiv */ | |
221 | COSTS_N_INSNS (72), /* idivX */ | |
222 | COSTS_N_INSNS (1), /* movcc/movr */ | |
223 | 0, /* shift penalty */ | |
224 | }; | |
225 | ||
9eeaed6e DM |
226 | static const |
227 | struct processor_costs niagara2_costs = { | |
228 | COSTS_N_INSNS (3), /* int load */ | |
229 | COSTS_N_INSNS (3), /* int signed load */ | |
230 | COSTS_N_INSNS (3), /* int zeroed load */ | |
231 | COSTS_N_INSNS (3), /* float load */ | |
232 | COSTS_N_INSNS (6), /* fmov, fneg, fabs */ | |
233 | COSTS_N_INSNS (6), /* fadd, fsub */ | |
234 | COSTS_N_INSNS (6), /* fcmp */ | |
235 | COSTS_N_INSNS (6), /* fmov, fmovr */ | |
236 | COSTS_N_INSNS (6), /* fmul */ | |
237 | COSTS_N_INSNS (19), /* fdivs */ | |
238 | COSTS_N_INSNS (33), /* fdivd */ | |
239 | COSTS_N_INSNS (19), /* fsqrts */ | |
240 | COSTS_N_INSNS (33), /* fsqrtd */ | |
241 | COSTS_N_INSNS (5), /* imul */ | |
242 | COSTS_N_INSNS (5), /* imulX */ | |
243 | 0, /* imul bit factor */ | |
244 | COSTS_N_INSNS (31), /* idiv, average of 12 - 41 cycle range */ | |
245 | COSTS_N_INSNS (31), /* idivX, average of 12 - 41 cycle range */ | |
246 | COSTS_N_INSNS (1), /* movcc/movr */ | |
247 | 0, /* shift penalty */ | |
248 | }; | |
249 | ||
18df6de9 DM |
250 | const struct processor_costs *sparc_costs = &cypress_costs; |
251 | ||
9ac617d4 EB |
252 | #ifdef HAVE_AS_RELAX_OPTION |
253 | /* If 'as' and 'ld' are relaxing tail call insns into branch always, use | |
254 | "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized. | |
255 | With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if | |
256 | somebody does not branch between the sethi and jmp. */ | |
6244b14e | 257 | #define LEAF_SIBCALL_SLOT_RESERVED_P 1 |
9ac617d4 | 258 | #else |
6244b14e EB |
259 | #define LEAF_SIBCALL_SLOT_RESERVED_P \ |
260 | ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic) | |
9ac617d4 EB |
261 | #endif |
262 | ||
ab835497 RK |
263 | /* Global variables for machine-dependent things. */ |
264 | ||
61a55e8b | 265 | /* Size of frame. Need to know this to emit return insns from leaf procedures. |
9ac617d4 EB |
266 | ACTUAL_FSIZE is set by sparc_compute_frame_size() which is called during the |
267 | reload pass. This is important as the value is later used for scheduling | |
268 | (to see what can go in a delay slot). | |
61a55e8b DE |
269 | APPARENT_FSIZE is the size of the stack less the register save area and less |
270 | the outgoing argument area. It is used when saving call preserved regs. */ | |
6a151f87 EB |
271 | static HOST_WIDE_INT apparent_fsize; |
272 | static HOST_WIDE_INT actual_fsize; | |
61a55e8b | 273 | |
33074e5f RH |
274 | /* Number of live general or floating point registers needed to be |
275 | saved (as 4-byte quantities). */ | |
e48addee JJ |
276 | static int num_gfregs; |
277 | ||
9ac617d4 | 278 | /* The alias set for prologue/epilogue register save/restore. */ |
4862826d | 279 | static GTY(()) alias_set_type sparc_sr_alias_set; |
9ac617d4 | 280 | |
1da4c3a7 | 281 | /* The alias set for the structure return value. */ |
4862826d | 282 | static GTY(()) alias_set_type struct_value_alias_set; |
1da4c3a7 | 283 | |
563c12b0 RH |
284 | /* Vector to say how input registers are mapped to output registers. |
285 | HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to | |
286 | eliminate it. You must use -fomit-frame-pointer to get that. */ | |
d70e94ec | 287 | char leaf_reg_remap[] = |
ab835497 RK |
288 | { 0, 1, 2, 3, 4, 5, 6, 7, |
289 | -1, -1, -1, -1, -1, -1, 14, -1, | |
290 | -1, -1, -1, -1, -1, -1, -1, -1, | |
291 | 8, 9, 10, 11, 12, 13, -1, 15, | |
292 | ||
293 | 32, 33, 34, 35, 36, 37, 38, 39, | |
294 | 40, 41, 42, 43, 44, 45, 46, 47, | |
295 | 48, 49, 50, 51, 52, 53, 54, 55, | |
61a55e8b DE |
296 | 56, 57, 58, 59, 60, 61, 62, 63, |
297 | 64, 65, 66, 67, 68, 69, 70, 71, | |
298 | 72, 73, 74, 75, 76, 77, 78, 79, | |
299 | 80, 81, 82, 83, 84, 85, 86, 87, | |
300 | 88, 89, 90, 91, 92, 93, 94, 95, | |
c4ce6853 | 301 | 96, 97, 98, 99, 100}; |
ab835497 | 302 | |
7d167afd JJ |
303 | /* Vector, indexed by hard register number, which contains 1 |
304 | for a register that is allowable in a candidate for leaf | |
305 | function treatment. */ | |
306 | char sparc_leaf_regs[] = | |
307 | { 1, 1, 1, 1, 1, 1, 1, 1, | |
308 | 0, 0, 0, 0, 0, 0, 1, 0, | |
309 | 0, 0, 0, 0, 0, 0, 0, 0, | |
310 | 1, 1, 1, 1, 1, 1, 0, 1, | |
311 | 1, 1, 1, 1, 1, 1, 1, 1, | |
312 | 1, 1, 1, 1, 1, 1, 1, 1, | |
313 | 1, 1, 1, 1, 1, 1, 1, 1, | |
314 | 1, 1, 1, 1, 1, 1, 1, 1, | |
315 | 1, 1, 1, 1, 1, 1, 1, 1, | |
316 | 1, 1, 1, 1, 1, 1, 1, 1, | |
317 | 1, 1, 1, 1, 1, 1, 1, 1, | |
318 | 1, 1, 1, 1, 1, 1, 1, 1, | |
319 | 1, 1, 1, 1, 1}; | |
320 | ||
d1b38208 | 321 | struct GTY(()) machine_function |
5751a10b JJ |
322 | { |
323 | /* Some local-dynamic TLS symbol name. */ | |
324 | const char *some_ld_name; | |
5be9b7a1 EB |
325 | |
326 | /* True if the current function is leaf and uses only leaf regs, | |
327 | so that the SPARC leaf function optimization can be applied. | |
328 | Private version of current_function_uses_only_leaf_regs, see | |
329 | sparc_expand_prologue for the rationale. */ | |
330 | int leaf_function_p; | |
331 | ||
332 | /* True if the data calculated by sparc_expand_prologue are valid. */ | |
333 | bool prologue_data_valid_p; | |
5751a10b JJ |
334 | }; |
335 | ||
5be9b7a1 EB |
336 | #define sparc_leaf_function_p cfun->machine->leaf_function_p |
337 | #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p | |
338 | ||
9ac617d4 EB |
339 | /* Register we pretend to think the frame pointer is allocated to. |
340 | Normally, this is %fp, but if we are in a leaf procedure, this | |
341 | is %sp+"something". We record "something" separately as it may | |
342 | be too big for reg+constant addressing. */ | |
9ac617d4 | 343 | static rtx frame_base_reg; |
6a151f87 | 344 | static HOST_WIDE_INT frame_base_offset; |
ab835497 | 345 | |
4e5b002b EB |
346 | /* 1 if the next opcode is to be specially indented. */ |
347 | int sparc_indent_opcode = 0; | |
348 | ||
fe609b0f | 349 | static bool sparc_handle_option (size_t, const char *, int); |
e80d5f80 | 350 | static void sparc_init_modes (void); |
5c4f4b18 | 351 | static void scan_record_type (tree, int *, int *, int *); |
e80d5f80 KG |
352 | static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode, |
353 | tree, int, int, int *, int *); | |
354 | ||
355 | static int supersparc_adjust_cost (rtx, rtx, rtx, int); | |
356 | static int hypersparc_adjust_cost (rtx, rtx, rtx, int); | |
357 | ||
358 | static void sparc_output_addr_vec (rtx); | |
359 | static void sparc_output_addr_diff_vec (rtx); | |
360 | static void sparc_output_deferred_case_vectors (void); | |
c6c3dba9 | 361 | static bool sparc_legitimate_address_p (enum machine_mode, rtx, bool); |
3e2cc1d1 | 362 | static rtx sparc_builtin_saveregs (void); |
e80d5f80 KG |
363 | static int epilogue_renumber (rtx *, int); |
364 | static bool sparc_assemble_integer (rtx, unsigned int, int); | |
365 | static int set_extends (rtx); | |
9fd52c6c | 366 | static void load_pic_register (void); |
9ac617d4 | 367 | static int save_or_restore_regs (int, int, rtx, int, int); |
f8ece000 | 368 | static void emit_save_or_restore_regs (int); |
9ac617d4 EB |
369 | static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT); |
370 | static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT); | |
d1ced155 EB |
371 | static void sparc_solaris_elf_asm_named_section (const char *, unsigned int, |
372 | tree) ATTRIBUTE_UNUSED; | |
e80d5f80 KG |
373 | static int sparc_adjust_cost (rtx, rtx, rtx, int); |
374 | static int sparc_issue_rate (void); | |
375 | static void sparc_sched_init (FILE *, int, int); | |
e80d5f80 KG |
376 | static int sparc_use_sched_lookahead (void); |
377 | ||
378 | static void emit_soft_tfmode_libcall (const char *, int, rtx *); | |
379 | static void emit_soft_tfmode_binop (enum rtx_code, rtx *); | |
380 | static void emit_soft_tfmode_unop (enum rtx_code, rtx *); | |
381 | static void emit_soft_tfmode_cvt (enum rtx_code, rtx *); | |
382 | static void emit_hard_tfmode_operation (enum rtx_code, rtx *); | |
383 | ||
384 | static bool sparc_function_ok_for_sibcall (tree, tree); | |
c15c90bb | 385 | static void sparc_init_libfuncs (void); |
16f59241 JM |
386 | static void sparc_init_builtins (void); |
387 | static void sparc_vis_init_builtins (void); | |
388 | static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
f311c3b4 | 389 | static tree sparc_fold_builtin (tree, int, tree *, bool); |
8b0096b4 JM |
390 | static int sparc_vis_mul8x16 (int, int); |
391 | static tree sparc_handle_vis_mul8x16 (int, tree, tree, tree); | |
e80d5f80 KG |
392 | static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, |
393 | HOST_WIDE_INT, tree); | |
3101faab KG |
394 | static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT, |
395 | HOST_WIDE_INT, const_tree); | |
5751a10b JJ |
396 | static struct machine_function * sparc_init_machine_status (void); |
397 | static bool sparc_cannot_force_const_mem (rtx); | |
398 | static rtx sparc_tls_get_addr (void); | |
399 | static rtx sparc_tls_got (void); | |
400 | static const char *get_some_local_dynamic_name (void); | |
401 | static int get_some_local_dynamic_name_1 (rtx *, void *); | |
f40751dd | 402 | static bool sparc_rtx_costs (rtx, int, int, int *, bool); |
586de218 | 403 | static bool sparc_promote_prototypes (const_tree); |
3e2cc1d1 | 404 | static rtx sparc_struct_value_rtx (tree, int); |
cde0f3fd PB |
405 | static enum machine_mode sparc_promote_function_mode (const_tree, enum machine_mode, |
406 | int *, const_tree, int); | |
586de218 | 407 | static bool sparc_return_in_memory (const_tree, const_tree); |
3e2cc1d1 | 408 | static bool sparc_strict_argument_naming (CUMULATIVE_ARGS *); |
d7bd8aeb | 409 | static void sparc_va_start (tree, rtx); |
726a989a | 410 | static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *); |
c75d6010 | 411 | static bool sparc_vector_mode_supported_p (enum machine_mode); |
208384fd EB |
412 | static bool sparc_tls_referenced_p (rtx); |
413 | static rtx legitimize_tls_address (rtx); | |
414 | static rtx legitimize_pic_address (rtx, rtx); | |
506d7b68 | 415 | static rtx sparc_legitimize_address (rtx, rtx, enum machine_mode); |
8cd5a4e0 | 416 | static bool sparc_pass_by_reference (CUMULATIVE_ARGS *, |
586de218 | 417 | enum machine_mode, const_tree, bool); |
78a52f11 RH |
418 | static int sparc_arg_partial_bytes (CUMULATIVE_ARGS *, |
419 | enum machine_mode, tree, bool); | |
e54c7471 | 420 | static void sparc_dwarf_handle_frame_unspec (const char *, rtx, int); |
fdbe66f2 | 421 | static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED; |
5e77eb53 | 422 | static void sparc_file_end (void); |
b52b1749 | 423 | static bool sparc_frame_pointer_required (void); |
7b5cbb57 | 424 | static bool sparc_can_eliminate (const int, const int); |
7269aee7 | 425 | #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING |
3101faab | 426 | static const char *sparc_mangle_type (const_tree); |
7269aee7 | 427 | #endif |
809e78cc | 428 | static void sparc_trampoline_init (rtx, tree, rtx); |
5a82ecd9 | 429 | \f |
07a43492 | 430 | #ifdef SUBTARGET_ATTRIBUTE_TABLE |
5a82ecd9 ILT |
431 | /* Table of valid machine attributes. */ |
432 | static const struct attribute_spec sparc_attribute_table[] = | |
433 | { | |
434 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ | |
435 | SUBTARGET_ATTRIBUTE_TABLE, | |
436 | { NULL, 0, 0, false, false, false, NULL } | |
437 | }; | |
07a43492 | 438 | #endif |
61a55e8b DE |
439 | \f |
440 | /* Option handling. */ | |
441 | ||
a0a301fc DE |
442 | /* Parsed value. */ |
443 | enum cmodel sparc_cmodel; | |
444 | ||
1cb36a98 RH |
445 | char sparc_hard_reg_printed[8]; |
446 | ||
16956f6e DE |
447 | struct sparc_cpu_select sparc_select[] = |
448 | { | |
449 | /* switch name, tune arch */ | |
450 | { (char *)0, "default", 1, 1 }, | |
451 | { (char *)0, "-mcpu=", 1, 1 }, | |
452 | { (char *)0, "-mtune=", 1, 0 }, | |
db3d4438 | 453 | { 0, 0, 0, 0 } |
16956f6e | 454 | }; |
733f53f5 | 455 | |
16956f6e DE |
456 | /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */ |
457 | enum processor_type sparc_cpu; | |
fe609b0f EB |
458 | |
459 | /* Whether\fan FPU option was specified. */ | |
460 | static bool fpu_option_set = false; | |
461 | ||
672a6f42 | 462 | /* Initialize the GCC target structure. */ |
301d03af | 463 | |
d1ced155 | 464 | /* The default is to use .half rather than .short for aligned HI objects. */ |
301d03af RS |
465 | #undef TARGET_ASM_ALIGNED_HI_OP |
466 | #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t" | |
301d03af RS |
467 | |
468 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
469 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t" | |
470 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
471 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t" | |
472 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
473 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t" | |
474 | ||
475 | /* The target hook has to handle DI-mode values. */ | |
476 | #undef TARGET_ASM_INTEGER | |
477 | #define TARGET_ASM_INTEGER sparc_assemble_integer | |
478 | ||
08c148a8 | 479 | #undef TARGET_ASM_FUNCTION_PROLOGUE |
9ac617d4 | 480 | #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue |
08c148a8 | 481 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
9ac617d4 | 482 | #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue |
733f53f5 | 483 | |
c237e94a ZW |
484 | #undef TARGET_SCHED_ADJUST_COST |
485 | #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost | |
486 | #undef TARGET_SCHED_ISSUE_RATE | |
487 | #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate | |
c237e94a ZW |
488 | #undef TARGET_SCHED_INIT |
489 | #define TARGET_SCHED_INIT sparc_sched_init | |
fae15c93 VM |
490 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD |
491 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead | |
c237e94a | 492 | |
4977bab6 ZW |
493 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
494 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall | |
495 | ||
c15c90bb ZW |
496 | #undef TARGET_INIT_LIBFUNCS |
497 | #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs | |
16f59241 JM |
498 | #undef TARGET_INIT_BUILTINS |
499 | #define TARGET_INIT_BUILTINS sparc_init_builtins | |
500 | ||
506d7b68 PB |
501 | #undef TARGET_LEGITIMIZE_ADDRESS |
502 | #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address | |
503 | ||
16f59241 JM |
504 | #undef TARGET_EXPAND_BUILTIN |
505 | #define TARGET_EXPAND_BUILTIN sparc_expand_builtin | |
8b0096b4 JM |
506 | #undef TARGET_FOLD_BUILTIN |
507 | #define TARGET_FOLD_BUILTIN sparc_fold_builtin | |
c15c90bb | 508 | |
fdbe66f2 | 509 | #if TARGET_TLS |
5751a10b JJ |
510 | #undef TARGET_HAVE_TLS |
511 | #define TARGET_HAVE_TLS true | |
512 | #endif | |
fdbe66f2 | 513 | |
5751a10b JJ |
514 | #undef TARGET_CANNOT_FORCE_CONST_MEM |
515 | #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem | |
516 | ||
c590b625 RH |
517 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
518 | #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk | |
3961e8fe | 519 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK |
b9bf58e5 | 520 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk |
c590b625 | 521 | |
3c50106f RH |
522 | #undef TARGET_RTX_COSTS |
523 | #define TARGET_RTX_COSTS sparc_rtx_costs | |
dcefdf67 | 524 | #undef TARGET_ADDRESS_COST |
f40751dd | 525 | #define TARGET_ADDRESS_COST hook_int_rtx_bool_0 |
3c50106f | 526 | |
cde0f3fd PB |
527 | #undef TARGET_PROMOTE_FUNCTION_MODE |
528 | #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode | |
3e2cc1d1 KH |
529 | |
530 | #undef TARGET_PROMOTE_PROTOTYPES | |
531 | #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes | |
532 | ||
533 | #undef TARGET_STRUCT_VALUE_RTX | |
534 | #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx | |
535 | #undef TARGET_RETURN_IN_MEMORY | |
536 | #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory | |
fe984136 RH |
537 | #undef TARGET_MUST_PASS_IN_STACK |
538 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8cd5a4e0 RH |
539 | #undef TARGET_PASS_BY_REFERENCE |
540 | #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference | |
78a52f11 RH |
541 | #undef TARGET_ARG_PARTIAL_BYTES |
542 | #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes | |
3e2cc1d1 KH |
543 | |
544 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
545 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs | |
546 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
547 | #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming | |
548 | ||
d7bd8aeb JJ |
549 | #undef TARGET_EXPAND_BUILTIN_VA_START |
550 | #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start | |
5f393b25 JM |
551 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
552 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg | |
553 | ||
c75d6010 JM |
554 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
555 | #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p | |
556 | ||
e54c7471 EB |
557 | #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC |
558 | #define TARGET_DWARF_HANDLE_FRAME_UNSPEC sparc_dwarf_handle_frame_unspec | |
559 | ||
07a43492 DJ |
560 | #ifdef SUBTARGET_INSERT_ATTRIBUTES |
561 | #undef TARGET_INSERT_ATTRIBUTES | |
562 | #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES | |
563 | #endif | |
564 | ||
565 | #ifdef SUBTARGET_ATTRIBUTE_TABLE | |
566 | #undef TARGET_ATTRIBUTE_TABLE | |
567 | #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table | |
568 | #endif | |
569 | ||
0e4e8dab | 570 | #undef TARGET_RELAXED_ORDERING |
548a6c4f | 571 | #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING |
445cf5eb | 572 | |
fe609b0f EB |
573 | #undef TARGET_DEFAULT_TARGET_FLAGS |
574 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
575 | #undef TARGET_HANDLE_OPTION | |
576 | #define TARGET_HANDLE_OPTION sparc_handle_option | |
577 | ||
f06fb22b | 578 | #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL) |
fdbe66f2 EB |
579 | #undef TARGET_ASM_OUTPUT_DWARF_DTPREL |
580 | #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel | |
581 | #endif | |
582 | ||
c81739f7 | 583 | #undef TARGET_ASM_FILE_END |
5e77eb53 EB |
584 | #define TARGET_ASM_FILE_END sparc_file_end |
585 | ||
b52b1749 AS |
586 | #undef TARGET_FRAME_POINTER_REQUIRED |
587 | #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required | |
588 | ||
7b5cbb57 AS |
589 | #undef TARGET_CAN_ELIMINATE |
590 | #define TARGET_CAN_ELIMINATE sparc_can_eliminate | |
591 | ||
7269aee7 | 592 | #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING |
608063c3 JB |
593 | #undef TARGET_MANGLE_TYPE |
594 | #define TARGET_MANGLE_TYPE sparc_mangle_type | |
7269aee7 AH |
595 | #endif |
596 | ||
c6c3dba9 PB |
597 | #undef TARGET_LEGITIMATE_ADDRESS_P |
598 | #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p | |
599 | ||
809e78cc RH |
600 | #undef TARGET_TRAMPOLINE_INIT |
601 | #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init | |
602 | ||
f6897b10 | 603 | struct gcc_target targetm = TARGET_INITIALIZER; |
fe609b0f EB |
604 | |
605 | /* Implement TARGET_HANDLE_OPTION. */ | |
606 | ||
607 | static bool | |
608 | sparc_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED) | |
609 | { | |
610 | switch (code) | |
611 | { | |
612 | case OPT_mfpu: | |
613 | case OPT_mhard_float: | |
614 | case OPT_msoft_float: | |
615 | fpu_option_set = true; | |
616 | break; | |
617 | ||
618 | case OPT_mcpu_: | |
619 | sparc_select[1].string = arg; | |
620 | break; | |
621 | ||
622 | case OPT_mtune_: | |
623 | sparc_select[2].string = arg; | |
624 | break; | |
fe609b0f EB |
625 | } |
626 | ||
627 | return true; | |
628 | } | |
629 | ||
61a55e8b DE |
630 | /* Validate and override various options, and do some machine dependent |
631 | initialization. */ | |
632 | ||
633 | void | |
fc27d102 | 634 | sparc_override_options (void) |
61a55e8b | 635 | { |
a0a301fc | 636 | static struct code_model { |
8b60264b | 637 | const char *const name; |
32e8bb8e | 638 | const enum cmodel value; |
8b60264b | 639 | } const cmodels[] = { |
a0a301fc DE |
640 | { "32", CM_32 }, |
641 | { "medlow", CM_MEDLOW }, | |
642 | { "medmid", CM_MEDMID }, | |
643 | { "medany", CM_MEDANY }, | |
644 | { "embmedany", CM_EMBMEDANY }, | |
32e8bb8e | 645 | { NULL, (enum cmodel) 0 } |
a0a301fc | 646 | }; |
8b60264b | 647 | const struct code_model *cmodel; |
16956f6e | 648 | /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */ |
733f53f5 | 649 | static struct cpu_default { |
8b60264b KG |
650 | const int cpu; |
651 | const char *const name; | |
652 | } const cpu_default[] = { | |
a0a301fc | 653 | /* There must be one entry here for each TARGET_CPU value. */ |
16956f6e | 654 | { TARGET_CPU_sparc, "cypress" }, |
3592ea0d | 655 | { TARGET_CPU_sparclet, "tsc701" }, |
16956f6e | 656 | { TARGET_CPU_sparclite, "f930" }, |
a0a301fc | 657 | { TARGET_CPU_v8, "v8" }, |
8947065c RH |
658 | { TARGET_CPU_hypersparc, "hypersparc" }, |
659 | { TARGET_CPU_sparclite86x, "sparclite86x" }, | |
a0a301fc DE |
660 | { TARGET_CPU_supersparc, "supersparc" }, |
661 | { TARGET_CPU_v9, "v9" }, | |
ba542005 | 662 | { TARGET_CPU_ultrasparc, "ultrasparc" }, |
fae15c93 | 663 | { TARGET_CPU_ultrasparc3, "ultrasparc3" }, |
4c837a1e | 664 | { TARGET_CPU_niagara, "niagara" }, |
9eeaed6e | 665 | { TARGET_CPU_niagara2, "niagara2" }, |
db3d4438 | 666 | { 0, 0 } |
733f53f5 | 667 | }; |
8b60264b | 668 | const struct cpu_default *def; |
16956f6e | 669 | /* Table of values for -m{cpu,tune}=. */ |
733f53f5 | 670 | static struct cpu_table { |
8b60264b KG |
671 | const char *const name; |
672 | const enum processor_type processor; | |
673 | const int disable; | |
674 | const int enable; | |
675 | } const cpu_table[] = { | |
16956f6e DE |
676 | { "v7", PROCESSOR_V7, MASK_ISA, 0 }, |
677 | { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 }, | |
678 | { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 }, | |
733f53f5 | 679 | /* TI TMS390Z55 supersparc */ |
16956f6e DE |
680 | { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 }, |
681 | { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE }, | |
733f53f5 DE |
682 | /* The Fujitsu MB86930 is the original sparclite chip, with no fpu. |
683 | The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */ | |
16956f6e DE |
684 | { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE }, |
685 | { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU }, | |
8947065c | 686 | { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU }, |
d3ec6b06 VM |
687 | { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU, |
688 | MASK_SPARCLITE }, | |
16956f6e | 689 | { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET }, |
733f53f5 | 690 | /* TEMIC sparclet */ |
3592ea0d | 691 | { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET }, |
16956f6e | 692 | { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 }, |
5cb01b65 JJ |
693 | /* TI ultrasparc I, II, IIi */ |
694 | { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9 | |
695 | /* Although insns using %y are deprecated, it is a clear win on current | |
80ffc95e | 696 | ultrasparcs. */ |
3276910d | 697 | |MASK_DEPRECATED_V8_INSNS}, |
fae15c93 VM |
698 | /* TI ultrasparc III */ |
699 | /* ??? Check if %y issue still holds true in ultra3. */ | |
700 | { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS}, | |
4c837a1e DM |
701 | /* UltraSPARC T1 */ |
702 | { "niagara", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS}, | |
9eeaed6e | 703 | { "niagara2", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9}, |
81f40b79 | 704 | { 0, (enum processor_type) 0, 0, 0 } |
733f53f5 | 705 | }; |
8b60264b KG |
706 | const struct cpu_table *cpu; |
707 | const struct sparc_cpu_select *sel; | |
a0a301fc | 708 | int fpu; |
4900aa61 | 709 | |
a0a301fc DE |
710 | #ifndef SPARC_BI_ARCH |
711 | /* Check for unsupported architecture size. */ | |
712 | if (! TARGET_64BIT != DEFAULT_ARCH32_P) | |
3276910d RK |
713 | error ("%s is not supported by this configuration", |
714 | DEFAULT_ARCH32_P ? "-m64" : "-m32"); | |
a0a301fc DE |
715 | #endif |
716 | ||
4710d3eb JJ |
717 | /* We force all 64bit archs to use 128 bit long double */ |
718 | if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128) | |
719 | { | |
720 | error ("-mlong-double-64 not allowed with -m64"); | |
721 | target_flags |= MASK_LONG_DOUBLE_128; | |
722 | } | |
723 | ||
a0a301fc DE |
724 | /* Code model selection. */ |
725 | sparc_cmodel = SPARC_DEFAULT_CMODEL; | |
4900aa61 | 726 | |
345a6161 DM |
727 | #ifdef SPARC_BI_ARCH |
728 | if (TARGET_ARCH32) | |
729 | sparc_cmodel = CM_32; | |
730 | #endif | |
731 | ||
a0a301fc DE |
732 | if (sparc_cmodel_string != NULL) |
733 | { | |
734 | if (TARGET_ARCH64) | |
735 | { | |
736 | for (cmodel = &cmodels[0]; cmodel->name; cmodel++) | |
737 | if (strcmp (sparc_cmodel_string, cmodel->name) == 0) | |
738 | break; | |
739 | if (cmodel->name == NULL) | |
740 | error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string); | |
741 | else | |
742 | sparc_cmodel = cmodel->value; | |
743 | } | |
744 | else | |
745 | error ("-mcmodel= is not supported on 32 bit systems"); | |
746 | } | |
16956f6e | 747 | |
06063fa2 | 748 | fpu = target_flags & MASK_FPU; /* save current -mfpu status */ |
1f1406b4 | 749 | |
a0a301fc | 750 | /* Set the default CPU. */ |
16956f6e DE |
751 | for (def = &cpu_default[0]; def->name; ++def) |
752 | if (def->cpu == TARGET_CPU_DEFAULT) | |
753 | break; | |
f5f7d171 | 754 | gcc_assert (def->name); |
16956f6e DE |
755 | sparc_select[0].string = def->name; |
756 | ||
757 | for (sel = &sparc_select[0]; sel->name; ++sel) | |
61a55e8b | 758 | { |
16956f6e | 759 | if (sel->string) |
733f53f5 | 760 | { |
16956f6e DE |
761 | for (cpu = &cpu_table[0]; cpu->name; ++cpu) |
762 | if (! strcmp (sel->string, cpu->name)) | |
763 | { | |
764 | if (sel->set_tune_p) | |
765 | sparc_cpu = cpu->processor; | |
766 | ||
767 | if (sel->set_arch_p) | |
768 | { | |
769 | target_flags &= ~cpu->disable; | |
770 | target_flags |= cpu->enable; | |
771 | } | |
772 | break; | |
773 | } | |
774 | ||
775 | if (! cpu->name) | |
776 | error ("bad value (%s) for %s switch", sel->string, sel->name); | |
733f53f5 | 777 | } |
61a55e8b | 778 | } |
97da85b7 | 779 | |
1f1406b4 | 780 | /* If -mfpu or -mno-fpu was explicitly used, don't override with |
fe609b0f EB |
781 | the processor default. */ |
782 | if (fpu_option_set) | |
783 | target_flags = (target_flags & ~MASK_FPU) | fpu; | |
1f1406b4 | 784 | |
a5774a7d JJ |
785 | /* Don't allow -mvis if FPU is disabled. */ |
786 | if (! TARGET_FPU) | |
787 | target_flags &= ~MASK_VIS; | |
788 | ||
789 | /* -mvis assumes UltraSPARC+, so we are sure v9 instructions | |
790 | are available. | |
791 | -m64 also implies v9. */ | |
792 | if (TARGET_VIS || TARGET_ARCH64) | |
c4031a04 JJ |
793 | { |
794 | target_flags |= MASK_V9; | |
795 | target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE); | |
796 | } | |
a5774a7d | 797 | |
16956f6e DE |
798 | /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */ |
799 | if (TARGET_V9 && TARGET_ARCH32) | |
733f53f5 | 800 | target_flags |= MASK_DEPRECATED_V8_INSNS; |
61a55e8b | 801 | |
345a6161 DM |
802 | /* V8PLUS requires V9, makes no sense in 64 bit mode. */ |
803 | if (! TARGET_V9 || TARGET_ARCH64) | |
284d86e9 JC |
804 | target_flags &= ~MASK_V8PLUS; |
805 | ||
82d6b402 RH |
806 | /* Don't use stack biasing in 32 bit mode. */ |
807 | if (TARGET_ARCH32) | |
808 | target_flags &= ~MASK_STACK_BIAS; | |
4900aa61 | 809 | |
efa3896a | 810 | /* Supply a default value for align_functions. */ |
fae15c93 VM |
811 | if (align_functions == 0 |
812 | && (sparc_cpu == PROCESSOR_ULTRASPARC | |
4c837a1e | 813 | || sparc_cpu == PROCESSOR_ULTRASPARC3 |
9eeaed6e DM |
814 | || sparc_cpu == PROCESSOR_NIAGARA |
815 | || sparc_cpu == PROCESSOR_NIAGARA2)) | |
efa3896a | 816 | align_functions = 32; |
bf62bbf1 | 817 | |
82d6b402 RH |
818 | /* Validate PCC_STRUCT_RETURN. */ |
819 | if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN) | |
820 | flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1); | |
821 | ||
301d03af RS |
822 | /* Only use .uaxword when compiling for a 64-bit target. */ |
823 | if (!TARGET_ARCH64) | |
824 | targetm.asm_out.unaligned_op.di = NULL; | |
825 | ||
61a55e8b DE |
826 | /* Do various machine dependent initializations. */ |
827 | sparc_init_modes (); | |
5751a10b | 828 | |
1da4c3a7 | 829 | /* Acquire unique alias sets for our private stuff. */ |
9ac617d4 | 830 | sparc_sr_alias_set = new_alias_set (); |
1da4c3a7 | 831 | struct_value_alias_set = new_alias_set (); |
9ac617d4 | 832 | |
5751a10b JJ |
833 | /* Set up function hooks. */ |
834 | init_machine_status = sparc_init_machine_status; | |
18df6de9 DM |
835 | |
836 | switch (sparc_cpu) | |
837 | { | |
838 | case PROCESSOR_V7: | |
839 | case PROCESSOR_CYPRESS: | |
840 | sparc_costs = &cypress_costs; | |
841 | break; | |
842 | case PROCESSOR_V8: | |
843 | case PROCESSOR_SPARCLITE: | |
844 | case PROCESSOR_SUPERSPARC: | |
845 | sparc_costs = &supersparc_costs; | |
846 | break; | |
847 | case PROCESSOR_F930: | |
848 | case PROCESSOR_F934: | |
849 | case PROCESSOR_HYPERSPARC: | |
850 | case PROCESSOR_SPARCLITE86X: | |
851 | sparc_costs = &hypersparc_costs; | |
852 | break; | |
853 | case PROCESSOR_SPARCLET: | |
854 | case PROCESSOR_TSC701: | |
855 | sparc_costs = &sparclet_costs; | |
856 | break; | |
857 | case PROCESSOR_V9: | |
858 | case PROCESSOR_ULTRASPARC: | |
859 | sparc_costs = &ultrasparc_costs; | |
860 | break; | |
861 | case PROCESSOR_ULTRASPARC3: | |
862 | sparc_costs = &ultrasparc3_costs; | |
863 | break; | |
4c837a1e DM |
864 | case PROCESSOR_NIAGARA: |
865 | sparc_costs = &niagara_costs; | |
866 | break; | |
9eeaed6e DM |
867 | case PROCESSOR_NIAGARA2: |
868 | sparc_costs = &niagara2_costs; | |
869 | break; | |
18df6de9 | 870 | }; |
ed965309 JJ |
871 | |
872 | #ifdef TARGET_DEFAULT_LONG_DOUBLE_128 | |
873 | if (!(target_flags_explicit & MASK_LONG_DOUBLE_128)) | |
874 | target_flags |= MASK_LONG_DOUBLE_128; | |
875 | #endif | |
47eb5b32 ZD |
876 | |
877 | if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES)) | |
878 | set_param_value ("simultaneous-prefetches", | |
879 | ((sparc_cpu == PROCESSOR_ULTRASPARC | |
9eeaed6e DM |
880 | || sparc_cpu == PROCESSOR_NIAGARA |
881 | || sparc_cpu == PROCESSOR_NIAGARA2) | |
47eb5b32 ZD |
882 | ? 2 |
883 | : (sparc_cpu == PROCESSOR_ULTRASPARC3 | |
884 | ? 8 : 3))); | |
885 | if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE)) | |
4900aa61 | 886 | set_param_value ("l1-cache-line-size", |
47eb5b32 ZD |
887 | ((sparc_cpu == PROCESSOR_ULTRASPARC |
888 | || sparc_cpu == PROCESSOR_ULTRASPARC3 | |
9eeaed6e DM |
889 | || sparc_cpu == PROCESSOR_NIAGARA |
890 | || sparc_cpu == PROCESSOR_NIAGARA2) | |
47eb5b32 | 891 | ? 64 : 32)); |
61a55e8b DE |
892 | } |
893 | \f | |
cd5fb1ee DE |
894 | /* Miscellaneous utilities. */ |
895 | ||
896 | /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move | |
897 | or branch on register contents instructions. */ | |
898 | ||
899 | int | |
fc27d102 | 900 | v9_regcmp_p (enum rtx_code code) |
cd5fb1ee DE |
901 | { |
902 | return (code == EQ || code == NE || code == GE || code == LT | |
903 | || code == LE || code == GT); | |
904 | } | |
bfd6bc60 | 905 | |
62190128 DM |
906 | /* Nonzero if OP is a floating point constant which can |
907 | be loaded into an integer register using a single | |
908 | sethi instruction. */ | |
909 | ||
910 | int | |
fc27d102 | 911 | fp_sethi_p (rtx op) |
62190128 DM |
912 | { |
913 | if (GET_CODE (op) == CONST_DOUBLE) | |
914 | { | |
915 | REAL_VALUE_TYPE r; | |
916 | long i; | |
917 | ||
918 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
62190128 | 919 | REAL_VALUE_TO_TARGET_SINGLE (r, i); |
9a9e266b | 920 | return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i); |
62190128 DM |
921 | } |
922 | ||
923 | return 0; | |
924 | } | |
925 | ||
926 | /* Nonzero if OP is a floating point constant which can | |
927 | be loaded into an integer register using a single | |
928 | mov instruction. */ | |
929 | ||
930 | int | |
fc27d102 | 931 | fp_mov_p (rtx op) |
62190128 DM |
932 | { |
933 | if (GET_CODE (op) == CONST_DOUBLE) | |
934 | { | |
935 | REAL_VALUE_TYPE r; | |
936 | long i; | |
937 | ||
938 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
62190128 | 939 | REAL_VALUE_TO_TARGET_SINGLE (r, i); |
9a9e266b | 940 | return SPARC_SIMM13_P (i); |
62190128 DM |
941 | } |
942 | ||
943 | return 0; | |
944 | } | |
945 | ||
946 | /* Nonzero if OP is a floating point constant which can | |
947 | be loaded into an integer register using a high/losum | |
948 | instruction sequence. */ | |
949 | ||
950 | int | |
fc27d102 | 951 | fp_high_losum_p (rtx op) |
62190128 DM |
952 | { |
953 | /* The constraints calling this should only be in | |
954 | SFmode move insns, so any constant which cannot | |
955 | be moved using a single insn will do. */ | |
956 | if (GET_CODE (op) == CONST_DOUBLE) | |
957 | { | |
958 | REAL_VALUE_TYPE r; | |
959 | long i; | |
960 | ||
961 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
62190128 | 962 | REAL_VALUE_TO_TARGET_SINGLE (r, i); |
9a9e266b | 963 | return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i); |
62190128 DM |
964 | } |
965 | ||
966 | return 0; | |
967 | } | |
968 | ||
bea5071f EB |
969 | /* Expand a move instruction. Return true if all work is done. */ |
970 | ||
971 | bool | |
972 | sparc_expand_move (enum machine_mode mode, rtx *operands) | |
973 | { | |
974 | /* Handle sets of MEM first. */ | |
975 | if (GET_CODE (operands[0]) == MEM) | |
976 | { | |
977 | /* 0 is a register (or a pair of registers) on SPARC. */ | |
978 | if (register_or_zero_operand (operands[1], mode)) | |
979 | return false; | |
980 | ||
981 | if (!reload_in_progress) | |
982 | { | |
983 | operands[0] = validize_mem (operands[0]); | |
984 | operands[1] = force_reg (mode, operands[1]); | |
985 | } | |
986 | } | |
987 | ||
988 | /* Fixup TLS cases. */ | |
38fc66ba EB |
989 | if (TARGET_HAVE_TLS |
990 | && CONSTANT_P (operands[1]) | |
38fc66ba EB |
991 | && sparc_tls_referenced_p (operands [1])) |
992 | { | |
208384fd EB |
993 | operands[1] = legitimize_tls_address (operands[1]); |
994 | return false; | |
38fc66ba | 995 | } |
4900aa61 | 996 | |
bea5071f EB |
997 | /* Fixup PIC cases. */ |
998 | if (flag_pic && CONSTANT_P (operands[1])) | |
999 | { | |
1000 | if (pic_address_needs_scratch (operands[1])) | |
208384fd | 1001 | operands[1] = legitimize_pic_address (operands[1], NULL_RTX); |
bea5071f | 1002 | |
1910440e RS |
1003 | /* VxWorks does not impose a fixed gap between segments; the run-time |
1004 | gap can be different from the object-file gap. We therefore can't | |
1005 | assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we | |
1006 | are absolutely sure that X is in the same segment as the GOT. | |
1007 | Unfortunately, the flexibility of linker scripts means that we | |
1008 | can't be sure of that in general, so assume that _G_O_T_-relative | |
1009 | accesses are never valid on VxWorks. */ | |
1010 | if (GET_CODE (operands[1]) == LABEL_REF && !TARGET_VXWORKS_RTP) | |
bea5071f | 1011 | { |
1910440e RS |
1012 | if (mode == SImode) |
1013 | { | |
1014 | emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1])); | |
1015 | return true; | |
1016 | } | |
bea5071f | 1017 | |
1910440e RS |
1018 | if (mode == DImode) |
1019 | { | |
1020 | gcc_assert (TARGET_ARCH64); | |
1021 | emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1])); | |
1022 | return true; | |
1023 | } | |
bea5071f EB |
1024 | } |
1025 | ||
1026 | if (symbolic_operand (operands[1], mode)) | |
1027 | { | |
1028 | operands[1] = legitimize_pic_address (operands[1], | |
208384fd EB |
1029 | reload_in_progress |
1030 | ? operands[0] : NULL_RTX); | |
bea5071f EB |
1031 | return false; |
1032 | } | |
1033 | } | |
1034 | ||
1035 | /* If we are trying to toss an integer constant into FP registers, | |
1036 | or loading a FP or vector constant, force it into memory. */ | |
1037 | if (CONSTANT_P (operands[1]) | |
1038 | && REG_P (operands[0]) | |
1039 | && (SPARC_FP_REG_P (REGNO (operands[0])) | |
1040 | || SCALAR_FLOAT_MODE_P (mode) | |
1041 | || VECTOR_MODE_P (mode))) | |
1042 | { | |
1043 | /* emit_group_store will send such bogosity to us when it is | |
1044 | not storing directly into memory. So fix this up to avoid | |
1045 | crashes in output_constant_pool. */ | |
1046 | if (operands [1] == const0_rtx) | |
1047 | operands[1] = CONST0_RTX (mode); | |
1048 | ||
1049 | /* We can clear FP registers if TARGET_VIS, and always other regs. */ | |
1050 | if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG) | |
1051 | && const_zero_operand (operands[1], mode)) | |
1052 | return false; | |
1053 | ||
1054 | if (REGNO (operands[0]) < SPARC_FIRST_FP_REG | |
1055 | /* We are able to build any SF constant in integer registers | |
1056 | with at most 2 instructions. */ | |
1057 | && (mode == SFmode | |
1058 | /* And any DF constant in integer registers. */ | |
1059 | || (mode == DFmode | |
1060 | && (reload_completed || reload_in_progress)))) | |
1061 | return false; | |
1062 | ||
1063 | operands[1] = force_const_mem (mode, operands[1]); | |
1064 | if (!reload_in_progress) | |
1065 | operands[1] = validize_mem (operands[1]); | |
1066 | return false; | |
1067 | } | |
1068 | ||
1069 | /* Accept non-constants and valid constants unmodified. */ | |
1070 | if (!CONSTANT_P (operands[1]) | |
1071 | || GET_CODE (operands[1]) == HIGH | |
1072 | || input_operand (operands[1], mode)) | |
1073 | return false; | |
1074 | ||
1075 | switch (mode) | |
1076 | { | |
1077 | case QImode: | |
1078 | /* All QImode constants require only one insn, so proceed. */ | |
1079 | break; | |
1080 | ||
1081 | case HImode: | |
1082 | case SImode: | |
1083 | sparc_emit_set_const32 (operands[0], operands[1]); | |
1084 | return true; | |
1085 | ||
1086 | case DImode: | |
1087 | /* input_operand should have filtered out 32-bit mode. */ | |
1088 | sparc_emit_set_const64 (operands[0], operands[1]); | |
1089 | return true; | |
4900aa61 | 1090 | |
bea5071f EB |
1091 | default: |
1092 | gcc_unreachable (); | |
1093 | } | |
1094 | ||
1095 | return false; | |
1096 | } | |
1097 | ||
1098 | /* Load OP1, a 32-bit constant, into OP0, a register. | |
1099 | We know it can't be done in one insn when we get | |
1100 | here, the move expander guarantees this. */ | |
1101 | ||
e0d80184 | 1102 | void |
fc27d102 | 1103 | sparc_emit_set_const32 (rtx op0, rtx op1) |
e0d80184 DM |
1104 | { |
1105 | enum machine_mode mode = GET_MODE (op0); | |
1106 | rtx temp; | |
1107 | ||
e0d80184 DM |
1108 | if (reload_in_progress || reload_completed) |
1109 | temp = op0; | |
284d86e9 | 1110 | else |
e0d80184 DM |
1111 | temp = gen_reg_rtx (mode); |
1112 | ||
71648202 DM |
1113 | if (GET_CODE (op1) == CONST_INT) |
1114 | { | |
0e5d569c EB |
1115 | gcc_assert (!small_int_operand (op1, mode) |
1116 | && !const_high_operand (op1, mode)); | |
1117 | ||
71648202 DM |
1118 | /* Emit them as real moves instead of a HIGH/LO_SUM, |
1119 | this way CSE can see everything and reuse intermediate | |
1120 | values if it wants. */ | |
0e5d569c EB |
1121 | emit_insn (gen_rtx_SET (VOIDmode, temp, |
1122 | GEN_INT (INTVAL (op1) | |
1123 | & ~(HOST_WIDE_INT)0x3ff))); | |
f2827474 | 1124 | |
71648202 DM |
1125 | emit_insn (gen_rtx_SET (VOIDmode, |
1126 | op0, | |
f2827474 | 1127 | gen_rtx_IOR (mode, temp, |
71648202 DM |
1128 | GEN_INT (INTVAL (op1) & 0x3ff)))); |
1129 | } | |
1130 | else | |
1131 | { | |
1132 | /* A symbol, emit in the traditional way. */ | |
f2827474 RK |
1133 | emit_insn (gen_rtx_SET (VOIDmode, temp, |
1134 | gen_rtx_HIGH (mode, op1))); | |
71648202 | 1135 | emit_insn (gen_rtx_SET (VOIDmode, |
f2827474 | 1136 | op0, gen_rtx_LO_SUM (mode, temp, op1))); |
71648202 | 1137 | } |
e0d80184 DM |
1138 | } |
1139 | ||
1d511ed5 | 1140 | /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register. |
0e40b5f2 | 1141 | If TEMP is nonzero, we are forbidden to use any other scratch |
1d511ed5 EB |
1142 | registers. Otherwise, we are allowed to generate them as needed. |
1143 | ||
1144 | Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY | |
1145 | or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */ | |
bea5071f | 1146 | |
e0d80184 | 1147 | void |
1d511ed5 | 1148 | sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp) |
e0d80184 | 1149 | { |
1d511ed5 EB |
1150 | rtx temp1, temp2, temp3, temp4, temp5; |
1151 | rtx ti_temp = 0; | |
3968de80 | 1152 | |
1d511ed5 | 1153 | if (temp && GET_MODE (temp) == TImode) |
3968de80 | 1154 | { |
1d511ed5 EB |
1155 | ti_temp = temp; |
1156 | temp = gen_rtx_REG (DImode, REGNO (temp)); | |
3968de80 DD |
1157 | } |
1158 | ||
1d511ed5 | 1159 | /* SPARC-V9 code-model support. */ |
e0d80184 | 1160 | switch (sparc_cmodel) |
284d86e9 | 1161 | { |
e0d80184 DM |
1162 | case CM_MEDLOW: |
1163 | /* The range spanned by all instructions in the object is less | |
1164 | than 2^31 bytes (2GB) and the distance from any instruction | |
1165 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1166 | than 2^31 bytes (2GB). | |
ab835497 | 1167 | |
e0d80184 DM |
1168 | The executable must be in the low 4TB of the virtual address |
1169 | space. | |
61a55e8b | 1170 | |
1d511ed5 EB |
1171 | sethi %hi(symbol), %temp1 |
1172 | or %temp1, %lo(symbol), %reg */ | |
1173 | if (temp) | |
1174 | temp1 = temp; /* op0 is allowed. */ | |
1175 | else | |
1176 | temp1 = gen_reg_rtx (DImode); | |
1177 | ||
9208e4b2 DM |
1178 | emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1))); |
1179 | emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1))); | |
e0d80184 | 1180 | break; |
61a55e8b | 1181 | |
e0d80184 DM |
1182 | case CM_MEDMID: |
1183 | /* The range spanned by all instructions in the object is less | |
1184 | than 2^31 bytes (2GB) and the distance from any instruction | |
1185 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1186 | than 2^31 bytes (2GB). | |
1187 | ||
1188 | The executable must be in the low 16TB of the virtual address | |
1189 | space. | |
1190 | ||
1191 | sethi %h44(symbol), %temp1 | |
1192 | or %temp1, %m44(symbol), %temp2 | |
1193 | sllx %temp2, 12, %temp3 | |
1194 | or %temp3, %l44(symbol), %reg */ | |
1d511ed5 EB |
1195 | if (temp) |
1196 | { | |
1197 | temp1 = op0; | |
1198 | temp2 = op0; | |
1199 | temp3 = temp; /* op0 is allowed. */ | |
1200 | } | |
1201 | else | |
1202 | { | |
1203 | temp1 = gen_reg_rtx (DImode); | |
1204 | temp2 = gen_reg_rtx (DImode); | |
1205 | temp3 = gen_reg_rtx (DImode); | |
1206 | } | |
1207 | ||
1208 | emit_insn (gen_seth44 (temp1, op1)); | |
1209 | emit_insn (gen_setm44 (temp2, temp1, op1)); | |
1210 | emit_insn (gen_rtx_SET (VOIDmode, temp3, | |
1211 | gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12)))); | |
1212 | emit_insn (gen_setl44 (op0, temp3, op1)); | |
e0d80184 | 1213 | break; |
61a55e8b | 1214 | |
e0d80184 DM |
1215 | case CM_MEDANY: |
1216 | /* The range spanned by all instructions in the object is less | |
1217 | than 2^31 bytes (2GB) and the distance from any instruction | |
1218 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1219 | than 2^31 bytes (2GB). | |
1220 | ||
1221 | The executable can be placed anywhere in the virtual address | |
1222 | space. | |
1223 | ||
1224 | sethi %hh(symbol), %temp1 | |
1225 | sethi %lm(symbol), %temp2 | |
1226 | or %temp1, %hm(symbol), %temp3 | |
1d511ed5 EB |
1227 | sllx %temp3, 32, %temp4 |
1228 | or %temp4, %temp2, %temp5 | |
1229 | or %temp5, %lo(symbol), %reg */ | |
1230 | if (temp) | |
3968de80 | 1231 | { |
1d511ed5 EB |
1232 | /* It is possible that one of the registers we got for operands[2] |
1233 | might coincide with that of operands[0] (which is why we made | |
1234 | it TImode). Pick the other one to use as our scratch. */ | |
1235 | if (rtx_equal_p (temp, op0)) | |
1236 | { | |
f5f7d171 JM |
1237 | gcc_assert (ti_temp); |
1238 | temp = gen_rtx_REG (DImode, REGNO (temp) + 1); | |
1d511ed5 EB |
1239 | } |
1240 | temp1 = op0; | |
1241 | temp2 = temp; /* op0 is _not_ allowed, see above. */ | |
1242 | temp3 = op0; | |
1243 | temp4 = op0; | |
1244 | temp5 = op0; | |
1245 | } | |
1246 | else | |
1247 | { | |
1248 | temp1 = gen_reg_rtx (DImode); | |
1249 | temp2 = gen_reg_rtx (DImode); | |
1250 | temp3 = gen_reg_rtx (DImode); | |
1251 | temp4 = gen_reg_rtx (DImode); | |
1252 | temp5 = gen_reg_rtx (DImode); | |
3968de80 | 1253 | } |
61a55e8b | 1254 | |
1d511ed5 EB |
1255 | emit_insn (gen_sethh (temp1, op1)); |
1256 | emit_insn (gen_setlm (temp2, op1)); | |
1257 | emit_insn (gen_sethm (temp3, temp1, op1)); | |
1258 | emit_insn (gen_rtx_SET (VOIDmode, temp4, | |
1259 | gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32)))); | |
1260 | emit_insn (gen_rtx_SET (VOIDmode, temp5, | |
1261 | gen_rtx_PLUS (DImode, temp4, temp2))); | |
1262 | emit_insn (gen_setlo (op0, temp5, op1)); | |
e0d80184 | 1263 | break; |
61a55e8b | 1264 | |
e0d80184 DM |
1265 | case CM_EMBMEDANY: |
1266 | /* Old old old backwards compatibility kruft here. | |
1267 | Essentially it is MEDLOW with a fixed 64-bit | |
1268 | virtual base added to all data segment addresses. | |
1269 | Text-segment stuff is computed like MEDANY, we can't | |
1270 | reuse the code above because the relocation knobs | |
1271 | look different. | |
1272 | ||
1273 | Data segment: sethi %hi(symbol), %temp1 | |
1d511ed5 EB |
1274 | add %temp1, EMBMEDANY_BASE_REG, %temp2 |
1275 | or %temp2, %lo(symbol), %reg */ | |
e0d80184 DM |
1276 | if (data_segment_operand (op1, GET_MODE (op1))) |
1277 | { | |
1d511ed5 EB |
1278 | if (temp) |
1279 | { | |
1280 | temp1 = temp; /* op0 is allowed. */ | |
1281 | temp2 = op0; | |
1282 | } | |
1283 | else | |
1284 | { | |
1285 | temp1 = gen_reg_rtx (DImode); | |
1286 | temp2 = gen_reg_rtx (DImode); | |
1287 | } | |
1288 | ||
e0d80184 | 1289 | emit_insn (gen_embmedany_sethi (temp1, op1)); |
1d511ed5 EB |
1290 | emit_insn (gen_embmedany_brsum (temp2, temp1)); |
1291 | emit_insn (gen_embmedany_losum (op0, temp2, op1)); | |
e0d80184 | 1292 | } |
1d511ed5 EB |
1293 | |
1294 | /* Text segment: sethi %uhi(symbol), %temp1 | |
1295 | sethi %hi(symbol), %temp2 | |
1296 | or %temp1, %ulo(symbol), %temp3 | |
1297 | sllx %temp3, 32, %temp4 | |
1298 | or %temp4, %temp2, %temp5 | |
1299 | or %temp5, %lo(symbol), %reg */ | |
e0d80184 DM |
1300 | else |
1301 | { | |
1d511ed5 | 1302 | if (temp) |
3968de80 | 1303 | { |
1d511ed5 EB |
1304 | /* It is possible that one of the registers we got for operands[2] |
1305 | might coincide with that of operands[0] (which is why we made | |
1306 | it TImode). Pick the other one to use as our scratch. */ | |
1307 | if (rtx_equal_p (temp, op0)) | |
1308 | { | |
f5f7d171 JM |
1309 | gcc_assert (ti_temp); |
1310 | temp = gen_rtx_REG (DImode, REGNO (temp) + 1); | |
1d511ed5 EB |
1311 | } |
1312 | temp1 = op0; | |
1313 | temp2 = temp; /* op0 is _not_ allowed, see above. */ | |
1314 | temp3 = op0; | |
1315 | temp4 = op0; | |
1316 | temp5 = op0; | |
1317 | } | |
1318 | else | |
1319 | { | |
1320 | temp1 = gen_reg_rtx (DImode); | |
1321 | temp2 = gen_reg_rtx (DImode); | |
1322 | temp3 = gen_reg_rtx (DImode); | |
1323 | temp4 = gen_reg_rtx (DImode); | |
1324 | temp5 = gen_reg_rtx (DImode); | |
3968de80 | 1325 | } |
61a55e8b | 1326 | |
1d511ed5 EB |
1327 | emit_insn (gen_embmedany_textuhi (temp1, op1)); |
1328 | emit_insn (gen_embmedany_texthi (temp2, op1)); | |
1329 | emit_insn (gen_embmedany_textulo (temp3, temp1, op1)); | |
1330 | emit_insn (gen_rtx_SET (VOIDmode, temp4, | |
1331 | gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32)))); | |
1332 | emit_insn (gen_rtx_SET (VOIDmode, temp5, | |
1333 | gen_rtx_PLUS (DImode, temp4, temp2))); | |
1334 | emit_insn (gen_embmedany_textlo (op0, temp5, op1)); | |
e0d80184 DM |
1335 | } |
1336 | break; | |
1337 | ||
1338 | default: | |
f5f7d171 | 1339 | gcc_unreachable (); |
e0d80184 DM |
1340 | } |
1341 | } | |
1342 | ||
0e5d569c EB |
1343 | #if HOST_BITS_PER_WIDE_INT == 32 |
1344 | void | |
1345 | sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED) | |
1346 | { | |
1347 | gcc_unreachable (); | |
1348 | } | |
1349 | #else | |
9208e4b2 DM |
1350 | /* These avoid problems when cross compiling. If we do not |
1351 | go through all this hair then the optimizer will see | |
1352 | invalid REG_EQUAL notes or in some cases none at all. */ | |
9bb3323d | 1353 | static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT); |
e80d5f80 KG |
1354 | static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT); |
1355 | static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT); | |
1356 | static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT); | |
9208e4b2 | 1357 | |
9208e4b2 DM |
1358 | /* The optimizer is not to assume anything about exactly |
1359 | which bits are set for a HIGH, they are unspecified. | |
1360 | Unfortunately this leads to many missed optimizations | |
1361 | during CSE. We mask out the non-HIGH bits, and matches | |
1362 | a plain movdi, to alleviate this problem. */ | |
9bb3323d EB |
1363 | static rtx |
1364 | gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val) | |
9208e4b2 | 1365 | { |
9bb3323d | 1366 | return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff)); |
9208e4b2 | 1367 | } |
2a01c939 | 1368 | |
e0d80184 | 1369 | static rtx |
fc27d102 | 1370 | gen_safe_SET64 (rtx dest, HOST_WIDE_INT val) |
61a55e8b | 1371 | { |
0e5d569c | 1372 | return gen_rtx_SET (VOIDmode, dest, GEN_INT (val)); |
9208e4b2 DM |
1373 | } |
1374 | ||
1375 | static rtx | |
fc27d102 | 1376 | gen_safe_OR64 (rtx src, HOST_WIDE_INT val) |
9208e4b2 | 1377 | { |
0e5d569c | 1378 | return gen_rtx_IOR (DImode, src, GEN_INT (val)); |
9208e4b2 DM |
1379 | } |
1380 | ||
1381 | static rtx | |
fc27d102 | 1382 | gen_safe_XOR64 (rtx src, HOST_WIDE_INT val) |
9208e4b2 | 1383 | { |
0e5d569c | 1384 | return gen_rtx_XOR (DImode, src, GEN_INT (val)); |
e0d80184 | 1385 | } |
61a55e8b | 1386 | |
e0d80184 DM |
1387 | /* Worker routines for 64-bit constant formation on arch64. |
1388 | One of the key things to be doing in these emissions is | |
1389 | to create as many temp REGs as possible. This makes it | |
1390 | possible for half-built constants to be used later when | |
1391 | such values are similar to something required later on. | |
1392 | Without doing this, the optimizer cannot see such | |
1393 | opportunities. */ | |
2a01c939 | 1394 | |
e80d5f80 KG |
1395 | static void sparc_emit_set_const64_quick1 (rtx, rtx, |
1396 | unsigned HOST_WIDE_INT, int); | |
2a01c939 | 1397 | |
e0d80184 | 1398 | static void |
fc27d102 KG |
1399 | sparc_emit_set_const64_quick1 (rtx op0, rtx temp, |
1400 | unsigned HOST_WIDE_INT low_bits, int is_neg) | |
e0d80184 | 1401 | { |
2a01c939 | 1402 | unsigned HOST_WIDE_INT high_bits; |
a9e27770 | 1403 | |
e0d80184 | 1404 | if (is_neg) |
2a01c939 | 1405 | high_bits = (~low_bits) & 0xffffffff; |
31bc9eba | 1406 | else |
e0d80184 | 1407 | high_bits = low_bits; |
31bc9eba | 1408 | |
9bb3323d | 1409 | emit_insn (gen_safe_HIGH64 (temp, high_bits)); |
e0d80184 | 1410 | if (!is_neg) |
5108f3e5 DM |
1411 | { |
1412 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1413 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); | |
1414 | } | |
e0d80184 | 1415 | else |
5108f3e5 DM |
1416 | { |
1417 | /* If we are XOR'ing with -1, then we should emit a one's complement | |
1418 | instead. This way the combiner will notice logical operations | |
1419 | such as ANDN later on and substitute. */ | |
1420 | if ((low_bits & 0x3ff) == 0x3ff) | |
1421 | { | |
1422 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1423 | gen_rtx_NOT (DImode, temp))); | |
1424 | } | |
1425 | else | |
1426 | { | |
1427 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1428 | gen_safe_XOR64 (temp, | |
9e0625a3 AO |
1429 | (-(HOST_WIDE_INT)0x400 |
1430 | | (low_bits & 0x3ff))))); | |
5108f3e5 DM |
1431 | } |
1432 | } | |
e0d80184 | 1433 | } |
61a55e8b | 1434 | |
e80d5f80 KG |
1435 | static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT, |
1436 | unsigned HOST_WIDE_INT, int); | |
2a01c939 | 1437 | |
e0d80184 | 1438 | static void |
fc27d102 KG |
1439 | sparc_emit_set_const64_quick2 (rtx op0, rtx temp, |
1440 | unsigned HOST_WIDE_INT high_bits, | |
1441 | unsigned HOST_WIDE_INT low_immediate, | |
1442 | int shift_count) | |
e0d80184 DM |
1443 | { |
1444 | rtx temp2 = op0; | |
61a55e8b | 1445 | |
e0d80184 DM |
1446 | if ((high_bits & 0xfffffc00) != 0) |
1447 | { | |
9bb3323d | 1448 | emit_insn (gen_safe_HIGH64 (temp, high_bits)); |
e0d80184 | 1449 | if ((high_bits & ~0xfffffc00) != 0) |
9208e4b2 DM |
1450 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
1451 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); | |
61a55e8b | 1452 | else |
e0d80184 | 1453 | temp2 = temp; |
61a55e8b DE |
1454 | } |
1455 | else | |
1456 | { | |
f710f868 | 1457 | emit_insn (gen_safe_SET64 (temp, high_bits)); |
e0d80184 DM |
1458 | temp2 = temp; |
1459 | } | |
61a55e8b | 1460 | |
80ffc95e | 1461 | /* Now shift it up into place. */ |
9208e4b2 | 1462 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1463 | gen_rtx_ASHIFT (DImode, temp2, |
1464 | GEN_INT (shift_count)))); | |
61a55e8b | 1465 | |
e0d80184 DM |
1466 | /* If there is a low immediate part piece, finish up by |
1467 | putting that in as well. */ | |
1468 | if (low_immediate != 0) | |
9208e4b2 DM |
1469 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
1470 | gen_safe_OR64 (op0, low_immediate))); | |
ab835497 RK |
1471 | } |
1472 | ||
e80d5f80 KG |
1473 | static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT, |
1474 | unsigned HOST_WIDE_INT); | |
2a01c939 | 1475 | |
e0d80184 DM |
1476 | /* Full 64-bit constant decomposition. Even though this is the |
1477 | 'worst' case, we still optimize a few things away. */ | |
1478 | static void | |
fc27d102 KG |
1479 | sparc_emit_set_const64_longway (rtx op0, rtx temp, |
1480 | unsigned HOST_WIDE_INT high_bits, | |
1481 | unsigned HOST_WIDE_INT low_bits) | |
ab835497 | 1482 | { |
e0d80184 | 1483 | rtx sub_temp; |
4bf7135b | 1484 | |
e0d80184 DM |
1485 | if (reload_in_progress || reload_completed) |
1486 | sub_temp = op0; | |
1487 | else | |
1488 | sub_temp = gen_reg_rtx (DImode); | |
1489 | ||
1490 | if ((high_bits & 0xfffffc00) != 0) | |
1491 | { | |
9bb3323d | 1492 | emit_insn (gen_safe_HIGH64 (temp, high_bits)); |
e0d80184 | 1493 | if ((high_bits & ~0xfffffc00) != 0) |
9208e4b2 | 1494 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1495 | sub_temp, |
9208e4b2 | 1496 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); |
e0d80184 DM |
1497 | else |
1498 | sub_temp = temp; | |
1499 | } | |
1500 | else | |
ab835497 | 1501 | { |
f710f868 | 1502 | emit_insn (gen_safe_SET64 (temp, high_bits)); |
e0d80184 | 1503 | sub_temp = temp; |
ab835497 | 1504 | } |
4bf7135b | 1505 | |
e0d80184 | 1506 | if (!reload_in_progress && !reload_completed) |
284d86e9 | 1507 | { |
e0d80184 DM |
1508 | rtx temp2 = gen_reg_rtx (DImode); |
1509 | rtx temp3 = gen_reg_rtx (DImode); | |
1510 | rtx temp4 = gen_reg_rtx (DImode); | |
284d86e9 | 1511 | |
9208e4b2 | 1512 | emit_insn (gen_rtx_SET (VOIDmode, temp4, |
e0d80184 DM |
1513 | gen_rtx_ASHIFT (DImode, sub_temp, |
1514 | GEN_INT (32)))); | |
4bf7135b | 1515 | |
9bb3323d | 1516 | emit_insn (gen_safe_HIGH64 (temp2, low_bits)); |
e0d80184 | 1517 | if ((low_bits & ~0xfffffc00) != 0) |
f710f868 DM |
1518 | { |
1519 | emit_insn (gen_rtx_SET (VOIDmode, temp3, | |
1520 | gen_safe_OR64 (temp2, (low_bits & 0x3ff)))); | |
1521 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1522 | gen_rtx_PLUS (DImode, temp4, temp3))); | |
1523 | } | |
1524 | else | |
1525 | { | |
1526 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1527 | gen_rtx_PLUS (DImode, temp4, temp2))); | |
1528 | } | |
284d86e9 | 1529 | } |
e0d80184 DM |
1530 | else |
1531 | { | |
9208e4b2 DM |
1532 | rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff); |
1533 | rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff); | |
1534 | rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff); | |
e0d80184 | 1535 | int to_shift = 12; |
284d86e9 | 1536 | |
e0d80184 DM |
1537 | /* We are in the middle of reload, so this is really |
1538 | painful. However we do still make an attempt to | |
5108f3e5 | 1539 | avoid emitting truly stupid code. */ |
e0d80184 DM |
1540 | if (low1 != const0_rtx) |
1541 | { | |
9208e4b2 | 1542 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 | 1543 | gen_rtx_ASHIFT (DImode, sub_temp, |
2a01c939 | 1544 | GEN_INT (to_shift)))); |
9208e4b2 | 1545 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1546 | gen_rtx_IOR (DImode, op0, low1))); |
1547 | sub_temp = op0; | |
1548 | to_shift = 12; | |
1549 | } | |
1550 | else | |
1551 | { | |
1552 | to_shift += 12; | |
1553 | } | |
1554 | if (low2 != const0_rtx) | |
1555 | { | |
9208e4b2 | 1556 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1557 | gen_rtx_ASHIFT (DImode, sub_temp, |
1558 | GEN_INT (to_shift)))); | |
9208e4b2 | 1559 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1560 | gen_rtx_IOR (DImode, op0, low2))); |
1561 | sub_temp = op0; | |
1562 | to_shift = 8; | |
1563 | } | |
1564 | else | |
1565 | { | |
1566 | to_shift += 8; | |
1567 | } | |
9208e4b2 | 1568 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1569 | gen_rtx_ASHIFT (DImode, sub_temp, |
1570 | GEN_INT (to_shift)))); | |
1571 | if (low3 != const0_rtx) | |
9208e4b2 | 1572 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 | 1573 | gen_rtx_IOR (DImode, op0, low3))); |
80ffc95e | 1574 | /* phew... */ |
e0d80184 | 1575 | } |
284d86e9 JC |
1576 | } |
1577 | ||
80ffc95e | 1578 | /* Analyze a 64-bit constant for certain properties. */ |
e80d5f80 KG |
1579 | static void analyze_64bit_constant (unsigned HOST_WIDE_INT, |
1580 | unsigned HOST_WIDE_INT, | |
1581 | int *, int *, int *); | |
2a01c939 | 1582 | |
e0d80184 | 1583 | static void |
fc27d102 KG |
1584 | analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits, |
1585 | unsigned HOST_WIDE_INT low_bits, | |
1586 | int *hbsp, int *lbsp, int *abbasp) | |
ab835497 | 1587 | { |
e0d80184 DM |
1588 | int lowest_bit_set, highest_bit_set, all_bits_between_are_set; |
1589 | int i; | |
ab835497 | 1590 | |
e0d80184 DM |
1591 | lowest_bit_set = highest_bit_set = -1; |
1592 | i = 0; | |
1593 | do | |
ab835497 | 1594 | { |
e0d80184 DM |
1595 | if ((lowest_bit_set == -1) |
1596 | && ((low_bits >> i) & 1)) | |
1597 | lowest_bit_set = i; | |
1598 | if ((highest_bit_set == -1) | |
1599 | && ((high_bits >> (32 - i - 1)) & 1)) | |
1600 | highest_bit_set = (64 - i - 1); | |
1601 | } | |
1602 | while (++i < 32 | |
1603 | && ((highest_bit_set == -1) | |
1604 | || (lowest_bit_set == -1))); | |
1605 | if (i == 32) | |
1606 | { | |
1607 | i = 0; | |
1608 | do | |
ab835497 | 1609 | { |
e0d80184 DM |
1610 | if ((lowest_bit_set == -1) |
1611 | && ((high_bits >> i) & 1)) | |
1612 | lowest_bit_set = i + 32; | |
1613 | if ((highest_bit_set == -1) | |
1614 | && ((low_bits >> (32 - i - 1)) & 1)) | |
1615 | highest_bit_set = 32 - i - 1; | |
ab835497 | 1616 | } |
e0d80184 DM |
1617 | while (++i < 32 |
1618 | && ((highest_bit_set == -1) | |
1619 | || (lowest_bit_set == -1))); | |
1620 | } | |
1621 | /* If there are no bits set this should have gone out | |
1622 | as one instruction! */ | |
f5f7d171 | 1623 | gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1); |
e0d80184 DM |
1624 | all_bits_between_are_set = 1; |
1625 | for (i = lowest_bit_set; i <= highest_bit_set; i++) | |
1626 | { | |
1627 | if (i < 32) | |
1628 | { | |
1629 | if ((low_bits & (1 << i)) != 0) | |
1630 | continue; | |
1631 | } | |
1632 | else | |
1633 | { | |
1634 | if ((high_bits & (1 << (i - 32))) != 0) | |
1635 | continue; | |
1636 | } | |
1637 | all_bits_between_are_set = 0; | |
1638 | break; | |
ab835497 | 1639 | } |
e0d80184 DM |
1640 | *hbsp = highest_bit_set; |
1641 | *lbsp = lowest_bit_set; | |
1642 | *abbasp = all_bits_between_are_set; | |
ab835497 | 1643 | } |
6a4bb1fa | 1644 | |
e80d5f80 | 1645 | static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT); |
2a01c939 | 1646 | |
e0d80184 | 1647 | static int |
fc27d102 KG |
1648 | const64_is_2insns (unsigned HOST_WIDE_INT high_bits, |
1649 | unsigned HOST_WIDE_INT low_bits) | |
e0d80184 DM |
1650 | { |
1651 | int highest_bit_set, lowest_bit_set, all_bits_between_are_set; | |
284d86e9 | 1652 | |
e0d80184 | 1653 | if (high_bits == 0 |
f451b552 | 1654 | || high_bits == 0xffffffff) |
e0d80184 | 1655 | return 1; |
6a4bb1fa | 1656 | |
e0d80184 DM |
1657 | analyze_64bit_constant (high_bits, low_bits, |
1658 | &highest_bit_set, &lowest_bit_set, | |
1659 | &all_bits_between_are_set); | |
6a4bb1fa | 1660 | |
f710f868 DM |
1661 | if ((highest_bit_set == 63 |
1662 | || lowest_bit_set == 0) | |
e0d80184 DM |
1663 | && all_bits_between_are_set != 0) |
1664 | return 1; | |
6a4bb1fa | 1665 | |
9208e4b2 DM |
1666 | if ((highest_bit_set - lowest_bit_set) < 21) |
1667 | return 1; | |
1668 | ||
6a4bb1fa DE |
1669 | return 0; |
1670 | } | |
1671 | ||
e80d5f80 KG |
1672 | static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT, |
1673 | unsigned HOST_WIDE_INT, | |
1674 | int, int); | |
2a01c939 DM |
1675 | |
1676 | static unsigned HOST_WIDE_INT | |
fc27d102 KG |
1677 | create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits, |
1678 | unsigned HOST_WIDE_INT low_bits, | |
1679 | int lowest_bit_set, int shift) | |
ab835497 | 1680 | { |
f710f868 | 1681 | HOST_WIDE_INT hi, lo; |
ab835497 | 1682 | |
e0d80184 DM |
1683 | if (lowest_bit_set < 32) |
1684 | { | |
1685 | lo = (low_bits >> lowest_bit_set) << shift; | |
1686 | hi = ((high_bits << (32 - lowest_bit_set)) << shift); | |
1687 | } | |
1688 | else | |
1689 | { | |
1690 | lo = 0; | |
1691 | hi = ((high_bits >> (lowest_bit_set - 32)) << shift); | |
1692 | } | |
f5f7d171 | 1693 | gcc_assert (! (hi & lo)); |
e0d80184 DM |
1694 | return (hi | lo); |
1695 | } | |
ab835497 | 1696 | |
e0d80184 DM |
1697 | /* Here we are sure to be arch64 and this is an integer constant |
1698 | being loaded into a register. Emit the most efficient | |
1699 | insn sequence possible. Detection of all the 1-insn cases | |
1700 | has been done already. */ | |
1701 | void | |
fc27d102 | 1702 | sparc_emit_set_const64 (rtx op0, rtx op1) |
e0d80184 | 1703 | { |
2a01c939 | 1704 | unsigned HOST_WIDE_INT high_bits, low_bits; |
e0d80184 DM |
1705 | int lowest_bit_set, highest_bit_set; |
1706 | int all_bits_between_are_set; | |
1d511ed5 | 1707 | rtx temp = 0; |
ab835497 | 1708 | |
e0d80184 | 1709 | /* Sanity check that we know what we are working with. */ |
bea5071f EB |
1710 | gcc_assert (TARGET_ARCH64 |
1711 | && (GET_CODE (op0) == SUBREG | |
1712 | || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0))))); | |
380f6ad3 | 1713 | |
f710f868 DM |
1714 | if (reload_in_progress || reload_completed) |
1715 | temp = op0; | |
f710f868 | 1716 | |
0e5d569c | 1717 | if (GET_CODE (op1) != CONST_INT) |
e0d80184 | 1718 | { |
5f78aa30 JC |
1719 | sparc_emit_set_symbolic_const64 (op0, op1, temp); |
1720 | return; | |
e0d80184 | 1721 | } |
ab835497 | 1722 | |
1d511ed5 EB |
1723 | if (! temp) |
1724 | temp = gen_reg_rtx (DImode); | |
1725 | ||
0e5d569c EB |
1726 | high_bits = ((INTVAL (op1) >> 32) & 0xffffffff); |
1727 | low_bits = (INTVAL (op1) & 0xffffffff); | |
ab835497 | 1728 | |
e0d80184 DM |
1729 | /* low_bits bits 0 --> 31 |
1730 | high_bits bits 32 --> 63 */ | |
ab835497 | 1731 | |
e0d80184 DM |
1732 | analyze_64bit_constant (high_bits, low_bits, |
1733 | &highest_bit_set, &lowest_bit_set, | |
1734 | &all_bits_between_are_set); | |
1735 | ||
1736 | /* First try for a 2-insn sequence. */ | |
1737 | ||
1738 | /* These situations are preferred because the optimizer can | |
1739 | * do more things with them: | |
1740 | * 1) mov -1, %reg | |
1741 | * sllx %reg, shift, %reg | |
1742 | * 2) mov -1, %reg | |
1743 | * srlx %reg, shift, %reg | |
1744 | * 3) mov some_small_const, %reg | |
1745 | * sllx %reg, shift, %reg | |
1746 | */ | |
1747 | if (((highest_bit_set == 63 | |
1748 | || lowest_bit_set == 0) | |
1749 | && all_bits_between_are_set != 0) | |
2a01c939 | 1750 | || ((highest_bit_set - lowest_bit_set) < 12)) |
e0d80184 | 1751 | { |
9208e4b2 | 1752 | HOST_WIDE_INT the_const = -1; |
e0d80184 DM |
1753 | int shift = lowest_bit_set; |
1754 | ||
f710f868 DM |
1755 | if ((highest_bit_set != 63 |
1756 | && lowest_bit_set != 0) | |
1757 | || all_bits_between_are_set == 0) | |
ab835497 | 1758 | { |
e0d80184 | 1759 | the_const = |
9208e4b2 | 1760 | create_simple_focus_bits (high_bits, low_bits, |
9208e4b2 | 1761 | lowest_bit_set, 0); |
ab835497 | 1762 | } |
e0d80184 | 1763 | else if (lowest_bit_set == 0) |
9208e4b2 | 1764 | shift = -(63 - highest_bit_set); |
e0d80184 | 1765 | |
f5f7d171 JM |
1766 | gcc_assert (SPARC_SIMM13_P (the_const)); |
1767 | gcc_assert (shift != 0); | |
f710f868 | 1768 | |
9208e4b2 | 1769 | emit_insn (gen_safe_SET64 (temp, the_const)); |
e0d80184 | 1770 | if (shift > 0) |
9208e4b2 | 1771 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 DM |
1772 | op0, |
1773 | gen_rtx_ASHIFT (DImode, | |
1774 | temp, | |
1775 | GEN_INT (shift)))); | |
1776 | else if (shift < 0) | |
9208e4b2 | 1777 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1778 | op0, |
9208e4b2 | 1779 | gen_rtx_LSHIFTRT (DImode, |
e0d80184 DM |
1780 | temp, |
1781 | GEN_INT (-shift)))); | |
e0d80184 DM |
1782 | return; |
1783 | } | |
ab835497 | 1784 | |
e0d80184 DM |
1785 | /* Now a range of 22 or less bits set somewhere. |
1786 | * 1) sethi %hi(focus_bits), %reg | |
1787 | * sllx %reg, shift, %reg | |
1788 | * 2) sethi %hi(focus_bits), %reg | |
1789 | * srlx %reg, shift, %reg | |
1790 | */ | |
2a01c939 | 1791 | if ((highest_bit_set - lowest_bit_set) < 21) |
e0d80184 | 1792 | { |
2a01c939 | 1793 | unsigned HOST_WIDE_INT focus_bits = |
e0d80184 | 1794 | create_simple_focus_bits (high_bits, low_bits, |
ddf80874 | 1795 | lowest_bit_set, 10); |
f710f868 | 1796 | |
f5f7d171 JM |
1797 | gcc_assert (SPARC_SETHI_P (focus_bits)); |
1798 | gcc_assert (lowest_bit_set != 10); | |
f710f868 | 1799 | |
9bb3323d | 1800 | emit_insn (gen_safe_HIGH64 (temp, focus_bits)); |
e0d80184 | 1801 | |
9208e4b2 | 1802 | /* If lowest_bit_set == 10 then a sethi alone could have done it. */ |
e0d80184 | 1803 | if (lowest_bit_set < 10) |
9208e4b2 | 1804 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1805 | op0, |
9208e4b2 | 1806 | gen_rtx_LSHIFTRT (DImode, temp, |
e0d80184 | 1807 | GEN_INT (10 - lowest_bit_set)))); |
9208e4b2 DM |
1808 | else if (lowest_bit_set > 10) |
1809 | emit_insn (gen_rtx_SET (VOIDmode, | |
e0d80184 DM |
1810 | op0, |
1811 | gen_rtx_ASHIFT (DImode, temp, | |
1812 | GEN_INT (lowest_bit_set - 10)))); | |
e0d80184 DM |
1813 | return; |
1814 | } | |
1815 | ||
1816 | /* 1) sethi %hi(low_bits), %reg | |
1817 | * or %reg, %lo(low_bits), %reg | |
1818 | * 2) sethi %hi(~low_bits), %reg | |
1819 | * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg | |
1820 | */ | |
1821 | if (high_bits == 0 | |
2a01c939 | 1822 | || high_bits == 0xffffffff) |
5f78aa30 JC |
1823 | { |
1824 | sparc_emit_set_const64_quick1 (op0, temp, low_bits, | |
1825 | (high_bits == 0xffffffff)); | |
1826 | return; | |
1827 | } | |
e0d80184 | 1828 | |
f710f868 DM |
1829 | /* Now, try 3-insn sequences. */ |
1830 | ||
e0d80184 DM |
1831 | /* 1) sethi %hi(high_bits), %reg |
1832 | * or %reg, %lo(high_bits), %reg | |
1833 | * sllx %reg, 32, %reg | |
1834 | */ | |
f710f868 | 1835 | if (low_bits == 0) |
5f78aa30 | 1836 | { |
f710f868 | 1837 | sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32); |
5f78aa30 JC |
1838 | return; |
1839 | } | |
e0d80184 | 1840 | |
f710f868 DM |
1841 | /* We may be able to do something quick |
1842 | when the constant is negated, so try that. */ | |
e0d80184 DM |
1843 | if (const64_is_2insns ((~high_bits) & 0xffffffff, |
1844 | (~low_bits) & 0xfffffc00)) | |
1845 | { | |
5108f3e5 DM |
1846 | /* NOTE: The trailing bits get XOR'd so we need the |
1847 | non-negated bits, not the negated ones. */ | |
1848 | unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff; | |
e0d80184 DM |
1849 | |
1850 | if ((((~high_bits) & 0xffffffff) == 0 | |
1851 | && ((~low_bits) & 0x80000000) == 0) | |
1852 | || (((~high_bits) & 0xffffffff) == 0xffffffff | |
1853 | && ((~low_bits) & 0x80000000) != 0)) | |
ab835497 | 1854 | { |
0e5d569c | 1855 | unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff); |
e0d80184 | 1856 | |
5108f3e5 DM |
1857 | if ((SPARC_SETHI_P (fast_int) |
1858 | && (~high_bits & 0xffffffff) == 0) | |
1859 | || SPARC_SIMM13_P (fast_int)) | |
9208e4b2 | 1860 | emit_insn (gen_safe_SET64 (temp, fast_int)); |
ab835497 | 1861 | else |
0e5d569c | 1862 | sparc_emit_set_const64 (temp, GEN_INT (fast_int)); |
ab835497 | 1863 | } |
e0d80184 DM |
1864 | else |
1865 | { | |
1866 | rtx negated_const; | |
e0d80184 DM |
1867 | negated_const = GEN_INT (((~low_bits) & 0xfffffc00) | |
1868 | (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32)); | |
e0d80184 DM |
1869 | sparc_emit_set_const64 (temp, negated_const); |
1870 | } | |
5108f3e5 DM |
1871 | |
1872 | /* If we are XOR'ing with -1, then we should emit a one's complement | |
1873 | instead. This way the combiner will notice logical operations | |
1874 | such as ANDN later on and substitute. */ | |
1875 | if (trailing_bits == 0x3ff) | |
1876 | { | |
1877 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1878 | gen_rtx_NOT (DImode, temp))); | |
1879 | } | |
1880 | else | |
1881 | { | |
1882 | emit_insn (gen_rtx_SET (VOIDmode, | |
1883 | op0, | |
1884 | gen_safe_XOR64 (temp, | |
1885 | (-0x400 | trailing_bits)))); | |
1886 | } | |
e0d80184 | 1887 | return; |
ab835497 | 1888 | } |
ab835497 | 1889 | |
e0d80184 DM |
1890 | /* 1) sethi %hi(xxx), %reg |
1891 | * or %reg, %lo(xxx), %reg | |
1892 | * sllx %reg, yyy, %reg | |
f710f868 DM |
1893 | * |
1894 | * ??? This is just a generalized version of the low_bits==0 | |
1895 | * thing above, FIXME... | |
e0d80184 DM |
1896 | */ |
1897 | if ((highest_bit_set - lowest_bit_set) < 32) | |
1898 | { | |
f710f868 DM |
1899 | unsigned HOST_WIDE_INT focus_bits = |
1900 | create_simple_focus_bits (high_bits, low_bits, | |
ddf80874 | 1901 | lowest_bit_set, 0); |
ab835497 | 1902 | |
e0d80184 | 1903 | /* We can't get here in this state. */ |
f5f7d171 | 1904 | gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32); |
95726648 | 1905 | |
e0d80184 DM |
1906 | /* So what we know is that the set bits straddle the |
1907 | middle of the 64-bit word. */ | |
5f78aa30 JC |
1908 | sparc_emit_set_const64_quick2 (op0, temp, |
1909 | focus_bits, 0, | |
1910 | lowest_bit_set); | |
1911 | return; | |
e0d80184 | 1912 | } |
c85f7c16 | 1913 | |
f710f868 DM |
1914 | /* 1) sethi %hi(high_bits), %reg |
1915 | * or %reg, %lo(high_bits), %reg | |
1916 | * sllx %reg, 32, %reg | |
1917 | * or %reg, low_bits, %reg | |
1918 | */ | |
1919 | if (SPARC_SIMM13_P(low_bits) | |
1920 | && ((int)low_bits > 0)) | |
1921 | { | |
1922 | sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32); | |
1923 | return; | |
1924 | } | |
1925 | ||
80ffc95e | 1926 | /* The easiest way when all else fails, is full decomposition. */ |
e0d80184 DM |
1927 | #if 0 |
1928 | printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n", | |
1929 | high_bits, low_bits, ~high_bits, ~low_bits); | |
1930 | #endif | |
1931 | sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits); | |
c85f7c16 | 1932 | } |
0e5d569c | 1933 | #endif /* HOST_BITS_PER_WIDE_INT == 32 */ |
c85f7c16 | 1934 | |
e267e177 RH |
1935 | /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE, |
1936 | return the mode to be used for the comparison. For floating-point, | |
1937 | CCFP[E]mode is used. CC_NOOVmode should be used when the first operand | |
1938 | is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special | |
1939 | processing is needed. */ | |
1940 | ||
1941 | enum machine_mode | |
fc27d102 | 1942 | select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED) |
e267e177 RH |
1943 | { |
1944 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1945 | { | |
1946 | switch (op) | |
1947 | { | |
1948 | case EQ: | |
1949 | case NE: | |
1950 | case UNORDERED: | |
1951 | case ORDERED: | |
1952 | case UNLT: | |
1953 | case UNLE: | |
1954 | case UNGT: | |
1955 | case UNGE: | |
1956 | case UNEQ: | |
7913f3d0 | 1957 | case LTGT: |
e267e177 RH |
1958 | return CCFPmode; |
1959 | ||
1960 | case LT: | |
1961 | case LE: | |
1962 | case GT: | |
1963 | case GE: | |
1964 | return CCFPEmode; | |
1965 | ||
1966 | default: | |
f5f7d171 | 1967 | gcc_unreachable (); |
e267e177 RH |
1968 | } |
1969 | } | |
1970 | else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS | |
1971 | || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT) | |
1972 | { | |
1973 | if (TARGET_ARCH64 && GET_MODE (x) == DImode) | |
1974 | return CCX_NOOVmode; | |
1975 | else | |
1976 | return CC_NOOVmode; | |
1977 | } | |
1978 | else | |
1979 | { | |
1980 | if (TARGET_ARCH64 && GET_MODE (x) == DImode) | |
1981 | return CCXmode; | |
1982 | else | |
1983 | return CCmode; | |
1984 | } | |
1985 | } | |
1986 | ||
f90b7a5a PB |
1987 | /* Emit the compare insn and return the CC reg for a CODE comparison |
1988 | with operands X and Y. */ | |
c85f7c16 | 1989 | |
f90b7a5a PB |
1990 | static rtx |
1991 | gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y) | |
c85f7c16 | 1992 | { |
65c5ae8b | 1993 | enum machine_mode mode; |
f90b7a5a | 1994 | rtx cc_reg; |
c85f7c16 | 1995 | |
f90b7a5a PB |
1996 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC) |
1997 | return x; | |
65c5ae8b | 1998 | |
65c5ae8b | 1999 | mode = SELECT_CC_MODE (code, x, y); |
8ec11fe9 | 2000 | |
e0d80184 DM |
2001 | /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the |
2002 | fcc regs (cse can't tell they're really call clobbered regs and will | |
2003 | remove a duplicate comparison even if there is an intervening function | |
2004 | call - it will then try to reload the cc reg via an int reg which is why | |
2005 | we need the movcc patterns). It is possible to provide the movcc | |
2006 | patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two | |
2007 | registers (say %g1,%g5) and it takes about 6 insns. A better fix would be | |
2008 | to tell cse that CCFPE mode registers (even pseudos) are call | |
2009 | clobbered. */ | |
c85f7c16 | 2010 | |
e0d80184 DM |
2011 | /* ??? This is an experiment. Rather than making changes to cse which may |
2012 | or may not be easy/clean, we do our own cse. This is possible because | |
2013 | we will generate hard registers. Cse knows they're call clobbered (it | |
2014 | doesn't know the same thing about pseudos). If we guess wrong, no big | |
2015 | deal, but if we win, great! */ | |
c85f7c16 | 2016 | |
e0d80184 DM |
2017 | if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) |
2018 | #if 1 /* experiment */ | |
284d86e9 | 2019 | { |
e0d80184 DM |
2020 | int reg; |
2021 | /* We cycle through the registers to ensure they're all exercised. */ | |
2022 | static int next_fcc_reg = 0; | |
2023 | /* Previous x,y for each fcc reg. */ | |
2024 | static rtx prev_args[4][2]; | |
284d86e9 | 2025 | |
e0d80184 DM |
2026 | /* Scan prev_args for x,y. */ |
2027 | for (reg = 0; reg < 4; reg++) | |
2028 | if (prev_args[reg][0] == x && prev_args[reg][1] == y) | |
2029 | break; | |
2030 | if (reg == 4) | |
2031 | { | |
2032 | reg = next_fcc_reg; | |
2033 | prev_args[reg][0] = x; | |
2034 | prev_args[reg][1] = y; | |
2035 | next_fcc_reg = (next_fcc_reg + 1) & 3; | |
2036 | } | |
2037 | cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG); | |
284d86e9 | 2038 | } |
e0d80184 DM |
2039 | #else |
2040 | cc_reg = gen_reg_rtx (mode); | |
2041 | #endif /* ! experiment */ | |
2042 | else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
2043 | cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG); | |
2044 | else | |
2045 | cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG); | |
284d86e9 | 2046 | |
19047e4a EB |
2047 | /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this |
2048 | will only result in an unrecognizable insn so no point in asserting. */ | |
2049 | emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y))); | |
c85f7c16 | 2050 | |
e0d80184 DM |
2051 | return cc_reg; |
2052 | } | |
c85f7c16 | 2053 | |
f90b7a5a PB |
2054 | |
2055 | /* Emit the compare insn and return the CC reg for the comparison in CMP. */ | |
19047e4a EB |
2056 | |
2057 | rtx | |
f90b7a5a | 2058 | gen_compare_reg (rtx cmp) |
19047e4a | 2059 | { |
f90b7a5a | 2060 | return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1)); |
19047e4a EB |
2061 | } |
2062 | ||
e0d80184 | 2063 | /* This function is used for v9 only. |
f90b7a5a | 2064 | DEST is the target of the Scc insn. |
e0d80184 | 2065 | CODE is the code for an Scc's comparison. |
f90b7a5a | 2066 | X and Y are the values we compare. |
c85f7c16 | 2067 | |
e0d80184 | 2068 | This function is needed to turn |
ab835497 | 2069 | |
e0d80184 DM |
2070 | (set (reg:SI 110) |
2071 | (gt (reg:CCX 100 %icc) | |
2072 | (const_int 0))) | |
2073 | into | |
2074 | (set (reg:SI 110) | |
2075 | (gt:DI (reg:CCX 100 %icc) | |
2076 | (const_int 0))) | |
2077 | ||
2078 | IE: The instruction recognizer needs to see the mode of the comparison to | |
2079 | find the right instruction. We could use "gt:DI" right in the | |
f90b7a5a | 2080 | define_expand, but leaving it out allows us to handle DI, SI, etc. */ |
ab835497 | 2081 | |
f90b7a5a PB |
2082 | static int |
2083 | gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y) | |
ab835497 | 2084 | { |
e0d80184 | 2085 | if (! TARGET_ARCH64 |
f90b7a5a PB |
2086 | && (GET_MODE (x) == DImode |
2087 | || GET_MODE (dest) == DImode)) | |
e0d80184 | 2088 | return 0; |
7e2feebf | 2089 | |
e0d80184 DM |
2090 | /* Try to use the movrCC insns. */ |
2091 | if (TARGET_ARCH64 | |
f90b7a5a PB |
2092 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT |
2093 | && y == const0_rtx | |
e0d80184 | 2094 | && v9_regcmp_p (compare_code)) |
ab835497 | 2095 | { |
f90b7a5a | 2096 | rtx op0 = x; |
7e04157d DM |
2097 | rtx temp; |
2098 | ||
e0d80184 | 2099 | /* Special case for op0 != 0. This can be done with one instruction if |
f90b7a5a | 2100 | dest == x. */ |
e0d80184 DM |
2101 | |
2102 | if (compare_code == NE | |
f90b7a5a PB |
2103 | && GET_MODE (dest) == DImode |
2104 | && rtx_equal_p (op0, dest)) | |
ab835497 | 2105 | { |
f90b7a5a | 2106 | emit_insn (gen_rtx_SET (VOIDmode, dest, |
e0d80184 DM |
2107 | gen_rtx_IF_THEN_ELSE (DImode, |
2108 | gen_rtx_fmt_ee (compare_code, DImode, | |
2109 | op0, const0_rtx), | |
2110 | const1_rtx, | |
f90b7a5a | 2111 | dest))); |
ab835497 RK |
2112 | return 1; |
2113 | } | |
e0d80184 | 2114 | |
f90b7a5a | 2115 | if (reg_overlap_mentioned_p (dest, op0)) |
a42519be | 2116 | { |
f90b7a5a | 2117 | /* Handle the case where dest == x. |
a42519be | 2118 | We "early clobber" the result. */ |
f90b7a5a PB |
2119 | op0 = gen_reg_rtx (GET_MODE (x)); |
2120 | emit_move_insn (op0, x); | |
a42519be JJ |
2121 | } |
2122 | ||
f90b7a5a | 2123 | emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx)); |
e0d80184 | 2124 | if (GET_MODE (op0) != DImode) |
ab835497 | 2125 | { |
e0d80184 DM |
2126 | temp = gen_reg_rtx (DImode); |
2127 | convert_move (temp, op0, 0); | |
ab835497 | 2128 | } |
e0d80184 DM |
2129 | else |
2130 | temp = op0; | |
f90b7a5a PB |
2131 | emit_insn (gen_rtx_SET (VOIDmode, dest, |
2132 | gen_rtx_IF_THEN_ELSE (GET_MODE (dest), | |
e0d80184 DM |
2133 | gen_rtx_fmt_ee (compare_code, DImode, |
2134 | temp, const0_rtx), | |
2135 | const1_rtx, | |
f90b7a5a | 2136 | dest))); |
95726648 DE |
2137 | return 1; |
2138 | } | |
e0d80184 | 2139 | else |
ab835497 | 2140 | { |
f90b7a5a PB |
2141 | x = gen_compare_reg_1 (compare_code, x, y); |
2142 | y = const0_rtx; | |
ab835497 | 2143 | |
f90b7a5a PB |
2144 | gcc_assert (GET_MODE (x) != CC_NOOVmode |
2145 | && GET_MODE (x) != CCX_NOOVmode); | |
2146 | ||
2147 | emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx)); | |
2148 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
2149 | gen_rtx_IF_THEN_ELSE (GET_MODE (dest), | |
e0d80184 | 2150 | gen_rtx_fmt_ee (compare_code, |
f90b7a5a PB |
2151 | GET_MODE (x), x, y), |
2152 | const1_rtx, dest))); | |
e0d80184 | 2153 | return 1; |
ab835497 | 2154 | } |
ab835497 | 2155 | } |
4fb4e4b8 | 2156 | |
f90b7a5a PB |
2157 | |
2158 | /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this | |
2159 | without jumps using the addx/subx instructions. */ | |
2160 | ||
2161 | bool | |
2162 | emit_scc_insn (rtx operands[]) | |
2163 | { | |
2164 | rtx tem; | |
2165 | rtx x; | |
2166 | rtx y; | |
2167 | enum rtx_code code; | |
2168 | ||
2169 | /* The quad-word fp compare library routines all return nonzero to indicate | |
2170 | true, which is different from the equivalent libgcc routines, so we must | |
2171 | handle them specially here. */ | |
2172 | if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD) | |
2173 | { | |
2174 | operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3], | |
2175 | GET_CODE (operands[1])); | |
2176 | operands[2] = XEXP (operands[1], 0); | |
2177 | operands[3] = XEXP (operands[1], 1); | |
2178 | } | |
2179 | ||
2180 | code = GET_CODE (operands[1]); | |
2181 | x = operands[2]; | |
2182 | y = operands[3]; | |
2183 | ||
2184 | /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has | |
2185 | more applications). The exception to this is "reg != 0" which can | |
2186 | be done in one instruction on v9 (so we do it). */ | |
2187 | if (code == EQ) | |
2188 | { | |
2189 | if (GET_MODE (x) == SImode) | |
2190 | { | |
2191 | rtx pat = gen_seqsi_special (operands[0], x, y); | |
2192 | emit_insn (pat); | |
2193 | return true; | |
2194 | } | |
2195 | else if (GET_MODE (x) == DImode) | |
2196 | { | |
2197 | rtx pat = gen_seqdi_special (operands[0], x, y); | |
2198 | emit_insn (pat); | |
2199 | return true; | |
2200 | } | |
2201 | } | |
2202 | ||
2203 | if (code == NE) | |
2204 | { | |
2205 | if (GET_MODE (x) == SImode) | |
2206 | { | |
2207 | rtx pat = gen_snesi_special (operands[0], x, y); | |
2208 | emit_insn (pat); | |
2209 | return true; | |
2210 | } | |
2211 | else if (GET_MODE (x) == DImode) | |
2212 | { | |
2213 | rtx pat = gen_snedi_special (operands[0], x, y); | |
2214 | emit_insn (pat); | |
2215 | return true; | |
2216 | } | |
2217 | } | |
2218 | ||
2219 | /* For the rest, on v9 we can use conditional moves. */ | |
2220 | ||
2221 | if (TARGET_V9) | |
2222 | { | |
2223 | if (gen_v9_scc (operands[0], code, x, y)) | |
2224 | return true; | |
2225 | } | |
2226 | ||
2227 | /* We can do LTU and GEU using the addx/subx instructions too. And | |
2228 | for GTU/LEU, if both operands are registers swap them and fall | |
2229 | back to the easy case. */ | |
2230 | if (code == GTU || code == LEU) | |
2231 | { | |
2232 | if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG) | |
2233 | && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG)) | |
2234 | { | |
2235 | tem = x; | |
2236 | x = y; | |
2237 | y = tem; | |
2238 | code = swap_condition (code); | |
2239 | } | |
2240 | } | |
2241 | ||
2242 | if (code == LTU || code == GEU) | |
2243 | { | |
2244 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
4900aa61 | 2245 | gen_rtx_fmt_ee (code, SImode, |
f90b7a5a PB |
2246 | gen_compare_reg_1 (code, x, y), |
2247 | const0_rtx))); | |
2248 | return true; | |
2249 | } | |
2250 | ||
2251 | /* Nope, do branches. */ | |
2252 | return false; | |
2253 | } | |
2254 | ||
e0d80184 DM |
2255 | /* Emit a conditional jump insn for the v9 architecture using comparison code |
2256 | CODE and jump target LABEL. | |
2257 | This function exists to take advantage of the v9 brxx insns. */ | |
ab835497 | 2258 | |
f90b7a5a | 2259 | static void |
fc27d102 | 2260 | emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label) |
ab835497 | 2261 | { |
e0d80184 DM |
2262 | emit_jump_insn (gen_rtx_SET (VOIDmode, |
2263 | pc_rtx, | |
2264 | gen_rtx_IF_THEN_ELSE (VOIDmode, | |
2265 | gen_rtx_fmt_ee (code, GET_MODE (op0), | |
2266 | op0, const0_rtx), | |
2267 | gen_rtx_LABEL_REF (VOIDmode, label), | |
2268 | pc_rtx))); | |
2269 | } | |
7b1ac798 | 2270 | |
f90b7a5a PB |
2271 | void |
2272 | emit_conditional_branch_insn (rtx operands[]) | |
2273 | { | |
2274 | /* The quad-word fp compare library routines all return nonzero to indicate | |
2275 | true, which is different from the equivalent libgcc routines, so we must | |
2276 | handle them specially here. */ | |
2277 | if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD) | |
2278 | { | |
2279 | operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2], | |
2280 | GET_CODE (operands[0])); | |
2281 | operands[1] = XEXP (operands[0], 0); | |
2282 | operands[2] = XEXP (operands[0], 1); | |
2283 | } | |
2284 | ||
2285 | if (TARGET_ARCH64 && operands[2] == const0_rtx | |
2286 | && GET_CODE (operands[1]) == REG | |
2287 | && GET_MODE (operands[1]) == DImode) | |
2288 | { | |
2289 | emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]); | |
2290 | return; | |
2291 | } | |
2292 | ||
2293 | operands[1] = gen_compare_reg (operands[0]); | |
2294 | operands[2] = const0_rtx; | |
2295 | operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode, | |
2296 | operands[1], operands[2]); | |
2297 | emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2], | |
2298 | operands[3])); | |
2299 | } | |
2300 | ||
2301 | ||
7b1ac798 JJ |
2302 | /* Generate a DFmode part of a hard TFmode register. |
2303 | REG is the TFmode hard register, LOW is 1 for the | |
2304 | low 64bit of the register and 0 otherwise. | |
2305 | */ | |
2306 | rtx | |
fc27d102 | 2307 | gen_df_reg (rtx reg, int low) |
7b1ac798 JJ |
2308 | { |
2309 | int regno = REGNO (reg); | |
2310 | ||
2311 | if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0)) | |
2aac41d6 | 2312 | regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2; |
7b1ac798 JJ |
2313 | return gen_rtx_REG (DFmode, regno); |
2314 | } | |
e0d80184 | 2315 | \f |
73985940 RH |
2316 | /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value. |
2317 | Unlike normal calls, TFmode operands are passed by reference. It is | |
2318 | assumed that no more than 3 operands are required. */ | |
2319 | ||
2320 | static void | |
fc27d102 | 2321 | emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands) |
73985940 RH |
2322 | { |
2323 | rtx ret_slot = NULL, arg[3], func_sym; | |
4e7d5d27 | 2324 | int i; |
73985940 RH |
2325 | |
2326 | /* We only expect to be called for conversions, unary, and binary ops. */ | |
f5f7d171 | 2327 | gcc_assert (nargs == 2 || nargs == 3); |
73985940 RH |
2328 | |
2329 | for (i = 0; i < nargs; ++i) | |
2330 | { | |
2331 | rtx this_arg = operands[i]; | |
2332 | rtx this_slot; | |
2333 | ||
2334 | /* TFmode arguments and return values are passed by reference. */ | |
2335 | if (GET_MODE (this_arg) == TFmode) | |
2336 | { | |
4e7d5d27 DM |
2337 | int force_stack_temp; |
2338 | ||
2339 | force_stack_temp = 0; | |
2340 | if (TARGET_BUGGY_QP_LIB && i == 0) | |
2341 | force_stack_temp = 1; | |
2342 | ||
2343 | if (GET_CODE (this_arg) == MEM | |
2344 | && ! force_stack_temp) | |
2345 | this_arg = XEXP (this_arg, 0); | |
2346 | else if (CONSTANT_P (this_arg) | |
2347 | && ! force_stack_temp) | |
73985940 RH |
2348 | { |
2349 | this_slot = force_const_mem (TFmode, this_arg); | |
2350 | this_arg = XEXP (this_slot, 0); | |
2351 | } | |
2352 | else | |
2353 | { | |
2354 | this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0); | |
2355 | ||
2356 | /* Operand 0 is the return value. We'll copy it out later. */ | |
2357 | if (i > 0) | |
2358 | emit_move_insn (this_slot, this_arg); | |
2359 | else | |
2360 | ret_slot = this_slot; | |
2361 | ||
2362 | this_arg = XEXP (this_slot, 0); | |
2363 | } | |
2364 | } | |
2365 | ||
2366 | arg[i] = this_arg; | |
2367 | } | |
2368 | ||
2369 | func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name); | |
2370 | ||
2371 | if (GET_MODE (operands[0]) == TFmode) | |
2372 | { | |
2373 | if (nargs == 2) | |
2374 | emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2, | |
2375 | arg[0], GET_MODE (arg[0]), | |
2376 | arg[1], GET_MODE (arg[1])); | |
2377 | else | |
2378 | emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3, | |
2379 | arg[0], GET_MODE (arg[0]), | |
2380 | arg[1], GET_MODE (arg[1]), | |
2381 | arg[2], GET_MODE (arg[2])); | |
2382 | ||
2383 | if (ret_slot) | |
2384 | emit_move_insn (operands[0], ret_slot); | |
2385 | } | |
2386 | else | |
2387 | { | |
2388 | rtx ret; | |
2389 | ||
f5f7d171 | 2390 | gcc_assert (nargs == 2); |
73985940 RH |
2391 | |
2392 | ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL, | |
2393 | GET_MODE (operands[0]), 1, | |
2394 | arg[1], GET_MODE (arg[1])); | |
2395 | ||
2396 | if (ret != operands[0]) | |
2397 | emit_move_insn (operands[0], ret); | |
2398 | } | |
2399 | } | |
2400 | ||
2401 | /* Expand soft-float TFmode calls to sparc abi routines. */ | |
2402 | ||
2403 | static void | |
fc27d102 | 2404 | emit_soft_tfmode_binop (enum rtx_code code, rtx *operands) |
73985940 RH |
2405 | { |
2406 | const char *func; | |
2407 | ||
2408 | switch (code) | |
2409 | { | |
2410 | case PLUS: | |
2411 | func = "_Qp_add"; | |
2412 | break; | |
2413 | case MINUS: | |
2414 | func = "_Qp_sub"; | |
2415 | break; | |
2416 | case MULT: | |
2417 | func = "_Qp_mul"; | |
2418 | break; | |
2419 | case DIV: | |
2420 | func = "_Qp_div"; | |
2421 | break; | |
2422 | default: | |
f5f7d171 | 2423 | gcc_unreachable (); |
73985940 RH |
2424 | } |
2425 | ||
2426 | emit_soft_tfmode_libcall (func, 3, operands); | |
2427 | } | |
2428 | ||
2429 | static void | |
fc27d102 | 2430 | emit_soft_tfmode_unop (enum rtx_code code, rtx *operands) |
73985940 RH |
2431 | { |
2432 | const char *func; | |
2433 | ||
f5f7d171 JM |
2434 | gcc_assert (code == SQRT); |
2435 | func = "_Qp_sqrt"; | |
73985940 RH |
2436 | |
2437 | emit_soft_tfmode_libcall (func, 2, operands); | |
2438 | } | |
2439 | ||
2440 | static void | |
fc27d102 | 2441 | emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands) |
73985940 RH |
2442 | { |
2443 | const char *func; | |
2444 | ||
2445 | switch (code) | |
2446 | { | |
2447 | case FLOAT_EXTEND: | |
2448 | switch (GET_MODE (operands[1])) | |
2449 | { | |
2450 | case SFmode: | |
2451 | func = "_Qp_stoq"; | |
2452 | break; | |
2453 | case DFmode: | |
2454 | func = "_Qp_dtoq"; | |
2455 | break; | |
2456 | default: | |
f5f7d171 | 2457 | gcc_unreachable (); |
73985940 RH |
2458 | } |
2459 | break; | |
2460 | ||
2461 | case FLOAT_TRUNCATE: | |
2462 | switch (GET_MODE (operands[0])) | |
2463 | { | |
2464 | case SFmode: | |
2465 | func = "_Qp_qtos"; | |
2466 | break; | |
2467 | case DFmode: | |
2468 | func = "_Qp_qtod"; | |
2469 | break; | |
2470 | default: | |
f5f7d171 | 2471 | gcc_unreachable (); |
73985940 RH |
2472 | } |
2473 | break; | |
2474 | ||
2475 | case FLOAT: | |
2476 | switch (GET_MODE (operands[1])) | |
2477 | { | |
2478 | case SImode: | |
2479 | func = "_Qp_itoq"; | |
92693034 JM |
2480 | if (TARGET_ARCH64) |
2481 | operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]); | |
73985940 RH |
2482 | break; |
2483 | case DImode: | |
2484 | func = "_Qp_xtoq"; | |
2485 | break; | |
2486 | default: | |
f5f7d171 | 2487 | gcc_unreachable (); |
73985940 RH |
2488 | } |
2489 | break; | |
2490 | ||
2491 | case UNSIGNED_FLOAT: | |
2492 | switch (GET_MODE (operands[1])) | |
2493 | { | |
2494 | case SImode: | |
2495 | func = "_Qp_uitoq"; | |
92693034 JM |
2496 | if (TARGET_ARCH64) |
2497 | operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]); | |
73985940 RH |
2498 | break; |
2499 | case DImode: | |
2500 | func = "_Qp_uxtoq"; | |
2501 | break; | |
2502 | default: | |
f5f7d171 | 2503 | gcc_unreachable (); |
73985940 RH |
2504 | } |
2505 | break; | |
2506 | ||
2507 | case FIX: | |
2508 | switch (GET_MODE (operands[0])) | |
2509 | { | |
2510 | case SImode: | |
2511 | func = "_Qp_qtoi"; | |
2512 | break; | |
2513 | case DImode: | |
2514 | func = "_Qp_qtox"; | |
2515 | break; | |
2516 | default: | |
f5f7d171 | 2517 | gcc_unreachable (); |
73985940 RH |
2518 | } |
2519 | break; | |
2520 | ||
2521 | case UNSIGNED_FIX: | |
2522 | switch (GET_MODE (operands[0])) | |
2523 | { | |
2524 | case SImode: | |
2525 | func = "_Qp_qtoui"; | |
2526 | break; | |
2527 | case DImode: | |
2528 | func = "_Qp_qtoux"; | |
2529 | break; | |
2530 | default: | |
f5f7d171 | 2531 | gcc_unreachable (); |
73985940 RH |
2532 | } |
2533 | break; | |
2534 | ||
2535 | default: | |
f5f7d171 | 2536 | gcc_unreachable (); |
73985940 RH |
2537 | } |
2538 | ||
2539 | emit_soft_tfmode_libcall (func, 2, operands); | |
2540 | } | |
2541 | ||
2542 | /* Expand a hard-float tfmode operation. All arguments must be in | |
2543 | registers. */ | |
2544 | ||
2545 | static void | |
fc27d102 | 2546 | emit_hard_tfmode_operation (enum rtx_code code, rtx *operands) |
73985940 RH |
2547 | { |
2548 | rtx op, dest; | |
2549 | ||
ec8e098d | 2550 | if (GET_RTX_CLASS (code) == RTX_UNARY) |
73985940 RH |
2551 | { |
2552 | operands[1] = force_reg (GET_MODE (operands[1]), operands[1]); | |
2553 | op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]); | |
2554 | } | |
2555 | else | |
2556 | { | |
2557 | operands[1] = force_reg (GET_MODE (operands[1]), operands[1]); | |
2558 | operands[2] = force_reg (GET_MODE (operands[2]), operands[2]); | |
2559 | op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]), | |
2560 | operands[1], operands[2]); | |
2561 | } | |
2562 | ||
2563 | if (register_operand (operands[0], VOIDmode)) | |
2564 | dest = operands[0]; | |
2565 | else | |
2566 | dest = gen_reg_rtx (GET_MODE (operands[0])); | |
2567 | ||
2568 | emit_insn (gen_rtx_SET (VOIDmode, dest, op)); | |
2569 | ||
2570 | if (dest != operands[0]) | |
2571 | emit_move_insn (operands[0], dest); | |
2572 | } | |
2573 | ||
2574 | void | |
fc27d102 | 2575 | emit_tfmode_binop (enum rtx_code code, rtx *operands) |
73985940 RH |
2576 | { |
2577 | if (TARGET_HARD_QUAD) | |
2578 | emit_hard_tfmode_operation (code, operands); | |
2579 | else | |
2580 | emit_soft_tfmode_binop (code, operands); | |
2581 | } | |
2582 | ||
2583 | void | |
fc27d102 | 2584 | emit_tfmode_unop (enum rtx_code code, rtx *operands) |
73985940 RH |
2585 | { |
2586 | if (TARGET_HARD_QUAD) | |
2587 | emit_hard_tfmode_operation (code, operands); | |
2588 | else | |
2589 | emit_soft_tfmode_unop (code, operands); | |
2590 | } | |
2591 | ||
2592 | void | |
fc27d102 | 2593 | emit_tfmode_cvt (enum rtx_code code, rtx *operands) |
73985940 RH |
2594 | { |
2595 | if (TARGET_HARD_QUAD) | |
2596 | emit_hard_tfmode_operation (code, operands); | |
2597 | else | |
2598 | emit_soft_tfmode_cvt (code, operands); | |
2599 | } | |
2600 | \f | |
f17f9332 JJ |
2601 | /* Return nonzero if a branch/jump/call instruction will be emitting |
2602 | nop into its delay slot. */ | |
2603 | ||
2604 | int | |
fc27d102 | 2605 | empty_delay_slot (rtx insn) |
f17f9332 JJ |
2606 | { |
2607 | rtx seq; | |
2608 | ||
2609 | /* If no previous instruction (should not happen), return true. */ | |
2610 | if (PREV_INSN (insn) == NULL) | |
2611 | return 1; | |
2612 | ||
2613 | seq = NEXT_INSN (PREV_INSN (insn)); | |
2614 | if (GET_CODE (PATTERN (seq)) == SEQUENCE) | |
2615 | return 0; | |
2616 | ||
2617 | return 1; | |
2618 | } | |
2619 | ||
9ac617d4 | 2620 | /* Return nonzero if TRIAL can go into the call delay slot. */ |
ab835497 | 2621 | |
e0d80184 | 2622 | int |
9ac617d4 | 2623 | tls_call_delay (rtx trial) |
e0d80184 | 2624 | { |
e10fbf96 | 2625 | rtx pat; |
1150a841 | 2626 | |
9ac617d4 | 2627 | /* Binutils allows |
e10fbf96 EB |
2628 | call __tls_get_addr, %tgd_call (foo) |
2629 | add %l7, %o0, %o0, %tgd_add (foo) | |
9ac617d4 EB |
2630 | while Sun as/ld does not. */ |
2631 | if (TARGET_GNU_TLS || !TARGET_TLS) | |
2632 | return 1; | |
4fb4e4b8 | 2633 | |
e0d80184 | 2634 | pat = PATTERN (trial); |
795068a4 | 2635 | |
e10fbf96 EB |
2636 | /* We must reject tgd_add{32|64}, i.e. |
2637 | (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD))) | |
2638 | and tldm_add{32|64}, i.e. | |
2639 | (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM))) | |
2640 | for Sun as/ld. */ | |
2641 | if (GET_CODE (pat) == SET | |
2642 | && GET_CODE (SET_SRC (pat)) == PLUS) | |
2643 | { | |
2644 | rtx unspec = XEXP (SET_SRC (pat), 1); | |
795068a4 | 2645 | |
e10fbf96 EB |
2646 | if (GET_CODE (unspec) == UNSPEC |
2647 | && (XINT (unspec, 1) == UNSPEC_TLSGD | |
2648 | || XINT (unspec, 1) == UNSPEC_TLSLDM)) | |
2649 | return 0; | |
2650 | } | |
2651 | ||
2652 | return 1; | |
9ac617d4 EB |
2653 | } |
2654 | ||
2655 | /* Return nonzero if TRIAL, an insn, can be combined with a 'restore' | |
2656 | instruction. RETURN_P is true if the v9 variant 'return' is to be | |
2657 | considered in the test too. | |
c137830f | 2658 | |
9ac617d4 EB |
2659 | TRIAL must be a SET whose destination is a REG appropriate for the |
2660 | 'restore' instruction or, if RETURN_P is true, for the 'return' | |
2661 | instruction. */ | |
795068a4 | 2662 | |
9ac617d4 EB |
2663 | static int |
2664 | eligible_for_restore_insn (rtx trial, bool return_p) | |
2665 | { | |
2666 | rtx pat = PATTERN (trial); | |
2667 | rtx src = SET_SRC (pat); | |
795068a4 | 2668 | |
9ac617d4 | 2669 | /* The 'restore src,%g0,dest' pattern for word mode and below. */ |
3ec6f1a9 RK |
2670 | if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT |
2671 | && arith_operand (src, GET_MODE (src))) | |
3c6088d2 JJ |
2672 | { |
2673 | if (TARGET_ARCH64) | |
2674 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode); | |
2675 | else | |
2676 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode); | |
2677 | } | |
e48addee | 2678 | |
9ac617d4 | 2679 | /* The 'restore src,%g0,dest' pattern for double-word mode. */ |
3ec6f1a9 RK |
2680 | else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT |
2681 | && arith_double_operand (src, GET_MODE (src))) | |
e0d80184 | 2682 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode); |
795068a4 | 2683 | |
9ac617d4 EB |
2684 | /* The 'restore src,%g0,dest' pattern for float if no FPU. */ |
2685 | else if (! TARGET_FPU && register_operand (src, SFmode)) | |
2686 | return 1; | |
2687 | ||
2688 | /* The 'restore src,%g0,dest' pattern for double if no FPU. */ | |
2689 | else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode)) | |
e0d80184 | 2690 | return 1; |
f5da225e | 2691 | |
9ac617d4 | 2692 | /* If we have the 'return' instruction, anything that does not use |
e48addee | 2693 | local or output registers and can go into a delay slot wins. */ |
9ac617d4 EB |
2694 | else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1) |
2695 | && (get_attr_in_uncond_branch_delay (trial) | |
2696 | == IN_UNCOND_BRANCH_DELAY_TRUE)) | |
e48addee JJ |
2697 | return 1; |
2698 | ||
9ac617d4 | 2699 | /* The 'restore src1,src2,dest' pattern for SImode. */ |
e0d80184 | 2700 | else if (GET_CODE (src) == PLUS |
9ac617d4 EB |
2701 | && register_operand (XEXP (src, 0), SImode) |
2702 | && arith_operand (XEXP (src, 1), SImode)) | |
e0d80184 DM |
2703 | return 1; |
2704 | ||
9ac617d4 | 2705 | /* The 'restore src1,src2,dest' pattern for DImode. */ |
e0d80184 | 2706 | else if (GET_CODE (src) == PLUS |
9ac617d4 EB |
2707 | && register_operand (XEXP (src, 0), DImode) |
2708 | && arith_double_operand (XEXP (src, 1), DImode)) | |
795068a4 | 2709 | return 1; |
795068a4 | 2710 | |
9ac617d4 | 2711 | /* The 'restore src1,%lo(src2),dest' pattern. */ |
e48addee JJ |
2712 | else if (GET_CODE (src) == LO_SUM |
2713 | && ! TARGET_CM_MEDMID | |
2714 | && ((register_operand (XEXP (src, 0), SImode) | |
2715 | && immediate_operand (XEXP (src, 1), SImode)) | |
2716 | || (TARGET_ARCH64 | |
2717 | && register_operand (XEXP (src, 0), DImode) | |
2718 | && immediate_operand (XEXP (src, 1), DImode)))) | |
2719 | return 1; | |
2720 | ||
9ac617d4 | 2721 | /* The 'restore src,src,dest' pattern. */ |
e48addee JJ |
2722 | else if (GET_CODE (src) == ASHIFT |
2723 | && (register_operand (XEXP (src, 0), SImode) | |
2724 | || register_operand (XEXP (src, 0), DImode)) | |
2725 | && XEXP (src, 1) == const1_rtx) | |
2726 | return 1; | |
2727 | ||
795068a4 JW |
2728 | return 0; |
2729 | } | |
2730 | ||
9ac617d4 EB |
2731 | /* Return nonzero if TRIAL can go into the function return's |
2732 | delay slot. */ | |
2733 | ||
5751a10b | 2734 | int |
9ac617d4 | 2735 | eligible_for_return_delay (rtx trial) |
5751a10b | 2736 | { |
9ac617d4 | 2737 | rtx pat; |
5751a10b | 2738 | |
9ac617d4 EB |
2739 | if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET) |
2740 | return 0; | |
2741 | ||
2742 | if (get_attr_length (trial) != 1) | |
2743 | return 0; | |
2744 | ||
2745 | /* If there are any call-saved registers, we should scan TRIAL if it | |
2746 | does not reference them. For now just make it easy. */ | |
2747 | if (num_gfregs) | |
2748 | return 0; | |
2749 | ||
2750 | /* If the function uses __builtin_eh_return, the eh_return machinery | |
2751 | occupies the delay slot. */ | |
e3b5732b | 2752 | if (crtl->calls_eh_return) |
9ac617d4 EB |
2753 | return 0; |
2754 | ||
2755 | /* In the case of a true leaf function, anything can go into the slot. */ | |
5be9b7a1 | 2756 | if (sparc_leaf_function_p) |
9ac617d4 EB |
2757 | return get_attr_in_uncond_branch_delay (trial) |
2758 | == IN_UNCOND_BRANCH_DELAY_TRUE; | |
5751a10b JJ |
2759 | |
2760 | pat = PATTERN (trial); | |
5751a10b | 2761 | |
9ac617d4 EB |
2762 | /* Otherwise, only operations which can be done in tandem with |
2763 | a `restore' or `return' insn can go into the delay slot. */ | |
2764 | if (GET_CODE (SET_DEST (pat)) != REG | |
2765 | || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)) | |
2766 | return 0; | |
5751a10b | 2767 | |
9ac617d4 EB |
2768 | /* If this instruction sets up floating point register and we have a return |
2769 | instruction, it can probably go in. But restore will not work | |
2770 | with FP_REGS. */ | |
2771 | if (REGNO (SET_DEST (pat)) >= 32) | |
2772 | return (TARGET_V9 | |
2773 | && ! epilogue_renumber (&pat, 1) | |
2774 | && (get_attr_in_uncond_branch_delay (trial) | |
2775 | == IN_UNCOND_BRANCH_DELAY_TRUE)); | |
2776 | ||
2777 | return eligible_for_restore_insn (trial, true); | |
5751a10b JJ |
2778 | } |
2779 | ||
9ac617d4 | 2780 | /* Return nonzero if TRIAL can go into the sibling call's |
7d167afd JJ |
2781 | delay slot. */ |
2782 | ||
2783 | int | |
fc27d102 | 2784 | eligible_for_sibcall_delay (rtx trial) |
7d167afd | 2785 | { |
9ac617d4 | 2786 | rtx pat; |
7d167afd JJ |
2787 | |
2788 | if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET) | |
2789 | return 0; | |
2790 | ||
8456b95a | 2791 | if (get_attr_length (trial) != 1) |
7d167afd JJ |
2792 | return 0; |
2793 | ||
2794 | pat = PATTERN (trial); | |
2795 | ||
5be9b7a1 | 2796 | if (sparc_leaf_function_p) |
7d167afd JJ |
2797 | { |
2798 | /* If the tail call is done using the call instruction, | |
2799 | we have to restore %o7 in the delay slot. */ | |
6244b14e | 2800 | if (LEAF_SIBCALL_SLOT_RESERVED_P) |
7d167afd JJ |
2801 | return 0; |
2802 | ||
2803 | /* %g1 is used to build the function address */ | |
2804 | if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat)) | |
2805 | return 0; | |
2806 | ||
2807 | return 1; | |
2808 | } | |
2809 | ||
2810 | /* Otherwise, only operations which can be done in tandem with | |
2811 | a `restore' insn can go into the delay slot. */ | |
2812 | if (GET_CODE (SET_DEST (pat)) != REG | |
9ac617d4 | 2813 | || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24) |
7d167afd JJ |
2814 | || REGNO (SET_DEST (pat)) >= 32) |
2815 | return 0; | |
2816 | ||
2817 | /* If it mentions %o7, it can't go in, because sibcall will clobber it | |
2818 | in most cases. */ | |
2819 | if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat)) | |
2820 | return 0; | |
2821 | ||
9ac617d4 | 2822 | return eligible_for_restore_insn (trial, false); |
e0d80184 | 2823 | } |
ab835497 | 2824 | |
e0d80184 | 2825 | int |
fc27d102 | 2826 | short_branch (int uid1, int uid2) |
e0d80184 | 2827 | { |
9d98a694 | 2828 | int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2); |
77b91512 RK |
2829 | |
2830 | /* Leave a few words of "slop". */ | |
2831 | if (delta >= -1023 && delta <= 1022) | |
e0d80184 | 2832 | return 1; |
77b91512 | 2833 | |
e0d80184 DM |
2834 | return 0; |
2835 | } | |
ab835497 | 2836 | |
5e7a8ee0 | 2837 | /* Return nonzero if REG is not used after INSN. |
e0d80184 DM |
2838 | We assume REG is a reload reg, and therefore does |
2839 | not live past labels or calls or jumps. */ | |
2840 | int | |
fc27d102 | 2841 | reg_unused_after (rtx reg, rtx insn) |
e0d80184 DM |
2842 | { |
2843 | enum rtx_code code, prev_code = UNKNOWN; | |
ab835497 | 2844 | |
e0d80184 | 2845 | while ((insn = NEXT_INSN (insn))) |
61a55e8b | 2846 | { |
e0d80184 DM |
2847 | if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)]) |
2848 | return 1; | |
ab835497 | 2849 | |
e0d80184 DM |
2850 | code = GET_CODE (insn); |
2851 | if (GET_CODE (insn) == CODE_LABEL) | |
2852 | return 1; | |
bc961ed7 | 2853 | |
ec8e098d | 2854 | if (INSN_P (insn)) |
61a55e8b | 2855 | { |
e0d80184 DM |
2856 | rtx set = single_set (insn); |
2857 | int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set)); | |
2858 | if (set && in_src) | |
2859 | return 0; | |
2860 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
2861 | return 1; | |
2862 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
2863 | return 0; | |
61a55e8b | 2864 | } |
e0d80184 | 2865 | prev_code = code; |
61a55e8b | 2866 | } |
e0d80184 DM |
2867 | return 1; |
2868 | } | |
2869 | \f | |
5751a10b JJ |
2870 | /* Determine if it's legal to put X into the constant pool. This |
2871 | is not possible if X contains the address of a symbol that is | |
2872 | not constant (TLS) or not known at final link time (PIC). */ | |
2873 | ||
2874 | static bool | |
2875 | sparc_cannot_force_const_mem (rtx x) | |
2876 | { | |
2877 | switch (GET_CODE (x)) | |
2878 | { | |
2879 | case CONST_INT: | |
2880 | case CONST_DOUBLE: | |
c75d6010 | 2881 | case CONST_VECTOR: |
5751a10b JJ |
2882 | /* Accept all non-symbolic constants. */ |
2883 | return false; | |
2884 | ||
2885 | case LABEL_REF: | |
2886 | /* Labels are OK iff we are non-PIC. */ | |
2887 | return flag_pic != 0; | |
2888 | ||
2889 | case SYMBOL_REF: | |
2890 | /* 'Naked' TLS symbol references are never OK, | |
2891 | non-TLS symbols are OK iff we are non-PIC. */ | |
2892 | if (SYMBOL_REF_TLS_MODEL (x)) | |
2893 | return true; | |
2894 | else | |
2895 | return flag_pic != 0; | |
2896 | ||
2897 | case CONST: | |
2898 | return sparc_cannot_force_const_mem (XEXP (x, 0)); | |
2899 | case PLUS: | |
2900 | case MINUS: | |
2901 | return sparc_cannot_force_const_mem (XEXP (x, 0)) | |
2902 | || sparc_cannot_force_const_mem (XEXP (x, 1)); | |
2903 | case UNSPEC: | |
2904 | return true; | |
2905 | default: | |
f5f7d171 | 2906 | gcc_unreachable (); |
5751a10b JJ |
2907 | } |
2908 | } | |
2909 | \f | |
5e77eb53 | 2910 | /* PIC support. */ |
9fd52c6c | 2911 | static GTY(()) bool pic_helper_needed = false; |
5e77eb53 | 2912 | static GTY(()) rtx pic_helper_symbol; |
e2500fed | 2913 | static GTY(()) rtx global_offset_table; |
bc961ed7 | 2914 | |
e0d80184 | 2915 | /* Ensure that we are not using patterns that are not OK with PIC. */ |
795068a4 | 2916 | |
e0d80184 | 2917 | int |
fc27d102 | 2918 | check_pic (int i) |
e0d80184 DM |
2919 | { |
2920 | switch (flag_pic) | |
ab835497 | 2921 | { |
e0d80184 | 2922 | case 1: |
f5f7d171 JM |
2923 | gcc_assert (GET_CODE (recog_data.operand[i]) != SYMBOL_REF |
2924 | && (GET_CODE (recog_data.operand[i]) != CONST | |
2925 | || (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS | |
2926 | && (XEXP (XEXP (recog_data.operand[i], 0), 0) | |
2927 | == global_offset_table) | |
2928 | && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1)) | |
2929 | == CONST)))); | |
e0d80184 DM |
2930 | case 2: |
2931 | default: | |
2932 | return 1; | |
ab835497 | 2933 | } |
ab835497 | 2934 | } |
795068a4 | 2935 | |
4900aa61 | 2936 | /* Return true if X is an address which needs a temporary register when |
e0d80184 | 2937 | reloaded while generating PIC code. */ |
795068a4 | 2938 | |
e0d80184 | 2939 | int |
fc27d102 | 2940 | pic_address_needs_scratch (rtx x) |
795068a4 | 2941 | { |
e0d80184 DM |
2942 | /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */ |
2943 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS | |
2944 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF | |
2945 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2946 | && ! SMALL_INT (XEXP (XEXP (x, 0), 1))) | |
2947 | return 1; | |
795068a4 | 2948 | |
e0d80184 DM |
2949 | return 0; |
2950 | } | |
795068a4 | 2951 | |
5751a10b JJ |
2952 | /* Determine if a given RTX is a valid constant. We already know this |
2953 | satisfies CONSTANT_P. */ | |
2954 | ||
2955 | bool | |
2956 | legitimate_constant_p (rtx x) | |
2957 | { | |
5751a10b JJ |
2958 | switch (GET_CODE (x)) |
2959 | { | |
5751a10b | 2960 | case CONST: |
208384fd EB |
2961 | case SYMBOL_REF: |
2962 | if (sparc_tls_referenced_p (x)) | |
5751a10b JJ |
2963 | return false; |
2964 | break; | |
2965 | ||
2966 | case CONST_DOUBLE: | |
2967 | if (GET_MODE (x) == VOIDmode) | |
2968 | return true; | |
2969 | ||
2970 | /* Floating point constants are generally not ok. | |
2971 | The only exception is 0.0 in VIS. */ | |
2972 | if (TARGET_VIS | |
bea5071f EB |
2973 | && SCALAR_FLOAT_MODE_P (GET_MODE (x)) |
2974 | && const_zero_operand (x, GET_MODE (x))) | |
2975 | return true; | |
2976 | ||
2977 | return false; | |
2978 | ||
2979 | case CONST_VECTOR: | |
2980 | /* Vector constants are generally not ok. | |
2981 | The only exception is 0 in VIS. */ | |
2982 | if (TARGET_VIS | |
0e5d569c | 2983 | && const_zero_operand (x, GET_MODE (x))) |
5751a10b JJ |
2984 | return true; |
2985 | ||
2986 | return false; | |
2987 | ||
2988 | default: | |
2989 | break; | |
2990 | } | |
2991 | ||
2992 | return true; | |
2993 | } | |
2994 | ||
2995 | /* Determine if a given RTX is a valid constant address. */ | |
2996 | ||
2997 | bool | |
2998 | constant_address_p (rtx x) | |
2999 | { | |
3000 | switch (GET_CODE (x)) | |
3001 | { | |
3002 | case LABEL_REF: | |
3003 | case CONST_INT: | |
3004 | case HIGH: | |
3005 | return true; | |
3006 | ||
3007 | case CONST: | |
3008 | if (flag_pic && pic_address_needs_scratch (x)) | |
3009 | return false; | |
3010 | return legitimate_constant_p (x); | |
3011 | ||
3012 | case SYMBOL_REF: | |
3013 | return !flag_pic && legitimate_constant_p (x); | |
3014 | ||
3015 | default: | |
3016 | return false; | |
3017 | } | |
3018 | } | |
3019 | ||
3020 | /* Nonzero if the constant value X is a legitimate general operand | |
3021 | when generating PIC code. It is given that flag_pic is on and | |
3022 | that X satisfies CONSTANT_P or is a CONST_DOUBLE. */ | |
3023 | ||
3024 | bool | |
3025 | legitimate_pic_operand_p (rtx x) | |
3026 | { | |
3027 | if (pic_address_needs_scratch (x)) | |
3028 | return false; | |
208384fd | 3029 | if (sparc_tls_referenced_p (x)) |
5751a10b JJ |
3030 | return false; |
3031 | return true; | |
3032 | } | |
3033 | ||
3034 | /* Return nonzero if ADDR is a valid memory address. | |
3035 | STRICT specifies whether strict register checking applies. */ | |
4900aa61 | 3036 | |
c6c3dba9 PB |
3037 | static bool |
3038 | sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict) | |
5751a10b | 3039 | { |
f1b259d5 | 3040 | rtx rs1 = NULL, rs2 = NULL, imm1 = NULL; |
5751a10b JJ |
3041 | |
3042 | if (REG_P (addr) || GET_CODE (addr) == SUBREG) | |
3043 | rs1 = addr; | |
3044 | else if (GET_CODE (addr) == PLUS) | |
3045 | { | |
3046 | rs1 = XEXP (addr, 0); | |
3047 | rs2 = XEXP (addr, 1); | |
3048 | ||
3049 | /* Canonicalize. REG comes first, if there are no regs, | |
3050 | LO_SUM comes first. */ | |
3051 | if (!REG_P (rs1) | |
3052 | && GET_CODE (rs1) != SUBREG | |
3053 | && (REG_P (rs2) | |
3054 | || GET_CODE (rs2) == SUBREG | |
3055 | || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM))) | |
3056 | { | |
3057 | rs1 = XEXP (addr, 1); | |
3058 | rs2 = XEXP (addr, 0); | |
3059 | } | |
3060 | ||
3061 | if ((flag_pic == 1 | |
3062 | && rs1 == pic_offset_table_rtx | |
3063 | && !REG_P (rs2) | |
3064 | && GET_CODE (rs2) != SUBREG | |
3065 | && GET_CODE (rs2) != LO_SUM | |
3066 | && GET_CODE (rs2) != MEM | |
208384fd | 3067 | && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2)) |
5751a10b JJ |
3068 | && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode) |
3069 | && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2))) | |
3070 | || ((REG_P (rs1) | |
3071 | || GET_CODE (rs1) == SUBREG) | |
3072 | && RTX_OK_FOR_OFFSET_P (rs2))) | |
3073 | { | |
3074 | imm1 = rs2; | |
3075 | rs2 = NULL; | |
3076 | } | |
3077 | else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG) | |
3078 | && (REG_P (rs2) || GET_CODE (rs2) == SUBREG)) | |
3079 | { | |
39ba2f92 EB |
3080 | /* We prohibit REG + REG for TFmode when there are no quad move insns |
3081 | and we consequently need to split. We do this because REG+REG | |
3082 | is not an offsettable address. If we get the situation in reload | |
5751a10b JJ |
3083 | where source and destination of a movtf pattern are both MEMs with |
3084 | REG+REG address, then only one of them gets converted to an | |
39ba2f92 | 3085 | offsettable address. */ |
5751a10b | 3086 | if (mode == TFmode |
39ba2f92 | 3087 | && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD)) |
5751a10b JJ |
3088 | return 0; |
3089 | ||
3090 | /* We prohibit REG + REG on ARCH32 if not optimizing for | |
3091 | DFmode/DImode because then mem_min_alignment is likely to be zero | |
3092 | after reload and the forced split would lack a matching splitter | |
3093 | pattern. */ | |
3094 | if (TARGET_ARCH32 && !optimize | |
3095 | && (mode == DFmode || mode == DImode)) | |
3096 | return 0; | |
3097 | } | |
3098 | else if (USE_AS_OFFSETABLE_LO10 | |
3099 | && GET_CODE (rs1) == LO_SUM | |
3100 | && TARGET_ARCH64 | |
3101 | && ! TARGET_CM_MEDMID | |
3102 | && RTX_OK_FOR_OLO10_P (rs2)) | |
3103 | { | |
5751a10b JJ |
3104 | rs2 = NULL; |
3105 | imm1 = XEXP (rs1, 1); | |
3106 | rs1 = XEXP (rs1, 0); | |
208384fd EB |
3107 | if (!CONSTANT_P (imm1) |
3108 | || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1))) | |
5751a10b JJ |
3109 | return 0; |
3110 | } | |
3111 | } | |
3112 | else if (GET_CODE (addr) == LO_SUM) | |
3113 | { | |
3114 | rs1 = XEXP (addr, 0); | |
3115 | imm1 = XEXP (addr, 1); | |
3116 | ||
208384fd EB |
3117 | if (!CONSTANT_P (imm1) |
3118 | || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1))) | |
5751a10b JJ |
3119 | return 0; |
3120 | ||
f1b259d5 EB |
3121 | /* We can't allow TFmode in 32-bit mode, because an offset greater |
3122 | than the alignment (8) may cause the LO_SUM to overflow. */ | |
d24088cc | 3123 | if (mode == TFmode && TARGET_ARCH32) |
f1b259d5 | 3124 | return 0; |
5751a10b JJ |
3125 | } |
3126 | else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr)) | |
3127 | return 1; | |
3128 | else | |
3129 | return 0; | |
3130 | ||
3131 | if (GET_CODE (rs1) == SUBREG) | |
3132 | rs1 = SUBREG_REG (rs1); | |
3133 | if (!REG_P (rs1)) | |
3134 | return 0; | |
3135 | ||
3136 | if (rs2) | |
3137 | { | |
3138 | if (GET_CODE (rs2) == SUBREG) | |
3139 | rs2 = SUBREG_REG (rs2); | |
3140 | if (!REG_P (rs2)) | |
3141 | return 0; | |
3142 | } | |
3143 | ||
3144 | if (strict) | |
3145 | { | |
3146 | if (!REGNO_OK_FOR_BASE_P (REGNO (rs1)) | |
3147 | || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2)))) | |
3148 | return 0; | |
3149 | } | |
3150 | else | |
3151 | { | |
3152 | if ((REGNO (rs1) >= 32 | |
3153 | && REGNO (rs1) != FRAME_POINTER_REGNUM | |
3154 | && REGNO (rs1) < FIRST_PSEUDO_REGISTER) | |
3155 | || (rs2 | |
3156 | && (REGNO (rs2) >= 32 | |
3157 | && REGNO (rs2) != FRAME_POINTER_REGNUM | |
3158 | && REGNO (rs2) < FIRST_PSEUDO_REGISTER))) | |
3159 | return 0; | |
3160 | } | |
3161 | return 1; | |
3162 | } | |
3163 | ||
3164 | /* Construct the SYMBOL_REF for the tls_get_offset function. */ | |
3165 | ||
3166 | static GTY(()) rtx sparc_tls_symbol; | |
38fc66ba | 3167 | |
5751a10b JJ |
3168 | static rtx |
3169 | sparc_tls_get_addr (void) | |
3170 | { | |
3171 | if (!sparc_tls_symbol) | |
3172 | sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr"); | |
3173 | ||
3174 | return sparc_tls_symbol; | |
3175 | } | |
3176 | ||
3177 | static rtx | |
3178 | sparc_tls_got (void) | |
3179 | { | |
3180 | rtx temp; | |
3181 | if (flag_pic) | |
3182 | { | |
e3b5732b | 3183 | crtl->uses_pic_offset_table = 1; |
5751a10b JJ |
3184 | return pic_offset_table_rtx; |
3185 | } | |
3186 | ||
3187 | if (!global_offset_table) | |
3188 | global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
3189 | temp = gen_reg_rtx (Pmode); | |
3190 | emit_move_insn (temp, global_offset_table); | |
3191 | return temp; | |
3192 | } | |
3193 | ||
208384fd | 3194 | /* Return true if X contains a thread-local symbol. */ |
38fc66ba | 3195 | |
208384fd | 3196 | static bool |
38fc66ba EB |
3197 | sparc_tls_referenced_p (rtx x) |
3198 | { | |
3199 | if (!TARGET_HAVE_TLS) | |
3200 | return false; | |
3201 | ||
208384fd EB |
3202 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS) |
3203 | x = XEXP (XEXP (x, 0), 0); | |
3204 | ||
3205 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x)) | |
3206 | return true; | |
3207 | ||
3208 | /* That's all we handle in legitimize_tls_address for now. */ | |
3209 | return false; | |
38fc66ba | 3210 | } |
5751a10b JJ |
3211 | |
3212 | /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute | |
3213 | this (thread-local) address. */ | |
3214 | ||
208384fd | 3215 | static rtx |
5751a10b JJ |
3216 | legitimize_tls_address (rtx addr) |
3217 | { | |
3218 | rtx temp1, temp2, temp3, ret, o0, got, insn; | |
3219 | ||
b3a13419 | 3220 | gcc_assert (can_create_pseudo_p ()); |
5751a10b JJ |
3221 | |
3222 | if (GET_CODE (addr) == SYMBOL_REF) | |
3223 | switch (SYMBOL_REF_TLS_MODEL (addr)) | |
3224 | { | |
3225 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
3226 | start_sequence (); | |
3227 | temp1 = gen_reg_rtx (SImode); | |
3228 | temp2 = gen_reg_rtx (SImode); | |
3229 | ret = gen_reg_rtx (Pmode); | |
3230 | o0 = gen_rtx_REG (Pmode, 8); | |
3231 | got = sparc_tls_got (); | |
3232 | emit_insn (gen_tgd_hi22 (temp1, addr)); | |
3233 | emit_insn (gen_tgd_lo10 (temp2, temp1, addr)); | |
3234 | if (TARGET_ARCH32) | |
3235 | { | |
3236 | emit_insn (gen_tgd_add32 (o0, got, temp2, addr)); | |
3237 | insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (), | |
3238 | addr, const1_rtx)); | |
3239 | } | |
3240 | else | |
3241 | { | |
3242 | emit_insn (gen_tgd_add64 (o0, got, temp2, addr)); | |
3243 | insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (), | |
3244 | addr, const1_rtx)); | |
3245 | } | |
3246 | CALL_INSN_FUNCTION_USAGE (insn) | |
3247 | = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0), | |
3248 | CALL_INSN_FUNCTION_USAGE (insn)); | |
3249 | insn = get_insns (); | |
3250 | end_sequence (); | |
3251 | emit_libcall_block (insn, ret, o0, addr); | |
3252 | break; | |
3253 | ||
3254 | case TLS_MODEL_LOCAL_DYNAMIC: | |
3255 | start_sequence (); | |
3256 | temp1 = gen_reg_rtx (SImode); | |
3257 | temp2 = gen_reg_rtx (SImode); | |
3258 | temp3 = gen_reg_rtx (Pmode); | |
3259 | ret = gen_reg_rtx (Pmode); | |
3260 | o0 = gen_rtx_REG (Pmode, 8); | |
3261 | got = sparc_tls_got (); | |
3262 | emit_insn (gen_tldm_hi22 (temp1)); | |
3263 | emit_insn (gen_tldm_lo10 (temp2, temp1)); | |
3264 | if (TARGET_ARCH32) | |
3265 | { | |
3266 | emit_insn (gen_tldm_add32 (o0, got, temp2)); | |
3267 | insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (), | |
3268 | const1_rtx)); | |
3269 | } | |
3270 | else | |
3271 | { | |
3272 | emit_insn (gen_tldm_add64 (o0, got, temp2)); | |
3273 | insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (), | |
3274 | const1_rtx)); | |
3275 | } | |
3276 | CALL_INSN_FUNCTION_USAGE (insn) | |
3277 | = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0), | |
3278 | CALL_INSN_FUNCTION_USAGE (insn)); | |
3279 | insn = get_insns (); | |
3280 | end_sequence (); | |
3281 | emit_libcall_block (insn, temp3, o0, | |
3282 | gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), | |
3283 | UNSPEC_TLSLD_BASE)); | |
3284 | temp1 = gen_reg_rtx (SImode); | |
3285 | temp2 = gen_reg_rtx (SImode); | |
3286 | emit_insn (gen_tldo_hix22 (temp1, addr)); | |
3287 | emit_insn (gen_tldo_lox10 (temp2, temp1, addr)); | |
3288 | if (TARGET_ARCH32) | |
3289 | emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr)); | |
3290 | else | |
3291 | emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr)); | |
3292 | break; | |
3293 | ||
3294 | case TLS_MODEL_INITIAL_EXEC: | |
3295 | temp1 = gen_reg_rtx (SImode); | |
3296 | temp2 = gen_reg_rtx (SImode); | |
3297 | temp3 = gen_reg_rtx (Pmode); | |
3298 | got = sparc_tls_got (); | |
3299 | emit_insn (gen_tie_hi22 (temp1, addr)); | |
3300 | emit_insn (gen_tie_lo10 (temp2, temp1, addr)); | |
3301 | if (TARGET_ARCH32) | |
3302 | emit_insn (gen_tie_ld32 (temp3, got, temp2, addr)); | |
3303 | else | |
3304 | emit_insn (gen_tie_ld64 (temp3, got, temp2, addr)); | |
3305 | if (TARGET_SUN_TLS) | |
3306 | { | |
3307 | ret = gen_reg_rtx (Pmode); | |
3308 | if (TARGET_ARCH32) | |
3309 | emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7), | |
3310 | temp3, addr)); | |
3311 | else | |
3312 | emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7), | |
3313 | temp3, addr)); | |
3314 | } | |
3315 | else | |
3316 | ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3); | |
3317 | break; | |
3318 | ||
3319 | case TLS_MODEL_LOCAL_EXEC: | |
3320 | temp1 = gen_reg_rtx (Pmode); | |
3321 | temp2 = gen_reg_rtx (Pmode); | |
3322 | if (TARGET_ARCH32) | |
3323 | { | |
3324 | emit_insn (gen_tle_hix22_sp32 (temp1, addr)); | |
3325 | emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr)); | |
3326 | } | |
3327 | else | |
3328 | { | |
3329 | emit_insn (gen_tle_hix22_sp64 (temp1, addr)); | |
3330 | emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr)); | |
3331 | } | |
3332 | ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2); | |
3333 | break; | |
3334 | ||
3335 | default: | |
f5f7d171 | 3336 | gcc_unreachable (); |
5751a10b JJ |
3337 | } |
3338 | ||
208384fd EB |
3339 | else if (GET_CODE (addr) == CONST) |
3340 | { | |
3341 | rtx base, offset; | |
3342 | ||
3343 | gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS); | |
3344 | ||
3345 | base = legitimize_tls_address (XEXP (XEXP (addr, 0), 0)); | |
3346 | offset = XEXP (XEXP (addr, 0), 1); | |
3347 | ||
3348 | base = force_operand (base, NULL_RTX); | |
3349 | if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset))) | |
3350 | offset = force_reg (Pmode, offset); | |
3351 | ret = gen_rtx_PLUS (Pmode, base, offset); | |
3352 | } | |
3353 | ||
5751a10b | 3354 | else |
f5f7d171 | 3355 | gcc_unreachable (); /* for now ... */ |
5751a10b JJ |
3356 | |
3357 | return ret; | |
3358 | } | |
3359 | ||
e0d80184 DM |
3360 | /* Legitimize PIC addresses. If the address is already position-independent, |
3361 | we return ORIG. Newly generated position-independent addresses go into a | |
88cad84b | 3362 | reg. This is REG if nonzero, otherwise we allocate register(s) as |
e0d80184 | 3363 | necessary. */ |
795068a4 | 3364 | |
208384fd EB |
3365 | static rtx |
3366 | legitimize_pic_address (rtx orig, rtx reg) | |
e0d80184 | 3367 | { |
878ee0ab DM |
3368 | bool gotdata_op = false; |
3369 | ||
1910440e RS |
3370 | if (GET_CODE (orig) == SYMBOL_REF |
3371 | /* See the comment in sparc_expand_move. */ | |
3372 | || (TARGET_VXWORKS_RTP && GET_CODE (orig) == LABEL_REF)) | |
82d6b402 | 3373 | { |
e0d80184 DM |
3374 | rtx pic_ref, address; |
3375 | rtx insn; | |
82d6b402 | 3376 | |
e0d80184 | 3377 | if (reg == 0) |
82d6b402 | 3378 | { |
f5f7d171 JM |
3379 | gcc_assert (! reload_in_progress && ! reload_completed); |
3380 | reg = gen_reg_rtx (Pmode); | |
82d6b402 | 3381 | } |
795068a4 | 3382 | |
e0d80184 | 3383 | if (flag_pic == 2) |
4fb4e4b8 | 3384 | { |
e0d80184 DM |
3385 | /* If not during reload, allocate another temp reg here for loading |
3386 | in the address, so that these instructions can be optimized | |
3387 | properly. */ | |
3388 | rtx temp_reg = ((reload_in_progress || reload_completed) | |
3389 | ? reg : gen_reg_rtx (Pmode)); | |
3390 | ||
3391 | /* Must put the SYMBOL_REF inside an UNSPEC here so that cse | |
3392 | won't get confused into thinking that these two instructions | |
3393 | are loading in the true address of the symbol. If in the | |
3394 | future a PIC rtx exists, that should be used instead. */ | |
bea5071f | 3395 | if (TARGET_ARCH64) |
be3f1ff5 | 3396 | { |
bea5071f EB |
3397 | emit_insn (gen_movdi_high_pic (temp_reg, orig)); |
3398 | emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig)); | |
be3f1ff5 DM |
3399 | } |
3400 | else | |
3401 | { | |
bea5071f EB |
3402 | emit_insn (gen_movsi_high_pic (temp_reg, orig)); |
3403 | emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig)); | |
be3f1ff5 | 3404 | } |
e0d80184 | 3405 | address = temp_reg; |
878ee0ab | 3406 | gotdata_op = true; |
4fb4e4b8 DE |
3407 | } |
3408 | else | |
e0d80184 DM |
3409 | address = orig; |
3410 | ||
e3b5732b | 3411 | crtl->uses_pic_offset_table = 1; |
878ee0ab DM |
3412 | if (gotdata_op) |
3413 | { | |
3414 | if (TARGET_ARCH64) | |
3415 | insn = emit_insn (gen_movdi_pic_gotdata_op (reg, pic_offset_table_rtx, | |
3416 | address, orig)); | |
3417 | else | |
3418 | insn = emit_insn (gen_movsi_pic_gotdata_op (reg, pic_offset_table_rtx, | |
3419 | address, orig)); | |
3420 | } | |
3421 | else | |
3422 | { | |
3423 | pic_ref = gen_const_mem (Pmode, | |
3424 | gen_rtx_PLUS (Pmode, | |
3425 | pic_offset_table_rtx, address)); | |
3426 | insn = emit_move_insn (reg, pic_ref); | |
3427 | } | |
e0d80184 DM |
3428 | /* Put a REG_EQUAL note on this insn, so that it can be optimized |
3429 | by loop. */ | |
bd94cb6e | 3430 | set_unique_reg_note (insn, REG_EQUAL, orig); |
e0d80184 | 3431 | return reg; |
795068a4 | 3432 | } |
e0d80184 | 3433 | else if (GET_CODE (orig) == CONST) |
795068a4 | 3434 | { |
e0d80184 | 3435 | rtx base, offset; |
795068a4 | 3436 | |
e0d80184 DM |
3437 | if (GET_CODE (XEXP (orig, 0)) == PLUS |
3438 | && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx) | |
3439 | return orig; | |
82d6b402 | 3440 | |
e0d80184 | 3441 | if (reg == 0) |
4fb4e4b8 | 3442 | { |
f5f7d171 JM |
3443 | gcc_assert (! reload_in_progress && ! reload_completed); |
3444 | reg = gen_reg_rtx (Pmode); | |
4fb4e4b8 | 3445 | } |
e0d80184 | 3446 | |
f5f7d171 | 3447 | gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS); |
208384fd EB |
3448 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg); |
3449 | offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), | |
3450 | base == reg ? NULL_RTX : reg); | |
61a66555 | 3451 | |
e0d80184 DM |
3452 | if (GET_CODE (offset) == CONST_INT) |
3453 | { | |
3454 | if (SMALL_INT (offset)) | |
ed8908e7 | 3455 | return plus_constant (base, INTVAL (offset)); |
e0d80184 DM |
3456 | else if (! reload_in_progress && ! reload_completed) |
3457 | offset = force_reg (Pmode, offset); | |
3458 | else | |
3459 | /* If we reach here, then something is seriously wrong. */ | |
f5f7d171 | 3460 | gcc_unreachable (); |
e0d80184 DM |
3461 | } |
3462 | return gen_rtx_PLUS (Pmode, base, offset); | |
795068a4 | 3463 | } |
e0d80184 DM |
3464 | else if (GET_CODE (orig) == LABEL_REF) |
3465 | /* ??? Why do we do this? */ | |
3466 | /* Now movsi_pic_label_ref uses it, but we ought to be checking that | |
3467 | the register is live instead, in case it is eliminated. */ | |
e3b5732b | 3468 | crtl->uses_pic_offset_table = 1; |
795068a4 | 3469 | |
e0d80184 DM |
3470 | return orig; |
3471 | } | |
795068a4 | 3472 | |
5751a10b JJ |
3473 | /* Try machine-dependent ways of modifying an illegitimate address X |
3474 | to be legitimate. If we find one, return the new, valid address. | |
3475 | ||
3476 | OLDX is the address as it was before break_out_memory_refs was called. | |
3477 | In some cases it is useful to look at this to decide what needs to be done. | |
3478 | ||
506d7b68 PB |
3479 | MODE is the mode of the operand pointed to by X. |
3480 | ||
3481 | On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */ | |
5751a10b | 3482 | |
208384fd | 3483 | static rtx |
506d7b68 PB |
3484 | sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, |
3485 | enum machine_mode mode) | |
5751a10b JJ |
3486 | { |
3487 | rtx orig_x = x; | |
3488 | ||
3489 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT) | |
3490 | x = gen_rtx_PLUS (Pmode, XEXP (x, 1), | |
3491 | force_operand (XEXP (x, 0), NULL_RTX)); | |
3492 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT) | |
3493 | x = gen_rtx_PLUS (Pmode, XEXP (x, 0), | |
3494 | force_operand (XEXP (x, 1), NULL_RTX)); | |
3495 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS) | |
3496 | x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX), | |
3497 | XEXP (x, 1)); | |
3498 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS) | |
3499 | x = gen_rtx_PLUS (Pmode, XEXP (x, 0), | |
3500 | force_operand (XEXP (x, 1), NULL_RTX)); | |
3501 | ||
c6c3dba9 | 3502 | if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE)) |
5751a10b JJ |
3503 | return x; |
3504 | ||
208384fd | 3505 | if (sparc_tls_referenced_p (x)) |
5751a10b JJ |
3506 | x = legitimize_tls_address (x); |
3507 | else if (flag_pic) | |
208384fd | 3508 | x = legitimize_pic_address (x, NULL_RTX); |
5751a10b JJ |
3509 | else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1))) |
3510 | x = gen_rtx_PLUS (Pmode, XEXP (x, 0), | |
3511 | copy_to_mode_reg (Pmode, XEXP (x, 1))); | |
3512 | else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0))) | |
3513 | x = gen_rtx_PLUS (Pmode, XEXP (x, 1), | |
3514 | copy_to_mode_reg (Pmode, XEXP (x, 0))); | |
3515 | else if (GET_CODE (x) == SYMBOL_REF | |
3516 | || GET_CODE (x) == CONST | |
208384fd | 3517 | || GET_CODE (x) == LABEL_REF) |
5751a10b | 3518 | x = copy_to_suggested_reg (x, NULL_RTX, Pmode); |
208384fd | 3519 | |
5751a10b JJ |
3520 | return x; |
3521 | } | |
3522 | ||
9fd52c6c DM |
3523 | #ifdef HAVE_GAS_HIDDEN |
3524 | # define USE_HIDDEN_LINKONCE 1 | |
3525 | #else | |
3526 | # define USE_HIDDEN_LINKONCE 0 | |
3527 | #endif | |
795068a4 | 3528 | |
9ac617d4 | 3529 | static void |
9fd52c6c | 3530 | get_pc_thunk_name (char name[32], unsigned int regno) |
e0d80184 | 3531 | { |
9fd52c6c | 3532 | const char *pic_name = reg_names[regno]; |
4fb4e4b8 | 3533 | |
9fd52c6c DM |
3534 | /* Skip the leading '%' as that cannot be used in a |
3535 | symbol name. */ | |
3536 | pic_name += 1; | |
82d6b402 | 3537 | |
9fd52c6c DM |
3538 | if (USE_HIDDEN_LINKONCE) |
3539 | sprintf (name, "__sparc_get_pc_thunk.%s", pic_name); | |
5e77eb53 | 3540 | else |
9fd52c6c | 3541 | ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno); |
5e77eb53 EB |
3542 | } |
3543 | ||
3544 | /* Emit code to load the PIC register. */ | |
3545 | ||
3546 | static void | |
9fd52c6c | 3547 | load_pic_register (void) |
5e77eb53 EB |
3548 | { |
3549 | int orig_flag_pic = flag_pic; | |
3550 | ||
1910440e RS |
3551 | if (TARGET_VXWORKS_RTP) |
3552 | { | |
3553 | emit_insn (gen_vxworks_load_got ()); | |
c41c1387 | 3554 | emit_use (pic_offset_table_rtx); |
1910440e RS |
3555 | return; |
3556 | } | |
3557 | ||
5e77eb53 | 3558 | /* If we haven't initialized the special PIC symbols, do so now. */ |
9fd52c6c | 3559 | if (!pic_helper_needed) |
5e77eb53 | 3560 | { |
9fd52c6c DM |
3561 | char name[32]; |
3562 | ||
3563 | pic_helper_needed = true; | |
3564 | ||
3565 | get_pc_thunk_name (name, REGNO (pic_offset_table_rtx)); | |
3566 | pic_helper_symbol = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name)); | |
3567 | ||
5e77eb53 | 3568 | global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); |
e0d80184 | 3569 | } |
82d6b402 | 3570 | |
ef573211 | 3571 | flag_pic = 0; |
eb7a7903 EB |
3572 | if (TARGET_ARCH64) |
3573 | emit_insn (gen_load_pcrel_symdi (pic_offset_table_rtx, global_offset_table, | |
5e77eb53 | 3574 | pic_helper_symbol)); |
eb7a7903 EB |
3575 | else |
3576 | emit_insn (gen_load_pcrel_symsi (pic_offset_table_rtx, global_offset_table, | |
5e77eb53 | 3577 | pic_helper_symbol)); |
e0d80184 | 3578 | flag_pic = orig_flag_pic; |
82d6b402 | 3579 | |
e0d80184 DM |
3580 | /* Need to emit this whether or not we obey regdecls, |
3581 | since setjmp/longjmp can cause life info to screw up. | |
3582 | ??? In the case where we don't obey regdecls, this is not sufficient | |
3583 | since we may not fall out the bottom. */ | |
c41c1387 | 3584 | emit_use (pic_offset_table_rtx); |
795068a4 | 3585 | } |
1910440e RS |
3586 | |
3587 | /* Emit a call instruction with the pattern given by PAT. ADDR is the | |
3588 | address of the call target. */ | |
3589 | ||
3590 | void | |
3591 | sparc_emit_call_insn (rtx pat, rtx addr) | |
3592 | { | |
3593 | rtx insn; | |
3594 | ||
3595 | insn = emit_call_insn (pat); | |
3596 | ||
3597 | /* The PIC register is live on entry to VxWorks PIC PLT entries. */ | |
3598 | if (TARGET_VXWORKS_RTP | |
3599 | && flag_pic | |
3600 | && GET_CODE (addr) == SYMBOL_REF | |
3601 | && (SYMBOL_REF_DECL (addr) | |
3602 | ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr)) | |
3603 | : !SYMBOL_REF_LOCAL_P (addr))) | |
3604 | { | |
3605 | use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx); | |
e3b5732b | 3606 | crtl->uses_pic_offset_table = 1; |
1910440e RS |
3607 | } |
3608 | } | |
ab835497 | 3609 | \f |
e0d80184 | 3610 | /* Return 1 if RTX is a MEM which is known to be aligned to at |
7a31a340 | 3611 | least a DESIRED byte boundary. */ |
a3ee5899 | 3612 | |
e0d80184 | 3613 | int |
fc27d102 | 3614 | mem_min_alignment (rtx mem, int desired) |
ab835497 | 3615 | { |
e0d80184 DM |
3616 | rtx addr, base, offset; |
3617 | ||
3618 | /* If it's not a MEM we can't accept it. */ | |
3619 | if (GET_CODE (mem) != MEM) | |
3620 | return 0; | |
3621 | ||
167a1354 | 3622 | /* Obviously... */ |
0f95e914 EB |
3623 | if (!TARGET_UNALIGNED_DOUBLES |
3624 | && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired) | |
167a1354 EB |
3625 | return 1; |
3626 | ||
3627 | /* ??? The rest of the function predates MEM_ALIGN so | |
3628 | there is probably a bit of redundancy. */ | |
e0d80184 DM |
3629 | addr = XEXP (mem, 0); |
3630 | base = offset = NULL_RTX; | |
3631 | if (GET_CODE (addr) == PLUS) | |
ab835497 | 3632 | { |
e0d80184 | 3633 | if (GET_CODE (XEXP (addr, 0)) == REG) |
61a55e8b | 3634 | { |
e0d80184 DM |
3635 | base = XEXP (addr, 0); |
3636 | ||
3637 | /* What we are saying here is that if the base | |
3638 | REG is aligned properly, the compiler will make | |
3639 | sure any REG based index upon it will be so | |
3640 | as well. */ | |
3641 | if (GET_CODE (XEXP (addr, 1)) == CONST_INT) | |
3642 | offset = XEXP (addr, 1); | |
61a55e8b | 3643 | else |
e0d80184 | 3644 | offset = const0_rtx; |
61a55e8b | 3645 | } |
ab835497 | 3646 | } |
e0d80184 | 3647 | else if (GET_CODE (addr) == REG) |
ab835497 | 3648 | { |
e0d80184 DM |
3649 | base = addr; |
3650 | offset = const0_rtx; | |
ab835497 | 3651 | } |
82d6b402 | 3652 | |
e0d80184 | 3653 | if (base != NULL_RTX) |
795068a4 | 3654 | { |
e0d80184 DM |
3655 | int regno = REGNO (base); |
3656 | ||
563c12b0 | 3657 | if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM) |
61a55e8b | 3658 | { |
e0d80184 DM |
3659 | /* Check if the compiler has recorded some information |
3660 | about the alignment of the base REG. If reload has | |
021e2ecc JJ |
3661 | completed, we already matched with proper alignments. |
3662 | If not running global_alloc, reload might give us | |
3663 | unaligned pointer to local stack though. */ | |
bdb429a5 RK |
3664 | if (((cfun != 0 |
3665 | && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT) | |
021e2ecc | 3666 | || (optimize && reload_completed)) |
bdb429a5 | 3667 | && (INTVAL (offset) & (desired - 1)) == 0) |
e0d80184 | 3668 | return 1; |
61a55e8b | 3669 | } |
795068a4 | 3670 | else |
e0d80184 DM |
3671 | { |
3672 | if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0) | |
3673 | return 1; | |
3674 | } | |
795068a4 | 3675 | } |
e0d80184 DM |
3676 | else if (! TARGET_UNALIGNED_DOUBLES |
3677 | || CONSTANT_P (addr) | |
3678 | || GET_CODE (addr) == LO_SUM) | |
795068a4 | 3679 | { |
e0d80184 DM |
3680 | /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES |
3681 | is true, in which case we can only assume that an access is aligned if | |
3682 | it is to a constant address, or the address involves a LO_SUM. */ | |
3683 | return 1; | |
795068a4 | 3684 | } |
4900aa61 | 3685 | |
e0d80184 DM |
3686 | /* An obviously unaligned address. */ |
3687 | return 0; | |
795068a4 | 3688 | } |
e0d80184 | 3689 | |
ab835497 | 3690 | \f |
61a55e8b | 3691 | /* Vectors to keep interesting information about registers where it can easily |
01d3224a | 3692 | be got. We used to use the actual mode value as the bit number, but there |
61a55e8b DE |
3693 | are more than 32 modes now. Instead we use two tables: one indexed by |
3694 | hard register number, and one indexed by mode. */ | |
3695 | ||
3696 | /* The purpose of sparc_mode_class is to shrink the range of modes so that | |
a7b376ee | 3697 | they all fit (as bit numbers) in a 32-bit word (again). Each real mode is |
61a55e8b DE |
3698 | mapped into one sparc_mode_class mode. */ |
3699 | ||
3700 | enum sparc_mode_class { | |
61a55e8b | 3701 | S_MODE, D_MODE, T_MODE, O_MODE, |
c4ce6853 DE |
3702 | SF_MODE, DF_MODE, TF_MODE, OF_MODE, |
3703 | CC_MODE, CCFP_MODE | |
61a55e8b | 3704 | }; |
ab835497 | 3705 | |
61a55e8b DE |
3706 | /* Modes for single-word and smaller quantities. */ |
3707 | #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE)) | |
3708 | ||
3709 | /* Modes for double-word and smaller quantities. */ | |
3710 | #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE)) | |
3711 | ||
3712 | /* Modes for quad-word and smaller quantities. */ | |
3713 | #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE)) | |
ab835497 | 3714 | |
d0cee9ba JJ |
3715 | /* Modes for 8-word and smaller quantities. */ |
3716 | #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE)) | |
3717 | ||
1e4eb8d1 JW |
3718 | /* Modes for single-float quantities. We must allow any single word or |
3719 | smaller quantity. This is because the fix/float conversion instructions | |
3720 | take integer inputs/outputs from the float registers. */ | |
3721 | #define SF_MODES (S_MODES) | |
ab835497 | 3722 | |
61a55e8b | 3723 | /* Modes for double-float and smaller quantities. */ |
948bf106 | 3724 | #define DF_MODES (D_MODES) |
61a55e8b DE |
3725 | |
3726 | /* Modes for quad-float and smaller quantities. */ | |
948bf106 | 3727 | #define TF_MODES (DF_MODES | (1 << (int) TF_MODE)) |
61a55e8b | 3728 | |
948bf106 EB |
3729 | /* Modes for quad-float pairs and smaller quantities. */ |
3730 | #define OF_MODES (TF_MODES | (1 << (int) OF_MODE)) | |
d0cee9ba | 3731 | |
948bf106 EB |
3732 | /* Modes for double-float only quantities. */ |
3733 | #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE)) | |
d0cee9ba | 3734 | |
948bf106 EB |
3735 | /* Modes for quad-float and double-float only quantities. */ |
3736 | #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE)) | |
d0cee9ba | 3737 | |
948bf106 EB |
3738 | /* Modes for quad-float pairs and double-float only quantities. */ |
3739 | #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE)) | |
ab835497 | 3740 | |
c4ce6853 DE |
3741 | /* Modes for condition codes. */ |
3742 | #define CC_MODES (1 << (int) CC_MODE) | |
3743 | #define CCFP_MODES (1 << (int) CCFP_MODE) | |
3744 | ||
ab835497 RK |
3745 | /* Value is 1 if register/mode pair is acceptable on sparc. |
3746 | The funny mixture of D and T modes is because integer operations | |
3747 | do not specially operate on tetra quantities, so non-quad-aligned | |
3748 | registers can hold quadword quantities (except %o4 and %i4 because | |
c4ce6853 | 3749 | they cross fixed registers). */ |
ab835497 | 3750 | |
61a55e8b | 3751 | /* This points to either the 32 bit or the 64 bit version. */ |
0b5826ac | 3752 | const int *hard_regno_mode_classes; |
61a55e8b | 3753 | |
0b5826ac | 3754 | static const int hard_32bit_mode_classes[] = { |
c4ce6853 | 3755 | S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, |
ab835497 RK |
3756 | T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES, |
3757 | T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, | |
3758 | T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES, | |
3759 | ||
d0cee9ba JJ |
3760 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, |
3761 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
3762 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
3763 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
a9e27770 DE |
3764 | |
3765 | /* FP regs f32 to f63. Only the even numbered registers actually exist, | |
3766 | and none can hold SFmode/SImode values. */ | |
d0cee9ba JJ |
3767 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, |
3768 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
3769 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
3770 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
a9e27770 DE |
3771 | |
3772 | /* %fcc[0123] */ | |
c4ce6853 DE |
3773 | CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES, |
3774 | ||
3775 | /* %icc */ | |
3776 | CC_MODES | |
61a55e8b DE |
3777 | }; |
3778 | ||
0b5826ac | 3779 | static const int hard_64bit_mode_classes[] = { |
c4ce6853 | 3780 | D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
d0cee9ba | 3781 | O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
61a55e8b | 3782 | T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
d0cee9ba | 3783 | O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
61a55e8b | 3784 | |
d0cee9ba JJ |
3785 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, |
3786 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
3787 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
3788 | OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
61a55e8b | 3789 | |
a9e27770 | 3790 | /* FP regs f32 to f63. Only the even numbered registers actually exist, |
61a55e8b | 3791 | and none can hold SFmode/SImode values. */ |
d0cee9ba JJ |
3792 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, |
3793 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
3794 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
3795 | OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0, | |
61a55e8b DE |
3796 | |
3797 | /* %fcc[0123] */ | |
c4ce6853 DE |
3798 | CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES, |
3799 | ||
3800 | /* %icc */ | |
3801 | CC_MODES | |
61a55e8b DE |
3802 | }; |
3803 | ||
3804 | int sparc_mode_class [NUM_MACHINE_MODES]; | |
3805 | ||
c4ce6853 DE |
3806 | enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER]; |
3807 | ||
61a55e8b | 3808 | static void |
fc27d102 | 3809 | sparc_init_modes (void) |
61a55e8b DE |
3810 | { |
3811 | int i; | |
3812 | ||
61a55e8b DE |
3813 | for (i = 0; i < NUM_MACHINE_MODES; i++) |
3814 | { | |
3815 | switch (GET_MODE_CLASS (i)) | |
3816 | { | |
3817 | case MODE_INT: | |
3818 | case MODE_PARTIAL_INT: | |
3819 | case MODE_COMPLEX_INT: | |
3820 | if (GET_MODE_SIZE (i) <= 4) | |
3821 | sparc_mode_class[i] = 1 << (int) S_MODE; | |
3822 | else if (GET_MODE_SIZE (i) == 8) | |
3823 | sparc_mode_class[i] = 1 << (int) D_MODE; | |
3824 | else if (GET_MODE_SIZE (i) == 16) | |
3825 | sparc_mode_class[i] = 1 << (int) T_MODE; | |
3826 | else if (GET_MODE_SIZE (i) == 32) | |
3827 | sparc_mode_class[i] = 1 << (int) O_MODE; | |
4900aa61 | 3828 | else |
61a55e8b DE |
3829 | sparc_mode_class[i] = 0; |
3830 | break; | |
c75d6010 JM |
3831 | case MODE_VECTOR_INT: |
3832 | if (GET_MODE_SIZE (i) <= 4) | |
3833 | sparc_mode_class[i] = 1 << (int)SF_MODE; | |
3834 | else if (GET_MODE_SIZE (i) == 8) | |
3835 | sparc_mode_class[i] = 1 << (int)DF_MODE; | |
3836 | break; | |
61a55e8b DE |
3837 | case MODE_FLOAT: |
3838 | case MODE_COMPLEX_FLOAT: | |
3839 | if (GET_MODE_SIZE (i) <= 4) | |
3840 | sparc_mode_class[i] = 1 << (int) SF_MODE; | |
3841 | else if (GET_MODE_SIZE (i) == 8) | |
3842 | sparc_mode_class[i] = 1 << (int) DF_MODE; | |
3843 | else if (GET_MODE_SIZE (i) == 16) | |
3844 | sparc_mode_class[i] = 1 << (int) TF_MODE; | |
3845 | else if (GET_MODE_SIZE (i) == 32) | |
3846 | sparc_mode_class[i] = 1 << (int) OF_MODE; | |
4900aa61 | 3847 | else |
61a55e8b DE |
3848 | sparc_mode_class[i] = 0; |
3849 | break; | |
3850 | case MODE_CC: | |
61a55e8b DE |
3851 | if (i == (int) CCFPmode || i == (int) CCFPEmode) |
3852 | sparc_mode_class[i] = 1 << (int) CCFP_MODE; | |
61a55e8b | 3853 | else |
94134f42 ZW |
3854 | sparc_mode_class[i] = 1 << (int) CC_MODE; |
3855 | break; | |
3856 | default: | |
3857 | sparc_mode_class[i] = 0; | |
61a55e8b DE |
3858 | break; |
3859 | } | |
3860 | } | |
3861 | ||
a9e27770 | 3862 | if (TARGET_ARCH64) |
61a55e8b DE |
3863 | hard_regno_mode_classes = hard_64bit_mode_classes; |
3864 | else | |
3865 | hard_regno_mode_classes = hard_32bit_mode_classes; | |
c4ce6853 DE |
3866 | |
3867 | /* Initialize the array used by REGNO_REG_CLASS. */ | |
3868 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
3869 | { | |
284d86e9 JC |
3870 | if (i < 16 && TARGET_V8PLUS) |
3871 | sparc_regno_reg_class[i] = I64_REGS; | |
563c12b0 | 3872 | else if (i < 32 || i == FRAME_POINTER_REGNUM) |
c4ce6853 DE |
3873 | sparc_regno_reg_class[i] = GENERAL_REGS; |
3874 | else if (i < 64) | |
3875 | sparc_regno_reg_class[i] = FP_REGS; | |
3876 | else if (i < 96) | |
3877 | sparc_regno_reg_class[i] = EXTRA_FP_REGS; | |
3878 | else if (i < 100) | |
3879 | sparc_regno_reg_class[i] = FPCC_REGS; | |
3880 | else | |
3881 | sparc_regno_reg_class[i] = NO_REGS; | |
3882 | } | |
61a55e8b | 3883 | } |
ab835497 | 3884 | \f |
61a55e8b | 3885 | /* Compute the frame size required by the function. This function is called |
9ac617d4 | 3886 | during the reload pass and also by sparc_expand_prologue. */ |
ab835497 | 3887 | |
6a151f87 | 3888 | HOST_WIDE_INT |
9ac617d4 | 3889 | sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p) |
ab835497 | 3890 | { |
38173d38 | 3891 | int outgoing_args_size = (crtl->outgoing_args_size |
4fb4e4b8 | 3892 | + REG_PARM_STACK_SPACE (current_function_decl)); |
9ac617d4 EB |
3893 | int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */ |
3894 | int i; | |
61a55e8b | 3895 | |
33074e5f RH |
3896 | if (TARGET_ARCH64) |
3897 | { | |
3898 | for (i = 0; i < 8; i++) | |
6fb5fa3c | 3899 | if (df_regs_ever_live_p (i) && ! call_used_regs[i]) |
33074e5f RH |
3900 | n_regs += 2; |
3901 | } | |
3902 | else | |
3903 | { | |
3904 | for (i = 0; i < 8; i += 2) | |
6fb5fa3c DB |
3905 | if ((df_regs_ever_live_p (i) && ! call_used_regs[i]) |
3906 | || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1])) | |
61a55e8b | 3907 | n_regs += 2; |
ab835497 RK |
3908 | } |
3909 | ||
33074e5f | 3910 | for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2) |
6fb5fa3c DB |
3911 | if ((df_regs_ever_live_p (i) && ! call_used_regs[i]) |
3912 | || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1])) | |
33074e5f RH |
3913 | n_regs += 2; |
3914 | ||
9ac617d4 | 3915 | /* Set up values for use in prologue and epilogue. */ |
61a55e8b | 3916 | num_gfregs = n_regs; |
ab835497 | 3917 | |
9ac617d4 EB |
3918 | if (leaf_function_p |
3919 | && n_regs == 0 | |
3920 | && size == 0 | |
38173d38 | 3921 | && crtl->outgoing_args_size == 0) |
9ac617d4 | 3922 | actual_fsize = apparent_fsize = 0; |
61a55e8b DE |
3923 | else |
3924 | { | |
225909c3 JS |
3925 | /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */ |
3926 | apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8; | |
61a55e8b DE |
3927 | apparent_fsize += n_regs * 4; |
3928 | actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8); | |
3929 | } | |
ab835497 RK |
3930 | |
3931 | /* Make sure nothing can clobber our register windows. | |
3932 | If a SAVE must be done, or there is a stack-local variable, | |
f8ece000 | 3933 | the register window area must be allocated. */ |
9ac617d4 | 3934 | if (! leaf_function_p || size > 0) |
f8ece000 | 3935 | actual_fsize += FIRST_PARM_OFFSET (current_function_decl); |
ab835497 | 3936 | |
61a55e8b DE |
3937 | return SPARC_STACK_ALIGN (actual_fsize); |
3938 | } | |
3939 | ||
9ac617d4 | 3940 | /* Output any necessary .register pseudo-ops. */ |
61a55e8b | 3941 | |
9ac617d4 EB |
3942 | void |
3943 | sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED) | |
61a55e8b | 3944 | { |
9ac617d4 EB |
3945 | #ifdef HAVE_AS_REGISTER_PSEUDO_OP |
3946 | int i; | |
6a151f87 | 3947 | |
9ac617d4 EB |
3948 | if (TARGET_ARCH32) |
3949 | return; | |
3950 | ||
3951 | /* Check if %g[2367] were used without | |
3952 | .register being printed for them already. */ | |
3953 | for (i = 2; i < 8; i++) | |
61a55e8b | 3954 | { |
6fb5fa3c | 3955 | if (df_regs_ever_live_p (i) |
9ac617d4 EB |
3956 | && ! sparc_hard_reg_printed [i]) |
3957 | { | |
3958 | sparc_hard_reg_printed [i] = 1; | |
392e5756 JJ |
3959 | /* %g7 is used as TLS base register, use #ignore |
3960 | for it instead of #scratch. */ | |
3961 | fprintf (file, "\t.register\t%%g%d, #%s\n", i, | |
3962 | i == 7 ? "ignore" : "scratch"); | |
9ac617d4 EB |
3963 | } |
3964 | if (i == 3) i = 5; | |
61a55e8b | 3965 | } |
6a151f87 | 3966 | #endif |
9ac617d4 | 3967 | } |
6a151f87 | 3968 | |
9ac617d4 EB |
3969 | /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET |
3970 | as needed. LOW should be double-word aligned for 32-bit registers. | |
3971 | Return the new OFFSET. */ | |
6a151f87 | 3972 | |
9ac617d4 EB |
3973 | #define SORR_SAVE 0 |
3974 | #define SORR_RESTORE 1 | |
3975 | ||
3976 | static int | |
3977 | save_or_restore_regs (int low, int high, rtx base, int offset, int action) | |
3978 | { | |
3979 | rtx mem, insn; | |
3980 | int i; | |
3981 | ||
3982 | if (TARGET_ARCH64 && high <= 32) | |
3983 | { | |
3984 | for (i = low; i < high; i++) | |
6a151f87 | 3985 | { |
6fb5fa3c | 3986 | if (df_regs_ever_live_p (i) && ! call_used_regs[i]) |
9ac617d4 EB |
3987 | { |
3988 | mem = gen_rtx_MEM (DImode, plus_constant (base, offset)); | |
3989 | set_mem_alias_set (mem, sparc_sr_alias_set); | |
3990 | if (action == SORR_SAVE) | |
3991 | { | |
3992 | insn = emit_move_insn (mem, gen_rtx_REG (DImode, i)); | |
3993 | RTX_FRAME_RELATED_P (insn) = 1; | |
3994 | } | |
3995 | else /* action == SORR_RESTORE */ | |
3996 | emit_move_insn (gen_rtx_REG (DImode, i), mem); | |
3997 | offset += 8; | |
3998 | } | |
6a151f87 | 3999 | } |
9ac617d4 EB |
4000 | } |
4001 | else | |
4002 | { | |
4003 | for (i = low; i < high; i += 2) | |
6a151f87 | 4004 | { |
6fb5fa3c DB |
4005 | bool reg0 = df_regs_ever_live_p (i) && ! call_used_regs[i]; |
4006 | bool reg1 = df_regs_ever_live_p (i+1) && ! call_used_regs[i+1]; | |
9ac617d4 EB |
4007 | enum machine_mode mode; |
4008 | int regno; | |
4009 | ||
4010 | if (reg0 && reg1) | |
4011 | { | |
4012 | mode = i < 32 ? DImode : DFmode; | |
4013 | regno = i; | |
4014 | } | |
4015 | else if (reg0) | |
4016 | { | |
4017 | mode = i < 32 ? SImode : SFmode; | |
4018 | regno = i; | |
4019 | } | |
4020 | else if (reg1) | |
4021 | { | |
4022 | mode = i < 32 ? SImode : SFmode; | |
4023 | regno = i + 1; | |
4024 | offset += 4; | |
4025 | } | |
4026 | else | |
4027 | continue; | |
4028 | ||
4029 | mem = gen_rtx_MEM (mode, plus_constant (base, offset)); | |
4030 | set_mem_alias_set (mem, sparc_sr_alias_set); | |
4031 | if (action == SORR_SAVE) | |
4032 | { | |
4033 | insn = emit_move_insn (mem, gen_rtx_REG (mode, regno)); | |
4034 | RTX_FRAME_RELATED_P (insn) = 1; | |
4035 | } | |
4036 | else /* action == SORR_RESTORE */ | |
4037 | emit_move_insn (gen_rtx_REG (mode, regno), mem); | |
4038 | ||
4039 | /* Always preserve double-word alignment. */ | |
4040 | offset = (offset + 7) & -8; | |
6a151f87 | 4041 | } |
6a151f87 | 4042 | } |
9ac617d4 EB |
4043 | |
4044 | return offset; | |
ab835497 RK |
4045 | } |
4046 | ||
9ac617d4 EB |
4047 | /* Emit code to save call-saved registers. */ |
4048 | ||
4049 | static void | |
f8ece000 | 4050 | emit_save_or_restore_regs (int action) |
1cb36a98 | 4051 | { |
9ac617d4 EB |
4052 | HOST_WIDE_INT offset; |
4053 | rtx base; | |
1cb36a98 | 4054 | |
9ac617d4 | 4055 | offset = frame_base_offset - apparent_fsize; |
1cb36a98 | 4056 | |
f8ece000 | 4057 | if (offset < -4096 || offset + num_gfregs * 4 > 4095) |
1cb36a98 | 4058 | { |
9ac617d4 EB |
4059 | /* ??? This might be optimized a little as %g1 might already have a |
4060 | value close enough that a single add insn will do. */ | |
4061 | /* ??? Although, all of this is probably only a temporary fix | |
4062 | because if %g1 can hold a function result, then | |
4063 | sparc_expand_epilogue will lose (the result will be | |
4064 | clobbered). */ | |
4065 | base = gen_rtx_REG (Pmode, 1); | |
4066 | emit_move_insn (base, GEN_INT (offset)); | |
4067 | emit_insn (gen_rtx_SET (VOIDmode, | |
4068 | base, | |
4069 | gen_rtx_PLUS (Pmode, frame_base_reg, base))); | |
4070 | offset = 0; | |
1cb36a98 | 4071 | } |
9ac617d4 EB |
4072 | else |
4073 | base = frame_base_reg; | |
4074 | ||
f8ece000 EB |
4075 | offset = save_or_restore_regs (0, 8, base, offset, action); |
4076 | save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, action); | |
9ac617d4 | 4077 | } |
08c148a8 | 4078 | |
e54c7471 | 4079 | /* Generate a save_register_window insn. */ |
08c148a8 | 4080 | |
e54c7471 EB |
4081 | static rtx |
4082 | gen_save_register_window (rtx increment) | |
08c148a8 | 4083 | { |
9ac617d4 | 4084 | if (TARGET_ARCH64) |
e54c7471 | 4085 | return gen_save_register_windowdi (increment); |
9ac617d4 | 4086 | else |
e54c7471 | 4087 | return gen_save_register_windowsi (increment); |
08c148a8 NB |
4088 | } |
4089 | ||
e54c7471 | 4090 | /* Generate an increment for the stack pointer. */ |
915f619f | 4091 | |
e54c7471 EB |
4092 | static rtx |
4093 | gen_stack_pointer_inc (rtx increment) | |
ab835497 | 4094 | { |
6deb0339 EB |
4095 | return gen_rtx_SET (VOIDmode, |
4096 | stack_pointer_rtx, | |
4097 | gen_rtx_PLUS (Pmode, | |
4098 | stack_pointer_rtx, | |
4099 | increment)); | |
e54c7471 EB |
4100 | } |
4101 | ||
4102 | /* Generate a decrement for the stack pointer. */ | |
4103 | ||
4104 | static rtx | |
4105 | gen_stack_pointer_dec (rtx decrement) | |
4106 | { | |
6deb0339 EB |
4107 | return gen_rtx_SET (VOIDmode, |
4108 | stack_pointer_rtx, | |
4109 | gen_rtx_MINUS (Pmode, | |
4110 | stack_pointer_rtx, | |
4111 | decrement)); | |
9ac617d4 EB |
4112 | } |
4113 | ||
4114 | /* Expand the function prologue. The prologue is responsible for reserving | |
4115 | storage for the frame, saving the call-saved registers and loading the | |
4116 | PIC register if needed. */ | |
4117 | ||
4118 | void | |
4119 | sparc_expand_prologue (void) | |
4120 | { | |
e54c7471 EB |
4121 | rtx insn; |
4122 | int i; | |
4123 | ||
5be9b7a1 EB |
4124 | /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying |
4125 | on the final value of the flag means deferring the prologue/epilogue | |
4126 | expansion until just before the second scheduling pass, which is too | |
4127 | late to emit multiple epilogues or return insns. | |
4128 | ||
4129 | Of course we are making the assumption that the value of the flag | |
4130 | will not change between now and its final value. Of the three parts | |
4131 | of the formula, only the last one can reasonably vary. Let's take a | |
4132 | closer look, after assuming that the first two ones are set to true | |
4133 | (otherwise the last value is effectively silenced). | |
4134 | ||
4135 | If only_leaf_regs_used returns false, the global predicate will also | |
4136 | be false so the actual frame size calculated below will be positive. | |
4137 | As a consequence, the save_register_window insn will be emitted in | |
4138 | the instruction stream; now this insn explicitly references %fp | |
4139 | which is not a leaf register so only_leaf_regs_used will always | |
4140 | return false subsequently. | |
4141 | ||
4142 | If only_leaf_regs_used returns true, we hope that the subsequent | |
4143 | optimization passes won't cause non-leaf registers to pop up. For | |
4144 | example, the regrename pass has special provisions to not rename to | |
4145 | non-leaf registers in a leaf function. */ | |
4146 | sparc_leaf_function_p | |
4147 | = optimize > 0 && leaf_function_p () && only_leaf_regs_used (); | |
1cb36a98 | 4148 | |
915f619f JW |
4149 | /* Need to use actual_fsize, since we are also allocating |
4150 | space for our callee (and our own register save area). */ | |
5be9b7a1 EB |
4151 | actual_fsize |
4152 | = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p); | |
4153 | ||
4154 | /* Advertise that the data calculated just above are now valid. */ | |
4155 | sparc_prologue_data_valid_p = true; | |
ab835497 | 4156 | |
5be9b7a1 | 4157 | if (sparc_leaf_function_p) |
61a55e8b | 4158 | { |
9ac617d4 | 4159 | frame_base_reg = stack_pointer_rtx; |
61a55e8b DE |
4160 | frame_base_offset = actual_fsize + SPARC_STACK_BIAS; |
4161 | } | |
4162 | else | |
4163 | { | |
9ac617d4 | 4164 | frame_base_reg = hard_frame_pointer_rtx; |
61a55e8b DE |
4165 | frame_base_offset = SPARC_STACK_BIAS; |
4166 | } | |
4167 | ||
915f619f JW |
4168 | if (actual_fsize == 0) |
4169 | /* do nothing. */ ; | |
5be9b7a1 | 4170 | else if (sparc_leaf_function_p) |
ab835497 | 4171 | { |
3592ea0d | 4172 | if (actual_fsize <= 4096) |
e54c7471 | 4173 | insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-actual_fsize))); |
3592ea0d | 4174 | else if (actual_fsize <= 8192) |
915f619f | 4175 | { |
e54c7471 EB |
4176 | insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096))); |
4177 | /* %sp is still the CFA register. */ | |
4178 | RTX_FRAME_RELATED_P (insn) = 1; | |
4179 | insn | |
4180 | = emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize))); | |
915f619f JW |
4181 | } |
4182 | else | |
4183 | { | |
9ac617d4 EB |
4184 | rtx reg = gen_rtx_REG (Pmode, 1); |
4185 | emit_move_insn (reg, GEN_INT (-actual_fsize)); | |
e54c7471 | 4186 | insn = emit_insn (gen_stack_pointer_inc (reg)); |
bbbbb16a ILT |
4187 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, |
4188 | gen_stack_pointer_inc (GEN_INT (-actual_fsize))); | |
915f619f | 4189 | } |
e54c7471 EB |
4190 | |
4191 | RTX_FRAME_RELATED_P (insn) = 1; | |
ab835497 | 4192 | } |
9ac617d4 | 4193 | else |
5f4241d5 | 4194 | { |
3592ea0d | 4195 | if (actual_fsize <= 4096) |
e54c7471 | 4196 | insn = emit_insn (gen_save_register_window (GEN_INT (-actual_fsize))); |
3592ea0d DE |
4197 | else if (actual_fsize <= 8192) |
4198 | { | |
e54c7471 EB |
4199 | insn = emit_insn (gen_save_register_window (GEN_INT (-4096))); |
4200 | /* %sp is not the CFA register anymore. */ | |
4201 | emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize))); | |
3592ea0d | 4202 | } |
915f619f | 4203 | else |
3592ea0d | 4204 | { |
9ac617d4 EB |
4205 | rtx reg = gen_rtx_REG (Pmode, 1); |
4206 | emit_move_insn (reg, GEN_INT (-actual_fsize)); | |
e54c7471 | 4207 | insn = emit_insn (gen_save_register_window (reg)); |
3592ea0d | 4208 | } |
e54c7471 EB |
4209 | |
4210 | RTX_FRAME_RELATED_P (insn) = 1; | |
4211 | for (i=0; i < XVECLEN (PATTERN (insn), 0); i++) | |
4212 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, i)) = 1; | |
ab835497 RK |
4213 | } |
4214 | ||
9ac617d4 | 4215 | if (num_gfregs) |
f8ece000 | 4216 | emit_save_or_restore_regs (SORR_SAVE); |
9ac617d4 EB |
4217 | |
4218 | /* Load the PIC register if needed. */ | |
e3b5732b | 4219 | if (flag_pic && crtl->uses_pic_offset_table) |
9fd52c6c | 4220 | load_pic_register (); |
9ac617d4 | 4221 | } |
4900aa61 | 4222 | |
9ac617d4 | 4223 | /* This function generates the assembly code for function entry, which boils |
f8ece000 | 4224 | down to emitting the necessary .register directives. */ |
9ac617d4 EB |
4225 | |
4226 | static void | |
4227 | sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
4228 | { | |
5be9b7a1 | 4229 | /* Check that the assumption we made in sparc_expand_prologue is valid. */ |
f5f7d171 | 4230 | gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs); |
9ac617d4 EB |
4231 | |
4232 | sparc_output_scratch_registers (file); | |
9ac617d4 EB |
4233 | } |
4234 | ||
4235 | /* Expand the function epilogue, either normal or part of a sibcall. | |
4236 | We emit all the instructions except the return or the call. */ | |
c53aa195 | 4237 | |
9ac617d4 EB |
4238 | void |
4239 | sparc_expand_epilogue (void) | |
4240 | { | |
61a55e8b | 4241 | if (num_gfregs) |
f8ece000 | 4242 | emit_save_or_restore_regs (SORR_RESTORE); |
ab835497 | 4243 | |
9ac617d4 EB |
4244 | if (actual_fsize == 0) |
4245 | /* do nothing. */ ; | |
5be9b7a1 | 4246 | else if (sparc_leaf_function_p) |
9ac617d4 EB |
4247 | { |
4248 | if (actual_fsize <= 4096) | |
e54c7471 | 4249 | emit_insn (gen_stack_pointer_dec (GEN_INT (- actual_fsize))); |
9ac617d4 | 4250 | else if (actual_fsize <= 8192) |
61a55e8b | 4251 | { |
e54c7471 EB |
4252 | emit_insn (gen_stack_pointer_dec (GEN_INT (-4096))); |
4253 | emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - actual_fsize))); | |
61a55e8b | 4254 | } |
ab835497 | 4255 | else |
61a55e8b | 4256 | { |
9ac617d4 EB |
4257 | rtx reg = gen_rtx_REG (Pmode, 1); |
4258 | emit_move_insn (reg, GEN_INT (-actual_fsize)); | |
e54c7471 | 4259 | emit_insn (gen_stack_pointer_dec (reg)); |
61a55e8b | 4260 | } |
ab835497 RK |
4261 | } |
4262 | } | |
5be9b7a1 EB |
4263 | |
4264 | /* Return true if it is appropriate to emit `return' instructions in the | |
4265 | body of a function. */ | |
4266 | ||
4267 | bool | |
4268 | sparc_can_use_return_insn_p (void) | |
4269 | { | |
4270 | return sparc_prologue_data_valid_p | |
4271 | && (actual_fsize == 0 || !sparc_leaf_function_p); | |
4272 | } | |
4900aa61 | 4273 | |
9ac617d4 | 4274 | /* This function generates the assembly code for function exit. */ |
4900aa61 | 4275 | |
9ac617d4 EB |
4276 | static void |
4277 | sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
4278 | { | |
4279 | /* If code does not drop into the epilogue, we have to still output | |
4280 | a dummy nop for the sake of sane backtraces. Otherwise, if the | |
4281 | last two instructions of a function were "call foo; dslot;" this | |
19525b57 | 4282 | can make the return PC of foo (i.e. address of call instruction |
9ac617d4 | 4283 | plus 8) point to the first instruction in the next function. */ |
ab835497 | 4284 | |
9ac617d4 | 4285 | rtx insn, last_real_insn; |
7d167afd | 4286 | |
9ac617d4 EB |
4287 | insn = get_last_insn (); |
4288 | ||
4289 | last_real_insn = prev_real_insn (insn); | |
4290 | if (last_real_insn | |
4291 | && GET_CODE (last_real_insn) == INSN | |
4292 | && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE) | |
4293 | last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0); | |
4294 | ||
4295 | if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN) | |
4296 | fputs("\tnop\n", file); | |
4297 | ||
4298 | sparc_output_deferred_case_vectors (); | |
4299 | } | |
4900aa61 | 4300 | |
9ac617d4 | 4301 | /* Output a 'restore' instruction. */ |
4900aa61 | 4302 | |
7d167afd | 4303 | static void |
6244b14e | 4304 | output_restore (rtx pat) |
7d167afd | 4305 | { |
6244b14e | 4306 | rtx operands[3]; |
7d167afd | 4307 | |
6244b14e | 4308 | if (! pat) |
7d167afd | 4309 | { |
9ac617d4 EB |
4310 | fputs ("\t restore\n", asm_out_file); |
4311 | return; | |
7d167afd JJ |
4312 | } |
4313 | ||
f5f7d171 | 4314 | gcc_assert (GET_CODE (pat) == SET); |
08c148a8 | 4315 | |
9ac617d4 EB |
4316 | operands[0] = SET_DEST (pat); |
4317 | pat = SET_SRC (pat); | |
08c148a8 | 4318 | |
9ac617d4 EB |
4319 | switch (GET_CODE (pat)) |
4320 | { | |
4321 | case PLUS: | |
4322 | operands[1] = XEXP (pat, 0); | |
4323 | operands[2] = XEXP (pat, 1); | |
4324 | output_asm_insn (" restore %r1, %2, %Y0", operands); | |
4325 | break; | |
4326 | case LO_SUM: | |
4327 | operands[1] = XEXP (pat, 0); | |
4328 | operands[2] = XEXP (pat, 1); | |
4329 | output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands); | |
4330 | break; | |
4331 | case ASHIFT: | |
4332 | operands[1] = XEXP (pat, 0); | |
f5f7d171 | 4333 | gcc_assert (XEXP (pat, 1) == const1_rtx); |
9ac617d4 EB |
4334 | output_asm_insn (" restore %r1, %r1, %Y0", operands); |
4335 | break; | |
4336 | default: | |
4337 | operands[1] = pat; | |
4338 | output_asm_insn (" restore %%g0, %1, %Y0", operands); | |
4339 | break; | |
4340 | } | |
08c148a8 | 4341 | } |
4900aa61 | 4342 | |
9ac617d4 | 4343 | /* Output a return. */ |
08c148a8 | 4344 | |
9ac617d4 EB |
4345 | const char * |
4346 | output_return (rtx insn) | |
ab835497 | 4347 | { |
5be9b7a1 | 4348 | if (sparc_leaf_function_p) |
e48addee | 4349 | { |
9ac617d4 EB |
4350 | /* This is a leaf function so we don't have to bother restoring the |
4351 | register window, which frees us from dealing with the convoluted | |
4352 | semantics of restore/return. We simply output the jump to the | |
4e5b002b | 4353 | return address and the insn in the delay slot (if any). */ |
7a31a340 | 4354 | |
e3b5732b | 4355 | gcc_assert (! crtl->calls_eh_return); |
ac59ed37 | 4356 | |
4e5b002b | 4357 | return "jmp\t%%o7+%)%#"; |
bfd6bc60 | 4358 | } |
ab835497 | 4359 | else |
ab835497 | 4360 | { |
9ac617d4 EB |
4361 | /* This is a regular function so we have to restore the register window. |
4362 | We may have a pending insn for the delay slot, which will be either | |
4363 | combined with the 'restore' instruction or put in the delay slot of | |
4364 | the 'return' instruction. */ | |
4365 | ||
e3b5732b | 4366 | if (crtl->calls_eh_return) |
33074e5f | 4367 | { |
9ac617d4 EB |
4368 | /* If the function uses __builtin_eh_return, the eh_return |
4369 | machinery occupies the delay slot. */ | |
f5f7d171 | 4370 | gcc_assert (! final_sequence); |
ab835497 | 4371 | |
951661a1 EB |
4372 | if (! flag_delayed_branch) |
4373 | fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file); | |
4374 | ||
9ac617d4 EB |
4375 | if (TARGET_V9) |
4376 | fputs ("\treturn\t%i7+8\n", asm_out_file); | |
4377 | else | |
4378 | fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file); | |
4379 | ||
951661a1 EB |
4380 | if (flag_delayed_branch) |
4381 | fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file); | |
4382 | else | |
4383 | fputs ("\t nop\n", asm_out_file); | |
33074e5f | 4384 | } |
4e5b002b | 4385 | else if (final_sequence) |
ef8200df | 4386 | { |
9ac617d4 EB |
4387 | rtx delay, pat; |
4388 | ||
4389 | delay = NEXT_INSN (insn); | |
f5f7d171 | 4390 | gcc_assert (delay); |
1150a841 | 4391 | |
9ac617d4 EB |
4392 | pat = PATTERN (delay); |
4393 | ||
4394 | if (TARGET_V9 && ! epilogue_renumber (&pat, 1)) | |
33074e5f | 4395 | { |
9ac617d4 | 4396 | epilogue_renumber (&pat, 0); |
4e5b002b | 4397 | return "return\t%%i7+%)%#"; |
1150a841 | 4398 | } |
33074e5f | 4399 | else |
ab835497 | 4400 | { |
4e5b002b | 4401 | output_asm_insn ("jmp\t%%i7+%)", NULL); |
6244b14e | 4402 | output_restore (pat); |
4e5b002b EB |
4403 | PATTERN (delay) = gen_blockage (); |
4404 | INSN_CODE (delay) = -1; | |
ab835497 | 4405 | } |
ef8200df | 4406 | } |
33074e5f | 4407 | else |
9ac617d4 EB |
4408 | { |
4409 | /* The delay slot is empty. */ | |
4410 | if (TARGET_V9) | |
4e5b002b | 4411 | return "return\t%%i7+%)\n\t nop"; |
951661a1 | 4412 | else if (flag_delayed_branch) |
4e5b002b | 4413 | return "jmp\t%%i7+%)\n\t restore"; |
951661a1 | 4414 | else |
4e5b002b | 4415 | return "restore\n\tjmp\t%%o7+%)\n\t nop"; |
9ac617d4 | 4416 | } |
6a151f87 | 4417 | } |
e0d80184 | 4418 | |
9ac617d4 | 4419 | return ""; |
ab835497 | 4420 | } |
7d167afd JJ |
4421 | |
4422 | /* Output a sibling call. */ | |
4423 | ||
4424 | const char * | |
fc27d102 | 4425 | output_sibcall (rtx insn, rtx call_operand) |
7d167afd | 4426 | { |
9ac617d4 | 4427 | rtx operands[1]; |
7d167afd | 4428 | |
f5f7d171 | 4429 | gcc_assert (flag_delayed_branch); |
951661a1 | 4430 | |
7d167afd JJ |
4431 | operands[0] = call_operand; |
4432 | ||
5be9b7a1 | 4433 | if (sparc_leaf_function_p) |
7d167afd | 4434 | { |
9ac617d4 EB |
4435 | /* This is a leaf function so we don't have to bother restoring the |
4436 | register window. We simply output the jump to the function and | |
4437 | the insn in the delay slot (if any). */ | |
7d167afd | 4438 | |
f5f7d171 | 4439 | gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence)); |
9ac617d4 | 4440 | |
4e5b002b EB |
4441 | if (final_sequence) |
4442 | output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#", | |
4443 | operands); | |
7d167afd | 4444 | else |
4e5b002b EB |
4445 | /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize |
4446 | it into branch if possible. */ | |
4447 | output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7", | |
4448 | operands); | |
7d167afd | 4449 | } |
9ac617d4 | 4450 | else |
7d167afd | 4451 | { |
9ac617d4 EB |
4452 | /* This is a regular function so we have to restore the register window. |
4453 | We may have a pending insn for the delay slot, which will be combined | |
4454 | with the 'restore' instruction. */ | |
7d167afd | 4455 | |
9ac617d4 | 4456 | output_asm_insn ("call\t%a0, 0", operands); |
7d167afd | 4457 | |
4e5b002b | 4458 | if (final_sequence) |
7d167afd | 4459 | { |
9ac617d4 | 4460 | rtx delay = NEXT_INSN (insn); |
f5f7d171 | 4461 | gcc_assert (delay); |
9ac617d4 | 4462 | |
6244b14e | 4463 | output_restore (PATTERN (delay)); |
9ac617d4 EB |
4464 | |
4465 | PATTERN (delay) = gen_blockage (); | |
4466 | INSN_CODE (delay) = -1; | |
7d167afd | 4467 | } |
9ac617d4 | 4468 | else |
6244b14e | 4469 | output_restore (NULL_RTX); |
7d167afd | 4470 | } |
9ac617d4 | 4471 | |
7d167afd JJ |
4472 | return ""; |
4473 | } | |
4fb4e4b8 DE |
4474 | \f |
4475 | /* Functions for handling argument passing. | |
4476 | ||
1b5963f7 | 4477 | For 32-bit, the first 6 args are normally in registers and the rest are |
4fb4e4b8 DE |
4478 | pushed. Any arg that starts within the first 6 words is at least |
4479 | partially passed in a register unless its data type forbids. | |
4480 | ||
1b5963f7 | 4481 | For 64-bit, the argument registers are laid out as an array of 16 elements |
4fb4e4b8 DE |
4482 | and arguments are added sequentially. The first 6 int args and up to the |
4483 | first 16 fp args (depending on size) are passed in regs. | |
4484 | ||
4485 | Slot Stack Integral Float Float in structure Double Long Double | |
4486 | ---- ----- -------- ----- ------------------ ------ ----------- | |
4487 | 15 [SP+248] %f31 %f30,%f31 %d30 | |
4488 | 14 [SP+240] %f29 %f28,%f29 %d28 %q28 | |
4489 | 13 [SP+232] %f27 %f26,%f27 %d26 | |
4490 | 12 [SP+224] %f25 %f24,%f25 %d24 %q24 | |
4491 | 11 [SP+216] %f23 %f22,%f23 %d22 | |
4492 | 10 [SP+208] %f21 %f20,%f21 %d20 %q20 | |
4493 | 9 [SP+200] %f19 %f18,%f19 %d18 | |
4494 | 8 [SP+192] %f17 %f16,%f17 %d16 %q16 | |
4495 | 7 [SP+184] %f15 %f14,%f15 %d14 | |
4496 | 6 [SP+176] %f13 %f12,%f13 %d12 %q12 | |
4497 | 5 [SP+168] %o5 %f11 %f10,%f11 %d10 | |
4498 | 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8 | |
4499 | 3 [SP+152] %o3 %f7 %f6,%f7 %d6 | |
4500 | 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4 | |
4501 | 1 [SP+136] %o1 %f3 %f2,%f3 %d2 | |
4502 | 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0 | |
4503 | ||
4504 | Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise. | |
4505 | ||
1b5963f7 | 4506 | Integral arguments are always passed as 64-bit quantities appropriately |
4fb4e4b8 DE |
4507 | extended. |
4508 | ||
4509 | Passing of floating point values is handled as follows. | |
4510 | If a prototype is in scope: | |
4511 | If the value is in a named argument (i.e. not a stdarg function or a | |
4512 | value not part of the `...') then the value is passed in the appropriate | |
4513 | fp reg. | |
4514 | If the value is part of the `...' and is passed in one of the first 6 | |
4515 | slots then the value is passed in the appropriate int reg. | |
4516 | If the value is part of the `...' and is not passed in one of the first 6 | |
4517 | slots then the value is passed in memory. | |
4518 | If a prototype is not in scope: | |
4519 | If the value is one of the first 6 arguments the value is passed in the | |
4520 | appropriate integer reg and the appropriate fp reg. | |
4521 | If the value is not one of the first 6 arguments the value is passed in | |
4522 | the appropriate fp reg and in memory. | |
1b5963f7 EB |
4523 | |
4524 | ||
4900aa61 | 4525 | Summary of the calling conventions implemented by GCC on the SPARC: |
1b5963f7 EB |
4526 | |
4527 | 32-bit ABI: | |
4528 | size argument return value | |
4529 | ||
4530 | small integer <4 int. reg. int. reg. | |
4531 | word 4 int. reg. int. reg. | |
4532 | double word 8 int. reg. int. reg. | |
4533 | ||
4534 | _Complex small integer <8 int. reg. int. reg. | |
4535 | _Complex word 8 int. reg. int. reg. | |
4536 | _Complex double word 16 memory int. reg. | |
4537 | ||
4538 | vector integer <=8 int. reg. FP reg. | |
4539 | vector integer >8 memory memory | |
4540 | ||
4541 | float 4 int. reg. FP reg. | |
4542 | double 8 int. reg. FP reg. | |
4543 | long double 16 memory memory | |
4544 | ||
4545 | _Complex float 8 memory FP reg. | |
4546 | _Complex double 16 memory FP reg. | |
4547 | _Complex long double 32 memory FP reg. | |
4548 | ||
45261626 | 4549 | vector float any memory memory |
1b5963f7 EB |
4550 | |
4551 | aggregate any memory memory | |
4552 | ||
4553 | ||
4554 | ||
4555 | 64-bit ABI: | |
4556 | size argument return value | |
4557 | ||
4558 | small integer <8 int. reg. int. reg. | |
4559 | word 8 int. reg. int. reg. | |
4560 | double word 16 int. reg. int. reg. | |
4561 | ||
4562 | _Complex small integer <16 int. reg. int. reg. | |
4563 | _Complex word 16 int. reg. int. reg. | |
4564 | _Complex double word 32 memory int. reg. | |
4565 | ||
4566 | vector integer <=16 FP reg. FP reg. | |
4567 | vector integer 16<s<=32 memory FP reg. | |
4568 | vector integer >32 memory memory | |
4569 | ||
4570 | float 4 FP reg. FP reg. | |
4571 | double 8 FP reg. FP reg. | |
4572 | long double 16 FP reg. FP reg. | |
4573 | ||
4574 | _Complex float 8 FP reg. FP reg. | |
4575 | _Complex double 16 FP reg. FP reg. | |
4576 | _Complex long double 32 memory FP reg. | |
4577 | ||
4578 | vector float <=16 FP reg. FP reg. | |
4579 | vector float 16<s<=32 memory FP reg. | |
4580 | vector float >32 memory memory | |
4581 | ||
4582 | aggregate <=16 reg. reg. | |
4583 | aggregate 16<s<=32 memory reg. | |
4584 | aggregate >32 memory memory | |
4585 | ||
4586 | ||
4587 | ||
4588 | Note #1: complex floating-point types follow the extended SPARC ABIs as | |
4589 | implemented by the Sun compiler. | |
4590 | ||
4591 | Note #2: integral vector types follow the scalar floating-point types | |
4592 | conventions to match what is implemented by the Sun VIS SDK. | |
4593 | ||
4900aa61 | 4594 | Note #3: floating-point vector types follow the aggregate types |
45261626 | 4595 | conventions. */ |
1b5963f7 | 4596 | |
4fb4e4b8 DE |
4597 | |
4598 | /* Maximum number of int regs for args. */ | |
4599 | #define SPARC_INT_ARG_MAX 6 | |
4600 | /* Maximum number of fp regs for args. */ | |
4601 | #define SPARC_FP_ARG_MAX 16 | |
4602 | ||
4603 | #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD) | |
4604 | ||
4605 | /* Handle the INIT_CUMULATIVE_ARGS macro. | |
4606 | Initialize a variable CUM of type CUMULATIVE_ARGS | |
4607 | for a call to a function whose data type is FNTYPE. | |
4608 | For a library call, FNTYPE is 0. */ | |
3ea1fdd3 | 4609 | |
4fb4e4b8 | 4610 | void |
fc27d102 KG |
4611 | init_cumulative_args (struct sparc_args *cum, tree fntype, |
4612 | rtx libname ATTRIBUTE_UNUSED, | |
4613 | tree fndecl ATTRIBUTE_UNUSED) | |
4fb4e4b8 DE |
4614 | { |
4615 | cum->words = 0; | |
4616 | cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype); | |
4617 | cum->libcall_p = fntype == 0; | |
4618 | } | |
61a55e8b | 4619 | |
bd5bd7ac | 4620 | /* Handle the TARGET_PROMOTE_PROTOTYPES target hook. |
3e2cc1d1 KH |
4621 | When a prototype says `char' or `short', really pass an `int'. */ |
4622 | ||
4623 | static bool | |
586de218 | 4624 | sparc_promote_prototypes (const_tree fntype ATTRIBUTE_UNUSED) |
3e2cc1d1 KH |
4625 | { |
4626 | return TARGET_ARCH32 ? true : false; | |
4627 | } | |
4628 | ||
cde0f3fd PB |
4629 | /* Handle promotion of pointer and integer arguments. */ |
4630 | ||
4631 | static enum machine_mode | |
4632 | sparc_promote_function_mode (const_tree type ATTRIBUTE_UNUSED, | |
4633 | enum machine_mode mode, | |
4634 | int *punsignedp ATTRIBUTE_UNUSED, | |
4635 | const_tree fntype ATTRIBUTE_UNUSED, | |
4636 | int for_return ATTRIBUTE_UNUSED) | |
4637 | { | |
4638 | if (POINTER_TYPE_P (type)) | |
4639 | { | |
4640 | *punsignedp = POINTERS_EXTEND_UNSIGNED; | |
4641 | return Pmode; | |
4642 | } | |
4643 | ||
4644 | /* For TARGET_ARCH64 we need this, as we don't have instructions | |
4645 | for arithmetic operations which do zero/sign extension at the same time, | |
4646 | so without this we end up with a srl/sra after every assignment to an | |
4647 | user variable, which means very very bad code. */ | |
cde0f3fd PB |
4648 | if (TARGET_ARCH64 |
4649 | && GET_MODE_CLASS (mode) == MODE_INT | |
4650 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD) | |
4651 | return word_mode; | |
4652 | ||
4653 | return mode; | |
4654 | } | |
4655 | ||
bd5bd7ac | 4656 | /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */ |
3e2cc1d1 KH |
4657 | |
4658 | static bool | |
4659 | sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) | |
4660 | { | |
1b5963f7 | 4661 | return TARGET_ARCH64 ? true : false; |
3e2cc1d1 KH |
4662 | } |
4663 | ||
5c4f4b18 EB |
4664 | /* Scan the record type TYPE and return the following predicates: |
4665 | - INTREGS_P: the record contains at least one field or sub-field | |
4666 | that is eligible for promotion in integer registers. | |
4667 | - FP_REGS_P: the record contains at least one field or sub-field | |
4668 | that is eligible for promotion in floating-point registers. | |
4669 | - PACKED_P: the record contains at least one field that is packed. | |
4670 | ||
4671 | Sub-fields are not taken into account for the PACKED_P predicate. */ | |
4672 | ||
4673 | static void | |
a7bba6ca | 4674 | scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p) |
5c4f4b18 EB |
4675 | { |
4676 | tree field; | |
4677 | ||
4678 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
4679 | { | |
4680 | if (TREE_CODE (field) == FIELD_DECL) | |
4681 | { | |
4682 | if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE) | |
4683 | scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0); | |
45261626 EB |
4684 | else if ((FLOAT_TYPE_P (TREE_TYPE (field)) |
4685 | || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE) | |
4686 | && TARGET_FPU) | |
5c4f4b18 EB |
4687 | *fpregs_p = 1; |
4688 | else | |
4689 | *intregs_p = 1; | |
4690 | ||
4691 | if (packed_p && DECL_PACKED (field)) | |
4692 | *packed_p = 1; | |
4693 | } | |
4694 | } | |
4695 | } | |
4696 | ||
4fb4e4b8 | 4697 | /* Compute the slot number to pass an argument in. |
5c4f4b18 | 4698 | Return the slot number or -1 if passing on the stack. |
4fb4e4b8 DE |
4699 | |
4700 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
4701 | the preceding args and about the function being called. | |
4702 | MODE is the argument's machine mode. | |
4703 | TYPE is the data type of the argument (as a tree). | |
4704 | This is null for libcalls where that information may | |
4705 | not be available. | |
4706 | NAMED is nonzero if this argument is a named parameter | |
4707 | (otherwise it is an extra parameter matching an ellipsis). | |
4708 | INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. | |
4709 | *PREGNO records the register number to use if scalar type. | |
4710 | *PPADDING records the amount of padding needed in words. */ | |
4711 | ||
4712 | static int | |
fc27d102 KG |
4713 | function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode, |
4714 | tree type, int named, int incoming_p, | |
4715 | int *pregno, int *ppadding) | |
4fb4e4b8 DE |
4716 | { |
4717 | int regbase = (incoming_p | |
4718 | ? SPARC_INCOMING_INT_ARG_FIRST | |
4719 | : SPARC_OUTGOING_INT_ARG_FIRST); | |
4720 | int slotno = cum->words; | |
45261626 | 4721 | enum mode_class mclass; |
4fb4e4b8 DE |
4722 | int regno; |
4723 | ||
4724 | *ppadding = 0; | |
4725 | ||
1b5963f7 | 4726 | if (type && TREE_ADDRESSABLE (type)) |
4fb4e4b8 | 4727 | return -1; |
1b5963f7 | 4728 | |
4fb4e4b8 | 4729 | if (TARGET_ARCH32 |
1b5963f7 EB |
4730 | && mode == BLKmode |
4731 | && type | |
4fb4e4b8 DE |
4732 | && TYPE_ALIGN (type) % PARM_BOUNDARY != 0) |
4733 | return -1; | |
4734 | ||
1b5963f7 EB |
4735 | /* For SPARC64, objects requiring 16-byte alignment get it. */ |
4736 | if (TARGET_ARCH64 | |
7617ec02 | 4737 | && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128 |
1b5963f7 EB |
4738 | && (slotno & 1) != 0) |
4739 | slotno++, *ppadding = 1; | |
4fb4e4b8 | 4740 | |
45261626 EB |
4741 | mclass = GET_MODE_CLASS (mode); |
4742 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
4743 | { | |
4744 | /* Vector types deserve special treatment because they are | |
4745 | polymorphic wrt their mode, depending upon whether VIS | |
4746 | instructions are enabled. */ | |
4747 | if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE) | |
4748 | { | |
4749 | /* The SPARC port defines no floating-point vector modes. */ | |
f5f7d171 | 4750 | gcc_assert (mode == BLKmode); |
45261626 EB |
4751 | } |
4752 | else | |
4753 | { | |
4754 | /* Integral vector types should either have a vector | |
4755 | mode or an integral mode, because we are guaranteed | |
4756 | by pass_by_reference that their size is not greater | |
4757 | than 16 bytes and TImode is 16-byte wide. */ | |
f5f7d171 | 4758 | gcc_assert (mode != BLKmode); |
45261626 EB |
4759 | |
4760 | /* Vector integers are handled like floats according to | |
4761 | the Sun VIS SDK. */ | |
4762 | mclass = MODE_FLOAT; | |
4763 | } | |
4764 | } | |
4765 | ||
4766 | switch (mclass) | |
1b5963f7 EB |
4767 | { |
4768 | case MODE_FLOAT: | |
4769 | case MODE_COMPLEX_FLOAT: | |
0dd8c9bc | 4770 | case MODE_VECTOR_INT: |
1b5963f7 EB |
4771 | if (TARGET_ARCH64 && TARGET_FPU && named) |
4772 | { | |
4773 | if (slotno >= SPARC_FP_ARG_MAX) | |
4774 | return -1; | |
4775 | regno = SPARC_FP_ARG_FIRST + slotno * 2; | |
4776 | /* Arguments filling only one single FP register are | |
4777 | right-justified in the outer double FP register. */ | |
4778 | if (GET_MODE_SIZE (mode) <= 4) | |
4779 | regno++; | |
4780 | break; | |
4781 | } | |
d560b0aa EB |
4782 | /* fallthrough */ |
4783 | ||
1b5963f7 EB |
4784 | case MODE_INT: |
4785 | case MODE_COMPLEX_INT: | |
4fb4e4b8 DE |
4786 | if (slotno >= SPARC_INT_ARG_MAX) |
4787 | return -1; | |
4788 | regno = regbase + slotno; | |
4789 | break; | |
4790 | ||
1b5963f7 EB |
4791 | case MODE_RANDOM: |
4792 | if (mode == VOIDmode) | |
4793 | /* MODE is VOIDmode when generating the actual call. */ | |
4794 | return -1; | |
d560b0aa | 4795 | |
f5f7d171 | 4796 | gcc_assert (mode == BLKmode); |
4fb4e4b8 | 4797 | |
ef6843ed EB |
4798 | if (TARGET_ARCH32 |
4799 | || !type | |
4800 | || (TREE_CODE (type) != VECTOR_TYPE | |
4801 | && TREE_CODE (type) != RECORD_TYPE)) | |
4fb4e4b8 DE |
4802 | { |
4803 | if (slotno >= SPARC_INT_ARG_MAX) | |
4804 | return -1; | |
4805 | regno = regbase + slotno; | |
4806 | } | |
45261626 | 4807 | else /* TARGET_ARCH64 && type */ |
4fb4e4b8 | 4808 | { |
5c4f4b18 EB |
4809 | int intregs_p = 0, fpregs_p = 0, packed_p = 0; |
4810 | ||
4811 | /* First see what kinds of registers we would need. */ | |
45261626 EB |
4812 | if (TREE_CODE (type) == VECTOR_TYPE) |
4813 | fpregs_p = 1; | |
4814 | else | |
4815 | scan_record_type (type, &intregs_p, &fpregs_p, &packed_p); | |
5c4f4b18 EB |
4816 | |
4817 | /* The ABI obviously doesn't specify how packed structures | |
4818 | are passed. These are defined to be passed in int regs | |
4819 | if possible, otherwise memory. */ | |
4fb4e4b8 DE |
4820 | if (packed_p || !named) |
4821 | fpregs_p = 0, intregs_p = 1; | |
4822 | ||
4823 | /* If all arg slots are filled, then must pass on stack. */ | |
4824 | if (fpregs_p && slotno >= SPARC_FP_ARG_MAX) | |
4825 | return -1; | |
1b5963f7 | 4826 | |
4fb4e4b8 DE |
4827 | /* If there are only int args and all int arg slots are filled, |
4828 | then must pass on stack. */ | |
4829 | if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX) | |
4830 | return -1; | |
1b5963f7 | 4831 | |
4fb4e4b8 DE |
4832 | /* Note that even if all int arg slots are filled, fp members may |
4833 | still be passed in regs if such regs are available. | |
4834 | *PREGNO isn't set because there may be more than one, it's up | |
4835 | to the caller to compute them. */ | |
4836 | return slotno; | |
4837 | } | |
4838 | break; | |
4839 | ||
4840 | default : | |
f5f7d171 | 4841 | gcc_unreachable (); |
4fb4e4b8 DE |
4842 | } |
4843 | ||
4844 | *pregno = regno; | |
4845 | return slotno; | |
4846 | } | |
4847 | ||
82d6b402 RH |
4848 | /* Handle recursive register counting for structure field layout. */ |
4849 | ||
4850 | struct function_arg_record_value_parms | |
4851 | { | |
48028e21 EB |
4852 | rtx ret; /* return expression being built. */ |
4853 | int slotno; /* slot number of the argument. */ | |
4854 | int named; /* whether the argument is named. */ | |
4855 | int regbase; /* regno of the base register. */ | |
4856 | int stack; /* 1 if part of the argument is on the stack. */ | |
1fd05073 | 4857 | int intoffset; /* offset of the first pending integer field. */ |
48028e21 | 4858 | unsigned int nregs; /* number of words passed in registers. */ |
82d6b402 RH |
4859 | }; |
4860 | ||
2a01c939 | 4861 | static void function_arg_record_value_3 |
e80d5f80 | 4862 | (HOST_WIDE_INT, struct function_arg_record_value_parms *); |
2a01c939 | 4863 | static void function_arg_record_value_2 |
586de218 | 4864 | (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool); |
b1474bb7 | 4865 | static void function_arg_record_value_1 |
586de218 KG |
4866 | (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool); |
4867 | static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int); | |
22d8d627 | 4868 | static rtx function_arg_union_value (int, enum machine_mode, int, int); |
2a01c939 | 4869 | |
0020b823 | 4870 | /* A subroutine of function_arg_record_value. Traverse the structure |
fae778eb | 4871 | recursively and determine how many registers will be required. */ |
0020b823 | 4872 | |
82d6b402 | 4873 | static void |
586de218 | 4874 | function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos, |
cde85594 EB |
4875 | struct function_arg_record_value_parms *parms, |
4876 | bool packed_p) | |
82d6b402 RH |
4877 | { |
4878 | tree field; | |
4879 | ||
82d6b402 RH |
4880 | /* We need to compute how many registers are needed so we can |
4881 | allocate the PARALLEL but before we can do that we need to know | |
cde85594 EB |
4882 | whether there are any packed fields. The ABI obviously doesn't |
4883 | specify how structures are passed in this case, so they are | |
4884 | defined to be passed in int regs if possible, otherwise memory, | |
4885 | regardless of whether there are fp values present. */ | |
4886 | ||
4887 | if (! packed_p) | |
4888 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
4889 | { | |
4890 | if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field)) | |
4891 | { | |
4892 | packed_p = true; | |
4893 | break; | |
4894 | } | |
4895 | } | |
82d6b402 RH |
4896 | |
4897 | /* Compute how many registers we need. */ | |
4898 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
4899 | { | |
4900 | if (TREE_CODE (field) == FIELD_DECL) | |
4901 | { | |
75131237 RK |
4902 | HOST_WIDE_INT bitpos = startbitpos; |
4903 | ||
bb05e341 EB |
4904 | if (DECL_SIZE (field) != 0) |
4905 | { | |
4906 | if (integer_zerop (DECL_SIZE (field))) | |
4907 | continue; | |
4908 | ||
4909 | if (host_integerp (bit_position (field), 1)) | |
4910 | bitpos += int_bit_position (field); | |
4911 | } | |
75131237 | 4912 | |
82d6b402 RH |
4913 | /* ??? FIXME: else assume zero offset. */ |
4914 | ||
4915 | if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE) | |
cde85594 EB |
4916 | function_arg_record_value_1 (TREE_TYPE (field), |
4917 | bitpos, | |
4918 | parms, | |
4919 | packed_p); | |
1b5963f7 EB |
4920 | else if ((FLOAT_TYPE_P (TREE_TYPE (field)) |
4921 | || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE) | |
8a98812c EB |
4922 | && TARGET_FPU |
4923 | && parms->named | |
4924 | && ! packed_p) | |
82d6b402 RH |
4925 | { |
4926 | if (parms->intoffset != -1) | |
4927 | { | |
b2f24c85 | 4928 | unsigned int startbit, endbit; |
82d6b402 RH |
4929 | int intslots, this_slotno; |
4930 | ||
b2f24c85 EB |
4931 | startbit = parms->intoffset & -BITS_PER_WORD; |
4932 | endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD; | |
4933 | ||
4934 | intslots = (endbit - startbit) / BITS_PER_WORD; | |
82d6b402 RH |
4935 | this_slotno = parms->slotno + parms->intoffset |
4936 | / BITS_PER_WORD; | |
4937 | ||
48028e21 EB |
4938 | if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno) |
4939 | { | |
4940 | intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno); | |
4941 | /* We need to pass this field on the stack. */ | |
4942 | parms->stack = 1; | |
4943 | } | |
4944 | ||
82d6b402 RH |
4945 | parms->nregs += intslots; |
4946 | parms->intoffset = -1; | |
4947 | } | |
4948 | ||
4949 | /* There's no need to check this_slotno < SPARC_FP_ARG MAX. | |
4950 | If it wasn't true we wouldn't be here. */ | |
45261626 EB |
4951 | if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE |
4952 | && DECL_MODE (field) == BLKmode) | |
4953 | parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field)); | |
4954 | else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE) | |
4955 | parms->nregs += 2; | |
4956 | else | |
105b2084 | 4957 | parms->nregs += 1; |
82d6b402 RH |
4958 | } |
4959 | else | |
4960 | { | |
4961 | if (parms->intoffset == -1) | |
4962 | parms->intoffset = bitpos; | |
4963 | } | |
4964 | } | |
4965 | } | |
4966 | } | |
4967 | ||
0020b823 RH |
4968 | /* A subroutine of function_arg_record_value. Assign the bits of the |
4969 | structure between parms->intoffset and bitpos to integer registers. */ | |
82d6b402 | 4970 | |
4900aa61 | 4971 | static void |
fc27d102 KG |
4972 | function_arg_record_value_3 (HOST_WIDE_INT bitpos, |
4973 | struct function_arg_record_value_parms *parms) | |
82d6b402 RH |
4974 | { |
4975 | enum machine_mode mode; | |
75131237 | 4976 | unsigned int regno; |
0020b823 | 4977 | unsigned int startbit, endbit; |
75131237 | 4978 | int this_slotno, intslots, intoffset; |
82d6b402 RH |
4979 | rtx reg; |
4980 | ||
4981 | if (parms->intoffset == -1) | |
4982 | return; | |
75131237 | 4983 | |
82d6b402 RH |
4984 | intoffset = parms->intoffset; |
4985 | parms->intoffset = -1; | |
4986 | ||
0020b823 RH |
4987 | startbit = intoffset & -BITS_PER_WORD; |
4988 | endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD; | |
4989 | intslots = (endbit - startbit) / BITS_PER_WORD; | |
82d6b402 RH |
4990 | this_slotno = parms->slotno + intoffset / BITS_PER_WORD; |
4991 | ||
4992 | intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno); | |
4993 | if (intslots <= 0) | |
4994 | return; | |
4995 | ||
4996 | /* If this is the trailing part of a word, only load that much into | |
4997 | the register. Otherwise load the whole register. Note that in | |
4998 | the latter case we may pick up unwanted bits. It's not a problem | |
4999 | at the moment but may wish to revisit. */ | |
5000 | ||
5001 | if (intoffset % BITS_PER_WORD != 0) | |
bb05e341 EB |
5002 | mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD, |
5003 | MODE_INT); | |
82d6b402 RH |
5004 | else |
5005 | mode = word_mode; | |
5006 | ||
5007 | intoffset /= BITS_PER_UNIT; | |
5008 | do | |
5009 | { | |
5010 | regno = parms->regbase + this_slotno; | |
254110c2 | 5011 | reg = gen_rtx_REG (mode, regno); |
48028e21 | 5012 | XVECEXP (parms->ret, 0, parms->stack + parms->nregs) |
254110c2 | 5013 | = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset)); |
82d6b402 RH |
5014 | |
5015 | this_slotno += 1; | |
5016 | intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1; | |
1fd05073 | 5017 | mode = word_mode; |
82d6b402 RH |
5018 | parms->nregs += 1; |
5019 | intslots -= 1; | |
5020 | } | |
5021 | while (intslots > 0); | |
5022 | } | |
5023 | ||
0020b823 RH |
5024 | /* A subroutine of function_arg_record_value. Traverse the structure |
5025 | recursively and assign bits to floating point registers. Track which | |
5026 | bits in between need integer registers; invoke function_arg_record_value_3 | |
5027 | to make that happen. */ | |
5028 | ||
82d6b402 | 5029 | static void |
586de218 | 5030 | function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos, |
cde85594 EB |
5031 | struct function_arg_record_value_parms *parms, |
5032 | bool packed_p) | |
82d6b402 RH |
5033 | { |
5034 | tree field; | |
82d6b402 | 5035 | |
cde85594 EB |
5036 | if (! packed_p) |
5037 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
5038 | { | |
5039 | if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field)) | |
5040 | { | |
5041 | packed_p = true; | |
5042 | break; | |
5043 | } | |
5044 | } | |
82d6b402 RH |
5045 | |
5046 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
5047 | { | |
5048 | if (TREE_CODE (field) == FIELD_DECL) | |
5049 | { | |
75131237 RK |
5050 | HOST_WIDE_INT bitpos = startbitpos; |
5051 | ||
bb05e341 EB |
5052 | if (DECL_SIZE (field) != 0) |
5053 | { | |
5054 | if (integer_zerop (DECL_SIZE (field))) | |
5055 | continue; | |
5056 | ||
5057 | if (host_integerp (bit_position (field), 1)) | |
5058 | bitpos += int_bit_position (field); | |
5059 | } | |
75131237 | 5060 | |
82d6b402 RH |
5061 | /* ??? FIXME: else assume zero offset. */ |
5062 | ||
5063 | if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE) | |
cde85594 EB |
5064 | function_arg_record_value_2 (TREE_TYPE (field), |
5065 | bitpos, | |
5066 | parms, | |
5067 | packed_p); | |
1b5963f7 EB |
5068 | else if ((FLOAT_TYPE_P (TREE_TYPE (field)) |
5069 | || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE) | |
8a98812c EB |
5070 | && TARGET_FPU |
5071 | && parms->named | |
5072 | && ! packed_p) | |
82d6b402 RH |
5073 | { |
5074 | int this_slotno = parms->slotno + bitpos / BITS_PER_WORD; | |
45261626 | 5075 | int regno, nregs, pos; |
105b2084 | 5076 | enum machine_mode mode = DECL_MODE (field); |
82d6b402 RH |
5077 | rtx reg; |
5078 | ||
5079 | function_arg_record_value_3 (bitpos, parms); | |
45261626 EB |
5080 | |
5081 | if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE | |
5082 | && mode == BLKmode) | |
5083 | { | |
5084 | mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field))); | |
5085 | nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field)); | |
105b2084 | 5086 | } |
45261626 EB |
5087 | else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE) |
5088 | { | |
5089 | mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field))); | |
5090 | nregs = 2; | |
5091 | } | |
5092 | else | |
5093 | nregs = 1; | |
5094 | ||
1b5963f7 EB |
5095 | regno = SPARC_FP_ARG_FIRST + this_slotno * 2; |
5096 | if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0) | |
5097 | regno++; | |
105b2084 | 5098 | reg = gen_rtx_REG (mode, regno); |
45261626 | 5099 | pos = bitpos / BITS_PER_UNIT; |
48028e21 | 5100 | XVECEXP (parms->ret, 0, parms->stack + parms->nregs) |
45261626 | 5101 | = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos)); |
82d6b402 | 5102 | parms->nregs += 1; |
45261626 | 5103 | while (--nregs > 0) |
105b2084 JJ |
5104 | { |
5105 | regno += GET_MODE_SIZE (mode) / 4; | |
5106 | reg = gen_rtx_REG (mode, regno); | |
45261626 | 5107 | pos += GET_MODE_SIZE (mode); |
48028e21 | 5108 | XVECEXP (parms->ret, 0, parms->stack + parms->nregs) |
45261626 | 5109 | = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos)); |
105b2084 JJ |
5110 | parms->nregs += 1; |
5111 | } | |
82d6b402 RH |
5112 | } |
5113 | else | |
5114 | { | |
5115 | if (parms->intoffset == -1) | |
5116 | parms->intoffset = bitpos; | |
5117 | } | |
5118 | } | |
5119 | } | |
5120 | } | |
5121 | ||
0020b823 | 5122 | /* Used by function_arg and function_value to implement the complex |
48028e21 EB |
5123 | conventions of the 64-bit ABI for passing and returning structures. |
5124 | Return an expression valid as a return value for the two macros | |
5125 | FUNCTION_ARG and FUNCTION_VALUE. | |
0020b823 | 5126 | |
48028e21 EB |
5127 | TYPE is the data type of the argument (as a tree). |
5128 | This is null for libcalls where that information may | |
5129 | not be available. | |
5130 | MODE is the argument's machine mode. | |
5131 | SLOTNO is the index number of the argument's slot in the parameter array. | |
5132 | NAMED is nonzero if this argument is a named parameter | |
5133 | (otherwise it is an extra parameter matching an ellipsis). | |
5134 | REGBASE is the regno of the base register for the parameter array. */ | |
4900aa61 | 5135 | |
82d6b402 | 5136 | static rtx |
586de218 | 5137 | function_arg_record_value (const_tree type, enum machine_mode mode, |
fc27d102 | 5138 | int slotno, int named, int regbase) |
82d6b402 RH |
5139 | { |
5140 | HOST_WIDE_INT typesize = int_size_in_bytes (type); | |
5141 | struct function_arg_record_value_parms parms; | |
75131237 | 5142 | unsigned int nregs; |
82d6b402 RH |
5143 | |
5144 | parms.ret = NULL_RTX; | |
5145 | parms.slotno = slotno; | |
5146 | parms.named = named; | |
5147 | parms.regbase = regbase; | |
48028e21 | 5148 | parms.stack = 0; |
82d6b402 RH |
5149 | |
5150 | /* Compute how many registers we need. */ | |
5151 | parms.nregs = 0; | |
5152 | parms.intoffset = 0; | |
cde85594 | 5153 | function_arg_record_value_1 (type, 0, &parms, false); |
82d6b402 | 5154 | |
1fd05073 | 5155 | /* Take into account pending integer fields. */ |
82d6b402 RH |
5156 | if (parms.intoffset != -1) |
5157 | { | |
0020b823 | 5158 | unsigned int startbit, endbit; |
82d6b402 RH |
5159 | int intslots, this_slotno; |
5160 | ||
0020b823 RH |
5161 | startbit = parms.intoffset & -BITS_PER_WORD; |
5162 | endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD; | |
5163 | intslots = (endbit - startbit) / BITS_PER_WORD; | |
82d6b402 RH |
5164 | this_slotno = slotno + parms.intoffset / BITS_PER_WORD; |
5165 | ||
48028e21 EB |
5166 | if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno) |
5167 | { | |
5168 | intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno); | |
5169 | /* We need to pass this field on the stack. */ | |
5170 | parms.stack = 1; | |
5171 | } | |
82d6b402 RH |
5172 | |
5173 | parms.nregs += intslots; | |
5174 | } | |
5175 | nregs = parms.nregs; | |
5176 | ||
5177 | /* Allocate the vector and handle some annoying special cases. */ | |
5178 | if (nregs == 0) | |
5179 | { | |
5180 | /* ??? Empty structure has no value? Duh? */ | |
5181 | if (typesize <= 0) | |
5182 | { | |
5183 | /* Though there's nothing really to store, return a word register | |
5184 | anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL | |
5185 | leads to breakage due to the fact that there are zero bytes to | |
5186 | load. */ | |
1eac9f59 | 5187 | return gen_rtx_REG (mode, regbase); |
82d6b402 RH |
5188 | } |
5189 | else | |
5190 | { | |
5191 | /* ??? C++ has structures with no fields, and yet a size. Give up | |
5192 | for now and pass everything back in integer registers. */ | |
5193 | nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
5194 | } | |
5195 | if (nregs + slotno > SPARC_INT_ARG_MAX) | |
5196 | nregs = SPARC_INT_ARG_MAX - slotno; | |
5197 | } | |
f5f7d171 | 5198 | gcc_assert (nregs != 0); |
82d6b402 | 5199 | |
48028e21 EB |
5200 | parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs)); |
5201 | ||
5202 | /* If at least one field must be passed on the stack, generate | |
5203 | (parallel [(expr_list (nil) ...) ...]) so that all fields will | |
5204 | also be passed on the stack. We can't do much better because the | |
78a52f11 | 5205 | semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case |
48028e21 EB |
5206 | of structures for which the fields passed exclusively in registers |
5207 | are not at the beginning of the structure. */ | |
5208 | if (parms.stack) | |
5209 | XVECEXP (parms.ret, 0, 0) | |
5210 | = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx); | |
82d6b402 RH |
5211 | |
5212 | /* Fill in the entries. */ | |
5213 | parms.nregs = 0; | |
5214 | parms.intoffset = 0; | |
cde85594 | 5215 | function_arg_record_value_2 (type, 0, &parms, false); |
82d6b402 RH |
5216 | function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms); |
5217 | ||
f5f7d171 | 5218 | gcc_assert (parms.nregs == nregs); |
82d6b402 RH |
5219 | |
5220 | return parms.ret; | |
5221 | } | |
5222 | ||
a7bba6ca EB |
5223 | /* Used by function_arg and function_value to implement the conventions |
5224 | of the 64-bit ABI for passing and returning unions. | |
5225 | Return an expression valid as a return value for the two macros | |
5226 | FUNCTION_ARG and FUNCTION_VALUE. | |
5227 | ||
5228 | SIZE is the size in bytes of the union. | |
85bbb21f | 5229 | MODE is the argument's machine mode. |
a7bba6ca EB |
5230 | REGNO is the hard register the union will be passed in. */ |
5231 | ||
5232 | static rtx | |
22d8d627 EB |
5233 | function_arg_union_value (int size, enum machine_mode mode, int slotno, |
5234 | int regno) | |
a7bba6ca | 5235 | { |
85bbb21f EB |
5236 | int nwords = ROUND_ADVANCE (size), i; |
5237 | rtx regs; | |
a7bba6ca | 5238 | |
bb05e341 EB |
5239 | /* See comment in previous function for empty structures. */ |
5240 | if (nwords == 0) | |
5241 | return gen_rtx_REG (mode, regno); | |
5242 | ||
22d8d627 EB |
5243 | if (slotno == SPARC_INT_ARG_MAX - 1) |
5244 | nwords = 1; | |
5245 | ||
85bbb21f | 5246 | regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords)); |
a7bba6ca | 5247 | |
85bbb21f | 5248 | for (i = 0; i < nwords; i++) |
45261626 EB |
5249 | { |
5250 | /* Unions are passed left-justified. */ | |
5251 | XVECEXP (regs, 0, i) | |
5252 | = gen_rtx_EXPR_LIST (VOIDmode, | |
5253 | gen_rtx_REG (word_mode, regno), | |
5254 | GEN_INT (UNITS_PER_WORD * i)); | |
5255 | regno++; | |
5256 | } | |
5257 | ||
5258 | return regs; | |
5259 | } | |
5260 | ||
5261 | /* Used by function_arg and function_value to implement the conventions | |
5262 | for passing and returning large (BLKmode) vectors. | |
5263 | Return an expression valid as a return value for the two macros | |
5264 | FUNCTION_ARG and FUNCTION_VALUE. | |
5265 | ||
26237a50 | 5266 | SIZE is the size in bytes of the vector (at least 8 bytes). |
45261626 EB |
5267 | REGNO is the FP hard register the vector will be passed in. */ |
5268 | ||
5269 | static rtx | |
26237a50 | 5270 | function_arg_vector_value (int size, int regno) |
45261626 | 5271 | { |
26237a50 | 5272 | int i, nregs = size / 8; |
45261626 EB |
5273 | rtx regs; |
5274 | ||
5275 | regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs)); | |
5276 | ||
5277 | for (i = 0; i < nregs; i++) | |
5278 | { | |
5279 | XVECEXP (regs, 0, i) | |
5280 | = gen_rtx_EXPR_LIST (VOIDmode, | |
26237a50 EB |
5281 | gen_rtx_REG (DImode, regno + 2*i), |
5282 | GEN_INT (i*8)); | |
45261626 | 5283 | } |
a7bba6ca | 5284 | |
85bbb21f | 5285 | return regs; |
a7bba6ca EB |
5286 | } |
5287 | ||
4fb4e4b8 DE |
5288 | /* Handle the FUNCTION_ARG macro. |
5289 | Determine where to put an argument to a function. | |
5290 | Value is zero to push the argument on the stack, | |
5291 | or a hard register in which to store the argument. | |
5292 | ||
5293 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
5294 | the preceding args and about the function being called. | |
5295 | MODE is the argument's machine mode. | |
5296 | TYPE is the data type of the argument (as a tree). | |
5297 | This is null for libcalls where that information may | |
5298 | not be available. | |
5299 | NAMED is nonzero if this argument is a named parameter | |
5300 | (otherwise it is an extra parameter matching an ellipsis). | |
5301 | INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */ | |
61a55e8b DE |
5302 | |
5303 | rtx | |
fc27d102 KG |
5304 | function_arg (const struct sparc_args *cum, enum machine_mode mode, |
5305 | tree type, int named, int incoming_p) | |
61a55e8b | 5306 | { |
4fb4e4b8 DE |
5307 | int regbase = (incoming_p |
5308 | ? SPARC_INCOMING_INT_ARG_FIRST | |
5309 | : SPARC_OUTGOING_INT_ARG_FIRST); | |
5310 | int slotno, regno, padding; | |
45261626 | 5311 | enum mode_class mclass = GET_MODE_CLASS (mode); |
4fb4e4b8 DE |
5312 | |
5313 | slotno = function_arg_slotno (cum, mode, type, named, incoming_p, | |
5314 | ®no, &padding); | |
4fb4e4b8 DE |
5315 | if (slotno == -1) |
5316 | return 0; | |
5317 | ||
ef6843ed EB |
5318 | /* Vector types deserve special treatment because they are polymorphic wrt |
5319 | their mode, depending upon whether VIS instructions are enabled. */ | |
5320 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
4fb4e4b8 | 5321 | { |
ef6843ed EB |
5322 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5323 | gcc_assert ((TARGET_ARCH32 && size <= 8) | |
5324 | || (TARGET_ARCH64 && size <= 16)); | |
5325 | ||
5326 | if (mode == BLKmode) | |
5327 | return function_arg_vector_value (size, | |
ef6843ed EB |
5328 | SPARC_FP_ARG_FIRST + 2*slotno); |
5329 | else | |
5330 | mclass = MODE_FLOAT; | |
4fb4e4b8 | 5331 | } |
4043d6fb | 5332 | |
ef6843ed EB |
5333 | if (TARGET_ARCH32) |
5334 | return gen_rtx_REG (mode, regno); | |
4043d6fb | 5335 | |
ef6843ed EB |
5336 | /* Structures up to 16 bytes in size are passed in arg slots on the stack |
5337 | and are promoted to registers if possible. */ | |
5338 | if (type && TREE_CODE (type) == RECORD_TYPE) | |
4043d6fb | 5339 | { |
a7bba6ca | 5340 | HOST_WIDE_INT size = int_size_in_bytes (type); |
f5f7d171 | 5341 | gcc_assert (size <= 16); |
4fb4e4b8 | 5342 | |
ef6843ed | 5343 | return function_arg_record_value (type, mode, slotno, named, regbase); |
4043d6fb | 5344 | } |
ef6843ed EB |
5345 | |
5346 | /* Unions up to 16 bytes in size are passed in integer registers. */ | |
5347 | else if (type && TREE_CODE (type) == UNION_TYPE) | |
45261626 | 5348 | { |
45261626 | 5349 | HOST_WIDE_INT size = int_size_in_bytes (type); |
f5f7d171 | 5350 | gcc_assert (size <= 16); |
45261626 | 5351 | |
ef6843ed | 5352 | return function_arg_union_value (size, mode, slotno, regno); |
45261626 EB |
5353 | } |
5354 | ||
4fb4e4b8 DE |
5355 | /* v9 fp args in reg slots beyond the int reg slots get passed in regs |
5356 | but also have the slot allocated for them. | |
5357 | If no prototype is in scope fp values in register slots get passed | |
5358 | in two places, either fp regs and int regs or fp regs and memory. */ | |
ef6843ed EB |
5359 | else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) |
5360 | && SPARC_FP_REG_P (regno)) | |
4fb4e4b8 | 5361 | { |
ef6843ed | 5362 | rtx reg = gen_rtx_REG (mode, regno); |
4fb4e4b8 DE |
5363 | if (cum->prototype_p || cum->libcall_p) |
5364 | { | |
5365 | /* "* 2" because fp reg numbers are recorded in 4 byte | |
5366 | quantities. */ | |
82d6b402 | 5367 | #if 0 |
4fb4e4b8 DE |
5368 | /* ??? This will cause the value to be passed in the fp reg and |
5369 | in the stack. When a prototype exists we want to pass the | |
5370 | value in the reg but reserve space on the stack. That's an | |
956d6950 | 5371 | optimization, and is deferred [for a bit]. */ |
4fb4e4b8 | 5372 | if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2) |
254110c2 | 5373 | return gen_rtx_PARALLEL (mode, |
4fb4e4b8 | 5374 | gen_rtvec (2, |
254110c2 | 5375 | gen_rtx_EXPR_LIST (VOIDmode, |
4fb4e4b8 | 5376 | NULL_RTX, const0_rtx), |
254110c2 | 5377 | gen_rtx_EXPR_LIST (VOIDmode, |
4fb4e4b8 DE |
5378 | reg, const0_rtx))); |
5379 | else | |
82d6b402 RH |
5380 | #else |
5381 | /* ??? It seems that passing back a register even when past | |
5382 | the area declared by REG_PARM_STACK_SPACE will allocate | |
5383 | space appropriately, and will not copy the data onto the | |
5384 | stack, exactly as we desire. | |
5385 | ||
5386 | This is due to locate_and_pad_parm being called in | |
5387 | expand_call whenever reg_parm_stack_space > 0, which | |
fae778eb | 5388 | while beneficial to our example here, would seem to be |
82d6b402 RH |
5389 | in error from what had been intended. Ho hum... -- r~ */ |
5390 | #endif | |
4fb4e4b8 DE |
5391 | return reg; |
5392 | } | |
5393 | else | |
5394 | { | |
82d6b402 RH |
5395 | rtx v0, v1; |
5396 | ||
4fb4e4b8 DE |
5397 | if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2) |
5398 | { | |
82d6b402 RH |
5399 | int intreg; |
5400 | ||
5401 | /* On incoming, we don't need to know that the value | |
5402 | is passed in %f0 and %i0, and it confuses other parts | |
5403 | causing needless spillage even on the simplest cases. */ | |
5404 | if (incoming_p) | |
5405 | return reg; | |
5406 | ||
5407 | intreg = (SPARC_OUTGOING_INT_ARG_FIRST | |
5408 | + (regno - SPARC_FP_ARG_FIRST) / 2); | |
5409 | ||
5410 | v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx); | |
5411 | v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg), | |
5412 | const0_rtx); | |
5413 | return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1)); | |
4fb4e4b8 DE |
5414 | } |
5415 | else | |
82d6b402 RH |
5416 | { |
5417 | v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx); | |
5418 | v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx); | |
5419 | return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1)); | |
5420 | } | |
4fb4e4b8 DE |
5421 | } |
5422 | } | |
ef6843ed EB |
5423 | |
5424 | /* All other aggregate types are passed in an integer register in a mode | |
5425 | corresponding to the size of the type. */ | |
5426 | else if (type && AGGREGATE_TYPE_P (type)) | |
4fb4e4b8 | 5427 | { |
ef6843ed EB |
5428 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5429 | gcc_assert (size <= 16); | |
5430 | ||
5431 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
4fb4e4b8 DE |
5432 | } |
5433 | ||
ef6843ed | 5434 | return gen_rtx_REG (mode, regno); |
4fb4e4b8 | 5435 | } |
a7acd911 | 5436 | |
78a52f11 RH |
5437 | /* For an arg passed partly in registers and partly in memory, |
5438 | this is the number of bytes of registers used. | |
4fb4e4b8 | 5439 | For args passed entirely in registers or entirely in memory, zero. |
61a55e8b | 5440 | |
4fb4e4b8 DE |
5441 | Any arg that starts in the first 6 regs but won't entirely fit in them |
5442 | needs partial registers on v8. On v9, structures with integer | |
5443 | values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp | |
5444 | values that begin in the last fp reg [where "last fp reg" varies with the | |
5445 | mode] will be split between that reg and memory. */ | |
61a55e8b | 5446 | |
78a52f11 RH |
5447 | static int |
5448 | sparc_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
5449 | tree type, bool named) | |
4fb4e4b8 DE |
5450 | { |
5451 | int slotno, regno, padding; | |
61a55e8b | 5452 | |
4fb4e4b8 DE |
5453 | /* We pass 0 for incoming_p here, it doesn't matter. */ |
5454 | slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding); | |
5455 | ||
5456 | if (slotno == -1) | |
5457 | return 0; | |
5458 | ||
5459 | if (TARGET_ARCH32) | |
bf62bbf1 | 5460 | { |
4fb4e4b8 DE |
5461 | if ((slotno + (mode == BLKmode |
5462 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
5463 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))) | |
1b5963f7 | 5464 | > SPARC_INT_ARG_MAX) |
78a52f11 | 5465 | return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD; |
4fb4e4b8 DE |
5466 | } |
5467 | else | |
5468 | { | |
8cd5a4e0 | 5469 | /* We are guaranteed by pass_by_reference that the size of the |
78a52f11 RH |
5470 | argument is not greater than 16 bytes, so we only need to return |
5471 | one word if the argument is partially passed in registers. */ | |
1b5963f7 | 5472 | |
4fb4e4b8 DE |
5473 | if (type && AGGREGATE_TYPE_P (type)) |
5474 | { | |
5475 | int size = int_size_in_bytes (type); | |
bf62bbf1 | 5476 | |
1b5963f7 | 5477 | if (size > UNITS_PER_WORD |
4fb4e4b8 | 5478 | && slotno == SPARC_INT_ARG_MAX - 1) |
78a52f11 | 5479 | return UNITS_PER_WORD; |
4fb4e4b8 | 5480 | } |
0a9e65f9 | 5481 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT |
4fb4e4b8 | 5482 | || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT |
c3427c7d | 5483 | && ! (TARGET_FPU && named))) |
4fb4e4b8 | 5484 | { |
0a9e65f9 | 5485 | /* The complex types are passed as packed types. */ |
1b5963f7 EB |
5486 | if (GET_MODE_SIZE (mode) > UNITS_PER_WORD |
5487 | && slotno == SPARC_INT_ARG_MAX - 1) | |
78a52f11 | 5488 | return UNITS_PER_WORD; |
4fb4e4b8 DE |
5489 | } |
5490 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
5491 | { | |
4fb4e4b8 DE |
5492 | if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD) |
5493 | > SPARC_FP_ARG_MAX) | |
78a52f11 | 5494 | return UNITS_PER_WORD; |
4fb4e4b8 | 5495 | } |
bf62bbf1 | 5496 | } |
1b5963f7 EB |
5497 | |
5498 | return 0; | |
4fb4e4b8 | 5499 | } |
61a55e8b | 5500 | |
45261626 EB |
5501 | /* Handle the TARGET_PASS_BY_REFERENCE target hook. |
5502 | Specify whether to pass the argument by reference. */ | |
4fb4e4b8 | 5503 | |
8cd5a4e0 RH |
5504 | static bool |
5505 | sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED, | |
586de218 | 5506 | enum machine_mode mode, const_tree type, |
8cd5a4e0 | 5507 | bool named ATTRIBUTE_UNUSED) |
4fb4e4b8 DE |
5508 | { |
5509 | if (TARGET_ARCH32) | |
45261626 EB |
5510 | /* Original SPARC 32-bit ABI says that structures and unions, |
5511 | and quad-precision floats are passed by reference. For Pascal, | |
5512 | also pass arrays by reference. All other base types are passed | |
5513 | in registers. | |
5514 | ||
5515 | Extended ABI (as implemented by the Sun compiler) says that all | |
5516 | complex floats are passed by reference. Pass complex integers | |
5517 | in registers up to 8 bytes. More generally, enforce the 2-word | |
5518 | cap for passing arguments in registers. | |
5519 | ||
5520 | Vector ABI (as implemented by the Sun VIS SDK) says that vector | |
5521 | integers are passed like floats of the same size, that is in | |
5522 | registers up to 8 bytes. Pass all vector floats by reference | |
5523 | like structure and unions. */ | |
ef6843ed EB |
5524 | return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type))) |
5525 | || mode == SCmode | |
5526 | /* Catch CDImode, TFmode, DCmode and TCmode. */ | |
5527 | || GET_MODE_SIZE (mode) > 8 | |
5528 | || (type | |
5529 | && TREE_CODE (type) == VECTOR_TYPE | |
5530 | && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8)); | |
4fb4e4b8 | 5531 | else |
45261626 EB |
5532 | /* Original SPARC 64-bit ABI says that structures and unions |
5533 | smaller than 16 bytes are passed in registers, as well as | |
ef6843ed | 5534 | all other base types. |
4900aa61 | 5535 | |
45261626 EB |
5536 | Extended ABI (as implemented by the Sun compiler) says that |
5537 | complex floats are passed in registers up to 16 bytes. Pass | |
5538 | all complex integers in registers up to 16 bytes. More generally, | |
5539 | enforce the 2-word cap for passing arguments in registers. | |
5540 | ||
5541 | Vector ABI (as implemented by the Sun VIS SDK) says that vector | |
5542 | integers are passed like floats of the same size, that is in | |
5543 | registers (up to 16 bytes). Pass all vector floats like structure | |
5544 | and unions. */ | |
ef6843ed EB |
5545 | return ((type |
5546 | && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE) | |
5547 | && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16) | |
5548 | /* Catch CTImode and TCmode. */ | |
5549 | || GET_MODE_SIZE (mode) > 16); | |
4fb4e4b8 DE |
5550 | } |
5551 | ||
5552 | /* Handle the FUNCTION_ARG_ADVANCE macro. | |
5553 | Update the data in CUM to advance over an argument | |
5554 | of mode MODE and data type TYPE. | |
5555 | TYPE is null for libcalls where that information may not be available. */ | |
5556 | ||
5557 | void | |
fc27d102 KG |
5558 | function_arg_advance (struct sparc_args *cum, enum machine_mode mode, |
5559 | tree type, int named) | |
4fb4e4b8 DE |
5560 | { |
5561 | int slotno, regno, padding; | |
5562 | ||
5563 | /* We pass 0 for incoming_p here, it doesn't matter. */ | |
5564 | slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding); | |
5565 | ||
5566 | /* If register required leading padding, add it. */ | |
5567 | if (slotno != -1) | |
5568 | cum->words += padding; | |
eadceb59 | 5569 | |
4fb4e4b8 DE |
5570 | if (TARGET_ARCH32) |
5571 | { | |
5572 | cum->words += (mode != BLKmode | |
5573 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
5574 | : ROUND_ADVANCE (int_size_in_bytes (type))); | |
5575 | } | |
5576 | else | |
5577 | { | |
5578 | if (type && AGGREGATE_TYPE_P (type)) | |
5579 | { | |
5580 | int size = int_size_in_bytes (type); | |
5581 | ||
5582 | if (size <= 8) | |
5583 | ++cum->words; | |
5584 | else if (size <= 16) | |
5585 | cum->words += 2; | |
5586 | else /* passed by reference */ | |
5587 | ++cum->words; | |
5588 | } | |
4fb4e4b8 DE |
5589 | else |
5590 | { | |
5591 | cum->words += (mode != BLKmode | |
5592 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
5593 | : ROUND_ADVANCE (int_size_in_bytes (type))); | |
5594 | } | |
eadceb59 | 5595 | } |
4fb4e4b8 | 5596 | } |
eadceb59 | 5597 | |
4fb4e4b8 DE |
5598 | /* Handle the FUNCTION_ARG_PADDING macro. |
5599 | For the 64 bit ABI structs are always stored left shifted in their | |
5600 | argument slot. */ | |
5601 | ||
5602 | enum direction | |
586de218 | 5603 | function_arg_padding (enum machine_mode mode, const_tree type) |
4fb4e4b8 | 5604 | { |
c85f7c16 JL |
5605 | if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type)) |
5606 | return upward; | |
61a55e8b | 5607 | |
d3704c46 KH |
5608 | /* Fall back to the default. */ |
5609 | return DEFAULT_FUNCTION_ARG_PADDING (mode, type); | |
61a55e8b | 5610 | } |
82d6b402 | 5611 | |
1b5963f7 EB |
5612 | /* Handle the TARGET_RETURN_IN_MEMORY target hook. |
5613 | Specify whether to return the return value in memory. */ | |
5614 | ||
5615 | static bool | |
586de218 | 5616 | sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) |
1b5963f7 EB |
5617 | { |
5618 | if (TARGET_ARCH32) | |
45261626 EB |
5619 | /* Original SPARC 32-bit ABI says that structures and unions, |
5620 | and quad-precision floats are returned in memory. All other | |
5621 | base types are returned in registers. | |
5622 | ||
5623 | Extended ABI (as implemented by the Sun compiler) says that | |
5624 | all complex floats are returned in registers (8 FP registers | |
5625 | at most for '_Complex long double'). Return all complex integers | |
5626 | in registers (4 at most for '_Complex long long'). | |
5627 | ||
5628 | Vector ABI (as implemented by the Sun VIS SDK) says that vector | |
5629 | integers are returned like floats of the same size, that is in | |
5630 | registers up to 8 bytes and in memory otherwise. Return all | |
5631 | vector floats in memory like structure and unions; note that | |
5632 | they always have BLKmode like the latter. */ | |
1b5963f7 EB |
5633 | return (TYPE_MODE (type) == BLKmode |
5634 | || TYPE_MODE (type) == TFmode | |
45261626 EB |
5635 | || (TREE_CODE (type) == VECTOR_TYPE |
5636 | && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8)); | |
1b5963f7 EB |
5637 | else |
5638 | /* Original SPARC 64-bit ABI says that structures and unions | |
45261626 EB |
5639 | smaller than 32 bytes are returned in registers, as well as |
5640 | all other base types. | |
4900aa61 | 5641 | |
45261626 EB |
5642 | Extended ABI (as implemented by the Sun compiler) says that all |
5643 | complex floats are returned in registers (8 FP registers at most | |
5644 | for '_Complex long double'). Return all complex integers in | |
5645 | registers (4 at most for '_Complex TItype'). | |
5646 | ||
5647 | Vector ABI (as implemented by the Sun VIS SDK) says that vector | |
5648 | integers are returned like floats of the same size, that is in | |
5649 | registers. Return all vector floats like structure and unions; | |
5650 | note that they always have BLKmode like the latter. */ | |
1b5963f7 | 5651 | return ((TYPE_MODE (type) == BLKmode |
45261626 | 5652 | && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32)); |
1b5963f7 EB |
5653 | } |
5654 | ||
5655 | /* Handle the TARGET_STRUCT_VALUE target hook. | |
5656 | Return where to find the structure return value address. */ | |
5657 | ||
5658 | static rtx | |
2225b57c | 5659 | sparc_struct_value_rtx (tree fndecl, int incoming) |
1b5963f7 EB |
5660 | { |
5661 | if (TARGET_ARCH64) | |
5662 | return 0; | |
5663 | else | |
5664 | { | |
1da4c3a7 EB |
5665 | rtx mem; |
5666 | ||
1b5963f7 | 5667 | if (incoming) |
1da4c3a7 EB |
5668 | mem = gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx, |
5669 | STRUCT_VALUE_OFFSET)); | |
1b5963f7 | 5670 | else |
1da4c3a7 EB |
5671 | mem = gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx, |
5672 | STRUCT_VALUE_OFFSET)); | |
5673 | ||
4900aa61 EB |
5674 | /* Only follow the SPARC ABI for fixed-size structure returns. |
5675 | Variable size structure returns are handled per the normal | |
2225b57c | 5676 | procedures in GCC. This is enabled by -mstd-struct-return */ |
4900aa61 | 5677 | if (incoming == 2 |
2225b57c CD |
5678 | && sparc_std_struct_return |
5679 | && TYPE_SIZE_UNIT (TREE_TYPE (fndecl)) | |
5680 | && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST) | |
5681 | { | |
5682 | /* We must check and adjust the return address, as it is | |
5683 | optional as to whether the return object is really | |
5684 | provided. */ | |
5685 | rtx ret_rtx = gen_rtx_REG (Pmode, 31); | |
5686 | rtx scratch = gen_reg_rtx (SImode); | |
4900aa61 | 5687 | rtx endlab = gen_label_rtx (); |
2225b57c CD |
5688 | |
5689 | /* Calculate the return object size */ | |
5690 | tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl)); | |
5691 | rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff); | |
5692 | /* Construct a temporary return value */ | |
5693 | rtx temp_val = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0); | |
5694 | ||
5695 | /* Implement SPARC 32-bit psABI callee returns struck checking | |
4900aa61 EB |
5696 | requirements: |
5697 | ||
2225b57c CD |
5698 | Fetch the instruction where we will return to and see if |
5699 | it's an unimp instruction (the most significant 10 bits | |
5700 | will be zero). */ | |
5701 | emit_move_insn (scratch, gen_rtx_MEM (SImode, | |
5702 | plus_constant (ret_rtx, 8))); | |
5703 | /* Assume the size is valid and pre-adjust */ | |
5704 | emit_insn (gen_add3_insn (ret_rtx, ret_rtx, GEN_INT (4))); | |
5705 | emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode, 0, endlab); | |
5706 | emit_insn (gen_sub3_insn (ret_rtx, ret_rtx, GEN_INT (4))); | |
4900aa61 | 5707 | /* Assign stack temp: |
2225b57c CD |
5708 | Write the address of the memory pointed to by temp_val into |
5709 | the memory pointed to by mem */ | |
5710 | emit_move_insn (mem, XEXP (temp_val, 0)); | |
5711 | emit_label (endlab); | |
5712 | } | |
5713 | ||
1da4c3a7 EB |
5714 | set_mem_alias_set (mem, struct_value_alias_set); |
5715 | return mem; | |
1b5963f7 EB |
5716 | } |
5717 | } | |
5718 | ||
82d6b402 RH |
5719 | /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros. |
5720 | For v9, function return values are subject to the same rules as arguments, | |
1b5963f7 | 5721 | except that up to 32 bytes may be returned in registers. */ |
82d6b402 RH |
5722 | |
5723 | rtx | |
586de218 | 5724 | function_value (const_tree type, enum machine_mode mode, int incoming_p) |
82d6b402 | 5725 | { |
1b5963f7 EB |
5726 | /* Beware that the two values are swapped here wrt function_arg. */ |
5727 | int regbase = (incoming_p | |
5728 | ? SPARC_OUTGOING_INT_ARG_FIRST | |
5729 | : SPARC_INCOMING_INT_ARG_FIRST); | |
45261626 | 5730 | enum mode_class mclass = GET_MODE_CLASS (mode); |
82d6b402 | 5731 | int regno; |
82d6b402 | 5732 | |
ef6843ed EB |
5733 | /* Vector types deserve special treatment because they are polymorphic wrt |
5734 | their mode, depending upon whether VIS instructions are enabled. */ | |
45261626 EB |
5735 | if (type && TREE_CODE (type) == VECTOR_TYPE) |
5736 | { | |
45261626 | 5737 | HOST_WIDE_INT size = int_size_in_bytes (type); |
f5f7d171 JM |
5738 | gcc_assert ((TARGET_ARCH32 && size <= 8) |
5739 | || (TARGET_ARCH64 && size <= 32)); | |
45261626 EB |
5740 | |
5741 | if (mode == BLKmode) | |
5742 | return function_arg_vector_value (size, | |
45261626 EB |
5743 | SPARC_FP_ARG_FIRST); |
5744 | else | |
5745 | mclass = MODE_FLOAT; | |
5746 | } | |
ef6843ed EB |
5747 | |
5748 | if (TARGET_ARCH64 && type) | |
82d6b402 | 5749 | { |
ef6843ed | 5750 | /* Structures up to 32 bytes in size are returned in registers. */ |
82d6b402 RH |
5751 | if (TREE_CODE (type) == RECORD_TYPE) |
5752 | { | |
ef6843ed EB |
5753 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5754 | gcc_assert (size <= 32); | |
82d6b402 | 5755 | |
1eac9f59 | 5756 | return function_arg_record_value (type, mode, 0, 1, regbase); |
82d6b402 | 5757 | } |
ef6843ed EB |
5758 | |
5759 | /* Unions up to 32 bytes in size are returned in integer registers. */ | |
a7bba6ca EB |
5760 | else if (TREE_CODE (type) == UNION_TYPE) |
5761 | { | |
5762 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
f5f7d171 | 5763 | gcc_assert (size <= 32); |
a7bba6ca | 5764 | |
22d8d627 | 5765 | return function_arg_union_value (size, mode, 0, regbase); |
a7bba6ca | 5766 | } |
ef6843ed EB |
5767 | |
5768 | /* Objects that require it are returned in FP registers. */ | |
5769 | else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) | |
5770 | ; | |
5771 | ||
5772 | /* All other aggregate types are returned in an integer register in a | |
5773 | mode corresponding to the size of the type. */ | |
8a7199ad | 5774 | else if (AGGREGATE_TYPE_P (type)) |
82d6b402 | 5775 | { |
8a7199ad RK |
5776 | /* All other aggregate types are passed in an integer register |
5777 | in a mode corresponding to the size of the type. */ | |
ef6843ed EB |
5778 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5779 | gcc_assert (size <= 32); | |
82d6b402 | 5780 | |
ef6843ed | 5781 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); |
c74fd3f1 EB |
5782 | |
5783 | /* ??? We probably should have made the same ABI change in | |
5784 | 3.4.0 as the one we made for unions. The latter was | |
5785 | required by the SCD though, while the former is not | |
5786 | specified, so we favored compatibility and efficiency. | |
5787 | ||
5788 | Now we're stuck for aggregates larger than 16 bytes, | |
5789 | because OImode vanished in the meantime. Let's not | |
5790 | try to be unduly clever, and simply follow the ABI | |
5791 | for unions in that case. */ | |
5792 | if (mode == BLKmode) | |
ef6843ed | 5793 | return function_arg_union_value (size, mode, 0, regbase); |
45261626 EB |
5794 | else |
5795 | mclass = MODE_INT; | |
82d6b402 | 5796 | } |
ef6843ed | 5797 | |
cde0f3fd PB |
5798 | /* This must match sparc_promote_function_mode. |
5799 | ??? Maybe 32-bit pointers should actually remain in Pmode? */ | |
ef6843ed | 5800 | else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD) |
a7bba6ca | 5801 | mode = word_mode; |
82d6b402 RH |
5802 | } |
5803 | ||
ef6843ed | 5804 | if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU) |
1b5963f7 | 5805 | regno = SPARC_FP_ARG_FIRST; |
82d6b402 | 5806 | else |
1b5963f7 | 5807 | regno = regbase; |
82d6b402 | 5808 | |
254110c2 | 5809 | return gen_rtx_REG (mode, regno); |
82d6b402 RH |
5810 | } |
5811 | ||
648d2ffc RH |
5812 | /* Do what is necessary for `va_start'. We look at the current function |
5813 | to determine if stdarg or varargs is used and return the address of | |
5814 | the first unnamed parameter. */ | |
3ea1fdd3 | 5815 | |
3e2cc1d1 | 5816 | static rtx |
fc27d102 | 5817 | sparc_builtin_saveregs (void) |
3ea1fdd3 | 5818 | { |
38173d38 | 5819 | int first_reg = crtl->args.info.words; |
3ea1fdd3 JW |
5820 | rtx address; |
5821 | int regno; | |
5822 | ||
1b5963f7 | 5823 | for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++) |
254110c2 | 5824 | emit_move_insn (gen_rtx_MEM (word_mode, |
8ac61af7 RK |
5825 | gen_rtx_PLUS (Pmode, |
5826 | frame_pointer_rtx, | |
563c12b0 | 5827 | GEN_INT (FIRST_PARM_OFFSET (0) |
8ac61af7 RK |
5828 | + (UNITS_PER_WORD |
5829 | * regno)))), | |
254110c2 | 5830 | gen_rtx_REG (word_mode, |
1b5963f7 | 5831 | SPARC_INCOMING_INT_ARG_FIRST + regno)); |
3ea1fdd3 | 5832 | |
254110c2 | 5833 | address = gen_rtx_PLUS (Pmode, |
8ac61af7 | 5834 | frame_pointer_rtx, |
563c12b0 | 5835 | GEN_INT (FIRST_PARM_OFFSET (0) |
8ac61af7 | 5836 | + UNITS_PER_WORD * first_reg)); |
3ea1fdd3 JW |
5837 | |
5838 | return address; | |
5839 | } | |
a8b2c8a1 | 5840 | |
1b5963f7 | 5841 | /* Implement `va_start' for stdarg. */ |
a8b2c8a1 | 5842 | |
d7bd8aeb | 5843 | static void |
fc27d102 | 5844 | sparc_va_start (tree valist, rtx nextarg) |
a8b2c8a1 RH |
5845 | { |
5846 | nextarg = expand_builtin_saveregs (); | |
e5faf155 | 5847 | std_expand_builtin_va_start (valist, nextarg); |
a8b2c8a1 RH |
5848 | } |
5849 | ||
1b5963f7 | 5850 | /* Implement `va_arg' for stdarg. */ |
a8b2c8a1 | 5851 | |
35652573 | 5852 | static tree |
726a989a RB |
5853 | sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p, |
5854 | gimple_seq *post_p) | |
5f393b25 JM |
5855 | { |
5856 | HOST_WIDE_INT size, rsize, align; | |
5857 | tree addr, incr; | |
5858 | bool indirect; | |
5859 | tree ptrtype = build_pointer_type (type); | |
5860 | ||
22d66adf | 5861 | if (pass_by_reference (NULL, TYPE_MODE (type), type, false)) |
5f393b25 JM |
5862 | { |
5863 | indirect = true; | |
5864 | size = rsize = UNITS_PER_WORD; | |
5865 | align = 0; | |
5866 | } | |
5867 | else | |
5868 | { | |
5869 | indirect = false; | |
5870 | size = int_size_in_bytes (type); | |
5871 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
5872 | align = 0; | |
3764d512 | 5873 | |
5f393b25 JM |
5874 | if (TARGET_ARCH64) |
5875 | { | |
5876 | /* For SPARC64, objects requiring 16-byte alignment get it. */ | |
5877 | if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD) | |
5878 | align = 2 * UNITS_PER_WORD; | |
5879 | ||
5880 | /* SPARC-V9 ABI states that structures up to 16 bytes in size | |
22d66adf | 5881 | are left-justified in their slots. */ |
5f393b25 JM |
5882 | if (AGGREGATE_TYPE_P (type)) |
5883 | { | |
5884 | if (size == 0) | |
5885 | size = rsize = UNITS_PER_WORD; | |
5886 | else | |
5887 | size = rsize; | |
5888 | } | |
5889 | } | |
5890 | } | |
5891 | ||
5892 | incr = valist; | |
5893 | if (align) | |
5894 | { | |
5be014d5 AP |
5895 | incr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr, |
5896 | size_int (align - 1)); | |
5897 | incr = fold_convert (sizetype, incr); | |
5898 | incr = fold_build2 (BIT_AND_EXPR, sizetype, incr, | |
5899 | size_int (-align)); | |
5900 | incr = fold_convert (ptr_type_node, incr); | |
5f393b25 JM |
5901 | } |
5902 | ||
5903 | gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue); | |
5904 | addr = incr; | |
5905 | ||
5906 | if (BYTES_BIG_ENDIAN && size < rsize) | |
5be014d5 AP |
5907 | addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr, |
5908 | size_int (rsize - size)); | |
5f393b25 JM |
5909 | |
5910 | if (indirect) | |
5911 | { | |
5912 | addr = fold_convert (build_pointer_type (ptrtype), addr); | |
22d66adf | 5913 | addr = build_va_arg_indirect_ref (addr); |
5f393b25 | 5914 | } |
3764d512 EB |
5915 | |
5916 | /* If the address isn't aligned properly for the type, we need a temporary. | |
5917 | FIXME: This is inefficient, usually we can do this in registers. */ | |
5918 | else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD) | |
5f393b25 JM |
5919 | { |
5920 | tree tmp = create_tmp_var (type, "va_arg_tmp"); | |
5921 | tree dest_addr = build_fold_addr_expr (tmp); | |
3764d512 EB |
5922 | tree copy = build_call_expr (implicit_built_in_decls[BUILT_IN_MEMCPY], |
5923 | 3, dest_addr, addr, size_int (rsize)); | |
5924 | TREE_ADDRESSABLE (tmp) = 1; | |
5f393b25 JM |
5925 | gimplify_and_add (copy, pre_p); |
5926 | addr = dest_addr; | |
5927 | } | |
3764d512 | 5928 | |
5f393b25 JM |
5929 | else |
5930 | addr = fold_convert (ptrtype, addr); | |
5931 | ||
3764d512 EB |
5932 | incr |
5933 | = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr, size_int (rsize)); | |
726a989a | 5934 | gimplify_assign (valist, incr, post_p); |
5f393b25 | 5935 | |
22d66adf | 5936 | return build_va_arg_indirect_ref (addr); |
5f393b25 | 5937 | } |
ab835497 | 5938 | \f |
c75d6010 JM |
5939 | /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook. |
5940 | Specify whether the vector mode is supported by the hardware. */ | |
5941 | ||
5942 | static bool | |
5943 | sparc_vector_mode_supported_p (enum machine_mode mode) | |
5944 | { | |
5945 | return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false; | |
5946 | } | |
5947 | \f | |
1b3c2c2d EB |
5948 | /* Return the string to output an unconditional branch to LABEL, which is |
5949 | the operand number of the label. | |
5950 | ||
5951 | DEST is the destination insn (i.e. the label), INSN is the source. */ | |
5952 | ||
5953 | const char * | |
5954 | output_ubranch (rtx dest, int label, rtx insn) | |
5955 | { | |
5956 | static char string[64]; | |
4e5b002b | 5957 | bool v9_form = false; |
1b3c2c2d EB |
5958 | char *p; |
5959 | ||
4e5b002b | 5960 | if (TARGET_V9 && INSN_ADDRESSES_SET_P ()) |
1b3c2c2d | 5961 | { |
4e5b002b EB |
5962 | int delta = (INSN_ADDRESSES (INSN_UID (dest)) |
5963 | - INSN_ADDRESSES (INSN_UID (insn))); | |
5964 | /* Leave some instructions for "slop". */ | |
5965 | if (delta >= -260000 && delta < 260000) | |
5966 | v9_form = true; | |
1b3c2c2d | 5967 | } |
1b3c2c2d | 5968 | |
4e5b002b EB |
5969 | if (v9_form) |
5970 | strcpy (string, "ba%*,pt\t%%xcc, "); | |
5971 | else | |
5972 | strcpy (string, "b%*\t"); | |
1b3c2c2d EB |
5973 | |
5974 | p = strchr (string, '\0'); | |
5975 | *p++ = '%'; | |
5976 | *p++ = 'l'; | |
5977 | *p++ = '0' + label; | |
5978 | *p++ = '%'; | |
4e5b002b | 5979 | *p++ = '('; |
1b3c2c2d EB |
5980 | *p = '\0'; |
5981 | ||
5982 | return string; | |
5983 | } | |
5984 | ||
ab835497 | 5985 | /* Return the string to output a conditional branch to LABEL, which is |
c4ce6853 DE |
5986 | the operand number of the label. OP is the conditional expression. |
5987 | XEXP (OP, 0) is assumed to be a condition code register (integer or | |
5988 | floating point) and its mode specifies what kind of comparison we made. | |
61a55e8b | 5989 | |
1b3c2c2d EB |
5990 | DEST is the destination insn (i.e. the label), INSN is the source. |
5991 | ||
5e7a8ee0 | 5992 | REVERSED is nonzero if we should reverse the sense of the comparison. |
ab835497 | 5993 | |
4e5b002b | 5994 | ANNUL is nonzero if we should generate an annulling branch. */ |
ab835497 | 5995 | |
a6163c22 | 5996 | const char * |
fc27d102 | 5997 | output_cbranch (rtx op, rtx dest, int label, int reversed, int annul, |
4e5b002b | 5998 | rtx insn) |
ab835497 | 5999 | { |
a6163c22 | 6000 | static char string[64]; |
ab835497 | 6001 | enum rtx_code code = GET_CODE (op); |
c4ce6853 DE |
6002 | rtx cc_reg = XEXP (op, 0); |
6003 | enum machine_mode mode = GET_MODE (cc_reg); | |
0b82d204 JJ |
6004 | const char *labelno, *branch; |
6005 | int spaces = 8, far; | |
6006 | char *p; | |
6007 | ||
6008 | /* v9 branches are limited to +-1MB. If it is too far away, | |
6009 | change | |
6010 | ||
6011 | bne,pt %xcc, .LC30 | |
6012 | ||
6013 | to | |
6014 | ||
6015 | be,pn %xcc, .+12 | |
a6163c22 | 6016 | nop |
0b82d204 JJ |
6017 | ba .LC30 |
6018 | ||
6019 | and | |
6020 | ||
6021 | fbne,a,pn %fcc2, .LC29 | |
61a55e8b | 6022 | |
0b82d204 JJ |
6023 | to |
6024 | ||
6025 | fbe,pt %fcc2, .+16 | |
a6163c22 | 6026 | nop |
0b82d204 JJ |
6027 | ba .LC29 */ |
6028 | ||
883d9e0c | 6029 | far = TARGET_V9 && (get_attr_length (insn) >= 3); |
0b82d204 | 6030 | if (reversed ^ far) |
ab835497 | 6031 | { |
e267e177 RH |
6032 | /* Reversal of FP compares takes care -- an ordered compare |
6033 | becomes an unordered compare and vice versa. */ | |
4d449554 | 6034 | if (mode == CCFPmode || mode == CCFPEmode) |
7913f3d0 | 6035 | code = reverse_condition_maybe_unordered (code); |
ab835497 | 6036 | else |
e267e177 | 6037 | code = reverse_condition (code); |
ab835497 RK |
6038 | } |
6039 | ||
e267e177 RH |
6040 | /* Start by writing the branch condition. */ |
6041 | if (mode == CCFPmode || mode == CCFPEmode) | |
5850dc00 RH |
6042 | { |
6043 | switch (code) | |
6044 | { | |
6045 | case NE: | |
6046 | branch = "fbne"; | |
6047 | break; | |
6048 | case EQ: | |
6049 | branch = "fbe"; | |
6050 | break; | |
6051 | case GE: | |
6052 | branch = "fbge"; | |
6053 | break; | |
6054 | case GT: | |
6055 | branch = "fbg"; | |
6056 | break; | |
6057 | case LE: | |
6058 | branch = "fble"; | |
6059 | break; | |
6060 | case LT: | |
6061 | branch = "fbl"; | |
6062 | break; | |
6063 | case UNORDERED: | |
6064 | branch = "fbu"; | |
6065 | break; | |
6066 | case ORDERED: | |
6067 | branch = "fbo"; | |
6068 | break; | |
6069 | case UNGT: | |
6070 | branch = "fbug"; | |
6071 | break; | |
6072 | case UNLT: | |
6073 | branch = "fbul"; | |
6074 | break; | |
6075 | case UNEQ: | |
6076 | branch = "fbue"; | |
6077 | break; | |
6078 | case UNGE: | |
6079 | branch = "fbuge"; | |
6080 | break; | |
6081 | case UNLE: | |
6082 | branch = "fbule"; | |
6083 | break; | |
6084 | case LTGT: | |
6085 | branch = "fblg"; | |
6086 | break; | |
6087 | ||
6088 | default: | |
f5f7d171 | 6089 | gcc_unreachable (); |
5850dc00 RH |
6090 | } |
6091 | ||
6092 | /* ??? !v9: FP branches cannot be preceded by another floating point | |
6093 | insn. Because there is currently no concept of pre-delay slots, | |
6094 | we can fix this only by always emitting a nop before a floating | |
6095 | point branch. */ | |
6096 | ||
6097 | string[0] = '\0'; | |
6098 | if (! TARGET_V9) | |
6099 | strcpy (string, "nop\n\t"); | |
6100 | strcat (string, branch); | |
6101 | } | |
e267e177 | 6102 | else |
5850dc00 RH |
6103 | { |
6104 | switch (code) | |
6105 | { | |
6106 | case NE: | |
6107 | branch = "bne"; | |
6108 | break; | |
6109 | case EQ: | |
6110 | branch = "be"; | |
6111 | break; | |
6112 | case GE: | |
0b82d204 | 6113 | if (mode == CC_NOOVmode || mode == CCX_NOOVmode) |
5850dc00 RH |
6114 | branch = "bpos"; |
6115 | else | |
6116 | branch = "bge"; | |
6117 | break; | |
6118 | case GT: | |
6119 | branch = "bg"; | |
6120 | break; | |
6121 | case LE: | |
6122 | branch = "ble"; | |
6123 | break; | |
6124 | case LT: | |
0b82d204 | 6125 | if (mode == CC_NOOVmode || mode == CCX_NOOVmode) |
5850dc00 RH |
6126 | branch = "bneg"; |
6127 | else | |
6128 | branch = "bl"; | |
6129 | break; | |
6130 | case GEU: | |
6131 | branch = "bgeu"; | |
6132 | break; | |
6133 | case GTU: | |
6134 | branch = "bgu"; | |
6135 | break; | |
6136 | case LEU: | |
6137 | branch = "bleu"; | |
6138 | break; | |
6139 | case LTU: | |
6140 | branch = "blu"; | |
6141 | break; | |
6142 | ||
6143 | default: | |
f5f7d171 | 6144 | gcc_unreachable (); |
5850dc00 RH |
6145 | } |
6146 | strcpy (string, branch); | |
6147 | } | |
e267e177 | 6148 | spaces -= strlen (branch); |
0b82d204 | 6149 | p = strchr (string, '\0'); |
e267e177 | 6150 | |
ab835497 | 6151 | /* Now add the annulling, the label, and a possible noop. */ |
0b82d204 | 6152 | if (annul && ! far) |
e0d80184 | 6153 | { |
0b82d204 JJ |
6154 | strcpy (p, ",a"); |
6155 | p += 2; | |
e0d80184 DM |
6156 | spaces -= 2; |
6157 | } | |
ab835497 | 6158 | |
883d9e0c | 6159 | if (TARGET_V9) |
61a55e8b | 6160 | { |
c6b0465b | 6161 | rtx note; |
0b82d204 | 6162 | int v8 = 0; |
c6b0465b | 6163 | |
0b82d204 | 6164 | if (! far && insn && INSN_ADDRESSES_SET_P ()) |
e0d80184 | 6165 | { |
0b82d204 JJ |
6166 | int delta = (INSN_ADDRESSES (INSN_UID (dest)) |
6167 | - INSN_ADDRESSES (INSN_UID (insn))); | |
6168 | /* Leave some instructions for "slop". */ | |
6169 | if (delta < -260000 || delta >= 260000) | |
6170 | v8 = 1; | |
e0d80184 | 6171 | } |
c6b0465b | 6172 | |
61a55e8b DE |
6173 | if (mode == CCFPmode || mode == CCFPEmode) |
6174 | { | |
0b82d204 | 6175 | static char v9_fcc_labelno[] = "%%fccX, "; |
61a55e8b | 6176 | /* Set the char indicating the number of the fcc reg to use. */ |
0b82d204 JJ |
6177 | v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0'; |
6178 | labelno = v9_fcc_labelno; | |
6179 | if (v8) | |
6180 | { | |
f5f7d171 JM |
6181 | gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG); |
6182 | labelno = ""; | |
0b82d204 | 6183 | } |
61a55e8b DE |
6184 | } |
6185 | else if (mode == CCXmode || mode == CCX_NOOVmode) | |
0b82d204 JJ |
6186 | { |
6187 | labelno = "%%xcc, "; | |
f5f7d171 | 6188 | gcc_assert (! v8); |
0b82d204 | 6189 | } |
61a55e8b | 6190 | else |
0b82d204 JJ |
6191 | { |
6192 | labelno = "%%icc, "; | |
6193 | if (v8) | |
6194 | labelno = ""; | |
6195 | } | |
6196 | ||
4f31cce8 | 6197 | if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX))) |
0b82d204 JJ |
6198 | { |
6199 | strcpy (p, | |
4f31cce8 | 6200 | ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far) |
0b82d204 JJ |
6201 | ? ",pt" : ",pn"); |
6202 | p += 3; | |
6203 | spaces -= 3; | |
6204 | } | |
61a55e8b | 6205 | } |
883d9e0c EB |
6206 | else |
6207 | labelno = ""; | |
6208 | ||
e0d80184 | 6209 | if (spaces > 0) |
0b82d204 | 6210 | *p++ = '\t'; |
e0d80184 | 6211 | else |
0b82d204 JJ |
6212 | *p++ = ' '; |
6213 | strcpy (p, labelno); | |
6214 | p = strchr (p, '\0'); | |
6215 | if (far) | |
6216 | { | |
a6163c22 | 6217 | strcpy (p, ".+12\n\t nop\n\tb\t"); |
4e5b002b EB |
6218 | /* Skip the next insn if requested or |
6219 | if we know that it will be a nop. */ | |
6220 | if (annul || ! final_sequence) | |
0b82d204 | 6221 | p[3] = '6'; |
a6163c22 | 6222 | p += 14; |
0b82d204 JJ |
6223 | } |
6224 | *p++ = '%'; | |
6225 | *p++ = 'l'; | |
0b82d204 | 6226 | *p++ = label + '0'; |
4e5b002b EB |
6227 | *p++ = '%'; |
6228 | *p++ = '#'; | |
0b82d204 | 6229 | *p = '\0'; |
61a55e8b DE |
6230 | |
6231 | return string; | |
6232 | } | |
6233 | ||
47ac041c | 6234 | /* Emit a library call comparison between floating point X and Y. |
19047e4a EB |
6235 | COMPARISON is the operator to compare with (EQ, NE, GT, etc). |
6236 | Return the new operator to be used in the comparison sequence. | |
6237 | ||
47ac041c JJ |
6238 | TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode |
6239 | values as arguments instead of the TFmode registers themselves, | |
6240 | that's why we cannot call emit_float_lib_cmp. */ | |
19047e4a | 6241 | |
f90b7a5a | 6242 | rtx |
fc27d102 | 6243 | sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison) |
47ac041c | 6244 | { |
e77d72cb | 6245 | const char *qpfunc; |
390b17c2 | 6246 | rtx slot0, slot1, result, tem, tem2, libfunc; |
5c5c34a4 | 6247 | enum machine_mode mode; |
19047e4a | 6248 | enum rtx_code new_comparison; |
47ac041c JJ |
6249 | |
6250 | switch (comparison) | |
6251 | { | |
6252 | case EQ: | |
19047e4a | 6253 | qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq"); |
47ac041c JJ |
6254 | break; |
6255 | ||
6256 | case NE: | |
19047e4a | 6257 | qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne"); |
47ac041c JJ |
6258 | break; |
6259 | ||
6260 | case GT: | |
19047e4a | 6261 | qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt"); |
47ac041c JJ |
6262 | break; |
6263 | ||
6264 | case GE: | |
19047e4a | 6265 | qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge"); |
47ac041c JJ |
6266 | break; |
6267 | ||
6268 | case LT: | |
19047e4a | 6269 | qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt"); |
47ac041c JJ |
6270 | break; |
6271 | ||
6272 | case LE: | |
19047e4a | 6273 | qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle"); |
5c5c34a4 JJ |
6274 | break; |
6275 | ||
6276 | case ORDERED: | |
6277 | case UNORDERED: | |
6278 | case UNGT: | |
6279 | case UNLT: | |
6280 | case UNEQ: | |
6281 | case UNGE: | |
6282 | case UNLE: | |
6283 | case LTGT: | |
19047e4a | 6284 | qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp"); |
47ac041c JJ |
6285 | break; |
6286 | ||
6287 | default: | |
f5f7d171 | 6288 | gcc_unreachable (); |
47ac041c JJ |
6289 | } |
6290 | ||
5c5c34a4 | 6291 | if (TARGET_ARCH64) |
47ac041c | 6292 | { |
19047e4a EB |
6293 | if (MEM_P (x)) |
6294 | slot0 = x; | |
6295 | else | |
5c5c34a4 JJ |
6296 | { |
6297 | slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0); | |
c5d74498 | 6298 | emit_move_insn (slot0, x); |
5c5c34a4 | 6299 | } |
5c5c34a4 | 6300 | |
19047e4a EB |
6301 | if (MEM_P (y)) |
6302 | slot1 = y; | |
6303 | else | |
5c5c34a4 JJ |
6304 | { |
6305 | slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0); | |
c5d74498 | 6306 | emit_move_insn (slot1, y); |
5c5c34a4 | 6307 | } |
47ac041c | 6308 | |
390b17c2 RE |
6309 | libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc); |
6310 | emit_library_call (libfunc, LCT_NORMAL, | |
5c5c34a4 JJ |
6311 | DImode, 2, |
6312 | XEXP (slot0, 0), Pmode, | |
6313 | XEXP (slot1, 0), Pmode); | |
5c5c34a4 JJ |
6314 | mode = DImode; |
6315 | } | |
6316 | else | |
47ac041c | 6317 | { |
390b17c2 RE |
6318 | libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc); |
6319 | emit_library_call (libfunc, LCT_NORMAL, | |
5c5c34a4 JJ |
6320 | SImode, 2, |
6321 | x, TFmode, y, TFmode); | |
5c5c34a4 | 6322 | mode = SImode; |
47ac041c JJ |
6323 | } |
6324 | ||
47ac041c JJ |
6325 | |
6326 | /* Immediately move the result of the libcall into a pseudo | |
6327 | register so reload doesn't clobber the value if it needs | |
6328 | the return register for a spill reg. */ | |
5c5c34a4 | 6329 | result = gen_reg_rtx (mode); |
390b17c2 | 6330 | emit_move_insn (result, hard_libcall_value (mode, libfunc)); |
47ac041c | 6331 | |
5c5c34a4 JJ |
6332 | switch (comparison) |
6333 | { | |
6334 | default: | |
f90b7a5a | 6335 | return gen_rtx_NE (VOIDmode, result, const0_rtx); |
5c5c34a4 JJ |
6336 | case ORDERED: |
6337 | case UNORDERED: | |
19047e4a | 6338 | new_comparison = (comparison == UNORDERED ? EQ : NE); |
f90b7a5a | 6339 | return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3)); |
5c5c34a4 JJ |
6340 | case UNGT: |
6341 | case UNGE: | |
19047e4a | 6342 | new_comparison = (comparison == UNGT ? GT : NE); |
f90b7a5a | 6343 | return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx); |
5c5c34a4 | 6344 | case UNLE: |
f90b7a5a | 6345 | return gen_rtx_NE (VOIDmode, result, const2_rtx); |
5c5c34a4 JJ |
6346 | case UNLT: |
6347 | tem = gen_reg_rtx (mode); | |
6348 | if (TARGET_ARCH32) | |
6349 | emit_insn (gen_andsi3 (tem, result, const1_rtx)); | |
6350 | else | |
6351 | emit_insn (gen_anddi3 (tem, result, const1_rtx)); | |
f90b7a5a | 6352 | return gen_rtx_NE (VOIDmode, tem, const0_rtx); |
5c5c34a4 JJ |
6353 | case UNEQ: |
6354 | case LTGT: | |
6355 | tem = gen_reg_rtx (mode); | |
6356 | if (TARGET_ARCH32) | |
6357 | emit_insn (gen_addsi3 (tem, result, const1_rtx)); | |
6358 | else | |
6359 | emit_insn (gen_adddi3 (tem, result, const1_rtx)); | |
6360 | tem2 = gen_reg_rtx (mode); | |
6361 | if (TARGET_ARCH32) | |
6362 | emit_insn (gen_andsi3 (tem2, tem, const2_rtx)); | |
6363 | else | |
6364 | emit_insn (gen_anddi3 (tem2, tem, const2_rtx)); | |
19047e4a | 6365 | new_comparison = (comparison == UNEQ ? EQ : NE); |
f90b7a5a | 6366 | return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx); |
5c5c34a4 | 6367 | } |
19047e4a | 6368 | |
f90b7a5a | 6369 | gcc_unreachable (); |
47ac041c | 6370 | } |
5c5c34a4 | 6371 | |
d88e57d1 RH |
6372 | /* Generate an unsigned DImode to FP conversion. This is the same code |
6373 | optabs would emit if we didn't have TFmode patterns. */ | |
6374 | ||
6375 | void | |
b178305d | 6376 | sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode) |
d88e57d1 RH |
6377 | { |
6378 | rtx neglab, donelab, i0, i1, f0, in, out; | |
d88e57d1 RH |
6379 | |
6380 | out = operands[0]; | |
6381 | in = force_reg (DImode, operands[1]); | |
d88e57d1 RH |
6382 | neglab = gen_label_rtx (); |
6383 | donelab = gen_label_rtx (); | |
6384 | i0 = gen_reg_rtx (DImode); | |
6385 | i1 = gen_reg_rtx (DImode); | |
6386 | f0 = gen_reg_rtx (mode); | |
6387 | ||
6388 | emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab); | |
6389 | ||
6390 | emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in))); | |
6391 | emit_jump_insn (gen_jump (donelab)); | |
6392 | emit_barrier (); | |
6393 | ||
6394 | emit_label (neglab); | |
6395 | ||
6396 | emit_insn (gen_lshrdi3 (i0, in, const1_rtx)); | |
6397 | emit_insn (gen_anddi3 (i1, in, const1_rtx)); | |
6398 | emit_insn (gen_iordi3 (i0, i0, i1)); | |
6399 | emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0))); | |
6400 | emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0))); | |
6401 | ||
6402 | emit_label (donelab); | |
6403 | } | |
6404 | ||
b178305d EB |
6405 | /* Generate an FP to unsigned DImode conversion. This is the same code |
6406 | optabs would emit if we didn't have TFmode patterns. */ | |
6407 | ||
6408 | void | |
6409 | sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode) | |
6410 | { | |
6411 | rtx neglab, donelab, i0, i1, f0, in, out, limit; | |
6412 | ||
6413 | out = operands[0]; | |
6414 | in = force_reg (mode, operands[1]); | |
6415 | neglab = gen_label_rtx (); | |
6416 | donelab = gen_label_rtx (); | |
6417 | i0 = gen_reg_rtx (DImode); | |
6418 | i1 = gen_reg_rtx (DImode); | |
6419 | limit = gen_reg_rtx (mode); | |
6420 | f0 = gen_reg_rtx (mode); | |
6421 | ||
6422 | emit_move_insn (limit, | |
6423 | CONST_DOUBLE_FROM_REAL_VALUE ( | |
6424 | REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode)); | |
6425 | emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab); | |
6426 | ||
6427 | emit_insn (gen_rtx_SET (VOIDmode, | |
6428 | out, | |
6429 | gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in)))); | |
6430 | emit_jump_insn (gen_jump (donelab)); | |
6431 | emit_barrier (); | |
6432 | ||
6433 | emit_label (neglab); | |
6434 | ||
6435 | emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit))); | |
6436 | emit_insn (gen_rtx_SET (VOIDmode, | |
6437 | i0, | |
6438 | gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0)))); | |
6439 | emit_insn (gen_movdi (i1, const1_rtx)); | |
6440 | emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63))); | |
6441 | emit_insn (gen_xordi3 (out, i0, i1)); | |
6442 | ||
6443 | emit_label (donelab); | |
6444 | } | |
6445 | ||
61a55e8b DE |
6446 | /* Return the string to output a conditional branch to LABEL, testing |
6447 | register REG. LABEL is the operand number of the label; REG is the | |
6448 | operand number of the reg. OP is the conditional expression. The mode | |
6449 | of REG says what kind of comparison we made. | |
6450 | ||
1b3c2c2d EB |
6451 | DEST is the destination insn (i.e. the label), INSN is the source. |
6452 | ||
5e7a8ee0 | 6453 | REVERSED is nonzero if we should reverse the sense of the comparison. |
61a55e8b | 6454 | |
4e5b002b | 6455 | ANNUL is nonzero if we should generate an annulling branch. */ |
61a55e8b | 6456 | |
a6163c22 | 6457 | const char * |
fc27d102 | 6458 | output_v9branch (rtx op, rtx dest, int reg, int label, int reversed, |
4e5b002b | 6459 | int annul, rtx insn) |
61a55e8b | 6460 | { |
a6163c22 | 6461 | static char string[64]; |
61a55e8b DE |
6462 | enum rtx_code code = GET_CODE (op); |
6463 | enum machine_mode mode = GET_MODE (XEXP (op, 0)); | |
e0d80184 | 6464 | rtx note; |
0b82d204 JJ |
6465 | int far; |
6466 | char *p; | |
6467 | ||
6468 | /* branch on register are limited to +-128KB. If it is too far away, | |
6469 | change | |
4900aa61 | 6470 | |
0b82d204 | 6471 | brnz,pt %g1, .LC30 |
4900aa61 | 6472 | |
0b82d204 | 6473 | to |
4900aa61 | 6474 | |
0b82d204 | 6475 | brz,pn %g1, .+12 |
a6163c22 | 6476 | nop |
0b82d204 | 6477 | ba,pt %xcc, .LC30 |
4900aa61 | 6478 | |
0b82d204 | 6479 | and |
4900aa61 | 6480 | |
0b82d204 | 6481 | brgez,a,pn %o1, .LC29 |
4900aa61 | 6482 | |
0b82d204 | 6483 | to |
4900aa61 | 6484 | |
0b82d204 | 6485 | brlz,pt %o1, .+16 |
a6163c22 | 6486 | nop |
0b82d204 JJ |
6487 | ba,pt %xcc, .LC29 */ |
6488 | ||
6489 | far = get_attr_length (insn) >= 3; | |
61a55e8b DE |
6490 | |
6491 | /* If not floating-point or if EQ or NE, we can just reverse the code. */ | |
0b82d204 JJ |
6492 | if (reversed ^ far) |
6493 | code = reverse_condition (code); | |
61a55e8b DE |
6494 | |
6495 | /* Only 64 bit versions of these instructions exist. */ | |
f5f7d171 | 6496 | gcc_assert (mode == DImode); |
61a55e8b DE |
6497 | |
6498 | /* Start by writing the branch condition. */ | |
6499 | ||
6500 | switch (code) | |
6501 | { | |
6502 | case NE: | |
6503 | strcpy (string, "brnz"); | |
6504 | break; | |
6505 | ||
6506 | case EQ: | |
6507 | strcpy (string, "brz"); | |
6508 | break; | |
6509 | ||
6510 | case GE: | |
6511 | strcpy (string, "brgez"); | |
6512 | break; | |
6513 | ||
6514 | case LT: | |
6515 | strcpy (string, "brlz"); | |
6516 | break; | |
6517 | ||
6518 | case LE: | |
6519 | strcpy (string, "brlez"); | |
6520 | break; | |
6521 | ||
6522 | case GT: | |
6523 | strcpy (string, "brgz"); | |
6524 | break; | |
6525 | ||
6526 | default: | |
f5f7d171 | 6527 | gcc_unreachable (); |
61a55e8b DE |
6528 | } |
6529 | ||
0b82d204 JJ |
6530 | p = strchr (string, '\0'); |
6531 | ||
61a55e8b | 6532 | /* Now add the annulling, reg, label, and nop. */ |
0b82d204 | 6533 | if (annul && ! far) |
e0d80184 | 6534 | { |
0b82d204 JJ |
6535 | strcpy (p, ",a"); |
6536 | p += 2; | |
e0d80184 | 6537 | } |
61a55e8b | 6538 | |
4f31cce8 | 6539 | if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX))) |
e0d80184 | 6540 | { |
0b82d204 | 6541 | strcpy (p, |
4f31cce8 | 6542 | ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far) |
0b82d204 JJ |
6543 | ? ",pt" : ",pn"); |
6544 | p += 3; | |
e0d80184 | 6545 | } |
61a55e8b | 6546 | |
0b82d204 JJ |
6547 | *p = p < string + 8 ? '\t' : ' '; |
6548 | p++; | |
6549 | *p++ = '%'; | |
6550 | *p++ = '0' + reg; | |
6551 | *p++ = ','; | |
6552 | *p++ = ' '; | |
6553 | if (far) | |
6554 | { | |
6555 | int veryfar = 1, delta; | |
6556 | ||
6557 | if (INSN_ADDRESSES_SET_P ()) | |
6558 | { | |
6559 | delta = (INSN_ADDRESSES (INSN_UID (dest)) | |
6560 | - INSN_ADDRESSES (INSN_UID (insn))); | |
6561 | /* Leave some instructions for "slop". */ | |
6562 | if (delta >= -260000 && delta < 260000) | |
6563 | veryfar = 0; | |
6564 | } | |
6565 | ||
a6163c22 | 6566 | strcpy (p, ".+12\n\t nop\n\t"); |
4e5b002b EB |
6567 | /* Skip the next insn if requested or |
6568 | if we know that it will be a nop. */ | |
6569 | if (annul || ! final_sequence) | |
0b82d204 | 6570 | p[3] = '6'; |
a6163c22 | 6571 | p += 12; |
0b82d204 JJ |
6572 | if (veryfar) |
6573 | { | |
6574 | strcpy (p, "b\t"); | |
6575 | p += 2; | |
6576 | } | |
6577 | else | |
6578 | { | |
6579 | strcpy (p, "ba,pt\t%%xcc, "); | |
6580 | p += 13; | |
6581 | } | |
6582 | } | |
6583 | *p++ = '%'; | |
6584 | *p++ = 'l'; | |
6585 | *p++ = '0' + label; | |
4e5b002b EB |
6586 | *p++ = '%'; |
6587 | *p++ = '#'; | |
0b82d204 | 6588 | *p = '\0'; |
ab835497 | 6589 | |
ab835497 RK |
6590 | return string; |
6591 | } | |
6592 | ||
e48addee JJ |
6593 | /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7]. |
6594 | Such instructions cannot be used in the delay slot of return insn on v9. | |
6595 | If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts. | |
6596 | */ | |
284d86e9 | 6597 | |
e48addee | 6598 | static int |
fc27d102 | 6599 | epilogue_renumber (register rtx *where, int test) |
284d86e9 | 6600 | { |
e48addee JJ |
6601 | register const char *fmt; |
6602 | register int i; | |
6603 | register enum rtx_code code; | |
6604 | ||
6605 | if (*where == 0) | |
6606 | return 0; | |
6607 | ||
6608 | code = GET_CODE (*where); | |
284d86e9 JC |
6609 | |
6610 | switch (code) | |
6611 | { | |
284d86e9 | 6612 | case REG: |
e48addee JJ |
6613 | if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */ |
6614 | return 1; | |
6615 | if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32) | |
f1c25d3b | 6616 | *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where))); |
e48addee JJ |
6617 | case SCRATCH: |
6618 | case CC0: | |
6619 | case PC: | |
284d86e9 JC |
6620 | case CONST_INT: |
6621 | case CONST_DOUBLE: | |
e48addee | 6622 | return 0; |
76a7c776 | 6623 | |
09ebda1b SC |
6624 | /* Do not replace the frame pointer with the stack pointer because |
6625 | it can cause the delayed instruction to load below the stack. | |
6626 | This occurs when instructions like: | |
6627 | ||
6628 | (set (reg/i:SI 24 %i0) | |
6629 | (mem/f:SI (plus:SI (reg/f:SI 30 %fp) | |
6630 | (const_int -20 [0xffffffec])) 0)) | |
6631 | ||
6632 | are in the return delayed slot. */ | |
6633 | case PLUS: | |
6634 | if (GET_CODE (XEXP (*where, 0)) == REG | |
563c12b0 | 6635 | && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM |
09ebda1b SC |
6636 | && (GET_CODE (XEXP (*where, 1)) != CONST_INT |
6637 | || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS)) | |
6638 | return 1; | |
6639 | break; | |
6640 | ||
6641 | case MEM: | |
6642 | if (SPARC_STACK_BIAS | |
6643 | && GET_CODE (XEXP (*where, 0)) == REG | |
563c12b0 | 6644 | && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM) |
09ebda1b SC |
6645 | return 1; |
6646 | break; | |
6647 | ||
76a7c776 DM |
6648 | default: |
6649 | break; | |
e48addee | 6650 | } |
284d86e9 | 6651 | |
e48addee | 6652 | fmt = GET_RTX_FORMAT (code); |
284d86e9 | 6653 | |
e48addee JJ |
6654 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
6655 | { | |
6656 | if (fmt[i] == 'E') | |
6657 | { | |
6658 | register int j; | |
6659 | for (j = XVECLEN (*where, i) - 1; j >= 0; j--) | |
6660 | if (epilogue_renumber (&(XVECEXP (*where, i, j)), test)) | |
6661 | return 1; | |
6662 | } | |
6663 | else if (fmt[i] == 'e' | |
6664 | && epilogue_renumber (&(XEXP (*where, i)), test)) | |
6665 | return 1; | |
284d86e9 | 6666 | } |
e48addee | 6667 | return 0; |
284d86e9 | 6668 | } |
ab835497 RK |
6669 | \f |
6670 | /* Leaf functions and non-leaf functions have different needs. */ | |
6671 | ||
8b60264b | 6672 | static const int |
ab835497 RK |
6673 | reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER; |
6674 | ||
8b60264b | 6675 | static const int |
ab835497 RK |
6676 | reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER; |
6677 | ||
8b60264b | 6678 | static const int *const reg_alloc_orders[] = { |
ab835497 RK |
6679 | reg_leaf_alloc_order, |
6680 | reg_nonleaf_alloc_order}; | |
6681 | ||
6682 | void | |
fc27d102 | 6683 | order_regs_for_local_alloc (void) |
ab835497 RK |
6684 | { |
6685 | static int last_order_nonleaf = 1; | |
6686 | ||
6fb5fa3c | 6687 | if (df_regs_ever_live_p (15) != last_order_nonleaf) |
ab835497 RK |
6688 | { |
6689 | last_order_nonleaf = !last_order_nonleaf; | |
2e09e75a | 6690 | memcpy ((char *) reg_alloc_order, |
8b60264b | 6691 | (const char *) reg_alloc_orders[last_order_nonleaf], |
2e09e75a | 6692 | FIRST_PSEUDO_REGISTER * sizeof (int)); |
ab835497 RK |
6693 | } |
6694 | } | |
6695 | \f | |
e0d80184 DM |
6696 | /* Return 1 if REG and MEM are legitimate enough to allow the various |
6697 | mem<-->reg splits to be run. */ | |
6698 | ||
6699 | int | |
fc27d102 | 6700 | sparc_splitdi_legitimate (rtx reg, rtx mem) |
e0d80184 | 6701 | { |
e0d80184 | 6702 | /* Punt if we are here by mistake. */ |
f5f7d171 | 6703 | gcc_assert (reload_completed); |
e0d80184 DM |
6704 | |
6705 | /* We must have an offsettable memory reference. */ | |
6706 | if (! offsettable_memref_p (mem)) | |
6707 | return 0; | |
6708 | ||
6709 | /* If we have legitimate args for ldd/std, we do not want | |
6710 | the split to happen. */ | |
6711 | if ((REGNO (reg) % 2) == 0 | |
6712 | && mem_min_alignment (mem, 8)) | |
6713 | return 0; | |
6714 | ||
6715 | /* Success. */ | |
6716 | return 1; | |
6717 | } | |
6718 | ||
e61c29e9 | 6719 | /* Return 1 if x and y are some kind of REG and they refer to |
fae778eb | 6720 | different hard registers. This test is guaranteed to be |
e61c29e9 DM |
6721 | run after reload. */ |
6722 | ||
6723 | int | |
fc27d102 | 6724 | sparc_absnegfloat_split_legitimate (rtx x, rtx y) |
e61c29e9 | 6725 | { |
e61c29e9 DM |
6726 | if (GET_CODE (x) != REG) |
6727 | return 0; | |
e61c29e9 DM |
6728 | if (GET_CODE (y) != REG) |
6729 | return 0; | |
6730 | if (REGNO (x) == REGNO (y)) | |
6731 | return 0; | |
6732 | return 1; | |
6733 | } | |
6734 | ||
35016322 | 6735 | /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1. |
4900aa61 | 6736 | This makes them candidates for using ldd and std insns. |
35016322 | 6737 | |
bfd6bc60 | 6738 | Note reg1 and reg2 *must* be hard registers. */ |
35016322 JW |
6739 | |
6740 | int | |
fc27d102 | 6741 | registers_ok_for_ldd_peep (rtx reg1, rtx reg2) |
35016322 | 6742 | { |
35016322 | 6743 | /* We might have been passed a SUBREG. */ |
4900aa61 | 6744 | if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG) |
35016322 JW |
6745 | return 0; |
6746 | ||
35016322 JW |
6747 | if (REGNO (reg1) % 2 != 0) |
6748 | return 0; | |
6749 | ||
4900aa61 EB |
6750 | /* Integer ldd is deprecated in SPARC V9 */ |
6751 | if (TARGET_V9 && REGNO (reg1) < 32) | |
6752 | return 0; | |
bfd6bc60 | 6753 | |
35016322 | 6754 | return (REGNO (reg1) == REGNO (reg2) - 1); |
35016322 JW |
6755 | } |
6756 | ||
2aad5d68 | 6757 | /* Return 1 if the addresses in mem1 and mem2 are suitable for use in |
303f8933 | 6758 | an ldd or std insn. |
4900aa61 | 6759 | |
2aad5d68 DN |
6760 | This can only happen when addr1 and addr2, the addresses in mem1 |
6761 | and mem2, are consecutive memory locations (addr1 + 4 == addr2). | |
303f8933 DN |
6762 | addr1 must also be aligned on a 64-bit boundary. |
6763 | ||
6764 | Also iff dependent_reg_rtx is not null it should not be used to | |
6765 | compute the address for mem1, i.e. we cannot optimize a sequence | |
6766 | like: | |
6767 | ld [%o0], %o0 | |
6768 | ld [%o0 + 4], %o1 | |
6769 | to | |
bcdd764b | 6770 | ldd [%o0], %o0 |
0acf409f DM |
6771 | nor: |
6772 | ld [%g3 + 4], %g3 | |
6773 | ld [%g3], %g2 | |
6774 | to | |
6775 | ldd [%g3], %g2 | |
6776 | ||
6777 | But, note that the transformation from: | |
6778 | ld [%g2 + 4], %g3 | |
6779 | ld [%g2], %g2 | |
6780 | to | |
6781 | ldd [%g2], %g2 | |
6782 | is perfectly fine. Thus, the peephole2 patterns always pass us | |
6783 | the destination register of the first load, never the second one. | |
6784 | ||
303f8933 DN |
6785 | For stores we don't have a similar problem, so dependent_reg_rtx is |
6786 | NULL_RTX. */ | |
35016322 JW |
6787 | |
6788 | int | |
fc27d102 | 6789 | mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx) |
35016322 | 6790 | { |
2aad5d68 | 6791 | rtx addr1, addr2; |
0d587737 | 6792 | unsigned int reg1; |
6a151f87 | 6793 | HOST_WIDE_INT offset1; |
35016322 | 6794 | |
303f8933 DN |
6795 | /* The mems cannot be volatile. */ |
6796 | if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2)) | |
6797 | return 0; | |
2aad5d68 | 6798 | |
303f8933 | 6799 | /* MEM1 should be aligned on a 64-bit boundary. */ |
2aad5d68 DN |
6800 | if (MEM_ALIGN (mem1) < 64) |
6801 | return 0; | |
4900aa61 | 6802 | |
303f8933 DN |
6803 | addr1 = XEXP (mem1, 0); |
6804 | addr2 = XEXP (mem2, 0); | |
4900aa61 | 6805 | |
35016322 JW |
6806 | /* Extract a register number and offset (if used) from the first addr. */ |
6807 | if (GET_CODE (addr1) == PLUS) | |
6808 | { | |
6809 | /* If not a REG, return zero. */ | |
6810 | if (GET_CODE (XEXP (addr1, 0)) != REG) | |
6811 | return 0; | |
6812 | else | |
6813 | { | |
6814 | reg1 = REGNO (XEXP (addr1, 0)); | |
6815 | /* The offset must be constant! */ | |
6816 | if (GET_CODE (XEXP (addr1, 1)) != CONST_INT) | |
6817 | return 0; | |
6818 | offset1 = INTVAL (XEXP (addr1, 1)); | |
6819 | } | |
6820 | } | |
6821 | else if (GET_CODE (addr1) != REG) | |
6822 | return 0; | |
6823 | else | |
6824 | { | |
6825 | reg1 = REGNO (addr1); | |
6826 | /* This was a simple (mem (reg)) expression. Offset is 0. */ | |
6827 | offset1 = 0; | |
6828 | } | |
6829 | ||
6830 | /* Make sure the second address is a (mem (plus (reg) (const_int). */ | |
6831 | if (GET_CODE (addr2) != PLUS) | |
6832 | return 0; | |
6833 | ||
6834 | if (GET_CODE (XEXP (addr2, 0)) != REG | |
6835 | || GET_CODE (XEXP (addr2, 1)) != CONST_INT) | |
6836 | return 0; | |
6837 | ||
35016322 JW |
6838 | if (reg1 != REGNO (XEXP (addr2, 0))) |
6839 | return 0; | |
6840 | ||
303f8933 DN |
6841 | if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx)) |
6842 | return 0; | |
4900aa61 EB |
6843 | |
6844 | /* The first offset must be evenly divisible by 8 to ensure the | |
35016322 JW |
6845 | address is 64 bit aligned. */ |
6846 | if (offset1 % 8 != 0) | |
6847 | return 0; | |
6848 | ||
6849 | /* The offset for the second addr must be 4 more than the first addr. */ | |
6850 | if (INTVAL (XEXP (addr2, 1)) != offset1 + 4) | |
6851 | return 0; | |
6852 | ||
6853 | /* All the tests passed. addr1 and addr2 are valid for ldd and std | |
6854 | instructions. */ | |
6855 | return 1; | |
6856 | } | |
7c56249d | 6857 | |
157891a3 EB |
6858 | /* Return 1 if reg is a pseudo, or is the first register in |
6859 | a hard register pair. This makes it suitable for use in | |
7c56249d JL |
6860 | ldd and std insns. */ |
6861 | ||
6862 | int | |
fc27d102 | 6863 | register_ok_for_ldd (rtx reg) |
7c56249d | 6864 | { |
7c56249d | 6865 | /* We might have been passed a SUBREG. */ |
157891a3 | 6866 | if (!REG_P (reg)) |
7c56249d JL |
6867 | return 0; |
6868 | ||
6869 | if (REGNO (reg) < FIRST_PSEUDO_REGISTER) | |
6870 | return (REGNO (reg) % 2 == 0); | |
157891a3 EB |
6871 | |
6872 | return 1; | |
6873 | } | |
6874 | ||
6875 | /* Return 1 if OP is a memory whose address is known to be | |
6876 | aligned to 8-byte boundary, or a pseudo during reload. | |
6877 | This makes it suitable for use in ldd and std insns. */ | |
6878 | ||
6879 | int | |
6880 | memory_ok_for_ldd (rtx op) | |
6881 | { | |
6882 | if (MEM_P (op)) | |
6883 | { | |
6884 | /* In 64-bit mode, we assume that the address is word-aligned. */ | |
6885 | if (TARGET_ARCH32 && !mem_min_alignment (op, 8)) | |
6886 | return 0; | |
6887 | ||
6888 | if ((reload_in_progress || reload_completed) | |
6889 | && !strict_memory_address_p (Pmode, XEXP (op, 0))) | |
6890 | return 0; | |
6891 | } | |
6892 | else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER) | |
6893 | { | |
6894 | if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0)) | |
6895 | return 0; | |
6896 | } | |
6897 | else | |
6898 | return 0; | |
6899 | ||
6900 | return 1; | |
7c56249d | 6901 | } |
ab835497 | 6902 | \f |
ab835497 RK |
6903 | /* Print operand X (an rtx) in assembler syntax to file FILE. |
6904 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
6905 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
6906 | ||
6907 | void | |
fc27d102 | 6908 | print_operand (FILE *file, rtx x, int code) |
ab835497 RK |
6909 | { |
6910 | switch (code) | |
6911 | { | |
6912 | case '#': | |
4e5b002b EB |
6913 | /* Output an insn in a delay slot. */ |
6914 | if (final_sequence) | |
6915 | sparc_indent_opcode = 1; | |
6916 | else | |
e0d80184 | 6917 | fputs ("\n\t nop", file); |
ab835497 RK |
6918 | return; |
6919 | case '*': | |
c6ce0969 | 6920 | /* Output an annul flag if there's nothing for the delay slot and we |
4e5b002b EB |
6921 | are optimizing. This is always used with '(' below. |
6922 | Sun OS 4.1.1 dbx can't handle an annulled unconditional branch; | |
6923 | this is a dbx bug. So, we only do this when optimizing. | |
6924 | On UltraSPARC, a branch in a delay slot causes a pipeline flush. | |
bfd6bc60 | 6925 | Always emit a nop in case the next instruction is a branch. */ |
4e5b002b | 6926 | if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9)) |
c6ce0969 JW |
6927 | fputs (",a", file); |
6928 | return; | |
6929 | case '(': | |
6930 | /* Output a 'nop' if there's nothing for the delay slot and we are | |
6931 | not optimizing. This is always used with '*' above. */ | |
4e5b002b | 6932 | if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9)) |
e0d80184 | 6933 | fputs ("\n\t nop", file); |
4e5b002b EB |
6934 | else if (final_sequence) |
6935 | sparc_indent_opcode = 1; | |
6936 | return; | |
6937 | case ')': | |
6938 | /* Output the right displacement from the saved PC on function return. | |
6939 | The caller may have placed an "unimp" insn immediately after the call | |
6940 | so we have to account for it. This insn is used in the 32-bit ABI | |
66695571 | 6941 | when calling a function that returns a non zero-sized structure. The |
4e5b002b | 6942 | 64-bit ABI doesn't have it. Be careful to have this test be the same |
66695571 EB |
6943 | as that for the call. The exception is when sparc_std_struct_return |
6944 | is enabled, the psABI is followed exactly and the adjustment is made | |
6945 | by the code in sparc_struct_value_rtx. The call emitted is the same | |
6946 | when sparc_std_struct_return is enabled. */ | |
6947 | if (!TARGET_ARCH64 | |
e3b5732b | 6948 | && cfun->returns_struct |
66695571 EB |
6949 | && !sparc_std_struct_return |
6950 | && DECL_SIZE (DECL_RESULT (current_function_decl)) | |
6951 | && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl))) | |
6952 | == INTEGER_CST | |
6953 | && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl)))) | |
4e5b002b EB |
6954 | fputs ("12", file); |
6955 | else | |
6956 | fputc ('8', file); | |
ab835497 | 6957 | return; |
61a55e8b | 6958 | case '_': |
a0a301fc DE |
6959 | /* Output the Embedded Medium/Anywhere code model base register. */ |
6960 | fputs (EMBMEDANY_BASE_REG, file); | |
61a55e8b | 6961 | return; |
5751a10b JJ |
6962 | case '&': |
6963 | /* Print some local dynamic TLS name. */ | |
6964 | assemble_name (file, get_some_local_dynamic_name ()); | |
6965 | return; | |
4e5b002b | 6966 | |
ab835497 RK |
6967 | case 'Y': |
6968 | /* Adjust the operand to take into account a RESTORE operation. */ | |
bfd6bc60 JC |
6969 | if (GET_CODE (x) == CONST_INT) |
6970 | break; | |
6971 | else if (GET_CODE (x) != REG) | |
c725bd79 | 6972 | output_operand_lossage ("invalid %%Y operand"); |
b3b1e8bd | 6973 | else if (REGNO (x) < 8) |
ab835497 RK |
6974 | fputs (reg_names[REGNO (x)], file); |
6975 | else if (REGNO (x) >= 24 && REGNO (x) < 32) | |
6976 | fputs (reg_names[REGNO (x)-16], file); | |
6977 | else | |
c725bd79 | 6978 | output_operand_lossage ("invalid %%Y operand"); |
ab835497 | 6979 | return; |
0f39a9aa DE |
6980 | case 'L': |
6981 | /* Print out the low order register name of a register pair. */ | |
6982 | if (WORDS_BIG_ENDIAN) | |
6983 | fputs (reg_names[REGNO (x)+1], file); | |
6984 | else | |
6985 | fputs (reg_names[REGNO (x)], file); | |
6986 | return; | |
6987 | case 'H': | |
6988 | /* Print out the high order register name of a register pair. */ | |
6989 | if (WORDS_BIG_ENDIAN) | |
6990 | fputs (reg_names[REGNO (x)], file); | |
6991 | else | |
6992 | fputs (reg_names[REGNO (x)+1], file); | |
6993 | return; | |
ab835497 | 6994 | case 'R': |
795068a4 | 6995 | /* Print out the second register name of a register pair or quad. |
ab835497 RK |
6996 | I.e., R (%o0) => %o1. */ |
6997 | fputs (reg_names[REGNO (x)+1], file); | |
6998 | return; | |
795068a4 JW |
6999 | case 'S': |
7000 | /* Print out the third register name of a register quad. | |
7001 | I.e., S (%o0) => %o2. */ | |
7002 | fputs (reg_names[REGNO (x)+2], file); | |
7003 | return; | |
7004 | case 'T': | |
7005 | /* Print out the fourth register name of a register quad. | |
7006 | I.e., T (%o0) => %o3. */ | |
7007 | fputs (reg_names[REGNO (x)+3], file); | |
7008 | return; | |
304b7a23 DE |
7009 | case 'x': |
7010 | /* Print a condition code register. */ | |
7011 | if (REGNO (x) == SPARC_ICC_REG) | |
7012 | { | |
7013 | /* We don't handle CC[X]_NOOVmode because they're not supposed | |
7014 | to occur here. */ | |
7015 | if (GET_MODE (x) == CCmode) | |
7016 | fputs ("%icc", file); | |
7017 | else if (GET_MODE (x) == CCXmode) | |
7018 | fputs ("%xcc", file); | |
7019 | else | |
f5f7d171 | 7020 | gcc_unreachable (); |
304b7a23 DE |
7021 | } |
7022 | else | |
7023 | /* %fccN register */ | |
7024 | fputs (reg_names[REGNO (x)], file); | |
7025 | return; | |
ab835497 RK |
7026 | case 'm': |
7027 | /* Print the operand's address only. */ | |
7028 | output_address (XEXP (x, 0)); | |
7029 | return; | |
7030 | case 'r': | |
7031 | /* In this case we need a register. Use %g0 if the | |
77a02b01 | 7032 | operand is const0_rtx. */ |
76052e74 RS |
7033 | if (x == const0_rtx |
7034 | || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x)))) | |
ab835497 RK |
7035 | { |
7036 | fputs ("%g0", file); | |
7037 | return; | |
7038 | } | |
7039 | else | |
7040 | break; | |
7041 | ||
61a55e8b | 7042 | case 'A': |
ab835497 RK |
7043 | switch (GET_CODE (x)) |
7044 | { | |
7045 | case IOR: fputs ("or", file); break; | |
7046 | case AND: fputs ("and", file); break; | |
7047 | case XOR: fputs ("xor", file); break; | |
c725bd79 | 7048 | default: output_operand_lossage ("invalid %%A operand"); |
ab835497 RK |
7049 | } |
7050 | return; | |
7051 | ||
7052 | case 'B': | |
7053 | switch (GET_CODE (x)) | |
7054 | { | |
7055 | case IOR: fputs ("orn", file); break; | |
7056 | case AND: fputs ("andn", file); break; | |
7057 | case XOR: fputs ("xnor", file); break; | |
c725bd79 | 7058 | default: output_operand_lossage ("invalid %%B operand"); |
ab835497 RK |
7059 | } |
7060 | return; | |
7061 | ||
304b7a23 DE |
7062 | /* These are used by the conditional move instructions. */ |
7063 | case 'c' : | |
61a55e8b | 7064 | case 'C': |
304b7a23 | 7065 | { |
5c5c34a4 JJ |
7066 | enum rtx_code rc = GET_CODE (x); |
7067 | ||
7068 | if (code == 'c') | |
7069 | { | |
7070 | enum machine_mode mode = GET_MODE (XEXP (x, 0)); | |
7071 | if (mode == CCFPmode || mode == CCFPEmode) | |
7072 | rc = reverse_condition_maybe_unordered (GET_CODE (x)); | |
7073 | else | |
7074 | rc = reverse_condition (GET_CODE (x)); | |
7075 | } | |
304b7a23 DE |
7076 | switch (rc) |
7077 | { | |
7078 | case NE: fputs ("ne", file); break; | |
7079 | case EQ: fputs ("e", file); break; | |
7080 | case GE: fputs ("ge", file); break; | |
7081 | case GT: fputs ("g", file); break; | |
7082 | case LE: fputs ("le", file); break; | |
7083 | case LT: fputs ("l", file); break; | |
7084 | case GEU: fputs ("geu", file); break; | |
7085 | case GTU: fputs ("gu", file); break; | |
7086 | case LEU: fputs ("leu", file); break; | |
7087 | case LTU: fputs ("lu", file); break; | |
5c5c34a4 JJ |
7088 | case LTGT: fputs ("lg", file); break; |
7089 | case UNORDERED: fputs ("u", file); break; | |
7090 | case ORDERED: fputs ("o", file); break; | |
7091 | case UNLT: fputs ("ul", file); break; | |
7092 | case UNLE: fputs ("ule", file); break; | |
7093 | case UNGT: fputs ("ug", file); break; | |
7094 | case UNGE: fputs ("uge", file); break; | |
7095 | case UNEQ: fputs ("ue", file); break; | |
304b7a23 | 7096 | default: output_operand_lossage (code == 'c' |
c725bd79 NB |
7097 | ? "invalid %%c operand" |
7098 | : "invalid %%C operand"); | |
304b7a23 DE |
7099 | } |
7100 | return; | |
7101 | } | |
7102 | ||
7103 | /* These are used by the movr instruction pattern. */ | |
7104 | case 'd': | |
61a55e8b | 7105 | case 'D': |
304b7a23 DE |
7106 | { |
7107 | enum rtx_code rc = (code == 'd' | |
7108 | ? reverse_condition (GET_CODE (x)) | |
7109 | : GET_CODE (x)); | |
7110 | switch (rc) | |
7111 | { | |
7112 | case NE: fputs ("ne", file); break; | |
7113 | case EQ: fputs ("e", file); break; | |
7114 | case GE: fputs ("gez", file); break; | |
7115 | case LT: fputs ("lz", file); break; | |
7116 | case LE: fputs ("lez", file); break; | |
7117 | case GT: fputs ("gz", file); break; | |
7118 | default: output_operand_lossage (code == 'd' | |
c725bd79 NB |
7119 | ? "invalid %%d operand" |
7120 | : "invalid %%D operand"); | |
304b7a23 DE |
7121 | } |
7122 | return; | |
7123 | } | |
61a55e8b | 7124 | |
ab835497 RK |
7125 | case 'b': |
7126 | { | |
7127 | /* Print a sign-extended character. */ | |
9e0625a3 | 7128 | int i = trunc_int_for_mode (INTVAL (x), QImode); |
ab835497 RK |
7129 | fprintf (file, "%d", i); |
7130 | return; | |
7131 | } | |
7132 | ||
d2889939 RK |
7133 | case 'f': |
7134 | /* Operand must be a MEM; write its address. */ | |
7135 | if (GET_CODE (x) != MEM) | |
c725bd79 | 7136 | output_operand_lossage ("invalid %%f operand"); |
d2889939 RK |
7137 | output_address (XEXP (x, 0)); |
7138 | return; | |
7139 | ||
8707fe93 RH |
7140 | case 's': |
7141 | { | |
7142 | /* Print a sign-extended 32-bit value. */ | |
7143 | HOST_WIDE_INT i; | |
7144 | if (GET_CODE(x) == CONST_INT) | |
7145 | i = INTVAL (x); | |
7146 | else if (GET_CODE(x) == CONST_DOUBLE) | |
7147 | i = CONST_DOUBLE_LOW (x); | |
7148 | else | |
ffa916ba KG |
7149 | { |
7150 | output_operand_lossage ("invalid %%s operand"); | |
7151 | return; | |
7152 | } | |
8707fe93 RH |
7153 | i = trunc_int_for_mode (i, SImode); |
7154 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, i); | |
7155 | return; | |
7156 | } | |
7157 | ||
ab835497 RK |
7158 | case 0: |
7159 | /* Do nothing special. */ | |
7160 | break; | |
7161 | ||
7162 | default: | |
7163 | /* Undocumented flag. */ | |
415f583e | 7164 | output_operand_lossage ("invalid operand output code"); |
ab835497 RK |
7165 | } |
7166 | ||
7167 | if (GET_CODE (x) == REG) | |
7168 | fputs (reg_names[REGNO (x)], file); | |
7169 | else if (GET_CODE (x) == MEM) | |
7170 | { | |
7171 | fputc ('[', file); | |
ab835497 | 7172 | /* Poor Sun assembler doesn't understand absolute addressing. */ |
e6c1be7e | 7173 | if (CONSTANT_P (XEXP (x, 0))) |
ab835497 RK |
7174 | fputs ("%g0+", file); |
7175 | output_address (XEXP (x, 0)); | |
7176 | fputc (']', file); | |
7177 | } | |
7178 | else if (GET_CODE (x) == HIGH) | |
7179 | { | |
7180 | fputs ("%hi(", file); | |
7181 | output_addr_const (file, XEXP (x, 0)); | |
7182 | fputc (')', file); | |
7183 | } | |
7184 | else if (GET_CODE (x) == LO_SUM) | |
7185 | { | |
7186 | print_operand (file, XEXP (x, 0), 0); | |
e0d80184 DM |
7187 | if (TARGET_CM_MEDMID) |
7188 | fputs ("+%l44(", file); | |
7189 | else | |
7190 | fputs ("+%lo(", file); | |
ab835497 RK |
7191 | output_addr_const (file, XEXP (x, 1)); |
7192 | fputc (')', file); | |
7193 | } | |
e601abce JW |
7194 | else if (GET_CODE (x) == CONST_DOUBLE |
7195 | && (GET_MODE (x) == VOIDmode | |
7196 | || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT)) | |
ab835497 RK |
7197 | { |
7198 | if (CONST_DOUBLE_HIGH (x) == 0) | |
0d9484c5 | 7199 | fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x)); |
ab835497 RK |
7200 | else if (CONST_DOUBLE_HIGH (x) == -1 |
7201 | && CONST_DOUBLE_LOW (x) < 0) | |
0d9484c5 | 7202 | fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x)); |
ab835497 | 7203 | else |
e601abce | 7204 | output_operand_lossage ("long long constant not a valid immediate operand"); |
ab835497 | 7205 | } |
e601abce JW |
7206 | else if (GET_CODE (x) == CONST_DOUBLE) |
7207 | output_operand_lossage ("floating point constant not a valid immediate operand"); | |
ab835497 RK |
7208 | else { output_addr_const (file, x); } |
7209 | } | |
7210 | \f | |
301d03af RS |
7211 | /* Target hook for assembling integer objects. The sparc version has |
7212 | special handling for aligned DI-mode objects. */ | |
ab835497 | 7213 | |
301d03af | 7214 | static bool |
fc27d102 | 7215 | sparc_assemble_integer (rtx x, unsigned int size, int aligned_p) |
ab835497 | 7216 | { |
301d03af RS |
7217 | /* ??? We only output .xword's for symbols and only then in environments |
7218 | where the assembler can handle them. */ | |
7219 | if (aligned_p && size == 8 | |
7220 | && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE)) | |
ab835497 | 7221 | { |
301d03af | 7222 | if (TARGET_V9) |
61a55e8b | 7223 | { |
301d03af RS |
7224 | assemble_integer_with_op ("\t.xword\t", x); |
7225 | return true; | |
61a55e8b DE |
7226 | } |
7227 | else | |
7228 | { | |
301d03af RS |
7229 | assemble_aligned_integer (4, const0_rtx); |
7230 | assemble_aligned_integer (4, x); | |
7231 | return true; | |
61a55e8b | 7232 | } |
ab835497 | 7233 | } |
301d03af | 7234 | return default_assemble_integer (x, size, aligned_p); |
ab835497 | 7235 | } |
210aa14a | 7236 | \f |
b0468b84 RK |
7237 | /* Return the value of a code used in the .proc pseudo-op that says |
7238 | what kind of result this function returns. For non-C types, we pick | |
7239 | the closest C type. */ | |
7240 | ||
77a02b01 JW |
7241 | #ifndef SHORT_TYPE_SIZE |
7242 | #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2) | |
7243 | #endif | |
7244 | ||
7245 | #ifndef INT_TYPE_SIZE | |
7246 | #define INT_TYPE_SIZE BITS_PER_WORD | |
7247 | #endif | |
7248 | ||
7249 | #ifndef LONG_TYPE_SIZE | |
7250 | #define LONG_TYPE_SIZE BITS_PER_WORD | |
7251 | #endif | |
7252 | ||
7253 | #ifndef LONG_LONG_TYPE_SIZE | |
7254 | #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2) | |
7255 | #endif | |
7256 | ||
7257 | #ifndef FLOAT_TYPE_SIZE | |
7258 | #define FLOAT_TYPE_SIZE BITS_PER_WORD | |
7259 | #endif | |
7260 | ||
7261 | #ifndef DOUBLE_TYPE_SIZE | |
7262 | #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2) | |
7263 | #endif | |
7264 | ||
7265 | #ifndef LONG_DOUBLE_TYPE_SIZE | |
7266 | #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2) | |
7267 | #endif | |
210aa14a RS |
7268 | |
7269 | unsigned long | |
fc27d102 | 7270 | sparc_type_code (register tree type) |
210aa14a RS |
7271 | { |
7272 | register unsigned long qualifiers = 0; | |
e5e809f4 | 7273 | register unsigned shift; |
210aa14a | 7274 | |
abc95ed3 | 7275 | /* Only the first 30 bits of the qualifier are valid. We must refrain from |
aee2c3c5 JW |
7276 | setting more, since some assemblers will give an error for this. Also, |
7277 | we must be careful to avoid shifts of 32 bits or more to avoid getting | |
7278 | unpredictable results. */ | |
7279 | ||
e5e809f4 | 7280 | for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type)) |
210aa14a RS |
7281 | { |
7282 | switch (TREE_CODE (type)) | |
7283 | { | |
7284 | case ERROR_MARK: | |
7285 | return qualifiers; | |
4900aa61 | 7286 | |
210aa14a | 7287 | case ARRAY_TYPE: |
e5e809f4 | 7288 | qualifiers |= (3 << shift); |
210aa14a RS |
7289 | break; |
7290 | ||
7291 | case FUNCTION_TYPE: | |
7292 | case METHOD_TYPE: | |
e5e809f4 | 7293 | qualifiers |= (2 << shift); |
210aa14a RS |
7294 | break; |
7295 | ||
7296 | case POINTER_TYPE: | |
7297 | case REFERENCE_TYPE: | |
7298 | case OFFSET_TYPE: | |
e5e809f4 | 7299 | qualifiers |= (1 << shift); |
210aa14a | 7300 | break; |
ab835497 | 7301 | |
210aa14a RS |
7302 | case RECORD_TYPE: |
7303 | return (qualifiers | 8); | |
7304 | ||
7305 | case UNION_TYPE: | |
b0468b84 | 7306 | case QUAL_UNION_TYPE: |
210aa14a RS |
7307 | return (qualifiers | 9); |
7308 | ||
7309 | case ENUMERAL_TYPE: | |
7310 | return (qualifiers | 10); | |
7311 | ||
7312 | case VOID_TYPE: | |
7313 | return (qualifiers | 16); | |
7314 | ||
7315 | case INTEGER_TYPE: | |
654209e6 JW |
7316 | /* If this is a range type, consider it to be the underlying |
7317 | type. */ | |
7318 | if (TREE_TYPE (type) != 0) | |
e5e809f4 | 7319 | break; |
654209e6 | 7320 | |
77a02b01 | 7321 | /* Carefully distinguish all the standard types of C, |
b0468b84 | 7322 | without messing up if the language is not C. We do this by |
8df83eae | 7323 | testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to |
b0468b84 RK |
7324 | look at both the names and the above fields, but that's redundant. |
7325 | Any type whose size is between two C types will be considered | |
7326 | to be the wider of the two types. Also, we do not have a | |
7327 | special code to use for "long long", so anything wider than | |
7328 | long is treated the same. Note that we can't distinguish | |
7329 | between "int" and "long" in this code if they are the same | |
7330 | size, but that's fine, since neither can the assembler. */ | |
7331 | ||
7332 | if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE) | |
8df83eae | 7333 | return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2)); |
4900aa61 | 7334 | |
b0468b84 | 7335 | else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE) |
8df83eae | 7336 | return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3)); |
4900aa61 | 7337 | |
b0468b84 | 7338 | else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE) |
8df83eae | 7339 | return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4)); |
4900aa61 | 7340 | |
b0468b84 | 7341 | else |
8df83eae | 7342 | return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5)); |
4900aa61 | 7343 | |
210aa14a | 7344 | case REAL_TYPE: |
e5e809f4 JL |
7345 | /* If this is a range type, consider it to be the underlying |
7346 | type. */ | |
7347 | if (TREE_TYPE (type) != 0) | |
7348 | break; | |
7349 | ||
77a02b01 JW |
7350 | /* Carefully distinguish all the standard types of C, |
7351 | without messing up if the language is not C. */ | |
b0468b84 | 7352 | |
77a02b01 | 7353 | if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE) |
210aa14a | 7354 | return (qualifiers | 6); |
b0468b84 | 7355 | |
4900aa61 | 7356 | else |
b0468b84 | 7357 | return (qualifiers | 7); |
4900aa61 | 7358 | |
210aa14a | 7359 | case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */ |
13d39dbc | 7360 | /* ??? We need to distinguish between double and float complex types, |
c82aa69a JW |
7361 | but I don't know how yet because I can't reach this code from |
7362 | existing front-ends. */ | |
7363 | return (qualifiers | 7); /* Who knows? */ | |
7364 | ||
1b5963f7 | 7365 | case VECTOR_TYPE: |
62de55f7 | 7366 | case BOOLEAN_TYPE: /* Boolean truth value type. */ |
210aa14a | 7367 | case LANG_TYPE: /* ? */ |
26902ae0 | 7368 | return qualifiers; |
4900aa61 | 7369 | |
210aa14a | 7370 | default: |
f5f7d171 | 7371 | gcc_unreachable (); /* Not a type! */ |
210aa14a RS |
7372 | } |
7373 | } | |
e5e809f4 JL |
7374 | |
7375 | return qualifiers; | |
210aa14a | 7376 | } |
ead69aea | 7377 | \f |
61a55e8b DE |
7378 | /* Nested function support. */ |
7379 | ||
7380 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
7381 | FNADDR is an RTX for the address of the function's pure code. | |
7382 | CXT is an RTX for the static chain value for the function. | |
7383 | ||
7384 | This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi | |
7385 | (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes | |
7386 | (to store insns). This is a bit excessive. Perhaps a different | |
7387 | mechanism would be better here. | |
7388 | ||
849a528d | 7389 | Emit enough FLUSH insns to synchronize the data and instruction caches. */ |
61a55e8b | 7390 | |
809e78cc RH |
7391 | static void |
7392 | sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt) | |
61a55e8b | 7393 | { |
1ec0c607 | 7394 | /* SPARC 32-bit trampoline: |
c6b0465b | 7395 | |
e0d80184 DM |
7396 | sethi %hi(fn), %g1 |
7397 | sethi %hi(static), %g2 | |
7398 | jmp %g1+%lo(fn) | |
7399 | or %g2, %lo(static), %g2 | |
c6b0465b JC |
7400 | |
7401 | SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii | |
7402 | JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii | |
7403 | */ | |
7404 | ||
35aa3c1c | 7405 | emit_move_insn |
809e78cc | 7406 | (adjust_address (m_tramp, SImode, 0), |
35aa3c1c RK |
7407 | expand_binop (SImode, ior_optab, |
7408 | expand_shift (RSHIFT_EXPR, SImode, fnaddr, | |
7409 | size_int (10), 0, 1), | |
7410 | GEN_INT (trunc_int_for_mode (0x03000000, SImode)), | |
7411 | NULL_RTX, 1, OPTAB_DIRECT)); | |
7412 | ||
7413 | emit_move_insn | |
809e78cc | 7414 | (adjust_address (m_tramp, SImode, 4), |
35aa3c1c RK |
7415 | expand_binop (SImode, ior_optab, |
7416 | expand_shift (RSHIFT_EXPR, SImode, cxt, | |
7417 | size_int (10), 0, 1), | |
7418 | GEN_INT (trunc_int_for_mode (0x05000000, SImode)), | |
7419 | NULL_RTX, 1, OPTAB_DIRECT)); | |
7420 | ||
7421 | emit_move_insn | |
809e78cc | 7422 | (adjust_address (m_tramp, SImode, 8), |
35aa3c1c RK |
7423 | expand_binop (SImode, ior_optab, |
7424 | expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX), | |
7425 | GEN_INT (trunc_int_for_mode (0x81c06000, SImode)), | |
7426 | NULL_RTX, 1, OPTAB_DIRECT)); | |
7427 | ||
7428 | emit_move_insn | |
809e78cc | 7429 | (adjust_address (m_tramp, SImode, 12), |
35aa3c1c RK |
7430 | expand_binop (SImode, ior_optab, |
7431 | expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX), | |
7432 | GEN_INT (trunc_int_for_mode (0x8410a000, SImode)), | |
7433 | NULL_RTX, 1, OPTAB_DIRECT)); | |
c6b0465b | 7434 | |
c6b0465b JC |
7435 | /* On UltraSPARC a flush flushes an entire cache line. The trampoline is |
7436 | aligned on a 16 byte boundary so one flush clears it all. */ | |
809e78cc | 7437 | emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0)))); |
84643cbf | 7438 | if (sparc_cpu != PROCESSOR_ULTRASPARC |
4c837a1e | 7439 | && sparc_cpu != PROCESSOR_ULTRASPARC3 |
9eeaed6e DM |
7440 | && sparc_cpu != PROCESSOR_NIAGARA |
7441 | && sparc_cpu != PROCESSOR_NIAGARA2) | |
809e78cc | 7442 | emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8)))); |
1ec0c607 WH |
7443 | |
7444 | /* Call __enable_execute_stack after writing onto the stack to make sure | |
7445 | the stack address is accessible. */ | |
e7a742ec | 7446 | #ifdef ENABLE_EXECUTE_STACK |
f1c25d3b | 7447 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"), |
809e78cc | 7448 | LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode); |
1ec0c607 WH |
7449 | #endif |
7450 | ||
61a55e8b DE |
7451 | } |
7452 | ||
1ec0c607 | 7453 | /* The 64-bit version is simpler because it makes more sense to load the |
849a528d RK |
7454 | values as "immediate" data out of the trampoline. It's also easier since |
7455 | we can read the PC without clobbering a register. */ | |
7456 | ||
809e78cc RH |
7457 | static void |
7458 | sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt) | |
61a55e8b | 7459 | { |
1ec0c607 | 7460 | /* SPARC 64-bit trampoline: |
345a6161 | 7461 | |
e0d80184 DM |
7462 | rd %pc, %g1 |
7463 | ldx [%g1+24], %g5 | |
7464 | jmp %g5 | |
7465 | ldx [%g1+16], %g5 | |
c6b0465b JC |
7466 | +16 bytes data |
7467 | */ | |
7468 | ||
809e78cc | 7469 | emit_move_insn (adjust_address (m_tramp, SImode, 0), |
e3aaacf4 | 7470 | GEN_INT (trunc_int_for_mode (0x83414000, SImode))); |
809e78cc | 7471 | emit_move_insn (adjust_address (m_tramp, SImode, 4), |
e3aaacf4 | 7472 | GEN_INT (trunc_int_for_mode (0xca586018, SImode))); |
809e78cc | 7473 | emit_move_insn (adjust_address (m_tramp, SImode, 8), |
e3aaacf4 | 7474 | GEN_INT (trunc_int_for_mode (0x81c14000, SImode))); |
809e78cc | 7475 | emit_move_insn (adjust_address (m_tramp, SImode, 12), |
e3aaacf4 | 7476 | GEN_INT (trunc_int_for_mode (0xca586010, SImode))); |
809e78cc RH |
7477 | emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt); |
7478 | emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr); | |
7479 | emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0)))); | |
345a6161 | 7480 | |
fae15c93 | 7481 | if (sparc_cpu != PROCESSOR_ULTRASPARC |
4c837a1e | 7482 | && sparc_cpu != PROCESSOR_ULTRASPARC3 |
9eeaed6e DM |
7483 | && sparc_cpu != PROCESSOR_NIAGARA |
7484 | && sparc_cpu != PROCESSOR_NIAGARA2) | |
809e78cc | 7485 | emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8)))); |
1ec0c607 WH |
7486 | |
7487 | /* Call __enable_execute_stack after writing onto the stack to make sure | |
7488 | the stack address is accessible. */ | |
e7a742ec | 7489 | #ifdef ENABLE_EXECUTE_STACK |
f1c25d3b | 7490 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"), |
809e78cc | 7491 | LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode); |
1ec0c607 | 7492 | #endif |
61a55e8b | 7493 | } |
809e78cc RH |
7494 | |
7495 | /* Worker for TARGET_TRAMPOLINE_INIT. */ | |
7496 | ||
7497 | static void | |
7498 | sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt) | |
7499 | { | |
7500 | rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0)); | |
7501 | cxt = force_reg (Pmode, cxt); | |
7502 | if (TARGET_ARCH64) | |
7503 | sparc64_initialize_trampoline (m_tramp, fnaddr, cxt); | |
7504 | else | |
7505 | sparc32_initialize_trampoline (m_tramp, fnaddr, cxt); | |
7506 | } | |
61a55e8b | 7507 | \f |
dbb54862 JW |
7508 | /* Adjust the cost of a scheduling dependency. Return the new cost of |
7509 | a dependency LINK or INSN on DEP_INSN. COST is the current cost. */ | |
7510 | ||
8947065c | 7511 | static int |
fc27d102 | 7512 | supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost) |
dbb54862 JW |
7513 | { |
7514 | enum attr_type insn_type; | |
7515 | ||
7516 | if (! recog_memoized (insn)) | |
7517 | return 0; | |
7518 | ||
7519 | insn_type = get_attr_type (insn); | |
7520 | ||
7521 | if (REG_NOTE_KIND (link) == 0) | |
7522 | { | |
7523 | /* Data dependency; DEP_INSN writes a register that INSN reads some | |
7524 | cycles later. */ | |
7525 | ||
7526 | /* if a load, then the dependence must be on the memory address; | |
3bc8b61e DM |
7527 | add an extra "cycle". Note that the cost could be two cycles |
7528 | if the reg was written late in an instruction group; we ca not tell | |
dbb54862 JW |
7529 | here. */ |
7530 | if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD) | |
7531 | return cost + 3; | |
7532 | ||
7533 | /* Get the delay only if the address of the store is the dependence. */ | |
7534 | if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE) | |
7535 | { | |
7536 | rtx pat = PATTERN(insn); | |
7537 | rtx dep_pat = PATTERN (dep_insn); | |
7538 | ||
7539 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
3bc8b61e | 7540 | return cost; /* This should not happen! */ |
dbb54862 JW |
7541 | |
7542 | /* The dependency between the two instructions was on the data that | |
7543 | is being stored. Assume that this implies that the address of the | |
7544 | store is not dependent. */ | |
7545 | if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat))) | |
7546 | return cost; | |
7547 | ||
7548 | return cost + 3; /* An approximation. */ | |
7549 | } | |
7550 | ||
7551 | /* A shift instruction cannot receive its data from an instruction | |
7552 | in the same cycle; add a one cycle penalty. */ | |
7553 | if (insn_type == TYPE_SHIFT) | |
7554 | return cost + 3; /* Split before cascade into shift. */ | |
7555 | } | |
7556 | else | |
7557 | { | |
7558 | /* Anti- or output- dependency; DEP_INSN reads/writes a register that | |
7559 | INSN writes some cycles later. */ | |
7560 | ||
7561 | /* These are only significant for the fpu unit; writing a fp reg before | |
7562 | the fpu has finished with it stalls the processor. */ | |
7563 | ||
7564 | /* Reusing an integer register causes no problems. */ | |
7565 | if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT) | |
7566 | return 0; | |
7567 | } | |
7568 | ||
7569 | return cost; | |
7570 | } | |
bfd6bc60 | 7571 | |
8947065c | 7572 | static int |
fc27d102 | 7573 | hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost) |
8947065c RH |
7574 | { |
7575 | enum attr_type insn_type, dep_type; | |
7576 | rtx pat = PATTERN(insn); | |
7577 | rtx dep_pat = PATTERN (dep_insn); | |
7578 | ||
7579 | if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0) | |
7580 | return cost; | |
7581 | ||
7582 | insn_type = get_attr_type (insn); | |
7583 | dep_type = get_attr_type (dep_insn); | |
7584 | ||
7585 | switch (REG_NOTE_KIND (link)) | |
7586 | { | |
7587 | case 0: | |
7588 | /* Data dependency; DEP_INSN writes a register that INSN reads some | |
7589 | cycles later. */ | |
7590 | ||
7591 | switch (insn_type) | |
7592 | { | |
7593 | case TYPE_STORE: | |
7594 | case TYPE_FPSTORE: | |
80ffc95e | 7595 | /* Get the delay iff the address of the store is the dependence. */ |
8947065c RH |
7596 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) |
7597 | return cost; | |
7598 | ||
7599 | if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat))) | |
7600 | return cost; | |
7601 | return cost + 3; | |
7602 | ||
7603 | case TYPE_LOAD: | |
7604 | case TYPE_SLOAD: | |
7605 | case TYPE_FPLOAD: | |
7606 | /* If a load, then the dependence must be on the memory address. If | |
7607 | the addresses aren't equal, then it might be a false dependency */ | |
7608 | if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE) | |
7609 | { | |
7610 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET | |
4900aa61 | 7611 | || GET_CODE (SET_DEST (dep_pat)) != MEM |
8947065c RH |
7612 | || GET_CODE (SET_SRC (pat)) != MEM |
7613 | || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0), | |
7614 | XEXP (SET_SRC (pat), 0))) | |
7615 | return cost + 2; | |
7616 | ||
4900aa61 | 7617 | return cost + 8; |
8947065c RH |
7618 | } |
7619 | break; | |
7620 | ||
7621 | case TYPE_BRANCH: | |
7622 | /* Compare to branch latency is 0. There is no benefit from | |
7623 | separating compare and branch. */ | |
7624 | if (dep_type == TYPE_COMPARE) | |
7625 | return 0; | |
7626 | /* Floating point compare to branch latency is less than | |
7627 | compare to conditional move. */ | |
7628 | if (dep_type == TYPE_FPCMP) | |
7629 | return cost - 1; | |
7630 | break; | |
3bb5de61 KG |
7631 | default: |
7632 | break; | |
8947065c RH |
7633 | } |
7634 | break; | |
7635 | ||
7636 | case REG_DEP_ANTI: | |
80ffc95e | 7637 | /* Anti-dependencies only penalize the fpu unit. */ |
8947065c RH |
7638 | if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT) |
7639 | return 0; | |
7640 | break; | |
7641 | ||
7642 | default: | |
7643 | break; | |
4900aa61 | 7644 | } |
8947065c RH |
7645 | |
7646 | return cost; | |
7647 | } | |
7648 | ||
c237e94a | 7649 | static int |
fc27d102 | 7650 | sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost) |
8947065c RH |
7651 | { |
7652 | switch (sparc_cpu) | |
7653 | { | |
7654 | case PROCESSOR_SUPERSPARC: | |
7655 | cost = supersparc_adjust_cost (insn, link, dep, cost); | |
7656 | break; | |
7657 | case PROCESSOR_HYPERSPARC: | |
7658 | case PROCESSOR_SPARCLITE86X: | |
7659 | cost = hypersparc_adjust_cost (insn, link, dep, cost); | |
7660 | break; | |
8947065c RH |
7661 | default: |
7662 | break; | |
7663 | } | |
7664 | return cost; | |
7665 | } | |
7666 | ||
fae15c93 | 7667 | static void |
fc27d102 KG |
7668 | sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED, |
7669 | int sched_verbose ATTRIBUTE_UNUSED, | |
7670 | int max_ready ATTRIBUTE_UNUSED) | |
4900aa61 EB |
7671 | {} |
7672 | ||
fae15c93 | 7673 | static int |
fc27d102 | 7674 | sparc_use_sched_lookahead (void) |
fae15c93 | 7675 | { |
9eeaed6e DM |
7676 | if (sparc_cpu == PROCESSOR_NIAGARA |
7677 | || sparc_cpu == PROCESSOR_NIAGARA2) | |
4c837a1e | 7678 | return 0; |
fae15c93 VM |
7679 | if (sparc_cpu == PROCESSOR_ULTRASPARC |
7680 | || sparc_cpu == PROCESSOR_ULTRASPARC3) | |
7681 | return 4; | |
7682 | if ((1 << sparc_cpu) & | |
7683 | ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) | | |
7684 | (1 << PROCESSOR_SPARCLITE86X))) | |
7685 | return 3; | |
3bc8b61e DM |
7686 | return 0; |
7687 | } | |
7688 | ||
fae15c93 | 7689 | static int |
fc27d102 | 7690 | sparc_issue_rate (void) |
bfd6bc60 JC |
7691 | { |
7692 | switch (sparc_cpu) | |
7693 | { | |
4c837a1e | 7694 | case PROCESSOR_NIAGARA: |
9eeaed6e | 7695 | case PROCESSOR_NIAGARA2: |
fae15c93 VM |
7696 | default: |
7697 | return 1; | |
7698 | case PROCESSOR_V9: | |
284d86e9 | 7699 | /* Assume V9 processors are capable of at least dual-issue. */ |
bfd6bc60 | 7700 | return 2; |
fae15c93 VM |
7701 | case PROCESSOR_SUPERSPARC: |
7702 | return 3; | |
393049a9 RH |
7703 | case PROCESSOR_HYPERSPARC: |
7704 | case PROCESSOR_SPARCLITE86X: | |
7705 | return 2; | |
fae15c93 VM |
7706 | case PROCESSOR_ULTRASPARC: |
7707 | case PROCESSOR_ULTRASPARC3: | |
7708 | return 4; | |
bfd6bc60 JC |
7709 | } |
7710 | } | |
284d86e9 JC |
7711 | |
7712 | static int | |
fc27d102 | 7713 | set_extends (rtx insn) |
284d86e9 JC |
7714 | { |
7715 | register rtx pat = PATTERN (insn); | |
7716 | ||
7717 | switch (GET_CODE (SET_SRC (pat))) | |
7718 | { | |
80ffc95e | 7719 | /* Load and some shift instructions zero extend. */ |
284d86e9 JC |
7720 | case MEM: |
7721 | case ZERO_EXTEND: | |
7722 | /* sethi clears the high bits */ | |
7723 | case HIGH: | |
7724 | /* LO_SUM is used with sethi. sethi cleared the high | |
7725 | bits and the values used with lo_sum are positive */ | |
7726 | case LO_SUM: | |
284d86e9 JC |
7727 | /* Store flag stores 0 or 1 */ |
7728 | case LT: case LTU: | |
7729 | case GT: case GTU: | |
7730 | case LE: case LEU: | |
7731 | case GE: case GEU: | |
7732 | case EQ: | |
7733 | case NE: | |
7734 | return 1; | |
7735 | case AND: | |
7736 | { | |
4df1190a | 7737 | rtx op0 = XEXP (SET_SRC (pat), 0); |
284d86e9 JC |
7738 | rtx op1 = XEXP (SET_SRC (pat), 1); |
7739 | if (GET_CODE (op1) == CONST_INT) | |
7740 | return INTVAL (op1) >= 0; | |
4df1190a JJ |
7741 | if (GET_CODE (op0) != REG) |
7742 | return 0; | |
7743 | if (sparc_check_64 (op0, insn) == 1) | |
284d86e9 | 7744 | return 1; |
4df1190a JJ |
7745 | return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1); |
7746 | } | |
7747 | case IOR: | |
7748 | case XOR: | |
7749 | { | |
7750 | rtx op0 = XEXP (SET_SRC (pat), 0); | |
7751 | rtx op1 = XEXP (SET_SRC (pat), 1); | |
7752 | if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0) | |
7753 | return 0; | |
7754 | if (GET_CODE (op1) == CONST_INT) | |
7755 | return INTVAL (op1) >= 0; | |
7756 | return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1); | |
284d86e9 | 7757 | } |
284d86e9 JC |
7758 | case LSHIFTRT: |
7759 | return GET_MODE (SET_SRC (pat)) == SImode; | |
80ffc95e | 7760 | /* Positive integers leave the high bits zero. */ |
284d86e9 | 7761 | case CONST_DOUBLE: |
4df1190a | 7762 | return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000); |
284d86e9 | 7763 | case CONST_INT: |
4df1190a | 7764 | return ! (INTVAL (SET_SRC (pat)) & 0x80000000); |
284d86e9 JC |
7765 | case ASHIFTRT: |
7766 | case SIGN_EXTEND: | |
7767 | return - (GET_MODE (SET_SRC (pat)) == SImode); | |
4df1190a JJ |
7768 | case REG: |
7769 | return sparc_check_64 (SET_SRC (pat), insn); | |
284d86e9 JC |
7770 | default: |
7771 | return 0; | |
7772 | } | |
7773 | } | |
7774 | ||
80ffc95e | 7775 | /* We _ought_ to have only one kind per function, but... */ |
e2500fed GK |
7776 | static GTY(()) rtx sparc_addr_diff_list; |
7777 | static GTY(()) rtx sparc_addr_list; | |
e0d80184 DM |
7778 | |
7779 | void | |
fc27d102 | 7780 | sparc_defer_case_vector (rtx lab, rtx vec, int diff) |
e0d80184 DM |
7781 | { |
7782 | vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec); | |
7783 | if (diff) | |
7784 | sparc_addr_diff_list | |
7785 | = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list); | |
7786 | else | |
7787 | sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list); | |
7788 | } | |
7789 | ||
4900aa61 | 7790 | static void |
fc27d102 | 7791 | sparc_output_addr_vec (rtx vec) |
e0d80184 DM |
7792 | { |
7793 | rtx lab = XEXP (vec, 0), body = XEXP (vec, 1); | |
7794 | int idx, vlen = XVECLEN (body, 0); | |
7795 | ||
4900aa61 | 7796 | #ifdef ASM_OUTPUT_ADDR_VEC_START |
d1accaa3 JJ |
7797 | ASM_OUTPUT_ADDR_VEC_START (asm_out_file); |
7798 | #endif | |
7799 | ||
e0d80184 DM |
7800 | #ifdef ASM_OUTPUT_CASE_LABEL |
7801 | ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab), | |
7802 | NEXT_INSN (lab)); | |
7803 | #else | |
4977bab6 | 7804 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab)); |
e0d80184 DM |
7805 | #endif |
7806 | ||
7807 | for (idx = 0; idx < vlen; idx++) | |
7808 | { | |
7809 | ASM_OUTPUT_ADDR_VEC_ELT | |
7810 | (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0))); | |
7811 | } | |
4900aa61 | 7812 | |
d1accaa3 JJ |
7813 | #ifdef ASM_OUTPUT_ADDR_VEC_END |
7814 | ASM_OUTPUT_ADDR_VEC_END (asm_out_file); | |
7815 | #endif | |
e0d80184 DM |
7816 | } |
7817 | ||
4900aa61 | 7818 | static void |
fc27d102 | 7819 | sparc_output_addr_diff_vec (rtx vec) |
e0d80184 DM |
7820 | { |
7821 | rtx lab = XEXP (vec, 0), body = XEXP (vec, 1); | |
7822 | rtx base = XEXP (XEXP (body, 0), 0); | |
7823 | int idx, vlen = XVECLEN (body, 1); | |
7824 | ||
4900aa61 | 7825 | #ifdef ASM_OUTPUT_ADDR_VEC_START |
d1accaa3 JJ |
7826 | ASM_OUTPUT_ADDR_VEC_START (asm_out_file); |
7827 | #endif | |
7828 | ||
e0d80184 DM |
7829 | #ifdef ASM_OUTPUT_CASE_LABEL |
7830 | ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab), | |
7831 | NEXT_INSN (lab)); | |
7832 | #else | |
4977bab6 | 7833 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab)); |
e0d80184 DM |
7834 | #endif |
7835 | ||
7836 | for (idx = 0; idx < vlen; idx++) | |
7837 | { | |
7838 | ASM_OUTPUT_ADDR_DIFF_ELT | |
7839 | (asm_out_file, | |
7840 | body, | |
7841 | CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)), | |
7842 | CODE_LABEL_NUMBER (base)); | |
7843 | } | |
4900aa61 | 7844 | |
d1accaa3 JJ |
7845 | #ifdef ASM_OUTPUT_ADDR_VEC_END |
7846 | ASM_OUTPUT_ADDR_VEC_END (asm_out_file); | |
7847 | #endif | |
e0d80184 DM |
7848 | } |
7849 | ||
7850 | static void | |
fc27d102 | 7851 | sparc_output_deferred_case_vectors (void) |
e0d80184 DM |
7852 | { |
7853 | rtx t; | |
d676da68 | 7854 | int align; |
e0d80184 | 7855 | |
f3b8847b DM |
7856 | if (sparc_addr_list == NULL_RTX |
7857 | && sparc_addr_diff_list == NULL_RTX) | |
7858 | return; | |
7859 | ||
e0d80184 | 7860 | /* Align to cache line in the function's code section. */ |
d6b5193b | 7861 | switch_to_section (current_function_section ()); |
d676da68 DM |
7862 | |
7863 | align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT); | |
7864 | if (align > 0) | |
f3b8847b | 7865 | ASM_OUTPUT_ALIGN (asm_out_file, align); |
4900aa61 | 7866 | |
e0d80184 DM |
7867 | for (t = sparc_addr_list; t ; t = XEXP (t, 1)) |
7868 | sparc_output_addr_vec (XEXP (t, 0)); | |
7869 | for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1)) | |
7870 | sparc_output_addr_diff_vec (XEXP (t, 0)); | |
7871 | ||
7872 | sparc_addr_list = sparc_addr_diff_list = NULL_RTX; | |
7873 | } | |
7874 | ||
284d86e9 JC |
7875 | /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are |
7876 | unknown. Return 1 if the high bits are zero, -1 if the register is | |
7877 | sign extended. */ | |
7878 | int | |
fc27d102 | 7879 | sparc_check_64 (rtx x, rtx insn) |
284d86e9 JC |
7880 | { |
7881 | /* If a register is set only once it is safe to ignore insns this | |
7882 | code does not know how to handle. The loop will either recognize | |
7883 | the single set and return the correct value or fail to recognize | |
7884 | it and return 0. */ | |
7885 | int set_once = 0; | |
4df1190a JJ |
7886 | rtx y = x; |
7887 | ||
f5f7d171 | 7888 | gcc_assert (GET_CODE (x) == REG); |
4df1190a JJ |
7889 | |
7890 | if (GET_MODE (x) == DImode) | |
7891 | y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN); | |
284d86e9 | 7892 | |
4df1190a | 7893 | if (flag_expensive_optimizations |
b718216c | 7894 | && df && DF_REG_DEF_COUNT (REGNO (y)) == 1) |
284d86e9 JC |
7895 | set_once = 1; |
7896 | ||
7897 | if (insn == 0) | |
d6f4ec51 KG |
7898 | { |
7899 | if (set_once) | |
7900 | insn = get_last_insn_anywhere (); | |
7901 | else | |
7902 | return 0; | |
7903 | } | |
284d86e9 | 7904 | |
d6f4ec51 | 7905 | while ((insn = PREV_INSN (insn))) |
284d86e9 JC |
7906 | { |
7907 | switch (GET_CODE (insn)) | |
7908 | { | |
7909 | case JUMP_INSN: | |
7910 | case NOTE: | |
7911 | break; | |
7912 | case CODE_LABEL: | |
7913 | case CALL_INSN: | |
7914 | default: | |
7915 | if (! set_once) | |
7916 | return 0; | |
7917 | break; | |
7918 | case INSN: | |
7919 | { | |
7920 | rtx pat = PATTERN (insn); | |
7921 | if (GET_CODE (pat) != SET) | |
7922 | return 0; | |
7923 | if (rtx_equal_p (x, SET_DEST (pat))) | |
4df1190a JJ |
7924 | return set_extends (insn); |
7925 | if (y && rtx_equal_p (y, SET_DEST (pat))) | |
7926 | return set_extends (insn); | |
7927 | if (reg_overlap_mentioned_p (SET_DEST (pat), y)) | |
284d86e9 JC |
7928 | return 0; |
7929 | } | |
7930 | } | |
7931 | } | |
7932 | return 0; | |
7933 | } | |
7934 | ||
01d3224a EB |
7935 | /* Returns assembly code to perform a DImode shift using |
7936 | a 64-bit global or out register on SPARC-V8+. */ | |
a6163c22 EB |
7937 | const char * |
7938 | output_v8plus_shift (rtx *operands, rtx insn, const char *opcode) | |
284d86e9 JC |
7939 | { |
7940 | static char asm_code[60]; | |
7941 | ||
01d3224a EB |
7942 | /* The scratch register is only required when the destination |
7943 | register is not a 64-bit global or out register. */ | |
7944 | if (which_alternative != 2) | |
284d86e9 | 7945 | operands[3] = operands[0]; |
01d3224a | 7946 | |
1c8b4e29 EB |
7947 | /* We can only shift by constants <= 63. */ |
7948 | if (GET_CODE (operands[2]) == CONST_INT) | |
7949 | operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f); | |
7950 | ||
c6b0465b JC |
7951 | if (GET_CODE (operands[1]) == CONST_INT) |
7952 | { | |
4df1190a | 7953 | output_asm_insn ("mov\t%1, %3", operands); |
c6b0465b JC |
7954 | } |
7955 | else | |
7956 | { | |
4df1190a | 7957 | output_asm_insn ("sllx\t%H1, 32, %3", operands); |
c6b0465b | 7958 | if (sparc_check_64 (operands[1], insn) <= 0) |
4df1190a JJ |
7959 | output_asm_insn ("srl\t%L1, 0, %L1", operands); |
7960 | output_asm_insn ("or\t%L1, %3, %3", operands); | |
c6b0465b | 7961 | } |
284d86e9 JC |
7962 | |
7963 | strcpy(asm_code, opcode); | |
01d3224a | 7964 | |
284d86e9 | 7965 | if (which_alternative != 2) |
4df1190a | 7966 | return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0"); |
284d86e9 | 7967 | else |
4df1190a | 7968 | return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0"); |
284d86e9 | 7969 | } |
2be15d0f | 7970 | \f |
89a8b315 RH |
7971 | /* Output rtl to increment the profiler label LABELNO |
7972 | for profiling a function entry. */ | |
2be15d0f RH |
7973 | |
7974 | void | |
fc27d102 | 7975 | sparc_profile_hook (int labelno) |
2be15d0f RH |
7976 | { |
7977 | char buf[32]; | |
89a8b315 | 7978 | rtx lab, fun; |
2be15d0f | 7979 | |
89a8b315 | 7980 | fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION); |
cc44abe6 DM |
7981 | if (NO_PROFILE_COUNTERS) |
7982 | { | |
7983 | emit_library_call (fun, LCT_NORMAL, VOIDmode, 0); | |
7984 | } | |
7985 | else | |
7986 | { | |
7987 | ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno); | |
7988 | lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf)); | |
7989 | emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode); | |
7990 | } | |
2be15d0f | 7991 | } |
89a8b315 | 7992 | \f |
d1ced155 EB |
7993 | /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */ |
7994 | ||
7c262518 | 7995 | static void |
d1ced155 EB |
7996 | sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags, |
7997 | tree decl ATTRIBUTE_UNUSED) | |
7c262518 RH |
7998 | { |
7999 | fprintf (asm_out_file, "\t.section\t\"%s\"", name); | |
8000 | ||
8001 | if (!(flags & SECTION_DEBUG)) | |
8002 | fputs (",#alloc", asm_out_file); | |
8003 | if (flags & SECTION_WRITE) | |
8004 | fputs (",#write", asm_out_file); | |
2f3321ca EB |
8005 | if (flags & SECTION_TLS) |
8006 | fputs (",#tls", asm_out_file); | |
7c262518 RH |
8007 | if (flags & SECTION_CODE) |
8008 | fputs (",#execinstr", asm_out_file); | |
8009 | ||
8010 | /* ??? Handle SECTION_BSS. */ | |
8011 | ||
8012 | fputc ('\n', asm_out_file); | |
8013 | } | |
f451b552 | 8014 | |
60b85c4c | 8015 | /* We do not allow indirect calls to be optimized into sibling calls. |
951661a1 EB |
8016 | |
8017 | We cannot use sibling calls when delayed branches are disabled | |
8018 | because they will likely require the call delay slot to be filled. | |
8019 | ||
60b85c4c | 8020 | Also, on SPARC 32-bit we cannot emit a sibling call when the |
4977bab6 ZW |
8021 | current function returns a structure. This is because the "unimp |
8022 | after call" convention would cause the callee to return to the | |
8023 | wrong place. The generic code already disallows cases where the | |
8024 | function being called returns a structure. | |
8025 | ||
8026 | It may seem strange how this last case could occur. Usually there | |
8027 | is code after the call which jumps to epilogue code which dumps the | |
8028 | return value into the struct return area. That ought to invalidate | |
951661a1 | 8029 | the sibling call right? Well, in the C++ case we can end up passing |
4977bab6 ZW |
8030 | the pointer to the struct return area to a constructor (which returns |
8031 | void) and then nothing else happens. Such a sibling call would look | |
1910440e RS |
8032 | valid without the added check here. |
8033 | ||
8034 | VxWorks PIC PLT entries require the global pointer to be initialized | |
8035 | on entry. We therefore can't emit sibling calls to them. */ | |
4977bab6 | 8036 | static bool |
fc27d102 | 8037 | sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
4977bab6 | 8038 | { |
951661a1 EB |
8039 | return (decl |
8040 | && flag_delayed_branch | |
e3b5732b | 8041 | && (TARGET_ARCH64 || ! cfun->returns_struct) |
1910440e RS |
8042 | && !(TARGET_VXWORKS_RTP |
8043 | && flag_pic | |
8044 | && !targetm.binds_local_p (decl))); | |
4977bab6 | 8045 | } |
c15c90bb ZW |
8046 | \f |
8047 | /* libfunc renaming. */ | |
8048 | #include "config/gofast.h" | |
8049 | ||
8050 | static void | |
8051 | sparc_init_libfuncs (void) | |
8052 | { | |
8053 | if (TARGET_ARCH32) | |
8054 | { | |
8055 | /* Use the subroutines that Sun's library provides for integer | |
8056 | multiply and divide. The `*' prevents an underscore from | |
8057 | being prepended by the compiler. .umul is a little faster | |
ff482c8d | 8058 | than .mul. */ |
c15c90bb ZW |
8059 | set_optab_libfunc (smul_optab, SImode, "*.umul"); |
8060 | set_optab_libfunc (sdiv_optab, SImode, "*.div"); | |
8061 | set_optab_libfunc (udiv_optab, SImode, "*.udiv"); | |
8062 | set_optab_libfunc (smod_optab, SImode, "*.rem"); | |
8063 | set_optab_libfunc (umod_optab, SImode, "*.urem"); | |
8064 | ||
8065 | /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */ | |
8066 | set_optab_libfunc (add_optab, TFmode, "_Q_add"); | |
8067 | set_optab_libfunc (sub_optab, TFmode, "_Q_sub"); | |
8068 | set_optab_libfunc (neg_optab, TFmode, "_Q_neg"); | |
8069 | set_optab_libfunc (smul_optab, TFmode, "_Q_mul"); | |
8070 | set_optab_libfunc (sdiv_optab, TFmode, "_Q_div"); | |
8071 | ||
8072 | /* We can define the TFmode sqrt optab only if TARGET_FPU. This | |
8073 | is because with soft-float, the SFmode and DFmode sqrt | |
8074 | instructions will be absent, and the compiler will notice and | |
8075 | try to use the TFmode sqrt instruction for calls to the | |
8076 | builtin function sqrt, but this fails. */ | |
8077 | if (TARGET_FPU) | |
8078 | set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt"); | |
8079 | ||
c9034561 ZW |
8080 | set_optab_libfunc (eq_optab, TFmode, "_Q_feq"); |
8081 | set_optab_libfunc (ne_optab, TFmode, "_Q_fne"); | |
8082 | set_optab_libfunc (gt_optab, TFmode, "_Q_fgt"); | |
8083 | set_optab_libfunc (ge_optab, TFmode, "_Q_fge"); | |
8084 | set_optab_libfunc (lt_optab, TFmode, "_Q_flt"); | |
8085 | set_optab_libfunc (le_optab, TFmode, "_Q_fle"); | |
c15c90bb | 8086 | |
85363ca0 ZW |
8087 | set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq"); |
8088 | set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq"); | |
8089 | set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos"); | |
8090 | set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod"); | |
8091 | ||
8092 | set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi"); | |
8093 | set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou"); | |
8094 | set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq"); | |
9b336fa4 | 8095 | set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq"); |
c15c90bb | 8096 | |
45dcc026 JJ |
8097 | if (DITF_CONVERSION_LIBFUNCS) |
8098 | { | |
8099 | set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll"); | |
8100 | set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull"); | |
8101 | set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq"); | |
9b336fa4 | 8102 | set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq"); |
45dcc026 JJ |
8103 | } |
8104 | ||
c15c90bb ZW |
8105 | if (SUN_CONVERSION_LIBFUNCS) |
8106 | { | |
85363ca0 ZW |
8107 | set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll"); |
8108 | set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull"); | |
8109 | set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll"); | |
8110 | set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull"); | |
c15c90bb ZW |
8111 | } |
8112 | } | |
8113 | if (TARGET_ARCH64) | |
8114 | { | |
8115 | /* In the SPARC 64bit ABI, SImode multiply and divide functions | |
8116 | do not exist in the library. Make sure the compiler does not | |
8117 | emit calls to them by accident. (It should always use the | |
8118 | hardware instructions.) */ | |
8119 | set_optab_libfunc (smul_optab, SImode, 0); | |
8120 | set_optab_libfunc (sdiv_optab, SImode, 0); | |
8121 | set_optab_libfunc (udiv_optab, SImode, 0); | |
8122 | set_optab_libfunc (smod_optab, SImode, 0); | |
8123 | set_optab_libfunc (umod_optab, SImode, 0); | |
8124 | ||
8125 | if (SUN_INTEGER_MULTIPLY_64) | |
8126 | { | |
8127 | set_optab_libfunc (smul_optab, DImode, "__mul64"); | |
8128 | set_optab_libfunc (sdiv_optab, DImode, "__div64"); | |
8129 | set_optab_libfunc (udiv_optab, DImode, "__udiv64"); | |
8130 | set_optab_libfunc (smod_optab, DImode, "__rem64"); | |
8131 | set_optab_libfunc (umod_optab, DImode, "__urem64"); | |
8132 | } | |
8133 | ||
8134 | if (SUN_CONVERSION_LIBFUNCS) | |
8135 | { | |
85363ca0 ZW |
8136 | set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol"); |
8137 | set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul"); | |
8138 | set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol"); | |
8139 | set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul"); | |
c15c90bb ZW |
8140 | } |
8141 | } | |
4977bab6 | 8142 | |
c15c90bb ZW |
8143 | gofast_maybe_init_libfuncs (); |
8144 | } | |
8145 | \f | |
16f59241 | 8146 | #define def_builtin(NAME, CODE, TYPE) \ |
c79efc4d RÁE |
8147 | add_builtin_function((NAME), (TYPE), (CODE), BUILT_IN_MD, NULL, \ |
8148 | NULL_TREE) | |
16f59241 JM |
8149 | |
8150 | /* Implement the TARGET_INIT_BUILTINS target hook. | |
8151 | Create builtin functions for special SPARC instructions. */ | |
8152 | ||
8153 | static void | |
8154 | sparc_init_builtins (void) | |
8155 | { | |
8156 | if (TARGET_VIS) | |
8157 | sparc_vis_init_builtins (); | |
8158 | } | |
8159 | ||
8160 | /* Create builtin functions for VIS 1.0 instructions. */ | |
8161 | ||
8162 | static void | |
8163 | sparc_vis_init_builtins (void) | |
8164 | { | |
8165 | tree v4qi = build_vector_type (unsigned_intQI_type_node, 4); | |
8166 | tree v8qi = build_vector_type (unsigned_intQI_type_node, 8); | |
8167 | tree v4hi = build_vector_type (intHI_type_node, 4); | |
8168 | tree v2hi = build_vector_type (intHI_type_node, 2); | |
8169 | tree v2si = build_vector_type (intSI_type_node, 2); | |
8170 | ||
8171 | tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0); | |
8172 | tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0); | |
8173 | tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0); | |
8174 | tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0); | |
8175 | tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0); | |
8176 | tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0); | |
8177 | tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0); | |
8178 | tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0); | |
8179 | tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0); | |
8180 | tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0); | |
8181 | tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0); | |
8182 | tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0); | |
8183 | tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node, | |
8184 | v8qi, v8qi, | |
8185 | intDI_type_node, 0); | |
8186 | tree di_ftype_di_di = build_function_type_list (intDI_type_node, | |
8187 | intDI_type_node, | |
8188 | intDI_type_node, 0); | |
8189 | tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node, | |
8190 | ptr_type_node, | |
8191 | intSI_type_node, 0); | |
8192 | tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node, | |
8193 | ptr_type_node, | |
8194 | intDI_type_node, 0); | |
8195 | ||
8196 | /* Packing and expanding vectors. */ | |
8197 | def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis, v4qi_ftype_v4hi); | |
8198 | def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis, | |
8199 | v8qi_ftype_v2si_v8qi); | |
8200 | def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis, | |
8201 | v2hi_ftype_v2si); | |
8202 | def_builtin ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis, v4hi_ftype_v4qi); | |
8203 | def_builtin ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis, | |
8204 | v8qi_ftype_v4qi_v4qi); | |
8205 | ||
8206 | /* Multiplications. */ | |
8207 | def_builtin ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis, | |
8208 | v4hi_ftype_v4qi_v4hi); | |
8209 | def_builtin ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis, | |
8210 | v4hi_ftype_v4qi_v2hi); | |
8211 | def_builtin ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis, | |
8212 | v4hi_ftype_v4qi_v2hi); | |
8213 | def_builtin ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis, | |
8214 | v4hi_ftype_v8qi_v4hi); | |
8215 | def_builtin ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis, | |
8216 | v4hi_ftype_v8qi_v4hi); | |
8217 | def_builtin ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis, | |
8218 | v2si_ftype_v4qi_v2hi); | |
8219 | def_builtin ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis, | |
8220 | v2si_ftype_v4qi_v2hi); | |
8221 | ||
8222 | /* Data aligning. */ | |
8223 | def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis, | |
8224 | v4hi_ftype_v4hi_v4hi); | |
8225 | def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis, | |
8226 | v8qi_ftype_v8qi_v8qi); | |
8227 | def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis, | |
8228 | v2si_ftype_v2si_v2si); | |
8229 | def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatadi_vis, | |
8230 | di_ftype_di_di); | |
8231 | if (TARGET_ARCH64) | |
8232 | def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis, | |
8233 | ptr_ftype_ptr_di); | |
8234 | else | |
8235 | def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis, | |
8236 | ptr_ftype_ptr_si); | |
8237 | ||
8238 | /* Pixel distance. */ | |
8239 | def_builtin ("__builtin_vis_pdist", CODE_FOR_pdist_vis, | |
8240 | di_ftype_v8qi_v8qi_di); | |
8241 | } | |
8242 | ||
8243 | /* Handle TARGET_EXPAND_BUILTIN target hook. | |
0fa2e4df | 8244 | Expand builtin functions for sparc intrinsics. */ |
16f59241 JM |
8245 | |
8246 | static rtx | |
724d7c55 EB |
8247 | sparc_expand_builtin (tree exp, rtx target, |
8248 | rtx subtarget ATTRIBUTE_UNUSED, | |
8249 | enum machine_mode tmode ATTRIBUTE_UNUSED, | |
8250 | int ignore ATTRIBUTE_UNUSED) | |
16f59241 | 8251 | { |
5039610b SL |
8252 | tree arg; |
8253 | call_expr_arg_iterator iter; | |
8254 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
16f59241 JM |
8255 | unsigned int icode = DECL_FUNCTION_CODE (fndecl); |
8256 | rtx pat, op[4]; | |
8257 | enum machine_mode mode[4]; | |
8258 | int arg_count = 0; | |
8259 | ||
724d7c55 EB |
8260 | mode[0] = insn_data[icode].operand[0].mode; |
8261 | if (!target | |
8262 | || GET_MODE (target) != mode[0] | |
8263 | || ! (*insn_data[icode].operand[0].predicate) (target, mode[0])) | |
8264 | op[0] = gen_reg_rtx (mode[0]); | |
16f59241 | 8265 | else |
724d7c55 | 8266 | op[0] = target; |
16f59241 | 8267 | |
5039610b | 8268 | FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) |
16f59241 | 8269 | { |
16f59241 JM |
8270 | arg_count++; |
8271 | mode[arg_count] = insn_data[icode].operand[arg_count].mode; | |
84217346 | 8272 | op[arg_count] = expand_normal (arg); |
16f59241 JM |
8273 | |
8274 | if (! (*insn_data[icode].operand[arg_count].predicate) (op[arg_count], | |
8275 | mode[arg_count])) | |
8276 | op[arg_count] = copy_to_mode_reg (mode[arg_count], op[arg_count]); | |
8277 | } | |
8278 | ||
8279 | switch (arg_count) | |
8280 | { | |
8281 | case 1: | |
8282 | pat = GEN_FCN (icode) (op[0], op[1]); | |
8283 | break; | |
8284 | case 2: | |
8285 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
8286 | break; | |
8287 | case 3: | |
8288 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
8289 | break; | |
8290 | default: | |
8291 | gcc_unreachable (); | |
8292 | } | |
8293 | ||
8294 | if (!pat) | |
8295 | return NULL_RTX; | |
8296 | ||
8297 | emit_insn (pat); | |
8298 | ||
8299 | return op[0]; | |
8300 | } | |
8b0096b4 JM |
8301 | |
8302 | static int | |
8303 | sparc_vis_mul8x16 (int e8, int e16) | |
8304 | { | |
8305 | return (e8 * e16 + 128) / 256; | |
8306 | } | |
8307 | ||
8308 | /* Multiply the vector elements in ELTS0 to the elements in ELTS1 as specified | |
8309 | by FNCODE. All of the elements in ELTS0 and ELTS1 lists must be integer | |
8310 | constants. A tree list with the results of the multiplications is returned, | |
8311 | and each element in the list is of INNER_TYPE. */ | |
8312 | ||
8313 | static tree | |
8314 | sparc_handle_vis_mul8x16 (int fncode, tree inner_type, tree elts0, tree elts1) | |
8315 | { | |
8316 | tree n_elts = NULL_TREE; | |
8317 | int scale; | |
8318 | ||
8319 | switch (fncode) | |
8320 | { | |
8321 | case CODE_FOR_fmul8x16_vis: | |
8322 | for (; elts0 && elts1; | |
8323 | elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1)) | |
8324 | { | |
8325 | int val | |
8326 | = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)), | |
8327 | TREE_INT_CST_LOW (TREE_VALUE (elts1))); | |
8328 | n_elts = tree_cons (NULL_TREE, | |
8329 | build_int_cst (inner_type, val), | |
8330 | n_elts); | |
8331 | } | |
8332 | break; | |
8333 | ||
8334 | case CODE_FOR_fmul8x16au_vis: | |
8335 | scale = TREE_INT_CST_LOW (TREE_VALUE (elts1)); | |
8336 | ||
8337 | for (; elts0; elts0 = TREE_CHAIN (elts0)) | |
8338 | { | |
8339 | int val | |
8340 | = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)), | |
8341 | scale); | |
8342 | n_elts = tree_cons (NULL_TREE, | |
8343 | build_int_cst (inner_type, val), | |
8344 | n_elts); | |
8345 | } | |
8346 | break; | |
8347 | ||
8348 | case CODE_FOR_fmul8x16al_vis: | |
8349 | scale = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (elts1))); | |
8350 | ||
8351 | for (; elts0; elts0 = TREE_CHAIN (elts0)) | |
8352 | { | |
8353 | int val | |
8354 | = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)), | |
8355 | scale); | |
8356 | n_elts = tree_cons (NULL_TREE, | |
8357 | build_int_cst (inner_type, val), | |
8358 | n_elts); | |
8359 | } | |
8360 | break; | |
8361 | ||
8362 | default: | |
8363 | gcc_unreachable (); | |
8364 | } | |
8365 | ||
8366 | return nreverse (n_elts); | |
8367 | ||
8368 | } | |
8369 | /* Handle TARGET_FOLD_BUILTIN target hook. | |
607fb860 | 8370 | Fold builtin functions for SPARC intrinsics. If IGNORE is true the |
8b0096b4 JM |
8371 | result of the function call is ignored. NULL_TREE is returned if the |
8372 | function could not be folded. */ | |
8373 | ||
8374 | static tree | |
f311c3b4 NF |
8375 | sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED, |
8376 | tree *args, bool ignore) | |
8b0096b4 JM |
8377 | { |
8378 | tree arg0, arg1, arg2; | |
8379 | tree rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
bbbbb16a | 8380 | enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl); |
8b0096b4 | 8381 | |
724d7c55 | 8382 | if (ignore |
bbbbb16a ILT |
8383 | && icode != CODE_FOR_alignaddrsi_vis |
8384 | && icode != CODE_FOR_alignaddrdi_vis) | |
724d7c55 | 8385 | return fold_convert (rtype, integer_zero_node); |
8b0096b4 | 8386 | |
bbbbb16a | 8387 | switch (icode) |
8b0096b4 JM |
8388 | { |
8389 | case CODE_FOR_fexpand_vis: | |
f311c3b4 | 8390 | arg0 = args[0]; |
8b0096b4 JM |
8391 | STRIP_NOPS (arg0); |
8392 | ||
8393 | if (TREE_CODE (arg0) == VECTOR_CST) | |
8394 | { | |
8395 | tree inner_type = TREE_TYPE (rtype); | |
8396 | tree elts = TREE_VECTOR_CST_ELTS (arg0); | |
8397 | tree n_elts = NULL_TREE; | |
8398 | ||
8399 | for (; elts; elts = TREE_CHAIN (elts)) | |
8400 | { | |
8401 | unsigned int val = TREE_INT_CST_LOW (TREE_VALUE (elts)) << 4; | |
8402 | n_elts = tree_cons (NULL_TREE, | |
8403 | build_int_cst (inner_type, val), | |
8404 | n_elts); | |
8405 | } | |
8406 | return build_vector (rtype, nreverse (n_elts)); | |
8407 | } | |
8408 | break; | |
8409 | ||
8410 | case CODE_FOR_fmul8x16_vis: | |
8411 | case CODE_FOR_fmul8x16au_vis: | |
8412 | case CODE_FOR_fmul8x16al_vis: | |
f311c3b4 NF |
8413 | arg0 = args[0]; |
8414 | arg1 = args[1]; | |
8b0096b4 JM |
8415 | STRIP_NOPS (arg0); |
8416 | STRIP_NOPS (arg1); | |
8417 | ||
8418 | if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST) | |
8419 | { | |
8420 | tree inner_type = TREE_TYPE (rtype); | |
8421 | tree elts0 = TREE_VECTOR_CST_ELTS (arg0); | |
8422 | tree elts1 = TREE_VECTOR_CST_ELTS (arg1); | |
bbbbb16a ILT |
8423 | tree n_elts = sparc_handle_vis_mul8x16 (icode, inner_type, elts0, |
8424 | elts1); | |
8b0096b4 JM |
8425 | |
8426 | return build_vector (rtype, n_elts); | |
8427 | } | |
8428 | break; | |
8429 | ||
8430 | case CODE_FOR_fpmerge_vis: | |
f311c3b4 NF |
8431 | arg0 = args[0]; |
8432 | arg1 = args[1]; | |
8b0096b4 JM |
8433 | STRIP_NOPS (arg0); |
8434 | STRIP_NOPS (arg1); | |
8435 | ||
8436 | if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST) | |
8437 | { | |
8438 | tree elts0 = TREE_VECTOR_CST_ELTS (arg0); | |
8439 | tree elts1 = TREE_VECTOR_CST_ELTS (arg1); | |
8440 | tree n_elts = NULL_TREE; | |
8441 | ||
8442 | for (; elts0 && elts1; | |
8443 | elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1)) | |
8444 | { | |
8445 | n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts0), n_elts); | |
8446 | n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts1), n_elts); | |
8447 | } | |
8448 | ||
8449 | return build_vector (rtype, nreverse (n_elts)); | |
8450 | } | |
8451 | break; | |
8452 | ||
8453 | case CODE_FOR_pdist_vis: | |
f311c3b4 NF |
8454 | arg0 = args[0]; |
8455 | arg1 = args[1]; | |
8456 | arg2 = args[2]; | |
8b0096b4 JM |
8457 | STRIP_NOPS (arg0); |
8458 | STRIP_NOPS (arg1); | |
8459 | STRIP_NOPS (arg2); | |
8460 | ||
8461 | if (TREE_CODE (arg0) == VECTOR_CST | |
8462 | && TREE_CODE (arg1) == VECTOR_CST | |
8463 | && TREE_CODE (arg2) == INTEGER_CST) | |
8464 | { | |
8465 | int overflow = 0; | |
8466 | unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2); | |
8467 | HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2); | |
8468 | tree elts0 = TREE_VECTOR_CST_ELTS (arg0); | |
8469 | tree elts1 = TREE_VECTOR_CST_ELTS (arg1); | |
8470 | ||
8471 | for (; elts0 && elts1; | |
8472 | elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1)) | |
8473 | { | |
8474 | unsigned HOST_WIDE_INT | |
8475 | low0 = TREE_INT_CST_LOW (TREE_VALUE (elts0)), | |
8476 | low1 = TREE_INT_CST_LOW (TREE_VALUE (elts1)); | |
8477 | HOST_WIDE_INT high0 = TREE_INT_CST_HIGH (TREE_VALUE (elts0)); | |
8478 | HOST_WIDE_INT high1 = TREE_INT_CST_HIGH (TREE_VALUE (elts1)); | |
8479 | ||
8480 | unsigned HOST_WIDE_INT l; | |
8481 | HOST_WIDE_INT h; | |
8482 | ||
8483 | overflow |= neg_double (low1, high1, &l, &h); | |
8484 | overflow |= add_double (low0, high0, l, h, &l, &h); | |
8485 | if (h < 0) | |
8486 | overflow |= neg_double (l, h, &l, &h); | |
8487 | ||
8488 | overflow |= add_double (low, high, l, h, &low, &high); | |
8489 | } | |
8490 | ||
8491 | gcc_assert (overflow == 0); | |
8492 | ||
8493 | return build_int_cst_wide (rtype, low, high); | |
8494 | } | |
8495 | ||
8496 | default: | |
8497 | break; | |
8498 | } | |
724d7c55 | 8499 | |
8b0096b4 JM |
8500 | return NULL_TREE; |
8501 | } | |
16f59241 | 8502 | \f |
214c6394 DM |
8503 | /* ??? This duplicates information provided to the compiler by the |
8504 | ??? scheduler description. Some day, teach genautomata to output | |
8505 | ??? the latencies and then CSE will just use that. */ | |
8506 | ||
3c50106f | 8507 | static bool |
f40751dd JH |
8508 | sparc_rtx_costs (rtx x, int code, int outer_code, int *total, |
8509 | bool speed ATTRIBUTE_UNUSED) | |
2ce04b6b | 8510 | { |
18df6de9 DM |
8511 | enum machine_mode mode = GET_MODE (x); |
8512 | bool float_mode_p = FLOAT_MODE_P (mode); | |
8513 | ||
2ce04b6b DM |
8514 | switch (code) |
8515 | { | |
18df6de9 DM |
8516 | case CONST_INT: |
8517 | if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000) | |
214c6394 | 8518 | { |
18df6de9 | 8519 | *total = 0; |
3c50106f | 8520 | return true; |
214c6394 | 8521 | } |
18df6de9 | 8522 | /* FALLTHRU */ |
214c6394 | 8523 | |
18df6de9 DM |
8524 | case HIGH: |
8525 | *total = 2; | |
8526 | return true; | |
214c6394 | 8527 | |
18df6de9 DM |
8528 | case CONST: |
8529 | case LABEL_REF: | |
8530 | case SYMBOL_REF: | |
8531 | *total = 4; | |
8532 | return true; | |
214c6394 | 8533 | |
18df6de9 | 8534 | case CONST_DOUBLE: |
ca075703 DM |
8535 | if (GET_MODE (x) == VOIDmode |
8536 | && ((CONST_DOUBLE_HIGH (x) == 0 | |
8537 | && CONST_DOUBLE_LOW (x) < 0x1000) | |
8538 | || (CONST_DOUBLE_HIGH (x) == -1 | |
8539 | && CONST_DOUBLE_LOW (x) < 0 | |
8540 | && CONST_DOUBLE_LOW (x) >= -0x1000))) | |
18df6de9 DM |
8541 | *total = 0; |
8542 | else | |
8543 | *total = 8; | |
3c50106f | 8544 | return true; |
214c6394 | 8545 | |
18df6de9 DM |
8546 | case MEM: |
8547 | /* If outer-code was a sign or zero extension, a cost | |
8548 | of COSTS_N_INSNS (1) was already added in. This is | |
8549 | why we are subtracting it back out. */ | |
8550 | if (outer_code == ZERO_EXTEND) | |
214c6394 | 8551 | { |
18df6de9 | 8552 | *total = sparc_costs->int_zload - COSTS_N_INSNS (1); |
214c6394 | 8553 | } |
18df6de9 | 8554 | else if (outer_code == SIGN_EXTEND) |
3c50106f | 8555 | { |
18df6de9 | 8556 | *total = sparc_costs->int_sload - COSTS_N_INSNS (1); |
3c50106f | 8557 | } |
18df6de9 | 8558 | else if (float_mode_p) |
3c50106f | 8559 | { |
18df6de9 | 8560 | *total = sparc_costs->float_load; |
3c50106f | 8561 | } |
18df6de9 | 8562 | else |
3c50106f | 8563 | { |
18df6de9 | 8564 | *total = sparc_costs->int_load; |
3c50106f | 8565 | } |
214c6394 | 8566 | |
3c50106f | 8567 | return true; |
2ce04b6b | 8568 | |
18df6de9 DM |
8569 | case PLUS: |
8570 | case MINUS: | |
8571 | if (float_mode_p) | |
8572 | *total = sparc_costs->float_plusminus; | |
8573 | else | |
8574 | *total = COSTS_N_INSNS (1); | |
8575 | return false; | |
214c6394 | 8576 | |
18df6de9 DM |
8577 | case MULT: |
8578 | if (float_mode_p) | |
8579 | *total = sparc_costs->float_mul; | |
e08eb33c | 8580 | else if (! TARGET_HARD_MUL) |
bf6c40e9 | 8581 | *total = COSTS_N_INSNS (25); |
18df6de9 DM |
8582 | else |
8583 | { | |
8584 | int bit_cost; | |
214c6394 | 8585 | |
18df6de9 DM |
8586 | bit_cost = 0; |
8587 | if (sparc_costs->int_mul_bit_factor) | |
8588 | { | |
8589 | int nbits; | |
214c6394 | 8590 | |
18df6de9 DM |
8591 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) |
8592 | { | |
8593 | unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1)); | |
8594 | for (nbits = 0; value != 0; value &= value - 1) | |
8595 | nbits++; | |
8596 | } | |
8597 | else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE | |
ca075703 | 8598 | && GET_MODE (XEXP (x, 1)) == VOIDmode) |
18df6de9 DM |
8599 | { |
8600 | rtx x1 = XEXP (x, 1); | |
ca075703 DM |
8601 | unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1); |
8602 | unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1); | |
18df6de9 DM |
8603 | |
8604 | for (nbits = 0; value1 != 0; value1 &= value1 - 1) | |
8605 | nbits++; | |
8606 | for (; value2 != 0; value2 &= value2 - 1) | |
8607 | nbits++; | |
8608 | } | |
214c6394 | 8609 | else |
18df6de9 | 8610 | nbits = 7; |
214c6394 | 8611 | |
18df6de9 DM |
8612 | if (nbits < 3) |
8613 | nbits = 3; | |
8614 | bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor; | |
ffe384ff | 8615 | bit_cost = COSTS_N_INSNS (bit_cost); |
214c6394 | 8616 | } |
18df6de9 DM |
8617 | |
8618 | if (mode == DImode) | |
ffe384ff | 8619 | *total = sparc_costs->int_mulX + bit_cost; |
18df6de9 | 8620 | else |
ffe384ff | 8621 | *total = sparc_costs->int_mul + bit_cost; |
214c6394 | 8622 | } |
18df6de9 | 8623 | return false; |
214c6394 | 8624 | |
18df6de9 DM |
8625 | case ASHIFT: |
8626 | case ASHIFTRT: | |
8627 | case LSHIFTRT: | |
8628 | *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty; | |
8629 | return false; | |
2ce04b6b | 8630 | |
18df6de9 DM |
8631 | case DIV: |
8632 | case UDIV: | |
8633 | case MOD: | |
8634 | case UMOD: | |
8635 | if (float_mode_p) | |
214c6394 | 8636 | { |
18df6de9 DM |
8637 | if (mode == DFmode) |
8638 | *total = sparc_costs->float_div_df; | |
214c6394 | 8639 | else |
18df6de9 | 8640 | *total = sparc_costs->float_div_sf; |
214c6394 | 8641 | } |
18df6de9 | 8642 | else |
214c6394 | 8643 | { |
18df6de9 DM |
8644 | if (mode == DImode) |
8645 | *total = sparc_costs->int_divX; | |
214c6394 | 8646 | else |
18df6de9 | 8647 | *total = sparc_costs->int_div; |
214c6394 | 8648 | } |
18df6de9 | 8649 | return false; |
2ce04b6b | 8650 | |
18df6de9 DM |
8651 | case NEG: |
8652 | if (! float_mode_p) | |
3c50106f | 8653 | { |
18df6de9 DM |
8654 | *total = COSTS_N_INSNS (1); |
8655 | return false; | |
3c50106f | 8656 | } |
5efb1046 | 8657 | /* FALLTHRU */ |
2ce04b6b | 8658 | |
18df6de9 DM |
8659 | case ABS: |
8660 | case FLOAT: | |
8661 | case UNSIGNED_FLOAT: | |
8662 | case FIX: | |
8663 | case UNSIGNED_FIX: | |
8664 | case FLOAT_EXTEND: | |
8665 | case FLOAT_TRUNCATE: | |
8666 | *total = sparc_costs->float_move; | |
8667 | return false; | |
2ce04b6b | 8668 | |
18df6de9 DM |
8669 | case SQRT: |
8670 | if (mode == DFmode) | |
8671 | *total = sparc_costs->float_sqrt_df; | |
8672 | else | |
8673 | *total = sparc_costs->float_sqrt_sf; | |
8674 | return false; | |
2ce04b6b | 8675 | |
18df6de9 DM |
8676 | case COMPARE: |
8677 | if (float_mode_p) | |
8678 | *total = sparc_costs->float_cmp; | |
3c50106f | 8679 | else |
18df6de9 DM |
8680 | *total = COSTS_N_INSNS (1); |
8681 | return false; | |
8682 | ||
8683 | case IF_THEN_ELSE: | |
8684 | if (float_mode_p) | |
8685 | *total = sparc_costs->float_cmove; | |
8686 | else | |
8687 | *total = sparc_costs->int_cmove; | |
8688 | return false; | |
2ce04b6b | 8689 | |
893e18a5 JM |
8690 | case IOR: |
8691 | /* Handle the NAND vector patterns. */ | |
8692 | if (sparc_vector_mode_supported_p (GET_MODE (x)) | |
8693 | && GET_CODE (XEXP (x, 0)) == NOT | |
8694 | && GET_CODE (XEXP (x, 1)) == NOT) | |
8695 | { | |
8696 | *total = COSTS_N_INSNS (1); | |
8697 | return true; | |
8698 | } | |
8699 | else | |
8700 | return false; | |
8701 | ||
2ce04b6b | 8702 | default: |
3c50106f RH |
8703 | return false; |
8704 | } | |
2ce04b6b | 8705 | } |
fb49053f | 8706 | |
68ba34f3 AS |
8707 | /* Emit the sequence of insns SEQ while preserving the registers REG and REG2. |
8708 | This is achieved by means of a manual dynamic stack space allocation in | |
8709 | the current frame. We make the assumption that SEQ doesn't contain any | |
8710 | function calls, with the possible exception of calls to the PIC helper. */ | |
951661a1 EB |
8711 | |
8712 | static void | |
1b970c28 | 8713 | emit_and_preserve (rtx seq, rtx reg, rtx reg2) |
951661a1 | 8714 | { |
68ba34f3 AS |
8715 | /* We must preserve the lowest 16 words for the register save area. */ |
8716 | HOST_WIDE_INT offset = 16*UNITS_PER_WORD; | |
8717 | /* We really need only 2 words of fresh stack space. */ | |
8718 | HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD); | |
951661a1 | 8719 | |
68ba34f3 AS |
8720 | rtx slot |
8721 | = gen_rtx_MEM (word_mode, plus_constant (stack_pointer_rtx, | |
8722 | SPARC_STACK_BIAS + offset)); | |
8723 | ||
8724 | emit_insn (gen_stack_pointer_dec (GEN_INT (size))); | |
951661a1 | 8725 | emit_insn (gen_rtx_SET (VOIDmode, slot, reg)); |
1b970c28 EB |
8726 | if (reg2) |
8727 | emit_insn (gen_rtx_SET (VOIDmode, | |
8728 | adjust_address (slot, word_mode, UNITS_PER_WORD), | |
8729 | reg2)); | |
951661a1 | 8730 | emit_insn (seq); |
1b970c28 EB |
8731 | if (reg2) |
8732 | emit_insn (gen_rtx_SET (VOIDmode, | |
8733 | reg2, | |
8734 | adjust_address (slot, word_mode, UNITS_PER_WORD))); | |
951661a1 | 8735 | emit_insn (gen_rtx_SET (VOIDmode, reg, slot)); |
68ba34f3 | 8736 | emit_insn (gen_stack_pointer_inc (GEN_INT (size))); |
951661a1 EB |
8737 | } |
8738 | ||
b9bf58e5 EB |
8739 | /* Output the assembler code for a thunk function. THUNK_DECL is the |
8740 | declaration for the thunk function itself, FUNCTION is the decl for | |
8741 | the target function. DELTA is an immediate constant offset to be | |
8742 | added to THIS. If VCALL_OFFSET is nonzero, the word at address | |
8743 | (*THIS + VCALL_OFFSET) should be additionally added to THIS. */ | |
e133041b | 8744 | |
c590b625 | 8745 | static void |
fc27d102 | 8746 | sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
b9bf58e5 | 8747 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, |
fc27d102 | 8748 | tree function) |
e133041b | 8749 | { |
0a2aaacc | 8750 | rtx this_rtx, insn, funexp; |
951661a1 | 8751 | unsigned int int_arg_first; |
e133041b RH |
8752 | |
8753 | reload_completed = 1; | |
fe3ad572 | 8754 | epilogue_completed = 1; |
e133041b | 8755 | |
2e040219 | 8756 | emit_note (NOTE_INSN_PROLOGUE_END); |
e133041b | 8757 | |
951661a1 EB |
8758 | if (flag_delayed_branch) |
8759 | { | |
8760 | /* We will emit a regular sibcall below, so we need to instruct | |
8761 | output_sibcall that we are in a leaf function. */ | |
22d66adf | 8762 | sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1; |
951661a1 EB |
8763 | |
8764 | /* This will cause final.c to invoke leaf_renumber_regs so we | |
8765 | must behave as if we were in a not-yet-leafified function. */ | |
8766 | int_arg_first = SPARC_INCOMING_INT_ARG_FIRST; | |
8767 | } | |
8768 | else | |
8769 | { | |
8770 | /* We will emit the sibcall manually below, so we will need to | |
8771 | manually spill non-leaf registers. */ | |
22d66adf | 8772 | sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0; |
951661a1 EB |
8773 | |
8774 | /* We really are in a leaf function. */ | |
8775 | int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST; | |
8776 | } | |
8777 | ||
e133041b RH |
8778 | /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function |
8779 | returns a structure, the structure return pointer is there instead. */ | |
4900aa61 EB |
8780 | if (TARGET_ARCH64 |
8781 | && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) | |
0a2aaacc | 8782 | this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1); |
e133041b | 8783 | else |
0a2aaacc | 8784 | this_rtx = gen_rtx_REG (Pmode, int_arg_first); |
e133041b RH |
8785 | |
8786 | /* Add DELTA. When possible use a plain add, otherwise load it into | |
8787 | a register first. */ | |
b9bf58e5 | 8788 | if (delta) |
e133041b | 8789 | { |
b9bf58e5 EB |
8790 | rtx delta_rtx = GEN_INT (delta); |
8791 | ||
8792 | if (! SPARC_SIMM13_P (delta)) | |
8793 | { | |
8794 | rtx scratch = gen_rtx_REG (Pmode, 1); | |
8795 | emit_move_insn (scratch, delta_rtx); | |
8796 | delta_rtx = scratch; | |
8797 | } | |
8798 | ||
0a2aaacc KG |
8799 | /* THIS_RTX += DELTA. */ |
8800 | emit_insn (gen_add2_insn (this_rtx, delta_rtx)); | |
b9bf58e5 EB |
8801 | } |
8802 | ||
0a2aaacc | 8803 | /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */ |
b9bf58e5 EB |
8804 | if (vcall_offset) |
8805 | { | |
8806 | rtx vcall_offset_rtx = GEN_INT (vcall_offset); | |
e133041b | 8807 | rtx scratch = gen_rtx_REG (Pmode, 1); |
ed1fe829 | 8808 | |
f5f7d171 | 8809 | gcc_assert (vcall_offset < 0); |
b9bf58e5 | 8810 | |
0a2aaacc KG |
8811 | /* SCRATCH = *THIS_RTX. */ |
8812 | emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx)); | |
b9bf58e5 EB |
8813 | |
8814 | /* Prepare for adding VCALL_OFFSET. The difficulty is that we | |
8815 | may not have any available scratch register at this point. */ | |
8816 | if (SPARC_SIMM13_P (vcall_offset)) | |
8817 | ; | |
8818 | /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */ | |
8819 | else if (! fixed_regs[5] | |
8820 | /* The below sequence is made up of at least 2 insns, | |
8821 | while the default method may need only one. */ | |
8822 | && vcall_offset < -8192) | |
8823 | { | |
8824 | rtx scratch2 = gen_rtx_REG (Pmode, 5); | |
8825 | emit_move_insn (scratch2, vcall_offset_rtx); | |
8826 | vcall_offset_rtx = scratch2; | |
8827 | } | |
e133041b | 8828 | else |
ed1fe829 | 8829 | { |
b9bf58e5 EB |
8830 | rtx increment = GEN_INT (-4096); |
8831 | ||
8832 | /* VCALL_OFFSET is a negative number whose typical range can be | |
8833 | estimated as -32768..0 in 32-bit mode. In almost all cases | |
8834 | it is therefore cheaper to emit multiple add insns than | |
8835 | spilling and loading the constant into a register (at least | |
8836 | 6 insns). */ | |
8837 | while (! SPARC_SIMM13_P (vcall_offset)) | |
8838 | { | |
8839 | emit_insn (gen_add2_insn (scratch, increment)); | |
8840 | vcall_offset += 4096; | |
8841 | } | |
8842 | vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */ | |
ed1fe829 EB |
8843 | } |
8844 | ||
0a2aaacc | 8845 | /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */ |
b9bf58e5 EB |
8846 | emit_move_insn (scratch, gen_rtx_MEM (Pmode, |
8847 | gen_rtx_PLUS (Pmode, | |
8848 | scratch, | |
8849 | vcall_offset_rtx))); | |
e133041b | 8850 | |
0a2aaacc KG |
8851 | /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */ |
8852 | emit_insn (gen_add2_insn (this_rtx, scratch)); | |
b9bf58e5 | 8853 | } |
e133041b RH |
8854 | |
8855 | /* Generate a tail call to the target function. */ | |
8856 | if (! TREE_USED (function)) | |
8857 | { | |
8858 | assemble_external (function); | |
8859 | TREE_USED (function) = 1; | |
8860 | } | |
8861 | funexp = XEXP (DECL_RTL (function), 0); | |
951661a1 EB |
8862 | |
8863 | if (flag_delayed_branch) | |
8864 | { | |
8865 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
8866 | insn = emit_call_insn (gen_sibcall (funexp)); | |
8867 | SIBLING_CALL_P (insn) = 1; | |
8868 | } | |
8869 | else | |
8870 | { | |
8871 | /* The hoops we have to jump through in order to generate a sibcall | |
8872 | without using delay slots... */ | |
1b970c28 | 8873 | rtx spill_reg, spill_reg2, seq, scratch = gen_rtx_REG (Pmode, 1); |
951661a1 EB |
8874 | |
8875 | if (flag_pic) | |
8876 | { | |
8877 | spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */ | |
1b970c28 | 8878 | spill_reg2 = gen_rtx_REG (word_mode, PIC_OFFSET_TABLE_REGNUM); |
951661a1 | 8879 | start_sequence (); |
5e77eb53 EB |
8880 | /* Delay emitting the PIC helper function because it needs to |
8881 | change the section and we are emitting assembly code. */ | |
9fd52c6c | 8882 | load_pic_register (); /* clobbers %o7 */ |
208384fd | 8883 | scratch = legitimize_pic_address (funexp, scratch); |
951661a1 EB |
8884 | seq = get_insns (); |
8885 | end_sequence (); | |
1b970c28 | 8886 | emit_and_preserve (seq, spill_reg, spill_reg2); |
951661a1 EB |
8887 | } |
8888 | else if (TARGET_ARCH32) | |
8889 | { | |
8890 | emit_insn (gen_rtx_SET (VOIDmode, | |
8891 | scratch, | |
8892 | gen_rtx_HIGH (SImode, funexp))); | |
8893 | emit_insn (gen_rtx_SET (VOIDmode, | |
8894 | scratch, | |
8895 | gen_rtx_LO_SUM (SImode, scratch, funexp))); | |
8896 | } | |
8897 | else /* TARGET_ARCH64 */ | |
8898 | { | |
8899 | switch (sparc_cmodel) | |
8900 | { | |
8901 | case CM_MEDLOW: | |
8902 | case CM_MEDMID: | |
8903 | /* The destination can serve as a temporary. */ | |
8904 | sparc_emit_set_symbolic_const64 (scratch, funexp, scratch); | |
8905 | break; | |
8906 | ||
8907 | case CM_MEDANY: | |
8908 | case CM_EMBMEDANY: | |
8909 | /* The destination cannot serve as a temporary. */ | |
8910 | spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */ | |
8911 | start_sequence (); | |
8912 | sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg); | |
8913 | seq = get_insns (); | |
8914 | end_sequence (); | |
1b970c28 | 8915 | emit_and_preserve (seq, spill_reg, 0); |
951661a1 EB |
8916 | break; |
8917 | ||
8918 | default: | |
f5f7d171 | 8919 | gcc_unreachable (); |
951661a1 EB |
8920 | } |
8921 | } | |
8922 | ||
8923 | emit_jump_insn (gen_indirect_jump (scratch)); | |
8924 | } | |
8925 | ||
e133041b RH |
8926 | emit_barrier (); |
8927 | ||
8928 | /* Run just enough of rest_of_compilation to get the insns emitted. | |
8929 | There's not really enough bulk here to make other passes such as | |
8930 | instruction scheduling worth while. Note that use_thunk calls | |
8931 | assemble_start_function and assemble_end_function. */ | |
8932 | insn = get_insns (); | |
55e092c4 | 8933 | insn_locators_alloc (); |
e133041b RH |
8934 | shorten_branches (insn); |
8935 | final_start_function (insn, file, 1); | |
c9d691e9 | 8936 | final (insn, file, 1); |
e133041b RH |
8937 | final_end_function (); |
8938 | ||
8939 | reload_completed = 0; | |
fe3ad572 | 8940 | epilogue_completed = 0; |
e133041b | 8941 | } |
e2500fed | 8942 | |
b9bf58e5 EB |
8943 | /* Return true if sparc_output_mi_thunk would be able to output the |
8944 | assembler code for the thunk function specified by the arguments | |
8945 | it is passed, and false otherwise. */ | |
8946 | static bool | |
3101faab | 8947 | sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED, |
b9bf58e5 EB |
8948 | HOST_WIDE_INT delta ATTRIBUTE_UNUSED, |
8949 | HOST_WIDE_INT vcall_offset, | |
3101faab | 8950 | const_tree function ATTRIBUTE_UNUSED) |
b9bf58e5 EB |
8951 | { |
8952 | /* Bound the loop used in the default method above. */ | |
8953 | return (vcall_offset >= -32768 || ! fixed_regs[5]); | |
8954 | } | |
8955 | ||
5751a10b JJ |
8956 | /* How to allocate a 'struct machine_function'. */ |
8957 | ||
8958 | static struct machine_function * | |
8959 | sparc_init_machine_status (void) | |
8960 | { | |
5ead67f6 | 8961 | return GGC_CNEW (struct machine_function); |
5751a10b JJ |
8962 | } |
8963 | ||
8964 | /* Locate some local-dynamic symbol still in use by this function | |
8965 | so that we can print its name in local-dynamic base patterns. */ | |
8966 | ||
8967 | static const char * | |
8968 | get_some_local_dynamic_name (void) | |
8969 | { | |
8970 | rtx insn; | |
8971 | ||
8972 | if (cfun->machine->some_ld_name) | |
8973 | return cfun->machine->some_ld_name; | |
8974 | ||
8975 | for (insn = get_insns (); insn ; insn = NEXT_INSN (insn)) | |
8976 | if (INSN_P (insn) | |
8977 | && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0)) | |
8978 | return cfun->machine->some_ld_name; | |
8979 | ||
f5f7d171 | 8980 | gcc_unreachable (); |
5751a10b JJ |
8981 | } |
8982 | ||
8983 | static int | |
8984 | get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED) | |
8985 | { | |
8986 | rtx x = *px; | |
8987 | ||
8988 | if (x | |
8989 | && GET_CODE (x) == SYMBOL_REF | |
8990 | && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC) | |
8991 | { | |
8992 | cfun->machine->some_ld_name = XSTR (x, 0); | |
8993 | return 1; | |
8994 | } | |
8995 | ||
8996 | return 0; | |
8997 | } | |
8998 | ||
e54c7471 EB |
8999 | /* Handle the TARGET_DWARF_HANDLE_FRAME_UNSPEC hook. |
9000 | This is called from dwarf2out.c to emit call frame instructions | |
9001 | for frame-related insns containing UNSPECs and UNSPEC_VOLATILEs. */ | |
9002 | static void | |
9003 | sparc_dwarf_handle_frame_unspec (const char *label, | |
9004 | rtx pattern ATTRIBUTE_UNUSED, | |
9005 | int index ATTRIBUTE_UNUSED) | |
9006 | { | |
9007 | gcc_assert (index == UNSPECV_SAVEW); | |
9008 | dwarf2out_window_save (label); | |
9009 | } | |
9010 | ||
fdbe66f2 | 9011 | /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL. |
5751a10b JJ |
9012 | We need to emit DTP-relative relocations. */ |
9013 | ||
fdbe66f2 | 9014 | static void |
5751a10b JJ |
9015 | sparc_output_dwarf_dtprel (FILE *file, int size, rtx x) |
9016 | { | |
9017 | switch (size) | |
9018 | { | |
9019 | case 4: | |
9020 | fputs ("\t.word\t%r_tls_dtpoff32(", file); | |
9021 | break; | |
9022 | case 8: | |
9023 | fputs ("\t.xword\t%r_tls_dtpoff64(", file); | |
9024 | break; | |
9025 | default: | |
f5f7d171 | 9026 | gcc_unreachable (); |
5751a10b JJ |
9027 | } |
9028 | output_addr_const (file, x); | |
9029 | fputs (")", file); | |
9030 | } | |
9031 | ||
fdbe66f2 EB |
9032 | /* Do whatever processing is required at the end of a file. */ |
9033 | ||
9034 | static void | |
9035 | sparc_file_end (void) | |
5e77eb53 | 9036 | { |
9fd52c6c DM |
9037 | /* If need to emit the special PIC helper function, do so now. */ |
9038 | if (pic_helper_needed) | |
9039 | { | |
9040 | unsigned int regno = REGNO (pic_offset_table_rtx); | |
9041 | const char *pic_name = reg_names[regno]; | |
9042 | char name[32]; | |
9043 | #ifdef DWARF2_UNWIND_INFO | |
9044 | bool do_cfi; | |
9045 | #endif | |
9046 | ||
9047 | get_pc_thunk_name (name, regno); | |
9048 | if (USE_HIDDEN_LINKONCE) | |
9049 | { | |
9050 | tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL, | |
9051 | get_identifier (name), | |
9052 | build_function_type (void_type_node, | |
9053 | void_list_node)); | |
9054 | DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL, | |
9055 | NULL_TREE, void_type_node); | |
9056 | TREE_STATIC (decl) = 1; | |
9057 | make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl)); | |
9058 | DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN; | |
9059 | DECL_VISIBILITY_SPECIFIED (decl) = 1; | |
9060 | allocate_struct_function (decl, true); | |
9061 | current_function_decl = decl; | |
9062 | init_varasm_status (); | |
9063 | assemble_start_function (decl, name); | |
9064 | } | |
9065 | else | |
9066 | { | |
9067 | const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT); | |
9068 | switch_to_section (text_section); | |
9069 | if (align > 0) | |
9070 | ASM_OUTPUT_ALIGN (asm_out_file, align); | |
9071 | ASM_OUTPUT_LABEL (asm_out_file, name); | |
9072 | } | |
9073 | ||
9074 | #ifdef DWARF2_UNWIND_INFO | |
9075 | do_cfi = dwarf2out_do_cfi_asm (); | |
9076 | if (do_cfi) | |
9077 | fprintf (asm_out_file, "\t.cfi_startproc\n"); | |
9078 | #endif | |
9079 | if (flag_delayed_branch) | |
9080 | fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n", | |
9081 | pic_name, pic_name); | |
9082 | else | |
9083 | fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n", | |
9084 | pic_name, pic_name); | |
9085 | #ifdef DWARF2_UNWIND_INFO | |
9086 | if (do_cfi) | |
9087 | fprintf (asm_out_file, "\t.cfi_endproc\n"); | |
9088 | #endif | |
9089 | } | |
5e77eb53 EB |
9090 | |
9091 | if (NEED_INDICATE_EXEC_STACK) | |
9092 | file_end_indicate_exec_stack (); | |
9093 | } | |
9094 | ||
7269aee7 | 9095 | #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING |
608063c3 | 9096 | /* Implement TARGET_MANGLE_TYPE. */ |
7269aee7 AH |
9097 | |
9098 | static const char * | |
3101faab | 9099 | sparc_mangle_type (const_tree type) |
7269aee7 AH |
9100 | { |
9101 | if (!TARGET_64BIT | |
9102 | && TYPE_MAIN_VARIANT (type) == long_double_type_node | |
9103 | && TARGET_LONG_DOUBLE_128) | |
9104 | return "g"; | |
9105 | ||
9106 | /* For all other types, use normal C++ mangling. */ | |
9107 | return NULL; | |
9108 | } | |
9109 | #endif | |
9110 | ||
396b535a JJ |
9111 | /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit |
9112 | compare and swap on the word containing the byte or half-word. */ | |
9113 | ||
9114 | void | |
9115 | sparc_expand_compare_and_swap_12 (rtx result, rtx mem, rtx oldval, rtx newval) | |
9116 | { | |
9117 | rtx addr1 = force_reg (Pmode, XEXP (mem, 0)); | |
9118 | rtx addr = gen_reg_rtx (Pmode); | |
9119 | rtx off = gen_reg_rtx (SImode); | |
9120 | rtx oldv = gen_reg_rtx (SImode); | |
9121 | rtx newv = gen_reg_rtx (SImode); | |
9122 | rtx oldvalue = gen_reg_rtx (SImode); | |
9123 | rtx newvalue = gen_reg_rtx (SImode); | |
9124 | rtx res = gen_reg_rtx (SImode); | |
9125 | rtx resv = gen_reg_rtx (SImode); | |
9126 | rtx memsi, val, mask, end_label, loop_label, cc; | |
9127 | ||
9128 | emit_insn (gen_rtx_SET (VOIDmode, addr, | |
9129 | gen_rtx_AND (Pmode, addr1, GEN_INT (-4)))); | |
9130 | ||
9131 | if (Pmode != SImode) | |
9132 | addr1 = gen_lowpart (SImode, addr1); | |
9133 | emit_insn (gen_rtx_SET (VOIDmode, off, | |
9134 | gen_rtx_AND (SImode, addr1, GEN_INT (3)))); | |
9135 | ||
9136 | memsi = gen_rtx_MEM (SImode, addr); | |
44d64274 | 9137 | set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER); |
396b535a JJ |
9138 | MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem); |
9139 | ||
9140 | val = force_reg (SImode, memsi); | |
9141 | ||
9142 | emit_insn (gen_rtx_SET (VOIDmode, off, | |
9143 | gen_rtx_XOR (SImode, off, | |
9144 | GEN_INT (GET_MODE (mem) == QImode | |
9145 | ? 3 : 2)))); | |
9146 | ||
9147 | emit_insn (gen_rtx_SET (VOIDmode, off, | |
9148 | gen_rtx_ASHIFT (SImode, off, GEN_INT (3)))); | |
9149 | ||
9150 | if (GET_MODE (mem) == QImode) | |
9151 | mask = force_reg (SImode, GEN_INT (0xff)); | |
9152 | else | |
9153 | mask = force_reg (SImode, GEN_INT (0xffff)); | |
9154 | ||
9155 | emit_insn (gen_rtx_SET (VOIDmode, mask, | |
9156 | gen_rtx_ASHIFT (SImode, mask, off))); | |
9157 | ||
9158 | emit_insn (gen_rtx_SET (VOIDmode, val, | |
9159 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
9160 | val))); | |
9161 | ||
9162 | oldval = gen_lowpart (SImode, oldval); | |
9163 | emit_insn (gen_rtx_SET (VOIDmode, oldv, | |
9164 | gen_rtx_ASHIFT (SImode, oldval, off))); | |
9165 | ||
9166 | newval = gen_lowpart_common (SImode, newval); | |
9167 | emit_insn (gen_rtx_SET (VOIDmode, newv, | |
9168 | gen_rtx_ASHIFT (SImode, newval, off))); | |
9169 | ||
9170 | emit_insn (gen_rtx_SET (VOIDmode, oldv, | |
9171 | gen_rtx_AND (SImode, oldv, mask))); | |
9172 | ||
9173 | emit_insn (gen_rtx_SET (VOIDmode, newv, | |
9174 | gen_rtx_AND (SImode, newv, mask))); | |
9175 | ||
9176 | end_label = gen_label_rtx (); | |
9177 | loop_label = gen_label_rtx (); | |
9178 | emit_label (loop_label); | |
9179 | ||
9180 | emit_insn (gen_rtx_SET (VOIDmode, oldvalue, | |
9181 | gen_rtx_IOR (SImode, oldv, val))); | |
9182 | ||
9183 | emit_insn (gen_rtx_SET (VOIDmode, newvalue, | |
9184 | gen_rtx_IOR (SImode, newv, val))); | |
9185 | ||
9186 | emit_insn (gen_sync_compare_and_swapsi (res, memsi, oldvalue, newvalue)); | |
9187 | ||
9188 | emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label); | |
9189 | ||
9190 | emit_insn (gen_rtx_SET (VOIDmode, resv, | |
9191 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
9192 | res))); | |
9193 | ||
f90b7a5a | 9194 | cc = gen_compare_reg_1 (NE, resv, val); |
396b535a JJ |
9195 | emit_insn (gen_rtx_SET (VOIDmode, val, resv)); |
9196 | ||
f90b7a5a PB |
9197 | /* Use cbranchcc4 to separate the compare and branch! */ |
9198 | emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx), | |
9199 | cc, const0_rtx, loop_label)); | |
396b535a JJ |
9200 | |
9201 | emit_label (end_label); | |
9202 | ||
9203 | emit_insn (gen_rtx_SET (VOIDmode, res, | |
9204 | gen_rtx_AND (SImode, res, mask))); | |
9205 | ||
9206 | emit_insn (gen_rtx_SET (VOIDmode, res, | |
9207 | gen_rtx_LSHIFTRT (SImode, res, off))); | |
9208 | ||
9209 | emit_move_insn (result, gen_lowpart (GET_MODE (result), res)); | |
9210 | } | |
9211 | ||
b52b1749 AS |
9212 | /* Implement TARGET_FRAME_POINTER_REQUIRED. */ |
9213 | ||
9214 | bool | |
9215 | sparc_frame_pointer_required (void) | |
9216 | { | |
9217 | return !(leaf_function_p () && only_leaf_regs_used ()); | |
9218 | } | |
9219 | ||
9220 | /* The way this is structured, we can't eliminate SFP in favor of SP | |
9221 | if the frame pointer is required: we want to use the SFP->HFP elimination | |
9222 | in that case. But the test in update_eliminables doesn't know we are | |
9223 | assuming below that we only do the former elimination. */ | |
9224 | ||
9225 | bool | |
9226 | sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to) | |
9227 | { | |
4900aa61 | 9228 | return (to == HARD_FRAME_POINTER_REGNUM |
b52b1749 AS |
9229 | || !targetm.frame_pointer_required ()); |
9230 | } | |
9231 | ||
e2500fed | 9232 | #include "gt-sparc.h" |