]>
Commit | Line | Data |
---|---|---|
26cd9add | 1 | /* Redundant Extension Elimination pass for the GNU compiler. |
23a5b65a | 2 | Copyright (C) 2010-2014 Free Software Foundation, Inc. |
282bc7b4 | 3 | Contributed by Ilya Enkovich (ilya.enkovich@intel.com) |
26cd9add | 4 | |
282bc7b4 EB |
5 | Based on the Redundant Zero-extension elimination pass contributed by |
6 | Sriraman Tallam (tmsriram@google.com) and Silvius Rus (rus@google.com). | |
87a0ebfd ST |
7 | |
8 | This file is part of GCC. | |
9 | ||
10 | GCC is free software; you can redistribute it and/or modify it under | |
11 | the terms of the GNU General Public License as published by the Free | |
12 | Software Foundation; either version 3, or (at your option) any later | |
13 | version. | |
14 | ||
15 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
16 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
17 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
18 | for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
21 | along with GCC; see the file COPYING3. If not see | |
22 | <http://www.gnu.org/licenses/>. */ | |
23 | ||
24 | ||
25 | /* Problem Description : | |
26 | -------------------- | |
26cd9add EI |
27 | This pass is intended to remove redundant extension instructions. |
28 | Such instructions appear for different reasons. We expect some of | |
29 | them due to implicit zero-extension in 64-bit registers after writing | |
30 | to their lower 32-bit half (e.g. for the x86-64 architecture). | |
31 | Another possible reason is a type cast which follows a load (for | |
32 | instance a register restore) and which can be combined into a single | |
33 | instruction, and for which earlier local passes, e.g. the combiner, | |
34 | weren't able to optimize. | |
87a0ebfd ST |
35 | |
36 | How does this pass work ? | |
37 | -------------------------- | |
38 | ||
39 | This pass is run after register allocation. Hence, all registers that | |
26cd9add EI |
40 | this pass deals with are hard registers. This pass first looks for an |
41 | extension instruction that could possibly be redundant. Such extension | |
42 | instructions show up in RTL with the pattern : | |
43 | (set (reg:<SWI248> x) (any_extend:<SWI248> (reg:<SWI124> x))), | |
44 | where x can be any hard register. | |
87a0ebfd | 45 | Now, this pass tries to eliminate this instruction by merging the |
26cd9add | 46 | extension with the definitions of register x. For instance, if |
87a0ebfd ST |
47 | one of the definitions of register x was : |
48 | (set (reg:SI x) (plus:SI (reg:SI z1) (reg:SI z2))), | |
26cd9add EI |
49 | followed by extension : |
50 | (set (reg:DI x) (zero_extend:DI (reg:SI x))) | |
87a0ebfd ST |
51 | then the combination converts this into : |
52 | (set (reg:DI x) (zero_extend:DI (plus:SI (reg:SI z1) (reg:SI z2)))). | |
53 | If all the merged definitions are recognizable assembly instructions, | |
26cd9add EI |
54 | the extension is effectively eliminated. |
55 | ||
56 | For example, for the x86-64 architecture, implicit zero-extensions | |
57 | are captured with appropriate patterns in the i386.md file. Hence, | |
58 | these merged definition can be matched to a single assembly instruction. | |
59 | The original extension instruction is then deleted if all the | |
60 | definitions can be merged. | |
87a0ebfd ST |
61 | |
62 | However, there are cases where the definition instruction cannot be | |
26cd9add EI |
63 | merged with an extension. Examples are CALL instructions. In such |
64 | cases, the original extension is not redundant and this pass does | |
87a0ebfd ST |
65 | not delete it. |
66 | ||
67 | Handling conditional moves : | |
68 | ---------------------------- | |
69 | ||
26cd9add EI |
70 | Architectures like x86-64 support conditional moves whose semantics for |
71 | extension differ from the other instructions. For instance, the | |
87a0ebfd ST |
72 | instruction *cmov ebx, eax* |
73 | zero-extends eax onto rax only when the move from ebx to eax happens. | |
282bc7b4 | 74 | Otherwise, eax may not be zero-extended. Consider conditional moves as |
87a0ebfd ST |
75 | RTL instructions of the form |
76 | (set (reg:SI x) (if_then_else (cond) (reg:SI y) (reg:SI z))). | |
26cd9add | 77 | This pass tries to merge an extension with a conditional move by |
282bc7b4 | 78 | actually merging the definitions of y and z with an extension and then |
87a0ebfd ST |
79 | converting the conditional move into : |
80 | (set (reg:DI x) (if_then_else (cond) (reg:DI y) (reg:DI z))). | |
26cd9add EI |
81 | Since registers y and z are extended, register x will also be extended |
82 | after the conditional move. Note that this step has to be done | |
83 | transitively since the definition of a conditional copy can be | |
87a0ebfd ST |
84 | another conditional copy. |
85 | ||
86 | Motivating Example I : | |
87 | --------------------- | |
88 | For this program : | |
89 | ********************************************** | |
90 | bad_code.c | |
91 | ||
92 | int mask[1000]; | |
93 | ||
94 | int foo(unsigned x) | |
95 | { | |
96 | if (x < 10) | |
97 | x = x * 45; | |
98 | else | |
99 | x = x * 78; | |
100 | return mask[x]; | |
101 | } | |
102 | ********************************************** | |
103 | ||
26cd9add | 104 | $ gcc -O2 bad_code.c |
87a0ebfd ST |
105 | ........ |
106 | 400315: b8 4e 00 00 00 mov $0x4e,%eax | |
107 | 40031a: 0f af f8 imul %eax,%edi | |
282bc7b4 | 108 | 40031d: 89 ff mov %edi,%edi - useless extension |
87a0ebfd ST |
109 | 40031f: 8b 04 bd 60 19 40 00 mov 0x401960(,%rdi,4),%eax |
110 | 400326: c3 retq | |
111 | ...... | |
112 | 400330: ba 2d 00 00 00 mov $0x2d,%edx | |
113 | 400335: 0f af fa imul %edx,%edi | |
282bc7b4 | 114 | 400338: 89 ff mov %edi,%edi - useless extension |
87a0ebfd ST |
115 | 40033a: 8b 04 bd 60 19 40 00 mov 0x401960(,%rdi,4),%eax |
116 | 400341: c3 retq | |
117 | ||
26cd9add | 118 | $ gcc -O2 -free bad_code.c |
87a0ebfd ST |
119 | ...... |
120 | 400315: 6b ff 4e imul $0x4e,%edi,%edi | |
121 | 400318: 8b 04 bd 40 19 40 00 mov 0x401940(,%rdi,4),%eax | |
122 | 40031f: c3 retq | |
123 | 400320: 6b ff 2d imul $0x2d,%edi,%edi | |
124 | 400323: 8b 04 bd 40 19 40 00 mov 0x401940(,%rdi,4),%eax | |
125 | 40032a: c3 retq | |
126 | ||
127 | Motivating Example II : | |
128 | --------------------- | |
129 | ||
130 | Here is an example with a conditional move. | |
131 | ||
132 | For this program : | |
133 | ********************************************** | |
134 | ||
135 | unsigned long long foo(unsigned x , unsigned y) | |
136 | { | |
137 | unsigned z; | |
138 | if (x > 100) | |
139 | z = x + y; | |
140 | else | |
141 | z = x - y; | |
142 | return (unsigned long long)(z); | |
143 | } | |
144 | ||
26cd9add | 145 | $ gcc -O2 bad_code.c |
87a0ebfd ST |
146 | ............ |
147 | 400360: 8d 14 3e lea (%rsi,%rdi,1),%edx | |
148 | 400363: 89 f8 mov %edi,%eax | |
149 | 400365: 29 f0 sub %esi,%eax | |
150 | 400367: 83 ff 65 cmp $0x65,%edi | |
151 | 40036a: 0f 43 c2 cmovae %edx,%eax | |
282bc7b4 | 152 | 40036d: 89 c0 mov %eax,%eax - useless extension |
87a0ebfd ST |
153 | 40036f: c3 retq |
154 | ||
26cd9add | 155 | $ gcc -O2 -free bad_code.c |
87a0ebfd ST |
156 | ............. |
157 | 400360: 89 fa mov %edi,%edx | |
158 | 400362: 8d 04 3e lea (%rsi,%rdi,1),%eax | |
159 | 400365: 29 f2 sub %esi,%edx | |
160 | 400367: 83 ff 65 cmp $0x65,%edi | |
161 | 40036a: 89 d6 mov %edx,%esi | |
162 | 40036c: 48 0f 42 c6 cmovb %rsi,%rax | |
163 | 400370: c3 retq | |
164 | ||
26cd9add EI |
165 | Motivating Example III : |
166 | --------------------- | |
167 | ||
168 | Here is an example with a type cast. | |
169 | ||
170 | For this program : | |
171 | ********************************************** | |
172 | ||
173 | void test(int size, unsigned char *in, unsigned char *out) | |
174 | { | |
175 | int i; | |
176 | unsigned char xr, xg, xy=0; | |
177 | ||
178 | for (i = 0; i < size; i++) { | |
179 | xr = *in++; | |
180 | xg = *in++; | |
181 | xy = (unsigned char) ((19595*xr + 38470*xg) >> 16); | |
182 | *out++ = xy; | |
183 | } | |
184 | } | |
185 | ||
186 | $ gcc -O2 bad_code.c | |
187 | ............ | |
188 | 10: 0f b6 0e movzbl (%rsi),%ecx | |
189 | 13: 0f b6 46 01 movzbl 0x1(%rsi),%eax | |
190 | 17: 48 83 c6 02 add $0x2,%rsi | |
282bc7b4 EB |
191 | 1b: 0f b6 c9 movzbl %cl,%ecx - useless extension |
192 | 1e: 0f b6 c0 movzbl %al,%eax - useless extension | |
26cd9add EI |
193 | 21: 69 c9 8b 4c 00 00 imul $0x4c8b,%ecx,%ecx |
194 | 27: 69 c0 46 96 00 00 imul $0x9646,%eax,%eax | |
195 | ||
196 | $ gcc -O2 -free bad_code.c | |
197 | ............. | |
198 | 10: 0f b6 0e movzbl (%rsi),%ecx | |
199 | 13: 0f b6 46 01 movzbl 0x1(%rsi),%eax | |
200 | 17: 48 83 c6 02 add $0x2,%rsi | |
201 | 1b: 69 c9 8b 4c 00 00 imul $0x4c8b,%ecx,%ecx | |
202 | 21: 69 c0 46 96 00 00 imul $0x9646,%eax,%eax | |
87a0ebfd ST |
203 | |
204 | Usefulness : | |
205 | ---------- | |
206 | ||
26cd9add EI |
207 | The original redundant zero-extension elimination pass reported reduction |
208 | of the dynamic instruction count of a compression benchmark by 2.8% and | |
209 | improvement of its run time by about 1%. | |
87a0ebfd | 210 | |
26cd9add EI |
211 | The additional performance gain with the enhanced pass is mostly expected |
212 | on in-order architectures where redundancy cannot be compensated by out of | |
213 | order execution. Measurements showed up to 10% performance gain (reduced | |
214 | run time) on EEMBC 2.0 benchmarks on Atom processor with geomean performance | |
215 | gain 1%. */ | |
87a0ebfd ST |
216 | |
217 | ||
218 | #include "config.h" | |
219 | #include "system.h" | |
220 | #include "coretypes.h" | |
221 | #include "tm.h" | |
222 | #include "rtl.h" | |
223 | #include "tree.h" | |
224 | #include "tm_p.h" | |
225 | #include "flags.h" | |
226 | #include "regs.h" | |
227 | #include "hard-reg-set.h" | |
228 | #include "basic-block.h" | |
229 | #include "insn-config.h" | |
83685514 AM |
230 | #include "hashtab.h" |
231 | #include "hash-set.h" | |
232 | #include "vec.h" | |
233 | #include "machmode.h" | |
234 | #include "input.h" | |
87a0ebfd ST |
235 | #include "function.h" |
236 | #include "expr.h" | |
237 | #include "insn-attr.h" | |
238 | #include "recog.h" | |
718f9c0f | 239 | #include "diagnostic-core.h" |
87a0ebfd | 240 | #include "target.h" |
87a0ebfd ST |
241 | #include "optabs.h" |
242 | #include "insn-codes.h" | |
243 | #include "rtlhooks-def.h" | |
87a0ebfd | 244 | #include "params.h" |
87a0ebfd ST |
245 | #include "tree-pass.h" |
246 | #include "df.h" | |
247 | #include "cgraph.h" | |
248 | ||
282bc7b4 | 249 | /* This structure represents a candidate for elimination. */ |
87a0ebfd | 250 | |
8a1239ac | 251 | typedef struct ext_cand |
87a0ebfd | 252 | { |
282bc7b4 EB |
253 | /* The expression. */ |
254 | const_rtx expr; | |
87a0ebfd | 255 | |
282bc7b4 EB |
256 | /* The kind of extension. */ |
257 | enum rtx_code code; | |
26cd9add | 258 | |
282bc7b4 EB |
259 | /* The destination mode. */ |
260 | enum machine_mode mode; | |
261 | ||
262 | /* The instruction where it lives. */ | |
59a0c032 | 263 | rtx_insn *insn; |
282bc7b4 | 264 | } ext_cand; |
26cd9add | 265 | |
26cd9add | 266 | |
87a0ebfd ST |
267 | static int max_insn_uid; |
268 | ||
26cd9add EI |
269 | /* Given a insn (CURR_INSN), an extension candidate for removal (CAND) |
270 | and a pointer to the SET rtx (ORIG_SET) that needs to be modified, | |
271 | this code modifies the SET rtx to a new SET rtx that extends the | |
272 | right hand expression into a register on the left hand side. Note | |
273 | that multiple assumptions are made about the nature of the set that | |
274 | needs to be true for this to work and is called from merge_def_and_ext. | |
87a0ebfd ST |
275 | |
276 | Original : | |
26cd9add | 277 | (set (reg a) (expression)) |
87a0ebfd ST |
278 | |
279 | Transform : | |
282bc7b4 | 280 | (set (reg a) (any_extend (expression))) |
87a0ebfd ST |
281 | |
282 | Special Cases : | |
282bc7b4 | 283 | If the expression is a constant or another extension, then directly |
26cd9add | 284 | assign it to the register. */ |
87a0ebfd ST |
285 | |
286 | static bool | |
59a0c032 | 287 | combine_set_extension (ext_cand *cand, rtx_insn *curr_insn, rtx *orig_set) |
87a0ebfd | 288 | { |
282bc7b4 | 289 | rtx orig_src = SET_SRC (*orig_set); |
282bc7b4 | 290 | rtx new_set; |
3c92da90 JL |
291 | rtx cand_pat = PATTERN (cand->insn); |
292 | ||
293 | /* If the extension's source/destination registers are not the same | |
294 | then we need to change the original load to reference the destination | |
295 | of the extension. Then we need to emit a copy from that destination | |
296 | to the original destination of the load. */ | |
297 | rtx new_reg; | |
298 | bool copy_needed | |
299 | = (REGNO (SET_DEST (cand_pat)) != REGNO (XEXP (SET_SRC (cand_pat), 0))); | |
300 | if (copy_needed) | |
301 | new_reg = gen_rtx_REG (cand->mode, REGNO (SET_DEST (cand_pat))); | |
302 | else | |
303 | new_reg = gen_rtx_REG (cand->mode, REGNO (SET_DEST (*orig_set))); | |
87a0ebfd | 304 | |
2043135a JL |
305 | #if 0 |
306 | /* Rethinking test. Temporarily disabled. */ | |
a6a2d67b JL |
307 | /* We're going to be widening the result of DEF_INSN, ensure that doing so |
308 | doesn't change the number of hard registers needed for the result. */ | |
309 | if (HARD_REGNO_NREGS (REGNO (new_reg), cand->mode) | |
2043135a JL |
310 | != HARD_REGNO_NREGS (REGNO (SET_DEST (*orig_set)), |
311 | GET_MODE (SET_DEST (*orig_set)))) | |
a6a2d67b | 312 | return false; |
2043135a | 313 | #endif |
a6a2d67b | 314 | |
282bc7b4 EB |
315 | /* Merge constants by directly moving the constant into the register under |
316 | some conditions. Recall that RTL constants are sign-extended. */ | |
26cd9add | 317 | if (GET_CODE (orig_src) == CONST_INT |
282bc7b4 | 318 | && HOST_BITS_PER_WIDE_INT >= GET_MODE_BITSIZE (cand->mode)) |
26cd9add | 319 | { |
282bc7b4 EB |
320 | if (INTVAL (orig_src) >= 0 || cand->code == SIGN_EXTEND) |
321 | new_set = gen_rtx_SET (VOIDmode, new_reg, orig_src); | |
87a0ebfd | 322 | else |
26cd9add EI |
323 | { |
324 | /* Zero-extend the negative constant by masking out the bits outside | |
325 | the source mode. */ | |
326 | enum machine_mode src_mode = GET_MODE (SET_DEST (*orig_set)); | |
282bc7b4 | 327 | rtx new_const_int |
69db2d57 RS |
328 | = gen_int_mode (INTVAL (orig_src) & GET_MODE_MASK (src_mode), |
329 | GET_MODE (new_reg)); | |
282bc7b4 | 330 | new_set = gen_rtx_SET (VOIDmode, new_reg, new_const_int); |
26cd9add EI |
331 | } |
332 | } | |
333 | else if (GET_MODE (orig_src) == VOIDmode) | |
334 | { | |
282bc7b4 | 335 | /* This is mostly due to a call insn that should not be optimized. */ |
26cd9add | 336 | return false; |
87a0ebfd | 337 | } |
282bc7b4 | 338 | else if (GET_CODE (orig_src) == cand->code) |
87a0ebfd | 339 | { |
282bc7b4 EB |
340 | /* Here is a sequence of two extensions. Try to merge them. */ |
341 | rtx temp_extension | |
342 | = gen_rtx_fmt_e (cand->code, cand->mode, XEXP (orig_src, 0)); | |
343 | rtx simplified_temp_extension = simplify_rtx (temp_extension); | |
87a0ebfd ST |
344 | if (simplified_temp_extension) |
345 | temp_extension = simplified_temp_extension; | |
282bc7b4 | 346 | new_set = gen_rtx_SET (VOIDmode, new_reg, temp_extension); |
87a0ebfd ST |
347 | } |
348 | else if (GET_CODE (orig_src) == IF_THEN_ELSE) | |
349 | { | |
26cd9add | 350 | /* Only IF_THEN_ELSE of phi-type copies are combined. Otherwise, |
87a0ebfd | 351 | in general, IF_THEN_ELSE should not be combined. */ |
87a0ebfd ST |
352 | return false; |
353 | } | |
354 | else | |
355 | { | |
282bc7b4 EB |
356 | /* This is the normal case. */ |
357 | rtx temp_extension | |
358 | = gen_rtx_fmt_e (cand->code, cand->mode, orig_src); | |
359 | rtx simplified_temp_extension = simplify_rtx (temp_extension); | |
87a0ebfd ST |
360 | if (simplified_temp_extension) |
361 | temp_extension = simplified_temp_extension; | |
282bc7b4 | 362 | new_set = gen_rtx_SET (VOIDmode, new_reg, temp_extension); |
87a0ebfd ST |
363 | } |
364 | ||
26cd9add | 365 | /* This change is a part of a group of changes. Hence, |
87a0ebfd | 366 | validate_change will not try to commit the change. */ |
87a0ebfd ST |
367 | if (validate_change (curr_insn, orig_set, new_set, true)) |
368 | { | |
369 | if (dump_file) | |
370 | { | |
ca10595c | 371 | fprintf (dump_file, |
3c92da90 JL |
372 | "Tentatively merged extension with definition %s:\n", |
373 | (copy_needed) ? "(copy needed)" : ""); | |
87a0ebfd ST |
374 | print_rtl_single (dump_file, curr_insn); |
375 | } | |
376 | return true; | |
377 | } | |
282bc7b4 | 378 | |
87a0ebfd ST |
379 | return false; |
380 | } | |
381 | ||
87a0ebfd | 382 | /* Treat if_then_else insns, where the operands of both branches |
26cd9add | 383 | are registers, as copies. For instance, |
87a0ebfd ST |
384 | Original : |
385 | (set (reg:SI a) (if_then_else (cond) (reg:SI b) (reg:SI c))) | |
386 | Transformed : | |
387 | (set (reg:DI a) (if_then_else (cond) (reg:DI b) (reg:DI c))) | |
388 | DEF_INSN is the if_then_else insn. */ | |
389 | ||
390 | static bool | |
59a0c032 | 391 | transform_ifelse (ext_cand *cand, rtx_insn *def_insn) |
87a0ebfd ST |
392 | { |
393 | rtx set_insn = PATTERN (def_insn); | |
394 | rtx srcreg, dstreg, srcreg2; | |
395 | rtx map_srcreg, map_dstreg, map_srcreg2; | |
396 | rtx ifexpr; | |
397 | rtx cond; | |
398 | rtx new_set; | |
399 | ||
400 | gcc_assert (GET_CODE (set_insn) == SET); | |
282bc7b4 | 401 | |
87a0ebfd ST |
402 | cond = XEXP (SET_SRC (set_insn), 0); |
403 | dstreg = SET_DEST (set_insn); | |
404 | srcreg = XEXP (SET_SRC (set_insn), 1); | |
405 | srcreg2 = XEXP (SET_SRC (set_insn), 2); | |
b57cca0b JJ |
406 | /* If the conditional move already has the right or wider mode, |
407 | there is nothing to do. */ | |
408 | if (GET_MODE_SIZE (GET_MODE (dstreg)) >= GET_MODE_SIZE (cand->mode)) | |
409 | return true; | |
410 | ||
282bc7b4 EB |
411 | map_srcreg = gen_rtx_REG (cand->mode, REGNO (srcreg)); |
412 | map_srcreg2 = gen_rtx_REG (cand->mode, REGNO (srcreg2)); | |
413 | map_dstreg = gen_rtx_REG (cand->mode, REGNO (dstreg)); | |
414 | ifexpr = gen_rtx_IF_THEN_ELSE (cand->mode, cond, map_srcreg, map_srcreg2); | |
87a0ebfd ST |
415 | new_set = gen_rtx_SET (VOIDmode, map_dstreg, ifexpr); |
416 | ||
417 | if (validate_change (def_insn, &PATTERN (def_insn), new_set, true)) | |
418 | { | |
419 | if (dump_file) | |
420 | { | |
282bc7b4 EB |
421 | fprintf (dump_file, |
422 | "Mode of conditional move instruction extended:\n"); | |
87a0ebfd ST |
423 | print_rtl_single (dump_file, def_insn); |
424 | } | |
425 | return true; | |
426 | } | |
282bc7b4 EB |
427 | |
428 | return false; | |
87a0ebfd ST |
429 | } |
430 | ||
282bc7b4 EB |
431 | /* Get all the reaching definitions of an instruction. The definitions are |
432 | desired for REG used in INSN. Return the definition list or NULL if a | |
433 | definition is missing. If DEST is non-NULL, additionally push the INSN | |
434 | of the definitions onto DEST. */ | |
87a0ebfd | 435 | |
282bc7b4 | 436 | static struct df_link * |
59a0c032 | 437 | get_defs (rtx_insn *insn, rtx reg, vec<rtx_insn *> *dest) |
87a0ebfd | 438 | { |
bfac633a | 439 | df_ref use; |
282bc7b4 | 440 | struct df_link *ref_chain, *ref_link; |
87a0ebfd | 441 | |
bfac633a | 442 | FOR_EACH_INSN_USE (use, insn) |
87a0ebfd | 443 | { |
bfac633a | 444 | if (GET_CODE (DF_REF_REG (use)) == SUBREG) |
282bc7b4 | 445 | return NULL; |
bfac633a RS |
446 | if (REGNO (DF_REF_REG (use)) == REGNO (reg)) |
447 | break; | |
87a0ebfd ST |
448 | } |
449 | ||
bfac633a | 450 | gcc_assert (use != NULL); |
87a0ebfd | 451 | |
bfac633a | 452 | ref_chain = DF_REF_CHAIN (use); |
282bc7b4 EB |
453 | |
454 | for (ref_link = ref_chain; ref_link; ref_link = ref_link->next) | |
87a0ebfd ST |
455 | { |
456 | /* Problem getting some definition for this instruction. */ | |
282bc7b4 EB |
457 | if (ref_link->ref == NULL) |
458 | return NULL; | |
459 | if (DF_REF_INSN_INFO (ref_link->ref) == NULL) | |
460 | return NULL; | |
87a0ebfd ST |
461 | } |
462 | ||
282bc7b4 EB |
463 | if (dest) |
464 | for (ref_link = ref_chain; ref_link; ref_link = ref_link->next) | |
9771b263 | 465 | dest->safe_push (DF_REF_INSN (ref_link->ref)); |
87a0ebfd | 466 | |
282bc7b4 | 467 | return ref_chain; |
87a0ebfd ST |
468 | } |
469 | ||
282bc7b4 EB |
470 | /* Return true if INSN is |
471 | (SET (reg REGNO (def_reg)) (if_then_else (cond) (REG x1) (REG x2))) | |
472 | and store x1 and x2 in REG_1 and REG_2. */ | |
87a0ebfd | 473 | |
282bc7b4 | 474 | static bool |
59a0c032 | 475 | is_cond_copy_insn (rtx_insn *insn, rtx *reg1, rtx *reg2) |
87a0ebfd | 476 | { |
282bc7b4 | 477 | rtx expr = single_set (insn); |
87a0ebfd | 478 | |
282bc7b4 EB |
479 | if (expr != NULL_RTX |
480 | && GET_CODE (expr) == SET | |
87a0ebfd | 481 | && GET_CODE (SET_DEST (expr)) == REG |
87a0ebfd ST |
482 | && GET_CODE (SET_SRC (expr)) == IF_THEN_ELSE |
483 | && GET_CODE (XEXP (SET_SRC (expr), 1)) == REG | |
26cd9add | 484 | && GET_CODE (XEXP (SET_SRC (expr), 2)) == REG) |
87a0ebfd | 485 | { |
282bc7b4 EB |
486 | *reg1 = XEXP (SET_SRC (expr), 1); |
487 | *reg2 = XEXP (SET_SRC (expr), 2); | |
488 | return true; | |
87a0ebfd ST |
489 | } |
490 | ||
282bc7b4 | 491 | return false; |
87a0ebfd ST |
492 | } |
493 | ||
b57cca0b JJ |
494 | enum ext_modified_kind |
495 | { | |
496 | /* The insn hasn't been modified by ree pass yet. */ | |
497 | EXT_MODIFIED_NONE, | |
498 | /* Changed into zero extension. */ | |
499 | EXT_MODIFIED_ZEXT, | |
500 | /* Changed into sign extension. */ | |
501 | EXT_MODIFIED_SEXT | |
502 | }; | |
503 | ||
925e30ff | 504 | struct ATTRIBUTE_PACKED ext_modified |
b57cca0b JJ |
505 | { |
506 | /* Mode from which ree has zero or sign extended the destination. */ | |
507 | ENUM_BITFIELD(machine_mode) mode : 8; | |
508 | ||
509 | /* Kind of modification of the insn. */ | |
510 | ENUM_BITFIELD(ext_modified_kind) kind : 2; | |
511 | ||
0d732cca JL |
512 | unsigned int do_not_reextend : 1; |
513 | ||
b57cca0b JJ |
514 | /* True if the insn is scheduled to be deleted. */ |
515 | unsigned int deleted : 1; | |
516 | }; | |
517 | ||
518 | /* Vectors used by combine_reaching_defs and its helpers. */ | |
519 | typedef struct ext_state | |
520 | { | |
9771b263 | 521 | /* In order to avoid constant alloc/free, we keep these |
b57cca0b | 522 | 4 vectors live through the entire find_and_remove_re and just |
9771b263 | 523 | truncate them each time. */ |
59a0c032 DM |
524 | vec<rtx_insn *> defs_list; |
525 | vec<rtx_insn *> copies_list; | |
526 | vec<rtx_insn *> modified_list; | |
527 | vec<rtx_insn *> work_list; | |
b57cca0b JJ |
528 | |
529 | /* For instructions that have been successfully modified, this is | |
530 | the original mode from which the insn is extending and | |
531 | kind of extension. */ | |
532 | struct ext_modified *modified; | |
533 | } ext_state; | |
534 | ||
26cd9add EI |
535 | /* Reaching Definitions of the extended register could be conditional copies |
536 | or regular definitions. This function separates the two types into two | |
b57cca0b JJ |
537 | lists, STATE->DEFS_LIST and STATE->COPIES_LIST. This is necessary because, |
538 | if a reaching definition is a conditional copy, merging the extension with | |
539 | this definition is wrong. Conditional copies are merged by transitively | |
540 | merging their definitions. The defs_list is populated with all the reaching | |
541 | definitions of the extension instruction (EXTEND_INSN) which must be merged | |
542 | with an extension. The copies_list contains all the conditional moves that | |
543 | will later be extended into a wider mode conditional move if all the merges | |
544 | are successful. The function returns false upon failure, true upon | |
545 | success. */ | |
546 | ||
547 | static bool | |
59a0c032 | 548 | make_defs_and_copies_lists (rtx_insn *extend_insn, const_rtx set_pat, |
b57cca0b | 549 | ext_state *state) |
87a0ebfd | 550 | { |
282bc7b4 | 551 | rtx src_reg = XEXP (SET_SRC (set_pat), 0); |
87a0ebfd | 552 | bool *is_insn_visited; |
b57cca0b JJ |
553 | bool ret = true; |
554 | ||
9771b263 | 555 | state->work_list.truncate (0); |
87a0ebfd | 556 | |
282bc7b4 | 557 | /* Initialize the work list. */ |
b57cca0b JJ |
558 | if (!get_defs (extend_insn, src_reg, &state->work_list)) |
559 | gcc_unreachable (); | |
87a0ebfd | 560 | |
282bc7b4 | 561 | is_insn_visited = XCNEWVEC (bool, max_insn_uid); |
87a0ebfd ST |
562 | |
563 | /* Perform transitive closure for conditional copies. */ | |
9771b263 | 564 | while (!state->work_list.is_empty ()) |
87a0ebfd | 565 | { |
59a0c032 | 566 | rtx_insn *def_insn = state->work_list.pop (); |
282bc7b4 EB |
567 | rtx reg1, reg2; |
568 | ||
87a0ebfd ST |
569 | gcc_assert (INSN_UID (def_insn) < max_insn_uid); |
570 | ||
571 | if (is_insn_visited[INSN_UID (def_insn)]) | |
282bc7b4 | 572 | continue; |
87a0ebfd | 573 | is_insn_visited[INSN_UID (def_insn)] = true; |
87a0ebfd | 574 | |
282bc7b4 EB |
575 | if (is_cond_copy_insn (def_insn, ®1, ®2)) |
576 | { | |
577 | /* Push it onto the copy list first. */ | |
9771b263 | 578 | state->copies_list.safe_push (def_insn); |
282bc7b4 EB |
579 | |
580 | /* Now perform the transitive closure. */ | |
b57cca0b JJ |
581 | if (!get_defs (def_insn, reg1, &state->work_list) |
582 | || !get_defs (def_insn, reg2, &state->work_list)) | |
282bc7b4 | 583 | { |
b57cca0b | 584 | ret = false; |
282bc7b4 EB |
585 | break; |
586 | } | |
87a0ebfd ST |
587 | } |
588 | else | |
9771b263 | 589 | state->defs_list.safe_push (def_insn); |
87a0ebfd ST |
590 | } |
591 | ||
87a0ebfd | 592 | XDELETEVEC (is_insn_visited); |
282bc7b4 EB |
593 | |
594 | return ret; | |
87a0ebfd ST |
595 | } |
596 | ||
650c4c85 JL |
597 | /* If DEF_INSN has single SET expression, possibly buried inside |
598 | a PARALLEL, return the address of the SET expression, else | |
599 | return NULL. This is similar to single_set, except that | |
600 | single_set allows multiple SETs when all but one is dead. */ | |
601 | static rtx * | |
59a0c032 | 602 | get_sub_rtx (rtx_insn *def_insn) |
87a0ebfd | 603 | { |
650c4c85 JL |
604 | enum rtx_code code = GET_CODE (PATTERN (def_insn)); |
605 | rtx *sub_rtx = NULL; | |
87a0ebfd ST |
606 | |
607 | if (code == PARALLEL) | |
608 | { | |
650c4c85 | 609 | for (int i = 0; i < XVECLEN (PATTERN (def_insn), 0); i++) |
87a0ebfd | 610 | { |
650c4c85 | 611 | rtx s_expr = XVECEXP (PATTERN (def_insn), 0, i); |
87a0ebfd ST |
612 | if (GET_CODE (s_expr) != SET) |
613 | continue; | |
614 | ||
615 | if (sub_rtx == NULL) | |
616 | sub_rtx = &XVECEXP (PATTERN (def_insn), 0, i); | |
617 | else | |
618 | { | |
619 | /* PARALLEL with multiple SETs. */ | |
650c4c85 | 620 | return NULL; |
87a0ebfd ST |
621 | } |
622 | } | |
623 | } | |
624 | else if (code == SET) | |
625 | sub_rtx = &PATTERN (def_insn); | |
626 | else | |
627 | { | |
628 | /* It is not a PARALLEL or a SET, what could it be ? */ | |
650c4c85 | 629 | return NULL; |
87a0ebfd ST |
630 | } |
631 | ||
632 | gcc_assert (sub_rtx != NULL); | |
650c4c85 JL |
633 | return sub_rtx; |
634 | } | |
635 | ||
636 | /* Merge the DEF_INSN with an extension. Calls combine_set_extension | |
637 | on the SET pattern. */ | |
638 | ||
639 | static bool | |
59a0c032 | 640 | merge_def_and_ext (ext_cand *cand, rtx_insn *def_insn, ext_state *state) |
650c4c85 JL |
641 | { |
642 | enum machine_mode ext_src_mode; | |
643 | rtx *sub_rtx; | |
644 | ||
645 | ext_src_mode = GET_MODE (XEXP (SET_SRC (cand->expr), 0)); | |
646 | sub_rtx = get_sub_rtx (def_insn); | |
647 | ||
648 | if (sub_rtx == NULL) | |
649 | return false; | |
87a0ebfd | 650 | |
b57cca0b JJ |
651 | if (REG_P (SET_DEST (*sub_rtx)) |
652 | && (GET_MODE (SET_DEST (*sub_rtx)) == ext_src_mode | |
653 | || ((state->modified[INSN_UID (def_insn)].kind | |
654 | == (cand->code == ZERO_EXTEND | |
655 | ? EXT_MODIFIED_ZEXT : EXT_MODIFIED_SEXT)) | |
656 | && state->modified[INSN_UID (def_insn)].mode | |
657 | == ext_src_mode))) | |
87a0ebfd | 658 | { |
b57cca0b JJ |
659 | if (GET_MODE_SIZE (GET_MODE (SET_DEST (*sub_rtx))) |
660 | >= GET_MODE_SIZE (cand->mode)) | |
661 | return true; | |
662 | /* If def_insn is already scheduled to be deleted, don't attempt | |
663 | to modify it. */ | |
664 | if (state->modified[INSN_UID (def_insn)].deleted) | |
665 | return false; | |
666 | if (combine_set_extension (cand, def_insn, sub_rtx)) | |
667 | { | |
668 | if (state->modified[INSN_UID (def_insn)].kind == EXT_MODIFIED_NONE) | |
669 | state->modified[INSN_UID (def_insn)].mode = ext_src_mode; | |
670 | return true; | |
671 | } | |
87a0ebfd | 672 | } |
26cd9add EI |
673 | |
674 | return false; | |
87a0ebfd ST |
675 | } |
676 | ||
059742a4 JL |
677 | /* Given SRC, which should be one or more extensions of a REG, strip |
678 | away the extensions and return the REG. */ | |
679 | ||
680 | static inline rtx | |
681 | get_extended_src_reg (rtx src) | |
682 | { | |
683 | while (GET_CODE (src) == SIGN_EXTEND || GET_CODE (src) == ZERO_EXTEND) | |
684 | src = XEXP (src, 0); | |
685 | gcc_assert (REG_P (src)); | |
686 | return src; | |
687 | } | |
688 | ||
87a0ebfd | 689 | /* This function goes through all reaching defs of the source |
26cd9add EI |
690 | of the candidate for elimination (CAND) and tries to combine |
691 | the extension with the definition instruction. The changes | |
692 | are made as a group so that even if one definition cannot be | |
693 | merged, all reaching definitions end up not being merged. | |
694 | When a conditional copy is encountered, merging is attempted | |
695 | transitively on its definitions. It returns true upon success | |
696 | and false upon failure. */ | |
87a0ebfd ST |
697 | |
698 | static bool | |
089dacc5 | 699 | combine_reaching_defs (ext_cand *cand, const_rtx set_pat, ext_state *state) |
87a0ebfd | 700 | { |
59a0c032 | 701 | rtx_insn *def_insn; |
87a0ebfd ST |
702 | bool merge_successful = true; |
703 | int i; | |
704 | int defs_ix; | |
b57cca0b | 705 | bool outcome; |
87a0ebfd | 706 | |
9771b263 DN |
707 | state->defs_list.truncate (0); |
708 | state->copies_list.truncate (0); | |
87a0ebfd | 709 | |
b57cca0b | 710 | outcome = make_defs_and_copies_lists (cand->insn, set_pat, state); |
87a0ebfd | 711 | |
b57cca0b JJ |
712 | if (!outcome) |
713 | return false; | |
87a0ebfd | 714 | |
3c92da90 JL |
715 | /* If the destination operand of the extension is a different |
716 | register than the source operand, then additional restrictions | |
059742a4 JL |
717 | are needed. Note we have to handle cases where we have nested |
718 | extensions in the source operand. */ | |
0d732cca JL |
719 | bool copy_needed |
720 | = (REGNO (SET_DEST (PATTERN (cand->insn))) | |
721 | != REGNO (get_extended_src_reg (SET_SRC (PATTERN (cand->insn))))); | |
722 | if (copy_needed) | |
3c92da90 JL |
723 | { |
724 | /* In theory we could handle more than one reaching def, it | |
725 | just makes the code to update the insn stream more complex. */ | |
726 | if (state->defs_list.length () != 1) | |
727 | return false; | |
728 | ||
059742a4 JL |
729 | /* We require the candidate not already be modified. It may, |
730 | for example have been changed from a (sign_extend (reg)) | |
0d732cca | 731 | into (zero_extend (sign_extend (reg))). |
059742a4 JL |
732 | |
733 | Handling that case shouldn't be terribly difficult, but the code | |
734 | here and the code to emit copies would need auditing. Until | |
735 | we see a need, this is the safe thing to do. */ | |
3c92da90 JL |
736 | if (state->modified[INSN_UID (cand->insn)].kind != EXT_MODIFIED_NONE) |
737 | return false; | |
738 | ||
0d6d7b9a JJ |
739 | /* Transformation of |
740 | (set (reg1) (expression)) | |
741 | (set (reg2) (any_extend (reg1))) | |
742 | into | |
743 | (set (reg2) (any_extend (expression))) | |
744 | (set (reg1) (reg2)) | |
745 | is only valid for scalar integral modes, as it relies on the low | |
746 | subreg of reg1 to have the value of (expression), which is not true | |
747 | e.g. for vector modes. */ | |
748 | if (!SCALAR_INT_MODE_P (GET_MODE (SET_DEST (PATTERN (cand->insn))))) | |
749 | return false; | |
750 | ||
e533e26c WD |
751 | enum machine_mode dst_mode = GET_MODE (SET_DEST (PATTERN (cand->insn))); |
752 | rtx src_reg = get_extended_src_reg (SET_SRC (PATTERN (cand->insn))); | |
753 | ||
754 | /* Ensure the number of hard registers of the copy match. */ | |
755 | if (HARD_REGNO_NREGS (REGNO (src_reg), dst_mode) | |
756 | != HARD_REGNO_NREGS (REGNO (src_reg), GET_MODE (src_reg))) | |
757 | return false; | |
758 | ||
3c92da90 | 759 | /* There's only one reaching def. */ |
59a0c032 | 760 | rtx_insn *def_insn = state->defs_list[0]; |
3c92da90 JL |
761 | |
762 | /* The defining statement must not have been modified either. */ | |
763 | if (state->modified[INSN_UID (def_insn)].kind != EXT_MODIFIED_NONE) | |
764 | return false; | |
765 | ||
766 | /* The defining statement and candidate insn must be in the same block. | |
767 | This is merely to keep the test for safety and updating the insn | |
7e41c852 JL |
768 | stream simple. Also ensure that within the block the candidate |
769 | follows the defining insn. */ | |
770 | if (BLOCK_FOR_INSN (cand->insn) != BLOCK_FOR_INSN (def_insn) | |
771 | || DF_INSN_LUID (def_insn) > DF_INSN_LUID (cand->insn)) | |
3c92da90 JL |
772 | return false; |
773 | ||
774 | /* If there is an overlap between the destination of DEF_INSN and | |
775 | CAND->insn, then this transformation is not safe. Note we have | |
776 | to test in the widened mode. */ | |
650c4c85 JL |
777 | rtx *dest_sub_rtx = get_sub_rtx (def_insn); |
778 | if (dest_sub_rtx == NULL | |
779 | || !REG_P (SET_DEST (*dest_sub_rtx))) | |
780 | return false; | |
781 | ||
3c92da90 | 782 | rtx tmp_reg = gen_rtx_REG (GET_MODE (SET_DEST (PATTERN (cand->insn))), |
650c4c85 | 783 | REGNO (SET_DEST (*dest_sub_rtx))); |
3c92da90 JL |
784 | if (reg_overlap_mentioned_p (tmp_reg, SET_DEST (PATTERN (cand->insn)))) |
785 | return false; | |
786 | ||
787 | /* The destination register of the extension insn must not be | |
788 | used or set between the def_insn and cand->insn exclusive. */ | |
789 | if (reg_used_between_p (SET_DEST (PATTERN (cand->insn)), | |
790 | def_insn, cand->insn) | |
791 | || reg_set_between_p (SET_DEST (PATTERN (cand->insn)), | |
792 | def_insn, cand->insn)) | |
793 | return false; | |
0d732cca JL |
794 | |
795 | /* We must be able to copy between the two registers. Generate, | |
796 | recognize and verify constraints of the copy. Also fail if this | |
797 | generated more than one insn. | |
798 | ||
799 | This generates garbage since we throw away the insn when we're | |
c7ece684 JL |
800 | done, only to recreate it later if this test was successful. |
801 | ||
802 | Make sure to get the mode from the extension (cand->insn). This | |
803 | is different than in the code to emit the copy as we have not | |
804 | modified the defining insn yet. */ | |
0d732cca | 805 | start_sequence (); |
0d732cca | 806 | rtx pat = PATTERN (cand->insn); |
c7ece684 | 807 | rtx new_dst = gen_rtx_REG (GET_MODE (SET_DEST (pat)), |
e533e26c | 808 | REGNO (get_extended_src_reg (SET_SRC (pat)))); |
c7ece684 | 809 | rtx new_src = gen_rtx_REG (GET_MODE (SET_DEST (pat)), |
0d732cca JL |
810 | REGNO (SET_DEST (pat))); |
811 | emit_move_insn (new_dst, new_src); | |
812 | ||
b32d5189 | 813 | rtx_insn *insn = get_insns(); |
0d732cca JL |
814 | end_sequence (); |
815 | if (NEXT_INSN (insn)) | |
816 | return false; | |
817 | if (recog_memoized (insn) == -1) | |
818 | return false; | |
819 | extract_insn (insn); | |
820 | if (!constrain_operands (1)) | |
821 | return false; | |
3c92da90 JL |
822 | } |
823 | ||
824 | ||
6aae324c JJ |
825 | /* If cand->insn has been already modified, update cand->mode to a wider |
826 | mode if possible, or punt. */ | |
827 | if (state->modified[INSN_UID (cand->insn)].kind != EXT_MODIFIED_NONE) | |
828 | { | |
829 | enum machine_mode mode; | |
830 | rtx set; | |
831 | ||
832 | if (state->modified[INSN_UID (cand->insn)].kind | |
833 | != (cand->code == ZERO_EXTEND | |
834 | ? EXT_MODIFIED_ZEXT : EXT_MODIFIED_SEXT) | |
835 | || state->modified[INSN_UID (cand->insn)].mode != cand->mode | |
836 | || (set = single_set (cand->insn)) == NULL_RTX) | |
837 | return false; | |
838 | mode = GET_MODE (SET_DEST (set)); | |
839 | gcc_assert (GET_MODE_SIZE (mode) >= GET_MODE_SIZE (cand->mode)); | |
840 | cand->mode = mode; | |
841 | } | |
842 | ||
87a0ebfd ST |
843 | merge_successful = true; |
844 | ||
845 | /* Go through the defs vector and try to merge all the definitions | |
846 | in this vector. */ | |
9771b263 DN |
847 | state->modified_list.truncate (0); |
848 | FOR_EACH_VEC_ELT (state->defs_list, defs_ix, def_insn) | |
87a0ebfd | 849 | { |
b57cca0b | 850 | if (merge_def_and_ext (cand, def_insn, state)) |
9771b263 | 851 | state->modified_list.safe_push (def_insn); |
87a0ebfd ST |
852 | else |
853 | { | |
854 | merge_successful = false; | |
855 | break; | |
856 | } | |
857 | } | |
858 | ||
859 | /* Now go through the conditional copies vector and try to merge all | |
860 | the copies in this vector. */ | |
87a0ebfd ST |
861 | if (merge_successful) |
862 | { | |
9771b263 | 863 | FOR_EACH_VEC_ELT (state->copies_list, i, def_insn) |
87a0ebfd | 864 | { |
26cd9add | 865 | if (transform_ifelse (cand, def_insn)) |
9771b263 | 866 | state->modified_list.safe_push (def_insn); |
87a0ebfd ST |
867 | else |
868 | { | |
869 | merge_successful = false; | |
870 | break; | |
871 | } | |
872 | } | |
873 | } | |
874 | ||
875 | if (merge_successful) | |
876 | { | |
282bc7b4 EB |
877 | /* Commit the changes here if possible |
878 | FIXME: It's an all-or-nothing scenario. Even if only one definition | |
879 | cannot be merged, we entirely give up. In the future, we should allow | |
880 | extensions to be partially eliminated along those paths where the | |
881 | definitions could be merged. */ | |
87a0ebfd ST |
882 | if (apply_change_group ()) |
883 | { | |
884 | if (dump_file) | |
282bc7b4 | 885 | fprintf (dump_file, "All merges were successful.\n"); |
87a0ebfd | 886 | |
9771b263 | 887 | FOR_EACH_VEC_ELT (state->modified_list, i, def_insn) |
0d732cca JL |
888 | { |
889 | ext_modified *modified = &state->modified[INSN_UID (def_insn)]; | |
890 | if (modified->kind == EXT_MODIFIED_NONE) | |
891 | modified->kind = (cand->code == ZERO_EXTEND ? EXT_MODIFIED_ZEXT | |
892 | : EXT_MODIFIED_SEXT); | |
b57cca0b | 893 | |
0d732cca JL |
894 | if (copy_needed) |
895 | modified->do_not_reextend = 1; | |
896 | } | |
87a0ebfd ST |
897 | return true; |
898 | } | |
899 | else | |
900 | { | |
0acba2b4 EB |
901 | /* Changes need not be cancelled explicitly as apply_change_group |
902 | does it. Print list of definitions in the dump_file for debug | |
26cd9add | 903 | purposes. This extension cannot be deleted. */ |
87a0ebfd ST |
904 | if (dump_file) |
905 | { | |
ca10595c EB |
906 | fprintf (dump_file, |
907 | "Merge cancelled, non-mergeable definitions:\n"); | |
9771b263 | 908 | FOR_EACH_VEC_ELT (state->modified_list, i, def_insn) |
ca10595c | 909 | print_rtl_single (dump_file, def_insn); |
87a0ebfd ST |
910 | } |
911 | } | |
912 | } | |
913 | else | |
914 | { | |
915 | /* Cancel any changes that have been made so far. */ | |
916 | cancel_changes (0); | |
917 | } | |
918 | ||
87a0ebfd ST |
919 | return false; |
920 | } | |
921 | ||
089dacc5 | 922 | /* Add an extension pattern that could be eliminated. */ |
0acba2b4 EB |
923 | |
924 | static void | |
59a0c032 | 925 | add_removable_extension (const_rtx expr, rtx_insn *insn, |
9771b263 | 926 | vec<ext_cand> *insn_list, |
68c8a824 | 927 | unsigned *def_map) |
0acba2b4 | 928 | { |
282bc7b4 EB |
929 | enum rtx_code code; |
930 | enum machine_mode mode; | |
68c8a824 | 931 | unsigned int idx; |
0acba2b4 EB |
932 | rtx src, dest; |
933 | ||
282bc7b4 | 934 | /* We are looking for SET (REG N) (ANY_EXTEND (REG N)). */ |
0acba2b4 EB |
935 | if (GET_CODE (expr) != SET) |
936 | return; | |
937 | ||
938 | src = SET_SRC (expr); | |
282bc7b4 | 939 | code = GET_CODE (src); |
0acba2b4 | 940 | dest = SET_DEST (expr); |
282bc7b4 | 941 | mode = GET_MODE (dest); |
0acba2b4 EB |
942 | |
943 | if (REG_P (dest) | |
282bc7b4 | 944 | && (code == SIGN_EXTEND || code == ZERO_EXTEND) |
3c92da90 | 945 | && REG_P (XEXP (src, 0))) |
0acba2b4 | 946 | { |
282bc7b4 EB |
947 | struct df_link *defs, *def; |
948 | ext_cand *cand; | |
949 | ||
950 | /* First, make sure we can get all the reaching definitions. */ | |
089dacc5 | 951 | defs = get_defs (insn, XEXP (src, 0), NULL); |
282bc7b4 | 952 | if (!defs) |
0acba2b4 | 953 | { |
282bc7b4 EB |
954 | if (dump_file) |
955 | { | |
956 | fprintf (dump_file, "Cannot eliminate extension:\n"); | |
089dacc5 | 957 | print_rtl_single (dump_file, insn); |
282bc7b4 EB |
958 | fprintf (dump_file, " because of missing definition(s)\n"); |
959 | } | |
960 | return; | |
0acba2b4 | 961 | } |
282bc7b4 EB |
962 | |
963 | /* Second, make sure the reaching definitions don't feed another and | |
964 | different extension. FIXME: this obviously can be improved. */ | |
965 | for (def = defs; def; def = def->next) | |
c3284718 | 966 | if ((idx = def_map[INSN_UID (DF_REF_INSN (def->ref))]) |
9771b263 | 967 | && (cand = &(*insn_list)[idx - 1]) |
ca3f371f | 968 | && cand->code != code) |
282bc7b4 EB |
969 | { |
970 | if (dump_file) | |
971 | { | |
972 | fprintf (dump_file, "Cannot eliminate extension:\n"); | |
089dacc5 | 973 | print_rtl_single (dump_file, insn); |
282bc7b4 EB |
974 | fprintf (dump_file, " because of other extension\n"); |
975 | } | |
976 | return; | |
977 | } | |
978 | ||
979 | /* Then add the candidate to the list and insert the reaching definitions | |
980 | into the definition map. */ | |
f32682ca | 981 | ext_cand e = {expr, code, mode, insn}; |
9771b263 DN |
982 | insn_list->safe_push (e); |
983 | idx = insn_list->length (); | |
282bc7b4 EB |
984 | |
985 | for (def = defs; def; def = def->next) | |
c3284718 | 986 | def_map[INSN_UID (DF_REF_INSN (def->ref))] = idx; |
0acba2b4 EB |
987 | } |
988 | } | |
989 | ||
26cd9add | 990 | /* Traverse the instruction stream looking for extensions and return the |
0acba2b4 | 991 | list of candidates. */ |
87a0ebfd | 992 | |
9771b263 | 993 | static vec<ext_cand> |
26cd9add | 994 | find_removable_extensions (void) |
87a0ebfd | 995 | { |
6e1aa848 | 996 | vec<ext_cand> insn_list = vNULL; |
0acba2b4 | 997 | basic_block bb; |
59a0c032 DM |
998 | rtx_insn *insn; |
999 | rtx set; | |
68c8a824 | 1000 | unsigned *def_map = XCNEWVEC (unsigned, max_insn_uid); |
87a0ebfd | 1001 | |
11cd3bed | 1002 | FOR_EACH_BB_FN (bb, cfun) |
0acba2b4 EB |
1003 | FOR_BB_INSNS (bb, insn) |
1004 | { | |
1005 | if (!NONDEBUG_INSN_P (insn)) | |
1006 | continue; | |
87a0ebfd | 1007 | |
089dacc5 JJ |
1008 | set = single_set (insn); |
1009 | if (set == NULL_RTX) | |
1010 | continue; | |
1011 | add_removable_extension (set, insn, &insn_list, def_map); | |
0acba2b4 EB |
1012 | } |
1013 | ||
089dacc5 | 1014 | XDELETEVEC (def_map); |
282bc7b4 | 1015 | |
089dacc5 | 1016 | return insn_list; |
87a0ebfd ST |
1017 | } |
1018 | ||
1019 | /* This is the main function that checks the insn stream for redundant | |
26cd9add | 1020 | extensions and tries to remove them if possible. */ |
87a0ebfd | 1021 | |
282bc7b4 | 1022 | static void |
26cd9add | 1023 | find_and_remove_re (void) |
87a0ebfd | 1024 | { |
282bc7b4 | 1025 | ext_cand *curr_cand; |
59a0c032 | 1026 | rtx_insn *curr_insn = NULL; |
282bc7b4 | 1027 | int num_re_opportunities = 0, num_realized = 0, i; |
9771b263 | 1028 | vec<ext_cand> reinsn_list; |
59a0c032 DM |
1029 | auto_vec<rtx_insn *> reinsn_del_list; |
1030 | auto_vec<rtx_insn *> reinsn_copy_list; | |
b57cca0b | 1031 | ext_state state; |
87a0ebfd ST |
1032 | |
1033 | /* Construct DU chain to get all reaching definitions of each | |
26cd9add | 1034 | extension instruction. */ |
7b19209f | 1035 | df_set_flags (DF_RD_PRUNE_DEAD_DEFS); |
87a0ebfd ST |
1036 | df_chain_add_problem (DF_UD_CHAIN + DF_DU_CHAIN); |
1037 | df_analyze (); | |
b57cca0b | 1038 | df_set_flags (DF_DEFER_INSN_RESCAN); |
87a0ebfd ST |
1039 | |
1040 | max_insn_uid = get_max_uid (); | |
26cd9add | 1041 | reinsn_list = find_removable_extensions (); |
9771b263 DN |
1042 | state.defs_list.create (0); |
1043 | state.copies_list.create (0); | |
1044 | state.modified_list.create (0); | |
1045 | state.work_list.create (0); | |
1046 | if (reinsn_list.is_empty ()) | |
b57cca0b JJ |
1047 | state.modified = NULL; |
1048 | else | |
1049 | state.modified = XCNEWVEC (struct ext_modified, max_insn_uid); | |
87a0ebfd | 1050 | |
9771b263 | 1051 | FOR_EACH_VEC_ELT (reinsn_list, i, curr_cand) |
87a0ebfd | 1052 | { |
26cd9add | 1053 | num_re_opportunities++; |
87a0ebfd | 1054 | |
282bc7b4 | 1055 | /* Try to combine the extension with the definition. */ |
87a0ebfd ST |
1056 | if (dump_file) |
1057 | { | |
282bc7b4 EB |
1058 | fprintf (dump_file, "Trying to eliminate extension:\n"); |
1059 | print_rtl_single (dump_file, curr_cand->insn); | |
87a0ebfd ST |
1060 | } |
1061 | ||
089dacc5 | 1062 | if (combine_reaching_defs (curr_cand, curr_cand->expr, &state)) |
87a0ebfd ST |
1063 | { |
1064 | if (dump_file) | |
282bc7b4 | 1065 | fprintf (dump_file, "Eliminated the extension.\n"); |
87a0ebfd | 1066 | num_realized++; |
059742a4 JL |
1067 | /* If the RHS of the current candidate is not (extend (reg)), then |
1068 | we do not allow the optimization of extensions where | |
1069 | the source and destination registers do not match. Thus | |
1070 | checking REG_P here is correct. */ | |
1071 | if (REG_P (XEXP (SET_SRC (PATTERN (curr_cand->insn)), 0)) | |
1072 | && (REGNO (SET_DEST (PATTERN (curr_cand->insn))) | |
1073 | != REGNO (XEXP (SET_SRC (PATTERN (curr_cand->insn)), 0)))) | |
3c92da90 JL |
1074 | { |
1075 | reinsn_copy_list.safe_push (curr_cand->insn); | |
1076 | reinsn_copy_list.safe_push (state.defs_list[0]); | |
1077 | } | |
1078 | reinsn_del_list.safe_push (curr_cand->insn); | |
b57cca0b | 1079 | state.modified[INSN_UID (curr_cand->insn)].deleted = 1; |
87a0ebfd ST |
1080 | } |
1081 | } | |
1082 | ||
3c92da90 JL |
1083 | /* The copy list contains pairs of insns which describe copies we |
1084 | need to insert into the INSN stream. | |
1085 | ||
1086 | The first insn in each pair is the extension insn, from which | |
1087 | we derive the source and destination of the copy. | |
1088 | ||
1089 | The second insn in each pair is the memory reference where the | |
1090 | extension will ultimately happen. We emit the new copy | |
1091 | immediately after this insn. | |
1092 | ||
1093 | It may first appear that the arguments for the copy are reversed. | |
1094 | Remember that the memory reference will be changed to refer to the | |
1095 | destination of the extention. So we're actually emitting a copy | |
1096 | from the new destination to the old destination. */ | |
1097 | for (unsigned int i = 0; i < reinsn_copy_list.length (); i += 2) | |
1098 | { | |
59a0c032 DM |
1099 | rtx_insn *curr_insn = reinsn_copy_list[i]; |
1100 | rtx_insn *def_insn = reinsn_copy_list[i + 1]; | |
a6a2d67b JL |
1101 | |
1102 | /* Use the mode of the destination of the defining insn | |
1103 | for the mode of the copy. This is necessary if the | |
1104 | defining insn was used to eliminate a second extension | |
1105 | that was wider than the first. */ | |
1106 | rtx sub_rtx = *get_sub_rtx (def_insn); | |
3c92da90 | 1107 | rtx pat = PATTERN (curr_insn); |
a6a2d67b | 1108 | rtx new_dst = gen_rtx_REG (GET_MODE (SET_DEST (sub_rtx)), |
3c92da90 | 1109 | REGNO (XEXP (SET_SRC (pat), 0))); |
a6a2d67b JL |
1110 | rtx new_src = gen_rtx_REG (GET_MODE (SET_DEST (sub_rtx)), |
1111 | REGNO (SET_DEST (pat))); | |
1112 | rtx set = gen_rtx_SET (VOIDmode, new_dst, new_src); | |
1113 | emit_insn_after (set, def_insn); | |
3c92da90 JL |
1114 | } |
1115 | ||
26cd9add | 1116 | /* Delete all useless extensions here in one sweep. */ |
9771b263 | 1117 | FOR_EACH_VEC_ELT (reinsn_del_list, i, curr_insn) |
0acba2b4 | 1118 | delete_insn (curr_insn); |
87a0ebfd | 1119 | |
9771b263 | 1120 | reinsn_list.release (); |
9771b263 DN |
1121 | state.defs_list.release (); |
1122 | state.copies_list.release (); | |
1123 | state.modified_list.release (); | |
1124 | state.work_list.release (); | |
b57cca0b | 1125 | XDELETEVEC (state.modified); |
87a0ebfd | 1126 | |
26cd9add | 1127 | if (dump_file && num_re_opportunities > 0) |
282bc7b4 EB |
1128 | fprintf (dump_file, "Elimination opportunities = %d realized = %d\n", |
1129 | num_re_opportunities, num_realized); | |
87a0ebfd ST |
1130 | } |
1131 | ||
26cd9add | 1132 | /* Find and remove redundant extensions. */ |
87a0ebfd ST |
1133 | |
1134 | static unsigned int | |
26cd9add | 1135 | rest_of_handle_ree (void) |
87a0ebfd | 1136 | { |
26cd9add EI |
1137 | timevar_push (TV_REE); |
1138 | find_and_remove_re (); | |
1139 | timevar_pop (TV_REE); | |
87a0ebfd ST |
1140 | return 0; |
1141 | } | |
1142 | ||
27a4cd48 DM |
1143 | namespace { |
1144 | ||
1145 | const pass_data pass_data_ree = | |
87a0ebfd | 1146 | { |
27a4cd48 DM |
1147 | RTL_PASS, /* type */ |
1148 | "ree", /* name */ | |
1149 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1150 | TV_REE, /* tv_id */ |
1151 | 0, /* properties_required */ | |
1152 | 0, /* properties_provided */ | |
1153 | 0, /* properties_destroyed */ | |
1154 | 0, /* todo_flags_start */ | |
3bea341f | 1155 | TODO_df_finish, /* todo_flags_finish */ |
87a0ebfd | 1156 | }; |
27a4cd48 DM |
1157 | |
1158 | class pass_ree : public rtl_opt_pass | |
1159 | { | |
1160 | public: | |
c3284718 RS |
1161 | pass_ree (gcc::context *ctxt) |
1162 | : rtl_opt_pass (pass_data_ree, ctxt) | |
27a4cd48 DM |
1163 | {} |
1164 | ||
1165 | /* opt_pass methods: */ | |
1a3d085c | 1166 | virtual bool gate (function *) { return (optimize > 0 && flag_ree); } |
be55bfe6 | 1167 | virtual unsigned int execute (function *) { return rest_of_handle_ree (); } |
27a4cd48 DM |
1168 | |
1169 | }; // class pass_ree | |
1170 | ||
1171 | } // anon namespace | |
1172 | ||
1173 | rtl_opt_pass * | |
1174 | make_pass_ree (gcc::context *ctxt) | |
1175 | { | |
1176 | return new pass_ree (ctxt); | |
1177 | } |