Mercurial > hg > CbC > CbC_gcc
annotate gcc/config/mcore/mcore.c @ 63:b7f97abdc517 gcc-4.6-20100522
update gcc from gcc-4.5.0 to gcc-4.6
author | ryoma <e075725@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 24 May 2010 12:47:05 +0900 |
parents | 77e2b8dfacca |
children | f6334be47118 |
rev | line source |
---|---|
0 | 1 /* Output routines for Motorola MCore processor |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2 Copyright (C) 1993, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3 2009 Free Software Foundation, Inc. |
0 | 4 |
5 This file is part of GCC. | |
6 | |
7 GCC is free software; you can redistribute it and/or modify it | |
8 under the terms of the GNU General Public License as published | |
9 by the Free Software Foundation; either version 3, or (at your | |
10 option) any later version. | |
11 | |
12 GCC is distributed in the hope that it will be useful, but WITHOUT | |
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
14 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
15 License for more details. | |
16 | |
17 You should have received a copy of the GNU General Public License | |
18 along with GCC; see the file COPYING3. If not see | |
19 <http://www.gnu.org/licenses/>. */ | |
20 | |
21 #include "config.h" | |
22 #include "system.h" | |
23 #include "coretypes.h" | |
24 #include "tm.h" | |
25 #include "rtl.h" | |
26 #include "tree.h" | |
27 #include "tm_p.h" | |
28 #include "assert.h" | |
29 #include "mcore.h" | |
30 #include "regs.h" | |
31 #include "hard-reg-set.h" | |
32 #include "insn-config.h" | |
33 #include "conditions.h" | |
34 #include "output.h" | |
35 #include "insn-attr.h" | |
36 #include "flags.h" | |
37 #include "obstack.h" | |
38 #include "expr.h" | |
39 #include "reload.h" | |
40 #include "recog.h" | |
41 #include "function.h" | |
42 #include "ggc.h" | |
43 #include "toplev.h" | |
44 #include "target.h" | |
45 #include "target-def.h" | |
46 #include "df.h" | |
47 | |
48 /* Maximum size we are allowed to grow the stack in a single operation. | |
49 If we want more, we must do it in increments of at most this size. | |
50 If this value is 0, we don't check at all. */ | |
51 int mcore_stack_increment = STACK_UNITS_MAXSTEP; | |
52 | |
53 /* For dumping information about frame sizes. */ | |
54 char * mcore_current_function_name = 0; | |
55 long mcore_current_compilation_timestamp = 0; | |
56 | |
57 /* Global variables for machine-dependent things. */ | |
58 | |
59 /* Provides the class number of the smallest class containing | |
60 reg number. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
61 const enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
0 | 62 { |
63 GENERAL_REGS, ONLYR1_REGS, LRW_REGS, LRW_REGS, | |
64 LRW_REGS, LRW_REGS, LRW_REGS, LRW_REGS, | |
65 LRW_REGS, LRW_REGS, LRW_REGS, LRW_REGS, | |
66 LRW_REGS, LRW_REGS, LRW_REGS, GENERAL_REGS, | |
67 GENERAL_REGS, C_REGS, NO_REGS, NO_REGS, | |
68 }; | |
69 | |
70 /* Provide reg_class from a letter such as appears in the machine | |
71 description. */ | |
72 const enum reg_class reg_class_from_letter[] = | |
73 { | |
74 /* a */ LRW_REGS, /* b */ ONLYR1_REGS, /* c */ C_REGS, /* d */ NO_REGS, | |
75 /* e */ NO_REGS, /* f */ NO_REGS, /* g */ NO_REGS, /* h */ NO_REGS, | |
76 /* i */ NO_REGS, /* j */ NO_REGS, /* k */ NO_REGS, /* l */ NO_REGS, | |
77 /* m */ NO_REGS, /* n */ NO_REGS, /* o */ NO_REGS, /* p */ NO_REGS, | |
78 /* q */ NO_REGS, /* r */ GENERAL_REGS, /* s */ NO_REGS, /* t */ NO_REGS, | |
79 /* u */ NO_REGS, /* v */ NO_REGS, /* w */ NO_REGS, /* x */ ALL_REGS, | |
80 /* y */ NO_REGS, /* z */ NO_REGS | |
81 }; | |
82 | |
83 struct mcore_frame | |
84 { | |
85 int arg_size; /* Stdarg spills (bytes). */ | |
86 int reg_size; /* Non-volatile reg saves (bytes). */ | |
87 int reg_mask; /* Non-volatile reg saves. */ | |
88 int local_size; /* Locals. */ | |
89 int outbound_size; /* Arg overflow on calls out. */ | |
90 int pad_outbound; | |
91 int pad_local; | |
92 int pad_reg; | |
93 /* Describe the steps we'll use to grow it. */ | |
94 #define MAX_STACK_GROWS 4 /* Gives us some spare space. */ | |
95 int growth[MAX_STACK_GROWS]; | |
96 int arg_offset; | |
97 int reg_offset; | |
98 int reg_growth; | |
99 int local_growth; | |
100 }; | |
101 | |
102 typedef enum | |
103 { | |
104 COND_NO, | |
105 COND_MOV_INSN, | |
106 COND_CLR_INSN, | |
107 COND_INC_INSN, | |
108 COND_DEC_INSN, | |
109 COND_BRANCH_INSN | |
110 } | |
111 cond_type; | |
112 | |
113 static void output_stack_adjust (int, int); | |
114 static int calc_live_regs (int *); | |
115 static int try_constant_tricks (long, HOST_WIDE_INT *, HOST_WIDE_INT *); | |
116 static const char * output_inline_const (enum machine_mode, rtx *); | |
117 static void layout_mcore_frame (struct mcore_frame *); | |
118 static void mcore_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
119 static cond_type is_cond_candidate (rtx); | |
120 static rtx emit_new_cond_insn (rtx, int); | |
121 static rtx conditionalize_block (rtx); | |
122 static void conditionalize_optimization (void); | |
123 static void mcore_reorg (void); | |
124 static rtx handle_structs_in_regs (enum machine_mode, const_tree, int); | |
125 static void mcore_mark_dllexport (tree); | |
126 static void mcore_mark_dllimport (tree); | |
127 static int mcore_dllexport_p (tree); | |
128 static int mcore_dllimport_p (tree); | |
129 static tree mcore_handle_naked_attribute (tree *, tree, tree, int, bool *); | |
130 #ifdef OBJECT_FORMAT_ELF | |
131 static void mcore_asm_named_section (const char *, | |
132 unsigned int, tree); | |
133 #endif | |
134 static void mcore_unique_section (tree, int); | |
135 static void mcore_encode_section_info (tree, rtx, int); | |
136 static const char *mcore_strip_name_encoding (const char *); | |
137 static int mcore_const_costs (rtx, RTX_CODE); | |
138 static int mcore_and_cost (rtx); | |
139 static int mcore_ior_cost (rtx); | |
140 static bool mcore_rtx_costs (rtx, int, int, int *, bool); | |
141 static void mcore_external_libcall (rtx); | |
142 static bool mcore_return_in_memory (const_tree, const_tree); | |
143 static int mcore_arg_partial_bytes (CUMULATIVE_ARGS *, | |
144 enum machine_mode, | |
145 tree, bool); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
146 static void mcore_asm_trampoline_template (FILE *); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
147 static void mcore_trampoline_init (rtx, tree, rtx); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
148 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
149 /* MCore specific attributes. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
150 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
151 static const struct attribute_spec mcore_attribute_table[] = |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
152 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
153 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
154 { "dllexport", 0, 0, true, false, false, NULL }, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
155 { "dllimport", 0, 0, true, false, false, NULL }, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
156 { "naked", 0, 0, true, false, false, mcore_handle_naked_attribute }, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
157 { NULL, 0, 0, false, false, false, NULL } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
158 }; |
0 | 159 |
160 /* Initialize the GCC target structure. */ | |
161 #undef TARGET_ASM_EXTERNAL_LIBCALL | |
162 #define TARGET_ASM_EXTERNAL_LIBCALL mcore_external_libcall | |
163 | |
164 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES | |
165 #undef TARGET_MERGE_DECL_ATTRIBUTES | |
166 #define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes | |
167 #endif | |
168 | |
169 #ifdef OBJECT_FORMAT_ELF | |
170 #undef TARGET_ASM_UNALIGNED_HI_OP | |
171 #define TARGET_ASM_UNALIGNED_HI_OP "\t.short\t" | |
172 #undef TARGET_ASM_UNALIGNED_SI_OP | |
173 #define TARGET_ASM_UNALIGNED_SI_OP "\t.long\t" | |
174 #endif | |
175 | |
176 #undef TARGET_ATTRIBUTE_TABLE | |
177 #define TARGET_ATTRIBUTE_TABLE mcore_attribute_table | |
178 #undef TARGET_ASM_UNIQUE_SECTION | |
179 #define TARGET_ASM_UNIQUE_SECTION mcore_unique_section | |
180 #undef TARGET_ASM_FUNCTION_RODATA_SECTION | |
181 #define TARGET_ASM_FUNCTION_RODATA_SECTION default_no_function_rodata_section | |
182 #undef TARGET_DEFAULT_TARGET_FLAGS | |
183 #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
184 #undef TARGET_ENCODE_SECTION_INFO | |
185 #define TARGET_ENCODE_SECTION_INFO mcore_encode_section_info | |
186 #undef TARGET_STRIP_NAME_ENCODING | |
187 #define TARGET_STRIP_NAME_ENCODING mcore_strip_name_encoding | |
188 #undef TARGET_RTX_COSTS | |
189 #define TARGET_RTX_COSTS mcore_rtx_costs | |
190 #undef TARGET_ADDRESS_COST | |
191 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0 | |
192 #undef TARGET_MACHINE_DEPENDENT_REORG | |
193 #define TARGET_MACHINE_DEPENDENT_REORG mcore_reorg | |
194 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
195 #undef TARGET_PROMOTE_FUNCTION_MODE |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
196 #define TARGET_PROMOTE_FUNCTION_MODE default_promote_function_mode_always_promote |
0 | 197 #undef TARGET_PROMOTE_PROTOTYPES |
198 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true | |
199 | |
200 #undef TARGET_RETURN_IN_MEMORY | |
201 #define TARGET_RETURN_IN_MEMORY mcore_return_in_memory | |
202 #undef TARGET_MUST_PASS_IN_STACK | |
203 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
204 #undef TARGET_PASS_BY_REFERENCE | |
205 #define TARGET_PASS_BY_REFERENCE hook_pass_by_reference_must_pass_in_stack | |
206 #undef TARGET_ARG_PARTIAL_BYTES | |
207 #define TARGET_ARG_PARTIAL_BYTES mcore_arg_partial_bytes | |
208 | |
209 #undef TARGET_SETUP_INCOMING_VARARGS | |
210 #define TARGET_SETUP_INCOMING_VARARGS mcore_setup_incoming_varargs | |
211 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
212 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
213 #define TARGET_ASM_TRAMPOLINE_TEMPLATE mcore_asm_trampoline_template |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
214 #undef TARGET_TRAMPOLINE_INIT |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
215 #define TARGET_TRAMPOLINE_INIT mcore_trampoline_init |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
216 |
0 | 217 struct gcc_target targetm = TARGET_INITIALIZER; |
218 | |
219 /* Adjust the stack and return the number of bytes taken to do it. */ | |
220 static void | |
221 output_stack_adjust (int direction, int size) | |
222 { | |
223 /* If extending stack a lot, we do it incrementally. */ | |
224 if (direction < 0 && size > mcore_stack_increment && mcore_stack_increment > 0) | |
225 { | |
226 rtx tmp = gen_rtx_REG (SImode, 1); | |
227 rtx memref; | |
228 | |
229 emit_insn (gen_movsi (tmp, GEN_INT (mcore_stack_increment))); | |
230 do | |
231 { | |
232 emit_insn (gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, tmp)); | |
233 memref = gen_rtx_MEM (SImode, stack_pointer_rtx); | |
234 MEM_VOLATILE_P (memref) = 1; | |
235 emit_insn (gen_movsi (memref, stack_pointer_rtx)); | |
236 size -= mcore_stack_increment; | |
237 } | |
238 while (size > mcore_stack_increment); | |
239 | |
240 /* SIZE is now the residual for the last adjustment, | |
241 which doesn't require a probe. */ | |
242 } | |
243 | |
244 if (size) | |
245 { | |
246 rtx insn; | |
247 rtx val = GEN_INT (size); | |
248 | |
249 if (size > 32) | |
250 { | |
251 rtx nval = gen_rtx_REG (SImode, 1); | |
252 emit_insn (gen_movsi (nval, val)); | |
253 val = nval; | |
254 } | |
255 | |
256 if (direction > 0) | |
257 insn = gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, val); | |
258 else | |
259 insn = gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, val); | |
260 | |
261 emit_insn (insn); | |
262 } | |
263 } | |
264 | |
265 /* Work out the registers which need to be saved, | |
266 both as a mask and a count. */ | |
267 | |
268 static int | |
269 calc_live_regs (int * count) | |
270 { | |
271 int reg; | |
272 int live_regs_mask = 0; | |
273 | |
274 * count = 0; | |
275 | |
276 for (reg = 0; reg < FIRST_PSEUDO_REGISTER; reg++) | |
277 { | |
278 if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) | |
279 { | |
280 (*count)++; | |
281 live_regs_mask |= (1 << reg); | |
282 } | |
283 } | |
284 | |
285 return live_regs_mask; | |
286 } | |
287 | |
288 /* Print the operand address in x to the stream. */ | |
289 | |
290 void | |
291 mcore_print_operand_address (FILE * stream, rtx x) | |
292 { | |
293 switch (GET_CODE (x)) | |
294 { | |
295 case REG: | |
296 fprintf (stream, "(%s)", reg_names[REGNO (x)]); | |
297 break; | |
298 | |
299 case PLUS: | |
300 { | |
301 rtx base = XEXP (x, 0); | |
302 rtx index = XEXP (x, 1); | |
303 | |
304 if (GET_CODE (base) != REG) | |
305 { | |
306 /* Ensure that BASE is a register (one of them must be). */ | |
307 rtx temp = base; | |
308 base = index; | |
309 index = temp; | |
310 } | |
311 | |
312 switch (GET_CODE (index)) | |
313 { | |
314 case CONST_INT: | |
315 fprintf (stream, "(%s," HOST_WIDE_INT_PRINT_DEC ")", | |
316 reg_names[REGNO(base)], INTVAL (index)); | |
317 break; | |
318 | |
319 default: | |
320 gcc_unreachable (); | |
321 } | |
322 } | |
323 | |
324 break; | |
325 | |
326 default: | |
327 output_addr_const (stream, x); | |
328 break; | |
329 } | |
330 } | |
331 | |
332 /* Print operand x (an rtx) in assembler syntax to file stream | |
333 according to modifier code. | |
334 | |
335 'R' print the next register or memory location along, i.e. the lsw in | |
336 a double word value | |
337 'O' print a constant without the # | |
338 'M' print a constant as its negative | |
339 'P' print log2 of a power of two | |
340 'Q' print log2 of an inverse of a power of two | |
341 'U' print register for ldm/stm instruction | |
342 'X' print byte number for xtrbN instruction. */ | |
343 | |
344 void | |
345 mcore_print_operand (FILE * stream, rtx x, int code) | |
346 { | |
347 switch (code) | |
348 { | |
349 case 'N': | |
350 if (INTVAL(x) == -1) | |
351 fprintf (asm_out_file, "32"); | |
352 else | |
353 fprintf (asm_out_file, "%d", exact_log2 (INTVAL (x) + 1)); | |
354 break; | |
355 case 'P': | |
356 fprintf (asm_out_file, "%d", exact_log2 (INTVAL (x) & 0xffffffff)); | |
357 break; | |
358 case 'Q': | |
359 fprintf (asm_out_file, "%d", exact_log2 (~INTVAL (x))); | |
360 break; | |
361 case 'O': | |
362 fprintf (asm_out_file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
363 break; | |
364 case 'M': | |
365 fprintf (asm_out_file, HOST_WIDE_INT_PRINT_DEC, - INTVAL (x)); | |
366 break; | |
367 case 'R': | |
368 /* Next location along in memory or register. */ | |
369 switch (GET_CODE (x)) | |
370 { | |
371 case REG: | |
372 fputs (reg_names[REGNO (x) + 1], (stream)); | |
373 break; | |
374 case MEM: | |
375 mcore_print_operand_address | |
376 (stream, XEXP (adjust_address (x, SImode, 4), 0)); | |
377 break; | |
378 default: | |
379 gcc_unreachable (); | |
380 } | |
381 break; | |
382 case 'U': | |
383 fprintf (asm_out_file, "%s-%s", reg_names[REGNO (x)], | |
384 reg_names[REGNO (x) + 3]); | |
385 break; | |
386 case 'x': | |
387 fprintf (asm_out_file, HOST_WIDE_INT_PRINT_HEX, INTVAL (x)); | |
388 break; | |
389 case 'X': | |
390 fprintf (asm_out_file, HOST_WIDE_INT_PRINT_DEC, 3 - INTVAL (x) / 8); | |
391 break; | |
392 | |
393 default: | |
394 switch (GET_CODE (x)) | |
395 { | |
396 case REG: | |
397 fputs (reg_names[REGNO (x)], (stream)); | |
398 break; | |
399 case MEM: | |
400 output_address (XEXP (x, 0)); | |
401 break; | |
402 default: | |
403 output_addr_const (stream, x); | |
404 break; | |
405 } | |
406 break; | |
407 } | |
408 } | |
409 | |
410 /* What does a constant cost ? */ | |
411 | |
412 static int | |
413 mcore_const_costs (rtx exp, enum rtx_code code) | |
414 { | |
415 HOST_WIDE_INT val = INTVAL (exp); | |
416 | |
417 /* Easy constants. */ | |
418 if ( CONST_OK_FOR_I (val) | |
419 || CONST_OK_FOR_M (val) | |
420 || CONST_OK_FOR_N (val) | |
421 || (code == PLUS && CONST_OK_FOR_L (val))) | |
422 return 1; | |
423 else if (code == AND | |
424 && ( CONST_OK_FOR_M (~val) | |
425 || CONST_OK_FOR_N (~val))) | |
426 return 2; | |
427 else if (code == PLUS | |
428 && ( CONST_OK_FOR_I (-val) | |
429 || CONST_OK_FOR_M (-val) | |
430 || CONST_OK_FOR_N (-val))) | |
431 return 2; | |
432 | |
433 return 5; | |
434 } | |
435 | |
436 /* What does an and instruction cost - we do this b/c immediates may | |
437 have been relaxed. We want to ensure that cse will cse relaxed immeds | |
438 out. Otherwise we'll get bad code (multiple reloads of the same const). */ | |
439 | |
440 static int | |
441 mcore_and_cost (rtx x) | |
442 { | |
443 HOST_WIDE_INT val; | |
444 | |
445 if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
446 return 2; | |
447 | |
448 val = INTVAL (XEXP (x, 1)); | |
449 | |
450 /* Do it directly. */ | |
451 if (CONST_OK_FOR_K (val) || CONST_OK_FOR_M (~val)) | |
452 return 2; | |
453 /* Takes one instruction to load. */ | |
454 else if (const_ok_for_mcore (val)) | |
455 return 3; | |
456 /* Takes two instructions to load. */ | |
457 else if (TARGET_HARDLIT && mcore_const_ok_for_inline (val)) | |
458 return 4; | |
459 | |
460 /* Takes a lrw to load. */ | |
461 return 5; | |
462 } | |
463 | |
464 /* What does an or cost - see and_cost(). */ | |
465 | |
466 static int | |
467 mcore_ior_cost (rtx x) | |
468 { | |
469 HOST_WIDE_INT val; | |
470 | |
471 if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
472 return 2; | |
473 | |
474 val = INTVAL (XEXP (x, 1)); | |
475 | |
476 /* Do it directly with bclri. */ | |
477 if (CONST_OK_FOR_M (val)) | |
478 return 2; | |
479 /* Takes one instruction to load. */ | |
480 else if (const_ok_for_mcore (val)) | |
481 return 3; | |
482 /* Takes two instructions to load. */ | |
483 else if (TARGET_HARDLIT && mcore_const_ok_for_inline (val)) | |
484 return 4; | |
485 | |
486 /* Takes a lrw to load. */ | |
487 return 5; | |
488 } | |
489 | |
490 static bool | |
491 mcore_rtx_costs (rtx x, int code, int outer_code, int * total, | |
492 bool speed ATTRIBUTE_UNUSED) | |
493 { | |
494 switch (code) | |
495 { | |
496 case CONST_INT: | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
497 *total = mcore_const_costs (x, (enum rtx_code) outer_code); |
0 | 498 return true; |
499 case CONST: | |
500 case LABEL_REF: | |
501 case SYMBOL_REF: | |
502 *total = 5; | |
503 return true; | |
504 case CONST_DOUBLE: | |
505 *total = 10; | |
506 return true; | |
507 | |
508 case AND: | |
509 *total = COSTS_N_INSNS (mcore_and_cost (x)); | |
510 return true; | |
511 | |
512 case IOR: | |
513 *total = COSTS_N_INSNS (mcore_ior_cost (x)); | |
514 return true; | |
515 | |
516 case DIV: | |
517 case UDIV: | |
518 case MOD: | |
519 case UMOD: | |
520 case FLOAT: | |
521 case FIX: | |
522 *total = COSTS_N_INSNS (100); | |
523 return true; | |
524 | |
525 default: | |
526 return false; | |
527 } | |
528 } | |
529 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
530 /* Prepare the operands for a comparison. Return whether the branch/setcc |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
531 should reverse the operands. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
532 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
533 bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
534 mcore_gen_compare (enum rtx_code code, rtx op0, rtx op1) |
0 | 535 { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
536 rtx cc_reg = gen_rtx_REG (CCmode, CC_REG); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
537 bool invert; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
538 |
0 | 539 if (GET_CODE (op1) == CONST_INT) |
540 { | |
541 HOST_WIDE_INT val = INTVAL (op1); | |
542 | |
543 switch (code) | |
544 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
545 case GTU: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
546 /* Unsigned > 0 is the same as != 0; everything else is converted |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
547 below to LEU (reversed cmphs). */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
548 if (val == 0) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
549 code = NE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
550 break; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
551 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
552 /* Check whether (LE A imm) can become (LT A imm + 1), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
553 or (GT A imm) can become (GE A imm + 1). */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
554 case GT: |
0 | 555 case LE: |
556 if (CONST_OK_FOR_J (val + 1)) | |
557 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
558 op1 = GEN_INT (val + 1); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
559 code = code == LE ? LT : GE; |
0 | 560 } |
561 break; | |
562 | |
563 default: | |
564 break; | |
565 } | |
566 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
567 |
0 | 568 if (CONSTANT_P (op1) && GET_CODE (op1) != CONST_INT) |
569 op1 = force_reg (SImode, op1); | |
570 | |
571 /* cmpnei: 0-31 (K immediate) | |
572 cmplti: 1-32 (J immediate, 0 using btsti x,31). */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
573 invert = false; |
0 | 574 switch (code) |
575 { | |
576 case EQ: /* Use inverted condition, cmpne. */ | |
577 code = NE; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
578 invert = true; |
0 | 579 /* Drop through. */ |
580 | |
581 case NE: /* Use normal condition, cmpne. */ | |
582 if (GET_CODE (op1) == CONST_INT && ! CONST_OK_FOR_K (INTVAL (op1))) | |
583 op1 = force_reg (SImode, op1); | |
584 break; | |
585 | |
586 case LE: /* Use inverted condition, reversed cmplt. */ | |
587 code = GT; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
588 invert = true; |
0 | 589 /* Drop through. */ |
590 | |
591 case GT: /* Use normal condition, reversed cmplt. */ | |
592 if (GET_CODE (op1) == CONST_INT) | |
593 op1 = force_reg (SImode, op1); | |
594 break; | |
595 | |
596 case GE: /* Use inverted condition, cmplt. */ | |
597 code = LT; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
598 invert = true; |
0 | 599 /* Drop through. */ |
600 | |
601 case LT: /* Use normal condition, cmplt. */ | |
602 if (GET_CODE (op1) == CONST_INT && | |
603 /* covered by btsti x,31. */ | |
604 INTVAL (op1) != 0 && | |
605 ! CONST_OK_FOR_J (INTVAL (op1))) | |
606 op1 = force_reg (SImode, op1); | |
607 break; | |
608 | |
609 case GTU: /* Use inverted condition, cmple. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
610 /* We coped with unsigned > 0 above. */ |
0 | 611 gcc_assert (GET_CODE (op1) != CONST_INT || INTVAL (op1) != 0); |
612 code = LEU; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
613 invert = true; |
0 | 614 /* Drop through. */ |
615 | |
616 case LEU: /* Use normal condition, reversed cmphs. */ | |
617 if (GET_CODE (op1) == CONST_INT && INTVAL (op1) != 0) | |
618 op1 = force_reg (SImode, op1); | |
619 break; | |
620 | |
621 case LTU: /* Use inverted condition, cmphs. */ | |
622 code = GEU; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
623 invert = true; |
0 | 624 /* Drop through. */ |
625 | |
626 case GEU: /* Use normal condition, cmphs. */ | |
627 if (GET_CODE (op1) == CONST_INT && INTVAL (op1) != 0) | |
628 op1 = force_reg (SImode, op1); | |
629 break; | |
630 | |
631 default: | |
632 break; | |
633 } | |
634 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
635 emit_insn (gen_rtx_SET (VOIDmode, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
636 cc_reg, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
637 gen_rtx_fmt_ee (code, CCmode, op0, op1))); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
638 return invert; |
0 | 639 } |
640 | |
641 int | |
642 mcore_symbolic_address_p (rtx x) | |
643 { | |
644 switch (GET_CODE (x)) | |
645 { | |
646 case SYMBOL_REF: | |
647 case LABEL_REF: | |
648 return 1; | |
649 case CONST: | |
650 x = XEXP (x, 0); | |
651 return ( (GET_CODE (XEXP (x, 0)) == SYMBOL_REF | |
652 || GET_CODE (XEXP (x, 0)) == LABEL_REF) | |
653 && GET_CODE (XEXP (x, 1)) == CONST_INT); | |
654 default: | |
655 return 0; | |
656 } | |
657 } | |
658 | |
659 /* Functions to output assembly code for a function call. */ | |
660 | |
661 char * | |
662 mcore_output_call (rtx operands[], int index) | |
663 { | |
664 static char buffer[20]; | |
665 rtx addr = operands [index]; | |
666 | |
667 if (REG_P (addr)) | |
668 { | |
669 if (TARGET_CG_DATA) | |
670 { | |
671 gcc_assert (mcore_current_function_name); | |
672 | |
673 ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, | |
674 "unknown", 1); | |
675 } | |
676 | |
677 sprintf (buffer, "jsr\t%%%d", index); | |
678 } | |
679 else | |
680 { | |
681 if (TARGET_CG_DATA) | |
682 { | |
683 gcc_assert (mcore_current_function_name); | |
684 gcc_assert (GET_CODE (addr) == SYMBOL_REF); | |
685 | |
686 ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, | |
687 XSTR (addr, 0), 0); | |
688 } | |
689 | |
690 sprintf (buffer, "jbsr\t%%%d", index); | |
691 } | |
692 | |
693 return buffer; | |
694 } | |
695 | |
696 /* Can we load a constant with a single instruction ? */ | |
697 | |
698 int | |
699 const_ok_for_mcore (HOST_WIDE_INT value) | |
700 { | |
701 if (value >= 0 && value <= 127) | |
702 return 1; | |
703 | |
704 /* Try exact power of two. */ | |
705 if (CONST_OK_FOR_M (value)) | |
706 return 1; | |
707 | |
708 /* Try exact power of two - 1. */ | |
709 if (CONST_OK_FOR_N (value) && value != -1) | |
710 return 1; | |
711 | |
712 return 0; | |
713 } | |
714 | |
715 /* Can we load a constant inline with up to 2 instructions ? */ | |
716 | |
717 int | |
718 mcore_const_ok_for_inline (HOST_WIDE_INT value) | |
719 { | |
720 HOST_WIDE_INT x, y; | |
721 | |
722 return try_constant_tricks (value, & x, & y) > 0; | |
723 } | |
724 | |
725 /* Are we loading the constant using a not ? */ | |
726 | |
727 int | |
728 mcore_const_trick_uses_not (HOST_WIDE_INT value) | |
729 { | |
730 HOST_WIDE_INT x, y; | |
731 | |
732 return try_constant_tricks (value, & x, & y) == 2; | |
733 } | |
734 | |
735 /* Try tricks to load a constant inline and return the trick number if | |
736 success (0 is non-inlinable). | |
737 | |
738 0: not inlinable | |
739 1: single instruction (do the usual thing) | |
740 2: single insn followed by a 'not' | |
741 3: single insn followed by a subi | |
742 4: single insn followed by an addi | |
743 5: single insn followed by rsubi | |
744 6: single insn followed by bseti | |
745 7: single insn followed by bclri | |
746 8: single insn followed by rotli | |
747 9: single insn followed by lsli | |
748 10: single insn followed by ixh | |
749 11: single insn followed by ixw. */ | |
750 | |
751 static int | |
752 try_constant_tricks (HOST_WIDE_INT value, HOST_WIDE_INT * x, HOST_WIDE_INT * y) | |
753 { | |
754 HOST_WIDE_INT i; | |
755 unsigned HOST_WIDE_INT bit, shf, rot; | |
756 | |
757 if (const_ok_for_mcore (value)) | |
758 return 1; /* Do the usual thing. */ | |
759 | |
760 if (! TARGET_HARDLIT) | |
761 return 0; | |
762 | |
763 if (const_ok_for_mcore (~value)) | |
764 { | |
765 *x = ~value; | |
766 return 2; | |
767 } | |
768 | |
769 for (i = 1; i <= 32; i++) | |
770 { | |
771 if (const_ok_for_mcore (value - i)) | |
772 { | |
773 *x = value - i; | |
774 *y = i; | |
775 | |
776 return 3; | |
777 } | |
778 | |
779 if (const_ok_for_mcore (value + i)) | |
780 { | |
781 *x = value + i; | |
782 *y = i; | |
783 | |
784 return 4; | |
785 } | |
786 } | |
787 | |
788 bit = 0x80000000ULL; | |
789 | |
790 for (i = 0; i <= 31; i++) | |
791 { | |
792 if (const_ok_for_mcore (i - value)) | |
793 { | |
794 *x = i - value; | |
795 *y = i; | |
796 | |
797 return 5; | |
798 } | |
799 | |
800 if (const_ok_for_mcore (value & ~bit)) | |
801 { | |
802 *y = bit; | |
803 *x = value & ~bit; | |
804 return 6; | |
805 } | |
806 | |
807 if (const_ok_for_mcore (value | bit)) | |
808 { | |
809 *y = ~bit; | |
810 *x = value | bit; | |
811 | |
812 return 7; | |
813 } | |
814 | |
815 bit >>= 1; | |
816 } | |
817 | |
818 shf = value; | |
819 rot = value; | |
820 | |
821 for (i = 1; i < 31; i++) | |
822 { | |
823 int c; | |
824 | |
825 /* MCore has rotate left. */ | |
826 c = rot << 31; | |
827 rot >>= 1; | |
828 rot &= 0x7FFFFFFF; | |
829 rot |= c; /* Simulate rotate. */ | |
830 | |
831 if (const_ok_for_mcore (rot)) | |
832 { | |
833 *y = i; | |
834 *x = rot; | |
835 | |
836 return 8; | |
837 } | |
838 | |
839 if (shf & 1) | |
840 shf = 0; /* Can't use logical shift, low order bit is one. */ | |
841 | |
842 shf >>= 1; | |
843 | |
844 if (shf != 0 && const_ok_for_mcore (shf)) | |
845 { | |
846 *y = i; | |
847 *x = shf; | |
848 | |
849 return 9; | |
850 } | |
851 } | |
852 | |
853 if ((value % 3) == 0 && const_ok_for_mcore (value / 3)) | |
854 { | |
855 *x = value / 3; | |
856 | |
857 return 10; | |
858 } | |
859 | |
860 if ((value % 5) == 0 && const_ok_for_mcore (value / 5)) | |
861 { | |
862 *x = value / 5; | |
863 | |
864 return 11; | |
865 } | |
866 | |
867 return 0; | |
868 } | |
869 | |
870 /* Check whether reg is dead at first. This is done by searching ahead | |
871 for either the next use (i.e., reg is live), a death note, or a set of | |
872 reg. Don't just use dead_or_set_p() since reload does not always mark | |
873 deaths (especially if PRESERVE_DEATH_NOTES_REGNO_P is not defined). We | |
874 can ignore subregs by extracting the actual register. BRC */ | |
875 | |
876 int | |
877 mcore_is_dead (rtx first, rtx reg) | |
878 { | |
879 rtx insn; | |
880 | |
881 /* For mcore, subregs can't live independently of their parent regs. */ | |
882 if (GET_CODE (reg) == SUBREG) | |
883 reg = SUBREG_REG (reg); | |
884 | |
885 /* Dies immediately. */ | |
886 if (dead_or_set_p (first, reg)) | |
887 return 1; | |
888 | |
889 /* Look for conclusive evidence of live/death, otherwise we have | |
890 to assume that it is live. */ | |
891 for (insn = NEXT_INSN (first); insn; insn = NEXT_INSN (insn)) | |
892 { | |
893 if (GET_CODE (insn) == JUMP_INSN) | |
894 return 0; /* We lose track, assume it is alive. */ | |
895 | |
896 else if (GET_CODE(insn) == CALL_INSN) | |
897 { | |
898 /* Call's might use it for target or register parms. */ | |
899 if (reg_referenced_p (reg, PATTERN (insn)) | |
900 || find_reg_fusage (insn, USE, reg)) | |
901 return 0; | |
902 else if (dead_or_set_p (insn, reg)) | |
903 return 1; | |
904 } | |
905 else if (GET_CODE (insn) == INSN) | |
906 { | |
907 if (reg_referenced_p (reg, PATTERN (insn))) | |
908 return 0; | |
909 else if (dead_or_set_p (insn, reg)) | |
910 return 1; | |
911 } | |
912 } | |
913 | |
914 /* No conclusive evidence either way, we cannot take the chance | |
915 that control flow hid the use from us -- "I'm not dead yet". */ | |
916 return 0; | |
917 } | |
918 | |
919 /* Count the number of ones in mask. */ | |
920 | |
921 int | |
922 mcore_num_ones (HOST_WIDE_INT mask) | |
923 { | |
924 /* A trick to count set bits recently posted on comp.compilers. */ | |
925 mask = (mask >> 1 & 0x55555555) + (mask & 0x55555555); | |
926 mask = ((mask >> 2) & 0x33333333) + (mask & 0x33333333); | |
927 mask = ((mask >> 4) + mask) & 0x0f0f0f0f; | |
928 mask = ((mask >> 8) + mask); | |
929 | |
930 return (mask + (mask >> 16)) & 0xff; | |
931 } | |
932 | |
933 /* Count the number of zeros in mask. */ | |
934 | |
935 int | |
936 mcore_num_zeros (HOST_WIDE_INT mask) | |
937 { | |
938 return 32 - mcore_num_ones (mask); | |
939 } | |
940 | |
941 /* Determine byte being masked. */ | |
942 | |
943 int | |
944 mcore_byte_offset (unsigned int mask) | |
945 { | |
946 if (mask == 0x00ffffffL) | |
947 return 0; | |
948 else if (mask == 0xff00ffffL) | |
949 return 1; | |
950 else if (mask == 0xffff00ffL) | |
951 return 2; | |
952 else if (mask == 0xffffff00L) | |
953 return 3; | |
954 | |
955 return -1; | |
956 } | |
957 | |
958 /* Determine halfword being masked. */ | |
959 | |
960 int | |
961 mcore_halfword_offset (unsigned int mask) | |
962 { | |
963 if (mask == 0x0000ffffL) | |
964 return 0; | |
965 else if (mask == 0xffff0000L) | |
966 return 1; | |
967 | |
968 return -1; | |
969 } | |
970 | |
971 /* Output a series of bseti's corresponding to mask. */ | |
972 | |
973 const char * | |
974 mcore_output_bseti (rtx dst, int mask) | |
975 { | |
976 rtx out_operands[2]; | |
977 int bit; | |
978 | |
979 out_operands[0] = dst; | |
980 | |
981 for (bit = 0; bit < 32; bit++) | |
982 { | |
983 if ((mask & 0x1) == 0x1) | |
984 { | |
985 out_operands[1] = GEN_INT (bit); | |
986 | |
987 output_asm_insn ("bseti\t%0,%1", out_operands); | |
988 } | |
989 mask >>= 1; | |
990 } | |
991 | |
992 return ""; | |
993 } | |
994 | |
995 /* Output a series of bclri's corresponding to mask. */ | |
996 | |
997 const char * | |
998 mcore_output_bclri (rtx dst, int mask) | |
999 { | |
1000 rtx out_operands[2]; | |
1001 int bit; | |
1002 | |
1003 out_operands[0] = dst; | |
1004 | |
1005 for (bit = 0; bit < 32; bit++) | |
1006 { | |
1007 if ((mask & 0x1) == 0x0) | |
1008 { | |
1009 out_operands[1] = GEN_INT (bit); | |
1010 | |
1011 output_asm_insn ("bclri\t%0,%1", out_operands); | |
1012 } | |
1013 | |
1014 mask >>= 1; | |
1015 } | |
1016 | |
1017 return ""; | |
1018 } | |
1019 | |
1020 /* Output a conditional move of two constants that are +/- 1 within each | |
1021 other. See the "movtK" patterns in mcore.md. I'm not sure this is | |
1022 really worth the effort. */ | |
1023 | |
1024 const char * | |
1025 mcore_output_cmov (rtx operands[], int cmp_t, const char * test) | |
1026 { | |
1027 HOST_WIDE_INT load_value; | |
1028 HOST_WIDE_INT adjust_value; | |
1029 rtx out_operands[4]; | |
1030 | |
1031 out_operands[0] = operands[0]; | |
1032 | |
1033 /* Check to see which constant is loadable. */ | |
1034 if (const_ok_for_mcore (INTVAL (operands[1]))) | |
1035 { | |
1036 out_operands[1] = operands[1]; | |
1037 out_operands[2] = operands[2]; | |
1038 } | |
1039 else if (const_ok_for_mcore (INTVAL (operands[2]))) | |
1040 { | |
1041 out_operands[1] = operands[2]; | |
1042 out_operands[2] = operands[1]; | |
1043 | |
1044 /* Complement test since constants are swapped. */ | |
1045 cmp_t = (cmp_t == 0); | |
1046 } | |
1047 load_value = INTVAL (out_operands[1]); | |
1048 adjust_value = INTVAL (out_operands[2]); | |
1049 | |
1050 /* First output the test if folded into the pattern. */ | |
1051 | |
1052 if (test) | |
1053 output_asm_insn (test, operands); | |
1054 | |
1055 /* Load the constant - for now, only support constants that can be | |
1056 generated with a single instruction. maybe add general inlinable | |
1057 constants later (this will increase the # of patterns since the | |
1058 instruction sequence has a different length attribute). */ | |
1059 if (load_value >= 0 && load_value <= 127) | |
1060 output_asm_insn ("movi\t%0,%1", out_operands); | |
1061 else if (CONST_OK_FOR_M (load_value)) | |
1062 output_asm_insn ("bgeni\t%0,%P1", out_operands); | |
1063 else if (CONST_OK_FOR_N (load_value)) | |
1064 output_asm_insn ("bmaski\t%0,%N1", out_operands); | |
1065 | |
1066 /* Output the constant adjustment. */ | |
1067 if (load_value > adjust_value) | |
1068 { | |
1069 if (cmp_t) | |
1070 output_asm_insn ("decf\t%0", out_operands); | |
1071 else | |
1072 output_asm_insn ("dect\t%0", out_operands); | |
1073 } | |
1074 else | |
1075 { | |
1076 if (cmp_t) | |
1077 output_asm_insn ("incf\t%0", out_operands); | |
1078 else | |
1079 output_asm_insn ("inct\t%0", out_operands); | |
1080 } | |
1081 | |
1082 return ""; | |
1083 } | |
1084 | |
1085 /* Outputs the peephole for moving a constant that gets not'ed followed | |
1086 by an and (i.e. combine the not and the and into andn). BRC */ | |
1087 | |
1088 const char * | |
1089 mcore_output_andn (rtx insn ATTRIBUTE_UNUSED, rtx operands[]) | |
1090 { | |
1091 HOST_WIDE_INT x, y; | |
1092 rtx out_operands[3]; | |
1093 const char * load_op; | |
1094 char buf[256]; | |
1095 int trick_no; | |
1096 | |
1097 trick_no = try_constant_tricks (INTVAL (operands[1]), &x, &y); | |
1098 gcc_assert (trick_no == 2); | |
1099 | |
1100 out_operands[0] = operands[0]; | |
1101 out_operands[1] = GEN_INT (x); | |
1102 out_operands[2] = operands[2]; | |
1103 | |
1104 if (x >= 0 && x <= 127) | |
1105 load_op = "movi\t%0,%1"; | |
1106 | |
1107 /* Try exact power of two. */ | |
1108 else if (CONST_OK_FOR_M (x)) | |
1109 load_op = "bgeni\t%0,%P1"; | |
1110 | |
1111 /* Try exact power of two - 1. */ | |
1112 else if (CONST_OK_FOR_N (x)) | |
1113 load_op = "bmaski\t%0,%N1"; | |
1114 | |
1115 else | |
1116 { | |
1117 load_op = "BADMOVI-andn\t%0, %1"; | |
1118 gcc_unreachable (); | |
1119 } | |
1120 | |
1121 sprintf (buf, "%s\n\tandn\t%%2,%%0", load_op); | |
1122 output_asm_insn (buf, out_operands); | |
1123 | |
1124 return ""; | |
1125 } | |
1126 | |
1127 /* Output an inline constant. */ | |
1128 | |
1129 static const char * | |
1130 output_inline_const (enum machine_mode mode, rtx operands[]) | |
1131 { | |
1132 HOST_WIDE_INT x = 0, y = 0; | |
1133 int trick_no; | |
1134 rtx out_operands[3]; | |
1135 char buf[256]; | |
1136 char load_op[256]; | |
1137 const char *dst_fmt; | |
1138 HOST_WIDE_INT value; | |
1139 | |
1140 value = INTVAL (operands[1]); | |
1141 | |
1142 trick_no = try_constant_tricks (value, &x, &y); | |
1143 /* lrw's are handled separately: Large inlinable constants never get | |
1144 turned into lrw's. Our caller uses try_constant_tricks to back | |
1145 off to an lrw rather than calling this routine. */ | |
1146 gcc_assert (trick_no != 0); | |
1147 | |
1148 if (trick_no == 1) | |
1149 x = value; | |
1150 | |
1151 /* operands: 0 = dst, 1 = load immed., 2 = immed. adjustment. */ | |
1152 out_operands[0] = operands[0]; | |
1153 out_operands[1] = GEN_INT (x); | |
1154 | |
1155 if (trick_no > 2) | |
1156 out_operands[2] = GEN_INT (y); | |
1157 | |
1158 /* Select dst format based on mode. */ | |
1159 if (mode == DImode && (! TARGET_LITTLE_END)) | |
1160 dst_fmt = "%R0"; | |
1161 else | |
1162 dst_fmt = "%0"; | |
1163 | |
1164 if (x >= 0 && x <= 127) | |
1165 sprintf (load_op, "movi\t%s,%%1", dst_fmt); | |
1166 | |
1167 /* Try exact power of two. */ | |
1168 else if (CONST_OK_FOR_M (x)) | |
1169 sprintf (load_op, "bgeni\t%s,%%P1", dst_fmt); | |
1170 | |
1171 /* Try exact power of two - 1. */ | |
1172 else if (CONST_OK_FOR_N (x)) | |
1173 sprintf (load_op, "bmaski\t%s,%%N1", dst_fmt); | |
1174 | |
1175 else | |
1176 { | |
1177 sprintf (load_op, "BADMOVI-inline_const %s, %%1", dst_fmt); | |
1178 gcc_unreachable (); | |
1179 } | |
1180 | |
1181 switch (trick_no) | |
1182 { | |
1183 case 1: | |
1184 strcpy (buf, load_op); | |
1185 break; | |
1186 case 2: /* not */ | |
1187 sprintf (buf, "%s\n\tnot\t%s\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1188 break; | |
1189 case 3: /* add */ | |
1190 sprintf (buf, "%s\n\taddi\t%s,%%2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1191 break; | |
1192 case 4: /* sub */ | |
1193 sprintf (buf, "%s\n\tsubi\t%s,%%2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1194 break; | |
1195 case 5: /* rsub */ | |
1196 /* Never happens unless -mrsubi, see try_constant_tricks(). */ | |
1197 sprintf (buf, "%s\n\trsubi\t%s,%%2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1198 break; | |
1199 case 6: /* bseti */ | |
1200 sprintf (buf, "%s\n\tbseti\t%s,%%P2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1201 break; | |
1202 case 7: /* bclr */ | |
1203 sprintf (buf, "%s\n\tbclri\t%s,%%Q2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1204 break; | |
1205 case 8: /* rotl */ | |
1206 sprintf (buf, "%s\n\trotli\t%s,%%2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1207 break; | |
1208 case 9: /* lsl */ | |
1209 sprintf (buf, "%s\n\tlsli\t%s,%%2\t// %ld 0x%lx", load_op, dst_fmt, value, value); | |
1210 break; | |
1211 case 10: /* ixh */ | |
1212 sprintf (buf, "%s\n\tixh\t%s,%s\t// %ld 0x%lx", load_op, dst_fmt, dst_fmt, value, value); | |
1213 break; | |
1214 case 11: /* ixw */ | |
1215 sprintf (buf, "%s\n\tixw\t%s,%s\t// %ld 0x%lx", load_op, dst_fmt, dst_fmt, value, value); | |
1216 break; | |
1217 default: | |
1218 return ""; | |
1219 } | |
1220 | |
1221 output_asm_insn (buf, out_operands); | |
1222 | |
1223 return ""; | |
1224 } | |
1225 | |
1226 /* Output a move of a word or less value. */ | |
1227 | |
1228 const char * | |
1229 mcore_output_move (rtx insn ATTRIBUTE_UNUSED, rtx operands[], | |
1230 enum machine_mode mode ATTRIBUTE_UNUSED) | |
1231 { | |
1232 rtx dst = operands[0]; | |
1233 rtx src = operands[1]; | |
1234 | |
1235 if (GET_CODE (dst) == REG) | |
1236 { | |
1237 if (GET_CODE (src) == REG) | |
1238 { | |
1239 if (REGNO (src) == CC_REG) /* r-c */ | |
1240 return "mvc\t%0"; | |
1241 else | |
1242 return "mov\t%0,%1"; /* r-r*/ | |
1243 } | |
1244 else if (GET_CODE (src) == MEM) | |
1245 { | |
1246 if (GET_CODE (XEXP (src, 0)) == LABEL_REF) | |
1247 return "lrw\t%0,[%1]"; /* a-R */ | |
1248 else | |
1249 switch (GET_MODE (src)) /* r-m */ | |
1250 { | |
1251 case SImode: | |
1252 return "ldw\t%0,%1"; | |
1253 case HImode: | |
1254 return "ld.h\t%0,%1"; | |
1255 case QImode: | |
1256 return "ld.b\t%0,%1"; | |
1257 default: | |
1258 gcc_unreachable (); | |
1259 } | |
1260 } | |
1261 else if (GET_CODE (src) == CONST_INT) | |
1262 { | |
1263 HOST_WIDE_INT x, y; | |
1264 | |
1265 if (CONST_OK_FOR_I (INTVAL (src))) /* r-I */ | |
1266 return "movi\t%0,%1"; | |
1267 else if (CONST_OK_FOR_M (INTVAL (src))) /* r-M */ | |
1268 return "bgeni\t%0,%P1\t// %1 %x1"; | |
1269 else if (CONST_OK_FOR_N (INTVAL (src))) /* r-N */ | |
1270 return "bmaski\t%0,%N1\t// %1 %x1"; | |
1271 else if (try_constant_tricks (INTVAL (src), &x, &y)) /* R-P */ | |
1272 return output_inline_const (SImode, operands); /* 1-2 insns */ | |
1273 else | |
1274 return "lrw\t%0,%x1\t// %1"; /* Get it from literal pool. */ | |
1275 } | |
1276 else | |
1277 return "lrw\t%0, %1"; /* Into the literal pool. */ | |
1278 } | |
1279 else if (GET_CODE (dst) == MEM) /* m-r */ | |
1280 switch (GET_MODE (dst)) | |
1281 { | |
1282 case SImode: | |
1283 return "stw\t%1,%0"; | |
1284 case HImode: | |
1285 return "st.h\t%1,%0"; | |
1286 case QImode: | |
1287 return "st.b\t%1,%0"; | |
1288 default: | |
1289 gcc_unreachable (); | |
1290 } | |
1291 | |
1292 gcc_unreachable (); | |
1293 } | |
1294 | |
1295 /* Return a sequence of instructions to perform DI or DF move. | |
1296 Since the MCORE cannot move a DI or DF in one instruction, we have | |
1297 to take care when we see overlapping source and dest registers. */ | |
1298 | |
1299 const char * | |
1300 mcore_output_movedouble (rtx operands[], enum machine_mode mode ATTRIBUTE_UNUSED) | |
1301 { | |
1302 rtx dst = operands[0]; | |
1303 rtx src = operands[1]; | |
1304 | |
1305 if (GET_CODE (dst) == REG) | |
1306 { | |
1307 if (GET_CODE (src) == REG) | |
1308 { | |
1309 int dstreg = REGNO (dst); | |
1310 int srcreg = REGNO (src); | |
1311 | |
1312 /* Ensure the second source not overwritten. */ | |
1313 if (srcreg + 1 == dstreg) | |
1314 return "mov %R0,%R1\n\tmov %0,%1"; | |
1315 else | |
1316 return "mov %0,%1\n\tmov %R0,%R1"; | |
1317 } | |
1318 else if (GET_CODE (src) == MEM) | |
1319 { | |
1320 rtx memexp = memexp = XEXP (src, 0); | |
1321 int dstreg = REGNO (dst); | |
1322 int basereg = -1; | |
1323 | |
1324 if (GET_CODE (memexp) == LABEL_REF) | |
1325 return "lrw\t%0,[%1]\n\tlrw\t%R0,[%R1]"; | |
1326 else if (GET_CODE (memexp) == REG) | |
1327 basereg = REGNO (memexp); | |
1328 else if (GET_CODE (memexp) == PLUS) | |
1329 { | |
1330 if (GET_CODE (XEXP (memexp, 0)) == REG) | |
1331 basereg = REGNO (XEXP (memexp, 0)); | |
1332 else if (GET_CODE (XEXP (memexp, 1)) == REG) | |
1333 basereg = REGNO (XEXP (memexp, 1)); | |
1334 else | |
1335 gcc_unreachable (); | |
1336 } | |
1337 else | |
1338 gcc_unreachable (); | |
1339 | |
1340 /* ??? length attribute is wrong here. */ | |
1341 if (dstreg == basereg) | |
1342 { | |
1343 /* Just load them in reverse order. */ | |
1344 return "ldw\t%R0,%R1\n\tldw\t%0,%1"; | |
1345 | |
1346 /* XXX: alternative: move basereg to basereg+1 | |
1347 and then fall through. */ | |
1348 } | |
1349 else | |
1350 return "ldw\t%0,%1\n\tldw\t%R0,%R1"; | |
1351 } | |
1352 else if (GET_CODE (src) == CONST_INT) | |
1353 { | |
1354 if (TARGET_LITTLE_END) | |
1355 { | |
1356 if (CONST_OK_FOR_I (INTVAL (src))) | |
1357 output_asm_insn ("movi %0,%1", operands); | |
1358 else if (CONST_OK_FOR_M (INTVAL (src))) | |
1359 output_asm_insn ("bgeni %0,%P1", operands); | |
1360 else if (CONST_OK_FOR_N (INTVAL (src))) | |
1361 output_asm_insn ("bmaski %0,%N1", operands); | |
1362 else | |
1363 gcc_unreachable (); | |
1364 | |
1365 if (INTVAL (src) < 0) | |
1366 return "bmaski %R0,32"; | |
1367 else | |
1368 return "movi %R0,0"; | |
1369 } | |
1370 else | |
1371 { | |
1372 if (CONST_OK_FOR_I (INTVAL (src))) | |
1373 output_asm_insn ("movi %R0,%1", operands); | |
1374 else if (CONST_OK_FOR_M (INTVAL (src))) | |
1375 output_asm_insn ("bgeni %R0,%P1", operands); | |
1376 else if (CONST_OK_FOR_N (INTVAL (src))) | |
1377 output_asm_insn ("bmaski %R0,%N1", operands); | |
1378 else | |
1379 gcc_unreachable (); | |
1380 | |
1381 if (INTVAL (src) < 0) | |
1382 return "bmaski %0,32"; | |
1383 else | |
1384 return "movi %0,0"; | |
1385 } | |
1386 } | |
1387 else | |
1388 gcc_unreachable (); | |
1389 } | |
1390 else if (GET_CODE (dst) == MEM && GET_CODE (src) == REG) | |
1391 return "stw\t%1,%0\n\tstw\t%R1,%R0"; | |
1392 else | |
1393 gcc_unreachable (); | |
1394 } | |
1395 | |
1396 /* Predicates used by the templates. */ | |
1397 | |
1398 int | |
1399 mcore_arith_S_operand (rtx op) | |
1400 { | |
1401 if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_M (~INTVAL (op))) | |
1402 return 1; | |
1403 | |
1404 return 0; | |
1405 } | |
1406 | |
1407 /* Expand insert bit field. BRC */ | |
1408 | |
1409 int | |
1410 mcore_expand_insv (rtx operands[]) | |
1411 { | |
1412 int width = INTVAL (operands[1]); | |
1413 int posn = INTVAL (operands[2]); | |
1414 int mask; | |
1415 rtx mreg, sreg, ereg; | |
1416 | |
1417 /* To get width 1 insv, the test in store_bit_field() (expmed.c, line 191) | |
1418 for width==1 must be removed. Look around line 368. This is something | |
1419 we really want the md part to do. */ | |
1420 if (width == 1 && GET_CODE (operands[3]) == CONST_INT) | |
1421 { | |
1422 /* Do directly with bseti or bclri. */ | |
1423 /* RBE: 2/97 consider only low bit of constant. */ | |
1424 if ((INTVAL (operands[3]) & 1) == 0) | |
1425 { | |
1426 mask = ~(1 << posn); | |
1427 emit_insn (gen_rtx_SET (SImode, operands[0], | |
1428 gen_rtx_AND (SImode, operands[0], GEN_INT (mask)))); | |
1429 } | |
1430 else | |
1431 { | |
1432 mask = 1 << posn; | |
1433 emit_insn (gen_rtx_SET (SImode, operands[0], | |
1434 gen_rtx_IOR (SImode, operands[0], GEN_INT (mask)))); | |
1435 } | |
1436 | |
1437 return 1; | |
1438 } | |
1439 | |
1440 /* Look at some bit-field placements that we aren't interested | |
1441 in handling ourselves, unless specifically directed to do so. */ | |
1442 if (! TARGET_W_FIELD) | |
1443 return 0; /* Generally, give up about now. */ | |
1444 | |
1445 if (width == 8 && posn % 8 == 0) | |
1446 /* Byte sized and aligned; let caller break it up. */ | |
1447 return 0; | |
1448 | |
1449 if (width == 16 && posn % 16 == 0) | |
1450 /* Short sized and aligned; let caller break it up. */ | |
1451 return 0; | |
1452 | |
1453 /* The general case - we can do this a little bit better than what the | |
1454 machine independent part tries. This will get rid of all the subregs | |
1455 that mess up constant folding in combine when working with relaxed | |
1456 immediates. */ | |
1457 | |
1458 /* If setting the entire field, do it directly. */ | |
1459 if (GET_CODE (operands[3]) == CONST_INT | |
1460 && INTVAL (operands[3]) == ((1 << width) - 1)) | |
1461 { | |
1462 mreg = force_reg (SImode, GEN_INT (INTVAL (operands[3]) << posn)); | |
1463 emit_insn (gen_rtx_SET (SImode, operands[0], | |
1464 gen_rtx_IOR (SImode, operands[0], mreg))); | |
1465 return 1; | |
1466 } | |
1467 | |
1468 /* Generate the clear mask. */ | |
1469 mreg = force_reg (SImode, GEN_INT (~(((1 << width) - 1) << posn))); | |
1470 | |
1471 /* Clear the field, to overlay it later with the source. */ | |
1472 emit_insn (gen_rtx_SET (SImode, operands[0], | |
1473 gen_rtx_AND (SImode, operands[0], mreg))); | |
1474 | |
1475 /* If the source is constant 0, we've nothing to add back. */ | |
1476 if (GET_CODE (operands[3]) == CONST_INT && INTVAL (operands[3]) == 0) | |
1477 return 1; | |
1478 | |
1479 /* XXX: Should we worry about more games with constant values? | |
1480 We've covered the high profile: set/clear single-bit and many-bit | |
1481 fields. How often do we see "arbitrary bit pattern" constants? */ | |
1482 sreg = copy_to_mode_reg (SImode, operands[3]); | |
1483 | |
1484 /* Extract src as same width as dst (needed for signed values). We | |
1485 always have to do this since we widen everything to SImode. | |
1486 We don't have to mask if we're shifting this up against the | |
1487 MSB of the register (e.g., the shift will push out any hi-order | |
1488 bits. */ | |
1489 if (width + posn != (int) GET_MODE_SIZE (SImode)) | |
1490 { | |
1491 ereg = force_reg (SImode, GEN_INT ((1 << width) - 1)); | |
1492 emit_insn (gen_rtx_SET (SImode, sreg, | |
1493 gen_rtx_AND (SImode, sreg, ereg))); | |
1494 } | |
1495 | |
1496 /* Insert source value in dest. */ | |
1497 if (posn != 0) | |
1498 emit_insn (gen_rtx_SET (SImode, sreg, | |
1499 gen_rtx_ASHIFT (SImode, sreg, GEN_INT (posn)))); | |
1500 | |
1501 emit_insn (gen_rtx_SET (SImode, operands[0], | |
1502 gen_rtx_IOR (SImode, operands[0], sreg))); | |
1503 | |
1504 return 1; | |
1505 } | |
1506 | |
1507 /* ??? Block move stuff stolen from m88k. This code has not been | |
1508 verified for correctness. */ | |
1509 | |
1510 /* Emit code to perform a block move. Choose the best method. | |
1511 | |
1512 OPERANDS[0] is the destination. | |
1513 OPERANDS[1] is the source. | |
1514 OPERANDS[2] is the size. | |
1515 OPERANDS[3] is the alignment safe to use. */ | |
1516 | |
1517 /* Emit code to perform a block move with an offset sequence of ldw/st | |
1518 instructions (..., ldw 0, stw 1, ldw 1, stw 0, ...). SIZE and ALIGN are | |
1519 known constants. DEST and SRC are registers. OFFSET is the known | |
1520 starting point for the output pattern. */ | |
1521 | |
1522 static const enum machine_mode mode_from_align[] = | |
1523 { | |
1524 VOIDmode, QImode, HImode, VOIDmode, SImode, | |
1525 }; | |
1526 | |
1527 static void | |
1528 block_move_sequence (rtx dst_mem, rtx src_mem, int size, int align) | |
1529 { | |
1530 rtx temp[2]; | |
1531 enum machine_mode mode[2]; | |
1532 int amount[2]; | |
1533 bool active[2]; | |
1534 int phase = 0; | |
1535 int next; | |
1536 int offset_ld = 0; | |
1537 int offset_st = 0; | |
1538 rtx x; | |
1539 | |
1540 x = XEXP (dst_mem, 0); | |
1541 if (!REG_P (x)) | |
1542 { | |
1543 x = force_reg (Pmode, x); | |
1544 dst_mem = replace_equiv_address (dst_mem, x); | |
1545 } | |
1546 | |
1547 x = XEXP (src_mem, 0); | |
1548 if (!REG_P (x)) | |
1549 { | |
1550 x = force_reg (Pmode, x); | |
1551 src_mem = replace_equiv_address (src_mem, x); | |
1552 } | |
1553 | |
1554 active[0] = active[1] = false; | |
1555 | |
1556 do | |
1557 { | |
1558 next = phase; | |
1559 phase ^= 1; | |
1560 | |
1561 if (size > 0) | |
1562 { | |
1563 int next_amount; | |
1564 | |
1565 next_amount = (size >= 4 ? 4 : (size >= 2 ? 2 : 1)); | |
1566 next_amount = MIN (next_amount, align); | |
1567 | |
1568 amount[next] = next_amount; | |
1569 mode[next] = mode_from_align[next_amount]; | |
1570 temp[next] = gen_reg_rtx (mode[next]); | |
1571 | |
1572 x = adjust_address (src_mem, mode[next], offset_ld); | |
1573 emit_insn (gen_rtx_SET (VOIDmode, temp[next], x)); | |
1574 | |
1575 offset_ld += next_amount; | |
1576 size -= next_amount; | |
1577 active[next] = true; | |
1578 } | |
1579 | |
1580 if (active[phase]) | |
1581 { | |
1582 active[phase] = false; | |
1583 | |
1584 x = adjust_address (dst_mem, mode[phase], offset_st); | |
1585 emit_insn (gen_rtx_SET (VOIDmode, x, temp[phase])); | |
1586 | |
1587 offset_st += amount[phase]; | |
1588 } | |
1589 } | |
1590 while (active[next]); | |
1591 } | |
1592 | |
1593 bool | |
1594 mcore_expand_block_move (rtx *operands) | |
1595 { | |
1596 HOST_WIDE_INT align, bytes, max; | |
1597 | |
1598 if (GET_CODE (operands[2]) != CONST_INT) | |
1599 return false; | |
1600 | |
1601 bytes = INTVAL (operands[2]); | |
1602 align = INTVAL (operands[3]); | |
1603 | |
1604 if (bytes <= 0) | |
1605 return false; | |
1606 if (align > 4) | |
1607 align = 4; | |
1608 | |
1609 switch (align) | |
1610 { | |
1611 case 4: | |
1612 if (bytes & 1) | |
1613 max = 4*4; | |
1614 else if (bytes & 3) | |
1615 max = 8*4; | |
1616 else | |
1617 max = 16*4; | |
1618 break; | |
1619 case 2: | |
1620 max = 4*2; | |
1621 break; | |
1622 case 1: | |
1623 max = 4*1; | |
1624 break; | |
1625 default: | |
1626 gcc_unreachable (); | |
1627 } | |
1628 | |
1629 if (bytes <= max) | |
1630 { | |
1631 block_move_sequence (operands[0], operands[1], bytes, align); | |
1632 return true; | |
1633 } | |
1634 | |
1635 return false; | |
1636 } | |
1637 | |
1638 | |
1639 /* Code to generate prologue and epilogue sequences. */ | |
1640 static int number_of_regs_before_varargs; | |
1641 | |
1642 /* Set by TARGET_SETUP_INCOMING_VARARGS to indicate to prolog that this is | |
1643 for a varargs function. */ | |
1644 static int current_function_anonymous_args; | |
1645 | |
1646 #define STACK_BYTES (STACK_BOUNDARY/BITS_PER_UNIT) | |
1647 #define STORE_REACH (64) /* Maximum displace of word store + 4. */ | |
1648 #define ADDI_REACH (32) /* Maximum addi operand. */ | |
1649 | |
1650 static void | |
1651 layout_mcore_frame (struct mcore_frame * infp) | |
1652 { | |
1653 int n; | |
1654 unsigned int i; | |
1655 int nbytes; | |
1656 int regarg; | |
1657 int localregarg; | |
1658 int localreg; | |
1659 int outbounds; | |
1660 unsigned int growths; | |
1661 int step; | |
1662 | |
1663 /* Might have to spill bytes to re-assemble a big argument that | |
1664 was passed partially in registers and partially on the stack. */ | |
1665 nbytes = crtl->args.pretend_args_size; | |
1666 | |
1667 /* Determine how much space for spilled anonymous args (e.g., stdarg). */ | |
1668 if (current_function_anonymous_args) | |
1669 nbytes += (NPARM_REGS - number_of_regs_before_varargs) * UNITS_PER_WORD; | |
1670 | |
1671 infp->arg_size = nbytes; | |
1672 | |
1673 /* How much space to save non-volatile registers we stomp. */ | |
1674 infp->reg_mask = calc_live_regs (& n); | |
1675 infp->reg_size = n * 4; | |
1676 | |
1677 /* And the rest of it... locals and space for overflowed outbounds. */ | |
1678 infp->local_size = get_frame_size (); | |
1679 infp->outbound_size = crtl->outgoing_args_size; | |
1680 | |
1681 /* Make sure we have a whole number of words for the locals. */ | |
1682 if (infp->local_size % STACK_BYTES) | |
1683 infp->local_size = (infp->local_size + STACK_BYTES - 1) & ~ (STACK_BYTES -1); | |
1684 | |
1685 /* Only thing we know we have to pad is the outbound space, since | |
1686 we've aligned our locals assuming that base of locals is aligned. */ | |
1687 infp->pad_local = 0; | |
1688 infp->pad_reg = 0; | |
1689 infp->pad_outbound = 0; | |
1690 if (infp->outbound_size % STACK_BYTES) | |
1691 infp->pad_outbound = STACK_BYTES - (infp->outbound_size % STACK_BYTES); | |
1692 | |
1693 /* Now we see how we want to stage the prologue so that it does | |
1694 the most appropriate stack growth and register saves to either: | |
1695 (1) run fast, | |
1696 (2) reduce instruction space, or | |
1697 (3) reduce stack space. */ | |
1698 for (i = 0; i < ARRAY_SIZE (infp->growth); i++) | |
1699 infp->growth[i] = 0; | |
1700 | |
1701 regarg = infp->reg_size + infp->arg_size; | |
1702 localregarg = infp->local_size + regarg; | |
1703 localreg = infp->local_size + infp->reg_size; | |
1704 outbounds = infp->outbound_size + infp->pad_outbound; | |
1705 growths = 0; | |
1706 | |
1707 /* XXX: Consider one where we consider localregarg + outbound too! */ | |
1708 | |
1709 /* Frame of <= 32 bytes and using stm would get <= 2 registers. | |
1710 use stw's with offsets and buy the frame in one shot. */ | |
1711 if (localregarg <= ADDI_REACH | |
1712 && (infp->reg_size <= 8 || (infp->reg_mask & 0xc000) != 0xc000)) | |
1713 { | |
1714 /* Make sure we'll be aligned. */ | |
1715 if (localregarg % STACK_BYTES) | |
1716 infp->pad_reg = STACK_BYTES - (localregarg % STACK_BYTES); | |
1717 | |
1718 step = localregarg + infp->pad_reg; | |
1719 infp->reg_offset = infp->local_size; | |
1720 | |
1721 if (outbounds + step <= ADDI_REACH && !frame_pointer_needed) | |
1722 { | |
1723 step += outbounds; | |
1724 infp->reg_offset += outbounds; | |
1725 outbounds = 0; | |
1726 } | |
1727 | |
1728 infp->arg_offset = step - 4; | |
1729 infp->growth[growths++] = step; | |
1730 infp->reg_growth = growths; | |
1731 infp->local_growth = growths; | |
1732 | |
1733 /* If we haven't already folded it in. */ | |
1734 if (outbounds) | |
1735 infp->growth[growths++] = outbounds; | |
1736 | |
1737 goto finish; | |
1738 } | |
1739 | |
1740 /* Frame can't be done with a single subi, but can be done with 2 | |
1741 insns. If the 'stm' is getting <= 2 registers, we use stw's and | |
1742 shift some of the stack purchase into the first subi, so both are | |
1743 single instructions. */ | |
1744 if (localregarg <= STORE_REACH | |
1745 && (infp->local_size > ADDI_REACH) | |
1746 && (infp->reg_size <= 8 || (infp->reg_mask & 0xc000) != 0xc000)) | |
1747 { | |
1748 int all; | |
1749 | |
1750 /* Make sure we'll be aligned; use either pad_reg or pad_local. */ | |
1751 if (localregarg % STACK_BYTES) | |
1752 infp->pad_reg = STACK_BYTES - (localregarg % STACK_BYTES); | |
1753 | |
1754 all = localregarg + infp->pad_reg + infp->pad_local; | |
1755 step = ADDI_REACH; /* As much up front as we can. */ | |
1756 if (step > all) | |
1757 step = all; | |
1758 | |
1759 /* XXX: Consider whether step will still be aligned; we believe so. */ | |
1760 infp->arg_offset = step - 4; | |
1761 infp->growth[growths++] = step; | |
1762 infp->reg_growth = growths; | |
1763 infp->reg_offset = step - infp->pad_reg - infp->reg_size; | |
1764 all -= step; | |
1765 | |
1766 /* Can we fold in any space required for outbounds? */ | |
1767 if (outbounds + all <= ADDI_REACH && !frame_pointer_needed) | |
1768 { | |
1769 all += outbounds; | |
1770 outbounds = 0; | |
1771 } | |
1772 | |
1773 /* Get the rest of the locals in place. */ | |
1774 step = all; | |
1775 infp->growth[growths++] = step; | |
1776 infp->local_growth = growths; | |
1777 all -= step; | |
1778 | |
1779 assert (all == 0); | |
1780 | |
1781 /* Finish off if we need to do so. */ | |
1782 if (outbounds) | |
1783 infp->growth[growths++] = outbounds; | |
1784 | |
1785 goto finish; | |
1786 } | |
1787 | |
1788 /* Registers + args is nicely aligned, so we'll buy that in one shot. | |
1789 Then we buy the rest of the frame in 1 or 2 steps depending on | |
1790 whether we need a frame pointer. */ | |
1791 if ((regarg % STACK_BYTES) == 0) | |
1792 { | |
1793 infp->growth[growths++] = regarg; | |
1794 infp->reg_growth = growths; | |
1795 infp->arg_offset = regarg - 4; | |
1796 infp->reg_offset = 0; | |
1797 | |
1798 if (infp->local_size % STACK_BYTES) | |
1799 infp->pad_local = STACK_BYTES - (infp->local_size % STACK_BYTES); | |
1800 | |
1801 step = infp->local_size + infp->pad_local; | |
1802 | |
1803 if (!frame_pointer_needed) | |
1804 { | |
1805 step += outbounds; | |
1806 outbounds = 0; | |
1807 } | |
1808 | |
1809 infp->growth[growths++] = step; | |
1810 infp->local_growth = growths; | |
1811 | |
1812 /* If there's any left to be done. */ | |
1813 if (outbounds) | |
1814 infp->growth[growths++] = outbounds; | |
1815 | |
1816 goto finish; | |
1817 } | |
1818 | |
1819 /* XXX: optimizations that we'll want to play with.... | |
1820 -- regarg is not aligned, but it's a small number of registers; | |
1821 use some of localsize so that regarg is aligned and then | |
1822 save the registers. */ | |
1823 | |
1824 /* Simple encoding; plods down the stack buying the pieces as it goes. | |
1825 -- does not optimize space consumption. | |
1826 -- does not attempt to optimize instruction counts. | |
1827 -- but it is safe for all alignments. */ | |
1828 if (regarg % STACK_BYTES != 0) | |
1829 infp->pad_reg = STACK_BYTES - (regarg % STACK_BYTES); | |
1830 | |
1831 infp->growth[growths++] = infp->arg_size + infp->reg_size + infp->pad_reg; | |
1832 infp->reg_growth = growths; | |
1833 infp->arg_offset = infp->growth[0] - 4; | |
1834 infp->reg_offset = 0; | |
1835 | |
1836 if (frame_pointer_needed) | |
1837 { | |
1838 if (infp->local_size % STACK_BYTES != 0) | |
1839 infp->pad_local = STACK_BYTES - (infp->local_size % STACK_BYTES); | |
1840 | |
1841 infp->growth[growths++] = infp->local_size + infp->pad_local; | |
1842 infp->local_growth = growths; | |
1843 | |
1844 infp->growth[growths++] = outbounds; | |
1845 } | |
1846 else | |
1847 { | |
1848 if ((infp->local_size + outbounds) % STACK_BYTES != 0) | |
1849 infp->pad_local = STACK_BYTES - ((infp->local_size + outbounds) % STACK_BYTES); | |
1850 | |
1851 infp->growth[growths++] = infp->local_size + infp->pad_local + outbounds; | |
1852 infp->local_growth = growths; | |
1853 } | |
1854 | |
1855 /* Anything else that we've forgotten?, plus a few consistency checks. */ | |
1856 finish: | |
1857 assert (infp->reg_offset >= 0); | |
1858 assert (growths <= MAX_STACK_GROWS); | |
1859 | |
1860 for (i = 0; i < growths; i++) | |
1861 gcc_assert (!(infp->growth[i] % STACK_BYTES)); | |
1862 } | |
1863 | |
1864 /* Define the offset between two registers, one to be eliminated, and | |
1865 the other its replacement, at the start of a routine. */ | |
1866 | |
1867 int | |
1868 mcore_initial_elimination_offset (int from, int to) | |
1869 { | |
1870 int above_frame; | |
1871 int below_frame; | |
1872 struct mcore_frame fi; | |
1873 | |
1874 layout_mcore_frame (& fi); | |
1875 | |
1876 /* fp to ap */ | |
1877 above_frame = fi.local_size + fi.pad_local + fi.reg_size + fi.pad_reg; | |
1878 /* sp to fp */ | |
1879 below_frame = fi.outbound_size + fi.pad_outbound; | |
1880 | |
1881 if (from == ARG_POINTER_REGNUM && to == FRAME_POINTER_REGNUM) | |
1882 return above_frame; | |
1883 | |
1884 if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
1885 return above_frame + below_frame; | |
1886 | |
1887 if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
1888 return below_frame; | |
1889 | |
1890 gcc_unreachable (); | |
1891 } | |
1892 | |
1893 /* Keep track of some information about varargs for the prolog. */ | |
1894 | |
1895 static void | |
1896 mcore_setup_incoming_varargs (CUMULATIVE_ARGS *args_so_far, | |
1897 enum machine_mode mode, tree type, | |
1898 int * ptr_pretend_size ATTRIBUTE_UNUSED, | |
1899 int second_time ATTRIBUTE_UNUSED) | |
1900 { | |
1901 current_function_anonymous_args = 1; | |
1902 | |
1903 /* We need to know how many argument registers are used before | |
1904 the varargs start, so that we can push the remaining argument | |
1905 registers during the prologue. */ | |
1906 number_of_regs_before_varargs = *args_so_far + mcore_num_arg_regs (mode, type); | |
1907 | |
1908 /* There is a bug somewhere in the arg handling code. | |
1909 Until I can find it this workaround always pushes the | |
1910 last named argument onto the stack. */ | |
1911 number_of_regs_before_varargs = *args_so_far; | |
1912 | |
1913 /* The last named argument may be split between argument registers | |
1914 and the stack. Allow for this here. */ | |
1915 if (number_of_regs_before_varargs > NPARM_REGS) | |
1916 number_of_regs_before_varargs = NPARM_REGS; | |
1917 } | |
1918 | |
1919 void | |
1920 mcore_expand_prolog (void) | |
1921 { | |
1922 struct mcore_frame fi; | |
1923 int space_allocated = 0; | |
1924 int growth = 0; | |
1925 | |
1926 /* Find out what we're doing. */ | |
1927 layout_mcore_frame (&fi); | |
1928 | |
1929 space_allocated = fi.arg_size + fi.reg_size + fi.local_size + | |
1930 fi.outbound_size + fi.pad_outbound + fi.pad_local + fi.pad_reg; | |
1931 | |
1932 if (TARGET_CG_DATA) | |
1933 { | |
1934 /* Emit a symbol for this routine's frame size. */ | |
1935 rtx x; | |
1936 | |
1937 x = DECL_RTL (current_function_decl); | |
1938 | |
1939 gcc_assert (GET_CODE (x) == MEM); | |
1940 | |
1941 x = XEXP (x, 0); | |
1942 | |
1943 gcc_assert (GET_CODE (x) == SYMBOL_REF); | |
1944 | |
1945 if (mcore_current_function_name) | |
1946 free (mcore_current_function_name); | |
1947 | |
1948 mcore_current_function_name = xstrdup (XSTR (x, 0)); | |
1949 | |
1950 ASM_OUTPUT_CG_NODE (asm_out_file, mcore_current_function_name, space_allocated); | |
1951 | |
1952 if (cfun->calls_alloca) | |
1953 ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, "alloca", 1); | |
1954 | |
1955 /* 970425: RBE: | |
1956 We're looking at how the 8byte alignment affects stack layout | |
1957 and where we had to pad things. This emits information we can | |
1958 extract which tells us about frame sizes and the like. */ | |
1959 fprintf (asm_out_file, | |
1960 "\t.equ\t__$frame$info$_%s_$_%d_%d_x%x_%d_%d_%d,0\n", | |
1961 mcore_current_function_name, | |
1962 fi.arg_size, fi.reg_size, fi.reg_mask, | |
1963 fi.local_size, fi.outbound_size, | |
1964 frame_pointer_needed); | |
1965 } | |
1966 | |
1967 if (mcore_naked_function_p ()) | |
1968 return; | |
1969 | |
1970 /* Handle stdarg+regsaves in one shot: can't be more than 64 bytes. */ | |
1971 output_stack_adjust (-1, fi.growth[growth++]); /* Grows it. */ | |
1972 | |
1973 /* If we have a parameter passed partially in regs and partially in memory, | |
1974 the registers will have been stored to memory already in function.c. So | |
1975 we only need to do something here for varargs functions. */ | |
1976 if (fi.arg_size != 0 && crtl->args.pretend_args_size == 0) | |
1977 { | |
1978 int offset; | |
1979 int rn = FIRST_PARM_REG + NPARM_REGS - 1; | |
1980 int remaining = fi.arg_size; | |
1981 | |
1982 for (offset = fi.arg_offset; remaining >= 4; offset -= 4, rn--, remaining -= 4) | |
1983 { | |
1984 emit_insn (gen_movsi | |
1985 (gen_rtx_MEM (SImode, | |
1986 plus_constant (stack_pointer_rtx, offset)), | |
1987 gen_rtx_REG (SImode, rn))); | |
1988 } | |
1989 } | |
1990 | |
1991 /* Do we need another stack adjustment before we do the register saves? */ | |
1992 if (growth < fi.reg_growth) | |
1993 output_stack_adjust (-1, fi.growth[growth++]); /* Grows it. */ | |
1994 | |
1995 if (fi.reg_size != 0) | |
1996 { | |
1997 int i; | |
1998 int offs = fi.reg_offset; | |
1999 | |
2000 for (i = 15; i >= 0; i--) | |
2001 { | |
2002 if (offs == 0 && i == 15 && ((fi.reg_mask & 0xc000) == 0xc000)) | |
2003 { | |
2004 int first_reg = 15; | |
2005 | |
2006 while (fi.reg_mask & (1 << first_reg)) | |
2007 first_reg--; | |
2008 first_reg++; | |
2009 | |
2010 emit_insn (gen_store_multiple (gen_rtx_MEM (SImode, stack_pointer_rtx), | |
2011 gen_rtx_REG (SImode, first_reg), | |
2012 GEN_INT (16 - first_reg))); | |
2013 | |
2014 i -= (15 - first_reg); | |
2015 offs += (16 - first_reg) * 4; | |
2016 } | |
2017 else if (fi.reg_mask & (1 << i)) | |
2018 { | |
2019 emit_insn (gen_movsi | |
2020 (gen_rtx_MEM (SImode, | |
2021 plus_constant (stack_pointer_rtx, offs)), | |
2022 gen_rtx_REG (SImode, i))); | |
2023 offs += 4; | |
2024 } | |
2025 } | |
2026 } | |
2027 | |
2028 /* Figure the locals + outbounds. */ | |
2029 if (frame_pointer_needed) | |
2030 { | |
2031 /* If we haven't already purchased to 'fp'. */ | |
2032 if (growth < fi.local_growth) | |
2033 output_stack_adjust (-1, fi.growth[growth++]); /* Grows it. */ | |
2034 | |
2035 emit_insn (gen_movsi (frame_pointer_rtx, stack_pointer_rtx)); | |
2036 | |
2037 /* ... and then go any remaining distance for outbounds, etc. */ | |
2038 if (fi.growth[growth]) | |
2039 output_stack_adjust (-1, fi.growth[growth++]); | |
2040 } | |
2041 else | |
2042 { | |
2043 if (growth < fi.local_growth) | |
2044 output_stack_adjust (-1, fi.growth[growth++]); /* Grows it. */ | |
2045 if (fi.growth[growth]) | |
2046 output_stack_adjust (-1, fi.growth[growth++]); | |
2047 } | |
2048 } | |
2049 | |
2050 void | |
2051 mcore_expand_epilog (void) | |
2052 { | |
2053 struct mcore_frame fi; | |
2054 int i; | |
2055 int offs; | |
2056 int growth = MAX_STACK_GROWS - 1 ; | |
2057 | |
2058 | |
2059 /* Find out what we're doing. */ | |
2060 layout_mcore_frame(&fi); | |
2061 | |
2062 if (mcore_naked_function_p ()) | |
2063 return; | |
2064 | |
2065 /* If we had a frame pointer, restore the sp from that. */ | |
2066 if (frame_pointer_needed) | |
2067 { | |
2068 emit_insn (gen_movsi (stack_pointer_rtx, frame_pointer_rtx)); | |
2069 growth = fi.local_growth - 1; | |
2070 } | |
2071 else | |
2072 { | |
2073 /* XXX: while loop should accumulate and do a single sell. */ | |
2074 while (growth >= fi.local_growth) | |
2075 { | |
2076 if (fi.growth[growth] != 0) | |
2077 output_stack_adjust (1, fi.growth[growth]); | |
2078 growth--; | |
2079 } | |
2080 } | |
2081 | |
2082 /* Make sure we've shrunk stack back to the point where the registers | |
2083 were laid down. This is typically 0/1 iterations. Then pull the | |
2084 register save information back off the stack. */ | |
2085 while (growth >= fi.reg_growth) | |
2086 output_stack_adjust ( 1, fi.growth[growth--]); | |
2087 | |
2088 offs = fi.reg_offset; | |
2089 | |
2090 for (i = 15; i >= 0; i--) | |
2091 { | |
2092 if (offs == 0 && i == 15 && ((fi.reg_mask & 0xc000) == 0xc000)) | |
2093 { | |
2094 int first_reg; | |
2095 | |
2096 /* Find the starting register. */ | |
2097 first_reg = 15; | |
2098 | |
2099 while (fi.reg_mask & (1 << first_reg)) | |
2100 first_reg--; | |
2101 | |
2102 first_reg++; | |
2103 | |
2104 emit_insn (gen_load_multiple (gen_rtx_REG (SImode, first_reg), | |
2105 gen_rtx_MEM (SImode, stack_pointer_rtx), | |
2106 GEN_INT (16 - first_reg))); | |
2107 | |
2108 i -= (15 - first_reg); | |
2109 offs += (16 - first_reg) * 4; | |
2110 } | |
2111 else if (fi.reg_mask & (1 << i)) | |
2112 { | |
2113 emit_insn (gen_movsi | |
2114 (gen_rtx_REG (SImode, i), | |
2115 gen_rtx_MEM (SImode, | |
2116 plus_constant (stack_pointer_rtx, offs)))); | |
2117 offs += 4; | |
2118 } | |
2119 } | |
2120 | |
2121 /* Give back anything else. */ | |
2122 /* XXX: Should accumulate total and then give it back. */ | |
2123 while (growth >= 0) | |
2124 output_stack_adjust ( 1, fi.growth[growth--]); | |
2125 } | |
2126 | |
2127 /* This code is borrowed from the SH port. */ | |
2128 | |
2129 /* The MCORE cannot load a large constant into a register, constants have to | |
2130 come from a pc relative load. The reference of a pc relative load | |
2131 instruction must be less than 1k in front of the instruction. This | |
2132 means that we often have to dump a constant inside a function, and | |
2133 generate code to branch around it. | |
2134 | |
2135 It is important to minimize this, since the branches will slow things | |
2136 down and make things bigger. | |
2137 | |
2138 Worst case code looks like: | |
2139 | |
2140 lrw L1,r0 | |
2141 br L2 | |
2142 align | |
2143 L1: .long value | |
2144 L2: | |
2145 .. | |
2146 | |
2147 lrw L3,r0 | |
2148 br L4 | |
2149 align | |
2150 L3: .long value | |
2151 L4: | |
2152 .. | |
2153 | |
2154 We fix this by performing a scan before scheduling, which notices which | |
2155 instructions need to have their operands fetched from the constant table | |
2156 and builds the table. | |
2157 | |
2158 The algorithm is: | |
2159 | |
2160 scan, find an instruction which needs a pcrel move. Look forward, find the | |
2161 last barrier which is within MAX_COUNT bytes of the requirement. | |
2162 If there isn't one, make one. Process all the instructions between | |
2163 the find and the barrier. | |
2164 | |
2165 In the above example, we can tell that L3 is within 1k of L1, so | |
2166 the first move can be shrunk from the 2 insn+constant sequence into | |
2167 just 1 insn, and the constant moved to L3 to make: | |
2168 | |
2169 lrw L1,r0 | |
2170 .. | |
2171 lrw L3,r0 | |
2172 bra L4 | |
2173 align | |
2174 L3:.long value | |
2175 L4:.long value | |
2176 | |
2177 Then the second move becomes the target for the shortening process. */ | |
2178 | |
2179 typedef struct | |
2180 { | |
2181 rtx value; /* Value in table. */ | |
2182 rtx label; /* Label of value. */ | |
2183 } pool_node; | |
2184 | |
2185 /* The maximum number of constants that can fit into one pool, since | |
2186 the pc relative range is 0...1020 bytes and constants are at least 4 | |
2187 bytes long. We subtract 4 from the range to allow for the case where | |
2188 we need to add a branch/align before the constant pool. */ | |
2189 | |
2190 #define MAX_COUNT 1016 | |
2191 #define MAX_POOL_SIZE (MAX_COUNT/4) | |
2192 static pool_node pool_vector[MAX_POOL_SIZE]; | |
2193 static int pool_size; | |
2194 | |
2195 /* Dump out any constants accumulated in the final pass. These | |
2196 will only be labels. */ | |
2197 | |
2198 const char * | |
2199 mcore_output_jump_label_table (void) | |
2200 { | |
2201 int i; | |
2202 | |
2203 if (pool_size) | |
2204 { | |
2205 fprintf (asm_out_file, "\t.align 2\n"); | |
2206 | |
2207 for (i = 0; i < pool_size; i++) | |
2208 { | |
2209 pool_node * p = pool_vector + i; | |
2210 | |
2211 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (p->label)); | |
2212 | |
2213 output_asm_insn (".long %0", &p->value); | |
2214 } | |
2215 | |
2216 pool_size = 0; | |
2217 } | |
2218 | |
2219 return ""; | |
2220 } | |
2221 | |
2222 /* Check whether insn is a candidate for a conditional. */ | |
2223 | |
2224 static cond_type | |
2225 is_cond_candidate (rtx insn) | |
2226 { | |
2227 /* The only things we conditionalize are those that can be directly | |
2228 changed into a conditional. Only bother with SImode items. If | |
2229 we wanted to be a little more aggressive, we could also do other | |
2230 modes such as DImode with reg-reg move or load 0. */ | |
2231 if (GET_CODE (insn) == INSN) | |
2232 { | |
2233 rtx pat = PATTERN (insn); | |
2234 rtx src, dst; | |
2235 | |
2236 if (GET_CODE (pat) != SET) | |
2237 return COND_NO; | |
2238 | |
2239 dst = XEXP (pat, 0); | |
2240 | |
2241 if ((GET_CODE (dst) != REG && | |
2242 GET_CODE (dst) != SUBREG) || | |
2243 GET_MODE (dst) != SImode) | |
2244 return COND_NO; | |
2245 | |
2246 src = XEXP (pat, 1); | |
2247 | |
2248 if ((GET_CODE (src) == REG || | |
2249 (GET_CODE (src) == SUBREG && | |
2250 GET_CODE (SUBREG_REG (src)) == REG)) && | |
2251 GET_MODE (src) == SImode) | |
2252 return COND_MOV_INSN; | |
2253 else if (GET_CODE (src) == CONST_INT && | |
2254 INTVAL (src) == 0) | |
2255 return COND_CLR_INSN; | |
2256 else if (GET_CODE (src) == PLUS && | |
2257 (GET_CODE (XEXP (src, 0)) == REG || | |
2258 (GET_CODE (XEXP (src, 0)) == SUBREG && | |
2259 GET_CODE (SUBREG_REG (XEXP (src, 0))) == REG)) && | |
2260 GET_MODE (XEXP (src, 0)) == SImode && | |
2261 GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2262 INTVAL (XEXP (src, 1)) == 1) | |
2263 return COND_INC_INSN; | |
2264 else if (((GET_CODE (src) == MINUS && | |
2265 GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2266 INTVAL( XEXP (src, 1)) == 1) || | |
2267 (GET_CODE (src) == PLUS && | |
2268 GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2269 INTVAL (XEXP (src, 1)) == -1)) && | |
2270 (GET_CODE (XEXP (src, 0)) == REG || | |
2271 (GET_CODE (XEXP (src, 0)) == SUBREG && | |
2272 GET_CODE (SUBREG_REG (XEXP (src, 0))) == REG)) && | |
2273 GET_MODE (XEXP (src, 0)) == SImode) | |
2274 return COND_DEC_INSN; | |
2275 | |
2276 /* Some insns that we don't bother with: | |
2277 (set (rx:DI) (ry:DI)) | |
2278 (set (rx:DI) (const_int 0)) | |
2279 */ | |
2280 | |
2281 } | |
2282 else if (GET_CODE (insn) == JUMP_INSN && | |
2283 GET_CODE (PATTERN (insn)) == SET && | |
2284 GET_CODE (XEXP (PATTERN (insn), 1)) == LABEL_REF) | |
2285 return COND_BRANCH_INSN; | |
2286 | |
2287 return COND_NO; | |
2288 } | |
2289 | |
2290 /* Emit a conditional version of insn and replace the old insn with the | |
2291 new one. Return the new insn if emitted. */ | |
2292 | |
2293 static rtx | |
2294 emit_new_cond_insn (rtx insn, int cond) | |
2295 { | |
2296 rtx c_insn = 0; | |
2297 rtx pat, dst, src; | |
2298 cond_type num; | |
2299 | |
2300 if ((num = is_cond_candidate (insn)) == COND_NO) | |
2301 return NULL; | |
2302 | |
2303 pat = PATTERN (insn); | |
2304 | |
2305 if (GET_CODE (insn) == INSN) | |
2306 { | |
2307 dst = SET_DEST (pat); | |
2308 src = SET_SRC (pat); | |
2309 } | |
2310 else | |
2311 { | |
2312 dst = JUMP_LABEL (insn); | |
2313 src = NULL_RTX; | |
2314 } | |
2315 | |
2316 switch (num) | |
2317 { | |
2318 case COND_MOV_INSN: | |
2319 case COND_CLR_INSN: | |
2320 if (cond) | |
2321 c_insn = gen_movt0 (dst, src, dst); | |
2322 else | |
2323 c_insn = gen_movt0 (dst, dst, src); | |
2324 break; | |
2325 | |
2326 case COND_INC_INSN: | |
2327 if (cond) | |
2328 c_insn = gen_incscc (dst, dst); | |
2329 else | |
2330 c_insn = gen_incscc_false (dst, dst); | |
2331 break; | |
2332 | |
2333 case COND_DEC_INSN: | |
2334 if (cond) | |
2335 c_insn = gen_decscc (dst, dst); | |
2336 else | |
2337 c_insn = gen_decscc_false (dst, dst); | |
2338 break; | |
2339 | |
2340 case COND_BRANCH_INSN: | |
2341 if (cond) | |
2342 c_insn = gen_branch_true (dst); | |
2343 else | |
2344 c_insn = gen_branch_false (dst); | |
2345 break; | |
2346 | |
2347 default: | |
2348 return NULL; | |
2349 } | |
2350 | |
2351 /* Only copy the notes if they exist. */ | |
2352 if (rtx_length [GET_CODE (c_insn)] >= 7 && rtx_length [GET_CODE (insn)] >= 7) | |
2353 { | |
2354 /* We really don't need to bother with the notes and links at this | |
2355 point, but go ahead and save the notes. This will help is_dead() | |
2356 when applying peepholes (links don't matter since they are not | |
2357 used any more beyond this point for the mcore). */ | |
2358 REG_NOTES (c_insn) = REG_NOTES (insn); | |
2359 } | |
2360 | |
2361 if (num == COND_BRANCH_INSN) | |
2362 { | |
2363 /* For jumps, we need to be a little bit careful and emit the new jump | |
2364 before the old one and to update the use count for the target label. | |
2365 This way, the barrier following the old (uncond) jump will get | |
2366 deleted, but the label won't. */ | |
2367 c_insn = emit_jump_insn_before (c_insn, insn); | |
2368 | |
2369 ++ LABEL_NUSES (dst); | |
2370 | |
2371 JUMP_LABEL (c_insn) = dst; | |
2372 } | |
2373 else | |
2374 c_insn = emit_insn_after (c_insn, insn); | |
2375 | |
2376 delete_insn (insn); | |
2377 | |
2378 return c_insn; | |
2379 } | |
2380 | |
2381 /* Attempt to change a basic block into a series of conditional insns. This | |
2382 works by taking the branch at the end of the 1st block and scanning for the | |
2383 end of the 2nd block. If all instructions in the 2nd block have cond. | |
2384 versions and the label at the start of block 3 is the same as the target | |
2385 from the branch at block 1, then conditionalize all insn in block 2 using | |
2386 the inverse condition of the branch at block 1. (Note I'm bending the | |
2387 definition of basic block here.) | |
2388 | |
2389 e.g., change: | |
2390 | |
2391 bt L2 <-- end of block 1 (delete) | |
2392 mov r7,r8 | |
2393 addu r7,1 | |
2394 br L3 <-- end of block 2 | |
2395 | |
2396 L2: ... <-- start of block 3 (NUSES==1) | |
2397 L3: ... | |
2398 | |
2399 to: | |
2400 | |
2401 movf r7,r8 | |
2402 incf r7 | |
2403 bf L3 | |
2404 | |
2405 L3: ... | |
2406 | |
2407 we can delete the L2 label if NUSES==1 and re-apply the optimization | |
2408 starting at the last instruction of block 2. This may allow an entire | |
2409 if-then-else statement to be conditionalized. BRC */ | |
2410 static rtx | |
2411 conditionalize_block (rtx first) | |
2412 { | |
2413 rtx insn; | |
2414 rtx br_pat; | |
2415 rtx end_blk_1_br = 0; | |
2416 rtx end_blk_2_insn = 0; | |
2417 rtx start_blk_3_lab = 0; | |
2418 int cond; | |
2419 int br_lab_num; | |
2420 int blk_size = 0; | |
2421 | |
2422 | |
2423 /* Check that the first insn is a candidate conditional jump. This is | |
2424 the one that we'll eliminate. If not, advance to the next insn to | |
2425 try. */ | |
2426 if (GET_CODE (first) != JUMP_INSN || | |
2427 GET_CODE (PATTERN (first)) != SET || | |
2428 GET_CODE (XEXP (PATTERN (first), 1)) != IF_THEN_ELSE) | |
2429 return NEXT_INSN (first); | |
2430 | |
2431 /* Extract some information we need. */ | |
2432 end_blk_1_br = first; | |
2433 br_pat = PATTERN (end_blk_1_br); | |
2434 | |
2435 /* Complement the condition since we use the reverse cond. for the insns. */ | |
2436 cond = (GET_CODE (XEXP (XEXP (br_pat, 1), 0)) == EQ); | |
2437 | |
2438 /* Determine what kind of branch we have. */ | |
2439 if (GET_CODE (XEXP (XEXP (br_pat, 1), 1)) == LABEL_REF) | |
2440 { | |
2441 /* A normal branch, so extract label out of first arm. */ | |
2442 br_lab_num = CODE_LABEL_NUMBER (XEXP (XEXP (XEXP (br_pat, 1), 1), 0)); | |
2443 } | |
2444 else | |
2445 { | |
2446 /* An inverse branch, so extract the label out of the 2nd arm | |
2447 and complement the condition. */ | |
2448 cond = (cond == 0); | |
2449 br_lab_num = CODE_LABEL_NUMBER (XEXP (XEXP (XEXP (br_pat, 1), 2), 0)); | |
2450 } | |
2451 | |
2452 /* Scan forward for the start of block 2: it must start with a | |
2453 label and that label must be the same as the branch target | |
2454 label from block 1. We don't care about whether block 2 actually | |
2455 ends with a branch or a label (an uncond. branch is | |
2456 conditionalizable). */ | |
2457 for (insn = NEXT_INSN (first); insn; insn = NEXT_INSN (insn)) | |
2458 { | |
2459 enum rtx_code code; | |
2460 | |
2461 code = GET_CODE (insn); | |
2462 | |
2463 /* Look for the label at the start of block 3. */ | |
2464 if (code == CODE_LABEL && CODE_LABEL_NUMBER (insn) == br_lab_num) | |
2465 break; | |
2466 | |
2467 /* Skip barriers, notes, and conditionalizable insns. If the | |
2468 insn is not conditionalizable or makes this optimization fail, | |
2469 just return the next insn so we can start over from that point. */ | |
2470 if (code != BARRIER && code != NOTE && !is_cond_candidate (insn)) | |
2471 return NEXT_INSN (insn); | |
2472 | |
2473 /* Remember the last real insn before the label (i.e. end of block 2). */ | |
2474 if (code == JUMP_INSN || code == INSN) | |
2475 { | |
2476 blk_size ++; | |
2477 end_blk_2_insn = insn; | |
2478 } | |
2479 } | |
2480 | |
2481 if (!insn) | |
2482 return insn; | |
2483 | |
2484 /* It is possible for this optimization to slow performance if the blocks | |
2485 are long. This really depends upon whether the branch is likely taken | |
2486 or not. If the branch is taken, we slow performance in many cases. But, | |
2487 if the branch is not taken, we always help performance (for a single | |
2488 block, but for a double block (i.e. when the optimization is re-applied) | |
2489 this is not true since the 'right thing' depends on the overall length of | |
2490 the collapsed block). As a compromise, don't apply this optimization on | |
2491 blocks larger than size 2 (unlikely for the mcore) when speed is important. | |
2492 the best threshold depends on the latencies of the instructions (i.e., | |
2493 the branch penalty). */ | |
2494 if (optimize > 1 && blk_size > 2) | |
2495 return insn; | |
2496 | |
2497 /* At this point, we've found the start of block 3 and we know that | |
2498 it is the destination of the branch from block 1. Also, all | |
2499 instructions in the block 2 are conditionalizable. So, apply the | |
2500 conditionalization and delete the branch. */ | |
2501 start_blk_3_lab = insn; | |
2502 | |
2503 for (insn = NEXT_INSN (end_blk_1_br); insn != start_blk_3_lab; | |
2504 insn = NEXT_INSN (insn)) | |
2505 { | |
2506 rtx newinsn; | |
2507 | |
2508 if (INSN_DELETED_P (insn)) | |
2509 continue; | |
2510 | |
2511 /* Try to form a conditional variant of the instruction and emit it. */ | |
2512 if ((newinsn = emit_new_cond_insn (insn, cond))) | |
2513 { | |
2514 if (end_blk_2_insn == insn) | |
2515 end_blk_2_insn = newinsn; | |
2516 | |
2517 insn = newinsn; | |
2518 } | |
2519 } | |
2520 | |
2521 /* Note whether we will delete the label starting blk 3 when the jump | |
2522 gets deleted. If so, we want to re-apply this optimization at the | |
2523 last real instruction right before the label. */ | |
2524 if (LABEL_NUSES (start_blk_3_lab) == 1) | |
2525 { | |
2526 start_blk_3_lab = 0; | |
2527 } | |
2528 | |
2529 /* ??? we probably should redistribute the death notes for this insn, esp. | |
2530 the death of cc, but it doesn't really matter this late in the game. | |
2531 The peepholes all use is_dead() which will find the correct death | |
2532 regardless of whether there is a note. */ | |
2533 delete_insn (end_blk_1_br); | |
2534 | |
2535 if (! start_blk_3_lab) | |
2536 return end_blk_2_insn; | |
2537 | |
2538 /* Return the insn right after the label at the start of block 3. */ | |
2539 return NEXT_INSN (start_blk_3_lab); | |
2540 } | |
2541 | |
2542 /* Apply the conditionalization of blocks optimization. This is the | |
2543 outer loop that traverses through the insns scanning for a branch | |
2544 that signifies an opportunity to apply the optimization. Note that | |
2545 this optimization is applied late. If we could apply it earlier, | |
2546 say before cse 2, it may expose more optimization opportunities. | |
2547 but, the pay back probably isn't really worth the effort (we'd have | |
2548 to update all reg/flow/notes/links/etc to make it work - and stick it | |
2549 in before cse 2). */ | |
2550 | |
2551 static void | |
2552 conditionalize_optimization (void) | |
2553 { | |
2554 rtx insn; | |
2555 | |
2556 for (insn = get_insns (); insn; insn = conditionalize_block (insn)) | |
2557 continue; | |
2558 } | |
2559 | |
2560 static int saved_warn_return_type = -1; | |
2561 static int saved_warn_return_type_count = 0; | |
2562 | |
2563 /* This is to handle loads from the constant pool. */ | |
2564 | |
2565 static void | |
2566 mcore_reorg (void) | |
2567 { | |
2568 /* Reset this variable. */ | |
2569 current_function_anonymous_args = 0; | |
2570 | |
2571 /* Restore the warn_return_type if it has been altered. */ | |
2572 if (saved_warn_return_type != -1) | |
2573 { | |
2574 /* Only restore the value if we have reached another function. | |
2575 The test of warn_return_type occurs in final_function () in | |
2576 c-decl.c a long time after the code for the function is generated, | |
2577 so we need a counter to tell us when we have finished parsing that | |
2578 function and can restore the flag. */ | |
2579 if (--saved_warn_return_type_count == 0) | |
2580 { | |
2581 warn_return_type = saved_warn_return_type; | |
2582 saved_warn_return_type = -1; | |
2583 } | |
2584 } | |
2585 | |
2586 if (optimize == 0) | |
2587 return; | |
2588 | |
2589 /* Conditionalize blocks where we can. */ | |
2590 conditionalize_optimization (); | |
2591 | |
2592 /* Literal pool generation is now pushed off until the assembler. */ | |
2593 } | |
2594 | |
2595 | |
2596 /* Return true if X is something that can be moved directly into r15. */ | |
2597 | |
2598 bool | |
2599 mcore_r15_operand_p (rtx x) | |
2600 { | |
2601 switch (GET_CODE (x)) | |
2602 { | |
2603 case CONST_INT: | |
2604 return mcore_const_ok_for_inline (INTVAL (x)); | |
2605 | |
2606 case REG: | |
2607 case SUBREG: | |
2608 case MEM: | |
2609 return 1; | |
2610 | |
2611 default: | |
2612 return 0; | |
2613 } | |
2614 } | |
2615 | |
2616 /* Implement SECONDARY_RELOAD_CLASS. If RCLASS contains r15, and we can't | |
2617 directly move X into it, use r1-r14 as a temporary. */ | |
2618 | |
2619 enum reg_class | |
2620 mcore_secondary_reload_class (enum reg_class rclass, | |
2621 enum machine_mode mode ATTRIBUTE_UNUSED, rtx x) | |
2622 { | |
2623 if (TEST_HARD_REG_BIT (reg_class_contents[rclass], 15) | |
2624 && !mcore_r15_operand_p (x)) | |
2625 return LRW_REGS; | |
2626 return NO_REGS; | |
2627 } | |
2628 | |
2629 /* Return the reg_class to use when reloading the rtx X into the class | |
2630 RCLASS. If X is too complex to move directly into r15, prefer to | |
2631 use LRW_REGS instead. */ | |
2632 | |
2633 enum reg_class | |
2634 mcore_reload_class (rtx x, enum reg_class rclass) | |
2635 { | |
2636 if (reg_class_subset_p (LRW_REGS, rclass) && !mcore_r15_operand_p (x)) | |
2637 return LRW_REGS; | |
2638 | |
2639 return rclass; | |
2640 } | |
2641 | |
2642 /* Tell me if a pair of reg/subreg rtx's actually refer to the same | |
2643 register. Note that the current version doesn't worry about whether | |
2644 they are the same mode or note (e.g., a QImode in r2 matches an HImode | |
2645 in r2 matches an SImode in r2. Might think in the future about whether | |
2646 we want to be able to say something about modes. */ | |
2647 | |
2648 int | |
2649 mcore_is_same_reg (rtx x, rtx y) | |
2650 { | |
2651 /* Strip any and all of the subreg wrappers. */ | |
2652 while (GET_CODE (x) == SUBREG) | |
2653 x = SUBREG_REG (x); | |
2654 | |
2655 while (GET_CODE (y) == SUBREG) | |
2656 y = SUBREG_REG (y); | |
2657 | |
2658 if (GET_CODE(x) == REG && GET_CODE(y) == REG && REGNO(x) == REGNO(y)) | |
2659 return 1; | |
2660 | |
2661 return 0; | |
2662 } | |
2663 | |
2664 void | |
2665 mcore_override_options (void) | |
2666 { | |
2667 /* Only the m340 supports little endian code. */ | |
2668 if (TARGET_LITTLE_END && ! TARGET_M340) | |
2669 target_flags |= MASK_M340; | |
2670 } | |
2671 | |
2672 /* Compute the number of word sized registers needed to | |
2673 hold a function argument of mode MODE and type TYPE. */ | |
2674 | |
2675 int | |
2676 mcore_num_arg_regs (enum machine_mode mode, const_tree type) | |
2677 { | |
2678 int size; | |
2679 | |
2680 if (targetm.calls.must_pass_in_stack (mode, type)) | |
2681 return 0; | |
2682 | |
2683 if (type && mode == BLKmode) | |
2684 size = int_size_in_bytes (type); | |
2685 else | |
2686 size = GET_MODE_SIZE (mode); | |
2687 | |
2688 return ROUND_ADVANCE (size); | |
2689 } | |
2690 | |
2691 static rtx | |
2692 handle_structs_in_regs (enum machine_mode mode, const_tree type, int reg) | |
2693 { | |
2694 int size; | |
2695 | |
2696 /* The MCore ABI defines that a structure whose size is not a whole multiple | |
2697 of bytes is passed packed into registers (or spilled onto the stack if | |
2698 not enough registers are available) with the last few bytes of the | |
2699 structure being packed, left-justified, into the last register/stack slot. | |
2700 GCC handles this correctly if the last word is in a stack slot, but we | |
2701 have to generate a special, PARALLEL RTX if the last word is in an | |
2702 argument register. */ | |
2703 if (type | |
2704 && TYPE_MODE (type) == BLKmode | |
2705 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST | |
2706 && (size = int_size_in_bytes (type)) > UNITS_PER_WORD | |
2707 && (size % UNITS_PER_WORD != 0) | |
2708 && (reg + mcore_num_arg_regs (mode, type) <= (FIRST_PARM_REG + NPARM_REGS))) | |
2709 { | |
2710 rtx arg_regs [NPARM_REGS]; | |
2711 int nregs; | |
2712 rtx result; | |
2713 rtvec rtvec; | |
2714 | |
2715 for (nregs = 0; size > 0; size -= UNITS_PER_WORD) | |
2716 { | |
2717 arg_regs [nregs] = | |
2718 gen_rtx_EXPR_LIST (SImode, gen_rtx_REG (SImode, reg ++), | |
2719 GEN_INT (nregs * UNITS_PER_WORD)); | |
2720 nregs ++; | |
2721 } | |
2722 | |
2723 /* We assume here that NPARM_REGS == 6. The assert checks this. */ | |
2724 assert (ARRAY_SIZE (arg_regs) == 6); | |
2725 rtvec = gen_rtvec (nregs, arg_regs[0], arg_regs[1], arg_regs[2], | |
2726 arg_regs[3], arg_regs[4], arg_regs[5]); | |
2727 | |
2728 result = gen_rtx_PARALLEL (mode, rtvec); | |
2729 return result; | |
2730 } | |
2731 | |
2732 return gen_rtx_REG (mode, reg); | |
2733 } | |
2734 | |
2735 rtx | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2736 mcore_function_value (const_tree valtype, const_tree func) |
0 | 2737 { |
2738 enum machine_mode mode; | |
2739 int unsigned_p; | |
2740 | |
2741 mode = TYPE_MODE (valtype); | |
2742 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2743 /* Since we promote return types, we must promote the mode here too. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2744 mode = promote_function_mode (valtype, mode, &unsigned_p, func, 1); |
0 | 2745 |
2746 return handle_structs_in_regs (mode, valtype, FIRST_RET_REG); | |
2747 } | |
2748 | |
2749 /* Define where to put the arguments to a function. | |
2750 Value is zero to push the argument on the stack, | |
2751 or a hard register in which to store the argument. | |
2752 | |
2753 MODE is the argument's machine mode. | |
2754 TYPE is the data type of the argument (as a tree). | |
2755 This is null for libcalls where that information may | |
2756 not be available. | |
2757 CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
2758 the preceding args and about the function being called. | |
2759 NAMED is nonzero if this argument is a named parameter | |
2760 (otherwise it is an extra parameter matching an ellipsis). | |
2761 | |
2762 On MCore the first args are normally in registers | |
2763 and the rest are pushed. Any arg that starts within the first | |
2764 NPARM_REGS words is at least partially passed in a register unless | |
2765 its data type forbids. */ | |
2766 | |
2767 rtx | |
2768 mcore_function_arg (CUMULATIVE_ARGS cum, enum machine_mode mode, | |
2769 tree type, int named) | |
2770 { | |
2771 int arg_reg; | |
2772 | |
2773 if (! named || mode == VOIDmode) | |
2774 return 0; | |
2775 | |
2776 if (targetm.calls.must_pass_in_stack (mode, type)) | |
2777 return 0; | |
2778 | |
2779 arg_reg = ROUND_REG (cum, mode); | |
2780 | |
2781 if (arg_reg < NPARM_REGS) | |
2782 return handle_structs_in_regs (mode, type, FIRST_PARM_REG + arg_reg); | |
2783 | |
2784 return 0; | |
2785 } | |
2786 | |
2787 /* Returns the number of bytes of argument registers required to hold *part* | |
2788 of a parameter of machine mode MODE and type TYPE (which may be NULL if | |
2789 the type is not known). If the argument fits entirely in the argument | |
2790 registers, or entirely on the stack, then 0 is returned. CUM is the | |
2791 number of argument registers already used by earlier parameters to | |
2792 the function. */ | |
2793 | |
2794 static int | |
2795 mcore_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
2796 tree type, bool named) | |
2797 { | |
2798 int reg = ROUND_REG (*cum, mode); | |
2799 | |
2800 if (named == 0) | |
2801 return 0; | |
2802 | |
2803 if (targetm.calls.must_pass_in_stack (mode, type)) | |
2804 return 0; | |
2805 | |
2806 /* REG is not the *hardware* register number of the register that holds | |
2807 the argument, it is the *argument* register number. So for example, | |
2808 the first argument to a function goes in argument register 0, which | |
2809 translates (for the MCore) into hardware register 2. The second | |
2810 argument goes into argument register 1, which translates into hardware | |
2811 register 3, and so on. NPARM_REGS is the number of argument registers | |
2812 supported by the target, not the maximum hardware register number of | |
2813 the target. */ | |
2814 if (reg >= NPARM_REGS) | |
2815 return 0; | |
2816 | |
2817 /* If the argument fits entirely in registers, return 0. */ | |
2818 if (reg + mcore_num_arg_regs (mode, type) <= NPARM_REGS) | |
2819 return 0; | |
2820 | |
2821 /* The argument overflows the number of available argument registers. | |
2822 Compute how many argument registers have not yet been assigned to | |
2823 hold an argument. */ | |
2824 reg = NPARM_REGS - reg; | |
2825 | |
2826 /* Return partially in registers and partially on the stack. */ | |
2827 return reg * UNITS_PER_WORD; | |
2828 } | |
2829 | |
2830 /* Return nonzero if SYMBOL is marked as being dllexport'd. */ | |
2831 | |
2832 int | |
2833 mcore_dllexport_name_p (const char * symbol) | |
2834 { | |
2835 return symbol[0] == '@' && symbol[1] == 'e' && symbol[2] == '.'; | |
2836 } | |
2837 | |
2838 /* Return nonzero if SYMBOL is marked as being dllimport'd. */ | |
2839 | |
2840 int | |
2841 mcore_dllimport_name_p (const char * symbol) | |
2842 { | |
2843 return symbol[0] == '@' && symbol[1] == 'i' && symbol[2] == '.'; | |
2844 } | |
2845 | |
2846 /* Mark a DECL as being dllexport'd. */ | |
2847 | |
2848 static void | |
2849 mcore_mark_dllexport (tree decl) | |
2850 { | |
2851 const char * oldname; | |
2852 char * newname; | |
2853 rtx rtlname; | |
2854 tree idp; | |
2855 | |
2856 rtlname = XEXP (DECL_RTL (decl), 0); | |
2857 | |
2858 if (GET_CODE (rtlname) == MEM) | |
2859 rtlname = XEXP (rtlname, 0); | |
2860 gcc_assert (GET_CODE (rtlname) == SYMBOL_REF); | |
2861 oldname = XSTR (rtlname, 0); | |
2862 | |
2863 if (mcore_dllexport_name_p (oldname)) | |
2864 return; /* Already done. */ | |
2865 | |
2866 newname = XALLOCAVEC (char, strlen (oldname) + 4); | |
2867 sprintf (newname, "@e.%s", oldname); | |
2868 | |
2869 /* We pass newname through get_identifier to ensure it has a unique | |
2870 address. RTL processing can sometimes peek inside the symbol ref | |
2871 and compare the string's addresses to see if two symbols are | |
2872 identical. */ | |
2873 /* ??? At least I think that's why we do this. */ | |
2874 idp = get_identifier (newname); | |
2875 | |
2876 XEXP (DECL_RTL (decl), 0) = | |
2877 gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (idp)); | |
2878 } | |
2879 | |
2880 /* Mark a DECL as being dllimport'd. */ | |
2881 | |
2882 static void | |
2883 mcore_mark_dllimport (tree decl) | |
2884 { | |
2885 const char * oldname; | |
2886 char * newname; | |
2887 tree idp; | |
2888 rtx rtlname; | |
2889 rtx newrtl; | |
2890 | |
2891 rtlname = XEXP (DECL_RTL (decl), 0); | |
2892 | |
2893 if (GET_CODE (rtlname) == MEM) | |
2894 rtlname = XEXP (rtlname, 0); | |
2895 gcc_assert (GET_CODE (rtlname) == SYMBOL_REF); | |
2896 oldname = XSTR (rtlname, 0); | |
2897 | |
2898 gcc_assert (!mcore_dllexport_name_p (oldname)); | |
2899 if (mcore_dllimport_name_p (oldname)) | |
2900 return; /* Already done. */ | |
2901 | |
2902 /* ??? One can well ask why we're making these checks here, | |
2903 and that would be a good question. */ | |
2904 | |
2905 /* Imported variables can't be initialized. */ | |
2906 if (TREE_CODE (decl) == VAR_DECL | |
2907 && !DECL_VIRTUAL_P (decl) | |
2908 && DECL_INITIAL (decl)) | |
2909 { | |
2910 error ("initialized variable %q+D is marked dllimport", decl); | |
2911 return; | |
2912 } | |
2913 | |
2914 /* `extern' needn't be specified with dllimport. | |
2915 Specify `extern' now and hope for the best. Sigh. */ | |
2916 if (TREE_CODE (decl) == VAR_DECL | |
2917 /* ??? Is this test for vtables needed? */ | |
2918 && !DECL_VIRTUAL_P (decl)) | |
2919 { | |
2920 DECL_EXTERNAL (decl) = 1; | |
2921 TREE_PUBLIC (decl) = 1; | |
2922 } | |
2923 | |
2924 newname = XALLOCAVEC (char, strlen (oldname) + 11); | |
2925 sprintf (newname, "@i.__imp_%s", oldname); | |
2926 | |
2927 /* We pass newname through get_identifier to ensure it has a unique | |
2928 address. RTL processing can sometimes peek inside the symbol ref | |
2929 and compare the string's addresses to see if two symbols are | |
2930 identical. */ | |
2931 /* ??? At least I think that's why we do this. */ | |
2932 idp = get_identifier (newname); | |
2933 | |
2934 newrtl = gen_rtx_MEM (Pmode, | |
2935 gen_rtx_SYMBOL_REF (Pmode, | |
2936 IDENTIFIER_POINTER (idp))); | |
2937 XEXP (DECL_RTL (decl), 0) = newrtl; | |
2938 } | |
2939 | |
2940 static int | |
2941 mcore_dllexport_p (tree decl) | |
2942 { | |
2943 if ( TREE_CODE (decl) != VAR_DECL | |
2944 && TREE_CODE (decl) != FUNCTION_DECL) | |
2945 return 0; | |
2946 | |
2947 return lookup_attribute ("dllexport", DECL_ATTRIBUTES (decl)) != 0; | |
2948 } | |
2949 | |
2950 static int | |
2951 mcore_dllimport_p (tree decl) | |
2952 { | |
2953 if ( TREE_CODE (decl) != VAR_DECL | |
2954 && TREE_CODE (decl) != FUNCTION_DECL) | |
2955 return 0; | |
2956 | |
2957 return lookup_attribute ("dllimport", DECL_ATTRIBUTES (decl)) != 0; | |
2958 } | |
2959 | |
2960 /* We must mark dll symbols specially. Definitions of dllexport'd objects | |
2961 install some info in the .drective (PE) or .exports (ELF) sections. */ | |
2962 | |
2963 static void | |
2964 mcore_encode_section_info (tree decl, rtx rtl ATTRIBUTE_UNUSED, int first ATTRIBUTE_UNUSED) | |
2965 { | |
2966 /* Mark the decl so we can tell from the rtl whether the object is | |
2967 dllexport'd or dllimport'd. */ | |
2968 if (mcore_dllexport_p (decl)) | |
2969 mcore_mark_dllexport (decl); | |
2970 else if (mcore_dllimport_p (decl)) | |
2971 mcore_mark_dllimport (decl); | |
2972 | |
2973 /* It might be that DECL has already been marked as dllimport, but | |
2974 a subsequent definition nullified that. The attribute is gone | |
2975 but DECL_RTL still has @i.__imp_foo. We need to remove that. */ | |
2976 else if ((TREE_CODE (decl) == FUNCTION_DECL | |
2977 || TREE_CODE (decl) == VAR_DECL) | |
2978 && DECL_RTL (decl) != NULL_RTX | |
2979 && GET_CODE (DECL_RTL (decl)) == MEM | |
2980 && GET_CODE (XEXP (DECL_RTL (decl), 0)) == MEM | |
2981 && GET_CODE (XEXP (XEXP (DECL_RTL (decl), 0), 0)) == SYMBOL_REF | |
2982 && mcore_dllimport_name_p (XSTR (XEXP (XEXP (DECL_RTL (decl), 0), 0), 0))) | |
2983 { | |
2984 const char * oldname = XSTR (XEXP (XEXP (DECL_RTL (decl), 0), 0), 0); | |
2985 tree idp = get_identifier (oldname + 9); | |
2986 rtx newrtl = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (idp)); | |
2987 | |
2988 XEXP (DECL_RTL (decl), 0) = newrtl; | |
2989 | |
2990 /* We previously set TREE_PUBLIC and DECL_EXTERNAL. | |
2991 ??? We leave these alone for now. */ | |
2992 } | |
2993 } | |
2994 | |
2995 /* Undo the effects of the above. */ | |
2996 | |
2997 static const char * | |
2998 mcore_strip_name_encoding (const char * str) | |
2999 { | |
3000 return str + (str[0] == '@' ? 3 : 0); | |
3001 } | |
3002 | |
3003 /* MCore specific attribute support. | |
3004 dllexport - for exporting a function/variable that will live in a dll | |
3005 dllimport - for importing a function/variable from a dll | |
3006 naked - do not create a function prologue/epilogue. */ | |
3007 | |
3008 /* Handle a "naked" attribute; arguments as in | |
3009 struct attribute_spec.handler. */ | |
3010 | |
3011 static tree | |
3012 mcore_handle_naked_attribute (tree * node, tree name, tree args ATTRIBUTE_UNUSED, | |
3013 int flags ATTRIBUTE_UNUSED, bool * no_add_attrs) | |
3014 { | |
3015 if (TREE_CODE (*node) == FUNCTION_DECL) | |
3016 { | |
3017 /* PR14310 - don't complain about lack of return statement | |
3018 in naked functions. The solution here is a gross hack | |
3019 but this is the only way to solve the problem without | |
3020 adding a new feature to GCC. I did try submitting a patch | |
3021 that would add such a new feature, but it was (rightfully) | |
3022 rejected on the grounds that it was creeping featurism, | |
3023 so hence this code. */ | |
3024 if (warn_return_type) | |
3025 { | |
3026 saved_warn_return_type = warn_return_type; | |
3027 warn_return_type = 0; | |
3028 saved_warn_return_type_count = 2; | |
3029 } | |
3030 else if (saved_warn_return_type_count) | |
3031 saved_warn_return_type_count = 2; | |
3032 } | |
3033 else | |
3034 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3035 warning (OPT_Wattributes, "%qE attribute only applies to functions", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3036 name); |
0 | 3037 *no_add_attrs = true; |
3038 } | |
3039 | |
3040 return NULL_TREE; | |
3041 } | |
3042 | |
3043 /* ??? It looks like this is PE specific? Oh well, this is what the | |
3044 old code did as well. */ | |
3045 | |
3046 static void | |
3047 mcore_unique_section (tree decl, int reloc ATTRIBUTE_UNUSED) | |
3048 { | |
3049 int len; | |
3050 const char * name; | |
3051 char * string; | |
3052 const char * prefix; | |
3053 | |
3054 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)); | |
3055 | |
3056 /* Strip off any encoding in name. */ | |
3057 name = (* targetm.strip_name_encoding) (name); | |
3058 | |
3059 /* The object is put in, for example, section .text$foo. | |
3060 The linker will then ultimately place them in .text | |
3061 (everything from the $ on is stripped). */ | |
3062 if (TREE_CODE (decl) == FUNCTION_DECL) | |
3063 prefix = ".text$"; | |
3064 /* For compatibility with EPOC, we ignore the fact that the | |
3065 section might have relocs against it. */ | |
3066 else if (decl_readonly_section (decl, 0)) | |
3067 prefix = ".rdata$"; | |
3068 else | |
3069 prefix = ".data$"; | |
3070 | |
3071 len = strlen (name) + strlen (prefix); | |
3072 string = XALLOCAVEC (char, len + 1); | |
3073 | |
3074 sprintf (string, "%s%s", prefix, name); | |
3075 | |
3076 DECL_SECTION_NAME (decl) = build_string (len, string); | |
3077 } | |
3078 | |
3079 int | |
3080 mcore_naked_function_p (void) | |
3081 { | |
3082 return lookup_attribute ("naked", DECL_ATTRIBUTES (current_function_decl)) != NULL_TREE; | |
3083 } | |
3084 | |
3085 #ifdef OBJECT_FORMAT_ELF | |
3086 static void | |
3087 mcore_asm_named_section (const char *name, | |
3088 unsigned int flags ATTRIBUTE_UNUSED, | |
3089 tree decl ATTRIBUTE_UNUSED) | |
3090 { | |
3091 fprintf (asm_out_file, "\t.section %s\n", name); | |
3092 } | |
3093 #endif /* OBJECT_FORMAT_ELF */ | |
3094 | |
3095 /* Worker function for TARGET_ASM_EXTERNAL_LIBCALL. */ | |
3096 | |
3097 static void | |
3098 mcore_external_libcall (rtx fun) | |
3099 { | |
3100 fprintf (asm_out_file, "\t.import\t"); | |
3101 assemble_name (asm_out_file, XSTR (fun, 0)); | |
3102 fprintf (asm_out_file, "\n"); | |
3103 } | |
3104 | |
3105 /* Worker function for TARGET_RETURN_IN_MEMORY. */ | |
3106 | |
3107 static bool | |
3108 mcore_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) | |
3109 { | |
3110 const HOST_WIDE_INT size = int_size_in_bytes (type); | |
3111 return (size == -1 || size > 2 * UNITS_PER_WORD); | |
3112 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3113 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3114 /* Worker function for TARGET_ASM_TRAMPOLINE_TEMPLATE. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3115 Output assembler code for a block containing the constant parts |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3116 of a trampoline, leaving space for the variable parts. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3117 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3118 On the MCore, the trampoline looks like: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3119 lrw r1, function |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3120 lrw r13, area |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3121 jmp r13 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3122 or r0, r0 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3123 .literals */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3124 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3125 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3126 mcore_asm_trampoline_template (FILE *f) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3127 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3128 fprintf (f, "\t.short 0x7102\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3129 fprintf (f, "\t.short 0x7d02\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3130 fprintf (f, "\t.short 0x00cd\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3131 fprintf (f, "\t.short 0x1e00\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3132 fprintf (f, "\t.long 0\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3133 fprintf (f, "\t.long 0\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3134 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3135 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3136 /* Worker function for TARGET_TRAMPOLINE_INIT. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3137 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3138 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3139 mcore_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3140 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3141 rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3142 rtx mem; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3143 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3144 emit_block_move (m_tramp, assemble_trampoline_template (), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3145 GEN_INT (2*UNITS_PER_WORD), BLOCK_OP_NORMAL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3146 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3147 mem = adjust_address (m_tramp, SImode, 8); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3148 emit_move_insn (mem, chain_value); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3149 mem = adjust_address (m_tramp, SImode, 12); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3150 emit_move_insn (mem, fnaddr); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3151 } |