]>
Commit | Line | Data |
---|---|---|
526b7aee | 1 | /* Subroutines used for code generation on the Synopsys DesignWare ARC cpu. |
85ec4feb | 2 | Copyright (C) 1994-2018 Free Software Foundation, Inc. |
526b7aee SV |
3 | |
4 | Sources derived from work done by Sankhya Technologies (www.sankhya.com) on | |
5 | behalf of Synopsys Inc. | |
6 | ||
7 | Position Independent Code support added,Code cleaned up, | |
8 | Comments and Support For ARC700 instructions added by | |
9 | Saurabh Verma (saurabh.verma@codito.com) | |
10 | Ramana Radhakrishnan(ramana.radhakrishnan@codito.com) | |
11 | ||
12 | Fixing ABI inconsistencies, optimizations for ARC600 / ARC700 pipelines, | |
13 | profiling support added by Joern Rennecke <joern.rennecke@embecosm.com> | |
14 | ||
15 | This file is part of GCC. | |
16 | ||
17 | GCC is free software; you can redistribute it and/or modify | |
18 | it under the terms of the GNU General Public License as published by | |
19 | the Free Software Foundation; either version 3, or (at your option) | |
20 | any later version. | |
21 | ||
22 | GCC is distributed in the hope that it will be useful, | |
23 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
24 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
25 | GNU General Public License for more details. | |
26 | ||
27 | You should have received a copy of the GNU General Public License | |
28 | along with GCC; see the file COPYING3. If not see | |
29 | <http://www.gnu.org/licenses/>. */ | |
30 | ||
8fcc61f8 RS |
31 | #define IN_TARGET_CODE 1 |
32 | ||
526b7aee | 33 | #include "config.h" |
526b7aee SV |
34 | #include "system.h" |
35 | #include "coretypes.h" | |
4d0cdd0c | 36 | #include "memmodel.h" |
c7131fb2 | 37 | #include "backend.h" |
e11c4407 | 38 | #include "target.h" |
c7131fb2 | 39 | #include "rtl.h" |
e11c4407 AM |
40 | #include "tree.h" |
41 | #include "cfghooks.h" | |
c7131fb2 | 42 | #include "df.h" |
e11c4407 AM |
43 | #include "tm_p.h" |
44 | #include "stringpool.h" | |
314e6352 | 45 | #include "attribs.h" |
e11c4407 AM |
46 | #include "optabs.h" |
47 | #include "regs.h" | |
48 | #include "emit-rtl.h" | |
49 | #include "recog.h" | |
50 | #include "diagnostic.h" | |
40e23961 | 51 | #include "fold-const.h" |
d8a2d370 DN |
52 | #include "varasm.h" |
53 | #include "stor-layout.h" | |
d8a2d370 | 54 | #include "calls.h" |
526b7aee SV |
55 | #include "output.h" |
56 | #include "insn-attr.h" | |
57 | #include "flags.h" | |
36566b39 | 58 | #include "explow.h" |
526b7aee | 59 | #include "expr.h" |
526b7aee | 60 | #include "langhooks.h" |
526b7aee SV |
61 | #include "tm-constrs.h" |
62 | #include "reload.h" /* For operands_match_p */ | |
60393bbc | 63 | #include "cfgrtl.h" |
526b7aee SV |
64 | #include "tree-pass.h" |
65 | #include "context.h" | |
9b2b7279 | 66 | #include "builtins.h" |
6733978e | 67 | #include "rtl-iter.h" |
b8a64b7f | 68 | #include "alias.h" |
41453183 | 69 | #include "opts.h" |
a2de90a4 | 70 | #include "hw-doloop.h" |
526b7aee | 71 | |
fb155425 | 72 | /* Which cpu we're compiling for (ARC600, ARC601, ARC700). */ |
f9ccf899 CZ |
73 | static char arc_cpu_name[10] = ""; |
74 | static const char *arc_cpu_string = arc_cpu_name; | |
526b7aee | 75 | |
66825a30 CZ |
76 | /* Track which regs are set fixed/call saved/call used from commnad line. */ |
77 | HARD_REG_SET overrideregs; | |
78 | ||
a2de90a4 CZ |
79 | /* Maximum size of a loop. */ |
80 | #define ARC_MAX_LOOP_LENGTH 4095 | |
81 | ||
526b7aee SV |
82 | /* ??? Loads can handle any constant, stores can only handle small ones. */ |
83 | /* OTOH, LIMMs cost extra, so their usefulness is limited. */ | |
84 | #define RTX_OK_FOR_OFFSET_P(MODE, X) \ | |
85 | (GET_CODE (X) == CONST_INT \ | |
86 | && SMALL_INT_RANGE (INTVAL (X), (GET_MODE_SIZE (MODE) - 1) & -4, \ | |
87 | (INTVAL (X) & (GET_MODE_SIZE (MODE) - 1) & 3 \ | |
88 | ? 0 \ | |
89 | : -(-GET_MODE_SIZE (MODE) | -4) >> 1))) | |
90 | ||
9f532472 CZ |
91 | #define LEGITIMATE_SMALL_DATA_OFFSET_P(X) \ |
92 | (GET_CODE (X) == CONST \ | |
93 | && GET_CODE (XEXP ((X), 0)) == PLUS \ | |
94 | && GET_CODE (XEXP (XEXP ((X), 0), 0)) == SYMBOL_REF \ | |
95 | && SYMBOL_REF_SMALL_P (XEXP (XEXP ((X), 0), 0)) \ | |
96 | && GET_CODE (XEXP(XEXP ((X), 0), 1)) == CONST_INT \ | |
97 | && INTVAL (XEXP (XEXP ((X), 0), 1)) <= g_switch_value) | |
98 | ||
99 | #define LEGITIMATE_SMALL_DATA_ADDRESS_P(X) \ | |
100 | (GET_CODE (X) == PLUS \ | |
101 | && REG_P (XEXP ((X), 0)) \ | |
102 | && REGNO (XEXP ((X), 0)) == SDATA_BASE_REGNUM \ | |
103 | && ((GET_CODE (XEXP ((X), 1)) == SYMBOL_REF \ | |
104 | && SYMBOL_REF_SMALL_P (XEXP ((X), 1))) \ | |
105 | || LEGITIMATE_SMALL_DATA_OFFSET_P (XEXP ((X), 1)))) | |
526b7aee SV |
106 | |
107 | /* Array of valid operand punctuation characters. */ | |
108 | char arc_punct_chars[256]; | |
109 | ||
110 | /* State used by arc_ccfsm_advance to implement conditional execution. */ | |
111 | struct GTY (()) arc_ccfsm | |
112 | { | |
113 | int state; | |
114 | int cc; | |
115 | rtx cond; | |
b3458f61 | 116 | rtx_insn *target_insn; |
526b7aee SV |
117 | int target_label; |
118 | }; | |
119 | ||
41453183 CZ |
120 | /* Status of the IRQ_CTRL_AUX register. */ |
121 | typedef struct irq_ctrl_saved_t | |
122 | { | |
123 | /* Last register number used by IRQ_CTRL_SAVED aux_reg. */ | |
124 | short irq_save_last_reg; | |
125 | /* True if BLINK is automatically saved. */ | |
126 | bool irq_save_blink; | |
127 | /* True if LPCOUNT is automatically saved. */ | |
128 | bool irq_save_lpcount; | |
129 | } irq_ctrl_saved_t; | |
130 | static irq_ctrl_saved_t irq_ctrl_saved; | |
131 | ||
132 | #define ARC_AUTOBLINK_IRQ_P(FNTYPE) \ | |
c7314bc1 CZ |
133 | ((ARC_INTERRUPT_P (FNTYPE) \ |
134 | && irq_ctrl_saved.irq_save_blink) \ | |
135 | || (ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
136 | && rgf_banked_register_count > 8)) | |
137 | ||
138 | #define ARC_AUTOFP_IRQ_P(FNTYPE) \ | |
139 | ((ARC_INTERRUPT_P (FNTYPE) \ | |
140 | && (irq_ctrl_saved.irq_save_last_reg > 26)) \ | |
141 | || (ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
142 | && rgf_banked_register_count > 8)) | |
143 | ||
144 | #define ARC_AUTO_IRQ_P(FNTYPE) \ | |
145 | (ARC_INTERRUPT_P (FNTYPE) && !ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
146 | && (irq_ctrl_saved.irq_save_blink \ | |
41453183 CZ |
147 | || (irq_ctrl_saved.irq_save_last_reg >= 0))) |
148 | ||
c7314bc1 CZ |
149 | /* Number of registers in second bank for FIRQ support. */ |
150 | static int rgf_banked_register_count; | |
151 | ||
526b7aee SV |
152 | #define arc_ccfsm_current cfun->machine->ccfsm_current |
153 | ||
154 | #define ARC_CCFSM_BRANCH_DELETED_P(STATE) \ | |
155 | ((STATE)->state == 1 || (STATE)->state == 2) | |
156 | ||
157 | /* Indicate we're conditionalizing insns now. */ | |
158 | #define ARC_CCFSM_RECORD_BRANCH_DELETED(STATE) \ | |
159 | ((STATE)->state += 2) | |
160 | ||
161 | #define ARC_CCFSM_COND_EXEC_P(STATE) \ | |
162 | ((STATE)->state == 3 || (STATE)->state == 4 || (STATE)->state == 5 \ | |
163 | || current_insn_predicate) | |
164 | ||
165 | /* Check if INSN has a 16 bit opcode considering struct arc_ccfsm *STATE. */ | |
166 | #define CCFSM_ISCOMPACT(INSN,STATE) \ | |
167 | (ARC_CCFSM_COND_EXEC_P (STATE) \ | |
168 | ? (get_attr_iscompact (INSN) == ISCOMPACT_TRUE \ | |
169 | || get_attr_iscompact (INSN) == ISCOMPACT_TRUE_LIMM) \ | |
170 | : get_attr_iscompact (INSN) != ISCOMPACT_FALSE) | |
171 | ||
172 | /* Likewise, but also consider that INSN might be in a delay slot of JUMP. */ | |
173 | #define CCFSM_DBR_ISCOMPACT(INSN,JUMP,STATE) \ | |
174 | ((ARC_CCFSM_COND_EXEC_P (STATE) \ | |
175 | || (JUMP_P (JUMP) \ | |
176 | && INSN_ANNULLED_BRANCH_P (JUMP) \ | |
177 | && (TARGET_AT_DBR_CONDEXEC || INSN_FROM_TARGET_P (INSN)))) \ | |
178 | ? (get_attr_iscompact (INSN) == ISCOMPACT_TRUE \ | |
179 | || get_attr_iscompact (INSN) == ISCOMPACT_TRUE_LIMM) \ | |
180 | : get_attr_iscompact (INSN) != ISCOMPACT_FALSE) | |
181 | ||
182 | /* The maximum number of insns skipped which will be conditionalised if | |
183 | possible. */ | |
184 | /* When optimizing for speed: | |
185 | Let p be the probability that the potentially skipped insns need to | |
186 | be executed, pn the cost of a correctly predicted non-taken branch, | |
187 | mt the cost of a mis/non-predicted taken branch, | |
188 | mn mispredicted non-taken, pt correctly predicted taken ; | |
189 | costs expressed in numbers of instructions like the ones considered | |
190 | skipping. | |
191 | Unfortunately we don't have a measure of predictability - this | |
192 | is linked to probability only in that in the no-eviction-scenario | |
193 | there is a lower bound 1 - 2 * min (p, 1-p), and a somewhat larger | |
194 | value that can be assumed *if* the distribution is perfectly random. | |
195 | A predictability of 1 is perfectly plausible not matter what p is, | |
196 | because the decision could be dependent on an invocation parameter | |
197 | of the program. | |
198 | For large p, we want MAX_INSNS_SKIPPED == pn/(1-p) + mt - pn | |
199 | For small p, we want MAX_INSNS_SKIPPED == pt | |
200 | ||
201 | When optimizing for size: | |
202 | We want to skip insn unless we could use 16 opcodes for the | |
203 | non-conditionalized insn to balance the branch length or more. | |
204 | Performance can be tie-breaker. */ | |
205 | /* If the potentially-skipped insns are likely to be executed, we'll | |
206 | generally save one non-taken branch | |
207 | o | |
208 | this to be no less than the 1/p */ | |
209 | #define MAX_INSNS_SKIPPED 3 | |
210 | ||
526b7aee SV |
211 | /* A nop is needed between a 4 byte insn that sets the condition codes and |
212 | a branch that uses them (the same isn't true for an 8 byte insn that sets | |
213 | the condition codes). Set by arc_ccfsm_advance. Used by | |
214 | arc_print_operand. */ | |
215 | ||
216 | static int get_arc_condition_code (rtx); | |
217 | ||
218 | static tree arc_handle_interrupt_attribute (tree *, tree, tree, int, bool *); | |
1825c61e | 219 | static tree arc_handle_fndecl_attribute (tree *, tree, tree, int, bool *); |
526b7aee SV |
220 | |
221 | /* Initialized arc_attribute_table to NULL since arc doesnot have any | |
222 | machine specific supported attributes. */ | |
223 | const struct attribute_spec arc_attribute_table[] = | |
224 | { | |
4849deb1 JJ |
225 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, |
226 | affects_type_identity, handler, exclude } */ | |
227 | { "interrupt", 1, 1, true, false, false, true, | |
228 | arc_handle_interrupt_attribute, NULL }, | |
526b7aee SV |
229 | /* Function calls made to this symbol must be done indirectly, because |
230 | it may lie outside of the 21/25 bit addressing range of a normal function | |
231 | call. */ | |
4849deb1 | 232 | { "long_call", 0, 0, false, true, true, false, NULL, NULL }, |
526b7aee SV |
233 | /* Whereas these functions are always known to reside within the 25 bit |
234 | addressing range of unconditionalized bl. */ | |
4849deb1 | 235 | { "medium_call", 0, 0, false, true, true, false, NULL, NULL }, |
526b7aee SV |
236 | /* And these functions are always known to reside within the 21 bit |
237 | addressing range of blcc. */ | |
4849deb1 | 238 | { "short_call", 0, 0, false, true, true, false, NULL, NULL }, |
1825c61e CZ |
239 | /* Function which are not having the prologue and epilogue generated |
240 | by the compiler. */ | |
4849deb1 | 241 | { "naked", 0, 0, true, false, false, false, arc_handle_fndecl_attribute, |
5d9ae53d | 242 | NULL }, |
4849deb1 | 243 | { NULL, 0, 0, false, false, false, false, NULL, NULL } |
526b7aee SV |
244 | }; |
245 | static int arc_comp_type_attributes (const_tree, const_tree); | |
246 | static void arc_file_start (void); | |
247 | static void arc_internal_label (FILE *, const char *, unsigned long); | |
248 | static void arc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, | |
249 | tree); | |
ef4bddc2 | 250 | static int arc_address_cost (rtx, machine_mode, addr_space_t, bool); |
526b7aee SV |
251 | static void arc_encode_section_info (tree decl, rtx rtl, int first); |
252 | ||
253 | static void arc_init_builtins (void); | |
ef4bddc2 | 254 | static rtx arc_expand_builtin (tree, rtx, rtx, machine_mode, int); |
526b7aee SV |
255 | |
256 | static int branch_dest (rtx); | |
257 | ||
258 | static void arc_output_pic_addr_const (FILE *, rtx, int); | |
526b7aee SV |
259 | static bool arc_function_ok_for_sibcall (tree, tree); |
260 | static rtx arc_function_value (const_tree, const_tree, bool); | |
261 | const char * output_shift (rtx *); | |
262 | static void arc_reorg (void); | |
263 | static bool arc_in_small_data_p (const_tree); | |
264 | ||
265 | static void arc_init_reg_tables (void); | |
266 | static bool arc_return_in_memory (const_tree, const_tree); | |
ef4bddc2 | 267 | static bool arc_vector_mode_supported_p (machine_mode); |
526b7aee | 268 | |
807e902e KZ |
269 | static bool arc_can_use_doloop_p (const widest_int &, const widest_int &, |
270 | unsigned int, bool); | |
ac44248e | 271 | static const char *arc_invalid_within_doloop (const rtx_insn *); |
526b7aee SV |
272 | |
273 | static void output_short_suffix (FILE *file); | |
274 | ||
275 | static bool arc_frame_pointer_required (void); | |
276 | ||
445d7826 | 277 | static bool arc_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT, |
ad23f5d4 JG |
278 | unsigned int, |
279 | enum by_pieces_operation op, | |
280 | bool); | |
281 | ||
c3bde35a AB |
282 | /* Globally visible information about currently selected cpu. */ |
283 | const arc_cpu_t *arc_selected_cpu; | |
f9ccf899 | 284 | |
9f532472 CZ |
285 | static bool |
286 | legitimate_scaled_address_p (machine_mode mode, rtx op, bool strict) | |
287 | { | |
288 | if (GET_CODE (op) != PLUS) | |
289 | return false; | |
290 | ||
291 | if (GET_CODE (XEXP (op, 0)) != MULT) | |
292 | return false; | |
293 | ||
294 | /* Check multiplication operands. */ | |
295 | if (!RTX_OK_FOR_INDEX_P (XEXP (XEXP (op, 0), 0), strict)) | |
296 | return false; | |
297 | ||
298 | if (!CONST_INT_P (XEXP (XEXP (op, 0), 1))) | |
299 | return false; | |
300 | ||
301 | switch (GET_MODE_SIZE (mode)) | |
302 | { | |
303 | case 2: | |
304 | if (INTVAL (XEXP (XEXP (op, 0), 1)) != 2) | |
305 | return false; | |
306 | break; | |
307 | case 8: | |
308 | if (!TARGET_LL64) | |
309 | return false; | |
310 | /* Fall through. */ | |
311 | case 4: | |
312 | if (INTVAL (XEXP (XEXP (op, 0), 1)) != 4) | |
313 | return false; | |
41bc2c0b | 314 | /* Fall through. */ |
9f532472 CZ |
315 | default: |
316 | return false; | |
317 | } | |
318 | ||
319 | /* Check the base. */ | |
320 | if (RTX_OK_FOR_BASE_P (XEXP (op, 1), (strict))) | |
321 | return true; | |
322 | ||
323 | if (flag_pic) | |
324 | { | |
325 | if (CONST_INT_P (XEXP (op, 1))) | |
326 | return true; | |
327 | return false; | |
328 | } | |
329 | if (CONSTANT_P (XEXP (op, 1))) | |
330 | { | |
331 | /* Scalled addresses for sdata is done other places. */ | |
332 | if (GET_CODE (XEXP (op, 1)) == SYMBOL_REF | |
333 | && SYMBOL_REF_SMALL_P (XEXP (op, 1))) | |
334 | return false; | |
335 | return true; | |
336 | } | |
337 | ||
338 | return false; | |
339 | } | |
340 | ||
ac2e1a51 CZ |
341 | /* Check for constructions like REG + OFFS, where OFFS can be a |
342 | register, an immediate or an long immediate. */ | |
343 | ||
344 | static bool | |
b8506a8a | 345 | legitimate_offset_address_p (machine_mode mode, rtx x, bool index, bool strict) |
ac2e1a51 CZ |
346 | { |
347 | if (GET_CODE (x) != PLUS) | |
348 | return false; | |
349 | ||
350 | if (!RTX_OK_FOR_BASE_P (XEXP (x, 0), (strict))) | |
351 | return false; | |
352 | ||
353 | /* Check for: [Rx + small offset] or [Rx + Ry]. */ | |
354 | if (((index && RTX_OK_FOR_INDEX_P (XEXP (x, 1), (strict)) | |
355 | && GET_MODE_SIZE ((mode)) <= 4) | |
356 | || RTX_OK_FOR_OFFSET_P (mode, XEXP (x, 1)))) | |
357 | return true; | |
358 | ||
359 | /* Check for [Rx + symbol]. */ | |
360 | if (!flag_pic | |
361 | && (GET_CODE (XEXP (x, 1)) == SYMBOL_REF) | |
362 | /* Avoid this type of address for double or larger modes. */ | |
363 | && (GET_MODE_SIZE (mode) <= 4) | |
364 | /* Avoid small data which ends in something like GP + | |
365 | symb@sda. */ | |
9f532472 | 366 | && (!SYMBOL_REF_SMALL_P (XEXP (x, 1)))) |
ac2e1a51 CZ |
367 | return true; |
368 | ||
369 | return false; | |
370 | } | |
371 | ||
526b7aee SV |
372 | /* Implements target hook vector_mode_supported_p. */ |
373 | ||
374 | static bool | |
ef4bddc2 | 375 | arc_vector_mode_supported_p (machine_mode mode) |
526b7aee | 376 | { |
00c072ae CZ |
377 | switch (mode) |
378 | { | |
4e10a5a7 | 379 | case E_V2HImode: |
00c072ae | 380 | return TARGET_PLUS_DMPY; |
4e10a5a7 RS |
381 | case E_V4HImode: |
382 | case E_V2SImode: | |
00c072ae | 383 | return TARGET_PLUS_QMACW; |
4e10a5a7 RS |
384 | case E_V4SImode: |
385 | case E_V8HImode: | |
00c072ae | 386 | return TARGET_SIMD_SET; |
526b7aee | 387 | |
00c072ae CZ |
388 | default: |
389 | return false; | |
390 | } | |
391 | } | |
526b7aee | 392 | |
00c072ae CZ |
393 | /* Implements target hook TARGET_VECTORIZE_PREFERRED_SIMD_MODE. */ |
394 | ||
cd1e4d41 | 395 | static machine_mode |
005ba29c | 396 | arc_preferred_simd_mode (scalar_mode mode) |
00c072ae CZ |
397 | { |
398 | switch (mode) | |
399 | { | |
4e10a5a7 | 400 | case E_HImode: |
00c072ae | 401 | return TARGET_PLUS_QMACW ? V4HImode : V2HImode; |
4e10a5a7 | 402 | case E_SImode: |
00c072ae CZ |
403 | return V2SImode; |
404 | ||
405 | default: | |
406 | return word_mode; | |
407 | } | |
526b7aee SV |
408 | } |
409 | ||
00c072ae CZ |
410 | /* Implements target hook |
411 | TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES. */ | |
412 | ||
86e36728 RS |
413 | static void |
414 | arc_autovectorize_vector_sizes (vector_sizes *sizes) | |
00c072ae | 415 | { |
86e36728 RS |
416 | if (TARGET_PLUS_QMACW) |
417 | { | |
418 | sizes->quick_push (8); | |
419 | sizes->quick_push (4); | |
420 | } | |
00c072ae | 421 | } |
526b7aee SV |
422 | |
423 | /* TARGET_PRESERVE_RELOAD_P is still awaiting patch re-evaluation / review. */ | |
424 | static bool arc_preserve_reload_p (rtx in) ATTRIBUTE_UNUSED; | |
425 | static rtx arc_delegitimize_address (rtx); | |
c1ce59ab DM |
426 | static bool arc_can_follow_jump (const rtx_insn *follower, |
427 | const rtx_insn *followee); | |
526b7aee SV |
428 | |
429 | static rtx frame_insn (rtx); | |
ef4bddc2 | 430 | static void arc_function_arg_advance (cumulative_args_t, machine_mode, |
526b7aee | 431 | const_tree, bool); |
ef4bddc2 | 432 | static rtx arc_legitimize_address_0 (rtx, rtx, machine_mode mode); |
526b7aee SV |
433 | |
434 | static void arc_finalize_pic (void); | |
435 | ||
436 | /* initialize the GCC target structure. */ | |
437 | #undef TARGET_COMP_TYPE_ATTRIBUTES | |
438 | #define TARGET_COMP_TYPE_ATTRIBUTES arc_comp_type_attributes | |
439 | #undef TARGET_ASM_FILE_START | |
440 | #define TARGET_ASM_FILE_START arc_file_start | |
441 | #undef TARGET_ATTRIBUTE_TABLE | |
442 | #define TARGET_ATTRIBUTE_TABLE arc_attribute_table | |
443 | #undef TARGET_ASM_INTERNAL_LABEL | |
444 | #define TARGET_ASM_INTERNAL_LABEL arc_internal_label | |
445 | #undef TARGET_RTX_COSTS | |
446 | #define TARGET_RTX_COSTS arc_rtx_costs | |
447 | #undef TARGET_ADDRESS_COST | |
448 | #define TARGET_ADDRESS_COST arc_address_cost | |
449 | ||
450 | #undef TARGET_ENCODE_SECTION_INFO | |
451 | #define TARGET_ENCODE_SECTION_INFO arc_encode_section_info | |
452 | ||
453 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
454 | #define TARGET_CANNOT_FORCE_CONST_MEM arc_cannot_force_const_mem | |
455 | ||
456 | #undef TARGET_INIT_BUILTINS | |
457 | #define TARGET_INIT_BUILTINS arc_init_builtins | |
458 | ||
459 | #undef TARGET_EXPAND_BUILTIN | |
460 | #define TARGET_EXPAND_BUILTIN arc_expand_builtin | |
461 | ||
c69899f0 CZ |
462 | #undef TARGET_BUILTIN_DECL |
463 | #define TARGET_BUILTIN_DECL arc_builtin_decl | |
464 | ||
526b7aee SV |
465 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
466 | #define TARGET_ASM_OUTPUT_MI_THUNK arc_output_mi_thunk | |
467 | ||
468 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
469 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true | |
470 | ||
471 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
472 | #define TARGET_FUNCTION_OK_FOR_SIBCALL arc_function_ok_for_sibcall | |
473 | ||
474 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
475 | #define TARGET_MACHINE_DEPENDENT_REORG arc_reorg | |
476 | ||
477 | #undef TARGET_IN_SMALL_DATA_P | |
478 | #define TARGET_IN_SMALL_DATA_P arc_in_small_data_p | |
479 | ||
480 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
481 | #define TARGET_PROMOTE_FUNCTION_MODE \ | |
482 | default_promote_function_mode_always_promote | |
483 | ||
484 | #undef TARGET_PROMOTE_PROTOTYPES | |
485 | #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true | |
486 | ||
487 | #undef TARGET_RETURN_IN_MEMORY | |
488 | #define TARGET_RETURN_IN_MEMORY arc_return_in_memory | |
489 | #undef TARGET_PASS_BY_REFERENCE | |
490 | #define TARGET_PASS_BY_REFERENCE arc_pass_by_reference | |
491 | ||
492 | #undef TARGET_SETUP_INCOMING_VARARGS | |
493 | #define TARGET_SETUP_INCOMING_VARARGS arc_setup_incoming_varargs | |
494 | ||
495 | #undef TARGET_ARG_PARTIAL_BYTES | |
496 | #define TARGET_ARG_PARTIAL_BYTES arc_arg_partial_bytes | |
497 | ||
498 | #undef TARGET_MUST_PASS_IN_STACK | |
499 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
500 | ||
501 | #undef TARGET_FUNCTION_VALUE | |
502 | #define TARGET_FUNCTION_VALUE arc_function_value | |
503 | ||
504 | #undef TARGET_SCHED_ADJUST_PRIORITY | |
505 | #define TARGET_SCHED_ADJUST_PRIORITY arc_sched_adjust_priority | |
506 | ||
507 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
508 | #define TARGET_VECTOR_MODE_SUPPORTED_P arc_vector_mode_supported_p | |
509 | ||
00c072ae CZ |
510 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
511 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE arc_preferred_simd_mode | |
512 | ||
513 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES | |
514 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES arc_autovectorize_vector_sizes | |
515 | ||
1d0216c8 RS |
516 | #undef TARGET_CAN_USE_DOLOOP_P |
517 | #define TARGET_CAN_USE_DOLOOP_P arc_can_use_doloop_p | |
518 | ||
526b7aee SV |
519 | #undef TARGET_INVALID_WITHIN_DOLOOP |
520 | #define TARGET_INVALID_WITHIN_DOLOOP arc_invalid_within_doloop | |
521 | ||
522 | #undef TARGET_PRESERVE_RELOAD_P | |
523 | #define TARGET_PRESERVE_RELOAD_P arc_preserve_reload_p | |
524 | ||
525 | #undef TARGET_CAN_FOLLOW_JUMP | |
526 | #define TARGET_CAN_FOLLOW_JUMP arc_can_follow_jump | |
527 | ||
528 | #undef TARGET_DELEGITIMIZE_ADDRESS | |
529 | #define TARGET_DELEGITIMIZE_ADDRESS arc_delegitimize_address | |
530 | ||
ad23f5d4 JG |
531 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
532 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
533 | arc_use_by_pieces_infrastructure_p | |
534 | ||
526b7aee SV |
535 | /* Usually, we will be able to scale anchor offsets. |
536 | When this fails, we want LEGITIMIZE_ADDRESS to kick in. */ | |
537 | #undef TARGET_MIN_ANCHOR_OFFSET | |
538 | #define TARGET_MIN_ANCHOR_OFFSET (-1024) | |
539 | #undef TARGET_MAX_ANCHOR_OFFSET | |
540 | #define TARGET_MAX_ANCHOR_OFFSET (1020) | |
541 | ||
542 | #undef TARGET_SECONDARY_RELOAD | |
543 | #define TARGET_SECONDARY_RELOAD arc_secondary_reload | |
544 | ||
545 | #define TARGET_OPTION_OVERRIDE arc_override_options | |
546 | ||
547 | #define TARGET_CONDITIONAL_REGISTER_USAGE arc_conditional_register_usage | |
548 | ||
549 | #define TARGET_TRAMPOLINE_INIT arc_initialize_trampoline | |
550 | ||
526b7aee SV |
551 | #define TARGET_CAN_ELIMINATE arc_can_eliminate |
552 | ||
553 | #define TARGET_FRAME_POINTER_REQUIRED arc_frame_pointer_required | |
554 | ||
555 | #define TARGET_FUNCTION_ARG arc_function_arg | |
556 | ||
557 | #define TARGET_FUNCTION_ARG_ADVANCE arc_function_arg_advance | |
558 | ||
559 | #define TARGET_LEGITIMATE_CONSTANT_P arc_legitimate_constant_p | |
560 | ||
561 | #define TARGET_LEGITIMATE_ADDRESS_P arc_legitimate_address_p | |
562 | ||
563 | #define TARGET_MODE_DEPENDENT_ADDRESS_P arc_mode_dependent_address_p | |
564 | ||
565 | #define TARGET_LEGITIMIZE_ADDRESS arc_legitimize_address | |
566 | ||
567 | #define TARGET_ADJUST_INSN_LENGTH arc_adjust_insn_length | |
568 | ||
569 | #define TARGET_INSN_LENGTH_PARAMETERS arc_insn_length_parameters | |
570 | ||
bf9e9dc5 CZ |
571 | #undef TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P |
572 | #define TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P \ | |
573 | arc_no_speculation_in_delay_slots_p | |
574 | ||
53c8d5a7 | 575 | #undef TARGET_LRA_P |
526b7aee SV |
576 | #define TARGET_LRA_P arc_lra_p |
577 | #define TARGET_REGISTER_PRIORITY arc_register_priority | |
578 | /* Stores with scaled offsets have different displacement ranges. */ | |
579 | #define TARGET_DIFFERENT_ADDR_DISPLACEMENT_P hook_bool_void_true | |
580 | #define TARGET_SPILL_CLASS arc_spill_class | |
581 | ||
1825c61e CZ |
582 | #undef TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS |
583 | #define TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS arc_allocate_stack_slots_for_args | |
584 | ||
585 | #undef TARGET_WARN_FUNC_RETURN | |
586 | #define TARGET_WARN_FUNC_RETURN arc_warn_func_return | |
587 | ||
526b7aee SV |
588 | #include "target-def.h" |
589 | ||
590 | #undef TARGET_ASM_ALIGNED_HI_OP | |
591 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
592 | #undef TARGET_ASM_ALIGNED_SI_OP | |
593 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
594 | ||
28633bbd CZ |
595 | #ifdef HAVE_AS_TLS |
596 | #undef TARGET_HAVE_TLS | |
597 | #define TARGET_HAVE_TLS HAVE_AS_TLS | |
598 | #endif | |
599 | ||
d34a0fdc CZ |
600 | #undef TARGET_DWARF_REGISTER_SPAN |
601 | #define TARGET_DWARF_REGISTER_SPAN arc_dwarf_register_span | |
602 | ||
c43f4279 RS |
603 | #undef TARGET_HARD_REGNO_NREGS |
604 | #define TARGET_HARD_REGNO_NREGS arc_hard_regno_nregs | |
f939c3e6 RS |
605 | #undef TARGET_HARD_REGNO_MODE_OK |
606 | #define TARGET_HARD_REGNO_MODE_OK arc_hard_regno_mode_ok | |
607 | ||
99e1629f RS |
608 | #undef TARGET_MODES_TIEABLE_P |
609 | #define TARGET_MODES_TIEABLE_P arc_modes_tieable_p | |
6fe5e235 CZ |
610 | #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE |
611 | #define TARGET_BUILTIN_SETJMP_FRAME_VALUE arc_builtin_setjmp_frame_value | |
99e1629f | 612 | |
526b7aee SV |
613 | /* Try to keep the (mov:DF _, reg) as early as possible so |
614 | that the d<add/sub/mul>h-lr insns appear together and can | |
615 | use the peephole2 pattern. */ | |
616 | ||
617 | static int | |
ac44248e | 618 | arc_sched_adjust_priority (rtx_insn *insn, int priority) |
526b7aee SV |
619 | { |
620 | rtx set = single_set (insn); | |
621 | if (set | |
622 | && GET_MODE (SET_SRC(set)) == DFmode | |
623 | && GET_CODE (SET_SRC(set)) == REG) | |
624 | { | |
625 | /* Incrementing priority by 20 (empirically derived). */ | |
626 | return priority + 20; | |
627 | } | |
628 | ||
629 | return priority; | |
630 | } | |
631 | ||
f50bb868 CZ |
632 | /* For ARC base register + offset addressing, the validity of the |
633 | address is mode-dependent for most of the offset range, as the | |
634 | offset can be scaled by the access size. | |
635 | We don't expose these as mode-dependent addresses in the | |
636 | mode_dependent_address_p target hook, because that would disable | |
637 | lots of optimizations, and most uses of these addresses are for 32 | |
638 | or 64 bit accesses anyways, which are fine. | |
639 | However, that leaves some addresses for 8 / 16 bit values not | |
640 | properly reloaded by the generic code, which is why we have to | |
641 | schedule secondary reloads for these. */ | |
642 | ||
526b7aee | 643 | static reg_class_t |
f50bb868 CZ |
644 | arc_secondary_reload (bool in_p, |
645 | rtx x, | |
646 | reg_class_t cl, | |
647 | machine_mode mode, | |
648 | secondary_reload_info *sri) | |
526b7aee | 649 | { |
f50bb868 CZ |
650 | enum rtx_code code = GET_CODE (x); |
651 | ||
526b7aee SV |
652 | if (cl == DOUBLE_REGS) |
653 | return GENERAL_REGS; | |
654 | ||
655 | /* The loop counter register can be stored, but not loaded directly. */ | |
656 | if ((cl == LPCOUNT_REG || cl == WRITABLE_CORE_REGS) | |
657 | && in_p && MEM_P (x)) | |
658 | return GENERAL_REGS; | |
f50bb868 CZ |
659 | |
660 | /* If we have a subreg (reg), where reg is a pseudo (that will end in | |
661 | a memory location), then we may need a scratch register to handle | |
662 | the fp/sp+largeoffset address. */ | |
663 | if (code == SUBREG) | |
664 | { | |
665 | rtx addr = NULL_RTX; | |
666 | x = SUBREG_REG (x); | |
667 | ||
668 | if (REG_P (x)) | |
669 | { | |
670 | int regno = REGNO (x); | |
671 | if (regno >= FIRST_PSEUDO_REGISTER) | |
672 | regno = reg_renumber[regno]; | |
673 | ||
674 | if (regno != -1) | |
675 | return NO_REGS; | |
676 | ||
677 | /* It is a pseudo that ends in a stack location. */ | |
678 | if (reg_equiv_mem (REGNO (x))) | |
679 | { | |
680 | /* Get the equivalent address and check the range of the | |
681 | offset. */ | |
682 | rtx mem = reg_equiv_mem (REGNO (x)); | |
683 | addr = find_replacement (&XEXP (mem, 0)); | |
684 | } | |
685 | } | |
686 | else | |
687 | { | |
688 | gcc_assert (MEM_P (x)); | |
689 | addr = XEXP (x, 0); | |
690 | addr = simplify_rtx (addr); | |
691 | } | |
692 | if (addr && GET_CODE (addr) == PLUS | |
693 | && CONST_INT_P (XEXP (addr, 1)) | |
694 | && (!RTX_OK_FOR_OFFSET_P (mode, XEXP (addr, 1)))) | |
695 | { | |
696 | switch (mode) | |
697 | { | |
4e10a5a7 | 698 | case E_QImode: |
f50bb868 CZ |
699 | sri->icode = |
700 | in_p ? CODE_FOR_reload_qi_load : CODE_FOR_reload_qi_store; | |
701 | break; | |
4e10a5a7 | 702 | case E_HImode: |
f50bb868 CZ |
703 | sri->icode = |
704 | in_p ? CODE_FOR_reload_hi_load : CODE_FOR_reload_hi_store; | |
705 | break; | |
706 | default: | |
707 | break; | |
708 | } | |
709 | } | |
710 | } | |
526b7aee SV |
711 | return NO_REGS; |
712 | } | |
713 | ||
f50bb868 CZ |
714 | /* Convert reloads using offsets that are too large to use indirect |
715 | addressing. */ | |
716 | ||
717 | void | |
718 | arc_secondary_reload_conv (rtx reg, rtx mem, rtx scratch, bool store_p) | |
719 | { | |
720 | rtx addr; | |
721 | ||
722 | gcc_assert (GET_CODE (mem) == MEM); | |
723 | addr = XEXP (mem, 0); | |
724 | ||
725 | /* Large offset: use a move. FIXME: ld ops accepts limms as | |
726 | offsets. Hence, the following move insn is not required. */ | |
727 | emit_move_insn (scratch, addr); | |
728 | mem = replace_equiv_address_nv (mem, scratch); | |
729 | ||
730 | /* Now create the move. */ | |
731 | if (store_p) | |
732 | emit_insn (gen_rtx_SET (mem, reg)); | |
733 | else | |
734 | emit_insn (gen_rtx_SET (reg, mem)); | |
735 | ||
736 | return; | |
737 | } | |
738 | ||
526b7aee SV |
739 | static unsigned arc_ifcvt (void); |
740 | ||
741 | namespace { | |
742 | ||
743 | const pass_data pass_data_arc_ifcvt = | |
744 | { | |
745 | RTL_PASS, | |
746 | "arc_ifcvt", /* name */ | |
747 | OPTGROUP_NONE, /* optinfo_flags */ | |
526b7aee SV |
748 | TV_IFCVT2, /* tv_id */ |
749 | 0, /* properties_required */ | |
750 | 0, /* properties_provided */ | |
751 | 0, /* properties_destroyed */ | |
752 | 0, /* todo_flags_start */ | |
753 | TODO_df_finish /* todo_flags_finish */ | |
754 | }; | |
755 | ||
756 | class pass_arc_ifcvt : public rtl_opt_pass | |
757 | { | |
758 | public: | |
759 | pass_arc_ifcvt(gcc::context *ctxt) | |
760 | : rtl_opt_pass(pass_data_arc_ifcvt, ctxt) | |
761 | {} | |
762 | ||
763 | /* opt_pass methods: */ | |
cd4dd8f0 | 764 | opt_pass * clone () { return new pass_arc_ifcvt (m_ctxt); } |
be55bfe6 | 765 | virtual unsigned int execute (function *) { return arc_ifcvt (); } |
526b7aee SV |
766 | }; |
767 | ||
768 | } // anon namespace | |
769 | ||
770 | rtl_opt_pass * | |
771 | make_pass_arc_ifcvt (gcc::context *ctxt) | |
772 | { | |
773 | return new pass_arc_ifcvt (ctxt); | |
774 | } | |
775 | ||
0bc69b81 JR |
776 | static unsigned arc_predicate_delay_insns (void); |
777 | ||
778 | namespace { | |
779 | ||
780 | const pass_data pass_data_arc_predicate_delay_insns = | |
781 | { | |
782 | RTL_PASS, | |
783 | "arc_predicate_delay_insns", /* name */ | |
784 | OPTGROUP_NONE, /* optinfo_flags */ | |
0bc69b81 JR |
785 | TV_IFCVT2, /* tv_id */ |
786 | 0, /* properties_required */ | |
787 | 0, /* properties_provided */ | |
788 | 0, /* properties_destroyed */ | |
789 | 0, /* todo_flags_start */ | |
790 | TODO_df_finish /* todo_flags_finish */ | |
791 | }; | |
792 | ||
793 | class pass_arc_predicate_delay_insns : public rtl_opt_pass | |
794 | { | |
795 | public: | |
796 | pass_arc_predicate_delay_insns(gcc::context *ctxt) | |
797 | : rtl_opt_pass(pass_data_arc_predicate_delay_insns, ctxt) | |
798 | {} | |
799 | ||
800 | /* opt_pass methods: */ | |
be55bfe6 TS |
801 | virtual unsigned int execute (function *) |
802 | { | |
803 | return arc_predicate_delay_insns (); | |
804 | } | |
0bc69b81 JR |
805 | }; |
806 | ||
807 | } // anon namespace | |
808 | ||
809 | rtl_opt_pass * | |
810 | make_pass_arc_predicate_delay_insns (gcc::context *ctxt) | |
811 | { | |
812 | return new pass_arc_predicate_delay_insns (ctxt); | |
813 | } | |
814 | ||
526b7aee SV |
815 | /* Called by OVERRIDE_OPTIONS to initialize various things. */ |
816 | ||
f9ccf899 | 817 | static void |
526b7aee SV |
818 | arc_init (void) |
819 | { | |
0e5172eb CZ |
820 | if (TARGET_V2) |
821 | { | |
822 | /* I have the multiplier, then use it*/ | |
823 | if (TARGET_MPYW || TARGET_MULTI) | |
824 | arc_multcost = COSTS_N_INSNS (1); | |
825 | } | |
526b7aee SV |
826 | /* Note: arc_multcost is only used in rtx_cost if speed is true. */ |
827 | if (arc_multcost < 0) | |
828 | switch (arc_tune) | |
829 | { | |
830 | case TUNE_ARC700_4_2_STD: | |
831 | /* latency 7; | |
832 | max throughput (1 multiply + 4 other insns) / 5 cycles. */ | |
833 | arc_multcost = COSTS_N_INSNS (4); | |
834 | if (TARGET_NOMPY_SET) | |
835 | arc_multcost = COSTS_N_INSNS (30); | |
836 | break; | |
837 | case TUNE_ARC700_4_2_XMAC: | |
838 | /* latency 5; | |
839 | max throughput (1 multiply + 2 other insns) / 3 cycles. */ | |
840 | arc_multcost = COSTS_N_INSNS (3); | |
841 | if (TARGET_NOMPY_SET) | |
842 | arc_multcost = COSTS_N_INSNS (30); | |
843 | break; | |
844 | case TUNE_ARC600: | |
845 | if (TARGET_MUL64_SET) | |
846 | { | |
847 | arc_multcost = COSTS_N_INSNS (4); | |
848 | break; | |
849 | } | |
850 | /* Fall through. */ | |
851 | default: | |
852 | arc_multcost = COSTS_N_INSNS (30); | |
853 | break; | |
854 | } | |
855 | ||
f50bb868 CZ |
856 | /* MPY instructions valid only for ARC700 or ARCv2. */ |
857 | if (TARGET_NOMPY_SET && TARGET_ARC600_FAMILY) | |
858 | error ("-mno-mpy supported only for ARC700 or ARCv2"); | |
526b7aee | 859 | |
526b7aee | 860 | if (!TARGET_DPFP && TARGET_DPFP_DISABLE_LRSR) |
2236746b | 861 | error ("-mno-dpfp-lrsr supported only with -mdpfp"); |
526b7aee SV |
862 | |
863 | /* FPX-1. No fast and compact together. */ | |
864 | if ((TARGET_DPFP_FAST_SET && TARGET_DPFP_COMPACT_SET) | |
865 | || (TARGET_SPFP_FAST_SET && TARGET_SPFP_COMPACT_SET)) | |
866 | error ("FPX fast and compact options cannot be specified together"); | |
867 | ||
868 | /* FPX-2. No fast-spfp for arc600 or arc601. */ | |
f50bb868 | 869 | if (TARGET_SPFP_FAST_SET && TARGET_ARC600_FAMILY) |
526b7aee SV |
870 | error ("-mspfp_fast not available on ARC600 or ARC601"); |
871 | ||
f9ccf899 CZ |
872 | /* FPX-4. No FPX extensions mixed with FPU extensions. */ |
873 | if ((TARGET_DPFP_FAST_SET || TARGET_DPFP_COMPACT_SET || TARGET_SPFP) | |
874 | && TARGET_HARD_FLOAT) | |
8f3304d0 CZ |
875 | error ("No FPX/FPU mixing allowed"); |
876 | ||
526b7aee | 877 | /* Warn for unimplemented PIC in pre-ARC700 cores, and disable flag_pic. */ |
f50bb868 | 878 | if (flag_pic && TARGET_ARC600_FAMILY) |
526b7aee | 879 | { |
f50bb868 CZ |
880 | warning (DK_WARNING, |
881 | "PIC is not supported for %s. Generating non-PIC code only..", | |
882 | arc_cpu_string); | |
526b7aee SV |
883 | flag_pic = 0; |
884 | } | |
885 | ||
886 | arc_init_reg_tables (); | |
887 | ||
888 | /* Initialize array for PRINT_OPERAND_PUNCT_VALID_P. */ | |
889 | memset (arc_punct_chars, 0, sizeof (arc_punct_chars)); | |
890 | arc_punct_chars['#'] = 1; | |
891 | arc_punct_chars['*'] = 1; | |
892 | arc_punct_chars['?'] = 1; | |
893 | arc_punct_chars['!'] = 1; | |
894 | arc_punct_chars['^'] = 1; | |
895 | arc_punct_chars['&'] = 1; | |
f50bb868 CZ |
896 | arc_punct_chars['+'] = 1; |
897 | arc_punct_chars['_'] = 1; | |
526b7aee SV |
898 | |
899 | if (optimize > 1 && !TARGET_NO_COND_EXEC) | |
900 | { | |
901 | /* There are two target-independent ifcvt passes, and arc_reorg may do | |
902 | one or more arc_ifcvt calls. */ | |
903 | opt_pass *pass_arc_ifcvt_4 = make_pass_arc_ifcvt (g); | |
904 | struct register_pass_info arc_ifcvt4_info | |
905 | = { pass_arc_ifcvt_4, "dbr", 1, PASS_POS_INSERT_AFTER }; | |
906 | struct register_pass_info arc_ifcvt5_info | |
907 | = { pass_arc_ifcvt_4->clone (), "shorten", 1, PASS_POS_INSERT_BEFORE }; | |
908 | ||
909 | register_pass (&arc_ifcvt4_info); | |
910 | register_pass (&arc_ifcvt5_info); | |
911 | } | |
0bc69b81 JR |
912 | |
913 | if (flag_delayed_branch) | |
914 | { | |
915 | opt_pass *pass_arc_predicate_delay_insns | |
916 | = make_pass_arc_predicate_delay_insns (g); | |
917 | struct register_pass_info arc_predicate_delay_info | |
918 | = { pass_arc_predicate_delay_insns, "dbr", 1, PASS_POS_INSERT_AFTER }; | |
919 | ||
920 | register_pass (&arc_predicate_delay_info); | |
921 | } | |
526b7aee SV |
922 | } |
923 | ||
41453183 CZ |
924 | /* Parse -mirq-ctrl-saved=RegisterRange, blink, lp_copunt. The |
925 | register range is specified as two registers separated by a dash. | |
926 | It always starts with r0, and its upper limit is fp register. | |
927 | blink and lp_count registers are optional. */ | |
928 | ||
929 | static void | |
930 | irq_range (const char *cstr) | |
931 | { | |
932 | int i, first, last, blink, lpcount, xreg; | |
933 | char *str, *dash, *comma; | |
934 | ||
935 | i = strlen (cstr); | |
936 | str = (char *) alloca (i + 1); | |
937 | memcpy (str, cstr, i + 1); | |
938 | blink = -1; | |
939 | lpcount = -1; | |
940 | ||
941 | dash = strchr (str, '-'); | |
942 | if (!dash) | |
943 | { | |
944 | warning (0, "value of -mirq-ctrl-saved must have form R0-REGx"); | |
945 | return; | |
946 | } | |
947 | *dash = '\0'; | |
948 | ||
949 | comma = strchr (dash + 1, ','); | |
950 | if (comma) | |
951 | *comma = '\0'; | |
952 | ||
953 | first = decode_reg_name (str); | |
954 | if (first != 0) | |
955 | { | |
956 | warning (0, "first register must be R0"); | |
957 | return; | |
958 | } | |
959 | ||
960 | /* At this moment we do not have the register names initialized | |
961 | accordingly. */ | |
962 | if (!strcmp (dash + 1, "ilink")) | |
963 | last = 29; | |
964 | else | |
965 | last = decode_reg_name (dash + 1); | |
966 | ||
967 | if (last < 0) | |
968 | { | |
969 | warning (0, "unknown register name: %s", dash + 1); | |
970 | return; | |
971 | } | |
972 | ||
973 | if (!(last & 0x01)) | |
974 | { | |
975 | warning (0, "last register name %s must be an odd register", dash + 1); | |
976 | return; | |
977 | } | |
978 | ||
979 | *dash = '-'; | |
980 | ||
981 | if (first > last) | |
982 | { | |
983 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
984 | return; | |
985 | } | |
986 | ||
987 | while (comma) | |
988 | { | |
989 | *comma = ','; | |
990 | str = comma + 1; | |
991 | ||
992 | comma = strchr (str, ','); | |
993 | if (comma) | |
994 | *comma = '\0'; | |
995 | ||
996 | xreg = decode_reg_name (str); | |
997 | switch (xreg) | |
998 | { | |
999 | case 31: | |
1000 | blink = 31; | |
1001 | break; | |
1002 | ||
1003 | case 60: | |
1004 | lpcount = 60; | |
1005 | break; | |
1006 | ||
1007 | default: | |
1008 | warning (0, "unknown register name: %s", str); | |
1009 | return; | |
1010 | } | |
1011 | } | |
1012 | ||
1013 | irq_ctrl_saved.irq_save_last_reg = last; | |
1014 | irq_ctrl_saved.irq_save_blink = (blink == 31) || (last == 31); | |
1015 | irq_ctrl_saved.irq_save_lpcount = (lpcount == 60); | |
1016 | } | |
1017 | ||
c7314bc1 CZ |
1018 | /* Parse -mrgf-banked-regs=NUM option string. Valid values for NUM are 4, |
1019 | 8, 16, or 32. */ | |
1020 | ||
1021 | static void | |
1022 | parse_mrgf_banked_regs_option (const char *arg) | |
1023 | { | |
1024 | long int val; | |
1025 | char *end_ptr; | |
1026 | ||
1027 | errno = 0; | |
1028 | val = strtol (arg, &end_ptr, 10); | |
1029 | if (errno != 0 || *arg == '\0' || *end_ptr != '\0' | |
1030 | || (val != 0 && val != 4 && val != 8 && val != 16 && val != 32)) | |
1031 | { | |
1032 | error ("invalid number in -mrgf-banked-regs=%s " | |
1033 | "valid values are 0, 4, 8, 16, or 32", arg); | |
1034 | return; | |
1035 | } | |
1036 | rgf_banked_register_count = (int) val; | |
1037 | } | |
1038 | ||
526b7aee SV |
1039 | /* Check ARC options, generate derived target attributes. */ |
1040 | ||
1041 | static void | |
1042 | arc_override_options (void) | |
1043 | { | |
41453183 CZ |
1044 | unsigned int i; |
1045 | cl_deferred_option *opt; | |
1046 | vec<cl_deferred_option> *vopt | |
1047 | = (vec<cl_deferred_option> *) arc_deferred_options; | |
1048 | ||
526b7aee | 1049 | if (arc_cpu == PROCESSOR_NONE) |
f9ccf899 CZ |
1050 | arc_cpu = TARGET_CPU_DEFAULT; |
1051 | ||
1052 | /* Set the default cpu options. */ | |
1053 | arc_selected_cpu = &arc_cpu_types[(int) arc_cpu]; | |
f9ccf899 CZ |
1054 | |
1055 | /* Set the architectures. */ | |
c3bde35a | 1056 | switch (arc_selected_cpu->arch_info->arch_id) |
f9ccf899 CZ |
1057 | { |
1058 | case BASE_ARCH_em: | |
1059 | arc_cpu_string = "EM"; | |
1060 | break; | |
1061 | case BASE_ARCH_hs: | |
1062 | arc_cpu_string = "HS"; | |
1063 | break; | |
1064 | case BASE_ARCH_700: | |
1065 | if (arc_selected_cpu->processor == PROCESSOR_nps400) | |
1066 | arc_cpu_string = "NPS400"; | |
1067 | else | |
1068 | arc_cpu_string = "ARC700"; | |
1069 | break; | |
1070 | case BASE_ARCH_6xx: | |
1071 | arc_cpu_string = "ARC600"; | |
1072 | break; | |
1073 | default: | |
1074 | gcc_unreachable (); | |
1075 | } | |
1076 | ||
41453183 CZ |
1077 | irq_ctrl_saved.irq_save_last_reg = -1; |
1078 | irq_ctrl_saved.irq_save_blink = false; | |
1079 | irq_ctrl_saved.irq_save_lpcount = false; | |
1080 | ||
c7314bc1 CZ |
1081 | rgf_banked_register_count = 0; |
1082 | ||
41453183 CZ |
1083 | /* Handle the deferred options. */ |
1084 | if (vopt) | |
1085 | FOR_EACH_VEC_ELT (*vopt, i, opt) | |
1086 | { | |
1087 | switch (opt->opt_index) | |
1088 | { | |
1089 | case OPT_mirq_ctrl_saved_: | |
1090 | if (TARGET_V2) | |
1091 | irq_range (opt->arg); | |
1092 | else | |
1093 | warning (0, "option -mirq-ctrl-saved valid only for ARC v2 processors"); | |
1094 | break; | |
1095 | ||
c7314bc1 CZ |
1096 | case OPT_mrgf_banked_regs_: |
1097 | if (TARGET_V2) | |
1098 | parse_mrgf_banked_regs_option (opt->arg); | |
1099 | else | |
1100 | warning (0, "option -mrgf-banked-regs valid only for ARC v2 processors"); | |
1101 | break; | |
1102 | ||
41453183 CZ |
1103 | default: |
1104 | gcc_unreachable(); | |
1105 | } | |
1106 | } | |
1107 | ||
66825a30 CZ |
1108 | CLEAR_HARD_REG_SET (overrideregs); |
1109 | if (common_deferred_options) | |
1110 | { | |
1111 | vec<cl_deferred_option> v = | |
1112 | *((vec<cl_deferred_option> *) common_deferred_options); | |
1113 | int reg, nregs, j; | |
1114 | ||
1115 | FOR_EACH_VEC_ELT (v, i, opt) | |
1116 | { | |
1117 | switch (opt->opt_index) | |
1118 | { | |
1119 | case OPT_ffixed_: | |
1120 | case OPT_fcall_used_: | |
1121 | case OPT_fcall_saved_: | |
1122 | if ((reg = decode_reg_name_and_count (opt->arg, &nregs)) >= 0) | |
1123 | for (j = reg; j < reg + nregs; j++) | |
1124 | SET_HARD_REG_BIT (overrideregs, j); | |
1125 | break; | |
1126 | default: | |
1127 | break; | |
1128 | } | |
1129 | } | |
1130 | } | |
1131 | ||
f9ccf899 CZ |
1132 | /* Set cpu flags accordingly to architecture/selected cpu. The cpu |
1133 | specific flags are set in arc-common.c. The architecture forces | |
1134 | the default hardware configurations in, regardless what command | |
1135 | line options are saying. The CPU optional hw options can be | |
1136 | turned on or off. */ | |
1137 | #define ARC_OPT(NAME, CODE, MASK, DOC) \ | |
1138 | do { \ | |
1139 | if ((arc_selected_cpu->flags & CODE) \ | |
1140 | && ((target_flags_explicit & MASK) == 0)) \ | |
1141 | target_flags |= MASK; \ | |
c3bde35a | 1142 | if (arc_selected_cpu->arch_info->dflags & CODE) \ |
f9ccf899 CZ |
1143 | target_flags |= MASK; \ |
1144 | } while (0); | |
c3bde35a AB |
1145 | #define ARC_OPTX(NAME, CODE, VAR, VAL, DOC) \ |
1146 | do { \ | |
1147 | if ((arc_selected_cpu->flags & CODE) \ | |
1148 | && (VAR == DEFAULT_##VAR)) \ | |
1149 | VAR = VAL; \ | |
1150 | if (arc_selected_cpu->arch_info->dflags & CODE) \ | |
1151 | VAR = VAL; \ | |
f9ccf899 CZ |
1152 | } while (0); |
1153 | ||
1154 | #include "arc-options.def" | |
1155 | ||
1156 | #undef ARC_OPTX | |
1157 | #undef ARC_OPT | |
1158 | ||
1159 | /* Check options against architecture options. Throw an error if | |
1160 | option is not allowed. */ | |
1161 | #define ARC_OPTX(NAME, CODE, VAR, VAL, DOC) \ | |
1162 | do { \ | |
1163 | if ((VAR == VAL) \ | |
c3bde35a | 1164 | && (!(arc_selected_cpu->arch_info->flags & CODE))) \ |
f9ccf899 CZ |
1165 | { \ |
1166 | error ("%s is not available for %s architecture", \ | |
c3bde35a | 1167 | DOC, arc_selected_cpu->arch_info->name); \ |
f9ccf899 CZ |
1168 | } \ |
1169 | } while (0); | |
1170 | #define ARC_OPT(NAME, CODE, MASK, DOC) \ | |
1171 | do { \ | |
1172 | if ((target_flags & MASK) \ | |
c3bde35a | 1173 | && (!(arc_selected_cpu->arch_info->flags & CODE))) \ |
f9ccf899 | 1174 | error ("%s is not available for %s architecture", \ |
c3bde35a | 1175 | DOC, arc_selected_cpu->arch_info->name); \ |
f9ccf899 CZ |
1176 | } while (0); |
1177 | ||
1178 | #include "arc-options.def" | |
1179 | ||
1180 | #undef ARC_OPTX | |
1181 | #undef ARC_OPT | |
1182 | ||
1183 | /* Set Tune option. */ | |
1184 | if (arc_tune == TUNE_NONE) | |
1185 | arc_tune = (enum attr_tune) arc_selected_cpu->tune; | |
526b7aee SV |
1186 | |
1187 | if (arc_size_opt_level == 3) | |
1188 | optimize_size = 1; | |
1189 | ||
30102051 | 1190 | /* Compact casesi is not a valid option for ARCv2 family. */ |
6323c981 | 1191 | if (TARGET_V2) |
30102051 | 1192 | { |
6323c981 CZ |
1193 | if (TARGET_COMPACT_CASESI) |
1194 | { | |
1195 | warning (0, "compact-casesi is not applicable to ARCv2"); | |
1196 | TARGET_COMPACT_CASESI = 0; | |
1197 | } | |
30102051 CZ |
1198 | } |
1199 | else if (optimize_size == 1 | |
1200 | && !global_options_set.x_TARGET_COMPACT_CASESI) | |
1201 | TARGET_COMPACT_CASESI = 1; | |
1202 | ||
526b7aee SV |
1203 | if (flag_pic) |
1204 | target_flags |= MASK_NO_SDATA_SET; | |
1205 | ||
1206 | if (flag_no_common == 255) | |
1207 | flag_no_common = !TARGET_NO_SDATA_SET; | |
1208 | ||
f50bb868 | 1209 | /* TARGET_COMPACT_CASESI needs the "q" register class. */ |
526b7aee SV |
1210 | if (TARGET_MIXED_CODE) |
1211 | TARGET_Q_CLASS = 1; | |
1212 | if (!TARGET_Q_CLASS) | |
1213 | TARGET_COMPACT_CASESI = 0; | |
1214 | if (TARGET_COMPACT_CASESI) | |
1215 | TARGET_CASE_VECTOR_PC_RELATIVE = 1; | |
1216 | ||
9f532472 CZ |
1217 | /* Check for small data option */ |
1218 | if (!global_options_set.x_g_switch_value && !TARGET_NO_SDATA_SET) | |
1219 | g_switch_value = TARGET_LL64 ? 8 : 4; | |
1220 | ||
526b7aee SV |
1221 | /* These need to be done at start up. It's convenient to do them here. */ |
1222 | arc_init (); | |
1223 | } | |
1224 | ||
1225 | /* The condition codes of the ARC, and the inverse function. */ | |
1226 | /* For short branches, the "c" / "nc" names are not defined in the ARC | |
1227 | Programmers manual, so we have to use "lo" / "hs"" instead. */ | |
1228 | static const char *arc_condition_codes[] = | |
1229 | { | |
1230 | "al", 0, "eq", "ne", "p", "n", "lo", "hs", "v", "nv", | |
1231 | "gt", "le", "ge", "lt", "hi", "ls", "pnz", 0 | |
1232 | }; | |
1233 | ||
1234 | enum arc_cc_code_index | |
1235 | { | |
1236 | ARC_CC_AL, ARC_CC_EQ = ARC_CC_AL+2, ARC_CC_NE, ARC_CC_P, ARC_CC_N, | |
1237 | ARC_CC_C, ARC_CC_NC, ARC_CC_V, ARC_CC_NV, | |
1238 | ARC_CC_GT, ARC_CC_LE, ARC_CC_GE, ARC_CC_LT, ARC_CC_HI, ARC_CC_LS, ARC_CC_PNZ, | |
1239 | ARC_CC_LO = ARC_CC_C, ARC_CC_HS = ARC_CC_NC | |
1240 | }; | |
1241 | ||
1242 | #define ARC_INVERSE_CONDITION_CODE(X) ((X) ^ 1) | |
1243 | ||
1244 | /* Returns the index of the ARC condition code string in | |
1245 | `arc_condition_codes'. COMPARISON should be an rtx like | |
1246 | `(eq (...) (...))'. */ | |
1247 | ||
1248 | static int | |
1249 | get_arc_condition_code (rtx comparison) | |
1250 | { | |
1251 | switch (GET_MODE (XEXP (comparison, 0))) | |
1252 | { | |
4e10a5a7 RS |
1253 | case E_CCmode: |
1254 | case E_SImode: /* For BRcc. */ | |
526b7aee SV |
1255 | switch (GET_CODE (comparison)) |
1256 | { | |
1257 | case EQ : return ARC_CC_EQ; | |
1258 | case NE : return ARC_CC_NE; | |
1259 | case GT : return ARC_CC_GT; | |
1260 | case LE : return ARC_CC_LE; | |
1261 | case GE : return ARC_CC_GE; | |
1262 | case LT : return ARC_CC_LT; | |
1263 | case GTU : return ARC_CC_HI; | |
1264 | case LEU : return ARC_CC_LS; | |
1265 | case LTU : return ARC_CC_LO; | |
1266 | case GEU : return ARC_CC_HS; | |
1267 | default : gcc_unreachable (); | |
1268 | } | |
4e10a5a7 | 1269 | case E_CC_ZNmode: |
526b7aee SV |
1270 | switch (GET_CODE (comparison)) |
1271 | { | |
1272 | case EQ : return ARC_CC_EQ; | |
1273 | case NE : return ARC_CC_NE; | |
1274 | case GE: return ARC_CC_P; | |
1275 | case LT: return ARC_CC_N; | |
1276 | case GT : return ARC_CC_PNZ; | |
1277 | default : gcc_unreachable (); | |
1278 | } | |
4e10a5a7 | 1279 | case E_CC_Zmode: |
526b7aee SV |
1280 | switch (GET_CODE (comparison)) |
1281 | { | |
1282 | case EQ : return ARC_CC_EQ; | |
1283 | case NE : return ARC_CC_NE; | |
1284 | default : gcc_unreachable (); | |
1285 | } | |
4e10a5a7 | 1286 | case E_CC_Cmode: |
526b7aee SV |
1287 | switch (GET_CODE (comparison)) |
1288 | { | |
1289 | case LTU : return ARC_CC_C; | |
1290 | case GEU : return ARC_CC_NC; | |
1291 | default : gcc_unreachable (); | |
1292 | } | |
4e10a5a7 | 1293 | case E_CC_FP_GTmode: |
526b7aee SV |
1294 | if (TARGET_ARGONAUT_SET && TARGET_SPFP) |
1295 | switch (GET_CODE (comparison)) | |
1296 | { | |
1297 | case GT : return ARC_CC_N; | |
1298 | case UNLE: return ARC_CC_P; | |
1299 | default : gcc_unreachable (); | |
1300 | } | |
1301 | else | |
1302 | switch (GET_CODE (comparison)) | |
1303 | { | |
1304 | case GT : return ARC_CC_HI; | |
1305 | case UNLE : return ARC_CC_LS; | |
1306 | default : gcc_unreachable (); | |
1307 | } | |
4e10a5a7 | 1308 | case E_CC_FP_GEmode: |
526b7aee SV |
1309 | /* Same for FPX and non-FPX. */ |
1310 | switch (GET_CODE (comparison)) | |
1311 | { | |
1312 | case GE : return ARC_CC_HS; | |
1313 | case UNLT : return ARC_CC_LO; | |
1314 | default : gcc_unreachable (); | |
1315 | } | |
4e10a5a7 | 1316 | case E_CC_FP_UNEQmode: |
526b7aee SV |
1317 | switch (GET_CODE (comparison)) |
1318 | { | |
1319 | case UNEQ : return ARC_CC_EQ; | |
1320 | case LTGT : return ARC_CC_NE; | |
1321 | default : gcc_unreachable (); | |
1322 | } | |
4e10a5a7 | 1323 | case E_CC_FP_ORDmode: |
526b7aee SV |
1324 | switch (GET_CODE (comparison)) |
1325 | { | |
1326 | case UNORDERED : return ARC_CC_C; | |
1327 | case ORDERED : return ARC_CC_NC; | |
1328 | default : gcc_unreachable (); | |
1329 | } | |
4e10a5a7 | 1330 | case E_CC_FPXmode: |
526b7aee SV |
1331 | switch (GET_CODE (comparison)) |
1332 | { | |
1333 | case EQ : return ARC_CC_EQ; | |
1334 | case NE : return ARC_CC_NE; | |
1335 | case UNORDERED : return ARC_CC_C; | |
1336 | case ORDERED : return ARC_CC_NC; | |
1337 | case LTGT : return ARC_CC_HI; | |
1338 | case UNEQ : return ARC_CC_LS; | |
1339 | default : gcc_unreachable (); | |
1340 | } | |
4e10a5a7 | 1341 | case E_CC_FPUmode: |
8f3304d0 CZ |
1342 | switch (GET_CODE (comparison)) |
1343 | { | |
1344 | case EQ : return ARC_CC_EQ; | |
1345 | case NE : return ARC_CC_NE; | |
1346 | case GT : return ARC_CC_GT; | |
1347 | case GE : return ARC_CC_GE; | |
1348 | case LT : return ARC_CC_C; | |
1349 | case LE : return ARC_CC_LS; | |
1350 | case UNORDERED : return ARC_CC_V; | |
1351 | case ORDERED : return ARC_CC_NV; | |
1352 | case UNGT : return ARC_CC_HI; | |
1353 | case UNGE : return ARC_CC_HS; | |
1354 | case UNLT : return ARC_CC_LT; | |
1355 | case UNLE : return ARC_CC_LE; | |
1356 | /* UNEQ and LTGT do not have representation. */ | |
1357 | case LTGT : /* Fall through. */ | |
1358 | case UNEQ : /* Fall through. */ | |
1359 | default : gcc_unreachable (); | |
1360 | } | |
4e10a5a7 | 1361 | case E_CC_FPU_UNEQmode: |
8f3304d0 CZ |
1362 | switch (GET_CODE (comparison)) |
1363 | { | |
1364 | case LTGT : return ARC_CC_NE; | |
1365 | case UNEQ : return ARC_CC_EQ; | |
1366 | default : gcc_unreachable (); | |
1367 | } | |
526b7aee SV |
1368 | default : gcc_unreachable (); |
1369 | } | |
1370 | /*NOTREACHED*/ | |
1371 | return (42); | |
1372 | } | |
1373 | ||
1374 | /* Return true if COMPARISON has a short form that can accomodate OFFSET. */ | |
1375 | ||
1376 | bool | |
1377 | arc_short_comparison_p (rtx comparison, int offset) | |
1378 | { | |
1379 | gcc_assert (ARC_CC_NC == ARC_CC_HS); | |
1380 | gcc_assert (ARC_CC_C == ARC_CC_LO); | |
1381 | switch (get_arc_condition_code (comparison)) | |
1382 | { | |
1383 | case ARC_CC_EQ: case ARC_CC_NE: | |
1384 | return offset >= -512 && offset <= 506; | |
1385 | case ARC_CC_GT: case ARC_CC_LE: case ARC_CC_GE: case ARC_CC_LT: | |
1386 | case ARC_CC_HI: case ARC_CC_LS: case ARC_CC_LO: case ARC_CC_HS: | |
1387 | return offset >= -64 && offset <= 58; | |
1388 | default: | |
1389 | return false; | |
1390 | } | |
1391 | } | |
1392 | ||
1393 | /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE, | |
1394 | return the mode to be used for the comparison. */ | |
1395 | ||
ef4bddc2 | 1396 | machine_mode |
526b7aee SV |
1397 | arc_select_cc_mode (enum rtx_code op, rtx x, rtx y) |
1398 | { | |
ef4bddc2 | 1399 | machine_mode mode = GET_MODE (x); |
526b7aee SV |
1400 | rtx x1; |
1401 | ||
1402 | /* For an operation that sets the condition codes as a side-effect, the | |
1403 | C and V flags is not set as for cmp, so we can only use comparisons where | |
1404 | this doesn't matter. (For LT and GE we can use "mi" and "pl" | |
1405 | instead.) */ | |
1406 | /* ??? We could use "pnz" for greater than zero, however, we could then | |
1407 | get into trouble because the comparison could not be reversed. */ | |
1408 | if (GET_MODE_CLASS (mode) == MODE_INT | |
1409 | && y == const0_rtx | |
1410 | && (op == EQ || op == NE | |
486c559b | 1411 | || ((op == LT || op == GE) && GET_MODE_SIZE (GET_MODE (x)) <= 4))) |
526b7aee SV |
1412 | return CC_ZNmode; |
1413 | ||
1414 | /* add.f for if (a+b) */ | |
1415 | if (mode == SImode | |
1416 | && GET_CODE (y) == NEG | |
1417 | && (op == EQ || op == NE)) | |
1418 | return CC_ZNmode; | |
1419 | ||
1420 | /* Check if this is a test suitable for bxor.f . */ | |
1421 | if (mode == SImode && (op == EQ || op == NE) && CONST_INT_P (y) | |
1422 | && ((INTVAL (y) - 1) & INTVAL (y)) == 0 | |
1423 | && INTVAL (y)) | |
1424 | return CC_Zmode; | |
1425 | ||
1426 | /* Check if this is a test suitable for add / bmsk.f . */ | |
1427 | if (mode == SImode && (op == EQ || op == NE) && CONST_INT_P (y) | |
1428 | && GET_CODE (x) == AND && CONST_INT_P ((x1 = XEXP (x, 1))) | |
1429 | && ((INTVAL (x1) + 1) & INTVAL (x1)) == 0 | |
1430 | && (~INTVAL (x1) | INTVAL (y)) < 0 | |
1431 | && (~INTVAL (x1) | INTVAL (y)) > -0x800) | |
1432 | return CC_Zmode; | |
1433 | ||
1434 | if (GET_MODE (x) == SImode && (op == LTU || op == GEU) | |
1435 | && GET_CODE (x) == PLUS | |
1436 | && (rtx_equal_p (XEXP (x, 0), y) || rtx_equal_p (XEXP (x, 1), y))) | |
1437 | return CC_Cmode; | |
1438 | ||
1439 | if (TARGET_ARGONAUT_SET | |
1440 | && ((mode == SFmode && TARGET_SPFP) || (mode == DFmode && TARGET_DPFP))) | |
1441 | switch (op) | |
1442 | { | |
1443 | case EQ: case NE: case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
1444 | return CC_FPXmode; | |
1445 | case LT: case UNGE: case GT: case UNLE: | |
1446 | return CC_FP_GTmode; | |
1447 | case LE: case UNGT: case GE: case UNLT: | |
1448 | return CC_FP_GEmode; | |
1449 | default: gcc_unreachable (); | |
1450 | } | |
8f3304d0 CZ |
1451 | else if (TARGET_HARD_FLOAT |
1452 | && ((mode == SFmode && TARGET_FP_SP_BASE) | |
1453 | || (mode == DFmode && TARGET_FP_DP_BASE))) | |
526b7aee SV |
1454 | switch (op) |
1455 | { | |
8f3304d0 CZ |
1456 | case EQ: |
1457 | case NE: | |
1458 | case UNORDERED: | |
1459 | case ORDERED: | |
1460 | case UNLT: | |
1461 | case UNLE: | |
1462 | case UNGT: | |
1463 | case UNGE: | |
1464 | case LT: | |
1465 | case LE: | |
1466 | case GT: | |
1467 | case GE: | |
1468 | return CC_FPUmode; | |
1469 | ||
1470 | case LTGT: | |
1471 | case UNEQ: | |
1472 | return CC_FPU_UNEQmode; | |
526b7aee | 1473 | |
8f3304d0 CZ |
1474 | default: |
1475 | gcc_unreachable (); | |
1476 | } | |
1477 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT && TARGET_OPTFPE) | |
1478 | { | |
1479 | switch (op) | |
1480 | { | |
1481 | case EQ: case NE: return CC_Zmode; | |
1482 | case LT: case UNGE: | |
1483 | case GT: case UNLE: return CC_FP_GTmode; | |
1484 | case LE: case UNGT: | |
1485 | case GE: case UNLT: return CC_FP_GEmode; | |
1486 | case UNEQ: case LTGT: return CC_FP_UNEQmode; | |
1487 | case ORDERED: case UNORDERED: return CC_FP_ORDmode; | |
1488 | default: gcc_unreachable (); | |
1489 | } | |
1490 | } | |
526b7aee SV |
1491 | return CCmode; |
1492 | } | |
1493 | ||
1494 | /* Vectors to keep interesting information about registers where it can easily | |
1495 | be got. We use to use the actual mode value as the bit number, but there | |
1496 | is (or may be) more than 32 modes now. Instead we use two tables: one | |
1497 | indexed by hard register number, and one indexed by mode. */ | |
1498 | ||
1499 | /* The purpose of arc_mode_class is to shrink the range of modes so that | |
1500 | they all fit (as bit numbers) in a 32-bit word (again). Each real mode is | |
1501 | mapped into one arc_mode_class mode. */ | |
1502 | ||
1503 | enum arc_mode_class { | |
1504 | C_MODE, | |
1505 | S_MODE, D_MODE, T_MODE, O_MODE, | |
1506 | SF_MODE, DF_MODE, TF_MODE, OF_MODE, | |
1507 | V_MODE | |
1508 | }; | |
1509 | ||
1510 | /* Modes for condition codes. */ | |
1511 | #define C_MODES (1 << (int) C_MODE) | |
1512 | ||
1513 | /* Modes for single-word and smaller quantities. */ | |
1514 | #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE)) | |
1515 | ||
1516 | /* Modes for double-word and smaller quantities. */ | |
1517 | #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE)) | |
1518 | ||
1519 | /* Mode for 8-byte DF values only. */ | |
1520 | #define DF_MODES (1 << DF_MODE) | |
1521 | ||
1522 | /* Modes for quad-word and smaller quantities. */ | |
1523 | #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE)) | |
1524 | ||
1525 | /* Modes for 128-bit vectors. */ | |
1526 | #define V_MODES (1 << (int) V_MODE) | |
1527 | ||
1528 | /* Value is 1 if register/mode pair is acceptable on arc. */ | |
1529 | ||
f939c3e6 | 1530 | static unsigned int arc_hard_regno_modes[] = { |
526b7aee SV |
1531 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, |
1532 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, | |
1533 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, D_MODES, | |
1534 | D_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1535 | ||
1536 | /* ??? Leave these as S_MODES for now. */ | |
1537 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1538 | DF_MODES, 0, DF_MODES, 0, S_MODES, S_MODES, S_MODES, S_MODES, | |
1539 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1540 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, C_MODES, S_MODES, | |
1541 | ||
1542 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1543 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1544 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1545 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1546 | ||
1547 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1548 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1549 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1550 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1551 | ||
1552 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1553 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES | |
1554 | }; | |
1555 | ||
f939c3e6 | 1556 | static unsigned int arc_mode_class [NUM_MACHINE_MODES]; |
526b7aee SV |
1557 | |
1558 | enum reg_class arc_regno_reg_class[FIRST_PSEUDO_REGISTER]; | |
1559 | ||
1560 | enum reg_class | |
1561 | arc_preferred_reload_class (rtx, enum reg_class cl) | |
1562 | { | |
1563 | if ((cl) == CHEAP_CORE_REGS || (cl) == WRITABLE_CORE_REGS) | |
1564 | return GENERAL_REGS; | |
1565 | return cl; | |
1566 | } | |
1567 | ||
1568 | /* Initialize the arc_mode_class array. */ | |
1569 | ||
1570 | static void | |
1571 | arc_init_reg_tables (void) | |
1572 | { | |
1573 | int i; | |
1574 | ||
1575 | for (i = 0; i < NUM_MACHINE_MODES; i++) | |
1576 | { | |
ef4bddc2 | 1577 | machine_mode m = (machine_mode) i; |
f8d91e80 NC |
1578 | |
1579 | switch (GET_MODE_CLASS (m)) | |
526b7aee SV |
1580 | { |
1581 | case MODE_INT: | |
1582 | case MODE_PARTIAL_INT: | |
1583 | case MODE_COMPLEX_INT: | |
f8d91e80 | 1584 | if (GET_MODE_SIZE (m) <= 4) |
526b7aee | 1585 | arc_mode_class[i] = 1 << (int) S_MODE; |
f8d91e80 | 1586 | else if (GET_MODE_SIZE (m) == 8) |
526b7aee | 1587 | arc_mode_class[i] = 1 << (int) D_MODE; |
f8d91e80 | 1588 | else if (GET_MODE_SIZE (m) == 16) |
526b7aee | 1589 | arc_mode_class[i] = 1 << (int) T_MODE; |
f8d91e80 | 1590 | else if (GET_MODE_SIZE (m) == 32) |
526b7aee SV |
1591 | arc_mode_class[i] = 1 << (int) O_MODE; |
1592 | else | |
1593 | arc_mode_class[i] = 0; | |
1594 | break; | |
1595 | case MODE_FLOAT: | |
1596 | case MODE_COMPLEX_FLOAT: | |
f8d91e80 | 1597 | if (GET_MODE_SIZE (m) <= 4) |
526b7aee | 1598 | arc_mode_class[i] = 1 << (int) SF_MODE; |
f8d91e80 | 1599 | else if (GET_MODE_SIZE (m) == 8) |
526b7aee | 1600 | arc_mode_class[i] = 1 << (int) DF_MODE; |
f8d91e80 | 1601 | else if (GET_MODE_SIZE (m) == 16) |
526b7aee | 1602 | arc_mode_class[i] = 1 << (int) TF_MODE; |
f8d91e80 | 1603 | else if (GET_MODE_SIZE (m) == 32) |
526b7aee SV |
1604 | arc_mode_class[i] = 1 << (int) OF_MODE; |
1605 | else | |
1606 | arc_mode_class[i] = 0; | |
1607 | break; | |
1608 | case MODE_VECTOR_INT: | |
00c072ae CZ |
1609 | if (GET_MODE_SIZE (m) == 4) |
1610 | arc_mode_class[i] = (1 << (int) S_MODE); | |
1611 | else if (GET_MODE_SIZE (m) == 8) | |
1612 | arc_mode_class[i] = (1 << (int) D_MODE); | |
1613 | else | |
1614 | arc_mode_class[i] = (1 << (int) V_MODE); | |
526b7aee SV |
1615 | break; |
1616 | case MODE_CC: | |
1617 | default: | |
1618 | /* mode_class hasn't been initialized yet for EXTRA_CC_MODES, so | |
1619 | we must explicitly check for them here. */ | |
1620 | if (i == (int) CCmode || i == (int) CC_ZNmode || i == (int) CC_Zmode | |
1621 | || i == (int) CC_Cmode | |
8f3304d0 CZ |
1622 | || i == CC_FP_GTmode || i == CC_FP_GEmode || i == CC_FP_ORDmode |
1623 | || i == CC_FPUmode || i == CC_FPU_UNEQmode) | |
526b7aee SV |
1624 | arc_mode_class[i] = 1 << (int) C_MODE; |
1625 | else | |
1626 | arc_mode_class[i] = 0; | |
1627 | break; | |
1628 | } | |
1629 | } | |
1630 | } | |
1631 | ||
1632 | /* Core registers 56..59 are used for multiply extension options. | |
1633 | The dsp option uses r56 and r57, these are then named acc1 and acc2. | |
1634 | acc1 is the highpart, and acc2 the lowpart, so which register gets which | |
1635 | number depends on endianness. | |
1636 | The mul64 multiplier options use r57 for mlo, r58 for mmid and r59 for mhi. | |
1637 | Because mlo / mhi form a 64 bit value, we use different gcc internal | |
1638 | register numbers to make them form a register pair as the gcc internals | |
1639 | know it. mmid gets number 57, if still available, and mlo / mhi get | |
1640 | number 58 and 59, depending on endianness. We use DBX_REGISTER_NUMBER | |
1641 | to map this back. */ | |
1642 | char rname56[5] = "r56"; | |
1643 | char rname57[5] = "r57"; | |
1644 | char rname58[5] = "r58"; | |
1645 | char rname59[5] = "r59"; | |
f50bb868 CZ |
1646 | char rname29[7] = "ilink1"; |
1647 | char rname30[7] = "ilink2"; | |
526b7aee SV |
1648 | |
1649 | static void | |
1650 | arc_conditional_register_usage (void) | |
1651 | { | |
1652 | int regno; | |
1653 | int i; | |
1654 | int fix_start = 60, fix_end = 55; | |
1655 | ||
f50bb868 CZ |
1656 | if (TARGET_V2) |
1657 | { | |
1658 | /* For ARCv2 the core register set is changed. */ | |
1659 | strcpy (rname29, "ilink"); | |
1660 | strcpy (rname30, "r30"); | |
66825a30 CZ |
1661 | |
1662 | if (!TEST_HARD_REG_BIT (overrideregs, 30)) | |
1663 | { | |
1664 | /* No user interference. Set the r30 to be used by the | |
1665 | compiler. */ | |
1666 | call_used_regs[30] = 1; | |
1667 | fixed_regs[30] = 0; | |
1668 | ||
1669 | arc_regno_reg_class[30] = WRITABLE_CORE_REGS; | |
1670 | SET_HARD_REG_BIT (reg_class_contents[WRITABLE_CORE_REGS], 30); | |
1671 | SET_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], 30); | |
1672 | SET_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], 30); | |
1673 | SET_HARD_REG_BIT (reg_class_contents[MPY_WRITABLE_CORE_REGS], 30); | |
1674 | } | |
f50bb868 CZ |
1675 | } |
1676 | ||
526b7aee SV |
1677 | if (TARGET_MUL64_SET) |
1678 | { | |
1679 | fix_start = 57; | |
1680 | fix_end = 59; | |
1681 | ||
1682 | /* We don't provide a name for mmed. In rtl / assembly resource lists, | |
1683 | you are supposed to refer to it as mlo & mhi, e.g | |
1684 | (zero_extract:SI (reg:DI 58) (const_int 32) (16)) . | |
1685 | In an actual asm instruction, you are of course use mmed. | |
1686 | The point of avoiding having a separate register for mmed is that | |
1687 | this way, we don't have to carry clobbers of that reg around in every | |
1688 | isntruction that modifies mlo and/or mhi. */ | |
1689 | strcpy (rname57, ""); | |
1690 | strcpy (rname58, TARGET_BIG_ENDIAN ? "mhi" : "mlo"); | |
1691 | strcpy (rname59, TARGET_BIG_ENDIAN ? "mlo" : "mhi"); | |
1692 | } | |
28633bbd CZ |
1693 | |
1694 | /* The nature of arc_tp_regno is actually something more like a global | |
1695 | register, however globalize_reg requires a declaration. | |
1696 | We use EPILOGUE_USES to compensate so that sets from | |
1697 | __builtin_set_frame_pointer are not deleted. */ | |
1698 | if (arc_tp_regno != -1) | |
1699 | fixed_regs[arc_tp_regno] = call_used_regs[arc_tp_regno] = 1; | |
1700 | ||
526b7aee SV |
1701 | if (TARGET_MULMAC_32BY16_SET) |
1702 | { | |
1703 | fix_start = 56; | |
1704 | fix_end = fix_end > 57 ? fix_end : 57; | |
1705 | strcpy (rname56, TARGET_BIG_ENDIAN ? "acc1" : "acc2"); | |
1706 | strcpy (rname57, TARGET_BIG_ENDIAN ? "acc2" : "acc1"); | |
1707 | } | |
1708 | for (regno = fix_start; regno <= fix_end; regno++) | |
1709 | { | |
1710 | if (!fixed_regs[regno]) | |
1711 | warning (0, "multiply option implies r%d is fixed", regno); | |
1712 | fixed_regs [regno] = call_used_regs[regno] = 1; | |
1713 | } | |
1714 | if (TARGET_Q_CLASS) | |
1715 | { | |
0e5172eb CZ |
1716 | if (optimize_size) |
1717 | { | |
1718 | reg_alloc_order[0] = 0; | |
1719 | reg_alloc_order[1] = 1; | |
1720 | reg_alloc_order[2] = 2; | |
1721 | reg_alloc_order[3] = 3; | |
1722 | reg_alloc_order[4] = 12; | |
1723 | reg_alloc_order[5] = 13; | |
1724 | reg_alloc_order[6] = 14; | |
1725 | reg_alloc_order[7] = 15; | |
1726 | reg_alloc_order[8] = 4; | |
1727 | reg_alloc_order[9] = 5; | |
1728 | reg_alloc_order[10] = 6; | |
1729 | reg_alloc_order[11] = 7; | |
1730 | reg_alloc_order[12] = 8; | |
1731 | reg_alloc_order[13] = 9; | |
1732 | reg_alloc_order[14] = 10; | |
1733 | reg_alloc_order[15] = 11; | |
1734 | } | |
1735 | else | |
1736 | { | |
1737 | reg_alloc_order[2] = 12; | |
1738 | reg_alloc_order[3] = 13; | |
1739 | reg_alloc_order[4] = 14; | |
1740 | reg_alloc_order[5] = 15; | |
1741 | reg_alloc_order[6] = 1; | |
1742 | reg_alloc_order[7] = 0; | |
1743 | reg_alloc_order[8] = 4; | |
1744 | reg_alloc_order[9] = 5; | |
1745 | reg_alloc_order[10] = 6; | |
1746 | reg_alloc_order[11] = 7; | |
1747 | reg_alloc_order[12] = 8; | |
1748 | reg_alloc_order[13] = 9; | |
1749 | reg_alloc_order[14] = 10; | |
1750 | reg_alloc_order[15] = 11; | |
1751 | } | |
526b7aee SV |
1752 | } |
1753 | if (TARGET_SIMD_SET) | |
1754 | { | |
1755 | int i; | |
6462fab0 JR |
1756 | for (i = ARC_FIRST_SIMD_VR_REG; i <= ARC_LAST_SIMD_VR_REG; i++) |
1757 | reg_alloc_order [i] = i; | |
1758 | for (i = ARC_FIRST_SIMD_DMA_CONFIG_REG; | |
1759 | i <= ARC_LAST_SIMD_DMA_CONFIG_REG; i++) | |
526b7aee SV |
1760 | reg_alloc_order [i] = i; |
1761 | } | |
a2de90a4 | 1762 | |
526b7aee SV |
1763 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) |
1764 | if (!call_used_regs[regno]) | |
1765 | CLEAR_HARD_REG_BIT (reg_class_contents[SIBCALL_REGS], regno); | |
1766 | for (regno = 32; regno < 60; regno++) | |
1767 | if (!fixed_regs[regno]) | |
1768 | SET_HARD_REG_BIT (reg_class_contents[WRITABLE_CORE_REGS], regno); | |
f50bb868 | 1769 | if (!TARGET_ARC600_FAMILY) |
526b7aee SV |
1770 | { |
1771 | for (regno = 32; regno <= 60; regno++) | |
1772 | CLEAR_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], regno); | |
1773 | ||
1774 | /* If they have used -ffixed-lp_count, make sure it takes | |
1775 | effect. */ | |
1776 | if (fixed_regs[LP_COUNT]) | |
1777 | { | |
1778 | CLEAR_HARD_REG_BIT (reg_class_contents[LPCOUNT_REG], LP_COUNT); | |
1779 | CLEAR_HARD_REG_BIT (reg_class_contents[SIBCALL_REGS], LP_COUNT); | |
1780 | CLEAR_HARD_REG_BIT (reg_class_contents[WRITABLE_CORE_REGS], LP_COUNT); | |
1781 | ||
1782 | /* Instead of taking out SF_MODE like below, forbid it outright. */ | |
f939c3e6 | 1783 | arc_hard_regno_modes[60] = 0; |
526b7aee SV |
1784 | } |
1785 | else | |
f939c3e6 | 1786 | arc_hard_regno_modes[60] = 1 << (int) S_MODE; |
526b7aee SV |
1787 | } |
1788 | ||
8f3304d0 CZ |
1789 | /* ARCHS has 64-bit data-path which makes use of the even-odd paired |
1790 | registers. */ | |
1791 | if (TARGET_HS) | |
1792 | { | |
1793 | for (regno = 1; regno < 32; regno +=2) | |
1794 | { | |
f939c3e6 | 1795 | arc_hard_regno_modes[regno] = S_MODES; |
8f3304d0 CZ |
1796 | } |
1797 | } | |
1798 | ||
526b7aee SV |
1799 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
1800 | { | |
1801 | if (i < 29) | |
1802 | { | |
ceaaa9fe JR |
1803 | if ((TARGET_Q_CLASS || TARGET_RRQ_CLASS) |
1804 | && ((i <= 3) || ((i >= 12) && (i <= 15)))) | |
526b7aee SV |
1805 | arc_regno_reg_class[i] = ARCOMPACT16_REGS; |
1806 | else | |
1807 | arc_regno_reg_class[i] = GENERAL_REGS; | |
1808 | } | |
1809 | else if (i < 60) | |
1810 | arc_regno_reg_class[i] | |
1811 | = (fixed_regs[i] | |
1812 | ? (TEST_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], i) | |
1813 | ? CHEAP_CORE_REGS : ALL_CORE_REGS) | |
f50bb868 | 1814 | : (((!TARGET_ARC600_FAMILY) |
526b7aee SV |
1815 | && TEST_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], i)) |
1816 | ? CHEAP_CORE_REGS : WRITABLE_CORE_REGS)); | |
1817 | else | |
1818 | arc_regno_reg_class[i] = NO_REGS; | |
1819 | } | |
1820 | ||
ceaaa9fe JR |
1821 | /* ARCOMPACT16_REGS is empty, if TARGET_Q_CLASS / TARGET_RRQ_CLASS |
1822 | has not been activated. */ | |
1823 | if (!TARGET_Q_CLASS && !TARGET_RRQ_CLASS) | |
1824 | CLEAR_HARD_REG_SET(reg_class_contents [ARCOMPACT16_REGS]); | |
526b7aee | 1825 | if (!TARGET_Q_CLASS) |
ceaaa9fe | 1826 | CLEAR_HARD_REG_SET(reg_class_contents [AC16_BASE_REGS]); |
526b7aee SV |
1827 | |
1828 | gcc_assert (FIRST_PSEUDO_REGISTER >= 144); | |
1829 | ||
1830 | /* Handle Special Registers. */ | |
1831 | arc_regno_reg_class[29] = LINK_REGS; /* ilink1 register. */ | |
f50bb868 CZ |
1832 | if (!TARGET_V2) |
1833 | arc_regno_reg_class[30] = LINK_REGS; /* ilink2 register. */ | |
526b7aee SV |
1834 | arc_regno_reg_class[31] = LINK_REGS; /* blink register. */ |
1835 | arc_regno_reg_class[60] = LPCOUNT_REG; | |
1836 | arc_regno_reg_class[61] = NO_REGS; /* CC_REG: must be NO_REGS. */ | |
1837 | arc_regno_reg_class[62] = GENERAL_REGS; | |
1838 | ||
1839 | if (TARGET_DPFP) | |
1840 | { | |
1841 | for (i = 40; i < 44; ++i) | |
1842 | { | |
1843 | arc_regno_reg_class[i] = DOUBLE_REGS; | |
1844 | ||
1845 | /* Unless they want us to do 'mov d1, 0x00000000' make sure | |
1846 | no attempt is made to use such a register as a destination | |
1847 | operand in *movdf_insn. */ | |
1848 | if (!TARGET_ARGONAUT_SET) | |
1849 | { | |
1850 | /* Make sure no 'c', 'w', 'W', or 'Rac' constraint is | |
1851 | interpreted to mean they can use D1 or D2 in their insn. */ | |
1852 | CLEAR_HARD_REG_BIT(reg_class_contents[CHEAP_CORE_REGS ], i); | |
1853 | CLEAR_HARD_REG_BIT(reg_class_contents[ALL_CORE_REGS ], i); | |
1854 | CLEAR_HARD_REG_BIT(reg_class_contents[WRITABLE_CORE_REGS ], i); | |
1855 | CLEAR_HARD_REG_BIT(reg_class_contents[MPY_WRITABLE_CORE_REGS], i); | |
1856 | } | |
1857 | } | |
1858 | } | |
1859 | else | |
1860 | { | |
1861 | /* Disable all DOUBLE_REGISTER settings, | |
1862 | if not generating DPFP code. */ | |
1863 | arc_regno_reg_class[40] = ALL_REGS; | |
1864 | arc_regno_reg_class[41] = ALL_REGS; | |
1865 | arc_regno_reg_class[42] = ALL_REGS; | |
1866 | arc_regno_reg_class[43] = ALL_REGS; | |
1867 | ||
ad3d6e77 CZ |
1868 | fixed_regs[40] = 1; |
1869 | fixed_regs[41] = 1; | |
1870 | fixed_regs[42] = 1; | |
1871 | fixed_regs[43] = 1; | |
1872 | ||
f939c3e6 RS |
1873 | arc_hard_regno_modes[40] = 0; |
1874 | arc_hard_regno_modes[42] = 0; | |
526b7aee SV |
1875 | |
1876 | CLEAR_HARD_REG_SET(reg_class_contents [DOUBLE_REGS]); | |
1877 | } | |
1878 | ||
1879 | if (TARGET_SIMD_SET) | |
1880 | { | |
1881 | gcc_assert (ARC_FIRST_SIMD_VR_REG == 64); | |
1882 | gcc_assert (ARC_LAST_SIMD_VR_REG == 127); | |
1883 | ||
1884 | for (i = ARC_FIRST_SIMD_VR_REG; i <= ARC_LAST_SIMD_VR_REG; i++) | |
1885 | arc_regno_reg_class [i] = SIMD_VR_REGS; | |
1886 | ||
1887 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_REG == 128); | |
1888 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_IN_REG == 128); | |
1889 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_OUT_REG == 136); | |
1890 | gcc_assert (ARC_LAST_SIMD_DMA_CONFIG_REG == 143); | |
1891 | ||
1892 | for (i = ARC_FIRST_SIMD_DMA_CONFIG_REG; | |
1893 | i <= ARC_LAST_SIMD_DMA_CONFIG_REG; i++) | |
1894 | arc_regno_reg_class [i] = SIMD_DMA_CONFIG_REGS; | |
1895 | } | |
1896 | ||
1897 | /* pc : r63 */ | |
1898 | arc_regno_reg_class[PROGRAM_COUNTER_REGNO] = GENERAL_REGS; | |
8f3304d0 CZ |
1899 | |
1900 | /*ARCV2 Accumulator. */ | |
79557bae CZ |
1901 | if ((TARGET_V2 |
1902 | && (TARGET_FP_DP_FUSED || TARGET_FP_SP_FUSED)) | |
1903 | || TARGET_PLUS_DMPY) | |
8f3304d0 CZ |
1904 | { |
1905 | arc_regno_reg_class[ACCL_REGNO] = WRITABLE_CORE_REGS; | |
1906 | arc_regno_reg_class[ACCH_REGNO] = WRITABLE_CORE_REGS; | |
1907 | SET_HARD_REG_BIT (reg_class_contents[WRITABLE_CORE_REGS], ACCL_REGNO); | |
1908 | SET_HARD_REG_BIT (reg_class_contents[WRITABLE_CORE_REGS], ACCH_REGNO); | |
1909 | SET_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], ACCL_REGNO); | |
1910 | SET_HARD_REG_BIT (reg_class_contents[CHEAP_CORE_REGS], ACCH_REGNO); | |
8b22ef6a CZ |
1911 | SET_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], ACCL_REGNO); |
1912 | SET_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], ACCH_REGNO); | |
1913 | SET_HARD_REG_BIT (reg_class_contents[MPY_WRITABLE_CORE_REGS], ACCL_REGNO); | |
1914 | SET_HARD_REG_BIT (reg_class_contents[MPY_WRITABLE_CORE_REGS], ACCH_REGNO); | |
1915 | ||
66825a30 CZ |
1916 | /* Allow the compiler to freely use them. */ |
1917 | if (!TEST_HARD_REG_BIT (overrideregs, ACCL_REGNO)) | |
1918 | fixed_regs[ACCL_REGNO] = 0; | |
1919 | if (!TEST_HARD_REG_BIT (overrideregs, ACCH_REGNO)) | |
1920 | fixed_regs[ACCH_REGNO] = 0; | |
8b22ef6a | 1921 | |
66825a30 CZ |
1922 | if (!fixed_regs[ACCH_REGNO] && !fixed_regs[ACCL_REGNO]) |
1923 | arc_hard_regno_modes[ACC_REG_FIRST] = D_MODES; | |
8f3304d0 | 1924 | } |
526b7aee SV |
1925 | } |
1926 | ||
c43f4279 RS |
1927 | /* Implement TARGET_HARD_REGNO_NREGS. */ |
1928 | ||
1929 | static unsigned int | |
1930 | arc_hard_regno_nregs (unsigned int regno, machine_mode mode) | |
1931 | { | |
1932 | if (GET_MODE_SIZE (mode) == 16 | |
1933 | && regno >= ARC_FIRST_SIMD_VR_REG | |
1934 | && regno <= ARC_LAST_SIMD_VR_REG) | |
1935 | return 1; | |
1936 | ||
1937 | return CEIL (GET_MODE_SIZE (mode), UNITS_PER_WORD); | |
1938 | } | |
1939 | ||
f939c3e6 RS |
1940 | /* Implement TARGET_HARD_REGNO_MODE_OK. */ |
1941 | ||
1942 | static bool | |
1943 | arc_hard_regno_mode_ok (unsigned int regno, machine_mode mode) | |
1944 | { | |
1945 | return (arc_hard_regno_modes[regno] & arc_mode_class[mode]) != 0; | |
1946 | } | |
1947 | ||
99e1629f RS |
1948 | /* Implement TARGET_MODES_TIEABLE_P. Tie QI/HI/SI modes together. */ |
1949 | ||
1950 | static bool | |
1951 | arc_modes_tieable_p (machine_mode mode1, machine_mode mode2) | |
1952 | { | |
1953 | return (GET_MODE_CLASS (mode1) == MODE_INT | |
1954 | && GET_MODE_CLASS (mode2) == MODE_INT | |
1955 | && GET_MODE_SIZE (mode1) <= UNITS_PER_WORD | |
1956 | && GET_MODE_SIZE (mode2) <= UNITS_PER_WORD); | |
1957 | } | |
1958 | ||
526b7aee SV |
1959 | /* Handle an "interrupt" attribute; arguments as in |
1960 | struct attribute_spec.handler. */ | |
1961 | ||
1962 | static tree | |
1963 | arc_handle_interrupt_attribute (tree *, tree name, tree args, int, | |
1964 | bool *no_add_attrs) | |
1965 | { | |
1966 | gcc_assert (args); | |
1967 | ||
1968 | tree value = TREE_VALUE (args); | |
1969 | ||
1970 | if (TREE_CODE (value) != STRING_CST) | |
1971 | { | |
1972 | warning (OPT_Wattributes, | |
1973 | "argument of %qE attribute is not a string constant", | |
1974 | name); | |
1975 | *no_add_attrs = true; | |
1976 | } | |
c7314bc1 CZ |
1977 | else if (!TARGET_V2 |
1978 | && strcmp (TREE_STRING_POINTER (value), "ilink1") | |
1979 | && strcmp (TREE_STRING_POINTER (value), "ilink2")) | |
526b7aee SV |
1980 | { |
1981 | warning (OPT_Wattributes, | |
1982 | "argument of %qE attribute is not \"ilink1\" or \"ilink2\"", | |
1983 | name); | |
1984 | *no_add_attrs = true; | |
1985 | } | |
f50bb868 | 1986 | else if (TARGET_V2 |
c7314bc1 CZ |
1987 | && strcmp (TREE_STRING_POINTER (value), "ilink") |
1988 | && strcmp (TREE_STRING_POINTER (value), "firq")) | |
f50bb868 CZ |
1989 | { |
1990 | warning (OPT_Wattributes, | |
c7314bc1 | 1991 | "argument of %qE attribute is not \"ilink\" or \"firq\"", |
f50bb868 CZ |
1992 | name); |
1993 | *no_add_attrs = true; | |
1994 | } | |
1995 | ||
526b7aee SV |
1996 | return NULL_TREE; |
1997 | } | |
1998 | ||
1825c61e CZ |
1999 | static tree |
2000 | arc_handle_fndecl_attribute (tree *node, tree name, tree args ATTRIBUTE_UNUSED, | |
2001 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
2002 | { | |
2003 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
2004 | { | |
2005 | warning (OPT_Wattributes, "%qE attribute only applies to functions", | |
2006 | name); | |
2007 | *no_add_attrs = true; | |
2008 | } | |
2009 | ||
2010 | return NULL_TREE; | |
2011 | } | |
2012 | ||
2013 | /* Implement `TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS' */ | |
2014 | ||
2015 | static bool | |
2016 | arc_allocate_stack_slots_for_args (void) | |
2017 | { | |
2018 | /* Naked functions should not allocate stack slots for arguments. */ | |
2019 | unsigned int fn_type = arc_compute_function_type (cfun); | |
2020 | ||
2021 | return !ARC_NAKED_P(fn_type); | |
2022 | } | |
2023 | ||
2024 | /* Implement `TARGET_WARN_FUNC_RETURN'. */ | |
2025 | ||
2026 | static bool | |
2027 | arc_warn_func_return (tree decl) | |
2028 | { | |
2029 | struct function *func = DECL_STRUCT_FUNCTION (decl); | |
2030 | unsigned int fn_type = arc_compute_function_type (func); | |
2031 | ||
2032 | return !ARC_NAKED_P (fn_type); | |
2033 | } | |
2034 | ||
526b7aee SV |
2035 | /* Return zero if TYPE1 and TYPE are incompatible, one if they are compatible, |
2036 | and two if they are nearly compatible (which causes a warning to be | |
2037 | generated). */ | |
2038 | ||
2039 | static int | |
2040 | arc_comp_type_attributes (const_tree type1, | |
2041 | const_tree type2) | |
2042 | { | |
2043 | int l1, l2, m1, m2, s1, s2; | |
2044 | ||
2045 | /* Check for mismatch of non-default calling convention. */ | |
2046 | if (TREE_CODE (type1) != FUNCTION_TYPE) | |
2047 | return 1; | |
2048 | ||
2049 | /* Check for mismatched call attributes. */ | |
2050 | l1 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2051 | l2 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2052 | m1 = lookup_attribute ("medium_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2053 | m2 = lookup_attribute ("medium_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2054 | s1 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2055 | s2 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2056 | ||
2057 | /* Only bother to check if an attribute is defined. */ | |
2058 | if (l1 | l2 | m1 | m2 | s1 | s2) | |
2059 | { | |
2060 | /* If one type has an attribute, the other must have the same attribute. */ | |
2061 | if ((l1 != l2) || (m1 != m2) || (s1 != s2)) | |
2062 | return 0; | |
2063 | ||
2064 | /* Disallow mixed attributes. */ | |
2065 | if (l1 + m1 + s1 > 1) | |
2066 | return 0; | |
2067 | } | |
2068 | ||
2069 | ||
2070 | return 1; | |
2071 | } | |
2072 | ||
2073 | /* Set the default attributes for TYPE. */ | |
2074 | ||
2075 | void | |
2076 | arc_set_default_type_attributes (tree type ATTRIBUTE_UNUSED) | |
2077 | { | |
2078 | gcc_unreachable(); | |
2079 | } | |
2080 | ||
2081 | /* Misc. utilities. */ | |
2082 | ||
2083 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
2084 | return the rtx for the cc reg in the proper mode. */ | |
2085 | ||
2086 | rtx | |
ef4bddc2 | 2087 | gen_compare_reg (rtx comparison, machine_mode omode) |
526b7aee SV |
2088 | { |
2089 | enum rtx_code code = GET_CODE (comparison); | |
2090 | rtx x = XEXP (comparison, 0); | |
2091 | rtx y = XEXP (comparison, 1); | |
2092 | rtx tmp, cc_reg; | |
ef4bddc2 | 2093 | machine_mode mode, cmode; |
526b7aee SV |
2094 | |
2095 | ||
2096 | cmode = GET_MODE (x); | |
2097 | if (cmode == VOIDmode) | |
2098 | cmode = GET_MODE (y); | |
2099 | gcc_assert (cmode == SImode || cmode == SFmode || cmode == DFmode); | |
2100 | if (cmode == SImode) | |
2101 | { | |
2102 | if (!register_operand (x, SImode)) | |
2103 | { | |
2104 | if (register_operand (y, SImode)) | |
2105 | { | |
2106 | tmp = x; | |
2107 | x = y; | |
2108 | y = tmp; | |
2109 | code = swap_condition (code); | |
2110 | } | |
2111 | else | |
2112 | x = copy_to_mode_reg (SImode, x); | |
2113 | } | |
2114 | if (GET_CODE (y) == SYMBOL_REF && flag_pic) | |
2115 | y = copy_to_mode_reg (SImode, y); | |
2116 | } | |
2117 | else | |
2118 | { | |
2119 | x = force_reg (cmode, x); | |
2120 | y = force_reg (cmode, y); | |
2121 | } | |
2122 | mode = SELECT_CC_MODE (code, x, y); | |
2123 | ||
2124 | cc_reg = gen_rtx_REG (mode, CC_REG); | |
2125 | ||
2126 | /* ??? FIXME (x-y)==0, as done by both cmpsfpx_raw and | |
2127 | cmpdfpx_raw, is not a correct comparison for floats: | |
2128 | http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm | |
2129 | */ | |
2130 | if (TARGET_ARGONAUT_SET | |
2131 | && ((cmode == SFmode && TARGET_SPFP) || (cmode == DFmode && TARGET_DPFP))) | |
2132 | { | |
2133 | switch (code) | |
2134 | { | |
2135 | case NE: case EQ: case LT: case UNGE: case LE: case UNGT: | |
2136 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2137 | break; | |
2138 | case GT: case UNLE: case GE: case UNLT: | |
2139 | code = swap_condition (code); | |
2140 | tmp = x; | |
2141 | x = y; | |
2142 | y = tmp; | |
2143 | break; | |
2144 | default: | |
2145 | gcc_unreachable (); | |
2146 | } | |
2147 | if (cmode == SFmode) | |
2148 | { | |
2149 | emit_insn (gen_cmpsfpx_raw (x, y)); | |
2150 | } | |
2151 | else /* DFmode */ | |
2152 | { | |
2153 | /* Accepts Dx regs directly by insns. */ | |
2154 | emit_insn (gen_cmpdfpx_raw (x, y)); | |
2155 | } | |
2156 | ||
2157 | if (mode != CC_FPXmode) | |
f7df4a84 | 2158 | emit_insn (gen_rtx_SET (cc_reg, |
526b7aee SV |
2159 | gen_rtx_COMPARE (mode, |
2160 | gen_rtx_REG (CC_FPXmode, 61), | |
2161 | const0_rtx))); | |
2162 | } | |
c4014855 CZ |
2163 | else if (TARGET_FPX_QUARK && (cmode == SFmode)) |
2164 | { | |
2165 | switch (code) | |
2166 | { | |
2167 | case NE: case EQ: case GT: case UNLE: case GE: case UNLT: | |
2168 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2169 | break; | |
2170 | case LT: case UNGE: case LE: case UNGT: | |
2171 | code = swap_condition (code); | |
2172 | tmp = x; | |
2173 | x = y; | |
2174 | y = tmp; | |
2175 | break; | |
2176 | default: | |
2177 | gcc_unreachable (); | |
2178 | } | |
2179 | ||
2180 | emit_insn (gen_cmp_quark (cc_reg, | |
2181 | gen_rtx_COMPARE (mode, x, y))); | |
2182 | } | |
8f3304d0 CZ |
2183 | else if (TARGET_HARD_FLOAT |
2184 | && ((cmode == SFmode && TARGET_FP_SP_BASE) | |
2185 | || (cmode == DFmode && TARGET_FP_DP_BASE))) | |
2186 | emit_insn (gen_rtx_SET (cc_reg, gen_rtx_COMPARE (mode, x, y))); | |
526b7aee SV |
2187 | else if (GET_MODE_CLASS (cmode) == MODE_FLOAT && TARGET_OPTFPE) |
2188 | { | |
2189 | rtx op0 = gen_rtx_REG (cmode, 0); | |
2190 | rtx op1 = gen_rtx_REG (cmode, GET_MODE_SIZE (cmode) / UNITS_PER_WORD); | |
b1a82751 | 2191 | bool swap = false; |
526b7aee SV |
2192 | |
2193 | switch (code) | |
2194 | { | |
2195 | case NE: case EQ: case GT: case UNLE: case GE: case UNLT: | |
2196 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2197 | break; | |
2198 | case LT: case UNGE: case LE: case UNGT: | |
2199 | code = swap_condition (code); | |
b1a82751 | 2200 | swap = true; |
526b7aee SV |
2201 | break; |
2202 | default: | |
2203 | gcc_unreachable (); | |
2204 | } | |
2205 | if (currently_expanding_to_rtl) | |
2206 | { | |
b1a82751 CZ |
2207 | if (swap) |
2208 | { | |
2209 | tmp = x; | |
2210 | x = y; | |
2211 | y = tmp; | |
2212 | } | |
526b7aee SV |
2213 | emit_move_insn (op0, x); |
2214 | emit_move_insn (op1, y); | |
2215 | } | |
2216 | else | |
2217 | { | |
2218 | gcc_assert (rtx_equal_p (op0, x)); | |
2219 | gcc_assert (rtx_equal_p (op1, y)); | |
b1a82751 CZ |
2220 | if (swap) |
2221 | { | |
2222 | op0 = y; | |
2223 | op1 = x; | |
2224 | } | |
526b7aee SV |
2225 | } |
2226 | emit_insn (gen_cmp_float (cc_reg, gen_rtx_COMPARE (mode, op0, op1))); | |
2227 | } | |
2228 | else | |
f7df4a84 | 2229 | emit_insn (gen_rtx_SET (cc_reg, gen_rtx_COMPARE (mode, x, y))); |
526b7aee SV |
2230 | return gen_rtx_fmt_ee (code, omode, cc_reg, const0_rtx); |
2231 | } | |
2232 | ||
2233 | /* Return true if VALUE, a const_double, will fit in a limm (4 byte number). | |
2234 | We assume the value can be either signed or unsigned. */ | |
2235 | ||
2236 | bool | |
2237 | arc_double_limm_p (rtx value) | |
2238 | { | |
2239 | HOST_WIDE_INT low, high; | |
2240 | ||
2241 | gcc_assert (GET_CODE (value) == CONST_DOUBLE); | |
2242 | ||
2243 | if (TARGET_DPFP) | |
2244 | return true; | |
2245 | ||
2246 | low = CONST_DOUBLE_LOW (value); | |
2247 | high = CONST_DOUBLE_HIGH (value); | |
2248 | ||
2249 | if (low & 0x80000000) | |
2250 | { | |
2251 | return (((unsigned HOST_WIDE_INT) low <= 0xffffffff && high == 0) | |
2252 | || (((low & - (unsigned HOST_WIDE_INT) 0x80000000) | |
2253 | == - (unsigned HOST_WIDE_INT) 0x80000000) | |
2254 | && high == -1)); | |
2255 | } | |
2256 | else | |
2257 | { | |
2258 | return (unsigned HOST_WIDE_INT) low <= 0x7fffffff && high == 0; | |
2259 | } | |
2260 | } | |
2261 | ||
2262 | /* Do any needed setup for a variadic function. For the ARC, we must | |
2263 | create a register parameter block, and then copy any anonymous arguments | |
2264 | in registers to memory. | |
2265 | ||
2266 | CUM has not been updated for the last named argument which has type TYPE | |
2267 | and mode MODE, and we rely on this fact. */ | |
2268 | ||
2269 | static void | |
2270 | arc_setup_incoming_varargs (cumulative_args_t args_so_far, | |
ef4bddc2 | 2271 | machine_mode mode, tree type, |
526b7aee SV |
2272 | int *pretend_size, int no_rtl) |
2273 | { | |
2274 | int first_anon_arg; | |
2275 | CUMULATIVE_ARGS next_cum; | |
2276 | ||
2277 | /* We must treat `__builtin_va_alist' as an anonymous arg. */ | |
2278 | ||
2279 | next_cum = *get_cumulative_args (args_so_far); | |
8f3304d0 CZ |
2280 | arc_function_arg_advance (pack_cumulative_args (&next_cum), |
2281 | mode, type, true); | |
526b7aee SV |
2282 | first_anon_arg = next_cum; |
2283 | ||
8f3304d0 | 2284 | if (FUNCTION_ARG_REGNO_P (first_anon_arg)) |
526b7aee SV |
2285 | { |
2286 | /* First anonymous (unnamed) argument is in a reg. */ | |
2287 | ||
2288 | /* Note that first_reg_offset < MAX_ARC_PARM_REGS. */ | |
2289 | int first_reg_offset = first_anon_arg; | |
2290 | ||
2291 | if (!no_rtl) | |
2292 | { | |
2293 | rtx regblock | |
2294 | = gen_rtx_MEM (BLKmode, plus_constant (Pmode, arg_pointer_rtx, | |
2295 | FIRST_PARM_OFFSET (0))); | |
2296 | move_block_from_reg (first_reg_offset, regblock, | |
2297 | MAX_ARC_PARM_REGS - first_reg_offset); | |
2298 | } | |
2299 | ||
2300 | *pretend_size | |
2301 | = ((MAX_ARC_PARM_REGS - first_reg_offset ) * UNITS_PER_WORD); | |
2302 | } | |
2303 | } | |
2304 | ||
2305 | /* Cost functions. */ | |
2306 | ||
2307 | /* Provide the costs of an addressing mode that contains ADDR. | |
2308 | If ADDR is not a valid address, its cost is irrelevant. */ | |
2309 | ||
2310 | int | |
ef4bddc2 | 2311 | arc_address_cost (rtx addr, machine_mode, addr_space_t, bool speed) |
526b7aee SV |
2312 | { |
2313 | switch (GET_CODE (addr)) | |
2314 | { | |
2315 | case REG : | |
2316 | return speed || satisfies_constraint_Rcq (addr) ? 0 : 1; | |
2317 | case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: | |
2318 | case PRE_MODIFY: case POST_MODIFY: | |
2319 | return !speed; | |
2320 | ||
2321 | case LABEL_REF : | |
2322 | case SYMBOL_REF : | |
2323 | case CONST : | |
4d03dc2f JR |
2324 | if (TARGET_NPS_CMEM && cmem_address (addr, SImode)) |
2325 | return 0; | |
526b7aee SV |
2326 | /* Most likely needs a LIMM. */ |
2327 | return COSTS_N_INSNS (1); | |
2328 | ||
2329 | case PLUS : | |
2330 | { | |
2331 | register rtx plus0 = XEXP (addr, 0); | |
2332 | register rtx plus1 = XEXP (addr, 1); | |
2333 | ||
2334 | if (GET_CODE (plus0) != REG | |
2335 | && (GET_CODE (plus0) != MULT | |
2336 | || !CONST_INT_P (XEXP (plus0, 1)) | |
2337 | || (INTVAL (XEXP (plus0, 1)) != 2 | |
2338 | && INTVAL (XEXP (plus0, 1)) != 4))) | |
2339 | break; | |
2340 | ||
2341 | switch (GET_CODE (plus1)) | |
2342 | { | |
2343 | case CONST_INT : | |
2344 | return (!RTX_OK_FOR_OFFSET_P (SImode, plus1) | |
2345 | ? COSTS_N_INSNS (1) | |
2346 | : speed | |
2347 | ? 0 | |
2348 | : (satisfies_constraint_Rcq (plus0) | |
2349 | && satisfies_constraint_O (plus1)) | |
2350 | ? 0 | |
2351 | : 1); | |
2352 | case REG: | |
2353 | return (speed < 1 ? 0 | |
2354 | : (satisfies_constraint_Rcq (plus0) | |
2355 | && satisfies_constraint_Rcq (plus1)) | |
2356 | ? 0 : 1); | |
2357 | case CONST : | |
2358 | case SYMBOL_REF : | |
2359 | case LABEL_REF : | |
2360 | return COSTS_N_INSNS (1); | |
2361 | default: | |
2362 | break; | |
2363 | } | |
2364 | break; | |
2365 | } | |
2366 | default: | |
2367 | break; | |
2368 | } | |
2369 | ||
2370 | return 4; | |
2371 | } | |
2372 | ||
2373 | /* Emit instruction X with the frame related bit set. */ | |
2374 | ||
2375 | static rtx | |
2376 | frame_insn (rtx x) | |
2377 | { | |
2378 | x = emit_insn (x); | |
2379 | RTX_FRAME_RELATED_P (x) = 1; | |
2380 | return x; | |
2381 | } | |
2382 | ||
2383 | /* Emit a frame insn to move SRC to DST. */ | |
2384 | ||
2385 | static rtx | |
2386 | frame_move (rtx dst, rtx src) | |
2387 | { | |
67a96300 CZ |
2388 | rtx tmp = gen_rtx_SET (dst, src); |
2389 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2390 | return frame_insn (tmp); | |
526b7aee SV |
2391 | } |
2392 | ||
2393 | /* Like frame_move, but add a REG_INC note for REG if ADDR contains an | |
2394 | auto increment address, or is zero. */ | |
2395 | ||
2396 | static rtx | |
2397 | frame_move_inc (rtx dst, rtx src, rtx reg, rtx addr) | |
2398 | { | |
2399 | rtx insn = frame_move (dst, src); | |
2400 | ||
2401 | if (!addr | |
2402 | || GET_CODE (addr) == PRE_DEC || GET_CODE (addr) == POST_INC | |
2403 | || GET_CODE (addr) == PRE_MODIFY || GET_CODE (addr) == POST_MODIFY) | |
2404 | add_reg_note (insn, REG_INC, reg); | |
2405 | return insn; | |
2406 | } | |
2407 | ||
2408 | /* Emit a frame insn which adjusts a frame address register REG by OFFSET. */ | |
2409 | ||
2410 | static rtx | |
2411 | frame_add (rtx reg, HOST_WIDE_INT offset) | |
2412 | { | |
2413 | gcc_assert ((offset & 0x3) == 0); | |
2414 | if (!offset) | |
2415 | return NULL_RTX; | |
2416 | return frame_move (reg, plus_constant (Pmode, reg, offset)); | |
2417 | } | |
2418 | ||
2419 | /* Emit a frame insn which adjusts stack pointer by OFFSET. */ | |
2420 | ||
2421 | static rtx | |
2422 | frame_stack_add (HOST_WIDE_INT offset) | |
2423 | { | |
2424 | return frame_add (stack_pointer_rtx, offset); | |
2425 | } | |
2426 | ||
2427 | /* Traditionally, we push saved registers first in the prologue, | |
2428 | then we allocate the rest of the frame - and reverse in the epilogue. | |
2429 | This has still its merits for ease of debugging, or saving code size | |
2430 | or even execution time if the stack frame is so large that some accesses | |
2431 | can't be encoded anymore with offsets in the instruction code when using | |
2432 | a different scheme. | |
2433 | Also, it would be a good starting point if we got instructions to help | |
2434 | with register save/restore. | |
2435 | ||
2436 | However, often stack frames are small, and the pushing / popping has | |
2437 | some costs: | |
2438 | - the stack modification prevents a lot of scheduling. | |
2439 | - frame allocation / deallocation needs extra instructions. | |
2440 | - unless we know that we compile ARC700 user code, we need to put | |
2441 | a memory barrier after frame allocation / before deallocation to | |
2442 | prevent interrupts clobbering our data in the frame. | |
2443 | In particular, we don't have any such guarantees for library functions, | |
2444 | which tend to, on the other hand, to have small frames. | |
2445 | ||
2446 | Thus, for small frames, we'd like to use a different scheme: | |
2447 | - The frame is allocated in full with the first prologue instruction, | |
2448 | and deallocated in full with the last epilogue instruction. | |
2449 | Thus, the instructions in-betwen can be freely scheduled. | |
2450 | - If the function has no outgoing arguments on the stack, we can allocate | |
2451 | one register save slot at the top of the stack. This register can then | |
2452 | be saved simultanously with frame allocation, and restored with | |
2453 | frame deallocation. | |
2454 | This register can be picked depending on scheduling considerations, | |
2455 | although same though should go into having some set of registers | |
2456 | to be potentially lingering after a call, and others to be available | |
2457 | immediately - i.e. in the absence of interprocedual optimization, we | |
2458 | can use an ABI-like convention for register allocation to reduce | |
2459 | stalls after function return. */ | |
2460 | /* Function prologue/epilogue handlers. */ | |
2461 | ||
2462 | /* ARCompact stack frames look like: | |
2463 | ||
2464 | Before call After call | |
2465 | high +-----------------------+ +-----------------------+ | |
2466 | mem | reg parm save area | | reg parm save area | | |
2467 | | only created for | | only created for | | |
2468 | | variable arg fns | | variable arg fns | | |
2469 | AP +-----------------------+ +-----------------------+ | |
2470 | | return addr register | | return addr register | | |
2471 | | (if required) | | (if required) | | |
2472 | +-----------------------+ +-----------------------+ | |
2473 | | | | | | |
2474 | | reg save area | | reg save area | | |
2475 | | | | | | |
2476 | +-----------------------+ +-----------------------+ | |
2477 | | frame pointer | | frame pointer | | |
2478 | | (if required) | | (if required) | | |
2479 | FP +-----------------------+ +-----------------------+ | |
2480 | | | | | | |
2481 | | local/temp variables | | local/temp variables | | |
2482 | | | | | | |
2483 | +-----------------------+ +-----------------------+ | |
2484 | | | | | | |
2485 | | arguments on stack | | arguments on stack | | |
2486 | | | | | | |
2487 | SP +-----------------------+ +-----------------------+ | |
2488 | | reg parm save area | | |
2489 | | only created for | | |
2490 | | variable arg fns | | |
2491 | AP +-----------------------+ | |
2492 | | return addr register | | |
2493 | | (if required) | | |
2494 | +-----------------------+ | |
2495 | | | | |
2496 | | reg save area | | |
2497 | | | | |
2498 | +-----------------------+ | |
2499 | | frame pointer | | |
2500 | | (if required) | | |
2501 | FP +-----------------------+ | |
2502 | | | | |
2503 | | local/temp variables | | |
2504 | | | | |
2505 | +-----------------------+ | |
2506 | | | | |
2507 | | arguments on stack | | |
2508 | low | | | |
2509 | mem SP +-----------------------+ | |
2510 | ||
2511 | Notes: | |
2512 | 1) The "reg parm save area" does not exist for non variable argument fns. | |
2513 | The "reg parm save area" can be eliminated completely if we created our | |
2514 | own va-arc.h, but that has tradeoffs as well (so it's not done). */ | |
2515 | ||
2516 | /* Structure to be filled in by arc_compute_frame_size with register | |
2517 | save masks, and offsets for the current function. */ | |
6cdfeeb4 | 2518 | struct GTY (()) arc_frame_info |
526b7aee SV |
2519 | { |
2520 | unsigned int total_size; /* # bytes that the entire frame takes up. */ | |
2521 | unsigned int extra_size; /* # bytes of extra stuff. */ | |
2522 | unsigned int pretend_size; /* # bytes we push and pretend caller did. */ | |
2523 | unsigned int args_size; /* # bytes that outgoing arguments take up. */ | |
2524 | unsigned int reg_size; /* # bytes needed to store regs. */ | |
2525 | unsigned int var_size; /* # bytes that variables take up. */ | |
2526 | unsigned int reg_offset; /* Offset from new sp to store regs. */ | |
2527 | unsigned int gmask; /* Mask of saved gp registers. */ | |
2528 | int initialized; /* Nonzero if frame size already calculated. */ | |
2529 | short millicode_start_reg; | |
2530 | short millicode_end_reg; | |
2531 | bool save_return_addr; | |
2532 | }; | |
2533 | ||
2534 | /* Defining data structures for per-function information. */ | |
2535 | ||
2536 | typedef struct GTY (()) machine_function | |
2537 | { | |
1825c61e | 2538 | unsigned int fn_type; |
526b7aee SV |
2539 | struct arc_frame_info frame_info; |
2540 | /* To keep track of unalignment caused by short insns. */ | |
2541 | int unalign; | |
2542 | int force_short_suffix; /* Used when disgorging return delay slot insns. */ | |
2543 | const char *size_reason; | |
2544 | struct arc_ccfsm ccfsm_current; | |
2545 | /* Map from uid to ccfsm state during branch shortening. */ | |
2546 | rtx ccfsm_current_insn; | |
2547 | char arc_reorg_started; | |
2548 | char prescan_initialized; | |
2549 | } machine_function; | |
2550 | ||
2551 | /* Type of function DECL. | |
2552 | ||
2553 | The result is cached. To reset the cache at the end of a function, | |
2554 | call with DECL = NULL_TREE. */ | |
2555 | ||
1825c61e | 2556 | unsigned int |
526b7aee SV |
2557 | arc_compute_function_type (struct function *fun) |
2558 | { | |
1825c61e CZ |
2559 | tree attr, decl = fun->decl; |
2560 | unsigned int fn_type = fun->machine->fn_type; | |
526b7aee SV |
2561 | |
2562 | if (fn_type != ARC_FUNCTION_UNKNOWN) | |
2563 | return fn_type; | |
2564 | ||
1825c61e CZ |
2565 | /* Check if it is a naked function. */ |
2566 | if (lookup_attribute ("naked", DECL_ATTRIBUTES (decl)) != NULL_TREE) | |
2567 | fn_type |= ARC_FUNCTION_NAKED; | |
2568 | else | |
2569 | fn_type |= ARC_FUNCTION_NORMAL; | |
526b7aee SV |
2570 | |
2571 | /* Now see if this is an interrupt handler. */ | |
1825c61e CZ |
2572 | attr = lookup_attribute ("interrupt", DECL_ATTRIBUTES (decl)); |
2573 | if (attr != NULL_TREE) | |
2574 | { | |
2575 | tree value, args = TREE_VALUE (attr); | |
2576 | ||
2577 | gcc_assert (list_length (args) == 1); | |
2578 | value = TREE_VALUE (args); | |
2579 | gcc_assert (TREE_CODE (value) == STRING_CST); | |
2580 | ||
2581 | if (!strcmp (TREE_STRING_POINTER (value), "ilink1") | |
2582 | || !strcmp (TREE_STRING_POINTER (value), "ilink")) | |
2583 | fn_type |= ARC_FUNCTION_ILINK1; | |
2584 | else if (!strcmp (TREE_STRING_POINTER (value), "ilink2")) | |
2585 | fn_type |= ARC_FUNCTION_ILINK2; | |
2586 | else if (!strcmp (TREE_STRING_POINTER (value), "firq")) | |
2587 | fn_type |= ARC_FUNCTION_FIRQ; | |
2588 | else | |
2589 | gcc_unreachable (); | |
526b7aee SV |
2590 | } |
2591 | ||
2592 | return fun->machine->fn_type = fn_type; | |
2593 | } | |
2594 | ||
2595 | #define FRAME_POINTER_MASK (1 << (FRAME_POINTER_REGNUM)) | |
2596 | #define RETURN_ADDR_MASK (1 << (RETURN_ADDR_REGNUM)) | |
2597 | ||
2598 | /* Tell prologue and epilogue if register REGNO should be saved / restored. | |
2599 | The return address and frame pointer are treated separately. | |
2600 | Don't consider them here. | |
2601 | Addition for pic: The gp register needs to be saved if the current | |
2602 | function changes it to access gotoff variables. | |
2603 | FIXME: This will not be needed if we used some arbitrary register | |
6fe5e235 | 2604 | instead of r26. */ |
526b7aee | 2605 | |
41453183 CZ |
2606 | static bool |
2607 | arc_must_save_register (int regno, struct function *func) | |
2608 | { | |
1825c61e | 2609 | unsigned int fn_type = arc_compute_function_type (func); |
41453183 | 2610 | bool irq_auto_save_p = ((irq_ctrl_saved.irq_save_last_reg >= regno) |
c7314bc1 CZ |
2611 | && ARC_AUTO_IRQ_P (fn_type)); |
2612 | bool firq_auto_save_p = ARC_FAST_INTERRUPT_P (fn_type); | |
2613 | ||
2614 | switch (rgf_banked_register_count) | |
2615 | { | |
2616 | case 4: | |
2617 | firq_auto_save_p &= (regno < 4); | |
2618 | break; | |
2619 | case 8: | |
2620 | firq_auto_save_p &= ((regno < 4) || ((regno > 11) && (regno < 16))); | |
2621 | break; | |
2622 | case 16: | |
2623 | firq_auto_save_p &= ((regno < 4) || ((regno > 9) && (regno < 16)) | |
2624 | || ((regno > 25) && (regno < 29)) | |
2625 | || ((regno > 29) && (regno < 32))); | |
2626 | break; | |
2627 | case 32: | |
2628 | firq_auto_save_p &= (regno != 29) && (regno < 32); | |
2629 | break; | |
2630 | default: | |
2631 | firq_auto_save_p = false; | |
2632 | break; | |
2633 | } | |
41453183 CZ |
2634 | |
2635 | if ((regno) != RETURN_ADDR_REGNUM | |
2636 | && (regno) != FRAME_POINTER_REGNUM | |
2637 | && df_regs_ever_live_p (regno) | |
2638 | && (!call_used_regs[regno] | |
2639 | || ARC_INTERRUPT_P (fn_type)) | |
2640 | /* Do not emit code for auto saved regs. */ | |
c7314bc1 CZ |
2641 | && !irq_auto_save_p |
2642 | && !firq_auto_save_p) | |
41453183 CZ |
2643 | return true; |
2644 | ||
2645 | if (flag_pic && crtl->uses_pic_offset_table | |
2646 | && regno == PIC_OFFSET_TABLE_REGNUM) | |
2647 | return true; | |
2648 | ||
2649 | return false; | |
2650 | } | |
2651 | ||
2652 | /* Return true if the return address must be saved in the current function, | |
2653 | otherwise return false. */ | |
2654 | ||
2655 | static bool | |
2656 | arc_must_save_return_addr (struct function *func) | |
2657 | { | |
2658 | if (func->machine->frame_info.save_return_addr) | |
2659 | return true; | |
2660 | ||
2661 | return false; | |
2662 | } | |
2663 | ||
2664 | /* Helper function to wrap FRAME_POINTER_NEEDED. We do this as | |
2665 | FRAME_POINTER_NEEDED will not be true until the IRA (Integrated | |
2666 | Register Allocator) pass, while we want to get the frame size | |
6fe5e235 CZ |
2667 | correct earlier than the IRA pass. |
2668 | ||
2669 | When a function uses eh_return we must ensure that the fp register | |
2670 | is saved and then restored so that the unwinder can restore the | |
2671 | correct value for the frame we are going to jump to. | |
2672 | ||
2673 | To do this we force all frames that call eh_return to require a | |
2674 | frame pointer (see arc_frame_pointer_required), this | |
2675 | will ensure that the previous frame pointer is stored on entry to | |
2676 | the function, and will then be reloaded at function exit. | |
2677 | ||
2678 | As the frame pointer is handled as a special case in our prologue | |
2679 | and epilogue code it must not be saved and restored using the | |
2680 | MUST_SAVE_REGISTER mechanism otherwise we run into issues where GCC | |
2681 | believes that the function is not using a frame pointer and that | |
2682 | the value in the fp register is the frame pointer, while the | |
2683 | prologue and epilogue are busy saving and restoring the fp | |
2684 | register. | |
2685 | ||
2686 | During compilation of a function the frame size is evaluated | |
2687 | multiple times, it is not until the reload pass is complete the the | |
2688 | frame size is considered fixed (it is at this point that space for | |
2689 | all spills has been allocated). However the frame_pointer_needed | |
2690 | variable is not set true until the register allocation pass, as a | |
2691 | result in the early stages the frame size does not include space | |
2692 | for the frame pointer to be spilled. | |
2693 | ||
2694 | The problem that this causes is that the rtl generated for | |
2695 | EH_RETURN_HANDLER_RTX uses the details of the frame size to compute | |
2696 | the offset from the frame pointer at which the return address | |
2697 | lives. However, in early passes GCC has not yet realised we need a | |
2698 | frame pointer, and so has not included space for the frame pointer | |
2699 | in the frame size, and so gets the offset of the return address | |
2700 | wrong. This should not be an issue as in later passes GCC has | |
2701 | realised that the frame pointer needs to be spilled, and has | |
2702 | increased the frame size. However, the rtl for the | |
2703 | EH_RETURN_HANDLER_RTX is not regenerated to use the newer, larger | |
2704 | offset, and the wrong smaller offset is used. */ | |
2705 | ||
41453183 CZ |
2706 | static bool |
2707 | arc_frame_pointer_needed (void) | |
2708 | { | |
6fe5e235 | 2709 | return (frame_pointer_needed || crtl->calls_eh_return); |
41453183 CZ |
2710 | } |
2711 | ||
526b7aee SV |
2712 | /* Return non-zero if there are registers to be saved or loaded using |
2713 | millicode thunks. We can only use consecutive sequences starting | |
2714 | with r13, and not going beyond r25. | |
2715 | GMASK is a bitmask of registers to save. This function sets | |
2716 | FRAME->millicod_start_reg .. FRAME->millicode_end_reg to the range | |
2717 | of registers to be saved / restored with a millicode call. */ | |
2718 | ||
2719 | static int | |
2720 | arc_compute_millicode_save_restore_regs (unsigned int gmask, | |
2721 | struct arc_frame_info *frame) | |
2722 | { | |
2723 | int regno; | |
2724 | ||
2725 | int start_reg = 13, end_reg = 25; | |
2726 | ||
2727 | for (regno = start_reg; regno <= end_reg && (gmask & (1L << regno));) | |
2728 | regno++; | |
2729 | end_reg = regno - 1; | |
2730 | /* There is no point in using millicode thunks if we don't save/restore | |
2731 | at least three registers. For non-leaf functions we also have the | |
2732 | blink restore. */ | |
2733 | if (regno - start_reg >= 3 - (crtl->is_leaf == 0)) | |
2734 | { | |
2735 | frame->millicode_start_reg = 13; | |
2736 | frame->millicode_end_reg = regno - 1; | |
2737 | return 1; | |
2738 | } | |
2739 | return 0; | |
2740 | } | |
2741 | ||
6fe5e235 CZ |
2742 | /* Return the bytes needed to compute the frame pointer from the |
2743 | current stack pointer. */ | |
526b7aee | 2744 | |
6fe5e235 CZ |
2745 | static unsigned int |
2746 | arc_compute_frame_size (void) | |
526b7aee SV |
2747 | { |
2748 | int regno; | |
2749 | unsigned int total_size, var_size, args_size, pretend_size, extra_size; | |
2750 | unsigned int reg_size, reg_offset; | |
2751 | unsigned int gmask; | |
6fe5e235 CZ |
2752 | struct arc_frame_info *frame_info; |
2753 | int size; | |
2754 | ||
2755 | /* The answer might already be known. */ | |
2756 | if (cfun->machine->frame_info.initialized) | |
2757 | return cfun->machine->frame_info.total_size; | |
526b7aee | 2758 | |
6fe5e235 CZ |
2759 | frame_info = &cfun->machine->frame_info; |
2760 | size = ARC_STACK_ALIGN (get_frame_size ()); | |
526b7aee | 2761 | |
6fe5e235 | 2762 | /* 1) Size of locals and temporaries. */ |
526b7aee SV |
2763 | var_size = size; |
2764 | ||
6fe5e235 | 2765 | /* 2) Size of outgoing arguments. */ |
526b7aee SV |
2766 | args_size = crtl->outgoing_args_size; |
2767 | ||
2768 | /* 3) Calculate space needed for saved registers. | |
2769 | ??? We ignore the extension registers for now. */ | |
2770 | ||
2771 | /* See if this is an interrupt handler. Call used registers must be saved | |
2772 | for them too. */ | |
2773 | ||
2774 | reg_size = 0; | |
2775 | gmask = 0; | |
526b7aee SV |
2776 | |
2777 | for (regno = 0; regno <= 31; regno++) | |
2778 | { | |
41453183 | 2779 | if (arc_must_save_register (regno, cfun)) |
526b7aee SV |
2780 | { |
2781 | reg_size += UNITS_PER_WORD; | |
41453183 | 2782 | gmask |= 1L << regno; |
526b7aee SV |
2783 | } |
2784 | } | |
2785 | ||
6fe5e235 CZ |
2786 | /* In a frame that calls __builtin_eh_return two data registers are |
2787 | used to pass values back to the exception handler. | |
2788 | ||
2789 | Ensure that these registers are spilled to the stack so that the | |
2790 | exception throw code can find them, and update the saved values. | |
2791 | The handling code will then consume these reloaded values to | |
2792 | handle the exception. */ | |
2793 | if (crtl->calls_eh_return) | |
2794 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
2795 | { | |
2796 | reg_size += UNITS_PER_WORD; | |
2797 | gmask |= 1 << regno; | |
2798 | } | |
2799 | ||
526b7aee SV |
2800 | /* 4) Space for back trace data structure. |
2801 | <return addr reg size> (if required) + <fp size> (if required). */ | |
2802 | frame_info->save_return_addr | |
6fe5e235 CZ |
2803 | = (!crtl->is_leaf || df_regs_ever_live_p (RETURN_ADDR_REGNUM) |
2804 | || crtl->calls_eh_return); | |
526b7aee | 2805 | /* Saving blink reg in case of leaf function for millicode thunk calls. */ |
6fe5e235 CZ |
2806 | if (optimize_size |
2807 | && !TARGET_NO_MILLICODE_THUNK_SET | |
2808 | && !crtl->calls_eh_return) | |
526b7aee SV |
2809 | { |
2810 | if (arc_compute_millicode_save_restore_regs (gmask, frame_info)) | |
2811 | frame_info->save_return_addr = true; | |
2812 | } | |
2813 | ||
2814 | extra_size = 0; | |
41453183 | 2815 | if (arc_must_save_return_addr (cfun)) |
526b7aee | 2816 | extra_size = 4; |
41453183 | 2817 | if (arc_frame_pointer_needed ()) |
526b7aee SV |
2818 | extra_size += 4; |
2819 | ||
2820 | /* 5) Space for variable arguments passed in registers */ | |
2821 | pretend_size = crtl->args.pretend_args_size; | |
2822 | ||
2823 | /* Ensure everything before the locals is aligned appropriately. */ | |
2824 | { | |
2825 | unsigned int extra_plus_reg_size; | |
2826 | unsigned int extra_plus_reg_size_aligned; | |
2827 | ||
2828 | extra_plus_reg_size = extra_size + reg_size; | |
2829 | extra_plus_reg_size_aligned = ARC_STACK_ALIGN(extra_plus_reg_size); | |
2830 | reg_size = extra_plus_reg_size_aligned - extra_size; | |
2831 | } | |
2832 | ||
2833 | /* Compute total frame size. */ | |
2834 | total_size = var_size + args_size + extra_size + pretend_size + reg_size; | |
2835 | ||
6fe5e235 CZ |
2836 | /* It used to be the case that the alignment was forced at this |
2837 | point. However, that is dangerous, calculations based on | |
2838 | total_size would be wrong. Given that this has never cropped up | |
2839 | as an issue I've changed this to an assert for now. */ | |
2840 | gcc_assert (total_size == ARC_STACK_ALIGN (total_size)); | |
526b7aee SV |
2841 | |
2842 | /* Compute offset of register save area from stack pointer: | |
fb155425 | 2843 | Frame: pretend_size <blink> reg_size <fp> var_size args_size <--sp |
526b7aee SV |
2844 | */ |
2845 | reg_offset = (total_size - (pretend_size + reg_size + extra_size) | |
41453183 | 2846 | + (arc_frame_pointer_needed () ? 4 : 0)); |
526b7aee SV |
2847 | |
2848 | /* Save computed information. */ | |
2849 | frame_info->total_size = total_size; | |
2850 | frame_info->extra_size = extra_size; | |
2851 | frame_info->pretend_size = pretend_size; | |
2852 | frame_info->var_size = var_size; | |
2853 | frame_info->args_size = args_size; | |
2854 | frame_info->reg_size = reg_size; | |
2855 | frame_info->reg_offset = reg_offset; | |
2856 | frame_info->gmask = gmask; | |
2857 | frame_info->initialized = reload_completed; | |
2858 | ||
2859 | /* Ok, we're done. */ | |
2860 | return total_size; | |
2861 | } | |
2862 | ||
2863 | /* Common code to save/restore registers. */ | |
2864 | /* BASE_REG is the base register to use for addressing and to adjust. | |
2865 | GMASK is a bitmask of general purpose registers to save/restore. | |
2866 | epilogue_p 0: prologue 1:epilogue 2:epilogue, sibling thunk | |
2867 | If *FIRST_OFFSET is non-zero, add it first to BASE_REG - preferably | |
2868 | using a pre-modify for the first memory access. *FIRST_OFFSET is then | |
2869 | zeroed. */ | |
2870 | ||
2871 | static void | |
2872 | arc_save_restore (rtx base_reg, | |
2873 | unsigned int gmask, int epilogue_p, int *first_offset) | |
2874 | { | |
2875 | unsigned int offset = 0; | |
2876 | int regno; | |
2877 | struct arc_frame_info *frame = &cfun->machine->frame_info; | |
2878 | rtx sibthunk_insn = NULL_RTX; | |
2879 | ||
2880 | if (gmask) | |
2881 | { | |
2882 | /* Millicode thunks implementation: | |
2883 | Generates calls to millicodes for registers starting from r13 to r25 | |
2884 | Present Limitations: | |
2885 | - Only one range supported. The remaining regs will have the ordinary | |
2886 | st and ld instructions for store and loads. Hence a gmask asking | |
2887 | to store r13-14, r16-r25 will only generate calls to store and | |
2888 | load r13 to r14 while store and load insns will be generated for | |
2889 | r16 to r25 in the prologue and epilogue respectively. | |
2890 | ||
2891 | - Presently library only supports register ranges starting from r13. | |
2892 | */ | |
2893 | if (epilogue_p == 2 || frame->millicode_end_reg > 14) | |
2894 | { | |
2895 | int start_call = frame->millicode_start_reg; | |
2896 | int end_call = frame->millicode_end_reg; | |
2897 | int n_regs = end_call - start_call + 1; | |
2898 | int i = 0, r, off = 0; | |
2899 | rtx insn; | |
2900 | rtx ret_addr = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
2901 | ||
2902 | if (*first_offset) | |
2903 | { | |
2904 | /* "reg_size" won't be more than 127 . */ | |
6ace1161 | 2905 | gcc_assert (epilogue_p || abs (*first_offset) <= 127); |
526b7aee SV |
2906 | frame_add (base_reg, *first_offset); |
2907 | *first_offset = 0; | |
2908 | } | |
2909 | insn = gen_rtx_PARALLEL | |
2910 | (VOIDmode, rtvec_alloc ((epilogue_p == 2) + n_regs + 1)); | |
2911 | if (epilogue_p == 2) | |
2912 | i += 2; | |
2913 | else | |
2914 | XVECEXP (insn, 0, n_regs) = gen_rtx_CLOBBER (VOIDmode, ret_addr); | |
2915 | for (r = start_call; r <= end_call; r++, off += UNITS_PER_WORD, i++) | |
2916 | { | |
2917 | rtx reg = gen_rtx_REG (SImode, r); | |
2918 | rtx mem | |
2919 | = gen_frame_mem (SImode, plus_constant (Pmode, base_reg, off)); | |
2920 | ||
2921 | if (epilogue_p) | |
f7df4a84 | 2922 | XVECEXP (insn, 0, i) = gen_rtx_SET (reg, mem); |
526b7aee | 2923 | else |
f7df4a84 | 2924 | XVECEXP (insn, 0, i) = gen_rtx_SET (mem, reg); |
526b7aee SV |
2925 | gmask = gmask & ~(1L << r); |
2926 | } | |
2927 | if (epilogue_p == 2) | |
2928 | sibthunk_insn = insn; | |
2929 | else | |
67a96300 CZ |
2930 | { |
2931 | insn = frame_insn (insn); | |
b0c7ddf8 CM |
2932 | for (r = start_call, off = 0; |
2933 | r <= end_call; | |
2934 | r++, off += UNITS_PER_WORD) | |
2935 | { | |
2936 | rtx reg = gen_rtx_REG (SImode, r); | |
2937 | if (epilogue_p) | |
2938 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
2939 | else | |
2940 | { | |
2941 | rtx mem = gen_rtx_MEM (SImode, plus_constant (Pmode, | |
2942 | base_reg, | |
2943 | off)); | |
2944 | ||
2945 | add_reg_note (insn, REG_CFA_OFFSET, | |
2946 | gen_rtx_SET (mem, reg)); | |
2947 | } | |
2948 | } | |
67a96300 | 2949 | } |
526b7aee SV |
2950 | offset += off; |
2951 | } | |
2952 | ||
2953 | for (regno = 0; regno <= 31; regno++) | |
2954 | { | |
cd1e4d41 | 2955 | machine_mode mode = SImode; |
d34a0fdc CZ |
2956 | bool found = false; |
2957 | ||
2958 | if (TARGET_LL64 | |
2959 | && (regno % 2 == 0) | |
2960 | && ((gmask & (1L << regno)) != 0) | |
2961 | && ((gmask & (1L << (regno+1))) != 0)) | |
2962 | { | |
2963 | found = true; | |
2964 | mode = DImode; | |
2965 | } | |
2966 | else if ((gmask & (1L << regno)) != 0) | |
526b7aee | 2967 | { |
d34a0fdc CZ |
2968 | found = true; |
2969 | mode = SImode; | |
2970 | } | |
2971 | ||
2972 | if (found) | |
2973 | { | |
2974 | rtx reg = gen_rtx_REG (mode, regno); | |
526b7aee | 2975 | rtx addr, mem; |
67a96300 | 2976 | int cfa_adjust = *first_offset; |
526b7aee SV |
2977 | |
2978 | if (*first_offset) | |
2979 | { | |
2980 | gcc_assert (!offset); | |
2981 | addr = plus_constant (Pmode, base_reg, *first_offset); | |
2982 | addr = gen_rtx_PRE_MODIFY (Pmode, base_reg, addr); | |
2983 | *first_offset = 0; | |
2984 | } | |
2985 | else | |
2986 | { | |
2987 | gcc_assert (SMALL_INT (offset)); | |
2988 | addr = plus_constant (Pmode, base_reg, offset); | |
2989 | } | |
d34a0fdc | 2990 | mem = gen_frame_mem (mode, addr); |
526b7aee | 2991 | if (epilogue_p) |
67a96300 CZ |
2992 | { |
2993 | rtx insn = | |
2994 | frame_move_inc (reg, mem, base_reg, addr); | |
2995 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
2996 | if (cfa_adjust) | |
2997 | { | |
2998 | enum reg_note note = REG_CFA_ADJUST_CFA; | |
2999 | add_reg_note (insn, note, | |
3000 | gen_rtx_SET (stack_pointer_rtx, | |
3001 | plus_constant (Pmode, | |
3002 | stack_pointer_rtx, | |
3003 | cfa_adjust))); | |
3004 | } | |
3005 | } | |
526b7aee SV |
3006 | else |
3007 | frame_move_inc (mem, reg, base_reg, addr); | |
3008 | offset += UNITS_PER_WORD; | |
d34a0fdc CZ |
3009 | if (mode == DImode) |
3010 | { | |
3011 | offset += UNITS_PER_WORD; | |
3012 | ++regno; | |
3013 | } | |
526b7aee SV |
3014 | } /* if */ |
3015 | } /* for */ | |
3016 | }/* if */ | |
3017 | if (sibthunk_insn) | |
3018 | { | |
67a96300 CZ |
3019 | int start_call = frame->millicode_start_reg; |
3020 | int end_call = frame->millicode_end_reg; | |
3021 | int r; | |
3022 | ||
526b7aee SV |
3023 | rtx r12 = gen_rtx_REG (Pmode, 12); |
3024 | ||
f7df4a84 | 3025 | frame_insn (gen_rtx_SET (r12, GEN_INT (offset))); |
526b7aee SV |
3026 | XVECEXP (sibthunk_insn, 0, 0) = ret_rtx; |
3027 | XVECEXP (sibthunk_insn, 0, 1) | |
f7df4a84 | 3028 | = gen_rtx_SET (stack_pointer_rtx, |
526b7aee SV |
3029 | gen_rtx_PLUS (Pmode, stack_pointer_rtx, r12)); |
3030 | sibthunk_insn = emit_jump_insn (sibthunk_insn); | |
3031 | RTX_FRAME_RELATED_P (sibthunk_insn) = 1; | |
67a96300 CZ |
3032 | |
3033 | /* Would be nice if we could do this earlier, when the PARALLEL | |
3034 | is populated, but these need to be attached after the | |
3035 | emit. */ | |
3036 | for (r = start_call; r <= end_call; r++) | |
3037 | { | |
3038 | rtx reg = gen_rtx_REG (SImode, r); | |
3039 | add_reg_note (sibthunk_insn, REG_CFA_RESTORE, reg); | |
3040 | } | |
526b7aee SV |
3041 | } |
3042 | } /* arc_save_restore */ | |
3043 | ||
41453183 CZ |
3044 | /* Build dwarf information when the context is saved via AUX_IRQ_CTRL |
3045 | mechanism. */ | |
3046 | ||
3047 | static void | |
3048 | arc_dwarf_emit_irq_save_regs (void) | |
3049 | { | |
3050 | rtx tmp, par, insn, reg; | |
3051 | int i, offset, j; | |
3052 | ||
3053 | par = gen_rtx_SEQUENCE (VOIDmode, | |
3054 | rtvec_alloc (irq_ctrl_saved.irq_save_last_reg + 1 | |
3055 | + irq_ctrl_saved.irq_save_blink | |
3056 | + irq_ctrl_saved.irq_save_lpcount | |
3057 | + 1)); | |
3058 | ||
3059 | /* Build the stack adjustment note for unwind info. */ | |
3060 | j = 0; | |
3061 | offset = UNITS_PER_WORD * (irq_ctrl_saved.irq_save_last_reg + 1 | |
3062 | + irq_ctrl_saved.irq_save_blink | |
3063 | + irq_ctrl_saved.irq_save_lpcount); | |
3064 | tmp = plus_constant (Pmode, stack_pointer_rtx, -1 * offset); | |
3065 | tmp = gen_rtx_SET (stack_pointer_rtx, tmp); | |
3066 | RTX_FRAME_RELATED_P (tmp) = 1; | |
3067 | XVECEXP (par, 0, j++) = tmp; | |
3068 | ||
3069 | offset -= UNITS_PER_WORD; | |
3070 | ||
3071 | /* 1st goes LP_COUNT. */ | |
3072 | if (irq_ctrl_saved.irq_save_lpcount) | |
3073 | { | |
3074 | reg = gen_rtx_REG (SImode, 60); | |
3075 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
3076 | tmp = gen_frame_mem (SImode, tmp); | |
3077 | tmp = gen_rtx_SET (tmp, reg); | |
3078 | RTX_FRAME_RELATED_P (tmp) = 1; | |
3079 | XVECEXP (par, 0, j++) = tmp; | |
3080 | offset -= UNITS_PER_WORD; | |
3081 | } | |
3082 | ||
3083 | /* 2nd goes BLINK. */ | |
3084 | if (irq_ctrl_saved.irq_save_blink) | |
3085 | { | |
3086 | reg = gen_rtx_REG (SImode, 31); | |
3087 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
3088 | tmp = gen_frame_mem (SImode, tmp); | |
3089 | tmp = gen_rtx_SET (tmp, reg); | |
3090 | RTX_FRAME_RELATED_P (tmp) = 1; | |
3091 | XVECEXP (par, 0, j++) = tmp; | |
3092 | offset -= UNITS_PER_WORD; | |
3093 | } | |
3094 | ||
3095 | /* Build the parallel of the remaining registers recorded as saved | |
3096 | for unwind. */ | |
3097 | for (i = irq_ctrl_saved.irq_save_last_reg; i >= 0; i--) | |
3098 | { | |
3099 | reg = gen_rtx_REG (SImode, i); | |
3100 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
3101 | tmp = gen_frame_mem (SImode, tmp); | |
3102 | tmp = gen_rtx_SET (tmp, reg); | |
3103 | RTX_FRAME_RELATED_P (tmp) = 1; | |
3104 | XVECEXP (par, 0, j++) = tmp; | |
3105 | offset -= UNITS_PER_WORD; | |
3106 | } | |
3107 | ||
3108 | /* Dummy insn used to anchor the dwarf info. */ | |
3109 | insn = emit_insn (gen_stack_irq_dwarf()); | |
3110 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, par); | |
3111 | RTX_FRAME_RELATED_P (insn) = 1; | |
3112 | } | |
3113 | ||
526b7aee SV |
3114 | /* Set up the stack and frame pointer (if desired) for the function. */ |
3115 | ||
3116 | void | |
3117 | arc_expand_prologue (void) | |
3118 | { | |
6fe5e235 | 3119 | int size; |
526b7aee SV |
3120 | unsigned int gmask = cfun->machine->frame_info.gmask; |
3121 | /* unsigned int frame_pointer_offset;*/ | |
3122 | unsigned int frame_size_to_allocate; | |
3123 | /* (FIXME: The first store will use a PRE_MODIFY; this will usually be r13. | |
3124 | Change the stack layout so that we rather store a high register with the | |
3125 | PRE_MODIFY, thus enabling more short insn generation.) */ | |
3126 | int first_offset = 0; | |
1825c61e CZ |
3127 | unsigned int fn_type = arc_compute_function_type (cfun); |
3128 | ||
3129 | /* Naked functions don't have prologue. */ | |
3130 | if (ARC_NAKED_P (fn_type)) | |
3131 | return; | |
526b7aee | 3132 | |
6fe5e235 CZ |
3133 | /* Compute total frame size. */ |
3134 | size = arc_compute_frame_size (); | |
526b7aee SV |
3135 | |
3136 | if (flag_stack_usage_info) | |
3137 | current_function_static_stack_size = size; | |
3138 | ||
3139 | /* Keep track of frame size to be allocated. */ | |
3140 | frame_size_to_allocate = size; | |
3141 | ||
3142 | /* These cases shouldn't happen. Catch them now. */ | |
3143 | gcc_assert (!(size == 0 && gmask)); | |
3144 | ||
3145 | /* Allocate space for register arguments if this is a variadic function. */ | |
3146 | if (cfun->machine->frame_info.pretend_size != 0) | |
3147 | { | |
3148 | /* Ensure pretend_size is maximum of 8 * word_size. */ | |
3149 | gcc_assert (cfun->machine->frame_info.pretend_size <= 32); | |
3150 | ||
3151 | frame_stack_add (-(HOST_WIDE_INT)cfun->machine->frame_info.pretend_size); | |
3152 | frame_size_to_allocate -= cfun->machine->frame_info.pretend_size; | |
3153 | } | |
3154 | ||
41453183 CZ |
3155 | /* IRQ using automatic save mechanism will save the register before |
3156 | anything we do. */ | |
c7314bc1 CZ |
3157 | if (ARC_AUTO_IRQ_P (fn_type) |
3158 | && !ARC_FAST_INTERRUPT_P (fn_type)) | |
41453183 CZ |
3159 | { |
3160 | arc_dwarf_emit_irq_save_regs (); | |
3161 | } | |
3162 | ||
526b7aee | 3163 | /* The home-grown ABI says link register is saved first. */ |
41453183 CZ |
3164 | if (arc_must_save_return_addr (cfun) |
3165 | && !ARC_AUTOBLINK_IRQ_P (fn_type)) | |
526b7aee SV |
3166 | { |
3167 | rtx ra = gen_rtx_REG (SImode, RETURN_ADDR_REGNUM); | |
41453183 CZ |
3168 | rtx mem = gen_frame_mem (Pmode, |
3169 | gen_rtx_PRE_DEC (Pmode, | |
3170 | stack_pointer_rtx)); | |
526b7aee SV |
3171 | |
3172 | frame_move_inc (mem, ra, stack_pointer_rtx, 0); | |
3173 | frame_size_to_allocate -= UNITS_PER_WORD; | |
41453183 | 3174 | } |
526b7aee SV |
3175 | |
3176 | /* Save any needed call-saved regs (and call-used if this is an | |
3177 | interrupt handler) for ARCompact ISA. */ | |
3178 | if (cfun->machine->frame_info.reg_size) | |
3179 | { | |
3180 | first_offset = -cfun->machine->frame_info.reg_size; | |
3181 | /* N.B. FRAME_POINTER_MASK and RETURN_ADDR_MASK are cleared in gmask. */ | |
3182 | arc_save_restore (stack_pointer_rtx, gmask, 0, &first_offset); | |
3183 | frame_size_to_allocate -= cfun->machine->frame_info.reg_size; | |
3184 | } | |
3185 | ||
b0c7ddf8 CM |
3186 | /* In the case of millicode thunk, we need to restore the clobbered |
3187 | blink register. */ | |
3188 | if (cfun->machine->frame_info.millicode_end_reg > 0 | |
3189 | && arc_must_save_return_addr (cfun)) | |
3190 | { | |
3191 | HOST_WIDE_INT tmp = cfun->machine->frame_info.reg_size; | |
3192 | emit_insn (gen_rtx_SET (gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM), | |
3193 | gen_rtx_MEM (Pmode, | |
3194 | plus_constant (Pmode, | |
3195 | stack_pointer_rtx, | |
3196 | tmp)))); | |
3197 | } | |
3198 | ||
41453183 CZ |
3199 | /* Save frame pointer if needed. First save the FP on stack, if not |
3200 | autosaved. */ | |
3201 | if (arc_frame_pointer_needed () | |
3202 | && !ARC_AUTOFP_IRQ_P (fn_type)) | |
526b7aee SV |
3203 | { |
3204 | rtx addr = gen_rtx_PLUS (Pmode, stack_pointer_rtx, | |
3205 | GEN_INT (-UNITS_PER_WORD + first_offset)); | |
3206 | rtx mem = gen_frame_mem (Pmode, gen_rtx_PRE_MODIFY (Pmode, | |
3207 | stack_pointer_rtx, | |
3208 | addr)); | |
3209 | frame_move_inc (mem, frame_pointer_rtx, stack_pointer_rtx, 0); | |
3210 | frame_size_to_allocate -= UNITS_PER_WORD; | |
3211 | first_offset = 0; | |
41453183 CZ |
3212 | } |
3213 | ||
3214 | /* Emit mov fp,sp. */ | |
3215 | if (arc_frame_pointer_needed ()) | |
3216 | { | |
526b7aee SV |
3217 | frame_move (frame_pointer_rtx, stack_pointer_rtx); |
3218 | } | |
3219 | ||
3220 | /* ??? We don't handle the case where the saved regs are more than 252 | |
3221 | bytes away from sp. This can be handled by decrementing sp once, saving | |
3222 | the regs, and then decrementing it again. The epilogue doesn't have this | |
3223 | problem as the `ld' insn takes reg+limm values (though it would be more | |
3224 | efficient to avoid reg+limm). */ | |
3225 | ||
3226 | frame_size_to_allocate -= first_offset; | |
3227 | /* Allocate the stack frame. */ | |
3228 | if (frame_size_to_allocate > 0) | |
2daad50b CZ |
3229 | { |
3230 | frame_stack_add ((HOST_WIDE_INT) 0 - frame_size_to_allocate); | |
3231 | /* If the frame pointer is needed, emit a special barrier that | |
3232 | will prevent the scheduler from moving stores to the frame | |
3233 | before the stack adjustment. */ | |
3234 | if (arc_frame_pointer_needed ()) | |
3235 | emit_insn (gen_stack_tie (stack_pointer_rtx, | |
3236 | hard_frame_pointer_rtx)); | |
3237 | } | |
526b7aee SV |
3238 | |
3239 | /* Setup the gp register, if needed. */ | |
3240 | if (crtl->uses_pic_offset_table) | |
3241 | arc_finalize_pic (); | |
3242 | } | |
3243 | ||
3244 | /* Do any necessary cleanup after a function to restore stack, frame, | |
3245 | and regs. */ | |
3246 | ||
3247 | void | |
3248 | arc_expand_epilogue (int sibcall_p) | |
3249 | { | |
6fe5e235 | 3250 | int size; |
1825c61e | 3251 | unsigned int fn_type = arc_compute_function_type (cfun); |
526b7aee | 3252 | |
6fe5e235 | 3253 | size = arc_compute_frame_size (); |
526b7aee | 3254 | |
5719867d JR |
3255 | unsigned int pretend_size = cfun->machine->frame_info.pretend_size; |
3256 | unsigned int frame_size; | |
3257 | unsigned int size_to_deallocate; | |
3258 | int restored; | |
3259 | int can_trust_sp_p = !cfun->calls_alloca; | |
3260 | int first_offset = 0; | |
3261 | int millicode_p = cfun->machine->frame_info.millicode_end_reg > 0; | |
67a96300 | 3262 | rtx insn; |
526b7aee | 3263 | |
1825c61e CZ |
3264 | /* Naked functions don't have epilogue. */ |
3265 | if (ARC_NAKED_P (fn_type)) | |
3266 | return; | |
3267 | ||
5719867d | 3268 | size_to_deallocate = size; |
526b7aee | 3269 | |
5719867d JR |
3270 | frame_size = size - (pretend_size + |
3271 | cfun->machine->frame_info.reg_size + | |
3272 | cfun->machine->frame_info.extra_size); | |
526b7aee | 3273 | |
5719867d JR |
3274 | /* ??? There are lots of optimizations that can be done here. |
3275 | EG: Use fp to restore regs if it's closer. | |
3276 | Maybe in time we'll do them all. For now, always restore regs from | |
3277 | sp, but don't restore sp if we don't have to. */ | |
526b7aee | 3278 | |
5719867d | 3279 | if (!can_trust_sp_p) |
41453183 | 3280 | gcc_assert (arc_frame_pointer_needed ()); |
526b7aee | 3281 | |
5719867d JR |
3282 | /* Restore stack pointer to the beginning of saved register area for |
3283 | ARCompact ISA. */ | |
3284 | if (frame_size) | |
3285 | { | |
41453183 | 3286 | if (arc_frame_pointer_needed ()) |
5719867d JR |
3287 | frame_move (stack_pointer_rtx, frame_pointer_rtx); |
3288 | else | |
3289 | first_offset = frame_size; | |
3290 | size_to_deallocate -= frame_size; | |
3291 | } | |
3292 | else if (!can_trust_sp_p) | |
3293 | frame_stack_add (-frame_size); | |
526b7aee | 3294 | |
526b7aee | 3295 | |
5719867d | 3296 | /* Restore any saved registers. */ |
41453183 CZ |
3297 | if (arc_frame_pointer_needed () |
3298 | && !ARC_AUTOFP_IRQ_P (fn_type)) | |
5719867d | 3299 | { |
67a96300 CZ |
3300 | rtx addr = gen_rtx_POST_INC (Pmode, stack_pointer_rtx); |
3301 | ||
3302 | insn = frame_move_inc (frame_pointer_rtx, gen_frame_mem (Pmode, addr), | |
3303 | stack_pointer_rtx, 0); | |
3304 | add_reg_note (insn, REG_CFA_RESTORE, frame_pointer_rtx); | |
3305 | add_reg_note (insn, REG_CFA_DEF_CFA, | |
3306 | plus_constant (SImode, stack_pointer_rtx, | |
3307 | 4)); | |
3308 | size_to_deallocate -= UNITS_PER_WORD; | |
5719867d JR |
3309 | } |
3310 | ||
3311 | /* Load blink after the calls to thunk calls in case of optimize size. */ | |
3312 | if (millicode_p) | |
3313 | { | |
3314 | int sibthunk_p = (!sibcall_p | |
3315 | && fn_type == ARC_FUNCTION_NORMAL | |
3316 | && !cfun->machine->frame_info.pretend_size); | |
3317 | ||
3318 | gcc_assert (!(cfun->machine->frame_info.gmask | |
3319 | & (FRAME_POINTER_MASK | RETURN_ADDR_MASK))); | |
3320 | arc_save_restore (stack_pointer_rtx, | |
3321 | cfun->machine->frame_info.gmask, | |
3322 | 1 + sibthunk_p, &first_offset); | |
3323 | if (sibthunk_p) | |
67a96300 | 3324 | return; |
5719867d JR |
3325 | } |
3326 | /* If we are to restore registers, and first_offset would require | |
3327 | a limm to be encoded in a PRE_MODIFY, yet we can add it with a | |
3328 | fast add to the stack pointer, do this now. */ | |
3329 | if ((!SMALL_INT (first_offset) | |
3330 | && cfun->machine->frame_info.gmask | |
3331 | && ((TARGET_ARC700 && !optimize_size) | |
3332 | ? first_offset <= 0x800 | |
3333 | : satisfies_constraint_C2a (GEN_INT (first_offset)))) | |
3334 | /* Also do this if we have both gprs and return | |
3335 | address to restore, and they both would need a LIMM. */ | |
41453183 CZ |
3336 | || (arc_must_save_return_addr (cfun) |
3337 | && !SMALL_INT ((cfun->machine->frame_info.reg_size + first_offset) >> 2) | |
3338 | && cfun->machine->frame_info.gmask)) | |
5719867d JR |
3339 | { |
3340 | frame_stack_add (first_offset); | |
3341 | first_offset = 0; | |
3342 | } | |
41453183 CZ |
3343 | if (arc_must_save_return_addr (cfun) |
3344 | && !ARC_AUTOBLINK_IRQ_P (fn_type)) | |
5719867d JR |
3345 | { |
3346 | rtx ra = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3347 | int ra_offs = cfun->machine->frame_info.reg_size + first_offset; | |
3348 | rtx addr = plus_constant (Pmode, stack_pointer_rtx, ra_offs); | |
67a96300 | 3349 | HOST_WIDE_INT cfa_adjust = 0; |
5719867d JR |
3350 | |
3351 | /* If the load of blink would need a LIMM, but we can add | |
3352 | the offset quickly to sp, do the latter. */ | |
3353 | if (!SMALL_INT (ra_offs >> 2) | |
3354 | && !cfun->machine->frame_info.gmask | |
3355 | && ((TARGET_ARC700 && !optimize_size) | |
3356 | ? ra_offs <= 0x800 | |
3357 | : satisfies_constraint_C2a (GEN_INT (ra_offs)))) | |
526b7aee | 3358 | { |
5719867d JR |
3359 | size_to_deallocate -= ra_offs - first_offset; |
3360 | first_offset = 0; | |
3361 | frame_stack_add (ra_offs); | |
3362 | ra_offs = 0; | |
3363 | addr = stack_pointer_rtx; | |
526b7aee | 3364 | } |
5719867d JR |
3365 | /* See if we can combine the load of the return address with the |
3366 | final stack adjustment. | |
3367 | We need a separate load if there are still registers to | |
3368 | restore. We also want a separate load if the combined insn | |
3369 | would need a limm, but a separate load doesn't. */ | |
3370 | if (ra_offs | |
3371 | && !cfun->machine->frame_info.gmask | |
3372 | && (SMALL_INT (ra_offs) || !SMALL_INT (ra_offs >> 2))) | |
526b7aee | 3373 | { |
5719867d | 3374 | addr = gen_rtx_PRE_MODIFY (Pmode, stack_pointer_rtx, addr); |
67a96300 | 3375 | cfa_adjust = ra_offs; |
526b7aee | 3376 | first_offset = 0; |
5719867d | 3377 | size_to_deallocate -= cfun->machine->frame_info.reg_size; |
526b7aee | 3378 | } |
5719867d | 3379 | else if (!ra_offs && size_to_deallocate == UNITS_PER_WORD) |
526b7aee | 3380 | { |
5719867d | 3381 | addr = gen_rtx_POST_INC (Pmode, addr); |
67a96300 | 3382 | cfa_adjust = GET_MODE_SIZE (Pmode); |
5719867d | 3383 | size_to_deallocate = 0; |
526b7aee | 3384 | } |
67a96300 CZ |
3385 | |
3386 | insn = frame_move_inc (ra, gen_frame_mem (Pmode, addr), | |
3387 | stack_pointer_rtx, addr); | |
3388 | if (cfa_adjust) | |
3389 | { | |
3390 | enum reg_note note = REG_CFA_ADJUST_CFA; | |
3391 | ||
3392 | add_reg_note (insn, note, | |
3393 | gen_rtx_SET (stack_pointer_rtx, | |
3394 | plus_constant (SImode, stack_pointer_rtx, | |
3395 | cfa_adjust))); | |
3396 | } | |
3397 | add_reg_note (insn, REG_CFA_RESTORE, ra); | |
5719867d | 3398 | } |
526b7aee | 3399 | |
5719867d JR |
3400 | if (!millicode_p) |
3401 | { | |
3402 | if (cfun->machine->frame_info.reg_size) | |
3403 | arc_save_restore (stack_pointer_rtx, | |
3404 | /* The zeroing of these two bits is unnecessary, but leave this in for clarity. */ | |
3405 | cfun->machine->frame_info.gmask | |
3406 | & ~(FRAME_POINTER_MASK | RETURN_ADDR_MASK), 1, &first_offset); | |
3407 | } | |
526b7aee | 3408 | |
5719867d JR |
3409 | /* The rest of this function does the following: |
3410 | ARCompact : handle epilogue_delay, restore sp (phase-2), return | |
3411 | */ | |
526b7aee | 3412 | |
5719867d JR |
3413 | /* Keep track of how much of the stack pointer we've restored. |
3414 | It makes the following a lot more readable. */ | |
3415 | size_to_deallocate += first_offset; | |
3416 | restored = size - size_to_deallocate; | |
526b7aee | 3417 | |
5719867d JR |
3418 | if (size > restored) |
3419 | frame_stack_add (size - restored); | |
67a96300 | 3420 | |
6fe5e235 CZ |
3421 | /* For frames that use __builtin_eh_return, the register defined by |
3422 | EH_RETURN_STACKADJ_RTX is set to 0 for all standard return paths. | |
3423 | On eh_return paths however, the register is set to the value that | |
3424 | should be added to the stack pointer in order to restore the | |
3425 | correct stack pointer for the exception handling frame. | |
3426 | ||
3427 | For ARC we are going to use r2 for EH_RETURN_STACKADJ_RTX, add | |
3428 | this onto the stack for eh_return frames. */ | |
3429 | if (crtl->calls_eh_return) | |
3430 | emit_insn (gen_add2_insn (stack_pointer_rtx, | |
3431 | EH_RETURN_STACKADJ_RTX)); | |
3432 | ||
5719867d JR |
3433 | /* Emit the return instruction. */ |
3434 | if (sibcall_p == FALSE) | |
3435 | emit_jump_insn (gen_simple_return ()); | |
526b7aee SV |
3436 | } |
3437 | ||
6fe5e235 CZ |
3438 | /* Return rtx for the location of the return address on the stack, |
3439 | suitable for use in __builtin_eh_return. The new return address | |
3440 | will be written to this location in order to redirect the return to | |
3441 | the exception handler. */ | |
526b7aee | 3442 | |
6fe5e235 CZ |
3443 | rtx |
3444 | arc_eh_return_address_location (void) | |
3445 | { | |
3446 | rtx mem; | |
3447 | int offset; | |
3448 | struct arc_frame_info *afi; | |
3449 | ||
3450 | arc_compute_frame_size (); | |
3451 | afi = &cfun->machine->frame_info; | |
3452 | ||
3453 | gcc_assert (crtl->calls_eh_return); | |
3454 | gcc_assert (afi->save_return_addr); | |
3455 | gcc_assert (afi->extra_size >= 4); | |
3456 | ||
3457 | /* The '-4' removes the size of the return address, which is | |
3458 | included in the 'extra_size' field. */ | |
3459 | offset = afi->reg_size + afi->extra_size - 4; | |
3460 | mem = gen_frame_mem (Pmode, | |
3461 | plus_constant (Pmode, frame_pointer_rtx, offset)); | |
3462 | ||
3463 | /* The following should not be needed, and is, really a hack. The | |
3464 | issue being worked around here is that the DSE (Dead Store | |
3465 | Elimination) pass will remove this write to the stack as it sees | |
3466 | a single store and no corresponding read. The read however | |
3467 | occurs in the epilogue code, which is not added into the function | |
3468 | rtl until a later pass. So, at the time of DSE, the decision to | |
3469 | remove this store seems perfectly sensible. Marking the memory | |
3470 | address as volatile obviously has the effect of preventing DSE | |
3471 | from removing the store. */ | |
3472 | MEM_VOLATILE_P (mem) = 1; | |
3473 | return mem; | |
526b7aee SV |
3474 | } |
3475 | ||
3476 | /* PIC */ | |
3477 | ||
5a5c5784 CZ |
3478 | /* Helper to generate unspec constant. */ |
3479 | ||
3480 | static rtx | |
3481 | arc_unspec_offset (rtx loc, int unspec) | |
3482 | { | |
3483 | return gen_rtx_CONST (Pmode, gen_rtx_UNSPEC (Pmode, gen_rtvec (1, loc), | |
3484 | unspec)); | |
3485 | } | |
3486 | ||
526b7aee SV |
3487 | /* Emit special PIC prologues and epilogues. */ |
3488 | /* If the function has any GOTOFF relocations, then the GOTBASE | |
3489 | register has to be setup in the prologue | |
3490 | The instruction needed at the function start for setting up the | |
3491 | GOTBASE register is | |
3492 | add rdest, pc, | |
3493 | ---------------------------------------------------------- | |
3494 | The rtl to be emitted for this should be: | |
3495 | set (reg basereg) | |
3496 | (plus (reg pc) | |
3497 | (const (unspec (symref _DYNAMIC) 3))) | |
3498 | ---------------------------------------------------------- */ | |
3499 | ||
3500 | static void | |
3501 | arc_finalize_pic (void) | |
3502 | { | |
3503 | rtx pat; | |
3504 | rtx baseptr_rtx = gen_rtx_REG (Pmode, PIC_OFFSET_TABLE_REGNUM); | |
3505 | ||
3506 | if (crtl->uses_pic_offset_table == 0) | |
3507 | return; | |
3508 | ||
3509 | gcc_assert (flag_pic != 0); | |
3510 | ||
3511 | pat = gen_rtx_SYMBOL_REF (Pmode, "_DYNAMIC"); | |
5a5c5784 | 3512 | pat = arc_unspec_offset (pat, ARC_UNSPEC_GOT); |
f7df4a84 | 3513 | pat = gen_rtx_SET (baseptr_rtx, pat); |
526b7aee SV |
3514 | |
3515 | emit_insn (pat); | |
3516 | } | |
3517 | \f | |
3518 | /* !TARGET_BARREL_SHIFTER support. */ | |
3519 | /* Emit a shift insn to set OP0 to OP1 shifted by OP2; CODE specifies what | |
3520 | kind of shift. */ | |
3521 | ||
3522 | void | |
3523 | emit_shift (enum rtx_code code, rtx op0, rtx op1, rtx op2) | |
3524 | { | |
3525 | rtx shift = gen_rtx_fmt_ee (code, SImode, op1, op2); | |
3526 | rtx pat | |
3527 | = ((shift4_operator (shift, SImode) ? gen_shift_si3 : gen_shift_si3_loop) | |
3528 | (op0, op1, op2, shift)); | |
3529 | emit_insn (pat); | |
3530 | } | |
3531 | ||
3532 | /* Output the assembler code for doing a shift. | |
3533 | We go to a bit of trouble to generate efficient code as the ARC601 only has | |
3534 | single bit shifts. This is taken from the h8300 port. We only have one | |
3535 | mode of shifting and can't access individual bytes like the h8300 can, so | |
3536 | this is greatly simplified (at the expense of not generating hyper- | |
3537 | efficient code). | |
3538 | ||
3539 | This function is not used if the variable shift insns are present. */ | |
3540 | ||
3541 | /* FIXME: This probably can be done using a define_split in arc.md. | |
3542 | Alternately, generate rtx rather than output instructions. */ | |
3543 | ||
3544 | const char * | |
3545 | output_shift (rtx *operands) | |
3546 | { | |
3547 | /* static int loopend_lab;*/ | |
3548 | rtx shift = operands[3]; | |
ef4bddc2 | 3549 | machine_mode mode = GET_MODE (shift); |
526b7aee SV |
3550 | enum rtx_code code = GET_CODE (shift); |
3551 | const char *shift_one; | |
3552 | ||
3553 | gcc_assert (mode == SImode); | |
3554 | ||
3555 | switch (code) | |
3556 | { | |
3557 | case ASHIFT: shift_one = "add %0,%1,%1"; break; | |
3558 | case ASHIFTRT: shift_one = "asr %0,%1"; break; | |
3559 | case LSHIFTRT: shift_one = "lsr %0,%1"; break; | |
3560 | default: gcc_unreachable (); | |
3561 | } | |
3562 | ||
3563 | if (GET_CODE (operands[2]) != CONST_INT) | |
3564 | { | |
3565 | output_asm_insn ("and.f lp_count,%2, 0x1f", operands); | |
3566 | goto shiftloop; | |
3567 | } | |
3568 | else | |
3569 | { | |
3570 | int n; | |
3571 | ||
3572 | n = INTVAL (operands[2]); | |
3573 | ||
3574 | /* Only consider the lower 5 bits of the shift count. */ | |
3575 | n = n & 0x1f; | |
3576 | ||
3577 | /* First see if we can do them inline. */ | |
3578 | /* ??? We could get better scheduling & shorter code (using short insns) | |
3579 | by using splitters. Alas, that'd be even more verbose. */ | |
3580 | if (code == ASHIFT && n <= 9 && n > 2 | |
3581 | && dest_reg_operand (operands[4], SImode)) | |
3582 | { | |
3583 | output_asm_insn ("mov %4,0\n\tadd3 %0,%4,%1", operands); | |
3584 | for (n -=3 ; n >= 3; n -= 3) | |
3585 | output_asm_insn ("add3 %0,%4,%0", operands); | |
3586 | if (n == 2) | |
3587 | output_asm_insn ("add2 %0,%4,%0", operands); | |
3588 | else if (n) | |
3589 | output_asm_insn ("add %0,%0,%0", operands); | |
3590 | } | |
3591 | else if (n <= 4) | |
3592 | { | |
3593 | while (--n >= 0) | |
3594 | { | |
3595 | output_asm_insn (shift_one, operands); | |
3596 | operands[1] = operands[0]; | |
3597 | } | |
3598 | } | |
3599 | /* See if we can use a rotate/and. */ | |
3600 | else if (n == BITS_PER_WORD - 1) | |
3601 | { | |
3602 | switch (code) | |
3603 | { | |
3604 | case ASHIFT : | |
3605 | output_asm_insn ("and %0,%1,1\n\tror %0,%0", operands); | |
3606 | break; | |
3607 | case ASHIFTRT : | |
3608 | /* The ARC doesn't have a rol insn. Use something else. */ | |
3609 | output_asm_insn ("add.f 0,%1,%1\n\tsbc %0,%0,%0", operands); | |
3610 | break; | |
3611 | case LSHIFTRT : | |
3612 | /* The ARC doesn't have a rol insn. Use something else. */ | |
3613 | output_asm_insn ("add.f 0,%1,%1\n\trlc %0,0", operands); | |
3614 | break; | |
3615 | default: | |
3616 | break; | |
3617 | } | |
3618 | } | |
3619 | else if (n == BITS_PER_WORD - 2 && dest_reg_operand (operands[4], SImode)) | |
3620 | { | |
3621 | switch (code) | |
3622 | { | |
3623 | case ASHIFT : | |
3624 | output_asm_insn ("and %0,%1,3\n\tror %0,%0\n\tror %0,%0", operands); | |
3625 | break; | |
3626 | case ASHIFTRT : | |
3627 | #if 1 /* Need some scheduling comparisons. */ | |
3628 | output_asm_insn ("add.f %4,%1,%1\n\tsbc %0,%0,%0\n\t" | |
3629 | "add.f 0,%4,%4\n\trlc %0,%0", operands); | |
3630 | #else | |
3631 | output_asm_insn ("add.f %4,%1,%1\n\tbxor %0,%4,31\n\t" | |
3632 | "sbc.f %0,%0,%4\n\trlc %0,%0", operands); | |
3633 | #endif | |
3634 | break; | |
3635 | case LSHIFTRT : | |
3636 | #if 1 | |
3637 | output_asm_insn ("add.f %4,%1,%1\n\trlc %0,0\n\t" | |
3638 | "add.f 0,%4,%4\n\trlc %0,%0", operands); | |
3639 | #else | |
3640 | output_asm_insn ("add.f %0,%1,%1\n\trlc.f %0,0\n\t" | |
3641 | "and %0,%0,1\n\trlc %0,%0", operands); | |
3642 | #endif | |
3643 | break; | |
3644 | default: | |
3645 | break; | |
3646 | } | |
3647 | } | |
3648 | else if (n == BITS_PER_WORD - 3 && code == ASHIFT) | |
3649 | output_asm_insn ("and %0,%1,7\n\tror %0,%0\n\tror %0,%0\n\tror %0,%0", | |
3650 | operands); | |
3651 | /* Must loop. */ | |
3652 | else | |
3653 | { | |
3654 | operands[2] = GEN_INT (n); | |
3655 | output_asm_insn ("mov.f lp_count, %2", operands); | |
3656 | ||
3657 | shiftloop: | |
3658 | { | |
3659 | output_asm_insn ("lpnz\t2f", operands); | |
3660 | output_asm_insn (shift_one, operands); | |
3661 | output_asm_insn ("nop", operands); | |
3662 | fprintf (asm_out_file, "2:\t%s end single insn loop\n", | |
3663 | ASM_COMMENT_START); | |
3664 | } | |
3665 | } | |
3666 | } | |
3667 | ||
3668 | return ""; | |
3669 | } | |
3670 | \f | |
3671 | /* Nested function support. */ | |
3672 | ||
efcc2e30 CZ |
3673 | /* Output assembler code for a block containing the constant parts of |
3674 | a trampoline, leaving space for variable parts. A trampoline looks | |
3675 | like this: | |
3676 | ||
3677 | ld_s r12,[pcl,8] | |
3678 | ld r11,[pcl,12] | |
3679 | j_s [r12] | |
3680 | .word function's address | |
3681 | .word static chain value | |
3682 | ||
3683 | */ | |
526b7aee SV |
3684 | |
3685 | static void | |
efcc2e30 | 3686 | arc_asm_trampoline_template (FILE *f) |
526b7aee | 3687 | { |
efcc2e30 CZ |
3688 | asm_fprintf (f, "\tld_s\t%s,[pcl,8]\n", ARC_TEMP_SCRATCH_REG); |
3689 | asm_fprintf (f, "\tld\t%s,[pcl,12]\n", reg_names[STATIC_CHAIN_REGNUM]); | |
3690 | asm_fprintf (f, "\tj_s\t[%s]\n", ARC_TEMP_SCRATCH_REG); | |
3691 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
3692 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
526b7aee SV |
3693 | } |
3694 | ||
3695 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
efcc2e30 CZ |
3696 | FNADDR is an RTX for the address of the function's pure code. CXT |
3697 | is an RTX for the static chain value for the function. | |
526b7aee SV |
3698 | |
3699 | The fastest trampoline to execute for trampolines within +-8KB of CTX | |
3700 | would be: | |
efcc2e30 | 3701 | |
526b7aee SV |
3702 | add2 r11,pcl,s12 |
3703 | j [limm] 0x20200f80 limm | |
efcc2e30 CZ |
3704 | |
3705 | and that would also be faster to write to the stack by computing | |
3706 | the offset from CTX to TRAMP at compile time. However, it would | |
3707 | really be better to get rid of the high cost of cache invalidation | |
3708 | when generating trampolines, which requires that the code part of | |
3709 | trampolines stays constant, and additionally either making sure | |
3710 | that no executable code but trampolines is on the stack, no icache | |
3711 | entries linger for the area of the stack from when before the stack | |
3712 | was allocated, and allocating trampolines in trampoline-only cache | |
3713 | lines or allocate trampolines fram a special pool of pre-allocated | |
3714 | trampolines. */ | |
526b7aee SV |
3715 | |
3716 | static void | |
3717 | arc_initialize_trampoline (rtx tramp, tree fndecl, rtx cxt) | |
3718 | { | |
3719 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); | |
3720 | ||
efcc2e30 CZ |
3721 | emit_block_move (tramp, assemble_trampoline_template (), |
3722 | GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL); | |
3723 | emit_move_insn (adjust_address (tramp, SImode, 8), fnaddr); | |
3724 | emit_move_insn (adjust_address (tramp, SImode, 12), cxt); | |
3725 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
3726 | LCT_NORMAL, VOIDmode, XEXP (tramp, 0), Pmode, | |
3727 | plus_constant (Pmode, XEXP (tramp, 0), TRAMPOLINE_SIZE), | |
3728 | Pmode); | |
526b7aee SV |
3729 | } |
3730 | ||
3731 | /* This is set briefly to 1 when we output a ".as" address modifer, and then | |
3732 | reset when we output the scaled address. */ | |
3733 | static int output_scaled = 0; | |
3734 | ||
3735 | /* Print operand X (an rtx) in assembler syntax to file FILE. | |
3736 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
3737 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
3738 | /* In final.c:output_asm_insn: | |
3739 | 'l' : label | |
3740 | 'a' : address | |
3741 | 'c' : constant address if CONSTANT_ADDRESS_P | |
3742 | 'n' : negative | |
3743 | Here: | |
3744 | 'Z': log2(x+1)-1 | |
3745 | 'z': log2 | |
3746 | 'M': log2(~x) | |
ceaaa9fe JR |
3747 | 'p': bit Position of lsb |
3748 | 's': size of bit field | |
526b7aee SV |
3749 | '#': condbranch delay slot suffix |
3750 | '*': jump delay slot suffix | |
3751 | '?' : nonjump-insn suffix for conditional execution or short instruction | |
3752 | '!' : jump / call suffix for conditional execution or short instruction | |
3753 | '`': fold constant inside unary o-perator, re-recognize, and emit. | |
3754 | 'd' | |
3755 | 'D' | |
3756 | 'R': Second word | |
3757 | 'S' | |
3758 | 'B': Branch comparison operand - suppress sda reference | |
3759 | 'H': Most significant word | |
3760 | 'L': Least significant word | |
3761 | 'A': ASCII decimal representation of floating point value | |
3762 | 'U': Load/store update or scaling indicator | |
3763 | 'V': cache bypass indicator for volatile | |
3764 | 'P' | |
3765 | 'F' | |
3766 | '^' | |
3767 | 'O': Operator | |
3768 | 'o': original symbol - no @ prepending. */ | |
3769 | ||
3770 | void | |
3771 | arc_print_operand (FILE *file, rtx x, int code) | |
3772 | { | |
3773 | switch (code) | |
3774 | { | |
3775 | case 'Z': | |
3776 | if (GET_CODE (x) == CONST_INT) | |
3777 | fprintf (file, "%d",exact_log2(INTVAL (x) + 1) - 1 ); | |
3778 | else | |
3779 | output_operand_lossage ("invalid operand to %%Z code"); | |
3780 | ||
3781 | return; | |
3782 | ||
3783 | case 'z': | |
3784 | if (GET_CODE (x) == CONST_INT) | |
3785 | fprintf (file, "%d",exact_log2(INTVAL (x)) ); | |
3786 | else | |
3787 | output_operand_lossage ("invalid operand to %%z code"); | |
3788 | ||
3789 | return; | |
3790 | ||
1e466f04 GM |
3791 | case 'c': |
3792 | if (GET_CODE (x) == CONST_INT) | |
41bc2c0b | 3793 | fprintf (file, "%ld", INTVAL (x) ); |
1e466f04 GM |
3794 | else |
3795 | output_operand_lossage ("invalid operands to %%c code"); | |
3796 | ||
3797 | return; | |
3798 | ||
526b7aee SV |
3799 | case 'M': |
3800 | if (GET_CODE (x) == CONST_INT) | |
3801 | fprintf (file, "%d",exact_log2(~INTVAL (x)) ); | |
3802 | else | |
3803 | output_operand_lossage ("invalid operand to %%M code"); | |
3804 | ||
3805 | return; | |
3806 | ||
ceaaa9fe JR |
3807 | case 'p': |
3808 | if (GET_CODE (x) == CONST_INT) | |
3809 | fprintf (file, "%d", exact_log2 (INTVAL (x) & -INTVAL (x))); | |
3810 | else | |
3811 | output_operand_lossage ("invalid operand to %%p code"); | |
3812 | return; | |
3813 | ||
3814 | case 's': | |
3815 | if (GET_CODE (x) == CONST_INT) | |
3816 | { | |
3817 | HOST_WIDE_INT i = INTVAL (x); | |
3818 | HOST_WIDE_INT s = exact_log2 (i & -i); | |
3819 | fprintf (file, "%d", exact_log2 (((0xffffffffUL & i) >> s) + 1)); | |
3820 | } | |
3821 | else | |
3822 | output_operand_lossage ("invalid operand to %%s code"); | |
3823 | return; | |
3824 | ||
526b7aee SV |
3825 | case '#' : |
3826 | /* Conditional branches depending on condition codes. | |
3827 | Note that this is only for branches that were known to depend on | |
3828 | condition codes before delay slot scheduling; | |
3829 | out-of-range brcc / bbit expansions should use '*'. | |
3830 | This distinction is important because of the different | |
3831 | allowable delay slot insns and the output of the delay suffix | |
3832 | for TARGET_AT_DBR_COND_EXEC. */ | |
3833 | case '*' : | |
3834 | /* Unconditional branches / branches not depending on condition codes. | |
3835 | This could also be a CALL_INSN. | |
3836 | Output the appropriate delay slot suffix. */ | |
84034c69 | 3837 | if (final_sequence && final_sequence->len () != 1) |
526b7aee | 3838 | { |
84034c69 DM |
3839 | rtx_insn *jump = final_sequence->insn (0); |
3840 | rtx_insn *delay = final_sequence->insn (1); | |
526b7aee SV |
3841 | |
3842 | /* For TARGET_PAD_RETURN we might have grabbed the delay insn. */ | |
4654c0cf | 3843 | if (delay->deleted ()) |
526b7aee SV |
3844 | return; |
3845 | if (JUMP_P (jump) && INSN_ANNULLED_BRANCH_P (jump)) | |
3846 | fputs (INSN_FROM_TARGET_P (delay) ? ".d" | |
3847 | : TARGET_AT_DBR_CONDEXEC && code == '#' ? ".d" | |
3848 | : get_attr_type (jump) == TYPE_RETURN && code == '#' ? "" | |
3849 | : ".nd", | |
3850 | file); | |
3851 | else | |
3852 | fputs (".d", file); | |
3853 | } | |
3854 | return; | |
3855 | case '?' : /* with leading "." */ | |
3856 | case '!' : /* without leading "." */ | |
3857 | /* This insn can be conditionally executed. See if the ccfsm machinery | |
3858 | says it should be conditionalized. | |
3859 | If it shouldn't, we'll check the compact attribute if this insn | |
3860 | has a short variant, which may be used depending on code size and | |
3861 | alignment considerations. */ | |
3862 | if (current_insn_predicate) | |
3863 | arc_ccfsm_current.cc | |
3864 | = get_arc_condition_code (current_insn_predicate); | |
3865 | if (ARC_CCFSM_COND_EXEC_P (&arc_ccfsm_current)) | |
3866 | { | |
3867 | /* Is this insn in a delay slot sequence? */ | |
3868 | if (!final_sequence || XVECLEN (final_sequence, 0) < 2 | |
3869 | || current_insn_predicate | |
68a1a6c0 DM |
3870 | || CALL_P (final_sequence->insn (0)) |
3871 | || simplejump_p (final_sequence->insn (0))) | |
526b7aee SV |
3872 | { |
3873 | /* This insn isn't in a delay slot sequence, or conditionalized | |
3874 | independently of its position in a delay slot. */ | |
3875 | fprintf (file, "%s%s", | |
3876 | code == '?' ? "." : "", | |
3877 | arc_condition_codes[arc_ccfsm_current.cc]); | |
3878 | /* If this is a jump, there are still short variants. However, | |
3879 | only beq_s / bne_s have the same offset range as b_s, | |
3880 | and the only short conditional returns are jeq_s and jne_s. */ | |
3881 | if (code == '!' | |
3882 | && (arc_ccfsm_current.cc == ARC_CC_EQ | |
3883 | || arc_ccfsm_current.cc == ARC_CC_NE | |
3884 | || 0 /* FIXME: check if branch in 7 bit range. */)) | |
3885 | output_short_suffix (file); | |
3886 | } | |
3887 | else if (code == '!') /* Jump with delay slot. */ | |
3888 | fputs (arc_condition_codes[arc_ccfsm_current.cc], file); | |
3889 | else /* An Instruction in a delay slot of a jump or call. */ | |
3890 | { | |
3891 | rtx jump = XVECEXP (final_sequence, 0, 0); | |
3892 | rtx insn = XVECEXP (final_sequence, 0, 1); | |
3893 | ||
3894 | /* If the insn is annulled and is from the target path, we need | |
3895 | to inverse the condition test. */ | |
3896 | if (JUMP_P (jump) && INSN_ANNULLED_BRANCH_P (jump)) | |
3897 | { | |
3898 | if (INSN_FROM_TARGET_P (insn)) | |
3899 | fprintf (file, "%s%s", | |
3900 | code == '?' ? "." : "", | |
3901 | arc_condition_codes[ARC_INVERSE_CONDITION_CODE (arc_ccfsm_current.cc)]); | |
3902 | else | |
3903 | fprintf (file, "%s%s", | |
3904 | code == '?' ? "." : "", | |
3905 | arc_condition_codes[arc_ccfsm_current.cc]); | |
3906 | if (arc_ccfsm_current.state == 5) | |
3907 | arc_ccfsm_current.state = 0; | |
3908 | } | |
3909 | else | |
3910 | /* This insn is executed for either path, so don't | |
3911 | conditionalize it at all. */ | |
3912 | output_short_suffix (file); | |
3913 | ||
3914 | } | |
3915 | } | |
3916 | else | |
3917 | output_short_suffix (file); | |
3918 | return; | |
3919 | case'`': | |
3920 | /* FIXME: fold constant inside unary operator, re-recognize, and emit. */ | |
3921 | gcc_unreachable (); | |
3922 | case 'd' : | |
3923 | fputs (arc_condition_codes[get_arc_condition_code (x)], file); | |
3924 | return; | |
3925 | case 'D' : | |
3926 | fputs (arc_condition_codes[ARC_INVERSE_CONDITION_CODE | |
3927 | (get_arc_condition_code (x))], | |
3928 | file); | |
3929 | return; | |
3930 | case 'R' : | |
3931 | /* Write second word of DImode or DFmode reference, | |
3932 | register or memory. */ | |
3933 | if (GET_CODE (x) == REG) | |
3934 | fputs (reg_names[REGNO (x)+1], file); | |
3935 | else if (GET_CODE (x) == MEM) | |
3936 | { | |
3937 | fputc ('[', file); | |
3938 | ||
3939 | /* Handle possible auto-increment. For PRE_INC / PRE_DEC / | |
3940 | PRE_MODIFY, we will have handled the first word already; | |
3941 | For POST_INC / POST_DEC / POST_MODIFY, the access to the | |
3942 | first word will be done later. In either case, the access | |
3943 | to the first word will do the modify, and we only have | |
3944 | to add an offset of four here. */ | |
3945 | if (GET_CODE (XEXP (x, 0)) == PRE_INC | |
3946 | || GET_CODE (XEXP (x, 0)) == PRE_DEC | |
3947 | || GET_CODE (XEXP (x, 0)) == PRE_MODIFY | |
3948 | || GET_CODE (XEXP (x, 0)) == POST_INC | |
3949 | || GET_CODE (XEXP (x, 0)) == POST_DEC | |
3950 | || GET_CODE (XEXP (x, 0)) == POST_MODIFY) | |
cc8ca59e JB |
3951 | output_address (VOIDmode, |
3952 | plus_constant (Pmode, XEXP (XEXP (x, 0), 0), 4)); | |
526b7aee SV |
3953 | else if (output_scaled) |
3954 | { | |
3955 | rtx addr = XEXP (x, 0); | |
3956 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
3957 | ||
cc8ca59e JB |
3958 | output_address (VOIDmode, |
3959 | plus_constant (Pmode, XEXP (addr, 0), | |
526b7aee SV |
3960 | ((INTVAL (XEXP (addr, 1)) + 4) |
3961 | >> (size == 2 ? 1 : 2)))); | |
3962 | output_scaled = 0; | |
3963 | } | |
3964 | else | |
cc8ca59e JB |
3965 | output_address (VOIDmode, |
3966 | plus_constant (Pmode, XEXP (x, 0), 4)); | |
526b7aee SV |
3967 | fputc (']', file); |
3968 | } | |
3969 | else | |
3970 | output_operand_lossage ("invalid operand to %%R code"); | |
3971 | return; | |
3972 | case 'S' : | |
3973 | /* FIXME: remove %S option. */ | |
3974 | break; | |
3975 | case 'B' /* Branch or other LIMM ref - must not use sda references. */ : | |
3976 | if (CONSTANT_P (x)) | |
3977 | { | |
3978 | output_addr_const (file, x); | |
3979 | return; | |
3980 | } | |
3981 | break; | |
3982 | case 'H' : | |
3983 | case 'L' : | |
3984 | if (GET_CODE (x) == REG) | |
3985 | { | |
3986 | /* L = least significant word, H = most significant word. */ | |
3987 | if ((WORDS_BIG_ENDIAN != 0) ^ (code == 'L')) | |
3988 | fputs (reg_names[REGNO (x)], file); | |
3989 | else | |
3990 | fputs (reg_names[REGNO (x)+1], file); | |
3991 | } | |
3992 | else if (GET_CODE (x) == CONST_INT | |
3993 | || GET_CODE (x) == CONST_DOUBLE) | |
3994 | { | |
8ad9df62 | 3995 | rtx first, second, word; |
526b7aee SV |
3996 | |
3997 | split_double (x, &first, &second); | |
3998 | ||
3999 | if((WORDS_BIG_ENDIAN) == 0) | |
8ad9df62 | 4000 | word = (code == 'L' ? first : second); |
526b7aee | 4001 | else |
8ad9df62 | 4002 | word = (code == 'L' ? second : first); |
526b7aee | 4003 | |
8ad9df62 JR |
4004 | fprintf (file, "0x%08" PRIx32, ((uint32_t) INTVAL (word))); |
4005 | } | |
526b7aee SV |
4006 | else |
4007 | output_operand_lossage ("invalid operand to %%H/%%L code"); | |
4008 | return; | |
4009 | case 'A' : | |
4010 | { | |
4011 | char str[30]; | |
4012 | ||
4013 | gcc_assert (GET_CODE (x) == CONST_DOUBLE | |
4014 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT); | |
4015 | ||
4016 | real_to_decimal (str, CONST_DOUBLE_REAL_VALUE (x), sizeof (str), 0, 1); | |
4017 | fprintf (file, "%s", str); | |
4018 | return; | |
4019 | } | |
4020 | case 'U' : | |
4021 | /* Output a load/store with update indicator if appropriate. */ | |
4022 | if (GET_CODE (x) == MEM) | |
4023 | { | |
4024 | rtx addr = XEXP (x, 0); | |
4025 | switch (GET_CODE (addr)) | |
4026 | { | |
4027 | case PRE_INC: case PRE_DEC: case PRE_MODIFY: | |
4028 | fputs (".a", file); break; | |
4029 | case POST_INC: case POST_DEC: case POST_MODIFY: | |
4030 | fputs (".ab", file); break; | |
4031 | case PLUS: | |
4032 | /* Are we using a scaled index? */ | |
4033 | if (GET_CODE (XEXP (addr, 0)) == MULT) | |
4034 | fputs (".as", file); | |
4035 | /* Can we use a scaled offset? */ | |
4036 | else if (CONST_INT_P (XEXP (addr, 1)) | |
4037 | && GET_MODE_SIZE (GET_MODE (x)) > 1 | |
4038 | && (!(INTVAL (XEXP (addr, 1)) | |
4039 | & (GET_MODE_SIZE (GET_MODE (x)) - 1) & 3)) | |
4040 | /* Does it make a difference? */ | |
4041 | && !SMALL_INT_RANGE(INTVAL (XEXP (addr, 1)), | |
4042 | GET_MODE_SIZE (GET_MODE (x)) - 2, 0)) | |
4043 | { | |
4044 | fputs (".as", file); | |
4045 | output_scaled = 1; | |
4046 | } | |
b6fb7933 CZ |
4047 | else if (LEGITIMATE_SMALL_DATA_ADDRESS_P (addr) |
4048 | && GET_MODE_SIZE (GET_MODE (x)) > 1) | |
4049 | { | |
4050 | tree decl = NULL_TREE; | |
4051 | int align = 0; | |
4052 | if (GET_CODE (XEXP (addr, 1)) == SYMBOL_REF) | |
4053 | decl = SYMBOL_REF_DECL (XEXP (addr, 1)); | |
4054 | else if (GET_CODE (XEXP (XEXP (XEXP (addr, 1), 0), 0)) | |
4055 | == SYMBOL_REF) | |
4056 | decl = SYMBOL_REF_DECL (XEXP (XEXP (XEXP (addr, 1), 0), 0)); | |
4057 | if (decl) | |
4058 | align = DECL_ALIGN (decl); | |
4059 | align = align / BITS_PER_UNIT; | |
4060 | if ((GET_MODE_SIZE (GET_MODE (x)) == 2) | |
4061 | && align && ((align & 1) == 0)) | |
4062 | fputs (".as", file); | |
4063 | if ((GET_MODE_SIZE (GET_MODE (x)) >= 4) | |
4064 | && align && ((align & 3) == 0)) | |
4065 | fputs (".as", file); | |
4066 | } | |
526b7aee SV |
4067 | break; |
4068 | case REG: | |
4069 | break; | |
4070 | default: | |
4071 | gcc_assert (CONSTANT_P (addr)); break; | |
4072 | } | |
4073 | } | |
4074 | else | |
4075 | output_operand_lossage ("invalid operand to %%U code"); | |
4076 | return; | |
4077 | case 'V' : | |
4078 | /* Output cache bypass indicator for a load/store insn. Volatile memory | |
4079 | refs are defined to use the cache bypass mechanism. */ | |
4080 | if (GET_CODE (x) == MEM) | |
4081 | { | |
4082 | if (MEM_VOLATILE_P (x) && !TARGET_VOLATILE_CACHE_SET ) | |
4083 | fputs (".di", file); | |
4084 | } | |
4085 | else | |
4086 | output_operand_lossage ("invalid operand to %%V code"); | |
4087 | return; | |
4088 | /* plt code. */ | |
4089 | case 'P': | |
4090 | case 0 : | |
4091 | /* Do nothing special. */ | |
4092 | break; | |
4093 | case 'F': | |
4094 | fputs (reg_names[REGNO (x)]+1, file); | |
4095 | return; | |
4096 | case '^': | |
4097 | /* This punctuation character is needed because label references are | |
4098 | printed in the output template using %l. This is a front end | |
4099 | character, and when we want to emit a '@' before it, we have to use | |
4100 | this '^'. */ | |
4101 | ||
4102 | fputc('@',file); | |
4103 | return; | |
4104 | case 'O': | |
4105 | /* Output an operator. */ | |
4106 | switch (GET_CODE (x)) | |
4107 | { | |
4108 | case PLUS: fputs ("add", file); return; | |
4109 | case SS_PLUS: fputs ("adds", file); return; | |
4110 | case AND: fputs ("and", file); return; | |
4111 | case IOR: fputs ("or", file); return; | |
4112 | case XOR: fputs ("xor", file); return; | |
4113 | case MINUS: fputs ("sub", file); return; | |
4114 | case SS_MINUS: fputs ("subs", file); return; | |
4115 | case ASHIFT: fputs ("asl", file); return; | |
4116 | case ASHIFTRT: fputs ("asr", file); return; | |
4117 | case LSHIFTRT: fputs ("lsr", file); return; | |
4118 | case ROTATERT: fputs ("ror", file); return; | |
4119 | case MULT: fputs ("mpy", file); return; | |
4120 | case ABS: fputs ("abs", file); return; /* Unconditional. */ | |
4121 | case NEG: fputs ("neg", file); return; | |
4122 | case SS_NEG: fputs ("negs", file); return; | |
4123 | case NOT: fputs ("not", file); return; /* Unconditional. */ | |
4124 | case ZERO_EXTEND: | |
4125 | fputs ("ext", file); /* bmsk allows predication. */ | |
4126 | goto size_suffix; | |
4127 | case SIGN_EXTEND: /* Unconditional. */ | |
4128 | fputs ("sex", file); | |
4129 | size_suffix: | |
4130 | switch (GET_MODE (XEXP (x, 0))) | |
4131 | { | |
4e10a5a7 RS |
4132 | case E_QImode: fputs ("b", file); return; |
4133 | case E_HImode: fputs ("w", file); return; | |
526b7aee SV |
4134 | default: break; |
4135 | } | |
4136 | break; | |
4137 | case SS_TRUNCATE: | |
4138 | if (GET_MODE (x) != HImode) | |
4139 | break; | |
4140 | fputs ("sat16", file); | |
4141 | default: break; | |
4142 | } | |
4143 | output_operand_lossage ("invalid operand to %%O code"); return; | |
4144 | case 'o': | |
4145 | if (GET_CODE (x) == SYMBOL_REF) | |
4146 | { | |
4147 | assemble_name (file, XSTR (x, 0)); | |
4148 | return; | |
4149 | } | |
4150 | break; | |
4151 | case '&': | |
4152 | if (TARGET_ANNOTATE_ALIGN && cfun->machine->size_reason) | |
4153 | fprintf (file, "; unalign: %d", cfun->machine->unalign); | |
4154 | return; | |
f50bb868 CZ |
4155 | case '+': |
4156 | if (TARGET_V2) | |
4157 | fputs ("m", file); | |
4158 | else | |
4159 | fputs ("h", file); | |
4160 | return; | |
4161 | case '_': | |
4162 | if (TARGET_V2) | |
4163 | fputs ("h", file); | |
4164 | else | |
4165 | fputs ("w", file); | |
4166 | return; | |
526b7aee SV |
4167 | default : |
4168 | /* Unknown flag. */ | |
4169 | output_operand_lossage ("invalid operand output code"); | |
4170 | } | |
4171 | ||
4172 | switch (GET_CODE (x)) | |
4173 | { | |
4174 | case REG : | |
4175 | fputs (reg_names[REGNO (x)], file); | |
4176 | break; | |
4177 | case MEM : | |
4178 | { | |
4179 | rtx addr = XEXP (x, 0); | |
4180 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
4181 | ||
4182 | fputc ('[', file); | |
4183 | ||
4184 | switch (GET_CODE (addr)) | |
4185 | { | |
4186 | case PRE_INC: case POST_INC: | |
cc8ca59e JB |
4187 | output_address (VOIDmode, |
4188 | plus_constant (Pmode, XEXP (addr, 0), size)); break; | |
526b7aee | 4189 | case PRE_DEC: case POST_DEC: |
cc8ca59e JB |
4190 | output_address (VOIDmode, |
4191 | plus_constant (Pmode, XEXP (addr, 0), -size)); | |
526b7aee SV |
4192 | break; |
4193 | case PRE_MODIFY: case POST_MODIFY: | |
cc8ca59e | 4194 | output_address (VOIDmode, XEXP (addr, 1)); break; |
526b7aee SV |
4195 | case PLUS: |
4196 | if (output_scaled) | |
4197 | { | |
cc8ca59e JB |
4198 | output_address (VOIDmode, |
4199 | plus_constant (Pmode, XEXP (addr, 0), | |
526b7aee SV |
4200 | (INTVAL (XEXP (addr, 1)) |
4201 | >> (size == 2 ? 1 : 2)))); | |
4202 | output_scaled = 0; | |
4203 | } | |
4204 | else | |
cc8ca59e | 4205 | output_address (VOIDmode, addr); |
526b7aee SV |
4206 | break; |
4207 | default: | |
4208 | if (flag_pic && CONSTANT_ADDRESS_P (addr)) | |
4209 | arc_output_pic_addr_const (file, addr, code); | |
4210 | else | |
cc8ca59e | 4211 | output_address (VOIDmode, addr); |
526b7aee SV |
4212 | break; |
4213 | } | |
4214 | fputc (']', file); | |
4215 | break; | |
4216 | } | |
4217 | case CONST_DOUBLE : | |
4218 | /* We handle SFmode constants here as output_addr_const doesn't. */ | |
4219 | if (GET_MODE (x) == SFmode) | |
4220 | { | |
526b7aee SV |
4221 | long l; |
4222 | ||
34a72c33 | 4223 | REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (x), l); |
526b7aee SV |
4224 | fprintf (file, "0x%08lx", l); |
4225 | break; | |
4226 | } | |
3bbe0b82 JL |
4227 | /* FALLTHRU */ |
4228 | /* Let output_addr_const deal with it. */ | |
526b7aee | 4229 | default : |
28633bbd CZ |
4230 | if (flag_pic |
4231 | || (GET_CODE (x) == CONST | |
4232 | && GET_CODE (XEXP (x, 0)) == UNSPEC | |
4233 | && (XINT (XEXP (x, 0), 1) == UNSPEC_TLS_OFF | |
4234 | || XINT (XEXP (x, 0), 1) == UNSPEC_TLS_GD)) | |
4235 | || (GET_CODE (x) == CONST | |
4236 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4237 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == UNSPEC | |
4238 | && (XINT (XEXP (XEXP (x, 0), 0), 1) == UNSPEC_TLS_OFF | |
4239 | || XINT (XEXP (XEXP (x, 0), 0), 1) == UNSPEC_TLS_GD))) | |
526b7aee SV |
4240 | arc_output_pic_addr_const (file, x, code); |
4241 | else | |
4242 | { | |
4243 | /* FIXME: Dirty way to handle @var@sda+const. Shd be handled | |
4244 | with asm_output_symbol_ref */ | |
4245 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS) | |
4246 | { | |
4247 | x = XEXP (x, 0); | |
4248 | output_addr_const (file, XEXP (x, 0)); | |
4249 | if (GET_CODE (XEXP (x, 0)) == SYMBOL_REF && SYMBOL_REF_SMALL_P (XEXP (x, 0))) | |
4250 | fprintf (file, "@sda"); | |
4251 | ||
4252 | if (GET_CODE (XEXP (x, 1)) != CONST_INT | |
4253 | || INTVAL (XEXP (x, 1)) >= 0) | |
4254 | fprintf (file, "+"); | |
4255 | output_addr_const (file, XEXP (x, 1)); | |
4256 | } | |
4257 | else | |
4258 | output_addr_const (file, x); | |
4259 | } | |
4260 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_SMALL_P (x)) | |
4261 | fprintf (file, "@sda"); | |
4262 | break; | |
4263 | } | |
4264 | } | |
4265 | ||
4266 | /* Print a memory address as an operand to reference that memory location. */ | |
4267 | ||
4268 | void | |
4269 | arc_print_operand_address (FILE *file , rtx addr) | |
4270 | { | |
4271 | register rtx base, index = 0; | |
4272 | ||
4273 | switch (GET_CODE (addr)) | |
4274 | { | |
4275 | case REG : | |
4276 | fputs (reg_names[REGNO (addr)], file); | |
4277 | break; | |
4278 | case SYMBOL_REF : | |
4279 | output_addr_const (file, addr); | |
4280 | if (SYMBOL_REF_SMALL_P (addr)) | |
4281 | fprintf (file, "@sda"); | |
4282 | break; | |
4283 | case PLUS : | |
4284 | if (GET_CODE (XEXP (addr, 0)) == MULT) | |
4285 | index = XEXP (XEXP (addr, 0), 0), base = XEXP (addr, 1); | |
4286 | else if (CONST_INT_P (XEXP (addr, 0))) | |
4287 | index = XEXP (addr, 0), base = XEXP (addr, 1); | |
4288 | else | |
4289 | base = XEXP (addr, 0), index = XEXP (addr, 1); | |
4290 | ||
4291 | gcc_assert (OBJECT_P (base)); | |
4292 | arc_print_operand_address (file, base); | |
4293 | if (CONSTANT_P (base) && CONST_INT_P (index)) | |
4294 | fputc ('+', file); | |
4295 | else | |
4296 | fputc (',', file); | |
4297 | gcc_assert (OBJECT_P (index)); | |
4298 | arc_print_operand_address (file, index); | |
4299 | break; | |
4300 | case CONST: | |
4301 | { | |
4302 | rtx c = XEXP (addr, 0); | |
4303 | ||
28633bbd CZ |
4304 | if ((GET_CODE (c) == UNSPEC |
4305 | && (XINT (c, 1) == UNSPEC_TLS_OFF | |
4306 | || XINT (c, 1) == UNSPEC_TLS_IE)) | |
4307 | || (GET_CODE (c) == PLUS | |
4308 | && GET_CODE (XEXP (c, 0)) == UNSPEC | |
f5e336b1 CZ |
4309 | && (XINT (XEXP (c, 0), 1) == UNSPEC_TLS_OFF |
4310 | || XINT (XEXP (c, 0), 1) == ARC_UNSPEC_GOTOFFPC))) | |
28633bbd CZ |
4311 | { |
4312 | arc_output_pic_addr_const (file, c, 0); | |
4313 | break; | |
4314 | } | |
4315 | gcc_assert (GET_CODE (c) == PLUS); | |
526b7aee SV |
4316 | gcc_assert (GET_CODE (XEXP (c, 0)) == SYMBOL_REF); |
4317 | gcc_assert (GET_CODE (XEXP (c, 1)) == CONST_INT); | |
4318 | ||
cc8ca59e | 4319 | output_address (VOIDmode, XEXP (addr, 0)); |
526b7aee SV |
4320 | |
4321 | break; | |
4322 | } | |
4323 | case PRE_INC : | |
4324 | case PRE_DEC : | |
4325 | /* We shouldn't get here as we've lost the mode of the memory object | |
4326 | (which says how much to inc/dec by. */ | |
4327 | gcc_unreachable (); | |
4328 | break; | |
4329 | default : | |
4330 | if (flag_pic) | |
4331 | arc_output_pic_addr_const (file, addr, 0); | |
4332 | else | |
4333 | output_addr_const (file, addr); | |
4334 | break; | |
4335 | } | |
4336 | } | |
4337 | ||
526b7aee SV |
4338 | /* Conditional execution support. |
4339 | ||
4340 | This is based on the ARM port but for now is much simpler. | |
4341 | ||
4342 | A finite state machine takes care of noticing whether or not instructions | |
4343 | can be conditionally executed, and thus decrease execution time and code | |
4344 | size by deleting branch instructions. The fsm is controlled by | |
4345 | arc_ccfsm_advance (called by arc_final_prescan_insn), and controls the | |
4346 | actions of PRINT_OPERAND. The patterns in the .md file for the branch | |
4347 | insns also have a hand in this. */ | |
4348 | /* The way we leave dealing with non-anulled or annull-false delay slot | |
4349 | insns to the consumer is awkward. */ | |
4350 | ||
4351 | /* The state of the fsm controlling condition codes are: | |
4352 | 0: normal, do nothing special | |
4353 | 1: don't output this insn | |
4354 | 2: don't output this insn | |
4355 | 3: make insns conditional | |
4356 | 4: make insns conditional | |
4357 | 5: make insn conditional (only for outputting anulled delay slot insns) | |
4358 | ||
4359 | special value for cfun->machine->uid_ccfsm_state: | |
4360 | 6: return with but one insn before it since function start / call | |
4361 | ||
4362 | State transitions (state->state by whom, under what condition): | |
4363 | 0 -> 1 arc_ccfsm_advance, if insn is a conditional branch skipping over | |
4364 | some instructions. | |
4365 | 0 -> 2 arc_ccfsm_advance, if insn is a conditional branch followed | |
4366 | by zero or more non-jump insns and an unconditional branch with | |
4367 | the same target label as the condbranch. | |
4368 | 1 -> 3 branch patterns, after having not output the conditional branch | |
4369 | 2 -> 4 branch patterns, after having not output the conditional branch | |
4370 | 0 -> 5 branch patterns, for anulled delay slot insn. | |
4371 | 3 -> 0 ASM_OUTPUT_INTERNAL_LABEL, if the `target' label is reached | |
4372 | (the target label has CODE_LABEL_NUMBER equal to | |
4373 | arc_ccfsm_target_label). | |
4374 | 4 -> 0 arc_ccfsm_advance, if `target' unconditional branch is reached | |
4375 | 3 -> 1 arc_ccfsm_advance, finding an 'else' jump skipping over some insns. | |
4376 | 5 -> 0 when outputting the delay slot insn | |
4377 | ||
4378 | If the jump clobbers the conditions then we use states 2 and 4. | |
4379 | ||
4380 | A similar thing can be done with conditional return insns. | |
4381 | ||
4382 | We also handle separating branches from sets of the condition code. | |
4383 | This is done here because knowledge of the ccfsm state is required, | |
4384 | we may not be outputting the branch. */ | |
4385 | ||
4386 | /* arc_final_prescan_insn calls arc_ccfsm_advance to adjust arc_ccfsm_current, | |
4387 | before letting final output INSN. */ | |
4388 | ||
4389 | static void | |
b3458f61 | 4390 | arc_ccfsm_advance (rtx_insn *insn, struct arc_ccfsm *state) |
526b7aee SV |
4391 | { |
4392 | /* BODY will hold the body of INSN. */ | |
4393 | register rtx body; | |
4394 | ||
4395 | /* This will be 1 if trying to repeat the trick (ie: do the `else' part of | |
4396 | an if/then/else), and things need to be reversed. */ | |
4397 | int reverse = 0; | |
4398 | ||
4399 | /* If we start with a return insn, we only succeed if we find another one. */ | |
4400 | int seeking_return = 0; | |
4401 | ||
4402 | /* START_INSN will hold the insn from where we start looking. This is the | |
4403 | first insn after the following code_label if REVERSE is true. */ | |
b3458f61 | 4404 | rtx_insn *start_insn = insn; |
526b7aee SV |
4405 | |
4406 | /* Type of the jump_insn. Brcc insns don't affect ccfsm changes, | |
4407 | since they don't rely on a cmp preceding the. */ | |
4408 | enum attr_type jump_insn_type; | |
4409 | ||
4410 | /* Allow -mdebug-ccfsm to turn this off so we can see how well it does. | |
4411 | We can't do this in macro FINAL_PRESCAN_INSN because its called from | |
4412 | final_scan_insn which has `optimize' as a local. */ | |
4413 | if (optimize < 2 || TARGET_NO_COND_EXEC) | |
4414 | return; | |
4415 | ||
4416 | /* Ignore notes and labels. */ | |
4417 | if (!INSN_P (insn)) | |
4418 | return; | |
4419 | body = PATTERN (insn); | |
4420 | /* If in state 4, check if the target branch is reached, in order to | |
4421 | change back to state 0. */ | |
4422 | if (state->state == 4) | |
4423 | { | |
4424 | if (insn == state->target_insn) | |
4425 | { | |
4426 | state->target_insn = NULL; | |
4427 | state->state = 0; | |
4428 | } | |
4429 | return; | |
4430 | } | |
4431 | ||
4432 | /* If in state 3, it is possible to repeat the trick, if this insn is an | |
4433 | unconditional branch to a label, and immediately following this branch | |
4434 | is the previous target label which is only used once, and the label this | |
4435 | branch jumps to is not too far off. Or in other words "we've done the | |
4436 | `then' part, see if we can do the `else' part." */ | |
4437 | if (state->state == 3) | |
4438 | { | |
4439 | if (simplejump_p (insn)) | |
4440 | { | |
4441 | start_insn = next_nonnote_insn (start_insn); | |
4442 | if (GET_CODE (start_insn) == BARRIER) | |
4443 | { | |
4444 | /* ??? Isn't this always a barrier? */ | |
4445 | start_insn = next_nonnote_insn (start_insn); | |
4446 | } | |
4447 | if (GET_CODE (start_insn) == CODE_LABEL | |
4448 | && CODE_LABEL_NUMBER (start_insn) == state->target_label | |
4449 | && LABEL_NUSES (start_insn) == 1) | |
4450 | reverse = TRUE; | |
4451 | else | |
4452 | return; | |
4453 | } | |
4454 | else if (GET_CODE (body) == SIMPLE_RETURN) | |
4455 | { | |
4456 | start_insn = next_nonnote_insn (start_insn); | |
4457 | if (GET_CODE (start_insn) == BARRIER) | |
4458 | start_insn = next_nonnote_insn (start_insn); | |
4459 | if (GET_CODE (start_insn) == CODE_LABEL | |
4460 | && CODE_LABEL_NUMBER (start_insn) == state->target_label | |
4461 | && LABEL_NUSES (start_insn) == 1) | |
4462 | { | |
4463 | reverse = TRUE; | |
4464 | seeking_return = 1; | |
4465 | } | |
4466 | else | |
4467 | return; | |
4468 | } | |
4469 | else | |
4470 | return; | |
4471 | } | |
4472 | ||
4473 | if (GET_CODE (insn) != JUMP_INSN | |
4474 | || GET_CODE (PATTERN (insn)) == ADDR_VEC | |
4475 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
4476 | return; | |
4477 | ||
4478 | /* We can't predicate BRCC or loop ends. | |
4479 | Also, when generating PIC code, and considering a medium range call, | |
4480 | we can't predicate the call. */ | |
4481 | jump_insn_type = get_attr_type (insn); | |
4482 | if (jump_insn_type == TYPE_BRCC | |
4483 | || jump_insn_type == TYPE_BRCC_NO_DELAY_SLOT | |
4484 | || jump_insn_type == TYPE_LOOP_END | |
4485 | || (jump_insn_type == TYPE_CALL && !get_attr_predicable (insn))) | |
4486 | return; | |
4487 | ||
4488 | /* This jump might be paralleled with a clobber of the condition codes, | |
4489 | the jump should always come first. */ | |
4490 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
4491 | body = XVECEXP (body, 0, 0); | |
4492 | ||
4493 | if (reverse | |
4494 | || (GET_CODE (body) == SET && GET_CODE (SET_DEST (body)) == PC | |
4495 | && GET_CODE (SET_SRC (body)) == IF_THEN_ELSE)) | |
4496 | { | |
4497 | int insns_skipped = 0, fail = FALSE, succeed = FALSE; | |
4498 | /* Flag which part of the IF_THEN_ELSE is the LABEL_REF. */ | |
4499 | int then_not_else = TRUE; | |
4500 | /* Nonzero if next insn must be the target label. */ | |
4501 | int next_must_be_target_label_p; | |
b3458f61 DM |
4502 | rtx_insn *this_insn = start_insn; |
4503 | rtx label = 0; | |
526b7aee SV |
4504 | |
4505 | /* Register the insn jumped to. */ | |
4506 | if (reverse) | |
4507 | { | |
4508 | if (!seeking_return) | |
4509 | label = XEXP (SET_SRC (body), 0); | |
4510 | } | |
4511 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == LABEL_REF) | |
4512 | label = XEXP (XEXP (SET_SRC (body), 1), 0); | |
4513 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == LABEL_REF) | |
4514 | { | |
4515 | label = XEXP (XEXP (SET_SRC (body), 2), 0); | |
4516 | then_not_else = FALSE; | |
4517 | } | |
4518 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == SIMPLE_RETURN) | |
4519 | seeking_return = 1; | |
4520 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == SIMPLE_RETURN) | |
4521 | { | |
4522 | seeking_return = 1; | |
4523 | then_not_else = FALSE; | |
4524 | } | |
4525 | else | |
4526 | gcc_unreachable (); | |
4527 | ||
4528 | /* If this is a non-annulled branch with a delay slot, there is | |
4529 | no need to conditionalize the delay slot. */ | |
782bdf21 | 4530 | if ((GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) == SEQUENCE) |
526b7aee SV |
4531 | && state->state == 0 && !INSN_ANNULLED_BRANCH_P (insn)) |
4532 | { | |
4533 | this_insn = NEXT_INSN (this_insn); | |
526b7aee SV |
4534 | } |
4535 | /* See how many insns this branch skips, and what kind of insns. If all | |
4536 | insns are okay, and the label or unconditional branch to the same | |
4537 | label is not too far away, succeed. */ | |
4538 | for (insns_skipped = 0, next_must_be_target_label_p = FALSE; | |
4539 | !fail && !succeed && insns_skipped < MAX_INSNS_SKIPPED; | |
4540 | insns_skipped++) | |
4541 | { | |
4542 | rtx scanbody; | |
4543 | ||
4544 | this_insn = next_nonnote_insn (this_insn); | |
4545 | if (!this_insn) | |
4546 | break; | |
4547 | ||
4548 | if (next_must_be_target_label_p) | |
4549 | { | |
4550 | if (GET_CODE (this_insn) == BARRIER) | |
4551 | continue; | |
4552 | if (GET_CODE (this_insn) == CODE_LABEL | |
4553 | && this_insn == label) | |
4554 | { | |
4555 | state->state = 1; | |
4556 | succeed = TRUE; | |
4557 | } | |
4558 | else | |
4559 | fail = TRUE; | |
4560 | break; | |
4561 | } | |
4562 | ||
526b7aee SV |
4563 | switch (GET_CODE (this_insn)) |
4564 | { | |
4565 | case CODE_LABEL: | |
4566 | /* Succeed if it is the target label, otherwise fail since | |
4567 | control falls in from somewhere else. */ | |
4568 | if (this_insn == label) | |
4569 | { | |
4570 | state->state = 1; | |
4571 | succeed = TRUE; | |
4572 | } | |
4573 | else | |
4574 | fail = TRUE; | |
4575 | break; | |
4576 | ||
4577 | case BARRIER: | |
4578 | /* Succeed if the following insn is the target label. | |
4579 | Otherwise fail. | |
4580 | If return insns are used then the last insn in a function | |
4581 | will be a barrier. */ | |
4582 | next_must_be_target_label_p = TRUE; | |
4583 | break; | |
4584 | ||
4585 | case CALL_INSN: | |
4586 | /* Can handle a call insn if there are no insns after it. | |
4587 | IE: The next "insn" is the target label. We don't have to | |
4588 | worry about delay slots as such insns are SEQUENCE's inside | |
4589 | INSN's. ??? It is possible to handle such insns though. */ | |
4590 | if (get_attr_cond (this_insn) == COND_CANUSE) | |
4591 | next_must_be_target_label_p = TRUE; | |
4592 | else | |
4593 | fail = TRUE; | |
4594 | break; | |
4595 | ||
4596 | case JUMP_INSN: | |
4173ddaf SB |
4597 | scanbody = PATTERN (this_insn); |
4598 | ||
526b7aee SV |
4599 | /* If this is an unconditional branch to the same label, succeed. |
4600 | If it is to another label, do nothing. If it is conditional, | |
4601 | fail. */ | |
4602 | /* ??? Probably, the test for the SET and the PC are | |
4603 | unnecessary. */ | |
4604 | ||
4605 | if (GET_CODE (scanbody) == SET | |
4606 | && GET_CODE (SET_DEST (scanbody)) == PC) | |
4607 | { | |
4608 | if (GET_CODE (SET_SRC (scanbody)) == LABEL_REF | |
4609 | && XEXP (SET_SRC (scanbody), 0) == label && !reverse) | |
4610 | { | |
4611 | state->state = 2; | |
4612 | succeed = TRUE; | |
4613 | } | |
4614 | else if (GET_CODE (SET_SRC (scanbody)) == IF_THEN_ELSE) | |
4615 | fail = TRUE; | |
4616 | else if (get_attr_cond (this_insn) != COND_CANUSE) | |
4617 | fail = TRUE; | |
4618 | } | |
4619 | else if (GET_CODE (scanbody) == SIMPLE_RETURN | |
4620 | && seeking_return) | |
4621 | { | |
4622 | state->state = 2; | |
4623 | succeed = TRUE; | |
4624 | } | |
4625 | else if (GET_CODE (scanbody) == PARALLEL) | |
4626 | { | |
4627 | if (get_attr_cond (this_insn) != COND_CANUSE) | |
4628 | fail = TRUE; | |
4629 | } | |
4630 | break; | |
4631 | ||
4632 | case INSN: | |
4173ddaf SB |
4633 | scanbody = PATTERN (this_insn); |
4634 | ||
526b7aee SV |
4635 | /* We can only do this with insns that can use the condition |
4636 | codes (and don't set them). */ | |
4637 | if (GET_CODE (scanbody) == SET | |
4638 | || GET_CODE (scanbody) == PARALLEL) | |
4639 | { | |
4640 | if (get_attr_cond (this_insn) != COND_CANUSE) | |
4641 | fail = TRUE; | |
4642 | } | |
4643 | /* We can't handle other insns like sequences. */ | |
4644 | else | |
4645 | fail = TRUE; | |
4646 | break; | |
4647 | ||
4648 | default: | |
4649 | break; | |
4650 | } | |
4651 | } | |
4652 | ||
4653 | if (succeed) | |
4654 | { | |
4655 | if ((!seeking_return) && (state->state == 1 || reverse)) | |
4656 | state->target_label = CODE_LABEL_NUMBER (label); | |
4657 | else if (seeking_return || state->state == 2) | |
4658 | { | |
4659 | while (this_insn && GET_CODE (PATTERN (this_insn)) == USE) | |
4660 | { | |
4661 | this_insn = next_nonnote_insn (this_insn); | |
4662 | ||
4663 | gcc_assert (!this_insn || | |
4664 | (GET_CODE (this_insn) != BARRIER | |
4665 | && GET_CODE (this_insn) != CODE_LABEL)); | |
4666 | } | |
4667 | if (!this_insn) | |
4668 | { | |
4669 | /* Oh dear! we ran off the end, give up. */ | |
4670 | extract_insn_cached (insn); | |
4671 | state->state = 0; | |
4672 | state->target_insn = NULL; | |
4673 | return; | |
4674 | } | |
4675 | state->target_insn = this_insn; | |
4676 | } | |
4677 | else | |
4678 | gcc_unreachable (); | |
4679 | ||
4680 | /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from | |
4681 | what it was. */ | |
4682 | if (!reverse) | |
4683 | { | |
4684 | state->cond = XEXP (SET_SRC (body), 0); | |
4685 | state->cc = get_arc_condition_code (XEXP (SET_SRC (body), 0)); | |
4686 | } | |
4687 | ||
4688 | if (reverse || then_not_else) | |
4689 | state->cc = ARC_INVERSE_CONDITION_CODE (state->cc); | |
4690 | } | |
4691 | ||
4692 | /* Restore recog_operand. Getting the attributes of other insns can | |
4693 | destroy this array, but final.c assumes that it remains intact | |
4694 | across this call; since the insn has been recognized already we | |
4695 | call insn_extract direct. */ | |
4696 | extract_insn_cached (insn); | |
4697 | } | |
4698 | } | |
4699 | ||
4700 | /* Record that we are currently outputting label NUM with prefix PREFIX. | |
4701 | It it's the label we're looking for, reset the ccfsm machinery. | |
4702 | ||
4703 | Called from ASM_OUTPUT_INTERNAL_LABEL. */ | |
4704 | ||
4705 | static void | |
4706 | arc_ccfsm_at_label (const char *prefix, int num, struct arc_ccfsm *state) | |
4707 | { | |
4708 | if (state->state == 3 && state->target_label == num | |
4709 | && !strcmp (prefix, "L")) | |
4710 | { | |
4711 | state->state = 0; | |
b3458f61 | 4712 | state->target_insn = NULL; |
526b7aee SV |
4713 | } |
4714 | } | |
4715 | ||
4716 | /* We are considering a conditional branch with the condition COND. | |
4717 | Check if we want to conditionalize a delay slot insn, and if so modify | |
4718 | the ccfsm state accordingly. | |
4719 | REVERSE says branch will branch when the condition is false. */ | |
4720 | void | |
b32d5189 | 4721 | arc_ccfsm_record_condition (rtx cond, bool reverse, rtx_insn *jump, |
526b7aee SV |
4722 | struct arc_ccfsm *state) |
4723 | { | |
b3458f61 | 4724 | rtx_insn *seq_insn = NEXT_INSN (PREV_INSN (jump)); |
526b7aee SV |
4725 | if (!state) |
4726 | state = &arc_ccfsm_current; | |
4727 | ||
4728 | gcc_assert (state->state == 0); | |
4729 | if (seq_insn != jump) | |
4730 | { | |
4731 | rtx insn = XVECEXP (PATTERN (seq_insn), 0, 1); | |
4732 | ||
4654c0cf | 4733 | if (!as_a<rtx_insn *> (insn)->deleted () |
526b7aee SV |
4734 | && INSN_ANNULLED_BRANCH_P (jump) |
4735 | && (TARGET_AT_DBR_CONDEXEC || INSN_FROM_TARGET_P (insn))) | |
4736 | { | |
4737 | state->cond = cond; | |
4738 | state->cc = get_arc_condition_code (cond); | |
4739 | if (!reverse) | |
4740 | arc_ccfsm_current.cc | |
4741 | = ARC_INVERSE_CONDITION_CODE (state->cc); | |
4742 | rtx pat = PATTERN (insn); | |
4743 | if (GET_CODE (pat) == COND_EXEC) | |
4744 | gcc_assert ((INSN_FROM_TARGET_P (insn) | |
4745 | ? ARC_INVERSE_CONDITION_CODE (state->cc) : state->cc) | |
4746 | == get_arc_condition_code (XEXP (pat, 0))); | |
4747 | else | |
4748 | state->state = 5; | |
4749 | } | |
4750 | } | |
4751 | } | |
4752 | ||
4753 | /* Update *STATE as we would when we emit INSN. */ | |
4754 | ||
4755 | static void | |
b3458f61 | 4756 | arc_ccfsm_post_advance (rtx_insn *insn, struct arc_ccfsm *state) |
526b7aee | 4757 | { |
53ea364f JR |
4758 | enum attr_type type; |
4759 | ||
526b7aee SV |
4760 | if (LABEL_P (insn)) |
4761 | arc_ccfsm_at_label ("L", CODE_LABEL_NUMBER (insn), state); | |
4762 | else if (JUMP_P (insn) | |
4763 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
4764 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
53ea364f | 4765 | && ((type = get_attr_type (insn)) == TYPE_BRANCH |
6c28e6ae CZ |
4766 | || ((type == TYPE_UNCOND_BRANCH |
4767 | || type == TYPE_RETURN) | |
53ea364f | 4768 | && ARC_CCFSM_BRANCH_DELETED_P (state)))) |
526b7aee SV |
4769 | { |
4770 | if (ARC_CCFSM_BRANCH_DELETED_P (state)) | |
4771 | ARC_CCFSM_RECORD_BRANCH_DELETED (state); | |
4772 | else | |
4773 | { | |
4774 | rtx src = SET_SRC (PATTERN (insn)); | |
4775 | arc_ccfsm_record_condition (XEXP (src, 0), XEXP (src, 1) == pc_rtx, | |
4776 | insn, state); | |
4777 | } | |
4778 | } | |
4779 | else if (arc_ccfsm_current.state == 5) | |
4780 | arc_ccfsm_current.state = 0; | |
4781 | } | |
4782 | ||
4783 | /* Return true if the current insn, which is a conditional branch, is to be | |
4784 | deleted. */ | |
4785 | ||
4786 | bool | |
4787 | arc_ccfsm_branch_deleted_p (void) | |
4788 | { | |
4789 | return ARC_CCFSM_BRANCH_DELETED_P (&arc_ccfsm_current); | |
4790 | } | |
4791 | ||
4792 | /* Record a branch isn't output because subsequent insns can be | |
4793 | conditionalized. */ | |
4794 | ||
4795 | void | |
4796 | arc_ccfsm_record_branch_deleted (void) | |
4797 | { | |
4798 | ARC_CCFSM_RECORD_BRANCH_DELETED (&arc_ccfsm_current); | |
4799 | } | |
4800 | ||
4801 | /* During insn output, indicate if the current insn is predicated. */ | |
4802 | ||
4803 | bool | |
4804 | arc_ccfsm_cond_exec_p (void) | |
4805 | { | |
4806 | return (cfun->machine->prescan_initialized | |
4807 | && ARC_CCFSM_COND_EXEC_P (&arc_ccfsm_current)); | |
4808 | } | |
4809 | ||
4810 | /* Like next_active_insn, but return NULL if we find an ADDR_(DIFF_)VEC, | |
4811 | and look inside SEQUENCEs. */ | |
4812 | ||
b3458f61 DM |
4813 | static rtx_insn * |
4814 | arc_next_active_insn (rtx_insn *insn, struct arc_ccfsm *statep) | |
526b7aee SV |
4815 | { |
4816 | rtx pat; | |
4817 | ||
4818 | do | |
4819 | { | |
4820 | if (statep) | |
4821 | arc_ccfsm_post_advance (insn, statep); | |
4822 | insn = NEXT_INSN (insn); | |
4823 | if (!insn || BARRIER_P (insn)) | |
b3458f61 | 4824 | return NULL; |
526b7aee SV |
4825 | if (statep) |
4826 | arc_ccfsm_advance (insn, statep); | |
4827 | } | |
4828 | while (NOTE_P (insn) | |
4829 | || (cfun->machine->arc_reorg_started | |
4830 | && LABEL_P (insn) && !label_to_alignment (insn)) | |
4831 | || (NONJUMP_INSN_P (insn) | |
4832 | && (GET_CODE (PATTERN (insn)) == USE | |
4833 | || GET_CODE (PATTERN (insn)) == CLOBBER))); | |
4834 | if (!LABEL_P (insn)) | |
4835 | { | |
4836 | gcc_assert (INSN_P (insn)); | |
4837 | pat = PATTERN (insn); | |
4838 | if (GET_CODE (pat) == ADDR_VEC || GET_CODE (pat) == ADDR_DIFF_VEC) | |
b3458f61 | 4839 | return NULL; |
526b7aee | 4840 | if (GET_CODE (pat) == SEQUENCE) |
b3458f61 | 4841 | return as_a <rtx_insn *> (XVECEXP (pat, 0, 0)); |
526b7aee SV |
4842 | } |
4843 | return insn; | |
4844 | } | |
4845 | ||
4846 | /* When deciding if an insn should be output short, we want to know something | |
4847 | about the following insns: | |
4848 | - if another insn follows which we know we can output as a short insn | |
4849 | before an alignment-sensitive point, we can output this insn short: | |
4850 | the decision about the eventual alignment can be postponed. | |
4851 | - if a to-be-aligned label comes next, we should output this insn such | |
4852 | as to get / preserve 4-byte alignment. | |
4853 | - if a likely branch without delay slot insn, or a call with an immediately | |
4854 | following short insn comes next, we should out output this insn such as to | |
4855 | get / preserve 2 mod 4 unalignment. | |
4856 | - do the same for a not completely unlikely branch with a short insn | |
4857 | following before any other branch / label. | |
4858 | - in order to decide if we are actually looking at a branch, we need to | |
4859 | call arc_ccfsm_advance. | |
4860 | - in order to decide if we are looking at a short insn, we should know | |
4861 | if it is conditionalized. To a first order of approximation this is | |
4862 | the case if the state from arc_ccfsm_advance from before this insn | |
4863 | indicates the insn is conditionalized. However, a further refinement | |
4864 | could be to not conditionalize an insn if the destination register(s) | |
4865 | is/are dead in the non-executed case. */ | |
4866 | /* Return non-zero if INSN should be output as a short insn. UNALIGN is | |
4867 | zero if the current insn is aligned to a 4-byte-boundary, two otherwise. | |
4868 | If CHECK_ATTR is greater than 0, check the iscompact attribute first. */ | |
4869 | ||
4870 | int | |
b3458f61 | 4871 | arc_verify_short (rtx_insn *insn, int, int check_attr) |
526b7aee SV |
4872 | { |
4873 | enum attr_iscompact iscompact; | |
4874 | struct machine_function *machine; | |
4875 | ||
4876 | if (check_attr > 0) | |
4877 | { | |
4878 | iscompact = get_attr_iscompact (insn); | |
4879 | if (iscompact == ISCOMPACT_FALSE) | |
4880 | return 0; | |
4881 | } | |
4882 | machine = cfun->machine; | |
4883 | ||
4884 | if (machine->force_short_suffix >= 0) | |
4885 | return machine->force_short_suffix; | |
4886 | ||
4887 | return (get_attr_length (insn) & 2) != 0; | |
4888 | } | |
4889 | ||
4890 | /* When outputting an instruction (alternative) that can potentially be short, | |
4891 | output the short suffix if the insn is in fact short, and update | |
4892 | cfun->machine->unalign accordingly. */ | |
4893 | ||
4894 | static void | |
4895 | output_short_suffix (FILE *file) | |
4896 | { | |
b3458f61 | 4897 | rtx_insn *insn = current_output_insn; |
526b7aee SV |
4898 | |
4899 | if (arc_verify_short (insn, cfun->machine->unalign, 1)) | |
4900 | { | |
4901 | fprintf (file, "_s"); | |
4902 | cfun->machine->unalign ^= 2; | |
4903 | } | |
4904 | /* Restore recog_operand. */ | |
4905 | extract_insn_cached (insn); | |
4906 | } | |
4907 | ||
4908 | /* Implement FINAL_PRESCAN_INSN. */ | |
4909 | ||
4910 | void | |
b3458f61 | 4911 | arc_final_prescan_insn (rtx_insn *insn, rtx *opvec ATTRIBUTE_UNUSED, |
526b7aee SV |
4912 | int noperands ATTRIBUTE_UNUSED) |
4913 | { | |
4914 | if (TARGET_DUMPISIZE) | |
4915 | fprintf (asm_out_file, "\n; at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); | |
4916 | ||
4917 | /* Output a nop if necessary to prevent a hazard. | |
4918 | Don't do this for delay slots: inserting a nop would | |
4919 | alter semantics, and the only time we would find a hazard is for a | |
4920 | call function result - and in that case, the hazard is spurious to | |
4921 | start with. */ | |
4922 | if (PREV_INSN (insn) | |
4923 | && PREV_INSN (NEXT_INSN (insn)) == insn | |
4924 | && arc_hazard (prev_real_insn (insn), insn)) | |
4925 | { | |
4926 | current_output_insn = | |
4927 | emit_insn_before (gen_nop (), NEXT_INSN (PREV_INSN (insn))); | |
4928 | final_scan_insn (current_output_insn, asm_out_file, optimize, 1, NULL); | |
b3458f61 | 4929 | current_output_insn = insn; |
526b7aee SV |
4930 | } |
4931 | /* Restore extraction data which might have been clobbered by arc_hazard. */ | |
4932 | extract_constrain_insn_cached (insn); | |
4933 | ||
4934 | if (!cfun->machine->prescan_initialized) | |
4935 | { | |
4936 | /* Clear lingering state from branch shortening. */ | |
4937 | memset (&arc_ccfsm_current, 0, sizeof arc_ccfsm_current); | |
4938 | cfun->machine->prescan_initialized = 1; | |
4939 | } | |
4940 | arc_ccfsm_advance (insn, &arc_ccfsm_current); | |
4941 | ||
4942 | cfun->machine->size_reason = 0; | |
4943 | } | |
4944 | ||
4945 | /* Given FROM and TO register numbers, say whether this elimination is allowed. | |
4946 | Frame pointer elimination is automatically handled. | |
4947 | ||
4948 | All eliminations are permissible. If we need a frame | |
4949 | pointer, we must eliminate ARG_POINTER_REGNUM into | |
4950 | FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
4951 | ||
4952 | static bool | |
4953 | arc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to) | |
4954 | { | |
19dc4752 | 4955 | return ((to == FRAME_POINTER_REGNUM) || !arc_frame_pointer_needed ()); |
526b7aee SV |
4956 | } |
4957 | ||
4958 | /* Define the offset between two registers, one to be eliminated, and | |
4959 | the other its replacement, at the start of a routine. */ | |
4960 | ||
4961 | int | |
4962 | arc_initial_elimination_offset (int from, int to) | |
4963 | { | |
6fe5e235 CZ |
4964 | if (!cfun->machine->frame_info.initialized) |
4965 | arc_compute_frame_size (); | |
526b7aee SV |
4966 | |
4967 | if (from == ARG_POINTER_REGNUM && to == FRAME_POINTER_REGNUM) | |
4968 | { | |
4969 | return (cfun->machine->frame_info.extra_size | |
4970 | + cfun->machine->frame_info.reg_size); | |
4971 | } | |
4972 | ||
4973 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
4974 | { | |
4975 | return (cfun->machine->frame_info.total_size | |
4976 | - cfun->machine->frame_info.pretend_size); | |
4977 | } | |
4978 | ||
4979 | if ((from == FRAME_POINTER_REGNUM) && (to == STACK_POINTER_REGNUM)) | |
4980 | { | |
4981 | return (cfun->machine->frame_info.total_size | |
4982 | - (cfun->machine->frame_info.pretend_size | |
4983 | + cfun->machine->frame_info.extra_size | |
4984 | + cfun->machine->frame_info.reg_size)); | |
4985 | } | |
4986 | ||
4987 | gcc_unreachable (); | |
4988 | } | |
4989 | ||
4990 | static bool | |
4991 | arc_frame_pointer_required (void) | |
4992 | { | |
6fe5e235 | 4993 | return cfun->calls_alloca || crtl->calls_eh_return; |
526b7aee SV |
4994 | } |
4995 | ||
4996 | ||
4997 | /* Return the destination address of a branch. */ | |
4998 | ||
4999 | int | |
5000 | branch_dest (rtx branch) | |
5001 | { | |
5002 | rtx pat = PATTERN (branch); | |
5003 | rtx dest = (GET_CODE (pat) == PARALLEL | |
5004 | ? SET_SRC (XVECEXP (pat, 0, 0)) : SET_SRC (pat)); | |
5005 | int dest_uid; | |
5006 | ||
5007 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
5008 | dest = XEXP (dest, XEXP (dest, 1) == pc_rtx ? 2 : 1); | |
5009 | ||
5010 | dest = XEXP (dest, 0); | |
5011 | dest_uid = INSN_UID (dest); | |
5012 | ||
5013 | return INSN_ADDRESSES (dest_uid); | |
5014 | } | |
5015 | ||
5016 | ||
5719867d | 5017 | /* Implement TARGET_ENCODE_SECTION_INFO hook. */ |
526b7aee SV |
5018 | |
5019 | static void | |
5020 | arc_encode_section_info (tree decl, rtx rtl, int first) | |
5021 | { | |
5022 | /* For sdata, SYMBOL_FLAG_LOCAL and SYMBOL_FLAG_FUNCTION. | |
5023 | This clears machine specific flags, so has to come first. */ | |
5024 | default_encode_section_info (decl, rtl, first); | |
5025 | ||
5026 | /* Check if it is a function, and whether it has the | |
5027 | [long/medium/short]_call attribute specified. */ | |
5028 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
5029 | { | |
5030 | rtx symbol = XEXP (rtl, 0); | |
5031 | int flags = SYMBOL_REF_FLAGS (symbol); | |
5032 | ||
5033 | tree attr = (TREE_TYPE (decl) != error_mark_node | |
5034 | ? TYPE_ATTRIBUTES (TREE_TYPE (decl)) : NULL_TREE); | |
5035 | tree long_call_attr = lookup_attribute ("long_call", attr); | |
5036 | tree medium_call_attr = lookup_attribute ("medium_call", attr); | |
5037 | tree short_call_attr = lookup_attribute ("short_call", attr); | |
5038 | ||
5039 | if (long_call_attr != NULL_TREE) | |
5040 | flags |= SYMBOL_FLAG_LONG_CALL; | |
5041 | else if (medium_call_attr != NULL_TREE) | |
5042 | flags |= SYMBOL_FLAG_MEDIUM_CALL; | |
5043 | else if (short_call_attr != NULL_TREE) | |
5044 | flags |= SYMBOL_FLAG_SHORT_CALL; | |
5045 | ||
5046 | SYMBOL_REF_FLAGS (symbol) = flags; | |
5047 | } | |
4d03dc2f JR |
5048 | else if (TREE_CODE (decl) == VAR_DECL) |
5049 | { | |
5050 | rtx symbol = XEXP (rtl, 0); | |
5051 | ||
5052 | tree attr = (TREE_TYPE (decl) != error_mark_node | |
5053 | ? DECL_ATTRIBUTES (decl) : NULL_TREE); | |
5054 | ||
5055 | tree sec_attr = lookup_attribute ("section", attr); | |
5056 | if (sec_attr) | |
5057 | { | |
5058 | const char *sec_name | |
5059 | = TREE_STRING_POINTER (TREE_VALUE (TREE_VALUE (sec_attr))); | |
5060 | if (strcmp (sec_name, ".cmem") == 0 | |
5061 | || strcmp (sec_name, ".cmem_shared") == 0 | |
5062 | || strcmp (sec_name, ".cmem_private") == 0) | |
5063 | SYMBOL_REF_FLAGS (symbol) |= SYMBOL_FLAG_CMEM; | |
5064 | } | |
5065 | } | |
526b7aee SV |
5066 | } |
5067 | ||
5068 | /* This is how to output a definition of an internal numbered label where | |
5069 | PREFIX is the class of label and NUM is the number within the class. */ | |
5070 | ||
5071 | static void arc_internal_label (FILE *stream, const char *prefix, unsigned long labelno) | |
5072 | { | |
5073 | if (cfun) | |
5074 | arc_ccfsm_at_label (prefix, labelno, &arc_ccfsm_current); | |
5075 | default_internal_label (stream, prefix, labelno); | |
5076 | } | |
5077 | ||
5078 | /* Set the cpu type and print out other fancy things, | |
5079 | at the top of the file. */ | |
5080 | ||
5081 | static void arc_file_start (void) | |
5082 | { | |
5083 | default_file_start (); | |
5084 | fprintf (asm_out_file, "\t.cpu %s\n", arc_cpu_string); | |
5085 | } | |
5086 | ||
5087 | /* Cost functions. */ | |
5088 | ||
5089 | /* Compute a (partial) cost for rtx X. Return true if the complete | |
5090 | cost has been computed, and false if subexpressions should be | |
5091 | scanned. In either case, *TOTAL contains the cost result. */ | |
5092 | ||
5093 | static bool | |
e548c9df AM |
5094 | arc_rtx_costs (rtx x, machine_mode mode, int outer_code, |
5095 | int opno ATTRIBUTE_UNUSED, int *total, bool speed) | |
526b7aee | 5096 | { |
e548c9df AM |
5097 | int code = GET_CODE (x); |
5098 | ||
526b7aee SV |
5099 | switch (code) |
5100 | { | |
5101 | /* Small integers are as cheap as registers. */ | |
5102 | case CONST_INT: | |
5103 | { | |
5104 | bool nolimm = false; /* Can we do without long immediate? */ | |
5105 | bool fast = false; /* Is the result available immediately? */ | |
5106 | bool condexec = false; /* Does this allow conditiobnal execution? */ | |
5107 | bool compact = false; /* Is a 16 bit opcode available? */ | |
5108 | /* CONDEXEC also implies that we can have an unconditional | |
5109 | 3-address operation. */ | |
5110 | ||
5111 | nolimm = compact = condexec = false; | |
5112 | if (UNSIGNED_INT6 (INTVAL (x))) | |
5113 | nolimm = condexec = compact = true; | |
5114 | else | |
5115 | { | |
5116 | if (SMALL_INT (INTVAL (x))) | |
5117 | nolimm = fast = true; | |
5118 | switch (outer_code) | |
5119 | { | |
5120 | case AND: /* bclr, bmsk, ext[bw] */ | |
5121 | if (satisfies_constraint_Ccp (x) /* bclr */ | |
5122 | || satisfies_constraint_C1p (x) /* bmsk */) | |
5123 | nolimm = fast = condexec = compact = true; | |
5124 | break; | |
5125 | case IOR: /* bset */ | |
5126 | if (satisfies_constraint_C0p (x)) /* bset */ | |
5127 | nolimm = fast = condexec = compact = true; | |
5128 | break; | |
5129 | case XOR: | |
5130 | if (satisfies_constraint_C0p (x)) /* bxor */ | |
5131 | nolimm = fast = condexec = true; | |
5132 | break; | |
5133 | case SET: | |
5134 | if (satisfies_constraint_Crr (x)) /* ror b,u6 */ | |
5135 | nolimm = true; | |
5136 | default: | |
5137 | break; | |
5138 | } | |
5139 | } | |
5140 | /* FIXME: Add target options to attach a small cost if | |
5141 | condexec / compact is not true. */ | |
5142 | if (nolimm) | |
5143 | { | |
5144 | *total = 0; | |
5145 | return true; | |
5146 | } | |
5147 | } | |
5148 | /* FALLTHRU */ | |
5149 | ||
5150 | /* 4 byte values can be fetched as immediate constants - | |
5151 | let's give that the cost of an extra insn. */ | |
5152 | case CONST: | |
5153 | case LABEL_REF: | |
5154 | case SYMBOL_REF: | |
5155 | *total = COSTS_N_INSNS (1); | |
5156 | return true; | |
5157 | ||
5158 | case CONST_DOUBLE: | |
5159 | { | |
7d81a567 | 5160 | rtx first, second; |
526b7aee SV |
5161 | |
5162 | if (TARGET_DPFP) | |
5163 | { | |
5164 | *total = COSTS_N_INSNS (1); | |
5165 | return true; | |
5166 | } | |
7d81a567 CZ |
5167 | split_double (x, &first, &second); |
5168 | *total = COSTS_N_INSNS (!SMALL_INT (INTVAL (first)) | |
5169 | + !SMALL_INT (INTVAL (second))); | |
526b7aee SV |
5170 | return true; |
5171 | } | |
5172 | ||
5173 | /* Encourage synth_mult to find a synthetic multiply when reasonable. | |
5174 | If we need more than 12 insns to do a multiply, then go out-of-line, | |
5175 | since the call overhead will be < 10% of the cost of the multiply. */ | |
5176 | case ASHIFT: | |
5177 | case ASHIFTRT: | |
5178 | case LSHIFTRT: | |
5179 | if (TARGET_BARREL_SHIFTER) | |
5180 | { | |
5181 | /* If we want to shift a constant, we need a LIMM. */ | |
5182 | /* ??? when the optimizers want to know if a constant should be | |
5183 | hoisted, they ask for the cost of the constant. OUTER_CODE is | |
5184 | insufficient context for shifts since we don't know which operand | |
5185 | we are looking at. */ | |
5186 | if (CONSTANT_P (XEXP (x, 0))) | |
5187 | { | |
5188 | *total += (COSTS_N_INSNS (2) | |
e548c9df AM |
5189 | + rtx_cost (XEXP (x, 1), mode, (enum rtx_code) code, |
5190 | 0, speed)); | |
526b7aee SV |
5191 | return true; |
5192 | } | |
5193 | *total = COSTS_N_INSNS (1); | |
5194 | } | |
5195 | else if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5196 | *total = COSTS_N_INSNS (16); | |
5197 | else | |
5198 | { | |
5199 | *total = COSTS_N_INSNS (INTVAL (XEXP ((x), 1))); | |
5200 | /* ??? want_to_gcse_p can throw negative shift counts at us, | |
5201 | and then panics when it gets a negative cost as result. | |
5202 | Seen for gcc.c-torture/compile/20020710-1.c -Os . */ | |
5203 | if (*total < 0) | |
5204 | *total = 0; | |
5205 | } | |
5206 | return false; | |
5207 | ||
5208 | case DIV: | |
5209 | case UDIV: | |
5210 | if (speed) | |
5211 | *total = COSTS_N_INSNS(30); | |
5212 | else | |
5213 | *total = COSTS_N_INSNS(1); | |
5214 | return false; | |
5215 | ||
5216 | case MULT: | |
5217 | if ((TARGET_DPFP && GET_MODE (x) == DFmode)) | |
5218 | *total = COSTS_N_INSNS (1); | |
5219 | else if (speed) | |
5220 | *total= arc_multcost; | |
5221 | /* We do not want synth_mult sequences when optimizing | |
5222 | for size. */ | |
f50bb868 | 5223 | else if (TARGET_MUL64_SET || TARGET_ARC700_MPY) |
526b7aee SV |
5224 | *total = COSTS_N_INSNS (1); |
5225 | else | |
5226 | *total = COSTS_N_INSNS (2); | |
5227 | return false; | |
5228 | case PLUS: | |
1e466f04 GM |
5229 | if ((GET_CODE (XEXP (x, 0)) == ASHIFT |
5230 | && _1_2_3_operand (XEXP (XEXP (x, 0), 1), VOIDmode)) | |
5231 | || (GET_CODE (XEXP (x, 0)) == MULT | |
5232 | && _2_4_8_operand (XEXP (XEXP (x, 0), 1), VOIDmode))) | |
526b7aee | 5233 | { |
e548c9df AM |
5234 | *total += (rtx_cost (XEXP (x, 1), mode, PLUS, 0, speed) |
5235 | + rtx_cost (XEXP (XEXP (x, 0), 0), mode, PLUS, 1, speed)); | |
526b7aee SV |
5236 | return true; |
5237 | } | |
5238 | return false; | |
5239 | case MINUS: | |
1e466f04 GM |
5240 | if ((GET_CODE (XEXP (x, 1)) == ASHIFT |
5241 | && _1_2_3_operand (XEXP (XEXP (x, 1), 1), VOIDmode)) | |
5242 | || (GET_CODE (XEXP (x, 1)) == MULT | |
5243 | && _2_4_8_operand (XEXP (XEXP (x, 1), 1), VOIDmode))) | |
526b7aee | 5244 | { |
e548c9df AM |
5245 | *total += (rtx_cost (XEXP (x, 0), mode, PLUS, 0, speed) |
5246 | + rtx_cost (XEXP (XEXP (x, 1), 0), mode, PLUS, 1, speed)); | |
526b7aee SV |
5247 | return true; |
5248 | } | |
5249 | return false; | |
5250 | case COMPARE: | |
5251 | { | |
5252 | rtx op0 = XEXP (x, 0); | |
5253 | rtx op1 = XEXP (x, 1); | |
5254 | ||
5255 | if (GET_CODE (op0) == ZERO_EXTRACT && op1 == const0_rtx | |
5256 | && XEXP (op0, 1) == const1_rtx) | |
5257 | { | |
5258 | /* btst / bbit0 / bbit1: | |
5259 | Small integers and registers are free; everything else can | |
5260 | be put in a register. */ | |
e548c9df AM |
5261 | mode = GET_MODE (XEXP (op0, 0)); |
5262 | *total = (rtx_cost (XEXP (op0, 0), mode, SET, 1, speed) | |
5263 | + rtx_cost (XEXP (op0, 2), mode, SET, 1, speed)); | |
526b7aee SV |
5264 | return true; |
5265 | } | |
5266 | if (GET_CODE (op0) == AND && op1 == const0_rtx | |
5267 | && satisfies_constraint_C1p (XEXP (op0, 1))) | |
5268 | { | |
5269 | /* bmsk.f */ | |
e548c9df | 5270 | *total = rtx_cost (XEXP (op0, 0), VOIDmode, SET, 1, speed); |
526b7aee SV |
5271 | return true; |
5272 | } | |
5273 | /* add.f */ | |
5274 | if (GET_CODE (op1) == NEG) | |
5275 | { | |
5276 | /* op0 might be constant, the inside of op1 is rather | |
5277 | unlikely to be so. So swapping the operands might lower | |
5278 | the cost. */ | |
e548c9df AM |
5279 | mode = GET_MODE (op0); |
5280 | *total = (rtx_cost (op0, mode, PLUS, 1, speed) | |
5281 | + rtx_cost (XEXP (op1, 0), mode, PLUS, 0, speed)); | |
526b7aee SV |
5282 | } |
5283 | return false; | |
5284 | } | |
5285 | case EQ: case NE: | |
5286 | if (outer_code == IF_THEN_ELSE | |
5287 | && GET_CODE (XEXP (x, 0)) == ZERO_EXTRACT | |
5288 | && XEXP (x, 1) == const0_rtx | |
5289 | && XEXP (XEXP (x, 0), 1) == const1_rtx) | |
5290 | { | |
5291 | /* btst / bbit0 / bbit1: | |
5292 | Small integers and registers are free; everything else can | |
5293 | be put in a register. */ | |
5294 | rtx op0 = XEXP (x, 0); | |
5295 | ||
e548c9df AM |
5296 | mode = GET_MODE (XEXP (op0, 0)); |
5297 | *total = (rtx_cost (XEXP (op0, 0), mode, SET, 1, speed) | |
5298 | + rtx_cost (XEXP (op0, 2), mode, SET, 1, speed)); | |
526b7aee SV |
5299 | return true; |
5300 | } | |
5301 | /* Fall through. */ | |
5302 | /* scc_insn expands into two insns. */ | |
5303 | case GTU: case GEU: case LEU: | |
e548c9df | 5304 | if (mode == SImode) |
526b7aee SV |
5305 | *total += COSTS_N_INSNS (1); |
5306 | return false; | |
5307 | case LTU: /* might use adc. */ | |
e548c9df | 5308 | if (mode == SImode) |
526b7aee SV |
5309 | *total += COSTS_N_INSNS (1) - 1; |
5310 | return false; | |
5311 | default: | |
5312 | return false; | |
5313 | } | |
5314 | } | |
5315 | ||
526b7aee SV |
5316 | /* Return true if ADDR is a valid pic address. |
5317 | A valid pic address on arc should look like | |
5318 | const (unspec (SYMBOL_REF/LABEL) (ARC_UNSPEC_GOTOFF/ARC_UNSPEC_GOT)) */ | |
5319 | ||
5320 | bool | |
5321 | arc_legitimate_pic_addr_p (rtx addr) | |
5322 | { | |
526b7aee SV |
5323 | if (GET_CODE (addr) != CONST) |
5324 | return false; | |
5325 | ||
5326 | addr = XEXP (addr, 0); | |
5327 | ||
5328 | ||
5329 | if (GET_CODE (addr) == PLUS) | |
5330 | { | |
5331 | if (GET_CODE (XEXP (addr, 1)) != CONST_INT) | |
5332 | return false; | |
5333 | addr = XEXP (addr, 0); | |
5334 | } | |
5335 | ||
5336 | if (GET_CODE (addr) != UNSPEC | |
5337 | || XVECLEN (addr, 0) != 1) | |
5338 | return false; | |
5339 | ||
f5e336b1 | 5340 | /* Must be one of @GOT, @GOTOFF, @GOTOFFPC, @tlsgd, tlsie. */ |
526b7aee | 5341 | if (XINT (addr, 1) != ARC_UNSPEC_GOT |
28633bbd | 5342 | && XINT (addr, 1) != ARC_UNSPEC_GOTOFF |
f5e336b1 | 5343 | && XINT (addr, 1) != ARC_UNSPEC_GOTOFFPC |
28633bbd CZ |
5344 | && XINT (addr, 1) != UNSPEC_TLS_GD |
5345 | && XINT (addr, 1) != UNSPEC_TLS_IE) | |
526b7aee SV |
5346 | return false; |
5347 | ||
5348 | if (GET_CODE (XVECEXP (addr, 0, 0)) != SYMBOL_REF | |
5349 | && GET_CODE (XVECEXP (addr, 0, 0)) != LABEL_REF) | |
5350 | return false; | |
5351 | ||
5352 | return true; | |
5353 | } | |
5354 | ||
5355 | ||
5356 | ||
5357 | /* Return true if OP contains a symbol reference. */ | |
5358 | ||
5359 | static bool | |
5360 | symbolic_reference_mentioned_p (rtx op) | |
5361 | { | |
5362 | register const char *fmt; | |
5363 | register int i; | |
5364 | ||
5365 | if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF) | |
5366 | return true; | |
5367 | ||
5368 | fmt = GET_RTX_FORMAT (GET_CODE (op)); | |
5369 | for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--) | |
5370 | { | |
5371 | if (fmt[i] == 'E') | |
5372 | { | |
5373 | register int j; | |
5374 | ||
5375 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) | |
5376 | if (symbolic_reference_mentioned_p (XVECEXP (op, i, j))) | |
5377 | return true; | |
5378 | } | |
5379 | ||
5380 | else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i))) | |
5381 | return true; | |
5382 | } | |
5383 | ||
5384 | return false; | |
5385 | } | |
5386 | ||
5387 | /* Return true if OP contains a SYMBOL_REF that is not wrapped in an unspec. | |
5388 | If SKIP_LOCAL is true, skip symbols that bind locally. | |
5389 | This is used further down in this file, and, without SKIP_LOCAL, | |
5390 | in the addsi3 / subsi3 expanders when generating PIC code. */ | |
5391 | ||
5392 | bool | |
5393 | arc_raw_symbolic_reference_mentioned_p (rtx op, bool skip_local) | |
5394 | { | |
5395 | register const char *fmt; | |
5396 | register int i; | |
5397 | ||
5398 | if (GET_CODE(op) == UNSPEC) | |
5399 | return false; | |
5400 | ||
5401 | if (GET_CODE (op) == SYMBOL_REF) | |
5402 | { | |
28633bbd CZ |
5403 | if (SYMBOL_REF_TLS_MODEL (op)) |
5404 | return true; | |
5405 | if (!flag_pic) | |
5406 | return false; | |
526b7aee SV |
5407 | tree decl = SYMBOL_REF_DECL (op); |
5408 | return !skip_local || !decl || !default_binds_local_p (decl); | |
5409 | } | |
5410 | ||
5411 | fmt = GET_RTX_FORMAT (GET_CODE (op)); | |
5412 | for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--) | |
5413 | { | |
5414 | if (fmt[i] == 'E') | |
5415 | { | |
5416 | register int j; | |
5417 | ||
5418 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) | |
5419 | if (arc_raw_symbolic_reference_mentioned_p (XVECEXP (op, i, j), | |
5420 | skip_local)) | |
5421 | return true; | |
5422 | } | |
5423 | ||
5424 | else if (fmt[i] == 'e' | |
5425 | && arc_raw_symbolic_reference_mentioned_p (XEXP (op, i), | |
5426 | skip_local)) | |
5427 | return true; | |
5428 | } | |
5429 | ||
5430 | return false; | |
5431 | } | |
5432 | ||
28633bbd CZ |
5433 | /* Get the thread pointer. */ |
5434 | ||
5435 | static rtx | |
5436 | arc_get_tp (void) | |
5437 | { | |
5438 | /* If arc_tp_regno has been set, we can use that hard register | |
5439 | directly as a base register. */ | |
5440 | if (arc_tp_regno != -1) | |
5441 | return gen_rtx_REG (Pmode, arc_tp_regno); | |
5442 | ||
5443 | /* Otherwise, call __read_tp. Copy the result to a pseudo to avoid | |
5444 | conflicts with function arguments / results. */ | |
5445 | rtx reg = gen_reg_rtx (Pmode); | |
5446 | emit_insn (gen_tls_load_tp_soft ()); | |
5447 | emit_move_insn (reg, gen_rtx_REG (Pmode, R0_REG)); | |
5448 | return reg; | |
5449 | } | |
5450 | ||
5451 | /* Helper to be used by TLS Global dynamic model. */ | |
5452 | ||
5453 | static rtx | |
5454 | arc_emit_call_tls_get_addr (rtx sym, int reloc, rtx eqv) | |
5455 | { | |
5456 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); | |
28633bbd CZ |
5457 | rtx call_fusage = NULL_RTX; |
5458 | ||
5459 | start_sequence (); | |
5460 | ||
5a5c5784 | 5461 | rtx x = arc_unspec_offset (sym, reloc); |
28633bbd CZ |
5462 | emit_move_insn (r0, x); |
5463 | use_reg (&call_fusage, r0); | |
5464 | ||
5465 | gcc_assert (reloc == UNSPEC_TLS_GD); | |
5466 | rtx call_insn = emit_call_insn (gen_tls_gd_get_addr (sym)); | |
5467 | /* Should we set RTL_CONST_CALL_P? We read memory, but not in a | |
5468 | way that the application should care. */ | |
5469 | RTL_PURE_CALL_P (call_insn) = 1; | |
5470 | add_function_usage_to (call_insn, call_fusage); | |
5471 | ||
9b2ea071 | 5472 | rtx_insn *insns = get_insns (); |
28633bbd CZ |
5473 | end_sequence (); |
5474 | ||
5475 | rtx dest = gen_reg_rtx (Pmode); | |
5476 | emit_libcall_block (insns, dest, r0, eqv); | |
5477 | return dest; | |
5478 | } | |
5479 | ||
5480 | #define DTPOFF_ZERO_SYM ".tdata" | |
5481 | ||
5482 | /* Return a legitimized address for ADDR, | |
5483 | which is a SYMBOL_REF with tls_model MODEL. */ | |
5484 | ||
5485 | static rtx | |
5486 | arc_legitimize_tls_address (rtx addr, enum tls_model model) | |
5487 | { | |
5488 | if (!flag_pic && model == TLS_MODEL_LOCAL_DYNAMIC) | |
5489 | model = TLS_MODEL_LOCAL_EXEC; | |
5490 | ||
5491 | switch (model) | |
5492 | { | |
5493 | case TLS_MODEL_LOCAL_DYNAMIC: | |
5494 | rtx base; | |
5495 | tree decl; | |
5496 | const char *base_name; | |
5497 | rtvec v; | |
5498 | ||
5499 | decl = SYMBOL_REF_DECL (addr); | |
5500 | base_name = DTPOFF_ZERO_SYM; | |
5501 | if (decl && bss_initializer_p (decl)) | |
5502 | base_name = ".tbss"; | |
5503 | ||
5504 | base = gen_rtx_SYMBOL_REF (Pmode, base_name); | |
5505 | if (strcmp (base_name, DTPOFF_ZERO_SYM) == 0) | |
5506 | { | |
5507 | if (!flag_pic) | |
5508 | goto local_exec; | |
5509 | v = gen_rtvec (1, addr); | |
5510 | } | |
5511 | else | |
5512 | v = gen_rtvec (2, addr, base); | |
5513 | addr = gen_rtx_UNSPEC (Pmode, v, UNSPEC_TLS_OFF); | |
5514 | addr = gen_rtx_CONST (Pmode, addr); | |
5515 | base = arc_legitimize_tls_address (base, TLS_MODEL_GLOBAL_DYNAMIC); | |
5516 | return gen_rtx_PLUS (Pmode, force_reg (Pmode, base), addr); | |
5a5c5784 | 5517 | |
28633bbd CZ |
5518 | case TLS_MODEL_GLOBAL_DYNAMIC: |
5519 | return arc_emit_call_tls_get_addr (addr, UNSPEC_TLS_GD, addr); | |
5a5c5784 | 5520 | |
28633bbd | 5521 | case TLS_MODEL_INITIAL_EXEC: |
5a5c5784 | 5522 | addr = arc_unspec_offset (addr, UNSPEC_TLS_IE); |
28633bbd CZ |
5523 | addr = copy_to_mode_reg (Pmode, gen_const_mem (Pmode, addr)); |
5524 | return gen_rtx_PLUS (Pmode, arc_get_tp (), addr); | |
5a5c5784 | 5525 | |
28633bbd CZ |
5526 | case TLS_MODEL_LOCAL_EXEC: |
5527 | local_exec: | |
5a5c5784 | 5528 | addr = arc_unspec_offset (addr, UNSPEC_TLS_OFF); |
28633bbd CZ |
5529 | return gen_rtx_PLUS (Pmode, arc_get_tp (), addr); |
5530 | default: | |
5531 | gcc_unreachable (); | |
5532 | } | |
5533 | } | |
5534 | ||
526b7aee SV |
5535 | /* Legitimize a pic address reference in ORIG. |
5536 | The return value is the legitimated address. | |
5537 | If OLDX is non-zero, it is the target to assign the address to first. */ | |
5538 | ||
28633bbd | 5539 | static rtx |
526b7aee SV |
5540 | arc_legitimize_pic_address (rtx orig, rtx oldx) |
5541 | { | |
5542 | rtx addr = orig; | |
5543 | rtx pat = orig; | |
5544 | rtx base; | |
5545 | ||
5546 | if (oldx == orig) | |
5547 | oldx = NULL; | |
5548 | ||
5549 | if (GET_CODE (addr) == LABEL_REF) | |
5550 | ; /* Do nothing. */ | |
28633bbd | 5551 | else if (GET_CODE (addr) == SYMBOL_REF) |
526b7aee | 5552 | { |
28633bbd CZ |
5553 | enum tls_model model = SYMBOL_REF_TLS_MODEL (addr); |
5554 | if (model != 0) | |
5555 | return arc_legitimize_tls_address (addr, model); | |
5556 | else if (!flag_pic) | |
5557 | return orig; | |
5558 | else if (CONSTANT_POOL_ADDRESS_P (addr) || SYMBOL_REF_LOCAL_P (addr)) | |
5a5c5784 | 5559 | return arc_unspec_offset (addr, ARC_UNSPEC_GOTOFFPC); |
f5e336b1 CZ |
5560 | |
5561 | /* This symbol must be referenced via a load from the Global | |
5562 | Offset Table (@GOTPC). */ | |
5a5c5784 | 5563 | pat = arc_unspec_offset (addr, ARC_UNSPEC_GOT); |
f5e336b1 | 5564 | pat = gen_const_mem (Pmode, pat); |
526b7aee | 5565 | |
28633bbd | 5566 | if (oldx == NULL) |
526b7aee SV |
5567 | oldx = gen_reg_rtx (Pmode); |
5568 | ||
5569 | emit_move_insn (oldx, pat); | |
5570 | pat = oldx; | |
5571 | } | |
5572 | else | |
5573 | { | |
5574 | if (GET_CODE (addr) == CONST) | |
5575 | { | |
5576 | addr = XEXP (addr, 0); | |
5577 | if (GET_CODE (addr) == UNSPEC) | |
5578 | { | |
5579 | /* Check that the unspec is one of the ones we generate? */ | |
f5e336b1 | 5580 | return orig; |
526b7aee | 5581 | } |
14555394 CZ |
5582 | /* fwprop is placing in the REG_EQUIV notes constant pic |
5583 | unspecs expressions. Then, loop may use these notes for | |
5584 | optimizations resulting in complex patterns that are not | |
5585 | supported by the current implementation. The following | |
5586 | two if-cases are simplifying the complex patters to | |
5587 | simpler ones. */ | |
5588 | else if (GET_CODE (addr) == MINUS) | |
5589 | { | |
5590 | rtx op0 = XEXP (addr, 0); | |
5591 | rtx op1 = XEXP (addr, 1); | |
5592 | gcc_assert (oldx); | |
5593 | gcc_assert (GET_CODE (op1) == UNSPEC); | |
5594 | ||
5595 | emit_move_insn (oldx, | |
5596 | gen_rtx_CONST (SImode, | |
5597 | arc_legitimize_pic_address (op1, | |
5598 | NULL_RTX))); | |
5599 | emit_insn (gen_rtx_SET (oldx, gen_rtx_MINUS (SImode, op0, oldx))); | |
5600 | return oldx; | |
5601 | ||
5602 | } | |
5603 | else if (GET_CODE (addr) != PLUS) | |
5604 | { | |
5605 | rtx tmp = XEXP (addr, 0); | |
5606 | enum rtx_code code = GET_CODE (addr); | |
5607 | ||
5608 | /* It only works for UNARY operations. */ | |
5609 | gcc_assert (UNARY_P (addr)); | |
5610 | gcc_assert (GET_CODE (tmp) == UNSPEC); | |
5611 | gcc_assert (oldx); | |
5612 | ||
5613 | emit_move_insn | |
5614 | (oldx, | |
5615 | gen_rtx_CONST (SImode, | |
5616 | arc_legitimize_pic_address (tmp, | |
5617 | NULL_RTX))); | |
5618 | ||
5619 | emit_insn (gen_rtx_SET (oldx, | |
5620 | gen_rtx_fmt_ee (code, SImode, | |
5621 | oldx, const0_rtx))); | |
5622 | ||
5623 | return oldx; | |
5624 | } | |
526b7aee | 5625 | else |
14555394 CZ |
5626 | { |
5627 | gcc_assert (GET_CODE (addr) == PLUS); | |
5628 | if (GET_CODE (XEXP (addr, 0)) == UNSPEC) | |
5629 | return orig; | |
5630 | } | |
526b7aee SV |
5631 | } |
5632 | ||
5633 | if (GET_CODE (addr) == PLUS) | |
5634 | { | |
5635 | rtx op0 = XEXP (addr, 0), op1 = XEXP (addr, 1); | |
5636 | ||
28633bbd CZ |
5637 | base = arc_legitimize_pic_address (op0, oldx); |
5638 | pat = arc_legitimize_pic_address (op1, | |
526b7aee SV |
5639 | base == oldx ? NULL_RTX : oldx); |
5640 | ||
28633bbd CZ |
5641 | if (base == op0 && pat == op1) |
5642 | return orig; | |
5643 | ||
5644 | if (GET_CODE (pat) == CONST_INT) | |
5645 | pat = plus_constant (Pmode, base, INTVAL (pat)); | |
5646 | else | |
5647 | { | |
5648 | if (GET_CODE (pat) == PLUS && CONSTANT_P (XEXP (pat, 1))) | |
526b7aee | 5649 | { |
28633bbd CZ |
5650 | base = gen_rtx_PLUS (Pmode, base, XEXP (pat, 0)); |
5651 | pat = XEXP (pat, 1); | |
526b7aee | 5652 | } |
28633bbd | 5653 | pat = gen_rtx_PLUS (Pmode, base, pat); |
526b7aee SV |
5654 | } |
5655 | } | |
5656 | } | |
5657 | ||
5658 | return pat; | |
5659 | } | |
5660 | ||
5661 | /* Output address constant X to FILE, taking PIC into account. */ | |
5662 | ||
9f532472 | 5663 | static void |
526b7aee SV |
5664 | arc_output_pic_addr_const (FILE * file, rtx x, int code) |
5665 | { | |
5666 | char buf[256]; | |
5667 | ||
5668 | restart: | |
5669 | switch (GET_CODE (x)) | |
5670 | { | |
5671 | case PC: | |
5672 | if (flag_pic) | |
5673 | putc ('.', file); | |
5674 | else | |
5675 | gcc_unreachable (); | |
5676 | break; | |
5677 | ||
5678 | case SYMBOL_REF: | |
5679 | output_addr_const (file, x); | |
5680 | ||
5681 | /* Local functions do not get references through the PLT. */ | |
5682 | if (code == 'P' && ! SYMBOL_REF_LOCAL_P (x)) | |
5683 | fputs ("@plt", file); | |
5684 | break; | |
5685 | ||
5686 | case LABEL_REF: | |
5687 | ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (XEXP (x, 0))); | |
5688 | assemble_name (file, buf); | |
5689 | break; | |
5690 | ||
5691 | case CODE_LABEL: | |
5692 | ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x)); | |
5693 | assemble_name (file, buf); | |
5694 | break; | |
5695 | ||
5696 | case CONST_INT: | |
5697 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
5698 | break; | |
5699 | ||
5700 | case CONST: | |
5701 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
5702 | break; | |
5703 | ||
5704 | case CONST_DOUBLE: | |
5705 | if (GET_MODE (x) == VOIDmode) | |
5706 | { | |
5707 | /* We can use %d if the number is one word and positive. */ | |
5708 | if (CONST_DOUBLE_HIGH (x)) | |
5709 | fprintf (file, HOST_WIDE_INT_PRINT_DOUBLE_HEX, | |
5710 | CONST_DOUBLE_HIGH (x), CONST_DOUBLE_LOW (x)); | |
5711 | else if (CONST_DOUBLE_LOW (x) < 0) | |
5712 | fprintf (file, HOST_WIDE_INT_PRINT_HEX, CONST_DOUBLE_LOW (x)); | |
5713 | else | |
5714 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, CONST_DOUBLE_LOW (x)); | |
5715 | } | |
5716 | else | |
5717 | /* We can't handle floating point constants; | |
5718 | PRINT_OPERAND must handle them. */ | |
5719 | output_operand_lossage ("floating constant misused"); | |
5720 | break; | |
5721 | ||
5722 | case PLUS: | |
5723 | /* FIXME: Not needed here. */ | |
5724 | /* Some assemblers need integer constants to appear last (eg masm). */ | |
5725 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) | |
5726 | { | |
5727 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
5728 | fprintf (file, "+"); | |
5729 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
5730 | } | |
5731 | else if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5732 | { | |
5733 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
5734 | if (INTVAL (XEXP (x, 1)) >= 0) | |
5735 | fprintf (file, "+"); | |
5736 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
5737 | } | |
5738 | else | |
5739 | gcc_unreachable(); | |
5740 | break; | |
5741 | ||
5742 | case MINUS: | |
5743 | /* Avoid outputting things like x-x or x+5-x, | |
5744 | since some assemblers can't handle that. */ | |
5745 | x = simplify_subtraction (x); | |
5746 | if (GET_CODE (x) != MINUS) | |
5747 | goto restart; | |
5748 | ||
5749 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
5750 | fprintf (file, "-"); | |
5751 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
5752 | && INTVAL (XEXP (x, 1)) < 0) | |
5753 | { | |
5754 | fprintf (file, "("); | |
5755 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
5756 | fprintf (file, ")"); | |
5757 | } | |
5758 | else | |
5759 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
5760 | break; | |
5761 | ||
5762 | case ZERO_EXTEND: | |
5763 | case SIGN_EXTEND: | |
5764 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
5765 | break; | |
5766 | ||
5767 | ||
5768 | case UNSPEC: | |
28633bbd CZ |
5769 | const char *suffix; |
5770 | bool pcrel; pcrel = false; | |
5771 | rtx base; base = NULL; | |
5772 | gcc_assert (XVECLEN (x, 0) >= 1); | |
526b7aee SV |
5773 | switch (XINT (x, 1)) |
5774 | { | |
5775 | case ARC_UNSPEC_GOT: | |
28633bbd | 5776 | suffix = "@gotpc", pcrel = true; |
526b7aee SV |
5777 | break; |
5778 | case ARC_UNSPEC_GOTOFF: | |
28633bbd | 5779 | suffix = "@gotoff"; |
526b7aee | 5780 | break; |
f5e336b1 CZ |
5781 | case ARC_UNSPEC_GOTOFFPC: |
5782 | suffix = "@pcl", pcrel = true; | |
5783 | break; | |
526b7aee | 5784 | case ARC_UNSPEC_PLT: |
28633bbd CZ |
5785 | suffix = "@plt"; |
5786 | break; | |
5787 | case UNSPEC_TLS_GD: | |
5788 | suffix = "@tlsgd", pcrel = true; | |
5789 | break; | |
5790 | case UNSPEC_TLS_IE: | |
5791 | suffix = "@tlsie", pcrel = true; | |
5792 | break; | |
5793 | case UNSPEC_TLS_OFF: | |
5794 | if (XVECLEN (x, 0) == 2) | |
5795 | base = XVECEXP (x, 0, 1); | |
5796 | if (SYMBOL_REF_TLS_MODEL (XVECEXP (x, 0, 0)) == TLS_MODEL_LOCAL_EXEC | |
5797 | || (!flag_pic && !base)) | |
5798 | suffix = "@tpoff"; | |
5799 | else | |
5800 | suffix = "@dtpoff"; | |
526b7aee SV |
5801 | break; |
5802 | default: | |
cd1e4d41 | 5803 | suffix = "@invalid"; |
526b7aee SV |
5804 | output_operand_lossage ("invalid UNSPEC as operand: %d", XINT (x,1)); |
5805 | break; | |
5806 | } | |
28633bbd CZ |
5807 | if (pcrel) |
5808 | fputs ("pcl,", file); | |
5809 | arc_output_pic_addr_const (file, XVECEXP (x, 0, 0), code); | |
5810 | fputs (suffix, file); | |
5811 | if (base) | |
5812 | arc_output_pic_addr_const (file, base, code); | |
5813 | break; | |
526b7aee SV |
5814 | |
5815 | default: | |
5816 | output_operand_lossage ("invalid expression as operand"); | |
5817 | } | |
5818 | } | |
5819 | ||
5820 | #define SYMBOLIC_CONST(X) \ | |
5821 | (GET_CODE (X) == SYMBOL_REF \ | |
5822 | || GET_CODE (X) == LABEL_REF \ | |
5823 | || (GET_CODE (X) == CONST && symbolic_reference_mentioned_p (X))) | |
5824 | ||
5825 | /* Emit insns to move operands[1] into operands[0]. */ | |
5826 | ||
28633bbd CZ |
5827 | static void |
5828 | prepare_pic_move (rtx *operands, machine_mode) | |
526b7aee | 5829 | { |
28633bbd CZ |
5830 | if (GET_CODE (operands[0]) == MEM && SYMBOLIC_CONST (operands[1]) |
5831 | && flag_pic) | |
526b7aee SV |
5832 | operands[1] = force_reg (Pmode, operands[1]); |
5833 | else | |
28633bbd CZ |
5834 | { |
5835 | rtx temp = (reload_in_progress ? operands[0] | |
5836 | : flag_pic? gen_reg_rtx (Pmode) : NULL_RTX); | |
5837 | operands[1] = arc_legitimize_pic_address (operands[1], temp); | |
5838 | } | |
526b7aee SV |
5839 | } |
5840 | ||
5841 | ||
5842 | /* The function returning the number of words, at the beginning of an | |
5843 | argument, must be put in registers. The returned value must be | |
5844 | zero for arguments that are passed entirely in registers or that | |
5845 | are entirely pushed on the stack. | |
5846 | ||
5847 | On some machines, certain arguments must be passed partially in | |
5848 | registers and partially in memory. On these machines, typically | |
5849 | the first N words of arguments are passed in registers, and the | |
5850 | rest on the stack. If a multi-word argument (a `double' or a | |
5851 | structure) crosses that boundary, its first few words must be | |
5852 | passed in registers and the rest must be pushed. This function | |
5853 | tells the compiler when this occurs, and how many of the words | |
5854 | should go in registers. | |
5855 | ||
5856 | `FUNCTION_ARG' for these arguments should return the first register | |
5857 | to be used by the caller for this argument; likewise | |
5858 | `FUNCTION_INCOMING_ARG', for the called function. | |
5859 | ||
5860 | The function is used to implement macro FUNCTION_ARG_PARTIAL_NREGS. */ | |
5861 | ||
5862 | /* If REGNO is the least arg reg available then what is the total number of arg | |
5863 | regs available. */ | |
5864 | #define GPR_REST_ARG_REGS(REGNO) \ | |
5865 | ((REGNO) <= MAX_ARC_PARM_REGS ? MAX_ARC_PARM_REGS - (REGNO) : 0 ) | |
5866 | ||
5867 | /* Since arc parm regs are contiguous. */ | |
5868 | #define ARC_NEXT_ARG_REG(REGNO) ( (REGNO) + 1 ) | |
5869 | ||
5870 | /* Implement TARGET_ARG_PARTIAL_BYTES. */ | |
5871 | ||
5872 | static int | |
ef4bddc2 | 5873 | arc_arg_partial_bytes (cumulative_args_t cum_v, machine_mode mode, |
526b7aee SV |
5874 | tree type, bool named ATTRIBUTE_UNUSED) |
5875 | { | |
5876 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
5877 | int bytes = (mode == BLKmode | |
5878 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode)); | |
5879 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
5880 | int arg_num = *cum; | |
5881 | int ret; | |
5882 | ||
5883 | arg_num = ROUND_ADVANCE_CUM (arg_num, mode, type); | |
5884 | ret = GPR_REST_ARG_REGS (arg_num); | |
5885 | ||
5886 | /* ICEd at function.c:2361, and ret is copied to data->partial */ | |
5887 | ret = (ret >= words ? 0 : ret * UNITS_PER_WORD); | |
5888 | ||
5889 | return ret; | |
5890 | } | |
5891 | ||
526b7aee SV |
5892 | /* This function is used to control a function argument is passed in a |
5893 | register, and which register. | |
5894 | ||
5895 | The arguments are CUM, of type CUMULATIVE_ARGS, which summarizes | |
5896 | (in a way defined by INIT_CUMULATIVE_ARGS and FUNCTION_ARG_ADVANCE) | |
5897 | all of the previous arguments so far passed in registers; MODE, the | |
5898 | machine mode of the argument; TYPE, the data type of the argument | |
5899 | as a tree node or 0 if that is not known (which happens for C | |
5900 | support library functions); and NAMED, which is 1 for an ordinary | |
5901 | argument and 0 for nameless arguments that correspond to `...' in | |
5902 | the called function's prototype. | |
5903 | ||
5904 | The returned value should either be a `reg' RTX for the hard | |
5905 | register in which to pass the argument, or zero to pass the | |
5906 | argument on the stack. | |
5907 | ||
5908 | For machines like the Vax and 68000, where normally all arguments | |
5909 | are pushed, zero suffices as a definition. | |
5910 | ||
5911 | The usual way to make the ANSI library `stdarg.h' work on a machine | |
5912 | where some arguments are usually passed in registers, is to cause | |
5913 | nameless arguments to be passed on the stack instead. This is done | |
5914 | by making the function return 0 whenever NAMED is 0. | |
5915 | ||
5916 | You may use the macro `MUST_PASS_IN_STACK (MODE, TYPE)' in the | |
5917 | definition of this function to determine if this argument is of a | |
5918 | type that must be passed in the stack. If `REG_PARM_STACK_SPACE' | |
5919 | is not defined and the function returns non-zero for such an | |
5920 | argument, the compiler will abort. If `REG_PARM_STACK_SPACE' is | |
5921 | defined, the argument will be computed in the stack and then loaded | |
5922 | into a register. | |
5923 | ||
5924 | The function is used to implement macro FUNCTION_ARG. */ | |
5925 | /* On the ARC the first MAX_ARC_PARM_REGS args are normally in registers | |
5926 | and the rest are pushed. */ | |
5927 | ||
5928 | static rtx | |
8f3304d0 CZ |
5929 | arc_function_arg (cumulative_args_t cum_v, |
5930 | machine_mode mode, | |
5931 | const_tree type ATTRIBUTE_UNUSED, | |
5932 | bool named ATTRIBUTE_UNUSED) | |
526b7aee SV |
5933 | { |
5934 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
5935 | int arg_num = *cum; | |
5936 | rtx ret; | |
5937 | const char *debstr ATTRIBUTE_UNUSED; | |
5938 | ||
5939 | arg_num = ROUND_ADVANCE_CUM (arg_num, mode, type); | |
5940 | /* Return a marker for use in the call instruction. */ | |
5941 | if (mode == VOIDmode) | |
5942 | { | |
5943 | ret = const0_rtx; | |
5944 | debstr = "<0>"; | |
5945 | } | |
5946 | else if (GPR_REST_ARG_REGS (arg_num) > 0) | |
5947 | { | |
5948 | ret = gen_rtx_REG (mode, arg_num); | |
5949 | debstr = reg_names [arg_num]; | |
5950 | } | |
5951 | else | |
5952 | { | |
5953 | ret = NULL_RTX; | |
5954 | debstr = "memory"; | |
5955 | } | |
5956 | return ret; | |
5957 | } | |
5958 | ||
5959 | /* The function to update the summarizer variable *CUM to advance past | |
5960 | an argument in the argument list. The values MODE, TYPE and NAMED | |
5961 | describe that argument. Once this is done, the variable *CUM is | |
5962 | suitable for analyzing the *following* argument with | |
5963 | `FUNCTION_ARG', etc. | |
5964 | ||
5965 | This function need not do anything if the argument in question was | |
5966 | passed on the stack. The compiler knows how to track the amount of | |
5967 | stack space used for arguments without any special help. | |
5968 | ||
5969 | The function is used to implement macro FUNCTION_ARG_ADVANCE. */ | |
5970 | /* For the ARC: the cum set here is passed on to function_arg where we | |
5971 | look at its value and say which reg to use. Strategy: advance the | |
5972 | regnumber here till we run out of arg regs, then set *cum to last | |
5973 | reg. In function_arg, since *cum > last arg reg we would return 0 | |
5974 | and thus the arg will end up on the stack. For straddling args of | |
5975 | course function_arg_partial_nregs will come into play. */ | |
5976 | ||
5977 | static void | |
8f3304d0 CZ |
5978 | arc_function_arg_advance (cumulative_args_t cum_v, |
5979 | machine_mode mode, | |
5980 | const_tree type, | |
5981 | bool named ATTRIBUTE_UNUSED) | |
526b7aee SV |
5982 | { |
5983 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
5984 | int bytes = (mode == BLKmode | |
5985 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode)); | |
5986 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
5987 | int i; | |
5988 | ||
5989 | if (words) | |
5990 | *cum = ROUND_ADVANCE_CUM (*cum, mode, type); | |
5991 | for (i = 0; i < words; i++) | |
5992 | *cum = ARC_NEXT_ARG_REG (*cum); | |
5993 | ||
5994 | } | |
5995 | ||
5996 | /* Define how to find the value returned by a function. | |
5997 | VALTYPE is the data type of the value (as a tree). | |
5998 | If the precise function being called is known, FN_DECL_OR_TYPE is its | |
5999 | FUNCTION_DECL; otherwise, FN_DECL_OR_TYPE is its type. */ | |
6000 | ||
6001 | static rtx | |
6002 | arc_function_value (const_tree valtype, | |
6003 | const_tree fn_decl_or_type ATTRIBUTE_UNUSED, | |
6004 | bool outgoing ATTRIBUTE_UNUSED) | |
6005 | { | |
ef4bddc2 | 6006 | machine_mode mode = TYPE_MODE (valtype); |
526b7aee SV |
6007 | int unsignedp ATTRIBUTE_UNUSED; |
6008 | ||
6009 | unsignedp = TYPE_UNSIGNED (valtype); | |
6010 | if (INTEGRAL_TYPE_P (valtype) || TREE_CODE (valtype) == OFFSET_TYPE) | |
6011 | PROMOTE_MODE (mode, unsignedp, valtype); | |
6012 | return gen_rtx_REG (mode, 0); | |
6013 | } | |
6014 | ||
6015 | /* Returns the return address that is used by builtin_return_address. */ | |
6016 | ||
6017 | rtx | |
6018 | arc_return_addr_rtx (int count, ATTRIBUTE_UNUSED rtx frame) | |
6019 | { | |
6020 | if (count != 0) | |
6021 | return const0_rtx; | |
6022 | ||
6023 | return get_hard_reg_initial_val (Pmode , RETURN_ADDR_REGNUM); | |
6024 | } | |
6025 | ||
526b7aee SV |
6026 | /* Determine if a given RTX is a valid constant. We already know this |
6027 | satisfies CONSTANT_P. */ | |
6028 | ||
6029 | bool | |
28633bbd | 6030 | arc_legitimate_constant_p (machine_mode mode, rtx x) |
526b7aee | 6031 | { |
526b7aee SV |
6032 | switch (GET_CODE (x)) |
6033 | { | |
6034 | case CONST: | |
b6c354eb | 6035 | if (flag_pic) |
526b7aee | 6036 | { |
b6c354eb | 6037 | if (arc_legitimate_pic_addr_p (x)) |
526b7aee | 6038 | return true; |
b6c354eb CZ |
6039 | } |
6040 | return arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
526b7aee | 6041 | |
526b7aee | 6042 | case SYMBOL_REF: |
28633bbd CZ |
6043 | if (SYMBOL_REF_TLS_MODEL (x)) |
6044 | return false; | |
6045 | /* Fall through. */ | |
6046 | case LABEL_REF: | |
6047 | if (flag_pic) | |
6048 | return false; | |
6049 | /* Fall through. */ | |
b6c354eb CZ |
6050 | case CONST_INT: |
6051 | case CONST_DOUBLE: | |
6052 | return true; | |
6053 | ||
6054 | case NEG: | |
6055 | return arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
6056 | ||
6057 | case PLUS: | |
6058 | case MINUS: | |
6059 | { | |
6060 | bool t1 = arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
6061 | bool t2 = arc_legitimate_constant_p (mode, XEXP (x, 1)); | |
6062 | ||
6063 | return (t1 && t2); | |
6064 | } | |
6065 | ||
6066 | case CONST_VECTOR: | |
6067 | switch (mode) | |
6068 | { | |
4e10a5a7 | 6069 | case E_V2HImode: |
b6c354eb | 6070 | return TARGET_PLUS_DMPY; |
4e10a5a7 RS |
6071 | case E_V2SImode: |
6072 | case E_V4HImode: | |
b6c354eb CZ |
6073 | return TARGET_PLUS_QMACW; |
6074 | default: | |
6075 | return false; | |
6076 | } | |
6077 | ||
6078 | case UNSPEC: | |
6079 | switch (XINT (x, 1)) | |
6080 | { | |
6081 | case UNSPEC_TLS_GD: | |
6082 | case UNSPEC_TLS_OFF: | |
6083 | case UNSPEC_TLS_IE: | |
6084 | return true; | |
6085 | default: | |
6086 | /* Any other unspec ending here are pic related, hence the above | |
6087 | constant pic address checking returned false. */ | |
6088 | return false; | |
6089 | } | |
6090 | /* Fall through. */ | |
526b7aee SV |
6091 | |
6092 | default: | |
b6c354eb | 6093 | fatal_insn ("unrecognized supposed constant", x); |
526b7aee SV |
6094 | } |
6095 | ||
b6c354eb | 6096 | gcc_unreachable (); |
526b7aee SV |
6097 | } |
6098 | ||
6099 | static bool | |
ef4bddc2 | 6100 | arc_legitimate_address_p (machine_mode mode, rtx x, bool strict) |
526b7aee SV |
6101 | { |
6102 | if (RTX_OK_FOR_BASE_P (x, strict)) | |
6103 | return true; | |
ac2e1a51 | 6104 | if (legitimate_offset_address_p (mode, x, TARGET_INDEXED_LOADS, strict)) |
526b7aee | 6105 | return true; |
9f532472 | 6106 | if (legitimate_scaled_address_p (mode, x, strict)) |
526b7aee SV |
6107 | return true; |
6108 | if (LEGITIMATE_SMALL_DATA_ADDRESS_P (x)) | |
6109 | return true; | |
6110 | if (GET_CODE (x) == CONST_INT && LARGE_INT (INTVAL (x))) | |
6111 | return true; | |
28633bbd CZ |
6112 | |
6113 | /* When we compile for size avoid const (@sym + offset) | |
6114 | addresses. */ | |
6115 | if (!flag_pic && optimize_size && !reload_completed | |
6116 | && (GET_CODE (x) == CONST) | |
6117 | && (GET_CODE (XEXP (x, 0)) == PLUS) | |
6118 | && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
6119 | && SYMBOL_REF_TLS_MODEL (XEXP (XEXP (x, 0), 0)) == 0 | |
6120 | && !SYMBOL_REF_FUNCTION_P (XEXP (XEXP (x, 0), 0))) | |
526b7aee | 6121 | { |
28633bbd CZ |
6122 | rtx addend = XEXP (XEXP (x, 0), 1); |
6123 | gcc_assert (CONST_INT_P (addend)); | |
6124 | HOST_WIDE_INT offset = INTVAL (addend); | |
6125 | ||
6126 | /* Allow addresses having a large offset to pass. Anyhow they | |
6127 | will end in a limm. */ | |
6128 | return !(offset > -1024 && offset < 1020); | |
6129 | } | |
6130 | ||
6131 | if ((GET_MODE_SIZE (mode) != 16) && CONSTANT_P (x)) | |
6132 | { | |
b6c354eb | 6133 | return arc_legitimate_constant_p (mode, x); |
526b7aee SV |
6134 | } |
6135 | if ((GET_CODE (x) == PRE_DEC || GET_CODE (x) == PRE_INC | |
6136 | || GET_CODE (x) == POST_DEC || GET_CODE (x) == POST_INC) | |
6137 | && RTX_OK_FOR_BASE_P (XEXP (x, 0), strict)) | |
6138 | return true; | |
6139 | /* We're restricted here by the `st' insn. */ | |
6140 | if ((GET_CODE (x) == PRE_MODIFY || GET_CODE (x) == POST_MODIFY) | |
6141 | && GET_CODE (XEXP ((x), 1)) == PLUS | |
6142 | && rtx_equal_p (XEXP ((x), 0), XEXP (XEXP (x, 1), 0)) | |
ac2e1a51 | 6143 | && legitimate_offset_address_p (QImode, XEXP (x, 1), |
526b7aee SV |
6144 | TARGET_AUTO_MODIFY_REG, strict)) |
6145 | return true; | |
6146 | return false; | |
6147 | } | |
6148 | ||
6149 | /* Return true iff ADDR (a legitimate address expression) | |
6150 | has an effect that depends on the machine mode it is used for. */ | |
6151 | ||
6152 | static bool | |
6153 | arc_mode_dependent_address_p (const_rtx addr, addr_space_t) | |
6154 | { | |
6155 | /* SYMBOL_REF is not mode dependent: it is either a small data reference, | |
6156 | which is valid for loads and stores, or a limm offset, which is valid for | |
1fccdd40 | 6157 | loads. Scaled indices are scaled by the access mode. */ |
526b7aee | 6158 | if (GET_CODE (addr) == PLUS |
1fccdd40 | 6159 | && GET_CODE (XEXP ((addr), 0)) == MULT) |
526b7aee SV |
6160 | return true; |
6161 | return false; | |
6162 | } | |
6163 | ||
6164 | /* Determine if it's legal to put X into the constant pool. */ | |
6165 | ||
6166 | static bool | |
ef4bddc2 | 6167 | arc_cannot_force_const_mem (machine_mode mode, rtx x) |
526b7aee SV |
6168 | { |
6169 | return !arc_legitimate_constant_p (mode, x); | |
6170 | } | |
6171 | ||
c69899f0 CZ |
6172 | /* IDs for all the ARC builtins. */ |
6173 | ||
6174 | enum arc_builtin_id | |
6175 | { | |
6176 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6177 | ARC_BUILTIN_ ## NAME, | |
6178 | #include "builtins.def" | |
6179 | #undef DEF_BUILTIN | |
6180 | ||
6181 | ARC_BUILTIN_COUNT | |
6182 | }; | |
6183 | ||
6184 | struct GTY(()) arc_builtin_description | |
6185 | { | |
6186 | enum insn_code icode; | |
6187 | int n_args; | |
6188 | tree fndecl; | |
6189 | }; | |
6190 | ||
6191 | static GTY(()) struct arc_builtin_description | |
6192 | arc_bdesc[ARC_BUILTIN_COUNT] = | |
6193 | { | |
6194 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6195 | { (enum insn_code) CODE_FOR_ ## ICODE, N_ARGS, NULL_TREE }, | |
6196 | #include "builtins.def" | |
6197 | #undef DEF_BUILTIN | |
6198 | }; | |
6199 | ||
6200 | /* Transform UP into lowercase and write the result to LO. | |
6201 | You must provide enough space for LO. Return LO. */ | |
6202 | ||
6203 | static char* | |
6204 | arc_tolower (char *lo, const char *up) | |
6205 | { | |
6206 | char *lo0 = lo; | |
6207 | ||
6208 | for (; *up; up++, lo++) | |
6209 | *lo = TOLOWER (*up); | |
6210 | ||
6211 | *lo = '\0'; | |
6212 | ||
6213 | return lo0; | |
6214 | } | |
6215 | ||
6216 | /* Implement `TARGET_BUILTIN_DECL'. */ | |
526b7aee | 6217 | |
c69899f0 CZ |
6218 | static tree |
6219 | arc_builtin_decl (unsigned id, bool initialize_p ATTRIBUTE_UNUSED) | |
6220 | { | |
6221 | if (id < ARC_BUILTIN_COUNT) | |
6222 | return arc_bdesc[id].fndecl; | |
526b7aee | 6223 | |
c69899f0 CZ |
6224 | return error_mark_node; |
6225 | } | |
526b7aee SV |
6226 | |
6227 | static void | |
6228 | arc_init_builtins (void) | |
6229 | { | |
00c072ae CZ |
6230 | tree V4HI_type_node; |
6231 | tree V2SI_type_node; | |
6232 | tree V2HI_type_node; | |
6233 | ||
6234 | /* Vector types based on HS SIMD elements. */ | |
6235 | V4HI_type_node = build_vector_type_for_mode (intHI_type_node, V4HImode); | |
6236 | V2SI_type_node = build_vector_type_for_mode (intSI_type_node, V2SImode); | |
6237 | V2HI_type_node = build_vector_type_for_mode (intHI_type_node, V2HImode); | |
6238 | ||
c69899f0 CZ |
6239 | tree pcvoid_type_node |
6240 | = build_pointer_type (build_qualified_type (void_type_node, | |
6241 | TYPE_QUAL_CONST)); | |
6242 | tree V8HI_type_node = build_vector_type_for_mode (intHI_type_node, | |
6243 | V8HImode); | |
6244 | ||
6245 | tree void_ftype_void | |
6246 | = build_function_type_list (void_type_node, NULL_TREE); | |
6247 | tree int_ftype_int | |
6248 | = build_function_type_list (integer_type_node, integer_type_node, | |
6249 | NULL_TREE); | |
6250 | tree int_ftype_pcvoid_int | |
6251 | = build_function_type_list (integer_type_node, pcvoid_type_node, | |
6252 | integer_type_node, NULL_TREE); | |
6253 | tree void_ftype_usint_usint | |
6254 | = build_function_type_list (void_type_node, long_unsigned_type_node, | |
6255 | long_unsigned_type_node, NULL_TREE); | |
6256 | tree int_ftype_int_int | |
6257 | = build_function_type_list (integer_type_node, integer_type_node, | |
6258 | integer_type_node, NULL_TREE); | |
6259 | tree usint_ftype_usint | |
6260 | = build_function_type_list (long_unsigned_type_node, | |
6261 | long_unsigned_type_node, NULL_TREE); | |
6262 | tree void_ftype_usint | |
6263 | = build_function_type_list (void_type_node, long_unsigned_type_node, | |
6264 | NULL_TREE); | |
6265 | tree int_ftype_void | |
6266 | = build_function_type_list (integer_type_node, void_type_node, | |
6267 | NULL_TREE); | |
6268 | tree void_ftype_int | |
6269 | = build_function_type_list (void_type_node, integer_type_node, | |
6270 | NULL_TREE); | |
6271 | tree int_ftype_short | |
6272 | = build_function_type_list (integer_type_node, short_integer_type_node, | |
6273 | NULL_TREE); | |
6274 | ||
6275 | /* Old ARC SIMD types. */ | |
6276 | tree v8hi_ftype_v8hi_v8hi | |
6277 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6278 | V8HI_type_node, NULL_TREE); | |
6279 | tree v8hi_ftype_v8hi_int | |
6280 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6281 | integer_type_node, NULL_TREE); | |
6282 | tree v8hi_ftype_v8hi_int_int | |
6283 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6284 | integer_type_node, integer_type_node, | |
6285 | NULL_TREE); | |
6286 | tree void_ftype_v8hi_int_int | |
6287 | = build_function_type_list (void_type_node, V8HI_type_node, | |
6288 | integer_type_node, integer_type_node, | |
6289 | NULL_TREE); | |
6290 | tree void_ftype_v8hi_int_int_int | |
6291 | = build_function_type_list (void_type_node, V8HI_type_node, | |
6292 | integer_type_node, integer_type_node, | |
6293 | integer_type_node, NULL_TREE); | |
6294 | tree v8hi_ftype_int_int | |
6295 | = build_function_type_list (V8HI_type_node, integer_type_node, | |
6296 | integer_type_node, NULL_TREE); | |
6297 | tree void_ftype_int_int | |
6298 | = build_function_type_list (void_type_node, integer_type_node, | |
6299 | integer_type_node, NULL_TREE); | |
6300 | tree v8hi_ftype_v8hi | |
6301 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6302 | NULL_TREE); | |
00c072ae CZ |
6303 | /* ARCv2 SIMD types. */ |
6304 | tree long_ftype_v4hi_v4hi | |
6305 | = build_function_type_list (long_long_integer_type_node, | |
6306 | V4HI_type_node, V4HI_type_node, NULL_TREE); | |
6307 | tree int_ftype_v2hi_v2hi | |
6308 | = build_function_type_list (integer_type_node, | |
6309 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6310 | tree v2si_ftype_v2hi_v2hi | |
6311 | = build_function_type_list (V2SI_type_node, | |
6312 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6313 | tree v2hi_ftype_v2hi_v2hi | |
6314 | = build_function_type_list (V2HI_type_node, | |
6315 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6316 | tree v2si_ftype_v2si_v2si | |
6317 | = build_function_type_list (V2SI_type_node, | |
6318 | V2SI_type_node, V2SI_type_node, NULL_TREE); | |
6319 | tree v4hi_ftype_v4hi_v4hi | |
6320 | = build_function_type_list (V4HI_type_node, | |
6321 | V4HI_type_node, V4HI_type_node, NULL_TREE); | |
6322 | tree long_ftype_v2si_v2hi | |
6323 | = build_function_type_list (long_long_integer_type_node, | |
6324 | V2SI_type_node, V2HI_type_node, NULL_TREE); | |
c69899f0 CZ |
6325 | |
6326 | /* Add the builtins. */ | |
6327 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6328 | { \ | |
6329 | int id = ARC_BUILTIN_ ## NAME; \ | |
6330 | const char *Name = "__builtin_arc_" #NAME; \ | |
6331 | char *name = (char*) alloca (1 + strlen (Name)); \ | |
6332 | \ | |
6333 | gcc_assert (id < ARC_BUILTIN_COUNT); \ | |
6334 | if (MASK) \ | |
6335 | arc_bdesc[id].fndecl \ | |
6336 | = add_builtin_function (arc_tolower(name, Name), TYPE, id, \ | |
6337 | BUILT_IN_MD, NULL, NULL_TREE); \ | |
6338 | } | |
6339 | #include "builtins.def" | |
6340 | #undef DEF_BUILTIN | |
6341 | } | |
6342 | ||
6343 | /* Helper to expand __builtin_arc_aligned (void* val, int | |
6344 | alignval). */ | |
6345 | ||
6346 | static rtx | |
6347 | arc_expand_builtin_aligned (tree exp) | |
6348 | { | |
6349 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
6350 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
6351 | fold (arg1); | |
6352 | rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
6353 | rtx op1 = expand_expr (arg1, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
6354 | ||
6355 | if (!CONST_INT_P (op1)) | |
6356 | { | |
6357 | /* If we can't fold the alignment to a constant integer | |
6358 | whilst optimizing, this is probably a user error. */ | |
6359 | if (optimize) | |
6360 | warning (0, "__builtin_arc_aligned with non-constant alignment"); | |
6361 | } | |
6362 | else | |
6363 | { | |
6364 | HOST_WIDE_INT alignTest = INTVAL (op1); | |
6365 | /* Check alignTest is positive, and a power of two. */ | |
6366 | if (alignTest <= 0 || alignTest != (alignTest & -alignTest)) | |
6367 | { | |
6368 | error ("invalid alignment value for __builtin_arc_aligned"); | |
6369 | return NULL_RTX; | |
6370 | } | |
6371 | ||
6372 | if (CONST_INT_P (op0)) | |
6373 | { | |
6374 | HOST_WIDE_INT pnt = INTVAL (op0); | |
6375 | ||
6376 | if ((pnt & (alignTest - 1)) == 0) | |
6377 | return const1_rtx; | |
6378 | } | |
6379 | else | |
6380 | { | |
6381 | unsigned align = get_pointer_alignment (arg0); | |
6382 | unsigned numBits = alignTest * BITS_PER_UNIT; | |
6383 | ||
6384 | if (align && align >= numBits) | |
6385 | return const1_rtx; | |
6386 | /* Another attempt to ascertain alignment. Check the type | |
6387 | we are pointing to. */ | |
6388 | if (POINTER_TYPE_P (TREE_TYPE (arg0)) | |
6389 | && TYPE_ALIGN (TREE_TYPE (TREE_TYPE (arg0))) >= numBits) | |
6390 | return const1_rtx; | |
6391 | } | |
6392 | } | |
6393 | ||
6394 | /* Default to false. */ | |
6395 | return const0_rtx; | |
6396 | } | |
6397 | ||
6398 | /* Helper arc_expand_builtin, generates a pattern for the given icode | |
6399 | and arguments. */ | |
6400 | ||
6401 | static rtx_insn * | |
6402 | apply_GEN_FCN (enum insn_code icode, rtx *arg) | |
6403 | { | |
6404 | switch (insn_data[icode].n_generator_args) | |
6405 | { | |
6406 | case 0: | |
6407 | return GEN_FCN (icode) (); | |
6408 | case 1: | |
6409 | return GEN_FCN (icode) (arg[0]); | |
6410 | case 2: | |
6411 | return GEN_FCN (icode) (arg[0], arg[1]); | |
6412 | case 3: | |
6413 | return GEN_FCN (icode) (arg[0], arg[1], arg[2]); | |
6414 | case 4: | |
6415 | return GEN_FCN (icode) (arg[0], arg[1], arg[2], arg[3]); | |
6416 | case 5: | |
6417 | return GEN_FCN (icode) (arg[0], arg[1], arg[2], arg[3], arg[4]); | |
6418 | default: | |
6419 | gcc_unreachable (); | |
6420 | } | |
6421 | } | |
526b7aee SV |
6422 | |
6423 | /* Expand an expression EXP that calls a built-in function, | |
6424 | with result going to TARGET if that's convenient | |
6425 | (and in mode MODE if that's convenient). | |
6426 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
6427 | IGNORE is nonzero if the value is to be ignored. */ | |
6428 | ||
6429 | static rtx | |
6430 | arc_expand_builtin (tree exp, | |
6431 | rtx target, | |
c69899f0 CZ |
6432 | rtx subtarget ATTRIBUTE_UNUSED, |
6433 | machine_mode mode ATTRIBUTE_UNUSED, | |
6434 | int ignore ATTRIBUTE_UNUSED) | |
6435 | { | |
6436 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
6437 | unsigned int id = DECL_FUNCTION_CODE (fndecl); | |
6438 | const struct arc_builtin_description *d = &arc_bdesc[id]; | |
6439 | int i, j, n_args = call_expr_nargs (exp); | |
6440 | rtx pat = NULL_RTX; | |
6441 | rtx xop[5]; | |
6442 | enum insn_code icode = d->icode; | |
6443 | machine_mode tmode = insn_data[icode].operand[0].mode; | |
6444 | int nonvoid; | |
6445 | tree arg0; | |
6446 | tree arg1; | |
6447 | tree arg2; | |
6448 | tree arg3; | |
6449 | rtx op0; | |
6450 | rtx op1; | |
6451 | rtx op2; | |
6452 | rtx op3; | |
6453 | rtx op4; | |
ef4bddc2 RS |
6454 | machine_mode mode0; |
6455 | machine_mode mode1; | |
c69899f0 CZ |
6456 | machine_mode mode2; |
6457 | machine_mode mode3; | |
6458 | machine_mode mode4; | |
526b7aee | 6459 | |
c69899f0 CZ |
6460 | if (id >= ARC_BUILTIN_COUNT) |
6461 | internal_error ("bad builtin fcode"); | |
526b7aee | 6462 | |
c69899f0 CZ |
6463 | /* 1st part: Expand special builtins. */ |
6464 | switch (id) | |
526b7aee SV |
6465 | { |
6466 | case ARC_BUILTIN_NOP: | |
c69899f0 | 6467 | emit_insn (gen_nopv ()); |
526b7aee SV |
6468 | return NULL_RTX; |
6469 | ||
c69899f0 CZ |
6470 | case ARC_BUILTIN_RTIE: |
6471 | case ARC_BUILTIN_SYNC: | |
6472 | case ARC_BUILTIN_BRK: | |
6473 | case ARC_BUILTIN_SWI: | |
6474 | case ARC_BUILTIN_UNIMP_S: | |
6475 | gcc_assert (icode != 0); | |
6476 | emit_insn (GEN_FCN (icode) (const1_rtx)); | |
6477 | return NULL_RTX; | |
526b7aee | 6478 | |
c69899f0 CZ |
6479 | case ARC_BUILTIN_ALIGNED: |
6480 | return arc_expand_builtin_aligned (exp); | |
526b7aee | 6481 | |
c69899f0 CZ |
6482 | case ARC_BUILTIN_CLRI: |
6483 | target = gen_reg_rtx (SImode); | |
6484 | emit_insn (gen_clri (target, const1_rtx)); | |
526b7aee SV |
6485 | return target; |
6486 | ||
c69899f0 CZ |
6487 | case ARC_BUILTIN_TRAP_S: |
6488 | case ARC_BUILTIN_SLEEP: | |
526b7aee | 6489 | arg0 = CALL_EXPR_ARG (exp, 0); |
c69899f0 | 6490 | fold (arg0); |
526b7aee | 6491 | op0 = expand_expr (arg0, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
526b7aee | 6492 | |
c69899f0 CZ |
6493 | if (!CONST_INT_P (op0) || !satisfies_constraint_L (op0)) |
6494 | { | |
6495 | error ("builtin operand should be an unsigned 6-bit value"); | |
6496 | return NULL_RTX; | |
6497 | } | |
6498 | gcc_assert (icode != 0); | |
6499 | emit_insn (GEN_FCN (icode) (op0)); | |
6500 | return NULL_RTX; | |
526b7aee | 6501 | |
c69899f0 CZ |
6502 | case ARC_BUILTIN_VDORUN: |
6503 | case ARC_BUILTIN_VDIRUN: | |
526b7aee SV |
6504 | arg0 = CALL_EXPR_ARG (exp, 0); |
6505 | arg1 = CALL_EXPR_ARG (exp, 1); | |
c69899f0 CZ |
6506 | op0 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); |
6507 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee | 6508 | |
c69899f0 CZ |
6509 | target = gen_rtx_REG (SImode, (id == ARC_BUILTIN_VDIRUN) ? 131 : 139); |
6510 | ||
6511 | mode0 = insn_data[icode].operand[1].mode; | |
6512 | mode1 = insn_data[icode].operand[2].mode; | |
526b7aee | 6513 | |
c69899f0 | 6514 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) |
526b7aee SV |
6515 | op0 = copy_to_mode_reg (mode0, op0); |
6516 | ||
c69899f0 | 6517 | if (!insn_data[icode].operand[2].predicate (op1, mode1)) |
526b7aee SV |
6518 | op1 = copy_to_mode_reg (mode1, op1); |
6519 | ||
c69899f0 CZ |
6520 | pat = GEN_FCN (icode) (target, op0, op1); |
6521 | if (!pat) | |
6522 | return NULL_RTX; | |
6523 | ||
6524 | emit_insn (pat); | |
526b7aee SV |
6525 | return NULL_RTX; |
6526 | ||
c69899f0 CZ |
6527 | case ARC_BUILTIN_VDIWR: |
6528 | case ARC_BUILTIN_VDOWR: | |
526b7aee SV |
6529 | arg0 = CALL_EXPR_ARG (exp, 0); |
6530 | arg1 = CALL_EXPR_ARG (exp, 1); | |
c69899f0 CZ |
6531 | op0 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); |
6532 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
6533 | ||
6534 | if (!CONST_INT_P (op0) | |
6535 | || !(UNSIGNED_INT3 (INTVAL (op0)))) | |
6536 | error ("operand 1 should be an unsigned 3-bit immediate"); | |
526b7aee | 6537 | |
526b7aee SV |
6538 | mode1 = insn_data[icode].operand[1].mode; |
6539 | ||
c69899f0 CZ |
6540 | if (icode == CODE_FOR_vdiwr_insn) |
6541 | target = gen_rtx_REG (SImode, | |
6542 | ARC_FIRST_SIMD_DMA_CONFIG_IN_REG + INTVAL (op0)); | |
6543 | else if (icode == CODE_FOR_vdowr_insn) | |
6544 | target = gen_rtx_REG (SImode, | |
6545 | ARC_FIRST_SIMD_DMA_CONFIG_OUT_REG + INTVAL (op0)); | |
6546 | else | |
6547 | gcc_unreachable (); | |
526b7aee | 6548 | |
c69899f0 | 6549 | if (!insn_data[icode].operand[2].predicate (op1, mode1)) |
526b7aee SV |
6550 | op1 = copy_to_mode_reg (mode1, op1); |
6551 | ||
c69899f0 CZ |
6552 | pat = GEN_FCN (icode) (target, op1); |
6553 | if (!pat) | |
6554 | return NULL_RTX; | |
526b7aee | 6555 | |
c69899f0 | 6556 | emit_insn (pat); |
526b7aee SV |
6557 | return NULL_RTX; |
6558 | ||
c69899f0 CZ |
6559 | case ARC_BUILTIN_VASRW: |
6560 | case ARC_BUILTIN_VSR8: | |
6561 | case ARC_BUILTIN_VSR8AW: | |
526b7aee | 6562 | arg0 = CALL_EXPR_ARG (exp, 0); |
c69899f0 CZ |
6563 | arg1 = CALL_EXPR_ARG (exp, 1); |
6564 | op0 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
6565 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
6566 | op2 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
6567 | ||
6568 | target = gen_reg_rtx (V8HImode); | |
526b7aee | 6569 | mode0 = insn_data[icode].operand[1].mode; |
c69899f0 | 6570 | mode1 = insn_data[icode].operand[2].mode; |
526b7aee | 6571 | |
c69899f0 | 6572 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) |
526b7aee SV |
6573 | op0 = copy_to_mode_reg (mode0, op0); |
6574 | ||
c69899f0 CZ |
6575 | if ((!insn_data[icode].operand[2].predicate (op1, mode1)) |
6576 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
6577 | error ("operand 2 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 6578 | |
c69899f0 CZ |
6579 | pat = GEN_FCN (icode) (target, op0, op1, op2); |
6580 | if (!pat) | |
6581 | return NULL_RTX; | |
526b7aee | 6582 | |
c69899f0 CZ |
6583 | emit_insn (pat); |
6584 | return target; | |
526b7aee | 6585 | |
c69899f0 CZ |
6586 | case ARC_BUILTIN_VLD32WH: |
6587 | case ARC_BUILTIN_VLD32WL: | |
6588 | case ARC_BUILTIN_VLD64: | |
6589 | case ARC_BUILTIN_VLD32: | |
6590 | rtx src_vreg; | |
6591 | icode = d->icode; | |
6592 | arg0 = CALL_EXPR_ARG (exp, 0); /* source vreg. */ | |
6593 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
6594 | arg2 = CALL_EXPR_ARG (exp, 2); /* u8. */ | |
526b7aee | 6595 | |
c69899f0 CZ |
6596 | src_vreg = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); |
6597 | op0 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
6598 | op1 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
6599 | op2 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
526b7aee | 6600 | |
c69899f0 CZ |
6601 | /* target <- src vreg. */ |
6602 | emit_insn (gen_move_insn (target, src_vreg)); | |
526b7aee | 6603 | |
c69899f0 CZ |
6604 | /* target <- vec_concat: target, mem (Ib, u8). */ |
6605 | mode0 = insn_data[icode].operand[3].mode; | |
6606 | mode1 = insn_data[icode].operand[1].mode; | |
526b7aee | 6607 | |
c69899f0 CZ |
6608 | if ((!insn_data[icode].operand[3].predicate (op0, mode0)) |
6609 | || !(UNSIGNED_INT3 (INTVAL (op0)))) | |
6610 | error ("operand 1 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 6611 | |
c69899f0 CZ |
6612 | if ((!insn_data[icode].operand[1].predicate (op1, mode1)) |
6613 | || !(UNSIGNED_INT8 (INTVAL (op1)))) | |
6614 | error ("operand 2 should be an unsigned 8-bit value"); | |
526b7aee | 6615 | |
c69899f0 CZ |
6616 | pat = GEN_FCN (icode) (target, op1, op2, op0); |
6617 | if (!pat) | |
6618 | return NULL_RTX; | |
526b7aee | 6619 | |
c69899f0 CZ |
6620 | emit_insn (pat); |
6621 | return target; | |
526b7aee | 6622 | |
c69899f0 CZ |
6623 | case ARC_BUILTIN_VLD64W: |
6624 | case ARC_BUILTIN_VLD128: | |
6625 | arg0 = CALL_EXPR_ARG (exp, 0); /* dest vreg. */ | |
6626 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
526b7aee | 6627 | |
c69899f0 CZ |
6628 | op0 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); |
6629 | op1 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); | |
6630 | op2 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee | 6631 | |
c69899f0 CZ |
6632 | /* target <- src vreg. */ |
6633 | target = gen_reg_rtx (V8HImode); | |
526b7aee | 6634 | |
c69899f0 CZ |
6635 | /* target <- vec_concat: target, mem (Ib, u8). */ |
6636 | mode0 = insn_data[icode].operand[1].mode; | |
6637 | mode1 = insn_data[icode].operand[2].mode; | |
6638 | mode2 = insn_data[icode].operand[3].mode; | |
526b7aee | 6639 | |
c69899f0 CZ |
6640 | if ((!insn_data[icode].operand[2].predicate (op1, mode1)) |
6641 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
6642 | error ("operand 1 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 6643 | |
c69899f0 CZ |
6644 | if ((!insn_data[icode].operand[3].predicate (op2, mode2)) |
6645 | || !(UNSIGNED_INT8 (INTVAL (op2)))) | |
6646 | error ("operand 2 should be an unsigned 8-bit value"); | |
526b7aee | 6647 | |
c69899f0 | 6648 | pat = GEN_FCN (icode) (target, op0, op1, op2); |
526b7aee | 6649 | |
c69899f0 CZ |
6650 | if (!pat) |
6651 | return NULL_RTX; | |
526b7aee | 6652 | |
c69899f0 | 6653 | emit_insn (pat); |
526b7aee SV |
6654 | return target; |
6655 | ||
c69899f0 CZ |
6656 | case ARC_BUILTIN_VST128: |
6657 | case ARC_BUILTIN_VST64: | |
6658 | arg0 = CALL_EXPR_ARG (exp, 0); /* src vreg. */ | |
6659 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
6660 | arg2 = CALL_EXPR_ARG (exp, 2); /* u8. */ | |
526b7aee | 6661 | |
c69899f0 CZ |
6662 | op0 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); |
6663 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
6664 | op2 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
6665 | op3 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
526b7aee SV |
6666 | |
6667 | mode0 = insn_data[icode].operand[0].mode; | |
6668 | mode1 = insn_data[icode].operand[1].mode; | |
c69899f0 CZ |
6669 | mode2 = insn_data[icode].operand[2].mode; |
6670 | mode3 = insn_data[icode].operand[3].mode; | |
526b7aee | 6671 | |
c69899f0 CZ |
6672 | if ((!insn_data[icode].operand[1].predicate (op1, mode1)) |
6673 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
6674 | error ("operand 2 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 6675 | |
c69899f0 CZ |
6676 | if ((!insn_data[icode].operand[2].predicate (op2, mode2)) |
6677 | || !(UNSIGNED_INT8 (INTVAL (op2)))) | |
6678 | error ("operand 3 should be an unsigned 8-bit value"); | |
526b7aee | 6679 | |
c69899f0 CZ |
6680 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) |
6681 | op3 = copy_to_mode_reg (mode3, op3); | |
526b7aee | 6682 | |
c69899f0 CZ |
6683 | pat = GEN_FCN (icode) (op0, op1, op2, op3); |
6684 | if (!pat) | |
6685 | return NULL_RTX; | |
526b7aee | 6686 | |
c69899f0 CZ |
6687 | emit_insn (pat); |
6688 | return NULL_RTX; | |
526b7aee | 6689 | |
c69899f0 CZ |
6690 | case ARC_BUILTIN_VST16_N: |
6691 | case ARC_BUILTIN_VST32_N: | |
6692 | arg0 = CALL_EXPR_ARG (exp, 0); /* source vreg. */ | |
6693 | arg1 = CALL_EXPR_ARG (exp, 1); /* u3. */ | |
6694 | arg2 = CALL_EXPR_ARG (exp, 2); /* [I]0-7. */ | |
6695 | arg3 = CALL_EXPR_ARG (exp, 3); /* u8. */ | |
526b7aee | 6696 | |
c69899f0 CZ |
6697 | op0 = expand_expr (arg3, NULL_RTX, SImode, EXPAND_NORMAL); |
6698 | op1 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
6699 | op2 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
6700 | op3 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
6701 | op4 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee SV |
6702 | |
6703 | mode0 = insn_data[icode].operand[0].mode; | |
c69899f0 CZ |
6704 | mode2 = insn_data[icode].operand[2].mode; |
6705 | mode3 = insn_data[icode].operand[3].mode; | |
6706 | mode4 = insn_data[icode].operand[4].mode; | |
526b7aee | 6707 | |
c69899f0 CZ |
6708 | /* Do some correctness checks for the operands. */ |
6709 | if ((!insn_data[icode].operand[0].predicate (op0, mode0)) | |
6710 | || !(UNSIGNED_INT8 (INTVAL (op0)))) | |
6711 | error ("operand 4 should be an unsigned 8-bit value (0-255)"); | |
526b7aee | 6712 | |
c69899f0 CZ |
6713 | if ((!insn_data[icode].operand[2].predicate (op2, mode2)) |
6714 | || !(UNSIGNED_INT3 (INTVAL (op2)))) | |
6715 | error ("operand 3 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 6716 | |
c69899f0 CZ |
6717 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) |
6718 | op3 = copy_to_mode_reg (mode3, op3); | |
526b7aee | 6719 | |
c69899f0 CZ |
6720 | if ((!insn_data[icode].operand[4].predicate (op4, mode4)) |
6721 | || !(UNSIGNED_INT3 (INTVAL (op4)))) | |
6722 | error ("operand 2 should be an unsigned 3-bit value (subreg 0-7)"); | |
6723 | else if (icode == CODE_FOR_vst32_n_insn | |
6724 | && ((INTVAL (op4) % 2) != 0)) | |
6725 | error ("operand 2 should be an even 3-bit value (subreg 0,2,4,6)"); | |
526b7aee | 6726 | |
c69899f0 CZ |
6727 | pat = GEN_FCN (icode) (op0, op1, op2, op3, op4); |
6728 | if (!pat) | |
6729 | return NULL_RTX; | |
526b7aee | 6730 | |
c69899f0 | 6731 | emit_insn (pat); |
526b7aee SV |
6732 | return NULL_RTX; |
6733 | ||
c69899f0 CZ |
6734 | default: |
6735 | break; | |
6736 | } | |
6737 | ||
6738 | /* 2nd part: Expand regular builtins. */ | |
6739 | if (icode == 0) | |
6740 | internal_error ("bad builtin fcode"); | |
6741 | ||
6742 | nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node; | |
6743 | j = 0; | |
526b7aee | 6744 | |
c69899f0 CZ |
6745 | if (nonvoid) |
6746 | { | |
6747 | if (target == NULL_RTX | |
6748 | || GET_MODE (target) != tmode | |
6749 | || !insn_data[icode].operand[0].predicate (target, tmode)) | |
526b7aee | 6750 | { |
c69899f0 | 6751 | target = gen_reg_rtx (tmode); |
526b7aee | 6752 | } |
c69899f0 CZ |
6753 | xop[j++] = target; |
6754 | } | |
6755 | ||
6756 | gcc_assert (n_args <= 4); | |
6757 | for (i = 0; i < n_args; i++, j++) | |
6758 | { | |
6759 | tree arg = CALL_EXPR_ARG (exp, i); | |
6760 | machine_mode mode = insn_data[icode].operand[j].mode; | |
6761 | rtx op = expand_expr (arg, NULL_RTX, mode, EXPAND_NORMAL); | |
6762 | machine_mode opmode = GET_MODE (op); | |
6763 | char c = insn_data[icode].operand[j].constraint[0]; | |
6764 | ||
6765 | /* SIMD extension requires exact immediate operand match. */ | |
6766 | if ((id > ARC_BUILTIN_SIMD_BEGIN) | |
6767 | && (id < ARC_BUILTIN_SIMD_END) | |
6768 | && (c != 'v') | |
6769 | && (c != 'r')) | |
526b7aee | 6770 | { |
c69899f0 CZ |
6771 | if (!CONST_INT_P (op)) |
6772 | error ("builtin requires an immediate for operand %d", j); | |
6773 | switch (c) | |
526b7aee | 6774 | { |
c69899f0 CZ |
6775 | case 'L': |
6776 | if (!satisfies_constraint_L (op)) | |
6777 | error ("operand %d should be a 6 bit unsigned immediate", j); | |
6778 | break; | |
6779 | case 'P': | |
6780 | if (!satisfies_constraint_P (op)) | |
6781 | error ("operand %d should be a 8 bit unsigned immediate", j); | |
6782 | break; | |
6783 | case 'K': | |
6784 | if (!satisfies_constraint_K (op)) | |
6785 | error ("operand %d should be a 3 bit unsigned immediate", j); | |
6786 | break; | |
6787 | default: | |
6788 | error ("unknown builtin immediate operand type for operand %d", | |
6789 | j); | |
526b7aee | 6790 | } |
c69899f0 | 6791 | } |
526b7aee | 6792 | |
c69899f0 CZ |
6793 | if (CONST_INT_P (op)) |
6794 | opmode = mode; | |
526b7aee | 6795 | |
c69899f0 CZ |
6796 | if ((opmode == SImode) && (mode == HImode)) |
6797 | { | |
6798 | opmode = HImode; | |
6799 | op = gen_lowpart (HImode, op); | |
526b7aee SV |
6800 | } |
6801 | ||
c69899f0 CZ |
6802 | /* In case the insn wants input operands in modes different from |
6803 | the result, abort. */ | |
6804 | gcc_assert (opmode == mode || opmode == VOIDmode); | |
526b7aee | 6805 | |
c69899f0 CZ |
6806 | if (!insn_data[icode].operand[i + nonvoid].predicate (op, mode)) |
6807 | op = copy_to_mode_reg (mode, op); | |
6808 | ||
6809 | xop[j] = op; | |
526b7aee SV |
6810 | } |
6811 | ||
c69899f0 CZ |
6812 | pat = apply_GEN_FCN (icode, xop); |
6813 | if (pat == NULL_RTX) | |
6814 | return NULL_RTX; | |
6815 | ||
6816 | emit_insn (pat); | |
6817 | ||
6818 | if (nonvoid) | |
6819 | return target; | |
6820 | else | |
6821 | return const0_rtx; | |
526b7aee SV |
6822 | } |
6823 | ||
6824 | /* Returns true if the operands[opno] is a valid compile-time constant to be | |
6825 | used as register number in the code for builtins. Else it flags an error | |
6826 | and returns false. */ | |
6827 | ||
6828 | bool | |
6829 | check_if_valid_regno_const (rtx *operands, int opno) | |
6830 | { | |
6831 | ||
6832 | switch (GET_CODE (operands[opno])) | |
6833 | { | |
6834 | case SYMBOL_REF : | |
6835 | case CONST : | |
6836 | case CONST_INT : | |
6837 | return true; | |
6838 | default: | |
6839 | error ("register number must be a compile-time constant. Try giving higher optimization levels"); | |
6840 | break; | |
6841 | } | |
6842 | return false; | |
6843 | } | |
6844 | ||
6845 | /* Check that after all the constant folding, whether the operand to | |
6846 | __builtin_arc_sleep is an unsigned int of 6 bits. If not, flag an error. */ | |
6847 | ||
6848 | bool | |
6849 | check_if_valid_sleep_operand (rtx *operands, int opno) | |
6850 | { | |
6851 | switch (GET_CODE (operands[opno])) | |
6852 | { | |
6853 | case CONST : | |
6854 | case CONST_INT : | |
6855 | if( UNSIGNED_INT6 (INTVAL (operands[opno]))) | |
6856 | return true; | |
3bbe0b82 | 6857 | /* FALLTHRU */ |
526b7aee | 6858 | default: |
40fecdd6 JM |
6859 | fatal_error (input_location, |
6860 | "operand for sleep instruction must be an unsigned 6 bit compile-time constant"); | |
526b7aee SV |
6861 | break; |
6862 | } | |
6863 | return false; | |
6864 | } | |
6865 | ||
6866 | /* Return true if it is ok to make a tail-call to DECL. */ | |
6867 | ||
6868 | static bool | |
6869 | arc_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, | |
6870 | tree exp ATTRIBUTE_UNUSED) | |
6871 | { | |
6872 | /* Never tailcall from an ISR routine - it needs a special exit sequence. */ | |
6873 | if (ARC_INTERRUPT_P (arc_compute_function_type (cfun))) | |
6874 | return false; | |
6875 | ||
6876 | /* Everything else is ok. */ | |
6877 | return true; | |
6878 | } | |
6879 | ||
6880 | /* Output code to add DELTA to the first argument, and then jump | |
6881 | to FUNCTION. Used for C++ multiple inheritance. */ | |
6882 | ||
6883 | static void | |
6884 | arc_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
6885 | HOST_WIDE_INT delta, | |
6886 | HOST_WIDE_INT vcall_offset, | |
6887 | tree function) | |
6888 | { | |
6889 | int mi_delta = delta; | |
6890 | const char *const mi_op = mi_delta < 0 ? "sub" : "add"; | |
6891 | int shift = 0; | |
6892 | int this_regno | |
6893 | = aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) ? 1 : 0; | |
6894 | rtx fnaddr; | |
6895 | ||
6896 | if (mi_delta < 0) | |
6897 | mi_delta = - mi_delta; | |
6898 | ||
6899 | /* Add DELTA. When possible use a plain add, otherwise load it into | |
6900 | a register first. */ | |
6901 | ||
6902 | while (mi_delta != 0) | |
6903 | { | |
6904 | if ((mi_delta & (3 << shift)) == 0) | |
6905 | shift += 2; | |
6906 | else | |
6907 | { | |
6908 | asm_fprintf (file, "\t%s\t%s, %s, %d\n", | |
6909 | mi_op, reg_names[this_regno], reg_names[this_regno], | |
6910 | mi_delta & (0xff << shift)); | |
6911 | mi_delta &= ~(0xff << shift); | |
6912 | shift += 8; | |
6913 | } | |
6914 | } | |
6915 | ||
6916 | /* If needed, add *(*THIS + VCALL_OFFSET) to THIS. */ | |
6917 | if (vcall_offset != 0) | |
6918 | { | |
6919 | /* ld r12,[this] --> temp = *this | |
6920 | add r12,r12,vcall_offset --> temp = *(*this + vcall_offset) | |
6921 | ld r12,[r12] | |
6922 | add this,this,r12 --> this+ = *(*this + vcall_offset) */ | |
6923 | asm_fprintf (file, "\tld\t%s, [%s]\n", | |
6924 | ARC_TEMP_SCRATCH_REG, reg_names[this_regno]); | |
dfca07ea | 6925 | asm_fprintf (file, "\tadd\t%s, %s, " HOST_WIDE_INT_PRINT_DEC "\n", |
526b7aee SV |
6926 | ARC_TEMP_SCRATCH_REG, ARC_TEMP_SCRATCH_REG, vcall_offset); |
6927 | asm_fprintf (file, "\tld\t%s, [%s]\n", | |
6928 | ARC_TEMP_SCRATCH_REG, ARC_TEMP_SCRATCH_REG); | |
6929 | asm_fprintf (file, "\tadd\t%s, %s, %s\n", reg_names[this_regno], | |
6930 | reg_names[this_regno], ARC_TEMP_SCRATCH_REG); | |
6931 | } | |
6932 | ||
6933 | fnaddr = XEXP (DECL_RTL (function), 0); | |
6934 | ||
6935 | if (arc_is_longcall_p (fnaddr)) | |
1f8876c7 CZ |
6936 | { |
6937 | if (flag_pic) | |
6938 | { | |
6939 | asm_fprintf (file, "\tld\t%s, [pcl, @", | |
6940 | ARC_TEMP_SCRATCH_REG); | |
6941 | assemble_name (file, XSTR (fnaddr, 0)); | |
6942 | fputs ("@gotpc]\n", file); | |
6943 | asm_fprintf (file, "\tj\t[%s]", ARC_TEMP_SCRATCH_REG); | |
6944 | } | |
6945 | else | |
6946 | { | |
6947 | fputs ("\tj\t@", file); | |
6948 | assemble_name (file, XSTR (fnaddr, 0)); | |
6949 | } | |
6950 | } | |
526b7aee | 6951 | else |
1f8876c7 CZ |
6952 | { |
6953 | fputs ("\tb\t@", file); | |
6954 | assemble_name (file, XSTR (fnaddr, 0)); | |
6955 | if (flag_pic) | |
6956 | fputs ("@plt\n", file); | |
6957 | } | |
526b7aee SV |
6958 | fputc ('\n', file); |
6959 | } | |
6960 | ||
6961 | /* Return true if a 32 bit "long_call" should be generated for | |
6962 | this calling SYM_REF. We generate a long_call if the function: | |
6963 | ||
6964 | a. has an __attribute__((long call)) | |
6965 | or b. the -mlong-calls command line switch has been specified | |
6966 | ||
6967 | However we do not generate a long call if the function has an | |
6968 | __attribute__ ((short_call)) or __attribute__ ((medium_call)) | |
6969 | ||
6970 | This function will be called by C fragments contained in the machine | |
6971 | description file. */ | |
6972 | ||
6973 | bool | |
6974 | arc_is_longcall_p (rtx sym_ref) | |
6975 | { | |
6976 | if (GET_CODE (sym_ref) != SYMBOL_REF) | |
6977 | return false; | |
6978 | ||
6979 | return (SYMBOL_REF_LONG_CALL_P (sym_ref) | |
6980 | || (TARGET_LONG_CALLS_SET | |
6981 | && !SYMBOL_REF_SHORT_CALL_P (sym_ref) | |
6982 | && !SYMBOL_REF_MEDIUM_CALL_P (sym_ref))); | |
6983 | ||
6984 | } | |
6985 | ||
6986 | /* Likewise for short calls. */ | |
6987 | ||
6988 | bool | |
6989 | arc_is_shortcall_p (rtx sym_ref) | |
6990 | { | |
6991 | if (GET_CODE (sym_ref) != SYMBOL_REF) | |
6992 | return false; | |
6993 | ||
6994 | return (SYMBOL_REF_SHORT_CALL_P (sym_ref) | |
6995 | || (!TARGET_LONG_CALLS_SET && !TARGET_MEDIUM_CALLS | |
6996 | && !SYMBOL_REF_LONG_CALL_P (sym_ref) | |
6997 | && !SYMBOL_REF_MEDIUM_CALL_P (sym_ref))); | |
6998 | ||
6999 | } | |
7000 | ||
526b7aee SV |
7001 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
7002 | ||
7003 | static bool | |
7004 | arc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) | |
7005 | { | |
7006 | if (AGGREGATE_TYPE_P (type) || TREE_ADDRESSABLE (type)) | |
7007 | return true; | |
7008 | else | |
7009 | { | |
7010 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
f50bb868 | 7011 | return (size == -1 || size > (TARGET_V2 ? 16 : 8)); |
526b7aee SV |
7012 | } |
7013 | } | |
7014 | ||
7015 | ||
7016 | /* This was in rtlanal.c, and can go in there when we decide we want | |
7017 | to submit the change for inclusion in the GCC tree. */ | |
7018 | /* Like note_stores, but allow the callback to have side effects on the rtl | |
7019 | (like the note_stores of yore): | |
7020 | Call FUN on each register or MEM that is stored into or clobbered by X. | |
7021 | (X would be the pattern of an insn). DATA is an arbitrary pointer, | |
7022 | ignored by note_stores, but passed to FUN. | |
7023 | FUN may alter parts of the RTL. | |
7024 | ||
7025 | FUN receives three arguments: | |
7026 | 1. the REG, MEM, CC0 or PC being stored in or clobbered, | |
7027 | 2. the SET or CLOBBER rtx that does the store, | |
7028 | 3. the pointer DATA provided to note_stores. | |
7029 | ||
7030 | If the item being stored in or clobbered is a SUBREG of a hard register, | |
7031 | the SUBREG will be passed. */ | |
7032 | ||
7033 | /* For now. */ static | |
7034 | void | |
7035 | walk_stores (rtx x, void (*fun) (rtx, rtx, void *), void *data) | |
7036 | { | |
7037 | int i; | |
7038 | ||
7039 | if (GET_CODE (x) == COND_EXEC) | |
7040 | x = COND_EXEC_CODE (x); | |
7041 | ||
7042 | if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER) | |
7043 | { | |
7044 | rtx dest = SET_DEST (x); | |
7045 | ||
7046 | while ((GET_CODE (dest) == SUBREG | |
7047 | && (!REG_P (SUBREG_REG (dest)) | |
7048 | || REGNO (SUBREG_REG (dest)) >= FIRST_PSEUDO_REGISTER)) | |
7049 | || GET_CODE (dest) == ZERO_EXTRACT | |
7050 | || GET_CODE (dest) == STRICT_LOW_PART) | |
7051 | dest = XEXP (dest, 0); | |
7052 | ||
7053 | /* If we have a PARALLEL, SET_DEST is a list of EXPR_LIST expressions, | |
7054 | each of whose first operand is a register. */ | |
7055 | if (GET_CODE (dest) == PARALLEL) | |
7056 | { | |
7057 | for (i = XVECLEN (dest, 0) - 1; i >= 0; i--) | |
7058 | if (XEXP (XVECEXP (dest, 0, i), 0) != 0) | |
7059 | (*fun) (XEXP (XVECEXP (dest, 0, i), 0), x, data); | |
7060 | } | |
7061 | else | |
7062 | (*fun) (dest, x, data); | |
7063 | } | |
7064 | ||
7065 | else if (GET_CODE (x) == PARALLEL) | |
7066 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
7067 | walk_stores (XVECEXP (x, 0, i), fun, data); | |
7068 | } | |
7069 | ||
7070 | static bool | |
7071 | arc_pass_by_reference (cumulative_args_t ca_v ATTRIBUTE_UNUSED, | |
ef4bddc2 | 7072 | machine_mode mode ATTRIBUTE_UNUSED, |
526b7aee SV |
7073 | const_tree type, |
7074 | bool named ATTRIBUTE_UNUSED) | |
7075 | { | |
7076 | return (type != 0 | |
7077 | && (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST | |
7078 | || TREE_ADDRESSABLE (type))); | |
7079 | } | |
7080 | ||
1d0216c8 RS |
7081 | /* Implement TARGET_CAN_USE_DOLOOP_P. */ |
7082 | ||
7083 | static bool | |
a2de90a4 CZ |
7084 | arc_can_use_doloop_p (const widest_int &, |
7085 | const widest_int &iterations_max, | |
1d0216c8 RS |
7086 | unsigned int loop_depth, bool entered_at_top) |
7087 | { | |
a2de90a4 CZ |
7088 | /* Considering limitations in the hardware, only use doloop |
7089 | for innermost loops which must be entered from the top. */ | |
7090 | if (loop_depth > 1 || !entered_at_top) | |
1d0216c8 | 7091 | return false; |
a2de90a4 CZ |
7092 | |
7093 | /* Check for lp_count width boundary. */ | |
7094 | if (arc_lpcwidth != 32 | |
7095 | && (wi::gtu_p (iterations_max, ((1 << arc_lpcwidth) - 1)) | |
7096 | || wi::eq_p (iterations_max, 0))) | |
1d0216c8 RS |
7097 | return false; |
7098 | return true; | |
7099 | } | |
526b7aee | 7100 | |
a2de90a4 CZ |
7101 | /* NULL if INSN insn is valid within a low-overhead loop. Otherwise |
7102 | return why doloop cannot be applied. */ | |
526b7aee SV |
7103 | |
7104 | static const char * | |
ac44248e | 7105 | arc_invalid_within_doloop (const rtx_insn *insn) |
526b7aee SV |
7106 | { |
7107 | if (CALL_P (insn)) | |
7108 | return "Function call in the loop."; | |
a2de90a4 CZ |
7109 | |
7110 | /* FIXME! add here all the ZOL exceptions. */ | |
526b7aee SV |
7111 | return NULL; |
7112 | } | |
7113 | ||
e9472c81 AB |
7114 | /* Return true if a load instruction (CONSUMER) uses the same address as a |
7115 | store instruction (PRODUCER). This function is used to avoid st/ld | |
7116 | address hazard in ARC700 cores. */ | |
7117 | bool | |
7118 | arc_store_addr_hazard_p (rtx_insn* producer, rtx_insn* consumer) | |
7119 | { | |
7120 | rtx in_set, out_set; | |
7121 | rtx out_addr, in_addr; | |
7122 | ||
7123 | if (!producer) | |
7124 | return false; | |
7125 | ||
7126 | if (!consumer) | |
7127 | return false; | |
7128 | ||
7129 | /* Peel the producer and the consumer for the address. */ | |
7130 | out_set = single_set (producer); | |
7131 | if (out_set) | |
7132 | { | |
7133 | out_addr = SET_DEST (out_set); | |
7134 | if (!out_addr) | |
7135 | return false; | |
7136 | if (GET_CODE (out_addr) == ZERO_EXTEND | |
7137 | || GET_CODE (out_addr) == SIGN_EXTEND) | |
7138 | out_addr = XEXP (out_addr, 0); | |
7139 | ||
7140 | if (!MEM_P (out_addr)) | |
7141 | return false; | |
7142 | ||
7143 | in_set = single_set (consumer); | |
7144 | if (in_set) | |
7145 | { | |
7146 | in_addr = SET_SRC (in_set); | |
7147 | if (!in_addr) | |
7148 | return false; | |
7149 | if (GET_CODE (in_addr) == ZERO_EXTEND | |
7150 | || GET_CODE (in_addr) == SIGN_EXTEND) | |
7151 | in_addr = XEXP (in_addr, 0); | |
7152 | ||
7153 | if (!MEM_P (in_addr)) | |
7154 | return false; | |
7155 | /* Get rid of the MEM and check if the addresses are | |
7156 | equivalent. */ | |
7157 | in_addr = XEXP (in_addr, 0); | |
7158 | out_addr = XEXP (out_addr, 0); | |
7159 | ||
7160 | return exp_equiv_p (in_addr, out_addr, 0, true); | |
7161 | } | |
7162 | } | |
7163 | return false; | |
7164 | } | |
7165 | ||
f50bb868 CZ |
7166 | /* The same functionality as arc_hazard. It is called in machine |
7167 | reorg before any other optimization. Hence, the NOP size is taken | |
7168 | into account when doing branch shortening. */ | |
7169 | ||
7170 | static void | |
7171 | workaround_arc_anomaly (void) | |
7172 | { | |
7173 | rtx_insn *insn, *succ0; | |
7174 | ||
7175 | /* For any architecture: call arc_hazard here. */ | |
7176 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
7177 | { | |
7178 | succ0 = next_real_insn (insn); | |
7179 | if (arc_hazard (insn, succ0)) | |
7180 | { | |
7181 | emit_insn_before (gen_nopv (), succ0); | |
7182 | } | |
7183 | } | |
e9472c81 AB |
7184 | |
7185 | if (TARGET_ARC700) | |
7186 | { | |
7187 | rtx_insn *succ1; | |
7188 | ||
7189 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
7190 | { | |
7191 | succ0 = next_real_insn (insn); | |
7192 | if (arc_store_addr_hazard_p (insn, succ0)) | |
7193 | { | |
7194 | emit_insn_after (gen_nopv (), insn); | |
7195 | emit_insn_after (gen_nopv (), insn); | |
7196 | continue; | |
7197 | } | |
7198 | ||
7199 | /* Avoid adding nops if the instruction between the ST and LD is | |
7200 | a call or jump. */ | |
7201 | succ1 = next_real_insn (succ0); | |
7202 | if (succ0 && !JUMP_P (succ0) && !CALL_P (succ0) | |
7203 | && arc_store_addr_hazard_p (insn, succ1)) | |
7204 | emit_insn_after (gen_nopv (), insn); | |
7205 | } | |
7206 | } | |
f50bb868 CZ |
7207 | } |
7208 | ||
a2de90a4 CZ |
7209 | /* A callback for the hw-doloop pass. Called when a loop we have discovered |
7210 | turns out not to be optimizable; we have to split the loop_end pattern into | |
7211 | a subtract and a test. */ | |
7212 | ||
7213 | static void | |
7214 | hwloop_fail (hwloop_info loop) | |
7215 | { | |
7216 | rtx test; | |
7217 | rtx insn = loop->loop_end; | |
7218 | ||
7219 | if (TARGET_V2 | |
7220 | && (loop->length && (loop->length <= ARC_MAX_LOOP_LENGTH)) | |
7221 | && REG_P (loop->iter_reg)) | |
7222 | { | |
7223 | /* TARGET_V2 has dbnz instructions. */ | |
7224 | test = gen_dbnz (loop->iter_reg, loop->start_label); | |
7225 | insn = emit_jump_insn_before (test, loop->loop_end); | |
7226 | } | |
7227 | else if (REG_P (loop->iter_reg) && (REGNO (loop->iter_reg) == LP_COUNT)) | |
7228 | { | |
7229 | /* We have the lp_count as loop iterator, try to use it. */ | |
7230 | emit_insn_before (gen_loop_fail (), loop->loop_end); | |
7231 | test = gen_rtx_NE (VOIDmode, gen_rtx_REG (CC_ZNmode, CC_REG), | |
7232 | const0_rtx); | |
7233 | test = gen_rtx_IF_THEN_ELSE (VOIDmode, test, | |
7234 | gen_rtx_LABEL_REF (Pmode, loop->start_label), | |
7235 | pc_rtx); | |
7236 | insn = emit_jump_insn_before (gen_rtx_SET (pc_rtx, test), | |
7237 | loop->loop_end); | |
7238 | } | |
7239 | else | |
7240 | { | |
7241 | emit_insn_before (gen_addsi3 (loop->iter_reg, | |
7242 | loop->iter_reg, | |
7243 | constm1_rtx), | |
7244 | loop->loop_end); | |
7245 | test = gen_rtx_NE (VOIDmode, loop->iter_reg, const0_rtx); | |
7246 | insn = emit_jump_insn_before (gen_cbranchsi4 (test, | |
7247 | loop->iter_reg, | |
7248 | const0_rtx, | |
7249 | loop->start_label), | |
7250 | loop->loop_end); | |
7251 | } | |
7252 | JUMP_LABEL (insn) = loop->start_label; | |
7253 | LABEL_NUSES (loop->start_label)++; | |
7254 | delete_insn (loop->loop_end); | |
7255 | } | |
7256 | ||
7257 | /* Optimize LOOP. */ | |
7258 | ||
7259 | static bool | |
7260 | hwloop_optimize (hwloop_info loop) | |
7261 | { | |
7262 | int i; | |
7263 | edge entry_edge; | |
7264 | basic_block entry_bb, bb; | |
7265 | rtx iter_reg, end_label; | |
7266 | rtx_insn *insn, *seq, *entry_after, *last_insn; | |
7267 | unsigned int length; | |
7268 | bool need_fix = false; | |
7269 | rtx lp_reg = gen_rtx_REG (SImode, LP_COUNT); | |
7270 | ||
7271 | if (loop->depth > 1) | |
7272 | { | |
7273 | if (dump_file) | |
7274 | fprintf (dump_file, ";; loop %d is not innermost\n", | |
7275 | loop->loop_no); | |
7276 | return false; | |
7277 | } | |
7278 | ||
7279 | if (!loop->incoming_dest) | |
7280 | { | |
7281 | if (dump_file) | |
7282 | fprintf (dump_file, ";; loop %d has more than one entry\n", | |
7283 | loop->loop_no); | |
7284 | return false; | |
7285 | } | |
7286 | ||
7287 | if (loop->incoming_dest != loop->head) | |
7288 | { | |
7289 | if (dump_file) | |
7290 | fprintf (dump_file, ";; loop %d is not entered from head\n", | |
7291 | loop->loop_no); | |
7292 | return false; | |
7293 | } | |
7294 | ||
7295 | if (loop->has_call || loop->has_asm) | |
7296 | { | |
7297 | if (dump_file) | |
7298 | fprintf (dump_file, ";; loop %d has invalid insn\n", | |
7299 | loop->loop_no); | |
7300 | return false; | |
7301 | } | |
7302 | ||
7303 | /* Scan all the blocks to make sure they don't use iter_reg. */ | |
7304 | if (loop->iter_reg_used || loop->iter_reg_used_outside) | |
7305 | { | |
7306 | if (dump_file) | |
7307 | fprintf (dump_file, ";; loop %d uses iterator\n", | |
7308 | loop->loop_no); | |
7309 | return false; | |
7310 | } | |
7311 | ||
7312 | /* Check if start_label appears before doloop_end. */ | |
7313 | length = 0; | |
7314 | for (insn = loop->start_label; | |
7315 | insn && insn != loop->loop_end; | |
7316 | insn = NEXT_INSN (insn)) | |
7317 | length += NONDEBUG_INSN_P (insn) ? get_attr_length (insn) : 0; | |
7318 | ||
7319 | if (!insn) | |
7320 | { | |
7321 | if (dump_file) | |
7322 | fprintf (dump_file, ";; loop %d start_label not before loop_end\n", | |
7323 | loop->loop_no); | |
7324 | return false; | |
7325 | } | |
7326 | ||
7327 | loop->length = length; | |
7328 | if (loop->length > ARC_MAX_LOOP_LENGTH) | |
7329 | { | |
7330 | if (dump_file) | |
7331 | fprintf (dump_file, ";; loop %d too long\n", loop->loop_no); | |
7332 | return false; | |
7333 | } | |
5b5905bb CZ |
7334 | else if (!loop->length) |
7335 | { | |
7336 | if (dump_file) | |
7337 | fprintf (dump_file, ";; loop %d is empty\n", loop->loop_no); | |
7338 | return false; | |
7339 | } | |
a2de90a4 CZ |
7340 | |
7341 | /* Check if we use a register or not. */ | |
7342 | if (!REG_P (loop->iter_reg)) | |
7343 | { | |
7344 | if (dump_file) | |
7345 | fprintf (dump_file, ";; loop %d iterator is MEM\n", | |
7346 | loop->loop_no); | |
7347 | return false; | |
7348 | } | |
7349 | ||
7350 | /* Check if loop register is lpcount. */ | |
7351 | if (REG_P (loop->iter_reg) && (REGNO (loop->iter_reg)) != LP_COUNT) | |
7352 | { | |
7353 | if (dump_file) | |
7354 | fprintf (dump_file, ";; loop %d doesn't use lp_count as loop" | |
7355 | " iterator\n", | |
7356 | loop->loop_no); | |
7357 | /* This loop doesn't use the lp_count, check though if we can | |
7358 | fix it. */ | |
7359 | if (TEST_HARD_REG_BIT (loop->regs_set_in_loop, LP_COUNT) | |
7360 | /* In very unique cases we may have LP_COUNT alive. */ | |
7361 | || (loop->incoming_src | |
7362 | && REGNO_REG_SET_P (df_get_live_out (loop->incoming_src), | |
7363 | LP_COUNT))) | |
7364 | return false; | |
7365 | else | |
7366 | need_fix = true; | |
7367 | } | |
7368 | ||
7369 | /* Check for control like instruction as the last instruction of a | |
7370 | ZOL. */ | |
7371 | bb = loop->tail; | |
7372 | last_insn = PREV_INSN (loop->loop_end); | |
7373 | ||
7374 | while (1) | |
7375 | { | |
7376 | for (; last_insn != BB_HEAD (bb); | |
7377 | last_insn = PREV_INSN (last_insn)) | |
7378 | if (NONDEBUG_INSN_P (last_insn)) | |
7379 | break; | |
7380 | ||
7381 | if (last_insn != BB_HEAD (bb)) | |
7382 | break; | |
7383 | ||
7384 | if (single_pred_p (bb) | |
7385 | && single_pred_edge (bb)->flags & EDGE_FALLTHRU | |
7386 | && single_pred (bb) != ENTRY_BLOCK_PTR_FOR_FN (cfun)) | |
7387 | { | |
7388 | bb = single_pred (bb); | |
7389 | last_insn = BB_END (bb); | |
7390 | continue; | |
7391 | } | |
7392 | else | |
7393 | { | |
7394 | last_insn = NULL; | |
7395 | break; | |
7396 | } | |
7397 | } | |
7398 | ||
7399 | if (!last_insn) | |
7400 | { | |
7401 | if (dump_file) | |
7402 | fprintf (dump_file, ";; loop %d has no last instruction\n", | |
7403 | loop->loop_no); | |
7404 | return false; | |
7405 | } | |
7406 | ||
7407 | if ((TARGET_ARC600_FAMILY || TARGET_HS) | |
7408 | && INSN_P (last_insn) | |
7409 | && (JUMP_P (last_insn) || CALL_P (last_insn) | |
7410 | || GET_CODE (PATTERN (last_insn)) == SEQUENCE | |
5b5905bb CZ |
7411 | /* At this stage we can have (insn (clobber (mem:BLK |
7412 | (reg)))) instructions, ignore them. */ | |
7413 | || (GET_CODE (PATTERN (last_insn)) != CLOBBER | |
7414 | && (get_attr_type (last_insn) == TYPE_BRCC | |
7415 | || get_attr_type (last_insn) == TYPE_BRCC_NO_DELAY_SLOT)))) | |
a2de90a4 CZ |
7416 | { |
7417 | if (loop->length + 2 > ARC_MAX_LOOP_LENGTH) | |
7418 | { | |
7419 | if (dump_file) | |
7420 | fprintf (dump_file, ";; loop %d too long\n", loop->loop_no); | |
7421 | return false; | |
7422 | } | |
7423 | if (dump_file) | |
7424 | fprintf (dump_file, ";; loop %d has a control like last insn;" | |
7425 | "add a nop\n", | |
7426 | loop->loop_no); | |
7427 | ||
7428 | last_insn = emit_insn_after (gen_nopv (), last_insn); | |
7429 | } | |
7430 | ||
7431 | if (LABEL_P (last_insn)) | |
7432 | { | |
7433 | if (dump_file) | |
7434 | fprintf (dump_file, ";; loop %d has a label as last insn;" | |
7435 | "add a nop\n", | |
7436 | loop->loop_no); | |
7437 | last_insn = emit_insn_after (gen_nopv (), last_insn); | |
7438 | } | |
a0920243 CZ |
7439 | |
7440 | /* SAVE_NOTE is used by haifa scheduler. However, we are after it | |
7441 | and we can use it to indicate the last ZOL instruction cannot be | |
7442 | part of a delay slot. */ | |
7443 | add_reg_note (last_insn, REG_SAVE_NOTE, GEN_INT (2)); | |
7444 | ||
a2de90a4 CZ |
7445 | loop->last_insn = last_insn; |
7446 | ||
7447 | /* Get the loop iteration register. */ | |
7448 | iter_reg = loop->iter_reg; | |
7449 | ||
7450 | gcc_assert (REG_P (iter_reg)); | |
7451 | ||
7452 | entry_edge = NULL; | |
7453 | ||
7454 | FOR_EACH_VEC_SAFE_ELT (loop->incoming, i, entry_edge) | |
7455 | if (entry_edge->flags & EDGE_FALLTHRU) | |
7456 | break; | |
7457 | ||
7458 | if (entry_edge == NULL) | |
7459 | { | |
7460 | if (dump_file) | |
7461 | fprintf (dump_file, ";; loop %d has no fallthru edge jumping" | |
7462 | "into the loop\n", | |
7463 | loop->loop_no); | |
7464 | return false; | |
7465 | } | |
7466 | /* The loop is good. */ | |
7467 | end_label = gen_label_rtx (); | |
7468 | loop->end_label = end_label; | |
7469 | ||
7470 | /* Place the zero_cost_loop_start instruction before the loop. */ | |
7471 | entry_bb = entry_edge->src; | |
7472 | ||
7473 | start_sequence (); | |
7474 | ||
7475 | if (need_fix) | |
7476 | { | |
7477 | /* The loop uses a R-register, but the lp_count is free, thus | |
7478 | use lp_count. */ | |
7479 | emit_insn (gen_movsi (lp_reg, iter_reg)); | |
7480 | SET_HARD_REG_BIT (loop->regs_set_in_loop, LP_COUNT); | |
7481 | iter_reg = lp_reg; | |
7482 | if (dump_file) | |
7483 | { | |
7484 | fprintf (dump_file, ";; fix loop %d to use lp_count\n", | |
7485 | loop->loop_no); | |
7486 | } | |
7487 | } | |
7488 | ||
7489 | insn = emit_insn (gen_arc_lp (iter_reg, | |
7490 | loop->start_label, | |
7491 | loop->end_label)); | |
7492 | ||
7493 | seq = get_insns (); | |
7494 | end_sequence (); | |
7495 | ||
7496 | entry_after = BB_END (entry_bb); | |
7497 | if (!single_succ_p (entry_bb) || vec_safe_length (loop->incoming) > 1 | |
7498 | || !entry_after) | |
7499 | { | |
7500 | basic_block new_bb; | |
7501 | edge e; | |
7502 | edge_iterator ei; | |
7503 | ||
7504 | emit_insn_before (seq, BB_HEAD (loop->head)); | |
7505 | seq = emit_label_before (gen_label_rtx (), seq); | |
7506 | new_bb = create_basic_block (seq, insn, entry_bb); | |
7507 | FOR_EACH_EDGE (e, ei, loop->incoming) | |
7508 | { | |
7509 | if (!(e->flags & EDGE_FALLTHRU)) | |
7510 | redirect_edge_and_branch_force (e, new_bb); | |
7511 | else | |
7512 | redirect_edge_succ (e, new_bb); | |
7513 | } | |
7514 | ||
7515 | make_edge (new_bb, loop->head, 0); | |
7516 | } | |
7517 | else | |
7518 | { | |
7519 | #if 0 | |
7520 | while (DEBUG_INSN_P (entry_after) | |
7521 | || (NOTE_P (entry_after) | |
7522 | && NOTE_KIND (entry_after) != NOTE_INSN_BASIC_BLOCK | |
7523 | /* Make sure we don't split a call and its corresponding | |
7524 | CALL_ARG_LOCATION note. */ | |
7525 | && NOTE_KIND (entry_after) != NOTE_INSN_CALL_ARG_LOCATION)) | |
7526 | entry_after = NEXT_INSN (entry_after); | |
7527 | #endif | |
c483db37 | 7528 | entry_after = next_nonnote_nondebug_insn_bb (entry_after); |
a2de90a4 CZ |
7529 | |
7530 | gcc_assert (entry_after); | |
7531 | emit_insn_before (seq, entry_after); | |
7532 | } | |
7533 | ||
7534 | delete_insn (loop->loop_end); | |
7535 | /* Insert the loop end label before the last instruction of the | |
7536 | loop. */ | |
7537 | emit_label_after (end_label, loop->last_insn); | |
5d4c34aa CZ |
7538 | /* Make sure we mark the begining and end label as used. */ |
7539 | LABEL_NUSES (loop->end_label)++; | |
7540 | LABEL_NUSES (loop->start_label)++; | |
a2de90a4 CZ |
7541 | |
7542 | return true; | |
7543 | } | |
7544 | ||
7545 | /* A callback for the hw-doloop pass. This function examines INSN; if | |
7546 | it is a loop_end pattern we recognize, return the reg rtx for the | |
7547 | loop counter. Otherwise, return NULL_RTX. */ | |
7548 | ||
7549 | static rtx | |
7550 | hwloop_pattern_reg (rtx_insn *insn) | |
7551 | { | |
7552 | rtx reg; | |
7553 | ||
7554 | if (!JUMP_P (insn) || recog_memoized (insn) != CODE_FOR_loop_end) | |
7555 | return NULL_RTX; | |
7556 | ||
7557 | reg = SET_DEST (XVECEXP (PATTERN (insn), 0, 1)); | |
7558 | if (!REG_P (reg)) | |
7559 | return NULL_RTX; | |
7560 | return reg; | |
7561 | } | |
7562 | ||
7563 | static struct hw_doloop_hooks arc_doloop_hooks = | |
7564 | { | |
7565 | hwloop_pattern_reg, | |
7566 | hwloop_optimize, | |
7567 | hwloop_fail | |
7568 | }; | |
7569 | ||
7570 | /* Run from machine_dependent_reorg, this pass looks for doloop_end insns | |
7571 | and tries to rewrite the RTL of these loops so that proper Blackfin | |
7572 | hardware loops are generated. */ | |
7573 | ||
7574 | static void | |
7575 | arc_reorg_loops (void) | |
7576 | { | |
7577 | reorg_loops (true, &arc_doloop_hooks); | |
7578 | } | |
7579 | ||
526b7aee SV |
7580 | static int arc_reorg_in_progress = 0; |
7581 | ||
7582 | /* ARC's machince specific reorg function. */ | |
7583 | ||
7584 | static void | |
7585 | arc_reorg (void) | |
7586 | { | |
b3458f61 DM |
7587 | rtx_insn *insn; |
7588 | rtx pattern; | |
526b7aee SV |
7589 | rtx pc_target; |
7590 | long offset; | |
7591 | int changed; | |
7592 | ||
7593 | cfun->machine->arc_reorg_started = 1; | |
7594 | arc_reorg_in_progress = 1; | |
7595 | ||
a2de90a4 | 7596 | compute_bb_for_insn (); |
526b7aee | 7597 | |
a2de90a4 | 7598 | df_analyze (); |
526b7aee | 7599 | |
a2de90a4 CZ |
7600 | /* Doloop optimization. */ |
7601 | arc_reorg_loops (); | |
526b7aee | 7602 | |
a2de90a4 | 7603 | workaround_arc_anomaly (); |
526b7aee SV |
7604 | |
7605 | /* FIXME: should anticipate ccfsm action, generate special patterns for | |
7606 | to-be-deleted branches that have no delay slot and have at least the | |
7607 | length of the size increase forced on other insns that are conditionalized. | |
7608 | This can also have an insn_list inside that enumerates insns which are | |
7609 | not actually conditionalized because the destinations are dead in the | |
7610 | not-execute case. | |
7611 | Could also tag branches that we want to be unaligned if they get no delay | |
7612 | slot, or even ones that we don't want to do delay slot sheduling for | |
7613 | because we can unalign them. | |
7614 | ||
7615 | However, there are cases when conditional execution is only possible after | |
7616 | delay slot scheduling: | |
7617 | ||
7618 | - If a delay slot is filled with a nocond/set insn from above, the previous | |
7619 | basic block can become elegible for conditional execution. | |
7620 | - If a delay slot is filled with a nocond insn from the fall-through path, | |
7621 | the branch with that delay slot can become eligble for conditional | |
7622 | execution (however, with the same sort of data flow analysis that dbr | |
7623 | does, we could have figured out before that we don't need to | |
7624 | conditionalize this insn.) | |
7625 | - If a delay slot insn is filled with an insn from the target, the | |
7626 | target label gets its uses decremented (even deleted if falling to zero), | |
7627 | thus possibly creating more condexec opportunities there. | |
7628 | Therefore, we should still be prepared to apply condexec optimization on | |
7629 | non-prepared branches if the size increase of conditionalized insns is no | |
7630 | more than the size saved from eliminating the branch. An invocation option | |
7631 | could also be used to reserve a bit of extra size for condbranches so that | |
7632 | this'll work more often (could also test in arc_reorg if the block is | |
7633 | 'close enough' to be eligible for condexec to make this likely, and | |
7634 | estimate required size increase). */ | |
7635 | /* Generate BRcc insns, by combining cmp and Bcc insns wherever possible. */ | |
7636 | if (TARGET_NO_BRCC_SET) | |
7637 | return; | |
7638 | ||
7639 | do | |
7640 | { | |
7641 | init_insn_lengths(); | |
7642 | changed = 0; | |
7643 | ||
7644 | if (optimize > 1 && !TARGET_NO_COND_EXEC) | |
7645 | { | |
7646 | arc_ifcvt (); | |
7647 | unsigned int flags = pass_data_arc_ifcvt.todo_flags_finish; | |
7648 | df_finish_pass ((flags & TODO_df_verify) != 0); | |
782bdf21 CZ |
7649 | |
7650 | if (dump_file) | |
7651 | { | |
7652 | fprintf (dump_file, ";; After if conversion:\n\n"); | |
7653 | print_rtl (dump_file, get_insns ()); | |
7654 | } | |
526b7aee SV |
7655 | } |
7656 | ||
7657 | /* Call shorten_branches to calculate the insn lengths. */ | |
7658 | shorten_branches (get_insns()); | |
7659 | cfun->machine->ccfsm_current_insn = NULL_RTX; | |
7660 | ||
7661 | if (!INSN_ADDRESSES_SET_P()) | |
40fecdd6 | 7662 | fatal_error (input_location, "Insn addresses not set after shorten_branches"); |
526b7aee SV |
7663 | |
7664 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
7665 | { | |
7666 | rtx label; | |
7667 | enum attr_type insn_type; | |
7668 | ||
7669 | /* If a non-jump insn (or a casesi jump table), continue. */ | |
7670 | if (GET_CODE (insn) != JUMP_INSN || | |
7671 | GET_CODE (PATTERN (insn)) == ADDR_VEC | |
7672 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
7673 | continue; | |
7674 | ||
7675 | /* If we already have a brcc, note if it is suitable for brcc_s. | |
7676 | Be a bit generous with the brcc_s range so that we can take | |
7677 | advantage of any code shortening from delay slot scheduling. */ | |
7678 | if (recog_memoized (insn) == CODE_FOR_cbranchsi4_scratch) | |
7679 | { | |
7680 | rtx pat = PATTERN (insn); | |
7681 | rtx op = XEXP (SET_SRC (XVECEXP (pat, 0, 0)), 0); | |
7682 | rtx *ccp = &XEXP (XVECEXP (pat, 0, 1), 0); | |
7683 | ||
7684 | offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); | |
7685 | if ((offset >= -140 && offset < 140) | |
7686 | && rtx_equal_p (XEXP (op, 1), const0_rtx) | |
7687 | && compact_register_operand (XEXP (op, 0), VOIDmode) | |
7688 | && equality_comparison_operator (op, VOIDmode)) | |
7689 | PUT_MODE (*ccp, CC_Zmode); | |
7690 | else if (GET_MODE (*ccp) == CC_Zmode) | |
7691 | PUT_MODE (*ccp, CC_ZNmode); | |
7692 | continue; | |
7693 | } | |
7694 | if ((insn_type = get_attr_type (insn)) == TYPE_BRCC | |
7695 | || insn_type == TYPE_BRCC_NO_DELAY_SLOT) | |
7696 | continue; | |
7697 | ||
7698 | /* OK. so we have a jump insn. */ | |
7699 | /* We need to check that it is a bcc. */ | |
7700 | /* Bcc => set (pc) (if_then_else ) */ | |
7701 | pattern = PATTERN (insn); | |
7702 | if (GET_CODE (pattern) != SET | |
7703 | || GET_CODE (SET_SRC (pattern)) != IF_THEN_ELSE | |
7704 | || ANY_RETURN_P (XEXP (SET_SRC (pattern), 1))) | |
7705 | continue; | |
7706 | ||
7707 | /* Now check if the jump is beyond the s9 range. */ | |
339ba33b | 7708 | if (CROSSING_JUMP_P (insn)) |
526b7aee SV |
7709 | continue; |
7710 | offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); | |
7711 | ||
7712 | if(offset > 253 || offset < -254) | |
7713 | continue; | |
7714 | ||
7715 | pc_target = SET_SRC (pattern); | |
7716 | ||
8f3304d0 CZ |
7717 | /* Avoid FPU instructions. */ |
7718 | if ((GET_MODE (XEXP (XEXP (pc_target, 0), 0)) == CC_FPUmode) | |
7719 | || (GET_MODE (XEXP (XEXP (pc_target, 0), 0)) == CC_FPU_UNEQmode)) | |
7720 | continue; | |
7721 | ||
526b7aee SV |
7722 | /* Now go back and search for the set cc insn. */ |
7723 | ||
7724 | label = XEXP (pc_target, 1); | |
7725 | ||
7726 | { | |
b3458f61 DM |
7727 | rtx pat; |
7728 | rtx_insn *scan, *link_insn = NULL; | |
526b7aee SV |
7729 | |
7730 | for (scan = PREV_INSN (insn); | |
7731 | scan && GET_CODE (scan) != CODE_LABEL; | |
7732 | scan = PREV_INSN (scan)) | |
7733 | { | |
7734 | if (! INSN_P (scan)) | |
7735 | continue; | |
7736 | pat = PATTERN (scan); | |
7737 | if (GET_CODE (pat) == SET | |
7738 | && cc_register (SET_DEST (pat), VOIDmode)) | |
7739 | { | |
7740 | link_insn = scan; | |
7741 | break; | |
7742 | } | |
7743 | } | |
8f3304d0 | 7744 | if (!link_insn) |
526b7aee SV |
7745 | continue; |
7746 | else | |
7747 | /* Check if this is a data dependency. */ | |
7748 | { | |
7749 | rtx op, cc_clob_rtx, op0, op1, brcc_insn, note; | |
7750 | rtx cmp0, cmp1; | |
7751 | ||
7752 | /* Ok this is the set cc. copy args here. */ | |
7753 | op = XEXP (pc_target, 0); | |
7754 | ||
7755 | op0 = cmp0 = XEXP (SET_SRC (pat), 0); | |
7756 | op1 = cmp1 = XEXP (SET_SRC (pat), 1); | |
7757 | if (GET_CODE (op0) == ZERO_EXTRACT | |
7758 | && XEXP (op0, 1) == const1_rtx | |
7759 | && (GET_CODE (op) == EQ | |
7760 | || GET_CODE (op) == NE)) | |
7761 | { | |
7762 | /* btst / b{eq,ne} -> bbit{0,1} */ | |
7763 | op0 = XEXP (cmp0, 0); | |
7764 | op1 = XEXP (cmp0, 2); | |
7765 | } | |
7766 | else if (!register_operand (op0, VOIDmode) | |
7767 | || !general_operand (op1, VOIDmode)) | |
7768 | continue; | |
7769 | /* Be careful not to break what cmpsfpx_raw is | |
7770 | trying to create for checking equality of | |
7771 | single-precision floats. */ | |
7772 | else if (TARGET_SPFP | |
7773 | && GET_MODE (op0) == SFmode | |
7774 | && GET_MODE (op1) == SFmode) | |
7775 | continue; | |
7776 | ||
7777 | /* None of the two cmp operands should be set between the | |
7778 | cmp and the branch. */ | |
7779 | if (reg_set_between_p (op0, link_insn, insn)) | |
7780 | continue; | |
7781 | ||
7782 | if (reg_set_between_p (op1, link_insn, insn)) | |
7783 | continue; | |
7784 | ||
7785 | /* Since the MODE check does not work, check that this is | |
7786 | CC reg's last set location before insn, and also no | |
7787 | instruction between the cmp and branch uses the | |
7788 | condition codes. */ | |
7789 | if ((reg_set_between_p (SET_DEST (pat), link_insn, insn)) | |
7790 | || (reg_used_between_p (SET_DEST (pat), link_insn, insn))) | |
7791 | continue; | |
7792 | ||
7793 | /* CC reg should be dead after insn. */ | |
7794 | if (!find_regno_note (insn, REG_DEAD, CC_REG)) | |
7795 | continue; | |
7796 | ||
7797 | op = gen_rtx_fmt_ee (GET_CODE (op), | |
7798 | GET_MODE (op), cmp0, cmp1); | |
7799 | /* If we create a LIMM where there was none before, | |
7800 | we only benefit if we can avoid a scheduling bubble | |
7801 | for the ARC600. Otherwise, we'd only forgo chances | |
7802 | at short insn generation, and risk out-of-range | |
7803 | branches. */ | |
7804 | if (!brcc_nolimm_operator (op, VOIDmode) | |
7805 | && !long_immediate_operand (op1, VOIDmode) | |
7806 | && (TARGET_ARC700 | |
7807 | || next_active_insn (link_insn) != insn)) | |
7808 | continue; | |
7809 | ||
7810 | /* Emit bbit / brcc (or brcc_s if possible). | |
7811 | CC_Zmode indicates that brcc_s is possible. */ | |
7812 | ||
7813 | if (op0 != cmp0) | |
7814 | cc_clob_rtx = gen_rtx_REG (CC_ZNmode, CC_REG); | |
7815 | else if ((offset >= -140 && offset < 140) | |
7816 | && rtx_equal_p (op1, const0_rtx) | |
7817 | && compact_register_operand (op0, VOIDmode) | |
7818 | && (GET_CODE (op) == EQ | |
7819 | || GET_CODE (op) == NE)) | |
7820 | cc_clob_rtx = gen_rtx_REG (CC_Zmode, CC_REG); | |
7821 | else | |
7822 | cc_clob_rtx = gen_rtx_REG (CCmode, CC_REG); | |
7823 | ||
7824 | brcc_insn | |
7825 | = gen_rtx_IF_THEN_ELSE (VOIDmode, op, label, pc_rtx); | |
f7df4a84 | 7826 | brcc_insn = gen_rtx_SET (pc_rtx, brcc_insn); |
526b7aee SV |
7827 | cc_clob_rtx = gen_rtx_CLOBBER (VOIDmode, cc_clob_rtx); |
7828 | brcc_insn | |
7829 | = gen_rtx_PARALLEL | |
7830 | (VOIDmode, gen_rtvec (2, brcc_insn, cc_clob_rtx)); | |
7831 | brcc_insn = emit_jump_insn_before (brcc_insn, insn); | |
7832 | ||
7833 | JUMP_LABEL (brcc_insn) = JUMP_LABEL (insn); | |
7834 | note = find_reg_note (insn, REG_BR_PROB, 0); | |
7835 | if (note) | |
7836 | { | |
7837 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
7838 | REG_NOTES (brcc_insn) = note; | |
7839 | } | |
7840 | note = find_reg_note (link_insn, REG_DEAD, op0); | |
7841 | if (note) | |
7842 | { | |
7843 | remove_note (link_insn, note); | |
7844 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
7845 | REG_NOTES (brcc_insn) = note; | |
7846 | } | |
7847 | note = find_reg_note (link_insn, REG_DEAD, op1); | |
7848 | if (note) | |
7849 | { | |
7850 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
7851 | REG_NOTES (brcc_insn) = note; | |
7852 | } | |
7853 | ||
7854 | changed = 1; | |
7855 | ||
7856 | /* Delete the bcc insn. */ | |
7857 | set_insn_deleted (insn); | |
7858 | ||
7859 | /* Delete the cmp insn. */ | |
7860 | set_insn_deleted (link_insn); | |
7861 | ||
7862 | } | |
7863 | } | |
7864 | } | |
7865 | /* Clear out insn_addresses. */ | |
7866 | INSN_ADDRESSES_FREE (); | |
7867 | ||
7868 | } while (changed); | |
7869 | ||
7870 | if (INSN_ADDRESSES_SET_P()) | |
40fecdd6 | 7871 | fatal_error (input_location, "insn addresses not freed"); |
526b7aee SV |
7872 | |
7873 | arc_reorg_in_progress = 0; | |
7874 | } | |
7875 | ||
7876 | /* Check if the operands are valid for BRcc.d generation | |
7877 | Valid Brcc.d patterns are | |
7878 | Brcc.d b, c, s9 | |
7879 | Brcc.d b, u6, s9 | |
7880 | ||
7881 | For cc={GT, LE, GTU, LEU}, u6=63 can not be allowed, | |
7882 | since they are encoded by the assembler as {GE, LT, HS, LS} 64, which | |
7883 | does not have a delay slot | |
7884 | ||
7885 | Assumed precondition: Second operand is either a register or a u6 value. */ | |
7886 | ||
7887 | bool | |
7888 | valid_brcc_with_delay_p (rtx *operands) | |
7889 | { | |
7890 | if (optimize_size && GET_MODE (operands[4]) == CC_Zmode) | |
7891 | return false; | |
7892 | return brcc_nolimm_operator (operands[0], VOIDmode); | |
7893 | } | |
7894 | ||
526b7aee SV |
7895 | /* Implement TARGET_IN_SMALL_DATA_P. Return true if it would be safe to |
7896 | access DECL using %gp_rel(...)($gp). */ | |
7897 | ||
7898 | static bool | |
7899 | arc_in_small_data_p (const_tree decl) | |
7900 | { | |
7901 | HOST_WIDE_INT size; | |
7902 | ||
9f532472 CZ |
7903 | /* Only variables are going into small data area. */ |
7904 | if (TREE_CODE (decl) != VAR_DECL) | |
526b7aee SV |
7905 | return false; |
7906 | ||
526b7aee SV |
7907 | if (TARGET_NO_SDATA_SET) |
7908 | return false; | |
7909 | ||
526b7aee SV |
7910 | /* Disable sdata references to weak variables. */ |
7911 | if (DECL_WEAK (decl)) | |
7912 | return false; | |
7913 | ||
9f532472 CZ |
7914 | /* Don't put constants into the small data section: we want them to |
7915 | be in ROM rather than RAM. */ | |
7916 | if (TREE_READONLY (decl)) | |
7917 | return false; | |
7918 | ||
7919 | /* To ensure -mvolatile-cache works ld.di does not have a | |
7920 | gp-relative variant. */ | |
7921 | if (!TARGET_VOLATILE_CACHE_SET | |
7922 | && TREE_THIS_VOLATILE (decl)) | |
7923 | return false; | |
526b7aee | 7924 | |
9f532472 CZ |
7925 | if (DECL_SECTION_NAME (decl) != 0) |
7926 | { | |
7927 | const char *name = DECL_SECTION_NAME (decl); | |
7928 | if (strcmp (name, ".sdata") == 0 | |
7929 | || strcmp (name, ".sbss") == 0) | |
7930 | return true; | |
7931 | } | |
7932 | /* If it's not public, there's no need to put it in the small data | |
7933 | section. */ | |
7934 | else if (TREE_PUBLIC (decl)) | |
7935 | { | |
7936 | size = int_size_in_bytes (TREE_TYPE (decl)); | |
7937 | return (size > 0 && size <= g_switch_value); | |
7938 | } | |
7939 | return false; | |
526b7aee SV |
7940 | } |
7941 | ||
7942 | /* Return true if X is a small data address that can be rewritten | |
7943 | as a gp+symref. */ | |
7944 | ||
7945 | static bool | |
752ae22f | 7946 | arc_rewrite_small_data_p (const_rtx x) |
526b7aee SV |
7947 | { |
7948 | if (GET_CODE (x) == CONST) | |
7949 | x = XEXP (x, 0); | |
7950 | ||
7951 | if (GET_CODE (x) == PLUS) | |
7952 | { | |
7953 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
7954 | x = XEXP (x, 0); | |
7955 | } | |
7956 | ||
28633bbd CZ |
7957 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_SMALL_P (x)) |
7958 | { | |
7959 | gcc_assert (SYMBOL_REF_TLS_MODEL (x) == 0); | |
7960 | return true; | |
7961 | } | |
7962 | return false; | |
526b7aee SV |
7963 | } |
7964 | ||
526b7aee SV |
7965 | /* If possible, rewrite OP so that it refers to small data using |
7966 | explicit relocations. */ | |
7967 | ||
9f532472 CZ |
7968 | static rtx |
7969 | arc_rewrite_small_data_1 (rtx op) | |
526b7aee | 7970 | { |
9f532472 | 7971 | rtx rgp = gen_rtx_REG (Pmode, SDATA_BASE_REGNUM); |
526b7aee | 7972 | op = copy_insn (op); |
6733978e RS |
7973 | subrtx_ptr_iterator::array_type array; |
7974 | FOR_EACH_SUBRTX_PTR (iter, array, &op, ALL) | |
7975 | { | |
7976 | rtx *loc = *iter; | |
7977 | if (arc_rewrite_small_data_p (*loc)) | |
7978 | { | |
9f532472 | 7979 | *loc = gen_rtx_PLUS (Pmode, rgp, *loc); |
6733978e RS |
7980 | iter.skip_subrtxes (); |
7981 | } | |
7982 | else if (GET_CODE (*loc) == PLUS | |
9f532472 | 7983 | && rtx_equal_p (XEXP (*loc, 0), rgp)) |
6733978e RS |
7984 | iter.skip_subrtxes (); |
7985 | } | |
526b7aee SV |
7986 | return op; |
7987 | } | |
7988 | ||
9f532472 CZ |
7989 | rtx |
7990 | arc_rewrite_small_data (rtx op) | |
7991 | { | |
7992 | op = arc_rewrite_small_data_1 (op); | |
7993 | ||
7994 | /* Check if we fit small data constraints. */ | |
7995 | if (MEM_P (op) | |
7996 | && !LEGITIMATE_SMALL_DATA_ADDRESS_P (XEXP (op, 0))) | |
7997 | { | |
7998 | rtx addr = XEXP (op, 0); | |
7999 | rtx tmp = gen_reg_rtx (Pmode); | |
8000 | emit_move_insn (tmp, addr); | |
8001 | op = replace_equiv_address_nv (op, tmp); | |
8002 | } | |
8003 | return op; | |
8004 | } | |
8005 | ||
526b7aee SV |
8006 | /* Return true if OP refers to small data symbols directly, not through |
8007 | a PLUS. */ | |
8008 | ||
8009 | bool | |
ef4bddc2 | 8010 | small_data_pattern (rtx op, machine_mode) |
526b7aee | 8011 | { |
752ae22f RS |
8012 | if (GET_CODE (op) == SEQUENCE) |
8013 | return false; | |
9f532472 CZ |
8014 | |
8015 | rtx rgp = gen_rtx_REG (Pmode, SDATA_BASE_REGNUM); | |
752ae22f RS |
8016 | subrtx_iterator::array_type array; |
8017 | FOR_EACH_SUBRTX (iter, array, op, ALL) | |
8018 | { | |
8019 | const_rtx x = *iter; | |
8020 | if (GET_CODE (x) == PLUS | |
9f532472 | 8021 | && rtx_equal_p (XEXP (x, 0), rgp)) |
752ae22f RS |
8022 | iter.skip_subrtxes (); |
8023 | else if (arc_rewrite_small_data_p (x)) | |
8024 | return true; | |
8025 | } | |
8026 | return false; | |
526b7aee SV |
8027 | } |
8028 | ||
8029 | /* Return true if OP is an acceptable memory operand for ARCompact | |
8030 | 16-bit gp-relative load instructions. | |
8031 | op shd look like : [r26, symref@sda] | |
8032 | i.e. (mem (plus (reg 26) (symref with smalldata flag set)) | |
8033 | */ | |
8034 | /* volatile cache option still to be handled. */ | |
8035 | ||
8036 | bool | |
b6fb7933 | 8037 | compact_sda_memory_operand (rtx op, machine_mode mode, bool short_p) |
526b7aee SV |
8038 | { |
8039 | rtx addr; | |
8040 | int size; | |
b6fb7933 CZ |
8041 | tree decl = NULL_TREE; |
8042 | int align = 0; | |
8043 | int mask = 0; | |
526b7aee SV |
8044 | |
8045 | /* Eliminate non-memory operations. */ | |
8046 | if (GET_CODE (op) != MEM) | |
8047 | return false; | |
8048 | ||
8049 | if (mode == VOIDmode) | |
8050 | mode = GET_MODE (op); | |
8051 | ||
8052 | size = GET_MODE_SIZE (mode); | |
8053 | ||
8054 | /* dword operations really put out 2 instructions, so eliminate them. */ | |
8055 | if (size > UNITS_PER_WORD) | |
8056 | return false; | |
8057 | ||
8058 | /* Decode the address now. */ | |
8059 | addr = XEXP (op, 0); | |
8060 | ||
b6fb7933 CZ |
8061 | if (!LEGITIMATE_SMALL_DATA_ADDRESS_P (addr)) |
8062 | return false; | |
8063 | ||
8064 | if (!short_p || size == 1) | |
8065 | return true; | |
8066 | ||
8067 | /* Now check for the alignment, the short loads using gp require the | |
8068 | addresses to be aligned. */ | |
8069 | if (GET_CODE (XEXP (addr, 1)) == SYMBOL_REF) | |
8070 | decl = SYMBOL_REF_DECL (XEXP (addr, 1)); | |
8071 | else if (GET_CODE (XEXP (XEXP (XEXP (addr, 1), 0), 0)) == SYMBOL_REF) | |
8072 | decl = SYMBOL_REF_DECL (XEXP (XEXP (XEXP (addr, 1), 0), 0)); | |
8073 | if (decl) | |
8074 | align = DECL_ALIGN (decl); | |
8075 | align = align / BITS_PER_UNIT; | |
8076 | ||
8077 | switch (mode) | |
8078 | { | |
8079 | case E_HImode: | |
8080 | mask = 1; | |
8081 | break; | |
8082 | default: | |
8083 | mask = 3; | |
8084 | break; | |
8085 | } | |
8086 | ||
8087 | if (align && ((align & mask) == 0)) | |
8088 | return true; | |
8089 | return false; | |
526b7aee SV |
8090 | } |
8091 | ||
8092 | /* Implement ASM_OUTPUT_ALIGNED_DECL_LOCAL. */ | |
8093 | ||
8094 | void | |
8095 | arc_asm_output_aligned_decl_local (FILE * stream, tree decl, const char * name, | |
8096 | unsigned HOST_WIDE_INT size, | |
8097 | unsigned HOST_WIDE_INT align, | |
8098 | unsigned HOST_WIDE_INT globalize_p) | |
8099 | { | |
8100 | int in_small_data = arc_in_small_data_p (decl); | |
8101 | ||
8102 | if (in_small_data) | |
8103 | switch_to_section (get_named_section (NULL, ".sbss", 0)); | |
8104 | /* named_section (0,".sbss",0); */ | |
8105 | else | |
8106 | switch_to_section (bss_section); | |
8107 | ||
8108 | if (globalize_p) | |
8109 | (*targetm.asm_out.globalize_label) (stream, name); | |
8110 | ||
8111 | ASM_OUTPUT_ALIGN (stream, floor_log2 ((align) / BITS_PER_UNIT)); | |
8112 | ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "object"); | |
8113 | ASM_OUTPUT_SIZE_DIRECTIVE (stream, name, size); | |
8114 | ASM_OUTPUT_LABEL (stream, name); | |
8115 | ||
8116 | if (size != 0) | |
8117 | ASM_OUTPUT_SKIP (stream, size); | |
8118 | } | |
8119 | ||
526b7aee SV |
8120 | static bool |
8121 | arc_preserve_reload_p (rtx in) | |
8122 | { | |
8123 | return (GET_CODE (in) == PLUS | |
8124 | && RTX_OK_FOR_BASE_P (XEXP (in, 0), true) | |
8125 | && CONST_INT_P (XEXP (in, 1)) | |
8126 | && !((INTVAL (XEXP (in, 1)) & 511))); | |
8127 | } | |
8128 | ||
8129 | int | |
ef4bddc2 | 8130 | arc_register_move_cost (machine_mode, |
526b7aee SV |
8131 | enum reg_class from_class, enum reg_class to_class) |
8132 | { | |
8133 | /* The ARC600 has no bypass for extension registers, hence a nop might be | |
8134 | needed to be inserted after a write so that reads are safe. */ | |
8135 | if (TARGET_ARC600) | |
8136 | { | |
8137 | if (to_class == MPY_WRITABLE_CORE_REGS) | |
8138 | return 3; | |
8139 | /* Instructions modifying LP_COUNT need 4 additional cycles before | |
8140 | the register will actually contain the value. */ | |
8141 | else if (to_class == LPCOUNT_REG) | |
8142 | return 6; | |
8143 | else if (to_class == WRITABLE_CORE_REGS) | |
8144 | return 6; | |
8145 | } | |
8146 | ||
a2de90a4 CZ |
8147 | /* Using lp_count as scratch reg is a VERY bad idea. */ |
8148 | if (from_class == LPCOUNT_REG) | |
8149 | return 1000; | |
8150 | if (to_class == LPCOUNT_REG) | |
8151 | return 6; | |
526b7aee SV |
8152 | |
8153 | /* Force an attempt to 'mov Dy,Dx' to spill. */ | |
c4014855 | 8154 | if ((TARGET_ARC700 || TARGET_EM) && TARGET_DPFP |
526b7aee SV |
8155 | && from_class == DOUBLE_REGS && to_class == DOUBLE_REGS) |
8156 | return 100; | |
8157 | ||
8158 | return 2; | |
8159 | } | |
8160 | ||
8161 | /* Emit code for an addsi3 instruction with OPERANDS. | |
8162 | COND_P indicates if this will use conditional execution. | |
8163 | Return the length of the instruction. | |
8164 | If OUTPUT_P is false, don't actually output the instruction, just return | |
8165 | its length. */ | |
8166 | int | |
8167 | arc_output_addsi (rtx *operands, bool cond_p, bool output_p) | |
8168 | { | |
3bbe0b82 | 8169 | char format[35]; |
526b7aee SV |
8170 | |
8171 | int match = operands_match_p (operands[0], operands[1]); | |
8172 | int match2 = operands_match_p (operands[0], operands[2]); | |
8173 | int intval = (REG_P (operands[2]) ? 1 | |
8174 | : CONST_INT_P (operands[2]) ? INTVAL (operands[2]) : 0xbadc057); | |
8175 | int neg_intval = -intval; | |
8176 | int short_0 = satisfies_constraint_Rcq (operands[0]); | |
8177 | int short_p = (!cond_p && short_0 && satisfies_constraint_Rcq (operands[1])); | |
8178 | int ret = 0; | |
8179 | ||
a0caeef6 CZ |
8180 | #define REG_H_P(OP) (REG_P (OP) && ((TARGET_V2 && REGNO (OP) <= 31 \ |
8181 | && REGNO (OP) != 30) \ | |
8182 | || !TARGET_V2)) | |
8183 | ||
526b7aee SV |
8184 | #define ADDSI_OUTPUT1(FORMAT) do {\ |
8185 | if (output_p) \ | |
8186 | output_asm_insn (FORMAT, operands);\ | |
8187 | return ret; \ | |
8188 | } while (0) | |
8189 | #define ADDSI_OUTPUT(LIST) do {\ | |
8190 | if (output_p) \ | |
8191 | sprintf LIST;\ | |
8192 | ADDSI_OUTPUT1 (format);\ | |
8193 | return ret; \ | |
8194 | } while (0) | |
8195 | ||
8196 | /* First try to emit a 16 bit insn. */ | |
8197 | ret = 2; | |
8198 | if (!cond_p | |
8199 | /* If we are actually about to output this insn, don't try a 16 bit | |
8200 | variant if we already decided that we don't want that | |
8201 | (I.e. we upsized this insn to align some following insn.) | |
8202 | E.g. add_s r0,sp,70 is 16 bit, but add r0,sp,70 requires a LIMM - | |
8203 | but add1 r0,sp,35 doesn't. */ | |
8204 | && (!output_p || (get_attr_length (current_output_insn) & 2))) | |
8205 | { | |
a0caeef6 CZ |
8206 | /* Generate add_s a,b,c; add_s b,b,u7; add_s c,b,u3; add_s b,b,h |
8207 | patterns. */ | |
526b7aee | 8208 | if (short_p |
a0caeef6 CZ |
8209 | && ((REG_H_P (operands[2]) |
8210 | && (match || satisfies_constraint_Rcq (operands[2]))) | |
8211 | || (CONST_INT_P (operands[2]) | |
8212 | && ((unsigned) intval <= (match ? 127 : 7))))) | |
8213 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;1"); | |
8214 | ||
8215 | /* Generate add_s b,b,h patterns. */ | |
8216 | if (short_0 && match2 && REG_H_P (operands[1])) | |
8217 | ADDSI_OUTPUT1 ("add%? %0,%2,%1 ;2"); | |
8218 | ||
8219 | /* Generate add_s b,sp,u7; add_s sp,sp,u7 patterns. */ | |
526b7aee SV |
8220 | if ((short_0 || REGNO (operands[0]) == STACK_POINTER_REGNUM) |
8221 | && REGNO (operands[1]) == STACK_POINTER_REGNUM && !(intval & ~124)) | |
a0caeef6 | 8222 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;3"); |
526b7aee SV |
8223 | |
8224 | if ((short_p && (unsigned) neg_intval <= (match ? 31 : 7)) | |
8225 | || (REGNO (operands[0]) == STACK_POINTER_REGNUM | |
8226 | && match && !(neg_intval & ~124))) | |
a0caeef6 | 8227 | ADDSI_OUTPUT1 ("sub%? %0,%1,%n2 ;4"); |
fa9c1b3c | 8228 | |
a0caeef6 CZ |
8229 | /* Generate add_s h,h,s3 patterns. */ |
8230 | if (REG_H_P (operands[0]) && match && TARGET_V2 | |
8231 | && CONST_INT_P (operands[2]) && ((intval>= -1) && (intval <= 6))) | |
8232 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;5"); | |
fa9c1b3c | 8233 | |
a0caeef6 CZ |
8234 | /* Generate add_s r0,b,u6; add_s r1,b,u6 patterns. */ |
8235 | if (TARGET_CODE_DENSITY && REG_P (operands[0]) && REG_P (operands[1]) | |
8236 | && ((REGNO (operands[0]) == 0) || (REGNO (operands[0]) == 1)) | |
fa9c1b3c CZ |
8237 | && satisfies_constraint_Rcq (operands[1]) |
8238 | && satisfies_constraint_L (operands[2])) | |
a0caeef6 | 8239 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;6"); |
526b7aee SV |
8240 | } |
8241 | ||
8242 | /* Now try to emit a 32 bit insn without long immediate. */ | |
8243 | ret = 4; | |
8244 | if (!match && match2 && REG_P (operands[1])) | |
8245 | ADDSI_OUTPUT1 ("add%? %0,%2,%1"); | |
8246 | if (match || !cond_p) | |
8247 | { | |
8248 | int limit = (match && !cond_p) ? 0x7ff : 0x3f; | |
8249 | int range_factor = neg_intval & intval; | |
8250 | int shift; | |
8251 | ||
c419f71c | 8252 | if (intval == (HOST_WIDE_INT) (HOST_WIDE_INT_M1U << 31)) |
526b7aee SV |
8253 | ADDSI_OUTPUT1 ("bxor%? %0,%1,31"); |
8254 | ||
8255 | /* If we can use a straight add / sub instead of a {add,sub}[123] of | |
8256 | same size, do, so - the insn latency is lower. */ | |
8257 | /* -0x800 is a 12-bit constant for add /add3 / sub / sub3, but | |
8258 | 0x800 is not. */ | |
8259 | if ((intval >= 0 && intval <= limit) | |
8260 | || (intval == -0x800 && limit == 0x7ff)) | |
8261 | ADDSI_OUTPUT1 ("add%? %0,%1,%2"); | |
8262 | else if ((intval < 0 && neg_intval <= limit) | |
8263 | || (intval == 0x800 && limit == 0x7ff)) | |
8264 | ADDSI_OUTPUT1 ("sub%? %0,%1,%n2"); | |
8265 | shift = range_factor >= 8 ? 3 : (range_factor >> 1); | |
8266 | gcc_assert (shift == 0 || shift == 1 || shift == 2 || shift == 3); | |
8267 | gcc_assert ((((1 << shift) - 1) & intval) == 0); | |
8268 | if (((intval < 0 && intval != -0x4000) | |
8269 | /* sub[123] is slower than add_s / sub, only use it if it | |
8270 | avoids a long immediate. */ | |
8271 | && neg_intval <= limit << shift) | |
8272 | || (intval == 0x4000 && limit == 0x7ff)) | |
8273 | ADDSI_OUTPUT ((format, "sub%d%%? %%0,%%1,%d", | |
8274 | shift, neg_intval >> shift)); | |
8275 | else if ((intval >= 0 && intval <= limit << shift) | |
8276 | || (intval == -0x4000 && limit == 0x7ff)) | |
8277 | ADDSI_OUTPUT ((format, "add%d%%? %%0,%%1,%d", shift, intval >> shift)); | |
8278 | } | |
8279 | /* Try to emit a 16 bit opcode with long immediate. */ | |
8280 | ret = 6; | |
8281 | if (short_p && match) | |
8282 | ADDSI_OUTPUT1 ("add%? %0,%1,%S2"); | |
8283 | ||
8284 | /* We have to use a 32 bit opcode, and with a long immediate. */ | |
8285 | ret = 8; | |
8286 | ADDSI_OUTPUT1 (intval < 0 ? "sub%? %0,%1,%n2" : "add%? %0,%1,%S2"); | |
8287 | } | |
8288 | ||
8289 | /* Emit code for an commutative_cond_exec instruction with OPERANDS. | |
8290 | Return the length of the instruction. | |
8291 | If OUTPUT_P is false, don't actually output the instruction, just return | |
8292 | its length. */ | |
8293 | int | |
8294 | arc_output_commutative_cond_exec (rtx *operands, bool output_p) | |
8295 | { | |
8296 | enum rtx_code commutative_op = GET_CODE (operands[3]); | |
8297 | const char *pat = NULL; | |
8298 | ||
8299 | /* Canonical rtl should not have a constant in the first operand position. */ | |
8300 | gcc_assert (!CONSTANT_P (operands[1])); | |
8301 | ||
8302 | switch (commutative_op) | |
8303 | { | |
8304 | case AND: | |
8305 | if (satisfies_constraint_C1p (operands[2])) | |
8306 | pat = "bmsk%? %0,%1,%Z2"; | |
fc1c2d04 CZ |
8307 | else if (satisfies_constraint_C2p (operands[2])) |
8308 | { | |
8309 | operands[2] = GEN_INT ((~INTVAL (operands[2]))); | |
8310 | pat = "bmskn%? %0,%1,%Z2"; | |
8311 | } | |
526b7aee SV |
8312 | else if (satisfies_constraint_Ccp (operands[2])) |
8313 | pat = "bclr%? %0,%1,%M2"; | |
8314 | else if (satisfies_constraint_CnL (operands[2])) | |
8315 | pat = "bic%? %0,%1,%n2-1"; | |
8316 | break; | |
8317 | case IOR: | |
8318 | if (satisfies_constraint_C0p (operands[2])) | |
8319 | pat = "bset%? %0,%1,%z2"; | |
8320 | break; | |
8321 | case XOR: | |
8322 | if (satisfies_constraint_C0p (operands[2])) | |
8323 | pat = "bxor%? %0,%1,%z2"; | |
8324 | break; | |
8325 | case PLUS: | |
8326 | return arc_output_addsi (operands, true, output_p); | |
8327 | default: break; | |
8328 | } | |
8329 | if (output_p) | |
8330 | output_asm_insn (pat ? pat : "%O3.%d5 %0,%1,%2", operands); | |
8331 | if (pat || REG_P (operands[2]) || satisfies_constraint_L (operands[2])) | |
8332 | return 4; | |
8333 | return 8; | |
8334 | } | |
8335 | ||
8336 | /* Helper function of arc_expand_movmem. ADDR points to a chunk of memory. | |
8337 | Emit code and return an potentially modified address such that offsets | |
8338 | up to SIZE are can be added to yield a legitimate address. | |
8339 | if REUSE is set, ADDR is a register that may be modified. */ | |
8340 | ||
8341 | static rtx | |
8342 | force_offsettable (rtx addr, HOST_WIDE_INT size, bool reuse) | |
8343 | { | |
8344 | rtx base = addr; | |
8345 | rtx offs = const0_rtx; | |
8346 | ||
8347 | if (GET_CODE (base) == PLUS) | |
8348 | { | |
8349 | offs = XEXP (base, 1); | |
8350 | base = XEXP (base, 0); | |
8351 | } | |
8352 | if (!REG_P (base) | |
8353 | || (REGNO (base) != STACK_POINTER_REGNUM | |
4173ddaf | 8354 | && REGNO_PTR_FRAME_P (REGNO (base))) |
526b7aee SV |
8355 | || !CONST_INT_P (offs) || !SMALL_INT (INTVAL (offs)) |
8356 | || !SMALL_INT (INTVAL (offs) + size)) | |
8357 | { | |
8358 | if (reuse) | |
8359 | emit_insn (gen_add2_insn (addr, offs)); | |
8360 | else | |
8361 | addr = copy_to_mode_reg (Pmode, addr); | |
8362 | } | |
8363 | return addr; | |
8364 | } | |
8365 | ||
d34a0fdc CZ |
8366 | /* Like move_by_pieces, but take account of load latency, and actual |
8367 | offset ranges. Return true on success. */ | |
526b7aee SV |
8368 | |
8369 | bool | |
8370 | arc_expand_movmem (rtx *operands) | |
8371 | { | |
8372 | rtx dst = operands[0]; | |
8373 | rtx src = operands[1]; | |
8374 | rtx dst_addr, src_addr; | |
8375 | HOST_WIDE_INT size; | |
8376 | int align = INTVAL (operands[3]); | |
8377 | unsigned n_pieces; | |
8378 | int piece = align; | |
8379 | rtx store[2]; | |
8380 | rtx tmpx[2]; | |
8381 | int i; | |
8382 | ||
8383 | if (!CONST_INT_P (operands[2])) | |
8384 | return false; | |
8385 | size = INTVAL (operands[2]); | |
8386 | /* move_by_pieces_ninsns is static, so we can't use it. */ | |
8387 | if (align >= 4) | |
d34a0fdc CZ |
8388 | { |
8389 | if (TARGET_LL64) | |
8390 | n_pieces = (size + 4) / 8U + ((size >> 1) & 1) + (size & 1); | |
8391 | else | |
8392 | n_pieces = (size + 2) / 4U + (size & 1); | |
8393 | } | |
526b7aee SV |
8394 | else if (align == 2) |
8395 | n_pieces = (size + 1) / 2U; | |
8396 | else | |
8397 | n_pieces = size; | |
8398 | if (n_pieces >= (unsigned int) (optimize_size ? 3 : 15)) | |
8399 | return false; | |
d34a0fdc CZ |
8400 | /* Force 32 bit aligned and larger datum to use 64 bit transfers, if |
8401 | possible. */ | |
8402 | if (TARGET_LL64 && (piece >= 4) && (size >= 8)) | |
8403 | piece = 8; | |
8404 | else if (piece > 4) | |
526b7aee SV |
8405 | piece = 4; |
8406 | dst_addr = force_offsettable (XEXP (operands[0], 0), size, 0); | |
8407 | src_addr = force_offsettable (XEXP (operands[1], 0), size, 0); | |
8408 | store[0] = store[1] = NULL_RTX; | |
8409 | tmpx[0] = tmpx[1] = NULL_RTX; | |
8410 | for (i = 0; size > 0; i ^= 1, size -= piece) | |
8411 | { | |
8412 | rtx tmp; | |
ef4bddc2 | 8413 | machine_mode mode; |
526b7aee | 8414 | |
d34a0fdc CZ |
8415 | while (piece > size) |
8416 | piece >>= 1; | |
f67f4dff | 8417 | mode = smallest_int_mode_for_size (piece * BITS_PER_UNIT); |
526b7aee SV |
8418 | /* If we don't re-use temporaries, the scheduler gets carried away, |
8419 | and the register pressure gets unnecessarily high. */ | |
8420 | if (0 && tmpx[i] && GET_MODE (tmpx[i]) == mode) | |
8421 | tmp = tmpx[i]; | |
8422 | else | |
8423 | tmpx[i] = tmp = gen_reg_rtx (mode); | |
8424 | dst_addr = force_offsettable (dst_addr, piece, 1); | |
8425 | src_addr = force_offsettable (src_addr, piece, 1); | |
8426 | if (store[i]) | |
8427 | emit_insn (store[i]); | |
8428 | emit_move_insn (tmp, change_address (src, mode, src_addr)); | |
8429 | store[i] = gen_move_insn (change_address (dst, mode, dst_addr), tmp); | |
8430 | dst_addr = plus_constant (Pmode, dst_addr, piece); | |
8431 | src_addr = plus_constant (Pmode, src_addr, piece); | |
8432 | } | |
8433 | if (store[i]) | |
8434 | emit_insn (store[i]); | |
8435 | if (store[i^1]) | |
8436 | emit_insn (store[i^1]); | |
8437 | return true; | |
8438 | } | |
8439 | ||
8440 | /* Prepare operands for move in MODE. Return true iff the move has | |
8441 | been emitted. */ | |
8442 | ||
8443 | bool | |
ef4bddc2 | 8444 | prepare_move_operands (rtx *operands, machine_mode mode) |
526b7aee SV |
8445 | { |
8446 | /* We used to do this only for MODE_INT Modes, but addresses to floating | |
8447 | point variables may well be in the small data section. */ | |
28633bbd CZ |
8448 | if (!TARGET_NO_SDATA_SET && small_data_pattern (operands[0], Pmode)) |
8449 | operands[0] = arc_rewrite_small_data (operands[0]); | |
8450 | ||
8451 | if (mode == SImode && SYMBOLIC_CONST (operands[1])) | |
526b7aee | 8452 | { |
28633bbd | 8453 | prepare_pic_move (operands, SImode); |
526b7aee | 8454 | |
28633bbd CZ |
8455 | /* Disable any REG_EQUALs associated with the symref |
8456 | otherwise the optimization pass undoes the work done | |
8457 | here and references the variable directly. */ | |
8458 | } | |
8459 | ||
8460 | if (GET_CODE (operands[0]) != MEM | |
8461 | && !TARGET_NO_SDATA_SET | |
8462 | && small_data_pattern (operands[1], Pmode)) | |
8463 | { | |
8464 | /* This is to take care of address calculations involving sdata | |
8465 | variables. */ | |
8466 | operands[1] = arc_rewrite_small_data (operands[1]); | |
8467 | ||
8468 | emit_insn (gen_rtx_SET (operands[0],operands[1])); | |
8469 | /* ??? This note is useless, since it only restates the set itself. | |
8470 | We should rather use the original SYMBOL_REF. However, there is | |
8471 | the problem that we are lying to the compiler about these | |
8472 | SYMBOL_REFs to start with. symbol@sda should be encoded specially | |
8473 | so that we can tell it apart from an actual symbol. */ | |
8474 | set_unique_reg_note (get_last_insn (), REG_EQUAL, operands[1]); | |
8475 | ||
8476 | /* Take care of the REG_EQUAL note that will be attached to mark the | |
8477 | output reg equal to the initial symbol_ref after this code is | |
8478 | executed. */ | |
8479 | emit_move_insn (operands[0], operands[0]); | |
8480 | return true; | |
526b7aee SV |
8481 | } |
8482 | ||
8483 | if (MEM_P (operands[0]) | |
8484 | && !(reload_in_progress || reload_completed)) | |
8485 | { | |
8486 | operands[1] = force_reg (mode, operands[1]); | |
8487 | if (!move_dest_operand (operands[0], mode)) | |
8488 | { | |
8489 | rtx addr = copy_to_mode_reg (Pmode, XEXP (operands[0], 0)); | |
8490 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
8491 | except that we can't use that function because it is static. */ | |
8492 | rtx pat = change_address (operands[0], mode, addr); | |
8493 | MEM_COPY_ATTRIBUTES (pat, operands[0]); | |
8494 | operands[0] = pat; | |
8495 | } | |
8496 | if (!cse_not_expected) | |
8497 | { | |
8498 | rtx pat = XEXP (operands[0], 0); | |
8499 | ||
8500 | pat = arc_legitimize_address_0 (pat, pat, mode); | |
8501 | if (pat) | |
8502 | { | |
8503 | pat = change_address (operands[0], mode, pat); | |
8504 | MEM_COPY_ATTRIBUTES (pat, operands[0]); | |
8505 | operands[0] = pat; | |
8506 | } | |
8507 | } | |
8508 | } | |
8509 | ||
8510 | if (MEM_P (operands[1]) && !cse_not_expected) | |
8511 | { | |
8512 | rtx pat = XEXP (operands[1], 0); | |
8513 | ||
8514 | pat = arc_legitimize_address_0 (pat, pat, mode); | |
8515 | if (pat) | |
8516 | { | |
8517 | pat = change_address (operands[1], mode, pat); | |
8518 | MEM_COPY_ATTRIBUTES (pat, operands[1]); | |
8519 | operands[1] = pat; | |
8520 | } | |
8521 | } | |
8522 | ||
8523 | return false; | |
8524 | } | |
8525 | ||
8526 | /* Prepare OPERANDS for an extension using CODE to OMODE. | |
8527 | Return true iff the move has been emitted. */ | |
8528 | ||
8529 | bool | |
8530 | prepare_extend_operands (rtx *operands, enum rtx_code code, | |
ef4bddc2 | 8531 | machine_mode omode) |
526b7aee SV |
8532 | { |
8533 | if (!TARGET_NO_SDATA_SET && small_data_pattern (operands[1], Pmode)) | |
8534 | { | |
8535 | /* This is to take care of address calculations involving sdata | |
8536 | variables. */ | |
8537 | operands[1] | |
8538 | = gen_rtx_fmt_e (code, omode, arc_rewrite_small_data (operands[1])); | |
f7df4a84 | 8539 | emit_insn (gen_rtx_SET (operands[0], operands[1])); |
526b7aee SV |
8540 | set_unique_reg_note (get_last_insn (), REG_EQUAL, operands[1]); |
8541 | ||
8542 | /* Take care of the REG_EQUAL note that will be attached to mark the | |
8543 | output reg equal to the initial extension after this code is | |
8544 | executed. */ | |
8545 | emit_move_insn (operands[0], operands[0]); | |
8546 | return true; | |
8547 | } | |
8548 | return false; | |
8549 | } | |
8550 | ||
8551 | /* Output a library call to a function called FNAME that has been arranged | |
8552 | to be local to any dso. */ | |
8553 | ||
8554 | const char * | |
8555 | arc_output_libcall (const char *fname) | |
8556 | { | |
8557 | unsigned len = strlen (fname); | |
8558 | static char buf[64]; | |
8559 | ||
8560 | gcc_assert (len < sizeof buf - 35); | |
8561 | if (TARGET_LONG_CALLS_SET | |
8562 | || (TARGET_MEDIUM_CALLS && arc_ccfsm_cond_exec_p ())) | |
8563 | { | |
8564 | if (flag_pic) | |
f5e336b1 | 8565 | sprintf (buf, "add r12,pcl,@%s@pcl\n\tjl%%!%%* [r12]", fname); |
526b7aee SV |
8566 | else |
8567 | sprintf (buf, "jl%%! @%s", fname); | |
8568 | } | |
8569 | else | |
8570 | sprintf (buf, "bl%%!%%* @%s", fname); | |
8571 | return buf; | |
8572 | } | |
8573 | ||
8574 | /* Return the SImode highpart of the DImode value IN. */ | |
8575 | ||
8576 | rtx | |
8577 | disi_highpart (rtx in) | |
8578 | { | |
8579 | return simplify_gen_subreg (SImode, in, DImode, TARGET_BIG_ENDIAN ? 0 : 4); | |
8580 | } | |
8581 | ||
526b7aee SV |
8582 | /* Return length adjustment for INSN. |
8583 | For ARC600: | |
8584 | A write to a core reg greater or equal to 32 must not be immediately | |
8585 | followed by a use. Anticipate the length requirement to insert a nop | |
8586 | between PRED and SUCC to prevent a hazard. */ | |
8587 | ||
8588 | static int | |
647d790d | 8589 | arc600_corereg_hazard (rtx_insn *pred, rtx_insn *succ) |
526b7aee SV |
8590 | { |
8591 | if (!TARGET_ARC600) | |
8592 | return 0; | |
526b7aee | 8593 | if (GET_CODE (PATTERN (pred)) == SEQUENCE) |
647d790d | 8594 | pred = as_a <rtx_sequence *> (PATTERN (pred))->insn (1); |
526b7aee | 8595 | if (GET_CODE (PATTERN (succ)) == SEQUENCE) |
647d790d | 8596 | succ = as_a <rtx_sequence *> (PATTERN (succ))->insn (0); |
526b7aee SV |
8597 | if (recog_memoized (pred) == CODE_FOR_mulsi_600 |
8598 | || recog_memoized (pred) == CODE_FOR_umul_600 | |
8599 | || recog_memoized (pred) == CODE_FOR_mac_600 | |
8600 | || recog_memoized (pred) == CODE_FOR_mul64_600 | |
8601 | || recog_memoized (pred) == CODE_FOR_mac64_600 | |
8602 | || recog_memoized (pred) == CODE_FOR_umul64_600 | |
8603 | || recog_memoized (pred) == CODE_FOR_umac64_600) | |
8604 | return 0; | |
36cc6254 RS |
8605 | subrtx_iterator::array_type array; |
8606 | FOR_EACH_SUBRTX (iter, array, PATTERN (pred), NONCONST) | |
8607 | { | |
8608 | const_rtx x = *iter; | |
8609 | switch (GET_CODE (x)) | |
8610 | { | |
8611 | case SET: case POST_INC: case POST_DEC: case PRE_INC: case PRE_DEC: | |
8612 | break; | |
8613 | default: | |
8614 | /* This is also fine for PRE/POST_MODIFY, because they | |
8615 | contain a SET. */ | |
8616 | continue; | |
8617 | } | |
8618 | rtx dest = XEXP (x, 0); | |
8619 | /* Check if this sets a an extension register. N.B. we use 61 for the | |
8620 | condition codes, which is definitely not an extension register. */ | |
8621 | if (REG_P (dest) && REGNO (dest) >= 32 && REGNO (dest) < 61 | |
8622 | /* Check if the same register is used by the PAT. */ | |
8623 | && (refers_to_regno_p | |
8624 | (REGNO (dest), | |
8625 | REGNO (dest) + (GET_MODE_SIZE (GET_MODE (dest)) + 3) / 4U, | |
8626 | PATTERN (succ), 0))) | |
8627 | return 4; | |
8628 | } | |
8629 | return 0; | |
526b7aee SV |
8630 | } |
8631 | ||
f50bb868 CZ |
8632 | /* Given a rtx, check if it is an assembly instruction or not. */ |
8633 | ||
8634 | static int | |
8635 | arc_asm_insn_p (rtx x) | |
8636 | { | |
8637 | int i, j; | |
8638 | ||
8639 | if (x == 0) | |
8640 | return 0; | |
8641 | ||
8642 | switch (GET_CODE (x)) | |
8643 | { | |
8644 | case ASM_OPERANDS: | |
8645 | case ASM_INPUT: | |
8646 | return 1; | |
8647 | ||
8648 | case SET: | |
8649 | return arc_asm_insn_p (SET_SRC (x)); | |
8650 | ||
8651 | case PARALLEL: | |
8652 | j = 0; | |
8653 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
8654 | j += arc_asm_insn_p (XVECEXP (x, 0, i)); | |
8655 | if ( j > 0) | |
8656 | return 1; | |
8657 | break; | |
8658 | ||
8659 | default: | |
8660 | break; | |
8661 | } | |
8662 | ||
8663 | return 0; | |
8664 | } | |
8665 | ||
526b7aee SV |
8666 | /* For ARC600: |
8667 | A write to a core reg greater or equal to 32 must not be immediately | |
8668 | followed by a use. Anticipate the length requirement to insert a nop | |
8669 | between PRED and SUCC to prevent a hazard. */ | |
8670 | ||
8671 | int | |
647d790d | 8672 | arc_hazard (rtx_insn *pred, rtx_insn *succ) |
526b7aee | 8673 | { |
526b7aee SV |
8674 | if (!pred || !INSN_P (pred) || !succ || !INSN_P (succ)) |
8675 | return 0; | |
f50bb868 | 8676 | |
f50bb868 CZ |
8677 | if (TARGET_ARC600) |
8678 | return arc600_corereg_hazard (pred, succ); | |
8679 | ||
8680 | return 0; | |
526b7aee SV |
8681 | } |
8682 | ||
8683 | /* Return length adjustment for INSN. */ | |
8684 | ||
8685 | int | |
647d790d | 8686 | arc_adjust_insn_length (rtx_insn *insn, int len, bool) |
526b7aee SV |
8687 | { |
8688 | if (!INSN_P (insn)) | |
8689 | return len; | |
8690 | /* We already handle sequences by ignoring the delay sequence flag. */ | |
8691 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) | |
8692 | return len; | |
8693 | ||
526b7aee SV |
8694 | /* Check for return with but one preceding insn since function |
8695 | start / call. */ | |
8696 | if (TARGET_PAD_RETURN | |
8697 | && JUMP_P (insn) | |
8698 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
8699 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
8700 | && get_attr_type (insn) == TYPE_RETURN) | |
8701 | { | |
84034c69 | 8702 | rtx_insn *prev = prev_active_insn (insn); |
526b7aee SV |
8703 | |
8704 | if (!prev || !(prev = prev_active_insn (prev)) | |
8705 | || ((NONJUMP_INSN_P (prev) | |
8706 | && GET_CODE (PATTERN (prev)) == SEQUENCE) | |
84034c69 DM |
8707 | ? CALL_ATTR (as_a <rtx_sequence *> (PATTERN (prev))->insn (0), |
8708 | NON_SIBCALL) | |
526b7aee SV |
8709 | : CALL_ATTR (prev, NON_SIBCALL))) |
8710 | return len + 4; | |
8711 | } | |
8712 | if (TARGET_ARC600) | |
8713 | { | |
b3458f61 | 8714 | rtx_insn *succ = next_real_insn (insn); |
526b7aee SV |
8715 | |
8716 | /* One the ARC600, a write to an extension register must be separated | |
8717 | from a read. */ | |
8718 | if (succ && INSN_P (succ)) | |
8719 | len += arc600_corereg_hazard (insn, succ); | |
8720 | } | |
8721 | ||
8722 | /* Restore extracted operands - otherwise splitters like the addsi3_mixed one | |
8723 | can go awry. */ | |
8724 | extract_constrain_insn_cached (insn); | |
8725 | ||
8726 | return len; | |
8727 | } | |
8728 | ||
8729 | /* Values for length_sensitive. */ | |
8730 | enum | |
8731 | { | |
8732 | ARC_LS_NONE,// Jcc | |
8733 | ARC_LS_25, // 25 bit offset, B | |
8734 | ARC_LS_21, // 21 bit offset, Bcc | |
8735 | ARC_LS_U13,// 13 bit unsigned offset, LP | |
8736 | ARC_LS_10, // 10 bit offset, B_s, Beq_s, Bne_s | |
8737 | ARC_LS_9, // 9 bit offset, BRcc | |
8738 | ARC_LS_8, // 8 bit offset, BRcc_s | |
8739 | ARC_LS_U7, // 7 bit unsigned offset, LPcc | |
8740 | ARC_LS_7 // 7 bit offset, Bcc_s | |
8741 | }; | |
8742 | ||
8743 | /* While the infrastructure patch is waiting for review, duplicate the | |
8744 | struct definitions, to allow this file to compile. */ | |
8745 | #if 1 | |
8746 | typedef struct | |
8747 | { | |
8748 | unsigned align_set; | |
8749 | /* Cost as a branch / call target or call return address. */ | |
8750 | int target_cost; | |
8751 | int fallthrough_cost; | |
8752 | int branch_cost; | |
8753 | int length; | |
8754 | /* 0 for not length sensitive, 1 for largest offset range, | |
8755 | * 2 for next smaller etc. */ | |
8756 | unsigned length_sensitive : 8; | |
8757 | bool enabled; | |
8758 | } insn_length_variant_t; | |
8759 | ||
8760 | typedef struct insn_length_parameters_s | |
8761 | { | |
8762 | int align_unit_log; | |
8763 | int align_base_log; | |
8764 | int max_variants; | |
647d790d | 8765 | int (*get_variants) (rtx_insn *, int, bool, bool, insn_length_variant_t *); |
526b7aee SV |
8766 | } insn_length_parameters_t; |
8767 | ||
8768 | static void | |
8769 | arc_insn_length_parameters (insn_length_parameters_t *ilp) ATTRIBUTE_UNUSED; | |
8770 | #endif | |
8771 | ||
8772 | static int | |
647d790d | 8773 | arc_get_insn_variants (rtx_insn *insn, int len, bool, bool target_p, |
526b7aee SV |
8774 | insn_length_variant_t *ilv) |
8775 | { | |
8776 | if (!NONDEBUG_INSN_P (insn)) | |
8777 | return 0; | |
8778 | enum attr_type type; | |
8779 | /* shorten_branches doesn't take optimize_size into account yet for the | |
8780 | get_variants mechanism, so turn this off for now. */ | |
8781 | if (optimize_size) | |
8782 | return 0; | |
647d790d | 8783 | if (rtx_sequence *pat = dyn_cast <rtx_sequence *> (PATTERN (insn))) |
526b7aee SV |
8784 | { |
8785 | /* The interaction of a short delay slot insn with a short branch is | |
8786 | too weird for shorten_branches to piece together, so describe the | |
8787 | entire SEQUENCE. */ | |
647d790d | 8788 | rtx_insn *inner; |
526b7aee | 8789 | if (TARGET_UPSIZE_DBR |
84034c69 | 8790 | && get_attr_length (pat->insn (1)) <= 2 |
647d790d | 8791 | && (((type = get_attr_type (inner = pat->insn (0))) |
526b7aee SV |
8792 | == TYPE_UNCOND_BRANCH) |
8793 | || type == TYPE_BRANCH) | |
8794 | && get_attr_delay_slot_filled (inner) == DELAY_SLOT_FILLED_YES) | |
8795 | { | |
8796 | int n_variants | |
8797 | = arc_get_insn_variants (inner, get_attr_length (inner), true, | |
8798 | target_p, ilv+1); | |
8799 | /* The short variant gets split into a higher-cost aligned | |
8800 | and a lower cost unaligned variant. */ | |
8801 | gcc_assert (n_variants); | |
8802 | gcc_assert (ilv[1].length_sensitive == ARC_LS_7 | |
8803 | || ilv[1].length_sensitive == ARC_LS_10); | |
8804 | gcc_assert (ilv[1].align_set == 3); | |
8805 | ilv[0] = ilv[1]; | |
8806 | ilv[0].align_set = 1; | |
8807 | ilv[0].branch_cost += 1; | |
8808 | ilv[1].align_set = 2; | |
8809 | n_variants++; | |
8810 | for (int i = 0; i < n_variants; i++) | |
8811 | ilv[i].length += 2; | |
8812 | /* In case an instruction with aligned size is wanted, and | |
8813 | the short variants are unavailable / too expensive, add | |
8814 | versions of long branch + long delay slot. */ | |
8815 | for (int i = 2, end = n_variants; i < end; i++, n_variants++) | |
8816 | { | |
8817 | ilv[n_variants] = ilv[i]; | |
8818 | ilv[n_variants].length += 2; | |
8819 | } | |
8820 | return n_variants; | |
8821 | } | |
8822 | return 0; | |
8823 | } | |
8824 | insn_length_variant_t *first_ilv = ilv; | |
8825 | type = get_attr_type (insn); | |
8826 | bool delay_filled | |
8827 | = (get_attr_delay_slot_filled (insn) == DELAY_SLOT_FILLED_YES); | |
8828 | int branch_align_cost = delay_filled ? 0 : 1; | |
8829 | int branch_unalign_cost = delay_filled ? 0 : TARGET_UNALIGN_BRANCH ? 0 : 1; | |
8830 | /* If the previous instruction is an sfunc call, this insn is always | |
8831 | a target, even though the middle-end is unaware of this. */ | |
8832 | bool force_target = false; | |
b3458f61 | 8833 | rtx_insn *prev = prev_active_insn (insn); |
526b7aee SV |
8834 | if (prev && arc_next_active_insn (prev, 0) == insn |
8835 | && ((NONJUMP_INSN_P (prev) && GET_CODE (PATTERN (prev)) == SEQUENCE) | |
84034c69 DM |
8836 | ? CALL_ATTR (as_a <rtx_sequence *> (PATTERN (prev))->insn (0), |
8837 | NON_SIBCALL) | |
526b7aee SV |
8838 | : (CALL_ATTR (prev, NON_SIBCALL) |
8839 | && NEXT_INSN (PREV_INSN (prev)) == prev))) | |
8840 | force_target = true; | |
8841 | ||
8842 | switch (type) | |
8843 | { | |
8844 | case TYPE_BRCC: | |
8845 | /* Short BRCC only comes in no-delay-slot version, and without limm */ | |
8846 | if (!delay_filled) | |
8847 | { | |
8848 | ilv->align_set = 3; | |
8849 | ilv->length = 2; | |
8850 | ilv->branch_cost = 1; | |
8851 | ilv->enabled = (len == 2); | |
8852 | ilv->length_sensitive = ARC_LS_8; | |
8853 | ilv++; | |
8854 | } | |
8855 | /* Fall through. */ | |
8856 | case TYPE_BRCC_NO_DELAY_SLOT: | |
8857 | /* doloop_fallback* patterns are TYPE_BRCC_NO_DELAY_SLOT for | |
8858 | (delay slot) scheduling purposes, but they are longer. */ | |
8859 | if (GET_CODE (PATTERN (insn)) == PARALLEL | |
8860 | && GET_CODE (XVECEXP (PATTERN (insn), 0, 1)) == SET) | |
8861 | return 0; | |
8862 | /* Standard BRCC: 4 bytes, or 8 bytes with limm. */ | |
8863 | ilv->length = ((type == TYPE_BRCC) ? 4 : 8); | |
8864 | ilv->align_set = 3; | |
8865 | ilv->branch_cost = branch_align_cost; | |
8866 | ilv->enabled = (len <= ilv->length); | |
8867 | ilv->length_sensitive = ARC_LS_9; | |
8868 | if ((target_p || force_target) | |
8869 | || (!delay_filled && TARGET_UNALIGN_BRANCH)) | |
8870 | { | |
8871 | ilv[1] = *ilv; | |
8872 | ilv->align_set = 1; | |
8873 | ilv++; | |
8874 | ilv->align_set = 2; | |
8875 | ilv->target_cost = 1; | |
8876 | ilv->branch_cost = branch_unalign_cost; | |
8877 | } | |
8878 | ilv++; | |
8879 | ||
8880 | rtx op, op0; | |
8881 | op = XEXP (SET_SRC (XVECEXP (PATTERN (insn), 0, 0)), 0); | |
8882 | op0 = XEXP (op, 0); | |
8883 | ||
8884 | if (GET_CODE (op0) == ZERO_EXTRACT | |
8885 | && satisfies_constraint_L (XEXP (op0, 2))) | |
8886 | op0 = XEXP (op0, 0); | |
8887 | if (satisfies_constraint_Rcq (op0)) | |
8888 | { | |
8889 | ilv->length = ((type == TYPE_BRCC) ? 6 : 10); | |
8890 | ilv->align_set = 3; | |
8891 | ilv->branch_cost = 1 + branch_align_cost; | |
8892 | ilv->fallthrough_cost = 1; | |
8893 | ilv->enabled = true; | |
8894 | ilv->length_sensitive = ARC_LS_21; | |
8895 | if (!delay_filled && TARGET_UNALIGN_BRANCH) | |
8896 | { | |
8897 | ilv[1] = *ilv; | |
8898 | ilv->align_set = 1; | |
8899 | ilv++; | |
8900 | ilv->align_set = 2; | |
8901 | ilv->branch_cost = 1 + branch_unalign_cost; | |
8902 | } | |
8903 | ilv++; | |
8904 | } | |
8905 | ilv->length = ((type == TYPE_BRCC) ? 8 : 12); | |
8906 | ilv->align_set = 3; | |
8907 | ilv->branch_cost = 1 + branch_align_cost; | |
8908 | ilv->fallthrough_cost = 1; | |
8909 | ilv->enabled = true; | |
8910 | ilv->length_sensitive = ARC_LS_21; | |
8911 | if ((target_p || force_target) | |
8912 | || (!delay_filled && TARGET_UNALIGN_BRANCH)) | |
8913 | { | |
8914 | ilv[1] = *ilv; | |
8915 | ilv->align_set = 1; | |
8916 | ilv++; | |
8917 | ilv->align_set = 2; | |
8918 | ilv->target_cost = 1; | |
8919 | ilv->branch_cost = 1 + branch_unalign_cost; | |
8920 | } | |
8921 | ilv++; | |
8922 | break; | |
8923 | ||
8924 | case TYPE_SFUNC: | |
8925 | ilv->length = 12; | |
8926 | goto do_call; | |
8927 | case TYPE_CALL_NO_DELAY_SLOT: | |
8928 | ilv->length = 8; | |
8929 | goto do_call; | |
8930 | case TYPE_CALL: | |
8931 | ilv->length = 4; | |
8932 | ilv->length_sensitive | |
8933 | = GET_CODE (PATTERN (insn)) == COND_EXEC ? ARC_LS_21 : ARC_LS_25; | |
8934 | do_call: | |
8935 | ilv->align_set = 3; | |
8936 | ilv->fallthrough_cost = branch_align_cost; | |
8937 | ilv->enabled = true; | |
8938 | if ((target_p || force_target) | |
8939 | || (!delay_filled && TARGET_UNALIGN_BRANCH)) | |
8940 | { | |
8941 | ilv[1] = *ilv; | |
8942 | ilv->align_set = 1; | |
8943 | ilv++; | |
8944 | ilv->align_set = 2; | |
8945 | ilv->target_cost = 1; | |
8946 | ilv->fallthrough_cost = branch_unalign_cost; | |
8947 | } | |
8948 | ilv++; | |
8949 | break; | |
8950 | case TYPE_UNCOND_BRANCH: | |
8951 | /* Strictly speaking, this should be ARC_LS_10 for equality comparisons, | |
8952 | but that makes no difference at the moment. */ | |
8953 | ilv->length_sensitive = ARC_LS_7; | |
8954 | ilv[1].length_sensitive = ARC_LS_25; | |
8955 | goto do_branch; | |
8956 | case TYPE_BRANCH: | |
8957 | ilv->length_sensitive = ARC_LS_10; | |
8958 | ilv[1].length_sensitive = ARC_LS_21; | |
8959 | do_branch: | |
8960 | ilv->align_set = 3; | |
8961 | ilv->length = 2; | |
8962 | ilv->branch_cost = branch_align_cost; | |
8963 | ilv->enabled = (len == ilv->length); | |
8964 | ilv++; | |
8965 | ilv->length = 4; | |
8966 | ilv->align_set = 3; | |
8967 | ilv->branch_cost = branch_align_cost; | |
8968 | ilv->enabled = true; | |
8969 | if ((target_p || force_target) | |
8970 | || (!delay_filled && TARGET_UNALIGN_BRANCH)) | |
8971 | { | |
8972 | ilv[1] = *ilv; | |
8973 | ilv->align_set = 1; | |
8974 | ilv++; | |
8975 | ilv->align_set = 2; | |
8976 | ilv->target_cost = 1; | |
8977 | ilv->branch_cost = branch_unalign_cost; | |
8978 | } | |
8979 | ilv++; | |
8980 | break; | |
8981 | case TYPE_JUMP: | |
8982 | return 0; | |
8983 | default: | |
8984 | /* For every short insn, there is generally also a long insn. | |
8985 | trap_s is an exception. */ | |
8986 | if ((len & 2) == 0 || recog_memoized (insn) == CODE_FOR_trap_s) | |
8987 | return 0; | |
8988 | ilv->align_set = 3; | |
8989 | ilv->length = len; | |
8990 | ilv->enabled = 1; | |
8991 | ilv++; | |
8992 | ilv->align_set = 3; | |
8993 | ilv->length = len + 2; | |
8994 | ilv->enabled = 1; | |
8995 | if (target_p || force_target) | |
8996 | { | |
8997 | ilv[1] = *ilv; | |
8998 | ilv->align_set = 1; | |
8999 | ilv++; | |
9000 | ilv->align_set = 2; | |
9001 | ilv->target_cost = 1; | |
9002 | } | |
9003 | ilv++; | |
9004 | } | |
9005 | /* If the previous instruction is an sfunc call, this insn is always | |
9006 | a target, even though the middle-end is unaware of this. | |
9007 | Therefore, if we have a call predecessor, transfer the target cost | |
9008 | to the fallthrough and branch costs. */ | |
9009 | if (force_target) | |
9010 | { | |
9011 | for (insn_length_variant_t *p = first_ilv; p < ilv; p++) | |
9012 | { | |
9013 | p->fallthrough_cost += p->target_cost; | |
9014 | p->branch_cost += p->target_cost; | |
9015 | p->target_cost = 0; | |
9016 | } | |
9017 | } | |
9018 | ||
9019 | return ilv - first_ilv; | |
9020 | } | |
9021 | ||
9022 | static void | |
9023 | arc_insn_length_parameters (insn_length_parameters_t *ilp) | |
9024 | { | |
9025 | ilp->align_unit_log = 1; | |
9026 | ilp->align_base_log = 1; | |
9027 | ilp->max_variants = 7; | |
9028 | ilp->get_variants = arc_get_insn_variants; | |
9029 | } | |
9030 | ||
9031 | /* Return a copy of COND from *STATEP, inverted if that is indicated by the | |
9032 | CC field of *STATEP. */ | |
9033 | ||
9034 | static rtx | |
9035 | arc_get_ccfsm_cond (struct arc_ccfsm *statep, bool reverse) | |
9036 | { | |
9037 | rtx cond = statep->cond; | |
9038 | int raw_cc = get_arc_condition_code (cond); | |
9039 | if (reverse) | |
9040 | raw_cc = ARC_INVERSE_CONDITION_CODE (raw_cc); | |
9041 | ||
9042 | if (statep->cc == raw_cc) | |
9043 | return copy_rtx (cond); | |
9044 | ||
9045 | gcc_assert (ARC_INVERSE_CONDITION_CODE (raw_cc) == statep->cc); | |
9046 | ||
ef4bddc2 | 9047 | machine_mode ccm = GET_MODE (XEXP (cond, 0)); |
526b7aee SV |
9048 | enum rtx_code code = reverse_condition (GET_CODE (cond)); |
9049 | if (code == UNKNOWN || ccm == CC_FP_GTmode || ccm == CC_FP_GEmode) | |
9050 | code = reverse_condition_maybe_unordered (GET_CODE (cond)); | |
9051 | ||
9052 | return gen_rtx_fmt_ee (code, GET_MODE (cond), | |
9053 | copy_rtx (XEXP (cond, 0)), copy_rtx (XEXP (cond, 1))); | |
9054 | } | |
9055 | ||
bae56bbb JR |
9056 | /* Return version of PAT conditionalized with COND, which is part of INSN. |
9057 | ANNULLED indicates if INSN is an annulled delay-slot insn. | |
9058 | Register further changes if necessary. */ | |
9059 | static rtx | |
9060 | conditionalize_nonjump (rtx pat, rtx cond, rtx insn, bool annulled) | |
9061 | { | |
9062 | /* For commutative operators, we generally prefer to have | |
9063 | the first source match the destination. */ | |
9064 | if (GET_CODE (pat) == SET) | |
9065 | { | |
9066 | rtx src = SET_SRC (pat); | |
9067 | ||
9068 | if (COMMUTATIVE_P (src)) | |
9069 | { | |
9070 | rtx src0 = XEXP (src, 0); | |
9071 | rtx src1 = XEXP (src, 1); | |
9072 | rtx dst = SET_DEST (pat); | |
9073 | ||
9074 | if (rtx_equal_p (src1, dst) && !rtx_equal_p (src0, dst) | |
9075 | /* Leave add_n alone - the canonical form is to | |
9076 | have the complex summand first. */ | |
9077 | && REG_P (src0)) | |
f7df4a84 | 9078 | pat = gen_rtx_SET (dst, |
bae56bbb JR |
9079 | gen_rtx_fmt_ee (GET_CODE (src), GET_MODE (src), |
9080 | src1, src0)); | |
9081 | } | |
9082 | } | |
9083 | ||
9084 | /* dwarf2out.c:dwarf2out_frame_debug_expr doesn't know | |
9085 | what to do with COND_EXEC. */ | |
9086 | if (RTX_FRAME_RELATED_P (insn)) | |
9087 | { | |
9088 | /* If this is the delay slot insn of an anulled branch, | |
9089 | dwarf2out.c:scan_trace understands the anulling semantics | |
9090 | without the COND_EXEC. */ | |
9091 | gcc_assert (annulled); | |
9092 | rtx note = alloc_reg_note (REG_FRAME_RELATED_EXPR, pat, | |
9093 | REG_NOTES (insn)); | |
9094 | validate_change (insn, ®_NOTES (insn), note, 1); | |
9095 | } | |
9096 | pat = gen_rtx_COND_EXEC (VOIDmode, cond, pat); | |
9097 | return pat; | |
9098 | } | |
9099 | ||
526b7aee SV |
9100 | /* Use the ccfsm machinery to do if conversion. */ |
9101 | ||
9102 | static unsigned | |
9103 | arc_ifcvt (void) | |
9104 | { | |
9105 | struct arc_ccfsm *statep = &cfun->machine->ccfsm_current; | |
526b7aee SV |
9106 | |
9107 | memset (statep, 0, sizeof *statep); | |
b3458f61 | 9108 | for (rtx_insn *insn = get_insns (); insn; insn = next_insn (insn)) |
526b7aee SV |
9109 | { |
9110 | arc_ccfsm_advance (insn, statep); | |
9111 | ||
9112 | switch (statep->state) | |
9113 | { | |
9114 | case 0: | |
526b7aee SV |
9115 | break; |
9116 | case 1: case 2: | |
9117 | { | |
9118 | /* Deleted branch. */ | |
526b7aee | 9119 | arc_ccfsm_post_advance (insn, statep); |
53ea364f | 9120 | gcc_assert (!IN_RANGE (statep->state, 1, 2)); |
b3458f61 | 9121 | rtx_insn *seq = NEXT_INSN (PREV_INSN (insn)); |
782bdf21 | 9122 | if (GET_CODE (PATTERN (seq)) == SEQUENCE) |
526b7aee SV |
9123 | { |
9124 | rtx slot = XVECEXP (PATTERN (seq), 0, 1); | |
9125 | rtx pat = PATTERN (slot); | |
9126 | if (INSN_ANNULLED_BRANCH_P (insn)) | |
9127 | { | |
9128 | rtx cond | |
9129 | = arc_get_ccfsm_cond (statep, INSN_FROM_TARGET_P (slot)); | |
9130 | pat = gen_rtx_COND_EXEC (VOIDmode, cond, pat); | |
9131 | } | |
9132 | if (!validate_change (seq, &PATTERN (seq), pat, 0)) | |
9133 | gcc_unreachable (); | |
9134 | PUT_CODE (slot, NOTE); | |
9135 | NOTE_KIND (slot) = NOTE_INSN_DELETED; | |
526b7aee SV |
9136 | } |
9137 | else | |
9138 | { | |
782bdf21 | 9139 | set_insn_deleted (insn); |
526b7aee SV |
9140 | } |
9141 | continue; | |
9142 | } | |
9143 | case 3: | |
9144 | if (LABEL_P (insn) | |
9145 | && statep->target_label == CODE_LABEL_NUMBER (insn)) | |
9146 | { | |
9147 | arc_ccfsm_post_advance (insn, statep); | |
782bdf21 CZ |
9148 | if (--LABEL_NUSES (insn) == 0) |
9149 | delete_insn (insn); | |
526b7aee SV |
9150 | continue; |
9151 | } | |
9152 | /* Fall through. */ | |
9153 | case 4: case 5: | |
9154 | if (!NONDEBUG_INSN_P (insn)) | |
9155 | break; | |
9156 | ||
9157 | /* Conditionalized insn. */ | |
9158 | ||
b3458f61 DM |
9159 | rtx_insn *prev, *pprev; |
9160 | rtx *patp, pat, cond; | |
bae56bbb | 9161 | bool annulled; annulled = false; |
526b7aee SV |
9162 | |
9163 | /* If this is a delay slot insn in a non-annulled branch, | |
9164 | don't conditionalize it. N.B., this should be fine for | |
9165 | conditional return too. However, don't do this for | |
9166 | unconditional branches, as these would be encountered when | |
9167 | processing an 'else' part. */ | |
9168 | prev = PREV_INSN (insn); | |
9169 | pprev = PREV_INSN (prev); | |
9170 | if (pprev && NEXT_INSN (NEXT_INSN (pprev)) == NEXT_INSN (insn) | |
bae56bbb JR |
9171 | && JUMP_P (prev) && get_attr_cond (prev) == COND_USE) |
9172 | { | |
9173 | if (!INSN_ANNULLED_BRANCH_P (prev)) | |
9174 | break; | |
9175 | annulled = true; | |
9176 | } | |
526b7aee SV |
9177 | |
9178 | patp = &PATTERN (insn); | |
9179 | pat = *patp; | |
9180 | cond = arc_get_ccfsm_cond (statep, INSN_FROM_TARGET_P (insn)); | |
9181 | if (NONJUMP_INSN_P (insn) || CALL_P (insn)) | |
9182 | { | |
9183 | /* ??? don't conditionalize if all side effects are dead | |
9184 | in the not-execute case. */ | |
9bf218f9 | 9185 | |
bae56bbb | 9186 | pat = conditionalize_nonjump (pat, cond, insn, annulled); |
526b7aee SV |
9187 | } |
9188 | else if (simplejump_p (insn)) | |
9189 | { | |
9190 | patp = &SET_SRC (pat); | |
9191 | pat = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, *patp, pc_rtx); | |
9192 | } | |
9193 | else if (JUMP_P (insn) && ANY_RETURN_P (PATTERN (insn))) | |
9194 | { | |
9195 | pat = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, pat, pc_rtx); | |
f7df4a84 | 9196 | pat = gen_rtx_SET (pc_rtx, pat); |
526b7aee SV |
9197 | } |
9198 | else | |
9199 | gcc_unreachable (); | |
9200 | validate_change (insn, patp, pat, 1); | |
9201 | if (!apply_change_group ()) | |
9202 | gcc_unreachable (); | |
9203 | if (JUMP_P (insn)) | |
9204 | { | |
b3458f61 | 9205 | rtx_insn *next = next_nonnote_insn (insn); |
526b7aee SV |
9206 | if (GET_CODE (next) == BARRIER) |
9207 | delete_insn (next); | |
9208 | if (statep->state == 3) | |
9209 | continue; | |
9210 | } | |
9211 | break; | |
9212 | default: | |
9213 | gcc_unreachable (); | |
9214 | } | |
9215 | arc_ccfsm_post_advance (insn, statep); | |
9216 | } | |
9217 | return 0; | |
9218 | } | |
9219 | ||
0bc69b81 JR |
9220 | /* Find annulled delay insns and convert them to use the appropriate predicate. |
9221 | This allows branch shortening to size up these insns properly. */ | |
9222 | ||
9223 | static unsigned | |
9224 | arc_predicate_delay_insns (void) | |
9225 | { | |
b3458f61 | 9226 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
0bc69b81 JR |
9227 | { |
9228 | rtx pat, jump, dlay, src, cond, *patp; | |
9229 | int reverse; | |
9230 | ||
9231 | if (!NONJUMP_INSN_P (insn) | |
9232 | || GET_CODE (pat = PATTERN (insn)) != SEQUENCE) | |
9233 | continue; | |
9234 | jump = XVECEXP (pat, 0, 0); | |
9235 | dlay = XVECEXP (pat, 0, 1); | |
9236 | if (!JUMP_P (jump) || !INSN_ANNULLED_BRANCH_P (jump)) | |
9237 | continue; | |
9238 | /* If the branch insn does the annulling, leave the delay insn alone. */ | |
9239 | if (!TARGET_AT_DBR_CONDEXEC && !INSN_FROM_TARGET_P (dlay)) | |
9240 | continue; | |
9241 | /* ??? Could also leave DLAY un-conditionalized if its target is dead | |
9242 | on the other path. */ | |
9243 | gcc_assert (GET_CODE (PATTERN (jump)) == SET); | |
9244 | gcc_assert (SET_DEST (PATTERN (jump)) == pc_rtx); | |
9245 | src = SET_SRC (PATTERN (jump)); | |
9246 | gcc_assert (GET_CODE (src) == IF_THEN_ELSE); | |
9247 | cond = XEXP (src, 0); | |
9248 | if (XEXP (src, 2) == pc_rtx) | |
9249 | reverse = 0; | |
9250 | else if (XEXP (src, 1) == pc_rtx) | |
9251 | reverse = 1; | |
9252 | else | |
9253 | gcc_unreachable (); | |
9af539fe | 9254 | if (reverse != !INSN_FROM_TARGET_P (dlay)) |
0bc69b81 | 9255 | { |
ef4bddc2 | 9256 | machine_mode ccm = GET_MODE (XEXP (cond, 0)); |
0bc69b81 JR |
9257 | enum rtx_code code = reverse_condition (GET_CODE (cond)); |
9258 | if (code == UNKNOWN || ccm == CC_FP_GTmode || ccm == CC_FP_GEmode) | |
9259 | code = reverse_condition_maybe_unordered (GET_CODE (cond)); | |
9260 | ||
9261 | cond = gen_rtx_fmt_ee (code, GET_MODE (cond), | |
9262 | copy_rtx (XEXP (cond, 0)), | |
9263 | copy_rtx (XEXP (cond, 1))); | |
9264 | } | |
9265 | else | |
9266 | cond = copy_rtx (cond); | |
9267 | patp = &PATTERN (dlay); | |
9268 | pat = *patp; | |
eeac7d15 | 9269 | pat = conditionalize_nonjump (pat, cond, dlay, true); |
0bc69b81 JR |
9270 | validate_change (dlay, patp, pat, 1); |
9271 | if (!apply_change_group ()) | |
9272 | gcc_unreachable (); | |
9273 | } | |
9274 | return 0; | |
9275 | } | |
9276 | ||
526b7aee SV |
9277 | /* For ARC600: If a write to a core reg >=32 appears in a delay slot |
9278 | (other than of a forward brcc), it creates a hazard when there is a read | |
9279 | of the same register at the branch target. We can't know what is at the | |
9280 | branch target of calls, and for branches, we don't really know before the | |
9281 | end of delay slot scheduling, either. Not only can individual instruction | |
9282 | be hoisted out into a delay slot, a basic block can also be emptied this | |
9283 | way, and branch and/or fall through targets be redirected. Hence we don't | |
9284 | want such writes in a delay slot. */ | |
526b7aee SV |
9285 | |
9286 | /* Return nonzreo iff INSN writes to an extension core register. */ | |
9287 | ||
9288 | int | |
9289 | arc_write_ext_corereg (rtx insn) | |
9290 | { | |
24dbe738 RS |
9291 | subrtx_iterator::array_type array; |
9292 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
9293 | { | |
9294 | const_rtx x = *iter; | |
9295 | switch (GET_CODE (x)) | |
9296 | { | |
9297 | case SET: case POST_INC: case POST_DEC: case PRE_INC: case PRE_DEC: | |
9298 | break; | |
9299 | default: | |
9300 | /* This is also fine for PRE/POST_MODIFY, because they | |
9301 | contain a SET. */ | |
9302 | continue; | |
9303 | } | |
9304 | const_rtx dest = XEXP (x, 0); | |
9305 | if (REG_P (dest) && REGNO (dest) >= 32 && REGNO (dest) < 61) | |
9306 | return 1; | |
9307 | } | |
9308 | return 0; | |
526b7aee SV |
9309 | } |
9310 | ||
9311 | /* This is like the hook, but returns NULL when it can't / won't generate | |
9312 | a legitimate address. */ | |
9313 | ||
9314 | static rtx | |
9315 | arc_legitimize_address_0 (rtx x, rtx oldx ATTRIBUTE_UNUSED, | |
ef4bddc2 | 9316 | machine_mode mode) |
526b7aee SV |
9317 | { |
9318 | rtx addr, inner; | |
9319 | ||
9320 | if (flag_pic && SYMBOLIC_CONST (x)) | |
9321 | (x) = arc_legitimize_pic_address (x, 0); | |
9322 | addr = x; | |
9323 | if (GET_CODE (addr) == CONST) | |
9324 | addr = XEXP (addr, 0); | |
9325 | if (GET_CODE (addr) == PLUS | |
9326 | && CONST_INT_P (XEXP (addr, 1)) | |
9327 | && ((GET_CODE (XEXP (addr, 0)) == SYMBOL_REF | |
9328 | && !SYMBOL_REF_FUNCTION_P (XEXP (addr, 0))) | |
9329 | || (REG_P (XEXP (addr, 0)) | |
9330 | && (INTVAL (XEXP (addr, 1)) & 252)))) | |
9331 | { | |
9332 | HOST_WIDE_INT offs, upper; | |
9333 | int size = GET_MODE_SIZE (mode); | |
9334 | ||
9335 | offs = INTVAL (XEXP (addr, 1)); | |
9336 | upper = (offs + 256 * size) & ~511 * size; | |
9337 | inner = plus_constant (Pmode, XEXP (addr, 0), upper); | |
9338 | #if 0 /* ??? this produces worse code for EEMBC idctrn01 */ | |
9339 | if (GET_CODE (x) == CONST) | |
9340 | inner = gen_rtx_CONST (Pmode, inner); | |
9341 | #endif | |
9342 | addr = plus_constant (Pmode, force_reg (Pmode, inner), offs - upper); | |
9343 | x = addr; | |
9344 | } | |
9345 | else if (GET_CODE (addr) == SYMBOL_REF && !SYMBOL_REF_FUNCTION_P (addr)) | |
9346 | x = force_reg (Pmode, x); | |
ef4bddc2 | 9347 | if (memory_address_p ((machine_mode) mode, x)) |
526b7aee SV |
9348 | return x; |
9349 | return NULL_RTX; | |
9350 | } | |
9351 | ||
9352 | static rtx | |
ef4bddc2 | 9353 | arc_legitimize_address (rtx orig_x, rtx oldx, machine_mode mode) |
526b7aee | 9354 | { |
28633bbd CZ |
9355 | if (GET_CODE (orig_x) == SYMBOL_REF) |
9356 | { | |
9357 | enum tls_model model = SYMBOL_REF_TLS_MODEL (orig_x); | |
9358 | if (model != 0) | |
9359 | return arc_legitimize_tls_address (orig_x, model); | |
9360 | } | |
9361 | ||
526b7aee SV |
9362 | rtx new_x = arc_legitimize_address_0 (orig_x, oldx, mode); |
9363 | ||
9364 | if (new_x) | |
9365 | return new_x; | |
9366 | return orig_x; | |
9367 | } | |
9368 | ||
9369 | static rtx | |
9370 | arc_delegitimize_address_0 (rtx x) | |
9371 | { | |
f5e336b1 | 9372 | rtx u, gp, p; |
526b7aee SV |
9373 | |
9374 | if (GET_CODE (x) == CONST && GET_CODE (u = XEXP (x, 0)) == UNSPEC) | |
9375 | { | |
f5e336b1 CZ |
9376 | if (XINT (u, 1) == ARC_UNSPEC_GOT |
9377 | || XINT (u, 1) == ARC_UNSPEC_GOTOFFPC) | |
526b7aee SV |
9378 | return XVECEXP (u, 0, 0); |
9379 | } | |
f5e336b1 CZ |
9380 | else if (GET_CODE (x) == CONST && GET_CODE (p = XEXP (x, 0)) == PLUS |
9381 | && GET_CODE (u = XEXP (p, 0)) == UNSPEC | |
9382 | && (XINT (u, 1) == ARC_UNSPEC_GOT | |
9383 | || XINT (u, 1) == ARC_UNSPEC_GOTOFFPC)) | |
9384 | return gen_rtx_CONST | |
9385 | (GET_MODE (x), | |
9386 | gen_rtx_PLUS (GET_MODE (p), XVECEXP (u, 0, 0), XEXP (p, 1))); | |
526b7aee SV |
9387 | else if (GET_CODE (x) == PLUS |
9388 | && ((REG_P (gp = XEXP (x, 0)) | |
9389 | && REGNO (gp) == PIC_OFFSET_TABLE_REGNUM) | |
9390 | || (GET_CODE (gp) == CONST | |
9391 | && GET_CODE (u = XEXP (gp, 0)) == UNSPEC | |
9392 | && XINT (u, 1) == ARC_UNSPEC_GOT | |
9393 | && GET_CODE (XVECEXP (u, 0, 0)) == SYMBOL_REF | |
9394 | && !strcmp (XSTR (XVECEXP (u, 0, 0), 0), "_DYNAMIC"))) | |
9395 | && GET_CODE (XEXP (x, 1)) == CONST | |
9396 | && GET_CODE (u = XEXP (XEXP (x, 1), 0)) == UNSPEC | |
9397 | && XINT (u, 1) == ARC_UNSPEC_GOTOFF) | |
9398 | return XVECEXP (u, 0, 0); | |
9399 | else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS | |
9400 | && ((REG_P (gp = XEXP (XEXP (x, 0), 1)) | |
9401 | && REGNO (gp) == PIC_OFFSET_TABLE_REGNUM) | |
9402 | || (GET_CODE (gp) == CONST | |
9403 | && GET_CODE (u = XEXP (gp, 0)) == UNSPEC | |
9404 | && XINT (u, 1) == ARC_UNSPEC_GOT | |
9405 | && GET_CODE (XVECEXP (u, 0, 0)) == SYMBOL_REF | |
9406 | && !strcmp (XSTR (XVECEXP (u, 0, 0), 0), "_DYNAMIC"))) | |
9407 | && GET_CODE (XEXP (x, 1)) == CONST | |
9408 | && GET_CODE (u = XEXP (XEXP (x, 1), 0)) == UNSPEC | |
9409 | && XINT (u, 1) == ARC_UNSPEC_GOTOFF) | |
9410 | return gen_rtx_PLUS (GET_MODE (x), XEXP (XEXP (x, 0), 0), | |
9411 | XVECEXP (u, 0, 0)); | |
9412 | else if (GET_CODE (x) == PLUS | |
9413 | && (u = arc_delegitimize_address_0 (XEXP (x, 1)))) | |
9414 | return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0), u); | |
9415 | return NULL_RTX; | |
9416 | } | |
9417 | ||
9418 | static rtx | |
9419 | arc_delegitimize_address (rtx x) | |
9420 | { | |
9421 | rtx orig_x = x = delegitimize_mem_from_attrs (x); | |
9422 | if (GET_CODE (x) == MEM) | |
9423 | x = XEXP (x, 0); | |
9424 | x = arc_delegitimize_address_0 (x); | |
9425 | if (x) | |
9426 | { | |
9427 | if (MEM_P (orig_x)) | |
9428 | x = replace_equiv_address_nv (orig_x, x); | |
9429 | return x; | |
9430 | } | |
9431 | return orig_x; | |
9432 | } | |
9433 | ||
9434 | /* Return a REG rtx for acc1. N.B. the gcc-internal representation may | |
9435 | differ from the hardware register number in order to allow the generic | |
9436 | code to correctly split the concatenation of acc1 and acc2. */ | |
9437 | ||
9438 | rtx | |
9439 | gen_acc1 (void) | |
9440 | { | |
9441 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 56: 57); | |
9442 | } | |
9443 | ||
9444 | /* Return a REG rtx for acc2. N.B. the gcc-internal representation may | |
9445 | differ from the hardware register number in order to allow the generic | |
9446 | code to correctly split the concatenation of acc1 and acc2. */ | |
9447 | ||
9448 | rtx | |
9449 | gen_acc2 (void) | |
9450 | { | |
9451 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 57: 56); | |
9452 | } | |
9453 | ||
9454 | /* Return a REG rtx for mlo. N.B. the gcc-internal representation may | |
9455 | differ from the hardware register number in order to allow the generic | |
9456 | code to correctly split the concatenation of mhi and mlo. */ | |
9457 | ||
9458 | rtx | |
9459 | gen_mlo (void) | |
9460 | { | |
9461 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 59: 58); | |
9462 | } | |
9463 | ||
9464 | /* Return a REG rtx for mhi. N.B. the gcc-internal representation may | |
9465 | differ from the hardware register number in order to allow the generic | |
9466 | code to correctly split the concatenation of mhi and mlo. */ | |
9467 | ||
9468 | rtx | |
9469 | gen_mhi (void) | |
9470 | { | |
9471 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 58: 59); | |
9472 | } | |
9473 | ||
9474 | /* FIXME: a parameter should be added, and code added to final.c, | |
9475 | to reproduce this functionality in shorten_branches. */ | |
9476 | #if 0 | |
9477 | /* Return nonzero iff BRANCH should be unaligned if possible by upsizing | |
9478 | a previous instruction. */ | |
9479 | int | |
9480 | arc_unalign_branch_p (rtx branch) | |
9481 | { | |
9482 | rtx note; | |
9483 | ||
9484 | if (!TARGET_UNALIGN_BRANCH) | |
9485 | return 0; | |
9486 | /* Do not do this if we have a filled delay slot. */ | |
9487 | if (get_attr_delay_slot_filled (branch) == DELAY_SLOT_FILLED_YES | |
4654c0cf | 9488 | && !NEXT_INSN (branch)->deleted ()) |
526b7aee SV |
9489 | return 0; |
9490 | note = find_reg_note (branch, REG_BR_PROB, 0); | |
9491 | return (!note | |
9492 | || (arc_unalign_prob_threshold && !br_prob_note_reliable_p (note)) | |
9493 | || INTVAL (XEXP (note, 0)) < arc_unalign_prob_threshold); | |
9494 | } | |
9495 | #endif | |
9496 | ||
9497 | /* When estimating sizes during arc_reorg, when optimizing for speed, there | |
9498 | are three reasons why we need to consider branches to be length 6: | |
9499 | - annull-false delay slot insns are implemented using conditional execution, | |
9500 | thus preventing short insn formation where used. | |
9501 | - for ARC600: annul-true delay slot insns are implemented where possible | |
9502 | using conditional execution, preventing short insn formation where used. | |
9503 | - for ARC700: likely or somewhat likely taken branches are made long and | |
9504 | unaligned if possible to avoid branch penalty. */ | |
9505 | ||
9506 | bool | |
9507 | arc_branch_size_unknown_p (void) | |
9508 | { | |
9509 | return !optimize_size && arc_reorg_in_progress; | |
9510 | } | |
9511 | ||
9512 | /* We are about to output a return insn. Add padding if necessary to avoid | |
9513 | a mispredict. A return could happen immediately after the function | |
9514 | start, but after a call we know that there will be at least a blink | |
9515 | restore. */ | |
9516 | ||
9517 | void | |
9518 | arc_pad_return (void) | |
9519 | { | |
fa7af581 | 9520 | rtx_insn *insn = current_output_insn; |
b3458f61 | 9521 | rtx_insn *prev = prev_active_insn (insn); |
526b7aee SV |
9522 | int want_long; |
9523 | ||
9524 | if (!prev) | |
9525 | { | |
9526 | fputs ("\tnop_s\n", asm_out_file); | |
9527 | cfun->machine->unalign ^= 2; | |
9528 | want_long = 1; | |
9529 | } | |
9530 | /* If PREV is a sequence, we know it must be a branch / jump or a tailcall, | |
9531 | because after a call, we'd have to restore blink first. */ | |
9532 | else if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
9533 | return; | |
9534 | else | |
9535 | { | |
9536 | want_long = (get_attr_length (prev) == 2); | |
9537 | prev = prev_active_insn (prev); | |
9538 | } | |
9539 | if (!prev | |
9540 | || ((NONJUMP_INSN_P (prev) && GET_CODE (PATTERN (prev)) == SEQUENCE) | |
84034c69 DM |
9541 | ? CALL_ATTR (as_a <rtx_sequence *> (PATTERN (prev))->insn (0), |
9542 | NON_SIBCALL) | |
526b7aee SV |
9543 | : CALL_ATTR (prev, NON_SIBCALL))) |
9544 | { | |
9545 | if (want_long) | |
9546 | cfun->machine->size_reason | |
9547 | = "call/return and return/return must be 6 bytes apart to avoid mispredict"; | |
9548 | else if (TARGET_UNALIGN_BRANCH && cfun->machine->unalign) | |
9549 | { | |
9550 | cfun->machine->size_reason | |
9551 | = "Long unaligned jump avoids non-delay slot penalty"; | |
9552 | want_long = 1; | |
9553 | } | |
9554 | /* Disgorge delay insn, if there is any, and it may be moved. */ | |
9555 | if (final_sequence | |
9556 | /* ??? Annulled would be OK if we can and do conditionalize | |
9557 | the delay slot insn accordingly. */ | |
9558 | && !INSN_ANNULLED_BRANCH_P (insn) | |
9559 | && (get_attr_cond (insn) != COND_USE | |
9560 | || !reg_set_p (gen_rtx_REG (CCmode, CC_REG), | |
9561 | XVECEXP (final_sequence, 0, 1)))) | |
9562 | { | |
b3458f61 | 9563 | prev = as_a <rtx_insn *> (XVECEXP (final_sequence, 0, 1)); |
526b7aee SV |
9564 | gcc_assert (!prev_real_insn (insn) |
9565 | || !arc_hazard (prev_real_insn (insn), prev)); | |
9566 | cfun->machine->force_short_suffix = !want_long; | |
9567 | rtx save_pred = current_insn_predicate; | |
9568 | final_scan_insn (prev, asm_out_file, optimize, 1, NULL); | |
9569 | cfun->machine->force_short_suffix = -1; | |
4654c0cf | 9570 | prev->set_deleted (); |
526b7aee SV |
9571 | current_output_insn = insn; |
9572 | current_insn_predicate = save_pred; | |
9573 | } | |
9574 | else if (want_long) | |
9575 | fputs ("\tnop\n", asm_out_file); | |
9576 | else | |
9577 | { | |
9578 | fputs ("\tnop_s\n", asm_out_file); | |
9579 | cfun->machine->unalign ^= 2; | |
9580 | } | |
9581 | } | |
9582 | return; | |
9583 | } | |
9584 | ||
9585 | /* The usual; we set up our machine_function data. */ | |
9586 | ||
9587 | static struct machine_function * | |
9588 | arc_init_machine_status (void) | |
9589 | { | |
9590 | struct machine_function *machine; | |
766090c2 | 9591 | machine = ggc_cleared_alloc<machine_function> (); |
526b7aee SV |
9592 | machine->fn_type = ARC_FUNCTION_UNKNOWN; |
9593 | machine->force_short_suffix = -1; | |
9594 | ||
9595 | return machine; | |
9596 | } | |
9597 | ||
9598 | /* Implements INIT_EXPANDERS. We just set up to call the above | |
9599 | function. */ | |
9600 | ||
9601 | void | |
9602 | arc_init_expanders (void) | |
9603 | { | |
9604 | init_machine_status = arc_init_machine_status; | |
9605 | } | |
9606 | ||
9607 | /* Check if OP is a proper parallel of a millicode call pattern. OFFSET | |
9608 | indicates a number of elements to ignore - that allows to have a | |
9609 | sibcall pattern that starts with (return). LOAD_P is zero for store | |
9610 | multiple (for prologues), and one for load multiples (for epilogues), | |
9611 | and two for load multiples where no final clobber of blink is required. | |
9612 | We also skip the first load / store element since this is supposed to | |
9613 | be checked in the instruction pattern. */ | |
9614 | ||
9615 | int | |
9616 | arc_check_millicode (rtx op, int offset, int load_p) | |
9617 | { | |
9618 | int len = XVECLEN (op, 0) - offset; | |
9619 | int i; | |
9620 | ||
9621 | if (load_p == 2) | |
9622 | { | |
9623 | if (len < 2 || len > 13) | |
9624 | return 0; | |
9625 | load_p = 1; | |
9626 | } | |
9627 | else | |
9628 | { | |
9629 | rtx elt = XVECEXP (op, 0, --len); | |
9630 | ||
9631 | if (GET_CODE (elt) != CLOBBER | |
9632 | || !REG_P (XEXP (elt, 0)) | |
9633 | || REGNO (XEXP (elt, 0)) != RETURN_ADDR_REGNUM | |
9634 | || len < 3 || len > 13) | |
9635 | return 0; | |
9636 | } | |
9637 | for (i = 1; i < len; i++) | |
9638 | { | |
9639 | rtx elt = XVECEXP (op, 0, i + offset); | |
9640 | rtx reg, mem, addr; | |
9641 | ||
9642 | if (GET_CODE (elt) != SET) | |
9643 | return 0; | |
9644 | mem = XEXP (elt, load_p); | |
9645 | reg = XEXP (elt, 1-load_p); | |
9646 | if (!REG_P (reg) || REGNO (reg) != 13U+i || !MEM_P (mem)) | |
9647 | return 0; | |
9648 | addr = XEXP (mem, 0); | |
9649 | if (GET_CODE (addr) != PLUS | |
9650 | || !rtx_equal_p (stack_pointer_rtx, XEXP (addr, 0)) | |
9651 | || !CONST_INT_P (XEXP (addr, 1)) || INTVAL (XEXP (addr, 1)) != i*4) | |
9652 | return 0; | |
9653 | } | |
9654 | return 1; | |
9655 | } | |
9656 | ||
9657 | /* Accessor functions for cfun->machine->unalign. */ | |
9658 | ||
9659 | int | |
9660 | arc_get_unalign (void) | |
9661 | { | |
9662 | return cfun->machine->unalign; | |
9663 | } | |
9664 | ||
9665 | void | |
9666 | arc_clear_unalign (void) | |
9667 | { | |
9668 | if (cfun) | |
9669 | cfun->machine->unalign = 0; | |
9670 | } | |
9671 | ||
9672 | void | |
9673 | arc_toggle_unalign (void) | |
9674 | { | |
9675 | cfun->machine->unalign ^= 2; | |
9676 | } | |
9677 | ||
9678 | /* Operands 0..2 are the operands of a addsi which uses a 12 bit | |
9679 | constant in operand 2, but which would require a LIMM because of | |
9680 | operand mismatch. | |
9681 | operands 3 and 4 are new SET_SRCs for operands 0. */ | |
9682 | ||
9683 | void | |
9684 | split_addsi (rtx *operands) | |
9685 | { | |
9686 | int val = INTVAL (operands[2]); | |
9687 | ||
9688 | /* Try for two short insns first. Lengths being equal, we prefer | |
9689 | expansions with shorter register lifetimes. */ | |
9690 | if (val > 127 && val <= 255 | |
9691 | && satisfies_constraint_Rcq (operands[0])) | |
9692 | { | |
9693 | operands[3] = operands[2]; | |
9694 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[1]); | |
9695 | } | |
9696 | else | |
9697 | { | |
9698 | operands[3] = operands[1]; | |
9699 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[2]); | |
9700 | } | |
9701 | } | |
9702 | ||
9703 | /* Operands 0..2 are the operands of a subsi which uses a 12 bit | |
9704 | constant in operand 1, but which would require a LIMM because of | |
9705 | operand mismatch. | |
9706 | operands 3 and 4 are new SET_SRCs for operands 0. */ | |
9707 | ||
9708 | void | |
9709 | split_subsi (rtx *operands) | |
9710 | { | |
9711 | int val = INTVAL (operands[1]); | |
9712 | ||
9713 | /* Try for two short insns first. Lengths being equal, we prefer | |
9714 | expansions with shorter register lifetimes. */ | |
9715 | if (satisfies_constraint_Rcq (operands[0]) | |
9716 | && satisfies_constraint_Rcq (operands[2])) | |
9717 | { | |
9718 | if (val >= -31 && val <= 127) | |
9719 | { | |
9720 | operands[3] = gen_rtx_NEG (SImode, operands[2]); | |
9721 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[1]); | |
9722 | return; | |
9723 | } | |
9724 | else if (val >= 0 && val < 255) | |
9725 | { | |
9726 | operands[3] = operands[1]; | |
9727 | operands[4] = gen_rtx_MINUS (SImode, operands[0], operands[2]); | |
9728 | return; | |
9729 | } | |
9730 | } | |
9731 | /* If the destination is not an ARCompact16 register, we might | |
9732 | still have a chance to make a short insn if the source is; | |
9733 | we need to start with a reg-reg move for this. */ | |
9734 | operands[3] = operands[2]; | |
9735 | operands[4] = gen_rtx_MINUS (SImode, operands[1], operands[0]); | |
9736 | } | |
9737 | ||
9738 | /* Handle DOUBLE_REGS uses. | |
9739 | Operand 0: destination register | |
9740 | Operand 1: source register */ | |
9741 | ||
d34a0fdc | 9742 | static bool |
526b7aee SV |
9743 | arc_process_double_reg_moves (rtx *operands) |
9744 | { | |
9745 | rtx dest = operands[0]; | |
9746 | rtx src = operands[1]; | |
526b7aee SV |
9747 | |
9748 | enum usesDxState { none, srcDx, destDx, maxDx }; | |
9749 | enum usesDxState state = none; | |
9750 | ||
9751 | if (refers_to_regno_p (40, 44, src, 0)) | |
9752 | state = srcDx; | |
9753 | if (refers_to_regno_p (40, 44, dest, 0)) | |
9754 | { | |
9755 | /* Via arc_register_move_cost, we should never see D,D moves. */ | |
9756 | gcc_assert (state == none); | |
9757 | state = destDx; | |
9758 | } | |
9759 | ||
9760 | if (state == none) | |
d34a0fdc | 9761 | return false; |
526b7aee SV |
9762 | |
9763 | if (state == srcDx) | |
9764 | { | |
9765 | /* Without the LR insn, we need to split this into a | |
9766 | sequence of insns which will use the DEXCLx and DADDHxy | |
9767 | insns to be able to read the Dx register in question. */ | |
9768 | if (TARGET_DPFP_DISABLE_LRSR) | |
9769 | { | |
9770 | /* gen *movdf_insn_nolrsr */ | |
f7df4a84 | 9771 | rtx set = gen_rtx_SET (dest, src); |
526b7aee SV |
9772 | rtx use1 = gen_rtx_USE (VOIDmode, const1_rtx); |
9773 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set, use1))); | |
9774 | } | |
9775 | else | |
9776 | { | |
9777 | /* When we have 'mov D, r' or 'mov D, D' then get the target | |
9778 | register pair for use with LR insn. */ | |
7d81a567 CZ |
9779 | rtx destHigh = simplify_gen_subreg (SImode, dest, DFmode, |
9780 | TARGET_BIG_ENDIAN ? 0 : 4); | |
9781 | rtx destLow = simplify_gen_subreg (SImode, dest, DFmode, | |
9782 | TARGET_BIG_ENDIAN ? 4 : 0); | |
526b7aee SV |
9783 | |
9784 | /* Produce the two LR insns to get the high and low parts. */ | |
f7df4a84 | 9785 | emit_insn (gen_rtx_SET (destHigh, |
c69899f0 CZ |
9786 | gen_rtx_UNSPEC_VOLATILE (Pmode, |
9787 | gen_rtvec (1, src), | |
9788 | VUNSPEC_ARC_LR_HIGH))); | |
f7df4a84 | 9789 | emit_insn (gen_rtx_SET (destLow, |
c69899f0 CZ |
9790 | gen_rtx_UNSPEC_VOLATILE (Pmode, |
9791 | gen_rtvec (1, src), | |
9792 | VUNSPEC_ARC_LR))); | |
526b7aee SV |
9793 | } |
9794 | } | |
9795 | else if (state == destDx) | |
9796 | { | |
9797 | /* When we have 'mov r, D' or 'mov D, D' and we have access to the | |
9798 | LR insn get the target register pair. */ | |
7d81a567 CZ |
9799 | rtx srcHigh = simplify_gen_subreg (SImode, src, DFmode, |
9800 | TARGET_BIG_ENDIAN ? 0 : 4); | |
9801 | rtx srcLow = simplify_gen_subreg (SImode, src, DFmode, | |
9802 | TARGET_BIG_ENDIAN ? 4 : 0); | |
526b7aee | 9803 | |
491483b0 | 9804 | emit_insn (gen_dexcl_2op (dest, srcHigh, srcLow)); |
526b7aee SV |
9805 | } |
9806 | else | |
9807 | gcc_unreachable (); | |
9808 | ||
d34a0fdc | 9809 | return true; |
526b7aee SV |
9810 | } |
9811 | ||
9812 | /* operands 0..1 are the operands of a 64 bit move instruction. | |
9813 | split it into two moves with operands 2/3 and 4/5. */ | |
9814 | ||
d34a0fdc | 9815 | void |
526b7aee SV |
9816 | arc_split_move (rtx *operands) |
9817 | { | |
ef4bddc2 | 9818 | machine_mode mode = GET_MODE (operands[0]); |
526b7aee SV |
9819 | int i; |
9820 | int swap = 0; | |
9821 | rtx xop[4]; | |
526b7aee SV |
9822 | |
9823 | if (TARGET_DPFP) | |
9824 | { | |
d34a0fdc CZ |
9825 | if (arc_process_double_reg_moves (operands)) |
9826 | return; | |
526b7aee SV |
9827 | } |
9828 | ||
d34a0fdc CZ |
9829 | if (TARGET_LL64 |
9830 | && ((memory_operand (operands[0], mode) | |
9831 | && even_register_operand (operands[1], mode)) | |
9832 | || (memory_operand (operands[1], mode) | |
9833 | && even_register_operand (operands[0], mode)))) | |
9834 | { | |
9835 | emit_move_insn (operands[0], operands[1]); | |
9836 | return; | |
9837 | } | |
9838 | ||
00c072ae CZ |
9839 | if (TARGET_PLUS_QMACW |
9840 | && GET_CODE (operands[1]) == CONST_VECTOR) | |
9841 | { | |
9842 | HOST_WIDE_INT intval0, intval1; | |
9843 | if (GET_MODE (operands[1]) == V2SImode) | |
9844 | { | |
9845 | intval0 = INTVAL (XVECEXP (operands[1], 0, 0)); | |
9846 | intval1 = INTVAL (XVECEXP (operands[1], 0, 1)); | |
9847 | } | |
9848 | else | |
9849 | { | |
9850 | intval1 = INTVAL (XVECEXP (operands[1], 0, 3)) << 16; | |
9851 | intval1 |= INTVAL (XVECEXP (operands[1], 0, 2)) & 0xFFFF; | |
9852 | intval0 = INTVAL (XVECEXP (operands[1], 0, 1)) << 16; | |
9853 | intval0 |= INTVAL (XVECEXP (operands[1], 0, 0)) & 0xFFFF; | |
9854 | } | |
9855 | xop[0] = gen_rtx_REG (SImode, REGNO (operands[0])); | |
9856 | xop[3] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); | |
9857 | xop[2] = GEN_INT (trunc_int_for_mode (intval0, SImode)); | |
9858 | xop[1] = GEN_INT (trunc_int_for_mode (intval1, SImode)); | |
9859 | emit_move_insn (xop[0], xop[2]); | |
9860 | emit_move_insn (xop[3], xop[1]); | |
9861 | return; | |
9862 | } | |
9863 | ||
526b7aee SV |
9864 | for (i = 0; i < 2; i++) |
9865 | { | |
9866 | if (MEM_P (operands[i]) && auto_inc_p (XEXP (operands[i], 0))) | |
9867 | { | |
9868 | rtx addr = XEXP (operands[i], 0); | |
9869 | rtx r, o; | |
9870 | enum rtx_code code; | |
9871 | ||
9872 | gcc_assert (!reg_overlap_mentioned_p (operands[0], addr)); | |
9873 | switch (GET_CODE (addr)) | |
9874 | { | |
9875 | case PRE_DEC: o = GEN_INT (-8); goto pre_modify; | |
9876 | case PRE_INC: o = GEN_INT (8); goto pre_modify; | |
9877 | case PRE_MODIFY: o = XEXP (XEXP (addr, 1), 1); | |
9878 | pre_modify: | |
9879 | code = PRE_MODIFY; | |
9880 | break; | |
9881 | case POST_DEC: o = GEN_INT (-8); goto post_modify; | |
9882 | case POST_INC: o = GEN_INT (8); goto post_modify; | |
9883 | case POST_MODIFY: o = XEXP (XEXP (addr, 1), 1); | |
9884 | post_modify: | |
9885 | code = POST_MODIFY; | |
9886 | swap = 2; | |
9887 | break; | |
9888 | default: | |
9889 | gcc_unreachable (); | |
9890 | } | |
9891 | r = XEXP (addr, 0); | |
9892 | xop[0+i] = adjust_automodify_address_nv | |
9893 | (operands[i], SImode, | |
9894 | gen_rtx_fmt_ee (code, Pmode, r, | |
9895 | gen_rtx_PLUS (Pmode, r, o)), | |
9896 | 0); | |
9897 | xop[2+i] = adjust_automodify_address_nv | |
9898 | (operands[i], SImode, plus_constant (Pmode, r, 4), 4); | |
9899 | } | |
9900 | else | |
9901 | { | |
9902 | xop[0+i] = operand_subword (operands[i], 0, 0, mode); | |
9903 | xop[2+i] = operand_subword (operands[i], 1, 0, mode); | |
9904 | } | |
9905 | } | |
9906 | if (reg_overlap_mentioned_p (xop[0], xop[3])) | |
9907 | { | |
9908 | swap = 2; | |
9909 | gcc_assert (!reg_overlap_mentioned_p (xop[2], xop[1])); | |
9910 | } | |
526b7aee | 9911 | |
d34a0fdc CZ |
9912 | emit_move_insn (xop[0 + swap], xop[1 + swap]); |
9913 | emit_move_insn (xop[2 - swap], xop[3 - swap]); | |
526b7aee | 9914 | |
526b7aee SV |
9915 | } |
9916 | ||
9917 | /* Select between the instruction output templates s_tmpl (for short INSNs) | |
9918 | and l_tmpl (for long INSNs). */ | |
9919 | ||
9920 | const char * | |
b3458f61 | 9921 | arc_short_long (rtx_insn *insn, const char *s_tmpl, const char *l_tmpl) |
526b7aee SV |
9922 | { |
9923 | int is_short = arc_verify_short (insn, cfun->machine->unalign, -1); | |
9924 | ||
9925 | extract_constrain_insn_cached (insn); | |
9926 | return is_short ? s_tmpl : l_tmpl; | |
9927 | } | |
9928 | ||
9929 | /* Searches X for any reference to REGNO, returning the rtx of the | |
9930 | reference found if any. Otherwise, returns NULL_RTX. */ | |
9931 | ||
9932 | rtx | |
9933 | arc_regno_use_in (unsigned int regno, rtx x) | |
9934 | { | |
9935 | const char *fmt; | |
9936 | int i, j; | |
9937 | rtx tem; | |
9938 | ||
c9bd6bcd | 9939 | if (REG_P (x) && refers_to_regno_p (regno, x)) |
526b7aee SV |
9940 | return x; |
9941 | ||
9942 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
9943 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
9944 | { | |
9945 | if (fmt[i] == 'e') | |
9946 | { | |
9947 | if ((tem = regno_use_in (regno, XEXP (x, i)))) | |
9948 | return tem; | |
9949 | } | |
9950 | else if (fmt[i] == 'E') | |
9951 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9952 | if ((tem = regno_use_in (regno , XVECEXP (x, i, j)))) | |
9953 | return tem; | |
9954 | } | |
9955 | ||
9956 | return NULL_RTX; | |
9957 | } | |
9958 | ||
9959 | /* Return the integer value of the "type" attribute for INSN, or -1 if | |
9960 | INSN can't have attributes. */ | |
9961 | ||
9962 | int | |
84034c69 | 9963 | arc_attr_type (rtx_insn *insn) |
526b7aee SV |
9964 | { |
9965 | if (NONJUMP_INSN_P (insn) | |
9966 | ? (GET_CODE (PATTERN (insn)) == USE | |
9967 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
9968 | : JUMP_P (insn) | |
9969 | ? (GET_CODE (PATTERN (insn)) == ADDR_VEC | |
9970 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
9971 | : !CALL_P (insn)) | |
9972 | return -1; | |
9973 | return get_attr_type (insn); | |
9974 | } | |
9975 | ||
9976 | /* Return true if insn sets the condition codes. */ | |
9977 | ||
9978 | bool | |
84034c69 | 9979 | arc_sets_cc_p (rtx_insn *insn) |
526b7aee | 9980 | { |
84034c69 DM |
9981 | if (NONJUMP_INSN_P (insn)) |
9982 | if (rtx_sequence *seq = dyn_cast <rtx_sequence *> (PATTERN (insn))) | |
9983 | insn = seq->insn (seq->len () - 1); | |
526b7aee SV |
9984 | return arc_attr_type (insn) == TYPE_COMPARE; |
9985 | } | |
9986 | ||
9987 | /* Return true if INSN is an instruction with a delay slot we may want | |
9988 | to fill. */ | |
9989 | ||
9990 | bool | |
b3458f61 | 9991 | arc_need_delay (rtx_insn *insn) |
526b7aee | 9992 | { |
b3458f61 | 9993 | rtx_insn *next; |
526b7aee SV |
9994 | |
9995 | if (!flag_delayed_branch) | |
9996 | return false; | |
9997 | /* The return at the end of a function needs a delay slot. */ | |
9998 | if (NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE | |
9999 | && (!(next = next_active_insn (insn)) | |
10000 | || ((!NONJUMP_INSN_P (next) || GET_CODE (PATTERN (next)) != SEQUENCE) | |
10001 | && arc_attr_type (next) == TYPE_RETURN)) | |
10002 | && (!TARGET_PAD_RETURN | |
10003 | || (prev_active_insn (insn) | |
10004 | && prev_active_insn (prev_active_insn (insn)) | |
10005 | && prev_active_insn (prev_active_insn (prev_active_insn (insn)))))) | |
10006 | return true; | |
10007 | if (NONJUMP_INSN_P (insn) | |
10008 | ? (GET_CODE (PATTERN (insn)) == USE | |
10009 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
10010 | || GET_CODE (PATTERN (insn)) == SEQUENCE) | |
10011 | : JUMP_P (insn) | |
10012 | ? (GET_CODE (PATTERN (insn)) == ADDR_VEC | |
10013 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
10014 | : !CALL_P (insn)) | |
10015 | return false; | |
10016 | return num_delay_slots (insn) != 0; | |
10017 | } | |
10018 | ||
10019 | /* Return true if the scheduling pass(es) has/have already run, | |
10020 | i.e. where possible, we should try to mitigate high latencies | |
10021 | by different instruction selection. */ | |
10022 | ||
10023 | bool | |
10024 | arc_scheduling_not_expected (void) | |
10025 | { | |
10026 | return cfun->machine->arc_reorg_started; | |
10027 | } | |
10028 | ||
526b7aee | 10029 | int |
82082f65 | 10030 | arc_label_align (rtx_insn *label) |
526b7aee | 10031 | { |
526b7aee SV |
10032 | /* Code has a minimum p2 alignment of 1, which we must restore after an |
10033 | ADDR_DIFF_VEC. */ | |
10034 | if (align_labels_log < 1) | |
10035 | { | |
b3458f61 | 10036 | rtx_insn *next = next_nonnote_nondebug_insn (label); |
526b7aee SV |
10037 | if (INSN_P (next) && recog_memoized (next) >= 0) |
10038 | return 1; | |
10039 | } | |
10040 | return align_labels_log; | |
10041 | } | |
10042 | ||
10043 | /* Return true if LABEL is in executable code. */ | |
10044 | ||
10045 | bool | |
b32d5189 | 10046 | arc_text_label (rtx_insn *label) |
526b7aee | 10047 | { |
b3458f61 | 10048 | rtx_insn *next; |
526b7aee SV |
10049 | |
10050 | /* ??? We use deleted labels like they were still there, see | |
10051 | gcc.c-torture/compile/20000326-2.c . */ | |
10052 | gcc_assert (GET_CODE (label) == CODE_LABEL | |
10053 | || (GET_CODE (label) == NOTE | |
10054 | && NOTE_KIND (label) == NOTE_INSN_DELETED_LABEL)); | |
10055 | next = next_nonnote_insn (label); | |
10056 | if (next) | |
10057 | return (!JUMP_TABLE_DATA_P (next) | |
10058 | || GET_CODE (PATTERN (next)) != ADDR_VEC); | |
10059 | else if (!PREV_INSN (label)) | |
10060 | /* ??? sometimes text labels get inserted very late, see | |
10061 | gcc.dg/torture/stackalign/comp-goto-1.c */ | |
10062 | return true; | |
10063 | return false; | |
10064 | } | |
10065 | ||
526b7aee SV |
10066 | /* Without this, gcc.dg/tree-prof/bb-reorg.c fails to assemble |
10067 | when compiling with -O2 -freorder-blocks-and-partition -fprofile-use | |
339ba33b | 10068 | -D_PROFILE_USE; delay branch scheduling then follows a crossing jump |
526b7aee SV |
10069 | to redirect two breqs. */ |
10070 | ||
10071 | static bool | |
c1ce59ab | 10072 | arc_can_follow_jump (const rtx_insn *follower, const rtx_insn *followee) |
526b7aee SV |
10073 | { |
10074 | /* ??? get_attr_type is declared to take an rtx. */ | |
c1ce59ab | 10075 | union { const rtx_insn *c; rtx_insn *r; } u; |
526b7aee SV |
10076 | |
10077 | u.c = follower; | |
339ba33b | 10078 | if (CROSSING_JUMP_P (followee)) |
526b7aee SV |
10079 | switch (get_attr_type (u.r)) |
10080 | { | |
28f4ff35 CZ |
10081 | case TYPE_BRANCH: |
10082 | if (get_attr_length (u.r) != 2) | |
10083 | break; | |
41bc2c0b | 10084 | /* Fall through. */ |
526b7aee SV |
10085 | case TYPE_BRCC: |
10086 | case TYPE_BRCC_NO_DELAY_SLOT: | |
10087 | return false; | |
10088 | default: | |
10089 | return true; | |
10090 | } | |
10091 | return true; | |
10092 | } | |
10093 | ||
1825c61e CZ |
10094 | /* Return the register number of the register holding the return address |
10095 | for a function of type TYPE. */ | |
10096 | ||
10097 | int | |
10098 | arc_return_address_register (unsigned int fn_type) | |
10099 | { | |
10100 | int regno = 0; | |
10101 | ||
10102 | if (ARC_INTERRUPT_P (fn_type)) | |
10103 | { | |
10104 | if (((fn_type & ARC_FUNCTION_ILINK1) | ARC_FUNCTION_FIRQ) != 0) | |
10105 | regno = ILINK1_REGNUM; | |
10106 | else if ((fn_type & ARC_FUNCTION_ILINK2) != 0) | |
10107 | regno = ILINK2_REGNUM; | |
10108 | else | |
10109 | gcc_unreachable (); | |
10110 | } | |
10111 | else if (ARC_NORMAL_P (fn_type) || ARC_NAKED_P (fn_type)) | |
10112 | regno = RETURN_ADDR_REGNUM; | |
10113 | ||
10114 | gcc_assert (regno != 0); | |
10115 | return regno; | |
10116 | } | |
c7314bc1 | 10117 | |
1825c61e | 10118 | /* Implement EPILOGUE_USES. |
526b7aee SV |
10119 | Return true if REGNO should be added to the deemed uses of the epilogue. |
10120 | ||
1825c61e CZ |
10121 | We have to make sure all the register restore instructions are |
10122 | known to be live in interrupt functions, plus the blink register if | |
10123 | it is clobbered by the isr. */ | |
526b7aee SV |
10124 | |
10125 | bool | |
10126 | arc_epilogue_uses (int regno) | |
10127 | { | |
1825c61e CZ |
10128 | unsigned int fn_type; |
10129 | ||
28633bbd CZ |
10130 | if (regno == arc_tp_regno) |
10131 | return true; | |
1825c61e CZ |
10132 | |
10133 | fn_type = arc_compute_function_type (cfun); | |
526b7aee SV |
10134 | if (reload_completed) |
10135 | { | |
10136 | if (ARC_INTERRUPT_P (cfun->machine->fn_type)) | |
10137 | { | |
10138 | if (!fixed_regs[regno]) | |
10139 | return true; | |
1825c61e | 10140 | return ((regno == arc_return_address_register (fn_type)) |
84804c5b | 10141 | || (regno == RETURN_ADDR_REGNUM)); |
526b7aee SV |
10142 | } |
10143 | else | |
10144 | return regno == RETURN_ADDR_REGNUM; | |
10145 | } | |
10146 | else | |
1825c61e | 10147 | return regno == arc_return_address_register (fn_type); |
526b7aee SV |
10148 | } |
10149 | ||
28633bbd CZ |
10150 | /* Helper for EH_USES macro. */ |
10151 | ||
10152 | bool | |
10153 | arc_eh_uses (int regno) | |
10154 | { | |
10155 | if (regno == arc_tp_regno) | |
10156 | return true; | |
10157 | return false; | |
10158 | } | |
10159 | ||
526b7aee SV |
10160 | #ifndef TARGET_NO_LRA |
10161 | #define TARGET_NO_LRA !TARGET_LRA | |
10162 | #endif | |
10163 | ||
10164 | static bool | |
10165 | arc_lra_p (void) | |
10166 | { | |
10167 | return !TARGET_NO_LRA; | |
10168 | } | |
10169 | ||
10170 | /* ??? Should we define TARGET_REGISTER_PRIORITY? We might perfer to use | |
10171 | Rcq registers, because some insn are shorter with them. OTOH we already | |
10172 | have separate alternatives for this purpose, and other insns don't | |
10173 | mind, so maybe we should rather prefer the other registers? | |
10174 | We need more data, and we can only get that if we allow people to | |
10175 | try all options. */ | |
10176 | static int | |
10177 | arc_register_priority (int r) | |
10178 | { | |
10179 | switch (arc_lra_priority_tag) | |
10180 | { | |
10181 | case ARC_LRA_PRIORITY_NONE: | |
10182 | return 0; | |
10183 | case ARC_LRA_PRIORITY_NONCOMPACT: | |
10184 | return ((((r & 7) ^ 4) - 4) & 15) != r; | |
10185 | case ARC_LRA_PRIORITY_COMPACT: | |
10186 | return ((((r & 7) ^ 4) - 4) & 15) == r; | |
10187 | default: | |
10188 | gcc_unreachable (); | |
10189 | } | |
10190 | } | |
10191 | ||
10192 | static reg_class_t | |
ef4bddc2 | 10193 | arc_spill_class (reg_class_t /* orig_class */, machine_mode) |
526b7aee SV |
10194 | { |
10195 | return GENERAL_REGS; | |
10196 | } | |
10197 | ||
10198 | bool | |
ef4bddc2 | 10199 | arc_legitimize_reload_address (rtx *p, machine_mode mode, int opnum, |
526b7aee SV |
10200 | int itype) |
10201 | { | |
10202 | rtx x = *p; | |
10203 | enum reload_type type = (enum reload_type) itype; | |
10204 | ||
10205 | if (GET_CODE (x) == PLUS | |
10206 | && CONST_INT_P (XEXP (x, 1)) | |
10207 | && (RTX_OK_FOR_BASE_P (XEXP (x, 0), true) | |
10208 | || (REG_P (XEXP (x, 0)) | |
10209 | && reg_equiv_constant (REGNO (XEXP (x, 0)))))) | |
10210 | { | |
10211 | int scale = GET_MODE_SIZE (mode); | |
10212 | int shift; | |
10213 | rtx index_rtx = XEXP (x, 1); | |
10214 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
10215 | rtx reg, sum, sum2; | |
10216 | ||
10217 | if (scale > 4) | |
10218 | scale = 4; | |
10219 | if ((scale-1) & offset) | |
10220 | scale = 1; | |
10221 | shift = scale >> 1; | |
c419f71c JL |
10222 | offset_base |
10223 | = ((offset + (256 << shift)) | |
4e671509 | 10224 | & ((HOST_WIDE_INT)((unsigned HOST_WIDE_INT) -512 << shift))); |
526b7aee SV |
10225 | /* Sometimes the normal form does not suit DImode. We |
10226 | could avoid that by using smaller ranges, but that | |
10227 | would give less optimized code when SImode is | |
10228 | prevalent. */ | |
10229 | if (GET_MODE_SIZE (mode) + offset - offset_base <= (256 << shift)) | |
10230 | { | |
10231 | int regno; | |
10232 | ||
10233 | reg = XEXP (x, 0); | |
10234 | regno = REGNO (reg); | |
10235 | sum2 = sum = plus_constant (Pmode, reg, offset_base); | |
10236 | ||
10237 | if (reg_equiv_constant (regno)) | |
10238 | { | |
10239 | sum2 = plus_constant (Pmode, reg_equiv_constant (regno), | |
10240 | offset_base); | |
10241 | if (GET_CODE (sum2) == PLUS) | |
10242 | sum2 = gen_rtx_CONST (Pmode, sum2); | |
10243 | } | |
10244 | *p = gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
10245 | push_reload (sum2, NULL_RTX, &XEXP (*p, 0), NULL, | |
10246 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, | |
10247 | type); | |
10248 | return true; | |
10249 | } | |
10250 | } | |
10251 | /* We must re-recognize what we created before. */ | |
10252 | else if (GET_CODE (x) == PLUS | |
10253 | && GET_CODE (XEXP (x, 0)) == PLUS | |
10254 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
10255 | && REG_P (XEXP (XEXP (x, 0), 0)) | |
10256 | && CONST_INT_P (XEXP (x, 1))) | |
10257 | { | |
10258 | /* Because this address is so complex, we know it must have | |
10259 | been created by LEGITIMIZE_RELOAD_ADDRESS before; thus, | |
10260 | it is already unshared, and needs no further unsharing. */ | |
10261 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
10262 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
10263 | return true; | |
10264 | } | |
10265 | return false; | |
10266 | } | |
10267 | ||
ad23f5d4 JG |
10268 | /* Implement TARGET_USE_BY_PIECES_INFRASTRUCTURE_P. */ |
10269 | ||
10270 | static bool | |
445d7826 | 10271 | arc_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
ad23f5d4 JG |
10272 | unsigned int align, |
10273 | enum by_pieces_operation op, | |
10274 | bool speed_p) | |
10275 | { | |
10276 | /* Let the movmem expander handle small block moves. */ | |
10277 | if (op == MOVE_BY_PIECES) | |
10278 | return false; | |
10279 | ||
10280 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
10281 | } | |
10282 | ||
b8a64b7f CZ |
10283 | /* Emit a (pre) memory barrier around an atomic sequence according to |
10284 | MODEL. */ | |
10285 | ||
10286 | static void | |
10287 | arc_pre_atomic_barrier (enum memmodel model) | |
10288 | { | |
10289 | if (need_atomic_barrier_p (model, true)) | |
10290 | emit_insn (gen_memory_barrier ()); | |
10291 | } | |
10292 | ||
10293 | /* Emit a (post) memory barrier around an atomic sequence according to | |
10294 | MODEL. */ | |
10295 | ||
10296 | static void | |
10297 | arc_post_atomic_barrier (enum memmodel model) | |
10298 | { | |
10299 | if (need_atomic_barrier_p (model, false)) | |
10300 | emit_insn (gen_memory_barrier ()); | |
10301 | } | |
10302 | ||
10303 | /* Expand a compare and swap pattern. */ | |
10304 | ||
10305 | static void | |
10306 | emit_unlikely_jump (rtx insn) | |
10307 | { | |
f370536c | 10308 | rtx_insn *jump = emit_jump_insn (insn); |
5fa396ad | 10309 | add_reg_br_prob_note (jump, profile_probability::very_unlikely ()); |
b8a64b7f CZ |
10310 | } |
10311 | ||
10312 | /* Expand code to perform a 8 or 16-bit compare and swap by doing | |
10313 | 32-bit compare and swap on the word containing the byte or | |
10314 | half-word. The difference between a weak and a strong CAS is that | |
10315 | the weak version may simply fail. The strong version relies on two | |
10316 | loops, one checks if the SCOND op is succsfully or not, the other | |
10317 | checks if the 32 bit accessed location which contains the 8 or 16 | |
10318 | bit datum is not changed by other thread. The first loop is | |
10319 | implemented by the atomic_compare_and_swapsi_1 pattern. The second | |
10320 | loops is implemented by this routine. */ | |
10321 | ||
10322 | static void | |
10323 | arc_expand_compare_and_swap_qh (rtx bool_result, rtx result, rtx mem, | |
10324 | rtx oldval, rtx newval, rtx weak, | |
10325 | rtx mod_s, rtx mod_f) | |
10326 | { | |
10327 | rtx addr1 = force_reg (Pmode, XEXP (mem, 0)); | |
10328 | rtx addr = gen_reg_rtx (Pmode); | |
10329 | rtx off = gen_reg_rtx (SImode); | |
10330 | rtx oldv = gen_reg_rtx (SImode); | |
10331 | rtx newv = gen_reg_rtx (SImode); | |
10332 | rtx oldvalue = gen_reg_rtx (SImode); | |
10333 | rtx newvalue = gen_reg_rtx (SImode); | |
10334 | rtx res = gen_reg_rtx (SImode); | |
10335 | rtx resv = gen_reg_rtx (SImode); | |
10336 | rtx memsi, val, mask, end_label, loop_label, cc, x; | |
10337 | machine_mode mode; | |
10338 | bool is_weak = (weak != const0_rtx); | |
10339 | ||
10340 | /* Truncate the address. */ | |
10341 | emit_insn (gen_rtx_SET (addr, | |
10342 | gen_rtx_AND (Pmode, addr1, GEN_INT (-4)))); | |
10343 | ||
10344 | /* Compute the datum offset. */ | |
10345 | emit_insn (gen_rtx_SET (off, | |
10346 | gen_rtx_AND (SImode, addr1, GEN_INT (3)))); | |
10347 | if (TARGET_BIG_ENDIAN) | |
10348 | emit_insn (gen_rtx_SET (off, | |
10349 | gen_rtx_MINUS (SImode, | |
10350 | (GET_MODE (mem) == QImode) ? | |
10351 | GEN_INT (3) : GEN_INT (2), off))); | |
10352 | ||
10353 | /* Normal read from truncated address. */ | |
10354 | memsi = gen_rtx_MEM (SImode, addr); | |
10355 | set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER); | |
10356 | MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem); | |
10357 | ||
10358 | val = copy_to_reg (memsi); | |
10359 | ||
10360 | /* Convert the offset in bits. */ | |
10361 | emit_insn (gen_rtx_SET (off, | |
10362 | gen_rtx_ASHIFT (SImode, off, GEN_INT (3)))); | |
10363 | ||
10364 | /* Get the proper mask. */ | |
10365 | if (GET_MODE (mem) == QImode) | |
10366 | mask = force_reg (SImode, GEN_INT (0xff)); | |
10367 | else | |
10368 | mask = force_reg (SImode, GEN_INT (0xffff)); | |
10369 | ||
10370 | emit_insn (gen_rtx_SET (mask, | |
10371 | gen_rtx_ASHIFT (SImode, mask, off))); | |
10372 | ||
10373 | /* Prepare the old and new values. */ | |
10374 | emit_insn (gen_rtx_SET (val, | |
10375 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
10376 | val))); | |
10377 | ||
10378 | oldval = gen_lowpart (SImode, oldval); | |
10379 | emit_insn (gen_rtx_SET (oldv, | |
10380 | gen_rtx_ASHIFT (SImode, oldval, off))); | |
10381 | ||
10382 | newval = gen_lowpart_common (SImode, newval); | |
10383 | emit_insn (gen_rtx_SET (newv, | |
10384 | gen_rtx_ASHIFT (SImode, newval, off))); | |
10385 | ||
10386 | emit_insn (gen_rtx_SET (oldv, | |
10387 | gen_rtx_AND (SImode, oldv, mask))); | |
10388 | ||
10389 | emit_insn (gen_rtx_SET (newv, | |
10390 | gen_rtx_AND (SImode, newv, mask))); | |
10391 | ||
10392 | if (!is_weak) | |
10393 | { | |
10394 | end_label = gen_label_rtx (); | |
10395 | loop_label = gen_label_rtx (); | |
10396 | emit_label (loop_label); | |
10397 | } | |
10398 | ||
10399 | /* Make the old and new values. */ | |
10400 | emit_insn (gen_rtx_SET (oldvalue, | |
10401 | gen_rtx_IOR (SImode, oldv, val))); | |
10402 | ||
10403 | emit_insn (gen_rtx_SET (newvalue, | |
10404 | gen_rtx_IOR (SImode, newv, val))); | |
10405 | ||
10406 | /* Try an 32bit atomic compare and swap. It clobbers the CC | |
10407 | register. */ | |
10408 | emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue, | |
10409 | weak, mod_s, mod_f)); | |
10410 | ||
10411 | /* Regardless of the weakness of the operation, a proper boolean | |
10412 | result needs to be provided. */ | |
10413 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10414 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
10415 | emit_insn (gen_rtx_SET (bool_result, x)); | |
10416 | ||
10417 | if (!is_weak) | |
10418 | { | |
10419 | /* Check the results: if the atomic op is successfully the goto | |
10420 | to end label. */ | |
10421 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10422 | x = gen_rtx_EQ (VOIDmode, x, const0_rtx); | |
10423 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10424 | gen_rtx_LABEL_REF (Pmode, end_label), pc_rtx); | |
10425 | emit_jump_insn (gen_rtx_SET (pc_rtx, x)); | |
10426 | ||
10427 | /* Wait for the right moment when the accessed 32-bit location | |
10428 | is stable. */ | |
10429 | emit_insn (gen_rtx_SET (resv, | |
10430 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
10431 | res))); | |
10432 | mode = SELECT_CC_MODE (NE, resv, val); | |
10433 | cc = gen_rtx_REG (mode, CC_REG); | |
10434 | emit_insn (gen_rtx_SET (cc, gen_rtx_COMPARE (mode, resv, val))); | |
10435 | ||
10436 | /* Set the new value of the 32 bit location, proper masked. */ | |
10437 | emit_insn (gen_rtx_SET (val, resv)); | |
10438 | ||
10439 | /* Try again if location is unstable. Fall through if only | |
10440 | scond op failed. */ | |
10441 | x = gen_rtx_NE (VOIDmode, cc, const0_rtx); | |
10442 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10443 | gen_rtx_LABEL_REF (Pmode, loop_label), pc_rtx); | |
10444 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10445 | ||
10446 | emit_label (end_label); | |
10447 | } | |
10448 | ||
10449 | /* End: proper return the result for the given mode. */ | |
10450 | emit_insn (gen_rtx_SET (res, | |
10451 | gen_rtx_AND (SImode, res, mask))); | |
10452 | ||
10453 | emit_insn (gen_rtx_SET (res, | |
10454 | gen_rtx_LSHIFTRT (SImode, res, off))); | |
10455 | ||
10456 | emit_move_insn (result, gen_lowpart (GET_MODE (result), res)); | |
10457 | } | |
10458 | ||
10459 | /* Helper function used by "atomic_compare_and_swap" expand | |
10460 | pattern. */ | |
10461 | ||
10462 | void | |
10463 | arc_expand_compare_and_swap (rtx operands[]) | |
10464 | { | |
10465 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
10466 | machine_mode mode; | |
10467 | ||
10468 | bval = operands[0]; | |
10469 | rval = operands[1]; | |
10470 | mem = operands[2]; | |
10471 | oldval = operands[3]; | |
10472 | newval = operands[4]; | |
10473 | is_weak = operands[5]; | |
10474 | mod_s = operands[6]; | |
10475 | mod_f = operands[7]; | |
10476 | mode = GET_MODE (mem); | |
10477 | ||
10478 | if (reg_overlap_mentioned_p (rval, oldval)) | |
10479 | oldval = copy_to_reg (oldval); | |
10480 | ||
10481 | if (mode == SImode) | |
10482 | { | |
10483 | emit_insn (gen_atomic_compare_and_swapsi_1 (rval, mem, oldval, newval, | |
10484 | is_weak, mod_s, mod_f)); | |
10485 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10486 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
10487 | emit_insn (gen_rtx_SET (bval, x)); | |
10488 | } | |
10489 | else | |
10490 | { | |
10491 | arc_expand_compare_and_swap_qh (bval, rval, mem, oldval, newval, | |
10492 | is_weak, mod_s, mod_f); | |
10493 | } | |
10494 | } | |
10495 | ||
10496 | /* Helper function used by the "atomic_compare_and_swapsi_1" | |
10497 | pattern. */ | |
10498 | ||
10499 | void | |
10500 | arc_split_compare_and_swap (rtx operands[]) | |
10501 | { | |
10502 | rtx rval, mem, oldval, newval; | |
10503 | machine_mode mode; | |
10504 | enum memmodel mod_s, mod_f; | |
10505 | bool is_weak; | |
10506 | rtx label1, label2, x, cond; | |
10507 | ||
10508 | rval = operands[0]; | |
10509 | mem = operands[1]; | |
10510 | oldval = operands[2]; | |
10511 | newval = operands[3]; | |
10512 | is_weak = (operands[4] != const0_rtx); | |
10513 | mod_s = (enum memmodel) INTVAL (operands[5]); | |
10514 | mod_f = (enum memmodel) INTVAL (operands[6]); | |
10515 | mode = GET_MODE (mem); | |
10516 | ||
10517 | /* ARC atomic ops work only with 32-bit aligned memories. */ | |
10518 | gcc_assert (mode == SImode); | |
10519 | ||
10520 | arc_pre_atomic_barrier (mod_s); | |
10521 | ||
10522 | label1 = NULL_RTX; | |
10523 | if (!is_weak) | |
10524 | { | |
10525 | label1 = gen_label_rtx (); | |
10526 | emit_label (label1); | |
10527 | } | |
10528 | label2 = gen_label_rtx (); | |
10529 | ||
10530 | /* Load exclusive. */ | |
10531 | emit_insn (gen_arc_load_exclusivesi (rval, mem)); | |
10532 | ||
10533 | /* Check if it is oldval. */ | |
10534 | mode = SELECT_CC_MODE (NE, rval, oldval); | |
10535 | cond = gen_rtx_REG (mode, CC_REG); | |
10536 | emit_insn (gen_rtx_SET (cond, gen_rtx_COMPARE (mode, rval, oldval))); | |
10537 | ||
10538 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10539 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10540 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
10541 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10542 | ||
10543 | /* Exclusively store new item. Store clobbers CC reg. */ | |
10544 | emit_insn (gen_arc_store_exclusivesi (mem, newval)); | |
10545 | ||
10546 | if (!is_weak) | |
10547 | { | |
10548 | /* Check the result of the store. */ | |
10549 | cond = gen_rtx_REG (CC_Zmode, CC_REG); | |
10550 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10551 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10552 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
10553 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10554 | } | |
10555 | ||
10556 | if (mod_f != MEMMODEL_RELAXED) | |
10557 | emit_label (label2); | |
10558 | ||
10559 | arc_post_atomic_barrier (mod_s); | |
10560 | ||
10561 | if (mod_f == MEMMODEL_RELAXED) | |
10562 | emit_label (label2); | |
10563 | } | |
10564 | ||
10565 | /* Expand an atomic fetch-and-operate pattern. CODE is the binary operation | |
10566 | to perform. MEM is the memory on which to operate. VAL is the second | |
10567 | operand of the binary operator. BEFORE and AFTER are optional locations to | |
10568 | return the value of MEM either before of after the operation. MODEL_RTX | |
10569 | is a CONST_INT containing the memory model to use. */ | |
10570 | ||
10571 | void | |
10572 | arc_expand_atomic_op (enum rtx_code code, rtx mem, rtx val, | |
10573 | rtx orig_before, rtx orig_after, rtx model_rtx) | |
10574 | { | |
10575 | enum memmodel model = (enum memmodel) INTVAL (model_rtx); | |
10576 | machine_mode mode = GET_MODE (mem); | |
10577 | rtx label, x, cond; | |
10578 | rtx before = orig_before, after = orig_after; | |
10579 | ||
10580 | /* ARC atomic ops work only with 32-bit aligned memories. */ | |
10581 | gcc_assert (mode == SImode); | |
10582 | ||
10583 | arc_pre_atomic_barrier (model); | |
10584 | ||
10585 | label = gen_label_rtx (); | |
10586 | emit_label (label); | |
10587 | label = gen_rtx_LABEL_REF (VOIDmode, label); | |
10588 | ||
10589 | if (before == NULL_RTX) | |
10590 | before = gen_reg_rtx (mode); | |
10591 | ||
10592 | if (after == NULL_RTX) | |
10593 | after = gen_reg_rtx (mode); | |
10594 | ||
10595 | /* Load exclusive. */ | |
10596 | emit_insn (gen_arc_load_exclusivesi (before, mem)); | |
10597 | ||
10598 | switch (code) | |
10599 | { | |
10600 | case NOT: | |
10601 | x = gen_rtx_AND (mode, before, val); | |
10602 | emit_insn (gen_rtx_SET (after, x)); | |
10603 | x = gen_rtx_NOT (mode, after); | |
10604 | emit_insn (gen_rtx_SET (after, x)); | |
10605 | break; | |
10606 | ||
10607 | case MINUS: | |
10608 | if (CONST_INT_P (val)) | |
10609 | { | |
10610 | val = GEN_INT (-INTVAL (val)); | |
10611 | code = PLUS; | |
10612 | } | |
10613 | ||
10614 | /* FALLTHRU. */ | |
10615 | default: | |
10616 | x = gen_rtx_fmt_ee (code, mode, before, val); | |
10617 | emit_insn (gen_rtx_SET (after, x)); | |
10618 | break; | |
10619 | } | |
10620 | ||
10621 | /* Exclusively store new item. Store clobbers CC reg. */ | |
10622 | emit_insn (gen_arc_store_exclusivesi (mem, after)); | |
10623 | ||
10624 | /* Check the result of the store. */ | |
10625 | cond = gen_rtx_REG (CC_Zmode, CC_REG); | |
10626 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10627 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10628 | label, pc_rtx); | |
10629 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10630 | ||
10631 | arc_post_atomic_barrier (model); | |
10632 | } | |
10633 | ||
bf9e9dc5 CZ |
10634 | /* Implement TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P. */ |
10635 | ||
10636 | static bool | |
10637 | arc_no_speculation_in_delay_slots_p () | |
10638 | { | |
10639 | return true; | |
10640 | } | |
10641 | ||
d34a0fdc CZ |
10642 | /* Return a parallel of registers to represent where to find the |
10643 | register pieces if required, otherwise NULL_RTX. */ | |
10644 | ||
10645 | static rtx | |
10646 | arc_dwarf_register_span (rtx rtl) | |
10647 | { | |
cd1e4d41 | 10648 | machine_mode mode = GET_MODE (rtl); |
d34a0fdc CZ |
10649 | unsigned regno; |
10650 | rtx p; | |
10651 | ||
10652 | if (GET_MODE_SIZE (mode) != 8) | |
10653 | return NULL_RTX; | |
10654 | ||
10655 | p = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2)); | |
10656 | regno = REGNO (rtl); | |
10657 | XVECEXP (p, 0, 0) = gen_rtx_REG (SImode, regno); | |
10658 | XVECEXP (p, 0, 1) = gen_rtx_REG (SImode, regno + 1); | |
10659 | ||
10660 | return p; | |
10661 | } | |
10662 | ||
fc1c2d04 CZ |
10663 | /* Return true if OP is an acceptable memory operand for ARCompact |
10664 | 16-bit load instructions of MODE. | |
10665 | ||
10666 | AV2SHORT: TRUE if address needs to fit into the new ARCv2 short | |
10667 | non scaled instructions. | |
10668 | ||
10669 | SCALED: TRUE if address can be scaled. */ | |
10670 | ||
10671 | bool | |
10672 | compact_memory_operand_p (rtx op, machine_mode mode, | |
10673 | bool av2short, bool scaled) | |
10674 | { | |
10675 | rtx addr, plus0, plus1; | |
10676 | int size, off; | |
10677 | ||
10678 | /* Eliminate non-memory operations. */ | |
10679 | if (GET_CODE (op) != MEM) | |
10680 | return 0; | |
10681 | ||
10682 | /* .di instructions have no 16-bit form. */ | |
10683 | if (MEM_VOLATILE_P (op) && !TARGET_VOLATILE_CACHE_SET) | |
10684 | return false; | |
10685 | ||
10686 | if (mode == VOIDmode) | |
10687 | mode = GET_MODE (op); | |
10688 | ||
10689 | size = GET_MODE_SIZE (mode); | |
10690 | ||
10691 | /* dword operations really put out 2 instructions, so eliminate | |
10692 | them. */ | |
10693 | if (size > UNITS_PER_WORD) | |
10694 | return false; | |
10695 | ||
10696 | /* Decode the address now. */ | |
10697 | addr = XEXP (op, 0); | |
10698 | switch (GET_CODE (addr)) | |
10699 | { | |
10700 | case REG: | |
10701 | return (REGNO (addr) >= FIRST_PSEUDO_REGISTER | |
10702 | || COMPACT_GP_REG_P (REGNO (addr)) | |
10703 | || (SP_REG_P (REGNO (addr)) && (size != 2))); | |
10704 | case PLUS: | |
10705 | plus0 = XEXP (addr, 0); | |
10706 | plus1 = XEXP (addr, 1); | |
10707 | ||
10708 | if ((GET_CODE (plus0) == REG) | |
10709 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
10710 | || COMPACT_GP_REG_P (REGNO (plus0))) | |
10711 | && ((GET_CODE (plus1) == REG) | |
10712 | && ((REGNO (plus1) >= FIRST_PSEUDO_REGISTER) | |
10713 | || COMPACT_GP_REG_P (REGNO (plus1))))) | |
10714 | { | |
10715 | return !av2short; | |
10716 | } | |
10717 | ||
10718 | if ((GET_CODE (plus0) == REG) | |
10719 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
10720 | || (COMPACT_GP_REG_P (REGNO (plus0)) && !av2short) | |
10721 | || (IN_RANGE (REGNO (plus0), 0, 31) && av2short)) | |
10722 | && (GET_CODE (plus1) == CONST_INT)) | |
10723 | { | |
10724 | bool valid = false; | |
10725 | ||
10726 | off = INTVAL (plus1); | |
10727 | ||
10728 | /* Negative offset is not supported in 16-bit load/store insns. */ | |
10729 | if (off < 0) | |
10730 | return 0; | |
10731 | ||
10732 | /* Only u5 immediates allowed in code density instructions. */ | |
10733 | if (av2short) | |
10734 | { | |
10735 | switch (size) | |
10736 | { | |
10737 | case 1: | |
10738 | return false; | |
10739 | case 2: | |
10740 | /* This is an ldh_s.x instruction, check the u6 | |
10741 | immediate. */ | |
10742 | if (COMPACT_GP_REG_P (REGNO (plus0))) | |
10743 | valid = true; | |
10744 | break; | |
10745 | case 4: | |
10746 | /* Only u5 immediates allowed in 32bit access code | |
10747 | density instructions. */ | |
10748 | if (REGNO (plus0) <= 31) | |
10749 | return ((off < 32) && (off % 4 == 0)); | |
10750 | break; | |
10751 | default: | |
10752 | return false; | |
10753 | } | |
10754 | } | |
10755 | else | |
10756 | if (COMPACT_GP_REG_P (REGNO (plus0))) | |
10757 | valid = true; | |
10758 | ||
10759 | if (valid) | |
10760 | { | |
10761 | ||
10762 | switch (size) | |
10763 | { | |
10764 | case 1: | |
10765 | return (off < 32); | |
10766 | case 2: | |
10767 | /* The 6-bit constant get shifted to fit the real | |
10768 | 5-bits field. Check also for the alignment. */ | |
10769 | return ((off < 64) && (off % 2 == 0)); | |
10770 | case 4: | |
10771 | return ((off < 128) && (off % 4 == 0)); | |
10772 | default: | |
10773 | return false; | |
10774 | } | |
10775 | } | |
10776 | } | |
10777 | ||
10778 | if (REG_P (plus0) && CONST_INT_P (plus1) | |
10779 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
10780 | || SP_REG_P (REGNO (plus0))) | |
10781 | && !av2short) | |
10782 | { | |
10783 | off = INTVAL (plus1); | |
10784 | return ((size != 2) && (off >= 0 && off < 128) && (off % 4 == 0)); | |
10785 | } | |
10786 | ||
10787 | if ((GET_CODE (plus0) == MULT) | |
10788 | && (GET_CODE (XEXP (plus0, 0)) == REG) | |
10789 | && ((REGNO (XEXP (plus0, 0)) >= FIRST_PSEUDO_REGISTER) | |
10790 | || COMPACT_GP_REG_P (REGNO (XEXP (plus0, 0)))) | |
10791 | && (GET_CODE (plus1) == REG) | |
10792 | && ((REGNO (plus1) >= FIRST_PSEUDO_REGISTER) | |
10793 | || COMPACT_GP_REG_P (REGNO (plus1)))) | |
10794 | return scaled; | |
10795 | default: | |
10796 | break ; | |
10797 | /* TODO: 'gp' and 'pcl' are to supported as base address operand | |
10798 | for 16-bit load instructions. */ | |
10799 | } | |
10800 | return false; | |
10801 | } | |
10802 | ||
6fe5e235 CZ |
10803 | /* Return the frame pointer value to be backed up in the setjmp buffer. */ |
10804 | ||
10805 | static rtx | |
10806 | arc_builtin_setjmp_frame_value (void) | |
10807 | { | |
10808 | /* We always want to preserve whatever value is currently in the frame | |
10809 | pointer register. For frames that are using the frame pointer the new | |
10810 | value of the frame pointer register will have already been computed | |
10811 | (as part of the prologue). For frames that are not using the frame | |
10812 | pointer it is important that we backup whatever value is in the frame | |
10813 | pointer register, as earlier (more outer) frames may have placed a | |
10814 | value into the frame pointer register. It might be tempting to try | |
10815 | and use `frame_pointer_rtx` here, however, this is not what we want. | |
10816 | For frames that are using the frame pointer this will give the | |
10817 | correct value. However, for frames that are not using the frame | |
10818 | pointer this will still give the value that _would_ have been the | |
10819 | frame pointer value for this frame (if the use of the frame pointer | |
10820 | had not been removed). We really do want the raw frame pointer | |
10821 | register value. */ | |
10822 | return gen_raw_REG (Pmode, FRAME_POINTER_REGNUM); | |
10823 | } | |
10824 | ||
7cfbf676 CZ |
10825 | /* Implement TARGET_USE_ANCHORS_FOR_SYMBOL_P. We don't want to use |
10826 | anchors for small data: the GP register acts as an anchor in that | |
10827 | case. We also don't want to use them for PC-relative accesses, | |
10828 | where the PC acts as an anchor. Prohibit also TLS symbols to use | |
10829 | anchors. */ | |
10830 | ||
10831 | static bool | |
10832 | arc_use_anchors_for_symbol_p (const_rtx symbol) | |
10833 | { | |
10834 | if (SYMBOL_REF_TLS_MODEL (symbol)) | |
10835 | return false; | |
10836 | ||
10837 | if (flag_pic) | |
10838 | return false; | |
10839 | ||
10840 | if (SYMBOL_REF_SMALL_P (symbol)) | |
10841 | return false; | |
10842 | ||
10843 | return default_use_anchors_for_symbol_p (symbol); | |
10844 | } | |
10845 | ||
31e72f4f CZ |
10846 | /* Return true if SUBST can't safely replace its equivalent during RA. */ |
10847 | static bool | |
10848 | arc_cannot_substitute_mem_equiv_p (rtx) | |
10849 | { | |
10850 | /* If SUBST is mem[base+index], the address may not fit ISA, | |
10851 | thus return true. */ | |
10852 | return true; | |
10853 | } | |
10854 | ||
7cfbf676 CZ |
10855 | #undef TARGET_USE_ANCHORS_FOR_SYMBOL_P |
10856 | #define TARGET_USE_ANCHORS_FOR_SYMBOL_P arc_use_anchors_for_symbol_p | |
10857 | ||
58e17cf8 RS |
10858 | #undef TARGET_CONSTANT_ALIGNMENT |
10859 | #define TARGET_CONSTANT_ALIGNMENT constant_alignment_word_strings | |
10860 | ||
31e72f4f CZ |
10861 | #undef TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P |
10862 | #define TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P arc_cannot_substitute_mem_equiv_p | |
10863 | ||
efcc2e30 CZ |
10864 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE |
10865 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE arc_asm_trampoline_template | |
10866 | ||
526b7aee SV |
10867 | struct gcc_target targetm = TARGET_INITIALIZER; |
10868 | ||
10869 | #include "gt-arc.h" |