1 /* The Blackfin code generation auxiliary output file.
2 Copyright (C) 2005, 2006, 2007, 2008, 2009 Free Software Foundation, Inc.
3 Contributed by Analog Devices.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published
9 by the Free Software Foundation; either version 3, or (at your
10 option) any later version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
15 License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
27 #include "hard-reg-set.h"
29 #include "insn-config.h"
30 #include "insn-codes.h"
31 #include "conditions.h"
32 #include "insn-flags.h"
34 #include "insn-attr.h"
41 #include "target-def.h"
47 #include "integrate.h"
49 #include "langhooks.h"
50 #include "bfin-protos.h"
52 #include "tm-constrs.h"
54 #include "basic-block.h"
55 #include "cfglayout.h"
59 /* A C structure for machine-specific, per-function data.
60 This is added to the cfun structure. */
61 struct GTY(()) machine_function
63 /* Set if we are notified by the doloop pass that a hardware loop
65 int has_hardware_loops;
66 /* Set if we create a memcpy pattern that uses loop registers. */
67 int has_loopreg_clobber;
70 /* RTX for condition code flag register and RETS register */
71 extern GTY(()) rtx bfin_cc_rtx;
72 extern GTY(()) rtx bfin_rets_rtx;
73 rtx bfin_cc_rtx, bfin_rets_rtx;
75 int max_arg_registers = 0;
77 /* Arrays used when emitting register names. */
78 const char *short_reg_names[] = SHORT_REGISTER_NAMES;
79 const char *high_reg_names[] = HIGH_REGISTER_NAMES;
80 const char *dregs_pair_names[] = DREGS_PAIR_NAMES;
81 const char *byte_reg_names[] = BYTE_REGISTER_NAMES;
83 static int arg_regs[] = FUNCTION_ARG_REGISTERS;
85 /* Nonzero if -mshared-library-id was given. */
86 static int bfin_lib_id_given;
88 /* Nonzero if -fschedule-insns2 was given. We override it and
89 call the scheduler ourselves during reorg. */
90 static int bfin_flag_schedule_insns2;
92 /* Determines whether we run variable tracking in machine dependent
94 static int bfin_flag_var_tracking;
97 bfin_cpu_t bfin_cpu_type = BFIN_CPU_UNKNOWN;
99 /* -msi-revision support. There are three special values:
100 -1 -msi-revision=none.
101 0xffff -msi-revision=any. */
102 int bfin_si_revision;
104 /* The workarounds enabled */
105 unsigned int bfin_workarounds = 0;
112 unsigned int workarounds;
115 struct bfin_cpu bfin_cpus[] =
117 {"bf512", BFIN_CPU_BF512, 0x0000,
118 WA_SPECULATIVE_LOADS},
120 {"bf514", BFIN_CPU_BF514, 0x0000,
121 WA_SPECULATIVE_LOADS},
123 {"bf516", BFIN_CPU_BF516, 0x0000,
124 WA_SPECULATIVE_LOADS},
126 {"bf518", BFIN_CPU_BF518, 0x0000,
127 WA_SPECULATIVE_LOADS},
129 {"bf522", BFIN_CPU_BF522, 0x0002,
130 WA_SPECULATIVE_LOADS},
131 {"bf522", BFIN_CPU_BF522, 0x0001,
132 WA_SPECULATIVE_LOADS | WA_RETS},
133 {"bf522", BFIN_CPU_BF522, 0x0000,
134 WA_SPECULATIVE_LOADS | WA_RETS},
136 {"bf523", BFIN_CPU_BF523, 0x0002,
137 WA_SPECULATIVE_LOADS},
138 {"bf523", BFIN_CPU_BF523, 0x0001,
139 WA_SPECULATIVE_LOADS | WA_RETS},
140 {"bf523", BFIN_CPU_BF523, 0x0000,
141 WA_SPECULATIVE_LOADS | WA_RETS},
143 {"bf524", BFIN_CPU_BF524, 0x0002,
144 WA_SPECULATIVE_LOADS},
145 {"bf524", BFIN_CPU_BF524, 0x0001,
146 WA_SPECULATIVE_LOADS | WA_RETS},
147 {"bf524", BFIN_CPU_BF524, 0x0000,
148 WA_SPECULATIVE_LOADS | WA_RETS},
150 {"bf525", BFIN_CPU_BF525, 0x0002,
151 WA_SPECULATIVE_LOADS},
152 {"bf525", BFIN_CPU_BF525, 0x0001,
153 WA_SPECULATIVE_LOADS | WA_RETS},
154 {"bf525", BFIN_CPU_BF525, 0x0000,
155 WA_SPECULATIVE_LOADS | WA_RETS},
157 {"bf526", BFIN_CPU_BF526, 0x0002,
158 WA_SPECULATIVE_LOADS},
159 {"bf526", BFIN_CPU_BF526, 0x0001,
160 WA_SPECULATIVE_LOADS | WA_RETS},
161 {"bf526", BFIN_CPU_BF526, 0x0000,
162 WA_SPECULATIVE_LOADS | WA_RETS},
164 {"bf527", BFIN_CPU_BF527, 0x0002,
165 WA_SPECULATIVE_LOADS},
166 {"bf527", BFIN_CPU_BF527, 0x0001,
167 WA_SPECULATIVE_LOADS | WA_RETS},
168 {"bf527", BFIN_CPU_BF527, 0x0000,
169 WA_SPECULATIVE_LOADS | WA_RETS},
171 {"bf531", BFIN_CPU_BF531, 0x0006,
172 WA_SPECULATIVE_LOADS | WA_LOAD_LCREGS},
173 {"bf531", BFIN_CPU_BF531, 0x0005,
174 WA_SPECULATIVE_LOADS | WA_RETS | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
175 {"bf531", BFIN_CPU_BF531, 0x0004,
176 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
177 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
178 {"bf531", BFIN_CPU_BF531, 0x0003,
179 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
180 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
182 {"bf532", BFIN_CPU_BF532, 0x0006,
183 WA_SPECULATIVE_LOADS | WA_LOAD_LCREGS},
184 {"bf532", BFIN_CPU_BF532, 0x0005,
185 WA_SPECULATIVE_LOADS | WA_RETS | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
186 {"bf532", BFIN_CPU_BF532, 0x0004,
187 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
188 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
189 {"bf532", BFIN_CPU_BF532, 0x0003,
190 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
191 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
193 {"bf533", BFIN_CPU_BF533, 0x0006,
194 WA_SPECULATIVE_LOADS | WA_LOAD_LCREGS},
195 {"bf533", BFIN_CPU_BF533, 0x0005,
196 WA_SPECULATIVE_LOADS | WA_RETS | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
197 {"bf533", BFIN_CPU_BF533, 0x0004,
198 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
199 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
200 {"bf533", BFIN_CPU_BF533, 0x0003,
201 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
202 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
204 {"bf534", BFIN_CPU_BF534, 0x0003,
205 WA_SPECULATIVE_LOADS | WA_RETS | WA_LOAD_LCREGS},
206 {"bf534", BFIN_CPU_BF534, 0x0002,
207 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
208 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
209 {"bf534", BFIN_CPU_BF534, 0x0001,
210 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
211 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
213 {"bf536", BFIN_CPU_BF536, 0x0003,
214 WA_SPECULATIVE_LOADS | WA_RETS | WA_LOAD_LCREGS},
215 {"bf536", BFIN_CPU_BF536, 0x0002,
216 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
217 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
218 {"bf536", BFIN_CPU_BF536, 0x0001,
219 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
220 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
222 {"bf537", BFIN_CPU_BF537, 0x0003,
223 WA_SPECULATIVE_LOADS | WA_RETS | WA_LOAD_LCREGS},
224 {"bf537", BFIN_CPU_BF537, 0x0002,
225 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
226 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
227 {"bf537", BFIN_CPU_BF537, 0x0001,
228 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
229 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
231 {"bf538", BFIN_CPU_BF538, 0x0005,
232 WA_SPECULATIVE_LOADS | WA_LOAD_LCREGS},
233 {"bf538", BFIN_CPU_BF538, 0x0004,
234 WA_SPECULATIVE_LOADS | WA_RETS | WA_LOAD_LCREGS},
235 {"bf538", BFIN_CPU_BF538, 0x0003,
236 WA_SPECULATIVE_LOADS | WA_RETS
237 | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
238 {"bf538", BFIN_CPU_BF538, 0x0002,
239 WA_SPECULATIVE_LOADS | WA_RETS
240 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
242 {"bf539", BFIN_CPU_BF539, 0x0005,
243 WA_SPECULATIVE_LOADS | WA_LOAD_LCREGS},
244 {"bf539", BFIN_CPU_BF539, 0x0004,
245 WA_SPECULATIVE_LOADS | WA_RETS | WA_LOAD_LCREGS},
246 {"bf539", BFIN_CPU_BF539, 0x0003,
247 WA_SPECULATIVE_LOADS | WA_RETS
248 | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
249 {"bf539", BFIN_CPU_BF539, 0x0002,
250 WA_SPECULATIVE_LOADS | WA_RETS
251 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
253 {"bf542", BFIN_CPU_BF542, 0x0002,
254 WA_SPECULATIVE_LOADS | WA_INDIRECT_CALLS},
255 {"bf542", BFIN_CPU_BF542, 0x0001,
256 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS},
257 {"bf542", BFIN_CPU_BF542, 0x0000,
258 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS | WA_LOAD_LCREGS},
260 {"bf544", BFIN_CPU_BF544, 0x0002,
261 WA_SPECULATIVE_LOADS | WA_INDIRECT_CALLS},
262 {"bf544", BFIN_CPU_BF544, 0x0001,
263 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS},
264 {"bf544", BFIN_CPU_BF544, 0x0000,
265 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS | WA_LOAD_LCREGS},
267 {"bf547", BFIN_CPU_BF547, 0x0002,
268 WA_SPECULATIVE_LOADS | WA_INDIRECT_CALLS},
269 {"bf547", BFIN_CPU_BF547, 0x0001,
270 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS},
271 {"bf547", BFIN_CPU_BF547, 0x0000,
272 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS | WA_LOAD_LCREGS},
274 {"bf548", BFIN_CPU_BF548, 0x0002,
275 WA_SPECULATIVE_LOADS | WA_INDIRECT_CALLS},
276 {"bf548", BFIN_CPU_BF548, 0x0001,
277 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS},
278 {"bf548", BFIN_CPU_BF548, 0x0000,
279 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS | WA_LOAD_LCREGS},
281 {"bf549", BFIN_CPU_BF549, 0x0002,
282 WA_SPECULATIVE_LOADS | WA_INDIRECT_CALLS},
283 {"bf549", BFIN_CPU_BF549, 0x0001,
284 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS},
285 {"bf549", BFIN_CPU_BF549, 0x0000,
286 WA_SPECULATIVE_LOADS | WA_RETS | WA_INDIRECT_CALLS | WA_LOAD_LCREGS},
288 {"bf561", BFIN_CPU_BF561, 0x0005, WA_RETS
289 | WA_05000283 | WA_05000315 | WA_LOAD_LCREGS},
290 {"bf561", BFIN_CPU_BF561, 0x0003,
291 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
292 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
293 {"bf561", BFIN_CPU_BF561, 0x0002,
294 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS
295 | WA_05000283 | WA_05000257 | WA_05000315 | WA_LOAD_LCREGS},
300 int splitting_for_sched, splitting_loops;
303 bfin_globalize_label (FILE *stream, const char *name)
305 fputs (".global ", stream);
306 assemble_name (stream, name);
312 output_file_start (void)
314 FILE *file = asm_out_file;
317 /* Variable tracking should be run after all optimizations which change order
318 of insns. It also needs a valid CFG. This can't be done in
319 override_options, because flag_var_tracking is finalized after
321 bfin_flag_var_tracking = flag_var_tracking;
322 flag_var_tracking = 0;
324 fprintf (file, ".file \"%s\";\n", input_filename);
326 for (i = 0; arg_regs[i] >= 0; i++)
328 max_arg_registers = i; /* how many arg reg used */
331 /* Called early in the compilation to conditionally modify
332 fixed_regs/call_used_regs. */
335 conditional_register_usage (void)
337 /* initialize condition code flag register rtx */
338 bfin_cc_rtx = gen_rtx_REG (BImode, REG_CC);
339 bfin_rets_rtx = gen_rtx_REG (Pmode, REG_RETS);
342 /* Examine machine-dependent attributes of function type FUNTYPE and return its
343 type. See the definition of E_FUNKIND. */
346 funkind (const_tree funtype)
348 tree attrs = TYPE_ATTRIBUTES (funtype);
349 if (lookup_attribute ("interrupt_handler", attrs))
350 return INTERRUPT_HANDLER;
351 else if (lookup_attribute ("exception_handler", attrs))
352 return EXCPT_HANDLER;
353 else if (lookup_attribute ("nmi_handler", attrs))
359 /* Legitimize PIC addresses. If the address is already position-independent,
360 we return ORIG. Newly generated position-independent addresses go into a
361 reg. This is REG if nonzero, otherwise we allocate register(s) as
362 necessary. PICREG is the register holding the pointer to the PIC offset
366 legitimize_pic_address (rtx orig, rtx reg, rtx picreg)
371 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
376 if (TARGET_ID_SHARED_LIBRARY)
377 unspec = UNSPEC_MOVE_PIC;
378 else if (GET_CODE (addr) == SYMBOL_REF
379 && SYMBOL_REF_FUNCTION_P (addr))
380 unspec = UNSPEC_FUNCDESC_GOT17M4;
382 unspec = UNSPEC_MOVE_FDPIC;
386 gcc_assert (can_create_pseudo_p ());
387 reg = gen_reg_rtx (Pmode);
390 tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), unspec);
391 new_rtx = gen_const_mem (Pmode, gen_rtx_PLUS (Pmode, picreg, tmp));
393 emit_move_insn (reg, new_rtx);
394 if (picreg == pic_offset_table_rtx)
395 crtl->uses_pic_offset_table = 1;
399 else if (GET_CODE (addr) == CONST || GET_CODE (addr) == PLUS)
403 if (GET_CODE (addr) == CONST)
405 addr = XEXP (addr, 0);
406 gcc_assert (GET_CODE (addr) == PLUS);
409 if (XEXP (addr, 0) == picreg)
414 gcc_assert (can_create_pseudo_p ());
415 reg = gen_reg_rtx (Pmode);
418 base = legitimize_pic_address (XEXP (addr, 0), reg, picreg);
419 addr = legitimize_pic_address (XEXP (addr, 1),
420 base == reg ? NULL_RTX : reg,
423 if (GET_CODE (addr) == CONST_INT)
425 gcc_assert (! reload_in_progress && ! reload_completed);
426 addr = force_reg (Pmode, addr);
429 if (GET_CODE (addr) == PLUS && CONSTANT_P (XEXP (addr, 1)))
431 base = gen_rtx_PLUS (Pmode, base, XEXP (addr, 0));
432 addr = XEXP (addr, 1);
435 return gen_rtx_PLUS (Pmode, base, addr);
441 /* Stack frame layout. */
443 /* For a given REGNO, determine whether it must be saved in the function
444 prologue. IS_INTHANDLER specifies whether we're generating a normal
445 prologue or an interrupt/exception one. */
447 must_save_p (bool is_inthandler, unsigned regno)
449 if (D_REGNO_P (regno))
451 bool is_eh_return_reg = false;
452 if (crtl->calls_eh_return)
457 unsigned test = EH_RETURN_DATA_REGNO (j);
458 if (test == INVALID_REGNUM)
461 is_eh_return_reg = true;
465 return (is_eh_return_reg
466 || (df_regs_ever_live_p (regno)
467 && !fixed_regs[regno]
468 && (is_inthandler || !call_used_regs[regno])));
470 else if (P_REGNO_P (regno))
472 return ((df_regs_ever_live_p (regno)
473 && !fixed_regs[regno]
474 && (is_inthandler || !call_used_regs[regno]))
476 && (ENABLE_WA_05000283 || ENABLE_WA_05000315)
479 && regno == PIC_OFFSET_TABLE_REGNUM
480 && (crtl->uses_pic_offset_table
481 || (TARGET_ID_SHARED_LIBRARY && !current_function_is_leaf))));
484 return ((is_inthandler || !call_used_regs[regno])
485 && (df_regs_ever_live_p (regno)
486 || (!leaf_function_p () && call_used_regs[regno])));
490 /* Compute the number of DREGS to save with a push_multiple operation.
491 This could include registers that aren't modified in the function,
492 since push_multiple only takes a range of registers.
493 If IS_INTHANDLER, then everything that is live must be saved, even
494 if normally call-clobbered.
495 If CONSECUTIVE, return the number of registers we can save in one
496 instruction with a push/pop multiple instruction. */
499 n_dregs_to_save (bool is_inthandler, bool consecutive)
504 for (i = REG_R7 + 1; i-- != REG_R0;)
506 if (must_save_p (is_inthandler, i))
508 else if (consecutive)
514 /* Like n_dregs_to_save, but compute number of PREGS to save. */
517 n_pregs_to_save (bool is_inthandler, bool consecutive)
522 for (i = REG_P5 + 1; i-- != REG_P0;)
523 if (must_save_p (is_inthandler, i))
525 else if (consecutive)
530 /* Determine if we are going to save the frame pointer in the prologue. */
533 must_save_fp_p (void)
535 return frame_pointer_needed || df_regs_ever_live_p (REG_FP);
539 stack_frame_needed_p (void)
541 /* EH return puts a new return address into the frame using an
542 address relative to the frame pointer. */
543 if (crtl->calls_eh_return)
545 return frame_pointer_needed;
548 /* Emit code to save registers in the prologue. SAVEALL is nonzero if we
549 must save all registers; this is used for interrupt handlers.
550 SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
551 this for an interrupt (or exception) handler. */
554 expand_prologue_reg_save (rtx spreg, int saveall, bool is_inthandler)
556 rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
557 rtx predec = gen_rtx_MEM (SImode, predec1);
558 int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
559 int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
560 int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
561 int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
563 int total_consec = ndregs_consec + npregs_consec;
566 if (saveall || is_inthandler)
568 rtx insn = emit_move_insn (predec, gen_rtx_REG (SImode, REG_ASTAT));
570 RTX_FRAME_RELATED_P (insn) = 1;
571 for (dregno = REG_LT0; dregno <= REG_LB1; dregno++)
572 if (! current_function_is_leaf
573 || cfun->machine->has_hardware_loops
574 || cfun->machine->has_loopreg_clobber
575 || (ENABLE_WA_05000257
576 && (dregno == REG_LC0 || dregno == REG_LC1)))
578 insn = emit_move_insn (predec, gen_rtx_REG (SImode, dregno));
579 RTX_FRAME_RELATED_P (insn) = 1;
583 if (total_consec != 0)
586 rtx val = GEN_INT (-total_consec * 4);
587 rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 2));
589 XVECEXP (pat, 0, 0) = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, val),
590 UNSPEC_PUSH_MULTIPLE);
591 XVECEXP (pat, 0, total_consec + 1) = gen_rtx_SET (VOIDmode, spreg,
595 RTX_FRAME_RELATED_P (XVECEXP (pat, 0, total_consec + 1)) = 1;
596 d_to_save = ndregs_consec;
597 dregno = REG_R7 + 1 - ndregs_consec;
598 pregno = REG_P5 + 1 - npregs_consec;
599 for (i = 0; i < total_consec; i++)
601 rtx memref = gen_rtx_MEM (word_mode,
602 gen_rtx_PLUS (Pmode, spreg,
603 GEN_INT (- i * 4 - 4)));
607 subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
613 subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
616 XVECEXP (pat, 0, i + 1) = subpat;
617 RTX_FRAME_RELATED_P (subpat) = 1;
619 insn = emit_insn (pat);
620 RTX_FRAME_RELATED_P (insn) = 1;
623 for (dregno = REG_R0; ndregs != ndregs_consec; dregno++)
625 if (must_save_p (is_inthandler, dregno))
627 rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, dregno));
628 RTX_FRAME_RELATED_P (insn) = 1;
632 for (pregno = REG_P0; npregs != npregs_consec; pregno++)
634 if (must_save_p (is_inthandler, pregno))
636 rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, pregno));
637 RTX_FRAME_RELATED_P (insn) = 1;
641 for (i = REG_P7 + 1; i < REG_CC; i++)
644 && (df_regs_ever_live_p (i)
645 || (!leaf_function_p () && call_used_regs[i]))))
648 if (i == REG_A0 || i == REG_A1)
649 insn = emit_move_insn (gen_rtx_MEM (PDImode, predec1),
650 gen_rtx_REG (PDImode, i));
652 insn = emit_move_insn (predec, gen_rtx_REG (SImode, i));
653 RTX_FRAME_RELATED_P (insn) = 1;
657 /* Emit code to restore registers in the epilogue. SAVEALL is nonzero if we
658 must save all registers; this is used for interrupt handlers.
659 SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
660 this for an interrupt (or exception) handler. */
663 expand_epilogue_reg_restore (rtx spreg, bool saveall, bool is_inthandler)
665 rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
666 rtx postinc = gen_rtx_MEM (SImode, postinc1);
668 int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
669 int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
670 int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
671 int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
672 int total_consec = ndregs_consec + npregs_consec;
676 /* A slightly crude technique to stop flow from trying to delete "dead"
678 MEM_VOLATILE_P (postinc) = 1;
680 for (i = REG_CC - 1; i > REG_P7; i--)
683 && (df_regs_ever_live_p (i)
684 || (!leaf_function_p () && call_used_regs[i]))))
686 if (i == REG_A0 || i == REG_A1)
688 rtx mem = gen_rtx_MEM (PDImode, postinc1);
689 MEM_VOLATILE_P (mem) = 1;
690 emit_move_insn (gen_rtx_REG (PDImode, i), mem);
693 emit_move_insn (gen_rtx_REG (SImode, i), postinc);
696 regno = REG_P5 - npregs_consec;
697 for (; npregs != npregs_consec; regno--)
699 if (must_save_p (is_inthandler, regno))
701 emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
705 regno = REG_R7 - ndregs_consec;
706 for (; ndregs != ndregs_consec; regno--)
708 if (must_save_p (is_inthandler, regno))
710 emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
715 if (total_consec != 0)
717 rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 1));
719 = gen_rtx_SET (VOIDmode, spreg,
720 gen_rtx_PLUS (Pmode, spreg,
721 GEN_INT (total_consec * 4)));
723 if (npregs_consec > 0)
728 for (i = 0; i < total_consec; i++)
731 ? gen_rtx_PLUS (Pmode, spreg, GEN_INT (i * 4))
733 rtx memref = gen_rtx_MEM (word_mode, addr);
736 XVECEXP (pat, 0, i + 1)
737 = gen_rtx_SET (VOIDmode, gen_rtx_REG (word_mode, regno), memref);
739 if (npregs_consec > 0)
741 if (--npregs_consec == 0)
746 insn = emit_insn (pat);
747 RTX_FRAME_RELATED_P (insn) = 1;
749 if (saveall || is_inthandler)
751 for (regno = REG_LB1; regno >= REG_LT0; regno--)
752 if (! current_function_is_leaf
753 || cfun->machine->has_hardware_loops
754 || cfun->machine->has_loopreg_clobber
755 || (ENABLE_WA_05000257 && (regno == REG_LC0 || regno == REG_LC1)))
756 emit_move_insn (gen_rtx_REG (SImode, regno), postinc);
758 emit_move_insn (gen_rtx_REG (SImode, REG_ASTAT), postinc);
762 /* Perform any needed actions needed for a function that is receiving a
763 variable number of arguments.
767 MODE and TYPE are the mode and type of the current parameter.
769 PRETEND_SIZE is a variable that should be set to the amount of stack
770 that must be pushed by the prolog to pretend that our caller pushed
773 Normally, this macro will push all remaining incoming registers on the
774 stack and set PRETEND_SIZE to the length of the registers pushed.
777 - VDSP C compiler manual (our ABI) says that a variable args function
778 should save the R0, R1 and R2 registers in the stack.
779 - The caller will always leave space on the stack for the
780 arguments that are passed in registers, so we dont have
781 to leave any extra space.
782 - now, the vastart pointer can access all arguments from the stack. */
785 setup_incoming_varargs (CUMULATIVE_ARGS *cum,
786 enum machine_mode mode ATTRIBUTE_UNUSED,
787 tree type ATTRIBUTE_UNUSED, int *pretend_size,
796 /* The move for named arguments will be generated automatically by the
797 compiler. We need to generate the move rtx for the unnamed arguments
798 if they are in the first 3 words. We assume at least 1 named argument
799 exists, so we never generate [ARGP] = R0 here. */
801 for (i = cum->words + 1; i < max_arg_registers; i++)
803 mem = gen_rtx_MEM (Pmode,
804 plus_constant (arg_pointer_rtx, (i * UNITS_PER_WORD)));
805 emit_move_insn (mem, gen_rtx_REG (Pmode, i));
811 /* Value should be nonzero if functions must have frame pointers.
812 Zero means the frame pointer need not be set up (and parms may
813 be accessed via the stack pointer) in functions that seem suitable. */
816 bfin_frame_pointer_required (void)
818 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
820 if (fkind != SUBROUTINE)
823 /* We turn on -fomit-frame-pointer if -momit-leaf-frame-pointer is used,
824 so we have to override it for non-leaf functions. */
825 if (TARGET_OMIT_LEAF_FRAME_POINTER && ! current_function_is_leaf)
831 /* Return the number of registers pushed during the prologue. */
834 n_regs_saved_by_prologue (void)
836 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
837 bool is_inthandler = fkind != SUBROUTINE;
838 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
839 bool all = (lookup_attribute ("saveall", attrs) != NULL_TREE
840 || (is_inthandler && !current_function_is_leaf));
841 int ndregs = all ? 8 : n_dregs_to_save (is_inthandler, false);
842 int npregs = all ? 6 : n_pregs_to_save (is_inthandler, false);
843 int n = ndregs + npregs;
846 if (all || stack_frame_needed_p ())
847 /* We use a LINK instruction in this case. */
851 if (must_save_fp_p ())
853 if (! current_function_is_leaf)
857 if (fkind != SUBROUTINE || all)
859 /* Increment once for ASTAT. */
861 if (! current_function_is_leaf
862 || cfun->machine->has_hardware_loops
863 || cfun->machine->has_loopreg_clobber)
869 if (fkind != SUBROUTINE)
872 if (lookup_attribute ("nesting", attrs))
876 for (i = REG_P7 + 1; i < REG_CC; i++)
878 || (fkind != SUBROUTINE
879 && (df_regs_ever_live_p (i)
880 || (!leaf_function_p () && call_used_regs[i]))))
881 n += i == REG_A0 || i == REG_A1 ? 2 : 1;
886 /* Return the offset between two registers, one to be eliminated, and the other
887 its replacement, at the start of a routine. */
890 bfin_initial_elimination_offset (int from, int to)
892 HOST_WIDE_INT offset = 0;
894 if (from == ARG_POINTER_REGNUM)
895 offset = n_regs_saved_by_prologue () * 4;
897 if (to == STACK_POINTER_REGNUM)
899 if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
900 offset += crtl->outgoing_args_size;
901 else if (crtl->outgoing_args_size)
902 offset += FIXED_STACK_AREA;
904 offset += get_frame_size ();
910 /* Emit code to load a constant CONSTANT into register REG; setting
911 RTX_FRAME_RELATED_P on all insns we generate if RELATED is true.
912 Make sure that the insns we generate need not be split. */
915 frame_related_constant_load (rtx reg, HOST_WIDE_INT constant, bool related)
918 rtx cst = GEN_INT (constant);
920 if (constant >= -32768 && constant < 65536)
921 insn = emit_move_insn (reg, cst);
924 /* We don't call split_load_immediate here, since dwarf2out.c can get
925 confused about some of the more clever sequences it can generate. */
926 insn = emit_insn (gen_movsi_high (reg, cst));
928 RTX_FRAME_RELATED_P (insn) = 1;
929 insn = emit_insn (gen_movsi_low (reg, reg, cst));
932 RTX_FRAME_RELATED_P (insn) = 1;
935 /* Generate efficient code to add a value to a P register.
936 Set RTX_FRAME_RELATED_P on the generated insns if FRAME is nonzero.
937 EPILOGUE_P is zero if this function is called for prologue,
938 otherwise it's nonzero. And it's less than zero if this is for
942 add_to_reg (rtx reg, HOST_WIDE_INT value, int frame, int epilogue_p)
947 /* Choose whether to use a sequence using a temporary register, or
948 a sequence with multiple adds. We can add a signed 7-bit value
949 in one instruction. */
950 if (value > 120 || value < -120)
958 /* For prologue or normal epilogue, P1 can be safely used
959 as the temporary register. For sibcall epilogue, we try to find
960 a call used P register, which will be restored in epilogue.
961 If we cannot find such a P register, we have to use one I register
965 tmpreg = gen_rtx_REG (SImode, REG_P1);
969 for (i = REG_P0; i <= REG_P5; i++)
970 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
972 && i == PIC_OFFSET_TABLE_REGNUM
973 && (crtl->uses_pic_offset_table
974 || (TARGET_ID_SHARED_LIBRARY
975 && ! current_function_is_leaf))))
978 tmpreg = gen_rtx_REG (SImode, i);
981 tmpreg = gen_rtx_REG (SImode, REG_P1);
982 tmpreg2 = gen_rtx_REG (SImode, REG_I0);
983 emit_move_insn (tmpreg2, tmpreg);
988 frame_related_constant_load (tmpreg, value, TRUE);
990 insn = emit_move_insn (tmpreg, GEN_INT (value));
992 insn = emit_insn (gen_addsi3 (reg, reg, tmpreg));
994 RTX_FRAME_RELATED_P (insn) = 1;
996 if (tmpreg2 != NULL_RTX)
997 emit_move_insn (tmpreg, tmpreg2);
1007 else if (size < -60)
1008 /* We could use -62, but that would leave the stack unaligned, so
1012 insn = emit_insn (gen_addsi3 (reg, reg, GEN_INT (size)));
1014 RTX_FRAME_RELATED_P (insn) = 1;
1020 /* Generate a LINK insn for a frame sized FRAME_SIZE. If this constant
1021 is too large, generate a sequence of insns that has the same effect.
1022 SPREG contains (reg:SI REG_SP). */
1025 emit_link_insn (rtx spreg, HOST_WIDE_INT frame_size)
1027 HOST_WIDE_INT link_size = frame_size;
1031 if (link_size > 262140)
1034 /* Use a LINK insn with as big a constant as possible, then subtract
1035 any remaining size from the SP. */
1036 insn = emit_insn (gen_link (GEN_INT (-8 - link_size)));
1037 RTX_FRAME_RELATED_P (insn) = 1;
1039 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1041 rtx set = XVECEXP (PATTERN (insn), 0, i);
1042 gcc_assert (GET_CODE (set) == SET);
1043 RTX_FRAME_RELATED_P (set) = 1;
1046 frame_size -= link_size;
1050 /* Must use a call-clobbered PREG that isn't the static chain. */
1051 rtx tmpreg = gen_rtx_REG (Pmode, REG_P1);
1053 frame_related_constant_load (tmpreg, -frame_size, TRUE);
1054 insn = emit_insn (gen_addsi3 (spreg, spreg, tmpreg));
1055 RTX_FRAME_RELATED_P (insn) = 1;
1059 /* Return the number of bytes we must reserve for outgoing arguments
1060 in the current function's stack frame. */
1062 static HOST_WIDE_INT
1063 arg_area_size (void)
1065 if (crtl->outgoing_args_size)
1067 if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
1068 return crtl->outgoing_args_size;
1070 return FIXED_STACK_AREA;
1075 /* Save RETS and FP, and allocate a stack frame. ALL is true if the
1076 function must save all its registers (true only for certain interrupt
1080 do_link (rtx spreg, HOST_WIDE_INT frame_size, bool all)
1082 frame_size += arg_area_size ();
1084 if (all || stack_frame_needed_p ()
1085 || (must_save_fp_p () && ! current_function_is_leaf))
1086 emit_link_insn (spreg, frame_size);
1089 if (! current_function_is_leaf)
1091 rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
1092 gen_rtx_PRE_DEC (Pmode, spreg)),
1094 rtx insn = emit_insn (pat);
1095 RTX_FRAME_RELATED_P (insn) = 1;
1097 if (must_save_fp_p ())
1099 rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
1100 gen_rtx_PRE_DEC (Pmode, spreg)),
1101 gen_rtx_REG (Pmode, REG_FP));
1102 rtx insn = emit_insn (pat);
1103 RTX_FRAME_RELATED_P (insn) = 1;
1105 add_to_reg (spreg, -frame_size, 1, 0);
1109 /* Like do_link, but used for epilogues to deallocate the stack frame.
1110 EPILOGUE_P is zero if this function is called for prologue,
1111 otherwise it's nonzero. And it's less than zero if this is for
1112 sibcall epilogue. */
1115 do_unlink (rtx spreg, HOST_WIDE_INT frame_size, bool all, int epilogue_p)
1117 frame_size += arg_area_size ();
1119 if (all || stack_frame_needed_p ())
1120 emit_insn (gen_unlink ());
1123 rtx postinc = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, spreg));
1125 add_to_reg (spreg, frame_size, 0, epilogue_p);
1126 if (must_save_fp_p ())
1128 rtx fpreg = gen_rtx_REG (Pmode, REG_FP);
1129 emit_move_insn (fpreg, postinc);
1132 if (! current_function_is_leaf)
1134 emit_move_insn (bfin_rets_rtx, postinc);
1135 emit_use (bfin_rets_rtx);
1140 /* Generate a prologue suitable for a function of kind FKIND. This is
1141 called for interrupt and exception handler prologues.
1142 SPREG contains (reg:SI REG_SP). */
1145 expand_interrupt_handler_prologue (rtx spreg, e_funkind fkind, bool all)
1147 HOST_WIDE_INT frame_size = get_frame_size ();
1148 rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
1149 rtx predec = gen_rtx_MEM (SImode, predec1);
1151 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1152 tree kspisusp = lookup_attribute ("kspisusp", attrs);
1156 insn = emit_move_insn (spreg, gen_rtx_REG (Pmode, REG_USP));
1157 RTX_FRAME_RELATED_P (insn) = 1;
1160 /* We need space on the stack in case we need to save the argument
1162 if (fkind == EXCPT_HANDLER)
1164 insn = emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (-12)));
1165 RTX_FRAME_RELATED_P (insn) = 1;
1168 /* If we're calling other functions, they won't save their call-clobbered
1169 registers, so we must save everything here. */
1170 if (!current_function_is_leaf)
1172 expand_prologue_reg_save (spreg, all, true);
1174 if (ENABLE_WA_05000283 || ENABLE_WA_05000315)
1176 rtx chipid = GEN_INT (trunc_int_for_mode (0xFFC00014, SImode));
1177 rtx p5reg = gen_rtx_REG (Pmode, REG_P5);
1178 emit_insn (gen_movbi (bfin_cc_rtx, const1_rtx));
1179 emit_insn (gen_movsi_high (p5reg, chipid));
1180 emit_insn (gen_movsi_low (p5reg, p5reg, chipid));
1181 emit_insn (gen_dummy_load (p5reg, bfin_cc_rtx));
1184 if (lookup_attribute ("nesting", attrs))
1186 rtx srcreg = gen_rtx_REG (Pmode, (fkind == EXCPT_HANDLER ? REG_RETX
1187 : fkind == NMI_HANDLER ? REG_RETN
1189 insn = emit_move_insn (predec, srcreg);
1190 RTX_FRAME_RELATED_P (insn) = 1;
1193 do_link (spreg, frame_size, all);
1195 if (fkind == EXCPT_HANDLER)
1197 rtx r0reg = gen_rtx_REG (SImode, REG_R0);
1198 rtx r1reg = gen_rtx_REG (SImode, REG_R1);
1199 rtx r2reg = gen_rtx_REG (SImode, REG_R2);
1202 insn = emit_move_insn (r0reg, gen_rtx_REG (SImode, REG_SEQSTAT));
1203 insn = emit_insn (gen_ashrsi3 (r0reg, r0reg, GEN_INT (26)));
1204 insn = emit_insn (gen_ashlsi3 (r0reg, r0reg, GEN_INT (26)));
1205 insn = emit_move_insn (r1reg, spreg);
1206 insn = emit_move_insn (r2reg, gen_rtx_REG (Pmode, REG_FP));
1207 insn = emit_insn (gen_addsi3 (r2reg, r2reg, GEN_INT (8)));
1211 /* Generate an epilogue suitable for a function of kind FKIND. This is
1212 called for interrupt and exception handler epilogues.
1213 SPREG contains (reg:SI REG_SP). */
1216 expand_interrupt_handler_epilogue (rtx spreg, e_funkind fkind, bool all)
1218 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1219 rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
1220 rtx postinc = gen_rtx_MEM (SImode, postinc1);
1222 /* A slightly crude technique to stop flow from trying to delete "dead"
1224 MEM_VOLATILE_P (postinc) = 1;
1226 do_unlink (spreg, get_frame_size (), all, 1);
1228 if (lookup_attribute ("nesting", attrs))
1230 rtx srcreg = gen_rtx_REG (Pmode, (fkind == EXCPT_HANDLER ? REG_RETX
1231 : fkind == NMI_HANDLER ? REG_RETN
1233 emit_move_insn (srcreg, postinc);
1236 /* If we're calling other functions, they won't save their call-clobbered
1237 registers, so we must save (and restore) everything here. */
1238 if (!current_function_is_leaf)
1241 expand_epilogue_reg_restore (spreg, all, true);
1243 /* Deallocate any space we left on the stack in case we needed to save the
1244 argument registers. */
1245 if (fkind == EXCPT_HANDLER)
1246 emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (12)));
1248 emit_jump_insn (gen_return_internal (GEN_INT (fkind)));
1251 /* Used while emitting the prologue to generate code to load the correct value
1252 into the PIC register, which is passed in DEST. */
1255 bfin_load_pic_reg (rtx dest)
1257 struct cgraph_local_info *i = NULL;
1260 i = cgraph_local_info (current_function_decl);
1262 /* Functions local to the translation unit don't need to reload the
1263 pic reg, since the caller always passes a usable one. */
1265 return pic_offset_table_rtx;
1267 if (bfin_lib_id_given)
1268 addr = plus_constant (pic_offset_table_rtx, -4 - bfin_library_id * 4);
1270 addr = gen_rtx_PLUS (Pmode, pic_offset_table_rtx,
1271 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
1272 UNSPEC_LIBRARY_OFFSET));
1273 insn = emit_insn (gen_movsi (dest, gen_rtx_MEM (Pmode, addr)));
1277 /* Generate RTL for the prologue of the current function. */
1280 bfin_expand_prologue (void)
1282 HOST_WIDE_INT frame_size = get_frame_size ();
1283 rtx spreg = gen_rtx_REG (Pmode, REG_SP);
1284 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
1285 rtx pic_reg_loaded = NULL_RTX;
1286 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1287 bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
1289 if (fkind != SUBROUTINE)
1291 expand_interrupt_handler_prologue (spreg, fkind, all);
1295 if (crtl->limit_stack
1296 || (TARGET_STACK_CHECK_L1
1297 && !DECL_NO_LIMIT_STACK (current_function_decl)))
1299 HOST_WIDE_INT offset
1300 = bfin_initial_elimination_offset (ARG_POINTER_REGNUM,
1301 STACK_POINTER_REGNUM);
1302 rtx lim = crtl->limit_stack ? stack_limit_rtx : NULL_RTX;
1303 rtx p2reg = gen_rtx_REG (Pmode, REG_P2);
1307 emit_move_insn (p2reg, gen_int_mode (0xFFB00000, SImode));
1308 emit_move_insn (p2reg, gen_rtx_MEM (Pmode, p2reg));
1311 if (GET_CODE (lim) == SYMBOL_REF)
1313 if (TARGET_ID_SHARED_LIBRARY)
1315 rtx p1reg = gen_rtx_REG (Pmode, REG_P1);
1317 pic_reg_loaded = bfin_load_pic_reg (p2reg);
1318 val = legitimize_pic_address (stack_limit_rtx, p1reg,
1320 emit_move_insn (p1reg, val);
1321 frame_related_constant_load (p2reg, offset, FALSE);
1322 emit_insn (gen_addsi3 (p2reg, p2reg, p1reg));
1327 rtx limit = plus_constant (lim, offset);
1328 emit_move_insn (p2reg, limit);
1335 emit_move_insn (p2reg, lim);
1336 add_to_reg (p2reg, offset, 0, 0);
1339 emit_insn (gen_compare_lt (bfin_cc_rtx, spreg, lim));
1340 emit_insn (gen_trapifcc ());
1342 expand_prologue_reg_save (spreg, all, false);
1344 do_link (spreg, frame_size, false);
1346 if (TARGET_ID_SHARED_LIBRARY
1348 && (crtl->uses_pic_offset_table
1349 || !current_function_is_leaf))
1350 bfin_load_pic_reg (pic_offset_table_rtx);
1353 /* Generate RTL for the epilogue of the current function. NEED_RETURN is zero
1354 if this is for a sibcall. EH_RETURN is nonzero if we're expanding an
1355 eh_return pattern. SIBCALL_P is true if this is a sibcall epilogue,
1359 bfin_expand_epilogue (int need_return, int eh_return, bool sibcall_p)
1361 rtx spreg = gen_rtx_REG (Pmode, REG_SP);
1362 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
1363 int e = sibcall_p ? -1 : 1;
1364 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1365 bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
1367 if (fkind != SUBROUTINE)
1369 expand_interrupt_handler_epilogue (spreg, fkind, all);
1373 do_unlink (spreg, get_frame_size (), false, e);
1375 expand_epilogue_reg_restore (spreg, all, false);
1377 /* Omit the return insn if this is for a sibcall. */
1382 emit_insn (gen_addsi3 (spreg, spreg, gen_rtx_REG (Pmode, REG_P2)));
1384 emit_jump_insn (gen_return_internal (GEN_INT (SUBROUTINE)));
1387 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
1390 bfin_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED,
1391 unsigned int new_reg)
1393 /* Interrupt functions can only use registers that have already been
1394 saved by the prologue, even if they would normally be
1397 if (funkind (TREE_TYPE (current_function_decl)) != SUBROUTINE
1398 && !df_regs_ever_live_p (new_reg))
1404 /* Return the value of the return address for the frame COUNT steps up
1405 from the current frame, after the prologue.
1406 We punt for everything but the current frame by returning const0_rtx. */
1409 bfin_return_addr_rtx (int count)
1414 return get_hard_reg_initial_val (Pmode, REG_RETS);
1418 bfin_delegitimize_address (rtx orig_x)
1422 if (GET_CODE (x) != MEM)
1426 if (GET_CODE (x) == PLUS
1427 && GET_CODE (XEXP (x, 1)) == UNSPEC
1428 && XINT (XEXP (x, 1), 1) == UNSPEC_MOVE_PIC
1429 && GET_CODE (XEXP (x, 0)) == REG
1430 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM)
1431 return XVECEXP (XEXP (x, 1), 0, 0);
1436 /* This predicate is used to compute the length of a load/store insn.
1437 OP is a MEM rtx, we return nonzero if its addressing mode requires a
1438 32-bit instruction. */
1441 effective_address_32bit_p (rtx op, enum machine_mode mode)
1443 HOST_WIDE_INT offset;
1445 mode = GET_MODE (op);
1448 if (GET_CODE (op) != PLUS)
1450 gcc_assert (REG_P (op) || GET_CODE (op) == POST_INC
1451 || GET_CODE (op) == PRE_DEC || GET_CODE (op) == POST_DEC);
1455 if (GET_CODE (XEXP (op, 1)) == UNSPEC)
1458 offset = INTVAL (XEXP (op, 1));
1460 /* All byte loads use a 16-bit offset. */
1461 if (GET_MODE_SIZE (mode) == 1)
1464 if (GET_MODE_SIZE (mode) == 4)
1466 /* Frame pointer relative loads can use a negative offset, all others
1467 are restricted to a small positive one. */
1468 if (XEXP (op, 0) == frame_pointer_rtx)
1469 return offset < -128 || offset > 60;
1470 return offset < 0 || offset > 60;
1473 /* Must be HImode now. */
1474 return offset < 0 || offset > 30;
1477 /* Returns true if X is a memory reference using an I register. */
1479 bfin_dsp_memref_p (rtx x)
1484 if (GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_INC
1485 || GET_CODE (x) == POST_DEC || GET_CODE (x) == PRE_DEC)
1490 /* Return cost of the memory address ADDR.
1491 All addressing modes are equally cheap on the Blackfin. */
1494 bfin_address_cost (rtx addr ATTRIBUTE_UNUSED, bool speed ATTRIBUTE_UNUSED)
1499 /* Subroutine of print_operand; used to print a memory reference X to FILE. */
1502 print_address_operand (FILE *file, rtx x)
1504 switch (GET_CODE (x))
1507 output_address (XEXP (x, 0));
1508 fprintf (file, "+");
1509 output_address (XEXP (x, 1));
1513 fprintf (file, "--");
1514 output_address (XEXP (x, 0));
1517 output_address (XEXP (x, 0));
1518 fprintf (file, "++");
1521 output_address (XEXP (x, 0));
1522 fprintf (file, "--");
1526 gcc_assert (GET_CODE (x) != MEM);
1527 print_operand (file, x, 0);
1532 /* Adding intp DImode support by Tony
1538 print_operand (FILE *file, rtx x, char code)
1540 enum machine_mode mode;
1544 if (GET_MODE (current_output_insn) == SImode)
1545 fprintf (file, " ||");
1547 fprintf (file, ";");
1551 mode = GET_MODE (x);
1556 switch (GET_CODE (x))
1559 fprintf (file, "e");
1562 fprintf (file, "ne");
1565 fprintf (file, "g");
1568 fprintf (file, "l");
1571 fprintf (file, "ge");
1574 fprintf (file, "le");
1577 fprintf (file, "g");
1580 fprintf (file, "l");
1583 fprintf (file, "ge");
1586 fprintf (file, "le");
1589 output_operand_lossage ("invalid %%j value");
1593 case 'J': /* reverse logic */
1594 switch (GET_CODE(x))
1597 fprintf (file, "ne");
1600 fprintf (file, "e");
1603 fprintf (file, "le");
1606 fprintf (file, "ge");
1609 fprintf (file, "l");
1612 fprintf (file, "g");
1615 fprintf (file, "le");
1618 fprintf (file, "ge");
1621 fprintf (file, "l");
1624 fprintf (file, "g");
1627 output_operand_lossage ("invalid %%J value");
1632 switch (GET_CODE (x))
1638 fprintf (file, "%s", short_reg_names[REGNO (x)]);
1640 output_operand_lossage ("invalid operand for code '%c'", code);
1642 else if (code == 'd')
1645 fprintf (file, "%s", high_reg_names[REGNO (x)]);
1647 output_operand_lossage ("invalid operand for code '%c'", code);
1649 else if (code == 'w')
1651 if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
1652 fprintf (file, "%s.w", reg_names[REGNO (x)]);
1654 output_operand_lossage ("invalid operand for code '%c'", code);
1656 else if (code == 'x')
1658 if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
1659 fprintf (file, "%s.x", reg_names[REGNO (x)]);
1661 output_operand_lossage ("invalid operand for code '%c'", code);
1663 else if (code == 'v')
1665 if (REGNO (x) == REG_A0)
1666 fprintf (file, "AV0");
1667 else if (REGNO (x) == REG_A1)
1668 fprintf (file, "AV1");
1670 output_operand_lossage ("invalid operand for code '%c'", code);
1672 else if (code == 'D')
1674 if (D_REGNO_P (REGNO (x)))
1675 fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
1677 output_operand_lossage ("invalid operand for code '%c'", code);
1679 else if (code == 'H')
1681 if ((mode == DImode || mode == DFmode) && REG_P (x))
1682 fprintf (file, "%s", reg_names[REGNO (x) + 1]);
1684 output_operand_lossage ("invalid operand for code '%c'", code);
1686 else if (code == 'T')
1688 if (D_REGNO_P (REGNO (x)))
1689 fprintf (file, "%s", byte_reg_names[REGNO (x)]);
1691 output_operand_lossage ("invalid operand for code '%c'", code);
1694 fprintf (file, "%s", reg_names[REGNO (x)]);
1700 print_address_operand (file, x);
1712 fputs ("(FU)", file);
1715 fputs ("(T)", file);
1718 fputs ("(TFU)", file);
1721 fputs ("(W32)", file);
1724 fputs ("(IS)", file);
1727 fputs ("(IU)", file);
1730 fputs ("(IH)", file);
1733 fputs ("(M)", file);
1736 fputs ("(IS,M)", file);
1739 fputs ("(ISS2)", file);
1742 fputs ("(S2RND)", file);
1749 else if (code == 'b')
1751 if (INTVAL (x) == 0)
1753 else if (INTVAL (x) == 1)
1759 /* Moves to half registers with d or h modifiers always use unsigned
1761 else if (code == 'd')
1762 x = GEN_INT ((INTVAL (x) >> 16) & 0xffff);
1763 else if (code == 'h')
1764 x = GEN_INT (INTVAL (x) & 0xffff);
1765 else if (code == 'N')
1766 x = GEN_INT (-INTVAL (x));
1767 else if (code == 'X')
1768 x = GEN_INT (exact_log2 (0xffffffff & INTVAL (x)));
1769 else if (code == 'Y')
1770 x = GEN_INT (exact_log2 (0xffffffff & ~INTVAL (x)));
1771 else if (code == 'Z')
1772 /* Used for LINK insns. */
1773 x = GEN_INT (-8 - INTVAL (x));
1778 output_addr_const (file, x);
1782 output_operand_lossage ("invalid const_double operand");
1786 switch (XINT (x, 1))
1788 case UNSPEC_MOVE_PIC:
1789 output_addr_const (file, XVECEXP (x, 0, 0));
1790 fprintf (file, "@GOT");
1793 case UNSPEC_MOVE_FDPIC:
1794 output_addr_const (file, XVECEXP (x, 0, 0));
1795 fprintf (file, "@GOT17M4");
1798 case UNSPEC_FUNCDESC_GOT17M4:
1799 output_addr_const (file, XVECEXP (x, 0, 0));
1800 fprintf (file, "@FUNCDESC_GOT17M4");
1803 case UNSPEC_LIBRARY_OFFSET:
1804 fprintf (file, "_current_shared_library_p5_offset_");
1813 output_addr_const (file, x);
1818 /* Argument support functions. */
1820 /* Initialize a variable CUM of type CUMULATIVE_ARGS
1821 for a call to a function whose data type is FNTYPE.
1822 For a library call, FNTYPE is 0.
1823 VDSP C Compiler manual, our ABI says that
1824 first 3 words of arguments will use R0, R1 and R2.
1828 init_cumulative_args (CUMULATIVE_ARGS *cum, tree fntype,
1829 rtx libname ATTRIBUTE_UNUSED)
1831 static CUMULATIVE_ARGS zero_cum;
1835 /* Set up the number of registers to use for passing arguments. */
1837 cum->nregs = max_arg_registers;
1838 cum->arg_regs = arg_regs;
1840 cum->call_cookie = CALL_NORMAL;
1841 /* Check for a longcall attribute. */
1842 if (fntype && lookup_attribute ("shortcall", TYPE_ATTRIBUTES (fntype)))
1843 cum->call_cookie |= CALL_SHORT;
1844 else if (fntype && lookup_attribute ("longcall", TYPE_ATTRIBUTES (fntype)))
1845 cum->call_cookie |= CALL_LONG;
1850 /* Update the data in CUM to advance over an argument
1851 of mode MODE and data type TYPE.
1852 (TYPE is null for libcalls where that information may not be available.) */
1855 function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
1856 int named ATTRIBUTE_UNUSED)
1858 int count, bytes, words;
1860 bytes = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1861 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1863 cum->words += words;
1864 cum->nregs -= words;
1866 if (cum->nregs <= 0)
1869 cum->arg_regs = NULL;
1873 for (count = 1; count <= words; count++)
1880 /* Define where to put the arguments to a function.
1881 Value is zero to push the argument on the stack,
1882 or a hard register in which to store the argument.
1884 MODE is the argument's machine mode.
1885 TYPE is the data type of the argument (as a tree).
1886 This is null for libcalls where that information may
1888 CUM is a variable of type CUMULATIVE_ARGS which gives info about
1889 the preceding args and about the function being called.
1890 NAMED is nonzero if this argument is a named parameter
1891 (otherwise it is an extra parameter matching an ellipsis). */
1894 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
1895 int named ATTRIBUTE_UNUSED)
1898 = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1900 if (mode == VOIDmode)
1901 /* Compute operand 2 of the call insn. */
1902 return GEN_INT (cum->call_cookie);
1908 return gen_rtx_REG (mode, *(cum->arg_regs));
1913 /* For an arg passed partly in registers and partly in memory,
1914 this is the number of bytes passed in registers.
1915 For args passed entirely in registers or entirely in memory, zero.
1917 Refer VDSP C Compiler manual, our ABI.
1918 First 3 words are in registers. So, if an argument is larger
1919 than the registers available, it will span the register and
1923 bfin_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
1924 tree type ATTRIBUTE_UNUSED,
1925 bool named ATTRIBUTE_UNUSED)
1928 = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1929 int bytes_left = cum->nregs * UNITS_PER_WORD;
1934 if (bytes_left == 0)
1936 if (bytes > bytes_left)
1941 /* Variable sized types are passed by reference. */
1944 bfin_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
1945 enum machine_mode mode ATTRIBUTE_UNUSED,
1946 const_tree type, bool named ATTRIBUTE_UNUSED)
1948 return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
1951 /* Decide whether a type should be returned in memory (true)
1952 or in a register (false). This is called by the macro
1953 TARGET_RETURN_IN_MEMORY. */
1956 bfin_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
1958 int size = int_size_in_bytes (type);
1959 return size > 2 * UNITS_PER_WORD || size == -1;
1962 /* Register in which address to store a structure value
1963 is passed to a function. */
1965 bfin_struct_value_rtx (tree fntype ATTRIBUTE_UNUSED,
1966 int incoming ATTRIBUTE_UNUSED)
1968 return gen_rtx_REG (Pmode, REG_P0);
1971 /* Return true when register may be used to pass function parameters. */
1974 function_arg_regno_p (int n)
1977 for (i = 0; arg_regs[i] != -1; i++)
1978 if (n == arg_regs[i])
1983 /* Returns 1 if OP contains a symbol reference */
1986 symbolic_reference_mentioned_p (rtx op)
1988 register const char *fmt;
1991 if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
1994 fmt = GET_RTX_FORMAT (GET_CODE (op));
1995 for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--)
2001 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
2002 if (symbolic_reference_mentioned_p (XVECEXP (op, i, j)))
2006 else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i)))
2013 /* Decide whether we can make a sibling call to a function. DECL is the
2014 declaration of the function being targeted by the call and EXP is the
2015 CALL_EXPR representing the call. */
2018 bfin_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
2019 tree exp ATTRIBUTE_UNUSED)
2021 struct cgraph_local_info *this_func, *called_func;
2022 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
2023 if (fkind != SUBROUTINE)
2025 if (!TARGET_ID_SHARED_LIBRARY || TARGET_SEP_DATA)
2028 /* When compiling for ID shared libraries, can't sibcall a local function
2029 from a non-local function, because the local function thinks it does
2030 not need to reload P5 in the prologue, but the sibcall wil pop P5 in the
2031 sibcall epilogue, and we end up with the wrong value in P5. */
2034 /* Not enough information. */
2037 this_func = cgraph_local_info (current_function_decl);
2038 called_func = cgraph_local_info (decl);
2039 return !called_func->local || this_func->local;
2042 /* Emit RTL insns to initialize the variable parts of a trampoline at
2043 TRAMP. FNADDR is an RTX for the address of the function's pure
2044 code. CXT is an RTX for the static chain value for the function. */
2047 initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
2049 rtx t1 = copy_to_reg (fnaddr);
2050 rtx t2 = copy_to_reg (cxt);
2056 rtx a = memory_address (Pmode, plus_constant (tramp, 8));
2057 addr = memory_address (Pmode, tramp);
2058 emit_move_insn (gen_rtx_MEM (SImode, addr), a);
2062 addr = memory_address (Pmode, plus_constant (tramp, i + 2));
2063 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t1));
2064 emit_insn (gen_ashrsi3 (t1, t1, GEN_INT (16)));
2065 addr = memory_address (Pmode, plus_constant (tramp, i + 6));
2066 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t1));
2068 addr = memory_address (Pmode, plus_constant (tramp, i + 10));
2069 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t2));
2070 emit_insn (gen_ashrsi3 (t2, t2, GEN_INT (16)));
2071 addr = memory_address (Pmode, plus_constant (tramp, i + 14));
2072 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t2));
2075 /* Emit insns to move operands[1] into operands[0]. */
2078 emit_pic_move (rtx *operands, enum machine_mode mode ATTRIBUTE_UNUSED)
2080 rtx temp = reload_in_progress ? operands[0] : gen_reg_rtx (Pmode);
2082 gcc_assert (!TARGET_FDPIC || !(reload_in_progress || reload_completed));
2083 if (GET_CODE (operands[0]) == MEM && SYMBOLIC_CONST (operands[1]))
2084 operands[1] = force_reg (SImode, operands[1]);
2086 operands[1] = legitimize_pic_address (operands[1], temp,
2087 TARGET_FDPIC ? OUR_FDPIC_REG
2088 : pic_offset_table_rtx);
2091 /* Expand a move operation in mode MODE. The operands are in OPERANDS.
2092 Returns true if no further code must be generated, false if the caller
2093 should generate an insn to move OPERANDS[1] to OPERANDS[0]. */
2096 expand_move (rtx *operands, enum machine_mode mode)
2098 rtx op = operands[1];
2099 if ((TARGET_ID_SHARED_LIBRARY || TARGET_FDPIC)
2100 && SYMBOLIC_CONST (op))
2101 emit_pic_move (operands, mode);
2102 else if (mode == SImode && GET_CODE (op) == CONST
2103 && GET_CODE (XEXP (op, 0)) == PLUS
2104 && GET_CODE (XEXP (XEXP (op, 0), 0)) == SYMBOL_REF
2105 && !bfin_legitimate_constant_p (op))
2107 rtx dest = operands[0];
2109 gcc_assert (!reload_in_progress && !reload_completed);
2111 op0 = force_reg (mode, XEXP (op, 0));
2113 if (!insn_data[CODE_FOR_addsi3].operand[2].predicate (op1, mode))
2114 op1 = force_reg (mode, op1);
2115 if (GET_CODE (dest) == MEM)
2116 dest = gen_reg_rtx (mode);
2117 emit_insn (gen_addsi3 (dest, op0, op1));
2118 if (dest == operands[0])
2122 /* Don't generate memory->memory or constant->memory moves, go through a
2124 else if ((reload_in_progress | reload_completed) == 0
2125 && GET_CODE (operands[0]) == MEM
2126 && GET_CODE (operands[1]) != REG)
2127 operands[1] = force_reg (mode, operands[1]);
2131 /* Split one or more DImode RTL references into pairs of SImode
2132 references. The RTL can be REG, offsettable MEM, integer constant, or
2133 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
2134 split and "num" is its length. lo_half and hi_half are output arrays
2135 that parallel "operands". */
2138 split_di (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
2142 rtx op = operands[num];
2144 /* simplify_subreg refuse to split volatile memory addresses,
2145 but we still have to handle it. */
2146 if (GET_CODE (op) == MEM)
2148 lo_half[num] = adjust_address (op, SImode, 0);
2149 hi_half[num] = adjust_address (op, SImode, 4);
2153 lo_half[num] = simplify_gen_subreg (SImode, op,
2154 GET_MODE (op) == VOIDmode
2155 ? DImode : GET_MODE (op), 0);
2156 hi_half[num] = simplify_gen_subreg (SImode, op,
2157 GET_MODE (op) == VOIDmode
2158 ? DImode : GET_MODE (op), 4);
2164 bfin_longcall_p (rtx op, int call_cookie)
2166 gcc_assert (GET_CODE (op) == SYMBOL_REF);
2167 if (call_cookie & CALL_SHORT)
2169 if (call_cookie & CALL_LONG)
2171 if (TARGET_LONG_CALLS)
2176 /* Expand a call instruction. FNADDR is the call target, RETVAL the return value.
2177 COOKIE is a CONST_INT holding the call_cookie prepared init_cumulative_args.
2178 SIBCALL is nonzero if this is a sibling call. */
2181 bfin_expand_call (rtx retval, rtx fnaddr, rtx callarg1, rtx cookie, int sibcall)
2183 rtx use = NULL, call;
2184 rtx callee = XEXP (fnaddr, 0);
2185 int nelts = 2 + !!sibcall;
2187 rtx picreg = get_hard_reg_initial_val (SImode, FDPIC_REGNO);
2190 /* In an untyped call, we can get NULL for operand 2. */
2191 if (cookie == NULL_RTX)
2192 cookie = const0_rtx;
2194 /* Static functions and indirect calls don't need the pic register. */
2195 if (!TARGET_FDPIC && flag_pic
2196 && GET_CODE (callee) == SYMBOL_REF
2197 && !SYMBOL_REF_LOCAL_P (callee))
2198 use_reg (&use, pic_offset_table_rtx);
2202 int caller_has_l1_text, callee_has_l1_text;
2204 caller_has_l1_text = callee_has_l1_text = 0;
2206 if (lookup_attribute ("l1_text",
2207 DECL_ATTRIBUTES (cfun->decl)) != NULL_TREE)
2208 caller_has_l1_text = 1;
2210 if (GET_CODE (callee) == SYMBOL_REF
2211 && SYMBOL_REF_DECL (callee) && DECL_P (SYMBOL_REF_DECL (callee))
2214 DECL_ATTRIBUTES (SYMBOL_REF_DECL (callee))) != NULL_TREE)
2215 callee_has_l1_text = 1;
2217 if (GET_CODE (callee) != SYMBOL_REF
2218 || bfin_longcall_p (callee, INTVAL (cookie))
2219 || (GET_CODE (callee) == SYMBOL_REF
2220 && !SYMBOL_REF_LOCAL_P (callee)
2221 && TARGET_INLINE_PLT)
2222 || caller_has_l1_text != callee_has_l1_text
2223 || (caller_has_l1_text && callee_has_l1_text
2224 && (GET_CODE (callee) != SYMBOL_REF
2225 || !SYMBOL_REF_LOCAL_P (callee))))
2228 if (! address_operand (addr, Pmode))
2229 addr = force_reg (Pmode, addr);
2231 fnaddr = gen_reg_rtx (SImode);
2232 emit_insn (gen_load_funcdescsi (fnaddr, addr));
2233 fnaddr = gen_rtx_MEM (Pmode, fnaddr);
2235 picreg = gen_reg_rtx (SImode);
2236 emit_insn (gen_load_funcdescsi (picreg,
2237 plus_constant (addr, 4)));
2242 else if ((!register_no_elim_operand (callee, Pmode)
2243 && GET_CODE (callee) != SYMBOL_REF)
2244 || (GET_CODE (callee) == SYMBOL_REF
2245 && ((TARGET_ID_SHARED_LIBRARY && !TARGET_LEAF_ID_SHARED_LIBRARY)
2246 || bfin_longcall_p (callee, INTVAL (cookie)))))
2248 callee = copy_to_mode_reg (Pmode, callee);
2249 fnaddr = gen_rtx_MEM (Pmode, callee);
2251 call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1);
2254 call = gen_rtx_SET (VOIDmode, retval, call);
2256 pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nelts));
2258 XVECEXP (pat, 0, n++) = call;
2260 XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, picreg);
2261 XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, cookie);
2263 XVECEXP (pat, 0, n++) = gen_rtx_RETURN (VOIDmode);
2264 call = emit_call_insn (pat);
2266 CALL_INSN_FUNCTION_USAGE (call) = use;
2269 /* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
2272 hard_regno_mode_ok (int regno, enum machine_mode mode)
2274 /* Allow only dregs to store value of mode HI or QI */
2275 enum reg_class rclass = REGNO_REG_CLASS (regno);
2280 if (mode == V2HImode)
2281 return D_REGNO_P (regno);
2282 if (rclass == CCREGS)
2283 return mode == BImode;
2284 if (mode == PDImode || mode == V2PDImode)
2285 return regno == REG_A0 || regno == REG_A1;
2287 /* Allow all normal 32-bit regs, except REG_M3, in case regclass ever comes
2288 up with a bad register class (such as ALL_REGS) for DImode. */
2290 return regno < REG_M3;
2293 && TEST_HARD_REG_BIT (reg_class_contents[PROLOGUE_REGS], regno))
2296 return TEST_HARD_REG_BIT (reg_class_contents[MOST_REGS], regno);
2299 /* Implements target hook vector_mode_supported_p. */
2302 bfin_vector_mode_supported_p (enum machine_mode mode)
2304 return mode == V2HImode;
2307 /* Return the cost of moving data from a register in class CLASS1 to
2308 one in class CLASS2. A cost of 2 is the default. */
2311 bfin_register_move_cost (enum machine_mode mode,
2312 enum reg_class class1, enum reg_class class2)
2314 /* These need secondary reloads, so they're more expensive. */
2315 if ((class1 == CCREGS && !reg_class_subset_p (class2, DREGS))
2316 || (class2 == CCREGS && !reg_class_subset_p (class1, DREGS)))
2319 /* If optimizing for size, always prefer reg-reg over reg-memory moves. */
2323 if (GET_MODE_CLASS (mode) == MODE_INT)
2325 /* Discourage trying to use the accumulators. */
2326 if (TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A0)
2327 || TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A1)
2328 || TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A0)
2329 || TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A1))
2335 /* Return the cost of moving data of mode M between a
2336 register and memory. A value of 2 is the default; this cost is
2337 relative to those in `REGISTER_MOVE_COST'.
2339 ??? In theory L1 memory has single-cycle latency. We should add a switch
2340 that tells the compiler whether we expect to use only L1 memory for the
2341 program; it'll make the costs more accurate. */
2344 bfin_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
2345 enum reg_class rclass,
2346 int in ATTRIBUTE_UNUSED)
2348 /* Make memory accesses slightly more expensive than any register-register
2349 move. Also, penalize non-DP registers, since they need secondary
2350 reloads to load and store. */
2351 if (! reg_class_subset_p (rclass, DPREGS))
2357 /* Inform reload about cases where moving X with a mode MODE to a register in
2358 RCLASS requires an extra scratch register. Return the class needed for the
2359 scratch register. */
2361 static enum reg_class
2362 bfin_secondary_reload (bool in_p, rtx x, enum reg_class rclass,
2363 enum machine_mode mode, secondary_reload_info *sri)
2365 /* If we have HImode or QImode, we can only use DREGS as secondary registers;
2366 in most other cases we can also use PREGS. */
2367 enum reg_class default_class = GET_MODE_SIZE (mode) >= 4 ? DPREGS : DREGS;
2368 enum reg_class x_class = NO_REGS;
2369 enum rtx_code code = GET_CODE (x);
2372 x = SUBREG_REG (x), code = GET_CODE (x);
2375 int regno = REGNO (x);
2376 if (regno >= FIRST_PSEUDO_REGISTER)
2377 regno = reg_renumber[regno];
2382 x_class = REGNO_REG_CLASS (regno);
2385 /* We can be asked to reload (plus (FP) (large_constant)) into a DREG.
2386 This happens as a side effect of register elimination, and we need
2387 a scratch register to do it. */
2388 if (fp_plus_const_operand (x, mode))
2390 rtx op2 = XEXP (x, 1);
2391 int large_constant_p = ! satisfies_constraint_Ks7 (op2);
2393 if (rclass == PREGS || rclass == PREGS_CLOBBERED)
2395 /* If destination is a DREG, we can do this without a scratch register
2396 if the constant is valid for an add instruction. */
2397 if ((rclass == DREGS || rclass == DPREGS)
2398 && ! large_constant_p)
2400 /* Reloading to anything other than a DREG? Use a PREG scratch
2402 sri->icode = CODE_FOR_reload_insi;
2406 /* Data can usually be moved freely between registers of most classes.
2407 AREGS are an exception; they can only move to or from another register
2408 in AREGS or one in DREGS. They can also be assigned the constant 0. */
2409 if (x_class == AREGS || x_class == EVEN_AREGS || x_class == ODD_AREGS)
2410 return (rclass == DREGS || rclass == AREGS || rclass == EVEN_AREGS
2411 || rclass == ODD_AREGS
2414 if (rclass == AREGS || rclass == EVEN_AREGS || rclass == ODD_AREGS)
2418 sri->icode = in_p ? CODE_FOR_reload_inpdi : CODE_FOR_reload_outpdi;
2422 if (x != const0_rtx && x_class != DREGS)
2430 /* CCREGS can only be moved from/to DREGS. */
2431 if (rclass == CCREGS && x_class != DREGS)
2433 if (x_class == CCREGS && rclass != DREGS)
2436 /* All registers other than AREGS can load arbitrary constants. The only
2437 case that remains is MEM. */
2439 if (! reg_class_subset_p (rclass, default_class))
2440 return default_class;
2445 /* Implement TARGET_HANDLE_OPTION. */
2448 bfin_handle_option (size_t code, const char *arg, int value)
2452 case OPT_mshared_library_id_:
2453 if (value > MAX_LIBRARY_ID)
2454 error ("-mshared-library-id=%s is not between 0 and %d",
2455 arg, MAX_LIBRARY_ID);
2456 bfin_lib_id_given = 1;
2465 while ((p = bfin_cpus[i].name) != NULL)
2467 if (strncmp (arg, p, strlen (p)) == 0)
2474 error ("-mcpu=%s is not valid", arg);
2478 bfin_cpu_type = bfin_cpus[i].type;
2480 q = arg + strlen (p);
2484 bfin_si_revision = bfin_cpus[i].si_revision;
2485 bfin_workarounds |= bfin_cpus[i].workarounds;
2487 else if (strcmp (q, "-none") == 0)
2488 bfin_si_revision = -1;
2489 else if (strcmp (q, "-any") == 0)
2491 bfin_si_revision = 0xffff;
2492 while (bfin_cpus[i].type == bfin_cpu_type)
2494 bfin_workarounds |= bfin_cpus[i].workarounds;
2500 unsigned int si_major, si_minor;
2503 rev_len = strlen (q);
2505 if (sscanf (q, "-%u.%u%n", &si_major, &si_minor, &n) != 2
2507 || si_major > 0xff || si_minor > 0xff)
2509 invalid_silicon_revision:
2510 error ("-mcpu=%s has invalid silicon revision", arg);
2514 bfin_si_revision = (si_major << 8) | si_minor;
2516 while (bfin_cpus[i].type == bfin_cpu_type
2517 && bfin_cpus[i].si_revision != bfin_si_revision)
2520 if (bfin_cpus[i].type != bfin_cpu_type)
2521 goto invalid_silicon_revision;
2523 bfin_workarounds |= bfin_cpus[i].workarounds;
2534 static struct machine_function *
2535 bfin_init_machine_status (void)
2537 struct machine_function *f;
2539 f = GGC_CNEW (struct machine_function);
2544 /* Implement the macro OVERRIDE_OPTIONS. */
2547 override_options (void)
2549 /* If processor type is not specified, enable all workarounds. */
2550 if (bfin_cpu_type == BFIN_CPU_UNKNOWN)
2554 for (i = 0; bfin_cpus[i].name != NULL; i++)
2555 bfin_workarounds |= bfin_cpus[i].workarounds;
2557 bfin_si_revision = 0xffff;
2560 if (bfin_csync_anomaly == 1)
2561 bfin_workarounds |= WA_SPECULATIVE_SYNCS;
2562 else if (bfin_csync_anomaly == 0)
2563 bfin_workarounds &= ~WA_SPECULATIVE_SYNCS;
2565 if (bfin_specld_anomaly == 1)
2566 bfin_workarounds |= WA_SPECULATIVE_LOADS;
2567 else if (bfin_specld_anomaly == 0)
2568 bfin_workarounds &= ~WA_SPECULATIVE_LOADS;
2570 if (TARGET_OMIT_LEAF_FRAME_POINTER)
2571 flag_omit_frame_pointer = 1;
2573 /* Library identification */
2574 if (bfin_lib_id_given && ! TARGET_ID_SHARED_LIBRARY)
2575 error ("-mshared-library-id= specified without -mid-shared-library");
2577 if (stack_limit_rtx && TARGET_STACK_CHECK_L1)
2578 error ("Can't use multiple stack checking methods together.");
2580 if (TARGET_ID_SHARED_LIBRARY && TARGET_FDPIC)
2581 error ("ID shared libraries and FD-PIC mode can't be used together.");
2583 /* Don't allow the user to specify -mid-shared-library and -msep-data
2584 together, as it makes little sense from a user's point of view... */
2585 if (TARGET_SEP_DATA && TARGET_ID_SHARED_LIBRARY)
2586 error ("cannot specify both -msep-data and -mid-shared-library");
2587 /* ... internally, however, it's nearly the same. */
2588 if (TARGET_SEP_DATA)
2589 target_flags |= MASK_ID_SHARED_LIBRARY | MASK_LEAF_ID_SHARED_LIBRARY;
2591 if (TARGET_ID_SHARED_LIBRARY && flag_pic == 0)
2594 /* There is no single unaligned SI op for PIC code. Sometimes we
2595 need to use ".4byte" and sometimes we need to use ".picptr".
2596 See bfin_assemble_integer for details. */
2598 targetm.asm_out.unaligned_op.si = 0;
2600 /* Silently turn off flag_pic if not doing FDPIC or ID shared libraries,
2601 since we don't support it and it'll just break. */
2602 if (flag_pic && !TARGET_FDPIC && !TARGET_ID_SHARED_LIBRARY)
2605 if (TARGET_MULTICORE && bfin_cpu_type != BFIN_CPU_BF561)
2606 error ("-mmulticore can only be used with BF561");
2608 if (TARGET_COREA && !TARGET_MULTICORE)
2609 error ("-mcorea should be used with -mmulticore");
2611 if (TARGET_COREB && !TARGET_MULTICORE)
2612 error ("-mcoreb should be used with -mmulticore");
2614 if (TARGET_COREA && TARGET_COREB)
2615 error ("-mcorea and -mcoreb can't be used together");
2617 flag_schedule_insns = 0;
2619 /* Passes after sched2 can break the helpful TImode annotations that
2620 haifa-sched puts on every insn. Just do scheduling in reorg. */
2621 bfin_flag_schedule_insns2 = flag_schedule_insns_after_reload;
2622 flag_schedule_insns_after_reload = 0;
2624 init_machine_status = bfin_init_machine_status;
2627 /* Return the destination address of BRANCH.
2628 We need to use this instead of get_attr_length, because the
2629 cbranch_with_nops pattern conservatively sets its length to 6, and
2630 we still prefer to use shorter sequences. */
2633 branch_dest (rtx branch)
2637 rtx pat = PATTERN (branch);
2638 if (GET_CODE (pat) == PARALLEL)
2639 pat = XVECEXP (pat, 0, 0);
2640 dest = SET_SRC (pat);
2641 if (GET_CODE (dest) == IF_THEN_ELSE)
2642 dest = XEXP (dest, 1);
2643 dest = XEXP (dest, 0);
2644 dest_uid = INSN_UID (dest);
2645 return INSN_ADDRESSES (dest_uid);
2648 /* Return nonzero if INSN is annotated with a REG_BR_PROB note that indicates
2649 it's a branch that's predicted taken. */
2652 cbranch_predicted_taken_p (rtx insn)
2654 rtx x = find_reg_note (insn, REG_BR_PROB, 0);
2658 int pred_val = INTVAL (XEXP (x, 0));
2660 return pred_val >= REG_BR_PROB_BASE / 2;
2666 /* Templates for use by asm_conditional_branch. */
2668 static const char *ccbranch_templates[][3] = {
2669 { "if !cc jump %3;", "if cc jump 4 (bp); jump.s %3;", "if cc jump 6 (bp); jump.l %3;" },
2670 { "if cc jump %3;", "if !cc jump 4 (bp); jump.s %3;", "if !cc jump 6 (bp); jump.l %3;" },
2671 { "if !cc jump %3 (bp);", "if cc jump 4; jump.s %3;", "if cc jump 6; jump.l %3;" },
2672 { "if cc jump %3 (bp);", "if !cc jump 4; jump.s %3;", "if !cc jump 6; jump.l %3;" },
2675 /* Output INSN, which is a conditional branch instruction with operands
2678 We deal with the various forms of conditional branches that can be generated
2679 by bfin_reorg to prevent the hardware from doing speculative loads, by
2680 - emitting a sufficient number of nops, if N_NOPS is nonzero, or
2681 - always emitting the branch as predicted taken, if PREDICT_TAKEN is true.
2682 Either of these is only necessary if the branch is short, otherwise the
2683 template we use ends in an unconditional jump which flushes the pipeline
2687 asm_conditional_branch (rtx insn, rtx *operands, int n_nops, int predict_taken)
2689 int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn));
2690 /* Note : offset for instructions like if cc jmp; jump.[sl] offset
2691 is to be taken from start of if cc rather than jump.
2692 Range for jump.s is (-4094, 4096) instead of (-4096, 4094)
2694 int len = (offset >= -1024 && offset <= 1022 ? 0
2695 : offset >= -4094 && offset <= 4096 ? 1
2697 int bp = predict_taken && len == 0 ? 1 : cbranch_predicted_taken_p (insn);
2698 int idx = (bp << 1) | (GET_CODE (operands[0]) == EQ ? BRF : BRT);
2699 output_asm_insn (ccbranch_templates[idx][len], operands);
2700 gcc_assert (n_nops == 0 || !bp);
2702 while (n_nops-- > 0)
2703 output_asm_insn ("nop;", NULL);
2706 /* Emit rtl for a comparison operation CMP in mode MODE. Operands have been
2707 stored in bfin_compare_op0 and bfin_compare_op1 already. */
2710 bfin_gen_compare (rtx cmp, enum machine_mode mode ATTRIBUTE_UNUSED)
2712 enum rtx_code code1, code2;
2713 rtx op0 = XEXP (cmp, 0), op1 = XEXP (cmp, 1);
2714 rtx tem = bfin_cc_rtx;
2715 enum rtx_code code = GET_CODE (cmp);
2717 /* If we have a BImode input, then we already have a compare result, and
2718 do not need to emit another comparison. */
2719 if (GET_MODE (op0) == BImode)
2721 gcc_assert ((code == NE || code == EQ) && op1 == const0_rtx);
2722 tem = op0, code2 = code;
2727 /* bfin has these conditions */
2737 code1 = reverse_condition (code);
2741 emit_insn (gen_rtx_SET (VOIDmode, tem,
2742 gen_rtx_fmt_ee (code1, BImode, op0, op1)));
2745 return gen_rtx_fmt_ee (code2, BImode, tem, CONST0_RTX (BImode));
2748 /* Return nonzero iff C has exactly one bit set if it is interpreted
2749 as a 32-bit constant. */
2752 log2constp (unsigned HOST_WIDE_INT c)
2755 return c != 0 && (c & (c-1)) == 0;
2758 /* Returns the number of consecutive least significant zeros in the binary
2759 representation of *V.
2760 We modify *V to contain the original value arithmetically shifted right by
2761 the number of zeroes. */
2764 shiftr_zero (HOST_WIDE_INT *v)
2766 unsigned HOST_WIDE_INT tmp = *v;
2767 unsigned HOST_WIDE_INT sgn;
2773 sgn = tmp & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1));
2774 while ((tmp & 0x1) == 0 && n <= 32)
2776 tmp = (tmp >> 1) | sgn;
2783 /* After reload, split the load of an immediate constant. OPERANDS are the
2784 operands of the movsi_insn pattern which we are splitting. We return
2785 nonzero if we emitted a sequence to load the constant, zero if we emitted
2786 nothing because we want to use the splitter's default sequence. */
2789 split_load_immediate (rtx operands[])
2791 HOST_WIDE_INT val = INTVAL (operands[1]);
2793 HOST_WIDE_INT shifted = val;
2794 HOST_WIDE_INT shifted_compl = ~val;
2795 int num_zero = shiftr_zero (&shifted);
2796 int num_compl_zero = shiftr_zero (&shifted_compl);
2797 unsigned int regno = REGNO (operands[0]);
2799 /* This case takes care of single-bit set/clear constants, which we could
2800 also implement with BITSET/BITCLR. */
2802 && shifted >= -32768 && shifted < 65536
2803 && (D_REGNO_P (regno)
2804 || (regno >= REG_P0 && regno <= REG_P7 && num_zero <= 2)))
2806 emit_insn (gen_movsi (operands[0], GEN_INT (shifted)));
2807 emit_insn (gen_ashlsi3 (operands[0], operands[0], GEN_INT (num_zero)));
2812 tmp |= -(tmp & 0x8000);
2814 /* If high word has one bit set or clear, try to use a bit operation. */
2815 if (D_REGNO_P (regno))
2817 if (log2constp (val & 0xFFFF0000))
2819 emit_insn (gen_movsi (operands[0], GEN_INT (val & 0xFFFF)));
2820 emit_insn (gen_iorsi3 (operands[0], operands[0], GEN_INT (val & 0xFFFF0000)));
2823 else if (log2constp (val | 0xFFFF) && (val & 0x8000) != 0)
2825 emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
2826 emit_insn (gen_andsi3 (operands[0], operands[0], GEN_INT (val | 0xFFFF)));
2830 if (D_REGNO_P (regno))
2832 if (tmp >= -64 && tmp <= 63)
2834 emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
2835 emit_insn (gen_movstricthi_high (operands[0], GEN_INT (val & -65536)));
2839 if ((val & 0xFFFF0000) == 0)
2841 emit_insn (gen_movsi (operands[0], const0_rtx));
2842 emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
2846 if ((val & 0xFFFF0000) == 0xFFFF0000)
2848 emit_insn (gen_movsi (operands[0], constm1_rtx));
2849 emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
2854 /* Need DREGs for the remaining case. */
2859 && num_compl_zero && shifted_compl >= -64 && shifted_compl <= 63)
2861 /* If optimizing for size, generate a sequence that has more instructions
2863 emit_insn (gen_movsi (operands[0], GEN_INT (shifted_compl)));
2864 emit_insn (gen_ashlsi3 (operands[0], operands[0],
2865 GEN_INT (num_compl_zero)));
2866 emit_insn (gen_one_cmplsi2 (operands[0], operands[0]));
2872 /* Return true if the legitimate memory address for a memory operand of mode
2873 MODE. Return false if not. */
2876 bfin_valid_add (enum machine_mode mode, HOST_WIDE_INT value)
2878 unsigned HOST_WIDE_INT v = value > 0 ? value : -value;
2879 int sz = GET_MODE_SIZE (mode);
2880 int shift = sz == 1 ? 0 : sz == 2 ? 1 : 2;
2881 /* The usual offsettable_memref machinery doesn't work so well for this
2882 port, so we deal with the problem here. */
2883 if (value > 0 && sz == 8)
2885 return (v & ~(0x7fff << shift)) == 0;
2889 bfin_valid_reg_p (unsigned int regno, int strict, enum machine_mode mode,
2890 enum rtx_code outer_code)
2893 return REGNO_OK_FOR_BASE_STRICT_P (regno, mode, outer_code, SCRATCH);
2895 return REGNO_OK_FOR_BASE_NONSTRICT_P (regno, mode, outer_code, SCRATCH);
2899 bfin_legitimate_address_p (enum machine_mode mode, rtx x, int strict)
2901 switch (GET_CODE (x)) {
2903 if (bfin_valid_reg_p (REGNO (x), strict, mode, MEM))
2907 if (REG_P (XEXP (x, 0))
2908 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PLUS)
2909 && ((GET_CODE (XEXP (x, 1)) == UNSPEC && mode == SImode)
2910 || (GET_CODE (XEXP (x, 1)) == CONST_INT
2911 && bfin_valid_add (mode, INTVAL (XEXP (x, 1))))))
2916 if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
2917 && REG_P (XEXP (x, 0))
2918 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, POST_INC))
2921 if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
2922 && XEXP (x, 0) == stack_pointer_rtx
2923 && REG_P (XEXP (x, 0))
2924 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PRE_DEC))
2933 /* Decide whether we can force certain constants to memory. If we
2934 decide we can't, the caller should be able to cope with it in
2938 bfin_cannot_force_const_mem (rtx x ATTRIBUTE_UNUSED)
2940 /* We have only one class of non-legitimate constants, and our movsi
2941 expander knows how to handle them. Dropping these constants into the
2942 data section would only shift the problem - we'd still get relocs
2943 outside the object, in the data section rather than the text section. */
2947 /* Ensure that for any constant of the form symbol + offset, the offset
2948 remains within the object. Any other constants are ok.
2949 This ensures that flat binaries never have to deal with relocations
2950 crossing section boundaries. */
2953 bfin_legitimate_constant_p (rtx x)
2956 HOST_WIDE_INT offset;
2958 if (GET_CODE (x) != CONST)
2962 gcc_assert (GET_CODE (x) == PLUS);
2966 if (GET_CODE (sym) != SYMBOL_REF
2967 || GET_CODE (x) != CONST_INT)
2969 offset = INTVAL (x);
2971 if (SYMBOL_REF_DECL (sym) == 0)
2974 || offset >= int_size_in_bytes (TREE_TYPE (SYMBOL_REF_DECL (sym))))
2981 bfin_rtx_costs (rtx x, int code, int outer_code, int *total, bool speed)
2983 int cost2 = COSTS_N_INSNS (1);
2989 if (outer_code == SET || outer_code == PLUS)
2990 *total = satisfies_constraint_Ks7 (x) ? 0 : cost2;
2991 else if (outer_code == AND)
2992 *total = log2constp (~INTVAL (x)) ? 0 : cost2;
2993 else if (outer_code == LE || outer_code == LT || outer_code == EQ)
2994 *total = (INTVAL (x) >= -4 && INTVAL (x) <= 3) ? 0 : cost2;
2995 else if (outer_code == LEU || outer_code == LTU)
2996 *total = (INTVAL (x) >= 0 && INTVAL (x) <= 7) ? 0 : cost2;
2997 else if (outer_code == MULT)
2998 *total = (INTVAL (x) == 2 || INTVAL (x) == 4) ? 0 : cost2;
2999 else if (outer_code == ASHIFT && (INTVAL (x) == 1 || INTVAL (x) == 2))
3001 else if (outer_code == ASHIFT || outer_code == ASHIFTRT
3002 || outer_code == LSHIFTRT)
3003 *total = (INTVAL (x) >= 0 && INTVAL (x) <= 31) ? 0 : cost2;
3004 else if (outer_code == IOR || outer_code == XOR)
3005 *total = (INTVAL (x) & (INTVAL (x) - 1)) == 0 ? 0 : cost2;
3014 *total = COSTS_N_INSNS (2);
3020 if (GET_MODE (x) == SImode)
3022 if (GET_CODE (op0) == MULT
3023 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
3025 HOST_WIDE_INT val = INTVAL (XEXP (op0, 1));
3026 if (val == 2 || val == 4)
3029 *total += rtx_cost (XEXP (op0, 0), outer_code, speed);
3030 *total += rtx_cost (op1, outer_code, speed);
3035 if (GET_CODE (op0) != REG
3036 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3037 *total += rtx_cost (op0, SET, speed);
3038 #if 0 /* We'd like to do this for accuracy, but it biases the loop optimizer
3039 towards creating too many induction variables. */
3040 if (!reg_or_7bit_operand (op1, SImode))
3041 *total += rtx_cost (op1, SET, speed);
3044 else if (GET_MODE (x) == DImode)
3047 if (GET_CODE (op1) != CONST_INT
3048 || !satisfies_constraint_Ks7 (op1))
3049 *total += rtx_cost (op1, PLUS, speed);
3050 if (GET_CODE (op0) != REG
3051 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3052 *total += rtx_cost (op0, PLUS, speed);
3057 if (GET_MODE (x) == DImode)
3066 if (GET_MODE (x) == DImode)
3073 if (GET_CODE (op0) != REG
3074 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3075 *total += rtx_cost (op0, code, speed);
3085 /* Handle special cases of IOR: rotates, ALIGN insns, movstricthi_high. */
3088 if ((GET_CODE (op0) == LSHIFTRT && GET_CODE (op1) == ASHIFT)
3089 || (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == ZERO_EXTEND)
3090 || (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT)
3091 || (GET_CODE (op0) == AND && GET_CODE (op1) == CONST_INT))
3098 if (GET_CODE (op0) != REG
3099 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3100 *total += rtx_cost (op0, code, speed);
3102 if (GET_MODE (x) == DImode)
3108 if (GET_MODE (x) != SImode)
3113 if (! rhs_andsi3_operand (XEXP (x, 1), SImode))
3114 *total += rtx_cost (XEXP (x, 1), code, speed);
3118 if (! regorlog2_operand (XEXP (x, 1), SImode))
3119 *total += rtx_cost (XEXP (x, 1), code, speed);
3126 if (outer_code == SET
3127 && XEXP (x, 1) == const1_rtx
3128 && GET_CODE (XEXP (x, 2)) == CONST_INT)
3144 if (GET_CODE (op0) == GET_CODE (op1)
3145 && (GET_CODE (op0) == ZERO_EXTEND
3146 || GET_CODE (op0) == SIGN_EXTEND))
3148 *total = COSTS_N_INSNS (1);
3149 op0 = XEXP (op0, 0);
3150 op1 = XEXP (op1, 0);
3153 *total = COSTS_N_INSNS (1);
3155 *total = COSTS_N_INSNS (3);
3157 if (GET_CODE (op0) != REG
3158 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3159 *total += rtx_cost (op0, MULT, speed);
3160 if (GET_CODE (op1) != REG
3161 && (GET_CODE (op1) != SUBREG || GET_CODE (SUBREG_REG (op1)) != REG))
3162 *total += rtx_cost (op1, MULT, speed);
3168 *total = COSTS_N_INSNS (32);
3173 if (outer_code == SET)
3182 /* Used for communication between {push,pop}_multiple_operation (which
3183 we use not only as a predicate) and the corresponding output functions. */
3184 static int first_preg_to_save, first_dreg_to_save;
3185 static int n_regs_to_save;
3188 push_multiple_operation (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
3190 int lastdreg = 8, lastpreg = 6;
3193 first_preg_to_save = lastpreg;
3194 first_dreg_to_save = lastdreg;
3195 for (i = 1, group = 0; i < XVECLEN (op, 0) - 1; i++)
3197 rtx t = XVECEXP (op, 0, i);
3201 if (GET_CODE (t) != SET)
3205 dest = SET_DEST (t);
3206 if (GET_CODE (dest) != MEM || ! REG_P (src))
3208 dest = XEXP (dest, 0);
3209 if (GET_CODE (dest) != PLUS
3210 || ! REG_P (XEXP (dest, 0))
3211 || REGNO (XEXP (dest, 0)) != REG_SP
3212 || GET_CODE (XEXP (dest, 1)) != CONST_INT
3213 || INTVAL (XEXP (dest, 1)) != -i * 4)
3216 regno = REGNO (src);
3219 if (D_REGNO_P (regno))
3222 first_dreg_to_save = lastdreg = regno - REG_R0;
3224 else if (regno >= REG_P0 && regno <= REG_P7)
3227 first_preg_to_save = lastpreg = regno - REG_P0;
3237 if (regno >= REG_P0 && regno <= REG_P7)
3240 first_preg_to_save = lastpreg = regno - REG_P0;
3242 else if (regno != REG_R0 + lastdreg + 1)
3247 else if (group == 2)
3249 if (regno != REG_P0 + lastpreg + 1)
3254 n_regs_to_save = 8 - first_dreg_to_save + 6 - first_preg_to_save;
3259 pop_multiple_operation (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
3261 int lastdreg = 8, lastpreg = 6;
3264 for (i = 1, group = 0; i < XVECLEN (op, 0); i++)
3266 rtx t = XVECEXP (op, 0, i);
3270 if (GET_CODE (t) != SET)
3274 dest = SET_DEST (t);
3275 if (GET_CODE (src) != MEM || ! REG_P (dest))
3277 src = XEXP (src, 0);
3281 if (! REG_P (src) || REGNO (src) != REG_SP)
3284 else if (GET_CODE (src) != PLUS
3285 || ! REG_P (XEXP (src, 0))
3286 || REGNO (XEXP (src, 0)) != REG_SP
3287 || GET_CODE (XEXP (src, 1)) != CONST_INT
3288 || INTVAL (XEXP (src, 1)) != (i - 1) * 4)
3291 regno = REGNO (dest);
3294 if (regno == REG_R7)
3299 else if (regno != REG_P0 + lastpreg - 1)
3304 else if (group == 1)
3306 if (regno != REG_R0 + lastdreg - 1)
3312 first_dreg_to_save = lastdreg;
3313 first_preg_to_save = lastpreg;
3314 n_regs_to_save = 8 - first_dreg_to_save + 6 - first_preg_to_save;
3318 /* Emit assembly code for one multi-register push described by INSN, with
3319 operands in OPERANDS. */
3322 output_push_multiple (rtx insn, rtx *operands)
3327 /* Validate the insn again, and compute first_[dp]reg_to_save. */
3328 ok = push_multiple_operation (PATTERN (insn), VOIDmode);
3331 if (first_dreg_to_save == 8)
3332 sprintf (buf, "[--sp] = ( p5:%d );\n", first_preg_to_save);
3333 else if (first_preg_to_save == 6)
3334 sprintf (buf, "[--sp] = ( r7:%d );\n", first_dreg_to_save);
3336 sprintf (buf, "[--sp] = ( r7:%d, p5:%d );\n",
3337 first_dreg_to_save, first_preg_to_save);
3339 output_asm_insn (buf, operands);
3342 /* Emit assembly code for one multi-register pop described by INSN, with
3343 operands in OPERANDS. */
3346 output_pop_multiple (rtx insn, rtx *operands)
3351 /* Validate the insn again, and compute first_[dp]reg_to_save. */
3352 ok = pop_multiple_operation (PATTERN (insn), VOIDmode);
3355 if (first_dreg_to_save == 8)
3356 sprintf (buf, "( p5:%d ) = [sp++];\n", first_preg_to_save);
3357 else if (first_preg_to_save == 6)
3358 sprintf (buf, "( r7:%d ) = [sp++];\n", first_dreg_to_save);
3360 sprintf (buf, "( r7:%d, p5:%d ) = [sp++];\n",
3361 first_dreg_to_save, first_preg_to_save);
3363 output_asm_insn (buf, operands);
3366 /* Adjust DST and SRC by OFFSET bytes, and generate one move in mode MODE. */
3369 single_move_for_movmem (rtx dst, rtx src, enum machine_mode mode, HOST_WIDE_INT offset)
3371 rtx scratch = gen_reg_rtx (mode);
3374 srcmem = adjust_address_nv (src, mode, offset);
3375 dstmem = adjust_address_nv (dst, mode, offset);
3376 emit_move_insn (scratch, srcmem);
3377 emit_move_insn (dstmem, scratch);
3380 /* Expand a string move operation of COUNT_EXP bytes from SRC to DST, with
3381 alignment ALIGN_EXP. Return true if successful, false if we should fall
3382 back on a different method. */
3385 bfin_expand_movmem (rtx dst, rtx src, rtx count_exp, rtx align_exp)
3387 rtx srcreg, destreg, countreg;
3388 HOST_WIDE_INT align = 0;
3389 unsigned HOST_WIDE_INT count = 0;
3391 if (GET_CODE (align_exp) == CONST_INT)
3392 align = INTVAL (align_exp);
3393 if (GET_CODE (count_exp) == CONST_INT)
3395 count = INTVAL (count_exp);
3397 if (!TARGET_INLINE_ALL_STRINGOPS && count > 64)
3402 /* If optimizing for size, only do single copies inline. */
3405 if (count == 2 && align < 2)
3407 if (count == 4 && align < 4)
3409 if (count != 1 && count != 2 && count != 4)
3412 if (align < 2 && count != 1)
3415 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
3416 if (destreg != XEXP (dst, 0))
3417 dst = replace_equiv_address_nv (dst, destreg);
3418 srcreg = copy_to_mode_reg (Pmode, XEXP (src, 0));
3419 if (srcreg != XEXP (src, 0))
3420 src = replace_equiv_address_nv (src, srcreg);
3422 if (count != 0 && align >= 2)
3424 unsigned HOST_WIDE_INT offset = 0;
3428 if ((count & ~3) == 4)
3430 single_move_for_movmem (dst, src, SImode, offset);
3433 else if (count & ~3)
3435 HOST_WIDE_INT new_count = ((count >> 2) & 0x3fffffff) - 1;
3436 countreg = copy_to_mode_reg (Pmode, GEN_INT (new_count));
3438 emit_insn (gen_rep_movsi (destreg, srcreg, countreg, destreg, srcreg));
3439 cfun->machine->has_loopreg_clobber = true;
3443 single_move_for_movmem (dst, src, HImode, offset);
3449 if ((count & ~1) == 2)
3451 single_move_for_movmem (dst, src, HImode, offset);
3454 else if (count & ~1)
3456 HOST_WIDE_INT new_count = ((count >> 1) & 0x7fffffff) - 1;
3457 countreg = copy_to_mode_reg (Pmode, GEN_INT (new_count));
3459 emit_insn (gen_rep_movhi (destreg, srcreg, countreg, destreg, srcreg));
3460 cfun->machine->has_loopreg_clobber = true;
3465 single_move_for_movmem (dst, src, QImode, offset);
3472 /* Compute the alignment for a local variable.
3473 TYPE is the data type, and ALIGN is the alignment that
3474 the object would ordinarily have. The value of this macro is used
3475 instead of that alignment to align the object. */
3478 bfin_local_alignment (tree type, int align)
3480 /* Increasing alignment for (relatively) big types allows the builtin
3481 memcpy can use 32 bit loads/stores. */
3482 if (TYPE_SIZE (type)
3483 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
3484 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) > 8
3485 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 32)
3490 /* Implement TARGET_SCHED_ISSUE_RATE. */
3493 bfin_issue_rate (void)
3499 bfin_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
3501 enum attr_type insn_type, dep_insn_type;
3502 int dep_insn_code_number;
3504 /* Anti and output dependencies have zero cost. */
3505 if (REG_NOTE_KIND (link) != 0)
3508 dep_insn_code_number = recog_memoized (dep_insn);
3510 /* If we can't recognize the insns, we can't really do anything. */
3511 if (dep_insn_code_number < 0 || recog_memoized (insn) < 0)
3514 insn_type = get_attr_type (insn);
3515 dep_insn_type = get_attr_type (dep_insn);
3517 if (dep_insn_type == TYPE_MOVE || dep_insn_type == TYPE_MCLD)
3519 rtx pat = PATTERN (dep_insn);
3520 if (GET_CODE (pat) == PARALLEL)
3521 pat = XVECEXP (pat, 0, 0);
3522 rtx dest = SET_DEST (pat);
3523 rtx src = SET_SRC (pat);
3524 if (! ADDRESS_REGNO_P (REGNO (dest))
3525 || ! (MEM_P (src) || D_REGNO_P (REGNO (src))))
3527 return cost + (dep_insn_type == TYPE_MOVE ? 4 : 3);
3533 /* This function acts like NEXT_INSN, but is aware of three-insn bundles and
3534 skips all subsequent parallel instructions if INSN is the start of such
3537 find_next_insn_start (rtx insn)
3539 if (GET_MODE (insn) == SImode)
3541 while (GET_MODE (insn) != QImode)
3542 insn = NEXT_INSN (insn);
3544 return NEXT_INSN (insn);
3547 /* This function acts like PREV_INSN, but is aware of three-insn bundles and
3548 skips all subsequent parallel instructions if INSN is the start of such
3551 find_prev_insn_start (rtx insn)
3553 insn = PREV_INSN (insn);
3554 gcc_assert (GET_MODE (insn) != SImode);
3555 if (GET_MODE (insn) == QImode)
3557 while (GET_MODE (PREV_INSN (insn)) == SImode)
3558 insn = PREV_INSN (insn);
3563 /* Increment the counter for the number of loop instructions in the
3564 current function. */
3567 bfin_hardware_loop (void)
3569 cfun->machine->has_hardware_loops++;
3572 /* Maximum loop nesting depth. */
3573 #define MAX_LOOP_DEPTH 2
3575 /* Maximum size of a loop. */
3576 #define MAX_LOOP_LENGTH 2042
3578 /* Maximum distance of the LSETUP instruction from the loop start. */
3579 #define MAX_LSETUP_DISTANCE 30
3581 /* We need to keep a vector of loops */
3582 typedef struct loop_info *loop_info;
3583 DEF_VEC_P (loop_info);
3584 DEF_VEC_ALLOC_P (loop_info,heap);
3586 /* Information about a loop we have found (or are in the process of
3588 struct GTY (()) loop_info
3590 /* loop number, for dumps */
3593 /* All edges that jump into and out of the loop. */
3594 VEC(edge,gc) *incoming;
3596 /* We can handle two cases: all incoming edges have the same destination
3597 block, or all incoming edges have the same source block. These two
3598 members are set to the common source or destination we found, or NULL
3599 if different blocks were found. If both are NULL the loop can't be
3601 basic_block incoming_src;
3602 basic_block incoming_dest;
3604 /* First block in the loop. This is the one branched to by the loop_end
3608 /* Last block in the loop (the one with the loop_end insn). */
3611 /* The successor block of the loop. This is the one the loop_end insn
3613 basic_block successor;
3615 /* The last instruction in the tail. */
3618 /* The loop_end insn. */
3621 /* The iteration register. */
3624 /* The new label placed at the beginning of the loop. */
3627 /* The new label placed at the end of the loop. */
3630 /* The length of the loop. */
3633 /* The nesting depth of the loop. */
3636 /* Nonzero if we can't optimize this loop. */
3639 /* True if we have visited this loop. */
3642 /* True if this loop body clobbers any of LC0, LT0, or LB0. */
3645 /* True if this loop body clobbers any of LC1, LT1, or LB1. */
3648 /* Next loop in the graph. */
3649 struct loop_info *next;
3651 /* Immediate outer loop of this loop. */
3652 struct loop_info *outer;
3654 /* Vector of blocks only within the loop, including those within
3656 VEC (basic_block,heap) *blocks;
3658 /* Same information in a bitmap. */
3659 bitmap block_bitmap;
3661 /* Vector of inner loops within this loop */
3662 VEC (loop_info,heap) *loops;
3666 bfin_dump_loops (loop_info loops)
3670 for (loop = loops; loop; loop = loop->next)
3676 fprintf (dump_file, ";; loop %d: ", loop->loop_no);
3678 fprintf (dump_file, "(bad) ");
3679 fprintf (dump_file, "{head:%d, depth:%d}", loop->head->index, loop->depth);
3681 fprintf (dump_file, " blocks: [ ");
3682 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, b); ix++)
3683 fprintf (dump_file, "%d ", b->index);
3684 fprintf (dump_file, "] ");
3686 fprintf (dump_file, " inner loops: [ ");
3687 for (ix = 0; VEC_iterate (loop_info, loop->loops, ix, i); ix++)
3688 fprintf (dump_file, "%d ", i->loop_no);
3689 fprintf (dump_file, "]\n");
3691 fprintf (dump_file, "\n");
3694 /* Scan the blocks of LOOP (and its inferiors) looking for basic block
3695 BB. Return true, if we find it. */
3698 bfin_bb_in_loop (loop_info loop, basic_block bb)
3700 return bitmap_bit_p (loop->block_bitmap, bb->index);
3703 /* Scan the blocks of LOOP (and its inferiors) looking for uses of
3704 REG. Return true, if we find any. Don't count the loop's loop_end
3705 insn if it matches LOOP_END. */
3708 bfin_scan_loop (loop_info loop, rtx reg, rtx loop_end)
3713 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, bb); ix++)
3717 for (insn = BB_HEAD (bb);
3718 insn != NEXT_INSN (BB_END (bb));
3719 insn = NEXT_INSN (insn))
3723 if (insn == loop_end)
3725 if (reg_mentioned_p (reg, PATTERN (insn)))
3732 /* Estimate the length of INSN conservatively. */
3735 length_for_loop (rtx insn)
3738 if (JUMP_P (insn) && any_condjump_p (insn) && !optimize_size)
3740 if (ENABLE_WA_SPECULATIVE_SYNCS)
3742 else if (ENABLE_WA_SPECULATIVE_LOADS)
3745 else if (LABEL_P (insn))
3747 if (ENABLE_WA_SPECULATIVE_SYNCS)
3752 length += get_attr_length (insn);
3757 /* Optimize LOOP. */
3760 bfin_optimize_loop (loop_info loop)
3764 rtx insn, last_insn;
3765 rtx loop_init, start_label, end_label;
3766 rtx reg_lc0, reg_lc1, reg_lt0, reg_lt1, reg_lb0, reg_lb1;
3767 rtx iter_reg, scratchreg, scratch_init, scratch_init_insn;
3768 rtx lc_reg, lt_reg, lb_reg;
3772 int inner_depth = 0;
3782 fprintf (dump_file, ";; loop %d bad when found\n", loop->loop_no);
3786 /* Every loop contains in its list of inner loops every loop nested inside
3787 it, even if there are intermediate loops. This works because we're doing
3788 a depth-first search here and never visit a loop more than once. */
3789 for (ix = 0; VEC_iterate (loop_info, loop->loops, ix, inner); ix++)
3791 bfin_optimize_loop (inner);
3793 if (!inner->bad && inner_depth < inner->depth)
3795 inner_depth = inner->depth;
3797 loop->clobber_loop0 |= inner->clobber_loop0;
3798 loop->clobber_loop1 |= inner->clobber_loop1;
3802 loop->depth = inner_depth + 1;
3803 if (loop->depth > MAX_LOOP_DEPTH)
3806 fprintf (dump_file, ";; loop %d too deep\n", loop->loop_no);
3810 /* Get the loop iteration register. */
3811 iter_reg = loop->iter_reg;
3813 if (!REG_P (iter_reg))
3816 fprintf (dump_file, ";; loop %d iteration count not in a register\n",
3820 scratchreg = NULL_RTX;
3821 scratch_init = iter_reg;
3822 scratch_init_insn = NULL_RTX;
3823 if (!PREG_P (iter_reg) && loop->incoming_src)
3825 basic_block bb_in = loop->incoming_src;
3827 for (i = REG_P0; i <= REG_P5; i++)
3828 if ((df_regs_ever_live_p (i)
3829 || (funkind (TREE_TYPE (current_function_decl)) == SUBROUTINE
3830 && call_used_regs[i]))
3831 && !REGNO_REG_SET_P (df_get_live_out (bb_in), i))
3833 scratchreg = gen_rtx_REG (SImode, i);
3836 for (insn = BB_END (bb_in); insn != BB_HEAD (bb_in);
3837 insn = PREV_INSN (insn))
3840 if (NOTE_P (insn) || BARRIER_P (insn))
3842 set = single_set (insn);
3843 if (set && rtx_equal_p (SET_DEST (set), iter_reg))
3845 if (CONSTANT_P (SET_SRC (set)))
3847 scratch_init = SET_SRC (set);
3848 scratch_init_insn = insn;
3852 else if (reg_mentioned_p (iter_reg, PATTERN (insn)))
3857 if (loop->incoming_src)
3859 /* Make sure the predecessor is before the loop start label, as required by
3860 the LSETUP instruction. */
3862 insn = BB_END (loop->incoming_src);
3863 /* If we have to insert the LSETUP before a jump, count that jump in the
3865 if (VEC_length (edge, loop->incoming) > 1
3866 || !(VEC_last (edge, loop->incoming)->flags & EDGE_FALLTHRU))
3868 gcc_assert (JUMP_P (insn));
3869 insn = PREV_INSN (insn);
3872 for (; insn && insn != loop->start_label; insn = NEXT_INSN (insn))
3873 length += length_for_loop (insn);
3878 fprintf (dump_file, ";; loop %d lsetup not before loop_start\n",
3883 /* Account for the pop of a scratch register where necessary. */
3884 if (!PREG_P (iter_reg) && scratchreg == NULL_RTX
3885 && ENABLE_WA_LOAD_LCREGS)
3888 if (length > MAX_LSETUP_DISTANCE)
3891 fprintf (dump_file, ";; loop %d lsetup too far away\n", loop->loop_no);
3896 /* Check if start_label appears before loop_end and calculate the
3897 offset between them. We calculate the length of instructions
3900 for (insn = loop->start_label;
3901 insn && insn != loop->loop_end;
3902 insn = NEXT_INSN (insn))
3903 length += length_for_loop (insn);
3908 fprintf (dump_file, ";; loop %d start_label not before loop_end\n",
3913 loop->length = length;
3914 if (loop->length > MAX_LOOP_LENGTH)
3917 fprintf (dump_file, ";; loop %d too long\n", loop->loop_no);
3921 /* Scan all the blocks to make sure they don't use iter_reg. */
3922 if (bfin_scan_loop (loop, iter_reg, loop->loop_end))
3925 fprintf (dump_file, ";; loop %d uses iterator\n", loop->loop_no);
3929 /* Scan all the insns to see if the loop body clobber
3930 any hardware loop registers. */
3932 reg_lc0 = gen_rtx_REG (SImode, REG_LC0);
3933 reg_lc1 = gen_rtx_REG (SImode, REG_LC1);
3934 reg_lt0 = gen_rtx_REG (SImode, REG_LT0);
3935 reg_lt1 = gen_rtx_REG (SImode, REG_LT1);
3936 reg_lb0 = gen_rtx_REG (SImode, REG_LB0);
3937 reg_lb1 = gen_rtx_REG (SImode, REG_LB1);
3939 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, bb); ix++)
3943 for (insn = BB_HEAD (bb);
3944 insn != NEXT_INSN (BB_END (bb));
3945 insn = NEXT_INSN (insn))
3950 if (reg_set_p (reg_lc0, insn)
3951 || reg_set_p (reg_lt0, insn)
3952 || reg_set_p (reg_lb0, insn))
3953 loop->clobber_loop0 = 1;
3955 if (reg_set_p (reg_lc1, insn)
3956 || reg_set_p (reg_lt1, insn)
3957 || reg_set_p (reg_lb1, insn))
3958 loop->clobber_loop1 |= 1;
3962 if ((loop->clobber_loop0 && loop->clobber_loop1)
3963 || (loop->depth == MAX_LOOP_DEPTH && loop->clobber_loop0))
3965 loop->depth = MAX_LOOP_DEPTH + 1;
3967 fprintf (dump_file, ";; loop %d no loop reg available\n",
3972 /* There should be an instruction before the loop_end instruction
3973 in the same basic block. And the instruction must not be
3975 - CONDITIONAL BRANCH
3979 - Returns (RTS, RTN, etc.) */
3982 last_insn = find_prev_insn_start (loop->loop_end);
3986 for (; last_insn != BB_HEAD (bb);
3987 last_insn = find_prev_insn_start (last_insn))
3988 if (INSN_P (last_insn))
3991 if (last_insn != BB_HEAD (bb))
3994 if (single_pred_p (bb)
3995 && single_pred_edge (bb)->flags & EDGE_FALLTHRU
3996 && single_pred (bb) != ENTRY_BLOCK_PTR)
3998 bb = single_pred (bb);
3999 last_insn = BB_END (bb);
4004 last_insn = NULL_RTX;
4012 fprintf (dump_file, ";; loop %d has no last instruction\n",
4017 if (JUMP_P (last_insn) && !any_condjump_p (last_insn))
4020 fprintf (dump_file, ";; loop %d has bad last instruction\n",
4024 /* In all other cases, try to replace a bad last insn with a nop. */
4025 else if (JUMP_P (last_insn)
4026 || CALL_P (last_insn)
4027 || get_attr_type (last_insn) == TYPE_SYNC
4028 || get_attr_type (last_insn) == TYPE_CALL
4029 || get_attr_seq_insns (last_insn) == SEQ_INSNS_MULTI
4030 || recog_memoized (last_insn) == CODE_FOR_return_internal
4031 || GET_CODE (PATTERN (last_insn)) == ASM_INPUT
4032 || asm_noperands (PATTERN (last_insn)) >= 0)
4034 if (loop->length + 2 > MAX_LOOP_LENGTH)
4037 fprintf (dump_file, ";; loop %d too long\n", loop->loop_no);
4041 fprintf (dump_file, ";; loop %d has bad last insn; replace with nop\n",
4044 last_insn = emit_insn_after (gen_forced_nop (), last_insn);
4047 loop->last_insn = last_insn;
4049 /* The loop is good for replacement. */
4050 start_label = loop->start_label;
4051 end_label = gen_label_rtx ();
4052 iter_reg = loop->iter_reg;
4054 if (loop->depth == 1 && !loop->clobber_loop1)
4059 loop->clobber_loop1 = 1;
4066 loop->clobber_loop0 = 1;
4069 loop->end_label = end_label;
4071 /* Create a sequence containing the loop setup. */
4074 /* LSETUP only accepts P registers. If we have one, we can use it,
4075 otherwise there are several ways of working around the problem.
4076 If we're not affected by anomaly 312, we can load the LC register
4077 from any iteration register, and use LSETUP without initialization.
4078 If we've found a P scratch register that's not live here, we can
4079 instead copy the iter_reg into that and use an initializing LSETUP.
4080 If all else fails, push and pop P0 and use it as a scratch. */
4081 if (P_REGNO_P (REGNO (iter_reg)))
4083 loop_init = gen_lsetup_with_autoinit (lt_reg, start_label,
4086 seq_end = emit_insn (loop_init);
4088 else if (!ENABLE_WA_LOAD_LCREGS && DPREG_P (iter_reg))
4090 emit_insn (gen_movsi (lc_reg, iter_reg));
4091 loop_init = gen_lsetup_without_autoinit (lt_reg, start_label,
4094 seq_end = emit_insn (loop_init);
4096 else if (scratchreg != NULL_RTX)
4098 emit_insn (gen_movsi (scratchreg, scratch_init));
4099 loop_init = gen_lsetup_with_autoinit (lt_reg, start_label,
4101 lc_reg, scratchreg);
4102 seq_end = emit_insn (loop_init);
4103 if (scratch_init_insn != NULL_RTX)
4104 delete_insn (scratch_init_insn);
4108 rtx p0reg = gen_rtx_REG (SImode, REG_P0);
4109 rtx push = gen_frame_mem (SImode,
4110 gen_rtx_PRE_DEC (SImode, stack_pointer_rtx));
4111 rtx pop = gen_frame_mem (SImode,
4112 gen_rtx_POST_INC (SImode, stack_pointer_rtx));
4113 emit_insn (gen_movsi (push, p0reg));
4114 emit_insn (gen_movsi (p0reg, scratch_init));
4115 loop_init = gen_lsetup_with_autoinit (lt_reg, start_label,
4118 emit_insn (loop_init);
4119 seq_end = emit_insn (gen_movsi (p0reg, pop));
4120 if (scratch_init_insn != NULL_RTX)
4121 delete_insn (scratch_init_insn);
4126 fprintf (dump_file, ";; replacing loop %d initializer with\n",
4128 print_rtl_single (dump_file, loop_init);
4129 fprintf (dump_file, ";; replacing loop %d terminator with\n",
4131 print_rtl_single (dump_file, loop->loop_end);
4134 /* If the loop isn't entered at the top, also create a jump to the entry
4136 if (!loop->incoming_src && loop->head != loop->incoming_dest)
4138 rtx label = BB_HEAD (loop->incoming_dest);
4139 /* If we're jumping to the final basic block in the loop, and there's
4140 only one cheap instruction before the end (typically an increment of
4141 an induction variable), we can just emit a copy here instead of a
4143 if (loop->incoming_dest == loop->tail
4144 && next_real_insn (label) == last_insn
4145 && asm_noperands (last_insn) < 0
4146 && GET_CODE (PATTERN (last_insn)) == SET)
4148 seq_end = emit_insn (copy_rtx (PATTERN (last_insn)));
4151 seq_end = emit_jump_insn (gen_jump (label));
4157 if (loop->incoming_src)
4159 rtx prev = BB_END (loop->incoming_src);
4160 if (VEC_length (edge, loop->incoming) > 1
4161 || !(VEC_last (edge, loop->incoming)->flags & EDGE_FALLTHRU))
4163 gcc_assert (JUMP_P (prev));
4164 prev = PREV_INSN (prev);
4166 emit_insn_after (seq, prev);
4174 #ifdef ENABLE_CHECKING
4175 if (loop->head != loop->incoming_dest)
4177 /* We aren't entering the loop at the top. Since we've established
4178 that the loop is entered only at one point, this means there
4179 can't be fallthru edges into the head. Any such fallthru edges
4180 would become invalid when we insert the new block, so verify
4181 that this does not in fact happen. */
4182 FOR_EACH_EDGE (e, ei, loop->head->preds)
4183 gcc_assert (!(e->flags & EDGE_FALLTHRU));
4187 emit_insn_before (seq, BB_HEAD (loop->head));
4188 seq = emit_label_before (gen_label_rtx (), seq);
4190 new_bb = create_basic_block (seq, seq_end, loop->head->prev_bb);
4191 FOR_EACH_EDGE (e, ei, loop->incoming)
4193 if (!(e->flags & EDGE_FALLTHRU)
4194 || e->dest != loop->head)
4195 redirect_edge_and_branch_force (e, new_bb);
4197 redirect_edge_succ (e, new_bb);
4201 delete_insn (loop->loop_end);
4202 /* Insert the loop end label before the last instruction of the loop. */
4203 emit_label_before (loop->end_label, loop->last_insn);
4210 fprintf (dump_file, ";; loop %d is bad\n", loop->loop_no);
4214 if (DPREG_P (loop->iter_reg))
4216 /* If loop->iter_reg is a DREG or PREG, we can split it here
4217 without scratch register. */
4220 emit_insn_before (gen_addsi3 (loop->iter_reg,
4225 test = gen_rtx_NE (VOIDmode, loop->iter_reg, const0_rtx);
4226 insn = emit_jump_insn_before (gen_cbranchsi4 (test,
4227 loop->iter_reg, const0_rtx,
4231 JUMP_LABEL (insn) = loop->start_label;
4232 LABEL_NUSES (loop->start_label)++;
4233 delete_insn (loop->loop_end);
4237 /* Called from bfin_reorg_loops when a potential loop end is found. LOOP is
4238 a newly set up structure describing the loop, it is this function's
4239 responsibility to fill most of it. TAIL_BB and TAIL_INSN point to the
4240 loop_end insn and its enclosing basic block. */
4243 bfin_discover_loop (loop_info loop, basic_block tail_bb, rtx tail_insn)
4247 VEC (basic_block,heap) *works = VEC_alloc (basic_block,heap,20);
4249 loop->tail = tail_bb;
4250 loop->head = BRANCH_EDGE (tail_bb)->dest;
4251 loop->successor = FALLTHRU_EDGE (tail_bb)->dest;
4252 loop->loop_end = tail_insn;
4253 loop->last_insn = NULL_RTX;
4254 loop->iter_reg = SET_DEST (XVECEXP (PATTERN (tail_insn), 0, 1));
4255 loop->depth = loop->length = 0;
4257 loop->clobber_loop0 = loop->clobber_loop1 = 0;
4260 loop->incoming = VEC_alloc (edge, gc, 2);
4261 loop->start_label = XEXP (XEXP (SET_SRC (XVECEXP (PATTERN (tail_insn), 0, 0)), 1), 0);
4262 loop->end_label = NULL_RTX;
4265 VEC_safe_push (basic_block, heap, works, loop->head);
4267 while (VEC_iterate (basic_block, works, dwork++, bb))
4271 if (bb == EXIT_BLOCK_PTR)
4273 /* We've reached the exit block. The loop must be bad. */
4276 ";; Loop is bad - reached exit block while scanning\n");
4281 if (bitmap_bit_p (loop->block_bitmap, bb->index))
4284 /* We've not seen this block before. Add it to the loop's
4285 list and then add each successor to the work list. */
4287 VEC_safe_push (basic_block, heap, loop->blocks, bb);
4288 bitmap_set_bit (loop->block_bitmap, bb->index);
4292 FOR_EACH_EDGE (e, ei, bb->succs)
4294 basic_block succ = EDGE_SUCC (bb, ei.index)->dest;
4295 if (!REGNO_REG_SET_P (df_get_live_in (succ),
4296 REGNO (loop->iter_reg)))
4298 if (!VEC_space (basic_block, works, 1))
4302 VEC_block_remove (basic_block, works, 0, dwork);
4306 VEC_reserve (basic_block, heap, works, 1);
4308 VEC_quick_push (basic_block, works, succ);
4313 /* Find the predecessor, and make sure nothing else jumps into this loop. */
4317 for (dwork = 0; VEC_iterate (basic_block, loop->blocks, dwork, bb); dwork++)
4321 FOR_EACH_EDGE (e, ei, bb->preds)
4323 basic_block pred = e->src;
4325 if (!bfin_bb_in_loop (loop, pred))
4328 fprintf (dump_file, ";; Loop %d: incoming edge %d -> %d\n",
4329 loop->loop_no, pred->index,
4331 VEC_safe_push (edge, gc, loop->incoming, e);
4336 for (pass = 0, retry = 1; retry && pass < 2; pass++)
4343 FOR_EACH_EDGE (e, ei, loop->incoming)
4347 loop->incoming_src = e->src;
4348 loop->incoming_dest = e->dest;
4353 if (e->dest != loop->incoming_dest)
4354 loop->incoming_dest = NULL;
4355 if (e->src != loop->incoming_src)
4356 loop->incoming_src = NULL;
4358 if (loop->incoming_src == NULL && loop->incoming_dest == NULL)
4364 ";; retrying loop %d with forwarder blocks\n",
4372 ";; can't find suitable entry for loop %d\n",
4380 FOR_EACH_EDGE (e, ei, loop->incoming)
4382 if (forwarder_block_p (e->src))
4389 ";; Adding forwarder block %d to loop %d and retrying\n",
4390 e->src->index, loop->loop_no);
4391 VEC_safe_push (basic_block, heap, loop->blocks, e->src);
4392 bitmap_set_bit (loop->block_bitmap, e->src->index);
4393 FOR_EACH_EDGE (e2, ei2, e->src->preds)
4394 VEC_safe_push (edge, gc, loop->incoming, e2);
4395 VEC_unordered_remove (edge, loop->incoming, ei.index);
4403 fprintf (dump_file, ";; No forwarder blocks found\n");
4411 VEC_free (basic_block, heap, works);
4414 /* Analyze the structure of the loops in the current function. Use STACK
4415 for bitmap allocations. Returns all the valid candidates for hardware
4416 loops found in this function. */
4418 bfin_discover_loops (bitmap_obstack *stack, FILE *dump_file)
4420 loop_info loops = NULL;
4426 /* Find all the possible loop tails. This means searching for every
4427 loop_end instruction. For each one found, create a loop_info
4428 structure and add the head block to the work list. */
4431 rtx tail = BB_END (bb);
4433 while (GET_CODE (tail) == NOTE)
4434 tail = PREV_INSN (tail);
4438 if (INSN_P (tail) && recog_memoized (tail) == CODE_FOR_loop_end)
4441 /* A possible loop end */
4443 /* There's a degenerate case we can handle - an empty loop consisting
4444 of only a back branch. Handle that by deleting the branch. */
4445 insn = BB_HEAD (BRANCH_EDGE (bb)->dest);
4446 if (next_real_insn (insn) == tail)
4450 fprintf (dump_file, ";; degenerate loop ending at\n");
4451 print_rtl_single (dump_file, tail);
4453 delete_insn_and_edges (tail);
4457 loop = XNEW (struct loop_info);
4460 loop->loop_no = nloops++;
4461 loop->blocks = VEC_alloc (basic_block, heap, 20);
4462 loop->block_bitmap = BITMAP_ALLOC (stack);
4467 fprintf (dump_file, ";; potential loop %d ending at\n",
4469 print_rtl_single (dump_file, tail);
4472 bfin_discover_loop (loop, bb, tail);
4476 tmp_bitmap = BITMAP_ALLOC (stack);
4477 /* Compute loop nestings. */
4478 for (loop = loops; loop; loop = loop->next)
4484 for (other = loop->next; other; other = other->next)
4489 bitmap_and (tmp_bitmap, other->block_bitmap, loop->block_bitmap);
4490 if (bitmap_empty_p (tmp_bitmap))
4492 if (bitmap_equal_p (tmp_bitmap, other->block_bitmap))
4494 other->outer = loop;
4495 VEC_safe_push (loop_info, heap, loop->loops, other);
4497 else if (bitmap_equal_p (tmp_bitmap, loop->block_bitmap))
4499 loop->outer = other;
4500 VEC_safe_push (loop_info, heap, other->loops, loop);
4506 ";; can't find suitable nesting for loops %d and %d\n",
4507 loop->loop_no, other->loop_no);
4508 loop->bad = other->bad = 1;
4512 BITMAP_FREE (tmp_bitmap);
4517 /* Free up the loop structures in LOOPS. */
4519 free_loops (loop_info loops)
4523 loop_info loop = loops;
4525 VEC_free (loop_info, heap, loop->loops);
4526 VEC_free (basic_block, heap, loop->blocks);
4527 BITMAP_FREE (loop->block_bitmap);
4532 #define BB_AUX_INDEX(BB) ((unsigned)(BB)->aux)
4534 /* The taken-branch edge from the loop end can actually go forward. Since the
4535 Blackfin's LSETUP instruction requires that the loop end be after the loop
4536 start, try to reorder a loop's basic blocks when we find such a case. */
4538 bfin_reorder_loops (loop_info loops, FILE *dump_file)
4545 cfg_layout_initialize (0);
4547 for (loop = loops; loop; loop = loop->next)
4557 /* Recreate an index for basic blocks that represents their order. */
4558 for (bb = ENTRY_BLOCK_PTR->next_bb, index = 0;
4559 bb != EXIT_BLOCK_PTR;
4560 bb = bb->next_bb, index++)
4561 bb->aux = (PTR) index;
4563 if (BB_AUX_INDEX (loop->head) < BB_AUX_INDEX (loop->tail))
4566 FOR_EACH_EDGE (e, ei, loop->head->succs)
4568 if (bitmap_bit_p (loop->block_bitmap, e->dest->index)
4569 && BB_AUX_INDEX (e->dest) < BB_AUX_INDEX (loop->tail))
4571 basic_block start_bb = e->dest;
4572 basic_block start_prev_bb = start_bb->prev_bb;
4575 fprintf (dump_file, ";; Moving block %d before block %d\n",
4576 loop->head->index, start_bb->index);
4577 loop->head->prev_bb->next_bb = loop->head->next_bb;
4578 loop->head->next_bb->prev_bb = loop->head->prev_bb;
4580 loop->head->prev_bb = start_prev_bb;
4581 loop->head->next_bb = start_bb;
4582 start_prev_bb->next_bb = start_bb->prev_bb = loop->head;
4586 loops = loops->next;
4591 if (bb->next_bb != EXIT_BLOCK_PTR)
4592 bb->aux = bb->next_bb;
4596 cfg_layout_finalize ();
4600 /* Run from machine_dependent_reorg, this pass looks for doloop_end insns
4601 and tries to rewrite the RTL of these loops so that proper Blackfin
4602 hardware loops are generated. */
4605 bfin_reorg_loops (FILE *dump_file)
4607 loop_info loops = NULL;
4610 bitmap_obstack stack;
4612 bitmap_obstack_initialize (&stack);
4615 fprintf (dump_file, ";; Find loops, first pass\n\n");
4617 loops = bfin_discover_loops (&stack, dump_file);
4620 bfin_dump_loops (loops);
4622 bfin_reorder_loops (loops, dump_file);
4626 fprintf (dump_file, ";; Find loops, second pass\n\n");
4628 loops = bfin_discover_loops (&stack, dump_file);
4631 fprintf (dump_file, ";; All loops found:\n\n");
4632 bfin_dump_loops (loops);
4635 /* Now apply the optimizations. */
4636 for (loop = loops; loop; loop = loop->next)
4637 bfin_optimize_loop (loop);
4641 fprintf (dump_file, ";; After hardware loops optimization:\n\n");
4642 bfin_dump_loops (loops);
4648 print_rtl (dump_file, get_insns ());
4653 splitting_loops = 1;
4656 rtx insn = BB_END (bb);
4660 try_split (PATTERN (insn), insn, 1);
4662 splitting_loops = 0;
4665 /* Possibly generate a SEQUENCE out of three insns found in SLOT.
4666 Returns true if we modified the insn chain, false otherwise. */
4668 gen_one_bundle (rtx slot[3])
4670 gcc_assert (slot[1] != NULL_RTX);
4672 /* Don't add extra NOPs if optimizing for size. */
4674 && (slot[0] == NULL_RTX || slot[2] == NULL_RTX))
4677 /* Verify that we really can do the multi-issue. */
4680 rtx t = NEXT_INSN (slot[0]);
4681 while (t != slot[1])
4683 if (GET_CODE (t) != NOTE
4684 || NOTE_KIND (t) != NOTE_INSN_DELETED)
4691 rtx t = NEXT_INSN (slot[1]);
4692 while (t != slot[2])
4694 if (GET_CODE (t) != NOTE
4695 || NOTE_KIND (t) != NOTE_INSN_DELETED)
4701 if (slot[0] == NULL_RTX)
4703 slot[0] = emit_insn_before (gen_mnop (), slot[1]);
4704 df_insn_rescan (slot[0]);
4706 if (slot[2] == NULL_RTX)
4708 slot[2] = emit_insn_after (gen_forced_nop (), slot[1]);
4709 df_insn_rescan (slot[2]);
4712 /* Avoid line number information being printed inside one bundle. */
4713 if (INSN_LOCATOR (slot[1])
4714 && INSN_LOCATOR (slot[1]) != INSN_LOCATOR (slot[0]))
4715 INSN_LOCATOR (slot[1]) = INSN_LOCATOR (slot[0]);
4716 if (INSN_LOCATOR (slot[2])
4717 && INSN_LOCATOR (slot[2]) != INSN_LOCATOR (slot[0]))
4718 INSN_LOCATOR (slot[2]) = INSN_LOCATOR (slot[0]);
4720 /* Terminate them with "|| " instead of ";" in the output. */
4721 PUT_MODE (slot[0], SImode);
4722 PUT_MODE (slot[1], SImode);
4723 /* Terminate the bundle, for the benefit of reorder_var_tracking_notes. */
4724 PUT_MODE (slot[2], QImode);
4728 /* Go through all insns, and use the information generated during scheduling
4729 to generate SEQUENCEs to represent bundles of instructions issued
4733 bfin_gen_bundles (void)
4742 slot[0] = slot[1] = slot[2] = NULL_RTX;
4743 for (insn = BB_HEAD (bb);; insn = next)
4748 if (get_attr_type (insn) == TYPE_DSP32)
4750 else if (slot[1] == NULL_RTX)
4757 next = NEXT_INSN (insn);
4758 while (next && insn != BB_END (bb)
4760 && GET_CODE (PATTERN (next)) != USE
4761 && GET_CODE (PATTERN (next)) != CLOBBER))
4764 next = NEXT_INSN (insn);
4767 /* BB_END can change due to emitting extra NOPs, so check here. */
4768 at_end = insn == BB_END (bb);
4769 if (at_end || GET_MODE (next) == TImode)
4772 || !gen_one_bundle (slot))
4773 && slot[0] != NULL_RTX)
4775 rtx pat = PATTERN (slot[0]);
4776 if (GET_CODE (pat) == SET
4777 && GET_CODE (SET_SRC (pat)) == UNSPEC
4778 && XINT (SET_SRC (pat), 1) == UNSPEC_32BIT)
4780 SET_SRC (pat) = XVECEXP (SET_SRC (pat), 0, 0);
4781 INSN_CODE (slot[0]) = -1;
4782 df_insn_rescan (slot[0]);
4786 slot[0] = slot[1] = slot[2] = NULL_RTX;
4794 /* Ensure that no var tracking notes are emitted in the middle of a
4795 three-instruction bundle. */
4798 reorder_var_tracking_notes (void)
4804 rtx queue = NULL_RTX;
4805 bool in_bundle = false;
4807 for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = next)
4809 next = NEXT_INSN (insn);
4813 /* Emit queued up notes at the last instruction of a bundle. */
4814 if (GET_MODE (insn) == QImode)
4818 rtx next_queue = PREV_INSN (queue);
4819 PREV_INSN (NEXT_INSN (insn)) = queue;
4820 NEXT_INSN (queue) = NEXT_INSN (insn);
4821 NEXT_INSN (insn) = queue;
4822 PREV_INSN (queue) = insn;
4827 else if (GET_MODE (insn) == SImode)
4830 else if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_VAR_LOCATION)
4834 rtx prev = PREV_INSN (insn);
4835 PREV_INSN (next) = prev;
4836 NEXT_INSN (prev) = next;
4838 PREV_INSN (insn) = queue;
4846 /* On some silicon revisions, functions shorter than a certain number of cycles
4847 can cause unpredictable behaviour. Work around this by adding NOPs as
4850 workaround_rts_anomaly (void)
4852 rtx insn, first_insn = NULL_RTX;
4855 if (! ENABLE_WA_RETS)
4858 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
4862 if (BARRIER_P (insn))
4865 if (NOTE_P (insn) || LABEL_P (insn))
4868 if (first_insn == NULL_RTX)
4870 pat = PATTERN (insn);
4871 if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
4872 || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
4873 || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
4881 if (recog_memoized (insn) == CODE_FOR_return_internal)
4884 /* Nothing to worry about for direct jumps. */
4885 if (!any_condjump_p (insn))
4891 else if (INSN_P (insn))
4893 rtx pat = PATTERN (insn);
4894 int this_cycles = 1;
4896 if (GET_CODE (pat) == PARALLEL)
4898 if (push_multiple_operation (pat, VOIDmode)
4899 || pop_multiple_operation (pat, VOIDmode))
4900 this_cycles = n_regs_to_save;
4904 enum insn_code icode = recog_memoized (insn);
4905 if (icode == CODE_FOR_link)
4907 else if (icode == CODE_FOR_unlink)
4909 else if (icode == CODE_FOR_mulsi3)
4912 if (this_cycles >= cycles)
4915 cycles -= this_cycles;
4920 emit_insn_before (gen_nop (), first_insn);
4925 /* Return an insn type for INSN that can be used by the caller for anomaly
4926 workarounds. This differs from plain get_attr_type in that it handles
4929 static enum attr_type
4930 type_for_anomaly (rtx insn)
4932 rtx pat = PATTERN (insn);
4933 if (GET_CODE (pat) == SEQUENCE)
4936 t = get_attr_type (XVECEXP (pat, 0, 1));
4939 t = get_attr_type (XVECEXP (pat, 0, 2));
4945 return get_attr_type (insn);
4948 /* Return nonzero if INSN contains any loads that may trap. It handles
4949 SEQUENCEs correctly. */
4952 trapping_loads_p (rtx insn)
4954 rtx pat = PATTERN (insn);
4955 if (GET_CODE (pat) == SEQUENCE)
4958 t = get_attr_type (XVECEXP (pat, 0, 1));
4960 && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 1)))))
4962 t = get_attr_type (XVECEXP (pat, 0, 2));
4964 && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 2)))))
4969 return may_trap_p (SET_SRC (single_set (insn)));
4972 /* Return INSN if it is of TYPE_MCLD. Alternatively, if INSN is the start of
4973 a three-insn bundle, see if one of them is a load and return that if so.
4974 Return NULL_RTX if the insn does not contain loads. */
4976 find_load (rtx insn)
4978 if (get_attr_type (insn) == TYPE_MCLD)
4980 if (GET_MODE (insn) != SImode)
4983 insn = NEXT_INSN (insn);
4984 if ((GET_MODE (insn) == SImode || GET_MODE (insn) == QImode)
4985 && get_attr_type (insn) == TYPE_MCLD)
4987 } while (GET_MODE (insn) != QImode);
4991 /* Determine whether PAT is an indirect call pattern. */
4993 indirect_call_p (rtx pat)
4995 if (GET_CODE (pat) == PARALLEL)
4996 pat = XVECEXP (pat, 0, 0);
4997 if (GET_CODE (pat) == SET)
4998 pat = SET_SRC (pat);
4999 gcc_assert (GET_CODE (pat) == CALL);
5000 pat = XEXP (pat, 0);
5001 gcc_assert (GET_CODE (pat) == MEM);
5002 pat = XEXP (pat, 0);
5008 workaround_speculation (void)
5011 rtx last_condjump = NULL_RTX;
5012 int cycles_since_jump = INT_MAX;
5013 int delay_added = 0;
5015 if (! ENABLE_WA_SPECULATIVE_LOADS && ! ENABLE_WA_SPECULATIVE_SYNCS
5016 && ! ENABLE_WA_INDIRECT_CALLS)
5019 /* First pass: find predicted-false branches; if something after them
5020 needs nops, insert them or change the branch to predict true. */
5021 for (insn = get_insns (); insn; insn = next)
5024 int delay_needed = 0;
5026 next = find_next_insn_start (insn);
5028 if (NOTE_P (insn) || BARRIER_P (insn) || LABEL_P (insn))
5031 pat = PATTERN (insn);
5032 if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
5033 || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
5034 || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
5039 if (any_condjump_p (insn)
5040 && ! cbranch_predicted_taken_p (insn))
5042 last_condjump = insn;
5044 cycles_since_jump = 0;
5047 cycles_since_jump = INT_MAX;
5049 else if (CALL_P (insn))
5051 if (cycles_since_jump < INT_MAX)
5052 cycles_since_jump++;
5053 if (indirect_call_p (pat) && ENABLE_WA_INDIRECT_CALLS)
5058 else if (INSN_P (insn))
5060 rtx load_insn = find_load (insn);
5061 enum attr_type type = type_for_anomaly (insn);
5063 if (cycles_since_jump < INT_MAX)
5064 cycles_since_jump++;
5066 if (load_insn && ENABLE_WA_SPECULATIVE_LOADS)
5068 if (trapping_loads_p (load_insn))
5071 else if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
5075 if (delay_needed > cycles_since_jump
5076 && (delay_needed - cycles_since_jump) > delay_added)
5080 rtx *op = recog_data.operand;
5082 delay_needed -= cycles_since_jump;
5084 extract_insn (last_condjump);
5087 pat1 = gen_cbranch_predicted_taken (op[0], op[1], op[2],
5089 cycles_since_jump = INT_MAX;
5093 /* Do not adjust cycles_since_jump in this case, so that
5094 we'll increase the number of NOPs for a subsequent insn
5096 pat1 = gen_cbranch_with_nops (op[0], op[1], op[2], op[3],
5097 GEN_INT (delay_needed));
5098 delay_added = delay_needed;
5100 PATTERN (last_condjump) = pat1;
5101 INSN_CODE (last_condjump) = recog (pat1, insn, &num_clobbers);
5105 cycles_since_jump = INT_MAX;
5110 /* Second pass: for predicted-true branches, see if anything at the
5111 branch destination needs extra nops. */
5112 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
5114 int cycles_since_jump;
5116 && any_condjump_p (insn)
5117 && (INSN_CODE (insn) == CODE_FOR_cbranch_predicted_taken
5118 || cbranch_predicted_taken_p (insn)))
5120 rtx target = JUMP_LABEL (insn);
5124 cycles_since_jump = 0;
5125 for (; target && cycles_since_jump < 3; target = next_tgt)
5129 next_tgt = find_next_insn_start (target);
5131 if (NOTE_P (target) || BARRIER_P (target) || LABEL_P (target))
5134 pat = PATTERN (target);
5135 if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
5136 || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
5137 || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
5140 if (INSN_P (target))
5142 rtx load_insn = find_load (target);
5143 enum attr_type type = type_for_anomaly (target);
5144 int delay_needed = 0;
5145 if (cycles_since_jump < INT_MAX)
5146 cycles_since_jump++;
5148 if (load_insn && ENABLE_WA_SPECULATIVE_LOADS)
5150 if (trapping_loads_p (load_insn))
5153 else if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
5156 if (delay_needed > cycles_since_jump)
5158 rtx prev = prev_real_insn (label);
5159 delay_needed -= cycles_since_jump;
5161 fprintf (dump_file, "Adding %d nops after %d\n",
5162 delay_needed, INSN_UID (label));
5164 && INSN_CODE (prev) == CODE_FOR_cbranch_with_nops)
5171 "Reducing nops on insn %d.\n",
5174 x = XVECEXP (x, 0, 1);
5175 v = INTVAL (XVECEXP (x, 0, 0)) - delay_needed;
5176 XVECEXP (x, 0, 0) = GEN_INT (v);
5178 while (delay_needed-- > 0)
5179 emit_insn_after (gen_nop (), label);
5188 /* We use the machine specific reorg pass for emitting CSYNC instructions
5189 after conditional branches as needed.
5191 The Blackfin is unusual in that a code sequence like
5194 may speculatively perform the load even if the condition isn't true. This
5195 happens for a branch that is predicted not taken, because the pipeline
5196 isn't flushed or stalled, so the early stages of the following instructions,
5197 which perform the memory reference, are allowed to execute before the
5198 jump condition is evaluated.
5199 Therefore, we must insert additional instructions in all places where this
5200 could lead to incorrect behavior. The manual recommends CSYNC, while
5201 VDSP seems to use NOPs (even though its corresponding compiler option is
5204 When optimizing for speed, we emit NOPs, which seems faster than a CSYNC.
5205 When optimizing for size, we turn the branch into a predicted taken one.
5206 This may be slower due to mispredicts, but saves code size. */
5211 /* We are freeing block_for_insn in the toplev to keep compatibility
5212 with old MDEP_REORGS that are not CFG based. Recompute it now. */
5213 compute_bb_for_insn ();
5215 if (bfin_flag_schedule_insns2)
5217 splitting_for_sched = 1;
5219 splitting_for_sched = 0;
5221 timevar_push (TV_SCHED2);
5223 timevar_pop (TV_SCHED2);
5225 /* Examine the schedule and insert nops as necessary for 64-bit parallel
5227 bfin_gen_bundles ();
5232 /* Doloop optimization */
5233 if (cfun->machine->has_hardware_loops)
5234 bfin_reorg_loops (dump_file);
5236 workaround_speculation ();
5238 if (bfin_flag_var_tracking)
5240 timevar_push (TV_VAR_TRACKING);
5241 variable_tracking_main ();
5242 reorder_var_tracking_notes ();
5243 timevar_pop (TV_VAR_TRACKING);
5246 df_finish_pass (false);
5248 workaround_rts_anomaly ();
5251 /* Handle interrupt_handler, exception_handler and nmi_handler function
5252 attributes; arguments as in struct attribute_spec.handler. */
5255 handle_int_attribute (tree *node, tree name,
5256 tree args ATTRIBUTE_UNUSED,
5257 int flags ATTRIBUTE_UNUSED,
5261 if (TREE_CODE (x) == FUNCTION_DECL)
5264 if (TREE_CODE (x) != FUNCTION_TYPE)
5266 warning (OPT_Wattributes, "%qE attribute only applies to functions",
5268 *no_add_attrs = true;
5270 else if (funkind (x) != SUBROUTINE)
5271 error ("multiple function type attributes specified");
5276 /* Return 0 if the attributes for two types are incompatible, 1 if they
5277 are compatible, and 2 if they are nearly compatible (which causes a
5278 warning to be generated). */
5281 bfin_comp_type_attributes (const_tree type1, const_tree type2)
5283 e_funkind kind1, kind2;
5285 if (TREE_CODE (type1) != FUNCTION_TYPE)
5288 kind1 = funkind (type1);
5289 kind2 = funkind (type2);
5294 /* Check for mismatched modifiers */
5295 if (!lookup_attribute ("nesting", TYPE_ATTRIBUTES (type1))
5296 != !lookup_attribute ("nesting", TYPE_ATTRIBUTES (type2)))
5299 if (!lookup_attribute ("saveall", TYPE_ATTRIBUTES (type1))
5300 != !lookup_attribute ("saveall", TYPE_ATTRIBUTES (type2)))
5303 if (!lookup_attribute ("kspisusp", TYPE_ATTRIBUTES (type1))
5304 != !lookup_attribute ("kspisusp", TYPE_ATTRIBUTES (type2)))
5307 if (!lookup_attribute ("longcall", TYPE_ATTRIBUTES (type1))
5308 != !lookup_attribute ("longcall", TYPE_ATTRIBUTES (type2)))
5314 /* Handle a "longcall" or "shortcall" attribute; arguments as in
5315 struct attribute_spec.handler. */
5318 bfin_handle_longcall_attribute (tree *node, tree name,
5319 tree args ATTRIBUTE_UNUSED,
5320 int flags ATTRIBUTE_UNUSED,
5323 if (TREE_CODE (*node) != FUNCTION_TYPE
5324 && TREE_CODE (*node) != FIELD_DECL
5325 && TREE_CODE (*node) != TYPE_DECL)
5327 warning (OPT_Wattributes, "%qE attribute only applies to functions",
5329 *no_add_attrs = true;
5332 if ((strcmp (IDENTIFIER_POINTER (name), "longcall") == 0
5333 && lookup_attribute ("shortcall", TYPE_ATTRIBUTES (*node)))
5334 || (strcmp (IDENTIFIER_POINTER (name), "shortcall") == 0
5335 && lookup_attribute ("longcall", TYPE_ATTRIBUTES (*node))))
5337 warning (OPT_Wattributes,
5338 "can't apply both longcall and shortcall attributes to the same function");
5339 *no_add_attrs = true;
5345 /* Handle a "l1_text" attribute; arguments as in
5346 struct attribute_spec.handler. */
5349 bfin_handle_l1_text_attribute (tree *node, tree name, tree ARG_UNUSED (args),
5350 int ARG_UNUSED (flags), bool *no_add_attrs)
5354 if (TREE_CODE (decl) != FUNCTION_DECL)
5356 error ("%qE attribute only applies to functions",
5358 *no_add_attrs = true;
5361 /* The decl may have already been given a section attribute
5362 from a previous declaration. Ensure they match. */
5363 else if (DECL_SECTION_NAME (decl) != NULL_TREE
5364 && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
5367 error ("section of %q+D conflicts with previous declaration",
5369 *no_add_attrs = true;
5372 DECL_SECTION_NAME (decl) = build_string (9, ".l1.text");
5377 /* Handle a "l1_data", "l1_data_A" or "l1_data_B" attribute;
5378 arguments as in struct attribute_spec.handler. */
5381 bfin_handle_l1_data_attribute (tree *node, tree name, tree ARG_UNUSED (args),
5382 int ARG_UNUSED (flags), bool *no_add_attrs)
5386 if (TREE_CODE (decl) != VAR_DECL)
5388 error ("%qE attribute only applies to variables",
5390 *no_add_attrs = true;
5392 else if (current_function_decl != NULL_TREE
5393 && !TREE_STATIC (decl))
5395 error ("%qE attribute cannot be specified for local variables",
5397 *no_add_attrs = true;
5401 const char *section_name;
5403 if (strcmp (IDENTIFIER_POINTER (name), "l1_data") == 0)
5404 section_name = ".l1.data";
5405 else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_A") == 0)
5406 section_name = ".l1.data.A";
5407 else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_B") == 0)
5408 section_name = ".l1.data.B";
5412 /* The decl may have already been given a section attribute
5413 from a previous declaration. Ensure they match. */
5414 if (DECL_SECTION_NAME (decl) != NULL_TREE
5415 && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
5418 error ("section of %q+D conflicts with previous declaration",
5420 *no_add_attrs = true;
5423 DECL_SECTION_NAME (decl)
5424 = build_string (strlen (section_name) + 1, section_name);
5430 /* Table of valid machine attributes. */
5431 const struct attribute_spec bfin_attribute_table[] =
5433 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
5434 { "interrupt_handler", 0, 0, false, true, true, handle_int_attribute },
5435 { "exception_handler", 0, 0, false, true, true, handle_int_attribute },
5436 { "nmi_handler", 0, 0, false, true, true, handle_int_attribute },
5437 { "nesting", 0, 0, false, true, true, NULL },
5438 { "kspisusp", 0, 0, false, true, true, NULL },
5439 { "saveall", 0, 0, false, true, true, NULL },
5440 { "longcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
5441 { "shortcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
5442 { "l1_text", 0, 0, true, false, false, bfin_handle_l1_text_attribute },
5443 { "l1_data", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5444 { "l1_data_A", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5445 { "l1_data_B", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5446 { NULL, 0, 0, false, false, false, NULL }
5449 /* Implementation of TARGET_ASM_INTEGER. When using FD-PIC, we need to
5450 tell the assembler to generate pointers to function descriptors in
5454 bfin_assemble_integer (rtx value, unsigned int size, int aligned_p)
5456 if (TARGET_FDPIC && size == UNITS_PER_WORD)
5458 if (GET_CODE (value) == SYMBOL_REF
5459 && SYMBOL_REF_FUNCTION_P (value))
5461 fputs ("\t.picptr\tfuncdesc(", asm_out_file);
5462 output_addr_const (asm_out_file, value);
5463 fputs (")\n", asm_out_file);
5468 /* We've set the unaligned SI op to NULL, so we always have to
5469 handle the unaligned case here. */
5470 assemble_integer_with_op ("\t.4byte\t", value);
5474 return default_assemble_integer (value, size, aligned_p);
5477 /* Output the assembler code for a thunk function. THUNK_DECL is the
5478 declaration for the thunk function itself, FUNCTION is the decl for
5479 the target function. DELTA is an immediate constant offset to be
5480 added to THIS. If VCALL_OFFSET is nonzero, the word at
5481 *(*this + vcall_offset) should be added to THIS. */
5484 bfin_output_mi_thunk (FILE *file ATTRIBUTE_UNUSED,
5485 tree thunk ATTRIBUTE_UNUSED, HOST_WIDE_INT delta,
5486 HOST_WIDE_INT vcall_offset, tree function)
5489 /* The this parameter is passed as the first argument. */
5490 rtx this_rtx = gen_rtx_REG (Pmode, REG_R0);
5492 /* Adjust the this parameter by a fixed constant. */
5496 if (delta >= -64 && delta <= 63)
5498 xops[0] = GEN_INT (delta);
5499 output_asm_insn ("%1 += %0;", xops);
5501 else if (delta >= -128 && delta < -64)
5503 xops[0] = GEN_INT (delta + 64);
5504 output_asm_insn ("%1 += -64; %1 += %0;", xops);
5506 else if (delta > 63 && delta <= 126)
5508 xops[0] = GEN_INT (delta - 63);
5509 output_asm_insn ("%1 += 63; %1 += %0;", xops);
5513 xops[0] = GEN_INT (delta);
5514 output_asm_insn ("r3.l = %h0; r3.h = %d0; %1 = %1 + r3;", xops);
5518 /* Adjust the this parameter by a value stored in the vtable. */
5521 rtx p2tmp = gen_rtx_REG (Pmode, REG_P2);
5522 rtx tmp = gen_rtx_REG (Pmode, REG_R3);
5526 output_asm_insn ("%2 = r0; %2 = [%2];", xops);
5528 /* Adjust the this parameter. */
5529 xops[0] = gen_rtx_MEM (Pmode, plus_constant (p2tmp, vcall_offset));
5530 if (!memory_operand (xops[0], Pmode))
5532 rtx tmp2 = gen_rtx_REG (Pmode, REG_P1);
5533 xops[0] = GEN_INT (vcall_offset);
5535 output_asm_insn ("%h1 = %h0; %d1 = %d0; %2 = %2 + %1", xops);
5536 xops[0] = gen_rtx_MEM (Pmode, p2tmp);
5539 output_asm_insn ("%1 = %0; %2 = %2 + %1;", xops);
5542 xops[0] = XEXP (DECL_RTL (function), 0);
5543 if (1 || !flag_pic || (*targetm.binds_local_p) (function))
5544 output_asm_insn ("jump.l\t%P0", xops);
5547 /* Codes for all the Blackfin builtins. */
5553 BFIN_BUILTIN_COMPOSE_2X16,
5554 BFIN_BUILTIN_EXTRACTLO,
5555 BFIN_BUILTIN_EXTRACTHI,
5557 BFIN_BUILTIN_SSADD_2X16,
5558 BFIN_BUILTIN_SSSUB_2X16,
5559 BFIN_BUILTIN_SSADDSUB_2X16,
5560 BFIN_BUILTIN_SSSUBADD_2X16,
5561 BFIN_BUILTIN_MULT_2X16,
5562 BFIN_BUILTIN_MULTR_2X16,
5563 BFIN_BUILTIN_NEG_2X16,
5564 BFIN_BUILTIN_ABS_2X16,
5565 BFIN_BUILTIN_MIN_2X16,
5566 BFIN_BUILTIN_MAX_2X16,
5568 BFIN_BUILTIN_SSADD_1X16,
5569 BFIN_BUILTIN_SSSUB_1X16,
5570 BFIN_BUILTIN_MULT_1X16,
5571 BFIN_BUILTIN_MULTR_1X16,
5572 BFIN_BUILTIN_NORM_1X16,
5573 BFIN_BUILTIN_NEG_1X16,
5574 BFIN_BUILTIN_ABS_1X16,
5575 BFIN_BUILTIN_MIN_1X16,
5576 BFIN_BUILTIN_MAX_1X16,
5578 BFIN_BUILTIN_SUM_2X16,
5579 BFIN_BUILTIN_DIFFHL_2X16,
5580 BFIN_BUILTIN_DIFFLH_2X16,
5582 BFIN_BUILTIN_SSADD_1X32,
5583 BFIN_BUILTIN_SSSUB_1X32,
5584 BFIN_BUILTIN_NORM_1X32,
5585 BFIN_BUILTIN_ROUND_1X32,
5586 BFIN_BUILTIN_NEG_1X32,
5587 BFIN_BUILTIN_ABS_1X32,
5588 BFIN_BUILTIN_MIN_1X32,
5589 BFIN_BUILTIN_MAX_1X32,
5590 BFIN_BUILTIN_MULT_1X32,
5591 BFIN_BUILTIN_MULT_1X32X32,
5592 BFIN_BUILTIN_MULT_1X32X32NS,
5594 BFIN_BUILTIN_MULHISILL,
5595 BFIN_BUILTIN_MULHISILH,
5596 BFIN_BUILTIN_MULHISIHL,
5597 BFIN_BUILTIN_MULHISIHH,
5599 BFIN_BUILTIN_LSHIFT_1X16,
5600 BFIN_BUILTIN_LSHIFT_2X16,
5601 BFIN_BUILTIN_SSASHIFT_1X16,
5602 BFIN_BUILTIN_SSASHIFT_2X16,
5603 BFIN_BUILTIN_SSASHIFT_1X32,
5605 BFIN_BUILTIN_CPLX_MUL_16,
5606 BFIN_BUILTIN_CPLX_MAC_16,
5607 BFIN_BUILTIN_CPLX_MSU_16,
5609 BFIN_BUILTIN_CPLX_MUL_16_S40,
5610 BFIN_BUILTIN_CPLX_MAC_16_S40,
5611 BFIN_BUILTIN_CPLX_MSU_16_S40,
5613 BFIN_BUILTIN_CPLX_SQU,
5615 BFIN_BUILTIN_LOADBYTES,
5620 #define def_builtin(NAME, TYPE, CODE) \
5622 add_builtin_function ((NAME), (TYPE), (CODE), BUILT_IN_MD, \
5626 /* Set up all builtin functions for this target. */
5628 bfin_init_builtins (void)
5630 tree V2HI_type_node = build_vector_type_for_mode (intHI_type_node, V2HImode);
5631 tree void_ftype_void
5632 = build_function_type (void_type_node, void_list_node);
5633 tree short_ftype_short
5634 = build_function_type_list (short_integer_type_node, short_integer_type_node,
5636 tree short_ftype_int_int
5637 = build_function_type_list (short_integer_type_node, integer_type_node,
5638 integer_type_node, NULL_TREE);
5639 tree int_ftype_int_int
5640 = build_function_type_list (integer_type_node, integer_type_node,
5641 integer_type_node, NULL_TREE);
5643 = build_function_type_list (integer_type_node, integer_type_node,
5645 tree short_ftype_int
5646 = build_function_type_list (short_integer_type_node, integer_type_node,
5648 tree int_ftype_v2hi_v2hi
5649 = build_function_type_list (integer_type_node, V2HI_type_node,
5650 V2HI_type_node, NULL_TREE);
5651 tree v2hi_ftype_v2hi_v2hi
5652 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5653 V2HI_type_node, NULL_TREE);
5654 tree v2hi_ftype_v2hi_v2hi_v2hi
5655 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5656 V2HI_type_node, V2HI_type_node, NULL_TREE);
5657 tree v2hi_ftype_int_int
5658 = build_function_type_list (V2HI_type_node, integer_type_node,
5659 integer_type_node, NULL_TREE);
5660 tree v2hi_ftype_v2hi_int
5661 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5662 integer_type_node, NULL_TREE);
5663 tree int_ftype_short_short
5664 = build_function_type_list (integer_type_node, short_integer_type_node,
5665 short_integer_type_node, NULL_TREE);
5666 tree v2hi_ftype_v2hi
5667 = build_function_type_list (V2HI_type_node, V2HI_type_node, NULL_TREE);
5668 tree short_ftype_v2hi
5669 = build_function_type_list (short_integer_type_node, V2HI_type_node,
5672 = build_function_type_list (integer_type_node,
5673 build_pointer_type (integer_type_node),
5676 /* Add the remaining MMX insns with somewhat more complicated types. */
5677 def_builtin ("__builtin_bfin_csync", void_ftype_void, BFIN_BUILTIN_CSYNC);
5678 def_builtin ("__builtin_bfin_ssync", void_ftype_void, BFIN_BUILTIN_SSYNC);
5680 def_builtin ("__builtin_bfin_ones", short_ftype_int, BFIN_BUILTIN_ONES);
5682 def_builtin ("__builtin_bfin_compose_2x16", v2hi_ftype_int_int,
5683 BFIN_BUILTIN_COMPOSE_2X16);
5684 def_builtin ("__builtin_bfin_extract_hi", short_ftype_v2hi,
5685 BFIN_BUILTIN_EXTRACTHI);
5686 def_builtin ("__builtin_bfin_extract_lo", short_ftype_v2hi,
5687 BFIN_BUILTIN_EXTRACTLO);
5689 def_builtin ("__builtin_bfin_min_fr2x16", v2hi_ftype_v2hi_v2hi,
5690 BFIN_BUILTIN_MIN_2X16);
5691 def_builtin ("__builtin_bfin_max_fr2x16", v2hi_ftype_v2hi_v2hi,
5692 BFIN_BUILTIN_MAX_2X16);
5694 def_builtin ("__builtin_bfin_add_fr2x16", v2hi_ftype_v2hi_v2hi,
5695 BFIN_BUILTIN_SSADD_2X16);
5696 def_builtin ("__builtin_bfin_sub_fr2x16", v2hi_ftype_v2hi_v2hi,
5697 BFIN_BUILTIN_SSSUB_2X16);
5698 def_builtin ("__builtin_bfin_dspaddsubsat", v2hi_ftype_v2hi_v2hi,
5699 BFIN_BUILTIN_SSADDSUB_2X16);
5700 def_builtin ("__builtin_bfin_dspsubaddsat", v2hi_ftype_v2hi_v2hi,
5701 BFIN_BUILTIN_SSSUBADD_2X16);
5702 def_builtin ("__builtin_bfin_mult_fr2x16", v2hi_ftype_v2hi_v2hi,
5703 BFIN_BUILTIN_MULT_2X16);
5704 def_builtin ("__builtin_bfin_multr_fr2x16", v2hi_ftype_v2hi_v2hi,
5705 BFIN_BUILTIN_MULTR_2X16);
5706 def_builtin ("__builtin_bfin_negate_fr2x16", v2hi_ftype_v2hi,
5707 BFIN_BUILTIN_NEG_2X16);
5708 def_builtin ("__builtin_bfin_abs_fr2x16", v2hi_ftype_v2hi,
5709 BFIN_BUILTIN_ABS_2X16);
5711 def_builtin ("__builtin_bfin_min_fr1x16", short_ftype_int_int,
5712 BFIN_BUILTIN_MIN_1X16);
5713 def_builtin ("__builtin_bfin_max_fr1x16", short_ftype_int_int,
5714 BFIN_BUILTIN_MAX_1X16);
5716 def_builtin ("__builtin_bfin_add_fr1x16", short_ftype_int_int,
5717 BFIN_BUILTIN_SSADD_1X16);
5718 def_builtin ("__builtin_bfin_sub_fr1x16", short_ftype_int_int,
5719 BFIN_BUILTIN_SSSUB_1X16);
5720 def_builtin ("__builtin_bfin_mult_fr1x16", short_ftype_int_int,
5721 BFIN_BUILTIN_MULT_1X16);
5722 def_builtin ("__builtin_bfin_multr_fr1x16", short_ftype_int_int,
5723 BFIN_BUILTIN_MULTR_1X16);
5724 def_builtin ("__builtin_bfin_negate_fr1x16", short_ftype_short,
5725 BFIN_BUILTIN_NEG_1X16);
5726 def_builtin ("__builtin_bfin_abs_fr1x16", short_ftype_short,
5727 BFIN_BUILTIN_ABS_1X16);
5728 def_builtin ("__builtin_bfin_norm_fr1x16", short_ftype_int,
5729 BFIN_BUILTIN_NORM_1X16);
5731 def_builtin ("__builtin_bfin_sum_fr2x16", short_ftype_v2hi,
5732 BFIN_BUILTIN_SUM_2X16);
5733 def_builtin ("__builtin_bfin_diff_hl_fr2x16", short_ftype_v2hi,
5734 BFIN_BUILTIN_DIFFHL_2X16);
5735 def_builtin ("__builtin_bfin_diff_lh_fr2x16", short_ftype_v2hi,
5736 BFIN_BUILTIN_DIFFLH_2X16);
5738 def_builtin ("__builtin_bfin_mulhisill", int_ftype_v2hi_v2hi,
5739 BFIN_BUILTIN_MULHISILL);
5740 def_builtin ("__builtin_bfin_mulhisihl", int_ftype_v2hi_v2hi,
5741 BFIN_BUILTIN_MULHISIHL);
5742 def_builtin ("__builtin_bfin_mulhisilh", int_ftype_v2hi_v2hi,
5743 BFIN_BUILTIN_MULHISILH);
5744 def_builtin ("__builtin_bfin_mulhisihh", int_ftype_v2hi_v2hi,
5745 BFIN_BUILTIN_MULHISIHH);
5747 def_builtin ("__builtin_bfin_min_fr1x32", int_ftype_int_int,
5748 BFIN_BUILTIN_MIN_1X32);
5749 def_builtin ("__builtin_bfin_max_fr1x32", int_ftype_int_int,
5750 BFIN_BUILTIN_MAX_1X32);
5752 def_builtin ("__builtin_bfin_add_fr1x32", int_ftype_int_int,
5753 BFIN_BUILTIN_SSADD_1X32);
5754 def_builtin ("__builtin_bfin_sub_fr1x32", int_ftype_int_int,
5755 BFIN_BUILTIN_SSSUB_1X32);
5756 def_builtin ("__builtin_bfin_negate_fr1x32", int_ftype_int,
5757 BFIN_BUILTIN_NEG_1X32);
5758 def_builtin ("__builtin_bfin_abs_fr1x32", int_ftype_int,
5759 BFIN_BUILTIN_ABS_1X32);
5760 def_builtin ("__builtin_bfin_norm_fr1x32", short_ftype_int,
5761 BFIN_BUILTIN_NORM_1X32);
5762 def_builtin ("__builtin_bfin_round_fr1x32", short_ftype_int,
5763 BFIN_BUILTIN_ROUND_1X32);
5764 def_builtin ("__builtin_bfin_mult_fr1x32", int_ftype_short_short,
5765 BFIN_BUILTIN_MULT_1X32);
5766 def_builtin ("__builtin_bfin_mult_fr1x32x32", int_ftype_int_int,
5767 BFIN_BUILTIN_MULT_1X32X32);
5768 def_builtin ("__builtin_bfin_mult_fr1x32x32NS", int_ftype_int_int,
5769 BFIN_BUILTIN_MULT_1X32X32NS);
5772 def_builtin ("__builtin_bfin_shl_fr1x16", short_ftype_int_int,
5773 BFIN_BUILTIN_SSASHIFT_1X16);
5774 def_builtin ("__builtin_bfin_shl_fr2x16", v2hi_ftype_v2hi_int,
5775 BFIN_BUILTIN_SSASHIFT_2X16);
5776 def_builtin ("__builtin_bfin_lshl_fr1x16", short_ftype_int_int,
5777 BFIN_BUILTIN_LSHIFT_1X16);
5778 def_builtin ("__builtin_bfin_lshl_fr2x16", v2hi_ftype_v2hi_int,
5779 BFIN_BUILTIN_LSHIFT_2X16);
5780 def_builtin ("__builtin_bfin_shl_fr1x32", int_ftype_int_int,
5781 BFIN_BUILTIN_SSASHIFT_1X32);
5783 /* Complex numbers. */
5784 def_builtin ("__builtin_bfin_cmplx_add", v2hi_ftype_v2hi_v2hi,
5785 BFIN_BUILTIN_SSADD_2X16);
5786 def_builtin ("__builtin_bfin_cmplx_sub", v2hi_ftype_v2hi_v2hi,
5787 BFIN_BUILTIN_SSSUB_2X16);
5788 def_builtin ("__builtin_bfin_cmplx_mul", v2hi_ftype_v2hi_v2hi,
5789 BFIN_BUILTIN_CPLX_MUL_16);
5790 def_builtin ("__builtin_bfin_cmplx_mac", v2hi_ftype_v2hi_v2hi_v2hi,
5791 BFIN_BUILTIN_CPLX_MAC_16);
5792 def_builtin ("__builtin_bfin_cmplx_msu", v2hi_ftype_v2hi_v2hi_v2hi,
5793 BFIN_BUILTIN_CPLX_MSU_16);
5794 def_builtin ("__builtin_bfin_cmplx_mul_s40", v2hi_ftype_v2hi_v2hi,
5795 BFIN_BUILTIN_CPLX_MUL_16_S40);
5796 def_builtin ("__builtin_bfin_cmplx_mac_s40", v2hi_ftype_v2hi_v2hi_v2hi,
5797 BFIN_BUILTIN_CPLX_MAC_16_S40);
5798 def_builtin ("__builtin_bfin_cmplx_msu_s40", v2hi_ftype_v2hi_v2hi_v2hi,
5799 BFIN_BUILTIN_CPLX_MSU_16_S40);
5800 def_builtin ("__builtin_bfin_csqu_fr16", v2hi_ftype_v2hi,
5801 BFIN_BUILTIN_CPLX_SQU);
5803 /* "Unaligned" load. */
5804 def_builtin ("__builtin_bfin_loadbytes", int_ftype_pint,
5805 BFIN_BUILTIN_LOADBYTES);
5810 struct builtin_description
5812 const enum insn_code icode;
5813 const char *const name;
5814 const enum bfin_builtins code;
5818 static const struct builtin_description bdesc_2arg[] =
5820 { CODE_FOR_composev2hi, "__builtin_bfin_compose_2x16", BFIN_BUILTIN_COMPOSE_2X16, -1 },
5822 { CODE_FOR_ssashiftv2hi3, "__builtin_bfin_shl_fr2x16", BFIN_BUILTIN_SSASHIFT_2X16, -1 },
5823 { CODE_FOR_ssashifthi3, "__builtin_bfin_shl_fr1x16", BFIN_BUILTIN_SSASHIFT_1X16, -1 },
5824 { CODE_FOR_lshiftv2hi3, "__builtin_bfin_lshl_fr2x16", BFIN_BUILTIN_LSHIFT_2X16, -1 },
5825 { CODE_FOR_lshifthi3, "__builtin_bfin_lshl_fr1x16", BFIN_BUILTIN_LSHIFT_1X16, -1 },
5826 { CODE_FOR_ssashiftsi3, "__builtin_bfin_shl_fr1x32", BFIN_BUILTIN_SSASHIFT_1X32, -1 },
5828 { CODE_FOR_sminhi3, "__builtin_bfin_min_fr1x16", BFIN_BUILTIN_MIN_1X16, -1 },
5829 { CODE_FOR_smaxhi3, "__builtin_bfin_max_fr1x16", BFIN_BUILTIN_MAX_1X16, -1 },
5830 { CODE_FOR_ssaddhi3, "__builtin_bfin_add_fr1x16", BFIN_BUILTIN_SSADD_1X16, -1 },
5831 { CODE_FOR_sssubhi3, "__builtin_bfin_sub_fr1x16", BFIN_BUILTIN_SSSUB_1X16, -1 },
5833 { CODE_FOR_sminsi3, "__builtin_bfin_min_fr1x32", BFIN_BUILTIN_MIN_1X32, -1 },
5834 { CODE_FOR_smaxsi3, "__builtin_bfin_max_fr1x32", BFIN_BUILTIN_MAX_1X32, -1 },
5835 { CODE_FOR_ssaddsi3, "__builtin_bfin_add_fr1x32", BFIN_BUILTIN_SSADD_1X32, -1 },
5836 { CODE_FOR_sssubsi3, "__builtin_bfin_sub_fr1x32", BFIN_BUILTIN_SSSUB_1X32, -1 },
5838 { CODE_FOR_sminv2hi3, "__builtin_bfin_min_fr2x16", BFIN_BUILTIN_MIN_2X16, -1 },
5839 { CODE_FOR_smaxv2hi3, "__builtin_bfin_max_fr2x16", BFIN_BUILTIN_MAX_2X16, -1 },
5840 { CODE_FOR_ssaddv2hi3, "__builtin_bfin_add_fr2x16", BFIN_BUILTIN_SSADD_2X16, -1 },
5841 { CODE_FOR_sssubv2hi3, "__builtin_bfin_sub_fr2x16", BFIN_BUILTIN_SSSUB_2X16, -1 },
5842 { CODE_FOR_ssaddsubv2hi3, "__builtin_bfin_dspaddsubsat", BFIN_BUILTIN_SSADDSUB_2X16, -1 },
5843 { CODE_FOR_sssubaddv2hi3, "__builtin_bfin_dspsubaddsat", BFIN_BUILTIN_SSSUBADD_2X16, -1 },
5845 { CODE_FOR_flag_mulhisi, "__builtin_bfin_mult_fr1x32", BFIN_BUILTIN_MULT_1X32, MACFLAG_NONE },
5846 { CODE_FOR_flag_mulhi, "__builtin_bfin_mult_fr1x16", BFIN_BUILTIN_MULT_1X16, MACFLAG_T },
5847 { CODE_FOR_flag_mulhi, "__builtin_bfin_multr_fr1x16", BFIN_BUILTIN_MULTR_1X16, MACFLAG_NONE },
5848 { CODE_FOR_flag_mulv2hi, "__builtin_bfin_mult_fr2x16", BFIN_BUILTIN_MULT_2X16, MACFLAG_T },
5849 { CODE_FOR_flag_mulv2hi, "__builtin_bfin_multr_fr2x16", BFIN_BUILTIN_MULTR_2X16, MACFLAG_NONE },
5851 { CODE_FOR_mulhisi_ll, "__builtin_bfin_mulhisill", BFIN_BUILTIN_MULHISILL, -1 },
5852 { CODE_FOR_mulhisi_lh, "__builtin_bfin_mulhisilh", BFIN_BUILTIN_MULHISILH, -1 },
5853 { CODE_FOR_mulhisi_hl, "__builtin_bfin_mulhisihl", BFIN_BUILTIN_MULHISIHL, -1 },
5854 { CODE_FOR_mulhisi_hh, "__builtin_bfin_mulhisihh", BFIN_BUILTIN_MULHISIHH, -1 }
5858 static const struct builtin_description bdesc_1arg[] =
5860 { CODE_FOR_loadbytes, "__builtin_bfin_loadbytes", BFIN_BUILTIN_LOADBYTES, 0 },
5862 { CODE_FOR_ones, "__builtin_bfin_ones", BFIN_BUILTIN_ONES, 0 },
5864 { CODE_FOR_signbitshi2, "__builtin_bfin_norm_fr1x16", BFIN_BUILTIN_NORM_1X16, 0 },
5865 { CODE_FOR_ssneghi2, "__builtin_bfin_negate_fr1x16", BFIN_BUILTIN_NEG_1X16, 0 },
5866 { CODE_FOR_abshi2, "__builtin_bfin_abs_fr1x16", BFIN_BUILTIN_ABS_1X16, 0 },
5868 { CODE_FOR_signbitssi2, "__builtin_bfin_norm_fr1x32", BFIN_BUILTIN_NORM_1X32, 0 },
5869 { CODE_FOR_ssroundsi2, "__builtin_bfin_round_fr1x32", BFIN_BUILTIN_ROUND_1X32, 0 },
5870 { CODE_FOR_ssnegsi2, "__builtin_bfin_negate_fr1x32", BFIN_BUILTIN_NEG_1X32, 0 },
5871 { CODE_FOR_ssabssi2, "__builtin_bfin_abs_fr1x32", BFIN_BUILTIN_ABS_1X32, 0 },
5873 { CODE_FOR_movv2hi_hi_low, "__builtin_bfin_extract_lo", BFIN_BUILTIN_EXTRACTLO, 0 },
5874 { CODE_FOR_movv2hi_hi_high, "__builtin_bfin_extract_hi", BFIN_BUILTIN_EXTRACTHI, 0 },
5875 { CODE_FOR_ssnegv2hi2, "__builtin_bfin_negate_fr2x16", BFIN_BUILTIN_NEG_2X16, 0 },
5876 { CODE_FOR_ssabsv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
5879 /* Errors in the source file can cause expand_expr to return const0_rtx
5880 where we expect a vector. To avoid crashing, use one of the vector
5881 clear instructions. */
5883 safe_vector_operand (rtx x, enum machine_mode mode)
5885 if (x != const0_rtx)
5887 x = gen_reg_rtx (SImode);
5889 emit_insn (gen_movsi (x, CONST0_RTX (SImode)));
5890 return gen_lowpart (mode, x);
5893 /* Subroutine of bfin_expand_builtin to take care of binop insns. MACFLAG is -1
5894 if this is a normal binary op, or one of the MACFLAG_xxx constants. */
5897 bfin_expand_binop_builtin (enum insn_code icode, tree exp, rtx target,
5901 tree arg0 = CALL_EXPR_ARG (exp, 0);
5902 tree arg1 = CALL_EXPR_ARG (exp, 1);
5903 rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5904 rtx op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
5905 enum machine_mode op0mode = GET_MODE (op0);
5906 enum machine_mode op1mode = GET_MODE (op1);
5907 enum machine_mode tmode = insn_data[icode].operand[0].mode;
5908 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
5909 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
5911 if (VECTOR_MODE_P (mode0))
5912 op0 = safe_vector_operand (op0, mode0);
5913 if (VECTOR_MODE_P (mode1))
5914 op1 = safe_vector_operand (op1, mode1);
5917 || GET_MODE (target) != tmode
5918 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
5919 target = gen_reg_rtx (tmode);
5921 if ((op0mode == SImode || op0mode == VOIDmode) && mode0 == HImode)
5924 op0 = gen_lowpart (HImode, op0);
5926 if ((op1mode == SImode || op1mode == VOIDmode) && mode1 == HImode)
5929 op1 = gen_lowpart (HImode, op1);
5931 /* In case the insn wants input operands in modes different from
5932 the result, abort. */
5933 gcc_assert ((op0mode == mode0 || op0mode == VOIDmode)
5934 && (op1mode == mode1 || op1mode == VOIDmode));
5936 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
5937 op0 = copy_to_mode_reg (mode0, op0);
5938 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
5939 op1 = copy_to_mode_reg (mode1, op1);
5942 pat = GEN_FCN (icode) (target, op0, op1);
5944 pat = GEN_FCN (icode) (target, op0, op1, GEN_INT (macflag));
5952 /* Subroutine of bfin_expand_builtin to take care of unop insns. */
5955 bfin_expand_unop_builtin (enum insn_code icode, tree exp,
5959 tree arg0 = CALL_EXPR_ARG (exp, 0);
5960 rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5961 enum machine_mode op0mode = GET_MODE (op0);
5962 enum machine_mode tmode = insn_data[icode].operand[0].mode;
5963 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
5966 || GET_MODE (target) != tmode
5967 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
5968 target = gen_reg_rtx (tmode);
5970 if (VECTOR_MODE_P (mode0))
5971 op0 = safe_vector_operand (op0, mode0);
5973 if (op0mode == SImode && mode0 == HImode)
5976 op0 = gen_lowpart (HImode, op0);
5978 gcc_assert (op0mode == mode0 || op0mode == VOIDmode);
5980 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
5981 op0 = copy_to_mode_reg (mode0, op0);
5983 pat = GEN_FCN (icode) (target, op0);
5990 /* Expand an expression EXP that calls a built-in function,
5991 with result going to TARGET if that's convenient
5992 (and in mode MODE if that's convenient).
5993 SUBTARGET may be used as the target for computing one of EXP's operands.
5994 IGNORE is nonzero if the value is to be ignored. */
5997 bfin_expand_builtin (tree exp, rtx target ATTRIBUTE_UNUSED,
5998 rtx subtarget ATTRIBUTE_UNUSED,
5999 enum machine_mode mode ATTRIBUTE_UNUSED,
6000 int ignore ATTRIBUTE_UNUSED)
6003 enum insn_code icode;
6004 const struct builtin_description *d;
6005 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
6006 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
6007 tree arg0, arg1, arg2;
6008 rtx op0, op1, op2, accvec, pat, tmp1, tmp2, a0reg, a1reg;
6009 enum machine_mode tmode, mode0;
6013 case BFIN_BUILTIN_CSYNC:
6014 emit_insn (gen_csync ());
6016 case BFIN_BUILTIN_SSYNC:
6017 emit_insn (gen_ssync ());
6020 case BFIN_BUILTIN_DIFFHL_2X16:
6021 case BFIN_BUILTIN_DIFFLH_2X16:
6022 case BFIN_BUILTIN_SUM_2X16:
6023 arg0 = CALL_EXPR_ARG (exp, 0);
6024 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
6025 icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16 ? CODE_FOR_subhilov2hi3
6026 : fcode == BFIN_BUILTIN_DIFFLH_2X16 ? CODE_FOR_sublohiv2hi3
6027 : CODE_FOR_ssaddhilov2hi3);
6028 tmode = insn_data[icode].operand[0].mode;
6029 mode0 = insn_data[icode].operand[1].mode;
6032 || GET_MODE (target) != tmode
6033 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6034 target = gen_reg_rtx (tmode);
6036 if (VECTOR_MODE_P (mode0))
6037 op0 = safe_vector_operand (op0, mode0);
6039 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
6040 op0 = copy_to_mode_reg (mode0, op0);
6042 pat = GEN_FCN (icode) (target, op0, op0);
6048 case BFIN_BUILTIN_MULT_1X32X32:
6049 case BFIN_BUILTIN_MULT_1X32X32NS:
6050 arg0 = CALL_EXPR_ARG (exp, 0);
6051 arg1 = CALL_EXPR_ARG (exp, 1);
6052 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
6053 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
6055 || !register_operand (target, SImode))
6056 target = gen_reg_rtx (SImode);
6058 a1reg = gen_rtx_REG (PDImode, REG_A1);
6059 a0reg = gen_rtx_REG (PDImode, REG_A0);
6060 tmp1 = gen_lowpart (V2HImode, op0);
6061 tmp2 = gen_lowpart (V2HImode, op1);
6062 emit_insn (gen_flag_macinit1hi (a1reg,
6063 gen_lowpart (HImode, op0),
6064 gen_lowpart (HImode, op1),
6065 GEN_INT (MACFLAG_FU)));
6066 emit_insn (gen_lshrpdi3 (a1reg, a1reg, GEN_INT (16)));
6068 if (fcode == BFIN_BUILTIN_MULT_1X32X32)
6069 emit_insn (gen_flag_mul_macv2hi_parts_acconly (a0reg, a1reg, tmp1, tmp2,
6070 const1_rtx, const1_rtx,
6071 const1_rtx, const0_rtx, a1reg,
6072 const0_rtx, GEN_INT (MACFLAG_NONE),
6073 GEN_INT (MACFLAG_M)));
6076 /* For saturating multiplication, there's exactly one special case
6077 to be handled: multiplying the smallest negative value with
6078 itself. Due to shift correction in fractional multiplies, this
6079 can overflow. Iff this happens, OP2 will contain 1, which, when
6080 added in 32 bits to the smallest negative, wraps to the largest
6081 positive, which is the result we want. */
6082 op2 = gen_reg_rtx (V2HImode);
6083 emit_insn (gen_packv2hi (op2, tmp1, tmp2, const0_rtx, const0_rtx));
6084 emit_insn (gen_movsibi (gen_rtx_REG (BImode, REG_CC),
6085 gen_lowpart (SImode, op2)));
6086 emit_insn (gen_flag_mul_macv2hi_parts_acconly_andcc0 (a0reg, a1reg, tmp1, tmp2,
6087 const1_rtx, const1_rtx,
6088 const1_rtx, const0_rtx, a1reg,
6089 const0_rtx, GEN_INT (MACFLAG_NONE),
6090 GEN_INT (MACFLAG_M)));
6091 op2 = gen_reg_rtx (SImode);
6092 emit_insn (gen_movbisi (op2, gen_rtx_REG (BImode, REG_CC)));
6094 emit_insn (gen_flag_machi_parts_acconly (a1reg, tmp2, tmp1,
6095 const1_rtx, const0_rtx,
6096 a1reg, const0_rtx, GEN_INT (MACFLAG_M)));
6097 emit_insn (gen_ashrpdi3 (a1reg, a1reg, GEN_INT (15)));
6098 emit_insn (gen_sum_of_accumulators (target, a0reg, a0reg, a1reg));
6099 if (fcode == BFIN_BUILTIN_MULT_1X32X32NS)
6100 emit_insn (gen_addsi3 (target, target, op2));
6103 case BFIN_BUILTIN_CPLX_MUL_16:
6104 case BFIN_BUILTIN_CPLX_MUL_16_S40:
6105 arg0 = CALL_EXPR_ARG (exp, 0);
6106 arg1 = CALL_EXPR_ARG (exp, 1);
6107 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
6108 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
6109 accvec = gen_reg_rtx (V2PDImode);
6112 || GET_MODE (target) != V2HImode
6113 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
6114 target = gen_reg_rtx (tmode);
6115 if (! register_operand (op0, GET_MODE (op0)))
6116 op0 = copy_to_mode_reg (GET_MODE (op0), op0);
6117 if (! register_operand (op1, GET_MODE (op1)))
6118 op1 = copy_to_mode_reg (GET_MODE (op1), op1);
6120 if (fcode == BFIN_BUILTIN_CPLX_MUL_16)
6121 emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
6122 const0_rtx, const0_rtx,
6123 const1_rtx, GEN_INT (MACFLAG_W32)));
6125 emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
6126 const0_rtx, const0_rtx,
6127 const1_rtx, GEN_INT (MACFLAG_NONE)));
6128 emit_insn (gen_flag_macv2hi_parts (target, op0, op1, const1_rtx,
6129 const1_rtx, const1_rtx,
6130 const0_rtx, accvec, const1_rtx, const0_rtx,
6131 GEN_INT (MACFLAG_NONE), accvec));
6135 case BFIN_BUILTIN_CPLX_MAC_16:
6136 case BFIN_BUILTIN_CPLX_MSU_16:
6137 case BFIN_BUILTIN_CPLX_MAC_16_S40:
6138 case BFIN_BUILTIN_CPLX_MSU_16_S40:
6139 arg0 = CALL_EXPR_ARG (exp, 0);
6140 arg1 = CALL_EXPR_ARG (exp, 1);
6141 arg2 = CALL_EXPR_ARG (exp, 2);
6142 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
6143 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
6144 op2 = expand_expr (arg2, NULL_RTX, VOIDmode, 0);
6145 accvec = gen_reg_rtx (V2PDImode);
6148 || GET_MODE (target) != V2HImode
6149 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
6150 target = gen_reg_rtx (tmode);
6151 if (! register_operand (op1, GET_MODE (op1)))
6152 op1 = copy_to_mode_reg (GET_MODE (op1), op1);
6153 if (! register_operand (op2, GET_MODE (op2)))
6154 op2 = copy_to_mode_reg (GET_MODE (op2), op2);
6156 tmp1 = gen_reg_rtx (SImode);
6157 tmp2 = gen_reg_rtx (SImode);
6158 emit_insn (gen_ashlsi3 (tmp1, gen_lowpart (SImode, op0), GEN_INT (16)));
6159 emit_move_insn (tmp2, gen_lowpart (SImode, op0));
6160 emit_insn (gen_movstricthi_1 (gen_lowpart (HImode, tmp2), const0_rtx));
6161 emit_insn (gen_load_accumulator_pair (accvec, tmp1, tmp2));
6162 if (fcode == BFIN_BUILTIN_CPLX_MAC_16
6163 || fcode == BFIN_BUILTIN_CPLX_MSU_16)
6164 emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
6165 const0_rtx, const0_rtx,
6166 const1_rtx, accvec, const0_rtx,
6168 GEN_INT (MACFLAG_W32)));
6170 emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
6171 const0_rtx, const0_rtx,
6172 const1_rtx, accvec, const0_rtx,
6174 GEN_INT (MACFLAG_NONE)));
6175 if (fcode == BFIN_BUILTIN_CPLX_MAC_16
6176 || fcode == BFIN_BUILTIN_CPLX_MAC_16_S40)
6186 emit_insn (gen_flag_macv2hi_parts (target, op1, op2, const1_rtx,
6187 const1_rtx, const1_rtx,
6188 const0_rtx, accvec, tmp1, tmp2,
6189 GEN_INT (MACFLAG_NONE), accvec));
6193 case BFIN_BUILTIN_CPLX_SQU:
6194 arg0 = CALL_EXPR_ARG (exp, 0);
6195 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
6196 accvec = gen_reg_rtx (V2PDImode);
6197 icode = CODE_FOR_flag_mulv2hi;
6198 tmp1 = gen_reg_rtx (V2HImode);
6199 tmp2 = gen_reg_rtx (V2HImode);
6202 || GET_MODE (target) != V2HImode
6203 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
6204 target = gen_reg_rtx (V2HImode);
6205 if (! register_operand (op0, GET_MODE (op0)))
6206 op0 = copy_to_mode_reg (GET_MODE (op0), op0);
6208 emit_insn (gen_flag_mulv2hi (tmp1, op0, op0, GEN_INT (MACFLAG_NONE)));
6210 emit_insn (gen_flag_mulhi_parts (gen_lowpart (HImode, tmp2), op0, op0,
6211 const0_rtx, const1_rtx,
6212 GEN_INT (MACFLAG_NONE)));
6214 emit_insn (gen_ssaddhi3_high_parts (target, tmp2, tmp2, tmp2, const0_rtx,
6216 emit_insn (gen_sssubhi3_low_parts (target, target, tmp1, tmp1,
6217 const0_rtx, const1_rtx));
6225 for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++)
6226 if (d->code == fcode)
6227 return bfin_expand_binop_builtin (d->icode, exp, target,
6230 for (i = 0, d = bdesc_1arg; i < ARRAY_SIZE (bdesc_1arg); i++, d++)
6231 if (d->code == fcode)
6232 return bfin_expand_unop_builtin (d->icode, exp, target);
6237 #undef TARGET_INIT_BUILTINS
6238 #define TARGET_INIT_BUILTINS bfin_init_builtins
6240 #undef TARGET_EXPAND_BUILTIN
6241 #define TARGET_EXPAND_BUILTIN bfin_expand_builtin
6243 #undef TARGET_ASM_GLOBALIZE_LABEL
6244 #define TARGET_ASM_GLOBALIZE_LABEL bfin_globalize_label
6246 #undef TARGET_ASM_FILE_START
6247 #define TARGET_ASM_FILE_START output_file_start
6249 #undef TARGET_ATTRIBUTE_TABLE
6250 #define TARGET_ATTRIBUTE_TABLE bfin_attribute_table
6252 #undef TARGET_COMP_TYPE_ATTRIBUTES
6253 #define TARGET_COMP_TYPE_ATTRIBUTES bfin_comp_type_attributes
6255 #undef TARGET_RTX_COSTS
6256 #define TARGET_RTX_COSTS bfin_rtx_costs
6258 #undef TARGET_ADDRESS_COST
6259 #define TARGET_ADDRESS_COST bfin_address_cost
6261 #undef TARGET_ASM_INTEGER
6262 #define TARGET_ASM_INTEGER bfin_assemble_integer
6264 #undef TARGET_MACHINE_DEPENDENT_REORG
6265 #define TARGET_MACHINE_DEPENDENT_REORG bfin_reorg
6267 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
6268 #define TARGET_FUNCTION_OK_FOR_SIBCALL bfin_function_ok_for_sibcall
6270 #undef TARGET_ASM_OUTPUT_MI_THUNK
6271 #define TARGET_ASM_OUTPUT_MI_THUNK bfin_output_mi_thunk
6272 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
6273 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
6275 #undef TARGET_SCHED_ADJUST_COST
6276 #define TARGET_SCHED_ADJUST_COST bfin_adjust_cost
6278 #undef TARGET_SCHED_ISSUE_RATE
6279 #define TARGET_SCHED_ISSUE_RATE bfin_issue_rate
6281 #undef TARGET_PROMOTE_PROTOTYPES
6282 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
6283 #undef TARGET_PROMOTE_FUNCTION_ARGS
6284 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
6285 #undef TARGET_PROMOTE_FUNCTION_RETURN
6286 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
6288 #undef TARGET_ARG_PARTIAL_BYTES
6289 #define TARGET_ARG_PARTIAL_BYTES bfin_arg_partial_bytes
6291 #undef TARGET_PASS_BY_REFERENCE
6292 #define TARGET_PASS_BY_REFERENCE bfin_pass_by_reference
6294 #undef TARGET_SETUP_INCOMING_VARARGS
6295 #define TARGET_SETUP_INCOMING_VARARGS setup_incoming_varargs
6297 #undef TARGET_STRUCT_VALUE_RTX
6298 #define TARGET_STRUCT_VALUE_RTX bfin_struct_value_rtx
6300 #undef TARGET_VECTOR_MODE_SUPPORTED_P
6301 #define TARGET_VECTOR_MODE_SUPPORTED_P bfin_vector_mode_supported_p
6303 #undef TARGET_HANDLE_OPTION
6304 #define TARGET_HANDLE_OPTION bfin_handle_option
6306 #undef TARGET_DEFAULT_TARGET_FLAGS
6307 #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT
6309 #undef TARGET_SECONDARY_RELOAD
6310 #define TARGET_SECONDARY_RELOAD bfin_secondary_reload
6312 #undef TARGET_DELEGITIMIZE_ADDRESS
6313 #define TARGET_DELEGITIMIZE_ADDRESS bfin_delegitimize_address
6315 #undef TARGET_CANNOT_FORCE_CONST_MEM
6316 #define TARGET_CANNOT_FORCE_CONST_MEM bfin_cannot_force_const_mem
6318 #undef TARGET_RETURN_IN_MEMORY
6319 #define TARGET_RETURN_IN_MEMORY bfin_return_in_memory
6321 struct gcc_target targetm = TARGET_INITIALIZER;