]>
Commit | Line | Data |
---|---|---|
25583c4f | 1 | /* Internal functions. |
7adcbafe | 2 | Copyright (C) 2011-2022 Free Software Foundation, Inc. |
25583c4f RS |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | 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" | |
c7131fb2 | 23 | #include "backend.h" |
957060b5 AM |
24 | #include "target.h" |
25 | #include "rtl.h" | |
25583c4f | 26 | #include "tree.h" |
c7131fb2 | 27 | #include "gimple.h" |
957060b5 AM |
28 | #include "predict.h" |
29 | #include "stringpool.h" | |
f90aa46c | 30 | #include "tree-vrp.h" |
957060b5 AM |
31 | #include "tree-ssanames.h" |
32 | #include "expmed.h" | |
4d0cdd0c | 33 | #include "memmodel.h" |
957060b5 AM |
34 | #include "optabs.h" |
35 | #include "emit-rtl.h" | |
36 | #include "diagnostic-core.h" | |
40e23961 | 37 | #include "fold-const.h" |
0e37a2f3 | 38 | #include "internal-fn.h" |
d8a2d370 | 39 | #include "stor-layout.h" |
36566b39 | 40 | #include "dojump.h" |
25583c4f | 41 | #include "expr.h" |
314e6352 ML |
42 | #include "stringpool.h" |
43 | #include "attribs.h" | |
e3174bdf | 44 | #include "asan.h" |
31e071ae | 45 | #include "ubsan.h" |
686ee971 | 46 | #include "recog.h" |
adedd5c1 | 47 | #include "builtins.h" |
1705cebd | 48 | #include "optabs-tree.h" |
0b99f253 JJ |
49 | #include "gimple-ssa.h" |
50 | #include "tree-phinodes.h" | |
51 | #include "ssa-iterators.h" | |
502d63b6 | 52 | #include "explow.h" |
bf510679 | 53 | #include "rtl-iter.h" |
45f4e2b0 | 54 | #include "gimple-range.h" |
25583c4f | 55 | |
a25e0b5e | 56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
58 | ||
25583c4f RS |
59 | /* The names of each internal function, indexed by function number. */ |
60 | const char *const internal_fn_name_array[] = { | |
b78475cf | 61 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) #CODE, |
25583c4f | 62 | #include "internal-fn.def" |
25583c4f RS |
63 | "<invalid-fn>" |
64 | }; | |
65 | ||
66 | /* The ECF_* flags of each internal function, indexed by function number. */ | |
67 | const int internal_fn_flags_array[] = { | |
b78475cf | 68 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) FLAGS, |
25583c4f | 69 | #include "internal-fn.def" |
25583c4f RS |
70 | 0 |
71 | }; | |
72 | ||
e4f81565 RS |
73 | /* Return the internal function called NAME, or IFN_LAST if there's |
74 | no such function. */ | |
75 | ||
76 | internal_fn | |
77 | lookup_internal_fn (const char *name) | |
78 | { | |
79 | typedef hash_map<nofree_string_hash, internal_fn> name_to_fn_map_type; | |
80 | static name_to_fn_map_type *name_to_fn_map; | |
81 | ||
82 | if (!name_to_fn_map) | |
83 | { | |
84 | name_to_fn_map = new name_to_fn_map_type (IFN_LAST); | |
85 | for (unsigned int i = 0; i < IFN_LAST; ++i) | |
86 | name_to_fn_map->put (internal_fn_name (internal_fn (i)), | |
87 | internal_fn (i)); | |
88 | } | |
89 | internal_fn *entry = name_to_fn_map->get (name); | |
90 | return entry ? *entry : IFN_LAST; | |
91 | } | |
92 | ||
b78475cf YG |
93 | /* Fnspec of each internal function, indexed by function number. */ |
94 | const_tree internal_fn_fnspec_array[IFN_LAST + 1]; | |
95 | ||
96 | void | |
97 | init_internal_fns () | |
98 | { | |
99 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
100 | if (FNSPEC) internal_fn_fnspec_array[IFN_##CODE] = \ | |
ba663ac1 | 101 | build_string ((int) sizeof (FNSPEC) - 1, FNSPEC ? FNSPEC : ""); |
b78475cf | 102 | #include "internal-fn.def" |
b78475cf YG |
103 | internal_fn_fnspec_array[IFN_LAST] = 0; |
104 | } | |
105 | ||
ab23f5d9 RS |
106 | /* Create static initializers for the information returned by |
107 | direct_internal_fn. */ | |
70439f0d | 108 | #define not_direct { -2, -2, false } |
bcde3345 | 109 | #define mask_load_direct { -1, 2, false } |
70439f0d | 110 | #define load_lanes_direct { -1, -1, false } |
7e11fc7f | 111 | #define mask_load_lanes_direct { -1, -1, false } |
09eb042a | 112 | #define gather_load_direct { 3, 1, false } |
d496134a | 113 | #define len_load_direct { -1, -1, false } |
70439f0d RS |
114 | #define mask_store_direct { 3, 2, false } |
115 | #define store_lanes_direct { 0, 0, false } | |
7e11fc7f | 116 | #define mask_store_lanes_direct { 0, 0, false } |
298e76e6 RS |
117 | #define vec_cond_mask_direct { 1, 0, false } |
118 | #define vec_cond_direct { 2, 0, false } | |
09eb042a | 119 | #define scatter_store_direct { 3, 1, false } |
d496134a | 120 | #define len_store_direct { 3, 3, false } |
683e55fa | 121 | #define vec_set_direct { 3, 3, false } |
70439f0d RS |
122 | #define unary_direct { 0, 0, true } |
123 | #define binary_direct { 0, 0, true } | |
c566cc9f | 124 | #define ternary_direct { 0, 0, true } |
bfe1bb57 | 125 | #define cond_unary_direct { 1, 1, true } |
0972596e | 126 | #define cond_binary_direct { 1, 1, true } |
b41d1f6e | 127 | #define cond_ternary_direct { 1, 1, true } |
7cfb4d93 | 128 | #define while_direct { 0, 2, false } |
bb6c2b68 | 129 | #define fold_extract_direct { 2, 2, false } |
b781a135 | 130 | #define fold_left_direct { 1, 1, false } |
bce29d65 | 131 | #define mask_fold_left_direct { 1, 1, false } |
58c036c8 | 132 | #define check_ptrs_direct { 0, 0, false } |
ab23f5d9 RS |
133 | |
134 | const direct_internal_fn_info direct_internal_fn_array[IFN_LAST + 1] = { | |
135 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) not_direct, | |
136 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) TYPE##_direct, | |
16d24520 RS |
137 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
138 | UNSIGNED_OPTAB, TYPE) TYPE##_direct, | |
ab23f5d9 RS |
139 | #include "internal-fn.def" |
140 | not_direct | |
141 | }; | |
142 | ||
272c6793 | 143 | /* ARRAY_TYPE is an array of vector modes. Return the associated insn |
ab23f5d9 | 144 | for load-lanes-style optab OPTAB, or CODE_FOR_nothing if none. */ |
272c6793 RS |
145 | |
146 | static enum insn_code | |
147 | get_multi_vector_move (tree array_type, convert_optab optab) | |
148 | { | |
ef4bddc2 RS |
149 | machine_mode imode; |
150 | machine_mode vmode; | |
272c6793 RS |
151 | |
152 | gcc_assert (TREE_CODE (array_type) == ARRAY_TYPE); | |
153 | imode = TYPE_MODE (array_type); | |
154 | vmode = TYPE_MODE (TREE_TYPE (array_type)); | |
155 | ||
ab23f5d9 | 156 | return convert_optab_handler (optab, imode, vmode); |
272c6793 RS |
157 | } |
158 | ||
ab23f5d9 | 159 | /* Expand LOAD_LANES call STMT using optab OPTAB. */ |
272c6793 RS |
160 | |
161 | static void | |
4cfe7a6c | 162 | expand_load_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
272c6793 | 163 | { |
99b1c316 | 164 | class expand_operand ops[2]; |
272c6793 RS |
165 | tree type, lhs, rhs; |
166 | rtx target, mem; | |
167 | ||
168 | lhs = gimple_call_lhs (stmt); | |
169 | rhs = gimple_call_arg (stmt, 0); | |
170 | type = TREE_TYPE (lhs); | |
171 | ||
172 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
173 | mem = expand_normal (rhs); | |
174 | ||
175 | gcc_assert (MEM_P (mem)); | |
176 | PUT_MODE (mem, TYPE_MODE (type)); | |
177 | ||
178 | create_output_operand (&ops[0], target, TYPE_MODE (type)); | |
179 | create_fixed_operand (&ops[1], mem); | |
ab23f5d9 | 180 | expand_insn (get_multi_vector_move (type, optab), 2, ops); |
3af3bec2 RS |
181 | if (!rtx_equal_p (target, ops[0].value)) |
182 | emit_move_insn (target, ops[0].value); | |
272c6793 RS |
183 | } |
184 | ||
ab23f5d9 | 185 | /* Expand STORE_LANES call STMT using optab OPTAB. */ |
272c6793 RS |
186 | |
187 | static void | |
4cfe7a6c | 188 | expand_store_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
272c6793 | 189 | { |
99b1c316 | 190 | class expand_operand ops[2]; |
272c6793 RS |
191 | tree type, lhs, rhs; |
192 | rtx target, reg; | |
193 | ||
194 | lhs = gimple_call_lhs (stmt); | |
195 | rhs = gimple_call_arg (stmt, 0); | |
196 | type = TREE_TYPE (rhs); | |
197 | ||
198 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
199 | reg = expand_normal (rhs); | |
200 | ||
201 | gcc_assert (MEM_P (target)); | |
202 | PUT_MODE (target, TYPE_MODE (type)); | |
203 | ||
204 | create_fixed_operand (&ops[0], target); | |
205 | create_input_operand (&ops[1], reg, TYPE_MODE (type)); | |
ab23f5d9 | 206 | expand_insn (get_multi_vector_move (type, optab), 2, ops); |
272c6793 RS |
207 | } |
208 | ||
8170608b | 209 | static void |
4cfe7a6c | 210 | expand_ANNOTATE (internal_fn, gcall *) |
8170608b TB |
211 | { |
212 | gcc_unreachable (); | |
213 | } | |
214 | ||
6c7509bc JJ |
215 | /* This should get expanded in omp_device_lower pass. */ |
216 | ||
217 | static void | |
218 | expand_GOMP_USE_SIMT (internal_fn, gcall *) | |
219 | { | |
220 | gcc_unreachable (); | |
221 | } | |
222 | ||
0c6b03b5 AM |
223 | /* This should get expanded in omp_device_lower pass. */ |
224 | ||
225 | static void | |
226 | expand_GOMP_SIMT_ENTER (internal_fn, gcall *) | |
227 | { | |
228 | gcc_unreachable (); | |
229 | } | |
230 | ||
231 | /* Allocate per-lane storage and begin non-uniform execution region. */ | |
232 | ||
233 | static void | |
234 | expand_GOMP_SIMT_ENTER_ALLOC (internal_fn, gcall *stmt) | |
235 | { | |
236 | rtx target; | |
237 | tree lhs = gimple_call_lhs (stmt); | |
238 | if (lhs) | |
239 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
240 | else | |
241 | target = gen_reg_rtx (Pmode); | |
242 | rtx size = expand_normal (gimple_call_arg (stmt, 0)); | |
243 | rtx align = expand_normal (gimple_call_arg (stmt, 1)); | |
99b1c316 | 244 | class expand_operand ops[3]; |
0c6b03b5 AM |
245 | create_output_operand (&ops[0], target, Pmode); |
246 | create_input_operand (&ops[1], size, Pmode); | |
247 | create_input_operand (&ops[2], align, Pmode); | |
248 | gcc_assert (targetm.have_omp_simt_enter ()); | |
249 | expand_insn (targetm.code_for_omp_simt_enter, 3, ops); | |
4d7c874e TV |
250 | if (!rtx_equal_p (target, ops[0].value)) |
251 | emit_move_insn (target, ops[0].value); | |
0c6b03b5 AM |
252 | } |
253 | ||
254 | /* Deallocate per-lane storage and leave non-uniform execution region. */ | |
255 | ||
256 | static void | |
257 | expand_GOMP_SIMT_EXIT (internal_fn, gcall *stmt) | |
258 | { | |
259 | gcc_checking_assert (!gimple_call_lhs (stmt)); | |
260 | rtx arg = expand_normal (gimple_call_arg (stmt, 0)); | |
99b1c316 | 261 | class expand_operand ops[1]; |
0c6b03b5 AM |
262 | create_input_operand (&ops[0], arg, Pmode); |
263 | gcc_assert (targetm.have_omp_simt_exit ()); | |
264 | expand_insn (targetm.code_for_omp_simt_exit, 1, ops); | |
265 | } | |
266 | ||
9669b00b AM |
267 | /* Lane index on SIMT targets: thread index in the warp on NVPTX. On targets |
268 | without SIMT execution this should be expanded in omp_device_lower pass. */ | |
269 | ||
270 | static void | |
271 | expand_GOMP_SIMT_LANE (internal_fn, gcall *stmt) | |
272 | { | |
273 | tree lhs = gimple_call_lhs (stmt); | |
274 | if (!lhs) | |
275 | return; | |
276 | ||
277 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
278 | gcc_assert (targetm.have_omp_simt_lane ()); | |
279 | emit_insn (targetm.gen_omp_simt_lane (target)); | |
280 | } | |
281 | ||
282 | /* This should get expanded in omp_device_lower pass. */ | |
283 | ||
284 | static void | |
285 | expand_GOMP_SIMT_VF (internal_fn, gcall *) | |
286 | { | |
287 | gcc_unreachable (); | |
288 | } | |
289 | ||
290 | /* Lane index of the first SIMT lane that supplies a non-zero argument. | |
291 | This is a SIMT counterpart to GOMP_SIMD_LAST_LANE, used to represent the | |
292 | lane that executed the last iteration for handling OpenMP lastprivate. */ | |
293 | ||
294 | static void | |
295 | expand_GOMP_SIMT_LAST_LANE (internal_fn, gcall *stmt) | |
296 | { | |
297 | tree lhs = gimple_call_lhs (stmt); | |
298 | if (!lhs) | |
299 | return; | |
300 | ||
301 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
302 | rtx cond = expand_normal (gimple_call_arg (stmt, 0)); | |
303 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
99b1c316 | 304 | class expand_operand ops[2]; |
9669b00b AM |
305 | create_output_operand (&ops[0], target, mode); |
306 | create_input_operand (&ops[1], cond, mode); | |
307 | gcc_assert (targetm.have_omp_simt_last_lane ()); | |
308 | expand_insn (targetm.code_for_omp_simt_last_lane, 2, ops); | |
4d7c874e TV |
309 | if (!rtx_equal_p (target, ops[0].value)) |
310 | emit_move_insn (target, ops[0].value); | |
9669b00b AM |
311 | } |
312 | ||
313 | /* Non-transparent predicate used in SIMT lowering of OpenMP "ordered". */ | |
314 | ||
315 | static void | |
316 | expand_GOMP_SIMT_ORDERED_PRED (internal_fn, gcall *stmt) | |
317 | { | |
318 | tree lhs = gimple_call_lhs (stmt); | |
319 | if (!lhs) | |
320 | return; | |
321 | ||
322 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
323 | rtx ctr = expand_normal (gimple_call_arg (stmt, 0)); | |
324 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
99b1c316 | 325 | class expand_operand ops[2]; |
9669b00b AM |
326 | create_output_operand (&ops[0], target, mode); |
327 | create_input_operand (&ops[1], ctr, mode); | |
328 | gcc_assert (targetm.have_omp_simt_ordered ()); | |
329 | expand_insn (targetm.code_for_omp_simt_ordered, 2, ops); | |
4d7c874e TV |
330 | if (!rtx_equal_p (target, ops[0].value)) |
331 | emit_move_insn (target, ops[0].value); | |
9669b00b AM |
332 | } |
333 | ||
334 | /* "Or" boolean reduction across SIMT lanes: return non-zero in all lanes if | |
335 | any lane supplies a non-zero argument. */ | |
336 | ||
337 | static void | |
338 | expand_GOMP_SIMT_VOTE_ANY (internal_fn, gcall *stmt) | |
339 | { | |
340 | tree lhs = gimple_call_lhs (stmt); | |
341 | if (!lhs) | |
342 | return; | |
343 | ||
344 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
345 | rtx cond = expand_normal (gimple_call_arg (stmt, 0)); | |
346 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
99b1c316 | 347 | class expand_operand ops[2]; |
9669b00b AM |
348 | create_output_operand (&ops[0], target, mode); |
349 | create_input_operand (&ops[1], cond, mode); | |
350 | gcc_assert (targetm.have_omp_simt_vote_any ()); | |
351 | expand_insn (targetm.code_for_omp_simt_vote_any, 2, ops); | |
4d7c874e TV |
352 | if (!rtx_equal_p (target, ops[0].value)) |
353 | emit_move_insn (target, ops[0].value); | |
9669b00b AM |
354 | } |
355 | ||
356 | /* Exchange between SIMT lanes with a "butterfly" pattern: source lane index | |
357 | is destination lane index XOR given offset. */ | |
358 | ||
359 | static void | |
360 | expand_GOMP_SIMT_XCHG_BFLY (internal_fn, gcall *stmt) | |
361 | { | |
362 | tree lhs = gimple_call_lhs (stmt); | |
363 | if (!lhs) | |
364 | return; | |
365 | ||
366 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
367 | rtx src = expand_normal (gimple_call_arg (stmt, 0)); | |
368 | rtx idx = expand_normal (gimple_call_arg (stmt, 1)); | |
369 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
99b1c316 | 370 | class expand_operand ops[3]; |
9669b00b AM |
371 | create_output_operand (&ops[0], target, mode); |
372 | create_input_operand (&ops[1], src, mode); | |
373 | create_input_operand (&ops[2], idx, SImode); | |
374 | gcc_assert (targetm.have_omp_simt_xchg_bfly ()); | |
375 | expand_insn (targetm.code_for_omp_simt_xchg_bfly, 3, ops); | |
4d7c874e TV |
376 | if (!rtx_equal_p (target, ops[0].value)) |
377 | emit_move_insn (target, ops[0].value); | |
9669b00b AM |
378 | } |
379 | ||
380 | /* Exchange between SIMT lanes according to given source lane index. */ | |
381 | ||
382 | static void | |
383 | expand_GOMP_SIMT_XCHG_IDX (internal_fn, gcall *stmt) | |
384 | { | |
385 | tree lhs = gimple_call_lhs (stmt); | |
386 | if (!lhs) | |
387 | return; | |
388 | ||
389 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
390 | rtx src = expand_normal (gimple_call_arg (stmt, 0)); | |
391 | rtx idx = expand_normal (gimple_call_arg (stmt, 1)); | |
392 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
99b1c316 | 393 | class expand_operand ops[3]; |
9669b00b AM |
394 | create_output_operand (&ops[0], target, mode); |
395 | create_input_operand (&ops[1], src, mode); | |
396 | create_input_operand (&ops[2], idx, SImode); | |
397 | gcc_assert (targetm.have_omp_simt_xchg_idx ()); | |
398 | expand_insn (targetm.code_for_omp_simt_xchg_idx, 3, ops); | |
4d7c874e TV |
399 | if (!rtx_equal_p (target, ops[0].value)) |
400 | emit_move_insn (target, ops[0].value); | |
9669b00b AM |
401 | } |
402 | ||
74bf76ed JJ |
403 | /* This should get expanded in adjust_simduid_builtins. */ |
404 | ||
405 | static void | |
4cfe7a6c | 406 | expand_GOMP_SIMD_LANE (internal_fn, gcall *) |
74bf76ed JJ |
407 | { |
408 | gcc_unreachable (); | |
409 | } | |
410 | ||
411 | /* This should get expanded in adjust_simduid_builtins. */ | |
412 | ||
413 | static void | |
4cfe7a6c | 414 | expand_GOMP_SIMD_VF (internal_fn, gcall *) |
74bf76ed JJ |
415 | { |
416 | gcc_unreachable (); | |
417 | } | |
418 | ||
419 | /* This should get expanded in adjust_simduid_builtins. */ | |
420 | ||
421 | static void | |
4cfe7a6c | 422 | expand_GOMP_SIMD_LAST_LANE (internal_fn, gcall *) |
74bf76ed JJ |
423 | { |
424 | gcc_unreachable (); | |
425 | } | |
426 | ||
d9a6bd32 JJ |
427 | /* This should get expanded in adjust_simduid_builtins. */ |
428 | ||
429 | static void | |
4cfe7a6c | 430 | expand_GOMP_SIMD_ORDERED_START (internal_fn, gcall *) |
d9a6bd32 JJ |
431 | { |
432 | gcc_unreachable (); | |
433 | } | |
434 | ||
435 | /* This should get expanded in adjust_simduid_builtins. */ | |
436 | ||
437 | static void | |
4cfe7a6c | 438 | expand_GOMP_SIMD_ORDERED_END (internal_fn, gcall *) |
d9a6bd32 JJ |
439 | { |
440 | gcc_unreachable (); | |
441 | } | |
442 | ||
b9a55b13 MP |
443 | /* This should get expanded in the sanopt pass. */ |
444 | ||
445 | static void | |
4cfe7a6c | 446 | expand_UBSAN_NULL (internal_fn, gcall *) |
b9a55b13 MP |
447 | { |
448 | gcc_unreachable (); | |
449 | } | |
450 | ||
0e37a2f3 MP |
451 | /* This should get expanded in the sanopt pass. */ |
452 | ||
453 | static void | |
4cfe7a6c | 454 | expand_UBSAN_BOUNDS (internal_fn, gcall *) |
0e82f089 MP |
455 | { |
456 | gcc_unreachable (); | |
457 | } | |
458 | ||
459 | /* This should get expanded in the sanopt pass. */ | |
460 | ||
461 | static void | |
4cfe7a6c | 462 | expand_UBSAN_VPTR (internal_fn, gcall *) |
0e37a2f3 MP |
463 | { |
464 | gcc_unreachable (); | |
465 | } | |
466 | ||
c62ccb9a YG |
467 | /* This should get expanded in the sanopt pass. */ |
468 | ||
c9b39a49 JJ |
469 | static void |
470 | expand_UBSAN_PTR (internal_fn, gcall *) | |
471 | { | |
472 | gcc_unreachable (); | |
473 | } | |
474 | ||
475 | /* This should get expanded in the sanopt pass. */ | |
476 | ||
c62ccb9a | 477 | static void |
4cfe7a6c | 478 | expand_UBSAN_OBJECT_SIZE (internal_fn, gcall *) |
35228ac7 JJ |
479 | { |
480 | gcc_unreachable (); | |
481 | } | |
482 | ||
483 | /* This should get expanded in the sanopt pass. */ | |
484 | ||
93a73251 MM |
485 | static void |
486 | expand_HWASAN_CHECK (internal_fn, gcall *) | |
487 | { | |
488 | gcc_unreachable (); | |
489 | } | |
490 | ||
491 | /* For hwasan stack tagging: | |
492 | Clear tags on the dynamically allocated space. | |
493 | For use after an object dynamically allocated on the stack goes out of | |
494 | scope. */ | |
495 | static void | |
496 | expand_HWASAN_ALLOCA_UNPOISON (internal_fn, gcall *gc) | |
497 | { | |
498 | gcc_assert (Pmode == ptr_mode); | |
499 | tree restored_position = gimple_call_arg (gc, 0); | |
500 | rtx restored_rtx = expand_expr (restored_position, NULL_RTX, VOIDmode, | |
501 | EXPAND_NORMAL); | |
502 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); | |
503 | rtx off = expand_simple_binop (Pmode, MINUS, restored_rtx, | |
504 | stack_pointer_rtx, NULL_RTX, 0, | |
505 | OPTAB_WIDEN); | |
506 | emit_library_call_value (func, NULL_RTX, LCT_NORMAL, VOIDmode, | |
507 | virtual_stack_dynamic_rtx, Pmode, | |
508 | HWASAN_STACK_BACKGROUND, QImode, | |
509 | off, Pmode); | |
510 | } | |
511 | ||
512 | /* For hwasan stack tagging: | |
513 | Return a tag to be used for a dynamic allocation. */ | |
514 | static void | |
515 | expand_HWASAN_CHOOSE_TAG (internal_fn, gcall *gc) | |
516 | { | |
517 | tree tag = gimple_call_lhs (gc); | |
518 | rtx target = expand_expr (tag, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
519 | machine_mode mode = GET_MODE (target); | |
520 | gcc_assert (mode == QImode); | |
521 | ||
522 | rtx base_tag = targetm.memtag.extract_tag (hwasan_frame_base (), NULL_RTX); | |
523 | gcc_assert (base_tag); | |
524 | rtx tag_offset = gen_int_mode (hwasan_current_frame_tag (), QImode); | |
525 | rtx chosen_tag = expand_simple_binop (QImode, PLUS, base_tag, tag_offset, | |
526 | target, /* unsignedp = */1, | |
527 | OPTAB_WIDEN); | |
528 | chosen_tag = hwasan_truncate_to_tag_size (chosen_tag, target); | |
529 | ||
530 | /* Really need to put the tag into the `target` RTX. */ | |
531 | if (chosen_tag != target) | |
532 | { | |
533 | rtx temp = chosen_tag; | |
534 | gcc_assert (GET_MODE (chosen_tag) == mode); | |
535 | emit_move_insn (target, temp); | |
536 | } | |
537 | ||
538 | hwasan_increment_frame_tag (); | |
539 | } | |
540 | ||
541 | /* For hwasan stack tagging: | |
542 | Tag a region of space in the shadow stack according to the base pointer of | |
543 | an object on the stack. N.b. the length provided in the internal call is | |
544 | required to be aligned to HWASAN_TAG_GRANULE_SIZE. */ | |
545 | static void | |
546 | expand_HWASAN_MARK (internal_fn, gcall *gc) | |
547 | { | |
548 | gcc_assert (ptr_mode == Pmode); | |
549 | HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gc, 0)); | |
550 | bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON; | |
551 | ||
552 | tree base = gimple_call_arg (gc, 1); | |
553 | gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR); | |
554 | rtx base_rtx = expand_normal (base); | |
555 | ||
556 | rtx tag = is_poison ? HWASAN_STACK_BACKGROUND | |
557 | : targetm.memtag.extract_tag (base_rtx, NULL_RTX); | |
558 | rtx address = targetm.memtag.untagged_pointer (base_rtx, NULL_RTX); | |
559 | ||
560 | tree len = gimple_call_arg (gc, 2); | |
561 | rtx r_len = expand_normal (len); | |
562 | ||
563 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); | |
564 | emit_library_call (func, LCT_NORMAL, VOIDmode, address, Pmode, | |
565 | tag, QImode, r_len, Pmode); | |
566 | } | |
567 | ||
568 | /* For hwasan stack tagging: | |
569 | Store a tag into a pointer. */ | |
570 | static void | |
571 | expand_HWASAN_SET_TAG (internal_fn, gcall *gc) | |
572 | { | |
573 | gcc_assert (ptr_mode == Pmode); | |
574 | tree g_target = gimple_call_lhs (gc); | |
575 | tree g_ptr = gimple_call_arg (gc, 0); | |
576 | tree g_tag = gimple_call_arg (gc, 1); | |
577 | ||
578 | rtx ptr = expand_normal (g_ptr); | |
579 | rtx tag = expand_expr (g_tag, NULL_RTX, QImode, EXPAND_NORMAL); | |
580 | rtx target = expand_normal (g_target); | |
581 | ||
582 | rtx untagged = targetm.memtag.untagged_pointer (ptr, target); | |
583 | rtx tagged_value = targetm.memtag.set_tag (untagged, tag, target); | |
584 | if (tagged_value != target) | |
585 | emit_move_insn (target, tagged_value); | |
586 | } | |
587 | ||
588 | /* This should get expanded in the sanopt pass. */ | |
589 | ||
35228ac7 | 590 | static void |
4cfe7a6c | 591 | expand_ASAN_CHECK (internal_fn, gcall *) |
c62ccb9a YG |
592 | { |
593 | gcc_unreachable (); | |
594 | } | |
595 | ||
6dc4a604 ML |
596 | /* This should get expanded in the sanopt pass. */ |
597 | ||
598 | static void | |
599 | expand_ASAN_MARK (internal_fn, gcall *) | |
600 | { | |
601 | gcc_unreachable (); | |
602 | } | |
603 | ||
c7775327 ML |
604 | /* This should get expanded in the sanopt pass. */ |
605 | ||
606 | static void | |
607 | expand_ASAN_POISON (internal_fn, gcall *) | |
608 | { | |
609 | gcc_unreachable (); | |
610 | } | |
6dc4a604 | 611 | |
f6b9f2ff ML |
612 | /* This should get expanded in the sanopt pass. */ |
613 | ||
614 | static void | |
615 | expand_ASAN_POISON_USE (internal_fn, gcall *) | |
616 | { | |
617 | gcc_unreachable (); | |
618 | } | |
619 | ||
fca4adf2 JJ |
620 | /* This should get expanded in the tsan pass. */ |
621 | ||
622 | static void | |
4cfe7a6c | 623 | expand_TSAN_FUNC_EXIT (internal_fn, gcall *) |
fca4adf2 JJ |
624 | { |
625 | gcc_unreachable (); | |
626 | } | |
627 | ||
81fea426 MP |
628 | /* This should get expanded in the lower pass. */ |
629 | ||
630 | static void | |
631 | expand_FALLTHROUGH (internal_fn, gcall *call) | |
632 | { | |
633 | error_at (gimple_location (call), | |
634 | "invalid use of attribute %<fallthrough%>"); | |
635 | } | |
636 | ||
1304953e JJ |
637 | /* Return minimum precision needed to represent all values |
638 | of ARG in SIGNed integral type. */ | |
639 | ||
640 | static int | |
641 | get_min_precision (tree arg, signop sign) | |
642 | { | |
643 | int prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
644 | int cnt = 0; | |
645 | signop orig_sign = sign; | |
646 | if (TREE_CODE (arg) == INTEGER_CST) | |
647 | { | |
648 | int p; | |
649 | if (TYPE_SIGN (TREE_TYPE (arg)) != sign) | |
650 | { | |
651 | widest_int w = wi::to_widest (arg); | |
652 | w = wi::ext (w, prec, sign); | |
653 | p = wi::min_precision (w, sign); | |
654 | } | |
655 | else | |
8e6cdc90 | 656 | p = wi::min_precision (wi::to_wide (arg), sign); |
1304953e JJ |
657 | return MIN (p, prec); |
658 | } | |
659 | while (CONVERT_EXPR_P (arg) | |
660 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) | |
661 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) <= prec) | |
662 | { | |
663 | arg = TREE_OPERAND (arg, 0); | |
664 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) | |
665 | { | |
666 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) | |
667 | sign = UNSIGNED; | |
668 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) | |
669 | return prec + (orig_sign != sign); | |
670 | prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
671 | } | |
672 | if (++cnt > 30) | |
673 | return prec + (orig_sign != sign); | |
674 | } | |
049ce9d2 JJ |
675 | if (CONVERT_EXPR_P (arg) |
676 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) | |
677 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) > prec) | |
678 | { | |
679 | /* We have e.g. (unsigned short) y_2 where int y_2 = (int) x_1(D); | |
680 | If y_2's min precision is smaller than prec, return that. */ | |
681 | int oprec = get_min_precision (TREE_OPERAND (arg, 0), sign); | |
682 | if (oprec < prec) | |
683 | return oprec + (orig_sign != sign); | |
684 | } | |
1304953e JJ |
685 | if (TREE_CODE (arg) != SSA_NAME) |
686 | return prec + (orig_sign != sign); | |
45f4e2b0 AH |
687 | value_range r; |
688 | while (!get_global_range_query ()->range_of_expr (r, arg) | |
689 | || r.kind () != VR_RANGE) | |
1304953e | 690 | { |
355fe088 | 691 | gimple *g = SSA_NAME_DEF_STMT (arg); |
1304953e JJ |
692 | if (is_gimple_assign (g) |
693 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (g))) | |
694 | { | |
695 | tree t = gimple_assign_rhs1 (g); | |
696 | if (INTEGRAL_TYPE_P (TREE_TYPE (t)) | |
697 | && TYPE_PRECISION (TREE_TYPE (t)) <= prec) | |
698 | { | |
699 | arg = t; | |
700 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) | |
701 | { | |
702 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) | |
703 | sign = UNSIGNED; | |
704 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) | |
705 | return prec + (orig_sign != sign); | |
706 | prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
707 | } | |
708 | if (++cnt > 30) | |
709 | return prec + (orig_sign != sign); | |
710 | continue; | |
711 | } | |
712 | } | |
713 | return prec + (orig_sign != sign); | |
714 | } | |
715 | if (sign == TYPE_SIGN (TREE_TYPE (arg))) | |
716 | { | |
45f4e2b0 AH |
717 | int p1 = wi::min_precision (r.lower_bound (), sign); |
718 | int p2 = wi::min_precision (r.upper_bound (), sign); | |
1304953e JJ |
719 | p1 = MAX (p1, p2); |
720 | prec = MIN (prec, p1); | |
721 | } | |
45f4e2b0 | 722 | else if (sign == UNSIGNED && !wi::neg_p (r.lower_bound (), SIGNED)) |
1304953e | 723 | { |
45f4e2b0 | 724 | int p = wi::min_precision (r.upper_bound (), UNSIGNED); |
1304953e JJ |
725 | prec = MIN (prec, p); |
726 | } | |
727 | return prec + (orig_sign != sign); | |
728 | } | |
729 | ||
a86451b9 JJ |
730 | /* Helper for expand_*_overflow. Set the __imag__ part to true |
731 | (1 except for signed:1 type, in which case store -1). */ | |
732 | ||
733 | static void | |
734 | expand_arith_set_overflow (tree lhs, rtx target) | |
735 | { | |
736 | if (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))) == 1 | |
737 | && !TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs)))) | |
738 | write_complex_part (target, constm1_rtx, true); | |
739 | else | |
740 | write_complex_part (target, const1_rtx, true); | |
741 | } | |
742 | ||
1304953e JJ |
743 | /* Helper for expand_*_overflow. Store RES into the __real__ part |
744 | of TARGET. If RES has larger MODE than __real__ part of TARGET, | |
a86451b9 JJ |
745 | set the __imag__ part to 1 if RES doesn't fit into it. Similarly |
746 | if LHS has smaller precision than its mode. */ | |
1304953e JJ |
747 | |
748 | static void | |
749 | expand_arith_overflow_result_store (tree lhs, rtx target, | |
095a2d76 | 750 | scalar_int_mode mode, rtx res) |
1304953e | 751 | { |
c7ad039d RS |
752 | scalar_int_mode tgtmode |
753 | = as_a <scalar_int_mode> (GET_MODE_INNER (GET_MODE (target))); | |
1304953e JJ |
754 | rtx lres = res; |
755 | if (tgtmode != mode) | |
756 | { | |
757 | rtx_code_label *done_label = gen_label_rtx (); | |
758 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); | |
759 | lres = convert_modes (tgtmode, mode, res, uns); | |
760 | gcc_assert (GET_MODE_PRECISION (tgtmode) < GET_MODE_PRECISION (mode)); | |
92344ed0 | 761 | do_compare_rtx_and_jump (res, convert_modes (mode, tgtmode, lres, uns), |
1476d1bd | 762 | EQ, true, mode, NULL_RTX, NULL, done_label, |
357067f2 | 763 | profile_probability::very_likely ()); |
a86451b9 JJ |
764 | expand_arith_set_overflow (lhs, target); |
765 | emit_label (done_label); | |
766 | } | |
767 | int prec = TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))); | |
768 | int tgtprec = GET_MODE_PRECISION (tgtmode); | |
769 | if (prec < tgtprec) | |
770 | { | |
771 | rtx_code_label *done_label = gen_label_rtx (); | |
772 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); | |
773 | res = lres; | |
774 | if (uns) | |
775 | { | |
776 | rtx mask | |
777 | = immed_wide_int_const (wi::shifted_mask (0, prec, false, tgtprec), | |
778 | tgtmode); | |
779 | lres = expand_simple_binop (tgtmode, AND, res, mask, NULL_RTX, | |
95ef39f4 | 780 | true, OPTAB_LIB_WIDEN); |
a86451b9 JJ |
781 | } |
782 | else | |
783 | { | |
784 | lres = expand_shift (LSHIFT_EXPR, tgtmode, res, tgtprec - prec, | |
785 | NULL_RTX, 1); | |
786 | lres = expand_shift (RSHIFT_EXPR, tgtmode, lres, tgtprec - prec, | |
787 | NULL_RTX, 0); | |
788 | } | |
789 | do_compare_rtx_and_jump (res, lres, | |
790 | EQ, true, tgtmode, NULL_RTX, NULL, done_label, | |
357067f2 | 791 | profile_probability::very_likely ()); |
a86451b9 | 792 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
793 | emit_label (done_label); |
794 | } | |
795 | write_complex_part (target, lres, false); | |
796 | } | |
797 | ||
5620052d JJ |
798 | /* Helper for expand_*_overflow. Store RES into TARGET. */ |
799 | ||
800 | static void | |
801 | expand_ubsan_result_store (rtx target, rtx res) | |
802 | { | |
803 | if (GET_CODE (target) == SUBREG && SUBREG_PROMOTED_VAR_P (target)) | |
804 | /* If this is a scalar in a register that is stored in a wider mode | |
805 | than the declared mode, compute the result into its declared mode | |
806 | and then convert to the wider mode. Our value is the computed | |
807 | expression. */ | |
808 | convert_move (SUBREG_REG (target), res, SUBREG_PROMOTED_SIGN (target)); | |
809 | else | |
810 | emit_move_insn (target, res); | |
811 | } | |
812 | ||
31e071ae MP |
813 | /* Add sub/add overflow checking to the statement STMT. |
814 | CODE says whether the operation is +, or -. */ | |
815 | ||
4d87bd39 | 816 | void |
1304953e JJ |
817 | expand_addsub_overflow (location_t loc, tree_code code, tree lhs, |
818 | tree arg0, tree arg1, bool unsr_p, bool uns0_p, | |
1705cebd | 819 | bool uns1_p, bool is_ubsan, tree *datap) |
31e071ae | 820 | { |
1304953e JJ |
821 | rtx res, target = NULL_RTX; |
822 | tree fn; | |
823 | rtx_code_label *done_label = gen_label_rtx (); | |
824 | rtx_code_label *do_error = gen_label_rtx (); | |
31e071ae | 825 | do_pending_stack_adjust (); |
1304953e JJ |
826 | rtx op0 = expand_normal (arg0); |
827 | rtx op1 = expand_normal (arg1); | |
7a504f33 | 828 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1304953e JJ |
829 | int prec = GET_MODE_PRECISION (mode); |
830 | rtx sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); | |
831 | bool do_xor = false; | |
832 | ||
833 | if (is_ubsan) | |
834 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); | |
835 | ||
31e071ae | 836 | if (lhs) |
1304953e JJ |
837 | { |
838 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
839 | if (!is_ubsan) | |
840 | write_complex_part (target, const0_rtx, true); | |
841 | } | |
842 | ||
843 | /* We assume both operands and result have the same precision | |
844 | here (GET_MODE_BITSIZE (mode)), S stands for signed type | |
845 | with that precision, U for unsigned type with that precision, | |
846 | sgn for unsigned most significant bit in that precision. | |
847 | s1 is signed first operand, u1 is unsigned first operand, | |
848 | s2 is signed second operand, u2 is unsigned second operand, | |
849 | sr is signed result, ur is unsigned result and the following | |
850 | rules say how to compute result (which is always result of | |
851 | the operands as if both were unsigned, cast to the right | |
852 | signedness) and how to compute whether operation overflowed. | |
853 | ||
854 | s1 + s2 -> sr | |
855 | res = (S) ((U) s1 + (U) s2) | |
856 | ovf = s2 < 0 ? res > s1 : res < s1 (or jump on overflow) | |
857 | s1 - s2 -> sr | |
858 | res = (S) ((U) s1 - (U) s2) | |
859 | ovf = s2 < 0 ? res < s1 : res > s2 (or jump on overflow) | |
860 | u1 + u2 -> ur | |
861 | res = u1 + u2 | |
862 | ovf = res < u1 (or jump on carry, but RTL opts will handle it) | |
863 | u1 - u2 -> ur | |
864 | res = u1 - u2 | |
865 | ovf = res > u1 (or jump on carry, but RTL opts will handle it) | |
866 | s1 + u2 -> sr | |
867 | res = (S) ((U) s1 + u2) | |
868 | ovf = ((U) res ^ sgn) < u2 | |
869 | s1 + u2 -> ur | |
870 | t1 = (S) (u2 ^ sgn) | |
871 | t2 = s1 + t1 | |
872 | res = (U) t2 ^ sgn | |
873 | ovf = t1 < 0 ? t2 > s1 : t2 < s1 (or jump on overflow) | |
874 | s1 - u2 -> sr | |
875 | res = (S) ((U) s1 - u2) | |
876 | ovf = u2 > ((U) s1 ^ sgn) | |
877 | s1 - u2 -> ur | |
878 | res = (U) s1 - u2 | |
879 | ovf = s1 < 0 || u2 > (U) s1 | |
880 | u1 - s2 -> sr | |
881 | res = u1 - (U) s2 | |
882 | ovf = u1 >= ((U) s2 ^ sgn) | |
883 | u1 - s2 -> ur | |
884 | t1 = u1 ^ sgn | |
885 | t2 = t1 - (U) s2 | |
886 | res = t2 ^ sgn | |
887 | ovf = s2 < 0 ? (S) t2 < (S) t1 : (S) t2 > (S) t1 (or jump on overflow) | |
888 | s1 + s2 -> ur | |
889 | res = (U) s1 + (U) s2 | |
890 | ovf = s2 < 0 ? (s1 | (S) res) < 0) : (s1 & (S) res) < 0) | |
891 | u1 + u2 -> sr | |
892 | res = (S) (u1 + u2) | |
893 | ovf = (U) res < u2 || res < 0 | |
894 | u1 - u2 -> sr | |
895 | res = (S) (u1 - u2) | |
896 | ovf = u1 >= u2 ? res < 0 : res >= 0 | |
897 | s1 - s2 -> ur | |
898 | res = (U) s1 - (U) s2 | |
899 | ovf = s2 >= 0 ? ((s1 | (S) res) < 0) : ((s1 & (S) res) < 0) */ | |
900 | ||
901 | if (code == PLUS_EXPR && uns0_p && !uns1_p) | |
902 | { | |
903 | /* PLUS_EXPR is commutative, if operand signedness differs, | |
904 | canonicalize to the first operand being signed and second | |
905 | unsigned to simplify following code. */ | |
6b4db501 MM |
906 | std::swap (op0, op1); |
907 | std::swap (arg0, arg1); | |
908 | uns0_p = false; | |
909 | uns1_p = true; | |
1304953e JJ |
910 | } |
911 | ||
912 | /* u1 +- u2 -> ur */ | |
913 | if (uns0_p && uns1_p && unsr_p) | |
914 | { | |
cde9d596 RH |
915 | insn_code icode = optab_handler (code == PLUS_EXPR ? uaddv4_optab |
916 | : usubv4_optab, mode); | |
917 | if (icode != CODE_FOR_nothing) | |
918 | { | |
99b1c316 | 919 | class expand_operand ops[4]; |
cde9d596 RH |
920 | rtx_insn *last = get_last_insn (); |
921 | ||
922 | res = gen_reg_rtx (mode); | |
923 | create_output_operand (&ops[0], res, mode); | |
924 | create_input_operand (&ops[1], op0, mode); | |
925 | create_input_operand (&ops[2], op1, mode); | |
926 | create_fixed_operand (&ops[3], do_error); | |
927 | if (maybe_expand_insn (icode, 4, ops)) | |
928 | { | |
929 | last = get_last_insn (); | |
930 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT | |
931 | && JUMP_P (last) | |
932 | && any_condjump_p (last) | |
933 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
934 | add_reg_br_prob_note (last, |
935 | profile_probability::very_unlikely ()); | |
cde9d596 RH |
936 | emit_jump (done_label); |
937 | goto do_error_label; | |
938 | } | |
939 | ||
940 | delete_insns_since (last); | |
941 | } | |
942 | ||
1304953e JJ |
943 | /* Compute the operation. On RTL level, the addition is always |
944 | unsigned. */ | |
945 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
946 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
947 | rtx tem = op0; | |
948 | /* For PLUS_EXPR, the operation is commutative, so we can pick | |
949 | operand to compare against. For prec <= BITS_PER_WORD, I think | |
950 | preferring REG operand is better over CONST_INT, because | |
951 | the CONST_INT might enlarge the instruction or CSE would need | |
952 | to figure out we'd already loaded it into a register before. | |
953 | For prec > BITS_PER_WORD, I think CONST_INT might be more beneficial, | |
954 | as then the multi-word comparison can be perhaps simplified. */ | |
955 | if (code == PLUS_EXPR | |
956 | && (prec <= BITS_PER_WORD | |
957 | ? (CONST_SCALAR_INT_P (op0) && REG_P (op1)) | |
958 | : CONST_SCALAR_INT_P (op1))) | |
959 | tem = op1; | |
92344ed0 | 960 | do_compare_rtx_and_jump (res, tem, code == PLUS_EXPR ? GEU : LEU, |
1476d1bd | 961 | true, mode, NULL_RTX, NULL, done_label, |
357067f2 | 962 | profile_probability::very_likely ()); |
1304953e JJ |
963 | goto do_error_label; |
964 | } | |
965 | ||
966 | /* s1 +- u2 -> sr */ | |
967 | if (!uns0_p && uns1_p && !unsr_p) | |
968 | { | |
969 | /* Compute the operation. On RTL level, the addition is always | |
970 | unsigned. */ | |
971 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
972 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
973 | rtx tem = expand_binop (mode, add_optab, | |
974 | code == PLUS_EXPR ? res : op0, sgn, | |
975 | NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1476d1bd | 976 | do_compare_rtx_and_jump (tem, op1, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 977 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
978 | goto do_error_label; |
979 | } | |
980 | ||
981 | /* s1 + u2 -> ur */ | |
982 | if (code == PLUS_EXPR && !uns0_p && uns1_p && unsr_p) | |
983 | { | |
984 | op1 = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, | |
985 | OPTAB_LIB_WIDEN); | |
986 | /* As we've changed op1, we have to avoid using the value range | |
987 | for the original argument. */ | |
988 | arg1 = error_mark_node; | |
989 | do_xor = true; | |
990 | goto do_signed; | |
991 | } | |
992 | ||
993 | /* u1 - s2 -> ur */ | |
994 | if (code == MINUS_EXPR && uns0_p && !uns1_p && unsr_p) | |
995 | { | |
996 | op0 = expand_binop (mode, add_optab, op0, sgn, NULL_RTX, false, | |
997 | OPTAB_LIB_WIDEN); | |
998 | /* As we've changed op0, we have to avoid using the value range | |
999 | for the original argument. */ | |
1000 | arg0 = error_mark_node; | |
1001 | do_xor = true; | |
1002 | goto do_signed; | |
1003 | } | |
1004 | ||
1005 | /* s1 - u2 -> ur */ | |
1006 | if (code == MINUS_EXPR && !uns0_p && uns1_p && unsr_p) | |
1007 | { | |
1008 | /* Compute the operation. On RTL level, the addition is always | |
1009 | unsigned. */ | |
1010 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1011 | OPTAB_LIB_WIDEN); | |
1012 | int pos_neg = get_range_pos_neg (arg0); | |
1013 | if (pos_neg == 2) | |
1014 | /* If ARG0 is known to be always negative, this is always overflow. */ | |
1015 | emit_jump (do_error); | |
1016 | else if (pos_neg == 3) | |
1017 | /* If ARG0 is not known to be always positive, check at runtime. */ | |
92344ed0 | 1018 | do_compare_rtx_and_jump (op0, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1019 | NULL, do_error, profile_probability::very_unlikely ()); |
1476d1bd | 1020 | do_compare_rtx_and_jump (op1, op0, LEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1021 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1022 | goto do_error_label; |
1023 | } | |
1024 | ||
1025 | /* u1 - s2 -> sr */ | |
1026 | if (code == MINUS_EXPR && uns0_p && !uns1_p && !unsr_p) | |
1027 | { | |
1028 | /* Compute the operation. On RTL level, the addition is always | |
1029 | unsigned. */ | |
1030 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1031 | OPTAB_LIB_WIDEN); | |
1032 | rtx tem = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, | |
1033 | OPTAB_LIB_WIDEN); | |
1476d1bd | 1034 | do_compare_rtx_and_jump (op0, tem, LTU, true, mode, NULL_RTX, NULL, |
357067f2 | 1035 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1036 | goto do_error_label; |
1037 | } | |
1038 | ||
1039 | /* u1 + u2 -> sr */ | |
1040 | if (code == PLUS_EXPR && uns0_p && uns1_p && !unsr_p) | |
1041 | { | |
1042 | /* Compute the operation. On RTL level, the addition is always | |
1043 | unsigned. */ | |
1044 | res = expand_binop (mode, add_optab, op0, op1, NULL_RTX, false, | |
1045 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1046 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1047 | NULL, do_error, profile_probability::very_unlikely ()); |
1304953e JJ |
1048 | rtx tem = op1; |
1049 | /* The operation is commutative, so we can pick operand to compare | |
1050 | against. For prec <= BITS_PER_WORD, I think preferring REG operand | |
1051 | is better over CONST_INT, because the CONST_INT might enlarge the | |
1052 | instruction or CSE would need to figure out we'd already loaded it | |
1053 | into a register before. For prec > BITS_PER_WORD, I think CONST_INT | |
1054 | might be more beneficial, as then the multi-word comparison can be | |
1055 | perhaps simplified. */ | |
1056 | if (prec <= BITS_PER_WORD | |
1057 | ? (CONST_SCALAR_INT_P (op1) && REG_P (op0)) | |
1058 | : CONST_SCALAR_INT_P (op0)) | |
1059 | tem = op0; | |
1476d1bd | 1060 | do_compare_rtx_and_jump (res, tem, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1061 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1062 | goto do_error_label; |
1063 | } | |
1064 | ||
1065 | /* s1 +- s2 -> ur */ | |
1066 | if (!uns0_p && !uns1_p && unsr_p) | |
1067 | { | |
1068 | /* Compute the operation. On RTL level, the addition is always | |
1069 | unsigned. */ | |
1070 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1071 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1072 | int pos_neg = get_range_pos_neg (arg1); | |
1073 | if (code == PLUS_EXPR) | |
1074 | { | |
1075 | int pos_neg0 = get_range_pos_neg (arg0); | |
1076 | if (pos_neg0 != 3 && pos_neg == 3) | |
1077 | { | |
6b4db501 | 1078 | std::swap (op0, op1); |
1304953e JJ |
1079 | pos_neg = pos_neg0; |
1080 | } | |
1081 | } | |
1082 | rtx tem; | |
1083 | if (pos_neg != 3) | |
1084 | { | |
1085 | tem = expand_binop (mode, ((pos_neg == 1) ^ (code == MINUS_EXPR)) | |
1086 | ? and_optab : ior_optab, | |
1087 | op0, res, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1476d1bd | 1088 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL, |
357067f2 | 1089 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1090 | } |
1091 | else | |
1092 | { | |
1093 | rtx_code_label *do_ior_label = gen_label_rtx (); | |
92344ed0 JJ |
1094 | do_compare_rtx_and_jump (op1, const0_rtx, |
1095 | code == MINUS_EXPR ? GE : LT, false, mode, | |
1476d1bd | 1096 | NULL_RTX, NULL, do_ior_label, |
357067f2 | 1097 | profile_probability::even ()); |
1304953e JJ |
1098 | tem = expand_binop (mode, and_optab, op0, res, NULL_RTX, false, |
1099 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1100 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1101 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1102 | emit_jump (do_error); |
1103 | emit_label (do_ior_label); | |
1104 | tem = expand_binop (mode, ior_optab, op0, res, NULL_RTX, false, | |
1105 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1106 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1107 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1108 | } |
1109 | goto do_error_label; | |
1110 | } | |
1111 | ||
1112 | /* u1 - u2 -> sr */ | |
1113 | if (code == MINUS_EXPR && uns0_p && uns1_p && !unsr_p) | |
1114 | { | |
1115 | /* Compute the operation. On RTL level, the addition is always | |
1116 | unsigned. */ | |
1117 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1118 | OPTAB_LIB_WIDEN); | |
1119 | rtx_code_label *op0_geu_op1 = gen_label_rtx (); | |
1476d1bd | 1120 | do_compare_rtx_and_jump (op0, op1, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1121 | op0_geu_op1, profile_probability::even ()); |
92344ed0 | 1122 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1123 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1124 | emit_jump (do_error); |
1125 | emit_label (op0_geu_op1); | |
92344ed0 | 1126 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1127 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1128 | goto do_error_label; |
1129 | } | |
31e071ae | 1130 | |
1304953e JJ |
1131 | gcc_assert (!uns0_p && !uns1_p && !unsr_p); |
1132 | ||
1133 | /* s1 +- s2 -> sr */ | |
cde9d596 RH |
1134 | do_signed: |
1135 | { | |
1136 | insn_code icode = optab_handler (code == PLUS_EXPR ? addv4_optab | |
1137 | : subv4_optab, mode); | |
1138 | if (icode != CODE_FOR_nothing) | |
1139 | { | |
99b1c316 | 1140 | class expand_operand ops[4]; |
cde9d596 RH |
1141 | rtx_insn *last = get_last_insn (); |
1142 | ||
1143 | res = gen_reg_rtx (mode); | |
1144 | create_output_operand (&ops[0], res, mode); | |
1145 | create_input_operand (&ops[1], op0, mode); | |
1146 | create_input_operand (&ops[2], op1, mode); | |
1147 | create_fixed_operand (&ops[3], do_error); | |
1148 | if (maybe_expand_insn (icode, 4, ops)) | |
1149 | { | |
1150 | last = get_last_insn (); | |
1151 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT | |
1152 | && JUMP_P (last) | |
1153 | && any_condjump_p (last) | |
1154 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1155 | add_reg_br_prob_note (last, |
1156 | profile_probability::very_unlikely ()); | |
cde9d596 RH |
1157 | emit_jump (done_label); |
1158 | goto do_error_label; | |
1159 | } | |
1160 | ||
1161 | delete_insns_since (last); | |
1162 | } | |
1163 | ||
cde9d596 RH |
1164 | /* Compute the operation. On RTL level, the addition is always |
1165 | unsigned. */ | |
1166 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1167 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1168 | ||
47135167 | 1169 | /* If we can prove that one of the arguments (for MINUS_EXPR only |
cde9d596 RH |
1170 | the second operand, as subtraction is not commutative) is always |
1171 | non-negative or always negative, we can do just one comparison | |
47135167 EB |
1172 | and conditional jump. */ |
1173 | int pos_neg = get_range_pos_neg (arg1); | |
1174 | if (code == PLUS_EXPR) | |
cde9d596 | 1175 | { |
47135167 EB |
1176 | int pos_neg0 = get_range_pos_neg (arg0); |
1177 | if (pos_neg0 != 3 && pos_neg == 3) | |
1178 | { | |
1179 | std::swap (op0, op1); | |
1180 | pos_neg = pos_neg0; | |
1181 | } | |
cde9d596 | 1182 | } |
cde9d596 | 1183 | |
47135167 EB |
1184 | /* Addition overflows if and only if the two operands have the same sign, |
1185 | and the result has the opposite sign. Subtraction overflows if and | |
1186 | only if the two operands have opposite sign, and the subtrahend has | |
1187 | the same sign as the result. Here 0 is counted as positive. */ | |
cde9d596 | 1188 | if (pos_neg == 3) |
47135167 EB |
1189 | { |
1190 | /* Compute op0 ^ op1 (operands have opposite sign). */ | |
1191 | rtx op_xor = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, | |
1192 | OPTAB_LIB_WIDEN); | |
cde9d596 | 1193 | |
47135167 EB |
1194 | /* Compute res ^ op1 (result and 2nd operand have opposite sign). */ |
1195 | rtx res_xor = expand_binop (mode, xor_optab, res, op1, NULL_RTX, false, | |
1196 | OPTAB_LIB_WIDEN); | |
97286431 | 1197 | |
47135167 EB |
1198 | rtx tem; |
1199 | if (code == PLUS_EXPR) | |
1200 | { | |
1201 | /* Compute (res ^ op1) & ~(op0 ^ op1). */ | |
1202 | tem = expand_unop (mode, one_cmpl_optab, op_xor, NULL_RTX, false); | |
1203 | tem = expand_binop (mode, and_optab, res_xor, tem, NULL_RTX, false, | |
1204 | OPTAB_LIB_WIDEN); | |
1205 | } | |
1206 | else | |
1207 | { | |
1208 | /* Compute (op0 ^ op1) & ~(res ^ op1). */ | |
1209 | tem = expand_unop (mode, one_cmpl_optab, res_xor, NULL_RTX, false); | |
1210 | tem = expand_binop (mode, and_optab, op_xor, tem, NULL_RTX, false, | |
1211 | OPTAB_LIB_WIDEN); | |
1212 | } | |
1213 | ||
1214 | /* No overflow if the result has bit sign cleared. */ | |
1215 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, | |
357067f2 | 1216 | NULL, done_label, profile_probability::very_likely ()); |
cde9d596 | 1217 | } |
31e071ae | 1218 | |
47135167 EB |
1219 | /* Compare the result of the operation with the first operand. |
1220 | No overflow for addition if second operand is positive and result | |
1221 | is larger or second operand is negative and result is smaller. | |
1222 | Likewise for subtraction with sign of second operand flipped. */ | |
1223 | else | |
1224 | do_compare_rtx_and_jump (res, op0, | |
1225 | (pos_neg == 1) ^ (code == MINUS_EXPR) ? GE : LE, | |
cde9d596 | 1226 | false, mode, NULL_RTX, NULL, done_label, |
357067f2 | 1227 | profile_probability::very_likely ()); |
cde9d596 | 1228 | } |
31e071ae | 1229 | |
1304953e | 1230 | do_error_label: |
1769415d | 1231 | emit_label (do_error); |
1304953e JJ |
1232 | if (is_ubsan) |
1233 | { | |
1234 | /* Expand the ubsan builtin call. */ | |
1235 | push_temp_slots (); | |
1236 | fn = ubsan_build_overflow_builtin (code, loc, TREE_TYPE (arg0), | |
1705cebd | 1237 | arg0, arg1, datap); |
1304953e JJ |
1238 | expand_normal (fn); |
1239 | pop_temp_slots (); | |
1240 | do_pending_stack_adjust (); | |
1241 | } | |
1242 | else if (lhs) | |
a86451b9 | 1243 | expand_arith_set_overflow (lhs, target); |
31e071ae | 1244 | |
1769415d MP |
1245 | /* We're done. */ |
1246 | emit_label (done_label); | |
31e071ae MP |
1247 | |
1248 | if (lhs) | |
1304953e JJ |
1249 | { |
1250 | if (is_ubsan) | |
5620052d | 1251 | expand_ubsan_result_store (target, res); |
1304953e JJ |
1252 | else |
1253 | { | |
1254 | if (do_xor) | |
1255 | res = expand_binop (mode, add_optab, res, sgn, NULL_RTX, false, | |
1256 | OPTAB_LIB_WIDEN); | |
1257 | ||
1258 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
1259 | } | |
1260 | } | |
31e071ae MP |
1261 | } |
1262 | ||
1263 | /* Add negate overflow checking to the statement STMT. */ | |
1264 | ||
1304953e | 1265 | static void |
1705cebd JJ |
1266 | expand_neg_overflow (location_t loc, tree lhs, tree arg1, bool is_ubsan, |
1267 | tree *datap) | |
31e071ae MP |
1268 | { |
1269 | rtx res, op1; | |
1304953e | 1270 | tree fn; |
da664544 DM |
1271 | rtx_code_label *done_label, *do_error; |
1272 | rtx target = NULL_RTX; | |
31e071ae | 1273 | |
31e071ae MP |
1274 | done_label = gen_label_rtx (); |
1275 | do_error = gen_label_rtx (); | |
31e071ae MP |
1276 | |
1277 | do_pending_stack_adjust (); | |
1278 | op1 = expand_normal (arg1); | |
1279 | ||
7a504f33 | 1280 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg1)); |
31e071ae | 1281 | if (lhs) |
1304953e JJ |
1282 | { |
1283 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
1284 | if (!is_ubsan) | |
1285 | write_complex_part (target, const0_rtx, true); | |
1286 | } | |
31e071ae MP |
1287 | |
1288 | enum insn_code icode = optab_handler (negv3_optab, mode); | |
1289 | if (icode != CODE_FOR_nothing) | |
1290 | { | |
99b1c316 | 1291 | class expand_operand ops[3]; |
da664544 | 1292 | rtx_insn *last = get_last_insn (); |
31e071ae MP |
1293 | |
1294 | res = gen_reg_rtx (mode); | |
1295 | create_output_operand (&ops[0], res, mode); | |
1296 | create_input_operand (&ops[1], op1, mode); | |
1297 | create_fixed_operand (&ops[2], do_error); | |
1298 | if (maybe_expand_insn (icode, 3, ops)) | |
1299 | { | |
1300 | last = get_last_insn (); | |
0a6a6ac9 | 1301 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
31e071ae MP |
1302 | && JUMP_P (last) |
1303 | && any_condjump_p (last) | |
1304 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1305 | add_reg_br_prob_note (last, |
1306 | profile_probability::very_unlikely ()); | |
31e071ae MP |
1307 | emit_jump (done_label); |
1308 | } | |
1309 | else | |
1310 | { | |
1311 | delete_insns_since (last); | |
1312 | icode = CODE_FOR_nothing; | |
1313 | } | |
1314 | } | |
1315 | ||
1316 | if (icode == CODE_FOR_nothing) | |
1317 | { | |
1318 | /* Compute the operation. On RTL level, the addition is always | |
1319 | unsigned. */ | |
1320 | res = expand_unop (mode, neg_optab, op1, NULL_RTX, false); | |
1321 | ||
1322 | /* Compare the operand with the most negative value. */ | |
1323 | rtx minv = expand_normal (TYPE_MIN_VALUE (TREE_TYPE (arg1))); | |
1476d1bd | 1324 | do_compare_rtx_and_jump (op1, minv, NE, true, mode, NULL_RTX, NULL, |
357067f2 | 1325 | done_label, profile_probability::very_likely ()); |
31e071ae MP |
1326 | } |
1327 | ||
1328 | emit_label (do_error); | |
1304953e JJ |
1329 | if (is_ubsan) |
1330 | { | |
1331 | /* Expand the ubsan builtin call. */ | |
1332 | push_temp_slots (); | |
1333 | fn = ubsan_build_overflow_builtin (NEGATE_EXPR, loc, TREE_TYPE (arg1), | |
1705cebd | 1334 | arg1, NULL_TREE, datap); |
1304953e JJ |
1335 | expand_normal (fn); |
1336 | pop_temp_slots (); | |
1337 | do_pending_stack_adjust (); | |
1338 | } | |
1339 | else if (lhs) | |
a86451b9 | 1340 | expand_arith_set_overflow (lhs, target); |
31e071ae MP |
1341 | |
1342 | /* We're done. */ | |
1343 | emit_label (done_label); | |
1344 | ||
1345 | if (lhs) | |
1304953e JJ |
1346 | { |
1347 | if (is_ubsan) | |
5620052d | 1348 | expand_ubsan_result_store (target, res); |
1304953e JJ |
1349 | else |
1350 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
1351 | } | |
31e071ae MP |
1352 | } |
1353 | ||
0b99f253 JJ |
1354 | /* Return true if UNS WIDEN_MULT_EXPR with result mode WMODE and operand |
1355 | mode MODE can be expanded without using a libcall. */ | |
1356 | ||
1357 | static bool | |
1358 | can_widen_mult_without_libcall (scalar_int_mode wmode, scalar_int_mode mode, | |
1359 | rtx op0, rtx op1, bool uns) | |
1360 | { | |
1361 | if (find_widening_optab_handler (umul_widen_optab, wmode, mode) | |
1362 | != CODE_FOR_nothing) | |
1363 | return true; | |
1364 | ||
1365 | if (find_widening_optab_handler (smul_widen_optab, wmode, mode) | |
1366 | != CODE_FOR_nothing) | |
1367 | return true; | |
1368 | ||
1369 | rtx_insn *last = get_last_insn (); | |
1370 | if (CONSTANT_P (op0)) | |
1371 | op0 = convert_modes (wmode, mode, op0, uns); | |
1372 | else | |
1373 | op0 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 1); | |
1374 | if (CONSTANT_P (op1)) | |
1375 | op1 = convert_modes (wmode, mode, op1, uns); | |
1376 | else | |
1377 | op1 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 2); | |
1378 | rtx ret = expand_mult (wmode, op0, op1, NULL_RTX, uns, true); | |
1379 | delete_insns_since (last); | |
1380 | return ret != NULL_RTX; | |
1381 | } | |
1382 | ||
31e071ae MP |
1383 | /* Add mul overflow checking to the statement STMT. */ |
1384 | ||
1304953e JJ |
1385 | static void |
1386 | expand_mul_overflow (location_t loc, tree lhs, tree arg0, tree arg1, | |
1705cebd JJ |
1387 | bool unsr_p, bool uns0_p, bool uns1_p, bool is_ubsan, |
1388 | tree *datap) | |
31e071ae MP |
1389 | { |
1390 | rtx res, op0, op1; | |
1304953e | 1391 | tree fn, type; |
da664544 DM |
1392 | rtx_code_label *done_label, *do_error; |
1393 | rtx target = NULL_RTX; | |
1304953e JJ |
1394 | signop sign; |
1395 | enum insn_code icode; | |
31e071ae | 1396 | |
31e071ae MP |
1397 | done_label = gen_label_rtx (); |
1398 | do_error = gen_label_rtx (); | |
31e071ae MP |
1399 | |
1400 | do_pending_stack_adjust (); | |
1401 | op0 = expand_normal (arg0); | |
1402 | op1 = expand_normal (arg1); | |
1403 | ||
7a504f33 | 1404 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1304953e | 1405 | bool uns = unsr_p; |
31e071ae | 1406 | if (lhs) |
1304953e JJ |
1407 | { |
1408 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
1409 | if (!is_ubsan) | |
1410 | write_complex_part (target, const0_rtx, true); | |
1411 | } | |
1412 | ||
1413 | if (is_ubsan) | |
1414 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); | |
1415 | ||
1416 | /* We assume both operands and result have the same precision | |
1417 | here (GET_MODE_BITSIZE (mode)), S stands for signed type | |
1418 | with that precision, U for unsigned type with that precision, | |
1419 | sgn for unsigned most significant bit in that precision. | |
1420 | s1 is signed first operand, u1 is unsigned first operand, | |
1421 | s2 is signed second operand, u2 is unsigned second operand, | |
1422 | sr is signed result, ur is unsigned result and the following | |
1423 | rules say how to compute result (which is always result of | |
1424 | the operands as if both were unsigned, cast to the right | |
1425 | signedness) and how to compute whether operation overflowed. | |
1426 | main_ovf (false) stands for jump on signed multiplication | |
1427 | overflow or the main algorithm with uns == false. | |
1428 | main_ovf (true) stands for jump on unsigned multiplication | |
1429 | overflow or the main algorithm with uns == true. | |
1430 | ||
1431 | s1 * s2 -> sr | |
1432 | res = (S) ((U) s1 * (U) s2) | |
1433 | ovf = main_ovf (false) | |
1434 | u1 * u2 -> ur | |
1435 | res = u1 * u2 | |
1436 | ovf = main_ovf (true) | |
1437 | s1 * u2 -> ur | |
1438 | res = (U) s1 * u2 | |
1439 | ovf = (s1 < 0 && u2) || main_ovf (true) | |
1440 | u1 * u2 -> sr | |
1441 | res = (S) (u1 * u2) | |
1442 | ovf = res < 0 || main_ovf (true) | |
1443 | s1 * u2 -> sr | |
1444 | res = (S) ((U) s1 * u2) | |
1445 | ovf = (S) u2 >= 0 ? main_ovf (false) | |
1446 | : (s1 != 0 && (s1 != -1 || u2 != (U) res)) | |
1447 | s1 * s2 -> ur | |
1448 | t1 = (s1 & s2) < 0 ? (-(U) s1) : ((U) s1) | |
1449 | t2 = (s1 & s2) < 0 ? (-(U) s2) : ((U) s2) | |
1450 | res = t1 * t2 | |
1451 | ovf = (s1 ^ s2) < 0 ? (s1 && s2) : main_ovf (true) */ | |
1452 | ||
1453 | if (uns0_p && !uns1_p) | |
1454 | { | |
1455 | /* Multiplication is commutative, if operand signedness differs, | |
1456 | canonicalize to the first operand being signed and second | |
1457 | unsigned to simplify following code. */ | |
6b4db501 MM |
1458 | std::swap (op0, op1); |
1459 | std::swap (arg0, arg1); | |
1460 | uns0_p = false; | |
1461 | uns1_p = true; | |
1304953e JJ |
1462 | } |
1463 | ||
1464 | int pos_neg0 = get_range_pos_neg (arg0); | |
1465 | int pos_neg1 = get_range_pos_neg (arg1); | |
1466 | ||
1467 | /* s1 * u2 -> ur */ | |
1468 | if (!uns0_p && uns1_p && unsr_p) | |
1469 | { | |
1470 | switch (pos_neg0) | |
1471 | { | |
1472 | case 1: | |
1473 | /* If s1 is non-negative, just perform normal u1 * u2 -> ur. */ | |
1474 | goto do_main; | |
1475 | case 2: | |
1476 | /* If s1 is negative, avoid the main code, just multiply and | |
1477 | signal overflow if op1 is not 0. */ | |
1478 | struct separate_ops ops; | |
1479 | ops.code = MULT_EXPR; | |
1480 | ops.type = TREE_TYPE (arg1); | |
1481 | ops.op0 = make_tree (ops.type, op0); | |
1482 | ops.op1 = make_tree (ops.type, op1); | |
1483 | ops.op2 = NULL_TREE; | |
1484 | ops.location = loc; | |
1485 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
92344ed0 | 1486 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1487 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1488 | goto do_error_label; |
1489 | case 3: | |
049ce9d2 JJ |
1490 | if (get_min_precision (arg1, UNSIGNED) |
1491 | + get_min_precision (arg0, SIGNED) <= GET_MODE_PRECISION (mode)) | |
1492 | { | |
1493 | /* If the first operand is sign extended from narrower type, the | |
1494 | second operand is zero extended from narrower type and | |
1495 | the sum of the two precisions is smaller or equal to the | |
1496 | result precision: if the first argument is at runtime | |
1497 | non-negative, maximum result will be 0x7e81 or 0x7f..fe80..01 | |
1498 | and there will be no overflow, if the first argument is | |
1499 | negative and the second argument zero, the result will be | |
1500 | 0 and there will be no overflow, if the first argument is | |
1501 | negative and the second argument positive, the result when | |
1502 | treated as signed will be negative (minimum -0x7f80 or | |
1503 | -0x7f..f80..0) there there will be always overflow. So, do | |
1504 | res = (U) (s1 * u2) | |
1505 | ovf = (S) res < 0 */ | |
1506 | struct separate_ops ops; | |
1507 | ops.code = MULT_EXPR; | |
1508 | ops.type | |
1509 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1510 | 1); | |
1511 | ops.op0 = make_tree (ops.type, op0); | |
1512 | ops.op1 = make_tree (ops.type, op1); | |
1513 | ops.op2 = NULL_TREE; | |
1514 | ops.location = loc; | |
1515 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1516 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, | |
1517 | mode, NULL_RTX, NULL, done_label, | |
1518 | profile_probability::very_likely ()); | |
1519 | goto do_error_label; | |
1520 | } | |
1304953e JJ |
1521 | rtx_code_label *do_main_label; |
1522 | do_main_label = gen_label_rtx (); | |
92344ed0 | 1523 | do_compare_rtx_and_jump (op0, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1524 | NULL, do_main_label, profile_probability::very_likely ()); |
92344ed0 | 1525 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1526 | NULL, do_main_label, profile_probability::very_likely ()); |
a86451b9 | 1527 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
1528 | emit_label (do_main_label); |
1529 | goto do_main; | |
1530 | default: | |
1531 | gcc_unreachable (); | |
1532 | } | |
1533 | } | |
1534 | ||
1535 | /* u1 * u2 -> sr */ | |
1536 | if (uns0_p && uns1_p && !unsr_p) | |
1537 | { | |
049ce9d2 JJ |
1538 | if ((pos_neg0 | pos_neg1) == 1) |
1539 | { | |
1540 | /* If both arguments are zero extended from narrower types, | |
1541 | the MSB will be clear on both and so we can pretend it is | |
1542 | a normal s1 * s2 -> sr multiplication. */ | |
1543 | uns0_p = false; | |
1544 | uns1_p = false; | |
1545 | } | |
1546 | else | |
1547 | uns = true; | |
1304953e JJ |
1548 | /* Rest of handling of this case after res is computed. */ |
1549 | goto do_main; | |
1550 | } | |
1551 | ||
1552 | /* s1 * u2 -> sr */ | |
1553 | if (!uns0_p && uns1_p && !unsr_p) | |
1554 | { | |
1555 | switch (pos_neg1) | |
1556 | { | |
1557 | case 1: | |
1558 | goto do_main; | |
1559 | case 2: | |
1560 | /* If (S) u2 is negative (i.e. u2 is larger than maximum of S, | |
1561 | avoid the main code, just multiply and signal overflow | |
1562 | unless 0 * u2 or -1 * ((U) Smin). */ | |
1563 | struct separate_ops ops; | |
1564 | ops.code = MULT_EXPR; | |
1565 | ops.type = TREE_TYPE (arg1); | |
1566 | ops.op0 = make_tree (ops.type, op0); | |
1567 | ops.op1 = make_tree (ops.type, op1); | |
1568 | ops.op2 = NULL_TREE; | |
1569 | ops.location = loc; | |
1570 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
92344ed0 | 1571 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1572 | NULL, done_label, profile_probability::very_likely ()); |
92344ed0 | 1573 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
357067f2 | 1574 | NULL, do_error, profile_probability::very_unlikely ()); |
1304953e JJ |
1575 | int prec; |
1576 | prec = GET_MODE_PRECISION (mode); | |
1577 | rtx sgn; | |
1578 | sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); | |
92344ed0 | 1579 | do_compare_rtx_and_jump (op1, sgn, EQ, true, mode, NULL_RTX, |
357067f2 | 1580 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1581 | goto do_error_label; |
1582 | case 3: | |
1583 | /* Rest of handling of this case after res is computed. */ | |
1584 | goto do_main; | |
1585 | default: | |
1586 | gcc_unreachable (); | |
1587 | } | |
1588 | } | |
31e071ae | 1589 | |
1304953e JJ |
1590 | /* s1 * s2 -> ur */ |
1591 | if (!uns0_p && !uns1_p && unsr_p) | |
1592 | { | |
beeac4c2 | 1593 | rtx tem; |
1304953e JJ |
1594 | switch (pos_neg0 | pos_neg1) |
1595 | { | |
1596 | case 1: /* Both operands known to be non-negative. */ | |
1597 | goto do_main; | |
1598 | case 2: /* Both operands known to be negative. */ | |
1599 | op0 = expand_unop (mode, neg_optab, op0, NULL_RTX, false); | |
1600 | op1 = expand_unop (mode, neg_optab, op1, NULL_RTX, false); | |
1601 | /* Avoid looking at arg0/arg1 ranges, as we've changed | |
1602 | the arguments. */ | |
1603 | arg0 = error_mark_node; | |
1604 | arg1 = error_mark_node; | |
1605 | goto do_main; | |
1606 | case 3: | |
1607 | if ((pos_neg0 ^ pos_neg1) == 3) | |
1608 | { | |
1609 | /* If one operand is known to be negative and the other | |
1610 | non-negative, this overflows always, unless the non-negative | |
1611 | one is 0. Just do normal multiply and set overflow | |
1612 | unless one of the operands is 0. */ | |
1613 | struct separate_ops ops; | |
1614 | ops.code = MULT_EXPR; | |
1615 | ops.type | |
1616 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1617 | 1); | |
1618 | ops.op0 = make_tree (ops.type, op0); | |
1619 | ops.op1 = make_tree (ops.type, op1); | |
1620 | ops.op2 = NULL_TREE; | |
1621 | ops.location = loc; | |
1622 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
beeac4c2 JJ |
1623 | do_compare_rtx_and_jump (pos_neg0 == 1 ? op0 : op1, const0_rtx, EQ, |
1624 | true, mode, NULL_RTX, NULL, done_label, | |
357067f2 | 1625 | profile_probability::very_likely ()); |
1304953e JJ |
1626 | goto do_error_label; |
1627 | } | |
049ce9d2 JJ |
1628 | if (get_min_precision (arg0, SIGNED) |
1629 | + get_min_precision (arg1, SIGNED) <= GET_MODE_PRECISION (mode)) | |
1630 | { | |
1631 | /* If both operands are sign extended from narrower types and | |
1632 | the sum of the two precisions is smaller or equal to the | |
1633 | result precision: if both arguments are at runtime | |
1634 | non-negative, maximum result will be 0x3f01 or 0x3f..f0..01 | |
1635 | and there will be no overflow, if both arguments are negative, | |
1636 | maximum result will be 0x40..00 and there will be no overflow | |
1637 | either, if one argument is positive and the other argument | |
1638 | negative, the result when treated as signed will be negative | |
1639 | and there will be always overflow, and if one argument is | |
1640 | zero and the other negative the result will be zero and no | |
1641 | overflow. So, do | |
1642 | res = (U) (s1 * s2) | |
1643 | ovf = (S) res < 0 */ | |
1644 | struct separate_ops ops; | |
1645 | ops.code = MULT_EXPR; | |
1646 | ops.type | |
1647 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1648 | 1); | |
1649 | ops.op0 = make_tree (ops.type, op0); | |
1650 | ops.op1 = make_tree (ops.type, op1); | |
1651 | ops.op2 = NULL_TREE; | |
1652 | ops.location = loc; | |
1653 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1654 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, | |
1655 | mode, NULL_RTX, NULL, done_label, | |
1656 | profile_probability::very_likely ()); | |
1657 | goto do_error_label; | |
1658 | } | |
1304953e JJ |
1659 | /* The general case, do all the needed comparisons at runtime. */ |
1660 | rtx_code_label *do_main_label, *after_negate_label; | |
1661 | rtx rop0, rop1; | |
1662 | rop0 = gen_reg_rtx (mode); | |
1663 | rop1 = gen_reg_rtx (mode); | |
1664 | emit_move_insn (rop0, op0); | |
1665 | emit_move_insn (rop1, op1); | |
1666 | op0 = rop0; | |
1667 | op1 = rop1; | |
1668 | do_main_label = gen_label_rtx (); | |
1669 | after_negate_label = gen_label_rtx (); | |
1670 | tem = expand_binop (mode, and_optab, op0, op1, NULL_RTX, false, | |
1671 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1672 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1673 | NULL, after_negate_label, profile_probability::very_likely ()); |
1304953e JJ |
1674 | /* Both arguments negative here, negate them and continue with |
1675 | normal unsigned overflow checking multiplication. */ | |
1676 | emit_move_insn (op0, expand_unop (mode, neg_optab, op0, | |
1677 | NULL_RTX, false)); | |
1678 | emit_move_insn (op1, expand_unop (mode, neg_optab, op1, | |
1679 | NULL_RTX, false)); | |
1680 | /* Avoid looking at arg0/arg1 ranges, as we might have changed | |
1681 | the arguments. */ | |
1682 | arg0 = error_mark_node; | |
1683 | arg1 = error_mark_node; | |
1684 | emit_jump (do_main_label); | |
1685 | emit_label (after_negate_label); | |
beeac4c2 JJ |
1686 | tem = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1687 | OPTAB_LIB_WIDEN); | |
1688 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, | |
1689 | NULL, do_main_label, | |
1690 | profile_probability::very_likely ()); | |
1304953e JJ |
1691 | /* One argument is negative here, the other positive. This |
1692 | overflows always, unless one of the arguments is 0. But | |
1693 | if e.g. s2 is 0, (U) s1 * 0 doesn't overflow, whatever s1 | |
1694 | is, thus we can keep do_main code oring in overflow as is. */ | |
beeac4c2 JJ |
1695 | if (pos_neg0 != 2) |
1696 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, | |
1697 | NULL, do_main_label, | |
1698 | profile_probability::very_unlikely ()); | |
1699 | if (pos_neg1 != 2) | |
1700 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, | |
1701 | NULL, do_main_label, | |
1702 | profile_probability::very_unlikely ()); | |
a86451b9 | 1703 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
1704 | emit_label (do_main_label); |
1705 | goto do_main; | |
1706 | default: | |
1707 | gcc_unreachable (); | |
1708 | } | |
1709 | } | |
1710 | ||
1711 | do_main: | |
1712 | type = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), uns); | |
1713 | sign = uns ? UNSIGNED : SIGNED; | |
1714 | icode = optab_handler (uns ? umulv4_optab : mulv4_optab, mode); | |
89b1427f JJ |
1715 | if (uns |
1716 | && (integer_pow2p (arg0) || integer_pow2p (arg1)) | |
1717 | && (optimize_insn_for_speed_p () || icode == CODE_FOR_nothing)) | |
1718 | { | |
1719 | /* Optimize unsigned multiplication by power of 2 constant | |
1720 | using 2 shifts, one for result, one to extract the shifted | |
1721 | out bits to see if they are all zero. | |
1722 | Don't do this if optimizing for size and we have umulv4_optab, | |
1723 | in that case assume multiplication will be shorter. | |
1724 | This is heuristics based on the single target that provides | |
1725 | umulv4 right now (i?86/x86_64), if further targets add it, this | |
1726 | might need to be revisited. | |
1727 | Cases where both operands are constant should be folded already | |
1728 | during GIMPLE, and cases where one operand is constant but not | |
1729 | power of 2 are questionable, either the WIDEN_MULT_EXPR case | |
1730 | below can be done without multiplication, just by shifts and adds, | |
1731 | or we'd need to divide the result (and hope it actually doesn't | |
1732 | really divide nor multiply) and compare the result of the division | |
1733 | with the original operand. */ | |
1734 | rtx opn0 = op0; | |
1735 | rtx opn1 = op1; | |
1736 | tree argn0 = arg0; | |
1737 | tree argn1 = arg1; | |
1738 | if (integer_pow2p (arg0)) | |
1739 | { | |
1740 | std::swap (opn0, opn1); | |
1741 | std::swap (argn0, argn1); | |
1742 | } | |
1743 | int cnt = tree_log2 (argn1); | |
1744 | if (cnt >= 0 && cnt < GET_MODE_PRECISION (mode)) | |
1745 | { | |
1746 | rtx upper = const0_rtx; | |
1747 | res = expand_shift (LSHIFT_EXPR, mode, opn0, cnt, NULL_RTX, uns); | |
1748 | if (cnt != 0) | |
1749 | upper = expand_shift (RSHIFT_EXPR, mode, opn0, | |
1750 | GET_MODE_PRECISION (mode) - cnt, | |
1751 | NULL_RTX, uns); | |
1752 | do_compare_rtx_and_jump (upper, const0_rtx, EQ, true, mode, | |
1753 | NULL_RTX, NULL, done_label, | |
1754 | profile_probability::very_likely ()); | |
1755 | goto do_error_label; | |
1756 | } | |
1757 | } | |
31e071ae MP |
1758 | if (icode != CODE_FOR_nothing) |
1759 | { | |
99b1c316 | 1760 | class expand_operand ops[4]; |
da664544 | 1761 | rtx_insn *last = get_last_insn (); |
31e071ae MP |
1762 | |
1763 | res = gen_reg_rtx (mode); | |
1764 | create_output_operand (&ops[0], res, mode); | |
1765 | create_input_operand (&ops[1], op0, mode); | |
1766 | create_input_operand (&ops[2], op1, mode); | |
1767 | create_fixed_operand (&ops[3], do_error); | |
1768 | if (maybe_expand_insn (icode, 4, ops)) | |
1769 | { | |
1770 | last = get_last_insn (); | |
0a6a6ac9 | 1771 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
31e071ae MP |
1772 | && JUMP_P (last) |
1773 | && any_condjump_p (last) | |
1774 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1775 | add_reg_br_prob_note (last, |
1776 | profile_probability::very_unlikely ()); | |
31e071ae MP |
1777 | emit_jump (done_label); |
1778 | } | |
1779 | else | |
1780 | { | |
1781 | delete_insns_since (last); | |
1782 | icode = CODE_FOR_nothing; | |
1783 | } | |
1784 | } | |
1785 | ||
1786 | if (icode == CODE_FOR_nothing) | |
1787 | { | |
1788 | struct separate_ops ops; | |
1304953e | 1789 | int prec = GET_MODE_PRECISION (mode); |
095a2d76 | 1790 | scalar_int_mode hmode, wmode; |
1304953e JJ |
1791 | ops.op0 = make_tree (type, op0); |
1792 | ops.op1 = make_tree (type, op1); | |
31e071ae | 1793 | ops.op2 = NULL_TREE; |
1304953e | 1794 | ops.location = loc; |
0b99f253 JJ |
1795 | |
1796 | /* Optimize unsigned overflow check where we don't use the | |
1797 | multiplication result, just whether overflow happened. | |
1798 | If we can do MULT_HIGHPART_EXPR, that followed by | |
1799 | comparison of the result against zero is cheapest. | |
1800 | We'll still compute res, but it should be DCEd later. */ | |
1801 | use_operand_p use; | |
1802 | gimple *use_stmt; | |
1803 | if (!is_ubsan | |
1804 | && lhs | |
1805 | && uns | |
1806 | && !(uns0_p && uns1_p && !unsr_p) | |
1807 | && can_mult_highpart_p (mode, uns) == 1 | |
1808 | && single_imm_use (lhs, &use, &use_stmt) | |
1809 | && is_gimple_assign (use_stmt) | |
1810 | && gimple_assign_rhs_code (use_stmt) == IMAGPART_EXPR) | |
1811 | goto highpart; | |
1812 | ||
490d0f6c | 1813 | if (GET_MODE_2XWIDER_MODE (mode).exists (&wmode) |
0b99f253 JJ |
1814 | && targetm.scalar_mode_supported_p (wmode) |
1815 | && can_widen_mult_without_libcall (wmode, mode, op0, op1, uns)) | |
31e071ae | 1816 | { |
0b99f253 | 1817 | twoxwider: |
31e071ae MP |
1818 | ops.code = WIDEN_MULT_EXPR; |
1819 | ops.type | |
1304953e | 1820 | = build_nonstandard_integer_type (GET_MODE_PRECISION (wmode), uns); |
31e071ae MP |
1821 | |
1822 | res = expand_expr_real_2 (&ops, NULL_RTX, wmode, EXPAND_NORMAL); | |
1304953e JJ |
1823 | rtx hipart = expand_shift (RSHIFT_EXPR, wmode, res, prec, |
1824 | NULL_RTX, uns); | |
4ed543bc KC |
1825 | hipart = convert_modes (mode, wmode, hipart, uns); |
1826 | res = convert_modes (mode, wmode, res, uns); | |
1304953e JJ |
1827 | if (uns) |
1828 | /* For the unsigned multiplication, there was overflow if | |
1829 | HIPART is non-zero. */ | |
92344ed0 | 1830 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
1476d1bd | 1831 | NULL_RTX, NULL, done_label, |
357067f2 | 1832 | profile_probability::very_likely ()); |
1304953e JJ |
1833 | else |
1834 | { | |
2b4bbc17 RS |
1835 | /* RES is used more than once, place it in a pseudo. */ |
1836 | res = force_reg (mode, res); | |
1837 | ||
1304953e JJ |
1838 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
1839 | NULL_RTX, 0); | |
1840 | /* RES is low half of the double width result, HIPART | |
1841 | the high half. There was overflow if | |
1842 | HIPART is different from RES < 0 ? -1 : 0. */ | |
92344ed0 | 1843 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
1476d1bd | 1844 | NULL_RTX, NULL, done_label, |
357067f2 | 1845 | profile_probability::very_likely ()); |
1304953e | 1846 | } |
31e071ae | 1847 | } |
0b99f253 JJ |
1848 | else if (can_mult_highpart_p (mode, uns) == 1) |
1849 | { | |
1850 | highpart: | |
1851 | ops.code = MULT_HIGHPART_EXPR; | |
1852 | ops.type = type; | |
1853 | ||
1854 | rtx hipart = expand_expr_real_2 (&ops, NULL_RTX, mode, | |
1855 | EXPAND_NORMAL); | |
1856 | ops.code = MULT_EXPR; | |
1857 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1858 | if (uns) | |
1859 | /* For the unsigned multiplication, there was overflow if | |
1860 | HIPART is non-zero. */ | |
1861 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, | |
1862 | NULL_RTX, NULL, done_label, | |
1863 | profile_probability::very_likely ()); | |
1864 | else | |
1865 | { | |
1866 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, | |
1867 | NULL_RTX, 0); | |
1868 | /* RES is low half of the double width result, HIPART | |
1869 | the high half. There was overflow if | |
1870 | HIPART is different from RES < 0 ? -1 : 0. */ | |
1871 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, | |
1872 | NULL_RTX, NULL, done_label, | |
1873 | profile_probability::very_likely ()); | |
1874 | } | |
1875 | ||
1876 | } | |
fffbab82 RS |
1877 | else if (int_mode_for_size (prec / 2, 1).exists (&hmode) |
1878 | && 2 * GET_MODE_PRECISION (hmode) == prec) | |
d5fa9cc9 | 1879 | { |
da664544 DM |
1880 | rtx_code_label *large_op0 = gen_label_rtx (); |
1881 | rtx_code_label *small_op0_large_op1 = gen_label_rtx (); | |
1882 | rtx_code_label *one_small_one_large = gen_label_rtx (); | |
1883 | rtx_code_label *both_ops_large = gen_label_rtx (); | |
1304953e JJ |
1884 | rtx_code_label *after_hipart_neg = uns ? NULL : gen_label_rtx (); |
1885 | rtx_code_label *after_lopart_neg = uns ? NULL : gen_label_rtx (); | |
da664544 | 1886 | rtx_code_label *do_overflow = gen_label_rtx (); |
1304953e | 1887 | rtx_code_label *hipart_different = uns ? NULL : gen_label_rtx (); |
d5fa9cc9 | 1888 | |
807e902e | 1889 | unsigned int hprec = GET_MODE_PRECISION (hmode); |
d5fa9cc9 | 1890 | rtx hipart0 = expand_shift (RSHIFT_EXPR, mode, op0, hprec, |
1304953e | 1891 | NULL_RTX, uns); |
4ed543bc KC |
1892 | hipart0 = convert_modes (hmode, mode, hipart0, uns); |
1893 | rtx lopart0 = convert_modes (hmode, mode, op0, uns); | |
1304953e JJ |
1894 | rtx signbit0 = const0_rtx; |
1895 | if (!uns) | |
1896 | signbit0 = expand_shift (RSHIFT_EXPR, hmode, lopart0, hprec - 1, | |
1897 | NULL_RTX, 0); | |
d5fa9cc9 | 1898 | rtx hipart1 = expand_shift (RSHIFT_EXPR, mode, op1, hprec, |
1304953e | 1899 | NULL_RTX, uns); |
4ed543bc KC |
1900 | hipart1 = convert_modes (hmode, mode, hipart1, uns); |
1901 | rtx lopart1 = convert_modes (hmode, mode, op1, uns); | |
1304953e JJ |
1902 | rtx signbit1 = const0_rtx; |
1903 | if (!uns) | |
1904 | signbit1 = expand_shift (RSHIFT_EXPR, hmode, lopart1, hprec - 1, | |
1905 | NULL_RTX, 0); | |
d5fa9cc9 JJ |
1906 | |
1907 | res = gen_reg_rtx (mode); | |
1908 | ||
1909 | /* True if op0 resp. op1 are known to be in the range of | |
1910 | halfstype. */ | |
1911 | bool op0_small_p = false; | |
1912 | bool op1_small_p = false; | |
1913 | /* True if op0 resp. op1 are known to have all zeros or all ones | |
1914 | in the upper half of bits, but are not known to be | |
1915 | op{0,1}_small_p. */ | |
1916 | bool op0_medium_p = false; | |
1917 | bool op1_medium_p = false; | |
1918 | /* -1 if op{0,1} is known to be negative, 0 if it is known to be | |
1919 | nonnegative, 1 if unknown. */ | |
1920 | int op0_sign = 1; | |
1921 | int op1_sign = 1; | |
1922 | ||
1304953e JJ |
1923 | if (pos_neg0 == 1) |
1924 | op0_sign = 0; | |
1925 | else if (pos_neg0 == 2) | |
1926 | op0_sign = -1; | |
1927 | if (pos_neg1 == 1) | |
1928 | op1_sign = 0; | |
1929 | else if (pos_neg1 == 2) | |
1930 | op1_sign = -1; | |
1931 | ||
1932 | unsigned int mprec0 = prec; | |
1933 | if (arg0 != error_mark_node) | |
1934 | mprec0 = get_min_precision (arg0, sign); | |
1935 | if (mprec0 <= hprec) | |
1936 | op0_small_p = true; | |
1937 | else if (!uns && mprec0 <= hprec + 1) | |
1938 | op0_medium_p = true; | |
1939 | unsigned int mprec1 = prec; | |
1940 | if (arg1 != error_mark_node) | |
1941 | mprec1 = get_min_precision (arg1, sign); | |
1942 | if (mprec1 <= hprec) | |
1943 | op1_small_p = true; | |
1944 | else if (!uns && mprec1 <= hprec + 1) | |
1945 | op1_medium_p = true; | |
d5fa9cc9 JJ |
1946 | |
1947 | int smaller_sign = 1; | |
1948 | int larger_sign = 1; | |
1949 | if (op0_small_p) | |
1950 | { | |
1951 | smaller_sign = op0_sign; | |
1952 | larger_sign = op1_sign; | |
1953 | } | |
1954 | else if (op1_small_p) | |
1955 | { | |
1956 | smaller_sign = op1_sign; | |
1957 | larger_sign = op0_sign; | |
1958 | } | |
1959 | else if (op0_sign == op1_sign) | |
1960 | { | |
1961 | smaller_sign = op0_sign; | |
1962 | larger_sign = op0_sign; | |
1963 | } | |
1964 | ||
1965 | if (!op0_small_p) | |
92344ed0 | 1966 | do_compare_rtx_and_jump (signbit0, hipart0, NE, true, hmode, |
1476d1bd | 1967 | NULL_RTX, NULL, large_op0, |
357067f2 | 1968 | profile_probability::unlikely ()); |
d5fa9cc9 JJ |
1969 | |
1970 | if (!op1_small_p) | |
92344ed0 | 1971 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
1476d1bd | 1972 | NULL_RTX, NULL, small_op0_large_op1, |
357067f2 | 1973 | profile_probability::unlikely ()); |
d5fa9cc9 | 1974 | |
1304953e JJ |
1975 | /* If both op0 and op1 are sign (!uns) or zero (uns) extended from |
1976 | hmode to mode, the multiplication will never overflow. We can | |
1977 | do just one hmode x hmode => mode widening multiplication. */ | |
1304953e | 1978 | tree halfstype = build_nonstandard_integer_type (hprec, uns); |
0764a0d2 JJ |
1979 | ops.op0 = make_tree (halfstype, lopart0); |
1980 | ops.op1 = make_tree (halfstype, lopart1); | |
d5fa9cc9 | 1981 | ops.code = WIDEN_MULT_EXPR; |
1304953e | 1982 | ops.type = type; |
d5fa9cc9 JJ |
1983 | rtx thisres |
1984 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1985 | emit_move_insn (res, thisres); | |
1986 | emit_jump (done_label); | |
1987 | ||
1988 | emit_label (small_op0_large_op1); | |
1989 | ||
1304953e JJ |
1990 | /* If op0 is sign (!uns) or zero (uns) extended from hmode to mode, |
1991 | but op1 is not, just swap the arguments and handle it as op1 | |
1992 | sign/zero extended, op0 not. */ | |
d5fa9cc9 JJ |
1993 | rtx larger = gen_reg_rtx (mode); |
1994 | rtx hipart = gen_reg_rtx (hmode); | |
1995 | rtx lopart = gen_reg_rtx (hmode); | |
1996 | emit_move_insn (larger, op1); | |
1997 | emit_move_insn (hipart, hipart1); | |
1998 | emit_move_insn (lopart, lopart0); | |
1999 | emit_jump (one_small_one_large); | |
2000 | ||
2001 | emit_label (large_op0); | |
2002 | ||
2003 | if (!op1_small_p) | |
92344ed0 | 2004 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
1476d1bd | 2005 | NULL_RTX, NULL, both_ops_large, |
357067f2 | 2006 | profile_probability::unlikely ()); |
d5fa9cc9 | 2007 | |
1304953e JJ |
2008 | /* If op1 is sign (!uns) or zero (uns) extended from hmode to mode, |
2009 | but op0 is not, prepare larger, hipart and lopart pseudos and | |
2010 | handle it together with small_op0_large_op1. */ | |
d5fa9cc9 JJ |
2011 | emit_move_insn (larger, op0); |
2012 | emit_move_insn (hipart, hipart0); | |
2013 | emit_move_insn (lopart, lopart1); | |
2014 | ||
2015 | emit_label (one_small_one_large); | |
2016 | ||
2017 | /* lopart is the low part of the operand that is sign extended | |
026c3cfd | 2018 | to mode, larger is the other operand, hipart is the |
d5fa9cc9 JJ |
2019 | high part of larger and lopart0 and lopart1 are the low parts |
2020 | of both operands. | |
2021 | We perform lopart0 * lopart1 and lopart * hipart widening | |
2022 | multiplications. */ | |
2023 | tree halfutype = build_nonstandard_integer_type (hprec, 1); | |
2024 | ops.op0 = make_tree (halfutype, lopart0); | |
2025 | ops.op1 = make_tree (halfutype, lopart1); | |
2026 | rtx lo0xlo1 | |
2027 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2028 | ||
2029 | ops.op0 = make_tree (halfutype, lopart); | |
2030 | ops.op1 = make_tree (halfutype, hipart); | |
2031 | rtx loxhi = gen_reg_rtx (mode); | |
2032 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2033 | emit_move_insn (loxhi, tem); | |
2034 | ||
1304953e JJ |
2035 | if (!uns) |
2036 | { | |
2037 | /* if (hipart < 0) loxhi -= lopart << (bitsize / 2); */ | |
2038 | if (larger_sign == 0) | |
2039 | emit_jump (after_hipart_neg); | |
2040 | else if (larger_sign != -1) | |
92344ed0 | 2041 | do_compare_rtx_and_jump (hipart, const0_rtx, GE, false, hmode, |
1476d1bd | 2042 | NULL_RTX, NULL, after_hipart_neg, |
357067f2 | 2043 | profile_probability::even ()); |
1304953e JJ |
2044 | |
2045 | tem = convert_modes (mode, hmode, lopart, 1); | |
2046 | tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1); | |
2047 | tem = expand_simple_binop (mode, MINUS, loxhi, tem, NULL_RTX, | |
644f0132 | 2048 | 1, OPTAB_WIDEN); |
1304953e JJ |
2049 | emit_move_insn (loxhi, tem); |
2050 | ||
2051 | emit_label (after_hipart_neg); | |
2052 | ||
2053 | /* if (lopart < 0) loxhi -= larger; */ | |
2054 | if (smaller_sign == 0) | |
2055 | emit_jump (after_lopart_neg); | |
2056 | else if (smaller_sign != -1) | |
92344ed0 | 2057 | do_compare_rtx_and_jump (lopart, const0_rtx, GE, false, hmode, |
1476d1bd | 2058 | NULL_RTX, NULL, after_lopart_neg, |
357067f2 | 2059 | profile_probability::even ()); |
1304953e JJ |
2060 | |
2061 | tem = expand_simple_binop (mode, MINUS, loxhi, larger, NULL_RTX, | |
644f0132 | 2062 | 1, OPTAB_WIDEN); |
1304953e JJ |
2063 | emit_move_insn (loxhi, tem); |
2064 | ||
2065 | emit_label (after_lopart_neg); | |
2066 | } | |
d5fa9cc9 JJ |
2067 | |
2068 | /* loxhi += (uns) lo0xlo1 >> (bitsize / 2); */ | |
2069 | tem = expand_shift (RSHIFT_EXPR, mode, lo0xlo1, hprec, NULL_RTX, 1); | |
2070 | tem = expand_simple_binop (mode, PLUS, loxhi, tem, NULL_RTX, | |
644f0132 | 2071 | 1, OPTAB_WIDEN); |
d5fa9cc9 JJ |
2072 | emit_move_insn (loxhi, tem); |
2073 | ||
2074 | /* if (loxhi >> (bitsize / 2) | |
1304953e JJ |
2075 | == (hmode) loxhi >> (bitsize / 2 - 1)) (if !uns) |
2076 | if (loxhi >> (bitsize / 2) == 0 (if uns). */ | |
d5fa9cc9 JJ |
2077 | rtx hipartloxhi = expand_shift (RSHIFT_EXPR, mode, loxhi, hprec, |
2078 | NULL_RTX, 0); | |
4ed543bc | 2079 | hipartloxhi = convert_modes (hmode, mode, hipartloxhi, 0); |
1304953e JJ |
2080 | rtx signbitloxhi = const0_rtx; |
2081 | if (!uns) | |
2082 | signbitloxhi = expand_shift (RSHIFT_EXPR, hmode, | |
4ed543bc KC |
2083 | convert_modes (hmode, mode, |
2084 | loxhi, 0), | |
1304953e | 2085 | hprec - 1, NULL_RTX, 0); |
d5fa9cc9 | 2086 | |
92344ed0 | 2087 | do_compare_rtx_and_jump (signbitloxhi, hipartloxhi, NE, true, hmode, |
1476d1bd | 2088 | NULL_RTX, NULL, do_overflow, |
357067f2 | 2089 | profile_probability::very_unlikely ()); |
d5fa9cc9 JJ |
2090 | |
2091 | /* res = (loxhi << (bitsize / 2)) | (hmode) lo0xlo1; */ | |
2092 | rtx loxhishifted = expand_shift (LSHIFT_EXPR, mode, loxhi, hprec, | |
2093 | NULL_RTX, 1); | |
4ed543bc KC |
2094 | tem = convert_modes (mode, hmode, |
2095 | convert_modes (hmode, mode, lo0xlo1, 1), 1); | |
d5fa9cc9 JJ |
2096 | |
2097 | tem = expand_simple_binop (mode, IOR, loxhishifted, tem, res, | |
644f0132 | 2098 | 1, OPTAB_WIDEN); |
d5fa9cc9 JJ |
2099 | if (tem != res) |
2100 | emit_move_insn (res, tem); | |
2101 | emit_jump (done_label); | |
2102 | ||
2103 | emit_label (both_ops_large); | |
2104 | ||
1304953e JJ |
2105 | /* If both operands are large (not sign (!uns) or zero (uns) |
2106 | extended from hmode), then perform the full multiplication | |
2107 | which will be the result of the operation. | |
2108 | The only cases which don't overflow are for signed multiplication | |
2109 | some cases where both hipart0 and highpart1 are 0 or -1. | |
2110 | For unsigned multiplication when high parts are both non-zero | |
2111 | this overflows always. */ | |
d5fa9cc9 | 2112 | ops.code = MULT_EXPR; |
1304953e JJ |
2113 | ops.op0 = make_tree (type, op0); |
2114 | ops.op1 = make_tree (type, op1); | |
d5fa9cc9 JJ |
2115 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2116 | emit_move_insn (res, tem); | |
2117 | ||
1304953e | 2118 | if (!uns) |
d5fa9cc9 | 2119 | { |
1304953e JJ |
2120 | if (!op0_medium_p) |
2121 | { | |
2122 | tem = expand_simple_binop (hmode, PLUS, hipart0, const1_rtx, | |
644f0132 | 2123 | NULL_RTX, 1, OPTAB_WIDEN); |
92344ed0 | 2124 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
1476d1bd | 2125 | NULL_RTX, NULL, do_error, |
357067f2 | 2126 | profile_probability::very_unlikely ()); |
1304953e | 2127 | } |
d5fa9cc9 | 2128 | |
1304953e JJ |
2129 | if (!op1_medium_p) |
2130 | { | |
2131 | tem = expand_simple_binop (hmode, PLUS, hipart1, const1_rtx, | |
644f0132 | 2132 | NULL_RTX, 1, OPTAB_WIDEN); |
92344ed0 | 2133 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
1476d1bd | 2134 | NULL_RTX, NULL, do_error, |
357067f2 | 2135 | profile_probability::very_unlikely ()); |
1304953e | 2136 | } |
d5fa9cc9 | 2137 | |
1304953e | 2138 | /* At this point hipart{0,1} are both in [-1, 0]. If they are |
e7176f75 JJ |
2139 | the same, overflow happened if res is non-positive, if they |
2140 | are different, overflow happened if res is positive. */ | |
1304953e JJ |
2141 | if (op0_sign != 1 && op1_sign != 1 && op0_sign != op1_sign) |
2142 | emit_jump (hipart_different); | |
2143 | else if (op0_sign == 1 || op1_sign == 1) | |
92344ed0 | 2144 | do_compare_rtx_and_jump (hipart0, hipart1, NE, true, hmode, |
1476d1bd | 2145 | NULL_RTX, NULL, hipart_different, |
357067f2 | 2146 | profile_probability::even ()); |
d5fa9cc9 | 2147 | |
e7176f75 | 2148 | do_compare_rtx_and_jump (res, const0_rtx, LE, false, mode, |
1476d1bd | 2149 | NULL_RTX, NULL, do_error, |
357067f2 | 2150 | profile_probability::very_unlikely ()); |
1304953e | 2151 | emit_jump (done_label); |
d5fa9cc9 | 2152 | |
1304953e JJ |
2153 | emit_label (hipart_different); |
2154 | ||
92344ed0 | 2155 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, |
1476d1bd | 2156 | NULL_RTX, NULL, do_error, |
357067f2 | 2157 | profile_probability::very_unlikely ()); |
1304953e JJ |
2158 | emit_jump (done_label); |
2159 | } | |
d5fa9cc9 JJ |
2160 | |
2161 | emit_label (do_overflow); | |
2162 | ||
2163 | /* Overflow, do full multiplication and fallthru into do_error. */ | |
1304953e JJ |
2164 | ops.op0 = make_tree (type, op0); |
2165 | ops.op1 = make_tree (type, op1); | |
d5fa9cc9 JJ |
2166 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2167 | emit_move_insn (res, tem); | |
2168 | } | |
0b99f253 JJ |
2169 | else if (GET_MODE_2XWIDER_MODE (mode).exists (&wmode) |
2170 | && targetm.scalar_mode_supported_p (wmode)) | |
2171 | /* Even emitting a libcall is better than not detecting overflow | |
2172 | at all. */ | |
2173 | goto twoxwider; | |
31e071ae MP |
2174 | else |
2175 | { | |
1304953e | 2176 | gcc_assert (!is_ubsan); |
31e071ae | 2177 | ops.code = MULT_EXPR; |
1304953e | 2178 | ops.type = type; |
31e071ae MP |
2179 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2180 | emit_jump (done_label); | |
2181 | } | |
2182 | } | |
2183 | ||
1304953e | 2184 | do_error_label: |
31e071ae | 2185 | emit_label (do_error); |
1304953e JJ |
2186 | if (is_ubsan) |
2187 | { | |
2188 | /* Expand the ubsan builtin call. */ | |
2189 | push_temp_slots (); | |
2190 | fn = ubsan_build_overflow_builtin (MULT_EXPR, loc, TREE_TYPE (arg0), | |
1705cebd | 2191 | arg0, arg1, datap); |
1304953e JJ |
2192 | expand_normal (fn); |
2193 | pop_temp_slots (); | |
2194 | do_pending_stack_adjust (); | |
2195 | } | |
2196 | else if (lhs) | |
a86451b9 | 2197 | expand_arith_set_overflow (lhs, target); |
31e071ae MP |
2198 | |
2199 | /* We're done. */ | |
2200 | emit_label (done_label); | |
2201 | ||
1304953e JJ |
2202 | /* u1 * u2 -> sr */ |
2203 | if (uns0_p && uns1_p && !unsr_p) | |
2204 | { | |
2205 | rtx_code_label *all_done_label = gen_label_rtx (); | |
92344ed0 | 2206 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 2207 | NULL, all_done_label, profile_probability::very_likely ()); |
a86451b9 | 2208 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
2209 | emit_label (all_done_label); |
2210 | } | |
2211 | ||
2212 | /* s1 * u2 -> sr */ | |
2213 | if (!uns0_p && uns1_p && !unsr_p && pos_neg1 == 3) | |
2214 | { | |
2215 | rtx_code_label *all_done_label = gen_label_rtx (); | |
2216 | rtx_code_label *set_noovf = gen_label_rtx (); | |
92344ed0 | 2217 | do_compare_rtx_and_jump (op1, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 2218 | NULL, all_done_label, profile_probability::very_likely ()); |
a86451b9 | 2219 | expand_arith_set_overflow (lhs, target); |
92344ed0 | 2220 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 2221 | NULL, set_noovf, profile_probability::very_likely ()); |
92344ed0 | 2222 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
357067f2 | 2223 | NULL, all_done_label, profile_probability::very_unlikely ()); |
1476d1bd | 2224 | do_compare_rtx_and_jump (op1, res, NE, true, mode, NULL_RTX, NULL, |
357067f2 | 2225 | all_done_label, profile_probability::very_unlikely ()); |
1304953e JJ |
2226 | emit_label (set_noovf); |
2227 | write_complex_part (target, const0_rtx, true); | |
2228 | emit_label (all_done_label); | |
2229 | } | |
2230 | ||
31e071ae | 2231 | if (lhs) |
1304953e JJ |
2232 | { |
2233 | if (is_ubsan) | |
5620052d | 2234 | expand_ubsan_result_store (target, res); |
1304953e JJ |
2235 | else |
2236 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
2237 | } | |
31e071ae MP |
2238 | } |
2239 | ||
1705cebd JJ |
2240 | /* Expand UBSAN_CHECK_* internal function if it has vector operands. */ |
2241 | ||
2242 | static void | |
2243 | expand_vector_ubsan_overflow (location_t loc, enum tree_code code, tree lhs, | |
2244 | tree arg0, tree arg1) | |
2245 | { | |
07626e49 | 2246 | poly_uint64 cnt = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)); |
1705cebd JJ |
2247 | rtx_code_label *loop_lab = NULL; |
2248 | rtx cntvar = NULL_RTX; | |
2249 | tree cntv = NULL_TREE; | |
2250 | tree eltype = TREE_TYPE (TREE_TYPE (arg0)); | |
2251 | tree sz = TYPE_SIZE (eltype); | |
2252 | tree data = NULL_TREE; | |
2253 | tree resv = NULL_TREE; | |
2254 | rtx lhsr = NULL_RTX; | |
2255 | rtx resvr = NULL_RTX; | |
07626e49 RS |
2256 | unsigned HOST_WIDE_INT const_cnt = 0; |
2257 | bool use_loop_p = (!cnt.is_constant (&const_cnt) || const_cnt > 4); | |
1705cebd JJ |
2258 | |
2259 | if (lhs) | |
2260 | { | |
2261 | optab op; | |
2262 | lhsr = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
6a5cdb0e | 2263 | if (!VECTOR_MODE_P (GET_MODE (lhsr)) |
1705cebd JJ |
2264 | || (op = optab_for_tree_code (code, TREE_TYPE (arg0), |
2265 | optab_default)) == unknown_optab | |
2266 | || (optab_handler (op, TYPE_MODE (TREE_TYPE (arg0))) | |
2267 | == CODE_FOR_nothing)) | |
2268 | { | |
2269 | if (MEM_P (lhsr)) | |
2270 | resv = make_tree (TREE_TYPE (lhs), lhsr); | |
2271 | else | |
2272 | { | |
2273 | resvr = assign_temp (TREE_TYPE (lhs), 1, 1); | |
2274 | resv = make_tree (TREE_TYPE (lhs), resvr); | |
2275 | } | |
2276 | } | |
2277 | } | |
07626e49 | 2278 | if (use_loop_p) |
1705cebd JJ |
2279 | { |
2280 | do_pending_stack_adjust (); | |
2281 | loop_lab = gen_label_rtx (); | |
2282 | cntvar = gen_reg_rtx (TYPE_MODE (sizetype)); | |
2283 | cntv = make_tree (sizetype, cntvar); | |
2284 | emit_move_insn (cntvar, const0_rtx); | |
2285 | emit_label (loop_lab); | |
2286 | } | |
2287 | if (TREE_CODE (arg0) != VECTOR_CST) | |
2288 | { | |
2289 | rtx arg0r = expand_normal (arg0); | |
2290 | arg0 = make_tree (TREE_TYPE (arg0), arg0r); | |
2291 | } | |
2292 | if (TREE_CODE (arg1) != VECTOR_CST) | |
2293 | { | |
2294 | rtx arg1r = expand_normal (arg1); | |
2295 | arg1 = make_tree (TREE_TYPE (arg1), arg1r); | |
2296 | } | |
07626e49 | 2297 | for (unsigned int i = 0; i < (use_loop_p ? 1 : const_cnt); i++) |
1705cebd JJ |
2298 | { |
2299 | tree op0, op1, res = NULL_TREE; | |
07626e49 | 2300 | if (use_loop_p) |
1705cebd JJ |
2301 | { |
2302 | tree atype = build_array_type_nelts (eltype, cnt); | |
4b48e883 JJ |
2303 | op0 = uniform_vector_p (arg0); |
2304 | if (op0 == NULL_TREE) | |
2305 | { | |
2306 | op0 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg0); | |
2307 | op0 = build4_loc (loc, ARRAY_REF, eltype, op0, cntv, | |
2308 | NULL_TREE, NULL_TREE); | |
2309 | } | |
2310 | op1 = uniform_vector_p (arg1); | |
2311 | if (op1 == NULL_TREE) | |
2312 | { | |
2313 | op1 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg1); | |
2314 | op1 = build4_loc (loc, ARRAY_REF, eltype, op1, cntv, | |
2315 | NULL_TREE, NULL_TREE); | |
2316 | } | |
1705cebd JJ |
2317 | if (resv) |
2318 | { | |
2319 | res = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, resv); | |
2320 | res = build4_loc (loc, ARRAY_REF, eltype, res, cntv, | |
2321 | NULL_TREE, NULL_TREE); | |
2322 | } | |
2323 | } | |
2324 | else | |
2325 | { | |
2326 | tree bitpos = bitsize_int (tree_to_uhwi (sz) * i); | |
2327 | op0 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg0, sz, bitpos); | |
2328 | op1 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg1, sz, bitpos); | |
2329 | if (resv) | |
2330 | res = fold_build3_loc (loc, BIT_FIELD_REF, eltype, resv, sz, | |
2331 | bitpos); | |
2332 | } | |
2333 | switch (code) | |
2334 | { | |
2335 | case PLUS_EXPR: | |
2336 | expand_addsub_overflow (loc, PLUS_EXPR, res, op0, op1, | |
2337 | false, false, false, true, &data); | |
2338 | break; | |
2339 | case MINUS_EXPR: | |
07626e49 | 2340 | if (use_loop_p ? integer_zerop (arg0) : integer_zerop (op0)) |
1705cebd JJ |
2341 | expand_neg_overflow (loc, res, op1, true, &data); |
2342 | else | |
2343 | expand_addsub_overflow (loc, MINUS_EXPR, res, op0, op1, | |
2344 | false, false, false, true, &data); | |
2345 | break; | |
2346 | case MULT_EXPR: | |
2347 | expand_mul_overflow (loc, res, op0, op1, false, false, false, | |
2348 | true, &data); | |
2349 | break; | |
2350 | default: | |
2351 | gcc_unreachable (); | |
2352 | } | |
2353 | } | |
07626e49 | 2354 | if (use_loop_p) |
1705cebd JJ |
2355 | { |
2356 | struct separate_ops ops; | |
2357 | ops.code = PLUS_EXPR; | |
2358 | ops.type = TREE_TYPE (cntv); | |
2359 | ops.op0 = cntv; | |
2360 | ops.op1 = build_int_cst (TREE_TYPE (cntv), 1); | |
2361 | ops.op2 = NULL_TREE; | |
2362 | ops.location = loc; | |
2363 | rtx ret = expand_expr_real_2 (&ops, cntvar, TYPE_MODE (sizetype), | |
2364 | EXPAND_NORMAL); | |
2365 | if (ret != cntvar) | |
2366 | emit_move_insn (cntvar, ret); | |
07626e49 RS |
2367 | rtx cntrtx = gen_int_mode (cnt, TYPE_MODE (sizetype)); |
2368 | do_compare_rtx_and_jump (cntvar, cntrtx, NE, false, | |
1705cebd | 2369 | TYPE_MODE (sizetype), NULL_RTX, NULL, loop_lab, |
357067f2 | 2370 | profile_probability::very_likely ()); |
1705cebd JJ |
2371 | } |
2372 | if (lhs && resv == NULL_TREE) | |
2373 | { | |
2374 | struct separate_ops ops; | |
2375 | ops.code = code; | |
2376 | ops.type = TREE_TYPE (arg0); | |
2377 | ops.op0 = arg0; | |
2378 | ops.op1 = arg1; | |
2379 | ops.op2 = NULL_TREE; | |
2380 | ops.location = loc; | |
2381 | rtx ret = expand_expr_real_2 (&ops, lhsr, TYPE_MODE (TREE_TYPE (arg0)), | |
2382 | EXPAND_NORMAL); | |
2383 | if (ret != lhsr) | |
2384 | emit_move_insn (lhsr, ret); | |
2385 | } | |
2386 | else if (resvr) | |
2387 | emit_move_insn (lhsr, resvr); | |
2388 | } | |
2389 | ||
31e071ae MP |
2390 | /* Expand UBSAN_CHECK_ADD call STMT. */ |
2391 | ||
2392 | static void | |
4cfe7a6c | 2393 | expand_UBSAN_CHECK_ADD (internal_fn, gcall *stmt) |
31e071ae | 2394 | { |
1304953e JJ |
2395 | location_t loc = gimple_location (stmt); |
2396 | tree lhs = gimple_call_lhs (stmt); | |
2397 | tree arg0 = gimple_call_arg (stmt, 0); | |
2398 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2399 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2400 | expand_vector_ubsan_overflow (loc, PLUS_EXPR, lhs, arg0, arg1); | |
2401 | else | |
2402 | expand_addsub_overflow (loc, PLUS_EXPR, lhs, arg0, arg1, | |
2403 | false, false, false, true, NULL); | |
31e071ae MP |
2404 | } |
2405 | ||
2406 | /* Expand UBSAN_CHECK_SUB call STMT. */ | |
2407 | ||
2408 | static void | |
4cfe7a6c | 2409 | expand_UBSAN_CHECK_SUB (internal_fn, gcall *stmt) |
31e071ae | 2410 | { |
1304953e JJ |
2411 | location_t loc = gimple_location (stmt); |
2412 | tree lhs = gimple_call_lhs (stmt); | |
2413 | tree arg0 = gimple_call_arg (stmt, 0); | |
2414 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2415 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2416 | expand_vector_ubsan_overflow (loc, MINUS_EXPR, lhs, arg0, arg1); | |
2417 | else if (integer_zerop (arg0)) | |
2418 | expand_neg_overflow (loc, lhs, arg1, true, NULL); | |
31e071ae | 2419 | else |
1304953e | 2420 | expand_addsub_overflow (loc, MINUS_EXPR, lhs, arg0, arg1, |
1705cebd | 2421 | false, false, false, true, NULL); |
31e071ae MP |
2422 | } |
2423 | ||
2424 | /* Expand UBSAN_CHECK_MUL call STMT. */ | |
2425 | ||
2426 | static void | |
4cfe7a6c | 2427 | expand_UBSAN_CHECK_MUL (internal_fn, gcall *stmt) |
31e071ae | 2428 | { |
1304953e JJ |
2429 | location_t loc = gimple_location (stmt); |
2430 | tree lhs = gimple_call_lhs (stmt); | |
2431 | tree arg0 = gimple_call_arg (stmt, 0); | |
2432 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2433 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2434 | expand_vector_ubsan_overflow (loc, MULT_EXPR, lhs, arg0, arg1); | |
2435 | else | |
2436 | expand_mul_overflow (loc, lhs, arg0, arg1, false, false, false, true, | |
2437 | NULL); | |
1304953e JJ |
2438 | } |
2439 | ||
2440 | /* Helper function for {ADD,SUB,MUL}_OVERFLOW call stmt expansion. */ | |
2441 | ||
2442 | static void | |
355fe088 | 2443 | expand_arith_overflow (enum tree_code code, gimple *stmt) |
1304953e JJ |
2444 | { |
2445 | tree lhs = gimple_call_lhs (stmt); | |
2446 | if (lhs == NULL_TREE) | |
2447 | return; | |
2448 | tree arg0 = gimple_call_arg (stmt, 0); | |
2449 | tree arg1 = gimple_call_arg (stmt, 1); | |
2450 | tree type = TREE_TYPE (TREE_TYPE (lhs)); | |
2451 | int uns0_p = TYPE_UNSIGNED (TREE_TYPE (arg0)); | |
2452 | int uns1_p = TYPE_UNSIGNED (TREE_TYPE (arg1)); | |
2453 | int unsr_p = TYPE_UNSIGNED (type); | |
2454 | int prec0 = TYPE_PRECISION (TREE_TYPE (arg0)); | |
2455 | int prec1 = TYPE_PRECISION (TREE_TYPE (arg1)); | |
2456 | int precres = TYPE_PRECISION (type); | |
2457 | location_t loc = gimple_location (stmt); | |
2458 | if (!uns0_p && get_range_pos_neg (arg0) == 1) | |
2459 | uns0_p = true; | |
2460 | if (!uns1_p && get_range_pos_neg (arg1) == 1) | |
2461 | uns1_p = true; | |
2462 | int pr = get_min_precision (arg0, uns0_p ? UNSIGNED : SIGNED); | |
2463 | prec0 = MIN (prec0, pr); | |
2464 | pr = get_min_precision (arg1, uns1_p ? UNSIGNED : SIGNED); | |
2465 | prec1 = MIN (prec1, pr); | |
2466 | ||
2467 | /* If uns0_p && uns1_p, precop is minimum needed precision | |
2468 | of unsigned type to hold the exact result, otherwise | |
2469 | precop is minimum needed precision of signed type to | |
2470 | hold the exact result. */ | |
2471 | int precop; | |
2472 | if (code == MULT_EXPR) | |
2473 | precop = prec0 + prec1 + (uns0_p != uns1_p); | |
2474 | else | |
2475 | { | |
2476 | if (uns0_p == uns1_p) | |
2477 | precop = MAX (prec0, prec1) + 1; | |
2478 | else if (uns0_p) | |
2479 | precop = MAX (prec0 + 1, prec1) + 1; | |
2480 | else | |
2481 | precop = MAX (prec0, prec1 + 1) + 1; | |
2482 | } | |
2483 | int orig_precres = precres; | |
2484 | ||
2485 | do | |
2486 | { | |
2487 | if ((uns0_p && uns1_p) | |
2488 | ? ((precop + !unsr_p) <= precres | |
2489 | /* u1 - u2 -> ur can overflow, no matter what precision | |
2490 | the result has. */ | |
2491 | && (code != MINUS_EXPR || !unsr_p)) | |
2492 | : (!unsr_p && precop <= precres)) | |
2493 | { | |
2494 | /* The infinity precision result will always fit into result. */ | |
2495 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2496 | write_complex_part (target, const0_rtx, true); | |
7a504f33 | 2497 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (type); |
1304953e JJ |
2498 | struct separate_ops ops; |
2499 | ops.code = code; | |
2500 | ops.type = type; | |
2501 | ops.op0 = fold_convert_loc (loc, type, arg0); | |
2502 | ops.op1 = fold_convert_loc (loc, type, arg1); | |
2503 | ops.op2 = NULL_TREE; | |
2504 | ops.location = loc; | |
2505 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2506 | expand_arith_overflow_result_store (lhs, target, mode, tem); | |
2507 | return; | |
2508 | } | |
2509 | ||
894d8b41 EB |
2510 | /* For operations with low precision, if target doesn't have them, start |
2511 | with precres widening right away, otherwise do it only if the most | |
2512 | simple cases can't be used. */ | |
2513 | const int min_precision = targetm.min_arithmetic_precision (); | |
2514 | if (orig_precres == precres && precres < min_precision) | |
1304953e | 2515 | ; |
9e11bfef TS |
2516 | else if ((uns0_p && uns1_p && unsr_p && prec0 <= precres |
2517 | && prec1 <= precres) | |
1304953e JJ |
2518 | || ((!uns0_p || !uns1_p) && !unsr_p |
2519 | && prec0 + uns0_p <= precres | |
2520 | && prec1 + uns1_p <= precres)) | |
2521 | { | |
2522 | arg0 = fold_convert_loc (loc, type, arg0); | |
2523 | arg1 = fold_convert_loc (loc, type, arg1); | |
2524 | switch (code) | |
2525 | { | |
2526 | case MINUS_EXPR: | |
2527 | if (integer_zerop (arg0) && !unsr_p) | |
7d704548 | 2528 | { |
1705cebd | 2529 | expand_neg_overflow (loc, lhs, arg1, false, NULL); |
7d704548 JJ |
2530 | return; |
2531 | } | |
1304953e JJ |
2532 | /* FALLTHRU */ |
2533 | case PLUS_EXPR: | |
1705cebd JJ |
2534 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2535 | unsr_p, unsr_p, false, NULL); | |
1304953e JJ |
2536 | return; |
2537 | case MULT_EXPR: | |
1705cebd JJ |
2538 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2539 | unsr_p, unsr_p, false, NULL); | |
1304953e JJ |
2540 | return; |
2541 | default: | |
2542 | gcc_unreachable (); | |
2543 | } | |
2544 | } | |
2545 | ||
2546 | /* For sub-word operations, retry with a wider type first. */ | |
2547 | if (orig_precres == precres && precop <= BITS_PER_WORD) | |
2548 | { | |
894d8b41 | 2549 | int p = MAX (min_precision, precop); |
f67f4dff | 2550 | scalar_int_mode m = smallest_int_mode_for_size (p); |
1304953e JJ |
2551 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (m), |
2552 | uns0_p && uns1_p | |
2553 | && unsr_p); | |
2554 | p = TYPE_PRECISION (optype); | |
2555 | if (p > precres) | |
2556 | { | |
2557 | precres = p; | |
2558 | unsr_p = TYPE_UNSIGNED (optype); | |
2559 | type = optype; | |
2560 | continue; | |
2561 | } | |
2562 | } | |
2563 | ||
2564 | if (prec0 <= precres && prec1 <= precres) | |
2565 | { | |
2566 | tree types[2]; | |
2567 | if (unsr_p) | |
2568 | { | |
2569 | types[0] = build_nonstandard_integer_type (precres, 0); | |
2570 | types[1] = type; | |
2571 | } | |
2572 | else | |
2573 | { | |
2574 | types[0] = type; | |
2575 | types[1] = build_nonstandard_integer_type (precres, 1); | |
2576 | } | |
2577 | arg0 = fold_convert_loc (loc, types[uns0_p], arg0); | |
2578 | arg1 = fold_convert_loc (loc, types[uns1_p], arg1); | |
2579 | if (code != MULT_EXPR) | |
2580 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, | |
1705cebd | 2581 | uns0_p, uns1_p, false, NULL); |
1304953e JJ |
2582 | else |
2583 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, | |
1705cebd | 2584 | uns0_p, uns1_p, false, NULL); |
1304953e JJ |
2585 | return; |
2586 | } | |
2587 | ||
2588 | /* Retry with a wider type. */ | |
2589 | if (orig_precres == precres) | |
2590 | { | |
2591 | int p = MAX (prec0, prec1); | |
f67f4dff | 2592 | scalar_int_mode m = smallest_int_mode_for_size (p); |
1304953e JJ |
2593 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (m), |
2594 | uns0_p && uns1_p | |
2595 | && unsr_p); | |
2596 | p = TYPE_PRECISION (optype); | |
2597 | if (p > precres) | |
2598 | { | |
2599 | precres = p; | |
2600 | unsr_p = TYPE_UNSIGNED (optype); | |
2601 | type = optype; | |
2602 | continue; | |
2603 | } | |
2604 | } | |
2605 | ||
2606 | gcc_unreachable (); | |
2607 | } | |
2608 | while (1); | |
2609 | } | |
2610 | ||
2611 | /* Expand ADD_OVERFLOW STMT. */ | |
2612 | ||
2613 | static void | |
4cfe7a6c | 2614 | expand_ADD_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2615 | { |
2616 | expand_arith_overflow (PLUS_EXPR, stmt); | |
2617 | } | |
2618 | ||
2619 | /* Expand SUB_OVERFLOW STMT. */ | |
2620 | ||
2621 | static void | |
4cfe7a6c | 2622 | expand_SUB_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2623 | { |
2624 | expand_arith_overflow (MINUS_EXPR, stmt); | |
2625 | } | |
2626 | ||
2627 | /* Expand MUL_OVERFLOW STMT. */ | |
2628 | ||
2629 | static void | |
4cfe7a6c | 2630 | expand_MUL_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2631 | { |
2632 | expand_arith_overflow (MULT_EXPR, stmt); | |
31e071ae MP |
2633 | } |
2634 | ||
5ce9450f JJ |
2635 | /* This should get folded in tree-vectorizer.c. */ |
2636 | ||
2637 | static void | |
4cfe7a6c | 2638 | expand_LOOP_VECTORIZED (internal_fn, gcall *) |
5ce9450f JJ |
2639 | { |
2640 | gcc_unreachable (); | |
2641 | } | |
2642 | ||
542e7230 BC |
2643 | /* This should get folded in tree-vectorizer.c. */ |
2644 | ||
2645 | static void | |
2646 | expand_LOOP_DIST_ALIAS (internal_fn, gcall *) | |
2647 | { | |
2648 | gcc_unreachable (); | |
2649 | } | |
2650 | ||
65dd1346 RS |
2651 | /* Return a memory reference of type TYPE for argument INDEX of STMT. |
2652 | Use argument INDEX + 1 to derive the second (TBAA) operand. */ | |
2653 | ||
2654 | static tree | |
2655 | expand_call_mem_ref (tree type, gcall *stmt, int index) | |
2656 | { | |
2657 | tree addr = gimple_call_arg (stmt, index); | |
2658 | tree alias_ptr_type = TREE_TYPE (gimple_call_arg (stmt, index + 1)); | |
2659 | unsigned int align = tree_to_shwi (gimple_call_arg (stmt, index + 1)); | |
2660 | if (TYPE_ALIGN (type) != align) | |
2661 | type = build_aligned_type (type, align); | |
2662 | ||
2663 | tree tmp = addr; | |
2664 | if (TREE_CODE (tmp) == SSA_NAME) | |
2665 | { | |
2666 | gimple *def = SSA_NAME_DEF_STMT (tmp); | |
2667 | if (gimple_assign_single_p (def)) | |
2668 | tmp = gimple_assign_rhs1 (def); | |
2669 | } | |
2670 | ||
2671 | if (TREE_CODE (tmp) == ADDR_EXPR) | |
2672 | { | |
2673 | tree mem = TREE_OPERAND (tmp, 0); | |
2674 | if (TREE_CODE (mem) == TARGET_MEM_REF | |
2675 | && types_compatible_p (TREE_TYPE (mem), type)) | |
2676 | { | |
2677 | tree offset = TMR_OFFSET (mem); | |
9bd958c5 RS |
2678 | if (type != TREE_TYPE (mem) |
2679 | || alias_ptr_type != TREE_TYPE (offset) | |
2680 | || !integer_zerop (offset)) | |
65dd1346 RS |
2681 | { |
2682 | mem = copy_node (mem); | |
2683 | TMR_OFFSET (mem) = wide_int_to_tree (alias_ptr_type, | |
2684 | wi::to_poly_wide (offset)); | |
9bd958c5 | 2685 | TREE_TYPE (mem) = type; |
65dd1346 RS |
2686 | } |
2687 | return mem; | |
2688 | } | |
2689 | } | |
2690 | ||
2691 | return fold_build2 (MEM_REF, type, addr, build_int_cst (alias_ptr_type, 0)); | |
2692 | } | |
2693 | ||
d496134a | 2694 | /* Expand MASK_LOAD{,_LANES} or LEN_LOAD call STMT using optab OPTAB. */ |
ab23f5d9 | 2695 | |
5ce9450f | 2696 | static void |
d496134a | 2697 | expand_partial_load_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
5ce9450f | 2698 | { |
b0e51639 RD |
2699 | class expand_operand ops[4]; |
2700 | tree type, lhs, rhs, maskt, biast; | |
2701 | rtx mem, target, mask, bias; | |
7e11fc7f | 2702 | insn_code icode; |
5ce9450f JJ |
2703 | |
2704 | maskt = gimple_call_arg (stmt, 2); | |
2705 | lhs = gimple_call_lhs (stmt); | |
8e91d222 JJ |
2706 | if (lhs == NULL_TREE) |
2707 | return; | |
5ce9450f | 2708 | type = TREE_TYPE (lhs); |
65dd1346 | 2709 | rhs = expand_call_mem_ref (type, stmt, 0); |
5ce9450f | 2710 | |
7e11fc7f RS |
2711 | if (optab == vec_mask_load_lanes_optab) |
2712 | icode = get_multi_vector_move (type, optab); | |
d496134a KL |
2713 | else if (optab == len_load_optab) |
2714 | icode = direct_optab_handler (optab, TYPE_MODE (type)); | |
7e11fc7f RS |
2715 | else |
2716 | icode = convert_optab_handler (optab, TYPE_MODE (type), | |
2717 | TYPE_MODE (TREE_TYPE (maskt))); | |
2718 | ||
5ce9450f JJ |
2719 | mem = expand_expr (rhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
2720 | gcc_assert (MEM_P (mem)); | |
2721 | mask = expand_normal (maskt); | |
2722 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2723 | create_output_operand (&ops[0], target, TYPE_MODE (type)); | |
2724 | create_fixed_operand (&ops[1], mem); | |
d496134a | 2725 | if (optab == len_load_optab) |
b0e51639 RD |
2726 | { |
2727 | create_convert_operand_from (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt)), | |
2728 | TYPE_UNSIGNED (TREE_TYPE (maskt))); | |
2729 | biast = gimple_call_arg (stmt, 3); | |
2730 | bias = expand_normal (biast); | |
2731 | create_input_operand (&ops[3], bias, QImode); | |
2732 | expand_insn (icode, 4, ops); | |
2733 | } | |
d496134a | 2734 | else |
b0e51639 RD |
2735 | { |
2736 | create_input_operand (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt))); | |
2737 | expand_insn (icode, 3, ops); | |
2738 | } | |
2739 | ||
3af3bec2 RS |
2740 | if (!rtx_equal_p (target, ops[0].value)) |
2741 | emit_move_insn (target, ops[0].value); | |
5ce9450f JJ |
2742 | } |
2743 | ||
d496134a | 2744 | #define expand_mask_load_optab_fn expand_partial_load_optab_fn |
7e11fc7f | 2745 | #define expand_mask_load_lanes_optab_fn expand_mask_load_optab_fn |
d496134a | 2746 | #define expand_len_load_optab_fn expand_partial_load_optab_fn |
7e11fc7f | 2747 | |
d496134a | 2748 | /* Expand MASK_STORE{,_LANES} or LEN_STORE call STMT using optab OPTAB. */ |
ab23f5d9 | 2749 | |
5ce9450f | 2750 | static void |
d496134a | 2751 | expand_partial_store_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
5ce9450f | 2752 | { |
b0e51639 RD |
2753 | class expand_operand ops[4]; |
2754 | tree type, lhs, rhs, maskt, biast; | |
2755 | rtx mem, reg, mask, bias; | |
7e11fc7f | 2756 | insn_code icode; |
5ce9450f JJ |
2757 | |
2758 | maskt = gimple_call_arg (stmt, 2); | |
2759 | rhs = gimple_call_arg (stmt, 3); | |
2760 | type = TREE_TYPE (rhs); | |
65dd1346 | 2761 | lhs = expand_call_mem_ref (type, stmt, 0); |
5ce9450f | 2762 | |
7e11fc7f RS |
2763 | if (optab == vec_mask_store_lanes_optab) |
2764 | icode = get_multi_vector_move (type, optab); | |
d496134a KL |
2765 | else if (optab == len_store_optab) |
2766 | icode = direct_optab_handler (optab, TYPE_MODE (type)); | |
7e11fc7f RS |
2767 | else |
2768 | icode = convert_optab_handler (optab, TYPE_MODE (type), | |
2769 | TYPE_MODE (TREE_TYPE (maskt))); | |
2770 | ||
5ce9450f JJ |
2771 | mem = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
2772 | gcc_assert (MEM_P (mem)); | |
2773 | mask = expand_normal (maskt); | |
2774 | reg = expand_normal (rhs); | |
2775 | create_fixed_operand (&ops[0], mem); | |
2776 | create_input_operand (&ops[1], reg, TYPE_MODE (type)); | |
d496134a | 2777 | if (optab == len_store_optab) |
b0e51639 RD |
2778 | { |
2779 | create_convert_operand_from (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt)), | |
2780 | TYPE_UNSIGNED (TREE_TYPE (maskt))); | |
2781 | biast = gimple_call_arg (stmt, 4); | |
2782 | bias = expand_normal (biast); | |
2783 | create_input_operand (&ops[3], bias, QImode); | |
2784 | expand_insn (icode, 4, ops); | |
2785 | } | |
d496134a | 2786 | else |
b0e51639 RD |
2787 | { |
2788 | create_input_operand (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt))); | |
2789 | expand_insn (icode, 3, ops); | |
2790 | } | |
5ce9450f JJ |
2791 | } |
2792 | ||
d496134a | 2793 | #define expand_mask_store_optab_fn expand_partial_store_optab_fn |
7e11fc7f | 2794 | #define expand_mask_store_lanes_optab_fn expand_mask_store_optab_fn |
d496134a | 2795 | #define expand_len_store_optab_fn expand_partial_store_optab_fn |
7e11fc7f | 2796 | |
502d63b6 ML |
2797 | /* Expand VCOND, VCONDU and VCONDEQ optab internal functions. |
2798 | The expansion of STMT happens based on OPTAB table associated. */ | |
2799 | ||
2800 | static void | |
298e76e6 | 2801 | expand_vec_cond_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
502d63b6 ML |
2802 | { |
2803 | class expand_operand ops[6]; | |
2804 | insn_code icode; | |
2805 | tree lhs = gimple_call_lhs (stmt); | |
2806 | tree op0a = gimple_call_arg (stmt, 0); | |
2807 | tree op0b = gimple_call_arg (stmt, 1); | |
2808 | tree op1 = gimple_call_arg (stmt, 2); | |
2809 | tree op2 = gimple_call_arg (stmt, 3); | |
2810 | enum tree_code tcode = (tree_code) int_cst_value (gimple_call_arg (stmt, 4)); | |
2811 | ||
2812 | tree vec_cond_type = TREE_TYPE (lhs); | |
2813 | tree op_mode = TREE_TYPE (op0a); | |
2814 | bool unsignedp = TYPE_UNSIGNED (op_mode); | |
2815 | ||
2816 | machine_mode mode = TYPE_MODE (vec_cond_type); | |
2817 | machine_mode cmp_op_mode = TYPE_MODE (op_mode); | |
2818 | ||
2819 | icode = convert_optab_handler (optab, mode, cmp_op_mode); | |
2820 | rtx comparison | |
2821 | = vector_compare_rtx (VOIDmode, tcode, op0a, op0b, unsignedp, icode, 4); | |
2822 | rtx rtx_op1 = expand_normal (op1); | |
2823 | rtx rtx_op2 = expand_normal (op2); | |
2824 | ||
2825 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2826 | create_output_operand (&ops[0], target, mode); | |
2827 | create_input_operand (&ops[1], rtx_op1, mode); | |
2828 | create_input_operand (&ops[2], rtx_op2, mode); | |
2829 | create_fixed_operand (&ops[3], comparison); | |
2830 | create_fixed_operand (&ops[4], XEXP (comparison, 0)); | |
2831 | create_fixed_operand (&ops[5], XEXP (comparison, 1)); | |
2832 | expand_insn (icode, 6, ops); | |
eccc3d43 RB |
2833 | if (!rtx_equal_p (ops[0].value, target)) |
2834 | emit_move_insn (target, ops[0].value); | |
502d63b6 ML |
2835 | } |
2836 | ||
502d63b6 ML |
2837 | /* Expand VCOND_MASK optab internal function. |
2838 | The expansion of STMT happens based on OPTAB table associated. */ | |
2839 | ||
2840 | static void | |
298e76e6 | 2841 | expand_vec_cond_mask_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
502d63b6 ML |
2842 | { |
2843 | class expand_operand ops[4]; | |
2844 | ||
2845 | tree lhs = gimple_call_lhs (stmt); | |
2846 | tree op0 = gimple_call_arg (stmt, 0); | |
2847 | tree op1 = gimple_call_arg (stmt, 1); | |
2848 | tree op2 = gimple_call_arg (stmt, 2); | |
2849 | tree vec_cond_type = TREE_TYPE (lhs); | |
2850 | ||
2851 | machine_mode mode = TYPE_MODE (vec_cond_type); | |
2852 | machine_mode mask_mode = TYPE_MODE (TREE_TYPE (op0)); | |
2853 | enum insn_code icode = convert_optab_handler (optab, mode, mask_mode); | |
2854 | rtx mask, rtx_op1, rtx_op2; | |
2855 | ||
2856 | gcc_assert (icode != CODE_FOR_nothing); | |
2857 | ||
2858 | mask = expand_normal (op0); | |
2859 | rtx_op1 = expand_normal (op1); | |
2860 | rtx_op2 = expand_normal (op2); | |
2861 | ||
2862 | mask = force_reg (mask_mode, mask); | |
bc909324 | 2863 | rtx_op1 = force_reg (mode, rtx_op1); |
502d63b6 ML |
2864 | |
2865 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2866 | create_output_operand (&ops[0], target, mode); | |
2867 | create_input_operand (&ops[1], rtx_op1, mode); | |
2868 | create_input_operand (&ops[2], rtx_op2, mode); | |
2869 | create_input_operand (&ops[3], mask, mask_mode); | |
2870 | expand_insn (icode, 4, ops); | |
eccc3d43 RB |
2871 | if (!rtx_equal_p (ops[0].value, target)) |
2872 | emit_move_insn (target, ops[0].value); | |
502d63b6 ML |
2873 | } |
2874 | ||
683e55fa XL |
2875 | /* Expand VEC_SET internal functions. */ |
2876 | ||
2877 | static void | |
2878 | expand_vec_set_optab_fn (internal_fn, gcall *stmt, convert_optab optab) | |
2879 | { | |
2880 | tree lhs = gimple_call_lhs (stmt); | |
2881 | tree op0 = gimple_call_arg (stmt, 0); | |
2882 | tree op1 = gimple_call_arg (stmt, 1); | |
2883 | tree op2 = gimple_call_arg (stmt, 2); | |
2884 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2885 | rtx src = expand_normal (op0); | |
2886 | ||
2887 | machine_mode outermode = TYPE_MODE (TREE_TYPE (op0)); | |
2888 | scalar_mode innermode = GET_MODE_INNER (outermode); | |
2889 | ||
2890 | rtx value = expand_normal (op1); | |
2891 | rtx pos = expand_normal (op2); | |
2892 | ||
2893 | class expand_operand ops[3]; | |
2894 | enum insn_code icode = optab_handler (optab, outermode); | |
2895 | ||
2896 | if (icode != CODE_FOR_nothing) | |
2897 | { | |
2898 | rtx temp = gen_reg_rtx (outermode); | |
2899 | emit_move_insn (temp, src); | |
2900 | ||
2901 | create_fixed_operand (&ops[0], temp); | |
2902 | create_input_operand (&ops[1], value, innermode); | |
2903 | create_convert_operand_from (&ops[2], pos, TYPE_MODE (TREE_TYPE (op2)), | |
2904 | true); | |
2905 | if (maybe_expand_insn (icode, 3, ops)) | |
2906 | { | |
2907 | emit_move_insn (target, temp); | |
2908 | return; | |
2909 | } | |
2910 | } | |
2911 | gcc_unreachable (); | |
2912 | } | |
2913 | ||
09b22f48 | 2914 | static void |
4cfe7a6c | 2915 | expand_ABNORMAL_DISPATCHER (internal_fn, gcall *) |
09b22f48 JJ |
2916 | { |
2917 | } | |
2918 | ||
ed9c79e1 | 2919 | static void |
4cfe7a6c | 2920 | expand_BUILTIN_EXPECT (internal_fn, gcall *stmt) |
ed9c79e1 JJ |
2921 | { |
2922 | /* When guessing was done, the hints should be already stripped away. */ | |
2923 | gcc_assert (!flag_guess_branch_prob || optimize == 0 || seen_error ()); | |
2924 | ||
2925 | rtx target; | |
2926 | tree lhs = gimple_call_lhs (stmt); | |
2927 | if (lhs) | |
2928 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2929 | else | |
2930 | target = const0_rtx; | |
2931 | rtx val = expand_expr (gimple_call_arg (stmt, 0), target, VOIDmode, EXPAND_NORMAL); | |
2932 | if (lhs && val != target) | |
2933 | emit_move_insn (target, val); | |
2934 | } | |
2935 | ||
f8e89441 TV |
2936 | /* IFN_VA_ARG is supposed to be expanded at pass_stdarg. So this dummy function |
2937 | should never be called. */ | |
2938 | ||
2939 | static void | |
4cfe7a6c | 2940 | expand_VA_ARG (internal_fn, gcall *) |
d8fcab68 JJ |
2941 | { |
2942 | gcc_unreachable (); | |
2943 | } | |
2944 | ||
2945 | /* IFN_VEC_CONVERT is supposed to be expanded at pass_lower_vector. So this | |
2946 | dummy function should never be called. */ | |
2947 | ||
2948 | static void | |
2949 | expand_VEC_CONVERT (internal_fn, gcall *) | |
f8e89441 TV |
2950 | { |
2951 | gcc_unreachable (); | |
2952 | } | |
2953 | ||
6f966f06 SSF |
2954 | /* Expand IFN_RAWMEMCHAR internal function. */ |
2955 | ||
2956 | void | |
2957 | expand_RAWMEMCHR (internal_fn, gcall *stmt) | |
2958 | { | |
2959 | expand_operand ops[3]; | |
2960 | ||
2961 | tree lhs = gimple_call_lhs (stmt); | |
2962 | if (!lhs) | |
2963 | return; | |
2964 | machine_mode lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); | |
2965 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2966 | create_output_operand (&ops[0], lhs_rtx, lhs_mode); | |
2967 | ||
2968 | tree mem = gimple_call_arg (stmt, 0); | |
2969 | rtx mem_rtx = get_memory_rtx (mem, NULL); | |
2970 | create_fixed_operand (&ops[1], mem_rtx); | |
2971 | ||
2972 | tree pattern = gimple_call_arg (stmt, 1); | |
2973 | machine_mode mode = TYPE_MODE (TREE_TYPE (pattern)); | |
2974 | rtx pattern_rtx = expand_normal (pattern); | |
2975 | create_input_operand (&ops[2], pattern_rtx, mode); | |
2976 | ||
2977 | insn_code icode = direct_optab_handler (rawmemchr_optab, mode); | |
2978 | ||
2979 | expand_insn (icode, 3, ops); | |
2980 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) | |
2981 | emit_move_insn (lhs_rtx, ops[0].value); | |
2982 | } | |
2983 | ||
8ab78162 NS |
2984 | /* Expand the IFN_UNIQUE function according to its first argument. */ |
2985 | ||
2986 | static void | |
4cfe7a6c | 2987 | expand_UNIQUE (internal_fn, gcall *stmt) |
8ab78162 NS |
2988 | { |
2989 | rtx pattern = NULL_RTX; | |
2990 | enum ifn_unique_kind kind | |
2991 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (stmt, 0)); | |
2992 | ||
2993 | switch (kind) | |
2994 | { | |
2995 | default: | |
2996 | gcc_unreachable (); | |
2997 | ||
2998 | case IFN_UNIQUE_UNSPEC: | |
2999 | if (targetm.have_unique ()) | |
3000 | pattern = targetm.gen_unique (); | |
3001 | break; | |
9bd46bc9 NS |
3002 | |
3003 | case IFN_UNIQUE_OACC_FORK: | |
3004 | case IFN_UNIQUE_OACC_JOIN: | |
3005 | if (targetm.have_oacc_fork () && targetm.have_oacc_join ()) | |
3006 | { | |
3007 | tree lhs = gimple_call_lhs (stmt); | |
3008 | rtx target = const0_rtx; | |
3009 | ||
3010 | if (lhs) | |
3011 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3012 | ||
3013 | rtx data_dep = expand_normal (gimple_call_arg (stmt, 1)); | |
3014 | rtx axis = expand_normal (gimple_call_arg (stmt, 2)); | |
3015 | ||
3016 | if (kind == IFN_UNIQUE_OACC_FORK) | |
3017 | pattern = targetm.gen_oacc_fork (target, data_dep, axis); | |
3018 | else | |
3019 | pattern = targetm.gen_oacc_join (target, data_dep, axis); | |
3020 | } | |
3021 | else | |
3022 | gcc_unreachable (); | |
3023 | break; | |
8ab78162 NS |
3024 | } |
3025 | ||
3026 | if (pattern) | |
3027 | emit_insn (pattern); | |
3028 | } | |
3029 | ||
a25e0b5e | 3030 | /* Expand the IFN_DEFERRED_INIT function: |
6c98c8b4 | 3031 | LHS = DEFERRED_INIT (SIZE of the DECL, INIT_TYPE, NAME of the DECL); |
a25e0b5e | 3032 | |
3033 | Initialize the LHS with zero/pattern according to its second argument | |
3034 | INIT_TYPE: | |
3035 | if INIT_TYPE is AUTO_INIT_ZERO, use zeroes to initialize; | |
3036 | if INIT_TYPE is AUTO_INIT_PATTERN, use 0xFE byte-repeatable pattern | |
3037 | to initialize; | |
3038 | The LHS variable is initialized including paddings. | |
3039 | The reasons to choose 0xFE for pattern initialization are: | |
3040 | 1. It is a non-canonical virtual address on x86_64, and at the | |
3041 | high end of the i386 kernel address space. | |
3042 | 2. It is a very large float value (-1.694739530317379e+38). | |
3043 | 3. It is also an unusual number for integers. */ | |
3044 | #define INIT_PATTERN_VALUE 0xFE | |
3045 | static void | |
3046 | expand_DEFERRED_INIT (internal_fn, gcall *stmt) | |
3047 | { | |
3048 | tree lhs = gimple_call_lhs (stmt); | |
3049 | tree var_size = gimple_call_arg (stmt, 0); | |
3050 | enum auto_init_type init_type | |
3051 | = (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1)); | |
a25e0b5e | 3052 | bool reg_lhs = true; |
3053 | ||
3054 | tree var_type = TREE_TYPE (lhs); | |
3055 | gcc_assert (init_type > AUTO_INIT_UNINITIALIZED); | |
3056 | ||
79f488de | 3057 | if (TREE_CODE (lhs) == SSA_NAME) |
a25e0b5e | 3058 | reg_lhs = true; |
3059 | else | |
3060 | { | |
7e0c0500 RB |
3061 | tree lhs_base = lhs; |
3062 | while (handled_component_p (lhs_base)) | |
3063 | lhs_base = TREE_OPERAND (lhs_base, 0); | |
3064 | reg_lhs = (mem_ref_refers_to_non_mem_p (lhs_base) | |
3065 | || non_mem_decl_p (lhs_base)); | |
06b8cdc8 RB |
3066 | /* If this expands to a register and the underlying decl is wrapped in |
3067 | a MEM_REF that just serves as an access type change expose the decl | |
3068 | if it is of correct size. This avoids a situation as in PR103271 | |
3069 | if the target does not support a direct move to the registers mode. */ | |
3070 | if (reg_lhs | |
3071 | && TREE_CODE (lhs_base) == MEM_REF | |
3072 | && TREE_CODE (TREE_OPERAND (lhs_base, 0)) == ADDR_EXPR | |
3073 | && DECL_P (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)) | |
3074 | && integer_zerop (TREE_OPERAND (lhs_base, 1)) | |
3075 | && tree_fits_uhwi_p (var_size) | |
3076 | && tree_int_cst_equal | |
3077 | (var_size, | |
3078 | DECL_SIZE_UNIT (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)))) | |
3079 | { | |
3080 | lhs = TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0); | |
3081 | var_type = TREE_TYPE (lhs); | |
3082 | } | |
a25e0b5e | 3083 | } |
3084 | ||
a25e0b5e | 3085 | if (!reg_lhs) |
3086 | { | |
6c98c8b4 QZ |
3087 | /* If the variable is not in register, expand to a memset |
3088 | to initialize it. */ | |
a25e0b5e | 3089 | mark_addressable (lhs); |
3090 | tree var_addr = build_fold_addr_expr (lhs); | |
3091 | ||
1c04af34 QZ |
3092 | tree value = (init_type == AUTO_INIT_PATTERN) |
3093 | ? build_int_cst (integer_type_node, | |
3094 | INIT_PATTERN_VALUE) | |
3095 | : integer_zero_node; | |
a25e0b5e | 3096 | tree m_call = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMSET), |
3097 | 3, var_addr, value, var_size); | |
3098 | /* Expand this memset call. */ | |
3099 | expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type)); | |
3100 | } | |
604459a0 | 3101 | else |
a25e0b5e | 3102 | { |
c081d0a3 RB |
3103 | /* If this variable is in a register use expand_assignment. |
3104 | For boolean scalars force zero-init. */ | |
604459a0 | 3105 | tree init; |
1c04af34 | 3106 | scalar_int_mode var_mode; |
c081d0a3 RB |
3107 | if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE |
3108 | && tree_fits_uhwi_p (var_size) | |
604459a0 | 3109 | && (init_type == AUTO_INIT_PATTERN |
33872565 RB |
3110 | || !is_gimple_reg_type (var_type)) |
3111 | && int_mode_for_size (tree_to_uhwi (var_size) * BITS_PER_UNIT, | |
1c04af34 QZ |
3112 | 0).exists (&var_mode) |
3113 | && have_insn_for (SET, var_mode)) | |
a25e0b5e | 3114 | { |
604459a0 | 3115 | unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size); |
1c04af34 | 3116 | unsigned char *buf = XALLOCAVEC (unsigned char, total_bytes); |
604459a0 RB |
3117 | memset (buf, (init_type == AUTO_INIT_PATTERN |
3118 | ? INIT_PATTERN_VALUE : 0), total_bytes); | |
c081d0a3 RB |
3119 | tree itype = build_nonstandard_integer_type |
3120 | (total_bytes * BITS_PER_UNIT, 1); | |
3121 | wide_int w = wi::from_buffer (buf, total_bytes); | |
3122 | init = wide_int_to_tree (itype, w); | |
3123 | /* Pun the LHS to make sure its type has constant size | |
3124 | unless it is an SSA name where that's already known. */ | |
3125 | if (TREE_CODE (lhs) != SSA_NAME) | |
3126 | lhs = build1 (VIEW_CONVERT_EXPR, itype, lhs); | |
8d6b12b2 | 3127 | else |
c081d0a3 | 3128 | init = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (lhs), init); |
a25e0b5e | 3129 | } |
604459a0 RB |
3130 | else |
3131 | /* Use zero-init also for variable-length sizes. */ | |
3132 | init = build_zero_cst (var_type); | |
a25e0b5e | 3133 | |
3134 | expand_assignment (lhs, init, false); | |
3135 | } | |
3136 | } | |
3137 | ||
9bd46bc9 NS |
3138 | /* The size of an OpenACC compute dimension. */ |
3139 | ||
3140 | static void | |
4cfe7a6c | 3141 | expand_GOACC_DIM_SIZE (internal_fn, gcall *stmt) |
9bd46bc9 NS |
3142 | { |
3143 | tree lhs = gimple_call_lhs (stmt); | |
3144 | ||
3145 | if (!lhs) | |
3146 | return; | |
3147 | ||
3148 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3149 | if (targetm.have_oacc_dim_size ()) | |
3150 | { | |
3151 | rtx dim = expand_expr (gimple_call_arg (stmt, 0), NULL_RTX, | |
3152 | VOIDmode, EXPAND_NORMAL); | |
3153 | emit_insn (targetm.gen_oacc_dim_size (target, dim)); | |
3154 | } | |
3155 | else | |
3156 | emit_move_insn (target, GEN_INT (1)); | |
3157 | } | |
3158 | ||
3159 | /* The position of an OpenACC execution engine along one compute axis. */ | |
3160 | ||
3161 | static void | |
4cfe7a6c | 3162 | expand_GOACC_DIM_POS (internal_fn, gcall *stmt) |
9bd46bc9 NS |
3163 | { |
3164 | tree lhs = gimple_call_lhs (stmt); | |
3165 | ||
3166 | if (!lhs) | |
3167 | return; | |
3168 | ||
3169 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3170 | if (targetm.have_oacc_dim_pos ()) | |
3171 | { | |
3172 | rtx dim = expand_expr (gimple_call_arg (stmt, 0), NULL_RTX, | |
3173 | VOIDmode, EXPAND_NORMAL); | |
3174 | emit_insn (targetm.gen_oacc_dim_pos (target, dim)); | |
3175 | } | |
3176 | else | |
3177 | emit_move_insn (target, const0_rtx); | |
3178 | } | |
3179 | ||
3180 | /* This is expanded by oacc_device_lower pass. */ | |
3181 | ||
3182 | static void | |
4cfe7a6c | 3183 | expand_GOACC_LOOP (internal_fn, gcall *) |
9bd46bc9 NS |
3184 | { |
3185 | gcc_unreachable (); | |
3186 | } | |
3187 | ||
e5014671 NS |
3188 | /* This is expanded by oacc_device_lower pass. */ |
3189 | ||
3190 | static void | |
4cfe7a6c | 3191 | expand_GOACC_REDUCTION (internal_fn, gcall *) |
e5014671 NS |
3192 | { |
3193 | gcc_unreachable (); | |
02889d23 CLT |
3194 | } |
3195 | ||
3196 | /* This is expanded by oacc_device_lower pass. */ | |
3197 | ||
3198 | static void | |
3199 | expand_GOACC_TILE (internal_fn, gcall *) | |
3200 | { | |
3201 | gcc_unreachable (); | |
e5014671 NS |
3202 | } |
3203 | ||
883cabde RS |
3204 | /* Set errno to EDOM. */ |
3205 | ||
3206 | static void | |
3207 | expand_SET_EDOM (internal_fn, gcall *) | |
3208 | { | |
3209 | #ifdef TARGET_EDOM | |
3210 | #ifdef GEN_ERRNO_RTX | |
3211 | rtx errno_rtx = GEN_ERRNO_RTX; | |
3212 | #else | |
3213 | rtx errno_rtx = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno")); | |
3214 | #endif | |
3215 | emit_move_insn (errno_rtx, | |
3216 | gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); | |
3217 | #else | |
3218 | gcc_unreachable (); | |
3219 | #endif | |
3220 | } | |
3221 | ||
adedd5c1 JJ |
3222 | /* Expand atomic bit test and set. */ |
3223 | ||
3224 | static void | |
3225 | expand_ATOMIC_BIT_TEST_AND_SET (internal_fn, gcall *call) | |
3226 | { | |
3227 | expand_ifn_atomic_bit_test_and (call); | |
3228 | } | |
3229 | ||
3230 | /* Expand atomic bit test and complement. */ | |
3231 | ||
3232 | static void | |
3233 | expand_ATOMIC_BIT_TEST_AND_COMPLEMENT (internal_fn, gcall *call) | |
3234 | { | |
3235 | expand_ifn_atomic_bit_test_and (call); | |
3236 | } | |
3237 | ||
3238 | /* Expand atomic bit test and reset. */ | |
3239 | ||
3240 | static void | |
3241 | expand_ATOMIC_BIT_TEST_AND_RESET (internal_fn, gcall *call) | |
3242 | { | |
3243 | expand_ifn_atomic_bit_test_and (call); | |
3244 | } | |
3245 | ||
849a76a5 JJ |
3246 | /* Expand atomic bit test and set. */ |
3247 | ||
3248 | static void | |
3249 | expand_ATOMIC_COMPARE_EXCHANGE (internal_fn, gcall *call) | |
3250 | { | |
3251 | expand_ifn_atomic_compare_exchange (call); | |
3252 | } | |
3253 | ||
6362627b JJ |
3254 | /* Expand atomic add fetch and cmp with 0. */ |
3255 | ||
3256 | static void | |
3257 | expand_ATOMIC_ADD_FETCH_CMP_0 (internal_fn, gcall *call) | |
3258 | { | |
3259 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3260 | } | |
3261 | ||
3262 | /* Expand atomic sub fetch and cmp with 0. */ | |
3263 | ||
3264 | static void | |
3265 | expand_ATOMIC_SUB_FETCH_CMP_0 (internal_fn, gcall *call) | |
3266 | { | |
3267 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3268 | } | |
3269 | ||
3270 | /* Expand atomic and fetch and cmp with 0. */ | |
3271 | ||
3272 | static void | |
3273 | expand_ATOMIC_AND_FETCH_CMP_0 (internal_fn, gcall *call) | |
3274 | { | |
3275 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3276 | } | |
3277 | ||
3278 | /* Expand atomic or fetch and cmp with 0. */ | |
3279 | ||
3280 | static void | |
3281 | expand_ATOMIC_OR_FETCH_CMP_0 (internal_fn, gcall *call) | |
3282 | { | |
3283 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3284 | } | |
3285 | ||
3286 | /* Expand atomic xor fetch and cmp with 0. */ | |
3287 | ||
3288 | static void | |
3289 | expand_ATOMIC_XOR_FETCH_CMP_0 (internal_fn, gcall *call) | |
3290 | { | |
3291 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3292 | } | |
3293 | ||
e16f1cc7 JJ |
3294 | /* Expand LAUNDER to assignment, lhs = arg0. */ |
3295 | ||
3296 | static void | |
3297 | expand_LAUNDER (internal_fn, gcall *call) | |
3298 | { | |
3299 | tree lhs = gimple_call_lhs (call); | |
3300 | ||
3301 | if (!lhs) | |
3302 | return; | |
3303 | ||
3304 | expand_assignment (lhs, gimple_call_arg (call, 0), false); | |
3305 | } | |
3306 | ||
f307441a RS |
3307 | /* Expand {MASK_,}SCATTER_STORE{S,U} call CALL using optab OPTAB. */ |
3308 | ||
3309 | static void | |
3310 | expand_scatter_store_optab_fn (internal_fn, gcall *stmt, direct_optab optab) | |
3311 | { | |
3312 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
3313 | int rhs_index = internal_fn_stored_value_index (ifn); | |
3314 | int mask_index = internal_fn_mask_index (ifn); | |
3315 | tree base = gimple_call_arg (stmt, 0); | |
3316 | tree offset = gimple_call_arg (stmt, 1); | |
3317 | tree scale = gimple_call_arg (stmt, 2); | |
3318 | tree rhs = gimple_call_arg (stmt, rhs_index); | |
3319 | ||
3320 | rtx base_rtx = expand_normal (base); | |
3321 | rtx offset_rtx = expand_normal (offset); | |
3322 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); | |
3323 | rtx rhs_rtx = expand_normal (rhs); | |
3324 | ||
99b1c316 | 3325 | class expand_operand ops[6]; |
f307441a RS |
3326 | int i = 0; |
3327 | create_address_operand (&ops[i++], base_rtx); | |
3328 | create_input_operand (&ops[i++], offset_rtx, TYPE_MODE (TREE_TYPE (offset))); | |
3329 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); | |
3330 | create_integer_operand (&ops[i++], scale_int); | |
3331 | create_input_operand (&ops[i++], rhs_rtx, TYPE_MODE (TREE_TYPE (rhs))); | |
3332 | if (mask_index >= 0) | |
3333 | { | |
3334 | tree mask = gimple_call_arg (stmt, mask_index); | |
3335 | rtx mask_rtx = expand_normal (mask); | |
3336 | create_input_operand (&ops[i++], mask_rtx, TYPE_MODE (TREE_TYPE (mask))); | |
3337 | } | |
3338 | ||
09eb042a RS |
3339 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (TREE_TYPE (rhs)), |
3340 | TYPE_MODE (TREE_TYPE (offset))); | |
f307441a RS |
3341 | expand_insn (icode, i, ops); |
3342 | } | |
3343 | ||
bfaa08b7 RS |
3344 | /* Expand {MASK_,}GATHER_LOAD call CALL using optab OPTAB. */ |
3345 | ||
3346 | static void | |
3347 | expand_gather_load_optab_fn (internal_fn, gcall *stmt, direct_optab optab) | |
3348 | { | |
3349 | tree lhs = gimple_call_lhs (stmt); | |
3350 | tree base = gimple_call_arg (stmt, 0); | |
3351 | tree offset = gimple_call_arg (stmt, 1); | |
3352 | tree scale = gimple_call_arg (stmt, 2); | |
3353 | ||
3354 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3355 | rtx base_rtx = expand_normal (base); | |
3356 | rtx offset_rtx = expand_normal (offset); | |
3357 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); | |
3358 | ||
3359 | int i = 0; | |
99b1c316 | 3360 | class expand_operand ops[6]; |
bfaa08b7 RS |
3361 | create_output_operand (&ops[i++], lhs_rtx, TYPE_MODE (TREE_TYPE (lhs))); |
3362 | create_address_operand (&ops[i++], base_rtx); | |
3363 | create_input_operand (&ops[i++], offset_rtx, TYPE_MODE (TREE_TYPE (offset))); | |
3364 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); | |
3365 | create_integer_operand (&ops[i++], scale_int); | |
3366 | if (optab == mask_gather_load_optab) | |
3367 | { | |
09eb042a | 3368 | tree mask = gimple_call_arg (stmt, 4); |
bfaa08b7 RS |
3369 | rtx mask_rtx = expand_normal (mask); |
3370 | create_input_operand (&ops[i++], mask_rtx, TYPE_MODE (TREE_TYPE (mask))); | |
3371 | } | |
09eb042a RS |
3372 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (TREE_TYPE (lhs)), |
3373 | TYPE_MODE (TREE_TYPE (offset))); | |
bfaa08b7 | 3374 | expand_insn (icode, i, ops); |
3af3bec2 RS |
3375 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3376 | emit_move_insn (lhs_rtx, ops[0].value); | |
bfaa08b7 RS |
3377 | } |
3378 | ||
bf510679 JJ |
3379 | /* Helper for expand_DIVMOD. Return true if the sequence starting with |
3380 | INSN contains any call insns or insns with {,U}{DIV,MOD} rtxes. */ | |
3381 | ||
3382 | static bool | |
3383 | contains_call_div_mod (rtx_insn *insn) | |
3384 | { | |
3385 | subrtx_iterator::array_type array; | |
3386 | for (; insn; insn = NEXT_INSN (insn)) | |
3387 | if (CALL_P (insn)) | |
3388 | return true; | |
3389 | else if (INSN_P (insn)) | |
3390 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
3391 | switch (GET_CODE (*iter)) | |
3392 | { | |
3393 | case CALL: | |
3394 | case DIV: | |
3395 | case UDIV: | |
3396 | case MOD: | |
3397 | case UMOD: | |
3398 | return true; | |
3399 | default: | |
3400 | break; | |
3401 | } | |
3402 | return false; | |
3403 | } | |
3404 | ||
e72531b9 PK |
3405 | /* Expand DIVMOD() using: |
3406 | a) optab handler for udivmod/sdivmod if it is available. | |
3407 | b) If optab_handler doesn't exist, generate call to | |
3408 | target-specific divmod libfunc. */ | |
3409 | ||
3410 | static void | |
3411 | expand_DIVMOD (internal_fn, gcall *call_stmt) | |
3412 | { | |
3413 | tree lhs = gimple_call_lhs (call_stmt); | |
3414 | tree arg0 = gimple_call_arg (call_stmt, 0); | |
3415 | tree arg1 = gimple_call_arg (call_stmt, 1); | |
3416 | ||
3417 | gcc_assert (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE); | |
3418 | tree type = TREE_TYPE (TREE_TYPE (lhs)); | |
3419 | machine_mode mode = TYPE_MODE (type); | |
3420 | bool unsignedp = TYPE_UNSIGNED (type); | |
3421 | optab tab = (unsignedp) ? udivmod_optab : sdivmod_optab; | |
3422 | ||
3423 | rtx op0 = expand_normal (arg0); | |
3424 | rtx op1 = expand_normal (arg1); | |
3425 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3426 | ||
bf510679 JJ |
3427 | rtx quotient = NULL_RTX, remainder = NULL_RTX; |
3428 | rtx_insn *insns = NULL; | |
3429 | ||
3430 | if (TREE_CODE (arg1) == INTEGER_CST) | |
3431 | { | |
3432 | /* For DIVMOD by integral constants, there could be efficient code | |
3433 | expanded inline e.g. using shifts and plus/minus. Try to expand | |
3434 | the division and modulo and if it emits any library calls or any | |
3435 | {,U}{DIV,MOD} rtxes throw it away and use a divmod optab or | |
3436 | divmod libcall. */ | |
037fe26e JJ |
3437 | scalar_int_mode int_mode; |
3438 | if (remainder == NULL_RTX | |
3439 | && optimize | |
3440 | && CONST_INT_P (op1) | |
3441 | && !pow2p_hwi (INTVAL (op1)) | |
3442 | && is_int_mode (TYPE_MODE (type), &int_mode) | |
3443 | && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD | |
3444 | && optab_handler (and_optab, word_mode) != CODE_FOR_nothing | |
3445 | && optab_handler (add_optab, word_mode) != CODE_FOR_nothing | |
3446 | && optimize_insn_for_speed_p ()) | |
3447 | { | |
3448 | rtx_insn *last = get_last_insn (); | |
3449 | remainder = NULL_RTX; | |
3450 | quotient = expand_doubleword_divmod (int_mode, op0, op1, &remainder, | |
3451 | TYPE_UNSIGNED (type)); | |
3452 | if (quotient != NULL_RTX) | |
3453 | { | |
3454 | if (optab_handler (mov_optab, int_mode) != CODE_FOR_nothing) | |
3455 | { | |
3456 | rtx_insn *move = emit_move_insn (quotient, quotient); | |
3457 | set_dst_reg_note (move, REG_EQUAL, | |
3458 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) | |
3459 | ? UDIV : DIV, int_mode, | |
3460 | copy_rtx (op0), op1), | |
3461 | quotient); | |
3462 | move = emit_move_insn (remainder, remainder); | |
3463 | set_dst_reg_note (move, REG_EQUAL, | |
3464 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) | |
3465 | ? UMOD : MOD, int_mode, | |
3466 | copy_rtx (op0), op1), | |
3467 | quotient); | |
3468 | } | |
3469 | } | |
3470 | else | |
3471 | delete_insns_since (last); | |
3472 | } | |
3473 | ||
3474 | if (remainder == NULL_RTX) | |
bf510679 | 3475 | { |
037fe26e JJ |
3476 | struct separate_ops ops; |
3477 | ops.code = TRUNC_DIV_EXPR; | |
3478 | ops.type = type; | |
3479 | ops.op0 = make_tree (ops.type, op0); | |
3480 | ops.op1 = arg1; | |
3481 | ops.op2 = NULL_TREE; | |
3482 | ops.location = gimple_location (call_stmt); | |
3483 | start_sequence (); | |
3484 | quotient = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
bf510679 | 3485 | if (contains_call_div_mod (get_insns ())) |
037fe26e JJ |
3486 | quotient = NULL_RTX; |
3487 | else | |
3488 | { | |
3489 | ops.code = TRUNC_MOD_EXPR; | |
3490 | remainder = expand_expr_real_2 (&ops, NULL_RTX, mode, | |
3491 | EXPAND_NORMAL); | |
3492 | if (contains_call_div_mod (get_insns ())) | |
3493 | remainder = NULL_RTX; | |
3494 | } | |
3495 | if (remainder) | |
3496 | insns = get_insns (); | |
3497 | end_sequence (); | |
bf510679 | 3498 | } |
bf510679 JJ |
3499 | } |
3500 | ||
3501 | if (remainder) | |
3502 | emit_insn (insns); | |
e72531b9 PK |
3503 | |
3504 | /* Check if optab_handler exists for divmod_optab for given mode. */ | |
bf510679 | 3505 | else if (optab_handler (tab, mode) != CODE_FOR_nothing) |
e72531b9 PK |
3506 | { |
3507 | quotient = gen_reg_rtx (mode); | |
3508 | remainder = gen_reg_rtx (mode); | |
3509 | expand_twoval_binop (tab, op0, op1, quotient, remainder, unsignedp); | |
3510 | } | |
3511 | ||
3512 | /* Generate call to divmod libfunc if it exists. */ | |
bf510679 | 3513 | else if (rtx libfunc = optab_libfunc (tab, mode)) |
e72531b9 PK |
3514 | targetm.expand_divmod_libfunc (libfunc, mode, op0, op1, |
3515 | "ient, &remainder); | |
3516 | ||
3517 | else | |
3518 | gcc_unreachable (); | |
3519 | ||
3520 | /* Wrap the return value (quotient, remainder) within COMPLEX_EXPR. */ | |
3521 | expand_expr (build2 (COMPLEX_EXPR, TREE_TYPE (lhs), | |
3522 | make_tree (TREE_TYPE (arg0), quotient), | |
3523 | make_tree (TREE_TYPE (arg1), remainder)), | |
0b99f253 | 3524 | target, VOIDmode, EXPAND_NORMAL); |
e72531b9 PK |
3525 | } |
3526 | ||
87a5e0e8 RB |
3527 | /* Expand a NOP. */ |
3528 | ||
3529 | static void | |
3530 | expand_NOP (internal_fn, gcall *) | |
3531 | { | |
3532 | /* Nothing. But it shouldn't really prevail. */ | |
3533 | } | |
3534 | ||
49789fd0 IS |
3535 | /* Coroutines, all should have been processed at this stage. */ |
3536 | ||
3537 | static void | |
3538 | expand_CO_FRAME (internal_fn, gcall *) | |
3539 | { | |
3540 | gcc_unreachable (); | |
3541 | } | |
3542 | ||
3543 | static void | |
3544 | expand_CO_YIELD (internal_fn, gcall *) | |
3545 | { | |
3546 | gcc_unreachable (); | |
3547 | } | |
3548 | ||
3549 | static void | |
3550 | expand_CO_SUSPN (internal_fn, gcall *) | |
3551 | { | |
3552 | gcc_unreachable (); | |
3553 | } | |
3554 | ||
3555 | static void | |
3556 | expand_CO_ACTOR (internal_fn, gcall *) | |
3557 | { | |
3558 | gcc_unreachable (); | |
3559 | } | |
3560 | ||
4cfe7a6c RS |
3561 | /* Expand a call to FN using the operands in STMT. FN has a single |
3562 | output operand and NARGS input operands. */ | |
686ee971 RS |
3563 | |
3564 | static void | |
4cfe7a6c RS |
3565 | expand_direct_optab_fn (internal_fn fn, gcall *stmt, direct_optab optab, |
3566 | unsigned int nargs) | |
686ee971 RS |
3567 | { |
3568 | expand_operand *ops = XALLOCAVEC (expand_operand, nargs + 1); | |
3569 | ||
686ee971 RS |
3570 | tree_pair types = direct_internal_fn_types (fn, stmt); |
3571 | insn_code icode = direct_optab_handler (optab, TYPE_MODE (types.first)); | |
41241199 | 3572 | gcc_assert (icode != CODE_FOR_nothing); |
686ee971 RS |
3573 | |
3574 | tree lhs = gimple_call_lhs (stmt); | |
440ed9f8 RS |
3575 | rtx lhs_rtx = NULL_RTX; |
3576 | if (lhs) | |
3577 | lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
b7753f75 RS |
3578 | |
3579 | /* Do not assign directly to a promoted subreg, since there is no | |
3580 | guarantee that the instruction will leave the upper bits of the | |
3581 | register in the state required by SUBREG_PROMOTED_SIGN. */ | |
3582 | rtx dest = lhs_rtx; | |
440ed9f8 | 3583 | if (dest && GET_CODE (dest) == SUBREG && SUBREG_PROMOTED_VAR_P (dest)) |
b7753f75 RS |
3584 | dest = NULL_RTX; |
3585 | ||
3586 | create_output_operand (&ops[0], dest, insn_data[icode].operand[0].mode); | |
686ee971 RS |
3587 | |
3588 | for (unsigned int i = 0; i < nargs; ++i) | |
3589 | { | |
3590 | tree rhs = gimple_call_arg (stmt, i); | |
3591 | tree rhs_type = TREE_TYPE (rhs); | |
3592 | rtx rhs_rtx = expand_normal (rhs); | |
3593 | if (INTEGRAL_TYPE_P (rhs_type)) | |
3594 | create_convert_operand_from (&ops[i + 1], rhs_rtx, | |
3595 | TYPE_MODE (rhs_type), | |
3596 | TYPE_UNSIGNED (rhs_type)); | |
3597 | else | |
3598 | create_input_operand (&ops[i + 1], rhs_rtx, TYPE_MODE (rhs_type)); | |
3599 | } | |
3600 | ||
3601 | expand_insn (icode, nargs + 1, ops); | |
440ed9f8 | 3602 | if (lhs_rtx && !rtx_equal_p (lhs_rtx, ops[0].value)) |
686ee971 | 3603 | { |
ee132692 RS |
3604 | /* If the return value has an integral type, convert the instruction |
3605 | result to that type. This is useful for things that return an | |
3606 | int regardless of the size of the input. If the instruction result | |
3607 | is smaller than required, assume that it is signed. | |
3608 | ||
3609 | If the return value has a nonintegral type, its mode must match | |
3610 | the instruction result. */ | |
3611 | if (GET_CODE (lhs_rtx) == SUBREG && SUBREG_PROMOTED_VAR_P (lhs_rtx)) | |
3612 | { | |
3613 | /* If this is a scalar in a register that is stored in a wider | |
3614 | mode than the declared mode, compute the result into its | |
3615 | declared mode and then convert to the wider mode. */ | |
440ed9f8 | 3616 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
ee132692 RS |
3617 | rtx tmp = convert_to_mode (GET_MODE (lhs_rtx), ops[0].value, 0); |
3618 | convert_move (SUBREG_REG (lhs_rtx), tmp, | |
3619 | SUBREG_PROMOTED_SIGN (lhs_rtx)); | |
3620 | } | |
3621 | else if (GET_MODE (lhs_rtx) == GET_MODE (ops[0].value)) | |
686ee971 | 3622 | emit_move_insn (lhs_rtx, ops[0].value); |
ee132692 RS |
3623 | else |
3624 | { | |
440ed9f8 | 3625 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
ee132692 RS |
3626 | convert_move (lhs_rtx, ops[0].value, 0); |
3627 | } | |
686ee971 RS |
3628 | } |
3629 | } | |
3630 | ||
7cfb4d93 RS |
3631 | /* Expand WHILE_ULT call STMT using optab OPTAB. */ |
3632 | ||
3633 | static void | |
3634 | expand_while_optab_fn (internal_fn, gcall *stmt, convert_optab optab) | |
3635 | { | |
3636 | expand_operand ops[3]; | |
3637 | tree rhs_type[2]; | |
3638 | ||
3639 | tree lhs = gimple_call_lhs (stmt); | |
3640 | tree lhs_type = TREE_TYPE (lhs); | |
3641 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3642 | create_output_operand (&ops[0], lhs_rtx, TYPE_MODE (lhs_type)); | |
3643 | ||
3644 | for (unsigned int i = 0; i < 2; ++i) | |
3645 | { | |
3646 | tree rhs = gimple_call_arg (stmt, i); | |
3647 | rhs_type[i] = TREE_TYPE (rhs); | |
3648 | rtx rhs_rtx = expand_normal (rhs); | |
3649 | create_input_operand (&ops[i + 1], rhs_rtx, TYPE_MODE (rhs_type[i])); | |
3650 | } | |
3651 | ||
3652 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (rhs_type[0]), | |
3653 | TYPE_MODE (lhs_type)); | |
3654 | ||
3655 | expand_insn (icode, 3, ops); | |
3656 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) | |
3657 | emit_move_insn (lhs_rtx, ops[0].value); | |
3658 | } | |
3659 | ||
686ee971 RS |
3660 | /* Expanders for optabs that can use expand_direct_optab_fn. */ |
3661 | ||
4cfe7a6c RS |
3662 | #define expand_unary_optab_fn(FN, STMT, OPTAB) \ |
3663 | expand_direct_optab_fn (FN, STMT, OPTAB, 1) | |
686ee971 | 3664 | |
4cfe7a6c RS |
3665 | #define expand_binary_optab_fn(FN, STMT, OPTAB) \ |
3666 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) | |
686ee971 | 3667 | |
c566cc9f RS |
3668 | #define expand_ternary_optab_fn(FN, STMT, OPTAB) \ |
3669 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3670 | ||
bfe1bb57 | 3671 | #define expand_cond_unary_optab_fn(FN, STMT, OPTAB) \ |
9d4ac06e | 3672 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
bfe1bb57 | 3673 | |
0972596e | 3674 | #define expand_cond_binary_optab_fn(FN, STMT, OPTAB) \ |
9d4ac06e | 3675 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
0972596e | 3676 | |
b41d1f6e RS |
3677 | #define expand_cond_ternary_optab_fn(FN, STMT, OPTAB) \ |
3678 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) | |
3679 | ||
bb6c2b68 RS |
3680 | #define expand_fold_extract_optab_fn(FN, STMT, OPTAB) \ |
3681 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3682 | ||
b781a135 RS |
3683 | #define expand_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3684 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) | |
3685 | ||
bce29d65 AM |
3686 | #define expand_mask_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3687 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3688 | ||
58c036c8 RS |
3689 | #define expand_check_ptrs_optab_fn(FN, STMT, OPTAB) \ |
3690 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) | |
3691 | ||
ab23f5d9 RS |
3692 | /* RETURN_TYPE and ARGS are a return type and argument list that are |
3693 | in principle compatible with FN (which satisfies direct_internal_fn_p). | |
3694 | Return the types that should be used to determine whether the | |
3695 | target supports FN. */ | |
3696 | ||
3697 | tree_pair | |
3698 | direct_internal_fn_types (internal_fn fn, tree return_type, tree *args) | |
3699 | { | |
3700 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
3701 | tree type0 = (info.type0 < 0 ? return_type : TREE_TYPE (args[info.type0])); | |
3702 | tree type1 = (info.type1 < 0 ? return_type : TREE_TYPE (args[info.type1])); | |
3703 | return tree_pair (type0, type1); | |
3704 | } | |
3705 | ||
3706 | /* CALL is a call whose return type and arguments are in principle | |
3707 | compatible with FN (which satisfies direct_internal_fn_p). Return the | |
3708 | types that should be used to determine whether the target supports FN. */ | |
3709 | ||
3710 | tree_pair | |
3711 | direct_internal_fn_types (internal_fn fn, gcall *call) | |
3712 | { | |
3713 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
3714 | tree op0 = (info.type0 < 0 | |
3715 | ? gimple_call_lhs (call) | |
3716 | : gimple_call_arg (call, info.type0)); | |
3717 | tree op1 = (info.type1 < 0 | |
3718 | ? gimple_call_lhs (call) | |
3719 | : gimple_call_arg (call, info.type1)); | |
3720 | return tree_pair (TREE_TYPE (op0), TREE_TYPE (op1)); | |
3721 | } | |
3722 | ||
3723 | /* Return true if OPTAB is supported for TYPES (whose modes should be | |
d95ab70a RS |
3724 | the same) when the optimization type is OPT_TYPE. Used for simple |
3725 | direct optabs. */ | |
ab23f5d9 RS |
3726 | |
3727 | static bool | |
d95ab70a RS |
3728 | direct_optab_supported_p (direct_optab optab, tree_pair types, |
3729 | optimization_type opt_type) | |
ab23f5d9 RS |
3730 | { |
3731 | machine_mode mode = TYPE_MODE (types.first); | |
3732 | gcc_checking_assert (mode == TYPE_MODE (types.second)); | |
d95ab70a | 3733 | return direct_optab_handler (optab, mode, opt_type) != CODE_FOR_nothing; |
ab23f5d9 RS |
3734 | } |
3735 | ||
7cfb4d93 RS |
3736 | /* Return true if OPTAB is supported for TYPES, where the first type |
3737 | is the destination and the second type is the source. Used for | |
3738 | convert optabs. */ | |
3739 | ||
3740 | static bool | |
3741 | convert_optab_supported_p (convert_optab optab, tree_pair types, | |
3742 | optimization_type opt_type) | |
3743 | { | |
3744 | return (convert_optab_handler (optab, TYPE_MODE (types.first), | |
3745 | TYPE_MODE (types.second), opt_type) | |
3746 | != CODE_FOR_nothing); | |
3747 | } | |
3748 | ||
ab23f5d9 | 3749 | /* Return true if load/store lanes optab OPTAB is supported for |
d95ab70a | 3750 | array type TYPES.first when the optimization type is OPT_TYPE. */ |
ab23f5d9 RS |
3751 | |
3752 | static bool | |
d95ab70a RS |
3753 | multi_vector_optab_supported_p (convert_optab optab, tree_pair types, |
3754 | optimization_type opt_type) | |
ab23f5d9 | 3755 | { |
d95ab70a RS |
3756 | gcc_assert (TREE_CODE (types.first) == ARRAY_TYPE); |
3757 | machine_mode imode = TYPE_MODE (types.first); | |
3758 | machine_mode vmode = TYPE_MODE (TREE_TYPE (types.first)); | |
3759 | return (convert_optab_handler (optab, imode, vmode, opt_type) | |
3760 | != CODE_FOR_nothing); | |
ab23f5d9 RS |
3761 | } |
3762 | ||
686ee971 RS |
3763 | #define direct_unary_optab_supported_p direct_optab_supported_p |
3764 | #define direct_binary_optab_supported_p direct_optab_supported_p | |
c566cc9f | 3765 | #define direct_ternary_optab_supported_p direct_optab_supported_p |
bfe1bb57 | 3766 | #define direct_cond_unary_optab_supported_p direct_optab_supported_p |
0972596e | 3767 | #define direct_cond_binary_optab_supported_p direct_optab_supported_p |
b41d1f6e | 3768 | #define direct_cond_ternary_optab_supported_p direct_optab_supported_p |
ef8d1da1 | 3769 | #define direct_mask_load_optab_supported_p convert_optab_supported_p |
ab23f5d9 | 3770 | #define direct_load_lanes_optab_supported_p multi_vector_optab_supported_p |
7e11fc7f | 3771 | #define direct_mask_load_lanes_optab_supported_p multi_vector_optab_supported_p |
09eb042a | 3772 | #define direct_gather_load_optab_supported_p convert_optab_supported_p |
d496134a | 3773 | #define direct_len_load_optab_supported_p direct_optab_supported_p |
ef8d1da1 | 3774 | #define direct_mask_store_optab_supported_p convert_optab_supported_p |
ab23f5d9 | 3775 | #define direct_store_lanes_optab_supported_p multi_vector_optab_supported_p |
7e11fc7f | 3776 | #define direct_mask_store_lanes_optab_supported_p multi_vector_optab_supported_p |
298e76e6 RS |
3777 | #define direct_vec_cond_mask_optab_supported_p convert_optab_supported_p |
3778 | #define direct_vec_cond_optab_supported_p convert_optab_supported_p | |
09eb042a | 3779 | #define direct_scatter_store_optab_supported_p convert_optab_supported_p |
d496134a | 3780 | #define direct_len_store_optab_supported_p direct_optab_supported_p |
7cfb4d93 | 3781 | #define direct_while_optab_supported_p convert_optab_supported_p |
bb6c2b68 | 3782 | #define direct_fold_extract_optab_supported_p direct_optab_supported_p |
b781a135 | 3783 | #define direct_fold_left_optab_supported_p direct_optab_supported_p |
bce29d65 | 3784 | #define direct_mask_fold_left_optab_supported_p direct_optab_supported_p |
58c036c8 | 3785 | #define direct_check_ptrs_optab_supported_p direct_optab_supported_p |
683e55fa | 3786 | #define direct_vec_set_optab_supported_p direct_optab_supported_p |
ab23f5d9 | 3787 | |
16d24520 RS |
3788 | /* Return the optab used by internal function FN. */ |
3789 | ||
3790 | static optab | |
3791 | direct_internal_fn_optab (internal_fn fn, tree_pair types) | |
3792 | { | |
3793 | switch (fn) | |
3794 | { | |
3795 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
3796 | case IFN_##CODE: break; | |
3797 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
3798 | case IFN_##CODE: return OPTAB##_optab; | |
3799 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ | |
3800 | UNSIGNED_OPTAB, TYPE) \ | |
3801 | case IFN_##CODE: return (TYPE_UNSIGNED (types.SELECTOR) \ | |
3802 | ? UNSIGNED_OPTAB ## _optab \ | |
3803 | : SIGNED_OPTAB ## _optab); | |
3804 | #include "internal-fn.def" | |
3805 | ||
3806 | case IFN_LAST: | |
3807 | break; | |
3808 | } | |
3809 | gcc_unreachable (); | |
3810 | } | |
3811 | ||
bfaa08b7 RS |
3812 | /* Return the optab used by internal function FN. */ |
3813 | ||
3814 | static optab | |
3815 | direct_internal_fn_optab (internal_fn fn) | |
3816 | { | |
3817 | switch (fn) | |
3818 | { | |
3819 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
3820 | case IFN_##CODE: break; | |
3821 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
3822 | case IFN_##CODE: return OPTAB##_optab; | |
3823 | #include "internal-fn.def" | |
3824 | ||
3825 | case IFN_LAST: | |
3826 | break; | |
3827 | } | |
3828 | gcc_unreachable (); | |
3829 | } | |
3830 | ||
d95ab70a RS |
3831 | /* Return true if FN is supported for the types in TYPES when the |
3832 | optimization type is OPT_TYPE. The types are those associated with | |
3833 | the "type0" and "type1" fields of FN's direct_internal_fn_info | |
3834 | structure. */ | |
ab23f5d9 RS |
3835 | |
3836 | bool | |
d95ab70a RS |
3837 | direct_internal_fn_supported_p (internal_fn fn, tree_pair types, |
3838 | optimization_type opt_type) | |
ab23f5d9 RS |
3839 | { |
3840 | switch (fn) | |
3841 | { | |
3842 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
3843 | case IFN_##CODE: break; | |
3844 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
3845 | case IFN_##CODE: \ | |
d95ab70a RS |
3846 | return direct_##TYPE##_optab_supported_p (OPTAB##_optab, types, \ |
3847 | opt_type); | |
16d24520 RS |
3848 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
3849 | UNSIGNED_OPTAB, TYPE) \ | |
3850 | case IFN_##CODE: \ | |
3851 | { \ | |
3852 | optab which_optab = (TYPE_UNSIGNED (types.SELECTOR) \ | |
3853 | ? UNSIGNED_OPTAB ## _optab \ | |
3854 | : SIGNED_OPTAB ## _optab); \ | |
3855 | return direct_##TYPE##_optab_supported_p (which_optab, types, \ | |
3856 | opt_type); \ | |
3857 | } | |
ab23f5d9 RS |
3858 | #include "internal-fn.def" |
3859 | ||
3860 | case IFN_LAST: | |
3861 | break; | |
3862 | } | |
3863 | gcc_unreachable (); | |
3864 | } | |
3865 | ||
d95ab70a RS |
3866 | /* Return true if FN is supported for type TYPE when the optimization |
3867 | type is OPT_TYPE. The caller knows that the "type0" and "type1" | |
3868 | fields of FN's direct_internal_fn_info structure are the same. */ | |
ab23f5d9 RS |
3869 | |
3870 | bool | |
d95ab70a RS |
3871 | direct_internal_fn_supported_p (internal_fn fn, tree type, |
3872 | optimization_type opt_type) | |
ab23f5d9 RS |
3873 | { |
3874 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
3875 | gcc_checking_assert (info.type0 == info.type1); | |
d95ab70a | 3876 | return direct_internal_fn_supported_p (fn, tree_pair (type, type), opt_type); |
ab23f5d9 RS |
3877 | } |
3878 | ||
41241199 RL |
3879 | /* Return true if the STMT is supported when the optimization type is OPT_TYPE, |
3880 | given that STMT is a call to a direct internal function. */ | |
3881 | ||
3882 | bool | |
3883 | direct_internal_fn_supported_p (gcall *stmt, optimization_type opt_type) | |
3884 | { | |
3885 | internal_fn fn = gimple_call_internal_fn (stmt); | |
3886 | tree_pair types = direct_internal_fn_types (fn, stmt); | |
3887 | return direct_internal_fn_supported_p (fn, types, opt_type); | |
3888 | } | |
3889 | ||
30213ae9 | 3890 | /* Return true if FN is a binary operation and if FN is commutative. */ |
0246112a | 3891 | |
30213ae9 RS |
3892 | bool |
3893 | commutative_binary_fn_p (internal_fn fn) | |
0246112a RS |
3894 | { |
3895 | switch (fn) | |
3896 | { | |
0246112a RS |
3897 | case IFN_AVG_FLOOR: |
3898 | case IFN_AVG_CEIL: | |
a1d27560 | 3899 | case IFN_MULH: |
58cc9876 YW |
3900 | case IFN_MULHS: |
3901 | case IFN_MULHRS: | |
0246112a RS |
3902 | case IFN_FMIN: |
3903 | case IFN_FMAX: | |
71207246 | 3904 | case IFN_COMPLEX_MUL: |
12e38012 RS |
3905 | case IFN_UBSAN_CHECK_ADD: |
3906 | case IFN_UBSAN_CHECK_MUL: | |
b1d15146 RS |
3907 | case IFN_ADD_OVERFLOW: |
3908 | case IFN_MUL_OVERFLOW: | |
30213ae9 RS |
3909 | return true; |
3910 | ||
3911 | default: | |
3912 | return false; | |
3913 | } | |
3914 | } | |
3915 | ||
3916 | /* Return true if FN is a ternary operation and if its first two arguments | |
3917 | are commutative. */ | |
3918 | ||
3919 | bool | |
3920 | commutative_ternary_fn_p (internal_fn fn) | |
3921 | { | |
3922 | switch (fn) | |
3923 | { | |
3924 | case IFN_FMA: | |
3925 | case IFN_FMS: | |
3926 | case IFN_FNMA: | |
3927 | case IFN_FNMS: | |
3928 | return true; | |
3929 | ||
3930 | default: | |
3931 | return false; | |
3932 | } | |
3933 | } | |
3934 | ||
3935 | /* Return true if FN is an associative binary operation. */ | |
3936 | ||
3937 | bool | |
3938 | associative_binary_fn_p (internal_fn fn) | |
3939 | { | |
3940 | switch (fn) | |
3941 | { | |
3942 | case IFN_FMIN: | |
3943 | case IFN_FMAX: | |
3944 | return true; | |
0246112a | 3945 | |
30213ae9 RS |
3946 | default: |
3947 | return false; | |
3948 | } | |
3949 | } | |
3950 | ||
3951 | /* If FN is commutative in two consecutive arguments, return the | |
3952 | index of the first, otherwise return -1. */ | |
3953 | ||
3954 | int | |
3955 | first_commutative_argument (internal_fn fn) | |
3956 | { | |
3957 | switch (fn) | |
3958 | { | |
0246112a RS |
3959 | case IFN_COND_ADD: |
3960 | case IFN_COND_MUL: | |
3961 | case IFN_COND_MIN: | |
3962 | case IFN_COND_MAX: | |
70613000 RS |
3963 | case IFN_COND_FMIN: |
3964 | case IFN_COND_FMAX: | |
0246112a RS |
3965 | case IFN_COND_AND: |
3966 | case IFN_COND_IOR: | |
3967 | case IFN_COND_XOR: | |
3968 | case IFN_COND_FMA: | |
3969 | case IFN_COND_FMS: | |
3970 | case IFN_COND_FNMA: | |
3971 | case IFN_COND_FNMS: | |
3972 | return 1; | |
3973 | ||
3974 | default: | |
30213ae9 RS |
3975 | if (commutative_binary_fn_p (fn) |
3976 | || commutative_ternary_fn_p (fn)) | |
3977 | return 0; | |
0246112a RS |
3978 | return -1; |
3979 | } | |
3980 | } | |
3981 | ||
883cabde RS |
3982 | /* Return true if IFN_SET_EDOM is supported. */ |
3983 | ||
3984 | bool | |
3985 | set_edom_supported_p (void) | |
3986 | { | |
3987 | #ifdef TARGET_EDOM | |
3988 | return true; | |
3989 | #else | |
3990 | return false; | |
3991 | #endif | |
3992 | } | |
3993 | ||
ab23f5d9 RS |
3994 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
3995 | static void \ | |
4cfe7a6c | 3996 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
ab23f5d9 | 3997 | { \ |
4cfe7a6c | 3998 | expand_##TYPE##_optab_fn (fn, stmt, OPTAB##_optab); \ |
ab23f5d9 | 3999 | } |
16d24520 RS |
4000 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4001 | UNSIGNED_OPTAB, TYPE) \ | |
4002 | static void \ | |
4003 | expand_##CODE (internal_fn fn, gcall *stmt) \ | |
4004 | { \ | |
4005 | tree_pair types = direct_internal_fn_types (fn, stmt); \ | |
4006 | optab which_optab = direct_internal_fn_optab (fn, types); \ | |
4007 | expand_##TYPE##_optab_fn (fn, stmt, which_optab); \ | |
4008 | } | |
ab23f5d9 RS |
4009 | #include "internal-fn.def" |
4010 | ||
25583c4f RS |
4011 | /* Routines to expand each internal function, indexed by function number. |
4012 | Each routine has the prototype: | |
4013 | ||
538dd0b7 | 4014 | expand_<NAME> (gcall *stmt) |
25583c4f RS |
4015 | |
4016 | where STMT is the statement that performs the call. */ | |
4cfe7a6c | 4017 | static void (*const internal_fn_expanders[]) (internal_fn, gcall *) = { |
b78475cf | 4018 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) expand_##CODE, |
25583c4f | 4019 | #include "internal-fn.def" |
25583c4f RS |
4020 | 0 |
4021 | }; | |
4022 | ||
6a86928d RS |
4023 | /* Invoke T(CODE, IFN) for each conditional function IFN that maps to a |
4024 | tree code CODE. */ | |
4025 | #define FOR_EACH_CODE_MAPPING(T) \ | |
4026 | T (PLUS_EXPR, IFN_COND_ADD) \ | |
4027 | T (MINUS_EXPR, IFN_COND_SUB) \ | |
4028 | T (MULT_EXPR, IFN_COND_MUL) \ | |
4029 | T (TRUNC_DIV_EXPR, IFN_COND_DIV) \ | |
4030 | T (TRUNC_MOD_EXPR, IFN_COND_MOD) \ | |
4031 | T (RDIV_EXPR, IFN_COND_RDIV) \ | |
4032 | T (MIN_EXPR, IFN_COND_MIN) \ | |
4033 | T (MAX_EXPR, IFN_COND_MAX) \ | |
4034 | T (BIT_AND_EXPR, IFN_COND_AND) \ | |
4035 | T (BIT_IOR_EXPR, IFN_COND_IOR) \ | |
20103c0e RS |
4036 | T (BIT_XOR_EXPR, IFN_COND_XOR) \ |
4037 | T (LSHIFT_EXPR, IFN_COND_SHL) \ | |
20dcda98 | 4038 | T (RSHIFT_EXPR, IFN_COND_SHR) \ |
4039 | T (NEGATE_EXPR, IFN_COND_NEG) | |
6a86928d | 4040 | |
9d4ac06e RS |
4041 | /* Return a function that only performs CODE when a certain condition is met |
4042 | and that uses a given fallback value otherwise. For example, if CODE is | |
4043 | a binary operation associated with conditional function FN: | |
4044 | ||
4045 | LHS = FN (COND, A, B, ELSE) | |
4046 | ||
4047 | is equivalent to the C expression: | |
4048 | ||
4049 | LHS = COND ? A CODE B : ELSE; | |
0972596e | 4050 | |
9d4ac06e | 4051 | operating elementwise if the operands are vectors. |
0972596e | 4052 | |
9d4ac06e | 4053 | Return IFN_LAST if no such function exists. */ |
0972596e RS |
4054 | |
4055 | internal_fn | |
4056 | get_conditional_internal_fn (tree_code code) | |
4057 | { | |
4058 | switch (code) | |
4059 | { | |
6a86928d RS |
4060 | #define CASE(CODE, IFN) case CODE: return IFN; |
4061 | FOR_EACH_CODE_MAPPING(CASE) | |
4062 | #undef CASE | |
0972596e RS |
4063 | default: |
4064 | return IFN_LAST; | |
4065 | } | |
4066 | } | |
4067 | ||
6a86928d RS |
4068 | /* If IFN implements the conditional form of a tree code, return that |
4069 | tree code, otherwise return ERROR_MARK. */ | |
4070 | ||
4071 | tree_code | |
4072 | conditional_internal_fn_code (internal_fn ifn) | |
4073 | { | |
4074 | switch (ifn) | |
4075 | { | |
4076 | #define CASE(CODE, IFN) case IFN: return CODE; | |
4077 | FOR_EACH_CODE_MAPPING(CASE) | |
4078 | #undef CASE | |
4079 | default: | |
4080 | return ERROR_MARK; | |
4081 | } | |
4082 | } | |
4083 | ||
b41d1f6e RS |
4084 | /* Invoke T(IFN) for each internal function IFN that also has an |
4085 | IFN_COND_* form. */ | |
4086 | #define FOR_EACH_COND_FN_PAIR(T) \ | |
70613000 RS |
4087 | T (FMAX) \ |
4088 | T (FMIN) \ | |
b41d1f6e RS |
4089 | T (FMA) \ |
4090 | T (FMS) \ | |
4091 | T (FNMA) \ | |
4092 | T (FNMS) | |
4093 | ||
4094 | /* Return a function that only performs internal function FN when a | |
4095 | certain condition is met and that uses a given fallback value otherwise. | |
4096 | In other words, the returned function FN' is such that: | |
4097 | ||
4098 | LHS = FN' (COND, A1, ... An, ELSE) | |
4099 | ||
4100 | is equivalent to the C expression: | |
4101 | ||
4102 | LHS = COND ? FN (A1, ..., An) : ELSE; | |
4103 | ||
4104 | operating elementwise if the operands are vectors. | |
4105 | ||
4106 | Return IFN_LAST if no such function exists. */ | |
4107 | ||
4108 | internal_fn | |
4109 | get_conditional_internal_fn (internal_fn fn) | |
4110 | { | |
4111 | switch (fn) | |
4112 | { | |
4113 | #define CASE(NAME) case IFN_##NAME: return IFN_COND_##NAME; | |
4114 | FOR_EACH_COND_FN_PAIR(CASE) | |
4115 | #undef CASE | |
4116 | default: | |
4117 | return IFN_LAST; | |
4118 | } | |
4119 | } | |
4120 | ||
4121 | /* If IFN implements the conditional form of an unconditional internal | |
4122 | function, return that unconditional function, otherwise return IFN_LAST. */ | |
4123 | ||
4124 | internal_fn | |
4125 | get_unconditional_internal_fn (internal_fn ifn) | |
4126 | { | |
4127 | switch (ifn) | |
4128 | { | |
4129 | #define CASE(NAME) case IFN_COND_##NAME: return IFN_##NAME; | |
4130 | FOR_EACH_COND_FN_PAIR(CASE) | |
4131 | #undef CASE | |
4132 | default: | |
4133 | return IFN_LAST; | |
4134 | } | |
4135 | } | |
4136 | ||
0936858f RS |
4137 | /* Return true if STMT can be interpreted as a conditional tree code |
4138 | operation of the form: | |
4139 | ||
4140 | LHS = COND ? OP (RHS1, ...) : ELSE; | |
4141 | ||
4142 | operating elementwise if the operands are vectors. This includes | |
4143 | the case of an all-true COND, so that the operation always happens. | |
4144 | ||
4145 | When returning true, set: | |
4146 | ||
4147 | - *COND_OUT to the condition COND, or to NULL_TREE if the condition | |
4148 | is known to be all-true | |
4149 | - *CODE_OUT to the tree code | |
4150 | - OPS[I] to operand I of *CODE_OUT | |
4151 | - *ELSE_OUT to the fallback value ELSE, or to NULL_TREE if the | |
4152 | condition is known to be all true. */ | |
4153 | ||
4154 | bool | |
4155 | can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out, | |
4156 | tree_code *code_out, | |
4157 | tree (&ops)[3], tree *else_out) | |
4158 | { | |
4159 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
4160 | { | |
4161 | *cond_out = NULL_TREE; | |
4162 | *code_out = gimple_assign_rhs_code (assign); | |
4163 | ops[0] = gimple_assign_rhs1 (assign); | |
4164 | ops[1] = gimple_assign_rhs2 (assign); | |
4165 | ops[2] = gimple_assign_rhs3 (assign); | |
4166 | *else_out = NULL_TREE; | |
4167 | return true; | |
4168 | } | |
4169 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
4170 | if (gimple_call_internal_p (call)) | |
4171 | { | |
4172 | internal_fn ifn = gimple_call_internal_fn (call); | |
4173 | tree_code code = conditional_internal_fn_code (ifn); | |
4174 | if (code != ERROR_MARK) | |
4175 | { | |
4176 | *cond_out = gimple_call_arg (call, 0); | |
4177 | *code_out = code; | |
4178 | unsigned int nops = gimple_call_num_args (call) - 2; | |
4179 | for (unsigned int i = 0; i < 3; ++i) | |
4180 | ops[i] = i < nops ? gimple_call_arg (call, i + 1) : NULL_TREE; | |
4181 | *else_out = gimple_call_arg (call, nops + 1); | |
4182 | if (integer_truep (*cond_out)) | |
4183 | { | |
4184 | *cond_out = NULL_TREE; | |
4185 | *else_out = NULL_TREE; | |
4186 | } | |
4187 | return true; | |
4188 | } | |
4189 | } | |
4190 | return false; | |
4191 | } | |
4192 | ||
bfaa08b7 RS |
4193 | /* Return true if IFN is some form of load from memory. */ |
4194 | ||
4195 | bool | |
4196 | internal_load_fn_p (internal_fn fn) | |
4197 | { | |
4198 | switch (fn) | |
4199 | { | |
4200 | case IFN_MASK_LOAD: | |
4201 | case IFN_LOAD_LANES: | |
4202 | case IFN_MASK_LOAD_LANES: | |
4203 | case IFN_GATHER_LOAD: | |
4204 | case IFN_MASK_GATHER_LOAD: | |
d496134a | 4205 | case IFN_LEN_LOAD: |
bfaa08b7 RS |
4206 | return true; |
4207 | ||
4208 | default: | |
4209 | return false; | |
4210 | } | |
4211 | } | |
4212 | ||
f307441a RS |
4213 | /* Return true if IFN is some form of store to memory. */ |
4214 | ||
4215 | bool | |
4216 | internal_store_fn_p (internal_fn fn) | |
4217 | { | |
4218 | switch (fn) | |
4219 | { | |
4220 | case IFN_MASK_STORE: | |
4221 | case IFN_STORE_LANES: | |
4222 | case IFN_MASK_STORE_LANES: | |
4223 | case IFN_SCATTER_STORE: | |
4224 | case IFN_MASK_SCATTER_STORE: | |
d496134a | 4225 | case IFN_LEN_STORE: |
f307441a RS |
4226 | return true; |
4227 | ||
4228 | default: | |
4229 | return false; | |
4230 | } | |
4231 | } | |
4232 | ||
bfaa08b7 RS |
4233 | /* Return true if IFN is some form of gather load or scatter store. */ |
4234 | ||
4235 | bool | |
4236 | internal_gather_scatter_fn_p (internal_fn fn) | |
4237 | { | |
4238 | switch (fn) | |
4239 | { | |
4240 | case IFN_GATHER_LOAD: | |
4241 | case IFN_MASK_GATHER_LOAD: | |
f307441a RS |
4242 | case IFN_SCATTER_STORE: |
4243 | case IFN_MASK_SCATTER_STORE: | |
bfaa08b7 RS |
4244 | return true; |
4245 | ||
4246 | default: | |
4247 | return false; | |
4248 | } | |
4249 | } | |
4250 | ||
4251 | /* If FN takes a vector mask argument, return the index of that argument, | |
4252 | otherwise return -1. */ | |
4253 | ||
4254 | int | |
4255 | internal_fn_mask_index (internal_fn fn) | |
4256 | { | |
4257 | switch (fn) | |
4258 | { | |
4259 | case IFN_MASK_LOAD: | |
4260 | case IFN_MASK_LOAD_LANES: | |
4261 | case IFN_MASK_STORE: | |
4262 | case IFN_MASK_STORE_LANES: | |
4263 | return 2; | |
4264 | ||
4265 | case IFN_MASK_GATHER_LOAD: | |
f307441a RS |
4266 | case IFN_MASK_SCATTER_STORE: |
4267 | return 4; | |
4268 | ||
4269 | default: | |
2c58d42c RS |
4270 | return (conditional_internal_fn_code (fn) != ERROR_MARK |
4271 | || get_unconditional_internal_fn (fn) != IFN_LAST ? 0 : -1); | |
f307441a RS |
4272 | } |
4273 | } | |
4274 | ||
4275 | /* If FN takes a value that should be stored to memory, return the index | |
4276 | of that argument, otherwise return -1. */ | |
4277 | ||
4278 | int | |
4279 | internal_fn_stored_value_index (internal_fn fn) | |
4280 | { | |
4281 | switch (fn) | |
4282 | { | |
4283 | case IFN_MASK_STORE: | |
bd28b730 | 4284 | case IFN_MASK_STORE_LANES: |
f307441a RS |
4285 | case IFN_SCATTER_STORE: |
4286 | case IFN_MASK_SCATTER_STORE: | |
d496134a | 4287 | case IFN_LEN_STORE: |
f307441a RS |
4288 | return 3; |
4289 | ||
bfaa08b7 RS |
4290 | default: |
4291 | return -1; | |
4292 | } | |
4293 | } | |
4294 | ||
4295 | /* Return true if the target supports gather load or scatter store function | |
4296 | IFN. For loads, VECTOR_TYPE is the vector type of the load result, | |
4297 | while for stores it is the vector type of the stored data argument. | |
4298 | MEMORY_ELEMENT_TYPE is the type of the memory elements being loaded | |
09eb042a RS |
4299 | or stored. OFFSET_VECTOR_TYPE is the vector type that holds the |
4300 | offset from the shared base address of each loaded or stored element. | |
4301 | SCALE is the amount by which these offsets should be multiplied | |
4302 | *after* they have been extended to address width. */ | |
bfaa08b7 RS |
4303 | |
4304 | bool | |
4305 | internal_gather_scatter_fn_supported_p (internal_fn ifn, tree vector_type, | |
4306 | tree memory_element_type, | |
09eb042a | 4307 | tree offset_vector_type, int scale) |
bfaa08b7 RS |
4308 | { |
4309 | if (!tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vector_type)), | |
4310 | TYPE_SIZE (memory_element_type))) | |
4311 | return false; | |
09eb042a RS |
4312 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vector_type), |
4313 | TYPE_VECTOR_SUBPARTS (offset_vector_type))) | |
4314 | return false; | |
bfaa08b7 | 4315 | optab optab = direct_internal_fn_optab (ifn); |
09eb042a RS |
4316 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (vector_type), |
4317 | TYPE_MODE (offset_vector_type)); | |
f307441a | 4318 | int output_ops = internal_load_fn_p (ifn) ? 1 : 0; |
09eb042a | 4319 | bool unsigned_p = TYPE_UNSIGNED (TREE_TYPE (offset_vector_type)); |
bfaa08b7 | 4320 | return (icode != CODE_FOR_nothing |
09eb042a RS |
4321 | && insn_operand_matches (icode, 2 + output_ops, GEN_INT (unsigned_p)) |
4322 | && insn_operand_matches (icode, 3 + output_ops, GEN_INT (scale))); | |
bfaa08b7 RS |
4323 | } |
4324 | ||
58c036c8 RS |
4325 | /* Return true if the target supports IFN_CHECK_{RAW,WAR}_PTRS function IFN |
4326 | for pointers of type TYPE when the accesses have LENGTH bytes and their | |
4327 | common byte alignment is ALIGN. */ | |
4328 | ||
4329 | bool | |
4330 | internal_check_ptrs_fn_supported_p (internal_fn ifn, tree type, | |
4331 | poly_uint64 length, unsigned int align) | |
4332 | { | |
4333 | machine_mode mode = TYPE_MODE (type); | |
4334 | optab optab = direct_internal_fn_optab (ifn); | |
4335 | insn_code icode = direct_optab_handler (optab, mode); | |
4336 | if (icode == CODE_FOR_nothing) | |
4337 | return false; | |
4338 | rtx length_rtx = immed_wide_int_const (length, mode); | |
4339 | return (insn_operand_matches (icode, 3, length_rtx) | |
4340 | && insn_operand_matches (icode, 4, GEN_INT (align))); | |
4341 | } | |
4342 | ||
b0e51639 RD |
4343 | /* Return the supported bias for IFN which is either IFN_LEN_LOAD |
4344 | or IFN_LEN_STORE. For now we only support the biases of 0 and -1 | |
4345 | (in case 0 is not an allowable length for len_load or len_store). | |
4346 | If none of the biases match what the backend provides, return | |
4347 | VECT_PARTIAL_BIAS_UNSUPPORTED. */ | |
4348 | ||
4349 | signed char | |
4350 | internal_len_load_store_bias (internal_fn ifn, machine_mode mode) | |
4351 | { | |
4352 | optab optab = direct_internal_fn_optab (ifn); | |
4353 | insn_code icode = direct_optab_handler (optab, mode); | |
4354 | ||
4355 | if (icode != CODE_FOR_nothing) | |
4356 | { | |
4357 | /* For now we only support biases of 0 or -1. Try both of them. */ | |
4358 | if (insn_operand_matches (icode, 3, GEN_INT (0))) | |
4359 | return 0; | |
4360 | if (insn_operand_matches (icode, 3, GEN_INT (-1))) | |
4361 | return -1; | |
4362 | } | |
4363 | ||
4364 | return VECT_PARTIAL_BIAS_UNSUPPORTED; | |
4365 | } | |
4366 | ||
4cfe7a6c RS |
4367 | /* Expand STMT as though it were a call to internal function FN. */ |
4368 | ||
4369 | void | |
4370 | expand_internal_call (internal_fn fn, gcall *stmt) | |
4371 | { | |
4372 | internal_fn_expanders[fn] (fn, stmt); | |
4373 | } | |
4374 | ||
25583c4f RS |
4375 | /* Expand STMT, which is a call to internal function FN. */ |
4376 | ||
4377 | void | |
538dd0b7 | 4378 | expand_internal_call (gcall *stmt) |
25583c4f | 4379 | { |
4cfe7a6c | 4380 | expand_internal_call (gimple_call_internal_fn (stmt), stmt); |
25583c4f | 4381 | } |
1ee62b92 | 4382 | |
2c58d42c RS |
4383 | /* If TYPE is a vector type, return true if IFN is a direct internal |
4384 | function that is supported for that type. If TYPE is a scalar type, | |
4385 | return true if IFN is a direct internal function that is supported for | |
4386 | the target's preferred vector version of TYPE. */ | |
4387 | ||
4388 | bool | |
4389 | vectorized_internal_fn_supported_p (internal_fn ifn, tree type) | |
4390 | { | |
0ae469e8 RS |
4391 | if (VECTOR_MODE_P (TYPE_MODE (type))) |
4392 | return direct_internal_fn_supported_p (ifn, type, OPTIMIZE_FOR_SPEED); | |
4393 | ||
2c58d42c | 4394 | scalar_mode smode; |
0ae469e8 RS |
4395 | if (!is_a <scalar_mode> (TYPE_MODE (type), &smode)) |
4396 | return false; | |
4397 | ||
4398 | machine_mode vmode = targetm.vectorize.preferred_simd_mode (smode); | |
4399 | if (VECTOR_MODE_P (vmode)) | |
2c58d42c | 4400 | { |
0ae469e8 RS |
4401 | tree vectype = build_vector_type_for_mode (type, vmode); |
4402 | if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED)) | |
4403 | return true; | |
2c58d42c RS |
4404 | } |
4405 | ||
0ae469e8 RS |
4406 | auto_vector_modes vector_modes; |
4407 | targetm.vectorize.autovectorize_vector_modes (&vector_modes, true); | |
4408 | for (machine_mode base_mode : vector_modes) | |
4409 | if (related_vector_mode (base_mode, smode).exists (&vmode)) | |
4410 | { | |
4411 | tree vectype = build_vector_type_for_mode (type, vmode); | |
4412 | if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED)) | |
4413 | return true; | |
4414 | } | |
4415 | ||
4416 | return false; | |
2c58d42c RS |
4417 | } |
4418 | ||
ef8176e0 RB |
4419 | void |
4420 | expand_SHUFFLEVECTOR (internal_fn, gcall *) | |
4421 | { | |
4422 | gcc_unreachable (); | |
4423 | } | |
4424 | ||
1ee62b92 PG |
4425 | void |
4426 | expand_PHI (internal_fn, gcall *) | |
4427 | { | |
4428 | gcc_unreachable (); | |
4429 | } |