1 | /* Internal functions. |
2 | Copyright (C) 2011-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it 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" |
23 | #include "backend.h" |
24 | #include "target.h" |
25 | #include "rtl.h" |
26 | #include "tree.h" |
27 | #include "gimple.h" |
28 | #include "predict.h" |
29 | #include "stringpool.h" |
30 | #include "tree-vrp.h" |
31 | #include "tree-ssanames.h" |
32 | #include "expmed.h" |
33 | #include "memmodel.h" |
34 | #include "optabs.h" |
35 | #include "emit-rtl.h" |
36 | #include "diagnostic-core.h" |
37 | #include "fold-const.h" |
38 | #include "internal-fn.h" |
39 | #include "stor-layout.h" |
40 | #include "dojump.h" |
41 | #include "expr.h" |
42 | #include "stringpool.h" |
43 | #include "attribs.h" |
44 | #include "asan.h" |
45 | #include "ubsan.h" |
46 | #include "recog.h" |
47 | #include "builtins.h" |
48 | #include "optabs-tree.h" |
49 | #include "gimple-ssa.h" |
50 | #include "tree-phinodes.h" |
51 | #include "ssa-iterators.h" |
52 | #include "explow.h" |
53 | #include "rtl-iter.h" |
54 | #include "gimple-range.h" |
55 | |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" |
58 | |
59 | /* The names of each internal function, indexed by function number. */ |
60 | const char *const internal_fn_name_array[] = { |
61 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) #CODE, |
62 | #include "internal-fn.def" |
63 | "<invalid-fn>" |
64 | }; |
65 | |
66 | /* The ECF_* flags of each internal function, indexed by function number. */ |
67 | const int internal_fn_flags_array[] = { |
68 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) FLAGS, |
69 | #include "internal-fn.def" |
70 | 0 |
71 | }; |
72 | |
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 (k: internal_fn_name (fn: internal_fn (i)), |
87 | v: internal_fn (i)); |
88 | } |
89 | internal_fn *entry = name_to_fn_map->get (k: name); |
90 | return entry ? *entry : IFN_LAST; |
91 | } |
92 | |
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 | |
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] = \ |
155 | build_string ((int) sizeof (FNSPEC) - 1, FNSPEC ? FNSPEC : ""); |
156 | #include "internal-fn.def" |
157 | internal_fn_fnspec_array[IFN_LAST] = 0; |
158 | } |
159 | |
160 | /* Create static initializers for the information returned by |
161 | direct_internal_fn. */ |
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 } |
168 | #define mask_len_load_direct { -1, 4, false } |
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_mask_len_direct { 1, 1, false } |
174 | #define vec_cond_direct { 2, 0, false } |
175 | #define scatter_store_direct { 3, 1, false } |
176 | #define len_store_direct { 3, 3, false } |
177 | #define mask_len_store_direct { 4, 5, false } |
178 | #define vec_set_direct { 3, 3, false } |
179 | #define { 0, -1, false } |
180 | #define unary_direct { 0, 0, true } |
181 | #define unary_convert_direct { -1, 0, true } |
182 | #define binary_direct { 0, 0, true } |
183 | #define ternary_direct { 0, 0, true } |
184 | #define cond_unary_direct { 1, 1, true } |
185 | #define cond_binary_direct { 1, 1, true } |
186 | #define cond_ternary_direct { 1, 1, true } |
187 | #define cond_len_unary_direct { 1, 1, true } |
188 | #define cond_len_binary_direct { 1, 1, true } |
189 | #define cond_len_ternary_direct { 1, 1, true } |
190 | #define while_direct { 0, 2, false } |
191 | #define { 2, 2, false } |
192 | #define { 2, 2, false } |
193 | #define fold_left_direct { 1, 1, false } |
194 | #define mask_fold_left_direct { 1, 1, false } |
195 | #define mask_len_fold_left_direct { 1, 1, false } |
196 | #define check_ptrs_direct { 0, 0, false } |
197 | |
198 | const direct_internal_fn_info direct_internal_fn_array[IFN_LAST + 1] = { |
199 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) not_direct, |
200 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) TYPE##_direct, |
201 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
202 | UNSIGNED_OPTAB, TYPE) TYPE##_direct, |
203 | #include "internal-fn.def" |
204 | not_direct |
205 | }; |
206 | |
207 | /* Expand STMT using instruction ICODE. The instruction has NOUTPUTS |
208 | output operands and NINPUTS input operands, where NOUTPUTS is either |
209 | 0 or 1. The output operand (if any) comes first, followed by the |
210 | NINPUTS input operands. */ |
211 | |
212 | static void |
213 | expand_fn_using_insn (gcall *stmt, insn_code icode, unsigned int noutputs, |
214 | unsigned int ninputs) |
215 | { |
216 | gcc_assert (icode != CODE_FOR_nothing); |
217 | |
218 | expand_operand *ops = XALLOCAVEC (expand_operand, noutputs + ninputs); |
219 | unsigned int opno = 0; |
220 | rtx lhs_rtx = NULL_RTX; |
221 | tree lhs = gimple_call_lhs (gs: stmt); |
222 | |
223 | if (noutputs) |
224 | { |
225 | gcc_assert (noutputs == 1); |
226 | if (lhs) |
227 | lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
228 | |
229 | /* Do not assign directly to a promoted subreg, since there is no |
230 | guarantee that the instruction will leave the upper bits of the |
231 | register in the state required by SUBREG_PROMOTED_SIGN. */ |
232 | rtx dest = lhs_rtx; |
233 | if (dest && GET_CODE (dest) == SUBREG && SUBREG_PROMOTED_VAR_P (dest)) |
234 | dest = NULL_RTX; |
235 | create_output_operand (op: &ops[opno], x: dest, |
236 | mode: insn_data[icode].operand[opno].mode); |
237 | opno += 1; |
238 | } |
239 | else |
240 | gcc_assert (!lhs); |
241 | |
242 | for (unsigned int i = 0; i < ninputs; ++i) |
243 | { |
244 | tree rhs = gimple_call_arg (gs: stmt, index: i); |
245 | tree rhs_type = TREE_TYPE (rhs); |
246 | rtx rhs_rtx = expand_normal (exp: rhs); |
247 | if (INTEGRAL_TYPE_P (rhs_type)) |
248 | create_convert_operand_from (op: &ops[opno], value: rhs_rtx, |
249 | TYPE_MODE (rhs_type), |
250 | TYPE_UNSIGNED (rhs_type)); |
251 | else if (TREE_CODE (rhs) == SSA_NAME |
252 | && SSA_NAME_IS_DEFAULT_DEF (rhs) |
253 | && VAR_P (SSA_NAME_VAR (rhs))) |
254 | create_undefined_input_operand (op: &ops[opno], TYPE_MODE (rhs_type)); |
255 | else |
256 | create_input_operand (op: &ops[opno], value: rhs_rtx, TYPE_MODE (rhs_type)); |
257 | opno += 1; |
258 | } |
259 | |
260 | gcc_assert (opno == noutputs + ninputs); |
261 | expand_insn (icode, nops: opno, ops); |
262 | if (lhs_rtx && !rtx_equal_p (lhs_rtx, ops[0].value)) |
263 | { |
264 | /* If the return value has an integral type, convert the instruction |
265 | result to that type. This is useful for things that return an |
266 | int regardless of the size of the input. If the instruction result |
267 | is smaller than required, assume that it is signed. |
268 | |
269 | If the return value has a nonintegral type, its mode must match |
270 | the instruction result. */ |
271 | if (GET_CODE (lhs_rtx) == SUBREG && SUBREG_PROMOTED_VAR_P (lhs_rtx)) |
272 | { |
273 | /* If this is a scalar in a register that is stored in a wider |
274 | mode than the declared mode, compute the result into its |
275 | declared mode and then convert to the wider mode. */ |
276 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
277 | rtx tmp = convert_to_mode (GET_MODE (lhs_rtx), ops[0].value, 0); |
278 | convert_move (SUBREG_REG (lhs_rtx), tmp, |
279 | SUBREG_PROMOTED_SIGN (lhs_rtx)); |
280 | } |
281 | else if (GET_MODE (lhs_rtx) == GET_MODE (ops[0].value)) |
282 | emit_move_insn (lhs_rtx, ops[0].value); |
283 | else |
284 | { |
285 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
286 | convert_move (lhs_rtx, ops[0].value, 0); |
287 | } |
288 | } |
289 | } |
290 | |
291 | /* ARRAY_TYPE is an array of vector modes. Return the associated insn |
292 | for load-lanes-style optab OPTAB, or CODE_FOR_nothing if none. */ |
293 | |
294 | static enum insn_code |
295 | get_multi_vector_move (tree array_type, convert_optab optab) |
296 | { |
297 | machine_mode imode; |
298 | machine_mode vmode; |
299 | |
300 | gcc_assert (TREE_CODE (array_type) == ARRAY_TYPE); |
301 | imode = TYPE_MODE (array_type); |
302 | vmode = TYPE_MODE (TREE_TYPE (array_type)); |
303 | |
304 | return convert_optab_handler (op: optab, to_mode: imode, from_mode: vmode); |
305 | } |
306 | |
307 | /* Add mask and len arguments according to the STMT. */ |
308 | |
309 | static unsigned int |
310 | add_mask_and_len_args (expand_operand *ops, unsigned int opno, gcall *stmt) |
311 | { |
312 | internal_fn ifn = gimple_call_internal_fn (gs: stmt); |
313 | int len_index = internal_fn_len_index (ifn); |
314 | /* BIAS is always consecutive next of LEN. */ |
315 | int bias_index = len_index + 1; |
316 | int mask_index = internal_fn_mask_index (ifn); |
317 | /* The order of arguments are always {len,bias,mask}. */ |
318 | if (mask_index >= 0) |
319 | { |
320 | tree mask = gimple_call_arg (gs: stmt, index: mask_index); |
321 | rtx mask_rtx = expand_normal (exp: mask); |
322 | create_input_operand (op: &ops[opno++], value: mask_rtx, |
323 | TYPE_MODE (TREE_TYPE (mask))); |
324 | } |
325 | if (len_index >= 0) |
326 | { |
327 | tree len = gimple_call_arg (gs: stmt, index: len_index); |
328 | rtx len_rtx = expand_normal (exp: len); |
329 | create_convert_operand_from (op: &ops[opno++], value: len_rtx, |
330 | TYPE_MODE (TREE_TYPE (len)), |
331 | TYPE_UNSIGNED (TREE_TYPE (len))); |
332 | tree biast = gimple_call_arg (gs: stmt, index: bias_index); |
333 | rtx bias = expand_normal (exp: biast); |
334 | create_input_operand (op: &ops[opno++], value: bias, QImode); |
335 | } |
336 | return opno; |
337 | } |
338 | |
339 | /* Expand LOAD_LANES call STMT using optab OPTAB. */ |
340 | |
341 | static void |
342 | expand_load_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
343 | { |
344 | class expand_operand ops[2]; |
345 | tree type, lhs, rhs; |
346 | rtx target, mem; |
347 | |
348 | lhs = gimple_call_lhs (gs: stmt); |
349 | rhs = gimple_call_arg (gs: stmt, index: 0); |
350 | type = TREE_TYPE (lhs); |
351 | |
352 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
353 | mem = expand_normal (exp: rhs); |
354 | |
355 | gcc_assert (MEM_P (mem)); |
356 | PUT_MODE (x: mem, TYPE_MODE (type)); |
357 | |
358 | create_output_operand (op: &ops[0], x: target, TYPE_MODE (type)); |
359 | create_fixed_operand (op: &ops[1], x: mem); |
360 | expand_insn (icode: get_multi_vector_move (array_type: type, optab), nops: 2, ops); |
361 | if (!rtx_equal_p (target, ops[0].value)) |
362 | emit_move_insn (target, ops[0].value); |
363 | } |
364 | |
365 | /* Expand STORE_LANES call STMT using optab OPTAB. */ |
366 | |
367 | static void |
368 | expand_store_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
369 | { |
370 | class expand_operand ops[2]; |
371 | tree type, lhs, rhs; |
372 | rtx target, reg; |
373 | |
374 | lhs = gimple_call_lhs (gs: stmt); |
375 | rhs = gimple_call_arg (gs: stmt, index: 0); |
376 | type = TREE_TYPE (rhs); |
377 | |
378 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
379 | reg = expand_normal (exp: rhs); |
380 | |
381 | gcc_assert (MEM_P (target)); |
382 | PUT_MODE (x: target, TYPE_MODE (type)); |
383 | |
384 | create_fixed_operand (op: &ops[0], x: target); |
385 | create_input_operand (op: &ops[1], value: reg, TYPE_MODE (type)); |
386 | expand_insn (icode: get_multi_vector_move (array_type: type, optab), nops: 2, ops); |
387 | } |
388 | |
389 | static void |
390 | expand_ANNOTATE (internal_fn, gcall *) |
391 | { |
392 | gcc_unreachable (); |
393 | } |
394 | |
395 | /* This should get expanded in omp_device_lower pass. */ |
396 | |
397 | static void |
398 | expand_GOMP_USE_SIMT (internal_fn, gcall *) |
399 | { |
400 | gcc_unreachable (); |
401 | } |
402 | |
403 | /* This should get expanded in omp_device_lower pass. */ |
404 | |
405 | static void |
406 | expand_GOMP_SIMT_ENTER (internal_fn, gcall *) |
407 | { |
408 | gcc_unreachable (); |
409 | } |
410 | |
411 | /* Allocate per-lane storage and begin non-uniform execution region. */ |
412 | |
413 | static void |
414 | expand_GOMP_SIMT_ENTER_ALLOC (internal_fn, gcall *stmt) |
415 | { |
416 | rtx target; |
417 | tree lhs = gimple_call_lhs (gs: stmt); |
418 | if (lhs) |
419 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
420 | else |
421 | target = gen_reg_rtx (Pmode); |
422 | rtx size = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
423 | rtx align = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
424 | class expand_operand ops[3]; |
425 | create_output_operand (op: &ops[0], x: target, Pmode); |
426 | create_input_operand (op: &ops[1], value: size, Pmode); |
427 | create_input_operand (op: &ops[2], value: align, Pmode); |
428 | gcc_assert (targetm.have_omp_simt_enter ()); |
429 | expand_insn (icode: targetm.code_for_omp_simt_enter, nops: 3, ops); |
430 | if (!rtx_equal_p (target, ops[0].value)) |
431 | emit_move_insn (target, ops[0].value); |
432 | } |
433 | |
434 | /* Deallocate per-lane storage and leave non-uniform execution region. */ |
435 | |
436 | static void |
437 | expand_GOMP_SIMT_EXIT (internal_fn, gcall *stmt) |
438 | { |
439 | gcc_checking_assert (!gimple_call_lhs (stmt)); |
440 | rtx arg = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
441 | class expand_operand ops[1]; |
442 | create_input_operand (op: &ops[0], value: arg, Pmode); |
443 | gcc_assert (targetm.have_omp_simt_exit ()); |
444 | expand_insn (icode: targetm.code_for_omp_simt_exit, nops: 1, ops); |
445 | } |
446 | |
447 | /* Lane index on SIMT targets: thread index in the warp on NVPTX. On targets |
448 | without SIMT execution this should be expanded in omp_device_lower pass. */ |
449 | |
450 | static void |
451 | expand_GOMP_SIMT_LANE (internal_fn, gcall *stmt) |
452 | { |
453 | tree lhs = gimple_call_lhs (gs: stmt); |
454 | if (!lhs) |
455 | return; |
456 | |
457 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
458 | gcc_assert (targetm.have_omp_simt_lane ()); |
459 | emit_insn (targetm.gen_omp_simt_lane (target)); |
460 | } |
461 | |
462 | /* This should get expanded in omp_device_lower pass. */ |
463 | |
464 | static void |
465 | expand_GOMP_SIMT_VF (internal_fn, gcall *) |
466 | { |
467 | gcc_unreachable (); |
468 | } |
469 | |
470 | /* This should get expanded in omp_device_lower pass. */ |
471 | |
472 | static void |
473 | expand_GOMP_TARGET_REV (internal_fn, gcall *) |
474 | { |
475 | gcc_unreachable (); |
476 | } |
477 | |
478 | /* Lane index of the first SIMT lane that supplies a non-zero argument. |
479 | This is a SIMT counterpart to GOMP_SIMD_LAST_LANE, used to represent the |
480 | lane that executed the last iteration for handling OpenMP lastprivate. */ |
481 | |
482 | static void |
483 | expand_GOMP_SIMT_LAST_LANE (internal_fn, gcall *stmt) |
484 | { |
485 | tree lhs = gimple_call_lhs (gs: stmt); |
486 | if (!lhs) |
487 | return; |
488 | |
489 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
490 | rtx cond = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
491 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
492 | class expand_operand ops[2]; |
493 | create_output_operand (op: &ops[0], x: target, mode); |
494 | create_input_operand (op: &ops[1], value: cond, mode); |
495 | gcc_assert (targetm.have_omp_simt_last_lane ()); |
496 | expand_insn (icode: targetm.code_for_omp_simt_last_lane, nops: 2, ops); |
497 | if (!rtx_equal_p (target, ops[0].value)) |
498 | emit_move_insn (target, ops[0].value); |
499 | } |
500 | |
501 | /* Non-transparent predicate used in SIMT lowering of OpenMP "ordered". */ |
502 | |
503 | static void |
504 | expand_GOMP_SIMT_ORDERED_PRED (internal_fn, gcall *stmt) |
505 | { |
506 | tree lhs = gimple_call_lhs (gs: stmt); |
507 | if (!lhs) |
508 | return; |
509 | |
510 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
511 | rtx ctr = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
512 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
513 | class expand_operand ops[2]; |
514 | create_output_operand (op: &ops[0], x: target, mode); |
515 | create_input_operand (op: &ops[1], value: ctr, mode); |
516 | gcc_assert (targetm.have_omp_simt_ordered ()); |
517 | expand_insn (icode: targetm.code_for_omp_simt_ordered, nops: 2, ops); |
518 | if (!rtx_equal_p (target, ops[0].value)) |
519 | emit_move_insn (target, ops[0].value); |
520 | } |
521 | |
522 | /* "Or" boolean reduction across SIMT lanes: return non-zero in all lanes if |
523 | any lane supplies a non-zero argument. */ |
524 | |
525 | static void |
526 | expand_GOMP_SIMT_VOTE_ANY (internal_fn, gcall *stmt) |
527 | { |
528 | tree lhs = gimple_call_lhs (gs: stmt); |
529 | if (!lhs) |
530 | return; |
531 | |
532 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
533 | rtx cond = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
534 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
535 | class expand_operand ops[2]; |
536 | create_output_operand (op: &ops[0], x: target, mode); |
537 | create_input_operand (op: &ops[1], value: cond, mode); |
538 | gcc_assert (targetm.have_omp_simt_vote_any ()); |
539 | expand_insn (icode: targetm.code_for_omp_simt_vote_any, nops: 2, ops); |
540 | if (!rtx_equal_p (target, ops[0].value)) |
541 | emit_move_insn (target, ops[0].value); |
542 | } |
543 | |
544 | /* Exchange between SIMT lanes with a "butterfly" pattern: source lane index |
545 | is destination lane index XOR given offset. */ |
546 | |
547 | static void |
548 | expand_GOMP_SIMT_XCHG_BFLY (internal_fn, gcall *stmt) |
549 | { |
550 | tree lhs = gimple_call_lhs (gs: stmt); |
551 | if (!lhs) |
552 | return; |
553 | |
554 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
555 | rtx src = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
556 | rtx idx = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
557 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
558 | class expand_operand ops[3]; |
559 | create_output_operand (op: &ops[0], x: target, mode); |
560 | create_input_operand (op: &ops[1], value: src, mode); |
561 | create_input_operand (op: &ops[2], value: idx, SImode); |
562 | gcc_assert (targetm.have_omp_simt_xchg_bfly ()); |
563 | expand_insn (icode: targetm.code_for_omp_simt_xchg_bfly, nops: 3, ops); |
564 | if (!rtx_equal_p (target, ops[0].value)) |
565 | emit_move_insn (target, ops[0].value); |
566 | } |
567 | |
568 | /* Exchange between SIMT lanes according to given source lane index. */ |
569 | |
570 | static void |
571 | expand_GOMP_SIMT_XCHG_IDX (internal_fn, gcall *stmt) |
572 | { |
573 | tree lhs = gimple_call_lhs (gs: stmt); |
574 | if (!lhs) |
575 | return; |
576 | |
577 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
578 | rtx src = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
579 | rtx idx = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
580 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
581 | class expand_operand ops[3]; |
582 | create_output_operand (op: &ops[0], x: target, mode); |
583 | create_input_operand (op: &ops[1], value: src, mode); |
584 | create_input_operand (op: &ops[2], value: idx, SImode); |
585 | gcc_assert (targetm.have_omp_simt_xchg_idx ()); |
586 | expand_insn (icode: targetm.code_for_omp_simt_xchg_idx, nops: 3, ops); |
587 | if (!rtx_equal_p (target, ops[0].value)) |
588 | emit_move_insn (target, ops[0].value); |
589 | } |
590 | |
591 | /* This should get expanded in adjust_simduid_builtins. */ |
592 | |
593 | static void |
594 | expand_GOMP_SIMD_LANE (internal_fn, gcall *) |
595 | { |
596 | gcc_unreachable (); |
597 | } |
598 | |
599 | /* This should get expanded in adjust_simduid_builtins. */ |
600 | |
601 | static void |
602 | expand_GOMP_SIMD_VF (internal_fn, gcall *) |
603 | { |
604 | gcc_unreachable (); |
605 | } |
606 | |
607 | /* This should get expanded in adjust_simduid_builtins. */ |
608 | |
609 | static void |
610 | expand_GOMP_SIMD_LAST_LANE (internal_fn, gcall *) |
611 | { |
612 | gcc_unreachable (); |
613 | } |
614 | |
615 | /* This should get expanded in adjust_simduid_builtins. */ |
616 | |
617 | static void |
618 | expand_GOMP_SIMD_ORDERED_START (internal_fn, gcall *) |
619 | { |
620 | gcc_unreachable (); |
621 | } |
622 | |
623 | /* This should get expanded in adjust_simduid_builtins. */ |
624 | |
625 | static void |
626 | expand_GOMP_SIMD_ORDERED_END (internal_fn, gcall *) |
627 | { |
628 | gcc_unreachable (); |
629 | } |
630 | |
631 | /* This should get expanded in the sanopt pass. */ |
632 | |
633 | static void |
634 | expand_UBSAN_NULL (internal_fn, gcall *) |
635 | { |
636 | gcc_unreachable (); |
637 | } |
638 | |
639 | /* This should get expanded in the sanopt pass. */ |
640 | |
641 | static void |
642 | expand_UBSAN_BOUNDS (internal_fn, gcall *) |
643 | { |
644 | gcc_unreachable (); |
645 | } |
646 | |
647 | /* This should get expanded in the sanopt pass. */ |
648 | |
649 | static void |
650 | expand_UBSAN_VPTR (internal_fn, gcall *) |
651 | { |
652 | gcc_unreachable (); |
653 | } |
654 | |
655 | /* This should get expanded in the sanopt pass. */ |
656 | |
657 | static void |
658 | expand_UBSAN_PTR (internal_fn, gcall *) |
659 | { |
660 | gcc_unreachable (); |
661 | } |
662 | |
663 | /* This should get expanded in the sanopt pass. */ |
664 | |
665 | static void |
666 | expand_UBSAN_OBJECT_SIZE (internal_fn, gcall *) |
667 | { |
668 | gcc_unreachable (); |
669 | } |
670 | |
671 | /* This should get expanded in the sanopt pass. */ |
672 | |
673 | static void |
674 | expand_HWASAN_CHECK (internal_fn, gcall *) |
675 | { |
676 | gcc_unreachable (); |
677 | } |
678 | |
679 | /* For hwasan stack tagging: |
680 | Clear tags on the dynamically allocated space. |
681 | For use after an object dynamically allocated on the stack goes out of |
682 | scope. */ |
683 | static void |
684 | expand_HWASAN_ALLOCA_UNPOISON (internal_fn, gcall *gc) |
685 | { |
686 | gcc_assert (Pmode == ptr_mode); |
687 | tree restored_position = gimple_call_arg (gs: gc, index: 0); |
688 | rtx restored_rtx = expand_expr (exp: restored_position, NULL_RTX, VOIDmode, |
689 | modifier: EXPAND_NORMAL); |
690 | rtx func = init_one_libfunc ("__hwasan_tag_memory" ); |
691 | rtx off = expand_simple_binop (Pmode, MINUS, restored_rtx, |
692 | stack_pointer_rtx, NULL_RTX, 0, |
693 | OPTAB_WIDEN); |
694 | emit_library_call_value (fun: func, NULL_RTX, fn_type: LCT_NORMAL, VOIDmode, |
695 | virtual_stack_dynamic_rtx, Pmode, |
696 | HWASAN_STACK_BACKGROUND, QImode, |
697 | arg3: off, Pmode); |
698 | } |
699 | |
700 | /* For hwasan stack tagging: |
701 | Return a tag to be used for a dynamic allocation. */ |
702 | static void |
703 | expand_HWASAN_CHOOSE_TAG (internal_fn, gcall *gc) |
704 | { |
705 | tree tag = gimple_call_lhs (gs: gc); |
706 | rtx target = expand_expr (exp: tag, NULL_RTX, VOIDmode, modifier: EXPAND_NORMAL); |
707 | machine_mode mode = GET_MODE (target); |
708 | gcc_assert (mode == QImode); |
709 | |
710 | rtx base_tag = targetm.memtag.extract_tag (hwasan_frame_base (), NULL_RTX); |
711 | gcc_assert (base_tag); |
712 | rtx tag_offset = gen_int_mode (hwasan_current_frame_tag (), QImode); |
713 | rtx chosen_tag = expand_simple_binop (QImode, PLUS, base_tag, tag_offset, |
714 | target, /* unsignedp = */1, |
715 | OPTAB_WIDEN); |
716 | chosen_tag = hwasan_truncate_to_tag_size (chosen_tag, target); |
717 | |
718 | /* Really need to put the tag into the `target` RTX. */ |
719 | if (chosen_tag != target) |
720 | { |
721 | rtx temp = chosen_tag; |
722 | gcc_assert (GET_MODE (chosen_tag) == mode); |
723 | emit_move_insn (target, temp); |
724 | } |
725 | |
726 | hwasan_increment_frame_tag (); |
727 | } |
728 | |
729 | /* For hwasan stack tagging: |
730 | Tag a region of space in the shadow stack according to the base pointer of |
731 | an object on the stack. N.b. the length provided in the internal call is |
732 | required to be aligned to HWASAN_TAG_GRANULE_SIZE. */ |
733 | static void |
734 | expand_HWASAN_MARK (internal_fn, gcall *gc) |
735 | { |
736 | gcc_assert (ptr_mode == Pmode); |
737 | HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gs: gc, index: 0)); |
738 | bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON; |
739 | |
740 | tree base = gimple_call_arg (gs: gc, index: 1); |
741 | gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR); |
742 | rtx base_rtx = expand_normal (exp: base); |
743 | |
744 | rtx tag = is_poison ? HWASAN_STACK_BACKGROUND |
745 | : targetm.memtag.extract_tag (base_rtx, NULL_RTX); |
746 | rtx address = targetm.memtag.untagged_pointer (base_rtx, NULL_RTX); |
747 | |
748 | tree len = gimple_call_arg (gs: gc, index: 2); |
749 | rtx r_len = expand_normal (exp: len); |
750 | |
751 | rtx func = init_one_libfunc ("__hwasan_tag_memory" ); |
752 | emit_library_call (fun: func, fn_type: LCT_NORMAL, VOIDmode, arg1: address, Pmode, |
753 | arg2: tag, QImode, arg3: r_len, Pmode); |
754 | } |
755 | |
756 | /* For hwasan stack tagging: |
757 | Store a tag into a pointer. */ |
758 | static void |
759 | expand_HWASAN_SET_TAG (internal_fn, gcall *gc) |
760 | { |
761 | gcc_assert (ptr_mode == Pmode); |
762 | tree g_target = gimple_call_lhs (gs: gc); |
763 | tree g_ptr = gimple_call_arg (gs: gc, index: 0); |
764 | tree g_tag = gimple_call_arg (gs: gc, index: 1); |
765 | |
766 | rtx ptr = expand_normal (exp: g_ptr); |
767 | rtx tag = expand_expr (exp: g_tag, NULL_RTX, QImode, modifier: EXPAND_NORMAL); |
768 | rtx target = expand_normal (exp: g_target); |
769 | |
770 | rtx untagged = targetm.memtag.untagged_pointer (ptr, target); |
771 | rtx tagged_value = targetm.memtag.set_tag (untagged, tag, target); |
772 | if (tagged_value != target) |
773 | emit_move_insn (target, tagged_value); |
774 | } |
775 | |
776 | /* This should get expanded in the sanopt pass. */ |
777 | |
778 | static void |
779 | expand_ASAN_CHECK (internal_fn, gcall *) |
780 | { |
781 | gcc_unreachable (); |
782 | } |
783 | |
784 | /* This should get expanded in the sanopt pass. */ |
785 | |
786 | static void |
787 | expand_ASAN_MARK (internal_fn, gcall *) |
788 | { |
789 | gcc_unreachable (); |
790 | } |
791 | |
792 | /* This should get expanded in the sanopt pass. */ |
793 | |
794 | static void |
795 | expand_ASAN_POISON (internal_fn, gcall *) |
796 | { |
797 | gcc_unreachable (); |
798 | } |
799 | |
800 | /* This should get expanded in the sanopt pass. */ |
801 | |
802 | static void |
803 | expand_ASAN_POISON_USE (internal_fn, gcall *) |
804 | { |
805 | gcc_unreachable (); |
806 | } |
807 | |
808 | /* This should get expanded in the tsan pass. */ |
809 | |
810 | static void |
811 | expand_TSAN_FUNC_EXIT (internal_fn, gcall *) |
812 | { |
813 | gcc_unreachable (); |
814 | } |
815 | |
816 | /* This should get expanded in the lower pass. */ |
817 | |
818 | static void |
819 | expand_FALLTHROUGH (internal_fn, gcall *call) |
820 | { |
821 | error_at (gimple_location (g: call), |
822 | "invalid use of attribute %<fallthrough%>" ); |
823 | } |
824 | |
825 | /* Return minimum precision needed to represent all values |
826 | of ARG in SIGNed integral type. */ |
827 | |
828 | static int |
829 | get_min_precision (tree arg, signop sign) |
830 | { |
831 | int prec = TYPE_PRECISION (TREE_TYPE (arg)); |
832 | int cnt = 0; |
833 | signop orig_sign = sign; |
834 | if (TREE_CODE (arg) == INTEGER_CST) |
835 | { |
836 | int p; |
837 | if (TYPE_SIGN (TREE_TYPE (arg)) != sign) |
838 | { |
839 | widest_int w = wi::to_widest (t: arg); |
840 | w = wi::ext (x: w, offset: prec, sgn: sign); |
841 | p = wi::min_precision (x: w, sgn: sign); |
842 | } |
843 | else |
844 | p = wi::min_precision (x: wi::to_wide (t: arg), sgn: sign); |
845 | return MIN (p, prec); |
846 | } |
847 | while (CONVERT_EXPR_P (arg) |
848 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) |
849 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) <= prec) |
850 | { |
851 | arg = TREE_OPERAND (arg, 0); |
852 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) |
853 | { |
854 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) |
855 | sign = UNSIGNED; |
856 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) |
857 | return prec + (orig_sign != sign); |
858 | prec = TYPE_PRECISION (TREE_TYPE (arg)); |
859 | } |
860 | if (++cnt > 30) |
861 | return prec + (orig_sign != sign); |
862 | } |
863 | if (CONVERT_EXPR_P (arg) |
864 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) |
865 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) > prec) |
866 | { |
867 | /* We have e.g. (unsigned short) y_2 where int y_2 = (int) x_1(D); |
868 | If y_2's min precision is smaller than prec, return that. */ |
869 | int oprec = get_min_precision (TREE_OPERAND (arg, 0), sign); |
870 | if (oprec < prec) |
871 | return oprec + (orig_sign != sign); |
872 | } |
873 | if (TREE_CODE (arg) != SSA_NAME) |
874 | return prec + (orig_sign != sign); |
875 | value_range r; |
876 | while (!get_global_range_query ()->range_of_expr (r, expr: arg) |
877 | || r.varying_p () |
878 | || r.undefined_p ()) |
879 | { |
880 | gimple *g = SSA_NAME_DEF_STMT (arg); |
881 | if (is_gimple_assign (gs: g) |
882 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (g))) |
883 | { |
884 | tree t = gimple_assign_rhs1 (gs: g); |
885 | if (INTEGRAL_TYPE_P (TREE_TYPE (t)) |
886 | && TYPE_PRECISION (TREE_TYPE (t)) <= prec) |
887 | { |
888 | arg = t; |
889 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) |
890 | { |
891 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) |
892 | sign = UNSIGNED; |
893 | else if (sign == UNSIGNED && get_range_pos_neg (arg) != 1) |
894 | return prec + (orig_sign != sign); |
895 | prec = TYPE_PRECISION (TREE_TYPE (arg)); |
896 | } |
897 | if (++cnt > 30) |
898 | return prec + (orig_sign != sign); |
899 | continue; |
900 | } |
901 | } |
902 | return prec + (orig_sign != sign); |
903 | } |
904 | if (sign == TYPE_SIGN (TREE_TYPE (arg))) |
905 | { |
906 | int p1 = wi::min_precision (x: r.lower_bound (), sgn: sign); |
907 | int p2 = wi::min_precision (x: r.upper_bound (), sgn: sign); |
908 | p1 = MAX (p1, p2); |
909 | prec = MIN (prec, p1); |
910 | } |
911 | else if (sign == UNSIGNED && !wi::neg_p (x: r.lower_bound (), sgn: SIGNED)) |
912 | { |
913 | int p = wi::min_precision (x: r.upper_bound (), sgn: UNSIGNED); |
914 | prec = MIN (prec, p); |
915 | } |
916 | return prec + (orig_sign != sign); |
917 | } |
918 | |
919 | /* Helper for expand_*_overflow. Set the __imag__ part to true |
920 | (1 except for signed:1 type, in which case store -1). */ |
921 | |
922 | static void |
923 | expand_arith_set_overflow (tree lhs, rtx target) |
924 | { |
925 | if (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))) == 1 |
926 | && !TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs)))) |
927 | write_complex_part (target, constm1_rtx, true, false); |
928 | else |
929 | write_complex_part (target, const1_rtx, true, false); |
930 | } |
931 | |
932 | /* Helper for expand_*_overflow. Store RES into the __real__ part |
933 | of TARGET. If RES has larger MODE than __real__ part of TARGET, |
934 | set the __imag__ part to 1 if RES doesn't fit into it. Similarly |
935 | if LHS has smaller precision than its mode. */ |
936 | |
937 | static void |
938 | expand_arith_overflow_result_store (tree lhs, rtx target, |
939 | scalar_int_mode mode, rtx res) |
940 | { |
941 | scalar_int_mode tgtmode |
942 | = as_a <scalar_int_mode> (GET_MODE_INNER (GET_MODE (target))); |
943 | rtx lres = res; |
944 | if (tgtmode != mode) |
945 | { |
946 | rtx_code_label *done_label = gen_label_rtx (); |
947 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); |
948 | lres = convert_modes (mode: tgtmode, oldmode: mode, x: res, unsignedp: uns); |
949 | gcc_assert (GET_MODE_PRECISION (tgtmode) < GET_MODE_PRECISION (mode)); |
950 | do_compare_rtx_and_jump (res, convert_modes (mode, oldmode: tgtmode, x: lres, unsignedp: uns), |
951 | EQ, true, mode, NULL_RTX, NULL, done_label, |
952 | profile_probability::very_likely ()); |
953 | expand_arith_set_overflow (lhs, target); |
954 | emit_label (done_label); |
955 | } |
956 | int prec = TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))); |
957 | int tgtprec = GET_MODE_PRECISION (mode: tgtmode); |
958 | if (prec < tgtprec) |
959 | { |
960 | rtx_code_label *done_label = gen_label_rtx (); |
961 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); |
962 | res = lres; |
963 | if (uns) |
964 | { |
965 | rtx mask |
966 | = immed_wide_int_const (wi::shifted_mask (start: 0, width: prec, negate_p: false, precision: tgtprec), |
967 | tgtmode); |
968 | lres = expand_simple_binop (tgtmode, AND, res, mask, NULL_RTX, |
969 | true, OPTAB_LIB_WIDEN); |
970 | } |
971 | else |
972 | { |
973 | lres = expand_shift (LSHIFT_EXPR, tgtmode, res, tgtprec - prec, |
974 | NULL_RTX, 1); |
975 | lres = expand_shift (RSHIFT_EXPR, tgtmode, lres, tgtprec - prec, |
976 | NULL_RTX, 0); |
977 | } |
978 | do_compare_rtx_and_jump (res, lres, |
979 | EQ, true, tgtmode, NULL_RTX, NULL, done_label, |
980 | profile_probability::very_likely ()); |
981 | expand_arith_set_overflow (lhs, target); |
982 | emit_label (done_label); |
983 | } |
984 | write_complex_part (target, lres, false, false); |
985 | } |
986 | |
987 | /* Helper for expand_*_overflow. Store RES into TARGET. */ |
988 | |
989 | static void |
990 | expand_ubsan_result_store (tree lhs, rtx target, scalar_int_mode mode, |
991 | rtx res, rtx_code_label *do_error) |
992 | { |
993 | if (TREE_CODE (TREE_TYPE (lhs)) == BITINT_TYPE |
994 | && TYPE_PRECISION (TREE_TYPE (lhs)) < GET_MODE_PRECISION (mode)) |
995 | { |
996 | int uns = TYPE_UNSIGNED (TREE_TYPE (lhs)); |
997 | int prec = TYPE_PRECISION (TREE_TYPE (lhs)); |
998 | int tgtprec = GET_MODE_PRECISION (mode); |
999 | rtx resc = gen_reg_rtx (mode), lres; |
1000 | emit_move_insn (resc, res); |
1001 | if (uns) |
1002 | { |
1003 | rtx mask |
1004 | = immed_wide_int_const (wi::shifted_mask (start: 0, width: prec, negate_p: false, precision: tgtprec), |
1005 | mode); |
1006 | lres = expand_simple_binop (mode, AND, res, mask, NULL_RTX, |
1007 | true, OPTAB_LIB_WIDEN); |
1008 | } |
1009 | else |
1010 | { |
1011 | lres = expand_shift (LSHIFT_EXPR, mode, res, tgtprec - prec, |
1012 | NULL_RTX, 1); |
1013 | lres = expand_shift (RSHIFT_EXPR, mode, lres, tgtprec - prec, |
1014 | NULL_RTX, 0); |
1015 | } |
1016 | if (lres != res) |
1017 | emit_move_insn (res, lres); |
1018 | do_compare_rtx_and_jump (res, resc, |
1019 | NE, true, mode, NULL_RTX, NULL, do_error, |
1020 | profile_probability::very_unlikely ()); |
1021 | } |
1022 | if (GET_CODE (target) == SUBREG && SUBREG_PROMOTED_VAR_P (target)) |
1023 | /* If this is a scalar in a register that is stored in a wider mode |
1024 | than the declared mode, compute the result into its declared mode |
1025 | and then convert to the wider mode. Our value is the computed |
1026 | expression. */ |
1027 | convert_move (SUBREG_REG (target), res, SUBREG_PROMOTED_SIGN (target)); |
1028 | else |
1029 | emit_move_insn (target, res); |
1030 | } |
1031 | |
1032 | /* Add sub/add overflow checking to the statement STMT. |
1033 | CODE says whether the operation is +, or -. */ |
1034 | |
1035 | void |
1036 | expand_addsub_overflow (location_t loc, tree_code code, tree lhs, |
1037 | tree arg0, tree arg1, bool unsr_p, bool uns0_p, |
1038 | bool uns1_p, bool is_ubsan, tree *datap) |
1039 | { |
1040 | rtx res, target = NULL_RTX; |
1041 | tree fn; |
1042 | rtx_code_label *done_label = gen_label_rtx (); |
1043 | rtx_code_label *do_error = gen_label_rtx (); |
1044 | do_pending_stack_adjust (); |
1045 | rtx op0 = expand_normal (exp: arg0); |
1046 | rtx op1 = expand_normal (exp: arg1); |
1047 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1048 | int prec = GET_MODE_PRECISION (mode); |
1049 | rtx sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); |
1050 | bool do_xor = false; |
1051 | |
1052 | if (is_ubsan) |
1053 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); |
1054 | |
1055 | if (lhs) |
1056 | { |
1057 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1058 | if (!is_ubsan) |
1059 | write_complex_part (target, const0_rtx, true, false); |
1060 | } |
1061 | |
1062 | /* We assume both operands and result have the same precision |
1063 | here (GET_MODE_BITSIZE (mode)), S stands for signed type |
1064 | with that precision, U for unsigned type with that precision, |
1065 | sgn for unsigned most significant bit in that precision. |
1066 | s1 is signed first operand, u1 is unsigned first operand, |
1067 | s2 is signed second operand, u2 is unsigned second operand, |
1068 | sr is signed result, ur is unsigned result and the following |
1069 | rules say how to compute result (which is always result of |
1070 | the operands as if both were unsigned, cast to the right |
1071 | signedness) and how to compute whether operation overflowed. |
1072 | |
1073 | s1 + s2 -> sr |
1074 | res = (S) ((U) s1 + (U) s2) |
1075 | ovf = s2 < 0 ? res > s1 : res < s1 (or jump on overflow) |
1076 | s1 - s2 -> sr |
1077 | res = (S) ((U) s1 - (U) s2) |
1078 | ovf = s2 < 0 ? res < s1 : res > s2 (or jump on overflow) |
1079 | u1 + u2 -> ur |
1080 | res = u1 + u2 |
1081 | ovf = res < u1 (or jump on carry, but RTL opts will handle it) |
1082 | u1 - u2 -> ur |
1083 | res = u1 - u2 |
1084 | ovf = res > u1 (or jump on carry, but RTL opts will handle it) |
1085 | s1 + u2 -> sr |
1086 | res = (S) ((U) s1 + u2) |
1087 | ovf = ((U) res ^ sgn) < u2 |
1088 | s1 + u2 -> ur |
1089 | t1 = (S) (u2 ^ sgn) |
1090 | t2 = s1 + t1 |
1091 | res = (U) t2 ^ sgn |
1092 | ovf = t1 < 0 ? t2 > s1 : t2 < s1 (or jump on overflow) |
1093 | s1 - u2 -> sr |
1094 | res = (S) ((U) s1 - u2) |
1095 | ovf = u2 > ((U) s1 ^ sgn) |
1096 | s1 - u2 -> ur |
1097 | res = (U) s1 - u2 |
1098 | ovf = s1 < 0 || u2 > (U) s1 |
1099 | u1 - s2 -> sr |
1100 | res = u1 - (U) s2 |
1101 | ovf = u1 >= ((U) s2 ^ sgn) |
1102 | u1 - s2 -> ur |
1103 | t1 = u1 ^ sgn |
1104 | t2 = t1 - (U) s2 |
1105 | res = t2 ^ sgn |
1106 | ovf = s2 < 0 ? (S) t2 < (S) t1 : (S) t2 > (S) t1 (or jump on overflow) |
1107 | s1 + s2 -> ur |
1108 | res = (U) s1 + (U) s2 |
1109 | ovf = s2 < 0 ? (s1 | (S) res) < 0) : (s1 & (S) res) < 0) |
1110 | u1 + u2 -> sr |
1111 | res = (S) (u1 + u2) |
1112 | ovf = (U) res < u2 || res < 0 |
1113 | u1 - u2 -> sr |
1114 | res = (S) (u1 - u2) |
1115 | ovf = u1 >= u2 ? res < 0 : res >= 0 |
1116 | s1 - s2 -> ur |
1117 | res = (U) s1 - (U) s2 |
1118 | ovf = s2 >= 0 ? ((s1 | (S) res) < 0) : ((s1 & (S) res) < 0) */ |
1119 | |
1120 | if (code == PLUS_EXPR && uns0_p && !uns1_p) |
1121 | { |
1122 | /* PLUS_EXPR is commutative, if operand signedness differs, |
1123 | canonicalize to the first operand being signed and second |
1124 | unsigned to simplify following code. */ |
1125 | std::swap (a&: op0, b&: op1); |
1126 | std::swap (a&: arg0, b&: arg1); |
1127 | uns0_p = false; |
1128 | uns1_p = true; |
1129 | } |
1130 | |
1131 | /* u1 +- u2 -> ur */ |
1132 | if (uns0_p && uns1_p && unsr_p) |
1133 | { |
1134 | insn_code icode = optab_handler (op: code == PLUS_EXPR ? uaddv4_optab |
1135 | : usubv4_optab, mode); |
1136 | if (icode != CODE_FOR_nothing) |
1137 | { |
1138 | class expand_operand ops[4]; |
1139 | rtx_insn *last = get_last_insn (); |
1140 | |
1141 | res = gen_reg_rtx (mode); |
1142 | create_output_operand (op: &ops[0], x: res, mode); |
1143 | create_input_operand (op: &ops[1], value: op0, mode); |
1144 | create_input_operand (op: &ops[2], value: op1, mode); |
1145 | create_fixed_operand (op: &ops[3], x: do_error); |
1146 | if (maybe_expand_insn (icode, nops: 4, ops)) |
1147 | { |
1148 | last = get_last_insn (); |
1149 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1150 | && JUMP_P (last) |
1151 | && any_condjump_p (last) |
1152 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1153 | add_reg_br_prob_note (last, |
1154 | profile_probability::very_unlikely ()); |
1155 | emit_jump (done_label); |
1156 | goto do_error_label; |
1157 | } |
1158 | |
1159 | delete_insns_since (last); |
1160 | } |
1161 | |
1162 | /* Compute the operation. On RTL level, the addition is always |
1163 | unsigned. */ |
1164 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1165 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1166 | rtx tem = op0; |
1167 | /* For PLUS_EXPR, the operation is commutative, so we can pick |
1168 | operand to compare against. For prec <= BITS_PER_WORD, I think |
1169 | preferring REG operand is better over CONST_INT, because |
1170 | the CONST_INT might enlarge the instruction or CSE would need |
1171 | to figure out we'd already loaded it into a register before. |
1172 | For prec > BITS_PER_WORD, I think CONST_INT might be more beneficial, |
1173 | as then the multi-word comparison can be perhaps simplified. */ |
1174 | if (code == PLUS_EXPR |
1175 | && (prec <= BITS_PER_WORD |
1176 | ? (CONST_SCALAR_INT_P (op0) && REG_P (op1)) |
1177 | : CONST_SCALAR_INT_P (op1))) |
1178 | tem = op1; |
1179 | do_compare_rtx_and_jump (res, tem, code == PLUS_EXPR ? GEU : LEU, |
1180 | true, mode, NULL_RTX, NULL, done_label, |
1181 | profile_probability::very_likely ()); |
1182 | goto do_error_label; |
1183 | } |
1184 | |
1185 | /* s1 +- u2 -> sr */ |
1186 | if (!uns0_p && uns1_p && !unsr_p) |
1187 | { |
1188 | /* Compute the operation. On RTL level, the addition is always |
1189 | unsigned. */ |
1190 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1191 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1192 | rtx tem = expand_binop (mode, add_optab, |
1193 | code == PLUS_EXPR ? res : op0, sgn, |
1194 | NULL_RTX, false, OPTAB_LIB_WIDEN); |
1195 | do_compare_rtx_and_jump (tem, op1, GEU, true, mode, NULL_RTX, NULL, |
1196 | done_label, profile_probability::very_likely ()); |
1197 | goto do_error_label; |
1198 | } |
1199 | |
1200 | /* s1 + u2 -> ur */ |
1201 | if (code == PLUS_EXPR && !uns0_p && uns1_p && unsr_p) |
1202 | { |
1203 | op1 = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, |
1204 | OPTAB_LIB_WIDEN); |
1205 | /* As we've changed op1, we have to avoid using the value range |
1206 | for the original argument. */ |
1207 | arg1 = error_mark_node; |
1208 | do_xor = true; |
1209 | goto do_signed; |
1210 | } |
1211 | |
1212 | /* u1 - s2 -> ur */ |
1213 | if (code == MINUS_EXPR && uns0_p && !uns1_p && unsr_p) |
1214 | { |
1215 | op0 = expand_binop (mode, add_optab, op0, sgn, NULL_RTX, false, |
1216 | OPTAB_LIB_WIDEN); |
1217 | /* As we've changed op0, we have to avoid using the value range |
1218 | for the original argument. */ |
1219 | arg0 = error_mark_node; |
1220 | do_xor = true; |
1221 | goto do_signed; |
1222 | } |
1223 | |
1224 | /* s1 - u2 -> ur */ |
1225 | if (code == MINUS_EXPR && !uns0_p && uns1_p && unsr_p) |
1226 | { |
1227 | /* Compute the operation. On RTL level, the addition is always |
1228 | unsigned. */ |
1229 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1230 | OPTAB_LIB_WIDEN); |
1231 | int pos_neg = get_range_pos_neg (arg0); |
1232 | if (pos_neg == 2) |
1233 | /* If ARG0 is known to be always negative, this is always overflow. */ |
1234 | emit_jump (do_error); |
1235 | else if (pos_neg == 3) |
1236 | /* If ARG0 is not known to be always positive, check at runtime. */ |
1237 | do_compare_rtx_and_jump (op0, const0_rtx, LT, false, mode, NULL_RTX, |
1238 | NULL, do_error, profile_probability::very_unlikely ()); |
1239 | do_compare_rtx_and_jump (op1, op0, LEU, true, mode, NULL_RTX, NULL, |
1240 | done_label, profile_probability::very_likely ()); |
1241 | goto do_error_label; |
1242 | } |
1243 | |
1244 | /* u1 - s2 -> sr */ |
1245 | if (code == MINUS_EXPR && uns0_p && !uns1_p && !unsr_p) |
1246 | { |
1247 | /* Compute the operation. On RTL level, the addition is always |
1248 | unsigned. */ |
1249 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1250 | OPTAB_LIB_WIDEN); |
1251 | rtx tem = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, |
1252 | OPTAB_LIB_WIDEN); |
1253 | do_compare_rtx_and_jump (op0, tem, LTU, true, mode, NULL_RTX, NULL, |
1254 | done_label, profile_probability::very_likely ()); |
1255 | goto do_error_label; |
1256 | } |
1257 | |
1258 | /* u1 + u2 -> sr */ |
1259 | if (code == PLUS_EXPR && uns0_p && uns1_p && !unsr_p) |
1260 | { |
1261 | /* Compute the operation. On RTL level, the addition is always |
1262 | unsigned. */ |
1263 | res = expand_binop (mode, add_optab, op0, op1, NULL_RTX, false, |
1264 | OPTAB_LIB_WIDEN); |
1265 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
1266 | NULL, do_error, profile_probability::very_unlikely ()); |
1267 | rtx tem = op1; |
1268 | /* The operation is commutative, so we can pick operand to compare |
1269 | against. For prec <= BITS_PER_WORD, I think preferring REG operand |
1270 | is better over CONST_INT, because the CONST_INT might enlarge the |
1271 | instruction or CSE would need to figure out we'd already loaded it |
1272 | into a register before. For prec > BITS_PER_WORD, I think CONST_INT |
1273 | might be more beneficial, as then the multi-word comparison can be |
1274 | perhaps simplified. */ |
1275 | if (prec <= BITS_PER_WORD |
1276 | ? (CONST_SCALAR_INT_P (op1) && REG_P (op0)) |
1277 | : CONST_SCALAR_INT_P (op0)) |
1278 | tem = op0; |
1279 | do_compare_rtx_and_jump (res, tem, GEU, true, mode, NULL_RTX, NULL, |
1280 | done_label, profile_probability::very_likely ()); |
1281 | goto do_error_label; |
1282 | } |
1283 | |
1284 | /* s1 +- s2 -> ur */ |
1285 | if (!uns0_p && !uns1_p && unsr_p) |
1286 | { |
1287 | /* Compute the operation. On RTL level, the addition is always |
1288 | unsigned. */ |
1289 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1290 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1291 | int pos_neg = get_range_pos_neg (arg1); |
1292 | if (code == PLUS_EXPR) |
1293 | { |
1294 | int pos_neg0 = get_range_pos_neg (arg0); |
1295 | if (pos_neg0 != 3 && pos_neg == 3) |
1296 | { |
1297 | std::swap (a&: op0, b&: op1); |
1298 | pos_neg = pos_neg0; |
1299 | } |
1300 | } |
1301 | rtx tem; |
1302 | if (pos_neg != 3) |
1303 | { |
1304 | tem = expand_binop (mode, ((pos_neg == 1) ^ (code == MINUS_EXPR)) |
1305 | ? and_optab : ior_optab, |
1306 | op0, res, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1307 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL, |
1308 | NULL, done_label, profile_probability::very_likely ()); |
1309 | } |
1310 | else |
1311 | { |
1312 | rtx_code_label *do_ior_label = gen_label_rtx (); |
1313 | do_compare_rtx_and_jump (op1, const0_rtx, |
1314 | code == MINUS_EXPR ? GE : LT, false, mode, |
1315 | NULL_RTX, NULL, do_ior_label, |
1316 | profile_probability::even ()); |
1317 | tem = expand_binop (mode, and_optab, op0, res, NULL_RTX, false, |
1318 | OPTAB_LIB_WIDEN); |
1319 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1320 | NULL, done_label, profile_probability::very_likely ()); |
1321 | emit_jump (do_error); |
1322 | emit_label (do_ior_label); |
1323 | tem = expand_binop (mode, ior_optab, op0, res, NULL_RTX, false, |
1324 | OPTAB_LIB_WIDEN); |
1325 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1326 | NULL, done_label, profile_probability::very_likely ()); |
1327 | } |
1328 | goto do_error_label; |
1329 | } |
1330 | |
1331 | /* u1 - u2 -> sr */ |
1332 | if (code == MINUS_EXPR && uns0_p && uns1_p && !unsr_p) |
1333 | { |
1334 | /* Compute the operation. On RTL level, the addition is always |
1335 | unsigned. */ |
1336 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1337 | OPTAB_LIB_WIDEN); |
1338 | rtx_code_label *op0_geu_op1 = gen_label_rtx (); |
1339 | do_compare_rtx_and_jump (op0, op1, GEU, true, mode, NULL_RTX, NULL, |
1340 | op0_geu_op1, profile_probability::even ()); |
1341 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
1342 | NULL, done_label, profile_probability::very_likely ()); |
1343 | emit_jump (do_error); |
1344 | emit_label (op0_geu_op1); |
1345 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
1346 | NULL, done_label, profile_probability::very_likely ()); |
1347 | goto do_error_label; |
1348 | } |
1349 | |
1350 | gcc_assert (!uns0_p && !uns1_p && !unsr_p); |
1351 | |
1352 | /* s1 +- s2 -> sr */ |
1353 | do_signed: |
1354 | { |
1355 | insn_code icode = optab_handler (op: code == PLUS_EXPR ? addv4_optab |
1356 | : subv4_optab, mode); |
1357 | if (icode != CODE_FOR_nothing) |
1358 | { |
1359 | class expand_operand ops[4]; |
1360 | rtx_insn *last = get_last_insn (); |
1361 | |
1362 | res = gen_reg_rtx (mode); |
1363 | create_output_operand (op: &ops[0], x: res, mode); |
1364 | create_input_operand (op: &ops[1], value: op0, mode); |
1365 | create_input_operand (op: &ops[2], value: op1, mode); |
1366 | create_fixed_operand (op: &ops[3], x: do_error); |
1367 | if (maybe_expand_insn (icode, nops: 4, ops)) |
1368 | { |
1369 | last = get_last_insn (); |
1370 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1371 | && JUMP_P (last) |
1372 | && any_condjump_p (last) |
1373 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1374 | add_reg_br_prob_note (last, |
1375 | profile_probability::very_unlikely ()); |
1376 | emit_jump (done_label); |
1377 | goto do_error_label; |
1378 | } |
1379 | |
1380 | delete_insns_since (last); |
1381 | } |
1382 | |
1383 | /* Compute the operation. On RTL level, the addition is always |
1384 | unsigned. */ |
1385 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1386 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1387 | |
1388 | /* If we can prove that one of the arguments (for MINUS_EXPR only |
1389 | the second operand, as subtraction is not commutative) is always |
1390 | non-negative or always negative, we can do just one comparison |
1391 | and conditional jump. */ |
1392 | int pos_neg = get_range_pos_neg (arg1); |
1393 | if (code == PLUS_EXPR) |
1394 | { |
1395 | int pos_neg0 = get_range_pos_neg (arg0); |
1396 | if (pos_neg0 != 3 && pos_neg == 3) |
1397 | { |
1398 | std::swap (a&: op0, b&: op1); |
1399 | pos_neg = pos_neg0; |
1400 | } |
1401 | } |
1402 | |
1403 | /* Addition overflows if and only if the two operands have the same sign, |
1404 | and the result has the opposite sign. Subtraction overflows if and |
1405 | only if the two operands have opposite sign, and the subtrahend has |
1406 | the same sign as the result. Here 0 is counted as positive. */ |
1407 | if (pos_neg == 3) |
1408 | { |
1409 | /* Compute op0 ^ op1 (operands have opposite sign). */ |
1410 | rtx op_xor = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1411 | OPTAB_LIB_WIDEN); |
1412 | |
1413 | /* Compute res ^ op1 (result and 2nd operand have opposite sign). */ |
1414 | rtx res_xor = expand_binop (mode, xor_optab, res, op1, NULL_RTX, false, |
1415 | OPTAB_LIB_WIDEN); |
1416 | |
1417 | rtx tem; |
1418 | if (code == PLUS_EXPR) |
1419 | { |
1420 | /* Compute (res ^ op1) & ~(op0 ^ op1). */ |
1421 | tem = expand_unop (mode, one_cmpl_optab, op_xor, NULL_RTX, false); |
1422 | tem = expand_binop (mode, and_optab, res_xor, tem, NULL_RTX, false, |
1423 | OPTAB_LIB_WIDEN); |
1424 | } |
1425 | else |
1426 | { |
1427 | /* Compute (op0 ^ op1) & ~(res ^ op1). */ |
1428 | tem = expand_unop (mode, one_cmpl_optab, res_xor, NULL_RTX, false); |
1429 | tem = expand_binop (mode, and_optab, op_xor, tem, NULL_RTX, false, |
1430 | OPTAB_LIB_WIDEN); |
1431 | } |
1432 | |
1433 | /* No overflow if the result has bit sign cleared. */ |
1434 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1435 | NULL, done_label, profile_probability::very_likely ()); |
1436 | } |
1437 | |
1438 | /* Compare the result of the operation with the first operand. |
1439 | No overflow for addition if second operand is positive and result |
1440 | is larger or second operand is negative and result is smaller. |
1441 | Likewise for subtraction with sign of second operand flipped. */ |
1442 | else |
1443 | do_compare_rtx_and_jump (res, op0, |
1444 | (pos_neg == 1) ^ (code == MINUS_EXPR) ? GE : LE, |
1445 | false, mode, NULL_RTX, NULL, done_label, |
1446 | profile_probability::very_likely ()); |
1447 | } |
1448 | |
1449 | do_error_label: |
1450 | emit_label (do_error); |
1451 | if (is_ubsan) |
1452 | { |
1453 | /* Expand the ubsan builtin call. */ |
1454 | push_temp_slots (); |
1455 | fn = ubsan_build_overflow_builtin (code, loc, TREE_TYPE (arg0), |
1456 | arg0, arg1, datap); |
1457 | expand_normal (exp: fn); |
1458 | pop_temp_slots (); |
1459 | do_pending_stack_adjust (); |
1460 | } |
1461 | else if (lhs) |
1462 | expand_arith_set_overflow (lhs, target); |
1463 | |
1464 | /* We're done. */ |
1465 | emit_label (done_label); |
1466 | |
1467 | if (lhs) |
1468 | { |
1469 | if (is_ubsan) |
1470 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
1471 | else |
1472 | { |
1473 | if (do_xor) |
1474 | res = expand_binop (mode, add_optab, res, sgn, NULL_RTX, false, |
1475 | OPTAB_LIB_WIDEN); |
1476 | |
1477 | expand_arith_overflow_result_store (lhs, target, mode, res); |
1478 | } |
1479 | } |
1480 | } |
1481 | |
1482 | /* Add negate overflow checking to the statement STMT. */ |
1483 | |
1484 | static void |
1485 | expand_neg_overflow (location_t loc, tree lhs, tree arg1, bool is_ubsan, |
1486 | tree *datap) |
1487 | { |
1488 | rtx res, op1; |
1489 | tree fn; |
1490 | rtx_code_label *done_label, *do_error; |
1491 | rtx target = NULL_RTX; |
1492 | |
1493 | done_label = gen_label_rtx (); |
1494 | do_error = gen_label_rtx (); |
1495 | |
1496 | do_pending_stack_adjust (); |
1497 | op1 = expand_normal (exp: arg1); |
1498 | |
1499 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg1)); |
1500 | if (lhs) |
1501 | { |
1502 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1503 | if (!is_ubsan) |
1504 | write_complex_part (target, const0_rtx, true, false); |
1505 | } |
1506 | |
1507 | enum insn_code icode = optab_handler (op: negv3_optab, mode); |
1508 | if (icode != CODE_FOR_nothing) |
1509 | { |
1510 | class expand_operand ops[3]; |
1511 | rtx_insn *last = get_last_insn (); |
1512 | |
1513 | res = gen_reg_rtx (mode); |
1514 | create_output_operand (op: &ops[0], x: res, mode); |
1515 | create_input_operand (op: &ops[1], value: op1, mode); |
1516 | create_fixed_operand (op: &ops[2], x: do_error); |
1517 | if (maybe_expand_insn (icode, nops: 3, ops)) |
1518 | { |
1519 | last = get_last_insn (); |
1520 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1521 | && JUMP_P (last) |
1522 | && any_condjump_p (last) |
1523 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1524 | add_reg_br_prob_note (last, |
1525 | profile_probability::very_unlikely ()); |
1526 | emit_jump (done_label); |
1527 | } |
1528 | else |
1529 | { |
1530 | delete_insns_since (last); |
1531 | icode = CODE_FOR_nothing; |
1532 | } |
1533 | } |
1534 | |
1535 | if (icode == CODE_FOR_nothing) |
1536 | { |
1537 | /* Compute the operation. On RTL level, the addition is always |
1538 | unsigned. */ |
1539 | res = expand_unop (mode, neg_optab, op1, NULL_RTX, false); |
1540 | |
1541 | /* Compare the operand with the most negative value. */ |
1542 | rtx minv = expand_normal (TYPE_MIN_VALUE (TREE_TYPE (arg1))); |
1543 | do_compare_rtx_and_jump (op1, minv, NE, true, mode, NULL_RTX, NULL, |
1544 | done_label, profile_probability::very_likely ()); |
1545 | } |
1546 | |
1547 | emit_label (do_error); |
1548 | if (is_ubsan) |
1549 | { |
1550 | /* Expand the ubsan builtin call. */ |
1551 | push_temp_slots (); |
1552 | fn = ubsan_build_overflow_builtin (NEGATE_EXPR, loc, TREE_TYPE (arg1), |
1553 | arg1, NULL_TREE, datap); |
1554 | expand_normal (exp: fn); |
1555 | pop_temp_slots (); |
1556 | do_pending_stack_adjust (); |
1557 | } |
1558 | else if (lhs) |
1559 | expand_arith_set_overflow (lhs, target); |
1560 | |
1561 | /* We're done. */ |
1562 | emit_label (done_label); |
1563 | |
1564 | if (lhs) |
1565 | { |
1566 | if (is_ubsan) |
1567 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
1568 | else |
1569 | expand_arith_overflow_result_store (lhs, target, mode, res); |
1570 | } |
1571 | } |
1572 | |
1573 | /* Return true if UNS WIDEN_MULT_EXPR with result mode WMODE and operand |
1574 | mode MODE can be expanded without using a libcall. */ |
1575 | |
1576 | static bool |
1577 | can_widen_mult_without_libcall (scalar_int_mode wmode, scalar_int_mode mode, |
1578 | rtx op0, rtx op1, bool uns) |
1579 | { |
1580 | if (find_widening_optab_handler (umul_widen_optab, wmode, mode) |
1581 | != CODE_FOR_nothing) |
1582 | return true; |
1583 | |
1584 | if (find_widening_optab_handler (smul_widen_optab, wmode, mode) |
1585 | != CODE_FOR_nothing) |
1586 | return true; |
1587 | |
1588 | rtx_insn *last = get_last_insn (); |
1589 | if (CONSTANT_P (op0)) |
1590 | op0 = convert_modes (mode: wmode, oldmode: mode, x: op0, unsignedp: uns); |
1591 | else |
1592 | op0 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 1); |
1593 | if (CONSTANT_P (op1)) |
1594 | op1 = convert_modes (mode: wmode, oldmode: mode, x: op1, unsignedp: uns); |
1595 | else |
1596 | op1 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 2); |
1597 | rtx ret = expand_mult (wmode, op0, op1, NULL_RTX, uns, true); |
1598 | delete_insns_since (last); |
1599 | return ret != NULL_RTX; |
1600 | } |
1601 | |
1602 | /* Add mul overflow checking to the statement STMT. */ |
1603 | |
1604 | static void |
1605 | expand_mul_overflow (location_t loc, tree lhs, tree arg0, tree arg1, |
1606 | bool unsr_p, bool uns0_p, bool uns1_p, bool is_ubsan, |
1607 | tree *datap) |
1608 | { |
1609 | rtx res, op0, op1; |
1610 | tree fn, type; |
1611 | rtx_code_label *done_label, *do_error; |
1612 | rtx target = NULL_RTX; |
1613 | signop sign; |
1614 | enum insn_code icode; |
1615 | |
1616 | done_label = gen_label_rtx (); |
1617 | do_error = gen_label_rtx (); |
1618 | |
1619 | do_pending_stack_adjust (); |
1620 | op0 = expand_normal (exp: arg0); |
1621 | op1 = expand_normal (exp: arg1); |
1622 | |
1623 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1624 | bool uns = unsr_p; |
1625 | if (lhs) |
1626 | { |
1627 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1628 | if (!is_ubsan) |
1629 | write_complex_part (target, const0_rtx, true, false); |
1630 | } |
1631 | |
1632 | if (is_ubsan) |
1633 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); |
1634 | |
1635 | /* We assume both operands and result have the same precision |
1636 | here (GET_MODE_BITSIZE (mode)), S stands for signed type |
1637 | with that precision, U for unsigned type with that precision, |
1638 | sgn for unsigned most significant bit in that precision. |
1639 | s1 is signed first operand, u1 is unsigned first operand, |
1640 | s2 is signed second operand, u2 is unsigned second operand, |
1641 | sr is signed result, ur is unsigned result and the following |
1642 | rules say how to compute result (which is always result of |
1643 | the operands as if both were unsigned, cast to the right |
1644 | signedness) and how to compute whether operation overflowed. |
1645 | main_ovf (false) stands for jump on signed multiplication |
1646 | overflow or the main algorithm with uns == false. |
1647 | main_ovf (true) stands for jump on unsigned multiplication |
1648 | overflow or the main algorithm with uns == true. |
1649 | |
1650 | s1 * s2 -> sr |
1651 | res = (S) ((U) s1 * (U) s2) |
1652 | ovf = main_ovf (false) |
1653 | u1 * u2 -> ur |
1654 | res = u1 * u2 |
1655 | ovf = main_ovf (true) |
1656 | s1 * u2 -> ur |
1657 | res = (U) s1 * u2 |
1658 | ovf = (s1 < 0 && u2) || main_ovf (true) |
1659 | u1 * u2 -> sr |
1660 | res = (S) (u1 * u2) |
1661 | ovf = res < 0 || main_ovf (true) |
1662 | s1 * u2 -> sr |
1663 | res = (S) ((U) s1 * u2) |
1664 | ovf = (S) u2 >= 0 ? main_ovf (false) |
1665 | : (s1 != 0 && (s1 != -1 || u2 != (U) res)) |
1666 | s1 * s2 -> ur |
1667 | t1 = (s1 & s2) < 0 ? (-(U) s1) : ((U) s1) |
1668 | t2 = (s1 & s2) < 0 ? (-(U) s2) : ((U) s2) |
1669 | res = t1 * t2 |
1670 | ovf = (s1 ^ s2) < 0 ? (s1 && s2) : main_ovf (true) */ |
1671 | |
1672 | if (uns0_p && !uns1_p) |
1673 | { |
1674 | /* Multiplication is commutative, if operand signedness differs, |
1675 | canonicalize to the first operand being signed and second |
1676 | unsigned to simplify following code. */ |
1677 | std::swap (a&: op0, b&: op1); |
1678 | std::swap (a&: arg0, b&: arg1); |
1679 | uns0_p = false; |
1680 | uns1_p = true; |
1681 | } |
1682 | |
1683 | int pos_neg0 = get_range_pos_neg (arg0); |
1684 | int pos_neg1 = get_range_pos_neg (arg1); |
1685 | /* Unsigned types with smaller than mode precision, even if they have most |
1686 | significant bit set, are still zero-extended. */ |
1687 | if (uns0_p && TYPE_PRECISION (TREE_TYPE (arg0)) < GET_MODE_PRECISION (mode)) |
1688 | pos_neg0 = 1; |
1689 | if (uns1_p && TYPE_PRECISION (TREE_TYPE (arg1)) < GET_MODE_PRECISION (mode)) |
1690 | pos_neg1 = 1; |
1691 | |
1692 | /* s1 * u2 -> ur */ |
1693 | if (!uns0_p && uns1_p && unsr_p) |
1694 | { |
1695 | switch (pos_neg0) |
1696 | { |
1697 | case 1: |
1698 | /* If s1 is non-negative, just perform normal u1 * u2 -> ur. */ |
1699 | goto do_main; |
1700 | case 2: |
1701 | /* If s1 is negative, avoid the main code, just multiply and |
1702 | signal overflow if op1 is not 0. */ |
1703 | struct separate_ops ops; |
1704 | ops.code = MULT_EXPR; |
1705 | ops.type = TREE_TYPE (arg1); |
1706 | ops.op0 = make_tree (ops.type, op0); |
1707 | ops.op1 = make_tree (ops.type, op1); |
1708 | ops.op2 = NULL_TREE; |
1709 | ops.location = loc; |
1710 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1711 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
1712 | NULL, done_label, profile_probability::very_likely ()); |
1713 | goto do_error_label; |
1714 | case 3: |
1715 | if (get_min_precision (arg: arg1, sign: UNSIGNED) |
1716 | + get_min_precision (arg: arg0, sign: SIGNED) <= GET_MODE_PRECISION (mode)) |
1717 | { |
1718 | /* If the first operand is sign extended from narrower type, the |
1719 | second operand is zero extended from narrower type and |
1720 | the sum of the two precisions is smaller or equal to the |
1721 | result precision: if the first argument is at runtime |
1722 | non-negative, maximum result will be 0x7e81 or 0x7f..fe80..01 |
1723 | and there will be no overflow, if the first argument is |
1724 | negative and the second argument zero, the result will be |
1725 | 0 and there will be no overflow, if the first argument is |
1726 | negative and the second argument positive, the result when |
1727 | treated as signed will be negative (minimum -0x7f80 or |
1728 | -0x7f..f80..0) there will be always overflow. So, do |
1729 | res = (U) (s1 * u2) |
1730 | ovf = (S) res < 0 */ |
1731 | struct separate_ops ops; |
1732 | ops.code = MULT_EXPR; |
1733 | ops.type |
1734 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1735 | 1); |
1736 | ops.op0 = make_tree (ops.type, op0); |
1737 | ops.op1 = make_tree (ops.type, op1); |
1738 | ops.op2 = NULL_TREE; |
1739 | ops.location = loc; |
1740 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1741 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, |
1742 | mode, NULL_RTX, NULL, done_label, |
1743 | profile_probability::very_likely ()); |
1744 | goto do_error_label; |
1745 | } |
1746 | rtx_code_label *do_main_label; |
1747 | do_main_label = gen_label_rtx (); |
1748 | do_compare_rtx_and_jump (op0, const0_rtx, GE, false, mode, NULL_RTX, |
1749 | NULL, do_main_label, profile_probability::very_likely ()); |
1750 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
1751 | NULL, do_main_label, profile_probability::very_likely ()); |
1752 | expand_arith_set_overflow (lhs, target); |
1753 | emit_label (do_main_label); |
1754 | goto do_main; |
1755 | default: |
1756 | gcc_unreachable (); |
1757 | } |
1758 | } |
1759 | |
1760 | /* u1 * u2 -> sr */ |
1761 | if (uns0_p && uns1_p && !unsr_p) |
1762 | { |
1763 | if ((pos_neg0 | pos_neg1) == 1) |
1764 | { |
1765 | /* If both arguments are zero extended from narrower types, |
1766 | the MSB will be clear on both and so we can pretend it is |
1767 | a normal s1 * s2 -> sr multiplication. */ |
1768 | uns0_p = false; |
1769 | uns1_p = false; |
1770 | } |
1771 | else |
1772 | uns = true; |
1773 | /* Rest of handling of this case after res is computed. */ |
1774 | goto do_main; |
1775 | } |
1776 | |
1777 | /* s1 * u2 -> sr */ |
1778 | if (!uns0_p && uns1_p && !unsr_p) |
1779 | { |
1780 | switch (pos_neg1) |
1781 | { |
1782 | case 1: |
1783 | goto do_main; |
1784 | case 2: |
1785 | /* If (S) u2 is negative (i.e. u2 is larger than maximum of S, |
1786 | avoid the main code, just multiply and signal overflow |
1787 | unless 0 * u2 or -1 * ((U) Smin). */ |
1788 | struct separate_ops ops; |
1789 | ops.code = MULT_EXPR; |
1790 | ops.type = TREE_TYPE (arg1); |
1791 | ops.op0 = make_tree (ops.type, op0); |
1792 | ops.op1 = make_tree (ops.type, op1); |
1793 | ops.op2 = NULL_TREE; |
1794 | ops.location = loc; |
1795 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1796 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
1797 | NULL, done_label, profile_probability::very_likely ()); |
1798 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
1799 | NULL, do_error, profile_probability::very_unlikely ()); |
1800 | int prec; |
1801 | prec = GET_MODE_PRECISION (mode); |
1802 | rtx sgn; |
1803 | sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); |
1804 | do_compare_rtx_and_jump (op1, sgn, EQ, true, mode, NULL_RTX, |
1805 | NULL, done_label, profile_probability::very_likely ()); |
1806 | goto do_error_label; |
1807 | case 3: |
1808 | /* Rest of handling of this case after res is computed. */ |
1809 | goto do_main; |
1810 | default: |
1811 | gcc_unreachable (); |
1812 | } |
1813 | } |
1814 | |
1815 | /* s1 * s2 -> ur */ |
1816 | if (!uns0_p && !uns1_p && unsr_p) |
1817 | { |
1818 | rtx tem; |
1819 | switch (pos_neg0 | pos_neg1) |
1820 | { |
1821 | case 1: /* Both operands known to be non-negative. */ |
1822 | goto do_main; |
1823 | case 2: /* Both operands known to be negative. */ |
1824 | op0 = expand_unop (mode, neg_optab, op0, NULL_RTX, false); |
1825 | op1 = expand_unop (mode, neg_optab, op1, NULL_RTX, false); |
1826 | /* Avoid looking at arg0/arg1 ranges, as we've changed |
1827 | the arguments. */ |
1828 | arg0 = error_mark_node; |
1829 | arg1 = error_mark_node; |
1830 | goto do_main; |
1831 | case 3: |
1832 | if ((pos_neg0 ^ pos_neg1) == 3) |
1833 | { |
1834 | /* If one operand is known to be negative and the other |
1835 | non-negative, this overflows always, unless the non-negative |
1836 | one is 0. Just do normal multiply and set overflow |
1837 | unless one of the operands is 0. */ |
1838 | struct separate_ops ops; |
1839 | ops.code = MULT_EXPR; |
1840 | ops.type |
1841 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1842 | 1); |
1843 | ops.op0 = make_tree (ops.type, op0); |
1844 | ops.op1 = make_tree (ops.type, op1); |
1845 | ops.op2 = NULL_TREE; |
1846 | ops.location = loc; |
1847 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1848 | do_compare_rtx_and_jump (pos_neg0 == 1 ? op0 : op1, const0_rtx, EQ, |
1849 | true, mode, NULL_RTX, NULL, done_label, |
1850 | profile_probability::very_likely ()); |
1851 | goto do_error_label; |
1852 | } |
1853 | if (get_min_precision (arg: arg0, sign: SIGNED) |
1854 | + get_min_precision (arg: arg1, sign: SIGNED) <= GET_MODE_PRECISION (mode)) |
1855 | { |
1856 | /* If both operands are sign extended from narrower types and |
1857 | the sum of the two precisions is smaller or equal to the |
1858 | result precision: if both arguments are at runtime |
1859 | non-negative, maximum result will be 0x3f01 or 0x3f..f0..01 |
1860 | and there will be no overflow, if both arguments are negative, |
1861 | maximum result will be 0x40..00 and there will be no overflow |
1862 | either, if one argument is positive and the other argument |
1863 | negative, the result when treated as signed will be negative |
1864 | and there will be always overflow, and if one argument is |
1865 | zero and the other negative the result will be zero and no |
1866 | overflow. So, do |
1867 | res = (U) (s1 * s2) |
1868 | ovf = (S) res < 0 */ |
1869 | struct separate_ops ops; |
1870 | ops.code = MULT_EXPR; |
1871 | ops.type |
1872 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1873 | 1); |
1874 | ops.op0 = make_tree (ops.type, op0); |
1875 | ops.op1 = make_tree (ops.type, op1); |
1876 | ops.op2 = NULL_TREE; |
1877 | ops.location = loc; |
1878 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1879 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, |
1880 | mode, NULL_RTX, NULL, done_label, |
1881 | profile_probability::very_likely ()); |
1882 | goto do_error_label; |
1883 | } |
1884 | /* The general case, do all the needed comparisons at runtime. */ |
1885 | rtx_code_label *do_main_label, *after_negate_label; |
1886 | rtx rop0, rop1; |
1887 | rop0 = gen_reg_rtx (mode); |
1888 | rop1 = gen_reg_rtx (mode); |
1889 | emit_move_insn (rop0, op0); |
1890 | emit_move_insn (rop1, op1); |
1891 | op0 = rop0; |
1892 | op1 = rop1; |
1893 | do_main_label = gen_label_rtx (); |
1894 | after_negate_label = gen_label_rtx (); |
1895 | tem = expand_binop (mode, and_optab, op0, op1, NULL_RTX, false, |
1896 | OPTAB_LIB_WIDEN); |
1897 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1898 | NULL, after_negate_label, profile_probability::very_likely ()); |
1899 | /* Both arguments negative here, negate them and continue with |
1900 | normal unsigned overflow checking multiplication. */ |
1901 | emit_move_insn (op0, expand_unop (mode, neg_optab, op0, |
1902 | NULL_RTX, false)); |
1903 | emit_move_insn (op1, expand_unop (mode, neg_optab, op1, |
1904 | NULL_RTX, false)); |
1905 | /* Avoid looking at arg0/arg1 ranges, as we might have changed |
1906 | the arguments. */ |
1907 | arg0 = error_mark_node; |
1908 | arg1 = error_mark_node; |
1909 | emit_jump (do_main_label); |
1910 | emit_label (after_negate_label); |
1911 | tem = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1912 | OPTAB_LIB_WIDEN); |
1913 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1914 | NULL, do_main_label, |
1915 | profile_probability::very_likely ()); |
1916 | /* One argument is negative here, the other positive. This |
1917 | overflows always, unless one of the arguments is 0. But |
1918 | if e.g. s2 is 0, (U) s1 * 0 doesn't overflow, whatever s1 |
1919 | is, thus we can keep do_main code oring in overflow as is. */ |
1920 | if (pos_neg0 != 2) |
1921 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
1922 | NULL, do_main_label, |
1923 | profile_probability::very_unlikely ()); |
1924 | if (pos_neg1 != 2) |
1925 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
1926 | NULL, do_main_label, |
1927 | profile_probability::very_unlikely ()); |
1928 | expand_arith_set_overflow (lhs, target); |
1929 | emit_label (do_main_label); |
1930 | goto do_main; |
1931 | default: |
1932 | gcc_unreachable (); |
1933 | } |
1934 | } |
1935 | |
1936 | do_main: |
1937 | type = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), uns); |
1938 | sign = uns ? UNSIGNED : SIGNED; |
1939 | icode = optab_handler (op: uns ? umulv4_optab : mulv4_optab, mode); |
1940 | if (uns |
1941 | && (integer_pow2p (arg0) || integer_pow2p (arg1)) |
1942 | && (optimize_insn_for_speed_p () || icode == CODE_FOR_nothing)) |
1943 | { |
1944 | /* Optimize unsigned multiplication by power of 2 constant |
1945 | using 2 shifts, one for result, one to extract the shifted |
1946 | out bits to see if they are all zero. |
1947 | Don't do this if optimizing for size and we have umulv4_optab, |
1948 | in that case assume multiplication will be shorter. |
1949 | This is heuristics based on the single target that provides |
1950 | umulv4 right now (i?86/x86_64), if further targets add it, this |
1951 | might need to be revisited. |
1952 | Cases where both operands are constant should be folded already |
1953 | during GIMPLE, and cases where one operand is constant but not |
1954 | power of 2 are questionable, either the WIDEN_MULT_EXPR case |
1955 | below can be done without multiplication, just by shifts and adds, |
1956 | or we'd need to divide the result (and hope it actually doesn't |
1957 | really divide nor multiply) and compare the result of the division |
1958 | with the original operand. */ |
1959 | rtx opn0 = op0; |
1960 | rtx opn1 = op1; |
1961 | tree argn0 = arg0; |
1962 | tree argn1 = arg1; |
1963 | if (integer_pow2p (arg0)) |
1964 | { |
1965 | std::swap (a&: opn0, b&: opn1); |
1966 | std::swap (a&: argn0, b&: argn1); |
1967 | } |
1968 | int cnt = tree_log2 (argn1); |
1969 | if (cnt >= 0 && cnt < GET_MODE_PRECISION (mode)) |
1970 | { |
1971 | rtx upper = const0_rtx; |
1972 | res = expand_shift (LSHIFT_EXPR, mode, opn0, cnt, NULL_RTX, uns); |
1973 | if (cnt != 0) |
1974 | upper = expand_shift (RSHIFT_EXPR, mode, opn0, |
1975 | GET_MODE_PRECISION (mode) - cnt, |
1976 | NULL_RTX, uns); |
1977 | do_compare_rtx_and_jump (upper, const0_rtx, EQ, true, mode, |
1978 | NULL_RTX, NULL, done_label, |
1979 | profile_probability::very_likely ()); |
1980 | goto do_error_label; |
1981 | } |
1982 | } |
1983 | if (icode != CODE_FOR_nothing) |
1984 | { |
1985 | class expand_operand ops[4]; |
1986 | rtx_insn *last = get_last_insn (); |
1987 | |
1988 | res = gen_reg_rtx (mode); |
1989 | create_output_operand (op: &ops[0], x: res, mode); |
1990 | create_input_operand (op: &ops[1], value: op0, mode); |
1991 | create_input_operand (op: &ops[2], value: op1, mode); |
1992 | create_fixed_operand (op: &ops[3], x: do_error); |
1993 | if (maybe_expand_insn (icode, nops: 4, ops)) |
1994 | { |
1995 | last = get_last_insn (); |
1996 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1997 | && JUMP_P (last) |
1998 | && any_condjump_p (last) |
1999 | && !find_reg_note (last, REG_BR_PROB, 0)) |
2000 | add_reg_br_prob_note (last, |
2001 | profile_probability::very_unlikely ()); |
2002 | emit_jump (done_label); |
2003 | } |
2004 | else |
2005 | { |
2006 | delete_insns_since (last); |
2007 | icode = CODE_FOR_nothing; |
2008 | } |
2009 | } |
2010 | |
2011 | if (icode == CODE_FOR_nothing) |
2012 | { |
2013 | struct separate_ops ops; |
2014 | int prec = GET_MODE_PRECISION (mode); |
2015 | scalar_int_mode hmode, wmode; |
2016 | ops.op0 = make_tree (type, op0); |
2017 | ops.op1 = make_tree (type, op1); |
2018 | ops.op2 = NULL_TREE; |
2019 | ops.location = loc; |
2020 | |
2021 | /* Optimize unsigned overflow check where we don't use the |
2022 | multiplication result, just whether overflow happened. |
2023 | If we can do MULT_HIGHPART_EXPR, that followed by |
2024 | comparison of the result against zero is cheapest. |
2025 | We'll still compute res, but it should be DCEd later. */ |
2026 | use_operand_p use; |
2027 | gimple *use_stmt; |
2028 | if (!is_ubsan |
2029 | && lhs |
2030 | && uns |
2031 | && !(uns0_p && uns1_p && !unsr_p) |
2032 | && can_mult_highpart_p (mode, uns) == 1 |
2033 | && single_imm_use (var: lhs, use_p: &use, stmt: &use_stmt) |
2034 | && is_gimple_assign (gs: use_stmt) |
2035 | && gimple_assign_rhs_code (gs: use_stmt) == IMAGPART_EXPR) |
2036 | goto highpart; |
2037 | |
2038 | if (GET_MODE_2XWIDER_MODE (m: mode).exists (mode: &wmode) |
2039 | && targetm.scalar_mode_supported_p (wmode) |
2040 | && can_widen_mult_without_libcall (wmode, mode, op0, op1, uns)) |
2041 | { |
2042 | twoxwider: |
2043 | ops.code = WIDEN_MULT_EXPR; |
2044 | ops.type |
2045 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: wmode), uns); |
2046 | |
2047 | res = expand_expr_real_2 (&ops, NULL_RTX, wmode, EXPAND_NORMAL); |
2048 | rtx hipart = expand_shift (RSHIFT_EXPR, wmode, res, prec, |
2049 | NULL_RTX, uns); |
2050 | hipart = convert_modes (mode, oldmode: wmode, x: hipart, unsignedp: uns); |
2051 | res = convert_modes (mode, oldmode: wmode, x: res, unsignedp: uns); |
2052 | if (uns) |
2053 | /* For the unsigned multiplication, there was overflow if |
2054 | HIPART is non-zero. */ |
2055 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
2056 | NULL_RTX, NULL, done_label, |
2057 | profile_probability::very_likely ()); |
2058 | else |
2059 | { |
2060 | /* RES is used more than once, place it in a pseudo. */ |
2061 | res = force_reg (mode, res); |
2062 | |
2063 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
2064 | NULL_RTX, 0); |
2065 | /* RES is low half of the double width result, HIPART |
2066 | the high half. There was overflow if |
2067 | HIPART is different from RES < 0 ? -1 : 0. */ |
2068 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
2069 | NULL_RTX, NULL, done_label, |
2070 | profile_probability::very_likely ()); |
2071 | } |
2072 | } |
2073 | else if (can_mult_highpart_p (mode, uns) == 1) |
2074 | { |
2075 | highpart: |
2076 | ops.code = MULT_HIGHPART_EXPR; |
2077 | ops.type = type; |
2078 | |
2079 | rtx hipart = expand_expr_real_2 (&ops, NULL_RTX, mode, |
2080 | EXPAND_NORMAL); |
2081 | ops.code = MULT_EXPR; |
2082 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2083 | if (uns) |
2084 | /* For the unsigned multiplication, there was overflow if |
2085 | HIPART is non-zero. */ |
2086 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
2087 | NULL_RTX, NULL, done_label, |
2088 | profile_probability::very_likely ()); |
2089 | else |
2090 | { |
2091 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
2092 | NULL_RTX, 0); |
2093 | /* RES is low half of the double width result, HIPART |
2094 | the high half. There was overflow if |
2095 | HIPART is different from RES < 0 ? -1 : 0. */ |
2096 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
2097 | NULL_RTX, NULL, done_label, |
2098 | profile_probability::very_likely ()); |
2099 | } |
2100 | |
2101 | } |
2102 | else if (int_mode_for_size (size: prec / 2, limit: 1).exists (mode: &hmode) |
2103 | && 2 * GET_MODE_PRECISION (mode: hmode) == prec) |
2104 | { |
2105 | rtx_code_label *large_op0 = gen_label_rtx (); |
2106 | rtx_code_label *small_op0_large_op1 = gen_label_rtx (); |
2107 | rtx_code_label *one_small_one_large = gen_label_rtx (); |
2108 | rtx_code_label *both_ops_large = gen_label_rtx (); |
2109 | rtx_code_label *after_hipart_neg = uns ? NULL : gen_label_rtx (); |
2110 | rtx_code_label *after_lopart_neg = uns ? NULL : gen_label_rtx (); |
2111 | rtx_code_label *do_overflow = gen_label_rtx (); |
2112 | rtx_code_label *hipart_different = uns ? NULL : gen_label_rtx (); |
2113 | |
2114 | unsigned int hprec = GET_MODE_PRECISION (mode: hmode); |
2115 | rtx hipart0 = expand_shift (RSHIFT_EXPR, mode, op0, hprec, |
2116 | NULL_RTX, uns); |
2117 | hipart0 = convert_modes (mode: hmode, oldmode: mode, x: hipart0, unsignedp: uns); |
2118 | rtx lopart0 = convert_modes (mode: hmode, oldmode: mode, x: op0, unsignedp: uns); |
2119 | rtx signbit0 = const0_rtx; |
2120 | if (!uns) |
2121 | signbit0 = expand_shift (RSHIFT_EXPR, hmode, lopart0, hprec - 1, |
2122 | NULL_RTX, 0); |
2123 | rtx hipart1 = expand_shift (RSHIFT_EXPR, mode, op1, hprec, |
2124 | NULL_RTX, uns); |
2125 | hipart1 = convert_modes (mode: hmode, oldmode: mode, x: hipart1, unsignedp: uns); |
2126 | rtx lopart1 = convert_modes (mode: hmode, oldmode: mode, x: op1, unsignedp: uns); |
2127 | rtx signbit1 = const0_rtx; |
2128 | if (!uns) |
2129 | signbit1 = expand_shift (RSHIFT_EXPR, hmode, lopart1, hprec - 1, |
2130 | NULL_RTX, 0); |
2131 | |
2132 | res = gen_reg_rtx (mode); |
2133 | |
2134 | /* True if op0 resp. op1 are known to be in the range of |
2135 | halfstype. */ |
2136 | bool op0_small_p = false; |
2137 | bool op1_small_p = false; |
2138 | /* True if op0 resp. op1 are known to have all zeros or all ones |
2139 | in the upper half of bits, but are not known to be |
2140 | op{0,1}_small_p. */ |
2141 | bool op0_medium_p = false; |
2142 | bool op1_medium_p = false; |
2143 | /* -1 if op{0,1} is known to be negative, 0 if it is known to be |
2144 | nonnegative, 1 if unknown. */ |
2145 | int op0_sign = 1; |
2146 | int op1_sign = 1; |
2147 | |
2148 | if (pos_neg0 == 1) |
2149 | op0_sign = 0; |
2150 | else if (pos_neg0 == 2) |
2151 | op0_sign = -1; |
2152 | if (pos_neg1 == 1) |
2153 | op1_sign = 0; |
2154 | else if (pos_neg1 == 2) |
2155 | op1_sign = -1; |
2156 | |
2157 | unsigned int mprec0 = prec; |
2158 | if (arg0 != error_mark_node) |
2159 | mprec0 = get_min_precision (arg: arg0, sign); |
2160 | if (mprec0 <= hprec) |
2161 | op0_small_p = true; |
2162 | else if (!uns && mprec0 <= hprec + 1) |
2163 | op0_medium_p = true; |
2164 | unsigned int mprec1 = prec; |
2165 | if (arg1 != error_mark_node) |
2166 | mprec1 = get_min_precision (arg: arg1, sign); |
2167 | if (mprec1 <= hprec) |
2168 | op1_small_p = true; |
2169 | else if (!uns && mprec1 <= hprec + 1) |
2170 | op1_medium_p = true; |
2171 | |
2172 | int smaller_sign = 1; |
2173 | int larger_sign = 1; |
2174 | if (op0_small_p) |
2175 | { |
2176 | smaller_sign = op0_sign; |
2177 | larger_sign = op1_sign; |
2178 | } |
2179 | else if (op1_small_p) |
2180 | { |
2181 | smaller_sign = op1_sign; |
2182 | larger_sign = op0_sign; |
2183 | } |
2184 | else if (op0_sign == op1_sign) |
2185 | { |
2186 | smaller_sign = op0_sign; |
2187 | larger_sign = op0_sign; |
2188 | } |
2189 | |
2190 | if (!op0_small_p) |
2191 | do_compare_rtx_and_jump (signbit0, hipart0, NE, true, hmode, |
2192 | NULL_RTX, NULL, large_op0, |
2193 | profile_probability::unlikely ()); |
2194 | |
2195 | if (!op1_small_p) |
2196 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
2197 | NULL_RTX, NULL, small_op0_large_op1, |
2198 | profile_probability::unlikely ()); |
2199 | |
2200 | /* If both op0 and op1 are sign (!uns) or zero (uns) extended from |
2201 | hmode to mode, the multiplication will never overflow. We can |
2202 | do just one hmode x hmode => mode widening multiplication. */ |
2203 | tree halfstype = build_nonstandard_integer_type (hprec, uns); |
2204 | ops.op0 = make_tree (halfstype, lopart0); |
2205 | ops.op1 = make_tree (halfstype, lopart1); |
2206 | ops.code = WIDEN_MULT_EXPR; |
2207 | ops.type = type; |
2208 | rtx thisres |
2209 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2210 | emit_move_insn (res, thisres); |
2211 | emit_jump (done_label); |
2212 | |
2213 | emit_label (small_op0_large_op1); |
2214 | |
2215 | /* If op0 is sign (!uns) or zero (uns) extended from hmode to mode, |
2216 | but op1 is not, just swap the arguments and handle it as op1 |
2217 | sign/zero extended, op0 not. */ |
2218 | rtx larger = gen_reg_rtx (mode); |
2219 | rtx hipart = gen_reg_rtx (hmode); |
2220 | rtx lopart = gen_reg_rtx (hmode); |
2221 | emit_move_insn (larger, op1); |
2222 | emit_move_insn (hipart, hipart1); |
2223 | emit_move_insn (lopart, lopart0); |
2224 | emit_jump (one_small_one_large); |
2225 | |
2226 | emit_label (large_op0); |
2227 | |
2228 | if (!op1_small_p) |
2229 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
2230 | NULL_RTX, NULL, both_ops_large, |
2231 | profile_probability::unlikely ()); |
2232 | |
2233 | /* If op1 is sign (!uns) or zero (uns) extended from hmode to mode, |
2234 | but op0 is not, prepare larger, hipart and lopart pseudos and |
2235 | handle it together with small_op0_large_op1. */ |
2236 | emit_move_insn (larger, op0); |
2237 | emit_move_insn (hipart, hipart0); |
2238 | emit_move_insn (lopart, lopart1); |
2239 | |
2240 | emit_label (one_small_one_large); |
2241 | |
2242 | /* lopart is the low part of the operand that is sign extended |
2243 | to mode, larger is the other operand, hipart is the |
2244 | high part of larger and lopart0 and lopart1 are the low parts |
2245 | of both operands. |
2246 | We perform lopart0 * lopart1 and lopart * hipart widening |
2247 | multiplications. */ |
2248 | tree halfutype = build_nonstandard_integer_type (hprec, 1); |
2249 | ops.op0 = make_tree (halfutype, lopart0); |
2250 | ops.op1 = make_tree (halfutype, lopart1); |
2251 | rtx lo0xlo1 |
2252 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2253 | |
2254 | ops.op0 = make_tree (halfutype, lopart); |
2255 | ops.op1 = make_tree (halfutype, hipart); |
2256 | rtx loxhi = gen_reg_rtx (mode); |
2257 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2258 | emit_move_insn (loxhi, tem); |
2259 | |
2260 | if (!uns) |
2261 | { |
2262 | /* if (hipart < 0) loxhi -= lopart << (bitsize / 2); */ |
2263 | if (larger_sign == 0) |
2264 | emit_jump (after_hipart_neg); |
2265 | else if (larger_sign != -1) |
2266 | do_compare_rtx_and_jump (hipart, const0_rtx, GE, false, hmode, |
2267 | NULL_RTX, NULL, after_hipart_neg, |
2268 | profile_probability::even ()); |
2269 | |
2270 | tem = convert_modes (mode, oldmode: hmode, x: lopart, unsignedp: 1); |
2271 | tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1); |
2272 | tem = expand_simple_binop (mode, MINUS, loxhi, tem, NULL_RTX, |
2273 | 1, OPTAB_WIDEN); |
2274 | emit_move_insn (loxhi, tem); |
2275 | |
2276 | emit_label (after_hipart_neg); |
2277 | |
2278 | /* if (lopart < 0) loxhi -= larger; */ |
2279 | if (smaller_sign == 0) |
2280 | emit_jump (after_lopart_neg); |
2281 | else if (smaller_sign != -1) |
2282 | do_compare_rtx_and_jump (lopart, const0_rtx, GE, false, hmode, |
2283 | NULL_RTX, NULL, after_lopart_neg, |
2284 | profile_probability::even ()); |
2285 | |
2286 | tem = expand_simple_binop (mode, MINUS, loxhi, larger, NULL_RTX, |
2287 | 1, OPTAB_WIDEN); |
2288 | emit_move_insn (loxhi, tem); |
2289 | |
2290 | emit_label (after_lopart_neg); |
2291 | } |
2292 | |
2293 | /* loxhi += (uns) lo0xlo1 >> (bitsize / 2); */ |
2294 | tem = expand_shift (RSHIFT_EXPR, mode, lo0xlo1, hprec, NULL_RTX, 1); |
2295 | tem = expand_simple_binop (mode, PLUS, loxhi, tem, NULL_RTX, |
2296 | 1, OPTAB_WIDEN); |
2297 | emit_move_insn (loxhi, tem); |
2298 | |
2299 | /* if (loxhi >> (bitsize / 2) |
2300 | == (hmode) loxhi >> (bitsize / 2 - 1)) (if !uns) |
2301 | if (loxhi >> (bitsize / 2) == 0 (if uns). */ |
2302 | rtx hipartloxhi = expand_shift (RSHIFT_EXPR, mode, loxhi, hprec, |
2303 | NULL_RTX, 0); |
2304 | hipartloxhi = convert_modes (mode: hmode, oldmode: mode, x: hipartloxhi, unsignedp: 0); |
2305 | rtx signbitloxhi = const0_rtx; |
2306 | if (!uns) |
2307 | signbitloxhi = expand_shift (RSHIFT_EXPR, hmode, |
2308 | convert_modes (mode: hmode, oldmode: mode, |
2309 | x: loxhi, unsignedp: 0), |
2310 | hprec - 1, NULL_RTX, 0); |
2311 | |
2312 | do_compare_rtx_and_jump (signbitloxhi, hipartloxhi, NE, true, hmode, |
2313 | NULL_RTX, NULL, do_overflow, |
2314 | profile_probability::very_unlikely ()); |
2315 | |
2316 | /* res = (loxhi << (bitsize / 2)) | (hmode) lo0xlo1; */ |
2317 | rtx loxhishifted = expand_shift (LSHIFT_EXPR, mode, loxhi, hprec, |
2318 | NULL_RTX, 1); |
2319 | tem = convert_modes (mode, oldmode: hmode, |
2320 | x: convert_modes (mode: hmode, oldmode: mode, x: lo0xlo1, unsignedp: 1), unsignedp: 1); |
2321 | |
2322 | tem = expand_simple_binop (mode, IOR, loxhishifted, tem, res, |
2323 | 1, OPTAB_WIDEN); |
2324 | if (tem != res) |
2325 | emit_move_insn (res, tem); |
2326 | emit_jump (done_label); |
2327 | |
2328 | emit_label (both_ops_large); |
2329 | |
2330 | /* If both operands are large (not sign (!uns) or zero (uns) |
2331 | extended from hmode), then perform the full multiplication |
2332 | which will be the result of the operation. |
2333 | The only cases which don't overflow are for signed multiplication |
2334 | some cases where both hipart0 and highpart1 are 0 or -1. |
2335 | For unsigned multiplication when high parts are both non-zero |
2336 | this overflows always. */ |
2337 | ops.code = MULT_EXPR; |
2338 | ops.op0 = make_tree (type, op0); |
2339 | ops.op1 = make_tree (type, op1); |
2340 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2341 | emit_move_insn (res, tem); |
2342 | |
2343 | if (!uns) |
2344 | { |
2345 | if (!op0_medium_p) |
2346 | { |
2347 | tem = expand_simple_binop (hmode, PLUS, hipart0, const1_rtx, |
2348 | NULL_RTX, 1, OPTAB_WIDEN); |
2349 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
2350 | NULL_RTX, NULL, do_error, |
2351 | profile_probability::very_unlikely ()); |
2352 | } |
2353 | |
2354 | if (!op1_medium_p) |
2355 | { |
2356 | tem = expand_simple_binop (hmode, PLUS, hipart1, const1_rtx, |
2357 | NULL_RTX, 1, OPTAB_WIDEN); |
2358 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
2359 | NULL_RTX, NULL, do_error, |
2360 | profile_probability::very_unlikely ()); |
2361 | } |
2362 | |
2363 | /* At this point hipart{0,1} are both in [-1, 0]. If they are |
2364 | the same, overflow happened if res is non-positive, if they |
2365 | are different, overflow happened if res is positive. */ |
2366 | if (op0_sign != 1 && op1_sign != 1 && op0_sign != op1_sign) |
2367 | emit_jump (hipart_different); |
2368 | else if (op0_sign == 1 || op1_sign == 1) |
2369 | do_compare_rtx_and_jump (hipart0, hipart1, NE, true, hmode, |
2370 | NULL_RTX, NULL, hipart_different, |
2371 | profile_probability::even ()); |
2372 | |
2373 | do_compare_rtx_and_jump (res, const0_rtx, LE, false, mode, |
2374 | NULL_RTX, NULL, do_error, |
2375 | profile_probability::very_unlikely ()); |
2376 | emit_jump (done_label); |
2377 | |
2378 | emit_label (hipart_different); |
2379 | |
2380 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, |
2381 | NULL_RTX, NULL, do_error, |
2382 | profile_probability::very_unlikely ()); |
2383 | emit_jump (done_label); |
2384 | } |
2385 | |
2386 | emit_label (do_overflow); |
2387 | |
2388 | /* Overflow, do full multiplication and fallthru into do_error. */ |
2389 | ops.op0 = make_tree (type, op0); |
2390 | ops.op1 = make_tree (type, op1); |
2391 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2392 | emit_move_insn (res, tem); |
2393 | } |
2394 | else if (GET_MODE_2XWIDER_MODE (m: mode).exists (mode: &wmode) |
2395 | && targetm.scalar_mode_supported_p (wmode)) |
2396 | /* Even emitting a libcall is better than not detecting overflow |
2397 | at all. */ |
2398 | goto twoxwider; |
2399 | else |
2400 | { |
2401 | gcc_assert (!is_ubsan); |
2402 | ops.code = MULT_EXPR; |
2403 | ops.type = type; |
2404 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2405 | emit_jump (done_label); |
2406 | } |
2407 | } |
2408 | |
2409 | do_error_label: |
2410 | emit_label (do_error); |
2411 | if (is_ubsan) |
2412 | { |
2413 | /* Expand the ubsan builtin call. */ |
2414 | push_temp_slots (); |
2415 | fn = ubsan_build_overflow_builtin (MULT_EXPR, loc, TREE_TYPE (arg0), |
2416 | arg0, arg1, datap); |
2417 | expand_normal (exp: fn); |
2418 | pop_temp_slots (); |
2419 | do_pending_stack_adjust (); |
2420 | } |
2421 | else if (lhs) |
2422 | expand_arith_set_overflow (lhs, target); |
2423 | |
2424 | /* We're done. */ |
2425 | emit_label (done_label); |
2426 | |
2427 | /* u1 * u2 -> sr */ |
2428 | if (uns0_p && uns1_p && !unsr_p) |
2429 | { |
2430 | rtx_code_label *all_done_label = gen_label_rtx (); |
2431 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
2432 | NULL, all_done_label, profile_probability::very_likely ()); |
2433 | expand_arith_set_overflow (lhs, target); |
2434 | emit_label (all_done_label); |
2435 | } |
2436 | |
2437 | /* s1 * u2 -> sr */ |
2438 | if (!uns0_p && uns1_p && !unsr_p && pos_neg1 == 3) |
2439 | { |
2440 | rtx_code_label *all_done_label = gen_label_rtx (); |
2441 | rtx_code_label *set_noovf = gen_label_rtx (); |
2442 | do_compare_rtx_and_jump (op1, const0_rtx, GE, false, mode, NULL_RTX, |
2443 | NULL, all_done_label, profile_probability::very_likely ()); |
2444 | expand_arith_set_overflow (lhs, target); |
2445 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
2446 | NULL, set_noovf, profile_probability::very_likely ()); |
2447 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
2448 | NULL, all_done_label, profile_probability::very_unlikely ()); |
2449 | do_compare_rtx_and_jump (op1, res, NE, true, mode, NULL_RTX, NULL, |
2450 | all_done_label, profile_probability::very_unlikely ()); |
2451 | emit_label (set_noovf); |
2452 | write_complex_part (target, const0_rtx, true, false); |
2453 | emit_label (all_done_label); |
2454 | } |
2455 | |
2456 | if (lhs) |
2457 | { |
2458 | if (is_ubsan) |
2459 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
2460 | else |
2461 | expand_arith_overflow_result_store (lhs, target, mode, res); |
2462 | } |
2463 | } |
2464 | |
2465 | /* Expand UBSAN_CHECK_* internal function if it has vector operands. */ |
2466 | |
2467 | static void |
2468 | expand_vector_ubsan_overflow (location_t loc, enum tree_code code, tree lhs, |
2469 | tree arg0, tree arg1) |
2470 | { |
2471 | poly_uint64 cnt = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)); |
2472 | rtx_code_label *loop_lab = NULL; |
2473 | rtx cntvar = NULL_RTX; |
2474 | tree cntv = NULL_TREE; |
2475 | tree eltype = TREE_TYPE (TREE_TYPE (arg0)); |
2476 | tree sz = TYPE_SIZE (eltype); |
2477 | tree data = NULL_TREE; |
2478 | tree resv = NULL_TREE; |
2479 | rtx lhsr = NULL_RTX; |
2480 | rtx resvr = NULL_RTX; |
2481 | unsigned HOST_WIDE_INT const_cnt = 0; |
2482 | bool use_loop_p = (!cnt.is_constant (const_value: &const_cnt) || const_cnt > 4); |
2483 | |
2484 | if (lhs) |
2485 | { |
2486 | optab op; |
2487 | lhsr = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2488 | if (!VECTOR_MODE_P (GET_MODE (lhsr)) |
2489 | || (op = optab_for_tree_code (code, TREE_TYPE (arg0), |
2490 | optab_default)) == unknown_optab |
2491 | || (optab_handler (op, TYPE_MODE (TREE_TYPE (arg0))) |
2492 | == CODE_FOR_nothing)) |
2493 | { |
2494 | if (MEM_P (lhsr)) |
2495 | resv = make_tree (TREE_TYPE (lhs), lhsr); |
2496 | else |
2497 | { |
2498 | resvr = assign_temp (TREE_TYPE (lhs), 1, 1); |
2499 | resv = make_tree (TREE_TYPE (lhs), resvr); |
2500 | } |
2501 | } |
2502 | } |
2503 | if (use_loop_p) |
2504 | { |
2505 | do_pending_stack_adjust (); |
2506 | loop_lab = gen_label_rtx (); |
2507 | cntvar = gen_reg_rtx (TYPE_MODE (sizetype)); |
2508 | cntv = make_tree (sizetype, cntvar); |
2509 | emit_move_insn (cntvar, const0_rtx); |
2510 | emit_label (loop_lab); |
2511 | } |
2512 | if (TREE_CODE (arg0) != VECTOR_CST) |
2513 | { |
2514 | rtx arg0r = expand_normal (exp: arg0); |
2515 | arg0 = make_tree (TREE_TYPE (arg0), arg0r); |
2516 | } |
2517 | if (TREE_CODE (arg1) != VECTOR_CST) |
2518 | { |
2519 | rtx arg1r = expand_normal (exp: arg1); |
2520 | arg1 = make_tree (TREE_TYPE (arg1), arg1r); |
2521 | } |
2522 | for (unsigned int i = 0; i < (use_loop_p ? 1 : const_cnt); i++) |
2523 | { |
2524 | tree op0, op1, res = NULL_TREE; |
2525 | if (use_loop_p) |
2526 | { |
2527 | tree atype = build_array_type_nelts (eltype, cnt); |
2528 | op0 = uniform_vector_p (arg0); |
2529 | if (op0 == NULL_TREE) |
2530 | { |
2531 | op0 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg0); |
2532 | op0 = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: op0, arg1: cntv, |
2533 | NULL_TREE, NULL_TREE); |
2534 | } |
2535 | op1 = uniform_vector_p (arg1); |
2536 | if (op1 == NULL_TREE) |
2537 | { |
2538 | op1 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg1); |
2539 | op1 = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: op1, arg1: cntv, |
2540 | NULL_TREE, NULL_TREE); |
2541 | } |
2542 | if (resv) |
2543 | { |
2544 | res = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, resv); |
2545 | res = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: res, arg1: cntv, |
2546 | NULL_TREE, NULL_TREE); |
2547 | } |
2548 | } |
2549 | else |
2550 | { |
2551 | tree bitpos = bitsize_int (tree_to_uhwi (sz) * i); |
2552 | op0 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg0, sz, bitpos); |
2553 | op1 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg1, sz, bitpos); |
2554 | if (resv) |
2555 | res = fold_build3_loc (loc, BIT_FIELD_REF, eltype, resv, sz, |
2556 | bitpos); |
2557 | } |
2558 | switch (code) |
2559 | { |
2560 | case PLUS_EXPR: |
2561 | expand_addsub_overflow (loc, code: PLUS_EXPR, lhs: res, arg0: op0, arg1: op1, |
2562 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, datap: &data); |
2563 | break; |
2564 | case MINUS_EXPR: |
2565 | if (use_loop_p ? integer_zerop (arg0) : integer_zerop (op0)) |
2566 | expand_neg_overflow (loc, lhs: res, arg1: op1, is_ubsan: true, datap: &data); |
2567 | else |
2568 | expand_addsub_overflow (loc, code: MINUS_EXPR, lhs: res, arg0: op0, arg1: op1, |
2569 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, datap: &data); |
2570 | break; |
2571 | case MULT_EXPR: |
2572 | expand_mul_overflow (loc, lhs: res, arg0: op0, arg1: op1, unsr_p: false, uns0_p: false, uns1_p: false, |
2573 | is_ubsan: true, datap: &data); |
2574 | break; |
2575 | default: |
2576 | gcc_unreachable (); |
2577 | } |
2578 | } |
2579 | if (use_loop_p) |
2580 | { |
2581 | struct separate_ops ops; |
2582 | ops.code = PLUS_EXPR; |
2583 | ops.type = TREE_TYPE (cntv); |
2584 | ops.op0 = cntv; |
2585 | ops.op1 = build_int_cst (TREE_TYPE (cntv), 1); |
2586 | ops.op2 = NULL_TREE; |
2587 | ops.location = loc; |
2588 | rtx ret = expand_expr_real_2 (&ops, cntvar, TYPE_MODE (sizetype), |
2589 | EXPAND_NORMAL); |
2590 | if (ret != cntvar) |
2591 | emit_move_insn (cntvar, ret); |
2592 | rtx cntrtx = gen_int_mode (cnt, TYPE_MODE (sizetype)); |
2593 | do_compare_rtx_and_jump (cntvar, cntrtx, NE, false, |
2594 | TYPE_MODE (sizetype), NULL_RTX, NULL, loop_lab, |
2595 | profile_probability::very_likely ()); |
2596 | } |
2597 | if (lhs && resv == NULL_TREE) |
2598 | { |
2599 | struct separate_ops ops; |
2600 | ops.code = code; |
2601 | ops.type = TREE_TYPE (arg0); |
2602 | ops.op0 = arg0; |
2603 | ops.op1 = arg1; |
2604 | ops.op2 = NULL_TREE; |
2605 | ops.location = loc; |
2606 | rtx ret = expand_expr_real_2 (&ops, lhsr, TYPE_MODE (TREE_TYPE (arg0)), |
2607 | EXPAND_NORMAL); |
2608 | if (ret != lhsr) |
2609 | emit_move_insn (lhsr, ret); |
2610 | } |
2611 | else if (resvr) |
2612 | emit_move_insn (lhsr, resvr); |
2613 | } |
2614 | |
2615 | /* Expand UBSAN_CHECK_ADD call STMT. */ |
2616 | |
2617 | static void |
2618 | expand_UBSAN_CHECK_ADD (internal_fn, gcall *stmt) |
2619 | { |
2620 | location_t loc = gimple_location (g: stmt); |
2621 | tree lhs = gimple_call_lhs (gs: stmt); |
2622 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2623 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2624 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2625 | expand_vector_ubsan_overflow (loc, code: PLUS_EXPR, lhs, arg0, arg1); |
2626 | else |
2627 | expand_addsub_overflow (loc, code: PLUS_EXPR, lhs, arg0, arg1, |
2628 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, NULL); |
2629 | } |
2630 | |
2631 | /* Expand UBSAN_CHECK_SUB call STMT. */ |
2632 | |
2633 | static void |
2634 | expand_UBSAN_CHECK_SUB (internal_fn, gcall *stmt) |
2635 | { |
2636 | location_t loc = gimple_location (g: stmt); |
2637 | tree lhs = gimple_call_lhs (gs: stmt); |
2638 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2639 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2640 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2641 | expand_vector_ubsan_overflow (loc, code: MINUS_EXPR, lhs, arg0, arg1); |
2642 | else if (integer_zerop (arg0)) |
2643 | expand_neg_overflow (loc, lhs, arg1, is_ubsan: true, NULL); |
2644 | else |
2645 | expand_addsub_overflow (loc, code: MINUS_EXPR, lhs, arg0, arg1, |
2646 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, NULL); |
2647 | } |
2648 | |
2649 | /* Expand UBSAN_CHECK_MUL call STMT. */ |
2650 | |
2651 | static void |
2652 | expand_UBSAN_CHECK_MUL (internal_fn, gcall *stmt) |
2653 | { |
2654 | location_t loc = gimple_location (g: stmt); |
2655 | tree lhs = gimple_call_lhs (gs: stmt); |
2656 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2657 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2658 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2659 | expand_vector_ubsan_overflow (loc, code: MULT_EXPR, lhs, arg0, arg1); |
2660 | else |
2661 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, |
2662 | NULL); |
2663 | } |
2664 | |
2665 | /* Helper function for {ADD,SUB,MUL}_OVERFLOW call stmt expansion. */ |
2666 | |
2667 | static void |
2668 | expand_arith_overflow (enum tree_code code, gimple *stmt) |
2669 | { |
2670 | tree lhs = gimple_call_lhs (gs: stmt); |
2671 | if (lhs == NULL_TREE) |
2672 | return; |
2673 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2674 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2675 | tree type = TREE_TYPE (TREE_TYPE (lhs)); |
2676 | int uns0_p = TYPE_UNSIGNED (TREE_TYPE (arg0)); |
2677 | int uns1_p = TYPE_UNSIGNED (TREE_TYPE (arg1)); |
2678 | int unsr_p = TYPE_UNSIGNED (type); |
2679 | int prec0 = TYPE_PRECISION (TREE_TYPE (arg0)); |
2680 | int prec1 = TYPE_PRECISION (TREE_TYPE (arg1)); |
2681 | int precres = TYPE_PRECISION (type); |
2682 | location_t loc = gimple_location (g: stmt); |
2683 | if (!uns0_p && get_range_pos_neg (arg0) == 1) |
2684 | uns0_p = true; |
2685 | if (!uns1_p && get_range_pos_neg (arg1) == 1) |
2686 | uns1_p = true; |
2687 | int pr = get_min_precision (arg: arg0, sign: uns0_p ? UNSIGNED : SIGNED); |
2688 | prec0 = MIN (prec0, pr); |
2689 | pr = get_min_precision (arg: arg1, sign: uns1_p ? UNSIGNED : SIGNED); |
2690 | prec1 = MIN (prec1, pr); |
2691 | |
2692 | /* If uns0_p && uns1_p, precop is minimum needed precision |
2693 | of unsigned type to hold the exact result, otherwise |
2694 | precop is minimum needed precision of signed type to |
2695 | hold the exact result. */ |
2696 | int precop; |
2697 | if (code == MULT_EXPR) |
2698 | precop = prec0 + prec1 + (uns0_p != uns1_p); |
2699 | else |
2700 | { |
2701 | if (uns0_p == uns1_p) |
2702 | precop = MAX (prec0, prec1) + 1; |
2703 | else if (uns0_p) |
2704 | precop = MAX (prec0 + 1, prec1) + 1; |
2705 | else |
2706 | precop = MAX (prec0, prec1 + 1) + 1; |
2707 | } |
2708 | int orig_precres = precres; |
2709 | |
2710 | do |
2711 | { |
2712 | if ((uns0_p && uns1_p) |
2713 | ? ((precop + !unsr_p) <= precres |
2714 | /* u1 - u2 -> ur can overflow, no matter what precision |
2715 | the result has. */ |
2716 | && (code != MINUS_EXPR || !unsr_p)) |
2717 | : (!unsr_p && precop <= precres)) |
2718 | { |
2719 | /* The infinity precision result will always fit into result. */ |
2720 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2721 | write_complex_part (target, const0_rtx, true, false); |
2722 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (type); |
2723 | struct separate_ops ops; |
2724 | ops.code = code; |
2725 | ops.type = type; |
2726 | ops.op0 = fold_convert_loc (loc, type, arg0); |
2727 | ops.op1 = fold_convert_loc (loc, type, arg1); |
2728 | ops.op2 = NULL_TREE; |
2729 | ops.location = loc; |
2730 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2731 | expand_arith_overflow_result_store (lhs, target, mode, res: tem); |
2732 | return; |
2733 | } |
2734 | |
2735 | /* For operations with low precision, if target doesn't have them, start |
2736 | with precres widening right away, otherwise do it only if the most |
2737 | simple cases can't be used. */ |
2738 | const int min_precision = targetm.min_arithmetic_precision (); |
2739 | if (orig_precres == precres && precres < min_precision) |
2740 | ; |
2741 | else if ((uns0_p && uns1_p && unsr_p && prec0 <= precres |
2742 | && prec1 <= precres) |
2743 | || ((!uns0_p || !uns1_p) && !unsr_p |
2744 | && prec0 + uns0_p <= precres |
2745 | && prec1 + uns1_p <= precres)) |
2746 | { |
2747 | arg0 = fold_convert_loc (loc, type, arg0); |
2748 | arg1 = fold_convert_loc (loc, type, arg1); |
2749 | switch (code) |
2750 | { |
2751 | case MINUS_EXPR: |
2752 | if (integer_zerop (arg0) && !unsr_p) |
2753 | { |
2754 | expand_neg_overflow (loc, lhs, arg1, is_ubsan: false, NULL); |
2755 | return; |
2756 | } |
2757 | /* FALLTHRU */ |
2758 | case PLUS_EXPR: |
2759 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2760 | uns0_p: unsr_p, uns1_p: unsr_p, is_ubsan: false, NULL); |
2761 | return; |
2762 | case MULT_EXPR: |
2763 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2764 | uns0_p: unsr_p, uns1_p: unsr_p, is_ubsan: false, NULL); |
2765 | return; |
2766 | default: |
2767 | gcc_unreachable (); |
2768 | } |
2769 | } |
2770 | |
2771 | /* For sub-word operations, retry with a wider type first. */ |
2772 | if (orig_precres == precres && precop <= BITS_PER_WORD) |
2773 | { |
2774 | int p = MAX (min_precision, precop); |
2775 | scalar_int_mode m = smallest_int_mode_for_size (size: p); |
2776 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: 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 | if (prec0 <= precres && prec1 <= precres) |
2790 | { |
2791 | tree types[2]; |
2792 | if (unsr_p) |
2793 | { |
2794 | types[0] = build_nonstandard_integer_type (precres, 0); |
2795 | types[1] = type; |
2796 | } |
2797 | else |
2798 | { |
2799 | types[0] = type; |
2800 | types[1] = build_nonstandard_integer_type (precres, 1); |
2801 | } |
2802 | arg0 = fold_convert_loc (loc, types[uns0_p], arg0); |
2803 | arg1 = fold_convert_loc (loc, types[uns1_p], arg1); |
2804 | if (code != MULT_EXPR) |
2805 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2806 | uns0_p, uns1_p, is_ubsan: false, NULL); |
2807 | else |
2808 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2809 | uns0_p, uns1_p, is_ubsan: false, NULL); |
2810 | return; |
2811 | } |
2812 | |
2813 | /* Retry with a wider type. */ |
2814 | if (orig_precres == precres) |
2815 | { |
2816 | int p = MAX (prec0, prec1); |
2817 | scalar_int_mode m = smallest_int_mode_for_size (size: p); |
2818 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: m), |
2819 | uns0_p && uns1_p |
2820 | && unsr_p); |
2821 | p = TYPE_PRECISION (optype); |
2822 | if (p > precres) |
2823 | { |
2824 | precres = p; |
2825 | unsr_p = TYPE_UNSIGNED (optype); |
2826 | type = optype; |
2827 | continue; |
2828 | } |
2829 | } |
2830 | |
2831 | gcc_unreachable (); |
2832 | } |
2833 | while (1); |
2834 | } |
2835 | |
2836 | /* Expand ADD_OVERFLOW STMT. */ |
2837 | |
2838 | static void |
2839 | expand_ADD_OVERFLOW (internal_fn, gcall *stmt) |
2840 | { |
2841 | expand_arith_overflow (code: PLUS_EXPR, stmt); |
2842 | } |
2843 | |
2844 | /* Expand SUB_OVERFLOW STMT. */ |
2845 | |
2846 | static void |
2847 | expand_SUB_OVERFLOW (internal_fn, gcall *stmt) |
2848 | { |
2849 | expand_arith_overflow (code: MINUS_EXPR, stmt); |
2850 | } |
2851 | |
2852 | /* Expand MUL_OVERFLOW STMT. */ |
2853 | |
2854 | static void |
2855 | expand_MUL_OVERFLOW (internal_fn, gcall *stmt) |
2856 | { |
2857 | expand_arith_overflow (code: MULT_EXPR, stmt); |
2858 | } |
2859 | |
2860 | /* Expand UADDC STMT. */ |
2861 | |
2862 | static void |
2863 | expand_UADDC (internal_fn ifn, gcall *stmt) |
2864 | { |
2865 | tree lhs = gimple_call_lhs (gs: stmt); |
2866 | tree arg1 = gimple_call_arg (gs: stmt, index: 0); |
2867 | tree arg2 = gimple_call_arg (gs: stmt, index: 1); |
2868 | tree arg3 = gimple_call_arg (gs: stmt, index: 2); |
2869 | tree type = TREE_TYPE (arg1); |
2870 | machine_mode mode = TYPE_MODE (type); |
2871 | insn_code icode = optab_handler (op: ifn == IFN_UADDC |
2872 | ? uaddc5_optab : usubc5_optab, mode); |
2873 | rtx op1 = expand_normal (exp: arg1); |
2874 | rtx op2 = expand_normal (exp: arg2); |
2875 | rtx op3 = expand_normal (exp: arg3); |
2876 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2877 | rtx re = gen_reg_rtx (mode); |
2878 | rtx im = gen_reg_rtx (mode); |
2879 | class expand_operand ops[5]; |
2880 | create_output_operand (op: &ops[0], x: re, mode); |
2881 | create_output_operand (op: &ops[1], x: im, mode); |
2882 | create_input_operand (op: &ops[2], value: op1, mode); |
2883 | create_input_operand (op: &ops[3], value: op2, mode); |
2884 | create_input_operand (op: &ops[4], value: op3, mode); |
2885 | expand_insn (icode, nops: 5, ops); |
2886 | write_complex_part (target, re, false, false); |
2887 | write_complex_part (target, im, true, false); |
2888 | } |
2889 | |
2890 | /* Expand USUBC STMT. */ |
2891 | |
2892 | static void |
2893 | expand_USUBC (internal_fn ifn, gcall *stmt) |
2894 | { |
2895 | expand_UADDC (ifn, stmt); |
2896 | } |
2897 | |
2898 | /* This should get folded in tree-vectorizer.cc. */ |
2899 | |
2900 | static void |
2901 | expand_LOOP_VECTORIZED (internal_fn, gcall *) |
2902 | { |
2903 | gcc_unreachable (); |
2904 | } |
2905 | |
2906 | /* This should get folded in tree-vectorizer.cc. */ |
2907 | |
2908 | static void |
2909 | expand_LOOP_DIST_ALIAS (internal_fn, gcall *) |
2910 | { |
2911 | gcc_unreachable (); |
2912 | } |
2913 | |
2914 | /* Return a memory reference of type TYPE for argument INDEX of STMT. |
2915 | Use argument INDEX + 1 to derive the second (TBAA) operand. */ |
2916 | |
2917 | static tree |
2918 | expand_call_mem_ref (tree type, gcall *stmt, int index) |
2919 | { |
2920 | tree addr = gimple_call_arg (gs: stmt, index); |
2921 | tree alias_ptr_type = TREE_TYPE (gimple_call_arg (stmt, index + 1)); |
2922 | unsigned int align = tree_to_shwi (gimple_call_arg (gs: stmt, index: index + 1)); |
2923 | if (TYPE_ALIGN (type) != align) |
2924 | type = build_aligned_type (type, align); |
2925 | |
2926 | tree tmp = addr; |
2927 | if (TREE_CODE (tmp) == SSA_NAME) |
2928 | { |
2929 | gimple *def = SSA_NAME_DEF_STMT (tmp); |
2930 | if (gimple_assign_single_p (gs: def)) |
2931 | tmp = gimple_assign_rhs1 (gs: def); |
2932 | } |
2933 | |
2934 | if (TREE_CODE (tmp) == ADDR_EXPR) |
2935 | { |
2936 | tree mem = TREE_OPERAND (tmp, 0); |
2937 | if (TREE_CODE (mem) == TARGET_MEM_REF |
2938 | && types_compatible_p (TREE_TYPE (mem), type2: type)) |
2939 | { |
2940 | tree offset = TMR_OFFSET (mem); |
2941 | if (type != TREE_TYPE (mem) |
2942 | || alias_ptr_type != TREE_TYPE (offset) |
2943 | || !integer_zerop (offset)) |
2944 | { |
2945 | mem = copy_node (mem); |
2946 | TMR_OFFSET (mem) = wide_int_to_tree (type: alias_ptr_type, |
2947 | cst: wi::to_poly_wide (t: offset)); |
2948 | TREE_TYPE (mem) = type; |
2949 | } |
2950 | return mem; |
2951 | } |
2952 | } |
2953 | |
2954 | return fold_build2 (MEM_REF, type, addr, build_int_cst (alias_ptr_type, 0)); |
2955 | } |
2956 | |
2957 | /* Expand MASK_LOAD{,_LANES}, MASK_LEN_LOAD or LEN_LOAD call STMT using optab |
2958 | * OPTAB. */ |
2959 | |
2960 | static void |
2961 | expand_partial_load_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
2962 | { |
2963 | int i = 0; |
2964 | class expand_operand ops[5]; |
2965 | tree type, lhs, rhs, maskt; |
2966 | rtx mem, target; |
2967 | insn_code icode; |
2968 | |
2969 | maskt = gimple_call_arg (gs: stmt, index: internal_fn_mask_index (ifn)); |
2970 | lhs = gimple_call_lhs (gs: stmt); |
2971 | if (lhs == NULL_TREE) |
2972 | return; |
2973 | type = TREE_TYPE (lhs); |
2974 | rhs = expand_call_mem_ref (type, stmt, index: 0); |
2975 | |
2976 | if (optab == vec_mask_load_lanes_optab |
2977 | || optab == vec_mask_len_load_lanes_optab) |
2978 | icode = get_multi_vector_move (array_type: type, optab); |
2979 | else if (optab == len_load_optab) |
2980 | icode = direct_optab_handler (op: optab, TYPE_MODE (type)); |
2981 | else |
2982 | icode = convert_optab_handler (op: optab, TYPE_MODE (type), |
2983 | TYPE_MODE (TREE_TYPE (maskt))); |
2984 | |
2985 | mem = expand_expr (exp: rhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2986 | gcc_assert (MEM_P (mem)); |
2987 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2988 | create_output_operand (op: &ops[i++], x: target, TYPE_MODE (type)); |
2989 | create_fixed_operand (op: &ops[i++], x: mem); |
2990 | i = add_mask_and_len_args (ops, opno: i, stmt); |
2991 | expand_insn (icode, nops: i, ops); |
2992 | |
2993 | if (!rtx_equal_p (target, ops[0].value)) |
2994 | emit_move_insn (target, ops[0].value); |
2995 | } |
2996 | |
2997 | #define expand_mask_load_optab_fn expand_partial_load_optab_fn |
2998 | #define expand_mask_load_lanes_optab_fn expand_mask_load_optab_fn |
2999 | #define expand_len_load_optab_fn expand_partial_load_optab_fn |
3000 | #define expand_mask_len_load_optab_fn expand_partial_load_optab_fn |
3001 | |
3002 | /* Expand MASK_STORE{,_LANES}, MASK_LEN_STORE or LEN_STORE call STMT using optab |
3003 | * OPTAB. */ |
3004 | |
3005 | static void |
3006 | expand_partial_store_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
3007 | { |
3008 | int i = 0; |
3009 | class expand_operand ops[5]; |
3010 | tree type, lhs, rhs, maskt; |
3011 | rtx mem, reg; |
3012 | insn_code icode; |
3013 | |
3014 | maskt = gimple_call_arg (gs: stmt, index: internal_fn_mask_index (ifn)); |
3015 | rhs = gimple_call_arg (gs: stmt, index: internal_fn_stored_value_index (ifn)); |
3016 | type = TREE_TYPE (rhs); |
3017 | lhs = expand_call_mem_ref (type, stmt, index: 0); |
3018 | |
3019 | if (optab == vec_mask_store_lanes_optab |
3020 | || optab == vec_mask_len_store_lanes_optab) |
3021 | icode = get_multi_vector_move (array_type: type, optab); |
3022 | else if (optab == len_store_optab) |
3023 | icode = direct_optab_handler (op: optab, TYPE_MODE (type)); |
3024 | else |
3025 | icode = convert_optab_handler (op: optab, TYPE_MODE (type), |
3026 | TYPE_MODE (TREE_TYPE (maskt))); |
3027 | |
3028 | mem = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3029 | gcc_assert (MEM_P (mem)); |
3030 | reg = expand_normal (exp: rhs); |
3031 | create_fixed_operand (op: &ops[i++], x: mem); |
3032 | create_input_operand (op: &ops[i++], value: reg, TYPE_MODE (type)); |
3033 | i = add_mask_and_len_args (ops, opno: i, stmt); |
3034 | expand_insn (icode, nops: i, ops); |
3035 | } |
3036 | |
3037 | #define expand_mask_store_optab_fn expand_partial_store_optab_fn |
3038 | #define expand_mask_store_lanes_optab_fn expand_mask_store_optab_fn |
3039 | #define expand_len_store_optab_fn expand_partial_store_optab_fn |
3040 | #define expand_mask_len_store_optab_fn expand_partial_store_optab_fn |
3041 | |
3042 | /* Expand VCOND, VCONDU and VCONDEQ optab internal functions. |
3043 | The expansion of STMT happens based on OPTAB table associated. */ |
3044 | |
3045 | static void |
3046 | expand_vec_cond_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3047 | { |
3048 | class expand_operand ops[6]; |
3049 | insn_code icode; |
3050 | tree lhs = gimple_call_lhs (gs: stmt); |
3051 | tree op0a = gimple_call_arg (gs: stmt, index: 0); |
3052 | tree op0b = gimple_call_arg (gs: stmt, index: 1); |
3053 | tree op1 = gimple_call_arg (gs: stmt, index: 2); |
3054 | tree op2 = gimple_call_arg (gs: stmt, index: 3); |
3055 | enum tree_code tcode = (tree_code) int_cst_value (gimple_call_arg (gs: stmt, index: 4)); |
3056 | |
3057 | tree vec_cond_type = TREE_TYPE (lhs); |
3058 | tree op_mode = TREE_TYPE (op0a); |
3059 | bool unsignedp = TYPE_UNSIGNED (op_mode); |
3060 | |
3061 | machine_mode mode = TYPE_MODE (vec_cond_type); |
3062 | machine_mode cmp_op_mode = TYPE_MODE (op_mode); |
3063 | |
3064 | icode = convert_optab_handler (op: optab, to_mode: mode, from_mode: cmp_op_mode); |
3065 | rtx comparison |
3066 | = vector_compare_rtx (VOIDmode, tcode, t_op0: op0a, t_op1: op0b, unsignedp, icode, opno: 4); |
3067 | /* vector_compare_rtx legitimizes operands, preserve equality when |
3068 | expanding op1/op2. */ |
3069 | rtx rtx_op1, rtx_op2; |
3070 | if (operand_equal_p (op1, op0a)) |
3071 | rtx_op1 = XEXP (comparison, 0); |
3072 | else if (operand_equal_p (op1, op0b)) |
3073 | rtx_op1 = XEXP (comparison, 1); |
3074 | else |
3075 | rtx_op1 = expand_normal (exp: op1); |
3076 | if (operand_equal_p (op2, op0a)) |
3077 | rtx_op2 = XEXP (comparison, 0); |
3078 | else if (operand_equal_p (op2, op0b)) |
3079 | rtx_op2 = XEXP (comparison, 1); |
3080 | else |
3081 | rtx_op2 = expand_normal (exp: op2); |
3082 | |
3083 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3084 | create_output_operand (op: &ops[0], x: target, mode); |
3085 | create_input_operand (op: &ops[1], value: rtx_op1, mode); |
3086 | create_input_operand (op: &ops[2], value: rtx_op2, mode); |
3087 | create_fixed_operand (op: &ops[3], x: comparison); |
3088 | create_fixed_operand (op: &ops[4], XEXP (comparison, 0)); |
3089 | create_fixed_operand (op: &ops[5], XEXP (comparison, 1)); |
3090 | expand_insn (icode, nops: 6, ops); |
3091 | if (!rtx_equal_p (ops[0].value, target)) |
3092 | emit_move_insn (target, ops[0].value); |
3093 | } |
3094 | |
3095 | /* Expand VCOND_MASK optab internal function. |
3096 | The expansion of STMT happens based on OPTAB table associated. */ |
3097 | |
3098 | static void |
3099 | expand_vec_cond_mask_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3100 | { |
3101 | class expand_operand ops[4]; |
3102 | |
3103 | tree lhs = gimple_call_lhs (gs: stmt); |
3104 | tree op0 = gimple_call_arg (gs: stmt, index: 0); |
3105 | tree op1 = gimple_call_arg (gs: stmt, index: 1); |
3106 | tree op2 = gimple_call_arg (gs: stmt, index: 2); |
3107 | tree vec_cond_type = TREE_TYPE (lhs); |
3108 | |
3109 | machine_mode mode = TYPE_MODE (vec_cond_type); |
3110 | machine_mode mask_mode = TYPE_MODE (TREE_TYPE (op0)); |
3111 | enum insn_code icode = convert_optab_handler (op: optab, to_mode: mode, from_mode: mask_mode); |
3112 | rtx mask, rtx_op1, rtx_op2; |
3113 | |
3114 | gcc_assert (icode != CODE_FOR_nothing); |
3115 | |
3116 | mask = expand_normal (exp: op0); |
3117 | rtx_op1 = expand_normal (exp: op1); |
3118 | rtx_op2 = expand_normal (exp: op2); |
3119 | |
3120 | mask = force_reg (mask_mode, mask); |
3121 | rtx_op1 = force_reg (mode, rtx_op1); |
3122 | |
3123 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3124 | create_output_operand (op: &ops[0], x: target, mode); |
3125 | create_input_operand (op: &ops[1], value: rtx_op1, mode); |
3126 | create_input_operand (op: &ops[2], value: rtx_op2, mode); |
3127 | create_input_operand (op: &ops[3], value: mask, mode: mask_mode); |
3128 | expand_insn (icode, nops: 4, ops); |
3129 | if (!rtx_equal_p (ops[0].value, target)) |
3130 | emit_move_insn (target, ops[0].value); |
3131 | } |
3132 | |
3133 | /* Expand VEC_SET internal functions. */ |
3134 | |
3135 | static void |
3136 | expand_vec_set_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3137 | { |
3138 | tree lhs = gimple_call_lhs (gs: stmt); |
3139 | tree op0 = gimple_call_arg (gs: stmt, index: 0); |
3140 | tree op1 = gimple_call_arg (gs: stmt, index: 1); |
3141 | tree op2 = gimple_call_arg (gs: stmt, index: 2); |
3142 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3143 | rtx src = expand_normal (exp: op0); |
3144 | |
3145 | machine_mode outermode = TYPE_MODE (TREE_TYPE (op0)); |
3146 | scalar_mode innermode = GET_MODE_INNER (outermode); |
3147 | |
3148 | rtx value = expand_normal (exp: op1); |
3149 | rtx pos = expand_normal (exp: op2); |
3150 | |
3151 | class expand_operand ops[3]; |
3152 | enum insn_code icode = optab_handler (op: optab, mode: outermode); |
3153 | |
3154 | if (icode != CODE_FOR_nothing) |
3155 | { |
3156 | rtx temp = gen_reg_rtx (outermode); |
3157 | emit_move_insn (temp, src); |
3158 | |
3159 | create_fixed_operand (op: &ops[0], x: temp); |
3160 | create_input_operand (op: &ops[1], value, mode: innermode); |
3161 | create_convert_operand_from (op: &ops[2], value: pos, TYPE_MODE (TREE_TYPE (op2)), |
3162 | unsigned_p: true); |
3163 | if (maybe_expand_insn (icode, nops: 3, ops)) |
3164 | { |
3165 | emit_move_insn (target, temp); |
3166 | return; |
3167 | } |
3168 | } |
3169 | gcc_unreachable (); |
3170 | } |
3171 | |
3172 | static void |
3173 | expand_ABNORMAL_DISPATCHER (internal_fn, gcall *) |
3174 | { |
3175 | } |
3176 | |
3177 | static void |
3178 | expand_BUILTIN_EXPECT (internal_fn, gcall *stmt) |
3179 | { |
3180 | /* When guessing was done, the hints should be already stripped away. */ |
3181 | gcc_assert (!flag_guess_branch_prob || optimize == 0 || seen_error ()); |
3182 | |
3183 | rtx target; |
3184 | tree lhs = gimple_call_lhs (gs: stmt); |
3185 | if (lhs) |
3186 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3187 | else |
3188 | target = const0_rtx; |
3189 | rtx val = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), target, VOIDmode, modifier: EXPAND_NORMAL); |
3190 | if (lhs && val != target) |
3191 | emit_move_insn (target, val); |
3192 | } |
3193 | |
3194 | /* IFN_VA_ARG is supposed to be expanded at pass_stdarg. So this dummy function |
3195 | should never be called. */ |
3196 | |
3197 | static void |
3198 | expand_VA_ARG (internal_fn, gcall *) |
3199 | { |
3200 | gcc_unreachable (); |
3201 | } |
3202 | |
3203 | /* IFN_VEC_CONVERT is supposed to be expanded at pass_lower_vector. So this |
3204 | dummy function should never be called. */ |
3205 | |
3206 | static void |
3207 | expand_VEC_CONVERT (internal_fn, gcall *) |
3208 | { |
3209 | gcc_unreachable (); |
3210 | } |
3211 | |
3212 | /* Expand IFN_RAWMEMCHR internal function. */ |
3213 | |
3214 | void |
3215 | expand_RAWMEMCHR (internal_fn, gcall *stmt) |
3216 | { |
3217 | expand_operand ops[3]; |
3218 | |
3219 | tree lhs = gimple_call_lhs (gs: stmt); |
3220 | if (!lhs) |
3221 | return; |
3222 | machine_mode lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); |
3223 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3224 | create_output_operand (op: &ops[0], x: lhs_rtx, mode: lhs_mode); |
3225 | |
3226 | tree mem = gimple_call_arg (gs: stmt, index: 0); |
3227 | rtx mem_rtx = get_memory_rtx (exp: mem, NULL); |
3228 | create_fixed_operand (op: &ops[1], x: mem_rtx); |
3229 | |
3230 | tree pattern = gimple_call_arg (gs: stmt, index: 1); |
3231 | machine_mode mode = TYPE_MODE (TREE_TYPE (pattern)); |
3232 | rtx pattern_rtx = expand_normal (exp: pattern); |
3233 | create_input_operand (op: &ops[2], value: pattern_rtx, mode); |
3234 | |
3235 | insn_code icode = direct_optab_handler (op: rawmemchr_optab, mode); |
3236 | |
3237 | expand_insn (icode, nops: 3, ops); |
3238 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3239 | emit_move_insn (lhs_rtx, ops[0].value); |
3240 | } |
3241 | |
3242 | /* Expand the IFN_UNIQUE function according to its first argument. */ |
3243 | |
3244 | static void |
3245 | expand_UNIQUE (internal_fn, gcall *stmt) |
3246 | { |
3247 | rtx pattern = NULL_RTX; |
3248 | enum ifn_unique_kind kind |
3249 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (stmt, 0)); |
3250 | |
3251 | switch (kind) |
3252 | { |
3253 | default: |
3254 | gcc_unreachable (); |
3255 | |
3256 | case IFN_UNIQUE_UNSPEC: |
3257 | if (targetm.have_unique ()) |
3258 | pattern = targetm.gen_unique (); |
3259 | break; |
3260 | |
3261 | case IFN_UNIQUE_OACC_FORK: |
3262 | case IFN_UNIQUE_OACC_JOIN: |
3263 | if (targetm.have_oacc_fork () && targetm.have_oacc_join ()) |
3264 | { |
3265 | tree lhs = gimple_call_lhs (gs: stmt); |
3266 | rtx target = const0_rtx; |
3267 | |
3268 | if (lhs) |
3269 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3270 | |
3271 | rtx data_dep = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
3272 | rtx axis = expand_normal (exp: gimple_call_arg (gs: stmt, index: 2)); |
3273 | |
3274 | if (kind == IFN_UNIQUE_OACC_FORK) |
3275 | pattern = targetm.gen_oacc_fork (target, data_dep, axis); |
3276 | else |
3277 | pattern = targetm.gen_oacc_join (target, data_dep, axis); |
3278 | } |
3279 | else |
3280 | gcc_unreachable (); |
3281 | break; |
3282 | } |
3283 | |
3284 | if (pattern) |
3285 | emit_insn (pattern); |
3286 | } |
3287 | |
3288 | /* Expand the IFN_DEFERRED_INIT function: |
3289 | LHS = DEFERRED_INIT (SIZE of the DECL, INIT_TYPE, NAME of the DECL); |
3290 | |
3291 | Initialize the LHS with zero/pattern according to its second argument |
3292 | INIT_TYPE: |
3293 | if INIT_TYPE is AUTO_INIT_ZERO, use zeroes to initialize; |
3294 | if INIT_TYPE is AUTO_INIT_PATTERN, use 0xFE byte-repeatable pattern |
3295 | to initialize; |
3296 | The LHS variable is initialized including paddings. |
3297 | The reasons to choose 0xFE for pattern initialization are: |
3298 | 1. It is a non-canonical virtual address on x86_64, and at the |
3299 | high end of the i386 kernel address space. |
3300 | 2. It is a very large float value (-1.694739530317379e+38). |
3301 | 3. It is also an unusual number for integers. */ |
3302 | #define INIT_PATTERN_VALUE 0xFE |
3303 | static void |
3304 | expand_DEFERRED_INIT (internal_fn, gcall *stmt) |
3305 | { |
3306 | tree lhs = gimple_call_lhs (gs: stmt); |
3307 | tree var_size = gimple_call_arg (gs: stmt, index: 0); |
3308 | enum auto_init_type init_type |
3309 | = (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1)); |
3310 | bool reg_lhs = true; |
3311 | |
3312 | tree var_type = TREE_TYPE (lhs); |
3313 | gcc_assert (init_type > AUTO_INIT_UNINITIALIZED); |
3314 | |
3315 | if (TREE_CODE (lhs) == SSA_NAME) |
3316 | reg_lhs = true; |
3317 | else |
3318 | { |
3319 | tree lhs_base = lhs; |
3320 | while (handled_component_p (t: lhs_base)) |
3321 | lhs_base = TREE_OPERAND (lhs_base, 0); |
3322 | reg_lhs = (mem_ref_refers_to_non_mem_p (lhs_base) |
3323 | || non_mem_decl_p (lhs_base)); |
3324 | /* If this expands to a register and the underlying decl is wrapped in |
3325 | a MEM_REF that just serves as an access type change expose the decl |
3326 | if it is of correct size. This avoids a situation as in PR103271 |
3327 | if the target does not support a direct move to the registers mode. */ |
3328 | if (reg_lhs |
3329 | && TREE_CODE (lhs_base) == MEM_REF |
3330 | && TREE_CODE (TREE_OPERAND (lhs_base, 0)) == ADDR_EXPR |
3331 | && DECL_P (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)) |
3332 | && integer_zerop (TREE_OPERAND (lhs_base, 1)) |
3333 | && tree_fits_uhwi_p (var_size) |
3334 | && tree_int_cst_equal |
3335 | (var_size, |
3336 | DECL_SIZE_UNIT (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)))) |
3337 | { |
3338 | lhs = TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0); |
3339 | var_type = TREE_TYPE (lhs); |
3340 | } |
3341 | } |
3342 | |
3343 | if (!reg_lhs) |
3344 | { |
3345 | /* If the variable is not in register, expand to a memset |
3346 | to initialize it. */ |
3347 | mark_addressable (lhs); |
3348 | tree var_addr = build_fold_addr_expr (lhs); |
3349 | |
3350 | tree value = (init_type == AUTO_INIT_PATTERN) |
3351 | ? build_int_cst (integer_type_node, |
3352 | INIT_PATTERN_VALUE) |
3353 | : integer_zero_node; |
3354 | tree m_call = build_call_expr (builtin_decl_implicit (fncode: BUILT_IN_MEMSET), |
3355 | 3, var_addr, value, var_size); |
3356 | /* Expand this memset call. */ |
3357 | expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type)); |
3358 | } |
3359 | else |
3360 | { |
3361 | /* If this variable is in a register use expand_assignment. |
3362 | For boolean scalars force zero-init. */ |
3363 | tree init; |
3364 | scalar_int_mode var_mode; |
3365 | if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE |
3366 | && tree_fits_uhwi_p (var_size) |
3367 | && (init_type == AUTO_INIT_PATTERN |
3368 | || !is_gimple_reg_type (type: var_type)) |
3369 | && int_mode_for_size (size: tree_to_uhwi (var_size) * BITS_PER_UNIT, |
3370 | limit: 0).exists (mode: &var_mode) |
3371 | && have_insn_for (SET, var_mode)) |
3372 | { |
3373 | unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size); |
3374 | unsigned char *buf = XALLOCAVEC (unsigned char, total_bytes); |
3375 | memset (s: buf, c: (init_type == AUTO_INIT_PATTERN |
3376 | ? INIT_PATTERN_VALUE : 0), n: total_bytes); |
3377 | tree itype = build_nonstandard_integer_type |
3378 | (total_bytes * BITS_PER_UNIT, 1); |
3379 | wide_int w = wi::from_buffer (buf, total_bytes); |
3380 | init = wide_int_to_tree (type: itype, cst: w); |
3381 | /* Pun the LHS to make sure its type has constant size |
3382 | unless it is an SSA name where that's already known. */ |
3383 | if (TREE_CODE (lhs) != SSA_NAME) |
3384 | lhs = build1 (VIEW_CONVERT_EXPR, itype, lhs); |
3385 | else |
3386 | init = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (lhs), init); |
3387 | } |
3388 | else |
3389 | /* Use zero-init also for variable-length sizes. */ |
3390 | init = build_zero_cst (var_type); |
3391 | |
3392 | expand_assignment (lhs, init, false); |
3393 | } |
3394 | } |
3395 | |
3396 | /* The size of an OpenACC compute dimension. */ |
3397 | |
3398 | static void |
3399 | expand_GOACC_DIM_SIZE (internal_fn, gcall *stmt) |
3400 | { |
3401 | tree lhs = gimple_call_lhs (gs: stmt); |
3402 | |
3403 | if (!lhs) |
3404 | return; |
3405 | |
3406 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3407 | if (targetm.have_oacc_dim_size ()) |
3408 | { |
3409 | rtx dim = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), NULL_RTX, |
3410 | VOIDmode, modifier: EXPAND_NORMAL); |
3411 | emit_insn (targetm.gen_oacc_dim_size (target, dim)); |
3412 | } |
3413 | else |
3414 | emit_move_insn (target, GEN_INT (1)); |
3415 | } |
3416 | |
3417 | /* The position of an OpenACC execution engine along one compute axis. */ |
3418 | |
3419 | static void |
3420 | expand_GOACC_DIM_POS (internal_fn, gcall *stmt) |
3421 | { |
3422 | tree lhs = gimple_call_lhs (gs: stmt); |
3423 | |
3424 | if (!lhs) |
3425 | return; |
3426 | |
3427 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3428 | if (targetm.have_oacc_dim_pos ()) |
3429 | { |
3430 | rtx dim = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), NULL_RTX, |
3431 | VOIDmode, modifier: EXPAND_NORMAL); |
3432 | emit_insn (targetm.gen_oacc_dim_pos (target, dim)); |
3433 | } |
3434 | else |
3435 | emit_move_insn (target, const0_rtx); |
3436 | } |
3437 | |
3438 | /* This is expanded by oacc_device_lower pass. */ |
3439 | |
3440 | static void |
3441 | expand_GOACC_LOOP (internal_fn, gcall *) |
3442 | { |
3443 | gcc_unreachable (); |
3444 | } |
3445 | |
3446 | /* This is expanded by oacc_device_lower pass. */ |
3447 | |
3448 | static void |
3449 | expand_GOACC_REDUCTION (internal_fn, gcall *) |
3450 | { |
3451 | gcc_unreachable (); |
3452 | } |
3453 | |
3454 | /* This is expanded by oacc_device_lower pass. */ |
3455 | |
3456 | static void |
3457 | expand_GOACC_TILE (internal_fn, gcall *) |
3458 | { |
3459 | gcc_unreachable (); |
3460 | } |
3461 | |
3462 | /* Set errno to EDOM. */ |
3463 | |
3464 | static void |
3465 | expand_SET_EDOM (internal_fn, gcall *) |
3466 | { |
3467 | #ifdef TARGET_EDOM |
3468 | #ifdef GEN_ERRNO_RTX |
3469 | rtx errno_rtx = GEN_ERRNO_RTX; |
3470 | #else |
3471 | rtx errno_rtx = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno" )); |
3472 | #endif |
3473 | emit_move_insn (errno_rtx, |
3474 | gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); |
3475 | #else |
3476 | gcc_unreachable (); |
3477 | #endif |
3478 | } |
3479 | |
3480 | /* Expand atomic bit test and set. */ |
3481 | |
3482 | static void |
3483 | expand_ATOMIC_BIT_TEST_AND_SET (internal_fn, gcall *call) |
3484 | { |
3485 | expand_ifn_atomic_bit_test_and (call); |
3486 | } |
3487 | |
3488 | /* Expand atomic bit test and complement. */ |
3489 | |
3490 | static void |
3491 | expand_ATOMIC_BIT_TEST_AND_COMPLEMENT (internal_fn, gcall *call) |
3492 | { |
3493 | expand_ifn_atomic_bit_test_and (call); |
3494 | } |
3495 | |
3496 | /* Expand atomic bit test and reset. */ |
3497 | |
3498 | static void |
3499 | expand_ATOMIC_BIT_TEST_AND_RESET (internal_fn, gcall *call) |
3500 | { |
3501 | expand_ifn_atomic_bit_test_and (call); |
3502 | } |
3503 | |
3504 | /* Expand atomic bit test and set. */ |
3505 | |
3506 | static void |
3507 | expand_ATOMIC_COMPARE_EXCHANGE (internal_fn, gcall *call) |
3508 | { |
3509 | expand_ifn_atomic_compare_exchange (call); |
3510 | } |
3511 | |
3512 | /* Expand atomic add fetch and cmp with 0. */ |
3513 | |
3514 | static void |
3515 | expand_ATOMIC_ADD_FETCH_CMP_0 (internal_fn, gcall *call) |
3516 | { |
3517 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3518 | } |
3519 | |
3520 | /* Expand atomic sub fetch and cmp with 0. */ |
3521 | |
3522 | static void |
3523 | expand_ATOMIC_SUB_FETCH_CMP_0 (internal_fn, gcall *call) |
3524 | { |
3525 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3526 | } |
3527 | |
3528 | /* Expand atomic and fetch and cmp with 0. */ |
3529 | |
3530 | static void |
3531 | expand_ATOMIC_AND_FETCH_CMP_0 (internal_fn, gcall *call) |
3532 | { |
3533 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3534 | } |
3535 | |
3536 | /* Expand atomic or fetch and cmp with 0. */ |
3537 | |
3538 | static void |
3539 | expand_ATOMIC_OR_FETCH_CMP_0 (internal_fn, gcall *call) |
3540 | { |
3541 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3542 | } |
3543 | |
3544 | /* Expand atomic xor fetch and cmp with 0. */ |
3545 | |
3546 | static void |
3547 | expand_ATOMIC_XOR_FETCH_CMP_0 (internal_fn, gcall *call) |
3548 | { |
3549 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3550 | } |
3551 | |
3552 | /* Expand LAUNDER to assignment, lhs = arg0. */ |
3553 | |
3554 | static void |
3555 | expand_LAUNDER (internal_fn, gcall *call) |
3556 | { |
3557 | tree lhs = gimple_call_lhs (gs: call); |
3558 | |
3559 | if (!lhs) |
3560 | return; |
3561 | |
3562 | expand_assignment (lhs, gimple_call_arg (gs: call, index: 0), false); |
3563 | } |
3564 | |
3565 | /* Expand {MASK_,}SCATTER_STORE{S,U} call CALL using optab OPTAB. */ |
3566 | |
3567 | static void |
3568 | expand_scatter_store_optab_fn (internal_fn, gcall *stmt, direct_optab optab) |
3569 | { |
3570 | internal_fn ifn = gimple_call_internal_fn (gs: stmt); |
3571 | int rhs_index = internal_fn_stored_value_index (ifn); |
3572 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3573 | tree offset = gimple_call_arg (gs: stmt, index: 1); |
3574 | tree scale = gimple_call_arg (gs: stmt, index: 2); |
3575 | tree rhs = gimple_call_arg (gs: stmt, index: rhs_index); |
3576 | |
3577 | rtx base_rtx = expand_normal (exp: base); |
3578 | rtx offset_rtx = expand_normal (exp: offset); |
3579 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); |
3580 | rtx rhs_rtx = expand_normal (exp: rhs); |
3581 | |
3582 | class expand_operand ops[8]; |
3583 | int i = 0; |
3584 | create_address_operand (op: &ops[i++], value: base_rtx); |
3585 | create_input_operand (op: &ops[i++], value: offset_rtx, TYPE_MODE (TREE_TYPE (offset))); |
3586 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); |
3587 | create_integer_operand (&ops[i++], scale_int); |
3588 | create_input_operand (op: &ops[i++], value: rhs_rtx, TYPE_MODE (TREE_TYPE (rhs))); |
3589 | i = add_mask_and_len_args (ops, opno: i, stmt); |
3590 | |
3591 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (TREE_TYPE (rhs)), |
3592 | TYPE_MODE (TREE_TYPE (offset))); |
3593 | expand_insn (icode, nops: i, ops); |
3594 | } |
3595 | |
3596 | /* Expand {MASK_,}GATHER_LOAD call CALL using optab OPTAB. */ |
3597 | |
3598 | static void |
3599 | expand_gather_load_optab_fn (internal_fn, gcall *stmt, direct_optab optab) |
3600 | { |
3601 | tree lhs = gimple_call_lhs (gs: stmt); |
3602 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3603 | tree offset = gimple_call_arg (gs: stmt, index: 1); |
3604 | tree scale = gimple_call_arg (gs: stmt, index: 2); |
3605 | |
3606 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3607 | rtx base_rtx = expand_normal (exp: base); |
3608 | rtx offset_rtx = expand_normal (exp: offset); |
3609 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); |
3610 | |
3611 | int i = 0; |
3612 | class expand_operand ops[8]; |
3613 | create_output_operand (op: &ops[i++], x: lhs_rtx, TYPE_MODE (TREE_TYPE (lhs))); |
3614 | create_address_operand (op: &ops[i++], value: base_rtx); |
3615 | create_input_operand (op: &ops[i++], value: offset_rtx, TYPE_MODE (TREE_TYPE (offset))); |
3616 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); |
3617 | create_integer_operand (&ops[i++], scale_int); |
3618 | i = add_mask_and_len_args (ops, opno: i, stmt); |
3619 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (TREE_TYPE (lhs)), |
3620 | TYPE_MODE (TREE_TYPE (offset))); |
3621 | expand_insn (icode, nops: i, ops); |
3622 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3623 | emit_move_insn (lhs_rtx, ops[0].value); |
3624 | } |
3625 | |
3626 | /* Helper for expand_DIVMOD. Return true if the sequence starting with |
3627 | INSN contains any call insns or insns with {,U}{DIV,MOD} rtxes. */ |
3628 | |
3629 | static bool |
3630 | contains_call_div_mod (rtx_insn *insn) |
3631 | { |
3632 | subrtx_iterator::array_type array; |
3633 | for (; insn; insn = NEXT_INSN (insn)) |
3634 | if (CALL_P (insn)) |
3635 | return true; |
3636 | else if (INSN_P (insn)) |
3637 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) |
3638 | switch (GET_CODE (*iter)) |
3639 | { |
3640 | case CALL: |
3641 | case DIV: |
3642 | case UDIV: |
3643 | case MOD: |
3644 | case UMOD: |
3645 | return true; |
3646 | default: |
3647 | break; |
3648 | } |
3649 | return false; |
3650 | } |
3651 | |
3652 | /* Expand DIVMOD() using: |
3653 | a) optab handler for udivmod/sdivmod if it is available. |
3654 | b) If optab_handler doesn't exist, generate call to |
3655 | target-specific divmod libfunc. */ |
3656 | |
3657 | static void |
3658 | expand_DIVMOD (internal_fn, gcall *call_stmt) |
3659 | { |
3660 | tree lhs = gimple_call_lhs (gs: call_stmt); |
3661 | tree arg0 = gimple_call_arg (gs: call_stmt, index: 0); |
3662 | tree arg1 = gimple_call_arg (gs: call_stmt, index: 1); |
3663 | |
3664 | gcc_assert (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE); |
3665 | tree type = TREE_TYPE (TREE_TYPE (lhs)); |
3666 | machine_mode mode = TYPE_MODE (type); |
3667 | bool unsignedp = TYPE_UNSIGNED (type); |
3668 | optab tab = (unsignedp) ? udivmod_optab : sdivmod_optab; |
3669 | |
3670 | rtx op0 = expand_normal (exp: arg0); |
3671 | rtx op1 = expand_normal (exp: arg1); |
3672 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3673 | |
3674 | rtx quotient = NULL_RTX, remainder = NULL_RTX; |
3675 | rtx_insn *insns = NULL; |
3676 | |
3677 | if (TREE_CODE (arg1) == INTEGER_CST) |
3678 | { |
3679 | /* For DIVMOD by integral constants, there could be efficient code |
3680 | expanded inline e.g. using shifts and plus/minus. Try to expand |
3681 | the division and modulo and if it emits any library calls or any |
3682 | {,U}{DIV,MOD} rtxes throw it away and use a divmod optab or |
3683 | divmod libcall. */ |
3684 | scalar_int_mode int_mode; |
3685 | if (remainder == NULL_RTX |
3686 | && optimize |
3687 | && CONST_INT_P (op1) |
3688 | && !pow2p_hwi (INTVAL (op1)) |
3689 | && is_int_mode (TYPE_MODE (type), int_mode: &int_mode) |
3690 | && GET_MODE_SIZE (mode: int_mode) == 2 * UNITS_PER_WORD |
3691 | && optab_handler (op: and_optab, mode: word_mode) != CODE_FOR_nothing |
3692 | && optab_handler (op: add_optab, mode: word_mode) != CODE_FOR_nothing |
3693 | && optimize_insn_for_speed_p ()) |
3694 | { |
3695 | rtx_insn *last = get_last_insn (); |
3696 | remainder = NULL_RTX; |
3697 | quotient = expand_doubleword_divmod (int_mode, op0, op1, &remainder, |
3698 | TYPE_UNSIGNED (type)); |
3699 | if (quotient != NULL_RTX) |
3700 | { |
3701 | if (optab_handler (op: mov_optab, mode: int_mode) != CODE_FOR_nothing) |
3702 | { |
3703 | rtx_insn *move = emit_move_insn (quotient, quotient); |
3704 | set_dst_reg_note (move, REG_EQUAL, |
3705 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) |
3706 | ? UDIV : DIV, int_mode, |
3707 | copy_rtx (op0), op1), |
3708 | quotient); |
3709 | move = emit_move_insn (remainder, remainder); |
3710 | set_dst_reg_note (move, REG_EQUAL, |
3711 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) |
3712 | ? UMOD : MOD, int_mode, |
3713 | copy_rtx (op0), op1), |
3714 | quotient); |
3715 | } |
3716 | } |
3717 | else |
3718 | delete_insns_since (last); |
3719 | } |
3720 | |
3721 | if (remainder == NULL_RTX) |
3722 | { |
3723 | struct separate_ops ops; |
3724 | ops.code = TRUNC_DIV_EXPR; |
3725 | ops.type = type; |
3726 | ops.op0 = make_tree (ops.type, op0); |
3727 | ops.op1 = arg1; |
3728 | ops.op2 = NULL_TREE; |
3729 | ops.location = gimple_location (g: call_stmt); |
3730 | start_sequence (); |
3731 | quotient = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
3732 | if (contains_call_div_mod (insn: get_insns ())) |
3733 | quotient = NULL_RTX; |
3734 | else |
3735 | { |
3736 | ops.code = TRUNC_MOD_EXPR; |
3737 | remainder = expand_expr_real_2 (&ops, NULL_RTX, mode, |
3738 | EXPAND_NORMAL); |
3739 | if (contains_call_div_mod (insn: get_insns ())) |
3740 | remainder = NULL_RTX; |
3741 | } |
3742 | if (remainder) |
3743 | insns = get_insns (); |
3744 | end_sequence (); |
3745 | } |
3746 | } |
3747 | |
3748 | if (remainder) |
3749 | emit_insn (insns); |
3750 | |
3751 | /* Check if optab_handler exists for divmod_optab for given mode. */ |
3752 | else if (optab_handler (op: tab, mode) != CODE_FOR_nothing) |
3753 | { |
3754 | quotient = gen_reg_rtx (mode); |
3755 | remainder = gen_reg_rtx (mode); |
3756 | expand_twoval_binop (tab, op0, op1, quotient, remainder, unsignedp); |
3757 | } |
3758 | |
3759 | /* Generate call to divmod libfunc if it exists. */ |
3760 | else if (rtx libfunc = optab_libfunc (tab, mode)) |
3761 | targetm.expand_divmod_libfunc (libfunc, mode, op0, op1, |
3762 | "ient, &remainder); |
3763 | |
3764 | else |
3765 | gcc_unreachable (); |
3766 | |
3767 | /* Wrap the return value (quotient, remainder) within COMPLEX_EXPR. */ |
3768 | expand_expr (exp: build2 (COMPLEX_EXPR, TREE_TYPE (lhs), |
3769 | make_tree (TREE_TYPE (arg0), quotient), |
3770 | make_tree (TREE_TYPE (arg1), remainder)), |
3771 | target, VOIDmode, modifier: EXPAND_NORMAL); |
3772 | } |
3773 | |
3774 | /* Expand a NOP. */ |
3775 | |
3776 | static void |
3777 | expand_NOP (internal_fn, gcall *) |
3778 | { |
3779 | /* Nothing. But it shouldn't really prevail. */ |
3780 | } |
3781 | |
3782 | /* Coroutines, all should have been processed at this stage. */ |
3783 | |
3784 | static void |
3785 | expand_CO_FRAME (internal_fn, gcall *) |
3786 | { |
3787 | gcc_unreachable (); |
3788 | } |
3789 | |
3790 | static void |
3791 | expand_CO_YIELD (internal_fn, gcall *) |
3792 | { |
3793 | gcc_unreachable (); |
3794 | } |
3795 | |
3796 | static void |
3797 | expand_CO_SUSPN (internal_fn, gcall *) |
3798 | { |
3799 | gcc_unreachable (); |
3800 | } |
3801 | |
3802 | static void |
3803 | expand_CO_ACTOR (internal_fn, gcall *) |
3804 | { |
3805 | gcc_unreachable (); |
3806 | } |
3807 | |
3808 | /* Expand a call to FN using the operands in STMT. FN has a single |
3809 | output operand and NARGS input operands. */ |
3810 | |
3811 | static void |
3812 | expand_direct_optab_fn (internal_fn fn, gcall *stmt, direct_optab optab, |
3813 | unsigned int nargs) |
3814 | { |
3815 | tree_pair types = direct_internal_fn_types (fn, stmt); |
3816 | insn_code icode = direct_optab_handler (op: optab, TYPE_MODE (types.first)); |
3817 | expand_fn_using_insn (stmt, icode, noutputs: 1, ninputs: nargs); |
3818 | } |
3819 | |
3820 | /* Expand WHILE_ULT call STMT using optab OPTAB. */ |
3821 | |
3822 | static void |
3823 | expand_while_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3824 | { |
3825 | expand_operand ops[4]; |
3826 | tree rhs_type[2]; |
3827 | |
3828 | tree lhs = gimple_call_lhs (gs: stmt); |
3829 | tree lhs_type = TREE_TYPE (lhs); |
3830 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3831 | create_output_operand (op: &ops[0], x: lhs_rtx, TYPE_MODE (lhs_type)); |
3832 | |
3833 | for (unsigned int i = 0; i < 2; ++i) |
3834 | { |
3835 | tree rhs = gimple_call_arg (gs: stmt, index: i); |
3836 | rhs_type[i] = TREE_TYPE (rhs); |
3837 | rtx rhs_rtx = expand_normal (exp: rhs); |
3838 | create_input_operand (op: &ops[i + 1], value: rhs_rtx, TYPE_MODE (rhs_type[i])); |
3839 | } |
3840 | |
3841 | int opcnt; |
3842 | if (!VECTOR_MODE_P (TYPE_MODE (lhs_type))) |
3843 | { |
3844 | /* When the mask is an integer mode the exact vector length may not |
3845 | be clear to the backend, so we pass it in operand[3]. |
3846 | Use the vector in arg2 for the most reliable intended size. */ |
3847 | tree type = TREE_TYPE (gimple_call_arg (stmt, 2)); |
3848 | create_integer_operand (&ops[3], TYPE_VECTOR_SUBPARTS (node: type)); |
3849 | opcnt = 4; |
3850 | } |
3851 | else |
3852 | /* The mask has a vector type so the length operand is unnecessary. */ |
3853 | opcnt = 3; |
3854 | |
3855 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (rhs_type[0]), |
3856 | TYPE_MODE (lhs_type)); |
3857 | |
3858 | expand_insn (icode, nops: opcnt, ops); |
3859 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3860 | emit_move_insn (lhs_rtx, ops[0].value); |
3861 | } |
3862 | |
3863 | /* Expand a call to a convert-like optab using the operands in STMT. |
3864 | FN has a single output operand and NARGS input operands. */ |
3865 | |
3866 | static void |
3867 | expand_convert_optab_fn (internal_fn fn, gcall *stmt, convert_optab optab, |
3868 | unsigned int nargs) |
3869 | { |
3870 | tree_pair types = direct_internal_fn_types (fn, stmt); |
3871 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (types.first), |
3872 | TYPE_MODE (types.second)); |
3873 | expand_fn_using_insn (stmt, icode, noutputs: 1, ninputs: nargs); |
3874 | } |
3875 | |
3876 | /* Expanders for optabs that can use expand_direct_optab_fn. */ |
3877 | |
3878 | #define expand_unary_optab_fn(FN, STMT, OPTAB) \ |
3879 | expand_direct_optab_fn (FN, STMT, OPTAB, 1) |
3880 | |
3881 | #define expand_binary_optab_fn(FN, STMT, OPTAB) \ |
3882 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) |
3883 | |
3884 | #define expand_ternary_optab_fn(FN, STMT, OPTAB) \ |
3885 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
3886 | |
3887 | #define expand_cond_unary_optab_fn(FN, STMT, OPTAB) \ |
3888 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
3889 | |
3890 | #define expand_cond_binary_optab_fn(FN, STMT, OPTAB) \ |
3891 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
3892 | |
3893 | #define expand_cond_ternary_optab_fn(FN, STMT, OPTAB) \ |
3894 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
3895 | |
3896 | #define expand_cond_len_unary_optab_fn(FN, STMT, OPTAB) \ |
3897 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
3898 | |
3899 | #define expand_cond_len_binary_optab_fn(FN, STMT, OPTAB) \ |
3900 | expand_direct_optab_fn (FN, STMT, OPTAB, 6) |
3901 | |
3902 | #define expand_cond_len_ternary_optab_fn(FN, STMT, OPTAB) \ |
3903 | expand_direct_optab_fn (FN, STMT, OPTAB, 7) |
3904 | |
3905 | #define expand_fold_extract_optab_fn(FN, STMT, OPTAB) \ |
3906 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
3907 | |
3908 | #define expand_fold_len_extract_optab_fn(FN, STMT, OPTAB) \ |
3909 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
3910 | |
3911 | #define expand_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3912 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) |
3913 | |
3914 | #define expand_mask_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3915 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
3916 | |
3917 | #define expand_mask_len_fold_left_optab_fn(FN, STMT, OPTAB) \ |
3918 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
3919 | |
3920 | #define expand_check_ptrs_optab_fn(FN, STMT, OPTAB) \ |
3921 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
3922 | |
3923 | /* Expanders for optabs that can use expand_convert_optab_fn. */ |
3924 | |
3925 | #define expand_unary_convert_optab_fn(FN, STMT, OPTAB) \ |
3926 | expand_convert_optab_fn (FN, STMT, OPTAB, 1) |
3927 | |
3928 | #define expand_vec_extract_optab_fn(FN, STMT, OPTAB) \ |
3929 | expand_convert_optab_fn (FN, STMT, OPTAB, 2) |
3930 | |
3931 | /* RETURN_TYPE and ARGS are a return type and argument list that are |
3932 | in principle compatible with FN (which satisfies direct_internal_fn_p). |
3933 | Return the types that should be used to determine whether the |
3934 | target supports FN. */ |
3935 | |
3936 | tree_pair |
3937 | direct_internal_fn_types (internal_fn fn, tree return_type, tree *args) |
3938 | { |
3939 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
3940 | tree type0 = (info.type0 < 0 ? return_type : TREE_TYPE (args[info.type0])); |
3941 | tree type1 = (info.type1 < 0 ? return_type : TREE_TYPE (args[info.type1])); |
3942 | return tree_pair (type0, type1); |
3943 | } |
3944 | |
3945 | /* CALL is a call whose return type and arguments are in principle |
3946 | compatible with FN (which satisfies direct_internal_fn_p). Return the |
3947 | types that should be used to determine whether the target supports FN. */ |
3948 | |
3949 | tree_pair |
3950 | direct_internal_fn_types (internal_fn fn, gcall *call) |
3951 | { |
3952 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
3953 | tree op0 = (info.type0 < 0 |
3954 | ? gimple_call_lhs (gs: call) |
3955 | : gimple_call_arg (gs: call, index: info.type0)); |
3956 | tree op1 = (info.type1 < 0 |
3957 | ? gimple_call_lhs (gs: call) |
3958 | : gimple_call_arg (gs: call, index: info.type1)); |
3959 | return tree_pair (TREE_TYPE (op0), TREE_TYPE (op1)); |
3960 | } |
3961 | |
3962 | /* Return true if OPTAB is supported for TYPES (whose modes should be |
3963 | the same) when the optimization type is OPT_TYPE. Used for simple |
3964 | direct optabs. */ |
3965 | |
3966 | static bool |
3967 | direct_optab_supported_p (direct_optab optab, tree_pair types, |
3968 | optimization_type opt_type) |
3969 | { |
3970 | machine_mode mode = TYPE_MODE (types.first); |
3971 | gcc_checking_assert (mode == TYPE_MODE (types.second)); |
3972 | return direct_optab_handler (optab, mode, opt_type) != CODE_FOR_nothing; |
3973 | } |
3974 | |
3975 | /* Return true if OPTAB is supported for TYPES, where the first type |
3976 | is the destination and the second type is the source. Used for |
3977 | convert optabs. */ |
3978 | |
3979 | static bool |
3980 | convert_optab_supported_p (convert_optab optab, tree_pair types, |
3981 | optimization_type opt_type) |
3982 | { |
3983 | return (convert_optab_handler (optab, TYPE_MODE (types.first), |
3984 | TYPE_MODE (types.second), opt_type) |
3985 | != CODE_FOR_nothing); |
3986 | } |
3987 | |
3988 | /* Return true if load/store lanes optab OPTAB is supported for |
3989 | array type TYPES.first when the optimization type is OPT_TYPE. */ |
3990 | |
3991 | static bool |
3992 | multi_vector_optab_supported_p (convert_optab optab, tree_pair types, |
3993 | optimization_type opt_type) |
3994 | { |
3995 | gcc_assert (TREE_CODE (types.first) == ARRAY_TYPE); |
3996 | machine_mode imode = TYPE_MODE (types.first); |
3997 | machine_mode vmode = TYPE_MODE (TREE_TYPE (types.first)); |
3998 | return (convert_optab_handler (optab, imode, vmode, opt_type) |
3999 | != CODE_FOR_nothing); |
4000 | } |
4001 | |
4002 | #define direct_unary_optab_supported_p direct_optab_supported_p |
4003 | #define direct_unary_convert_optab_supported_p convert_optab_supported_p |
4004 | #define direct_binary_optab_supported_p direct_optab_supported_p |
4005 | #define direct_ternary_optab_supported_p direct_optab_supported_p |
4006 | #define direct_cond_unary_optab_supported_p direct_optab_supported_p |
4007 | #define direct_cond_binary_optab_supported_p direct_optab_supported_p |
4008 | #define direct_cond_ternary_optab_supported_p direct_optab_supported_p |
4009 | #define direct_cond_len_unary_optab_supported_p direct_optab_supported_p |
4010 | #define direct_cond_len_binary_optab_supported_p direct_optab_supported_p |
4011 | #define direct_cond_len_ternary_optab_supported_p direct_optab_supported_p |
4012 | #define direct_mask_load_optab_supported_p convert_optab_supported_p |
4013 | #define direct_load_lanes_optab_supported_p multi_vector_optab_supported_p |
4014 | #define direct_mask_load_lanes_optab_supported_p multi_vector_optab_supported_p |
4015 | #define direct_gather_load_optab_supported_p convert_optab_supported_p |
4016 | #define direct_len_load_optab_supported_p direct_optab_supported_p |
4017 | #define direct_mask_len_load_optab_supported_p convert_optab_supported_p |
4018 | #define direct_mask_store_optab_supported_p convert_optab_supported_p |
4019 | #define direct_store_lanes_optab_supported_p multi_vector_optab_supported_p |
4020 | #define direct_mask_store_lanes_optab_supported_p multi_vector_optab_supported_p |
4021 | #define direct_vec_cond_mask_optab_supported_p convert_optab_supported_p |
4022 | #define direct_vec_cond_optab_supported_p convert_optab_supported_p |
4023 | #define direct_scatter_store_optab_supported_p convert_optab_supported_p |
4024 | #define direct_len_store_optab_supported_p direct_optab_supported_p |
4025 | #define direct_mask_len_store_optab_supported_p convert_optab_supported_p |
4026 | #define direct_while_optab_supported_p convert_optab_supported_p |
4027 | #define direct_optab_supported_p |
4028 | #define direct_optab_supported_p |
4029 | #define direct_fold_left_optab_supported_p direct_optab_supported_p |
4030 | #define direct_mask_fold_left_optab_supported_p direct_optab_supported_p |
4031 | #define direct_mask_len_fold_left_optab_supported_p direct_optab_supported_p |
4032 | #define direct_check_ptrs_optab_supported_p direct_optab_supported_p |
4033 | #define direct_vec_set_optab_supported_p direct_optab_supported_p |
4034 | #define convert_optab_supported_p |
4035 | |
4036 | /* Return the optab used by internal function FN. */ |
4037 | |
4038 | optab |
4039 | direct_internal_fn_optab (internal_fn fn, tree_pair types) |
4040 | { |
4041 | switch (fn) |
4042 | { |
4043 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4044 | case IFN_##CODE: break; |
4045 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4046 | case IFN_##CODE: return OPTAB##_optab; |
4047 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4048 | UNSIGNED_OPTAB, TYPE) \ |
4049 | case IFN_##CODE: return (TYPE_UNSIGNED (types.SELECTOR) \ |
4050 | ? UNSIGNED_OPTAB ## _optab \ |
4051 | : SIGNED_OPTAB ## _optab); |
4052 | #include "internal-fn.def" |
4053 | |
4054 | case IFN_LAST: |
4055 | break; |
4056 | } |
4057 | gcc_unreachable (); |
4058 | } |
4059 | |
4060 | /* Return the optab used by internal function FN. */ |
4061 | |
4062 | static optab |
4063 | direct_internal_fn_optab (internal_fn fn) |
4064 | { |
4065 | switch (fn) |
4066 | { |
4067 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4068 | case IFN_##CODE: break; |
4069 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4070 | case IFN_##CODE: return OPTAB##_optab; |
4071 | #include "internal-fn.def" |
4072 | |
4073 | case IFN_LAST: |
4074 | break; |
4075 | } |
4076 | gcc_unreachable (); |
4077 | } |
4078 | |
4079 | /* Return true if FN is supported for the types in TYPES when the |
4080 | optimization type is OPT_TYPE. The types are those associated with |
4081 | the "type0" and "type1" fields of FN's direct_internal_fn_info |
4082 | structure. */ |
4083 | |
4084 | bool |
4085 | direct_internal_fn_supported_p (internal_fn fn, tree_pair types, |
4086 | optimization_type opt_type) |
4087 | { |
4088 | switch (fn) |
4089 | { |
4090 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4091 | case IFN_##CODE: break; |
4092 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4093 | case IFN_##CODE: \ |
4094 | return direct_##TYPE##_optab_supported_p (OPTAB##_optab, types, \ |
4095 | opt_type); |
4096 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4097 | UNSIGNED_OPTAB, TYPE) \ |
4098 | case IFN_##CODE: \ |
4099 | { \ |
4100 | optab which_optab = (TYPE_UNSIGNED (types.SELECTOR) \ |
4101 | ? UNSIGNED_OPTAB ## _optab \ |
4102 | : SIGNED_OPTAB ## _optab); \ |
4103 | return direct_##TYPE##_optab_supported_p (which_optab, types, \ |
4104 | opt_type); \ |
4105 | } |
4106 | #include "internal-fn.def" |
4107 | |
4108 | case IFN_LAST: |
4109 | break; |
4110 | } |
4111 | gcc_unreachable (); |
4112 | } |
4113 | |
4114 | /* Return true if FN is supported for type TYPE when the optimization |
4115 | type is OPT_TYPE. The caller knows that the "type0" and "type1" |
4116 | fields of FN's direct_internal_fn_info structure are the same. */ |
4117 | |
4118 | bool |
4119 | direct_internal_fn_supported_p (internal_fn fn, tree type, |
4120 | optimization_type opt_type) |
4121 | { |
4122 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
4123 | gcc_checking_assert (info.type0 == info.type1); |
4124 | return direct_internal_fn_supported_p (fn, types: tree_pair (type, type), opt_type); |
4125 | } |
4126 | |
4127 | /* Return true if the STMT is supported when the optimization type is OPT_TYPE, |
4128 | given that STMT is a call to a direct internal function. */ |
4129 | |
4130 | bool |
4131 | direct_internal_fn_supported_p (gcall *stmt, optimization_type opt_type) |
4132 | { |
4133 | internal_fn fn = gimple_call_internal_fn (gs: stmt); |
4134 | tree_pair types = direct_internal_fn_types (fn, call: stmt); |
4135 | return direct_internal_fn_supported_p (fn, types, opt_type); |
4136 | } |
4137 | |
4138 | /* Return true if FN is a binary operation and if FN is commutative. */ |
4139 | |
4140 | bool |
4141 | commutative_binary_fn_p (internal_fn fn) |
4142 | { |
4143 | switch (fn) |
4144 | { |
4145 | case IFN_AVG_FLOOR: |
4146 | case IFN_AVG_CEIL: |
4147 | case IFN_MULH: |
4148 | case IFN_MULHS: |
4149 | case IFN_MULHRS: |
4150 | case IFN_FMIN: |
4151 | case IFN_FMAX: |
4152 | case IFN_COMPLEX_MUL: |
4153 | case IFN_UBSAN_CHECK_ADD: |
4154 | case IFN_UBSAN_CHECK_MUL: |
4155 | case IFN_ADD_OVERFLOW: |
4156 | case IFN_MUL_OVERFLOW: |
4157 | case IFN_VEC_WIDEN_PLUS: |
4158 | case IFN_VEC_WIDEN_PLUS_LO: |
4159 | case IFN_VEC_WIDEN_PLUS_HI: |
4160 | case IFN_VEC_WIDEN_PLUS_EVEN: |
4161 | case IFN_VEC_WIDEN_PLUS_ODD: |
4162 | return true; |
4163 | |
4164 | default: |
4165 | return false; |
4166 | } |
4167 | } |
4168 | |
4169 | /* Return true if FN is a ternary operation and if its first two arguments |
4170 | are commutative. */ |
4171 | |
4172 | bool |
4173 | commutative_ternary_fn_p (internal_fn fn) |
4174 | { |
4175 | switch (fn) |
4176 | { |
4177 | case IFN_FMA: |
4178 | case IFN_FMS: |
4179 | case IFN_FNMA: |
4180 | case IFN_FNMS: |
4181 | case IFN_UADDC: |
4182 | return true; |
4183 | |
4184 | default: |
4185 | return false; |
4186 | } |
4187 | } |
4188 | |
4189 | /* Return true if FN is an associative binary operation. */ |
4190 | |
4191 | bool |
4192 | associative_binary_fn_p (internal_fn fn) |
4193 | { |
4194 | switch (fn) |
4195 | { |
4196 | case IFN_FMIN: |
4197 | case IFN_FMAX: |
4198 | return true; |
4199 | |
4200 | default: |
4201 | return false; |
4202 | } |
4203 | } |
4204 | |
4205 | /* If FN is commutative in two consecutive arguments, return the |
4206 | index of the first, otherwise return -1. */ |
4207 | |
4208 | int |
4209 | first_commutative_argument (internal_fn fn) |
4210 | { |
4211 | switch (fn) |
4212 | { |
4213 | case IFN_COND_ADD: |
4214 | case IFN_COND_MUL: |
4215 | case IFN_COND_MIN: |
4216 | case IFN_COND_MAX: |
4217 | case IFN_COND_FMIN: |
4218 | case IFN_COND_FMAX: |
4219 | case IFN_COND_AND: |
4220 | case IFN_COND_IOR: |
4221 | case IFN_COND_XOR: |
4222 | case IFN_COND_FMA: |
4223 | case IFN_COND_FMS: |
4224 | case IFN_COND_FNMA: |
4225 | case IFN_COND_FNMS: |
4226 | case IFN_COND_LEN_ADD: |
4227 | case IFN_COND_LEN_MUL: |
4228 | case IFN_COND_LEN_MIN: |
4229 | case IFN_COND_LEN_MAX: |
4230 | case IFN_COND_LEN_FMIN: |
4231 | case IFN_COND_LEN_FMAX: |
4232 | case IFN_COND_LEN_AND: |
4233 | case IFN_COND_LEN_IOR: |
4234 | case IFN_COND_LEN_XOR: |
4235 | case IFN_COND_LEN_FMA: |
4236 | case IFN_COND_LEN_FMS: |
4237 | case IFN_COND_LEN_FNMA: |
4238 | case IFN_COND_LEN_FNMS: |
4239 | return 1; |
4240 | |
4241 | default: |
4242 | if (commutative_binary_fn_p (fn) |
4243 | || commutative_ternary_fn_p (fn)) |
4244 | return 0; |
4245 | return -1; |
4246 | } |
4247 | } |
4248 | |
4249 | /* Return true if this CODE describes an internal_fn that returns a vector with |
4250 | elements twice as wide as the element size of the input vectors. */ |
4251 | |
4252 | bool |
4253 | widening_fn_p (code_helper code) |
4254 | { |
4255 | if (!code.is_fn_code ()) |
4256 | return false; |
4257 | |
4258 | if (!internal_fn_p (code: (combined_fn) code)) |
4259 | return false; |
4260 | |
4261 | internal_fn fn = as_internal_fn (code: (combined_fn) code); |
4262 | switch (fn) |
4263 | { |
4264 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN |
4265 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ |
4266 | case IFN_##NAME: \ |
4267 | case IFN_##NAME##_HI: \ |
4268 | case IFN_##NAME##_LO: \ |
4269 | case IFN_##NAME##_EVEN: \ |
4270 | case IFN_##NAME##_ODD: \ |
4271 | return true; |
4272 | #include "internal-fn.def" |
4273 | #undef DEF_INTERNAL_WIDENING_OPTAB_FN |
4274 | |
4275 | default: |
4276 | return false; |
4277 | } |
4278 | } |
4279 | |
4280 | /* Return true if IFN_SET_EDOM is supported. */ |
4281 | |
4282 | bool |
4283 | set_edom_supported_p (void) |
4284 | { |
4285 | #ifdef TARGET_EDOM |
4286 | return true; |
4287 | #else |
4288 | return false; |
4289 | #endif |
4290 | } |
4291 | |
4292 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4293 | static void \ |
4294 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
4295 | { \ |
4296 | expand_##TYPE##_optab_fn (fn, stmt, OPTAB##_optab); \ |
4297 | } |
4298 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4299 | UNSIGNED_OPTAB, TYPE) \ |
4300 | static void \ |
4301 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
4302 | { \ |
4303 | tree_pair types = direct_internal_fn_types (fn, stmt); \ |
4304 | optab which_optab = direct_internal_fn_optab (fn, types); \ |
4305 | expand_##TYPE##_optab_fn (fn, stmt, which_optab); \ |
4306 | } |
4307 | #include "internal-fn.def" |
4308 | #undef DEF_INTERNAL_OPTAB_FN |
4309 | #undef DEF_INTERNAL_SIGNED_OPTAB_FN |
4310 | |
4311 | /* Routines to expand each internal function, indexed by function number. |
4312 | Each routine has the prototype: |
4313 | |
4314 | expand_<NAME> (gcall *stmt) |
4315 | |
4316 | where STMT is the statement that performs the call. */ |
4317 | static void (*const internal_fn_expanders[]) (internal_fn, gcall *) = { |
4318 | |
4319 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) expand_##CODE, |
4320 | #include "internal-fn.def" |
4321 | 0 |
4322 | }; |
4323 | |
4324 | /* Invoke T(CODE, SUFFIX) for each conditional function IFN_COND_##SUFFIX |
4325 | that maps to a tree code CODE. There is also an IFN_COND_LEN_##SUFFIX |
4326 | for each such IFN_COND_##SUFFIX. */ |
4327 | #define FOR_EACH_CODE_MAPPING(T) \ |
4328 | T (PLUS_EXPR, ADD) \ |
4329 | T (MINUS_EXPR, SUB) \ |
4330 | T (MULT_EXPR, MUL) \ |
4331 | T (TRUNC_DIV_EXPR, DIV) \ |
4332 | T (TRUNC_MOD_EXPR, MOD) \ |
4333 | T (RDIV_EXPR, RDIV) \ |
4334 | T (MIN_EXPR, MIN) \ |
4335 | T (MAX_EXPR, MAX) \ |
4336 | T (BIT_AND_EXPR, AND) \ |
4337 | T (BIT_IOR_EXPR, IOR) \ |
4338 | T (BIT_XOR_EXPR, XOR) \ |
4339 | T (LSHIFT_EXPR, SHL) \ |
4340 | T (RSHIFT_EXPR, SHR) \ |
4341 | T (NEGATE_EXPR, NEG) |
4342 | |
4343 | /* Return a function that only performs CODE when a certain condition is met |
4344 | and that uses a given fallback value otherwise. For example, if CODE is |
4345 | a binary operation associated with conditional function FN: |
4346 | |
4347 | LHS = FN (COND, A, B, ELSE) |
4348 | |
4349 | is equivalent to the C expression: |
4350 | |
4351 | LHS = COND ? A CODE B : ELSE; |
4352 | |
4353 | operating elementwise if the operands are vectors. |
4354 | |
4355 | Return IFN_LAST if no such function exists. */ |
4356 | |
4357 | internal_fn |
4358 | get_conditional_internal_fn (tree_code code) |
4359 | { |
4360 | switch (code) |
4361 | { |
4362 | #define CASE(CODE, IFN) case CODE: return IFN_COND_##IFN; |
4363 | FOR_EACH_CODE_MAPPING(CASE) |
4364 | #undef CASE |
4365 | default: |
4366 | return IFN_LAST; |
4367 | } |
4368 | } |
4369 | |
4370 | /* If IFN implements the conditional form of a tree code, return that |
4371 | tree code, otherwise return ERROR_MARK. */ |
4372 | |
4373 | tree_code |
4374 | conditional_internal_fn_code (internal_fn ifn) |
4375 | { |
4376 | switch (ifn) |
4377 | { |
4378 | #define CASE(CODE, IFN) \ |
4379 | case IFN_COND_##IFN: \ |
4380 | case IFN_COND_LEN_##IFN: \ |
4381 | return CODE; |
4382 | FOR_EACH_CODE_MAPPING (CASE) |
4383 | #undef CASE |
4384 | default: |
4385 | return ERROR_MARK; |
4386 | } |
4387 | } |
4388 | |
4389 | /* Like get_conditional_internal_fn, but return a function that |
4390 | additionally restricts the operation to the leading elements |
4391 | of a vector. The number of elements to process is given by a length |
4392 | and bias pair, as for IFN_LOAD_LEN. The values of the remaining |
4393 | elements are taken from the fallback ("else") argument. |
4394 | |
4395 | For example, if CODE is a binary operation associated with FN: |
4396 | |
4397 | LHS = FN (COND, A, B, ELSE, LEN, BIAS) |
4398 | |
4399 | is equivalent to the C code: |
4400 | |
4401 | for (int i = 0; i < NUNITS; i++) |
4402 | { |
4403 | if (i < LEN + BIAS && COND[i]) |
4404 | LHS[i] = A[i] CODE B[i]; |
4405 | else |
4406 | LHS[i] = ELSE[i]; |
4407 | } |
4408 | */ |
4409 | |
4410 | internal_fn |
4411 | get_conditional_len_internal_fn (tree_code code) |
4412 | { |
4413 | switch (code) |
4414 | { |
4415 | #define CASE(CODE, IFN) case CODE: return IFN_COND_LEN_##IFN; |
4416 | FOR_EACH_CODE_MAPPING(CASE) |
4417 | #undef CASE |
4418 | default: |
4419 | return IFN_LAST; |
4420 | } |
4421 | } |
4422 | |
4423 | /* Invoke T(IFN) for each internal function IFN that also has an |
4424 | IFN_COND_* form. */ |
4425 | #define FOR_EACH_COND_FN_PAIR(T) \ |
4426 | T (FMAX) \ |
4427 | T (FMIN) \ |
4428 | T (FMA) \ |
4429 | T (FMS) \ |
4430 | T (FNMA) \ |
4431 | T (FNMS) |
4432 | |
4433 | /* Return a function that only performs internal function FN when a |
4434 | certain condition is met and that uses a given fallback value otherwise. |
4435 | In other words, the returned function FN' is such that: |
4436 | |
4437 | LHS = FN' (COND, A1, ... An, ELSE) |
4438 | |
4439 | is equivalent to the C expression: |
4440 | |
4441 | LHS = COND ? FN (A1, ..., An) : ELSE; |
4442 | |
4443 | operating elementwise if the operands are vectors. |
4444 | |
4445 | Return IFN_LAST if no such function exists. */ |
4446 | |
4447 | internal_fn |
4448 | get_conditional_internal_fn (internal_fn fn) |
4449 | { |
4450 | switch (fn) |
4451 | { |
4452 | #define CASE(NAME) case IFN_##NAME: return IFN_COND_##NAME; |
4453 | FOR_EACH_COND_FN_PAIR(CASE) |
4454 | #undef CASE |
4455 | default: |
4456 | return IFN_LAST; |
4457 | } |
4458 | } |
4459 | |
4460 | /* If there exists an internal function like IFN that operates on vectors, |
4461 | but with additional length and bias parameters, return the internal_fn |
4462 | for that function, otherwise return IFN_LAST. */ |
4463 | internal_fn |
4464 | get_len_internal_fn (internal_fn fn) |
4465 | { |
4466 | switch (fn) |
4467 | { |
4468 | #undef DEF_INTERNAL_COND_FN |
4469 | #undef DEF_INTERNAL_SIGNED_COND_FN |
4470 | #define DEF_INTERNAL_COND_FN(NAME, ...) \ |
4471 | case IFN_COND_##NAME: \ |
4472 | return IFN_COND_LEN_##NAME; |
4473 | #define DEF_INTERNAL_SIGNED_COND_FN(NAME, ...) \ |
4474 | case IFN_COND_##NAME: \ |
4475 | return IFN_COND_LEN_##NAME; |
4476 | #include "internal-fn.def" |
4477 | #undef DEF_INTERNAL_COND_FN |
4478 | #undef DEF_INTERNAL_SIGNED_COND_FN |
4479 | default: |
4480 | return IFN_LAST; |
4481 | } |
4482 | } |
4483 | |
4484 | /* If IFN implements the conditional form of an unconditional internal |
4485 | function, return that unconditional function, otherwise return IFN_LAST. */ |
4486 | |
4487 | internal_fn |
4488 | get_unconditional_internal_fn (internal_fn ifn) |
4489 | { |
4490 | switch (ifn) |
4491 | { |
4492 | #define CASE(NAME) \ |
4493 | case IFN_COND_##NAME: \ |
4494 | case IFN_COND_LEN_##NAME: \ |
4495 | return IFN_##NAME; |
4496 | FOR_EACH_COND_FN_PAIR (CASE) |
4497 | #undef CASE |
4498 | default: |
4499 | return IFN_LAST; |
4500 | } |
4501 | } |
4502 | |
4503 | /* Return true if STMT can be interpreted as a conditional tree code |
4504 | operation of the form: |
4505 | |
4506 | LHS = COND ? OP (RHS1, ...) : ELSE; |
4507 | |
4508 | operating elementwise if the operands are vectors. This includes |
4509 | the case of an all-true COND, so that the operation always happens. |
4510 | |
4511 | There is an alternative approach to interpret the STMT when the operands |
4512 | are vectors which is the operation predicated by both conditional mask |
4513 | and loop control length, the equivalent C code: |
4514 | |
4515 | for (int i = 0; i < NUNTIS; i++) |
4516 | { |
4517 | if (i < LEN + BIAS && COND[i]) |
4518 | LHS[i] = A[i] CODE B[i]; |
4519 | else |
4520 | LHS[i] = ELSE[i]; |
4521 | } |
4522 | |
4523 | When returning true, set: |
4524 | |
4525 | - *COND_OUT to the condition COND, or to NULL_TREE if the condition |
4526 | is known to be all-true |
4527 | - *CODE_OUT to the tree code |
4528 | - OPS[I] to operand I of *CODE_OUT |
4529 | - *ELSE_OUT to the fallback value ELSE, or to NULL_TREE if the |
4530 | condition is known to be all true. |
4531 | - *LEN to the len argument if it COND_LEN_* operations or to NULL_TREE. |
4532 | - *BIAS to the bias argument if it COND_LEN_* operations or to NULL_TREE. */ |
4533 | |
4534 | bool |
4535 | can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out, |
4536 | tree_code *code_out, |
4537 | tree (&ops)[3], tree *else_out, |
4538 | tree *len, tree *bias) |
4539 | { |
4540 | *len = NULL_TREE; |
4541 | *bias = NULL_TREE; |
4542 | if (gassign *assign = dyn_cast <gassign *> (p: stmt)) |
4543 | { |
4544 | *cond_out = NULL_TREE; |
4545 | *code_out = gimple_assign_rhs_code (gs: assign); |
4546 | ops[0] = gimple_assign_rhs1 (gs: assign); |
4547 | ops[1] = gimple_assign_rhs2 (gs: assign); |
4548 | ops[2] = gimple_assign_rhs3 (gs: assign); |
4549 | *else_out = NULL_TREE; |
4550 | return true; |
4551 | } |
4552 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
4553 | if (gimple_call_internal_p (gs: call)) |
4554 | { |
4555 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
4556 | tree_code code = conditional_internal_fn_code (ifn); |
4557 | int len_index = internal_fn_len_index (ifn); |
4558 | int cond_nargs = len_index >= 0 ? 4 : 2; |
4559 | if (code != ERROR_MARK) |
4560 | { |
4561 | *cond_out = gimple_call_arg (gs: call, index: 0); |
4562 | *code_out = code; |
4563 | unsigned int nops = gimple_call_num_args (gs: call) - cond_nargs; |
4564 | for (unsigned int i = 0; i < 3; ++i) |
4565 | ops[i] = i < nops ? gimple_call_arg (gs: call, index: i + 1) : NULL_TREE; |
4566 | *else_out = gimple_call_arg (gs: call, index: nops + 1); |
4567 | if (len_index < 0) |
4568 | { |
4569 | if (integer_truep (*cond_out)) |
4570 | { |
4571 | *cond_out = NULL_TREE; |
4572 | *else_out = NULL_TREE; |
4573 | } |
4574 | } |
4575 | else |
4576 | { |
4577 | *len = gimple_call_arg (gs: call, index: len_index); |
4578 | *bias = gimple_call_arg (gs: call, index: len_index + 1); |
4579 | } |
4580 | return true; |
4581 | } |
4582 | } |
4583 | return false; |
4584 | } |
4585 | |
4586 | /* Return true if IFN is some form of load from memory. */ |
4587 | |
4588 | bool |
4589 | internal_load_fn_p (internal_fn fn) |
4590 | { |
4591 | switch (fn) |
4592 | { |
4593 | case IFN_MASK_LOAD: |
4594 | case IFN_LOAD_LANES: |
4595 | case IFN_MASK_LOAD_LANES: |
4596 | case IFN_MASK_LEN_LOAD_LANES: |
4597 | case IFN_GATHER_LOAD: |
4598 | case IFN_MASK_GATHER_LOAD: |
4599 | case IFN_MASK_LEN_GATHER_LOAD: |
4600 | case IFN_LEN_LOAD: |
4601 | case IFN_MASK_LEN_LOAD: |
4602 | return true; |
4603 | |
4604 | default: |
4605 | return false; |
4606 | } |
4607 | } |
4608 | |
4609 | /* Return true if IFN is some form of store to memory. */ |
4610 | |
4611 | bool |
4612 | internal_store_fn_p (internal_fn fn) |
4613 | { |
4614 | switch (fn) |
4615 | { |
4616 | case IFN_MASK_STORE: |
4617 | case IFN_STORE_LANES: |
4618 | case IFN_MASK_STORE_LANES: |
4619 | case IFN_MASK_LEN_STORE_LANES: |
4620 | case IFN_SCATTER_STORE: |
4621 | case IFN_MASK_SCATTER_STORE: |
4622 | case IFN_MASK_LEN_SCATTER_STORE: |
4623 | case IFN_LEN_STORE: |
4624 | case IFN_MASK_LEN_STORE: |
4625 | return true; |
4626 | |
4627 | default: |
4628 | return false; |
4629 | } |
4630 | } |
4631 | |
4632 | /* Return true if IFN is some form of gather load or scatter store. */ |
4633 | |
4634 | bool |
4635 | internal_gather_scatter_fn_p (internal_fn fn) |
4636 | { |
4637 | switch (fn) |
4638 | { |
4639 | case IFN_GATHER_LOAD: |
4640 | case IFN_MASK_GATHER_LOAD: |
4641 | case IFN_MASK_LEN_GATHER_LOAD: |
4642 | case IFN_SCATTER_STORE: |
4643 | case IFN_MASK_SCATTER_STORE: |
4644 | case IFN_MASK_LEN_SCATTER_STORE: |
4645 | return true; |
4646 | |
4647 | default: |
4648 | return false; |
4649 | } |
4650 | } |
4651 | |
4652 | /* If FN takes a vector len argument, return the index of that argument, |
4653 | otherwise return -1. */ |
4654 | |
4655 | int |
4656 | internal_fn_len_index (internal_fn fn) |
4657 | { |
4658 | switch (fn) |
4659 | { |
4660 | case IFN_LEN_LOAD: |
4661 | case IFN_LEN_STORE: |
4662 | return 2; |
4663 | |
4664 | case IFN_MASK_LEN_GATHER_LOAD: |
4665 | case IFN_MASK_LEN_SCATTER_STORE: |
4666 | case IFN_COND_LEN_FMA: |
4667 | case IFN_COND_LEN_FMS: |
4668 | case IFN_COND_LEN_FNMA: |
4669 | case IFN_COND_LEN_FNMS: |
4670 | return 5; |
4671 | |
4672 | case IFN_COND_LEN_ADD: |
4673 | case IFN_COND_LEN_SUB: |
4674 | case IFN_COND_LEN_MUL: |
4675 | case IFN_COND_LEN_DIV: |
4676 | case IFN_COND_LEN_MOD: |
4677 | case IFN_COND_LEN_RDIV: |
4678 | case IFN_COND_LEN_MIN: |
4679 | case IFN_COND_LEN_MAX: |
4680 | case IFN_COND_LEN_FMIN: |
4681 | case IFN_COND_LEN_FMAX: |
4682 | case IFN_COND_LEN_AND: |
4683 | case IFN_COND_LEN_IOR: |
4684 | case IFN_COND_LEN_XOR: |
4685 | case IFN_COND_LEN_SHL: |
4686 | case IFN_COND_LEN_SHR: |
4687 | return 4; |
4688 | |
4689 | case IFN_COND_LEN_NEG: |
4690 | case IFN_MASK_LEN_LOAD: |
4691 | case IFN_MASK_LEN_STORE: |
4692 | case IFN_MASK_LEN_LOAD_LANES: |
4693 | case IFN_MASK_LEN_STORE_LANES: |
4694 | case IFN_VCOND_MASK_LEN: |
4695 | return 3; |
4696 | |
4697 | default: |
4698 | return -1; |
4699 | } |
4700 | } |
4701 | |
4702 | /* If FN is an IFN_COND_* or IFN_COND_LEN_* function, return the index of the |
4703 | argument that is used when the condition is false. Return -1 otherwise. */ |
4704 | |
4705 | int |
4706 | internal_fn_else_index (internal_fn fn) |
4707 | { |
4708 | switch (fn) |
4709 | { |
4710 | case IFN_COND_NEG: |
4711 | case IFN_COND_NOT: |
4712 | case IFN_COND_LEN_NEG: |
4713 | case IFN_COND_LEN_NOT: |
4714 | return 2; |
4715 | |
4716 | case IFN_COND_ADD: |
4717 | case IFN_COND_SUB: |
4718 | case IFN_COND_MUL: |
4719 | case IFN_COND_DIV: |
4720 | case IFN_COND_MOD: |
4721 | case IFN_COND_MIN: |
4722 | case IFN_COND_MAX: |
4723 | case IFN_COND_FMIN: |
4724 | case IFN_COND_FMAX: |
4725 | case IFN_COND_AND: |
4726 | case IFN_COND_IOR: |
4727 | case IFN_COND_XOR: |
4728 | case IFN_COND_SHL: |
4729 | case IFN_COND_SHR: |
4730 | case IFN_COND_LEN_ADD: |
4731 | case IFN_COND_LEN_SUB: |
4732 | case IFN_COND_LEN_MUL: |
4733 | case IFN_COND_LEN_DIV: |
4734 | case IFN_COND_LEN_MOD: |
4735 | case IFN_COND_LEN_MIN: |
4736 | case IFN_COND_LEN_MAX: |
4737 | case IFN_COND_LEN_FMIN: |
4738 | case IFN_COND_LEN_FMAX: |
4739 | case IFN_COND_LEN_AND: |
4740 | case IFN_COND_LEN_IOR: |
4741 | case IFN_COND_LEN_XOR: |
4742 | case IFN_COND_LEN_SHL: |
4743 | case IFN_COND_LEN_SHR: |
4744 | return 3; |
4745 | |
4746 | case IFN_COND_FMA: |
4747 | case IFN_COND_FMS: |
4748 | case IFN_COND_FNMA: |
4749 | case IFN_COND_FNMS: |
4750 | case IFN_COND_LEN_FMA: |
4751 | case IFN_COND_LEN_FMS: |
4752 | case IFN_COND_LEN_FNMA: |
4753 | case IFN_COND_LEN_FNMS: |
4754 | return 4; |
4755 | |
4756 | default: |
4757 | return -1; |
4758 | } |
4759 | |
4760 | return -1; |
4761 | } |
4762 | |
4763 | /* If FN takes a vector mask argument, return the index of that argument, |
4764 | otherwise return -1. */ |
4765 | |
4766 | int |
4767 | internal_fn_mask_index (internal_fn fn) |
4768 | { |
4769 | switch (fn) |
4770 | { |
4771 | case IFN_MASK_LOAD: |
4772 | case IFN_MASK_LOAD_LANES: |
4773 | case IFN_MASK_LEN_LOAD_LANES: |
4774 | case IFN_MASK_STORE: |
4775 | case IFN_MASK_STORE_LANES: |
4776 | case IFN_MASK_LEN_STORE_LANES: |
4777 | case IFN_MASK_LEN_LOAD: |
4778 | case IFN_MASK_LEN_STORE: |
4779 | return 2; |
4780 | |
4781 | case IFN_MASK_GATHER_LOAD: |
4782 | case IFN_MASK_SCATTER_STORE: |
4783 | case IFN_MASK_LEN_GATHER_LOAD: |
4784 | case IFN_MASK_LEN_SCATTER_STORE: |
4785 | return 4; |
4786 | |
4787 | case IFN_VCOND_MASK_LEN: |
4788 | return 0; |
4789 | |
4790 | default: |
4791 | return (conditional_internal_fn_code (ifn: fn) != ERROR_MARK |
4792 | || get_unconditional_internal_fn (ifn: fn) != IFN_LAST ? 0 : -1); |
4793 | } |
4794 | } |
4795 | |
4796 | /* If FN takes a value that should be stored to memory, return the index |
4797 | of that argument, otherwise return -1. */ |
4798 | |
4799 | int |
4800 | internal_fn_stored_value_index (internal_fn fn) |
4801 | { |
4802 | switch (fn) |
4803 | { |
4804 | case IFN_MASK_STORE: |
4805 | case IFN_MASK_STORE_LANES: |
4806 | case IFN_SCATTER_STORE: |
4807 | case IFN_MASK_SCATTER_STORE: |
4808 | case IFN_MASK_LEN_SCATTER_STORE: |
4809 | return 3; |
4810 | |
4811 | case IFN_LEN_STORE: |
4812 | return 4; |
4813 | |
4814 | case IFN_MASK_LEN_STORE: |
4815 | case IFN_MASK_LEN_STORE_LANES: |
4816 | return 5; |
4817 | |
4818 | default: |
4819 | return -1; |
4820 | } |
4821 | } |
4822 | |
4823 | /* Return true if the target supports gather load or scatter store function |
4824 | IFN. For loads, VECTOR_TYPE is the vector type of the load result, |
4825 | while for stores it is the vector type of the stored data argument. |
4826 | MEMORY_ELEMENT_TYPE is the type of the memory elements being loaded |
4827 | or stored. OFFSET_VECTOR_TYPE is the vector type that holds the |
4828 | offset from the shared base address of each loaded or stored element. |
4829 | SCALE is the amount by which these offsets should be multiplied |
4830 | *after* they have been extended to address width. */ |
4831 | |
4832 | bool |
4833 | internal_gather_scatter_fn_supported_p (internal_fn ifn, tree vector_type, |
4834 | tree memory_element_type, |
4835 | tree offset_vector_type, int scale) |
4836 | { |
4837 | if (!tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vector_type)), |
4838 | TYPE_SIZE (memory_element_type))) |
4839 | return false; |
4840 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vector_type), |
4841 | b: TYPE_VECTOR_SUBPARTS (node: offset_vector_type))) |
4842 | return false; |
4843 | optab optab = direct_internal_fn_optab (fn: ifn); |
4844 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (vector_type), |
4845 | TYPE_MODE (offset_vector_type)); |
4846 | int output_ops = internal_load_fn_p (fn: ifn) ? 1 : 0; |
4847 | bool unsigned_p = TYPE_UNSIGNED (TREE_TYPE (offset_vector_type)); |
4848 | return (icode != CODE_FOR_nothing |
4849 | && insn_operand_matches (icode, opno: 2 + output_ops, GEN_INT (unsigned_p)) |
4850 | && insn_operand_matches (icode, opno: 3 + output_ops, GEN_INT (scale))); |
4851 | } |
4852 | |
4853 | /* Return true if the target supports IFN_CHECK_{RAW,WAR}_PTRS function IFN |
4854 | for pointers of type TYPE when the accesses have LENGTH bytes and their |
4855 | common byte alignment is ALIGN. */ |
4856 | |
4857 | bool |
4858 | internal_check_ptrs_fn_supported_p (internal_fn ifn, tree type, |
4859 | poly_uint64 length, unsigned int align) |
4860 | { |
4861 | machine_mode mode = TYPE_MODE (type); |
4862 | optab optab = direct_internal_fn_optab (fn: ifn); |
4863 | insn_code icode = direct_optab_handler (op: optab, mode); |
4864 | if (icode == CODE_FOR_nothing) |
4865 | return false; |
4866 | rtx length_rtx = immed_wide_int_const (length, mode); |
4867 | return (insn_operand_matches (icode, opno: 3, operand: length_rtx) |
4868 | && insn_operand_matches (icode, opno: 4, GEN_INT (align))); |
4869 | } |
4870 | |
4871 | /* Return the supported bias for IFN which is either IFN_{LEN_,MASK_LEN_,}LOAD |
4872 | or IFN_{LEN_,MASK_LEN_,}STORE. For now we only support the biases of 0 and |
4873 | -1 (in case 0 is not an allowable length for {len_,mask_len_}load or |
4874 | {len_,mask_len_}store). If none of the biases match what the backend |
4875 | provides, return VECT_PARTIAL_BIAS_UNSUPPORTED. */ |
4876 | |
4877 | signed char |
4878 | internal_len_load_store_bias (internal_fn ifn, machine_mode mode) |
4879 | { |
4880 | optab optab = direct_internal_fn_optab (fn: ifn); |
4881 | insn_code icode = direct_optab_handler (op: optab, mode); |
4882 | int bias_no = 3; |
4883 | |
4884 | if (icode == CODE_FOR_nothing) |
4885 | { |
4886 | machine_mode mask_mode; |
4887 | if (!targetm.vectorize.get_mask_mode (mode).exists (mode: &mask_mode)) |
4888 | return VECT_PARTIAL_BIAS_UNSUPPORTED; |
4889 | if (ifn == IFN_LEN_LOAD) |
4890 | { |
4891 | /* Try MASK_LEN_LOAD. */ |
4892 | optab = direct_internal_fn_optab (fn: IFN_MASK_LEN_LOAD); |
4893 | } |
4894 | else |
4895 | { |
4896 | /* Try MASK_LEN_STORE. */ |
4897 | optab = direct_internal_fn_optab (fn: IFN_MASK_LEN_STORE); |
4898 | } |
4899 | icode = convert_optab_handler (op: optab, to_mode: mode, from_mode: mask_mode); |
4900 | bias_no = 4; |
4901 | } |
4902 | |
4903 | if (icode != CODE_FOR_nothing) |
4904 | { |
4905 | /* For now we only support biases of 0 or -1. Try both of them. */ |
4906 | if (insn_operand_matches (icode, opno: bias_no, GEN_INT (0))) |
4907 | return 0; |
4908 | if (insn_operand_matches (icode, opno: bias_no, GEN_INT (-1))) |
4909 | return -1; |
4910 | } |
4911 | |
4912 | return VECT_PARTIAL_BIAS_UNSUPPORTED; |
4913 | } |
4914 | |
4915 | /* Expand STMT as though it were a call to internal function FN. */ |
4916 | |
4917 | void |
4918 | expand_internal_call (internal_fn fn, gcall *stmt) |
4919 | { |
4920 | internal_fn_expanders[fn] (fn, stmt); |
4921 | } |
4922 | |
4923 | /* Expand STMT, which is a call to internal function FN. */ |
4924 | |
4925 | void |
4926 | expand_internal_call (gcall *stmt) |
4927 | { |
4928 | expand_internal_call (fn: gimple_call_internal_fn (gs: stmt), stmt); |
4929 | } |
4930 | |
4931 | /* If TYPE is a vector type, return true if IFN is a direct internal |
4932 | function that is supported for that type. If TYPE is a scalar type, |
4933 | return true if IFN is a direct internal function that is supported for |
4934 | the target's preferred vector version of TYPE. */ |
4935 | |
4936 | bool |
4937 | vectorized_internal_fn_supported_p (internal_fn ifn, tree type) |
4938 | { |
4939 | if (VECTOR_MODE_P (TYPE_MODE (type))) |
4940 | return direct_internal_fn_supported_p (fn: ifn, type, opt_type: OPTIMIZE_FOR_SPEED); |
4941 | |
4942 | scalar_mode smode; |
4943 | if (VECTOR_TYPE_P (type) |
4944 | || !is_a <scalar_mode> (TYPE_MODE (type), result: &smode)) |
4945 | return false; |
4946 | |
4947 | machine_mode vmode = targetm.vectorize.preferred_simd_mode (smode); |
4948 | if (VECTOR_MODE_P (vmode)) |
4949 | { |
4950 | tree vectype = build_vector_type_for_mode (type, vmode); |
4951 | if (direct_internal_fn_supported_p (fn: ifn, type: vectype, opt_type: OPTIMIZE_FOR_SPEED)) |
4952 | return true; |
4953 | } |
4954 | |
4955 | auto_vector_modes vector_modes; |
4956 | targetm.vectorize.autovectorize_vector_modes (&vector_modes, true); |
4957 | for (machine_mode base_mode : vector_modes) |
4958 | if (related_vector_mode (base_mode, smode).exists (mode: &vmode)) |
4959 | { |
4960 | tree vectype = build_vector_type_for_mode (type, vmode); |
4961 | if (direct_internal_fn_supported_p (fn: ifn, type: vectype, opt_type: OPTIMIZE_FOR_SPEED)) |
4962 | return true; |
4963 | } |
4964 | |
4965 | return false; |
4966 | } |
4967 | |
4968 | void |
4969 | expand_SHUFFLEVECTOR (internal_fn, gcall *) |
4970 | { |
4971 | gcc_unreachable (); |
4972 | } |
4973 | |
4974 | void |
4975 | expand_PHI (internal_fn, gcall *) |
4976 | { |
4977 | gcc_unreachable (); |
4978 | } |
4979 | |
4980 | void |
4981 | expand_SPACESHIP (internal_fn, gcall *stmt) |
4982 | { |
4983 | tree lhs = gimple_call_lhs (gs: stmt); |
4984 | tree rhs1 = gimple_call_arg (gs: stmt, index: 0); |
4985 | tree rhs2 = gimple_call_arg (gs: stmt, index: 1); |
4986 | tree type = TREE_TYPE (rhs1); |
4987 | |
4988 | do_pending_stack_adjust (); |
4989 | |
4990 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
4991 | rtx op1 = expand_normal (exp: rhs1); |
4992 | rtx op2 = expand_normal (exp: rhs2); |
4993 | |
4994 | class expand_operand ops[3]; |
4995 | create_output_operand (op: &ops[0], x: target, TYPE_MODE (TREE_TYPE (lhs))); |
4996 | create_input_operand (op: &ops[1], value: op1, TYPE_MODE (type)); |
4997 | create_input_operand (op: &ops[2], value: op2, TYPE_MODE (type)); |
4998 | insn_code icode = optab_handler (op: spaceship_optab, TYPE_MODE (type)); |
4999 | expand_insn (icode, nops: 3, ops); |
5000 | if (!rtx_equal_p (target, ops[0].value)) |
5001 | emit_move_insn (target, ops[0].value); |
5002 | } |
5003 | |
5004 | void |
5005 | expand_ASSUME (internal_fn, gcall *) |
5006 | { |
5007 | } |
5008 | |
5009 | void |
5010 | expand_MASK_CALL (internal_fn, gcall *) |
5011 | { |
5012 | /* This IFN should only exist between ifcvt and vect passes. */ |
5013 | gcc_unreachable (); |
5014 | } |
5015 | |
5016 | void |
5017 | expand_MULBITINT (internal_fn, gcall *stmt) |
5018 | { |
5019 | rtx_mode_t args[6]; |
5020 | for (int i = 0; i < 6; i++) |
5021 | args[i] = rtx_mode_t (expand_normal (exp: gimple_call_arg (gs: stmt, index: i)), |
5022 | (i & 1) ? SImode : ptr_mode); |
5023 | rtx fun = init_one_libfunc ("__mulbitint3" ); |
5024 | emit_library_call_value_1 (0, fun, NULL_RTX, LCT_NORMAL, VOIDmode, 6, args); |
5025 | } |
5026 | |
5027 | void |
5028 | expand_DIVMODBITINT (internal_fn, gcall *stmt) |
5029 | { |
5030 | rtx_mode_t args[8]; |
5031 | for (int i = 0; i < 8; i++) |
5032 | args[i] = rtx_mode_t (expand_normal (exp: gimple_call_arg (gs: stmt, index: i)), |
5033 | (i & 1) ? SImode : ptr_mode); |
5034 | rtx fun = init_one_libfunc ("__divmodbitint4" ); |
5035 | emit_library_call_value_1 (0, fun, NULL_RTX, LCT_NORMAL, VOIDmode, 8, args); |
5036 | } |
5037 | |
5038 | void |
5039 | expand_FLOATTOBITINT (internal_fn, gcall *stmt) |
5040 | { |
5041 | machine_mode mode = TYPE_MODE (TREE_TYPE (gimple_call_arg (stmt, 2))); |
5042 | rtx arg0 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
5043 | rtx arg1 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
5044 | rtx arg2 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 2)); |
5045 | const char *mname = GET_MODE_NAME (mode); |
5046 | unsigned mname_len = strlen (s: mname); |
5047 | int len = 12 + mname_len; |
5048 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5049 | len += 4; |
5050 | char *libfunc_name = XALLOCAVEC (char, len); |
5051 | char *p = libfunc_name; |
5052 | const char *q; |
5053 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5054 | { |
5055 | #if ENABLE_DECIMAL_BID_FORMAT |
5056 | memcpy (dest: p, src: "__bid_fix" , n: 9); |
5057 | #else |
5058 | memcpy (p, "__dpd_fix" , 9); |
5059 | #endif |
5060 | p += 9; |
5061 | } |
5062 | else |
5063 | { |
5064 | memcpy (dest: p, src: "__fix" , n: 5); |
5065 | p += 5; |
5066 | } |
5067 | for (q = mname; *q; q++) |
5068 | *p++ = TOLOWER (*q); |
5069 | memcpy (dest: p, src: "bitint" , n: 7); |
5070 | rtx fun = init_one_libfunc (libfunc_name); |
5071 | emit_library_call (fun, fn_type: LCT_NORMAL, VOIDmode, arg1: arg0, arg1_mode: ptr_mode, arg2: arg1, |
5072 | SImode, arg3: arg2, arg3_mode: mode); |
5073 | } |
5074 | |
5075 | void |
5076 | expand_BITINTTOFLOAT (internal_fn, gcall *stmt) |
5077 | { |
5078 | tree lhs = gimple_call_lhs (gs: stmt); |
5079 | if (!lhs) |
5080 | return; |
5081 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
5082 | rtx arg0 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
5083 | rtx arg1 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
5084 | const char *mname = GET_MODE_NAME (mode); |
5085 | unsigned mname_len = strlen (s: mname); |
5086 | int len = 14 + mname_len; |
5087 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5088 | len += 4; |
5089 | char *libfunc_name = XALLOCAVEC (char, len); |
5090 | char *p = libfunc_name; |
5091 | const char *q; |
5092 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5093 | { |
5094 | #if ENABLE_DECIMAL_BID_FORMAT |
5095 | memcpy (dest: p, src: "__bid_floatbitint" , n: 17); |
5096 | #else |
5097 | memcpy (p, "__dpd_floatbitint" , 17); |
5098 | #endif |
5099 | p += 17; |
5100 | } |
5101 | else |
5102 | { |
5103 | memcpy (dest: p, src: "__floatbitint" , n: 13); |
5104 | p += 13; |
5105 | } |
5106 | for (q = mname; *q; q++) |
5107 | *p++ = TOLOWER (*q); |
5108 | *p = '\0'; |
5109 | rtx fun = init_one_libfunc (libfunc_name); |
5110 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5111 | rtx val = emit_library_call_value (fun, value: target, fn_type: LCT_PURE, outmode: mode, |
5112 | arg1: arg0, arg1_mode: ptr_mode, arg2: arg1, SImode); |
5113 | if (val != target) |
5114 | emit_move_insn (target, val); |
5115 | } |
5116 | |