1 | /* Lower vector operations to scalar operations. |
2 | Copyright (C) 2004-2024 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the |
8 | Free Software Foundation; either version 3, or (at your option) any |
9 | later version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include "config.h" |
21 | #include "system.h" |
22 | #include "coretypes.h" |
23 | #include "backend.h" |
24 | #include "rtl.h" |
25 | #include "tree.h" |
26 | #include "gimple.h" |
27 | #include "tree-pass.h" |
28 | #include "ssa.h" |
29 | #include "expmed.h" |
30 | #include "optabs-tree.h" |
31 | #include "diagnostic.h" |
32 | #include "fold-const.h" |
33 | #include "stor-layout.h" |
34 | #include "langhooks.h" |
35 | #include "tree-eh.h" |
36 | #include "gimple-iterator.h" |
37 | #include "gimplify-me.h" |
38 | #include "gimplify.h" |
39 | #include "tree-cfg.h" |
40 | #include "tree-vector-builder.h" |
41 | #include "vec-perm-indices.h" |
42 | #include "insn-config.h" |
43 | #include "tree-ssa-dce.h" |
44 | #include "gimple-fold.h" |
45 | #include "gimple-match.h" |
46 | #include "recog.h" /* FIXME: for insn_data */ |
47 | #include "optabs-libfuncs.h" |
48 | |
49 | |
50 | /* Build a ternary operation and gimplify it. Emit code before GSI. |
51 | Return the gimple_val holding the result. */ |
52 | |
53 | static tree |
54 | gimplify_build3 (gimple_stmt_iterator *gsi, enum tree_code code, |
55 | tree type, tree a, tree b, tree c) |
56 | { |
57 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
58 | return gimple_build (gsi, true, GSI_SAME_STMT, loc, code, type, a, b, c); |
59 | } |
60 | |
61 | /* Build a binary operation and gimplify it. Emit code before GSI. |
62 | Return the gimple_val holding the result. */ |
63 | |
64 | static tree |
65 | gimplify_build2 (gimple_stmt_iterator *gsi, enum tree_code code, |
66 | tree type, tree a, tree b) |
67 | { |
68 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
69 | return gimple_build (gsi, true, GSI_SAME_STMT, loc, code, type, a, b); |
70 | } |
71 | |
72 | /* Build a unary operation and gimplify it. Emit code before GSI. |
73 | Return the gimple_val holding the result. */ |
74 | |
75 | static tree |
76 | gimplify_build1 (gimple_stmt_iterator *gsi, enum tree_code code, tree type, |
77 | tree a) |
78 | { |
79 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
80 | return gimple_build (gsi, true, GSI_SAME_STMT, loc, code, type, a); |
81 | } |
82 | |
83 | |
84 | static void expand_vector_operations_1 (gimple_stmt_iterator *, bitmap); |
85 | |
86 | /* Return the number of elements in a vector type TYPE that we have |
87 | already decided needs to be expanded piecewise. We don't support |
88 | this kind of expansion for variable-length vectors, since we should |
89 | always check for target support before introducing uses of those. */ |
90 | static unsigned int |
91 | nunits_for_known_piecewise_op (const_tree type) |
92 | { |
93 | return TYPE_VECTOR_SUBPARTS (node: type).to_constant (); |
94 | } |
95 | |
96 | /* Return true if TYPE1 has more elements than TYPE2, where either |
97 | type may be a vector or a scalar. */ |
98 | |
99 | static inline bool |
100 | subparts_gt (tree type1, tree type2) |
101 | { |
102 | poly_uint64 n1 = VECTOR_TYPE_P (type1) ? TYPE_VECTOR_SUBPARTS (node: type1) : 1; |
103 | poly_uint64 n2 = VECTOR_TYPE_P (type2) ? TYPE_VECTOR_SUBPARTS (node: type2) : 1; |
104 | return known_gt (n1, n2); |
105 | } |
106 | |
107 | static GTY(()) tree vector_inner_type; |
108 | static GTY(()) tree vector_last_type; |
109 | static GTY(()) int vector_last_nunits; |
110 | |
111 | /* Return a suitable vector types made of SUBPARTS units each of mode |
112 | "word_mode" (the global variable). */ |
113 | static tree |
114 | build_word_mode_vector_type (int nunits) |
115 | { |
116 | if (!vector_inner_type) |
117 | vector_inner_type = lang_hooks.types.type_for_mode (word_mode, 1); |
118 | else if (vector_last_nunits == nunits) |
119 | { |
120 | gcc_assert (TREE_CODE (vector_last_type) == VECTOR_TYPE); |
121 | return vector_last_type; |
122 | } |
123 | |
124 | vector_last_nunits = nunits; |
125 | vector_last_type = build_vector_type (vector_inner_type, nunits); |
126 | return vector_last_type; |
127 | } |
128 | |
129 | typedef tree (*elem_op_func) (gimple_stmt_iterator *, |
130 | tree, tree, tree, tree, tree, enum tree_code, |
131 | tree); |
132 | |
133 | /* Extract the vector element of type TYPE at BITPOS with BITSIZE from T |
134 | and return it. */ |
135 | |
136 | tree |
137 | (gimple_stmt_iterator *gsi, tree type, |
138 | tree t, tree bitsize, tree bitpos) |
139 | { |
140 | /* We're using the resimplify API and maybe_push_res_to_seq to |
141 | simplify the BIT_FIELD_REF but restrict the simplification to |
142 | a single stmt while at the same time following SSA edges for |
143 | simplification with already emitted CTORs. */ |
144 | gimple_match_op opr; |
145 | opr.set_op (code_in: BIT_FIELD_REF, type_in: type, op0: t, op1: bitsize, op2: bitpos); |
146 | opr.resimplify (NULL, follow_all_ssa_edges); |
147 | gimple_seq stmts = NULL; |
148 | tree res = maybe_push_res_to_seq (&opr, &stmts); |
149 | if (!res) |
150 | { |
151 | /* This can happen if SSA_NAME_OCCURS_IN_ABNORMAL_PHI are |
152 | used. Build BIT_FIELD_REF manually otherwise. */ |
153 | t = build3 (BIT_FIELD_REF, type, t, bitsize, bitpos); |
154 | res = make_ssa_name (var: type); |
155 | gimple *g = gimple_build_assign (res, t); |
156 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
157 | return res; |
158 | } |
159 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
160 | return res; |
161 | } |
162 | |
163 | static tree |
164 | do_unop (gimple_stmt_iterator *gsi, tree inner_type, tree a, |
165 | tree b ATTRIBUTE_UNUSED, tree bitpos, tree bitsize, |
166 | enum tree_code code, tree type ATTRIBUTE_UNUSED) |
167 | { |
168 | a = tree_vec_extract (gsi, type: inner_type, t: a, bitsize, bitpos); |
169 | return gimplify_build1 (gsi, code, type: inner_type, a); |
170 | } |
171 | |
172 | static tree |
173 | do_binop (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b, |
174 | tree bitpos, tree bitsize, enum tree_code code, |
175 | tree type ATTRIBUTE_UNUSED) |
176 | { |
177 | if (VECTOR_TYPE_P (TREE_TYPE (a))) |
178 | a = tree_vec_extract (gsi, type: inner_type, t: a, bitsize, bitpos); |
179 | if (VECTOR_TYPE_P (TREE_TYPE (b))) |
180 | b = tree_vec_extract (gsi, type: inner_type, t: b, bitsize, bitpos); |
181 | return gimplify_build2 (gsi, code, type: inner_type, a, b); |
182 | } |
183 | |
184 | /* Construct expression (A[BITPOS] code B[BITPOS]) ? -1 : 0 |
185 | |
186 | INNER_TYPE is the type of A and B elements |
187 | |
188 | returned expression is of signed integer type with the |
189 | size equal to the size of INNER_TYPE. */ |
190 | static tree |
191 | do_compare (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b, |
192 | tree bitpos, tree bitsize, enum tree_code code, tree type) |
193 | { |
194 | tree stype = TREE_TYPE (type); |
195 | tree cst_false = build_zero_cst (stype); |
196 | tree cst_true = build_all_ones_cst (stype); |
197 | tree cmp; |
198 | |
199 | a = tree_vec_extract (gsi, type: inner_type, t: a, bitsize, bitpos); |
200 | b = tree_vec_extract (gsi, type: inner_type, t: b, bitsize, bitpos); |
201 | |
202 | cmp = build2 (code, boolean_type_node, a, b); |
203 | return gimplify_build3 (gsi, code: COND_EXPR, type: stype, a: cmp, b: cst_true, c: cst_false); |
204 | } |
205 | |
206 | /* Expand vector addition to scalars. This does bit twiddling |
207 | in order to increase parallelism: |
208 | |
209 | a + b = (((int) a & 0x7f7f7f7f) + ((int) b & 0x7f7f7f7f)) ^ |
210 | (a ^ b) & 0x80808080 |
211 | |
212 | a - b = (((int) a | 0x80808080) - ((int) b & 0x7f7f7f7f)) ^ |
213 | (a ^ ~b) & 0x80808080 |
214 | |
215 | -b = (0x80808080 - ((int) b & 0x7f7f7f7f)) ^ (~b & 0x80808080) |
216 | |
217 | This optimization should be done only if 4 vector items or more |
218 | fit into a word. */ |
219 | static tree |
220 | do_plus_minus (gimple_stmt_iterator *gsi, tree word_type, tree a, tree b, |
221 | tree bitpos ATTRIBUTE_UNUSED, tree bitsize ATTRIBUTE_UNUSED, |
222 | enum tree_code code, tree type ATTRIBUTE_UNUSED) |
223 | { |
224 | unsigned int width = vector_element_bits (TREE_TYPE (a)); |
225 | tree inner_type = TREE_TYPE (TREE_TYPE (a)); |
226 | unsigned HOST_WIDE_INT max; |
227 | tree low_bits, high_bits, a_low, b_low, result_low, signs; |
228 | |
229 | max = GET_MODE_MASK (TYPE_MODE (inner_type)); |
230 | low_bits = build_replicated_int_cst (word_type, width, max >> 1); |
231 | high_bits = build_replicated_int_cst (word_type, width, max & ~(max >> 1)); |
232 | |
233 | a = tree_vec_extract (gsi, type: word_type, t: a, bitsize, bitpos); |
234 | b = tree_vec_extract (gsi, type: word_type, t: b, bitsize, bitpos); |
235 | |
236 | signs = gimplify_build2 (gsi, code: BIT_XOR_EXPR, type: word_type, a, b); |
237 | b_low = gimplify_build2 (gsi, code: BIT_AND_EXPR, type: word_type, a: b, b: low_bits); |
238 | if (code == PLUS_EXPR) |
239 | a_low = gimplify_build2 (gsi, code: BIT_AND_EXPR, type: word_type, a, b: low_bits); |
240 | else |
241 | { |
242 | a_low = gimplify_build2 (gsi, code: BIT_IOR_EXPR, type: word_type, a, b: high_bits); |
243 | signs = gimplify_build1 (gsi, code: BIT_NOT_EXPR, type: word_type, a: signs); |
244 | } |
245 | |
246 | signs = gimplify_build2 (gsi, code: BIT_AND_EXPR, type: word_type, a: signs, b: high_bits); |
247 | result_low = gimplify_build2 (gsi, code, type: word_type, a: a_low, b: b_low); |
248 | return gimplify_build2 (gsi, code: BIT_XOR_EXPR, type: word_type, a: result_low, b: signs); |
249 | } |
250 | |
251 | static tree |
252 | do_negate (gimple_stmt_iterator *gsi, tree word_type, tree b, |
253 | tree unused ATTRIBUTE_UNUSED, tree bitpos ATTRIBUTE_UNUSED, |
254 | tree bitsize ATTRIBUTE_UNUSED, |
255 | enum tree_code code ATTRIBUTE_UNUSED, |
256 | tree type ATTRIBUTE_UNUSED) |
257 | { |
258 | unsigned int width = vector_element_bits (TREE_TYPE (b)); |
259 | tree inner_type = TREE_TYPE (TREE_TYPE (b)); |
260 | HOST_WIDE_INT max; |
261 | tree low_bits, high_bits, b_low, result_low, signs; |
262 | |
263 | max = GET_MODE_MASK (TYPE_MODE (inner_type)); |
264 | low_bits = build_replicated_int_cst (word_type, width, max >> 1); |
265 | high_bits = build_replicated_int_cst (word_type, width, max & ~(max >> 1)); |
266 | |
267 | b = tree_vec_extract (gsi, type: word_type, t: b, bitsize, bitpos); |
268 | |
269 | b_low = gimplify_build2 (gsi, code: BIT_AND_EXPR, type: word_type, a: b, b: low_bits); |
270 | signs = gimplify_build1 (gsi, code: BIT_NOT_EXPR, type: word_type, a: b); |
271 | signs = gimplify_build2 (gsi, code: BIT_AND_EXPR, type: word_type, a: signs, b: high_bits); |
272 | result_low = gimplify_build2 (gsi, code: MINUS_EXPR, type: word_type, a: high_bits, b: b_low); |
273 | return gimplify_build2 (gsi, code: BIT_XOR_EXPR, type: word_type, a: result_low, b: signs); |
274 | } |
275 | |
276 | /* Expand a vector operation to scalars, by using many operations |
277 | whose type is the vector type's inner type. */ |
278 | static tree |
279 | expand_vector_piecewise (gimple_stmt_iterator *gsi, elem_op_func f, |
280 | tree type, tree inner_type, |
281 | tree a, tree b, enum tree_code code, |
282 | bool parallel_p, tree ret_type = NULL_TREE) |
283 | { |
284 | vec<constructor_elt, va_gc> *v; |
285 | tree part_width = TYPE_SIZE (inner_type); |
286 | tree index = bitsize_int (0); |
287 | int nunits = nunits_for_known_piecewise_op (type); |
288 | int delta = tree_to_uhwi (part_width) / vector_element_bits (type); |
289 | int i; |
290 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
291 | |
292 | if (nunits == 1 |
293 | || warning_suppressed_p (gsi_stmt (i: *gsi), |
294 | OPT_Wvector_operation_performance)) |
295 | /* Do not diagnose decomposing single element vectors or when |
296 | decomposing vectorizer produced operations. */ |
297 | ; |
298 | else if (ret_type || !parallel_p) |
299 | warning_at (loc, OPT_Wvector_operation_performance, |
300 | "vector operation will be expanded piecewise" ); |
301 | else |
302 | warning_at (loc, OPT_Wvector_operation_performance, |
303 | "vector operation will be expanded in parallel" ); |
304 | |
305 | if (!ret_type) |
306 | ret_type = type; |
307 | vec_alloc (v, nelems: (nunits + delta - 1) / delta); |
308 | bool constant_p = true; |
309 | for (i = 0; i < nunits; |
310 | i += delta, index = int_const_binop (PLUS_EXPR, index, part_width)) |
311 | { |
312 | tree result = f (gsi, inner_type, a, b, index, part_width, code, |
313 | ret_type); |
314 | if (!CONSTANT_CLASS_P (result)) |
315 | constant_p = false; |
316 | constructor_elt ce = {NULL_TREE, .value: result}; |
317 | v->quick_push (obj: ce); |
318 | } |
319 | |
320 | if (constant_p) |
321 | return build_vector_from_ctor (ret_type, v); |
322 | else |
323 | return build_constructor (ret_type, v); |
324 | } |
325 | |
326 | /* Expand a vector operation to scalars with the freedom to use |
327 | a scalar integer type, or to use a different size for the items |
328 | in the vector type. */ |
329 | static tree |
330 | expand_vector_parallel (gimple_stmt_iterator *gsi, elem_op_func f, tree type, |
331 | tree a, tree b, enum tree_code code) |
332 | { |
333 | tree result, compute_type; |
334 | int n_words = tree_to_uhwi (TYPE_SIZE_UNIT (type)) / UNITS_PER_WORD; |
335 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
336 | |
337 | /* We have three strategies. If the type is already correct, just do |
338 | the operation an element at a time. Else, if the vector is wider than |
339 | one word, do it a word at a time; finally, if the vector is smaller |
340 | than one word, do it as a scalar. */ |
341 | if (TYPE_MODE (TREE_TYPE (type)) == word_mode) |
342 | return expand_vector_piecewise (gsi, f, |
343 | type, TREE_TYPE (type), |
344 | a, b, code, parallel_p: true); |
345 | else if (n_words > 1) |
346 | { |
347 | tree word_type = build_word_mode_vector_type (nunits: n_words); |
348 | result = expand_vector_piecewise (gsi, f, |
349 | type: word_type, TREE_TYPE (word_type), |
350 | a, b, code, parallel_p: true); |
351 | result = force_gimple_operand_gsi (gsi, result, true, NULL, true, |
352 | GSI_SAME_STMT); |
353 | } |
354 | else |
355 | { |
356 | /* Use a single scalar operation with a mode no wider than word_mode. */ |
357 | if (!warning_suppressed_p (gsi_stmt (i: *gsi), |
358 | OPT_Wvector_operation_performance)) |
359 | warning_at (loc, OPT_Wvector_operation_performance, |
360 | "vector operation will be expanded with a " |
361 | "single scalar operation" ); |
362 | scalar_int_mode mode |
363 | = int_mode_for_size (size: tree_to_uhwi (TYPE_SIZE (type)), limit: 0).require (); |
364 | compute_type = lang_hooks.types.type_for_mode (mode, 1); |
365 | result = f (gsi, compute_type, a, b, bitsize_zero_node, |
366 | TYPE_SIZE (compute_type), code, type); |
367 | } |
368 | |
369 | return result; |
370 | } |
371 | |
372 | /* Expand a vector operation to scalars; for integer types we can use |
373 | special bit twiddling tricks to do the sums a word at a time, using |
374 | function F_PARALLEL instead of F. These tricks are done only if |
375 | they can process at least four items, that is, only if the vector |
376 | holds at least four items and if a word can hold four items. */ |
377 | static tree |
378 | expand_vector_addition (gimple_stmt_iterator *gsi, |
379 | elem_op_func f, elem_op_func f_parallel, |
380 | tree type, tree a, tree b, enum tree_code code) |
381 | { |
382 | int parts_per_word = BITS_PER_WORD / vector_element_bits (type); |
383 | |
384 | if (INTEGRAL_TYPE_P (TREE_TYPE (type)) |
385 | && parts_per_word >= 4 |
386 | && nunits_for_known_piecewise_op (type) >= 4) |
387 | return expand_vector_parallel (gsi, f: f_parallel, |
388 | type, a, b, code); |
389 | else |
390 | return expand_vector_piecewise (gsi, f, |
391 | type, TREE_TYPE (type), |
392 | a, b, code, parallel_p: false); |
393 | } |
394 | |
395 | static bool |
396 | expand_vector_condition (gimple_stmt_iterator *gsi, bitmap dce_ssa_names); |
397 | |
398 | /* Try to expand vector comparison expression OP0 CODE OP1 by |
399 | querying optab if the following expression: |
400 | VEC_COND_EXPR< OP0 CODE OP1, {-1,...}, {0,...}> |
401 | can be expanded. */ |
402 | static tree |
403 | expand_vector_comparison (gimple_stmt_iterator *gsi, tree type, tree op0, |
404 | tree op1, enum tree_code code, |
405 | bitmap dce_ssa_names) |
406 | { |
407 | tree lhs = gimple_assign_lhs (gs: gsi_stmt (i: *gsi)); |
408 | use_operand_p use_p; |
409 | imm_use_iterator iterator; |
410 | bool vec_cond_expr_only = true; |
411 | |
412 | /* As seen in PR95830, we should not expand comparisons that are only |
413 | feeding a VEC_COND_EXPR statement. */ |
414 | auto_vec<gimple *> uses; |
415 | FOR_EACH_IMM_USE_FAST (use_p, iterator, lhs) |
416 | { |
417 | gimple *use = USE_STMT (use_p); |
418 | if (is_gimple_debug (gs: use)) |
419 | continue; |
420 | if (is_gimple_assign (gs: use) |
421 | && gimple_assign_rhs_code (gs: use) == VEC_COND_EXPR |
422 | && gimple_assign_rhs1 (gs: use) == lhs |
423 | && gimple_assign_rhs2 (gs: use) != lhs |
424 | && gimple_assign_rhs3 (gs: use) != lhs) |
425 | uses.safe_push (obj: use); |
426 | else |
427 | vec_cond_expr_only = false; |
428 | } |
429 | |
430 | if (vec_cond_expr_only) |
431 | for (gimple *use : uses) |
432 | { |
433 | gimple_stmt_iterator it = gsi_for_stmt (use); |
434 | if (!expand_vector_condition (gsi: &it, dce_ssa_names)) |
435 | { |
436 | vec_cond_expr_only = false; |
437 | break; |
438 | } |
439 | } |
440 | |
441 | if (!uses.is_empty () && vec_cond_expr_only) |
442 | return NULL_TREE; |
443 | |
444 | tree t; |
445 | if (!expand_vec_cmp_expr_p (TREE_TYPE (op0), type, code)) |
446 | { |
447 | if (VECTOR_BOOLEAN_TYPE_P (type) |
448 | && SCALAR_INT_MODE_P (TYPE_MODE (type)) |
449 | && known_lt (GET_MODE_BITSIZE (TYPE_MODE (type)), |
450 | TYPE_VECTOR_SUBPARTS (type) |
451 | * GET_MODE_BITSIZE (SCALAR_TYPE_MODE |
452 | (TREE_TYPE (type))))) |
453 | { |
454 | tree inner_type = TREE_TYPE (TREE_TYPE (op0)); |
455 | tree part_width = vector_element_bits_tree (TREE_TYPE (op0)); |
456 | tree index = bitsize_int (0); |
457 | int nunits = nunits_for_known_piecewise_op (TREE_TYPE (op0)); |
458 | int prec = GET_MODE_PRECISION (SCALAR_TYPE_MODE (type)); |
459 | tree ret_type = build_nonstandard_integer_type (prec, 1); |
460 | tree ret_inner_type = boolean_type_node; |
461 | int i; |
462 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
463 | t = build_zero_cst (ret_type); |
464 | |
465 | if (TYPE_PRECISION (ret_inner_type) != 1) |
466 | ret_inner_type = build_nonstandard_integer_type (1, 1); |
467 | if (!warning_suppressed_p (gsi_stmt (i: *gsi), |
468 | OPT_Wvector_operation_performance)) |
469 | warning_at (loc, OPT_Wvector_operation_performance, |
470 | "vector operation will be expanded piecewise" ); |
471 | for (i = 0; i < nunits; |
472 | i++, index = int_const_binop (PLUS_EXPR, index, part_width)) |
473 | { |
474 | tree a = tree_vec_extract (gsi, type: inner_type, t: op0, bitsize: part_width, |
475 | bitpos: index); |
476 | tree b = tree_vec_extract (gsi, type: inner_type, t: op1, bitsize: part_width, |
477 | bitpos: index); |
478 | tree result = gimplify_build2 (gsi, code, type: ret_inner_type, a, b); |
479 | t = gimplify_build3 (gsi, code: BIT_INSERT_EXPR, type: ret_type, a: t, b: result, |
480 | bitsize_int (i)); |
481 | } |
482 | t = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, type, a: t); |
483 | } |
484 | else |
485 | t = expand_vector_piecewise (gsi, f: do_compare, type, |
486 | TREE_TYPE (TREE_TYPE (op0)), a: op0, b: op1, |
487 | code, parallel_p: false); |
488 | } |
489 | else |
490 | t = NULL_TREE; |
491 | |
492 | return t; |
493 | } |
494 | |
495 | /* Helper function of expand_vector_divmod. Gimplify a RSHIFT_EXPR in type |
496 | of OP0 with shift counts in SHIFTCNTS array and return the temporary holding |
497 | the result if successful, otherwise return NULL_TREE. */ |
498 | static tree |
499 | add_rshift (gimple_stmt_iterator *gsi, tree type, tree op0, int *shiftcnts) |
500 | { |
501 | optab op; |
502 | unsigned int i, nunits = nunits_for_known_piecewise_op (type); |
503 | bool scalar_shift = true; |
504 | |
505 | for (i = 1; i < nunits; i++) |
506 | { |
507 | if (shiftcnts[i] != shiftcnts[0]) |
508 | scalar_shift = false; |
509 | } |
510 | |
511 | if (scalar_shift && shiftcnts[0] == 0) |
512 | return op0; |
513 | |
514 | if (scalar_shift) |
515 | { |
516 | op = optab_for_tree_code (RSHIFT_EXPR, type, optab_scalar); |
517 | if (op != unknown_optab |
518 | && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing) |
519 | return gimplify_build2 (gsi, code: RSHIFT_EXPR, type, a: op0, |
520 | b: build_int_cst (NULL_TREE, shiftcnts[0])); |
521 | } |
522 | |
523 | op = optab_for_tree_code (RSHIFT_EXPR, type, optab_vector); |
524 | if (op != unknown_optab |
525 | && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing) |
526 | { |
527 | tree_vector_builder vec (type, nunits, 1); |
528 | for (i = 0; i < nunits; i++) |
529 | vec.quick_push (obj: build_int_cst (TREE_TYPE (type), shiftcnts[i])); |
530 | return gimplify_build2 (gsi, code: RSHIFT_EXPR, type, a: op0, b: vec.build ()); |
531 | } |
532 | |
533 | return NULL_TREE; |
534 | } |
535 | |
536 | /* Try to expand integer vector division by constant using |
537 | widening multiply, shifts and additions. */ |
538 | static tree |
539 | expand_vector_divmod (gimple_stmt_iterator *gsi, tree type, tree op0, |
540 | tree op1, enum tree_code code) |
541 | { |
542 | bool use_pow2 = true; |
543 | bool has_vector_shift = true; |
544 | bool use_abs_op1 = false; |
545 | int mode = -1, this_mode; |
546 | int pre_shift = -1, post_shift; |
547 | unsigned int nunits = nunits_for_known_piecewise_op (type); |
548 | int *shifts = XALLOCAVEC (int, nunits * 4); |
549 | int *pre_shifts = shifts + nunits; |
550 | int *post_shifts = pre_shifts + nunits; |
551 | int *shift_temps = post_shifts + nunits; |
552 | unsigned HOST_WIDE_INT *mulc = XALLOCAVEC (unsigned HOST_WIDE_INT, nunits); |
553 | int prec = TYPE_PRECISION (TREE_TYPE (type)); |
554 | int dummy_int; |
555 | unsigned int i; |
556 | signop sign_p = TYPE_SIGN (TREE_TYPE (type)); |
557 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (TYPE_MODE (TREE_TYPE (type))); |
558 | tree cur_op, mulcst, tem; |
559 | optab op; |
560 | |
561 | if (prec > HOST_BITS_PER_WIDE_INT) |
562 | return NULL_TREE; |
563 | |
564 | op = optab_for_tree_code (RSHIFT_EXPR, type, optab_vector); |
565 | if (op == unknown_optab |
566 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
567 | has_vector_shift = false; |
568 | |
569 | /* Analysis phase. Determine if all op1 elements are either power |
570 | of two and it is possible to expand it using shifts (or for remainder |
571 | using masking). Additionally compute the multiplicative constants |
572 | and pre and post shifts if the division is to be expanded using |
573 | widening or high part multiplication plus shifts. */ |
574 | for (i = 0; i < nunits; i++) |
575 | { |
576 | tree cst = VECTOR_CST_ELT (op1, i); |
577 | unsigned HOST_WIDE_INT ml; |
578 | |
579 | if (TREE_CODE (cst) != INTEGER_CST || integer_zerop (cst)) |
580 | return NULL_TREE; |
581 | pre_shifts[i] = 0; |
582 | post_shifts[i] = 0; |
583 | mulc[i] = 0; |
584 | if (use_pow2 |
585 | && (!integer_pow2p (cst) || tree_int_cst_sgn (cst) != 1)) |
586 | use_pow2 = false; |
587 | if (use_pow2) |
588 | { |
589 | shifts[i] = tree_log2 (cst); |
590 | if (shifts[i] != shifts[0] |
591 | && code == TRUNC_DIV_EXPR |
592 | && !has_vector_shift) |
593 | use_pow2 = false; |
594 | } |
595 | if (mode == -2) |
596 | continue; |
597 | if (sign_p == UNSIGNED) |
598 | { |
599 | unsigned HOST_WIDE_INT mh; |
600 | unsigned HOST_WIDE_INT d = TREE_INT_CST_LOW (cst) & mask; |
601 | |
602 | if (d >= (HOST_WIDE_INT_1U << (prec - 1))) |
603 | /* FIXME: Can transform this into op0 >= op1 ? 1 : 0. */ |
604 | return NULL_TREE; |
605 | |
606 | if (d <= 1) |
607 | { |
608 | mode = -2; |
609 | continue; |
610 | } |
611 | |
612 | /* Find a suitable multiplier and right shift count |
613 | instead of multiplying with D. */ |
614 | mh = choose_multiplier (d, prec, prec, &ml, &post_shift, &dummy_int); |
615 | |
616 | /* If the suggested multiplier is more than SIZE bits, we can |
617 | do better for even divisors, using an initial right shift. */ |
618 | if ((mh != 0 && (d & 1) == 0) |
619 | || (!has_vector_shift && pre_shift != -1)) |
620 | { |
621 | if (has_vector_shift) |
622 | pre_shift = ctz_or_zero (x: d); |
623 | else if (pre_shift == -1) |
624 | { |
625 | unsigned int j; |
626 | for (j = 0; j < nunits; j++) |
627 | { |
628 | tree cst2 = VECTOR_CST_ELT (op1, j); |
629 | unsigned HOST_WIDE_INT d2; |
630 | int this_pre_shift; |
631 | |
632 | if (!tree_fits_uhwi_p (cst2)) |
633 | return NULL_TREE; |
634 | d2 = tree_to_uhwi (cst2) & mask; |
635 | if (d2 == 0) |
636 | return NULL_TREE; |
637 | this_pre_shift = floor_log2 (x: d2 & -d2); |
638 | if (pre_shift == -1 || this_pre_shift < pre_shift) |
639 | pre_shift = this_pre_shift; |
640 | } |
641 | if (i != 0 && pre_shift != 0) |
642 | { |
643 | /* Restart. */ |
644 | i = -1U; |
645 | mode = -1; |
646 | continue; |
647 | } |
648 | } |
649 | if (pre_shift != 0) |
650 | { |
651 | if ((d >> pre_shift) <= 1) |
652 | { |
653 | mode = -2; |
654 | continue; |
655 | } |
656 | mh = choose_multiplier (d >> pre_shift, prec, |
657 | prec - pre_shift, |
658 | &ml, &post_shift, &dummy_int); |
659 | gcc_assert (!mh); |
660 | pre_shifts[i] = pre_shift; |
661 | } |
662 | } |
663 | if (!mh) |
664 | this_mode = 0; |
665 | else |
666 | this_mode = 1; |
667 | } |
668 | else |
669 | { |
670 | HOST_WIDE_INT d = TREE_INT_CST_LOW (cst); |
671 | unsigned HOST_WIDE_INT abs_d; |
672 | |
673 | if (d == -1) |
674 | return NULL_TREE; |
675 | |
676 | /* Since d might be INT_MIN, we have to cast to |
677 | unsigned HOST_WIDE_INT before negating to avoid |
678 | undefined signed overflow. */ |
679 | abs_d = (d >= 0 |
680 | ? (unsigned HOST_WIDE_INT) d |
681 | : - (unsigned HOST_WIDE_INT) d); |
682 | |
683 | /* n rem d = n rem -d */ |
684 | if (code == TRUNC_MOD_EXPR && d < 0) |
685 | { |
686 | d = abs_d; |
687 | use_abs_op1 = true; |
688 | } |
689 | if (abs_d == HOST_WIDE_INT_1U << (prec - 1)) |
690 | { |
691 | /* This case is not handled correctly below. */ |
692 | mode = -2; |
693 | continue; |
694 | } |
695 | if (abs_d <= 1) |
696 | { |
697 | mode = -2; |
698 | continue; |
699 | } |
700 | |
701 | choose_multiplier (abs_d, prec, prec - 1, &ml, |
702 | &post_shift, &dummy_int); |
703 | if (ml >= HOST_WIDE_INT_1U << (prec - 1)) |
704 | { |
705 | this_mode = 4 + (d < 0); |
706 | ml |= HOST_WIDE_INT_M1U << (prec - 1); |
707 | } |
708 | else |
709 | this_mode = 2 + (d < 0); |
710 | } |
711 | mulc[i] = ml; |
712 | post_shifts[i] = post_shift; |
713 | if ((i && !has_vector_shift && post_shifts[0] != post_shift) |
714 | || post_shift >= prec |
715 | || pre_shifts[i] >= prec) |
716 | this_mode = -2; |
717 | |
718 | if (i == 0) |
719 | mode = this_mode; |
720 | else if (mode != this_mode) |
721 | mode = -2; |
722 | } |
723 | |
724 | if (use_pow2) |
725 | { |
726 | tree addend = NULL_TREE; |
727 | if (sign_p == SIGNED) |
728 | { |
729 | tree uns_type; |
730 | |
731 | /* Both division and remainder sequences need |
732 | op0 < 0 ? mask : 0 computed. It can be either computed as |
733 | (type) (((uns_type) (op0 >> (prec - 1))) >> (prec - shifts[i])) |
734 | if none of the shifts is 0, or as the conditional. */ |
735 | for (i = 0; i < nunits; i++) |
736 | if (shifts[i] == 0) |
737 | break; |
738 | uns_type |
739 | = build_vector_type (build_nonstandard_integer_type (prec, 1), |
740 | nunits); |
741 | if (i == nunits && TYPE_MODE (uns_type) == TYPE_MODE (type)) |
742 | { |
743 | for (i = 0; i < nunits; i++) |
744 | shift_temps[i] = prec - 1; |
745 | cur_op = add_rshift (gsi, type, op0, shiftcnts: shift_temps); |
746 | if (cur_op != NULL_TREE) |
747 | { |
748 | cur_op = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, |
749 | type: uns_type, a: cur_op); |
750 | for (i = 0; i < nunits; i++) |
751 | shift_temps[i] = prec - shifts[i]; |
752 | cur_op = add_rshift (gsi, type: uns_type, op0: cur_op, shiftcnts: shift_temps); |
753 | if (cur_op != NULL_TREE) |
754 | addend = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, |
755 | type, a: cur_op); |
756 | } |
757 | } |
758 | if (addend == NULL_TREE |
759 | && expand_vec_cond_expr_p (type, type, LT_EXPR)) |
760 | { |
761 | tree zero, cst, mask_type, mask; |
762 | gimple *stmt, *cond; |
763 | |
764 | mask_type = truth_type_for (type); |
765 | zero = build_zero_cst (type); |
766 | mask = make_ssa_name (var: mask_type); |
767 | cond = gimple_build_assign (mask, LT_EXPR, op0, zero); |
768 | gsi_insert_before (gsi, cond, GSI_SAME_STMT); |
769 | tree_vector_builder vec (type, nunits, 1); |
770 | for (i = 0; i < nunits; i++) |
771 | vec.quick_push (obj: build_int_cst (TREE_TYPE (type), |
772 | (HOST_WIDE_INT_1U |
773 | << shifts[i]) - 1)); |
774 | cst = vec.build (); |
775 | addend = make_ssa_name (var: type); |
776 | stmt |
777 | = gimple_build_assign (addend, VEC_COND_EXPR, mask, cst, zero); |
778 | gsi_insert_before (gsi, stmt, GSI_SAME_STMT); |
779 | } |
780 | } |
781 | if (code == TRUNC_DIV_EXPR) |
782 | { |
783 | if (sign_p == UNSIGNED) |
784 | { |
785 | /* q = op0 >> shift; */ |
786 | cur_op = add_rshift (gsi, type, op0, shiftcnts: shifts); |
787 | if (cur_op != NULL_TREE) |
788 | return cur_op; |
789 | } |
790 | else if (addend != NULL_TREE) |
791 | { |
792 | /* t1 = op0 + addend; |
793 | q = t1 >> shift; */ |
794 | op = optab_for_tree_code (PLUS_EXPR, type, optab_default); |
795 | if (op != unknown_optab |
796 | && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing) |
797 | { |
798 | cur_op = gimplify_build2 (gsi, code: PLUS_EXPR, type, a: op0, b: addend); |
799 | cur_op = add_rshift (gsi, type, op0: cur_op, shiftcnts: shifts); |
800 | if (cur_op != NULL_TREE) |
801 | return cur_op; |
802 | } |
803 | } |
804 | } |
805 | else |
806 | { |
807 | tree mask; |
808 | tree_vector_builder vec (type, nunits, 1); |
809 | for (i = 0; i < nunits; i++) |
810 | vec.quick_push (obj: build_int_cst (TREE_TYPE (type), |
811 | (HOST_WIDE_INT_1U |
812 | << shifts[i]) - 1)); |
813 | mask = vec.build (); |
814 | op = optab_for_tree_code (BIT_AND_EXPR, type, optab_default); |
815 | if (op != unknown_optab |
816 | && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing) |
817 | { |
818 | if (sign_p == UNSIGNED) |
819 | /* r = op0 & mask; */ |
820 | return gimplify_build2 (gsi, code: BIT_AND_EXPR, type, a: op0, b: mask); |
821 | else if (addend != NULL_TREE) |
822 | { |
823 | /* t1 = op0 + addend; |
824 | t2 = t1 & mask; |
825 | r = t2 - addend; */ |
826 | op = optab_for_tree_code (PLUS_EXPR, type, optab_default); |
827 | if (op != unknown_optab |
828 | && optab_handler (op, TYPE_MODE (type)) |
829 | != CODE_FOR_nothing) |
830 | { |
831 | cur_op = gimplify_build2 (gsi, code: PLUS_EXPR, type, a: op0, |
832 | b: addend); |
833 | cur_op = gimplify_build2 (gsi, code: BIT_AND_EXPR, type, |
834 | a: cur_op, b: mask); |
835 | op = optab_for_tree_code (MINUS_EXPR, type, |
836 | optab_default); |
837 | if (op != unknown_optab |
838 | && optab_handler (op, TYPE_MODE (type)) |
839 | != CODE_FOR_nothing) |
840 | return gimplify_build2 (gsi, code: MINUS_EXPR, type, |
841 | a: cur_op, b: addend); |
842 | } |
843 | } |
844 | } |
845 | } |
846 | } |
847 | |
848 | if (mode == -2 || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN) |
849 | return NULL_TREE; |
850 | |
851 | if (!can_mult_highpart_p (TYPE_MODE (type), TYPE_UNSIGNED (type))) |
852 | return NULL_TREE; |
853 | |
854 | cur_op = op0; |
855 | |
856 | switch (mode) |
857 | { |
858 | case 0: |
859 | gcc_assert (sign_p == UNSIGNED); |
860 | /* t1 = oprnd0 >> pre_shift; |
861 | t2 = t1 h* ml; |
862 | q = t2 >> post_shift; */ |
863 | cur_op = add_rshift (gsi, type, op0: cur_op, shiftcnts: pre_shifts); |
864 | if (cur_op == NULL_TREE) |
865 | return NULL_TREE; |
866 | break; |
867 | case 1: |
868 | gcc_assert (sign_p == UNSIGNED); |
869 | for (i = 0; i < nunits; i++) |
870 | { |
871 | shift_temps[i] = 1; |
872 | post_shifts[i]--; |
873 | } |
874 | break; |
875 | case 2: |
876 | case 3: |
877 | case 4: |
878 | case 5: |
879 | gcc_assert (sign_p == SIGNED); |
880 | for (i = 0; i < nunits; i++) |
881 | shift_temps[i] = prec - 1; |
882 | break; |
883 | default: |
884 | return NULL_TREE; |
885 | } |
886 | |
887 | tree_vector_builder vec (type, nunits, 1); |
888 | for (i = 0; i < nunits; i++) |
889 | vec.quick_push (obj: build_int_cst (TREE_TYPE (type), mulc[i])); |
890 | mulcst = vec.build (); |
891 | |
892 | cur_op = gimplify_build2 (gsi, code: MULT_HIGHPART_EXPR, type, a: cur_op, b: mulcst); |
893 | |
894 | switch (mode) |
895 | { |
896 | case 0: |
897 | /* t1 = oprnd0 >> pre_shift; |
898 | t2 = t1 h* ml; |
899 | q = t2 >> post_shift; */ |
900 | cur_op = add_rshift (gsi, type, op0: cur_op, shiftcnts: post_shifts); |
901 | break; |
902 | case 1: |
903 | /* t1 = oprnd0 h* ml; |
904 | t2 = oprnd0 - t1; |
905 | t3 = t2 >> 1; |
906 | t4 = t1 + t3; |
907 | q = t4 >> (post_shift - 1); */ |
908 | op = optab_for_tree_code (MINUS_EXPR, type, optab_default); |
909 | if (op == unknown_optab |
910 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
911 | return NULL_TREE; |
912 | tem = gimplify_build2 (gsi, code: MINUS_EXPR, type, a: op0, b: cur_op); |
913 | tem = add_rshift (gsi, type, op0: tem, shiftcnts: shift_temps); |
914 | op = optab_for_tree_code (PLUS_EXPR, type, optab_default); |
915 | if (op == unknown_optab |
916 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
917 | return NULL_TREE; |
918 | tem = gimplify_build2 (gsi, code: PLUS_EXPR, type, a: cur_op, b: tem); |
919 | cur_op = add_rshift (gsi, type, op0: tem, shiftcnts: post_shifts); |
920 | if (cur_op == NULL_TREE) |
921 | return NULL_TREE; |
922 | break; |
923 | case 2: |
924 | case 3: |
925 | case 4: |
926 | case 5: |
927 | /* t1 = oprnd0 h* ml; |
928 | t2 = t1; [ iff (mode & 2) != 0 ] |
929 | t2 = t1 + oprnd0; [ iff (mode & 2) == 0 ] |
930 | t3 = t2 >> post_shift; |
931 | t4 = oprnd0 >> (prec - 1); |
932 | q = t3 - t4; [ iff (mode & 1) == 0 ] |
933 | q = t4 - t3; [ iff (mode & 1) != 0 ] */ |
934 | if ((mode & 2) == 0) |
935 | { |
936 | op = optab_for_tree_code (PLUS_EXPR, type, optab_default); |
937 | if (op == unknown_optab |
938 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
939 | return NULL_TREE; |
940 | cur_op = gimplify_build2 (gsi, code: PLUS_EXPR, type, a: cur_op, b: op0); |
941 | } |
942 | cur_op = add_rshift (gsi, type, op0: cur_op, shiftcnts: post_shifts); |
943 | if (cur_op == NULL_TREE) |
944 | return NULL_TREE; |
945 | tem = add_rshift (gsi, type, op0, shiftcnts: shift_temps); |
946 | if (tem == NULL_TREE) |
947 | return NULL_TREE; |
948 | op = optab_for_tree_code (MINUS_EXPR, type, optab_default); |
949 | if (op == unknown_optab |
950 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
951 | return NULL_TREE; |
952 | if ((mode & 1) == 0) |
953 | cur_op = gimplify_build2 (gsi, code: MINUS_EXPR, type, a: cur_op, b: tem); |
954 | else |
955 | cur_op = gimplify_build2 (gsi, code: MINUS_EXPR, type, a: tem, b: cur_op); |
956 | break; |
957 | default: |
958 | gcc_unreachable (); |
959 | } |
960 | |
961 | if (code == TRUNC_DIV_EXPR) |
962 | return cur_op; |
963 | |
964 | /* We divided. Now finish by: |
965 | t1 = q * oprnd1; |
966 | r = oprnd0 - t1; */ |
967 | op = optab_for_tree_code (MULT_EXPR, type, optab_default); |
968 | if (op == unknown_optab |
969 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
970 | return NULL_TREE; |
971 | if (use_abs_op1) |
972 | { |
973 | tree_vector_builder elts; |
974 | if (!elts.new_unary_operation (shape: type, vec: op1, allow_stepped_p: false)) |
975 | return NULL_TREE; |
976 | unsigned int count = elts.encoded_nelts (); |
977 | for (unsigned int i = 0; i < count; ++i) |
978 | { |
979 | tree elem1 = VECTOR_CST_ELT (op1, i); |
980 | |
981 | tree elt = const_unop (ABS_EXPR, TREE_TYPE (elem1), elem1); |
982 | if (elt == NULL_TREE) |
983 | return NULL_TREE; |
984 | elts.quick_push (obj: elt); |
985 | } |
986 | op1 = elts.build (); |
987 | } |
988 | tem = gimplify_build2 (gsi, code: MULT_EXPR, type, a: cur_op, b: op1); |
989 | op = optab_for_tree_code (MINUS_EXPR, type, optab_default); |
990 | if (op == unknown_optab |
991 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
992 | return NULL_TREE; |
993 | return gimplify_build2 (gsi, code: MINUS_EXPR, type, a: op0, b: tem); |
994 | } |
995 | |
996 | /* Expand a vector condition to scalars, by using many conditions |
997 | on the vector's elements. */ |
998 | |
999 | static bool |
1000 | expand_vector_condition (gimple_stmt_iterator *gsi, bitmap dce_ssa_names) |
1001 | { |
1002 | gassign *stmt = as_a <gassign *> (p: gsi_stmt (i: *gsi)); |
1003 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); |
1004 | tree a = gimple_assign_rhs1 (gs: stmt); |
1005 | tree a1 = a; |
1006 | tree a2 = NULL_TREE; |
1007 | bool a_is_comparison = false; |
1008 | bool a_is_scalar_bitmask = false; |
1009 | tree b = gimple_assign_rhs2 (gs: stmt); |
1010 | tree c = gimple_assign_rhs3 (gs: stmt); |
1011 | vec<constructor_elt, va_gc> *v; |
1012 | tree constr; |
1013 | tree inner_type = TREE_TYPE (type); |
1014 | tree width = vector_element_bits_tree (type); |
1015 | tree cond_type = TREE_TYPE (TREE_TYPE (a)); |
1016 | tree comp_inner_type = cond_type; |
1017 | tree index = bitsize_int (0); |
1018 | tree comp_width = width; |
1019 | tree comp_index = index; |
1020 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
1021 | tree_code code = TREE_CODE (a); |
1022 | gassign *assign = NULL; |
1023 | |
1024 | if (code == SSA_NAME) |
1025 | { |
1026 | assign = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (a)); |
1027 | if (assign != NULL |
1028 | && TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison) |
1029 | { |
1030 | a_is_comparison = true; |
1031 | a1 = gimple_assign_rhs1 (gs: assign); |
1032 | a2 = gimple_assign_rhs2 (gs: assign); |
1033 | code = gimple_assign_rhs_code (gs: assign); |
1034 | comp_inner_type = TREE_TYPE (TREE_TYPE (a1)); |
1035 | comp_width = vector_element_bits_tree (TREE_TYPE (a1)); |
1036 | } |
1037 | } |
1038 | |
1039 | if (expand_vec_cond_expr_p (type, TREE_TYPE (a1), code) |
1040 | || (integer_all_onesp (b) && integer_zerop (c) |
1041 | && expand_vec_cmp_expr_p (type, TREE_TYPE (a1), code))) |
1042 | { |
1043 | gcc_assert (TREE_CODE (a) == SSA_NAME || TREE_CODE (a) == VECTOR_CST); |
1044 | return true; |
1045 | } |
1046 | |
1047 | /* If a has vector boolean type and is a comparison, above |
1048 | expand_vec_cond_expr_p might fail, even if both the comparison and |
1049 | VEC_COND_EXPR could be supported individually. See PR109176. */ |
1050 | if (a_is_comparison |
1051 | && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (a)) |
1052 | && expand_vec_cond_expr_p (type, TREE_TYPE (a), SSA_NAME) |
1053 | && expand_vec_cmp_expr_p (TREE_TYPE (a1), TREE_TYPE (a), code)) |
1054 | return true; |
1055 | |
1056 | /* Handle vector boolean types with bitmasks. If there is a comparison |
1057 | and we can expand the comparison into the vector boolean bitmask, |
1058 | or otherwise if it is compatible with type, we can transform |
1059 | vbfld_1 = x_2 < y_3 ? vbfld_4 : vbfld_5; |
1060 | into |
1061 | tmp_6 = x_2 < y_3; |
1062 | tmp_7 = tmp_6 & vbfld_4; |
1063 | tmp_8 = ~tmp_6; |
1064 | tmp_9 = tmp_8 & vbfld_5; |
1065 | vbfld_1 = tmp_7 | tmp_9; |
1066 | Similarly for vbfld_10 instead of x_2 < y_3. */ |
1067 | if (VECTOR_BOOLEAN_TYPE_P (type) |
1068 | && SCALAR_INT_MODE_P (TYPE_MODE (type)) |
1069 | && known_lt (GET_MODE_BITSIZE (TYPE_MODE (type)), |
1070 | TYPE_VECTOR_SUBPARTS (type) |
1071 | * GET_MODE_BITSIZE (SCALAR_TYPE_MODE (TREE_TYPE (type)))) |
1072 | && (a_is_comparison |
1073 | ? useless_type_conversion_p (type, TREE_TYPE (a)) |
1074 | : expand_vec_cmp_expr_p (TREE_TYPE (a1), type, TREE_CODE (a)))) |
1075 | { |
1076 | if (a_is_comparison) |
1077 | a = gimplify_build2 (gsi, code, type, a: a1, b: a2); |
1078 | a1 = gimplify_build2 (gsi, code: BIT_AND_EXPR, type, a, b); |
1079 | a2 = gimplify_build1 (gsi, code: BIT_NOT_EXPR, type, a); |
1080 | a2 = gimplify_build2 (gsi, code: BIT_AND_EXPR, type, a: a2, b: c); |
1081 | a = gimplify_build2 (gsi, code: BIT_IOR_EXPR, type, a: a1, b: a2); |
1082 | gimple_assign_set_rhs_from_tree (gsi, a); |
1083 | update_stmt (s: gsi_stmt (i: *gsi)); |
1084 | return true; |
1085 | } |
1086 | |
1087 | /* TODO: try and find a smaller vector type. */ |
1088 | |
1089 | if (!warning_suppressed_p (stmt, OPT_Wvector_operation_performance)) |
1090 | warning_at (loc, OPT_Wvector_operation_performance, |
1091 | "vector condition will be expanded piecewise" ); |
1092 | |
1093 | if (!a_is_comparison |
1094 | && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (a)) |
1095 | && SCALAR_INT_MODE_P (TYPE_MODE (TREE_TYPE (a))) |
1096 | && known_lt (GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (a))), |
1097 | TYPE_VECTOR_SUBPARTS (TREE_TYPE (a)) |
1098 | * GET_MODE_BITSIZE (SCALAR_TYPE_MODE |
1099 | (TREE_TYPE (TREE_TYPE (a)))))) |
1100 | { |
1101 | a_is_scalar_bitmask = true; |
1102 | int prec = GET_MODE_PRECISION (SCALAR_TYPE_MODE (TREE_TYPE (a))); |
1103 | tree atype = build_nonstandard_integer_type (prec, 1); |
1104 | a = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, type: atype, a); |
1105 | } |
1106 | else if (!a_is_comparison |
1107 | && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (a))) |
1108 | comp_width = vector_element_bits_tree (TREE_TYPE (a)); |
1109 | |
1110 | int nunits = nunits_for_known_piecewise_op (type); |
1111 | vec_alloc (v, nelems: nunits); |
1112 | bool constant_p = true; |
1113 | for (int i = 0; i < nunits; i++) |
1114 | { |
1115 | tree aa, result; |
1116 | tree bb = tree_vec_extract (gsi, type: inner_type, t: b, bitsize: width, bitpos: index); |
1117 | tree cc = tree_vec_extract (gsi, type: inner_type, t: c, bitsize: width, bitpos: index); |
1118 | if (a_is_comparison) |
1119 | { |
1120 | tree aa1 = tree_vec_extract (gsi, type: comp_inner_type, t: a1, |
1121 | bitsize: comp_width, bitpos: comp_index); |
1122 | tree aa2 = tree_vec_extract (gsi, type: comp_inner_type, t: a2, |
1123 | bitsize: comp_width, bitpos: comp_index); |
1124 | aa = gimplify_build2 (gsi, code, boolean_type_node, a: aa1, b: aa2); |
1125 | } |
1126 | else if (a_is_scalar_bitmask) |
1127 | { |
1128 | wide_int w = wi::set_bit_in_zero (bit: i, TYPE_PRECISION (TREE_TYPE (a))); |
1129 | result = gimplify_build2 (gsi, code: BIT_AND_EXPR, TREE_TYPE (a), |
1130 | a, b: wide_int_to_tree (TREE_TYPE (a), cst: w)); |
1131 | aa = gimplify_build2 (gsi, code: NE_EXPR, boolean_type_node, a: result, |
1132 | b: build_zero_cst (TREE_TYPE (a))); |
1133 | } |
1134 | else |
1135 | { |
1136 | result = tree_vec_extract (gsi, type: cond_type, t: a, bitsize: comp_width, bitpos: comp_index); |
1137 | aa = gimplify_build2 (gsi, code: NE_EXPR, boolean_type_node, a: result, |
1138 | b: build_zero_cst (cond_type)); |
1139 | } |
1140 | result = gimplify_build3 (gsi, code: COND_EXPR, type: inner_type, a: aa, b: bb, c: cc); |
1141 | if (!CONSTANT_CLASS_P (result)) |
1142 | constant_p = false; |
1143 | constructor_elt ce = {NULL_TREE, .value: result}; |
1144 | v->quick_push (obj: ce); |
1145 | index = int_const_binop (PLUS_EXPR, index, width); |
1146 | if (width == comp_width) |
1147 | comp_index = index; |
1148 | else |
1149 | comp_index = int_const_binop (PLUS_EXPR, comp_index, comp_width); |
1150 | } |
1151 | |
1152 | if (constant_p) |
1153 | constr = build_vector_from_ctor (type, v); |
1154 | else |
1155 | constr = build_constructor (type, v); |
1156 | gimple_assign_set_rhs_from_tree (gsi, constr); |
1157 | update_stmt (s: gsi_stmt (i: *gsi)); |
1158 | |
1159 | if (a_is_comparison) |
1160 | bitmap_set_bit (dce_ssa_names, |
1161 | SSA_NAME_VERSION (gimple_assign_lhs (assign))); |
1162 | |
1163 | return false; |
1164 | } |
1165 | |
1166 | static tree |
1167 | expand_vector_operation (gimple_stmt_iterator *gsi, tree type, tree compute_type, |
1168 | gassign *assign, enum tree_code code, |
1169 | bitmap dce_ssa_names) |
1170 | { |
1171 | machine_mode compute_mode = TYPE_MODE (compute_type); |
1172 | |
1173 | /* If the compute mode is not a vector mode (hence we are not decomposing |
1174 | a BLKmode vector to smaller, hardware-supported vectors), we may want |
1175 | to expand the operations in parallel. */ |
1176 | if (!VECTOR_MODE_P (compute_mode)) |
1177 | switch (code) |
1178 | { |
1179 | case PLUS_EXPR: |
1180 | case MINUS_EXPR: |
1181 | if (ANY_INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type)) |
1182 | return expand_vector_addition (gsi, f: do_binop, f_parallel: do_plus_minus, type, |
1183 | a: gimple_assign_rhs1 (gs: assign), |
1184 | b: gimple_assign_rhs2 (gs: assign), code); |
1185 | break; |
1186 | |
1187 | case NEGATE_EXPR: |
1188 | if (ANY_INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type)) |
1189 | return expand_vector_addition (gsi, f: do_unop, f_parallel: do_negate, type, |
1190 | a: gimple_assign_rhs1 (gs: assign), |
1191 | NULL_TREE, code); |
1192 | break; |
1193 | |
1194 | case BIT_AND_EXPR: |
1195 | case BIT_IOR_EXPR: |
1196 | case BIT_XOR_EXPR: |
1197 | return expand_vector_parallel (gsi, f: do_binop, type, |
1198 | a: gimple_assign_rhs1 (gs: assign), |
1199 | b: gimple_assign_rhs2 (gs: assign), code); |
1200 | |
1201 | case BIT_NOT_EXPR: |
1202 | return expand_vector_parallel (gsi, f: do_unop, type, |
1203 | a: gimple_assign_rhs1 (gs: assign), |
1204 | NULL_TREE, code); |
1205 | case EQ_EXPR: |
1206 | case NE_EXPR: |
1207 | case GT_EXPR: |
1208 | case LT_EXPR: |
1209 | case GE_EXPR: |
1210 | case LE_EXPR: |
1211 | case UNEQ_EXPR: |
1212 | case UNGT_EXPR: |
1213 | case UNLT_EXPR: |
1214 | case UNGE_EXPR: |
1215 | case UNLE_EXPR: |
1216 | case LTGT_EXPR: |
1217 | case ORDERED_EXPR: |
1218 | case UNORDERED_EXPR: |
1219 | { |
1220 | tree rhs1 = gimple_assign_rhs1 (gs: assign); |
1221 | tree rhs2 = gimple_assign_rhs2 (gs: assign); |
1222 | |
1223 | return expand_vector_comparison (gsi, type, op0: rhs1, op1: rhs2, code, |
1224 | dce_ssa_names); |
1225 | } |
1226 | |
1227 | case TRUNC_DIV_EXPR: |
1228 | case TRUNC_MOD_EXPR: |
1229 | { |
1230 | tree rhs1 = gimple_assign_rhs1 (gs: assign); |
1231 | tree rhs2 = gimple_assign_rhs2 (gs: assign); |
1232 | tree ret; |
1233 | |
1234 | if (!optimize |
1235 | || !VECTOR_INTEGER_TYPE_P (type) |
1236 | || TREE_CODE (rhs2) != VECTOR_CST |
1237 | || !VECTOR_MODE_P (TYPE_MODE (type))) |
1238 | break; |
1239 | |
1240 | ret = expand_vector_divmod (gsi, type, op0: rhs1, op1: rhs2, code); |
1241 | if (ret != NULL_TREE) |
1242 | return ret; |
1243 | break; |
1244 | } |
1245 | |
1246 | default: |
1247 | break; |
1248 | } |
1249 | |
1250 | if (TREE_CODE_CLASS (code) == tcc_unary) |
1251 | return expand_vector_piecewise (gsi, f: do_unop, type, inner_type: compute_type, |
1252 | a: gimple_assign_rhs1 (gs: assign), |
1253 | NULL_TREE, code, parallel_p: false); |
1254 | else |
1255 | return expand_vector_piecewise (gsi, f: do_binop, type, inner_type: compute_type, |
1256 | a: gimple_assign_rhs1 (gs: assign), |
1257 | b: gimple_assign_rhs2 (gs: assign), code, parallel_p: false); |
1258 | } |
1259 | |
1260 | /* Try to optimize |
1261 | a_5 = { b_7, b_7 + 3, b_7 + 6, b_7 + 9 }; |
1262 | style stmts into: |
1263 | _9 = { b_7, b_7, b_7, b_7 }; |
1264 | a_5 = _9 + { 0, 3, 6, 9 }; |
1265 | because vector splat operation is usually more efficient |
1266 | than piecewise initialization of the vector. */ |
1267 | |
1268 | static void |
1269 | optimize_vector_constructor (gimple_stmt_iterator *gsi) |
1270 | { |
1271 | gassign *stmt = as_a <gassign *> (p: gsi_stmt (i: *gsi)); |
1272 | tree lhs = gimple_assign_lhs (gs: stmt); |
1273 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
1274 | tree type = TREE_TYPE (rhs); |
1275 | unsigned int i, j; |
1276 | unsigned HOST_WIDE_INT nelts; |
1277 | bool all_same = true; |
1278 | constructor_elt *elt; |
1279 | gimple *g; |
1280 | tree base = NULL_TREE; |
1281 | optab op; |
1282 | |
1283 | if (!TYPE_VECTOR_SUBPARTS (node: type).is_constant (const_value: &nelts) |
1284 | || nelts <= 2 |
1285 | || CONSTRUCTOR_NELTS (rhs) != nelts) |
1286 | return; |
1287 | op = optab_for_tree_code (PLUS_EXPR, type, optab_default); |
1288 | if (op == unknown_optab |
1289 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing) |
1290 | return; |
1291 | FOR_EACH_VEC_SAFE_ELT (CONSTRUCTOR_ELTS (rhs), i, elt) |
1292 | if (TREE_CODE (elt->value) != SSA_NAME |
1293 | || TREE_CODE (TREE_TYPE (elt->value)) == VECTOR_TYPE) |
1294 | return; |
1295 | else |
1296 | { |
1297 | tree this_base = elt->value; |
1298 | if (this_base != CONSTRUCTOR_ELT (rhs, 0)->value) |
1299 | all_same = false; |
1300 | for (j = 0; j < nelts + 1; j++) |
1301 | { |
1302 | g = SSA_NAME_DEF_STMT (this_base); |
1303 | if (is_gimple_assign (gs: g) |
1304 | && gimple_assign_rhs_code (gs: g) == PLUS_EXPR |
1305 | && TREE_CODE (gimple_assign_rhs2 (g)) == INTEGER_CST |
1306 | && TREE_CODE (gimple_assign_rhs1 (g)) == SSA_NAME |
1307 | && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_assign_rhs1 (g))) |
1308 | this_base = gimple_assign_rhs1 (gs: g); |
1309 | else |
1310 | break; |
1311 | } |
1312 | if (i == 0) |
1313 | base = this_base; |
1314 | else if (this_base != base) |
1315 | return; |
1316 | } |
1317 | if (all_same) |
1318 | return; |
1319 | tree_vector_builder cst (type, nelts, 1); |
1320 | for (i = 0; i < nelts; i++) |
1321 | { |
1322 | tree this_base = CONSTRUCTOR_ELT (rhs, i)->value; |
1323 | tree elt = build_zero_cst (TREE_TYPE (base)); |
1324 | while (this_base != base) |
1325 | { |
1326 | g = SSA_NAME_DEF_STMT (this_base); |
1327 | elt = fold_binary (PLUS_EXPR, TREE_TYPE (base), |
1328 | elt, gimple_assign_rhs2 (g)); |
1329 | if (elt == NULL_TREE |
1330 | || TREE_CODE (elt) != INTEGER_CST |
1331 | || TREE_OVERFLOW (elt)) |
1332 | return; |
1333 | this_base = gimple_assign_rhs1 (gs: g); |
1334 | } |
1335 | cst.quick_push (obj: elt); |
1336 | } |
1337 | for (i = 0; i < nelts; i++) |
1338 | CONSTRUCTOR_ELT (rhs, i)->value = base; |
1339 | g = gimple_build_assign (make_ssa_name (var: type), rhs); |
1340 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
1341 | g = gimple_build_assign (lhs, PLUS_EXPR, gimple_assign_lhs (gs: g), |
1342 | cst.build ()); |
1343 | gsi_replace (gsi, g, false); |
1344 | } |
1345 | |
1346 | /* Return a type for the widest vector mode with the same element type as |
1347 | type ORIGINAL_VECTOR_TYPE, with at most the same number of elements as type |
1348 | ORIGINAL_VECTOR_TYPE and that is supported by the target for an operation |
1349 | with optab OP, or return NULL_TREE if none is found. */ |
1350 | |
1351 | static tree |
1352 | type_for_widest_vector_mode (tree original_vector_type, optab op) |
1353 | { |
1354 | gcc_assert (VECTOR_TYPE_P (original_vector_type)); |
1355 | tree type = TREE_TYPE (original_vector_type); |
1356 | machine_mode inner_mode = TYPE_MODE (type); |
1357 | machine_mode best_mode = VOIDmode, mode; |
1358 | poly_int64 best_nunits = 0; |
1359 | |
1360 | if (SCALAR_FLOAT_MODE_P (inner_mode)) |
1361 | mode = MIN_MODE_VECTOR_FLOAT; |
1362 | else if (SCALAR_FRACT_MODE_P (inner_mode)) |
1363 | mode = MIN_MODE_VECTOR_FRACT; |
1364 | else if (SCALAR_UFRACT_MODE_P (inner_mode)) |
1365 | mode = MIN_MODE_VECTOR_UFRACT; |
1366 | else if (SCALAR_ACCUM_MODE_P (inner_mode)) |
1367 | mode = MIN_MODE_VECTOR_ACCUM; |
1368 | else if (SCALAR_UACCUM_MODE_P (inner_mode)) |
1369 | mode = MIN_MODE_VECTOR_UACCUM; |
1370 | else if (inner_mode == BImode) |
1371 | mode = MIN_MODE_VECTOR_BOOL; |
1372 | else |
1373 | mode = MIN_MODE_VECTOR_INT; |
1374 | |
1375 | FOR_EACH_MODE_FROM (mode, mode) |
1376 | if (GET_MODE_INNER (mode) == inner_mode |
1377 | && maybe_gt (GET_MODE_NUNITS (mode), best_nunits) |
1378 | && optab_handler (op, mode) != CODE_FOR_nothing |
1379 | && known_le (GET_MODE_NUNITS (mode), |
1380 | TYPE_VECTOR_SUBPARTS (original_vector_type))) |
1381 | best_mode = mode, best_nunits = GET_MODE_NUNITS (mode); |
1382 | |
1383 | if (best_mode == VOIDmode) |
1384 | return NULL_TREE; |
1385 | else |
1386 | return build_vector_type_for_mode (type, best_mode); |
1387 | } |
1388 | |
1389 | |
1390 | /* Build a reference to the element of the vector VECT. Function |
1391 | returns either the element itself, either BIT_FIELD_REF, or an |
1392 | ARRAY_REF expression. |
1393 | |
1394 | GSI is required to insert temporary variables while building a |
1395 | refernece to the element of the vector VECT. |
1396 | |
1397 | PTMPVEC is a pointer to the temporary variable for caching |
1398 | purposes. In case when PTMPVEC is NULL new temporary variable |
1399 | will be created. */ |
1400 | static tree |
1401 | vector_element (gimple_stmt_iterator *gsi, tree vect, tree idx, tree *ptmpvec) |
1402 | { |
1403 | tree vect_type, vect_elt_type; |
1404 | gimple *asgn; |
1405 | tree tmpvec; |
1406 | tree arraytype; |
1407 | bool need_asgn = true; |
1408 | unsigned int elements; |
1409 | |
1410 | vect_type = TREE_TYPE (vect); |
1411 | vect_elt_type = TREE_TYPE (vect_type); |
1412 | elements = nunits_for_known_piecewise_op (type: vect_type); |
1413 | |
1414 | if (TREE_CODE (idx) == INTEGER_CST) |
1415 | { |
1416 | unsigned HOST_WIDE_INT index; |
1417 | |
1418 | /* Given that we're about to compute a binary modulus, |
1419 | we don't care about the high bits of the value. */ |
1420 | index = TREE_INT_CST_LOW (idx); |
1421 | if (!tree_fits_uhwi_p (idx) || index >= elements) |
1422 | { |
1423 | index &= elements - 1; |
1424 | idx = build_int_cst (TREE_TYPE (idx), index); |
1425 | } |
1426 | |
1427 | /* When lowering a vector statement sequence do some easy |
1428 | simplification by looking through intermediate vector results. */ |
1429 | if (TREE_CODE (vect) == SSA_NAME) |
1430 | { |
1431 | gimple *def_stmt = SSA_NAME_DEF_STMT (vect); |
1432 | if (is_gimple_assign (gs: def_stmt) |
1433 | && (gimple_assign_rhs_code (gs: def_stmt) == VECTOR_CST |
1434 | || gimple_assign_rhs_code (gs: def_stmt) == CONSTRUCTOR)) |
1435 | vect = gimple_assign_rhs1 (gs: def_stmt); |
1436 | } |
1437 | |
1438 | if (TREE_CODE (vect) == VECTOR_CST) |
1439 | return VECTOR_CST_ELT (vect, index); |
1440 | else if (TREE_CODE (vect) == CONSTRUCTOR |
1441 | && (CONSTRUCTOR_NELTS (vect) == 0 |
1442 | || TREE_CODE (TREE_TYPE (CONSTRUCTOR_ELT (vect, 0)->value)) |
1443 | != VECTOR_TYPE)) |
1444 | { |
1445 | if (index < CONSTRUCTOR_NELTS (vect)) |
1446 | return CONSTRUCTOR_ELT (vect, index)->value; |
1447 | return build_zero_cst (vect_elt_type); |
1448 | } |
1449 | else |
1450 | { |
1451 | tree size = vector_element_bits_tree (vect_type); |
1452 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, bitsize_int (index), |
1453 | size); |
1454 | return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos); |
1455 | } |
1456 | } |
1457 | |
1458 | if (!ptmpvec) |
1459 | tmpvec = create_tmp_var (vect_type, "vectmp" ); |
1460 | else if (!*ptmpvec) |
1461 | tmpvec = *ptmpvec = create_tmp_var (vect_type, "vectmp" ); |
1462 | else |
1463 | { |
1464 | tmpvec = *ptmpvec; |
1465 | need_asgn = false; |
1466 | } |
1467 | |
1468 | if (need_asgn) |
1469 | { |
1470 | TREE_ADDRESSABLE (tmpvec) = 1; |
1471 | asgn = gimple_build_assign (tmpvec, vect); |
1472 | gsi_insert_before (gsi, asgn, GSI_SAME_STMT); |
1473 | } |
1474 | |
1475 | arraytype = build_array_type_nelts (vect_elt_type, elements); |
1476 | return build4 (ARRAY_REF, vect_elt_type, |
1477 | build1 (VIEW_CONVERT_EXPR, arraytype, tmpvec), |
1478 | idx, NULL_TREE, NULL_TREE); |
1479 | } |
1480 | |
1481 | /* Check if VEC_PERM_EXPR within the given setting is supported |
1482 | by hardware, or lower it piecewise. |
1483 | |
1484 | When VEC_PERM_EXPR has the same first and second operands: |
1485 | VEC_PERM_EXPR <v0, v0, mask> the lowered version would be |
1486 | {v0[mask[0]], v0[mask[1]], ...} |
1487 | MASK and V0 must have the same number of elements. |
1488 | |
1489 | Otherwise VEC_PERM_EXPR <v0, v1, mask> is lowered to |
1490 | {mask[0] < len(v0) ? v0[mask[0]] : v1[mask[0]], ...} |
1491 | V0 and V1 must have the same type. MASK, V0, V1 must have the |
1492 | same number of arguments. */ |
1493 | |
1494 | static void |
1495 | lower_vec_perm (gimple_stmt_iterator *gsi) |
1496 | { |
1497 | gassign *stmt = as_a <gassign *> (p: gsi_stmt (i: *gsi)); |
1498 | tree mask = gimple_assign_rhs3 (gs: stmt); |
1499 | tree vec0 = gimple_assign_rhs1 (gs: stmt); |
1500 | tree vec1 = gimple_assign_rhs2 (gs: stmt); |
1501 | tree vect_type = TREE_TYPE (vec0); |
1502 | tree mask_type = TREE_TYPE (mask); |
1503 | tree vect_elt_type = TREE_TYPE (vect_type); |
1504 | tree mask_elt_type = TREE_TYPE (mask_type); |
1505 | unsigned HOST_WIDE_INT elements; |
1506 | vec<constructor_elt, va_gc> *v; |
1507 | tree constr, t, si, i_val; |
1508 | tree vec0tmp = NULL_TREE, vec1tmp = NULL_TREE, masktmp = NULL_TREE; |
1509 | bool two_operand_p = !operand_equal_p (vec0, vec1, flags: 0); |
1510 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
1511 | unsigned i; |
1512 | |
1513 | if (!TYPE_VECTOR_SUBPARTS (node: vect_type).is_constant (const_value: &elements)) |
1514 | return; |
1515 | |
1516 | if (TREE_CODE (mask) == SSA_NAME) |
1517 | { |
1518 | gimple *def_stmt = SSA_NAME_DEF_STMT (mask); |
1519 | if (is_gimple_assign (gs: def_stmt) |
1520 | && gimple_assign_rhs_code (gs: def_stmt) == VECTOR_CST) |
1521 | mask = gimple_assign_rhs1 (gs: def_stmt); |
1522 | } |
1523 | |
1524 | vec_perm_builder sel_int; |
1525 | |
1526 | if (TREE_CODE (mask) == VECTOR_CST |
1527 | && tree_to_vec_perm_builder (&sel_int, mask)) |
1528 | { |
1529 | vec_perm_indices indices (sel_int, 2, elements); |
1530 | machine_mode vmode = TYPE_MODE (vect_type); |
1531 | tree lhs_type = TREE_TYPE (gimple_assign_lhs (stmt)); |
1532 | machine_mode lhs_mode = TYPE_MODE (lhs_type); |
1533 | if (can_vec_perm_const_p (lhs_mode, vmode, indices)) |
1534 | { |
1535 | gimple_assign_set_rhs3 (gs: stmt, rhs: mask); |
1536 | update_stmt (s: stmt); |
1537 | return; |
1538 | } |
1539 | /* Also detect vec_shr pattern - VEC_PERM_EXPR with zero |
1540 | vector as VEC1 and a right element shift MASK. */ |
1541 | if (optab_handler (op: vec_shr_optab, TYPE_MODE (vect_type)) |
1542 | != CODE_FOR_nothing |
1543 | && TREE_CODE (vec1) == VECTOR_CST |
1544 | && initializer_zerop (vec1) |
1545 | && maybe_ne (a: indices[0], b: 0) |
1546 | && known_lt (poly_uint64 (indices[0]), elements)) |
1547 | { |
1548 | bool ok_p = indices.series_p (0, 1, indices[0], 1); |
1549 | if (!ok_p) |
1550 | { |
1551 | for (i = 1; i < elements; ++i) |
1552 | { |
1553 | poly_uint64 actual = indices[i]; |
1554 | poly_uint64 expected = i + indices[0]; |
1555 | /* Indices into the second vector are all equivalent. */ |
1556 | if (maybe_lt (a: actual, b: elements) |
1557 | ? maybe_ne (a: actual, b: expected) |
1558 | : maybe_lt (a: expected, b: elements)) |
1559 | break; |
1560 | } |
1561 | ok_p = i == elements; |
1562 | } |
1563 | if (ok_p) |
1564 | { |
1565 | gimple_assign_set_rhs3 (gs: stmt, rhs: mask); |
1566 | update_stmt (s: stmt); |
1567 | return; |
1568 | } |
1569 | } |
1570 | /* And similarly vec_shl pattern. */ |
1571 | if (optab_handler (op: vec_shl_optab, TYPE_MODE (vect_type)) |
1572 | != CODE_FOR_nothing |
1573 | && TREE_CODE (vec0) == VECTOR_CST |
1574 | && initializer_zerop (vec0)) |
1575 | { |
1576 | unsigned int first = 0; |
1577 | for (i = 0; i < elements; ++i) |
1578 | if (known_eq (poly_uint64 (indices[i]), elements)) |
1579 | { |
1580 | if (i == 0 || first) |
1581 | break; |
1582 | first = i; |
1583 | } |
1584 | else if (first |
1585 | ? maybe_ne (a: poly_uint64 (indices[i]), |
1586 | b: elements + i - first) |
1587 | : maybe_ge (poly_uint64 (indices[i]), elements)) |
1588 | break; |
1589 | if (first && i == elements) |
1590 | { |
1591 | gimple_assign_set_rhs3 (gs: stmt, rhs: mask); |
1592 | update_stmt (s: stmt); |
1593 | return; |
1594 | } |
1595 | } |
1596 | } |
1597 | else if (can_vec_perm_var_p (TYPE_MODE (vect_type))) |
1598 | return; |
1599 | |
1600 | if (!warning_suppressed_p (stmt, OPT_Wvector_operation_performance)) |
1601 | warning_at (loc, OPT_Wvector_operation_performance, |
1602 | "vector shuffling operation will be expanded piecewise" ); |
1603 | |
1604 | vec_alloc (v, nelems: elements); |
1605 | bool constant_p = true; |
1606 | for (i = 0; i < elements; i++) |
1607 | { |
1608 | si = size_int (i); |
1609 | i_val = vector_element (gsi, vect: mask, idx: si, ptmpvec: &masktmp); |
1610 | |
1611 | if (TREE_CODE (i_val) == INTEGER_CST) |
1612 | { |
1613 | unsigned HOST_WIDE_INT index; |
1614 | |
1615 | index = TREE_INT_CST_LOW (i_val); |
1616 | if (!tree_fits_uhwi_p (i_val) || index >= elements) |
1617 | i_val = build_int_cst (mask_elt_type, index & (elements - 1)); |
1618 | |
1619 | if (two_operand_p && (index & elements) != 0) |
1620 | t = vector_element (gsi, vect: vec1, idx: i_val, ptmpvec: &vec1tmp); |
1621 | else |
1622 | t = vector_element (gsi, vect: vec0, idx: i_val, ptmpvec: &vec0tmp); |
1623 | |
1624 | t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, |
1625 | true, GSI_SAME_STMT); |
1626 | } |
1627 | else |
1628 | { |
1629 | tree cond = NULL_TREE, v0_val; |
1630 | |
1631 | if (two_operand_p) |
1632 | { |
1633 | cond = fold_build2 (BIT_AND_EXPR, mask_elt_type, i_val, |
1634 | build_int_cst (mask_elt_type, elements)); |
1635 | cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, |
1636 | true, GSI_SAME_STMT); |
1637 | } |
1638 | |
1639 | i_val = fold_build2 (BIT_AND_EXPR, mask_elt_type, i_val, |
1640 | build_int_cst (mask_elt_type, elements - 1)); |
1641 | i_val = force_gimple_operand_gsi (gsi, i_val, true, NULL_TREE, |
1642 | true, GSI_SAME_STMT); |
1643 | |
1644 | v0_val = vector_element (gsi, vect: vec0, idx: i_val, ptmpvec: &vec0tmp); |
1645 | v0_val = force_gimple_operand_gsi (gsi, v0_val, true, NULL_TREE, |
1646 | true, GSI_SAME_STMT); |
1647 | |
1648 | if (two_operand_p) |
1649 | { |
1650 | tree v1_val; |
1651 | |
1652 | v1_val = vector_element (gsi, vect: vec1, idx: i_val, ptmpvec: &vec1tmp); |
1653 | v1_val = force_gimple_operand_gsi (gsi, v1_val, true, NULL_TREE, |
1654 | true, GSI_SAME_STMT); |
1655 | |
1656 | cond = fold_build2 (EQ_EXPR, boolean_type_node, |
1657 | cond, build_zero_cst (mask_elt_type)); |
1658 | cond = fold_build3 (COND_EXPR, vect_elt_type, |
1659 | cond, v0_val, v1_val); |
1660 | t = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE, |
1661 | true, GSI_SAME_STMT); |
1662 | } |
1663 | else |
1664 | t = v0_val; |
1665 | } |
1666 | |
1667 | if (!CONSTANT_CLASS_P (t)) |
1668 | constant_p = false; |
1669 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, t); |
1670 | } |
1671 | |
1672 | if (constant_p) |
1673 | constr = build_vector_from_ctor (vect_type, v); |
1674 | else |
1675 | constr = build_constructor (vect_type, v); |
1676 | gimple_assign_set_rhs_from_tree (gsi, constr); |
1677 | update_stmt (s: gsi_stmt (i: *gsi)); |
1678 | } |
1679 | |
1680 | /* If OP is a uniform vector return the element it is a splat from. */ |
1681 | |
1682 | static tree |
1683 | ssa_uniform_vector_p (tree op) |
1684 | { |
1685 | if (TREE_CODE (op) == VECTOR_CST |
1686 | || TREE_CODE (op) == VEC_DUPLICATE_EXPR |
1687 | || TREE_CODE (op) == CONSTRUCTOR) |
1688 | return uniform_vector_p (op); |
1689 | if (TREE_CODE (op) == SSA_NAME) |
1690 | { |
1691 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
1692 | if (gimple_assign_single_p (gs: def_stmt)) |
1693 | return uniform_vector_p (gimple_assign_rhs1 (gs: def_stmt)); |
1694 | } |
1695 | return NULL_TREE; |
1696 | } |
1697 | |
1698 | /* Return type in which CODE operation with optab OP can be |
1699 | computed. */ |
1700 | |
1701 | static tree |
1702 | get_compute_type (enum tree_code code, optab op, tree type) |
1703 | { |
1704 | /* For very wide vectors, try using a smaller vector mode. */ |
1705 | tree compute_type = type; |
1706 | if (op |
1707 | && (!VECTOR_MODE_P (TYPE_MODE (type)) |
1708 | || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)) |
1709 | { |
1710 | tree vector_compute_type |
1711 | = type_for_widest_vector_mode (original_vector_type: type, op); |
1712 | if (vector_compute_type != NULL_TREE |
1713 | && maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vector_compute_type), b: 1U) |
1714 | && (optab_handler (op, TYPE_MODE (vector_compute_type)) |
1715 | != CODE_FOR_nothing)) |
1716 | compute_type = vector_compute_type; |
1717 | } |
1718 | |
1719 | /* If we are breaking a BLKmode vector into smaller pieces, |
1720 | type_for_widest_vector_mode has already looked into the optab, |
1721 | so skip these checks. */ |
1722 | if (compute_type == type) |
1723 | { |
1724 | machine_mode compute_mode = TYPE_MODE (compute_type); |
1725 | if (VECTOR_MODE_P (compute_mode)) |
1726 | { |
1727 | if (op |
1728 | && (optab_handler (op, mode: compute_mode) != CODE_FOR_nothing |
1729 | || optab_libfunc (op, compute_mode))) |
1730 | return compute_type; |
1731 | if (code == MULT_HIGHPART_EXPR |
1732 | && can_mult_highpart_p (compute_mode, |
1733 | TYPE_UNSIGNED (compute_type))) |
1734 | return compute_type; |
1735 | } |
1736 | /* There is no operation in hardware, so fall back to scalars. */ |
1737 | compute_type = TREE_TYPE (type); |
1738 | } |
1739 | |
1740 | return compute_type; |
1741 | } |
1742 | |
1743 | static tree |
1744 | do_cond (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b, |
1745 | tree bitpos, tree bitsize, enum tree_code code, |
1746 | tree type ATTRIBUTE_UNUSED) |
1747 | { |
1748 | if (VECTOR_TYPE_P (TREE_TYPE (a))) |
1749 | a = tree_vec_extract (gsi, type: inner_type, t: a, bitsize, bitpos); |
1750 | if (VECTOR_TYPE_P (TREE_TYPE (b))) |
1751 | b = tree_vec_extract (gsi, type: inner_type, t: b, bitsize, bitpos); |
1752 | tree cond = gimple_assign_rhs1 (gs: gsi_stmt (i: *gsi)); |
1753 | return gimplify_build3 (gsi, code, type: inner_type, a: unshare_expr (cond), b: a, c: b); |
1754 | } |
1755 | |
1756 | /* Expand a vector COND_EXPR to scalars, piecewise. */ |
1757 | static void |
1758 | expand_vector_scalar_condition (gimple_stmt_iterator *gsi) |
1759 | { |
1760 | gassign *stmt = as_a <gassign *> (p: gsi_stmt (i: *gsi)); |
1761 | tree lhs = gimple_assign_lhs (gs: stmt); |
1762 | tree type = TREE_TYPE (lhs); |
1763 | tree compute_type = get_compute_type (code: COND_EXPR, op: mov_optab, type); |
1764 | machine_mode compute_mode = TYPE_MODE (compute_type); |
1765 | gcc_assert (compute_mode != BLKmode); |
1766 | tree rhs2 = gimple_assign_rhs2 (gs: stmt); |
1767 | tree rhs3 = gimple_assign_rhs3 (gs: stmt); |
1768 | tree new_rhs; |
1769 | |
1770 | /* If the compute mode is not a vector mode (hence we are not decomposing |
1771 | a BLKmode vector to smaller, hardware-supported vectors), we may want |
1772 | to expand the operations in parallel. */ |
1773 | if (!VECTOR_MODE_P (compute_mode)) |
1774 | new_rhs = expand_vector_parallel (gsi, f: do_cond, type, a: rhs2, b: rhs3, |
1775 | code: COND_EXPR); |
1776 | else |
1777 | new_rhs = expand_vector_piecewise (gsi, f: do_cond, type, inner_type: compute_type, |
1778 | a: rhs2, b: rhs3, code: COND_EXPR, parallel_p: false); |
1779 | if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (new_rhs))) |
1780 | new_rhs = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, TREE_TYPE (lhs), |
1781 | a: new_rhs); |
1782 | |
1783 | /* NOTE: We should avoid using gimple_assign_set_rhs_from_tree. One |
1784 | way to do it is change expand_vector_operation and its callees to |
1785 | return a tree_code, RHS1 and RHS2 instead of a tree. */ |
1786 | gimple_assign_set_rhs_from_tree (gsi, new_rhs); |
1787 | update_stmt (s: gsi_stmt (i: *gsi)); |
1788 | } |
1789 | |
1790 | /* Callback for expand_vector_piecewise to do VEC_CONVERT ifn call |
1791 | lowering. If INNER_TYPE is not a vector type, this is a scalar |
1792 | fallback. */ |
1793 | |
1794 | static tree |
1795 | do_vec_conversion (gimple_stmt_iterator *gsi, tree inner_type, tree a, |
1796 | tree decl, tree bitpos, tree bitsize, |
1797 | enum tree_code code, tree type) |
1798 | { |
1799 | a = tree_vec_extract (gsi, type: inner_type, t: a, bitsize, bitpos); |
1800 | if (!VECTOR_TYPE_P (inner_type)) |
1801 | return gimplify_build1 (gsi, code, TREE_TYPE (type), a); |
1802 | if (code == CALL_EXPR) |
1803 | { |
1804 | gimple *g = gimple_build_call (decl, 1, a); |
1805 | tree lhs = make_ssa_name (TREE_TYPE (TREE_TYPE (decl))); |
1806 | gimple_call_set_lhs (gs: g, lhs); |
1807 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
1808 | return lhs; |
1809 | } |
1810 | else |
1811 | { |
1812 | tree outer_type = build_vector_type (TREE_TYPE (type), |
1813 | TYPE_VECTOR_SUBPARTS (node: inner_type)); |
1814 | return gimplify_build1 (gsi, code, type: outer_type, a); |
1815 | } |
1816 | } |
1817 | |
1818 | /* Similarly, but for narrowing conversion. */ |
1819 | |
1820 | static tree |
1821 | do_vec_narrow_conversion (gimple_stmt_iterator *gsi, tree inner_type, tree a, |
1822 | tree, tree bitpos, tree, enum tree_code code, |
1823 | tree type) |
1824 | { |
1825 | tree itype = build_vector_type (TREE_TYPE (inner_type), |
1826 | exact_div (a: TYPE_VECTOR_SUBPARTS (node: inner_type), |
1827 | b: 2)); |
1828 | tree b = tree_vec_extract (gsi, type: itype, t: a, TYPE_SIZE (itype), bitpos); |
1829 | tree c = tree_vec_extract (gsi, type: itype, t: a, TYPE_SIZE (itype), |
1830 | bitpos: int_const_binop (PLUS_EXPR, bitpos, |
1831 | TYPE_SIZE (itype))); |
1832 | tree outer_type = build_vector_type (TREE_TYPE (type), |
1833 | TYPE_VECTOR_SUBPARTS (node: inner_type)); |
1834 | return gimplify_build2 (gsi, code, type: outer_type, a: b, b: c); |
1835 | } |
1836 | |
1837 | /* Expand VEC_CONVERT ifn call. */ |
1838 | |
1839 | static void |
1840 | expand_vector_conversion (gimple_stmt_iterator *gsi) |
1841 | { |
1842 | gimple *stmt = gsi_stmt (i: *gsi); |
1843 | gimple *g; |
1844 | tree lhs = gimple_call_lhs (gs: stmt); |
1845 | if (lhs == NULL_TREE) |
1846 | { |
1847 | g = gimple_build_nop (); |
1848 | gsi_replace (gsi, g, false); |
1849 | return; |
1850 | } |
1851 | tree arg = gimple_call_arg (gs: stmt, index: 0); |
1852 | tree ret_type = TREE_TYPE (lhs); |
1853 | tree arg_type = TREE_TYPE (arg); |
1854 | tree new_rhs, compute_type = TREE_TYPE (arg_type); |
1855 | enum tree_code code = NOP_EXPR; |
1856 | enum tree_code code1 = ERROR_MARK; |
1857 | enum { NARROW, NONE, WIDEN } modifier = NONE; |
1858 | optab optab1 = unknown_optab; |
1859 | |
1860 | gcc_checking_assert (VECTOR_TYPE_P (ret_type) && VECTOR_TYPE_P (arg_type)); |
1861 | if (INTEGRAL_TYPE_P (TREE_TYPE (ret_type)) |
1862 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (arg_type))) |
1863 | code = FIX_TRUNC_EXPR; |
1864 | else if (INTEGRAL_TYPE_P (TREE_TYPE (arg_type)) |
1865 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (ret_type))) |
1866 | code = FLOAT_EXPR; |
1867 | unsigned int ret_elt_bits = vector_element_bits (ret_type); |
1868 | unsigned int arg_elt_bits = vector_element_bits (arg_type); |
1869 | if (ret_elt_bits < arg_elt_bits) |
1870 | modifier = NARROW; |
1871 | else if (ret_elt_bits > arg_elt_bits) |
1872 | modifier = WIDEN; |
1873 | |
1874 | if (modifier == NONE && (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR)) |
1875 | { |
1876 | if (supportable_convert_operation (code, ret_type, arg_type, &code1)) |
1877 | { |
1878 | g = gimple_build_assign (lhs, code1, arg); |
1879 | gsi_replace (gsi, g, false); |
1880 | return; |
1881 | } |
1882 | /* Can't use get_compute_type here, as supportable_convert_operation |
1883 | doesn't necessarily use an optab and needs two arguments. */ |
1884 | tree vec_compute_type |
1885 | = type_for_widest_vector_mode (original_vector_type: arg_type, op: mov_optab); |
1886 | if (vec_compute_type |
1887 | && VECTOR_MODE_P (TYPE_MODE (vec_compute_type))) |
1888 | { |
1889 | unsigned HOST_WIDE_INT nelts |
1890 | = constant_lower_bound (a: TYPE_VECTOR_SUBPARTS (node: vec_compute_type)); |
1891 | while (nelts > 1) |
1892 | { |
1893 | tree ret1_type = build_vector_type (TREE_TYPE (ret_type), nelts); |
1894 | tree arg1_type = build_vector_type (TREE_TYPE (arg_type), nelts); |
1895 | if (supportable_convert_operation (code, ret1_type, arg1_type, |
1896 | &code1)) |
1897 | { |
1898 | new_rhs = expand_vector_piecewise (gsi, f: do_vec_conversion, |
1899 | type: ret_type, inner_type: arg1_type, a: arg, |
1900 | NULL_TREE, code: code1, parallel_p: false); |
1901 | g = gimple_build_assign (lhs, new_rhs); |
1902 | gsi_replace (gsi, g, false); |
1903 | return; |
1904 | } |
1905 | nelts = nelts / 2; |
1906 | } |
1907 | } |
1908 | } |
1909 | else if (modifier == NARROW) |
1910 | { |
1911 | switch (code) |
1912 | { |
1913 | CASE_CONVERT: |
1914 | code1 = VEC_PACK_TRUNC_EXPR; |
1915 | optab1 = optab_for_tree_code (code1, arg_type, optab_default); |
1916 | break; |
1917 | case FIX_TRUNC_EXPR: |
1918 | code1 = VEC_PACK_FIX_TRUNC_EXPR; |
1919 | /* The signedness is determined from output operand. */ |
1920 | optab1 = optab_for_tree_code (code1, ret_type, optab_default); |
1921 | break; |
1922 | case FLOAT_EXPR: |
1923 | code1 = VEC_PACK_FLOAT_EXPR; |
1924 | optab1 = optab_for_tree_code (code1, arg_type, optab_default); |
1925 | break; |
1926 | default: |
1927 | gcc_unreachable (); |
1928 | } |
1929 | |
1930 | if (optab1) |
1931 | compute_type = get_compute_type (code: code1, op: optab1, type: arg_type); |
1932 | enum insn_code icode1; |
1933 | if (VECTOR_TYPE_P (compute_type) |
1934 | && ((icode1 = optab_handler (op: optab1, TYPE_MODE (compute_type))) |
1935 | != CODE_FOR_nothing) |
1936 | && VECTOR_MODE_P (insn_data[icode1].operand[0].mode)) |
1937 | { |
1938 | tree cretd_type |
1939 | = build_vector_type (TREE_TYPE (ret_type), |
1940 | TYPE_VECTOR_SUBPARTS (node: compute_type) * 2); |
1941 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (cretd_type)) |
1942 | { |
1943 | if (compute_type == arg_type) |
1944 | { |
1945 | new_rhs = gimplify_build2 (gsi, code: code1, type: cretd_type, |
1946 | a: arg, b: build_zero_cst (arg_type)); |
1947 | new_rhs = tree_vec_extract (gsi, type: ret_type, t: new_rhs, |
1948 | TYPE_SIZE (ret_type), |
1949 | bitsize_int (0)); |
1950 | g = gimple_build_assign (lhs, new_rhs); |
1951 | gsi_replace (gsi, g, false); |
1952 | return; |
1953 | } |
1954 | tree dcompute_type |
1955 | = build_vector_type (TREE_TYPE (compute_type), |
1956 | TYPE_VECTOR_SUBPARTS (node: compute_type) * 2); |
1957 | if (TYPE_MAIN_VARIANT (dcompute_type) |
1958 | == TYPE_MAIN_VARIANT (arg_type)) |
1959 | new_rhs = do_vec_narrow_conversion (gsi, inner_type: dcompute_type, a: arg, |
1960 | NULL_TREE, bitsize_int (0), |
1961 | NULL_TREE, code: code1, |
1962 | type: ret_type); |
1963 | else |
1964 | new_rhs = expand_vector_piecewise (gsi, |
1965 | f: do_vec_narrow_conversion, |
1966 | type: arg_type, inner_type: dcompute_type, |
1967 | a: arg, NULL_TREE, code: code1, |
1968 | parallel_p: false, ret_type); |
1969 | g = gimple_build_assign (lhs, new_rhs); |
1970 | gsi_replace (gsi, g, false); |
1971 | return; |
1972 | } |
1973 | } |
1974 | } |
1975 | else if (modifier == WIDEN) |
1976 | { |
1977 | enum tree_code code2 = ERROR_MARK; |
1978 | optab optab2 = unknown_optab; |
1979 | switch (code) |
1980 | { |
1981 | CASE_CONVERT: |
1982 | code1 = VEC_UNPACK_LO_EXPR; |
1983 | code2 = VEC_UNPACK_HI_EXPR; |
1984 | break; |
1985 | case FIX_TRUNC_EXPR: |
1986 | code1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
1987 | code2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; |
1988 | break; |
1989 | case FLOAT_EXPR: |
1990 | code1 = VEC_UNPACK_FLOAT_LO_EXPR; |
1991 | code2 = VEC_UNPACK_FLOAT_HI_EXPR; |
1992 | break; |
1993 | default: |
1994 | gcc_unreachable (); |
1995 | } |
1996 | if (BYTES_BIG_ENDIAN) |
1997 | std::swap (a&: code1, b&: code2); |
1998 | |
1999 | if (code == FIX_TRUNC_EXPR) |
2000 | { |
2001 | /* The signedness is determined from output operand. */ |
2002 | optab1 = optab_for_tree_code (code1, ret_type, optab_default); |
2003 | optab2 = optab_for_tree_code (code2, ret_type, optab_default); |
2004 | } |
2005 | else |
2006 | { |
2007 | optab1 = optab_for_tree_code (code1, arg_type, optab_default); |
2008 | optab2 = optab_for_tree_code (code2, arg_type, optab_default); |
2009 | } |
2010 | |
2011 | if (optab1 && optab2) |
2012 | compute_type = get_compute_type (code: code1, op: optab1, type: arg_type); |
2013 | |
2014 | enum insn_code icode1, icode2; |
2015 | if (VECTOR_TYPE_P (compute_type) |
2016 | && ((icode1 = optab_handler (op: optab1, TYPE_MODE (compute_type))) |
2017 | != CODE_FOR_nothing) |
2018 | && ((icode2 = optab_handler (op: optab2, TYPE_MODE (compute_type))) |
2019 | != CODE_FOR_nothing) |
2020 | && VECTOR_MODE_P (insn_data[icode1].operand[0].mode) |
2021 | && (insn_data[icode1].operand[0].mode |
2022 | == insn_data[icode2].operand[0].mode)) |
2023 | { |
2024 | poly_uint64 nunits |
2025 | = exact_div (a: TYPE_VECTOR_SUBPARTS (node: compute_type), b: 2); |
2026 | tree cretd_type = build_vector_type (TREE_TYPE (ret_type), nunits); |
2027 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (cretd_type)) |
2028 | { |
2029 | vec<constructor_elt, va_gc> *v; |
2030 | tree part_width = TYPE_SIZE (compute_type); |
2031 | tree index = bitsize_int (0); |
2032 | int nunits = nunits_for_known_piecewise_op (type: arg_type); |
2033 | int delta = tree_to_uhwi (part_width) / arg_elt_bits; |
2034 | int i; |
2035 | location_t loc = gimple_location (g: gsi_stmt (i: *gsi)); |
2036 | |
2037 | if (compute_type != arg_type) |
2038 | { |
2039 | if (!warning_suppressed_p (gsi_stmt (i: *gsi), |
2040 | OPT_Wvector_operation_performance)) |
2041 | warning_at (loc, OPT_Wvector_operation_performance, |
2042 | "vector operation will be expanded piecewise" ); |
2043 | } |
2044 | else |
2045 | { |
2046 | nunits = 1; |
2047 | delta = 1; |
2048 | } |
2049 | |
2050 | vec_alloc (v, nelems: (nunits + delta - 1) / delta * 2); |
2051 | bool constant_p = true; |
2052 | for (i = 0; i < nunits; |
2053 | i += delta, index = int_const_binop (PLUS_EXPR, index, |
2054 | part_width)) |
2055 | { |
2056 | tree a = arg; |
2057 | if (compute_type != arg_type) |
2058 | a = tree_vec_extract (gsi, type: compute_type, t: a, bitsize: part_width, |
2059 | bitpos: index); |
2060 | tree result = gimplify_build1 (gsi, code: code1, type: cretd_type, a); |
2061 | constructor_elt ce = { NULL_TREE, .value: result }; |
2062 | if (!CONSTANT_CLASS_P (ce.value)) |
2063 | constant_p = false; |
2064 | v->quick_push (obj: ce); |
2065 | ce.value = gimplify_build1 (gsi, code: code2, type: cretd_type, a); |
2066 | if (!CONSTANT_CLASS_P (ce.value)) |
2067 | constant_p = false; |
2068 | v->quick_push (obj: ce); |
2069 | } |
2070 | |
2071 | if (constant_p) |
2072 | new_rhs = build_vector_from_ctor (ret_type, v); |
2073 | else |
2074 | new_rhs = build_constructor (ret_type, v); |
2075 | g = gimple_build_assign (lhs, new_rhs); |
2076 | gsi_replace (gsi, g, false); |
2077 | return; |
2078 | } |
2079 | } |
2080 | } |
2081 | |
2082 | new_rhs = expand_vector_piecewise (gsi, f: do_vec_conversion, type: arg_type, |
2083 | TREE_TYPE (arg_type), a: arg, |
2084 | NULL_TREE, code, parallel_p: false, ret_type); |
2085 | g = gimple_build_assign (lhs, new_rhs); |
2086 | gsi_replace (gsi, g, false); |
2087 | } |
2088 | |
2089 | /* Process one statement. If we identify a vector operation, expand it. */ |
2090 | |
2091 | static void |
2092 | expand_vector_operations_1 (gimple_stmt_iterator *gsi, |
2093 | bitmap dce_ssa_names) |
2094 | { |
2095 | tree lhs, rhs1, rhs2 = NULL, type, compute_type = NULL_TREE; |
2096 | enum tree_code code; |
2097 | optab op = unknown_optab; |
2098 | enum gimple_rhs_class rhs_class; |
2099 | tree new_rhs; |
2100 | |
2101 | /* Only consider code == GIMPLE_ASSIGN. */ |
2102 | gassign *stmt = dyn_cast <gassign *> (p: gsi_stmt (i: *gsi)); |
2103 | if (!stmt) |
2104 | { |
2105 | if (gimple_call_internal_p (gs: gsi_stmt (i: *gsi), fn: IFN_VEC_CONVERT)) |
2106 | expand_vector_conversion (gsi); |
2107 | return; |
2108 | } |
2109 | |
2110 | code = gimple_assign_rhs_code (gs: stmt); |
2111 | rhs_class = get_gimple_rhs_class (code); |
2112 | lhs = gimple_assign_lhs (gs: stmt); |
2113 | |
2114 | if (code == VEC_PERM_EXPR) |
2115 | { |
2116 | lower_vec_perm (gsi); |
2117 | return; |
2118 | } |
2119 | |
2120 | if (code == VEC_COND_EXPR) |
2121 | { |
2122 | expand_vector_condition (gsi, dce_ssa_names); |
2123 | return; |
2124 | } |
2125 | |
2126 | if (code == COND_EXPR |
2127 | && TREE_CODE (TREE_TYPE (gimple_assign_lhs (stmt))) == VECTOR_TYPE |
2128 | && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode) |
2129 | { |
2130 | expand_vector_scalar_condition (gsi); |
2131 | return; |
2132 | } |
2133 | |
2134 | if (code == CONSTRUCTOR |
2135 | && TREE_CODE (lhs) == SSA_NAME |
2136 | && VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (lhs))) |
2137 | && !gimple_clobber_p (s: stmt) |
2138 | && optimize) |
2139 | { |
2140 | optimize_vector_constructor (gsi); |
2141 | return; |
2142 | } |
2143 | |
2144 | if (rhs_class != GIMPLE_UNARY_RHS && rhs_class != GIMPLE_BINARY_RHS) |
2145 | return; |
2146 | |
2147 | rhs1 = gimple_assign_rhs1 (gs: stmt); |
2148 | if (rhs_class == GIMPLE_BINARY_RHS) |
2149 | rhs2 = gimple_assign_rhs2 (gs: stmt); |
2150 | |
2151 | type = TREE_TYPE (lhs); |
2152 | if (!VECTOR_TYPE_P (type) |
2153 | || !VECTOR_TYPE_P (TREE_TYPE (rhs1))) |
2154 | return; |
2155 | |
2156 | /* A scalar operation pretending to be a vector one. */ |
2157 | if (VECTOR_BOOLEAN_TYPE_P (type) |
2158 | && !VECTOR_MODE_P (TYPE_MODE (type)) |
2159 | && TYPE_MODE (type) != BLKmode |
2160 | && (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) != tcc_comparison |
2161 | || (VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1)) |
2162 | && !VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (rhs1))) |
2163 | && TYPE_MODE (TREE_TYPE (rhs1)) != BLKmode))) |
2164 | return; |
2165 | |
2166 | /* If the vector operation is operating on all same vector elements |
2167 | implement it with a scalar operation and a splat if the target |
2168 | supports the scalar operation. */ |
2169 | tree srhs1, srhs2 = NULL_TREE; |
2170 | if ((srhs1 = ssa_uniform_vector_p (op: rhs1)) != NULL_TREE |
2171 | && (rhs2 == NULL_TREE |
2172 | || (! VECTOR_TYPE_P (TREE_TYPE (rhs2)) |
2173 | && (srhs2 = rhs2)) |
2174 | || (srhs2 = ssa_uniform_vector_p (op: rhs2)) != NULL_TREE) |
2175 | /* As we query direct optabs restrict to non-convert operations. */ |
2176 | && TYPE_MODE (TREE_TYPE (type)) == TYPE_MODE (TREE_TYPE (srhs1))) |
2177 | { |
2178 | op = optab_for_tree_code (code, TREE_TYPE (type), optab_scalar); |
2179 | if (op >= FIRST_NORM_OPTAB && op <= LAST_NORM_OPTAB |
2180 | && optab_handler (op, TYPE_MODE (TREE_TYPE (type))) != CODE_FOR_nothing) |
2181 | { |
2182 | tree stype = TREE_TYPE (TREE_TYPE (lhs)); |
2183 | tree slhs = (rhs2 != NULL_TREE) |
2184 | ? gimplify_build2 (gsi, code, type: stype, a: srhs1, b: srhs2) |
2185 | : gimplify_build1 (gsi, code, type: stype, a: srhs1); |
2186 | gimple_assign_set_rhs_from_tree (gsi, |
2187 | build_vector_from_val (type, slhs)); |
2188 | update_stmt (s: stmt); |
2189 | return; |
2190 | } |
2191 | } |
2192 | |
2193 | if (CONVERT_EXPR_CODE_P (code) |
2194 | || code == FLOAT_EXPR |
2195 | || code == FIX_TRUNC_EXPR |
2196 | || code == VIEW_CONVERT_EXPR) |
2197 | return; |
2198 | |
2199 | /* The signedness is determined from input argument. */ |
2200 | if (code == VEC_UNPACK_FLOAT_HI_EXPR |
2201 | || code == VEC_UNPACK_FLOAT_LO_EXPR |
2202 | || code == VEC_PACK_FLOAT_EXPR) |
2203 | { |
2204 | /* We do not know how to scalarize those. */ |
2205 | return; |
2206 | } |
2207 | |
2208 | /* For widening/narrowing vector operations, the relevant type is of the |
2209 | arguments, not the widened result. VEC_UNPACK_FLOAT_*_EXPR is |
2210 | calculated in the same way above. */ |
2211 | if (code == WIDEN_SUM_EXPR |
2212 | || code == VEC_WIDEN_MULT_HI_EXPR |
2213 | || code == VEC_WIDEN_MULT_LO_EXPR |
2214 | || code == VEC_WIDEN_MULT_EVEN_EXPR |
2215 | || code == VEC_WIDEN_MULT_ODD_EXPR |
2216 | || code == VEC_UNPACK_HI_EXPR |
2217 | || code == VEC_UNPACK_LO_EXPR |
2218 | || code == VEC_UNPACK_FIX_TRUNC_HI_EXPR |
2219 | || code == VEC_UNPACK_FIX_TRUNC_LO_EXPR |
2220 | || code == VEC_PACK_TRUNC_EXPR |
2221 | || code == VEC_PACK_SAT_EXPR |
2222 | || code == VEC_PACK_FIX_TRUNC_EXPR |
2223 | || code == VEC_WIDEN_LSHIFT_HI_EXPR |
2224 | || code == VEC_WIDEN_LSHIFT_LO_EXPR) |
2225 | { |
2226 | /* We do not know how to scalarize those. */ |
2227 | return; |
2228 | } |
2229 | |
2230 | /* Choose between vector shift/rotate by vector and vector shift/rotate by |
2231 | scalar */ |
2232 | if (code == LSHIFT_EXPR |
2233 | || code == RSHIFT_EXPR |
2234 | || code == LROTATE_EXPR |
2235 | || code == RROTATE_EXPR) |
2236 | { |
2237 | optab opv; |
2238 | |
2239 | /* Check whether we have vector <op> {x,x,x,x} where x |
2240 | could be a scalar variable or a constant. Transform |
2241 | vector <op> {x,x,x,x} ==> vector <op> scalar. */ |
2242 | if (VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2))) |
2243 | { |
2244 | tree first; |
2245 | |
2246 | if ((first = ssa_uniform_vector_p (op: rhs2)) != NULL_TREE) |
2247 | { |
2248 | gimple_assign_set_rhs2 (gs: stmt, rhs: first); |
2249 | update_stmt (s: stmt); |
2250 | rhs2 = first; |
2251 | } |
2252 | } |
2253 | |
2254 | opv = optab_for_tree_code (code, type, optab_vector); |
2255 | if (VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2))) |
2256 | op = opv; |
2257 | else |
2258 | { |
2259 | op = optab_for_tree_code (code, type, optab_scalar); |
2260 | |
2261 | compute_type = get_compute_type (code, op, type); |
2262 | if (compute_type == type) |
2263 | return; |
2264 | /* The rtl expander will expand vector/scalar as vector/vector |
2265 | if necessary. Pick one with wider vector type. */ |
2266 | tree compute_vtype = get_compute_type (code, op: opv, type); |
2267 | if (subparts_gt (type1: compute_vtype, type2: compute_type)) |
2268 | { |
2269 | compute_type = compute_vtype; |
2270 | op = opv; |
2271 | } |
2272 | } |
2273 | |
2274 | if (code == LROTATE_EXPR || code == RROTATE_EXPR) |
2275 | { |
2276 | if (compute_type == NULL_TREE) |
2277 | compute_type = get_compute_type (code, op, type); |
2278 | if (compute_type == type) |
2279 | return; |
2280 | /* Before splitting vector rotates into scalar rotates, |
2281 | see if we can't use vector shifts and BIT_IOR_EXPR |
2282 | instead. For vector by vector rotates we'd also |
2283 | need to check BIT_AND_EXPR and NEGATE_EXPR, punt there |
2284 | for now, fold doesn't seem to create such rotates anyway. */ |
2285 | if (compute_type == TREE_TYPE (type) |
2286 | && !VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2))) |
2287 | { |
2288 | optab oplv = vashl_optab, opl = ashl_optab; |
2289 | optab oprv = vlshr_optab, opr = lshr_optab, opo = ior_optab; |
2290 | tree compute_lvtype = get_compute_type (code: LSHIFT_EXPR, op: oplv, type); |
2291 | tree compute_rvtype = get_compute_type (code: RSHIFT_EXPR, op: oprv, type); |
2292 | tree compute_otype = get_compute_type (code: BIT_IOR_EXPR, op: opo, type); |
2293 | tree compute_ltype = get_compute_type (code: LSHIFT_EXPR, op: opl, type); |
2294 | tree compute_rtype = get_compute_type (code: RSHIFT_EXPR, op: opr, type); |
2295 | /* The rtl expander will expand vector/scalar as vector/vector |
2296 | if necessary. Pick one with wider vector type. */ |
2297 | if (subparts_gt (type1: compute_lvtype, type2: compute_ltype)) |
2298 | { |
2299 | compute_ltype = compute_lvtype; |
2300 | opl = oplv; |
2301 | } |
2302 | if (subparts_gt (type1: compute_rvtype, type2: compute_rtype)) |
2303 | { |
2304 | compute_rtype = compute_rvtype; |
2305 | opr = oprv; |
2306 | } |
2307 | /* Pick the narrowest type from LSHIFT_EXPR, RSHIFT_EXPR and |
2308 | BIT_IOR_EXPR. */ |
2309 | compute_type = compute_ltype; |
2310 | if (subparts_gt (type1: compute_type, type2: compute_rtype)) |
2311 | compute_type = compute_rtype; |
2312 | if (subparts_gt (type1: compute_type, type2: compute_otype)) |
2313 | compute_type = compute_otype; |
2314 | /* Verify all 3 operations can be performed in that type. */ |
2315 | if (compute_type != TREE_TYPE (type)) |
2316 | { |
2317 | if (optab_handler (op: opl, TYPE_MODE (compute_type)) |
2318 | == CODE_FOR_nothing |
2319 | || optab_handler (op: opr, TYPE_MODE (compute_type)) |
2320 | == CODE_FOR_nothing |
2321 | || optab_handler (op: opo, TYPE_MODE (compute_type)) |
2322 | == CODE_FOR_nothing) |
2323 | compute_type = TREE_TYPE (type); |
2324 | } |
2325 | } |
2326 | } |
2327 | } |
2328 | else |
2329 | op = optab_for_tree_code (code, type, optab_default); |
2330 | |
2331 | /* Optabs will try converting a negation into a subtraction, so |
2332 | look for it as well. TODO: negation of floating-point vectors |
2333 | might be turned into an exclusive OR toggling the sign bit. */ |
2334 | if (op == unknown_optab |
2335 | && code == NEGATE_EXPR |
2336 | && INTEGRAL_TYPE_P (TREE_TYPE (type))) |
2337 | op = optab_for_tree_code (MINUS_EXPR, type, optab_default); |
2338 | |
2339 | if (compute_type == NULL_TREE) |
2340 | compute_type = get_compute_type (code, op, type); |
2341 | if (compute_type == type) |
2342 | return; |
2343 | |
2344 | new_rhs = expand_vector_operation (gsi, type, compute_type, assign: stmt, code, |
2345 | dce_ssa_names); |
2346 | |
2347 | /* Leave expression untouched for later expansion. */ |
2348 | if (new_rhs == NULL_TREE) |
2349 | return; |
2350 | |
2351 | if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (new_rhs))) |
2352 | new_rhs = gimplify_build1 (gsi, code: VIEW_CONVERT_EXPR, TREE_TYPE (lhs), |
2353 | a: new_rhs); |
2354 | |
2355 | /* NOTE: We should avoid using gimple_assign_set_rhs_from_tree. One |
2356 | way to do it is change expand_vector_operation and its callees to |
2357 | return a tree_code, RHS1 and RHS2 instead of a tree. */ |
2358 | gimple_assign_set_rhs_from_tree (gsi, new_rhs); |
2359 | update_stmt (s: gsi_stmt (i: *gsi)); |
2360 | } |
2361 | |
2362 | /* Use this to lower vector operations introduced by the vectorizer, |
2363 | if it may need the bit-twiddling tricks implemented in this file. */ |
2364 | |
2365 | static unsigned int |
2366 | expand_vector_operations (void) |
2367 | { |
2368 | gimple_stmt_iterator gsi; |
2369 | basic_block bb; |
2370 | bool cfg_changed = false; |
2371 | |
2372 | auto_bitmap dce_ssa_names; |
2373 | |
2374 | FOR_EACH_BB_FN (bb, cfun) |
2375 | { |
2376 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2377 | { |
2378 | expand_vector_operations_1 (gsi: &gsi, dce_ssa_names); |
2379 | /* ??? If we do not cleanup EH then we will ICE in |
2380 | verification. But in reality we have created wrong-code |
2381 | as we did not properly transition EH info and edges to |
2382 | the piecewise computations. */ |
2383 | if (maybe_clean_eh_stmt (gsi_stmt (i: gsi)) |
2384 | && gimple_purge_dead_eh_edges (bb)) |
2385 | cfg_changed = true; |
2386 | /* If a .LOOP_DIST_ALIAS call prevailed loops got elided |
2387 | before vectorization got a chance to get at them. Simply |
2388 | fold as if loop distribution wasn't performed. */ |
2389 | if (gimple_call_internal_p (gs: gsi_stmt (i: gsi), fn: IFN_LOOP_DIST_ALIAS)) |
2390 | { |
2391 | fold_loop_internal_call (gsi_stmt (i: gsi), boolean_false_node); |
2392 | cfg_changed = true; |
2393 | } |
2394 | } |
2395 | } |
2396 | |
2397 | simple_dce_from_worklist (dce_ssa_names); |
2398 | |
2399 | return cfg_changed ? TODO_cleanup_cfg : 0; |
2400 | } |
2401 | |
2402 | namespace { |
2403 | |
2404 | const pass_data pass_data_lower_vector = |
2405 | { |
2406 | .type: GIMPLE_PASS, /* type */ |
2407 | .name: "veclower" , /* name */ |
2408 | .optinfo_flags: OPTGROUP_VEC, /* optinfo_flags */ |
2409 | .tv_id: TV_NONE, /* tv_id */ |
2410 | PROP_cfg, /* properties_required */ |
2411 | PROP_gimple_lvec, /* properties_provided */ |
2412 | .properties_destroyed: 0, /* properties_destroyed */ |
2413 | .todo_flags_start: 0, /* todo_flags_start */ |
2414 | TODO_update_ssa, /* todo_flags_finish */ |
2415 | }; |
2416 | |
2417 | class pass_lower_vector : public gimple_opt_pass |
2418 | { |
2419 | public: |
2420 | pass_lower_vector (gcc::context *ctxt) |
2421 | : gimple_opt_pass (pass_data_lower_vector, ctxt) |
2422 | {} |
2423 | |
2424 | /* opt_pass methods: */ |
2425 | bool gate (function *fun) final override |
2426 | { |
2427 | return !(fun->curr_properties & PROP_gimple_lvec); |
2428 | } |
2429 | |
2430 | unsigned int execute (function *) final override |
2431 | { |
2432 | return expand_vector_operations (); |
2433 | } |
2434 | |
2435 | }; // class pass_lower_vector |
2436 | |
2437 | } // anon namespace |
2438 | |
2439 | gimple_opt_pass * |
2440 | make_pass_lower_vector (gcc::context *ctxt) |
2441 | { |
2442 | return new pass_lower_vector (ctxt); |
2443 | } |
2444 | |
2445 | namespace { |
2446 | |
2447 | const pass_data pass_data_lower_vector_ssa = |
2448 | { |
2449 | .type: GIMPLE_PASS, /* type */ |
2450 | .name: "veclower2" , /* name */ |
2451 | .optinfo_flags: OPTGROUP_VEC, /* optinfo_flags */ |
2452 | .tv_id: TV_NONE, /* tv_id */ |
2453 | PROP_cfg, /* properties_required */ |
2454 | PROP_gimple_lvec, /* properties_provided */ |
2455 | .properties_destroyed: 0, /* properties_destroyed */ |
2456 | .todo_flags_start: 0, /* todo_flags_start */ |
2457 | .todo_flags_finish: ( TODO_update_ssa |
2458 | | TODO_cleanup_cfg ), /* todo_flags_finish */ |
2459 | }; |
2460 | |
2461 | class pass_lower_vector_ssa : public gimple_opt_pass |
2462 | { |
2463 | public: |
2464 | pass_lower_vector_ssa (gcc::context *ctxt) |
2465 | : gimple_opt_pass (pass_data_lower_vector_ssa, ctxt) |
2466 | {} |
2467 | |
2468 | /* opt_pass methods: */ |
2469 | opt_pass * clone () final override |
2470 | { |
2471 | return new pass_lower_vector_ssa (m_ctxt); |
2472 | } |
2473 | unsigned int execute (function *) final override |
2474 | { |
2475 | return expand_vector_operations (); |
2476 | } |
2477 | |
2478 | }; // class pass_lower_vector_ssa |
2479 | |
2480 | } // anon namespace |
2481 | |
2482 | gimple_opt_pass * |
2483 | make_pass_lower_vector_ssa (gcc::context *ctxt) |
2484 | { |
2485 | return new pass_lower_vector_ssa (ctxt); |
2486 | } |
2487 | |
2488 | #include "gt-tree-vect-generic.h" |
2489 | |