]>
Commit | Line | Data |
---|---|---|
531b10fc SB |
1 | /* Lower GIMPLE_SWITCH expressions to something more efficient than |
2 | a jump table. | |
cbe34bb5 | 3 | Copyright (C) 2006-2017 Free Software Foundation, Inc. |
b6e99746 MJ |
4 | |
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published by the | |
9 | Free Software Foundation; either version 3, or (at your option) any | |
10 | later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 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, write to the Free | |
19 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA | |
20 | 02110-1301, USA. */ | |
21 | ||
531b10fc SB |
22 | /* This file handles the lowering of GIMPLE_SWITCH to an indexed |
23 | load, or a series of bit-test-and-branch expressions. */ | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
c7131fb2 | 28 | #include "backend.h" |
957060b5 AM |
29 | #include "insn-codes.h" |
30 | #include "rtl.h" | |
c7131fb2 AM |
31 | #include "tree.h" |
32 | #include "gimple.h" | |
957060b5 AM |
33 | #include "cfghooks.h" |
34 | #include "tree-pass.h" | |
c7131fb2 | 35 | #include "ssa.h" |
957060b5 AM |
36 | #include "optabs-tree.h" |
37 | #include "cgraph.h" | |
38 | #include "gimple-pretty-print.h" | |
531b10fc | 39 | #include "params.h" |
40e23961 | 40 | #include "fold-const.h" |
d8a2d370 DN |
41 | #include "varasm.h" |
42 | #include "stor-layout.h" | |
60393bbc | 43 | #include "cfganal.h" |
45b0be94 | 44 | #include "gimplify.h" |
5be5c238 | 45 | #include "gimple-iterator.h" |
18f429e2 | 46 | #include "gimplify-me.h" |
442b4905 | 47 | #include "tree-cfg.h" |
a9e0d843 | 48 | #include "cfgloop.h" |
9dc3d6a9 ML |
49 | #include "alloc-pool.h" |
50 | #include "target.h" | |
51 | #include "tree-into-ssa.h" | |
7ee2468b SB |
52 | |
53 | /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode | |
54 | type in the GIMPLE type system that is language-independent? */ | |
531b10fc SB |
55 | #include "langhooks.h" |
56 | ||
531b10fc SB |
57 | \f |
58 | /* Maximum number of case bit tests. | |
59 | FIXME: This should be derived from PARAM_CASE_VALUES_THRESHOLD and | |
60 | targetm.case_values_threshold(), or be its own param. */ | |
61 | #define MAX_CASE_BIT_TESTS 3 | |
62 | ||
63 | /* Split the basic block at the statement pointed to by GSIP, and insert | |
64 | a branch to the target basic block of E_TRUE conditional on tree | |
65 | expression COND. | |
66 | ||
67 | It is assumed that there is already an edge from the to-be-split | |
68 | basic block to E_TRUE->dest block. This edge is removed, and the | |
69 | profile information on the edge is re-used for the new conditional | |
70 | jump. | |
71 | ||
72 | The CFG is updated. The dominator tree will not be valid after | |
73 | this transformation, but the immediate dominators are updated if | |
74 | UPDATE_DOMINATORS is true. | |
75 | ||
76 | Returns the newly created basic block. */ | |
77 | ||
78 | static basic_block | |
79 | hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip, | |
80 | tree cond, edge e_true, | |
81 | bool update_dominators) | |
82 | { | |
83 | tree tmp; | |
538dd0b7 | 84 | gcond *cond_stmt; |
531b10fc SB |
85 | edge e_false; |
86 | basic_block new_bb, split_bb = gsi_bb (*gsip); | |
87 | bool dominated_e_true = false; | |
88 | ||
89 | gcc_assert (e_true->src == split_bb); | |
90 | ||
91 | if (update_dominators | |
92 | && get_immediate_dominator (CDI_DOMINATORS, e_true->dest) == split_bb) | |
93 | dominated_e_true = true; | |
94 | ||
95 | tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL, | |
96 | /*before=*/true, GSI_SAME_STMT); | |
97 | cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE); | |
98 | gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT); | |
99 | ||
100 | e_false = split_block (split_bb, cond_stmt); | |
101 | new_bb = e_false->dest; | |
102 | redirect_edge_pred (e_true, split_bb); | |
103 | ||
104 | e_true->flags &= ~EDGE_FALLTHRU; | |
105 | e_true->flags |= EDGE_TRUE_VALUE; | |
106 | ||
107 | e_false->flags &= ~EDGE_FALLTHRU; | |
108 | e_false->flags |= EDGE_FALSE_VALUE; | |
357067f2 | 109 | e_false->probability = e_true->probability.invert (); |
531b10fc SB |
110 | e_false->count = split_bb->count - e_true->count; |
111 | new_bb->count = e_false->count; | |
112 | ||
113 | if (update_dominators) | |
114 | { | |
115 | if (dominated_e_true) | |
116 | set_immediate_dominator (CDI_DOMINATORS, e_true->dest, split_bb); | |
117 | set_immediate_dominator (CDI_DOMINATORS, e_false->dest, split_bb); | |
118 | } | |
119 | ||
120 | return new_bb; | |
121 | } | |
122 | ||
123 | ||
531b10fc SB |
124 | /* Return true if a switch should be expanded as a bit test. |
125 | RANGE is the difference between highest and lowest case. | |
126 | UNIQ is number of unique case node targets, not counting the default case. | |
127 | COUNT is the number of comparisons needed, not counting the default case. */ | |
128 | ||
129 | static bool | |
130 | expand_switch_using_bit_tests_p (tree range, | |
131 | unsigned int uniq, | |
72798784 | 132 | unsigned int count, bool speed_p) |
531b10fc SB |
133 | { |
134 | return (((uniq == 1 && count >= 3) | |
135 | || (uniq == 2 && count >= 5) | |
136 | || (uniq == 3 && count >= 6)) | |
72798784 | 137 | && lshift_cheap_p (speed_p) |
531b10fc SB |
138 | && compare_tree_int (range, GET_MODE_BITSIZE (word_mode)) < 0 |
139 | && compare_tree_int (range, 0) > 0); | |
140 | } | |
141 | \f | |
142 | /* Implement switch statements with bit tests | |
143 | ||
144 | A GIMPLE switch statement can be expanded to a short sequence of bit-wise | |
145 | comparisons. "switch(x)" is converted into "if ((1 << (x-MINVAL)) & CST)" | |
146 | where CST and MINVAL are integer constants. This is better than a series | |
147 | of compare-and-banch insns in some cases, e.g. we can implement: | |
148 | ||
149 | if ((x==4) || (x==6) || (x==9) || (x==11)) | |
150 | ||
151 | as a single bit test: | |
152 | ||
153 | if ((1<<x) & ((1<<4)|(1<<6)|(1<<9)|(1<<11))) | |
154 | ||
155 | This transformation is only applied if the number of case targets is small, | |
34540577 | 156 | if CST constains at least 3 bits, and "1 << x" is cheap. The bit tests are |
531b10fc SB |
157 | performed in "word_mode". |
158 | ||
159 | The following example shows the code the transformation generates: | |
160 | ||
161 | int bar(int x) | |
162 | { | |
163 | switch (x) | |
164 | { | |
165 | case '0': case '1': case '2': case '3': case '4': | |
166 | case '5': case '6': case '7': case '8': case '9': | |
167 | case 'A': case 'B': case 'C': case 'D': case 'E': | |
168 | case 'F': | |
169 | return 1; | |
170 | } | |
171 | return 0; | |
172 | } | |
173 | ||
174 | ==> | |
175 | ||
176 | bar (int x) | |
177 | { | |
178 | tmp1 = x - 48; | |
179 | if (tmp1 > (70 - 48)) goto L2; | |
180 | tmp2 = 1 << tmp1; | |
181 | tmp3 = 0b11111100000001111111111; | |
182 | if ((tmp2 & tmp3) != 0) goto L1 ; else goto L2; | |
183 | L1: | |
184 | return 1; | |
185 | L2: | |
186 | return 0; | |
187 | } | |
188 | ||
189 | TODO: There are still some improvements to this transformation that could | |
190 | be implemented: | |
191 | ||
192 | * A narrower mode than word_mode could be used if that is cheaper, e.g. | |
193 | for x86_64 where a narrower-mode shift may result in smaller code. | |
194 | ||
195 | * The compounded constant could be shifted rather than the one. The | |
196 | test would be either on the sign bit or on the least significant bit, | |
197 | depending on the direction of the shift. On some machines, the test | |
198 | for the branch would be free if the bit to test is already set by the | |
199 | shift operation. | |
200 | ||
201 | This transformation was contributed by Roger Sayle, see this e-mail: | |
202 | http://gcc.gnu.org/ml/gcc-patches/2003-01/msg01950.html | |
203 | */ | |
204 | ||
205 | /* A case_bit_test represents a set of case nodes that may be | |
206 | selected from using a bit-wise comparison. HI and LO hold | |
207 | the integer to be tested against, TARGET_EDGE contains the | |
208 | edge to the basic block to jump to upon success and BITS | |
209 | counts the number of case nodes handled by this test, | |
210 | typically the number of bits set in HI:LO. The LABEL field | |
211 | is used to quickly identify all cases in this set without | |
212 | looking at label_to_block for every case label. */ | |
213 | ||
214 | struct case_bit_test | |
215 | { | |
aa79a1e1 | 216 | wide_int mask; |
531b10fc SB |
217 | edge target_edge; |
218 | tree label; | |
219 | int bits; | |
220 | }; | |
221 | ||
222 | /* Comparison function for qsort to order bit tests by decreasing | |
223 | probability of execution. Our best guess comes from a measured | |
224 | profile. If the profile counts are equal, break even on the | |
225 | number of case nodes, i.e. the node with the most cases gets | |
226 | tested first. | |
227 | ||
228 | TODO: Actually this currently runs before a profile is available. | |
229 | Therefore the case-as-bit-tests transformation should be done | |
230 | later in the pass pipeline, or something along the lines of | |
231 | "Efficient and effective branch reordering using profile data" | |
232 | (Yang et. al., 2002) should be implemented (although, how good | |
233 | is a paper is called "Efficient and effective ..." when the | |
234 | latter is implied by the former, but oh well...). */ | |
235 | ||
236 | static int | |
237 | case_bit_test_cmp (const void *p1, const void *p2) | |
238 | { | |
239 | const struct case_bit_test *const d1 = (const struct case_bit_test *) p1; | |
240 | const struct case_bit_test *const d2 = (const struct case_bit_test *) p2; | |
241 | ||
3995f3a2 JH |
242 | if (d2->target_edge->count < d1->target_edge->count) |
243 | return -1; | |
244 | if (d2->target_edge->count > d1->target_edge->count) | |
245 | return 1; | |
531b10fc SB |
246 | if (d2->bits != d1->bits) |
247 | return d2->bits - d1->bits; | |
248 | ||
249 | /* Stabilize the sort. */ | |
250 | return LABEL_DECL_UID (d2->label) - LABEL_DECL_UID (d1->label); | |
251 | } | |
252 | ||
253 | /* Expand a switch statement by a short sequence of bit-wise | |
254 | comparisons. "switch(x)" is effectively converted into | |
255 | "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are | |
256 | integer constants. | |
257 | ||
258 | INDEX_EXPR is the value being switched on. | |
259 | ||
260 | MINVAL is the lowest case value of in the case nodes, | |
261 | and RANGE is highest value minus MINVAL. MINVAL and RANGE | |
262 | are not guaranteed to be of the same type as INDEX_EXPR | |
263 | (the gimplifier doesn't change the type of case label values, | |
264 | and MINVAL and RANGE are derived from those values). | |
aa79a1e1 | 265 | MAXVAL is MINVAL + RANGE. |
531b10fc SB |
266 | |
267 | There *MUST* be MAX_CASE_BIT_TESTS or less unique case | |
268 | node targets. */ | |
269 | ||
270 | static void | |
538dd0b7 | 271 | emit_case_bit_tests (gswitch *swtch, tree index_expr, |
aa79a1e1 | 272 | tree minval, tree range, tree maxval) |
531b10fc | 273 | { |
66846c07 | 274 | struct case_bit_test test[MAX_CASE_BIT_TESTS] = { {} }; |
531b10fc SB |
275 | unsigned int i, j, k; |
276 | unsigned int count; | |
277 | ||
278 | basic_block switch_bb = gimple_bb (swtch); | |
279 | basic_block default_bb, new_default_bb, new_bb; | |
280 | edge default_edge; | |
281 | bool update_dom = dom_info_available_p (CDI_DOMINATORS); | |
282 | ||
6e1aa848 | 283 | vec<basic_block> bbs_to_fix_dom = vNULL; |
531b10fc SB |
284 | |
285 | tree index_type = TREE_TYPE (index_expr); | |
286 | tree unsigned_index_type = unsigned_type_for (index_type); | |
287 | unsigned int branch_num = gimple_switch_num_labels (swtch); | |
288 | ||
289 | gimple_stmt_iterator gsi; | |
538dd0b7 | 290 | gassign *shift_stmt; |
531b10fc SB |
291 | |
292 | tree idx, tmp, csui; | |
293 | tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1); | |
294 | tree word_mode_zero = fold_convert (word_type_node, integer_zero_node); | |
295 | tree word_mode_one = fold_convert (word_type_node, integer_one_node); | |
aa79a1e1 JJ |
296 | int prec = TYPE_PRECISION (word_type_node); |
297 | wide_int wone = wi::one (prec); | |
531b10fc | 298 | |
531b10fc | 299 | /* Get the edge for the default case. */ |
fd8d363e | 300 | tmp = gimple_switch_default_label (swtch); |
531b10fc SB |
301 | default_bb = label_to_block (CASE_LABEL (tmp)); |
302 | default_edge = find_edge (switch_bb, default_bb); | |
303 | ||
304 | /* Go through all case labels, and collect the case labels, profile | |
305 | counts, and other information we need to build the branch tests. */ | |
306 | count = 0; | |
307 | for (i = 1; i < branch_num; i++) | |
308 | { | |
309 | unsigned int lo, hi; | |
310 | tree cs = gimple_switch_label (swtch, i); | |
311 | tree label = CASE_LABEL (cs); | |
8166ff4d | 312 | edge e = find_edge (switch_bb, label_to_block (label)); |
531b10fc | 313 | for (k = 0; k < count; k++) |
8166ff4d | 314 | if (e == test[k].target_edge) |
531b10fc SB |
315 | break; |
316 | ||
317 | if (k == count) | |
318 | { | |
531b10fc | 319 | gcc_checking_assert (count < MAX_CASE_BIT_TESTS); |
aa79a1e1 | 320 | test[k].mask = wi::zero (prec); |
531b10fc SB |
321 | test[k].target_edge = e; |
322 | test[k].label = label; | |
323 | test[k].bits = 1; | |
324 | count++; | |
325 | } | |
326 | else | |
327 | test[k].bits++; | |
328 | ||
386b1f1f RS |
329 | lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, |
330 | CASE_LOW (cs), minval)); | |
531b10fc SB |
331 | if (CASE_HIGH (cs) == NULL_TREE) |
332 | hi = lo; | |
333 | else | |
386b1f1f RS |
334 | hi = tree_to_uhwi (int_const_binop (MINUS_EXPR, |
335 | CASE_HIGH (cs), minval)); | |
531b10fc SB |
336 | |
337 | for (j = lo; j <= hi; j++) | |
aa79a1e1 | 338 | test[k].mask |= wi::lshift (wone, j); |
531b10fc SB |
339 | } |
340 | ||
c3284718 | 341 | qsort (test, count, sizeof (*test), case_bit_test_cmp); |
531b10fc | 342 | |
aa79a1e1 JJ |
343 | /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of |
344 | the minval subtractions, but it might make the mask constants more | |
345 | expensive. So, compare the costs. */ | |
346 | if (compare_tree_int (minval, 0) > 0 | |
347 | && compare_tree_int (maxval, GET_MODE_BITSIZE (word_mode)) < 0) | |
348 | { | |
349 | int cost_diff; | |
350 | HOST_WIDE_INT m = tree_to_uhwi (minval); | |
351 | rtx reg = gen_raw_REG (word_mode, 10000); | |
352 | bool speed_p = optimize_bb_for_speed_p (gimple_bb (swtch)); | |
353 | cost_diff = set_rtx_cost (gen_rtx_PLUS (word_mode, reg, | |
354 | GEN_INT (-m)), speed_p); | |
355 | for (i = 0; i < count; i++) | |
356 | { | |
357 | rtx r = immed_wide_int_const (test[i].mask, word_mode); | |
e548c9df AM |
358 | cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r), |
359 | word_mode, speed_p); | |
aa79a1e1 | 360 | r = immed_wide_int_const (wi::lshift (test[i].mask, m), word_mode); |
e548c9df AM |
361 | cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r), |
362 | word_mode, speed_p); | |
aa79a1e1 JJ |
363 | } |
364 | if (cost_diff > 0) | |
365 | { | |
366 | for (i = 0; i < count; i++) | |
367 | test[i].mask = wi::lshift (test[i].mask, m); | |
368 | minval = build_zero_cst (TREE_TYPE (minval)); | |
369 | range = maxval; | |
370 | } | |
371 | } | |
372 | ||
531b10fc SB |
373 | /* We generate two jumps to the default case label. |
374 | Split the default edge, so that we don't have to do any PHI node | |
375 | updating. */ | |
376 | new_default_bb = split_edge (default_edge); | |
377 | ||
378 | if (update_dom) | |
379 | { | |
9771b263 DN |
380 | bbs_to_fix_dom.create (10); |
381 | bbs_to_fix_dom.quick_push (switch_bb); | |
382 | bbs_to_fix_dom.quick_push (default_bb); | |
383 | bbs_to_fix_dom.quick_push (new_default_bb); | |
531b10fc SB |
384 | } |
385 | ||
386 | /* Now build the test-and-branch code. */ | |
387 | ||
388 | gsi = gsi_last_bb (switch_bb); | |
389 | ||
d9e408de TV |
390 | /* idx = (unsigned)x - minval. */ |
391 | idx = fold_convert (unsigned_index_type, index_expr); | |
392 | idx = fold_build2 (MINUS_EXPR, unsigned_index_type, idx, | |
393 | fold_convert (unsigned_index_type, minval)); | |
531b10fc SB |
394 | idx = force_gimple_operand_gsi (&gsi, idx, |
395 | /*simple=*/true, NULL_TREE, | |
396 | /*before=*/true, GSI_SAME_STMT); | |
397 | ||
398 | /* if (idx > range) goto default */ | |
399 | range = force_gimple_operand_gsi (&gsi, | |
400 | fold_convert (unsigned_index_type, range), | |
401 | /*simple=*/true, NULL_TREE, | |
402 | /*before=*/true, GSI_SAME_STMT); | |
403 | tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range); | |
404 | new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, default_edge, update_dom); | |
405 | if (update_dom) | |
9771b263 | 406 | bbs_to_fix_dom.quick_push (new_bb); |
531b10fc SB |
407 | gcc_assert (gimple_bb (swtch) == new_bb); |
408 | gsi = gsi_last_bb (new_bb); | |
409 | ||
410 | /* Any blocks dominated by the GIMPLE_SWITCH, but that are not successors | |
411 | of NEW_BB, are still immediately dominated by SWITCH_BB. Make it so. */ | |
412 | if (update_dom) | |
413 | { | |
9771b263 | 414 | vec<basic_block> dom_bbs; |
531b10fc SB |
415 | basic_block dom_son; |
416 | ||
417 | dom_bbs = get_dominated_by (CDI_DOMINATORS, new_bb); | |
9771b263 | 418 | FOR_EACH_VEC_ELT (dom_bbs, i, dom_son) |
531b10fc SB |
419 | { |
420 | edge e = find_edge (new_bb, dom_son); | |
421 | if (e && single_pred_p (e->dest)) | |
422 | continue; | |
423 | set_immediate_dominator (CDI_DOMINATORS, dom_son, switch_bb); | |
9771b263 | 424 | bbs_to_fix_dom.safe_push (dom_son); |
531b10fc | 425 | } |
9771b263 | 426 | dom_bbs.release (); |
531b10fc SB |
427 | } |
428 | ||
429 | /* csui = (1 << (word_mode) idx) */ | |
b731b390 | 430 | csui = make_ssa_name (word_type_node); |
531b10fc SB |
431 | tmp = fold_build2 (LSHIFT_EXPR, word_type_node, word_mode_one, |
432 | fold_convert (word_type_node, idx)); | |
433 | tmp = force_gimple_operand_gsi (&gsi, tmp, | |
434 | /*simple=*/false, NULL_TREE, | |
435 | /*before=*/true, GSI_SAME_STMT); | |
436 | shift_stmt = gimple_build_assign (csui, tmp); | |
531b10fc SB |
437 | gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT); |
438 | update_stmt (shift_stmt); | |
439 | ||
440 | /* for each unique set of cases: | |
441 | if (const & csui) goto target */ | |
442 | for (k = 0; k < count; k++) | |
443 | { | |
aa79a1e1 | 444 | tmp = wide_int_to_tree (word_type_node, test[k].mask); |
531b10fc SB |
445 | tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp); |
446 | tmp = force_gimple_operand_gsi (&gsi, tmp, | |
447 | /*simple=*/true, NULL_TREE, | |
448 | /*before=*/true, GSI_SAME_STMT); | |
449 | tmp = fold_build2 (NE_EXPR, boolean_type_node, tmp, word_mode_zero); | |
450 | new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, test[k].target_edge, | |
451 | update_dom); | |
452 | if (update_dom) | |
9771b263 | 453 | bbs_to_fix_dom.safe_push (new_bb); |
531b10fc SB |
454 | gcc_assert (gimple_bb (swtch) == new_bb); |
455 | gsi = gsi_last_bb (new_bb); | |
456 | } | |
457 | ||
458 | /* We should have removed all edges now. */ | |
459 | gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0); | |
460 | ||
461 | /* If nothing matched, go to the default label. */ | |
462 | make_edge (gsi_bb (gsi), new_default_bb, EDGE_FALLTHRU); | |
463 | ||
464 | /* The GIMPLE_SWITCH is now redundant. */ | |
465 | gsi_remove (&gsi, true); | |
466 | ||
467 | if (update_dom) | |
468 | { | |
469 | /* Fix up the dominator tree. */ | |
470 | iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true); | |
9771b263 | 471 | bbs_to_fix_dom.release (); |
531b10fc SB |
472 | } |
473 | } | |
474 | \f | |
b6e99746 MJ |
475 | /* |
476 | Switch initialization conversion | |
477 | ||
478 | The following pass changes simple initializations of scalars in a switch | |
fade902a SB |
479 | statement into initializations from a static array. Obviously, the values |
480 | must be constant and known at compile time and a default branch must be | |
b6e99746 MJ |
481 | provided. For example, the following code: |
482 | ||
483 | int a,b; | |
484 | ||
485 | switch (argc) | |
486 | { | |
487 | case 1: | |
488 | case 2: | |
489 | a_1 = 8; | |
490 | b_1 = 6; | |
491 | break; | |
492 | case 3: | |
493 | a_2 = 9; | |
494 | b_2 = 5; | |
495 | break; | |
496 | case 12: | |
497 | a_3 = 10; | |
498 | b_3 = 4; | |
499 | break; | |
500 | default: | |
501 | a_4 = 16; | |
502 | b_4 = 1; | |
886cd84f | 503 | break; |
b6e99746 MJ |
504 | } |
505 | a_5 = PHI <a_1, a_2, a_3, a_4> | |
506 | b_5 = PHI <b_1, b_2, b_3, b_4> | |
507 | ||
508 | ||
509 | is changed into: | |
510 | ||
511 | static const int = CSWTCH01[] = {6, 6, 5, 1, 1, 1, 1, 1, 1, 1, 1, 4}; | |
512 | static const int = CSWTCH02[] = {8, 8, 9, 16, 16, 16, 16, 16, 16, 16, | |
513 | 16, 16, 10}; | |
514 | ||
515 | if (((unsigned) argc) - 1 < 11) | |
516 | { | |
517 | a_6 = CSWTCH02[argc - 1]; | |
518 | b_6 = CSWTCH01[argc - 1]; | |
519 | } | |
520 | else | |
521 | { | |
522 | a_7 = 16; | |
523 | b_7 = 1; | |
524 | } | |
886cd84f SB |
525 | a_5 = PHI <a_6, a_7> |
526 | b_b = PHI <b_6, b_7> | |
b6e99746 MJ |
527 | |
528 | There are further constraints. Specifically, the range of values across all | |
529 | case labels must not be bigger than SWITCH_CONVERSION_BRANCH_RATIO (default | |
531b10fc | 530 | eight) times the number of the actual switch branches. |
b6e99746 | 531 | |
531b10fc SB |
532 | This transformation was contributed by Martin Jambor, see this e-mail: |
533 | http://gcc.gnu.org/ml/gcc-patches/2008-07/msg00011.html */ | |
b6e99746 MJ |
534 | |
535 | /* The main structure of the pass. */ | |
536 | struct switch_conv_info | |
537 | { | |
886cd84f | 538 | /* The expression used to decide the switch branch. */ |
b6e99746 MJ |
539 | tree index_expr; |
540 | ||
886cd84f SB |
541 | /* The following integer constants store the minimum and maximum value |
542 | covered by the case labels. */ | |
b6e99746 | 543 | tree range_min; |
886cd84f | 544 | tree range_max; |
b6e99746 | 545 | |
886cd84f SB |
546 | /* The difference between the above two numbers. Stored here because it |
547 | is used in all the conversion heuristics, as well as for some of the | |
548 | transformation, and it is expensive to re-compute it all the time. */ | |
b6e99746 MJ |
549 | tree range_size; |
550 | ||
886cd84f | 551 | /* Basic block that contains the actual GIMPLE_SWITCH. */ |
b6e99746 MJ |
552 | basic_block switch_bb; |
553 | ||
886cd84f SB |
554 | /* Basic block that is the target of the default case. */ |
555 | basic_block default_bb; | |
556 | ||
557 | /* The single successor block of all branches out of the GIMPLE_SWITCH, | |
558 | if such a block exists. Otherwise NULL. */ | |
b6e99746 MJ |
559 | basic_block final_bb; |
560 | ||
886cd84f | 561 | /* The probability of the default edge in the replaced switch. */ |
357067f2 | 562 | profile_probability default_prob; |
886cd84f SB |
563 | |
564 | /* The count of the default edge in the replaced switch. */ | |
3995f3a2 | 565 | profile_count default_count; |
886cd84f SB |
566 | |
567 | /* Combined count of all other (non-default) edges in the replaced switch. */ | |
3995f3a2 | 568 | profile_count other_count; |
886cd84f | 569 | |
b6e99746 MJ |
570 | /* Number of phi nodes in the final bb (that we'll be replacing). */ |
571 | int phi_count; | |
572 | ||
b1ae1681 | 573 | /* Array of default values, in the same order as phi nodes. */ |
b6e99746 MJ |
574 | tree *default_values; |
575 | ||
576 | /* Constructors of new static arrays. */ | |
9771b263 | 577 | vec<constructor_elt, va_gc> **constructors; |
b6e99746 MJ |
578 | |
579 | /* Array of ssa names that are initialized with a value from a new static | |
580 | array. */ | |
581 | tree *target_inbound_names; | |
582 | ||
583 | /* Array of ssa names that are initialized with the default value if the | |
584 | switch expression is out of range. */ | |
585 | tree *target_outbound_names; | |
586 | ||
8dc6a926 JJ |
587 | /* VOP SSA_NAME. */ |
588 | tree target_vop; | |
589 | ||
b1ae1681 MJ |
590 | /* The first load statement that loads a temporary from a new static array. |
591 | */ | |
355fe088 | 592 | gimple *arr_ref_first; |
b6e99746 MJ |
593 | |
594 | /* The last load statement that loads a temporary from a new static array. */ | |
355fe088 | 595 | gimple *arr_ref_last; |
b6e99746 MJ |
596 | |
597 | /* String reason why the case wasn't a good candidate that is written to the | |
598 | dump file, if there is one. */ | |
599 | const char *reason; | |
8e97bc2b | 600 | |
18bfe940 JJ |
601 | /* True if default case is not used for any value between range_min and |
602 | range_max inclusive. */ | |
603 | bool contiguous_range; | |
604 | ||
605 | /* True if default case does not have the required shape for other case | |
606 | labels. */ | |
607 | bool default_case_nonstandard; | |
608 | ||
8e97bc2b JJ |
609 | /* Parameters for expand_switch_using_bit_tests. Should be computed |
610 | the same way as in expand_case. */ | |
886cd84f SB |
611 | unsigned int uniq; |
612 | unsigned int count; | |
b6e99746 MJ |
613 | }; |
614 | ||
886cd84f | 615 | /* Collect information about GIMPLE_SWITCH statement SWTCH into INFO. */ |
b6e99746 | 616 | |
886cd84f | 617 | static void |
538dd0b7 | 618 | collect_switch_conv_info (gswitch *swtch, struct switch_conv_info *info) |
b6e99746 | 619 | { |
726a989a | 620 | unsigned int branch_num = gimple_switch_num_labels (swtch); |
886cd84f SB |
621 | tree min_case, max_case; |
622 | unsigned int count, i; | |
18bfe940 | 623 | edge e, e_default, e_first; |
886cd84f | 624 | edge_iterator ei; |
18bfe940 | 625 | basic_block first; |
886cd84f SB |
626 | |
627 | memset (info, 0, sizeof (*info)); | |
b6e99746 MJ |
628 | |
629 | /* The gimplifier has already sorted the cases by CASE_LOW and ensured there | |
fd8d363e SB |
630 | is a default label which is the first in the vector. |
631 | Collect the bits we can deduce from the CFG. */ | |
886cd84f SB |
632 | info->index_expr = gimple_switch_index (swtch); |
633 | info->switch_bb = gimple_bb (swtch); | |
18bfe940 JJ |
634 | info->default_bb |
635 | = label_to_block (CASE_LABEL (gimple_switch_default_label (swtch))); | |
886cd84f SB |
636 | e_default = find_edge (info->switch_bb, info->default_bb); |
637 | info->default_prob = e_default->probability; | |
638 | info->default_count = e_default->count; | |
639 | FOR_EACH_EDGE (e, ei, info->switch_bb->succs) | |
640 | if (e != e_default) | |
641 | info->other_count += e->count; | |
b6e99746 | 642 | |
18bfe940 JJ |
643 | /* Get upper and lower bounds of case values, and the covered range. */ |
644 | min_case = gimple_switch_label (swtch, 1); | |
645 | max_case = gimple_switch_label (swtch, branch_num - 1); | |
646 | ||
647 | info->range_min = CASE_LOW (min_case); | |
648 | if (CASE_HIGH (max_case) != NULL_TREE) | |
649 | info->range_max = CASE_HIGH (max_case); | |
650 | else | |
651 | info->range_max = CASE_LOW (max_case); | |
652 | ||
653 | info->contiguous_range = true; | |
654 | tree last = CASE_HIGH (min_case) ? CASE_HIGH (min_case) : info->range_min; | |
655 | for (i = 2; i < branch_num; i++) | |
656 | { | |
657 | tree elt = gimple_switch_label (swtch, i); | |
658 | wide_int w = last; | |
659 | if (w + 1 != CASE_LOW (elt)) | |
660 | { | |
661 | info->contiguous_range = false; | |
662 | break; | |
663 | } | |
664 | last = CASE_HIGH (elt) ? CASE_HIGH (elt) : CASE_LOW (elt); | |
665 | } | |
666 | ||
667 | if (info->contiguous_range) | |
668 | { | |
669 | first = label_to_block (CASE_LABEL (gimple_switch_label (swtch, 1))); | |
670 | e_first = find_edge (info->switch_bb, first); | |
671 | } | |
672 | else | |
673 | { | |
674 | first = info->default_bb; | |
675 | e_first = e_default; | |
676 | } | |
677 | ||
886cd84f | 678 | /* See if there is one common successor block for all branch |
866f20d6 | 679 | targets. If it exists, record it in FINAL_BB. |
18bfe940 JJ |
680 | Start with the destination of the first non-default case |
681 | if the range is contiguous and default case otherwise as | |
682 | guess or its destination in case it is a forwarder block. */ | |
683 | if (! single_pred_p (e_first->dest)) | |
684 | info->final_bb = e_first->dest; | |
685 | else if (single_succ_p (e_first->dest) | |
686 | && ! single_pred_p (single_succ (e_first->dest))) | |
687 | info->final_bb = single_succ (e_first->dest); | |
866f20d6 | 688 | /* Require that all switch destinations are either that common |
18bfe940 JJ |
689 | FINAL_BB or a forwarder to it, except for the default |
690 | case if contiguous range. */ | |
886cd84f SB |
691 | if (info->final_bb) |
692 | FOR_EACH_EDGE (e, ei, info->switch_bb->succs) | |
693 | { | |
694 | if (e->dest == info->final_bb) | |
695 | continue; | |
696 | ||
697 | if (single_pred_p (e->dest) | |
698 | && single_succ_p (e->dest) | |
699 | && single_succ (e->dest) == info->final_bb) | |
700 | continue; | |
701 | ||
18bfe940 JJ |
702 | if (e == e_default && info->contiguous_range) |
703 | { | |
704 | info->default_case_nonstandard = true; | |
705 | continue; | |
706 | } | |
707 | ||
886cd84f SB |
708 | info->final_bb = NULL; |
709 | break; | |
710 | } | |
711 | ||
18bfe940 JJ |
712 | info->range_size |
713 | = int_const_binop (MINUS_EXPR, info->range_max, info->range_min); | |
b6e99746 | 714 | |
886cd84f SB |
715 | /* Get a count of the number of case labels. Single-valued case labels |
716 | simply count as one, but a case range counts double, since it may | |
717 | require two compares if it gets lowered as a branching tree. */ | |
718 | count = 0; | |
719 | for (i = 1; i < branch_num; i++) | |
720 | { | |
721 | tree elt = gimple_switch_label (swtch, i); | |
722 | count++; | |
723 | if (CASE_HIGH (elt) | |
724 | && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt))) | |
725 | count++; | |
726 | } | |
727 | info->count = count; | |
728 | ||
729 | /* Get the number of unique non-default targets out of the GIMPLE_SWITCH | |
730 | block. Assume a CFG cleanup would have already removed degenerate | |
731 | switch statements, this allows us to just use EDGE_COUNT. */ | |
732 | info->uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1; | |
733 | } | |
b6e99746 | 734 | |
886cd84f SB |
735 | /* Checks whether the range given by individual case statements of the SWTCH |
736 | switch statement isn't too big and whether the number of branches actually | |
737 | satisfies the size of the new array. */ | |
b6e99746 | 738 | |
886cd84f SB |
739 | static bool |
740 | check_range (struct switch_conv_info *info) | |
741 | { | |
fade902a | 742 | gcc_assert (info->range_size); |
cc269bb6 | 743 | if (!tree_fits_uhwi_p (info->range_size)) |
b6e99746 | 744 | { |
fade902a | 745 | info->reason = "index range way too large or otherwise unusable"; |
b6e99746 MJ |
746 | return false; |
747 | } | |
748 | ||
7d362f6c | 749 | if (tree_to_uhwi (info->range_size) |
886cd84f | 750 | > ((unsigned) info->count * SWITCH_CONVERSION_BRANCH_RATIO)) |
b6e99746 | 751 | { |
fade902a | 752 | info->reason = "the maximum range-branch ratio exceeded"; |
b6e99746 MJ |
753 | return false; |
754 | } | |
755 | ||
756 | return true; | |
757 | } | |
758 | ||
886cd84f | 759 | /* Checks whether all but the FINAL_BB basic blocks are empty. */ |
b6e99746 MJ |
760 | |
761 | static bool | |
886cd84f | 762 | check_all_empty_except_final (struct switch_conv_info *info) |
b6e99746 | 763 | { |
18bfe940 | 764 | edge e, e_default = find_edge (info->switch_bb, info->default_bb); |
886cd84f | 765 | edge_iterator ei; |
b6e99746 | 766 | |
886cd84f | 767 | FOR_EACH_EDGE (e, ei, info->switch_bb->succs) |
b6e99746 | 768 | { |
886cd84f SB |
769 | if (e->dest == info->final_bb) |
770 | continue; | |
b6e99746 | 771 | |
886cd84f | 772 | if (!empty_block_p (e->dest)) |
b6e99746 | 773 | { |
18bfe940 JJ |
774 | if (info->contiguous_range && e == e_default) |
775 | { | |
776 | info->default_case_nonstandard = true; | |
777 | continue; | |
778 | } | |
779 | ||
fade902a | 780 | info->reason = "bad case - a non-final BB not empty"; |
b6e99746 MJ |
781 | return false; |
782 | } | |
b6e99746 MJ |
783 | } |
784 | ||
785 | return true; | |
786 | } | |
787 | ||
788 | /* This function checks whether all required values in phi nodes in final_bb | |
789 | are constants. Required values are those that correspond to a basic block | |
790 | which is a part of the examined switch statement. It returns true if the | |
791 | phi nodes are OK, otherwise false. */ | |
792 | ||
793 | static bool | |
18bfe940 | 794 | check_final_bb (gswitch *swtch, struct switch_conv_info *info) |
b6e99746 | 795 | { |
538dd0b7 | 796 | gphi_iterator gsi; |
b6e99746 | 797 | |
fade902a SB |
798 | info->phi_count = 0; |
799 | for (gsi = gsi_start_phis (info->final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
b6e99746 | 800 | { |
538dd0b7 | 801 | gphi *phi = gsi.phi (); |
726a989a | 802 | unsigned int i; |
b6e99746 | 803 | |
18bfe940 JJ |
804 | if (virtual_operand_p (gimple_phi_result (phi))) |
805 | continue; | |
806 | ||
fade902a | 807 | info->phi_count++; |
b6e99746 | 808 | |
726a989a | 809 | for (i = 0; i < gimple_phi_num_args (phi); i++) |
b6e99746 | 810 | { |
726a989a | 811 | basic_block bb = gimple_phi_arg_edge (phi, i)->src; |
b6e99746 | 812 | |
fade902a | 813 | if (bb == info->switch_bb |
18bfe940 JJ |
814 | || (single_pred_p (bb) |
815 | && single_pred (bb) == info->switch_bb | |
816 | && (!info->default_case_nonstandard | |
817 | || empty_block_p (bb)))) | |
b6e99746 | 818 | { |
f6e6e990 | 819 | tree reloc, val; |
18bfe940 | 820 | const char *reason = NULL; |
f6e6e990 JJ |
821 | |
822 | val = gimple_phi_arg_def (phi, i); | |
823 | if (!is_gimple_ip_invariant (val)) | |
18bfe940 JJ |
824 | reason = "non-invariant value from a case"; |
825 | else | |
f6e6e990 | 826 | { |
18bfe940 JJ |
827 | reloc = initializer_constant_valid_p (val, TREE_TYPE (val)); |
828 | if ((flag_pic && reloc != null_pointer_node) | |
829 | || (!flag_pic && reloc == NULL_TREE)) | |
830 | { | |
831 | if (reloc) | |
832 | reason | |
833 | = "value from a case would need runtime relocations"; | |
834 | else | |
835 | reason | |
836 | = "value from a case is not a valid initializer"; | |
837 | } | |
f6e6e990 | 838 | } |
18bfe940 | 839 | if (reason) |
f6e6e990 | 840 | { |
18bfe940 JJ |
841 | /* For contiguous range, we can allow non-constant |
842 | or one that needs relocation, as long as it is | |
843 | only reachable from the default case. */ | |
844 | if (bb == info->switch_bb) | |
845 | bb = info->final_bb; | |
846 | if (!info->contiguous_range || bb != info->default_bb) | |
847 | { | |
848 | info->reason = reason; | |
849 | return false; | |
850 | } | |
851 | ||
852 | unsigned int branch_num = gimple_switch_num_labels (swtch); | |
853 | for (unsigned int i = 1; i < branch_num; i++) | |
854 | { | |
855 | tree lab = CASE_LABEL (gimple_switch_label (swtch, i)); | |
856 | if (label_to_block (lab) == bb) | |
857 | { | |
858 | info->reason = reason; | |
859 | return false; | |
860 | } | |
861 | } | |
862 | info->default_case_nonstandard = true; | |
f6e6e990 | 863 | } |
b6e99746 MJ |
864 | } |
865 | } | |
866 | } | |
867 | ||
868 | return true; | |
869 | } | |
870 | ||
871 | /* The following function allocates default_values, target_{in,out}_names and | |
872 | constructors arrays. The last one is also populated with pointers to | |
873 | vectors that will become constructors of new arrays. */ | |
874 | ||
875 | static void | |
fade902a | 876 | create_temp_arrays (struct switch_conv_info *info) |
b6e99746 MJ |
877 | { |
878 | int i; | |
879 | ||
fade902a | 880 | info->default_values = XCNEWVEC (tree, info->phi_count * 3); |
9771b263 DN |
881 | /* ??? Macros do not support multi argument templates in their |
882 | argument list. We create a typedef to work around that problem. */ | |
883 | typedef vec<constructor_elt, va_gc> *vec_constructor_elt_gc; | |
884 | info->constructors = XCNEWVEC (vec_constructor_elt_gc, info->phi_count); | |
fade902a SB |
885 | info->target_inbound_names = info->default_values + info->phi_count; |
886 | info->target_outbound_names = info->target_inbound_names + info->phi_count; | |
887 | for (i = 0; i < info->phi_count; i++) | |
ae7e9ddd | 888 | vec_alloc (info->constructors[i], tree_to_uhwi (info->range_size) + 1); |
b6e99746 MJ |
889 | } |
890 | ||
891 | /* Free the arrays created by create_temp_arrays(). The vectors that are | |
892 | created by that function are not freed here, however, because they have | |
893 | already become constructors and must be preserved. */ | |
894 | ||
895 | static void | |
fade902a | 896 | free_temp_arrays (struct switch_conv_info *info) |
b6e99746 | 897 | { |
fade902a SB |
898 | XDELETEVEC (info->constructors); |
899 | XDELETEVEC (info->default_values); | |
b6e99746 MJ |
900 | } |
901 | ||
902 | /* Populate the array of default values in the order of phi nodes. | |
18bfe940 JJ |
903 | DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch |
904 | if the range is non-contiguous or the default case has standard | |
905 | structure, otherwise it is the first non-default case instead. */ | |
b6e99746 MJ |
906 | |
907 | static void | |
fade902a | 908 | gather_default_values (tree default_case, struct switch_conv_info *info) |
b6e99746 | 909 | { |
538dd0b7 | 910 | gphi_iterator gsi; |
b6e99746 MJ |
911 | basic_block bb = label_to_block (CASE_LABEL (default_case)); |
912 | edge e; | |
726a989a | 913 | int i = 0; |
b6e99746 | 914 | |
18bfe940 JJ |
915 | gcc_assert (CASE_LOW (default_case) == NULL_TREE |
916 | || info->default_case_nonstandard); | |
b6e99746 | 917 | |
fade902a SB |
918 | if (bb == info->final_bb) |
919 | e = find_edge (info->switch_bb, bb); | |
b6e99746 MJ |
920 | else |
921 | e = single_succ_edge (bb); | |
922 | ||
fade902a | 923 | for (gsi = gsi_start_phis (info->final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
b6e99746 | 924 | { |
538dd0b7 | 925 | gphi *phi = gsi.phi (); |
18bfe940 JJ |
926 | if (virtual_operand_p (gimple_phi_result (phi))) |
927 | continue; | |
b6e99746 MJ |
928 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
929 | gcc_assert (val); | |
fade902a | 930 | info->default_values[i++] = val; |
b6e99746 MJ |
931 | } |
932 | } | |
933 | ||
934 | /* The following function populates the vectors in the constructors array with | |
935 | future contents of the static arrays. The vectors are populated in the | |
936 | order of phi nodes. SWTCH is the switch statement being converted. */ | |
937 | ||
938 | static void | |
538dd0b7 | 939 | build_constructors (gswitch *swtch, struct switch_conv_info *info) |
b6e99746 | 940 | { |
726a989a | 941 | unsigned i, branch_num = gimple_switch_num_labels (swtch); |
fade902a | 942 | tree pos = info->range_min; |
18bfe940 | 943 | tree pos_one = build_int_cst (TREE_TYPE (pos), 1); |
b6e99746 | 944 | |
726a989a | 945 | for (i = 1; i < branch_num; i++) |
b6e99746 | 946 | { |
726a989a | 947 | tree cs = gimple_switch_label (swtch, i); |
b6e99746 MJ |
948 | basic_block bb = label_to_block (CASE_LABEL (cs)); |
949 | edge e; | |
726a989a | 950 | tree high; |
538dd0b7 | 951 | gphi_iterator gsi; |
b6e99746 MJ |
952 | int j; |
953 | ||
fade902a SB |
954 | if (bb == info->final_bb) |
955 | e = find_edge (info->switch_bb, bb); | |
b6e99746 MJ |
956 | else |
957 | e = single_succ_edge (bb); | |
958 | gcc_assert (e); | |
959 | ||
960 | while (tree_int_cst_lt (pos, CASE_LOW (cs))) | |
961 | { | |
962 | int k; | |
18bfe940 | 963 | gcc_assert (!info->contiguous_range); |
fade902a | 964 | for (k = 0; k < info->phi_count; k++) |
b6e99746 | 965 | { |
f32682ca | 966 | constructor_elt elt; |
b6e99746 | 967 | |
f32682ca | 968 | elt.index = int_const_binop (MINUS_EXPR, pos, info->range_min); |
d1f98542 RB |
969 | elt.value |
970 | = unshare_expr_without_location (info->default_values[k]); | |
9771b263 | 971 | info->constructors[k]->quick_push (elt); |
b6e99746 MJ |
972 | } |
973 | ||
18bfe940 | 974 | pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
b6e99746 | 975 | } |
b1ae1681 | 976 | gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs))); |
b6e99746 MJ |
977 | |
978 | j = 0; | |
979 | if (CASE_HIGH (cs)) | |
980 | high = CASE_HIGH (cs); | |
981 | else | |
b1ae1681 | 982 | high = CASE_LOW (cs); |
fade902a | 983 | for (gsi = gsi_start_phis (info->final_bb); |
726a989a | 984 | !gsi_end_p (gsi); gsi_next (&gsi)) |
b6e99746 | 985 | { |
538dd0b7 | 986 | gphi *phi = gsi.phi (); |
18bfe940 JJ |
987 | if (virtual_operand_p (gimple_phi_result (phi))) |
988 | continue; | |
b6e99746 | 989 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
7f2a9982 | 990 | tree low = CASE_LOW (cs); |
b6e99746 MJ |
991 | pos = CASE_LOW (cs); |
992 | ||
b8698a0f | 993 | do |
b6e99746 | 994 | { |
f32682ca | 995 | constructor_elt elt; |
b6e99746 | 996 | |
f32682ca | 997 | elt.index = int_const_binop (MINUS_EXPR, pos, info->range_min); |
d1f98542 | 998 | elt.value = unshare_expr_without_location (val); |
9771b263 | 999 | info->constructors[j]->quick_push (elt); |
b6e99746 | 1000 | |
18bfe940 | 1001 | pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
7156c8ab MJ |
1002 | } while (!tree_int_cst_lt (high, pos) |
1003 | && tree_int_cst_lt (low, pos)); | |
b6e99746 MJ |
1004 | j++; |
1005 | } | |
1006 | } | |
1007 | } | |
1008 | ||
7156c8ab MJ |
1009 | /* If all values in the constructor vector are the same, return the value. |
1010 | Otherwise return NULL_TREE. Not supposed to be called for empty | |
1011 | vectors. */ | |
1012 | ||
1013 | static tree | |
9771b263 | 1014 | constructor_contains_same_values_p (vec<constructor_elt, va_gc> *vec) |
7156c8ab | 1015 | { |
8e97bc2b | 1016 | unsigned int i; |
7156c8ab | 1017 | tree prev = NULL_TREE; |
8e97bc2b | 1018 | constructor_elt *elt; |
7156c8ab | 1019 | |
9771b263 | 1020 | FOR_EACH_VEC_SAFE_ELT (vec, i, elt) |
7156c8ab | 1021 | { |
7156c8ab MJ |
1022 | if (!prev) |
1023 | prev = elt->value; | |
1024 | else if (!operand_equal_p (elt->value, prev, OEP_ONLY_CONST)) | |
1025 | return NULL_TREE; | |
1026 | } | |
1027 | return prev; | |
1028 | } | |
1029 | ||
f1b0632a OH |
1030 | /* Return type which should be used for array elements, either TYPE's |
1031 | main variant or, for integral types, some smaller integral type | |
1032 | that can still hold all the constants. */ | |
8e97bc2b JJ |
1033 | |
1034 | static tree | |
538dd0b7 | 1035 | array_value_type (gswitch *swtch, tree type, int num, |
fade902a | 1036 | struct switch_conv_info *info) |
8e97bc2b | 1037 | { |
9771b263 | 1038 | unsigned int i, len = vec_safe_length (info->constructors[num]); |
8e97bc2b | 1039 | constructor_elt *elt; |
8e97bc2b JJ |
1040 | int sign = 0; |
1041 | tree smaller_type; | |
1042 | ||
f1b0632a OH |
1043 | /* Types with alignments greater than their size can reach here, e.g. out of |
1044 | SRA. We couldn't use these as an array component type so get back to the | |
1045 | main variant first, which, for our purposes, is fine for other types as | |
1046 | well. */ | |
1047 | ||
1048 | type = TYPE_MAIN_VARIANT (type); | |
1049 | ||
8e97bc2b JJ |
1050 | if (!INTEGRAL_TYPE_P (type)) |
1051 | return type; | |
1052 | ||
ec35d572 RS |
1053 | machine_mode type_mode = TYPE_MODE (type); |
1054 | machine_mode mode = get_narrowest_mode (type_mode); | |
1055 | if (GET_MODE_SIZE (type_mode) <= GET_MODE_SIZE (mode)) | |
8e97bc2b JJ |
1056 | return type; |
1057 | ||
1058 | if (len < (optimize_bb_for_size_p (gimple_bb (swtch)) ? 2 : 32)) | |
1059 | return type; | |
1060 | ||
9771b263 | 1061 | FOR_EACH_VEC_SAFE_ELT (info->constructors[num], i, elt) |
8e97bc2b | 1062 | { |
807e902e | 1063 | wide_int cst; |
8e97bc2b JJ |
1064 | |
1065 | if (TREE_CODE (elt->value) != INTEGER_CST) | |
1066 | return type; | |
1067 | ||
807e902e | 1068 | cst = elt->value; |
8e97bc2b JJ |
1069 | while (1) |
1070 | { | |
1071 | unsigned int prec = GET_MODE_BITSIZE (mode); | |
1072 | if (prec > HOST_BITS_PER_WIDE_INT) | |
1073 | return type; | |
1074 | ||
807e902e | 1075 | if (sign >= 0 && cst == wi::zext (cst, prec)) |
8e97bc2b | 1076 | { |
807e902e | 1077 | if (sign == 0 && cst == wi::sext (cst, prec)) |
8e97bc2b JJ |
1078 | break; |
1079 | sign = 1; | |
1080 | break; | |
1081 | } | |
807e902e | 1082 | if (sign <= 0 && cst == wi::sext (cst, prec)) |
8e97bc2b JJ |
1083 | { |
1084 | sign = -1; | |
1085 | break; | |
1086 | } | |
1087 | ||
1088 | if (sign == 1) | |
1089 | sign = 0; | |
1090 | ||
1091 | mode = GET_MODE_WIDER_MODE (mode); | |
1092 | if (mode == VOIDmode | |
ec35d572 | 1093 | || GET_MODE_SIZE (mode) >= GET_MODE_SIZE (type_mode)) |
8e97bc2b JJ |
1094 | return type; |
1095 | } | |
1096 | } | |
1097 | ||
1098 | if (sign == 0) | |
1099 | sign = TYPE_UNSIGNED (type) ? 1 : -1; | |
1100 | smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0); | |
1101 | if (GET_MODE_SIZE (TYPE_MODE (type)) | |
1102 | <= GET_MODE_SIZE (TYPE_MODE (smaller_type))) | |
1103 | return type; | |
1104 | ||
1105 | return smaller_type; | |
1106 | } | |
1107 | ||
b6e99746 MJ |
1108 | /* Create an appropriate array type and declaration and assemble a static array |
1109 | variable. Also create a load statement that initializes the variable in | |
1110 | question with a value from the static array. SWTCH is the switch statement | |
1111 | being converted, NUM is the index to arrays of constructors, default values | |
1112 | and target SSA names for this particular array. ARR_INDEX_TYPE is the type | |
1113 | of the index of the new array, PHI is the phi node of the final BB that | |
1114 | corresponds to the value that will be loaded from the created array. TIDX | |
7156c8ab MJ |
1115 | is an ssa name of a temporary variable holding the index for loads from the |
1116 | new array. */ | |
b6e99746 MJ |
1117 | |
1118 | static void | |
538dd0b7 DM |
1119 | build_one_array (gswitch *swtch, int num, tree arr_index_type, |
1120 | gphi *phi, tree tidx, struct switch_conv_info *info) | |
b6e99746 | 1121 | { |
7156c8ab | 1122 | tree name, cst; |
355fe088 | 1123 | gimple *load; |
7156c8ab | 1124 | gimple_stmt_iterator gsi = gsi_for_stmt (swtch); |
c2255bc4 | 1125 | location_t loc = gimple_location (swtch); |
b6e99746 | 1126 | |
fade902a | 1127 | gcc_assert (info->default_values[num]); |
b6e99746 | 1128 | |
b731b390 | 1129 | name = copy_ssa_name (PHI_RESULT (phi)); |
fade902a | 1130 | info->target_inbound_names[num] = name; |
b6e99746 | 1131 | |
fade902a | 1132 | cst = constructor_contains_same_values_p (info->constructors[num]); |
7156c8ab MJ |
1133 | if (cst) |
1134 | load = gimple_build_assign (name, cst); | |
1135 | else | |
1136 | { | |
8e97bc2b | 1137 | tree array_type, ctor, decl, value_type, fetch, default_type; |
7156c8ab | 1138 | |
fade902a SB |
1139 | default_type = TREE_TYPE (info->default_values[num]); |
1140 | value_type = array_value_type (swtch, default_type, num, info); | |
7156c8ab | 1141 | array_type = build_array_type (value_type, arr_index_type); |
8e97bc2b JJ |
1142 | if (default_type != value_type) |
1143 | { | |
1144 | unsigned int i; | |
1145 | constructor_elt *elt; | |
1146 | ||
9771b263 | 1147 | FOR_EACH_VEC_SAFE_ELT (info->constructors[num], i, elt) |
8e97bc2b JJ |
1148 | elt->value = fold_convert (value_type, elt->value); |
1149 | } | |
fade902a | 1150 | ctor = build_constructor (array_type, info->constructors[num]); |
7156c8ab | 1151 | TREE_CONSTANT (ctor) = true; |
5f7ae6b6 | 1152 | TREE_STATIC (ctor) = true; |
7156c8ab | 1153 | |
c2255bc4 | 1154 | decl = build_decl (loc, VAR_DECL, NULL_TREE, array_type); |
7156c8ab MJ |
1155 | TREE_STATIC (decl) = 1; |
1156 | DECL_INITIAL (decl) = ctor; | |
1157 | ||
1158 | DECL_NAME (decl) = create_tmp_var_name ("CSWTCH"); | |
1159 | DECL_ARTIFICIAL (decl) = 1; | |
f8d851c6 | 1160 | DECL_IGNORED_P (decl) = 1; |
7156c8ab | 1161 | TREE_CONSTANT (decl) = 1; |
2e3b4885 | 1162 | TREE_READONLY (decl) = 1; |
d7438551 | 1163 | DECL_IGNORED_P (decl) = 1; |
9041d2e6 | 1164 | varpool_node::finalize_decl (decl); |
7156c8ab MJ |
1165 | |
1166 | fetch = build4 (ARRAY_REF, value_type, decl, tidx, NULL_TREE, | |
1167 | NULL_TREE); | |
8e97bc2b JJ |
1168 | if (default_type != value_type) |
1169 | { | |
1170 | fetch = fold_convert (default_type, fetch); | |
1171 | fetch = force_gimple_operand_gsi (&gsi, fetch, true, NULL_TREE, | |
1172 | true, GSI_SAME_STMT); | |
1173 | } | |
7156c8ab MJ |
1174 | load = gimple_build_assign (name, fetch); |
1175 | } | |
b6e99746 | 1176 | |
726a989a | 1177 | gsi_insert_before (&gsi, load, GSI_SAME_STMT); |
7156c8ab | 1178 | update_stmt (load); |
fade902a | 1179 | info->arr_ref_last = load; |
b6e99746 MJ |
1180 | } |
1181 | ||
1182 | /* Builds and initializes static arrays initialized with values gathered from | |
1183 | the SWTCH switch statement. Also creates statements that load values from | |
1184 | them. */ | |
1185 | ||
1186 | static void | |
538dd0b7 | 1187 | build_arrays (gswitch *swtch, struct switch_conv_info *info) |
b6e99746 MJ |
1188 | { |
1189 | tree arr_index_type; | |
83d5977e | 1190 | tree tidx, sub, utype; |
355fe088 | 1191 | gimple *stmt; |
726a989a | 1192 | gimple_stmt_iterator gsi; |
538dd0b7 | 1193 | gphi_iterator gpi; |
b6e99746 | 1194 | int i; |
db3927fb | 1195 | location_t loc = gimple_location (swtch); |
b6e99746 | 1196 | |
726a989a | 1197 | gsi = gsi_for_stmt (swtch); |
04e78aa9 | 1198 | |
edb9b69e | 1199 | /* Make sure we do not generate arithmetics in a subrange. */ |
fade902a | 1200 | utype = TREE_TYPE (info->index_expr); |
edb9b69e JJ |
1201 | if (TREE_TYPE (utype)) |
1202 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1); | |
1203 | else | |
1204 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1); | |
1205 | ||
fade902a | 1206 | arr_index_type = build_index_type (info->range_size); |
b731b390 | 1207 | tidx = make_ssa_name (utype); |
edb9b69e | 1208 | sub = fold_build2_loc (loc, MINUS_EXPR, utype, |
fade902a SB |
1209 | fold_convert_loc (loc, utype, info->index_expr), |
1210 | fold_convert_loc (loc, utype, info->range_min)); | |
fae1034e | 1211 | sub = force_gimple_operand_gsi (&gsi, sub, |
726a989a RB |
1212 | false, NULL, true, GSI_SAME_STMT); |
1213 | stmt = gimple_build_assign (tidx, sub); | |
b6e99746 | 1214 | |
726a989a | 1215 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
7156c8ab | 1216 | update_stmt (stmt); |
fade902a | 1217 | info->arr_ref_first = stmt; |
b6e99746 | 1218 | |
538dd0b7 | 1219 | for (gpi = gsi_start_phis (info->final_bb), i = 0; |
18bfe940 JJ |
1220 | !gsi_end_p (gpi); gsi_next (&gpi)) |
1221 | { | |
1222 | gphi *phi = gpi.phi (); | |
1223 | if (!virtual_operand_p (gimple_phi_result (phi))) | |
1224 | build_one_array (swtch, i++, arr_index_type, phi, tidx, info); | |
8dc6a926 JJ |
1225 | else |
1226 | { | |
1227 | edge e; | |
1228 | edge_iterator ei; | |
1229 | FOR_EACH_EDGE (e, ei, info->switch_bb->succs) | |
1230 | { | |
1231 | if (e->dest == info->final_bb) | |
1232 | break; | |
1233 | if (!info->default_case_nonstandard | |
1234 | || e->dest != info->default_bb) | |
1235 | { | |
1236 | e = single_succ_edge (e->dest); | |
1237 | break; | |
1238 | } | |
1239 | } | |
1240 | gcc_assert (e && e->dest == info->final_bb); | |
1241 | info->target_vop = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
1242 | } | |
18bfe940 | 1243 | } |
b6e99746 MJ |
1244 | } |
1245 | ||
1246 | /* Generates and appropriately inserts loads of default values at the position | |
1247 | given by BSI. Returns the last inserted statement. */ | |
1248 | ||
538dd0b7 | 1249 | static gassign * |
fade902a | 1250 | gen_def_assigns (gimple_stmt_iterator *gsi, struct switch_conv_info *info) |
b6e99746 MJ |
1251 | { |
1252 | int i; | |
538dd0b7 | 1253 | gassign *assign = NULL; |
b6e99746 | 1254 | |
fade902a | 1255 | for (i = 0; i < info->phi_count; i++) |
b6e99746 | 1256 | { |
b731b390 | 1257 | tree name = copy_ssa_name (info->target_inbound_names[i]); |
fade902a SB |
1258 | info->target_outbound_names[i] = name; |
1259 | assign = gimple_build_assign (name, info->default_values[i]); | |
726a989a | 1260 | gsi_insert_before (gsi, assign, GSI_SAME_STMT); |
7156c8ab | 1261 | update_stmt (assign); |
b6e99746 MJ |
1262 | } |
1263 | return assign; | |
1264 | } | |
1265 | ||
1266 | /* Deletes the unused bbs and edges that now contain the switch statement and | |
1267 | its empty branch bbs. BBD is the now dead BB containing the original switch | |
1268 | statement, FINAL is the last BB of the converted switch statement (in terms | |
1269 | of succession). */ | |
1270 | ||
1271 | static void | |
18bfe940 | 1272 | prune_bbs (basic_block bbd, basic_block final, basic_block default_bb) |
b6e99746 MJ |
1273 | { |
1274 | edge_iterator ei; | |
1275 | edge e; | |
1276 | ||
1277 | for (ei = ei_start (bbd->succs); (e = ei_safe_edge (ei)); ) | |
1278 | { | |
1279 | basic_block bb; | |
1280 | bb = e->dest; | |
1281 | remove_edge (e); | |
18bfe940 | 1282 | if (bb != final && bb != default_bb) |
b6e99746 MJ |
1283 | delete_basic_block (bb); |
1284 | } | |
1285 | delete_basic_block (bbd); | |
1286 | } | |
1287 | ||
1288 | /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge | |
1289 | from the basic block loading values from an array and E2F from the basic | |
1290 | block loading default values. BBF is the last switch basic block (see the | |
1291 | bbf description in the comment below). */ | |
1292 | ||
1293 | static void | |
fade902a SB |
1294 | fix_phi_nodes (edge e1f, edge e2f, basic_block bbf, |
1295 | struct switch_conv_info *info) | |
b6e99746 | 1296 | { |
538dd0b7 | 1297 | gphi_iterator gsi; |
b6e99746 MJ |
1298 | int i; |
1299 | ||
726a989a | 1300 | for (gsi = gsi_start_phis (bbf), i = 0; |
18bfe940 | 1301 | !gsi_end_p (gsi); gsi_next (&gsi)) |
b6e99746 | 1302 | { |
538dd0b7 | 1303 | gphi *phi = gsi.phi (); |
18bfe940 JJ |
1304 | tree inbound, outbound; |
1305 | if (virtual_operand_p (gimple_phi_result (phi))) | |
8dc6a926 | 1306 | inbound = outbound = info->target_vop; |
18bfe940 JJ |
1307 | else |
1308 | { | |
1309 | inbound = info->target_inbound_names[i]; | |
1310 | outbound = info->target_outbound_names[i++]; | |
1311 | } | |
1312 | add_phi_arg (phi, inbound, e1f, UNKNOWN_LOCATION); | |
1313 | if (!info->default_case_nonstandard) | |
1314 | add_phi_arg (phi, outbound, e2f, UNKNOWN_LOCATION); | |
b6e99746 | 1315 | } |
b6e99746 MJ |
1316 | } |
1317 | ||
1318 | /* Creates a check whether the switch expression value actually falls into the | |
1319 | range given by all the cases. If it does not, the temporaries are loaded | |
1320 | with default values instead. SWTCH is the switch statement being converted. | |
1321 | ||
1322 | bb0 is the bb with the switch statement, however, we'll end it with a | |
1323 | condition instead. | |
1324 | ||
1325 | bb1 is the bb to be used when the range check went ok. It is derived from | |
1326 | the switch BB | |
1327 | ||
1328 | bb2 is the bb taken when the expression evaluated outside of the range | |
1329 | covered by the created arrays. It is populated by loads of default | |
1330 | values. | |
1331 | ||
1332 | bbF is a fall through for both bb1 and bb2 and contains exactly what | |
1333 | originally followed the switch statement. | |
1334 | ||
1335 | bbD contains the switch statement (in the end). It is unreachable but we | |
1336 | still need to strip off its edges. | |
1337 | */ | |
1338 | ||
1339 | static void | |
538dd0b7 | 1340 | gen_inbound_check (gswitch *swtch, struct switch_conv_info *info) |
b6e99746 | 1341 | { |
c2255bc4 AH |
1342 | tree label_decl1 = create_artificial_label (UNKNOWN_LOCATION); |
1343 | tree label_decl2 = create_artificial_label (UNKNOWN_LOCATION); | |
1344 | tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION); | |
538dd0b7 | 1345 | glabel *label1, *label2, *label3; |
edb9b69e | 1346 | tree utype, tidx; |
b6e99746 MJ |
1347 | tree bound; |
1348 | ||
538dd0b7 | 1349 | gcond *cond_stmt; |
b6e99746 | 1350 | |
18bfe940 | 1351 | gassign *last_assign = NULL; |
726a989a | 1352 | gimple_stmt_iterator gsi; |
b6e99746 | 1353 | basic_block bb0, bb1, bb2, bbf, bbd; |
18bfe940 | 1354 | edge e01 = NULL, e02, e21, e1d, e1f, e2f; |
db3927fb | 1355 | location_t loc = gimple_location (swtch); |
b6e99746 | 1356 | |
fade902a | 1357 | gcc_assert (info->default_values); |
6ab1ab14 | 1358 | |
726a989a | 1359 | bb0 = gimple_bb (swtch); |
b6e99746 | 1360 | |
fade902a | 1361 | tidx = gimple_assign_lhs (info->arr_ref_first); |
edb9b69e | 1362 | utype = TREE_TYPE (tidx); |
145544ab | 1363 | |
b6e99746 | 1364 | /* (end of) block 0 */ |
fade902a | 1365 | gsi = gsi_for_stmt (info->arr_ref_first); |
edb9b69e | 1366 | gsi_next (&gsi); |
b6e99746 | 1367 | |
fade902a | 1368 | bound = fold_convert_loc (loc, utype, info->range_size); |
edb9b69e | 1369 | cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE); |
726a989a | 1370 | gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
7156c8ab | 1371 | update_stmt (cond_stmt); |
b6e99746 MJ |
1372 | |
1373 | /* block 2 */ | |
18bfe940 JJ |
1374 | if (!info->default_case_nonstandard) |
1375 | { | |
1376 | label2 = gimple_build_label (label_decl2); | |
1377 | gsi_insert_before (&gsi, label2, GSI_SAME_STMT); | |
1378 | last_assign = gen_def_assigns (&gsi, info); | |
1379 | } | |
b6e99746 MJ |
1380 | |
1381 | /* block 1 */ | |
726a989a RB |
1382 | label1 = gimple_build_label (label_decl1); |
1383 | gsi_insert_before (&gsi, label1, GSI_SAME_STMT); | |
b6e99746 MJ |
1384 | |
1385 | /* block F */ | |
fade902a | 1386 | gsi = gsi_start_bb (info->final_bb); |
726a989a RB |
1387 | label3 = gimple_build_label (label_decl3); |
1388 | gsi_insert_before (&gsi, label3, GSI_SAME_STMT); | |
b6e99746 MJ |
1389 | |
1390 | /* cfg fix */ | |
726a989a | 1391 | e02 = split_block (bb0, cond_stmt); |
b6e99746 MJ |
1392 | bb2 = e02->dest; |
1393 | ||
18bfe940 JJ |
1394 | if (info->default_case_nonstandard) |
1395 | { | |
1396 | bb1 = bb2; | |
1397 | bb2 = info->default_bb; | |
1398 | e01 = e02; | |
1399 | e01->flags = EDGE_TRUE_VALUE; | |
1400 | e02 = make_edge (bb0, bb2, EDGE_FALSE_VALUE); | |
1401 | edge e_default = find_edge (bb1, bb2); | |
1402 | for (gphi_iterator gsi = gsi_start_phis (bb2); | |
1403 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
1404 | { | |
1405 | gphi *phi = gsi.phi (); | |
1406 | tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_default); | |
1407 | add_phi_arg (phi, arg, e02, | |
1408 | gimple_phi_arg_location_from_edge (phi, e_default)); | |
1409 | } | |
1410 | /* Partially fix the dominator tree, if it is available. */ | |
1411 | if (dom_info_available_p (CDI_DOMINATORS)) | |
1412 | redirect_immediate_dominators (CDI_DOMINATORS, bb1, bb0); | |
1413 | } | |
1414 | else | |
1415 | { | |
1416 | e21 = split_block (bb2, last_assign); | |
1417 | bb1 = e21->dest; | |
1418 | remove_edge (e21); | |
1419 | } | |
b6e99746 | 1420 | |
fade902a | 1421 | e1d = split_block (bb1, info->arr_ref_last); |
b6e99746 MJ |
1422 | bbd = e1d->dest; |
1423 | remove_edge (e1d); | |
1424 | ||
1425 | /* flags and profiles of the edge for in-range values */ | |
18bfe940 JJ |
1426 | if (!info->default_case_nonstandard) |
1427 | e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE); | |
357067f2 | 1428 | e01->probability = info->default_prob.invert (); |
fade902a | 1429 | e01->count = info->other_count; |
b6e99746 MJ |
1430 | |
1431 | /* flags and profiles of the edge taking care of out-of-range values */ | |
1432 | e02->flags &= ~EDGE_FALLTHRU; | |
1433 | e02->flags |= EDGE_FALSE_VALUE; | |
fade902a SB |
1434 | e02->probability = info->default_prob; |
1435 | e02->count = info->default_count; | |
b6e99746 | 1436 | |
fade902a | 1437 | bbf = info->final_bb; |
b6e99746 MJ |
1438 | |
1439 | e1f = make_edge (bb1, bbf, EDGE_FALLTHRU); | |
357067f2 | 1440 | e1f->probability = profile_probability::always (); |
fade902a | 1441 | e1f->count = info->other_count; |
b6e99746 | 1442 | |
18bfe940 JJ |
1443 | if (info->default_case_nonstandard) |
1444 | e2f = NULL; | |
1445 | else | |
1446 | { | |
1447 | e2f = make_edge (bb2, bbf, EDGE_FALLTHRU); | |
357067f2 | 1448 | e2f->probability = profile_probability::always (); |
18bfe940 JJ |
1449 | e2f->count = info->default_count; |
1450 | } | |
b6e99746 MJ |
1451 | |
1452 | /* frequencies of the new BBs */ | |
1453 | bb1->frequency = EDGE_FREQUENCY (e01); | |
1454 | bb2->frequency = EDGE_FREQUENCY (e02); | |
18bfe940 JJ |
1455 | if (!info->default_case_nonstandard) |
1456 | bbf->frequency = EDGE_FREQUENCY (e1f) + EDGE_FREQUENCY (e2f); | |
b6e99746 | 1457 | |
6ab1ab14 | 1458 | /* Tidy blocks that have become unreachable. */ |
18bfe940 JJ |
1459 | prune_bbs (bbd, info->final_bb, |
1460 | info->default_case_nonstandard ? info->default_bb : NULL); | |
b6e99746 | 1461 | |
6ab1ab14 | 1462 | /* Fixup the PHI nodes in bbF. */ |
fade902a | 1463 | fix_phi_nodes (e1f, e2f, bbf, info); |
b6e99746 | 1464 | |
6ab1ab14 SB |
1465 | /* Fix the dominator tree, if it is available. */ |
1466 | if (dom_info_available_p (CDI_DOMINATORS)) | |
1467 | { | |
9771b263 | 1468 | vec<basic_block> bbs_to_fix_dom; |
6ab1ab14 SB |
1469 | |
1470 | set_immediate_dominator (CDI_DOMINATORS, bb1, bb0); | |
18bfe940 JJ |
1471 | if (!info->default_case_nonstandard) |
1472 | set_immediate_dominator (CDI_DOMINATORS, bb2, bb0); | |
531b10fc | 1473 | if (! get_immediate_dominator (CDI_DOMINATORS, bbf)) |
6ab1ab14 SB |
1474 | /* If bbD was the immediate dominator ... */ |
1475 | set_immediate_dominator (CDI_DOMINATORS, bbf, bb0); | |
1476 | ||
18bfe940 | 1477 | bbs_to_fix_dom.create (3 + (bb2 != bbf)); |
9771b263 DN |
1478 | bbs_to_fix_dom.quick_push (bb0); |
1479 | bbs_to_fix_dom.quick_push (bb1); | |
18bfe940 JJ |
1480 | if (bb2 != bbf) |
1481 | bbs_to_fix_dom.quick_push (bb2); | |
9771b263 | 1482 | bbs_to_fix_dom.quick_push (bbf); |
6ab1ab14 SB |
1483 | |
1484 | iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true); | |
9771b263 | 1485 | bbs_to_fix_dom.release (); |
6ab1ab14 | 1486 | } |
b6e99746 MJ |
1487 | } |
1488 | ||
1489 | /* The following function is invoked on every switch statement (the current one | |
1490 | is given in SWTCH) and runs the individual phases of switch conversion on it | |
fade902a SB |
1491 | one after another until one fails or the conversion is completed. |
1492 | Returns NULL on success, or a pointer to a string with the reason why the | |
1493 | conversion failed. */ | |
b6e99746 | 1494 | |
fade902a | 1495 | static const char * |
538dd0b7 | 1496 | process_switch (gswitch *swtch) |
b6e99746 | 1497 | { |
fade902a | 1498 | struct switch_conv_info info; |
b6e99746 | 1499 | |
238065a7 SB |
1500 | /* Group case labels so that we get the right results from the heuristics |
1501 | that decide on the code generation approach for this switch. */ | |
1502 | group_case_labels_stmt (swtch); | |
1503 | ||
1504 | /* If this switch is now a degenerate case with only a default label, | |
1505 | there is nothing left for us to do. */ | |
1506 | if (gimple_switch_num_labels (swtch) < 2) | |
1507 | return "switch is a degenerate case"; | |
886cd84f SB |
1508 | |
1509 | collect_switch_conv_info (swtch, &info); | |
1510 | ||
1511 | /* No error markers should reach here (they should be filtered out | |
1512 | during gimplification). */ | |
1513 | gcc_checking_assert (TREE_TYPE (info.index_expr) != error_mark_node); | |
1514 | ||
531b10fc SB |
1515 | /* A switch on a constant should have been optimized in tree-cfg-cleanup. */ |
1516 | gcc_checking_assert (! TREE_CONSTANT (info.index_expr)); | |
886cd84f | 1517 | |
531b10fc | 1518 | if (info.uniq <= MAX_CASE_BIT_TESTS) |
886cd84f | 1519 | { |
531b10fc | 1520 | if (expand_switch_using_bit_tests_p (info.range_size, |
72798784 RB |
1521 | info.uniq, info.count, |
1522 | optimize_bb_for_speed_p | |
1523 | (gimple_bb (swtch)))) | |
531b10fc SB |
1524 | { |
1525 | if (dump_file) | |
1526 | fputs (" expanding as bit test is preferable\n", dump_file); | |
aa79a1e1 JJ |
1527 | emit_case_bit_tests (swtch, info.index_expr, info.range_min, |
1528 | info.range_size, info.range_max); | |
726338f4 | 1529 | loops_state_set (LOOPS_NEED_FIXUP); |
531b10fc SB |
1530 | return NULL; |
1531 | } | |
1532 | ||
1533 | if (info.uniq <= 2) | |
1534 | /* This will be expanded as a decision tree in stmt.c:expand_case. */ | |
1535 | return " expanding as jumps is preferable"; | |
886cd84f | 1536 | } |
b6e99746 | 1537 | |
531b10fc SB |
1538 | /* If there is no common successor, we cannot do the transformation. */ |
1539 | if (! info.final_bb) | |
1540 | return "no common successor to all case label target blocks found"; | |
1541 | ||
b6e99746 | 1542 | /* Check the case label values are within reasonable range: */ |
886cd84f | 1543 | if (!check_range (&info)) |
fade902a SB |
1544 | { |
1545 | gcc_assert (info.reason); | |
1546 | return info.reason; | |
1547 | } | |
b6e99746 MJ |
1548 | |
1549 | /* For all the cases, see whether they are empty, the assignments they | |
1550 | represent constant and so on... */ | |
886cd84f | 1551 | if (! check_all_empty_except_final (&info)) |
8e97bc2b | 1552 | { |
886cd84f SB |
1553 | gcc_assert (info.reason); |
1554 | return info.reason; | |
8e97bc2b | 1555 | } |
18bfe940 | 1556 | if (!check_final_bb (swtch, &info)) |
fade902a SB |
1557 | { |
1558 | gcc_assert (info.reason); | |
1559 | return info.reason; | |
1560 | } | |
b6e99746 MJ |
1561 | |
1562 | /* At this point all checks have passed and we can proceed with the | |
1563 | transformation. */ | |
1564 | ||
fade902a | 1565 | create_temp_arrays (&info); |
18bfe940 JJ |
1566 | gather_default_values (info.default_case_nonstandard |
1567 | ? gimple_switch_label (swtch, 1) | |
1568 | : gimple_switch_default_label (swtch), &info); | |
fade902a | 1569 | build_constructors (swtch, &info); |
b6e99746 | 1570 | |
fade902a SB |
1571 | build_arrays (swtch, &info); /* Build the static arrays and assignments. */ |
1572 | gen_inbound_check (swtch, &info); /* Build the bounds check. */ | |
b6e99746 MJ |
1573 | |
1574 | /* Cleanup: */ | |
fade902a SB |
1575 | free_temp_arrays (&info); |
1576 | return NULL; | |
b6e99746 MJ |
1577 | } |
1578 | ||
1579 | /* The main function of the pass scans statements for switches and invokes | |
1580 | process_switch on them. */ | |
1581 | ||
be55bfe6 TS |
1582 | namespace { |
1583 | ||
1584 | const pass_data pass_data_convert_switch = | |
1585 | { | |
1586 | GIMPLE_PASS, /* type */ | |
1587 | "switchconv", /* name */ | |
1588 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
1589 | TV_TREE_SWITCH_CONVERSION, /* tv_id */ |
1590 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1591 | 0, /* properties_provided */ | |
1592 | 0, /* properties_destroyed */ | |
1593 | 0, /* todo_flags_start */ | |
3bea341f | 1594 | TODO_update_ssa, /* todo_flags_finish */ |
be55bfe6 TS |
1595 | }; |
1596 | ||
1597 | class pass_convert_switch : public gimple_opt_pass | |
1598 | { | |
1599 | public: | |
1600 | pass_convert_switch (gcc::context *ctxt) | |
1601 | : gimple_opt_pass (pass_data_convert_switch, ctxt) | |
1602 | {} | |
1603 | ||
1604 | /* opt_pass methods: */ | |
1605 | virtual bool gate (function *) { return flag_tree_switch_conversion != 0; } | |
1606 | virtual unsigned int execute (function *); | |
1607 | ||
1608 | }; // class pass_convert_switch | |
1609 | ||
1610 | unsigned int | |
1611 | pass_convert_switch::execute (function *fun) | |
b6e99746 MJ |
1612 | { |
1613 | basic_block bb; | |
1614 | ||
be55bfe6 | 1615 | FOR_EACH_BB_FN (bb, fun) |
b6e99746 | 1616 | { |
fade902a | 1617 | const char *failure_reason; |
355fe088 | 1618 | gimple *stmt = last_stmt (bb); |
726a989a | 1619 | if (stmt && gimple_code (stmt) == GIMPLE_SWITCH) |
b6e99746 | 1620 | { |
b6e99746 MJ |
1621 | if (dump_file) |
1622 | { | |
726a989a RB |
1623 | expanded_location loc = expand_location (gimple_location (stmt)); |
1624 | ||
b6e99746 MJ |
1625 | fprintf (dump_file, "beginning to process the following " |
1626 | "SWITCH statement (%s:%d) : ------- \n", | |
1627 | loc.file, loc.line); | |
726a989a | 1628 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
edb30094 | 1629 | putc ('\n', dump_file); |
b6e99746 MJ |
1630 | } |
1631 | ||
538dd0b7 | 1632 | failure_reason = process_switch (as_a <gswitch *> (stmt)); |
fade902a | 1633 | if (! failure_reason) |
b6e99746 MJ |
1634 | { |
1635 | if (dump_file) | |
1636 | { | |
edb30094 UB |
1637 | fputs ("Switch converted\n", dump_file); |
1638 | fputs ("--------------------------------\n", dump_file); | |
b6e99746 | 1639 | } |
531b10fc SB |
1640 | |
1641 | /* Make no effort to update the post-dominator tree. It is actually not | |
1642 | that hard for the transformations we have performed, but it is not | |
1643 | supported by iterate_fix_dominators. */ | |
1644 | free_dominance_info (CDI_POST_DOMINATORS); | |
b6e99746 MJ |
1645 | } |
1646 | else | |
1647 | { | |
1648 | if (dump_file) | |
1649 | { | |
edb30094 | 1650 | fputs ("Bailing out - ", dump_file); |
fade902a SB |
1651 | fputs (failure_reason, dump_file); |
1652 | fputs ("\n--------------------------------\n", dump_file); | |
b6e99746 MJ |
1653 | } |
1654 | } | |
1655 | } | |
1656 | } | |
1657 | ||
1658 | return 0; | |
1659 | } | |
1660 | ||
27a4cd48 DM |
1661 | } // anon namespace |
1662 | ||
1663 | gimple_opt_pass * | |
1664 | make_pass_convert_switch (gcc::context *ctxt) | |
1665 | { | |
1666 | return new pass_convert_switch (ctxt); | |
1667 | } | |
9dc3d6a9 ML |
1668 | |
1669 | struct case_node | |
1670 | { | |
1671 | case_node *left; /* Left son in binary tree. */ | |
1672 | case_node *right; /* Right son in binary tree; | |
1673 | also node chain. */ | |
1674 | case_node *parent; /* Parent of node in binary tree. */ | |
1675 | tree low; /* Lowest index value for this label. */ | |
1676 | tree high; /* Highest index value for this label. */ | |
1677 | basic_block case_bb; /* Label to jump to when node matches. */ | |
1678 | tree case_label; /* Label to jump to when node matches. */ | |
1679 | profile_probability prob; /* Probability of taking this case. */ | |
1680 | profile_probability subtree_prob; /* Probability of reaching subtree | |
1681 | rooted at this node. */ | |
1682 | }; | |
1683 | ||
1684 | typedef case_node *case_node_ptr; | |
1685 | ||
1686 | static basic_block emit_case_nodes (basic_block, tree, case_node_ptr, | |
1687 | basic_block, tree, profile_probability, | |
1688 | tree, hash_map<tree, tree> *); | |
1689 | static bool node_has_low_bound (case_node_ptr, tree); | |
1690 | static bool node_has_high_bound (case_node_ptr, tree); | |
1691 | static bool node_is_bounded (case_node_ptr, tree); | |
1692 | ||
1693 | /* Return the smallest number of different values for which it is best to use a | |
1694 | jump-table instead of a tree of conditional branches. */ | |
1695 | ||
1696 | static unsigned int | |
1697 | case_values_threshold (void) | |
1698 | { | |
1699 | unsigned int threshold = PARAM_VALUE (PARAM_CASE_VALUES_THRESHOLD); | |
1700 | ||
1701 | if (threshold == 0) | |
1702 | threshold = targetm.case_values_threshold (); | |
1703 | ||
1704 | return threshold; | |
1705 | } | |
1706 | ||
1707 | /* Reset the aux field of all outgoing edges of basic block BB. */ | |
1708 | ||
1709 | static inline void | |
1710 | reset_out_edges_aux (basic_block bb) | |
1711 | { | |
1712 | edge e; | |
1713 | edge_iterator ei; | |
1714 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1715 | e->aux = (void *) 0; | |
1716 | } | |
1717 | ||
1718 | /* Compute the number of case labels that correspond to each outgoing edge of | |
1719 | STMT. Record this information in the aux field of the edge. */ | |
1720 | ||
1721 | static inline void | |
1722 | compute_cases_per_edge (gswitch *stmt) | |
1723 | { | |
1724 | basic_block bb = gimple_bb (stmt); | |
1725 | reset_out_edges_aux (bb); | |
1726 | int ncases = gimple_switch_num_labels (stmt); | |
1727 | for (int i = ncases - 1; i >= 1; --i) | |
1728 | { | |
1729 | tree elt = gimple_switch_label (stmt, i); | |
1730 | tree lab = CASE_LABEL (elt); | |
1731 | basic_block case_bb = label_to_block_fn (cfun, lab); | |
1732 | edge case_edge = find_edge (bb, case_bb); | |
1733 | case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + 1); | |
1734 | } | |
1735 | } | |
1736 | ||
1737 | /* Do the insertion of a case label into case_list. The labels are | |
1738 | fed to us in descending order from the sorted vector of case labels used | |
1739 | in the tree part of the middle end. So the list we construct is | |
1740 | sorted in ascending order. | |
1741 | ||
1742 | LABEL is the case label to be inserted. LOW and HIGH are the bounds | |
1743 | against which the index is compared to jump to LABEL and PROB is the | |
1744 | estimated probability LABEL is reached from the switch statement. */ | |
1745 | ||
1746 | static case_node * | |
1747 | add_case_node (case_node *head, tree low, tree high, basic_block case_bb, | |
1748 | tree case_label, profile_probability prob, | |
1749 | object_allocator<case_node> &case_node_pool) | |
1750 | { | |
1751 | case_node *r; | |
1752 | ||
1753 | gcc_checking_assert (low); | |
1754 | gcc_checking_assert (high && (TREE_TYPE (low) == TREE_TYPE (high))); | |
1755 | ||
1756 | /* Add this label to the chain. */ | |
1757 | r = case_node_pool.allocate (); | |
1758 | r->low = low; | |
1759 | r->high = high; | |
1760 | r->case_bb = case_bb; | |
1761 | r->case_label = case_label; | |
1762 | r->parent = r->left = NULL; | |
1763 | r->prob = prob; | |
1764 | r->subtree_prob = prob; | |
1765 | r->right = head; | |
1766 | return r; | |
1767 | } | |
1768 | ||
1769 | /* Dump ROOT, a list or tree of case nodes, to file. */ | |
1770 | ||
1771 | static void | |
1772 | dump_case_nodes (FILE *f, case_node *root, int indent_step, int indent_level) | |
1773 | { | |
1774 | if (root == 0) | |
1775 | return; | |
1776 | indent_level++; | |
1777 | ||
1778 | dump_case_nodes (f, root->left, indent_step, indent_level); | |
1779 | ||
1780 | fputs (";; ", f); | |
1781 | fprintf (f, "%*s", indent_step * indent_level, ""); | |
1782 | print_dec (root->low, f, TYPE_SIGN (TREE_TYPE (root->low))); | |
1783 | if (!tree_int_cst_equal (root->low, root->high)) | |
1784 | { | |
1785 | fprintf (f, " ... "); | |
1786 | print_dec (root->high, f, TYPE_SIGN (TREE_TYPE (root->high))); | |
1787 | } | |
1788 | fputs ("\n", f); | |
1789 | ||
1790 | dump_case_nodes (f, root->right, indent_step, indent_level); | |
1791 | } | |
1792 | ||
1793 | /* Take an ordered list of case nodes | |
1794 | and transform them into a near optimal binary tree, | |
1795 | on the assumption that any target code selection value is as | |
1796 | likely as any other. | |
1797 | ||
1798 | The transformation is performed by splitting the ordered | |
1799 | list into two equal sections plus a pivot. The parts are | |
1800 | then attached to the pivot as left and right branches. Each | |
1801 | branch is then transformed recursively. */ | |
1802 | ||
1803 | static void | |
1804 | balance_case_nodes (case_node_ptr *head, case_node_ptr parent) | |
1805 | { | |
1806 | case_node_ptr np; | |
1807 | ||
1808 | np = *head; | |
1809 | if (np) | |
1810 | { | |
1811 | int i = 0; | |
1812 | int ranges = 0; | |
1813 | case_node_ptr *npp; | |
1814 | case_node_ptr left; | |
1815 | ||
1816 | /* Count the number of entries on branch. Also count the ranges. */ | |
1817 | ||
1818 | while (np) | |
1819 | { | |
1820 | if (!tree_int_cst_equal (np->low, np->high)) | |
1821 | ranges++; | |
1822 | ||
1823 | i++; | |
1824 | np = np->right; | |
1825 | } | |
1826 | ||
1827 | if (i > 2) | |
1828 | { | |
1829 | /* Split this list if it is long enough for that to help. */ | |
1830 | npp = head; | |
1831 | left = *npp; | |
1832 | ||
1833 | /* If there are just three nodes, split at the middle one. */ | |
1834 | if (i == 3) | |
1835 | npp = &(*npp)->right; | |
1836 | else | |
1837 | { | |
1838 | /* Find the place in the list that bisects the list's total cost, | |
1839 | where ranges count as 2. | |
1840 | Here I gets half the total cost. */ | |
1841 | i = (i + ranges + 1) / 2; | |
1842 | while (1) | |
1843 | { | |
1844 | /* Skip nodes while their cost does not reach that amount. */ | |
1845 | if (!tree_int_cst_equal ((*npp)->low, (*npp)->high)) | |
1846 | i--; | |
1847 | i--; | |
1848 | if (i <= 0) | |
1849 | break; | |
1850 | npp = &(*npp)->right; | |
1851 | } | |
1852 | } | |
1853 | *head = np = *npp; | |
1854 | *npp = 0; | |
1855 | np->parent = parent; | |
1856 | np->left = left; | |
1857 | ||
1858 | /* Optimize each of the two split parts. */ | |
1859 | balance_case_nodes (&np->left, np); | |
1860 | balance_case_nodes (&np->right, np); | |
1861 | np->subtree_prob = np->prob; | |
1862 | np->subtree_prob += np->left->subtree_prob; | |
1863 | np->subtree_prob += np->right->subtree_prob; | |
1864 | } | |
1865 | else | |
1866 | { | |
1867 | /* Else leave this branch as one level, | |
1868 | but fill in `parent' fields. */ | |
1869 | np = *head; | |
1870 | np->parent = parent; | |
1871 | np->subtree_prob = np->prob; | |
1872 | for (; np->right; np = np->right) | |
1873 | { | |
1874 | np->right->parent = np; | |
1875 | (*head)->subtree_prob += np->right->subtree_prob; | |
1876 | } | |
1877 | } | |
1878 | } | |
1879 | } | |
1880 | ||
1881 | /* Return true if a switch should be expanded as a decision tree. | |
1882 | RANGE is the difference between highest and lowest case. | |
1883 | UNIQ is number of unique case node targets, not counting the default case. | |
1884 | COUNT is the number of comparisons needed, not counting the default case. */ | |
1885 | ||
1886 | static bool | |
1887 | expand_switch_as_decision_tree_p (tree range, | |
1888 | unsigned int uniq ATTRIBUTE_UNUSED, | |
1889 | unsigned int count) | |
1890 | { | |
1891 | int max_ratio; | |
1892 | ||
1893 | /* If neither casesi or tablejump is available, or flag_jump_tables | |
1894 | over-ruled us, we really have no choice. */ | |
1895 | if (!targetm.have_casesi () && !targetm.have_tablejump ()) | |
1896 | return true; | |
1897 | if (!flag_jump_tables) | |
1898 | return true; | |
1899 | #ifndef ASM_OUTPUT_ADDR_DIFF_ELT | |
1900 | if (flag_pic) | |
1901 | return true; | |
1902 | #endif | |
1903 | ||
1904 | /* If the switch is relatively small such that the cost of one | |
1905 | indirect jump on the target are higher than the cost of a | |
1906 | decision tree, go with the decision tree. | |
1907 | ||
1908 | If range of values is much bigger than number of values, | |
1909 | or if it is too large to represent in a HOST_WIDE_INT, | |
1910 | make a sequence of conditional branches instead of a dispatch. | |
1911 | ||
1912 | The definition of "much bigger" depends on whether we are | |
1913 | optimizing for size or for speed. If the former, the maximum | |
1914 | ratio range/count = 3, because this was found to be the optimal | |
1915 | ratio for size on i686-pc-linux-gnu, see PR11823. The ratio | |
1916 | 10 is much older, and was probably selected after an extensive | |
1917 | benchmarking investigation on numerous platforms. Or maybe it | |
1918 | just made sense to someone at some point in the history of GCC, | |
1919 | who knows... */ | |
1920 | max_ratio = optimize_insn_for_size_p () ? 3 : 10; | |
1921 | if (count < case_values_threshold () || !tree_fits_uhwi_p (range) | |
1922 | || compare_tree_int (range, max_ratio * count) > 0) | |
1923 | return true; | |
1924 | ||
1925 | return false; | |
1926 | } | |
1927 | ||
1928 | static void | |
1929 | fix_phi_operands_for_edge (edge e, hash_map<tree, tree> *phi_mapping) | |
1930 | { | |
1931 | basic_block bb = e->dest; | |
1932 | gphi_iterator gsi; | |
1933 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1934 | { | |
1935 | gphi *phi = gsi.phi (); | |
1936 | ||
1937 | tree *definition = phi_mapping->get (gimple_phi_result (phi)); | |
1938 | if (definition) | |
1939 | add_phi_arg (phi, *definition, e, UNKNOWN_LOCATION); | |
1940 | } | |
1941 | } | |
1942 | ||
1943 | ||
1944 | /* Add an unconditional jump to CASE_BB that happens in basic block BB. */ | |
1945 | ||
1946 | static void | |
1947 | emit_jump (basic_block bb, basic_block case_bb, | |
1948 | hash_map<tree, tree> *phi_mapping) | |
1949 | { | |
1950 | edge e = single_succ_edge (bb); | |
1951 | redirect_edge_succ (e, case_bb); | |
1952 | fix_phi_operands_for_edge (e, phi_mapping); | |
1953 | } | |
1954 | ||
1955 | /* Generate a decision tree, switching on INDEX_EXPR and jumping to | |
1956 | one of the labels in CASE_LIST or to the DEFAULT_LABEL. | |
1957 | DEFAULT_PROB is the estimated probability that it jumps to | |
1958 | DEFAULT_LABEL. | |
1959 | ||
1960 | We generate a binary decision tree to select the appropriate target | |
1961 | code. */ | |
1962 | ||
1963 | static void | |
1964 | emit_case_decision_tree (gswitch *s, tree index_expr, tree index_type, | |
1965 | case_node_ptr case_list, basic_block default_bb, | |
1966 | tree default_label, profile_probability default_prob, | |
1967 | hash_map<tree, tree> *phi_mapping) | |
1968 | { | |
1969 | balance_case_nodes (&case_list, NULL); | |
1970 | ||
1971 | if (dump_file) | |
1972 | dump_function_to_file (current_function_decl, dump_file, dump_flags); | |
1973 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1974 | { | |
1975 | int indent_step = ceil_log2 (TYPE_PRECISION (index_type)) + 2; | |
1976 | fprintf (dump_file, ";; Expanding GIMPLE switch as decision tree:\n"); | |
1977 | dump_case_nodes (dump_file, case_list, indent_step, 0); | |
1978 | } | |
1979 | ||
1980 | basic_block bb = gimple_bb (s); | |
1981 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
1982 | edge e; | |
1983 | if (gsi_end_p (gsi)) | |
1984 | e = split_block_after_labels (bb); | |
1985 | else | |
1986 | { | |
1987 | gsi_prev (&gsi); | |
1988 | e = split_block (bb, gsi_stmt (gsi)); | |
1989 | } | |
1990 | bb = split_edge (e); | |
1991 | ||
1992 | bb = emit_case_nodes (bb, index_expr, case_list, default_bb, default_label, | |
1993 | default_prob, index_type, phi_mapping); | |
1994 | ||
1995 | if (bb) | |
1996 | emit_jump (bb, default_bb, phi_mapping); | |
1997 | ||
1998 | /* Remove all edges and do just an edge that will reach default_bb. */ | |
1999 | gsi = gsi_last_bb (gimple_bb (s)); | |
2000 | gsi_remove (&gsi, true); | |
2001 | } | |
2002 | ||
2003 | static void | |
2004 | record_phi_operand_mapping (const vec<basic_block> bbs, basic_block switch_bb, | |
2005 | hash_map <tree, tree> *map) | |
2006 | { | |
2007 | /* Record all PHI nodes that have to be fixed after conversion. */ | |
2008 | for (unsigned i = 0; i < bbs.length (); i++) | |
2009 | { | |
2010 | basic_block bb = bbs[i]; | |
2011 | ||
2012 | gphi_iterator gsi; | |
2013 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2014 | { | |
2015 | gphi *phi = gsi.phi (); | |
2016 | ||
2017 | for (unsigned i = 0; i < gimple_phi_num_args (phi); i++) | |
2018 | { | |
2019 | basic_block phi_src_bb = gimple_phi_arg_edge (phi, i)->src; | |
2020 | if (phi_src_bb == switch_bb) | |
2021 | { | |
2022 | tree def = gimple_phi_arg_def (phi, i); | |
2023 | tree result = gimple_phi_result (phi); | |
2024 | map->put (result, def); | |
2025 | break; | |
2026 | } | |
2027 | } | |
2028 | } | |
2029 | } | |
2030 | } | |
2031 | ||
2032 | /* Attempt to expand gimple switch STMT to a decision tree. */ | |
2033 | ||
2034 | static bool | |
2035 | try_switch_expansion (gswitch *stmt) | |
2036 | { | |
2037 | tree minval = NULL_TREE, maxval = NULL_TREE, range = NULL_TREE; | |
2038 | basic_block default_bb; | |
2039 | unsigned int count, uniq; | |
2040 | int i; | |
2041 | int ncases = gimple_switch_num_labels (stmt); | |
2042 | tree index_expr = gimple_switch_index (stmt); | |
2043 | tree index_type = TREE_TYPE (index_expr); | |
2044 | tree elt; | |
2045 | basic_block bb = gimple_bb (stmt); | |
2046 | ||
2047 | hash_map<tree, tree> phi_mapping; | |
2048 | auto_vec<basic_block> case_bbs; | |
2049 | ||
2050 | /* A list of case labels; it is first built as a list and it may then | |
2051 | be rearranged into a nearly balanced binary tree. */ | |
2052 | case_node *case_list = 0; | |
2053 | ||
2054 | /* A pool for case nodes. */ | |
2055 | object_allocator<case_node> case_node_pool ("struct case_node pool"); | |
2056 | ||
2057 | /* cleanup_tree_cfg removes all SWITCH_EXPR with their index | |
2058 | expressions being INTEGER_CST. */ | |
2059 | gcc_assert (TREE_CODE (index_expr) != INTEGER_CST); | |
2060 | ||
2061 | /* Optimization of switch statements with only one label has already | |
2062 | occurred, so we should never see them at this point. */ | |
2063 | gcc_assert (ncases > 1); | |
2064 | ||
2065 | /* Find the default case target label. */ | |
2066 | tree default_label = CASE_LABEL (gimple_switch_default_label (stmt)); | |
2067 | default_bb = label_to_block_fn (cfun, default_label); | |
2068 | edge default_edge = find_edge (bb, default_bb); | |
2069 | profile_probability default_prob = default_edge->probability; | |
2070 | case_bbs.safe_push (default_bb); | |
2071 | ||
2072 | /* Get upper and lower bounds of case values. */ | |
2073 | elt = gimple_switch_label (stmt, 1); | |
2074 | minval = fold_convert (index_type, CASE_LOW (elt)); | |
2075 | elt = gimple_switch_label (stmt, ncases - 1); | |
2076 | if (CASE_HIGH (elt)) | |
2077 | maxval = fold_convert (index_type, CASE_HIGH (elt)); | |
2078 | else | |
2079 | maxval = fold_convert (index_type, CASE_LOW (elt)); | |
2080 | ||
2081 | /* Compute span of values. */ | |
2082 | range = fold_build2 (MINUS_EXPR, index_type, maxval, minval); | |
2083 | ||
2084 | /* Listify the labels queue and gather some numbers to decide | |
2085 | how to expand this switch. */ | |
2086 | uniq = 0; | |
2087 | count = 0; | |
2088 | hash_set<tree> seen_labels; | |
2089 | compute_cases_per_edge (stmt); | |
2090 | ||
2091 | for (i = ncases - 1; i >= 1; --i) | |
2092 | { | |
2093 | elt = gimple_switch_label (stmt, i); | |
2094 | tree low = CASE_LOW (elt); | |
2095 | gcc_assert (low); | |
2096 | tree high = CASE_HIGH (elt); | |
2097 | gcc_assert (!high || tree_int_cst_lt (low, high)); | |
2098 | tree lab = CASE_LABEL (elt); | |
2099 | ||
2100 | /* Count the elements. | |
2101 | A range counts double, since it requires two compares. */ | |
2102 | count++; | |
2103 | if (high) | |
2104 | count++; | |
2105 | ||
2106 | /* If we have not seen this label yet, then increase the | |
2107 | number of unique case node targets seen. */ | |
2108 | if (!seen_labels.add (lab)) | |
2109 | uniq++; | |
2110 | ||
2111 | /* The bounds on the case range, LOW and HIGH, have to be converted | |
2112 | to case's index type TYPE. Note that the original type of the | |
2113 | case index in the source code is usually "lost" during | |
2114 | gimplification due to type promotion, but the case labels retain the | |
2115 | original type. Make sure to drop overflow flags. */ | |
2116 | low = fold_convert (index_type, low); | |
2117 | if (TREE_OVERFLOW (low)) | |
2118 | low = wide_int_to_tree (index_type, low); | |
2119 | ||
2120 | /* The canonical from of a case label in GIMPLE is that a simple case | |
2121 | has an empty CASE_HIGH. For the casesi and tablejump expanders, | |
2122 | the back ends want simple cases to have high == low. */ | |
2123 | if (!high) | |
2124 | high = low; | |
2125 | high = fold_convert (index_type, high); | |
2126 | if (TREE_OVERFLOW (high)) | |
2127 | high = wide_int_to_tree (index_type, high); | |
2128 | ||
2129 | basic_block case_bb = label_to_block_fn (cfun, lab); | |
2130 | edge case_edge = find_edge (bb, case_bb); | |
2131 | case_list = add_case_node ( | |
2132 | case_list, low, high, case_bb, lab, | |
2133 | case_edge->probability.apply_scale (1, (intptr_t) (case_edge->aux)), | |
2134 | case_node_pool); | |
2135 | ||
2136 | case_bbs.safe_push (case_bb); | |
2137 | } | |
2138 | reset_out_edges_aux (bb); | |
2139 | record_phi_operand_mapping (case_bbs, bb, &phi_mapping); | |
2140 | ||
2141 | /* cleanup_tree_cfg removes all SWITCH_EXPR with a single | |
2142 | destination, such as one with a default case only. | |
2143 | It also removes cases that are out of range for the switch | |
2144 | type, so we should never get a zero here. */ | |
2145 | gcc_assert (count > 0); | |
2146 | ||
2147 | /* Decide how to expand this switch. | |
2148 | The two options at this point are a dispatch table (casesi or | |
2149 | tablejump) or a decision tree. */ | |
2150 | ||
2151 | if (expand_switch_as_decision_tree_p (range, uniq, count)) | |
2152 | { | |
2153 | emit_case_decision_tree (stmt, index_expr, index_type, case_list, | |
2154 | default_bb, default_label, default_prob, | |
2155 | &phi_mapping); | |
2156 | return true; | |
2157 | } | |
2158 | ||
2159 | return false; | |
2160 | } | |
2161 | ||
2162 | /* The main function of the pass scans statements for switches and invokes | |
2163 | process_switch on them. */ | |
2164 | ||
2165 | namespace { | |
2166 | ||
2167 | const pass_data pass_data_lower_switch = | |
2168 | { | |
2169 | GIMPLE_PASS, /* type */ | |
2170 | "switchlower", /* name */ | |
2171 | OPTGROUP_NONE, /* optinfo_flags */ | |
2172 | TV_TREE_SWITCH_LOWERING, /* tv_id */ | |
2173 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2174 | 0, /* properties_provided */ | |
2175 | 0, /* properties_destroyed */ | |
2176 | 0, /* todo_flags_start */ | |
2177 | TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ | |
2178 | }; | |
2179 | ||
2180 | class pass_lower_switch : public gimple_opt_pass | |
2181 | { | |
2182 | public: | |
2183 | pass_lower_switch (gcc::context *ctxt) | |
2184 | : gimple_opt_pass (pass_data_lower_switch, ctxt) | |
2185 | {} | |
2186 | ||
2187 | /* opt_pass methods: */ | |
2188 | virtual bool gate (function *) { return true; } | |
2189 | virtual unsigned int execute (function *); | |
2190 | ||
2191 | }; // class pass_lower_switch | |
2192 | ||
2193 | unsigned int | |
2194 | pass_lower_switch::execute (function *fun) | |
2195 | { | |
2196 | basic_block bb; | |
2197 | bool expanded = false; | |
2198 | ||
2199 | FOR_EACH_BB_FN (bb, fun) | |
2200 | { | |
2201 | gimple *stmt = last_stmt (bb); | |
2202 | if (stmt && gimple_code (stmt) == GIMPLE_SWITCH) | |
2203 | { | |
2204 | if (dump_file) | |
2205 | { | |
2206 | expanded_location loc = expand_location (gimple_location (stmt)); | |
2207 | ||
2208 | fprintf (dump_file, "beginning to process the following " | |
2209 | "SWITCH statement (%s:%d) : ------- \n", | |
2210 | loc.file, loc.line); | |
2211 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
2212 | putc ('\n', dump_file); | |
2213 | } | |
2214 | ||
2215 | expanded |= try_switch_expansion (as_a<gswitch *> (stmt)); | |
2216 | } | |
2217 | } | |
2218 | ||
2219 | if (expanded) | |
2220 | { | |
2221 | free_dominance_info (CDI_DOMINATORS); | |
2222 | free_dominance_info (CDI_POST_DOMINATORS); | |
2223 | mark_virtual_operands_for_renaming (cfun); | |
2224 | } | |
2225 | ||
2226 | return 0; | |
2227 | } | |
2228 | ||
2229 | } // anon namespace | |
2230 | ||
2231 | gimple_opt_pass * | |
2232 | make_pass_lower_switch (gcc::context *ctxt) | |
2233 | { | |
2234 | return new pass_lower_switch (ctxt); | |
2235 | } | |
2236 | ||
2237 | /* Generate code to jump to LABEL if OP0 and OP1 are equal in mode MODE. | |
2238 | PROB is the probability of jumping to LABEL. */ | |
2239 | static basic_block | |
2240 | do_jump_if_equal (basic_block bb, tree op0, tree op1, basic_block label_bb, | |
2241 | profile_probability prob, hash_map<tree, tree> *phi_mapping) | |
2242 | { | |
2243 | gcond *cond = gimple_build_cond (EQ_EXPR, op0, op1, NULL_TREE, NULL_TREE); | |
2244 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2245 | gsi_insert_before (&gsi, cond, GSI_SAME_STMT); | |
2246 | ||
2247 | gcc_assert (single_succ_p (bb)); | |
2248 | ||
2249 | /* Make a new basic block where false branch will take place. */ | |
2250 | edge false_edge = split_block (bb, cond); | |
2251 | false_edge->flags = EDGE_FALSE_VALUE; | |
2252 | false_edge->probability = prob.invert (); | |
2253 | ||
2254 | edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
2255 | fix_phi_operands_for_edge (true_edge, phi_mapping); | |
2256 | true_edge->probability = prob; | |
2257 | ||
2258 | return false_edge->dest; | |
2259 | } | |
2260 | ||
2261 | /* Generate code to compare X with Y so that the condition codes are | |
2262 | set and to jump to LABEL if the condition is true. If X is a | |
2263 | constant and Y is not a constant, then the comparison is swapped to | |
2264 | ensure that the comparison RTL has the canonical form. | |
2265 | ||
2266 | UNSIGNEDP nonzero says that X and Y are unsigned; this matters if they | |
2267 | need to be widened. UNSIGNEDP is also used to select the proper | |
2268 | branch condition code. | |
2269 | ||
2270 | If X and Y have mode BLKmode, then SIZE specifies the size of both X and Y. | |
2271 | ||
2272 | MODE is the mode of the inputs (in case they are const_int). | |
2273 | ||
2274 | COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.). | |
2275 | It will be potentially converted into an unsigned variant based on | |
2276 | UNSIGNEDP to select a proper jump instruction. | |
2277 | ||
2278 | PROB is the probability of jumping to LABEL. */ | |
2279 | ||
2280 | static basic_block | |
2281 | emit_cmp_and_jump_insns (basic_block bb, tree op0, tree op1, | |
2282 | tree_code comparison, basic_block label_bb, | |
2283 | profile_probability prob, | |
2284 | hash_map<tree, tree> *phi_mapping) | |
2285 | { | |
2286 | gcond *cond = gimple_build_cond (comparison, op0, op1, NULL_TREE, NULL_TREE); | |
2287 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2288 | gsi_insert_after (&gsi, cond, GSI_NEW_STMT); | |
2289 | ||
2290 | gcc_assert (single_succ_p (bb)); | |
2291 | ||
2292 | /* Make a new basic block where false branch will take place. */ | |
2293 | edge false_edge = split_block (bb, cond); | |
2294 | false_edge->flags = EDGE_FALSE_VALUE; | |
2295 | false_edge->probability = prob.invert (); | |
2296 | ||
2297 | edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
2298 | fix_phi_operands_for_edge (true_edge, phi_mapping); | |
2299 | true_edge->probability = prob; | |
2300 | ||
2301 | return false_edge->dest; | |
2302 | } | |
2303 | ||
2304 | /* Computes the conditional probability of jumping to a target if the branch | |
2305 | instruction is executed. | |
2306 | TARGET_PROB is the estimated probability of jumping to a target relative | |
2307 | to some basic block BB. | |
2308 | BASE_PROB is the probability of reaching the branch instruction relative | |
2309 | to the same basic block BB. */ | |
2310 | ||
2311 | static inline profile_probability | |
2312 | conditional_probability (profile_probability target_prob, | |
2313 | profile_probability base_prob) | |
2314 | { | |
2315 | return target_prob / base_prob; | |
2316 | } | |
2317 | ||
2318 | /* Emit step-by-step code to select a case for the value of INDEX. | |
2319 | The thus generated decision tree follows the form of the | |
2320 | case-node binary tree NODE, whose nodes represent test conditions. | |
2321 | INDEX_TYPE is the type of the index of the switch. | |
2322 | ||
2323 | Care is taken to prune redundant tests from the decision tree | |
2324 | by detecting any boundary conditions already checked by | |
2325 | emitted rtx. (See node_has_high_bound, node_has_low_bound | |
2326 | and node_is_bounded, above.) | |
2327 | ||
2328 | Where the test conditions can be shown to be redundant we emit | |
2329 | an unconditional jump to the target code. As a further | |
2330 | optimization, the subordinates of a tree node are examined to | |
2331 | check for bounded nodes. In this case conditional and/or | |
2332 | unconditional jumps as a result of the boundary check for the | |
2333 | current node are arranged to target the subordinates associated | |
2334 | code for out of bound conditions on the current node. | |
2335 | ||
2336 | We can assume that when control reaches the code generated here, | |
2337 | the index value has already been compared with the parents | |
2338 | of this node, and determined to be on the same side of each parent | |
2339 | as this node is. Thus, if this node tests for the value 51, | |
2340 | and a parent tested for 52, we don't need to consider | |
2341 | the possibility of a value greater than 51. If another parent | |
2342 | tests for the value 50, then this node need not test anything. */ | |
2343 | ||
2344 | static basic_block | |
2345 | emit_case_nodes (basic_block bb, tree index, case_node_ptr node, | |
2346 | basic_block default_bb, tree default_label, | |
2347 | profile_probability default_prob, tree index_type, | |
2348 | hash_map<tree, tree> *phi_mapping) | |
2349 | { | |
2350 | /* If INDEX has an unsigned type, we must make unsigned branches. */ | |
2351 | profile_probability probability; | |
2352 | profile_probability prob = node->prob, subtree_prob = node->subtree_prob; | |
2353 | ||
2354 | /* See if our parents have already tested everything for us. | |
2355 | If they have, emit an unconditional jump for this node. */ | |
2356 | if (node_is_bounded (node, index_type)) | |
2357 | { | |
2358 | emit_jump (bb, node->case_bb, phi_mapping); | |
2359 | return NULL; | |
2360 | } | |
2361 | ||
2362 | else if (tree_int_cst_equal (node->low, node->high)) | |
2363 | { | |
2364 | probability = conditional_probability (prob, subtree_prob + default_prob); | |
2365 | /* Node is single valued. First see if the index expression matches | |
2366 | this node and then check our children, if any. */ | |
2367 | bb = do_jump_if_equal (bb, index, node->low, node->case_bb, probability, | |
2368 | phi_mapping); | |
2369 | /* Since this case is taken at this point, reduce its weight from | |
2370 | subtree_weight. */ | |
2371 | subtree_prob -= prob; | |
2372 | if (node->right != 0 && node->left != 0) | |
2373 | { | |
2374 | /* This node has children on both sides. | |
2375 | Dispatch to one side or the other | |
2376 | by comparing the index value with this node's value. | |
2377 | If one subtree is bounded, check that one first, | |
2378 | so we can avoid real branches in the tree. */ | |
2379 | ||
2380 | if (node_is_bounded (node->right, index_type)) | |
2381 | { | |
2382 | probability | |
2383 | = conditional_probability (node->right->prob, | |
2384 | subtree_prob + default_prob); | |
2385 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2386 | node->right->case_bb, probability, | |
2387 | phi_mapping); | |
2388 | bb = emit_case_nodes (bb, index, node->left, default_bb, | |
2389 | default_label, default_prob, index_type, | |
2390 | phi_mapping); | |
2391 | } | |
2392 | ||
2393 | else if (node_is_bounded (node->left, index_type)) | |
2394 | { | |
2395 | probability | |
2396 | = conditional_probability (node->left->prob, | |
2397 | subtree_prob + default_prob); | |
2398 | bb = emit_cmp_and_jump_insns (bb, index, node->high, LT_EXPR, | |
2399 | node->left->case_bb, probability, | |
2400 | phi_mapping); | |
2401 | bb = emit_case_nodes (bb, index, node->right, default_bb, | |
2402 | default_label, default_prob, index_type, | |
2403 | phi_mapping); | |
2404 | } | |
2405 | ||
2406 | /* If both children are single-valued cases with no | |
2407 | children, finish up all the work. This way, we can save | |
2408 | one ordered comparison. */ | |
2409 | else if (tree_int_cst_equal (node->right->low, node->right->high) | |
2410 | && node->right->left == 0 && node->right->right == 0 | |
2411 | && tree_int_cst_equal (node->left->low, node->left->high) | |
2412 | && node->left->left == 0 && node->left->right == 0) | |
2413 | { | |
2414 | /* Neither node is bounded. First distinguish the two sides; | |
2415 | then emit the code for one side at a time. */ | |
2416 | ||
2417 | /* See if the value matches what the right hand side | |
2418 | wants. */ | |
2419 | probability | |
2420 | = conditional_probability (node->right->prob, | |
2421 | subtree_prob + default_prob); | |
2422 | bb = do_jump_if_equal (bb, index, node->right->low, | |
2423 | node->right->case_bb, probability, | |
2424 | phi_mapping); | |
2425 | ||
2426 | /* See if the value matches what the left hand side | |
2427 | wants. */ | |
2428 | probability | |
2429 | = conditional_probability (node->left->prob, | |
2430 | subtree_prob + default_prob); | |
2431 | bb = do_jump_if_equal (bb, index, node->left->low, | |
2432 | node->left->case_bb, probability, | |
2433 | phi_mapping); | |
2434 | } | |
2435 | ||
2436 | else | |
2437 | { | |
2438 | /* Neither node is bounded. First distinguish the two sides; | |
2439 | then emit the code for one side at a time. */ | |
2440 | ||
2441 | basic_block test_bb = split_edge (single_succ_edge (bb)); | |
2442 | redirect_edge_succ (single_pred_edge (test_bb), | |
2443 | single_succ_edge (bb)->dest); | |
2444 | ||
2445 | /* The default label could be reached either through the right | |
2446 | subtree or the left subtree. Divide the probability | |
2447 | equally. */ | |
2448 | probability | |
2449 | = conditional_probability (node->right->subtree_prob | |
2450 | + default_prob.apply_scale (1, 2), | |
2451 | subtree_prob + default_prob); | |
2452 | /* See if the value is on the right. */ | |
2453 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2454 | test_bb, probability, phi_mapping); | |
2455 | default_prob = default_prob.apply_scale (1, 2); | |
2456 | ||
2457 | /* Value must be on the left. | |
2458 | Handle the left-hand subtree. */ | |
2459 | bb = emit_case_nodes (bb, index, node->left, default_bb, | |
2460 | default_label, default_prob, index_type, | |
2461 | phi_mapping); | |
2462 | /* If left-hand subtree does nothing, | |
2463 | go to default. */ | |
2464 | ||
2465 | if (bb && default_bb) | |
2466 | emit_jump (bb, default_bb, phi_mapping); | |
2467 | ||
2468 | /* Code branches here for the right-hand subtree. */ | |
2469 | bb = emit_case_nodes (test_bb, index, node->right, default_bb, | |
2470 | default_label, default_prob, index_type, | |
2471 | phi_mapping); | |
2472 | } | |
2473 | } | |
2474 | else if (node->right != 0 && node->left == 0) | |
2475 | { | |
2476 | /* Here we have a right child but no left so we issue a conditional | |
2477 | branch to default and process the right child. | |
2478 | ||
2479 | Omit the conditional branch to default if the right child | |
2480 | does not have any children and is single valued; it would | |
2481 | cost too much space to save so little time. */ | |
2482 | ||
2483 | if (node->right->right || node->right->left | |
2484 | || !tree_int_cst_equal (node->right->low, node->right->high)) | |
2485 | { | |
2486 | if (!node_has_low_bound (node, index_type)) | |
2487 | { | |
2488 | probability | |
2489 | = conditional_probability (default_prob.apply_scale (1, 2), | |
2490 | subtree_prob + default_prob); | |
2491 | bb = emit_cmp_and_jump_insns (bb, index, node->high, LT_EXPR, | |
2492 | default_bb, probability, | |
2493 | phi_mapping); | |
2494 | default_prob = default_prob.apply_scale (1, 2); | |
2495 | } | |
2496 | ||
2497 | bb = emit_case_nodes (bb, index, node->right, default_bb, | |
2498 | default_label, default_prob, index_type, | |
2499 | phi_mapping); | |
2500 | } | |
2501 | else | |
2502 | { | |
2503 | probability | |
2504 | = conditional_probability (node->right->subtree_prob, | |
2505 | subtree_prob + default_prob); | |
2506 | /* We cannot process node->right normally | |
2507 | since we haven't ruled out the numbers less than | |
2508 | this node's value. So handle node->right explicitly. */ | |
2509 | bb = do_jump_if_equal (bb, index, node->right->low, | |
2510 | node->right->case_bb, probability, | |
2511 | phi_mapping); | |
2512 | } | |
2513 | } | |
2514 | ||
2515 | else if (node->right == 0 && node->left != 0) | |
2516 | { | |
2517 | /* Just one subtree, on the left. */ | |
2518 | if (node->left->left || node->left->right | |
2519 | || !tree_int_cst_equal (node->left->low, node->left->high)) | |
2520 | { | |
2521 | if (!node_has_high_bound (node, index_type)) | |
2522 | { | |
2523 | probability | |
2524 | = conditional_probability (default_prob.apply_scale (1, 2), | |
2525 | subtree_prob + default_prob); | |
2526 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2527 | default_bb, probability, | |
2528 | phi_mapping); | |
2529 | default_prob = default_prob.apply_scale (1, 2); | |
2530 | } | |
2531 | ||
2532 | bb = emit_case_nodes (bb, index, node->left, default_bb, | |
2533 | default_label, default_prob, index_type, | |
2534 | phi_mapping); | |
2535 | } | |
2536 | else | |
2537 | { | |
2538 | probability | |
2539 | = conditional_probability (node->left->subtree_prob, | |
2540 | subtree_prob + default_prob); | |
2541 | /* We cannot process node->left normally | |
2542 | since we haven't ruled out the numbers less than | |
2543 | this node's value. So handle node->left explicitly. */ | |
2544 | do_jump_if_equal (bb, index, node->left->low, node->left->case_bb, | |
2545 | probability, phi_mapping); | |
2546 | } | |
2547 | } | |
2548 | } | |
2549 | else | |
2550 | { | |
2551 | /* Node is a range. These cases are very similar to those for a single | |
2552 | value, except that we do not start by testing whether this node | |
2553 | is the one to branch to. */ | |
2554 | ||
2555 | if (node->right != 0 && node->left != 0) | |
2556 | { | |
2557 | /* Node has subtrees on both sides. | |
2558 | If the right-hand subtree is bounded, | |
2559 | test for it first, since we can go straight there. | |
2560 | Otherwise, we need to make a branch in the control structure, | |
2561 | then handle the two subtrees. */ | |
2562 | basic_block test_bb = NULL; | |
2563 | ||
2564 | if (node_is_bounded (node->right, index_type)) | |
2565 | { | |
2566 | /* Right hand node is fully bounded so we can eliminate any | |
2567 | testing and branch directly to the target code. */ | |
2568 | probability | |
2569 | = conditional_probability (node->right->subtree_prob, | |
2570 | subtree_prob + default_prob); | |
2571 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2572 | node->right->case_bb, probability, | |
2573 | phi_mapping); | |
2574 | } | |
2575 | else | |
2576 | { | |
2577 | /* Right hand node requires testing. | |
2578 | Branch to a label where we will handle it later. */ | |
2579 | ||
2580 | test_bb = split_edge (single_succ_edge (bb)); | |
2581 | redirect_edge_succ (single_pred_edge (test_bb), | |
2582 | single_succ_edge (bb)->dest); | |
2583 | ||
2584 | probability | |
2585 | = conditional_probability (node->right->subtree_prob | |
2586 | + default_prob.apply_scale (1, 2), | |
2587 | subtree_prob + default_prob); | |
2588 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2589 | test_bb, probability, phi_mapping); | |
2590 | default_prob = default_prob.apply_scale (1, 2); | |
2591 | } | |
2592 | ||
2593 | /* Value belongs to this node or to the left-hand subtree. */ | |
2594 | ||
2595 | probability | |
2596 | = conditional_probability (prob, subtree_prob + default_prob); | |
2597 | bb = emit_cmp_and_jump_insns (bb, index, node->low, GE_EXPR, | |
2598 | node->case_bb, probability, | |
2599 | phi_mapping); | |
2600 | ||
2601 | /* Handle the left-hand subtree. */ | |
2602 | bb = emit_case_nodes (bb, index, node->left, default_bb, | |
2603 | default_label, default_prob, index_type, | |
2604 | phi_mapping); | |
2605 | ||
2606 | /* If right node had to be handled later, do that now. */ | |
2607 | if (test_bb) | |
2608 | { | |
2609 | /* If the left-hand subtree fell through, | |
2610 | don't let it fall into the right-hand subtree. */ | |
2611 | if (bb && default_bb) | |
2612 | emit_jump (bb, default_bb, phi_mapping); | |
2613 | ||
2614 | bb = emit_case_nodes (test_bb, index, node->right, default_bb, | |
2615 | default_label, default_prob, index_type, | |
2616 | phi_mapping); | |
2617 | } | |
2618 | } | |
2619 | ||
2620 | else if (node->right != 0 && node->left == 0) | |
2621 | { | |
2622 | /* Deal with values to the left of this node, | |
2623 | if they are possible. */ | |
2624 | if (!node_has_low_bound (node, index_type)) | |
2625 | { | |
2626 | probability | |
2627 | = conditional_probability (default_prob.apply_scale (1, 2), | |
2628 | subtree_prob + default_prob); | |
2629 | bb = emit_cmp_and_jump_insns (bb, index, node->low, LT_EXPR, | |
2630 | default_bb, probability, | |
2631 | phi_mapping); | |
2632 | default_prob = default_prob.apply_scale (1, 2); | |
2633 | } | |
2634 | ||
2635 | /* Value belongs to this node or to the right-hand subtree. */ | |
2636 | ||
2637 | probability | |
2638 | = conditional_probability (prob, subtree_prob + default_prob); | |
2639 | bb = emit_cmp_and_jump_insns (bb, index, node->high, LE_EXPR, | |
2640 | node->case_bb, probability, | |
2641 | phi_mapping); | |
2642 | ||
2643 | bb = emit_case_nodes (bb, index, node->right, default_bb, | |
2644 | default_label, default_prob, index_type, | |
2645 | phi_mapping); | |
2646 | } | |
2647 | ||
2648 | else if (node->right == 0 && node->left != 0) | |
2649 | { | |
2650 | /* Deal with values to the right of this node, | |
2651 | if they are possible. */ | |
2652 | if (!node_has_high_bound (node, index_type)) | |
2653 | { | |
2654 | probability | |
2655 | = conditional_probability (default_prob.apply_scale (1, 2), | |
2656 | subtree_prob + default_prob); | |
2657 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2658 | default_bb, probability, | |
2659 | phi_mapping); | |
2660 | default_prob = default_prob.apply_scale (1, 2); | |
2661 | } | |
2662 | ||
2663 | /* Value belongs to this node or to the left-hand subtree. */ | |
2664 | ||
2665 | probability | |
2666 | = conditional_probability (prob, subtree_prob + default_prob); | |
2667 | bb = emit_cmp_and_jump_insns (bb, index, node->low, GE_EXPR, | |
2668 | node->case_bb, probability, | |
2669 | phi_mapping); | |
2670 | ||
2671 | bb = emit_case_nodes (bb, index, node->left, default_bb, | |
2672 | default_label, default_prob, index_type, | |
2673 | phi_mapping); | |
2674 | } | |
2675 | ||
2676 | else | |
2677 | { | |
2678 | /* Node has no children so we check low and high bounds to remove | |
2679 | redundant tests. Only one of the bounds can exist, | |
2680 | since otherwise this node is bounded--a case tested already. */ | |
2681 | bool high_bound = node_has_high_bound (node, index_type); | |
2682 | bool low_bound = node_has_low_bound (node, index_type); | |
2683 | ||
2684 | if (!high_bound && low_bound) | |
2685 | { | |
2686 | probability | |
2687 | = conditional_probability (default_prob, | |
2688 | subtree_prob + default_prob); | |
2689 | bb = emit_cmp_and_jump_insns (bb, index, node->high, GT_EXPR, | |
2690 | default_bb, probability, | |
2691 | phi_mapping); | |
2692 | } | |
2693 | ||
2694 | else if (!low_bound && high_bound) | |
2695 | { | |
2696 | probability | |
2697 | = conditional_probability (default_prob, | |
2698 | subtree_prob + default_prob); | |
2699 | bb = emit_cmp_and_jump_insns (bb, index, node->low, LT_EXPR, | |
2700 | default_bb, probability, | |
2701 | phi_mapping); | |
2702 | } | |
2703 | else if (!low_bound && !high_bound) | |
2704 | { | |
2705 | tree type = TREE_TYPE (index); | |
2706 | tree utype = unsigned_type_for (type); | |
2707 | ||
2708 | tree lhs = make_ssa_name (type); | |
2709 | gassign *sub1 | |
2710 | = gimple_build_assign (lhs, MINUS_EXPR, index, node->low); | |
2711 | ||
2712 | tree converted = make_ssa_name (utype); | |
2713 | gassign *a = gimple_build_assign (converted, NOP_EXPR, lhs); | |
2714 | ||
2715 | tree rhs = fold_build2 (MINUS_EXPR, utype, | |
2716 | fold_convert (type, node->high), | |
2717 | fold_convert (type, node->low)); | |
2718 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2719 | gsi_insert_before (&gsi, sub1, GSI_SAME_STMT); | |
2720 | gsi_insert_before (&gsi, a, GSI_SAME_STMT); | |
2721 | ||
2722 | probability | |
2723 | = conditional_probability (default_prob, | |
2724 | subtree_prob + default_prob); | |
2725 | bb = emit_cmp_and_jump_insns (bb, converted, rhs, GT_EXPR, | |
2726 | default_bb, probability, | |
2727 | phi_mapping); | |
2728 | } | |
2729 | ||
2730 | emit_jump (bb, node->case_bb, phi_mapping); | |
2731 | return NULL; | |
2732 | } | |
2733 | } | |
2734 | ||
2735 | return bb; | |
2736 | } | |
2737 | ||
2738 | /* Search the parent sections of the case node tree | |
2739 | to see if a test for the lower bound of NODE would be redundant. | |
2740 | INDEX_TYPE is the type of the index expression. | |
2741 | ||
2742 | The instructions to generate the case decision tree are | |
2743 | output in the same order as nodes are processed so it is | |
2744 | known that if a parent node checks the range of the current | |
2745 | node minus one that the current node is bounded at its lower | |
2746 | span. Thus the test would be redundant. */ | |
2747 | ||
2748 | static bool | |
2749 | node_has_low_bound (case_node_ptr node, tree index_type) | |
2750 | { | |
2751 | tree low_minus_one; | |
2752 | case_node_ptr pnode; | |
2753 | ||
2754 | /* If the lower bound of this node is the lowest value in the index type, | |
2755 | we need not test it. */ | |
2756 | ||
2757 | if (tree_int_cst_equal (node->low, TYPE_MIN_VALUE (index_type))) | |
2758 | return true; | |
2759 | ||
2760 | /* If this node has a left branch, the value at the left must be less | |
2761 | than that at this node, so it cannot be bounded at the bottom and | |
2762 | we need not bother testing any further. */ | |
2763 | ||
2764 | if (node->left) | |
2765 | return false; | |
2766 | ||
2767 | low_minus_one = fold_build2 (MINUS_EXPR, TREE_TYPE (node->low), node->low, | |
2768 | build_int_cst (TREE_TYPE (node->low), 1)); | |
2769 | ||
2770 | /* If the subtraction above overflowed, we can't verify anything. | |
2771 | Otherwise, look for a parent that tests our value - 1. */ | |
2772 | ||
2773 | if (!tree_int_cst_lt (low_minus_one, node->low)) | |
2774 | return false; | |
2775 | ||
2776 | for (pnode = node->parent; pnode; pnode = pnode->parent) | |
2777 | if (tree_int_cst_equal (low_minus_one, pnode->high)) | |
2778 | return true; | |
2779 | ||
2780 | return false; | |
2781 | } | |
2782 | ||
2783 | /* Search the parent sections of the case node tree | |
2784 | to see if a test for the upper bound of NODE would be redundant. | |
2785 | INDEX_TYPE is the type of the index expression. | |
2786 | ||
2787 | The instructions to generate the case decision tree are | |
2788 | output in the same order as nodes are processed so it is | |
2789 | known that if a parent node checks the range of the current | |
2790 | node plus one that the current node is bounded at its upper | |
2791 | span. Thus the test would be redundant. */ | |
2792 | ||
2793 | static bool | |
2794 | node_has_high_bound (case_node_ptr node, tree index_type) | |
2795 | { | |
2796 | tree high_plus_one; | |
2797 | case_node_ptr pnode; | |
2798 | ||
2799 | /* If there is no upper bound, obviously no test is needed. */ | |
2800 | ||
2801 | if (TYPE_MAX_VALUE (index_type) == NULL) | |
2802 | return true; | |
2803 | ||
2804 | /* If the upper bound of this node is the highest value in the type | |
2805 | of the index expression, we need not test against it. */ | |
2806 | ||
2807 | if (tree_int_cst_equal (node->high, TYPE_MAX_VALUE (index_type))) | |
2808 | return true; | |
2809 | ||
2810 | /* If this node has a right branch, the value at the right must be greater | |
2811 | than that at this node, so it cannot be bounded at the top and | |
2812 | we need not bother testing any further. */ | |
2813 | ||
2814 | if (node->right) | |
2815 | return false; | |
2816 | ||
2817 | high_plus_one = fold_build2 (PLUS_EXPR, TREE_TYPE (node->high), node->high, | |
2818 | build_int_cst (TREE_TYPE (node->high), 1)); | |
2819 | ||
2820 | /* If the addition above overflowed, we can't verify anything. | |
2821 | Otherwise, look for a parent that tests our value + 1. */ | |
2822 | ||
2823 | if (!tree_int_cst_lt (node->high, high_plus_one)) | |
2824 | return false; | |
2825 | ||
2826 | for (pnode = node->parent; pnode; pnode = pnode->parent) | |
2827 | if (tree_int_cst_equal (high_plus_one, pnode->low)) | |
2828 | return true; | |
2829 | ||
2830 | return false; | |
2831 | } | |
2832 | ||
2833 | /* Search the parent sections of the | |
2834 | case node tree to see if both tests for the upper and lower | |
2835 | bounds of NODE would be redundant. */ | |
2836 | ||
2837 | static bool | |
2838 | node_is_bounded (case_node_ptr node, tree index_type) | |
2839 | { | |
2840 | return (node_has_low_bound (node, index_type) | |
2841 | && node_has_high_bound (node, index_type)); | |
2842 | } |