]>
Commit | Line | Data |
---|---|---|
25583c4f | 1 | /* Internal functions. |
aeee4812 | 2 | Copyright (C) 2011-2023 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 | ||
2f482a07 AV |
93 | /* Geven an internal_fn IFN that is a widening function, return its |
94 | corresponding LO and HI internal_fns. */ | |
95 | ||
96 | extern void | |
97 | lookup_hilo_internal_fn (internal_fn ifn, internal_fn *lo, internal_fn *hi) | |
98 | { | |
99 | gcc_assert (widening_fn_p (ifn)); | |
100 | ||
101 | switch (ifn) | |
102 | { | |
103 | default: | |
104 | gcc_unreachable (); | |
105 | #undef DEF_INTERNAL_FN | |
106 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
107 | #define DEF_INTERNAL_FN(NAME, FLAGS, TYPE) | |
108 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ | |
109 | case IFN_##NAME: \ | |
110 | *lo = internal_fn (IFN_##NAME##_LO); \ | |
111 | *hi = internal_fn (IFN_##NAME##_HI); \ | |
112 | break; | |
113 | #include "internal-fn.def" | |
114 | #undef DEF_INTERNAL_FN | |
115 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
116 | } | |
117 | } | |
118 | ||
119 | /* Given an internal_fn IFN that is a widening function, return its | |
120 | corresponding _EVEN and _ODD internal_fns in *EVEN and *ODD. */ | |
121 | ||
122 | extern void | |
123 | lookup_evenodd_internal_fn (internal_fn ifn, internal_fn *even, | |
124 | internal_fn *odd) | |
125 | { | |
126 | gcc_assert (widening_fn_p (ifn)); | |
127 | ||
128 | switch (ifn) | |
129 | { | |
130 | default: | |
131 | gcc_unreachable (); | |
132 | #undef DEF_INTERNAL_FN | |
133 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
134 | #define DEF_INTERNAL_FN(NAME, FLAGS, TYPE) | |
135 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ | |
136 | case IFN_##NAME: \ | |
137 | *even = internal_fn (IFN_##NAME##_EVEN); \ | |
138 | *odd = internal_fn (IFN_##NAME##_ODD); \ | |
139 | break; | |
140 | #include "internal-fn.def" | |
141 | #undef DEF_INTERNAL_FN | |
142 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
143 | } | |
144 | } | |
145 | ||
146 | ||
b78475cf YG |
147 | /* Fnspec of each internal function, indexed by function number. */ |
148 | const_tree internal_fn_fnspec_array[IFN_LAST + 1]; | |
149 | ||
150 | void | |
151 | init_internal_fns () | |
152 | { | |
153 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
154 | if (FNSPEC) internal_fn_fnspec_array[IFN_##CODE] = \ | |
ba663ac1 | 155 | build_string ((int) sizeof (FNSPEC) - 1, FNSPEC ? FNSPEC : ""); |
b78475cf | 156 | #include "internal-fn.def" |
b78475cf YG |
157 | internal_fn_fnspec_array[IFN_LAST] = 0; |
158 | } | |
159 | ||
ab23f5d9 RS |
160 | /* Create static initializers for the information returned by |
161 | direct_internal_fn. */ | |
9d2fe6d4 RS |
162 | #define not_direct { -2, -2, false } |
163 | #define mask_load_direct { -1, 2, false } | |
164 | #define load_lanes_direct { -1, -1, false } | |
165 | #define mask_load_lanes_direct { -1, -1, false } | |
166 | #define gather_load_direct { 3, 1, false } | |
167 | #define len_load_direct { -1, -1, false } | |
bd68b33f | 168 | #define mask_len_load_direct { -1, 4, false } |
9d2fe6d4 RS |
169 | #define mask_store_direct { 3, 2, false } |
170 | #define store_lanes_direct { 0, 0, false } | |
171 | #define mask_store_lanes_direct { 0, 0, false } | |
172 | #define vec_cond_mask_direct { 1, 0, false } | |
173 | #define vec_cond_direct { 2, 0, false } | |
174 | #define scatter_store_direct { 3, 1, false } | |
175 | #define len_store_direct { 3, 3, false } | |
bd68b33f | 176 | #define mask_len_store_direct { 4, 5, false } |
9d2fe6d4 | 177 | #define vec_set_direct { 3, 3, false } |
c30efd8c | 178 | #define vec_extract_direct { 3, 3, false } |
9d2fe6d4 | 179 | #define unary_direct { 0, 0, true } |
00eab0c6 | 180 | #define unary_convert_direct { -1, 0, true } |
9d2fe6d4 RS |
181 | #define binary_direct { 0, 0, true } |
182 | #define ternary_direct { 0, 0, true } | |
183 | #define cond_unary_direct { 1, 1, true } | |
184 | #define cond_binary_direct { 1, 1, true } | |
185 | #define cond_ternary_direct { 1, 1, true } | |
6c96d1e4 JZZ |
186 | #define cond_len_unary_direct { 1, 1, true } |
187 | #define cond_len_binary_direct { 1, 1, true } | |
188 | #define cond_len_ternary_direct { 1, 1, true } | |
9d2fe6d4 RS |
189 | #define while_direct { 0, 2, false } |
190 | #define fold_extract_direct { 2, 2, false } | |
191 | #define fold_left_direct { 1, 1, false } | |
192 | #define mask_fold_left_direct { 1, 1, false } | |
ba49332b | 193 | #define mask_len_fold_left_direct { 1, 1, false } |
9d2fe6d4 | 194 | #define check_ptrs_direct { 0, 0, false } |
ab23f5d9 RS |
195 | |
196 | const direct_internal_fn_info direct_internal_fn_array[IFN_LAST + 1] = { | |
197 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) not_direct, | |
198 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) TYPE##_direct, | |
16d24520 RS |
199 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
200 | UNSIGNED_OPTAB, TYPE) TYPE##_direct, | |
ab23f5d9 RS |
201 | #include "internal-fn.def" |
202 | not_direct | |
203 | }; | |
204 | ||
1d205dba RS |
205 | /* Expand STMT using instruction ICODE. The instruction has NOUTPUTS |
206 | output operands and NINPUTS input operands, where NOUTPUTS is either | |
207 | 0 or 1. The output operand (if any) comes first, followed by the | |
208 | NINPUTS input operands. */ | |
209 | ||
210 | static void | |
211 | expand_fn_using_insn (gcall *stmt, insn_code icode, unsigned int noutputs, | |
212 | unsigned int ninputs) | |
213 | { | |
214 | gcc_assert (icode != CODE_FOR_nothing); | |
215 | ||
216 | expand_operand *ops = XALLOCAVEC (expand_operand, noutputs + ninputs); | |
217 | unsigned int opno = 0; | |
218 | rtx lhs_rtx = NULL_RTX; | |
219 | tree lhs = gimple_call_lhs (stmt); | |
220 | ||
221 | if (noutputs) | |
222 | { | |
223 | gcc_assert (noutputs == 1); | |
224 | if (lhs) | |
225 | lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
226 | ||
227 | /* Do not assign directly to a promoted subreg, since there is no | |
228 | guarantee that the instruction will leave the upper bits of the | |
229 | register in the state required by SUBREG_PROMOTED_SIGN. */ | |
230 | rtx dest = lhs_rtx; | |
231 | if (dest && GET_CODE (dest) == SUBREG && SUBREG_PROMOTED_VAR_P (dest)) | |
232 | dest = NULL_RTX; | |
233 | create_output_operand (&ops[opno], dest, | |
234 | insn_data[icode].operand[opno].mode); | |
235 | opno += 1; | |
236 | } | |
237 | else | |
238 | gcc_assert (!lhs); | |
239 | ||
240 | for (unsigned int i = 0; i < ninputs; ++i) | |
241 | { | |
242 | tree rhs = gimple_call_arg (stmt, i); | |
243 | tree rhs_type = TREE_TYPE (rhs); | |
244 | rtx rhs_rtx = expand_normal (rhs); | |
245 | if (INTEGRAL_TYPE_P (rhs_type)) | |
246 | create_convert_operand_from (&ops[opno], rhs_rtx, | |
247 | TYPE_MODE (rhs_type), | |
248 | TYPE_UNSIGNED (rhs_type)); | |
249 | else | |
250 | create_input_operand (&ops[opno], rhs_rtx, TYPE_MODE (rhs_type)); | |
251 | opno += 1; | |
252 | } | |
253 | ||
254 | gcc_assert (opno == noutputs + ninputs); | |
255 | expand_insn (icode, opno, ops); | |
256 | if (lhs_rtx && !rtx_equal_p (lhs_rtx, ops[0].value)) | |
257 | { | |
258 | /* If the return value has an integral type, convert the instruction | |
259 | result to that type. This is useful for things that return an | |
260 | int regardless of the size of the input. If the instruction result | |
261 | is smaller than required, assume that it is signed. | |
262 | ||
263 | If the return value has a nonintegral type, its mode must match | |
264 | the instruction result. */ | |
265 | if (GET_CODE (lhs_rtx) == SUBREG && SUBREG_PROMOTED_VAR_P (lhs_rtx)) | |
266 | { | |
267 | /* If this is a scalar in a register that is stored in a wider | |
268 | mode than the declared mode, compute the result into its | |
269 | declared mode and then convert to the wider mode. */ | |
270 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); | |
271 | rtx tmp = convert_to_mode (GET_MODE (lhs_rtx), ops[0].value, 0); | |
272 | convert_move (SUBREG_REG (lhs_rtx), tmp, | |
273 | SUBREG_PROMOTED_SIGN (lhs_rtx)); | |
274 | } | |
275 | else if (GET_MODE (lhs_rtx) == GET_MODE (ops[0].value)) | |
276 | emit_move_insn (lhs_rtx, ops[0].value); | |
277 | else | |
278 | { | |
279 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); | |
280 | convert_move (lhs_rtx, ops[0].value, 0); | |
281 | } | |
282 | } | |
283 | } | |
284 | ||
272c6793 | 285 | /* ARRAY_TYPE is an array of vector modes. Return the associated insn |
ab23f5d9 | 286 | for load-lanes-style optab OPTAB, or CODE_FOR_nothing if none. */ |
272c6793 RS |
287 | |
288 | static enum insn_code | |
289 | get_multi_vector_move (tree array_type, convert_optab optab) | |
290 | { | |
ef4bddc2 RS |
291 | machine_mode imode; |
292 | machine_mode vmode; | |
272c6793 RS |
293 | |
294 | gcc_assert (TREE_CODE (array_type) == ARRAY_TYPE); | |
295 | imode = TYPE_MODE (array_type); | |
296 | vmode = TYPE_MODE (TREE_TYPE (array_type)); | |
297 | ||
ab23f5d9 | 298 | return convert_optab_handler (optab, imode, vmode); |
272c6793 RS |
299 | } |
300 | ||
363bb3dc | 301 | /* Add mask and len arguments according to the STMT. */ |
b8806f6f JZZ |
302 | |
303 | static unsigned int | |
363bb3dc | 304 | add_mask_and_len_args (expand_operand *ops, unsigned int opno, gcall *stmt) |
b8806f6f JZZ |
305 | { |
306 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
307 | int len_index = internal_fn_len_index (ifn); | |
308 | /* BIAS is always consecutive next of LEN. */ | |
309 | int bias_index = len_index + 1; | |
310 | int mask_index = internal_fn_mask_index (ifn); | |
311 | /* The order of arguments are always {len,bias,mask}. */ | |
363bb3dc JZ |
312 | if (mask_index >= 0) |
313 | { | |
314 | tree mask = gimple_call_arg (stmt, mask_index); | |
315 | rtx mask_rtx = expand_normal (mask); | |
316 | create_input_operand (&ops[opno++], mask_rtx, | |
317 | TYPE_MODE (TREE_TYPE (mask))); | |
318 | } | |
b8806f6f JZZ |
319 | if (len_index >= 0) |
320 | { | |
321 | tree len = gimple_call_arg (stmt, len_index); | |
322 | rtx len_rtx = expand_normal (len); | |
323 | create_convert_operand_from (&ops[opno++], len_rtx, | |
324 | TYPE_MODE (TREE_TYPE (len)), | |
325 | TYPE_UNSIGNED (TREE_TYPE (len))); | |
326 | tree biast = gimple_call_arg (stmt, bias_index); | |
327 | rtx bias = expand_normal (biast); | |
328 | create_input_operand (&ops[opno++], bias, QImode); | |
329 | } | |
b8806f6f JZZ |
330 | return opno; |
331 | } | |
332 | ||
ab23f5d9 | 333 | /* Expand LOAD_LANES call STMT using optab OPTAB. */ |
272c6793 RS |
334 | |
335 | static void | |
4cfe7a6c | 336 | expand_load_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
272c6793 | 337 | { |
99b1c316 | 338 | class expand_operand ops[2]; |
272c6793 RS |
339 | tree type, lhs, rhs; |
340 | rtx target, mem; | |
341 | ||
342 | lhs = gimple_call_lhs (stmt); | |
343 | rhs = gimple_call_arg (stmt, 0); | |
344 | type = TREE_TYPE (lhs); | |
345 | ||
346 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
347 | mem = expand_normal (rhs); | |
348 | ||
349 | gcc_assert (MEM_P (mem)); | |
350 | PUT_MODE (mem, TYPE_MODE (type)); | |
351 | ||
352 | create_output_operand (&ops[0], target, TYPE_MODE (type)); | |
353 | create_fixed_operand (&ops[1], mem); | |
ab23f5d9 | 354 | expand_insn (get_multi_vector_move (type, optab), 2, ops); |
3af3bec2 RS |
355 | if (!rtx_equal_p (target, ops[0].value)) |
356 | emit_move_insn (target, ops[0].value); | |
272c6793 RS |
357 | } |
358 | ||
ab23f5d9 | 359 | /* Expand STORE_LANES call STMT using optab OPTAB. */ |
272c6793 RS |
360 | |
361 | static void | |
4cfe7a6c | 362 | expand_store_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
272c6793 | 363 | { |
99b1c316 | 364 | class expand_operand ops[2]; |
272c6793 RS |
365 | tree type, lhs, rhs; |
366 | rtx target, reg; | |
367 | ||
368 | lhs = gimple_call_lhs (stmt); | |
369 | rhs = gimple_call_arg (stmt, 0); | |
370 | type = TREE_TYPE (rhs); | |
371 | ||
372 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
373 | reg = expand_normal (rhs); | |
374 | ||
375 | gcc_assert (MEM_P (target)); | |
376 | PUT_MODE (target, TYPE_MODE (type)); | |
377 | ||
378 | create_fixed_operand (&ops[0], target); | |
379 | create_input_operand (&ops[1], reg, TYPE_MODE (type)); | |
ab23f5d9 | 380 | expand_insn (get_multi_vector_move (type, optab), 2, ops); |
272c6793 RS |
381 | } |
382 | ||
8170608b | 383 | static void |
4cfe7a6c | 384 | expand_ANNOTATE (internal_fn, gcall *) |
8170608b TB |
385 | { |
386 | gcc_unreachable (); | |
387 | } | |
388 | ||
6c7509bc JJ |
389 | /* This should get expanded in omp_device_lower pass. */ |
390 | ||
391 | static void | |
392 | expand_GOMP_USE_SIMT (internal_fn, gcall *) | |
393 | { | |
394 | gcc_unreachable (); | |
395 | } | |
396 | ||
0c6b03b5 AM |
397 | /* This should get expanded in omp_device_lower pass. */ |
398 | ||
399 | static void | |
400 | expand_GOMP_SIMT_ENTER (internal_fn, gcall *) | |
401 | { | |
402 | gcc_unreachable (); | |
403 | } | |
404 | ||
9d2fe6d4 RS |
405 | /* Allocate per-lane storage and begin non-uniform execution region. */ |
406 | ||
407 | static void | |
408 | expand_GOMP_SIMT_ENTER_ALLOC (internal_fn, gcall *stmt) | |
409 | { | |
410 | rtx target; | |
411 | tree lhs = gimple_call_lhs (stmt); | |
412 | if (lhs) | |
413 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
414 | else | |
415 | target = gen_reg_rtx (Pmode); | |
416 | rtx size = expand_normal (gimple_call_arg (stmt, 0)); | |
417 | rtx align = expand_normal (gimple_call_arg (stmt, 1)); | |
418 | class expand_operand ops[3]; | |
419 | create_output_operand (&ops[0], target, Pmode); | |
420 | create_input_operand (&ops[1], size, Pmode); | |
421 | create_input_operand (&ops[2], align, Pmode); | |
422 | gcc_assert (targetm.have_omp_simt_enter ()); | |
423 | expand_insn (targetm.code_for_omp_simt_enter, 3, ops); | |
424 | if (!rtx_equal_p (target, ops[0].value)) | |
425 | emit_move_insn (target, ops[0].value); | |
426 | } | |
427 | ||
428 | /* Deallocate per-lane storage and leave non-uniform execution region. */ | |
429 | ||
430 | static void | |
431 | expand_GOMP_SIMT_EXIT (internal_fn, gcall *stmt) | |
432 | { | |
433 | gcc_checking_assert (!gimple_call_lhs (stmt)); | |
434 | rtx arg = expand_normal (gimple_call_arg (stmt, 0)); | |
435 | class expand_operand ops[1]; | |
436 | create_input_operand (&ops[0], arg, Pmode); | |
437 | gcc_assert (targetm.have_omp_simt_exit ()); | |
438 | expand_insn (targetm.code_for_omp_simt_exit, 1, ops); | |
439 | } | |
440 | ||
441 | /* Lane index on SIMT targets: thread index in the warp on NVPTX. On targets | |
442 | without SIMT execution this should be expanded in omp_device_lower pass. */ | |
443 | ||
444 | static void | |
445 | expand_GOMP_SIMT_LANE (internal_fn, gcall *stmt) | |
446 | { | |
447 | tree lhs = gimple_call_lhs (stmt); | |
448 | if (!lhs) | |
449 | return; | |
450 | ||
451 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
452 | gcc_assert (targetm.have_omp_simt_lane ()); | |
453 | emit_insn (targetm.gen_omp_simt_lane (target)); | |
454 | } | |
455 | ||
9669b00b AM |
456 | /* This should get expanded in omp_device_lower pass. */ |
457 | ||
458 | static void | |
459 | expand_GOMP_SIMT_VF (internal_fn, gcall *) | |
460 | { | |
461 | gcc_unreachable (); | |
462 | } | |
463 | ||
d6621a2f TB |
464 | /* This should get expanded in omp_device_lower pass. */ |
465 | ||
466 | static void | |
467 | expand_GOMP_TARGET_REV (internal_fn, gcall *) | |
468 | { | |
469 | gcc_unreachable (); | |
470 | } | |
471 | ||
9d2fe6d4 RS |
472 | /* Lane index of the first SIMT lane that supplies a non-zero argument. |
473 | This is a SIMT counterpart to GOMP_SIMD_LAST_LANE, used to represent the | |
474 | lane that executed the last iteration for handling OpenMP lastprivate. */ | |
475 | ||
476 | static void | |
477 | expand_GOMP_SIMT_LAST_LANE (internal_fn, gcall *stmt) | |
478 | { | |
479 | tree lhs = gimple_call_lhs (stmt); | |
480 | if (!lhs) | |
481 | return; | |
482 | ||
483 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
484 | rtx cond = expand_normal (gimple_call_arg (stmt, 0)); | |
485 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
486 | class expand_operand ops[2]; | |
487 | create_output_operand (&ops[0], target, mode); | |
488 | create_input_operand (&ops[1], cond, mode); | |
489 | gcc_assert (targetm.have_omp_simt_last_lane ()); | |
490 | expand_insn (targetm.code_for_omp_simt_last_lane, 2, ops); | |
491 | if (!rtx_equal_p (target, ops[0].value)) | |
492 | emit_move_insn (target, ops[0].value); | |
493 | } | |
494 | ||
495 | /* Non-transparent predicate used in SIMT lowering of OpenMP "ordered". */ | |
496 | ||
497 | static void | |
498 | expand_GOMP_SIMT_ORDERED_PRED (internal_fn, gcall *stmt) | |
499 | { | |
500 | tree lhs = gimple_call_lhs (stmt); | |
501 | if (!lhs) | |
502 | return; | |
503 | ||
504 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
505 | rtx ctr = expand_normal (gimple_call_arg (stmt, 0)); | |
506 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
507 | class expand_operand ops[2]; | |
508 | create_output_operand (&ops[0], target, mode); | |
509 | create_input_operand (&ops[1], ctr, mode); | |
510 | gcc_assert (targetm.have_omp_simt_ordered ()); | |
511 | expand_insn (targetm.code_for_omp_simt_ordered, 2, ops); | |
512 | if (!rtx_equal_p (target, ops[0].value)) | |
513 | emit_move_insn (target, ops[0].value); | |
514 | } | |
515 | ||
516 | /* "Or" boolean reduction across SIMT lanes: return non-zero in all lanes if | |
517 | any lane supplies a non-zero argument. */ | |
518 | ||
519 | static void | |
520 | expand_GOMP_SIMT_VOTE_ANY (internal_fn, gcall *stmt) | |
521 | { | |
522 | tree lhs = gimple_call_lhs (stmt); | |
523 | if (!lhs) | |
524 | return; | |
525 | ||
526 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
527 | rtx cond = expand_normal (gimple_call_arg (stmt, 0)); | |
528 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
529 | class expand_operand ops[2]; | |
530 | create_output_operand (&ops[0], target, mode); | |
531 | create_input_operand (&ops[1], cond, mode); | |
532 | gcc_assert (targetm.have_omp_simt_vote_any ()); | |
533 | expand_insn (targetm.code_for_omp_simt_vote_any, 2, ops); | |
534 | if (!rtx_equal_p (target, ops[0].value)) | |
535 | emit_move_insn (target, ops[0].value); | |
536 | } | |
537 | ||
538 | /* Exchange between SIMT lanes with a "butterfly" pattern: source lane index | |
539 | is destination lane index XOR given offset. */ | |
540 | ||
541 | static void | |
542 | expand_GOMP_SIMT_XCHG_BFLY (internal_fn, gcall *stmt) | |
543 | { | |
544 | tree lhs = gimple_call_lhs (stmt); | |
545 | if (!lhs) | |
546 | return; | |
547 | ||
548 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
549 | rtx src = expand_normal (gimple_call_arg (stmt, 0)); | |
550 | rtx idx = expand_normal (gimple_call_arg (stmt, 1)); | |
551 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
552 | class expand_operand ops[3]; | |
553 | create_output_operand (&ops[0], target, mode); | |
554 | create_input_operand (&ops[1], src, mode); | |
555 | create_input_operand (&ops[2], idx, SImode); | |
556 | gcc_assert (targetm.have_omp_simt_xchg_bfly ()); | |
557 | expand_insn (targetm.code_for_omp_simt_xchg_bfly, 3, ops); | |
558 | if (!rtx_equal_p (target, ops[0].value)) | |
559 | emit_move_insn (target, ops[0].value); | |
560 | } | |
561 | ||
562 | /* Exchange between SIMT lanes according to given source lane index. */ | |
563 | ||
564 | static void | |
565 | expand_GOMP_SIMT_XCHG_IDX (internal_fn, gcall *stmt) | |
566 | { | |
567 | tree lhs = gimple_call_lhs (stmt); | |
568 | if (!lhs) | |
569 | return; | |
570 | ||
571 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
572 | rtx src = expand_normal (gimple_call_arg (stmt, 0)); | |
573 | rtx idx = expand_normal (gimple_call_arg (stmt, 1)); | |
574 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); | |
575 | class expand_operand ops[3]; | |
576 | create_output_operand (&ops[0], target, mode); | |
577 | create_input_operand (&ops[1], src, mode); | |
578 | create_input_operand (&ops[2], idx, SImode); | |
579 | gcc_assert (targetm.have_omp_simt_xchg_idx ()); | |
580 | expand_insn (targetm.code_for_omp_simt_xchg_idx, 3, ops); | |
581 | if (!rtx_equal_p (target, ops[0].value)) | |
582 | emit_move_insn (target, ops[0].value); | |
583 | } | |
584 | ||
74bf76ed JJ |
585 | /* This should get expanded in adjust_simduid_builtins. */ |
586 | ||
587 | static void | |
4cfe7a6c | 588 | expand_GOMP_SIMD_LANE (internal_fn, gcall *) |
74bf76ed JJ |
589 | { |
590 | gcc_unreachable (); | |
591 | } | |
592 | ||
593 | /* This should get expanded in adjust_simduid_builtins. */ | |
594 | ||
595 | static void | |
4cfe7a6c | 596 | expand_GOMP_SIMD_VF (internal_fn, gcall *) |
74bf76ed JJ |
597 | { |
598 | gcc_unreachable (); | |
599 | } | |
600 | ||
601 | /* This should get expanded in adjust_simduid_builtins. */ | |
602 | ||
603 | static void | |
4cfe7a6c | 604 | expand_GOMP_SIMD_LAST_LANE (internal_fn, gcall *) |
74bf76ed JJ |
605 | { |
606 | gcc_unreachable (); | |
607 | } | |
608 | ||
d9a6bd32 JJ |
609 | /* This should get expanded in adjust_simduid_builtins. */ |
610 | ||
611 | static void | |
4cfe7a6c | 612 | expand_GOMP_SIMD_ORDERED_START (internal_fn, gcall *) |
d9a6bd32 JJ |
613 | { |
614 | gcc_unreachable (); | |
615 | } | |
616 | ||
617 | /* This should get expanded in adjust_simduid_builtins. */ | |
618 | ||
619 | static void | |
4cfe7a6c | 620 | expand_GOMP_SIMD_ORDERED_END (internal_fn, gcall *) |
d9a6bd32 JJ |
621 | { |
622 | gcc_unreachable (); | |
623 | } | |
624 | ||
b9a55b13 MP |
625 | /* This should get expanded in the sanopt pass. */ |
626 | ||
627 | static void | |
4cfe7a6c | 628 | expand_UBSAN_NULL (internal_fn, gcall *) |
b9a55b13 MP |
629 | { |
630 | gcc_unreachable (); | |
631 | } | |
632 | ||
0e37a2f3 MP |
633 | /* This should get expanded in the sanopt pass. */ |
634 | ||
635 | static void | |
4cfe7a6c | 636 | expand_UBSAN_BOUNDS (internal_fn, gcall *) |
0e82f089 MP |
637 | { |
638 | gcc_unreachable (); | |
639 | } | |
640 | ||
641 | /* This should get expanded in the sanopt pass. */ | |
642 | ||
643 | static void | |
4cfe7a6c | 644 | expand_UBSAN_VPTR (internal_fn, gcall *) |
0e37a2f3 MP |
645 | { |
646 | gcc_unreachable (); | |
647 | } | |
648 | ||
c62ccb9a YG |
649 | /* This should get expanded in the sanopt pass. */ |
650 | ||
c9b39a49 JJ |
651 | static void |
652 | expand_UBSAN_PTR (internal_fn, gcall *) | |
653 | { | |
654 | gcc_unreachable (); | |
655 | } | |
656 | ||
657 | /* This should get expanded in the sanopt pass. */ | |
658 | ||
c62ccb9a | 659 | static void |
4cfe7a6c | 660 | expand_UBSAN_OBJECT_SIZE (internal_fn, gcall *) |
35228ac7 JJ |
661 | { |
662 | gcc_unreachable (); | |
663 | } | |
664 | ||
665 | /* This should get expanded in the sanopt pass. */ | |
666 | ||
93a73251 MM |
667 | static void |
668 | expand_HWASAN_CHECK (internal_fn, gcall *) | |
669 | { | |
670 | gcc_unreachable (); | |
671 | } | |
672 | ||
673 | /* For hwasan stack tagging: | |
674 | Clear tags on the dynamically allocated space. | |
675 | For use after an object dynamically allocated on the stack goes out of | |
676 | scope. */ | |
677 | static void | |
678 | expand_HWASAN_ALLOCA_UNPOISON (internal_fn, gcall *gc) | |
679 | { | |
680 | gcc_assert (Pmode == ptr_mode); | |
681 | tree restored_position = gimple_call_arg (gc, 0); | |
682 | rtx restored_rtx = expand_expr (restored_position, NULL_RTX, VOIDmode, | |
683 | EXPAND_NORMAL); | |
684 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); | |
685 | rtx off = expand_simple_binop (Pmode, MINUS, restored_rtx, | |
686 | stack_pointer_rtx, NULL_RTX, 0, | |
687 | OPTAB_WIDEN); | |
688 | emit_library_call_value (func, NULL_RTX, LCT_NORMAL, VOIDmode, | |
689 | virtual_stack_dynamic_rtx, Pmode, | |
690 | HWASAN_STACK_BACKGROUND, QImode, | |
691 | off, Pmode); | |
692 | } | |
693 | ||
694 | /* For hwasan stack tagging: | |
695 | Return a tag to be used for a dynamic allocation. */ | |
696 | static void | |
697 | expand_HWASAN_CHOOSE_TAG (internal_fn, gcall *gc) | |
698 | { | |
699 | tree tag = gimple_call_lhs (gc); | |
700 | rtx target = expand_expr (tag, NULL_RTX, VOIDmode, EXPAND_NORMAL); | |
701 | machine_mode mode = GET_MODE (target); | |
702 | gcc_assert (mode == QImode); | |
703 | ||
704 | rtx base_tag = targetm.memtag.extract_tag (hwasan_frame_base (), NULL_RTX); | |
705 | gcc_assert (base_tag); | |
706 | rtx tag_offset = gen_int_mode (hwasan_current_frame_tag (), QImode); | |
707 | rtx chosen_tag = expand_simple_binop (QImode, PLUS, base_tag, tag_offset, | |
708 | target, /* unsignedp = */1, | |
709 | OPTAB_WIDEN); | |
710 | chosen_tag = hwasan_truncate_to_tag_size (chosen_tag, target); | |
711 | ||
712 | /* Really need to put the tag into the `target` RTX. */ | |
713 | if (chosen_tag != target) | |
714 | { | |
715 | rtx temp = chosen_tag; | |
716 | gcc_assert (GET_MODE (chosen_tag) == mode); | |
717 | emit_move_insn (target, temp); | |
718 | } | |
719 | ||
720 | hwasan_increment_frame_tag (); | |
721 | } | |
722 | ||
723 | /* For hwasan stack tagging: | |
724 | Tag a region of space in the shadow stack according to the base pointer of | |
725 | an object on the stack. N.b. the length provided in the internal call is | |
726 | required to be aligned to HWASAN_TAG_GRANULE_SIZE. */ | |
727 | static void | |
728 | expand_HWASAN_MARK (internal_fn, gcall *gc) | |
729 | { | |
730 | gcc_assert (ptr_mode == Pmode); | |
731 | HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gc, 0)); | |
732 | bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON; | |
733 | ||
734 | tree base = gimple_call_arg (gc, 1); | |
735 | gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR); | |
736 | rtx base_rtx = expand_normal (base); | |
737 | ||
738 | rtx tag = is_poison ? HWASAN_STACK_BACKGROUND | |
739 | : targetm.memtag.extract_tag (base_rtx, NULL_RTX); | |
740 | rtx address = targetm.memtag.untagged_pointer (base_rtx, NULL_RTX); | |
741 | ||
742 | tree len = gimple_call_arg (gc, 2); | |
743 | rtx r_len = expand_normal (len); | |
744 | ||
745 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); | |
746 | emit_library_call (func, LCT_NORMAL, VOIDmode, address, Pmode, | |
747 | tag, QImode, r_len, Pmode); | |
748 | } | |
749 | ||
750 | /* For hwasan stack tagging: | |
751 | Store a tag into a pointer. */ | |
752 | static void | |
753 | expand_HWASAN_SET_TAG (internal_fn, gcall *gc) | |
754 | { | |
755 | gcc_assert (ptr_mode == Pmode); | |
756 | tree g_target = gimple_call_lhs (gc); | |
757 | tree g_ptr = gimple_call_arg (gc, 0); | |
758 | tree g_tag = gimple_call_arg (gc, 1); | |
759 | ||
760 | rtx ptr = expand_normal (g_ptr); | |
761 | rtx tag = expand_expr (g_tag, NULL_RTX, QImode, EXPAND_NORMAL); | |
762 | rtx target = expand_normal (g_target); | |
763 | ||
764 | rtx untagged = targetm.memtag.untagged_pointer (ptr, target); | |
765 | rtx tagged_value = targetm.memtag.set_tag (untagged, tag, target); | |
766 | if (tagged_value != target) | |
767 | emit_move_insn (target, tagged_value); | |
768 | } | |
769 | ||
770 | /* This should get expanded in the sanopt pass. */ | |
771 | ||
35228ac7 | 772 | static void |
4cfe7a6c | 773 | expand_ASAN_CHECK (internal_fn, gcall *) |
c62ccb9a YG |
774 | { |
775 | gcc_unreachable (); | |
776 | } | |
777 | ||
6dc4a604 ML |
778 | /* This should get expanded in the sanopt pass. */ |
779 | ||
780 | static void | |
781 | expand_ASAN_MARK (internal_fn, gcall *) | |
782 | { | |
783 | gcc_unreachable (); | |
784 | } | |
785 | ||
c7775327 ML |
786 | /* This should get expanded in the sanopt pass. */ |
787 | ||
788 | static void | |
789 | expand_ASAN_POISON (internal_fn, gcall *) | |
790 | { | |
791 | gcc_unreachable (); | |
792 | } | |
6dc4a604 | 793 | |
f6b9f2ff ML |
794 | /* This should get expanded in the sanopt pass. */ |
795 | ||
796 | static void | |
797 | expand_ASAN_POISON_USE (internal_fn, gcall *) | |
798 | { | |
799 | gcc_unreachable (); | |
800 | } | |
801 | ||
fca4adf2 JJ |
802 | /* This should get expanded in the tsan pass. */ |
803 | ||
804 | static void | |
4cfe7a6c | 805 | expand_TSAN_FUNC_EXIT (internal_fn, gcall *) |
fca4adf2 JJ |
806 | { |
807 | gcc_unreachable (); | |
808 | } | |
809 | ||
81fea426 MP |
810 | /* This should get expanded in the lower pass. */ |
811 | ||
812 | static void | |
813 | expand_FALLTHROUGH (internal_fn, gcall *call) | |
814 | { | |
815 | error_at (gimple_location (call), | |
816 | "invalid use of attribute %<fallthrough%>"); | |
817 | } | |
818 | ||
1304953e JJ |
819 | /* Return minimum precision needed to represent all values |
820 | of ARG in SIGNed integral type. */ | |
821 | ||
822 | static int | |
823 | get_min_precision (tree arg, signop sign) | |
824 | { | |
825 | int prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
826 | int cnt = 0; | |
827 | signop orig_sign = sign; | |
828 | if (TREE_CODE (arg) == INTEGER_CST) | |
829 | { | |
830 | int p; | |
831 | if (TYPE_SIGN (TREE_TYPE (arg)) != sign) | |
832 | { | |
833 | widest_int w = wi::to_widest (arg); | |
834 | w = wi::ext (w, prec, sign); | |
835 | p = wi::min_precision (w, sign); | |
836 | } | |
837 | else | |
8e6cdc90 | 838 | p = wi::min_precision (wi::to_wide (arg), sign); |
1304953e JJ |
839 | return MIN (p, prec); |
840 | } | |
841 | while (CONVERT_EXPR_P (arg) | |
842 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) | |
843 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) <= prec) | |
844 | { | |
845 | arg = TREE_OPERAND (arg, 0); | |
846 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) | |
847 | { | |
848 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) | |
849 | sign = UNSIGNED; | |
850 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) | |
851 | return prec + (orig_sign != sign); | |
852 | prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
853 | } | |
854 | if (++cnt > 30) | |
855 | return prec + (orig_sign != sign); | |
856 | } | |
049ce9d2 JJ |
857 | if (CONVERT_EXPR_P (arg) |
858 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) | |
859 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) > prec) | |
860 | { | |
861 | /* We have e.g. (unsigned short) y_2 where int y_2 = (int) x_1(D); | |
862 | If y_2's min precision is smaller than prec, return that. */ | |
863 | int oprec = get_min_precision (TREE_OPERAND (arg, 0), sign); | |
864 | if (oprec < prec) | |
865 | return oprec + (orig_sign != sign); | |
866 | } | |
1304953e JJ |
867 | if (TREE_CODE (arg) != SSA_NAME) |
868 | return prec + (orig_sign != sign); | |
45f4e2b0 AH |
869 | value_range r; |
870 | while (!get_global_range_query ()->range_of_expr (r, arg) | |
637037f4 AH |
871 | || r.varying_p () |
872 | || r.undefined_p ()) | |
1304953e | 873 | { |
355fe088 | 874 | gimple *g = SSA_NAME_DEF_STMT (arg); |
1304953e JJ |
875 | if (is_gimple_assign (g) |
876 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (g))) | |
877 | { | |
878 | tree t = gimple_assign_rhs1 (g); | |
879 | if (INTEGRAL_TYPE_P (TREE_TYPE (t)) | |
880 | && TYPE_PRECISION (TREE_TYPE (t)) <= prec) | |
881 | { | |
882 | arg = t; | |
883 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) | |
884 | { | |
885 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) | |
886 | sign = UNSIGNED; | |
887 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) | |
888 | return prec + (orig_sign != sign); | |
889 | prec = TYPE_PRECISION (TREE_TYPE (arg)); | |
890 | } | |
891 | if (++cnt > 30) | |
892 | return prec + (orig_sign != sign); | |
893 | continue; | |
894 | } | |
895 | } | |
896 | return prec + (orig_sign != sign); | |
897 | } | |
898 | if (sign == TYPE_SIGN (TREE_TYPE (arg))) | |
899 | { | |
45f4e2b0 AH |
900 | int p1 = wi::min_precision (r.lower_bound (), sign); |
901 | int p2 = wi::min_precision (r.upper_bound (), sign); | |
1304953e JJ |
902 | p1 = MAX (p1, p2); |
903 | prec = MIN (prec, p1); | |
904 | } | |
45f4e2b0 | 905 | else if (sign == UNSIGNED && !wi::neg_p (r.lower_bound (), SIGNED)) |
1304953e | 906 | { |
45f4e2b0 | 907 | int p = wi::min_precision (r.upper_bound (), UNSIGNED); |
1304953e JJ |
908 | prec = MIN (prec, p); |
909 | } | |
910 | return prec + (orig_sign != sign); | |
911 | } | |
912 | ||
a86451b9 JJ |
913 | /* Helper for expand_*_overflow. Set the __imag__ part to true |
914 | (1 except for signed:1 type, in which case store -1). */ | |
915 | ||
916 | static void | |
917 | expand_arith_set_overflow (tree lhs, rtx target) | |
918 | { | |
919 | if (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))) == 1 | |
920 | && !TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs)))) | |
13f44099 | 921 | write_complex_part (target, constm1_rtx, true, false); |
a86451b9 | 922 | else |
13f44099 | 923 | write_complex_part (target, const1_rtx, true, false); |
a86451b9 JJ |
924 | } |
925 | ||
1304953e JJ |
926 | /* Helper for expand_*_overflow. Store RES into the __real__ part |
927 | of TARGET. If RES has larger MODE than __real__ part of TARGET, | |
a86451b9 JJ |
928 | set the __imag__ part to 1 if RES doesn't fit into it. Similarly |
929 | if LHS has smaller precision than its mode. */ | |
1304953e JJ |
930 | |
931 | static void | |
932 | expand_arith_overflow_result_store (tree lhs, rtx target, | |
095a2d76 | 933 | scalar_int_mode mode, rtx res) |
1304953e | 934 | { |
c7ad039d RS |
935 | scalar_int_mode tgtmode |
936 | = as_a <scalar_int_mode> (GET_MODE_INNER (GET_MODE (target))); | |
1304953e JJ |
937 | rtx lres = res; |
938 | if (tgtmode != mode) | |
939 | { | |
940 | rtx_code_label *done_label = gen_label_rtx (); | |
941 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); | |
942 | lres = convert_modes (tgtmode, mode, res, uns); | |
943 | gcc_assert (GET_MODE_PRECISION (tgtmode) < GET_MODE_PRECISION (mode)); | |
92344ed0 | 944 | do_compare_rtx_and_jump (res, convert_modes (mode, tgtmode, lres, uns), |
1476d1bd | 945 | EQ, true, mode, NULL_RTX, NULL, done_label, |
357067f2 | 946 | profile_probability::very_likely ()); |
a86451b9 JJ |
947 | expand_arith_set_overflow (lhs, target); |
948 | emit_label (done_label); | |
949 | } | |
950 | int prec = TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))); | |
951 | int tgtprec = GET_MODE_PRECISION (tgtmode); | |
952 | if (prec < tgtprec) | |
953 | { | |
954 | rtx_code_label *done_label = gen_label_rtx (); | |
955 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); | |
956 | res = lres; | |
957 | if (uns) | |
958 | { | |
959 | rtx mask | |
960 | = immed_wide_int_const (wi::shifted_mask (0, prec, false, tgtprec), | |
961 | tgtmode); | |
962 | lres = expand_simple_binop (tgtmode, AND, res, mask, NULL_RTX, | |
95ef39f4 | 963 | true, OPTAB_LIB_WIDEN); |
a86451b9 JJ |
964 | } |
965 | else | |
966 | { | |
967 | lres = expand_shift (LSHIFT_EXPR, tgtmode, res, tgtprec - prec, | |
968 | NULL_RTX, 1); | |
969 | lres = expand_shift (RSHIFT_EXPR, tgtmode, lres, tgtprec - prec, | |
970 | NULL_RTX, 0); | |
971 | } | |
972 | do_compare_rtx_and_jump (res, lres, | |
973 | EQ, true, tgtmode, NULL_RTX, NULL, done_label, | |
357067f2 | 974 | profile_probability::very_likely ()); |
a86451b9 | 975 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
976 | emit_label (done_label); |
977 | } | |
13f44099 | 978 | write_complex_part (target, lres, false, false); |
1304953e JJ |
979 | } |
980 | ||
5620052d JJ |
981 | /* Helper for expand_*_overflow. Store RES into TARGET. */ |
982 | ||
983 | static void | |
984 | expand_ubsan_result_store (rtx target, rtx res) | |
985 | { | |
986 | if (GET_CODE (target) == SUBREG && SUBREG_PROMOTED_VAR_P (target)) | |
987 | /* If this is a scalar in a register that is stored in a wider mode | |
988 | than the declared mode, compute the result into its declared mode | |
989 | and then convert to the wider mode. Our value is the computed | |
990 | expression. */ | |
991 | convert_move (SUBREG_REG (target), res, SUBREG_PROMOTED_SIGN (target)); | |
992 | else | |
993 | emit_move_insn (target, res); | |
994 | } | |
995 | ||
31e071ae MP |
996 | /* Add sub/add overflow checking to the statement STMT. |
997 | CODE says whether the operation is +, or -. */ | |
998 | ||
4d87bd39 | 999 | void |
1304953e JJ |
1000 | expand_addsub_overflow (location_t loc, tree_code code, tree lhs, |
1001 | tree arg0, tree arg1, bool unsr_p, bool uns0_p, | |
1705cebd | 1002 | bool uns1_p, bool is_ubsan, tree *datap) |
31e071ae | 1003 | { |
1304953e JJ |
1004 | rtx res, target = NULL_RTX; |
1005 | tree fn; | |
1006 | rtx_code_label *done_label = gen_label_rtx (); | |
1007 | rtx_code_label *do_error = gen_label_rtx (); | |
31e071ae | 1008 | do_pending_stack_adjust (); |
1304953e JJ |
1009 | rtx op0 = expand_normal (arg0); |
1010 | rtx op1 = expand_normal (arg1); | |
7a504f33 | 1011 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1304953e JJ |
1012 | int prec = GET_MODE_PRECISION (mode); |
1013 | rtx sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); | |
1014 | bool do_xor = false; | |
1015 | ||
1016 | if (is_ubsan) | |
1017 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); | |
1018 | ||
31e071ae | 1019 | if (lhs) |
1304953e JJ |
1020 | { |
1021 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
1022 | if (!is_ubsan) | |
13f44099 | 1023 | write_complex_part (target, const0_rtx, true, false); |
1304953e JJ |
1024 | } |
1025 | ||
1026 | /* We assume both operands and result have the same precision | |
1027 | here (GET_MODE_BITSIZE (mode)), S stands for signed type | |
1028 | with that precision, U for unsigned type with that precision, | |
1029 | sgn for unsigned most significant bit in that precision. | |
1030 | s1 is signed first operand, u1 is unsigned first operand, | |
1031 | s2 is signed second operand, u2 is unsigned second operand, | |
1032 | sr is signed result, ur is unsigned result and the following | |
1033 | rules say how to compute result (which is always result of | |
1034 | the operands as if both were unsigned, cast to the right | |
1035 | signedness) and how to compute whether operation overflowed. | |
1036 | ||
1037 | s1 + s2 -> sr | |
1038 | res = (S) ((U) s1 + (U) s2) | |
1039 | ovf = s2 < 0 ? res > s1 : res < s1 (or jump on overflow) | |
1040 | s1 - s2 -> sr | |
1041 | res = (S) ((U) s1 - (U) s2) | |
1042 | ovf = s2 < 0 ? res < s1 : res > s2 (or jump on overflow) | |
1043 | u1 + u2 -> ur | |
1044 | res = u1 + u2 | |
1045 | ovf = res < u1 (or jump on carry, but RTL opts will handle it) | |
1046 | u1 - u2 -> ur | |
1047 | res = u1 - u2 | |
1048 | ovf = res > u1 (or jump on carry, but RTL opts will handle it) | |
1049 | s1 + u2 -> sr | |
1050 | res = (S) ((U) s1 + u2) | |
1051 | ovf = ((U) res ^ sgn) < u2 | |
1052 | s1 + u2 -> ur | |
1053 | t1 = (S) (u2 ^ sgn) | |
1054 | t2 = s1 + t1 | |
1055 | res = (U) t2 ^ sgn | |
1056 | ovf = t1 < 0 ? t2 > s1 : t2 < s1 (or jump on overflow) | |
1057 | s1 - u2 -> sr | |
1058 | res = (S) ((U) s1 - u2) | |
1059 | ovf = u2 > ((U) s1 ^ sgn) | |
1060 | s1 - u2 -> ur | |
1061 | res = (U) s1 - u2 | |
1062 | ovf = s1 < 0 || u2 > (U) s1 | |
1063 | u1 - s2 -> sr | |
1064 | res = u1 - (U) s2 | |
1065 | ovf = u1 >= ((U) s2 ^ sgn) | |
1066 | u1 - s2 -> ur | |
1067 | t1 = u1 ^ sgn | |
1068 | t2 = t1 - (U) s2 | |
1069 | res = t2 ^ sgn | |
1070 | ovf = s2 < 0 ? (S) t2 < (S) t1 : (S) t2 > (S) t1 (or jump on overflow) | |
1071 | s1 + s2 -> ur | |
1072 | res = (U) s1 + (U) s2 | |
1073 | ovf = s2 < 0 ? (s1 | (S) res) < 0) : (s1 & (S) res) < 0) | |
1074 | u1 + u2 -> sr | |
1075 | res = (S) (u1 + u2) | |
1076 | ovf = (U) res < u2 || res < 0 | |
1077 | u1 - u2 -> sr | |
1078 | res = (S) (u1 - u2) | |
1079 | ovf = u1 >= u2 ? res < 0 : res >= 0 | |
1080 | s1 - s2 -> ur | |
1081 | res = (U) s1 - (U) s2 | |
1082 | ovf = s2 >= 0 ? ((s1 | (S) res) < 0) : ((s1 & (S) res) < 0) */ | |
1083 | ||
1084 | if (code == PLUS_EXPR && uns0_p && !uns1_p) | |
1085 | { | |
1086 | /* PLUS_EXPR is commutative, if operand signedness differs, | |
1087 | canonicalize to the first operand being signed and second | |
1088 | unsigned to simplify following code. */ | |
6b4db501 MM |
1089 | std::swap (op0, op1); |
1090 | std::swap (arg0, arg1); | |
1091 | uns0_p = false; | |
1092 | uns1_p = true; | |
1304953e JJ |
1093 | } |
1094 | ||
1095 | /* u1 +- u2 -> ur */ | |
1096 | if (uns0_p && uns1_p && unsr_p) | |
1097 | { | |
cde9d596 RH |
1098 | insn_code icode = optab_handler (code == PLUS_EXPR ? uaddv4_optab |
1099 | : usubv4_optab, mode); | |
1100 | if (icode != CODE_FOR_nothing) | |
1101 | { | |
99b1c316 | 1102 | class expand_operand ops[4]; |
cde9d596 RH |
1103 | rtx_insn *last = get_last_insn (); |
1104 | ||
1105 | res = gen_reg_rtx (mode); | |
1106 | create_output_operand (&ops[0], res, mode); | |
1107 | create_input_operand (&ops[1], op0, mode); | |
1108 | create_input_operand (&ops[2], op1, mode); | |
1109 | create_fixed_operand (&ops[3], do_error); | |
1110 | if (maybe_expand_insn (icode, 4, ops)) | |
1111 | { | |
1112 | last = get_last_insn (); | |
1113 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT | |
1114 | && JUMP_P (last) | |
1115 | && any_condjump_p (last) | |
1116 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1117 | add_reg_br_prob_note (last, |
1118 | profile_probability::very_unlikely ()); | |
cde9d596 RH |
1119 | emit_jump (done_label); |
1120 | goto do_error_label; | |
1121 | } | |
1122 | ||
1123 | delete_insns_since (last); | |
1124 | } | |
1125 | ||
1304953e JJ |
1126 | /* Compute the operation. On RTL level, the addition is always |
1127 | unsigned. */ | |
1128 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1129 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1130 | rtx tem = op0; | |
1131 | /* For PLUS_EXPR, the operation is commutative, so we can pick | |
1132 | operand to compare against. For prec <= BITS_PER_WORD, I think | |
1133 | preferring REG operand is better over CONST_INT, because | |
1134 | the CONST_INT might enlarge the instruction or CSE would need | |
1135 | to figure out we'd already loaded it into a register before. | |
1136 | For prec > BITS_PER_WORD, I think CONST_INT might be more beneficial, | |
1137 | as then the multi-word comparison can be perhaps simplified. */ | |
1138 | if (code == PLUS_EXPR | |
1139 | && (prec <= BITS_PER_WORD | |
1140 | ? (CONST_SCALAR_INT_P (op0) && REG_P (op1)) | |
1141 | : CONST_SCALAR_INT_P (op1))) | |
1142 | tem = op1; | |
92344ed0 | 1143 | do_compare_rtx_and_jump (res, tem, code == PLUS_EXPR ? GEU : LEU, |
1476d1bd | 1144 | true, mode, NULL_RTX, NULL, done_label, |
357067f2 | 1145 | profile_probability::very_likely ()); |
1304953e JJ |
1146 | goto do_error_label; |
1147 | } | |
1148 | ||
1149 | /* s1 +- u2 -> sr */ | |
1150 | if (!uns0_p && uns1_p && !unsr_p) | |
1151 | { | |
1152 | /* Compute the operation. On RTL level, the addition is always | |
1153 | unsigned. */ | |
1154 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1155 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1156 | rtx tem = expand_binop (mode, add_optab, | |
1157 | code == PLUS_EXPR ? res : op0, sgn, | |
1158 | NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1476d1bd | 1159 | do_compare_rtx_and_jump (tem, op1, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1160 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1161 | goto do_error_label; |
1162 | } | |
1163 | ||
1164 | /* s1 + u2 -> ur */ | |
1165 | if (code == PLUS_EXPR && !uns0_p && uns1_p && unsr_p) | |
1166 | { | |
1167 | op1 = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, | |
1168 | OPTAB_LIB_WIDEN); | |
1169 | /* As we've changed op1, we have to avoid using the value range | |
1170 | for the original argument. */ | |
1171 | arg1 = error_mark_node; | |
1172 | do_xor = true; | |
1173 | goto do_signed; | |
1174 | } | |
1175 | ||
1176 | /* u1 - s2 -> ur */ | |
1177 | if (code == MINUS_EXPR && uns0_p && !uns1_p && unsr_p) | |
1178 | { | |
1179 | op0 = expand_binop (mode, add_optab, op0, sgn, NULL_RTX, false, | |
1180 | OPTAB_LIB_WIDEN); | |
1181 | /* As we've changed op0, we have to avoid using the value range | |
1182 | for the original argument. */ | |
1183 | arg0 = error_mark_node; | |
1184 | do_xor = true; | |
1185 | goto do_signed; | |
1186 | } | |
1187 | ||
1188 | /* s1 - u2 -> ur */ | |
1189 | if (code == MINUS_EXPR && !uns0_p && uns1_p && unsr_p) | |
1190 | { | |
1191 | /* Compute the operation. On RTL level, the addition is always | |
1192 | unsigned. */ | |
1193 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1194 | OPTAB_LIB_WIDEN); | |
1195 | int pos_neg = get_range_pos_neg (arg0); | |
1196 | if (pos_neg == 2) | |
1197 | /* If ARG0 is known to be always negative, this is always overflow. */ | |
1198 | emit_jump (do_error); | |
1199 | else if (pos_neg == 3) | |
1200 | /* If ARG0 is not known to be always positive, check at runtime. */ | |
92344ed0 | 1201 | do_compare_rtx_and_jump (op0, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1202 | NULL, do_error, profile_probability::very_unlikely ()); |
1476d1bd | 1203 | do_compare_rtx_and_jump (op1, op0, LEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1204 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1205 | goto do_error_label; |
1206 | } | |
1207 | ||
1208 | /* u1 - s2 -> sr */ | |
1209 | if (code == MINUS_EXPR && uns0_p && !uns1_p && !unsr_p) | |
1210 | { | |
1211 | /* Compute the operation. On RTL level, the addition is always | |
1212 | unsigned. */ | |
1213 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1214 | OPTAB_LIB_WIDEN); | |
1215 | rtx tem = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, | |
1216 | OPTAB_LIB_WIDEN); | |
1476d1bd | 1217 | do_compare_rtx_and_jump (op0, tem, LTU, true, mode, NULL_RTX, NULL, |
357067f2 | 1218 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1219 | goto do_error_label; |
1220 | } | |
1221 | ||
1222 | /* u1 + u2 -> sr */ | |
1223 | if (code == PLUS_EXPR && uns0_p && uns1_p && !unsr_p) | |
1224 | { | |
1225 | /* Compute the operation. On RTL level, the addition is always | |
1226 | unsigned. */ | |
1227 | res = expand_binop (mode, add_optab, op0, op1, NULL_RTX, false, | |
1228 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1229 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1230 | NULL, do_error, profile_probability::very_unlikely ()); |
1304953e JJ |
1231 | rtx tem = op1; |
1232 | /* The operation is commutative, so we can pick operand to compare | |
1233 | against. For prec <= BITS_PER_WORD, I think preferring REG operand | |
1234 | is better over CONST_INT, because the CONST_INT might enlarge the | |
1235 | instruction or CSE would need to figure out we'd already loaded it | |
1236 | into a register before. For prec > BITS_PER_WORD, I think CONST_INT | |
1237 | might be more beneficial, as then the multi-word comparison can be | |
1238 | perhaps simplified. */ | |
1239 | if (prec <= BITS_PER_WORD | |
1240 | ? (CONST_SCALAR_INT_P (op1) && REG_P (op0)) | |
1241 | : CONST_SCALAR_INT_P (op0)) | |
1242 | tem = op0; | |
1476d1bd | 1243 | do_compare_rtx_and_jump (res, tem, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1244 | done_label, profile_probability::very_likely ()); |
1304953e JJ |
1245 | goto do_error_label; |
1246 | } | |
1247 | ||
1248 | /* s1 +- s2 -> ur */ | |
1249 | if (!uns0_p && !uns1_p && unsr_p) | |
1250 | { | |
1251 | /* Compute the operation. On RTL level, the addition is always | |
1252 | unsigned. */ | |
1253 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1254 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1255 | int pos_neg = get_range_pos_neg (arg1); | |
1256 | if (code == PLUS_EXPR) | |
1257 | { | |
1258 | int pos_neg0 = get_range_pos_neg (arg0); | |
1259 | if (pos_neg0 != 3 && pos_neg == 3) | |
1260 | { | |
6b4db501 | 1261 | std::swap (op0, op1); |
1304953e JJ |
1262 | pos_neg = pos_neg0; |
1263 | } | |
1264 | } | |
1265 | rtx tem; | |
1266 | if (pos_neg != 3) | |
1267 | { | |
1268 | tem = expand_binop (mode, ((pos_neg == 1) ^ (code == MINUS_EXPR)) | |
1269 | ? and_optab : ior_optab, | |
1270 | op0, res, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1476d1bd | 1271 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL, |
357067f2 | 1272 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1273 | } |
1274 | else | |
1275 | { | |
1276 | rtx_code_label *do_ior_label = gen_label_rtx (); | |
92344ed0 JJ |
1277 | do_compare_rtx_and_jump (op1, const0_rtx, |
1278 | code == MINUS_EXPR ? GE : LT, false, mode, | |
1476d1bd | 1279 | NULL_RTX, NULL, do_ior_label, |
357067f2 | 1280 | profile_probability::even ()); |
1304953e JJ |
1281 | tem = expand_binop (mode, and_optab, op0, res, NULL_RTX, false, |
1282 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1283 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1284 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1285 | emit_jump (do_error); |
1286 | emit_label (do_ior_label); | |
1287 | tem = expand_binop (mode, ior_optab, op0, res, NULL_RTX, false, | |
1288 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1289 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1290 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1291 | } |
1292 | goto do_error_label; | |
1293 | } | |
1294 | ||
1295 | /* u1 - u2 -> sr */ | |
1296 | if (code == MINUS_EXPR && uns0_p && uns1_p && !unsr_p) | |
1297 | { | |
1298 | /* Compute the operation. On RTL level, the addition is always | |
1299 | unsigned. */ | |
1300 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, | |
1301 | OPTAB_LIB_WIDEN); | |
1302 | rtx_code_label *op0_geu_op1 = gen_label_rtx (); | |
1476d1bd | 1303 | do_compare_rtx_and_jump (op0, op1, GEU, true, mode, NULL_RTX, NULL, |
357067f2 | 1304 | op0_geu_op1, profile_probability::even ()); |
92344ed0 | 1305 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
357067f2 | 1306 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1307 | emit_jump (do_error); |
1308 | emit_label (op0_geu_op1); | |
92344ed0 | 1309 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1310 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1311 | goto do_error_label; |
1312 | } | |
31e071ae | 1313 | |
1304953e JJ |
1314 | gcc_assert (!uns0_p && !uns1_p && !unsr_p); |
1315 | ||
1316 | /* s1 +- s2 -> sr */ | |
cde9d596 RH |
1317 | do_signed: |
1318 | { | |
1319 | insn_code icode = optab_handler (code == PLUS_EXPR ? addv4_optab | |
1320 | : subv4_optab, mode); | |
1321 | if (icode != CODE_FOR_nothing) | |
1322 | { | |
99b1c316 | 1323 | class expand_operand ops[4]; |
cde9d596 RH |
1324 | rtx_insn *last = get_last_insn (); |
1325 | ||
1326 | res = gen_reg_rtx (mode); | |
1327 | create_output_operand (&ops[0], res, mode); | |
1328 | create_input_operand (&ops[1], op0, mode); | |
1329 | create_input_operand (&ops[2], op1, mode); | |
1330 | create_fixed_operand (&ops[3], do_error); | |
1331 | if (maybe_expand_insn (icode, 4, ops)) | |
1332 | { | |
1333 | last = get_last_insn (); | |
1334 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT | |
1335 | && JUMP_P (last) | |
1336 | && any_condjump_p (last) | |
1337 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1338 | add_reg_br_prob_note (last, |
1339 | profile_probability::very_unlikely ()); | |
cde9d596 RH |
1340 | emit_jump (done_label); |
1341 | goto do_error_label; | |
1342 | } | |
1343 | ||
1344 | delete_insns_since (last); | |
1345 | } | |
1346 | ||
cde9d596 RH |
1347 | /* Compute the operation. On RTL level, the addition is always |
1348 | unsigned. */ | |
1349 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, | |
1350 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); | |
1351 | ||
47135167 | 1352 | /* If we can prove that one of the arguments (for MINUS_EXPR only |
cde9d596 RH |
1353 | the second operand, as subtraction is not commutative) is always |
1354 | non-negative or always negative, we can do just one comparison | |
47135167 EB |
1355 | and conditional jump. */ |
1356 | int pos_neg = get_range_pos_neg (arg1); | |
1357 | if (code == PLUS_EXPR) | |
cde9d596 | 1358 | { |
47135167 EB |
1359 | int pos_neg0 = get_range_pos_neg (arg0); |
1360 | if (pos_neg0 != 3 && pos_neg == 3) | |
1361 | { | |
1362 | std::swap (op0, op1); | |
1363 | pos_neg = pos_neg0; | |
1364 | } | |
cde9d596 | 1365 | } |
cde9d596 | 1366 | |
47135167 EB |
1367 | /* Addition overflows if and only if the two operands have the same sign, |
1368 | and the result has the opposite sign. Subtraction overflows if and | |
1369 | only if the two operands have opposite sign, and the subtrahend has | |
1370 | the same sign as the result. Here 0 is counted as positive. */ | |
cde9d596 | 1371 | if (pos_neg == 3) |
47135167 EB |
1372 | { |
1373 | /* Compute op0 ^ op1 (operands have opposite sign). */ | |
1374 | rtx op_xor = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, | |
1375 | OPTAB_LIB_WIDEN); | |
cde9d596 | 1376 | |
47135167 EB |
1377 | /* Compute res ^ op1 (result and 2nd operand have opposite sign). */ |
1378 | rtx res_xor = expand_binop (mode, xor_optab, res, op1, NULL_RTX, false, | |
1379 | OPTAB_LIB_WIDEN); | |
97286431 | 1380 | |
47135167 EB |
1381 | rtx tem; |
1382 | if (code == PLUS_EXPR) | |
1383 | { | |
1384 | /* Compute (res ^ op1) & ~(op0 ^ op1). */ | |
1385 | tem = expand_unop (mode, one_cmpl_optab, op_xor, NULL_RTX, false); | |
1386 | tem = expand_binop (mode, and_optab, res_xor, tem, NULL_RTX, false, | |
1387 | OPTAB_LIB_WIDEN); | |
1388 | } | |
1389 | else | |
1390 | { | |
1391 | /* Compute (op0 ^ op1) & ~(res ^ op1). */ | |
1392 | tem = expand_unop (mode, one_cmpl_optab, res_xor, NULL_RTX, false); | |
1393 | tem = expand_binop (mode, and_optab, op_xor, tem, NULL_RTX, false, | |
1394 | OPTAB_LIB_WIDEN); | |
1395 | } | |
1396 | ||
1397 | /* No overflow if the result has bit sign cleared. */ | |
1398 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, | |
357067f2 | 1399 | NULL, done_label, profile_probability::very_likely ()); |
cde9d596 | 1400 | } |
31e071ae | 1401 | |
47135167 EB |
1402 | /* Compare the result of the operation with the first operand. |
1403 | No overflow for addition if second operand is positive and result | |
1404 | is larger or second operand is negative and result is smaller. | |
1405 | Likewise for subtraction with sign of second operand flipped. */ | |
1406 | else | |
1407 | do_compare_rtx_and_jump (res, op0, | |
1408 | (pos_neg == 1) ^ (code == MINUS_EXPR) ? GE : LE, | |
cde9d596 | 1409 | false, mode, NULL_RTX, NULL, done_label, |
357067f2 | 1410 | profile_probability::very_likely ()); |
cde9d596 | 1411 | } |
31e071ae | 1412 | |
1304953e | 1413 | do_error_label: |
1769415d | 1414 | emit_label (do_error); |
1304953e JJ |
1415 | if (is_ubsan) |
1416 | { | |
1417 | /* Expand the ubsan builtin call. */ | |
1418 | push_temp_slots (); | |
1419 | fn = ubsan_build_overflow_builtin (code, loc, TREE_TYPE (arg0), | |
1705cebd | 1420 | arg0, arg1, datap); |
1304953e JJ |
1421 | expand_normal (fn); |
1422 | pop_temp_slots (); | |
1423 | do_pending_stack_adjust (); | |
1424 | } | |
1425 | else if (lhs) | |
a86451b9 | 1426 | expand_arith_set_overflow (lhs, target); |
31e071ae | 1427 | |
1769415d MP |
1428 | /* We're done. */ |
1429 | emit_label (done_label); | |
31e071ae MP |
1430 | |
1431 | if (lhs) | |
1304953e JJ |
1432 | { |
1433 | if (is_ubsan) | |
5620052d | 1434 | expand_ubsan_result_store (target, res); |
1304953e JJ |
1435 | else |
1436 | { | |
1437 | if (do_xor) | |
1438 | res = expand_binop (mode, add_optab, res, sgn, NULL_RTX, false, | |
1439 | OPTAB_LIB_WIDEN); | |
1440 | ||
1441 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
1442 | } | |
1443 | } | |
31e071ae MP |
1444 | } |
1445 | ||
1446 | /* Add negate overflow checking to the statement STMT. */ | |
1447 | ||
1304953e | 1448 | static void |
1705cebd JJ |
1449 | expand_neg_overflow (location_t loc, tree lhs, tree arg1, bool is_ubsan, |
1450 | tree *datap) | |
31e071ae MP |
1451 | { |
1452 | rtx res, op1; | |
1304953e | 1453 | tree fn; |
da664544 DM |
1454 | rtx_code_label *done_label, *do_error; |
1455 | rtx target = NULL_RTX; | |
31e071ae | 1456 | |
31e071ae MP |
1457 | done_label = gen_label_rtx (); |
1458 | do_error = gen_label_rtx (); | |
31e071ae MP |
1459 | |
1460 | do_pending_stack_adjust (); | |
1461 | op1 = expand_normal (arg1); | |
1462 | ||
7a504f33 | 1463 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg1)); |
31e071ae | 1464 | if (lhs) |
1304953e JJ |
1465 | { |
1466 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
1467 | if (!is_ubsan) | |
13f44099 | 1468 | write_complex_part (target, const0_rtx, true, false); |
1304953e | 1469 | } |
31e071ae MP |
1470 | |
1471 | enum insn_code icode = optab_handler (negv3_optab, mode); | |
1472 | if (icode != CODE_FOR_nothing) | |
1473 | { | |
99b1c316 | 1474 | class expand_operand ops[3]; |
da664544 | 1475 | rtx_insn *last = get_last_insn (); |
31e071ae MP |
1476 | |
1477 | res = gen_reg_rtx (mode); | |
1478 | create_output_operand (&ops[0], res, mode); | |
1479 | create_input_operand (&ops[1], op1, mode); | |
1480 | create_fixed_operand (&ops[2], do_error); | |
1481 | if (maybe_expand_insn (icode, 3, ops)) | |
1482 | { | |
1483 | last = get_last_insn (); | |
0a6a6ac9 | 1484 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
31e071ae MP |
1485 | && JUMP_P (last) |
1486 | && any_condjump_p (last) | |
1487 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1488 | add_reg_br_prob_note (last, |
1489 | profile_probability::very_unlikely ()); | |
31e071ae MP |
1490 | emit_jump (done_label); |
1491 | } | |
1492 | else | |
1493 | { | |
1494 | delete_insns_since (last); | |
1495 | icode = CODE_FOR_nothing; | |
1496 | } | |
1497 | } | |
1498 | ||
1499 | if (icode == CODE_FOR_nothing) | |
1500 | { | |
1501 | /* Compute the operation. On RTL level, the addition is always | |
1502 | unsigned. */ | |
1503 | res = expand_unop (mode, neg_optab, op1, NULL_RTX, false); | |
1504 | ||
1505 | /* Compare the operand with the most negative value. */ | |
1506 | rtx minv = expand_normal (TYPE_MIN_VALUE (TREE_TYPE (arg1))); | |
1476d1bd | 1507 | do_compare_rtx_and_jump (op1, minv, NE, true, mode, NULL_RTX, NULL, |
357067f2 | 1508 | done_label, profile_probability::very_likely ()); |
31e071ae MP |
1509 | } |
1510 | ||
1511 | emit_label (do_error); | |
1304953e JJ |
1512 | if (is_ubsan) |
1513 | { | |
1514 | /* Expand the ubsan builtin call. */ | |
1515 | push_temp_slots (); | |
1516 | fn = ubsan_build_overflow_builtin (NEGATE_EXPR, loc, TREE_TYPE (arg1), | |
1705cebd | 1517 | arg1, NULL_TREE, datap); |
1304953e JJ |
1518 | expand_normal (fn); |
1519 | pop_temp_slots (); | |
1520 | do_pending_stack_adjust (); | |
1521 | } | |
1522 | else if (lhs) | |
a86451b9 | 1523 | expand_arith_set_overflow (lhs, target); |
31e071ae MP |
1524 | |
1525 | /* We're done. */ | |
1526 | emit_label (done_label); | |
1527 | ||
1528 | if (lhs) | |
1304953e JJ |
1529 | { |
1530 | if (is_ubsan) | |
5620052d | 1531 | expand_ubsan_result_store (target, res); |
1304953e JJ |
1532 | else |
1533 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
1534 | } | |
31e071ae MP |
1535 | } |
1536 | ||
0b99f253 JJ |
1537 | /* Return true if UNS WIDEN_MULT_EXPR with result mode WMODE and operand |
1538 | mode MODE can be expanded without using a libcall. */ | |
1539 | ||
1540 | static bool | |
1541 | can_widen_mult_without_libcall (scalar_int_mode wmode, scalar_int_mode mode, | |
1542 | rtx op0, rtx op1, bool uns) | |
1543 | { | |
1544 | if (find_widening_optab_handler (umul_widen_optab, wmode, mode) | |
1545 | != CODE_FOR_nothing) | |
1546 | return true; | |
1547 | ||
1548 | if (find_widening_optab_handler (smul_widen_optab, wmode, mode) | |
1549 | != CODE_FOR_nothing) | |
1550 | return true; | |
1551 | ||
1552 | rtx_insn *last = get_last_insn (); | |
1553 | if (CONSTANT_P (op0)) | |
1554 | op0 = convert_modes (wmode, mode, op0, uns); | |
1555 | else | |
1556 | op0 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 1); | |
1557 | if (CONSTANT_P (op1)) | |
1558 | op1 = convert_modes (wmode, mode, op1, uns); | |
1559 | else | |
1560 | op1 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 2); | |
1561 | rtx ret = expand_mult (wmode, op0, op1, NULL_RTX, uns, true); | |
1562 | delete_insns_since (last); | |
1563 | return ret != NULL_RTX; | |
1564 | } | |
1565 | ||
31e071ae MP |
1566 | /* Add mul overflow checking to the statement STMT. */ |
1567 | ||
1304953e JJ |
1568 | static void |
1569 | expand_mul_overflow (location_t loc, tree lhs, tree arg0, tree arg1, | |
1705cebd JJ |
1570 | bool unsr_p, bool uns0_p, bool uns1_p, bool is_ubsan, |
1571 | tree *datap) | |
31e071ae MP |
1572 | { |
1573 | rtx res, op0, op1; | |
1304953e | 1574 | tree fn, type; |
da664544 DM |
1575 | rtx_code_label *done_label, *do_error; |
1576 | rtx target = NULL_RTX; | |
1304953e JJ |
1577 | signop sign; |
1578 | enum insn_code icode; | |
31e071ae | 1579 | |
31e071ae MP |
1580 | done_label = gen_label_rtx (); |
1581 | do_error = gen_label_rtx (); | |
31e071ae MP |
1582 | |
1583 | do_pending_stack_adjust (); | |
1584 | op0 = expand_normal (arg0); | |
1585 | op1 = expand_normal (arg1); | |
1586 | ||
7a504f33 | 1587 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1304953e | 1588 | bool uns = unsr_p; |
31e071ae | 1589 | if (lhs) |
1304953e JJ |
1590 | { |
1591 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
1592 | if (!is_ubsan) | |
13f44099 | 1593 | write_complex_part (target, const0_rtx, true, false); |
1304953e JJ |
1594 | } |
1595 | ||
1596 | if (is_ubsan) | |
1597 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); | |
1598 | ||
1599 | /* We assume both operands and result have the same precision | |
1600 | here (GET_MODE_BITSIZE (mode)), S stands for signed type | |
1601 | with that precision, U for unsigned type with that precision, | |
1602 | sgn for unsigned most significant bit in that precision. | |
1603 | s1 is signed first operand, u1 is unsigned first operand, | |
1604 | s2 is signed second operand, u2 is unsigned second operand, | |
1605 | sr is signed result, ur is unsigned result and the following | |
1606 | rules say how to compute result (which is always result of | |
1607 | the operands as if both were unsigned, cast to the right | |
1608 | signedness) and how to compute whether operation overflowed. | |
1609 | main_ovf (false) stands for jump on signed multiplication | |
1610 | overflow or the main algorithm with uns == false. | |
1611 | main_ovf (true) stands for jump on unsigned multiplication | |
1612 | overflow or the main algorithm with uns == true. | |
1613 | ||
1614 | s1 * s2 -> sr | |
1615 | res = (S) ((U) s1 * (U) s2) | |
1616 | ovf = main_ovf (false) | |
1617 | u1 * u2 -> ur | |
1618 | res = u1 * u2 | |
1619 | ovf = main_ovf (true) | |
1620 | s1 * u2 -> ur | |
1621 | res = (U) s1 * u2 | |
1622 | ovf = (s1 < 0 && u2) || main_ovf (true) | |
1623 | u1 * u2 -> sr | |
1624 | res = (S) (u1 * u2) | |
1625 | ovf = res < 0 || main_ovf (true) | |
1626 | s1 * u2 -> sr | |
1627 | res = (S) ((U) s1 * u2) | |
1628 | ovf = (S) u2 >= 0 ? main_ovf (false) | |
1629 | : (s1 != 0 && (s1 != -1 || u2 != (U) res)) | |
1630 | s1 * s2 -> ur | |
1631 | t1 = (s1 & s2) < 0 ? (-(U) s1) : ((U) s1) | |
1632 | t2 = (s1 & s2) < 0 ? (-(U) s2) : ((U) s2) | |
1633 | res = t1 * t2 | |
1634 | ovf = (s1 ^ s2) < 0 ? (s1 && s2) : main_ovf (true) */ | |
1635 | ||
1636 | if (uns0_p && !uns1_p) | |
1637 | { | |
1638 | /* Multiplication is commutative, if operand signedness differs, | |
1639 | canonicalize to the first operand being signed and second | |
1640 | unsigned to simplify following code. */ | |
6b4db501 MM |
1641 | std::swap (op0, op1); |
1642 | std::swap (arg0, arg1); | |
1643 | uns0_p = false; | |
1644 | uns1_p = true; | |
1304953e JJ |
1645 | } |
1646 | ||
1647 | int pos_neg0 = get_range_pos_neg (arg0); | |
1648 | int pos_neg1 = get_range_pos_neg (arg1); | |
1649 | ||
1650 | /* s1 * u2 -> ur */ | |
1651 | if (!uns0_p && uns1_p && unsr_p) | |
1652 | { | |
1653 | switch (pos_neg0) | |
1654 | { | |
1655 | case 1: | |
1656 | /* If s1 is non-negative, just perform normal u1 * u2 -> ur. */ | |
1657 | goto do_main; | |
1658 | case 2: | |
1659 | /* If s1 is negative, avoid the main code, just multiply and | |
1660 | signal overflow if op1 is not 0. */ | |
1661 | struct separate_ops ops; | |
1662 | ops.code = MULT_EXPR; | |
1663 | ops.type = TREE_TYPE (arg1); | |
1664 | ops.op0 = make_tree (ops.type, op0); | |
1665 | ops.op1 = make_tree (ops.type, op1); | |
1666 | ops.op2 = NULL_TREE; | |
1667 | ops.location = loc; | |
1668 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
92344ed0 | 1669 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1670 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1671 | goto do_error_label; |
1672 | case 3: | |
049ce9d2 JJ |
1673 | if (get_min_precision (arg1, UNSIGNED) |
1674 | + get_min_precision (arg0, SIGNED) <= GET_MODE_PRECISION (mode)) | |
1675 | { | |
1676 | /* If the first operand is sign extended from narrower type, the | |
1677 | second operand is zero extended from narrower type and | |
1678 | the sum of the two precisions is smaller or equal to the | |
1679 | result precision: if the first argument is at runtime | |
1680 | non-negative, maximum result will be 0x7e81 or 0x7f..fe80..01 | |
1681 | and there will be no overflow, if the first argument is | |
1682 | negative and the second argument zero, the result will be | |
1683 | 0 and there will be no overflow, if the first argument is | |
1684 | negative and the second argument positive, the result when | |
1685 | treated as signed will be negative (minimum -0x7f80 or | |
027e3041 | 1686 | -0x7f..f80..0) there will be always overflow. So, do |
049ce9d2 JJ |
1687 | res = (U) (s1 * u2) |
1688 | ovf = (S) res < 0 */ | |
1689 | struct separate_ops ops; | |
1690 | ops.code = MULT_EXPR; | |
1691 | ops.type | |
1692 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1693 | 1); | |
1694 | ops.op0 = make_tree (ops.type, op0); | |
1695 | ops.op1 = make_tree (ops.type, op1); | |
1696 | ops.op2 = NULL_TREE; | |
1697 | ops.location = loc; | |
1698 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1699 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, | |
1700 | mode, NULL_RTX, NULL, done_label, | |
1701 | profile_probability::very_likely ()); | |
1702 | goto do_error_label; | |
1703 | } | |
1304953e JJ |
1704 | rtx_code_label *do_main_label; |
1705 | do_main_label = gen_label_rtx (); | |
92344ed0 | 1706 | do_compare_rtx_and_jump (op0, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1707 | NULL, do_main_label, profile_probability::very_likely ()); |
92344ed0 | 1708 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1709 | NULL, do_main_label, profile_probability::very_likely ()); |
a86451b9 | 1710 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
1711 | emit_label (do_main_label); |
1712 | goto do_main; | |
1713 | default: | |
1714 | gcc_unreachable (); | |
1715 | } | |
1716 | } | |
1717 | ||
1718 | /* u1 * u2 -> sr */ | |
1719 | if (uns0_p && uns1_p && !unsr_p) | |
1720 | { | |
049ce9d2 JJ |
1721 | if ((pos_neg0 | pos_neg1) == 1) |
1722 | { | |
1723 | /* If both arguments are zero extended from narrower types, | |
1724 | the MSB will be clear on both and so we can pretend it is | |
1725 | a normal s1 * s2 -> sr multiplication. */ | |
1726 | uns0_p = false; | |
1727 | uns1_p = false; | |
1728 | } | |
1729 | else | |
1730 | uns = true; | |
1304953e JJ |
1731 | /* Rest of handling of this case after res is computed. */ |
1732 | goto do_main; | |
1733 | } | |
1734 | ||
1735 | /* s1 * u2 -> sr */ | |
1736 | if (!uns0_p && uns1_p && !unsr_p) | |
1737 | { | |
1738 | switch (pos_neg1) | |
1739 | { | |
1740 | case 1: | |
1741 | goto do_main; | |
1742 | case 2: | |
1743 | /* If (S) u2 is negative (i.e. u2 is larger than maximum of S, | |
1744 | avoid the main code, just multiply and signal overflow | |
1745 | unless 0 * u2 or -1 * ((U) Smin). */ | |
1746 | struct separate_ops ops; | |
1747 | ops.code = MULT_EXPR; | |
1748 | ops.type = TREE_TYPE (arg1); | |
1749 | ops.op0 = make_tree (ops.type, op0); | |
1750 | ops.op1 = make_tree (ops.type, op1); | |
1751 | ops.op2 = NULL_TREE; | |
1752 | ops.location = loc; | |
1753 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
92344ed0 | 1754 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 1755 | NULL, done_label, profile_probability::very_likely ()); |
92344ed0 | 1756 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
357067f2 | 1757 | NULL, do_error, profile_probability::very_unlikely ()); |
1304953e JJ |
1758 | int prec; |
1759 | prec = GET_MODE_PRECISION (mode); | |
1760 | rtx sgn; | |
1761 | sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); | |
92344ed0 | 1762 | do_compare_rtx_and_jump (op1, sgn, EQ, true, mode, NULL_RTX, |
357067f2 | 1763 | NULL, done_label, profile_probability::very_likely ()); |
1304953e JJ |
1764 | goto do_error_label; |
1765 | case 3: | |
1766 | /* Rest of handling of this case after res is computed. */ | |
1767 | goto do_main; | |
1768 | default: | |
1769 | gcc_unreachable (); | |
1770 | } | |
1771 | } | |
31e071ae | 1772 | |
1304953e JJ |
1773 | /* s1 * s2 -> ur */ |
1774 | if (!uns0_p && !uns1_p && unsr_p) | |
1775 | { | |
beeac4c2 | 1776 | rtx tem; |
1304953e JJ |
1777 | switch (pos_neg0 | pos_neg1) |
1778 | { | |
1779 | case 1: /* Both operands known to be non-negative. */ | |
1780 | goto do_main; | |
1781 | case 2: /* Both operands known to be negative. */ | |
1782 | op0 = expand_unop (mode, neg_optab, op0, NULL_RTX, false); | |
1783 | op1 = expand_unop (mode, neg_optab, op1, NULL_RTX, false); | |
1784 | /* Avoid looking at arg0/arg1 ranges, as we've changed | |
1785 | the arguments. */ | |
1786 | arg0 = error_mark_node; | |
1787 | arg1 = error_mark_node; | |
1788 | goto do_main; | |
1789 | case 3: | |
1790 | if ((pos_neg0 ^ pos_neg1) == 3) | |
1791 | { | |
1792 | /* If one operand is known to be negative and the other | |
1793 | non-negative, this overflows always, unless the non-negative | |
1794 | one is 0. Just do normal multiply and set overflow | |
1795 | unless one of the operands is 0. */ | |
1796 | struct separate_ops ops; | |
1797 | ops.code = MULT_EXPR; | |
1798 | ops.type | |
1799 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1800 | 1); | |
1801 | ops.op0 = make_tree (ops.type, op0); | |
1802 | ops.op1 = make_tree (ops.type, op1); | |
1803 | ops.op2 = NULL_TREE; | |
1804 | ops.location = loc; | |
1805 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
beeac4c2 JJ |
1806 | do_compare_rtx_and_jump (pos_neg0 == 1 ? op0 : op1, const0_rtx, EQ, |
1807 | true, mode, NULL_RTX, NULL, done_label, | |
357067f2 | 1808 | profile_probability::very_likely ()); |
1304953e JJ |
1809 | goto do_error_label; |
1810 | } | |
049ce9d2 JJ |
1811 | if (get_min_precision (arg0, SIGNED) |
1812 | + get_min_precision (arg1, SIGNED) <= GET_MODE_PRECISION (mode)) | |
1813 | { | |
1814 | /* If both operands are sign extended from narrower types and | |
1815 | the sum of the two precisions is smaller or equal to the | |
1816 | result precision: if both arguments are at runtime | |
1817 | non-negative, maximum result will be 0x3f01 or 0x3f..f0..01 | |
1818 | and there will be no overflow, if both arguments are negative, | |
1819 | maximum result will be 0x40..00 and there will be no overflow | |
1820 | either, if one argument is positive and the other argument | |
1821 | negative, the result when treated as signed will be negative | |
1822 | and there will be always overflow, and if one argument is | |
1823 | zero and the other negative the result will be zero and no | |
1824 | overflow. So, do | |
1825 | res = (U) (s1 * s2) | |
1826 | ovf = (S) res < 0 */ | |
1827 | struct separate_ops ops; | |
1828 | ops.code = MULT_EXPR; | |
1829 | ops.type | |
1830 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), | |
1831 | 1); | |
1832 | ops.op0 = make_tree (ops.type, op0); | |
1833 | ops.op1 = make_tree (ops.type, op1); | |
1834 | ops.op2 = NULL_TREE; | |
1835 | ops.location = loc; | |
1836 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
1837 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, | |
1838 | mode, NULL_RTX, NULL, done_label, | |
1839 | profile_probability::very_likely ()); | |
1840 | goto do_error_label; | |
1841 | } | |
1304953e JJ |
1842 | /* The general case, do all the needed comparisons at runtime. */ |
1843 | rtx_code_label *do_main_label, *after_negate_label; | |
1844 | rtx rop0, rop1; | |
1845 | rop0 = gen_reg_rtx (mode); | |
1846 | rop1 = gen_reg_rtx (mode); | |
1847 | emit_move_insn (rop0, op0); | |
1848 | emit_move_insn (rop1, op1); | |
1849 | op0 = rop0; | |
1850 | op1 = rop1; | |
1851 | do_main_label = gen_label_rtx (); | |
1852 | after_negate_label = gen_label_rtx (); | |
1853 | tem = expand_binop (mode, and_optab, op0, op1, NULL_RTX, false, | |
1854 | OPTAB_LIB_WIDEN); | |
92344ed0 | 1855 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 1856 | NULL, after_negate_label, profile_probability::very_likely ()); |
1304953e JJ |
1857 | /* Both arguments negative here, negate them and continue with |
1858 | normal unsigned overflow checking multiplication. */ | |
1859 | emit_move_insn (op0, expand_unop (mode, neg_optab, op0, | |
1860 | NULL_RTX, false)); | |
1861 | emit_move_insn (op1, expand_unop (mode, neg_optab, op1, | |
1862 | NULL_RTX, false)); | |
1863 | /* Avoid looking at arg0/arg1 ranges, as we might have changed | |
1864 | the arguments. */ | |
1865 | arg0 = error_mark_node; | |
1866 | arg1 = error_mark_node; | |
1867 | emit_jump (do_main_label); | |
1868 | emit_label (after_negate_label); | |
beeac4c2 JJ |
1869 | tem = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1870 | OPTAB_LIB_WIDEN); | |
1871 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, | |
1872 | NULL, do_main_label, | |
1873 | profile_probability::very_likely ()); | |
1304953e JJ |
1874 | /* One argument is negative here, the other positive. This |
1875 | overflows always, unless one of the arguments is 0. But | |
1876 | if e.g. s2 is 0, (U) s1 * 0 doesn't overflow, whatever s1 | |
1877 | is, thus we can keep do_main code oring in overflow as is. */ | |
beeac4c2 JJ |
1878 | if (pos_neg0 != 2) |
1879 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, | |
1880 | NULL, do_main_label, | |
1881 | profile_probability::very_unlikely ()); | |
1882 | if (pos_neg1 != 2) | |
1883 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, | |
1884 | NULL, do_main_label, | |
1885 | profile_probability::very_unlikely ()); | |
a86451b9 | 1886 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
1887 | emit_label (do_main_label); |
1888 | goto do_main; | |
1889 | default: | |
1890 | gcc_unreachable (); | |
1891 | } | |
1892 | } | |
1893 | ||
1894 | do_main: | |
1895 | type = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), uns); | |
1896 | sign = uns ? UNSIGNED : SIGNED; | |
1897 | icode = optab_handler (uns ? umulv4_optab : mulv4_optab, mode); | |
89b1427f JJ |
1898 | if (uns |
1899 | && (integer_pow2p (arg0) || integer_pow2p (arg1)) | |
1900 | && (optimize_insn_for_speed_p () || icode == CODE_FOR_nothing)) | |
1901 | { | |
1902 | /* Optimize unsigned multiplication by power of 2 constant | |
1903 | using 2 shifts, one for result, one to extract the shifted | |
1904 | out bits to see if they are all zero. | |
1905 | Don't do this if optimizing for size and we have umulv4_optab, | |
1906 | in that case assume multiplication will be shorter. | |
1907 | This is heuristics based on the single target that provides | |
1908 | umulv4 right now (i?86/x86_64), if further targets add it, this | |
1909 | might need to be revisited. | |
1910 | Cases where both operands are constant should be folded already | |
1911 | during GIMPLE, and cases where one operand is constant but not | |
1912 | power of 2 are questionable, either the WIDEN_MULT_EXPR case | |
1913 | below can be done without multiplication, just by shifts and adds, | |
1914 | or we'd need to divide the result (and hope it actually doesn't | |
1915 | really divide nor multiply) and compare the result of the division | |
1916 | with the original operand. */ | |
1917 | rtx opn0 = op0; | |
1918 | rtx opn1 = op1; | |
1919 | tree argn0 = arg0; | |
1920 | tree argn1 = arg1; | |
1921 | if (integer_pow2p (arg0)) | |
1922 | { | |
1923 | std::swap (opn0, opn1); | |
1924 | std::swap (argn0, argn1); | |
1925 | } | |
1926 | int cnt = tree_log2 (argn1); | |
1927 | if (cnt >= 0 && cnt < GET_MODE_PRECISION (mode)) | |
1928 | { | |
1929 | rtx upper = const0_rtx; | |
1930 | res = expand_shift (LSHIFT_EXPR, mode, opn0, cnt, NULL_RTX, uns); | |
1931 | if (cnt != 0) | |
1932 | upper = expand_shift (RSHIFT_EXPR, mode, opn0, | |
1933 | GET_MODE_PRECISION (mode) - cnt, | |
1934 | NULL_RTX, uns); | |
1935 | do_compare_rtx_and_jump (upper, const0_rtx, EQ, true, mode, | |
1936 | NULL_RTX, NULL, done_label, | |
1937 | profile_probability::very_likely ()); | |
1938 | goto do_error_label; | |
1939 | } | |
1940 | } | |
31e071ae MP |
1941 | if (icode != CODE_FOR_nothing) |
1942 | { | |
99b1c316 | 1943 | class expand_operand ops[4]; |
da664544 | 1944 | rtx_insn *last = get_last_insn (); |
31e071ae MP |
1945 | |
1946 | res = gen_reg_rtx (mode); | |
1947 | create_output_operand (&ops[0], res, mode); | |
1948 | create_input_operand (&ops[1], op0, mode); | |
1949 | create_input_operand (&ops[2], op1, mode); | |
1950 | create_fixed_operand (&ops[3], do_error); | |
1951 | if (maybe_expand_insn (icode, 4, ops)) | |
1952 | { | |
1953 | last = get_last_insn (); | |
0a6a6ac9 | 1954 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
31e071ae MP |
1955 | && JUMP_P (last) |
1956 | && any_condjump_p (last) | |
1957 | && !find_reg_note (last, REG_BR_PROB, 0)) | |
5fa396ad JH |
1958 | add_reg_br_prob_note (last, |
1959 | profile_probability::very_unlikely ()); | |
31e071ae MP |
1960 | emit_jump (done_label); |
1961 | } | |
1962 | else | |
1963 | { | |
1964 | delete_insns_since (last); | |
1965 | icode = CODE_FOR_nothing; | |
1966 | } | |
1967 | } | |
1968 | ||
1969 | if (icode == CODE_FOR_nothing) | |
1970 | { | |
1971 | struct separate_ops ops; | |
1304953e | 1972 | int prec = GET_MODE_PRECISION (mode); |
095a2d76 | 1973 | scalar_int_mode hmode, wmode; |
1304953e JJ |
1974 | ops.op0 = make_tree (type, op0); |
1975 | ops.op1 = make_tree (type, op1); | |
31e071ae | 1976 | ops.op2 = NULL_TREE; |
1304953e | 1977 | ops.location = loc; |
0b99f253 JJ |
1978 | |
1979 | /* Optimize unsigned overflow check where we don't use the | |
1980 | multiplication result, just whether overflow happened. | |
1981 | If we can do MULT_HIGHPART_EXPR, that followed by | |
1982 | comparison of the result against zero is cheapest. | |
1983 | We'll still compute res, but it should be DCEd later. */ | |
1984 | use_operand_p use; | |
1985 | gimple *use_stmt; | |
1986 | if (!is_ubsan | |
1987 | && lhs | |
1988 | && uns | |
1989 | && !(uns0_p && uns1_p && !unsr_p) | |
1990 | && can_mult_highpart_p (mode, uns) == 1 | |
1991 | && single_imm_use (lhs, &use, &use_stmt) | |
1992 | && is_gimple_assign (use_stmt) | |
1993 | && gimple_assign_rhs_code (use_stmt) == IMAGPART_EXPR) | |
1994 | goto highpart; | |
1995 | ||
490d0f6c | 1996 | if (GET_MODE_2XWIDER_MODE (mode).exists (&wmode) |
0b99f253 JJ |
1997 | && targetm.scalar_mode_supported_p (wmode) |
1998 | && can_widen_mult_without_libcall (wmode, mode, op0, op1, uns)) | |
31e071ae | 1999 | { |
0b99f253 | 2000 | twoxwider: |
31e071ae MP |
2001 | ops.code = WIDEN_MULT_EXPR; |
2002 | ops.type | |
1304953e | 2003 | = build_nonstandard_integer_type (GET_MODE_PRECISION (wmode), uns); |
31e071ae MP |
2004 | |
2005 | res = expand_expr_real_2 (&ops, NULL_RTX, wmode, EXPAND_NORMAL); | |
1304953e JJ |
2006 | rtx hipart = expand_shift (RSHIFT_EXPR, wmode, res, prec, |
2007 | NULL_RTX, uns); | |
4ed543bc KC |
2008 | hipart = convert_modes (mode, wmode, hipart, uns); |
2009 | res = convert_modes (mode, wmode, res, uns); | |
1304953e JJ |
2010 | if (uns) |
2011 | /* For the unsigned multiplication, there was overflow if | |
2012 | HIPART is non-zero. */ | |
92344ed0 | 2013 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
1476d1bd | 2014 | NULL_RTX, NULL, done_label, |
357067f2 | 2015 | profile_probability::very_likely ()); |
1304953e JJ |
2016 | else |
2017 | { | |
2b4bbc17 RS |
2018 | /* RES is used more than once, place it in a pseudo. */ |
2019 | res = force_reg (mode, res); | |
2020 | ||
1304953e JJ |
2021 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
2022 | NULL_RTX, 0); | |
2023 | /* RES is low half of the double width result, HIPART | |
2024 | the high half. There was overflow if | |
2025 | HIPART is different from RES < 0 ? -1 : 0. */ | |
92344ed0 | 2026 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
1476d1bd | 2027 | NULL_RTX, NULL, done_label, |
357067f2 | 2028 | profile_probability::very_likely ()); |
1304953e | 2029 | } |
31e071ae | 2030 | } |
0b99f253 JJ |
2031 | else if (can_mult_highpart_p (mode, uns) == 1) |
2032 | { | |
2033 | highpart: | |
2034 | ops.code = MULT_HIGHPART_EXPR; | |
2035 | ops.type = type; | |
2036 | ||
2037 | rtx hipart = expand_expr_real_2 (&ops, NULL_RTX, mode, | |
2038 | EXPAND_NORMAL); | |
2039 | ops.code = MULT_EXPR; | |
2040 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2041 | if (uns) | |
2042 | /* For the unsigned multiplication, there was overflow if | |
2043 | HIPART is non-zero. */ | |
2044 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, | |
2045 | NULL_RTX, NULL, done_label, | |
2046 | profile_probability::very_likely ()); | |
2047 | else | |
2048 | { | |
2049 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, | |
2050 | NULL_RTX, 0); | |
2051 | /* RES is low half of the double width result, HIPART | |
2052 | the high half. There was overflow if | |
2053 | HIPART is different from RES < 0 ? -1 : 0. */ | |
2054 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, | |
2055 | NULL_RTX, NULL, done_label, | |
2056 | profile_probability::very_likely ()); | |
2057 | } | |
2058 | ||
2059 | } | |
fffbab82 RS |
2060 | else if (int_mode_for_size (prec / 2, 1).exists (&hmode) |
2061 | && 2 * GET_MODE_PRECISION (hmode) == prec) | |
d5fa9cc9 | 2062 | { |
da664544 DM |
2063 | rtx_code_label *large_op0 = gen_label_rtx (); |
2064 | rtx_code_label *small_op0_large_op1 = gen_label_rtx (); | |
2065 | rtx_code_label *one_small_one_large = gen_label_rtx (); | |
2066 | rtx_code_label *both_ops_large = gen_label_rtx (); | |
1304953e JJ |
2067 | rtx_code_label *after_hipart_neg = uns ? NULL : gen_label_rtx (); |
2068 | rtx_code_label *after_lopart_neg = uns ? NULL : gen_label_rtx (); | |
da664544 | 2069 | rtx_code_label *do_overflow = gen_label_rtx (); |
1304953e | 2070 | rtx_code_label *hipart_different = uns ? NULL : gen_label_rtx (); |
d5fa9cc9 | 2071 | |
807e902e | 2072 | unsigned int hprec = GET_MODE_PRECISION (hmode); |
d5fa9cc9 | 2073 | rtx hipart0 = expand_shift (RSHIFT_EXPR, mode, op0, hprec, |
1304953e | 2074 | NULL_RTX, uns); |
4ed543bc KC |
2075 | hipart0 = convert_modes (hmode, mode, hipart0, uns); |
2076 | rtx lopart0 = convert_modes (hmode, mode, op0, uns); | |
1304953e JJ |
2077 | rtx signbit0 = const0_rtx; |
2078 | if (!uns) | |
2079 | signbit0 = expand_shift (RSHIFT_EXPR, hmode, lopart0, hprec - 1, | |
2080 | NULL_RTX, 0); | |
d5fa9cc9 | 2081 | rtx hipart1 = expand_shift (RSHIFT_EXPR, mode, op1, hprec, |
1304953e | 2082 | NULL_RTX, uns); |
4ed543bc KC |
2083 | hipart1 = convert_modes (hmode, mode, hipart1, uns); |
2084 | rtx lopart1 = convert_modes (hmode, mode, op1, uns); | |
1304953e JJ |
2085 | rtx signbit1 = const0_rtx; |
2086 | if (!uns) | |
2087 | signbit1 = expand_shift (RSHIFT_EXPR, hmode, lopart1, hprec - 1, | |
2088 | NULL_RTX, 0); | |
d5fa9cc9 JJ |
2089 | |
2090 | res = gen_reg_rtx (mode); | |
2091 | ||
2092 | /* True if op0 resp. op1 are known to be in the range of | |
2093 | halfstype. */ | |
2094 | bool op0_small_p = false; | |
2095 | bool op1_small_p = false; | |
2096 | /* True if op0 resp. op1 are known to have all zeros or all ones | |
2097 | in the upper half of bits, but are not known to be | |
2098 | op{0,1}_small_p. */ | |
2099 | bool op0_medium_p = false; | |
2100 | bool op1_medium_p = false; | |
2101 | /* -1 if op{0,1} is known to be negative, 0 if it is known to be | |
2102 | nonnegative, 1 if unknown. */ | |
2103 | int op0_sign = 1; | |
2104 | int op1_sign = 1; | |
2105 | ||
1304953e JJ |
2106 | if (pos_neg0 == 1) |
2107 | op0_sign = 0; | |
2108 | else if (pos_neg0 == 2) | |
2109 | op0_sign = -1; | |
2110 | if (pos_neg1 == 1) | |
2111 | op1_sign = 0; | |
2112 | else if (pos_neg1 == 2) | |
2113 | op1_sign = -1; | |
2114 | ||
2115 | unsigned int mprec0 = prec; | |
2116 | if (arg0 != error_mark_node) | |
2117 | mprec0 = get_min_precision (arg0, sign); | |
2118 | if (mprec0 <= hprec) | |
2119 | op0_small_p = true; | |
2120 | else if (!uns && mprec0 <= hprec + 1) | |
2121 | op0_medium_p = true; | |
2122 | unsigned int mprec1 = prec; | |
2123 | if (arg1 != error_mark_node) | |
2124 | mprec1 = get_min_precision (arg1, sign); | |
2125 | if (mprec1 <= hprec) | |
2126 | op1_small_p = true; | |
2127 | else if (!uns && mprec1 <= hprec + 1) | |
2128 | op1_medium_p = true; | |
d5fa9cc9 JJ |
2129 | |
2130 | int smaller_sign = 1; | |
2131 | int larger_sign = 1; | |
2132 | if (op0_small_p) | |
2133 | { | |
2134 | smaller_sign = op0_sign; | |
2135 | larger_sign = op1_sign; | |
2136 | } | |
2137 | else if (op1_small_p) | |
2138 | { | |
2139 | smaller_sign = op1_sign; | |
2140 | larger_sign = op0_sign; | |
2141 | } | |
2142 | else if (op0_sign == op1_sign) | |
2143 | { | |
2144 | smaller_sign = op0_sign; | |
2145 | larger_sign = op0_sign; | |
2146 | } | |
2147 | ||
2148 | if (!op0_small_p) | |
92344ed0 | 2149 | do_compare_rtx_and_jump (signbit0, hipart0, NE, true, hmode, |
1476d1bd | 2150 | NULL_RTX, NULL, large_op0, |
357067f2 | 2151 | profile_probability::unlikely ()); |
d5fa9cc9 JJ |
2152 | |
2153 | if (!op1_small_p) | |
92344ed0 | 2154 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
1476d1bd | 2155 | NULL_RTX, NULL, small_op0_large_op1, |
357067f2 | 2156 | profile_probability::unlikely ()); |
d5fa9cc9 | 2157 | |
1304953e JJ |
2158 | /* If both op0 and op1 are sign (!uns) or zero (uns) extended from |
2159 | hmode to mode, the multiplication will never overflow. We can | |
2160 | do just one hmode x hmode => mode widening multiplication. */ | |
1304953e | 2161 | tree halfstype = build_nonstandard_integer_type (hprec, uns); |
0764a0d2 JJ |
2162 | ops.op0 = make_tree (halfstype, lopart0); |
2163 | ops.op1 = make_tree (halfstype, lopart1); | |
d5fa9cc9 | 2164 | ops.code = WIDEN_MULT_EXPR; |
1304953e | 2165 | ops.type = type; |
d5fa9cc9 JJ |
2166 | rtx thisres |
2167 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2168 | emit_move_insn (res, thisres); | |
2169 | emit_jump (done_label); | |
2170 | ||
2171 | emit_label (small_op0_large_op1); | |
2172 | ||
1304953e JJ |
2173 | /* If op0 is sign (!uns) or zero (uns) extended from hmode to mode, |
2174 | but op1 is not, just swap the arguments and handle it as op1 | |
2175 | sign/zero extended, op0 not. */ | |
d5fa9cc9 JJ |
2176 | rtx larger = gen_reg_rtx (mode); |
2177 | rtx hipart = gen_reg_rtx (hmode); | |
2178 | rtx lopart = gen_reg_rtx (hmode); | |
2179 | emit_move_insn (larger, op1); | |
2180 | emit_move_insn (hipart, hipart1); | |
2181 | emit_move_insn (lopart, lopart0); | |
2182 | emit_jump (one_small_one_large); | |
2183 | ||
2184 | emit_label (large_op0); | |
2185 | ||
2186 | if (!op1_small_p) | |
92344ed0 | 2187 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
1476d1bd | 2188 | NULL_RTX, NULL, both_ops_large, |
357067f2 | 2189 | profile_probability::unlikely ()); |
d5fa9cc9 | 2190 | |
1304953e JJ |
2191 | /* If op1 is sign (!uns) or zero (uns) extended from hmode to mode, |
2192 | but op0 is not, prepare larger, hipart and lopart pseudos and | |
2193 | handle it together with small_op0_large_op1. */ | |
d5fa9cc9 JJ |
2194 | emit_move_insn (larger, op0); |
2195 | emit_move_insn (hipart, hipart0); | |
2196 | emit_move_insn (lopart, lopart1); | |
2197 | ||
2198 | emit_label (one_small_one_large); | |
2199 | ||
2200 | /* lopart is the low part of the operand that is sign extended | |
026c3cfd | 2201 | to mode, larger is the other operand, hipart is the |
d5fa9cc9 JJ |
2202 | high part of larger and lopart0 and lopart1 are the low parts |
2203 | of both operands. | |
2204 | We perform lopart0 * lopart1 and lopart * hipart widening | |
2205 | multiplications. */ | |
2206 | tree halfutype = build_nonstandard_integer_type (hprec, 1); | |
2207 | ops.op0 = make_tree (halfutype, lopart0); | |
2208 | ops.op1 = make_tree (halfutype, lopart1); | |
2209 | rtx lo0xlo1 | |
2210 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2211 | ||
2212 | ops.op0 = make_tree (halfutype, lopart); | |
2213 | ops.op1 = make_tree (halfutype, hipart); | |
2214 | rtx loxhi = gen_reg_rtx (mode); | |
2215 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2216 | emit_move_insn (loxhi, tem); | |
2217 | ||
1304953e JJ |
2218 | if (!uns) |
2219 | { | |
2220 | /* if (hipart < 0) loxhi -= lopart << (bitsize / 2); */ | |
2221 | if (larger_sign == 0) | |
2222 | emit_jump (after_hipart_neg); | |
2223 | else if (larger_sign != -1) | |
92344ed0 | 2224 | do_compare_rtx_and_jump (hipart, const0_rtx, GE, false, hmode, |
1476d1bd | 2225 | NULL_RTX, NULL, after_hipart_neg, |
357067f2 | 2226 | profile_probability::even ()); |
1304953e JJ |
2227 | |
2228 | tem = convert_modes (mode, hmode, lopart, 1); | |
2229 | tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1); | |
2230 | tem = expand_simple_binop (mode, MINUS, loxhi, tem, NULL_RTX, | |
644f0132 | 2231 | 1, OPTAB_WIDEN); |
1304953e JJ |
2232 | emit_move_insn (loxhi, tem); |
2233 | ||
2234 | emit_label (after_hipart_neg); | |
2235 | ||
2236 | /* if (lopart < 0) loxhi -= larger; */ | |
2237 | if (smaller_sign == 0) | |
2238 | emit_jump (after_lopart_neg); | |
2239 | else if (smaller_sign != -1) | |
92344ed0 | 2240 | do_compare_rtx_and_jump (lopart, const0_rtx, GE, false, hmode, |
1476d1bd | 2241 | NULL_RTX, NULL, after_lopart_neg, |
357067f2 | 2242 | profile_probability::even ()); |
1304953e JJ |
2243 | |
2244 | tem = expand_simple_binop (mode, MINUS, loxhi, larger, NULL_RTX, | |
644f0132 | 2245 | 1, OPTAB_WIDEN); |
1304953e JJ |
2246 | emit_move_insn (loxhi, tem); |
2247 | ||
2248 | emit_label (after_lopart_neg); | |
2249 | } | |
d5fa9cc9 JJ |
2250 | |
2251 | /* loxhi += (uns) lo0xlo1 >> (bitsize / 2); */ | |
2252 | tem = expand_shift (RSHIFT_EXPR, mode, lo0xlo1, hprec, NULL_RTX, 1); | |
2253 | tem = expand_simple_binop (mode, PLUS, loxhi, tem, NULL_RTX, | |
644f0132 | 2254 | 1, OPTAB_WIDEN); |
d5fa9cc9 JJ |
2255 | emit_move_insn (loxhi, tem); |
2256 | ||
2257 | /* if (loxhi >> (bitsize / 2) | |
1304953e JJ |
2258 | == (hmode) loxhi >> (bitsize / 2 - 1)) (if !uns) |
2259 | if (loxhi >> (bitsize / 2) == 0 (if uns). */ | |
d5fa9cc9 JJ |
2260 | rtx hipartloxhi = expand_shift (RSHIFT_EXPR, mode, loxhi, hprec, |
2261 | NULL_RTX, 0); | |
4ed543bc | 2262 | hipartloxhi = convert_modes (hmode, mode, hipartloxhi, 0); |
1304953e JJ |
2263 | rtx signbitloxhi = const0_rtx; |
2264 | if (!uns) | |
2265 | signbitloxhi = expand_shift (RSHIFT_EXPR, hmode, | |
4ed543bc KC |
2266 | convert_modes (hmode, mode, |
2267 | loxhi, 0), | |
1304953e | 2268 | hprec - 1, NULL_RTX, 0); |
d5fa9cc9 | 2269 | |
92344ed0 | 2270 | do_compare_rtx_and_jump (signbitloxhi, hipartloxhi, NE, true, hmode, |
1476d1bd | 2271 | NULL_RTX, NULL, do_overflow, |
357067f2 | 2272 | profile_probability::very_unlikely ()); |
d5fa9cc9 JJ |
2273 | |
2274 | /* res = (loxhi << (bitsize / 2)) | (hmode) lo0xlo1; */ | |
2275 | rtx loxhishifted = expand_shift (LSHIFT_EXPR, mode, loxhi, hprec, | |
2276 | NULL_RTX, 1); | |
4ed543bc KC |
2277 | tem = convert_modes (mode, hmode, |
2278 | convert_modes (hmode, mode, lo0xlo1, 1), 1); | |
d5fa9cc9 JJ |
2279 | |
2280 | tem = expand_simple_binop (mode, IOR, loxhishifted, tem, res, | |
644f0132 | 2281 | 1, OPTAB_WIDEN); |
d5fa9cc9 JJ |
2282 | if (tem != res) |
2283 | emit_move_insn (res, tem); | |
2284 | emit_jump (done_label); | |
2285 | ||
2286 | emit_label (both_ops_large); | |
2287 | ||
1304953e JJ |
2288 | /* If both operands are large (not sign (!uns) or zero (uns) |
2289 | extended from hmode), then perform the full multiplication | |
2290 | which will be the result of the operation. | |
2291 | The only cases which don't overflow are for signed multiplication | |
2292 | some cases where both hipart0 and highpart1 are 0 or -1. | |
2293 | For unsigned multiplication when high parts are both non-zero | |
2294 | this overflows always. */ | |
d5fa9cc9 | 2295 | ops.code = MULT_EXPR; |
1304953e JJ |
2296 | ops.op0 = make_tree (type, op0); |
2297 | ops.op1 = make_tree (type, op1); | |
d5fa9cc9 JJ |
2298 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2299 | emit_move_insn (res, tem); | |
2300 | ||
1304953e | 2301 | if (!uns) |
d5fa9cc9 | 2302 | { |
1304953e JJ |
2303 | if (!op0_medium_p) |
2304 | { | |
2305 | tem = expand_simple_binop (hmode, PLUS, hipart0, const1_rtx, | |
644f0132 | 2306 | NULL_RTX, 1, OPTAB_WIDEN); |
92344ed0 | 2307 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
1476d1bd | 2308 | NULL_RTX, NULL, do_error, |
357067f2 | 2309 | profile_probability::very_unlikely ()); |
1304953e | 2310 | } |
d5fa9cc9 | 2311 | |
1304953e JJ |
2312 | if (!op1_medium_p) |
2313 | { | |
2314 | tem = expand_simple_binop (hmode, PLUS, hipart1, const1_rtx, | |
644f0132 | 2315 | NULL_RTX, 1, OPTAB_WIDEN); |
92344ed0 | 2316 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
1476d1bd | 2317 | NULL_RTX, NULL, do_error, |
357067f2 | 2318 | profile_probability::very_unlikely ()); |
1304953e | 2319 | } |
d5fa9cc9 | 2320 | |
1304953e | 2321 | /* At this point hipart{0,1} are both in [-1, 0]. If they are |
e7176f75 JJ |
2322 | the same, overflow happened if res is non-positive, if they |
2323 | are different, overflow happened if res is positive. */ | |
1304953e JJ |
2324 | if (op0_sign != 1 && op1_sign != 1 && op0_sign != op1_sign) |
2325 | emit_jump (hipart_different); | |
2326 | else if (op0_sign == 1 || op1_sign == 1) | |
92344ed0 | 2327 | do_compare_rtx_and_jump (hipart0, hipart1, NE, true, hmode, |
1476d1bd | 2328 | NULL_RTX, NULL, hipart_different, |
357067f2 | 2329 | profile_probability::even ()); |
d5fa9cc9 | 2330 | |
e7176f75 | 2331 | do_compare_rtx_and_jump (res, const0_rtx, LE, false, mode, |
1476d1bd | 2332 | NULL_RTX, NULL, do_error, |
357067f2 | 2333 | profile_probability::very_unlikely ()); |
1304953e | 2334 | emit_jump (done_label); |
d5fa9cc9 | 2335 | |
1304953e JJ |
2336 | emit_label (hipart_different); |
2337 | ||
92344ed0 | 2338 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, |
1476d1bd | 2339 | NULL_RTX, NULL, do_error, |
357067f2 | 2340 | profile_probability::very_unlikely ()); |
1304953e JJ |
2341 | emit_jump (done_label); |
2342 | } | |
d5fa9cc9 JJ |
2343 | |
2344 | emit_label (do_overflow); | |
2345 | ||
2346 | /* Overflow, do full multiplication and fallthru into do_error. */ | |
1304953e JJ |
2347 | ops.op0 = make_tree (type, op0); |
2348 | ops.op1 = make_tree (type, op1); | |
d5fa9cc9 JJ |
2349 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2350 | emit_move_insn (res, tem); | |
2351 | } | |
0b99f253 JJ |
2352 | else if (GET_MODE_2XWIDER_MODE (mode).exists (&wmode) |
2353 | && targetm.scalar_mode_supported_p (wmode)) | |
2354 | /* Even emitting a libcall is better than not detecting overflow | |
2355 | at all. */ | |
2356 | goto twoxwider; | |
31e071ae MP |
2357 | else |
2358 | { | |
1304953e | 2359 | gcc_assert (!is_ubsan); |
31e071ae | 2360 | ops.code = MULT_EXPR; |
1304953e | 2361 | ops.type = type; |
31e071ae MP |
2362 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2363 | emit_jump (done_label); | |
2364 | } | |
2365 | } | |
2366 | ||
1304953e | 2367 | do_error_label: |
31e071ae | 2368 | emit_label (do_error); |
1304953e JJ |
2369 | if (is_ubsan) |
2370 | { | |
2371 | /* Expand the ubsan builtin call. */ | |
2372 | push_temp_slots (); | |
2373 | fn = ubsan_build_overflow_builtin (MULT_EXPR, loc, TREE_TYPE (arg0), | |
1705cebd | 2374 | arg0, arg1, datap); |
1304953e JJ |
2375 | expand_normal (fn); |
2376 | pop_temp_slots (); | |
2377 | do_pending_stack_adjust (); | |
2378 | } | |
2379 | else if (lhs) | |
a86451b9 | 2380 | expand_arith_set_overflow (lhs, target); |
31e071ae MP |
2381 | |
2382 | /* We're done. */ | |
2383 | emit_label (done_label); | |
2384 | ||
1304953e JJ |
2385 | /* u1 * u2 -> sr */ |
2386 | if (uns0_p && uns1_p && !unsr_p) | |
2387 | { | |
2388 | rtx_code_label *all_done_label = gen_label_rtx (); | |
92344ed0 | 2389 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 2390 | NULL, all_done_label, profile_probability::very_likely ()); |
a86451b9 | 2391 | expand_arith_set_overflow (lhs, target); |
1304953e JJ |
2392 | emit_label (all_done_label); |
2393 | } | |
2394 | ||
2395 | /* s1 * u2 -> sr */ | |
2396 | if (!uns0_p && uns1_p && !unsr_p && pos_neg1 == 3) | |
2397 | { | |
2398 | rtx_code_label *all_done_label = gen_label_rtx (); | |
2399 | rtx_code_label *set_noovf = gen_label_rtx (); | |
92344ed0 | 2400 | do_compare_rtx_and_jump (op1, const0_rtx, GE, false, mode, NULL_RTX, |
357067f2 | 2401 | NULL, all_done_label, profile_probability::very_likely ()); |
a86451b9 | 2402 | expand_arith_set_overflow (lhs, target); |
92344ed0 | 2403 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
357067f2 | 2404 | NULL, set_noovf, profile_probability::very_likely ()); |
92344ed0 | 2405 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
357067f2 | 2406 | NULL, all_done_label, profile_probability::very_unlikely ()); |
1476d1bd | 2407 | do_compare_rtx_and_jump (op1, res, NE, true, mode, NULL_RTX, NULL, |
357067f2 | 2408 | all_done_label, profile_probability::very_unlikely ()); |
1304953e | 2409 | emit_label (set_noovf); |
13f44099 | 2410 | write_complex_part (target, const0_rtx, true, false); |
1304953e JJ |
2411 | emit_label (all_done_label); |
2412 | } | |
2413 | ||
31e071ae | 2414 | if (lhs) |
1304953e JJ |
2415 | { |
2416 | if (is_ubsan) | |
5620052d | 2417 | expand_ubsan_result_store (target, res); |
1304953e JJ |
2418 | else |
2419 | expand_arith_overflow_result_store (lhs, target, mode, res); | |
2420 | } | |
31e071ae MP |
2421 | } |
2422 | ||
1705cebd JJ |
2423 | /* Expand UBSAN_CHECK_* internal function if it has vector operands. */ |
2424 | ||
2425 | static void | |
2426 | expand_vector_ubsan_overflow (location_t loc, enum tree_code code, tree lhs, | |
2427 | tree arg0, tree arg1) | |
2428 | { | |
07626e49 | 2429 | poly_uint64 cnt = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)); |
1705cebd JJ |
2430 | rtx_code_label *loop_lab = NULL; |
2431 | rtx cntvar = NULL_RTX; | |
2432 | tree cntv = NULL_TREE; | |
2433 | tree eltype = TREE_TYPE (TREE_TYPE (arg0)); | |
2434 | tree sz = TYPE_SIZE (eltype); | |
2435 | tree data = NULL_TREE; | |
2436 | tree resv = NULL_TREE; | |
2437 | rtx lhsr = NULL_RTX; | |
2438 | rtx resvr = NULL_RTX; | |
07626e49 RS |
2439 | unsigned HOST_WIDE_INT const_cnt = 0; |
2440 | bool use_loop_p = (!cnt.is_constant (&const_cnt) || const_cnt > 4); | |
1705cebd JJ |
2441 | |
2442 | if (lhs) | |
2443 | { | |
2444 | optab op; | |
2445 | lhsr = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
6a5cdb0e | 2446 | if (!VECTOR_MODE_P (GET_MODE (lhsr)) |
1705cebd JJ |
2447 | || (op = optab_for_tree_code (code, TREE_TYPE (arg0), |
2448 | optab_default)) == unknown_optab | |
2449 | || (optab_handler (op, TYPE_MODE (TREE_TYPE (arg0))) | |
2450 | == CODE_FOR_nothing)) | |
2451 | { | |
2452 | if (MEM_P (lhsr)) | |
2453 | resv = make_tree (TREE_TYPE (lhs), lhsr); | |
2454 | else | |
2455 | { | |
2456 | resvr = assign_temp (TREE_TYPE (lhs), 1, 1); | |
2457 | resv = make_tree (TREE_TYPE (lhs), resvr); | |
2458 | } | |
2459 | } | |
2460 | } | |
07626e49 | 2461 | if (use_loop_p) |
1705cebd JJ |
2462 | { |
2463 | do_pending_stack_adjust (); | |
2464 | loop_lab = gen_label_rtx (); | |
2465 | cntvar = gen_reg_rtx (TYPE_MODE (sizetype)); | |
2466 | cntv = make_tree (sizetype, cntvar); | |
2467 | emit_move_insn (cntvar, const0_rtx); | |
2468 | emit_label (loop_lab); | |
2469 | } | |
2470 | if (TREE_CODE (arg0) != VECTOR_CST) | |
2471 | { | |
2472 | rtx arg0r = expand_normal (arg0); | |
2473 | arg0 = make_tree (TREE_TYPE (arg0), arg0r); | |
2474 | } | |
2475 | if (TREE_CODE (arg1) != VECTOR_CST) | |
2476 | { | |
2477 | rtx arg1r = expand_normal (arg1); | |
2478 | arg1 = make_tree (TREE_TYPE (arg1), arg1r); | |
2479 | } | |
07626e49 | 2480 | for (unsigned int i = 0; i < (use_loop_p ? 1 : const_cnt); i++) |
1705cebd JJ |
2481 | { |
2482 | tree op0, op1, res = NULL_TREE; | |
07626e49 | 2483 | if (use_loop_p) |
1705cebd JJ |
2484 | { |
2485 | tree atype = build_array_type_nelts (eltype, cnt); | |
4b48e883 JJ |
2486 | op0 = uniform_vector_p (arg0); |
2487 | if (op0 == NULL_TREE) | |
2488 | { | |
2489 | op0 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg0); | |
2490 | op0 = build4_loc (loc, ARRAY_REF, eltype, op0, cntv, | |
2491 | NULL_TREE, NULL_TREE); | |
2492 | } | |
2493 | op1 = uniform_vector_p (arg1); | |
2494 | if (op1 == NULL_TREE) | |
2495 | { | |
2496 | op1 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg1); | |
2497 | op1 = build4_loc (loc, ARRAY_REF, eltype, op1, cntv, | |
2498 | NULL_TREE, NULL_TREE); | |
2499 | } | |
1705cebd JJ |
2500 | if (resv) |
2501 | { | |
2502 | res = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, resv); | |
2503 | res = build4_loc (loc, ARRAY_REF, eltype, res, cntv, | |
2504 | NULL_TREE, NULL_TREE); | |
2505 | } | |
2506 | } | |
2507 | else | |
2508 | { | |
2509 | tree bitpos = bitsize_int (tree_to_uhwi (sz) * i); | |
2510 | op0 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg0, sz, bitpos); | |
2511 | op1 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg1, sz, bitpos); | |
2512 | if (resv) | |
2513 | res = fold_build3_loc (loc, BIT_FIELD_REF, eltype, resv, sz, | |
2514 | bitpos); | |
2515 | } | |
2516 | switch (code) | |
2517 | { | |
2518 | case PLUS_EXPR: | |
2519 | expand_addsub_overflow (loc, PLUS_EXPR, res, op0, op1, | |
2520 | false, false, false, true, &data); | |
2521 | break; | |
2522 | case MINUS_EXPR: | |
07626e49 | 2523 | if (use_loop_p ? integer_zerop (arg0) : integer_zerop (op0)) |
1705cebd JJ |
2524 | expand_neg_overflow (loc, res, op1, true, &data); |
2525 | else | |
2526 | expand_addsub_overflow (loc, MINUS_EXPR, res, op0, op1, | |
2527 | false, false, false, true, &data); | |
2528 | break; | |
2529 | case MULT_EXPR: | |
2530 | expand_mul_overflow (loc, res, op0, op1, false, false, false, | |
2531 | true, &data); | |
2532 | break; | |
2533 | default: | |
2534 | gcc_unreachable (); | |
2535 | } | |
2536 | } | |
07626e49 | 2537 | if (use_loop_p) |
1705cebd JJ |
2538 | { |
2539 | struct separate_ops ops; | |
2540 | ops.code = PLUS_EXPR; | |
2541 | ops.type = TREE_TYPE (cntv); | |
2542 | ops.op0 = cntv; | |
2543 | ops.op1 = build_int_cst (TREE_TYPE (cntv), 1); | |
2544 | ops.op2 = NULL_TREE; | |
2545 | ops.location = loc; | |
2546 | rtx ret = expand_expr_real_2 (&ops, cntvar, TYPE_MODE (sizetype), | |
2547 | EXPAND_NORMAL); | |
2548 | if (ret != cntvar) | |
2549 | emit_move_insn (cntvar, ret); | |
07626e49 RS |
2550 | rtx cntrtx = gen_int_mode (cnt, TYPE_MODE (sizetype)); |
2551 | do_compare_rtx_and_jump (cntvar, cntrtx, NE, false, | |
1705cebd | 2552 | TYPE_MODE (sizetype), NULL_RTX, NULL, loop_lab, |
357067f2 | 2553 | profile_probability::very_likely ()); |
1705cebd JJ |
2554 | } |
2555 | if (lhs && resv == NULL_TREE) | |
2556 | { | |
2557 | struct separate_ops ops; | |
2558 | ops.code = code; | |
2559 | ops.type = TREE_TYPE (arg0); | |
2560 | ops.op0 = arg0; | |
2561 | ops.op1 = arg1; | |
2562 | ops.op2 = NULL_TREE; | |
2563 | ops.location = loc; | |
2564 | rtx ret = expand_expr_real_2 (&ops, lhsr, TYPE_MODE (TREE_TYPE (arg0)), | |
2565 | EXPAND_NORMAL); | |
2566 | if (ret != lhsr) | |
2567 | emit_move_insn (lhsr, ret); | |
2568 | } | |
2569 | else if (resvr) | |
2570 | emit_move_insn (lhsr, resvr); | |
2571 | } | |
2572 | ||
31e071ae MP |
2573 | /* Expand UBSAN_CHECK_ADD call STMT. */ |
2574 | ||
2575 | static void | |
4cfe7a6c | 2576 | expand_UBSAN_CHECK_ADD (internal_fn, gcall *stmt) |
31e071ae | 2577 | { |
1304953e JJ |
2578 | location_t loc = gimple_location (stmt); |
2579 | tree lhs = gimple_call_lhs (stmt); | |
2580 | tree arg0 = gimple_call_arg (stmt, 0); | |
2581 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2582 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2583 | expand_vector_ubsan_overflow (loc, PLUS_EXPR, lhs, arg0, arg1); | |
2584 | else | |
2585 | expand_addsub_overflow (loc, PLUS_EXPR, lhs, arg0, arg1, | |
2586 | false, false, false, true, NULL); | |
31e071ae MP |
2587 | } |
2588 | ||
2589 | /* Expand UBSAN_CHECK_SUB call STMT. */ | |
2590 | ||
2591 | static void | |
4cfe7a6c | 2592 | expand_UBSAN_CHECK_SUB (internal_fn, gcall *stmt) |
31e071ae | 2593 | { |
1304953e JJ |
2594 | location_t loc = gimple_location (stmt); |
2595 | tree lhs = gimple_call_lhs (stmt); | |
2596 | tree arg0 = gimple_call_arg (stmt, 0); | |
2597 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2598 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2599 | expand_vector_ubsan_overflow (loc, MINUS_EXPR, lhs, arg0, arg1); | |
2600 | else if (integer_zerop (arg0)) | |
2601 | expand_neg_overflow (loc, lhs, arg1, true, NULL); | |
31e071ae | 2602 | else |
1304953e | 2603 | expand_addsub_overflow (loc, MINUS_EXPR, lhs, arg0, arg1, |
1705cebd | 2604 | false, false, false, true, NULL); |
31e071ae MP |
2605 | } |
2606 | ||
2607 | /* Expand UBSAN_CHECK_MUL call STMT. */ | |
2608 | ||
2609 | static void | |
4cfe7a6c | 2610 | expand_UBSAN_CHECK_MUL (internal_fn, gcall *stmt) |
31e071ae | 2611 | { |
1304953e JJ |
2612 | location_t loc = gimple_location (stmt); |
2613 | tree lhs = gimple_call_lhs (stmt); | |
2614 | tree arg0 = gimple_call_arg (stmt, 0); | |
2615 | tree arg1 = gimple_call_arg (stmt, 1); | |
1705cebd JJ |
2616 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2617 | expand_vector_ubsan_overflow (loc, MULT_EXPR, lhs, arg0, arg1); | |
2618 | else | |
2619 | expand_mul_overflow (loc, lhs, arg0, arg1, false, false, false, true, | |
2620 | NULL); | |
1304953e JJ |
2621 | } |
2622 | ||
2623 | /* Helper function for {ADD,SUB,MUL}_OVERFLOW call stmt expansion. */ | |
2624 | ||
2625 | static void | |
355fe088 | 2626 | expand_arith_overflow (enum tree_code code, gimple *stmt) |
1304953e JJ |
2627 | { |
2628 | tree lhs = gimple_call_lhs (stmt); | |
2629 | if (lhs == NULL_TREE) | |
2630 | return; | |
2631 | tree arg0 = gimple_call_arg (stmt, 0); | |
2632 | tree arg1 = gimple_call_arg (stmt, 1); | |
2633 | tree type = TREE_TYPE (TREE_TYPE (lhs)); | |
2634 | int uns0_p = TYPE_UNSIGNED (TREE_TYPE (arg0)); | |
2635 | int uns1_p = TYPE_UNSIGNED (TREE_TYPE (arg1)); | |
2636 | int unsr_p = TYPE_UNSIGNED (type); | |
2637 | int prec0 = TYPE_PRECISION (TREE_TYPE (arg0)); | |
2638 | int prec1 = TYPE_PRECISION (TREE_TYPE (arg1)); | |
2639 | int precres = TYPE_PRECISION (type); | |
2640 | location_t loc = gimple_location (stmt); | |
2641 | if (!uns0_p && get_range_pos_neg (arg0) == 1) | |
2642 | uns0_p = true; | |
2643 | if (!uns1_p && get_range_pos_neg (arg1) == 1) | |
2644 | uns1_p = true; | |
2645 | int pr = get_min_precision (arg0, uns0_p ? UNSIGNED : SIGNED); | |
2646 | prec0 = MIN (prec0, pr); | |
2647 | pr = get_min_precision (arg1, uns1_p ? UNSIGNED : SIGNED); | |
2648 | prec1 = MIN (prec1, pr); | |
2649 | ||
2650 | /* If uns0_p && uns1_p, precop is minimum needed precision | |
2651 | of unsigned type to hold the exact result, otherwise | |
2652 | precop is minimum needed precision of signed type to | |
2653 | hold the exact result. */ | |
2654 | int precop; | |
2655 | if (code == MULT_EXPR) | |
2656 | precop = prec0 + prec1 + (uns0_p != uns1_p); | |
2657 | else | |
2658 | { | |
2659 | if (uns0_p == uns1_p) | |
2660 | precop = MAX (prec0, prec1) + 1; | |
2661 | else if (uns0_p) | |
2662 | precop = MAX (prec0 + 1, prec1) + 1; | |
2663 | else | |
2664 | precop = MAX (prec0, prec1 + 1) + 1; | |
2665 | } | |
2666 | int orig_precres = precres; | |
2667 | ||
2668 | do | |
2669 | { | |
2670 | if ((uns0_p && uns1_p) | |
2671 | ? ((precop + !unsr_p) <= precres | |
2672 | /* u1 - u2 -> ur can overflow, no matter what precision | |
2673 | the result has. */ | |
2674 | && (code != MINUS_EXPR || !unsr_p)) | |
2675 | : (!unsr_p && precop <= precres)) | |
2676 | { | |
2677 | /* The infinity precision result will always fit into result. */ | |
2678 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
13f44099 | 2679 | write_complex_part (target, const0_rtx, true, false); |
7a504f33 | 2680 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (type); |
1304953e JJ |
2681 | struct separate_ops ops; |
2682 | ops.code = code; | |
2683 | ops.type = type; | |
2684 | ops.op0 = fold_convert_loc (loc, type, arg0); | |
2685 | ops.op1 = fold_convert_loc (loc, type, arg1); | |
2686 | ops.op2 = NULL_TREE; | |
2687 | ops.location = loc; | |
2688 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
2689 | expand_arith_overflow_result_store (lhs, target, mode, tem); | |
2690 | return; | |
2691 | } | |
2692 | ||
894d8b41 EB |
2693 | /* For operations with low precision, if target doesn't have them, start |
2694 | with precres widening right away, otherwise do it only if the most | |
2695 | simple cases can't be used. */ | |
2696 | const int min_precision = targetm.min_arithmetic_precision (); | |
2697 | if (orig_precres == precres && precres < min_precision) | |
1304953e | 2698 | ; |
9e11bfef TS |
2699 | else if ((uns0_p && uns1_p && unsr_p && prec0 <= precres |
2700 | && prec1 <= precres) | |
1304953e JJ |
2701 | || ((!uns0_p || !uns1_p) && !unsr_p |
2702 | && prec0 + uns0_p <= precres | |
2703 | && prec1 + uns1_p <= precres)) | |
2704 | { | |
2705 | arg0 = fold_convert_loc (loc, type, arg0); | |
2706 | arg1 = fold_convert_loc (loc, type, arg1); | |
2707 | switch (code) | |
2708 | { | |
2709 | case MINUS_EXPR: | |
2710 | if (integer_zerop (arg0) && !unsr_p) | |
7d704548 | 2711 | { |
1705cebd | 2712 | expand_neg_overflow (loc, lhs, arg1, false, NULL); |
7d704548 JJ |
2713 | return; |
2714 | } | |
1304953e JJ |
2715 | /* FALLTHRU */ |
2716 | case PLUS_EXPR: | |
1705cebd JJ |
2717 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2718 | unsr_p, unsr_p, false, NULL); | |
1304953e JJ |
2719 | return; |
2720 | case MULT_EXPR: | |
1705cebd JJ |
2721 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2722 | unsr_p, unsr_p, false, NULL); | |
1304953e JJ |
2723 | return; |
2724 | default: | |
2725 | gcc_unreachable (); | |
2726 | } | |
2727 | } | |
2728 | ||
2729 | /* For sub-word operations, retry with a wider type first. */ | |
2730 | if (orig_precres == precres && precop <= BITS_PER_WORD) | |
2731 | { | |
894d8b41 | 2732 | int p = MAX (min_precision, precop); |
f67f4dff | 2733 | scalar_int_mode m = smallest_int_mode_for_size (p); |
1304953e JJ |
2734 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (m), |
2735 | uns0_p && uns1_p | |
2736 | && unsr_p); | |
2737 | p = TYPE_PRECISION (optype); | |
2738 | if (p > precres) | |
2739 | { | |
2740 | precres = p; | |
2741 | unsr_p = TYPE_UNSIGNED (optype); | |
2742 | type = optype; | |
2743 | continue; | |
2744 | } | |
2745 | } | |
2746 | ||
2747 | if (prec0 <= precres && prec1 <= precres) | |
2748 | { | |
2749 | tree types[2]; | |
2750 | if (unsr_p) | |
2751 | { | |
2752 | types[0] = build_nonstandard_integer_type (precres, 0); | |
2753 | types[1] = type; | |
2754 | } | |
2755 | else | |
2756 | { | |
2757 | types[0] = type; | |
2758 | types[1] = build_nonstandard_integer_type (precres, 1); | |
2759 | } | |
2760 | arg0 = fold_convert_loc (loc, types[uns0_p], arg0); | |
2761 | arg1 = fold_convert_loc (loc, types[uns1_p], arg1); | |
2762 | if (code != MULT_EXPR) | |
2763 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, | |
1705cebd | 2764 | uns0_p, uns1_p, false, NULL); |
1304953e JJ |
2765 | else |
2766 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, | |
1705cebd | 2767 | uns0_p, uns1_p, false, NULL); |
1304953e JJ |
2768 | return; |
2769 | } | |
2770 | ||
2771 | /* Retry with a wider type. */ | |
2772 | if (orig_precres == precres) | |
2773 | { | |
2774 | int p = MAX (prec0, prec1); | |
f67f4dff | 2775 | scalar_int_mode m = smallest_int_mode_for_size (p); |
1304953e JJ |
2776 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (m), |
2777 | uns0_p && uns1_p | |
2778 | && unsr_p); | |
2779 | p = TYPE_PRECISION (optype); | |
2780 | if (p > precres) | |
2781 | { | |
2782 | precres = p; | |
2783 | unsr_p = TYPE_UNSIGNED (optype); | |
2784 | type = optype; | |
2785 | continue; | |
2786 | } | |
2787 | } | |
2788 | ||
2789 | gcc_unreachable (); | |
2790 | } | |
2791 | while (1); | |
2792 | } | |
2793 | ||
2794 | /* Expand ADD_OVERFLOW STMT. */ | |
2795 | ||
2796 | static void | |
4cfe7a6c | 2797 | expand_ADD_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2798 | { |
2799 | expand_arith_overflow (PLUS_EXPR, stmt); | |
2800 | } | |
2801 | ||
2802 | /* Expand SUB_OVERFLOW STMT. */ | |
2803 | ||
2804 | static void | |
4cfe7a6c | 2805 | expand_SUB_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2806 | { |
2807 | expand_arith_overflow (MINUS_EXPR, stmt); | |
2808 | } | |
2809 | ||
2810 | /* Expand MUL_OVERFLOW STMT. */ | |
2811 | ||
2812 | static void | |
4cfe7a6c | 2813 | expand_MUL_OVERFLOW (internal_fn, gcall *stmt) |
1304953e JJ |
2814 | { |
2815 | expand_arith_overflow (MULT_EXPR, stmt); | |
31e071ae MP |
2816 | } |
2817 | ||
43a3252c JJ |
2818 | /* Expand UADDC STMT. */ |
2819 | ||
2820 | static void | |
2821 | expand_UADDC (internal_fn ifn, gcall *stmt) | |
2822 | { | |
2823 | tree lhs = gimple_call_lhs (stmt); | |
2824 | tree arg1 = gimple_call_arg (stmt, 0); | |
2825 | tree arg2 = gimple_call_arg (stmt, 1); | |
2826 | tree arg3 = gimple_call_arg (stmt, 2); | |
2827 | tree type = TREE_TYPE (arg1); | |
2828 | machine_mode mode = TYPE_MODE (type); | |
2829 | insn_code icode = optab_handler (ifn == IFN_UADDC | |
2830 | ? uaddc5_optab : usubc5_optab, mode); | |
2831 | rtx op1 = expand_normal (arg1); | |
2832 | rtx op2 = expand_normal (arg2); | |
2833 | rtx op3 = expand_normal (arg3); | |
2834 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
2835 | rtx re = gen_reg_rtx (mode); | |
2836 | rtx im = gen_reg_rtx (mode); | |
2837 | class expand_operand ops[5]; | |
2838 | create_output_operand (&ops[0], re, mode); | |
2839 | create_output_operand (&ops[1], im, mode); | |
2840 | create_input_operand (&ops[2], op1, mode); | |
2841 | create_input_operand (&ops[3], op2, mode); | |
2842 | create_input_operand (&ops[4], op3, mode); | |
2843 | expand_insn (icode, 5, ops); | |
2844 | write_complex_part (target, re, false, false); | |
2845 | write_complex_part (target, im, true, false); | |
2846 | } | |
2847 | ||
2848 | /* Expand USUBC STMT. */ | |
2849 | ||
2850 | static void | |
2851 | expand_USUBC (internal_fn ifn, gcall *stmt) | |
2852 | { | |
2853 | expand_UADDC (ifn, stmt); | |
2854 | } | |
2855 | ||
e53b6e56 | 2856 | /* This should get folded in tree-vectorizer.cc. */ |
5ce9450f JJ |
2857 | |
2858 | static void | |
4cfe7a6c | 2859 | expand_LOOP_VECTORIZED (internal_fn, gcall *) |
5ce9450f JJ |
2860 | { |
2861 | gcc_unreachable (); | |
2862 | } | |
2863 | ||
e53b6e56 | 2864 | /* This should get folded in tree-vectorizer.cc. */ |
542e7230 BC |
2865 | |
2866 | static void | |
2867 | expand_LOOP_DIST_ALIAS (internal_fn, gcall *) | |
2868 | { | |
2869 | gcc_unreachable (); | |
2870 | } | |
2871 | ||
65dd1346 RS |
2872 | /* Return a memory reference of type TYPE for argument INDEX of STMT. |
2873 | Use argument INDEX + 1 to derive the second (TBAA) operand. */ | |
2874 | ||
2875 | static tree | |
2876 | expand_call_mem_ref (tree type, gcall *stmt, int index) | |
2877 | { | |
2878 | tree addr = gimple_call_arg (stmt, index); | |
2879 | tree alias_ptr_type = TREE_TYPE (gimple_call_arg (stmt, index + 1)); | |
2880 | unsigned int align = tree_to_shwi (gimple_call_arg (stmt, index + 1)); | |
2881 | if (TYPE_ALIGN (type) != align) | |
2882 | type = build_aligned_type (type, align); | |
2883 | ||
2884 | tree tmp = addr; | |
2885 | if (TREE_CODE (tmp) == SSA_NAME) | |
2886 | { | |
2887 | gimple *def = SSA_NAME_DEF_STMT (tmp); | |
2888 | if (gimple_assign_single_p (def)) | |
2889 | tmp = gimple_assign_rhs1 (def); | |
2890 | } | |
2891 | ||
2892 | if (TREE_CODE (tmp) == ADDR_EXPR) | |
2893 | { | |
2894 | tree mem = TREE_OPERAND (tmp, 0); | |
2895 | if (TREE_CODE (mem) == TARGET_MEM_REF | |
2896 | && types_compatible_p (TREE_TYPE (mem), type)) | |
2897 | { | |
2898 | tree offset = TMR_OFFSET (mem); | |
9bd958c5 RS |
2899 | if (type != TREE_TYPE (mem) |
2900 | || alias_ptr_type != TREE_TYPE (offset) | |
2901 | || !integer_zerop (offset)) | |
65dd1346 RS |
2902 | { |
2903 | mem = copy_node (mem); | |
2904 | TMR_OFFSET (mem) = wide_int_to_tree (alias_ptr_type, | |
2905 | wi::to_poly_wide (offset)); | |
9bd958c5 | 2906 | TREE_TYPE (mem) = type; |
65dd1346 RS |
2907 | } |
2908 | return mem; | |
2909 | } | |
2910 | } | |
2911 | ||
2912 | return fold_build2 (MEM_REF, type, addr, build_int_cst (alias_ptr_type, 0)); | |
2913 | } | |
2914 | ||
bd68b33f | 2915 | /* Expand MASK_LOAD{,_LANES}, MASK_LEN_LOAD or LEN_LOAD call STMT using optab |
ccfdda34 | 2916 | * OPTAB. */ |
ab23f5d9 | 2917 | |
5ce9450f | 2918 | static void |
b8806f6f | 2919 | expand_partial_load_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
5ce9450f | 2920 | { |
b8806f6f | 2921 | int i = 0; |
ccfdda34 | 2922 | class expand_operand ops[5]; |
b8806f6f JZZ |
2923 | tree type, lhs, rhs, maskt; |
2924 | rtx mem, target; | |
7e11fc7f | 2925 | insn_code icode; |
5ce9450f | 2926 | |
b8806f6f | 2927 | maskt = gimple_call_arg (stmt, internal_fn_mask_index (ifn)); |
5ce9450f | 2928 | lhs = gimple_call_lhs (stmt); |
8e91d222 JJ |
2929 | if (lhs == NULL_TREE) |
2930 | return; | |
5ce9450f | 2931 | type = TREE_TYPE (lhs); |
65dd1346 | 2932 | rhs = expand_call_mem_ref (type, stmt, 0); |
5ce9450f | 2933 | |
7e11fc7f RS |
2934 | if (optab == vec_mask_load_lanes_optab) |
2935 | icode = get_multi_vector_move (type, optab); | |
d496134a KL |
2936 | else if (optab == len_load_optab) |
2937 | icode = direct_optab_handler (optab, TYPE_MODE (type)); | |
7e11fc7f RS |
2938 | else |
2939 | icode = convert_optab_handler (optab, TYPE_MODE (type), | |
2940 | TYPE_MODE (TREE_TYPE (maskt))); | |
2941 | ||
5ce9450f JJ |
2942 | mem = expand_expr (rhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
2943 | gcc_assert (MEM_P (mem)); | |
5ce9450f | 2944 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
b8806f6f JZZ |
2945 | create_output_operand (&ops[i++], target, TYPE_MODE (type)); |
2946 | create_fixed_operand (&ops[i++], mem); | |
363bb3dc | 2947 | i = add_mask_and_len_args (ops, i, stmt); |
b8806f6f | 2948 | expand_insn (icode, i, ops); |
b0e51639 | 2949 | |
3af3bec2 RS |
2950 | if (!rtx_equal_p (target, ops[0].value)) |
2951 | emit_move_insn (target, ops[0].value); | |
5ce9450f JJ |
2952 | } |
2953 | ||
d496134a | 2954 | #define expand_mask_load_optab_fn expand_partial_load_optab_fn |
7e11fc7f | 2955 | #define expand_mask_load_lanes_optab_fn expand_mask_load_optab_fn |
d496134a | 2956 | #define expand_len_load_optab_fn expand_partial_load_optab_fn |
bd68b33f | 2957 | #define expand_mask_len_load_optab_fn expand_partial_load_optab_fn |
7e11fc7f | 2958 | |
bd68b33f | 2959 | /* Expand MASK_STORE{,_LANES}, MASK_LEN_STORE or LEN_STORE call STMT using optab |
ccfdda34 | 2960 | * OPTAB. */ |
ab23f5d9 | 2961 | |
5ce9450f | 2962 | static void |
d39f4889 | 2963 | expand_partial_store_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
5ce9450f | 2964 | { |
b8806f6f | 2965 | int i = 0; |
ccfdda34 | 2966 | class expand_operand ops[5]; |
b8806f6f JZZ |
2967 | tree type, lhs, rhs, maskt; |
2968 | rtx mem, reg; | |
7e11fc7f | 2969 | insn_code icode; |
5ce9450f | 2970 | |
b8806f6f | 2971 | maskt = gimple_call_arg (stmt, internal_fn_mask_index (ifn)); |
d39f4889 | 2972 | rhs = gimple_call_arg (stmt, internal_fn_stored_value_index (ifn)); |
5ce9450f | 2973 | type = TREE_TYPE (rhs); |
65dd1346 | 2974 | lhs = expand_call_mem_ref (type, stmt, 0); |
5ce9450f | 2975 | |
7e11fc7f RS |
2976 | if (optab == vec_mask_store_lanes_optab) |
2977 | icode = get_multi_vector_move (type, optab); | |
d496134a KL |
2978 | else if (optab == len_store_optab) |
2979 | icode = direct_optab_handler (optab, TYPE_MODE (type)); | |
7e11fc7f RS |
2980 | else |
2981 | icode = convert_optab_handler (optab, TYPE_MODE (type), | |
2982 | TYPE_MODE (TREE_TYPE (maskt))); | |
2983 | ||
5ce9450f JJ |
2984 | mem = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
2985 | gcc_assert (MEM_P (mem)); | |
5ce9450f | 2986 | reg = expand_normal (rhs); |
b8806f6f JZZ |
2987 | create_fixed_operand (&ops[i++], mem); |
2988 | create_input_operand (&ops[i++], reg, TYPE_MODE (type)); | |
363bb3dc | 2989 | i = add_mask_and_len_args (ops, i, stmt); |
b8806f6f | 2990 | expand_insn (icode, i, ops); |
5ce9450f JJ |
2991 | } |
2992 | ||
d496134a | 2993 | #define expand_mask_store_optab_fn expand_partial_store_optab_fn |
7e11fc7f | 2994 | #define expand_mask_store_lanes_optab_fn expand_mask_store_optab_fn |
d496134a | 2995 | #define expand_len_store_optab_fn expand_partial_store_optab_fn |
bd68b33f | 2996 | #define expand_mask_len_store_optab_fn expand_partial_store_optab_fn |
7e11fc7f | 2997 | |
502d63b6 ML |
2998 | /* Expand VCOND, VCONDU and VCONDEQ optab internal functions. |
2999 | The expansion of STMT happens based on OPTAB table associated. */ | |
3000 | ||
3001 | static void | |
298e76e6 | 3002 | expand_vec_cond_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
502d63b6 ML |
3003 | { |
3004 | class expand_operand ops[6]; | |
3005 | insn_code icode; | |
3006 | tree lhs = gimple_call_lhs (stmt); | |
3007 | tree op0a = gimple_call_arg (stmt, 0); | |
3008 | tree op0b = gimple_call_arg (stmt, 1); | |
3009 | tree op1 = gimple_call_arg (stmt, 2); | |
3010 | tree op2 = gimple_call_arg (stmt, 3); | |
3011 | enum tree_code tcode = (tree_code) int_cst_value (gimple_call_arg (stmt, 4)); | |
3012 | ||
3013 | tree vec_cond_type = TREE_TYPE (lhs); | |
3014 | tree op_mode = TREE_TYPE (op0a); | |
3015 | bool unsignedp = TYPE_UNSIGNED (op_mode); | |
3016 | ||
3017 | machine_mode mode = TYPE_MODE (vec_cond_type); | |
3018 | machine_mode cmp_op_mode = TYPE_MODE (op_mode); | |
3019 | ||
3020 | icode = convert_optab_handler (optab, mode, cmp_op_mode); | |
3021 | rtx comparison | |
3022 | = vector_compare_rtx (VOIDmode, tcode, op0a, op0b, unsignedp, icode, 4); | |
ceae1400 RB |
3023 | /* vector_compare_rtx legitimizes operands, preserve equality when |
3024 | expanding op1/op2. */ | |
3025 | rtx rtx_op1, rtx_op2; | |
3026 | if (operand_equal_p (op1, op0a)) | |
3027 | rtx_op1 = XEXP (comparison, 0); | |
3028 | else if (operand_equal_p (op1, op0b)) | |
3029 | rtx_op1 = XEXP (comparison, 1); | |
3030 | else | |
3031 | rtx_op1 = expand_normal (op1); | |
3032 | if (operand_equal_p (op2, op0a)) | |
3033 | rtx_op2 = XEXP (comparison, 0); | |
3034 | else if (operand_equal_p (op2, op0b)) | |
3035 | rtx_op2 = XEXP (comparison, 1); | |
3036 | else | |
3037 | rtx_op2 = expand_normal (op2); | |
502d63b6 ML |
3038 | |
3039 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3040 | create_output_operand (&ops[0], target, mode); | |
3041 | create_input_operand (&ops[1], rtx_op1, mode); | |
3042 | create_input_operand (&ops[2], rtx_op2, mode); | |
3043 | create_fixed_operand (&ops[3], comparison); | |
3044 | create_fixed_operand (&ops[4], XEXP (comparison, 0)); | |
3045 | create_fixed_operand (&ops[5], XEXP (comparison, 1)); | |
3046 | expand_insn (icode, 6, ops); | |
eccc3d43 RB |
3047 | if (!rtx_equal_p (ops[0].value, target)) |
3048 | emit_move_insn (target, ops[0].value); | |
502d63b6 ML |
3049 | } |
3050 | ||
502d63b6 ML |
3051 | /* Expand VCOND_MASK optab internal function. |
3052 | The expansion of STMT happens based on OPTAB table associated. */ | |
3053 | ||
3054 | static void | |
298e76e6 | 3055 | expand_vec_cond_mask_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
502d63b6 ML |
3056 | { |
3057 | class expand_operand ops[4]; | |
3058 | ||
3059 | tree lhs = gimple_call_lhs (stmt); | |
3060 | tree op0 = gimple_call_arg (stmt, 0); | |
3061 | tree op1 = gimple_call_arg (stmt, 1); | |
3062 | tree op2 = gimple_call_arg (stmt, 2); | |
3063 | tree vec_cond_type = TREE_TYPE (lhs); | |
3064 | ||
3065 | machine_mode mode = TYPE_MODE (vec_cond_type); | |
3066 | machine_mode mask_mode = TYPE_MODE (TREE_TYPE (op0)); | |
3067 | enum insn_code icode = convert_optab_handler (optab, mode, mask_mode); | |
3068 | rtx mask, rtx_op1, rtx_op2; | |
3069 | ||
3070 | gcc_assert (icode != CODE_FOR_nothing); | |
3071 | ||
3072 | mask = expand_normal (op0); | |
3073 | rtx_op1 = expand_normal (op1); | |
3074 | rtx_op2 = expand_normal (op2); | |
3075 | ||
3076 | mask = force_reg (mask_mode, mask); | |
bc909324 | 3077 | rtx_op1 = force_reg (mode, rtx_op1); |
502d63b6 ML |
3078 | |
3079 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3080 | create_output_operand (&ops[0], target, mode); | |
3081 | create_input_operand (&ops[1], rtx_op1, mode); | |
3082 | create_input_operand (&ops[2], rtx_op2, mode); | |
3083 | create_input_operand (&ops[3], mask, mask_mode); | |
3084 | expand_insn (icode, 4, ops); | |
eccc3d43 RB |
3085 | if (!rtx_equal_p (ops[0].value, target)) |
3086 | emit_move_insn (target, ops[0].value); | |
502d63b6 ML |
3087 | } |
3088 | ||
683e55fa XL |
3089 | /* Expand VEC_SET internal functions. */ |
3090 | ||
3091 | static void | |
3092 | expand_vec_set_optab_fn (internal_fn, gcall *stmt, convert_optab optab) | |
3093 | { | |
3094 | tree lhs = gimple_call_lhs (stmt); | |
3095 | tree op0 = gimple_call_arg (stmt, 0); | |
3096 | tree op1 = gimple_call_arg (stmt, 1); | |
3097 | tree op2 = gimple_call_arg (stmt, 2); | |
3098 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3099 | rtx src = expand_normal (op0); | |
3100 | ||
3101 | machine_mode outermode = TYPE_MODE (TREE_TYPE (op0)); | |
3102 | scalar_mode innermode = GET_MODE_INNER (outermode); | |
3103 | ||
3104 | rtx value = expand_normal (op1); | |
3105 | rtx pos = expand_normal (op2); | |
3106 | ||
3107 | class expand_operand ops[3]; | |
3108 | enum insn_code icode = optab_handler (optab, outermode); | |
3109 | ||
3110 | if (icode != CODE_FOR_nothing) | |
3111 | { | |
3112 | rtx temp = gen_reg_rtx (outermode); | |
3113 | emit_move_insn (temp, src); | |
3114 | ||
3115 | create_fixed_operand (&ops[0], temp); | |
3116 | create_input_operand (&ops[1], value, innermode); | |
3117 | create_convert_operand_from (&ops[2], pos, TYPE_MODE (TREE_TYPE (op2)), | |
3118 | true); | |
3119 | if (maybe_expand_insn (icode, 3, ops)) | |
3120 | { | |
3121 | emit_move_insn (target, temp); | |
3122 | return; | |
3123 | } | |
3124 | } | |
3125 | gcc_unreachable (); | |
3126 | } | |
3127 | ||
c30efd8c RD |
3128 | /* Expand VEC_EXTRACT optab internal function. */ |
3129 | ||
3130 | static void | |
3131 | expand_vec_extract_optab_fn (internal_fn, gcall *stmt, convert_optab optab) | |
3132 | { | |
3133 | tree lhs = gimple_call_lhs (stmt); | |
3134 | tree op0 = gimple_call_arg (stmt, 0); | |
3135 | tree op1 = gimple_call_arg (stmt, 1); | |
3136 | ||
3137 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3138 | ||
3139 | machine_mode outermode = TYPE_MODE (TREE_TYPE (op0)); | |
3140 | machine_mode extract_mode = TYPE_MODE (TREE_TYPE (lhs)); | |
3141 | ||
3142 | rtx src = expand_normal (op0); | |
3143 | rtx pos = expand_normal (op1); | |
3144 | ||
3145 | class expand_operand ops[3]; | |
3146 | enum insn_code icode = convert_optab_handler (optab, outermode, | |
3147 | extract_mode); | |
3148 | ||
3149 | if (icode != CODE_FOR_nothing) | |
3150 | { | |
3151 | create_output_operand (&ops[0], target, extract_mode); | |
3152 | create_input_operand (&ops[1], src, outermode); | |
3153 | create_convert_operand_from (&ops[2], pos, | |
3154 | TYPE_MODE (TREE_TYPE (op1)), true); | |
3155 | if (maybe_expand_insn (icode, 3, ops)) | |
3156 | { | |
3157 | if (!rtx_equal_p (target, ops[0].value)) | |
3158 | emit_move_insn (target, ops[0].value); | |
3159 | return; | |
3160 | } | |
3161 | } | |
3162 | gcc_unreachable (); | |
3163 | } | |
3164 | ||
09b22f48 | 3165 | static void |
4cfe7a6c | 3166 | expand_ABNORMAL_DISPATCHER (internal_fn, gcall *) |
09b22f48 JJ |
3167 | { |
3168 | } | |
3169 | ||
ed9c79e1 | 3170 | static void |
4cfe7a6c | 3171 | expand_BUILTIN_EXPECT (internal_fn, gcall *stmt) |
ed9c79e1 JJ |
3172 | { |
3173 | /* When guessing was done, the hints should be already stripped away. */ | |
3174 | gcc_assert (!flag_guess_branch_prob || optimize == 0 || seen_error ()); | |
3175 | ||
3176 | rtx target; | |
3177 | tree lhs = gimple_call_lhs (stmt); | |
3178 | if (lhs) | |
3179 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3180 | else | |
3181 | target = const0_rtx; | |
3182 | rtx val = expand_expr (gimple_call_arg (stmt, 0), target, VOIDmode, EXPAND_NORMAL); | |
3183 | if (lhs && val != target) | |
3184 | emit_move_insn (target, val); | |
3185 | } | |
3186 | ||
f8e89441 TV |
3187 | /* IFN_VA_ARG is supposed to be expanded at pass_stdarg. So this dummy function |
3188 | should never be called. */ | |
3189 | ||
3190 | static void | |
4cfe7a6c | 3191 | expand_VA_ARG (internal_fn, gcall *) |
d8fcab68 JJ |
3192 | { |
3193 | gcc_unreachable (); | |
3194 | } | |
3195 | ||
3196 | /* IFN_VEC_CONVERT is supposed to be expanded at pass_lower_vector. So this | |
3197 | dummy function should never be called. */ | |
3198 | ||
3199 | static void | |
3200 | expand_VEC_CONVERT (internal_fn, gcall *) | |
f8e89441 TV |
3201 | { |
3202 | gcc_unreachable (); | |
3203 | } | |
3204 | ||
6f966f06 SSF |
3205 | /* Expand IFN_RAWMEMCHAR internal function. */ |
3206 | ||
3207 | void | |
3208 | expand_RAWMEMCHR (internal_fn, gcall *stmt) | |
3209 | { | |
3210 | expand_operand ops[3]; | |
3211 | ||
3212 | tree lhs = gimple_call_lhs (stmt); | |
3213 | if (!lhs) | |
3214 | return; | |
3215 | machine_mode lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); | |
3216 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3217 | create_output_operand (&ops[0], lhs_rtx, lhs_mode); | |
3218 | ||
3219 | tree mem = gimple_call_arg (stmt, 0); | |
3220 | rtx mem_rtx = get_memory_rtx (mem, NULL); | |
3221 | create_fixed_operand (&ops[1], mem_rtx); | |
3222 | ||
3223 | tree pattern = gimple_call_arg (stmt, 1); | |
3224 | machine_mode mode = TYPE_MODE (TREE_TYPE (pattern)); | |
3225 | rtx pattern_rtx = expand_normal (pattern); | |
3226 | create_input_operand (&ops[2], pattern_rtx, mode); | |
3227 | ||
3228 | insn_code icode = direct_optab_handler (rawmemchr_optab, mode); | |
3229 | ||
3230 | expand_insn (icode, 3, ops); | |
3231 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) | |
3232 | emit_move_insn (lhs_rtx, ops[0].value); | |
3233 | } | |
3234 | ||
8ab78162 NS |
3235 | /* Expand the IFN_UNIQUE function according to its first argument. */ |
3236 | ||
3237 | static void | |
4cfe7a6c | 3238 | expand_UNIQUE (internal_fn, gcall *stmt) |
8ab78162 NS |
3239 | { |
3240 | rtx pattern = NULL_RTX; | |
3241 | enum ifn_unique_kind kind | |
3242 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (stmt, 0)); | |
3243 | ||
3244 | switch (kind) | |
3245 | { | |
3246 | default: | |
3247 | gcc_unreachable (); | |
3248 | ||
3249 | case IFN_UNIQUE_UNSPEC: | |
3250 | if (targetm.have_unique ()) | |
3251 | pattern = targetm.gen_unique (); | |
3252 | break; | |
9bd46bc9 NS |
3253 | |
3254 | case IFN_UNIQUE_OACC_FORK: | |
3255 | case IFN_UNIQUE_OACC_JOIN: | |
3256 | if (targetm.have_oacc_fork () && targetm.have_oacc_join ()) | |
3257 | { | |
3258 | tree lhs = gimple_call_lhs (stmt); | |
3259 | rtx target = const0_rtx; | |
3260 | ||
3261 | if (lhs) | |
3262 | target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3263 | ||
3264 | rtx data_dep = expand_normal (gimple_call_arg (stmt, 1)); | |
3265 | rtx axis = expand_normal (gimple_call_arg (stmt, 2)); | |
3266 | ||
3267 | if (kind == IFN_UNIQUE_OACC_FORK) | |
3268 | pattern = targetm.gen_oacc_fork (target, data_dep, axis); | |
3269 | else | |
3270 | pattern = targetm.gen_oacc_join (target, data_dep, axis); | |
3271 | } | |
3272 | else | |
3273 | gcc_unreachable (); | |
3274 | break; | |
8ab78162 NS |
3275 | } |
3276 | ||
3277 | if (pattern) | |
3278 | emit_insn (pattern); | |
3279 | } | |
3280 | ||
a25e0b5e | 3281 | /* Expand the IFN_DEFERRED_INIT function: |
6c98c8b4 | 3282 | LHS = DEFERRED_INIT (SIZE of the DECL, INIT_TYPE, NAME of the DECL); |
a25e0b5e | 3283 | |
3284 | Initialize the LHS with zero/pattern according to its second argument | |
3285 | INIT_TYPE: | |
3286 | if INIT_TYPE is AUTO_INIT_ZERO, use zeroes to initialize; | |
3287 | if INIT_TYPE is AUTO_INIT_PATTERN, use 0xFE byte-repeatable pattern | |
3288 | to initialize; | |
3289 | The LHS variable is initialized including paddings. | |
3290 | The reasons to choose 0xFE for pattern initialization are: | |
3291 | 1. It is a non-canonical virtual address on x86_64, and at the | |
3292 | high end of the i386 kernel address space. | |
3293 | 2. It is a very large float value (-1.694739530317379e+38). | |
3294 | 3. It is also an unusual number for integers. */ | |
3295 | #define INIT_PATTERN_VALUE 0xFE | |
3296 | static void | |
3297 | expand_DEFERRED_INIT (internal_fn, gcall *stmt) | |
3298 | { | |
3299 | tree lhs = gimple_call_lhs (stmt); | |
3300 | tree var_size = gimple_call_arg (stmt, 0); | |
3301 | enum auto_init_type init_type | |
3302 | = (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1)); | |
a25e0b5e | 3303 | bool reg_lhs = true; |
3304 | ||
3305 | tree var_type = TREE_TYPE (lhs); | |
3306 | gcc_assert (init_type > AUTO_INIT_UNINITIALIZED); | |
3307 | ||
79f488de | 3308 | if (TREE_CODE (lhs) == SSA_NAME) |
a25e0b5e | 3309 | reg_lhs = true; |
3310 | else | |
3311 | { | |
7e0c0500 RB |
3312 | tree lhs_base = lhs; |
3313 | while (handled_component_p (lhs_base)) | |
3314 | lhs_base = TREE_OPERAND (lhs_base, 0); | |
3315 | reg_lhs = (mem_ref_refers_to_non_mem_p (lhs_base) | |
3316 | || non_mem_decl_p (lhs_base)); | |
06b8cdc8 RB |
3317 | /* If this expands to a register and the underlying decl is wrapped in |
3318 | a MEM_REF that just serves as an access type change expose the decl | |
3319 | if it is of correct size. This avoids a situation as in PR103271 | |
3320 | if the target does not support a direct move to the registers mode. */ | |
3321 | if (reg_lhs | |
3322 | && TREE_CODE (lhs_base) == MEM_REF | |
3323 | && TREE_CODE (TREE_OPERAND (lhs_base, 0)) == ADDR_EXPR | |
3324 | && DECL_P (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)) | |
3325 | && integer_zerop (TREE_OPERAND (lhs_base, 1)) | |
3326 | && tree_fits_uhwi_p (var_size) | |
3327 | && tree_int_cst_equal | |
3328 | (var_size, | |
3329 | DECL_SIZE_UNIT (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)))) | |
3330 | { | |
3331 | lhs = TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0); | |
3332 | var_type = TREE_TYPE (lhs); | |
3333 | } | |
a25e0b5e | 3334 | } |
3335 | ||
a25e0b5e | 3336 | if (!reg_lhs) |
3337 | { | |
6c98c8b4 QZ |
3338 | /* If the variable is not in register, expand to a memset |
3339 | to initialize it. */ | |
a25e0b5e | 3340 | mark_addressable (lhs); |
3341 | tree var_addr = build_fold_addr_expr (lhs); | |
3342 | ||
1c04af34 QZ |
3343 | tree value = (init_type == AUTO_INIT_PATTERN) |
3344 | ? build_int_cst (integer_type_node, | |
3345 | INIT_PATTERN_VALUE) | |
3346 | : integer_zero_node; | |
a25e0b5e | 3347 | tree m_call = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMSET), |
3348 | 3, var_addr, value, var_size); | |
3349 | /* Expand this memset call. */ | |
3350 | expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type)); | |
3351 | } | |
604459a0 | 3352 | else |
a25e0b5e | 3353 | { |
c081d0a3 RB |
3354 | /* If this variable is in a register use expand_assignment. |
3355 | For boolean scalars force zero-init. */ | |
604459a0 | 3356 | tree init; |
1c04af34 | 3357 | scalar_int_mode var_mode; |
c081d0a3 RB |
3358 | if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE |
3359 | && tree_fits_uhwi_p (var_size) | |
604459a0 | 3360 | && (init_type == AUTO_INIT_PATTERN |
33872565 RB |
3361 | || !is_gimple_reg_type (var_type)) |
3362 | && int_mode_for_size (tree_to_uhwi (var_size) * BITS_PER_UNIT, | |
1c04af34 QZ |
3363 | 0).exists (&var_mode) |
3364 | && have_insn_for (SET, var_mode)) | |
a25e0b5e | 3365 | { |
604459a0 | 3366 | unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size); |
1c04af34 | 3367 | unsigned char *buf = XALLOCAVEC (unsigned char, total_bytes); |
604459a0 RB |
3368 | memset (buf, (init_type == AUTO_INIT_PATTERN |
3369 | ? INIT_PATTERN_VALUE : 0), total_bytes); | |
c081d0a3 RB |
3370 | tree itype = build_nonstandard_integer_type |
3371 | (total_bytes * BITS_PER_UNIT, 1); | |
3372 | wide_int w = wi::from_buffer (buf, total_bytes); | |
3373 | init = wide_int_to_tree (itype, w); | |
3374 | /* Pun the LHS to make sure its type has constant size | |
3375 | unless it is an SSA name where that's already known. */ | |
3376 | if (TREE_CODE (lhs) != SSA_NAME) | |
3377 | lhs = build1 (VIEW_CONVERT_EXPR, itype, lhs); | |
8d6b12b2 | 3378 | else |
c081d0a3 | 3379 | init = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (lhs), init); |
a25e0b5e | 3380 | } |
604459a0 RB |
3381 | else |
3382 | /* Use zero-init also for variable-length sizes. */ | |
3383 | init = build_zero_cst (var_type); | |
a25e0b5e | 3384 | |
3385 | expand_assignment (lhs, init, false); | |
3386 | } | |
3387 | } | |
3388 | ||
9bd46bc9 NS |
3389 | /* The size of an OpenACC compute dimension. */ |
3390 | ||
3391 | static void | |
4cfe7a6c | 3392 | expand_GOACC_DIM_SIZE (internal_fn, gcall *stmt) |
9bd46bc9 NS |
3393 | { |
3394 | tree lhs = gimple_call_lhs (stmt); | |
3395 | ||
3396 | if (!lhs) | |
3397 | return; | |
3398 | ||
3399 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3400 | if (targetm.have_oacc_dim_size ()) | |
3401 | { | |
3402 | rtx dim = expand_expr (gimple_call_arg (stmt, 0), NULL_RTX, | |
3403 | VOIDmode, EXPAND_NORMAL); | |
3404 | emit_insn (targetm.gen_oacc_dim_size (target, dim)); | |
3405 | } | |
3406 | else | |
3407 | emit_move_insn (target, GEN_INT (1)); | |
3408 | } | |
3409 | ||
3410 | /* The position of an OpenACC execution engine along one compute axis. */ | |
3411 | ||
3412 | static void | |
4cfe7a6c | 3413 | expand_GOACC_DIM_POS (internal_fn, gcall *stmt) |
9bd46bc9 NS |
3414 | { |
3415 | tree lhs = gimple_call_lhs (stmt); | |
3416 | ||
3417 | if (!lhs) | |
3418 | return; | |
3419 | ||
3420 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3421 | if (targetm.have_oacc_dim_pos ()) | |
3422 | { | |
3423 | rtx dim = expand_expr (gimple_call_arg (stmt, 0), NULL_RTX, | |
3424 | VOIDmode, EXPAND_NORMAL); | |
3425 | emit_insn (targetm.gen_oacc_dim_pos (target, dim)); | |
3426 | } | |
3427 | else | |
3428 | emit_move_insn (target, const0_rtx); | |
3429 | } | |
3430 | ||
3431 | /* This is expanded by oacc_device_lower pass. */ | |
3432 | ||
3433 | static void | |
4cfe7a6c | 3434 | expand_GOACC_LOOP (internal_fn, gcall *) |
9bd46bc9 NS |
3435 | { |
3436 | gcc_unreachable (); | |
3437 | } | |
3438 | ||
e5014671 NS |
3439 | /* This is expanded by oacc_device_lower pass. */ |
3440 | ||
3441 | static void | |
4cfe7a6c | 3442 | expand_GOACC_REDUCTION (internal_fn, gcall *) |
e5014671 NS |
3443 | { |
3444 | gcc_unreachable (); | |
02889d23 CLT |
3445 | } |
3446 | ||
3447 | /* This is expanded by oacc_device_lower pass. */ | |
3448 | ||
3449 | static void | |
3450 | expand_GOACC_TILE (internal_fn, gcall *) | |
3451 | { | |
3452 | gcc_unreachable (); | |
e5014671 NS |
3453 | } |
3454 | ||
883cabde RS |
3455 | /* Set errno to EDOM. */ |
3456 | ||
3457 | static void | |
3458 | expand_SET_EDOM (internal_fn, gcall *) | |
3459 | { | |
3460 | #ifdef TARGET_EDOM | |
3461 | #ifdef GEN_ERRNO_RTX | |
3462 | rtx errno_rtx = GEN_ERRNO_RTX; | |
3463 | #else | |
3464 | rtx errno_rtx = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno")); | |
3465 | #endif | |
3466 | emit_move_insn (errno_rtx, | |
3467 | gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); | |
3468 | #else | |
3469 | gcc_unreachable (); | |
3470 | #endif | |
3471 | } | |
3472 | ||
adedd5c1 JJ |
3473 | /* Expand atomic bit test and set. */ |
3474 | ||
3475 | static void | |
3476 | expand_ATOMIC_BIT_TEST_AND_SET (internal_fn, gcall *call) | |
3477 | { | |
3478 | expand_ifn_atomic_bit_test_and (call); | |
3479 | } | |
3480 | ||
3481 | /* Expand atomic bit test and complement. */ | |
3482 | ||
3483 | static void | |
3484 | expand_ATOMIC_BIT_TEST_AND_COMPLEMENT (internal_fn, gcall *call) | |
3485 | { | |
3486 | expand_ifn_atomic_bit_test_and (call); | |
3487 | } | |
3488 | ||
3489 | /* Expand atomic bit test and reset. */ | |
3490 | ||
3491 | static void | |
3492 | expand_ATOMIC_BIT_TEST_AND_RESET (internal_fn, gcall *call) | |
3493 | { | |
3494 | expand_ifn_atomic_bit_test_and (call); | |
3495 | } | |
3496 | ||
849a76a5 JJ |
3497 | /* Expand atomic bit test and set. */ |
3498 | ||
3499 | static void | |
3500 | expand_ATOMIC_COMPARE_EXCHANGE (internal_fn, gcall *call) | |
3501 | { | |
3502 | expand_ifn_atomic_compare_exchange (call); | |
3503 | } | |
3504 | ||
6362627b JJ |
3505 | /* Expand atomic add fetch and cmp with 0. */ |
3506 | ||
3507 | static void | |
3508 | expand_ATOMIC_ADD_FETCH_CMP_0 (internal_fn, gcall *call) | |
3509 | { | |
3510 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3511 | } | |
3512 | ||
3513 | /* Expand atomic sub fetch and cmp with 0. */ | |
3514 | ||
3515 | static void | |
3516 | expand_ATOMIC_SUB_FETCH_CMP_0 (internal_fn, gcall *call) | |
3517 | { | |
3518 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3519 | } | |
3520 | ||
3521 | /* Expand atomic and fetch and cmp with 0. */ | |
3522 | ||
3523 | static void | |
3524 | expand_ATOMIC_AND_FETCH_CMP_0 (internal_fn, gcall *call) | |
3525 | { | |
3526 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3527 | } | |
3528 | ||
3529 | /* Expand atomic or fetch and cmp with 0. */ | |
3530 | ||
3531 | static void | |
3532 | expand_ATOMIC_OR_FETCH_CMP_0 (internal_fn, gcall *call) | |
3533 | { | |
3534 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3535 | } | |
3536 | ||
3537 | /* Expand atomic xor fetch and cmp with 0. */ | |
3538 | ||
3539 | static void | |
3540 | expand_ATOMIC_XOR_FETCH_CMP_0 (internal_fn, gcall *call) | |
3541 | { | |
3542 | expand_ifn_atomic_op_fetch_cmp_0 (call); | |
3543 | } | |
3544 | ||
e16f1cc7 JJ |
3545 | /* Expand LAUNDER to assignment, lhs = arg0. */ |
3546 | ||
3547 | static void | |
3548 | expand_LAUNDER (internal_fn, gcall *call) | |
3549 | { | |
3550 | tree lhs = gimple_call_lhs (call); | |
3551 | ||
3552 | if (!lhs) | |
3553 | return; | |
3554 | ||
3555 | expand_assignment (lhs, gimple_call_arg (call, 0), false); | |
3556 | } | |
3557 | ||
f307441a RS |
3558 | /* Expand {MASK_,}SCATTER_STORE{S,U} call CALL using optab OPTAB. */ |
3559 | ||
3560 | static void | |
3561 | expand_scatter_store_optab_fn (internal_fn, gcall *stmt, direct_optab optab) | |
3562 | { | |
3563 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
3564 | int rhs_index = internal_fn_stored_value_index (ifn); | |
f307441a RS |
3565 | tree base = gimple_call_arg (stmt, 0); |
3566 | tree offset = gimple_call_arg (stmt, 1); | |
3567 | tree scale = gimple_call_arg (stmt, 2); | |
3568 | tree rhs = gimple_call_arg (stmt, rhs_index); | |
3569 | ||
3570 | rtx base_rtx = expand_normal (base); | |
3571 | rtx offset_rtx = expand_normal (offset); | |
3572 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); | |
3573 | rtx rhs_rtx = expand_normal (rhs); | |
3574 | ||
db3efdaf | 3575 | class expand_operand ops[8]; |
f307441a RS |
3576 | int i = 0; |
3577 | create_address_operand (&ops[i++], base_rtx); | |
3578 | create_input_operand (&ops[i++], offset_rtx, TYPE_MODE (TREE_TYPE (offset))); | |
3579 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); | |
3580 | create_integer_operand (&ops[i++], scale_int); | |
3581 | create_input_operand (&ops[i++], rhs_rtx, TYPE_MODE (TREE_TYPE (rhs))); | |
363bb3dc | 3582 | i = add_mask_and_len_args (ops, i, stmt); |
f307441a | 3583 | |
09eb042a RS |
3584 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (TREE_TYPE (rhs)), |
3585 | TYPE_MODE (TREE_TYPE (offset))); | |
f307441a RS |
3586 | expand_insn (icode, i, ops); |
3587 | } | |
3588 | ||
bfaa08b7 RS |
3589 | /* Expand {MASK_,}GATHER_LOAD call CALL using optab OPTAB. */ |
3590 | ||
3591 | static void | |
3592 | expand_gather_load_optab_fn (internal_fn, gcall *stmt, direct_optab optab) | |
3593 | { | |
3594 | tree lhs = gimple_call_lhs (stmt); | |
3595 | tree base = gimple_call_arg (stmt, 0); | |
3596 | tree offset = gimple_call_arg (stmt, 1); | |
3597 | tree scale = gimple_call_arg (stmt, 2); | |
3598 | ||
3599 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3600 | rtx base_rtx = expand_normal (base); | |
3601 | rtx offset_rtx = expand_normal (offset); | |
3602 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); | |
3603 | ||
3604 | int i = 0; | |
db3efdaf | 3605 | class expand_operand ops[8]; |
bfaa08b7 RS |
3606 | create_output_operand (&ops[i++], lhs_rtx, TYPE_MODE (TREE_TYPE (lhs))); |
3607 | create_address_operand (&ops[i++], base_rtx); | |
3608 | create_input_operand (&ops[i++], offset_rtx, TYPE_MODE (TREE_TYPE (offset))); | |
3609 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); | |
3610 | create_integer_operand (&ops[i++], scale_int); | |
363bb3dc | 3611 | i = add_mask_and_len_args (ops, i, stmt); |
09eb042a RS |
3612 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (TREE_TYPE (lhs)), |
3613 | TYPE_MODE (TREE_TYPE (offset))); | |
bfaa08b7 | 3614 | expand_insn (icode, i, ops); |
3af3bec2 RS |
3615 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3616 | emit_move_insn (lhs_rtx, ops[0].value); | |
bfaa08b7 RS |
3617 | } |
3618 | ||
bf510679 JJ |
3619 | /* Helper for expand_DIVMOD. Return true if the sequence starting with |
3620 | INSN contains any call insns or insns with {,U}{DIV,MOD} rtxes. */ | |
3621 | ||
3622 | static bool | |
3623 | contains_call_div_mod (rtx_insn *insn) | |
3624 | { | |
3625 | subrtx_iterator::array_type array; | |
3626 | for (; insn; insn = NEXT_INSN (insn)) | |
3627 | if (CALL_P (insn)) | |
3628 | return true; | |
3629 | else if (INSN_P (insn)) | |
3630 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
3631 | switch (GET_CODE (*iter)) | |
3632 | { | |
3633 | case CALL: | |
3634 | case DIV: | |
3635 | case UDIV: | |
3636 | case MOD: | |
3637 | case UMOD: | |
3638 | return true; | |
3639 | default: | |
3640 | break; | |
3641 | } | |
3642 | return false; | |
3643 | } | |
3644 | ||
e72531b9 PK |
3645 | /* Expand DIVMOD() using: |
3646 | a) optab handler for udivmod/sdivmod if it is available. | |
3647 | b) If optab_handler doesn't exist, generate call to | |
3648 | target-specific divmod libfunc. */ | |
3649 | ||
3650 | static void | |
3651 | expand_DIVMOD (internal_fn, gcall *call_stmt) | |
3652 | { | |
3653 | tree lhs = gimple_call_lhs (call_stmt); | |
3654 | tree arg0 = gimple_call_arg (call_stmt, 0); | |
3655 | tree arg1 = gimple_call_arg (call_stmt, 1); | |
3656 | ||
3657 | gcc_assert (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE); | |
3658 | tree type = TREE_TYPE (TREE_TYPE (lhs)); | |
3659 | machine_mode mode = TYPE_MODE (type); | |
3660 | bool unsignedp = TYPE_UNSIGNED (type); | |
3661 | optab tab = (unsignedp) ? udivmod_optab : sdivmod_optab; | |
3662 | ||
3663 | rtx op0 = expand_normal (arg0); | |
3664 | rtx op1 = expand_normal (arg1); | |
3665 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3666 | ||
bf510679 JJ |
3667 | rtx quotient = NULL_RTX, remainder = NULL_RTX; |
3668 | rtx_insn *insns = NULL; | |
3669 | ||
3670 | if (TREE_CODE (arg1) == INTEGER_CST) | |
3671 | { | |
3672 | /* For DIVMOD by integral constants, there could be efficient code | |
3673 | expanded inline e.g. using shifts and plus/minus. Try to expand | |
3674 | the division and modulo and if it emits any library calls or any | |
3675 | {,U}{DIV,MOD} rtxes throw it away and use a divmod optab or | |
3676 | divmod libcall. */ | |
037fe26e JJ |
3677 | scalar_int_mode int_mode; |
3678 | if (remainder == NULL_RTX | |
3679 | && optimize | |
3680 | && CONST_INT_P (op1) | |
3681 | && !pow2p_hwi (INTVAL (op1)) | |
3682 | && is_int_mode (TYPE_MODE (type), &int_mode) | |
3683 | && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD | |
3684 | && optab_handler (and_optab, word_mode) != CODE_FOR_nothing | |
3685 | && optab_handler (add_optab, word_mode) != CODE_FOR_nothing | |
3686 | && optimize_insn_for_speed_p ()) | |
3687 | { | |
3688 | rtx_insn *last = get_last_insn (); | |
3689 | remainder = NULL_RTX; | |
3690 | quotient = expand_doubleword_divmod (int_mode, op0, op1, &remainder, | |
3691 | TYPE_UNSIGNED (type)); | |
3692 | if (quotient != NULL_RTX) | |
3693 | { | |
3694 | if (optab_handler (mov_optab, int_mode) != CODE_FOR_nothing) | |
3695 | { | |
3696 | rtx_insn *move = emit_move_insn (quotient, quotient); | |
3697 | set_dst_reg_note (move, REG_EQUAL, | |
3698 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) | |
3699 | ? UDIV : DIV, int_mode, | |
3700 | copy_rtx (op0), op1), | |
3701 | quotient); | |
3702 | move = emit_move_insn (remainder, remainder); | |
3703 | set_dst_reg_note (move, REG_EQUAL, | |
3704 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) | |
3705 | ? UMOD : MOD, int_mode, | |
3706 | copy_rtx (op0), op1), | |
3707 | quotient); | |
3708 | } | |
3709 | } | |
3710 | else | |
3711 | delete_insns_since (last); | |
3712 | } | |
3713 | ||
3714 | if (remainder == NULL_RTX) | |
bf510679 | 3715 | { |
037fe26e JJ |
3716 | struct separate_ops ops; |
3717 | ops.code = TRUNC_DIV_EXPR; | |
3718 | ops.type = type; | |
3719 | ops.op0 = make_tree (ops.type, op0); | |
3720 | ops.op1 = arg1; | |
3721 | ops.op2 = NULL_TREE; | |
3722 | ops.location = gimple_location (call_stmt); | |
3723 | start_sequence (); | |
3724 | quotient = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); | |
bf510679 | 3725 | if (contains_call_div_mod (get_insns ())) |
037fe26e JJ |
3726 | quotient = NULL_RTX; |
3727 | else | |
3728 | { | |
3729 | ops.code = TRUNC_MOD_EXPR; | |
3730 | remainder = expand_expr_real_2 (&ops, NULL_RTX, mode, | |
3731 | EXPAND_NORMAL); | |
3732 | if (contains_call_div_mod (get_insns ())) | |
3733 | remainder = NULL_RTX; | |
3734 | } | |
3735 | if (remainder) | |
3736 | insns = get_insns (); | |
3737 | end_sequence (); | |
bf510679 | 3738 | } |
bf510679 JJ |
3739 | } |
3740 | ||
3741 | if (remainder) | |
3742 | emit_insn (insns); | |
e72531b9 PK |
3743 | |
3744 | /* Check if optab_handler exists for divmod_optab for given mode. */ | |
bf510679 | 3745 | else if (optab_handler (tab, mode) != CODE_FOR_nothing) |
e72531b9 PK |
3746 | { |
3747 | quotient = gen_reg_rtx (mode); | |
3748 | remainder = gen_reg_rtx (mode); | |
3749 | expand_twoval_binop (tab, op0, op1, quotient, remainder, unsignedp); | |
3750 | } | |
3751 | ||
3752 | /* Generate call to divmod libfunc if it exists. */ | |
bf510679 | 3753 | else if (rtx libfunc = optab_libfunc (tab, mode)) |
e72531b9 PK |
3754 | targetm.expand_divmod_libfunc (libfunc, mode, op0, op1, |
3755 | "ient, &remainder); | |
3756 | ||
3757 | else | |
3758 | gcc_unreachable (); | |
3759 | ||
3760 | /* Wrap the return value (quotient, remainder) within COMPLEX_EXPR. */ | |
3761 | expand_expr (build2 (COMPLEX_EXPR, TREE_TYPE (lhs), | |
3762 | make_tree (TREE_TYPE (arg0), quotient), | |
3763 | make_tree (TREE_TYPE (arg1), remainder)), | |
0b99f253 | 3764 | target, VOIDmode, EXPAND_NORMAL); |
e72531b9 PK |
3765 | } |
3766 | ||
87a5e0e8 RB |
3767 | /* Expand a NOP. */ |
3768 | ||
3769 | static void | |
3770 | expand_NOP (internal_fn, gcall *) | |
3771 | { | |
3772 | /* Nothing. But it shouldn't really prevail. */ | |
3773 | } | |
3774 | ||
49789fd0 IS |
3775 | /* Coroutines, all should have been processed at this stage. */ |
3776 | ||
3777 | static void | |
3778 | expand_CO_FRAME (internal_fn, gcall *) | |
3779 | { | |
3780 | gcc_unreachable (); | |
3781 | } | |
3782 | ||
3783 | static void | |
3784 | expand_CO_YIELD (internal_fn, gcall *) | |
3785 | { | |
3786 | gcc_unreachable (); | |
3787 | } | |
3788 | ||
3789 | static void | |
3790 | expand_CO_SUSPN (internal_fn, gcall *) | |
3791 | { | |
3792 | gcc_unreachable (); | |
3793 | } | |
3794 | ||
3795 | static void | |
3796 | expand_CO_ACTOR (internal_fn, gcall *) | |
3797 | { | |
3798 | gcc_unreachable (); | |
3799 | } | |
3800 | ||
4cfe7a6c RS |
3801 | /* Expand a call to FN using the operands in STMT. FN has a single |
3802 | output operand and NARGS input operands. */ | |
686ee971 RS |
3803 | |
3804 | static void | |
4cfe7a6c RS |
3805 | expand_direct_optab_fn (internal_fn fn, gcall *stmt, direct_optab optab, |
3806 | unsigned int nargs) | |
686ee971 | 3807 | { |
686ee971 RS |
3808 | tree_pair types = direct_internal_fn_types (fn, stmt); |
3809 | insn_code icode = direct_optab_handler (optab, TYPE_MODE (types.first)); | |
1d205dba | 3810 | expand_fn_using_insn (stmt, icode, 1, nargs); |
686ee971 RS |
3811 | } |
3812 | ||
7cfb4d93 RS |
3813 | /* Expand WHILE_ULT call STMT using optab OPTAB. */ |
3814 | ||
3815 | static void | |
3816 | expand_while_optab_fn (internal_fn, gcall *stmt, convert_optab optab) | |
3817 | { | |
48960b68 | 3818 | expand_operand ops[4]; |
7cfb4d93 RS |
3819 | tree rhs_type[2]; |
3820 | ||
3821 | tree lhs = gimple_call_lhs (stmt); | |
3822 | tree lhs_type = TREE_TYPE (lhs); | |
3823 | rtx lhs_rtx = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
3824 | create_output_operand (&ops[0], lhs_rtx, TYPE_MODE (lhs_type)); | |
3825 | ||
3826 | for (unsigned int i = 0; i < 2; ++i) | |
3827 | { | |
3828 | tree rhs = gimple_call_arg (stmt, i); | |
3829 | rhs_type[i] = TREE_TYPE (rhs); | |
3830 | rtx rhs_rtx = expand_normal (rhs); | |
3831 | create_input_operand (&ops[i + 1], rhs_rtx, TYPE_MODE (rhs_type[i])); | |
3832 | } | |
3833 | ||
48960b68 AS |
3834 | int opcnt; |
3835 | if (!VECTOR_MODE_P (TYPE_MODE (lhs_type))) | |
3836 | { | |
3837 | /* When the mask is an integer mode the exact vector length may not | |
3838 | be clear to the backend, so we pass it in operand[3]. | |
3839 | Use the vector in arg2 for the most reliable intended size. */ | |
3840 | tree type = TREE_TYPE (gimple_call_arg (stmt, 2)); | |
3841 | create_integer_operand (&ops[3], TYPE_VECTOR_SUBPARTS (type)); | |
3842 | opcnt = 4; | |
3843 | } | |
3844 | else | |
3845 | /* The mask has a vector type so the length operand is unnecessary. */ | |
3846 | opcnt = 3; | |
3847 | ||
7cfb4d93 RS |
3848 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (rhs_type[0]), |
3849 | TYPE_MODE (lhs_type)); | |
3850 | ||
48960b68 | 3851 | expand_insn (icode, opcnt, ops); |
7cfb4d93 RS |
3852 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3853 | emit_move_insn (lhs_rtx, ops[0].value); | |
3854 | } | |
3855 | ||
00eab0c6 RS |
3856 | /* Expand a call to a convert-like optab using the operands in STMT. |
3857 | FN has a single output operand and NARGS input operands. */ | |
3858 | ||
3859 | static void | |
3860 | expand_convert_optab_fn (internal_fn fn, gcall *stmt, convert_optab optab, | |
3861 | unsigned int nargs) | |
3862 | { | |
3863 | tree_pair types = direct_internal_fn_types (fn, stmt); | |
3864 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (types.first), | |
3865 | TYPE_MODE (types.second)); | |
3866 | expand_fn_using_insn (stmt, icode, 1, nargs); | |
3867 | } | |
3868 | ||
686ee971 RS |
3869 | /* Expanders for optabs that can use expand_direct_optab_fn. */ |
3870 | ||
4cfe7a6c RS |
3871 | #define expand_unary_optab_fn(FN, STMT, OPTAB) \ |
3872 | expand_direct_optab_fn (FN, STMT, OPTAB, 1) | |
686ee971 | 3873 | |
4cfe7a6c RS |
3874 | #define expand_binary_optab_fn(FN, STMT, OPTAB) \ |
3875 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) | |
686ee971 | 3876 | |
c566cc9f RS |
3877 | #define expand_ternary_optab_fn(FN, STMT, OPTAB) \ |
3878 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3879 | ||
bfe1bb57 | 3880 | #define expand_cond_unary_optab_fn(FN, STMT, OPTAB) \ |
9d4ac06e | 3881 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
bfe1bb57 | 3882 | |
0972596e | 3883 | #define expand_cond_binary_optab_fn(FN, STMT, OPTAB) \ |
9d4ac06e | 3884 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
0972596e | 3885 | |
b41d1f6e RS |
3886 | #define expand_cond_ternary_optab_fn(FN, STMT, OPTAB) \ |
3887 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) | |
3888 | ||
6c96d1e4 JZZ |
3889 | #define expand_cond_len_unary_optab_fn(FN, STMT, OPTAB) \ |
3890 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) | |
3891 | ||
3892 | #define expand_cond_len_binary_optab_fn(FN, STMT, OPTAB) \ | |
3893 | expand_direct_optab_fn (FN, STMT, OPTAB, 6) | |
3894 | ||
3895 | #define expand_cond_len_ternary_optab_fn(FN, STMT, OPTAB) \ | |
3896 | expand_direct_optab_fn (FN, STMT, OPTAB, 7) | |
3897 | ||
bb6c2b68 RS |
3898 | #define expand_fold_extract_optab_fn(FN, STMT, OPTAB) \ |
3899 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3900 | ||
b781a135 RS |
3901 | #define expand_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3902 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) | |
3903 | ||
bce29d65 AM |
3904 | #define expand_mask_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3905 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) | |
3906 | ||
ba49332b JZZ |
3907 | #define expand_mask_len_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3908 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) | |
3909 | ||
58c036c8 RS |
3910 | #define expand_check_ptrs_optab_fn(FN, STMT, OPTAB) \ |
3911 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) | |
3912 | ||
00eab0c6 RS |
3913 | /* Expanders for optabs that can use expand_convert_optab_fn. */ |
3914 | ||
3915 | #define expand_unary_convert_optab_fn(FN, STMT, OPTAB) \ | |
3916 | expand_convert_optab_fn (FN, STMT, OPTAB, 1) | |
3917 | ||
ab23f5d9 RS |
3918 | /* RETURN_TYPE and ARGS are a return type and argument list that are |
3919 | in principle compatible with FN (which satisfies direct_internal_fn_p). | |
3920 | Return the types that should be used to determine whether the | |
3921 | target supports FN. */ | |
3922 | ||
3923 | tree_pair | |
3924 | direct_internal_fn_types (internal_fn fn, tree return_type, tree *args) | |
3925 | { | |
3926 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
3927 | tree type0 = (info.type0 < 0 ? return_type : TREE_TYPE (args[info.type0])); | |
3928 | tree type1 = (info.type1 < 0 ? return_type : TREE_TYPE (args[info.type1])); | |
3929 | return tree_pair (type0, type1); | |
3930 | } | |
3931 | ||
3932 | /* CALL is a call whose return type and arguments are in principle | |
3933 | compatible with FN (which satisfies direct_internal_fn_p). Return the | |
3934 | types that should be used to determine whether the target supports FN. */ | |
3935 | ||
3936 | tree_pair | |
3937 | direct_internal_fn_types (internal_fn fn, gcall *call) | |
3938 | { | |
3939 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
3940 | tree op0 = (info.type0 < 0 | |
3941 | ? gimple_call_lhs (call) | |
3942 | : gimple_call_arg (call, info.type0)); | |
3943 | tree op1 = (info.type1 < 0 | |
3944 | ? gimple_call_lhs (call) | |
3945 | : gimple_call_arg (call, info.type1)); | |
3946 | return tree_pair (TREE_TYPE (op0), TREE_TYPE (op1)); | |
3947 | } | |
3948 | ||
3949 | /* Return true if OPTAB is supported for TYPES (whose modes should be | |
d95ab70a RS |
3950 | the same) when the optimization type is OPT_TYPE. Used for simple |
3951 | direct optabs. */ | |
ab23f5d9 RS |
3952 | |
3953 | static bool | |
d95ab70a RS |
3954 | direct_optab_supported_p (direct_optab optab, tree_pair types, |
3955 | optimization_type opt_type) | |
ab23f5d9 RS |
3956 | { |
3957 | machine_mode mode = TYPE_MODE (types.first); | |
3958 | gcc_checking_assert (mode == TYPE_MODE (types.second)); | |
d95ab70a | 3959 | return direct_optab_handler (optab, mode, opt_type) != CODE_FOR_nothing; |
ab23f5d9 RS |
3960 | } |
3961 | ||
7cfb4d93 RS |
3962 | /* Return true if OPTAB is supported for TYPES, where the first type |
3963 | is the destination and the second type is the source. Used for | |
3964 | convert optabs. */ | |
3965 | ||
3966 | static bool | |
3967 | convert_optab_supported_p (convert_optab optab, tree_pair types, | |
3968 | optimization_type opt_type) | |
3969 | { | |
3970 | return (convert_optab_handler (optab, TYPE_MODE (types.first), | |
3971 | TYPE_MODE (types.second), opt_type) | |
3972 | != CODE_FOR_nothing); | |
3973 | } | |
3974 | ||
ab23f5d9 | 3975 | /* Return true if load/store lanes optab OPTAB is supported for |
d95ab70a | 3976 | array type TYPES.first when the optimization type is OPT_TYPE. */ |
ab23f5d9 RS |
3977 | |
3978 | static bool | |
d95ab70a RS |
3979 | multi_vector_optab_supported_p (convert_optab optab, tree_pair types, |
3980 | optimization_type opt_type) | |
ab23f5d9 | 3981 | { |
d95ab70a RS |
3982 | gcc_assert (TREE_CODE (types.first) == ARRAY_TYPE); |
3983 | machine_mode imode = TYPE_MODE (types.first); | |
3984 | machine_mode vmode = TYPE_MODE (TREE_TYPE (types.first)); | |
3985 | return (convert_optab_handler (optab, imode, vmode, opt_type) | |
3986 | != CODE_FOR_nothing); | |
ab23f5d9 RS |
3987 | } |
3988 | ||
686ee971 | 3989 | #define direct_unary_optab_supported_p direct_optab_supported_p |
00eab0c6 | 3990 | #define direct_unary_convert_optab_supported_p convert_optab_supported_p |
686ee971 | 3991 | #define direct_binary_optab_supported_p direct_optab_supported_p |
c566cc9f | 3992 | #define direct_ternary_optab_supported_p direct_optab_supported_p |
bfe1bb57 | 3993 | #define direct_cond_unary_optab_supported_p direct_optab_supported_p |
0972596e | 3994 | #define direct_cond_binary_optab_supported_p direct_optab_supported_p |
b41d1f6e | 3995 | #define direct_cond_ternary_optab_supported_p direct_optab_supported_p |
6c96d1e4 JZZ |
3996 | #define direct_cond_len_unary_optab_supported_p direct_optab_supported_p |
3997 | #define direct_cond_len_binary_optab_supported_p direct_optab_supported_p | |
3998 | #define direct_cond_len_ternary_optab_supported_p direct_optab_supported_p | |
ef8d1da1 | 3999 | #define direct_mask_load_optab_supported_p convert_optab_supported_p |
ab23f5d9 | 4000 | #define direct_load_lanes_optab_supported_p multi_vector_optab_supported_p |
7e11fc7f | 4001 | #define direct_mask_load_lanes_optab_supported_p multi_vector_optab_supported_p |
09eb042a | 4002 | #define direct_gather_load_optab_supported_p convert_optab_supported_p |
d496134a | 4003 | #define direct_len_load_optab_supported_p direct_optab_supported_p |
bd68b33f | 4004 | #define direct_mask_len_load_optab_supported_p convert_optab_supported_p |
ef8d1da1 | 4005 | #define direct_mask_store_optab_supported_p convert_optab_supported_p |
ab23f5d9 | 4006 | #define direct_store_lanes_optab_supported_p multi_vector_optab_supported_p |
7e11fc7f | 4007 | #define direct_mask_store_lanes_optab_supported_p multi_vector_optab_supported_p |
298e76e6 RS |
4008 | #define direct_vec_cond_mask_optab_supported_p convert_optab_supported_p |
4009 | #define direct_vec_cond_optab_supported_p convert_optab_supported_p | |
09eb042a | 4010 | #define direct_scatter_store_optab_supported_p convert_optab_supported_p |
d496134a | 4011 | #define direct_len_store_optab_supported_p direct_optab_supported_p |
bd68b33f | 4012 | #define direct_mask_len_store_optab_supported_p convert_optab_supported_p |
7cfb4d93 | 4013 | #define direct_while_optab_supported_p convert_optab_supported_p |
bb6c2b68 | 4014 | #define direct_fold_extract_optab_supported_p direct_optab_supported_p |
b781a135 | 4015 | #define direct_fold_left_optab_supported_p direct_optab_supported_p |
bce29d65 | 4016 | #define direct_mask_fold_left_optab_supported_p direct_optab_supported_p |
ba49332b | 4017 | #define direct_mask_len_fold_left_optab_supported_p direct_optab_supported_p |
58c036c8 | 4018 | #define direct_check_ptrs_optab_supported_p direct_optab_supported_p |
683e55fa | 4019 | #define direct_vec_set_optab_supported_p direct_optab_supported_p |
c30efd8c | 4020 | #define direct_vec_extract_optab_supported_p direct_optab_supported_p |
ab23f5d9 | 4021 | |
16d24520 RS |
4022 | /* Return the optab used by internal function FN. */ |
4023 | ||
2f482a07 | 4024 | optab |
16d24520 RS |
4025 | direct_internal_fn_optab (internal_fn fn, tree_pair types) |
4026 | { | |
4027 | switch (fn) | |
4028 | { | |
4029 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
4030 | case IFN_##CODE: break; | |
4031 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
4032 | case IFN_##CODE: return OPTAB##_optab; | |
4033 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ | |
4034 | UNSIGNED_OPTAB, TYPE) \ | |
4035 | case IFN_##CODE: return (TYPE_UNSIGNED (types.SELECTOR) \ | |
4036 | ? UNSIGNED_OPTAB ## _optab \ | |
4037 | : SIGNED_OPTAB ## _optab); | |
4038 | #include "internal-fn.def" | |
4039 | ||
4040 | case IFN_LAST: | |
4041 | break; | |
4042 | } | |
4043 | gcc_unreachable (); | |
4044 | } | |
4045 | ||
bfaa08b7 RS |
4046 | /* Return the optab used by internal function FN. */ |
4047 | ||
4048 | static optab | |
4049 | direct_internal_fn_optab (internal_fn fn) | |
4050 | { | |
4051 | switch (fn) | |
4052 | { | |
4053 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
4054 | case IFN_##CODE: break; | |
4055 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
4056 | case IFN_##CODE: return OPTAB##_optab; | |
4057 | #include "internal-fn.def" | |
4058 | ||
4059 | case IFN_LAST: | |
4060 | break; | |
4061 | } | |
4062 | gcc_unreachable (); | |
4063 | } | |
4064 | ||
d95ab70a RS |
4065 | /* Return true if FN is supported for the types in TYPES when the |
4066 | optimization type is OPT_TYPE. The types are those associated with | |
4067 | the "type0" and "type1" fields of FN's direct_internal_fn_info | |
4068 | structure. */ | |
ab23f5d9 RS |
4069 | |
4070 | bool | |
d95ab70a RS |
4071 | direct_internal_fn_supported_p (internal_fn fn, tree_pair types, |
4072 | optimization_type opt_type) | |
ab23f5d9 RS |
4073 | { |
4074 | switch (fn) | |
4075 | { | |
4076 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ | |
4077 | case IFN_##CODE: break; | |
4078 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ | |
4079 | case IFN_##CODE: \ | |
d95ab70a RS |
4080 | return direct_##TYPE##_optab_supported_p (OPTAB##_optab, types, \ |
4081 | opt_type); | |
16d24520 RS |
4082 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4083 | UNSIGNED_OPTAB, TYPE) \ | |
4084 | case IFN_##CODE: \ | |
4085 | { \ | |
4086 | optab which_optab = (TYPE_UNSIGNED (types.SELECTOR) \ | |
4087 | ? UNSIGNED_OPTAB ## _optab \ | |
4088 | : SIGNED_OPTAB ## _optab); \ | |
4089 | return direct_##TYPE##_optab_supported_p (which_optab, types, \ | |
4090 | opt_type); \ | |
4091 | } | |
ab23f5d9 RS |
4092 | #include "internal-fn.def" |
4093 | ||
4094 | case IFN_LAST: | |
4095 | break; | |
4096 | } | |
4097 | gcc_unreachable (); | |
4098 | } | |
4099 | ||
d95ab70a RS |
4100 | /* Return true if FN is supported for type TYPE when the optimization |
4101 | type is OPT_TYPE. The caller knows that the "type0" and "type1" | |
4102 | fields of FN's direct_internal_fn_info structure are the same. */ | |
ab23f5d9 RS |
4103 | |
4104 | bool | |
d95ab70a RS |
4105 | direct_internal_fn_supported_p (internal_fn fn, tree type, |
4106 | optimization_type opt_type) | |
ab23f5d9 RS |
4107 | { |
4108 | const direct_internal_fn_info &info = direct_internal_fn (fn); | |
4109 | gcc_checking_assert (info.type0 == info.type1); | |
d95ab70a | 4110 | return direct_internal_fn_supported_p (fn, tree_pair (type, type), opt_type); |
ab23f5d9 RS |
4111 | } |
4112 | ||
41241199 RL |
4113 | /* Return true if the STMT is supported when the optimization type is OPT_TYPE, |
4114 | given that STMT is a call to a direct internal function. */ | |
4115 | ||
4116 | bool | |
4117 | direct_internal_fn_supported_p (gcall *stmt, optimization_type opt_type) | |
4118 | { | |
4119 | internal_fn fn = gimple_call_internal_fn (stmt); | |
4120 | tree_pair types = direct_internal_fn_types (fn, stmt); | |
4121 | return direct_internal_fn_supported_p (fn, types, opt_type); | |
4122 | } | |
4123 | ||
30213ae9 | 4124 | /* Return true if FN is a binary operation and if FN is commutative. */ |
0246112a | 4125 | |
30213ae9 RS |
4126 | bool |
4127 | commutative_binary_fn_p (internal_fn fn) | |
0246112a RS |
4128 | { |
4129 | switch (fn) | |
4130 | { | |
0246112a RS |
4131 | case IFN_AVG_FLOOR: |
4132 | case IFN_AVG_CEIL: | |
a1d27560 | 4133 | case IFN_MULH: |
58cc9876 YW |
4134 | case IFN_MULHS: |
4135 | case IFN_MULHRS: | |
0246112a RS |
4136 | case IFN_FMIN: |
4137 | case IFN_FMAX: | |
71207246 | 4138 | case IFN_COMPLEX_MUL: |
12e38012 RS |
4139 | case IFN_UBSAN_CHECK_ADD: |
4140 | case IFN_UBSAN_CHECK_MUL: | |
b1d15146 RS |
4141 | case IFN_ADD_OVERFLOW: |
4142 | case IFN_MUL_OVERFLOW: | |
2f482a07 AV |
4143 | case IFN_VEC_WIDEN_PLUS: |
4144 | case IFN_VEC_WIDEN_PLUS_LO: | |
4145 | case IFN_VEC_WIDEN_PLUS_HI: | |
4146 | case IFN_VEC_WIDEN_PLUS_EVEN: | |
4147 | case IFN_VEC_WIDEN_PLUS_ODD: | |
30213ae9 RS |
4148 | return true; |
4149 | ||
4150 | default: | |
4151 | return false; | |
4152 | } | |
4153 | } | |
4154 | ||
4155 | /* Return true if FN is a ternary operation and if its first two arguments | |
4156 | are commutative. */ | |
4157 | ||
4158 | bool | |
4159 | commutative_ternary_fn_p (internal_fn fn) | |
4160 | { | |
4161 | switch (fn) | |
4162 | { | |
4163 | case IFN_FMA: | |
4164 | case IFN_FMS: | |
4165 | case IFN_FNMA: | |
4166 | case IFN_FNMS: | |
43a3252c | 4167 | case IFN_UADDC: |
30213ae9 RS |
4168 | return true; |
4169 | ||
4170 | default: | |
4171 | return false; | |
4172 | } | |
4173 | } | |
4174 | ||
4175 | /* Return true if FN is an associative binary operation. */ | |
4176 | ||
4177 | bool | |
4178 | associative_binary_fn_p (internal_fn fn) | |
4179 | { | |
4180 | switch (fn) | |
4181 | { | |
4182 | case IFN_FMIN: | |
4183 | case IFN_FMAX: | |
4184 | return true; | |
0246112a | 4185 | |
30213ae9 RS |
4186 | default: |
4187 | return false; | |
4188 | } | |
4189 | } | |
4190 | ||
4191 | /* If FN is commutative in two consecutive arguments, return the | |
4192 | index of the first, otherwise return -1. */ | |
4193 | ||
4194 | int | |
4195 | first_commutative_argument (internal_fn fn) | |
4196 | { | |
4197 | switch (fn) | |
4198 | { | |
0246112a RS |
4199 | case IFN_COND_ADD: |
4200 | case IFN_COND_MUL: | |
4201 | case IFN_COND_MIN: | |
4202 | case IFN_COND_MAX: | |
70613000 RS |
4203 | case IFN_COND_FMIN: |
4204 | case IFN_COND_FMAX: | |
0246112a RS |
4205 | case IFN_COND_AND: |
4206 | case IFN_COND_IOR: | |
4207 | case IFN_COND_XOR: | |
4208 | case IFN_COND_FMA: | |
4209 | case IFN_COND_FMS: | |
4210 | case IFN_COND_FNMA: | |
4211 | case IFN_COND_FNMS: | |
834f3494 JZZ |
4212 | case IFN_COND_LEN_ADD: |
4213 | case IFN_COND_LEN_MUL: | |
4214 | case IFN_COND_LEN_MIN: | |
4215 | case IFN_COND_LEN_MAX: | |
4216 | case IFN_COND_LEN_FMIN: | |
4217 | case IFN_COND_LEN_FMAX: | |
4218 | case IFN_COND_LEN_AND: | |
4219 | case IFN_COND_LEN_IOR: | |
4220 | case IFN_COND_LEN_XOR: | |
4221 | case IFN_COND_LEN_FMA: | |
4222 | case IFN_COND_LEN_FMS: | |
4223 | case IFN_COND_LEN_FNMA: | |
4224 | case IFN_COND_LEN_FNMS: | |
0246112a RS |
4225 | return 1; |
4226 | ||
4227 | default: | |
30213ae9 RS |
4228 | if (commutative_binary_fn_p (fn) |
4229 | || commutative_ternary_fn_p (fn)) | |
4230 | return 0; | |
0246112a RS |
4231 | return -1; |
4232 | } | |
4233 | } | |
4234 | ||
2f482a07 AV |
4235 | /* Return true if this CODE describes an internal_fn that returns a vector with |
4236 | elements twice as wide as the element size of the input vectors. */ | |
4237 | ||
4238 | bool | |
4239 | widening_fn_p (code_helper code) | |
4240 | { | |
4241 | if (!code.is_fn_code ()) | |
4242 | return false; | |
4243 | ||
4244 | if (!internal_fn_p ((combined_fn) code)) | |
4245 | return false; | |
4246 | ||
4247 | internal_fn fn = as_internal_fn ((combined_fn) code); | |
4248 | switch (fn) | |
4249 | { | |
4250 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
4251 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ | |
4252 | case IFN_##NAME: \ | |
4253 | case IFN_##NAME##_HI: \ | |
4254 | case IFN_##NAME##_LO: \ | |
4255 | case IFN_##NAME##_EVEN: \ | |
4256 | case IFN_##NAME##_ODD: \ | |
4257 | return true; | |
4258 | #include "internal-fn.def" | |
4259 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN | |
4260 | ||
4261 | default: | |
4262 | return false; | |
4263 | } | |
4264 | } | |
4265 | ||
883cabde RS |
4266 | /* Return true if IFN_SET_EDOM is supported. */ |
4267 | ||
4268 | bool | |
4269 | set_edom_supported_p (void) | |
4270 | { | |
4271 | #ifdef TARGET_EDOM | |
4272 | return true; | |
4273 | #else | |
4274 | return false; | |
4275 | #endif | |
4276 | } | |
4277 | ||
ab23f5d9 RS |
4278 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4279 | static void \ | |
4cfe7a6c | 4280 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
ab23f5d9 | 4281 | { \ |
4cfe7a6c | 4282 | expand_##TYPE##_optab_fn (fn, stmt, OPTAB##_optab); \ |
ab23f5d9 | 4283 | } |
16d24520 RS |
4284 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4285 | UNSIGNED_OPTAB, TYPE) \ | |
4286 | static void \ | |
4287 | expand_##CODE (internal_fn fn, gcall *stmt) \ | |
4288 | { \ | |
4289 | tree_pair types = direct_internal_fn_types (fn, stmt); \ | |
4290 | optab which_optab = direct_internal_fn_optab (fn, types); \ | |
4291 | expand_##TYPE##_optab_fn (fn, stmt, which_optab); \ | |
4292 | } | |
ab23f5d9 | 4293 | #include "internal-fn.def" |
2f482a07 AV |
4294 | #undef DEF_INTERNAL_OPTAB_FN |
4295 | #undef DEF_INTERNAL_SIGNED_OPTAB_FN | |
ab23f5d9 | 4296 | |
25583c4f RS |
4297 | /* Routines to expand each internal function, indexed by function number. |
4298 | Each routine has the prototype: | |
4299 | ||
538dd0b7 | 4300 | expand_<NAME> (gcall *stmt) |
25583c4f RS |
4301 | |
4302 | where STMT is the statement that performs the call. */ | |
4cfe7a6c | 4303 | static void (*const internal_fn_expanders[]) (internal_fn, gcall *) = { |
2f482a07 | 4304 | |
b78475cf | 4305 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) expand_##CODE, |
25583c4f | 4306 | #include "internal-fn.def" |
25583c4f RS |
4307 | 0 |
4308 | }; | |
4309 | ||
0d4dd7e0 JZZ |
4310 | /* Invoke T(CODE, SUFFIX) for each conditional function IFN_COND_##SUFFIX |
4311 | that maps to a tree code CODE. There is also an IFN_COND_LEN_##SUFFIX | |
4312 | for each such IFN_COND_##SUFFIX. */ | |
6a86928d | 4313 | #define FOR_EACH_CODE_MAPPING(T) \ |
0d4dd7e0 JZZ |
4314 | T (PLUS_EXPR, ADD) \ |
4315 | T (MINUS_EXPR, SUB) \ | |
4316 | T (MULT_EXPR, MUL) \ | |
4317 | T (TRUNC_DIV_EXPR, DIV) \ | |
4318 | T (TRUNC_MOD_EXPR, MOD) \ | |
4319 | T (RDIV_EXPR, RDIV) \ | |
4320 | T (MIN_EXPR, MIN) \ | |
4321 | T (MAX_EXPR, MAX) \ | |
4322 | T (BIT_AND_EXPR, AND) \ | |
4323 | T (BIT_IOR_EXPR, IOR) \ | |
4324 | T (BIT_XOR_EXPR, XOR) \ | |
4325 | T (LSHIFT_EXPR, SHL) \ | |
4326 | T (RSHIFT_EXPR, SHR) \ | |
4327 | T (NEGATE_EXPR, NEG) | |
6a86928d | 4328 | |
9d4ac06e RS |
4329 | /* Return a function that only performs CODE when a certain condition is met |
4330 | and that uses a given fallback value otherwise. For example, if CODE is | |
4331 | a binary operation associated with conditional function FN: | |
4332 | ||
4333 | LHS = FN (COND, A, B, ELSE) | |
4334 | ||
4335 | is equivalent to the C expression: | |
4336 | ||
4337 | LHS = COND ? A CODE B : ELSE; | |
0972596e | 4338 | |
9d4ac06e | 4339 | operating elementwise if the operands are vectors. |
0972596e | 4340 | |
9d4ac06e | 4341 | Return IFN_LAST if no such function exists. */ |
0972596e RS |
4342 | |
4343 | internal_fn | |
4344 | get_conditional_internal_fn (tree_code code) | |
4345 | { | |
4346 | switch (code) | |
4347 | { | |
0d4dd7e0 | 4348 | #define CASE(CODE, IFN) case CODE: return IFN_COND_##IFN; |
6a86928d RS |
4349 | FOR_EACH_CODE_MAPPING(CASE) |
4350 | #undef CASE | |
0972596e RS |
4351 | default: |
4352 | return IFN_LAST; | |
4353 | } | |
4354 | } | |
4355 | ||
6a86928d RS |
4356 | /* If IFN implements the conditional form of a tree code, return that |
4357 | tree code, otherwise return ERROR_MARK. */ | |
4358 | ||
4359 | tree_code | |
4360 | conditional_internal_fn_code (internal_fn ifn) | |
4361 | { | |
4362 | switch (ifn) | |
4363 | { | |
834f3494 JZZ |
4364 | #define CASE(CODE, IFN) \ |
4365 | case IFN_COND_##IFN: \ | |
4366 | case IFN_COND_LEN_##IFN: \ | |
4367 | return CODE; | |
4368 | FOR_EACH_CODE_MAPPING (CASE) | |
6a86928d | 4369 | #undef CASE |
834f3494 JZZ |
4370 | default: |
4371 | return ERROR_MARK; | |
6a86928d RS |
4372 | } |
4373 | } | |
4374 | ||
0d4dd7e0 JZZ |
4375 | /* Like get_conditional_internal_fn, but return a function that |
4376 | additionally restricts the operation to the leading elements | |
4377 | of a vector. The number of elements to process is given by a length | |
4378 | and bias pair, as for IFN_LOAD_LEN. The values of the remaining | |
4379 | elements are taken from the fallback ("else") argument. | |
4380 | ||
4381 | For example, if CODE is a binary operation associated with FN: | |
4382 | ||
4383 | LHS = FN (COND, A, B, ELSE, LEN, BIAS) | |
4384 | ||
4385 | is equivalent to the C code: | |
4386 | ||
4387 | for (int i = 0; i < NUNITS; i++) | |
4388 | { | |
4389 | if (i < LEN + BIAS && COND[i]) | |
4390 | LHS[i] = A[i] CODE B[i]; | |
4391 | else | |
4392 | LHS[i] = ELSE[i]; | |
4393 | } | |
4394 | */ | |
4395 | ||
4396 | internal_fn | |
4397 | get_conditional_len_internal_fn (tree_code code) | |
4398 | { | |
4399 | switch (code) | |
4400 | { | |
4401 | #define CASE(CODE, IFN) case CODE: return IFN_COND_LEN_##IFN; | |
4402 | FOR_EACH_CODE_MAPPING(CASE) | |
4403 | #undef CASE | |
4404 | default: | |
4405 | return IFN_LAST; | |
4406 | } | |
4407 | } | |
4408 | ||
b41d1f6e RS |
4409 | /* Invoke T(IFN) for each internal function IFN that also has an |
4410 | IFN_COND_* form. */ | |
4411 | #define FOR_EACH_COND_FN_PAIR(T) \ | |
70613000 RS |
4412 | T (FMAX) \ |
4413 | T (FMIN) \ | |
b41d1f6e RS |
4414 | T (FMA) \ |
4415 | T (FMS) \ | |
4416 | T (FNMA) \ | |
4417 | T (FNMS) | |
4418 | ||
4419 | /* Return a function that only performs internal function FN when a | |
4420 | certain condition is met and that uses a given fallback value otherwise. | |
4421 | In other words, the returned function FN' is such that: | |
4422 | ||
4423 | LHS = FN' (COND, A1, ... An, ELSE) | |
4424 | ||
4425 | is equivalent to the C expression: | |
4426 | ||
4427 | LHS = COND ? FN (A1, ..., An) : ELSE; | |
4428 | ||
4429 | operating elementwise if the operands are vectors. | |
4430 | ||
4431 | Return IFN_LAST if no such function exists. */ | |
4432 | ||
4433 | internal_fn | |
4434 | get_conditional_internal_fn (internal_fn fn) | |
4435 | { | |
4436 | switch (fn) | |
4437 | { | |
4438 | #define CASE(NAME) case IFN_##NAME: return IFN_COND_##NAME; | |
4439 | FOR_EACH_COND_FN_PAIR(CASE) | |
4440 | #undef CASE | |
4441 | default: | |
4442 | return IFN_LAST; | |
4443 | } | |
4444 | } | |
4445 | ||
02a015fa JZZ |
4446 | /* If there exists an internal function like IFN that operates on vectors, |
4447 | but with additional length and bias parameters, return the internal_fn | |
4448 | for that function, otherwise return IFN_LAST. */ | |
4449 | internal_fn | |
4450 | get_len_internal_fn (internal_fn fn) | |
4451 | { | |
4452 | switch (fn) | |
4453 | { | |
4454 | #undef DEF_INTERNAL_COND_FN | |
4455 | #undef DEF_INTERNAL_SIGNED_COND_FN | |
4456 | #define DEF_INTERNAL_COND_FN(NAME, ...) \ | |
4457 | case IFN_COND_##NAME: \ | |
4458 | return IFN_COND_LEN_##NAME; | |
4459 | #define DEF_INTERNAL_SIGNED_COND_FN(NAME, ...) \ | |
4460 | case IFN_COND_##NAME: \ | |
4461 | return IFN_COND_LEN_##NAME; | |
4462 | #include "internal-fn.def" | |
4463 | #undef DEF_INTERNAL_COND_FN | |
4464 | #undef DEF_INTERNAL_SIGNED_COND_FN | |
4465 | default: | |
4466 | return IFN_LAST; | |
4467 | } | |
4468 | } | |
4469 | ||
b41d1f6e RS |
4470 | /* If IFN implements the conditional form of an unconditional internal |
4471 | function, return that unconditional function, otherwise return IFN_LAST. */ | |
4472 | ||
4473 | internal_fn | |
4474 | get_unconditional_internal_fn (internal_fn ifn) | |
4475 | { | |
4476 | switch (ifn) | |
4477 | { | |
4478 | #define CASE(NAME) case IFN_COND_##NAME: return IFN_##NAME; | |
4479 | FOR_EACH_COND_FN_PAIR(CASE) | |
4480 | #undef CASE | |
4481 | default: | |
4482 | return IFN_LAST; | |
4483 | } | |
4484 | } | |
4485 | ||
0936858f RS |
4486 | /* Return true if STMT can be interpreted as a conditional tree code |
4487 | operation of the form: | |
4488 | ||
4489 | LHS = COND ? OP (RHS1, ...) : ELSE; | |
4490 | ||
4491 | operating elementwise if the operands are vectors. This includes | |
4492 | the case of an all-true COND, so that the operation always happens. | |
4493 | ||
834f3494 JZZ |
4494 | There is an alternative approach to interpret the STMT when the operands |
4495 | are vectors which is the operation predicated by both conditional mask | |
4496 | and loop control length, the equivalent C code: | |
4497 | ||
4498 | for (int i = 0; i < NUNTIS; i++) | |
4499 | { | |
4500 | if (i < LEN + BIAS && COND[i]) | |
4501 | LHS[i] = A[i] CODE B[i]; | |
4502 | else | |
4503 | LHS[i] = ELSE[i]; | |
4504 | } | |
4505 | ||
0936858f RS |
4506 | When returning true, set: |
4507 | ||
4508 | - *COND_OUT to the condition COND, or to NULL_TREE if the condition | |
4509 | is known to be all-true | |
4510 | - *CODE_OUT to the tree code | |
4511 | - OPS[I] to operand I of *CODE_OUT | |
4512 | - *ELSE_OUT to the fallback value ELSE, or to NULL_TREE if the | |
834f3494 JZZ |
4513 | condition is known to be all true. |
4514 | - *LEN to the len argument if it COND_LEN_* operations or to NULL_TREE. | |
4515 | - *BIAS to the bias argument if it COND_LEN_* operations or to NULL_TREE. */ | |
0936858f RS |
4516 | |
4517 | bool | |
4518 | can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out, | |
4519 | tree_code *code_out, | |
834f3494 JZZ |
4520 | tree (&ops)[3], tree *else_out, |
4521 | tree *len, tree *bias) | |
0936858f | 4522 | { |
834f3494 JZZ |
4523 | *len = NULL_TREE; |
4524 | *bias = NULL_TREE; | |
0936858f RS |
4525 | if (gassign *assign = dyn_cast <gassign *> (stmt)) |
4526 | { | |
4527 | *cond_out = NULL_TREE; | |
4528 | *code_out = gimple_assign_rhs_code (assign); | |
4529 | ops[0] = gimple_assign_rhs1 (assign); | |
4530 | ops[1] = gimple_assign_rhs2 (assign); | |
4531 | ops[2] = gimple_assign_rhs3 (assign); | |
4532 | *else_out = NULL_TREE; | |
4533 | return true; | |
4534 | } | |
4535 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
4536 | if (gimple_call_internal_p (call)) | |
4537 | { | |
4538 | internal_fn ifn = gimple_call_internal_fn (call); | |
4539 | tree_code code = conditional_internal_fn_code (ifn); | |
834f3494 JZZ |
4540 | int len_index = internal_fn_len_index (ifn); |
4541 | int cond_nargs = len_index >= 0 ? 4 : 2; | |
0936858f RS |
4542 | if (code != ERROR_MARK) |
4543 | { | |
4544 | *cond_out = gimple_call_arg (call, 0); | |
4545 | *code_out = code; | |
834f3494 | 4546 | unsigned int nops = gimple_call_num_args (call) - cond_nargs; |
0936858f RS |
4547 | for (unsigned int i = 0; i < 3; ++i) |
4548 | ops[i] = i < nops ? gimple_call_arg (call, i + 1) : NULL_TREE; | |
4549 | *else_out = gimple_call_arg (call, nops + 1); | |
834f3494 | 4550 | if (len_index < 0) |
0936858f | 4551 | { |
834f3494 JZZ |
4552 | if (integer_truep (*cond_out)) |
4553 | { | |
4554 | *cond_out = NULL_TREE; | |
4555 | *else_out = NULL_TREE; | |
4556 | } | |
4557 | } | |
4558 | else | |
4559 | { | |
4560 | *len = gimple_call_arg (call, len_index); | |
4561 | *bias = gimple_call_arg (call, len_index + 1); | |
0936858f RS |
4562 | } |
4563 | return true; | |
4564 | } | |
4565 | } | |
4566 | return false; | |
4567 | } | |
4568 | ||
bfaa08b7 RS |
4569 | /* Return true if IFN is some form of load from memory. */ |
4570 | ||
4571 | bool | |
4572 | internal_load_fn_p (internal_fn fn) | |
4573 | { | |
4574 | switch (fn) | |
4575 | { | |
4576 | case IFN_MASK_LOAD: | |
4577 | case IFN_LOAD_LANES: | |
4578 | case IFN_MASK_LOAD_LANES: | |
4579 | case IFN_GATHER_LOAD: | |
4580 | case IFN_MASK_GATHER_LOAD: | |
bd68b33f | 4581 | case IFN_MASK_LEN_GATHER_LOAD: |
d496134a | 4582 | case IFN_LEN_LOAD: |
bd68b33f | 4583 | case IFN_MASK_LEN_LOAD: |
bfaa08b7 RS |
4584 | return true; |
4585 | ||
4586 | default: | |
4587 | return false; | |
4588 | } | |
4589 | } | |
4590 | ||
f307441a RS |
4591 | /* Return true if IFN is some form of store to memory. */ |
4592 | ||
4593 | bool | |
4594 | internal_store_fn_p (internal_fn fn) | |
4595 | { | |
4596 | switch (fn) | |
4597 | { | |
4598 | case IFN_MASK_STORE: | |
4599 | case IFN_STORE_LANES: | |
4600 | case IFN_MASK_STORE_LANES: | |
4601 | case IFN_SCATTER_STORE: | |
4602 | case IFN_MASK_SCATTER_STORE: | |
bd68b33f | 4603 | case IFN_MASK_LEN_SCATTER_STORE: |
d496134a | 4604 | case IFN_LEN_STORE: |
bd68b33f | 4605 | case IFN_MASK_LEN_STORE: |
f307441a RS |
4606 | return true; |
4607 | ||
4608 | default: | |
4609 | return false; | |
4610 | } | |
4611 | } | |
4612 | ||
bfaa08b7 RS |
4613 | /* Return true if IFN is some form of gather load or scatter store. */ |
4614 | ||
4615 | bool | |
4616 | internal_gather_scatter_fn_p (internal_fn fn) | |
4617 | { | |
4618 | switch (fn) | |
4619 | { | |
4620 | case IFN_GATHER_LOAD: | |
4621 | case IFN_MASK_GATHER_LOAD: | |
bd68b33f | 4622 | case IFN_MASK_LEN_GATHER_LOAD: |
f307441a RS |
4623 | case IFN_SCATTER_STORE: |
4624 | case IFN_MASK_SCATTER_STORE: | |
bd68b33f | 4625 | case IFN_MASK_LEN_SCATTER_STORE: |
bfaa08b7 RS |
4626 | return true; |
4627 | ||
4628 | default: | |
4629 | return false; | |
4630 | } | |
4631 | } | |
4632 | ||
b8806f6f JZZ |
4633 | /* If FN takes a vector len argument, return the index of that argument, |
4634 | otherwise return -1. */ | |
4635 | ||
4636 | int | |
4637 | internal_fn_len_index (internal_fn fn) | |
4638 | { | |
4639 | switch (fn) | |
4640 | { | |
4641 | case IFN_LEN_LOAD: | |
4642 | case IFN_LEN_STORE: | |
b8806f6f JZZ |
4643 | return 2; |
4644 | ||
bd68b33f JZ |
4645 | case IFN_MASK_LEN_GATHER_LOAD: |
4646 | case IFN_MASK_LEN_SCATTER_STORE: | |
834f3494 JZZ |
4647 | case IFN_COND_LEN_FMA: |
4648 | case IFN_COND_LEN_FMS: | |
4649 | case IFN_COND_LEN_FNMA: | |
4650 | case IFN_COND_LEN_FNMS: | |
34c614b7 | 4651 | return 5; |
db3efdaf | 4652 | |
834f3494 JZZ |
4653 | case IFN_COND_LEN_ADD: |
4654 | case IFN_COND_LEN_SUB: | |
4655 | case IFN_COND_LEN_MUL: | |
4656 | case IFN_COND_LEN_DIV: | |
4657 | case IFN_COND_LEN_MOD: | |
4658 | case IFN_COND_LEN_RDIV: | |
4659 | case IFN_COND_LEN_MIN: | |
4660 | case IFN_COND_LEN_MAX: | |
4661 | case IFN_COND_LEN_FMIN: | |
4662 | case IFN_COND_LEN_FMAX: | |
4663 | case IFN_COND_LEN_AND: | |
4664 | case IFN_COND_LEN_IOR: | |
4665 | case IFN_COND_LEN_XOR: | |
4666 | case IFN_COND_LEN_SHL: | |
4667 | case IFN_COND_LEN_SHR: | |
4668 | return 4; | |
4669 | ||
4670 | case IFN_COND_LEN_NEG: | |
363bb3dc JZ |
4671 | case IFN_MASK_LEN_LOAD: |
4672 | case IFN_MASK_LEN_STORE: | |
834f3494 JZZ |
4673 | return 3; |
4674 | ||
b8806f6f JZZ |
4675 | default: |
4676 | return -1; | |
4677 | } | |
4678 | } | |
4679 | ||
bfaa08b7 RS |
4680 | /* If FN takes a vector mask argument, return the index of that argument, |
4681 | otherwise return -1. */ | |
4682 | ||
4683 | int | |
4684 | internal_fn_mask_index (internal_fn fn) | |
4685 | { | |
4686 | switch (fn) | |
4687 | { | |
4688 | case IFN_MASK_LOAD: | |
4689 | case IFN_MASK_LOAD_LANES: | |
4690 | case IFN_MASK_STORE: | |
4691 | case IFN_MASK_STORE_LANES: | |
363bb3dc JZ |
4692 | case IFN_MASK_LEN_LOAD: |
4693 | case IFN_MASK_LEN_STORE: | |
bfaa08b7 RS |
4694 | return 2; |
4695 | ||
4696 | case IFN_MASK_GATHER_LOAD: | |
f307441a | 4697 | case IFN_MASK_SCATTER_STORE: |
bd68b33f JZ |
4698 | case IFN_MASK_LEN_GATHER_LOAD: |
4699 | case IFN_MASK_LEN_SCATTER_STORE: | |
34c614b7 | 4700 | return 4; |
db3efdaf | 4701 | |
f307441a | 4702 | default: |
2c58d42c RS |
4703 | return (conditional_internal_fn_code (fn) != ERROR_MARK |
4704 | || get_unconditional_internal_fn (fn) != IFN_LAST ? 0 : -1); | |
f307441a RS |
4705 | } |
4706 | } | |
4707 | ||
4708 | /* If FN takes a value that should be stored to memory, return the index | |
4709 | of that argument, otherwise return -1. */ | |
4710 | ||
4711 | int | |
4712 | internal_fn_stored_value_index (internal_fn fn) | |
4713 | { | |
4714 | switch (fn) | |
4715 | { | |
4716 | case IFN_MASK_STORE: | |
bd28b730 | 4717 | case IFN_MASK_STORE_LANES: |
f307441a RS |
4718 | case IFN_SCATTER_STORE: |
4719 | case IFN_MASK_SCATTER_STORE: | |
bd68b33f | 4720 | case IFN_MASK_LEN_SCATTER_STORE: |
f307441a RS |
4721 | return 3; |
4722 | ||
b8806f6f | 4723 | case IFN_LEN_STORE: |
d39f4889 JZZ |
4724 | return 4; |
4725 | ||
bd68b33f | 4726 | case IFN_MASK_LEN_STORE: |
b8806f6f JZZ |
4727 | return 5; |
4728 | ||
bfaa08b7 RS |
4729 | default: |
4730 | return -1; | |
4731 | } | |
4732 | } | |
4733 | ||
4734 | /* Return true if the target supports gather load or scatter store function | |
4735 | IFN. For loads, VECTOR_TYPE is the vector type of the load result, | |
4736 | while for stores it is the vector type of the stored data argument. | |
4737 | MEMORY_ELEMENT_TYPE is the type of the memory elements being loaded | |
09eb042a RS |
4738 | or stored. OFFSET_VECTOR_TYPE is the vector type that holds the |
4739 | offset from the shared base address of each loaded or stored element. | |
4740 | SCALE is the amount by which these offsets should be multiplied | |
4741 | *after* they have been extended to address width. */ | |
bfaa08b7 RS |
4742 | |
4743 | bool | |
4744 | internal_gather_scatter_fn_supported_p (internal_fn ifn, tree vector_type, | |
4745 | tree memory_element_type, | |
09eb042a | 4746 | tree offset_vector_type, int scale) |
bfaa08b7 RS |
4747 | { |
4748 | if (!tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vector_type)), | |
4749 | TYPE_SIZE (memory_element_type))) | |
4750 | return false; | |
09eb042a RS |
4751 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vector_type), |
4752 | TYPE_VECTOR_SUBPARTS (offset_vector_type))) | |
4753 | return false; | |
bfaa08b7 | 4754 | optab optab = direct_internal_fn_optab (ifn); |
09eb042a RS |
4755 | insn_code icode = convert_optab_handler (optab, TYPE_MODE (vector_type), |
4756 | TYPE_MODE (offset_vector_type)); | |
f307441a | 4757 | int output_ops = internal_load_fn_p (ifn) ? 1 : 0; |
09eb042a | 4758 | bool unsigned_p = TYPE_UNSIGNED (TREE_TYPE (offset_vector_type)); |
bfaa08b7 | 4759 | return (icode != CODE_FOR_nothing |
09eb042a RS |
4760 | && insn_operand_matches (icode, 2 + output_ops, GEN_INT (unsigned_p)) |
4761 | && insn_operand_matches (icode, 3 + output_ops, GEN_INT (scale))); | |
bfaa08b7 RS |
4762 | } |
4763 | ||
58c036c8 RS |
4764 | /* Return true if the target supports IFN_CHECK_{RAW,WAR}_PTRS function IFN |
4765 | for pointers of type TYPE when the accesses have LENGTH bytes and their | |
4766 | common byte alignment is ALIGN. */ | |
4767 | ||
4768 | bool | |
4769 | internal_check_ptrs_fn_supported_p (internal_fn ifn, tree type, | |
4770 | poly_uint64 length, unsigned int align) | |
4771 | { | |
4772 | machine_mode mode = TYPE_MODE (type); | |
4773 | optab optab = direct_internal_fn_optab (ifn); | |
4774 | insn_code icode = direct_optab_handler (optab, mode); | |
4775 | if (icode == CODE_FOR_nothing) | |
4776 | return false; | |
4777 | rtx length_rtx = immed_wide_int_const (length, mode); | |
4778 | return (insn_operand_matches (icode, 3, length_rtx) | |
4779 | && insn_operand_matches (icode, 4, GEN_INT (align))); | |
4780 | } | |
4781 | ||
363bb3dc JZ |
4782 | /* Return the supported bias for IFN which is either IFN_{LEN_,MASK_LEN_,}LOAD |
4783 | or IFN_{LEN_,MASK_LEN_,}STORE. For now we only support the biases of 0 and | |
4784 | -1 (in case 0 is not an allowable length for {len_,mask_len_}load or | |
4785 | {len_,mask_len_}store). If none of the biases match what the backend | |
4786 | provides, return VECT_PARTIAL_BIAS_UNSUPPORTED. */ | |
b0e51639 RD |
4787 | |
4788 | signed char | |
4789 | internal_len_load_store_bias (internal_fn ifn, machine_mode mode) | |
4790 | { | |
4791 | optab optab = direct_internal_fn_optab (ifn); | |
4792 | insn_code icode = direct_optab_handler (optab, mode); | |
363bb3dc | 4793 | int bias_no = 3; |
d39f4889 JZZ |
4794 | |
4795 | if (icode == CODE_FOR_nothing) | |
4796 | { | |
4797 | machine_mode mask_mode; | |
4798 | if (!targetm.vectorize.get_mask_mode (mode).exists (&mask_mode)) | |
4799 | return VECT_PARTIAL_BIAS_UNSUPPORTED; | |
4800 | if (ifn == IFN_LEN_LOAD) | |
4801 | { | |
bd68b33f JZ |
4802 | /* Try MASK_LEN_LOAD. */ |
4803 | optab = direct_internal_fn_optab (IFN_MASK_LEN_LOAD); | |
d39f4889 JZZ |
4804 | } |
4805 | else | |
4806 | { | |
bd68b33f JZ |
4807 | /* Try MASK_LEN_STORE. */ |
4808 | optab = direct_internal_fn_optab (IFN_MASK_LEN_STORE); | |
d39f4889 JZZ |
4809 | } |
4810 | icode = convert_optab_handler (optab, mode, mask_mode); | |
363bb3dc | 4811 | bias_no = 4; |
d39f4889 | 4812 | } |
b0e51639 RD |
4813 | |
4814 | if (icode != CODE_FOR_nothing) | |
4815 | { | |
4816 | /* For now we only support biases of 0 or -1. Try both of them. */ | |
363bb3dc | 4817 | if (insn_operand_matches (icode, bias_no, GEN_INT (0))) |
b0e51639 | 4818 | return 0; |
363bb3dc | 4819 | if (insn_operand_matches (icode, bias_no, GEN_INT (-1))) |
b0e51639 RD |
4820 | return -1; |
4821 | } | |
4822 | ||
4823 | return VECT_PARTIAL_BIAS_UNSUPPORTED; | |
4824 | } | |
4825 | ||
4cfe7a6c RS |
4826 | /* Expand STMT as though it were a call to internal function FN. */ |
4827 | ||
4828 | void | |
4829 | expand_internal_call (internal_fn fn, gcall *stmt) | |
4830 | { | |
4831 | internal_fn_expanders[fn] (fn, stmt); | |
4832 | } | |
4833 | ||
25583c4f RS |
4834 | /* Expand STMT, which is a call to internal function FN. */ |
4835 | ||
4836 | void | |
538dd0b7 | 4837 | expand_internal_call (gcall *stmt) |
25583c4f | 4838 | { |
4cfe7a6c | 4839 | expand_internal_call (gimple_call_internal_fn (stmt), stmt); |
25583c4f | 4840 | } |
1ee62b92 | 4841 | |
2c58d42c RS |
4842 | /* If TYPE is a vector type, return true if IFN is a direct internal |
4843 | function that is supported for that type. If TYPE is a scalar type, | |
4844 | return true if IFN is a direct internal function that is supported for | |
4845 | the target's preferred vector version of TYPE. */ | |
4846 | ||
4847 | bool | |
4848 | vectorized_internal_fn_supported_p (internal_fn ifn, tree type) | |
4849 | { | |
0ae469e8 RS |
4850 | if (VECTOR_MODE_P (TYPE_MODE (type))) |
4851 | return direct_internal_fn_supported_p (ifn, type, OPTIMIZE_FOR_SPEED); | |
4852 | ||
2c58d42c | 4853 | scalar_mode smode; |
72f8d228 RB |
4854 | if (VECTOR_TYPE_P (type) |
4855 | || !is_a <scalar_mode> (TYPE_MODE (type), &smode)) | |
0ae469e8 RS |
4856 | return false; |
4857 | ||
4858 | machine_mode vmode = targetm.vectorize.preferred_simd_mode (smode); | |
4859 | if (VECTOR_MODE_P (vmode)) | |
2c58d42c | 4860 | { |
0ae469e8 RS |
4861 | tree vectype = build_vector_type_for_mode (type, vmode); |
4862 | if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED)) | |
4863 | return true; | |
2c58d42c RS |
4864 | } |
4865 | ||
0ae469e8 RS |
4866 | auto_vector_modes vector_modes; |
4867 | targetm.vectorize.autovectorize_vector_modes (&vector_modes, true); | |
4868 | for (machine_mode base_mode : vector_modes) | |
4869 | if (related_vector_mode (base_mode, smode).exists (&vmode)) | |
4870 | { | |
4871 | tree vectype = build_vector_type_for_mode (type, vmode); | |
4872 | if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED)) | |
4873 | return true; | |
4874 | } | |
4875 | ||
4876 | return false; | |
2c58d42c RS |
4877 | } |
4878 | ||
ef8176e0 RB |
4879 | void |
4880 | expand_SHUFFLEVECTOR (internal_fn, gcall *) | |
4881 | { | |
4882 | gcc_unreachable (); | |
4883 | } | |
4884 | ||
1ee62b92 PG |
4885 | void |
4886 | expand_PHI (internal_fn, gcall *) | |
4887 | { | |
463d9108 JJ |
4888 | gcc_unreachable (); |
4889 | } | |
4890 | ||
4891 | void | |
4892 | expand_SPACESHIP (internal_fn, gcall *stmt) | |
4893 | { | |
4894 | tree lhs = gimple_call_lhs (stmt); | |
4895 | tree rhs1 = gimple_call_arg (stmt, 0); | |
4896 | tree rhs2 = gimple_call_arg (stmt, 1); | |
4897 | tree type = TREE_TYPE (rhs1); | |
4898 | ||
526fbcfa JJ |
4899 | do_pending_stack_adjust (); |
4900 | ||
463d9108 JJ |
4901 | rtx target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); |
4902 | rtx op1 = expand_normal (rhs1); | |
4903 | rtx op2 = expand_normal (rhs2); | |
4904 | ||
4905 | class expand_operand ops[3]; | |
4906 | create_output_operand (&ops[0], target, TYPE_MODE (TREE_TYPE (lhs))); | |
4907 | create_input_operand (&ops[1], op1, TYPE_MODE (type)); | |
4908 | create_input_operand (&ops[2], op2, TYPE_MODE (type)); | |
4909 | insn_code icode = optab_handler (spaceship_optab, TYPE_MODE (type)); | |
4910 | expand_insn (icode, 3, ops); | |
4911 | if (!rtx_equal_p (target, ops[0].value)) | |
4912 | emit_move_insn (target, ops[0].value); | |
1ee62b92 | 4913 | } |
f64eb636 | 4914 | |
08b51bad JJ |
4915 | void |
4916 | expand_ASSUME (internal_fn, gcall *) | |
4917 | { | |
08b51bad | 4918 | } |
3da77f21 AS |
4919 | |
4920 | void | |
4921 | expand_MASK_CALL (internal_fn, gcall *) | |
4922 | { | |
4923 | /* This IFN should only exist between ifcvt and vect passes. */ | |
4924 | gcc_unreachable (); | |
4925 | } |