1 | /* Integrated Register Allocator (IRA) entry point. |
2 | Copyright (C) 2006-2023 Free Software Foundation, Inc. |
3 | Contributed by Vladimir Makarov <vmakarov@redhat.com>. |
4 | |
5 | This file is part of GCC. |
6 | |
7 | GCC is free software; you can redistribute it and/or modify it under |
8 | the terms of the GNU General Public License as published by the Free |
9 | Software Foundation; either version 3, or (at your option) any later |
10 | version. |
11 | |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
15 | 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 | /* The integrated register allocator (IRA) is a |
22 | regional register allocator performing graph coloring on a top-down |
23 | traversal of nested regions. Graph coloring in a region is based |
24 | on Chaitin-Briggs algorithm. It is called integrated because |
25 | register coalescing, register live range splitting, and choosing a |
26 | better hard register are done on-the-fly during coloring. Register |
27 | coalescing and choosing a cheaper hard register is done by hard |
28 | register preferencing during hard register assigning. The live |
29 | range splitting is a byproduct of the regional register allocation. |
30 | |
31 | Major IRA notions are: |
32 | |
33 | o *Region* is a part of CFG where graph coloring based on |
34 | Chaitin-Briggs algorithm is done. IRA can work on any set of |
35 | nested CFG regions forming a tree. Currently the regions are |
36 | the entire function for the root region and natural loops for |
37 | the other regions. Therefore data structure representing a |
38 | region is called loop_tree_node. |
39 | |
40 | o *Allocno class* is a register class used for allocation of |
41 | given allocno. It means that only hard register of given |
42 | register class can be assigned to given allocno. In reality, |
43 | even smaller subset of (*profitable*) hard registers can be |
44 | assigned. In rare cases, the subset can be even smaller |
45 | because our modification of Chaitin-Briggs algorithm requires |
46 | that sets of hard registers can be assigned to allocnos forms a |
47 | forest, i.e. the sets can be ordered in a way where any |
48 | previous set is not intersected with given set or is a superset |
49 | of given set. |
50 | |
51 | o *Pressure class* is a register class belonging to a set of |
52 | register classes containing all of the hard-registers available |
53 | for register allocation. The set of all pressure classes for a |
54 | target is defined in the corresponding machine-description file |
55 | according some criteria. Register pressure is calculated only |
56 | for pressure classes and it affects some IRA decisions as |
57 | forming allocation regions. |
58 | |
59 | o *Allocno* represents the live range of a pseudo-register in a |
60 | region. Besides the obvious attributes like the corresponding |
61 | pseudo-register number, allocno class, conflicting allocnos and |
62 | conflicting hard-registers, there are a few allocno attributes |
63 | which are important for understanding the allocation algorithm: |
64 | |
65 | - *Live ranges*. This is a list of ranges of *program points* |
66 | where the allocno lives. Program points represent places |
67 | where a pseudo can be born or become dead (there are |
68 | approximately two times more program points than the insns) |
69 | and they are represented by integers starting with 0. The |
70 | live ranges are used to find conflicts between allocnos. |
71 | They also play very important role for the transformation of |
72 | the IRA internal representation of several regions into a one |
73 | region representation. The later is used during the reload |
74 | pass work because each allocno represents all of the |
75 | corresponding pseudo-registers. |
76 | |
77 | - *Hard-register costs*. This is a vector of size equal to the |
78 | number of available hard-registers of the allocno class. The |
79 | cost of a callee-clobbered hard-register for an allocno is |
80 | increased by the cost of save/restore code around the calls |
81 | through the given allocno's life. If the allocno is a move |
82 | instruction operand and another operand is a hard-register of |
83 | the allocno class, the cost of the hard-register is decreased |
84 | by the move cost. |
85 | |
86 | When an allocno is assigned, the hard-register with minimal |
87 | full cost is used. Initially, a hard-register's full cost is |
88 | the corresponding value from the hard-register's cost vector. |
89 | If the allocno is connected by a *copy* (see below) to |
90 | another allocno which has just received a hard-register, the |
91 | cost of the hard-register is decreased. Before choosing a |
92 | hard-register for an allocno, the allocno's current costs of |
93 | the hard-registers are modified by the conflict hard-register |
94 | costs of all of the conflicting allocnos which are not |
95 | assigned yet. |
96 | |
97 | - *Conflict hard-register costs*. This is a vector of the same |
98 | size as the hard-register costs vector. To permit an |
99 | unassigned allocno to get a better hard-register, IRA uses |
100 | this vector to calculate the final full cost of the |
101 | available hard-registers. Conflict hard-register costs of an |
102 | unassigned allocno are also changed with a change of the |
103 | hard-register cost of the allocno when a copy involving the |
104 | allocno is processed as described above. This is done to |
105 | show other unassigned allocnos that a given allocno prefers |
106 | some hard-registers in order to remove the move instruction |
107 | corresponding to the copy. |
108 | |
109 | o *Cap*. If a pseudo-register does not live in a region but |
110 | lives in a nested region, IRA creates a special allocno called |
111 | a cap in the outer region. A region cap is also created for a |
112 | subregion cap. |
113 | |
114 | o *Copy*. Allocnos can be connected by copies. Copies are used |
115 | to modify hard-register costs for allocnos during coloring. |
116 | Such modifications reflects a preference to use the same |
117 | hard-register for the allocnos connected by copies. Usually |
118 | copies are created for move insns (in this case it results in |
119 | register coalescing). But IRA also creates copies for operands |
120 | of an insn which should be assigned to the same hard-register |
121 | due to constraints in the machine description (it usually |
122 | results in removing a move generated in reload to satisfy |
123 | the constraints) and copies referring to the allocno which is |
124 | the output operand of an instruction and the allocno which is |
125 | an input operand dying in the instruction (creation of such |
126 | copies results in less register shuffling). IRA *does not* |
127 | create copies between the same register allocnos from different |
128 | regions because we use another technique for propagating |
129 | hard-register preference on the borders of regions. |
130 | |
131 | Allocnos (including caps) for the upper region in the region tree |
132 | *accumulate* information important for coloring from allocnos with |
133 | the same pseudo-register from nested regions. This includes |
134 | hard-register and memory costs, conflicts with hard-registers, |
135 | allocno conflicts, allocno copies and more. *Thus, attributes for |
136 | allocnos in a region have the same values as if the region had no |
137 | subregions*. It means that attributes for allocnos in the |
138 | outermost region corresponding to the function have the same values |
139 | as though the allocation used only one region which is the entire |
140 | function. It also means that we can look at IRA work as if the |
141 | first IRA did allocation for all function then it improved the |
142 | allocation for loops then their subloops and so on. |
143 | |
144 | IRA major passes are: |
145 | |
146 | o Building IRA internal representation which consists of the |
147 | following subpasses: |
148 | |
149 | * First, IRA builds regions and creates allocnos (file |
150 | ira-build.cc) and initializes most of their attributes. |
151 | |
152 | * Then IRA finds an allocno class for each allocno and |
153 | calculates its initial (non-accumulated) cost of memory and |
154 | each hard-register of its allocno class (file ira-cost.c). |
155 | |
156 | * IRA creates live ranges of each allocno, calculates register |
157 | pressure for each pressure class in each region, sets up |
158 | conflict hard registers for each allocno and info about calls |
159 | the allocno lives through (file ira-lives.cc). |
160 | |
161 | * IRA removes low register pressure loops from the regions |
162 | mostly to speed IRA up (file ira-build.cc). |
163 | |
164 | * IRA propagates accumulated allocno info from lower region |
165 | allocnos to corresponding upper region allocnos (file |
166 | ira-build.cc). |
167 | |
168 | * IRA creates all caps (file ira-build.cc). |
169 | |
170 | * Having live-ranges of allocnos and their classes, IRA creates |
171 | conflicting allocnos for each allocno. Conflicting allocnos |
172 | are stored as a bit vector or array of pointers to the |
173 | conflicting allocnos whatever is more profitable (file |
174 | ira-conflicts.cc). At this point IRA creates allocno copies. |
175 | |
176 | o Coloring. Now IRA has all necessary info to start graph coloring |
177 | process. It is done in each region on top-down traverse of the |
178 | region tree (file ira-color.cc). There are following subpasses: |
179 | |
180 | * Finding profitable hard registers of corresponding allocno |
181 | class for each allocno. For example, only callee-saved hard |
182 | registers are frequently profitable for allocnos living |
183 | through colors. If the profitable hard register set of |
184 | allocno does not form a tree based on subset relation, we use |
185 | some approximation to form the tree. This approximation is |
186 | used to figure out trivial colorability of allocnos. The |
187 | approximation is a pretty rare case. |
188 | |
189 | * Putting allocnos onto the coloring stack. IRA uses Briggs |
190 | optimistic coloring which is a major improvement over |
191 | Chaitin's coloring. Therefore IRA does not spill allocnos at |
192 | this point. There is some freedom in the order of putting |
193 | allocnos on the stack which can affect the final result of |
194 | the allocation. IRA uses some heuristics to improve the |
195 | order. The major one is to form *threads* from colorable |
196 | allocnos and push them on the stack by threads. Thread is a |
197 | set of non-conflicting colorable allocnos connected by |
198 | copies. The thread contains allocnos from the colorable |
199 | bucket or colorable allocnos already pushed onto the coloring |
200 | stack. Pushing thread allocnos one after another onto the |
201 | stack increases chances of removing copies when the allocnos |
202 | get the same hard reg. |
203 | |
204 | We also use a modification of Chaitin-Briggs algorithm which |
205 | works for intersected register classes of allocnos. To |
206 | figure out trivial colorability of allocnos, the mentioned |
207 | above tree of hard register sets is used. To get an idea how |
208 | the algorithm works in i386 example, let us consider an |
209 | allocno to which any general hard register can be assigned. |
210 | If the allocno conflicts with eight allocnos to which only |
211 | EAX register can be assigned, given allocno is still |
212 | trivially colorable because all conflicting allocnos might be |
213 | assigned only to EAX and all other general hard registers are |
214 | still free. |
215 | |
216 | To get an idea of the used trivial colorability criterion, it |
217 | is also useful to read article "Graph-Coloring Register |
218 | Allocation for Irregular Architectures" by Michael D. Smith |
219 | and Glen Holloway. Major difference between the article |
220 | approach and approach used in IRA is that Smith's approach |
221 | takes register classes only from machine description and IRA |
222 | calculate register classes from intermediate code too |
223 | (e.g. an explicit usage of hard registers in RTL code for |
224 | parameter passing can result in creation of additional |
225 | register classes which contain or exclude the hard |
226 | registers). That makes IRA approach useful for improving |
227 | coloring even for architectures with regular register files |
228 | and in fact some benchmarking shows the improvement for |
229 | regular class architectures is even bigger than for irregular |
230 | ones. Another difference is that Smith's approach chooses |
231 | intersection of classes of all insn operands in which a given |
232 | pseudo occurs. IRA can use bigger classes if it is still |
233 | more profitable than memory usage. |
234 | |
235 | * Popping the allocnos from the stack and assigning them hard |
236 | registers. If IRA cannot assign a hard register to an |
237 | allocno and the allocno is coalesced, IRA undoes the |
238 | coalescing and puts the uncoalesced allocnos onto the stack in |
239 | the hope that some such allocnos will get a hard register |
240 | separately. If IRA fails to assign hard register or memory |
241 | is more profitable for it, IRA spills the allocno. IRA |
242 | assigns the allocno the hard-register with minimal full |
243 | allocation cost which reflects the cost of usage of the |
244 | hard-register for the allocno and cost of usage of the |
245 | hard-register for allocnos conflicting with given allocno. |
246 | |
247 | * Chaitin-Briggs coloring assigns as many pseudos as possible |
248 | to hard registers. After coloring we try to improve |
249 | allocation with cost point of view. We improve the |
250 | allocation by spilling some allocnos and assigning the freed |
251 | hard registers to other allocnos if it decreases the overall |
252 | allocation cost. |
253 | |
254 | * After allocno assigning in the region, IRA modifies the hard |
255 | register and memory costs for the corresponding allocnos in |
256 | the subregions to reflect the cost of possible loads, stores, |
257 | or moves on the border of the region and its subregions. |
258 | When default regional allocation algorithm is used |
259 | (-fira-algorithm=mixed), IRA just propagates the assignment |
260 | for allocnos if the register pressure in the region for the |
261 | corresponding pressure class is less than number of available |
262 | hard registers for given pressure class. |
263 | |
264 | o Spill/restore code moving. When IRA performs an allocation |
265 | by traversing regions in top-down order, it does not know what |
266 | happens below in the region tree. Therefore, sometimes IRA |
267 | misses opportunities to perform a better allocation. A simple |
268 | optimization tries to improve allocation in a region having |
269 | subregions and containing in another region. If the |
270 | corresponding allocnos in the subregion are spilled, it spills |
271 | the region allocno if it is profitable. The optimization |
272 | implements a simple iterative algorithm performing profitable |
273 | transformations while they are still possible. It is fast in |
274 | practice, so there is no real need for a better time complexity |
275 | algorithm. |
276 | |
277 | o Code change. After coloring, two allocnos representing the |
278 | same pseudo-register outside and inside a region respectively |
279 | may be assigned to different locations (hard-registers or |
280 | memory). In this case IRA creates and uses a new |
281 | pseudo-register inside the region and adds code to move allocno |
282 | values on the region's borders. This is done during top-down |
283 | traversal of the regions (file ira-emit.cc). In some |
284 | complicated cases IRA can create a new allocno to move allocno |
285 | values (e.g. when a swap of values stored in two hard-registers |
286 | is needed). At this stage, the new allocno is marked as |
287 | spilled. IRA still creates the pseudo-register and the moves |
288 | on the region borders even when both allocnos were assigned to |
289 | the same hard-register. If the reload pass spills a |
290 | pseudo-register for some reason, the effect will be smaller |
291 | because another allocno will still be in the hard-register. In |
292 | most cases, this is better then spilling both allocnos. If |
293 | reload does not change the allocation for the two |
294 | pseudo-registers, the trivial move will be removed by |
295 | post-reload optimizations. IRA does not generate moves for |
296 | allocnos assigned to the same hard register when the default |
297 | regional allocation algorithm is used and the register pressure |
298 | in the region for the corresponding pressure class is less than |
299 | number of available hard registers for given pressure class. |
300 | IRA also does some optimizations to remove redundant stores and |
301 | to reduce code duplication on the region borders. |
302 | |
303 | o Flattening internal representation. After changing code, IRA |
304 | transforms its internal representation for several regions into |
305 | one region representation (file ira-build.cc). This process is |
306 | called IR flattening. Such process is more complicated than IR |
307 | rebuilding would be, but is much faster. |
308 | |
309 | o After IR flattening, IRA tries to assign hard registers to all |
310 | spilled allocnos. This is implemented by a simple and fast |
311 | priority coloring algorithm (see function |
312 | ira_reassign_conflict_allocnos::ira-color.cc). Here new allocnos |
313 | created during the code change pass can be assigned to hard |
314 | registers. |
315 | |
316 | o At the end IRA calls the reload pass. The reload pass |
317 | communicates with IRA through several functions in file |
318 | ira-color.cc to improve its decisions in |
319 | |
320 | * sharing stack slots for the spilled pseudos based on IRA info |
321 | about pseudo-register conflicts. |
322 | |
323 | * reassigning hard-registers to all spilled pseudos at the end |
324 | of each reload iteration. |
325 | |
326 | * choosing a better hard-register to spill based on IRA info |
327 | about pseudo-register live ranges and the register pressure |
328 | in places where the pseudo-register lives. |
329 | |
330 | IRA uses a lot of data representing the target processors. These |
331 | data are initialized in file ira.cc. |
332 | |
333 | If function has no loops (or the loops are ignored when |
334 | -fira-algorithm=CB is used), we have classic Chaitin-Briggs |
335 | coloring (only instead of separate pass of coalescing, we use hard |
336 | register preferencing). In such case, IRA works much faster |
337 | because many things are not made (like IR flattening, the |
338 | spill/restore optimization, and the code change). |
339 | |
340 | Literature is worth to read for better understanding the code: |
341 | |
342 | o Preston Briggs, Keith D. Cooper, Linda Torczon. Improvements to |
343 | Graph Coloring Register Allocation. |
344 | |
345 | o David Callahan, Brian Koblenz. Register allocation via |
346 | hierarchical graph coloring. |
347 | |
348 | o Keith Cooper, Anshuman Dasgupta, Jason Eckhardt. Revisiting Graph |
349 | Coloring Register Allocation: A Study of the Chaitin-Briggs and |
350 | Callahan-Koblenz Algorithms. |
351 | |
352 | o Guei-Yuan Lueh, Thomas Gross, and Ali-Reza Adl-Tabatabai. Global |
353 | Register Allocation Based on Graph Fusion. |
354 | |
355 | o Michael D. Smith and Glenn Holloway. Graph-Coloring Register |
356 | Allocation for Irregular Architectures |
357 | |
358 | o Vladimir Makarov. The Integrated Register Allocator for GCC. |
359 | |
360 | o Vladimir Makarov. The top-down register allocator for irregular |
361 | register file architectures. |
362 | |
363 | */ |
364 | |
365 | |
366 | #include "config.h" |
367 | #include "system.h" |
368 | #include "coretypes.h" |
369 | #include "backend.h" |
370 | #include "target.h" |
371 | #include "rtl.h" |
372 | #include "tree.h" |
373 | #include "df.h" |
374 | #include "memmodel.h" |
375 | #include "tm_p.h" |
376 | #include "insn-config.h" |
377 | #include "regs.h" |
378 | #include "ira.h" |
379 | #include "ira-int.h" |
380 | #include "diagnostic-core.h" |
381 | #include "cfgrtl.h" |
382 | #include "cfgbuild.h" |
383 | #include "cfgcleanup.h" |
384 | #include "expr.h" |
385 | #include "tree-pass.h" |
386 | #include "output.h" |
387 | #include "reload.h" |
388 | #include "cfgloop.h" |
389 | #include "lra.h" |
390 | #include "dce.h" |
391 | #include "dbgcnt.h" |
392 | #include "rtl-iter.h" |
393 | #include "shrink-wrap.h" |
394 | #include "print-rtl.h" |
395 | |
396 | struct target_ira default_target_ira; |
397 | class target_ira_int default_target_ira_int; |
398 | #if SWITCHABLE_TARGET |
399 | struct target_ira *this_target_ira = &default_target_ira; |
400 | class target_ira_int *this_target_ira_int = &default_target_ira_int; |
401 | #endif |
402 | |
403 | /* A modified value of flag `-fira-verbose' used internally. */ |
404 | int internal_flag_ira_verbose; |
405 | |
406 | /* Dump file of the allocator if it is not NULL. */ |
407 | FILE *ira_dump_file; |
408 | |
409 | /* The number of elements in the following array. */ |
410 | int ira_spilled_reg_stack_slots_num; |
411 | |
412 | /* The following array contains info about spilled pseudo-registers |
413 | stack slots used in current function so far. */ |
414 | class ira_spilled_reg_stack_slot *ira_spilled_reg_stack_slots; |
415 | |
416 | /* Correspondingly overall cost of the allocation, overall cost before |
417 | reload, cost of the allocnos assigned to hard-registers, cost of |
418 | the allocnos assigned to memory, cost of loads, stores and register |
419 | move insns generated for pseudo-register live range splitting (see |
420 | ira-emit.cc). */ |
421 | int64_t ira_overall_cost, overall_cost_before; |
422 | int64_t ira_reg_cost, ira_mem_cost; |
423 | int64_t ira_load_cost, ira_store_cost, ira_shuffle_cost; |
424 | int ira_move_loops_num, ira_additional_jumps_num; |
425 | |
426 | /* All registers that can be eliminated. */ |
427 | |
428 | HARD_REG_SET eliminable_regset; |
429 | |
430 | /* Value of max_reg_num () before IRA work start. This value helps |
431 | us to recognize a situation when new pseudos were created during |
432 | IRA work. */ |
433 | static int max_regno_before_ira; |
434 | |
435 | /* Temporary hard reg set used for a different calculation. */ |
436 | static HARD_REG_SET temp_hard_regset; |
437 | |
438 | #define last_mode_for_init_move_cost \ |
439 | (this_target_ira_int->x_last_mode_for_init_move_cost) |
440 | |
441 | |
442 | /* The function sets up the map IRA_REG_MODE_HARD_REGSET. */ |
443 | static void |
444 | setup_reg_mode_hard_regset (void) |
445 | { |
446 | int i, m, hard_regno; |
447 | |
448 | for (m = 0; m < NUM_MACHINE_MODES; m++) |
449 | for (hard_regno = 0; hard_regno < FIRST_PSEUDO_REGISTER; hard_regno++) |
450 | { |
451 | CLEAR_HARD_REG_SET (ira_reg_mode_hard_regset[hard_regno][m]); |
452 | for (i = hard_regno_nregs (regno: hard_regno, mode: (machine_mode) m) - 1; |
453 | i >= 0; i--) |
454 | if (hard_regno + i < FIRST_PSEUDO_REGISTER) |
455 | SET_HARD_REG_BIT (ira_reg_mode_hard_regset[hard_regno][m], |
456 | bit: hard_regno + i); |
457 | } |
458 | } |
459 | |
460 | |
461 | #define no_unit_alloc_regs \ |
462 | (this_target_ira_int->x_no_unit_alloc_regs) |
463 | |
464 | /* The function sets up the three arrays declared above. */ |
465 | static void |
466 | setup_class_hard_regs (void) |
467 | { |
468 | int cl, i, hard_regno, n; |
469 | HARD_REG_SET processed_hard_reg_set; |
470 | |
471 | ira_assert (SHRT_MAX >= FIRST_PSEUDO_REGISTER); |
472 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
473 | { |
474 | temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs; |
475 | CLEAR_HARD_REG_SET (set&: processed_hard_reg_set); |
476 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
477 | { |
478 | ira_non_ordered_class_hard_regs[cl][i] = -1; |
479 | ira_class_hard_reg_index[cl][i] = -1; |
480 | } |
481 | for (n = 0, i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
482 | { |
483 | #ifdef REG_ALLOC_ORDER |
484 | hard_regno = reg_alloc_order[i]; |
485 | #else |
486 | hard_regno = i; |
487 | #endif |
488 | if (TEST_HARD_REG_BIT (set: processed_hard_reg_set, bit: hard_regno)) |
489 | continue; |
490 | SET_HARD_REG_BIT (set&: processed_hard_reg_set, bit: hard_regno); |
491 | if (! TEST_HARD_REG_BIT (set: temp_hard_regset, bit: hard_regno)) |
492 | ira_class_hard_reg_index[cl][hard_regno] = -1; |
493 | else |
494 | { |
495 | ira_class_hard_reg_index[cl][hard_regno] = n; |
496 | ira_class_hard_regs[cl][n++] = hard_regno; |
497 | } |
498 | } |
499 | ira_class_hard_regs_num[cl] = n; |
500 | for (n = 0, i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
501 | if (TEST_HARD_REG_BIT (set: temp_hard_regset, bit: i)) |
502 | ira_non_ordered_class_hard_regs[cl][n++] = i; |
503 | ira_assert (ira_class_hard_regs_num[cl] == n); |
504 | } |
505 | } |
506 | |
507 | /* Set up global variables defining info about hard registers for the |
508 | allocation. These depend on USE_HARD_FRAME_P whose TRUE value means |
509 | that we can use the hard frame pointer for the allocation. */ |
510 | static void |
511 | setup_alloc_regs (bool use_hard_frame_p) |
512 | { |
513 | #ifdef ADJUST_REG_ALLOC_ORDER |
514 | ADJUST_REG_ALLOC_ORDER; |
515 | #endif |
516 | no_unit_alloc_regs = fixed_nonglobal_reg_set; |
517 | if (! use_hard_frame_p) |
518 | add_to_hard_reg_set (regs: &no_unit_alloc_regs, Pmode, |
519 | HARD_FRAME_POINTER_REGNUM); |
520 | setup_class_hard_regs (); |
521 | } |
522 | |
523 | |
524 | |
525 | #define alloc_reg_class_subclasses \ |
526 | (this_target_ira_int->x_alloc_reg_class_subclasses) |
527 | |
528 | /* Initialize the table of subclasses of each reg class. */ |
529 | static void |
530 | setup_reg_subclasses (void) |
531 | { |
532 | int i, j; |
533 | HARD_REG_SET temp_hard_regset2; |
534 | |
535 | for (i = 0; i < N_REG_CLASSES; i++) |
536 | for (j = 0; j < N_REG_CLASSES; j++) |
537 | alloc_reg_class_subclasses[i][j] = LIM_REG_CLASSES; |
538 | |
539 | for (i = 0; i < N_REG_CLASSES; i++) |
540 | { |
541 | if (i == (int) NO_REGS) |
542 | continue; |
543 | |
544 | temp_hard_regset = reg_class_contents[i] & ~no_unit_alloc_regs; |
545 | if (hard_reg_set_empty_p (x: temp_hard_regset)) |
546 | continue; |
547 | for (j = 0; j < N_REG_CLASSES; j++) |
548 | if (i != j) |
549 | { |
550 | enum reg_class *p; |
551 | |
552 | temp_hard_regset2 = reg_class_contents[j] & ~no_unit_alloc_regs; |
553 | if (! hard_reg_set_subset_p (x: temp_hard_regset, |
554 | y: temp_hard_regset2)) |
555 | continue; |
556 | p = &alloc_reg_class_subclasses[j][0]; |
557 | while (*p != LIM_REG_CLASSES) p++; |
558 | *p = (enum reg_class) i; |
559 | } |
560 | } |
561 | } |
562 | |
563 | |
564 | |
565 | /* Set up IRA_MEMORY_MOVE_COST and IRA_MAX_MEMORY_MOVE_COST. */ |
566 | static void |
567 | setup_class_subset_and_memory_move_costs (void) |
568 | { |
569 | int cl, cl2, mode, cost; |
570 | HARD_REG_SET temp_hard_regset2; |
571 | |
572 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
573 | ira_memory_move_cost[mode][NO_REGS][0] |
574 | = ira_memory_move_cost[mode][NO_REGS][1] = SHRT_MAX; |
575 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
576 | { |
577 | if (cl != (int) NO_REGS) |
578 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
579 | { |
580 | ira_max_memory_move_cost[mode][cl][0] |
581 | = ira_memory_move_cost[mode][cl][0] |
582 | = memory_move_cost ((machine_mode) mode, |
583 | (reg_class_t) cl, false); |
584 | ira_max_memory_move_cost[mode][cl][1] |
585 | = ira_memory_move_cost[mode][cl][1] |
586 | = memory_move_cost ((machine_mode) mode, |
587 | (reg_class_t) cl, true); |
588 | /* Costs for NO_REGS are used in cost calculation on the |
589 | 1st pass when the preferred register classes are not |
590 | known yet. In this case we take the best scenario. */ |
591 | if (!targetm.hard_regno_mode_ok (ira_class_hard_regs[cl][0], |
592 | (machine_mode) mode)) |
593 | continue; |
594 | |
595 | if (ira_memory_move_cost[mode][NO_REGS][0] |
596 | > ira_memory_move_cost[mode][cl][0]) |
597 | ira_max_memory_move_cost[mode][NO_REGS][0] |
598 | = ira_memory_move_cost[mode][NO_REGS][0] |
599 | = ira_memory_move_cost[mode][cl][0]; |
600 | if (ira_memory_move_cost[mode][NO_REGS][1] |
601 | > ira_memory_move_cost[mode][cl][1]) |
602 | ira_max_memory_move_cost[mode][NO_REGS][1] |
603 | = ira_memory_move_cost[mode][NO_REGS][1] |
604 | = ira_memory_move_cost[mode][cl][1]; |
605 | } |
606 | } |
607 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
608 | for (cl2 = (int) N_REG_CLASSES - 1; cl2 >= 0; cl2--) |
609 | { |
610 | temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs; |
611 | temp_hard_regset2 = reg_class_contents[cl2] & ~no_unit_alloc_regs; |
612 | ira_class_subset_p[cl][cl2] |
613 | = hard_reg_set_subset_p (x: temp_hard_regset, y: temp_hard_regset2); |
614 | if (! hard_reg_set_empty_p (x: temp_hard_regset2) |
615 | && hard_reg_set_subset_p (reg_class_contents[cl2], |
616 | reg_class_contents[cl])) |
617 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
618 | { |
619 | cost = ira_memory_move_cost[mode][cl2][0]; |
620 | if (cost > ira_max_memory_move_cost[mode][cl][0]) |
621 | ira_max_memory_move_cost[mode][cl][0] = cost; |
622 | cost = ira_memory_move_cost[mode][cl2][1]; |
623 | if (cost > ira_max_memory_move_cost[mode][cl][1]) |
624 | ira_max_memory_move_cost[mode][cl][1] = cost; |
625 | } |
626 | } |
627 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
628 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
629 | { |
630 | ira_memory_move_cost[mode][cl][0] |
631 | = ira_max_memory_move_cost[mode][cl][0]; |
632 | ira_memory_move_cost[mode][cl][1] |
633 | = ira_max_memory_move_cost[mode][cl][1]; |
634 | } |
635 | setup_reg_subclasses (); |
636 | } |
637 | |
638 | |
639 | |
640 | /* Define the following macro if allocation through malloc if |
641 | preferable. */ |
642 | #define IRA_NO_OBSTACK |
643 | |
644 | #ifndef IRA_NO_OBSTACK |
645 | /* Obstack used for storing all dynamic data (except bitmaps) of the |
646 | IRA. */ |
647 | static struct obstack ira_obstack; |
648 | #endif |
649 | |
650 | /* Obstack used for storing all bitmaps of the IRA. */ |
651 | static struct bitmap_obstack ira_bitmap_obstack; |
652 | |
653 | /* Allocate memory of size LEN for IRA data. */ |
654 | void * |
655 | ira_allocate (size_t len) |
656 | { |
657 | void *res; |
658 | |
659 | #ifndef IRA_NO_OBSTACK |
660 | res = obstack_alloc (&ira_obstack, len); |
661 | #else |
662 | res = xmalloc (len); |
663 | #endif |
664 | return res; |
665 | } |
666 | |
667 | /* Free memory ADDR allocated for IRA data. */ |
668 | void |
669 | ira_free (void *addr ATTRIBUTE_UNUSED) |
670 | { |
671 | #ifndef IRA_NO_OBSTACK |
672 | /* do nothing */ |
673 | #else |
674 | free (ptr: addr); |
675 | #endif |
676 | } |
677 | |
678 | |
679 | /* Allocate and returns bitmap for IRA. */ |
680 | bitmap |
681 | ira_allocate_bitmap (void) |
682 | { |
683 | return BITMAP_ALLOC (obstack: &ira_bitmap_obstack); |
684 | } |
685 | |
686 | /* Free bitmap B allocated for IRA. */ |
687 | void |
688 | ira_free_bitmap (bitmap b ATTRIBUTE_UNUSED) |
689 | { |
690 | /* do nothing */ |
691 | } |
692 | |
693 | |
694 | |
695 | /* Output information about allocation of all allocnos (except for |
696 | caps) into file F. */ |
697 | void |
698 | ira_print_disposition (FILE *f) |
699 | { |
700 | int i, n, max_regno; |
701 | ira_allocno_t a; |
702 | basic_block bb; |
703 | |
704 | fprintf (stream: f, format: "Disposition:" ); |
705 | max_regno = max_reg_num (); |
706 | for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) |
707 | for (a = ira_regno_allocno_map[i]; |
708 | a != NULL; |
709 | a = ALLOCNO_NEXT_REGNO_ALLOCNO (a)) |
710 | { |
711 | if (n % 4 == 0) |
712 | fprintf (stream: f, format: "\n" ); |
713 | n++; |
714 | fprintf (stream: f, format: " %4d:r%-4d" , ALLOCNO_NUM (a), ALLOCNO_REGNO (a)); |
715 | if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL) |
716 | fprintf (stream: f, format: "b%-3d" , bb->index); |
717 | else |
718 | fprintf (stream: f, format: "l%-3d" , ALLOCNO_LOOP_TREE_NODE (a)->loop_num); |
719 | if (ALLOCNO_HARD_REGNO (a) >= 0) |
720 | fprintf (stream: f, format: " %3d" , ALLOCNO_HARD_REGNO (a)); |
721 | else |
722 | fprintf (stream: f, format: " mem" ); |
723 | } |
724 | fprintf (stream: f, format: "\n" ); |
725 | } |
726 | |
727 | /* Outputs information about allocation of all allocnos into |
728 | stderr. */ |
729 | void |
730 | ira_debug_disposition (void) |
731 | { |
732 | ira_print_disposition (stderr); |
733 | } |
734 | |
735 | |
736 | |
737 | /* Set up ira_stack_reg_pressure_class which is the biggest pressure |
738 | register class containing stack registers or NO_REGS if there are |
739 | no stack registers. To find this class, we iterate through all |
740 | register pressure classes and choose the first register pressure |
741 | class containing all the stack registers and having the biggest |
742 | size. */ |
743 | static void |
744 | setup_stack_reg_pressure_class (void) |
745 | { |
746 | ira_stack_reg_pressure_class = NO_REGS; |
747 | #ifdef STACK_REGS |
748 | { |
749 | int i, best, size; |
750 | enum reg_class cl; |
751 | HARD_REG_SET temp_hard_regset2; |
752 | |
753 | CLEAR_HARD_REG_SET (set&: temp_hard_regset); |
754 | for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++) |
755 | SET_HARD_REG_BIT (set&: temp_hard_regset, bit: i); |
756 | best = 0; |
757 | for (i = 0; i < ira_pressure_classes_num; i++) |
758 | { |
759 | cl = ira_pressure_classes[i]; |
760 | temp_hard_regset2 = temp_hard_regset & reg_class_contents[cl]; |
761 | size = hard_reg_set_size (set: temp_hard_regset2); |
762 | if (best < size) |
763 | { |
764 | best = size; |
765 | ira_stack_reg_pressure_class = cl; |
766 | } |
767 | } |
768 | } |
769 | #endif |
770 | } |
771 | |
772 | /* Find pressure classes which are register classes for which we |
773 | calculate register pressure in IRA, register pressure sensitive |
774 | insn scheduling, and register pressure sensitive loop invariant |
775 | motion. |
776 | |
777 | To make register pressure calculation easy, we always use |
778 | non-intersected register pressure classes. A move of hard |
779 | registers from one register pressure class is not more expensive |
780 | than load and store of the hard registers. Most likely an allocno |
781 | class will be a subset of a register pressure class and in many |
782 | cases a register pressure class. That makes usage of register |
783 | pressure classes a good approximation to find a high register |
784 | pressure. */ |
785 | static void |
786 | setup_pressure_classes (void) |
787 | { |
788 | int cost, i, n, curr; |
789 | int cl, cl2; |
790 | enum reg_class pressure_classes[N_REG_CLASSES]; |
791 | int m; |
792 | HARD_REG_SET temp_hard_regset2; |
793 | bool insert_p; |
794 | |
795 | if (targetm.compute_pressure_classes) |
796 | n = targetm.compute_pressure_classes (pressure_classes); |
797 | else |
798 | { |
799 | n = 0; |
800 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
801 | { |
802 | if (ira_class_hard_regs_num[cl] == 0) |
803 | continue; |
804 | if (ira_class_hard_regs_num[cl] != 1 |
805 | /* A register class without subclasses may contain a few |
806 | hard registers and movement between them is costly |
807 | (e.g. SPARC FPCC registers). We still should consider it |
808 | as a candidate for a pressure class. */ |
809 | && alloc_reg_class_subclasses[cl][0] < cl) |
810 | { |
811 | /* Check that the moves between any hard registers of the |
812 | current class are not more expensive for a legal mode |
813 | than load/store of the hard registers of the current |
814 | class. Such class is a potential candidate to be a |
815 | register pressure class. */ |
816 | for (m = 0; m < NUM_MACHINE_MODES; m++) |
817 | { |
818 | temp_hard_regset |
819 | = (reg_class_contents[cl] |
820 | & ~(no_unit_alloc_regs |
821 | | ira_prohibited_class_mode_regs[cl][m])); |
822 | if (hard_reg_set_empty_p (x: temp_hard_regset)) |
823 | continue; |
824 | ira_init_register_move_cost_if_necessary (mode: (machine_mode) m); |
825 | cost = ira_register_move_cost[m][cl][cl]; |
826 | if (cost <= ira_max_memory_move_cost[m][cl][1] |
827 | || cost <= ira_max_memory_move_cost[m][cl][0]) |
828 | break; |
829 | } |
830 | if (m >= NUM_MACHINE_MODES) |
831 | continue; |
832 | } |
833 | curr = 0; |
834 | insert_p = true; |
835 | temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs; |
836 | /* Remove so far added pressure classes which are subset of the |
837 | current candidate class. Prefer GENERAL_REGS as a pressure |
838 | register class to another class containing the same |
839 | allocatable hard registers. We do this because machine |
840 | dependent cost hooks might give wrong costs for the latter |
841 | class but always give the right cost for the former class |
842 | (GENERAL_REGS). */ |
843 | for (i = 0; i < n; i++) |
844 | { |
845 | cl2 = pressure_classes[i]; |
846 | temp_hard_regset2 = (reg_class_contents[cl2] |
847 | & ~no_unit_alloc_regs); |
848 | if (hard_reg_set_subset_p (x: temp_hard_regset, y: temp_hard_regset2) |
849 | && (temp_hard_regset != temp_hard_regset2 |
850 | || cl2 == (int) GENERAL_REGS)) |
851 | { |
852 | pressure_classes[curr++] = (enum reg_class) cl2; |
853 | insert_p = false; |
854 | continue; |
855 | } |
856 | if (hard_reg_set_subset_p (x: temp_hard_regset2, y: temp_hard_regset) |
857 | && (temp_hard_regset2 != temp_hard_regset |
858 | || cl == (int) GENERAL_REGS)) |
859 | continue; |
860 | if (temp_hard_regset2 == temp_hard_regset) |
861 | insert_p = false; |
862 | pressure_classes[curr++] = (enum reg_class) cl2; |
863 | } |
864 | /* If the current candidate is a subset of a so far added |
865 | pressure class, don't add it to the list of the pressure |
866 | classes. */ |
867 | if (insert_p) |
868 | pressure_classes[curr++] = (enum reg_class) cl; |
869 | n = curr; |
870 | } |
871 | } |
872 | #ifdef ENABLE_IRA_CHECKING |
873 | { |
874 | HARD_REG_SET ignore_hard_regs; |
875 | |
876 | /* Check pressure classes correctness: here we check that hard |
877 | registers from all register pressure classes contains all hard |
878 | registers available for the allocation. */ |
879 | CLEAR_HARD_REG_SET (set&: temp_hard_regset); |
880 | CLEAR_HARD_REG_SET (set&: temp_hard_regset2); |
881 | ignore_hard_regs = no_unit_alloc_regs; |
882 | for (cl = 0; cl < LIM_REG_CLASSES; cl++) |
883 | { |
884 | /* For some targets (like MIPS with MD_REGS), there are some |
885 | classes with hard registers available for allocation but |
886 | not able to hold value of any mode. */ |
887 | for (m = 0; m < NUM_MACHINE_MODES; m++) |
888 | if (contains_reg_of_mode[cl][m]) |
889 | break; |
890 | if (m >= NUM_MACHINE_MODES) |
891 | { |
892 | ignore_hard_regs |= reg_class_contents[cl]; |
893 | continue; |
894 | } |
895 | for (i = 0; i < n; i++) |
896 | if ((int) pressure_classes[i] == cl) |
897 | break; |
898 | temp_hard_regset2 |= reg_class_contents[cl]; |
899 | if (i < n) |
900 | temp_hard_regset |= reg_class_contents[cl]; |
901 | } |
902 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
903 | /* Some targets (like SPARC with ICC reg) have allocatable regs |
904 | for which no reg class is defined. */ |
905 | if (REGNO_REG_CLASS (i) == NO_REGS) |
906 | SET_HARD_REG_BIT (set&: ignore_hard_regs, bit: i); |
907 | temp_hard_regset &= ~ignore_hard_regs; |
908 | temp_hard_regset2 &= ~ignore_hard_regs; |
909 | ira_assert (hard_reg_set_subset_p (temp_hard_regset2, temp_hard_regset)); |
910 | } |
911 | #endif |
912 | ira_pressure_classes_num = 0; |
913 | for (i = 0; i < n; i++) |
914 | { |
915 | cl = (int) pressure_classes[i]; |
916 | ira_reg_pressure_class_p[cl] = true; |
917 | ira_pressure_classes[ira_pressure_classes_num++] = (enum reg_class) cl; |
918 | } |
919 | setup_stack_reg_pressure_class (); |
920 | } |
921 | |
922 | /* Set up IRA_UNIFORM_CLASS_P. Uniform class is a register class |
923 | whose register move cost between any registers of the class is the |
924 | same as for all its subclasses. We use the data to speed up the |
925 | 2nd pass of calculations of allocno costs. */ |
926 | static void |
927 | setup_uniform_class_p (void) |
928 | { |
929 | int i, cl, cl2, m; |
930 | |
931 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
932 | { |
933 | ira_uniform_class_p[cl] = false; |
934 | if (ira_class_hard_regs_num[cl] == 0) |
935 | continue; |
936 | /* We cannot use alloc_reg_class_subclasses here because move |
937 | cost hooks does not take into account that some registers are |
938 | unavailable for the subtarget. E.g. for i686, INT_SSE_REGS |
939 | is element of alloc_reg_class_subclasses for GENERAL_REGS |
940 | because SSE regs are unavailable. */ |
941 | for (i = 0; (cl2 = reg_class_subclasses[cl][i]) != LIM_REG_CLASSES; i++) |
942 | { |
943 | if (ira_class_hard_regs_num[cl2] == 0) |
944 | continue; |
945 | for (m = 0; m < NUM_MACHINE_MODES; m++) |
946 | if (contains_reg_of_mode[cl][m] && contains_reg_of_mode[cl2][m]) |
947 | { |
948 | ira_init_register_move_cost_if_necessary (mode: (machine_mode) m); |
949 | if (ira_register_move_cost[m][cl][cl] |
950 | != ira_register_move_cost[m][cl2][cl2]) |
951 | break; |
952 | } |
953 | if (m < NUM_MACHINE_MODES) |
954 | break; |
955 | } |
956 | if (cl2 == LIM_REG_CLASSES) |
957 | ira_uniform_class_p[cl] = true; |
958 | } |
959 | } |
960 | |
961 | /* Set up IRA_ALLOCNO_CLASSES, IRA_ALLOCNO_CLASSES_NUM, |
962 | IRA_IMPORTANT_CLASSES, and IRA_IMPORTANT_CLASSES_NUM. |
963 | |
964 | Target may have many subtargets and not all target hard registers can |
965 | be used for allocation, e.g. x86 port in 32-bit mode cannot use |
966 | hard registers introduced in x86-64 like r8-r15). Some classes |
967 | might have the same allocatable hard registers, e.g. INDEX_REGS |
968 | and GENERAL_REGS in x86 port in 32-bit mode. To decrease different |
969 | calculations efforts we introduce allocno classes which contain |
970 | unique non-empty sets of allocatable hard-registers. |
971 | |
972 | Pseudo class cost calculation in ira-costs.cc is very expensive. |
973 | Therefore we are trying to decrease number of classes involved in |
974 | such calculation. Register classes used in the cost calculation |
975 | are called important classes. They are allocno classes and other |
976 | non-empty classes whose allocatable hard register sets are inside |
977 | of an allocno class hard register set. From the first sight, it |
978 | looks like that they are just allocno classes. It is not true. In |
979 | example of x86-port in 32-bit mode, allocno classes will contain |
980 | GENERAL_REGS but not LEGACY_REGS (because allocatable hard |
981 | registers are the same for the both classes). The important |
982 | classes will contain GENERAL_REGS and LEGACY_REGS. It is done |
983 | because a machine description insn constraint may refers for |
984 | LEGACY_REGS and code in ira-costs.cc is mostly base on investigation |
985 | of the insn constraints. */ |
986 | static void |
987 | setup_allocno_and_important_classes (void) |
988 | { |
989 | int i, j, n, cl; |
990 | bool set_p; |
991 | HARD_REG_SET temp_hard_regset2; |
992 | static enum reg_class classes[LIM_REG_CLASSES + 1]; |
993 | |
994 | n = 0; |
995 | /* Collect classes which contain unique sets of allocatable hard |
996 | registers. Prefer GENERAL_REGS to other classes containing the |
997 | same set of hard registers. */ |
998 | for (i = 0; i < LIM_REG_CLASSES; i++) |
999 | { |
1000 | temp_hard_regset = reg_class_contents[i] & ~no_unit_alloc_regs; |
1001 | for (j = 0; j < n; j++) |
1002 | { |
1003 | cl = classes[j]; |
1004 | temp_hard_regset2 = reg_class_contents[cl] & ~no_unit_alloc_regs; |
1005 | if (temp_hard_regset == temp_hard_regset2) |
1006 | break; |
1007 | } |
1008 | if (j >= n || targetm.additional_allocno_class_p (i)) |
1009 | classes[n++] = (enum reg_class) i; |
1010 | else if (i == GENERAL_REGS) |
1011 | /* Prefer general regs. For i386 example, it means that |
1012 | we prefer GENERAL_REGS over INDEX_REGS or LEGACY_REGS |
1013 | (all of them consists of the same available hard |
1014 | registers). */ |
1015 | classes[j] = (enum reg_class) i; |
1016 | } |
1017 | classes[n] = LIM_REG_CLASSES; |
1018 | |
1019 | /* Set up classes which can be used for allocnos as classes |
1020 | containing non-empty unique sets of allocatable hard |
1021 | registers. */ |
1022 | ira_allocno_classes_num = 0; |
1023 | for (i = 0; (cl = classes[i]) != LIM_REG_CLASSES; i++) |
1024 | if (ira_class_hard_regs_num[cl] > 0) |
1025 | ira_allocno_classes[ira_allocno_classes_num++] = (enum reg_class) cl; |
1026 | ira_important_classes_num = 0; |
1027 | /* Add non-allocno classes containing to non-empty set of |
1028 | allocatable hard regs. */ |
1029 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1030 | if (ira_class_hard_regs_num[cl] > 0) |
1031 | { |
1032 | temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs; |
1033 | set_p = false; |
1034 | for (j = 0; j < ira_allocno_classes_num; j++) |
1035 | { |
1036 | temp_hard_regset2 = (reg_class_contents[ira_allocno_classes[j]] |
1037 | & ~no_unit_alloc_regs); |
1038 | if ((enum reg_class) cl == ira_allocno_classes[j]) |
1039 | break; |
1040 | else if (hard_reg_set_subset_p (x: temp_hard_regset, |
1041 | y: temp_hard_regset2)) |
1042 | set_p = true; |
1043 | } |
1044 | if (set_p && j >= ira_allocno_classes_num) |
1045 | ira_important_classes[ira_important_classes_num++] |
1046 | = (enum reg_class) cl; |
1047 | } |
1048 | /* Now add allocno classes to the important classes. */ |
1049 | for (j = 0; j < ira_allocno_classes_num; j++) |
1050 | ira_important_classes[ira_important_classes_num++] |
1051 | = ira_allocno_classes[j]; |
1052 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1053 | { |
1054 | ira_reg_allocno_class_p[cl] = false; |
1055 | ira_reg_pressure_class_p[cl] = false; |
1056 | } |
1057 | for (j = 0; j < ira_allocno_classes_num; j++) |
1058 | ira_reg_allocno_class_p[ira_allocno_classes[j]] = true; |
1059 | setup_pressure_classes (); |
1060 | setup_uniform_class_p (); |
1061 | } |
1062 | |
1063 | /* Setup translation in CLASS_TRANSLATE of all classes into a class |
1064 | given by array CLASSES of length CLASSES_NUM. The function is used |
1065 | make translation any reg class to an allocno class or to an |
1066 | pressure class. This translation is necessary for some |
1067 | calculations when we can use only allocno or pressure classes and |
1068 | such translation represents an approximate representation of all |
1069 | classes. |
1070 | |
1071 | The translation in case when allocatable hard register set of a |
1072 | given class is subset of allocatable hard register set of a class |
1073 | in CLASSES is pretty simple. We use smallest classes from CLASSES |
1074 | containing a given class. If allocatable hard register set of a |
1075 | given class is not a subset of any corresponding set of a class |
1076 | from CLASSES, we use the cheapest (with load/store point of view) |
1077 | class from CLASSES whose set intersects with given class set. */ |
1078 | static void |
1079 | setup_class_translate_array (enum reg_class *class_translate, |
1080 | int classes_num, enum reg_class *classes) |
1081 | { |
1082 | int cl, mode; |
1083 | enum reg_class aclass, best_class, *cl_ptr; |
1084 | int i, cost, min_cost, best_cost; |
1085 | |
1086 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1087 | class_translate[cl] = NO_REGS; |
1088 | |
1089 | for (i = 0; i < classes_num; i++) |
1090 | { |
1091 | aclass = classes[i]; |
1092 | for (cl_ptr = &alloc_reg_class_subclasses[aclass][0]; |
1093 | (cl = *cl_ptr) != LIM_REG_CLASSES; |
1094 | cl_ptr++) |
1095 | if (class_translate[cl] == NO_REGS) |
1096 | class_translate[cl] = aclass; |
1097 | class_translate[aclass] = aclass; |
1098 | } |
1099 | /* For classes which are not fully covered by one of given classes |
1100 | (in other words covered by more one given class), use the |
1101 | cheapest class. */ |
1102 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1103 | { |
1104 | if (cl == NO_REGS || class_translate[cl] != NO_REGS) |
1105 | continue; |
1106 | best_class = NO_REGS; |
1107 | best_cost = INT_MAX; |
1108 | for (i = 0; i < classes_num; i++) |
1109 | { |
1110 | aclass = classes[i]; |
1111 | temp_hard_regset = (reg_class_contents[aclass] |
1112 | & reg_class_contents[cl] |
1113 | & ~no_unit_alloc_regs); |
1114 | if (! hard_reg_set_empty_p (x: temp_hard_regset)) |
1115 | { |
1116 | min_cost = INT_MAX; |
1117 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
1118 | { |
1119 | cost = (ira_memory_move_cost[mode][aclass][0] |
1120 | + ira_memory_move_cost[mode][aclass][1]); |
1121 | if (min_cost > cost) |
1122 | min_cost = cost; |
1123 | } |
1124 | if (best_class == NO_REGS || best_cost > min_cost) |
1125 | { |
1126 | best_class = aclass; |
1127 | best_cost = min_cost; |
1128 | } |
1129 | } |
1130 | } |
1131 | class_translate[cl] = best_class; |
1132 | } |
1133 | } |
1134 | |
1135 | /* Set up array IRA_ALLOCNO_CLASS_TRANSLATE and |
1136 | IRA_PRESSURE_CLASS_TRANSLATE. */ |
1137 | static void |
1138 | setup_class_translate (void) |
1139 | { |
1140 | setup_class_translate_array (ira_allocno_class_translate, |
1141 | ira_allocno_classes_num, ira_allocno_classes); |
1142 | setup_class_translate_array (ira_pressure_class_translate, |
1143 | ira_pressure_classes_num, ira_pressure_classes); |
1144 | } |
1145 | |
1146 | /* Order numbers of allocno classes in original target allocno class |
1147 | array, -1 for non-allocno classes. */ |
1148 | static int allocno_class_order[N_REG_CLASSES]; |
1149 | |
1150 | /* The function used to sort the important classes. */ |
1151 | static int |
1152 | comp_reg_classes_func (const void *v1p, const void *v2p) |
1153 | { |
1154 | enum reg_class cl1 = *(const enum reg_class *) v1p; |
1155 | enum reg_class cl2 = *(const enum reg_class *) v2p; |
1156 | enum reg_class tcl1, tcl2; |
1157 | int diff; |
1158 | |
1159 | tcl1 = ira_allocno_class_translate[cl1]; |
1160 | tcl2 = ira_allocno_class_translate[cl2]; |
1161 | if (tcl1 != NO_REGS && tcl2 != NO_REGS |
1162 | && (diff = allocno_class_order[tcl1] - allocno_class_order[tcl2]) != 0) |
1163 | return diff; |
1164 | return (int) cl1 - (int) cl2; |
1165 | } |
1166 | |
1167 | /* For correct work of function setup_reg_class_relation we need to |
1168 | reorder important classes according to the order of their allocno |
1169 | classes. It places important classes containing the same |
1170 | allocatable hard register set adjacent to each other and allocno |
1171 | class with the allocatable hard register set right after the other |
1172 | important classes with the same set. |
1173 | |
1174 | In example from comments of function |
1175 | setup_allocno_and_important_classes, it places LEGACY_REGS and |
1176 | GENERAL_REGS close to each other and GENERAL_REGS is after |
1177 | LEGACY_REGS. */ |
1178 | static void |
1179 | reorder_important_classes (void) |
1180 | { |
1181 | int i; |
1182 | |
1183 | for (i = 0; i < N_REG_CLASSES; i++) |
1184 | allocno_class_order[i] = -1; |
1185 | for (i = 0; i < ira_allocno_classes_num; i++) |
1186 | allocno_class_order[ira_allocno_classes[i]] = i; |
1187 | qsort (ira_important_classes, ira_important_classes_num, |
1188 | sizeof (enum reg_class), comp_reg_classes_func); |
1189 | for (i = 0; i < ira_important_classes_num; i++) |
1190 | ira_important_class_nums[ira_important_classes[i]] = i; |
1191 | } |
1192 | |
1193 | /* Set up IRA_REG_CLASS_SUBUNION, IRA_REG_CLASS_SUPERUNION, |
1194 | IRA_REG_CLASS_SUPER_CLASSES, IRA_REG_CLASSES_INTERSECT, and |
1195 | IRA_REG_CLASSES_INTERSECT_P. For the meaning of the relations, |
1196 | please see corresponding comments in ira-int.h. */ |
1197 | static void |
1198 | setup_reg_class_relations (void) |
1199 | { |
1200 | int i, cl1, cl2, cl3; |
1201 | HARD_REG_SET intersection_set, union_set, temp_set2; |
1202 | bool important_class_p[N_REG_CLASSES]; |
1203 | |
1204 | memset (s: important_class_p, c: 0, n: sizeof (important_class_p)); |
1205 | for (i = 0; i < ira_important_classes_num; i++) |
1206 | important_class_p[ira_important_classes[i]] = true; |
1207 | for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++) |
1208 | { |
1209 | ira_reg_class_super_classes[cl1][0] = LIM_REG_CLASSES; |
1210 | for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++) |
1211 | { |
1212 | ira_reg_classes_intersect_p[cl1][cl2] = false; |
1213 | ira_reg_class_intersect[cl1][cl2] = NO_REGS; |
1214 | ira_reg_class_subset[cl1][cl2] = NO_REGS; |
1215 | temp_hard_regset = reg_class_contents[cl1] & ~no_unit_alloc_regs; |
1216 | temp_set2 = reg_class_contents[cl2] & ~no_unit_alloc_regs; |
1217 | if (hard_reg_set_empty_p (x: temp_hard_regset) |
1218 | && hard_reg_set_empty_p (x: temp_set2)) |
1219 | { |
1220 | /* The both classes have no allocatable hard registers |
1221 | -- take all class hard registers into account and use |
1222 | reg_class_subunion and reg_class_superunion. */ |
1223 | for (i = 0;; i++) |
1224 | { |
1225 | cl3 = reg_class_subclasses[cl1][i]; |
1226 | if (cl3 == LIM_REG_CLASSES) |
1227 | break; |
1228 | if (reg_class_subset_p (ira_reg_class_intersect[cl1][cl2], |
1229 | (enum reg_class) cl3)) |
1230 | ira_reg_class_intersect[cl1][cl2] = (enum reg_class) cl3; |
1231 | } |
1232 | ira_reg_class_subunion[cl1][cl2] = reg_class_subunion[cl1][cl2]; |
1233 | ira_reg_class_superunion[cl1][cl2] = reg_class_superunion[cl1][cl2]; |
1234 | continue; |
1235 | } |
1236 | ira_reg_classes_intersect_p[cl1][cl2] |
1237 | = hard_reg_set_intersect_p (x: temp_hard_regset, y: temp_set2); |
1238 | if (important_class_p[cl1] && important_class_p[cl2] |
1239 | && hard_reg_set_subset_p (x: temp_hard_regset, y: temp_set2)) |
1240 | { |
1241 | /* CL1 and CL2 are important classes and CL1 allocatable |
1242 | hard register set is inside of CL2 allocatable hard |
1243 | registers -- make CL1 a superset of CL2. */ |
1244 | enum reg_class *p; |
1245 | |
1246 | p = &ira_reg_class_super_classes[cl1][0]; |
1247 | while (*p != LIM_REG_CLASSES) |
1248 | p++; |
1249 | *p++ = (enum reg_class) cl2; |
1250 | *p = LIM_REG_CLASSES; |
1251 | } |
1252 | ira_reg_class_subunion[cl1][cl2] = NO_REGS; |
1253 | ira_reg_class_superunion[cl1][cl2] = NO_REGS; |
1254 | intersection_set = (reg_class_contents[cl1] |
1255 | & reg_class_contents[cl2] |
1256 | & ~no_unit_alloc_regs); |
1257 | union_set = ((reg_class_contents[cl1] | reg_class_contents[cl2]) |
1258 | & ~no_unit_alloc_regs); |
1259 | for (cl3 = 0; cl3 < N_REG_CLASSES; cl3++) |
1260 | { |
1261 | temp_hard_regset = reg_class_contents[cl3] & ~no_unit_alloc_regs; |
1262 | if (hard_reg_set_empty_p (x: temp_hard_regset)) |
1263 | continue; |
1264 | |
1265 | if (hard_reg_set_subset_p (x: temp_hard_regset, y: intersection_set)) |
1266 | { |
1267 | /* CL3 allocatable hard register set is inside of |
1268 | intersection of allocatable hard register sets |
1269 | of CL1 and CL2. */ |
1270 | if (important_class_p[cl3]) |
1271 | { |
1272 | temp_set2 |
1273 | = (reg_class_contents |
1274 | [ira_reg_class_intersect[cl1][cl2]]); |
1275 | temp_set2 &= ~no_unit_alloc_regs; |
1276 | if (! hard_reg_set_subset_p (x: temp_hard_regset, y: temp_set2) |
1277 | /* If the allocatable hard register sets are |
1278 | the same, prefer GENERAL_REGS or the |
1279 | smallest class for debugging |
1280 | purposes. */ |
1281 | || (temp_hard_regset == temp_set2 |
1282 | && (cl3 == GENERAL_REGS |
1283 | || ((ira_reg_class_intersect[cl1][cl2] |
1284 | != GENERAL_REGS) |
1285 | && hard_reg_set_subset_p |
1286 | (reg_class_contents[cl3], |
1287 | reg_class_contents |
1288 | [(int) |
1289 | ira_reg_class_intersect[cl1][cl2]]))))) |
1290 | ira_reg_class_intersect[cl1][cl2] = (enum reg_class) cl3; |
1291 | } |
1292 | temp_set2 |
1293 | = (reg_class_contents[ira_reg_class_subset[cl1][cl2]] |
1294 | & ~no_unit_alloc_regs); |
1295 | if (! hard_reg_set_subset_p (x: temp_hard_regset, y: temp_set2) |
1296 | /* Ignore unavailable hard registers and prefer |
1297 | smallest class for debugging purposes. */ |
1298 | || (temp_hard_regset == temp_set2 |
1299 | && hard_reg_set_subset_p |
1300 | (reg_class_contents[cl3], |
1301 | reg_class_contents |
1302 | [(int) ira_reg_class_subset[cl1][cl2]]))) |
1303 | ira_reg_class_subset[cl1][cl2] = (enum reg_class) cl3; |
1304 | } |
1305 | if (important_class_p[cl3] |
1306 | && hard_reg_set_subset_p (x: temp_hard_regset, y: union_set)) |
1307 | { |
1308 | /* CL3 allocatable hard register set is inside of |
1309 | union of allocatable hard register sets of CL1 |
1310 | and CL2. */ |
1311 | temp_set2 |
1312 | = (reg_class_contents[ira_reg_class_subunion[cl1][cl2]] |
1313 | & ~no_unit_alloc_regs); |
1314 | if (ira_reg_class_subunion[cl1][cl2] == NO_REGS |
1315 | || (hard_reg_set_subset_p (x: temp_set2, y: temp_hard_regset) |
1316 | |
1317 | && (temp_set2 != temp_hard_regset |
1318 | || cl3 == GENERAL_REGS |
1319 | /* If the allocatable hard register sets are the |
1320 | same, prefer GENERAL_REGS or the smallest |
1321 | class for debugging purposes. */ |
1322 | || (ira_reg_class_subunion[cl1][cl2] != GENERAL_REGS |
1323 | && hard_reg_set_subset_p |
1324 | (reg_class_contents[cl3], |
1325 | reg_class_contents |
1326 | [(int) ira_reg_class_subunion[cl1][cl2]]))))) |
1327 | ira_reg_class_subunion[cl1][cl2] = (enum reg_class) cl3; |
1328 | } |
1329 | if (hard_reg_set_subset_p (x: union_set, y: temp_hard_regset)) |
1330 | { |
1331 | /* CL3 allocatable hard register set contains union |
1332 | of allocatable hard register sets of CL1 and |
1333 | CL2. */ |
1334 | temp_set2 |
1335 | = (reg_class_contents[ira_reg_class_superunion[cl1][cl2]] |
1336 | & ~no_unit_alloc_regs); |
1337 | if (ira_reg_class_superunion[cl1][cl2] == NO_REGS |
1338 | || (hard_reg_set_subset_p (x: temp_hard_regset, y: temp_set2) |
1339 | |
1340 | && (temp_set2 != temp_hard_regset |
1341 | || cl3 == GENERAL_REGS |
1342 | /* If the allocatable hard register sets are the |
1343 | same, prefer GENERAL_REGS or the smallest |
1344 | class for debugging purposes. */ |
1345 | || (ira_reg_class_superunion[cl1][cl2] != GENERAL_REGS |
1346 | && hard_reg_set_subset_p |
1347 | (reg_class_contents[cl3], |
1348 | reg_class_contents |
1349 | [(int) ira_reg_class_superunion[cl1][cl2]]))))) |
1350 | ira_reg_class_superunion[cl1][cl2] = (enum reg_class) cl3; |
1351 | } |
1352 | } |
1353 | } |
1354 | } |
1355 | } |
1356 | |
1357 | /* Output all uniform and important classes into file F. */ |
1358 | static void |
1359 | print_uniform_and_important_classes (FILE *f) |
1360 | { |
1361 | int i, cl; |
1362 | |
1363 | fprintf (stream: f, format: "Uniform classes:\n" ); |
1364 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1365 | if (ira_uniform_class_p[cl]) |
1366 | fprintf (stream: f, format: " %s" , reg_class_names[cl]); |
1367 | fprintf (stream: f, format: "\nImportant classes:\n" ); |
1368 | for (i = 0; i < ira_important_classes_num; i++) |
1369 | fprintf (stream: f, format: " %s" , reg_class_names[ira_important_classes[i]]); |
1370 | fprintf (stream: f, format: "\n" ); |
1371 | } |
1372 | |
1373 | /* Output all possible allocno or pressure classes and their |
1374 | translation map into file F. */ |
1375 | static void |
1376 | print_translated_classes (FILE *f, bool pressure_p) |
1377 | { |
1378 | int classes_num = (pressure_p |
1379 | ? ira_pressure_classes_num : ira_allocno_classes_num); |
1380 | enum reg_class *classes = (pressure_p |
1381 | ? ira_pressure_classes : ira_allocno_classes); |
1382 | enum reg_class *class_translate = (pressure_p |
1383 | ? ira_pressure_class_translate |
1384 | : ira_allocno_class_translate); |
1385 | int i; |
1386 | |
1387 | fprintf (stream: f, format: "%s classes:\n" , pressure_p ? "Pressure" : "Allocno" ); |
1388 | for (i = 0; i < classes_num; i++) |
1389 | fprintf (stream: f, format: " %s" , reg_class_names[classes[i]]); |
1390 | fprintf (stream: f, format: "\nClass translation:\n" ); |
1391 | for (i = 0; i < N_REG_CLASSES; i++) |
1392 | fprintf (stream: f, format: " %s -> %s\n" , reg_class_names[i], |
1393 | reg_class_names[class_translate[i]]); |
1394 | } |
1395 | |
1396 | /* Output all possible allocno and translation classes and the |
1397 | translation maps into stderr. */ |
1398 | void |
1399 | ira_debug_allocno_classes (void) |
1400 | { |
1401 | print_uniform_and_important_classes (stderr); |
1402 | print_translated_classes (stderr, pressure_p: false); |
1403 | print_translated_classes (stderr, pressure_p: true); |
1404 | } |
1405 | |
1406 | /* Set up different arrays concerning class subsets, allocno and |
1407 | important classes. */ |
1408 | static void |
1409 | find_reg_classes (void) |
1410 | { |
1411 | setup_allocno_and_important_classes (); |
1412 | setup_class_translate (); |
1413 | reorder_important_classes (); |
1414 | setup_reg_class_relations (); |
1415 | } |
1416 | |
1417 | |
1418 | |
1419 | /* Set up the array above. */ |
1420 | static void |
1421 | setup_hard_regno_aclass (void) |
1422 | { |
1423 | int i; |
1424 | |
1425 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
1426 | { |
1427 | #if 1 |
1428 | ira_hard_regno_allocno_class[i] |
1429 | = (TEST_HARD_REG_BIT (no_unit_alloc_regs, bit: i) |
1430 | ? NO_REGS |
1431 | : ira_allocno_class_translate[REGNO_REG_CLASS (i)]); |
1432 | #else |
1433 | int j; |
1434 | enum reg_class cl; |
1435 | ira_hard_regno_allocno_class[i] = NO_REGS; |
1436 | for (j = 0; j < ira_allocno_classes_num; j++) |
1437 | { |
1438 | cl = ira_allocno_classes[j]; |
1439 | if (ira_class_hard_reg_index[cl][i] >= 0) |
1440 | { |
1441 | ira_hard_regno_allocno_class[i] = cl; |
1442 | break; |
1443 | } |
1444 | } |
1445 | #endif |
1446 | } |
1447 | } |
1448 | |
1449 | |
1450 | |
1451 | /* Form IRA_REG_CLASS_MAX_NREGS and IRA_REG_CLASS_MIN_NREGS maps. */ |
1452 | static void |
1453 | setup_reg_class_nregs (void) |
1454 | { |
1455 | int i, cl, cl2, m; |
1456 | |
1457 | for (m = 0; m < MAX_MACHINE_MODE; m++) |
1458 | { |
1459 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1460 | ira_reg_class_max_nregs[cl][m] |
1461 | = ira_reg_class_min_nregs[cl][m] |
1462 | = targetm.class_max_nregs ((reg_class_t) cl, (machine_mode) m); |
1463 | for (cl = 0; cl < N_REG_CLASSES; cl++) |
1464 | for (i = 0; |
1465 | (cl2 = alloc_reg_class_subclasses[cl][i]) != LIM_REG_CLASSES; |
1466 | i++) |
1467 | if (ira_reg_class_min_nregs[cl2][m] |
1468 | < ira_reg_class_min_nregs[cl][m]) |
1469 | ira_reg_class_min_nregs[cl][m] = ira_reg_class_min_nregs[cl2][m]; |
1470 | } |
1471 | } |
1472 | |
1473 | |
1474 | |
1475 | /* Set up IRA_PROHIBITED_CLASS_MODE_REGS, IRA_EXCLUDE_CLASS_MODE_REGS, and |
1476 | IRA_CLASS_SINGLETON. This function is called once IRA_CLASS_HARD_REGS has |
1477 | been initialized. */ |
1478 | static void |
1479 | setup_prohibited_and_exclude_class_mode_regs (void) |
1480 | { |
1481 | int j, k, hard_regno, cl, last_hard_regno, count; |
1482 | |
1483 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
1484 | { |
1485 | temp_hard_regset = reg_class_contents[cl] & ~no_unit_alloc_regs; |
1486 | for (j = 0; j < NUM_MACHINE_MODES; j++) |
1487 | { |
1488 | count = 0; |
1489 | last_hard_regno = -1; |
1490 | CLEAR_HARD_REG_SET (ira_prohibited_class_mode_regs[cl][j]); |
1491 | CLEAR_HARD_REG_SET (ira_exclude_class_mode_regs[cl][j]); |
1492 | for (k = ira_class_hard_regs_num[cl] - 1; k >= 0; k--) |
1493 | { |
1494 | hard_regno = ira_class_hard_regs[cl][k]; |
1495 | if (!targetm.hard_regno_mode_ok (hard_regno, (machine_mode) j)) |
1496 | SET_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j], |
1497 | bit: hard_regno); |
1498 | else if (in_hard_reg_set_p (regs: temp_hard_regset, |
1499 | mode: (machine_mode) j, regno: hard_regno)) |
1500 | { |
1501 | last_hard_regno = hard_regno; |
1502 | count++; |
1503 | } |
1504 | else |
1505 | { |
1506 | SET_HARD_REG_BIT (ira_exclude_class_mode_regs[cl][j], bit: hard_regno); |
1507 | } |
1508 | } |
1509 | ira_class_singleton[cl][j] = (count == 1 ? last_hard_regno : -1); |
1510 | } |
1511 | } |
1512 | } |
1513 | |
1514 | /* Clarify IRA_PROHIBITED_CLASS_MODE_REGS by excluding hard registers |
1515 | spanning from one register pressure class to another one. It is |
1516 | called after defining the pressure classes. */ |
1517 | static void |
1518 | clarify_prohibited_class_mode_regs (void) |
1519 | { |
1520 | int j, k, hard_regno, cl, pclass, nregs; |
1521 | |
1522 | for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--) |
1523 | for (j = 0; j < NUM_MACHINE_MODES; j++) |
1524 | { |
1525 | CLEAR_HARD_REG_SET (ira_useful_class_mode_regs[cl][j]); |
1526 | for (k = ira_class_hard_regs_num[cl] - 1; k >= 0; k--) |
1527 | { |
1528 | hard_regno = ira_class_hard_regs[cl][k]; |
1529 | if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j], bit: hard_regno)) |
1530 | continue; |
1531 | nregs = hard_regno_nregs (regno: hard_regno, mode: (machine_mode) j); |
1532 | if (hard_regno + nregs > FIRST_PSEUDO_REGISTER) |
1533 | { |
1534 | SET_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j], |
1535 | bit: hard_regno); |
1536 | continue; |
1537 | } |
1538 | pclass = ira_pressure_class_translate[REGNO_REG_CLASS (hard_regno)]; |
1539 | for (nregs-- ;nregs >= 0; nregs--) |
1540 | if (((enum reg_class) pclass |
1541 | != ira_pressure_class_translate[REGNO_REG_CLASS |
1542 | (hard_regno + nregs)])) |
1543 | { |
1544 | SET_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j], |
1545 | bit: hard_regno); |
1546 | break; |
1547 | } |
1548 | if (!TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j], |
1549 | bit: hard_regno)) |
1550 | add_to_hard_reg_set (regs: &ira_useful_class_mode_regs[cl][j], |
1551 | mode: (machine_mode) j, regno: hard_regno); |
1552 | } |
1553 | } |
1554 | } |
1555 | |
1556 | /* Allocate and initialize IRA_REGISTER_MOVE_COST, IRA_MAY_MOVE_IN_COST |
1557 | and IRA_MAY_MOVE_OUT_COST for MODE. */ |
1558 | void |
1559 | ira_init_register_move_cost (machine_mode mode) |
1560 | { |
1561 | static unsigned short last_move_cost[N_REG_CLASSES][N_REG_CLASSES]; |
1562 | bool all_match = true; |
1563 | unsigned int i, cl1, cl2; |
1564 | HARD_REG_SET ok_regs; |
1565 | |
1566 | ira_assert (ira_register_move_cost[mode] == NULL |
1567 | && ira_may_move_in_cost[mode] == NULL |
1568 | && ira_may_move_out_cost[mode] == NULL); |
1569 | CLEAR_HARD_REG_SET (set&: ok_regs); |
1570 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
1571 | if (targetm.hard_regno_mode_ok (i, mode)) |
1572 | SET_HARD_REG_BIT (set&: ok_regs, bit: i); |
1573 | |
1574 | /* Note that we might be asked about the move costs of modes that |
1575 | cannot be stored in any hard register, for example if an inline |
1576 | asm tries to create a register operand with an impossible mode. |
1577 | We therefore can't assert have_regs_of_mode[mode] here. */ |
1578 | for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++) |
1579 | for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++) |
1580 | { |
1581 | int cost; |
1582 | if (!hard_reg_set_intersect_p (x: ok_regs, reg_class_contents[cl1]) |
1583 | || !hard_reg_set_intersect_p (x: ok_regs, reg_class_contents[cl2])) |
1584 | { |
1585 | if ((ira_reg_class_max_nregs[cl1][mode] |
1586 | > ira_class_hard_regs_num[cl1]) |
1587 | || (ira_reg_class_max_nregs[cl2][mode] |
1588 | > ira_class_hard_regs_num[cl2])) |
1589 | cost = 65535; |
1590 | else |
1591 | cost = (ira_memory_move_cost[mode][cl1][0] |
1592 | + ira_memory_move_cost[mode][cl2][1]) * 2; |
1593 | } |
1594 | else |
1595 | { |
1596 | cost = register_move_cost (mode, (enum reg_class) cl1, |
1597 | (enum reg_class) cl2); |
1598 | ira_assert (cost < 65535); |
1599 | } |
1600 | all_match &= (last_move_cost[cl1][cl2] == cost); |
1601 | last_move_cost[cl1][cl2] = cost; |
1602 | } |
1603 | if (all_match && last_mode_for_init_move_cost != -1) |
1604 | { |
1605 | ira_register_move_cost[mode] |
1606 | = ira_register_move_cost[last_mode_for_init_move_cost]; |
1607 | ira_may_move_in_cost[mode] |
1608 | = ira_may_move_in_cost[last_mode_for_init_move_cost]; |
1609 | ira_may_move_out_cost[mode] |
1610 | = ira_may_move_out_cost[last_mode_for_init_move_cost]; |
1611 | return; |
1612 | } |
1613 | last_mode_for_init_move_cost = mode; |
1614 | ira_register_move_cost[mode] = XNEWVEC (move_table, N_REG_CLASSES); |
1615 | ira_may_move_in_cost[mode] = XNEWVEC (move_table, N_REG_CLASSES); |
1616 | ira_may_move_out_cost[mode] = XNEWVEC (move_table, N_REG_CLASSES); |
1617 | for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++) |
1618 | for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++) |
1619 | { |
1620 | int cost; |
1621 | enum reg_class *p1, *p2; |
1622 | |
1623 | if (last_move_cost[cl1][cl2] == 65535) |
1624 | { |
1625 | ira_register_move_cost[mode][cl1][cl2] = 65535; |
1626 | ira_may_move_in_cost[mode][cl1][cl2] = 65535; |
1627 | ira_may_move_out_cost[mode][cl1][cl2] = 65535; |
1628 | } |
1629 | else |
1630 | { |
1631 | cost = last_move_cost[cl1][cl2]; |
1632 | |
1633 | for (p2 = ®_class_subclasses[cl2][0]; |
1634 | *p2 != LIM_REG_CLASSES; p2++) |
1635 | if (ira_class_hard_regs_num[*p2] > 0 |
1636 | && (ira_reg_class_max_nregs[*p2][mode] |
1637 | <= ira_class_hard_regs_num[*p2])) |
1638 | cost = MAX (cost, ira_register_move_cost[mode][cl1][*p2]); |
1639 | |
1640 | for (p1 = ®_class_subclasses[cl1][0]; |
1641 | *p1 != LIM_REG_CLASSES; p1++) |
1642 | if (ira_class_hard_regs_num[*p1] > 0 |
1643 | && (ira_reg_class_max_nregs[*p1][mode] |
1644 | <= ira_class_hard_regs_num[*p1])) |
1645 | cost = MAX (cost, ira_register_move_cost[mode][*p1][cl2]); |
1646 | |
1647 | ira_assert (cost <= 65535); |
1648 | ira_register_move_cost[mode][cl1][cl2] = cost; |
1649 | |
1650 | if (ira_class_subset_p[cl1][cl2]) |
1651 | ira_may_move_in_cost[mode][cl1][cl2] = 0; |
1652 | else |
1653 | ira_may_move_in_cost[mode][cl1][cl2] = cost; |
1654 | |
1655 | if (ira_class_subset_p[cl2][cl1]) |
1656 | ira_may_move_out_cost[mode][cl1][cl2] = 0; |
1657 | else |
1658 | ira_may_move_out_cost[mode][cl1][cl2] = cost; |
1659 | } |
1660 | } |
1661 | } |
1662 | |
1663 | |
1664 | |
1665 | /* This is called once during compiler work. It sets up |
1666 | different arrays whose values don't depend on the compiled |
1667 | function. */ |
1668 | void |
1669 | ira_init_once (void) |
1670 | { |
1671 | ira_init_costs_once (); |
1672 | lra_init_once (); |
1673 | |
1674 | ira_use_lra_p = targetm.lra_p (); |
1675 | } |
1676 | |
1677 | /* Free ira_max_register_move_cost, ira_may_move_in_cost and |
1678 | ira_may_move_out_cost for each mode. */ |
1679 | void |
1680 | target_ira_int::free_register_move_costs (void) |
1681 | { |
1682 | int mode, i; |
1683 | |
1684 | /* Reset move_cost and friends, making sure we only free shared |
1685 | table entries once. */ |
1686 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
1687 | if (x_ira_register_move_cost[mode]) |
1688 | { |
1689 | for (i = 0; |
1690 | i < mode && (x_ira_register_move_cost[i] |
1691 | != x_ira_register_move_cost[mode]); |
1692 | i++) |
1693 | ; |
1694 | if (i == mode) |
1695 | { |
1696 | free (ptr: x_ira_register_move_cost[mode]); |
1697 | free (ptr: x_ira_may_move_in_cost[mode]); |
1698 | free (ptr: x_ira_may_move_out_cost[mode]); |
1699 | } |
1700 | } |
1701 | memset (s: x_ira_register_move_cost, c: 0, n: sizeof x_ira_register_move_cost); |
1702 | memset (s: x_ira_may_move_in_cost, c: 0, n: sizeof x_ira_may_move_in_cost); |
1703 | memset (s: x_ira_may_move_out_cost, c: 0, n: sizeof x_ira_may_move_out_cost); |
1704 | last_mode_for_init_move_cost = -1; |
1705 | } |
1706 | |
1707 | target_ira_int::~target_ira_int () |
1708 | { |
1709 | free_ira_costs (); |
1710 | free_register_move_costs (); |
1711 | } |
1712 | |
1713 | /* This is called every time when register related information is |
1714 | changed. */ |
1715 | void |
1716 | ira_init (void) |
1717 | { |
1718 | this_target_ira_int->free_register_move_costs (); |
1719 | setup_reg_mode_hard_regset (); |
1720 | setup_alloc_regs (flag_omit_frame_pointer != 0); |
1721 | setup_class_subset_and_memory_move_costs (); |
1722 | setup_reg_class_nregs (); |
1723 | setup_prohibited_and_exclude_class_mode_regs (); |
1724 | find_reg_classes (); |
1725 | clarify_prohibited_class_mode_regs (); |
1726 | setup_hard_regno_aclass (); |
1727 | ira_init_costs (); |
1728 | } |
1729 | |
1730 | |
1731 | #define ira_prohibited_mode_move_regs_initialized_p \ |
1732 | (this_target_ira_int->x_ira_prohibited_mode_move_regs_initialized_p) |
1733 | |
1734 | /* Set up IRA_PROHIBITED_MODE_MOVE_REGS. */ |
1735 | static void |
1736 | setup_prohibited_mode_move_regs (void) |
1737 | { |
1738 | int i, j; |
1739 | rtx test_reg1, test_reg2, move_pat; |
1740 | rtx_insn *move_insn; |
1741 | |
1742 | if (ira_prohibited_mode_move_regs_initialized_p) |
1743 | return; |
1744 | ira_prohibited_mode_move_regs_initialized_p = true; |
1745 | test_reg1 = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1); |
1746 | test_reg2 = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 2); |
1747 | move_pat = gen_rtx_SET (test_reg1, test_reg2); |
1748 | move_insn = gen_rtx_INSN (VOIDmode, prev_insn: 0, next_insn: 0, bb: 0, pattern: move_pat, location: 0, code: -1, reg_notes: 0); |
1749 | for (i = 0; i < NUM_MACHINE_MODES; i++) |
1750 | { |
1751 | SET_HARD_REG_SET (ira_prohibited_mode_move_regs[i]); |
1752 | for (j = 0; j < FIRST_PSEUDO_REGISTER; j++) |
1753 | { |
1754 | if (!targetm.hard_regno_mode_ok (j, (machine_mode) i)) |
1755 | continue; |
1756 | set_mode_and_regno (test_reg1, (machine_mode) i, j); |
1757 | set_mode_and_regno (test_reg2, (machine_mode) i, j); |
1758 | INSN_CODE (move_insn) = -1; |
1759 | recog_memoized (insn: move_insn); |
1760 | if (INSN_CODE (move_insn) < 0) |
1761 | continue; |
1762 | extract_insn (move_insn); |
1763 | /* We don't know whether the move will be in code that is optimized |
1764 | for size or speed, so consider all enabled alternatives. */ |
1765 | if (! constrain_operands (1, get_enabled_alternatives (move_insn))) |
1766 | continue; |
1767 | CLEAR_HARD_REG_BIT (ira_prohibited_mode_move_regs[i], bit: j); |
1768 | } |
1769 | } |
1770 | } |
1771 | |
1772 | |
1773 | |
1774 | /* Extract INSN and return the set of alternatives that we should consider. |
1775 | This excludes any alternatives whose constraints are obviously impossible |
1776 | to meet (e.g. because the constraint requires a constant and the operand |
1777 | is nonconstant). It also excludes alternatives that are bound to need |
1778 | a spill or reload, as long as we have other alternatives that match |
1779 | exactly. */ |
1780 | alternative_mask |
1781 | ira_setup_alts (rtx_insn *insn) |
1782 | { |
1783 | int nop, nalt; |
1784 | bool curr_swapped; |
1785 | const char *p; |
1786 | int commutative = -1; |
1787 | |
1788 | extract_insn (insn); |
1789 | preprocess_constraints (insn); |
1790 | alternative_mask preferred = get_preferred_alternatives (insn); |
1791 | alternative_mask alts = 0; |
1792 | alternative_mask exact_alts = 0; |
1793 | /* Check that the hard reg set is enough for holding all |
1794 | alternatives. It is hard to imagine the situation when the |
1795 | assertion is wrong. */ |
1796 | ira_assert (recog_data.n_alternatives |
1797 | <= (int) MAX (sizeof (HARD_REG_ELT_TYPE) * CHAR_BIT, |
1798 | FIRST_PSEUDO_REGISTER)); |
1799 | for (nop = 0; nop < recog_data.n_operands; nop++) |
1800 | if (recog_data.constraints[nop][0] == '%') |
1801 | { |
1802 | commutative = nop; |
1803 | break; |
1804 | } |
1805 | for (curr_swapped = false;; curr_swapped = true) |
1806 | { |
1807 | for (nalt = 0; nalt < recog_data.n_alternatives; nalt++) |
1808 | { |
1809 | if (!TEST_BIT (preferred, nalt) || TEST_BIT (exact_alts, nalt)) |
1810 | continue; |
1811 | |
1812 | const operand_alternative *op_alt |
1813 | = &recog_op_alt[nalt * recog_data.n_operands]; |
1814 | int this_reject = 0; |
1815 | for (nop = 0; nop < recog_data.n_operands; nop++) |
1816 | { |
1817 | int c, len; |
1818 | |
1819 | this_reject += op_alt[nop].reject; |
1820 | |
1821 | rtx op = recog_data.operand[nop]; |
1822 | p = op_alt[nop].constraint; |
1823 | if (*p == 0 || *p == ',') |
1824 | continue; |
1825 | |
1826 | bool win_p = false; |
1827 | do |
1828 | switch (c = *p, len = CONSTRAINT_LEN (c, p), c) |
1829 | { |
1830 | case '#': |
1831 | case ',': |
1832 | c = '\0'; |
1833 | /* FALLTHRU */ |
1834 | case '\0': |
1835 | len = 0; |
1836 | break; |
1837 | |
1838 | case '%': |
1839 | /* The commutative modifier is handled above. */ |
1840 | break; |
1841 | |
1842 | case '0': case '1': case '2': case '3': case '4': |
1843 | case '5': case '6': case '7': case '8': case '9': |
1844 | { |
1845 | char *end; |
1846 | unsigned long dup = strtoul (nptr: p, endptr: &end, base: 10); |
1847 | rtx other = recog_data.operand[dup]; |
1848 | len = end - p; |
1849 | if (MEM_P (other) |
1850 | ? rtx_equal_p (other, op) |
1851 | : REG_P (op) || SUBREG_P (op)) |
1852 | goto op_success; |
1853 | win_p = true; |
1854 | } |
1855 | break; |
1856 | |
1857 | case 'g': |
1858 | goto op_success; |
1859 | break; |
1860 | |
1861 | default: |
1862 | { |
1863 | enum constraint_num cn = lookup_constraint (p); |
1864 | rtx mem = NULL; |
1865 | switch (get_constraint_type (c: cn)) |
1866 | { |
1867 | case CT_REGISTER: |
1868 | if (reg_class_for_constraint (c: cn) != NO_REGS) |
1869 | { |
1870 | if (REG_P (op) || SUBREG_P (op)) |
1871 | goto op_success; |
1872 | win_p = true; |
1873 | } |
1874 | break; |
1875 | |
1876 | case CT_CONST_INT: |
1877 | if (CONST_INT_P (op) |
1878 | && (insn_const_int_ok_for_constraint |
1879 | (INTVAL (op), cn))) |
1880 | goto op_success; |
1881 | break; |
1882 | |
1883 | case CT_ADDRESS: |
1884 | goto op_success; |
1885 | |
1886 | case CT_MEMORY: |
1887 | case CT_RELAXED_MEMORY: |
1888 | mem = op; |
1889 | /* Fall through. */ |
1890 | case CT_SPECIAL_MEMORY: |
1891 | if (!mem) |
1892 | mem = extract_mem_from_operand (op); |
1893 | if (MEM_P (mem)) |
1894 | goto op_success; |
1895 | win_p = true; |
1896 | break; |
1897 | |
1898 | case CT_FIXED_FORM: |
1899 | if (constraint_satisfied_p (x: op, c: cn)) |
1900 | goto op_success; |
1901 | break; |
1902 | } |
1903 | break; |
1904 | } |
1905 | } |
1906 | while (p += len, c); |
1907 | if (!win_p) |
1908 | break; |
1909 | /* We can make the alternative match by spilling a register |
1910 | to memory or loading something into a register. Count a |
1911 | cost of one reload (the equivalent of the '?' constraint). */ |
1912 | this_reject += 6; |
1913 | op_success: |
1914 | ; |
1915 | } |
1916 | |
1917 | if (nop >= recog_data.n_operands) |
1918 | { |
1919 | alts |= ALTERNATIVE_BIT (nalt); |
1920 | if (this_reject == 0) |
1921 | exact_alts |= ALTERNATIVE_BIT (nalt); |
1922 | } |
1923 | } |
1924 | if (commutative < 0) |
1925 | break; |
1926 | /* Swap forth and back to avoid changing recog_data. */ |
1927 | std::swap (a&: recog_data.operand[commutative], |
1928 | b&: recog_data.operand[commutative + 1]); |
1929 | if (curr_swapped) |
1930 | break; |
1931 | } |
1932 | return exact_alts ? exact_alts : alts; |
1933 | } |
1934 | |
1935 | /* Return the number of the output non-early clobber operand which |
1936 | should be the same in any case as operand with number OP_NUM (or |
1937 | negative value if there is no such operand). ALTS is the mask |
1938 | of alternatives that we should consider. SINGLE_INPUT_OP_HAS_CSTR_P |
1939 | should be set in this function, it indicates whether there is only |
1940 | a single input operand which has the matching constraint on the |
1941 | output operand at the position specified in return value. If the |
1942 | pattern allows any one of several input operands holds the matching |
1943 | constraint, it's set as false, one typical case is destructive FMA |
1944 | instruction on target rs6000. Note that for a non-NO_REG preferred |
1945 | register class with no free register move copy, if the parameter |
1946 | PARAM_IRA_CONSIDER_DUP_IN_ALL_ALTS is set to one, this function |
1947 | will check all available alternatives for matching constraints, |
1948 | even if it has found or will find one alternative with non-NO_REG |
1949 | regclass, it can respect more cases with matching constraints. If |
1950 | PARAM_IRA_CONSIDER_DUP_IN_ALL_ALTS is set to zero, |
1951 | SINGLE_INPUT_OP_HAS_CSTR_P is always true, it will stop to find |
1952 | matching constraint relationship once it hits some alternative with |
1953 | some non-NO_REG regclass. */ |
1954 | int |
1955 | ira_get_dup_out_num (int op_num, alternative_mask alts, |
1956 | bool &single_input_op_has_cstr_p) |
1957 | { |
1958 | int curr_alt, c, original; |
1959 | bool ignore_p, use_commut_op_p; |
1960 | const char *str; |
1961 | |
1962 | if (op_num < 0 || recog_data.n_alternatives == 0) |
1963 | return -1; |
1964 | /* We should find duplications only for input operands. */ |
1965 | if (recog_data.operand_type[op_num] != OP_IN) |
1966 | return -1; |
1967 | str = recog_data.constraints[op_num]; |
1968 | use_commut_op_p = false; |
1969 | single_input_op_has_cstr_p = true; |
1970 | |
1971 | rtx op = recog_data.operand[op_num]; |
1972 | int op_regno = reg_or_subregno (op); |
1973 | enum reg_class op_pref_cl = reg_preferred_class (op_regno); |
1974 | machine_mode op_mode = GET_MODE (op); |
1975 | |
1976 | ira_init_register_move_cost_if_necessary (mode: op_mode); |
1977 | /* If the preferred regclass isn't NO_REG, continue to find the matching |
1978 | constraint in all available alternatives with preferred regclass, even |
1979 | if we have found or will find one alternative whose constraint stands |
1980 | for a REG (non-NO_REG) regclass. Note that it would be fine not to |
1981 | respect matching constraint if the register copy is free, so exclude |
1982 | it. */ |
1983 | bool respect_dup_despite_reg_cstr |
1984 | = param_ira_consider_dup_in_all_alts |
1985 | && op_pref_cl != NO_REGS |
1986 | && ira_register_move_cost[op_mode][op_pref_cl][op_pref_cl] > 0; |
1987 | |
1988 | /* Record the alternative whose constraint uses the same regclass as the |
1989 | preferred regclass, later if we find one matching constraint for this |
1990 | operand with preferred reclass, we will visit these recorded |
1991 | alternatives to check whether if there is one alternative in which no |
1992 | any INPUT operands have one matching constraint same as our candidate. |
1993 | If yes, it means there is one alternative which is perfectly fine |
1994 | without satisfying this matching constraint. If no, it means in any |
1995 | alternatives there is one other INPUT operand holding this matching |
1996 | constraint, it's fine to respect this matching constraint and further |
1997 | create this constraint copy since it would become harmless once some |
1998 | other takes preference and it's interfered. */ |
1999 | alternative_mask pref_cl_alts; |
2000 | |
2001 | for (;;) |
2002 | { |
2003 | pref_cl_alts = 0; |
2004 | |
2005 | for (curr_alt = 0, ignore_p = !TEST_BIT (alts, curr_alt), |
2006 | original = -1;;) |
2007 | { |
2008 | c = *str; |
2009 | if (c == '\0') |
2010 | break; |
2011 | if (c == '#') |
2012 | ignore_p = true; |
2013 | else if (c == ',') |
2014 | { |
2015 | curr_alt++; |
2016 | ignore_p = !TEST_BIT (alts, curr_alt); |
2017 | } |
2018 | else if (! ignore_p) |
2019 | switch (c) |
2020 | { |
2021 | case 'g': |
2022 | goto fail; |
2023 | default: |
2024 | { |
2025 | enum constraint_num cn = lookup_constraint (p: str); |
2026 | enum reg_class cl = reg_class_for_constraint (c: cn); |
2027 | if (cl != NO_REGS && !targetm.class_likely_spilled_p (cl)) |
2028 | { |
2029 | if (respect_dup_despite_reg_cstr) |
2030 | { |
2031 | /* If it's free to move from one preferred class to |
2032 | the one without matching constraint, it doesn't |
2033 | have to respect this constraint with costs. */ |
2034 | if (cl != op_pref_cl |
2035 | && (ira_reg_class_intersect[cl][op_pref_cl] |
2036 | != NO_REGS) |
2037 | && (ira_may_move_in_cost[op_mode][op_pref_cl][cl] |
2038 | == 0)) |
2039 | goto fail; |
2040 | else if (cl == op_pref_cl) |
2041 | pref_cl_alts |= ALTERNATIVE_BIT (curr_alt); |
2042 | } |
2043 | else |
2044 | goto fail; |
2045 | } |
2046 | if (constraint_satisfied_p (x: op, c: cn)) |
2047 | goto fail; |
2048 | break; |
2049 | } |
2050 | |
2051 | case '0': case '1': case '2': case '3': case '4': |
2052 | case '5': case '6': case '7': case '8': case '9': |
2053 | { |
2054 | char *end; |
2055 | int n = (int) strtoul (nptr: str, endptr: &end, base: 10); |
2056 | str = end; |
2057 | if (original != -1 && original != n) |
2058 | goto fail; |
2059 | gcc_assert (n < recog_data.n_operands); |
2060 | if (respect_dup_despite_reg_cstr) |
2061 | { |
2062 | const operand_alternative *op_alt |
2063 | = &recog_op_alt[curr_alt * recog_data.n_operands]; |
2064 | /* Only respect the one with preferred rclass, without |
2065 | respect_dup_despite_reg_cstr it's possible to get |
2066 | one whose regclass isn't preferred first before, |
2067 | but it would fail since there should be other |
2068 | alternatives with preferred regclass. */ |
2069 | if (op_alt[n].cl == op_pref_cl) |
2070 | original = n; |
2071 | } |
2072 | else |
2073 | original = n; |
2074 | continue; |
2075 | } |
2076 | } |
2077 | str += CONSTRAINT_LEN (c, str); |
2078 | } |
2079 | if (original == -1) |
2080 | goto fail; |
2081 | if (recog_data.operand_type[original] == OP_OUT) |
2082 | { |
2083 | if (pref_cl_alts == 0) |
2084 | return original; |
2085 | /* Visit these recorded alternatives to check whether |
2086 | there is one alternative in which no any INPUT operands |
2087 | have one matching constraint same as our candidate. |
2088 | Give up this candidate if so. */ |
2089 | int nop, nalt; |
2090 | for (nalt = 0; nalt < recog_data.n_alternatives; nalt++) |
2091 | { |
2092 | if (!TEST_BIT (pref_cl_alts, nalt)) |
2093 | continue; |
2094 | const operand_alternative *op_alt |
2095 | = &recog_op_alt[nalt * recog_data.n_operands]; |
2096 | bool dup_in_other = false; |
2097 | for (nop = 0; nop < recog_data.n_operands; nop++) |
2098 | { |
2099 | if (recog_data.operand_type[nop] != OP_IN) |
2100 | continue; |
2101 | if (nop == op_num) |
2102 | continue; |
2103 | if (op_alt[nop].matches == original) |
2104 | { |
2105 | dup_in_other = true; |
2106 | break; |
2107 | } |
2108 | } |
2109 | if (!dup_in_other) |
2110 | return -1; |
2111 | } |
2112 | single_input_op_has_cstr_p = false; |
2113 | return original; |
2114 | } |
2115 | fail: |
2116 | if (use_commut_op_p) |
2117 | break; |
2118 | use_commut_op_p = true; |
2119 | if (recog_data.constraints[op_num][0] == '%') |
2120 | str = recog_data.constraints[op_num + 1]; |
2121 | else if (op_num > 0 && recog_data.constraints[op_num - 1][0] == '%') |
2122 | str = recog_data.constraints[op_num - 1]; |
2123 | else |
2124 | break; |
2125 | } |
2126 | return -1; |
2127 | } |
2128 | |
2129 | |
2130 | |
2131 | /* Search forward to see if the source register of a copy insn dies |
2132 | before either it or the destination register is modified, but don't |
2133 | scan past the end of the basic block. If so, we can replace the |
2134 | source with the destination and let the source die in the copy |
2135 | insn. |
2136 | |
2137 | This will reduce the number of registers live in that range and may |
2138 | enable the destination and the source coalescing, thus often saving |
2139 | one register in addition to a register-register copy. */ |
2140 | |
2141 | static void |
2142 | decrease_live_ranges_number (void) |
2143 | { |
2144 | basic_block bb; |
2145 | rtx_insn *insn; |
2146 | rtx set, src, dest, dest_death, note; |
2147 | rtx_insn *p, *q; |
2148 | int sregno, dregno; |
2149 | |
2150 | if (! flag_expensive_optimizations) |
2151 | return; |
2152 | |
2153 | if (ira_dump_file) |
2154 | fprintf (stream: ira_dump_file, format: "Starting decreasing number of live ranges...\n" ); |
2155 | |
2156 | FOR_EACH_BB_FN (bb, cfun) |
2157 | FOR_BB_INSNS (bb, insn) |
2158 | { |
2159 | set = single_set (insn); |
2160 | if (! set) |
2161 | continue; |
2162 | src = SET_SRC (set); |
2163 | dest = SET_DEST (set); |
2164 | if (! REG_P (src) || ! REG_P (dest) |
2165 | || find_reg_note (insn, REG_DEAD, src)) |
2166 | continue; |
2167 | sregno = REGNO (src); |
2168 | dregno = REGNO (dest); |
2169 | |
2170 | /* We don't want to mess with hard regs if register classes |
2171 | are small. */ |
2172 | if (sregno == dregno |
2173 | || (targetm.small_register_classes_for_mode_p (GET_MODE (src)) |
2174 | && (sregno < FIRST_PSEUDO_REGISTER |
2175 | || dregno < FIRST_PSEUDO_REGISTER)) |
2176 | /* We don't see all updates to SP if they are in an |
2177 | auto-inc memory reference, so we must disallow this |
2178 | optimization on them. */ |
2179 | || sregno == STACK_POINTER_REGNUM |
2180 | || dregno == STACK_POINTER_REGNUM) |
2181 | continue; |
2182 | |
2183 | dest_death = NULL_RTX; |
2184 | |
2185 | for (p = NEXT_INSN (insn); p; p = NEXT_INSN (insn: p)) |
2186 | { |
2187 | if (! INSN_P (p)) |
2188 | continue; |
2189 | if (BLOCK_FOR_INSN (insn: p) != bb) |
2190 | break; |
2191 | |
2192 | if (reg_set_p (src, p) || reg_set_p (dest, p) |
2193 | /* If SRC is an asm-declared register, it must not be |
2194 | replaced in any asm. Unfortunately, the REG_EXPR |
2195 | tree for the asm variable may be absent in the SRC |
2196 | rtx, so we can't check the actual register |
2197 | declaration easily (the asm operand will have it, |
2198 | though). To avoid complicating the test for a rare |
2199 | case, we just don't perform register replacement |
2200 | for a hard reg mentioned in an asm. */ |
2201 | || (sregno < FIRST_PSEUDO_REGISTER |
2202 | && asm_noperands (PATTERN (insn: p)) >= 0 |
2203 | && reg_overlap_mentioned_p (src, PATTERN (insn: p))) |
2204 | /* Don't change hard registers used by a call. */ |
2205 | || (CALL_P (p) && sregno < FIRST_PSEUDO_REGISTER |
2206 | && find_reg_fusage (p, USE, src)) |
2207 | /* Don't change a USE of a register. */ |
2208 | || (GET_CODE (PATTERN (p)) == USE |
2209 | && reg_overlap_mentioned_p (src, XEXP (PATTERN (p), 0)))) |
2210 | break; |
2211 | |
2212 | /* See if all of SRC dies in P. This test is slightly |
2213 | more conservative than it needs to be. */ |
2214 | if ((note = find_regno_note (p, REG_DEAD, sregno)) |
2215 | && GET_MODE (XEXP (note, 0)) == GET_MODE (src)) |
2216 | { |
2217 | int failed = 0; |
2218 | |
2219 | /* We can do the optimization. Scan forward from INSN |
2220 | again, replacing regs as we go. Set FAILED if a |
2221 | replacement can't be done. In that case, we can't |
2222 | move the death note for SRC. This should be |
2223 | rare. */ |
2224 | |
2225 | /* Set to stop at next insn. */ |
2226 | for (q = next_real_insn (insn); |
2227 | q != next_real_insn (p); |
2228 | q = next_real_insn (q)) |
2229 | { |
2230 | if (reg_overlap_mentioned_p (src, PATTERN (insn: q))) |
2231 | { |
2232 | /* If SRC is a hard register, we might miss |
2233 | some overlapping registers with |
2234 | validate_replace_rtx, so we would have to |
2235 | undo it. We can't if DEST is present in |
2236 | the insn, so fail in that combination of |
2237 | cases. */ |
2238 | if (sregno < FIRST_PSEUDO_REGISTER |
2239 | && reg_mentioned_p (dest, PATTERN (insn: q))) |
2240 | failed = 1; |
2241 | |
2242 | /* Attempt to replace all uses. */ |
2243 | else if (!validate_replace_rtx (src, dest, q)) |
2244 | failed = 1; |
2245 | |
2246 | /* If this succeeded, but some part of the |
2247 | register is still present, undo the |
2248 | replacement. */ |
2249 | else if (sregno < FIRST_PSEUDO_REGISTER |
2250 | && reg_overlap_mentioned_p (src, PATTERN (insn: q))) |
2251 | { |
2252 | validate_replace_rtx (dest, src, q); |
2253 | failed = 1; |
2254 | } |
2255 | } |
2256 | |
2257 | /* If DEST dies here, remove the death note and |
2258 | save it for later. Make sure ALL of DEST dies |
2259 | here; again, this is overly conservative. */ |
2260 | if (! dest_death |
2261 | && (dest_death = find_regno_note (q, REG_DEAD, dregno))) |
2262 | { |
2263 | if (GET_MODE (XEXP (dest_death, 0)) == GET_MODE (dest)) |
2264 | remove_note (q, dest_death); |
2265 | else |
2266 | { |
2267 | failed = 1; |
2268 | dest_death = 0; |
2269 | } |
2270 | } |
2271 | } |
2272 | |
2273 | if (! failed) |
2274 | { |
2275 | /* Move death note of SRC from P to INSN. */ |
2276 | remove_note (p, note); |
2277 | XEXP (note, 1) = REG_NOTES (insn); |
2278 | REG_NOTES (insn) = note; |
2279 | } |
2280 | |
2281 | /* DEST is also dead if INSN has a REG_UNUSED note for |
2282 | DEST. */ |
2283 | if (! dest_death |
2284 | && (dest_death |
2285 | = find_regno_note (insn, REG_UNUSED, dregno))) |
2286 | { |
2287 | PUT_REG_NOTE_KIND (dest_death, REG_DEAD); |
2288 | remove_note (insn, dest_death); |
2289 | } |
2290 | |
2291 | /* Put death note of DEST on P if we saw it die. */ |
2292 | if (dest_death) |
2293 | { |
2294 | XEXP (dest_death, 1) = REG_NOTES (p); |
2295 | REG_NOTES (p) = dest_death; |
2296 | } |
2297 | break; |
2298 | } |
2299 | |
2300 | /* If SRC is a hard register which is set or killed in |
2301 | some other way, we can't do this optimization. */ |
2302 | else if (sregno < FIRST_PSEUDO_REGISTER && dead_or_set_p (p, src)) |
2303 | break; |
2304 | } |
2305 | } |
2306 | } |
2307 | |
2308 | |
2309 | |
2310 | /* Return nonzero if REGNO is a particularly bad choice for reloading X. */ |
2311 | static bool |
2312 | ira_bad_reload_regno_1 (int regno, rtx x) |
2313 | { |
2314 | int x_regno, n, i; |
2315 | ira_allocno_t a; |
2316 | enum reg_class pref; |
2317 | |
2318 | /* We only deal with pseudo regs. */ |
2319 | if (! x || GET_CODE (x) != REG) |
2320 | return false; |
2321 | |
2322 | x_regno = REGNO (x); |
2323 | if (x_regno < FIRST_PSEUDO_REGISTER) |
2324 | return false; |
2325 | |
2326 | /* If the pseudo prefers REGNO explicitly, then do not consider |
2327 | REGNO a bad spill choice. */ |
2328 | pref = reg_preferred_class (x_regno); |
2329 | if (reg_class_size[pref] == 1) |
2330 | return !TEST_HARD_REG_BIT (reg_class_contents[pref], bit: regno); |
2331 | |
2332 | /* If the pseudo conflicts with REGNO, then we consider REGNO a |
2333 | poor choice for a reload regno. */ |
2334 | a = ira_regno_allocno_map[x_regno]; |
2335 | n = ALLOCNO_NUM_OBJECTS (a); |
2336 | for (i = 0; i < n; i++) |
2337 | { |
2338 | ira_object_t obj = ALLOCNO_OBJECT (a, i); |
2339 | if (TEST_HARD_REG_BIT (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), bit: regno)) |
2340 | return true; |
2341 | } |
2342 | return false; |
2343 | } |
2344 | |
2345 | /* Return nonzero if REGNO is a particularly bad choice for reloading |
2346 | IN or OUT. */ |
2347 | bool |
2348 | ira_bad_reload_regno (int regno, rtx in, rtx out) |
2349 | { |
2350 | return (ira_bad_reload_regno_1 (regno, x: in) |
2351 | || ira_bad_reload_regno_1 (regno, x: out)); |
2352 | } |
2353 | |
2354 | /* Add register clobbers from asm statements. */ |
2355 | static void |
2356 | compute_regs_asm_clobbered (void) |
2357 | { |
2358 | basic_block bb; |
2359 | |
2360 | FOR_EACH_BB_FN (bb, cfun) |
2361 | { |
2362 | rtx_insn *insn; |
2363 | FOR_BB_INSNS_REVERSE (bb, insn) |
2364 | { |
2365 | df_ref def; |
2366 | |
2367 | if (NONDEBUG_INSN_P (insn) && asm_noperands (PATTERN (insn)) >= 0) |
2368 | FOR_EACH_INSN_DEF (def, insn) |
2369 | { |
2370 | unsigned int dregno = DF_REF_REGNO (def); |
2371 | if (HARD_REGISTER_NUM_P (dregno)) |
2372 | add_to_hard_reg_set (regs: &crtl->asm_clobbers, |
2373 | GET_MODE (DF_REF_REAL_REG (def)), |
2374 | regno: dregno); |
2375 | } |
2376 | } |
2377 | } |
2378 | } |
2379 | |
2380 | |
2381 | /* Set up ELIMINABLE_REGSET, IRA_NO_ALLOC_REGS, and |
2382 | REGS_EVER_LIVE. */ |
2383 | void |
2384 | ira_setup_eliminable_regset (void) |
2385 | { |
2386 | int i; |
2387 | static const struct {const int from, to; } eliminables[] = ELIMINABLE_REGS; |
2388 | int fp_reg_count = hard_regno_nregs (HARD_FRAME_POINTER_REGNUM, Pmode); |
2389 | |
2390 | /* Setup is_leaf as frame_pointer_required may use it. This function |
2391 | is called by sched_init before ira if scheduling is enabled. */ |
2392 | crtl->is_leaf = leaf_function_p (); |
2393 | |
2394 | /* FIXME: If EXIT_IGNORE_STACK is set, we will not save and restore |
2395 | sp for alloca. So we can't eliminate the frame pointer in that |
2396 | case. At some point, we should improve this by emitting the |
2397 | sp-adjusting insns for this case. */ |
2398 | frame_pointer_needed |
2399 | = (! flag_omit_frame_pointer |
2400 | || (cfun->calls_alloca && EXIT_IGNORE_STACK) |
2401 | /* We need the frame pointer to catch stack overflow exceptions if |
2402 | the stack pointer is moving (as for the alloca case just above). */ |
2403 | || (STACK_CHECK_MOVING_SP |
2404 | && flag_stack_check |
2405 | && flag_exceptions |
2406 | && cfun->can_throw_non_call_exceptions) |
2407 | || crtl->accesses_prior_frames |
2408 | || (SUPPORTS_STACK_ALIGNMENT && crtl->stack_realign_needed) |
2409 | || targetm.frame_pointer_required ()); |
2410 | |
2411 | /* The chance that FRAME_POINTER_NEEDED is changed from inspecting |
2412 | RTL is very small. So if we use frame pointer for RA and RTL |
2413 | actually prevents this, we will spill pseudos assigned to the |
2414 | frame pointer in LRA. */ |
2415 | |
2416 | if (frame_pointer_needed) |
2417 | for (i = 0; i < fp_reg_count; i++) |
2418 | df_set_regs_ever_live (HARD_FRAME_POINTER_REGNUM + i, true); |
2419 | |
2420 | ira_no_alloc_regs = no_unit_alloc_regs; |
2421 | CLEAR_HARD_REG_SET (set&: eliminable_regset); |
2422 | |
2423 | compute_regs_asm_clobbered (); |
2424 | |
2425 | /* Build the regset of all eliminable registers and show we can't |
2426 | use those that we already know won't be eliminated. */ |
2427 | for (i = 0; i < (int) ARRAY_SIZE (eliminables); i++) |
2428 | { |
2429 | bool cannot_elim |
2430 | = (! targetm.can_eliminate (eliminables[i].from, eliminables[i].to) |
2431 | || (eliminables[i].to == STACK_POINTER_REGNUM && frame_pointer_needed)); |
2432 | |
2433 | if (!TEST_HARD_REG_BIT (crtl->asm_clobbers, bit: eliminables[i].from)) |
2434 | { |
2435 | SET_HARD_REG_BIT (set&: eliminable_regset, bit: eliminables[i].from); |
2436 | |
2437 | if (cannot_elim) |
2438 | SET_HARD_REG_BIT (ira_no_alloc_regs, bit: eliminables[i].from); |
2439 | } |
2440 | else if (cannot_elim) |
2441 | error ("%s cannot be used in %<asm%> here" , |
2442 | reg_names[eliminables[i].from]); |
2443 | else |
2444 | df_set_regs_ever_live (eliminables[i].from, true); |
2445 | } |
2446 | if (!HARD_FRAME_POINTER_IS_FRAME_POINTER) |
2447 | { |
2448 | for (i = 0; i < fp_reg_count; i++) |
2449 | if (global_regs[HARD_FRAME_POINTER_REGNUM + i]) |
2450 | /* Nothing to do: the register is already treated as live |
2451 | where appropriate, and cannot be eliminated. */ |
2452 | ; |
2453 | else if (!TEST_HARD_REG_BIT (crtl->asm_clobbers, |
2454 | HARD_FRAME_POINTER_REGNUM + i)) |
2455 | { |
2456 | SET_HARD_REG_BIT (set&: eliminable_regset, |
2457 | HARD_FRAME_POINTER_REGNUM + i); |
2458 | if (frame_pointer_needed) |
2459 | SET_HARD_REG_BIT (ira_no_alloc_regs, |
2460 | HARD_FRAME_POINTER_REGNUM + i); |
2461 | } |
2462 | else if (frame_pointer_needed) |
2463 | error ("%s cannot be used in %<asm%> here" , |
2464 | reg_names[HARD_FRAME_POINTER_REGNUM + i]); |
2465 | else |
2466 | df_set_regs_ever_live (HARD_FRAME_POINTER_REGNUM + i, true); |
2467 | } |
2468 | } |
2469 | |
2470 | |
2471 | |
2472 | /* Vector of substitutions of register numbers, |
2473 | used to map pseudo regs into hardware regs. |
2474 | This is set up as a result of register allocation. |
2475 | Element N is the hard reg assigned to pseudo reg N, |
2476 | or is -1 if no hard reg was assigned. |
2477 | If N is a hard reg number, element N is N. */ |
2478 | short *reg_renumber; |
2479 | |
2480 | /* Set up REG_RENUMBER and CALLER_SAVE_NEEDED (used by reload) from |
2481 | the allocation found by IRA. */ |
2482 | static void |
2483 | setup_reg_renumber (void) |
2484 | { |
2485 | int regno, hard_regno; |
2486 | ira_allocno_t a; |
2487 | ira_allocno_iterator ai; |
2488 | |
2489 | caller_save_needed = 0; |
2490 | FOR_EACH_ALLOCNO (a, ai) |
2491 | { |
2492 | if (ira_use_lra_p && ALLOCNO_CAP_MEMBER (a) != NULL) |
2493 | continue; |
2494 | /* There are no caps at this point. */ |
2495 | ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL); |
2496 | if (! ALLOCNO_ASSIGNED_P (a)) |
2497 | /* It can happen if A is not referenced but partially anticipated |
2498 | somewhere in a region. */ |
2499 | ALLOCNO_ASSIGNED_P (a) = true; |
2500 | ira_free_allocno_updated_costs (a); |
2501 | hard_regno = ALLOCNO_HARD_REGNO (a); |
2502 | regno = ALLOCNO_REGNO (a); |
2503 | reg_renumber[regno] = (hard_regno < 0 ? -1 : hard_regno); |
2504 | if (hard_regno >= 0) |
2505 | { |
2506 | int i, nwords; |
2507 | enum reg_class pclass; |
2508 | ira_object_t obj; |
2509 | |
2510 | pclass = ira_pressure_class_translate[REGNO_REG_CLASS (hard_regno)]; |
2511 | nwords = ALLOCNO_NUM_OBJECTS (a); |
2512 | for (i = 0; i < nwords; i++) |
2513 | { |
2514 | obj = ALLOCNO_OBJECT (a, i); |
2515 | OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) |
2516 | |= ~reg_class_contents[pclass]; |
2517 | } |
2518 | if (ira_need_caller_save_p (a, regno: hard_regno)) |
2519 | { |
2520 | ira_assert (!optimize || flag_caller_saves |
2521 | || (ALLOCNO_CALLS_CROSSED_NUM (a) |
2522 | == ALLOCNO_CHEAP_CALLS_CROSSED_NUM (a)) |
2523 | || regno >= ira_reg_equiv_len |
2524 | || ira_equiv_no_lvalue_p (regno)); |
2525 | caller_save_needed = 1; |
2526 | } |
2527 | } |
2528 | } |
2529 | } |
2530 | |
2531 | /* Set up allocno assignment flags for further allocation |
2532 | improvements. */ |
2533 | static void |
2534 | setup_allocno_assignment_flags (void) |
2535 | { |
2536 | int hard_regno; |
2537 | ira_allocno_t a; |
2538 | ira_allocno_iterator ai; |
2539 | |
2540 | FOR_EACH_ALLOCNO (a, ai) |
2541 | { |
2542 | if (! ALLOCNO_ASSIGNED_P (a)) |
2543 | /* It can happen if A is not referenced but partially anticipated |
2544 | somewhere in a region. */ |
2545 | ira_free_allocno_updated_costs (a); |
2546 | hard_regno = ALLOCNO_HARD_REGNO (a); |
2547 | /* Don't assign hard registers to allocnos which are destination |
2548 | of removed store at the end of loop. It has no sense to keep |
2549 | the same value in different hard registers. It is also |
2550 | impossible to assign hard registers correctly to such |
2551 | allocnos because the cost info and info about intersected |
2552 | calls are incorrect for them. */ |
2553 | ALLOCNO_ASSIGNED_P (a) = (hard_regno >= 0 |
2554 | || ALLOCNO_EMIT_DATA (a)->mem_optimized_dest_p |
2555 | || (ALLOCNO_MEMORY_COST (a) |
2556 | - ALLOCNO_CLASS_COST (a)) < 0); |
2557 | ira_assert |
2558 | (hard_regno < 0 |
2559 | || ira_hard_reg_in_set_p (hard_regno, ALLOCNO_MODE (a), |
2560 | reg_class_contents[ALLOCNO_CLASS (a)])); |
2561 | } |
2562 | } |
2563 | |
2564 | /* Evaluate overall allocation cost and the costs for using hard |
2565 | registers and memory for allocnos. */ |
2566 | static void |
2567 | calculate_allocation_cost (void) |
2568 | { |
2569 | int hard_regno, cost; |
2570 | ira_allocno_t a; |
2571 | ira_allocno_iterator ai; |
2572 | |
2573 | ira_overall_cost = ira_reg_cost = ira_mem_cost = 0; |
2574 | FOR_EACH_ALLOCNO (a, ai) |
2575 | { |
2576 | hard_regno = ALLOCNO_HARD_REGNO (a); |
2577 | ira_assert (hard_regno < 0 |
2578 | || (ira_hard_reg_in_set_p |
2579 | (hard_regno, ALLOCNO_MODE (a), |
2580 | reg_class_contents[ALLOCNO_CLASS (a)]))); |
2581 | if (hard_regno < 0) |
2582 | { |
2583 | cost = ALLOCNO_MEMORY_COST (a); |
2584 | ira_mem_cost += cost; |
2585 | } |
2586 | else if (ALLOCNO_HARD_REG_COSTS (a) != NULL) |
2587 | { |
2588 | cost = (ALLOCNO_HARD_REG_COSTS (a) |
2589 | [ira_class_hard_reg_index |
2590 | [ALLOCNO_CLASS (a)][hard_regno]]); |
2591 | ira_reg_cost += cost; |
2592 | } |
2593 | else |
2594 | { |
2595 | cost = ALLOCNO_CLASS_COST (a); |
2596 | ira_reg_cost += cost; |
2597 | } |
2598 | ira_overall_cost += cost; |
2599 | } |
2600 | |
2601 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL) |
2602 | { |
2603 | fprintf (stream: ira_dump_file, |
2604 | format: "+++Costs: overall %" PRId64 |
2605 | ", reg %" PRId64 |
2606 | ", mem %" PRId64 |
2607 | ", ld %" PRId64 |
2608 | ", st %" PRId64 |
2609 | ", move %" PRId64, |
2610 | ira_overall_cost, ira_reg_cost, ira_mem_cost, |
2611 | ira_load_cost, ira_store_cost, ira_shuffle_cost); |
2612 | fprintf (stream: ira_dump_file, format: "\n+++ move loops %d, new jumps %d\n" , |
2613 | ira_move_loops_num, ira_additional_jumps_num); |
2614 | } |
2615 | |
2616 | } |
2617 | |
2618 | #ifdef ENABLE_IRA_CHECKING |
2619 | /* Check the correctness of the allocation. We do need this because |
2620 | of complicated code to transform more one region internal |
2621 | representation into one region representation. */ |
2622 | static void |
2623 | check_allocation (void) |
2624 | { |
2625 | ira_allocno_t a; |
2626 | int hard_regno, nregs, conflict_nregs; |
2627 | ira_allocno_iterator ai; |
2628 | |
2629 | FOR_EACH_ALLOCNO (a, ai) |
2630 | { |
2631 | int n = ALLOCNO_NUM_OBJECTS (a); |
2632 | int i; |
2633 | |
2634 | if (ALLOCNO_CAP_MEMBER (a) != NULL |
2635 | || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0) |
2636 | continue; |
2637 | nregs = hard_regno_nregs (regno: hard_regno, ALLOCNO_MODE (a)); |
2638 | if (nregs == 1) |
2639 | /* We allocated a single hard register. */ |
2640 | n = 1; |
2641 | else if (n > 1) |
2642 | /* We allocated multiple hard registers, and we will test |
2643 | conflicts in a granularity of single hard regs. */ |
2644 | nregs = 1; |
2645 | |
2646 | for (i = 0; i < n; i++) |
2647 | { |
2648 | ira_object_t obj = ALLOCNO_OBJECT (a, i); |
2649 | ira_object_t conflict_obj; |
2650 | ira_object_conflict_iterator oci; |
2651 | int this_regno = hard_regno; |
2652 | if (n > 1) |
2653 | { |
2654 | if (REG_WORDS_BIG_ENDIAN) |
2655 | this_regno += n - i - 1; |
2656 | else |
2657 | this_regno += i; |
2658 | } |
2659 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) |
2660 | { |
2661 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); |
2662 | int conflict_hard_regno = ALLOCNO_HARD_REGNO (conflict_a); |
2663 | if (conflict_hard_regno < 0) |
2664 | continue; |
2665 | if (ira_soft_conflict (a, conflict_a)) |
2666 | continue; |
2667 | |
2668 | conflict_nregs = hard_regno_nregs (regno: conflict_hard_regno, |
2669 | ALLOCNO_MODE (conflict_a)); |
2670 | |
2671 | if (ALLOCNO_NUM_OBJECTS (conflict_a) > 1 |
2672 | && conflict_nregs == ALLOCNO_NUM_OBJECTS (conflict_a)) |
2673 | { |
2674 | if (REG_WORDS_BIG_ENDIAN) |
2675 | conflict_hard_regno += (ALLOCNO_NUM_OBJECTS (conflict_a) |
2676 | - OBJECT_SUBWORD (conflict_obj) - 1); |
2677 | else |
2678 | conflict_hard_regno += OBJECT_SUBWORD (conflict_obj); |
2679 | conflict_nregs = 1; |
2680 | } |
2681 | |
2682 | if ((conflict_hard_regno <= this_regno |
2683 | && this_regno < conflict_hard_regno + conflict_nregs) |
2684 | || (this_regno <= conflict_hard_regno |
2685 | && conflict_hard_regno < this_regno + nregs)) |
2686 | { |
2687 | fprintf (stderr, format: "bad allocation for %d and %d\n" , |
2688 | ALLOCNO_REGNO (a), ALLOCNO_REGNO (conflict_a)); |
2689 | gcc_unreachable (); |
2690 | } |
2691 | } |
2692 | } |
2693 | } |
2694 | } |
2695 | #endif |
2696 | |
2697 | /* Allocate REG_EQUIV_INIT. Set up it from IRA_REG_EQUIV which should |
2698 | be already calculated. */ |
2699 | static void |
2700 | setup_reg_equiv_init (void) |
2701 | { |
2702 | int i; |
2703 | int max_regno = max_reg_num (); |
2704 | |
2705 | for (i = 0; i < max_regno; i++) |
2706 | reg_equiv_init (i) = ira_reg_equiv[i].init_insns; |
2707 | } |
2708 | |
2709 | /* Update equiv regno from movement of FROM_REGNO to TO_REGNO. INSNS |
2710 | are insns which were generated for such movement. It is assumed |
2711 | that FROM_REGNO and TO_REGNO always have the same value at the |
2712 | point of any move containing such registers. This function is used |
2713 | to update equiv info for register shuffles on the region borders |
2714 | and for caller save/restore insns. */ |
2715 | void |
2716 | ira_update_equiv_info_by_shuffle_insn (int to_regno, int from_regno, rtx_insn *insns) |
2717 | { |
2718 | rtx_insn *insn; |
2719 | rtx x, note; |
2720 | |
2721 | if (! ira_reg_equiv[from_regno].defined_p |
2722 | && (! ira_reg_equiv[to_regno].defined_p |
2723 | || ((x = ira_reg_equiv[to_regno].memory) != NULL_RTX |
2724 | && ! MEM_READONLY_P (x)))) |
2725 | return; |
2726 | insn = insns; |
2727 | if (NEXT_INSN (insn) != NULL_RTX) |
2728 | { |
2729 | if (! ira_reg_equiv[to_regno].defined_p) |
2730 | { |
2731 | ira_assert (ira_reg_equiv[to_regno].init_insns == NULL_RTX); |
2732 | return; |
2733 | } |
2734 | ira_reg_equiv[to_regno].defined_p = false; |
2735 | ira_reg_equiv[to_regno].caller_save_p = false; |
2736 | ira_reg_equiv[to_regno].memory |
2737 | = ira_reg_equiv[to_regno].constant |
2738 | = ira_reg_equiv[to_regno].invariant |
2739 | = ira_reg_equiv[to_regno].init_insns = NULL; |
2740 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2741 | fprintf (stream: ira_dump_file, |
2742 | format: " Invalidating equiv info for reg %d\n" , to_regno); |
2743 | return; |
2744 | } |
2745 | /* It is possible that FROM_REGNO still has no equivalence because |
2746 | in shuffles to_regno<-from_regno and from_regno<-to_regno the 2nd |
2747 | insn was not processed yet. */ |
2748 | if (ira_reg_equiv[from_regno].defined_p) |
2749 | { |
2750 | ira_reg_equiv[to_regno].defined_p = true; |
2751 | if ((x = ira_reg_equiv[from_regno].memory) != NULL_RTX) |
2752 | { |
2753 | ira_assert (ira_reg_equiv[from_regno].invariant == NULL_RTX |
2754 | && ira_reg_equiv[from_regno].constant == NULL_RTX); |
2755 | ira_assert (ira_reg_equiv[to_regno].memory == NULL_RTX |
2756 | || rtx_equal_p (ira_reg_equiv[to_regno].memory, x)); |
2757 | ira_reg_equiv[to_regno].memory = x; |
2758 | if (! MEM_READONLY_P (x)) |
2759 | /* We don't add the insn to insn init list because memory |
2760 | equivalence is just to say what memory is better to use |
2761 | when the pseudo is spilled. */ |
2762 | return; |
2763 | } |
2764 | else if ((x = ira_reg_equiv[from_regno].constant) != NULL_RTX) |
2765 | { |
2766 | ira_assert (ira_reg_equiv[from_regno].invariant == NULL_RTX); |
2767 | ira_assert (ira_reg_equiv[to_regno].constant == NULL_RTX |
2768 | || rtx_equal_p (ira_reg_equiv[to_regno].constant, x)); |
2769 | ira_reg_equiv[to_regno].constant = x; |
2770 | } |
2771 | else |
2772 | { |
2773 | x = ira_reg_equiv[from_regno].invariant; |
2774 | ira_assert (x != NULL_RTX); |
2775 | ira_assert (ira_reg_equiv[to_regno].invariant == NULL_RTX |
2776 | || rtx_equal_p (ira_reg_equiv[to_regno].invariant, x)); |
2777 | ira_reg_equiv[to_regno].invariant = x; |
2778 | } |
2779 | if (find_reg_note (insn, REG_EQUIV, x) == NULL_RTX) |
2780 | { |
2781 | note = set_unique_reg_note (insn, REG_EQUIV, copy_rtx (x)); |
2782 | gcc_assert (note != NULL_RTX); |
2783 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2784 | { |
2785 | fprintf (stream: ira_dump_file, |
2786 | format: " Adding equiv note to insn %u for reg %d " , |
2787 | INSN_UID (insn), to_regno); |
2788 | dump_value_slim (ira_dump_file, x, 1); |
2789 | fprintf (stream: ira_dump_file, format: "\n" ); |
2790 | } |
2791 | } |
2792 | } |
2793 | ira_reg_equiv[to_regno].init_insns |
2794 | = gen_rtx_INSN_LIST (VOIDmode, insn, |
2795 | ira_reg_equiv[to_regno].init_insns); |
2796 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2797 | fprintf (stream: ira_dump_file, |
2798 | format: " Adding equiv init move insn %u to reg %d\n" , |
2799 | INSN_UID (insn), to_regno); |
2800 | } |
2801 | |
2802 | /* Fix values of array REG_EQUIV_INIT after live range splitting done |
2803 | by IRA. */ |
2804 | static void |
2805 | fix_reg_equiv_init (void) |
2806 | { |
2807 | int max_regno = max_reg_num (); |
2808 | int i, new_regno, max; |
2809 | rtx set; |
2810 | rtx_insn_list *x, *next, *prev; |
2811 | rtx_insn *insn; |
2812 | |
2813 | if (max_regno_before_ira < max_regno) |
2814 | { |
2815 | max = vec_safe_length (v: reg_equivs); |
2816 | grow_reg_equivs (); |
2817 | for (i = FIRST_PSEUDO_REGISTER; i < max; i++) |
2818 | for (prev = NULL, x = reg_equiv_init (i); |
2819 | x != NULL_RTX; |
2820 | x = next) |
2821 | { |
2822 | next = x->next (); |
2823 | insn = x->insn (); |
2824 | set = single_set (insn); |
2825 | ira_assert (set != NULL_RTX |
2826 | && (REG_P (SET_DEST (set)) || REG_P (SET_SRC (set)))); |
2827 | if (REG_P (SET_DEST (set)) |
2828 | && ((int) REGNO (SET_DEST (set)) == i |
2829 | || (int) ORIGINAL_REGNO (SET_DEST (set)) == i)) |
2830 | new_regno = REGNO (SET_DEST (set)); |
2831 | else if (REG_P (SET_SRC (set)) |
2832 | && ((int) REGNO (SET_SRC (set)) == i |
2833 | || (int) ORIGINAL_REGNO (SET_SRC (set)) == i)) |
2834 | new_regno = REGNO (SET_SRC (set)); |
2835 | else |
2836 | gcc_unreachable (); |
2837 | if (new_regno == i) |
2838 | prev = x; |
2839 | else |
2840 | { |
2841 | /* Remove the wrong list element. */ |
2842 | if (prev == NULL_RTX) |
2843 | reg_equiv_init (i) = next; |
2844 | else |
2845 | XEXP (prev, 1) = next; |
2846 | XEXP (x, 1) = reg_equiv_init (new_regno); |
2847 | reg_equiv_init (new_regno) = x; |
2848 | } |
2849 | } |
2850 | } |
2851 | } |
2852 | |
2853 | #ifdef ENABLE_IRA_CHECKING |
2854 | /* Print redundant memory-memory copies. */ |
2855 | static void |
2856 | print_redundant_copies (void) |
2857 | { |
2858 | int hard_regno; |
2859 | ira_allocno_t a; |
2860 | ira_copy_t cp, next_cp; |
2861 | ira_allocno_iterator ai; |
2862 | |
2863 | FOR_EACH_ALLOCNO (a, ai) |
2864 | { |
2865 | if (ALLOCNO_CAP_MEMBER (a) != NULL) |
2866 | /* It is a cap. */ |
2867 | continue; |
2868 | hard_regno = ALLOCNO_HARD_REGNO (a); |
2869 | if (hard_regno >= 0) |
2870 | continue; |
2871 | for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp) |
2872 | if (cp->first == a) |
2873 | next_cp = cp->next_first_allocno_copy; |
2874 | else |
2875 | { |
2876 | next_cp = cp->next_second_allocno_copy; |
2877 | if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL |
2878 | && cp->insn != NULL_RTX |
2879 | && ALLOCNO_HARD_REGNO (cp->first) == hard_regno) |
2880 | fprintf (stream: ira_dump_file, |
2881 | format: " Redundant move from %d(freq %d):%d\n" , |
2882 | INSN_UID (insn: cp->insn), cp->freq, hard_regno); |
2883 | } |
2884 | } |
2885 | } |
2886 | #endif |
2887 | |
2888 | /* Setup preferred and alternative classes for new pseudo-registers |
2889 | created by IRA starting with START. */ |
2890 | static void |
2891 | setup_preferred_alternate_classes_for_new_pseudos (int start) |
2892 | { |
2893 | int i, old_regno; |
2894 | int max_regno = max_reg_num (); |
2895 | |
2896 | for (i = start; i < max_regno; i++) |
2897 | { |
2898 | old_regno = ORIGINAL_REGNO (regno_reg_rtx[i]); |
2899 | ira_assert (i != old_regno); |
2900 | setup_reg_classes (i, reg_preferred_class (old_regno), |
2901 | reg_alternate_class (old_regno), |
2902 | reg_allocno_class (old_regno)); |
2903 | if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL) |
2904 | fprintf (stream: ira_dump_file, |
2905 | format: " New r%d: setting preferred %s, alternative %s\n" , |
2906 | i, reg_class_names[reg_preferred_class (old_regno)], |
2907 | reg_class_names[reg_alternate_class (old_regno)]); |
2908 | } |
2909 | } |
2910 | |
2911 | |
2912 | /* The number of entries allocated in reg_info. */ |
2913 | static int allocated_reg_info_size; |
2914 | |
2915 | /* Regional allocation can create new pseudo-registers. This function |
2916 | expands some arrays for pseudo-registers. */ |
2917 | static void |
2918 | expand_reg_info (void) |
2919 | { |
2920 | int i; |
2921 | int size = max_reg_num (); |
2922 | |
2923 | resize_reg_info (); |
2924 | for (i = allocated_reg_info_size; i < size; i++) |
2925 | setup_reg_classes (i, GENERAL_REGS, ALL_REGS, GENERAL_REGS); |
2926 | setup_preferred_alternate_classes_for_new_pseudos (allocated_reg_info_size); |
2927 | allocated_reg_info_size = size; |
2928 | } |
2929 | |
2930 | /* Return TRUE if there is too high register pressure in the function. |
2931 | It is used to decide when stack slot sharing is worth to do. */ |
2932 | static bool |
2933 | too_high_register_pressure_p (void) |
2934 | { |
2935 | int i; |
2936 | enum reg_class pclass; |
2937 | |
2938 | for (i = 0; i < ira_pressure_classes_num; i++) |
2939 | { |
2940 | pclass = ira_pressure_classes[i]; |
2941 | if (ira_loop_tree_root->reg_pressure[pclass] > 10000) |
2942 | return true; |
2943 | } |
2944 | return false; |
2945 | } |
2946 | |
2947 | |
2948 | |
2949 | /* Indicate that hard register number FROM was eliminated and replaced with |
2950 | an offset from hard register number TO. The status of hard registers live |
2951 | at the start of a basic block is updated by replacing a use of FROM with |
2952 | a use of TO. */ |
2953 | |
2954 | void |
2955 | mark_elimination (int from, int to) |
2956 | { |
2957 | basic_block bb; |
2958 | bitmap r; |
2959 | |
2960 | FOR_EACH_BB_FN (bb, cfun) |
2961 | { |
2962 | r = DF_LR_IN (bb); |
2963 | if (bitmap_bit_p (r, from)) |
2964 | { |
2965 | bitmap_clear_bit (r, from); |
2966 | bitmap_set_bit (r, to); |
2967 | } |
2968 | if (! df_live) |
2969 | continue; |
2970 | r = DF_LIVE_IN (bb); |
2971 | if (bitmap_bit_p (r, from)) |
2972 | { |
2973 | bitmap_clear_bit (r, from); |
2974 | bitmap_set_bit (r, to); |
2975 | } |
2976 | } |
2977 | } |
2978 | |
2979 | |
2980 | |
2981 | /* The length of the following array. */ |
2982 | int ira_reg_equiv_len; |
2983 | |
2984 | /* Info about equiv. info for each register. */ |
2985 | struct ira_reg_equiv_s *ira_reg_equiv; |
2986 | |
2987 | /* Expand ira_reg_equiv if necessary. */ |
2988 | void |
2989 | ira_expand_reg_equiv (void) |
2990 | { |
2991 | int old = ira_reg_equiv_len; |
2992 | |
2993 | if (ira_reg_equiv_len > max_reg_num ()) |
2994 | return; |
2995 | ira_reg_equiv_len = max_reg_num () * 3 / 2 + 1; |
2996 | ira_reg_equiv |
2997 | = (struct ira_reg_equiv_s *) xrealloc (ira_reg_equiv, |
2998 | ira_reg_equiv_len |
2999 | * sizeof (struct ira_reg_equiv_s)); |
3000 | gcc_assert (old < ira_reg_equiv_len); |
3001 | memset (s: ira_reg_equiv + old, c: 0, |
3002 | n: sizeof (struct ira_reg_equiv_s) * (ira_reg_equiv_len - old)); |
3003 | } |
3004 | |
3005 | static void |
3006 | init_reg_equiv (void) |
3007 | { |
3008 | ira_reg_equiv_len = 0; |
3009 | ira_reg_equiv = NULL; |
3010 | ira_expand_reg_equiv (); |
3011 | } |
3012 | |
3013 | static void |
3014 | finish_reg_equiv (void) |
3015 | { |
3016 | free (ptr: ira_reg_equiv); |
3017 | } |
3018 | |
3019 | |
3020 | |
3021 | struct equivalence |
3022 | { |
3023 | /* Set when a REG_EQUIV note is found or created. Use to |
3024 | keep track of what memory accesses might be created later, |
3025 | e.g. by reload. */ |
3026 | rtx replacement; |
3027 | rtx *src_p; |
3028 | |
3029 | /* The list of each instruction which initializes this register. |
3030 | |
3031 | NULL indicates we know nothing about this register's equivalence |
3032 | properties. |
3033 | |
3034 | An INSN_LIST with a NULL insn indicates this pseudo is already |
3035 | known to not have a valid equivalence. */ |
3036 | rtx_insn_list *init_insns; |
3037 | |
3038 | /* Loop depth is used to recognize equivalences which appear |
3039 | to be present within the same loop (or in an inner loop). */ |
3040 | short loop_depth; |
3041 | /* Nonzero if this had a preexisting REG_EQUIV note. */ |
3042 | unsigned char is_arg_equivalence : 1; |
3043 | /* Set when an attempt should be made to replace a register |
3044 | with the associated src_p entry. */ |
3045 | unsigned char replace : 1; |
3046 | /* Set if this register has no known equivalence. */ |
3047 | unsigned char no_equiv : 1; |
3048 | /* Set if this register is mentioned in a paradoxical subreg. */ |
3049 | unsigned char pdx_subregs : 1; |
3050 | }; |
3051 | |
3052 | /* reg_equiv[N] (where N is a pseudo reg number) is the equivalence |
3053 | structure for that register. */ |
3054 | static struct equivalence *reg_equiv; |
3055 | |
3056 | /* Used for communication between the following two functions. */ |
3057 | struct equiv_mem_data |
3058 | { |
3059 | /* A MEM that we wish to ensure remains unchanged. */ |
3060 | rtx equiv_mem; |
3061 | |
3062 | /* Set true if EQUIV_MEM is modified. */ |
3063 | bool equiv_mem_modified; |
3064 | }; |
3065 | |
3066 | /* If EQUIV_MEM is modified by modifying DEST, indicate that it is modified. |
3067 | Called via note_stores. */ |
3068 | static void |
3069 | validate_equiv_mem_from_store (rtx dest, const_rtx set ATTRIBUTE_UNUSED, |
3070 | void *data) |
3071 | { |
3072 | struct equiv_mem_data *info = (struct equiv_mem_data *) data; |
3073 | |
3074 | if ((REG_P (dest) |
3075 | && reg_overlap_mentioned_p (dest, info->equiv_mem)) |
3076 | || (MEM_P (dest) |
3077 | && anti_dependence (info->equiv_mem, dest))) |
3078 | info->equiv_mem_modified = true; |
3079 | } |
3080 | |
3081 | static bool equiv_init_varies_p (rtx x); |
3082 | |
3083 | enum valid_equiv { valid_none, valid_combine, valid_reload }; |
3084 | |
3085 | /* Verify that no store between START and the death of REG invalidates |
3086 | MEMREF. MEMREF is invalidated by modifying a register used in MEMREF, |
3087 | by storing into an overlapping memory location, or with a non-const |
3088 | CALL_INSN. |
3089 | |
3090 | Return VALID_RELOAD if MEMREF remains valid for both reload and |
3091 | combine_and_move insns, VALID_COMBINE if only valid for |
3092 | combine_and_move_insns, and VALID_NONE otherwise. */ |
3093 | static enum valid_equiv |
3094 | validate_equiv_mem (rtx_insn *start, rtx reg, rtx memref) |
3095 | { |
3096 | rtx_insn *insn; |
3097 | rtx note; |
3098 | struct equiv_mem_data info = { .equiv_mem: memref, .equiv_mem_modified: false }; |
3099 | enum valid_equiv ret = valid_reload; |
3100 | |
3101 | /* If the memory reference has side effects or is volatile, it isn't a |
3102 | valid equivalence. */ |
3103 | if (side_effects_p (memref)) |
3104 | return valid_none; |
3105 | |
3106 | for (insn = start; insn; insn = NEXT_INSN (insn)) |
3107 | { |
3108 | if (!INSN_P (insn)) |
3109 | continue; |
3110 | |
3111 | if (find_reg_note (insn, REG_DEAD, reg)) |
3112 | return ret; |
3113 | |
3114 | if (CALL_P (insn)) |
3115 | { |
3116 | /* We can combine a reg def from one insn into a reg use in |
3117 | another over a call if the memory is readonly or the call |
3118 | const/pure. However, we can't set reg_equiv notes up for |
3119 | reload over any call. The problem is the equivalent form |
3120 | may reference a pseudo which gets assigned a call |
3121 | clobbered hard reg. When we later replace REG with its |
3122 | equivalent form, the value in the call-clobbered reg has |
3123 | been changed and all hell breaks loose. */ |
3124 | ret = valid_combine; |
3125 | if (!MEM_READONLY_P (memref) |
3126 | && (!RTL_CONST_OR_PURE_CALL_P (insn) |
3127 | || equiv_init_varies_p (XEXP (memref, 0)))) |
3128 | return valid_none; |
3129 | } |
3130 | |
3131 | note_stores (insn, validate_equiv_mem_from_store, &info); |
3132 | if (info.equiv_mem_modified) |
3133 | return valid_none; |
3134 | |
3135 | /* If a register mentioned in MEMREF is modified via an |
3136 | auto-increment, we lose the equivalence. Do the same if one |
3137 | dies; although we could extend the life, it doesn't seem worth |
3138 | the trouble. */ |
3139 | |
3140 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
3141 | if ((REG_NOTE_KIND (note) == REG_INC |
3142 | || REG_NOTE_KIND (note) == REG_DEAD) |
3143 | && REG_P (XEXP (note, 0)) |
3144 | && reg_overlap_mentioned_p (XEXP (note, 0), memref)) |
3145 | return valid_none; |
3146 | } |
3147 | |
3148 | return valid_none; |
3149 | } |
3150 | |
3151 | /* Returns false if X is known to be invariant. */ |
3152 | static bool |
3153 | equiv_init_varies_p (rtx x) |
3154 | { |
3155 | RTX_CODE code = GET_CODE (x); |
3156 | int i; |
3157 | const char *fmt; |
3158 | |
3159 | switch (code) |
3160 | { |
3161 | case MEM: |
3162 | return !MEM_READONLY_P (x) || equiv_init_varies_p (XEXP (x, 0)); |
3163 | |
3164 | case CONST: |
3165 | CASE_CONST_ANY: |
3166 | case SYMBOL_REF: |
3167 | case LABEL_REF: |
3168 | return false; |
3169 | |
3170 | case REG: |
3171 | return reg_equiv[REGNO (x)].replace == 0 && rtx_varies_p (x, 0); |
3172 | |
3173 | case ASM_OPERANDS: |
3174 | if (MEM_VOLATILE_P (x)) |
3175 | return true; |
3176 | |
3177 | /* Fall through. */ |
3178 | |
3179 | default: |
3180 | break; |
3181 | } |
3182 | |
3183 | fmt = GET_RTX_FORMAT (code); |
3184 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
3185 | if (fmt[i] == 'e') |
3186 | { |
3187 | if (equiv_init_varies_p (XEXP (x, i))) |
3188 | return true; |
3189 | } |
3190 | else if (fmt[i] == 'E') |
3191 | { |
3192 | int j; |
3193 | for (j = 0; j < XVECLEN (x, i); j++) |
3194 | if (equiv_init_varies_p (XVECEXP (x, i, j))) |
3195 | return true; |
3196 | } |
3197 | |
3198 | return false; |
3199 | } |
3200 | |
3201 | /* Returns true if X (used to initialize register REGNO) is movable. |
3202 | X is only movable if the registers it uses have equivalent initializations |
3203 | which appear to be within the same loop (or in an inner loop) and movable |
3204 | or if they are not candidates for local_alloc and don't vary. */ |
3205 | static bool |
3206 | equiv_init_movable_p (rtx x, int regno) |
3207 | { |
3208 | int i, j; |
3209 | const char *fmt; |
3210 | enum rtx_code code = GET_CODE (x); |
3211 | |
3212 | switch (code) |
3213 | { |
3214 | case SET: |
3215 | return equiv_init_movable_p (SET_SRC (x), regno); |
3216 | |
3217 | case CLOBBER: |
3218 | return false; |
3219 | |
3220 | case PRE_INC: |
3221 | case PRE_DEC: |
3222 | case POST_INC: |
3223 | case POST_DEC: |
3224 | case PRE_MODIFY: |
3225 | case POST_MODIFY: |
3226 | return false; |
3227 | |
3228 | case REG: |
3229 | return ((reg_equiv[REGNO (x)].loop_depth >= reg_equiv[regno].loop_depth |
3230 | && reg_equiv[REGNO (x)].replace) |
3231 | || (REG_BASIC_BLOCK (REGNO (x)) < NUM_FIXED_BLOCKS |
3232 | && ! rtx_varies_p (x, 0))); |
3233 | |
3234 | case UNSPEC_VOLATILE: |
3235 | return false; |
3236 | |
3237 | case ASM_OPERANDS: |
3238 | if (MEM_VOLATILE_P (x)) |
3239 | return false; |
3240 | |
3241 | /* Fall through. */ |
3242 | |
3243 | default: |
3244 | break; |
3245 | } |
3246 | |
3247 | fmt = GET_RTX_FORMAT (code); |
3248 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
3249 | switch (fmt[i]) |
3250 | { |
3251 | case 'e': |
3252 | if (! equiv_init_movable_p (XEXP (x, i), regno)) |
3253 | return false; |
3254 | break; |
3255 | case 'E': |
3256 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
3257 | if (! equiv_init_movable_p (XVECEXP (x, i, j), regno)) |
3258 | return false; |
3259 | break; |
3260 | } |
3261 | |
3262 | return true; |
3263 | } |
3264 | |
3265 | static bool memref_referenced_p (rtx memref, rtx x, bool read_p); |
3266 | |
3267 | /* Auxiliary function for memref_referenced_p. Process setting X for |
3268 | MEMREF store. */ |
3269 | static bool |
3270 | process_set_for_memref_referenced_p (rtx memref, rtx x) |
3271 | { |
3272 | /* If we are setting a MEM, it doesn't count (its address does), but any |
3273 | other SET_DEST that has a MEM in it is referencing the MEM. */ |
3274 | if (MEM_P (x)) |
3275 | { |
3276 | if (memref_referenced_p (memref, XEXP (x, 0), read_p: true)) |
3277 | return true; |
3278 | } |
3279 | else if (memref_referenced_p (memref, x, read_p: false)) |
3280 | return true; |
3281 | |
3282 | return false; |
3283 | } |
3284 | |
3285 | /* TRUE if X references a memory location (as a read if READ_P) that |
3286 | would be affected by a store to MEMREF. */ |
3287 | static bool |
3288 | memref_referenced_p (rtx memref, rtx x, bool read_p) |
3289 | { |
3290 | int i, j; |
3291 | const char *fmt; |
3292 | enum rtx_code code = GET_CODE (x); |
3293 | |
3294 | switch (code) |
3295 | { |
3296 | case CONST: |
3297 | case LABEL_REF: |
3298 | case SYMBOL_REF: |
3299 | CASE_CONST_ANY: |
3300 | case PC: |
3301 | case HIGH: |
3302 | case LO_SUM: |
3303 | return false; |
3304 | |
3305 | case REG: |
3306 | return (reg_equiv[REGNO (x)].replacement |
3307 | && memref_referenced_p (memref, |
3308 | x: reg_equiv[REGNO (x)].replacement, read_p)); |
3309 | |
3310 | case MEM: |
3311 | /* Memory X might have another effective type than MEMREF. */ |
3312 | if (read_p || true_dependence (memref, VOIDmode, x)) |
3313 | return true; |
3314 | break; |
3315 | |
3316 | case SET: |
3317 | if (process_set_for_memref_referenced_p (memref, SET_DEST (x))) |
3318 | return true; |
3319 | |
3320 | return memref_referenced_p (memref, SET_SRC (x), read_p: true); |
3321 | |
3322 | case CLOBBER: |
3323 | if (process_set_for_memref_referenced_p (memref, XEXP (x, 0))) |
3324 | return true; |
3325 | |
3326 | return false; |
3327 | |
3328 | case PRE_DEC: |
3329 | case POST_DEC: |
3330 | case PRE_INC: |
3331 | case POST_INC: |
3332 | if (process_set_for_memref_referenced_p (memref, XEXP (x, 0))) |
3333 | return true; |
3334 | |
3335 | return memref_referenced_p (memref, XEXP (x, 0), read_p: true); |
3336 | |
3337 | case POST_MODIFY: |
3338 | case PRE_MODIFY: |
3339 | /* op0 = op0 + op1 */ |
3340 | if (process_set_for_memref_referenced_p (memref, XEXP (x, 0))) |
3341 | return true; |
3342 | |
3343 | if (memref_referenced_p (memref, XEXP (x, 0), read_p: true)) |
3344 | return true; |
3345 | |
3346 | return memref_referenced_p (memref, XEXP (x, 1), read_p: true); |
3347 | |
3348 | default: |
3349 | break; |
3350 | } |
3351 | |
3352 | fmt = GET_RTX_FORMAT (code); |
3353 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
3354 | switch (fmt[i]) |
3355 | { |
3356 | case 'e': |
3357 | if (memref_referenced_p (memref, XEXP (x, i), read_p)) |
3358 | return true; |
3359 | break; |
3360 | case 'E': |
3361 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
3362 | if (memref_referenced_p (memref, XVECEXP (x, i, j), read_p)) |
3363 | return true; |
3364 | break; |
3365 | } |
3366 | |
3367 | return false; |
3368 | } |
3369 | |
3370 | /* TRUE if some insn in the range (START, END] references a memory location |
3371 | that would be affected by a store to MEMREF. |
3372 | |
3373 | Callers should not call this routine if START is after END in the |
3374 | RTL chain. */ |
3375 | |
3376 | static bool |
3377 | memref_used_between_p (rtx memref, rtx_insn *start, rtx_insn *end) |
3378 | { |
3379 | rtx_insn *insn; |
3380 | |
3381 | for (insn = NEXT_INSN (insn: start); |
3382 | insn && insn != NEXT_INSN (insn: end); |
3383 | insn = NEXT_INSN (insn)) |
3384 | { |
3385 | if (!NONDEBUG_INSN_P (insn)) |
3386 | continue; |
3387 | |
3388 | if (memref_referenced_p (memref, x: PATTERN (insn), read_p: false)) |
3389 | return true; |
3390 | |
3391 | /* Nonconst functions may access memory. */ |
3392 | if (CALL_P (insn) && (! RTL_CONST_CALL_P (insn))) |
3393 | return true; |
3394 | } |
3395 | |
3396 | gcc_assert (insn == NEXT_INSN (end)); |
3397 | return false; |
3398 | } |
3399 | |
3400 | /* Mark REG as having no known equivalence. |
3401 | Some instructions might have been processed before and furnished |
3402 | with REG_EQUIV notes for this register; these notes will have to be |
3403 | removed. |
3404 | STORE is the piece of RTL that does the non-constant / conflicting |
3405 | assignment - a SET, CLOBBER or REG_INC note. It is currently not used, |
3406 | but needs to be there because this function is called from note_stores. */ |
3407 | static void |
3408 | no_equiv (rtx reg, const_rtx store ATTRIBUTE_UNUSED, |
3409 | void *data ATTRIBUTE_UNUSED) |
3410 | { |
3411 | int regno; |
3412 | rtx_insn_list *list; |
3413 | |
3414 | if (!REG_P (reg)) |
3415 | return; |
3416 | regno = REGNO (reg); |
3417 | reg_equiv[regno].no_equiv = 1; |
3418 | list = reg_equiv[regno].init_insns; |
3419 | if (list && list->insn () == NULL) |
3420 | return; |
3421 | reg_equiv[regno].init_insns = gen_rtx_INSN_LIST (VOIDmode, NULL_RTX, NULL); |
3422 | reg_equiv[regno].replacement = NULL_RTX; |
3423 | /* This doesn't matter for equivalences made for argument registers, we |
3424 | should keep their initialization insns. */ |
3425 | if (reg_equiv[regno].is_arg_equivalence) |
3426 | return; |
3427 | ira_reg_equiv[regno].defined_p = false; |
3428 | ira_reg_equiv[regno].caller_save_p = false; |
3429 | ira_reg_equiv[regno].init_insns = NULL; |
3430 | for (; list; list = list->next ()) |
3431 | { |
3432 | rtx_insn *insn = list->insn (); |
3433 | remove_note (insn, find_reg_note (insn, REG_EQUIV, NULL_RTX)); |
3434 | } |
3435 | } |
3436 | |
3437 | /* Check whether the SUBREG is a paradoxical subreg and set the result |
3438 | in PDX_SUBREGS. */ |
3439 | |
3440 | static void |
3441 | set_paradoxical_subreg (rtx_insn *insn) |
3442 | { |
3443 | subrtx_iterator::array_type array; |
3444 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) |
3445 | { |
3446 | const_rtx subreg = *iter; |
3447 | if (GET_CODE (subreg) == SUBREG) |
3448 | { |
3449 | const_rtx reg = SUBREG_REG (subreg); |
3450 | if (REG_P (reg) && paradoxical_subreg_p (x: subreg)) |
3451 | reg_equiv[REGNO (reg)].pdx_subregs = true; |
3452 | } |
3453 | } |
3454 | } |
3455 | |
3456 | /* In DEBUG_INSN location adjust REGs from CLEARED_REGS bitmap to the |
3457 | equivalent replacement. */ |
3458 | |
3459 | static rtx |
3460 | adjust_cleared_regs (rtx loc, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data) |
3461 | { |
3462 | if (REG_P (loc)) |
3463 | { |
3464 | bitmap cleared_regs = (bitmap) data; |
3465 | if (bitmap_bit_p (cleared_regs, REGNO (loc))) |
3466 | return simplify_replace_fn_rtx (copy_rtx (*reg_equiv[REGNO (loc)].src_p), |
3467 | NULL_RTX, fn: adjust_cleared_regs, data); |
3468 | } |
3469 | return NULL_RTX; |
3470 | } |
3471 | |
3472 | /* Given register REGNO is set only once, return true if the defining |
3473 | insn dominates all uses. */ |
3474 | |
3475 | static bool |
3476 | def_dominates_uses (int regno) |
3477 | { |
3478 | df_ref def = DF_REG_DEF_CHAIN (regno); |
3479 | |
3480 | struct df_insn_info *def_info = DF_REF_INSN_INFO (def); |
3481 | /* If this is an artificial def (eh handler regs, hard frame pointer |
3482 | for non-local goto, regs defined on function entry) then def_info |
3483 | is NULL and the reg is always live before any use. We might |
3484 | reasonably return true in that case, but since the only call |
3485 | of this function is currently here in ira.cc when we are looking |
3486 | at a defining insn we can't have an artificial def as that would |
3487 | bump DF_REG_DEF_COUNT. */ |
3488 | gcc_assert (DF_REG_DEF_COUNT (regno) == 1 && def_info != NULL); |
3489 | |
3490 | rtx_insn *def_insn = DF_REF_INSN (def); |
3491 | basic_block def_bb = BLOCK_FOR_INSN (insn: def_insn); |
3492 | |
3493 | for (df_ref use = DF_REG_USE_CHAIN (regno); |
3494 | use; |
3495 | use = DF_REF_NEXT_REG (use)) |
3496 | { |
3497 | struct df_insn_info *use_info = DF_REF_INSN_INFO (use); |
3498 | /* Only check real uses, not artificial ones. */ |
3499 | if (use_info) |
3500 | { |
3501 | rtx_insn *use_insn = DF_REF_INSN (use); |
3502 | if (!DEBUG_INSN_P (use_insn)) |
3503 | { |
3504 | basic_block use_bb = BLOCK_FOR_INSN (insn: use_insn); |
3505 | if (use_bb != def_bb |
3506 | ? !dominated_by_p (CDI_DOMINATORS, use_bb, def_bb) |
3507 | : DF_INSN_INFO_LUID (use_info) < DF_INSN_INFO_LUID (def_info)) |
3508 | return false; |
3509 | } |
3510 | } |
3511 | } |
3512 | return true; |
3513 | } |
3514 | |
3515 | /* Scan the instructions before update_equiv_regs. Record which registers |
3516 | are referenced as paradoxical subregs. Also check for cases in which |
3517 | the current function needs to save a register that one of its call |
3518 | instructions clobbers. |
3519 | |
3520 | These things are logically unrelated, but it's more efficient to do |
3521 | them together. */ |
3522 | |
3523 | static void |
3524 | update_equiv_regs_prescan (void) |
3525 | { |
3526 | basic_block bb; |
3527 | rtx_insn *insn; |
3528 | function_abi_aggregator callee_abis; |
3529 | |
3530 | FOR_EACH_BB_FN (bb, cfun) |
3531 | FOR_BB_INSNS (bb, insn) |
3532 | if (NONDEBUG_INSN_P (insn)) |
3533 | { |
3534 | set_paradoxical_subreg (insn); |
3535 | if (CALL_P (insn)) |
3536 | callee_abis.note_callee_abi (abi: insn_callee_abi (insn)); |
3537 | } |
3538 | |
3539 | HARD_REG_SET = callee_abis.caller_save_regs (*crtl->abi); |
3540 | if (!hard_reg_set_empty_p (x: extra_caller_saves)) |
3541 | for (unsigned int regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno) |
3542 | if (TEST_HARD_REG_BIT (set: extra_caller_saves, bit: regno)) |
3543 | df_set_regs_ever_live (regno, true); |
3544 | } |
3545 | |
3546 | /* Find registers that are equivalent to a single value throughout the |
3547 | compilation (either because they can be referenced in memory or are |
3548 | set once from a single constant). Lower their priority for a |
3549 | register. |
3550 | |
3551 | If such a register is only referenced once, try substituting its |
3552 | value into the using insn. If it succeeds, we can eliminate the |
3553 | register completely. |
3554 | |
3555 | Initialize init_insns in ira_reg_equiv array. */ |
3556 | static void |
3557 | update_equiv_regs (void) |
3558 | { |
3559 | rtx_insn *insn; |
3560 | basic_block bb; |
3561 | |
3562 | /* Scan the insns and find which registers have equivalences. Do this |
3563 | in a separate scan of the insns because (due to -fcse-follow-jumps) |
3564 | a register can be set below its use. */ |
3565 | bitmap setjmp_crosses = regstat_get_setjmp_crosses (); |
3566 | FOR_EACH_BB_FN (bb, cfun) |
3567 | { |
3568 | int loop_depth = bb_loop_depth (bb); |
3569 | |
3570 | for (insn = BB_HEAD (bb); |
3571 | insn != NEXT_INSN (BB_END (bb)); |
3572 | insn = NEXT_INSN (insn)) |
3573 | { |
3574 | rtx note; |
3575 | rtx set; |
3576 | rtx dest, src; |
3577 | int regno; |
3578 | |
3579 | if (! INSN_P (insn)) |
3580 | continue; |
3581 | |
3582 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
3583 | if (REG_NOTE_KIND (note) == REG_INC) |
3584 | no_equiv (XEXP (note, 0), store: note, NULL); |
3585 | |
3586 | set = single_set (insn); |
3587 | |
3588 | /* If this insn contains more (or less) than a single SET, |
3589 | only mark all destinations as having no known equivalence. */ |
3590 | if (set == NULL_RTX |
3591 | || side_effects_p (SET_SRC (set))) |
3592 | { |
3593 | note_pattern_stores (PATTERN (insn), no_equiv, NULL); |
3594 | continue; |
3595 | } |
3596 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) |
3597 | { |
3598 | int i; |
3599 | |
3600 | for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) |
3601 | { |
3602 | rtx part = XVECEXP (PATTERN (insn), 0, i); |
3603 | if (part != set) |
3604 | note_pattern_stores (part, no_equiv, NULL); |
3605 | } |
3606 | } |
3607 | |
3608 | dest = SET_DEST (set); |
3609 | src = SET_SRC (set); |
3610 | |
3611 | /* See if this is setting up the equivalence between an argument |
3612 | register and its stack slot. */ |
3613 | note = find_reg_note (insn, REG_EQUIV, NULL_RTX); |
3614 | if (note) |
3615 | { |
3616 | gcc_assert (REG_P (dest)); |
3617 | regno = REGNO (dest); |
3618 | |
3619 | /* Note that we don't want to clear init_insns in |
3620 | ira_reg_equiv even if there are multiple sets of this |
3621 | register. */ |
3622 | reg_equiv[regno].is_arg_equivalence = 1; |
3623 | |
3624 | /* The insn result can have equivalence memory although |
3625 | the equivalence is not set up by the insn. We add |
3626 | this insn to init insns as it is a flag for now that |
3627 | regno has an equivalence. We will remove the insn |
3628 | from init insn list later. */ |
3629 | if (rtx_equal_p (src, XEXP (note, 0)) || MEM_P (XEXP (note, 0))) |
3630 | ira_reg_equiv[regno].init_insns |
3631 | = gen_rtx_INSN_LIST (VOIDmode, insn, |
3632 | ira_reg_equiv[regno].init_insns); |
3633 | |
3634 | /* Continue normally in case this is a candidate for |
3635 | replacements. */ |
3636 | } |
3637 | |
3638 | if (!optimize) |
3639 | continue; |
3640 | |
3641 | /* We only handle the case of a pseudo register being set |
3642 | once, or always to the same value. */ |
3643 | /* ??? The mn10200 port breaks if we add equivalences for |
3644 | values that need an ADDRESS_REGS register and set them equivalent |
3645 | to a MEM of a pseudo. The actual problem is in the over-conservative |
3646 | handling of INPADDR_ADDRESS / INPUT_ADDRESS / INPUT triples in |
3647 | calculate_needs, but we traditionally work around this problem |
3648 | here by rejecting equivalences when the destination is in a register |
3649 | that's likely spilled. This is fragile, of course, since the |
3650 | preferred class of a pseudo depends on all instructions that set |
3651 | or use it. */ |
3652 | |
3653 | if (!REG_P (dest) |
3654 | || (regno = REGNO (dest)) < FIRST_PSEUDO_REGISTER |
3655 | || (reg_equiv[regno].init_insns |
3656 | && reg_equiv[regno].init_insns->insn () == NULL) |
3657 | || (targetm.class_likely_spilled_p (reg_preferred_class (regno)) |
3658 | && MEM_P (src) && ! reg_equiv[regno].is_arg_equivalence)) |
3659 | { |
3660 | /* This might be setting a SUBREG of a pseudo, a pseudo that is |
3661 | also set somewhere else to a constant. */ |
3662 | note_pattern_stores (set, no_equiv, NULL); |
3663 | continue; |
3664 | } |
3665 | |
3666 | /* Don't set reg mentioned in a paradoxical subreg |
3667 | equivalent to a mem. */ |
3668 | if (MEM_P (src) && reg_equiv[regno].pdx_subregs) |
3669 | { |
3670 | note_pattern_stores (set, no_equiv, NULL); |
3671 | continue; |
3672 | } |
3673 | |
3674 | note = find_reg_note (insn, REG_EQUAL, NULL_RTX); |
3675 | |
3676 | /* cse sometimes generates function invariants, but doesn't put a |
3677 | REG_EQUAL note on the insn. Since this note would be redundant, |
3678 | there's no point creating it earlier than here. */ |
3679 | if (! note && ! rtx_varies_p (src, 0)) |
3680 | note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src)); |
3681 | |
3682 | /* Don't bother considering a REG_EQUAL note containing an EXPR_LIST |
3683 | since it represents a function call. */ |
3684 | if (note && GET_CODE (XEXP (note, 0)) == EXPR_LIST) |
3685 | note = NULL_RTX; |
3686 | |
3687 | if (DF_REG_DEF_COUNT (regno) != 1) |
3688 | { |
3689 | bool equal_p = true; |
3690 | rtx_insn_list *list; |
3691 | |
3692 | /* If we have already processed this pseudo and determined it |
3693 | cannot have an equivalence, then honor that decision. */ |
3694 | if (reg_equiv[regno].no_equiv) |
3695 | continue; |
3696 | |
3697 | if (! note |
3698 | || rtx_varies_p (XEXP (note, 0), 0) |
3699 | || (reg_equiv[regno].replacement |
3700 | && ! rtx_equal_p (XEXP (note, 0), |
3701 | reg_equiv[regno].replacement))) |
3702 | { |
3703 | no_equiv (reg: dest, store: set, NULL); |
3704 | continue; |
3705 | } |
3706 | |
3707 | list = reg_equiv[regno].init_insns; |
3708 | for (; list; list = list->next ()) |
3709 | { |
3710 | rtx note_tmp; |
3711 | rtx_insn *insn_tmp; |
3712 | |
3713 | insn_tmp = list->insn (); |
3714 | note_tmp = find_reg_note (insn_tmp, REG_EQUAL, NULL_RTX); |
3715 | gcc_assert (note_tmp); |
3716 | if (! rtx_equal_p (XEXP (note, 0), XEXP (note_tmp, 0))) |
3717 | { |
3718 | equal_p = false; |
3719 | break; |
3720 | } |
3721 | } |
3722 | |
3723 | if (! equal_p) |
3724 | { |
3725 | no_equiv (reg: dest, store: set, NULL); |
3726 | continue; |
3727 | } |
3728 | } |
3729 | |
3730 | /* Record this insn as initializing this register. */ |
3731 | reg_equiv[regno].init_insns |
3732 | = gen_rtx_INSN_LIST (VOIDmode, insn, reg_equiv[regno].init_insns); |
3733 | |
3734 | /* If this register is known to be equal to a constant, record that |
3735 | it is always equivalent to the constant. |
3736 | Note that it is possible to have a register use before |
3737 | the def in loops (see gcc.c-torture/execute/pr79286.c) |
3738 | where the reg is undefined on first use. If the def insn |
3739 | won't trap we can use it as an equivalence, effectively |
3740 | choosing the "undefined" value for the reg to be the |
3741 | same as the value set by the def. */ |
3742 | if (DF_REG_DEF_COUNT (regno) == 1 |
3743 | && note |
3744 | && !rtx_varies_p (XEXP (note, 0), 0) |
3745 | && (!may_trap_or_fault_p (XEXP (note, 0)) |
3746 | || def_dominates_uses (regno))) |
3747 | { |
3748 | rtx note_value = XEXP (note, 0); |
3749 | remove_note (insn, note); |
3750 | set_unique_reg_note (insn, REG_EQUIV, note_value); |
3751 | } |
3752 | |
3753 | /* If this insn introduces a "constant" register, decrease the priority |
3754 | of that register. Record this insn if the register is only used once |
3755 | more and the equivalence value is the same as our source. |
3756 | |
3757 | The latter condition is checked for two reasons: First, it is an |
3758 | indication that it may be more efficient to actually emit the insn |
3759 | as written (if no registers are available, reload will substitute |
3760 | the equivalence). Secondly, it avoids problems with any registers |
3761 | dying in this insn whose death notes would be missed. |
3762 | |
3763 | If we don't have a REG_EQUIV note, see if this insn is loading |
3764 | a register used only in one basic block from a MEM. If so, and the |
3765 | MEM remains unchanged for the life of the register, add a REG_EQUIV |
3766 | note. */ |
3767 | note = find_reg_note (insn, REG_EQUIV, NULL_RTX); |
3768 | |
3769 | rtx replacement = NULL_RTX; |
3770 | if (note) |
3771 | replacement = XEXP (note, 0); |
3772 | else if (REG_BASIC_BLOCK (regno) >= NUM_FIXED_BLOCKS |
3773 | && MEM_P (SET_SRC (set))) |
3774 | { |
3775 | enum valid_equiv validity; |
3776 | validity = validate_equiv_mem (start: insn, reg: dest, SET_SRC (set)); |
3777 | if (validity != valid_none) |
3778 | { |
3779 | replacement = copy_rtx (SET_SRC (set)); |
3780 | if (validity == valid_reload) |
3781 | { |
3782 | note = set_unique_reg_note (insn, REG_EQUIV, replacement); |
3783 | } |
3784 | else if (ira_use_lra_p) |
3785 | { |
3786 | /* We still can use this equivalence for caller save |
3787 | optimization in LRA. Mark this. */ |
3788 | ira_reg_equiv[regno].caller_save_p = true; |
3789 | ira_reg_equiv[regno].init_insns |
3790 | = gen_rtx_INSN_LIST (VOIDmode, insn, |
3791 | ira_reg_equiv[regno].init_insns); |
3792 | } |
3793 | } |
3794 | } |
3795 | |
3796 | /* If we haven't done so, record for reload that this is an |
3797 | equivalencing insn. */ |
3798 | if (note && !reg_equiv[regno].is_arg_equivalence) |
3799 | ira_reg_equiv[regno].init_insns |
3800 | = gen_rtx_INSN_LIST (VOIDmode, insn, |
3801 | ira_reg_equiv[regno].init_insns); |
3802 | |
3803 | if (replacement) |
3804 | { |
3805 | reg_equiv[regno].replacement = replacement; |
3806 | reg_equiv[regno].src_p = &SET_SRC (set); |
3807 | reg_equiv[regno].loop_depth = (short) loop_depth; |
3808 | |
3809 | /* Don't mess with things live during setjmp. */ |
3810 | if (optimize && !bitmap_bit_p (setjmp_crosses, regno)) |
3811 | { |
3812 | /* If the register is referenced exactly twice, meaning it is |
3813 | set once and used once, indicate that the reference may be |
3814 | replaced by the equivalence we computed above. Do this |
3815 | even if the register is only used in one block so that |
3816 | dependencies can be handled where the last register is |
3817 | used in a different block (i.e. HIGH / LO_SUM sequences) |
3818 | and to reduce the number of registers alive across |
3819 | calls. */ |
3820 | |
3821 | if (REG_N_REFS (regno) == 2 |
3822 | && (rtx_equal_p (replacement, src) |
3823 | || ! equiv_init_varies_p (x: src)) |
3824 | && NONJUMP_INSN_P (insn) |
3825 | && equiv_init_movable_p (x: PATTERN (insn), regno)) |
3826 | reg_equiv[regno].replace = 1; |
3827 | } |
3828 | } |
3829 | } |
3830 | } |
3831 | } |
3832 | |
3833 | /* For insns that set a MEM to the contents of a REG that is only used |
3834 | in a single basic block, see if the register is always equivalent |
3835 | to that memory location and if moving the store from INSN to the |
3836 | insn that sets REG is safe. If so, put a REG_EQUIV note on the |
3837 | initializing insn. */ |
3838 | static void |
3839 | add_store_equivs (void) |
3840 | { |
3841 | auto_bitmap seen_insns; |
3842 | |
3843 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
3844 | { |
3845 | rtx set, src, dest; |
3846 | unsigned regno; |
3847 | rtx_insn *init_insn; |
3848 | |
3849 | bitmap_set_bit (seen_insns, INSN_UID (insn)); |
3850 | |
3851 | if (! INSN_P (insn)) |
3852 | continue; |
3853 | |
3854 | set = single_set (insn); |
3855 | if (! set) |
3856 | continue; |
3857 | |
3858 | dest = SET_DEST (set); |
3859 | src = SET_SRC (set); |
3860 | |
3861 | /* Don't add a REG_EQUIV note if the insn already has one. The existing |
3862 | REG_EQUIV is likely more useful than the one we are adding. */ |
3863 | if (MEM_P (dest) && REG_P (src) |
3864 | && (regno = REGNO (src)) >= FIRST_PSEUDO_REGISTER |
3865 | && REG_BASIC_BLOCK (regno) >= NUM_FIXED_BLOCKS |
3866 | && DF_REG_DEF_COUNT (regno) == 1 |
3867 | && ! reg_equiv[regno].pdx_subregs |
3868 | && reg_equiv[regno].init_insns != NULL |
3869 | && (init_insn = reg_equiv[regno].init_insns->insn ()) != 0 |
3870 | && bitmap_bit_p (seen_insns, INSN_UID (insn: init_insn)) |
3871 | && ! find_reg_note (init_insn, REG_EQUIV, NULL_RTX) |
3872 | && validate_equiv_mem (start: init_insn, reg: src, memref: dest) == valid_reload |
3873 | && ! memref_used_between_p (memref: dest, start: init_insn, end: insn) |
3874 | /* Attaching a REG_EQUIV note will fail if INIT_INSN has |
3875 | multiple sets. */ |
3876 | && set_unique_reg_note (init_insn, REG_EQUIV, copy_rtx (dest))) |
3877 | { |
3878 | /* This insn makes the equivalence, not the one initializing |
3879 | the register. */ |
3880 | ira_reg_equiv[regno].init_insns |
3881 | = gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX); |
3882 | df_notes_rescan (init_insn); |
3883 | if (dump_file) |
3884 | fprintf (stream: dump_file, |
3885 | format: "Adding REG_EQUIV to insn %d for source of insn %d\n" , |
3886 | INSN_UID (insn: init_insn), |
3887 | INSN_UID (insn)); |
3888 | } |
3889 | } |
3890 | } |
3891 | |
3892 | /* Scan all regs killed in an insn to see if any of them are registers |
3893 | only used that once. If so, see if we can replace the reference |
3894 | with the equivalent form. If we can, delete the initializing |
3895 | reference and this register will go away. If we can't replace the |
3896 | reference, and the initializing reference is within the same loop |
3897 | (or in an inner loop), then move the register initialization just |
3898 | before the use, so that they are in the same basic block. */ |
3899 | static void |
3900 | combine_and_move_insns (void) |
3901 | { |
3902 | auto_bitmap cleared_regs; |
3903 | int max = max_reg_num (); |
3904 | |
3905 | for (int regno = FIRST_PSEUDO_REGISTER; regno < max; regno++) |
3906 | { |
3907 | if (!reg_equiv[regno].replace) |
3908 | continue; |
3909 | |
3910 | rtx_insn *use_insn = 0; |
3911 | for (df_ref use = DF_REG_USE_CHAIN (regno); |
3912 | use; |
3913 | use = DF_REF_NEXT_REG (use)) |
3914 | if (DF_REF_INSN_INFO (use)) |
3915 | { |
3916 | if (DEBUG_INSN_P (DF_REF_INSN (use))) |
3917 | continue; |
3918 | gcc_assert (!use_insn); |
3919 | use_insn = DF_REF_INSN (use); |
3920 | } |
3921 | gcc_assert (use_insn); |
3922 | |
3923 | /* Don't substitute into jumps. indirect_jump_optimize does |
3924 | this for anything we are prepared to handle. */ |
3925 | if (JUMP_P (use_insn)) |
3926 | continue; |
3927 | |
3928 | /* Also don't substitute into a conditional trap insn -- it can become |
3929 | an unconditional trap, and that is a flow control insn. */ |
3930 | if (GET_CODE (PATTERN (use_insn)) == TRAP_IF) |
3931 | continue; |
3932 | |
3933 | df_ref def = DF_REG_DEF_CHAIN (regno); |
3934 | gcc_assert (DF_REG_DEF_COUNT (regno) == 1 && DF_REF_INSN_INFO (def)); |
3935 | rtx_insn *def_insn = DF_REF_INSN (def); |
3936 | |
3937 | /* We may not move instructions that can throw, since that |
3938 | changes basic block boundaries and we are not prepared to |
3939 | adjust the CFG to match. */ |
3940 | if (can_throw_internal (def_insn)) |
3941 | continue; |
3942 | |
3943 | /* Instructions with multiple sets can only be moved if DF analysis is |
3944 | performed for all of the registers set. See PR91052. */ |
3945 | if (multiple_sets (def_insn)) |
3946 | continue; |
3947 | |
3948 | basic_block use_bb = BLOCK_FOR_INSN (insn: use_insn); |
3949 | basic_block def_bb = BLOCK_FOR_INSN (insn: def_insn); |
3950 | if (bb_loop_depth (use_bb) > bb_loop_depth (def_bb)) |
3951 | continue; |
3952 | |
3953 | if (asm_noperands (PATTERN (insn: def_insn)) < 0 |
3954 | && validate_replace_rtx (regno_reg_rtx[regno], |
3955 | *reg_equiv[regno].src_p, use_insn)) |
3956 | { |
3957 | rtx link; |
3958 | /* Append the REG_DEAD notes from def_insn. */ |
3959 | for (rtx *p = ®_NOTES (def_insn); (link = *p) != 0; ) |
3960 | { |
3961 | if (REG_NOTE_KIND (XEXP (link, 0)) == REG_DEAD) |
3962 | { |
3963 | *p = XEXP (link, 1); |
3964 | XEXP (link, 1) = REG_NOTES (use_insn); |
3965 | REG_NOTES (use_insn) = link; |
3966 | } |
3967 | else |
3968 | p = &XEXP (link, 1); |
3969 | } |
3970 | |
3971 | remove_death (regno, use_insn); |
3972 | SET_REG_N_REFS (regno, 0); |
3973 | REG_FREQ (regno) = 0; |
3974 | df_ref use; |
3975 | FOR_EACH_INSN_USE (use, def_insn) |
3976 | { |
3977 | unsigned int use_regno = DF_REF_REGNO (use); |
3978 | if (!HARD_REGISTER_NUM_P (use_regno)) |
3979 | reg_equiv[use_regno].replace = 0; |
3980 | } |
3981 | |
3982 | delete_insn (def_insn); |
3983 | |
3984 | reg_equiv[regno].init_insns = NULL; |
3985 | ira_reg_equiv[regno].init_insns = NULL; |
3986 | bitmap_set_bit (cleared_regs, regno); |
3987 | } |
3988 | |
3989 | /* Move the initialization of the register to just before |
3990 | USE_INSN. Update the flow information. */ |
3991 | else if (prev_nondebug_insn (use_insn) != def_insn) |
3992 | { |
3993 | rtx_insn *new_insn; |
3994 | |
3995 | new_insn = emit_insn_before (PATTERN (insn: def_insn), use_insn); |
3996 | REG_NOTES (new_insn) = REG_NOTES (def_insn); |
3997 | REG_NOTES (def_insn) = 0; |
3998 | /* Rescan it to process the notes. */ |
3999 | df_insn_rescan (new_insn); |
4000 | |
4001 | /* Make sure this insn is recognized before reload begins, |
4002 | otherwise eliminate_regs_in_insn will die. */ |
4003 | INSN_CODE (new_insn) = INSN_CODE (def_insn); |
4004 | |
4005 | delete_insn (def_insn); |
4006 | |
4007 | XEXP (reg_equiv[regno].init_insns, 0) = new_insn; |
4008 | |
4009 | REG_BASIC_BLOCK (regno) = use_bb->index; |
4010 | REG_N_CALLS_CROSSED (regno) = 0; |
4011 | |
4012 | if (use_insn == BB_HEAD (use_bb)) |
4013 | BB_HEAD (use_bb) = new_insn; |
4014 | |
4015 | /* We know regno dies in use_insn, but inside a loop |
4016 | REG_DEAD notes might be missing when def_insn was in |
4017 | another basic block. However, when we move def_insn into |
4018 | this bb we'll definitely get a REG_DEAD note and reload |
4019 | will see the death. It's possible that update_equiv_regs |
4020 | set up an equivalence referencing regno for a reg set by |
4021 | use_insn, when regno was seen as non-local. Now that |
4022 | regno is local to this block, and dies, such an |
4023 | equivalence is invalid. */ |
4024 | if (find_reg_note (use_insn, REG_EQUIV, regno_reg_rtx[regno])) |
4025 | { |
4026 | rtx set = single_set (insn: use_insn); |
4027 | if (set && REG_P (SET_DEST (set))) |
4028 | no_equiv (SET_DEST (set), store: set, NULL); |
4029 | } |
4030 | |
4031 | ira_reg_equiv[regno].init_insns |
4032 | = gen_rtx_INSN_LIST (VOIDmode, new_insn, NULL_RTX); |
4033 | bitmap_set_bit (cleared_regs, regno); |
4034 | } |
4035 | } |
4036 | |
4037 | if (!bitmap_empty_p (map: cleared_regs)) |
4038 | { |
4039 | basic_block bb; |
4040 | |
4041 | FOR_EACH_BB_FN (bb, cfun) |
4042 | { |
4043 | bitmap_and_compl_into (DF_LR_IN (bb), cleared_regs); |
4044 | bitmap_and_compl_into (DF_LR_OUT (bb), cleared_regs); |
4045 | if (!df_live) |
4046 | continue; |
4047 | bitmap_and_compl_into (DF_LIVE_IN (bb), cleared_regs); |
4048 | bitmap_and_compl_into (DF_LIVE_OUT (bb), cleared_regs); |
4049 | } |
4050 | |
4051 | /* Last pass - adjust debug insns referencing cleared regs. */ |
4052 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
4053 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
4054 | if (DEBUG_BIND_INSN_P (insn)) |
4055 | { |
4056 | rtx old_loc = INSN_VAR_LOCATION_LOC (insn); |
4057 | INSN_VAR_LOCATION_LOC (insn) |
4058 | = simplify_replace_fn_rtx (old_loc, NULL_RTX, |
4059 | fn: adjust_cleared_regs, |
4060 | (void *) cleared_regs); |
4061 | if (old_loc != INSN_VAR_LOCATION_LOC (insn)) |
4062 | df_insn_rescan (insn); |
4063 | } |
4064 | } |
4065 | } |
4066 | |
4067 | /* A pass over indirect jumps, converting simple cases to direct jumps. |
4068 | Combine does this optimization too, but only within a basic block. */ |
4069 | static void |
4070 | indirect_jump_optimize (void) |
4071 | { |
4072 | basic_block bb; |
4073 | bool rebuild_p = false; |
4074 | |
4075 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
4076 | { |
4077 | rtx_insn *insn = BB_END (bb); |
4078 | if (!JUMP_P (insn) |
4079 | || find_reg_note (insn, REG_NON_LOCAL_GOTO, NULL_RTX)) |
4080 | continue; |
4081 | |
4082 | rtx x = pc_set (insn); |
4083 | if (!x || !REG_P (SET_SRC (x))) |
4084 | continue; |
4085 | |
4086 | int regno = REGNO (SET_SRC (x)); |
4087 | if (DF_REG_DEF_COUNT (regno) == 1) |
4088 | { |
4089 | df_ref def = DF_REG_DEF_CHAIN (regno); |
4090 | if (!DF_REF_IS_ARTIFICIAL (def)) |
4091 | { |
4092 | rtx_insn *def_insn = DF_REF_INSN (def); |
4093 | rtx lab = NULL_RTX; |
4094 | rtx set = single_set (insn: def_insn); |
4095 | if (set && GET_CODE (SET_SRC (set)) == LABEL_REF) |
4096 | lab = SET_SRC (set); |
4097 | else |
4098 | { |
4099 | rtx eqnote = find_reg_note (def_insn, REG_EQUAL, NULL_RTX); |
4100 | if (eqnote && GET_CODE (XEXP (eqnote, 0)) == LABEL_REF) |
4101 | lab = XEXP (eqnote, 0); |
4102 | } |
4103 | if (lab && validate_replace_rtx (SET_SRC (x), lab, insn)) |
4104 | rebuild_p = true; |
4105 | } |
4106 | } |
4107 | } |
4108 | |
4109 | if (rebuild_p) |
4110 | { |
4111 | timevar_push (tv: TV_JUMP); |
4112 | rebuild_jump_labels (get_insns ()); |
4113 | if (purge_all_dead_edges ()) |
4114 | delete_unreachable_blocks (); |
4115 | timevar_pop (tv: TV_JUMP); |
4116 | } |
4117 | } |
4118 | |
4119 | /* Set up fields memory, constant, and invariant from init_insns in |
4120 | the structures of array ira_reg_equiv. */ |
4121 | static void |
4122 | setup_reg_equiv (void) |
4123 | { |
4124 | int i; |
4125 | rtx_insn_list *elem, *prev_elem, *next_elem; |
4126 | rtx_insn *insn; |
4127 | rtx set, x; |
4128 | |
4129 | for (i = FIRST_PSEUDO_REGISTER; i < ira_reg_equiv_len; i++) |
4130 | for (prev_elem = NULL, elem = ira_reg_equiv[i].init_insns; |
4131 | elem; |
4132 | prev_elem = elem, elem = next_elem) |
4133 | { |
4134 | next_elem = elem->next (); |
4135 | insn = elem->insn (); |
4136 | set = single_set (insn); |
4137 | |
4138 | /* Init insns can set up equivalence when the reg is a destination or |
4139 | a source (in this case the destination is memory). */ |
4140 | if (set != 0 && (REG_P (SET_DEST (set)) || REG_P (SET_SRC (set)))) |
4141 | { |
4142 | if ((x = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != NULL) |
4143 | { |
4144 | x = XEXP (x, 0); |
4145 | if (REG_P (SET_DEST (set)) |
4146 | && REGNO (SET_DEST (set)) == (unsigned int) i |
4147 | && ! rtx_equal_p (SET_SRC (set), x) && MEM_P (x)) |
4148 | { |
4149 | /* This insn reporting the equivalence but |
4150 | actually not setting it. Remove it from the |
4151 | list. */ |
4152 | if (prev_elem == NULL) |
4153 | ira_reg_equiv[i].init_insns = next_elem; |
4154 | else |
4155 | XEXP (prev_elem, 1) = next_elem; |
4156 | elem = prev_elem; |
4157 | } |
4158 | } |
4159 | else if (REG_P (SET_DEST (set)) |
4160 | && REGNO (SET_DEST (set)) == (unsigned int) i) |
4161 | x = SET_SRC (set); |
4162 | else |
4163 | { |
4164 | gcc_assert (REG_P (SET_SRC (set)) |
4165 | && REGNO (SET_SRC (set)) == (unsigned int) i); |
4166 | x = SET_DEST (set); |
4167 | } |
4168 | if (! function_invariant_p (x) |
4169 | || ! flag_pic |
4170 | /* A function invariant is often CONSTANT_P but may |
4171 | include a register. We promise to only pass |
4172 | CONSTANT_P objects to LEGITIMATE_PIC_OPERAND_P. */ |
4173 | || (CONSTANT_P (x) && LEGITIMATE_PIC_OPERAND_P (x))) |
4174 | { |
4175 | /* It can happen that a REG_EQUIV note contains a MEM |
4176 | that is not a legitimate memory operand. As later |
4177 | stages of reload assume that all addresses found in |
4178 | the lra_regno_equiv_* arrays were originally |
4179 | legitimate, we ignore such REG_EQUIV notes. */ |
4180 | if (memory_operand (x, VOIDmode)) |
4181 | { |
4182 | ira_reg_equiv[i].defined_p = !ira_reg_equiv[i].caller_save_p; |
4183 | ira_reg_equiv[i].memory = x; |
4184 | continue; |
4185 | } |
4186 | else if (function_invariant_p (x)) |
4187 | { |
4188 | machine_mode mode; |
4189 | |
4190 | mode = GET_MODE (SET_DEST (set)); |
4191 | if (GET_CODE (x) == PLUS |
4192 | || x == frame_pointer_rtx || x == arg_pointer_rtx) |
4193 | /* This is PLUS of frame pointer and a constant, |
4194 | or fp, or argp. */ |
4195 | ira_reg_equiv[i].invariant = x; |
4196 | else if (targetm.legitimate_constant_p (mode, x)) |
4197 | ira_reg_equiv[i].constant = x; |
4198 | else |
4199 | { |
4200 | ira_reg_equiv[i].memory = force_const_mem (mode, x); |
4201 | if (ira_reg_equiv[i].memory == NULL_RTX) |
4202 | { |
4203 | ira_reg_equiv[i].defined_p = false; |
4204 | ira_reg_equiv[i].caller_save_p = false; |
4205 | ira_reg_equiv[i].init_insns = NULL; |
4206 | break; |
4207 | } |
4208 | } |
4209 | ira_reg_equiv[i].defined_p = true; |
4210 | continue; |
4211 | } |
4212 | } |
4213 | } |
4214 | ira_reg_equiv[i].defined_p = false; |
4215 | ira_reg_equiv[i].caller_save_p = false; |
4216 | ira_reg_equiv[i].init_insns = NULL; |
4217 | break; |
4218 | } |
4219 | } |
4220 | |
4221 | |
4222 | |
4223 | /* Print chain C to FILE. */ |
4224 | static void |
4225 | print_insn_chain (FILE *file, class insn_chain *c) |
4226 | { |
4227 | fprintf (stream: file, format: "insn=%d, " , INSN_UID (insn: c->insn)); |
4228 | bitmap_print (file, &c->live_throughout, "live_throughout: " , ", " ); |
4229 | bitmap_print (file, &c->dead_or_set, "dead_or_set: " , "\n" ); |
4230 | } |
4231 | |
4232 | |
4233 | /* Print all reload_insn_chains to FILE. */ |
4234 | static void |
4235 | print_insn_chains (FILE *file) |
4236 | { |
4237 | class insn_chain *c; |
4238 | for (c = reload_insn_chain; c ; c = c->next) |
4239 | print_insn_chain (file, c); |
4240 | } |
4241 | |
4242 | /* Return true if pseudo REGNO should be added to set live_throughout |
4243 | or dead_or_set of the insn chains for reload consideration. */ |
4244 | static bool |
4245 | pseudo_for_reload_consideration_p (int regno) |
4246 | { |
4247 | /* Consider spilled pseudos too for IRA because they still have a |
4248 | chance to get hard-registers in the reload when IRA is used. */ |
4249 | return (reg_renumber[regno] >= 0 || ira_conflicts_p); |
4250 | } |
4251 | |
4252 | /* Return true if we can track the individual bytes of subreg X. |
4253 | When returning true, set *OUTER_SIZE to the number of bytes in |
4254 | X itself, *INNER_SIZE to the number of bytes in the inner register |
4255 | and *START to the offset of the first byte. */ |
4256 | static bool |
4257 | get_subreg_tracking_sizes (rtx x, HOST_WIDE_INT *outer_size, |
4258 | HOST_WIDE_INT *inner_size, HOST_WIDE_INT *start) |
4259 | { |
4260 | rtx reg = regno_reg_rtx[REGNO (SUBREG_REG (x))]; |
4261 | return (GET_MODE_SIZE (GET_MODE (x)).is_constant (const_value: outer_size) |
4262 | && GET_MODE_SIZE (GET_MODE (reg)).is_constant (const_value: inner_size) |
4263 | && SUBREG_BYTE (x).is_constant (const_value: start)); |
4264 | } |
4265 | |
4266 | /* Init LIVE_SUBREGS[ALLOCNUM] and LIVE_SUBREGS_USED[ALLOCNUM] for |
4267 | a register with SIZE bytes, making the register live if INIT_VALUE. */ |
4268 | static void |
4269 | init_live_subregs (bool init_value, sbitmap *live_subregs, |
4270 | bitmap live_subregs_used, int allocnum, int size) |
4271 | { |
4272 | gcc_assert (size > 0); |
4273 | |
4274 | /* Been there, done that. */ |
4275 | if (bitmap_bit_p (live_subregs_used, allocnum)) |
4276 | return; |
4277 | |
4278 | /* Create a new one. */ |
4279 | if (live_subregs[allocnum] == NULL) |
4280 | live_subregs[allocnum] = sbitmap_alloc (size); |
4281 | |
4282 | /* If the entire reg was live before blasting into subregs, we need |
4283 | to init all of the subregs to ones else init to 0. */ |
4284 | if (init_value) |
4285 | bitmap_ones (live_subregs[allocnum]); |
4286 | else |
4287 | bitmap_clear (live_subregs[allocnum]); |
4288 | |
4289 | bitmap_set_bit (live_subregs_used, allocnum); |
4290 | } |
4291 | |
4292 | /* Walk the insns of the current function and build reload_insn_chain, |
4293 | and record register life information. */ |
4294 | static void |
4295 | build_insn_chain (void) |
4296 | { |
4297 | unsigned int i; |
4298 | class insn_chain **p = &reload_insn_chain; |
4299 | basic_block bb; |
4300 | class insn_chain *c = NULL; |
4301 | class insn_chain *next = NULL; |
4302 | auto_bitmap live_relevant_regs; |
4303 | auto_bitmap elim_regset; |
4304 | /* live_subregs is a vector used to keep accurate information about |
4305 | which hardregs are live in multiword pseudos. live_subregs and |
4306 | live_subregs_used are indexed by pseudo number. The live_subreg |
4307 | entry for a particular pseudo is only used if the corresponding |
4308 | element is non zero in live_subregs_used. The sbitmap size of |
4309 | live_subreg[allocno] is number of bytes that the pseudo can |
4310 | occupy. */ |
4311 | sbitmap *live_subregs = XCNEWVEC (sbitmap, max_regno); |
4312 | auto_bitmap live_subregs_used; |
4313 | |
4314 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
4315 | if (TEST_HARD_REG_BIT (set: eliminable_regset, bit: i)) |
4316 | bitmap_set_bit (elim_regset, i); |
4317 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
4318 | { |
4319 | bitmap_iterator bi; |
4320 | rtx_insn *insn; |
4321 | |
4322 | CLEAR_REG_SET (live_relevant_regs); |
4323 | bitmap_clear (live_subregs_used); |
4324 | |
4325 | EXECUTE_IF_SET_IN_BITMAP (df_get_live_out (bb), 0, i, bi) |
4326 | { |
4327 | if (i >= FIRST_PSEUDO_REGISTER) |
4328 | break; |
4329 | bitmap_set_bit (live_relevant_regs, i); |
4330 | } |
4331 | |
4332 | EXECUTE_IF_SET_IN_BITMAP (df_get_live_out (bb), |
4333 | FIRST_PSEUDO_REGISTER, i, bi) |
4334 | { |
4335 | if (pseudo_for_reload_consideration_p (regno: i)) |
4336 | bitmap_set_bit (live_relevant_regs, i); |
4337 | } |
4338 | |
4339 | FOR_BB_INSNS_REVERSE (bb, insn) |
4340 | { |
4341 | if (!NOTE_P (insn) && !BARRIER_P (insn)) |
4342 | { |
4343 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
4344 | df_ref def, use; |
4345 | |
4346 | c = new_insn_chain (); |
4347 | c->next = next; |
4348 | next = c; |
4349 | *p = c; |
4350 | p = &c->prev; |
4351 | |
4352 | c->insn = insn; |
4353 | c->block = bb->index; |
4354 | |
4355 | if (NONDEBUG_INSN_P (insn)) |
4356 | FOR_EACH_INSN_INFO_DEF (def, insn_info) |
4357 | { |
4358 | unsigned int regno = DF_REF_REGNO (def); |
4359 | |
4360 | /* Ignore may clobbers because these are generated |
4361 | from calls. However, every other kind of def is |
4362 | added to dead_or_set. */ |
4363 | if (!DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) |
4364 | { |
4365 | if (regno < FIRST_PSEUDO_REGISTER) |
4366 | { |
4367 | if (!fixed_regs[regno]) |
4368 | bitmap_set_bit (&c->dead_or_set, regno); |
4369 | } |
4370 | else if (pseudo_for_reload_consideration_p (regno)) |
4371 | bitmap_set_bit (&c->dead_or_set, regno); |
4372 | } |
4373 | |
4374 | if ((regno < FIRST_PSEUDO_REGISTER |
4375 | || reg_renumber[regno] >= 0 |
4376 | || ira_conflicts_p) |
4377 | && (!DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL))) |
4378 | { |
4379 | rtx reg = DF_REF_REG (def); |
4380 | HOST_WIDE_INT outer_size, inner_size, start; |
4381 | |
4382 | /* We can usually track the liveness of individual |
4383 | bytes within a subreg. The only exceptions are |
4384 | subregs wrapped in ZERO_EXTRACTs and subregs whose |
4385 | size is not known; in those cases we need to be |
4386 | conservative and treat the definition as a partial |
4387 | definition of the full register rather than a full |
4388 | definition of a specific part of the register. */ |
4389 | if (GET_CODE (reg) == SUBREG |
4390 | && !DF_REF_FLAGS_IS_SET (def, DF_REF_ZERO_EXTRACT) |
4391 | && get_subreg_tracking_sizes (x: reg, outer_size: &outer_size, |
4392 | inner_size: &inner_size, start: &start)) |
4393 | { |
4394 | HOST_WIDE_INT last = start + outer_size; |
4395 | |
4396 | init_live_subregs |
4397 | (init_value: bitmap_bit_p (live_relevant_regs, regno), |
4398 | live_subregs, live_subregs_used, allocnum: regno, |
4399 | size: inner_size); |
4400 | |
4401 | if (!DF_REF_FLAGS_IS_SET |
4402 | (def, DF_REF_STRICT_LOW_PART)) |
4403 | { |
4404 | /* Expand the range to cover entire words. |
4405 | Bytes added here are "don't care". */ |
4406 | start |
4407 | = start / UNITS_PER_WORD * UNITS_PER_WORD; |
4408 | last = ((last + UNITS_PER_WORD - 1) |
4409 | / UNITS_PER_WORD * UNITS_PER_WORD); |
4410 | } |
4411 | |
4412 | /* Ignore the paradoxical bits. */ |
4413 | if (last > SBITMAP_SIZE (live_subregs[regno])) |
4414 | last = SBITMAP_SIZE (live_subregs[regno]); |
4415 | |
4416 | while (start < last) |
4417 | { |
4418 | bitmap_clear_bit (map: live_subregs[regno], bitno: start); |
4419 | start++; |
4420 | } |
4421 | |
4422 | if (bitmap_empty_p (live_subregs[regno])) |
4423 | { |
4424 | bitmap_clear_bit (live_subregs_used, regno); |
4425 | bitmap_clear_bit (live_relevant_regs, regno); |
4426 | } |
4427 | else |
4428 | /* Set live_relevant_regs here because |
4429 | that bit has to be true to get us to |
4430 | look at the live_subregs fields. */ |
4431 | bitmap_set_bit (live_relevant_regs, regno); |
4432 | } |
4433 | else |
4434 | { |
4435 | /* DF_REF_PARTIAL is generated for |
4436 | subregs, STRICT_LOW_PART, and |
4437 | ZERO_EXTRACT. We handle the subreg |
4438 | case above so here we have to keep from |
4439 | modeling the def as a killing def. */ |
4440 | if (!DF_REF_FLAGS_IS_SET (def, DF_REF_PARTIAL)) |
4441 | { |
4442 | bitmap_clear_bit (live_subregs_used, regno); |
4443 | bitmap_clear_bit (live_relevant_regs, regno); |
4444 | } |
4445 | } |
4446 | } |
4447 | } |
4448 | |
4449 | bitmap_and_compl_into (live_relevant_regs, elim_regset); |
4450 | bitmap_copy (&c->live_throughout, live_relevant_regs); |
4451 | |
4452 | if (NONDEBUG_INSN_P (insn)) |
4453 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
4454 | { |
4455 | unsigned int regno = DF_REF_REGNO (use); |
4456 | rtx reg = DF_REF_REG (use); |
4457 | |
4458 | /* DF_REF_READ_WRITE on a use means that this use |
4459 | is fabricated from a def that is a partial set |
4460 | to a multiword reg. Here, we only model the |
4461 | subreg case that is not wrapped in ZERO_EXTRACT |
4462 | precisely so we do not need to look at the |
4463 | fabricated use. */ |
4464 | if (DF_REF_FLAGS_IS_SET (use, DF_REF_READ_WRITE) |
4465 | && !DF_REF_FLAGS_IS_SET (use, DF_REF_ZERO_EXTRACT) |
4466 | && DF_REF_FLAGS_IS_SET (use, DF_REF_SUBREG)) |
4467 | continue; |
4468 | |
4469 | /* Add the last use of each var to dead_or_set. */ |
4470 | if (!bitmap_bit_p (live_relevant_regs, regno)) |
4471 | { |
4472 | if (regno < FIRST_PSEUDO_REGISTER) |
4473 | { |
4474 | if (!fixed_regs[regno]) |
4475 | bitmap_set_bit (&c->dead_or_set, regno); |
4476 | } |
4477 | else if (pseudo_for_reload_consideration_p (regno)) |
4478 | bitmap_set_bit (&c->dead_or_set, regno); |
4479 | } |
4480 | |
4481 | if (regno < FIRST_PSEUDO_REGISTER |
4482 | || pseudo_for_reload_consideration_p (regno)) |
4483 | { |
4484 | HOST_WIDE_INT outer_size, inner_size, start; |
4485 | if (GET_CODE (reg) == SUBREG |
4486 | && !DF_REF_FLAGS_IS_SET (use, |
4487 | DF_REF_SIGN_EXTRACT |
4488 | | DF_REF_ZERO_EXTRACT) |
4489 | && get_subreg_tracking_sizes (x: reg, outer_size: &outer_size, |
4490 | inner_size: &inner_size, start: &start)) |
4491 | { |
4492 | HOST_WIDE_INT last = start + outer_size; |
4493 | |
4494 | init_live_subregs |
4495 | (init_value: bitmap_bit_p (live_relevant_regs, regno), |
4496 | live_subregs, live_subregs_used, allocnum: regno, |
4497 | size: inner_size); |
4498 | |
4499 | /* Ignore the paradoxical bits. */ |
4500 | if (last > SBITMAP_SIZE (live_subregs[regno])) |
4501 | last = SBITMAP_SIZE (live_subregs[regno]); |
4502 | |
4503 | while (start < last) |
4504 | { |
4505 | bitmap_set_bit (map: live_subregs[regno], bitno: start); |
4506 | start++; |
4507 | } |
4508 | } |
4509 | else |
4510 | /* Resetting the live_subregs_used is |
4511 | effectively saying do not use the subregs |
4512 | because we are reading the whole |
4513 | pseudo. */ |
4514 | bitmap_clear_bit (live_subregs_used, regno); |
4515 | bitmap_set_bit (live_relevant_regs, regno); |
4516 | } |
4517 | } |
4518 | } |
4519 | } |
4520 | |
4521 | /* FIXME!! The following code is a disaster. Reload needs to see the |
4522 | labels and jump tables that are just hanging out in between |
4523 | the basic blocks. See pr33676. */ |
4524 | insn = BB_HEAD (bb); |
4525 | |
4526 | /* Skip over the barriers and cruft. */ |
4527 | while (insn && (BARRIER_P (insn) || NOTE_P (insn) |
4528 | || BLOCK_FOR_INSN (insn) == bb)) |
4529 | insn = PREV_INSN (insn); |
4530 | |
4531 | /* While we add anything except barriers and notes, the focus is |
4532 | to get the labels and jump tables into the |
4533 | reload_insn_chain. */ |
4534 | while (insn) |
4535 | { |
4536 | if (!NOTE_P (insn) && !BARRIER_P (insn)) |
4537 | { |
4538 | if (BLOCK_FOR_INSN (insn)) |
4539 | break; |
4540 | |
4541 | c = new_insn_chain (); |
4542 | c->next = next; |
4543 | next = c; |
4544 | *p = c; |
4545 | p = &c->prev; |
4546 | |
4547 | /* The block makes no sense here, but it is what the old |
4548 | code did. */ |
4549 | c->block = bb->index; |
4550 | c->insn = insn; |
4551 | bitmap_copy (&c->live_throughout, live_relevant_regs); |
4552 | } |
4553 | insn = PREV_INSN (insn); |
4554 | } |
4555 | } |
4556 | |
4557 | reload_insn_chain = c; |
4558 | *p = NULL; |
4559 | |
4560 | for (i = 0; i < (unsigned int) max_regno; i++) |
4561 | if (live_subregs[i] != NULL) |
4562 | sbitmap_free (map: live_subregs[i]); |
4563 | free (ptr: live_subregs); |
4564 | |
4565 | if (dump_file) |
4566 | print_insn_chains (file: dump_file); |
4567 | } |
4568 | |
4569 | /* Examine the rtx found in *LOC, which is read or written to as determined |
4570 | by TYPE. Return false if we find a reason why an insn containing this |
4571 | rtx should not be moved (such as accesses to non-constant memory), true |
4572 | otherwise. */ |
4573 | static bool |
4574 | rtx_moveable_p (rtx *loc, enum op_type type) |
4575 | { |
4576 | const char *fmt; |
4577 | rtx x = *loc; |
4578 | int i, j; |
4579 | |
4580 | enum rtx_code code = GET_CODE (x); |
4581 | switch (code) |
4582 | { |
4583 | case CONST: |
4584 | CASE_CONST_ANY: |
4585 | case SYMBOL_REF: |
4586 | case LABEL_REF: |
4587 | return true; |
4588 | |
4589 | case PC: |
4590 | return type == OP_IN; |
4591 | |
4592 | case REG: |
4593 | if (x == frame_pointer_rtx) |
4594 | return true; |
4595 | if (HARD_REGISTER_P (x)) |
4596 | return false; |
4597 | |
4598 | return true; |
4599 | |
4600 | case MEM: |
4601 | if (type == OP_IN && MEM_READONLY_P (x)) |
4602 | return rtx_moveable_p (loc: &XEXP (x, 0), type: OP_IN); |
4603 | return false; |
4604 | |
4605 | case SET: |
4606 | return (rtx_moveable_p (loc: &SET_SRC (x), type: OP_IN) |
4607 | && rtx_moveable_p (loc: &SET_DEST (x), type: OP_OUT)); |
4608 | |
4609 | case STRICT_LOW_PART: |
4610 | return rtx_moveable_p (loc: &XEXP (x, 0), type: OP_OUT); |
4611 | |
4612 | case ZERO_EXTRACT: |
4613 | case SIGN_EXTRACT: |
4614 | return (rtx_moveable_p (loc: &XEXP (x, 0), type) |
4615 | && rtx_moveable_p (loc: &XEXP (x, 1), type: OP_IN) |
4616 | && rtx_moveable_p (loc: &XEXP (x, 2), type: OP_IN)); |
4617 | |
4618 | case CLOBBER: |
4619 | return rtx_moveable_p (loc: &SET_DEST (x), type: OP_OUT); |
4620 | |
4621 | case UNSPEC_VOLATILE: |
4622 | /* It is a bad idea to consider insns with such rtl |
4623 | as moveable ones. The insn scheduler also considers them as barrier |
4624 | for a reason. */ |
4625 | return false; |
4626 | |
4627 | case ASM_OPERANDS: |
4628 | /* The same is true for volatile asm: it has unknown side effects, it |
4629 | cannot be moved at will. */ |
4630 | if (MEM_VOLATILE_P (x)) |
4631 | return false; |
4632 | |
4633 | default: |
4634 | break; |
4635 | } |
4636 | |
4637 | fmt = GET_RTX_FORMAT (code); |
4638 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
4639 | { |
4640 | if (fmt[i] == 'e') |
4641 | { |
4642 | if (!rtx_moveable_p (loc: &XEXP (x, i), type)) |
4643 | return false; |
4644 | } |
4645 | else if (fmt[i] == 'E') |
4646 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
4647 | { |
4648 | if (!rtx_moveable_p (loc: &XVECEXP (x, i, j), type)) |
4649 | return false; |
4650 | } |
4651 | } |
4652 | return true; |
4653 | } |
4654 | |
4655 | /* A wrapper around dominated_by_p, which uses the information in UID_LUID |
4656 | to give dominance relationships between two insns I1 and I2. */ |
4657 | static bool |
4658 | insn_dominated_by_p (rtx i1, rtx i2, int *uid_luid) |
4659 | { |
4660 | basic_block bb1 = BLOCK_FOR_INSN (insn: i1); |
4661 | basic_block bb2 = BLOCK_FOR_INSN (insn: i2); |
4662 | |
4663 | if (bb1 == bb2) |
4664 | return uid_luid[INSN_UID (insn: i2)] < uid_luid[INSN_UID (insn: i1)]; |
4665 | return dominated_by_p (CDI_DOMINATORS, bb1, bb2); |
4666 | } |
4667 | |
4668 | /* Record the range of register numbers added by find_moveable_pseudos. */ |
4669 | int first_moveable_pseudo, last_moveable_pseudo; |
4670 | |
4671 | /* These two vectors hold data for every register added by |
4672 | find_movable_pseudos, with index 0 holding data for the |
4673 | first_moveable_pseudo. */ |
4674 | /* The original home register. */ |
4675 | static vec<rtx> pseudo_replaced_reg; |
4676 | |
4677 | /* Look for instances where we have an instruction that is known to increase |
4678 | register pressure, and whose result is not used immediately. If it is |
4679 | possible to move the instruction downwards to just before its first use, |
4680 | split its lifetime into two ranges. We create a new pseudo to compute the |
4681 | value, and emit a move instruction just before the first use. If, after |
4682 | register allocation, the new pseudo remains unallocated, the function |
4683 | move_unallocated_pseudos then deletes the move instruction and places |
4684 | the computation just before the first use. |
4685 | |
4686 | Such a move is safe and profitable if all the input registers remain live |
4687 | and unchanged between the original computation and its first use. In such |
4688 | a situation, the computation is known to increase register pressure, and |
4689 | moving it is known to at least not worsen it. |
4690 | |
4691 | We restrict moves to only those cases where a register remains unallocated, |
4692 | in order to avoid interfering too much with the instruction schedule. As |
4693 | an exception, we may move insns which only modify their input register |
4694 | (typically induction variables), as this increases the freedom for our |
4695 | intended transformation, and does not limit the second instruction |
4696 | scheduler pass. */ |
4697 | |
4698 | static void |
4699 | find_moveable_pseudos (void) |
4700 | { |
4701 | unsigned i; |
4702 | int max_regs = max_reg_num (); |
4703 | int max_uid = get_max_uid (); |
4704 | basic_block bb; |
4705 | int *uid_luid = XNEWVEC (int, max_uid); |
4706 | rtx_insn **closest_uses = XNEWVEC (rtx_insn *, max_regs); |
4707 | /* A set of registers which are live but not modified throughout a block. */ |
4708 | bitmap_head *bb_transp_live = XNEWVEC (bitmap_head, |
4709 | last_basic_block_for_fn (cfun)); |
4710 | /* A set of registers which only exist in a given basic block. */ |
4711 | bitmap_head *bb_local = XNEWVEC (bitmap_head, |
4712 | last_basic_block_for_fn (cfun)); |
4713 | /* A set of registers which are set once, in an instruction that can be |
4714 | moved freely downwards, but are otherwise transparent to a block. */ |
4715 | bitmap_head *bb_moveable_reg_sets = XNEWVEC (bitmap_head, |
4716 | last_basic_block_for_fn (cfun)); |
4717 | auto_bitmap live, used, set, interesting, unusable_as_input; |
4718 | bitmap_iterator bi; |
4719 | |
4720 | first_moveable_pseudo = max_regs; |
4721 | pseudo_replaced_reg.release (); |
4722 | pseudo_replaced_reg.safe_grow_cleared (len: max_regs, exact: true); |
4723 | |
4724 | df_analyze (); |
4725 | calculate_dominance_info (CDI_DOMINATORS); |
4726 | |
4727 | i = 0; |
4728 | FOR_EACH_BB_FN (bb, cfun) |
4729 | { |
4730 | rtx_insn *insn; |
4731 | bitmap transp = bb_transp_live + bb->index; |
4732 | bitmap moveable = bb_moveable_reg_sets + bb->index; |
4733 | bitmap local = bb_local + bb->index; |
4734 | |
4735 | bitmap_initialize (head: local, obstack: 0); |
4736 | bitmap_initialize (head: transp, obstack: 0); |
4737 | bitmap_initialize (head: moveable, obstack: 0); |
4738 | bitmap_copy (live, df_get_live_out (bb)); |
4739 | bitmap_and_into (live, df_get_live_in (bb)); |
4740 | bitmap_copy (transp, live); |
4741 | bitmap_clear (moveable); |
4742 | bitmap_clear (live); |
4743 | bitmap_clear (used); |
4744 | bitmap_clear (set); |
4745 | FOR_BB_INSNS (bb, insn) |
4746 | if (NONDEBUG_INSN_P (insn)) |
4747 | { |
4748 | df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
4749 | df_ref def, use; |
4750 | |
4751 | uid_luid[INSN_UID (insn)] = i++; |
4752 | |
4753 | def = df_single_def (info: insn_info); |
4754 | use = df_single_use (info: insn_info); |
4755 | if (use |
4756 | && def |
4757 | && DF_REF_REGNO (use) == DF_REF_REGNO (def) |
4758 | && !bitmap_bit_p (set, DF_REF_REGNO (use)) |
4759 | && rtx_moveable_p (loc: &PATTERN (insn), type: OP_IN)) |
4760 | { |
4761 | unsigned regno = DF_REF_REGNO (use); |
4762 | bitmap_set_bit (moveable, regno); |
4763 | bitmap_set_bit (set, regno); |
4764 | bitmap_set_bit (used, regno); |
4765 | bitmap_clear_bit (transp, regno); |
4766 | continue; |
4767 | } |
4768 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
4769 | { |
4770 | unsigned regno = DF_REF_REGNO (use); |
4771 | bitmap_set_bit (used, regno); |
4772 | if (bitmap_clear_bit (moveable, regno)) |
4773 | bitmap_clear_bit (transp, regno); |
4774 | } |
4775 | |
4776 | FOR_EACH_INSN_INFO_DEF (def, insn_info) |
4777 | { |
4778 | unsigned regno = DF_REF_REGNO (def); |
4779 | bitmap_set_bit (set, regno); |
4780 | bitmap_clear_bit (transp, regno); |
4781 | bitmap_clear_bit (moveable, regno); |
4782 | } |
4783 | } |
4784 | } |
4785 | |
4786 | FOR_EACH_BB_FN (bb, cfun) |
4787 | { |
4788 | bitmap local = bb_local + bb->index; |
4789 | rtx_insn *insn; |
4790 | |
4791 | FOR_BB_INSNS (bb, insn) |
4792 | if (NONDEBUG_INSN_P (insn)) |
4793 | { |
4794 | df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
4795 | rtx_insn *def_insn; |
4796 | rtx closest_use, note; |
4797 | df_ref def, use; |
4798 | unsigned regno; |
4799 | bool all_dominated, all_local; |
4800 | machine_mode mode; |
4801 | |
4802 | def = df_single_def (info: insn_info); |
4803 | /* There must be exactly one def in this insn. */ |
4804 | if (!def || !single_set (insn)) |
4805 | continue; |
4806 | /* This must be the only definition of the reg. We also limit |
4807 | which modes we deal with so that we can assume we can generate |
4808 | move instructions. */ |
4809 | regno = DF_REF_REGNO (def); |
4810 | mode = GET_MODE (DF_REF_REG (def)); |
4811 | if (DF_REG_DEF_COUNT (regno) != 1 |
4812 | || !DF_REF_INSN_INFO (def) |
4813 | || HARD_REGISTER_NUM_P (regno) |
4814 | || DF_REG_EQ_USE_COUNT (regno) > 0 |
4815 | || (!INTEGRAL_MODE_P (mode) |
4816 | && !FLOAT_MODE_P (mode) |
4817 | && !OPAQUE_MODE_P (mode))) |
4818 | continue; |
4819 | def_insn = DF_REF_INSN (def); |
4820 | |
4821 | for (note = REG_NOTES (def_insn); note; note = XEXP (note, 1)) |
4822 | if (REG_NOTE_KIND (note) == REG_EQUIV && MEM_P (XEXP (note, 0))) |
4823 | break; |
4824 | |
4825 | if (note) |
4826 | { |
4827 | if (dump_file) |
4828 | fprintf (stream: dump_file, format: "Ignoring reg %d, has equiv memory\n" , |
4829 | regno); |
4830 | bitmap_set_bit (unusable_as_input, regno); |
4831 | continue; |
4832 | } |
4833 | |
4834 | use = DF_REG_USE_CHAIN (regno); |
4835 | all_dominated = true; |
4836 | all_local = true; |
4837 | closest_use = NULL_RTX; |
4838 | for (; use; use = DF_REF_NEXT_REG (use)) |
4839 | { |
4840 | rtx_insn *insn; |
4841 | if (!DF_REF_INSN_INFO (use)) |
4842 | { |
4843 | all_dominated = false; |
4844 | all_local = false; |
4845 | break; |
4846 | } |
4847 | insn = DF_REF_INSN (use); |
4848 | if (DEBUG_INSN_P (insn)) |
4849 | continue; |
4850 | if (BLOCK_FOR_INSN (insn) != BLOCK_FOR_INSN (insn: def_insn)) |
4851 | all_local = false; |
4852 | if (!insn_dominated_by_p (i1: insn, i2: def_insn, uid_luid)) |
4853 | all_dominated = false; |
4854 | if (closest_use != insn && closest_use != const0_rtx) |
4855 | { |
4856 | if (closest_use == NULL_RTX) |
4857 | closest_use = insn; |
4858 | else if (insn_dominated_by_p (i1: closest_use, i2: insn, uid_luid)) |
4859 | closest_use = insn; |
4860 | else if (!insn_dominated_by_p (i1: insn, i2: closest_use, uid_luid)) |
4861 | closest_use = const0_rtx; |
4862 | } |
4863 | } |
4864 | if (!all_dominated) |
4865 | { |
4866 | if (dump_file) |
4867 | fprintf (stream: dump_file, format: "Reg %d not all uses dominated by set\n" , |
4868 | regno); |
4869 | continue; |
4870 | } |
4871 | if (all_local) |
4872 | bitmap_set_bit (local, regno); |
4873 | if (closest_use == const0_rtx || closest_use == NULL |
4874 | || next_nonnote_nondebug_insn (def_insn) == closest_use) |
4875 | { |
4876 | if (dump_file) |
4877 | fprintf (stream: dump_file, format: "Reg %d uninteresting%s\n" , regno, |
4878 | closest_use == const0_rtx || closest_use == NULL |
4879 | ? " (no unique first use)" : "" ); |
4880 | continue; |
4881 | } |
4882 | |
4883 | bitmap_set_bit (interesting, regno); |
4884 | /* If we get here, we know closest_use is a non-NULL insn |
4885 | (as opposed to const_0_rtx). */ |
4886 | closest_uses[regno] = as_a <rtx_insn *> (p: closest_use); |
4887 | |
4888 | if (dump_file && (all_local || all_dominated)) |
4889 | { |
4890 | fprintf (stream: dump_file, format: "Reg %u:" , regno); |
4891 | if (all_local) |
4892 | fprintf (stream: dump_file, format: " local to bb %d" , bb->index); |
4893 | if (all_dominated) |
4894 | fprintf (stream: dump_file, format: " def dominates all uses" ); |
4895 | if (closest_use != const0_rtx) |
4896 | fprintf (stream: dump_file, format: " has unique first use" ); |
4897 | fputs (s: "\n" , stream: dump_file); |
4898 | } |
4899 | } |
4900 | } |
4901 | |
4902 | EXECUTE_IF_SET_IN_BITMAP (interesting, 0, i, bi) |
4903 | { |
4904 | df_ref def = DF_REG_DEF_CHAIN (i); |
4905 | rtx_insn *def_insn = DF_REF_INSN (def); |
4906 | basic_block def_block = BLOCK_FOR_INSN (insn: def_insn); |
4907 | bitmap def_bb_local = bb_local + def_block->index; |
4908 | bitmap def_bb_moveable = bb_moveable_reg_sets + def_block->index; |
4909 | bitmap def_bb_transp = bb_transp_live + def_block->index; |
4910 | bool local_to_bb_p = bitmap_bit_p (def_bb_local, i); |
4911 | rtx_insn *use_insn = closest_uses[i]; |
4912 | df_ref use; |
4913 | bool all_ok = true; |
4914 | bool all_transp = true; |
4915 | |
4916 | if (!REG_P (DF_REF_REG (def))) |
4917 | continue; |
4918 | |
4919 | if (!local_to_bb_p) |
4920 | { |
4921 | if (dump_file) |
4922 | fprintf (stream: dump_file, format: "Reg %u not local to one basic block\n" , |
4923 | i); |
4924 | continue; |
4925 | } |
4926 | if (reg_equiv_init (i) != NULL_RTX) |
4927 | { |
4928 | if (dump_file) |
4929 | fprintf (stream: dump_file, format: "Ignoring reg %u with equiv init insn\n" , |
4930 | i); |
4931 | continue; |
4932 | } |
4933 | if (!rtx_moveable_p (loc: &PATTERN (insn: def_insn), type: OP_IN)) |
4934 | { |
4935 | if (dump_file) |
4936 | fprintf (stream: dump_file, format: "Found def insn %d for %d to be not moveable\n" , |
4937 | INSN_UID (insn: def_insn), i); |
4938 | continue; |
4939 | } |
4940 | if (dump_file) |
4941 | fprintf (stream: dump_file, format: "Examining insn %d, def for %d\n" , |
4942 | INSN_UID (insn: def_insn), i); |
4943 | FOR_EACH_INSN_USE (use, def_insn) |
4944 | { |
4945 | unsigned regno = DF_REF_REGNO (use); |
4946 | if (bitmap_bit_p (unusable_as_input, regno)) |
4947 | { |
4948 | all_ok = false; |
4949 | if (dump_file) |
4950 | fprintf (stream: dump_file, format: " found unusable input reg %u.\n" , regno); |
4951 | break; |
4952 | } |
4953 | if (!bitmap_bit_p (def_bb_transp, regno)) |
4954 | { |
4955 | if (bitmap_bit_p (def_bb_moveable, regno) |
4956 | && !control_flow_insn_p (use_insn)) |
4957 | { |
4958 | if (modified_between_p (DF_REF_REG (use), def_insn, use_insn)) |
4959 | { |
4960 | rtx_insn *x = NEXT_INSN (insn: def_insn); |
4961 | while (!modified_in_p (DF_REF_REG (use), x)) |
4962 | { |
4963 | gcc_assert (x != use_insn); |
4964 | x = NEXT_INSN (insn: x); |
4965 | } |
4966 | if (dump_file) |
4967 | fprintf (stream: dump_file, format: " input reg %u modified but insn %d moveable\n" , |
4968 | regno, INSN_UID (insn: x)); |
4969 | emit_insn_after (PATTERN (insn: x), use_insn); |
4970 | set_insn_deleted (x); |
4971 | } |
4972 | else |
4973 | { |
4974 | if (dump_file) |
4975 | fprintf (stream: dump_file, format: " input reg %u modified between def and use\n" , |
4976 | regno); |
4977 | all_transp = false; |
4978 | } |
4979 | } |
4980 | else |
4981 | all_transp = false; |
4982 | } |
4983 | } |
4984 | if (!all_ok) |
4985 | continue; |
4986 | if (!dbg_cnt (index: ira_move)) |
4987 | break; |
4988 | if (dump_file) |
4989 | fprintf (stream: dump_file, format: " all ok%s\n" , all_transp ? " and transp" : "" ); |
4990 | |
4991 | if (all_transp) |
4992 | { |
4993 | rtx def_reg = DF_REF_REG (def); |
4994 | rtx newreg = ira_create_new_reg (def_reg); |
4995 | if (validate_change (def_insn, DF_REF_REAL_LOC (def), newreg, 0)) |
4996 | { |
4997 | unsigned nregno = REGNO (newreg); |
4998 | emit_insn_before (gen_move_insn (def_reg, newreg), use_insn); |
4999 | nregno -= max_regs; |
5000 | pseudo_replaced_reg[nregno] = def_reg; |
5001 | } |
5002 | } |
5003 | } |
5004 | |
5005 | FOR_EACH_BB_FN (bb, cfun) |
5006 | { |
5007 | bitmap_clear (bb_local + bb->index); |
5008 | bitmap_clear (bb_transp_live + bb->index); |
5009 | bitmap_clear (bb_moveable_reg_sets + bb->index); |
5010 | } |
5011 | free (ptr: uid_luid); |
5012 | free (ptr: closest_uses); |
5013 | free (ptr: bb_local); |
5014 | free (ptr: bb_transp_live); |
5015 | free (ptr: bb_moveable_reg_sets); |
5016 | |
5017 | last_moveable_pseudo = max_reg_num (); |
5018 | |
5019 | fix_reg_equiv_init (); |
5020 | expand_reg_info (); |
5021 | regstat_free_n_sets_and_refs (); |
5022 | regstat_free_ri (); |
5023 | regstat_init_n_sets_and_refs (); |
5024 | regstat_compute_ri (); |
5025 | free_dominance_info (CDI_DOMINATORS); |
5026 | } |
5027 | |
5028 | /* If SET pattern SET is an assignment from a hard register to a pseudo which |
5029 | is live at CALL_DOM (if non-NULL, otherwise this check is omitted), return |
5030 | the destination. Otherwise return NULL. */ |
5031 | |
5032 | static rtx |
5033 | interesting_dest_for_shprep_1 (rtx set, basic_block call_dom) |
5034 | { |
5035 | rtx src = SET_SRC (set); |
5036 | rtx dest = SET_DEST (set); |
5037 | if (!REG_P (src) || !HARD_REGISTER_P (src) |
5038 | || !REG_P (dest) || HARD_REGISTER_P (dest) |
5039 | || (call_dom && !bitmap_bit_p (df_get_live_in (bb: call_dom), REGNO (dest)))) |
5040 | return NULL; |
5041 | return dest; |
5042 | } |
5043 | |
5044 | /* If insn is interesting for parameter range-splitting shrink-wrapping |
5045 | preparation, i.e. it is a single set from a hard register to a pseudo, which |
5046 | is live at CALL_DOM (if non-NULL, otherwise this check is omitted), or a |
5047 | parallel statement with only one such statement, return the destination. |
5048 | Otherwise return NULL. */ |
5049 | |
5050 | static rtx |
5051 | interesting_dest_for_shprep (rtx_insn *insn, basic_block call_dom) |
5052 | { |
5053 | if (!INSN_P (insn)) |
5054 | return NULL; |
5055 | rtx pat = PATTERN (insn); |
5056 | if (GET_CODE (pat) == SET) |
5057 | return interesting_dest_for_shprep_1 (set: pat, call_dom); |
5058 | |
5059 | if (GET_CODE (pat) != PARALLEL) |
5060 | return NULL; |
5061 | rtx ret = NULL; |
5062 | for (int i = 0; i < XVECLEN (pat, 0); i++) |
5063 | { |
5064 | rtx sub = XVECEXP (pat, 0, i); |
5065 | if (GET_CODE (sub) == USE || GET_CODE (sub) == CLOBBER) |
5066 | continue; |
5067 | if (GET_CODE (sub) != SET |
5068 | || side_effects_p (sub)) |
5069 | return NULL; |
5070 | rtx dest = interesting_dest_for_shprep_1 (set: sub, call_dom); |
5071 | if (dest && ret) |
5072 | return NULL; |
5073 | if (dest) |
5074 | ret = dest; |
5075 | } |
5076 | return ret; |
5077 | } |
5078 | |
5079 | /* Split live ranges of pseudos that are loaded from hard registers in the |
5080 | first BB in a BB that dominates all non-sibling call if such a BB can be |
5081 | found and is not in a loop. Return true if the function has made any |
5082 | changes. */ |
5083 | |
5084 | static bool |
5085 | split_live_ranges_for_shrink_wrap (void) |
5086 | { |
5087 | basic_block bb, call_dom = NULL; |
5088 | basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
5089 | rtx_insn *insn, *last_interesting_insn = NULL; |
5090 | auto_bitmap need_new, reachable; |
5091 | vec<basic_block> queue; |
5092 | |
5093 | if (!SHRINK_WRAPPING_ENABLED) |
5094 | return false; |
5095 | |
5096 | queue.create (n_basic_blocks_for_fn (cfun)); |
5097 | |
5098 | FOR_EACH_BB_FN (bb, cfun) |
5099 | FOR_BB_INSNS (bb, insn) |
5100 | if (CALL_P (insn) && !SIBLING_CALL_P (insn)) |
5101 | { |
5102 | if (bb == first) |
5103 | { |
5104 | queue.release (); |
5105 | return false; |
5106 | } |
5107 | |
5108 | bitmap_set_bit (need_new, bb->index); |
5109 | bitmap_set_bit (reachable, bb->index); |
5110 | queue.quick_push (obj: bb); |
5111 | break; |
5112 | } |
5113 | |
5114 | if (queue.is_empty ()) |
5115 | { |
5116 | queue.release (); |
5117 | return false; |
5118 | } |
5119 | |
5120 | while (!queue.is_empty ()) |
5121 | { |
5122 | edge e; |
5123 | edge_iterator ei; |
5124 | |
5125 | bb = queue.pop (); |
5126 | FOR_EACH_EDGE (e, ei, bb->succs) |
5127 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
5128 | && bitmap_set_bit (reachable, e->dest->index)) |
5129 | queue.quick_push (obj: e->dest); |
5130 | } |
5131 | queue.release (); |
5132 | |
5133 | FOR_BB_INSNS (first, insn) |
5134 | { |
5135 | rtx dest = interesting_dest_for_shprep (insn, NULL); |
5136 | if (!dest) |
5137 | continue; |
5138 | |
5139 | if (DF_REG_DEF_COUNT (REGNO (dest)) > 1) |
5140 | return false; |
5141 | |
5142 | for (df_ref use = DF_REG_USE_CHAIN (REGNO(dest)); |
5143 | use; |
5144 | use = DF_REF_NEXT_REG (use)) |
5145 | { |
5146 | int ubbi = DF_REF_BB (use)->index; |
5147 | if (bitmap_bit_p (reachable, ubbi)) |
5148 | bitmap_set_bit (need_new, ubbi); |
5149 | } |
5150 | last_interesting_insn = insn; |
5151 | } |
5152 | |
5153 | if (!last_interesting_insn) |
5154 | return false; |
5155 | |
5156 | call_dom = nearest_common_dominator_for_set (CDI_DOMINATORS, need_new); |
5157 | if (call_dom == first) |
5158 | return false; |
5159 | |
5160 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); |
5161 | while (bb_loop_depth (call_dom) > 0) |
5162 | call_dom = get_immediate_dominator (CDI_DOMINATORS, call_dom); |
5163 | loop_optimizer_finalize (); |
5164 | |
5165 | if (call_dom == first) |
5166 | return false; |
5167 | |
5168 | calculate_dominance_info (CDI_POST_DOMINATORS); |
5169 | if (dominated_by_p (CDI_POST_DOMINATORS, first, call_dom)) |
5170 | { |
5171 | free_dominance_info (CDI_POST_DOMINATORS); |
5172 | return false; |
5173 | } |
5174 | free_dominance_info (CDI_POST_DOMINATORS); |
5175 | |
5176 | if (dump_file) |
5177 | fprintf (stream: dump_file, format: "Will split live ranges of parameters at BB %i\n" , |
5178 | call_dom->index); |
5179 | |
5180 | bool ret = false; |
5181 | FOR_BB_INSNS (first, insn) |
5182 | { |
5183 | rtx dest = interesting_dest_for_shprep (insn, call_dom); |
5184 | if (!dest || dest == pic_offset_table_rtx) |
5185 | continue; |
5186 | |
5187 | bool need_newreg = false; |
5188 | df_ref use, next; |
5189 | for (use = DF_REG_USE_CHAIN (REGNO (dest)); use; use = next) |
5190 | { |
5191 | rtx_insn *uin = DF_REF_INSN (use); |
5192 | next = DF_REF_NEXT_REG (use); |
5193 | |
5194 | if (DEBUG_INSN_P (uin)) |
5195 | continue; |
5196 | |
5197 | basic_block ubb = BLOCK_FOR_INSN (insn: uin); |
5198 | if (ubb == call_dom |
5199 | || dominated_by_p (CDI_DOMINATORS, ubb, call_dom)) |
5200 | { |
5201 | need_newreg = true; |
5202 | break; |
5203 | } |
5204 | } |
5205 | |
5206 | if (need_newreg) |
5207 | { |
5208 | rtx newreg = ira_create_new_reg (dest); |
5209 | |
5210 | for (use = DF_REG_USE_CHAIN (REGNO (dest)); use; use = next) |
5211 | { |
5212 | rtx_insn *uin = DF_REF_INSN (use); |
5213 | next = DF_REF_NEXT_REG (use); |
5214 | |
5215 | basic_block ubb = BLOCK_FOR_INSN (insn: uin); |
5216 | if (ubb == call_dom |
5217 | || dominated_by_p (CDI_DOMINATORS, ubb, call_dom)) |
5218 | validate_change (uin, DF_REF_REAL_LOC (use), newreg, true); |
5219 | } |
5220 | |
5221 | rtx_insn *new_move = gen_move_insn (newreg, dest); |
5222 | emit_insn_after (new_move, bb_note (call_dom)); |
5223 | if (dump_file) |
5224 | { |
5225 | fprintf (stream: dump_file, format: "Split live-range of register " ); |
5226 | print_rtl_single (dump_file, dest); |
5227 | } |
5228 | ret = true; |
5229 | } |
5230 | |
5231 | if (insn == last_interesting_insn) |
5232 | break; |
5233 | } |
5234 | apply_change_group (); |
5235 | return ret; |
5236 | } |
5237 | |
5238 | /* Perform the second half of the transformation started in |
5239 | find_moveable_pseudos. We look for instances where the newly introduced |
5240 | pseudo remains unallocated, and remove it by moving the definition to |
5241 | just before its use, replacing the move instruction generated by |
5242 | find_moveable_pseudos. */ |
5243 | static void |
5244 | move_unallocated_pseudos (void) |
5245 | { |
5246 | int i; |
5247 | for (i = first_moveable_pseudo; i < last_moveable_pseudo; i++) |
5248 | if (reg_renumber[i] < 0) |
5249 | { |
5250 | int idx = i - first_moveable_pseudo; |
5251 | rtx other_reg = pseudo_replaced_reg[idx]; |
5252 | /* The iterating range [first_moveable_pseudo, last_moveable_pseudo) |
5253 | covers every new pseudo created in find_moveable_pseudos, |
5254 | regardless of the validation with it is successful or not. |
5255 | So we need to skip the pseudos which were used in those failed |
5256 | validations to avoid unexpected DF info and consequent ICE. |
5257 | We only set pseudo_replaced_reg[] when the validation is successful |
5258 | in find_moveable_pseudos, it's enough to check it here. */ |
5259 | if (!other_reg) |
5260 | continue; |
5261 | rtx_insn *def_insn = DF_REF_INSN (DF_REG_DEF_CHAIN (i)); |
5262 | /* The use must follow all definitions of OTHER_REG, so we can |
5263 | insert the new definition immediately after any of them. */ |
5264 | df_ref other_def = DF_REG_DEF_CHAIN (REGNO (other_reg)); |
5265 | rtx_insn *move_insn = DF_REF_INSN (other_def); |
5266 | rtx_insn *newinsn = emit_insn_after (PATTERN (insn: def_insn), move_insn); |
5267 | rtx set; |
5268 | int success; |
5269 | |
5270 | if (dump_file) |
5271 | fprintf (stream: dump_file, format: "moving def of %d (insn %d now) " , |
5272 | REGNO (other_reg), INSN_UID (insn: def_insn)); |
5273 | |
5274 | delete_insn (move_insn); |
5275 | while ((other_def = DF_REG_DEF_CHAIN (REGNO (other_reg)))) |
5276 | delete_insn (DF_REF_INSN (other_def)); |
5277 | delete_insn (def_insn); |
5278 | |
5279 | set = single_set (insn: newinsn); |
5280 | success = validate_change (newinsn, &SET_DEST (set), other_reg, 0); |
5281 | gcc_assert (success); |
5282 | if (dump_file) |
5283 | fprintf (stream: dump_file, format: " %d) rather than keep unallocated replacement %d\n" , |
5284 | INSN_UID (insn: newinsn), i); |
5285 | SET_REG_N_REFS (i, 0); |
5286 | } |
5287 | |
5288 | first_moveable_pseudo = last_moveable_pseudo = 0; |
5289 | } |
5290 | |
5291 | |
5292 | |
5293 | /* Code dealing with scratches (changing them onto |
5294 | pseudos and restoring them from the pseudos). |
5295 | |
5296 | We change scratches into pseudos at the beginning of IRA to |
5297 | simplify dealing with them (conflicts, hard register assignments). |
5298 | |
5299 | If the pseudo denoting scratch was spilled it means that we do not |
5300 | need a hard register for it. Such pseudos are transformed back to |
5301 | scratches at the end of LRA. */ |
5302 | |
5303 | /* Description of location of a former scratch operand. */ |
5304 | struct sloc |
5305 | { |
5306 | rtx_insn *insn; /* Insn where the scratch was. */ |
5307 | int nop; /* Number of the operand which was a scratch. */ |
5308 | unsigned regno; /* regno gnerated instead of scratch */ |
5309 | int icode; /* Original icode from which scratch was removed. */ |
5310 | }; |
5311 | |
5312 | typedef struct sloc *sloc_t; |
5313 | |
5314 | /* Locations of the former scratches. */ |
5315 | static vec<sloc_t> scratches; |
5316 | |
5317 | /* Bitmap of scratch regnos. */ |
5318 | static bitmap_head scratch_bitmap; |
5319 | |
5320 | /* Bitmap of scratch operands. */ |
5321 | static bitmap_head scratch_operand_bitmap; |
5322 | |
5323 | /* Return true if pseudo REGNO is made of SCRATCH. */ |
5324 | bool |
5325 | ira_former_scratch_p (int regno) |
5326 | { |
5327 | return bitmap_bit_p (&scratch_bitmap, regno); |
5328 | } |
5329 | |
5330 | /* Return true if the operand NOP of INSN is a former scratch. */ |
5331 | bool |
5332 | ira_former_scratch_operand_p (rtx_insn *insn, int nop) |
5333 | { |
5334 | return bitmap_bit_p (&scratch_operand_bitmap, |
5335 | INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0; |
5336 | } |
5337 | |
5338 | /* Register operand NOP in INSN as a former scratch. It will be |
5339 | changed to scratch back, if it is necessary, at the LRA end. */ |
5340 | void |
5341 | ira_register_new_scratch_op (rtx_insn *insn, int nop, int icode) |
5342 | { |
5343 | rtx op = *recog_data.operand_loc[nop]; |
5344 | sloc_t loc = XNEW (struct sloc); |
5345 | ira_assert (REG_P (op)); |
5346 | loc->insn = insn; |
5347 | loc->nop = nop; |
5348 | loc->regno = REGNO (op); |
5349 | loc->icode = icode; |
5350 | scratches.safe_push (obj: loc); |
5351 | bitmap_set_bit (&scratch_bitmap, REGNO (op)); |
5352 | bitmap_set_bit (&scratch_operand_bitmap, |
5353 | INSN_UID (insn) * MAX_RECOG_OPERANDS + nop); |
5354 | add_reg_note (insn, REG_UNUSED, op); |
5355 | } |
5356 | |
5357 | /* Return true if string STR contains constraint 'X'. */ |
5358 | static bool |
5359 | contains_X_constraint_p (const char *str) |
5360 | { |
5361 | int c; |
5362 | |
5363 | while ((c = *str)) |
5364 | { |
5365 | str += CONSTRAINT_LEN (c, str); |
5366 | if (c == 'X') return true; |
5367 | } |
5368 | return false; |
5369 | } |
5370 | |
5371 | /* Change INSN's scratches into pseudos and save their location. |
5372 | Return true if we changed any scratch. */ |
5373 | bool |
5374 | ira_remove_insn_scratches (rtx_insn *insn, bool all_p, FILE *dump_file, |
5375 | rtx (*get_reg) (rtx original)) |
5376 | { |
5377 | int i; |
5378 | bool insn_changed_p; |
5379 | rtx reg, *loc; |
5380 | |
5381 | extract_insn (insn); |
5382 | insn_changed_p = false; |
5383 | for (i = 0; i < recog_data.n_operands; i++) |
5384 | { |
5385 | loc = recog_data.operand_loc[i]; |
5386 | if (GET_CODE (*loc) == SCRATCH && GET_MODE (*loc) != VOIDmode) |
5387 | { |
5388 | if (! all_p && contains_X_constraint_p (str: recog_data.constraints[i])) |
5389 | continue; |
5390 | insn_changed_p = true; |
5391 | *loc = reg = get_reg (*loc); |
5392 | ira_register_new_scratch_op (insn, nop: i, INSN_CODE (insn)); |
5393 | if (ira_dump_file != NULL) |
5394 | fprintf (stream: dump_file, |
5395 | format: "Removing SCRATCH to p%u in insn #%u (nop %d)\n" , |
5396 | REGNO (reg), INSN_UID (insn), i); |
5397 | } |
5398 | } |
5399 | return insn_changed_p; |
5400 | } |
5401 | |
5402 | /* Return new register of the same mode as ORIGINAL. Used in |
5403 | remove_scratches. */ |
5404 | static rtx |
5405 | get_scratch_reg (rtx original) |
5406 | { |
5407 | return gen_reg_rtx (GET_MODE (original)); |
5408 | } |
5409 | |
5410 | /* Change scratches into pseudos and save their location. Return true |
5411 | if we changed any scratch. */ |
5412 | static bool |
5413 | remove_scratches (void) |
5414 | { |
5415 | bool change_p = false; |
5416 | basic_block bb; |
5417 | rtx_insn *insn; |
5418 | |
5419 | scratches.create (nelems: get_max_uid ()); |
5420 | bitmap_initialize (head: &scratch_bitmap, obstack: ®_obstack); |
5421 | bitmap_initialize (head: &scratch_operand_bitmap, obstack: ®_obstack); |
5422 | FOR_EACH_BB_FN (bb, cfun) |
5423 | FOR_BB_INSNS (bb, insn) |
5424 | if (INSN_P (insn) |
5425 | && ira_remove_insn_scratches (insn, all_p: false, dump_file: ira_dump_file, get_reg: get_scratch_reg)) |
5426 | { |
5427 | /* Because we might use DF, we need to keep DF info up to date. */ |
5428 | df_insn_rescan (insn); |
5429 | change_p = true; |
5430 | } |
5431 | return change_p; |
5432 | } |
5433 | |
5434 | /* Changes pseudos created by function remove_scratches onto scratches. */ |
5435 | void |
5436 | ira_restore_scratches (FILE *dump_file) |
5437 | { |
5438 | int regno, n; |
5439 | unsigned i; |
5440 | rtx *op_loc; |
5441 | sloc_t loc; |
5442 | |
5443 | for (i = 0; scratches.iterate (ix: i, ptr: &loc); i++) |
5444 | { |
5445 | /* Ignore already deleted insns. */ |
5446 | if (NOTE_P (loc->insn) |
5447 | && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED) |
5448 | continue; |
5449 | extract_insn (loc->insn); |
5450 | if (loc->icode != INSN_CODE (loc->insn)) |
5451 | { |
5452 | /* The icode doesn't match, which means the insn has been modified |
5453 | (e.g. register elimination). The scratch cannot be restored. */ |
5454 | continue; |
5455 | } |
5456 | op_loc = recog_data.operand_loc[loc->nop]; |
5457 | if (REG_P (*op_loc) |
5458 | && ((regno = REGNO (*op_loc)) >= FIRST_PSEUDO_REGISTER) |
5459 | && reg_renumber[regno] < 0) |
5460 | { |
5461 | /* It should be only case when scratch register with chosen |
5462 | constraint 'X' did not get memory or hard register. */ |
5463 | ira_assert (ira_former_scratch_p (regno)); |
5464 | *op_loc = gen_rtx_SCRATCH (GET_MODE (*op_loc)); |
5465 | for (n = 0; n < recog_data.n_dups; n++) |
5466 | *recog_data.dup_loc[n] |
5467 | = *recog_data.operand_loc[(int) recog_data.dup_num[n]]; |
5468 | if (dump_file != NULL) |
5469 | fprintf (stream: dump_file, format: "Restoring SCRATCH in insn #%u(nop %d)\n" , |
5470 | INSN_UID (insn: loc->insn), loc->nop); |
5471 | } |
5472 | } |
5473 | for (i = 0; scratches.iterate (ix: i, ptr: &loc); i++) |
5474 | free (ptr: loc); |
5475 | scratches.release (); |
5476 | bitmap_clear (&scratch_bitmap); |
5477 | bitmap_clear (&scratch_operand_bitmap); |
5478 | } |
5479 | |
5480 | |
5481 | |
5482 | /* If the backend knows where to allocate pseudos for hard |
5483 | register initial values, register these allocations now. */ |
5484 | static void |
5485 | allocate_initial_values (void) |
5486 | { |
5487 | if (targetm.allocate_initial_value) |
5488 | { |
5489 | rtx hreg, preg, x; |
5490 | int i, regno; |
5491 | |
5492 | for (i = 0; HARD_REGISTER_NUM_P (i); i++) |
5493 | { |
5494 | if (! initial_value_entry (i, &hreg, &preg)) |
5495 | break; |
5496 | |
5497 | x = targetm.allocate_initial_value (hreg); |
5498 | regno = REGNO (preg); |
5499 | if (x && REG_N_SETS (regno) <= 1) |
5500 | { |
5501 | if (MEM_P (x)) |
5502 | reg_equiv_memory_loc (regno) = x; |
5503 | else |
5504 | { |
5505 | basic_block bb; |
5506 | int new_regno; |
5507 | |
5508 | gcc_assert (REG_P (x)); |
5509 | new_regno = REGNO (x); |
5510 | reg_renumber[regno] = new_regno; |
5511 | /* Poke the regno right into regno_reg_rtx so that even |
5512 | fixed regs are accepted. */ |
5513 | SET_REGNO (preg, new_regno); |
5514 | /* Update global register liveness information. */ |
5515 | FOR_EACH_BB_FN (bb, cfun) |
5516 | { |
5517 | if (REGNO_REG_SET_P (df_get_live_in (bb), regno)) |
5518 | SET_REGNO_REG_SET (df_get_live_in (bb), new_regno); |
5519 | if (REGNO_REG_SET_P (df_get_live_out (bb), regno)) |
5520 | SET_REGNO_REG_SET (df_get_live_out (bb), new_regno); |
5521 | } |
5522 | } |
5523 | } |
5524 | } |
5525 | |
5526 | gcc_checking_assert (! initial_value_entry (FIRST_PSEUDO_REGISTER, |
5527 | &hreg, &preg)); |
5528 | } |
5529 | } |
5530 | |
5531 | |
5532 | |
5533 | |
5534 | /* True when we use LRA instead of reload pass for the current |
5535 | function. */ |
5536 | bool ira_use_lra_p; |
5537 | |
5538 | /* True if we have allocno conflicts. It is false for non-optimized |
5539 | mode or when the conflict table is too big. */ |
5540 | bool ira_conflicts_p; |
5541 | |
5542 | /* Saved between IRA and reload. */ |
5543 | static int saved_flag_ira_share_spill_slots; |
5544 | |
5545 | /* Set to true while in IRA. */ |
5546 | bool ira_in_progress = false; |
5547 | |
5548 | /* This is the main entry of IRA. */ |
5549 | static void |
5550 | ira (FILE *f) |
5551 | { |
5552 | bool loops_p; |
5553 | int ira_max_point_before_emit; |
5554 | bool saved_flag_caller_saves = flag_caller_saves; |
5555 | enum ira_region saved_flag_ira_region = flag_ira_region; |
5556 | basic_block bb; |
5557 | edge_iterator ei; |
5558 | edge e; |
5559 | bool output_jump_reload_p = false; |
5560 | |
5561 | if (ira_use_lra_p) |
5562 | { |
5563 | /* First put potential jump output reloads on the output edges |
5564 | as USE which will be removed at the end of LRA. The major |
5565 | goal is actually to create BBs for critical edges for LRA and |
5566 | populate them later by live info. In LRA it will be |
5567 | difficult to do this. */ |
5568 | FOR_EACH_BB_FN (bb, cfun) |
5569 | { |
5570 | rtx_insn *end = BB_END (bb); |
5571 | if (!JUMP_P (end)) |
5572 | continue; |
5573 | extract_insn (end); |
5574 | for (int i = 0; i < recog_data.n_operands; i++) |
5575 | if (recog_data.operand_type[i] != OP_IN) |
5576 | { |
5577 | bool skip_p = false; |
5578 | FOR_EACH_EDGE (e, ei, bb->succs) |
5579 | if (EDGE_CRITICAL_P (e) |
5580 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
5581 | && (e->flags & EDGE_ABNORMAL)) |
5582 | { |
5583 | skip_p = true; |
5584 | break; |
5585 | } |
5586 | if (skip_p) |
5587 | break; |
5588 | output_jump_reload_p = true; |
5589 | FOR_EACH_EDGE (e, ei, bb->succs) |
5590 | if (EDGE_CRITICAL_P (e) |
5591 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
5592 | { |
5593 | start_sequence (); |
5594 | /* We need to put some no-op insn here. We can |
5595 | not put a note as commit_edges insertion will |
5596 | fail. */ |
5597 | emit_insn (gen_rtx_USE (VOIDmode, const1_rtx)); |
5598 | rtx_insn *insns = get_insns (); |
5599 | end_sequence (); |
5600 | insert_insn_on_edge (insns, e); |
5601 | } |
5602 | break; |
5603 | } |
5604 | } |
5605 | if (output_jump_reload_p) |
5606 | commit_edge_insertions (); |
5607 | } |
5608 | |
5609 | if (flag_ira_verbose < 10) |
5610 | { |
5611 | internal_flag_ira_verbose = flag_ira_verbose; |
5612 | ira_dump_file = f; |
5613 | } |
5614 | else |
5615 | { |
5616 | internal_flag_ira_verbose = flag_ira_verbose - 10; |
5617 | ira_dump_file = stderr; |
5618 | } |
5619 | |
5620 | clear_bb_flags (); |
5621 | |
5622 | /* Determine if the current function is a leaf before running IRA |
5623 | since this can impact optimizations done by the prologue and |
5624 | epilogue thus changing register elimination offsets. |
5625 | Other target callbacks may use crtl->is_leaf too, including |
5626 | SHRINK_WRAPPING_ENABLED, so initialize as early as possible. */ |
5627 | crtl->is_leaf = leaf_function_p (); |
5628 | |
5629 | /* Perform target specific PIC register initialization. */ |
5630 | targetm.init_pic_reg (); |
5631 | |
5632 | ira_conflicts_p = optimize > 0; |
5633 | |
5634 | /* Determine the number of pseudos actually requiring coloring. */ |
5635 | unsigned int num_used_regs = 0; |
5636 | for (unsigned int i = FIRST_PSEUDO_REGISTER; i < DF_REG_SIZE (df); i++) |
5637 | if (DF_REG_DEF_COUNT (i) || DF_REG_USE_COUNT (i)) |
5638 | num_used_regs++; |
5639 | |
5640 | /* If there are too many pseudos and/or basic blocks (e.g. 10K pseudos and |
5641 | 10K blocks or 100K pseudos and 1K blocks) or we have too many function |
5642 | insns, we will use simplified and faster algorithms in LRA. */ |
5643 | lra_simple_p |
5644 | = (ira_use_lra_p |
5645 | && (num_used_regs >= (1U << 26) / last_basic_block_for_fn (cfun) |
5646 | /* max uid is a good evaluation of the number of insns as most |
5647 | optimizations are done on tree-SSA level. */ |
5648 | || ((uint64_t) get_max_uid () |
5649 | > (uint64_t) param_ira_simple_lra_insn_threshold * 1000))); |
5650 | |
5651 | if (lra_simple_p) |
5652 | { |
5653 | /* It permits to skip live range splitting in LRA. */ |
5654 | flag_caller_saves = false; |
5655 | /* There is no sense to do regional allocation when we use |
5656 | simplified LRA. */ |
5657 | flag_ira_region = IRA_REGION_ONE; |
5658 | ira_conflicts_p = false; |
5659 | } |
5660 | |
5661 | #ifndef IRA_NO_OBSTACK |
5662 | gcc_obstack_init (&ira_obstack); |
5663 | #endif |
5664 | bitmap_obstack_initialize (&ira_bitmap_obstack); |
5665 | |
5666 | /* LRA uses its own infrastructure to handle caller save registers. */ |
5667 | if (flag_caller_saves && !ira_use_lra_p) |
5668 | init_caller_save (); |
5669 | |
5670 | setup_prohibited_mode_move_regs (); |
5671 | decrease_live_ranges_number (); |
5672 | df_note_add_problem (); |
5673 | |
5674 | /* DF_LIVE can't be used in the register allocator, too many other |
5675 | parts of the compiler depend on using the "classic" liveness |
5676 | interpretation of the DF_LR problem. See PR38711. |
5677 | Remove the problem, so that we don't spend time updating it in |
5678 | any of the df_analyze() calls during IRA/LRA. */ |
5679 | if (optimize > 1) |
5680 | df_remove_problem (df_live); |
5681 | gcc_checking_assert (df_live == NULL); |
5682 | |
5683 | if (flag_checking) |
5684 | df->changeable_flags |= DF_VERIFY_SCHEDULED; |
5685 | |
5686 | df_analyze (); |
5687 | |
5688 | init_reg_equiv (); |
5689 | if (ira_conflicts_p) |
5690 | { |
5691 | calculate_dominance_info (CDI_DOMINATORS); |
5692 | |
5693 | if (split_live_ranges_for_shrink_wrap ()) |
5694 | df_analyze (); |
5695 | |
5696 | free_dominance_info (CDI_DOMINATORS); |
5697 | } |
5698 | |
5699 | df_clear_flags (DF_NO_INSN_RESCAN); |
5700 | |
5701 | indirect_jump_optimize (); |
5702 | if (delete_trivially_dead_insns (get_insns (), max_reg_num ())) |
5703 | df_analyze (); |
5704 | |
5705 | regstat_init_n_sets_and_refs (); |
5706 | regstat_compute_ri (); |
5707 | |
5708 | /* If we are not optimizing, then this is the only place before |
5709 | register allocation where dataflow is done. And that is needed |
5710 | to generate these warnings. */ |
5711 | if (warn_clobbered) |
5712 | generate_setjmp_warnings (); |
5713 | |
5714 | /* update_equiv_regs can use reg classes of pseudos and they are set up in |
5715 | register pressure sensitive scheduling and loop invariant motion and in |
5716 | live range shrinking. This info can become obsolete if we add new pseudos |
5717 | since the last set up. Recalculate it again if the new pseudos were |
5718 | added. */ |
5719 | if (resize_reg_info () && (flag_sched_pressure || flag_live_range_shrinkage |
5720 | || flag_ira_loop_pressure)) |
5721 | ira_set_pseudo_classes (true, ira_dump_file); |
5722 | |
5723 | init_alias_analysis (); |
5724 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); |
5725 | reg_equiv = XCNEWVEC (struct equivalence, max_reg_num ()); |
5726 | update_equiv_regs_prescan (); |
5727 | update_equiv_regs (); |
5728 | |
5729 | /* Don't move insns if live range shrinkage or register |
5730 | pressure-sensitive scheduling were done because it will not |
5731 | improve allocation but likely worsen insn scheduling. */ |
5732 | if (optimize |
5733 | && !flag_live_range_shrinkage |
5734 | && !(flag_sched_pressure && flag_schedule_insns)) |
5735 | combine_and_move_insns (); |
5736 | |
5737 | /* Gather additional equivalences with memory. */ |
5738 | if (optimize) |
5739 | add_store_equivs (); |
5740 | |
5741 | loop_optimizer_finalize (); |
5742 | free_dominance_info (CDI_DOMINATORS); |
5743 | end_alias_analysis (); |
5744 | free (ptr: reg_equiv); |
5745 | |
5746 | /* Once max_regno changes, we need to free and re-init/re-compute |
5747 | some data structures like regstat_n_sets_and_refs and reg_info_p. */ |
5748 | auto regstat_recompute_for_max_regno = []() { |
5749 | regstat_free_n_sets_and_refs (); |
5750 | regstat_free_ri (); |
5751 | regstat_init_n_sets_and_refs (); |
5752 | regstat_compute_ri (); |
5753 | resize_reg_info (); |
5754 | }; |
5755 | |
5756 | int max_regno_before_rm = max_reg_num (); |
5757 | if (ira_use_lra_p && remove_scratches ()) |
5758 | { |
5759 | ira_expand_reg_equiv (); |
5760 | /* For now remove_scatches is supposed to create pseudos when it |
5761 | succeeds, assert this happens all the time. Once it doesn't |
5762 | hold, we should guard the regstat recompute for the case |
5763 | max_regno changes. */ |
5764 | gcc_assert (max_regno_before_rm != max_reg_num ()); |
5765 | regstat_recompute_for_max_regno (); |
5766 | } |
5767 | |
5768 | setup_reg_equiv (); |
5769 | grow_reg_equivs (); |
5770 | setup_reg_equiv_init (); |
5771 | |
5772 | allocated_reg_info_size = max_reg_num (); |
5773 | |
5774 | /* It is not worth to do such improvement when we use a simple |
5775 | allocation because of -O0 usage or because the function is too |
5776 | big. */ |
5777 | if (ira_conflicts_p) |
5778 | find_moveable_pseudos (); |
5779 | |
5780 | max_regno_before_ira = max_reg_num (); |
5781 | ira_setup_eliminable_regset (); |
5782 | |
5783 | ira_overall_cost = ira_reg_cost = ira_mem_cost = 0; |
5784 | ira_load_cost = ira_store_cost = ira_shuffle_cost = 0; |
5785 | ira_move_loops_num = ira_additional_jumps_num = 0; |
5786 | |
5787 | ira_assert (current_loops == NULL); |
5788 | if (flag_ira_region == IRA_REGION_ALL || flag_ira_region == IRA_REGION_MIXED) |
5789 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS | LOOPS_HAVE_RECORDED_EXITS); |
5790 | |
5791 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL) |
5792 | fprintf (stream: ira_dump_file, format: "Building IRA IR\n" ); |
5793 | loops_p = ira_build (); |
5794 | |
5795 | ira_assert (ira_conflicts_p || !loops_p); |
5796 | |
5797 | saved_flag_ira_share_spill_slots = flag_ira_share_spill_slots; |
5798 | if (too_high_register_pressure_p () || cfun->calls_setjmp) |
5799 | /* It is just wasting compiler's time to pack spilled pseudos into |
5800 | stack slots in this case -- prohibit it. We also do this if |
5801 | there is setjmp call because a variable not modified between |
5802 | setjmp and longjmp the compiler is required to preserve its |
5803 | value and sharing slots does not guarantee it. */ |
5804 | flag_ira_share_spill_slots = false; |
5805 | |
5806 | ira_color (); |
5807 | |
5808 | ira_max_point_before_emit = ira_max_point; |
5809 | |
5810 | ira_initiate_emit_data (); |
5811 | |
5812 | ira_emit (loops_p); |
5813 | |
5814 | max_regno = max_reg_num (); |
5815 | if (ira_conflicts_p) |
5816 | { |
5817 | if (! loops_p) |
5818 | { |
5819 | if (! ira_use_lra_p) |
5820 | ira_initiate_assign (); |
5821 | } |
5822 | else |
5823 | { |
5824 | expand_reg_info (); |
5825 | |
5826 | if (ira_use_lra_p) |
5827 | { |
5828 | ira_allocno_t a; |
5829 | ira_allocno_iterator ai; |
5830 | |
5831 | FOR_EACH_ALLOCNO (a, ai) |
5832 | { |
5833 | int old_regno = ALLOCNO_REGNO (a); |
5834 | int new_regno = REGNO (ALLOCNO_EMIT_DATA (a)->reg); |
5835 | |
5836 | ALLOCNO_REGNO (a) = new_regno; |
5837 | |
5838 | if (old_regno != new_regno) |
5839 | setup_reg_classes (new_regno, reg_preferred_class (old_regno), |
5840 | reg_alternate_class (old_regno), |
5841 | reg_allocno_class (old_regno)); |
5842 | } |
5843 | } |
5844 | else |
5845 | { |
5846 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL) |
5847 | fprintf (stream: ira_dump_file, format: "Flattening IR\n" ); |
5848 | ira_flattening (max_regno_before_ira, ira_max_point_before_emit); |
5849 | } |
5850 | /* New insns were generated: add notes and recalculate live |
5851 | info. */ |
5852 | df_analyze (); |
5853 | |
5854 | /* ??? Rebuild the loop tree, but why? Does the loop tree |
5855 | change if new insns were generated? Can that be handled |
5856 | by updating the loop tree incrementally? */ |
5857 | loop_optimizer_finalize (); |
5858 | free_dominance_info (CDI_DOMINATORS); |
5859 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS |
5860 | | LOOPS_HAVE_RECORDED_EXITS); |
5861 | |
5862 | if (! ira_use_lra_p) |
5863 | { |
5864 | setup_allocno_assignment_flags (); |
5865 | ira_initiate_assign (); |
5866 | ira_reassign_conflict_allocnos (max_regno); |
5867 | } |
5868 | } |
5869 | } |
5870 | |
5871 | ira_finish_emit_data (); |
5872 | |
5873 | setup_reg_renumber (); |
5874 | |
5875 | calculate_allocation_cost (); |
5876 | |
5877 | #ifdef ENABLE_IRA_CHECKING |
5878 | if (ira_conflicts_p && ! ira_use_lra_p) |
5879 | /* Opposite to reload pass, LRA does not use any conflict info |
5880 | from IRA. We don't rebuild conflict info for LRA (through |
5881 | ira_flattening call) and cannot use the check here. We could |
5882 | rebuild this info for LRA in the check mode but there is a risk |
5883 | that code generated with the check and without it will be a bit |
5884 | different. Calling ira_flattening in any mode would be a |
5885 | wasting CPU time. So do not check the allocation for LRA. */ |
5886 | check_allocation (); |
5887 | #endif |
5888 | |
5889 | if (max_regno != max_regno_before_ira) |
5890 | regstat_recompute_for_max_regno (); |
5891 | |
5892 | overall_cost_before = ira_overall_cost; |
5893 | if (! ira_conflicts_p) |
5894 | grow_reg_equivs (); |
5895 | else |
5896 | { |
5897 | fix_reg_equiv_init (); |
5898 | |
5899 | #ifdef ENABLE_IRA_CHECKING |
5900 | print_redundant_copies (); |
5901 | #endif |
5902 | if (! ira_use_lra_p) |
5903 | { |
5904 | ira_spilled_reg_stack_slots_num = 0; |
5905 | ira_spilled_reg_stack_slots |
5906 | = ((class ira_spilled_reg_stack_slot *) |
5907 | ira_allocate (len: max_regno |
5908 | * sizeof (class ira_spilled_reg_stack_slot))); |
5909 | memset (s: (void *)ira_spilled_reg_stack_slots, c: 0, |
5910 | n: max_regno * sizeof (class ira_spilled_reg_stack_slot)); |
5911 | } |
5912 | } |
5913 | allocate_initial_values (); |
5914 | |
5915 | /* See comment for find_moveable_pseudos call. */ |
5916 | if (ira_conflicts_p) |
5917 | move_unallocated_pseudos (); |
5918 | |
5919 | /* Restore original values. */ |
5920 | if (lra_simple_p) |
5921 | { |
5922 | flag_caller_saves = saved_flag_caller_saves; |
5923 | flag_ira_region = saved_flag_ira_region; |
5924 | } |
5925 | } |
5926 | |
5927 | /* Modify asm goto to avoid further trouble with this insn. We can |
5928 | not replace the insn by USE as in other asm insns as we still |
5929 | need to keep CFG consistency. */ |
5930 | void |
5931 | ira_nullify_asm_goto (rtx_insn *insn) |
5932 | { |
5933 | ira_assert (JUMP_P (insn) && INSN_CODE (insn) < 0); |
5934 | rtx tmp = extract_asm_operands (PATTERN (insn)); |
5935 | PATTERN (insn) = gen_rtx_ASM_OPERANDS (VOIDmode, ggc_strdup ("" ), "" , 0, |
5936 | rtvec_alloc (0), |
5937 | rtvec_alloc (0), |
5938 | ASM_OPERANDS_LABEL_VEC (tmp), |
5939 | ASM_OPERANDS_SOURCE_LOCATION(tmp)); |
5940 | } |
5941 | |
5942 | static void |
5943 | do_reload (void) |
5944 | { |
5945 | basic_block bb; |
5946 | bool need_dce; |
5947 | unsigned pic_offset_table_regno = INVALID_REGNUM; |
5948 | |
5949 | if (flag_ira_verbose < 10) |
5950 | ira_dump_file = dump_file; |
5951 | |
5952 | /* If pic_offset_table_rtx is a pseudo register, then keep it so |
5953 | after reload to avoid possible wrong usages of hard reg assigned |
5954 | to it. */ |
5955 | if (pic_offset_table_rtx |
5956 | && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER) |
5957 | pic_offset_table_regno = REGNO (pic_offset_table_rtx); |
5958 | |
5959 | timevar_push (tv: TV_RELOAD); |
5960 | if (ira_use_lra_p) |
5961 | { |
5962 | if (current_loops != NULL) |
5963 | { |
5964 | loop_optimizer_finalize (); |
5965 | free_dominance_info (CDI_DOMINATORS); |
5966 | } |
5967 | FOR_ALL_BB_FN (bb, cfun) |
5968 | bb->loop_father = NULL; |
5969 | current_loops = NULL; |
5970 | |
5971 | ira_destroy (); |
5972 | |
5973 | lra (ira_dump_file); |
5974 | /* ???!!! Move it before lra () when we use ira_reg_equiv in |
5975 | LRA. */ |
5976 | vec_free (v&: reg_equivs); |
5977 | reg_equivs = NULL; |
5978 | need_dce = false; |
5979 | } |
5980 | else |
5981 | { |
5982 | df_set_flags (DF_NO_INSN_RESCAN); |
5983 | build_insn_chain (); |
5984 | |
5985 | need_dce = reload (get_insns (), ira_conflicts_p); |
5986 | } |
5987 | |
5988 | timevar_pop (tv: TV_RELOAD); |
5989 | |
5990 | timevar_push (tv: TV_IRA); |
5991 | |
5992 | if (ira_conflicts_p && ! ira_use_lra_p) |
5993 | { |
5994 | ira_free (addr: ira_spilled_reg_stack_slots); |
5995 | ira_finish_assign (); |
5996 | } |
5997 | |
5998 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL |
5999 | && overall_cost_before != ira_overall_cost) |
6000 | fprintf (stream: ira_dump_file, format: "+++Overall after reload %" PRId64 "\n" , |
6001 | ira_overall_cost); |
6002 | |
6003 | flag_ira_share_spill_slots = saved_flag_ira_share_spill_slots; |
6004 | |
6005 | if (! ira_use_lra_p) |
6006 | { |
6007 | ira_destroy (); |
6008 | if (current_loops != NULL) |
6009 | { |
6010 | loop_optimizer_finalize (); |
6011 | free_dominance_info (CDI_DOMINATORS); |
6012 | } |
6013 | FOR_ALL_BB_FN (bb, cfun) |
6014 | bb->loop_father = NULL; |
6015 | current_loops = NULL; |
6016 | |
6017 | regstat_free_ri (); |
6018 | regstat_free_n_sets_and_refs (); |
6019 | } |
6020 | |
6021 | if (optimize) |
6022 | cleanup_cfg (CLEANUP_EXPENSIVE); |
6023 | |
6024 | finish_reg_equiv (); |
6025 | |
6026 | bitmap_obstack_release (&ira_bitmap_obstack); |
6027 | #ifndef IRA_NO_OBSTACK |
6028 | obstack_free (&ira_obstack, NULL); |
6029 | #endif |
6030 | |
6031 | /* The code after the reload has changed so much that at this point |
6032 | we might as well just rescan everything. Note that |
6033 | df_rescan_all_insns is not going to help here because it does not |
6034 | touch the artificial uses and defs. */ |
6035 | df_finish_pass (true); |
6036 | df_scan_alloc (NULL); |
6037 | df_scan_blocks (); |
6038 | |
6039 | if (optimize > 1) |
6040 | { |
6041 | df_live_add_problem (); |
6042 | df_live_set_all_dirty (); |
6043 | } |
6044 | |
6045 | if (optimize) |
6046 | df_analyze (); |
6047 | |
6048 | if (need_dce && optimize) |
6049 | run_fast_dce (); |
6050 | |
6051 | /* Diagnose uses of the hard frame pointer when it is used as a global |
6052 | register. Often we can get away with letting the user appropriate |
6053 | the frame pointer, but we should let them know when code generation |
6054 | makes that impossible. */ |
6055 | if (global_regs[HARD_FRAME_POINTER_REGNUM] && frame_pointer_needed) |
6056 | { |
6057 | tree decl = global_regs_decl[HARD_FRAME_POINTER_REGNUM]; |
6058 | error_at (DECL_SOURCE_LOCATION (current_function_decl), |
6059 | "frame pointer required, but reserved" ); |
6060 | inform (DECL_SOURCE_LOCATION (decl), "for %qD" , decl); |
6061 | } |
6062 | |
6063 | /* If we are doing generic stack checking, give a warning if this |
6064 | function's frame size is larger than we expect. */ |
6065 | if (flag_stack_check == GENERIC_STACK_CHECK) |
6066 | { |
6067 | poly_int64 size = get_frame_size () + STACK_CHECK_FIXED_FRAME_SIZE; |
6068 | |
6069 | for (int i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
6070 | if (df_regs_ever_live_p (i) |
6071 | && !fixed_regs[i] |
6072 | && !crtl->abi->clobbers_full_reg_p (regno: i)) |
6073 | size += UNITS_PER_WORD; |
6074 | |
6075 | if (constant_lower_bound (a: size) > STACK_CHECK_MAX_FRAME_SIZE) |
6076 | warning (0, "frame size too large for reliable stack checking" ); |
6077 | } |
6078 | |
6079 | if (pic_offset_table_regno != INVALID_REGNUM) |
6080 | pic_offset_table_rtx = gen_rtx_REG (Pmode, pic_offset_table_regno); |
6081 | |
6082 | timevar_pop (tv: TV_IRA); |
6083 | } |
6084 | |
6085 | /* Run the integrated register allocator. */ |
6086 | |
6087 | namespace { |
6088 | |
6089 | const pass_data pass_data_ira = |
6090 | { |
6091 | .type: RTL_PASS, /* type */ |
6092 | .name: "ira" , /* name */ |
6093 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6094 | .tv_id: TV_IRA, /* tv_id */ |
6095 | .properties_required: 0, /* properties_required */ |
6096 | .properties_provided: 0, /* properties_provided */ |
6097 | .properties_destroyed: 0, /* properties_destroyed */ |
6098 | .todo_flags_start: 0, /* todo_flags_start */ |
6099 | TODO_do_not_ggc_collect, /* todo_flags_finish */ |
6100 | }; |
6101 | |
6102 | class pass_ira : public rtl_opt_pass |
6103 | { |
6104 | public: |
6105 | pass_ira (gcc::context *ctxt) |
6106 | : rtl_opt_pass (pass_data_ira, ctxt) |
6107 | {} |
6108 | |
6109 | /* opt_pass methods: */ |
6110 | bool gate (function *) final override |
6111 | { |
6112 | return !targetm.no_register_allocation; |
6113 | } |
6114 | unsigned int execute (function *) final override |
6115 | { |
6116 | ira_in_progress = true; |
6117 | ira (f: dump_file); |
6118 | ira_in_progress = false; |
6119 | return 0; |
6120 | } |
6121 | |
6122 | }; // class pass_ira |
6123 | |
6124 | } // anon namespace |
6125 | |
6126 | rtl_opt_pass * |
6127 | make_pass_ira (gcc::context *ctxt) |
6128 | { |
6129 | return new pass_ira (ctxt); |
6130 | } |
6131 | |
6132 | namespace { |
6133 | |
6134 | const pass_data pass_data_reload = |
6135 | { |
6136 | .type: RTL_PASS, /* type */ |
6137 | .name: "reload" , /* name */ |
6138 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6139 | .tv_id: TV_RELOAD, /* tv_id */ |
6140 | .properties_required: 0, /* properties_required */ |
6141 | .properties_provided: 0, /* properties_provided */ |
6142 | .properties_destroyed: 0, /* properties_destroyed */ |
6143 | .todo_flags_start: 0, /* todo_flags_start */ |
6144 | .todo_flags_finish: 0, /* todo_flags_finish */ |
6145 | }; |
6146 | |
6147 | class pass_reload : public rtl_opt_pass |
6148 | { |
6149 | public: |
6150 | pass_reload (gcc::context *ctxt) |
6151 | : rtl_opt_pass (pass_data_reload, ctxt) |
6152 | {} |
6153 | |
6154 | /* opt_pass methods: */ |
6155 | bool gate (function *) final override |
6156 | { |
6157 | return !targetm.no_register_allocation; |
6158 | } |
6159 | unsigned int execute (function *) final override |
6160 | { |
6161 | do_reload (); |
6162 | return 0; |
6163 | } |
6164 | |
6165 | }; // class pass_reload |
6166 | |
6167 | } // anon namespace |
6168 | |
6169 | rtl_opt_pass * |
6170 | make_pass_reload (gcc::context *ctxt) |
6171 | { |
6172 | return new pass_reload (ctxt); |
6173 | } |
6174 | |