]>
Commit | Line | Data |
---|---|---|
526b7aee | 1 | /* Subroutines used for code generation on the Synopsys DesignWare ARC cpu. |
a5544970 | 2 | Copyright (C) 1994-2019 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 | |
6b55f8c9 CZ |
76 | typedef struct GTY (()) _arc_jli_section |
77 | { | |
78 | const char *name; | |
79 | struct _arc_jli_section *next; | |
80 | } arc_jli_section; | |
81 | ||
82 | static arc_jli_section *arc_jli_sections = NULL; | |
83 | ||
66825a30 CZ |
84 | /* Track which regs are set fixed/call saved/call used from commnad line. */ |
85 | HARD_REG_SET overrideregs; | |
86 | ||
a2de90a4 CZ |
87 | /* Maximum size of a loop. */ |
88 | #define ARC_MAX_LOOP_LENGTH 4095 | |
89 | ||
90b48013 CZ |
90 | /* Check if an rtx fits in the store instruction format. Loads can |
91 | handle any constant. */ | |
92 | #define RTX_OK_FOR_OFFSET_P(MODE, X) \ | |
93 | (GET_CODE (X) == CONST_INT \ | |
94 | && SMALL_INT_RANGE (INTVAL (X), (GET_MODE_SIZE (MODE) - 1) & (~0x03), \ | |
95 | (INTVAL (X) & (GET_MODE_SIZE (MODE) - 1) & 3 \ | |
96 | ? 0 \ | |
97 | : -(-GET_MODE_SIZE (MODE) | (~0x03)) >> 1))) | |
526b7aee | 98 | |
526b7aee SV |
99 | /* Array of valid operand punctuation characters. */ |
100 | char arc_punct_chars[256]; | |
101 | ||
102 | /* State used by arc_ccfsm_advance to implement conditional execution. */ | |
103 | struct GTY (()) arc_ccfsm | |
104 | { | |
105 | int state; | |
106 | int cc; | |
107 | rtx cond; | |
b3458f61 | 108 | rtx_insn *target_insn; |
526b7aee SV |
109 | int target_label; |
110 | }; | |
111 | ||
41453183 CZ |
112 | /* Status of the IRQ_CTRL_AUX register. */ |
113 | typedef struct irq_ctrl_saved_t | |
114 | { | |
115 | /* Last register number used by IRQ_CTRL_SAVED aux_reg. */ | |
116 | short irq_save_last_reg; | |
117 | /* True if BLINK is automatically saved. */ | |
118 | bool irq_save_blink; | |
119 | /* True if LPCOUNT is automatically saved. */ | |
120 | bool irq_save_lpcount; | |
121 | } irq_ctrl_saved_t; | |
122 | static irq_ctrl_saved_t irq_ctrl_saved; | |
123 | ||
124 | #define ARC_AUTOBLINK_IRQ_P(FNTYPE) \ | |
c7314bc1 CZ |
125 | ((ARC_INTERRUPT_P (FNTYPE) \ |
126 | && irq_ctrl_saved.irq_save_blink) \ | |
127 | || (ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
128 | && rgf_banked_register_count > 8)) | |
129 | ||
130 | #define ARC_AUTOFP_IRQ_P(FNTYPE) \ | |
131 | ((ARC_INTERRUPT_P (FNTYPE) \ | |
132 | && (irq_ctrl_saved.irq_save_last_reg > 26)) \ | |
133 | || (ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
134 | && rgf_banked_register_count > 8)) | |
135 | ||
136 | #define ARC_AUTO_IRQ_P(FNTYPE) \ | |
137 | (ARC_INTERRUPT_P (FNTYPE) && !ARC_FAST_INTERRUPT_P (FNTYPE) \ | |
138 | && (irq_ctrl_saved.irq_save_blink \ | |
41453183 CZ |
139 | || (irq_ctrl_saved.irq_save_last_reg >= 0))) |
140 | ||
c7314bc1 CZ |
141 | /* Number of registers in second bank for FIRQ support. */ |
142 | static int rgf_banked_register_count; | |
143 | ||
526b7aee SV |
144 | #define arc_ccfsm_current cfun->machine->ccfsm_current |
145 | ||
146 | #define ARC_CCFSM_BRANCH_DELETED_P(STATE) \ | |
147 | ((STATE)->state == 1 || (STATE)->state == 2) | |
148 | ||
149 | /* Indicate we're conditionalizing insns now. */ | |
150 | #define ARC_CCFSM_RECORD_BRANCH_DELETED(STATE) \ | |
151 | ((STATE)->state += 2) | |
152 | ||
153 | #define ARC_CCFSM_COND_EXEC_P(STATE) \ | |
154 | ((STATE)->state == 3 || (STATE)->state == 4 || (STATE)->state == 5 \ | |
155 | || current_insn_predicate) | |
156 | ||
157 | /* Check if INSN has a 16 bit opcode considering struct arc_ccfsm *STATE. */ | |
158 | #define CCFSM_ISCOMPACT(INSN,STATE) \ | |
159 | (ARC_CCFSM_COND_EXEC_P (STATE) \ | |
160 | ? (get_attr_iscompact (INSN) == ISCOMPACT_TRUE \ | |
161 | || get_attr_iscompact (INSN) == ISCOMPACT_TRUE_LIMM) \ | |
162 | : get_attr_iscompact (INSN) != ISCOMPACT_FALSE) | |
163 | ||
164 | /* Likewise, but also consider that INSN might be in a delay slot of JUMP. */ | |
165 | #define CCFSM_DBR_ISCOMPACT(INSN,JUMP,STATE) \ | |
166 | ((ARC_CCFSM_COND_EXEC_P (STATE) \ | |
167 | || (JUMP_P (JUMP) \ | |
168 | && INSN_ANNULLED_BRANCH_P (JUMP) \ | |
169 | && (TARGET_AT_DBR_CONDEXEC || INSN_FROM_TARGET_P (INSN)))) \ | |
170 | ? (get_attr_iscompact (INSN) == ISCOMPACT_TRUE \ | |
171 | || get_attr_iscompact (INSN) == ISCOMPACT_TRUE_LIMM) \ | |
172 | : get_attr_iscompact (INSN) != ISCOMPACT_FALSE) | |
173 | ||
90b48013 CZ |
174 | /* Start enter/leave register range. */ |
175 | #define ENTER_LEAVE_START_REG 13 | |
176 | ||
177 | /* End enter/leave register range. */ | |
178 | #define ENTER_LEAVE_END_REG 26 | |
179 | ||
526b7aee SV |
180 | /* The maximum number of insns skipped which will be conditionalised if |
181 | possible. */ | |
182 | /* When optimizing for speed: | |
183 | Let p be the probability that the potentially skipped insns need to | |
184 | be executed, pn the cost of a correctly predicted non-taken branch, | |
185 | mt the cost of a mis/non-predicted taken branch, | |
186 | mn mispredicted non-taken, pt correctly predicted taken ; | |
187 | costs expressed in numbers of instructions like the ones considered | |
188 | skipping. | |
189 | Unfortunately we don't have a measure of predictability - this | |
190 | is linked to probability only in that in the no-eviction-scenario | |
191 | there is a lower bound 1 - 2 * min (p, 1-p), and a somewhat larger | |
192 | value that can be assumed *if* the distribution is perfectly random. | |
193 | A predictability of 1 is perfectly plausible not matter what p is, | |
194 | because the decision could be dependent on an invocation parameter | |
195 | of the program. | |
196 | For large p, we want MAX_INSNS_SKIPPED == pn/(1-p) + mt - pn | |
197 | For small p, we want MAX_INSNS_SKIPPED == pt | |
198 | ||
199 | When optimizing for size: | |
200 | We want to skip insn unless we could use 16 opcodes for the | |
201 | non-conditionalized insn to balance the branch length or more. | |
202 | Performance can be tie-breaker. */ | |
203 | /* If the potentially-skipped insns are likely to be executed, we'll | |
204 | generally save one non-taken branch | |
205 | o | |
206 | this to be no less than the 1/p */ | |
207 | #define MAX_INSNS_SKIPPED 3 | |
208 | ||
ce9dbf20 CZ |
209 | /* ZOL control registers. */ |
210 | #define AUX_LP_START 0x02 | |
211 | #define AUX_LP_END 0x03 | |
212 | ||
213 | /* FPX AUX registers. */ | |
214 | #define AUX_DPFP_START 0x301 | |
215 | ||
526b7aee SV |
216 | /* A nop is needed between a 4 byte insn that sets the condition codes and |
217 | a branch that uses them (the same isn't true for an 8 byte insn that sets | |
218 | the condition codes). Set by arc_ccfsm_advance. Used by | |
219 | arc_print_operand. */ | |
220 | ||
221 | static int get_arc_condition_code (rtx); | |
222 | ||
223 | static tree arc_handle_interrupt_attribute (tree *, tree, tree, int, bool *); | |
1825c61e | 224 | static tree arc_handle_fndecl_attribute (tree *, tree, tree, int, bool *); |
6b55f8c9 | 225 | static tree arc_handle_jli_attribute (tree *, tree, tree, int, bool *); |
7778a1ad | 226 | static tree arc_handle_secure_attribute (tree *, tree, tree, int, bool *); |
8180c03f | 227 | static tree arc_handle_uncached_attribute (tree *, tree, tree, int, bool *); |
b6fb257b | 228 | static tree arc_handle_aux_attribute (tree *, tree, tree, int, bool *); |
526b7aee SV |
229 | |
230 | /* Initialized arc_attribute_table to NULL since arc doesnot have any | |
231 | machine specific supported attributes. */ | |
232 | const struct attribute_spec arc_attribute_table[] = | |
233 | { | |
4849deb1 JJ |
234 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, |
235 | affects_type_identity, handler, exclude } */ | |
236 | { "interrupt", 1, 1, true, false, false, true, | |
237 | arc_handle_interrupt_attribute, NULL }, | |
526b7aee SV |
238 | /* Function calls made to this symbol must be done indirectly, because |
239 | it may lie outside of the 21/25 bit addressing range of a normal function | |
240 | call. */ | |
4849deb1 | 241 | { "long_call", 0, 0, false, true, true, false, NULL, NULL }, |
526b7aee SV |
242 | /* Whereas these functions are always known to reside within the 25 bit |
243 | addressing range of unconditionalized bl. */ | |
4849deb1 | 244 | { "medium_call", 0, 0, false, true, true, false, NULL, NULL }, |
526b7aee SV |
245 | /* And these functions are always known to reside within the 21 bit |
246 | addressing range of blcc. */ | |
4849deb1 | 247 | { "short_call", 0, 0, false, true, true, false, NULL, NULL }, |
1825c61e CZ |
248 | /* Function which are not having the prologue and epilogue generated |
249 | by the compiler. */ | |
4849deb1 | 250 | { "naked", 0, 0, true, false, false, false, arc_handle_fndecl_attribute, |
5d9ae53d | 251 | NULL }, |
6b55f8c9 CZ |
252 | /* Functions calls made using jli instruction. The pointer in JLI |
253 | table is found latter. */ | |
8180c03f | 254 | { "jli_always", 0, 0, false, true, true, false, NULL, NULL }, |
6b55f8c9 CZ |
255 | /* Functions calls made using jli instruction. The pointer in JLI |
256 | table is given as input parameter. */ | |
8180c03f | 257 | { "jli_fixed", 1, 1, false, true, true, false, arc_handle_jli_attribute, |
6b55f8c9 | 258 | NULL }, |
7778a1ad | 259 | /* Call a function using secure-mode. */ |
8180c03f CZ |
260 | { "secure_call", 1, 1, false, true, true, false, arc_handle_secure_attribute, |
261 | NULL }, | |
262 | /* Bypass caches using .di flag. */ | |
263 | { "uncached", 0, 0, false, true, false, false, arc_handle_uncached_attribute, | |
7778a1ad | 264 | NULL }, |
b6fb257b | 265 | { "aux", 0, 1, true, false, false, false, arc_handle_aux_attribute, NULL }, |
4849deb1 | 266 | { NULL, 0, 0, false, false, false, false, NULL, NULL } |
526b7aee SV |
267 | }; |
268 | static int arc_comp_type_attributes (const_tree, const_tree); | |
269 | static void arc_file_start (void); | |
270 | static void arc_internal_label (FILE *, const char *, unsigned long); | |
271 | static void arc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, | |
272 | tree); | |
ef4bddc2 | 273 | static int arc_address_cost (rtx, machine_mode, addr_space_t, bool); |
526b7aee SV |
274 | static void arc_encode_section_info (tree decl, rtx rtl, int first); |
275 | ||
276 | static void arc_init_builtins (void); | |
ef4bddc2 | 277 | static rtx arc_expand_builtin (tree, rtx, rtx, machine_mode, int); |
526b7aee SV |
278 | |
279 | static int branch_dest (rtx); | |
280 | ||
281 | static void arc_output_pic_addr_const (FILE *, rtx, int); | |
526b7aee SV |
282 | static bool arc_function_ok_for_sibcall (tree, tree); |
283 | static rtx arc_function_value (const_tree, const_tree, bool); | |
284 | const char * output_shift (rtx *); | |
285 | static void arc_reorg (void); | |
286 | static bool arc_in_small_data_p (const_tree); | |
287 | ||
288 | static void arc_init_reg_tables (void); | |
289 | static bool arc_return_in_memory (const_tree, const_tree); | |
ef4bddc2 | 290 | static bool arc_vector_mode_supported_p (machine_mode); |
526b7aee | 291 | |
807e902e KZ |
292 | static bool arc_can_use_doloop_p (const widest_int &, const widest_int &, |
293 | unsigned int, bool); | |
ac44248e | 294 | static const char *arc_invalid_within_doloop (const rtx_insn *); |
526b7aee SV |
295 | |
296 | static void output_short_suffix (FILE *file); | |
297 | ||
298 | static bool arc_frame_pointer_required (void); | |
299 | ||
445d7826 | 300 | static bool arc_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT, |
ad23f5d4 JG |
301 | unsigned int, |
302 | enum by_pieces_operation op, | |
303 | bool); | |
304 | ||
c3bde35a AB |
305 | /* Globally visible information about currently selected cpu. */ |
306 | const arc_cpu_t *arc_selected_cpu; | |
f9ccf899 | 307 | |
ce9dbf20 CZ |
308 | /* Traditionally, we push saved registers first in the prologue, |
309 | then we allocate the rest of the frame - and reverse in the epilogue. | |
310 | This has still its merits for ease of debugging, or saving code size | |
311 | or even execution time if the stack frame is so large that some accesses | |
312 | can't be encoded anymore with offsets in the instruction code when using | |
313 | a different scheme. | |
314 | Also, it would be a good starting point if we got instructions to help | |
315 | with register save/restore. | |
316 | ||
317 | However, often stack frames are small, and the pushing / popping has | |
318 | some costs: | |
319 | - the stack modification prevents a lot of scheduling. | |
320 | - frame allocation / deallocation may need extra instructions. | |
321 | - we need to place a memory barrier after frame allocation to avoid | |
322 | the delay slot scheduler to reschedule a frame related info and | |
323 | messing up with dwarf unwinding. The barrier before deallocation | |
324 | is for flushing all pending sp operations. | |
325 | ||
326 | Thus, for small frames, we'd like to use a different scheme: | |
327 | - The frame is allocated in full with the first prologue instruction, | |
328 | and deallocated in full with the last epilogue instruction. | |
329 | Thus, the instructions in-between can be freely scheduled. | |
330 | - If the function has no outgoing arguments on the stack, we can allocate | |
331 | one register save slot at the top of the stack. This register can then | |
332 | be saved simultaneously with frame allocation, and restored with | |
333 | frame deallocation. | |
334 | This register can be picked depending on scheduling considerations, | |
335 | although same though should go into having some set of registers | |
336 | to be potentially lingering after a call, and others to be available | |
337 | immediately - i.e. in the absence of interprocedual optimization, we | |
338 | can use an ABI-like convention for register allocation to reduce | |
339 | stalls after function return. */ | |
340 | ||
341 | /* ARCompact stack frames look like: | |
342 | ||
343 | Before call After call | |
344 | high +-----------------------+ +-----------------------+ | |
345 | mem | reg parm save area | | reg parm save area | | |
346 | | only created for | | only created for | | |
347 | | variable arg fns | | variable arg fns | | |
348 | AP +-----------------------+ +-----------------------+ | |
349 | | return addr register | | return addr register | | |
350 | | (if required) | | (if required) | | |
351 | +-----------------------+ +-----------------------+ | |
352 | | | | | | |
353 | | reg save area | | reg save area | | |
354 | | | | | | |
355 | +-----------------------+ +-----------------------+ | |
356 | | frame pointer | | frame pointer | | |
357 | | (if required) | | (if required) | | |
358 | FP +-----------------------+ +-----------------------+ | |
359 | | | | | | |
360 | | local/temp variables | | local/temp variables | | |
361 | | | | | | |
362 | +-----------------------+ +-----------------------+ | |
363 | | | | | | |
364 | | arguments on stack | | arguments on stack | | |
365 | | | | | | |
366 | SP +-----------------------+ +-----------------------+ | |
367 | | reg parm save area | | |
368 | | only created for | | |
369 | | variable arg fns | | |
370 | AP +-----------------------+ | |
371 | | return addr register | | |
372 | | (if required) | | |
373 | +-----------------------+ | |
374 | | | | |
375 | | reg save area | | |
376 | | | | |
377 | +-----------------------+ | |
378 | | frame pointer | | |
379 | | (if required) | | |
380 | FP +-----------------------+ | |
381 | | | | |
382 | | local/temp variables | | |
383 | | | | |
384 | +-----------------------+ | |
385 | | | | |
386 | | arguments on stack | | |
387 | low | | | |
388 | mem SP +-----------------------+ | |
389 | ||
390 | Notes: | |
391 | 1) The "reg parm save area" does not exist for non variable argument fns. | |
392 | The "reg parm save area" can be eliminated completely if we created our | |
393 | own va-arc.h, but that has tradeoffs as well (so it's not done). */ | |
394 | ||
395 | /* Structure to be filled in by arc_compute_frame_size with register | |
396 | save masks, and offsets for the current function. */ | |
397 | struct GTY (()) arc_frame_info | |
398 | { | |
399 | unsigned int total_size; /* # bytes that the entire frame takes up. */ | |
400 | unsigned int extra_size; /* # bytes of extra stuff. */ | |
401 | unsigned int pretend_size; /* # bytes we push and pretend caller did. */ | |
402 | unsigned int args_size; /* # bytes that outgoing arguments take up. */ | |
403 | unsigned int reg_size; /* # bytes needed to store regs. */ | |
404 | unsigned int var_size; /* # bytes that variables take up. */ | |
405 | uint64_t gmask; /* Mask of saved gp registers. */ | |
406 | bool initialized; /* FALSE if frame size already calculated. */ | |
407 | short millicode_start_reg; | |
408 | short millicode_end_reg; | |
409 | bool save_return_addr; | |
410 | }; | |
411 | ||
412 | /* GMASK bit length -1. */ | |
413 | #define GMASK_LEN 63 | |
414 | ||
415 | /* Defining data structures for per-function information. */ | |
416 | ||
417 | typedef struct GTY (()) machine_function | |
418 | { | |
419 | unsigned int fn_type; | |
420 | struct arc_frame_info frame_info; | |
421 | /* To keep track of unalignment caused by short insns. */ | |
422 | int unalign; | |
423 | struct arc_ccfsm ccfsm_current; | |
424 | /* Map from uid to ccfsm state during branch shortening. */ | |
425 | rtx ccfsm_current_insn; | |
426 | char arc_reorg_started; | |
427 | char prescan_initialized; | |
428 | } machine_function; | |
429 | ||
430 | ||
e0be3321 CZ |
431 | /* Given a symbol RTX (const (symb <+ const_int>), returns its |
432 | alignment. */ | |
433 | ||
434 | static int | |
435 | get_symbol_alignment (rtx x) | |
436 | { | |
437 | tree decl = NULL_TREE; | |
438 | int align = 0; | |
439 | ||
440 | switch (GET_CODE (x)) | |
441 | { | |
442 | case SYMBOL_REF: | |
443 | decl = SYMBOL_REF_DECL (x); | |
444 | break; | |
445 | case CONST: | |
446 | return get_symbol_alignment (XEXP (x, 0)); | |
447 | case PLUS: | |
448 | gcc_assert (CONST_INT_P (XEXP (x, 1))); | |
449 | return get_symbol_alignment (XEXP (x, 0)); | |
450 | default: | |
451 | return 0; | |
452 | } | |
453 | ||
454 | if (decl) | |
455 | align = DECL_ALIGN (decl); | |
456 | align = align / BITS_PER_UNIT; | |
457 | return align; | |
458 | } | |
459 | ||
460 | /* Return true if x is ok to be used as a small data address. */ | |
461 | ||
462 | static bool | |
463 | legitimate_small_data_address_p (rtx x) | |
464 | { | |
465 | switch (GET_CODE (x)) | |
466 | { | |
467 | case CONST: | |
468 | return legitimate_small_data_address_p (XEXP (x, 0)); | |
469 | case SYMBOL_REF: | |
470 | return SYMBOL_REF_SMALL_P (x); | |
471 | case PLUS: | |
472 | { | |
473 | bool p0 = (GET_CODE (XEXP (x, 0)) == SYMBOL_REF) | |
474 | && SYMBOL_REF_SMALL_P (XEXP (x, 0)); | |
475 | bool p1 = CONST_INT_P (XEXP (x, 1)) | |
476 | && (INTVAL (XEXP (x, 1)) <= g_switch_value); | |
477 | return p0 && p1; | |
478 | } | |
479 | default: | |
480 | return false; | |
481 | } | |
482 | } | |
483 | ||
484 | /* TRUE if op is an scaled address. */ | |
9f532472 CZ |
485 | static bool |
486 | legitimate_scaled_address_p (machine_mode mode, rtx op, bool strict) | |
487 | { | |
488 | if (GET_CODE (op) != PLUS) | |
489 | return false; | |
490 | ||
491 | if (GET_CODE (XEXP (op, 0)) != MULT) | |
492 | return false; | |
493 | ||
494 | /* Check multiplication operands. */ | |
495 | if (!RTX_OK_FOR_INDEX_P (XEXP (XEXP (op, 0), 0), strict)) | |
496 | return false; | |
497 | ||
498 | if (!CONST_INT_P (XEXP (XEXP (op, 0), 1))) | |
499 | return false; | |
500 | ||
501 | switch (GET_MODE_SIZE (mode)) | |
502 | { | |
503 | case 2: | |
504 | if (INTVAL (XEXP (XEXP (op, 0), 1)) != 2) | |
505 | return false; | |
506 | break; | |
507 | case 8: | |
508 | if (!TARGET_LL64) | |
509 | return false; | |
510 | /* Fall through. */ | |
511 | case 4: | |
512 | if (INTVAL (XEXP (XEXP (op, 0), 1)) != 4) | |
513 | return false; | |
41bc2c0b | 514 | /* Fall through. */ |
9f532472 CZ |
515 | default: |
516 | return false; | |
517 | } | |
518 | ||
519 | /* Check the base. */ | |
520 | if (RTX_OK_FOR_BASE_P (XEXP (op, 1), (strict))) | |
521 | return true; | |
522 | ||
523 | if (flag_pic) | |
524 | { | |
525 | if (CONST_INT_P (XEXP (op, 1))) | |
526 | return true; | |
527 | return false; | |
528 | } | |
e0be3321 CZ |
529 | |
530 | /* Scalled addresses for sdata is done other places. */ | |
531 | if (legitimate_small_data_address_p (op)) | |
532 | return false; | |
533 | ||
9f532472 | 534 | if (CONSTANT_P (XEXP (op, 1))) |
9f532472 | 535 | return true; |
9f532472 CZ |
536 | |
537 | return false; | |
538 | } | |
539 | ||
ac2e1a51 CZ |
540 | /* Check for constructions like REG + OFFS, where OFFS can be a |
541 | register, an immediate or an long immediate. */ | |
542 | ||
543 | static bool | |
b8506a8a | 544 | legitimate_offset_address_p (machine_mode mode, rtx x, bool index, bool strict) |
ac2e1a51 CZ |
545 | { |
546 | if (GET_CODE (x) != PLUS) | |
547 | return false; | |
548 | ||
549 | if (!RTX_OK_FOR_BASE_P (XEXP (x, 0), (strict))) | |
550 | return false; | |
551 | ||
552 | /* Check for: [Rx + small offset] or [Rx + Ry]. */ | |
553 | if (((index && RTX_OK_FOR_INDEX_P (XEXP (x, 1), (strict)) | |
554 | && GET_MODE_SIZE ((mode)) <= 4) | |
555 | || RTX_OK_FOR_OFFSET_P (mode, XEXP (x, 1)))) | |
556 | return true; | |
557 | ||
558 | /* Check for [Rx + symbol]. */ | |
559 | if (!flag_pic | |
560 | && (GET_CODE (XEXP (x, 1)) == SYMBOL_REF) | |
561 | /* Avoid this type of address for double or larger modes. */ | |
562 | && (GET_MODE_SIZE (mode) <= 4) | |
563 | /* Avoid small data which ends in something like GP + | |
564 | symb@sda. */ | |
9f532472 | 565 | && (!SYMBOL_REF_SMALL_P (XEXP (x, 1)))) |
ac2e1a51 CZ |
566 | return true; |
567 | ||
568 | return false; | |
569 | } | |
570 | ||
526b7aee SV |
571 | /* Implements target hook vector_mode_supported_p. */ |
572 | ||
573 | static bool | |
ef4bddc2 | 574 | arc_vector_mode_supported_p (machine_mode mode) |
526b7aee | 575 | { |
00c072ae CZ |
576 | switch (mode) |
577 | { | |
4e10a5a7 | 578 | case E_V2HImode: |
00c072ae | 579 | return TARGET_PLUS_DMPY; |
4e10a5a7 RS |
580 | case E_V4HImode: |
581 | case E_V2SImode: | |
00c072ae | 582 | return TARGET_PLUS_QMACW; |
4e10a5a7 RS |
583 | case E_V4SImode: |
584 | case E_V8HImode: | |
00c072ae | 585 | return TARGET_SIMD_SET; |
526b7aee | 586 | |
00c072ae CZ |
587 | default: |
588 | return false; | |
589 | } | |
590 | } | |
526b7aee | 591 | |
00c072ae CZ |
592 | /* Implements target hook TARGET_VECTORIZE_PREFERRED_SIMD_MODE. */ |
593 | ||
cd1e4d41 | 594 | static machine_mode |
005ba29c | 595 | arc_preferred_simd_mode (scalar_mode mode) |
00c072ae CZ |
596 | { |
597 | switch (mode) | |
598 | { | |
4e10a5a7 | 599 | case E_HImode: |
00c072ae | 600 | return TARGET_PLUS_QMACW ? V4HImode : V2HImode; |
4e10a5a7 | 601 | case E_SImode: |
00c072ae CZ |
602 | return V2SImode; |
603 | ||
604 | default: | |
605 | return word_mode; | |
606 | } | |
526b7aee SV |
607 | } |
608 | ||
00c072ae CZ |
609 | /* Implements target hook |
610 | TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES. */ | |
611 | ||
86e36728 | 612 | static void |
f63445e5 | 613 | arc_autovectorize_vector_sizes (vector_sizes *sizes, bool) |
00c072ae | 614 | { |
86e36728 RS |
615 | if (TARGET_PLUS_QMACW) |
616 | { | |
617 | sizes->quick_push (8); | |
618 | sizes->quick_push (4); | |
619 | } | |
00c072ae | 620 | } |
526b7aee | 621 | |
43bb0fc2 CZ |
622 | |
623 | /* Implements target hook TARGET_SCHED_ISSUE_RATE. */ | |
624 | static int | |
625 | arc_sched_issue_rate (void) | |
626 | { | |
627 | switch (arc_tune) | |
628 | { | |
629 | case TUNE_ARCHS4X: | |
630 | case TUNE_ARCHS4XD: | |
631 | return 3; | |
632 | default: | |
633 | break; | |
634 | } | |
635 | return 1; | |
636 | } | |
637 | ||
526b7aee SV |
638 | /* TARGET_PRESERVE_RELOAD_P is still awaiting patch re-evaluation / review. */ |
639 | static bool arc_preserve_reload_p (rtx in) ATTRIBUTE_UNUSED; | |
640 | static rtx arc_delegitimize_address (rtx); | |
c1ce59ab DM |
641 | static bool arc_can_follow_jump (const rtx_insn *follower, |
642 | const rtx_insn *followee); | |
526b7aee SV |
643 | |
644 | static rtx frame_insn (rtx); | |
ef4bddc2 | 645 | static void arc_function_arg_advance (cumulative_args_t, machine_mode, |
526b7aee | 646 | const_tree, bool); |
ef4bddc2 | 647 | static rtx arc_legitimize_address_0 (rtx, rtx, machine_mode mode); |
526b7aee | 648 | |
526b7aee SV |
649 | /* initialize the GCC target structure. */ |
650 | #undef TARGET_COMP_TYPE_ATTRIBUTES | |
651 | #define TARGET_COMP_TYPE_ATTRIBUTES arc_comp_type_attributes | |
652 | #undef TARGET_ASM_FILE_START | |
653 | #define TARGET_ASM_FILE_START arc_file_start | |
654 | #undef TARGET_ATTRIBUTE_TABLE | |
655 | #define TARGET_ATTRIBUTE_TABLE arc_attribute_table | |
656 | #undef TARGET_ASM_INTERNAL_LABEL | |
657 | #define TARGET_ASM_INTERNAL_LABEL arc_internal_label | |
658 | #undef TARGET_RTX_COSTS | |
659 | #define TARGET_RTX_COSTS arc_rtx_costs | |
660 | #undef TARGET_ADDRESS_COST | |
661 | #define TARGET_ADDRESS_COST arc_address_cost | |
662 | ||
663 | #undef TARGET_ENCODE_SECTION_INFO | |
664 | #define TARGET_ENCODE_SECTION_INFO arc_encode_section_info | |
665 | ||
666 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
667 | #define TARGET_CANNOT_FORCE_CONST_MEM arc_cannot_force_const_mem | |
668 | ||
669 | #undef TARGET_INIT_BUILTINS | |
670 | #define TARGET_INIT_BUILTINS arc_init_builtins | |
671 | ||
672 | #undef TARGET_EXPAND_BUILTIN | |
673 | #define TARGET_EXPAND_BUILTIN arc_expand_builtin | |
674 | ||
c69899f0 CZ |
675 | #undef TARGET_BUILTIN_DECL |
676 | #define TARGET_BUILTIN_DECL arc_builtin_decl | |
677 | ||
526b7aee SV |
678 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
679 | #define TARGET_ASM_OUTPUT_MI_THUNK arc_output_mi_thunk | |
680 | ||
681 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
682 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true | |
683 | ||
684 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
685 | #define TARGET_FUNCTION_OK_FOR_SIBCALL arc_function_ok_for_sibcall | |
686 | ||
687 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
688 | #define TARGET_MACHINE_DEPENDENT_REORG arc_reorg | |
689 | ||
690 | #undef TARGET_IN_SMALL_DATA_P | |
691 | #define TARGET_IN_SMALL_DATA_P arc_in_small_data_p | |
692 | ||
693 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
694 | #define TARGET_PROMOTE_FUNCTION_MODE \ | |
695 | default_promote_function_mode_always_promote | |
696 | ||
697 | #undef TARGET_PROMOTE_PROTOTYPES | |
698 | #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true | |
699 | ||
700 | #undef TARGET_RETURN_IN_MEMORY | |
701 | #define TARGET_RETURN_IN_MEMORY arc_return_in_memory | |
702 | #undef TARGET_PASS_BY_REFERENCE | |
703 | #define TARGET_PASS_BY_REFERENCE arc_pass_by_reference | |
704 | ||
705 | #undef TARGET_SETUP_INCOMING_VARARGS | |
706 | #define TARGET_SETUP_INCOMING_VARARGS arc_setup_incoming_varargs | |
707 | ||
708 | #undef TARGET_ARG_PARTIAL_BYTES | |
709 | #define TARGET_ARG_PARTIAL_BYTES arc_arg_partial_bytes | |
710 | ||
711 | #undef TARGET_MUST_PASS_IN_STACK | |
712 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
713 | ||
714 | #undef TARGET_FUNCTION_VALUE | |
715 | #define TARGET_FUNCTION_VALUE arc_function_value | |
716 | ||
717 | #undef TARGET_SCHED_ADJUST_PRIORITY | |
718 | #define TARGET_SCHED_ADJUST_PRIORITY arc_sched_adjust_priority | |
719 | ||
43bb0fc2 CZ |
720 | #undef TARGET_SCHED_ISSUE_RATE |
721 | #define TARGET_SCHED_ISSUE_RATE arc_sched_issue_rate | |
722 | ||
526b7aee SV |
723 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
724 | #define TARGET_VECTOR_MODE_SUPPORTED_P arc_vector_mode_supported_p | |
725 | ||
00c072ae CZ |
726 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
727 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE arc_preferred_simd_mode | |
728 | ||
729 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES | |
730 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES arc_autovectorize_vector_sizes | |
731 | ||
1d0216c8 RS |
732 | #undef TARGET_CAN_USE_DOLOOP_P |
733 | #define TARGET_CAN_USE_DOLOOP_P arc_can_use_doloop_p | |
734 | ||
526b7aee SV |
735 | #undef TARGET_INVALID_WITHIN_DOLOOP |
736 | #define TARGET_INVALID_WITHIN_DOLOOP arc_invalid_within_doloop | |
737 | ||
738 | #undef TARGET_PRESERVE_RELOAD_P | |
739 | #define TARGET_PRESERVE_RELOAD_P arc_preserve_reload_p | |
740 | ||
741 | #undef TARGET_CAN_FOLLOW_JUMP | |
742 | #define TARGET_CAN_FOLLOW_JUMP arc_can_follow_jump | |
743 | ||
744 | #undef TARGET_DELEGITIMIZE_ADDRESS | |
745 | #define TARGET_DELEGITIMIZE_ADDRESS arc_delegitimize_address | |
746 | ||
ad23f5d4 JG |
747 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
748 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
749 | arc_use_by_pieces_infrastructure_p | |
750 | ||
526b7aee SV |
751 | /* Usually, we will be able to scale anchor offsets. |
752 | When this fails, we want LEGITIMIZE_ADDRESS to kick in. */ | |
753 | #undef TARGET_MIN_ANCHOR_OFFSET | |
754 | #define TARGET_MIN_ANCHOR_OFFSET (-1024) | |
755 | #undef TARGET_MAX_ANCHOR_OFFSET | |
756 | #define TARGET_MAX_ANCHOR_OFFSET (1020) | |
757 | ||
758 | #undef TARGET_SECONDARY_RELOAD | |
759 | #define TARGET_SECONDARY_RELOAD arc_secondary_reload | |
760 | ||
761 | #define TARGET_OPTION_OVERRIDE arc_override_options | |
762 | ||
763 | #define TARGET_CONDITIONAL_REGISTER_USAGE arc_conditional_register_usage | |
764 | ||
765 | #define TARGET_TRAMPOLINE_INIT arc_initialize_trampoline | |
766 | ||
526b7aee SV |
767 | #define TARGET_CAN_ELIMINATE arc_can_eliminate |
768 | ||
769 | #define TARGET_FRAME_POINTER_REQUIRED arc_frame_pointer_required | |
770 | ||
771 | #define TARGET_FUNCTION_ARG arc_function_arg | |
772 | ||
773 | #define TARGET_FUNCTION_ARG_ADVANCE arc_function_arg_advance | |
774 | ||
775 | #define TARGET_LEGITIMATE_CONSTANT_P arc_legitimate_constant_p | |
776 | ||
777 | #define TARGET_LEGITIMATE_ADDRESS_P arc_legitimate_address_p | |
778 | ||
779 | #define TARGET_MODE_DEPENDENT_ADDRESS_P arc_mode_dependent_address_p | |
780 | ||
781 | #define TARGET_LEGITIMIZE_ADDRESS arc_legitimize_address | |
782 | ||
bf9e9dc5 CZ |
783 | #undef TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P |
784 | #define TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P \ | |
785 | arc_no_speculation_in_delay_slots_p | |
786 | ||
53c8d5a7 | 787 | #undef TARGET_LRA_P |
526b7aee SV |
788 | #define TARGET_LRA_P arc_lra_p |
789 | #define TARGET_REGISTER_PRIORITY arc_register_priority | |
790 | /* Stores with scaled offsets have different displacement ranges. */ | |
791 | #define TARGET_DIFFERENT_ADDR_DISPLACEMENT_P hook_bool_void_true | |
792 | #define TARGET_SPILL_CLASS arc_spill_class | |
793 | ||
1825c61e CZ |
794 | #undef TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS |
795 | #define TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS arc_allocate_stack_slots_for_args | |
796 | ||
797 | #undef TARGET_WARN_FUNC_RETURN | |
798 | #define TARGET_WARN_FUNC_RETURN arc_warn_func_return | |
799 | ||
526b7aee SV |
800 | #include "target-def.h" |
801 | ||
802 | #undef TARGET_ASM_ALIGNED_HI_OP | |
803 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
804 | #undef TARGET_ASM_ALIGNED_SI_OP | |
805 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
806 | ||
28633bbd CZ |
807 | #ifdef HAVE_AS_TLS |
808 | #undef TARGET_HAVE_TLS | |
809 | #define TARGET_HAVE_TLS HAVE_AS_TLS | |
810 | #endif | |
811 | ||
d34a0fdc CZ |
812 | #undef TARGET_DWARF_REGISTER_SPAN |
813 | #define TARGET_DWARF_REGISTER_SPAN arc_dwarf_register_span | |
814 | ||
c43f4279 RS |
815 | #undef TARGET_HARD_REGNO_NREGS |
816 | #define TARGET_HARD_REGNO_NREGS arc_hard_regno_nregs | |
f939c3e6 RS |
817 | #undef TARGET_HARD_REGNO_MODE_OK |
818 | #define TARGET_HARD_REGNO_MODE_OK arc_hard_regno_mode_ok | |
819 | ||
99e1629f RS |
820 | #undef TARGET_MODES_TIEABLE_P |
821 | #define TARGET_MODES_TIEABLE_P arc_modes_tieable_p | |
822 | ||
526b7aee SV |
823 | /* Try to keep the (mov:DF _, reg) as early as possible so |
824 | that the d<add/sub/mul>h-lr insns appear together and can | |
825 | use the peephole2 pattern. */ | |
826 | ||
827 | static int | |
ac44248e | 828 | arc_sched_adjust_priority (rtx_insn *insn, int priority) |
526b7aee SV |
829 | { |
830 | rtx set = single_set (insn); | |
831 | if (set | |
832 | && GET_MODE (SET_SRC(set)) == DFmode | |
833 | && GET_CODE (SET_SRC(set)) == REG) | |
834 | { | |
835 | /* Incrementing priority by 20 (empirically derived). */ | |
836 | return priority + 20; | |
837 | } | |
838 | ||
839 | return priority; | |
840 | } | |
841 | ||
f50bb868 CZ |
842 | /* For ARC base register + offset addressing, the validity of the |
843 | address is mode-dependent for most of the offset range, as the | |
844 | offset can be scaled by the access size. | |
845 | We don't expose these as mode-dependent addresses in the | |
846 | mode_dependent_address_p target hook, because that would disable | |
847 | lots of optimizations, and most uses of these addresses are for 32 | |
848 | or 64 bit accesses anyways, which are fine. | |
849 | However, that leaves some addresses for 8 / 16 bit values not | |
850 | properly reloaded by the generic code, which is why we have to | |
851 | schedule secondary reloads for these. */ | |
852 | ||
526b7aee | 853 | static reg_class_t |
f50bb868 CZ |
854 | arc_secondary_reload (bool in_p, |
855 | rtx x, | |
856 | reg_class_t cl, | |
857 | machine_mode mode, | |
858 | secondary_reload_info *sri) | |
526b7aee | 859 | { |
f50bb868 CZ |
860 | enum rtx_code code = GET_CODE (x); |
861 | ||
526b7aee SV |
862 | if (cl == DOUBLE_REGS) |
863 | return GENERAL_REGS; | |
864 | ||
f50bb868 CZ |
865 | /* If we have a subreg (reg), where reg is a pseudo (that will end in |
866 | a memory location), then we may need a scratch register to handle | |
867 | the fp/sp+largeoffset address. */ | |
868 | if (code == SUBREG) | |
869 | { | |
870 | rtx addr = NULL_RTX; | |
871 | x = SUBREG_REG (x); | |
872 | ||
873 | if (REG_P (x)) | |
874 | { | |
875 | int regno = REGNO (x); | |
876 | if (regno >= FIRST_PSEUDO_REGISTER) | |
877 | regno = reg_renumber[regno]; | |
878 | ||
879 | if (regno != -1) | |
880 | return NO_REGS; | |
881 | ||
73dac59b CZ |
882 | /* It is a pseudo that ends in a stack location. This |
883 | procedure only works with the old reload step. */ | |
884 | if (reg_equiv_mem (REGNO (x)) && !lra_in_progress) | |
f50bb868 CZ |
885 | { |
886 | /* Get the equivalent address and check the range of the | |
887 | offset. */ | |
888 | rtx mem = reg_equiv_mem (REGNO (x)); | |
889 | addr = find_replacement (&XEXP (mem, 0)); | |
890 | } | |
891 | } | |
892 | else | |
893 | { | |
894 | gcc_assert (MEM_P (x)); | |
895 | addr = XEXP (x, 0); | |
896 | addr = simplify_rtx (addr); | |
897 | } | |
898 | if (addr && GET_CODE (addr) == PLUS | |
899 | && CONST_INT_P (XEXP (addr, 1)) | |
900 | && (!RTX_OK_FOR_OFFSET_P (mode, XEXP (addr, 1)))) | |
901 | { | |
902 | switch (mode) | |
903 | { | |
4e10a5a7 | 904 | case E_QImode: |
f50bb868 CZ |
905 | sri->icode = |
906 | in_p ? CODE_FOR_reload_qi_load : CODE_FOR_reload_qi_store; | |
907 | break; | |
4e10a5a7 | 908 | case E_HImode: |
f50bb868 CZ |
909 | sri->icode = |
910 | in_p ? CODE_FOR_reload_hi_load : CODE_FOR_reload_hi_store; | |
911 | break; | |
912 | default: | |
913 | break; | |
914 | } | |
915 | } | |
916 | } | |
526b7aee SV |
917 | return NO_REGS; |
918 | } | |
919 | ||
f50bb868 CZ |
920 | /* Convert reloads using offsets that are too large to use indirect |
921 | addressing. */ | |
922 | ||
923 | void | |
924 | arc_secondary_reload_conv (rtx reg, rtx mem, rtx scratch, bool store_p) | |
925 | { | |
926 | rtx addr; | |
927 | ||
928 | gcc_assert (GET_CODE (mem) == MEM); | |
929 | addr = XEXP (mem, 0); | |
930 | ||
931 | /* Large offset: use a move. FIXME: ld ops accepts limms as | |
932 | offsets. Hence, the following move insn is not required. */ | |
933 | emit_move_insn (scratch, addr); | |
934 | mem = replace_equiv_address_nv (mem, scratch); | |
935 | ||
936 | /* Now create the move. */ | |
937 | if (store_p) | |
938 | emit_insn (gen_rtx_SET (mem, reg)); | |
939 | else | |
940 | emit_insn (gen_rtx_SET (reg, mem)); | |
941 | ||
942 | return; | |
943 | } | |
944 | ||
526b7aee SV |
945 | static unsigned arc_ifcvt (void); |
946 | ||
947 | namespace { | |
948 | ||
949 | const pass_data pass_data_arc_ifcvt = | |
950 | { | |
951 | RTL_PASS, | |
952 | "arc_ifcvt", /* name */ | |
953 | OPTGROUP_NONE, /* optinfo_flags */ | |
526b7aee SV |
954 | TV_IFCVT2, /* tv_id */ |
955 | 0, /* properties_required */ | |
956 | 0, /* properties_provided */ | |
957 | 0, /* properties_destroyed */ | |
958 | 0, /* todo_flags_start */ | |
959 | TODO_df_finish /* todo_flags_finish */ | |
960 | }; | |
961 | ||
962 | class pass_arc_ifcvt : public rtl_opt_pass | |
963 | { | |
964 | public: | |
965 | pass_arc_ifcvt(gcc::context *ctxt) | |
966 | : rtl_opt_pass(pass_data_arc_ifcvt, ctxt) | |
967 | {} | |
968 | ||
969 | /* opt_pass methods: */ | |
cd4dd8f0 | 970 | opt_pass * clone () { return new pass_arc_ifcvt (m_ctxt); } |
be55bfe6 | 971 | virtual unsigned int execute (function *) { return arc_ifcvt (); } |
526b7aee SV |
972 | }; |
973 | ||
974 | } // anon namespace | |
975 | ||
976 | rtl_opt_pass * | |
977 | make_pass_arc_ifcvt (gcc::context *ctxt) | |
978 | { | |
979 | return new pass_arc_ifcvt (ctxt); | |
980 | } | |
981 | ||
0bc69b81 JR |
982 | static unsigned arc_predicate_delay_insns (void); |
983 | ||
984 | namespace { | |
985 | ||
986 | const pass_data pass_data_arc_predicate_delay_insns = | |
987 | { | |
988 | RTL_PASS, | |
989 | "arc_predicate_delay_insns", /* name */ | |
990 | OPTGROUP_NONE, /* optinfo_flags */ | |
0bc69b81 JR |
991 | TV_IFCVT2, /* tv_id */ |
992 | 0, /* properties_required */ | |
993 | 0, /* properties_provided */ | |
994 | 0, /* properties_destroyed */ | |
995 | 0, /* todo_flags_start */ | |
996 | TODO_df_finish /* todo_flags_finish */ | |
997 | }; | |
998 | ||
999 | class pass_arc_predicate_delay_insns : public rtl_opt_pass | |
1000 | { | |
1001 | public: | |
1002 | pass_arc_predicate_delay_insns(gcc::context *ctxt) | |
1003 | : rtl_opt_pass(pass_data_arc_predicate_delay_insns, ctxt) | |
1004 | {} | |
1005 | ||
1006 | /* opt_pass methods: */ | |
be55bfe6 TS |
1007 | virtual unsigned int execute (function *) |
1008 | { | |
1009 | return arc_predicate_delay_insns (); | |
1010 | } | |
0bc69b81 JR |
1011 | }; |
1012 | ||
1013 | } // anon namespace | |
1014 | ||
1015 | rtl_opt_pass * | |
1016 | make_pass_arc_predicate_delay_insns (gcc::context *ctxt) | |
1017 | { | |
1018 | return new pass_arc_predicate_delay_insns (ctxt); | |
1019 | } | |
1020 | ||
526b7aee SV |
1021 | /* Called by OVERRIDE_OPTIONS to initialize various things. */ |
1022 | ||
f9ccf899 | 1023 | static void |
526b7aee SV |
1024 | arc_init (void) |
1025 | { | |
0e5172eb CZ |
1026 | if (TARGET_V2) |
1027 | { | |
1028 | /* I have the multiplier, then use it*/ | |
1029 | if (TARGET_MPYW || TARGET_MULTI) | |
1030 | arc_multcost = COSTS_N_INSNS (1); | |
1031 | } | |
526b7aee SV |
1032 | /* Note: arc_multcost is only used in rtx_cost if speed is true. */ |
1033 | if (arc_multcost < 0) | |
1034 | switch (arc_tune) | |
1035 | { | |
62f26645 | 1036 | case ARC_TUNE_ARC700_4_2_STD: |
526b7aee SV |
1037 | /* latency 7; |
1038 | max throughput (1 multiply + 4 other insns) / 5 cycles. */ | |
1039 | arc_multcost = COSTS_N_INSNS (4); | |
1040 | if (TARGET_NOMPY_SET) | |
1041 | arc_multcost = COSTS_N_INSNS (30); | |
1042 | break; | |
62f26645 | 1043 | case ARC_TUNE_ARC700_4_2_XMAC: |
526b7aee SV |
1044 | /* latency 5; |
1045 | max throughput (1 multiply + 2 other insns) / 3 cycles. */ | |
1046 | arc_multcost = COSTS_N_INSNS (3); | |
1047 | if (TARGET_NOMPY_SET) | |
1048 | arc_multcost = COSTS_N_INSNS (30); | |
1049 | break; | |
62f26645 | 1050 | case ARC_TUNE_ARC600: |
526b7aee SV |
1051 | if (TARGET_MUL64_SET) |
1052 | { | |
1053 | arc_multcost = COSTS_N_INSNS (4); | |
1054 | break; | |
1055 | } | |
1056 | /* Fall through. */ | |
1057 | default: | |
1058 | arc_multcost = COSTS_N_INSNS (30); | |
1059 | break; | |
1060 | } | |
1061 | ||
f50bb868 CZ |
1062 | /* MPY instructions valid only for ARC700 or ARCv2. */ |
1063 | if (TARGET_NOMPY_SET && TARGET_ARC600_FAMILY) | |
a3f9f006 | 1064 | error ("%<-mno-mpy%> supported only for ARC700 or ARCv2"); |
526b7aee | 1065 | |
526b7aee | 1066 | if (!TARGET_DPFP && TARGET_DPFP_DISABLE_LRSR) |
a3f9f006 | 1067 | error ("%<-mno-dpfp-lrsr%> supported only with %<-mdpfp%>"); |
526b7aee SV |
1068 | |
1069 | /* FPX-1. No fast and compact together. */ | |
1070 | if ((TARGET_DPFP_FAST_SET && TARGET_DPFP_COMPACT_SET) | |
1071 | || (TARGET_SPFP_FAST_SET && TARGET_SPFP_COMPACT_SET)) | |
1072 | error ("FPX fast and compact options cannot be specified together"); | |
1073 | ||
1074 | /* FPX-2. No fast-spfp for arc600 or arc601. */ | |
f50bb868 | 1075 | if (TARGET_SPFP_FAST_SET && TARGET_ARC600_FAMILY) |
a3f9f006 | 1076 | error ("%<-mspfp_fast%> not available on ARC600 or ARC601"); |
526b7aee | 1077 | |
f9ccf899 CZ |
1078 | /* FPX-4. No FPX extensions mixed with FPU extensions. */ |
1079 | if ((TARGET_DPFP_FAST_SET || TARGET_DPFP_COMPACT_SET || TARGET_SPFP) | |
1080 | && TARGET_HARD_FLOAT) | |
2fa9c1f6 | 1081 | error ("no FPX/FPU mixing allowed"); |
8f3304d0 | 1082 | |
526b7aee | 1083 | /* Warn for unimplemented PIC in pre-ARC700 cores, and disable flag_pic. */ |
f50bb868 | 1084 | if (flag_pic && TARGET_ARC600_FAMILY) |
526b7aee | 1085 | { |
fe3ddee9 | 1086 | warning (0, "PIC is not supported for %qs", |
f50bb868 | 1087 | arc_cpu_string); |
526b7aee SV |
1088 | flag_pic = 0; |
1089 | } | |
1090 | ||
1091 | arc_init_reg_tables (); | |
1092 | ||
1093 | /* Initialize array for PRINT_OPERAND_PUNCT_VALID_P. */ | |
1094 | memset (arc_punct_chars, 0, sizeof (arc_punct_chars)); | |
1095 | arc_punct_chars['#'] = 1; | |
1096 | arc_punct_chars['*'] = 1; | |
1097 | arc_punct_chars['?'] = 1; | |
1098 | arc_punct_chars['!'] = 1; | |
1099 | arc_punct_chars['^'] = 1; | |
1100 | arc_punct_chars['&'] = 1; | |
f50bb868 CZ |
1101 | arc_punct_chars['+'] = 1; |
1102 | arc_punct_chars['_'] = 1; | |
526b7aee SV |
1103 | |
1104 | if (optimize > 1 && !TARGET_NO_COND_EXEC) | |
1105 | { | |
1106 | /* There are two target-independent ifcvt passes, and arc_reorg may do | |
1107 | one or more arc_ifcvt calls. */ | |
1108 | opt_pass *pass_arc_ifcvt_4 = make_pass_arc_ifcvt (g); | |
1109 | struct register_pass_info arc_ifcvt4_info | |
1110 | = { pass_arc_ifcvt_4, "dbr", 1, PASS_POS_INSERT_AFTER }; | |
1111 | struct register_pass_info arc_ifcvt5_info | |
1112 | = { pass_arc_ifcvt_4->clone (), "shorten", 1, PASS_POS_INSERT_BEFORE }; | |
1113 | ||
1114 | register_pass (&arc_ifcvt4_info); | |
1115 | register_pass (&arc_ifcvt5_info); | |
1116 | } | |
0bc69b81 JR |
1117 | |
1118 | if (flag_delayed_branch) | |
1119 | { | |
1120 | opt_pass *pass_arc_predicate_delay_insns | |
1121 | = make_pass_arc_predicate_delay_insns (g); | |
1122 | struct register_pass_info arc_predicate_delay_info | |
1123 | = { pass_arc_predicate_delay_insns, "dbr", 1, PASS_POS_INSERT_AFTER }; | |
1124 | ||
1125 | register_pass (&arc_predicate_delay_info); | |
1126 | } | |
526b7aee SV |
1127 | } |
1128 | ||
41453183 CZ |
1129 | /* Parse -mirq-ctrl-saved=RegisterRange, blink, lp_copunt. The |
1130 | register range is specified as two registers separated by a dash. | |
1131 | It always starts with r0, and its upper limit is fp register. | |
1132 | blink and lp_count registers are optional. */ | |
1133 | ||
1134 | static void | |
1135 | irq_range (const char *cstr) | |
1136 | { | |
1137 | int i, first, last, blink, lpcount, xreg; | |
1138 | char *str, *dash, *comma; | |
1139 | ||
1140 | i = strlen (cstr); | |
1141 | str = (char *) alloca (i + 1); | |
1142 | memcpy (str, cstr, i + 1); | |
1143 | blink = -1; | |
1144 | lpcount = -1; | |
1145 | ||
1146 | dash = strchr (str, '-'); | |
1147 | if (!dash) | |
1148 | { | |
d65485c5 | 1149 | warning (OPT_mirq_ctrl_saved_, "missing dash"); |
41453183 CZ |
1150 | return; |
1151 | } | |
1152 | *dash = '\0'; | |
1153 | ||
1154 | comma = strchr (dash + 1, ','); | |
1155 | if (comma) | |
1156 | *comma = '\0'; | |
1157 | ||
1158 | first = decode_reg_name (str); | |
1159 | if (first != 0) | |
1160 | { | |
d65485c5 | 1161 | warning (OPT_mirq_ctrl_saved_, "first register must be R0"); |
41453183 CZ |
1162 | return; |
1163 | } | |
1164 | ||
1165 | /* At this moment we do not have the register names initialized | |
1166 | accordingly. */ | |
1167 | if (!strcmp (dash + 1, "ilink")) | |
1168 | last = 29; | |
1169 | else | |
1170 | last = decode_reg_name (dash + 1); | |
1171 | ||
1172 | if (last < 0) | |
1173 | { | |
d65485c5 | 1174 | warning (OPT_mirq_ctrl_saved_, "unknown register name: %s", dash + 1); |
41453183 CZ |
1175 | return; |
1176 | } | |
1177 | ||
1178 | if (!(last & 0x01)) | |
1179 | { | |
d65485c5 CZ |
1180 | warning (OPT_mirq_ctrl_saved_, |
1181 | "last register name %s must be an odd register", dash + 1); | |
41453183 CZ |
1182 | return; |
1183 | } | |
1184 | ||
1185 | *dash = '-'; | |
1186 | ||
1187 | if (first > last) | |
1188 | { | |
d65485c5 CZ |
1189 | warning (OPT_mirq_ctrl_saved_, |
1190 | "%s-%s is an empty range", str, dash + 1); | |
41453183 CZ |
1191 | return; |
1192 | } | |
1193 | ||
1194 | while (comma) | |
1195 | { | |
1196 | *comma = ','; | |
1197 | str = comma + 1; | |
1198 | ||
1199 | comma = strchr (str, ','); | |
1200 | if (comma) | |
1201 | *comma = '\0'; | |
1202 | ||
1203 | xreg = decode_reg_name (str); | |
1204 | switch (xreg) | |
1205 | { | |
1206 | case 31: | |
1207 | blink = 31; | |
1208 | break; | |
1209 | ||
1210 | case 60: | |
1211 | lpcount = 60; | |
1212 | break; | |
1213 | ||
1214 | default: | |
d65485c5 CZ |
1215 | warning (OPT_mirq_ctrl_saved_, |
1216 | "unknown register name: %s", str); | |
41453183 CZ |
1217 | return; |
1218 | } | |
1219 | } | |
1220 | ||
1221 | irq_ctrl_saved.irq_save_last_reg = last; | |
1222 | irq_ctrl_saved.irq_save_blink = (blink == 31) || (last == 31); | |
1223 | irq_ctrl_saved.irq_save_lpcount = (lpcount == 60); | |
1224 | } | |
1225 | ||
c7314bc1 CZ |
1226 | /* Parse -mrgf-banked-regs=NUM option string. Valid values for NUM are 4, |
1227 | 8, 16, or 32. */ | |
1228 | ||
1229 | static void | |
1230 | parse_mrgf_banked_regs_option (const char *arg) | |
1231 | { | |
1232 | long int val; | |
1233 | char *end_ptr; | |
1234 | ||
1235 | errno = 0; | |
1236 | val = strtol (arg, &end_ptr, 10); | |
1237 | if (errno != 0 || *arg == '\0' || *end_ptr != '\0' | |
1238 | || (val != 0 && val != 4 && val != 8 && val != 16 && val != 32)) | |
1239 | { | |
a3f9f006 | 1240 | error ("invalid number in %<-mrgf-banked-regs=%s%> " |
c7314bc1 CZ |
1241 | "valid values are 0, 4, 8, 16, or 32", arg); |
1242 | return; | |
1243 | } | |
1244 | rgf_banked_register_count = (int) val; | |
1245 | } | |
1246 | ||
526b7aee SV |
1247 | /* Check ARC options, generate derived target attributes. */ |
1248 | ||
1249 | static void | |
1250 | arc_override_options (void) | |
1251 | { | |
41453183 CZ |
1252 | unsigned int i; |
1253 | cl_deferred_option *opt; | |
1254 | vec<cl_deferred_option> *vopt | |
1255 | = (vec<cl_deferred_option> *) arc_deferred_options; | |
1256 | ||
526b7aee | 1257 | if (arc_cpu == PROCESSOR_NONE) |
f9ccf899 CZ |
1258 | arc_cpu = TARGET_CPU_DEFAULT; |
1259 | ||
1260 | /* Set the default cpu options. */ | |
1261 | arc_selected_cpu = &arc_cpu_types[(int) arc_cpu]; | |
f9ccf899 CZ |
1262 | |
1263 | /* Set the architectures. */ | |
c3bde35a | 1264 | switch (arc_selected_cpu->arch_info->arch_id) |
f9ccf899 CZ |
1265 | { |
1266 | case BASE_ARCH_em: | |
1267 | arc_cpu_string = "EM"; | |
1268 | break; | |
1269 | case BASE_ARCH_hs: | |
1270 | arc_cpu_string = "HS"; | |
1271 | break; | |
1272 | case BASE_ARCH_700: | |
1273 | if (arc_selected_cpu->processor == PROCESSOR_nps400) | |
1274 | arc_cpu_string = "NPS400"; | |
1275 | else | |
1276 | arc_cpu_string = "ARC700"; | |
1277 | break; | |
1278 | case BASE_ARCH_6xx: | |
1279 | arc_cpu_string = "ARC600"; | |
1280 | break; | |
1281 | default: | |
1282 | gcc_unreachable (); | |
1283 | } | |
1284 | ||
41453183 CZ |
1285 | irq_ctrl_saved.irq_save_last_reg = -1; |
1286 | irq_ctrl_saved.irq_save_blink = false; | |
1287 | irq_ctrl_saved.irq_save_lpcount = false; | |
1288 | ||
c7314bc1 CZ |
1289 | rgf_banked_register_count = 0; |
1290 | ||
41453183 CZ |
1291 | /* Handle the deferred options. */ |
1292 | if (vopt) | |
1293 | FOR_EACH_VEC_ELT (*vopt, i, opt) | |
1294 | { | |
1295 | switch (opt->opt_index) | |
1296 | { | |
1297 | case OPT_mirq_ctrl_saved_: | |
1298 | if (TARGET_V2) | |
1299 | irq_range (opt->arg); | |
1300 | else | |
d65485c5 | 1301 | warning (OPT_mirq_ctrl_saved_, |
a3f9f006 ML |
1302 | "option %<-mirq-ctrl-saved%> valid only " |
1303 | "for ARC v2 processors"); | |
41453183 CZ |
1304 | break; |
1305 | ||
c7314bc1 CZ |
1306 | case OPT_mrgf_banked_regs_: |
1307 | if (TARGET_V2) | |
1308 | parse_mrgf_banked_regs_option (opt->arg); | |
1309 | else | |
d65485c5 | 1310 | warning (OPT_mrgf_banked_regs_, |
a3f9f006 ML |
1311 | "option %<-mrgf-banked-regs%> valid only for " |
1312 | "ARC v2 processors"); | |
c7314bc1 CZ |
1313 | break; |
1314 | ||
41453183 CZ |
1315 | default: |
1316 | gcc_unreachable(); | |
1317 | } | |
1318 | } | |
1319 | ||
66825a30 CZ |
1320 | CLEAR_HARD_REG_SET (overrideregs); |
1321 | if (common_deferred_options) | |
1322 | { | |
1323 | vec<cl_deferred_option> v = | |
1324 | *((vec<cl_deferred_option> *) common_deferred_options); | |
1325 | int reg, nregs, j; | |
1326 | ||
1327 | FOR_EACH_VEC_ELT (v, i, opt) | |
1328 | { | |
1329 | switch (opt->opt_index) | |
1330 | { | |
1331 | case OPT_ffixed_: | |
1332 | case OPT_fcall_used_: | |
1333 | case OPT_fcall_saved_: | |
1334 | if ((reg = decode_reg_name_and_count (opt->arg, &nregs)) >= 0) | |
1335 | for (j = reg; j < reg + nregs; j++) | |
1336 | SET_HARD_REG_BIT (overrideregs, j); | |
1337 | break; | |
1338 | default: | |
1339 | break; | |
1340 | } | |
1341 | } | |
1342 | } | |
1343 | ||
d65485c5 CZ |
1344 | /* Check options against architecture options. Throw an error if |
1345 | option is not allowed. Extra, check options against default | |
1346 | architecture/cpu flags and throw an warning if we find a | |
1347 | mismatch. */ | |
fe3ddee9 CZ |
1348 | /* TRANSLATORS: the DOC/DOC0/DOC1 are strings which shouldn't be |
1349 | translated. They are like keywords which one can relate with the | |
1350 | architectural choices taken for an ARC CPU implementation. */ | |
d65485c5 CZ |
1351 | #define ARC_OPTX(NAME, CODE, VAR, VAL, DOC0, DOC1) \ |
1352 | do { \ | |
1353 | if ((!(arc_selected_cpu->arch_info->flags & CODE)) \ | |
1354 | && (VAR == VAL)) \ | |
fe3ddee9 | 1355 | error ("option %<%s=%s%> is not available for %qs CPU", \ |
d65485c5 CZ |
1356 | DOC0, DOC1, arc_selected_cpu->name); \ |
1357 | if ((arc_selected_cpu->arch_info->dflags & CODE) \ | |
1358 | && (VAR != DEFAULT_##VAR) \ | |
1359 | && (VAR != VAL)) \ | |
fe3ddee9 CZ |
1360 | warning (0, "option %qs is ignored, the default value %qs" \ |
1361 | " is considered for %qs CPU", DOC0, DOC1, \ | |
d65485c5 CZ |
1362 | arc_selected_cpu->name); \ |
1363 | } while (0); | |
1364 | #define ARC_OPT(NAME, CODE, MASK, DOC) \ | |
1365 | do { \ | |
1366 | if ((!(arc_selected_cpu->arch_info->flags & CODE)) \ | |
1367 | && (target_flags & MASK)) \ | |
fe3ddee9 | 1368 | error ("option %qs is not available for %qs CPU", \ |
d65485c5 CZ |
1369 | DOC, arc_selected_cpu->name); \ |
1370 | if ((arc_selected_cpu->arch_info->dflags & CODE) \ | |
1371 | && (target_flags_explicit & MASK) \ | |
1372 | && (!(target_flags & MASK))) \ | |
fe3ddee9 CZ |
1373 | warning (0, "unset option %qs is ignored, it is always" \ |
1374 | " enabled for %qs CPU", DOC, \ | |
d65485c5 CZ |
1375 | arc_selected_cpu->name); \ |
1376 | } while (0); | |
1377 | ||
1378 | #include "arc-options.def" | |
1379 | ||
1380 | #undef ARC_OPTX | |
1381 | #undef ARC_OPT | |
1382 | ||
f9ccf899 CZ |
1383 | /* Set cpu flags accordingly to architecture/selected cpu. The cpu |
1384 | specific flags are set in arc-common.c. The architecture forces | |
1385 | the default hardware configurations in, regardless what command | |
1386 | line options are saying. The CPU optional hw options can be | |
1387 | turned on or off. */ | |
1388 | #define ARC_OPT(NAME, CODE, MASK, DOC) \ | |
1389 | do { \ | |
1390 | if ((arc_selected_cpu->flags & CODE) \ | |
1391 | && ((target_flags_explicit & MASK) == 0)) \ | |
1392 | target_flags |= MASK; \ | |
c3bde35a | 1393 | if (arc_selected_cpu->arch_info->dflags & CODE) \ |
f9ccf899 CZ |
1394 | target_flags |= MASK; \ |
1395 | } while (0); | |
d65485c5 | 1396 | #define ARC_OPTX(NAME, CODE, VAR, VAL, DOC0, DOC1) \ |
c3bde35a AB |
1397 | do { \ |
1398 | if ((arc_selected_cpu->flags & CODE) \ | |
1399 | && (VAR == DEFAULT_##VAR)) \ | |
1400 | VAR = VAL; \ | |
1401 | if (arc_selected_cpu->arch_info->dflags & CODE) \ | |
1402 | VAR = VAL; \ | |
f9ccf899 CZ |
1403 | } while (0); |
1404 | ||
1405 | #include "arc-options.def" | |
1406 | ||
f9ccf899 CZ |
1407 | #undef ARC_OPTX |
1408 | #undef ARC_OPT | |
1409 | ||
09d69286 CZ |
1410 | /* Set extras. */ |
1411 | switch (arc_selected_cpu->extra) | |
1412 | { | |
1413 | case HAS_LPCOUNT_16: | |
1414 | arc_lpcwidth = 16; | |
1415 | break; | |
1416 | default: | |
1417 | break; | |
1418 | } | |
1419 | ||
f9ccf899 | 1420 | /* Set Tune option. */ |
62f26645 CZ |
1421 | if (arc_tune == ARC_TUNE_NONE) |
1422 | arc_tune = (enum arc_tune_attr) arc_selected_cpu->tune; | |
526b7aee SV |
1423 | |
1424 | if (arc_size_opt_level == 3) | |
1425 | optimize_size = 1; | |
1426 | ||
f5d56cf9 CZ |
1427 | if (TARGET_V2 && optimize_size && (ATTRIBUTE_PCS == 2)) |
1428 | TARGET_CODE_DENSITY_FRAME = 1; | |
1429 | ||
526b7aee SV |
1430 | if (flag_pic) |
1431 | target_flags |= MASK_NO_SDATA_SET; | |
1432 | ||
1433 | if (flag_no_common == 255) | |
1434 | flag_no_common = !TARGET_NO_SDATA_SET; | |
1435 | ||
526b7aee SV |
1436 | if (TARGET_MIXED_CODE) |
1437 | TARGET_Q_CLASS = 1; | |
526b7aee | 1438 | |
9f532472 CZ |
1439 | /* Check for small data option */ |
1440 | if (!global_options_set.x_g_switch_value && !TARGET_NO_SDATA_SET) | |
1441 | g_switch_value = TARGET_LL64 ? 8 : 4; | |
1442 | ||
635aeaa2 CZ |
1443 | /* A7 has an issue with delay slots. */ |
1444 | if (TARGET_ARC700 && (arc_tune != ARC_TUNE_ARC7XX)) | |
1445 | flag_delayed_branch = 0; | |
1446 | ||
90b48013 CZ |
1447 | /* Millicode thunks doesn't work with long calls. */ |
1448 | if (TARGET_LONG_CALLS_SET) | |
1449 | target_flags &= ~MASK_MILLICODE_THUNK_SET; | |
1450 | ||
9f54ba8f CZ |
1451 | /* Set unaligned to all HS cpus. */ |
1452 | if (!global_options_set.x_unaligned_access && TARGET_HS) | |
1453 | unaligned_access = 1; | |
1454 | ||
526b7aee SV |
1455 | /* These need to be done at start up. It's convenient to do them here. */ |
1456 | arc_init (); | |
1457 | } | |
1458 | ||
1459 | /* The condition codes of the ARC, and the inverse function. */ | |
1460 | /* For short branches, the "c" / "nc" names are not defined in the ARC | |
1461 | Programmers manual, so we have to use "lo" / "hs"" instead. */ | |
1462 | static const char *arc_condition_codes[] = | |
1463 | { | |
1464 | "al", 0, "eq", "ne", "p", "n", "lo", "hs", "v", "nv", | |
1465 | "gt", "le", "ge", "lt", "hi", "ls", "pnz", 0 | |
1466 | }; | |
1467 | ||
1468 | enum arc_cc_code_index | |
1469 | { | |
1470 | ARC_CC_AL, ARC_CC_EQ = ARC_CC_AL+2, ARC_CC_NE, ARC_CC_P, ARC_CC_N, | |
1471 | ARC_CC_C, ARC_CC_NC, ARC_CC_V, ARC_CC_NV, | |
1472 | ARC_CC_GT, ARC_CC_LE, ARC_CC_GE, ARC_CC_LT, ARC_CC_HI, ARC_CC_LS, ARC_CC_PNZ, | |
1473 | ARC_CC_LO = ARC_CC_C, ARC_CC_HS = ARC_CC_NC | |
1474 | }; | |
1475 | ||
1476 | #define ARC_INVERSE_CONDITION_CODE(X) ((X) ^ 1) | |
1477 | ||
1478 | /* Returns the index of the ARC condition code string in | |
1479 | `arc_condition_codes'. COMPARISON should be an rtx like | |
1480 | `(eq (...) (...))'. */ | |
1481 | ||
1482 | static int | |
1483 | get_arc_condition_code (rtx comparison) | |
1484 | { | |
1485 | switch (GET_MODE (XEXP (comparison, 0))) | |
1486 | { | |
4e10a5a7 RS |
1487 | case E_CCmode: |
1488 | case E_SImode: /* For BRcc. */ | |
526b7aee SV |
1489 | switch (GET_CODE (comparison)) |
1490 | { | |
1491 | case EQ : return ARC_CC_EQ; | |
1492 | case NE : return ARC_CC_NE; | |
1493 | case GT : return ARC_CC_GT; | |
1494 | case LE : return ARC_CC_LE; | |
1495 | case GE : return ARC_CC_GE; | |
1496 | case LT : return ARC_CC_LT; | |
1497 | case GTU : return ARC_CC_HI; | |
1498 | case LEU : return ARC_CC_LS; | |
1499 | case LTU : return ARC_CC_LO; | |
1500 | case GEU : return ARC_CC_HS; | |
1501 | default : gcc_unreachable (); | |
1502 | } | |
4e10a5a7 | 1503 | case E_CC_ZNmode: |
526b7aee SV |
1504 | switch (GET_CODE (comparison)) |
1505 | { | |
1506 | case EQ : return ARC_CC_EQ; | |
1507 | case NE : return ARC_CC_NE; | |
1508 | case GE: return ARC_CC_P; | |
1509 | case LT: return ARC_CC_N; | |
1510 | case GT : return ARC_CC_PNZ; | |
1511 | default : gcc_unreachable (); | |
1512 | } | |
4e10a5a7 | 1513 | case E_CC_Zmode: |
526b7aee SV |
1514 | switch (GET_CODE (comparison)) |
1515 | { | |
1516 | case EQ : return ARC_CC_EQ; | |
1517 | case NE : return ARC_CC_NE; | |
1518 | default : gcc_unreachable (); | |
1519 | } | |
4e10a5a7 | 1520 | case E_CC_Cmode: |
526b7aee SV |
1521 | switch (GET_CODE (comparison)) |
1522 | { | |
1523 | case LTU : return ARC_CC_C; | |
1524 | case GEU : return ARC_CC_NC; | |
1525 | default : gcc_unreachable (); | |
1526 | } | |
4e10a5a7 | 1527 | case E_CC_FP_GTmode: |
526b7aee SV |
1528 | if (TARGET_ARGONAUT_SET && TARGET_SPFP) |
1529 | switch (GET_CODE (comparison)) | |
1530 | { | |
1531 | case GT : return ARC_CC_N; | |
1532 | case UNLE: return ARC_CC_P; | |
1533 | default : gcc_unreachable (); | |
1534 | } | |
1535 | else | |
1536 | switch (GET_CODE (comparison)) | |
1537 | { | |
1538 | case GT : return ARC_CC_HI; | |
1539 | case UNLE : return ARC_CC_LS; | |
1540 | default : gcc_unreachable (); | |
1541 | } | |
4e10a5a7 | 1542 | case E_CC_FP_GEmode: |
526b7aee SV |
1543 | /* Same for FPX and non-FPX. */ |
1544 | switch (GET_CODE (comparison)) | |
1545 | { | |
1546 | case GE : return ARC_CC_HS; | |
1547 | case UNLT : return ARC_CC_LO; | |
1548 | default : gcc_unreachable (); | |
1549 | } | |
4e10a5a7 | 1550 | case E_CC_FP_UNEQmode: |
526b7aee SV |
1551 | switch (GET_CODE (comparison)) |
1552 | { | |
1553 | case UNEQ : return ARC_CC_EQ; | |
1554 | case LTGT : return ARC_CC_NE; | |
1555 | default : gcc_unreachable (); | |
1556 | } | |
4e10a5a7 | 1557 | case E_CC_FP_ORDmode: |
526b7aee SV |
1558 | switch (GET_CODE (comparison)) |
1559 | { | |
1560 | case UNORDERED : return ARC_CC_C; | |
1561 | case ORDERED : return ARC_CC_NC; | |
1562 | default : gcc_unreachable (); | |
1563 | } | |
4e10a5a7 | 1564 | case E_CC_FPXmode: |
526b7aee SV |
1565 | switch (GET_CODE (comparison)) |
1566 | { | |
1567 | case EQ : return ARC_CC_EQ; | |
1568 | case NE : return ARC_CC_NE; | |
1569 | case UNORDERED : return ARC_CC_C; | |
1570 | case ORDERED : return ARC_CC_NC; | |
1571 | case LTGT : return ARC_CC_HI; | |
1572 | case UNEQ : return ARC_CC_LS; | |
1573 | default : gcc_unreachable (); | |
1574 | } | |
4e10a5a7 | 1575 | case E_CC_FPUmode: |
8f3304d0 CZ |
1576 | switch (GET_CODE (comparison)) |
1577 | { | |
1578 | case EQ : return ARC_CC_EQ; | |
1579 | case NE : return ARC_CC_NE; | |
1580 | case GT : return ARC_CC_GT; | |
1581 | case GE : return ARC_CC_GE; | |
1582 | case LT : return ARC_CC_C; | |
1583 | case LE : return ARC_CC_LS; | |
1584 | case UNORDERED : return ARC_CC_V; | |
1585 | case ORDERED : return ARC_CC_NV; | |
1586 | case UNGT : return ARC_CC_HI; | |
1587 | case UNGE : return ARC_CC_HS; | |
1588 | case UNLT : return ARC_CC_LT; | |
1589 | case UNLE : return ARC_CC_LE; | |
1590 | /* UNEQ and LTGT do not have representation. */ | |
1591 | case LTGT : /* Fall through. */ | |
1592 | case UNEQ : /* Fall through. */ | |
1593 | default : gcc_unreachable (); | |
1594 | } | |
4e10a5a7 | 1595 | case E_CC_FPU_UNEQmode: |
8f3304d0 CZ |
1596 | switch (GET_CODE (comparison)) |
1597 | { | |
1598 | case LTGT : return ARC_CC_NE; | |
1599 | case UNEQ : return ARC_CC_EQ; | |
1600 | default : gcc_unreachable (); | |
1601 | } | |
526b7aee SV |
1602 | default : gcc_unreachable (); |
1603 | } | |
1604 | /*NOTREACHED*/ | |
1605 | return (42); | |
1606 | } | |
1607 | ||
1608 | /* Return true if COMPARISON has a short form that can accomodate OFFSET. */ | |
1609 | ||
1610 | bool | |
1611 | arc_short_comparison_p (rtx comparison, int offset) | |
1612 | { | |
1613 | gcc_assert (ARC_CC_NC == ARC_CC_HS); | |
1614 | gcc_assert (ARC_CC_C == ARC_CC_LO); | |
1615 | switch (get_arc_condition_code (comparison)) | |
1616 | { | |
1617 | case ARC_CC_EQ: case ARC_CC_NE: | |
1618 | return offset >= -512 && offset <= 506; | |
1619 | case ARC_CC_GT: case ARC_CC_LE: case ARC_CC_GE: case ARC_CC_LT: | |
1620 | case ARC_CC_HI: case ARC_CC_LS: case ARC_CC_LO: case ARC_CC_HS: | |
1621 | return offset >= -64 && offset <= 58; | |
1622 | default: | |
1623 | return false; | |
1624 | } | |
1625 | } | |
1626 | ||
1627 | /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE, | |
1628 | return the mode to be used for the comparison. */ | |
1629 | ||
ef4bddc2 | 1630 | machine_mode |
526b7aee SV |
1631 | arc_select_cc_mode (enum rtx_code op, rtx x, rtx y) |
1632 | { | |
ef4bddc2 | 1633 | machine_mode mode = GET_MODE (x); |
526b7aee SV |
1634 | rtx x1; |
1635 | ||
1636 | /* For an operation that sets the condition codes as a side-effect, the | |
1637 | C and V flags is not set as for cmp, so we can only use comparisons where | |
1638 | this doesn't matter. (For LT and GE we can use "mi" and "pl" | |
1639 | instead.) */ | |
1640 | /* ??? We could use "pnz" for greater than zero, however, we could then | |
1641 | get into trouble because the comparison could not be reversed. */ | |
1642 | if (GET_MODE_CLASS (mode) == MODE_INT | |
1643 | && y == const0_rtx | |
1644 | && (op == EQ || op == NE | |
486c559b | 1645 | || ((op == LT || op == GE) && GET_MODE_SIZE (GET_MODE (x)) <= 4))) |
526b7aee SV |
1646 | return CC_ZNmode; |
1647 | ||
1648 | /* add.f for if (a+b) */ | |
1649 | if (mode == SImode | |
1650 | && GET_CODE (y) == NEG | |
1651 | && (op == EQ || op == NE)) | |
1652 | return CC_ZNmode; | |
1653 | ||
1654 | /* Check if this is a test suitable for bxor.f . */ | |
1655 | if (mode == SImode && (op == EQ || op == NE) && CONST_INT_P (y) | |
1656 | && ((INTVAL (y) - 1) & INTVAL (y)) == 0 | |
1657 | && INTVAL (y)) | |
1658 | return CC_Zmode; | |
1659 | ||
1660 | /* Check if this is a test suitable for add / bmsk.f . */ | |
1661 | if (mode == SImode && (op == EQ || op == NE) && CONST_INT_P (y) | |
1662 | && GET_CODE (x) == AND && CONST_INT_P ((x1 = XEXP (x, 1))) | |
1663 | && ((INTVAL (x1) + 1) & INTVAL (x1)) == 0 | |
1664 | && (~INTVAL (x1) | INTVAL (y)) < 0 | |
1665 | && (~INTVAL (x1) | INTVAL (y)) > -0x800) | |
1666 | return CC_Zmode; | |
1667 | ||
1668 | if (GET_MODE (x) == SImode && (op == LTU || op == GEU) | |
1669 | && GET_CODE (x) == PLUS | |
1670 | && (rtx_equal_p (XEXP (x, 0), y) || rtx_equal_p (XEXP (x, 1), y))) | |
1671 | return CC_Cmode; | |
1672 | ||
1673 | if (TARGET_ARGONAUT_SET | |
1674 | && ((mode == SFmode && TARGET_SPFP) || (mode == DFmode && TARGET_DPFP))) | |
1675 | switch (op) | |
1676 | { | |
1677 | case EQ: case NE: case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
1678 | return CC_FPXmode; | |
1679 | case LT: case UNGE: case GT: case UNLE: | |
1680 | return CC_FP_GTmode; | |
1681 | case LE: case UNGT: case GE: case UNLT: | |
1682 | return CC_FP_GEmode; | |
1683 | default: gcc_unreachable (); | |
1684 | } | |
8f3304d0 CZ |
1685 | else if (TARGET_HARD_FLOAT |
1686 | && ((mode == SFmode && TARGET_FP_SP_BASE) | |
1687 | || (mode == DFmode && TARGET_FP_DP_BASE))) | |
526b7aee SV |
1688 | switch (op) |
1689 | { | |
8f3304d0 CZ |
1690 | case EQ: |
1691 | case NE: | |
1692 | case UNORDERED: | |
1693 | case ORDERED: | |
1694 | case UNLT: | |
1695 | case UNLE: | |
1696 | case UNGT: | |
1697 | case UNGE: | |
1698 | case LT: | |
1699 | case LE: | |
1700 | case GT: | |
1701 | case GE: | |
1702 | return CC_FPUmode; | |
1703 | ||
1704 | case LTGT: | |
1705 | case UNEQ: | |
1706 | return CC_FPU_UNEQmode; | |
526b7aee | 1707 | |
8f3304d0 CZ |
1708 | default: |
1709 | gcc_unreachable (); | |
1710 | } | |
1711 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT && TARGET_OPTFPE) | |
1712 | { | |
1713 | switch (op) | |
1714 | { | |
1715 | case EQ: case NE: return CC_Zmode; | |
1716 | case LT: case UNGE: | |
1717 | case GT: case UNLE: return CC_FP_GTmode; | |
1718 | case LE: case UNGT: | |
1719 | case GE: case UNLT: return CC_FP_GEmode; | |
1720 | case UNEQ: case LTGT: return CC_FP_UNEQmode; | |
1721 | case ORDERED: case UNORDERED: return CC_FP_ORDmode; | |
1722 | default: gcc_unreachable (); | |
1723 | } | |
1724 | } | |
526b7aee SV |
1725 | return CCmode; |
1726 | } | |
1727 | ||
1728 | /* Vectors to keep interesting information about registers where it can easily | |
1729 | be got. We use to use the actual mode value as the bit number, but there | |
1730 | is (or may be) more than 32 modes now. Instead we use two tables: one | |
1731 | indexed by hard register number, and one indexed by mode. */ | |
1732 | ||
1733 | /* The purpose of arc_mode_class is to shrink the range of modes so that | |
1734 | they all fit (as bit numbers) in a 32-bit word (again). Each real mode is | |
1735 | mapped into one arc_mode_class mode. */ | |
1736 | ||
1737 | enum arc_mode_class { | |
1738 | C_MODE, | |
1739 | S_MODE, D_MODE, T_MODE, O_MODE, | |
1740 | SF_MODE, DF_MODE, TF_MODE, OF_MODE, | |
1741 | V_MODE | |
1742 | }; | |
1743 | ||
1744 | /* Modes for condition codes. */ | |
1745 | #define C_MODES (1 << (int) C_MODE) | |
1746 | ||
1747 | /* Modes for single-word and smaller quantities. */ | |
1748 | #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE)) | |
1749 | ||
1750 | /* Modes for double-word and smaller quantities. */ | |
1751 | #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE)) | |
1752 | ||
1753 | /* Mode for 8-byte DF values only. */ | |
1754 | #define DF_MODES (1 << DF_MODE) | |
1755 | ||
1756 | /* Modes for quad-word and smaller quantities. */ | |
1757 | #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE)) | |
1758 | ||
1759 | /* Modes for 128-bit vectors. */ | |
1760 | #define V_MODES (1 << (int) V_MODE) | |
1761 | ||
1762 | /* Value is 1 if register/mode pair is acceptable on arc. */ | |
1763 | ||
f939c3e6 | 1764 | static unsigned int arc_hard_regno_modes[] = { |
526b7aee SV |
1765 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, |
1766 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, | |
1767 | T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, T_MODES, D_MODES, | |
1768 | D_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1769 | ||
1770 | /* ??? Leave these as S_MODES for now. */ | |
1771 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1772 | DF_MODES, 0, DF_MODES, 0, S_MODES, S_MODES, S_MODES, S_MODES, | |
1773 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
1774 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, C_MODES, S_MODES, | |
1775 | ||
1776 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1777 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1778 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1779 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1780 | ||
1781 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1782 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1783 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1784 | V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, V_MODES, | |
1785 | ||
1786 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, | |
47d8cb23 CZ |
1787 | S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, S_MODES, |
1788 | S_MODES, S_MODES | |
526b7aee SV |
1789 | }; |
1790 | ||
f939c3e6 | 1791 | static unsigned int arc_mode_class [NUM_MACHINE_MODES]; |
526b7aee SV |
1792 | |
1793 | enum reg_class arc_regno_reg_class[FIRST_PSEUDO_REGISTER]; | |
1794 | ||
1795 | enum reg_class | |
1796 | arc_preferred_reload_class (rtx, enum reg_class cl) | |
1797 | { | |
526b7aee SV |
1798 | return cl; |
1799 | } | |
1800 | ||
1801 | /* Initialize the arc_mode_class array. */ | |
1802 | ||
1803 | static void | |
1804 | arc_init_reg_tables (void) | |
1805 | { | |
1806 | int i; | |
1807 | ||
1808 | for (i = 0; i < NUM_MACHINE_MODES; i++) | |
1809 | { | |
ef4bddc2 | 1810 | machine_mode m = (machine_mode) i; |
f8d91e80 NC |
1811 | |
1812 | switch (GET_MODE_CLASS (m)) | |
526b7aee SV |
1813 | { |
1814 | case MODE_INT: | |
1815 | case MODE_PARTIAL_INT: | |
1816 | case MODE_COMPLEX_INT: | |
f8d91e80 | 1817 | if (GET_MODE_SIZE (m) <= 4) |
526b7aee | 1818 | arc_mode_class[i] = 1 << (int) S_MODE; |
f8d91e80 | 1819 | else if (GET_MODE_SIZE (m) == 8) |
526b7aee | 1820 | arc_mode_class[i] = 1 << (int) D_MODE; |
f8d91e80 | 1821 | else if (GET_MODE_SIZE (m) == 16) |
526b7aee | 1822 | arc_mode_class[i] = 1 << (int) T_MODE; |
f8d91e80 | 1823 | else if (GET_MODE_SIZE (m) == 32) |
526b7aee SV |
1824 | arc_mode_class[i] = 1 << (int) O_MODE; |
1825 | else | |
1826 | arc_mode_class[i] = 0; | |
1827 | break; | |
1828 | case MODE_FLOAT: | |
1829 | case MODE_COMPLEX_FLOAT: | |
f8d91e80 | 1830 | if (GET_MODE_SIZE (m) <= 4) |
526b7aee | 1831 | arc_mode_class[i] = 1 << (int) SF_MODE; |
f8d91e80 | 1832 | else if (GET_MODE_SIZE (m) == 8) |
526b7aee | 1833 | arc_mode_class[i] = 1 << (int) DF_MODE; |
f8d91e80 | 1834 | else if (GET_MODE_SIZE (m) == 16) |
526b7aee | 1835 | arc_mode_class[i] = 1 << (int) TF_MODE; |
f8d91e80 | 1836 | else if (GET_MODE_SIZE (m) == 32) |
526b7aee SV |
1837 | arc_mode_class[i] = 1 << (int) OF_MODE; |
1838 | else | |
1839 | arc_mode_class[i] = 0; | |
1840 | break; | |
1841 | case MODE_VECTOR_INT: | |
00c072ae CZ |
1842 | if (GET_MODE_SIZE (m) == 4) |
1843 | arc_mode_class[i] = (1 << (int) S_MODE); | |
1844 | else if (GET_MODE_SIZE (m) == 8) | |
1845 | arc_mode_class[i] = (1 << (int) D_MODE); | |
1846 | else | |
1847 | arc_mode_class[i] = (1 << (int) V_MODE); | |
526b7aee SV |
1848 | break; |
1849 | case MODE_CC: | |
1850 | default: | |
1851 | /* mode_class hasn't been initialized yet for EXTRA_CC_MODES, so | |
1852 | we must explicitly check for them here. */ | |
1853 | if (i == (int) CCmode || i == (int) CC_ZNmode || i == (int) CC_Zmode | |
1854 | || i == (int) CC_Cmode | |
8f3304d0 CZ |
1855 | || i == CC_FP_GTmode || i == CC_FP_GEmode || i == CC_FP_ORDmode |
1856 | || i == CC_FPUmode || i == CC_FPU_UNEQmode) | |
526b7aee SV |
1857 | arc_mode_class[i] = 1 << (int) C_MODE; |
1858 | else | |
1859 | arc_mode_class[i] = 0; | |
1860 | break; | |
1861 | } | |
1862 | } | |
1863 | } | |
1864 | ||
1865 | /* Core registers 56..59 are used for multiply extension options. | |
1866 | The dsp option uses r56 and r57, these are then named acc1 and acc2. | |
1867 | acc1 is the highpart, and acc2 the lowpart, so which register gets which | |
1868 | number depends on endianness. | |
1869 | The mul64 multiplier options use r57 for mlo, r58 for mmid and r59 for mhi. | |
1870 | Because mlo / mhi form a 64 bit value, we use different gcc internal | |
1871 | register numbers to make them form a register pair as the gcc internals | |
1872 | know it. mmid gets number 57, if still available, and mlo / mhi get | |
1873 | number 58 and 59, depending on endianness. We use DBX_REGISTER_NUMBER | |
1874 | to map this back. */ | |
1875 | char rname56[5] = "r56"; | |
1876 | char rname57[5] = "r57"; | |
1877 | char rname58[5] = "r58"; | |
1878 | char rname59[5] = "r59"; | |
f50bb868 CZ |
1879 | char rname29[7] = "ilink1"; |
1880 | char rname30[7] = "ilink2"; | |
526b7aee SV |
1881 | |
1882 | static void | |
1883 | arc_conditional_register_usage (void) | |
1884 | { | |
1885 | int regno; | |
1886 | int i; | |
1887 | int fix_start = 60, fix_end = 55; | |
1888 | ||
f50bb868 CZ |
1889 | if (TARGET_V2) |
1890 | { | |
1891 | /* For ARCv2 the core register set is changed. */ | |
1892 | strcpy (rname29, "ilink"); | |
1893 | strcpy (rname30, "r30"); | |
66825a30 | 1894 | |
73dac59b | 1895 | if (!TEST_HARD_REG_BIT (overrideregs, R30_REG)) |
66825a30 CZ |
1896 | { |
1897 | /* No user interference. Set the r30 to be used by the | |
1898 | compiler. */ | |
73dac59b CZ |
1899 | call_used_regs[R30_REG] = 1; |
1900 | fixed_regs[R30_REG] = 0; | |
66825a30 | 1901 | |
73dac59b | 1902 | arc_regno_reg_class[R30_REG] = GENERAL_REGS; |
66825a30 | 1903 | } |
f50bb868 CZ |
1904 | } |
1905 | ||
526b7aee SV |
1906 | if (TARGET_MUL64_SET) |
1907 | { | |
73dac59b CZ |
1908 | fix_start = R57_REG; |
1909 | fix_end = R59_REG; | |
526b7aee SV |
1910 | |
1911 | /* We don't provide a name for mmed. In rtl / assembly resource lists, | |
1912 | you are supposed to refer to it as mlo & mhi, e.g | |
1913 | (zero_extract:SI (reg:DI 58) (const_int 32) (16)) . | |
1914 | In an actual asm instruction, you are of course use mmed. | |
1915 | The point of avoiding having a separate register for mmed is that | |
1916 | this way, we don't have to carry clobbers of that reg around in every | |
1917 | isntruction that modifies mlo and/or mhi. */ | |
1918 | strcpy (rname57, ""); | |
1919 | strcpy (rname58, TARGET_BIG_ENDIAN ? "mhi" : "mlo"); | |
1920 | strcpy (rname59, TARGET_BIG_ENDIAN ? "mlo" : "mhi"); | |
1921 | } | |
28633bbd CZ |
1922 | |
1923 | /* The nature of arc_tp_regno is actually something more like a global | |
1924 | register, however globalize_reg requires a declaration. | |
1925 | We use EPILOGUE_USES to compensate so that sets from | |
1926 | __builtin_set_frame_pointer are not deleted. */ | |
1927 | if (arc_tp_regno != -1) | |
1928 | fixed_regs[arc_tp_regno] = call_used_regs[arc_tp_regno] = 1; | |
1929 | ||
526b7aee SV |
1930 | if (TARGET_MULMAC_32BY16_SET) |
1931 | { | |
73dac59b CZ |
1932 | fix_start = MUL32x16_REG; |
1933 | fix_end = fix_end > R57_REG ? fix_end : R57_REG; | |
526b7aee SV |
1934 | strcpy (rname56, TARGET_BIG_ENDIAN ? "acc1" : "acc2"); |
1935 | strcpy (rname57, TARGET_BIG_ENDIAN ? "acc2" : "acc1"); | |
1936 | } | |
1937 | for (regno = fix_start; regno <= fix_end; regno++) | |
1938 | { | |
1939 | if (!fixed_regs[regno]) | |
1940 | warning (0, "multiply option implies r%d is fixed", regno); | |
1941 | fixed_regs [regno] = call_used_regs[regno] = 1; | |
1942 | } | |
a2de90a4 | 1943 | |
048c6a9a CZ |
1944 | /* Reduced configuration: don't use r4-r9, r16-r25. */ |
1945 | if (TARGET_RF16) | |
1946 | { | |
73dac59b CZ |
1947 | for (i = R4_REG; i <= R9_REG; i++) |
1948 | fixed_regs[i] = call_used_regs[i] = 1; | |
1949 | for (i = R16_REG; i <= R25_REG; i++) | |
1950 | fixed_regs[i] = call_used_regs[i] = 1; | |
526b7aee SV |
1951 | } |
1952 | ||
8f3304d0 CZ |
1953 | /* ARCHS has 64-bit data-path which makes use of the even-odd paired |
1954 | registers. */ | |
1955 | if (TARGET_HS) | |
73dac59b CZ |
1956 | for (regno = R1_REG; regno < R32_REG; regno +=2) |
1957 | arc_hard_regno_modes[regno] = S_MODES; | |
8f3304d0 | 1958 | |
526b7aee | 1959 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
73dac59b CZ |
1960 | if (i < ILINK1_REG) |
1961 | { | |
1962 | if ((TARGET_Q_CLASS || TARGET_RRQ_CLASS) | |
1963 | && ((i <= R3_REG) || ((i >= R12_REG) && (i <= R15_REG)))) | |
1964 | arc_regno_reg_class[i] = ARCOMPACT16_REGS; | |
1965 | else | |
1966 | arc_regno_reg_class[i] = GENERAL_REGS; | |
1967 | } | |
1968 | else if (i < LP_COUNT) | |
1969 | arc_regno_reg_class[i] = GENERAL_REGS; | |
1970 | else | |
1971 | arc_regno_reg_class[i] = NO_REGS; | |
526b7aee SV |
1972 | |
1973 | /* Handle Special Registers. */ | |
73dac59b | 1974 | arc_regno_reg_class[CC_REG] = NO_REGS; /* CC_REG: must be NO_REGS. */ |
47d8cb23 CZ |
1975 | arc_regno_reg_class[FRAME_POINTER_REGNUM] = GENERAL_REGS; |
1976 | arc_regno_reg_class[ARG_POINTER_REGNUM] = GENERAL_REGS; | |
526b7aee SV |
1977 | |
1978 | if (TARGET_DPFP) | |
73dac59b CZ |
1979 | for (i = R40_REG; i < R44_REG; ++i) |
1980 | { | |
1981 | arc_regno_reg_class[i] = DOUBLE_REGS; | |
1982 | if (!TARGET_ARGONAUT_SET) | |
1983 | CLEAR_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i); | |
1984 | } | |
526b7aee SV |
1985 | else |
1986 | { | |
73dac59b CZ |
1987 | /* Disable all DOUBLE_REGISTER settings, if not generating DPFP |
1988 | code. */ | |
1989 | arc_regno_reg_class[R40_REG] = ALL_REGS; | |
1990 | arc_regno_reg_class[R41_REG] = ALL_REGS; | |
1991 | arc_regno_reg_class[R42_REG] = ALL_REGS; | |
1992 | arc_regno_reg_class[R43_REG] = ALL_REGS; | |
526b7aee | 1993 | |
73dac59b CZ |
1994 | fixed_regs[R40_REG] = 1; |
1995 | fixed_regs[R41_REG] = 1; | |
1996 | fixed_regs[R42_REG] = 1; | |
1997 | fixed_regs[R43_REG] = 1; | |
ad3d6e77 | 1998 | |
73dac59b CZ |
1999 | arc_hard_regno_modes[R40_REG] = 0; |
2000 | arc_hard_regno_modes[R42_REG] = 0; | |
526b7aee SV |
2001 | } |
2002 | ||
2003 | if (TARGET_SIMD_SET) | |
2004 | { | |
2005 | gcc_assert (ARC_FIRST_SIMD_VR_REG == 64); | |
2006 | gcc_assert (ARC_LAST_SIMD_VR_REG == 127); | |
2007 | ||
2008 | for (i = ARC_FIRST_SIMD_VR_REG; i <= ARC_LAST_SIMD_VR_REG; i++) | |
2009 | arc_regno_reg_class [i] = SIMD_VR_REGS; | |
2010 | ||
2011 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_REG == 128); | |
2012 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_IN_REG == 128); | |
2013 | gcc_assert (ARC_FIRST_SIMD_DMA_CONFIG_OUT_REG == 136); | |
2014 | gcc_assert (ARC_LAST_SIMD_DMA_CONFIG_REG == 143); | |
2015 | ||
2016 | for (i = ARC_FIRST_SIMD_DMA_CONFIG_REG; | |
2017 | i <= ARC_LAST_SIMD_DMA_CONFIG_REG; i++) | |
2018 | arc_regno_reg_class [i] = SIMD_DMA_CONFIG_REGS; | |
2019 | } | |
2020 | ||
2021 | /* pc : r63 */ | |
73dac59b | 2022 | arc_regno_reg_class[PCL_REG] = NO_REGS; |
8f3304d0 CZ |
2023 | |
2024 | /*ARCV2 Accumulator. */ | |
79557bae CZ |
2025 | if ((TARGET_V2 |
2026 | && (TARGET_FP_DP_FUSED || TARGET_FP_SP_FUSED)) | |
2027 | || TARGET_PLUS_DMPY) | |
8f3304d0 | 2028 | { |
73dac59b CZ |
2029 | arc_regno_reg_class[ACCL_REGNO] = GENERAL_REGS; |
2030 | arc_regno_reg_class[ACCH_REGNO] = GENERAL_REGS; | |
8b22ef6a | 2031 | |
66825a30 CZ |
2032 | /* Allow the compiler to freely use them. */ |
2033 | if (!TEST_HARD_REG_BIT (overrideregs, ACCL_REGNO)) | |
2034 | fixed_regs[ACCL_REGNO] = 0; | |
2035 | if (!TEST_HARD_REG_BIT (overrideregs, ACCH_REGNO)) | |
2036 | fixed_regs[ACCH_REGNO] = 0; | |
8b22ef6a | 2037 | |
66825a30 CZ |
2038 | if (!fixed_regs[ACCH_REGNO] && !fixed_regs[ACCL_REGNO]) |
2039 | arc_hard_regno_modes[ACC_REG_FIRST] = D_MODES; | |
8f3304d0 | 2040 | } |
526b7aee SV |
2041 | } |
2042 | ||
c43f4279 RS |
2043 | /* Implement TARGET_HARD_REGNO_NREGS. */ |
2044 | ||
2045 | static unsigned int | |
2046 | arc_hard_regno_nregs (unsigned int regno, machine_mode mode) | |
2047 | { | |
2048 | if (GET_MODE_SIZE (mode) == 16 | |
2049 | && regno >= ARC_FIRST_SIMD_VR_REG | |
2050 | && regno <= ARC_LAST_SIMD_VR_REG) | |
2051 | return 1; | |
2052 | ||
2053 | return CEIL (GET_MODE_SIZE (mode), UNITS_PER_WORD); | |
2054 | } | |
2055 | ||
f939c3e6 RS |
2056 | /* Implement TARGET_HARD_REGNO_MODE_OK. */ |
2057 | ||
2058 | static bool | |
2059 | arc_hard_regno_mode_ok (unsigned int regno, machine_mode mode) | |
2060 | { | |
2061 | return (arc_hard_regno_modes[regno] & arc_mode_class[mode]) != 0; | |
2062 | } | |
2063 | ||
99e1629f RS |
2064 | /* Implement TARGET_MODES_TIEABLE_P. Tie QI/HI/SI modes together. */ |
2065 | ||
2066 | static bool | |
2067 | arc_modes_tieable_p (machine_mode mode1, machine_mode mode2) | |
2068 | { | |
2069 | return (GET_MODE_CLASS (mode1) == MODE_INT | |
2070 | && GET_MODE_CLASS (mode2) == MODE_INT | |
2071 | && GET_MODE_SIZE (mode1) <= UNITS_PER_WORD | |
2072 | && GET_MODE_SIZE (mode2) <= UNITS_PER_WORD); | |
2073 | } | |
2074 | ||
526b7aee SV |
2075 | /* Handle an "interrupt" attribute; arguments as in |
2076 | struct attribute_spec.handler. */ | |
2077 | ||
2078 | static tree | |
2079 | arc_handle_interrupt_attribute (tree *, tree name, tree args, int, | |
2080 | bool *no_add_attrs) | |
2081 | { | |
2082 | gcc_assert (args); | |
2083 | ||
2084 | tree value = TREE_VALUE (args); | |
2085 | ||
2086 | if (TREE_CODE (value) != STRING_CST) | |
2087 | { | |
2088 | warning (OPT_Wattributes, | |
2089 | "argument of %qE attribute is not a string constant", | |
2090 | name); | |
2091 | *no_add_attrs = true; | |
2092 | } | |
c7314bc1 CZ |
2093 | else if (!TARGET_V2 |
2094 | && strcmp (TREE_STRING_POINTER (value), "ilink1") | |
2095 | && strcmp (TREE_STRING_POINTER (value), "ilink2")) | |
526b7aee SV |
2096 | { |
2097 | warning (OPT_Wattributes, | |
2098 | "argument of %qE attribute is not \"ilink1\" or \"ilink2\"", | |
2099 | name); | |
2100 | *no_add_attrs = true; | |
2101 | } | |
f50bb868 | 2102 | else if (TARGET_V2 |
c7314bc1 CZ |
2103 | && strcmp (TREE_STRING_POINTER (value), "ilink") |
2104 | && strcmp (TREE_STRING_POINTER (value), "firq")) | |
f50bb868 CZ |
2105 | { |
2106 | warning (OPT_Wattributes, | |
c7314bc1 | 2107 | "argument of %qE attribute is not \"ilink\" or \"firq\"", |
f50bb868 CZ |
2108 | name); |
2109 | *no_add_attrs = true; | |
2110 | } | |
2111 | ||
526b7aee SV |
2112 | return NULL_TREE; |
2113 | } | |
2114 | ||
1825c61e CZ |
2115 | static tree |
2116 | arc_handle_fndecl_attribute (tree *node, tree name, tree args ATTRIBUTE_UNUSED, | |
2117 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
2118 | { | |
2119 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
2120 | { | |
2121 | warning (OPT_Wattributes, "%qE attribute only applies to functions", | |
2122 | name); | |
2123 | *no_add_attrs = true; | |
2124 | } | |
2125 | ||
2126 | return NULL_TREE; | |
2127 | } | |
2128 | ||
ce9dbf20 CZ |
2129 | /* Type of function DECL. |
2130 | ||
2131 | The result is cached. To reset the cache at the end of a function, | |
2132 | call with DECL = NULL_TREE. */ | |
2133 | ||
2134 | static unsigned int | |
2135 | arc_compute_function_type (struct function *fun) | |
2136 | { | |
2137 | tree attr, decl = fun->decl; | |
2138 | unsigned int fn_type = fun->machine->fn_type; | |
2139 | ||
2140 | if (fn_type != ARC_FUNCTION_UNKNOWN) | |
2141 | return fn_type; | |
2142 | ||
2143 | /* Check if it is a naked function. */ | |
2144 | if (lookup_attribute ("naked", DECL_ATTRIBUTES (decl)) != NULL_TREE) | |
2145 | fn_type |= ARC_FUNCTION_NAKED; | |
2146 | else | |
2147 | fn_type |= ARC_FUNCTION_NORMAL; | |
2148 | ||
2149 | /* Now see if this is an interrupt handler. */ | |
2150 | attr = lookup_attribute ("interrupt", DECL_ATTRIBUTES (decl)); | |
2151 | if (attr != NULL_TREE) | |
2152 | { | |
2153 | tree value, args = TREE_VALUE (attr); | |
2154 | ||
2155 | gcc_assert (list_length (args) == 1); | |
2156 | value = TREE_VALUE (args); | |
2157 | gcc_assert (TREE_CODE (value) == STRING_CST); | |
2158 | ||
2159 | if (!strcmp (TREE_STRING_POINTER (value), "ilink1") | |
2160 | || !strcmp (TREE_STRING_POINTER (value), "ilink")) | |
2161 | fn_type |= ARC_FUNCTION_ILINK1; | |
2162 | else if (!strcmp (TREE_STRING_POINTER (value), "ilink2")) | |
2163 | fn_type |= ARC_FUNCTION_ILINK2; | |
2164 | else if (!strcmp (TREE_STRING_POINTER (value), "firq")) | |
2165 | fn_type |= ARC_FUNCTION_FIRQ; | |
2166 | else | |
2167 | gcc_unreachable (); | |
2168 | } | |
2169 | ||
2170 | return fun->machine->fn_type = fn_type; | |
2171 | } | |
2172 | ||
1825c61e CZ |
2173 | /* Implement `TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS' */ |
2174 | ||
2175 | static bool | |
2176 | arc_allocate_stack_slots_for_args (void) | |
2177 | { | |
2178 | /* Naked functions should not allocate stack slots for arguments. */ | |
2179 | unsigned int fn_type = arc_compute_function_type (cfun); | |
2180 | ||
2181 | return !ARC_NAKED_P(fn_type); | |
2182 | } | |
2183 | ||
2184 | /* Implement `TARGET_WARN_FUNC_RETURN'. */ | |
2185 | ||
2186 | static bool | |
2187 | arc_warn_func_return (tree decl) | |
2188 | { | |
2189 | struct function *func = DECL_STRUCT_FUNCTION (decl); | |
2190 | unsigned int fn_type = arc_compute_function_type (func); | |
2191 | ||
2192 | return !ARC_NAKED_P (fn_type); | |
2193 | } | |
2194 | ||
526b7aee SV |
2195 | /* Return zero if TYPE1 and TYPE are incompatible, one if they are compatible, |
2196 | and two if they are nearly compatible (which causes a warning to be | |
2197 | generated). */ | |
2198 | ||
2199 | static int | |
2200 | arc_comp_type_attributes (const_tree type1, | |
2201 | const_tree type2) | |
2202 | { | |
2203 | int l1, l2, m1, m2, s1, s2; | |
2204 | ||
2205 | /* Check for mismatch of non-default calling convention. */ | |
2206 | if (TREE_CODE (type1) != FUNCTION_TYPE) | |
2207 | return 1; | |
2208 | ||
2209 | /* Check for mismatched call attributes. */ | |
2210 | l1 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2211 | l2 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2212 | m1 = lookup_attribute ("medium_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2213 | m2 = lookup_attribute ("medium_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2214 | s1 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
2215 | s2 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
2216 | ||
2217 | /* Only bother to check if an attribute is defined. */ | |
2218 | if (l1 | l2 | m1 | m2 | s1 | s2) | |
2219 | { | |
2220 | /* If one type has an attribute, the other must have the same attribute. */ | |
2221 | if ((l1 != l2) || (m1 != m2) || (s1 != s2)) | |
2222 | return 0; | |
2223 | ||
2224 | /* Disallow mixed attributes. */ | |
2225 | if (l1 + m1 + s1 > 1) | |
2226 | return 0; | |
2227 | } | |
2228 | ||
2229 | ||
2230 | return 1; | |
2231 | } | |
2232 | ||
526b7aee SV |
2233 | /* Misc. utilities. */ |
2234 | ||
2235 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
2236 | return the rtx for the cc reg in the proper mode. */ | |
2237 | ||
2238 | rtx | |
ef4bddc2 | 2239 | gen_compare_reg (rtx comparison, machine_mode omode) |
526b7aee SV |
2240 | { |
2241 | enum rtx_code code = GET_CODE (comparison); | |
2242 | rtx x = XEXP (comparison, 0); | |
2243 | rtx y = XEXP (comparison, 1); | |
2244 | rtx tmp, cc_reg; | |
ef4bddc2 | 2245 | machine_mode mode, cmode; |
526b7aee SV |
2246 | |
2247 | ||
2248 | cmode = GET_MODE (x); | |
2249 | if (cmode == VOIDmode) | |
2250 | cmode = GET_MODE (y); | |
2251 | gcc_assert (cmode == SImode || cmode == SFmode || cmode == DFmode); | |
2252 | if (cmode == SImode) | |
2253 | { | |
2254 | if (!register_operand (x, SImode)) | |
2255 | { | |
2256 | if (register_operand (y, SImode)) | |
2257 | { | |
2258 | tmp = x; | |
2259 | x = y; | |
2260 | y = tmp; | |
2261 | code = swap_condition (code); | |
2262 | } | |
2263 | else | |
2264 | x = copy_to_mode_reg (SImode, x); | |
2265 | } | |
2266 | if (GET_CODE (y) == SYMBOL_REF && flag_pic) | |
2267 | y = copy_to_mode_reg (SImode, y); | |
2268 | } | |
2269 | else | |
2270 | { | |
2271 | x = force_reg (cmode, x); | |
2272 | y = force_reg (cmode, y); | |
2273 | } | |
2274 | mode = SELECT_CC_MODE (code, x, y); | |
2275 | ||
2276 | cc_reg = gen_rtx_REG (mode, CC_REG); | |
2277 | ||
2278 | /* ??? FIXME (x-y)==0, as done by both cmpsfpx_raw and | |
2279 | cmpdfpx_raw, is not a correct comparison for floats: | |
2280 | http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm | |
2281 | */ | |
2282 | if (TARGET_ARGONAUT_SET | |
2283 | && ((cmode == SFmode && TARGET_SPFP) || (cmode == DFmode && TARGET_DPFP))) | |
2284 | { | |
2285 | switch (code) | |
2286 | { | |
2287 | case NE: case EQ: case LT: case UNGE: case LE: case UNGT: | |
2288 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2289 | break; | |
2290 | case GT: case UNLE: case GE: case UNLT: | |
2291 | code = swap_condition (code); | |
2292 | tmp = x; | |
2293 | x = y; | |
2294 | y = tmp; | |
2295 | break; | |
2296 | default: | |
2297 | gcc_unreachable (); | |
2298 | } | |
2299 | if (cmode == SFmode) | |
2300 | { | |
2301 | emit_insn (gen_cmpsfpx_raw (x, y)); | |
2302 | } | |
2303 | else /* DFmode */ | |
2304 | { | |
2305 | /* Accepts Dx regs directly by insns. */ | |
2306 | emit_insn (gen_cmpdfpx_raw (x, y)); | |
2307 | } | |
2308 | ||
2309 | if (mode != CC_FPXmode) | |
f7df4a84 | 2310 | emit_insn (gen_rtx_SET (cc_reg, |
526b7aee SV |
2311 | gen_rtx_COMPARE (mode, |
2312 | gen_rtx_REG (CC_FPXmode, 61), | |
2313 | const0_rtx))); | |
2314 | } | |
c4014855 CZ |
2315 | else if (TARGET_FPX_QUARK && (cmode == SFmode)) |
2316 | { | |
2317 | switch (code) | |
2318 | { | |
2319 | case NE: case EQ: case GT: case UNLE: case GE: case UNLT: | |
2320 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2321 | break; | |
2322 | case LT: case UNGE: case LE: case UNGT: | |
2323 | code = swap_condition (code); | |
2324 | tmp = x; | |
2325 | x = y; | |
2326 | y = tmp; | |
2327 | break; | |
2328 | default: | |
2329 | gcc_unreachable (); | |
2330 | } | |
2331 | ||
2332 | emit_insn (gen_cmp_quark (cc_reg, | |
2333 | gen_rtx_COMPARE (mode, x, y))); | |
2334 | } | |
8f3304d0 CZ |
2335 | else if (TARGET_HARD_FLOAT |
2336 | && ((cmode == SFmode && TARGET_FP_SP_BASE) | |
2337 | || (cmode == DFmode && TARGET_FP_DP_BASE))) | |
2338 | emit_insn (gen_rtx_SET (cc_reg, gen_rtx_COMPARE (mode, x, y))); | |
526b7aee SV |
2339 | else if (GET_MODE_CLASS (cmode) == MODE_FLOAT && TARGET_OPTFPE) |
2340 | { | |
2341 | rtx op0 = gen_rtx_REG (cmode, 0); | |
2342 | rtx op1 = gen_rtx_REG (cmode, GET_MODE_SIZE (cmode) / UNITS_PER_WORD); | |
b1a82751 | 2343 | bool swap = false; |
526b7aee SV |
2344 | |
2345 | switch (code) | |
2346 | { | |
2347 | case NE: case EQ: case GT: case UNLE: case GE: case UNLT: | |
2348 | case UNEQ: case LTGT: case ORDERED: case UNORDERED: | |
2349 | break; | |
2350 | case LT: case UNGE: case LE: case UNGT: | |
2351 | code = swap_condition (code); | |
b1a82751 | 2352 | swap = true; |
526b7aee SV |
2353 | break; |
2354 | default: | |
2355 | gcc_unreachable (); | |
2356 | } | |
2357 | if (currently_expanding_to_rtl) | |
2358 | { | |
b1a82751 CZ |
2359 | if (swap) |
2360 | { | |
2361 | tmp = x; | |
2362 | x = y; | |
2363 | y = tmp; | |
2364 | } | |
526b7aee SV |
2365 | emit_move_insn (op0, x); |
2366 | emit_move_insn (op1, y); | |
2367 | } | |
2368 | else | |
2369 | { | |
2370 | gcc_assert (rtx_equal_p (op0, x)); | |
2371 | gcc_assert (rtx_equal_p (op1, y)); | |
b1a82751 CZ |
2372 | if (swap) |
2373 | { | |
2374 | op0 = y; | |
2375 | op1 = x; | |
2376 | } | |
526b7aee SV |
2377 | } |
2378 | emit_insn (gen_cmp_float (cc_reg, gen_rtx_COMPARE (mode, op0, op1))); | |
2379 | } | |
2380 | else | |
f7df4a84 | 2381 | emit_insn (gen_rtx_SET (cc_reg, gen_rtx_COMPARE (mode, x, y))); |
526b7aee SV |
2382 | return gen_rtx_fmt_ee (code, omode, cc_reg, const0_rtx); |
2383 | } | |
2384 | ||
2385 | /* Return true if VALUE, a const_double, will fit in a limm (4 byte number). | |
2386 | We assume the value can be either signed or unsigned. */ | |
2387 | ||
2388 | bool | |
2389 | arc_double_limm_p (rtx value) | |
2390 | { | |
2391 | HOST_WIDE_INT low, high; | |
2392 | ||
2393 | gcc_assert (GET_CODE (value) == CONST_DOUBLE); | |
2394 | ||
2395 | if (TARGET_DPFP) | |
2396 | return true; | |
2397 | ||
2398 | low = CONST_DOUBLE_LOW (value); | |
2399 | high = CONST_DOUBLE_HIGH (value); | |
2400 | ||
2401 | if (low & 0x80000000) | |
2402 | { | |
2403 | return (((unsigned HOST_WIDE_INT) low <= 0xffffffff && high == 0) | |
2404 | || (((low & - (unsigned HOST_WIDE_INT) 0x80000000) | |
2405 | == - (unsigned HOST_WIDE_INT) 0x80000000) | |
2406 | && high == -1)); | |
2407 | } | |
2408 | else | |
2409 | { | |
2410 | return (unsigned HOST_WIDE_INT) low <= 0x7fffffff && high == 0; | |
2411 | } | |
2412 | } | |
2413 | ||
2414 | /* Do any needed setup for a variadic function. For the ARC, we must | |
2415 | create a register parameter block, and then copy any anonymous arguments | |
2416 | in registers to memory. | |
2417 | ||
2418 | CUM has not been updated for the last named argument which has type TYPE | |
2419 | and mode MODE, and we rely on this fact. */ | |
2420 | ||
2421 | static void | |
2422 | arc_setup_incoming_varargs (cumulative_args_t args_so_far, | |
ef4bddc2 | 2423 | machine_mode mode, tree type, |
526b7aee SV |
2424 | int *pretend_size, int no_rtl) |
2425 | { | |
2426 | int first_anon_arg; | |
2427 | CUMULATIVE_ARGS next_cum; | |
2428 | ||
2429 | /* We must treat `__builtin_va_alist' as an anonymous arg. */ | |
2430 | ||
2431 | next_cum = *get_cumulative_args (args_so_far); | |
8f3304d0 CZ |
2432 | arc_function_arg_advance (pack_cumulative_args (&next_cum), |
2433 | mode, type, true); | |
526b7aee SV |
2434 | first_anon_arg = next_cum; |
2435 | ||
8f3304d0 | 2436 | if (FUNCTION_ARG_REGNO_P (first_anon_arg)) |
526b7aee SV |
2437 | { |
2438 | /* First anonymous (unnamed) argument is in a reg. */ | |
2439 | ||
2440 | /* Note that first_reg_offset < MAX_ARC_PARM_REGS. */ | |
2441 | int first_reg_offset = first_anon_arg; | |
2442 | ||
2443 | if (!no_rtl) | |
2444 | { | |
2445 | rtx regblock | |
2446 | = gen_rtx_MEM (BLKmode, plus_constant (Pmode, arg_pointer_rtx, | |
2447 | FIRST_PARM_OFFSET (0))); | |
2448 | move_block_from_reg (first_reg_offset, regblock, | |
2449 | MAX_ARC_PARM_REGS - first_reg_offset); | |
2450 | } | |
2451 | ||
2452 | *pretend_size | |
2453 | = ((MAX_ARC_PARM_REGS - first_reg_offset ) * UNITS_PER_WORD); | |
2454 | } | |
2455 | } | |
2456 | ||
2457 | /* Cost functions. */ | |
2458 | ||
2459 | /* Provide the costs of an addressing mode that contains ADDR. | |
2460 | If ADDR is not a valid address, its cost is irrelevant. */ | |
2461 | ||
b51addd6 | 2462 | static int |
ef4bddc2 | 2463 | arc_address_cost (rtx addr, machine_mode, addr_space_t, bool speed) |
526b7aee SV |
2464 | { |
2465 | switch (GET_CODE (addr)) | |
2466 | { | |
2467 | case REG : | |
2468 | return speed || satisfies_constraint_Rcq (addr) ? 0 : 1; | |
2469 | case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: | |
2470 | case PRE_MODIFY: case POST_MODIFY: | |
2471 | return !speed; | |
2472 | ||
2473 | case LABEL_REF : | |
2474 | case SYMBOL_REF : | |
2475 | case CONST : | |
4d03dc2f JR |
2476 | if (TARGET_NPS_CMEM && cmem_address (addr, SImode)) |
2477 | return 0; | |
526b7aee SV |
2478 | /* Most likely needs a LIMM. */ |
2479 | return COSTS_N_INSNS (1); | |
2480 | ||
2481 | case PLUS : | |
2482 | { | |
2483 | register rtx plus0 = XEXP (addr, 0); | |
2484 | register rtx plus1 = XEXP (addr, 1); | |
2485 | ||
2486 | if (GET_CODE (plus0) != REG | |
2487 | && (GET_CODE (plus0) != MULT | |
2488 | || !CONST_INT_P (XEXP (plus0, 1)) | |
2489 | || (INTVAL (XEXP (plus0, 1)) != 2 | |
2490 | && INTVAL (XEXP (plus0, 1)) != 4))) | |
2491 | break; | |
2492 | ||
2493 | switch (GET_CODE (plus1)) | |
2494 | { | |
2495 | case CONST_INT : | |
2496 | return (!RTX_OK_FOR_OFFSET_P (SImode, plus1) | |
2497 | ? COSTS_N_INSNS (1) | |
2498 | : speed | |
2499 | ? 0 | |
2500 | : (satisfies_constraint_Rcq (plus0) | |
2501 | && satisfies_constraint_O (plus1)) | |
2502 | ? 0 | |
2503 | : 1); | |
2504 | case REG: | |
2505 | return (speed < 1 ? 0 | |
2506 | : (satisfies_constraint_Rcq (plus0) | |
2507 | && satisfies_constraint_Rcq (plus1)) | |
2508 | ? 0 : 1); | |
2509 | case CONST : | |
2510 | case SYMBOL_REF : | |
2511 | case LABEL_REF : | |
2512 | return COSTS_N_INSNS (1); | |
2513 | default: | |
2514 | break; | |
2515 | } | |
2516 | break; | |
2517 | } | |
2518 | default: | |
2519 | break; | |
2520 | } | |
2521 | ||
2522 | return 4; | |
2523 | } | |
2524 | ||
2525 | /* Emit instruction X with the frame related bit set. */ | |
2526 | ||
2527 | static rtx | |
2528 | frame_insn (rtx x) | |
2529 | { | |
2530 | x = emit_insn (x); | |
2531 | RTX_FRAME_RELATED_P (x) = 1; | |
2532 | return x; | |
2533 | } | |
2534 | ||
2535 | /* Emit a frame insn to move SRC to DST. */ | |
2536 | ||
2537 | static rtx | |
2538 | frame_move (rtx dst, rtx src) | |
2539 | { | |
67a96300 CZ |
2540 | rtx tmp = gen_rtx_SET (dst, src); |
2541 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2542 | return frame_insn (tmp); | |
526b7aee SV |
2543 | } |
2544 | ||
2545 | /* Like frame_move, but add a REG_INC note for REG if ADDR contains an | |
2546 | auto increment address, or is zero. */ | |
2547 | ||
2548 | static rtx | |
2549 | frame_move_inc (rtx dst, rtx src, rtx reg, rtx addr) | |
2550 | { | |
2551 | rtx insn = frame_move (dst, src); | |
2552 | ||
2553 | if (!addr | |
2554 | || GET_CODE (addr) == PRE_DEC || GET_CODE (addr) == POST_INC | |
2555 | || GET_CODE (addr) == PRE_MODIFY || GET_CODE (addr) == POST_MODIFY) | |
2556 | add_reg_note (insn, REG_INC, reg); | |
2557 | return insn; | |
2558 | } | |
2559 | ||
2560 | /* Emit a frame insn which adjusts a frame address register REG by OFFSET. */ | |
2561 | ||
2562 | static rtx | |
2563 | frame_add (rtx reg, HOST_WIDE_INT offset) | |
2564 | { | |
2565 | gcc_assert ((offset & 0x3) == 0); | |
2566 | if (!offset) | |
2567 | return NULL_RTX; | |
2568 | return frame_move (reg, plus_constant (Pmode, reg, offset)); | |
2569 | } | |
2570 | ||
2571 | /* Emit a frame insn which adjusts stack pointer by OFFSET. */ | |
2572 | ||
2573 | static rtx | |
2574 | frame_stack_add (HOST_WIDE_INT offset) | |
2575 | { | |
2576 | return frame_add (stack_pointer_rtx, offset); | |
2577 | } | |
2578 | ||
47d8cb23 CZ |
2579 | /* Helper function to wrap FRAME_POINTER_NEEDED. We do this as |
2580 | FRAME_POINTER_NEEDED will not be true until the IRA (Integrated | |
2581 | Register Allocator) pass, while we want to get the frame size | |
2582 | correct earlier than the IRA pass. | |
2583 | ||
2584 | When a function uses eh_return we must ensure that the fp register | |
2585 | is saved and then restored so that the unwinder can restore the | |
2586 | correct value for the frame we are going to jump to. | |
2587 | ||
2588 | To do this we force all frames that call eh_return to require a | |
2589 | frame pointer (see arc_frame_pointer_required), this | |
2590 | will ensure that the previous frame pointer is stored on entry to | |
2591 | the function, and will then be reloaded at function exit. | |
2592 | ||
2593 | As the frame pointer is handled as a special case in our prologue | |
2594 | and epilogue code it must not be saved and restored using the | |
2595 | MUST_SAVE_REGISTER mechanism otherwise we run into issues where GCC | |
2596 | believes that the function is not using a frame pointer and that | |
2597 | the value in the fp register is the frame pointer, while the | |
2598 | prologue and epilogue are busy saving and restoring the fp | |
2599 | register. | |
2600 | ||
2601 | During compilation of a function the frame size is evaluated | |
2602 | multiple times, it is not until the reload pass is complete the the | |
2603 | frame size is considered fixed (it is at this point that space for | |
2604 | all spills has been allocated). However the frame_pointer_needed | |
2605 | variable is not set true until the register allocation pass, as a | |
2606 | result in the early stages the frame size does not include space | |
2607 | for the frame pointer to be spilled. | |
2608 | ||
2609 | The problem that this causes is that the rtl generated for | |
2610 | EH_RETURN_HANDLER_RTX uses the details of the frame size to compute | |
2611 | the offset from the frame pointer at which the return address | |
2612 | lives. However, in early passes GCC has not yet realised we need a | |
2613 | frame pointer, and so has not included space for the frame pointer | |
2614 | in the frame size, and so gets the offset of the return address | |
2615 | wrong. This should not be an issue as in later passes GCC has | |
2616 | realised that the frame pointer needs to be spilled, and has | |
2617 | increased the frame size. However, the rtl for the | |
2618 | EH_RETURN_HANDLER_RTX is not regenerated to use the newer, larger | |
2619 | offset, and the wrong smaller offset is used. */ | |
2620 | ||
2621 | static bool | |
2622 | arc_frame_pointer_needed (void) | |
2623 | { | |
2624 | return (frame_pointer_needed || crtl->calls_eh_return); | |
2625 | } | |
526b7aee | 2626 | |
1ec86e1e | 2627 | /* Tell prologue and epilogue if register REGNO should be saved / |
ce9dbf20 CZ |
2628 | restored. The SPECIAL_P is true when the register may need special |
2629 | ld/st sequence. The return address, and stack pointer are treated | |
2630 | separately. Don't consider them here. */ | |
526b7aee | 2631 | |
41453183 | 2632 | static bool |
ce9dbf20 | 2633 | arc_must_save_register (int regno, struct function *func, bool special_p) |
41453183 | 2634 | { |
1825c61e | 2635 | unsigned int fn_type = arc_compute_function_type (func); |
41453183 | 2636 | bool irq_auto_save_p = ((irq_ctrl_saved.irq_save_last_reg >= regno) |
c7314bc1 CZ |
2637 | && ARC_AUTO_IRQ_P (fn_type)); |
2638 | bool firq_auto_save_p = ARC_FAST_INTERRUPT_P (fn_type); | |
2639 | ||
2640 | switch (rgf_banked_register_count) | |
2641 | { | |
2642 | case 4: | |
2643 | firq_auto_save_p &= (regno < 4); | |
2644 | break; | |
2645 | case 8: | |
2646 | firq_auto_save_p &= ((regno < 4) || ((regno > 11) && (regno < 16))); | |
2647 | break; | |
2648 | case 16: | |
2649 | firq_auto_save_p &= ((regno < 4) || ((regno > 9) && (regno < 16)) | |
2650 | || ((regno > 25) && (regno < 29)) | |
2651 | || ((regno > 29) && (regno < 32))); | |
2652 | break; | |
2653 | case 32: | |
2654 | firq_auto_save_p &= (regno != 29) && (regno < 32); | |
2655 | break; | |
2656 | default: | |
2657 | firq_auto_save_p = false; | |
2658 | break; | |
2659 | } | |
41453183 | 2660 | |
47d8cb23 CZ |
2661 | switch (regno) |
2662 | { | |
ce9dbf20 | 2663 | case ILINK1_REG: |
47d8cb23 CZ |
2664 | case RETURN_ADDR_REGNUM: |
2665 | case STACK_POINTER_REGNUM: | |
ce9dbf20 CZ |
2666 | /* The stack pointer and the return address are handled |
2667 | separately. */ | |
2668 | return false; | |
2669 | ||
2670 | case R30_REG: | |
2671 | /* r30 is either used as ilink2 by ARCv1 or as a free register | |
2672 | by ARCv2. */ | |
2673 | if (!TARGET_V2) | |
2674 | return false; | |
2675 | break; | |
2676 | ||
2677 | case R40_REG: | |
2678 | case R41_REG: | |
2679 | case R42_REG: | |
2680 | case R43_REG: | |
2681 | case R44_REG: | |
2682 | /* If those ones are used by the FPX machinery, we handle them | |
2683 | separately. */ | |
2684 | if (TARGET_DPFP && !special_p) | |
2685 | return false; | |
2686 | /* FALLTHRU. */ | |
2687 | ||
2688 | case R32_REG: | |
2689 | case R33_REG: | |
2690 | case R34_REG: | |
2691 | case R35_REG: | |
2692 | case R36_REG: | |
2693 | case R37_REG: | |
2694 | case R38_REG: | |
2695 | case R39_REG: | |
2696 | case R45_REG: | |
2697 | case R46_REG: | |
2698 | case R47_REG: | |
2699 | case R48_REG: | |
2700 | case R49_REG: | |
2701 | case R50_REG: | |
2702 | case R51_REG: | |
2703 | case R52_REG: | |
2704 | case R53_REG: | |
2705 | case R54_REG: | |
2706 | case R55_REG: | |
2707 | case R56_REG: | |
2708 | case R57_REG: | |
2709 | case R58_REG: | |
2710 | case R59_REG: | |
2711 | /* The Extension Registers. */ | |
2712 | if (ARC_INTERRUPT_P (fn_type) | |
2713 | && (df_regs_ever_live_p (RETURN_ADDR_REGNUM) | |
2714 | || df_regs_ever_live_p (regno)) | |
2715 | /* Not all extension registers are available, choose the | |
2716 | real ones. */ | |
2717 | && !fixed_regs[regno]) | |
2718 | return true; | |
2719 | return false; | |
2720 | ||
2721 | case 61: | |
2722 | case 62: | |
2723 | case 63: | |
2724 | /* Fixed/control register, nothing to do. LP_COUNT is | |
2725 | different. */ | |
47d8cb23 CZ |
2726 | return false; |
2727 | ||
2728 | case HARD_FRAME_POINTER_REGNUM: | |
2729 | /* If we need FP reg as a frame pointer then don't save it as a | |
2730 | regular reg. */ | |
2731 | if (arc_frame_pointer_needed ()) | |
2732 | return false; | |
ce9dbf20 | 2733 | break; |
47d8cb23 | 2734 | |
47d8cb23 | 2735 | default: |
ce9dbf20 | 2736 | break; |
47d8cb23 | 2737 | } |
41453183 | 2738 | |
ce9dbf20 CZ |
2739 | if (((df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
2740 | /* In an interrupt save everything. */ | |
2741 | || (ARC_INTERRUPT_P (fn_type) | |
2742 | && (df_regs_ever_live_p (RETURN_ADDR_REGNUM) | |
2743 | || df_regs_ever_live_p (regno)))) | |
2744 | /* Do not emit code for auto saved regs. */ | |
2745 | && !irq_auto_save_p | |
2746 | && !firq_auto_save_p) | |
2747 | return true; | |
41453183 CZ |
2748 | return false; |
2749 | } | |
2750 | ||
2751 | /* Return true if the return address must be saved in the current function, | |
2752 | otherwise return false. */ | |
2753 | ||
2754 | static bool | |
2755 | arc_must_save_return_addr (struct function *func) | |
2756 | { | |
2757 | if (func->machine->frame_info.save_return_addr) | |
2758 | return true; | |
2759 | ||
2760 | return false; | |
2761 | } | |
2762 | ||
526b7aee SV |
2763 | /* Return non-zero if there are registers to be saved or loaded using |
2764 | millicode thunks. We can only use consecutive sequences starting | |
2765 | with r13, and not going beyond r25. | |
2766 | GMASK is a bitmask of registers to save. This function sets | |
2767 | FRAME->millicod_start_reg .. FRAME->millicode_end_reg to the range | |
2768 | of registers to be saved / restored with a millicode call. */ | |
2769 | ||
2770 | static int | |
ce9dbf20 | 2771 | arc_compute_millicode_save_restore_regs (uint64_t gmask, |
526b7aee SV |
2772 | struct arc_frame_info *frame) |
2773 | { | |
2774 | int regno; | |
2775 | ||
2776 | int start_reg = 13, end_reg = 25; | |
2777 | ||
ce9dbf20 | 2778 | for (regno = start_reg; regno <= end_reg && (gmask & (1ULL << regno));) |
526b7aee SV |
2779 | regno++; |
2780 | end_reg = regno - 1; | |
2781 | /* There is no point in using millicode thunks if we don't save/restore | |
2782 | at least three registers. For non-leaf functions we also have the | |
2783 | blink restore. */ | |
2784 | if (regno - start_reg >= 3 - (crtl->is_leaf == 0)) | |
2785 | { | |
2786 | frame->millicode_start_reg = 13; | |
2787 | frame->millicode_end_reg = regno - 1; | |
2788 | return 1; | |
2789 | } | |
2790 | return 0; | |
2791 | } | |
2792 | ||
6fe5e235 CZ |
2793 | /* Return the bytes needed to compute the frame pointer from the |
2794 | current stack pointer. */ | |
526b7aee | 2795 | |
6fe5e235 CZ |
2796 | static unsigned int |
2797 | arc_compute_frame_size (void) | |
526b7aee SV |
2798 | { |
2799 | int regno; | |
2800 | unsigned int total_size, var_size, args_size, pretend_size, extra_size; | |
90b48013 | 2801 | unsigned int reg_size; |
ce9dbf20 | 2802 | uint64_t gmask; |
6fe5e235 CZ |
2803 | struct arc_frame_info *frame_info; |
2804 | int size; | |
90b48013 CZ |
2805 | unsigned int extra_plus_reg_size; |
2806 | unsigned int extra_plus_reg_size_aligned; | |
6fe5e235 CZ |
2807 | |
2808 | /* The answer might already be known. */ | |
2809 | if (cfun->machine->frame_info.initialized) | |
2810 | return cfun->machine->frame_info.total_size; | |
526b7aee | 2811 | |
6fe5e235 CZ |
2812 | frame_info = &cfun->machine->frame_info; |
2813 | size = ARC_STACK_ALIGN (get_frame_size ()); | |
526b7aee | 2814 | |
6fe5e235 | 2815 | /* 1) Size of locals and temporaries. */ |
526b7aee SV |
2816 | var_size = size; |
2817 | ||
6fe5e235 | 2818 | /* 2) Size of outgoing arguments. */ |
526b7aee SV |
2819 | args_size = crtl->outgoing_args_size; |
2820 | ||
2821 | /* 3) Calculate space needed for saved registers. | |
2822 | ??? We ignore the extension registers for now. */ | |
2823 | ||
2824 | /* See if this is an interrupt handler. Call used registers must be saved | |
2825 | for them too. */ | |
2826 | ||
2827 | reg_size = 0; | |
2828 | gmask = 0; | |
526b7aee | 2829 | |
ce9dbf20 CZ |
2830 | /* The last 4 regs are special, avoid them. */ |
2831 | for (regno = 0; regno <= (GMASK_LEN - 4); regno++) | |
526b7aee | 2832 | { |
ce9dbf20 | 2833 | if (arc_must_save_register (regno, cfun, false)) |
526b7aee SV |
2834 | { |
2835 | reg_size += UNITS_PER_WORD; | |
ce9dbf20 | 2836 | gmask |= 1ULL << regno; |
526b7aee SV |
2837 | } |
2838 | } | |
2839 | ||
6fe5e235 CZ |
2840 | /* In a frame that calls __builtin_eh_return two data registers are |
2841 | used to pass values back to the exception handler. | |
2842 | ||
2843 | Ensure that these registers are spilled to the stack so that the | |
2844 | exception throw code can find them, and update the saved values. | |
2845 | The handling code will then consume these reloaded values to | |
2846 | handle the exception. */ | |
2847 | if (crtl->calls_eh_return) | |
2848 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
2849 | { | |
2850 | reg_size += UNITS_PER_WORD; | |
ce9dbf20 | 2851 | gmask |= 1ULL << regno; |
6fe5e235 CZ |
2852 | } |
2853 | ||
90b48013 CZ |
2854 | /* Check if we need to save the return address. */ |
2855 | frame_info->save_return_addr = (!crtl->is_leaf | |
2856 | || df_regs_ever_live_p (RETURN_ADDR_REGNUM) | |
2857 | || crtl->calls_eh_return); | |
2858 | ||
2859 | /* Saving blink reg for millicode thunk calls. */ | |
2860 | if (TARGET_MILLICODE_THUNK_SET | |
ce9dbf20 CZ |
2861 | && !crtl->calls_eh_return |
2862 | && !ARC_INTERRUPT_P (arc_compute_function_type (cfun))) | |
526b7aee SV |
2863 | { |
2864 | if (arc_compute_millicode_save_restore_regs (gmask, frame_info)) | |
2865 | frame_info->save_return_addr = true; | |
2866 | } | |
2867 | ||
ce9dbf20 CZ |
2868 | /* Save lp_count, lp_start and lp_end. */ |
2869 | if (arc_lpcwidth != 0 && arc_must_save_register (LP_COUNT, cfun, true)) | |
2870 | reg_size += UNITS_PER_WORD * 3; | |
2871 | ||
2872 | /* Check for the special R40-R44 regs used by FPX extension. */ | |
2873 | if (arc_must_save_register (TARGET_BIG_ENDIAN ? R41_REG : R40_REG, | |
2874 | cfun, TARGET_DPFP)) | |
2875 | reg_size += UNITS_PER_WORD * 2; | |
2876 | if (arc_must_save_register (TARGET_BIG_ENDIAN ? R43_REG : R42_REG, | |
2877 | cfun, TARGET_DPFP)) | |
2878 | reg_size += UNITS_PER_WORD * 2; | |
2879 | ||
90b48013 | 2880 | /* 4) Calculate extra size made up of the blink + fp size. */ |
526b7aee | 2881 | extra_size = 0; |
41453183 | 2882 | if (arc_must_save_return_addr (cfun)) |
526b7aee | 2883 | extra_size = 4; |
41453183 | 2884 | if (arc_frame_pointer_needed ()) |
526b7aee SV |
2885 | extra_size += 4; |
2886 | ||
2887 | /* 5) Space for variable arguments passed in registers */ | |
2888 | pretend_size = crtl->args.pretend_args_size; | |
2889 | ||
2890 | /* Ensure everything before the locals is aligned appropriately. */ | |
90b48013 CZ |
2891 | extra_plus_reg_size = extra_size + reg_size; |
2892 | extra_plus_reg_size_aligned = ARC_STACK_ALIGN (extra_plus_reg_size); | |
2893 | reg_size = extra_plus_reg_size_aligned - extra_size; | |
526b7aee SV |
2894 | |
2895 | /* Compute total frame size. */ | |
2896 | total_size = var_size + args_size + extra_size + pretend_size + reg_size; | |
2897 | ||
6fe5e235 CZ |
2898 | /* It used to be the case that the alignment was forced at this |
2899 | point. However, that is dangerous, calculations based on | |
2900 | total_size would be wrong. Given that this has never cropped up | |
2901 | as an issue I've changed this to an assert for now. */ | |
2902 | gcc_assert (total_size == ARC_STACK_ALIGN (total_size)); | |
526b7aee | 2903 | |
526b7aee SV |
2904 | /* Save computed information. */ |
2905 | frame_info->total_size = total_size; | |
2906 | frame_info->extra_size = extra_size; | |
2907 | frame_info->pretend_size = pretend_size; | |
2908 | frame_info->var_size = var_size; | |
2909 | frame_info->args_size = args_size; | |
2910 | frame_info->reg_size = reg_size; | |
526b7aee SV |
2911 | frame_info->gmask = gmask; |
2912 | frame_info->initialized = reload_completed; | |
2913 | ||
2914 | /* Ok, we're done. */ | |
2915 | return total_size; | |
2916 | } | |
2917 | ||
41453183 CZ |
2918 | /* Build dwarf information when the context is saved via AUX_IRQ_CTRL |
2919 | mechanism. */ | |
2920 | ||
2921 | static void | |
2922 | arc_dwarf_emit_irq_save_regs (void) | |
2923 | { | |
2924 | rtx tmp, par, insn, reg; | |
2925 | int i, offset, j; | |
2926 | ||
2927 | par = gen_rtx_SEQUENCE (VOIDmode, | |
2928 | rtvec_alloc (irq_ctrl_saved.irq_save_last_reg + 1 | |
2929 | + irq_ctrl_saved.irq_save_blink | |
2930 | + irq_ctrl_saved.irq_save_lpcount | |
2931 | + 1)); | |
2932 | ||
2933 | /* Build the stack adjustment note for unwind info. */ | |
2934 | j = 0; | |
2935 | offset = UNITS_PER_WORD * (irq_ctrl_saved.irq_save_last_reg + 1 | |
2936 | + irq_ctrl_saved.irq_save_blink | |
2937 | + irq_ctrl_saved.irq_save_lpcount); | |
2938 | tmp = plus_constant (Pmode, stack_pointer_rtx, -1 * offset); | |
2939 | tmp = gen_rtx_SET (stack_pointer_rtx, tmp); | |
2940 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2941 | XVECEXP (par, 0, j++) = tmp; | |
2942 | ||
2943 | offset -= UNITS_PER_WORD; | |
2944 | ||
2945 | /* 1st goes LP_COUNT. */ | |
2946 | if (irq_ctrl_saved.irq_save_lpcount) | |
2947 | { | |
2948 | reg = gen_rtx_REG (SImode, 60); | |
2949 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
2950 | tmp = gen_frame_mem (SImode, tmp); | |
2951 | tmp = gen_rtx_SET (tmp, reg); | |
2952 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2953 | XVECEXP (par, 0, j++) = tmp; | |
2954 | offset -= UNITS_PER_WORD; | |
2955 | } | |
2956 | ||
2957 | /* 2nd goes BLINK. */ | |
2958 | if (irq_ctrl_saved.irq_save_blink) | |
2959 | { | |
2960 | reg = gen_rtx_REG (SImode, 31); | |
2961 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
2962 | tmp = gen_frame_mem (SImode, tmp); | |
2963 | tmp = gen_rtx_SET (tmp, reg); | |
2964 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2965 | XVECEXP (par, 0, j++) = tmp; | |
2966 | offset -= UNITS_PER_WORD; | |
2967 | } | |
2968 | ||
2969 | /* Build the parallel of the remaining registers recorded as saved | |
2970 | for unwind. */ | |
2971 | for (i = irq_ctrl_saved.irq_save_last_reg; i >= 0; i--) | |
2972 | { | |
2973 | reg = gen_rtx_REG (SImode, i); | |
2974 | tmp = plus_constant (Pmode, stack_pointer_rtx, offset); | |
2975 | tmp = gen_frame_mem (SImode, tmp); | |
2976 | tmp = gen_rtx_SET (tmp, reg); | |
2977 | RTX_FRAME_RELATED_P (tmp) = 1; | |
2978 | XVECEXP (par, 0, j++) = tmp; | |
2979 | offset -= UNITS_PER_WORD; | |
2980 | } | |
2981 | ||
2982 | /* Dummy insn used to anchor the dwarf info. */ | |
2983 | insn = emit_insn (gen_stack_irq_dwarf()); | |
2984 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, par); | |
2985 | RTX_FRAME_RELATED_P (insn) = 1; | |
2986 | } | |
2987 | ||
90b48013 CZ |
2988 | /* Helper for prologue: emit frame store with pre_modify or pre_dec to |
2989 | save register REG on stack. An initial offset OFFSET can be passed | |
2990 | to the function. */ | |
2991 | ||
2992 | static int | |
2993 | frame_save_reg (rtx reg, HOST_WIDE_INT offset) | |
2994 | { | |
2995 | rtx addr; | |
2996 | ||
2997 | if (offset) | |
2998 | { | |
2999 | rtx tmp = plus_constant (Pmode, stack_pointer_rtx, | |
3000 | offset - GET_MODE_SIZE (GET_MODE (reg))); | |
3001 | addr = gen_frame_mem (GET_MODE (reg), | |
3002 | gen_rtx_PRE_MODIFY (Pmode, | |
3003 | stack_pointer_rtx, | |
3004 | tmp)); | |
3005 | } | |
3006 | else | |
3007 | addr = gen_frame_mem (GET_MODE (reg), gen_rtx_PRE_DEC (Pmode, | |
3008 | stack_pointer_rtx)); | |
3009 | frame_move_inc (addr, reg, stack_pointer_rtx, 0); | |
3010 | ||
3011 | return GET_MODE_SIZE (GET_MODE (reg)) - offset; | |
3012 | } | |
3013 | ||
ce9dbf20 CZ |
3014 | /* Helper used when saving AUX regs during ISR. */ |
3015 | ||
3016 | static int | |
3017 | push_reg (rtx reg) | |
3018 | { | |
3019 | rtx stkslot = gen_rtx_MEM (GET_MODE (reg), gen_rtx_PRE_DEC (Pmode, | |
3020 | stack_pointer_rtx)); | |
3021 | rtx insn = emit_move_insn (stkslot, reg); | |
3022 | RTX_FRAME_RELATED_P (insn) = 1; | |
3023 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
3024 | gen_rtx_SET (stack_pointer_rtx, | |
3025 | plus_constant (Pmode, stack_pointer_rtx, | |
3026 | -GET_MODE_SIZE (GET_MODE (reg))))); | |
3027 | return GET_MODE_SIZE (GET_MODE (reg)); | |
3028 | } | |
3029 | ||
90b48013 CZ |
3030 | /* Helper for epilogue: emit frame load with post_modify or post_inc |
3031 | to restore register REG from stack. The initial offset is passed | |
3032 | via OFFSET. */ | |
3033 | ||
3034 | static int | |
3035 | frame_restore_reg (rtx reg, HOST_WIDE_INT offset) | |
3036 | { | |
3037 | rtx addr, insn; | |
3038 | ||
3039 | if (offset) | |
3040 | { | |
3041 | rtx tmp = plus_constant (Pmode, stack_pointer_rtx, | |
3042 | offset + GET_MODE_SIZE (GET_MODE (reg))); | |
3043 | addr = gen_frame_mem (GET_MODE (reg), | |
3044 | gen_rtx_POST_MODIFY (Pmode, | |
3045 | stack_pointer_rtx, | |
3046 | tmp)); | |
3047 | } | |
3048 | else | |
3049 | addr = gen_frame_mem (GET_MODE (reg), gen_rtx_POST_INC (Pmode, | |
3050 | stack_pointer_rtx)); | |
3051 | insn = frame_move_inc (reg, addr, stack_pointer_rtx, 0); | |
3052 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3053 | ||
47d8cb23 | 3054 | if (reg == hard_frame_pointer_rtx) |
90b48013 CZ |
3055 | add_reg_note (insn, REG_CFA_DEF_CFA, |
3056 | plus_constant (Pmode, stack_pointer_rtx, | |
3057 | GET_MODE_SIZE (GET_MODE (reg)) + offset)); | |
3058 | else | |
3059 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
3060 | gen_rtx_SET (stack_pointer_rtx, | |
3061 | plus_constant (Pmode, stack_pointer_rtx, | |
3062 | GET_MODE_SIZE (GET_MODE (reg)) | |
3063 | + offset))); | |
3064 | ||
3065 | return GET_MODE_SIZE (GET_MODE (reg)) + offset; | |
3066 | } | |
3067 | ||
ce9dbf20 CZ |
3068 | /* Helper used when restoring AUX regs during ISR. */ |
3069 | ||
3070 | static int | |
3071 | pop_reg (rtx reg) | |
3072 | { | |
3073 | rtx stkslot = gen_rtx_MEM (GET_MODE (reg), gen_rtx_POST_INC (Pmode, | |
3074 | stack_pointer_rtx)); | |
3075 | rtx insn = emit_move_insn (reg, stkslot); | |
3076 | RTX_FRAME_RELATED_P (insn) = 1; | |
3077 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
3078 | gen_rtx_SET (stack_pointer_rtx, | |
3079 | plus_constant (Pmode, stack_pointer_rtx, | |
3080 | GET_MODE_SIZE (GET_MODE (reg))))); | |
3081 | return GET_MODE_SIZE (GET_MODE (reg)); | |
3082 | } | |
3083 | ||
3084 | ||
90b48013 CZ |
3085 | /* Check if we have a continous range to be save/restored with the |
3086 | help of enter/leave instructions. A vaild register range starts | |
3087 | from $r13 and is up to (including) $r26. */ | |
3088 | ||
3089 | static bool | |
ce9dbf20 | 3090 | arc_enter_leave_p (uint64_t gmask) |
90b48013 CZ |
3091 | { |
3092 | int regno; | |
3093 | unsigned int rmask = 0; | |
3094 | ||
3095 | if (!gmask) | |
3096 | return false; | |
3097 | ||
3098 | for (regno = ENTER_LEAVE_START_REG; | |
ce9dbf20 CZ |
3099 | regno <= ENTER_LEAVE_END_REG && (gmask & (1ULL << regno)); regno++) |
3100 | rmask |= 1ULL << regno; | |
90b48013 CZ |
3101 | |
3102 | if (rmask ^ gmask) | |
3103 | return false; | |
3104 | ||
3105 | return true; | |
3106 | } | |
3107 | ||
3108 | /* ARC's prologue, save any needed call-saved regs (and call-used if | |
3109 | this is an interrupt handler) for ARCompact ISA, using ST/STD | |
3110 | instructions. */ | |
3111 | ||
3112 | static int | |
ce9dbf20 | 3113 | arc_save_callee_saves (uint64_t gmask, |
90b48013 CZ |
3114 | bool save_blink, |
3115 | bool save_fp, | |
3116 | HOST_WIDE_INT offset) | |
3117 | { | |
3118 | rtx reg; | |
3119 | int frame_allocated = 0; | |
ce9dbf20 | 3120 | int i; |
90b48013 CZ |
3121 | |
3122 | /* The home-grown ABI says link register is saved first. */ | |
3123 | if (save_blink) | |
3124 | { | |
3125 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3126 | frame_allocated += frame_save_reg (reg, offset); | |
3127 | offset = 0; | |
3128 | } | |
3129 | ||
3130 | /* N.B. FRAME_POINTER_MASK and RETURN_ADDR_MASK are cleared in gmask. */ | |
3131 | if (gmask) | |
ce9dbf20 | 3132 | for (i = GMASK_LEN; i >= 0; i--) |
90b48013 CZ |
3133 | { |
3134 | machine_mode save_mode = SImode; | |
3135 | ||
3136 | if (TARGET_LL64 | |
3137 | && ((i - 1) % 2 == 0) | |
ce9dbf20 CZ |
3138 | && ((gmask & (1ULL << i)) != 0) |
3139 | && ((gmask & (1ULL << (i - 1))) != 0)) | |
90b48013 CZ |
3140 | { |
3141 | save_mode = DImode; | |
3142 | --i; | |
3143 | } | |
ce9dbf20 | 3144 | else if ((gmask & (1ULL << i)) == 0) |
90b48013 CZ |
3145 | continue; |
3146 | ||
3147 | reg = gen_rtx_REG (save_mode, i); | |
3148 | frame_allocated += frame_save_reg (reg, offset); | |
3149 | offset = 0; | |
3150 | } | |
3151 | ||
ce9dbf20 CZ |
3152 | /* Check if we need to save the ZOL machinery. */ |
3153 | if (arc_lpcwidth != 0 && arc_must_save_register (LP_COUNT, cfun, true)) | |
3154 | { | |
3155 | rtx reg0 = gen_rtx_REG (SImode, R0_REG); | |
3156 | emit_insn (gen_rtx_SET (reg0, | |
3157 | gen_rtx_UNSPEC_VOLATILE | |
3158 | (Pmode, gen_rtvec (1, GEN_INT (AUX_LP_START)), | |
3159 | VUNSPEC_ARC_LR))); | |
3160 | frame_allocated += push_reg (reg0); | |
3161 | emit_insn (gen_rtx_SET (reg0, | |
3162 | gen_rtx_UNSPEC_VOLATILE | |
3163 | (Pmode, gen_rtvec (1, GEN_INT (AUX_LP_END)), | |
3164 | VUNSPEC_ARC_LR))); | |
3165 | frame_allocated += push_reg (reg0); | |
3166 | emit_move_insn (reg0, gen_rtx_REG (SImode, LP_COUNT)); | |
3167 | frame_allocated += push_reg (reg0); | |
3168 | } | |
3169 | ||
3170 | /* Save AUX regs used by FPX machinery. */ | |
3171 | if (arc_must_save_register (TARGET_BIG_ENDIAN ? R41_REG : R40_REG, | |
3172 | cfun, TARGET_DPFP)) | |
3173 | { | |
3174 | rtx reg0 = gen_rtx_REG (SImode, R0_REG); | |
3175 | ||
3176 | for (i = 0; i < 4; i++) | |
3177 | { | |
3178 | emit_insn (gen_rtx_SET (reg0, | |
3179 | gen_rtx_UNSPEC_VOLATILE | |
3180 | (Pmode, gen_rtvec (1, GEN_INT (AUX_DPFP_START | |
3181 | + i)), | |
3182 | VUNSPEC_ARC_LR))); | |
3183 | frame_allocated += push_reg (reg0); | |
3184 | } | |
3185 | } | |
3186 | ||
90b48013 CZ |
3187 | /* Save frame pointer if needed. First save the FP on stack, if not |
3188 | autosaved. Unfortunately, I cannot add it to gmask and use the | |
3189 | above loop to save fp because our ABI states fp goes aftert all | |
3190 | registers are saved. */ | |
3191 | if (save_fp) | |
3192 | { | |
47d8cb23 | 3193 | frame_allocated += frame_save_reg (hard_frame_pointer_rtx, offset); |
90b48013 CZ |
3194 | offset = 0; |
3195 | } | |
3196 | ||
3197 | /* Emit mov fp,sp. */ | |
3198 | if (arc_frame_pointer_needed ()) | |
47d8cb23 | 3199 | frame_move (hard_frame_pointer_rtx, stack_pointer_rtx); |
90b48013 CZ |
3200 | |
3201 | return frame_allocated; | |
3202 | } | |
3203 | ||
3204 | /* ARC's epilogue, restore any required call-saved regs (and call-used | |
3205 | if it is for an interrupt handler) using LD/LDD instructions. */ | |
3206 | ||
3207 | static int | |
ce9dbf20 | 3208 | arc_restore_callee_saves (uint64_t gmask, |
90b48013 CZ |
3209 | bool restore_blink, |
3210 | bool restore_fp, | |
3211 | HOST_WIDE_INT offset, | |
3212 | HOST_WIDE_INT allocated) | |
3213 | { | |
3214 | rtx reg; | |
3215 | int frame_deallocated = 0; | |
8f845213 CZ |
3216 | HOST_WIDE_INT offs = cfun->machine->frame_info.reg_size; |
3217 | bool early_blink_restore; | |
ce9dbf20 | 3218 | int i; |
90b48013 CZ |
3219 | |
3220 | /* Emit mov fp,sp. */ | |
3221 | if (arc_frame_pointer_needed () && offset) | |
3222 | { | |
47d8cb23 | 3223 | frame_move (stack_pointer_rtx, hard_frame_pointer_rtx); |
90b48013 CZ |
3224 | frame_deallocated += offset; |
3225 | offset = 0; | |
3226 | } | |
3227 | ||
3228 | if (restore_fp) | |
3229 | { | |
3230 | /* Any offset is taken care by previous if-statement. */ | |
3231 | gcc_assert (offset == 0); | |
47d8cb23 | 3232 | frame_deallocated += frame_restore_reg (hard_frame_pointer_rtx, 0); |
90b48013 CZ |
3233 | } |
3234 | ||
ce9dbf20 CZ |
3235 | /* Restore AUX-regs used by FPX machinery. */ |
3236 | if (arc_must_save_register (TARGET_BIG_ENDIAN ? R41_REG : R40_REG, | |
3237 | cfun, TARGET_DPFP)) | |
3238 | { | |
3239 | rtx reg0 = gen_rtx_REG (SImode, R0_REG); | |
3240 | ||
3241 | gcc_assert (offset == 0); | |
3242 | for (i = 0; i < 4; i++) | |
3243 | { | |
3244 | frame_deallocated += pop_reg (reg0); | |
3245 | emit_insn (gen_rtx_UNSPEC_VOLATILE | |
3246 | (VOIDmode, gen_rtvec (2, reg0, GEN_INT (AUX_DPFP_START | |
3247 | + i)), | |
3248 | VUNSPEC_ARC_SR)); | |
3249 | } | |
3250 | } | |
3251 | ||
3252 | /* Check if we need to restore the ZOL machinery. */ | |
3253 | if (arc_lpcwidth !=0 && arc_must_save_register (LP_COUNT, cfun, true)) | |
3254 | { | |
3255 | rtx reg0 = gen_rtx_REG (SImode, R0_REG); | |
3256 | ||
3257 | gcc_assert (offset == 0); | |
3258 | frame_deallocated += pop_reg (reg0); | |
3259 | emit_move_insn (gen_rtx_REG (SImode, LP_COUNT), reg0); | |
3260 | ||
3261 | frame_deallocated += pop_reg (reg0); | |
3262 | emit_insn (gen_rtx_UNSPEC_VOLATILE | |
3263 | (VOIDmode, gen_rtvec (2, reg0, GEN_INT (AUX_LP_END)), | |
3264 | VUNSPEC_ARC_SR)); | |
3265 | ||
3266 | frame_deallocated += pop_reg (reg0); | |
3267 | emit_insn (gen_rtx_UNSPEC_VOLATILE | |
3268 | (VOIDmode, gen_rtvec (2, reg0, GEN_INT (AUX_LP_START)), | |
3269 | VUNSPEC_ARC_SR)); | |
3270 | } | |
3271 | ||
90b48013 CZ |
3272 | if (offset) |
3273 | { | |
3274 | /* No $fp involved, we need to do an add to set the $sp to the | |
3275 | location of the first register. */ | |
3276 | frame_stack_add (offset); | |
3277 | frame_deallocated += offset; | |
3278 | offset = 0; | |
3279 | } | |
3280 | ||
8f845213 | 3281 | /* When we do not optimize for size, restore first blink. */ |
ce9dbf20 CZ |
3282 | early_blink_restore = restore_blink && !optimize_size && offs |
3283 | && !ARC_INTERRUPT_P (arc_compute_function_type (cfun)); | |
8f845213 CZ |
3284 | if (early_blink_restore) |
3285 | { | |
3286 | rtx addr = plus_constant (Pmode, stack_pointer_rtx, offs); | |
3287 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3288 | rtx insn = frame_move_inc (reg, gen_frame_mem (Pmode, addr), | |
3289 | stack_pointer_rtx, NULL_RTX); | |
3290 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3291 | restore_blink = false; | |
3292 | } | |
3293 | ||
90b48013 CZ |
3294 | /* N.B. FRAME_POINTER_MASK and RETURN_ADDR_MASK are cleared in gmask. */ |
3295 | if (gmask) | |
ce9dbf20 | 3296 | for (i = 0; i <= GMASK_LEN; i++) |
90b48013 CZ |
3297 | { |
3298 | machine_mode restore_mode = SImode; | |
3299 | ||
3300 | if (TARGET_LL64 | |
3301 | && ((i % 2) == 0) | |
ce9dbf20 CZ |
3302 | && ((gmask & (1ULL << i)) != 0) |
3303 | && ((gmask & (1ULL << (i + 1))) != 0)) | |
90b48013 | 3304 | restore_mode = DImode; |
ce9dbf20 | 3305 | else if ((gmask & (1ULL << i)) == 0) |
90b48013 CZ |
3306 | continue; |
3307 | ||
3308 | reg = gen_rtx_REG (restore_mode, i); | |
8f845213 CZ |
3309 | offs = 0; |
3310 | switch (restore_mode) | |
3311 | { | |
3312 | case E_DImode: | |
ce9dbf20 | 3313 | if ((GMASK_LEN - __builtin_clzll (gmask)) == (i + 1) |
8f845213 CZ |
3314 | && early_blink_restore) |
3315 | offs = 4; | |
3316 | break; | |
3317 | case E_SImode: | |
ce9dbf20 | 3318 | if ((GMASK_LEN - __builtin_clzll (gmask)) == i |
8f845213 CZ |
3319 | && early_blink_restore) |
3320 | offs = 4; | |
3321 | break; | |
3322 | default: | |
3323 | offs = 0; | |
3324 | } | |
3325 | frame_deallocated += frame_restore_reg (reg, offs); | |
90b48013 CZ |
3326 | offset = 0; |
3327 | ||
3328 | if (restore_mode == DImode) | |
3329 | i++; | |
3330 | } | |
3331 | ||
3332 | if (restore_blink) | |
3333 | { | |
3334 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3335 | frame_deallocated += frame_restore_reg (reg, allocated | |
3336 | - frame_deallocated | |
3337 | /* Consider as well the | |
3338 | current restored | |
3339 | register size. */ | |
3340 | - UNITS_PER_WORD); | |
3341 | } | |
3342 | ||
3343 | return frame_deallocated; | |
3344 | } | |
3345 | ||
3346 | /* ARC prologue, save the registers using enter instruction. Leave | |
3347 | instruction can also save $blink (SAVE_BLINK) and $fp (SAVE_FP) | |
3348 | register. */ | |
3349 | ||
3350 | static int | |
ce9dbf20 | 3351 | arc_save_callee_enter (uint64_t gmask, |
90b48013 CZ |
3352 | bool save_blink, |
3353 | bool save_fp, | |
3354 | HOST_WIDE_INT offset) | |
3355 | { | |
3356 | int start_reg = ENTER_LEAVE_START_REG; | |
3357 | int end_reg = ENTER_LEAVE_END_REG; | |
3358 | int regno, indx, off, nregs; | |
3359 | rtx insn, reg, mem; | |
3360 | int frame_allocated = 0; | |
3361 | ||
ce9dbf20 | 3362 | for (regno = start_reg; regno <= end_reg && (gmask & (1ULL << regno));) |
90b48013 CZ |
3363 | regno++; |
3364 | ||
3365 | end_reg = regno - 1; | |
3366 | nregs = end_reg - start_reg + 1; | |
3367 | nregs += save_blink ? 1 : 0; | |
3368 | nregs += save_fp ? 1 : 0; | |
3369 | ||
3370 | if (offset) | |
3371 | frame_stack_add (offset); | |
3372 | ||
3373 | insn = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nregs + (save_fp ? 1 : 0) | |
3374 | + 1)); | |
3375 | indx = 0; | |
3376 | ||
3377 | reg = gen_rtx_SET (stack_pointer_rtx, | |
3378 | plus_constant (Pmode, | |
3379 | stack_pointer_rtx, | |
f5d56cf9 | 3380 | -nregs * UNITS_PER_WORD)); |
90b48013 CZ |
3381 | RTX_FRAME_RELATED_P (reg) = 1; |
3382 | XVECEXP (insn, 0, indx++) = reg; | |
3383 | off = nregs * UNITS_PER_WORD; | |
3384 | ||
3385 | if (save_blink) | |
3386 | { | |
3387 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3388 | mem = gen_frame_mem (Pmode, plus_constant (Pmode, | |
3389 | stack_pointer_rtx, | |
3390 | off)); | |
3391 | XVECEXP (insn, 0, indx) = gen_rtx_SET (mem, reg); | |
3392 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx++)) = 1; | |
3393 | off -= UNITS_PER_WORD; | |
3394 | save_blink = false; | |
3395 | } | |
3396 | ||
3397 | for (regno = start_reg; | |
3398 | regno <= end_reg; | |
3399 | regno++, indx++, off -= UNITS_PER_WORD) | |
3400 | { | |
3401 | reg = gen_rtx_REG (SImode, regno); | |
3402 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3403 | stack_pointer_rtx, | |
3404 | off)); | |
3405 | XVECEXP (insn, 0, indx) = gen_rtx_SET (mem, reg); | |
3406 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx)) = 1; | |
ce9dbf20 | 3407 | gmask = gmask & ~(1ULL << regno); |
90b48013 CZ |
3408 | } |
3409 | ||
3410 | if (save_fp) | |
3411 | { | |
3412 | mem = gen_frame_mem (Pmode, plus_constant (Pmode, | |
3413 | stack_pointer_rtx, | |
3414 | off)); | |
47d8cb23 | 3415 | XVECEXP (insn, 0, indx) = gen_rtx_SET (mem, hard_frame_pointer_rtx); |
90b48013 CZ |
3416 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx++)) = 1; |
3417 | off -= UNITS_PER_WORD; | |
3418 | ||
47d8cb23 | 3419 | XVECEXP (insn, 0, indx) = gen_rtx_SET (hard_frame_pointer_rtx, |
90b48013 CZ |
3420 | stack_pointer_rtx); |
3421 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx++)) = 1; | |
3422 | save_fp = false; | |
3423 | } | |
3424 | ||
3425 | gcc_assert (off == 0); | |
3426 | insn = frame_insn (insn); | |
3427 | ||
3428 | add_reg_note (insn, REG_INC, stack_pointer_rtx); | |
3429 | ||
3430 | frame_allocated = nregs * UNITS_PER_WORD; | |
3431 | ||
3432 | /* offset is a negative number, make sure we add it. */ | |
3433 | return frame_allocated - offset; | |
3434 | } | |
3435 | ||
3436 | /* ARC epilogue, restore the registers using leave instruction. An | |
3437 | initial offset is passed in OFFSET. Besides restoring an register | |
3438 | range, leave can also restore $blink (RESTORE_BLINK), or $fp | |
3439 | (RESTORE_FP), and can automatic return (RETURN_P). */ | |
3440 | ||
3441 | static int | |
ce9dbf20 | 3442 | arc_restore_callee_leave (uint64_t gmask, |
90b48013 CZ |
3443 | bool restore_blink, |
3444 | bool restore_fp, | |
3445 | bool return_p, | |
3446 | HOST_WIDE_INT offset) | |
3447 | { | |
3448 | int start_reg = ENTER_LEAVE_START_REG; | |
3449 | int end_reg = ENTER_LEAVE_END_REG; | |
3450 | int regno, indx, off, nregs; | |
3451 | rtx insn, reg, mem; | |
3452 | int frame_allocated = 0; | |
3453 | ||
ce9dbf20 | 3454 | for (regno = start_reg; regno <= end_reg && (gmask & (1ULL << regno));) |
90b48013 CZ |
3455 | regno++; |
3456 | ||
3457 | end_reg = regno - 1; | |
3458 | nregs = end_reg - start_reg + 1; | |
3459 | nregs += restore_blink ? 1 : 0; | |
3460 | nregs += restore_fp ? 1 : 0; | |
3461 | ||
3462 | insn = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nregs + 1 | |
3463 | + (return_p ? 1 : 0))); | |
3464 | indx = 0; | |
3465 | ||
3466 | if (return_p) | |
3467 | XVECEXP (insn, 0, indx++) = ret_rtx; | |
3468 | ||
3469 | if (restore_fp) | |
3470 | { | |
3471 | /* I cannot emit set (sp, fp) here as cselib expects a single sp | |
3472 | set and not two. Thus, use the offset, and change sp adjust | |
3473 | value. */ | |
3474 | frame_allocated += offset; | |
3475 | } | |
3476 | ||
3477 | if (offset && !restore_fp) | |
3478 | { | |
3479 | /* This add is only emmited when we do not restore fp with leave | |
3480 | instruction. */ | |
3481 | frame_stack_add (offset); | |
3482 | frame_allocated += offset; | |
3483 | offset = 0; | |
3484 | } | |
3485 | ||
3486 | reg = gen_rtx_SET (stack_pointer_rtx, | |
3487 | plus_constant (Pmode, | |
3488 | stack_pointer_rtx, | |
3489 | offset + nregs * UNITS_PER_WORD)); | |
3490 | RTX_FRAME_RELATED_P (reg) = 1; | |
3491 | XVECEXP (insn, 0, indx++) = reg; | |
3492 | off = nregs * UNITS_PER_WORD; | |
3493 | ||
3494 | if (restore_blink) | |
3495 | { | |
3496 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3497 | mem = gen_frame_mem (Pmode, plus_constant (Pmode, | |
3498 | stack_pointer_rtx, | |
3499 | off)); | |
3500 | XVECEXP (insn, 0, indx) = gen_rtx_SET (reg, mem); | |
3501 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx++)) = 1; | |
3502 | off -= UNITS_PER_WORD; | |
3503 | } | |
3504 | ||
3505 | for (regno = start_reg; | |
3506 | regno <= end_reg; | |
3507 | regno++, indx++, off -= UNITS_PER_WORD) | |
3508 | { | |
3509 | reg = gen_rtx_REG (SImode, regno); | |
3510 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3511 | stack_pointer_rtx, | |
3512 | off)); | |
3513 | XVECEXP (insn, 0, indx) = gen_rtx_SET (reg, mem); | |
3514 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx)) = 1; | |
ce9dbf20 | 3515 | gmask = gmask & ~(1ULL << regno); |
90b48013 CZ |
3516 | } |
3517 | ||
3518 | if (restore_fp) | |
3519 | { | |
3520 | mem = gen_frame_mem (Pmode, plus_constant (Pmode, | |
3521 | stack_pointer_rtx, | |
3522 | off)); | |
47d8cb23 | 3523 | XVECEXP (insn, 0, indx) = gen_rtx_SET (hard_frame_pointer_rtx, mem); |
90b48013 CZ |
3524 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx++)) = 1; |
3525 | off -= UNITS_PER_WORD; | |
3526 | } | |
3527 | ||
3528 | gcc_assert (off == 0); | |
3529 | if (return_p) | |
3530 | { | |
3531 | insn = emit_jump_insn (insn); | |
3532 | RTX_FRAME_RELATED_P (insn) = 1; | |
3533 | } | |
3534 | else | |
3535 | insn = frame_insn (insn); | |
3536 | ||
3537 | add_reg_note (insn, REG_INC, stack_pointer_rtx); | |
3538 | ||
3539 | /* Dwarf related info. */ | |
3540 | if (restore_fp) | |
3541 | { | |
47d8cb23 | 3542 | add_reg_note (insn, REG_CFA_RESTORE, hard_frame_pointer_rtx); |
90b48013 CZ |
3543 | add_reg_note (insn, REG_CFA_DEF_CFA, |
3544 | plus_constant (Pmode, stack_pointer_rtx, | |
3545 | offset + nregs * UNITS_PER_WORD)); | |
3546 | } | |
3547 | else | |
3548 | { | |
3549 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
3550 | gen_rtx_SET (stack_pointer_rtx, | |
3551 | plus_constant (Pmode, stack_pointer_rtx, | |
3552 | nregs * UNITS_PER_WORD))); | |
3553 | } | |
3554 | if (restore_blink) | |
3555 | add_reg_note (insn, REG_CFA_RESTORE, | |
3556 | gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM)); | |
3557 | for (regno = start_reg; regno <= end_reg; regno++) | |
3558 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (SImode, regno)); | |
3559 | ||
3560 | frame_allocated += nregs * UNITS_PER_WORD; | |
3561 | ||
3562 | return frame_allocated; | |
3563 | } | |
3564 | ||
3565 | /* Millicode thunks implementation: | |
3566 | Generates calls to millicodes for registers starting from r13 to r25 | |
3567 | Present Limitations: | |
3568 | - Only one range supported. The remaining regs will have the ordinary | |
3569 | st and ld instructions for store and loads. Hence a gmask asking | |
3570 | to store r13-14, r16-r25 will only generate calls to store and | |
3571 | load r13 to r14 while store and load insns will be generated for | |
3572 | r16 to r25 in the prologue and epilogue respectively. | |
3573 | ||
3574 | - Presently library only supports register ranges starting from r13. | |
3575 | */ | |
3576 | ||
3577 | static int | |
ce9dbf20 | 3578 | arc_save_callee_milli (uint64_t gmask, |
90b48013 CZ |
3579 | bool save_blink, |
3580 | bool save_fp, | |
3581 | HOST_WIDE_INT offset, | |
3582 | HOST_WIDE_INT reg_size) | |
3583 | { | |
3584 | int start_reg = 13; | |
3585 | int end_reg = 25; | |
3586 | int regno, indx, off, nregs; | |
3587 | rtx insn, reg, mem; | |
3588 | int frame_allocated = 0; | |
3589 | ||
ce9dbf20 | 3590 | for (regno = start_reg; regno <= end_reg && (gmask & (1ULL << regno));) |
90b48013 CZ |
3591 | regno++; |
3592 | ||
3593 | end_reg = regno - 1; | |
3594 | nregs = end_reg - start_reg + 1; | |
3595 | gcc_assert (end_reg > 14); | |
3596 | ||
3597 | ||
3598 | /* Allocate space on stack for the registers, and take into account | |
3599 | also the initial offset. The registers will be saved using | |
3600 | offsets. N.B. OFFSET is a negative number. */ | |
3601 | if (save_blink) | |
3602 | { | |
3603 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3604 | frame_allocated += frame_save_reg (reg, offset); | |
3605 | offset = 0; | |
3606 | } | |
3607 | ||
3608 | if (reg_size || offset) | |
3609 | { | |
3610 | frame_stack_add (offset - reg_size); | |
3611 | frame_allocated += nregs * UNITS_PER_WORD - offset; | |
3612 | offset = 0; | |
3613 | } | |
3614 | ||
3615 | /* Start generate millicode call. */ | |
3616 | insn = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nregs + 1)); | |
3617 | indx = 0; | |
3618 | ||
3619 | /* This is a call, we clobber blink. */ | |
3620 | XVECEXP (insn, 0, nregs) = | |
3621 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM)); | |
3622 | ||
3623 | for (regno = start_reg, indx = 0, off = 0; | |
3624 | regno <= end_reg; | |
3625 | regno++, indx++, off += UNITS_PER_WORD) | |
3626 | { | |
3627 | reg = gen_rtx_REG (SImode, regno); | |
3628 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3629 | stack_pointer_rtx, | |
3630 | off)); | |
3631 | XVECEXP (insn, 0, indx) = gen_rtx_SET (mem, reg); | |
3632 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx)) = 1; | |
ce9dbf20 | 3633 | gmask = gmask & ~(1ULL << regno); |
90b48013 CZ |
3634 | } |
3635 | insn = frame_insn (insn); | |
3636 | ||
3637 | /* Add DWARF info. */ | |
3638 | for (regno = start_reg, off = 0; | |
3639 | regno <= end_reg; | |
3640 | regno++, off += UNITS_PER_WORD) | |
3641 | { | |
3642 | reg = gen_rtx_REG (SImode, regno); | |
3643 | mem = gen_rtx_MEM (SImode, plus_constant (Pmode, | |
3644 | stack_pointer_rtx, off)); | |
3645 | add_reg_note (insn, REG_CFA_OFFSET, gen_rtx_SET (mem, reg)); | |
3646 | ||
3647 | } | |
3648 | ||
3649 | /* In the case of millicode thunk, we need to restore the | |
3650 | clobbered blink register. */ | |
3651 | if (arc_must_save_return_addr (cfun)) | |
3652 | { | |
3653 | emit_insn (gen_rtx_SET (gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM), | |
3654 | gen_rtx_MEM (Pmode, | |
3655 | plus_constant (Pmode, | |
3656 | stack_pointer_rtx, | |
3657 | reg_size)))); | |
3658 | } | |
3659 | ||
3660 | /* Save remaining registers using st instructions. */ | |
ce9dbf20 | 3661 | for (regno = 0; regno <= GMASK_LEN; regno++) |
90b48013 | 3662 | { |
ce9dbf20 | 3663 | if ((gmask & (1ULL << regno)) == 0) |
90b48013 CZ |
3664 | continue; |
3665 | ||
3666 | reg = gen_rtx_REG (SImode, regno); | |
3667 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3668 | stack_pointer_rtx, | |
3669 | off)); | |
3670 | frame_move_inc (mem, reg, stack_pointer_rtx, 0); | |
3671 | frame_allocated += UNITS_PER_WORD; | |
3672 | off += UNITS_PER_WORD; | |
3673 | } | |
3674 | ||
3675 | /* Save frame pointer if needed. First save the FP on stack, if not | |
3676 | autosaved. Unfortunately, I cannot add it to gmask and use the | |
3677 | above loop to save fp because our ABI states fp goes aftert all | |
3678 | registers are saved. */ | |
3679 | if (save_fp) | |
47d8cb23 | 3680 | frame_allocated += frame_save_reg (hard_frame_pointer_rtx, offset); |
90b48013 CZ |
3681 | |
3682 | /* Emit mov fp,sp. */ | |
3683 | if (arc_frame_pointer_needed ()) | |
47d8cb23 | 3684 | frame_move (hard_frame_pointer_rtx, stack_pointer_rtx); |
90b48013 CZ |
3685 | |
3686 | return frame_allocated; | |
3687 | } | |
3688 | ||
3689 | /* Like the previous function but restore. */ | |
3690 | ||
3691 | static int | |
ce9dbf20 | 3692 | arc_restore_callee_milli (uint64_t gmask, |
90b48013 CZ |
3693 | bool restore_blink, |
3694 | bool restore_fp, | |
3695 | bool return_p, | |
3696 | HOST_WIDE_INT offset) | |
3697 | { | |
3698 | int start_reg = 13; | |
3699 | int end_reg = 25; | |
3700 | int regno, indx, off, nregs; | |
3701 | rtx insn, reg, mem; | |
3702 | int frame_allocated = 0; | |
3703 | ||
ce9dbf20 | 3704 | for (regno = start_reg; regno <= end_reg && (gmask & (1ULL << regno));) |
90b48013 CZ |
3705 | regno++; |
3706 | ||
3707 | end_reg = regno - 1; | |
3708 | nregs = end_reg - start_reg + 1; | |
3709 | gcc_assert (end_reg > 14); | |
3710 | ||
3711 | /* Emit mov fp,sp. */ | |
3712 | if (arc_frame_pointer_needed () && offset) | |
3713 | { | |
47d8cb23 | 3714 | frame_move (stack_pointer_rtx, hard_frame_pointer_rtx); |
90b48013 CZ |
3715 | frame_allocated = offset; |
3716 | offset = 0; | |
3717 | } | |
3718 | ||
3719 | if (restore_fp) | |
47d8cb23 | 3720 | frame_allocated += frame_restore_reg (hard_frame_pointer_rtx, 0); |
90b48013 CZ |
3721 | |
3722 | if (offset) | |
3723 | { | |
3724 | /* No fp involved, hence, we need to adjust the sp via an | |
3725 | add. */ | |
3726 | frame_stack_add (offset); | |
3727 | frame_allocated += offset; | |
3728 | offset = 0; | |
3729 | } | |
3730 | ||
3731 | /* Start generate millicode call. */ | |
3732 | insn = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc ((return_p ? 1 : 0) | |
3733 | + nregs + 1)); | |
3734 | indx = 0; | |
3735 | ||
3736 | if (return_p) | |
3737 | { | |
3738 | /* sibling call, the blink is restored with the help of the | |
3739 | value held into r12. */ | |
3740 | reg = gen_rtx_REG (Pmode, 12); | |
3741 | XVECEXP (insn, 0, indx++) = ret_rtx; | |
3742 | XVECEXP (insn, 0, indx++) = | |
3743 | gen_rtx_SET (stack_pointer_rtx, | |
3744 | gen_rtx_PLUS (Pmode, stack_pointer_rtx, reg)); | |
3745 | frame_allocated += UNITS_PER_WORD; | |
3746 | } | |
3747 | else | |
3748 | { | |
3749 | /* This is a call, we clobber blink. */ | |
3750 | XVECEXP (insn, 0, nregs) = | |
3751 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM)); | |
3752 | } | |
3753 | ||
3754 | for (regno = start_reg, off = 0; | |
3755 | regno <= end_reg; | |
3756 | regno++, indx++, off += UNITS_PER_WORD) | |
3757 | { | |
3758 | reg = gen_rtx_REG (SImode, regno); | |
3759 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3760 | stack_pointer_rtx, | |
3761 | off)); | |
3762 | XVECEXP (insn, 0, indx) = gen_rtx_SET (reg, mem); | |
3763 | RTX_FRAME_RELATED_P (XVECEXP (insn, 0, indx)) = 1; | |
ce9dbf20 | 3764 | gmask = gmask & ~(1ULL << regno); |
90b48013 CZ |
3765 | } |
3766 | ||
3767 | /* Restore remaining registers using LD instructions. */ | |
ce9dbf20 | 3768 | for (regno = 0; regno <= GMASK_LEN; regno++) |
90b48013 | 3769 | { |
ce9dbf20 | 3770 | if ((gmask & (1ULL << regno)) == 0) |
90b48013 CZ |
3771 | continue; |
3772 | ||
3773 | reg = gen_rtx_REG (SImode, regno); | |
3774 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
3775 | stack_pointer_rtx, | |
3776 | off)); | |
3777 | rtx tmp = frame_move_inc (reg, mem, stack_pointer_rtx, 0); | |
3778 | add_reg_note (tmp, REG_CFA_RESTORE, reg); | |
3779 | off += UNITS_PER_WORD; | |
3780 | } | |
3781 | ||
3782 | /* Emit millicode call. */ | |
3783 | if (return_p) | |
3784 | { | |
3785 | reg = gen_rtx_REG (Pmode, 12); | |
3786 | frame_insn (gen_rtx_SET (reg, GEN_INT (off))); | |
3787 | frame_allocated += off; | |
3788 | insn = emit_jump_insn (insn); | |
3789 | RTX_FRAME_RELATED_P (insn) = 1; | |
3790 | } | |
3791 | else | |
3792 | insn = frame_insn (insn); | |
3793 | ||
3794 | /* Add DWARF info. */ | |
544a4843 | 3795 | for (regno = start_reg; regno <= end_reg; regno++) |
90b48013 CZ |
3796 | { |
3797 | reg = gen_rtx_REG (SImode, regno); | |
3798 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3799 | ||
3800 | } | |
3801 | ||
3802 | if (restore_blink && !return_p) | |
3803 | { | |
3804 | reg = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM); | |
3805 | mem = gen_frame_mem (Pmode, plus_constant (Pmode, stack_pointer_rtx, | |
3806 | off)); | |
3807 | insn = frame_insn (gen_rtx_SET (reg, mem)); | |
3808 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3809 | } | |
3810 | ||
3811 | return frame_allocated; | |
3812 | } | |
3813 | ||
526b7aee SV |
3814 | /* Set up the stack and frame pointer (if desired) for the function. */ |
3815 | ||
3816 | void | |
3817 | arc_expand_prologue (void) | |
3818 | { | |
6fe5e235 | 3819 | int size; |
ce9dbf20 | 3820 | uint64_t gmask = cfun->machine->frame_info.gmask; |
90b48013 | 3821 | struct arc_frame_info *frame = &cfun->machine->frame_info; |
526b7aee | 3822 | unsigned int frame_size_to_allocate; |
526b7aee | 3823 | int first_offset = 0; |
1825c61e | 3824 | unsigned int fn_type = arc_compute_function_type (cfun); |
90b48013 CZ |
3825 | bool save_blink = false; |
3826 | bool save_fp = false; | |
1825c61e CZ |
3827 | |
3828 | /* Naked functions don't have prologue. */ | |
3829 | if (ARC_NAKED_P (fn_type)) | |
83b2a5f4 AB |
3830 | { |
3831 | if (flag_stack_usage_info) | |
3832 | current_function_static_stack_size = 0; | |
3833 | return; | |
3834 | } | |
526b7aee | 3835 | |
6fe5e235 CZ |
3836 | /* Compute total frame size. */ |
3837 | size = arc_compute_frame_size (); | |
526b7aee SV |
3838 | |
3839 | if (flag_stack_usage_info) | |
3840 | current_function_static_stack_size = size; | |
3841 | ||
3842 | /* Keep track of frame size to be allocated. */ | |
3843 | frame_size_to_allocate = size; | |
3844 | ||
3845 | /* These cases shouldn't happen. Catch them now. */ | |
3846 | gcc_assert (!(size == 0 && gmask)); | |
3847 | ||
3848 | /* Allocate space for register arguments if this is a variadic function. */ | |
90b48013 CZ |
3849 | if (frame->pretend_size != 0) |
3850 | first_offset = -frame->pretend_size; | |
526b7aee | 3851 | |
41453183 CZ |
3852 | /* IRQ using automatic save mechanism will save the register before |
3853 | anything we do. */ | |
c7314bc1 CZ |
3854 | if (ARC_AUTO_IRQ_P (fn_type) |
3855 | && !ARC_FAST_INTERRUPT_P (fn_type)) | |
41453183 | 3856 | { |
90b48013 | 3857 | frame_stack_add (first_offset); |
526b7aee | 3858 | first_offset = 0; |
90b48013 | 3859 | arc_dwarf_emit_irq_save_regs (); |
41453183 CZ |
3860 | } |
3861 | ||
90b48013 CZ |
3862 | save_blink = arc_must_save_return_addr (cfun) |
3863 | && !ARC_AUTOBLINK_IRQ_P (fn_type); | |
3864 | save_fp = arc_frame_pointer_needed () && !ARC_AUTOFP_IRQ_P (fn_type); | |
3865 | ||
3866 | /* Use enter/leave only for non-interrupt functions. */ | |
3867 | if (TARGET_CODE_DENSITY | |
3868 | && TARGET_CODE_DENSITY_FRAME | |
3869 | && !ARC_AUTOFP_IRQ_P (fn_type) | |
3870 | && !ARC_AUTOBLINK_IRQ_P (fn_type) | |
3871 | && !ARC_INTERRUPT_P (fn_type) | |
3872 | && arc_enter_leave_p (gmask)) | |
3873 | frame_size_to_allocate -= arc_save_callee_enter (gmask, save_blink, | |
3874 | save_fp, | |
3875 | first_offset); | |
3876 | else if (frame->millicode_end_reg > 14) | |
3877 | frame_size_to_allocate -= arc_save_callee_milli (gmask, save_blink, | |
3878 | save_fp, | |
3879 | first_offset, | |
3880 | frame->reg_size); | |
3881 | else | |
3882 | frame_size_to_allocate -= arc_save_callee_saves (gmask, save_blink, save_fp, | |
3883 | first_offset); | |
526b7aee | 3884 | |
526b7aee SV |
3885 | /* Allocate the stack frame. */ |
3886 | if (frame_size_to_allocate > 0) | |
1ec86e1e CZ |
3887 | frame_stack_add ((HOST_WIDE_INT) 0 - frame_size_to_allocate); |
3888 | ||
3889 | /* Emit a blockage to avoid delay slot scheduling. */ | |
3890 | emit_insn (gen_blockage ()); | |
526b7aee SV |
3891 | } |
3892 | ||
ce9dbf20 CZ |
3893 | /* Return the register number of the register holding the return address |
3894 | for a function of type TYPE. */ | |
3895 | ||
3896 | static int | |
3897 | arc_return_address_register (unsigned int fn_type) | |
3898 | { | |
3899 | int regno = 0; | |
3900 | ||
3901 | if (ARC_INTERRUPT_P (fn_type)) | |
3902 | { | |
3903 | if ((fn_type & (ARC_FUNCTION_ILINK1 | ARC_FUNCTION_FIRQ)) != 0) | |
3904 | regno = ILINK1_REG; | |
3905 | else if ((fn_type & ARC_FUNCTION_ILINK2) != 0) | |
3906 | regno = ILINK2_REG; | |
3907 | else | |
3908 | gcc_unreachable (); | |
3909 | } | |
3910 | else if (ARC_NORMAL_P (fn_type) || ARC_NAKED_P (fn_type)) | |
3911 | regno = RETURN_ADDR_REGNUM; | |
3912 | ||
3913 | gcc_assert (regno != 0); | |
3914 | return regno; | |
3915 | } | |
3916 | ||
526b7aee SV |
3917 | /* Do any necessary cleanup after a function to restore stack, frame, |
3918 | and regs. */ | |
3919 | ||
3920 | void | |
3921 | arc_expand_epilogue (int sibcall_p) | |
3922 | { | |
6fe5e235 | 3923 | int size; |
1825c61e | 3924 | unsigned int fn_type = arc_compute_function_type (cfun); |
5719867d JR |
3925 | unsigned int size_to_deallocate; |
3926 | int restored; | |
3927 | int can_trust_sp_p = !cfun->calls_alloca; | |
90b48013 CZ |
3928 | int first_offset; |
3929 | bool restore_fp = arc_frame_pointer_needed () && !ARC_AUTOFP_IRQ_P (fn_type); | |
3930 | bool restore_blink = arc_must_save_return_addr (cfun) | |
3931 | && !ARC_AUTOBLINK_IRQ_P (fn_type); | |
ce9dbf20 | 3932 | uint64_t gmask = cfun->machine->frame_info.gmask; |
90b48013 CZ |
3933 | bool return_p = !sibcall_p && fn_type == ARC_FUNCTION_NORMAL |
3934 | && !cfun->machine->frame_info.pretend_size; | |
3935 | struct arc_frame_info *frame = &cfun->machine->frame_info; | |
3936 | ||
1825c61e CZ |
3937 | /* Naked functions don't have epilogue. */ |
3938 | if (ARC_NAKED_P (fn_type)) | |
3939 | return; | |
3940 | ||
90b48013 | 3941 | size = arc_compute_frame_size (); |
5719867d | 3942 | size_to_deallocate = size; |
526b7aee | 3943 | |
90b48013 CZ |
3944 | first_offset = size - (frame->pretend_size + frame->reg_size |
3945 | + frame->extra_size); | |
526b7aee | 3946 | |
5719867d | 3947 | if (!can_trust_sp_p) |
41453183 | 3948 | gcc_assert (arc_frame_pointer_needed ()); |
526b7aee | 3949 | |
1ec86e1e CZ |
3950 | /* Emit a blockage to avoid/flush all pending sp operations. */ |
3951 | if (size) | |
3952 | emit_insn (gen_blockage ()); | |
3953 | ||
90b48013 CZ |
3954 | if (TARGET_CODE_DENSITY |
3955 | && TARGET_CODE_DENSITY_FRAME | |
3956 | && !ARC_AUTOFP_IRQ_P (fn_type) | |
3957 | && !ARC_AUTOBLINK_IRQ_P (fn_type) | |
3958 | && !ARC_INTERRUPT_P (fn_type) | |
3959 | && arc_enter_leave_p (gmask)) | |
3960 | { | |
3961 | /* Using leave instruction. */ | |
3962 | size_to_deallocate -= arc_restore_callee_leave (gmask, restore_blink, | |
3963 | restore_fp, | |
3964 | return_p, | |
3965 | first_offset); | |
3966 | if (return_p) | |
67a96300 | 3967 | { |
90b48013 CZ |
3968 | gcc_assert (size_to_deallocate == 0); |
3969 | return; | |
67a96300 | 3970 | } |
5719867d | 3971 | } |
90b48013 | 3972 | else if (frame->millicode_end_reg > 14) |
5719867d | 3973 | { |
90b48013 CZ |
3974 | /* Using millicode calls. */ |
3975 | size_to_deallocate -= arc_restore_callee_milli (gmask, restore_blink, | |
3976 | restore_fp, | |
3977 | return_p, | |
3978 | first_offset); | |
3979 | if (return_p) | |
3980 | { | |
3981 | gcc_assert (size_to_deallocate == 0); | |
3982 | return; | |
3983 | } | |
5719867d | 3984 | } |
90b48013 CZ |
3985 | else |
3986 | size_to_deallocate -= arc_restore_callee_saves (gmask, restore_blink, | |
3987 | restore_fp, | |
3988 | first_offset, | |
3989 | size_to_deallocate); | |
526b7aee | 3990 | |
90b48013 CZ |
3991 | /* Keep track of how much of the stack pointer we've restored. It |
3992 | makes the following a lot more readable. */ | |
5719867d | 3993 | restored = size - size_to_deallocate; |
526b7aee | 3994 | |
5719867d JR |
3995 | if (size > restored) |
3996 | frame_stack_add (size - restored); | |
67a96300 | 3997 | |
6fe5e235 CZ |
3998 | /* For frames that use __builtin_eh_return, the register defined by |
3999 | EH_RETURN_STACKADJ_RTX is set to 0 for all standard return paths. | |
4000 | On eh_return paths however, the register is set to the value that | |
4001 | should be added to the stack pointer in order to restore the | |
4002 | correct stack pointer for the exception handling frame. | |
4003 | ||
4004 | For ARC we are going to use r2 for EH_RETURN_STACKADJ_RTX, add | |
4005 | this onto the stack for eh_return frames. */ | |
4006 | if (crtl->calls_eh_return) | |
4007 | emit_insn (gen_add2_insn (stack_pointer_rtx, | |
4008 | EH_RETURN_STACKADJ_RTX)); | |
4009 | ||
5719867d | 4010 | /* Emit the return instruction. */ |
ce9dbf20 CZ |
4011 | if (ARC_INTERRUPT_P (fn_type)) |
4012 | { | |
4013 | rtx ra = gen_rtx_REG (Pmode, arc_return_address_register (fn_type)); | |
4014 | ||
4015 | if (TARGET_V2) | |
4016 | emit_jump_insn (gen_rtie ()); | |
4017 | else if (TARGET_ARC700) | |
4018 | emit_jump_insn (gen_rtie ()); | |
4019 | else | |
4020 | emit_jump_insn (gen_arc600_rtie (ra)); | |
4021 | } | |
4022 | else if (sibcall_p == FALSE) | |
5719867d | 4023 | emit_jump_insn (gen_simple_return ()); |
526b7aee SV |
4024 | } |
4025 | ||
90b48013 CZ |
4026 | /* Helper for {push/pop}_multi_operand: check if rtx OP is a suitable |
4027 | construct to match either enter or leave instruction. Which one | |
4028 | which is selected by PUSH_P argument. */ | |
4029 | ||
4030 | bool | |
4031 | arc_check_multi (rtx op, bool push_p) | |
4032 | { | |
4033 | HOST_WIDE_INT len = XVECLEN (op, 0); | |
4034 | unsigned int regno, i, start; | |
4035 | unsigned int memp = push_p ? 0 : 1; | |
4036 | rtx elt; | |
4037 | ||
4038 | if (len <= 1) | |
4039 | return false; | |
4040 | ||
4041 | start = 1; | |
4042 | elt = XVECEXP (op, 0, 0); | |
4043 | if (!push_p && GET_CODE (elt) == RETURN) | |
4044 | start = 2; | |
4045 | ||
4046 | for (i = start, regno = ENTER_LEAVE_START_REG; i < len; i++, regno++) | |
4047 | { | |
4048 | rtx elt = XVECEXP (op, 0, i); | |
4049 | rtx reg, mem, addr; | |
4050 | ||
4051 | if (GET_CODE (elt) != SET) | |
4052 | return false; | |
4053 | mem = XEXP (elt, memp); | |
4054 | reg = XEXP (elt, 1 - memp); | |
4055 | ||
4056 | if (!REG_P (reg) | |
4057 | || !MEM_P (mem)) | |
4058 | return false; | |
4059 | ||
4060 | /* Check for blink. */ | |
4061 | if (REGNO (reg) == RETURN_ADDR_REGNUM | |
4062 | && i == start) | |
4063 | regno = 12; | |
47d8cb23 | 4064 | else if (REGNO (reg) == HARD_FRAME_POINTER_REGNUM) |
90b48013 CZ |
4065 | ++i; |
4066 | else if (REGNO (reg) != regno) | |
4067 | return false; | |
4068 | ||
4069 | addr = XEXP (mem, 0); | |
4070 | if (GET_CODE (addr) == PLUS) | |
4071 | { | |
4072 | if (!rtx_equal_p (stack_pointer_rtx, XEXP (addr, 0)) | |
4073 | || !CONST_INT_P (XEXP (addr, 1))) | |
4074 | return false; | |
4075 | } | |
4076 | else | |
4077 | { | |
4078 | if (!rtx_equal_p (stack_pointer_rtx, addr)) | |
4079 | return false; | |
4080 | } | |
4081 | } | |
4082 | return true; | |
4083 | } | |
4084 | ||
6fe5e235 CZ |
4085 | /* Return rtx for the location of the return address on the stack, |
4086 | suitable for use in __builtin_eh_return. The new return address | |
4087 | will be written to this location in order to redirect the return to | |
3fd6ae8a CZ |
4088 | the exception handler. Our ABI says the blink is pushed first on |
4089 | stack followed by an unknown number of register saves, and finally | |
4090 | by fp. Hence we cannot use the EH_RETURN_ADDRESS macro as the | |
4091 | stack is not finalized. */ | |
526b7aee | 4092 | |
3fd6ae8a CZ |
4093 | void |
4094 | arc_eh_return_address_location (rtx source) | |
6fe5e235 CZ |
4095 | { |
4096 | rtx mem; | |
4097 | int offset; | |
4098 | struct arc_frame_info *afi; | |
4099 | ||
4100 | arc_compute_frame_size (); | |
4101 | afi = &cfun->machine->frame_info; | |
4102 | ||
4103 | gcc_assert (crtl->calls_eh_return); | |
4104 | gcc_assert (afi->save_return_addr); | |
4105 | gcc_assert (afi->extra_size >= 4); | |
4106 | ||
4107 | /* The '-4' removes the size of the return address, which is | |
4108 | included in the 'extra_size' field. */ | |
4109 | offset = afi->reg_size + afi->extra_size - 4; | |
4110 | mem = gen_frame_mem (Pmode, | |
47d8cb23 | 4111 | plus_constant (Pmode, hard_frame_pointer_rtx, offset)); |
6fe5e235 CZ |
4112 | |
4113 | /* The following should not be needed, and is, really a hack. The | |
4114 | issue being worked around here is that the DSE (Dead Store | |
4115 | Elimination) pass will remove this write to the stack as it sees | |
4116 | a single store and no corresponding read. The read however | |
4117 | occurs in the epilogue code, which is not added into the function | |
4118 | rtl until a later pass. So, at the time of DSE, the decision to | |
4119 | remove this store seems perfectly sensible. Marking the memory | |
4120 | address as volatile obviously has the effect of preventing DSE | |
4121 | from removing the store. */ | |
3fd6ae8a CZ |
4122 | MEM_VOLATILE_P (mem) = true; |
4123 | emit_move_insn (mem, source); | |
526b7aee SV |
4124 | } |
4125 | ||
4126 | /* PIC */ | |
4127 | ||
5a5c5784 CZ |
4128 | /* Helper to generate unspec constant. */ |
4129 | ||
4130 | static rtx | |
4131 | arc_unspec_offset (rtx loc, int unspec) | |
4132 | { | |
4133 | return gen_rtx_CONST (Pmode, gen_rtx_UNSPEC (Pmode, gen_rtvec (1, loc), | |
4134 | unspec)); | |
4135 | } | |
4136 | ||
526b7aee SV |
4137 | /* !TARGET_BARREL_SHIFTER support. */ |
4138 | /* Emit a shift insn to set OP0 to OP1 shifted by OP2; CODE specifies what | |
4139 | kind of shift. */ | |
4140 | ||
4141 | void | |
4142 | emit_shift (enum rtx_code code, rtx op0, rtx op1, rtx op2) | |
4143 | { | |
4144 | rtx shift = gen_rtx_fmt_ee (code, SImode, op1, op2); | |
4145 | rtx pat | |
4146 | = ((shift4_operator (shift, SImode) ? gen_shift_si3 : gen_shift_si3_loop) | |
4147 | (op0, op1, op2, shift)); | |
4148 | emit_insn (pat); | |
4149 | } | |
4150 | ||
4151 | /* Output the assembler code for doing a shift. | |
4152 | We go to a bit of trouble to generate efficient code as the ARC601 only has | |
4153 | single bit shifts. This is taken from the h8300 port. We only have one | |
4154 | mode of shifting and can't access individual bytes like the h8300 can, so | |
4155 | this is greatly simplified (at the expense of not generating hyper- | |
4156 | efficient code). | |
4157 | ||
4158 | This function is not used if the variable shift insns are present. */ | |
4159 | ||
4160 | /* FIXME: This probably can be done using a define_split in arc.md. | |
4161 | Alternately, generate rtx rather than output instructions. */ | |
4162 | ||
4163 | const char * | |
4164 | output_shift (rtx *operands) | |
4165 | { | |
4166 | /* static int loopend_lab;*/ | |
4167 | rtx shift = operands[3]; | |
ef4bddc2 | 4168 | machine_mode mode = GET_MODE (shift); |
526b7aee SV |
4169 | enum rtx_code code = GET_CODE (shift); |
4170 | const char *shift_one; | |
4171 | ||
4172 | gcc_assert (mode == SImode); | |
4173 | ||
4174 | switch (code) | |
4175 | { | |
4176 | case ASHIFT: shift_one = "add %0,%1,%1"; break; | |
4177 | case ASHIFTRT: shift_one = "asr %0,%1"; break; | |
4178 | case LSHIFTRT: shift_one = "lsr %0,%1"; break; | |
4179 | default: gcc_unreachable (); | |
4180 | } | |
4181 | ||
4182 | if (GET_CODE (operands[2]) != CONST_INT) | |
4183 | { | |
4184 | output_asm_insn ("and.f lp_count,%2, 0x1f", operands); | |
4185 | goto shiftloop; | |
4186 | } | |
4187 | else | |
4188 | { | |
4189 | int n; | |
4190 | ||
4191 | n = INTVAL (operands[2]); | |
4192 | ||
4193 | /* Only consider the lower 5 bits of the shift count. */ | |
4194 | n = n & 0x1f; | |
4195 | ||
4196 | /* First see if we can do them inline. */ | |
4197 | /* ??? We could get better scheduling & shorter code (using short insns) | |
4198 | by using splitters. Alas, that'd be even more verbose. */ | |
4199 | if (code == ASHIFT && n <= 9 && n > 2 | |
4200 | && dest_reg_operand (operands[4], SImode)) | |
4201 | { | |
4202 | output_asm_insn ("mov %4,0\n\tadd3 %0,%4,%1", operands); | |
4203 | for (n -=3 ; n >= 3; n -= 3) | |
4204 | output_asm_insn ("add3 %0,%4,%0", operands); | |
4205 | if (n == 2) | |
4206 | output_asm_insn ("add2 %0,%4,%0", operands); | |
4207 | else if (n) | |
4208 | output_asm_insn ("add %0,%0,%0", operands); | |
4209 | } | |
4210 | else if (n <= 4) | |
4211 | { | |
4212 | while (--n >= 0) | |
4213 | { | |
4214 | output_asm_insn (shift_one, operands); | |
4215 | operands[1] = operands[0]; | |
4216 | } | |
4217 | } | |
4218 | /* See if we can use a rotate/and. */ | |
4219 | else if (n == BITS_PER_WORD - 1) | |
4220 | { | |
4221 | switch (code) | |
4222 | { | |
4223 | case ASHIFT : | |
4224 | output_asm_insn ("and %0,%1,1\n\tror %0,%0", operands); | |
4225 | break; | |
4226 | case ASHIFTRT : | |
4227 | /* The ARC doesn't have a rol insn. Use something else. */ | |
4228 | output_asm_insn ("add.f 0,%1,%1\n\tsbc %0,%0,%0", operands); | |
4229 | break; | |
4230 | case LSHIFTRT : | |
4231 | /* The ARC doesn't have a rol insn. Use something else. */ | |
4232 | output_asm_insn ("add.f 0,%1,%1\n\trlc %0,0", operands); | |
4233 | break; | |
4234 | default: | |
4235 | break; | |
4236 | } | |
4237 | } | |
4238 | else if (n == BITS_PER_WORD - 2 && dest_reg_operand (operands[4], SImode)) | |
4239 | { | |
4240 | switch (code) | |
4241 | { | |
4242 | case ASHIFT : | |
4243 | output_asm_insn ("and %0,%1,3\n\tror %0,%0\n\tror %0,%0", operands); | |
4244 | break; | |
4245 | case ASHIFTRT : | |
4246 | #if 1 /* Need some scheduling comparisons. */ | |
4247 | output_asm_insn ("add.f %4,%1,%1\n\tsbc %0,%0,%0\n\t" | |
4248 | "add.f 0,%4,%4\n\trlc %0,%0", operands); | |
4249 | #else | |
4250 | output_asm_insn ("add.f %4,%1,%1\n\tbxor %0,%4,31\n\t" | |
4251 | "sbc.f %0,%0,%4\n\trlc %0,%0", operands); | |
4252 | #endif | |
4253 | break; | |
4254 | case LSHIFTRT : | |
4255 | #if 1 | |
4256 | output_asm_insn ("add.f %4,%1,%1\n\trlc %0,0\n\t" | |
4257 | "add.f 0,%4,%4\n\trlc %0,%0", operands); | |
4258 | #else | |
4259 | output_asm_insn ("add.f %0,%1,%1\n\trlc.f %0,0\n\t" | |
4260 | "and %0,%0,1\n\trlc %0,%0", operands); | |
4261 | #endif | |
4262 | break; | |
4263 | default: | |
4264 | break; | |
4265 | } | |
4266 | } | |
4267 | else if (n == BITS_PER_WORD - 3 && code == ASHIFT) | |
4268 | output_asm_insn ("and %0,%1,7\n\tror %0,%0\n\tror %0,%0\n\tror %0,%0", | |
4269 | operands); | |
4270 | /* Must loop. */ | |
4271 | else | |
4272 | { | |
4273 | operands[2] = GEN_INT (n); | |
4274 | output_asm_insn ("mov.f lp_count, %2", operands); | |
4275 | ||
4276 | shiftloop: | |
4277 | { | |
4278 | output_asm_insn ("lpnz\t2f", operands); | |
4279 | output_asm_insn (shift_one, operands); | |
4280 | output_asm_insn ("nop", operands); | |
4281 | fprintf (asm_out_file, "2:\t%s end single insn loop\n", | |
4282 | ASM_COMMENT_START); | |
4283 | } | |
4284 | } | |
4285 | } | |
4286 | ||
4287 | return ""; | |
4288 | } | |
4289 | \f | |
4290 | /* Nested function support. */ | |
4291 | ||
efcc2e30 CZ |
4292 | /* Output assembler code for a block containing the constant parts of |
4293 | a trampoline, leaving space for variable parts. A trampoline looks | |
4294 | like this: | |
4295 | ||
4296 | ld_s r12,[pcl,8] | |
4297 | ld r11,[pcl,12] | |
4298 | j_s [r12] | |
4299 | .word function's address | |
4300 | .word static chain value | |
4301 | ||
4302 | */ | |
526b7aee SV |
4303 | |
4304 | static void | |
efcc2e30 | 4305 | arc_asm_trampoline_template (FILE *f) |
526b7aee | 4306 | { |
efcc2e30 CZ |
4307 | asm_fprintf (f, "\tld_s\t%s,[pcl,8]\n", ARC_TEMP_SCRATCH_REG); |
4308 | asm_fprintf (f, "\tld\t%s,[pcl,12]\n", reg_names[STATIC_CHAIN_REGNUM]); | |
4309 | asm_fprintf (f, "\tj_s\t[%s]\n", ARC_TEMP_SCRATCH_REG); | |
4310 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
4311 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
526b7aee SV |
4312 | } |
4313 | ||
4314 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
efcc2e30 CZ |
4315 | FNADDR is an RTX for the address of the function's pure code. CXT |
4316 | is an RTX for the static chain value for the function. | |
526b7aee SV |
4317 | |
4318 | The fastest trampoline to execute for trampolines within +-8KB of CTX | |
4319 | would be: | |
efcc2e30 | 4320 | |
526b7aee SV |
4321 | add2 r11,pcl,s12 |
4322 | j [limm] 0x20200f80 limm | |
efcc2e30 CZ |
4323 | |
4324 | and that would also be faster to write to the stack by computing | |
4325 | the offset from CTX to TRAMP at compile time. However, it would | |
4326 | really be better to get rid of the high cost of cache invalidation | |
4327 | when generating trampolines, which requires that the code part of | |
4328 | trampolines stays constant, and additionally either making sure | |
4329 | that no executable code but trampolines is on the stack, no icache | |
4330 | entries linger for the area of the stack from when before the stack | |
4331 | was allocated, and allocating trampolines in trampoline-only cache | |
4332 | lines or allocate trampolines fram a special pool of pre-allocated | |
4333 | trampolines. */ | |
526b7aee SV |
4334 | |
4335 | static void | |
4336 | arc_initialize_trampoline (rtx tramp, tree fndecl, rtx cxt) | |
4337 | { | |
4338 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); | |
4339 | ||
efcc2e30 CZ |
4340 | emit_block_move (tramp, assemble_trampoline_template (), |
4341 | GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL); | |
4342 | emit_move_insn (adjust_address (tramp, SImode, 8), fnaddr); | |
4343 | emit_move_insn (adjust_address (tramp, SImode, 12), cxt); | |
4344 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
4345 | LCT_NORMAL, VOIDmode, XEXP (tramp, 0), Pmode, | |
4346 | plus_constant (Pmode, XEXP (tramp, 0), TRAMPOLINE_SIZE), | |
4347 | Pmode); | |
526b7aee SV |
4348 | } |
4349 | ||
7778a1ad CZ |
4350 | /* Add the given function declaration to emit code in JLI section. */ |
4351 | ||
4352 | static void | |
4353 | arc_add_jli_section (rtx pat) | |
4354 | { | |
4355 | const char *name; | |
4356 | tree attrs; | |
4357 | arc_jli_section *sec = arc_jli_sections, *new_section; | |
4358 | tree decl = SYMBOL_REF_DECL (pat); | |
4359 | ||
4360 | if (!pat) | |
4361 | return; | |
4362 | ||
4363 | if (decl) | |
4364 | { | |
4365 | /* For fixed locations do not generate the jli table entry. It | |
4366 | should be provided by the user as an asm file. */ | |
4367 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); | |
4368 | if (lookup_attribute ("jli_fixed", attrs)) | |
4369 | return; | |
4370 | } | |
4371 | ||
4372 | name = XSTR (pat, 0); | |
4373 | ||
4374 | /* Don't insert the same symbol twice. */ | |
4375 | while (sec != NULL) | |
4376 | { | |
4377 | if(strcmp (name, sec->name) == 0) | |
4378 | return; | |
4379 | sec = sec->next; | |
4380 | } | |
4381 | ||
4382 | /* New name, insert it. */ | |
4383 | new_section = (arc_jli_section *) xmalloc (sizeof (arc_jli_section)); | |
4384 | gcc_assert (new_section != NULL); | |
4385 | new_section->name = name; | |
4386 | new_section->next = arc_jli_sections; | |
4387 | arc_jli_sections = new_section; | |
4388 | } | |
4389 | ||
526b7aee SV |
4390 | /* This is set briefly to 1 when we output a ".as" address modifer, and then |
4391 | reset when we output the scaled address. */ | |
4392 | static int output_scaled = 0; | |
4393 | ||
e0be3321 CZ |
4394 | /* Set when we force sdata output. */ |
4395 | static int output_sdata = 0; | |
4396 | ||
526b7aee SV |
4397 | /* Print operand X (an rtx) in assembler syntax to file FILE. |
4398 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
4399 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
4400 | /* In final.c:output_asm_insn: | |
4401 | 'l' : label | |
4402 | 'a' : address | |
4403 | 'c' : constant address if CONSTANT_ADDRESS_P | |
4404 | 'n' : negative | |
4405 | Here: | |
4406 | 'Z': log2(x+1)-1 | |
4407 | 'z': log2 | |
4408 | 'M': log2(~x) | |
ceaaa9fe JR |
4409 | 'p': bit Position of lsb |
4410 | 's': size of bit field | |
526b7aee SV |
4411 | '#': condbranch delay slot suffix |
4412 | '*': jump delay slot suffix | |
4413 | '?' : nonjump-insn suffix for conditional execution or short instruction | |
4414 | '!' : jump / call suffix for conditional execution or short instruction | |
4415 | '`': fold constant inside unary o-perator, re-recognize, and emit. | |
4416 | 'd' | |
4417 | 'D' | |
4418 | 'R': Second word | |
6b55f8c9 | 4419 | 'S': JLI instruction |
7778a1ad | 4420 | 'j': used by mov instruction to properly emit jli related labels. |
526b7aee SV |
4421 | 'B': Branch comparison operand - suppress sda reference |
4422 | 'H': Most significant word | |
4423 | 'L': Least significant word | |
4424 | 'A': ASCII decimal representation of floating point value | |
4425 | 'U': Load/store update or scaling indicator | |
4426 | 'V': cache bypass indicator for volatile | |
4427 | 'P' | |
4428 | 'F' | |
4429 | '^' | |
4430 | 'O': Operator | |
4431 | 'o': original symbol - no @ prepending. */ | |
4432 | ||
4433 | void | |
4434 | arc_print_operand (FILE *file, rtx x, int code) | |
4435 | { | |
4436 | switch (code) | |
4437 | { | |
4438 | case 'Z': | |
4439 | if (GET_CODE (x) == CONST_INT) | |
4440 | fprintf (file, "%d",exact_log2(INTVAL (x) + 1) - 1 ); | |
4441 | else | |
4442 | output_operand_lossage ("invalid operand to %%Z code"); | |
4443 | ||
4444 | return; | |
4445 | ||
4446 | case 'z': | |
4447 | if (GET_CODE (x) == CONST_INT) | |
03301dcc | 4448 | fprintf (file, "%d",exact_log2 (INTVAL (x) & 0xffffffff)); |
526b7aee SV |
4449 | else |
4450 | output_operand_lossage ("invalid operand to %%z code"); | |
4451 | ||
4452 | return; | |
4453 | ||
1e466f04 GM |
4454 | case 'c': |
4455 | if (GET_CODE (x) == CONST_INT) | |
41bc2c0b | 4456 | fprintf (file, "%ld", INTVAL (x) ); |
1e466f04 GM |
4457 | else |
4458 | output_operand_lossage ("invalid operands to %%c code"); | |
4459 | ||
4460 | return; | |
4461 | ||
526b7aee SV |
4462 | case 'M': |
4463 | if (GET_CODE (x) == CONST_INT) | |
4464 | fprintf (file, "%d",exact_log2(~INTVAL (x)) ); | |
4465 | else | |
4466 | output_operand_lossage ("invalid operand to %%M code"); | |
4467 | ||
4468 | return; | |
4469 | ||
ceaaa9fe JR |
4470 | case 'p': |
4471 | if (GET_CODE (x) == CONST_INT) | |
4472 | fprintf (file, "%d", exact_log2 (INTVAL (x) & -INTVAL (x))); | |
4473 | else | |
4474 | output_operand_lossage ("invalid operand to %%p code"); | |
4475 | return; | |
4476 | ||
4477 | case 's': | |
4478 | if (GET_CODE (x) == CONST_INT) | |
4479 | { | |
4480 | HOST_WIDE_INT i = INTVAL (x); | |
4481 | HOST_WIDE_INT s = exact_log2 (i & -i); | |
4482 | fprintf (file, "%d", exact_log2 (((0xffffffffUL & i) >> s) + 1)); | |
4483 | } | |
4484 | else | |
4485 | output_operand_lossage ("invalid operand to %%s code"); | |
4486 | return; | |
4487 | ||
526b7aee SV |
4488 | case '#' : |
4489 | /* Conditional branches depending on condition codes. | |
4490 | Note that this is only for branches that were known to depend on | |
4491 | condition codes before delay slot scheduling; | |
4492 | out-of-range brcc / bbit expansions should use '*'. | |
4493 | This distinction is important because of the different | |
4494 | allowable delay slot insns and the output of the delay suffix | |
4495 | for TARGET_AT_DBR_COND_EXEC. */ | |
4496 | case '*' : | |
4497 | /* Unconditional branches / branches not depending on condition codes. | |
4498 | This could also be a CALL_INSN. | |
4499 | Output the appropriate delay slot suffix. */ | |
84034c69 | 4500 | if (final_sequence && final_sequence->len () != 1) |
526b7aee | 4501 | { |
84034c69 DM |
4502 | rtx_insn *jump = final_sequence->insn (0); |
4503 | rtx_insn *delay = final_sequence->insn (1); | |
526b7aee SV |
4504 | |
4505 | /* For TARGET_PAD_RETURN we might have grabbed the delay insn. */ | |
4654c0cf | 4506 | if (delay->deleted ()) |
526b7aee SV |
4507 | return; |
4508 | if (JUMP_P (jump) && INSN_ANNULLED_BRANCH_P (jump)) | |
4509 | fputs (INSN_FROM_TARGET_P (delay) ? ".d" | |
4510 | : TARGET_AT_DBR_CONDEXEC && code == '#' ? ".d" | |
4511 | : get_attr_type (jump) == TYPE_RETURN && code == '#' ? "" | |
4512 | : ".nd", | |
4513 | file); | |
4514 | else | |
4515 | fputs (".d", file); | |
4516 | } | |
4517 | return; | |
4518 | case '?' : /* with leading "." */ | |
4519 | case '!' : /* without leading "." */ | |
4520 | /* This insn can be conditionally executed. See if the ccfsm machinery | |
4521 | says it should be conditionalized. | |
4522 | If it shouldn't, we'll check the compact attribute if this insn | |
4523 | has a short variant, which may be used depending on code size and | |
4524 | alignment considerations. */ | |
4525 | if (current_insn_predicate) | |
4526 | arc_ccfsm_current.cc | |
4527 | = get_arc_condition_code (current_insn_predicate); | |
4528 | if (ARC_CCFSM_COND_EXEC_P (&arc_ccfsm_current)) | |
4529 | { | |
4530 | /* Is this insn in a delay slot sequence? */ | |
4531 | if (!final_sequence || XVECLEN (final_sequence, 0) < 2 | |
4532 | || current_insn_predicate | |
68a1a6c0 DM |
4533 | || CALL_P (final_sequence->insn (0)) |
4534 | || simplejump_p (final_sequence->insn (0))) | |
526b7aee SV |
4535 | { |
4536 | /* This insn isn't in a delay slot sequence, or conditionalized | |
4537 | independently of its position in a delay slot. */ | |
4538 | fprintf (file, "%s%s", | |
4539 | code == '?' ? "." : "", | |
4540 | arc_condition_codes[arc_ccfsm_current.cc]); | |
4541 | /* If this is a jump, there are still short variants. However, | |
4542 | only beq_s / bne_s have the same offset range as b_s, | |
4543 | and the only short conditional returns are jeq_s and jne_s. */ | |
4544 | if (code == '!' | |
4545 | && (arc_ccfsm_current.cc == ARC_CC_EQ | |
4546 | || arc_ccfsm_current.cc == ARC_CC_NE | |
4547 | || 0 /* FIXME: check if branch in 7 bit range. */)) | |
4548 | output_short_suffix (file); | |
4549 | } | |
4550 | else if (code == '!') /* Jump with delay slot. */ | |
4551 | fputs (arc_condition_codes[arc_ccfsm_current.cc], file); | |
4552 | else /* An Instruction in a delay slot of a jump or call. */ | |
4553 | { | |
4554 | rtx jump = XVECEXP (final_sequence, 0, 0); | |
4555 | rtx insn = XVECEXP (final_sequence, 0, 1); | |
4556 | ||
4557 | /* If the insn is annulled and is from the target path, we need | |
4558 | to inverse the condition test. */ | |
4559 | if (JUMP_P (jump) && INSN_ANNULLED_BRANCH_P (jump)) | |
4560 | { | |
4561 | if (INSN_FROM_TARGET_P (insn)) | |
4562 | fprintf (file, "%s%s", | |
4563 | code == '?' ? "." : "", | |
4564 | arc_condition_codes[ARC_INVERSE_CONDITION_CODE (arc_ccfsm_current.cc)]); | |
4565 | else | |
4566 | fprintf (file, "%s%s", | |
4567 | code == '?' ? "." : "", | |
4568 | arc_condition_codes[arc_ccfsm_current.cc]); | |
4569 | if (arc_ccfsm_current.state == 5) | |
4570 | arc_ccfsm_current.state = 0; | |
4571 | } | |
4572 | else | |
4573 | /* This insn is executed for either path, so don't | |
4574 | conditionalize it at all. */ | |
4575 | output_short_suffix (file); | |
4576 | ||
4577 | } | |
4578 | } | |
4579 | else | |
4580 | output_short_suffix (file); | |
4581 | return; | |
4582 | case'`': | |
4583 | /* FIXME: fold constant inside unary operator, re-recognize, and emit. */ | |
4584 | gcc_unreachable (); | |
4585 | case 'd' : | |
4586 | fputs (arc_condition_codes[get_arc_condition_code (x)], file); | |
4587 | return; | |
4588 | case 'D' : | |
4589 | fputs (arc_condition_codes[ARC_INVERSE_CONDITION_CODE | |
4590 | (get_arc_condition_code (x))], | |
4591 | file); | |
4592 | return; | |
4593 | case 'R' : | |
4594 | /* Write second word of DImode or DFmode reference, | |
4595 | register or memory. */ | |
4596 | if (GET_CODE (x) == REG) | |
4597 | fputs (reg_names[REGNO (x)+1], file); | |
4598 | else if (GET_CODE (x) == MEM) | |
4599 | { | |
4600 | fputc ('[', file); | |
4601 | ||
4602 | /* Handle possible auto-increment. For PRE_INC / PRE_DEC / | |
4603 | PRE_MODIFY, we will have handled the first word already; | |
4604 | For POST_INC / POST_DEC / POST_MODIFY, the access to the | |
4605 | first word will be done later. In either case, the access | |
4606 | to the first word will do the modify, and we only have | |
4607 | to add an offset of four here. */ | |
4608 | if (GET_CODE (XEXP (x, 0)) == PRE_INC | |
4609 | || GET_CODE (XEXP (x, 0)) == PRE_DEC | |
4610 | || GET_CODE (XEXP (x, 0)) == PRE_MODIFY | |
4611 | || GET_CODE (XEXP (x, 0)) == POST_INC | |
4612 | || GET_CODE (XEXP (x, 0)) == POST_DEC | |
4613 | || GET_CODE (XEXP (x, 0)) == POST_MODIFY) | |
cc8ca59e JB |
4614 | output_address (VOIDmode, |
4615 | plus_constant (Pmode, XEXP (XEXP (x, 0), 0), 4)); | |
526b7aee SV |
4616 | else if (output_scaled) |
4617 | { | |
4618 | rtx addr = XEXP (x, 0); | |
4619 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
4620 | ||
cc8ca59e JB |
4621 | output_address (VOIDmode, |
4622 | plus_constant (Pmode, XEXP (addr, 0), | |
526b7aee SV |
4623 | ((INTVAL (XEXP (addr, 1)) + 4) |
4624 | >> (size == 2 ? 1 : 2)))); | |
4625 | output_scaled = 0; | |
4626 | } | |
4627 | else | |
cc8ca59e JB |
4628 | output_address (VOIDmode, |
4629 | plus_constant (Pmode, XEXP (x, 0), 4)); | |
526b7aee SV |
4630 | fputc (']', file); |
4631 | } | |
4632 | else | |
4633 | output_operand_lossage ("invalid operand to %%R code"); | |
4634 | return; | |
7778a1ad | 4635 | case 'j': |
526b7aee | 4636 | case 'S' : |
6b55f8c9 CZ |
4637 | if (GET_CODE (x) == SYMBOL_REF |
4638 | && arc_is_jli_call_p (x)) | |
4639 | { | |
4640 | if (SYMBOL_REF_DECL (x)) | |
4641 | { | |
4642 | tree attrs = (TREE_TYPE (SYMBOL_REF_DECL (x)) != error_mark_node | |
4643 | ? TYPE_ATTRIBUTES (TREE_TYPE (SYMBOL_REF_DECL (x))) | |
4644 | : NULL_TREE); | |
4645 | if (lookup_attribute ("jli_fixed", attrs)) | |
4646 | { | |
7778a1ad CZ |
4647 | /* No special treatment for jli_fixed functions. */ |
4648 | if (code == 'j') | |
4649 | break; | |
6b55f8c9 CZ |
4650 | fprintf (file, "%ld\t; @", |
4651 | TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attrs)))); | |
4652 | assemble_name (file, XSTR (x, 0)); | |
4653 | return; | |
4654 | } | |
4655 | } | |
4656 | fprintf (file, "@__jli."); | |
4657 | assemble_name (file, XSTR (x, 0)); | |
7778a1ad CZ |
4658 | if (code == 'j') |
4659 | arc_add_jli_section (x); | |
4660 | return; | |
4661 | } | |
4662 | if (GET_CODE (x) == SYMBOL_REF | |
4663 | && arc_is_secure_call_p (x)) | |
4664 | { | |
4665 | /* No special treatment for secure functions. */ | |
4666 | if (code == 'j' ) | |
4667 | break; | |
4668 | tree attrs = (TREE_TYPE (SYMBOL_REF_DECL (x)) != error_mark_node | |
4669 | ? TYPE_ATTRIBUTES (TREE_TYPE (SYMBOL_REF_DECL (x))) | |
4670 | : NULL_TREE); | |
4671 | fprintf (file, "%ld\t; @", | |
4672 | TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attrs)))); | |
4673 | assemble_name (file, XSTR (x, 0)); | |
6b55f8c9 CZ |
4674 | return; |
4675 | } | |
4676 | break; | |
526b7aee SV |
4677 | case 'B' /* Branch or other LIMM ref - must not use sda references. */ : |
4678 | if (CONSTANT_P (x)) | |
4679 | { | |
4680 | output_addr_const (file, x); | |
4681 | return; | |
4682 | } | |
4683 | break; | |
4684 | case 'H' : | |
4685 | case 'L' : | |
4686 | if (GET_CODE (x) == REG) | |
4687 | { | |
4688 | /* L = least significant word, H = most significant word. */ | |
4689 | if ((WORDS_BIG_ENDIAN != 0) ^ (code == 'L')) | |
4690 | fputs (reg_names[REGNO (x)], file); | |
4691 | else | |
4692 | fputs (reg_names[REGNO (x)+1], file); | |
4693 | } | |
4694 | else if (GET_CODE (x) == CONST_INT | |
4695 | || GET_CODE (x) == CONST_DOUBLE) | |
4696 | { | |
8ad9df62 | 4697 | rtx first, second, word; |
526b7aee SV |
4698 | |
4699 | split_double (x, &first, &second); | |
4700 | ||
4701 | if((WORDS_BIG_ENDIAN) == 0) | |
8ad9df62 | 4702 | word = (code == 'L' ? first : second); |
526b7aee | 4703 | else |
8ad9df62 | 4704 | word = (code == 'L' ? second : first); |
526b7aee | 4705 | |
8ad9df62 JR |
4706 | fprintf (file, "0x%08" PRIx32, ((uint32_t) INTVAL (word))); |
4707 | } | |
526b7aee SV |
4708 | else |
4709 | output_operand_lossage ("invalid operand to %%H/%%L code"); | |
4710 | return; | |
4711 | case 'A' : | |
4712 | { | |
4713 | char str[30]; | |
4714 | ||
4715 | gcc_assert (GET_CODE (x) == CONST_DOUBLE | |
4716 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT); | |
4717 | ||
4718 | real_to_decimal (str, CONST_DOUBLE_REAL_VALUE (x), sizeof (str), 0, 1); | |
4719 | fprintf (file, "%s", str); | |
4720 | return; | |
4721 | } | |
4722 | case 'U' : | |
4723 | /* Output a load/store with update indicator if appropriate. */ | |
4724 | if (GET_CODE (x) == MEM) | |
4725 | { | |
4726 | rtx addr = XEXP (x, 0); | |
4727 | switch (GET_CODE (addr)) | |
4728 | { | |
4729 | case PRE_INC: case PRE_DEC: case PRE_MODIFY: | |
4730 | fputs (".a", file); break; | |
4731 | case POST_INC: case POST_DEC: case POST_MODIFY: | |
4732 | fputs (".ab", file); break; | |
4733 | case PLUS: | |
4734 | /* Are we using a scaled index? */ | |
4735 | if (GET_CODE (XEXP (addr, 0)) == MULT) | |
4736 | fputs (".as", file); | |
4737 | /* Can we use a scaled offset? */ | |
4738 | else if (CONST_INT_P (XEXP (addr, 1)) | |
4739 | && GET_MODE_SIZE (GET_MODE (x)) > 1 | |
4740 | && (!(INTVAL (XEXP (addr, 1)) | |
4741 | & (GET_MODE_SIZE (GET_MODE (x)) - 1) & 3)) | |
4742 | /* Does it make a difference? */ | |
4743 | && !SMALL_INT_RANGE(INTVAL (XEXP (addr, 1)), | |
4744 | GET_MODE_SIZE (GET_MODE (x)) - 2, 0)) | |
4745 | { | |
4746 | fputs (".as", file); | |
4747 | output_scaled = 1; | |
4748 | } | |
e0be3321 CZ |
4749 | break; |
4750 | case SYMBOL_REF: | |
4751 | case CONST: | |
4752 | if (legitimate_small_data_address_p (addr) | |
4753 | && GET_MODE_SIZE (GET_MODE (x)) > 1) | |
b6fb7933 | 4754 | { |
e0be3321 CZ |
4755 | int align = get_symbol_alignment (addr); |
4756 | int mask = 0; | |
4757 | switch (GET_MODE (x)) | |
4758 | { | |
4759 | case E_HImode: | |
4760 | mask = 1; | |
4761 | break; | |
4762 | default: | |
4763 | mask = 3; | |
4764 | break; | |
4765 | } | |
4766 | if (align && ((align & mask) == 0)) | |
b6fb7933 CZ |
4767 | fputs (".as", file); |
4768 | } | |
526b7aee SV |
4769 | break; |
4770 | case REG: | |
4771 | break; | |
4772 | default: | |
4773 | gcc_assert (CONSTANT_P (addr)); break; | |
4774 | } | |
4775 | } | |
4776 | else | |
4777 | output_operand_lossage ("invalid operand to %%U code"); | |
4778 | return; | |
4779 | case 'V' : | |
4780 | /* Output cache bypass indicator for a load/store insn. Volatile memory | |
4781 | refs are defined to use the cache bypass mechanism. */ | |
4782 | if (GET_CODE (x) == MEM) | |
4783 | { | |
8180c03f CZ |
4784 | if ((MEM_VOLATILE_P (x) && !TARGET_VOLATILE_CACHE_SET) |
4785 | || arc_is_uncached_mem_p (x)) | |
526b7aee SV |
4786 | fputs (".di", file); |
4787 | } | |
4788 | else | |
4789 | output_operand_lossage ("invalid operand to %%V code"); | |
4790 | return; | |
4791 | /* plt code. */ | |
4792 | case 'P': | |
4793 | case 0 : | |
4794 | /* Do nothing special. */ | |
4795 | break; | |
4796 | case 'F': | |
4797 | fputs (reg_names[REGNO (x)]+1, file); | |
4798 | return; | |
4799 | case '^': | |
4800 | /* This punctuation character is needed because label references are | |
4801 | printed in the output template using %l. This is a front end | |
4802 | character, and when we want to emit a '@' before it, we have to use | |
4803 | this '^'. */ | |
4804 | ||
4805 | fputc('@',file); | |
4806 | return; | |
4807 | case 'O': | |
4808 | /* Output an operator. */ | |
4809 | switch (GET_CODE (x)) | |
4810 | { | |
4811 | case PLUS: fputs ("add", file); return; | |
4812 | case SS_PLUS: fputs ("adds", file); return; | |
4813 | case AND: fputs ("and", file); return; | |
4814 | case IOR: fputs ("or", file); return; | |
4815 | case XOR: fputs ("xor", file); return; | |
4816 | case MINUS: fputs ("sub", file); return; | |
4817 | case SS_MINUS: fputs ("subs", file); return; | |
4818 | case ASHIFT: fputs ("asl", file); return; | |
4819 | case ASHIFTRT: fputs ("asr", file); return; | |
4820 | case LSHIFTRT: fputs ("lsr", file); return; | |
4821 | case ROTATERT: fputs ("ror", file); return; | |
4822 | case MULT: fputs ("mpy", file); return; | |
4823 | case ABS: fputs ("abs", file); return; /* Unconditional. */ | |
4824 | case NEG: fputs ("neg", file); return; | |
4825 | case SS_NEG: fputs ("negs", file); return; | |
4826 | case NOT: fputs ("not", file); return; /* Unconditional. */ | |
4827 | case ZERO_EXTEND: | |
4828 | fputs ("ext", file); /* bmsk allows predication. */ | |
4829 | goto size_suffix; | |
4830 | case SIGN_EXTEND: /* Unconditional. */ | |
4831 | fputs ("sex", file); | |
4832 | size_suffix: | |
4833 | switch (GET_MODE (XEXP (x, 0))) | |
4834 | { | |
4e10a5a7 RS |
4835 | case E_QImode: fputs ("b", file); return; |
4836 | case E_HImode: fputs ("w", file); return; | |
526b7aee SV |
4837 | default: break; |
4838 | } | |
4839 | break; | |
4840 | case SS_TRUNCATE: | |
4841 | if (GET_MODE (x) != HImode) | |
4842 | break; | |
4843 | fputs ("sat16", file); | |
4844 | default: break; | |
4845 | } | |
4846 | output_operand_lossage ("invalid operand to %%O code"); return; | |
4847 | case 'o': | |
4848 | if (GET_CODE (x) == SYMBOL_REF) | |
4849 | { | |
4850 | assemble_name (file, XSTR (x, 0)); | |
4851 | return; | |
4852 | } | |
4853 | break; | |
4854 | case '&': | |
16493b57 | 4855 | if (TARGET_ANNOTATE_ALIGN) |
526b7aee SV |
4856 | fprintf (file, "; unalign: %d", cfun->machine->unalign); |
4857 | return; | |
f50bb868 CZ |
4858 | case '+': |
4859 | if (TARGET_V2) | |
4860 | fputs ("m", file); | |
4861 | else | |
4862 | fputs ("h", file); | |
4863 | return; | |
4864 | case '_': | |
4865 | if (TARGET_V2) | |
4866 | fputs ("h", file); | |
4867 | else | |
4868 | fputs ("w", file); | |
4869 | return; | |
526b7aee SV |
4870 | default : |
4871 | /* Unknown flag. */ | |
4872 | output_operand_lossage ("invalid operand output code"); | |
4873 | } | |
4874 | ||
4875 | switch (GET_CODE (x)) | |
4876 | { | |
4877 | case REG : | |
4878 | fputs (reg_names[REGNO (x)], file); | |
4879 | break; | |
4880 | case MEM : | |
4881 | { | |
4882 | rtx addr = XEXP (x, 0); | |
4883 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
4884 | ||
e0be3321 CZ |
4885 | if (legitimate_small_data_address_p (addr)) |
4886 | output_sdata = 1; | |
4887 | ||
526b7aee SV |
4888 | fputc ('[', file); |
4889 | ||
4890 | switch (GET_CODE (addr)) | |
4891 | { | |
4892 | case PRE_INC: case POST_INC: | |
cc8ca59e JB |
4893 | output_address (VOIDmode, |
4894 | plus_constant (Pmode, XEXP (addr, 0), size)); break; | |
526b7aee | 4895 | case PRE_DEC: case POST_DEC: |
cc8ca59e JB |
4896 | output_address (VOIDmode, |
4897 | plus_constant (Pmode, XEXP (addr, 0), -size)); | |
526b7aee SV |
4898 | break; |
4899 | case PRE_MODIFY: case POST_MODIFY: | |
cc8ca59e | 4900 | output_address (VOIDmode, XEXP (addr, 1)); break; |
526b7aee SV |
4901 | case PLUS: |
4902 | if (output_scaled) | |
4903 | { | |
cc8ca59e JB |
4904 | output_address (VOIDmode, |
4905 | plus_constant (Pmode, XEXP (addr, 0), | |
526b7aee SV |
4906 | (INTVAL (XEXP (addr, 1)) |
4907 | >> (size == 2 ? 1 : 2)))); | |
4908 | output_scaled = 0; | |
4909 | } | |
4910 | else | |
cc8ca59e | 4911 | output_address (VOIDmode, addr); |
526b7aee SV |
4912 | break; |
4913 | default: | |
4914 | if (flag_pic && CONSTANT_ADDRESS_P (addr)) | |
4915 | arc_output_pic_addr_const (file, addr, code); | |
4916 | else | |
cc8ca59e | 4917 | output_address (VOIDmode, addr); |
526b7aee SV |
4918 | break; |
4919 | } | |
4920 | fputc (']', file); | |
4921 | break; | |
4922 | } | |
4923 | case CONST_DOUBLE : | |
4924 | /* We handle SFmode constants here as output_addr_const doesn't. */ | |
4925 | if (GET_MODE (x) == SFmode) | |
4926 | { | |
526b7aee SV |
4927 | long l; |
4928 | ||
34a72c33 | 4929 | REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (x), l); |
526b7aee SV |
4930 | fprintf (file, "0x%08lx", l); |
4931 | break; | |
4932 | } | |
3bbe0b82 JL |
4933 | /* FALLTHRU */ |
4934 | /* Let output_addr_const deal with it. */ | |
526b7aee | 4935 | default : |
28633bbd CZ |
4936 | if (flag_pic |
4937 | || (GET_CODE (x) == CONST | |
4938 | && GET_CODE (XEXP (x, 0)) == UNSPEC | |
4939 | && (XINT (XEXP (x, 0), 1) == UNSPEC_TLS_OFF | |
4940 | || XINT (XEXP (x, 0), 1) == UNSPEC_TLS_GD)) | |
4941 | || (GET_CODE (x) == CONST | |
4942 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4943 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == UNSPEC | |
4944 | && (XINT (XEXP (XEXP (x, 0), 0), 1) == UNSPEC_TLS_OFF | |
4945 | || XINT (XEXP (XEXP (x, 0), 0), 1) == UNSPEC_TLS_GD))) | |
526b7aee SV |
4946 | arc_output_pic_addr_const (file, x, code); |
4947 | else | |
e0be3321 | 4948 | output_addr_const (file, x); |
526b7aee SV |
4949 | break; |
4950 | } | |
4951 | } | |
4952 | ||
4953 | /* Print a memory address as an operand to reference that memory location. */ | |
4954 | ||
4955 | void | |
4956 | arc_print_operand_address (FILE *file , rtx addr) | |
4957 | { | |
4958 | register rtx base, index = 0; | |
4959 | ||
4960 | switch (GET_CODE (addr)) | |
4961 | { | |
4962 | case REG : | |
4963 | fputs (reg_names[REGNO (addr)], file); | |
4964 | break; | |
e0be3321 CZ |
4965 | case SYMBOL_REF: |
4966 | if (output_sdata) | |
4967 | fputs ("gp,", file); | |
526b7aee | 4968 | output_addr_const (file, addr); |
e0be3321 CZ |
4969 | if (output_sdata) |
4970 | fputs ("@sda", file); | |
4971 | output_sdata = 0; | |
526b7aee SV |
4972 | break; |
4973 | case PLUS : | |
4974 | if (GET_CODE (XEXP (addr, 0)) == MULT) | |
4975 | index = XEXP (XEXP (addr, 0), 0), base = XEXP (addr, 1); | |
4976 | else if (CONST_INT_P (XEXP (addr, 0))) | |
4977 | index = XEXP (addr, 0), base = XEXP (addr, 1); | |
4978 | else | |
4979 | base = XEXP (addr, 0), index = XEXP (addr, 1); | |
4980 | ||
4981 | gcc_assert (OBJECT_P (base)); | |
4982 | arc_print_operand_address (file, base); | |
4983 | if (CONSTANT_P (base) && CONST_INT_P (index)) | |
4984 | fputc ('+', file); | |
4985 | else | |
4986 | fputc (',', file); | |
4987 | gcc_assert (OBJECT_P (index)); | |
4988 | arc_print_operand_address (file, index); | |
4989 | break; | |
4990 | case CONST: | |
4991 | { | |
4992 | rtx c = XEXP (addr, 0); | |
4993 | ||
28633bbd CZ |
4994 | if ((GET_CODE (c) == UNSPEC |
4995 | && (XINT (c, 1) == UNSPEC_TLS_OFF | |
4996 | || XINT (c, 1) == UNSPEC_TLS_IE)) | |
4997 | || (GET_CODE (c) == PLUS | |
4998 | && GET_CODE (XEXP (c, 0)) == UNSPEC | |
f5e336b1 CZ |
4999 | && (XINT (XEXP (c, 0), 1) == UNSPEC_TLS_OFF |
5000 | || XINT (XEXP (c, 0), 1) == ARC_UNSPEC_GOTOFFPC))) | |
28633bbd CZ |
5001 | { |
5002 | arc_output_pic_addr_const (file, c, 0); | |
5003 | break; | |
5004 | } | |
5005 | gcc_assert (GET_CODE (c) == PLUS); | |
526b7aee SV |
5006 | gcc_assert (GET_CODE (XEXP (c, 0)) == SYMBOL_REF); |
5007 | gcc_assert (GET_CODE (XEXP (c, 1)) == CONST_INT); | |
5008 | ||
cc8ca59e | 5009 | output_address (VOIDmode, XEXP (addr, 0)); |
526b7aee SV |
5010 | |
5011 | break; | |
5012 | } | |
5013 | case PRE_INC : | |
5014 | case PRE_DEC : | |
5015 | /* We shouldn't get here as we've lost the mode of the memory object | |
5016 | (which says how much to inc/dec by. */ | |
5017 | gcc_unreachable (); | |
5018 | break; | |
5019 | default : | |
5020 | if (flag_pic) | |
5021 | arc_output_pic_addr_const (file, addr, 0); | |
5022 | else | |
5023 | output_addr_const (file, addr); | |
5024 | break; | |
5025 | } | |
5026 | } | |
5027 | ||
526b7aee SV |
5028 | /* Conditional execution support. |
5029 | ||
5030 | This is based on the ARM port but for now is much simpler. | |
5031 | ||
5032 | A finite state machine takes care of noticing whether or not instructions | |
5033 | can be conditionally executed, and thus decrease execution time and code | |
5034 | size by deleting branch instructions. The fsm is controlled by | |
5035 | arc_ccfsm_advance (called by arc_final_prescan_insn), and controls the | |
5036 | actions of PRINT_OPERAND. The patterns in the .md file for the branch | |
5037 | insns also have a hand in this. */ | |
5038 | /* The way we leave dealing with non-anulled or annull-false delay slot | |
5039 | insns to the consumer is awkward. */ | |
5040 | ||
5041 | /* The state of the fsm controlling condition codes are: | |
5042 | 0: normal, do nothing special | |
5043 | 1: don't output this insn | |
5044 | 2: don't output this insn | |
5045 | 3: make insns conditional | |
5046 | 4: make insns conditional | |
5047 | 5: make insn conditional (only for outputting anulled delay slot insns) | |
5048 | ||
5049 | special value for cfun->machine->uid_ccfsm_state: | |
5050 | 6: return with but one insn before it since function start / call | |
5051 | ||
5052 | State transitions (state->state by whom, under what condition): | |
5053 | 0 -> 1 arc_ccfsm_advance, if insn is a conditional branch skipping over | |
5054 | some instructions. | |
5055 | 0 -> 2 arc_ccfsm_advance, if insn is a conditional branch followed | |
5056 | by zero or more non-jump insns and an unconditional branch with | |
5057 | the same target label as the condbranch. | |
5058 | 1 -> 3 branch patterns, after having not output the conditional branch | |
5059 | 2 -> 4 branch patterns, after having not output the conditional branch | |
5060 | 0 -> 5 branch patterns, for anulled delay slot insn. | |
5061 | 3 -> 0 ASM_OUTPUT_INTERNAL_LABEL, if the `target' label is reached | |
5062 | (the target label has CODE_LABEL_NUMBER equal to | |
5063 | arc_ccfsm_target_label). | |
5064 | 4 -> 0 arc_ccfsm_advance, if `target' unconditional branch is reached | |
5065 | 3 -> 1 arc_ccfsm_advance, finding an 'else' jump skipping over some insns. | |
5066 | 5 -> 0 when outputting the delay slot insn | |
5067 | ||
5068 | If the jump clobbers the conditions then we use states 2 and 4. | |
5069 | ||
5070 | A similar thing can be done with conditional return insns. | |
5071 | ||
5072 | We also handle separating branches from sets of the condition code. | |
5073 | This is done here because knowledge of the ccfsm state is required, | |
5074 | we may not be outputting the branch. */ | |
5075 | ||
5076 | /* arc_final_prescan_insn calls arc_ccfsm_advance to adjust arc_ccfsm_current, | |
5077 | before letting final output INSN. */ | |
5078 | ||
5079 | static void | |
b3458f61 | 5080 | arc_ccfsm_advance (rtx_insn *insn, struct arc_ccfsm *state) |
526b7aee SV |
5081 | { |
5082 | /* BODY will hold the body of INSN. */ | |
5083 | register rtx body; | |
5084 | ||
5085 | /* This will be 1 if trying to repeat the trick (ie: do the `else' part of | |
5086 | an if/then/else), and things need to be reversed. */ | |
5087 | int reverse = 0; | |
5088 | ||
5089 | /* If we start with a return insn, we only succeed if we find another one. */ | |
5090 | int seeking_return = 0; | |
5091 | ||
5092 | /* START_INSN will hold the insn from where we start looking. This is the | |
5093 | first insn after the following code_label if REVERSE is true. */ | |
b3458f61 | 5094 | rtx_insn *start_insn = insn; |
526b7aee SV |
5095 | |
5096 | /* Type of the jump_insn. Brcc insns don't affect ccfsm changes, | |
5097 | since they don't rely on a cmp preceding the. */ | |
5098 | enum attr_type jump_insn_type; | |
5099 | ||
5100 | /* Allow -mdebug-ccfsm to turn this off so we can see how well it does. | |
5101 | We can't do this in macro FINAL_PRESCAN_INSN because its called from | |
5102 | final_scan_insn which has `optimize' as a local. */ | |
5103 | if (optimize < 2 || TARGET_NO_COND_EXEC) | |
5104 | return; | |
5105 | ||
5106 | /* Ignore notes and labels. */ | |
5107 | if (!INSN_P (insn)) | |
5108 | return; | |
5109 | body = PATTERN (insn); | |
5110 | /* If in state 4, check if the target branch is reached, in order to | |
5111 | change back to state 0. */ | |
5112 | if (state->state == 4) | |
5113 | { | |
5114 | if (insn == state->target_insn) | |
5115 | { | |
5116 | state->target_insn = NULL; | |
5117 | state->state = 0; | |
5118 | } | |
5119 | return; | |
5120 | } | |
5121 | ||
5122 | /* If in state 3, it is possible to repeat the trick, if this insn is an | |
5123 | unconditional branch to a label, and immediately following this branch | |
5124 | is the previous target label which is only used once, and the label this | |
5125 | branch jumps to is not too far off. Or in other words "we've done the | |
5126 | `then' part, see if we can do the `else' part." */ | |
5127 | if (state->state == 3) | |
5128 | { | |
5129 | if (simplejump_p (insn)) | |
5130 | { | |
5131 | start_insn = next_nonnote_insn (start_insn); | |
5132 | if (GET_CODE (start_insn) == BARRIER) | |
5133 | { | |
5134 | /* ??? Isn't this always a barrier? */ | |
5135 | start_insn = next_nonnote_insn (start_insn); | |
5136 | } | |
5137 | if (GET_CODE (start_insn) == CODE_LABEL | |
5138 | && CODE_LABEL_NUMBER (start_insn) == state->target_label | |
5139 | && LABEL_NUSES (start_insn) == 1) | |
5140 | reverse = TRUE; | |
5141 | else | |
5142 | return; | |
5143 | } | |
5144 | else if (GET_CODE (body) == SIMPLE_RETURN) | |
5145 | { | |
5146 | start_insn = next_nonnote_insn (start_insn); | |
5147 | if (GET_CODE (start_insn) == BARRIER) | |
5148 | start_insn = next_nonnote_insn (start_insn); | |
5149 | if (GET_CODE (start_insn) == CODE_LABEL | |
5150 | && CODE_LABEL_NUMBER (start_insn) == state->target_label | |
5151 | && LABEL_NUSES (start_insn) == 1) | |
5152 | { | |
5153 | reverse = TRUE; | |
5154 | seeking_return = 1; | |
5155 | } | |
5156 | else | |
5157 | return; | |
5158 | } | |
5159 | else | |
5160 | return; | |
5161 | } | |
5162 | ||
5163 | if (GET_CODE (insn) != JUMP_INSN | |
5164 | || GET_CODE (PATTERN (insn)) == ADDR_VEC | |
5165 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
5166 | return; | |
5167 | ||
5168 | /* We can't predicate BRCC or loop ends. | |
5169 | Also, when generating PIC code, and considering a medium range call, | |
5170 | we can't predicate the call. */ | |
5171 | jump_insn_type = get_attr_type (insn); | |
5172 | if (jump_insn_type == TYPE_BRCC | |
5173 | || jump_insn_type == TYPE_BRCC_NO_DELAY_SLOT | |
5174 | || jump_insn_type == TYPE_LOOP_END | |
5175 | || (jump_insn_type == TYPE_CALL && !get_attr_predicable (insn))) | |
5176 | return; | |
5177 | ||
5178 | /* This jump might be paralleled with a clobber of the condition codes, | |
5179 | the jump should always come first. */ | |
5180 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
5181 | body = XVECEXP (body, 0, 0); | |
5182 | ||
5183 | if (reverse | |
5184 | || (GET_CODE (body) == SET && GET_CODE (SET_DEST (body)) == PC | |
5185 | && GET_CODE (SET_SRC (body)) == IF_THEN_ELSE)) | |
5186 | { | |
5187 | int insns_skipped = 0, fail = FALSE, succeed = FALSE; | |
5188 | /* Flag which part of the IF_THEN_ELSE is the LABEL_REF. */ | |
5189 | int then_not_else = TRUE; | |
5190 | /* Nonzero if next insn must be the target label. */ | |
5191 | int next_must_be_target_label_p; | |
b3458f61 DM |
5192 | rtx_insn *this_insn = start_insn; |
5193 | rtx label = 0; | |
526b7aee SV |
5194 | |
5195 | /* Register the insn jumped to. */ | |
5196 | if (reverse) | |
5197 | { | |
5198 | if (!seeking_return) | |
5199 | label = XEXP (SET_SRC (body), 0); | |
5200 | } | |
5201 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == LABEL_REF) | |
5202 | label = XEXP (XEXP (SET_SRC (body), 1), 0); | |
5203 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == LABEL_REF) | |
5204 | { | |
5205 | label = XEXP (XEXP (SET_SRC (body), 2), 0); | |
5206 | then_not_else = FALSE; | |
5207 | } | |
5208 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == SIMPLE_RETURN) | |
5209 | seeking_return = 1; | |
5210 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == SIMPLE_RETURN) | |
5211 | { | |
5212 | seeking_return = 1; | |
5213 | then_not_else = FALSE; | |
5214 | } | |
5215 | else | |
5216 | gcc_unreachable (); | |
5217 | ||
5218 | /* If this is a non-annulled branch with a delay slot, there is | |
5219 | no need to conditionalize the delay slot. */ | |
782bdf21 | 5220 | if ((GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) == SEQUENCE) |
526b7aee SV |
5221 | && state->state == 0 && !INSN_ANNULLED_BRANCH_P (insn)) |
5222 | { | |
5223 | this_insn = NEXT_INSN (this_insn); | |
526b7aee SV |
5224 | } |
5225 | /* See how many insns this branch skips, and what kind of insns. If all | |
5226 | insns are okay, and the label or unconditional branch to the same | |
5227 | label is not too far away, succeed. */ | |
5228 | for (insns_skipped = 0, next_must_be_target_label_p = FALSE; | |
5229 | !fail && !succeed && insns_skipped < MAX_INSNS_SKIPPED; | |
5230 | insns_skipped++) | |
5231 | { | |
5232 | rtx scanbody; | |
5233 | ||
5234 | this_insn = next_nonnote_insn (this_insn); | |
5235 | if (!this_insn) | |
5236 | break; | |
5237 | ||
5238 | if (next_must_be_target_label_p) | |
5239 | { | |
5240 | if (GET_CODE (this_insn) == BARRIER) | |
5241 | continue; | |
5242 | if (GET_CODE (this_insn) == CODE_LABEL | |
5243 | && this_insn == label) | |
5244 | { | |
5245 | state->state = 1; | |
5246 | succeed = TRUE; | |
5247 | } | |
5248 | else | |
5249 | fail = TRUE; | |
5250 | break; | |
5251 | } | |
5252 | ||
526b7aee SV |
5253 | switch (GET_CODE (this_insn)) |
5254 | { | |
5255 | case CODE_LABEL: | |
5256 | /* Succeed if it is the target label, otherwise fail since | |
5257 | control falls in from somewhere else. */ | |
5258 | if (this_insn == label) | |
5259 | { | |
5260 | state->state = 1; | |
5261 | succeed = TRUE; | |
5262 | } | |
5263 | else | |
5264 | fail = TRUE; | |
5265 | break; | |
5266 | ||
5267 | case BARRIER: | |
5268 | /* Succeed if the following insn is the target label. | |
5269 | Otherwise fail. | |
5270 | If return insns are used then the last insn in a function | |
5271 | will be a barrier. */ | |
5272 | next_must_be_target_label_p = TRUE; | |
5273 | break; | |
5274 | ||
5275 | case CALL_INSN: | |
5276 | /* Can handle a call insn if there are no insns after it. | |
5277 | IE: The next "insn" is the target label. We don't have to | |
5278 | worry about delay slots as such insns are SEQUENCE's inside | |
5279 | INSN's. ??? It is possible to handle such insns though. */ | |
5280 | if (get_attr_cond (this_insn) == COND_CANUSE) | |
5281 | next_must_be_target_label_p = TRUE; | |
5282 | else | |
5283 | fail = TRUE; | |
5284 | break; | |
5285 | ||
5286 | case JUMP_INSN: | |
4173ddaf SB |
5287 | scanbody = PATTERN (this_insn); |
5288 | ||
526b7aee SV |
5289 | /* If this is an unconditional branch to the same label, succeed. |
5290 | If it is to another label, do nothing. If it is conditional, | |
5291 | fail. */ | |
5292 | /* ??? Probably, the test for the SET and the PC are | |
5293 | unnecessary. */ | |
5294 | ||
5295 | if (GET_CODE (scanbody) == SET | |
5296 | && GET_CODE (SET_DEST (scanbody)) == PC) | |
5297 | { | |
5298 | if (GET_CODE (SET_SRC (scanbody)) == LABEL_REF | |
5299 | && XEXP (SET_SRC (scanbody), 0) == label && !reverse) | |
5300 | { | |
5301 | state->state = 2; | |
5302 | succeed = TRUE; | |
5303 | } | |
5304 | else if (GET_CODE (SET_SRC (scanbody)) == IF_THEN_ELSE) | |
5305 | fail = TRUE; | |
5306 | else if (get_attr_cond (this_insn) != COND_CANUSE) | |
5307 | fail = TRUE; | |
5308 | } | |
5309 | else if (GET_CODE (scanbody) == SIMPLE_RETURN | |
5310 | && seeking_return) | |
5311 | { | |
5312 | state->state = 2; | |
5313 | succeed = TRUE; | |
5314 | } | |
5315 | else if (GET_CODE (scanbody) == PARALLEL) | |
5316 | { | |
5317 | if (get_attr_cond (this_insn) != COND_CANUSE) | |
5318 | fail = TRUE; | |
5319 | } | |
5320 | break; | |
5321 | ||
5322 | case INSN: | |
4173ddaf SB |
5323 | scanbody = PATTERN (this_insn); |
5324 | ||
526b7aee SV |
5325 | /* We can only do this with insns that can use the condition |
5326 | codes (and don't set them). */ | |
5327 | if (GET_CODE (scanbody) == SET | |
5328 | || GET_CODE (scanbody) == PARALLEL) | |
5329 | { | |
5330 | if (get_attr_cond (this_insn) != COND_CANUSE) | |
5331 | fail = TRUE; | |
5332 | } | |
5333 | /* We can't handle other insns like sequences. */ | |
5334 | else | |
5335 | fail = TRUE; | |
5336 | break; | |
5337 | ||
5338 | default: | |
5339 | break; | |
5340 | } | |
5341 | } | |
5342 | ||
5343 | if (succeed) | |
5344 | { | |
5345 | if ((!seeking_return) && (state->state == 1 || reverse)) | |
5346 | state->target_label = CODE_LABEL_NUMBER (label); | |
5347 | else if (seeking_return || state->state == 2) | |
5348 | { | |
5349 | while (this_insn && GET_CODE (PATTERN (this_insn)) == USE) | |
5350 | { | |
5351 | this_insn = next_nonnote_insn (this_insn); | |
5352 | ||
5353 | gcc_assert (!this_insn || | |
5354 | (GET_CODE (this_insn) != BARRIER | |
5355 | && GET_CODE (this_insn) != CODE_LABEL)); | |
5356 | } | |
5357 | if (!this_insn) | |
5358 | { | |
5359 | /* Oh dear! we ran off the end, give up. */ | |
5360 | extract_insn_cached (insn); | |
5361 | state->state = 0; | |
5362 | state->target_insn = NULL; | |
5363 | return; | |
5364 | } | |
5365 | state->target_insn = this_insn; | |
5366 | } | |
5367 | else | |
5368 | gcc_unreachable (); | |
5369 | ||
5370 | /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from | |
5371 | what it was. */ | |
5372 | if (!reverse) | |
5373 | { | |
5374 | state->cond = XEXP (SET_SRC (body), 0); | |
5375 | state->cc = get_arc_condition_code (XEXP (SET_SRC (body), 0)); | |
5376 | } | |
5377 | ||
5378 | if (reverse || then_not_else) | |
5379 | state->cc = ARC_INVERSE_CONDITION_CODE (state->cc); | |
5380 | } | |
5381 | ||
5382 | /* Restore recog_operand. Getting the attributes of other insns can | |
5383 | destroy this array, but final.c assumes that it remains intact | |
5384 | across this call; since the insn has been recognized already we | |
5385 | call insn_extract direct. */ | |
5386 | extract_insn_cached (insn); | |
5387 | } | |
5388 | } | |
5389 | ||
5390 | /* Record that we are currently outputting label NUM with prefix PREFIX. | |
5391 | It it's the label we're looking for, reset the ccfsm machinery. | |
5392 | ||
5393 | Called from ASM_OUTPUT_INTERNAL_LABEL. */ | |
5394 | ||
5395 | static void | |
5396 | arc_ccfsm_at_label (const char *prefix, int num, struct arc_ccfsm *state) | |
5397 | { | |
5398 | if (state->state == 3 && state->target_label == num | |
5399 | && !strcmp (prefix, "L")) | |
5400 | { | |
5401 | state->state = 0; | |
b3458f61 | 5402 | state->target_insn = NULL; |
526b7aee SV |
5403 | } |
5404 | } | |
5405 | ||
5406 | /* We are considering a conditional branch with the condition COND. | |
5407 | Check if we want to conditionalize a delay slot insn, and if so modify | |
5408 | the ccfsm state accordingly. | |
5409 | REVERSE says branch will branch when the condition is false. */ | |
5410 | void | |
b32d5189 | 5411 | arc_ccfsm_record_condition (rtx cond, bool reverse, rtx_insn *jump, |
526b7aee SV |
5412 | struct arc_ccfsm *state) |
5413 | { | |
b3458f61 | 5414 | rtx_insn *seq_insn = NEXT_INSN (PREV_INSN (jump)); |
526b7aee SV |
5415 | if (!state) |
5416 | state = &arc_ccfsm_current; | |
5417 | ||
5418 | gcc_assert (state->state == 0); | |
5419 | if (seq_insn != jump) | |
5420 | { | |
5421 | rtx insn = XVECEXP (PATTERN (seq_insn), 0, 1); | |
5422 | ||
4654c0cf | 5423 | if (!as_a<rtx_insn *> (insn)->deleted () |
526b7aee SV |
5424 | && INSN_ANNULLED_BRANCH_P (jump) |
5425 | && (TARGET_AT_DBR_CONDEXEC || INSN_FROM_TARGET_P (insn))) | |
5426 | { | |
5427 | state->cond = cond; | |
5428 | state->cc = get_arc_condition_code (cond); | |
5429 | if (!reverse) | |
5430 | arc_ccfsm_current.cc | |
5431 | = ARC_INVERSE_CONDITION_CODE (state->cc); | |
5432 | rtx pat = PATTERN (insn); | |
5433 | if (GET_CODE (pat) == COND_EXEC) | |
5434 | gcc_assert ((INSN_FROM_TARGET_P (insn) | |
5435 | ? ARC_INVERSE_CONDITION_CODE (state->cc) : state->cc) | |
5436 | == get_arc_condition_code (XEXP (pat, 0))); | |
5437 | else | |
5438 | state->state = 5; | |
5439 | } | |
5440 | } | |
5441 | } | |
5442 | ||
5443 | /* Update *STATE as we would when we emit INSN. */ | |
5444 | ||
5445 | static void | |
b3458f61 | 5446 | arc_ccfsm_post_advance (rtx_insn *insn, struct arc_ccfsm *state) |
526b7aee | 5447 | { |
53ea364f JR |
5448 | enum attr_type type; |
5449 | ||
526b7aee SV |
5450 | if (LABEL_P (insn)) |
5451 | arc_ccfsm_at_label ("L", CODE_LABEL_NUMBER (insn), state); | |
5452 | else if (JUMP_P (insn) | |
5453 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
5454 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
53ea364f | 5455 | && ((type = get_attr_type (insn)) == TYPE_BRANCH |
6c28e6ae CZ |
5456 | || ((type == TYPE_UNCOND_BRANCH |
5457 | || type == TYPE_RETURN) | |
53ea364f | 5458 | && ARC_CCFSM_BRANCH_DELETED_P (state)))) |
526b7aee SV |
5459 | { |
5460 | if (ARC_CCFSM_BRANCH_DELETED_P (state)) | |
5461 | ARC_CCFSM_RECORD_BRANCH_DELETED (state); | |
5462 | else | |
5463 | { | |
5464 | rtx src = SET_SRC (PATTERN (insn)); | |
5465 | arc_ccfsm_record_condition (XEXP (src, 0), XEXP (src, 1) == pc_rtx, | |
5466 | insn, state); | |
5467 | } | |
5468 | } | |
5469 | else if (arc_ccfsm_current.state == 5) | |
5470 | arc_ccfsm_current.state = 0; | |
5471 | } | |
5472 | ||
5473 | /* Return true if the current insn, which is a conditional branch, is to be | |
5474 | deleted. */ | |
5475 | ||
5476 | bool | |
5477 | arc_ccfsm_branch_deleted_p (void) | |
5478 | { | |
5479 | return ARC_CCFSM_BRANCH_DELETED_P (&arc_ccfsm_current); | |
5480 | } | |
5481 | ||
5482 | /* Record a branch isn't output because subsequent insns can be | |
5483 | conditionalized. */ | |
5484 | ||
5485 | void | |
5486 | arc_ccfsm_record_branch_deleted (void) | |
5487 | { | |
5488 | ARC_CCFSM_RECORD_BRANCH_DELETED (&arc_ccfsm_current); | |
5489 | } | |
5490 | ||
5491 | /* During insn output, indicate if the current insn is predicated. */ | |
5492 | ||
5493 | bool | |
5494 | arc_ccfsm_cond_exec_p (void) | |
5495 | { | |
5496 | return (cfun->machine->prescan_initialized | |
5497 | && ARC_CCFSM_COND_EXEC_P (&arc_ccfsm_current)); | |
5498 | } | |
5499 | ||
526b7aee SV |
5500 | /* When deciding if an insn should be output short, we want to know something |
5501 | about the following insns: | |
5502 | - if another insn follows which we know we can output as a short insn | |
5503 | before an alignment-sensitive point, we can output this insn short: | |
5504 | the decision about the eventual alignment can be postponed. | |
5505 | - if a to-be-aligned label comes next, we should output this insn such | |
5506 | as to get / preserve 4-byte alignment. | |
5507 | - if a likely branch without delay slot insn, or a call with an immediately | |
5508 | following short insn comes next, we should out output this insn such as to | |
5509 | get / preserve 2 mod 4 unalignment. | |
5510 | - do the same for a not completely unlikely branch with a short insn | |
5511 | following before any other branch / label. | |
5512 | - in order to decide if we are actually looking at a branch, we need to | |
5513 | call arc_ccfsm_advance. | |
5514 | - in order to decide if we are looking at a short insn, we should know | |
5515 | if it is conditionalized. To a first order of approximation this is | |
5516 | the case if the state from arc_ccfsm_advance from before this insn | |
5517 | indicates the insn is conditionalized. However, a further refinement | |
5518 | could be to not conditionalize an insn if the destination register(s) | |
5519 | is/are dead in the non-executed case. */ | |
5520 | /* Return non-zero if INSN should be output as a short insn. UNALIGN is | |
5521 | zero if the current insn is aligned to a 4-byte-boundary, two otherwise. | |
5522 | If CHECK_ATTR is greater than 0, check the iscompact attribute first. */ | |
5523 | ||
b51addd6 | 5524 | static int |
b3458f61 | 5525 | arc_verify_short (rtx_insn *insn, int, int check_attr) |
526b7aee SV |
5526 | { |
5527 | enum attr_iscompact iscompact; | |
526b7aee SV |
5528 | |
5529 | if (check_attr > 0) | |
5530 | { | |
5531 | iscompact = get_attr_iscompact (insn); | |
5532 | if (iscompact == ISCOMPACT_FALSE) | |
5533 | return 0; | |
5534 | } | |
526b7aee SV |
5535 | |
5536 | return (get_attr_length (insn) & 2) != 0; | |
5537 | } | |
5538 | ||
5539 | /* When outputting an instruction (alternative) that can potentially be short, | |
5540 | output the short suffix if the insn is in fact short, and update | |
5541 | cfun->machine->unalign accordingly. */ | |
5542 | ||
5543 | static void | |
5544 | output_short_suffix (FILE *file) | |
5545 | { | |
b3458f61 | 5546 | rtx_insn *insn = current_output_insn; |
526b7aee SV |
5547 | |
5548 | if (arc_verify_short (insn, cfun->machine->unalign, 1)) | |
5549 | { | |
5550 | fprintf (file, "_s"); | |
5551 | cfun->machine->unalign ^= 2; | |
5552 | } | |
5553 | /* Restore recog_operand. */ | |
5554 | extract_insn_cached (insn); | |
5555 | } | |
5556 | ||
5557 | /* Implement FINAL_PRESCAN_INSN. */ | |
5558 | ||
5559 | void | |
b3458f61 | 5560 | arc_final_prescan_insn (rtx_insn *insn, rtx *opvec ATTRIBUTE_UNUSED, |
526b7aee SV |
5561 | int noperands ATTRIBUTE_UNUSED) |
5562 | { | |
5563 | if (TARGET_DUMPISIZE) | |
5564 | fprintf (asm_out_file, "\n; at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); | |
5565 | ||
526b7aee SV |
5566 | if (!cfun->machine->prescan_initialized) |
5567 | { | |
5568 | /* Clear lingering state from branch shortening. */ | |
5569 | memset (&arc_ccfsm_current, 0, sizeof arc_ccfsm_current); | |
5570 | cfun->machine->prescan_initialized = 1; | |
5571 | } | |
5572 | arc_ccfsm_advance (insn, &arc_ccfsm_current); | |
526b7aee SV |
5573 | } |
5574 | ||
5575 | /* Given FROM and TO register numbers, say whether this elimination is allowed. | |
5576 | Frame pointer elimination is automatically handled. | |
5577 | ||
5578 | All eliminations are permissible. If we need a frame | |
5579 | pointer, we must eliminate ARG_POINTER_REGNUM into | |
5580 | FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
5581 | ||
5582 | static bool | |
5583 | arc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to) | |
5584 | { | |
47d8cb23 | 5585 | return ((to == HARD_FRAME_POINTER_REGNUM) || (to == STACK_POINTER_REGNUM)); |
526b7aee SV |
5586 | } |
5587 | ||
5588 | /* Define the offset between two registers, one to be eliminated, and | |
5589 | the other its replacement, at the start of a routine. */ | |
5590 | ||
5591 | int | |
5592 | arc_initial_elimination_offset (int from, int to) | |
5593 | { | |
6fe5e235 CZ |
5594 | if (!cfun->machine->frame_info.initialized) |
5595 | arc_compute_frame_size (); | |
526b7aee | 5596 | |
47d8cb23 | 5597 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
526b7aee SV |
5598 | { |
5599 | return (cfun->machine->frame_info.extra_size | |
5600 | + cfun->machine->frame_info.reg_size); | |
5601 | } | |
5602 | ||
5603 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
5604 | { | |
5605 | return (cfun->machine->frame_info.total_size | |
5606 | - cfun->machine->frame_info.pretend_size); | |
5607 | } | |
5608 | ||
5609 | if ((from == FRAME_POINTER_REGNUM) && (to == STACK_POINTER_REGNUM)) | |
5610 | { | |
5611 | return (cfun->machine->frame_info.total_size | |
5612 | - (cfun->machine->frame_info.pretend_size | |
5613 | + cfun->machine->frame_info.extra_size | |
5614 | + cfun->machine->frame_info.reg_size)); | |
5615 | } | |
47d8cb23 CZ |
5616 | if ((from == FRAME_POINTER_REGNUM) && (to == HARD_FRAME_POINTER_REGNUM)) |
5617 | return 0; | |
526b7aee SV |
5618 | |
5619 | gcc_unreachable (); | |
5620 | } | |
5621 | ||
5622 | static bool | |
5623 | arc_frame_pointer_required (void) | |
5624 | { | |
6fe5e235 | 5625 | return cfun->calls_alloca || crtl->calls_eh_return; |
526b7aee SV |
5626 | } |
5627 | ||
5628 | ||
5629 | /* Return the destination address of a branch. */ | |
5630 | ||
b51addd6 | 5631 | static int |
526b7aee SV |
5632 | branch_dest (rtx branch) |
5633 | { | |
5634 | rtx pat = PATTERN (branch); | |
5635 | rtx dest = (GET_CODE (pat) == PARALLEL | |
5636 | ? SET_SRC (XVECEXP (pat, 0, 0)) : SET_SRC (pat)); | |
5637 | int dest_uid; | |
5638 | ||
5639 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
5640 | dest = XEXP (dest, XEXP (dest, 1) == pc_rtx ? 2 : 1); | |
5641 | ||
5642 | dest = XEXP (dest, 0); | |
5643 | dest_uid = INSN_UID (dest); | |
5644 | ||
5645 | return INSN_ADDRESSES (dest_uid); | |
5646 | } | |
5647 | ||
5648 | ||
5719867d | 5649 | /* Implement TARGET_ENCODE_SECTION_INFO hook. */ |
526b7aee SV |
5650 | |
5651 | static void | |
5652 | arc_encode_section_info (tree decl, rtx rtl, int first) | |
5653 | { | |
5654 | /* For sdata, SYMBOL_FLAG_LOCAL and SYMBOL_FLAG_FUNCTION. | |
5655 | This clears machine specific flags, so has to come first. */ | |
5656 | default_encode_section_info (decl, rtl, first); | |
5657 | ||
5658 | /* Check if it is a function, and whether it has the | |
5659 | [long/medium/short]_call attribute specified. */ | |
5660 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
5661 | { | |
5662 | rtx symbol = XEXP (rtl, 0); | |
5663 | int flags = SYMBOL_REF_FLAGS (symbol); | |
5664 | ||
5665 | tree attr = (TREE_TYPE (decl) != error_mark_node | |
5666 | ? TYPE_ATTRIBUTES (TREE_TYPE (decl)) : NULL_TREE); | |
5667 | tree long_call_attr = lookup_attribute ("long_call", attr); | |
5668 | tree medium_call_attr = lookup_attribute ("medium_call", attr); | |
5669 | tree short_call_attr = lookup_attribute ("short_call", attr); | |
5670 | ||
5671 | if (long_call_attr != NULL_TREE) | |
5672 | flags |= SYMBOL_FLAG_LONG_CALL; | |
5673 | else if (medium_call_attr != NULL_TREE) | |
5674 | flags |= SYMBOL_FLAG_MEDIUM_CALL; | |
5675 | else if (short_call_attr != NULL_TREE) | |
5676 | flags |= SYMBOL_FLAG_SHORT_CALL; | |
5677 | ||
5678 | SYMBOL_REF_FLAGS (symbol) = flags; | |
5679 | } | |
4d03dc2f JR |
5680 | else if (TREE_CODE (decl) == VAR_DECL) |
5681 | { | |
5682 | rtx symbol = XEXP (rtl, 0); | |
5683 | ||
5684 | tree attr = (TREE_TYPE (decl) != error_mark_node | |
5685 | ? DECL_ATTRIBUTES (decl) : NULL_TREE); | |
5686 | ||
5687 | tree sec_attr = lookup_attribute ("section", attr); | |
5688 | if (sec_attr) | |
5689 | { | |
5690 | const char *sec_name | |
5691 | = TREE_STRING_POINTER (TREE_VALUE (TREE_VALUE (sec_attr))); | |
5692 | if (strcmp (sec_name, ".cmem") == 0 | |
5693 | || strcmp (sec_name, ".cmem_shared") == 0 | |
5694 | || strcmp (sec_name, ".cmem_private") == 0) | |
5695 | SYMBOL_REF_FLAGS (symbol) |= SYMBOL_FLAG_CMEM; | |
5696 | } | |
5697 | } | |
526b7aee SV |
5698 | } |
5699 | ||
5700 | /* This is how to output a definition of an internal numbered label where | |
5701 | PREFIX is the class of label and NUM is the number within the class. */ | |
5702 | ||
5703 | static void arc_internal_label (FILE *stream, const char *prefix, unsigned long labelno) | |
5704 | { | |
5705 | if (cfun) | |
5706 | arc_ccfsm_at_label (prefix, labelno, &arc_ccfsm_current); | |
5707 | default_internal_label (stream, prefix, labelno); | |
5708 | } | |
5709 | ||
5710 | /* Set the cpu type and print out other fancy things, | |
5711 | at the top of the file. */ | |
5712 | ||
5713 | static void arc_file_start (void) | |
5714 | { | |
5715 | default_file_start (); | |
5716 | fprintf (asm_out_file, "\t.cpu %s\n", arc_cpu_string); | |
048c6a9a CZ |
5717 | |
5718 | /* Set some want to have build attributes. */ | |
5719 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_PCS_config, %d\n", | |
5720 | ATTRIBUTE_PCS); | |
5721 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_ABI_rf16, %d\n", | |
5722 | TARGET_RF16 ? 1 : 0); | |
5723 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_ABI_pic, %d\n", | |
5724 | flag_pic ? 2 : 0); | |
5725 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_ABI_tls, %d\n", | |
5726 | (arc_tp_regno != -1) ? 1 : 0); | |
5727 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_ABI_sda, %d\n", | |
5728 | TARGET_NO_SDATA_SET ? 0 : 2); | |
5729 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_ABI_exceptions, %d\n", | |
5730 | TARGET_OPTFPE ? 1 : 0); | |
62f26645 CZ |
5731 | if (TARGET_V2) |
5732 | asm_fprintf (asm_out_file, "\t.arc_attribute Tag_ARC_CPU_variation, %d\n", | |
dd1fd744 VG |
5733 | (arc_tune < ARC_TUNE_CORE_3) ? 2 : |
5734 | (arc_tune == ARC_TUNE_CORE_3 ? 3 : 4)); | |
526b7aee SV |
5735 | } |
5736 | ||
6b55f8c9 CZ |
5737 | /* Implement `TARGET_ASM_FILE_END'. */ |
5738 | /* Outputs to the stdio stream FILE jli related text. */ | |
5739 | ||
5740 | void arc_file_end (void) | |
5741 | { | |
5742 | arc_jli_section *sec = arc_jli_sections; | |
5743 | ||
5744 | while (sec != NULL) | |
5745 | { | |
5746 | fprintf (asm_out_file, "\n"); | |
5747 | fprintf (asm_out_file, "# JLI entry for function "); | |
5748 | assemble_name (asm_out_file, sec->name); | |
5749 | fprintf (asm_out_file, "\n\t.section .jlitab, \"axG\", @progbits, " | |
5750 | ".jlitab."); | |
5751 | assemble_name (asm_out_file, sec->name); | |
5752 | fprintf (asm_out_file,", comdat\n"); | |
5753 | ||
5754 | fprintf (asm_out_file, "\t.align\t4\n"); | |
5755 | fprintf (asm_out_file, "__jli."); | |
5756 | assemble_name (asm_out_file, sec->name); | |
5757 | fprintf (asm_out_file, ":\n\t.weak __jli."); | |
5758 | assemble_name (asm_out_file, sec->name); | |
5759 | fprintf (asm_out_file, "\n\tb\t@"); | |
5760 | assemble_name (asm_out_file, sec->name); | |
5761 | fprintf (asm_out_file, "\n"); | |
5762 | sec = sec->next; | |
5763 | } | |
5764 | file_end_indicate_exec_stack (); | |
5765 | } | |
5766 | ||
526b7aee SV |
5767 | /* Cost functions. */ |
5768 | ||
5769 | /* Compute a (partial) cost for rtx X. Return true if the complete | |
5770 | cost has been computed, and false if subexpressions should be | |
5771 | scanned. In either case, *TOTAL contains the cost result. */ | |
5772 | ||
5773 | static bool | |
e548c9df AM |
5774 | arc_rtx_costs (rtx x, machine_mode mode, int outer_code, |
5775 | int opno ATTRIBUTE_UNUSED, int *total, bool speed) | |
526b7aee | 5776 | { |
e548c9df AM |
5777 | int code = GET_CODE (x); |
5778 | ||
526b7aee SV |
5779 | switch (code) |
5780 | { | |
5781 | /* Small integers are as cheap as registers. */ | |
5782 | case CONST_INT: | |
5783 | { | |
5784 | bool nolimm = false; /* Can we do without long immediate? */ | |
526b7aee | 5785 | |
d797b115 | 5786 | nolimm = false; |
526b7aee | 5787 | if (UNSIGNED_INT6 (INTVAL (x))) |
d797b115 | 5788 | nolimm = true; |
526b7aee SV |
5789 | else |
5790 | { | |
526b7aee SV |
5791 | switch (outer_code) |
5792 | { | |
5793 | case AND: /* bclr, bmsk, ext[bw] */ | |
5794 | if (satisfies_constraint_Ccp (x) /* bclr */ | |
5795 | || satisfies_constraint_C1p (x) /* bmsk */) | |
d797b115 | 5796 | nolimm = true; |
526b7aee SV |
5797 | break; |
5798 | case IOR: /* bset */ | |
5799 | if (satisfies_constraint_C0p (x)) /* bset */ | |
d797b115 | 5800 | nolimm = true; |
526b7aee SV |
5801 | break; |
5802 | case XOR: | |
5803 | if (satisfies_constraint_C0p (x)) /* bxor */ | |
d797b115 | 5804 | nolimm = true; |
526b7aee | 5805 | break; |
d797b115 CZ |
5806 | case SET: |
5807 | if (UNSIGNED_INT8 (INTVAL (x))) | |
5808 | nolimm = true; | |
5809 | if (satisfies_constraint_Chi (x)) | |
5810 | nolimm = true; | |
5811 | if (satisfies_constraint_Clo (x)) | |
5812 | nolimm = true; | |
526b7aee SV |
5813 | default: |
5814 | break; | |
5815 | } | |
5816 | } | |
807c3ab5 | 5817 | if (nolimm) |
526b7aee SV |
5818 | { |
5819 | *total = 0; | |
5820 | return true; | |
5821 | } | |
5822 | } | |
5823 | /* FALLTHRU */ | |
5824 | ||
5825 | /* 4 byte values can be fetched as immediate constants - | |
5826 | let's give that the cost of an extra insn. */ | |
5827 | case CONST: | |
5828 | case LABEL_REF: | |
5829 | case SYMBOL_REF: | |
d797b115 | 5830 | *total = speed ? COSTS_N_INSNS (1) : COSTS_N_INSNS (4); |
526b7aee SV |
5831 | return true; |
5832 | ||
5833 | case CONST_DOUBLE: | |
5834 | { | |
7d81a567 | 5835 | rtx first, second; |
526b7aee SV |
5836 | |
5837 | if (TARGET_DPFP) | |
5838 | { | |
5839 | *total = COSTS_N_INSNS (1); | |
5840 | return true; | |
5841 | } | |
7d81a567 CZ |
5842 | split_double (x, &first, &second); |
5843 | *total = COSTS_N_INSNS (!SMALL_INT (INTVAL (first)) | |
5844 | + !SMALL_INT (INTVAL (second))); | |
526b7aee SV |
5845 | return true; |
5846 | } | |
5847 | ||
5848 | /* Encourage synth_mult to find a synthetic multiply when reasonable. | |
5849 | If we need more than 12 insns to do a multiply, then go out-of-line, | |
5850 | since the call overhead will be < 10% of the cost of the multiply. */ | |
5851 | case ASHIFT: | |
5852 | case ASHIFTRT: | |
5853 | case LSHIFTRT: | |
5854 | if (TARGET_BARREL_SHIFTER) | |
5855 | { | |
526b7aee SV |
5856 | if (CONSTANT_P (XEXP (x, 0))) |
5857 | { | |
d797b115 CZ |
5858 | *total += rtx_cost (XEXP (x, 1), mode, (enum rtx_code) code, |
5859 | 0, speed); | |
526b7aee SV |
5860 | return true; |
5861 | } | |
5862 | *total = COSTS_N_INSNS (1); | |
5863 | } | |
5864 | else if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5865 | *total = COSTS_N_INSNS (16); | |
5866 | else | |
5867 | { | |
5868 | *total = COSTS_N_INSNS (INTVAL (XEXP ((x), 1))); | |
5869 | /* ??? want_to_gcse_p can throw negative shift counts at us, | |
5870 | and then panics when it gets a negative cost as result. | |
5871 | Seen for gcc.c-torture/compile/20020710-1.c -Os . */ | |
5872 | if (*total < 0) | |
5873 | *total = 0; | |
5874 | } | |
5875 | return false; | |
5876 | ||
5877 | case DIV: | |
5878 | case UDIV: | |
d797b115 CZ |
5879 | if (GET_MODE_CLASS (mode) == MODE_FLOAT |
5880 | && (TARGET_FP_SP_SQRT || TARGET_FP_DP_SQRT)) | |
5881 | *total = COSTS_N_INSNS(1); | |
5882 | else if (GET_MODE_CLASS (mode) == MODE_INT | |
5883 | && TARGET_DIVREM) | |
5884 | *total = COSTS_N_INSNS(1); | |
5885 | else if (speed) | |
526b7aee SV |
5886 | *total = COSTS_N_INSNS(30); |
5887 | else | |
5888 | *total = COSTS_N_INSNS(1); | |
5889 | return false; | |
5890 | ||
5891 | case MULT: | |
5892 | if ((TARGET_DPFP && GET_MODE (x) == DFmode)) | |
5893 | *total = COSTS_N_INSNS (1); | |
5894 | else if (speed) | |
5895 | *total= arc_multcost; | |
5896 | /* We do not want synth_mult sequences when optimizing | |
5897 | for size. */ | |
d797b115 | 5898 | else if (TARGET_ANY_MPY) |
526b7aee SV |
5899 | *total = COSTS_N_INSNS (1); |
5900 | else | |
5901 | *total = COSTS_N_INSNS (2); | |
5902 | return false; | |
d797b115 | 5903 | |
526b7aee | 5904 | case PLUS: |
d797b115 CZ |
5905 | if (outer_code == MEM && CONST_INT_P (XEXP (x, 1)) |
5906 | && RTX_OK_FOR_OFFSET_P (mode, XEXP (x, 1))) | |
5907 | { | |
5908 | *total = 0; | |
5909 | return true; | |
5910 | } | |
5911 | ||
1e466f04 GM |
5912 | if ((GET_CODE (XEXP (x, 0)) == ASHIFT |
5913 | && _1_2_3_operand (XEXP (XEXP (x, 0), 1), VOIDmode)) | |
5914 | || (GET_CODE (XEXP (x, 0)) == MULT | |
5915 | && _2_4_8_operand (XEXP (XEXP (x, 0), 1), VOIDmode))) | |
526b7aee | 5916 | { |
d797b115 CZ |
5917 | if (CONSTANT_P (XEXP (x, 1)) && !speed) |
5918 | *total += COSTS_N_INSNS (4); | |
5919 | *total += rtx_cost (XEXP (XEXP (x, 0), 0), mode, PLUS, 1, speed); | |
526b7aee SV |
5920 | return true; |
5921 | } | |
5922 | return false; | |
5923 | case MINUS: | |
1e466f04 GM |
5924 | if ((GET_CODE (XEXP (x, 1)) == ASHIFT |
5925 | && _1_2_3_operand (XEXP (XEXP (x, 1), 1), VOIDmode)) | |
5926 | || (GET_CODE (XEXP (x, 1)) == MULT | |
5927 | && _2_4_8_operand (XEXP (XEXP (x, 1), 1), VOIDmode))) | |
526b7aee | 5928 | { |
d797b115 CZ |
5929 | if (CONSTANT_P (XEXP (x, 0)) && !speed) |
5930 | *total += COSTS_N_INSNS (4); | |
5931 | *total += rtx_cost (XEXP (XEXP (x, 1), 0), mode, PLUS, 1, speed); | |
526b7aee SV |
5932 | return true; |
5933 | } | |
5934 | return false; | |
d797b115 | 5935 | |
526b7aee SV |
5936 | case COMPARE: |
5937 | { | |
5938 | rtx op0 = XEXP (x, 0); | |
5939 | rtx op1 = XEXP (x, 1); | |
5940 | ||
5941 | if (GET_CODE (op0) == ZERO_EXTRACT && op1 == const0_rtx | |
5942 | && XEXP (op0, 1) == const1_rtx) | |
5943 | { | |
5944 | /* btst / bbit0 / bbit1: | |
5945 | Small integers and registers are free; everything else can | |
5946 | be put in a register. */ | |
e548c9df AM |
5947 | mode = GET_MODE (XEXP (op0, 0)); |
5948 | *total = (rtx_cost (XEXP (op0, 0), mode, SET, 1, speed) | |
5949 | + rtx_cost (XEXP (op0, 2), mode, SET, 1, speed)); | |
526b7aee SV |
5950 | return true; |
5951 | } | |
5952 | if (GET_CODE (op0) == AND && op1 == const0_rtx | |
5953 | && satisfies_constraint_C1p (XEXP (op0, 1))) | |
5954 | { | |
5955 | /* bmsk.f */ | |
e548c9df | 5956 | *total = rtx_cost (XEXP (op0, 0), VOIDmode, SET, 1, speed); |
526b7aee SV |
5957 | return true; |
5958 | } | |
5959 | /* add.f */ | |
5960 | if (GET_CODE (op1) == NEG) | |
5961 | { | |
5962 | /* op0 might be constant, the inside of op1 is rather | |
5963 | unlikely to be so. So swapping the operands might lower | |
5964 | the cost. */ | |
e548c9df AM |
5965 | mode = GET_MODE (op0); |
5966 | *total = (rtx_cost (op0, mode, PLUS, 1, speed) | |
5967 | + rtx_cost (XEXP (op1, 0), mode, PLUS, 0, speed)); | |
526b7aee SV |
5968 | } |
5969 | return false; | |
5970 | } | |
5971 | case EQ: case NE: | |
5972 | if (outer_code == IF_THEN_ELSE | |
5973 | && GET_CODE (XEXP (x, 0)) == ZERO_EXTRACT | |
5974 | && XEXP (x, 1) == const0_rtx | |
5975 | && XEXP (XEXP (x, 0), 1) == const1_rtx) | |
5976 | { | |
5977 | /* btst / bbit0 / bbit1: | |
5978 | Small integers and registers are free; everything else can | |
5979 | be put in a register. */ | |
5980 | rtx op0 = XEXP (x, 0); | |
5981 | ||
e548c9df AM |
5982 | mode = GET_MODE (XEXP (op0, 0)); |
5983 | *total = (rtx_cost (XEXP (op0, 0), mode, SET, 1, speed) | |
5984 | + rtx_cost (XEXP (op0, 2), mode, SET, 1, speed)); | |
526b7aee SV |
5985 | return true; |
5986 | } | |
5987 | /* Fall through. */ | |
5988 | /* scc_insn expands into two insns. */ | |
5989 | case GTU: case GEU: case LEU: | |
e548c9df | 5990 | if (mode == SImode) |
526b7aee SV |
5991 | *total += COSTS_N_INSNS (1); |
5992 | return false; | |
5993 | case LTU: /* might use adc. */ | |
e548c9df | 5994 | if (mode == SImode) |
526b7aee SV |
5995 | *total += COSTS_N_INSNS (1) - 1; |
5996 | return false; | |
5997 | default: | |
5998 | return false; | |
5999 | } | |
6000 | } | |
6001 | ||
526b7aee SV |
6002 | /* Return true if ADDR is a valid pic address. |
6003 | A valid pic address on arc should look like | |
6004 | const (unspec (SYMBOL_REF/LABEL) (ARC_UNSPEC_GOTOFF/ARC_UNSPEC_GOT)) */ | |
6005 | ||
6006 | bool | |
6007 | arc_legitimate_pic_addr_p (rtx addr) | |
6008 | { | |
526b7aee SV |
6009 | if (GET_CODE (addr) != CONST) |
6010 | return false; | |
6011 | ||
6012 | addr = XEXP (addr, 0); | |
6013 | ||
6014 | ||
6015 | if (GET_CODE (addr) == PLUS) | |
6016 | { | |
6017 | if (GET_CODE (XEXP (addr, 1)) != CONST_INT) | |
6018 | return false; | |
6019 | addr = XEXP (addr, 0); | |
6020 | } | |
6021 | ||
6022 | if (GET_CODE (addr) != UNSPEC | |
6023 | || XVECLEN (addr, 0) != 1) | |
6024 | return false; | |
6025 | ||
f5e336b1 | 6026 | /* Must be one of @GOT, @GOTOFF, @GOTOFFPC, @tlsgd, tlsie. */ |
526b7aee | 6027 | if (XINT (addr, 1) != ARC_UNSPEC_GOT |
28633bbd | 6028 | && XINT (addr, 1) != ARC_UNSPEC_GOTOFF |
f5e336b1 | 6029 | && XINT (addr, 1) != ARC_UNSPEC_GOTOFFPC |
28633bbd CZ |
6030 | && XINT (addr, 1) != UNSPEC_TLS_GD |
6031 | && XINT (addr, 1) != UNSPEC_TLS_IE) | |
526b7aee SV |
6032 | return false; |
6033 | ||
6034 | if (GET_CODE (XVECEXP (addr, 0, 0)) != SYMBOL_REF | |
6035 | && GET_CODE (XVECEXP (addr, 0, 0)) != LABEL_REF) | |
6036 | return false; | |
6037 | ||
6038 | return true; | |
6039 | } | |
6040 | ||
6041 | ||
6042 | ||
6043 | /* Return true if OP contains a symbol reference. */ | |
6044 | ||
6045 | static bool | |
6046 | symbolic_reference_mentioned_p (rtx op) | |
6047 | { | |
6048 | register const char *fmt; | |
6049 | register int i; | |
6050 | ||
6051 | if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF) | |
6052 | return true; | |
6053 | ||
6054 | fmt = GET_RTX_FORMAT (GET_CODE (op)); | |
6055 | for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--) | |
6056 | { | |
6057 | if (fmt[i] == 'E') | |
6058 | { | |
6059 | register int j; | |
6060 | ||
6061 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) | |
6062 | if (symbolic_reference_mentioned_p (XVECEXP (op, i, j))) | |
6063 | return true; | |
6064 | } | |
6065 | ||
6066 | else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i))) | |
6067 | return true; | |
6068 | } | |
6069 | ||
6070 | return false; | |
6071 | } | |
6072 | ||
6073 | /* Return true if OP contains a SYMBOL_REF that is not wrapped in an unspec. | |
6074 | If SKIP_LOCAL is true, skip symbols that bind locally. | |
6075 | This is used further down in this file, and, without SKIP_LOCAL, | |
6076 | in the addsi3 / subsi3 expanders when generating PIC code. */ | |
6077 | ||
6078 | bool | |
6079 | arc_raw_symbolic_reference_mentioned_p (rtx op, bool skip_local) | |
6080 | { | |
6081 | register const char *fmt; | |
6082 | register int i; | |
6083 | ||
6084 | if (GET_CODE(op) == UNSPEC) | |
6085 | return false; | |
6086 | ||
6087 | if (GET_CODE (op) == SYMBOL_REF) | |
6088 | { | |
28633bbd CZ |
6089 | if (SYMBOL_REF_TLS_MODEL (op)) |
6090 | return true; | |
6091 | if (!flag_pic) | |
6092 | return false; | |
526b7aee SV |
6093 | tree decl = SYMBOL_REF_DECL (op); |
6094 | return !skip_local || !decl || !default_binds_local_p (decl); | |
6095 | } | |
6096 | ||
6097 | fmt = GET_RTX_FORMAT (GET_CODE (op)); | |
6098 | for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--) | |
6099 | { | |
6100 | if (fmt[i] == 'E') | |
6101 | { | |
6102 | register int j; | |
6103 | ||
6104 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) | |
6105 | if (arc_raw_symbolic_reference_mentioned_p (XVECEXP (op, i, j), | |
6106 | skip_local)) | |
6107 | return true; | |
6108 | } | |
6109 | ||
6110 | else if (fmt[i] == 'e' | |
6111 | && arc_raw_symbolic_reference_mentioned_p (XEXP (op, i), | |
6112 | skip_local)) | |
6113 | return true; | |
6114 | } | |
6115 | ||
6116 | return false; | |
6117 | } | |
6118 | ||
8efa18d6 CZ |
6119 | /* The __tls_get_attr symbol. */ |
6120 | static GTY(()) rtx arc_tls_symbol; | |
28633bbd | 6121 | |
8efa18d6 CZ |
6122 | /* Emit a call to __tls_get_addr. TI is the argument to this function. |
6123 | RET is an RTX for the return value location. The entire insn sequence | |
6124 | is returned. */ | |
28633bbd CZ |
6125 | |
6126 | static rtx | |
8efa18d6 | 6127 | arc_call_tls_get_addr (rtx ti) |
28633bbd | 6128 | { |
8efa18d6 CZ |
6129 | rtx arg = gen_rtx_REG (Pmode, R0_REG); |
6130 | rtx ret = gen_rtx_REG (Pmode, R0_REG); | |
6131 | rtx fn; | |
6132 | rtx_insn *insn; | |
6133 | ||
6134 | if (!arc_tls_symbol) | |
6135 | arc_tls_symbol = init_one_libfunc ("__tls_get_addr"); | |
6136 | ||
6137 | emit_move_insn (arg, ti); | |
6138 | fn = gen_rtx_MEM (SImode, arc_tls_symbol); | |
6139 | insn = emit_call_insn (gen_call_value (ret, fn, const0_rtx)); | |
6140 | RTL_CONST_CALL_P (insn) = 1; | |
6141 | use_reg (&CALL_INSN_FUNCTION_USAGE (insn), ret); | |
6142 | use_reg (&CALL_INSN_FUNCTION_USAGE (insn), arg); | |
6143 | ||
6144 | return ret; | |
28633bbd CZ |
6145 | } |
6146 | ||
6147 | #define DTPOFF_ZERO_SYM ".tdata" | |
6148 | ||
6149 | /* Return a legitimized address for ADDR, | |
6150 | which is a SYMBOL_REF with tls_model MODEL. */ | |
6151 | ||
6152 | static rtx | |
6153 | arc_legitimize_tls_address (rtx addr, enum tls_model model) | |
6154 | { | |
8efa18d6 CZ |
6155 | rtx tmp; |
6156 | ||
28633bbd CZ |
6157 | if (!flag_pic && model == TLS_MODEL_LOCAL_DYNAMIC) |
6158 | model = TLS_MODEL_LOCAL_EXEC; | |
6159 | ||
8efa18d6 CZ |
6160 | |
6161 | /* The TP pointer needs to be set. */ | |
6162 | gcc_assert (arc_tp_regno != -1); | |
6163 | ||
28633bbd CZ |
6164 | switch (model) |
6165 | { | |
8efa18d6 CZ |
6166 | case TLS_MODEL_GLOBAL_DYNAMIC: |
6167 | tmp = gen_reg_rtx (Pmode); | |
6168 | emit_move_insn (tmp, arc_unspec_offset (addr, UNSPEC_TLS_GD)); | |
6169 | return arc_call_tls_get_addr (tmp); | |
6170 | ||
28633bbd CZ |
6171 | case TLS_MODEL_LOCAL_DYNAMIC: |
6172 | rtx base; | |
6173 | tree decl; | |
6174 | const char *base_name; | |
28633bbd CZ |
6175 | |
6176 | decl = SYMBOL_REF_DECL (addr); | |
6177 | base_name = DTPOFF_ZERO_SYM; | |
6178 | if (decl && bss_initializer_p (decl)) | |
6179 | base_name = ".tbss"; | |
6180 | ||
6181 | base = gen_rtx_SYMBOL_REF (Pmode, base_name); | |
8efa18d6 CZ |
6182 | tmp = gen_reg_rtx (Pmode); |
6183 | emit_move_insn (tmp, arc_unspec_offset (base, UNSPEC_TLS_GD)); | |
6184 | base = arc_call_tls_get_addr (tmp); | |
6185 | return gen_rtx_PLUS (Pmode, force_reg (Pmode, base), | |
6186 | arc_unspec_offset (addr, UNSPEC_TLS_OFF)); | |
5a5c5784 | 6187 | |
28633bbd | 6188 | case TLS_MODEL_INITIAL_EXEC: |
5a5c5784 | 6189 | addr = arc_unspec_offset (addr, UNSPEC_TLS_IE); |
28633bbd | 6190 | addr = copy_to_mode_reg (Pmode, gen_const_mem (Pmode, addr)); |
8efa18d6 | 6191 | return gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, arc_tp_regno), addr); |
5a5c5784 | 6192 | |
28633bbd | 6193 | case TLS_MODEL_LOCAL_EXEC: |
5a5c5784 | 6194 | addr = arc_unspec_offset (addr, UNSPEC_TLS_OFF); |
8efa18d6 CZ |
6195 | return gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, arc_tp_regno), addr); |
6196 | ||
28633bbd CZ |
6197 | default: |
6198 | gcc_unreachable (); | |
6199 | } | |
6200 | } | |
6201 | ||
673f01b8 | 6202 | /* Return true if SYMBOL_REF X binds locally. */ |
526b7aee | 6203 | |
673f01b8 CZ |
6204 | static bool |
6205 | arc_symbol_binds_local_p (const_rtx x) | |
526b7aee | 6206 | { |
673f01b8 CZ |
6207 | return (SYMBOL_REF_DECL (x) |
6208 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
6209 | : SYMBOL_REF_LOCAL_P (x)); | |
6210 | } | |
6211 | ||
6212 | /* Legitimize a pic address reference in ADDR. The return value is | |
6213 | the legitimated address. */ | |
526b7aee | 6214 | |
673f01b8 CZ |
6215 | static rtx |
6216 | arc_legitimize_pic_address (rtx addr) | |
6217 | { | |
6218 | if (!flag_pic) | |
6219 | return addr; | |
526b7aee | 6220 | |
673f01b8 | 6221 | switch (GET_CODE (addr)) |
526b7aee | 6222 | { |
673f01b8 CZ |
6223 | case SYMBOL_REF: |
6224 | /* TLS symbols are handled in different place. */ | |
6225 | if (SYMBOL_REF_TLS_MODEL (addr)) | |
6226 | return addr; | |
f5e336b1 CZ |
6227 | |
6228 | /* This symbol must be referenced via a load from the Global | |
6229 | Offset Table (@GOTPC). */ | |
673f01b8 CZ |
6230 | if (!arc_symbol_binds_local_p (addr)) |
6231 | return gen_const_mem (Pmode, arc_unspec_offset (addr, ARC_UNSPEC_GOT)); | |
526b7aee | 6232 | |
673f01b8 CZ |
6233 | /* Local symb: use @pcl to access it. */ |
6234 | /* Fall through. */ | |
6235 | case LABEL_REF: | |
6236 | return arc_unspec_offset (addr, ARC_UNSPEC_GOTOFFPC); | |
28633bbd | 6237 | |
673f01b8 CZ |
6238 | default: |
6239 | break; | |
526b7aee SV |
6240 | } |
6241 | ||
673f01b8 | 6242 | return addr; |
526b7aee SV |
6243 | } |
6244 | ||
6245 | /* Output address constant X to FILE, taking PIC into account. */ | |
6246 | ||
9f532472 | 6247 | static void |
526b7aee SV |
6248 | arc_output_pic_addr_const (FILE * file, rtx x, int code) |
6249 | { | |
6250 | char buf[256]; | |
6251 | ||
6252 | restart: | |
6253 | switch (GET_CODE (x)) | |
6254 | { | |
6255 | case PC: | |
6256 | if (flag_pic) | |
6257 | putc ('.', file); | |
6258 | else | |
6259 | gcc_unreachable (); | |
6260 | break; | |
6261 | ||
6262 | case SYMBOL_REF: | |
6263 | output_addr_const (file, x); | |
6264 | ||
6265 | /* Local functions do not get references through the PLT. */ | |
6266 | if (code == 'P' && ! SYMBOL_REF_LOCAL_P (x)) | |
6267 | fputs ("@plt", file); | |
6268 | break; | |
6269 | ||
6270 | case LABEL_REF: | |
6271 | ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (XEXP (x, 0))); | |
6272 | assemble_name (file, buf); | |
6273 | break; | |
6274 | ||
6275 | case CODE_LABEL: | |
6276 | ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x)); | |
6277 | assemble_name (file, buf); | |
6278 | break; | |
6279 | ||
6280 | case CONST_INT: | |
6281 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
6282 | break; | |
6283 | ||
6284 | case CONST: | |
6285 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
6286 | break; | |
6287 | ||
6288 | case CONST_DOUBLE: | |
6289 | if (GET_MODE (x) == VOIDmode) | |
6290 | { | |
6291 | /* We can use %d if the number is one word and positive. */ | |
6292 | if (CONST_DOUBLE_HIGH (x)) | |
6293 | fprintf (file, HOST_WIDE_INT_PRINT_DOUBLE_HEX, | |
6294 | CONST_DOUBLE_HIGH (x), CONST_DOUBLE_LOW (x)); | |
6295 | else if (CONST_DOUBLE_LOW (x) < 0) | |
6296 | fprintf (file, HOST_WIDE_INT_PRINT_HEX, CONST_DOUBLE_LOW (x)); | |
6297 | else | |
6298 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, CONST_DOUBLE_LOW (x)); | |
6299 | } | |
6300 | else | |
6301 | /* We can't handle floating point constants; | |
6302 | PRINT_OPERAND must handle them. */ | |
6303 | output_operand_lossage ("floating constant misused"); | |
6304 | break; | |
6305 | ||
6306 | case PLUS: | |
6307 | /* FIXME: Not needed here. */ | |
6308 | /* Some assemblers need integer constants to appear last (eg masm). */ | |
6309 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) | |
6310 | { | |
6311 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
6312 | fprintf (file, "+"); | |
6313 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
6314 | } | |
6315 | else if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
6316 | { | |
6317 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
6318 | if (INTVAL (XEXP (x, 1)) >= 0) | |
6319 | fprintf (file, "+"); | |
6320 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
6321 | } | |
6322 | else | |
6323 | gcc_unreachable(); | |
6324 | break; | |
6325 | ||
6326 | case MINUS: | |
6327 | /* Avoid outputting things like x-x or x+5-x, | |
6328 | since some assemblers can't handle that. */ | |
6329 | x = simplify_subtraction (x); | |
6330 | if (GET_CODE (x) != MINUS) | |
6331 | goto restart; | |
6332 | ||
6333 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
6334 | fprintf (file, "-"); | |
6335 | if (GET_CODE (XEXP (x, 1)) == CONST_INT | |
6336 | && INTVAL (XEXP (x, 1)) < 0) | |
6337 | { | |
6338 | fprintf (file, "("); | |
6339 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
6340 | fprintf (file, ")"); | |
6341 | } | |
6342 | else | |
6343 | arc_output_pic_addr_const (file, XEXP (x, 1), code); | |
6344 | break; | |
6345 | ||
6346 | case ZERO_EXTEND: | |
6347 | case SIGN_EXTEND: | |
6348 | arc_output_pic_addr_const (file, XEXP (x, 0), code); | |
6349 | break; | |
6350 | ||
6351 | ||
6352 | case UNSPEC: | |
28633bbd CZ |
6353 | const char *suffix; |
6354 | bool pcrel; pcrel = false; | |
6355 | rtx base; base = NULL; | |
6356 | gcc_assert (XVECLEN (x, 0) >= 1); | |
526b7aee SV |
6357 | switch (XINT (x, 1)) |
6358 | { | |
6359 | case ARC_UNSPEC_GOT: | |
28633bbd | 6360 | suffix = "@gotpc", pcrel = true; |
526b7aee SV |
6361 | break; |
6362 | case ARC_UNSPEC_GOTOFF: | |
28633bbd | 6363 | suffix = "@gotoff"; |
526b7aee | 6364 | break; |
f5e336b1 CZ |
6365 | case ARC_UNSPEC_GOTOFFPC: |
6366 | suffix = "@pcl", pcrel = true; | |
6367 | break; | |
526b7aee | 6368 | case ARC_UNSPEC_PLT: |
28633bbd CZ |
6369 | suffix = "@plt"; |
6370 | break; | |
6371 | case UNSPEC_TLS_GD: | |
6372 | suffix = "@tlsgd", pcrel = true; | |
6373 | break; | |
6374 | case UNSPEC_TLS_IE: | |
6375 | suffix = "@tlsie", pcrel = true; | |
6376 | break; | |
6377 | case UNSPEC_TLS_OFF: | |
6378 | if (XVECLEN (x, 0) == 2) | |
6379 | base = XVECEXP (x, 0, 1); | |
6380 | if (SYMBOL_REF_TLS_MODEL (XVECEXP (x, 0, 0)) == TLS_MODEL_LOCAL_EXEC | |
6381 | || (!flag_pic && !base)) | |
6382 | suffix = "@tpoff"; | |
6383 | else | |
6384 | suffix = "@dtpoff"; | |
526b7aee SV |
6385 | break; |
6386 | default: | |
cd1e4d41 | 6387 | suffix = "@invalid"; |
526b7aee SV |
6388 | output_operand_lossage ("invalid UNSPEC as operand: %d", XINT (x,1)); |
6389 | break; | |
6390 | } | |
28633bbd CZ |
6391 | if (pcrel) |
6392 | fputs ("pcl,", file); | |
6393 | arc_output_pic_addr_const (file, XVECEXP (x, 0, 0), code); | |
6394 | fputs (suffix, file); | |
6395 | if (base) | |
6396 | arc_output_pic_addr_const (file, base, code); | |
6397 | break; | |
526b7aee SV |
6398 | |
6399 | default: | |
6400 | output_operand_lossage ("invalid expression as operand"); | |
6401 | } | |
6402 | } | |
6403 | ||
526b7aee SV |
6404 | /* The function returning the number of words, at the beginning of an |
6405 | argument, must be put in registers. The returned value must be | |
6406 | zero for arguments that are passed entirely in registers or that | |
6407 | are entirely pushed on the stack. | |
6408 | ||
6409 | On some machines, certain arguments must be passed partially in | |
6410 | registers and partially in memory. On these machines, typically | |
6411 | the first N words of arguments are passed in registers, and the | |
6412 | rest on the stack. If a multi-word argument (a `double' or a | |
6413 | structure) crosses that boundary, its first few words must be | |
6414 | passed in registers and the rest must be pushed. This function | |
6415 | tells the compiler when this occurs, and how many of the words | |
6416 | should go in registers. | |
6417 | ||
6418 | `FUNCTION_ARG' for these arguments should return the first register | |
6419 | to be used by the caller for this argument; likewise | |
6420 | `FUNCTION_INCOMING_ARG', for the called function. | |
6421 | ||
6422 | The function is used to implement macro FUNCTION_ARG_PARTIAL_NREGS. */ | |
6423 | ||
6424 | /* If REGNO is the least arg reg available then what is the total number of arg | |
6425 | regs available. */ | |
6426 | #define GPR_REST_ARG_REGS(REGNO) \ | |
6427 | ((REGNO) <= MAX_ARC_PARM_REGS ? MAX_ARC_PARM_REGS - (REGNO) : 0 ) | |
6428 | ||
6429 | /* Since arc parm regs are contiguous. */ | |
6430 | #define ARC_NEXT_ARG_REG(REGNO) ( (REGNO) + 1 ) | |
6431 | ||
6432 | /* Implement TARGET_ARG_PARTIAL_BYTES. */ | |
6433 | ||
6434 | static int | |
ef4bddc2 | 6435 | arc_arg_partial_bytes (cumulative_args_t cum_v, machine_mode mode, |
526b7aee SV |
6436 | tree type, bool named ATTRIBUTE_UNUSED) |
6437 | { | |
6438 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
6439 | int bytes = (mode == BLKmode | |
6440 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode)); | |
6441 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
6442 | int arg_num = *cum; | |
6443 | int ret; | |
6444 | ||
6445 | arg_num = ROUND_ADVANCE_CUM (arg_num, mode, type); | |
6446 | ret = GPR_REST_ARG_REGS (arg_num); | |
6447 | ||
6448 | /* ICEd at function.c:2361, and ret is copied to data->partial */ | |
6449 | ret = (ret >= words ? 0 : ret * UNITS_PER_WORD); | |
6450 | ||
6451 | return ret; | |
6452 | } | |
6453 | ||
526b7aee SV |
6454 | /* This function is used to control a function argument is passed in a |
6455 | register, and which register. | |
6456 | ||
6457 | The arguments are CUM, of type CUMULATIVE_ARGS, which summarizes | |
6458 | (in a way defined by INIT_CUMULATIVE_ARGS and FUNCTION_ARG_ADVANCE) | |
6459 | all of the previous arguments so far passed in registers; MODE, the | |
6460 | machine mode of the argument; TYPE, the data type of the argument | |
6461 | as a tree node or 0 if that is not known (which happens for C | |
6462 | support library functions); and NAMED, which is 1 for an ordinary | |
6463 | argument and 0 for nameless arguments that correspond to `...' in | |
6464 | the called function's prototype. | |
6465 | ||
6466 | The returned value should either be a `reg' RTX for the hard | |
6467 | register in which to pass the argument, or zero to pass the | |
6468 | argument on the stack. | |
6469 | ||
6470 | For machines like the Vax and 68000, where normally all arguments | |
6471 | are pushed, zero suffices as a definition. | |
6472 | ||
6473 | The usual way to make the ANSI library `stdarg.h' work on a machine | |
6474 | where some arguments are usually passed in registers, is to cause | |
6475 | nameless arguments to be passed on the stack instead. This is done | |
6476 | by making the function return 0 whenever NAMED is 0. | |
6477 | ||
6478 | You may use the macro `MUST_PASS_IN_STACK (MODE, TYPE)' in the | |
6479 | definition of this function to determine if this argument is of a | |
6480 | type that must be passed in the stack. If `REG_PARM_STACK_SPACE' | |
6481 | is not defined and the function returns non-zero for such an | |
6482 | argument, the compiler will abort. If `REG_PARM_STACK_SPACE' is | |
6483 | defined, the argument will be computed in the stack and then loaded | |
6484 | into a register. | |
6485 | ||
6486 | The function is used to implement macro FUNCTION_ARG. */ | |
6487 | /* On the ARC the first MAX_ARC_PARM_REGS args are normally in registers | |
6488 | and the rest are pushed. */ | |
6489 | ||
6490 | static rtx | |
8f3304d0 CZ |
6491 | arc_function_arg (cumulative_args_t cum_v, |
6492 | machine_mode mode, | |
6493 | const_tree type ATTRIBUTE_UNUSED, | |
6494 | bool named ATTRIBUTE_UNUSED) | |
526b7aee SV |
6495 | { |
6496 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
6497 | int arg_num = *cum; | |
6498 | rtx ret; | |
6499 | const char *debstr ATTRIBUTE_UNUSED; | |
6500 | ||
6501 | arg_num = ROUND_ADVANCE_CUM (arg_num, mode, type); | |
6502 | /* Return a marker for use in the call instruction. */ | |
6503 | if (mode == VOIDmode) | |
6504 | { | |
6505 | ret = const0_rtx; | |
6506 | debstr = "<0>"; | |
6507 | } | |
6508 | else if (GPR_REST_ARG_REGS (arg_num) > 0) | |
6509 | { | |
6510 | ret = gen_rtx_REG (mode, arg_num); | |
6511 | debstr = reg_names [arg_num]; | |
6512 | } | |
6513 | else | |
6514 | { | |
6515 | ret = NULL_RTX; | |
6516 | debstr = "memory"; | |
6517 | } | |
6518 | return ret; | |
6519 | } | |
6520 | ||
6521 | /* The function to update the summarizer variable *CUM to advance past | |
6522 | an argument in the argument list. The values MODE, TYPE and NAMED | |
6523 | describe that argument. Once this is done, the variable *CUM is | |
6524 | suitable for analyzing the *following* argument with | |
6525 | `FUNCTION_ARG', etc. | |
6526 | ||
6527 | This function need not do anything if the argument in question was | |
6528 | passed on the stack. The compiler knows how to track the amount of | |
6529 | stack space used for arguments without any special help. | |
6530 | ||
6531 | The function is used to implement macro FUNCTION_ARG_ADVANCE. */ | |
6532 | /* For the ARC: the cum set here is passed on to function_arg where we | |
6533 | look at its value and say which reg to use. Strategy: advance the | |
6534 | regnumber here till we run out of arg regs, then set *cum to last | |
6535 | reg. In function_arg, since *cum > last arg reg we would return 0 | |
6536 | and thus the arg will end up on the stack. For straddling args of | |
6537 | course function_arg_partial_nregs will come into play. */ | |
6538 | ||
6539 | static void | |
8f3304d0 CZ |
6540 | arc_function_arg_advance (cumulative_args_t cum_v, |
6541 | machine_mode mode, | |
6542 | const_tree type, | |
6543 | bool named ATTRIBUTE_UNUSED) | |
526b7aee SV |
6544 | { |
6545 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
6546 | int bytes = (mode == BLKmode | |
6547 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode)); | |
6548 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
6549 | int i; | |
6550 | ||
6551 | if (words) | |
6552 | *cum = ROUND_ADVANCE_CUM (*cum, mode, type); | |
6553 | for (i = 0; i < words; i++) | |
6554 | *cum = ARC_NEXT_ARG_REG (*cum); | |
6555 | ||
6556 | } | |
6557 | ||
6558 | /* Define how to find the value returned by a function. | |
6559 | VALTYPE is the data type of the value (as a tree). | |
6560 | If the precise function being called is known, FN_DECL_OR_TYPE is its | |
6561 | FUNCTION_DECL; otherwise, FN_DECL_OR_TYPE is its type. */ | |
6562 | ||
6563 | static rtx | |
6564 | arc_function_value (const_tree valtype, | |
6565 | const_tree fn_decl_or_type ATTRIBUTE_UNUSED, | |
6566 | bool outgoing ATTRIBUTE_UNUSED) | |
6567 | { | |
ef4bddc2 | 6568 | machine_mode mode = TYPE_MODE (valtype); |
526b7aee SV |
6569 | int unsignedp ATTRIBUTE_UNUSED; |
6570 | ||
6571 | unsignedp = TYPE_UNSIGNED (valtype); | |
6572 | if (INTEGRAL_TYPE_P (valtype) || TREE_CODE (valtype) == OFFSET_TYPE) | |
6573 | PROMOTE_MODE (mode, unsignedp, valtype); | |
6574 | return gen_rtx_REG (mode, 0); | |
6575 | } | |
6576 | ||
6577 | /* Returns the return address that is used by builtin_return_address. */ | |
6578 | ||
6579 | rtx | |
6580 | arc_return_addr_rtx (int count, ATTRIBUTE_UNUSED rtx frame) | |
6581 | { | |
6582 | if (count != 0) | |
6583 | return const0_rtx; | |
6584 | ||
6585 | return get_hard_reg_initial_val (Pmode , RETURN_ADDR_REGNUM); | |
6586 | } | |
6587 | ||
526b7aee SV |
6588 | /* Determine if a given RTX is a valid constant. We already know this |
6589 | satisfies CONSTANT_P. */ | |
6590 | ||
6591 | bool | |
28633bbd | 6592 | arc_legitimate_constant_p (machine_mode mode, rtx x) |
526b7aee | 6593 | { |
526b7aee SV |
6594 | switch (GET_CODE (x)) |
6595 | { | |
6596 | case CONST: | |
b6c354eb | 6597 | if (flag_pic) |
526b7aee | 6598 | { |
b6c354eb | 6599 | if (arc_legitimate_pic_addr_p (x)) |
526b7aee | 6600 | return true; |
b6c354eb CZ |
6601 | } |
6602 | return arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
526b7aee | 6603 | |
526b7aee | 6604 | case SYMBOL_REF: |
28633bbd CZ |
6605 | if (SYMBOL_REF_TLS_MODEL (x)) |
6606 | return false; | |
6607 | /* Fall through. */ | |
6608 | case LABEL_REF: | |
6609 | if (flag_pic) | |
6610 | return false; | |
6611 | /* Fall through. */ | |
b6c354eb CZ |
6612 | case CONST_INT: |
6613 | case CONST_DOUBLE: | |
6614 | return true; | |
6615 | ||
6616 | case NEG: | |
6617 | return arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
6618 | ||
6619 | case PLUS: | |
6620 | case MINUS: | |
6621 | { | |
6622 | bool t1 = arc_legitimate_constant_p (mode, XEXP (x, 0)); | |
6623 | bool t2 = arc_legitimate_constant_p (mode, XEXP (x, 1)); | |
6624 | ||
6625 | return (t1 && t2); | |
6626 | } | |
6627 | ||
6628 | case CONST_VECTOR: | |
6629 | switch (mode) | |
6630 | { | |
4e10a5a7 | 6631 | case E_V2HImode: |
b6c354eb | 6632 | return TARGET_PLUS_DMPY; |
4e10a5a7 RS |
6633 | case E_V2SImode: |
6634 | case E_V4HImode: | |
b6c354eb CZ |
6635 | return TARGET_PLUS_QMACW; |
6636 | default: | |
6637 | return false; | |
6638 | } | |
6639 | ||
6640 | case UNSPEC: | |
6641 | switch (XINT (x, 1)) | |
6642 | { | |
6643 | case UNSPEC_TLS_GD: | |
6644 | case UNSPEC_TLS_OFF: | |
6645 | case UNSPEC_TLS_IE: | |
6646 | return true; | |
6647 | default: | |
6648 | /* Any other unspec ending here are pic related, hence the above | |
6649 | constant pic address checking returned false. */ | |
6650 | return false; | |
6651 | } | |
6652 | /* Fall through. */ | |
526b7aee SV |
6653 | |
6654 | default: | |
b6c354eb | 6655 | fatal_insn ("unrecognized supposed constant", x); |
526b7aee SV |
6656 | } |
6657 | ||
b6c354eb | 6658 | gcc_unreachable (); |
526b7aee SV |
6659 | } |
6660 | ||
6661 | static bool | |
ef4bddc2 | 6662 | arc_legitimate_address_p (machine_mode mode, rtx x, bool strict) |
526b7aee SV |
6663 | { |
6664 | if (RTX_OK_FOR_BASE_P (x, strict)) | |
6665 | return true; | |
ac2e1a51 | 6666 | if (legitimate_offset_address_p (mode, x, TARGET_INDEXED_LOADS, strict)) |
526b7aee | 6667 | return true; |
9f532472 | 6668 | if (legitimate_scaled_address_p (mode, x, strict)) |
526b7aee | 6669 | return true; |
e0be3321 | 6670 | if (legitimate_small_data_address_p (x)) |
526b7aee SV |
6671 | return true; |
6672 | if (GET_CODE (x) == CONST_INT && LARGE_INT (INTVAL (x))) | |
6673 | return true; | |
28633bbd CZ |
6674 | |
6675 | /* When we compile for size avoid const (@sym + offset) | |
6676 | addresses. */ | |
6677 | if (!flag_pic && optimize_size && !reload_completed | |
6678 | && (GET_CODE (x) == CONST) | |
6679 | && (GET_CODE (XEXP (x, 0)) == PLUS) | |
6680 | && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
6681 | && SYMBOL_REF_TLS_MODEL (XEXP (XEXP (x, 0), 0)) == 0 | |
6682 | && !SYMBOL_REF_FUNCTION_P (XEXP (XEXP (x, 0), 0))) | |
526b7aee | 6683 | { |
28633bbd CZ |
6684 | rtx addend = XEXP (XEXP (x, 0), 1); |
6685 | gcc_assert (CONST_INT_P (addend)); | |
6686 | HOST_WIDE_INT offset = INTVAL (addend); | |
6687 | ||
6688 | /* Allow addresses having a large offset to pass. Anyhow they | |
6689 | will end in a limm. */ | |
6690 | return !(offset > -1024 && offset < 1020); | |
6691 | } | |
6692 | ||
6693 | if ((GET_MODE_SIZE (mode) != 16) && CONSTANT_P (x)) | |
6694 | { | |
b6c354eb | 6695 | return arc_legitimate_constant_p (mode, x); |
526b7aee SV |
6696 | } |
6697 | if ((GET_CODE (x) == PRE_DEC || GET_CODE (x) == PRE_INC | |
6698 | || GET_CODE (x) == POST_DEC || GET_CODE (x) == POST_INC) | |
6699 | && RTX_OK_FOR_BASE_P (XEXP (x, 0), strict)) | |
6700 | return true; | |
6701 | /* We're restricted here by the `st' insn. */ | |
6702 | if ((GET_CODE (x) == PRE_MODIFY || GET_CODE (x) == POST_MODIFY) | |
6703 | && GET_CODE (XEXP ((x), 1)) == PLUS | |
6704 | && rtx_equal_p (XEXP ((x), 0), XEXP (XEXP (x, 1), 0)) | |
ac2e1a51 | 6705 | && legitimate_offset_address_p (QImode, XEXP (x, 1), |
526b7aee SV |
6706 | TARGET_AUTO_MODIFY_REG, strict)) |
6707 | return true; | |
6708 | return false; | |
6709 | } | |
6710 | ||
6711 | /* Return true iff ADDR (a legitimate address expression) | |
6712 | has an effect that depends on the machine mode it is used for. */ | |
6713 | ||
6714 | static bool | |
6715 | arc_mode_dependent_address_p (const_rtx addr, addr_space_t) | |
6716 | { | |
6717 | /* SYMBOL_REF is not mode dependent: it is either a small data reference, | |
6718 | which is valid for loads and stores, or a limm offset, which is valid for | |
1fccdd40 | 6719 | loads. Scaled indices are scaled by the access mode. */ |
526b7aee | 6720 | if (GET_CODE (addr) == PLUS |
1fccdd40 | 6721 | && GET_CODE (XEXP ((addr), 0)) == MULT) |
526b7aee SV |
6722 | return true; |
6723 | return false; | |
6724 | } | |
6725 | ||
6726 | /* Determine if it's legal to put X into the constant pool. */ | |
6727 | ||
6728 | static bool | |
ef4bddc2 | 6729 | arc_cannot_force_const_mem (machine_mode mode, rtx x) |
526b7aee SV |
6730 | { |
6731 | return !arc_legitimate_constant_p (mode, x); | |
6732 | } | |
6733 | ||
c69899f0 CZ |
6734 | /* IDs for all the ARC builtins. */ |
6735 | ||
6736 | enum arc_builtin_id | |
6737 | { | |
6738 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6739 | ARC_BUILTIN_ ## NAME, | |
6740 | #include "builtins.def" | |
6741 | #undef DEF_BUILTIN | |
6742 | ||
6743 | ARC_BUILTIN_COUNT | |
6744 | }; | |
6745 | ||
6746 | struct GTY(()) arc_builtin_description | |
6747 | { | |
6748 | enum insn_code icode; | |
6749 | int n_args; | |
6750 | tree fndecl; | |
6751 | }; | |
6752 | ||
6753 | static GTY(()) struct arc_builtin_description | |
6754 | arc_bdesc[ARC_BUILTIN_COUNT] = | |
6755 | { | |
6756 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6757 | { (enum insn_code) CODE_FOR_ ## ICODE, N_ARGS, NULL_TREE }, | |
6758 | #include "builtins.def" | |
6759 | #undef DEF_BUILTIN | |
6760 | }; | |
6761 | ||
6762 | /* Transform UP into lowercase and write the result to LO. | |
6763 | You must provide enough space for LO. Return LO. */ | |
6764 | ||
6765 | static char* | |
6766 | arc_tolower (char *lo, const char *up) | |
6767 | { | |
6768 | char *lo0 = lo; | |
6769 | ||
6770 | for (; *up; up++, lo++) | |
6771 | *lo = TOLOWER (*up); | |
6772 | ||
6773 | *lo = '\0'; | |
6774 | ||
6775 | return lo0; | |
6776 | } | |
6777 | ||
6778 | /* Implement `TARGET_BUILTIN_DECL'. */ | |
526b7aee | 6779 | |
c69899f0 CZ |
6780 | static tree |
6781 | arc_builtin_decl (unsigned id, bool initialize_p ATTRIBUTE_UNUSED) | |
6782 | { | |
6783 | if (id < ARC_BUILTIN_COUNT) | |
6784 | return arc_bdesc[id].fndecl; | |
526b7aee | 6785 | |
c69899f0 CZ |
6786 | return error_mark_node; |
6787 | } | |
526b7aee SV |
6788 | |
6789 | static void | |
6790 | arc_init_builtins (void) | |
6791 | { | |
00c072ae CZ |
6792 | tree V4HI_type_node; |
6793 | tree V2SI_type_node; | |
6794 | tree V2HI_type_node; | |
6795 | ||
6796 | /* Vector types based on HS SIMD elements. */ | |
6797 | V4HI_type_node = build_vector_type_for_mode (intHI_type_node, V4HImode); | |
6798 | V2SI_type_node = build_vector_type_for_mode (intSI_type_node, V2SImode); | |
6799 | V2HI_type_node = build_vector_type_for_mode (intHI_type_node, V2HImode); | |
6800 | ||
c69899f0 CZ |
6801 | tree pcvoid_type_node |
6802 | = build_pointer_type (build_qualified_type (void_type_node, | |
6803 | TYPE_QUAL_CONST)); | |
6804 | tree V8HI_type_node = build_vector_type_for_mode (intHI_type_node, | |
6805 | V8HImode); | |
6806 | ||
6807 | tree void_ftype_void | |
6808 | = build_function_type_list (void_type_node, NULL_TREE); | |
6809 | tree int_ftype_int | |
6810 | = build_function_type_list (integer_type_node, integer_type_node, | |
6811 | NULL_TREE); | |
6812 | tree int_ftype_pcvoid_int | |
6813 | = build_function_type_list (integer_type_node, pcvoid_type_node, | |
6814 | integer_type_node, NULL_TREE); | |
6815 | tree void_ftype_usint_usint | |
6816 | = build_function_type_list (void_type_node, long_unsigned_type_node, | |
6817 | long_unsigned_type_node, NULL_TREE); | |
6818 | tree int_ftype_int_int | |
6819 | = build_function_type_list (integer_type_node, integer_type_node, | |
6820 | integer_type_node, NULL_TREE); | |
6821 | tree usint_ftype_usint | |
6822 | = build_function_type_list (long_unsigned_type_node, | |
6823 | long_unsigned_type_node, NULL_TREE); | |
6824 | tree void_ftype_usint | |
6825 | = build_function_type_list (void_type_node, long_unsigned_type_node, | |
6826 | NULL_TREE); | |
6827 | tree int_ftype_void | |
6828 | = build_function_type_list (integer_type_node, void_type_node, | |
6829 | NULL_TREE); | |
6830 | tree void_ftype_int | |
6831 | = build_function_type_list (void_type_node, integer_type_node, | |
6832 | NULL_TREE); | |
6833 | tree int_ftype_short | |
6834 | = build_function_type_list (integer_type_node, short_integer_type_node, | |
6835 | NULL_TREE); | |
6836 | ||
6837 | /* Old ARC SIMD types. */ | |
6838 | tree v8hi_ftype_v8hi_v8hi | |
6839 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6840 | V8HI_type_node, NULL_TREE); | |
6841 | tree v8hi_ftype_v8hi_int | |
6842 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6843 | integer_type_node, NULL_TREE); | |
6844 | tree v8hi_ftype_v8hi_int_int | |
6845 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6846 | integer_type_node, integer_type_node, | |
6847 | NULL_TREE); | |
6848 | tree void_ftype_v8hi_int_int | |
6849 | = build_function_type_list (void_type_node, V8HI_type_node, | |
6850 | integer_type_node, integer_type_node, | |
6851 | NULL_TREE); | |
6852 | tree void_ftype_v8hi_int_int_int | |
6853 | = build_function_type_list (void_type_node, V8HI_type_node, | |
6854 | integer_type_node, integer_type_node, | |
6855 | integer_type_node, NULL_TREE); | |
6856 | tree v8hi_ftype_int_int | |
6857 | = build_function_type_list (V8HI_type_node, integer_type_node, | |
6858 | integer_type_node, NULL_TREE); | |
6859 | tree void_ftype_int_int | |
6860 | = build_function_type_list (void_type_node, integer_type_node, | |
6861 | integer_type_node, NULL_TREE); | |
6862 | tree v8hi_ftype_v8hi | |
6863 | = build_function_type_list (V8HI_type_node, V8HI_type_node, | |
6864 | NULL_TREE); | |
00c072ae CZ |
6865 | /* ARCv2 SIMD types. */ |
6866 | tree long_ftype_v4hi_v4hi | |
6867 | = build_function_type_list (long_long_integer_type_node, | |
6868 | V4HI_type_node, V4HI_type_node, NULL_TREE); | |
6869 | tree int_ftype_v2hi_v2hi | |
6870 | = build_function_type_list (integer_type_node, | |
6871 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6872 | tree v2si_ftype_v2hi_v2hi | |
6873 | = build_function_type_list (V2SI_type_node, | |
6874 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6875 | tree v2hi_ftype_v2hi_v2hi | |
6876 | = build_function_type_list (V2HI_type_node, | |
6877 | V2HI_type_node, V2HI_type_node, NULL_TREE); | |
6878 | tree v2si_ftype_v2si_v2si | |
6879 | = build_function_type_list (V2SI_type_node, | |
6880 | V2SI_type_node, V2SI_type_node, NULL_TREE); | |
6881 | tree v4hi_ftype_v4hi_v4hi | |
6882 | = build_function_type_list (V4HI_type_node, | |
6883 | V4HI_type_node, V4HI_type_node, NULL_TREE); | |
6884 | tree long_ftype_v2si_v2hi | |
6885 | = build_function_type_list (long_long_integer_type_node, | |
6886 | V2SI_type_node, V2HI_type_node, NULL_TREE); | |
c69899f0 CZ |
6887 | |
6888 | /* Add the builtins. */ | |
6889 | #define DEF_BUILTIN(NAME, N_ARGS, TYPE, ICODE, MASK) \ | |
6890 | { \ | |
6891 | int id = ARC_BUILTIN_ ## NAME; \ | |
6892 | const char *Name = "__builtin_arc_" #NAME; \ | |
6893 | char *name = (char*) alloca (1 + strlen (Name)); \ | |
6894 | \ | |
6895 | gcc_assert (id < ARC_BUILTIN_COUNT); \ | |
6896 | if (MASK) \ | |
6897 | arc_bdesc[id].fndecl \ | |
6898 | = add_builtin_function (arc_tolower(name, Name), TYPE, id, \ | |
6899 | BUILT_IN_MD, NULL, NULL_TREE); \ | |
6900 | } | |
6901 | #include "builtins.def" | |
6902 | #undef DEF_BUILTIN | |
6903 | } | |
6904 | ||
6905 | /* Helper to expand __builtin_arc_aligned (void* val, int | |
6906 | alignval). */ | |
6907 | ||
6908 | static rtx | |
6909 | arc_expand_builtin_aligned (tree exp) | |
6910 | { | |
6911 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
6912 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
6913 | fold (arg1); | |
6914 | rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
6915 | rtx op1 = expand_expr (arg1, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
6916 | ||
6917 | if (!CONST_INT_P (op1)) | |
6918 | { | |
6919 | /* If we can't fold the alignment to a constant integer | |
6920 | whilst optimizing, this is probably a user error. */ | |
6921 | if (optimize) | |
a3f9f006 | 6922 | warning (0, "%<__builtin_arc_aligned%> with non-constant alignment"); |
c69899f0 CZ |
6923 | } |
6924 | else | |
6925 | { | |
6926 | HOST_WIDE_INT alignTest = INTVAL (op1); | |
6927 | /* Check alignTest is positive, and a power of two. */ | |
6928 | if (alignTest <= 0 || alignTest != (alignTest & -alignTest)) | |
6929 | { | |
a3f9f006 | 6930 | error ("invalid alignment value for %<__builtin_arc_aligned%>"); |
c69899f0 CZ |
6931 | return NULL_RTX; |
6932 | } | |
6933 | ||
6934 | if (CONST_INT_P (op0)) | |
6935 | { | |
6936 | HOST_WIDE_INT pnt = INTVAL (op0); | |
6937 | ||
6938 | if ((pnt & (alignTest - 1)) == 0) | |
6939 | return const1_rtx; | |
6940 | } | |
6941 | else | |
6942 | { | |
6943 | unsigned align = get_pointer_alignment (arg0); | |
6944 | unsigned numBits = alignTest * BITS_PER_UNIT; | |
6945 | ||
6946 | if (align && align >= numBits) | |
6947 | return const1_rtx; | |
6948 | /* Another attempt to ascertain alignment. Check the type | |
6949 | we are pointing to. */ | |
6950 | if (POINTER_TYPE_P (TREE_TYPE (arg0)) | |
6951 | && TYPE_ALIGN (TREE_TYPE (TREE_TYPE (arg0))) >= numBits) | |
6952 | return const1_rtx; | |
6953 | } | |
6954 | } | |
6955 | ||
6956 | /* Default to false. */ | |
6957 | return const0_rtx; | |
6958 | } | |
6959 | ||
6960 | /* Helper arc_expand_builtin, generates a pattern for the given icode | |
6961 | and arguments. */ | |
6962 | ||
6963 | static rtx_insn * | |
6964 | apply_GEN_FCN (enum insn_code icode, rtx *arg) | |
6965 | { | |
6966 | switch (insn_data[icode].n_generator_args) | |
6967 | { | |
6968 | case 0: | |
6969 | return GEN_FCN (icode) (); | |
6970 | case 1: | |
6971 | return GEN_FCN (icode) (arg[0]); | |
6972 | case 2: | |
6973 | return GEN_FCN (icode) (arg[0], arg[1]); | |
6974 | case 3: | |
6975 | return GEN_FCN (icode) (arg[0], arg[1], arg[2]); | |
6976 | case 4: | |
6977 | return GEN_FCN (icode) (arg[0], arg[1], arg[2], arg[3]); | |
6978 | case 5: | |
6979 | return GEN_FCN (icode) (arg[0], arg[1], arg[2], arg[3], arg[4]); | |
6980 | default: | |
6981 | gcc_unreachable (); | |
6982 | } | |
6983 | } | |
526b7aee SV |
6984 | |
6985 | /* Expand an expression EXP that calls a built-in function, | |
6986 | with result going to TARGET if that's convenient | |
6987 | (and in mode MODE if that's convenient). | |
6988 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
6989 | IGNORE is nonzero if the value is to be ignored. */ | |
6990 | ||
6991 | static rtx | |
6992 | arc_expand_builtin (tree exp, | |
6993 | rtx target, | |
c69899f0 CZ |
6994 | rtx subtarget ATTRIBUTE_UNUSED, |
6995 | machine_mode mode ATTRIBUTE_UNUSED, | |
6996 | int ignore ATTRIBUTE_UNUSED) | |
6997 | { | |
6998 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
4d732405 | 6999 | unsigned int id = DECL_MD_FUNCTION_CODE (fndecl); |
c69899f0 CZ |
7000 | const struct arc_builtin_description *d = &arc_bdesc[id]; |
7001 | int i, j, n_args = call_expr_nargs (exp); | |
7002 | rtx pat = NULL_RTX; | |
7003 | rtx xop[5]; | |
7004 | enum insn_code icode = d->icode; | |
7005 | machine_mode tmode = insn_data[icode].operand[0].mode; | |
7006 | int nonvoid; | |
7007 | tree arg0; | |
7008 | tree arg1; | |
7009 | tree arg2; | |
7010 | tree arg3; | |
7011 | rtx op0; | |
7012 | rtx op1; | |
7013 | rtx op2; | |
7014 | rtx op3; | |
7015 | rtx op4; | |
ef4bddc2 RS |
7016 | machine_mode mode0; |
7017 | machine_mode mode1; | |
c69899f0 CZ |
7018 | machine_mode mode2; |
7019 | machine_mode mode3; | |
7020 | machine_mode mode4; | |
526b7aee | 7021 | |
c69899f0 CZ |
7022 | if (id >= ARC_BUILTIN_COUNT) |
7023 | internal_error ("bad builtin fcode"); | |
526b7aee | 7024 | |
c69899f0 CZ |
7025 | /* 1st part: Expand special builtins. */ |
7026 | switch (id) | |
526b7aee SV |
7027 | { |
7028 | case ARC_BUILTIN_NOP: | |
c69899f0 | 7029 | emit_insn (gen_nopv ()); |
526b7aee SV |
7030 | return NULL_RTX; |
7031 | ||
c69899f0 CZ |
7032 | case ARC_BUILTIN_RTIE: |
7033 | case ARC_BUILTIN_SYNC: | |
7034 | case ARC_BUILTIN_BRK: | |
7035 | case ARC_BUILTIN_SWI: | |
7036 | case ARC_BUILTIN_UNIMP_S: | |
7037 | gcc_assert (icode != 0); | |
7038 | emit_insn (GEN_FCN (icode) (const1_rtx)); | |
7039 | return NULL_RTX; | |
526b7aee | 7040 | |
c69899f0 CZ |
7041 | case ARC_BUILTIN_ALIGNED: |
7042 | return arc_expand_builtin_aligned (exp); | |
526b7aee | 7043 | |
c69899f0 CZ |
7044 | case ARC_BUILTIN_CLRI: |
7045 | target = gen_reg_rtx (SImode); | |
7046 | emit_insn (gen_clri (target, const1_rtx)); | |
526b7aee SV |
7047 | return target; |
7048 | ||
c69899f0 CZ |
7049 | case ARC_BUILTIN_TRAP_S: |
7050 | case ARC_BUILTIN_SLEEP: | |
526b7aee | 7051 | arg0 = CALL_EXPR_ARG (exp, 0); |
c69899f0 | 7052 | fold (arg0); |
526b7aee | 7053 | op0 = expand_expr (arg0, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
526b7aee | 7054 | |
c69899f0 CZ |
7055 | gcc_assert (icode != 0); |
7056 | emit_insn (GEN_FCN (icode) (op0)); | |
7057 | return NULL_RTX; | |
526b7aee | 7058 | |
c69899f0 CZ |
7059 | case ARC_BUILTIN_VDORUN: |
7060 | case ARC_BUILTIN_VDIRUN: | |
526b7aee SV |
7061 | arg0 = CALL_EXPR_ARG (exp, 0); |
7062 | arg1 = CALL_EXPR_ARG (exp, 1); | |
c69899f0 CZ |
7063 | op0 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); |
7064 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee | 7065 | |
c69899f0 CZ |
7066 | target = gen_rtx_REG (SImode, (id == ARC_BUILTIN_VDIRUN) ? 131 : 139); |
7067 | ||
7068 | mode0 = insn_data[icode].operand[1].mode; | |
7069 | mode1 = insn_data[icode].operand[2].mode; | |
526b7aee | 7070 | |
c69899f0 | 7071 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) |
526b7aee SV |
7072 | op0 = copy_to_mode_reg (mode0, op0); |
7073 | ||
c69899f0 | 7074 | if (!insn_data[icode].operand[2].predicate (op1, mode1)) |
526b7aee SV |
7075 | op1 = copy_to_mode_reg (mode1, op1); |
7076 | ||
c69899f0 CZ |
7077 | pat = GEN_FCN (icode) (target, op0, op1); |
7078 | if (!pat) | |
7079 | return NULL_RTX; | |
7080 | ||
7081 | emit_insn (pat); | |
526b7aee SV |
7082 | return NULL_RTX; |
7083 | ||
c69899f0 CZ |
7084 | case ARC_BUILTIN_VDIWR: |
7085 | case ARC_BUILTIN_VDOWR: | |
526b7aee SV |
7086 | arg0 = CALL_EXPR_ARG (exp, 0); |
7087 | arg1 = CALL_EXPR_ARG (exp, 1); | |
c69899f0 CZ |
7088 | op0 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); |
7089 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
7090 | ||
7091 | if (!CONST_INT_P (op0) | |
7092 | || !(UNSIGNED_INT3 (INTVAL (op0)))) | |
7093 | error ("operand 1 should be an unsigned 3-bit immediate"); | |
526b7aee | 7094 | |
526b7aee SV |
7095 | mode1 = insn_data[icode].operand[1].mode; |
7096 | ||
c69899f0 CZ |
7097 | if (icode == CODE_FOR_vdiwr_insn) |
7098 | target = gen_rtx_REG (SImode, | |
7099 | ARC_FIRST_SIMD_DMA_CONFIG_IN_REG + INTVAL (op0)); | |
7100 | else if (icode == CODE_FOR_vdowr_insn) | |
7101 | target = gen_rtx_REG (SImode, | |
7102 | ARC_FIRST_SIMD_DMA_CONFIG_OUT_REG + INTVAL (op0)); | |
7103 | else | |
7104 | gcc_unreachable (); | |
526b7aee | 7105 | |
c69899f0 | 7106 | if (!insn_data[icode].operand[2].predicate (op1, mode1)) |
526b7aee SV |
7107 | op1 = copy_to_mode_reg (mode1, op1); |
7108 | ||
c69899f0 CZ |
7109 | pat = GEN_FCN (icode) (target, op1); |
7110 | if (!pat) | |
7111 | return NULL_RTX; | |
526b7aee | 7112 | |
c69899f0 | 7113 | emit_insn (pat); |
526b7aee SV |
7114 | return NULL_RTX; |
7115 | ||
c69899f0 CZ |
7116 | case ARC_BUILTIN_VASRW: |
7117 | case ARC_BUILTIN_VSR8: | |
7118 | case ARC_BUILTIN_VSR8AW: | |
526b7aee | 7119 | arg0 = CALL_EXPR_ARG (exp, 0); |
c69899f0 CZ |
7120 | arg1 = CALL_EXPR_ARG (exp, 1); |
7121 | op0 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
7122 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
7123 | op2 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
7124 | ||
7125 | target = gen_reg_rtx (V8HImode); | |
526b7aee | 7126 | mode0 = insn_data[icode].operand[1].mode; |
c69899f0 | 7127 | mode1 = insn_data[icode].operand[2].mode; |
526b7aee | 7128 | |
c69899f0 | 7129 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) |
526b7aee SV |
7130 | op0 = copy_to_mode_reg (mode0, op0); |
7131 | ||
c69899f0 CZ |
7132 | if ((!insn_data[icode].operand[2].predicate (op1, mode1)) |
7133 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
7134 | error ("operand 2 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 7135 | |
c69899f0 CZ |
7136 | pat = GEN_FCN (icode) (target, op0, op1, op2); |
7137 | if (!pat) | |
7138 | return NULL_RTX; | |
526b7aee | 7139 | |
c69899f0 CZ |
7140 | emit_insn (pat); |
7141 | return target; | |
526b7aee | 7142 | |
c69899f0 CZ |
7143 | case ARC_BUILTIN_VLD32WH: |
7144 | case ARC_BUILTIN_VLD32WL: | |
7145 | case ARC_BUILTIN_VLD64: | |
7146 | case ARC_BUILTIN_VLD32: | |
7147 | rtx src_vreg; | |
7148 | icode = d->icode; | |
7149 | arg0 = CALL_EXPR_ARG (exp, 0); /* source vreg. */ | |
7150 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
7151 | arg2 = CALL_EXPR_ARG (exp, 2); /* u8. */ | |
526b7aee | 7152 | |
c69899f0 CZ |
7153 | src_vreg = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); |
7154 | op0 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
7155 | op1 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
7156 | op2 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
526b7aee | 7157 | |
c69899f0 CZ |
7158 | /* target <- src vreg. */ |
7159 | emit_insn (gen_move_insn (target, src_vreg)); | |
526b7aee | 7160 | |
c69899f0 CZ |
7161 | /* target <- vec_concat: target, mem (Ib, u8). */ |
7162 | mode0 = insn_data[icode].operand[3].mode; | |
7163 | mode1 = insn_data[icode].operand[1].mode; | |
526b7aee | 7164 | |
c69899f0 CZ |
7165 | if ((!insn_data[icode].operand[3].predicate (op0, mode0)) |
7166 | || !(UNSIGNED_INT3 (INTVAL (op0)))) | |
7167 | error ("operand 1 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 7168 | |
c69899f0 CZ |
7169 | if ((!insn_data[icode].operand[1].predicate (op1, mode1)) |
7170 | || !(UNSIGNED_INT8 (INTVAL (op1)))) | |
7171 | error ("operand 2 should be an unsigned 8-bit value"); | |
526b7aee | 7172 | |
c69899f0 CZ |
7173 | pat = GEN_FCN (icode) (target, op1, op2, op0); |
7174 | if (!pat) | |
7175 | return NULL_RTX; | |
526b7aee | 7176 | |
c69899f0 CZ |
7177 | emit_insn (pat); |
7178 | return target; | |
526b7aee | 7179 | |
c69899f0 CZ |
7180 | case ARC_BUILTIN_VLD64W: |
7181 | case ARC_BUILTIN_VLD128: | |
7182 | arg0 = CALL_EXPR_ARG (exp, 0); /* dest vreg. */ | |
7183 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
526b7aee | 7184 | |
c69899f0 CZ |
7185 | op0 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); |
7186 | op1 = expand_expr (arg0, NULL_RTX, SImode, EXPAND_NORMAL); | |
7187 | op2 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee | 7188 | |
c69899f0 CZ |
7189 | /* target <- src vreg. */ |
7190 | target = gen_reg_rtx (V8HImode); | |
526b7aee | 7191 | |
c69899f0 CZ |
7192 | /* target <- vec_concat: target, mem (Ib, u8). */ |
7193 | mode0 = insn_data[icode].operand[1].mode; | |
7194 | mode1 = insn_data[icode].operand[2].mode; | |
7195 | mode2 = insn_data[icode].operand[3].mode; | |
526b7aee | 7196 | |
c69899f0 CZ |
7197 | if ((!insn_data[icode].operand[2].predicate (op1, mode1)) |
7198 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
7199 | error ("operand 1 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 7200 | |
c69899f0 CZ |
7201 | if ((!insn_data[icode].operand[3].predicate (op2, mode2)) |
7202 | || !(UNSIGNED_INT8 (INTVAL (op2)))) | |
7203 | error ("operand 2 should be an unsigned 8-bit value"); | |
526b7aee | 7204 | |
c69899f0 | 7205 | pat = GEN_FCN (icode) (target, op0, op1, op2); |
526b7aee | 7206 | |
c69899f0 CZ |
7207 | if (!pat) |
7208 | return NULL_RTX; | |
526b7aee | 7209 | |
c69899f0 | 7210 | emit_insn (pat); |
526b7aee SV |
7211 | return target; |
7212 | ||
c69899f0 CZ |
7213 | case ARC_BUILTIN_VST128: |
7214 | case ARC_BUILTIN_VST64: | |
7215 | arg0 = CALL_EXPR_ARG (exp, 0); /* src vreg. */ | |
7216 | arg1 = CALL_EXPR_ARG (exp, 1); /* [I]0-7. */ | |
7217 | arg2 = CALL_EXPR_ARG (exp, 2); /* u8. */ | |
526b7aee | 7218 | |
c69899f0 CZ |
7219 | op0 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); |
7220 | op1 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
7221 | op2 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
7222 | op3 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
526b7aee SV |
7223 | |
7224 | mode0 = insn_data[icode].operand[0].mode; | |
7225 | mode1 = insn_data[icode].operand[1].mode; | |
c69899f0 CZ |
7226 | mode2 = insn_data[icode].operand[2].mode; |
7227 | mode3 = insn_data[icode].operand[3].mode; | |
526b7aee | 7228 | |
c69899f0 CZ |
7229 | if ((!insn_data[icode].operand[1].predicate (op1, mode1)) |
7230 | || !(UNSIGNED_INT3 (INTVAL (op1)))) | |
7231 | error ("operand 2 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 7232 | |
c69899f0 CZ |
7233 | if ((!insn_data[icode].operand[2].predicate (op2, mode2)) |
7234 | || !(UNSIGNED_INT8 (INTVAL (op2)))) | |
7235 | error ("operand 3 should be an unsigned 8-bit value"); | |
526b7aee | 7236 | |
c69899f0 CZ |
7237 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) |
7238 | op3 = copy_to_mode_reg (mode3, op3); | |
526b7aee | 7239 | |
c69899f0 CZ |
7240 | pat = GEN_FCN (icode) (op0, op1, op2, op3); |
7241 | if (!pat) | |
7242 | return NULL_RTX; | |
526b7aee | 7243 | |
c69899f0 CZ |
7244 | emit_insn (pat); |
7245 | return NULL_RTX; | |
526b7aee | 7246 | |
c69899f0 CZ |
7247 | case ARC_BUILTIN_VST16_N: |
7248 | case ARC_BUILTIN_VST32_N: | |
7249 | arg0 = CALL_EXPR_ARG (exp, 0); /* source vreg. */ | |
7250 | arg1 = CALL_EXPR_ARG (exp, 1); /* u3. */ | |
7251 | arg2 = CALL_EXPR_ARG (exp, 2); /* [I]0-7. */ | |
7252 | arg3 = CALL_EXPR_ARG (exp, 3); /* u8. */ | |
526b7aee | 7253 | |
c69899f0 CZ |
7254 | op0 = expand_expr (arg3, NULL_RTX, SImode, EXPAND_NORMAL); |
7255 | op1 = gen_rtx_REG (V8HImode, ARC_FIRST_SIMD_VR_REG); | |
7256 | op2 = expand_expr (arg2, NULL_RTX, SImode, EXPAND_NORMAL); | |
7257 | op3 = expand_expr (arg0, NULL_RTX, V8HImode, EXPAND_NORMAL); | |
7258 | op4 = expand_expr (arg1, NULL_RTX, SImode, EXPAND_NORMAL); | |
526b7aee SV |
7259 | |
7260 | mode0 = insn_data[icode].operand[0].mode; | |
c69899f0 CZ |
7261 | mode2 = insn_data[icode].operand[2].mode; |
7262 | mode3 = insn_data[icode].operand[3].mode; | |
7263 | mode4 = insn_data[icode].operand[4].mode; | |
526b7aee | 7264 | |
c69899f0 CZ |
7265 | /* Do some correctness checks for the operands. */ |
7266 | if ((!insn_data[icode].operand[0].predicate (op0, mode0)) | |
7267 | || !(UNSIGNED_INT8 (INTVAL (op0)))) | |
7268 | error ("operand 4 should be an unsigned 8-bit value (0-255)"); | |
526b7aee | 7269 | |
c69899f0 CZ |
7270 | if ((!insn_data[icode].operand[2].predicate (op2, mode2)) |
7271 | || !(UNSIGNED_INT3 (INTVAL (op2)))) | |
7272 | error ("operand 3 should be an unsigned 3-bit value (I0-I7)"); | |
526b7aee | 7273 | |
c69899f0 CZ |
7274 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) |
7275 | op3 = copy_to_mode_reg (mode3, op3); | |
526b7aee | 7276 | |
c69899f0 CZ |
7277 | if ((!insn_data[icode].operand[4].predicate (op4, mode4)) |
7278 | || !(UNSIGNED_INT3 (INTVAL (op4)))) | |
7279 | error ("operand 2 should be an unsigned 3-bit value (subreg 0-7)"); | |
7280 | else if (icode == CODE_FOR_vst32_n_insn | |
7281 | && ((INTVAL (op4) % 2) != 0)) | |
7282 | error ("operand 2 should be an even 3-bit value (subreg 0,2,4,6)"); | |
526b7aee | 7283 | |
c69899f0 CZ |
7284 | pat = GEN_FCN (icode) (op0, op1, op2, op3, op4); |
7285 | if (!pat) | |
7286 | return NULL_RTX; | |
526b7aee | 7287 | |
c69899f0 | 7288 | emit_insn (pat); |
526b7aee SV |
7289 | return NULL_RTX; |
7290 | ||
c69899f0 CZ |
7291 | default: |
7292 | break; | |
7293 | } | |
7294 | ||
7295 | /* 2nd part: Expand regular builtins. */ | |
7296 | if (icode == 0) | |
7297 | internal_error ("bad builtin fcode"); | |
7298 | ||
7299 | nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node; | |
7300 | j = 0; | |
526b7aee | 7301 | |
c69899f0 CZ |
7302 | if (nonvoid) |
7303 | { | |
7304 | if (target == NULL_RTX | |
7305 | || GET_MODE (target) != tmode | |
7306 | || !insn_data[icode].operand[0].predicate (target, tmode)) | |
526b7aee | 7307 | { |
c69899f0 | 7308 | target = gen_reg_rtx (tmode); |
526b7aee | 7309 | } |
c69899f0 CZ |
7310 | xop[j++] = target; |
7311 | } | |
7312 | ||
7313 | gcc_assert (n_args <= 4); | |
7314 | for (i = 0; i < n_args; i++, j++) | |
7315 | { | |
7316 | tree arg = CALL_EXPR_ARG (exp, i); | |
7317 | machine_mode mode = insn_data[icode].operand[j].mode; | |
7318 | rtx op = expand_expr (arg, NULL_RTX, mode, EXPAND_NORMAL); | |
7319 | machine_mode opmode = GET_MODE (op); | |
7320 | char c = insn_data[icode].operand[j].constraint[0]; | |
7321 | ||
7322 | /* SIMD extension requires exact immediate operand match. */ | |
7323 | if ((id > ARC_BUILTIN_SIMD_BEGIN) | |
7324 | && (id < ARC_BUILTIN_SIMD_END) | |
7325 | && (c != 'v') | |
7326 | && (c != 'r')) | |
526b7aee | 7327 | { |
c69899f0 CZ |
7328 | if (!CONST_INT_P (op)) |
7329 | error ("builtin requires an immediate for operand %d", j); | |
7330 | switch (c) | |
526b7aee | 7331 | { |
c69899f0 CZ |
7332 | case 'L': |
7333 | if (!satisfies_constraint_L (op)) | |
7334 | error ("operand %d should be a 6 bit unsigned immediate", j); | |
7335 | break; | |
7336 | case 'P': | |
7337 | if (!satisfies_constraint_P (op)) | |
7338 | error ("operand %d should be a 8 bit unsigned immediate", j); | |
7339 | break; | |
7340 | case 'K': | |
7341 | if (!satisfies_constraint_K (op)) | |
7342 | error ("operand %d should be a 3 bit unsigned immediate", j); | |
7343 | break; | |
7344 | default: | |
7345 | error ("unknown builtin immediate operand type for operand %d", | |
7346 | j); | |
526b7aee | 7347 | } |
c69899f0 | 7348 | } |
526b7aee | 7349 | |
c69899f0 CZ |
7350 | if (CONST_INT_P (op)) |
7351 | opmode = mode; | |
526b7aee | 7352 | |
c69899f0 CZ |
7353 | if ((opmode == SImode) && (mode == HImode)) |
7354 | { | |
7355 | opmode = HImode; | |
7356 | op = gen_lowpart (HImode, op); | |
526b7aee SV |
7357 | } |
7358 | ||
c69899f0 CZ |
7359 | /* In case the insn wants input operands in modes different from |
7360 | the result, abort. */ | |
7361 | gcc_assert (opmode == mode || opmode == VOIDmode); | |
526b7aee | 7362 | |
c69899f0 CZ |
7363 | if (!insn_data[icode].operand[i + nonvoid].predicate (op, mode)) |
7364 | op = copy_to_mode_reg (mode, op); | |
7365 | ||
7366 | xop[j] = op; | |
526b7aee SV |
7367 | } |
7368 | ||
c69899f0 CZ |
7369 | pat = apply_GEN_FCN (icode, xop); |
7370 | if (pat == NULL_RTX) | |
7371 | return NULL_RTX; | |
7372 | ||
7373 | emit_insn (pat); | |
7374 | ||
7375 | if (nonvoid) | |
7376 | return target; | |
7377 | else | |
7378 | return const0_rtx; | |
526b7aee SV |
7379 | } |
7380 | ||
7381 | /* Returns true if the operands[opno] is a valid compile-time constant to be | |
7382 | used as register number in the code for builtins. Else it flags an error | |
7383 | and returns false. */ | |
7384 | ||
7385 | bool | |
7386 | check_if_valid_regno_const (rtx *operands, int opno) | |
7387 | { | |
7388 | ||
7389 | switch (GET_CODE (operands[opno])) | |
7390 | { | |
7391 | case SYMBOL_REF : | |
7392 | case CONST : | |
7393 | case CONST_INT : | |
7394 | return true; | |
7395 | default: | |
2fa9c1f6 CZ |
7396 | error ("register number must be a compile-time constant. " |
7397 | "Try giving higher optimization levels"); | |
526b7aee SV |
7398 | break; |
7399 | } | |
7400 | return false; | |
7401 | } | |
7402 | ||
526b7aee SV |
7403 | /* Return true if it is ok to make a tail-call to DECL. */ |
7404 | ||
7405 | static bool | |
6b55f8c9 | 7406 | arc_function_ok_for_sibcall (tree decl, |
526b7aee SV |
7407 | tree exp ATTRIBUTE_UNUSED) |
7408 | { | |
6b55f8c9 CZ |
7409 | tree attrs = NULL_TREE; |
7410 | ||
526b7aee SV |
7411 | /* Never tailcall from an ISR routine - it needs a special exit sequence. */ |
7412 | if (ARC_INTERRUPT_P (arc_compute_function_type (cfun))) | |
7413 | return false; | |
7414 | ||
6b55f8c9 CZ |
7415 | if (decl) |
7416 | { | |
7417 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); | |
7418 | ||
7419 | if (lookup_attribute ("jli_always", attrs)) | |
7420 | return false; | |
7421 | if (lookup_attribute ("jli_fixed", attrs)) | |
7422 | return false; | |
7778a1ad CZ |
7423 | if (lookup_attribute ("secure_call", attrs)) |
7424 | return false; | |
6b55f8c9 CZ |
7425 | } |
7426 | ||
526b7aee SV |
7427 | /* Everything else is ok. */ |
7428 | return true; | |
7429 | } | |
7430 | ||
7431 | /* Output code to add DELTA to the first argument, and then jump | |
7432 | to FUNCTION. Used for C++ multiple inheritance. */ | |
7433 | ||
7434 | static void | |
7435 | arc_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
7436 | HOST_WIDE_INT delta, | |
7437 | HOST_WIDE_INT vcall_offset, | |
7438 | tree function) | |
7439 | { | |
f7430263 | 7440 | const char *fnname = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (thunk)); |
526b7aee SV |
7441 | int mi_delta = delta; |
7442 | const char *const mi_op = mi_delta < 0 ? "sub" : "add"; | |
7443 | int shift = 0; | |
7444 | int this_regno | |
7445 | = aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) ? 1 : 0; | |
7446 | rtx fnaddr; | |
7447 | ||
f7430263 MF |
7448 | assemble_start_function (thunk, fnname); |
7449 | ||
526b7aee SV |
7450 | if (mi_delta < 0) |
7451 | mi_delta = - mi_delta; | |
7452 | ||
7453 | /* Add DELTA. When possible use a plain add, otherwise load it into | |
7454 | a register first. */ | |
7455 | ||
7456 | while (mi_delta != 0) | |
7457 | { | |
7458 | if ((mi_delta & (3 << shift)) == 0) | |
7459 | shift += 2; | |
7460 | else | |
7461 | { | |
7462 | asm_fprintf (file, "\t%s\t%s, %s, %d\n", | |
7463 | mi_op, reg_names[this_regno], reg_names[this_regno], | |
7464 | mi_delta & (0xff << shift)); | |
7465 | mi_delta &= ~(0xff << shift); | |
7466 | shift += 8; | |
7467 | } | |
7468 | } | |
7469 | ||
7470 | /* If needed, add *(*THIS + VCALL_OFFSET) to THIS. */ | |
7471 | if (vcall_offset != 0) | |
7472 | { | |
7473 | /* ld r12,[this] --> temp = *this | |
7474 | add r12,r12,vcall_offset --> temp = *(*this + vcall_offset) | |
7475 | ld r12,[r12] | |
7476 | add this,this,r12 --> this+ = *(*this + vcall_offset) */ | |
7477 | asm_fprintf (file, "\tld\t%s, [%s]\n", | |
7478 | ARC_TEMP_SCRATCH_REG, reg_names[this_regno]); | |
dfca07ea | 7479 | asm_fprintf (file, "\tadd\t%s, %s, " HOST_WIDE_INT_PRINT_DEC "\n", |
526b7aee SV |
7480 | ARC_TEMP_SCRATCH_REG, ARC_TEMP_SCRATCH_REG, vcall_offset); |
7481 | asm_fprintf (file, "\tld\t%s, [%s]\n", | |
7482 | ARC_TEMP_SCRATCH_REG, ARC_TEMP_SCRATCH_REG); | |
7483 | asm_fprintf (file, "\tadd\t%s, %s, %s\n", reg_names[this_regno], | |
7484 | reg_names[this_regno], ARC_TEMP_SCRATCH_REG); | |
7485 | } | |
7486 | ||
7487 | fnaddr = XEXP (DECL_RTL (function), 0); | |
7488 | ||
7489 | if (arc_is_longcall_p (fnaddr)) | |
1f8876c7 CZ |
7490 | { |
7491 | if (flag_pic) | |
7492 | { | |
7493 | asm_fprintf (file, "\tld\t%s, [pcl, @", | |
7494 | ARC_TEMP_SCRATCH_REG); | |
7495 | assemble_name (file, XSTR (fnaddr, 0)); | |
7496 | fputs ("@gotpc]\n", file); | |
7497 | asm_fprintf (file, "\tj\t[%s]", ARC_TEMP_SCRATCH_REG); | |
7498 | } | |
7499 | else | |
7500 | { | |
7501 | fputs ("\tj\t@", file); | |
7502 | assemble_name (file, XSTR (fnaddr, 0)); | |
7503 | } | |
7504 | } | |
526b7aee | 7505 | else |
1f8876c7 CZ |
7506 | { |
7507 | fputs ("\tb\t@", file); | |
7508 | assemble_name (file, XSTR (fnaddr, 0)); | |
7509 | if (flag_pic) | |
7510 | fputs ("@plt\n", file); | |
7511 | } | |
526b7aee | 7512 | fputc ('\n', file); |
f7430263 | 7513 | assemble_end_function (thunk, fnname); |
526b7aee SV |
7514 | } |
7515 | ||
7516 | /* Return true if a 32 bit "long_call" should be generated for | |
7517 | this calling SYM_REF. We generate a long_call if the function: | |
7518 | ||
7519 | a. has an __attribute__((long call)) | |
7520 | or b. the -mlong-calls command line switch has been specified | |
7521 | ||
7522 | However we do not generate a long call if the function has an | |
7523 | __attribute__ ((short_call)) or __attribute__ ((medium_call)) | |
7524 | ||
7525 | This function will be called by C fragments contained in the machine | |
7526 | description file. */ | |
7527 | ||
7528 | bool | |
7529 | arc_is_longcall_p (rtx sym_ref) | |
7530 | { | |
7531 | if (GET_CODE (sym_ref) != SYMBOL_REF) | |
7532 | return false; | |
7533 | ||
7534 | return (SYMBOL_REF_LONG_CALL_P (sym_ref) | |
7535 | || (TARGET_LONG_CALLS_SET | |
7536 | && !SYMBOL_REF_SHORT_CALL_P (sym_ref) | |
7537 | && !SYMBOL_REF_MEDIUM_CALL_P (sym_ref))); | |
7538 | ||
7539 | } | |
7540 | ||
7541 | /* Likewise for short calls. */ | |
7542 | ||
7543 | bool | |
7544 | arc_is_shortcall_p (rtx sym_ref) | |
7545 | { | |
7546 | if (GET_CODE (sym_ref) != SYMBOL_REF) | |
7547 | return false; | |
7548 | ||
7549 | return (SYMBOL_REF_SHORT_CALL_P (sym_ref) | |
7550 | || (!TARGET_LONG_CALLS_SET && !TARGET_MEDIUM_CALLS | |
7551 | && !SYMBOL_REF_LONG_CALL_P (sym_ref) | |
7552 | && !SYMBOL_REF_MEDIUM_CALL_P (sym_ref))); | |
7553 | ||
7554 | } | |
7555 | ||
526b7aee SV |
7556 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
7557 | ||
7558 | static bool | |
7559 | arc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) | |
7560 | { | |
7561 | if (AGGREGATE_TYPE_P (type) || TREE_ADDRESSABLE (type)) | |
7562 | return true; | |
7563 | else | |
7564 | { | |
7565 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
f50bb868 | 7566 | return (size == -1 || size > (TARGET_V2 ? 16 : 8)); |
526b7aee SV |
7567 | } |
7568 | } | |
7569 | ||
526b7aee SV |
7570 | static bool |
7571 | arc_pass_by_reference (cumulative_args_t ca_v ATTRIBUTE_UNUSED, | |
ef4bddc2 | 7572 | machine_mode mode ATTRIBUTE_UNUSED, |
526b7aee SV |
7573 | const_tree type, |
7574 | bool named ATTRIBUTE_UNUSED) | |
7575 | { | |
7576 | return (type != 0 | |
7577 | && (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST | |
7578 | || TREE_ADDRESSABLE (type))); | |
7579 | } | |
7580 | ||
1d0216c8 RS |
7581 | /* Implement TARGET_CAN_USE_DOLOOP_P. */ |
7582 | ||
7583 | static bool | |
a2de90a4 CZ |
7584 | arc_can_use_doloop_p (const widest_int &, |
7585 | const widest_int &iterations_max, | |
1d0216c8 RS |
7586 | unsigned int loop_depth, bool entered_at_top) |
7587 | { | |
a2de90a4 CZ |
7588 | /* Considering limitations in the hardware, only use doloop |
7589 | for innermost loops which must be entered from the top. */ | |
7590 | if (loop_depth > 1 || !entered_at_top) | |
1d0216c8 | 7591 | return false; |
a2de90a4 CZ |
7592 | |
7593 | /* Check for lp_count width boundary. */ | |
7594 | if (arc_lpcwidth != 32 | |
7595 | && (wi::gtu_p (iterations_max, ((1 << arc_lpcwidth) - 1)) | |
7596 | || wi::eq_p (iterations_max, 0))) | |
1d0216c8 RS |
7597 | return false; |
7598 | return true; | |
7599 | } | |
526b7aee | 7600 | |
a2de90a4 CZ |
7601 | /* NULL if INSN insn is valid within a low-overhead loop. Otherwise |
7602 | return why doloop cannot be applied. */ | |
526b7aee SV |
7603 | |
7604 | static const char * | |
ac44248e | 7605 | arc_invalid_within_doloop (const rtx_insn *insn) |
526b7aee SV |
7606 | { |
7607 | if (CALL_P (insn)) | |
7608 | return "Function call in the loop."; | |
a2de90a4 CZ |
7609 | |
7610 | /* FIXME! add here all the ZOL exceptions. */ | |
526b7aee SV |
7611 | return NULL; |
7612 | } | |
7613 | ||
635aeaa2 CZ |
7614 | /* Return the next active insn, skiping the inline assembly code. */ |
7615 | ||
7616 | static rtx_insn * | |
7617 | arc_active_insn (rtx_insn *insn) | |
7618 | { | |
7619 | rtx_insn *nxt = next_active_insn (insn); | |
7620 | ||
7621 | if (nxt && GET_CODE (PATTERN (nxt)) == ASM_INPUT) | |
7622 | nxt = next_active_insn (nxt); | |
7623 | return nxt; | |
7624 | } | |
7625 | ||
7626 | /* Search for a sequence made out of two stores and a given number of | |
7627 | loads, insert a nop if required. */ | |
7628 | ||
7629 | static void | |
7630 | check_store_cacheline_hazard (void) | |
7631 | { | |
7632 | rtx_insn *insn, *succ0, *insn1; | |
7633 | bool found = false; | |
7634 | ||
7635 | for (insn = get_insns (); insn; insn = arc_active_insn (insn)) | |
7636 | { | |
7637 | succ0 = arc_active_insn (insn); | |
7638 | ||
7639 | if (!succ0) | |
7640 | return; | |
7641 | ||
7642 | if (!single_set (insn) || !single_set (succ0)) | |
7643 | continue; | |
7644 | ||
7645 | if ((get_attr_type (insn) != TYPE_STORE) | |
7646 | || (get_attr_type (succ0) != TYPE_STORE)) | |
7647 | continue; | |
7648 | ||
7649 | /* Found at least two consecutive stores. Goto the end of the | |
7650 | store sequence. */ | |
7651 | for (insn1 = succ0; insn1; insn1 = arc_active_insn (insn1)) | |
7652 | if (!single_set (insn1) || get_attr_type (insn1) != TYPE_STORE) | |
7653 | break; | |
7654 | ||
7655 | /* Now, check the next two instructions for the following cases: | |
7656 | 1. next instruction is a LD => insert 2 nops between store | |
7657 | sequence and load. | |
7658 | 2. next-next instruction is a LD => inset 1 nop after the store | |
7659 | sequence. */ | |
7660 | if (insn1 && single_set (insn1) | |
7661 | && (get_attr_type (insn1) == TYPE_LOAD)) | |
7662 | { | |
7663 | found = true; | |
7664 | emit_insn_before (gen_nopv (), insn1); | |
7665 | emit_insn_before (gen_nopv (), insn1); | |
7666 | } | |
7667 | else | |
7668 | { | |
7669 | if (insn1 && (get_attr_type (insn1) == TYPE_COMPARE)) | |
7670 | { | |
7671 | /* REG_SAVE_NOTE is used by Haifa scheduler, we are in | |
7672 | reorg, so it is safe to reuse it for avoiding the | |
7673 | current compare insn to be part of a BRcc | |
7674 | optimization. */ | |
7675 | add_reg_note (insn1, REG_SAVE_NOTE, GEN_INT (3)); | |
7676 | } | |
7677 | insn1 = arc_active_insn (insn1); | |
7678 | if (insn1 && single_set (insn1) | |
7679 | && (get_attr_type (insn1) == TYPE_LOAD)) | |
7680 | { | |
7681 | found = true; | |
7682 | emit_insn_before (gen_nopv (), insn1); | |
7683 | } | |
7684 | } | |
7685 | ||
7686 | insn = insn1; | |
7687 | if (found) | |
7688 | found = false; | |
7689 | } | |
7690 | } | |
7691 | ||
e9472c81 AB |
7692 | /* Return true if a load instruction (CONSUMER) uses the same address as a |
7693 | store instruction (PRODUCER). This function is used to avoid st/ld | |
7694 | address hazard in ARC700 cores. */ | |
635aeaa2 CZ |
7695 | |
7696 | static bool | |
7697 | arc_store_addr_hazard_internal_p (rtx_insn* producer, rtx_insn* consumer) | |
e9472c81 AB |
7698 | { |
7699 | rtx in_set, out_set; | |
7700 | rtx out_addr, in_addr; | |
7701 | ||
7702 | if (!producer) | |
7703 | return false; | |
7704 | ||
7705 | if (!consumer) | |
7706 | return false; | |
7707 | ||
7708 | /* Peel the producer and the consumer for the address. */ | |
7709 | out_set = single_set (producer); | |
7710 | if (out_set) | |
7711 | { | |
7712 | out_addr = SET_DEST (out_set); | |
7713 | if (!out_addr) | |
7714 | return false; | |
7715 | if (GET_CODE (out_addr) == ZERO_EXTEND | |
7716 | || GET_CODE (out_addr) == SIGN_EXTEND) | |
7717 | out_addr = XEXP (out_addr, 0); | |
7718 | ||
7719 | if (!MEM_P (out_addr)) | |
7720 | return false; | |
7721 | ||
7722 | in_set = single_set (consumer); | |
7723 | if (in_set) | |
7724 | { | |
7725 | in_addr = SET_SRC (in_set); | |
7726 | if (!in_addr) | |
7727 | return false; | |
7728 | if (GET_CODE (in_addr) == ZERO_EXTEND | |
7729 | || GET_CODE (in_addr) == SIGN_EXTEND) | |
7730 | in_addr = XEXP (in_addr, 0); | |
7731 | ||
7732 | if (!MEM_P (in_addr)) | |
7733 | return false; | |
7734 | /* Get rid of the MEM and check if the addresses are | |
7735 | equivalent. */ | |
7736 | in_addr = XEXP (in_addr, 0); | |
7737 | out_addr = XEXP (out_addr, 0); | |
7738 | ||
7739 | return exp_equiv_p (in_addr, out_addr, 0, true); | |
7740 | } | |
7741 | } | |
7742 | return false; | |
7743 | } | |
7744 | ||
635aeaa2 CZ |
7745 | /* Return TRUE is we have an store address hazard. */ |
7746 | ||
7747 | bool | |
7748 | arc_store_addr_hazard_p (rtx_insn* producer, rtx_insn* consumer) | |
7749 | { | |
7750 | if (TARGET_ARC700 && (arc_tune != ARC_TUNE_ARC7XX)) | |
7751 | return true; | |
7752 | return arc_store_addr_hazard_internal_p (producer, consumer); | |
7753 | } | |
7754 | ||
f50bb868 CZ |
7755 | /* The same functionality as arc_hazard. It is called in machine |
7756 | reorg before any other optimization. Hence, the NOP size is taken | |
7757 | into account when doing branch shortening. */ | |
7758 | ||
7759 | static void | |
7760 | workaround_arc_anomaly (void) | |
7761 | { | |
7762 | rtx_insn *insn, *succ0; | |
635aeaa2 | 7763 | rtx_insn *succ1; |
f50bb868 CZ |
7764 | |
7765 | /* For any architecture: call arc_hazard here. */ | |
7766 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
7767 | { | |
7768 | succ0 = next_real_insn (insn); | |
7769 | if (arc_hazard (insn, succ0)) | |
7770 | { | |
7771 | emit_insn_before (gen_nopv (), succ0); | |
7772 | } | |
7773 | } | |
e9472c81 | 7774 | |
635aeaa2 CZ |
7775 | if (!TARGET_ARC700) |
7776 | return; | |
e9472c81 | 7777 | |
635aeaa2 CZ |
7778 | /* Old A7 are suffering of a cache hazard, and we need to insert two |
7779 | nops between any sequence of stores and a load. */ | |
7780 | if (arc_tune != ARC_TUNE_ARC7XX) | |
7781 | check_store_cacheline_hazard (); | |
e9472c81 | 7782 | |
635aeaa2 CZ |
7783 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
7784 | { | |
7785 | succ0 = next_real_insn (insn); | |
7786 | if (arc_store_addr_hazard_internal_p (insn, succ0)) | |
7787 | { | |
7788 | emit_insn_after (gen_nopv (), insn); | |
7789 | emit_insn_after (gen_nopv (), insn); | |
7790 | continue; | |
e9472c81 | 7791 | } |
635aeaa2 CZ |
7792 | |
7793 | /* Avoid adding nops if the instruction between the ST and LD is | |
7794 | a call or jump. */ | |
7795 | succ1 = next_real_insn (succ0); | |
7796 | if (succ0 && !JUMP_P (succ0) && !CALL_P (succ0) | |
7797 | && arc_store_addr_hazard_internal_p (insn, succ1)) | |
7798 | emit_insn_after (gen_nopv (), insn); | |
e9472c81 | 7799 | } |
f50bb868 CZ |
7800 | } |
7801 | ||
a2de90a4 CZ |
7802 | /* A callback for the hw-doloop pass. Called when a loop we have discovered |
7803 | turns out not to be optimizable; we have to split the loop_end pattern into | |
7804 | a subtract and a test. */ | |
7805 | ||
7806 | static void | |
7807 | hwloop_fail (hwloop_info loop) | |
7808 | { | |
7809 | rtx test; | |
7810 | rtx insn = loop->loop_end; | |
7811 | ||
62f26645 | 7812 | if (TARGET_DBNZ |
a2de90a4 CZ |
7813 | && (loop->length && (loop->length <= ARC_MAX_LOOP_LENGTH)) |
7814 | && REG_P (loop->iter_reg)) | |
7815 | { | |
62f26645 | 7816 | /* TARGET_V2 core3 has dbnz instructions. */ |
a2de90a4 CZ |
7817 | test = gen_dbnz (loop->iter_reg, loop->start_label); |
7818 | insn = emit_jump_insn_before (test, loop->loop_end); | |
7819 | } | |
7820 | else if (REG_P (loop->iter_reg) && (REGNO (loop->iter_reg) == LP_COUNT)) | |
7821 | { | |
7822 | /* We have the lp_count as loop iterator, try to use it. */ | |
7823 | emit_insn_before (gen_loop_fail (), loop->loop_end); | |
7824 | test = gen_rtx_NE (VOIDmode, gen_rtx_REG (CC_ZNmode, CC_REG), | |
7825 | const0_rtx); | |
7826 | test = gen_rtx_IF_THEN_ELSE (VOIDmode, test, | |
7827 | gen_rtx_LABEL_REF (Pmode, loop->start_label), | |
7828 | pc_rtx); | |
7829 | insn = emit_jump_insn_before (gen_rtx_SET (pc_rtx, test), | |
7830 | loop->loop_end); | |
7831 | } | |
7832 | else | |
7833 | { | |
7834 | emit_insn_before (gen_addsi3 (loop->iter_reg, | |
7835 | loop->iter_reg, | |
7836 | constm1_rtx), | |
7837 | loop->loop_end); | |
7838 | test = gen_rtx_NE (VOIDmode, loop->iter_reg, const0_rtx); | |
7839 | insn = emit_jump_insn_before (gen_cbranchsi4 (test, | |
7840 | loop->iter_reg, | |
7841 | const0_rtx, | |
7842 | loop->start_label), | |
7843 | loop->loop_end); | |
7844 | } | |
7845 | JUMP_LABEL (insn) = loop->start_label; | |
7846 | LABEL_NUSES (loop->start_label)++; | |
7847 | delete_insn (loop->loop_end); | |
7848 | } | |
7849 | ||
73dac59b CZ |
7850 | /* Return the next insn after INSN that is not a NOTE, but stop the |
7851 | search before we enter another basic block. This routine does not | |
7852 | look inside SEQUENCEs. */ | |
7853 | ||
7854 | static rtx_insn * | |
7855 | next_nonnote_insn_bb (rtx_insn *insn) | |
7856 | { | |
7857 | while (insn) | |
7858 | { | |
7859 | insn = NEXT_INSN (insn); | |
7860 | if (insn == 0 || !NOTE_P (insn)) | |
7861 | break; | |
7862 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
7863 | return NULL; | |
7864 | } | |
7865 | ||
7866 | return insn; | |
7867 | } | |
7868 | ||
a2de90a4 CZ |
7869 | /* Optimize LOOP. */ |
7870 | ||
7871 | static bool | |
7872 | hwloop_optimize (hwloop_info loop) | |
7873 | { | |
7874 | int i; | |
7875 | edge entry_edge; | |
7876 | basic_block entry_bb, bb; | |
4dea3bff DM |
7877 | rtx iter_reg; |
7878 | rtx_insn *insn, *seq, *entry_after, *last_insn, *end_label; | |
a2de90a4 CZ |
7879 | unsigned int length; |
7880 | bool need_fix = false; | |
7881 | rtx lp_reg = gen_rtx_REG (SImode, LP_COUNT); | |
7882 | ||
7883 | if (loop->depth > 1) | |
7884 | { | |
7885 | if (dump_file) | |
73dac59b CZ |
7886 | fprintf (dump_file, ";; loop %d is not innermost\n", |
7887 | loop->loop_no); | |
a2de90a4 CZ |
7888 | return false; |
7889 | } | |
7890 | ||
7891 | if (!loop->incoming_dest) | |
7892 | { | |
7893 | if (dump_file) | |
73dac59b CZ |
7894 | fprintf (dump_file, ";; loop %d has more than one entry\n", |
7895 | loop->loop_no); | |
a2de90a4 CZ |
7896 | return false; |
7897 | } | |
7898 | ||
7899 | if (loop->incoming_dest != loop->head) | |
7900 | { | |
7901 | if (dump_file) | |
73dac59b CZ |
7902 | fprintf (dump_file, ";; loop %d is not entered from head\n", |
7903 | loop->loop_no); | |
a2de90a4 CZ |
7904 | return false; |
7905 | } | |
7906 | ||
7907 | if (loop->has_call || loop->has_asm) | |
7908 | { | |
7909 | if (dump_file) | |
73dac59b CZ |
7910 | fprintf (dump_file, ";; loop %d has invalid insn\n", |
7911 | loop->loop_no); | |
a2de90a4 CZ |
7912 | return false; |
7913 | } | |
7914 | ||
7915 | /* Scan all the blocks to make sure they don't use iter_reg. */ | |
7916 | if (loop->iter_reg_used || loop->iter_reg_used_outside) | |
7917 | { | |
7918 | if (dump_file) | |
73dac59b CZ |
7919 | fprintf (dump_file, ";; loop %d uses iterator\n", |
7920 | loop->loop_no); | |
a2de90a4 CZ |
7921 | return false; |
7922 | } | |
7923 | ||
7924 | /* Check if start_label appears before doloop_end. */ | |
7925 | length = 0; | |
7926 | for (insn = loop->start_label; | |
7927 | insn && insn != loop->loop_end; | |
7928 | insn = NEXT_INSN (insn)) | |
dddc1815 CZ |
7929 | { |
7930 | length += NONDEBUG_INSN_P (insn) ? get_attr_length (insn) : 0; | |
7931 | if (JUMP_TABLES_IN_TEXT_SECTION | |
7932 | && JUMP_TABLE_DATA_P (insn)) | |
7933 | { | |
7934 | if (dump_file) | |
7935 | fprintf (dump_file, ";; loop %d has a jump table\n", | |
7936 | loop->loop_no); | |
7937 | return false; | |
7938 | } | |
7939 | } | |
a2de90a4 CZ |
7940 | |
7941 | if (!insn) | |
7942 | { | |
7943 | if (dump_file) | |
73dac59b CZ |
7944 | fprintf (dump_file, ";; loop %d start_label not before loop_end\n", |
7945 | loop->loop_no); | |
a2de90a4 CZ |
7946 | return false; |
7947 | } | |
7948 | ||
7949 | loop->length = length; | |
7950 | if (loop->length > ARC_MAX_LOOP_LENGTH) | |
7951 | { | |
7952 | if (dump_file) | |
7953 | fprintf (dump_file, ";; loop %d too long\n", loop->loop_no); | |
7954 | return false; | |
7955 | } | |
5b5905bb CZ |
7956 | else if (!loop->length) |
7957 | { | |
7958 | if (dump_file) | |
7959 | fprintf (dump_file, ";; loop %d is empty\n", loop->loop_no); | |
7960 | return false; | |
7961 | } | |
a2de90a4 | 7962 | |
73dac59b | 7963 | /* Check if we use a register or not. */ |
a2de90a4 CZ |
7964 | if (!REG_P (loop->iter_reg)) |
7965 | { | |
7966 | if (dump_file) | |
73dac59b CZ |
7967 | fprintf (dump_file, ";; loop %d iterator is MEM\n", |
7968 | loop->loop_no); | |
7969 | return false; | |
7970 | } | |
7971 | ||
7972 | /* Check if we use a register or not. */ | |
7973 | if (!REG_P (loop->iter_reg)) | |
7974 | { | |
7975 | if (dump_file) | |
7976 | fprintf (dump_file, ";; loop %d iterator is MEM\n", | |
7977 | loop->loop_no); | |
a2de90a4 CZ |
7978 | return false; |
7979 | } | |
7980 | ||
7981 | /* Check if loop register is lpcount. */ | |
7982 | if (REG_P (loop->iter_reg) && (REGNO (loop->iter_reg)) != LP_COUNT) | |
7983 | { | |
7984 | if (dump_file) | |
7985 | fprintf (dump_file, ";; loop %d doesn't use lp_count as loop" | |
7986 | " iterator\n", | |
7987 | loop->loop_no); | |
7988 | /* This loop doesn't use the lp_count, check though if we can | |
7989 | fix it. */ | |
7990 | if (TEST_HARD_REG_BIT (loop->regs_set_in_loop, LP_COUNT) | |
7991 | /* In very unique cases we may have LP_COUNT alive. */ | |
7992 | || (loop->incoming_src | |
7993 | && REGNO_REG_SET_P (df_get_live_out (loop->incoming_src), | |
7994 | LP_COUNT))) | |
73dac59b CZ |
7995 | { |
7996 | if (dump_file) | |
7997 | fprintf (dump_file, ";; loop %d, lp_count is alive", loop->loop_no); | |
7998 | return false; | |
7999 | } | |
a2de90a4 CZ |
8000 | else |
8001 | need_fix = true; | |
8002 | } | |
8003 | ||
8004 | /* Check for control like instruction as the last instruction of a | |
8005 | ZOL. */ | |
8006 | bb = loop->tail; | |
8007 | last_insn = PREV_INSN (loop->loop_end); | |
8008 | ||
8009 | while (1) | |
8010 | { | |
8011 | for (; last_insn != BB_HEAD (bb); | |
8012 | last_insn = PREV_INSN (last_insn)) | |
8013 | if (NONDEBUG_INSN_P (last_insn)) | |
8014 | break; | |
8015 | ||
8016 | if (last_insn != BB_HEAD (bb)) | |
8017 | break; | |
8018 | ||
8019 | if (single_pred_p (bb) | |
8020 | && single_pred_edge (bb)->flags & EDGE_FALLTHRU | |
8021 | && single_pred (bb) != ENTRY_BLOCK_PTR_FOR_FN (cfun)) | |
8022 | { | |
8023 | bb = single_pred (bb); | |
8024 | last_insn = BB_END (bb); | |
8025 | continue; | |
8026 | } | |
8027 | else | |
8028 | { | |
8029 | last_insn = NULL; | |
8030 | break; | |
8031 | } | |
8032 | } | |
8033 | ||
8034 | if (!last_insn) | |
8035 | { | |
8036 | if (dump_file) | |
8037 | fprintf (dump_file, ";; loop %d has no last instruction\n", | |
8038 | loop->loop_no); | |
8039 | return false; | |
8040 | } | |
8041 | ||
8042 | if ((TARGET_ARC600_FAMILY || TARGET_HS) | |
8043 | && INSN_P (last_insn) | |
8044 | && (JUMP_P (last_insn) || CALL_P (last_insn) | |
8045 | || GET_CODE (PATTERN (last_insn)) == SEQUENCE | |
5b5905bb CZ |
8046 | /* At this stage we can have (insn (clobber (mem:BLK |
8047 | (reg)))) instructions, ignore them. */ | |
8048 | || (GET_CODE (PATTERN (last_insn)) != CLOBBER | |
8049 | && (get_attr_type (last_insn) == TYPE_BRCC | |
8050 | || get_attr_type (last_insn) == TYPE_BRCC_NO_DELAY_SLOT)))) | |
a2de90a4 CZ |
8051 | { |
8052 | if (loop->length + 2 > ARC_MAX_LOOP_LENGTH) | |
8053 | { | |
8054 | if (dump_file) | |
8055 | fprintf (dump_file, ";; loop %d too long\n", loop->loop_no); | |
8056 | return false; | |
8057 | } | |
8058 | if (dump_file) | |
8059 | fprintf (dump_file, ";; loop %d has a control like last insn;" | |
8060 | "add a nop\n", | |
8061 | loop->loop_no); | |
8062 | ||
8063 | last_insn = emit_insn_after (gen_nopv (), last_insn); | |
8064 | } | |
8065 | ||
8066 | if (LABEL_P (last_insn)) | |
8067 | { | |
8068 | if (dump_file) | |
8069 | fprintf (dump_file, ";; loop %d has a label as last insn;" | |
8070 | "add a nop\n", | |
8071 | loop->loop_no); | |
8072 | last_insn = emit_insn_after (gen_nopv (), last_insn); | |
8073 | } | |
a0920243 CZ |
8074 | |
8075 | /* SAVE_NOTE is used by haifa scheduler. However, we are after it | |
8076 | and we can use it to indicate the last ZOL instruction cannot be | |
8077 | part of a delay slot. */ | |
8078 | add_reg_note (last_insn, REG_SAVE_NOTE, GEN_INT (2)); | |
8079 | ||
a2de90a4 CZ |
8080 | loop->last_insn = last_insn; |
8081 | ||
8082 | /* Get the loop iteration register. */ | |
8083 | iter_reg = loop->iter_reg; | |
8084 | ||
8085 | gcc_assert (REG_P (iter_reg)); | |
8086 | ||
8087 | entry_edge = NULL; | |
8088 | ||
8089 | FOR_EACH_VEC_SAFE_ELT (loop->incoming, i, entry_edge) | |
8090 | if (entry_edge->flags & EDGE_FALLTHRU) | |
8091 | break; | |
8092 | ||
8093 | if (entry_edge == NULL) | |
8094 | { | |
8095 | if (dump_file) | |
8096 | fprintf (dump_file, ";; loop %d has no fallthru edge jumping" | |
8097 | "into the loop\n", | |
8098 | loop->loop_no); | |
8099 | return false; | |
8100 | } | |
8101 | /* The loop is good. */ | |
8102 | end_label = gen_label_rtx (); | |
8103 | loop->end_label = end_label; | |
8104 | ||
8105 | /* Place the zero_cost_loop_start instruction before the loop. */ | |
8106 | entry_bb = entry_edge->src; | |
8107 | ||
8108 | start_sequence (); | |
8109 | ||
8110 | if (need_fix) | |
8111 | { | |
8112 | /* The loop uses a R-register, but the lp_count is free, thus | |
8113 | use lp_count. */ | |
73dac59b | 8114 | emit_insn (gen_rtx_SET (lp_reg, iter_reg)); |
a2de90a4 CZ |
8115 | SET_HARD_REG_BIT (loop->regs_set_in_loop, LP_COUNT); |
8116 | iter_reg = lp_reg; | |
8117 | if (dump_file) | |
8118 | { | |
8119 | fprintf (dump_file, ";; fix loop %d to use lp_count\n", | |
8120 | loop->loop_no); | |
8121 | } | |
8122 | } | |
8123 | ||
73dac59b | 8124 | insn = emit_insn (gen_arc_lp (loop->start_label, |
a2de90a4 CZ |
8125 | loop->end_label)); |
8126 | ||
8127 | seq = get_insns (); | |
8128 | end_sequence (); | |
8129 | ||
8130 | entry_after = BB_END (entry_bb); | |
8131 | if (!single_succ_p (entry_bb) || vec_safe_length (loop->incoming) > 1 | |
8132 | || !entry_after) | |
8133 | { | |
8134 | basic_block new_bb; | |
8135 | edge e; | |
8136 | edge_iterator ei; | |
8137 | ||
8138 | emit_insn_before (seq, BB_HEAD (loop->head)); | |
8139 | seq = emit_label_before (gen_label_rtx (), seq); | |
8140 | new_bb = create_basic_block (seq, insn, entry_bb); | |
8141 | FOR_EACH_EDGE (e, ei, loop->incoming) | |
73dac59b CZ |
8142 | { |
8143 | if (!(e->flags & EDGE_FALLTHRU)) | |
8144 | redirect_edge_and_branch_force (e, new_bb); | |
8145 | else | |
8146 | redirect_edge_succ (e, new_bb); | |
8147 | } | |
a2de90a4 CZ |
8148 | |
8149 | make_edge (new_bb, loop->head, 0); | |
8150 | } | |
8151 | else | |
8152 | { | |
8153 | #if 0 | |
8154 | while (DEBUG_INSN_P (entry_after) | |
73dac59b CZ |
8155 | || (NOTE_P (entry_after) |
8156 | && NOTE_KIND (entry_after) != NOTE_INSN_BASIC_BLOCK | |
8157 | /* Make sure we don't split a call and its corresponding | |
8158 | CALL_ARG_LOCATION note. */ | |
8159 | && NOTE_KIND (entry_after) != NOTE_INSN_CALL_ARG_LOCATION)) | |
a2de90a4 CZ |
8160 | entry_after = NEXT_INSN (entry_after); |
8161 | #endif | |
73dac59b | 8162 | entry_after = next_nonnote_insn_bb (entry_after); |
a2de90a4 CZ |
8163 | |
8164 | gcc_assert (entry_after); | |
8165 | emit_insn_before (seq, entry_after); | |
8166 | } | |
8167 | ||
a2de90a4 CZ |
8168 | /* Insert the loop end label before the last instruction of the |
8169 | loop. */ | |
8170 | emit_label_after (end_label, loop->last_insn); | |
5d4c34aa CZ |
8171 | /* Make sure we mark the begining and end label as used. */ |
8172 | LABEL_NUSES (loop->end_label)++; | |
8173 | LABEL_NUSES (loop->start_label)++; | |
a2de90a4 CZ |
8174 | |
8175 | return true; | |
8176 | } | |
8177 | ||
8178 | /* A callback for the hw-doloop pass. This function examines INSN; if | |
8179 | it is a loop_end pattern we recognize, return the reg rtx for the | |
8180 | loop counter. Otherwise, return NULL_RTX. */ | |
8181 | ||
8182 | static rtx | |
8183 | hwloop_pattern_reg (rtx_insn *insn) | |
8184 | { | |
8185 | rtx reg; | |
8186 | ||
8187 | if (!JUMP_P (insn) || recog_memoized (insn) != CODE_FOR_loop_end) | |
8188 | return NULL_RTX; | |
8189 | ||
8190 | reg = SET_DEST (XVECEXP (PATTERN (insn), 0, 1)); | |
8191 | if (!REG_P (reg)) | |
8192 | return NULL_RTX; | |
8193 | return reg; | |
8194 | } | |
8195 | ||
8196 | static struct hw_doloop_hooks arc_doloop_hooks = | |
8197 | { | |
8198 | hwloop_pattern_reg, | |
8199 | hwloop_optimize, | |
8200 | hwloop_fail | |
8201 | }; | |
8202 | ||
8203 | /* Run from machine_dependent_reorg, this pass looks for doloop_end insns | |
8204 | and tries to rewrite the RTL of these loops so that proper Blackfin | |
8205 | hardware loops are generated. */ | |
8206 | ||
8207 | static void | |
8208 | arc_reorg_loops (void) | |
8209 | { | |
8210 | reorg_loops (true, &arc_doloop_hooks); | |
8211 | } | |
8212 | ||
6b55f8c9 CZ |
8213 | /* Scan all calls and add symbols to be emitted in the jli section if |
8214 | needed. */ | |
8215 | ||
8216 | static void | |
8217 | jli_call_scan (void) | |
8218 | { | |
8219 | rtx_insn *insn; | |
8220 | ||
8221 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
8222 | { | |
8223 | if (!CALL_P (insn)) | |
8224 | continue; | |
8225 | ||
8226 | rtx pat = PATTERN (insn); | |
8227 | if (GET_CODE (pat) == COND_EXEC) | |
8228 | pat = COND_EXEC_CODE (pat); | |
8229 | pat = XVECEXP (pat, 0, 0); | |
8230 | if (GET_CODE (pat) == SET) | |
8231 | pat = SET_SRC (pat); | |
8232 | ||
8233 | pat = XEXP (XEXP (pat, 0), 0); | |
8234 | if (GET_CODE (pat) == SYMBOL_REF | |
8235 | && arc_is_jli_call_p (pat)) | |
8236 | arc_add_jli_section (pat); | |
8237 | } | |
8238 | } | |
8239 | ||
16493b57 CZ |
8240 | /* Add padding if necessary to avoid a mispredict. A return could |
8241 | happen immediately after the function start. A call/return and | |
8242 | return/return must be 6 bytes apart to avoid mispredict. */ | |
8243 | ||
8244 | static void | |
8245 | pad_return (void) | |
8246 | { | |
8247 | rtx_insn *insn; | |
8248 | long offset; | |
8249 | ||
8250 | if (!TARGET_PAD_RETURN) | |
8251 | return; | |
8252 | ||
8253 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
8254 | { | |
8255 | rtx_insn *prev0 = prev_active_insn (insn); | |
8256 | bool wantlong = false; | |
8257 | ||
8258 | if (!INSN_P (insn) || GET_CODE (PATTERN (insn)) != SIMPLE_RETURN) | |
8259 | continue; | |
8260 | ||
8261 | if (!prev0) | |
8262 | { | |
8263 | prev0 = emit_insn_before (gen_nopv (), insn); | |
8264 | /* REG_SAVE_NOTE is used by Haifa scheduler, we are in reorg | |
8265 | so it is safe to reuse it for forcing a particular length | |
8266 | for an instruction. */ | |
8267 | add_reg_note (prev0, REG_SAVE_NOTE, GEN_INT (1)); | |
8268 | emit_insn_before (gen_nopv (), insn); | |
8269 | continue; | |
8270 | } | |
8271 | offset = get_attr_length (prev0); | |
8272 | ||
8273 | if (get_attr_length (prev0) == 2 | |
8274 | && get_attr_iscompact (prev0) != ISCOMPACT_TRUE) | |
8275 | { | |
8276 | /* Force long version of the insn. */ | |
8277 | wantlong = true; | |
8278 | offset += 2; | |
8279 | } | |
8280 | ||
8281 | rtx_insn *prev = prev_active_insn (prev0); | |
8282 | if (prev) | |
8283 | offset += get_attr_length (prev); | |
8284 | ||
8285 | prev = prev_active_insn (prev); | |
8286 | if (prev) | |
8287 | offset += get_attr_length (prev); | |
8288 | ||
8289 | switch (offset) | |
8290 | { | |
8291 | case 2: | |
8292 | prev = emit_insn_before (gen_nopv (), insn); | |
8293 | add_reg_note (prev, REG_SAVE_NOTE, GEN_INT (1)); | |
8294 | break; | |
8295 | case 4: | |
8296 | emit_insn_before (gen_nopv (), insn); | |
8297 | break; | |
8298 | default: | |
8299 | continue; | |
8300 | } | |
8301 | ||
8302 | if (wantlong) | |
8303 | add_reg_note (prev0, REG_SAVE_NOTE, GEN_INT (1)); | |
8304 | ||
8305 | /* Emit a blockage to avoid delay slot scheduling. */ | |
8306 | emit_insn_before (gen_blockage (), insn); | |
8307 | } | |
8308 | } | |
8309 | ||
526b7aee SV |
8310 | static int arc_reorg_in_progress = 0; |
8311 | ||
8312 | /* ARC's machince specific reorg function. */ | |
8313 | ||
8314 | static void | |
8315 | arc_reorg (void) | |
8316 | { | |
b3458f61 DM |
8317 | rtx_insn *insn; |
8318 | rtx pattern; | |
526b7aee SV |
8319 | rtx pc_target; |
8320 | long offset; | |
8321 | int changed; | |
8322 | ||
8323 | cfun->machine->arc_reorg_started = 1; | |
8324 | arc_reorg_in_progress = 1; | |
8325 | ||
a2de90a4 | 8326 | compute_bb_for_insn (); |
526b7aee | 8327 | |
a2de90a4 | 8328 | df_analyze (); |
526b7aee | 8329 | |
a2de90a4 CZ |
8330 | /* Doloop optimization. */ |
8331 | arc_reorg_loops (); | |
526b7aee | 8332 | |
a2de90a4 | 8333 | workaround_arc_anomaly (); |
6b55f8c9 | 8334 | jli_call_scan (); |
16493b57 | 8335 | pad_return (); |
526b7aee SV |
8336 | |
8337 | /* FIXME: should anticipate ccfsm action, generate special patterns for | |
8338 | to-be-deleted branches that have no delay slot and have at least the | |
8339 | length of the size increase forced on other insns that are conditionalized. | |
8340 | This can also have an insn_list inside that enumerates insns which are | |
8341 | not actually conditionalized because the destinations are dead in the | |
8342 | not-execute case. | |
8343 | Could also tag branches that we want to be unaligned if they get no delay | |
8344 | slot, or even ones that we don't want to do delay slot sheduling for | |
8345 | because we can unalign them. | |
8346 | ||
8347 | However, there are cases when conditional execution is only possible after | |
8348 | delay slot scheduling: | |
8349 | ||
8350 | - If a delay slot is filled with a nocond/set insn from above, the previous | |
8351 | basic block can become elegible for conditional execution. | |
8352 | - If a delay slot is filled with a nocond insn from the fall-through path, | |
8353 | the branch with that delay slot can become eligble for conditional | |
8354 | execution (however, with the same sort of data flow analysis that dbr | |
8355 | does, we could have figured out before that we don't need to | |
8356 | conditionalize this insn.) | |
8357 | - If a delay slot insn is filled with an insn from the target, the | |
8358 | target label gets its uses decremented (even deleted if falling to zero), | |
8359 | thus possibly creating more condexec opportunities there. | |
8360 | Therefore, we should still be prepared to apply condexec optimization on | |
8361 | non-prepared branches if the size increase of conditionalized insns is no | |
8362 | more than the size saved from eliminating the branch. An invocation option | |
8363 | could also be used to reserve a bit of extra size for condbranches so that | |
8364 | this'll work more often (could also test in arc_reorg if the block is | |
8365 | 'close enough' to be eligible for condexec to make this likely, and | |
8366 | estimate required size increase). */ | |
8367 | /* Generate BRcc insns, by combining cmp and Bcc insns wherever possible. */ | |
8368 | if (TARGET_NO_BRCC_SET) | |
8369 | return; | |
8370 | ||
8371 | do | |
8372 | { | |
8373 | init_insn_lengths(); | |
8374 | changed = 0; | |
8375 | ||
8376 | if (optimize > 1 && !TARGET_NO_COND_EXEC) | |
8377 | { | |
8378 | arc_ifcvt (); | |
8379 | unsigned int flags = pass_data_arc_ifcvt.todo_flags_finish; | |
8380 | df_finish_pass ((flags & TODO_df_verify) != 0); | |
782bdf21 CZ |
8381 | |
8382 | if (dump_file) | |
8383 | { | |
8384 | fprintf (dump_file, ";; After if conversion:\n\n"); | |
8385 | print_rtl (dump_file, get_insns ()); | |
8386 | } | |
526b7aee SV |
8387 | } |
8388 | ||
8389 | /* Call shorten_branches to calculate the insn lengths. */ | |
8390 | shorten_branches (get_insns()); | |
8391 | cfun->machine->ccfsm_current_insn = NULL_RTX; | |
8392 | ||
8393 | if (!INSN_ADDRESSES_SET_P()) | |
2fa9c1f6 CZ |
8394 | fatal_error (input_location, |
8395 | "insn addresses not set after shorten_branches"); | |
526b7aee SV |
8396 | |
8397 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
8398 | { | |
8399 | rtx label; | |
8400 | enum attr_type insn_type; | |
8401 | ||
8402 | /* If a non-jump insn (or a casesi jump table), continue. */ | |
8403 | if (GET_CODE (insn) != JUMP_INSN || | |
8404 | GET_CODE (PATTERN (insn)) == ADDR_VEC | |
8405 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
8406 | continue; | |
8407 | ||
8408 | /* If we already have a brcc, note if it is suitable for brcc_s. | |
8409 | Be a bit generous with the brcc_s range so that we can take | |
8410 | advantage of any code shortening from delay slot scheduling. */ | |
8411 | if (recog_memoized (insn) == CODE_FOR_cbranchsi4_scratch) | |
8412 | { | |
8413 | rtx pat = PATTERN (insn); | |
8414 | rtx op = XEXP (SET_SRC (XVECEXP (pat, 0, 0)), 0); | |
8415 | rtx *ccp = &XEXP (XVECEXP (pat, 0, 1), 0); | |
8416 | ||
8417 | offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); | |
8418 | if ((offset >= -140 && offset < 140) | |
8419 | && rtx_equal_p (XEXP (op, 1), const0_rtx) | |
8420 | && compact_register_operand (XEXP (op, 0), VOIDmode) | |
8421 | && equality_comparison_operator (op, VOIDmode)) | |
8422 | PUT_MODE (*ccp, CC_Zmode); | |
8423 | else if (GET_MODE (*ccp) == CC_Zmode) | |
8424 | PUT_MODE (*ccp, CC_ZNmode); | |
8425 | continue; | |
8426 | } | |
8427 | if ((insn_type = get_attr_type (insn)) == TYPE_BRCC | |
8428 | || insn_type == TYPE_BRCC_NO_DELAY_SLOT) | |
8429 | continue; | |
8430 | ||
8431 | /* OK. so we have a jump insn. */ | |
8432 | /* We need to check that it is a bcc. */ | |
8433 | /* Bcc => set (pc) (if_then_else ) */ | |
8434 | pattern = PATTERN (insn); | |
8435 | if (GET_CODE (pattern) != SET | |
8436 | || GET_CODE (SET_SRC (pattern)) != IF_THEN_ELSE | |
8437 | || ANY_RETURN_P (XEXP (SET_SRC (pattern), 1))) | |
8438 | continue; | |
8439 | ||
8440 | /* Now check if the jump is beyond the s9 range. */ | |
339ba33b | 8441 | if (CROSSING_JUMP_P (insn)) |
526b7aee SV |
8442 | continue; |
8443 | offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); | |
8444 | ||
8445 | if(offset > 253 || offset < -254) | |
8446 | continue; | |
8447 | ||
8448 | pc_target = SET_SRC (pattern); | |
8449 | ||
8f3304d0 CZ |
8450 | /* Avoid FPU instructions. */ |
8451 | if ((GET_MODE (XEXP (XEXP (pc_target, 0), 0)) == CC_FPUmode) | |
8452 | || (GET_MODE (XEXP (XEXP (pc_target, 0), 0)) == CC_FPU_UNEQmode)) | |
8453 | continue; | |
8454 | ||
526b7aee SV |
8455 | /* Now go back and search for the set cc insn. */ |
8456 | ||
8457 | label = XEXP (pc_target, 1); | |
8458 | ||
8459 | { | |
b3458f61 DM |
8460 | rtx pat; |
8461 | rtx_insn *scan, *link_insn = NULL; | |
526b7aee SV |
8462 | |
8463 | for (scan = PREV_INSN (insn); | |
8464 | scan && GET_CODE (scan) != CODE_LABEL; | |
8465 | scan = PREV_INSN (scan)) | |
8466 | { | |
8467 | if (! INSN_P (scan)) | |
8468 | continue; | |
8469 | pat = PATTERN (scan); | |
8470 | if (GET_CODE (pat) == SET | |
8471 | && cc_register (SET_DEST (pat), VOIDmode)) | |
8472 | { | |
8473 | link_insn = scan; | |
8474 | break; | |
8475 | } | |
8476 | } | |
8f3304d0 | 8477 | if (!link_insn) |
526b7aee SV |
8478 | continue; |
8479 | else | |
526b7aee | 8480 | { |
635aeaa2 | 8481 | /* Check if this is a data dependency. */ |
526b7aee SV |
8482 | rtx op, cc_clob_rtx, op0, op1, brcc_insn, note; |
8483 | rtx cmp0, cmp1; | |
8484 | ||
635aeaa2 CZ |
8485 | /* Make sure we can use it for brcc insns. */ |
8486 | if (find_reg_note (link_insn, REG_SAVE_NOTE, GEN_INT (3))) | |
8487 | continue; | |
8488 | ||
526b7aee SV |
8489 | /* Ok this is the set cc. copy args here. */ |
8490 | op = XEXP (pc_target, 0); | |
8491 | ||
8492 | op0 = cmp0 = XEXP (SET_SRC (pat), 0); | |
8493 | op1 = cmp1 = XEXP (SET_SRC (pat), 1); | |
8494 | if (GET_CODE (op0) == ZERO_EXTRACT | |
8495 | && XEXP (op0, 1) == const1_rtx | |
8496 | && (GET_CODE (op) == EQ | |
8497 | || GET_CODE (op) == NE)) | |
8498 | { | |
8499 | /* btst / b{eq,ne} -> bbit{0,1} */ | |
8500 | op0 = XEXP (cmp0, 0); | |
8501 | op1 = XEXP (cmp0, 2); | |
8502 | } | |
8503 | else if (!register_operand (op0, VOIDmode) | |
8504 | || !general_operand (op1, VOIDmode)) | |
8505 | continue; | |
8506 | /* Be careful not to break what cmpsfpx_raw is | |
8507 | trying to create for checking equality of | |
8508 | single-precision floats. */ | |
8509 | else if (TARGET_SPFP | |
8510 | && GET_MODE (op0) == SFmode | |
8511 | && GET_MODE (op1) == SFmode) | |
8512 | continue; | |
8513 | ||
8514 | /* None of the two cmp operands should be set between the | |
8515 | cmp and the branch. */ | |
8516 | if (reg_set_between_p (op0, link_insn, insn)) | |
8517 | continue; | |
8518 | ||
8519 | if (reg_set_between_p (op1, link_insn, insn)) | |
8520 | continue; | |
8521 | ||
8522 | /* Since the MODE check does not work, check that this is | |
8523 | CC reg's last set location before insn, and also no | |
8524 | instruction between the cmp and branch uses the | |
8525 | condition codes. */ | |
8526 | if ((reg_set_between_p (SET_DEST (pat), link_insn, insn)) | |
8527 | || (reg_used_between_p (SET_DEST (pat), link_insn, insn))) | |
8528 | continue; | |
8529 | ||
8530 | /* CC reg should be dead after insn. */ | |
8531 | if (!find_regno_note (insn, REG_DEAD, CC_REG)) | |
8532 | continue; | |
8533 | ||
8534 | op = gen_rtx_fmt_ee (GET_CODE (op), | |
8535 | GET_MODE (op), cmp0, cmp1); | |
8536 | /* If we create a LIMM where there was none before, | |
8537 | we only benefit if we can avoid a scheduling bubble | |
8538 | for the ARC600. Otherwise, we'd only forgo chances | |
8539 | at short insn generation, and risk out-of-range | |
8540 | branches. */ | |
8541 | if (!brcc_nolimm_operator (op, VOIDmode) | |
8542 | && !long_immediate_operand (op1, VOIDmode) | |
8543 | && (TARGET_ARC700 | |
8544 | || next_active_insn (link_insn) != insn)) | |
8545 | continue; | |
8546 | ||
8547 | /* Emit bbit / brcc (or brcc_s if possible). | |
8548 | CC_Zmode indicates that brcc_s is possible. */ | |
8549 | ||
8550 | if (op0 != cmp0) | |
8551 | cc_clob_rtx = gen_rtx_REG (CC_ZNmode, CC_REG); | |
8552 | else if ((offset >= -140 && offset < 140) | |
8553 | && rtx_equal_p (op1, const0_rtx) | |
8554 | && compact_register_operand (op0, VOIDmode) | |
8555 | && (GET_CODE (op) == EQ | |
8556 | || GET_CODE (op) == NE)) | |
8557 | cc_clob_rtx = gen_rtx_REG (CC_Zmode, CC_REG); | |
8558 | else | |
8559 | cc_clob_rtx = gen_rtx_REG (CCmode, CC_REG); | |
8560 | ||
8561 | brcc_insn | |
8562 | = gen_rtx_IF_THEN_ELSE (VOIDmode, op, label, pc_rtx); | |
f7df4a84 | 8563 | brcc_insn = gen_rtx_SET (pc_rtx, brcc_insn); |
526b7aee SV |
8564 | cc_clob_rtx = gen_rtx_CLOBBER (VOIDmode, cc_clob_rtx); |
8565 | brcc_insn | |
8566 | = gen_rtx_PARALLEL | |
8567 | (VOIDmode, gen_rtvec (2, brcc_insn, cc_clob_rtx)); | |
8568 | brcc_insn = emit_jump_insn_before (brcc_insn, insn); | |
8569 | ||
8570 | JUMP_LABEL (brcc_insn) = JUMP_LABEL (insn); | |
8571 | note = find_reg_note (insn, REG_BR_PROB, 0); | |
8572 | if (note) | |
8573 | { | |
8574 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
8575 | REG_NOTES (brcc_insn) = note; | |
8576 | } | |
8577 | note = find_reg_note (link_insn, REG_DEAD, op0); | |
8578 | if (note) | |
8579 | { | |
8580 | remove_note (link_insn, note); | |
8581 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
8582 | REG_NOTES (brcc_insn) = note; | |
8583 | } | |
8584 | note = find_reg_note (link_insn, REG_DEAD, op1); | |
8585 | if (note) | |
8586 | { | |
8587 | XEXP (note, 1) = REG_NOTES (brcc_insn); | |
8588 | REG_NOTES (brcc_insn) = note; | |
8589 | } | |
8590 | ||
8591 | changed = 1; | |
8592 | ||
8593 | /* Delete the bcc insn. */ | |
8594 | set_insn_deleted (insn); | |
8595 | ||
8596 | /* Delete the cmp insn. */ | |
8597 | set_insn_deleted (link_insn); | |
8598 | ||
8599 | } | |
8600 | } | |
8601 | } | |
8602 | /* Clear out insn_addresses. */ | |
8603 | INSN_ADDRESSES_FREE (); | |
8604 | ||
8605 | } while (changed); | |
8606 | ||
8607 | if (INSN_ADDRESSES_SET_P()) | |
40fecdd6 | 8608 | fatal_error (input_location, "insn addresses not freed"); |
526b7aee SV |
8609 | |
8610 | arc_reorg_in_progress = 0; | |
8611 | } | |
8612 | ||
8613 | /* Check if the operands are valid for BRcc.d generation | |
8614 | Valid Brcc.d patterns are | |
8615 | Brcc.d b, c, s9 | |
8616 | Brcc.d b, u6, s9 | |
8617 | ||
67914693 | 8618 | For cc={GT, LE, GTU, LEU}, u6=63 cannot be allowed, |
526b7aee SV |
8619 | since they are encoded by the assembler as {GE, LT, HS, LS} 64, which |
8620 | does not have a delay slot | |
8621 | ||
8622 | Assumed precondition: Second operand is either a register or a u6 value. */ | |
8623 | ||
8624 | bool | |
8625 | valid_brcc_with_delay_p (rtx *operands) | |
8626 | { | |
8627 | if (optimize_size && GET_MODE (operands[4]) == CC_Zmode) | |
8628 | return false; | |
8629 | return brcc_nolimm_operator (operands[0], VOIDmode); | |
8630 | } | |
8631 | ||
526b7aee SV |
8632 | /* Implement TARGET_IN_SMALL_DATA_P. Return true if it would be safe to |
8633 | access DECL using %gp_rel(...)($gp). */ | |
8634 | ||
8635 | static bool | |
8636 | arc_in_small_data_p (const_tree decl) | |
8637 | { | |
8638 | HOST_WIDE_INT size; | |
8180c03f | 8639 | tree attr; |
526b7aee | 8640 | |
9f532472 CZ |
8641 | /* Only variables are going into small data area. */ |
8642 | if (TREE_CODE (decl) != VAR_DECL) | |
526b7aee SV |
8643 | return false; |
8644 | ||
526b7aee SV |
8645 | if (TARGET_NO_SDATA_SET) |
8646 | return false; | |
8647 | ||
526b7aee SV |
8648 | /* Disable sdata references to weak variables. */ |
8649 | if (DECL_WEAK (decl)) | |
8650 | return false; | |
8651 | ||
9f532472 CZ |
8652 | /* Don't put constants into the small data section: we want them to |
8653 | be in ROM rather than RAM. */ | |
8654 | if (TREE_READONLY (decl)) | |
8655 | return false; | |
8656 | ||
8657 | /* To ensure -mvolatile-cache works ld.di does not have a | |
8658 | gp-relative variant. */ | |
8659 | if (!TARGET_VOLATILE_CACHE_SET | |
8660 | && TREE_THIS_VOLATILE (decl)) | |
8661 | return false; | |
526b7aee | 8662 | |
8180c03f CZ |
8663 | /* Likewise for uncached data. */ |
8664 | attr = TYPE_ATTRIBUTES (TREE_TYPE (decl)); | |
8665 | if (lookup_attribute ("uncached", attr)) | |
8666 | return false; | |
8667 | ||
b6fb257b CZ |
8668 | /* and for aux regs. */ |
8669 | attr = DECL_ATTRIBUTES (decl); | |
8670 | if (lookup_attribute ("aux", attr)) | |
8671 | return false; | |
8672 | ||
9f532472 CZ |
8673 | if (DECL_SECTION_NAME (decl) != 0) |
8674 | { | |
8675 | const char *name = DECL_SECTION_NAME (decl); | |
8676 | if (strcmp (name, ".sdata") == 0 | |
8677 | || strcmp (name, ".sbss") == 0) | |
8678 | return true; | |
8679 | } | |
8680 | /* If it's not public, there's no need to put it in the small data | |
8681 | section. */ | |
8682 | else if (TREE_PUBLIC (decl)) | |
8683 | { | |
8684 | size = int_size_in_bytes (TREE_TYPE (decl)); | |
8685 | return (size > 0 && size <= g_switch_value); | |
8686 | } | |
8687 | return false; | |
526b7aee SV |
8688 | } |
8689 | ||
526b7aee SV |
8690 | /* Return true if OP is an acceptable memory operand for ARCompact |
8691 | 16-bit gp-relative load instructions. | |
e0be3321 | 8692 | */ |
526b7aee SV |
8693 | /* volatile cache option still to be handled. */ |
8694 | ||
8695 | bool | |
b6fb7933 | 8696 | compact_sda_memory_operand (rtx op, machine_mode mode, bool short_p) |
526b7aee SV |
8697 | { |
8698 | rtx addr; | |
8699 | int size; | |
b6fb7933 CZ |
8700 | int align = 0; |
8701 | int mask = 0; | |
526b7aee SV |
8702 | |
8703 | /* Eliminate non-memory operations. */ | |
8704 | if (GET_CODE (op) != MEM) | |
8705 | return false; | |
8706 | ||
8707 | if (mode == VOIDmode) | |
8708 | mode = GET_MODE (op); | |
8709 | ||
8710 | size = GET_MODE_SIZE (mode); | |
8711 | ||
8712 | /* dword operations really put out 2 instructions, so eliminate them. */ | |
8713 | if (size > UNITS_PER_WORD) | |
8714 | return false; | |
8715 | ||
8716 | /* Decode the address now. */ | |
8717 | addr = XEXP (op, 0); | |
8718 | ||
e0be3321 | 8719 | if (!legitimate_small_data_address_p (addr)) |
b6fb7933 CZ |
8720 | return false; |
8721 | ||
8722 | if (!short_p || size == 1) | |
8723 | return true; | |
8724 | ||
8725 | /* Now check for the alignment, the short loads using gp require the | |
8726 | addresses to be aligned. */ | |
e0be3321 | 8727 | align = get_symbol_alignment (addr); |
b6fb7933 CZ |
8728 | switch (mode) |
8729 | { | |
8730 | case E_HImode: | |
8731 | mask = 1; | |
8732 | break; | |
8733 | default: | |
8734 | mask = 3; | |
8735 | break; | |
8736 | } | |
8737 | ||
8738 | if (align && ((align & mask) == 0)) | |
8739 | return true; | |
8740 | return false; | |
526b7aee SV |
8741 | } |
8742 | ||
b6fb257b CZ |
8743 | /* Return TRUE if PAT is accessing an aux-reg. */ |
8744 | ||
8745 | static bool | |
8746 | arc_is_aux_reg_p (rtx pat) | |
8747 | { | |
8748 | tree attrs = NULL_TREE; | |
8749 | tree addr; | |
8750 | ||
8751 | if (!MEM_P (pat)) | |
8752 | return false; | |
8753 | ||
8754 | /* Get the memory attributes. */ | |
8755 | addr = MEM_EXPR (pat); | |
8756 | if (!addr) | |
8757 | return false; | |
8758 | ||
8759 | /* Get the attributes. */ | |
8760 | if (TREE_CODE (addr) == VAR_DECL) | |
8761 | attrs = DECL_ATTRIBUTES (addr); | |
8762 | else if (TREE_CODE (addr) == MEM_REF) | |
8763 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (TREE_OPERAND (addr, 0))); | |
8764 | else | |
8765 | return false; | |
8766 | ||
8767 | if (lookup_attribute ("aux", attrs)) | |
8768 | return true; | |
8769 | return false; | |
8770 | } | |
8771 | ||
526b7aee SV |
8772 | /* Implement ASM_OUTPUT_ALIGNED_DECL_LOCAL. */ |
8773 | ||
8774 | void | |
8775 | arc_asm_output_aligned_decl_local (FILE * stream, tree decl, const char * name, | |
8776 | unsigned HOST_WIDE_INT size, | |
8777 | unsigned HOST_WIDE_INT align, | |
8778 | unsigned HOST_WIDE_INT globalize_p) | |
8779 | { | |
b6fb257b CZ |
8780 | int in_small_data = arc_in_small_data_p (decl); |
8781 | rtx mem = decl == NULL_TREE ? NULL_RTX : DECL_RTL (decl); | |
8782 | ||
8783 | /* Don't output aux-reg symbols. */ | |
8784 | if (mem != NULL_RTX && MEM_P (mem) | |
8785 | && SYMBOL_REF_P (XEXP (mem, 0)) | |
8786 | && arc_is_aux_reg_p (mem)) | |
8787 | return; | |
526b7aee SV |
8788 | |
8789 | if (in_small_data) | |
8790 | switch_to_section (get_named_section (NULL, ".sbss", 0)); | |
8791 | /* named_section (0,".sbss",0); */ | |
8792 | else | |
8793 | switch_to_section (bss_section); | |
8794 | ||
8795 | if (globalize_p) | |
8796 | (*targetm.asm_out.globalize_label) (stream, name); | |
8797 | ||
8798 | ASM_OUTPUT_ALIGN (stream, floor_log2 ((align) / BITS_PER_UNIT)); | |
8799 | ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "object"); | |
8800 | ASM_OUTPUT_SIZE_DIRECTIVE (stream, name, size); | |
8801 | ASM_OUTPUT_LABEL (stream, name); | |
8802 | ||
8803 | if (size != 0) | |
8804 | ASM_OUTPUT_SKIP (stream, size); | |
8805 | } | |
8806 | ||
526b7aee SV |
8807 | static bool |
8808 | arc_preserve_reload_p (rtx in) | |
8809 | { | |
8810 | return (GET_CODE (in) == PLUS | |
8811 | && RTX_OK_FOR_BASE_P (XEXP (in, 0), true) | |
8812 | && CONST_INT_P (XEXP (in, 1)) | |
8813 | && !((INTVAL (XEXP (in, 1)) & 511))); | |
8814 | } | |
8815 | ||
b9bc3b12 CZ |
8816 | /* Implement TARGET_REGISTER_MOVE_COST. */ |
8817 | ||
8818 | static int | |
ef4bddc2 | 8819 | arc_register_move_cost (machine_mode, |
b9bc3b12 | 8820 | reg_class_t from_class, reg_class_t to_class) |
526b7aee | 8821 | { |
526b7aee | 8822 | /* Force an attempt to 'mov Dy,Dx' to spill. */ |
c4014855 | 8823 | if ((TARGET_ARC700 || TARGET_EM) && TARGET_DPFP |
526b7aee SV |
8824 | && from_class == DOUBLE_REGS && to_class == DOUBLE_REGS) |
8825 | return 100; | |
8826 | ||
8827 | return 2; | |
8828 | } | |
8829 | ||
8830 | /* Emit code for an addsi3 instruction with OPERANDS. | |
8831 | COND_P indicates if this will use conditional execution. | |
8832 | Return the length of the instruction. | |
8833 | If OUTPUT_P is false, don't actually output the instruction, just return | |
8834 | its length. */ | |
8835 | int | |
8836 | arc_output_addsi (rtx *operands, bool cond_p, bool output_p) | |
8837 | { | |
3bbe0b82 | 8838 | char format[35]; |
526b7aee SV |
8839 | |
8840 | int match = operands_match_p (operands[0], operands[1]); | |
8841 | int match2 = operands_match_p (operands[0], operands[2]); | |
8842 | int intval = (REG_P (operands[2]) ? 1 | |
8843 | : CONST_INT_P (operands[2]) ? INTVAL (operands[2]) : 0xbadc057); | |
8844 | int neg_intval = -intval; | |
8845 | int short_0 = satisfies_constraint_Rcq (operands[0]); | |
8846 | int short_p = (!cond_p && short_0 && satisfies_constraint_Rcq (operands[1])); | |
8847 | int ret = 0; | |
8848 | ||
a0caeef6 CZ |
8849 | #define REG_H_P(OP) (REG_P (OP) && ((TARGET_V2 && REGNO (OP) <= 31 \ |
8850 | && REGNO (OP) != 30) \ | |
8851 | || !TARGET_V2)) | |
8852 | ||
526b7aee SV |
8853 | #define ADDSI_OUTPUT1(FORMAT) do {\ |
8854 | if (output_p) \ | |
8855 | output_asm_insn (FORMAT, operands);\ | |
8856 | return ret; \ | |
8857 | } while (0) | |
8858 | #define ADDSI_OUTPUT(LIST) do {\ | |
8859 | if (output_p) \ | |
8860 | sprintf LIST;\ | |
8861 | ADDSI_OUTPUT1 (format);\ | |
8862 | return ret; \ | |
8863 | } while (0) | |
8864 | ||
8865 | /* First try to emit a 16 bit insn. */ | |
8866 | ret = 2; | |
8867 | if (!cond_p | |
8868 | /* If we are actually about to output this insn, don't try a 16 bit | |
8869 | variant if we already decided that we don't want that | |
8870 | (I.e. we upsized this insn to align some following insn.) | |
8871 | E.g. add_s r0,sp,70 is 16 bit, but add r0,sp,70 requires a LIMM - | |
8872 | but add1 r0,sp,35 doesn't. */ | |
8873 | && (!output_p || (get_attr_length (current_output_insn) & 2))) | |
8874 | { | |
a0caeef6 CZ |
8875 | /* Generate add_s a,b,c; add_s b,b,u7; add_s c,b,u3; add_s b,b,h |
8876 | patterns. */ | |
526b7aee | 8877 | if (short_p |
a0caeef6 CZ |
8878 | && ((REG_H_P (operands[2]) |
8879 | && (match || satisfies_constraint_Rcq (operands[2]))) | |
8880 | || (CONST_INT_P (operands[2]) | |
8881 | && ((unsigned) intval <= (match ? 127 : 7))))) | |
8882 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;1"); | |
8883 | ||
8884 | /* Generate add_s b,b,h patterns. */ | |
8885 | if (short_0 && match2 && REG_H_P (operands[1])) | |
8886 | ADDSI_OUTPUT1 ("add%? %0,%2,%1 ;2"); | |
8887 | ||
8888 | /* Generate add_s b,sp,u7; add_s sp,sp,u7 patterns. */ | |
526b7aee SV |
8889 | if ((short_0 || REGNO (operands[0]) == STACK_POINTER_REGNUM) |
8890 | && REGNO (operands[1]) == STACK_POINTER_REGNUM && !(intval & ~124)) | |
a0caeef6 | 8891 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;3"); |
526b7aee SV |
8892 | |
8893 | if ((short_p && (unsigned) neg_intval <= (match ? 31 : 7)) | |
8894 | || (REGNO (operands[0]) == STACK_POINTER_REGNUM | |
8895 | && match && !(neg_intval & ~124))) | |
a0caeef6 | 8896 | ADDSI_OUTPUT1 ("sub%? %0,%1,%n2 ;4"); |
fa9c1b3c | 8897 | |
a0caeef6 CZ |
8898 | /* Generate add_s h,h,s3 patterns. */ |
8899 | if (REG_H_P (operands[0]) && match && TARGET_V2 | |
8900 | && CONST_INT_P (operands[2]) && ((intval>= -1) && (intval <= 6))) | |
8901 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;5"); | |
fa9c1b3c | 8902 | |
a0caeef6 CZ |
8903 | /* Generate add_s r0,b,u6; add_s r1,b,u6 patterns. */ |
8904 | if (TARGET_CODE_DENSITY && REG_P (operands[0]) && REG_P (operands[1]) | |
8905 | && ((REGNO (operands[0]) == 0) || (REGNO (operands[0]) == 1)) | |
fa9c1b3c CZ |
8906 | && satisfies_constraint_Rcq (operands[1]) |
8907 | && satisfies_constraint_L (operands[2])) | |
a0caeef6 | 8908 | ADDSI_OUTPUT1 ("add%? %0,%1,%2 ;6"); |
526b7aee SV |
8909 | } |
8910 | ||
8911 | /* Now try to emit a 32 bit insn without long immediate. */ | |
8912 | ret = 4; | |
8913 | if (!match && match2 && REG_P (operands[1])) | |
8914 | ADDSI_OUTPUT1 ("add%? %0,%2,%1"); | |
8915 | if (match || !cond_p) | |
8916 | { | |
8917 | int limit = (match && !cond_p) ? 0x7ff : 0x3f; | |
8918 | int range_factor = neg_intval & intval; | |
8919 | int shift; | |
8920 | ||
c419f71c | 8921 | if (intval == (HOST_WIDE_INT) (HOST_WIDE_INT_M1U << 31)) |
526b7aee SV |
8922 | ADDSI_OUTPUT1 ("bxor%? %0,%1,31"); |
8923 | ||
8924 | /* If we can use a straight add / sub instead of a {add,sub}[123] of | |
8925 | same size, do, so - the insn latency is lower. */ | |
8926 | /* -0x800 is a 12-bit constant for add /add3 / sub / sub3, but | |
8927 | 0x800 is not. */ | |
8928 | if ((intval >= 0 && intval <= limit) | |
8929 | || (intval == -0x800 && limit == 0x7ff)) | |
8930 | ADDSI_OUTPUT1 ("add%? %0,%1,%2"); | |
8931 | else if ((intval < 0 && neg_intval <= limit) | |
8932 | || (intval == 0x800 && limit == 0x7ff)) | |
8933 | ADDSI_OUTPUT1 ("sub%? %0,%1,%n2"); | |
8934 | shift = range_factor >= 8 ? 3 : (range_factor >> 1); | |
8935 | gcc_assert (shift == 0 || shift == 1 || shift == 2 || shift == 3); | |
8936 | gcc_assert ((((1 << shift) - 1) & intval) == 0); | |
8937 | if (((intval < 0 && intval != -0x4000) | |
8938 | /* sub[123] is slower than add_s / sub, only use it if it | |
8939 | avoids a long immediate. */ | |
8940 | && neg_intval <= limit << shift) | |
8941 | || (intval == 0x4000 && limit == 0x7ff)) | |
8942 | ADDSI_OUTPUT ((format, "sub%d%%? %%0,%%1,%d", | |
8943 | shift, neg_intval >> shift)); | |
8944 | else if ((intval >= 0 && intval <= limit << shift) | |
8945 | || (intval == -0x4000 && limit == 0x7ff)) | |
8946 | ADDSI_OUTPUT ((format, "add%d%%? %%0,%%1,%d", shift, intval >> shift)); | |
8947 | } | |
8948 | /* Try to emit a 16 bit opcode with long immediate. */ | |
8949 | ret = 6; | |
8950 | if (short_p && match) | |
6b55f8c9 | 8951 | ADDSI_OUTPUT1 ("add%? %0,%1,%2"); |
526b7aee SV |
8952 | |
8953 | /* We have to use a 32 bit opcode, and with a long immediate. */ | |
8954 | ret = 8; | |
6b55f8c9 | 8955 | ADDSI_OUTPUT1 (intval < 0 ? "sub%? %0,%1,%n2" : "add%? %0,%1,%2"); |
526b7aee SV |
8956 | } |
8957 | ||
8958 | /* Emit code for an commutative_cond_exec instruction with OPERANDS. | |
8959 | Return the length of the instruction. | |
8960 | If OUTPUT_P is false, don't actually output the instruction, just return | |
8961 | its length. */ | |
8962 | int | |
8963 | arc_output_commutative_cond_exec (rtx *operands, bool output_p) | |
8964 | { | |
8965 | enum rtx_code commutative_op = GET_CODE (operands[3]); | |
8966 | const char *pat = NULL; | |
8967 | ||
8968 | /* Canonical rtl should not have a constant in the first operand position. */ | |
8969 | gcc_assert (!CONSTANT_P (operands[1])); | |
8970 | ||
8971 | switch (commutative_op) | |
8972 | { | |
8973 | case AND: | |
8974 | if (satisfies_constraint_C1p (operands[2])) | |
8975 | pat = "bmsk%? %0,%1,%Z2"; | |
fc1c2d04 CZ |
8976 | else if (satisfies_constraint_C2p (operands[2])) |
8977 | { | |
8978 | operands[2] = GEN_INT ((~INTVAL (operands[2]))); | |
8979 | pat = "bmskn%? %0,%1,%Z2"; | |
8980 | } | |
526b7aee SV |
8981 | else if (satisfies_constraint_Ccp (operands[2])) |
8982 | pat = "bclr%? %0,%1,%M2"; | |
8983 | else if (satisfies_constraint_CnL (operands[2])) | |
8984 | pat = "bic%? %0,%1,%n2-1"; | |
8985 | break; | |
8986 | case IOR: | |
8987 | if (satisfies_constraint_C0p (operands[2])) | |
8988 | pat = "bset%? %0,%1,%z2"; | |
8989 | break; | |
8990 | case XOR: | |
8991 | if (satisfies_constraint_C0p (operands[2])) | |
8992 | pat = "bxor%? %0,%1,%z2"; | |
8993 | break; | |
8994 | case PLUS: | |
8995 | return arc_output_addsi (operands, true, output_p); | |
8996 | default: break; | |
8997 | } | |
8998 | if (output_p) | |
8999 | output_asm_insn (pat ? pat : "%O3.%d5 %0,%1,%2", operands); | |
9000 | if (pat || REG_P (operands[2]) || satisfies_constraint_L (operands[2])) | |
9001 | return 4; | |
9002 | return 8; | |
9003 | } | |
9004 | ||
76715c32 | 9005 | /* Helper function of arc_expand_cpymem. ADDR points to a chunk of memory. |
526b7aee SV |
9006 | Emit code and return an potentially modified address such that offsets |
9007 | up to SIZE are can be added to yield a legitimate address. | |
9008 | if REUSE is set, ADDR is a register that may be modified. */ | |
9009 | ||
9010 | static rtx | |
9011 | force_offsettable (rtx addr, HOST_WIDE_INT size, bool reuse) | |
9012 | { | |
9013 | rtx base = addr; | |
9014 | rtx offs = const0_rtx; | |
9015 | ||
9016 | if (GET_CODE (base) == PLUS) | |
9017 | { | |
9018 | offs = XEXP (base, 1); | |
9019 | base = XEXP (base, 0); | |
9020 | } | |
9021 | if (!REG_P (base) | |
9022 | || (REGNO (base) != STACK_POINTER_REGNUM | |
4173ddaf | 9023 | && REGNO_PTR_FRAME_P (REGNO (base))) |
526b7aee SV |
9024 | || !CONST_INT_P (offs) || !SMALL_INT (INTVAL (offs)) |
9025 | || !SMALL_INT (INTVAL (offs) + size)) | |
9026 | { | |
9027 | if (reuse) | |
9028 | emit_insn (gen_add2_insn (addr, offs)); | |
9029 | else | |
9030 | addr = copy_to_mode_reg (Pmode, addr); | |
9031 | } | |
9032 | return addr; | |
9033 | } | |
9034 | ||
d34a0fdc CZ |
9035 | /* Like move_by_pieces, but take account of load latency, and actual |
9036 | offset ranges. Return true on success. */ | |
526b7aee SV |
9037 | |
9038 | bool | |
76715c32 | 9039 | arc_expand_cpymem (rtx *operands) |
526b7aee SV |
9040 | { |
9041 | rtx dst = operands[0]; | |
9042 | rtx src = operands[1]; | |
9043 | rtx dst_addr, src_addr; | |
9044 | HOST_WIDE_INT size; | |
9045 | int align = INTVAL (operands[3]); | |
9046 | unsigned n_pieces; | |
9047 | int piece = align; | |
9048 | rtx store[2]; | |
9049 | rtx tmpx[2]; | |
9050 | int i; | |
9051 | ||
9052 | if (!CONST_INT_P (operands[2])) | |
9053 | return false; | |
9054 | size = INTVAL (operands[2]); | |
9055 | /* move_by_pieces_ninsns is static, so we can't use it. */ | |
9056 | if (align >= 4) | |
d34a0fdc CZ |
9057 | { |
9058 | if (TARGET_LL64) | |
9059 | n_pieces = (size + 4) / 8U + ((size >> 1) & 1) + (size & 1); | |
9060 | else | |
9061 | n_pieces = (size + 2) / 4U + (size & 1); | |
9062 | } | |
526b7aee SV |
9063 | else if (align == 2) |
9064 | n_pieces = (size + 1) / 2U; | |
9065 | else | |
9066 | n_pieces = size; | |
9067 | if (n_pieces >= (unsigned int) (optimize_size ? 3 : 15)) | |
9068 | return false; | |
d34a0fdc CZ |
9069 | /* Force 32 bit aligned and larger datum to use 64 bit transfers, if |
9070 | possible. */ | |
9071 | if (TARGET_LL64 && (piece >= 4) && (size >= 8)) | |
9072 | piece = 8; | |
9073 | else if (piece > 4) | |
526b7aee SV |
9074 | piece = 4; |
9075 | dst_addr = force_offsettable (XEXP (operands[0], 0), size, 0); | |
9076 | src_addr = force_offsettable (XEXP (operands[1], 0), size, 0); | |
9077 | store[0] = store[1] = NULL_RTX; | |
9078 | tmpx[0] = tmpx[1] = NULL_RTX; | |
9079 | for (i = 0; size > 0; i ^= 1, size -= piece) | |
9080 | { | |
9081 | rtx tmp; | |
ef4bddc2 | 9082 | machine_mode mode; |
526b7aee | 9083 | |
d34a0fdc CZ |
9084 | while (piece > size) |
9085 | piece >>= 1; | |
f67f4dff | 9086 | mode = smallest_int_mode_for_size (piece * BITS_PER_UNIT); |
526b7aee SV |
9087 | /* If we don't re-use temporaries, the scheduler gets carried away, |
9088 | and the register pressure gets unnecessarily high. */ | |
9089 | if (0 && tmpx[i] && GET_MODE (tmpx[i]) == mode) | |
9090 | tmp = tmpx[i]; | |
9091 | else | |
9092 | tmpx[i] = tmp = gen_reg_rtx (mode); | |
9093 | dst_addr = force_offsettable (dst_addr, piece, 1); | |
9094 | src_addr = force_offsettable (src_addr, piece, 1); | |
9095 | if (store[i]) | |
9096 | emit_insn (store[i]); | |
9097 | emit_move_insn (tmp, change_address (src, mode, src_addr)); | |
9098 | store[i] = gen_move_insn (change_address (dst, mode, dst_addr), tmp); | |
9099 | dst_addr = plus_constant (Pmode, dst_addr, piece); | |
9100 | src_addr = plus_constant (Pmode, src_addr, piece); | |
9101 | } | |
9102 | if (store[i]) | |
9103 | emit_insn (store[i]); | |
9104 | if (store[i^1]) | |
9105 | emit_insn (store[i^1]); | |
9106 | return true; | |
9107 | } | |
9108 | ||
b6fb257b CZ |
9109 | static bool |
9110 | arc_get_aux_arg (rtx pat, int *auxr) | |
9111 | { | |
9112 | tree attr, addr = MEM_EXPR (pat); | |
9113 | if (TREE_CODE (addr) != VAR_DECL) | |
9114 | return false; | |
9115 | ||
9116 | attr = DECL_ATTRIBUTES (addr); | |
9117 | if (lookup_attribute ("aux", attr)) | |
9118 | { | |
9119 | tree arg = TREE_VALUE (attr); | |
9120 | if (arg) | |
9121 | { | |
9122 | *auxr = TREE_INT_CST_LOW (TREE_VALUE (arg)); | |
9123 | return true; | |
9124 | } | |
9125 | } | |
9126 | ||
9127 | return false; | |
9128 | } | |
9129 | ||
526b7aee SV |
9130 | /* Prepare operands for move in MODE. Return true iff the move has |
9131 | been emitted. */ | |
9132 | ||
9133 | bool | |
ef4bddc2 | 9134 | prepare_move_operands (rtx *operands, machine_mode mode) |
526b7aee | 9135 | { |
b6fb257b CZ |
9136 | /* First handle aux attribute. */ |
9137 | if (mode == SImode | |
9138 | && (MEM_P (operands[0]) || MEM_P (operands[1]))) | |
9139 | { | |
9140 | rtx tmp; | |
9141 | int auxr = 0; | |
9142 | if (MEM_P (operands[0]) && arc_is_aux_reg_p (operands[0])) | |
9143 | { | |
9144 | /* Save operation. */ | |
9145 | if (arc_get_aux_arg (operands[0], &auxr)) | |
9146 | { | |
9147 | tmp = gen_reg_rtx (SImode); | |
9148 | emit_move_insn (tmp, GEN_INT (auxr)); | |
9149 | } | |
9150 | else | |
9151 | { | |
9152 | tmp = XEXP (operands[0], 0); | |
9153 | } | |
9154 | ||
9155 | operands[1] = force_reg (SImode, operands[1]); | |
9156 | emit_insn (gen_rtx_UNSPEC_VOLATILE | |
9157 | (VOIDmode, gen_rtvec (2, operands[1], tmp), | |
9158 | VUNSPEC_ARC_SR)); | |
9159 | return true; | |
9160 | } | |
9161 | if (MEM_P (operands[1]) && arc_is_aux_reg_p (operands[1])) | |
9162 | { | |
9163 | if (arc_get_aux_arg (operands[1], &auxr)) | |
9164 | { | |
9165 | tmp = gen_reg_rtx (SImode); | |
9166 | emit_move_insn (tmp, GEN_INT (auxr)); | |
9167 | } | |
9168 | else | |
9169 | { | |
9170 | tmp = XEXP (operands[1], 0); | |
9171 | gcc_assert (GET_CODE (tmp) == SYMBOL_REF); | |
9172 | } | |
9173 | /* Load operation. */ | |
9174 | gcc_assert (REG_P (operands[0])); | |
9175 | emit_insn (gen_rtx_SET (operands[0], | |
9176 | gen_rtx_UNSPEC_VOLATILE | |
9177 | (SImode, gen_rtvec (1, tmp), | |
9178 | VUNSPEC_ARC_LR))); | |
9179 | return true; | |
9180 | } | |
9181 | } | |
9182 | ||
673f01b8 | 9183 | if (GET_CODE (operands[1]) == SYMBOL_REF) |
526b7aee | 9184 | { |
673f01b8 | 9185 | enum tls_model model = SYMBOL_REF_TLS_MODEL (operands[1]); |
4be6c9b9 | 9186 | if (MEM_P (operands[0])) |
673f01b8 CZ |
9187 | operands[1] = force_reg (mode, operands[1]); |
9188 | else if (model) | |
9189 | operands[1] = arc_legitimize_tls_address (operands[1], model); | |
28633bbd CZ |
9190 | } |
9191 | ||
673f01b8 CZ |
9192 | operands[1] = arc_legitimize_pic_address (operands[1]); |
9193 | ||
9194 | /* Store instructions are limited, they only accept as address an | |
9195 | immediate, a register or a register plus a small immediate. */ | |
526b7aee | 9196 | if (MEM_P (operands[0]) |
673f01b8 | 9197 | && !move_dest_operand (operands[0], mode)) |
526b7aee | 9198 | { |
673f01b8 CZ |
9199 | rtx tmp0 = copy_to_mode_reg (Pmode, XEXP (operands[0], 0)); |
9200 | rtx tmp1 = change_address (operands[0], mode, tmp0); | |
9201 | MEM_COPY_ATTRIBUTES (tmp1, operands[0]); | |
9202 | operands[0] = tmp1; | |
526b7aee SV |
9203 | } |
9204 | ||
673f01b8 CZ |
9205 | /* Check if it is constant but it is not legitimized. */ |
9206 | if (CONSTANT_P (operands[1]) | |
9207 | && !arc_legitimate_constant_p (mode, operands[1])) | |
9208 | operands[1] = force_reg (mode, XEXP (operands[1], 0)); | |
9209 | else if (MEM_P (operands[0]) | |
9210 | && ((CONSTANT_P (operands[1]) | |
9211 | && !satisfies_constraint_Cm3 (operands[1])) | |
9212 | || MEM_P (operands[1]))) | |
9213 | operands[1] = force_reg (mode, operands[1]); | |
9214 | ||
526b7aee SV |
9215 | return false; |
9216 | } | |
9217 | ||
526b7aee SV |
9218 | /* Output a library call to a function called FNAME that has been arranged |
9219 | to be local to any dso. */ | |
9220 | ||
9221 | const char * | |
9222 | arc_output_libcall (const char *fname) | |
9223 | { | |
9224 | unsigned len = strlen (fname); | |
9225 | static char buf[64]; | |
9226 | ||
9227 | gcc_assert (len < sizeof buf - 35); | |
9228 | if (TARGET_LONG_CALLS_SET | |
9229 | || (TARGET_MEDIUM_CALLS && arc_ccfsm_cond_exec_p ())) | |
9230 | { | |
9231 | if (flag_pic) | |
f5e336b1 | 9232 | sprintf (buf, "add r12,pcl,@%s@pcl\n\tjl%%!%%* [r12]", fname); |
526b7aee SV |
9233 | else |
9234 | sprintf (buf, "jl%%! @%s", fname); | |
9235 | } | |
9236 | else | |
9237 | sprintf (buf, "bl%%!%%* @%s", fname); | |
9238 | return buf; | |
9239 | } | |
9240 | ||
9241 | /* Return the SImode highpart of the DImode value IN. */ | |
9242 | ||
9243 | rtx | |
9244 | disi_highpart (rtx in) | |
9245 | { | |
9246 | return simplify_gen_subreg (SImode, in, DImode, TARGET_BIG_ENDIAN ? 0 : 4); | |
9247 | } | |
9248 | ||
526b7aee SV |
9249 | /* Return length adjustment for INSN. |
9250 | For ARC600: | |
9251 | A write to a core reg greater or equal to 32 must not be immediately | |
9252 | followed by a use. Anticipate the length requirement to insert a nop | |
9253 | between PRED and SUCC to prevent a hazard. */ | |
9254 | ||
9255 | static int | |
647d790d | 9256 | arc600_corereg_hazard (rtx_insn *pred, rtx_insn *succ) |
526b7aee SV |
9257 | { |
9258 | if (!TARGET_ARC600) | |
9259 | return 0; | |
526b7aee | 9260 | if (GET_CODE (PATTERN (pred)) == SEQUENCE) |
647d790d | 9261 | pred = as_a <rtx_sequence *> (PATTERN (pred))->insn (1); |
526b7aee | 9262 | if (GET_CODE (PATTERN (succ)) == SEQUENCE) |
647d790d | 9263 | succ = as_a <rtx_sequence *> (PATTERN (succ))->insn (0); |
526b7aee SV |
9264 | if (recog_memoized (pred) == CODE_FOR_mulsi_600 |
9265 | || recog_memoized (pred) == CODE_FOR_umul_600 | |
9266 | || recog_memoized (pred) == CODE_FOR_mac_600 | |
9267 | || recog_memoized (pred) == CODE_FOR_mul64_600 | |
9268 | || recog_memoized (pred) == CODE_FOR_mac64_600 | |
9269 | || recog_memoized (pred) == CODE_FOR_umul64_600 | |
9270 | || recog_memoized (pred) == CODE_FOR_umac64_600) | |
9271 | return 0; | |
36cc6254 RS |
9272 | subrtx_iterator::array_type array; |
9273 | FOR_EACH_SUBRTX (iter, array, PATTERN (pred), NONCONST) | |
9274 | { | |
9275 | const_rtx x = *iter; | |
9276 | switch (GET_CODE (x)) | |
9277 | { | |
9278 | case SET: case POST_INC: case POST_DEC: case PRE_INC: case PRE_DEC: | |
9279 | break; | |
9280 | default: | |
9281 | /* This is also fine for PRE/POST_MODIFY, because they | |
9282 | contain a SET. */ | |
9283 | continue; | |
9284 | } | |
9285 | rtx dest = XEXP (x, 0); | |
9286 | /* Check if this sets a an extension register. N.B. we use 61 for the | |
9287 | condition codes, which is definitely not an extension register. */ | |
9288 | if (REG_P (dest) && REGNO (dest) >= 32 && REGNO (dest) < 61 | |
9289 | /* Check if the same register is used by the PAT. */ | |
9290 | && (refers_to_regno_p | |
9291 | (REGNO (dest), | |
9292 | REGNO (dest) + (GET_MODE_SIZE (GET_MODE (dest)) + 3) / 4U, | |
9293 | PATTERN (succ), 0))) | |
9294 | return 4; | |
9295 | } | |
9296 | return 0; | |
526b7aee SV |
9297 | } |
9298 | ||
f50bb868 CZ |
9299 | /* Given a rtx, check if it is an assembly instruction or not. */ |
9300 | ||
9301 | static int | |
9302 | arc_asm_insn_p (rtx x) | |
9303 | { | |
9304 | int i, j; | |
9305 | ||
9306 | if (x == 0) | |
9307 | return 0; | |
9308 | ||
9309 | switch (GET_CODE (x)) | |
9310 | { | |
9311 | case ASM_OPERANDS: | |
9312 | case ASM_INPUT: | |
9313 | return 1; | |
9314 | ||
9315 | case SET: | |
9316 | return arc_asm_insn_p (SET_SRC (x)); | |
9317 | ||
9318 | case PARALLEL: | |
9319 | j = 0; | |
9320 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
9321 | j += arc_asm_insn_p (XVECEXP (x, 0, i)); | |
9322 | if ( j > 0) | |
9323 | return 1; | |
9324 | break; | |
9325 | ||
9326 | default: | |
9327 | break; | |
9328 | } | |
9329 | ||
9330 | return 0; | |
9331 | } | |
9332 | ||
526b7aee SV |
9333 | /* For ARC600: |
9334 | A write to a core reg greater or equal to 32 must not be immediately | |
9335 | followed by a use. Anticipate the length requirement to insert a nop | |
9336 | between PRED and SUCC to prevent a hazard. */ | |
9337 | ||
9338 | int | |
647d790d | 9339 | arc_hazard (rtx_insn *pred, rtx_insn *succ) |
526b7aee | 9340 | { |
526b7aee SV |
9341 | if (!pred || !INSN_P (pred) || !succ || !INSN_P (succ)) |
9342 | return 0; | |
f50bb868 | 9343 | |
f50bb868 CZ |
9344 | if (TARGET_ARC600) |
9345 | return arc600_corereg_hazard (pred, succ); | |
9346 | ||
9347 | return 0; | |
526b7aee SV |
9348 | } |
9349 | ||
9350 | /* Return length adjustment for INSN. */ | |
9351 | ||
9352 | int | |
647d790d | 9353 | arc_adjust_insn_length (rtx_insn *insn, int len, bool) |
526b7aee SV |
9354 | { |
9355 | if (!INSN_P (insn)) | |
9356 | return len; | |
9357 | /* We already handle sequences by ignoring the delay sequence flag. */ | |
9358 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) | |
9359 | return len; | |
9360 | ||
526b7aee SV |
9361 | /* Check for return with but one preceding insn since function |
9362 | start / call. */ | |
9363 | if (TARGET_PAD_RETURN | |
9364 | && JUMP_P (insn) | |
9365 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
9366 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
9367 | && get_attr_type (insn) == TYPE_RETURN) | |
9368 | { | |
84034c69 | 9369 | rtx_insn *prev = prev_active_insn (insn); |
526b7aee SV |
9370 | |
9371 | if (!prev || !(prev = prev_active_insn (prev)) | |
9372 | || ((NONJUMP_INSN_P (prev) | |
9373 | && GET_CODE (PATTERN (prev)) == SEQUENCE) | |
84034c69 DM |
9374 | ? CALL_ATTR (as_a <rtx_sequence *> (PATTERN (prev))->insn (0), |
9375 | NON_SIBCALL) | |
526b7aee SV |
9376 | : CALL_ATTR (prev, NON_SIBCALL))) |
9377 | return len + 4; | |
9378 | } | |
9379 | if (TARGET_ARC600) | |
9380 | { | |
b3458f61 | 9381 | rtx_insn *succ = next_real_insn (insn); |
526b7aee SV |
9382 | |
9383 | /* One the ARC600, a write to an extension register must be separated | |
9384 | from a read. */ | |
9385 | if (succ && INSN_P (succ)) | |
9386 | len += arc600_corereg_hazard (insn, succ); | |
9387 | } | |
9388 | ||
9389 | /* Restore extracted operands - otherwise splitters like the addsi3_mixed one | |
9390 | can go awry. */ | |
9391 | extract_constrain_insn_cached (insn); | |
9392 | ||
9393 | return len; | |
9394 | } | |
9395 | ||
526b7aee SV |
9396 | /* Return a copy of COND from *STATEP, inverted if that is indicated by the |
9397 | CC field of *STATEP. */ | |
9398 | ||
9399 | static rtx | |
9400 | arc_get_ccfsm_cond (struct arc_ccfsm *statep, bool reverse) | |
9401 | { | |
9402 | rtx cond = statep->cond; | |
9403 | int raw_cc = get_arc_condition_code (cond); | |
9404 | if (reverse) | |
9405 | raw_cc = ARC_INVERSE_CONDITION_CODE (raw_cc); | |
9406 | ||
9407 | if (statep->cc == raw_cc) | |
9408 | return copy_rtx (cond); | |
9409 | ||
9410 | gcc_assert (ARC_INVERSE_CONDITION_CODE (raw_cc) == statep->cc); | |
9411 | ||
ef4bddc2 | 9412 | machine_mode ccm = GET_MODE (XEXP (cond, 0)); |
526b7aee SV |
9413 | enum rtx_code code = reverse_condition (GET_CODE (cond)); |
9414 | if (code == UNKNOWN || ccm == CC_FP_GTmode || ccm == CC_FP_GEmode) | |
9415 | code = reverse_condition_maybe_unordered (GET_CODE (cond)); | |
9416 | ||
9417 | return gen_rtx_fmt_ee (code, GET_MODE (cond), | |
9418 | copy_rtx (XEXP (cond, 0)), copy_rtx (XEXP (cond, 1))); | |
9419 | } | |
9420 | ||
bae56bbb JR |
9421 | /* Return version of PAT conditionalized with COND, which is part of INSN. |
9422 | ANNULLED indicates if INSN is an annulled delay-slot insn. | |
9423 | Register further changes if necessary. */ | |
9424 | static rtx | |
9425 | conditionalize_nonjump (rtx pat, rtx cond, rtx insn, bool annulled) | |
9426 | { | |
9427 | /* For commutative operators, we generally prefer to have | |
9428 | the first source match the destination. */ | |
9429 | if (GET_CODE (pat) == SET) | |
9430 | { | |
9431 | rtx src = SET_SRC (pat); | |
9432 | ||
9433 | if (COMMUTATIVE_P (src)) | |
9434 | { | |
9435 | rtx src0 = XEXP (src, 0); | |
9436 | rtx src1 = XEXP (src, 1); | |
9437 | rtx dst = SET_DEST (pat); | |
9438 | ||
9439 | if (rtx_equal_p (src1, dst) && !rtx_equal_p (src0, dst) | |
9440 | /* Leave add_n alone - the canonical form is to | |
9441 | have the complex summand first. */ | |
9442 | && REG_P (src0)) | |
f7df4a84 | 9443 | pat = gen_rtx_SET (dst, |
bae56bbb JR |
9444 | gen_rtx_fmt_ee (GET_CODE (src), GET_MODE (src), |
9445 | src1, src0)); | |
9446 | } | |
9447 | } | |
9448 | ||
9449 | /* dwarf2out.c:dwarf2out_frame_debug_expr doesn't know | |
9450 | what to do with COND_EXEC. */ | |
9451 | if (RTX_FRAME_RELATED_P (insn)) | |
9452 | { | |
9453 | /* If this is the delay slot insn of an anulled branch, | |
9454 | dwarf2out.c:scan_trace understands the anulling semantics | |
9455 | without the COND_EXEC. */ | |
9456 | gcc_assert (annulled); | |
9457 | rtx note = alloc_reg_note (REG_FRAME_RELATED_EXPR, pat, | |
9458 | REG_NOTES (insn)); | |
9459 | validate_change (insn, ®_NOTES (insn), note, 1); | |
9460 | } | |
9461 | pat = gen_rtx_COND_EXEC (VOIDmode, cond, pat); | |
9462 | return pat; | |
9463 | } | |
9464 | ||
526b7aee SV |
9465 | /* Use the ccfsm machinery to do if conversion. */ |
9466 | ||
9467 | static unsigned | |
9468 | arc_ifcvt (void) | |
9469 | { | |
9470 | struct arc_ccfsm *statep = &cfun->machine->ccfsm_current; | |
526b7aee SV |
9471 | |
9472 | memset (statep, 0, sizeof *statep); | |
b3458f61 | 9473 | for (rtx_insn *insn = get_insns (); insn; insn = next_insn (insn)) |
526b7aee SV |
9474 | { |
9475 | arc_ccfsm_advance (insn, statep); | |
9476 | ||
9477 | switch (statep->state) | |
9478 | { | |
9479 | case 0: | |
526b7aee SV |
9480 | break; |
9481 | case 1: case 2: | |
9482 | { | |
9483 | /* Deleted branch. */ | |
526b7aee | 9484 | arc_ccfsm_post_advance (insn, statep); |
53ea364f | 9485 | gcc_assert (!IN_RANGE (statep->state, 1, 2)); |
b3458f61 | 9486 | rtx_insn *seq = NEXT_INSN (PREV_INSN (insn)); |
782bdf21 | 9487 | if (GET_CODE (PATTERN (seq)) == SEQUENCE) |
526b7aee SV |
9488 | { |
9489 | rtx slot = XVECEXP (PATTERN (seq), 0, 1); | |
9490 | rtx pat = PATTERN (slot); | |
9491 | if (INSN_ANNULLED_BRANCH_P (insn)) | |
9492 | { | |
9493 | rtx cond | |
9494 | = arc_get_ccfsm_cond (statep, INSN_FROM_TARGET_P (slot)); | |
9495 | pat = gen_rtx_COND_EXEC (VOIDmode, cond, pat); | |
9496 | } | |
9497 | if (!validate_change (seq, &PATTERN (seq), pat, 0)) | |
9498 | gcc_unreachable (); | |
9499 | PUT_CODE (slot, NOTE); | |
9500 | NOTE_KIND (slot) = NOTE_INSN_DELETED; | |
526b7aee SV |
9501 | } |
9502 | else | |
9503 | { | |
782bdf21 | 9504 | set_insn_deleted (insn); |
526b7aee SV |
9505 | } |
9506 | continue; | |
9507 | } | |
9508 | case 3: | |
9509 | if (LABEL_P (insn) | |
9510 | && statep->target_label == CODE_LABEL_NUMBER (insn)) | |
9511 | { | |
9512 | arc_ccfsm_post_advance (insn, statep); | |
782bdf21 CZ |
9513 | if (--LABEL_NUSES (insn) == 0) |
9514 | delete_insn (insn); | |
526b7aee SV |
9515 | continue; |
9516 | } | |
9517 | /* Fall through. */ | |
9518 | case 4: case 5: | |
9519 | if (!NONDEBUG_INSN_P (insn)) | |
9520 | break; | |
9521 | ||
9522 | /* Conditionalized insn. */ | |
9523 | ||
b3458f61 DM |
9524 | rtx_insn *prev, *pprev; |
9525 | rtx *patp, pat, cond; | |
bae56bbb | 9526 | bool annulled; annulled = false; |
526b7aee SV |
9527 | |
9528 | /* If this is a delay slot insn in a non-annulled branch, | |
9529 | don't conditionalize it. N.B., this should be fine for | |
9530 | conditional return too. However, don't do this for | |
9531 | unconditional branches, as these would be encountered when | |
9532 | processing an 'else' part. */ | |
9533 | prev = PREV_INSN (insn); | |
9534 | pprev = PREV_INSN (prev); | |
9535 | if (pprev && NEXT_INSN (NEXT_INSN (pprev)) == NEXT_INSN (insn) | |
bae56bbb JR |
9536 | && JUMP_P (prev) && get_attr_cond (prev) == COND_USE) |
9537 | { | |
9538 | if (!INSN_ANNULLED_BRANCH_P (prev)) | |
9539 | break; | |
9540 | annulled = true; | |
9541 | } | |
526b7aee SV |
9542 | |
9543 | patp = &PATTERN (insn); | |
9544 | pat = *patp; | |
9545 | cond = arc_get_ccfsm_cond (statep, INSN_FROM_TARGET_P (insn)); | |
9546 | if (NONJUMP_INSN_P (insn) || CALL_P (insn)) | |
9547 | { | |
9548 | /* ??? don't conditionalize if all side effects are dead | |
9549 | in the not-execute case. */ | |
9bf218f9 | 9550 | |
bae56bbb | 9551 | pat = conditionalize_nonjump (pat, cond, insn, annulled); |
526b7aee SV |
9552 | } |
9553 | else if (simplejump_p (insn)) | |
9554 | { | |
9555 | patp = &SET_SRC (pat); | |
9556 | pat = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, *patp, pc_rtx); | |
9557 | } | |
9558 | else if (JUMP_P (insn) && ANY_RETURN_P (PATTERN (insn))) | |
9559 | { | |
9560 | pat = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, pat, pc_rtx); | |
f7df4a84 | 9561 | pat = gen_rtx_SET (pc_rtx, pat); |
526b7aee SV |
9562 | } |
9563 | else | |
9564 | gcc_unreachable (); | |
9565 | validate_change (insn, patp, pat, 1); | |
9566 | if (!apply_change_group ()) | |
9567 | gcc_unreachable (); | |
9568 | if (JUMP_P (insn)) | |
9569 | { | |
b3458f61 | 9570 | rtx_insn *next = next_nonnote_insn (insn); |
526b7aee SV |
9571 | if (GET_CODE (next) == BARRIER) |
9572 | delete_insn (next); | |
9573 | if (statep->state == 3) | |
9574 | continue; | |
9575 | } | |
9576 | break; | |
9577 | default: | |
9578 | gcc_unreachable (); | |
9579 | } | |
9580 | arc_ccfsm_post_advance (insn, statep); | |
9581 | } | |
9582 | return 0; | |
9583 | } | |
9584 | ||
0bc69b81 JR |
9585 | /* Find annulled delay insns and convert them to use the appropriate predicate. |
9586 | This allows branch shortening to size up these insns properly. */ | |
9587 | ||
9588 | static unsigned | |
9589 | arc_predicate_delay_insns (void) | |
9590 | { | |
b3458f61 | 9591 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
0bc69b81 JR |
9592 | { |
9593 | rtx pat, jump, dlay, src, cond, *patp; | |
9594 | int reverse; | |
9595 | ||
9596 | if (!NONJUMP_INSN_P (insn) | |
9597 | || GET_CODE (pat = PATTERN (insn)) != SEQUENCE) | |
9598 | continue; | |
9599 | jump = XVECEXP (pat, 0, 0); | |
9600 | dlay = XVECEXP (pat, 0, 1); | |
9601 | if (!JUMP_P (jump) || !INSN_ANNULLED_BRANCH_P (jump)) | |
9602 | continue; | |
9603 | /* If the branch insn does the annulling, leave the delay insn alone. */ | |
9604 | if (!TARGET_AT_DBR_CONDEXEC && !INSN_FROM_TARGET_P (dlay)) | |
9605 | continue; | |
9606 | /* ??? Could also leave DLAY un-conditionalized if its target is dead | |
9607 | on the other path. */ | |
9608 | gcc_assert (GET_CODE (PATTERN (jump)) == SET); | |
9609 | gcc_assert (SET_DEST (PATTERN (jump)) == pc_rtx); | |
9610 | src = SET_SRC (PATTERN (jump)); | |
9611 | gcc_assert (GET_CODE (src) == IF_THEN_ELSE); | |
9612 | cond = XEXP (src, 0); | |
9613 | if (XEXP (src, 2) == pc_rtx) | |
9614 | reverse = 0; | |
9615 | else if (XEXP (src, 1) == pc_rtx) | |
9616 | reverse = 1; | |
9617 | else | |
9618 | gcc_unreachable (); | |
9af539fe | 9619 | if (reverse != !INSN_FROM_TARGET_P (dlay)) |
0bc69b81 | 9620 | { |
ef4bddc2 | 9621 | machine_mode ccm = GET_MODE (XEXP (cond, 0)); |
0bc69b81 JR |
9622 | enum rtx_code code = reverse_condition (GET_CODE (cond)); |
9623 | if (code == UNKNOWN || ccm == CC_FP_GTmode || ccm == CC_FP_GEmode) | |
9624 | code = reverse_condition_maybe_unordered (GET_CODE (cond)); | |
9625 | ||
9626 | cond = gen_rtx_fmt_ee (code, GET_MODE (cond), | |
9627 | copy_rtx (XEXP (cond, 0)), | |
9628 | copy_rtx (XEXP (cond, 1))); | |
9629 | } | |
9630 | else | |
9631 | cond = copy_rtx (cond); | |
9632 | patp = &PATTERN (dlay); | |
9633 | pat = *patp; | |
eeac7d15 | 9634 | pat = conditionalize_nonjump (pat, cond, dlay, true); |
0bc69b81 JR |
9635 | validate_change (dlay, patp, pat, 1); |
9636 | if (!apply_change_group ()) | |
9637 | gcc_unreachable (); | |
9638 | } | |
9639 | return 0; | |
9640 | } | |
9641 | ||
526b7aee SV |
9642 | /* For ARC600: If a write to a core reg >=32 appears in a delay slot |
9643 | (other than of a forward brcc), it creates a hazard when there is a read | |
9644 | of the same register at the branch target. We can't know what is at the | |
9645 | branch target of calls, and for branches, we don't really know before the | |
9646 | end of delay slot scheduling, either. Not only can individual instruction | |
9647 | be hoisted out into a delay slot, a basic block can also be emptied this | |
9648 | way, and branch and/or fall through targets be redirected. Hence we don't | |
9649 | want such writes in a delay slot. */ | |
526b7aee SV |
9650 | |
9651 | /* Return nonzreo iff INSN writes to an extension core register. */ | |
9652 | ||
9653 | int | |
9654 | arc_write_ext_corereg (rtx insn) | |
9655 | { | |
24dbe738 RS |
9656 | subrtx_iterator::array_type array; |
9657 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
9658 | { | |
9659 | const_rtx x = *iter; | |
9660 | switch (GET_CODE (x)) | |
9661 | { | |
9662 | case SET: case POST_INC: case POST_DEC: case PRE_INC: case PRE_DEC: | |
9663 | break; | |
9664 | default: | |
9665 | /* This is also fine for PRE/POST_MODIFY, because they | |
9666 | contain a SET. */ | |
9667 | continue; | |
9668 | } | |
9669 | const_rtx dest = XEXP (x, 0); | |
9670 | if (REG_P (dest) && REGNO (dest) >= 32 && REGNO (dest) < 61) | |
9671 | return 1; | |
9672 | } | |
9673 | return 0; | |
526b7aee SV |
9674 | } |
9675 | ||
9676 | /* This is like the hook, but returns NULL when it can't / won't generate | |
9677 | a legitimate address. */ | |
9678 | ||
9679 | static rtx | |
9680 | arc_legitimize_address_0 (rtx x, rtx oldx ATTRIBUTE_UNUSED, | |
ef4bddc2 | 9681 | machine_mode mode) |
526b7aee SV |
9682 | { |
9683 | rtx addr, inner; | |
9684 | ||
526b7aee SV |
9685 | addr = x; |
9686 | if (GET_CODE (addr) == CONST) | |
9687 | addr = XEXP (addr, 0); | |
673f01b8 | 9688 | |
526b7aee SV |
9689 | if (GET_CODE (addr) == PLUS |
9690 | && CONST_INT_P (XEXP (addr, 1)) | |
9691 | && ((GET_CODE (XEXP (addr, 0)) == SYMBOL_REF | |
9692 | && !SYMBOL_REF_FUNCTION_P (XEXP (addr, 0))) | |
9693 | || (REG_P (XEXP (addr, 0)) | |
9694 | && (INTVAL (XEXP (addr, 1)) & 252)))) | |
9695 | { | |
9696 | HOST_WIDE_INT offs, upper; | |
9697 | int size = GET_MODE_SIZE (mode); | |
9698 | ||
9699 | offs = INTVAL (XEXP (addr, 1)); | |
9700 | upper = (offs + 256 * size) & ~511 * size; | |
9701 | inner = plus_constant (Pmode, XEXP (addr, 0), upper); | |
9702 | #if 0 /* ??? this produces worse code for EEMBC idctrn01 */ | |
9703 | if (GET_CODE (x) == CONST) | |
9704 | inner = gen_rtx_CONST (Pmode, inner); | |
9705 | #endif | |
9706 | addr = plus_constant (Pmode, force_reg (Pmode, inner), offs - upper); | |
9707 | x = addr; | |
9708 | } | |
9709 | else if (GET_CODE (addr) == SYMBOL_REF && !SYMBOL_REF_FUNCTION_P (addr)) | |
9710 | x = force_reg (Pmode, x); | |
ef4bddc2 | 9711 | if (memory_address_p ((machine_mode) mode, x)) |
526b7aee SV |
9712 | return x; |
9713 | return NULL_RTX; | |
9714 | } | |
9715 | ||
9716 | static rtx | |
ef4bddc2 | 9717 | arc_legitimize_address (rtx orig_x, rtx oldx, machine_mode mode) |
526b7aee SV |
9718 | { |
9719 | rtx new_x = arc_legitimize_address_0 (orig_x, oldx, mode); | |
9720 | ||
9721 | if (new_x) | |
9722 | return new_x; | |
9723 | return orig_x; | |
9724 | } | |
9725 | ||
9726 | static rtx | |
20565692 CZ |
9727 | arc_delegitimize_address_0 (rtx op) |
9728 | { | |
9729 | switch (GET_CODE (op)) | |
9730 | { | |
9731 | case CONST: | |
9732 | return arc_delegitimize_address_0 (XEXP (op, 0)); | |
9733 | ||
9734 | case UNSPEC: | |
9735 | switch (XINT (op, 1)) | |
9736 | { | |
9737 | case ARC_UNSPEC_GOT: | |
9738 | case ARC_UNSPEC_GOTOFFPC: | |
9739 | return XVECEXP (op, 0, 0); | |
9740 | default: | |
9741 | break; | |
9742 | } | |
9743 | break; | |
9744 | ||
9745 | case PLUS: | |
9746 | { | |
9747 | rtx t1 = arc_delegitimize_address_0 (XEXP (op, 0)); | |
9748 | rtx t2 = XEXP (op, 1); | |
9749 | ||
9750 | if (t1 && t2) | |
9751 | return gen_rtx_PLUS (GET_MODE (op), t1, t2); | |
9752 | break; | |
9753 | } | |
9754 | ||
9755 | default: | |
9756 | break; | |
9757 | } | |
526b7aee SV |
9758 | return NULL_RTX; |
9759 | } | |
9760 | ||
9761 | static rtx | |
20565692 | 9762 | arc_delegitimize_address (rtx orig_x) |
526b7aee | 9763 | { |
20565692 CZ |
9764 | rtx x = orig_x; |
9765 | ||
9766 | if (MEM_P (x)) | |
526b7aee | 9767 | x = XEXP (x, 0); |
20565692 | 9768 | |
526b7aee | 9769 | x = arc_delegitimize_address_0 (x); |
20565692 CZ |
9770 | if (!x) |
9771 | return orig_x; | |
9772 | ||
9773 | if (MEM_P (orig_x)) | |
9774 | x = replace_equiv_address_nv (orig_x, x); | |
9775 | return x; | |
526b7aee SV |
9776 | } |
9777 | ||
9778 | /* Return a REG rtx for acc1. N.B. the gcc-internal representation may | |
9779 | differ from the hardware register number in order to allow the generic | |
9780 | code to correctly split the concatenation of acc1 and acc2. */ | |
9781 | ||
9782 | rtx | |
9783 | gen_acc1 (void) | |
9784 | { | |
9785 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 56: 57); | |
9786 | } | |
9787 | ||
9788 | /* Return a REG rtx for acc2. N.B. the gcc-internal representation may | |
9789 | differ from the hardware register number in order to allow the generic | |
9790 | code to correctly split the concatenation of acc1 and acc2. */ | |
9791 | ||
9792 | rtx | |
9793 | gen_acc2 (void) | |
9794 | { | |
9795 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 57: 56); | |
9796 | } | |
9797 | ||
9798 | /* Return a REG rtx for mlo. N.B. the gcc-internal representation may | |
9799 | differ from the hardware register number in order to allow the generic | |
9800 | code to correctly split the concatenation of mhi and mlo. */ | |
9801 | ||
9802 | rtx | |
9803 | gen_mlo (void) | |
9804 | { | |
9805 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 59: 58); | |
9806 | } | |
9807 | ||
9808 | /* Return a REG rtx for mhi. N.B. the gcc-internal representation may | |
9809 | differ from the hardware register number in order to allow the generic | |
9810 | code to correctly split the concatenation of mhi and mlo. */ | |
9811 | ||
9812 | rtx | |
9813 | gen_mhi (void) | |
9814 | { | |
9815 | return gen_rtx_REG (SImode, TARGET_BIG_ENDIAN ? 58: 59); | |
9816 | } | |
9817 | ||
9818 | /* FIXME: a parameter should be added, and code added to final.c, | |
9819 | to reproduce this functionality in shorten_branches. */ | |
9820 | #if 0 | |
9821 | /* Return nonzero iff BRANCH should be unaligned if possible by upsizing | |
9822 | a previous instruction. */ | |
9823 | int | |
9824 | arc_unalign_branch_p (rtx branch) | |
9825 | { | |
9826 | rtx note; | |
9827 | ||
9828 | if (!TARGET_UNALIGN_BRANCH) | |
9829 | return 0; | |
9830 | /* Do not do this if we have a filled delay slot. */ | |
9831 | if (get_attr_delay_slot_filled (branch) == DELAY_SLOT_FILLED_YES | |
4654c0cf | 9832 | && !NEXT_INSN (branch)->deleted ()) |
526b7aee SV |
9833 | return 0; |
9834 | note = find_reg_note (branch, REG_BR_PROB, 0); | |
9835 | return (!note | |
9836 | || (arc_unalign_prob_threshold && !br_prob_note_reliable_p (note)) | |
9837 | || INTVAL (XEXP (note, 0)) < arc_unalign_prob_threshold); | |
9838 | } | |
9839 | #endif | |
9840 | ||
9841 | /* When estimating sizes during arc_reorg, when optimizing for speed, there | |
9842 | are three reasons why we need to consider branches to be length 6: | |
9843 | - annull-false delay slot insns are implemented using conditional execution, | |
9844 | thus preventing short insn formation where used. | |
9845 | - for ARC600: annul-true delay slot insns are implemented where possible | |
9846 | using conditional execution, preventing short insn formation where used. | |
9847 | - for ARC700: likely or somewhat likely taken branches are made long and | |
9848 | unaligned if possible to avoid branch penalty. */ | |
9849 | ||
9850 | bool | |
9851 | arc_branch_size_unknown_p (void) | |
9852 | { | |
9853 | return !optimize_size && arc_reorg_in_progress; | |
9854 | } | |
9855 | ||
526b7aee SV |
9856 | /* The usual; we set up our machine_function data. */ |
9857 | ||
9858 | static struct machine_function * | |
9859 | arc_init_machine_status (void) | |
9860 | { | |
9861 | struct machine_function *machine; | |
766090c2 | 9862 | machine = ggc_cleared_alloc<machine_function> (); |
526b7aee | 9863 | machine->fn_type = ARC_FUNCTION_UNKNOWN; |
526b7aee SV |
9864 | |
9865 | return machine; | |
9866 | } | |
9867 | ||
9868 | /* Implements INIT_EXPANDERS. We just set up to call the above | |
9869 | function. */ | |
9870 | ||
9871 | void | |
9872 | arc_init_expanders (void) | |
9873 | { | |
9874 | init_machine_status = arc_init_machine_status; | |
9875 | } | |
9876 | ||
9877 | /* Check if OP is a proper parallel of a millicode call pattern. OFFSET | |
9878 | indicates a number of elements to ignore - that allows to have a | |
9879 | sibcall pattern that starts with (return). LOAD_P is zero for store | |
9880 | multiple (for prologues), and one for load multiples (for epilogues), | |
9881 | and two for load multiples where no final clobber of blink is required. | |
9882 | We also skip the first load / store element since this is supposed to | |
9883 | be checked in the instruction pattern. */ | |
9884 | ||
9885 | int | |
9886 | arc_check_millicode (rtx op, int offset, int load_p) | |
9887 | { | |
9888 | int len = XVECLEN (op, 0) - offset; | |
9889 | int i; | |
9890 | ||
9891 | if (load_p == 2) | |
9892 | { | |
9893 | if (len < 2 || len > 13) | |
9894 | return 0; | |
9895 | load_p = 1; | |
9896 | } | |
9897 | else | |
9898 | { | |
9899 | rtx elt = XVECEXP (op, 0, --len); | |
9900 | ||
9901 | if (GET_CODE (elt) != CLOBBER | |
9902 | || !REG_P (XEXP (elt, 0)) | |
9903 | || REGNO (XEXP (elt, 0)) != RETURN_ADDR_REGNUM | |
9904 | || len < 3 || len > 13) | |
9905 | return 0; | |
9906 | } | |
9907 | for (i = 1; i < len; i++) | |
9908 | { | |
9909 | rtx elt = XVECEXP (op, 0, i + offset); | |
9910 | rtx reg, mem, addr; | |
9911 | ||
9912 | if (GET_CODE (elt) != SET) | |
9913 | return 0; | |
9914 | mem = XEXP (elt, load_p); | |
9915 | reg = XEXP (elt, 1-load_p); | |
9916 | if (!REG_P (reg) || REGNO (reg) != 13U+i || !MEM_P (mem)) | |
9917 | return 0; | |
9918 | addr = XEXP (mem, 0); | |
9919 | if (GET_CODE (addr) != PLUS | |
9920 | || !rtx_equal_p (stack_pointer_rtx, XEXP (addr, 0)) | |
9921 | || !CONST_INT_P (XEXP (addr, 1)) || INTVAL (XEXP (addr, 1)) != i*4) | |
9922 | return 0; | |
9923 | } | |
9924 | return 1; | |
9925 | } | |
9926 | ||
9927 | /* Accessor functions for cfun->machine->unalign. */ | |
9928 | ||
526b7aee SV |
9929 | void |
9930 | arc_clear_unalign (void) | |
9931 | { | |
9932 | if (cfun) | |
9933 | cfun->machine->unalign = 0; | |
9934 | } | |
9935 | ||
9936 | void | |
9937 | arc_toggle_unalign (void) | |
9938 | { | |
9939 | cfun->machine->unalign ^= 2; | |
9940 | } | |
9941 | ||
9942 | /* Operands 0..2 are the operands of a addsi which uses a 12 bit | |
9943 | constant in operand 2, but which would require a LIMM because of | |
9944 | operand mismatch. | |
9945 | operands 3 and 4 are new SET_SRCs for operands 0. */ | |
9946 | ||
9947 | void | |
9948 | split_addsi (rtx *operands) | |
9949 | { | |
9950 | int val = INTVAL (operands[2]); | |
9951 | ||
9952 | /* Try for two short insns first. Lengths being equal, we prefer | |
9953 | expansions with shorter register lifetimes. */ | |
9954 | if (val > 127 && val <= 255 | |
9955 | && satisfies_constraint_Rcq (operands[0])) | |
9956 | { | |
9957 | operands[3] = operands[2]; | |
9958 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[1]); | |
9959 | } | |
9960 | else | |
9961 | { | |
9962 | operands[3] = operands[1]; | |
9963 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[2]); | |
9964 | } | |
9965 | } | |
9966 | ||
9967 | /* Operands 0..2 are the operands of a subsi which uses a 12 bit | |
9968 | constant in operand 1, but which would require a LIMM because of | |
9969 | operand mismatch. | |
9970 | operands 3 and 4 are new SET_SRCs for operands 0. */ | |
9971 | ||
9972 | void | |
9973 | split_subsi (rtx *operands) | |
9974 | { | |
9975 | int val = INTVAL (operands[1]); | |
9976 | ||
9977 | /* Try for two short insns first. Lengths being equal, we prefer | |
9978 | expansions with shorter register lifetimes. */ | |
9979 | if (satisfies_constraint_Rcq (operands[0]) | |
9980 | && satisfies_constraint_Rcq (operands[2])) | |
9981 | { | |
9982 | if (val >= -31 && val <= 127) | |
9983 | { | |
9984 | operands[3] = gen_rtx_NEG (SImode, operands[2]); | |
9985 | operands[4] = gen_rtx_PLUS (SImode, operands[0], operands[1]); | |
9986 | return; | |
9987 | } | |
9988 | else if (val >= 0 && val < 255) | |
9989 | { | |
9990 | operands[3] = operands[1]; | |
9991 | operands[4] = gen_rtx_MINUS (SImode, operands[0], operands[2]); | |
9992 | return; | |
9993 | } | |
9994 | } | |
9995 | /* If the destination is not an ARCompact16 register, we might | |
9996 | still have a chance to make a short insn if the source is; | |
9997 | we need to start with a reg-reg move for this. */ | |
9998 | operands[3] = operands[2]; | |
9999 | operands[4] = gen_rtx_MINUS (SImode, operands[1], operands[0]); | |
10000 | } | |
10001 | ||
10002 | /* Handle DOUBLE_REGS uses. | |
10003 | Operand 0: destination register | |
10004 | Operand 1: source register */ | |
10005 | ||
d34a0fdc | 10006 | static bool |
526b7aee SV |
10007 | arc_process_double_reg_moves (rtx *operands) |
10008 | { | |
526b7aee SV |
10009 | enum usesDxState { none, srcDx, destDx, maxDx }; |
10010 | enum usesDxState state = none; | |
73dac59b CZ |
10011 | rtx dest = operands[0]; |
10012 | rtx src = operands[1]; | |
526b7aee SV |
10013 | |
10014 | if (refers_to_regno_p (40, 44, src, 0)) | |
73dac59b CZ |
10015 | { |
10016 | state = srcDx; | |
10017 | gcc_assert (REG_P (dest)); | |
10018 | } | |
526b7aee SV |
10019 | if (refers_to_regno_p (40, 44, dest, 0)) |
10020 | { | |
10021 | /* Via arc_register_move_cost, we should never see D,D moves. */ | |
73dac59b | 10022 | gcc_assert (REG_P (src)); |
526b7aee SV |
10023 | gcc_assert (state == none); |
10024 | state = destDx; | |
10025 | } | |
10026 | ||
10027 | if (state == none) | |
d34a0fdc | 10028 | return false; |
526b7aee SV |
10029 | |
10030 | if (state == srcDx) | |
10031 | { | |
10032 | /* Without the LR insn, we need to split this into a | |
10033 | sequence of insns which will use the DEXCLx and DADDHxy | |
10034 | insns to be able to read the Dx register in question. */ | |
10035 | if (TARGET_DPFP_DISABLE_LRSR) | |
10036 | { | |
10037 | /* gen *movdf_insn_nolrsr */ | |
f7df4a84 | 10038 | rtx set = gen_rtx_SET (dest, src); |
526b7aee SV |
10039 | rtx use1 = gen_rtx_USE (VOIDmode, const1_rtx); |
10040 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set, use1))); | |
10041 | } | |
10042 | else | |
10043 | { | |
10044 | /* When we have 'mov D, r' or 'mov D, D' then get the target | |
10045 | register pair for use with LR insn. */ | |
7d81a567 CZ |
10046 | rtx destHigh = simplify_gen_subreg (SImode, dest, DFmode, |
10047 | TARGET_BIG_ENDIAN ? 0 : 4); | |
10048 | rtx destLow = simplify_gen_subreg (SImode, dest, DFmode, | |
10049 | TARGET_BIG_ENDIAN ? 4 : 0); | |
526b7aee SV |
10050 | |
10051 | /* Produce the two LR insns to get the high and low parts. */ | |
f7df4a84 | 10052 | emit_insn (gen_rtx_SET (destHigh, |
c69899f0 CZ |
10053 | gen_rtx_UNSPEC_VOLATILE (Pmode, |
10054 | gen_rtvec (1, src), | |
10055 | VUNSPEC_ARC_LR_HIGH))); | |
f7df4a84 | 10056 | emit_insn (gen_rtx_SET (destLow, |
c69899f0 CZ |
10057 | gen_rtx_UNSPEC_VOLATILE (Pmode, |
10058 | gen_rtvec (1, src), | |
10059 | VUNSPEC_ARC_LR))); | |
526b7aee SV |
10060 | } |
10061 | } | |
10062 | else if (state == destDx) | |
10063 | { | |
10064 | /* When we have 'mov r, D' or 'mov D, D' and we have access to the | |
10065 | LR insn get the target register pair. */ | |
7d81a567 CZ |
10066 | rtx srcHigh = simplify_gen_subreg (SImode, src, DFmode, |
10067 | TARGET_BIG_ENDIAN ? 0 : 4); | |
10068 | rtx srcLow = simplify_gen_subreg (SImode, src, DFmode, | |
10069 | TARGET_BIG_ENDIAN ? 4 : 0); | |
526b7aee | 10070 | |
491483b0 | 10071 | emit_insn (gen_dexcl_2op (dest, srcHigh, srcLow)); |
526b7aee SV |
10072 | } |
10073 | else | |
10074 | gcc_unreachable (); | |
10075 | ||
d34a0fdc | 10076 | return true; |
526b7aee SV |
10077 | } |
10078 | ||
10079 | /* operands 0..1 are the operands of a 64 bit move instruction. | |
10080 | split it into two moves with operands 2/3 and 4/5. */ | |
10081 | ||
d34a0fdc | 10082 | void |
526b7aee SV |
10083 | arc_split_move (rtx *operands) |
10084 | { | |
ef4bddc2 | 10085 | machine_mode mode = GET_MODE (operands[0]); |
526b7aee SV |
10086 | int i; |
10087 | int swap = 0; | |
10088 | rtx xop[4]; | |
526b7aee SV |
10089 | |
10090 | if (TARGET_DPFP) | |
10091 | { | |
d34a0fdc CZ |
10092 | if (arc_process_double_reg_moves (operands)) |
10093 | return; | |
526b7aee SV |
10094 | } |
10095 | ||
d34a0fdc CZ |
10096 | if (TARGET_LL64 |
10097 | && ((memory_operand (operands[0], mode) | |
2295aa75 CZ |
10098 | && (even_register_operand (operands[1], mode) |
10099 | || satisfies_constraint_Cm3 (operands[1]))) | |
d34a0fdc CZ |
10100 | || (memory_operand (operands[1], mode) |
10101 | && even_register_operand (operands[0], mode)))) | |
10102 | { | |
10103 | emit_move_insn (operands[0], operands[1]); | |
10104 | return; | |
10105 | } | |
10106 | ||
00c072ae CZ |
10107 | if (TARGET_PLUS_QMACW |
10108 | && GET_CODE (operands[1]) == CONST_VECTOR) | |
10109 | { | |
10110 | HOST_WIDE_INT intval0, intval1; | |
10111 | if (GET_MODE (operands[1]) == V2SImode) | |
10112 | { | |
10113 | intval0 = INTVAL (XVECEXP (operands[1], 0, 0)); | |
10114 | intval1 = INTVAL (XVECEXP (operands[1], 0, 1)); | |
10115 | } | |
10116 | else | |
10117 | { | |
10118 | intval1 = INTVAL (XVECEXP (operands[1], 0, 3)) << 16; | |
10119 | intval1 |= INTVAL (XVECEXP (operands[1], 0, 2)) & 0xFFFF; | |
10120 | intval0 = INTVAL (XVECEXP (operands[1], 0, 1)) << 16; | |
10121 | intval0 |= INTVAL (XVECEXP (operands[1], 0, 0)) & 0xFFFF; | |
10122 | } | |
10123 | xop[0] = gen_rtx_REG (SImode, REGNO (operands[0])); | |
10124 | xop[3] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); | |
10125 | xop[2] = GEN_INT (trunc_int_for_mode (intval0, SImode)); | |
10126 | xop[1] = GEN_INT (trunc_int_for_mode (intval1, SImode)); | |
10127 | emit_move_insn (xop[0], xop[2]); | |
10128 | emit_move_insn (xop[3], xop[1]); | |
10129 | return; | |
10130 | } | |
10131 | ||
526b7aee SV |
10132 | for (i = 0; i < 2; i++) |
10133 | { | |
10134 | if (MEM_P (operands[i]) && auto_inc_p (XEXP (operands[i], 0))) | |
10135 | { | |
10136 | rtx addr = XEXP (operands[i], 0); | |
10137 | rtx r, o; | |
10138 | enum rtx_code code; | |
10139 | ||
10140 | gcc_assert (!reg_overlap_mentioned_p (operands[0], addr)); | |
10141 | switch (GET_CODE (addr)) | |
10142 | { | |
10143 | case PRE_DEC: o = GEN_INT (-8); goto pre_modify; | |
10144 | case PRE_INC: o = GEN_INT (8); goto pre_modify; | |
10145 | case PRE_MODIFY: o = XEXP (XEXP (addr, 1), 1); | |
10146 | pre_modify: | |
10147 | code = PRE_MODIFY; | |
10148 | break; | |
10149 | case POST_DEC: o = GEN_INT (-8); goto post_modify; | |
10150 | case POST_INC: o = GEN_INT (8); goto post_modify; | |
10151 | case POST_MODIFY: o = XEXP (XEXP (addr, 1), 1); | |
10152 | post_modify: | |
10153 | code = POST_MODIFY; | |
10154 | swap = 2; | |
10155 | break; | |
10156 | default: | |
10157 | gcc_unreachable (); | |
10158 | } | |
10159 | r = XEXP (addr, 0); | |
10160 | xop[0+i] = adjust_automodify_address_nv | |
10161 | (operands[i], SImode, | |
10162 | gen_rtx_fmt_ee (code, Pmode, r, | |
10163 | gen_rtx_PLUS (Pmode, r, o)), | |
10164 | 0); | |
10165 | xop[2+i] = adjust_automodify_address_nv | |
10166 | (operands[i], SImode, plus_constant (Pmode, r, 4), 4); | |
10167 | } | |
10168 | else | |
10169 | { | |
10170 | xop[0+i] = operand_subword (operands[i], 0, 0, mode); | |
10171 | xop[2+i] = operand_subword (operands[i], 1, 0, mode); | |
10172 | } | |
10173 | } | |
10174 | if (reg_overlap_mentioned_p (xop[0], xop[3])) | |
10175 | { | |
10176 | swap = 2; | |
10177 | gcc_assert (!reg_overlap_mentioned_p (xop[2], xop[1])); | |
10178 | } | |
526b7aee | 10179 | |
d34a0fdc CZ |
10180 | emit_move_insn (xop[0 + swap], xop[1 + swap]); |
10181 | emit_move_insn (xop[2 - swap], xop[3 - swap]); | |
526b7aee | 10182 | |
526b7aee SV |
10183 | } |
10184 | ||
10185 | /* Select between the instruction output templates s_tmpl (for short INSNs) | |
10186 | and l_tmpl (for long INSNs). */ | |
10187 | ||
10188 | const char * | |
b3458f61 | 10189 | arc_short_long (rtx_insn *insn, const char *s_tmpl, const char *l_tmpl) |
526b7aee SV |
10190 | { |
10191 | int is_short = arc_verify_short (insn, cfun->machine->unalign, -1); | |
10192 | ||
10193 | extract_constrain_insn_cached (insn); | |
10194 | return is_short ? s_tmpl : l_tmpl; | |
10195 | } | |
10196 | ||
10197 | /* Searches X for any reference to REGNO, returning the rtx of the | |
10198 | reference found if any. Otherwise, returns NULL_RTX. */ | |
10199 | ||
10200 | rtx | |
10201 | arc_regno_use_in (unsigned int regno, rtx x) | |
10202 | { | |
10203 | const char *fmt; | |
10204 | int i, j; | |
10205 | rtx tem; | |
10206 | ||
c9bd6bcd | 10207 | if (REG_P (x) && refers_to_regno_p (regno, x)) |
526b7aee SV |
10208 | return x; |
10209 | ||
10210 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
10211 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
10212 | { | |
10213 | if (fmt[i] == 'e') | |
10214 | { | |
10215 | if ((tem = regno_use_in (regno, XEXP (x, i)))) | |
10216 | return tem; | |
10217 | } | |
10218 | else if (fmt[i] == 'E') | |
10219 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
10220 | if ((tem = regno_use_in (regno , XVECEXP (x, i, j)))) | |
10221 | return tem; | |
10222 | } | |
10223 | ||
10224 | return NULL_RTX; | |
10225 | } | |
10226 | ||
10227 | /* Return the integer value of the "type" attribute for INSN, or -1 if | |
10228 | INSN can't have attributes. */ | |
10229 | ||
b51addd6 | 10230 | static int |
84034c69 | 10231 | arc_attr_type (rtx_insn *insn) |
526b7aee SV |
10232 | { |
10233 | if (NONJUMP_INSN_P (insn) | |
10234 | ? (GET_CODE (PATTERN (insn)) == USE | |
10235 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
10236 | : JUMP_P (insn) | |
10237 | ? (GET_CODE (PATTERN (insn)) == ADDR_VEC | |
10238 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
10239 | : !CALL_P (insn)) | |
10240 | return -1; | |
10241 | return get_attr_type (insn); | |
10242 | } | |
10243 | ||
10244 | /* Return true if insn sets the condition codes. */ | |
10245 | ||
10246 | bool | |
84034c69 | 10247 | arc_sets_cc_p (rtx_insn *insn) |
526b7aee | 10248 | { |
84034c69 DM |
10249 | if (NONJUMP_INSN_P (insn)) |
10250 | if (rtx_sequence *seq = dyn_cast <rtx_sequence *> (PATTERN (insn))) | |
10251 | insn = seq->insn (seq->len () - 1); | |
526b7aee SV |
10252 | return arc_attr_type (insn) == TYPE_COMPARE; |
10253 | } | |
10254 | ||
10255 | /* Return true if INSN is an instruction with a delay slot we may want | |
10256 | to fill. */ | |
10257 | ||
10258 | bool | |
b3458f61 | 10259 | arc_need_delay (rtx_insn *insn) |
526b7aee | 10260 | { |
b3458f61 | 10261 | rtx_insn *next; |
526b7aee SV |
10262 | |
10263 | if (!flag_delayed_branch) | |
10264 | return false; | |
10265 | /* The return at the end of a function needs a delay slot. */ | |
10266 | if (NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE | |
10267 | && (!(next = next_active_insn (insn)) | |
10268 | || ((!NONJUMP_INSN_P (next) || GET_CODE (PATTERN (next)) != SEQUENCE) | |
10269 | && arc_attr_type (next) == TYPE_RETURN)) | |
10270 | && (!TARGET_PAD_RETURN | |
10271 | || (prev_active_insn (insn) | |
10272 | && prev_active_insn (prev_active_insn (insn)) | |
10273 | && prev_active_insn (prev_active_insn (prev_active_insn (insn)))))) | |
10274 | return true; | |
10275 | if (NONJUMP_INSN_P (insn) | |
10276 | ? (GET_CODE (PATTERN (insn)) == USE | |
10277 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
10278 | || GET_CODE (PATTERN (insn)) == SEQUENCE) | |
10279 | : JUMP_P (insn) | |
10280 | ? (GET_CODE (PATTERN (insn)) == ADDR_VEC | |
10281 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
10282 | : !CALL_P (insn)) | |
10283 | return false; | |
10284 | return num_delay_slots (insn) != 0; | |
10285 | } | |
10286 | ||
10287 | /* Return true if the scheduling pass(es) has/have already run, | |
10288 | i.e. where possible, we should try to mitigate high latencies | |
10289 | by different instruction selection. */ | |
10290 | ||
10291 | bool | |
10292 | arc_scheduling_not_expected (void) | |
10293 | { | |
10294 | return cfun->machine->arc_reorg_started; | |
10295 | } | |
10296 | ||
3f445374 CZ |
10297 | /* Code has a minimum p2 alignment of 1, which we must restore after |
10298 | an ADDR_DIFF_VEC. */ | |
10299 | ||
526b7aee | 10300 | int |
82082f65 | 10301 | arc_label_align (rtx_insn *label) |
526b7aee | 10302 | { |
3f445374 | 10303 | if (align_labels.levels[0].log < 1) |
526b7aee | 10304 | { |
b3458f61 | 10305 | rtx_insn *next = next_nonnote_nondebug_insn (label); |
526b7aee SV |
10306 | if (INSN_P (next) && recog_memoized (next) >= 0) |
10307 | return 1; | |
10308 | } | |
3f445374 | 10309 | return align_labels.levels[0].log; |
526b7aee SV |
10310 | } |
10311 | ||
10312 | /* Return true if LABEL is in executable code. */ | |
10313 | ||
10314 | bool | |
b32d5189 | 10315 | arc_text_label (rtx_insn *label) |
526b7aee | 10316 | { |
b3458f61 | 10317 | rtx_insn *next; |
526b7aee SV |
10318 | |
10319 | /* ??? We use deleted labels like they were still there, see | |
10320 | gcc.c-torture/compile/20000326-2.c . */ | |
10321 | gcc_assert (GET_CODE (label) == CODE_LABEL | |
10322 | || (GET_CODE (label) == NOTE | |
10323 | && NOTE_KIND (label) == NOTE_INSN_DELETED_LABEL)); | |
10324 | next = next_nonnote_insn (label); | |
10325 | if (next) | |
10326 | return (!JUMP_TABLE_DATA_P (next) | |
10327 | || GET_CODE (PATTERN (next)) != ADDR_VEC); | |
10328 | else if (!PREV_INSN (label)) | |
10329 | /* ??? sometimes text labels get inserted very late, see | |
10330 | gcc.dg/torture/stackalign/comp-goto-1.c */ | |
10331 | return true; | |
10332 | return false; | |
10333 | } | |
10334 | ||
526b7aee SV |
10335 | /* Without this, gcc.dg/tree-prof/bb-reorg.c fails to assemble |
10336 | when compiling with -O2 -freorder-blocks-and-partition -fprofile-use | |
339ba33b | 10337 | -D_PROFILE_USE; delay branch scheduling then follows a crossing jump |
526b7aee SV |
10338 | to redirect two breqs. */ |
10339 | ||
10340 | static bool | |
c1ce59ab | 10341 | arc_can_follow_jump (const rtx_insn *follower, const rtx_insn *followee) |
526b7aee SV |
10342 | { |
10343 | /* ??? get_attr_type is declared to take an rtx. */ | |
c1ce59ab | 10344 | union { const rtx_insn *c; rtx_insn *r; } u; |
526b7aee SV |
10345 | |
10346 | u.c = follower; | |
339ba33b | 10347 | if (CROSSING_JUMP_P (followee)) |
526b7aee SV |
10348 | switch (get_attr_type (u.r)) |
10349 | { | |
28f4ff35 CZ |
10350 | case TYPE_BRANCH: |
10351 | if (get_attr_length (u.r) != 2) | |
10352 | break; | |
41bc2c0b | 10353 | /* Fall through. */ |
526b7aee SV |
10354 | case TYPE_BRCC: |
10355 | case TYPE_BRCC_NO_DELAY_SLOT: | |
10356 | return false; | |
10357 | default: | |
10358 | return true; | |
10359 | } | |
10360 | return true; | |
10361 | } | |
10362 | ||
c7314bc1 | 10363 | |
1825c61e | 10364 | /* Implement EPILOGUE_USES. |
526b7aee SV |
10365 | Return true if REGNO should be added to the deemed uses of the epilogue. |
10366 | ||
1825c61e CZ |
10367 | We have to make sure all the register restore instructions are |
10368 | known to be live in interrupt functions, plus the blink register if | |
10369 | it is clobbered by the isr. */ | |
526b7aee SV |
10370 | |
10371 | bool | |
10372 | arc_epilogue_uses (int regno) | |
10373 | { | |
1825c61e | 10374 | unsigned int fn_type; |
ce9dbf20 | 10375 | fn_type = arc_compute_function_type (cfun); |
1825c61e | 10376 | |
28633bbd CZ |
10377 | if (regno == arc_tp_regno) |
10378 | return true; | |
1825c61e | 10379 | |
ce9dbf20 CZ |
10380 | if (regno == RETURN_ADDR_REGNUM) |
10381 | return true; | |
10382 | ||
10383 | if (regno == arc_return_address_register (fn_type)) | |
10384 | return true; | |
10385 | ||
10386 | if (epilogue_completed && ARC_INTERRUPT_P (fn_type)) | |
526b7aee | 10387 | { |
ce9dbf20 CZ |
10388 | /* An interrupt function restores more registers. */ |
10389 | if (df_regs_ever_live_p (regno) || call_used_regs[regno]) | |
10390 | return true; | |
526b7aee | 10391 | } |
ce9dbf20 CZ |
10392 | |
10393 | return false; | |
526b7aee SV |
10394 | } |
10395 | ||
28633bbd CZ |
10396 | /* Helper for EH_USES macro. */ |
10397 | ||
10398 | bool | |
10399 | arc_eh_uses (int regno) | |
10400 | { | |
10401 | if (regno == arc_tp_regno) | |
10402 | return true; | |
10403 | return false; | |
10404 | } | |
10405 | ||
73dac59b | 10406 | /* Return true if we use LRA instead of reload pass. */ |
526b7aee | 10407 | |
73dac59b | 10408 | bool |
526b7aee SV |
10409 | arc_lra_p (void) |
10410 | { | |
73dac59b | 10411 | return arc_lra_flag; |
526b7aee SV |
10412 | } |
10413 | ||
10414 | /* ??? Should we define TARGET_REGISTER_PRIORITY? We might perfer to use | |
10415 | Rcq registers, because some insn are shorter with them. OTOH we already | |
10416 | have separate alternatives for this purpose, and other insns don't | |
10417 | mind, so maybe we should rather prefer the other registers? | |
10418 | We need more data, and we can only get that if we allow people to | |
10419 | try all options. */ | |
10420 | static int | |
10421 | arc_register_priority (int r) | |
10422 | { | |
10423 | switch (arc_lra_priority_tag) | |
10424 | { | |
10425 | case ARC_LRA_PRIORITY_NONE: | |
10426 | return 0; | |
10427 | case ARC_LRA_PRIORITY_NONCOMPACT: | |
10428 | return ((((r & 7) ^ 4) - 4) & 15) != r; | |
10429 | case ARC_LRA_PRIORITY_COMPACT: | |
10430 | return ((((r & 7) ^ 4) - 4) & 15) == r; | |
10431 | default: | |
10432 | gcc_unreachable (); | |
10433 | } | |
10434 | } | |
10435 | ||
10436 | static reg_class_t | |
ef4bddc2 | 10437 | arc_spill_class (reg_class_t /* orig_class */, machine_mode) |
526b7aee SV |
10438 | { |
10439 | return GENERAL_REGS; | |
10440 | } | |
10441 | ||
10442 | bool | |
ef4bddc2 | 10443 | arc_legitimize_reload_address (rtx *p, machine_mode mode, int opnum, |
526b7aee SV |
10444 | int itype) |
10445 | { | |
10446 | rtx x = *p; | |
10447 | enum reload_type type = (enum reload_type) itype; | |
10448 | ||
10449 | if (GET_CODE (x) == PLUS | |
10450 | && CONST_INT_P (XEXP (x, 1)) | |
10451 | && (RTX_OK_FOR_BASE_P (XEXP (x, 0), true) | |
10452 | || (REG_P (XEXP (x, 0)) | |
10453 | && reg_equiv_constant (REGNO (XEXP (x, 0)))))) | |
10454 | { | |
10455 | int scale = GET_MODE_SIZE (mode); | |
10456 | int shift; | |
10457 | rtx index_rtx = XEXP (x, 1); | |
10458 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
10459 | rtx reg, sum, sum2; | |
10460 | ||
10461 | if (scale > 4) | |
10462 | scale = 4; | |
10463 | if ((scale-1) & offset) | |
10464 | scale = 1; | |
10465 | shift = scale >> 1; | |
c419f71c JL |
10466 | offset_base |
10467 | = ((offset + (256 << shift)) | |
4e671509 | 10468 | & ((HOST_WIDE_INT)((unsigned HOST_WIDE_INT) -512 << shift))); |
526b7aee SV |
10469 | /* Sometimes the normal form does not suit DImode. We |
10470 | could avoid that by using smaller ranges, but that | |
10471 | would give less optimized code when SImode is | |
10472 | prevalent. */ | |
10473 | if (GET_MODE_SIZE (mode) + offset - offset_base <= (256 << shift)) | |
10474 | { | |
10475 | int regno; | |
10476 | ||
10477 | reg = XEXP (x, 0); | |
10478 | regno = REGNO (reg); | |
10479 | sum2 = sum = plus_constant (Pmode, reg, offset_base); | |
10480 | ||
10481 | if (reg_equiv_constant (regno)) | |
10482 | { | |
10483 | sum2 = plus_constant (Pmode, reg_equiv_constant (regno), | |
10484 | offset_base); | |
10485 | if (GET_CODE (sum2) == PLUS) | |
10486 | sum2 = gen_rtx_CONST (Pmode, sum2); | |
10487 | } | |
10488 | *p = gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
10489 | push_reload (sum2, NULL_RTX, &XEXP (*p, 0), NULL, | |
10490 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, | |
10491 | type); | |
10492 | return true; | |
10493 | } | |
10494 | } | |
10495 | /* We must re-recognize what we created before. */ | |
10496 | else if (GET_CODE (x) == PLUS | |
10497 | && GET_CODE (XEXP (x, 0)) == PLUS | |
10498 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
10499 | && REG_P (XEXP (XEXP (x, 0), 0)) | |
10500 | && CONST_INT_P (XEXP (x, 1))) | |
10501 | { | |
10502 | /* Because this address is so complex, we know it must have | |
10503 | been created by LEGITIMIZE_RELOAD_ADDRESS before; thus, | |
10504 | it is already unshared, and needs no further unsharing. */ | |
10505 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
10506 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
10507 | return true; | |
10508 | } | |
10509 | return false; | |
10510 | } | |
10511 | ||
ad23f5d4 JG |
10512 | /* Implement TARGET_USE_BY_PIECES_INFRASTRUCTURE_P. */ |
10513 | ||
10514 | static bool | |
445d7826 | 10515 | arc_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
ad23f5d4 JG |
10516 | unsigned int align, |
10517 | enum by_pieces_operation op, | |
10518 | bool speed_p) | |
10519 | { | |
76715c32 | 10520 | /* Let the cpymem expander handle small block moves. */ |
ad23f5d4 JG |
10521 | if (op == MOVE_BY_PIECES) |
10522 | return false; | |
10523 | ||
10524 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
10525 | } | |
10526 | ||
b8a64b7f CZ |
10527 | /* Emit a (pre) memory barrier around an atomic sequence according to |
10528 | MODEL. */ | |
10529 | ||
10530 | static void | |
10531 | arc_pre_atomic_barrier (enum memmodel model) | |
10532 | { | |
10533 | if (need_atomic_barrier_p (model, true)) | |
10534 | emit_insn (gen_memory_barrier ()); | |
10535 | } | |
10536 | ||
10537 | /* Emit a (post) memory barrier around an atomic sequence according to | |
10538 | MODEL. */ | |
10539 | ||
10540 | static void | |
10541 | arc_post_atomic_barrier (enum memmodel model) | |
10542 | { | |
10543 | if (need_atomic_barrier_p (model, false)) | |
10544 | emit_insn (gen_memory_barrier ()); | |
10545 | } | |
10546 | ||
10547 | /* Expand a compare and swap pattern. */ | |
10548 | ||
10549 | static void | |
10550 | emit_unlikely_jump (rtx insn) | |
10551 | { | |
f370536c | 10552 | rtx_insn *jump = emit_jump_insn (insn); |
5fa396ad | 10553 | add_reg_br_prob_note (jump, profile_probability::very_unlikely ()); |
b8a64b7f CZ |
10554 | } |
10555 | ||
10556 | /* Expand code to perform a 8 or 16-bit compare and swap by doing | |
10557 | 32-bit compare and swap on the word containing the byte or | |
10558 | half-word. The difference between a weak and a strong CAS is that | |
10559 | the weak version may simply fail. The strong version relies on two | |
10560 | loops, one checks if the SCOND op is succsfully or not, the other | |
10561 | checks if the 32 bit accessed location which contains the 8 or 16 | |
10562 | bit datum is not changed by other thread. The first loop is | |
10563 | implemented by the atomic_compare_and_swapsi_1 pattern. The second | |
10564 | loops is implemented by this routine. */ | |
10565 | ||
10566 | static void | |
10567 | arc_expand_compare_and_swap_qh (rtx bool_result, rtx result, rtx mem, | |
10568 | rtx oldval, rtx newval, rtx weak, | |
10569 | rtx mod_s, rtx mod_f) | |
10570 | { | |
10571 | rtx addr1 = force_reg (Pmode, XEXP (mem, 0)); | |
10572 | rtx addr = gen_reg_rtx (Pmode); | |
10573 | rtx off = gen_reg_rtx (SImode); | |
10574 | rtx oldv = gen_reg_rtx (SImode); | |
10575 | rtx newv = gen_reg_rtx (SImode); | |
10576 | rtx oldvalue = gen_reg_rtx (SImode); | |
10577 | rtx newvalue = gen_reg_rtx (SImode); | |
10578 | rtx res = gen_reg_rtx (SImode); | |
10579 | rtx resv = gen_reg_rtx (SImode); | |
10580 | rtx memsi, val, mask, end_label, loop_label, cc, x; | |
10581 | machine_mode mode; | |
10582 | bool is_weak = (weak != const0_rtx); | |
10583 | ||
10584 | /* Truncate the address. */ | |
10585 | emit_insn (gen_rtx_SET (addr, | |
10586 | gen_rtx_AND (Pmode, addr1, GEN_INT (-4)))); | |
10587 | ||
10588 | /* Compute the datum offset. */ | |
10589 | emit_insn (gen_rtx_SET (off, | |
10590 | gen_rtx_AND (SImode, addr1, GEN_INT (3)))); | |
10591 | if (TARGET_BIG_ENDIAN) | |
10592 | emit_insn (gen_rtx_SET (off, | |
10593 | gen_rtx_MINUS (SImode, | |
10594 | (GET_MODE (mem) == QImode) ? | |
10595 | GEN_INT (3) : GEN_INT (2), off))); | |
10596 | ||
10597 | /* Normal read from truncated address. */ | |
10598 | memsi = gen_rtx_MEM (SImode, addr); | |
10599 | set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER); | |
10600 | MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem); | |
10601 | ||
10602 | val = copy_to_reg (memsi); | |
10603 | ||
10604 | /* Convert the offset in bits. */ | |
10605 | emit_insn (gen_rtx_SET (off, | |
10606 | gen_rtx_ASHIFT (SImode, off, GEN_INT (3)))); | |
10607 | ||
10608 | /* Get the proper mask. */ | |
10609 | if (GET_MODE (mem) == QImode) | |
10610 | mask = force_reg (SImode, GEN_INT (0xff)); | |
10611 | else | |
10612 | mask = force_reg (SImode, GEN_INT (0xffff)); | |
10613 | ||
10614 | emit_insn (gen_rtx_SET (mask, | |
10615 | gen_rtx_ASHIFT (SImode, mask, off))); | |
10616 | ||
10617 | /* Prepare the old and new values. */ | |
10618 | emit_insn (gen_rtx_SET (val, | |
10619 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
10620 | val))); | |
10621 | ||
10622 | oldval = gen_lowpart (SImode, oldval); | |
10623 | emit_insn (gen_rtx_SET (oldv, | |
10624 | gen_rtx_ASHIFT (SImode, oldval, off))); | |
10625 | ||
10626 | newval = gen_lowpart_common (SImode, newval); | |
10627 | emit_insn (gen_rtx_SET (newv, | |
10628 | gen_rtx_ASHIFT (SImode, newval, off))); | |
10629 | ||
10630 | emit_insn (gen_rtx_SET (oldv, | |
10631 | gen_rtx_AND (SImode, oldv, mask))); | |
10632 | ||
10633 | emit_insn (gen_rtx_SET (newv, | |
10634 | gen_rtx_AND (SImode, newv, mask))); | |
10635 | ||
10636 | if (!is_weak) | |
10637 | { | |
10638 | end_label = gen_label_rtx (); | |
10639 | loop_label = gen_label_rtx (); | |
10640 | emit_label (loop_label); | |
10641 | } | |
10642 | ||
10643 | /* Make the old and new values. */ | |
10644 | emit_insn (gen_rtx_SET (oldvalue, | |
10645 | gen_rtx_IOR (SImode, oldv, val))); | |
10646 | ||
10647 | emit_insn (gen_rtx_SET (newvalue, | |
10648 | gen_rtx_IOR (SImode, newv, val))); | |
10649 | ||
10650 | /* Try an 32bit atomic compare and swap. It clobbers the CC | |
10651 | register. */ | |
10652 | emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue, | |
10653 | weak, mod_s, mod_f)); | |
10654 | ||
10655 | /* Regardless of the weakness of the operation, a proper boolean | |
10656 | result needs to be provided. */ | |
10657 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10658 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
10659 | emit_insn (gen_rtx_SET (bool_result, x)); | |
10660 | ||
10661 | if (!is_weak) | |
10662 | { | |
10663 | /* Check the results: if the atomic op is successfully the goto | |
10664 | to end label. */ | |
10665 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10666 | x = gen_rtx_EQ (VOIDmode, x, const0_rtx); | |
10667 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10668 | gen_rtx_LABEL_REF (Pmode, end_label), pc_rtx); | |
10669 | emit_jump_insn (gen_rtx_SET (pc_rtx, x)); | |
10670 | ||
10671 | /* Wait for the right moment when the accessed 32-bit location | |
10672 | is stable. */ | |
10673 | emit_insn (gen_rtx_SET (resv, | |
10674 | gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask), | |
10675 | res))); | |
10676 | mode = SELECT_CC_MODE (NE, resv, val); | |
10677 | cc = gen_rtx_REG (mode, CC_REG); | |
10678 | emit_insn (gen_rtx_SET (cc, gen_rtx_COMPARE (mode, resv, val))); | |
10679 | ||
10680 | /* Set the new value of the 32 bit location, proper masked. */ | |
10681 | emit_insn (gen_rtx_SET (val, resv)); | |
10682 | ||
10683 | /* Try again if location is unstable. Fall through if only | |
10684 | scond op failed. */ | |
10685 | x = gen_rtx_NE (VOIDmode, cc, const0_rtx); | |
10686 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10687 | gen_rtx_LABEL_REF (Pmode, loop_label), pc_rtx); | |
10688 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10689 | ||
10690 | emit_label (end_label); | |
10691 | } | |
10692 | ||
10693 | /* End: proper return the result for the given mode. */ | |
10694 | emit_insn (gen_rtx_SET (res, | |
10695 | gen_rtx_AND (SImode, res, mask))); | |
10696 | ||
10697 | emit_insn (gen_rtx_SET (res, | |
10698 | gen_rtx_LSHIFTRT (SImode, res, off))); | |
10699 | ||
10700 | emit_move_insn (result, gen_lowpart (GET_MODE (result), res)); | |
10701 | } | |
10702 | ||
10703 | /* Helper function used by "atomic_compare_and_swap" expand | |
10704 | pattern. */ | |
10705 | ||
10706 | void | |
10707 | arc_expand_compare_and_swap (rtx operands[]) | |
10708 | { | |
10709 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
10710 | machine_mode mode; | |
10711 | ||
10712 | bval = operands[0]; | |
10713 | rval = operands[1]; | |
10714 | mem = operands[2]; | |
10715 | oldval = operands[3]; | |
10716 | newval = operands[4]; | |
10717 | is_weak = operands[5]; | |
10718 | mod_s = operands[6]; | |
10719 | mod_f = operands[7]; | |
10720 | mode = GET_MODE (mem); | |
10721 | ||
10722 | if (reg_overlap_mentioned_p (rval, oldval)) | |
10723 | oldval = copy_to_reg (oldval); | |
10724 | ||
10725 | if (mode == SImode) | |
10726 | { | |
10727 | emit_insn (gen_atomic_compare_and_swapsi_1 (rval, mem, oldval, newval, | |
10728 | is_weak, mod_s, mod_f)); | |
10729 | x = gen_rtx_REG (CC_Zmode, CC_REG); | |
10730 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
10731 | emit_insn (gen_rtx_SET (bval, x)); | |
10732 | } | |
10733 | else | |
10734 | { | |
10735 | arc_expand_compare_and_swap_qh (bval, rval, mem, oldval, newval, | |
10736 | is_weak, mod_s, mod_f); | |
10737 | } | |
10738 | } | |
10739 | ||
10740 | /* Helper function used by the "atomic_compare_and_swapsi_1" | |
10741 | pattern. */ | |
10742 | ||
10743 | void | |
10744 | arc_split_compare_and_swap (rtx operands[]) | |
10745 | { | |
10746 | rtx rval, mem, oldval, newval; | |
10747 | machine_mode mode; | |
10748 | enum memmodel mod_s, mod_f; | |
10749 | bool is_weak; | |
10750 | rtx label1, label2, x, cond; | |
10751 | ||
10752 | rval = operands[0]; | |
10753 | mem = operands[1]; | |
10754 | oldval = operands[2]; | |
10755 | newval = operands[3]; | |
10756 | is_weak = (operands[4] != const0_rtx); | |
10757 | mod_s = (enum memmodel) INTVAL (operands[5]); | |
10758 | mod_f = (enum memmodel) INTVAL (operands[6]); | |
10759 | mode = GET_MODE (mem); | |
10760 | ||
10761 | /* ARC atomic ops work only with 32-bit aligned memories. */ | |
10762 | gcc_assert (mode == SImode); | |
10763 | ||
10764 | arc_pre_atomic_barrier (mod_s); | |
10765 | ||
10766 | label1 = NULL_RTX; | |
10767 | if (!is_weak) | |
10768 | { | |
10769 | label1 = gen_label_rtx (); | |
10770 | emit_label (label1); | |
10771 | } | |
10772 | label2 = gen_label_rtx (); | |
10773 | ||
10774 | /* Load exclusive. */ | |
10775 | emit_insn (gen_arc_load_exclusivesi (rval, mem)); | |
10776 | ||
10777 | /* Check if it is oldval. */ | |
10778 | mode = SELECT_CC_MODE (NE, rval, oldval); | |
10779 | cond = gen_rtx_REG (mode, CC_REG); | |
10780 | emit_insn (gen_rtx_SET (cond, gen_rtx_COMPARE (mode, rval, oldval))); | |
10781 | ||
10782 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10783 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10784 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
10785 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10786 | ||
10787 | /* Exclusively store new item. Store clobbers CC reg. */ | |
10788 | emit_insn (gen_arc_store_exclusivesi (mem, newval)); | |
10789 | ||
10790 | if (!is_weak) | |
10791 | { | |
10792 | /* Check the result of the store. */ | |
10793 | cond = gen_rtx_REG (CC_Zmode, CC_REG); | |
10794 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10795 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10796 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
10797 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10798 | } | |
10799 | ||
10800 | if (mod_f != MEMMODEL_RELAXED) | |
10801 | emit_label (label2); | |
10802 | ||
10803 | arc_post_atomic_barrier (mod_s); | |
10804 | ||
10805 | if (mod_f == MEMMODEL_RELAXED) | |
10806 | emit_label (label2); | |
10807 | } | |
10808 | ||
10809 | /* Expand an atomic fetch-and-operate pattern. CODE is the binary operation | |
10810 | to perform. MEM is the memory on which to operate. VAL is the second | |
10811 | operand of the binary operator. BEFORE and AFTER are optional locations to | |
10812 | return the value of MEM either before of after the operation. MODEL_RTX | |
10813 | is a CONST_INT containing the memory model to use. */ | |
10814 | ||
10815 | void | |
10816 | arc_expand_atomic_op (enum rtx_code code, rtx mem, rtx val, | |
10817 | rtx orig_before, rtx orig_after, rtx model_rtx) | |
10818 | { | |
10819 | enum memmodel model = (enum memmodel) INTVAL (model_rtx); | |
10820 | machine_mode mode = GET_MODE (mem); | |
10821 | rtx label, x, cond; | |
10822 | rtx before = orig_before, after = orig_after; | |
10823 | ||
10824 | /* ARC atomic ops work only with 32-bit aligned memories. */ | |
10825 | gcc_assert (mode == SImode); | |
10826 | ||
10827 | arc_pre_atomic_barrier (model); | |
10828 | ||
10829 | label = gen_label_rtx (); | |
10830 | emit_label (label); | |
10831 | label = gen_rtx_LABEL_REF (VOIDmode, label); | |
10832 | ||
10833 | if (before == NULL_RTX) | |
10834 | before = gen_reg_rtx (mode); | |
10835 | ||
10836 | if (after == NULL_RTX) | |
10837 | after = gen_reg_rtx (mode); | |
10838 | ||
10839 | /* Load exclusive. */ | |
10840 | emit_insn (gen_arc_load_exclusivesi (before, mem)); | |
10841 | ||
10842 | switch (code) | |
10843 | { | |
10844 | case NOT: | |
10845 | x = gen_rtx_AND (mode, before, val); | |
10846 | emit_insn (gen_rtx_SET (after, x)); | |
10847 | x = gen_rtx_NOT (mode, after); | |
10848 | emit_insn (gen_rtx_SET (after, x)); | |
10849 | break; | |
10850 | ||
10851 | case MINUS: | |
10852 | if (CONST_INT_P (val)) | |
10853 | { | |
10854 | val = GEN_INT (-INTVAL (val)); | |
10855 | code = PLUS; | |
10856 | } | |
10857 | ||
10858 | /* FALLTHRU. */ | |
10859 | default: | |
10860 | x = gen_rtx_fmt_ee (code, mode, before, val); | |
10861 | emit_insn (gen_rtx_SET (after, x)); | |
10862 | break; | |
10863 | } | |
10864 | ||
10865 | /* Exclusively store new item. Store clobbers CC reg. */ | |
10866 | emit_insn (gen_arc_store_exclusivesi (mem, after)); | |
10867 | ||
10868 | /* Check the result of the store. */ | |
10869 | cond = gen_rtx_REG (CC_Zmode, CC_REG); | |
10870 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
10871 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
10872 | label, pc_rtx); | |
10873 | emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
10874 | ||
10875 | arc_post_atomic_barrier (model); | |
10876 | } | |
10877 | ||
bf9e9dc5 CZ |
10878 | /* Implement TARGET_NO_SPECULATION_IN_DELAY_SLOTS_P. */ |
10879 | ||
10880 | static bool | |
10881 | arc_no_speculation_in_delay_slots_p () | |
10882 | { | |
10883 | return true; | |
10884 | } | |
10885 | ||
d34a0fdc CZ |
10886 | /* Return a parallel of registers to represent where to find the |
10887 | register pieces if required, otherwise NULL_RTX. */ | |
10888 | ||
10889 | static rtx | |
10890 | arc_dwarf_register_span (rtx rtl) | |
10891 | { | |
cd1e4d41 | 10892 | machine_mode mode = GET_MODE (rtl); |
d34a0fdc CZ |
10893 | unsigned regno; |
10894 | rtx p; | |
10895 | ||
10896 | if (GET_MODE_SIZE (mode) != 8) | |
10897 | return NULL_RTX; | |
10898 | ||
10899 | p = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2)); | |
10900 | regno = REGNO (rtl); | |
10901 | XVECEXP (p, 0, 0) = gen_rtx_REG (SImode, regno); | |
10902 | XVECEXP (p, 0, 1) = gen_rtx_REG (SImode, regno + 1); | |
10903 | ||
10904 | return p; | |
10905 | } | |
10906 | ||
fc1c2d04 CZ |
10907 | /* Return true if OP is an acceptable memory operand for ARCompact |
10908 | 16-bit load instructions of MODE. | |
10909 | ||
10910 | AV2SHORT: TRUE if address needs to fit into the new ARCv2 short | |
10911 | non scaled instructions. | |
10912 | ||
10913 | SCALED: TRUE if address can be scaled. */ | |
10914 | ||
10915 | bool | |
10916 | compact_memory_operand_p (rtx op, machine_mode mode, | |
10917 | bool av2short, bool scaled) | |
10918 | { | |
10919 | rtx addr, plus0, plus1; | |
10920 | int size, off; | |
10921 | ||
10922 | /* Eliminate non-memory operations. */ | |
10923 | if (GET_CODE (op) != MEM) | |
10924 | return 0; | |
10925 | ||
10926 | /* .di instructions have no 16-bit form. */ | |
10927 | if (MEM_VOLATILE_P (op) && !TARGET_VOLATILE_CACHE_SET) | |
10928 | return false; | |
10929 | ||
3e4a5f54 CZ |
10930 | /* likewise for uncached types. */ |
10931 | if (arc_is_uncached_mem_p (op)) | |
10932 | return false; | |
10933 | ||
fc1c2d04 CZ |
10934 | if (mode == VOIDmode) |
10935 | mode = GET_MODE (op); | |
10936 | ||
10937 | size = GET_MODE_SIZE (mode); | |
10938 | ||
10939 | /* dword operations really put out 2 instructions, so eliminate | |
10940 | them. */ | |
10941 | if (size > UNITS_PER_WORD) | |
10942 | return false; | |
10943 | ||
10944 | /* Decode the address now. */ | |
10945 | addr = XEXP (op, 0); | |
10946 | switch (GET_CODE (addr)) | |
10947 | { | |
10948 | case REG: | |
10949 | return (REGNO (addr) >= FIRST_PSEUDO_REGISTER | |
10950 | || COMPACT_GP_REG_P (REGNO (addr)) | |
10951 | || (SP_REG_P (REGNO (addr)) && (size != 2))); | |
10952 | case PLUS: | |
10953 | plus0 = XEXP (addr, 0); | |
10954 | plus1 = XEXP (addr, 1); | |
10955 | ||
10956 | if ((GET_CODE (plus0) == REG) | |
10957 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
10958 | || COMPACT_GP_REG_P (REGNO (plus0))) | |
10959 | && ((GET_CODE (plus1) == REG) | |
10960 | && ((REGNO (plus1) >= FIRST_PSEUDO_REGISTER) | |
10961 | || COMPACT_GP_REG_P (REGNO (plus1))))) | |
10962 | { | |
10963 | return !av2short; | |
10964 | } | |
10965 | ||
10966 | if ((GET_CODE (plus0) == REG) | |
10967 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
10968 | || (COMPACT_GP_REG_P (REGNO (plus0)) && !av2short) | |
10969 | || (IN_RANGE (REGNO (plus0), 0, 31) && av2short)) | |
10970 | && (GET_CODE (plus1) == CONST_INT)) | |
10971 | { | |
10972 | bool valid = false; | |
10973 | ||
10974 | off = INTVAL (plus1); | |
10975 | ||
10976 | /* Negative offset is not supported in 16-bit load/store insns. */ | |
10977 | if (off < 0) | |
10978 | return 0; | |
10979 | ||
10980 | /* Only u5 immediates allowed in code density instructions. */ | |
10981 | if (av2short) | |
10982 | { | |
10983 | switch (size) | |
10984 | { | |
10985 | case 1: | |
10986 | return false; | |
10987 | case 2: | |
10988 | /* This is an ldh_s.x instruction, check the u6 | |
10989 | immediate. */ | |
10990 | if (COMPACT_GP_REG_P (REGNO (plus0))) | |
10991 | valid = true; | |
10992 | break; | |
10993 | case 4: | |
10994 | /* Only u5 immediates allowed in 32bit access code | |
10995 | density instructions. */ | |
10996 | if (REGNO (plus0) <= 31) | |
10997 | return ((off < 32) && (off % 4 == 0)); | |
10998 | break; | |
10999 | default: | |
11000 | return false; | |
11001 | } | |
11002 | } | |
11003 | else | |
11004 | if (COMPACT_GP_REG_P (REGNO (plus0))) | |
11005 | valid = true; | |
11006 | ||
11007 | if (valid) | |
11008 | { | |
11009 | ||
11010 | switch (size) | |
11011 | { | |
11012 | case 1: | |
11013 | return (off < 32); | |
11014 | case 2: | |
11015 | /* The 6-bit constant get shifted to fit the real | |
11016 | 5-bits field. Check also for the alignment. */ | |
11017 | return ((off < 64) && (off % 2 == 0)); | |
11018 | case 4: | |
11019 | return ((off < 128) && (off % 4 == 0)); | |
11020 | default: | |
11021 | return false; | |
11022 | } | |
11023 | } | |
11024 | } | |
11025 | ||
11026 | if (REG_P (plus0) && CONST_INT_P (plus1) | |
11027 | && ((REGNO (plus0) >= FIRST_PSEUDO_REGISTER) | |
11028 | || SP_REG_P (REGNO (plus0))) | |
11029 | && !av2short) | |
11030 | { | |
11031 | off = INTVAL (plus1); | |
11032 | return ((size != 2) && (off >= 0 && off < 128) && (off % 4 == 0)); | |
11033 | } | |
11034 | ||
11035 | if ((GET_CODE (plus0) == MULT) | |
11036 | && (GET_CODE (XEXP (plus0, 0)) == REG) | |
11037 | && ((REGNO (XEXP (plus0, 0)) >= FIRST_PSEUDO_REGISTER) | |
11038 | || COMPACT_GP_REG_P (REGNO (XEXP (plus0, 0)))) | |
11039 | && (GET_CODE (plus1) == REG) | |
11040 | && ((REGNO (plus1) >= FIRST_PSEUDO_REGISTER) | |
11041 | || COMPACT_GP_REG_P (REGNO (plus1)))) | |
11042 | return scaled; | |
11043 | default: | |
11044 | break ; | |
11045 | /* TODO: 'gp' and 'pcl' are to supported as base address operand | |
11046 | for 16-bit load instructions. */ | |
11047 | } | |
11048 | return false; | |
11049 | } | |
11050 | ||
6b55f8c9 CZ |
11051 | /* Return nonzero if a jli call should be generated for a call from |
11052 | the current function to DECL. */ | |
11053 | ||
11054 | bool | |
11055 | arc_is_jli_call_p (rtx pat) | |
11056 | { | |
11057 | tree attrs; | |
11058 | tree decl = SYMBOL_REF_DECL (pat); | |
11059 | ||
11060 | /* If it is not a well defined public function then return false. */ | |
11061 | if (!decl || !SYMBOL_REF_FUNCTION_P (pat) || !TREE_PUBLIC (decl)) | |
11062 | return false; | |
11063 | ||
11064 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); | |
11065 | if (lookup_attribute ("jli_always", attrs)) | |
11066 | return true; | |
11067 | ||
11068 | if (lookup_attribute ("jli_fixed", attrs)) | |
11069 | return true; | |
11070 | ||
11071 | return TARGET_JLI_ALWAYS; | |
11072 | } | |
11073 | ||
11074 | /* Handle and "jli" attribute; arguments as in struct | |
11075 | attribute_spec.handler. */ | |
11076 | ||
11077 | static tree | |
11078 | arc_handle_jli_attribute (tree *node ATTRIBUTE_UNUSED, | |
11079 | tree name, tree args, int, | |
11080 | bool *no_add_attrs) | |
11081 | { | |
11082 | if (!TARGET_V2) | |
11083 | { | |
11084 | warning (OPT_Wattributes, | |
11085 | "%qE attribute only valid for ARCv2 architecture", | |
11086 | name); | |
11087 | *no_add_attrs = true; | |
11088 | } | |
11089 | ||
11090 | if (args == NULL_TREE) | |
11091 | { | |
11092 | warning (OPT_Wattributes, | |
11093 | "argument of %qE attribute is missing", | |
11094 | name); | |
11095 | *no_add_attrs = true; | |
11096 | } | |
11097 | else | |
11098 | { | |
11099 | if (TREE_CODE (TREE_VALUE (args)) == NON_LVALUE_EXPR) | |
11100 | TREE_VALUE (args) = TREE_OPERAND (TREE_VALUE (args), 0); | |
11101 | tree arg = TREE_VALUE (args); | |
11102 | if (TREE_CODE (arg) != INTEGER_CST) | |
11103 | { | |
11104 | warning (0, "%qE attribute allows only an integer constant argument", | |
11105 | name); | |
11106 | *no_add_attrs = true; | |
11107 | } | |
11108 | /* FIXME! add range check. TREE_INT_CST_LOW (arg) */ | |
11109 | } | |
11110 | return NULL_TREE; | |
11111 | } | |
11112 | ||
7778a1ad CZ |
11113 | /* Handle and "scure" attribute; arguments as in struct |
11114 | attribute_spec.handler. */ | |
11115 | ||
11116 | static tree | |
11117 | arc_handle_secure_attribute (tree *node ATTRIBUTE_UNUSED, | |
11118 | tree name, tree args, int, | |
11119 | bool *no_add_attrs) | |
11120 | { | |
11121 | if (!TARGET_EM) | |
11122 | { | |
11123 | warning (OPT_Wattributes, | |
11124 | "%qE attribute only valid for ARC EM architecture", | |
11125 | name); | |
11126 | *no_add_attrs = true; | |
11127 | } | |
11128 | ||
11129 | if (args == NULL_TREE) | |
11130 | { | |
11131 | warning (OPT_Wattributes, | |
11132 | "argument of %qE attribute is missing", | |
11133 | name); | |
11134 | *no_add_attrs = true; | |
11135 | } | |
11136 | else | |
11137 | { | |
11138 | if (TREE_CODE (TREE_VALUE (args)) == NON_LVALUE_EXPR) | |
11139 | TREE_VALUE (args) = TREE_OPERAND (TREE_VALUE (args), 0); | |
11140 | tree arg = TREE_VALUE (args); | |
11141 | if (TREE_CODE (arg) != INTEGER_CST) | |
11142 | { | |
11143 | warning (0, "%qE attribute allows only an integer constant argument", | |
11144 | name); | |
11145 | *no_add_attrs = true; | |
11146 | } | |
11147 | } | |
11148 | return NULL_TREE; | |
11149 | } | |
11150 | ||
11151 | /* Return nonzero if the symbol is a secure function. */ | |
11152 | ||
11153 | bool | |
11154 | arc_is_secure_call_p (rtx pat) | |
11155 | { | |
11156 | tree attrs; | |
11157 | tree decl = SYMBOL_REF_DECL (pat); | |
11158 | ||
11159 | if (!decl) | |
11160 | return false; | |
11161 | ||
11162 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); | |
11163 | if (lookup_attribute ("secure_call", attrs)) | |
11164 | return true; | |
11165 | ||
11166 | return false; | |
11167 | } | |
11168 | ||
8180c03f CZ |
11169 | /* Handle "uncached" qualifier. */ |
11170 | ||
11171 | static tree | |
11172 | arc_handle_uncached_attribute (tree *node, | |
11173 | tree name, tree args, | |
11174 | int flags ATTRIBUTE_UNUSED, | |
11175 | bool *no_add_attrs) | |
11176 | { | |
11177 | if (DECL_P (*node) && TREE_CODE (*node) != TYPE_DECL) | |
11178 | { | |
11179 | error ("%qE attribute only applies to types", | |
11180 | name); | |
11181 | *no_add_attrs = true; | |
11182 | } | |
11183 | else if (args) | |
11184 | { | |
11185 | warning (OPT_Wattributes, "argument of %qE attribute ignored", name); | |
11186 | } | |
11187 | return NULL_TREE; | |
11188 | } | |
11189 | ||
11190 | /* Return TRUE if PAT is a memory addressing an uncached data. */ | |
11191 | ||
11192 | bool | |
11193 | arc_is_uncached_mem_p (rtx pat) | |
11194 | { | |
3e4a5f54 CZ |
11195 | tree attrs = NULL_TREE; |
11196 | tree addr; | |
8180c03f CZ |
11197 | |
11198 | if (!MEM_P (pat)) | |
11199 | return false; | |
11200 | ||
11201 | /* Get the memory attributes. */ | |
3e4a5f54 CZ |
11202 | addr = MEM_EXPR (pat); |
11203 | if (!addr) | |
8180c03f CZ |
11204 | return false; |
11205 | ||
3e4a5f54 CZ |
11206 | /* Get the attributes. */ |
11207 | if (TREE_CODE (addr) == MEM_REF) | |
11208 | { | |
11209 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (addr)); | |
11210 | if (lookup_attribute ("uncached", attrs)) | |
11211 | return true; | |
8180c03f | 11212 | |
3e4a5f54 CZ |
11213 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (TREE_OPERAND (addr, 0))); |
11214 | if (lookup_attribute ("uncached", attrs)) | |
11215 | return true; | |
11216 | } | |
11217 | ||
11218 | /* For COMPONENT_REF, use the FIELD_DECL from tree operand 1. */ | |
11219 | if (TREE_CODE (addr) == COMPONENT_REF) | |
11220 | { | |
11221 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (TREE_OPERAND (addr, 1))); | |
11222 | if (lookup_attribute ("uncached", attrs)) | |
11223 | return true; | |
11224 | } | |
8180c03f CZ |
11225 | return false; |
11226 | } | |
11227 | ||
b6fb257b CZ |
11228 | /* Handle aux attribute. The auxiliary registers are addressed using |
11229 | special instructions lr and sr. The attribute 'aux' indicates if a | |
11230 | variable refers to the aux-regs and what is the register number | |
11231 | desired. */ | |
11232 | ||
11233 | static tree | |
11234 | arc_handle_aux_attribute (tree *node, | |
11235 | tree name, tree args, int, | |
11236 | bool *no_add_attrs) | |
11237 | { | |
11238 | /* Isn't it better to use address spaces for the aux-regs? */ | |
11239 | if (DECL_P (*node)) | |
11240 | { | |
11241 | if (TREE_CODE (*node) != VAR_DECL) | |
11242 | { | |
11243 | error ("%qE attribute only applies to variables", name); | |
11244 | *no_add_attrs = true; | |
11245 | } | |
11246 | else if (args) | |
11247 | { | |
11248 | if (TREE_CODE (TREE_VALUE (args)) == NON_LVALUE_EXPR) | |
11249 | TREE_VALUE (args) = TREE_OPERAND (TREE_VALUE (args), 0); | |
11250 | tree arg = TREE_VALUE (args); | |
11251 | if (TREE_CODE (arg) != INTEGER_CST) | |
11252 | { | |
d65485c5 | 11253 | warning (OPT_Wattributes, "%qE attribute allows only an integer " |
b6fb257b CZ |
11254 | "constant argument", name); |
11255 | *no_add_attrs = true; | |
11256 | } | |
11257 | /* FIXME! add range check. TREE_INT_CST_LOW (arg) */ | |
11258 | } | |
11259 | ||
11260 | if (TREE_CODE (*node) == VAR_DECL) | |
11261 | { | |
11262 | tree fntype = TREE_TYPE (*node); | |
11263 | if (fntype && TREE_CODE (fntype) == POINTER_TYPE) | |
11264 | { | |
11265 | tree attrs = tree_cons (get_identifier ("aux"), NULL_TREE, | |
11266 | TYPE_ATTRIBUTES (fntype)); | |
11267 | TYPE_ATTRIBUTES (fntype) = attrs; | |
11268 | } | |
11269 | } | |
11270 | } | |
11271 | return NULL_TREE; | |
11272 | } | |
11273 | ||
7cfbf676 CZ |
11274 | /* Implement TARGET_USE_ANCHORS_FOR_SYMBOL_P. We don't want to use |
11275 | anchors for small data: the GP register acts as an anchor in that | |
11276 | case. We also don't want to use them for PC-relative accesses, | |
11277 | where the PC acts as an anchor. Prohibit also TLS symbols to use | |
11278 | anchors. */ | |
11279 | ||
11280 | static bool | |
11281 | arc_use_anchors_for_symbol_p (const_rtx symbol) | |
11282 | { | |
11283 | if (SYMBOL_REF_TLS_MODEL (symbol)) | |
11284 | return false; | |
11285 | ||
11286 | if (flag_pic) | |
11287 | return false; | |
11288 | ||
11289 | if (SYMBOL_REF_SMALL_P (symbol)) | |
11290 | return false; | |
11291 | ||
11292 | return default_use_anchors_for_symbol_p (symbol); | |
11293 | } | |
11294 | ||
31e72f4f CZ |
11295 | /* Return true if SUBST can't safely replace its equivalent during RA. */ |
11296 | static bool | |
11297 | arc_cannot_substitute_mem_equiv_p (rtx) | |
11298 | { | |
11299 | /* If SUBST is mem[base+index], the address may not fit ISA, | |
11300 | thus return true. */ | |
11301 | return true; | |
11302 | } | |
11303 | ||
8fa2c211 CZ |
11304 | /* Checks whether the operands are valid for use in an LDD/STD |
11305 | instruction. Assumes that RT, and RT2 are REG. This is guaranteed | |
11306 | by the patterns. Assumes that the address in the base register RN | |
11307 | is word aligned. Pattern guarantees that both memory accesses use | |
11308 | the same base register, the offsets are constants within the range, | |
11309 | and the gap between the offsets is 4. If reload complete then | |
11310 | check that registers are legal. */ | |
11311 | ||
11312 | static bool | |
11313 | operands_ok_ldd_std (rtx rt, rtx rt2, HOST_WIDE_INT offset) | |
11314 | { | |
11315 | unsigned int t, t2; | |
11316 | ||
11317 | if (!reload_completed) | |
11318 | return true; | |
11319 | ||
11320 | if (!(SMALL_INT_RANGE (offset, (GET_MODE_SIZE (DImode) - 1) & (~0x03), | |
11321 | (offset & (GET_MODE_SIZE (DImode) - 1) & 3 | |
11322 | ? 0 : -(-GET_MODE_SIZE (DImode) | (~0x03)) >> 1)))) | |
11323 | return false; | |
11324 | ||
11325 | t = REGNO (rt); | |
11326 | t2 = REGNO (rt2); | |
11327 | ||
73dac59b | 11328 | if ((t2 == PCL_REG) |
8fa2c211 CZ |
11329 | || (t % 2 != 0) /* First destination register is not even. */ |
11330 | || (t2 != t + 1)) | |
11331 | return false; | |
11332 | ||
11333 | return true; | |
11334 | } | |
11335 | ||
11336 | /* Helper for gen_operands_ldd_std. Returns true iff the memory | |
11337 | operand MEM's address contains an immediate offset from the base | |
11338 | register and has no side effects, in which case it sets BASE and | |
11339 | OFFSET accordingly. */ | |
11340 | ||
11341 | static bool | |
11342 | mem_ok_for_ldd_std (rtx mem, rtx *base, rtx *offset) | |
11343 | { | |
11344 | rtx addr; | |
11345 | ||
11346 | gcc_assert (base != NULL && offset != NULL); | |
11347 | ||
11348 | /* TODO: Handle more general memory operand patterns, such as | |
11349 | PRE_DEC and PRE_INC. */ | |
11350 | ||
11351 | if (side_effects_p (mem)) | |
11352 | return false; | |
11353 | ||
11354 | /* Can't deal with subregs. */ | |
11355 | if (GET_CODE (mem) == SUBREG) | |
11356 | return false; | |
11357 | ||
11358 | gcc_assert (MEM_P (mem)); | |
11359 | ||
11360 | *offset = const0_rtx; | |
11361 | ||
11362 | addr = XEXP (mem, 0); | |
11363 | ||
11364 | /* If addr isn't valid for DImode, then we can't handle it. */ | |
11365 | if (!arc_legitimate_address_p (DImode, addr, | |
11366 | reload_in_progress || reload_completed)) | |
11367 | return false; | |
11368 | ||
11369 | if (REG_P (addr)) | |
11370 | { | |
11371 | *base = addr; | |
11372 | return true; | |
11373 | } | |
11374 | else if (GET_CODE (addr) == PLUS || GET_CODE (addr) == MINUS) | |
11375 | { | |
11376 | *base = XEXP (addr, 0); | |
11377 | *offset = XEXP (addr, 1); | |
11378 | return (REG_P (*base) && CONST_INT_P (*offset)); | |
11379 | } | |
11380 | ||
11381 | return false; | |
11382 | } | |
11383 | ||
11384 | /* Called from peephole2 to replace two word-size accesses with a | |
11385 | single LDD/STD instruction. Returns true iff we can generate a new | |
11386 | instruction sequence. That is, both accesses use the same base | |
11387 | register and the gap between constant offsets is 4. OPERANDS are | |
11388 | the operands found by the peephole matcher; OPERANDS[0,1] are | |
11389 | register operands, and OPERANDS[2,3] are the corresponding memory | |
11390 | operands. LOAD indicates whether the access is load or store. */ | |
11391 | ||
11392 | bool | |
11393 | gen_operands_ldd_std (rtx *operands, bool load, bool commute) | |
11394 | { | |
11395 | int i, gap; | |
11396 | HOST_WIDE_INT offsets[2], offset; | |
11397 | int nops = 2; | |
11398 | rtx cur_base, cur_offset, tmp; | |
11399 | rtx base = NULL_RTX; | |
11400 | ||
11401 | /* Check that the memory references are immediate offsets from the | |
11402 | same base register. Extract the base register, the destination | |
11403 | registers, and the corresponding memory offsets. */ | |
11404 | for (i = 0; i < nops; i++) | |
11405 | { | |
11406 | if (!mem_ok_for_ldd_std (operands[nops+i], &cur_base, &cur_offset)) | |
11407 | return false; | |
11408 | ||
11409 | if (i == 0) | |
11410 | base = cur_base; | |
11411 | else if (REGNO (base) != REGNO (cur_base)) | |
11412 | return false; | |
11413 | ||
11414 | offsets[i] = INTVAL (cur_offset); | |
11415 | if (GET_CODE (operands[i]) == SUBREG) | |
11416 | { | |
11417 | tmp = SUBREG_REG (operands[i]); | |
11418 | gcc_assert (GET_MODE (operands[i]) == GET_MODE (tmp)); | |
11419 | operands[i] = tmp; | |
11420 | } | |
11421 | } | |
11422 | ||
11423 | /* Make sure there is no dependency between the individual loads. */ | |
11424 | if (load && REGNO (operands[0]) == REGNO (base)) | |
11425 | return false; /* RAW. */ | |
11426 | ||
11427 | if (load && REGNO (operands[0]) == REGNO (operands[1])) | |
11428 | return false; /* WAW. */ | |
11429 | ||
11430 | /* Make sure the instructions are ordered with lower memory access first. */ | |
11431 | if (offsets[0] > offsets[1]) | |
11432 | { | |
11433 | gap = offsets[0] - offsets[1]; | |
11434 | offset = offsets[1]; | |
11435 | ||
11436 | /* Swap the instructions such that lower memory is accessed first. */ | |
11437 | std::swap (operands[0], operands[1]); | |
11438 | std::swap (operands[2], operands[3]); | |
11439 | } | |
11440 | else | |
11441 | { | |
11442 | gap = offsets[1] - offsets[0]; | |
11443 | offset = offsets[0]; | |
11444 | } | |
11445 | ||
11446 | /* Make sure accesses are to consecutive memory locations. */ | |
11447 | if (gap != 4) | |
11448 | return false; | |
11449 | ||
11450 | /* Make sure we generate legal instructions. */ | |
11451 | if (operands_ok_ldd_std (operands[0], operands[1], offset)) | |
11452 | return true; | |
11453 | ||
11454 | if (load && commute) | |
11455 | { | |
11456 | /* Try reordering registers. */ | |
11457 | std::swap (operands[0], operands[1]); | |
11458 | if (operands_ok_ldd_std (operands[0], operands[1], offset)) | |
11459 | return true; | |
11460 | } | |
11461 | ||
11462 | return false; | |
11463 | } | |
11464 | ||
864e2eaa CZ |
11465 | /* This order of allocation is used when we compile for size. It |
11466 | allocates first the registers which are most probably to end up in | |
11467 | a short instruction. */ | |
11468 | static const int size_alloc_order[] = | |
11469 | { | |
11470 | 0, 1, 2, 3, 12, 13, 14, 15, | |
11471 | 4, 5, 6, 7, 8, 9, 10, 11 | |
11472 | }; | |
11473 | ||
11474 | /* Adjust register allocation order when compiling for size. */ | |
11475 | void | |
11476 | arc_adjust_reg_alloc_order (void) | |
11477 | { | |
11478 | const int arc_default_alloc_order[] = REG_ALLOC_ORDER; | |
11479 | memcpy (reg_alloc_order, arc_default_alloc_order, sizeof (reg_alloc_order)); | |
11480 | if (optimize_size) | |
11481 | memcpy (reg_alloc_order, size_alloc_order, sizeof (size_alloc_order)); | |
11482 | } | |
11483 | ||
b9bc3b12 CZ |
11484 | /* Implement TARGET_MEMORY_MOVE_COST. */ |
11485 | ||
11486 | static int | |
11487 | arc_memory_move_cost (machine_mode mode, | |
11488 | reg_class_t rclass ATTRIBUTE_UNUSED, | |
11489 | bool in ATTRIBUTE_UNUSED) | |
11490 | { | |
11491 | if ((GET_MODE_SIZE (mode) <= UNITS_PER_WORD) | |
11492 | || ((GET_MODE_SIZE (mode) <= UNITS_PER_WORD * 2) && TARGET_LL64)) | |
11493 | return 6; | |
11494 | ||
11495 | return (2 * GET_MODE_SIZE (mode)); | |
11496 | } | |
11497 | ||
03301dcc CZ |
11498 | /* Split an OR instruction into multiple BSET/OR instructions in a |
11499 | attempt to avoid long immediate constants. The next strategies are | |
11500 | employed when destination is 'q' reg. | |
11501 | ||
11502 | 1. if there are up to three bits set in the mask, a succession of | |
11503 | three bset instruction will be emitted: | |
11504 | OR rA, rB, mask -> | |
11505 | BSET(_S) rA,rB,mask1/BSET_S rA,rA,mask2/BSET_S rA,rA,mask3 | |
11506 | ||
11507 | 2. if the lower 6 bits of the mask is set and there is only one | |
11508 | bit set in the upper remaining bits then we will emit one bset and | |
11509 | one OR instruction: | |
11510 | OR rA, rB, mask -> OR rA,rB,mask1/BSET_S rA,mask2 | |
11511 | ||
11512 | 3. otherwise an OR with limm will be emmitted. */ | |
11513 | ||
11514 | void | |
11515 | arc_split_ior (rtx *operands) | |
11516 | { | |
11517 | unsigned HOST_WIDE_INT mask, maskx; | |
11518 | rtx op1 = operands[1]; | |
11519 | ||
11520 | gcc_assert (CONST_INT_P (operands[2])); | |
11521 | mask = INTVAL (operands[2]) & 0xffffffff; | |
11522 | ||
11523 | if (__builtin_popcount (mask) > 3 || (mask & 0x3f)) | |
11524 | { | |
11525 | maskx = mask & 0x3f; | |
11526 | emit_insn (gen_rtx_SET (operands[0], | |
11527 | gen_rtx_IOR (SImode, op1, GEN_INT (maskx)))); | |
11528 | op1 = operands[0]; | |
11529 | mask &= ~maskx; | |
11530 | } | |
11531 | ||
11532 | switch (__builtin_popcount (mask)) | |
11533 | { | |
11534 | case 3: | |
11535 | maskx = 1 << (__builtin_ffs (mask) - 1); | |
11536 | emit_insn (gen_rtx_SET (operands[0], | |
11537 | gen_rtx_IOR (SImode, op1, GEN_INT (maskx)))); | |
11538 | mask &= ~maskx; | |
11539 | op1 = operands[0]; | |
11540 | /* FALLTHRU */ | |
11541 | case 2: | |
11542 | maskx = 1 << (__builtin_ffs (mask) - 1); | |
11543 | emit_insn (gen_rtx_SET (operands[0], | |
11544 | gen_rtx_IOR (SImode, op1, GEN_INT (maskx)))); | |
11545 | mask &= ~maskx; | |
11546 | op1 = operands[0]; | |
11547 | /* FALLTHRU */ | |
11548 | case 1: | |
11549 | maskx = 1 << (__builtin_ffs (mask) - 1); | |
11550 | emit_insn (gen_rtx_SET (operands[0], | |
11551 | gen_rtx_IOR (SImode, op1, GEN_INT (maskx)))); | |
11552 | break; | |
11553 | default: | |
11554 | break; | |
11555 | } | |
11556 | } | |
11557 | ||
11558 | /* Helper to check C0x constraint. */ | |
11559 | ||
11560 | bool | |
11561 | arc_check_ior_const (HOST_WIDE_INT ival) | |
11562 | { | |
11563 | unsigned int mask = (unsigned int) (ival & 0xffffffff); | |
11564 | if (__builtin_popcount (mask) <= 3) | |
11565 | return true; | |
11566 | if (__builtin_popcount (mask & ~0x3f) <= 1) | |
11567 | return true; | |
11568 | return false; | |
11569 | } | |
11570 | ||
11571 | /* Split a mov with long immediate instruction into smaller, size | |
11572 | friendly instructions. */ | |
11573 | ||
11574 | bool | |
11575 | arc_split_mov_const (rtx *operands) | |
11576 | { | |
11577 | unsigned HOST_WIDE_INT ival; | |
11578 | HOST_WIDE_INT shimm; | |
11579 | machine_mode mode = GET_MODE (operands[0]); | |
11580 | ||
11581 | /* Manage a constant. */ | |
11582 | gcc_assert (CONST_INT_P (operands[1])); | |
11583 | ival = INTVAL (operands[1]) & 0xffffffff; | |
11584 | ||
11585 | if (SIGNED_INT12 (ival)) | |
11586 | return false; | |
11587 | ||
11588 | /* 1. Check if we can just rotate limm by 8 but using ROR8. */ | |
11589 | if (TARGET_BARREL_SHIFTER && TARGET_V2 | |
11590 | && ((ival & ~0x3f000000) == 0)) | |
11591 | { | |
11592 | shimm = (ival >> 24) & 0x3f; | |
11593 | emit_insn (gen_rtx_SET (operands[0], | |
11594 | gen_rtx_ROTATERT (mode, GEN_INT (shimm), | |
11595 | GEN_INT (8)))); | |
11596 | return true; | |
11597 | } | |
11598 | /* 2. Check if we can just shift by 8 to fit into the u6 of LSL8. */ | |
11599 | if (TARGET_BARREL_SHIFTER && TARGET_V2 | |
11600 | && ((ival & ~0x3f00) == 0)) | |
11601 | { | |
11602 | shimm = (ival >> 8) & 0x3f; | |
11603 | emit_insn (gen_rtx_SET (operands[0], | |
11604 | gen_rtx_ASHIFT (mode, GEN_INT (shimm), | |
11605 | GEN_INT (8)))); | |
11606 | return true; | |
11607 | } | |
11608 | ||
11609 | /* 3. Check if we can just shift by 16 to fit into the u6 of LSL16. */ | |
11610 | if (TARGET_BARREL_SHIFTER && TARGET_V2 | |
11611 | && ((ival & ~0x3f0000) == 0)) | |
11612 | { | |
11613 | shimm = (ival >> 16) & 0x3f; | |
11614 | emit_insn (gen_rtx_SET (operands[0], | |
11615 | gen_rtx_ASHIFT (mode, GEN_INT (shimm), | |
11616 | GEN_INT (16)))); | |
11617 | return true; | |
11618 | } | |
11619 | ||
11620 | /* 4. Check if we can do something like mov_s h,u8 / asl_s ra,h,#nb. */ | |
11621 | if (((ival >> (__builtin_ffs (ival) - 1)) & 0xffffff00) == 0 | |
11622 | && TARGET_BARREL_SHIFTER) | |
11623 | { | |
11624 | HOST_WIDE_INT shift = __builtin_ffs (ival); | |
11625 | shimm = (ival >> (shift - 1)) & 0xff; | |
11626 | emit_insn (gen_rtx_SET (operands[0], GEN_INT (shimm))); | |
11627 | emit_insn (gen_rtx_SET (operands[0], | |
11628 | gen_rtx_ASHIFT (mode, operands[0], | |
11629 | GEN_INT (shift - 1)))); | |
11630 | return true; | |
11631 | } | |
11632 | ||
11633 | /* 5. Check if we can just rotate the limm, useful when no barrel | |
11634 | shifter is present. */ | |
11635 | if ((ival & ~0x8000001f) == 0) | |
11636 | { | |
11637 | shimm = (ival * 2 + 1) & 0x3f; | |
11638 | emit_insn (gen_rtx_SET (operands[0], | |
11639 | gen_rtx_ROTATERT (mode, GEN_INT (shimm), | |
11640 | const1_rtx))); | |
11641 | return true; | |
11642 | } | |
11643 | ||
11644 | /* 6. Check if we can do something with bmask. */ | |
11645 | if (IS_POWEROF2_P (ival + 1)) | |
11646 | { | |
11647 | emit_insn (gen_rtx_SET (operands[0], constm1_rtx)); | |
11648 | emit_insn (gen_rtx_SET (operands[0], | |
11649 | gen_rtx_AND (mode, operands[0], | |
11650 | GEN_INT (ival)))); | |
11651 | return true; | |
11652 | } | |
11653 | ||
11654 | return false; | |
11655 | } | |
11656 | ||
11657 | /* Helper to check Cax constraint. */ | |
11658 | ||
11659 | bool | |
11660 | arc_check_mov_const (HOST_WIDE_INT ival) | |
11661 | { | |
11662 | ival = ival & 0xffffffff; | |
11663 | ||
11664 | if ((ival & ~0x8000001f) == 0) | |
11665 | return true; | |
11666 | ||
11667 | if (IS_POWEROF2_P (ival + 1)) | |
11668 | return true; | |
11669 | ||
11670 | /* The next rules requires a barrel shifter. */ | |
11671 | if (!TARGET_BARREL_SHIFTER) | |
11672 | return false; | |
11673 | ||
11674 | if (((ival >> (__builtin_ffs (ival) - 1)) & 0xffffff00) == 0) | |
11675 | return true; | |
11676 | ||
11677 | if ((ival & ~0x3f00) == 0) | |
11678 | return true; | |
11679 | ||
11680 | if ((ival & ~0x3f0000) == 0) | |
11681 | return true; | |
11682 | ||
11683 | if ((ival & ~0x3f000000) == 0) | |
11684 | return true; | |
11685 | ||
11686 | return false; | |
11687 | } | |
11688 | ||
ce9dbf20 CZ |
11689 | /* Return nonzero if this function is known to have a null epilogue. |
11690 | This allows the optimizer to omit jumps to jumps if no stack | |
11691 | was created. */ | |
11692 | ||
11693 | bool | |
11694 | arc_can_use_return_insn (void) | |
11695 | { | |
11696 | return (reload_completed && cfun->machine->frame_info.total_size == 0 | |
11697 | && !ARC_INTERRUPT_P (arc_compute_function_type (cfun))); | |
11698 | } | |
03301dcc | 11699 | |
7cfbf676 CZ |
11700 | #undef TARGET_USE_ANCHORS_FOR_SYMBOL_P |
11701 | #define TARGET_USE_ANCHORS_FOR_SYMBOL_P arc_use_anchors_for_symbol_p | |
11702 | ||
58e17cf8 RS |
11703 | #undef TARGET_CONSTANT_ALIGNMENT |
11704 | #define TARGET_CONSTANT_ALIGNMENT constant_alignment_word_strings | |
11705 | ||
31e72f4f CZ |
11706 | #undef TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P |
11707 | #define TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P arc_cannot_substitute_mem_equiv_p | |
11708 | ||
efcc2e30 CZ |
11709 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE |
11710 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE arc_asm_trampoline_template | |
11711 | ||
8e95721a CZ |
11712 | #undef TARGET_HAVE_SPECULATION_SAFE_VALUE |
11713 | #define TARGET_HAVE_SPECULATION_SAFE_VALUE speculation_safe_value_not_needed | |
11714 | ||
b9bc3b12 CZ |
11715 | #undef TARGET_REGISTER_MOVE_COST |
11716 | #define TARGET_REGISTER_MOVE_COST arc_register_move_cost | |
11717 | ||
11718 | #undef TARGET_MEMORY_MOVE_COST | |
11719 | #define TARGET_MEMORY_MOVE_COST arc_memory_move_cost | |
11720 | ||
526b7aee SV |
11721 | struct gcc_target targetm = TARGET_INITIALIZER; |
11722 | ||
11723 | #include "gt-arc.h" |