1 | /* Internal functions. |
---|---|
2 | Copyright (C) 2011-2025 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 | #define INCLUDE_MEMORY |
22 | #include "system.h" |
23 | #include "coretypes.h" |
24 | #include "backend.h" |
25 | #include "target.h" |
26 | #include "rtl.h" |
27 | #include "tree.h" |
28 | #include "gimple.h" |
29 | #include "predict.h" |
30 | #include "stringpool.h" |
31 | #include "tree-vrp.h" |
32 | #include "tree-ssanames.h" |
33 | #include "expmed.h" |
34 | #include "memmodel.h" |
35 | #include "optabs.h" |
36 | #include "emit-rtl.h" |
37 | #include "diagnostic-core.h" |
38 | #include "fold-const.h" |
39 | #include "internal-fn.h" |
40 | #include "stor-layout.h" |
41 | #include "dojump.h" |
42 | #include "expr.h" |
43 | #include "stringpool.h" |
44 | #include "attribs.h" |
45 | #include "asan.h" |
46 | #include "ubsan.h" |
47 | #include "recog.h" |
48 | #include "builtins.h" |
49 | #include "optabs-tree.h" |
50 | #include "gimple-ssa.h" |
51 | #include "tree-phinodes.h" |
52 | #include "ssa-iterators.h" |
53 | #include "explow.h" |
54 | #include "rtl-iter.h" |
55 | #include "gimple-range.h" |
56 | #include "fold-const-call.h" |
57 | #include "tree-ssa-live.h" |
58 | #include "tree-outof-ssa.h" |
59 | #include "gcc-urlifier.h" |
60 | |
61 | /* For lang_hooks.types.type_for_mode. */ |
62 | #include "langhooks.h" |
63 | |
64 | /* The names of each internal function, indexed by function number. */ |
65 | const char *const internal_fn_name_array[] = { |
66 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) #CODE, |
67 | #include "internal-fn.def" |
68 | "<invalid-fn>" |
69 | }; |
70 | |
71 | /* The ECF_* flags of each internal function, indexed by function number. */ |
72 | const int internal_fn_flags_array[] = { |
73 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) FLAGS, |
74 | #include "internal-fn.def" |
75 | 0 |
76 | }; |
77 | |
78 | /* Return the internal function called NAME, or IFN_LAST if there's |
79 | no such function. */ |
80 | |
81 | internal_fn |
82 | lookup_internal_fn (const char *name) |
83 | { |
84 | typedef hash_map<nofree_string_hash, internal_fn> name_to_fn_map_type; |
85 | static name_to_fn_map_type *name_to_fn_map; |
86 | |
87 | if (!name_to_fn_map) |
88 | { |
89 | name_to_fn_map = new name_to_fn_map_type (IFN_LAST); |
90 | for (unsigned int i = 0; i < IFN_LAST; ++i) |
91 | name_to_fn_map->put (k: internal_fn_name (fn: internal_fn (i)), |
92 | v: internal_fn (i)); |
93 | } |
94 | internal_fn *entry = name_to_fn_map->get (k: name); |
95 | return entry ? *entry : IFN_LAST; |
96 | } |
97 | |
98 | /* Geven an internal_fn IFN that is a widening function, return its |
99 | corresponding LO and HI internal_fns. */ |
100 | |
101 | extern void |
102 | lookup_hilo_internal_fn (internal_fn ifn, internal_fn *lo, internal_fn *hi) |
103 | { |
104 | gcc_assert (widening_fn_p (ifn)); |
105 | |
106 | switch (ifn) |
107 | { |
108 | default: |
109 | gcc_unreachable (); |
110 | #define DEF_INTERNAL_FN(NAME, FLAGS, TYPE) |
111 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ |
112 | case IFN_##NAME: \ |
113 | *lo = internal_fn (IFN_##NAME##_LO); \ |
114 | *hi = internal_fn (IFN_##NAME##_HI); \ |
115 | break; |
116 | #include "internal-fn.def" |
117 | } |
118 | } |
119 | |
120 | /* Given an internal_fn IFN that is a widening function, return its |
121 | corresponding _EVEN and _ODD internal_fns in *EVEN and *ODD. */ |
122 | |
123 | extern void |
124 | lookup_evenodd_internal_fn (internal_fn ifn, internal_fn *even, |
125 | internal_fn *odd) |
126 | { |
127 | gcc_assert (widening_fn_p (ifn)); |
128 | |
129 | switch (ifn) |
130 | { |
131 | default: |
132 | gcc_unreachable (); |
133 | #define DEF_INTERNAL_FN(NAME, FLAGS, TYPE) |
134 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ |
135 | case IFN_##NAME: \ |
136 | *even = internal_fn (IFN_##NAME##_EVEN); \ |
137 | *odd = internal_fn (IFN_##NAME##_ODD); \ |
138 | break; |
139 | #include "internal-fn.def" |
140 | } |
141 | } |
142 | |
143 | |
144 | /* Fnspec of each internal function, indexed by function number. */ |
145 | const_tree internal_fn_fnspec_array[IFN_LAST + 1]; |
146 | |
147 | void |
148 | init_internal_fns () |
149 | { |
150 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
151 | if (FNSPEC) internal_fn_fnspec_array[IFN_##CODE] = \ |
152 | build_string ((int) sizeof (FNSPEC) - 1, FNSPEC ? FNSPEC : ""); |
153 | #include "internal-fn.def" |
154 | internal_fn_fnspec_array[IFN_LAST] = 0; |
155 | } |
156 | |
157 | /* Create static initializers for the information returned by |
158 | direct_internal_fn. */ |
159 | #define not_direct { -2, -2, false } |
160 | #define mask_load_direct { -1, 2, false } |
161 | #define load_lanes_direct { -1, -1, false } |
162 | #define mask_load_lanes_direct { -1, -1, false } |
163 | #define gather_load_direct { 3, 1, false } |
164 | #define strided_load_direct { -1, -1, false } |
165 | #define len_load_direct { -1, -1, false } |
166 | #define mask_len_load_direct { -1, 4, false } |
167 | #define mask_store_direct { 3, 2, false } |
168 | #define store_lanes_direct { 0, 0, false } |
169 | #define mask_store_lanes_direct { 0, 0, false } |
170 | #define vec_cond_mask_direct { 1, 0, false } |
171 | #define vec_cond_mask_len_direct { 1, 1, false } |
172 | #define vec_cond_direct { 2, 0, false } |
173 | #define scatter_store_direct { 3, 1, false } |
174 | #define strided_store_direct { 1, 1, false } |
175 | #define len_store_direct { 3, 3, false } |
176 | #define mask_len_store_direct { 4, 5, false } |
177 | #define vec_set_direct { 3, 3, false } |
178 | #define vec_extract_direct { 0, -1, false } |
179 | #define unary_direct { 0, 0, true } |
180 | #define unary_convert_direct { -1, 0, true } |
181 | #define binary_direct { 0, 0, true } |
182 | #define ternary_direct { 0, 0, true } |
183 | #define cond_unary_direct { 1, 1, true } |
184 | #define cond_binary_direct { 1, 1, true } |
185 | #define cond_ternary_direct { 1, 1, true } |
186 | #define cond_len_unary_direct { 1, 1, true } |
187 | #define cond_len_binary_direct { 1, 1, true } |
188 | #define cond_len_ternary_direct { 1, 1, true } |
189 | #define while_direct { 0, 2, false } |
190 | #define fold_extract_direct { 2, 2, false } |
191 | #define fold_len_extract_direct { 2, 2, false } |
192 | #define fold_left_direct { 1, 1, false } |
193 | #define mask_fold_left_direct { 1, 1, false } |
194 | #define mask_len_fold_left_direct { 1, 1, false } |
195 | #define check_ptrs_direct { 0, 0, false } |
196 | #define crc_direct { 1, -1, true } |
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 | /* Like create_output_operand, but for callers that will use |
208 | assign_call_lhs afterwards. */ |
209 | |
210 | static void |
211 | create_call_lhs_operand (expand_operand *op, rtx lhs_rtx, machine_mode mode) |
212 | { |
213 | /* Do not assign directly to a promoted subreg, since there is no |
214 | guarantee that the instruction will leave the upper bits of the |
215 | register in the state required by SUBREG_PROMOTED_SIGN. */ |
216 | rtx dest = lhs_rtx; |
217 | if (dest && GET_CODE (dest) == SUBREG && SUBREG_PROMOTED_VAR_P (dest)) |
218 | dest = NULL_RTX; |
219 | create_output_operand (op, x: dest, mode); |
220 | } |
221 | |
222 | /* Move the result of an expanded instruction into the lhs of a gimple call. |
223 | LHS is the lhs of the call, LHS_RTX is its expanded form, and OP is the |
224 | result of the expanded instruction. OP should have been set up by |
225 | create_call_lhs_operand. */ |
226 | |
227 | static void |
228 | assign_call_lhs (tree lhs, rtx lhs_rtx, expand_operand *op) |
229 | { |
230 | if (rtx_equal_p (lhs_rtx, op->value)) |
231 | return; |
232 | |
233 | /* If the return value has an integral type, convert the instruction |
234 | result to that type. This is useful for things that return an |
235 | int regardless of the size of the input. If the instruction result |
236 | is smaller than required, assume that it is signed. |
237 | |
238 | If the return value has a nonintegral type, its mode must match |
239 | the instruction result. */ |
240 | if (GET_CODE (lhs_rtx) == SUBREG && SUBREG_PROMOTED_VAR_P (lhs_rtx)) |
241 | { |
242 | /* If this is a scalar in a register that is stored in a wider |
243 | mode than the declared mode, compute the result into its |
244 | declared mode and then convert to the wider mode. */ |
245 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
246 | rtx tmp = convert_to_mode (GET_MODE (lhs_rtx), op->value, 0); |
247 | convert_move (SUBREG_REG (lhs_rtx), tmp, |
248 | SUBREG_PROMOTED_SIGN (lhs_rtx)); |
249 | } |
250 | else if (GET_MODE (lhs_rtx) == GET_MODE (op->value)) |
251 | emit_move_insn (lhs_rtx, op->value); |
252 | else |
253 | { |
254 | gcc_checking_assert (INTEGRAL_TYPE_P (TREE_TYPE (lhs))); |
255 | convert_move (lhs_rtx, op->value, 0); |
256 | } |
257 | } |
258 | |
259 | /* Expand STMT using instruction ICODE. The instruction has NOUTPUTS |
260 | output operands and NINPUTS input operands, where NOUTPUTS is either |
261 | 0 or 1. The output operand (if any) comes first, followed by the |
262 | NINPUTS input operands. */ |
263 | |
264 | static void |
265 | expand_fn_using_insn (gcall *stmt, insn_code icode, unsigned int noutputs, |
266 | unsigned int ninputs) |
267 | { |
268 | gcc_assert (icode != CODE_FOR_nothing); |
269 | |
270 | expand_operand *ops = XALLOCAVEC (expand_operand, noutputs + ninputs); |
271 | unsigned int opno = 0; |
272 | rtx lhs_rtx = NULL_RTX; |
273 | tree lhs = gimple_call_lhs (gs: stmt); |
274 | |
275 | if (noutputs) |
276 | { |
277 | gcc_assert (noutputs == 1); |
278 | if (lhs) |
279 | lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
280 | create_call_lhs_operand (op: &ops[opno], lhs_rtx, |
281 | mode: insn_data[icode].operand[opno].mode); |
282 | opno += 1; |
283 | } |
284 | else |
285 | gcc_assert (!lhs); |
286 | |
287 | for (unsigned int i = 0; i < ninputs; ++i) |
288 | { |
289 | tree rhs = gimple_call_arg (gs: stmt, index: i); |
290 | tree rhs_type = TREE_TYPE (rhs); |
291 | rtx rhs_rtx = expand_normal (exp: rhs); |
292 | if (INTEGRAL_TYPE_P (rhs_type)) |
293 | create_convert_operand_from (op: &ops[opno], value: rhs_rtx, |
294 | TYPE_MODE (rhs_type), |
295 | TYPE_UNSIGNED (rhs_type)); |
296 | else if (TREE_CODE (rhs) == SSA_NAME |
297 | && SSA_NAME_IS_DEFAULT_DEF (rhs) |
298 | && VAR_P (SSA_NAME_VAR (rhs))) |
299 | create_undefined_input_operand (op: &ops[opno], TYPE_MODE (rhs_type)); |
300 | else if (VECTOR_BOOLEAN_TYPE_P (rhs_type) |
301 | && SCALAR_INT_MODE_P (TYPE_MODE (rhs_type)) |
302 | && maybe_ne (a: GET_MODE_PRECISION (TYPE_MODE (rhs_type)), |
303 | b: TYPE_VECTOR_SUBPARTS (node: rhs_type).to_constant ())) |
304 | { |
305 | /* Ensure that the vector bitmasks do not have excess bits. */ |
306 | int nunits = TYPE_VECTOR_SUBPARTS (node: rhs_type).to_constant (); |
307 | rtx tmp = expand_binop (TYPE_MODE (rhs_type), and_optab, rhs_rtx, |
308 | GEN_INT ((HOST_WIDE_INT_1U << nunits) - 1), |
309 | NULL_RTX, true, OPTAB_WIDEN); |
310 | create_input_operand (op: &ops[opno], value: tmp, TYPE_MODE (rhs_type)); |
311 | } |
312 | else |
313 | create_input_operand (op: &ops[opno], value: rhs_rtx, TYPE_MODE (rhs_type)); |
314 | opno += 1; |
315 | } |
316 | |
317 | gcc_assert (opno == noutputs + ninputs); |
318 | expand_insn (icode, nops: opno, ops); |
319 | if (lhs_rtx) |
320 | assign_call_lhs (lhs, lhs_rtx, op: &ops[0]); |
321 | } |
322 | |
323 | /* ARRAY_TYPE is an array of vector modes. Return the associated insn |
324 | for load-lanes-style optab OPTAB, or CODE_FOR_nothing if none. */ |
325 | |
326 | static enum insn_code |
327 | get_multi_vector_move (tree array_type, convert_optab optab) |
328 | { |
329 | machine_mode imode; |
330 | machine_mode vmode; |
331 | |
332 | gcc_assert (TREE_CODE (array_type) == ARRAY_TYPE); |
333 | imode = TYPE_MODE (array_type); |
334 | vmode = TYPE_MODE (TREE_TYPE (array_type)); |
335 | |
336 | return convert_optab_handler (op: optab, to_mode: imode, from_mode: vmode); |
337 | } |
338 | |
339 | /* Add mask, else, and len arguments according to the STMT. */ |
340 | |
341 | static unsigned int |
342 | add_mask_else_and_len_args (expand_operand *ops, unsigned int opno, gcall *stmt) |
343 | { |
344 | internal_fn ifn = gimple_call_internal_fn (gs: stmt); |
345 | int len_index = internal_fn_len_index (ifn); |
346 | /* BIAS is always consecutive next of LEN. */ |
347 | int bias_index = len_index + 1; |
348 | int mask_index = internal_fn_mask_index (ifn); |
349 | |
350 | /* The order of arguments is always {mask, else, len, bias}. */ |
351 | if (mask_index >= 0) |
352 | { |
353 | tree mask = gimple_call_arg (gs: stmt, index: mask_index); |
354 | rtx mask_rtx = expand_normal (exp: mask); |
355 | |
356 | tree mask_type = TREE_TYPE (mask); |
357 | if (VECTOR_BOOLEAN_TYPE_P (mask_type) |
358 | && SCALAR_INT_MODE_P (TYPE_MODE (mask_type)) |
359 | && maybe_ne (a: GET_MODE_PRECISION (TYPE_MODE (mask_type)), |
360 | b: TYPE_VECTOR_SUBPARTS (node: mask_type).to_constant ())) |
361 | { |
362 | /* Ensure that the vector bitmasks do not have excess bits. */ |
363 | int nunits = TYPE_VECTOR_SUBPARTS (node: mask_type).to_constant (); |
364 | mask_rtx = expand_binop (TYPE_MODE (mask_type), and_optab, mask_rtx, |
365 | GEN_INT ((HOST_WIDE_INT_1U << nunits) - 1), |
366 | NULL_RTX, true, OPTAB_WIDEN); |
367 | } |
368 | |
369 | create_input_operand (op: &ops[opno++], value: mask_rtx, |
370 | TYPE_MODE (TREE_TYPE (mask))); |
371 | } |
372 | |
373 | int els_index = internal_fn_else_index (ifn); |
374 | if (els_index >= 0) |
375 | { |
376 | tree els = gimple_call_arg (gs: stmt, index: els_index); |
377 | tree els_type = TREE_TYPE (els); |
378 | if (TREE_CODE (els) == SSA_NAME |
379 | && SSA_NAME_IS_DEFAULT_DEF (els) |
380 | && VAR_P (SSA_NAME_VAR (els))) |
381 | create_undefined_input_operand (op: &ops[opno++], TYPE_MODE (els_type)); |
382 | else |
383 | { |
384 | rtx els_rtx = expand_normal (exp: els); |
385 | create_input_operand (op: &ops[opno++], value: els_rtx, TYPE_MODE (els_type)); |
386 | } |
387 | } |
388 | if (len_index >= 0) |
389 | { |
390 | tree len = gimple_call_arg (gs: stmt, index: len_index); |
391 | rtx len_rtx = expand_normal (exp: len); |
392 | create_convert_operand_from (op: &ops[opno++], value: len_rtx, |
393 | TYPE_MODE (TREE_TYPE (len)), |
394 | TYPE_UNSIGNED (TREE_TYPE (len))); |
395 | tree biast = gimple_call_arg (gs: stmt, index: bias_index); |
396 | rtx bias = expand_normal (exp: biast); |
397 | create_input_operand (op: &ops[opno++], value: bias, QImode); |
398 | } |
399 | return opno; |
400 | } |
401 | |
402 | /* Expand LOAD_LANES call STMT using optab OPTAB. */ |
403 | |
404 | static void |
405 | expand_load_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
406 | { |
407 | class expand_operand ops[2]; |
408 | tree type, lhs, rhs; |
409 | rtx target, mem; |
410 | |
411 | lhs = gimple_call_lhs (gs: stmt); |
412 | rhs = gimple_call_arg (gs: stmt, index: 0); |
413 | type = TREE_TYPE (lhs); |
414 | |
415 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
416 | mem = expand_normal (exp: rhs); |
417 | |
418 | gcc_assert (MEM_P (mem)); |
419 | PUT_MODE (x: mem, TYPE_MODE (type)); |
420 | |
421 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, TYPE_MODE (type)); |
422 | create_fixed_operand (op: &ops[1], x: mem); |
423 | expand_insn (icode: get_multi_vector_move (array_type: type, optab), nops: 2, ops); |
424 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
425 | } |
426 | |
427 | /* Expand STORE_LANES call STMT using optab OPTAB. */ |
428 | |
429 | static void |
430 | expand_store_lanes_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
431 | { |
432 | class expand_operand ops[2]; |
433 | tree type, lhs, rhs; |
434 | rtx target, reg; |
435 | |
436 | lhs = gimple_call_lhs (gs: stmt); |
437 | rhs = gimple_call_arg (gs: stmt, index: 0); |
438 | type = TREE_TYPE (rhs); |
439 | |
440 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
441 | reg = expand_normal (exp: rhs); |
442 | |
443 | gcc_assert (MEM_P (target)); |
444 | PUT_MODE (x: target, TYPE_MODE (type)); |
445 | |
446 | create_fixed_operand (op: &ops[0], x: target); |
447 | create_input_operand (op: &ops[1], value: reg, TYPE_MODE (type)); |
448 | expand_insn (icode: get_multi_vector_move (array_type: type, optab), nops: 2, ops); |
449 | } |
450 | |
451 | static void |
452 | expand_ANNOTATE (internal_fn, gcall *) |
453 | { |
454 | gcc_unreachable (); |
455 | } |
456 | |
457 | /* This should get expanded in omp_device_lower pass. */ |
458 | |
459 | static void |
460 | expand_GOMP_USE_SIMT (internal_fn, gcall *) |
461 | { |
462 | gcc_unreachable (); |
463 | } |
464 | |
465 | /* This should get expanded in omp_device_lower pass. */ |
466 | |
467 | static void |
468 | expand_GOMP_SIMT_ENTER (internal_fn, gcall *) |
469 | { |
470 | gcc_unreachable (); |
471 | } |
472 | |
473 | /* Allocate per-lane storage and begin non-uniform execution region. */ |
474 | |
475 | static void |
476 | expand_GOMP_SIMT_ENTER_ALLOC (internal_fn, gcall *stmt) |
477 | { |
478 | rtx target; |
479 | tree lhs = gimple_call_lhs (gs: stmt); |
480 | if (lhs) |
481 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
482 | else |
483 | target = gen_reg_rtx (Pmode); |
484 | rtx size = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
485 | rtx align = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
486 | class expand_operand ops[3]; |
487 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, Pmode); |
488 | create_input_operand (op: &ops[1], value: size, Pmode); |
489 | create_input_operand (op: &ops[2], value: align, Pmode); |
490 | gcc_assert (targetm.have_omp_simt_enter ()); |
491 | expand_insn (icode: targetm.code_for_omp_simt_enter, nops: 3, ops); |
492 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
493 | } |
494 | |
495 | /* Deallocate per-lane storage and leave non-uniform execution region. */ |
496 | |
497 | static void |
498 | expand_GOMP_SIMT_EXIT (internal_fn, gcall *stmt) |
499 | { |
500 | gcc_checking_assert (!gimple_call_lhs (stmt)); |
501 | rtx arg = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
502 | class expand_operand ops[1]; |
503 | create_input_operand (op: &ops[0], value: arg, Pmode); |
504 | gcc_assert (targetm.have_omp_simt_exit ()); |
505 | expand_insn (icode: targetm.code_for_omp_simt_exit, nops: 1, ops); |
506 | } |
507 | |
508 | /* Lane index on SIMT targets: thread index in the warp on NVPTX. On targets |
509 | without SIMT execution this should be expanded in omp_device_lower pass. */ |
510 | |
511 | static void |
512 | expand_GOMP_SIMT_LANE (internal_fn, gcall *stmt) |
513 | { |
514 | tree lhs = gimple_call_lhs (gs: stmt); |
515 | if (!lhs) |
516 | return; |
517 | |
518 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
519 | gcc_assert (targetm.have_omp_simt_lane ()); |
520 | emit_insn (targetm.gen_omp_simt_lane (target)); |
521 | } |
522 | |
523 | /* This should get expanded in omp_device_lower pass. */ |
524 | |
525 | static void |
526 | expand_GOMP_SIMT_VF (internal_fn, gcall *) |
527 | { |
528 | gcc_unreachable (); |
529 | } |
530 | |
531 | /* This should get expanded in omp_device_lower pass. */ |
532 | |
533 | static void |
534 | expand_GOMP_MAX_VF (internal_fn, gcall *) |
535 | { |
536 | gcc_unreachable (); |
537 | } |
538 | |
539 | /* This should get expanded in omp_device_lower pass. */ |
540 | |
541 | static void |
542 | expand_GOMP_TARGET_REV (internal_fn, gcall *) |
543 | { |
544 | gcc_unreachable (); |
545 | } |
546 | |
547 | /* Lane index of the first SIMT lane that supplies a non-zero argument. |
548 | This is a SIMT counterpart to GOMP_SIMD_LAST_LANE, used to represent the |
549 | lane that executed the last iteration for handling OpenMP lastprivate. */ |
550 | |
551 | static void |
552 | expand_GOMP_SIMT_LAST_LANE (internal_fn, gcall *stmt) |
553 | { |
554 | tree lhs = gimple_call_lhs (gs: stmt); |
555 | if (!lhs) |
556 | return; |
557 | |
558 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
559 | rtx cond = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
560 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
561 | class expand_operand ops[2]; |
562 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
563 | create_input_operand (op: &ops[1], value: cond, mode); |
564 | gcc_assert (targetm.have_omp_simt_last_lane ()); |
565 | expand_insn (icode: targetm.code_for_omp_simt_last_lane, nops: 2, ops); |
566 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
567 | } |
568 | |
569 | /* Non-transparent predicate used in SIMT lowering of OpenMP "ordered". */ |
570 | |
571 | static void |
572 | expand_GOMP_SIMT_ORDERED_PRED (internal_fn, gcall *stmt) |
573 | { |
574 | tree lhs = gimple_call_lhs (gs: stmt); |
575 | if (!lhs) |
576 | return; |
577 | |
578 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
579 | rtx ctr = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
580 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
581 | class expand_operand ops[2]; |
582 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
583 | create_input_operand (op: &ops[1], value: ctr, mode); |
584 | gcc_assert (targetm.have_omp_simt_ordered ()); |
585 | expand_insn (icode: targetm.code_for_omp_simt_ordered, nops: 2, ops); |
586 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
587 | } |
588 | |
589 | /* "Or" boolean reduction across SIMT lanes: return non-zero in all lanes if |
590 | any lane supplies a non-zero argument. */ |
591 | |
592 | static void |
593 | expand_GOMP_SIMT_VOTE_ANY (internal_fn, gcall *stmt) |
594 | { |
595 | tree lhs = gimple_call_lhs (gs: stmt); |
596 | if (!lhs) |
597 | return; |
598 | |
599 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
600 | rtx cond = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
601 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
602 | class expand_operand ops[2]; |
603 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
604 | create_input_operand (op: &ops[1], value: cond, mode); |
605 | gcc_assert (targetm.have_omp_simt_vote_any ()); |
606 | expand_insn (icode: targetm.code_for_omp_simt_vote_any, nops: 2, ops); |
607 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
608 | } |
609 | |
610 | /* Exchange between SIMT lanes with a "butterfly" pattern: source lane index |
611 | is destination lane index XOR given offset. */ |
612 | |
613 | static void |
614 | expand_GOMP_SIMT_XCHG_BFLY (internal_fn, gcall *stmt) |
615 | { |
616 | tree lhs = gimple_call_lhs (gs: stmt); |
617 | if (!lhs) |
618 | return; |
619 | |
620 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
621 | rtx src = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
622 | rtx idx = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
623 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
624 | class expand_operand ops[3]; |
625 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
626 | create_input_operand (op: &ops[1], value: src, mode); |
627 | create_input_operand (op: &ops[2], value: idx, SImode); |
628 | gcc_assert (targetm.have_omp_simt_xchg_bfly ()); |
629 | expand_insn (icode: targetm.code_for_omp_simt_xchg_bfly, nops: 3, ops); |
630 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
631 | } |
632 | |
633 | /* Exchange between SIMT lanes according to given source lane index. */ |
634 | |
635 | static void |
636 | expand_GOMP_SIMT_XCHG_IDX (internal_fn, gcall *stmt) |
637 | { |
638 | tree lhs = gimple_call_lhs (gs: stmt); |
639 | if (!lhs) |
640 | return; |
641 | |
642 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
643 | rtx src = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
644 | rtx idx = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
645 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
646 | class expand_operand ops[3]; |
647 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
648 | create_input_operand (op: &ops[1], value: src, mode); |
649 | create_input_operand (op: &ops[2], value: idx, SImode); |
650 | gcc_assert (targetm.have_omp_simt_xchg_idx ()); |
651 | expand_insn (icode: targetm.code_for_omp_simt_xchg_idx, nops: 3, ops); |
652 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
653 | } |
654 | |
655 | /* This should get expanded in adjust_simduid_builtins. */ |
656 | |
657 | static void |
658 | expand_GOMP_SIMD_LANE (internal_fn, gcall *) |
659 | { |
660 | gcc_unreachable (); |
661 | } |
662 | |
663 | /* This should get expanded in adjust_simduid_builtins. */ |
664 | |
665 | static void |
666 | expand_GOMP_SIMD_VF (internal_fn, gcall *) |
667 | { |
668 | gcc_unreachable (); |
669 | } |
670 | |
671 | /* This should get expanded in adjust_simduid_builtins. */ |
672 | |
673 | static void |
674 | expand_GOMP_SIMD_LAST_LANE (internal_fn, gcall *) |
675 | { |
676 | gcc_unreachable (); |
677 | } |
678 | |
679 | /* This should get expanded in adjust_simduid_builtins. */ |
680 | |
681 | static void |
682 | expand_GOMP_SIMD_ORDERED_START (internal_fn, gcall *) |
683 | { |
684 | gcc_unreachable (); |
685 | } |
686 | |
687 | /* This should get expanded in adjust_simduid_builtins. */ |
688 | |
689 | static void |
690 | expand_GOMP_SIMD_ORDERED_END (internal_fn, gcall *) |
691 | { |
692 | gcc_unreachable (); |
693 | } |
694 | |
695 | /* This should get expanded in gimplify_omp_dispatch. */ |
696 | |
697 | static void |
698 | expand_GOMP_DISPATCH (internal_fn, gcall *) |
699 | { |
700 | gcc_unreachable (); |
701 | } |
702 | |
703 | /* This should get expanded in the sanopt pass. */ |
704 | |
705 | static void |
706 | expand_UBSAN_NULL (internal_fn, gcall *) |
707 | { |
708 | gcc_unreachable (); |
709 | } |
710 | |
711 | /* This should get expanded in the sanopt pass. */ |
712 | |
713 | static void |
714 | expand_UBSAN_BOUNDS (internal_fn, gcall *) |
715 | { |
716 | gcc_unreachable (); |
717 | } |
718 | |
719 | /* This should get expanded in the sanopt pass. */ |
720 | |
721 | static void |
722 | expand_UBSAN_VPTR (internal_fn, gcall *) |
723 | { |
724 | gcc_unreachable (); |
725 | } |
726 | |
727 | /* This should get expanded in the sanopt pass. */ |
728 | |
729 | static void |
730 | expand_UBSAN_PTR (internal_fn, gcall *) |
731 | { |
732 | gcc_unreachable (); |
733 | } |
734 | |
735 | /* This should get expanded in the sanopt pass. */ |
736 | |
737 | static void |
738 | expand_UBSAN_OBJECT_SIZE (internal_fn, gcall *) |
739 | { |
740 | gcc_unreachable (); |
741 | } |
742 | |
743 | /* This should get expanded in the sanopt pass. */ |
744 | |
745 | static void |
746 | expand_HWASAN_CHECK (internal_fn, gcall *) |
747 | { |
748 | gcc_unreachable (); |
749 | } |
750 | |
751 | /* For hwasan stack tagging: |
752 | Clear tags on the dynamically allocated space. |
753 | For use after an object dynamically allocated on the stack goes out of |
754 | scope. */ |
755 | static void |
756 | expand_HWASAN_ALLOCA_UNPOISON (internal_fn, gcall *gc) |
757 | { |
758 | gcc_assert (Pmode == ptr_mode); |
759 | tree restored_position = gimple_call_arg (gs: gc, index: 0); |
760 | rtx restored_rtx = expand_expr (exp: restored_position, NULL_RTX, VOIDmode, |
761 | modifier: EXPAND_NORMAL); |
762 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); |
763 | rtx off = expand_simple_binop (Pmode, MINUS, restored_rtx, |
764 | stack_pointer_rtx, NULL_RTX, 0, |
765 | OPTAB_WIDEN); |
766 | emit_library_call_value (fun: func, NULL_RTX, fn_type: LCT_NORMAL, VOIDmode, |
767 | virtual_stack_dynamic_rtx, Pmode, |
768 | HWASAN_STACK_BACKGROUND, QImode, |
769 | arg3: off, Pmode); |
770 | } |
771 | |
772 | /* For hwasan stack tagging: |
773 | Return a tag to be used for a dynamic allocation. */ |
774 | static void |
775 | expand_HWASAN_CHOOSE_TAG (internal_fn, gcall *gc) |
776 | { |
777 | tree tag = gimple_call_lhs (gs: gc); |
778 | rtx target = expand_expr (exp: tag, NULL_RTX, VOIDmode, modifier: EXPAND_NORMAL); |
779 | machine_mode mode = GET_MODE (target); |
780 | gcc_assert (mode == QImode); |
781 | |
782 | rtx base_tag = targetm.memtag.extract_tag (hwasan_frame_base (), NULL_RTX); |
783 | gcc_assert (base_tag); |
784 | rtx tag_offset = gen_int_mode (hwasan_current_frame_tag (), QImode); |
785 | rtx chosen_tag = expand_simple_binop (QImode, PLUS, base_tag, tag_offset, |
786 | target, /* unsignedp = */1, |
787 | OPTAB_WIDEN); |
788 | chosen_tag = hwasan_truncate_to_tag_size (chosen_tag, target); |
789 | |
790 | /* Really need to put the tag into the `target` RTX. */ |
791 | if (chosen_tag != target) |
792 | { |
793 | rtx temp = chosen_tag; |
794 | gcc_assert (GET_MODE (chosen_tag) == mode); |
795 | emit_move_insn (target, temp); |
796 | } |
797 | |
798 | hwasan_increment_frame_tag (); |
799 | } |
800 | |
801 | /* For hwasan stack tagging: |
802 | Tag a region of space in the shadow stack according to the base pointer of |
803 | an object on the stack. N.b. the length provided in the internal call is |
804 | required to be aligned to HWASAN_TAG_GRANULE_SIZE. */ |
805 | static void |
806 | expand_HWASAN_MARK (internal_fn, gcall *gc) |
807 | { |
808 | gcc_assert (ptr_mode == Pmode); |
809 | HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gs: gc, index: 0)); |
810 | bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON; |
811 | |
812 | tree base = gimple_call_arg (gs: gc, index: 1); |
813 | gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR); |
814 | rtx base_rtx = expand_normal (exp: base); |
815 | |
816 | rtx tag = is_poison ? HWASAN_STACK_BACKGROUND |
817 | : targetm.memtag.extract_tag (base_rtx, NULL_RTX); |
818 | rtx address = targetm.memtag.untagged_pointer (base_rtx, NULL_RTX); |
819 | |
820 | tree len = gimple_call_arg (gs: gc, index: 2); |
821 | rtx r_len = expand_normal (exp: len); |
822 | |
823 | rtx func = init_one_libfunc ("__hwasan_tag_memory"); |
824 | emit_library_call (fun: func, fn_type: LCT_NORMAL, VOIDmode, arg1: address, Pmode, |
825 | arg2: tag, QImode, arg3: r_len, Pmode); |
826 | } |
827 | |
828 | /* For hwasan stack tagging: |
829 | Store a tag into a pointer. */ |
830 | static void |
831 | expand_HWASAN_SET_TAG (internal_fn, gcall *gc) |
832 | { |
833 | gcc_assert (ptr_mode == Pmode); |
834 | tree g_target = gimple_call_lhs (gs: gc); |
835 | tree g_ptr = gimple_call_arg (gs: gc, index: 0); |
836 | tree g_tag = gimple_call_arg (gs: gc, index: 1); |
837 | |
838 | rtx ptr = expand_normal (exp: g_ptr); |
839 | rtx tag = expand_expr (exp: g_tag, NULL_RTX, QImode, modifier: EXPAND_NORMAL); |
840 | rtx target = expand_normal (exp: g_target); |
841 | |
842 | rtx untagged = targetm.memtag.untagged_pointer (ptr, target); |
843 | rtx tagged_value = targetm.memtag.set_tag (untagged, tag, target); |
844 | if (tagged_value != target) |
845 | emit_move_insn (target, tagged_value); |
846 | } |
847 | |
848 | /* This should get expanded in the sanopt pass. */ |
849 | |
850 | static void |
851 | expand_ASAN_CHECK (internal_fn, gcall *) |
852 | { |
853 | gcc_unreachable (); |
854 | } |
855 | |
856 | /* This should get expanded in the sanopt pass. */ |
857 | |
858 | static void |
859 | expand_ASAN_MARK (internal_fn, gcall *) |
860 | { |
861 | gcc_unreachable (); |
862 | } |
863 | |
864 | /* This should get expanded in the sanopt pass. */ |
865 | |
866 | static void |
867 | expand_ASAN_POISON (internal_fn, gcall *) |
868 | { |
869 | gcc_unreachable (); |
870 | } |
871 | |
872 | /* This should get expanded in the sanopt pass. */ |
873 | |
874 | static void |
875 | expand_ASAN_POISON_USE (internal_fn, gcall *) |
876 | { |
877 | gcc_unreachable (); |
878 | } |
879 | |
880 | /* This should get expanded in the tsan pass. */ |
881 | |
882 | static void |
883 | expand_TSAN_FUNC_EXIT (internal_fn, gcall *) |
884 | { |
885 | gcc_unreachable (); |
886 | } |
887 | |
888 | /* This should get expanded in the lower pass. */ |
889 | |
890 | static void |
891 | expand_FALLTHROUGH (internal_fn, gcall *call) |
892 | { |
893 | auto_urlify_attributes sentinel; |
894 | error_at (gimple_location (g: call), |
895 | "invalid use of attribute %<fallthrough%>"); |
896 | } |
897 | |
898 | /* Return minimum precision needed to represent all values |
899 | of ARG in SIGNed integral type. */ |
900 | |
901 | static int |
902 | get_min_precision (tree arg, signop sign) |
903 | { |
904 | int prec = TYPE_PRECISION (TREE_TYPE (arg)); |
905 | int cnt = 0; |
906 | signop orig_sign = sign; |
907 | if (TREE_CODE (arg) == INTEGER_CST) |
908 | { |
909 | int p; |
910 | if (TYPE_SIGN (TREE_TYPE (arg)) != sign) |
911 | { |
912 | widest_int w = wi::to_widest (t: arg); |
913 | w = wi::ext (x: w, offset: prec, sgn: sign); |
914 | p = wi::min_precision (x: w, sgn: sign); |
915 | } |
916 | else |
917 | p = wi::min_precision (x: wi::to_wide (t: arg), sgn: sign); |
918 | return MIN (p, prec); |
919 | } |
920 | while (CONVERT_EXPR_P (arg) |
921 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) |
922 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) <= prec) |
923 | { |
924 | arg = TREE_OPERAND (arg, 0); |
925 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) |
926 | { |
927 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) |
928 | sign = UNSIGNED; |
929 | else if (sign == UNSIGNED |
930 | && (get_range_pos_neg (arg, |
931 | currently_expanding_gimple_stmt) |
932 | != 1)) |
933 | return prec + (orig_sign != sign); |
934 | prec = TYPE_PRECISION (TREE_TYPE (arg)); |
935 | } |
936 | if (++cnt > 30) |
937 | return prec + (orig_sign != sign); |
938 | } |
939 | if (CONVERT_EXPR_P (arg) |
940 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (arg, 0))) |
941 | && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (arg, 0))) > prec) |
942 | { |
943 | /* We have e.g. (unsigned short) y_2 where int y_2 = (int) x_1(D); |
944 | If y_2's min precision is smaller than prec, return that. */ |
945 | int oprec = get_min_precision (TREE_OPERAND (arg, 0), sign); |
946 | if (oprec < prec) |
947 | return oprec + (orig_sign != sign); |
948 | } |
949 | if (TREE_CODE (arg) != SSA_NAME) |
950 | return prec + (orig_sign != sign); |
951 | int_range_max r; |
952 | gimple *cg = currently_expanding_gimple_stmt; |
953 | while (!get_range_query (cfun)->range_of_expr (r, expr: arg, cg) |
954 | || r.varying_p () |
955 | || r.undefined_p ()) |
956 | { |
957 | gimple *g = SSA_NAME_DEF_STMT (arg); |
958 | if (is_gimple_assign (gs: g) |
959 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (g))) |
960 | { |
961 | tree t = gimple_assign_rhs1 (gs: g); |
962 | if (INTEGRAL_TYPE_P (TREE_TYPE (t)) |
963 | && TYPE_PRECISION (TREE_TYPE (t)) <= prec) |
964 | { |
965 | arg = t; |
966 | if (TYPE_PRECISION (TREE_TYPE (arg)) < prec) |
967 | { |
968 | if (TYPE_UNSIGNED (TREE_TYPE (arg))) |
969 | sign = UNSIGNED; |
970 | else if (sign == UNSIGNED |
971 | && get_range_pos_neg (arg, g) != 1) |
972 | return prec + (orig_sign != sign); |
973 | prec = TYPE_PRECISION (TREE_TYPE (arg)); |
974 | } |
975 | if (++cnt > 30) |
976 | return prec + (orig_sign != sign); |
977 | continue; |
978 | } |
979 | } |
980 | return prec + (orig_sign != sign); |
981 | } |
982 | if (sign == TYPE_SIGN (TREE_TYPE (arg))) |
983 | { |
984 | int p1 = wi::min_precision (x: r.lower_bound (), sgn: sign); |
985 | int p2 = wi::min_precision (x: r.upper_bound (), sgn: sign); |
986 | p1 = MAX (p1, p2); |
987 | prec = MIN (prec, p1); |
988 | } |
989 | else if (sign == UNSIGNED && !wi::neg_p (x: r.lower_bound (), sgn: SIGNED)) |
990 | { |
991 | int p = wi::min_precision (x: r.upper_bound (), sgn: UNSIGNED); |
992 | prec = MIN (prec, p); |
993 | } |
994 | return prec + (orig_sign != sign); |
995 | } |
996 | |
997 | /* Helper for expand_*_overflow. Set the __imag__ part to true |
998 | (1 except for signed:1 type, in which case store -1). */ |
999 | |
1000 | static void |
1001 | expand_arith_set_overflow (tree lhs, rtx target) |
1002 | { |
1003 | if (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))) == 1 |
1004 | && !TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs)))) |
1005 | write_complex_part (target, constm1_rtx, true, false); |
1006 | else |
1007 | write_complex_part (target, const1_rtx, true, false); |
1008 | } |
1009 | |
1010 | /* Helper for expand_*_overflow. Store RES into the __real__ part |
1011 | of TARGET. If RES has larger MODE than __real__ part of TARGET, |
1012 | set the __imag__ part to 1 if RES doesn't fit into it. Similarly |
1013 | if LHS has smaller precision than its mode. */ |
1014 | |
1015 | static void |
1016 | expand_arith_overflow_result_store (tree lhs, rtx target, |
1017 | scalar_int_mode mode, rtx res) |
1018 | { |
1019 | scalar_int_mode tgtmode |
1020 | = as_a <scalar_int_mode> (GET_MODE_INNER (GET_MODE (target))); |
1021 | rtx lres = res; |
1022 | if (tgtmode != mode) |
1023 | { |
1024 | rtx_code_label *done_label = gen_label_rtx (); |
1025 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); |
1026 | lres = convert_modes (mode: tgtmode, oldmode: mode, x: res, unsignedp: uns); |
1027 | gcc_assert (GET_MODE_PRECISION (tgtmode) < GET_MODE_PRECISION (mode)); |
1028 | do_compare_rtx_and_jump (res, convert_modes (mode, oldmode: tgtmode, x: lres, unsignedp: uns), |
1029 | EQ, true, mode, NULL_RTX, NULL, done_label, |
1030 | profile_probability::very_likely ()); |
1031 | expand_arith_set_overflow (lhs, target); |
1032 | emit_label (done_label); |
1033 | } |
1034 | int prec = TYPE_PRECISION (TREE_TYPE (TREE_TYPE (lhs))); |
1035 | int tgtprec = GET_MODE_PRECISION (mode: tgtmode); |
1036 | if (prec < tgtprec) |
1037 | { |
1038 | rtx_code_label *done_label = gen_label_rtx (); |
1039 | int uns = TYPE_UNSIGNED (TREE_TYPE (TREE_TYPE (lhs))); |
1040 | res = lres; |
1041 | if (uns) |
1042 | { |
1043 | rtx mask |
1044 | = immed_wide_int_const (wi::shifted_mask (start: 0, width: prec, negate_p: false, precision: tgtprec), |
1045 | tgtmode); |
1046 | lres = expand_simple_binop (tgtmode, AND, res, mask, NULL_RTX, |
1047 | true, OPTAB_LIB_WIDEN); |
1048 | } |
1049 | else |
1050 | { |
1051 | lres = expand_shift (LSHIFT_EXPR, tgtmode, res, tgtprec - prec, |
1052 | NULL_RTX, 1); |
1053 | lres = expand_shift (RSHIFT_EXPR, tgtmode, lres, tgtprec - prec, |
1054 | NULL_RTX, 0); |
1055 | } |
1056 | do_compare_rtx_and_jump (res, lres, |
1057 | EQ, true, tgtmode, NULL_RTX, NULL, done_label, |
1058 | profile_probability::very_likely ()); |
1059 | expand_arith_set_overflow (lhs, target); |
1060 | emit_label (done_label); |
1061 | } |
1062 | write_complex_part (target, lres, false, false); |
1063 | } |
1064 | |
1065 | /* Helper for expand_*_overflow. Store RES into TARGET. */ |
1066 | |
1067 | static void |
1068 | expand_ubsan_result_store (tree lhs, rtx target, scalar_int_mode mode, |
1069 | rtx res, rtx_code_label *do_error) |
1070 | { |
1071 | if (TREE_CODE (TREE_TYPE (lhs)) == BITINT_TYPE |
1072 | && TYPE_PRECISION (TREE_TYPE (lhs)) < GET_MODE_PRECISION (mode)) |
1073 | { |
1074 | int uns = TYPE_UNSIGNED (TREE_TYPE (lhs)); |
1075 | int prec = TYPE_PRECISION (TREE_TYPE (lhs)); |
1076 | int tgtprec = GET_MODE_PRECISION (mode); |
1077 | rtx resc = gen_reg_rtx (mode), lres; |
1078 | emit_move_insn (resc, res); |
1079 | if (uns) |
1080 | { |
1081 | rtx mask |
1082 | = immed_wide_int_const (wi::shifted_mask (start: 0, width: prec, negate_p: false, precision: tgtprec), |
1083 | mode); |
1084 | lres = expand_simple_binop (mode, AND, res, mask, NULL_RTX, |
1085 | true, OPTAB_LIB_WIDEN); |
1086 | } |
1087 | else |
1088 | { |
1089 | lres = expand_shift (LSHIFT_EXPR, mode, res, tgtprec - prec, |
1090 | NULL_RTX, 1); |
1091 | lres = expand_shift (RSHIFT_EXPR, mode, lres, tgtprec - prec, |
1092 | NULL_RTX, 0); |
1093 | } |
1094 | if (lres != res) |
1095 | emit_move_insn (res, lres); |
1096 | do_compare_rtx_and_jump (res, resc, |
1097 | NE, true, mode, NULL_RTX, NULL, do_error, |
1098 | profile_probability::very_unlikely ()); |
1099 | } |
1100 | if (GET_CODE (target) == SUBREG && SUBREG_PROMOTED_VAR_P (target)) |
1101 | /* If this is a scalar in a register that is stored in a wider mode |
1102 | than the declared mode, compute the result into its declared mode |
1103 | and then convert to the wider mode. Our value is the computed |
1104 | expression. */ |
1105 | convert_move (SUBREG_REG (target), res, SUBREG_PROMOTED_SIGN (target)); |
1106 | else |
1107 | emit_move_insn (target, res); |
1108 | } |
1109 | |
1110 | /* Add sub/add overflow checking to the statement STMT. |
1111 | CODE says whether the operation is +, or -. */ |
1112 | |
1113 | void |
1114 | expand_addsub_overflow (location_t loc, tree_code code, tree lhs, |
1115 | tree arg0, tree arg1, bool unsr_p, bool uns0_p, |
1116 | bool uns1_p, bool is_ubsan, tree *datap) |
1117 | { |
1118 | rtx res, target = NULL_RTX; |
1119 | tree fn; |
1120 | rtx_code_label *done_label = gen_label_rtx (); |
1121 | rtx_code_label *do_error = gen_label_rtx (); |
1122 | do_pending_stack_adjust (); |
1123 | rtx op0 = expand_normal (exp: arg0); |
1124 | rtx op1 = expand_normal (exp: arg1); |
1125 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1126 | int prec = GET_MODE_PRECISION (mode); |
1127 | rtx sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); |
1128 | bool do_xor = false; |
1129 | |
1130 | if (is_ubsan) |
1131 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); |
1132 | |
1133 | if (lhs) |
1134 | { |
1135 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1136 | if (!is_ubsan) |
1137 | write_complex_part (target, const0_rtx, true, false); |
1138 | } |
1139 | |
1140 | /* We assume both operands and result have the same precision |
1141 | here (GET_MODE_BITSIZE (mode)), S stands for signed type |
1142 | with that precision, U for unsigned type with that precision, |
1143 | sgn for unsigned most significant bit in that precision. |
1144 | s1 is signed first operand, u1 is unsigned first operand, |
1145 | s2 is signed second operand, u2 is unsigned second operand, |
1146 | sr is signed result, ur is unsigned result and the following |
1147 | rules say how to compute result (which is always result of |
1148 | the operands as if both were unsigned, cast to the right |
1149 | signedness) and how to compute whether operation overflowed. |
1150 | |
1151 | s1 + s2 -> sr |
1152 | res = (S) ((U) s1 + (U) s2) |
1153 | ovf = s2 < 0 ? res > s1 : res < s1 (or jump on overflow) |
1154 | s1 - s2 -> sr |
1155 | res = (S) ((U) s1 - (U) s2) |
1156 | ovf = s2 < 0 ? res < s1 : res > s2 (or jump on overflow) |
1157 | u1 + u2 -> ur |
1158 | res = u1 + u2 |
1159 | ovf = res < u1 (or jump on carry, but RTL opts will handle it) |
1160 | u1 - u2 -> ur |
1161 | res = u1 - u2 |
1162 | ovf = res > u1 (or jump on carry, but RTL opts will handle it) |
1163 | s1 + u2 -> sr |
1164 | res = (S) ((U) s1 + u2) |
1165 | ovf = ((U) res ^ sgn) < u2 |
1166 | s1 + u2 -> ur |
1167 | t1 = (S) (u2 ^ sgn) |
1168 | t2 = s1 + t1 |
1169 | res = (U) t2 ^ sgn |
1170 | ovf = t1 < 0 ? t2 > s1 : t2 < s1 (or jump on overflow) |
1171 | s1 - u2 -> sr |
1172 | res = (S) ((U) s1 - u2) |
1173 | ovf = u2 > ((U) s1 ^ sgn) |
1174 | s1 - u2 -> ur |
1175 | res = (U) s1 - u2 |
1176 | ovf = s1 < 0 || u2 > (U) s1 |
1177 | u1 - s2 -> sr |
1178 | res = u1 - (U) s2 |
1179 | ovf = u1 >= ((U) s2 ^ sgn) |
1180 | u1 - s2 -> ur |
1181 | t1 = u1 ^ sgn |
1182 | t2 = t1 - (U) s2 |
1183 | res = t2 ^ sgn |
1184 | ovf = s2 < 0 ? (S) t2 < (S) t1 : (S) t2 > (S) t1 (or jump on overflow) |
1185 | s1 + s2 -> ur |
1186 | res = (U) s1 + (U) s2 |
1187 | ovf = s2 < 0 ? (s1 | (S) res) < 0) : (s1 & (S) res) < 0) |
1188 | u1 + u2 -> sr |
1189 | res = (S) (u1 + u2) |
1190 | ovf = (U) res < u2 || res < 0 |
1191 | u1 - u2 -> sr |
1192 | res = (S) (u1 - u2) |
1193 | ovf = u1 >= u2 ? res < 0 : res >= 0 |
1194 | s1 - s2 -> ur |
1195 | res = (U) s1 - (U) s2 |
1196 | ovf = s2 >= 0 ? ((s1 | (S) res) < 0) : ((s1 & (S) res) < 0) */ |
1197 | |
1198 | if (code == PLUS_EXPR && uns0_p && !uns1_p) |
1199 | { |
1200 | /* PLUS_EXPR is commutative, if operand signedness differs, |
1201 | canonicalize to the first operand being signed and second |
1202 | unsigned to simplify following code. */ |
1203 | std::swap (a&: op0, b&: op1); |
1204 | std::swap (a&: arg0, b&: arg1); |
1205 | uns0_p = false; |
1206 | uns1_p = true; |
1207 | } |
1208 | |
1209 | /* u1 +- u2 -> ur */ |
1210 | if (uns0_p && uns1_p && unsr_p) |
1211 | { |
1212 | insn_code icode = optab_handler (op: code == PLUS_EXPR ? uaddv4_optab |
1213 | : usubv4_optab, mode); |
1214 | if (icode != CODE_FOR_nothing) |
1215 | { |
1216 | class expand_operand ops[4]; |
1217 | rtx_insn *last = get_last_insn (); |
1218 | |
1219 | res = gen_reg_rtx (mode); |
1220 | create_output_operand (op: &ops[0], x: res, mode); |
1221 | create_input_operand (op: &ops[1], value: op0, mode); |
1222 | create_input_operand (op: &ops[2], value: op1, mode); |
1223 | create_fixed_operand (op: &ops[3], x: do_error); |
1224 | if (maybe_expand_insn (icode, nops: 4, ops)) |
1225 | { |
1226 | last = get_last_insn (); |
1227 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1228 | && JUMP_P (last) |
1229 | && any_condjump_p (last) |
1230 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1231 | add_reg_br_prob_note (last, |
1232 | profile_probability::very_unlikely ()); |
1233 | emit_jump (done_label); |
1234 | goto do_error_label; |
1235 | } |
1236 | |
1237 | delete_insns_since (last); |
1238 | } |
1239 | |
1240 | /* Compute the operation. On RTL level, the addition is always |
1241 | unsigned. */ |
1242 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1243 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1244 | rtx tem = op0; |
1245 | /* For PLUS_EXPR, the operation is commutative, so we can pick |
1246 | operand to compare against. For prec <= BITS_PER_WORD, I think |
1247 | preferring REG operand is better over CONST_INT, because |
1248 | the CONST_INT might enlarge the instruction or CSE would need |
1249 | to figure out we'd already loaded it into a register before. |
1250 | For prec > BITS_PER_WORD, I think CONST_INT might be more beneficial, |
1251 | as then the multi-word comparison can be perhaps simplified. */ |
1252 | if (code == PLUS_EXPR |
1253 | && (prec <= BITS_PER_WORD |
1254 | ? (CONST_SCALAR_INT_P (op0) && REG_P (op1)) |
1255 | : CONST_SCALAR_INT_P (op1))) |
1256 | tem = op1; |
1257 | do_compare_rtx_and_jump (res, tem, code == PLUS_EXPR ? GEU : LEU, |
1258 | true, mode, NULL_RTX, NULL, done_label, |
1259 | profile_probability::very_likely ()); |
1260 | goto do_error_label; |
1261 | } |
1262 | |
1263 | /* s1 +- u2 -> sr */ |
1264 | if (!uns0_p && uns1_p && !unsr_p) |
1265 | { |
1266 | /* Compute the operation. On RTL level, the addition is always |
1267 | unsigned. */ |
1268 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1269 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1270 | rtx tem = expand_binop (mode, add_optab, |
1271 | code == PLUS_EXPR ? res : op0, sgn, |
1272 | NULL_RTX, false, OPTAB_LIB_WIDEN); |
1273 | do_compare_rtx_and_jump (tem, op1, GEU, true, mode, NULL_RTX, NULL, |
1274 | done_label, profile_probability::very_likely ()); |
1275 | goto do_error_label; |
1276 | } |
1277 | |
1278 | /* s1 + u2 -> ur */ |
1279 | if (code == PLUS_EXPR && !uns0_p && uns1_p && unsr_p) |
1280 | { |
1281 | op1 = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, |
1282 | OPTAB_LIB_WIDEN); |
1283 | /* As we've changed op1, we have to avoid using the value range |
1284 | for the original argument. */ |
1285 | arg1 = error_mark_node; |
1286 | do_xor = true; |
1287 | goto do_signed; |
1288 | } |
1289 | |
1290 | /* u1 - s2 -> ur */ |
1291 | if (code == MINUS_EXPR && uns0_p && !uns1_p && unsr_p) |
1292 | { |
1293 | op0 = expand_binop (mode, add_optab, op0, sgn, NULL_RTX, false, |
1294 | OPTAB_LIB_WIDEN); |
1295 | /* As we've changed op0, we have to avoid using the value range |
1296 | for the original argument. */ |
1297 | arg0 = error_mark_node; |
1298 | do_xor = true; |
1299 | goto do_signed; |
1300 | } |
1301 | |
1302 | /* s1 - u2 -> ur */ |
1303 | if (code == MINUS_EXPR && !uns0_p && uns1_p && unsr_p) |
1304 | { |
1305 | /* Compute the operation. On RTL level, the addition is always |
1306 | unsigned. */ |
1307 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1308 | OPTAB_LIB_WIDEN); |
1309 | int pos_neg = get_range_pos_neg (arg0, currently_expanding_gimple_stmt); |
1310 | if (pos_neg == 2) |
1311 | /* If ARG0 is known to be always negative, this is always overflow. */ |
1312 | emit_jump (do_error); |
1313 | else if (pos_neg == 3) |
1314 | /* If ARG0 is not known to be always positive, check at runtime. */ |
1315 | do_compare_rtx_and_jump (op0, const0_rtx, LT, false, mode, NULL_RTX, |
1316 | NULL, do_error, profile_probability::very_unlikely ()); |
1317 | do_compare_rtx_and_jump (op1, op0, LEU, true, mode, NULL_RTX, NULL, |
1318 | done_label, profile_probability::very_likely ()); |
1319 | goto do_error_label; |
1320 | } |
1321 | |
1322 | /* u1 - s2 -> sr */ |
1323 | if (code == MINUS_EXPR && uns0_p && !uns1_p && !unsr_p) |
1324 | { |
1325 | /* Compute the operation. On RTL level, the addition is always |
1326 | unsigned. */ |
1327 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1328 | OPTAB_LIB_WIDEN); |
1329 | rtx tem = expand_binop (mode, add_optab, op1, sgn, NULL_RTX, false, |
1330 | OPTAB_LIB_WIDEN); |
1331 | do_compare_rtx_and_jump (op0, tem, LTU, true, mode, NULL_RTX, NULL, |
1332 | done_label, profile_probability::very_likely ()); |
1333 | goto do_error_label; |
1334 | } |
1335 | |
1336 | /* u1 + u2 -> sr */ |
1337 | if (code == PLUS_EXPR && uns0_p && uns1_p && !unsr_p) |
1338 | { |
1339 | /* Compute the operation. On RTL level, the addition is always |
1340 | unsigned. */ |
1341 | res = expand_binop (mode, add_optab, op0, op1, NULL_RTX, false, |
1342 | OPTAB_LIB_WIDEN); |
1343 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
1344 | NULL, do_error, profile_probability::very_unlikely ()); |
1345 | rtx tem = op1; |
1346 | /* The operation is commutative, so we can pick operand to compare |
1347 | against. For prec <= BITS_PER_WORD, I think preferring REG operand |
1348 | is better over CONST_INT, because the CONST_INT might enlarge the |
1349 | instruction or CSE would need to figure out we'd already loaded it |
1350 | into a register before. For prec > BITS_PER_WORD, I think CONST_INT |
1351 | might be more beneficial, as then the multi-word comparison can be |
1352 | perhaps simplified. */ |
1353 | if (prec <= BITS_PER_WORD |
1354 | ? (CONST_SCALAR_INT_P (op1) && REG_P (op0)) |
1355 | : CONST_SCALAR_INT_P (op0)) |
1356 | tem = op0; |
1357 | do_compare_rtx_and_jump (res, tem, GEU, true, mode, NULL_RTX, NULL, |
1358 | done_label, profile_probability::very_likely ()); |
1359 | goto do_error_label; |
1360 | } |
1361 | |
1362 | /* s1 +- s2 -> ur */ |
1363 | if (!uns0_p && !uns1_p && unsr_p) |
1364 | { |
1365 | /* Compute the operation. On RTL level, the addition is always |
1366 | unsigned. */ |
1367 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1368 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1369 | int pos_neg = get_range_pos_neg (arg1, currently_expanding_gimple_stmt); |
1370 | if (code == PLUS_EXPR) |
1371 | { |
1372 | int pos_neg0 = get_range_pos_neg (arg0, |
1373 | currently_expanding_gimple_stmt); |
1374 | if (pos_neg0 != 3 && pos_neg == 3) |
1375 | { |
1376 | std::swap (a&: op0, b&: op1); |
1377 | pos_neg = pos_neg0; |
1378 | } |
1379 | } |
1380 | rtx tem; |
1381 | if (pos_neg != 3) |
1382 | { |
1383 | tem = expand_binop (mode, ((pos_neg == 1) ^ (code == MINUS_EXPR)) |
1384 | ? and_optab : ior_optab, |
1385 | op0, res, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1386 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL, |
1387 | NULL, done_label, profile_probability::very_likely ()); |
1388 | } |
1389 | else |
1390 | { |
1391 | rtx_code_label *do_ior_label = gen_label_rtx (); |
1392 | do_compare_rtx_and_jump (op1, const0_rtx, |
1393 | code == MINUS_EXPR ? GE : LT, false, mode, |
1394 | NULL_RTX, NULL, do_ior_label, |
1395 | profile_probability::even ()); |
1396 | tem = expand_binop (mode, and_optab, op0, res, NULL_RTX, false, |
1397 | OPTAB_LIB_WIDEN); |
1398 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1399 | NULL, done_label, profile_probability::very_likely ()); |
1400 | emit_jump (do_error); |
1401 | emit_label (do_ior_label); |
1402 | tem = expand_binop (mode, ior_optab, op0, res, NULL_RTX, false, |
1403 | OPTAB_LIB_WIDEN); |
1404 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1405 | NULL, done_label, profile_probability::very_likely ()); |
1406 | } |
1407 | goto do_error_label; |
1408 | } |
1409 | |
1410 | /* u1 - u2 -> sr */ |
1411 | if (code == MINUS_EXPR && uns0_p && uns1_p && !unsr_p) |
1412 | { |
1413 | /* Compute the operation. On RTL level, the addition is always |
1414 | unsigned. */ |
1415 | res = expand_binop (mode, sub_optab, op0, op1, NULL_RTX, false, |
1416 | OPTAB_LIB_WIDEN); |
1417 | rtx_code_label *op0_geu_op1 = gen_label_rtx (); |
1418 | do_compare_rtx_and_jump (op0, op1, GEU, true, mode, NULL_RTX, NULL, |
1419 | op0_geu_op1, profile_probability::even ()); |
1420 | do_compare_rtx_and_jump (res, const0_rtx, LT, false, mode, NULL_RTX, |
1421 | NULL, done_label, profile_probability::very_likely ()); |
1422 | emit_jump (do_error); |
1423 | emit_label (op0_geu_op1); |
1424 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
1425 | NULL, done_label, profile_probability::very_likely ()); |
1426 | goto do_error_label; |
1427 | } |
1428 | |
1429 | gcc_assert (!uns0_p && !uns1_p && !unsr_p); |
1430 | |
1431 | /* s1 +- s2 -> sr */ |
1432 | do_signed: |
1433 | { |
1434 | insn_code icode = optab_handler (op: code == PLUS_EXPR ? addv4_optab |
1435 | : subv4_optab, mode); |
1436 | if (icode != CODE_FOR_nothing) |
1437 | { |
1438 | class expand_operand ops[4]; |
1439 | rtx_insn *last = get_last_insn (); |
1440 | |
1441 | res = gen_reg_rtx (mode); |
1442 | create_output_operand (op: &ops[0], x: res, mode); |
1443 | create_input_operand (op: &ops[1], value: op0, mode); |
1444 | create_input_operand (op: &ops[2], value: op1, mode); |
1445 | create_fixed_operand (op: &ops[3], x: do_error); |
1446 | if (maybe_expand_insn (icode, nops: 4, ops)) |
1447 | { |
1448 | last = get_last_insn (); |
1449 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1450 | && JUMP_P (last) |
1451 | && any_condjump_p (last) |
1452 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1453 | add_reg_br_prob_note (last, |
1454 | profile_probability::very_unlikely ()); |
1455 | emit_jump (done_label); |
1456 | goto do_error_label; |
1457 | } |
1458 | |
1459 | delete_insns_since (last); |
1460 | } |
1461 | |
1462 | /* Compute the operation. On RTL level, the addition is always |
1463 | unsigned. */ |
1464 | res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab, |
1465 | op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN); |
1466 | |
1467 | /* If we can prove that one of the arguments (for MINUS_EXPR only |
1468 | the second operand, as subtraction is not commutative) is always |
1469 | non-negative or always negative, we can do just one comparison |
1470 | and conditional jump. */ |
1471 | int pos_neg = get_range_pos_neg (arg1, currently_expanding_gimple_stmt); |
1472 | if (code == PLUS_EXPR) |
1473 | { |
1474 | int pos_neg0 = get_range_pos_neg (arg0, |
1475 | currently_expanding_gimple_stmt); |
1476 | if (pos_neg0 != 3 && pos_neg == 3) |
1477 | { |
1478 | std::swap (a&: op0, b&: op1); |
1479 | pos_neg = pos_neg0; |
1480 | } |
1481 | } |
1482 | |
1483 | /* Addition overflows if and only if the two operands have the same sign, |
1484 | and the result has the opposite sign. Subtraction overflows if and |
1485 | only if the two operands have opposite sign, and the subtrahend has |
1486 | the same sign as the result. Here 0 is counted as positive. */ |
1487 | if (pos_neg == 3) |
1488 | { |
1489 | /* Compute op0 ^ op1 (operands have opposite sign). */ |
1490 | rtx op_xor = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1491 | OPTAB_LIB_WIDEN); |
1492 | |
1493 | /* Compute res ^ op1 (result and 2nd operand have opposite sign). */ |
1494 | rtx res_xor = expand_binop (mode, xor_optab, res, op1, NULL_RTX, false, |
1495 | OPTAB_LIB_WIDEN); |
1496 | |
1497 | rtx tem; |
1498 | if (code == PLUS_EXPR) |
1499 | { |
1500 | /* Compute (res ^ op1) & ~(op0 ^ op1). */ |
1501 | tem = expand_unop (mode, one_cmpl_optab, op_xor, NULL_RTX, false); |
1502 | tem = expand_binop (mode, and_optab, res_xor, tem, NULL_RTX, false, |
1503 | OPTAB_LIB_WIDEN); |
1504 | } |
1505 | else |
1506 | { |
1507 | /* Compute (op0 ^ op1) & ~(res ^ op1). */ |
1508 | tem = expand_unop (mode, one_cmpl_optab, res_xor, NULL_RTX, false); |
1509 | tem = expand_binop (mode, and_optab, op_xor, tem, NULL_RTX, false, |
1510 | OPTAB_LIB_WIDEN); |
1511 | } |
1512 | |
1513 | /* No overflow if the result has bit sign cleared. */ |
1514 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1515 | NULL, done_label, profile_probability::very_likely ()); |
1516 | } |
1517 | |
1518 | /* Compare the result of the operation with the first operand. |
1519 | No overflow for addition if second operand is positive and result |
1520 | is larger or second operand is negative and result is smaller. |
1521 | Likewise for subtraction with sign of second operand flipped. */ |
1522 | else |
1523 | do_compare_rtx_and_jump (res, op0, |
1524 | (pos_neg == 1) ^ (code == MINUS_EXPR) ? GE : LE, |
1525 | false, mode, NULL_RTX, NULL, done_label, |
1526 | profile_probability::very_likely ()); |
1527 | } |
1528 | |
1529 | do_error_label: |
1530 | emit_label (do_error); |
1531 | if (is_ubsan) |
1532 | { |
1533 | /* Expand the ubsan builtin call. */ |
1534 | push_temp_slots (); |
1535 | fn = ubsan_build_overflow_builtin (code, loc, TREE_TYPE (arg0), |
1536 | arg0, arg1, datap); |
1537 | expand_normal (exp: fn); |
1538 | pop_temp_slots (); |
1539 | do_pending_stack_adjust (); |
1540 | } |
1541 | else if (lhs) |
1542 | expand_arith_set_overflow (lhs, target); |
1543 | |
1544 | /* We're done. */ |
1545 | emit_label (done_label); |
1546 | |
1547 | if (lhs) |
1548 | { |
1549 | if (is_ubsan) |
1550 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
1551 | else |
1552 | { |
1553 | if (do_xor) |
1554 | res = expand_binop (mode, add_optab, res, sgn, NULL_RTX, false, |
1555 | OPTAB_LIB_WIDEN); |
1556 | |
1557 | expand_arith_overflow_result_store (lhs, target, mode, res); |
1558 | } |
1559 | } |
1560 | } |
1561 | |
1562 | /* Add negate overflow checking to the statement STMT. */ |
1563 | |
1564 | static void |
1565 | expand_neg_overflow (location_t loc, tree lhs, tree arg1, bool is_ubsan, |
1566 | tree *datap) |
1567 | { |
1568 | rtx res, op1; |
1569 | tree fn; |
1570 | rtx_code_label *done_label, *do_error; |
1571 | rtx target = NULL_RTX; |
1572 | |
1573 | done_label = gen_label_rtx (); |
1574 | do_error = gen_label_rtx (); |
1575 | |
1576 | do_pending_stack_adjust (); |
1577 | op1 = expand_normal (exp: arg1); |
1578 | |
1579 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg1)); |
1580 | if (lhs) |
1581 | { |
1582 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1583 | if (!is_ubsan) |
1584 | write_complex_part (target, const0_rtx, true, false); |
1585 | } |
1586 | |
1587 | enum insn_code icode = optab_handler (op: negv3_optab, mode); |
1588 | if (icode != CODE_FOR_nothing) |
1589 | { |
1590 | class expand_operand ops[3]; |
1591 | rtx_insn *last = get_last_insn (); |
1592 | |
1593 | res = gen_reg_rtx (mode); |
1594 | create_output_operand (op: &ops[0], x: res, mode); |
1595 | create_input_operand (op: &ops[1], value: op1, mode); |
1596 | create_fixed_operand (op: &ops[2], x: do_error); |
1597 | if (maybe_expand_insn (icode, nops: 3, ops)) |
1598 | { |
1599 | last = get_last_insn (); |
1600 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
1601 | && JUMP_P (last) |
1602 | && any_condjump_p (last) |
1603 | && !find_reg_note (last, REG_BR_PROB, 0)) |
1604 | add_reg_br_prob_note (last, |
1605 | profile_probability::very_unlikely ()); |
1606 | emit_jump (done_label); |
1607 | } |
1608 | else |
1609 | { |
1610 | delete_insns_since (last); |
1611 | icode = CODE_FOR_nothing; |
1612 | } |
1613 | } |
1614 | |
1615 | if (icode == CODE_FOR_nothing) |
1616 | { |
1617 | /* Compute the operation. On RTL level, the addition is always |
1618 | unsigned. */ |
1619 | res = expand_unop (mode, neg_optab, op1, NULL_RTX, false); |
1620 | |
1621 | /* Compare the operand with the most negative value. */ |
1622 | rtx minv = expand_normal (TYPE_MIN_VALUE (TREE_TYPE (arg1))); |
1623 | do_compare_rtx_and_jump (op1, minv, NE, true, mode, NULL_RTX, NULL, |
1624 | done_label, profile_probability::very_likely ()); |
1625 | } |
1626 | |
1627 | emit_label (do_error); |
1628 | if (is_ubsan) |
1629 | { |
1630 | /* Expand the ubsan builtin call. */ |
1631 | push_temp_slots (); |
1632 | fn = ubsan_build_overflow_builtin (NEGATE_EXPR, loc, TREE_TYPE (arg1), |
1633 | arg1, NULL_TREE, datap); |
1634 | expand_normal (exp: fn); |
1635 | pop_temp_slots (); |
1636 | do_pending_stack_adjust (); |
1637 | } |
1638 | else if (lhs) |
1639 | expand_arith_set_overflow (lhs, target); |
1640 | |
1641 | /* We're done. */ |
1642 | emit_label (done_label); |
1643 | |
1644 | if (lhs) |
1645 | { |
1646 | if (is_ubsan) |
1647 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
1648 | else |
1649 | expand_arith_overflow_result_store (lhs, target, mode, res); |
1650 | } |
1651 | } |
1652 | |
1653 | /* Return true if UNS WIDEN_MULT_EXPR with result mode WMODE and operand |
1654 | mode MODE can be expanded without using a libcall. */ |
1655 | |
1656 | static bool |
1657 | can_widen_mult_without_libcall (scalar_int_mode wmode, scalar_int_mode mode, |
1658 | rtx op0, rtx op1, bool uns) |
1659 | { |
1660 | if (find_widening_optab_handler (umul_widen_optab, wmode, mode) |
1661 | != CODE_FOR_nothing) |
1662 | return true; |
1663 | |
1664 | if (find_widening_optab_handler (smul_widen_optab, wmode, mode) |
1665 | != CODE_FOR_nothing) |
1666 | return true; |
1667 | |
1668 | rtx_insn *last = get_last_insn (); |
1669 | if (CONSTANT_P (op0)) |
1670 | op0 = convert_modes (mode: wmode, oldmode: mode, x: op0, unsignedp: uns); |
1671 | else |
1672 | op0 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 1); |
1673 | if (CONSTANT_P (op1)) |
1674 | op1 = convert_modes (mode: wmode, oldmode: mode, x: op1, unsignedp: uns); |
1675 | else |
1676 | op1 = gen_raw_REG (wmode, LAST_VIRTUAL_REGISTER + 2); |
1677 | rtx ret = expand_mult (wmode, op0, op1, NULL_RTX, uns, true); |
1678 | delete_insns_since (last); |
1679 | return ret != NULL_RTX; |
1680 | } |
1681 | |
1682 | /* Add mul overflow checking to the statement STMT. */ |
1683 | |
1684 | static void |
1685 | expand_mul_overflow (location_t loc, tree lhs, tree arg0, tree arg1, |
1686 | bool unsr_p, bool uns0_p, bool uns1_p, bool is_ubsan, |
1687 | tree *datap) |
1688 | { |
1689 | rtx res, op0, op1; |
1690 | tree fn, type; |
1691 | rtx_code_label *done_label, *do_error; |
1692 | rtx target = NULL_RTX; |
1693 | signop sign; |
1694 | enum insn_code icode; |
1695 | int save_flag_trapv = flag_trapv; |
1696 | |
1697 | /* We don't want any __mulv?i3 etc. calls from the expansion of |
1698 | these internal functions, so disable -ftrapv temporarily. */ |
1699 | flag_trapv = 0; |
1700 | done_label = gen_label_rtx (); |
1701 | do_error = gen_label_rtx (); |
1702 | |
1703 | do_pending_stack_adjust (); |
1704 | op0 = expand_normal (exp: arg0); |
1705 | op1 = expand_normal (exp: arg1); |
1706 | |
1707 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (arg0)); |
1708 | bool uns = unsr_p; |
1709 | if (lhs) |
1710 | { |
1711 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
1712 | if (!is_ubsan) |
1713 | write_complex_part (target, const0_rtx, true, false); |
1714 | } |
1715 | |
1716 | if (is_ubsan) |
1717 | gcc_assert (!unsr_p && !uns0_p && !uns1_p); |
1718 | |
1719 | /* We assume both operands and result have the same precision |
1720 | here (GET_MODE_BITSIZE (mode)), S stands for signed type |
1721 | with that precision, U for unsigned type with that precision, |
1722 | sgn for unsigned most significant bit in that precision. |
1723 | s1 is signed first operand, u1 is unsigned first operand, |
1724 | s2 is signed second operand, u2 is unsigned second operand, |
1725 | sr is signed result, ur is unsigned result and the following |
1726 | rules say how to compute result (which is always result of |
1727 | the operands as if both were unsigned, cast to the right |
1728 | signedness) and how to compute whether operation overflowed. |
1729 | main_ovf (false) stands for jump on signed multiplication |
1730 | overflow or the main algorithm with uns == false. |
1731 | main_ovf (true) stands for jump on unsigned multiplication |
1732 | overflow or the main algorithm with uns == true. |
1733 | |
1734 | s1 * s2 -> sr |
1735 | res = (S) ((U) s1 * (U) s2) |
1736 | ovf = main_ovf (false) |
1737 | u1 * u2 -> ur |
1738 | res = u1 * u2 |
1739 | ovf = main_ovf (true) |
1740 | s1 * u2 -> ur |
1741 | res = (U) s1 * u2 |
1742 | ovf = (s1 < 0 && u2) || main_ovf (true) |
1743 | u1 * u2 -> sr |
1744 | res = (S) (u1 * u2) |
1745 | ovf = res < 0 || main_ovf (true) |
1746 | s1 * u2 -> sr |
1747 | res = (S) ((U) s1 * u2) |
1748 | ovf = (S) u2 >= 0 ? main_ovf (false) |
1749 | : (s1 != 0 && (s1 != -1 || u2 != (U) res)) |
1750 | s1 * s2 -> ur |
1751 | t1 = (s1 & s2) < 0 ? (-(U) s1) : ((U) s1) |
1752 | t2 = (s1 & s2) < 0 ? (-(U) s2) : ((U) s2) |
1753 | res = t1 * t2 |
1754 | ovf = (s1 ^ s2) < 0 ? (s1 && s2) : main_ovf (true) */ |
1755 | |
1756 | if (uns0_p && !uns1_p) |
1757 | { |
1758 | /* Multiplication is commutative, if operand signedness differs, |
1759 | canonicalize to the first operand being signed and second |
1760 | unsigned to simplify following code. */ |
1761 | std::swap (a&: op0, b&: op1); |
1762 | std::swap (a&: arg0, b&: arg1); |
1763 | uns0_p = false; |
1764 | uns1_p = true; |
1765 | } |
1766 | |
1767 | int pos_neg0 = get_range_pos_neg (arg0, currently_expanding_gimple_stmt); |
1768 | int pos_neg1 = get_range_pos_neg (arg1, currently_expanding_gimple_stmt); |
1769 | /* Unsigned types with smaller than mode precision, even if they have most |
1770 | significant bit set, are still zero-extended. */ |
1771 | if (uns0_p && TYPE_PRECISION (TREE_TYPE (arg0)) < GET_MODE_PRECISION (mode)) |
1772 | pos_neg0 = 1; |
1773 | if (uns1_p && TYPE_PRECISION (TREE_TYPE (arg1)) < GET_MODE_PRECISION (mode)) |
1774 | pos_neg1 = 1; |
1775 | |
1776 | /* s1 * u2 -> ur */ |
1777 | if (!uns0_p && uns1_p && unsr_p) |
1778 | { |
1779 | switch (pos_neg0) |
1780 | { |
1781 | case 1: |
1782 | /* If s1 is non-negative, just perform normal u1 * u2 -> ur. */ |
1783 | goto do_main; |
1784 | case 2: |
1785 | /* If s1 is negative, avoid the main code, just multiply and |
1786 | signal overflow if op1 is not 0. */ |
1787 | struct separate_ops ops; |
1788 | ops.code = MULT_EXPR; |
1789 | ops.type = TREE_TYPE (arg1); |
1790 | ops.op0 = make_tree (ops.type, op0); |
1791 | ops.op1 = make_tree (ops.type, op1); |
1792 | ops.op2 = NULL_TREE; |
1793 | ops.location = loc; |
1794 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1795 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
1796 | NULL, done_label, profile_probability::very_likely ()); |
1797 | goto do_error_label; |
1798 | case 3: |
1799 | if (get_min_precision (arg: arg1, sign: UNSIGNED) |
1800 | + get_min_precision (arg: arg0, sign: SIGNED) <= GET_MODE_PRECISION (mode)) |
1801 | { |
1802 | /* If the first operand is sign extended from narrower type, the |
1803 | second operand is zero extended from narrower type and |
1804 | the sum of the two precisions is smaller or equal to the |
1805 | result precision: if the first argument is at runtime |
1806 | non-negative, maximum result will be 0x7e81 or 0x7f..fe80..01 |
1807 | and there will be no overflow, if the first argument is |
1808 | negative and the second argument zero, the result will be |
1809 | 0 and there will be no overflow, if the first argument is |
1810 | negative and the second argument positive, the result when |
1811 | treated as signed will be negative (minimum -0x7f80 or |
1812 | -0x7f..f80..0) there will be always overflow. So, do |
1813 | res = (U) (s1 * u2) |
1814 | ovf = (S) res < 0 */ |
1815 | struct separate_ops ops; |
1816 | ops.code = MULT_EXPR; |
1817 | ops.type |
1818 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1819 | 1); |
1820 | ops.op0 = make_tree (ops.type, op0); |
1821 | ops.op1 = make_tree (ops.type, op1); |
1822 | ops.op2 = NULL_TREE; |
1823 | ops.location = loc; |
1824 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1825 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, |
1826 | mode, NULL_RTX, NULL, done_label, |
1827 | profile_probability::very_likely ()); |
1828 | goto do_error_label; |
1829 | } |
1830 | rtx_code_label *do_main_label; |
1831 | do_main_label = gen_label_rtx (); |
1832 | do_compare_rtx_and_jump (op0, const0_rtx, GE, false, mode, NULL_RTX, |
1833 | NULL, do_main_label, profile_probability::very_likely ()); |
1834 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
1835 | NULL, do_main_label, profile_probability::very_likely ()); |
1836 | expand_arith_set_overflow (lhs, target); |
1837 | emit_label (do_main_label); |
1838 | goto do_main; |
1839 | default: |
1840 | gcc_unreachable (); |
1841 | } |
1842 | } |
1843 | |
1844 | /* u1 * u2 -> sr */ |
1845 | if (uns0_p && uns1_p && !unsr_p) |
1846 | { |
1847 | if ((pos_neg0 | pos_neg1) == 1) |
1848 | { |
1849 | /* If both arguments are zero extended from narrower types, |
1850 | the MSB will be clear on both and so we can pretend it is |
1851 | a normal s1 * s2 -> sr multiplication. */ |
1852 | uns0_p = false; |
1853 | uns1_p = false; |
1854 | } |
1855 | else |
1856 | uns = true; |
1857 | /* Rest of handling of this case after res is computed. */ |
1858 | goto do_main; |
1859 | } |
1860 | |
1861 | /* s1 * u2 -> sr */ |
1862 | if (!uns0_p && uns1_p && !unsr_p) |
1863 | { |
1864 | switch (pos_neg1) |
1865 | { |
1866 | case 1: |
1867 | goto do_main; |
1868 | case 2: |
1869 | /* If (S) u2 is negative (i.e. u2 is larger than maximum of S, |
1870 | avoid the main code, just multiply and signal overflow |
1871 | unless 0 * u2 or -1 * ((U) Smin). */ |
1872 | struct separate_ops ops; |
1873 | ops.code = MULT_EXPR; |
1874 | ops.type = TREE_TYPE (arg1); |
1875 | ops.op0 = make_tree (ops.type, op0); |
1876 | ops.op1 = make_tree (ops.type, op1); |
1877 | ops.op2 = NULL_TREE; |
1878 | ops.location = loc; |
1879 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1880 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
1881 | NULL, done_label, profile_probability::very_likely ()); |
1882 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
1883 | NULL, do_error, profile_probability::very_unlikely ()); |
1884 | int prec; |
1885 | prec = GET_MODE_PRECISION (mode); |
1886 | rtx sgn; |
1887 | sgn = immed_wide_int_const (wi::min_value (prec, SIGNED), mode); |
1888 | do_compare_rtx_and_jump (op1, sgn, EQ, true, mode, NULL_RTX, |
1889 | NULL, done_label, profile_probability::very_likely ()); |
1890 | goto do_error_label; |
1891 | case 3: |
1892 | /* Rest of handling of this case after res is computed. */ |
1893 | goto do_main; |
1894 | default: |
1895 | gcc_unreachable (); |
1896 | } |
1897 | } |
1898 | |
1899 | /* s1 * s2 -> ur */ |
1900 | if (!uns0_p && !uns1_p && unsr_p) |
1901 | { |
1902 | rtx tem; |
1903 | switch (pos_neg0 | pos_neg1) |
1904 | { |
1905 | case 1: /* Both operands known to be non-negative. */ |
1906 | goto do_main; |
1907 | case 2: /* Both operands known to be negative. */ |
1908 | op0 = expand_unop (mode, neg_optab, op0, NULL_RTX, false); |
1909 | op1 = expand_unop (mode, neg_optab, op1, NULL_RTX, false); |
1910 | /* Avoid looking at arg0/arg1 ranges, as we've changed |
1911 | the arguments. */ |
1912 | arg0 = error_mark_node; |
1913 | arg1 = error_mark_node; |
1914 | goto do_main; |
1915 | case 3: |
1916 | if ((pos_neg0 ^ pos_neg1) == 3) |
1917 | { |
1918 | /* If one operand is known to be negative and the other |
1919 | non-negative, this overflows always, unless the non-negative |
1920 | one is 0. Just do normal multiply and set overflow |
1921 | unless one of the operands is 0. */ |
1922 | struct separate_ops ops; |
1923 | ops.code = MULT_EXPR; |
1924 | ops.type |
1925 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1926 | 1); |
1927 | ops.op0 = make_tree (ops.type, op0); |
1928 | ops.op1 = make_tree (ops.type, op1); |
1929 | ops.op2 = NULL_TREE; |
1930 | ops.location = loc; |
1931 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1932 | do_compare_rtx_and_jump (pos_neg0 == 1 ? op0 : op1, const0_rtx, EQ, |
1933 | true, mode, NULL_RTX, NULL, done_label, |
1934 | profile_probability::very_likely ()); |
1935 | goto do_error_label; |
1936 | } |
1937 | if (get_min_precision (arg: arg0, sign: SIGNED) |
1938 | + get_min_precision (arg: arg1, sign: SIGNED) <= GET_MODE_PRECISION (mode)) |
1939 | { |
1940 | /* If both operands are sign extended from narrower types and |
1941 | the sum of the two precisions is smaller or equal to the |
1942 | result precision: if both arguments are at runtime |
1943 | non-negative, maximum result will be 0x3f01 or 0x3f..f0..01 |
1944 | and there will be no overflow, if both arguments are negative, |
1945 | maximum result will be 0x40..00 and there will be no overflow |
1946 | either, if one argument is positive and the other argument |
1947 | negative, the result when treated as signed will be negative |
1948 | and there will be always overflow, and if one argument is |
1949 | zero and the other negative the result will be zero and no |
1950 | overflow. So, do |
1951 | res = (U) (s1 * s2) |
1952 | ovf = (S) res < 0 */ |
1953 | struct separate_ops ops; |
1954 | ops.code = MULT_EXPR; |
1955 | ops.type |
1956 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), |
1957 | 1); |
1958 | ops.op0 = make_tree (ops.type, op0); |
1959 | ops.op1 = make_tree (ops.type, op1); |
1960 | ops.op2 = NULL_TREE; |
1961 | ops.location = loc; |
1962 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
1963 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, |
1964 | mode, NULL_RTX, NULL, done_label, |
1965 | profile_probability::very_likely ()); |
1966 | goto do_error_label; |
1967 | } |
1968 | /* The general case, do all the needed comparisons at runtime. */ |
1969 | rtx_code_label *do_main_label, *after_negate_label; |
1970 | rtx rop0, rop1; |
1971 | rop0 = gen_reg_rtx (mode); |
1972 | rop1 = gen_reg_rtx (mode); |
1973 | emit_move_insn (rop0, op0); |
1974 | emit_move_insn (rop1, op1); |
1975 | op0 = rop0; |
1976 | op1 = rop1; |
1977 | do_main_label = gen_label_rtx (); |
1978 | after_negate_label = gen_label_rtx (); |
1979 | tem = expand_binop (mode, and_optab, op0, op1, NULL_RTX, false, |
1980 | OPTAB_LIB_WIDEN); |
1981 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1982 | NULL, after_negate_label, profile_probability::very_likely ()); |
1983 | /* Both arguments negative here, negate them and continue with |
1984 | normal unsigned overflow checking multiplication. */ |
1985 | emit_move_insn (op0, expand_unop (mode, neg_optab, op0, |
1986 | NULL_RTX, false)); |
1987 | emit_move_insn (op1, expand_unop (mode, neg_optab, op1, |
1988 | NULL_RTX, false)); |
1989 | /* Avoid looking at arg0/arg1 ranges, as we might have changed |
1990 | the arguments. */ |
1991 | arg0 = error_mark_node; |
1992 | arg1 = error_mark_node; |
1993 | emit_jump (do_main_label); |
1994 | emit_label (after_negate_label); |
1995 | tem = expand_binop (mode, xor_optab, op0, op1, NULL_RTX, false, |
1996 | OPTAB_LIB_WIDEN); |
1997 | do_compare_rtx_and_jump (tem, const0_rtx, GE, false, mode, NULL_RTX, |
1998 | NULL, do_main_label, |
1999 | profile_probability::very_likely ()); |
2000 | /* One argument is negative here, the other positive. This |
2001 | overflows always, unless one of the arguments is 0. But |
2002 | if e.g. s2 is 0, (U) s1 * 0 doesn't overflow, whatever s1 |
2003 | is, thus we can keep do_main code oring in overflow as is. */ |
2004 | if (pos_neg0 != 2) |
2005 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
2006 | NULL, do_main_label, |
2007 | profile_probability::very_unlikely ()); |
2008 | if (pos_neg1 != 2) |
2009 | do_compare_rtx_and_jump (op1, const0_rtx, EQ, true, mode, NULL_RTX, |
2010 | NULL, do_main_label, |
2011 | profile_probability::very_unlikely ()); |
2012 | expand_arith_set_overflow (lhs, target); |
2013 | emit_label (do_main_label); |
2014 | goto do_main; |
2015 | default: |
2016 | gcc_unreachable (); |
2017 | } |
2018 | } |
2019 | |
2020 | do_main: |
2021 | type = build_nonstandard_integer_type (GET_MODE_PRECISION (mode), uns); |
2022 | sign = uns ? UNSIGNED : SIGNED; |
2023 | icode = optab_handler (op: uns ? umulv4_optab : mulv4_optab, mode); |
2024 | if (uns |
2025 | && (integer_pow2p (arg0) || integer_pow2p (arg1)) |
2026 | && (optimize_insn_for_speed_p () || icode == CODE_FOR_nothing)) |
2027 | { |
2028 | /* Optimize unsigned multiplication by power of 2 constant |
2029 | using 2 shifts, one for result, one to extract the shifted |
2030 | out bits to see if they are all zero. |
2031 | Don't do this if optimizing for size and we have umulv4_optab, |
2032 | in that case assume multiplication will be shorter. |
2033 | This is heuristics based on the single target that provides |
2034 | umulv4 right now (i?86/x86_64), if further targets add it, this |
2035 | might need to be revisited. |
2036 | Cases where both operands are constant should be folded already |
2037 | during GIMPLE, and cases where one operand is constant but not |
2038 | power of 2 are questionable, either the WIDEN_MULT_EXPR case |
2039 | below can be done without multiplication, just by shifts and adds, |
2040 | or we'd need to divide the result (and hope it actually doesn't |
2041 | really divide nor multiply) and compare the result of the division |
2042 | with the original operand. */ |
2043 | rtx opn0 = op0; |
2044 | rtx opn1 = op1; |
2045 | tree argn0 = arg0; |
2046 | tree argn1 = arg1; |
2047 | if (integer_pow2p (arg0)) |
2048 | { |
2049 | std::swap (a&: opn0, b&: opn1); |
2050 | std::swap (a&: argn0, b&: argn1); |
2051 | } |
2052 | int cnt = tree_log2 (argn1); |
2053 | if (cnt >= 0 && cnt < GET_MODE_PRECISION (mode)) |
2054 | { |
2055 | rtx upper = const0_rtx; |
2056 | res = expand_shift (LSHIFT_EXPR, mode, opn0, cnt, NULL_RTX, uns); |
2057 | if (cnt != 0) |
2058 | upper = expand_shift (RSHIFT_EXPR, mode, opn0, |
2059 | GET_MODE_PRECISION (mode) - cnt, |
2060 | NULL_RTX, uns); |
2061 | do_compare_rtx_and_jump (upper, const0_rtx, EQ, true, mode, |
2062 | NULL_RTX, NULL, done_label, |
2063 | profile_probability::very_likely ()); |
2064 | goto do_error_label; |
2065 | } |
2066 | } |
2067 | if (icode != CODE_FOR_nothing) |
2068 | { |
2069 | class expand_operand ops[4]; |
2070 | rtx_insn *last = get_last_insn (); |
2071 | |
2072 | res = gen_reg_rtx (mode); |
2073 | create_output_operand (op: &ops[0], x: res, mode); |
2074 | create_input_operand (op: &ops[1], value: op0, mode); |
2075 | create_input_operand (op: &ops[2], value: op1, mode); |
2076 | create_fixed_operand (op: &ops[3], x: do_error); |
2077 | if (maybe_expand_insn (icode, nops: 4, ops)) |
2078 | { |
2079 | last = get_last_insn (); |
2080 | if (profile_status_for_fn (cfun) != PROFILE_ABSENT |
2081 | && JUMP_P (last) |
2082 | && any_condjump_p (last) |
2083 | && !find_reg_note (last, REG_BR_PROB, 0)) |
2084 | add_reg_br_prob_note (last, |
2085 | profile_probability::very_unlikely ()); |
2086 | emit_jump (done_label); |
2087 | } |
2088 | else |
2089 | { |
2090 | delete_insns_since (last); |
2091 | icode = CODE_FOR_nothing; |
2092 | } |
2093 | } |
2094 | |
2095 | if (icode == CODE_FOR_nothing) |
2096 | { |
2097 | struct separate_ops ops; |
2098 | int prec = GET_MODE_PRECISION (mode); |
2099 | scalar_int_mode hmode, wmode; |
2100 | ops.op0 = make_tree (type, op0); |
2101 | ops.op1 = make_tree (type, op1); |
2102 | ops.op2 = NULL_TREE; |
2103 | ops.location = loc; |
2104 | |
2105 | /* Optimize unsigned overflow check where we don't use the |
2106 | multiplication result, just whether overflow happened. |
2107 | If we can do MULT_HIGHPART_EXPR, that followed by |
2108 | comparison of the result against zero is cheapest. |
2109 | We'll still compute res, but it should be DCEd later. */ |
2110 | use_operand_p use; |
2111 | gimple *use_stmt; |
2112 | if (!is_ubsan |
2113 | && lhs |
2114 | && uns |
2115 | && !(uns0_p && uns1_p && !unsr_p) |
2116 | && can_mult_highpart_p (mode, uns) == 1 |
2117 | && single_imm_use (var: lhs, use_p: &use, stmt: &use_stmt) |
2118 | && is_gimple_assign (gs: use_stmt) |
2119 | && gimple_assign_rhs_code (gs: use_stmt) == IMAGPART_EXPR) |
2120 | goto highpart; |
2121 | |
2122 | if (GET_MODE_2XWIDER_MODE (m: mode).exists (mode: &wmode) |
2123 | && targetm.scalar_mode_supported_p (wmode) |
2124 | && can_widen_mult_without_libcall (wmode, mode, op0, op1, uns)) |
2125 | { |
2126 | twoxwider: |
2127 | ops.code = WIDEN_MULT_EXPR; |
2128 | ops.type |
2129 | = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: wmode), uns); |
2130 | |
2131 | res = expand_expr_real_2 (&ops, NULL_RTX, wmode, EXPAND_NORMAL); |
2132 | rtx hipart = expand_shift (RSHIFT_EXPR, wmode, res, prec, |
2133 | NULL_RTX, uns); |
2134 | hipart = convert_modes (mode, oldmode: wmode, x: hipart, unsignedp: uns); |
2135 | res = convert_modes (mode, oldmode: wmode, x: res, unsignedp: uns); |
2136 | if (uns) |
2137 | /* For the unsigned multiplication, there was overflow if |
2138 | HIPART is non-zero. */ |
2139 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
2140 | NULL_RTX, NULL, done_label, |
2141 | profile_probability::very_likely ()); |
2142 | else |
2143 | { |
2144 | /* RES is used more than once, place it in a pseudo. */ |
2145 | res = force_reg (mode, res); |
2146 | |
2147 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
2148 | NULL_RTX, 0); |
2149 | /* RES is low half of the double width result, HIPART |
2150 | the high half. There was overflow if |
2151 | HIPART is different from RES < 0 ? -1 : 0. */ |
2152 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
2153 | NULL_RTX, NULL, done_label, |
2154 | profile_probability::very_likely ()); |
2155 | } |
2156 | } |
2157 | else if (can_mult_highpart_p (mode, uns) == 1) |
2158 | { |
2159 | highpart: |
2160 | ops.code = MULT_HIGHPART_EXPR; |
2161 | ops.type = type; |
2162 | |
2163 | rtx hipart = expand_expr_real_2 (&ops, NULL_RTX, mode, |
2164 | EXPAND_NORMAL); |
2165 | ops.code = MULT_EXPR; |
2166 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2167 | if (uns) |
2168 | /* For the unsigned multiplication, there was overflow if |
2169 | HIPART is non-zero. */ |
2170 | do_compare_rtx_and_jump (hipart, const0_rtx, EQ, true, mode, |
2171 | NULL_RTX, NULL, done_label, |
2172 | profile_probability::very_likely ()); |
2173 | else |
2174 | { |
2175 | rtx signbit = expand_shift (RSHIFT_EXPR, mode, res, prec - 1, |
2176 | NULL_RTX, 0); |
2177 | /* RES is low half of the double width result, HIPART |
2178 | the high half. There was overflow if |
2179 | HIPART is different from RES < 0 ? -1 : 0. */ |
2180 | do_compare_rtx_and_jump (signbit, hipart, EQ, true, mode, |
2181 | NULL_RTX, NULL, done_label, |
2182 | profile_probability::very_likely ()); |
2183 | } |
2184 | |
2185 | } |
2186 | else if (int_mode_for_size (size: prec / 2, limit: 1).exists (mode: &hmode) |
2187 | && 2 * GET_MODE_PRECISION (mode: hmode) == prec) |
2188 | { |
2189 | rtx_code_label *large_op0 = gen_label_rtx (); |
2190 | rtx_code_label *small_op0_large_op1 = gen_label_rtx (); |
2191 | rtx_code_label *one_small_one_large = gen_label_rtx (); |
2192 | rtx_code_label *both_ops_large = gen_label_rtx (); |
2193 | rtx_code_label *after_hipart_neg = uns ? NULL : gen_label_rtx (); |
2194 | rtx_code_label *after_lopart_neg = uns ? NULL : gen_label_rtx (); |
2195 | rtx_code_label *do_overflow = gen_label_rtx (); |
2196 | rtx_code_label *hipart_different = uns ? NULL : gen_label_rtx (); |
2197 | |
2198 | unsigned int hprec = GET_MODE_PRECISION (mode: hmode); |
2199 | rtx hipart0 = expand_shift (RSHIFT_EXPR, mode, op0, hprec, |
2200 | NULL_RTX, uns); |
2201 | hipart0 = convert_modes (mode: hmode, oldmode: mode, x: hipart0, unsignedp: uns); |
2202 | rtx lopart0 = convert_modes (mode: hmode, oldmode: mode, x: op0, unsignedp: uns); |
2203 | rtx signbit0 = const0_rtx; |
2204 | if (!uns) |
2205 | signbit0 = expand_shift (RSHIFT_EXPR, hmode, lopart0, hprec - 1, |
2206 | NULL_RTX, 0); |
2207 | rtx hipart1 = expand_shift (RSHIFT_EXPR, mode, op1, hprec, |
2208 | NULL_RTX, uns); |
2209 | hipart1 = convert_modes (mode: hmode, oldmode: mode, x: hipart1, unsignedp: uns); |
2210 | rtx lopart1 = convert_modes (mode: hmode, oldmode: mode, x: op1, unsignedp: uns); |
2211 | rtx signbit1 = const0_rtx; |
2212 | if (!uns) |
2213 | signbit1 = expand_shift (RSHIFT_EXPR, hmode, lopart1, hprec - 1, |
2214 | NULL_RTX, 0); |
2215 | |
2216 | res = gen_reg_rtx (mode); |
2217 | |
2218 | /* True if op0 resp. op1 are known to be in the range of |
2219 | halfstype. */ |
2220 | bool op0_small_p = false; |
2221 | bool op1_small_p = false; |
2222 | /* True if op0 resp. op1 are known to have all zeros or all ones |
2223 | in the upper half of bits, but are not known to be |
2224 | op{0,1}_small_p. */ |
2225 | bool op0_medium_p = false; |
2226 | bool op1_medium_p = false; |
2227 | /* -1 if op{0,1} is known to be negative, 0 if it is known to be |
2228 | nonnegative, 1 if unknown. */ |
2229 | int op0_sign = 1; |
2230 | int op1_sign = 1; |
2231 | |
2232 | if (pos_neg0 == 1) |
2233 | op0_sign = 0; |
2234 | else if (pos_neg0 == 2) |
2235 | op0_sign = -1; |
2236 | if (pos_neg1 == 1) |
2237 | op1_sign = 0; |
2238 | else if (pos_neg1 == 2) |
2239 | op1_sign = -1; |
2240 | |
2241 | unsigned int mprec0 = prec; |
2242 | if (arg0 != error_mark_node) |
2243 | mprec0 = get_min_precision (arg: arg0, sign); |
2244 | if (mprec0 <= hprec) |
2245 | op0_small_p = true; |
2246 | else if (!uns && mprec0 <= hprec + 1) |
2247 | op0_medium_p = true; |
2248 | unsigned int mprec1 = prec; |
2249 | if (arg1 != error_mark_node) |
2250 | mprec1 = get_min_precision (arg: arg1, sign); |
2251 | if (mprec1 <= hprec) |
2252 | op1_small_p = true; |
2253 | else if (!uns && mprec1 <= hprec + 1) |
2254 | op1_medium_p = true; |
2255 | |
2256 | int smaller_sign = 1; |
2257 | int larger_sign = 1; |
2258 | if (op0_small_p) |
2259 | { |
2260 | smaller_sign = op0_sign; |
2261 | larger_sign = op1_sign; |
2262 | } |
2263 | else if (op1_small_p) |
2264 | { |
2265 | smaller_sign = op1_sign; |
2266 | larger_sign = op0_sign; |
2267 | } |
2268 | else if (op0_sign == op1_sign) |
2269 | { |
2270 | smaller_sign = op0_sign; |
2271 | larger_sign = op0_sign; |
2272 | } |
2273 | |
2274 | if (!op0_small_p) |
2275 | do_compare_rtx_and_jump (signbit0, hipart0, NE, true, hmode, |
2276 | NULL_RTX, NULL, large_op0, |
2277 | profile_probability::unlikely ()); |
2278 | |
2279 | if (!op1_small_p) |
2280 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
2281 | NULL_RTX, NULL, small_op0_large_op1, |
2282 | profile_probability::unlikely ()); |
2283 | |
2284 | /* If both op0 and op1 are sign (!uns) or zero (uns) extended from |
2285 | hmode to mode, the multiplication will never overflow. We can |
2286 | do just one hmode x hmode => mode widening multiplication. */ |
2287 | tree halfstype = build_nonstandard_integer_type (hprec, uns); |
2288 | ops.op0 = make_tree (halfstype, lopart0); |
2289 | ops.op1 = make_tree (halfstype, lopart1); |
2290 | ops.code = WIDEN_MULT_EXPR; |
2291 | ops.type = type; |
2292 | rtx thisres |
2293 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2294 | emit_move_insn (res, thisres); |
2295 | emit_jump (done_label); |
2296 | |
2297 | emit_label (small_op0_large_op1); |
2298 | |
2299 | /* If op0 is sign (!uns) or zero (uns) extended from hmode to mode, |
2300 | but op1 is not, just swap the arguments and handle it as op1 |
2301 | sign/zero extended, op0 not. */ |
2302 | rtx larger = gen_reg_rtx (mode); |
2303 | rtx hipart = gen_reg_rtx (hmode); |
2304 | rtx lopart = gen_reg_rtx (hmode); |
2305 | emit_move_insn (larger, op1); |
2306 | emit_move_insn (hipart, hipart1); |
2307 | emit_move_insn (lopart, lopart0); |
2308 | emit_jump (one_small_one_large); |
2309 | |
2310 | emit_label (large_op0); |
2311 | |
2312 | if (!op1_small_p) |
2313 | do_compare_rtx_and_jump (signbit1, hipart1, NE, true, hmode, |
2314 | NULL_RTX, NULL, both_ops_large, |
2315 | profile_probability::unlikely ()); |
2316 | |
2317 | /* If op1 is sign (!uns) or zero (uns) extended from hmode to mode, |
2318 | but op0 is not, prepare larger, hipart and lopart pseudos and |
2319 | handle it together with small_op0_large_op1. */ |
2320 | emit_move_insn (larger, op0); |
2321 | emit_move_insn (hipart, hipart0); |
2322 | emit_move_insn (lopart, lopart1); |
2323 | |
2324 | emit_label (one_small_one_large); |
2325 | |
2326 | /* lopart is the low part of the operand that is sign extended |
2327 | to mode, larger is the other operand, hipart is the |
2328 | high part of larger and lopart0 and lopart1 are the low parts |
2329 | of both operands. |
2330 | We perform lopart0 * lopart1 and lopart * hipart widening |
2331 | multiplications. */ |
2332 | tree halfutype = build_nonstandard_integer_type (hprec, 1); |
2333 | ops.op0 = make_tree (halfutype, lopart0); |
2334 | ops.op1 = make_tree (halfutype, lopart1); |
2335 | rtx lo0xlo1 |
2336 | = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2337 | |
2338 | ops.op0 = make_tree (halfutype, lopart); |
2339 | ops.op1 = make_tree (halfutype, hipart); |
2340 | rtx loxhi = gen_reg_rtx (mode); |
2341 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2342 | emit_move_insn (loxhi, tem); |
2343 | |
2344 | if (!uns) |
2345 | { |
2346 | /* if (hipart < 0) loxhi -= lopart << (bitsize / 2); */ |
2347 | if (larger_sign == 0) |
2348 | emit_jump (after_hipart_neg); |
2349 | else if (larger_sign != -1) |
2350 | do_compare_rtx_and_jump (hipart, const0_rtx, GE, false, hmode, |
2351 | NULL_RTX, NULL, after_hipart_neg, |
2352 | profile_probability::even ()); |
2353 | |
2354 | tem = convert_modes (mode, oldmode: hmode, x: lopart, unsignedp: 1); |
2355 | tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1); |
2356 | tem = expand_simple_binop (mode, MINUS, loxhi, tem, NULL_RTX, |
2357 | 1, OPTAB_WIDEN); |
2358 | emit_move_insn (loxhi, tem); |
2359 | |
2360 | emit_label (after_hipart_neg); |
2361 | |
2362 | /* if (lopart < 0) loxhi -= larger; */ |
2363 | if (smaller_sign == 0) |
2364 | emit_jump (after_lopart_neg); |
2365 | else if (smaller_sign != -1) |
2366 | do_compare_rtx_and_jump (lopart, const0_rtx, GE, false, hmode, |
2367 | NULL_RTX, NULL, after_lopart_neg, |
2368 | profile_probability::even ()); |
2369 | |
2370 | tem = expand_simple_binop (mode, MINUS, loxhi, larger, NULL_RTX, |
2371 | 1, OPTAB_WIDEN); |
2372 | emit_move_insn (loxhi, tem); |
2373 | |
2374 | emit_label (after_lopart_neg); |
2375 | } |
2376 | |
2377 | /* loxhi += (uns) lo0xlo1 >> (bitsize / 2); */ |
2378 | tem = expand_shift (RSHIFT_EXPR, mode, lo0xlo1, hprec, NULL_RTX, 1); |
2379 | tem = expand_simple_binop (mode, PLUS, loxhi, tem, NULL_RTX, |
2380 | 1, OPTAB_WIDEN); |
2381 | emit_move_insn (loxhi, tem); |
2382 | |
2383 | /* if (loxhi >> (bitsize / 2) |
2384 | == (hmode) loxhi >> (bitsize / 2 - 1)) (if !uns) |
2385 | if (loxhi >> (bitsize / 2) == 0 (if uns). */ |
2386 | rtx hipartloxhi = expand_shift (RSHIFT_EXPR, mode, loxhi, hprec, |
2387 | NULL_RTX, 0); |
2388 | hipartloxhi = convert_modes (mode: hmode, oldmode: mode, x: hipartloxhi, unsignedp: 0); |
2389 | rtx signbitloxhi = const0_rtx; |
2390 | if (!uns) |
2391 | signbitloxhi = expand_shift (RSHIFT_EXPR, hmode, |
2392 | convert_modes (mode: hmode, oldmode: mode, |
2393 | x: loxhi, unsignedp: 0), |
2394 | hprec - 1, NULL_RTX, 0); |
2395 | |
2396 | do_compare_rtx_and_jump (signbitloxhi, hipartloxhi, NE, true, hmode, |
2397 | NULL_RTX, NULL, do_overflow, |
2398 | profile_probability::very_unlikely ()); |
2399 | |
2400 | /* res = (loxhi << (bitsize / 2)) | (hmode) lo0xlo1; */ |
2401 | rtx loxhishifted = expand_shift (LSHIFT_EXPR, mode, loxhi, hprec, |
2402 | NULL_RTX, 1); |
2403 | tem = convert_modes (mode, oldmode: hmode, |
2404 | x: convert_modes (mode: hmode, oldmode: mode, x: lo0xlo1, unsignedp: 1), unsignedp: 1); |
2405 | |
2406 | tem = expand_simple_binop (mode, IOR, loxhishifted, tem, res, |
2407 | 1, OPTAB_WIDEN); |
2408 | if (tem != res) |
2409 | emit_move_insn (res, tem); |
2410 | emit_jump (done_label); |
2411 | |
2412 | emit_label (both_ops_large); |
2413 | |
2414 | /* If both operands are large (not sign (!uns) or zero (uns) |
2415 | extended from hmode), then perform the full multiplication |
2416 | which will be the result of the operation. |
2417 | The only cases which don't overflow are for signed multiplication |
2418 | some cases where both hipart0 and highpart1 are 0 or -1. |
2419 | For unsigned multiplication when high parts are both non-zero |
2420 | this overflows always. */ |
2421 | ops.code = MULT_EXPR; |
2422 | ops.op0 = make_tree (type, op0); |
2423 | ops.op1 = make_tree (type, op1); |
2424 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2425 | emit_move_insn (res, tem); |
2426 | |
2427 | if (!uns) |
2428 | { |
2429 | if (!op0_medium_p) |
2430 | { |
2431 | tem = expand_simple_binop (hmode, PLUS, hipart0, const1_rtx, |
2432 | NULL_RTX, 1, OPTAB_WIDEN); |
2433 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
2434 | NULL_RTX, NULL, do_error, |
2435 | profile_probability::very_unlikely ()); |
2436 | } |
2437 | |
2438 | if (!op1_medium_p) |
2439 | { |
2440 | tem = expand_simple_binop (hmode, PLUS, hipart1, const1_rtx, |
2441 | NULL_RTX, 1, OPTAB_WIDEN); |
2442 | do_compare_rtx_and_jump (tem, const1_rtx, GTU, true, hmode, |
2443 | NULL_RTX, NULL, do_error, |
2444 | profile_probability::very_unlikely ()); |
2445 | } |
2446 | |
2447 | /* At this point hipart{0,1} are both in [-1, 0]. If they are |
2448 | the same, overflow happened if res is non-positive, if they |
2449 | are different, overflow happened if res is positive. */ |
2450 | if (op0_sign != 1 && op1_sign != 1 && op0_sign != op1_sign) |
2451 | emit_jump (hipart_different); |
2452 | else if (op0_sign == 1 || op1_sign == 1) |
2453 | do_compare_rtx_and_jump (hipart0, hipart1, NE, true, hmode, |
2454 | NULL_RTX, NULL, hipart_different, |
2455 | profile_probability::even ()); |
2456 | |
2457 | do_compare_rtx_and_jump (res, const0_rtx, LE, false, mode, |
2458 | NULL_RTX, NULL, do_error, |
2459 | profile_probability::very_unlikely ()); |
2460 | emit_jump (done_label); |
2461 | |
2462 | emit_label (hipart_different); |
2463 | |
2464 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, |
2465 | NULL_RTX, NULL, do_error, |
2466 | profile_probability::very_unlikely ()); |
2467 | emit_jump (done_label); |
2468 | } |
2469 | |
2470 | emit_label (do_overflow); |
2471 | |
2472 | /* Overflow, do full multiplication and fallthru into do_error. */ |
2473 | ops.op0 = make_tree (type, op0); |
2474 | ops.op1 = make_tree (type, op1); |
2475 | tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2476 | emit_move_insn (res, tem); |
2477 | } |
2478 | else if (GET_MODE_2XWIDER_MODE (m: mode).exists (mode: &wmode) |
2479 | && targetm.scalar_mode_supported_p (wmode)) |
2480 | /* Even emitting a libcall is better than not detecting overflow |
2481 | at all. */ |
2482 | goto twoxwider; |
2483 | else |
2484 | { |
2485 | gcc_assert (!is_ubsan); |
2486 | ops.code = MULT_EXPR; |
2487 | ops.type = type; |
2488 | res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2489 | emit_jump (done_label); |
2490 | } |
2491 | } |
2492 | |
2493 | do_error_label: |
2494 | emit_label (do_error); |
2495 | if (is_ubsan) |
2496 | { |
2497 | /* Expand the ubsan builtin call. */ |
2498 | push_temp_slots (); |
2499 | fn = ubsan_build_overflow_builtin (MULT_EXPR, loc, TREE_TYPE (arg0), |
2500 | arg0, arg1, datap); |
2501 | expand_normal (exp: fn); |
2502 | pop_temp_slots (); |
2503 | do_pending_stack_adjust (); |
2504 | } |
2505 | else if (lhs) |
2506 | expand_arith_set_overflow (lhs, target); |
2507 | |
2508 | /* We're done. */ |
2509 | emit_label (done_label); |
2510 | |
2511 | /* u1 * u2 -> sr */ |
2512 | if (uns0_p && uns1_p && !unsr_p) |
2513 | { |
2514 | rtx_code_label *all_done_label = gen_label_rtx (); |
2515 | do_compare_rtx_and_jump (res, const0_rtx, GE, false, mode, NULL_RTX, |
2516 | NULL, all_done_label, profile_probability::very_likely ()); |
2517 | expand_arith_set_overflow (lhs, target); |
2518 | emit_label (all_done_label); |
2519 | } |
2520 | |
2521 | /* s1 * u2 -> sr */ |
2522 | if (!uns0_p && uns1_p && !unsr_p && pos_neg1 == 3) |
2523 | { |
2524 | rtx_code_label *all_done_label = gen_label_rtx (); |
2525 | rtx_code_label *set_noovf = gen_label_rtx (); |
2526 | do_compare_rtx_and_jump (op1, const0_rtx, GE, false, mode, NULL_RTX, |
2527 | NULL, all_done_label, profile_probability::very_likely ()); |
2528 | expand_arith_set_overflow (lhs, target); |
2529 | do_compare_rtx_and_jump (op0, const0_rtx, EQ, true, mode, NULL_RTX, |
2530 | NULL, set_noovf, profile_probability::very_likely ()); |
2531 | do_compare_rtx_and_jump (op0, constm1_rtx, NE, true, mode, NULL_RTX, |
2532 | NULL, all_done_label, profile_probability::very_unlikely ()); |
2533 | do_compare_rtx_and_jump (op1, res, NE, true, mode, NULL_RTX, NULL, |
2534 | all_done_label, profile_probability::very_unlikely ()); |
2535 | emit_label (set_noovf); |
2536 | write_complex_part (target, const0_rtx, true, false); |
2537 | emit_label (all_done_label); |
2538 | } |
2539 | |
2540 | if (lhs) |
2541 | { |
2542 | if (is_ubsan) |
2543 | expand_ubsan_result_store (lhs, target, mode, res, do_error); |
2544 | else |
2545 | expand_arith_overflow_result_store (lhs, target, mode, res); |
2546 | } |
2547 | flag_trapv = save_flag_trapv; |
2548 | } |
2549 | |
2550 | /* Expand UBSAN_CHECK_* internal function if it has vector operands. */ |
2551 | |
2552 | static void |
2553 | expand_vector_ubsan_overflow (location_t loc, enum tree_code code, tree lhs, |
2554 | tree arg0, tree arg1) |
2555 | { |
2556 | poly_uint64 cnt = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)); |
2557 | rtx_code_label *loop_lab = NULL; |
2558 | rtx cntvar = NULL_RTX; |
2559 | tree cntv = NULL_TREE; |
2560 | tree eltype = TREE_TYPE (TREE_TYPE (arg0)); |
2561 | tree sz = TYPE_SIZE (eltype); |
2562 | tree data = NULL_TREE; |
2563 | tree resv = NULL_TREE; |
2564 | rtx lhsr = NULL_RTX; |
2565 | rtx resvr = NULL_RTX; |
2566 | unsigned HOST_WIDE_INT const_cnt = 0; |
2567 | bool use_loop_p = (!cnt.is_constant (const_value: &const_cnt) || const_cnt > 4); |
2568 | int save_flag_trapv = flag_trapv; |
2569 | |
2570 | /* We don't want any __mulv?i3 etc. calls from the expansion of |
2571 | these internal functions, so disable -ftrapv temporarily. */ |
2572 | flag_trapv = 0; |
2573 | if (lhs) |
2574 | { |
2575 | optab op; |
2576 | lhsr = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2577 | if (!VECTOR_MODE_P (GET_MODE (lhsr)) |
2578 | || (op = optab_for_tree_code (code, TREE_TYPE (arg0), |
2579 | optab_default)) == unknown_optab |
2580 | || (optab_handler (op, TYPE_MODE (TREE_TYPE (arg0))) |
2581 | == CODE_FOR_nothing)) |
2582 | { |
2583 | if (MEM_P (lhsr)) |
2584 | resv = make_tree (TREE_TYPE (lhs), lhsr); |
2585 | else |
2586 | { |
2587 | resvr = assign_temp (TREE_TYPE (lhs), 1, 1); |
2588 | resv = make_tree (TREE_TYPE (lhs), resvr); |
2589 | } |
2590 | } |
2591 | } |
2592 | if (use_loop_p) |
2593 | { |
2594 | do_pending_stack_adjust (); |
2595 | loop_lab = gen_label_rtx (); |
2596 | cntvar = gen_reg_rtx (TYPE_MODE (sizetype)); |
2597 | cntv = make_tree (sizetype, cntvar); |
2598 | emit_move_insn (cntvar, const0_rtx); |
2599 | emit_label (loop_lab); |
2600 | } |
2601 | if (TREE_CODE (arg0) != VECTOR_CST) |
2602 | { |
2603 | rtx arg0r = expand_normal (exp: arg0); |
2604 | arg0 = make_tree (TREE_TYPE (arg0), arg0r); |
2605 | } |
2606 | if (TREE_CODE (arg1) != VECTOR_CST) |
2607 | { |
2608 | rtx arg1r = expand_normal (exp: arg1); |
2609 | arg1 = make_tree (TREE_TYPE (arg1), arg1r); |
2610 | } |
2611 | for (unsigned int i = 0; i < (use_loop_p ? 1 : const_cnt); i++) |
2612 | { |
2613 | tree op0, op1, res = NULL_TREE; |
2614 | if (use_loop_p) |
2615 | { |
2616 | tree atype = build_array_type_nelts (eltype, cnt); |
2617 | op0 = uniform_vector_p (arg0); |
2618 | if (op0 == NULL_TREE) |
2619 | { |
2620 | op0 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg0); |
2621 | op0 = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: op0, arg1: cntv, |
2622 | NULL_TREE, NULL_TREE); |
2623 | } |
2624 | op1 = uniform_vector_p (arg1); |
2625 | if (op1 == NULL_TREE) |
2626 | { |
2627 | op1 = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, arg1); |
2628 | op1 = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: op1, arg1: cntv, |
2629 | NULL_TREE, NULL_TREE); |
2630 | } |
2631 | if (resv) |
2632 | { |
2633 | res = fold_build1_loc (loc, VIEW_CONVERT_EXPR, atype, resv); |
2634 | res = build4_loc (loc, code: ARRAY_REF, type: eltype, arg0: res, arg1: cntv, |
2635 | NULL_TREE, NULL_TREE); |
2636 | } |
2637 | } |
2638 | else |
2639 | { |
2640 | tree bitpos = bitsize_int (tree_to_uhwi (sz) * i); |
2641 | op0 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg0, sz, bitpos); |
2642 | op1 = fold_build3_loc (loc, BIT_FIELD_REF, eltype, arg1, sz, bitpos); |
2643 | if (resv) |
2644 | res = fold_build3_loc (loc, BIT_FIELD_REF, eltype, resv, sz, |
2645 | bitpos); |
2646 | } |
2647 | switch (code) |
2648 | { |
2649 | case PLUS_EXPR: |
2650 | expand_addsub_overflow (loc, code: PLUS_EXPR, lhs: res, arg0: op0, arg1: op1, |
2651 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, datap: &data); |
2652 | break; |
2653 | case MINUS_EXPR: |
2654 | if (use_loop_p ? integer_zerop (arg0) : integer_zerop (op0)) |
2655 | expand_neg_overflow (loc, lhs: res, arg1: op1, is_ubsan: true, datap: &data); |
2656 | else |
2657 | expand_addsub_overflow (loc, code: MINUS_EXPR, lhs: res, arg0: op0, arg1: op1, |
2658 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, datap: &data); |
2659 | break; |
2660 | case MULT_EXPR: |
2661 | expand_mul_overflow (loc, lhs: res, arg0: op0, arg1: op1, unsr_p: false, uns0_p: false, uns1_p: false, |
2662 | is_ubsan: true, datap: &data); |
2663 | break; |
2664 | default: |
2665 | gcc_unreachable (); |
2666 | } |
2667 | } |
2668 | if (use_loop_p) |
2669 | { |
2670 | struct separate_ops ops; |
2671 | ops.code = PLUS_EXPR; |
2672 | ops.type = TREE_TYPE (cntv); |
2673 | ops.op0 = cntv; |
2674 | ops.op1 = build_int_cst (TREE_TYPE (cntv), 1); |
2675 | ops.op2 = NULL_TREE; |
2676 | ops.location = loc; |
2677 | rtx ret = expand_expr_real_2 (&ops, cntvar, TYPE_MODE (sizetype), |
2678 | EXPAND_NORMAL); |
2679 | if (ret != cntvar) |
2680 | emit_move_insn (cntvar, ret); |
2681 | rtx cntrtx = gen_int_mode (cnt, TYPE_MODE (sizetype)); |
2682 | do_compare_rtx_and_jump (cntvar, cntrtx, NE, false, |
2683 | TYPE_MODE (sizetype), NULL_RTX, NULL, loop_lab, |
2684 | profile_probability::very_likely ()); |
2685 | } |
2686 | if (lhs && resv == NULL_TREE) |
2687 | { |
2688 | struct separate_ops ops; |
2689 | ops.code = code; |
2690 | ops.type = TREE_TYPE (arg0); |
2691 | ops.op0 = arg0; |
2692 | ops.op1 = arg1; |
2693 | ops.op2 = NULL_TREE; |
2694 | ops.location = loc; |
2695 | rtx ret = expand_expr_real_2 (&ops, lhsr, TYPE_MODE (TREE_TYPE (arg0)), |
2696 | EXPAND_NORMAL); |
2697 | if (ret != lhsr) |
2698 | emit_move_insn (lhsr, ret); |
2699 | } |
2700 | else if (resvr) |
2701 | emit_move_insn (lhsr, resvr); |
2702 | flag_trapv = save_flag_trapv; |
2703 | } |
2704 | |
2705 | /* Expand UBSAN_CHECK_ADD call STMT. */ |
2706 | |
2707 | static void |
2708 | expand_UBSAN_CHECK_ADD (internal_fn, gcall *stmt) |
2709 | { |
2710 | location_t loc = gimple_location (g: stmt); |
2711 | tree lhs = gimple_call_lhs (gs: stmt); |
2712 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2713 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2714 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2715 | expand_vector_ubsan_overflow (loc, code: PLUS_EXPR, lhs, arg0, arg1); |
2716 | else |
2717 | expand_addsub_overflow (loc, code: PLUS_EXPR, lhs, arg0, arg1, |
2718 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, NULL); |
2719 | } |
2720 | |
2721 | /* Expand UBSAN_CHECK_SUB call STMT. */ |
2722 | |
2723 | static void |
2724 | expand_UBSAN_CHECK_SUB (internal_fn, gcall *stmt) |
2725 | { |
2726 | location_t loc = gimple_location (g: stmt); |
2727 | tree lhs = gimple_call_lhs (gs: stmt); |
2728 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2729 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2730 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2731 | expand_vector_ubsan_overflow (loc, code: MINUS_EXPR, lhs, arg0, arg1); |
2732 | else if (integer_zerop (arg0)) |
2733 | expand_neg_overflow (loc, lhs, arg1, is_ubsan: true, NULL); |
2734 | else |
2735 | expand_addsub_overflow (loc, code: MINUS_EXPR, lhs, arg0, arg1, |
2736 | unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, NULL); |
2737 | } |
2738 | |
2739 | /* Expand UBSAN_CHECK_MUL call STMT. */ |
2740 | |
2741 | static void |
2742 | expand_UBSAN_CHECK_MUL (internal_fn, gcall *stmt) |
2743 | { |
2744 | location_t loc = gimple_location (g: stmt); |
2745 | tree lhs = gimple_call_lhs (gs: stmt); |
2746 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2747 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2748 | if (VECTOR_TYPE_P (TREE_TYPE (arg0))) |
2749 | expand_vector_ubsan_overflow (loc, code: MULT_EXPR, lhs, arg0, arg1); |
2750 | else |
2751 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p: false, uns0_p: false, uns1_p: false, is_ubsan: true, |
2752 | NULL); |
2753 | } |
2754 | |
2755 | /* Helper function for {ADD,SUB,MUL}_OVERFLOW call stmt expansion. */ |
2756 | |
2757 | static void |
2758 | expand_arith_overflow (enum tree_code code, gimple *stmt) |
2759 | { |
2760 | tree lhs = gimple_call_lhs (gs: stmt); |
2761 | if (lhs == NULL_TREE) |
2762 | return; |
2763 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
2764 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
2765 | tree type = TREE_TYPE (TREE_TYPE (lhs)); |
2766 | int uns0_p = TYPE_UNSIGNED (TREE_TYPE (arg0)); |
2767 | int uns1_p = TYPE_UNSIGNED (TREE_TYPE (arg1)); |
2768 | int unsr_p = TYPE_UNSIGNED (type); |
2769 | int prec0 = TYPE_PRECISION (TREE_TYPE (arg0)); |
2770 | int prec1 = TYPE_PRECISION (TREE_TYPE (arg1)); |
2771 | int precres = TYPE_PRECISION (type); |
2772 | location_t loc = gimple_location (g: stmt); |
2773 | if (!uns0_p && get_range_pos_neg (arg0, stmt) == 1) |
2774 | uns0_p = true; |
2775 | if (!uns1_p && get_range_pos_neg (arg1, stmt) == 1) |
2776 | uns1_p = true; |
2777 | int pr = get_min_precision (arg: arg0, sign: uns0_p ? UNSIGNED : SIGNED); |
2778 | prec0 = MIN (prec0, pr); |
2779 | pr = get_min_precision (arg: arg1, sign: uns1_p ? UNSIGNED : SIGNED); |
2780 | prec1 = MIN (prec1, pr); |
2781 | int save_flag_trapv = flag_trapv; |
2782 | |
2783 | /* We don't want any __mulv?i3 etc. calls from the expansion of |
2784 | these internal functions, so disable -ftrapv temporarily. */ |
2785 | flag_trapv = 0; |
2786 | /* If uns0_p && uns1_p, precop is minimum needed precision |
2787 | of unsigned type to hold the exact result, otherwise |
2788 | precop is minimum needed precision of signed type to |
2789 | hold the exact result. */ |
2790 | int precop; |
2791 | if (code == MULT_EXPR) |
2792 | precop = prec0 + prec1 + (uns0_p != uns1_p); |
2793 | else |
2794 | { |
2795 | if (uns0_p == uns1_p) |
2796 | precop = MAX (prec0, prec1) + 1; |
2797 | else if (uns0_p) |
2798 | precop = MAX (prec0 + 1, prec1) + 1; |
2799 | else |
2800 | precop = MAX (prec0, prec1 + 1) + 1; |
2801 | } |
2802 | int orig_precres = precres; |
2803 | |
2804 | do |
2805 | { |
2806 | if ((uns0_p && uns1_p) |
2807 | ? ((precop + !unsr_p) <= precres |
2808 | /* u1 - u2 -> ur can overflow, no matter what precision |
2809 | the result has. */ |
2810 | && (code != MINUS_EXPR || !unsr_p)) |
2811 | : (!unsr_p && precop <= precres)) |
2812 | { |
2813 | /* The infinity precision result will always fit into result. */ |
2814 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2815 | write_complex_part (target, const0_rtx, true, false); |
2816 | scalar_int_mode mode = SCALAR_INT_TYPE_MODE (type); |
2817 | struct separate_ops ops; |
2818 | ops.code = code; |
2819 | ops.type = type; |
2820 | ops.op0 = fold_convert_loc (loc, type, arg0); |
2821 | ops.op1 = fold_convert_loc (loc, type, arg1); |
2822 | ops.op2 = NULL_TREE; |
2823 | ops.location = loc; |
2824 | rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
2825 | expand_arith_overflow_result_store (lhs, target, mode, res: tem); |
2826 | flag_trapv = save_flag_trapv; |
2827 | return; |
2828 | } |
2829 | |
2830 | /* For operations with low precision, if target doesn't have them, start |
2831 | with precres widening right away, otherwise do it only if the most |
2832 | simple cases can't be used. */ |
2833 | const int min_precision = targetm.min_arithmetic_precision (); |
2834 | if (orig_precres == precres && precres < min_precision) |
2835 | ; |
2836 | else if ((uns0_p && uns1_p && unsr_p && prec0 <= precres |
2837 | && prec1 <= precres) |
2838 | || ((!uns0_p || !uns1_p) && !unsr_p |
2839 | && prec0 + uns0_p <= precres |
2840 | && prec1 + uns1_p <= precres)) |
2841 | { |
2842 | arg0 = fold_convert_loc (loc, type, arg0); |
2843 | arg1 = fold_convert_loc (loc, type, arg1); |
2844 | switch (code) |
2845 | { |
2846 | case MINUS_EXPR: |
2847 | if (integer_zerop (arg0) && !unsr_p) |
2848 | { |
2849 | expand_neg_overflow (loc, lhs, arg1, is_ubsan: false, NULL); |
2850 | flag_trapv = save_flag_trapv; |
2851 | return; |
2852 | } |
2853 | /* FALLTHRU */ |
2854 | case PLUS_EXPR: |
2855 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2856 | uns0_p: unsr_p, uns1_p: unsr_p, is_ubsan: false, NULL); |
2857 | flag_trapv = save_flag_trapv; |
2858 | return; |
2859 | case MULT_EXPR: |
2860 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2861 | uns0_p: unsr_p, uns1_p: unsr_p, is_ubsan: false, NULL); |
2862 | flag_trapv = save_flag_trapv; |
2863 | return; |
2864 | default: |
2865 | gcc_unreachable (); |
2866 | } |
2867 | } |
2868 | |
2869 | /* For sub-word operations, retry with a wider type first. */ |
2870 | if (orig_precres == precres && precop <= BITS_PER_WORD) |
2871 | { |
2872 | int p = MAX (min_precision, precop); |
2873 | scalar_int_mode m = smallest_int_mode_for_size (size: p).require (); |
2874 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: m), |
2875 | uns0_p && uns1_p |
2876 | && unsr_p); |
2877 | p = TYPE_PRECISION (optype); |
2878 | if (p > precres) |
2879 | { |
2880 | precres = p; |
2881 | unsr_p = TYPE_UNSIGNED (optype); |
2882 | type = optype; |
2883 | continue; |
2884 | } |
2885 | } |
2886 | |
2887 | if (prec0 <= precres && prec1 <= precres) |
2888 | { |
2889 | tree types[2]; |
2890 | if (unsr_p) |
2891 | { |
2892 | types[0] = build_nonstandard_integer_type (precres, 0); |
2893 | types[1] = type; |
2894 | } |
2895 | else |
2896 | { |
2897 | types[0] = type; |
2898 | types[1] = build_nonstandard_integer_type (precres, 1); |
2899 | } |
2900 | arg0 = fold_convert_loc (loc, types[uns0_p], arg0); |
2901 | arg1 = fold_convert_loc (loc, types[uns1_p], arg1); |
2902 | if (code != MULT_EXPR) |
2903 | expand_addsub_overflow (loc, code, lhs, arg0, arg1, unsr_p, |
2904 | uns0_p, uns1_p, is_ubsan: false, NULL); |
2905 | else |
2906 | expand_mul_overflow (loc, lhs, arg0, arg1, unsr_p, |
2907 | uns0_p, uns1_p, is_ubsan: false, NULL); |
2908 | flag_trapv = save_flag_trapv; |
2909 | return; |
2910 | } |
2911 | |
2912 | /* Retry with a wider type. */ |
2913 | if (orig_precres == precres) |
2914 | { |
2915 | int p = MAX (prec0, prec1); |
2916 | scalar_int_mode m = smallest_int_mode_for_size (size: p).require (); |
2917 | tree optype = build_nonstandard_integer_type (GET_MODE_PRECISION (mode: m), |
2918 | uns0_p && uns1_p |
2919 | && unsr_p); |
2920 | p = TYPE_PRECISION (optype); |
2921 | if (p > precres) |
2922 | { |
2923 | precres = p; |
2924 | unsr_p = TYPE_UNSIGNED (optype); |
2925 | type = optype; |
2926 | continue; |
2927 | } |
2928 | } |
2929 | |
2930 | gcc_unreachable (); |
2931 | } |
2932 | while (1); |
2933 | } |
2934 | |
2935 | /* Expand ADD_OVERFLOW STMT. */ |
2936 | |
2937 | static void |
2938 | expand_ADD_OVERFLOW (internal_fn, gcall *stmt) |
2939 | { |
2940 | expand_arith_overflow (code: PLUS_EXPR, stmt); |
2941 | } |
2942 | |
2943 | /* Expand SUB_OVERFLOW STMT. */ |
2944 | |
2945 | static void |
2946 | expand_SUB_OVERFLOW (internal_fn, gcall *stmt) |
2947 | { |
2948 | expand_arith_overflow (code: MINUS_EXPR, stmt); |
2949 | } |
2950 | |
2951 | /* Expand MUL_OVERFLOW STMT. */ |
2952 | |
2953 | static void |
2954 | expand_MUL_OVERFLOW (internal_fn, gcall *stmt) |
2955 | { |
2956 | expand_arith_overflow (code: MULT_EXPR, stmt); |
2957 | } |
2958 | |
2959 | /* Expand UADDC STMT. */ |
2960 | |
2961 | static void |
2962 | expand_UADDC (internal_fn ifn, gcall *stmt) |
2963 | { |
2964 | tree lhs = gimple_call_lhs (gs: stmt); |
2965 | tree arg1 = gimple_call_arg (gs: stmt, index: 0); |
2966 | tree arg2 = gimple_call_arg (gs: stmt, index: 1); |
2967 | tree arg3 = gimple_call_arg (gs: stmt, index: 2); |
2968 | tree type = TREE_TYPE (arg1); |
2969 | machine_mode mode = TYPE_MODE (type); |
2970 | insn_code icode = optab_handler (op: ifn == IFN_UADDC |
2971 | ? uaddc5_optab : usubc5_optab, mode); |
2972 | rtx op1 = expand_normal (exp: arg1); |
2973 | rtx op2 = expand_normal (exp: arg2); |
2974 | rtx op3 = expand_normal (exp: arg3); |
2975 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
2976 | rtx re = gen_reg_rtx (mode); |
2977 | rtx im = gen_reg_rtx (mode); |
2978 | class expand_operand ops[5]; |
2979 | create_output_operand (op: &ops[0], x: re, mode); |
2980 | create_output_operand (op: &ops[1], x: im, mode); |
2981 | create_input_operand (op: &ops[2], value: op1, mode); |
2982 | create_input_operand (op: &ops[3], value: op2, mode); |
2983 | create_input_operand (op: &ops[4], value: op3, mode); |
2984 | expand_insn (icode, nops: 5, ops); |
2985 | write_complex_part (target, re, false, false); |
2986 | write_complex_part (target, im, true, false); |
2987 | } |
2988 | |
2989 | /* Expand USUBC STMT. */ |
2990 | |
2991 | static void |
2992 | expand_USUBC (internal_fn ifn, gcall *stmt) |
2993 | { |
2994 | expand_UADDC (ifn, stmt); |
2995 | } |
2996 | |
2997 | /* This should get folded in tree-vectorizer.cc. */ |
2998 | |
2999 | static void |
3000 | expand_LOOP_VECTORIZED (internal_fn, gcall *) |
3001 | { |
3002 | gcc_unreachable (); |
3003 | } |
3004 | |
3005 | /* This should get folded in tree-vectorizer.cc. */ |
3006 | |
3007 | static void |
3008 | expand_LOOP_DIST_ALIAS (internal_fn, gcall *) |
3009 | { |
3010 | gcc_unreachable (); |
3011 | } |
3012 | |
3013 | /* Return a memory reference of type TYPE for argument INDEX of STMT. |
3014 | Use argument INDEX + 1 to derive the second (TBAA) operand. */ |
3015 | |
3016 | static tree |
3017 | expand_call_mem_ref (tree type, gcall *stmt, int index) |
3018 | { |
3019 | tree addr = gimple_call_arg (gs: stmt, index); |
3020 | tree alias_ptr_type = TREE_TYPE (gimple_call_arg (stmt, index + 1)); |
3021 | unsigned int align = tree_to_shwi (gimple_call_arg (gs: stmt, index: index + 1)); |
3022 | if (TYPE_ALIGN (type) != align) |
3023 | type = build_aligned_type (type, align); |
3024 | |
3025 | tree tmp = addr; |
3026 | if (TREE_CODE (tmp) == SSA_NAME) |
3027 | { |
3028 | gimple *def = get_gimple_for_ssa_name (exp: tmp); |
3029 | if (def && gimple_assign_single_p (gs: def)) |
3030 | tmp = gimple_assign_rhs1 (gs: def); |
3031 | } |
3032 | |
3033 | if (TREE_CODE (tmp) == ADDR_EXPR) |
3034 | { |
3035 | tree mem = TREE_OPERAND (tmp, 0); |
3036 | if (TREE_CODE (mem) == TARGET_MEM_REF |
3037 | && types_compatible_p (TREE_TYPE (mem), type2: type)) |
3038 | { |
3039 | tree offset = TMR_OFFSET (mem); |
3040 | if (type != TREE_TYPE (mem) |
3041 | || alias_ptr_type != TREE_TYPE (offset) |
3042 | || !integer_zerop (offset)) |
3043 | { |
3044 | mem = copy_node (mem); |
3045 | TMR_OFFSET (mem) = wide_int_to_tree (type: alias_ptr_type, |
3046 | cst: wi::to_poly_wide (t: offset)); |
3047 | TREE_TYPE (mem) = type; |
3048 | } |
3049 | return mem; |
3050 | } |
3051 | } |
3052 | |
3053 | return fold_build2 (MEM_REF, type, addr, build_int_cst (alias_ptr_type, 0)); |
3054 | } |
3055 | |
3056 | /* Expand MASK_LOAD{,_LANES}, MASK_LEN_LOAD or LEN_LOAD call STMT using optab |
3057 | * OPTAB. */ |
3058 | |
3059 | static void |
3060 | expand_partial_load_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
3061 | { |
3062 | int i = 0; |
3063 | class expand_operand ops[6]; |
3064 | tree type, lhs, rhs, maskt; |
3065 | rtx mem, target; |
3066 | insn_code icode; |
3067 | |
3068 | maskt = gimple_call_arg (gs: stmt, index: internal_fn_mask_index (ifn)); |
3069 | lhs = gimple_call_lhs (gs: stmt); |
3070 | if (lhs == NULL_TREE) |
3071 | return; |
3072 | type = TREE_TYPE (lhs); |
3073 | rhs = expand_call_mem_ref (type, stmt, index: 0); |
3074 | |
3075 | if (optab == vec_mask_load_lanes_optab |
3076 | || optab == vec_mask_len_load_lanes_optab) |
3077 | icode = get_multi_vector_move (array_type: type, optab); |
3078 | else if (optab == len_load_optab) |
3079 | icode = direct_optab_handler (op: optab, TYPE_MODE (type)); |
3080 | else |
3081 | icode = convert_optab_handler (op: optab, TYPE_MODE (type), |
3082 | TYPE_MODE (TREE_TYPE (maskt))); |
3083 | |
3084 | mem = expand_expr (exp: rhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3085 | gcc_assert (MEM_P (mem)); |
3086 | /* The built MEM_REF does not accurately reflect that the load |
3087 | is only partial. Clear it. */ |
3088 | set_mem_expr (mem, NULL_TREE); |
3089 | clear_mem_offset (mem); |
3090 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3091 | create_call_lhs_operand (op: &ops[i++], lhs_rtx: target, TYPE_MODE (type)); |
3092 | create_fixed_operand (op: &ops[i++], x: mem); |
3093 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3094 | expand_insn (icode, nops: i, ops); |
3095 | |
3096 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
3097 | } |
3098 | |
3099 | #define expand_mask_load_optab_fn expand_partial_load_optab_fn |
3100 | #define expand_mask_load_lanes_optab_fn expand_mask_load_optab_fn |
3101 | #define expand_len_load_optab_fn expand_partial_load_optab_fn |
3102 | #define expand_mask_len_load_optab_fn expand_partial_load_optab_fn |
3103 | |
3104 | /* Expand MASK_STORE{,_LANES}, MASK_LEN_STORE or LEN_STORE call STMT using optab |
3105 | * OPTAB. */ |
3106 | |
3107 | static void |
3108 | expand_partial_store_optab_fn (internal_fn ifn, gcall *stmt, convert_optab optab) |
3109 | { |
3110 | int i = 0; |
3111 | class expand_operand ops[5]; |
3112 | tree type, lhs, rhs, maskt; |
3113 | rtx mem, reg; |
3114 | insn_code icode; |
3115 | |
3116 | maskt = gimple_call_arg (gs: stmt, index: internal_fn_mask_index (ifn)); |
3117 | rhs = gimple_call_arg (gs: stmt, index: internal_fn_stored_value_index (ifn)); |
3118 | type = TREE_TYPE (rhs); |
3119 | lhs = expand_call_mem_ref (type, stmt, index: 0); |
3120 | |
3121 | if (optab == vec_mask_store_lanes_optab |
3122 | || optab == vec_mask_len_store_lanes_optab) |
3123 | icode = get_multi_vector_move (array_type: type, optab); |
3124 | else if (optab == len_store_optab) |
3125 | icode = direct_optab_handler (op: optab, TYPE_MODE (type)); |
3126 | else |
3127 | icode = convert_optab_handler (op: optab, TYPE_MODE (type), |
3128 | TYPE_MODE (TREE_TYPE (maskt))); |
3129 | |
3130 | mem = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3131 | gcc_assert (MEM_P (mem)); |
3132 | /* The built MEM_REF does not accurately reflect that the store |
3133 | is only partial. Clear it. */ |
3134 | set_mem_expr (mem, NULL_TREE); |
3135 | clear_mem_offset (mem); |
3136 | reg = expand_normal (exp: rhs); |
3137 | create_fixed_operand (op: &ops[i++], x: mem); |
3138 | create_input_operand (op: &ops[i++], value: reg, TYPE_MODE (type)); |
3139 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3140 | expand_insn (icode, nops: i, ops); |
3141 | } |
3142 | |
3143 | #define expand_mask_store_optab_fn expand_partial_store_optab_fn |
3144 | #define expand_mask_store_lanes_optab_fn expand_mask_store_optab_fn |
3145 | #define expand_len_store_optab_fn expand_partial_store_optab_fn |
3146 | #define expand_mask_len_store_optab_fn expand_partial_store_optab_fn |
3147 | |
3148 | /* Expand VCOND_MASK optab internal function. |
3149 | The expansion of STMT happens based on OPTAB table associated. */ |
3150 | |
3151 | static void |
3152 | expand_vec_cond_mask_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3153 | { |
3154 | class expand_operand ops[4]; |
3155 | |
3156 | tree lhs = gimple_call_lhs (gs: stmt); |
3157 | tree op0 = gimple_call_arg (gs: stmt, index: 0); |
3158 | tree op1 = gimple_call_arg (gs: stmt, index: 1); |
3159 | tree op2 = gimple_call_arg (gs: stmt, index: 2); |
3160 | tree vec_cond_type = TREE_TYPE (lhs); |
3161 | |
3162 | machine_mode mode = TYPE_MODE (vec_cond_type); |
3163 | machine_mode mask_mode = TYPE_MODE (TREE_TYPE (op0)); |
3164 | enum insn_code icode = convert_optab_handler (op: optab, to_mode: mode, from_mode: mask_mode); |
3165 | rtx mask, rtx_op1, rtx_op2; |
3166 | |
3167 | gcc_assert (icode != CODE_FOR_nothing); |
3168 | |
3169 | mask = expand_normal (exp: op0); |
3170 | rtx_op1 = expand_normal (exp: op1); |
3171 | rtx_op2 = expand_normal (exp: op2); |
3172 | |
3173 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3174 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, mode); |
3175 | create_input_operand (op: &ops[1], value: rtx_op1, mode); |
3176 | create_input_operand (op: &ops[2], value: rtx_op2, mode); |
3177 | create_input_operand (op: &ops[3], value: mask, mode: mask_mode); |
3178 | expand_insn (icode, nops: 4, ops); |
3179 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
3180 | } |
3181 | |
3182 | /* Expand VEC_SET internal functions. */ |
3183 | |
3184 | static void |
3185 | expand_vec_set_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3186 | { |
3187 | tree lhs = gimple_call_lhs (gs: stmt); |
3188 | tree op0 = gimple_call_arg (gs: stmt, index: 0); |
3189 | tree op1 = gimple_call_arg (gs: stmt, index: 1); |
3190 | tree op2 = gimple_call_arg (gs: stmt, index: 2); |
3191 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3192 | rtx src = expand_normal (exp: op0); |
3193 | |
3194 | machine_mode outermode = TYPE_MODE (TREE_TYPE (op0)); |
3195 | scalar_mode innermode = GET_MODE_INNER (outermode); |
3196 | |
3197 | rtx value = expand_normal (exp: op1); |
3198 | rtx pos = expand_normal (exp: op2); |
3199 | |
3200 | class expand_operand ops[3]; |
3201 | enum insn_code icode = optab_handler (op: optab, mode: outermode); |
3202 | |
3203 | if (icode != CODE_FOR_nothing) |
3204 | { |
3205 | rtx temp = gen_reg_rtx (outermode); |
3206 | emit_move_insn (temp, src); |
3207 | |
3208 | create_fixed_operand (op: &ops[0], x: temp); |
3209 | create_input_operand (op: &ops[1], value, mode: innermode); |
3210 | create_convert_operand_from (op: &ops[2], value: pos, TYPE_MODE (TREE_TYPE (op2)), |
3211 | unsigned_p: true); |
3212 | if (maybe_expand_insn (icode, nops: 3, ops)) |
3213 | { |
3214 | emit_move_insn (target, temp); |
3215 | return; |
3216 | } |
3217 | } |
3218 | gcc_unreachable (); |
3219 | } |
3220 | |
3221 | static void |
3222 | expand_ABNORMAL_DISPATCHER (internal_fn, gcall *) |
3223 | { |
3224 | } |
3225 | |
3226 | static void |
3227 | expand_BUILTIN_EXPECT (internal_fn, gcall *stmt) |
3228 | { |
3229 | /* When guessing was done, the hints should be already stripped away. */ |
3230 | gcc_assert (!flag_guess_branch_prob || optimize == 0 || seen_error ()); |
3231 | |
3232 | rtx target; |
3233 | tree lhs = gimple_call_lhs (gs: stmt); |
3234 | if (lhs) |
3235 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3236 | else |
3237 | target = const0_rtx; |
3238 | rtx val = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), target, VOIDmode, modifier: EXPAND_NORMAL); |
3239 | if (lhs && val != target) |
3240 | emit_move_insn (target, val); |
3241 | } |
3242 | |
3243 | /* IFN_VA_ARG is supposed to be expanded at pass_stdarg. So this dummy function |
3244 | should never be called. */ |
3245 | |
3246 | static void |
3247 | expand_VA_ARG (internal_fn, gcall *) |
3248 | { |
3249 | gcc_unreachable (); |
3250 | } |
3251 | |
3252 | /* IFN_VEC_CONVERT is supposed to be expanded at pass_lower_vector. So this |
3253 | dummy function should never be called. */ |
3254 | |
3255 | static void |
3256 | expand_VEC_CONVERT (internal_fn, gcall *) |
3257 | { |
3258 | gcc_unreachable (); |
3259 | } |
3260 | |
3261 | /* Expand IFN_RAWMEMCHR internal function. */ |
3262 | |
3263 | void |
3264 | expand_RAWMEMCHR (internal_fn, gcall *stmt) |
3265 | { |
3266 | expand_operand ops[3]; |
3267 | |
3268 | tree lhs = gimple_call_lhs (gs: stmt); |
3269 | if (!lhs) |
3270 | return; |
3271 | machine_mode lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); |
3272 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3273 | create_call_lhs_operand (op: &ops[0], lhs_rtx, mode: lhs_mode); |
3274 | |
3275 | tree mem = gimple_call_arg (gs: stmt, index: 0); |
3276 | rtx mem_rtx = get_memory_rtx (exp: mem, NULL); |
3277 | create_fixed_operand (op: &ops[1], x: mem_rtx); |
3278 | |
3279 | tree pattern = gimple_call_arg (gs: stmt, index: 1); |
3280 | machine_mode mode = TYPE_MODE (TREE_TYPE (pattern)); |
3281 | rtx pattern_rtx = expand_normal (exp: pattern); |
3282 | create_input_operand (op: &ops[2], value: pattern_rtx, mode); |
3283 | |
3284 | insn_code icode = direct_optab_handler (op: rawmemchr_optab, mode); |
3285 | |
3286 | expand_insn (icode, nops: 3, ops); |
3287 | assign_call_lhs (lhs, lhs_rtx, op: &ops[0]); |
3288 | } |
3289 | |
3290 | /* Expand the IFN_UNIQUE function according to its first argument. */ |
3291 | |
3292 | static void |
3293 | expand_UNIQUE (internal_fn, gcall *stmt) |
3294 | { |
3295 | rtx pattern = NULL_RTX; |
3296 | enum ifn_unique_kind kind |
3297 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (stmt, 0)); |
3298 | |
3299 | switch (kind) |
3300 | { |
3301 | default: |
3302 | gcc_unreachable (); |
3303 | |
3304 | case IFN_UNIQUE_UNSPEC: |
3305 | if (targetm.have_unique ()) |
3306 | pattern = targetm.gen_unique (); |
3307 | break; |
3308 | |
3309 | case IFN_UNIQUE_OACC_FORK: |
3310 | case IFN_UNIQUE_OACC_JOIN: |
3311 | if (targetm.have_oacc_fork () && targetm.have_oacc_join ()) |
3312 | { |
3313 | tree lhs = gimple_call_lhs (gs: stmt); |
3314 | rtx target = const0_rtx; |
3315 | |
3316 | if (lhs) |
3317 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3318 | |
3319 | rtx data_dep = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
3320 | rtx axis = expand_normal (exp: gimple_call_arg (gs: stmt, index: 2)); |
3321 | |
3322 | if (kind == IFN_UNIQUE_OACC_FORK) |
3323 | pattern = targetm.gen_oacc_fork (target, data_dep, axis); |
3324 | else |
3325 | pattern = targetm.gen_oacc_join (target, data_dep, axis); |
3326 | } |
3327 | else |
3328 | gcc_unreachable (); |
3329 | break; |
3330 | } |
3331 | |
3332 | if (pattern) |
3333 | emit_insn (pattern); |
3334 | } |
3335 | |
3336 | /* Expand the IFN_DEFERRED_INIT function: |
3337 | LHS = DEFERRED_INIT (SIZE of the DECL, INIT_TYPE, NAME of the DECL); |
3338 | |
3339 | Initialize the LHS with zero/pattern according to its second argument |
3340 | INIT_TYPE: |
3341 | if INIT_TYPE is AUTO_INIT_ZERO, use zeroes to initialize; |
3342 | if INIT_TYPE is AUTO_INIT_PATTERN, use 0xFE byte-repeatable pattern |
3343 | to initialize; |
3344 | The LHS variable is initialized including paddings. |
3345 | The reasons to choose 0xFE for pattern initialization are: |
3346 | 1. It is a non-canonical virtual address on x86_64, and at the |
3347 | high end of the i386 kernel address space. |
3348 | 2. It is a very large float value (-1.694739530317379e+38). |
3349 | 3. It is also an unusual number for integers. */ |
3350 | #define INIT_PATTERN_VALUE 0xFE |
3351 | static void |
3352 | expand_DEFERRED_INIT (internal_fn, gcall *stmt) |
3353 | { |
3354 | tree lhs = gimple_call_lhs (gs: stmt); |
3355 | tree var_size = gimple_call_arg (gs: stmt, index: 0); |
3356 | enum auto_init_type init_type |
3357 | = (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1)); |
3358 | bool reg_lhs = true; |
3359 | |
3360 | tree var_type = TREE_TYPE (lhs); |
3361 | gcc_assert (init_type > AUTO_INIT_UNINITIALIZED); |
3362 | |
3363 | if (TREE_CODE (lhs) == SSA_NAME) |
3364 | reg_lhs = true; |
3365 | else |
3366 | { |
3367 | tree lhs_base = lhs; |
3368 | while (handled_component_p (t: lhs_base)) |
3369 | lhs_base = TREE_OPERAND (lhs_base, 0); |
3370 | reg_lhs = (mem_ref_refers_to_non_mem_p (lhs_base) |
3371 | || non_mem_decl_p (lhs_base)); |
3372 | /* If this expands to a register and the underlying decl is wrapped in |
3373 | a MEM_REF that just serves as an access type change expose the decl |
3374 | if it is of correct size. This avoids a situation as in PR103271 |
3375 | if the target does not support a direct move to the registers mode. */ |
3376 | if (reg_lhs |
3377 | && TREE_CODE (lhs_base) == MEM_REF |
3378 | && TREE_CODE (TREE_OPERAND (lhs_base, 0)) == ADDR_EXPR |
3379 | && DECL_P (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)) |
3380 | && integer_zerop (TREE_OPERAND (lhs_base, 1)) |
3381 | && tree_fits_uhwi_p (var_size) |
3382 | && tree_int_cst_equal |
3383 | (var_size, |
3384 | DECL_SIZE_UNIT (TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0)))) |
3385 | { |
3386 | lhs = TREE_OPERAND (TREE_OPERAND (lhs_base, 0), 0); |
3387 | var_type = TREE_TYPE (lhs); |
3388 | } |
3389 | } |
3390 | |
3391 | if (!reg_lhs) |
3392 | { |
3393 | /* If the variable is not in register, expand to a memset |
3394 | to initialize it. */ |
3395 | mark_addressable (lhs); |
3396 | tree var_addr = build_fold_addr_expr (lhs); |
3397 | |
3398 | tree value = (init_type == AUTO_INIT_PATTERN) |
3399 | ? build_int_cst (integer_type_node, |
3400 | INIT_PATTERN_VALUE) |
3401 | : integer_zero_node; |
3402 | tree m_call = build_call_expr (builtin_decl_implicit (fncode: BUILT_IN_MEMSET), |
3403 | 3, var_addr, value, var_size); |
3404 | /* Expand this memset call. */ |
3405 | expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type)); |
3406 | } |
3407 | else |
3408 | { |
3409 | /* If this variable is in a register use expand_assignment. |
3410 | For boolean scalars force zero-init. */ |
3411 | tree init; |
3412 | scalar_int_mode var_mode; |
3413 | if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE |
3414 | && tree_fits_uhwi_p (var_size) |
3415 | && (init_type == AUTO_INIT_PATTERN |
3416 | || !is_gimple_reg_type (type: var_type)) |
3417 | && int_mode_for_size (size: tree_to_uhwi (var_size) * BITS_PER_UNIT, |
3418 | limit: 0).exists (mode: &var_mode) |
3419 | && have_insn_for (SET, var_mode)) |
3420 | { |
3421 | unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size); |
3422 | unsigned char *buf = XALLOCAVEC (unsigned char, total_bytes); |
3423 | memset (s: buf, c: (init_type == AUTO_INIT_PATTERN |
3424 | ? INIT_PATTERN_VALUE : 0), n: total_bytes); |
3425 | tree itype = build_nonstandard_integer_type |
3426 | (total_bytes * BITS_PER_UNIT, 1); |
3427 | wide_int w = wi::from_buffer (buf, total_bytes); |
3428 | init = wide_int_to_tree (type: itype, cst: w); |
3429 | /* Pun the LHS to make sure its type has constant size |
3430 | unless it is an SSA name where that's already known. */ |
3431 | if (TREE_CODE (lhs) != SSA_NAME) |
3432 | lhs = build1 (VIEW_CONVERT_EXPR, itype, lhs); |
3433 | else |
3434 | init = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (lhs), init); |
3435 | } |
3436 | else |
3437 | /* Use zero-init also for variable-length sizes. */ |
3438 | init = build_zero_cst (var_type); |
3439 | |
3440 | expand_assignment (lhs, init, false); |
3441 | } |
3442 | } |
3443 | |
3444 | /* Expand the IFN_ACCESS_WITH_SIZE function: |
3445 | ACCESS_WITH_SIZE (REF_TO_OBJ, REF_TO_SIZE, CLASS_OF_SIZE, |
3446 | TYPE_OF_SIZE, ACCESS_MODE) |
3447 | which returns the REF_TO_OBJ same as the 1st argument; |
3448 | |
3449 | 1st argument REF_TO_OBJ: The reference to the object; |
3450 | 2nd argument REF_TO_SIZE: The reference to the size of the object, |
3451 | 3rd argument CLASS_OF_SIZE: The size referenced by the REF_TO_SIZE represents |
3452 | 0: the number of bytes. |
3453 | 1: the number of the elements of the object type; |
3454 | 4th argument TYPE_OF_SIZE: A constant 0 with its TYPE being the same as the TYPE |
3455 | of the object referenced by REF_TO_SIZE |
3456 | 5th argument ACCESS_MODE: |
3457 | -1: Unknown access semantics |
3458 | 0: none |
3459 | 1: read_only |
3460 | 2: write_only |
3461 | 3: read_write |
3462 | 6th argument: A constant 0 with the pointer TYPE to the original flexible |
3463 | array type. |
3464 | |
3465 | Both the return type and the type of the first argument of this |
3466 | function have been converted from the incomplete array type to |
3467 | the corresponding pointer type. |
3468 | |
3469 | For each call to a .ACCESS_WITH_SIZE, replace it with its 1st argument. */ |
3470 | |
3471 | static void |
3472 | expand_ACCESS_WITH_SIZE (internal_fn, gcall *stmt) |
3473 | { |
3474 | tree lhs = gimple_call_lhs (gs: stmt); |
3475 | tree ref_to_obj = gimple_call_arg (gs: stmt, index: 0); |
3476 | if (lhs) |
3477 | expand_assignment (lhs, ref_to_obj, false); |
3478 | } |
3479 | |
3480 | /* The size of an OpenACC compute dimension. */ |
3481 | |
3482 | static void |
3483 | expand_GOACC_DIM_SIZE (internal_fn, gcall *stmt) |
3484 | { |
3485 | tree lhs = gimple_call_lhs (gs: stmt); |
3486 | |
3487 | if (!lhs) |
3488 | return; |
3489 | |
3490 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3491 | if (targetm.have_oacc_dim_size ()) |
3492 | { |
3493 | rtx dim = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), NULL_RTX, |
3494 | VOIDmode, modifier: EXPAND_NORMAL); |
3495 | emit_insn (targetm.gen_oacc_dim_size (target, dim)); |
3496 | } |
3497 | else |
3498 | emit_move_insn (target, GEN_INT (1)); |
3499 | } |
3500 | |
3501 | /* The position of an OpenACC execution engine along one compute axis. */ |
3502 | |
3503 | static void |
3504 | expand_GOACC_DIM_POS (internal_fn, gcall *stmt) |
3505 | { |
3506 | tree lhs = gimple_call_lhs (gs: stmt); |
3507 | |
3508 | if (!lhs) |
3509 | return; |
3510 | |
3511 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3512 | if (targetm.have_oacc_dim_pos ()) |
3513 | { |
3514 | rtx dim = expand_expr (exp: gimple_call_arg (gs: stmt, index: 0), NULL_RTX, |
3515 | VOIDmode, modifier: EXPAND_NORMAL); |
3516 | emit_insn (targetm.gen_oacc_dim_pos (target, dim)); |
3517 | } |
3518 | else |
3519 | emit_move_insn (target, const0_rtx); |
3520 | } |
3521 | |
3522 | /* This is expanded by oacc_device_lower pass. */ |
3523 | |
3524 | static void |
3525 | expand_GOACC_LOOP (internal_fn, gcall *) |
3526 | { |
3527 | gcc_unreachable (); |
3528 | } |
3529 | |
3530 | /* This is expanded by oacc_device_lower pass. */ |
3531 | |
3532 | static void |
3533 | expand_GOACC_REDUCTION (internal_fn, gcall *) |
3534 | { |
3535 | gcc_unreachable (); |
3536 | } |
3537 | |
3538 | /* This is expanded by oacc_device_lower pass. */ |
3539 | |
3540 | static void |
3541 | expand_GOACC_TILE (internal_fn, gcall *) |
3542 | { |
3543 | gcc_unreachable (); |
3544 | } |
3545 | |
3546 | /* Set errno to EDOM. */ |
3547 | |
3548 | static void |
3549 | expand_SET_EDOM (internal_fn, gcall *) |
3550 | { |
3551 | #ifdef TARGET_EDOM |
3552 | #ifdef GEN_ERRNO_RTX |
3553 | rtx errno_rtx = GEN_ERRNO_RTX; |
3554 | #else |
3555 | rtx errno_rtx = gen_rtx_MEM (word_mode, gen_rtx_SYMBOL_REF (Pmode, "errno")); |
3556 | #endif |
3557 | emit_move_insn (errno_rtx, |
3558 | gen_int_mode (TARGET_EDOM, GET_MODE (errno_rtx))); |
3559 | #else |
3560 | gcc_unreachable (); |
3561 | #endif |
3562 | } |
3563 | |
3564 | /* Expand atomic bit test and set. */ |
3565 | |
3566 | static void |
3567 | expand_ATOMIC_BIT_TEST_AND_SET (internal_fn, gcall *call) |
3568 | { |
3569 | expand_ifn_atomic_bit_test_and (call); |
3570 | } |
3571 | |
3572 | /* Expand atomic bit test and complement. */ |
3573 | |
3574 | static void |
3575 | expand_ATOMIC_BIT_TEST_AND_COMPLEMENT (internal_fn, gcall *call) |
3576 | { |
3577 | expand_ifn_atomic_bit_test_and (call); |
3578 | } |
3579 | |
3580 | /* Expand atomic bit test and reset. */ |
3581 | |
3582 | static void |
3583 | expand_ATOMIC_BIT_TEST_AND_RESET (internal_fn, gcall *call) |
3584 | { |
3585 | expand_ifn_atomic_bit_test_and (call); |
3586 | } |
3587 | |
3588 | /* Expand atomic bit test and set. */ |
3589 | |
3590 | static void |
3591 | expand_ATOMIC_COMPARE_EXCHANGE (internal_fn, gcall *call) |
3592 | { |
3593 | expand_ifn_atomic_compare_exchange (call); |
3594 | } |
3595 | |
3596 | /* Expand atomic add fetch and cmp with 0. */ |
3597 | |
3598 | static void |
3599 | expand_ATOMIC_ADD_FETCH_CMP_0 (internal_fn, gcall *call) |
3600 | { |
3601 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3602 | } |
3603 | |
3604 | /* Expand atomic sub fetch and cmp with 0. */ |
3605 | |
3606 | static void |
3607 | expand_ATOMIC_SUB_FETCH_CMP_0 (internal_fn, gcall *call) |
3608 | { |
3609 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3610 | } |
3611 | |
3612 | /* Expand atomic and fetch and cmp with 0. */ |
3613 | |
3614 | static void |
3615 | expand_ATOMIC_AND_FETCH_CMP_0 (internal_fn, gcall *call) |
3616 | { |
3617 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3618 | } |
3619 | |
3620 | /* Expand atomic or fetch and cmp with 0. */ |
3621 | |
3622 | static void |
3623 | expand_ATOMIC_OR_FETCH_CMP_0 (internal_fn, gcall *call) |
3624 | { |
3625 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3626 | } |
3627 | |
3628 | /* Expand atomic xor fetch and cmp with 0. */ |
3629 | |
3630 | static void |
3631 | expand_ATOMIC_XOR_FETCH_CMP_0 (internal_fn, gcall *call) |
3632 | { |
3633 | expand_ifn_atomic_op_fetch_cmp_0 (call); |
3634 | } |
3635 | |
3636 | /* Expand LAUNDER to assignment, lhs = arg0. */ |
3637 | |
3638 | static void |
3639 | expand_LAUNDER (internal_fn, gcall *call) |
3640 | { |
3641 | tree lhs = gimple_call_lhs (gs: call); |
3642 | |
3643 | if (!lhs) |
3644 | return; |
3645 | |
3646 | expand_assignment (lhs, gimple_call_arg (gs: call, index: 0), false); |
3647 | } |
3648 | |
3649 | /* Expand {MASK_,}SCATTER_STORE{S,U} call CALL using optab OPTAB. */ |
3650 | |
3651 | static void |
3652 | expand_scatter_store_optab_fn (internal_fn, gcall *stmt, direct_optab optab) |
3653 | { |
3654 | internal_fn ifn = gimple_call_internal_fn (gs: stmt); |
3655 | int rhs_index = internal_fn_stored_value_index (ifn); |
3656 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3657 | tree offset = gimple_call_arg (gs: stmt, index: 1); |
3658 | tree scale = gimple_call_arg (gs: stmt, index: 2); |
3659 | tree rhs = gimple_call_arg (gs: stmt, index: rhs_index); |
3660 | |
3661 | rtx base_rtx = expand_normal (exp: base); |
3662 | rtx offset_rtx = expand_normal (exp: offset); |
3663 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); |
3664 | rtx rhs_rtx = expand_normal (exp: rhs); |
3665 | |
3666 | class expand_operand ops[8]; |
3667 | int i = 0; |
3668 | create_address_operand (op: &ops[i++], value: base_rtx); |
3669 | create_input_operand (op: &ops[i++], value: offset_rtx, TYPE_MODE (TREE_TYPE (offset))); |
3670 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); |
3671 | create_integer_operand (&ops[i++], scale_int); |
3672 | create_input_operand (op: &ops[i++], value: rhs_rtx, TYPE_MODE (TREE_TYPE (rhs))); |
3673 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3674 | |
3675 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (TREE_TYPE (rhs)), |
3676 | TYPE_MODE (TREE_TYPE (offset))); |
3677 | expand_insn (icode, nops: i, ops); |
3678 | } |
3679 | |
3680 | /* Expand {MASK_,}GATHER_LOAD call CALL using optab OPTAB. */ |
3681 | |
3682 | static void |
3683 | expand_gather_load_optab_fn (internal_fn, gcall *stmt, direct_optab optab) |
3684 | { |
3685 | tree lhs = gimple_call_lhs (gs: stmt); |
3686 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3687 | tree offset = gimple_call_arg (gs: stmt, index: 1); |
3688 | tree scale = gimple_call_arg (gs: stmt, index: 2); |
3689 | |
3690 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3691 | rtx base_rtx = expand_normal (exp: base); |
3692 | rtx offset_rtx = expand_normal (exp: offset); |
3693 | HOST_WIDE_INT scale_int = tree_to_shwi (scale); |
3694 | |
3695 | int i = 0; |
3696 | class expand_operand ops[9]; |
3697 | create_call_lhs_operand (op: &ops[i++], lhs_rtx, TYPE_MODE (TREE_TYPE (lhs))); |
3698 | create_address_operand (op: &ops[i++], value: base_rtx); |
3699 | create_input_operand (op: &ops[i++], value: offset_rtx, TYPE_MODE (TREE_TYPE (offset))); |
3700 | create_integer_operand (&ops[i++], TYPE_UNSIGNED (TREE_TYPE (offset))); |
3701 | create_integer_operand (&ops[i++], scale_int); |
3702 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3703 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (TREE_TYPE (lhs)), |
3704 | TYPE_MODE (TREE_TYPE (offset))); |
3705 | expand_insn (icode, nops: i, ops); |
3706 | assign_call_lhs (lhs, lhs_rtx, op: &ops[0]); |
3707 | } |
3708 | |
3709 | /* Expand MASK_LEN_STRIDED_LOAD call CALL by optab OPTAB. */ |
3710 | |
3711 | static void |
3712 | expand_strided_load_optab_fn (ATTRIBUTE_UNUSED internal_fn, gcall *stmt, |
3713 | direct_optab optab) |
3714 | { |
3715 | tree lhs = gimple_call_lhs (gs: stmt); |
3716 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3717 | tree stride = gimple_call_arg (gs: stmt, index: 1); |
3718 | |
3719 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3720 | rtx base_rtx = expand_normal (exp: base); |
3721 | rtx stride_rtx = expand_normal (exp: stride); |
3722 | |
3723 | unsigned i = 0; |
3724 | class expand_operand ops[7]; |
3725 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
3726 | |
3727 | create_output_operand (op: &ops[i++], x: lhs_rtx, mode); |
3728 | create_address_operand (op: &ops[i++], value: base_rtx); |
3729 | create_address_operand (op: &ops[i++], value: stride_rtx); |
3730 | |
3731 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3732 | expand_insn (icode: direct_optab_handler (op: optab, mode), nops: i, ops); |
3733 | |
3734 | if (!rtx_equal_p (lhs_rtx, ops[0].value)) |
3735 | emit_move_insn (lhs_rtx, ops[0].value); |
3736 | } |
3737 | |
3738 | /* Expand MASK_LEN_STRIDED_STORE call CALL by optab OPTAB. */ |
3739 | |
3740 | static void |
3741 | expand_strided_store_optab_fn (ATTRIBUTE_UNUSED internal_fn, gcall *stmt, |
3742 | direct_optab optab) |
3743 | { |
3744 | internal_fn fn = gimple_call_internal_fn (gs: stmt); |
3745 | int rhs_index = internal_fn_stored_value_index (fn); |
3746 | |
3747 | tree base = gimple_call_arg (gs: stmt, index: 0); |
3748 | tree stride = gimple_call_arg (gs: stmt, index: 1); |
3749 | tree rhs = gimple_call_arg (gs: stmt, index: rhs_index); |
3750 | |
3751 | rtx base_rtx = expand_normal (exp: base); |
3752 | rtx stride_rtx = expand_normal (exp: stride); |
3753 | rtx rhs_rtx = expand_normal (exp: rhs); |
3754 | |
3755 | unsigned i = 0; |
3756 | class expand_operand ops[6]; |
3757 | machine_mode mode = TYPE_MODE (TREE_TYPE (rhs)); |
3758 | |
3759 | create_address_operand (op: &ops[i++], value: base_rtx); |
3760 | create_address_operand (op: &ops[i++], value: stride_rtx); |
3761 | create_input_operand (op: &ops[i++], value: rhs_rtx, mode); |
3762 | |
3763 | i = add_mask_else_and_len_args (ops, opno: i, stmt); |
3764 | expand_insn (icode: direct_optab_handler (op: optab, mode), nops: i, ops); |
3765 | } |
3766 | |
3767 | /* Helper for expand_DIVMOD. Return true if the sequence starting with |
3768 | INSN contains any call insns or insns with {,U}{DIV,MOD} rtxes. */ |
3769 | |
3770 | static bool |
3771 | contains_call_div_mod (rtx_insn *insn) |
3772 | { |
3773 | subrtx_iterator::array_type array; |
3774 | for (; insn; insn = NEXT_INSN (insn)) |
3775 | if (CALL_P (insn)) |
3776 | return true; |
3777 | else if (INSN_P (insn)) |
3778 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) |
3779 | switch (GET_CODE (*iter)) |
3780 | { |
3781 | case CALL: |
3782 | case DIV: |
3783 | case UDIV: |
3784 | case MOD: |
3785 | case UMOD: |
3786 | return true; |
3787 | default: |
3788 | break; |
3789 | } |
3790 | return false; |
3791 | } |
3792 | |
3793 | /* Expand DIVMOD() using: |
3794 | a) optab handler for udivmod/sdivmod if it is available. |
3795 | b) If optab_handler doesn't exist, generate call to |
3796 | target-specific divmod libfunc. */ |
3797 | |
3798 | static void |
3799 | expand_DIVMOD (internal_fn, gcall *call_stmt) |
3800 | { |
3801 | tree lhs = gimple_call_lhs (gs: call_stmt); |
3802 | tree arg0 = gimple_call_arg (gs: call_stmt, index: 0); |
3803 | tree arg1 = gimple_call_arg (gs: call_stmt, index: 1); |
3804 | |
3805 | gcc_assert (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE); |
3806 | tree type = TREE_TYPE (TREE_TYPE (lhs)); |
3807 | machine_mode mode = TYPE_MODE (type); |
3808 | bool unsignedp = TYPE_UNSIGNED (type); |
3809 | optab tab = (unsignedp) ? udivmod_optab : sdivmod_optab; |
3810 | |
3811 | rtx op0 = expand_normal (exp: arg0); |
3812 | rtx op1 = expand_normal (exp: arg1); |
3813 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3814 | |
3815 | rtx quotient = NULL_RTX, remainder = NULL_RTX; |
3816 | rtx_insn *insns = NULL; |
3817 | |
3818 | if (TREE_CODE (arg1) == INTEGER_CST) |
3819 | { |
3820 | /* For DIVMOD by integral constants, there could be efficient code |
3821 | expanded inline e.g. using shifts and plus/minus. Try to expand |
3822 | the division and modulo and if it emits any library calls or any |
3823 | {,U}{DIV,MOD} rtxes throw it away and use a divmod optab or |
3824 | divmod libcall. */ |
3825 | scalar_int_mode int_mode; |
3826 | if (remainder == NULL_RTX |
3827 | && optimize |
3828 | && CONST_INT_P (op1) |
3829 | && !pow2p_hwi (INTVAL (op1)) |
3830 | && is_int_mode (TYPE_MODE (type), int_mode: &int_mode) |
3831 | && GET_MODE_SIZE (mode: int_mode) == 2 * UNITS_PER_WORD |
3832 | && optab_handler (op: and_optab, mode: word_mode) != CODE_FOR_nothing |
3833 | && optab_handler (op: add_optab, mode: word_mode) != CODE_FOR_nothing |
3834 | && optimize_insn_for_speed_p ()) |
3835 | { |
3836 | rtx_insn *last = get_last_insn (); |
3837 | remainder = NULL_RTX; |
3838 | quotient = expand_doubleword_divmod (int_mode, op0, op1, &remainder, |
3839 | TYPE_UNSIGNED (type)); |
3840 | if (quotient != NULL_RTX) |
3841 | { |
3842 | if (optab_handler (op: mov_optab, mode: int_mode) != CODE_FOR_nothing) |
3843 | { |
3844 | rtx_insn *move = emit_move_insn (quotient, quotient); |
3845 | set_dst_reg_note (move, REG_EQUAL, |
3846 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) |
3847 | ? UDIV : DIV, int_mode, |
3848 | copy_rtx (op0), op1), |
3849 | quotient); |
3850 | move = emit_move_insn (remainder, remainder); |
3851 | set_dst_reg_note (move, REG_EQUAL, |
3852 | gen_rtx_fmt_ee (TYPE_UNSIGNED (type) |
3853 | ? UMOD : MOD, int_mode, |
3854 | copy_rtx (op0), op1), |
3855 | quotient); |
3856 | } |
3857 | } |
3858 | else |
3859 | delete_insns_since (last); |
3860 | } |
3861 | |
3862 | if (remainder == NULL_RTX) |
3863 | { |
3864 | struct separate_ops ops; |
3865 | ops.code = TRUNC_DIV_EXPR; |
3866 | ops.type = type; |
3867 | ops.op0 = make_tree (ops.type, op0); |
3868 | ops.op1 = arg1; |
3869 | ops.op2 = NULL_TREE; |
3870 | ops.location = gimple_location (g: call_stmt); |
3871 | start_sequence (); |
3872 | quotient = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL); |
3873 | if (contains_call_div_mod (insn: get_insns ())) |
3874 | quotient = NULL_RTX; |
3875 | else |
3876 | { |
3877 | ops.code = TRUNC_MOD_EXPR; |
3878 | remainder = expand_expr_real_2 (&ops, NULL_RTX, mode, |
3879 | EXPAND_NORMAL); |
3880 | if (contains_call_div_mod (insn: get_insns ())) |
3881 | remainder = NULL_RTX; |
3882 | } |
3883 | if (remainder) |
3884 | insns = get_insns (); |
3885 | end_sequence (); |
3886 | } |
3887 | } |
3888 | |
3889 | if (remainder) |
3890 | emit_insn (insns); |
3891 | |
3892 | /* Check if optab_handler exists for divmod_optab for given mode. */ |
3893 | else if (optab_handler (op: tab, mode) != CODE_FOR_nothing) |
3894 | { |
3895 | quotient = gen_reg_rtx (mode); |
3896 | remainder = gen_reg_rtx (mode); |
3897 | expand_twoval_binop (tab, op0, op1, quotient, remainder, unsignedp); |
3898 | } |
3899 | |
3900 | /* Generate call to divmod libfunc if it exists. */ |
3901 | else if (rtx libfunc = optab_libfunc (tab, mode)) |
3902 | targetm.expand_divmod_libfunc (libfunc, mode, op0, op1, |
3903 | "ient, &remainder); |
3904 | |
3905 | else |
3906 | gcc_unreachable (); |
3907 | |
3908 | /* Wrap the return value (quotient, remainder) within COMPLEX_EXPR. */ |
3909 | expand_expr (exp: build2 (COMPLEX_EXPR, TREE_TYPE (lhs), |
3910 | make_tree (TREE_TYPE (arg0), quotient), |
3911 | make_tree (TREE_TYPE (arg1), remainder)), |
3912 | target, VOIDmode, modifier: EXPAND_NORMAL); |
3913 | } |
3914 | |
3915 | /* Expand a NOP. */ |
3916 | |
3917 | static void |
3918 | expand_NOP (internal_fn, gcall *) |
3919 | { |
3920 | /* Nothing. But it shouldn't really prevail. */ |
3921 | } |
3922 | |
3923 | /* Coroutines, all should have been processed at this stage. */ |
3924 | |
3925 | static void |
3926 | expand_CO_FRAME (internal_fn, gcall *) |
3927 | { |
3928 | gcc_unreachable (); |
3929 | } |
3930 | |
3931 | static void |
3932 | expand_CO_YIELD (internal_fn, gcall *) |
3933 | { |
3934 | gcc_unreachable (); |
3935 | } |
3936 | |
3937 | static void |
3938 | expand_CO_SUSPN (internal_fn, gcall *) |
3939 | { |
3940 | gcc_unreachable (); |
3941 | } |
3942 | |
3943 | static void |
3944 | expand_CO_ACTOR (internal_fn, gcall *) |
3945 | { |
3946 | gcc_unreachable (); |
3947 | } |
3948 | |
3949 | /* Expand a call to FN using the operands in STMT. FN has a single |
3950 | output operand and NARGS input operands. */ |
3951 | |
3952 | static void |
3953 | expand_direct_optab_fn (internal_fn fn, gcall *stmt, direct_optab optab, |
3954 | unsigned int nargs) |
3955 | { |
3956 | tree_pair types = direct_internal_fn_types (fn, stmt); |
3957 | insn_code icode = direct_optab_handler (op: optab, TYPE_MODE (types.first)); |
3958 | expand_fn_using_insn (stmt, icode, noutputs: 1, ninputs: nargs); |
3959 | } |
3960 | |
3961 | /* Expand WHILE_ULT call STMT using optab OPTAB. */ |
3962 | |
3963 | static void |
3964 | expand_while_optab_fn (internal_fn, gcall *stmt, convert_optab optab) |
3965 | { |
3966 | expand_operand ops[4]; |
3967 | tree rhs_type[2]; |
3968 | |
3969 | tree lhs = gimple_call_lhs (gs: stmt); |
3970 | tree lhs_type = TREE_TYPE (lhs); |
3971 | rtx lhs_rtx = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
3972 | create_call_lhs_operand (op: &ops[0], lhs_rtx, TYPE_MODE (lhs_type)); |
3973 | |
3974 | for (unsigned int i = 0; i < 2; ++i) |
3975 | { |
3976 | tree rhs = gimple_call_arg (gs: stmt, index: i); |
3977 | rhs_type[i] = TREE_TYPE (rhs); |
3978 | rtx rhs_rtx = expand_normal (exp: rhs); |
3979 | create_input_operand (op: &ops[i + 1], value: rhs_rtx, TYPE_MODE (rhs_type[i])); |
3980 | } |
3981 | |
3982 | int opcnt; |
3983 | if (!VECTOR_MODE_P (TYPE_MODE (lhs_type))) |
3984 | { |
3985 | /* When the mask is an integer mode the exact vector length may not |
3986 | be clear to the backend, so we pass it in operand[3]. |
3987 | Use the vector in arg2 for the most reliable intended size. */ |
3988 | tree type = TREE_TYPE (gimple_call_arg (stmt, 2)); |
3989 | create_integer_operand (&ops[3], TYPE_VECTOR_SUBPARTS (node: type)); |
3990 | opcnt = 4; |
3991 | } |
3992 | else |
3993 | /* The mask has a vector type so the length operand is unnecessary. */ |
3994 | opcnt = 3; |
3995 | |
3996 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (rhs_type[0]), |
3997 | TYPE_MODE (lhs_type)); |
3998 | |
3999 | expand_insn (icode, nops: opcnt, ops); |
4000 | assign_call_lhs (lhs, lhs_rtx, op: &ops[0]); |
4001 | } |
4002 | |
4003 | /* Expand a call to a convert-like optab using the operands in STMT. |
4004 | FN has a single output operand and NARGS input operands. */ |
4005 | |
4006 | static void |
4007 | expand_convert_optab_fn (internal_fn fn, gcall *stmt, convert_optab optab, |
4008 | unsigned int nargs) |
4009 | { |
4010 | tree_pair types = direct_internal_fn_types (fn, stmt); |
4011 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (types.first), |
4012 | TYPE_MODE (types.second)); |
4013 | expand_fn_using_insn (stmt, icode, noutputs: 1, ninputs: nargs); |
4014 | } |
4015 | |
4016 | /* Expand CRC call STMT. */ |
4017 | |
4018 | static void |
4019 | expand_crc_optab_fn (internal_fn fn, gcall *stmt, convert_optab optab) |
4020 | { |
4021 | tree lhs = gimple_call_lhs (gs: stmt); |
4022 | tree rhs1 = gimple_call_arg (gs: stmt, index: 0); // crc |
4023 | tree rhs2 = gimple_call_arg (gs: stmt, index: 1); // data |
4024 | tree rhs3 = gimple_call_arg (gs: stmt, index: 2); // polynomial |
4025 | |
4026 | tree result_type = TREE_TYPE (lhs); |
4027 | tree data_type = TREE_TYPE (rhs2); |
4028 | |
4029 | gcc_assert (TYPE_MODE (result_type) >= TYPE_MODE (data_type)); |
4030 | |
4031 | rtx dest = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
4032 | rtx crc = expand_normal (exp: rhs1); |
4033 | rtx data = expand_normal (exp: rhs2); |
4034 | gcc_assert (TREE_CODE (rhs3) == INTEGER_CST); |
4035 | rtx polynomial = gen_rtx_CONST_INT (TYPE_MODE (result_type), |
4036 | TREE_INT_CST_LOW (rhs3)); |
4037 | |
4038 | /* Use target specific expansion if it exists. |
4039 | Otherwise, generate table-based CRC. */ |
4040 | if (direct_internal_fn_supported_p (fn, tree_pair (data_type, result_type), |
4041 | OPTIMIZE_FOR_SPEED)) |
4042 | { |
4043 | class expand_operand ops[4]; |
4044 | |
4045 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4046 | { |
4047 | fprintf (stream: dump_file, |
4048 | format: ";; using optab for crc_%u_polynomial_" |
4049 | HOST_WIDE_INT_PRINT_HEX "\n", |
4050 | GET_MODE_BITSIZE (GET_MODE (dest)).to_constant (), |
4051 | TREE_INT_CST_LOW (rhs3)); |
4052 | } |
4053 | |
4054 | create_call_lhs_operand (op: &ops[0], lhs_rtx: dest, TYPE_MODE (result_type)); |
4055 | create_input_operand (op: &ops[1], value: crc, TYPE_MODE (result_type)); |
4056 | create_input_operand (op: &ops[2], value: data, TYPE_MODE (data_type)); |
4057 | create_input_operand (op: &ops[3], value: polynomial, TYPE_MODE (result_type)); |
4058 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (data_type), |
4059 | TYPE_MODE (result_type)); |
4060 | expand_insn (icode, nops: 4, ops); |
4061 | assign_call_lhs (lhs, lhs_rtx: dest, op: &ops[0]); |
4062 | } |
4063 | else |
4064 | { |
4065 | /* We're bypassing all the operand conversions that are done in the |
4066 | case when we get an icode, operands and pass that off to expand_insn. |
4067 | |
4068 | That path has special case handling for promoted return values which |
4069 | we must emulate here (is the same kind of special treatment ever |
4070 | needed for input arguments here?). |
4071 | |
4072 | In particular we do not want to store directly into a promoted |
4073 | SUBREG destination, instead store into a suitably sized pseudo. */ |
4074 | rtx orig_dest = dest; |
4075 | if (SUBREG_P (dest) && SUBREG_PROMOTED_VAR_P (dest)) |
4076 | dest = gen_reg_rtx (GET_MODE (dest)); |
4077 | |
4078 | /* If it's IFN_CRC generate bit-forward CRC. */ |
4079 | if (fn == IFN_CRC) |
4080 | expand_crc_table_based (dest, crc, data, polynomial, |
4081 | TYPE_MODE (data_type)); |
4082 | else |
4083 | /* If it's IFN_CRC_REV generate bit-reversed CRC. */ |
4084 | expand_reversed_crc_table_based (dest, crc, data, polynomial, |
4085 | TYPE_MODE (data_type), |
4086 | generate_reflecting_code_standard); |
4087 | |
4088 | /* Now get the return value where it needs to be, taking care to |
4089 | ensure it's promoted appropriately if the ABI demands it. |
4090 | |
4091 | Re-use assign_call_lhs to handle the details. */ |
4092 | class expand_operand ops[4]; |
4093 | create_call_lhs_operand (op: &ops[0], lhs_rtx: dest, TYPE_MODE (result_type)); |
4094 | ops[0].value = dest; |
4095 | assign_call_lhs (lhs, lhs_rtx: orig_dest, op: &ops[0]); |
4096 | } |
4097 | } |
4098 | |
4099 | /* Expanders for optabs that can use expand_direct_optab_fn. */ |
4100 | |
4101 | #define expand_unary_optab_fn(FN, STMT, OPTAB) \ |
4102 | expand_direct_optab_fn (FN, STMT, OPTAB, 1) |
4103 | |
4104 | #define expand_binary_optab_fn(FN, STMT, OPTAB) \ |
4105 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) |
4106 | |
4107 | #define expand_ternary_optab_fn(FN, STMT, OPTAB) \ |
4108 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
4109 | |
4110 | #define expand_cond_unary_optab_fn(FN, STMT, OPTAB) \ |
4111 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
4112 | |
4113 | #define expand_cond_binary_optab_fn(FN, STMT, OPTAB) \ |
4114 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
4115 | |
4116 | #define expand_cond_ternary_optab_fn(FN, STMT, OPTAB) \ |
4117 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
4118 | |
4119 | #define expand_cond_len_unary_optab_fn(FN, STMT, OPTAB) \ |
4120 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
4121 | |
4122 | #define expand_cond_len_binary_optab_fn(FN, STMT, OPTAB) \ |
4123 | expand_direct_optab_fn (FN, STMT, OPTAB, 6) |
4124 | |
4125 | #define expand_cond_len_ternary_optab_fn(FN, STMT, OPTAB) \ |
4126 | expand_direct_optab_fn (FN, STMT, OPTAB, 7) |
4127 | |
4128 | #define expand_fold_extract_optab_fn(FN, STMT, OPTAB) \ |
4129 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
4130 | |
4131 | #define expand_fold_len_extract_optab_fn(FN, STMT, OPTAB) \ |
4132 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
4133 | |
4134 | #define expand_fold_left_optab_fn(FN, STMT, OPTAB) \ |
4135 | expand_direct_optab_fn (FN, STMT, OPTAB, 2) |
4136 | |
4137 | #define expand_mask_fold_left_optab_fn(FN, STMT, OPTAB) \ |
4138 | expand_direct_optab_fn (FN, STMT, OPTAB, 3) |
4139 | |
4140 | #define expand_mask_len_fold_left_optab_fn(FN, STMT, OPTAB) \ |
4141 | expand_direct_optab_fn (FN, STMT, OPTAB, 5) |
4142 | |
4143 | #define expand_check_ptrs_optab_fn(FN, STMT, OPTAB) \ |
4144 | expand_direct_optab_fn (FN, STMT, OPTAB, 4) |
4145 | |
4146 | /* Expanders for optabs that can use expand_convert_optab_fn. */ |
4147 | |
4148 | #define expand_unary_convert_optab_fn(FN, STMT, OPTAB) \ |
4149 | expand_convert_optab_fn (FN, STMT, OPTAB, 1) |
4150 | |
4151 | #define expand_vec_extract_optab_fn(FN, STMT, OPTAB) \ |
4152 | expand_convert_optab_fn (FN, STMT, OPTAB, 2) |
4153 | |
4154 | /* RETURN_TYPE and ARGS are a return type and argument list that are |
4155 | in principle compatible with FN (which satisfies direct_internal_fn_p). |
4156 | Return the types that should be used to determine whether the |
4157 | target supports FN. */ |
4158 | |
4159 | tree_pair |
4160 | direct_internal_fn_types (internal_fn fn, tree return_type, tree *args) |
4161 | { |
4162 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
4163 | tree type0 = (info.type0 < 0 ? return_type : TREE_TYPE (args[info.type0])); |
4164 | tree type1 = (info.type1 < 0 ? return_type : TREE_TYPE (args[info.type1])); |
4165 | return tree_pair (type0, type1); |
4166 | } |
4167 | |
4168 | /* CALL is a call whose return type and arguments are in principle |
4169 | compatible with FN (which satisfies direct_internal_fn_p). Return the |
4170 | types that should be used to determine whether the target supports FN. */ |
4171 | |
4172 | tree_pair |
4173 | direct_internal_fn_types (internal_fn fn, gcall *call) |
4174 | { |
4175 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
4176 | tree op0 = (info.type0 < 0 |
4177 | ? gimple_call_lhs (gs: call) |
4178 | : gimple_call_arg (gs: call, index: info.type0)); |
4179 | tree op1 = (info.type1 < 0 |
4180 | ? gimple_call_lhs (gs: call) |
4181 | : gimple_call_arg (gs: call, index: info.type1)); |
4182 | return tree_pair (TREE_TYPE (op0), TREE_TYPE (op1)); |
4183 | } |
4184 | |
4185 | /* Return true if OPTAB is supported for TYPES (whose modes should be |
4186 | the same) when the optimization type is OPT_TYPE. Used for simple |
4187 | direct optabs. */ |
4188 | |
4189 | static bool |
4190 | direct_optab_supported_p (direct_optab optab, tree_pair types, |
4191 | optimization_type opt_type) |
4192 | { |
4193 | machine_mode mode = TYPE_MODE (types.first); |
4194 | gcc_checking_assert (mode == TYPE_MODE (types.second)); |
4195 | return direct_optab_handler (optab, mode, opt_type) != CODE_FOR_nothing; |
4196 | } |
4197 | |
4198 | /* Return true if OPTAB is supported for TYPES, where the first type |
4199 | is the destination and the second type is the source. Used for |
4200 | convert optabs. */ |
4201 | |
4202 | static bool |
4203 | convert_optab_supported_p (convert_optab optab, tree_pair types, |
4204 | optimization_type opt_type) |
4205 | { |
4206 | return (convert_optab_handler (optab, TYPE_MODE (types.first), |
4207 | TYPE_MODE (types.second), opt_type) |
4208 | != CODE_FOR_nothing); |
4209 | } |
4210 | |
4211 | /* Return true if load/store lanes optab OPTAB is supported for |
4212 | array type TYPES.first when the optimization type is OPT_TYPE. */ |
4213 | |
4214 | static bool |
4215 | multi_vector_optab_supported_p (convert_optab optab, tree_pair types, |
4216 | optimization_type opt_type) |
4217 | { |
4218 | gcc_assert (TREE_CODE (types.first) == ARRAY_TYPE); |
4219 | machine_mode imode = TYPE_MODE (types.first); |
4220 | machine_mode vmode = TYPE_MODE (TREE_TYPE (types.first)); |
4221 | return (convert_optab_handler (optab, imode, vmode, opt_type) |
4222 | != CODE_FOR_nothing); |
4223 | } |
4224 | |
4225 | #define direct_unary_optab_supported_p direct_optab_supported_p |
4226 | #define direct_unary_convert_optab_supported_p convert_optab_supported_p |
4227 | #define direct_binary_optab_supported_p direct_optab_supported_p |
4228 | #define direct_ternary_optab_supported_p direct_optab_supported_p |
4229 | #define direct_cond_unary_optab_supported_p direct_optab_supported_p |
4230 | #define direct_cond_binary_optab_supported_p direct_optab_supported_p |
4231 | #define direct_cond_ternary_optab_supported_p direct_optab_supported_p |
4232 | #define direct_cond_len_unary_optab_supported_p direct_optab_supported_p |
4233 | #define direct_cond_len_binary_optab_supported_p direct_optab_supported_p |
4234 | #define direct_cond_len_ternary_optab_supported_p direct_optab_supported_p |
4235 | #define direct_crc_optab_supported_p convert_optab_supported_p |
4236 | #define direct_mask_load_optab_supported_p convert_optab_supported_p |
4237 | #define direct_load_lanes_optab_supported_p multi_vector_optab_supported_p |
4238 | #define direct_mask_load_lanes_optab_supported_p multi_vector_optab_supported_p |
4239 | #define direct_gather_load_optab_supported_p convert_optab_supported_p |
4240 | #define direct_strided_load_optab_supported_p direct_optab_supported_p |
4241 | #define direct_len_load_optab_supported_p direct_optab_supported_p |
4242 | #define direct_mask_len_load_optab_supported_p convert_optab_supported_p |
4243 | #define direct_mask_store_optab_supported_p convert_optab_supported_p |
4244 | #define direct_store_lanes_optab_supported_p multi_vector_optab_supported_p |
4245 | #define direct_mask_store_lanes_optab_supported_p multi_vector_optab_supported_p |
4246 | #define direct_vec_cond_mask_optab_supported_p convert_optab_supported_p |
4247 | #define direct_vec_cond_optab_supported_p convert_optab_supported_p |
4248 | #define direct_scatter_store_optab_supported_p convert_optab_supported_p |
4249 | #define direct_strided_store_optab_supported_p direct_optab_supported_p |
4250 | #define direct_len_store_optab_supported_p direct_optab_supported_p |
4251 | #define direct_mask_len_store_optab_supported_p convert_optab_supported_p |
4252 | #define direct_while_optab_supported_p convert_optab_supported_p |
4253 | #define direct_fold_extract_optab_supported_p direct_optab_supported_p |
4254 | #define direct_fold_len_extract_optab_supported_p direct_optab_supported_p |
4255 | #define direct_fold_left_optab_supported_p direct_optab_supported_p |
4256 | #define direct_mask_fold_left_optab_supported_p direct_optab_supported_p |
4257 | #define direct_mask_len_fold_left_optab_supported_p direct_optab_supported_p |
4258 | #define direct_check_ptrs_optab_supported_p direct_optab_supported_p |
4259 | #define direct_vec_set_optab_supported_p direct_optab_supported_p |
4260 | #define direct_vec_extract_optab_supported_p convert_optab_supported_p |
4261 | |
4262 | /* Return the optab used by internal function FN. */ |
4263 | |
4264 | optab |
4265 | direct_internal_fn_optab (internal_fn fn, tree_pair types) |
4266 | { |
4267 | switch (fn) |
4268 | { |
4269 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4270 | case IFN_##CODE: break; |
4271 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4272 | case IFN_##CODE: return OPTAB##_optab; |
4273 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4274 | UNSIGNED_OPTAB, TYPE) \ |
4275 | case IFN_##CODE: return (TYPE_UNSIGNED (types.SELECTOR) \ |
4276 | ? UNSIGNED_OPTAB ## _optab \ |
4277 | : SIGNED_OPTAB ## _optab); |
4278 | #include "internal-fn.def" |
4279 | |
4280 | case IFN_LAST: |
4281 | break; |
4282 | } |
4283 | gcc_unreachable (); |
4284 | } |
4285 | |
4286 | /* Return the optab used by internal function FN. */ |
4287 | |
4288 | static optab |
4289 | direct_internal_fn_optab (internal_fn fn) |
4290 | { |
4291 | switch (fn) |
4292 | { |
4293 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4294 | case IFN_##CODE: break; |
4295 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4296 | case IFN_##CODE: return OPTAB##_optab; |
4297 | #include "internal-fn.def" |
4298 | |
4299 | case IFN_LAST: |
4300 | break; |
4301 | } |
4302 | gcc_unreachable (); |
4303 | } |
4304 | |
4305 | /* Return true if TYPE's mode has the same format as TYPE, and if there is |
4306 | a 1:1 correspondence between the values that the mode can store and the |
4307 | values that the type can store. */ |
4308 | |
4309 | static bool |
4310 | type_strictly_matches_mode_p (const_tree type) |
4311 | { |
4312 | /* The masked vector operations have both vector data operands and vector |
4313 | boolean operands. The vector data operands are expected to have a vector |
4314 | mode, but the vector boolean operands can be an integer mode rather than |
4315 | a vector mode, depending on how TARGET_VECTORIZE_GET_MASK_MODE is |
4316 | defined. PR116103. */ |
4317 | if (VECTOR_BOOLEAN_TYPE_P (type) |
4318 | && SCALAR_INT_MODE_P (TYPE_MODE (type)) |
4319 | && TYPE_PRECISION (TREE_TYPE (type)) == 1) |
4320 | return true; |
4321 | |
4322 | if (VECTOR_TYPE_P (type)) |
4323 | return VECTOR_MODE_P (TYPE_MODE (type)); |
4324 | |
4325 | if (INTEGRAL_TYPE_P (type)) |
4326 | return type_has_mode_precision_p (t: type); |
4327 | |
4328 | if (SCALAR_FLOAT_TYPE_P (type) || COMPLEX_FLOAT_TYPE_P (type)) |
4329 | return true; |
4330 | |
4331 | return false; |
4332 | } |
4333 | |
4334 | /* Returns true if both types of TYPE_PAIR strictly match their modes, |
4335 | else returns false. */ |
4336 | |
4337 | static bool |
4338 | type_pair_strictly_matches_mode_p (tree_pair type_pair) |
4339 | { |
4340 | return type_strictly_matches_mode_p (type: type_pair.first) |
4341 | && type_strictly_matches_mode_p (type: type_pair.second); |
4342 | } |
4343 | |
4344 | /* Return true if FN is supported for the types in TYPES when the |
4345 | optimization type is OPT_TYPE. The types are those associated with |
4346 | the "type0" and "type1" fields of FN's direct_internal_fn_info |
4347 | structure. */ |
4348 | |
4349 | bool |
4350 | direct_internal_fn_supported_p (internal_fn fn, tree_pair types, |
4351 | optimization_type opt_type) |
4352 | { |
4353 | if (!type_pair_strictly_matches_mode_p (type_pair: types)) |
4354 | return false; |
4355 | |
4356 | switch (fn) |
4357 | { |
4358 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) \ |
4359 | case IFN_##CODE: break; |
4360 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4361 | case IFN_##CODE: \ |
4362 | return direct_##TYPE##_optab_supported_p (OPTAB##_optab, types, \ |
4363 | opt_type); |
4364 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4365 | UNSIGNED_OPTAB, TYPE) \ |
4366 | case IFN_##CODE: \ |
4367 | { \ |
4368 | optab which_optab = (TYPE_UNSIGNED (types.SELECTOR) \ |
4369 | ? UNSIGNED_OPTAB ## _optab \ |
4370 | : SIGNED_OPTAB ## _optab); \ |
4371 | return direct_##TYPE##_optab_supported_p (which_optab, types, \ |
4372 | opt_type); \ |
4373 | } |
4374 | #include "internal-fn.def" |
4375 | |
4376 | case IFN_LAST: |
4377 | break; |
4378 | } |
4379 | gcc_unreachable (); |
4380 | } |
4381 | |
4382 | /* Return true if FN is supported for type TYPE when the optimization |
4383 | type is OPT_TYPE. The caller knows that the "type0" and "type1" |
4384 | fields of FN's direct_internal_fn_info structure are the same. */ |
4385 | |
4386 | bool |
4387 | direct_internal_fn_supported_p (internal_fn fn, tree type, |
4388 | optimization_type opt_type) |
4389 | { |
4390 | const direct_internal_fn_info &info = direct_internal_fn (fn); |
4391 | gcc_checking_assert (info.type0 == info.type1); |
4392 | return direct_internal_fn_supported_p (fn, types: tree_pair (type, type), opt_type); |
4393 | } |
4394 | |
4395 | /* Return true if the STMT is supported when the optimization type is OPT_TYPE, |
4396 | given that STMT is a call to a direct internal function. */ |
4397 | |
4398 | bool |
4399 | direct_internal_fn_supported_p (gcall *stmt, optimization_type opt_type) |
4400 | { |
4401 | internal_fn fn = gimple_call_internal_fn (gs: stmt); |
4402 | tree_pair types = direct_internal_fn_types (fn, call: stmt); |
4403 | return direct_internal_fn_supported_p (fn, types, opt_type); |
4404 | } |
4405 | |
4406 | /* Return true if FN is a binary operation and if FN is commutative. */ |
4407 | |
4408 | bool |
4409 | commutative_binary_fn_p (internal_fn fn) |
4410 | { |
4411 | switch (fn) |
4412 | { |
4413 | case IFN_AVG_FLOOR: |
4414 | case IFN_AVG_CEIL: |
4415 | case IFN_MULH: |
4416 | case IFN_MULHS: |
4417 | case IFN_MULHRS: |
4418 | case IFN_FMIN: |
4419 | case IFN_FMAX: |
4420 | case IFN_COMPLEX_MUL: |
4421 | case IFN_UBSAN_CHECK_ADD: |
4422 | case IFN_UBSAN_CHECK_MUL: |
4423 | case IFN_ADD_OVERFLOW: |
4424 | case IFN_MUL_OVERFLOW: |
4425 | case IFN_SAT_ADD: |
4426 | case IFN_VEC_WIDEN_PLUS: |
4427 | case IFN_VEC_WIDEN_PLUS_LO: |
4428 | case IFN_VEC_WIDEN_PLUS_HI: |
4429 | case IFN_VEC_WIDEN_PLUS_EVEN: |
4430 | case IFN_VEC_WIDEN_PLUS_ODD: |
4431 | return true; |
4432 | |
4433 | default: |
4434 | return false; |
4435 | } |
4436 | } |
4437 | |
4438 | /* Return true if FN is a ternary operation and if its first two arguments |
4439 | are commutative. */ |
4440 | |
4441 | bool |
4442 | commutative_ternary_fn_p (internal_fn fn) |
4443 | { |
4444 | switch (fn) |
4445 | { |
4446 | case IFN_FMA: |
4447 | case IFN_FMS: |
4448 | case IFN_FNMA: |
4449 | case IFN_FNMS: |
4450 | case IFN_UADDC: |
4451 | return true; |
4452 | |
4453 | default: |
4454 | return false; |
4455 | } |
4456 | } |
4457 | |
4458 | /* Return true if FN is an associative binary operation. */ |
4459 | |
4460 | bool |
4461 | associative_binary_fn_p (internal_fn fn) |
4462 | { |
4463 | switch (fn) |
4464 | { |
4465 | case IFN_FMIN: |
4466 | case IFN_FMAX: |
4467 | return true; |
4468 | |
4469 | default: |
4470 | return false; |
4471 | } |
4472 | } |
4473 | |
4474 | /* If FN is commutative in two consecutive arguments, return the |
4475 | index of the first, otherwise return -1. */ |
4476 | |
4477 | int |
4478 | first_commutative_argument (internal_fn fn) |
4479 | { |
4480 | switch (fn) |
4481 | { |
4482 | case IFN_COND_ADD: |
4483 | case IFN_COND_MUL: |
4484 | case IFN_COND_MIN: |
4485 | case IFN_COND_MAX: |
4486 | case IFN_COND_FMIN: |
4487 | case IFN_COND_FMAX: |
4488 | case IFN_COND_AND: |
4489 | case IFN_COND_IOR: |
4490 | case IFN_COND_XOR: |
4491 | case IFN_COND_FMA: |
4492 | case IFN_COND_FMS: |
4493 | case IFN_COND_FNMA: |
4494 | case IFN_COND_FNMS: |
4495 | case IFN_COND_LEN_ADD: |
4496 | case IFN_COND_LEN_MUL: |
4497 | case IFN_COND_LEN_MIN: |
4498 | case IFN_COND_LEN_MAX: |
4499 | case IFN_COND_LEN_FMIN: |
4500 | case IFN_COND_LEN_FMAX: |
4501 | case IFN_COND_LEN_AND: |
4502 | case IFN_COND_LEN_IOR: |
4503 | case IFN_COND_LEN_XOR: |
4504 | case IFN_COND_LEN_FMA: |
4505 | case IFN_COND_LEN_FMS: |
4506 | case IFN_COND_LEN_FNMA: |
4507 | case IFN_COND_LEN_FNMS: |
4508 | return 1; |
4509 | |
4510 | default: |
4511 | if (commutative_binary_fn_p (fn) |
4512 | || commutative_ternary_fn_p (fn)) |
4513 | return 0; |
4514 | return -1; |
4515 | } |
4516 | } |
4517 | |
4518 | /* Return true if this CODE describes an internal_fn that returns a vector with |
4519 | elements twice as wide as the element size of the input vectors. */ |
4520 | |
4521 | bool |
4522 | widening_fn_p (code_helper code) |
4523 | { |
4524 | if (!code.is_fn_code ()) |
4525 | return false; |
4526 | |
4527 | if (!internal_fn_p (code: (combined_fn) code)) |
4528 | return false; |
4529 | |
4530 | internal_fn fn = as_internal_fn (code: (combined_fn) code); |
4531 | switch (fn) |
4532 | { |
4533 | #define DEF_INTERNAL_WIDENING_OPTAB_FN(NAME, F, S, SO, UO, T) \ |
4534 | case IFN_##NAME: \ |
4535 | case IFN_##NAME##_HI: \ |
4536 | case IFN_##NAME##_LO: \ |
4537 | case IFN_##NAME##_EVEN: \ |
4538 | case IFN_##NAME##_ODD: \ |
4539 | return true; |
4540 | #include "internal-fn.def" |
4541 | |
4542 | default: |
4543 | return false; |
4544 | } |
4545 | } |
4546 | |
4547 | /* Return true if IFN_SET_EDOM is supported. */ |
4548 | |
4549 | bool |
4550 | set_edom_supported_p (void) |
4551 | { |
4552 | #ifdef TARGET_EDOM |
4553 | return true; |
4554 | #else |
4555 | return false; |
4556 | #endif |
4557 | } |
4558 | |
4559 | #define DEF_INTERNAL_OPTAB_FN(CODE, FLAGS, OPTAB, TYPE) \ |
4560 | static void \ |
4561 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
4562 | { \ |
4563 | expand_##TYPE##_optab_fn (fn, stmt, OPTAB##_optab); \ |
4564 | } |
4565 | #define DEF_INTERNAL_INT_EXT_FN(CODE, FLAGS, OPTAB, TYPE) |
4566 | #define DEF_INTERNAL_SIGNED_OPTAB_FN(CODE, FLAGS, SELECTOR, SIGNED_OPTAB, \ |
4567 | UNSIGNED_OPTAB, TYPE) \ |
4568 | static void \ |
4569 | expand_##CODE (internal_fn fn, gcall *stmt) \ |
4570 | { \ |
4571 | tree_pair types = direct_internal_fn_types (fn, stmt); \ |
4572 | optab which_optab = direct_internal_fn_optab (fn, types); \ |
4573 | expand_##TYPE##_optab_fn (fn, stmt, which_optab); \ |
4574 | } |
4575 | #include "internal-fn.def" |
4576 | |
4577 | /* Routines to expand each internal function, indexed by function number. |
4578 | Each routine has the prototype: |
4579 | |
4580 | expand_<NAME> (gcall *stmt) |
4581 | |
4582 | where STMT is the statement that performs the call. */ |
4583 | static void (*const internal_fn_expanders[]) (internal_fn, gcall *) = { |
4584 | |
4585 | #define DEF_INTERNAL_FN(CODE, FLAGS, FNSPEC) expand_##CODE, |
4586 | #include "internal-fn.def" |
4587 | 0 |
4588 | }; |
4589 | |
4590 | /* Invoke T(CODE, SUFFIX) for each conditional function IFN_COND_##SUFFIX |
4591 | that maps to a tree code CODE. There is also an IFN_COND_LEN_##SUFFIX |
4592 | for each such IFN_COND_##SUFFIX. */ |
4593 | #define FOR_EACH_CODE_MAPPING(T) \ |
4594 | T (PLUS_EXPR, ADD) \ |
4595 | T (MINUS_EXPR, SUB) \ |
4596 | T (MULT_EXPR, MUL) \ |
4597 | T (TRUNC_DIV_EXPR, DIV) \ |
4598 | T (TRUNC_MOD_EXPR, MOD) \ |
4599 | T (RDIV_EXPR, RDIV) \ |
4600 | T (MIN_EXPR, MIN) \ |
4601 | T (MAX_EXPR, MAX) \ |
4602 | T (BIT_AND_EXPR, AND) \ |
4603 | T (BIT_IOR_EXPR, IOR) \ |
4604 | T (BIT_XOR_EXPR, XOR) \ |
4605 | T (LSHIFT_EXPR, SHL) \ |
4606 | T (RSHIFT_EXPR, SHR) \ |
4607 | T (NEGATE_EXPR, NEG) |
4608 | |
4609 | /* Return a function that only performs CODE when a certain condition is met |
4610 | and that uses a given fallback value otherwise. For example, if CODE is |
4611 | a binary operation associated with conditional function FN: |
4612 | |
4613 | LHS = FN (COND, A, B, ELSE) |
4614 | |
4615 | is equivalent to the C expression: |
4616 | |
4617 | LHS = COND ? A CODE B : ELSE; |
4618 | |
4619 | operating elementwise if the operands are vectors. |
4620 | |
4621 | Return IFN_LAST if no such function exists. */ |
4622 | |
4623 | internal_fn |
4624 | get_conditional_internal_fn (tree_code code) |
4625 | { |
4626 | switch (code) |
4627 | { |
4628 | #define CASE(CODE, IFN) case CODE: return IFN_COND_##IFN; |
4629 | FOR_EACH_CODE_MAPPING(CASE) |
4630 | #undef CASE |
4631 | default: |
4632 | return IFN_LAST; |
4633 | } |
4634 | } |
4635 | |
4636 | /* If IFN implements the conditional form of a tree code, return that |
4637 | tree code, otherwise return ERROR_MARK. */ |
4638 | |
4639 | tree_code |
4640 | conditional_internal_fn_code (internal_fn ifn) |
4641 | { |
4642 | switch (ifn) |
4643 | { |
4644 | #define CASE(CODE, IFN) \ |
4645 | case IFN_COND_##IFN: \ |
4646 | case IFN_COND_LEN_##IFN: \ |
4647 | return CODE; |
4648 | FOR_EACH_CODE_MAPPING (CASE) |
4649 | #undef CASE |
4650 | default: |
4651 | return ERROR_MARK; |
4652 | } |
4653 | } |
4654 | |
4655 | /* Like get_conditional_internal_fn, but return a function that |
4656 | additionally restricts the operation to the leading elements |
4657 | of a vector. The number of elements to process is given by a length |
4658 | and bias pair, as for IFN_LOAD_LEN. The values of the remaining |
4659 | elements are taken from the fallback ("else") argument. |
4660 | |
4661 | For example, if CODE is a binary operation associated with FN: |
4662 | |
4663 | LHS = FN (COND, A, B, ELSE, LEN, BIAS) |
4664 | |
4665 | is equivalent to the C code: |
4666 | |
4667 | for (int i = 0; i < NUNITS; i++) |
4668 | { |
4669 | if (i < LEN + BIAS && COND[i]) |
4670 | LHS[i] = A[i] CODE B[i]; |
4671 | else |
4672 | LHS[i] = ELSE[i]; |
4673 | } |
4674 | */ |
4675 | |
4676 | internal_fn |
4677 | get_conditional_len_internal_fn (tree_code code) |
4678 | { |
4679 | switch (code) |
4680 | { |
4681 | #define CASE(CODE, IFN) case CODE: return IFN_COND_LEN_##IFN; |
4682 | FOR_EACH_CODE_MAPPING(CASE) |
4683 | #undef CASE |
4684 | default: |
4685 | return IFN_LAST; |
4686 | } |
4687 | } |
4688 | |
4689 | /* Invoke T(IFN) for each internal function IFN that also has an |
4690 | IFN_COND_* form. */ |
4691 | #define FOR_EACH_COND_FN_PAIR(T) \ |
4692 | T (FMAX) \ |
4693 | T (FMIN) \ |
4694 | T (FMA) \ |
4695 | T (FMS) \ |
4696 | T (FNMA) \ |
4697 | T (FNMS) |
4698 | |
4699 | /* Return a function that only performs internal function FN when a |
4700 | certain condition is met and that uses a given fallback value otherwise. |
4701 | In other words, the returned function FN' is such that: |
4702 | |
4703 | LHS = FN' (COND, A1, ... An, ELSE) |
4704 | |
4705 | is equivalent to the C expression: |
4706 | |
4707 | LHS = COND ? FN (A1, ..., An) : ELSE; |
4708 | |
4709 | operating elementwise if the operands are vectors. |
4710 | |
4711 | Return IFN_LAST if no such function exists. */ |
4712 | |
4713 | internal_fn |
4714 | get_conditional_internal_fn (internal_fn fn) |
4715 | { |
4716 | switch (fn) |
4717 | { |
4718 | #define CASE(NAME) case IFN_##NAME: return IFN_COND_##NAME; |
4719 | FOR_EACH_COND_FN_PAIR(CASE) |
4720 | #undef CASE |
4721 | default: |
4722 | return IFN_LAST; |
4723 | } |
4724 | } |
4725 | |
4726 | /* If there exists an internal function like IFN that operates on vectors, |
4727 | but with additional length and bias parameters, return the internal_fn |
4728 | for that function, otherwise return IFN_LAST. */ |
4729 | internal_fn |
4730 | get_len_internal_fn (internal_fn fn) |
4731 | { |
4732 | switch (fn) |
4733 | { |
4734 | #define DEF_INTERNAL_COND_FN(NAME, ...) \ |
4735 | case IFN_COND_##NAME: \ |
4736 | return IFN_COND_LEN_##NAME; |
4737 | #define DEF_INTERNAL_SIGNED_COND_FN(NAME, ...) \ |
4738 | case IFN_COND_##NAME: \ |
4739 | return IFN_COND_LEN_##NAME; |
4740 | #include "internal-fn.def" |
4741 | default: |
4742 | break; |
4743 | } |
4744 | |
4745 | switch (fn) |
4746 | { |
4747 | case IFN_MASK_LOAD: |
4748 | return IFN_MASK_LEN_LOAD; |
4749 | case IFN_MASK_LOAD_LANES: |
4750 | return IFN_MASK_LEN_LOAD_LANES; |
4751 | case IFN_MASK_GATHER_LOAD: |
4752 | return IFN_MASK_LEN_GATHER_LOAD; |
4753 | default: |
4754 | return IFN_LAST; |
4755 | } |
4756 | } |
4757 | |
4758 | /* If IFN implements the conditional form of an unconditional internal |
4759 | function, return that unconditional function, otherwise return IFN_LAST. */ |
4760 | |
4761 | internal_fn |
4762 | get_unconditional_internal_fn (internal_fn ifn) |
4763 | { |
4764 | switch (ifn) |
4765 | { |
4766 | #define CASE(NAME) \ |
4767 | case IFN_COND_##NAME: \ |
4768 | case IFN_COND_LEN_##NAME: \ |
4769 | return IFN_##NAME; |
4770 | FOR_EACH_COND_FN_PAIR (CASE) |
4771 | #undef CASE |
4772 | default: |
4773 | return IFN_LAST; |
4774 | } |
4775 | } |
4776 | |
4777 | /* Return true if STMT can be interpreted as a conditional tree code |
4778 | operation of the form: |
4779 | |
4780 | LHS = COND ? OP (RHS1, ...) : ELSE; |
4781 | |
4782 | operating elementwise if the operands are vectors. This includes |
4783 | the case of an all-true COND, so that the operation always happens. |
4784 | |
4785 | There is an alternative approach to interpret the STMT when the operands |
4786 | are vectors which is the operation predicated by both conditional mask |
4787 | and loop control length, the equivalent C code: |
4788 | |
4789 | for (int i = 0; i < NUNTIS; i++) |
4790 | { |
4791 | if (i < LEN + BIAS && COND[i]) |
4792 | LHS[i] = A[i] CODE B[i]; |
4793 | else |
4794 | LHS[i] = ELSE[i]; |
4795 | } |
4796 | |
4797 | When returning true, set: |
4798 | |
4799 | - *COND_OUT to the condition COND, or to NULL_TREE if the condition |
4800 | is known to be all-true |
4801 | - *CODE_OUT to the tree code |
4802 | - OPS[I] to operand I of *CODE_OUT |
4803 | - *ELSE_OUT to the fallback value ELSE, or to NULL_TREE if the |
4804 | condition is known to be all true. |
4805 | - *LEN to the len argument if it COND_LEN_* operations or to NULL_TREE. |
4806 | - *BIAS to the bias argument if it COND_LEN_* operations or to NULL_TREE. */ |
4807 | |
4808 | bool |
4809 | can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out, |
4810 | tree_code *code_out, |
4811 | tree (&ops)[3], tree *else_out, |
4812 | tree *len, tree *bias) |
4813 | { |
4814 | *len = NULL_TREE; |
4815 | *bias = NULL_TREE; |
4816 | if (gassign *assign = dyn_cast <gassign *> (p: stmt)) |
4817 | { |
4818 | *cond_out = NULL_TREE; |
4819 | *code_out = gimple_assign_rhs_code (gs: assign); |
4820 | ops[0] = gimple_assign_rhs1 (gs: assign); |
4821 | ops[1] = gimple_assign_rhs2 (gs: assign); |
4822 | ops[2] = gimple_assign_rhs3 (gs: assign); |
4823 | *else_out = NULL_TREE; |
4824 | return true; |
4825 | } |
4826 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
4827 | if (gimple_call_internal_p (gs: call)) |
4828 | { |
4829 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
4830 | tree_code code = conditional_internal_fn_code (ifn); |
4831 | int len_index = internal_fn_len_index (ifn); |
4832 | int cond_nargs = len_index >= 0 ? 4 : 2; |
4833 | if (code != ERROR_MARK) |
4834 | { |
4835 | *cond_out = gimple_call_arg (gs: call, index: 0); |
4836 | *code_out = code; |
4837 | unsigned int nops = gimple_call_num_args (gs: call) - cond_nargs; |
4838 | for (unsigned int i = 0; i < 3; ++i) |
4839 | ops[i] = i < nops ? gimple_call_arg (gs: call, index: i + 1) : NULL_TREE; |
4840 | *else_out = gimple_call_arg (gs: call, index: nops + 1); |
4841 | if (len_index < 0) |
4842 | { |
4843 | if (integer_truep (*cond_out)) |
4844 | { |
4845 | *cond_out = NULL_TREE; |
4846 | *else_out = NULL_TREE; |
4847 | } |
4848 | } |
4849 | else |
4850 | { |
4851 | *len = gimple_call_arg (gs: call, index: len_index); |
4852 | *bias = gimple_call_arg (gs: call, index: len_index + 1); |
4853 | } |
4854 | return true; |
4855 | } |
4856 | } |
4857 | return false; |
4858 | } |
4859 | |
4860 | /* Return true if IFN is some form of load from memory. */ |
4861 | |
4862 | bool |
4863 | internal_load_fn_p (internal_fn fn) |
4864 | { |
4865 | switch (fn) |
4866 | { |
4867 | case IFN_MASK_LOAD: |
4868 | case IFN_LOAD_LANES: |
4869 | case IFN_MASK_LOAD_LANES: |
4870 | case IFN_MASK_LEN_LOAD_LANES: |
4871 | case IFN_GATHER_LOAD: |
4872 | case IFN_MASK_GATHER_LOAD: |
4873 | case IFN_MASK_LEN_GATHER_LOAD: |
4874 | case IFN_LEN_LOAD: |
4875 | case IFN_MASK_LEN_LOAD: |
4876 | return true; |
4877 | |
4878 | default: |
4879 | return false; |
4880 | } |
4881 | } |
4882 | |
4883 | /* Return true if IFN is some form of store to memory. */ |
4884 | |
4885 | bool |
4886 | internal_store_fn_p (internal_fn fn) |
4887 | { |
4888 | switch (fn) |
4889 | { |
4890 | case IFN_MASK_STORE: |
4891 | case IFN_STORE_LANES: |
4892 | case IFN_MASK_STORE_LANES: |
4893 | case IFN_MASK_LEN_STORE_LANES: |
4894 | case IFN_SCATTER_STORE: |
4895 | case IFN_MASK_SCATTER_STORE: |
4896 | case IFN_MASK_LEN_SCATTER_STORE: |
4897 | case IFN_LEN_STORE: |
4898 | case IFN_MASK_LEN_STORE: |
4899 | return true; |
4900 | |
4901 | default: |
4902 | return false; |
4903 | } |
4904 | } |
4905 | |
4906 | /* Return true if IFN is some form of gather load or scatter store. */ |
4907 | |
4908 | bool |
4909 | internal_gather_scatter_fn_p (internal_fn fn) |
4910 | { |
4911 | switch (fn) |
4912 | { |
4913 | case IFN_GATHER_LOAD: |
4914 | case IFN_MASK_GATHER_LOAD: |
4915 | case IFN_MASK_LEN_GATHER_LOAD: |
4916 | case IFN_SCATTER_STORE: |
4917 | case IFN_MASK_SCATTER_STORE: |
4918 | case IFN_MASK_LEN_SCATTER_STORE: |
4919 | return true; |
4920 | |
4921 | default: |
4922 | return false; |
4923 | } |
4924 | } |
4925 | |
4926 | /* If FN takes a vector len argument, return the index of that argument, |
4927 | otherwise return -1. */ |
4928 | |
4929 | int |
4930 | internal_fn_len_index (internal_fn fn) |
4931 | { |
4932 | switch (fn) |
4933 | { |
4934 | case IFN_LEN_LOAD: |
4935 | case IFN_LEN_STORE: |
4936 | return 2; |
4937 | |
4938 | case IFN_MASK_LEN_SCATTER_STORE: |
4939 | case IFN_MASK_LEN_STRIDED_LOAD: |
4940 | return 5; |
4941 | |
4942 | case IFN_MASK_LEN_GATHER_LOAD: |
4943 | return 6; |
4944 | |
4945 | case IFN_COND_LEN_FMA: |
4946 | case IFN_COND_LEN_FMS: |
4947 | case IFN_COND_LEN_FNMA: |
4948 | case IFN_COND_LEN_FNMS: |
4949 | return 5; |
4950 | |
4951 | case IFN_COND_LEN_ADD: |
4952 | case IFN_COND_LEN_SUB: |
4953 | case IFN_COND_LEN_MUL: |
4954 | case IFN_COND_LEN_DIV: |
4955 | case IFN_COND_LEN_MOD: |
4956 | case IFN_COND_LEN_RDIV: |
4957 | case IFN_COND_LEN_MIN: |
4958 | case IFN_COND_LEN_MAX: |
4959 | case IFN_COND_LEN_FMIN: |
4960 | case IFN_COND_LEN_FMAX: |
4961 | case IFN_COND_LEN_AND: |
4962 | case IFN_COND_LEN_IOR: |
4963 | case IFN_COND_LEN_XOR: |
4964 | case IFN_COND_LEN_SHL: |
4965 | case IFN_COND_LEN_SHR: |
4966 | case IFN_MASK_LEN_STRIDED_STORE: |
4967 | return 4; |
4968 | |
4969 | case IFN_COND_LEN_NEG: |
4970 | case IFN_MASK_LEN_STORE: |
4971 | case IFN_MASK_LEN_STORE_LANES: |
4972 | case IFN_VCOND_MASK_LEN: |
4973 | return 3; |
4974 | |
4975 | case IFN_MASK_LEN_LOAD: |
4976 | case IFN_MASK_LEN_LOAD_LANES: |
4977 | return 4; |
4978 | |
4979 | default: |
4980 | return -1; |
4981 | } |
4982 | } |
4983 | |
4984 | /* If FN is an IFN_COND_* or IFN_COND_LEN_* function, return the index of the |
4985 | argument that is used when the condition is false. Return -1 otherwise. */ |
4986 | |
4987 | int |
4988 | internal_fn_else_index (internal_fn fn) |
4989 | { |
4990 | switch (fn) |
4991 | { |
4992 | case IFN_COND_NEG: |
4993 | case IFN_COND_NOT: |
4994 | case IFN_COND_LEN_NEG: |
4995 | case IFN_COND_LEN_NOT: |
4996 | return 2; |
4997 | |
4998 | case IFN_COND_ADD: |
4999 | case IFN_COND_SUB: |
5000 | case IFN_COND_MUL: |
5001 | case IFN_COND_DIV: |
5002 | case IFN_COND_MOD: |
5003 | case IFN_COND_MIN: |
5004 | case IFN_COND_MAX: |
5005 | case IFN_COND_FMIN: |
5006 | case IFN_COND_FMAX: |
5007 | case IFN_COND_AND: |
5008 | case IFN_COND_IOR: |
5009 | case IFN_COND_XOR: |
5010 | case IFN_COND_SHL: |
5011 | case IFN_COND_SHR: |
5012 | case IFN_COND_LEN_ADD: |
5013 | case IFN_COND_LEN_SUB: |
5014 | case IFN_COND_LEN_MUL: |
5015 | case IFN_COND_LEN_DIV: |
5016 | case IFN_COND_LEN_MOD: |
5017 | case IFN_COND_LEN_MIN: |
5018 | case IFN_COND_LEN_MAX: |
5019 | case IFN_COND_LEN_FMIN: |
5020 | case IFN_COND_LEN_FMAX: |
5021 | case IFN_COND_LEN_AND: |
5022 | case IFN_COND_LEN_IOR: |
5023 | case IFN_COND_LEN_XOR: |
5024 | case IFN_COND_LEN_SHL: |
5025 | case IFN_COND_LEN_SHR: |
5026 | return 3; |
5027 | |
5028 | case IFN_MASK_LOAD: |
5029 | case IFN_MASK_LEN_LOAD: |
5030 | case IFN_MASK_LOAD_LANES: |
5031 | case IFN_MASK_LEN_LOAD_LANES: |
5032 | return 3; |
5033 | |
5034 | case IFN_COND_FMA: |
5035 | case IFN_COND_FMS: |
5036 | case IFN_COND_FNMA: |
5037 | case IFN_COND_FNMS: |
5038 | case IFN_COND_LEN_FMA: |
5039 | case IFN_COND_LEN_FMS: |
5040 | case IFN_COND_LEN_FNMA: |
5041 | case IFN_COND_LEN_FNMS: |
5042 | case IFN_MASK_LEN_STRIDED_LOAD: |
5043 | return 4; |
5044 | |
5045 | case IFN_MASK_GATHER_LOAD: |
5046 | case IFN_MASK_LEN_GATHER_LOAD: |
5047 | return 5; |
5048 | |
5049 | default: |
5050 | return -1; |
5051 | } |
5052 | |
5053 | return -1; |
5054 | } |
5055 | |
5056 | /* If FN takes a vector mask argument, return the index of that argument, |
5057 | otherwise return -1. */ |
5058 | |
5059 | int |
5060 | internal_fn_mask_index (internal_fn fn) |
5061 | { |
5062 | switch (fn) |
5063 | { |
5064 | case IFN_MASK_LOAD: |
5065 | case IFN_MASK_LOAD_LANES: |
5066 | case IFN_MASK_LEN_LOAD_LANES: |
5067 | case IFN_MASK_STORE: |
5068 | case IFN_MASK_STORE_LANES: |
5069 | case IFN_MASK_LEN_STORE_LANES: |
5070 | case IFN_MASK_LEN_LOAD: |
5071 | case IFN_MASK_LEN_STORE: |
5072 | return 2; |
5073 | |
5074 | case IFN_MASK_LEN_STRIDED_LOAD: |
5075 | case IFN_MASK_LEN_STRIDED_STORE: |
5076 | return 3; |
5077 | |
5078 | case IFN_MASK_GATHER_LOAD: |
5079 | case IFN_MASK_SCATTER_STORE: |
5080 | case IFN_MASK_LEN_GATHER_LOAD: |
5081 | case IFN_MASK_LEN_SCATTER_STORE: |
5082 | return 4; |
5083 | |
5084 | case IFN_VCOND_MASK: |
5085 | case IFN_VCOND_MASK_LEN: |
5086 | return 0; |
5087 | |
5088 | default: |
5089 | return (conditional_internal_fn_code (ifn: fn) != ERROR_MARK |
5090 | || get_unconditional_internal_fn (ifn: fn) != IFN_LAST ? 0 : -1); |
5091 | } |
5092 | } |
5093 | |
5094 | /* If FN takes a value that should be stored to memory, return the index |
5095 | of that argument, otherwise return -1. */ |
5096 | |
5097 | int |
5098 | internal_fn_stored_value_index (internal_fn fn) |
5099 | { |
5100 | switch (fn) |
5101 | { |
5102 | case IFN_MASK_LEN_STRIDED_STORE: |
5103 | return 2; |
5104 | |
5105 | case IFN_MASK_STORE: |
5106 | case IFN_MASK_STORE_LANES: |
5107 | case IFN_SCATTER_STORE: |
5108 | case IFN_MASK_SCATTER_STORE: |
5109 | case IFN_MASK_LEN_SCATTER_STORE: |
5110 | return 3; |
5111 | |
5112 | case IFN_LEN_STORE: |
5113 | return 4; |
5114 | |
5115 | case IFN_MASK_LEN_STORE: |
5116 | case IFN_MASK_LEN_STORE_LANES: |
5117 | return 5; |
5118 | |
5119 | default: |
5120 | return -1; |
5121 | } |
5122 | } |
5123 | |
5124 | |
5125 | /* Store all supported else values for the optab referred to by ICODE |
5126 | in ELSE_VALS. The index of the else operand must be specified in |
5127 | ELSE_INDEX. */ |
5128 | |
5129 | void |
5130 | get_supported_else_vals (enum insn_code icode, unsigned else_index, |
5131 | vec<int> &else_vals) |
5132 | { |
5133 | const struct insn_data_d *data = &insn_data[icode]; |
5134 | if ((char)else_index >= data->n_operands) |
5135 | return; |
5136 | |
5137 | machine_mode else_mode = data->operand[else_index].mode; |
5138 | |
5139 | else_vals.truncate (size: 0); |
5140 | |
5141 | /* For now we only support else values of 0, -1, and "undefined". */ |
5142 | if (insn_operand_matches (icode, opno: else_index, CONST0_RTX (else_mode))) |
5143 | else_vals.safe_push (MASK_LOAD_ELSE_ZERO); |
5144 | |
5145 | if (insn_operand_matches (icode, opno: else_index, gen_rtx_SCRATCH (else_mode))) |
5146 | else_vals.safe_push (MASK_LOAD_ELSE_UNDEFINED); |
5147 | |
5148 | if (GET_MODE_CLASS (else_mode) == MODE_VECTOR_INT |
5149 | && insn_operand_matches (icode, opno: else_index, CONSTM1_RTX (else_mode))) |
5150 | else_vals.safe_push (MASK_LOAD_ELSE_M1); |
5151 | } |
5152 | |
5153 | /* Return true if the else value ELSE_VAL (one of MASK_LOAD_ELSE_ZERO, |
5154 | MASK_LOAD_ELSE_M1, and MASK_LOAD_ELSE_UNDEFINED) is valid fo the optab |
5155 | referred to by ICODE. The index of the else operand must be specified |
5156 | in ELSE_INDEX. */ |
5157 | |
5158 | bool |
5159 | supported_else_val_p (enum insn_code icode, unsigned else_index, int else_val) |
5160 | { |
5161 | if (else_val != MASK_LOAD_ELSE_ZERO && else_val != MASK_LOAD_ELSE_M1 |
5162 | && else_val != MASK_LOAD_ELSE_UNDEFINED) |
5163 | gcc_unreachable (); |
5164 | |
5165 | auto_vec<int> else_vals; |
5166 | get_supported_else_vals (icode, else_index, else_vals); |
5167 | return else_vals.contains (search: else_val); |
5168 | } |
5169 | |
5170 | /* Return true if the target supports gather load or scatter store function |
5171 | IFN. For loads, VECTOR_TYPE is the vector type of the load result, |
5172 | while for stores it is the vector type of the stored data argument. |
5173 | MEMORY_ELEMENT_TYPE is the type of the memory elements being loaded |
5174 | or stored. OFFSET_VECTOR_TYPE is the vector type that holds the |
5175 | offset from the shared base address of each loaded or stored element. |
5176 | SCALE is the amount by which these offsets should be multiplied |
5177 | *after* they have been extended to address width. |
5178 | If the target supports the gather load the supported else values |
5179 | will be added to the vector ELSVAL points to if it is nonzero. */ |
5180 | |
5181 | bool |
5182 | internal_gather_scatter_fn_supported_p (internal_fn ifn, tree vector_type, |
5183 | tree memory_element_type, |
5184 | tree offset_vector_type, int scale, |
5185 | vec<int> *elsvals) |
5186 | { |
5187 | if (!tree_int_cst_equal (TYPE_SIZE (TREE_TYPE (vector_type)), |
5188 | TYPE_SIZE (memory_element_type))) |
5189 | return false; |
5190 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vector_type), |
5191 | b: TYPE_VECTOR_SUBPARTS (node: offset_vector_type))) |
5192 | return false; |
5193 | optab optab = direct_internal_fn_optab (fn: ifn); |
5194 | insn_code icode = convert_optab_handler (op: optab, TYPE_MODE (vector_type), |
5195 | TYPE_MODE (offset_vector_type)); |
5196 | int output_ops = internal_load_fn_p (fn: ifn) ? 1 : 0; |
5197 | bool unsigned_p = TYPE_UNSIGNED (TREE_TYPE (offset_vector_type)); |
5198 | bool ok = icode != CODE_FOR_nothing |
5199 | && insn_operand_matches (icode, opno: 2 + output_ops, GEN_INT (unsigned_p)) |
5200 | && insn_operand_matches (icode, opno: 3 + output_ops, GEN_INT (scale)); |
5201 | |
5202 | /* For gather the optab's operand indices do not match the IFN's because |
5203 | the latter does not have the extension operand (operand 3). It is |
5204 | implicitly added during expansion so we use the IFN's else index + 1. |
5205 | */ |
5206 | if (ok && elsvals) |
5207 | get_supported_else_vals |
5208 | (icode, else_index: internal_fn_else_index (fn: IFN_MASK_GATHER_LOAD) + 1, else_vals&: *elsvals); |
5209 | |
5210 | return ok; |
5211 | } |
5212 | |
5213 | /* Return true if the target supports IFN_CHECK_{RAW,WAR}_PTRS function IFN |
5214 | for pointers of type TYPE when the accesses have LENGTH bytes and their |
5215 | common byte alignment is ALIGN. */ |
5216 | |
5217 | bool |
5218 | internal_check_ptrs_fn_supported_p (internal_fn ifn, tree type, |
5219 | poly_uint64 length, unsigned int align) |
5220 | { |
5221 | machine_mode mode = TYPE_MODE (type); |
5222 | optab optab = direct_internal_fn_optab (fn: ifn); |
5223 | insn_code icode = direct_optab_handler (op: optab, mode); |
5224 | if (icode == CODE_FOR_nothing) |
5225 | return false; |
5226 | rtx length_rtx = immed_wide_int_const (length, mode); |
5227 | return (insn_operand_matches (icode, opno: 3, operand: length_rtx) |
5228 | && insn_operand_matches (icode, opno: 4, GEN_INT (align))); |
5229 | } |
5230 | |
5231 | /* Return the supported bias for IFN which is either IFN_{LEN_,MASK_LEN_,}LOAD |
5232 | or IFN_{LEN_,MASK_LEN_,}STORE. For now we only support the biases of 0 and |
5233 | -1 (in case 0 is not an allowable length for {len_,mask_len_}load or |
5234 | {len_,mask_len_}store). If none of the biases match what the backend |
5235 | provides, return VECT_PARTIAL_BIAS_UNSUPPORTED. */ |
5236 | |
5237 | signed char |
5238 | internal_len_load_store_bias (internal_fn ifn, machine_mode mode) |
5239 | { |
5240 | optab optab = direct_internal_fn_optab (fn: ifn); |
5241 | insn_code icode = direct_optab_handler (op: optab, mode); |
5242 | int bias_no = 3; |
5243 | |
5244 | if (icode == CODE_FOR_nothing) |
5245 | { |
5246 | machine_mode mask_mode; |
5247 | if (!targetm.vectorize.get_mask_mode (mode).exists (mode: &mask_mode)) |
5248 | return VECT_PARTIAL_BIAS_UNSUPPORTED; |
5249 | if (ifn == IFN_LEN_LOAD) |
5250 | { |
5251 | /* Try MASK_LEN_LOAD. */ |
5252 | optab = direct_internal_fn_optab (fn: IFN_MASK_LEN_LOAD); |
5253 | } |
5254 | else |
5255 | { |
5256 | /* Try MASK_LEN_STORE. */ |
5257 | optab = direct_internal_fn_optab (fn: IFN_MASK_LEN_STORE); |
5258 | } |
5259 | icode = convert_optab_handler (op: optab, to_mode: mode, from_mode: mask_mode); |
5260 | bias_no = 4; |
5261 | } |
5262 | |
5263 | if (icode != CODE_FOR_nothing) |
5264 | { |
5265 | /* For now we only support biases of 0 or -1. Try both of them. */ |
5266 | if (insn_operand_matches (icode, opno: bias_no, GEN_INT (0))) |
5267 | return 0; |
5268 | if (insn_operand_matches (icode, opno: bias_no, GEN_INT (-1))) |
5269 | return -1; |
5270 | } |
5271 | |
5272 | return VECT_PARTIAL_BIAS_UNSUPPORTED; |
5273 | } |
5274 | |
5275 | /* Expand STMT as though it were a call to internal function FN. */ |
5276 | |
5277 | void |
5278 | expand_internal_call (internal_fn fn, gcall *stmt) |
5279 | { |
5280 | internal_fn_expanders[fn] (fn, stmt); |
5281 | } |
5282 | |
5283 | /* Expand STMT, which is a call to internal function FN. */ |
5284 | |
5285 | void |
5286 | expand_internal_call (gcall *stmt) |
5287 | { |
5288 | expand_internal_call (fn: gimple_call_internal_fn (gs: stmt), stmt); |
5289 | } |
5290 | |
5291 | /* If TYPE is a vector type, return true if IFN is a direct internal |
5292 | function that is supported for that type. If TYPE is a scalar type, |
5293 | return true if IFN is a direct internal function that is supported for |
5294 | the target's preferred vector version of TYPE. */ |
5295 | |
5296 | bool |
5297 | vectorized_internal_fn_supported_p (internal_fn ifn, tree type) |
5298 | { |
5299 | if (VECTOR_MODE_P (TYPE_MODE (type))) |
5300 | return direct_internal_fn_supported_p (fn: ifn, type, opt_type: OPTIMIZE_FOR_SPEED); |
5301 | |
5302 | scalar_mode smode; |
5303 | if (VECTOR_TYPE_P (type) |
5304 | || !is_a <scalar_mode> (TYPE_MODE (type), result: &smode)) |
5305 | return false; |
5306 | |
5307 | machine_mode vmode = targetm.vectorize.preferred_simd_mode (smode); |
5308 | if (VECTOR_MODE_P (vmode)) |
5309 | { |
5310 | tree vectype = build_vector_type_for_mode (type, vmode); |
5311 | if (direct_internal_fn_supported_p (fn: ifn, type: vectype, opt_type: OPTIMIZE_FOR_SPEED)) |
5312 | return true; |
5313 | } |
5314 | |
5315 | auto_vector_modes vector_modes; |
5316 | targetm.vectorize.autovectorize_vector_modes (&vector_modes, true); |
5317 | for (machine_mode base_mode : vector_modes) |
5318 | if (related_vector_mode (base_mode, smode).exists (mode: &vmode)) |
5319 | { |
5320 | tree vectype = build_vector_type_for_mode (type, vmode); |
5321 | if (direct_internal_fn_supported_p (fn: ifn, type: vectype, opt_type: OPTIMIZE_FOR_SPEED)) |
5322 | return true; |
5323 | } |
5324 | |
5325 | return false; |
5326 | } |
5327 | |
5328 | void |
5329 | expand_SHUFFLEVECTOR (internal_fn, gcall *) |
5330 | { |
5331 | gcc_unreachable (); |
5332 | } |
5333 | |
5334 | void |
5335 | expand_PHI (internal_fn, gcall *) |
5336 | { |
5337 | gcc_unreachable (); |
5338 | } |
5339 | |
5340 | void |
5341 | expand_SPACESHIP (internal_fn, gcall *stmt) |
5342 | { |
5343 | tree lhs = gimple_call_lhs (gs: stmt); |
5344 | tree rhs1 = gimple_call_arg (gs: stmt, index: 0); |
5345 | tree rhs2 = gimple_call_arg (gs: stmt, index: 1); |
5346 | tree rhs3 = gimple_call_arg (gs: stmt, index: 2); |
5347 | tree type = TREE_TYPE (rhs1); |
5348 | |
5349 | do_pending_stack_adjust (); |
5350 | |
5351 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5352 | rtx op1 = expand_normal (exp: rhs1); |
5353 | rtx op2 = expand_normal (exp: rhs2); |
5354 | rtx op3 = expand_normal (exp: rhs3); |
5355 | |
5356 | class expand_operand ops[4]; |
5357 | create_call_lhs_operand (op: &ops[0], lhs_rtx: target, TYPE_MODE (TREE_TYPE (lhs))); |
5358 | create_input_operand (op: &ops[1], value: op1, TYPE_MODE (type)); |
5359 | create_input_operand (op: &ops[2], value: op2, TYPE_MODE (type)); |
5360 | create_input_operand (op: &ops[3], value: op3, TYPE_MODE (TREE_TYPE (rhs3))); |
5361 | insn_code icode = optab_handler (op: spaceship_optab, TYPE_MODE (type)); |
5362 | expand_insn (icode, nops: 4, ops); |
5363 | assign_call_lhs (lhs, lhs_rtx: target, op: &ops[0]); |
5364 | } |
5365 | |
5366 | void |
5367 | expand_ASSUME (internal_fn, gcall *) |
5368 | { |
5369 | } |
5370 | |
5371 | void |
5372 | expand_MASK_CALL (internal_fn, gcall *) |
5373 | { |
5374 | /* This IFN should only exist between ifcvt and vect passes. */ |
5375 | gcc_unreachable (); |
5376 | } |
5377 | |
5378 | void |
5379 | expand_MULBITINT (internal_fn, gcall *stmt) |
5380 | { |
5381 | rtx_mode_t args[6]; |
5382 | for (int i = 0; i < 6; i++) |
5383 | args[i] = rtx_mode_t (expand_normal (exp: gimple_call_arg (gs: stmt, index: i)), |
5384 | (i & 1) ? SImode : ptr_mode); |
5385 | rtx fun = init_one_libfunc ("__mulbitint3"); |
5386 | emit_library_call_value_1 (0, fun, NULL_RTX, LCT_NORMAL, VOIDmode, 6, args); |
5387 | } |
5388 | |
5389 | void |
5390 | expand_DIVMODBITINT (internal_fn, gcall *stmt) |
5391 | { |
5392 | rtx_mode_t args[8]; |
5393 | for (int i = 0; i < 8; i++) |
5394 | args[i] = rtx_mode_t (expand_normal (exp: gimple_call_arg (gs: stmt, index: i)), |
5395 | (i & 1) ? SImode : ptr_mode); |
5396 | rtx fun = init_one_libfunc ("__divmodbitint4"); |
5397 | emit_library_call_value_1 (0, fun, NULL_RTX, LCT_NORMAL, VOIDmode, 8, args); |
5398 | } |
5399 | |
5400 | void |
5401 | expand_FLOATTOBITINT (internal_fn, gcall *stmt) |
5402 | { |
5403 | machine_mode mode = TYPE_MODE (TREE_TYPE (gimple_call_arg (stmt, 2))); |
5404 | rtx arg0 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
5405 | rtx arg1 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
5406 | rtx arg2 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 2)); |
5407 | const char *mname = GET_MODE_NAME (mode); |
5408 | unsigned mname_len = strlen (s: mname); |
5409 | int len = 12 + mname_len; |
5410 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5411 | len += 4; |
5412 | char *libfunc_name = XALLOCAVEC (char, len); |
5413 | char *p = libfunc_name; |
5414 | const char *q; |
5415 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5416 | { |
5417 | #if ENABLE_DECIMAL_BID_FORMAT |
5418 | memcpy (dest: p, src: "__bid_fix", n: 9); |
5419 | #else |
5420 | memcpy (p, "__dpd_fix", 9); |
5421 | #endif |
5422 | p += 9; |
5423 | } |
5424 | else |
5425 | { |
5426 | memcpy (dest: p, src: "__fix", n: 5); |
5427 | p += 5; |
5428 | } |
5429 | for (q = mname; *q; q++) |
5430 | *p++ = TOLOWER (*q); |
5431 | memcpy (dest: p, src: "bitint", n: 7); |
5432 | rtx fun = init_one_libfunc (libfunc_name); |
5433 | emit_library_call (fun, fn_type: LCT_NORMAL, VOIDmode, arg1: arg0, arg1_mode: ptr_mode, arg2: arg1, |
5434 | SImode, arg3: arg2, arg3_mode: mode); |
5435 | } |
5436 | |
5437 | void |
5438 | expand_BITINTTOFLOAT (internal_fn, gcall *stmt) |
5439 | { |
5440 | tree lhs = gimple_call_lhs (gs: stmt); |
5441 | if (!lhs) |
5442 | return; |
5443 | machine_mode mode = TYPE_MODE (TREE_TYPE (lhs)); |
5444 | rtx arg0 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 0)); |
5445 | rtx arg1 = expand_normal (exp: gimple_call_arg (gs: stmt, index: 1)); |
5446 | const char *mname = GET_MODE_NAME (mode); |
5447 | unsigned mname_len = strlen (s: mname); |
5448 | int len = 14 + mname_len; |
5449 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5450 | len += 4; |
5451 | char *libfunc_name = XALLOCAVEC (char, len); |
5452 | char *p = libfunc_name; |
5453 | const char *q; |
5454 | if (DECIMAL_FLOAT_MODE_P (mode)) |
5455 | { |
5456 | #if ENABLE_DECIMAL_BID_FORMAT |
5457 | memcpy (dest: p, src: "__bid_floatbitint", n: 17); |
5458 | #else |
5459 | memcpy (p, "__dpd_floatbitint", 17); |
5460 | #endif |
5461 | p += 17; |
5462 | } |
5463 | else |
5464 | { |
5465 | memcpy (dest: p, src: "__floatbitint", n: 13); |
5466 | p += 13; |
5467 | } |
5468 | for (q = mname; *q; q++) |
5469 | *p++ = TOLOWER (*q); |
5470 | *p = '\0'; |
5471 | rtx fun = init_one_libfunc (libfunc_name); |
5472 | rtx target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5473 | rtx val = emit_library_call_value (fun, value: target, fn_type: LCT_PURE, outmode: mode, |
5474 | arg1: arg0, arg1_mode: ptr_mode, arg2: arg1, SImode); |
5475 | if (val != target) |
5476 | emit_move_insn (target, val); |
5477 | } |
5478 | |
5479 | static bool |
5480 | expand_bitquery (internal_fn fn, gcall *stmt) |
5481 | { |
5482 | tree lhs = gimple_call_lhs (gs: stmt); |
5483 | if (lhs == NULL_TREE) |
5484 | return false; |
5485 | tree arg = gimple_call_arg (gs: stmt, index: 0); |
5486 | if (TREE_CODE (arg) == INTEGER_CST) |
5487 | { |
5488 | tree ret = fold_const_call (as_combined_fn (fn), TREE_TYPE (arg), arg); |
5489 | gcc_checking_assert (ret && TREE_CODE (ret) == INTEGER_CST); |
5490 | expand_assignment (lhs, ret, false); |
5491 | return false; |
5492 | } |
5493 | return true; |
5494 | } |
5495 | |
5496 | void |
5497 | expand_CLRSB (internal_fn fn, gcall *stmt) |
5498 | { |
5499 | if (expand_bitquery (fn, stmt)) |
5500 | expand_unary_optab_fn (fn, stmt, clrsb_optab); |
5501 | } |
5502 | |
5503 | void |
5504 | expand_CLZ (internal_fn fn, gcall *stmt) |
5505 | { |
5506 | if (expand_bitquery (fn, stmt)) |
5507 | expand_unary_optab_fn (fn, stmt, clz_optab); |
5508 | } |
5509 | |
5510 | void |
5511 | expand_CTZ (internal_fn fn, gcall *stmt) |
5512 | { |
5513 | if (expand_bitquery (fn, stmt)) |
5514 | expand_unary_optab_fn (fn, stmt, ctz_optab); |
5515 | } |
5516 | |
5517 | void |
5518 | expand_FFS (internal_fn fn, gcall *stmt) |
5519 | { |
5520 | if (expand_bitquery (fn, stmt)) |
5521 | expand_unary_optab_fn (fn, stmt, ffs_optab); |
5522 | } |
5523 | |
5524 | void |
5525 | expand_PARITY (internal_fn fn, gcall *stmt) |
5526 | { |
5527 | if (expand_bitquery (fn, stmt)) |
5528 | expand_unary_optab_fn (fn, stmt, parity_optab); |
5529 | } |
5530 | |
5531 | void |
5532 | expand_POPCOUNT (internal_fn fn, gcall *stmt) |
5533 | { |
5534 | if (!expand_bitquery (fn, stmt)) |
5535 | return; |
5536 | if (gimple_call_num_args (gs: stmt) == 1) |
5537 | { |
5538 | expand_unary_optab_fn (fn, stmt, popcount_optab); |
5539 | return; |
5540 | } |
5541 | /* If .POPCOUNT call has 2 arguments, match_single_bit_test marked it |
5542 | because the result is only used in an equality comparison against 1. |
5543 | Use rtx costs in that case to determine if .POPCOUNT (arg) == 1 |
5544 | or (arg ^ (arg - 1)) > arg - 1 is cheaper. |
5545 | If .POPCOUNT second argument is 0, we additionally know that arg |
5546 | is non-zero, so use arg & (arg - 1) == 0 instead. |
5547 | If .POPCOUNT second argument is -1, the comparison was either `<= 1` |
5548 | or `> 1`. */ |
5549 | bool speed_p = optimize_insn_for_speed_p (); |
5550 | tree lhs = gimple_call_lhs (gs: stmt); |
5551 | tree arg = gimple_call_arg (gs: stmt, index: 0); |
5552 | bool nonzero_arg = integer_zerop (gimple_call_arg (gs: stmt, index: 1)); |
5553 | bool was_le = integer_minus_onep (gimple_call_arg (gs: stmt, index: 1)); |
5554 | if (was_le) |
5555 | nonzero_arg = true; |
5556 | tree type = TREE_TYPE (arg); |
5557 | machine_mode mode = TYPE_MODE (type); |
5558 | machine_mode lhsmode = TYPE_MODE (TREE_TYPE (lhs)); |
5559 | do_pending_stack_adjust (); |
5560 | start_sequence (); |
5561 | expand_unary_optab_fn (fn, stmt, popcount_optab); |
5562 | rtx_insn *popcount_insns = end_sequence (); |
5563 | start_sequence (); |
5564 | rtx plhs = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5565 | rtx pcmp = emit_store_flag (NULL_RTX, EQ, plhs, const1_rtx, lhsmode, 0, 0); |
5566 | if (pcmp == NULL_RTX) |
5567 | { |
5568 | fail: |
5569 | end_sequence (); |
5570 | emit_insn (popcount_insns); |
5571 | return; |
5572 | } |
5573 | rtx_insn *popcount_cmp_insns = end_sequence (); |
5574 | start_sequence (); |
5575 | rtx op0 = expand_normal (exp: arg); |
5576 | rtx argm1 = expand_simple_binop (mode, PLUS, op0, constm1_rtx, NULL_RTX, |
5577 | 1, OPTAB_WIDEN); |
5578 | if (argm1 == NULL_RTX) |
5579 | goto fail; |
5580 | rtx argxorargm1 = expand_simple_binop (mode, nonzero_arg ? AND : XOR, op0, |
5581 | argm1, NULL_RTX, 1, OPTAB_WIDEN); |
5582 | if (argxorargm1 == NULL_RTX) |
5583 | goto fail; |
5584 | rtx cmp; |
5585 | if (nonzero_arg) |
5586 | cmp = emit_store_flag (NULL_RTX, EQ, argxorargm1, const0_rtx, mode, 1, 1); |
5587 | else |
5588 | cmp = emit_store_flag (NULL_RTX, GTU, argxorargm1, argm1, mode, 1, 1); |
5589 | if (cmp == NULL_RTX) |
5590 | goto fail; |
5591 | rtx_insn *cmp_insns = end_sequence (); |
5592 | unsigned popcount_cost = (seq_cost (popcount_insns, speed_p) |
5593 | + seq_cost (popcount_cmp_insns, speed_p)); |
5594 | unsigned cmp_cost = seq_cost (cmp_insns, speed_p); |
5595 | |
5596 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5597 | fprintf(stream: dump_file, format: "popcount == 1: popcount cost: %u; cmp cost: %u\n", |
5598 | popcount_cost, cmp_cost); |
5599 | |
5600 | if (popcount_cost <= cmp_cost) |
5601 | emit_insn (popcount_insns); |
5602 | else |
5603 | { |
5604 | start_sequence (); |
5605 | emit_insn (cmp_insns); |
5606 | plhs = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5607 | if (GET_MODE (cmp) != GET_MODE (plhs)) |
5608 | cmp = convert_to_mode (GET_MODE (plhs), cmp, 1); |
5609 | /* For `<= 1`, we need to produce `2 - cmp` or `cmp ? 1 : 2` as that |
5610 | then gets compared against 1 and we need the false case to be 2. */ |
5611 | if (was_le) |
5612 | { |
5613 | cmp = expand_simple_binop (GET_MODE (cmp), MINUS, const2_rtx, |
5614 | cmp, NULL_RTX, 1, OPTAB_WIDEN); |
5615 | if (!cmp) |
5616 | goto fail; |
5617 | } |
5618 | emit_move_insn (plhs, cmp); |
5619 | rtx_insn *all_insns = end_sequence (); |
5620 | emit_insn (all_insns); |
5621 | } |
5622 | } |
5623 |
Definitions
- internal_fn_name_array
- internal_fn_flags_array
- lookup_internal_fn
- lookup_hilo_internal_fn
- lookup_evenodd_internal_fn
- internal_fn_fnspec_array
- init_internal_fns
- direct_internal_fn_array
- create_call_lhs_operand
- assign_call_lhs
- expand_fn_using_insn
- get_multi_vector_move
- add_mask_else_and_len_args
- expand_load_lanes_optab_fn
- expand_store_lanes_optab_fn
- expand_ANNOTATE
- expand_GOMP_USE_SIMT
- expand_GOMP_SIMT_ENTER
- expand_GOMP_SIMT_ENTER_ALLOC
- expand_GOMP_SIMT_EXIT
- expand_GOMP_SIMT_LANE
- expand_GOMP_SIMT_VF
- expand_GOMP_MAX_VF
- expand_GOMP_TARGET_REV
- expand_GOMP_SIMT_LAST_LANE
- expand_GOMP_SIMT_ORDERED_PRED
- expand_GOMP_SIMT_VOTE_ANY
- expand_GOMP_SIMT_XCHG_BFLY
- expand_GOMP_SIMT_XCHG_IDX
- expand_GOMP_SIMD_LANE
- expand_GOMP_SIMD_VF
- expand_GOMP_SIMD_LAST_LANE
- expand_GOMP_SIMD_ORDERED_START
- expand_GOMP_SIMD_ORDERED_END
- expand_GOMP_DISPATCH
- expand_UBSAN_NULL
- expand_UBSAN_BOUNDS
- expand_UBSAN_VPTR
- expand_UBSAN_PTR
- expand_UBSAN_OBJECT_SIZE
- expand_HWASAN_CHECK
- expand_HWASAN_ALLOCA_UNPOISON
- expand_HWASAN_CHOOSE_TAG
- expand_HWASAN_MARK
- expand_HWASAN_SET_TAG
- expand_ASAN_CHECK
- expand_ASAN_MARK
- expand_ASAN_POISON
- expand_ASAN_POISON_USE
- expand_TSAN_FUNC_EXIT
- expand_FALLTHROUGH
- get_min_precision
- expand_arith_set_overflow
- expand_arith_overflow_result_store
- expand_ubsan_result_store
- expand_addsub_overflow
- expand_neg_overflow
- can_widen_mult_without_libcall
- expand_mul_overflow
- expand_vector_ubsan_overflow
- expand_UBSAN_CHECK_ADD
- expand_UBSAN_CHECK_SUB
- expand_UBSAN_CHECK_MUL
- expand_arith_overflow
- expand_ADD_OVERFLOW
- expand_SUB_OVERFLOW
- expand_MUL_OVERFLOW
- expand_UADDC
- expand_USUBC
- expand_LOOP_VECTORIZED
- expand_LOOP_DIST_ALIAS
- expand_call_mem_ref
- expand_partial_load_optab_fn
- expand_partial_store_optab_fn
- expand_vec_cond_mask_optab_fn
- expand_vec_set_optab_fn
- expand_ABNORMAL_DISPATCHER
- expand_BUILTIN_EXPECT
- expand_VA_ARG
- expand_VEC_CONVERT
- expand_RAWMEMCHR
- expand_UNIQUE
- expand_DEFERRED_INIT
- expand_ACCESS_WITH_SIZE
- expand_GOACC_DIM_SIZE
- expand_GOACC_DIM_POS
- expand_GOACC_LOOP
- expand_GOACC_REDUCTION
- expand_GOACC_TILE
- expand_SET_EDOM
- expand_ATOMIC_BIT_TEST_AND_SET
- expand_ATOMIC_BIT_TEST_AND_COMPLEMENT
- expand_ATOMIC_BIT_TEST_AND_RESET
- expand_ATOMIC_COMPARE_EXCHANGE
- expand_ATOMIC_ADD_FETCH_CMP_0
- expand_ATOMIC_SUB_FETCH_CMP_0
- expand_ATOMIC_AND_FETCH_CMP_0
- expand_ATOMIC_OR_FETCH_CMP_0
- expand_ATOMIC_XOR_FETCH_CMP_0
- expand_LAUNDER
- expand_scatter_store_optab_fn
- expand_gather_load_optab_fn
- expand_strided_load_optab_fn
- expand_strided_store_optab_fn
- contains_call_div_mod
- expand_DIVMOD
- expand_NOP
- expand_CO_FRAME
- expand_CO_YIELD
- expand_CO_SUSPN
- expand_CO_ACTOR
- expand_direct_optab_fn
- expand_while_optab_fn
- expand_convert_optab_fn
- expand_crc_optab_fn
- direct_internal_fn_types
- direct_internal_fn_types
- direct_optab_supported_p
- convert_optab_supported_p
- multi_vector_optab_supported_p
- direct_internal_fn_optab
- direct_internal_fn_optab
- type_strictly_matches_mode_p
- type_pair_strictly_matches_mode_p
- direct_internal_fn_supported_p
- direct_internal_fn_supported_p
- direct_internal_fn_supported_p
- commutative_binary_fn_p
- commutative_ternary_fn_p
- associative_binary_fn_p
- first_commutative_argument
- widening_fn_p
- set_edom_supported_p
- internal_fn_expanders
- get_conditional_internal_fn
- conditional_internal_fn_code
- get_conditional_len_internal_fn
- get_conditional_internal_fn
- get_len_internal_fn
- get_unconditional_internal_fn
- can_interpret_as_conditional_op_p
- internal_load_fn_p
- internal_store_fn_p
- internal_gather_scatter_fn_p
- internal_fn_len_index
- internal_fn_else_index
- internal_fn_mask_index
- internal_fn_stored_value_index
- get_supported_else_vals
- supported_else_val_p
- internal_gather_scatter_fn_supported_p
- internal_check_ptrs_fn_supported_p
- internal_len_load_store_bias
- expand_internal_call
- expand_internal_call
- vectorized_internal_fn_supported_p
- expand_SHUFFLEVECTOR
- expand_PHI
- expand_SPACESHIP
- expand_ASSUME
- expand_MASK_CALL
- expand_MULBITINT
- expand_DIVMODBITINT
- expand_FLOATTOBITINT
- expand_BITINTTOFLOAT
- expand_bitquery
- expand_CLRSB
- expand_CLZ
- expand_CTZ
- expand_FFS
- expand_PARITY
Update your C++ knowledge – Modern C++11/14/17 Training
Find out more