* alias.c, c-common.h, c-incpath.c, c-incpath.h, expr.c,
[official-gcc.git] / gcc / config / ia64 / ia64.c
blob76504c17be786a0158db9bcded23aaf24b9e547e
1 /* Definitions of target machine for GNU compiler.
2 Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005
3 Free Software Foundation, Inc.
4 Contributed by James E. Wilson <wilson@cygnus.com> and
5 David Mosberger <davidm@hpl.hp.com>.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 2, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING. If not, write to
21 the Free Software Foundation, 59 Temple Place - Suite 330,
22 Boston, MA 02111-1307, USA. */
24 #include "config.h"
25 #include "system.h"
26 #include "coretypes.h"
27 #include "tm.h"
28 #include "rtl.h"
29 #include "tree.h"
30 #include "regs.h"
31 #include "hard-reg-set.h"
32 #include "real.h"
33 #include "insn-config.h"
34 #include "conditions.h"
35 #include "output.h"
36 #include "insn-attr.h"
37 #include "flags.h"
38 #include "recog.h"
39 #include "expr.h"
40 #include "optabs.h"
41 #include "except.h"
42 #include "function.h"
43 #include "ggc.h"
44 #include "basic-block.h"
45 #include "toplev.h"
46 #include "sched-int.h"
47 #include "timevar.h"
48 #include "target.h"
49 #include "target-def.h"
50 #include "tm_p.h"
51 #include "hashtab.h"
52 #include "langhooks.h"
53 #include "cfglayout.h"
54 #include "tree-gimple.h"
56 /* This is used for communication between ASM_OUTPUT_LABEL and
57 ASM_OUTPUT_LABELREF. */
58 int ia64_asm_output_label = 0;
60 /* Define the information needed to generate branch and scc insns. This is
61 stored from the compare operation. */
62 struct rtx_def * ia64_compare_op0;
63 struct rtx_def * ia64_compare_op1;
65 /* Register names for ia64_expand_prologue. */
66 static const char * const ia64_reg_numbers[96] =
67 { "r32", "r33", "r34", "r35", "r36", "r37", "r38", "r39",
68 "r40", "r41", "r42", "r43", "r44", "r45", "r46", "r47",
69 "r48", "r49", "r50", "r51", "r52", "r53", "r54", "r55",
70 "r56", "r57", "r58", "r59", "r60", "r61", "r62", "r63",
71 "r64", "r65", "r66", "r67", "r68", "r69", "r70", "r71",
72 "r72", "r73", "r74", "r75", "r76", "r77", "r78", "r79",
73 "r80", "r81", "r82", "r83", "r84", "r85", "r86", "r87",
74 "r88", "r89", "r90", "r91", "r92", "r93", "r94", "r95",
75 "r96", "r97", "r98", "r99", "r100","r101","r102","r103",
76 "r104","r105","r106","r107","r108","r109","r110","r111",
77 "r112","r113","r114","r115","r116","r117","r118","r119",
78 "r120","r121","r122","r123","r124","r125","r126","r127"};
80 /* ??? These strings could be shared with REGISTER_NAMES. */
81 static const char * const ia64_input_reg_names[8] =
82 { "in0", "in1", "in2", "in3", "in4", "in5", "in6", "in7" };
84 /* ??? These strings could be shared with REGISTER_NAMES. */
85 static const char * const ia64_local_reg_names[80] =
86 { "loc0", "loc1", "loc2", "loc3", "loc4", "loc5", "loc6", "loc7",
87 "loc8", "loc9", "loc10","loc11","loc12","loc13","loc14","loc15",
88 "loc16","loc17","loc18","loc19","loc20","loc21","loc22","loc23",
89 "loc24","loc25","loc26","loc27","loc28","loc29","loc30","loc31",
90 "loc32","loc33","loc34","loc35","loc36","loc37","loc38","loc39",
91 "loc40","loc41","loc42","loc43","loc44","loc45","loc46","loc47",
92 "loc48","loc49","loc50","loc51","loc52","loc53","loc54","loc55",
93 "loc56","loc57","loc58","loc59","loc60","loc61","loc62","loc63",
94 "loc64","loc65","loc66","loc67","loc68","loc69","loc70","loc71",
95 "loc72","loc73","loc74","loc75","loc76","loc77","loc78","loc79" };
97 /* ??? These strings could be shared with REGISTER_NAMES. */
98 static const char * const ia64_output_reg_names[8] =
99 { "out0", "out1", "out2", "out3", "out4", "out5", "out6", "out7" };
101 /* String used with the -mfixed-range= option. */
102 const char *ia64_fixed_range_string;
104 /* Determines whether we use adds, addl, or movl to generate our
105 TLS immediate offsets. */
106 int ia64_tls_size = 22;
108 /* String used with the -mtls-size= option. */
109 const char *ia64_tls_size_string;
111 /* Which cpu are we scheduling for. */
112 enum processor_type ia64_tune;
114 /* String used with the -tune= option. */
115 const char *ia64_tune_string;
117 /* Determines whether we run our final scheduling pass or not. We always
118 avoid the normal second scheduling pass. */
119 static int ia64_flag_schedule_insns2;
121 /* Determines whether we run variable tracking in machine dependent
122 reorganization. */
123 static int ia64_flag_var_tracking;
125 /* Variables which are this size or smaller are put in the sdata/sbss
126 sections. */
128 unsigned int ia64_section_threshold;
130 /* The following variable is used by the DFA insn scheduler. The value is
131 TRUE if we do insn bundling instead of insn scheduling. */
132 int bundling_p = 0;
134 /* Structure to be filled in by ia64_compute_frame_size with register
135 save masks and offsets for the current function. */
137 struct ia64_frame_info
139 HOST_WIDE_INT total_size; /* size of the stack frame, not including
140 the caller's scratch area. */
141 HOST_WIDE_INT spill_cfa_off; /* top of the reg spill area from the cfa. */
142 HOST_WIDE_INT spill_size; /* size of the gr/br/fr spill area. */
143 HOST_WIDE_INT extra_spill_size; /* size of spill area for others. */
144 HARD_REG_SET mask; /* mask of saved registers. */
145 unsigned int gr_used_mask; /* mask of registers in use as gr spill
146 registers or long-term scratches. */
147 int n_spilled; /* number of spilled registers. */
148 int reg_fp; /* register for fp. */
149 int reg_save_b0; /* save register for b0. */
150 int reg_save_pr; /* save register for prs. */
151 int reg_save_ar_pfs; /* save register for ar.pfs. */
152 int reg_save_ar_unat; /* save register for ar.unat. */
153 int reg_save_ar_lc; /* save register for ar.lc. */
154 int reg_save_gp; /* save register for gp. */
155 int n_input_regs; /* number of input registers used. */
156 int n_local_regs; /* number of local registers used. */
157 int n_output_regs; /* number of output registers used. */
158 int n_rotate_regs; /* number of rotating registers used. */
160 char need_regstk; /* true if a .regstk directive needed. */
161 char initialized; /* true if the data is finalized. */
164 /* Current frame information calculated by ia64_compute_frame_size. */
165 static struct ia64_frame_info current_frame_info;
167 static int ia64_first_cycle_multipass_dfa_lookahead (void);
168 static void ia64_dependencies_evaluation_hook (rtx, rtx);
169 static void ia64_init_dfa_pre_cycle_insn (void);
170 static rtx ia64_dfa_pre_cycle_insn (void);
171 static int ia64_first_cycle_multipass_dfa_lookahead_guard (rtx);
172 static int ia64_dfa_new_cycle (FILE *, int, rtx, int, int, int *);
173 static rtx gen_tls_get_addr (void);
174 static rtx gen_thread_pointer (void);
175 static rtx ia64_expand_tls_address (enum tls_model, rtx, rtx);
176 static int find_gr_spill (int);
177 static int next_scratch_gr_reg (void);
178 static void mark_reg_gr_used_mask (rtx, void *);
179 static void ia64_compute_frame_size (HOST_WIDE_INT);
180 static void setup_spill_pointers (int, rtx, HOST_WIDE_INT);
181 static void finish_spill_pointers (void);
182 static rtx spill_restore_mem (rtx, HOST_WIDE_INT);
183 static void do_spill (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT, rtx);
184 static void do_restore (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT);
185 static rtx gen_movdi_x (rtx, rtx, rtx);
186 static rtx gen_fr_spill_x (rtx, rtx, rtx);
187 static rtx gen_fr_restore_x (rtx, rtx, rtx);
189 static enum machine_mode hfa_element_mode (tree, bool);
190 static void ia64_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode,
191 tree, int *, int);
192 static bool ia64_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
193 tree, bool);
194 static int ia64_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
195 tree, bool);
196 static bool ia64_function_ok_for_sibcall (tree, tree);
197 static bool ia64_return_in_memory (tree, tree);
198 static bool ia64_rtx_costs (rtx, int, int, int *);
199 static void fix_range (const char *);
200 static struct machine_function * ia64_init_machine_status (void);
201 static void emit_insn_group_barriers (FILE *);
202 static void emit_all_insn_group_barriers (FILE *);
203 static void final_emit_insn_group_barriers (FILE *);
204 static void emit_predicate_relation_info (void);
205 static void ia64_reorg (void);
206 static bool ia64_in_small_data_p (tree);
207 static void process_epilogue (void);
208 static int process_set (FILE *, rtx);
210 static rtx ia64_expand_fetch_and_op (optab, enum machine_mode, tree, rtx);
211 static rtx ia64_expand_op_and_fetch (optab, enum machine_mode, tree, rtx);
212 static rtx ia64_expand_compare_and_swap (enum machine_mode, enum machine_mode,
213 int, tree, rtx);
214 static rtx ia64_expand_lock_test_and_set (enum machine_mode, tree, rtx);
215 static rtx ia64_expand_lock_release (enum machine_mode, tree, rtx);
216 static bool ia64_assemble_integer (rtx, unsigned int, int);
217 static void ia64_output_function_prologue (FILE *, HOST_WIDE_INT);
218 static void ia64_output_function_epilogue (FILE *, HOST_WIDE_INT);
219 static void ia64_output_function_end_prologue (FILE *);
221 static int ia64_issue_rate (void);
222 static int ia64_adjust_cost (rtx, rtx, rtx, int);
223 static void ia64_sched_init (FILE *, int, int);
224 static void ia64_sched_finish (FILE *, int);
225 static int ia64_dfa_sched_reorder (FILE *, int, rtx *, int *, int, int);
226 static int ia64_sched_reorder (FILE *, int, rtx *, int *, int);
227 static int ia64_sched_reorder2 (FILE *, int, rtx *, int *, int);
228 static int ia64_variable_issue (FILE *, int, rtx, int);
230 static struct bundle_state *get_free_bundle_state (void);
231 static void free_bundle_state (struct bundle_state *);
232 static void initiate_bundle_states (void);
233 static void finish_bundle_states (void);
234 static unsigned bundle_state_hash (const void *);
235 static int bundle_state_eq_p (const void *, const void *);
236 static int insert_bundle_state (struct bundle_state *);
237 static void initiate_bundle_state_table (void);
238 static void finish_bundle_state_table (void);
239 static int try_issue_nops (struct bundle_state *, int);
240 static int try_issue_insn (struct bundle_state *, rtx);
241 static void issue_nops_and_insn (struct bundle_state *, int, rtx, int, int);
242 static int get_max_pos (state_t);
243 static int get_template (state_t, int);
245 static rtx get_next_important_insn (rtx, rtx);
246 static void bundling (FILE *, int, rtx, rtx);
248 static void ia64_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
249 HOST_WIDE_INT, tree);
250 static void ia64_file_start (void);
252 static void ia64_select_rtx_section (enum machine_mode, rtx,
253 unsigned HOST_WIDE_INT);
254 static void ia64_rwreloc_select_section (tree, int, unsigned HOST_WIDE_INT)
255 ATTRIBUTE_UNUSED;
256 static void ia64_rwreloc_unique_section (tree, int)
257 ATTRIBUTE_UNUSED;
258 static void ia64_rwreloc_select_rtx_section (enum machine_mode, rtx,
259 unsigned HOST_WIDE_INT)
260 ATTRIBUTE_UNUSED;
261 static unsigned int ia64_section_type_flags (tree, const char *, int);
262 static void ia64_hpux_add_extern_decl (tree decl)
263 ATTRIBUTE_UNUSED;
264 static void ia64_hpux_file_end (void)
265 ATTRIBUTE_UNUSED;
266 static void ia64_init_libfuncs (void)
267 ATTRIBUTE_UNUSED;
268 static void ia64_hpux_init_libfuncs (void)
269 ATTRIBUTE_UNUSED;
270 static void ia64_sysv4_init_libfuncs (void)
271 ATTRIBUTE_UNUSED;
272 static void ia64_vms_init_libfuncs (void)
273 ATTRIBUTE_UNUSED;
275 static tree ia64_handle_model_attribute (tree *, tree, tree, int, bool *);
276 static void ia64_encode_section_info (tree, rtx, int);
277 static rtx ia64_struct_value_rtx (tree, int);
278 static tree ia64_gimplify_va_arg (tree, tree, tree *, tree *);
279 static bool ia64_scalar_mode_supported_p (enum machine_mode mode);
280 static bool ia64_vector_mode_supported_p (enum machine_mode mode);
283 /* Table of valid machine attributes. */
284 static const struct attribute_spec ia64_attribute_table[] =
286 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
287 { "syscall_linkage", 0, 0, false, true, true, NULL },
288 { "model", 1, 1, true, false, false, ia64_handle_model_attribute },
289 { NULL, 0, 0, false, false, false, NULL }
292 /* Initialize the GCC target structure. */
293 #undef TARGET_ATTRIBUTE_TABLE
294 #define TARGET_ATTRIBUTE_TABLE ia64_attribute_table
296 #undef TARGET_INIT_BUILTINS
297 #define TARGET_INIT_BUILTINS ia64_init_builtins
299 #undef TARGET_EXPAND_BUILTIN
300 #define TARGET_EXPAND_BUILTIN ia64_expand_builtin
302 #undef TARGET_ASM_BYTE_OP
303 #define TARGET_ASM_BYTE_OP "\tdata1\t"
304 #undef TARGET_ASM_ALIGNED_HI_OP
305 #define TARGET_ASM_ALIGNED_HI_OP "\tdata2\t"
306 #undef TARGET_ASM_ALIGNED_SI_OP
307 #define TARGET_ASM_ALIGNED_SI_OP "\tdata4\t"
308 #undef TARGET_ASM_ALIGNED_DI_OP
309 #define TARGET_ASM_ALIGNED_DI_OP "\tdata8\t"
310 #undef TARGET_ASM_UNALIGNED_HI_OP
311 #define TARGET_ASM_UNALIGNED_HI_OP "\tdata2.ua\t"
312 #undef TARGET_ASM_UNALIGNED_SI_OP
313 #define TARGET_ASM_UNALIGNED_SI_OP "\tdata4.ua\t"
314 #undef TARGET_ASM_UNALIGNED_DI_OP
315 #define TARGET_ASM_UNALIGNED_DI_OP "\tdata8.ua\t"
316 #undef TARGET_ASM_INTEGER
317 #define TARGET_ASM_INTEGER ia64_assemble_integer
319 #undef TARGET_ASM_FUNCTION_PROLOGUE
320 #define TARGET_ASM_FUNCTION_PROLOGUE ia64_output_function_prologue
321 #undef TARGET_ASM_FUNCTION_END_PROLOGUE
322 #define TARGET_ASM_FUNCTION_END_PROLOGUE ia64_output_function_end_prologue
323 #undef TARGET_ASM_FUNCTION_EPILOGUE
324 #define TARGET_ASM_FUNCTION_EPILOGUE ia64_output_function_epilogue
326 #undef TARGET_IN_SMALL_DATA_P
327 #define TARGET_IN_SMALL_DATA_P ia64_in_small_data_p
329 #undef TARGET_SCHED_ADJUST_COST
330 #define TARGET_SCHED_ADJUST_COST ia64_adjust_cost
331 #undef TARGET_SCHED_ISSUE_RATE
332 #define TARGET_SCHED_ISSUE_RATE ia64_issue_rate
333 #undef TARGET_SCHED_VARIABLE_ISSUE
334 #define TARGET_SCHED_VARIABLE_ISSUE ia64_variable_issue
335 #undef TARGET_SCHED_INIT
336 #define TARGET_SCHED_INIT ia64_sched_init
337 #undef TARGET_SCHED_FINISH
338 #define TARGET_SCHED_FINISH ia64_sched_finish
339 #undef TARGET_SCHED_REORDER
340 #define TARGET_SCHED_REORDER ia64_sched_reorder
341 #undef TARGET_SCHED_REORDER2
342 #define TARGET_SCHED_REORDER2 ia64_sched_reorder2
344 #undef TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK
345 #define TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK ia64_dependencies_evaluation_hook
347 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
348 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD ia64_first_cycle_multipass_dfa_lookahead
350 #undef TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN
351 #define TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN ia64_init_dfa_pre_cycle_insn
352 #undef TARGET_SCHED_DFA_PRE_CYCLE_INSN
353 #define TARGET_SCHED_DFA_PRE_CYCLE_INSN ia64_dfa_pre_cycle_insn
355 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD
356 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD\
357 ia64_first_cycle_multipass_dfa_lookahead_guard
359 #undef TARGET_SCHED_DFA_NEW_CYCLE
360 #define TARGET_SCHED_DFA_NEW_CYCLE ia64_dfa_new_cycle
362 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
363 #define TARGET_FUNCTION_OK_FOR_SIBCALL ia64_function_ok_for_sibcall
364 #undef TARGET_PASS_BY_REFERENCE
365 #define TARGET_PASS_BY_REFERENCE ia64_pass_by_reference
366 #undef TARGET_ARG_PARTIAL_BYTES
367 #define TARGET_ARG_PARTIAL_BYTES ia64_arg_partial_bytes
369 #undef TARGET_ASM_OUTPUT_MI_THUNK
370 #define TARGET_ASM_OUTPUT_MI_THUNK ia64_output_mi_thunk
371 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
372 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
374 #undef TARGET_ASM_FILE_START
375 #define TARGET_ASM_FILE_START ia64_file_start
377 #undef TARGET_RTX_COSTS
378 #define TARGET_RTX_COSTS ia64_rtx_costs
379 #undef TARGET_ADDRESS_COST
380 #define TARGET_ADDRESS_COST hook_int_rtx_0
382 #undef TARGET_MACHINE_DEPENDENT_REORG
383 #define TARGET_MACHINE_DEPENDENT_REORG ia64_reorg
385 #undef TARGET_ENCODE_SECTION_INFO
386 #define TARGET_ENCODE_SECTION_INFO ia64_encode_section_info
388 #undef TARGET_SECTION_TYPE_FLAGS
389 #define TARGET_SECTION_TYPE_FLAGS ia64_section_type_flags
391 /* ??? ABI doesn't allow us to define this. */
392 #if 0
393 #undef TARGET_PROMOTE_FUNCTION_ARGS
394 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
395 #endif
397 /* ??? ABI doesn't allow us to define this. */
398 #if 0
399 #undef TARGET_PROMOTE_FUNCTION_RETURN
400 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
401 #endif
403 /* ??? Investigate. */
404 #if 0
405 #undef TARGET_PROMOTE_PROTOTYPES
406 #define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
407 #endif
409 #undef TARGET_STRUCT_VALUE_RTX
410 #define TARGET_STRUCT_VALUE_RTX ia64_struct_value_rtx
411 #undef TARGET_RETURN_IN_MEMORY
412 #define TARGET_RETURN_IN_MEMORY ia64_return_in_memory
413 #undef TARGET_SETUP_INCOMING_VARARGS
414 #define TARGET_SETUP_INCOMING_VARARGS ia64_setup_incoming_varargs
415 #undef TARGET_STRICT_ARGUMENT_NAMING
416 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
417 #undef TARGET_MUST_PASS_IN_STACK
418 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
420 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
421 #define TARGET_GIMPLIFY_VA_ARG_EXPR ia64_gimplify_va_arg
423 #undef TARGET_UNWIND_EMIT
424 #define TARGET_UNWIND_EMIT process_for_unwind_directive
426 #undef TARGET_SCALAR_MODE_SUPPORTED_P
427 #define TARGET_SCALAR_MODE_SUPPORTED_P ia64_scalar_mode_supported_p
428 #undef TARGET_VECTOR_MODE_SUPPORTED_P
429 #define TARGET_VECTOR_MODE_SUPPORTED_P ia64_vector_mode_supported_p
431 /* ia64 architecture manual 4.4.7: ... reads, writes, and flushes may occur
432 in an order different from the specified program order. */
433 #undef TARGET_RELAXED_ORDERING
434 #define TARGET_RELAXED_ORDERING true
436 struct gcc_target targetm = TARGET_INITIALIZER;
438 typedef enum
440 ADDR_AREA_NORMAL, /* normal address area */
441 ADDR_AREA_SMALL /* addressable by "addl" (-2MB < addr < 2MB) */
443 ia64_addr_area;
445 static GTY(()) tree small_ident1;
446 static GTY(()) tree small_ident2;
448 static void
449 init_idents (void)
451 if (small_ident1 == 0)
453 small_ident1 = get_identifier ("small");
454 small_ident2 = get_identifier ("__small__");
458 /* Retrieve the address area that has been chosen for the given decl. */
460 static ia64_addr_area
461 ia64_get_addr_area (tree decl)
463 tree model_attr;
465 model_attr = lookup_attribute ("model", DECL_ATTRIBUTES (decl));
466 if (model_attr)
468 tree id;
470 init_idents ();
471 id = TREE_VALUE (TREE_VALUE (model_attr));
472 if (id == small_ident1 || id == small_ident2)
473 return ADDR_AREA_SMALL;
475 return ADDR_AREA_NORMAL;
478 static tree
479 ia64_handle_model_attribute (tree *node, tree name, tree args,
480 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
482 ia64_addr_area addr_area = ADDR_AREA_NORMAL;
483 ia64_addr_area area;
484 tree arg, decl = *node;
486 init_idents ();
487 arg = TREE_VALUE (args);
488 if (arg == small_ident1 || arg == small_ident2)
490 addr_area = ADDR_AREA_SMALL;
492 else
494 warning ("invalid argument of %qs attribute",
495 IDENTIFIER_POINTER (name));
496 *no_add_attrs = true;
499 switch (TREE_CODE (decl))
501 case VAR_DECL:
502 if ((DECL_CONTEXT (decl) && TREE_CODE (DECL_CONTEXT (decl))
503 == FUNCTION_DECL)
504 && !TREE_STATIC (decl))
506 error ("%Jan address area attribute cannot be specified for "
507 "local variables", decl, decl);
508 *no_add_attrs = true;
510 area = ia64_get_addr_area (decl);
511 if (area != ADDR_AREA_NORMAL && addr_area != area)
513 error ("%Jaddress area of '%s' conflicts with previous "
514 "declaration", decl, decl);
515 *no_add_attrs = true;
517 break;
519 case FUNCTION_DECL:
520 error ("%Jaddress area attribute cannot be specified for functions",
521 decl, decl);
522 *no_add_attrs = true;
523 break;
525 default:
526 warning ("%qs attribute ignored", IDENTIFIER_POINTER (name));
527 *no_add_attrs = true;
528 break;
531 return NULL_TREE;
534 static void
535 ia64_encode_addr_area (tree decl, rtx symbol)
537 int flags;
539 flags = SYMBOL_REF_FLAGS (symbol);
540 switch (ia64_get_addr_area (decl))
542 case ADDR_AREA_NORMAL: break;
543 case ADDR_AREA_SMALL: flags |= SYMBOL_FLAG_SMALL_ADDR; break;
544 default: abort ();
546 SYMBOL_REF_FLAGS (symbol) = flags;
549 static void
550 ia64_encode_section_info (tree decl, rtx rtl, int first)
552 default_encode_section_info (decl, rtl, first);
554 /* Careful not to prod global register variables. */
555 if (TREE_CODE (decl) == VAR_DECL
556 && GET_CODE (DECL_RTL (decl)) == MEM
557 && GET_CODE (XEXP (DECL_RTL (decl), 0)) == SYMBOL_REF
558 && (TREE_STATIC (decl) || DECL_EXTERNAL (decl)))
559 ia64_encode_addr_area (decl, XEXP (rtl, 0));
562 /* Implement CONST_OK_FOR_LETTER_P. */
564 bool
565 ia64_const_ok_for_letter_p (HOST_WIDE_INT value, char c)
567 switch (c)
569 case 'I':
570 return CONST_OK_FOR_I (value);
571 case 'J':
572 return CONST_OK_FOR_J (value);
573 case 'K':
574 return CONST_OK_FOR_K (value);
575 case 'L':
576 return CONST_OK_FOR_L (value);
577 case 'M':
578 return CONST_OK_FOR_M (value);
579 case 'N':
580 return CONST_OK_FOR_N (value);
581 case 'O':
582 return CONST_OK_FOR_O (value);
583 case 'P':
584 return CONST_OK_FOR_P (value);
585 default:
586 return false;
590 /* Implement CONST_DOUBLE_OK_FOR_LETTER_P. */
592 bool
593 ia64_const_double_ok_for_letter_p (rtx value, char c)
595 switch (c)
597 case 'G':
598 return CONST_DOUBLE_OK_FOR_G (value);
599 default:
600 return false;
604 /* Implement EXTRA_CONSTRAINT. */
606 bool
607 ia64_extra_constraint (rtx value, char c)
609 switch (c)
611 case 'Q':
612 /* Non-volatile memory for FP_REG loads/stores. */
613 return memory_operand(value, VOIDmode) && !MEM_VOLATILE_P (value);
615 case 'R':
616 /* 1..4 for shladd arguments. */
617 return (GET_CODE (value) == CONST_INT
618 && INTVAL (value) >= 1 && INTVAL (value) <= 4);
620 case 'S':
621 /* Non-post-inc memory for asms and other unsavory creatures. */
622 return (GET_CODE (value) == MEM
623 && GET_RTX_CLASS (GET_CODE (XEXP (value, 0))) != RTX_AUTOINC
624 && (reload_in_progress || memory_operand (value, VOIDmode)));
626 case 'T':
627 /* Symbol ref to small-address-area. */
628 return (GET_CODE (value) == SYMBOL_REF
629 && SYMBOL_REF_SMALL_ADDR_P (value));
631 case 'U':
632 /* Vector zero. */
633 return value == CONST0_RTX (GET_MODE (value));
635 case 'W':
636 /* An integer vector, such that conversion to an integer yields a
637 value appropriate for an integer 'J' constraint. */
638 if (GET_CODE (value) == CONST_VECTOR
639 && GET_MODE_CLASS (GET_MODE (value)) == MODE_VECTOR_INT)
641 value = simplify_subreg (DImode, value, GET_MODE (value), 0);
642 return ia64_const_ok_for_letter_p (INTVAL (value), 'J');
644 return false;
646 case 'Y':
647 /* A V2SF vector containing elements that satisfy 'G'. */
648 return
649 (GET_CODE (value) == CONST_VECTOR
650 && GET_MODE (value) == V2SFmode
651 && ia64_const_double_ok_for_letter_p (XVECEXP (value, 0, 0), 'G')
652 && ia64_const_double_ok_for_letter_p (XVECEXP (value, 0, 1), 'G'));
654 default:
655 return false;
659 /* Return 1 if the operands of a move are ok. */
662 ia64_move_ok (rtx dst, rtx src)
664 /* If we're under init_recog_no_volatile, we'll not be able to use
665 memory_operand. So check the code directly and don't worry about
666 the validity of the underlying address, which should have been
667 checked elsewhere anyway. */
668 if (GET_CODE (dst) != MEM)
669 return 1;
670 if (GET_CODE (src) == MEM)
671 return 0;
672 if (register_operand (src, VOIDmode))
673 return 1;
675 /* Otherwise, this must be a constant, and that either 0 or 0.0 or 1.0. */
676 if (INTEGRAL_MODE_P (GET_MODE (dst)))
677 return src == const0_rtx;
678 else
679 return GET_CODE (src) == CONST_DOUBLE && CONST_DOUBLE_OK_FOR_G (src);
683 addp4_optimize_ok (rtx op1, rtx op2)
685 return (basereg_operand (op1, GET_MODE(op1)) !=
686 basereg_operand (op2, GET_MODE(op2)));
689 /* Check if OP is a mask suitable for use with SHIFT in a dep.z instruction.
690 Return the length of the field, or <= 0 on failure. */
693 ia64_depz_field_mask (rtx rop, rtx rshift)
695 unsigned HOST_WIDE_INT op = INTVAL (rop);
696 unsigned HOST_WIDE_INT shift = INTVAL (rshift);
698 /* Get rid of the zero bits we're shifting in. */
699 op >>= shift;
701 /* We must now have a solid block of 1's at bit 0. */
702 return exact_log2 (op + 1);
705 /* Expand a symbolic constant load. */
707 void
708 ia64_expand_load_address (rtx dest, rtx src)
710 if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (src))
711 abort ();
712 if (GET_CODE (dest) != REG)
713 abort ();
715 /* ILP32 mode still loads 64-bits of data from the GOT. This avoids
716 having to pointer-extend the value afterward. Other forms of address
717 computation below are also more natural to compute as 64-bit quantities.
718 If we've been given an SImode destination register, change it. */
719 if (GET_MODE (dest) != Pmode)
720 dest = gen_rtx_REG (Pmode, REGNO (dest));
722 if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_SMALL_ADDR_P (src))
724 emit_insn (gen_rtx_SET (VOIDmode, dest, src));
725 return;
727 else if (TARGET_AUTO_PIC)
729 emit_insn (gen_load_gprel64 (dest, src));
730 return;
732 else if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_FUNCTION_P (src))
734 emit_insn (gen_load_fptr (dest, src));
735 return;
737 else if (sdata_symbolic_operand (src, VOIDmode))
739 emit_insn (gen_load_gprel (dest, src));
740 return;
743 if (GET_CODE (src) == CONST
744 && GET_CODE (XEXP (src, 0)) == PLUS
745 && GET_CODE (XEXP (XEXP (src, 0), 1)) == CONST_INT
746 && (INTVAL (XEXP (XEXP (src, 0), 1)) & 0x3fff) != 0)
748 rtx sym = XEXP (XEXP (src, 0), 0);
749 HOST_WIDE_INT ofs, hi, lo;
751 /* Split the offset into a sign extended 14-bit low part
752 and a complementary high part. */
753 ofs = INTVAL (XEXP (XEXP (src, 0), 1));
754 lo = ((ofs & 0x3fff) ^ 0x2000) - 0x2000;
755 hi = ofs - lo;
757 ia64_expand_load_address (dest, plus_constant (sym, hi));
758 emit_insn (gen_adddi3 (dest, dest, GEN_INT (lo)));
760 else
762 rtx tmp;
764 tmp = gen_rtx_HIGH (Pmode, src);
765 tmp = gen_rtx_PLUS (Pmode, tmp, pic_offset_table_rtx);
766 emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
768 tmp = gen_rtx_LO_SUM (GET_MODE (dest), dest, src);
769 emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
773 static GTY(()) rtx gen_tls_tga;
774 static rtx
775 gen_tls_get_addr (void)
777 if (!gen_tls_tga)
778 gen_tls_tga = init_one_libfunc ("__tls_get_addr");
779 return gen_tls_tga;
782 static GTY(()) rtx thread_pointer_rtx;
783 static rtx
784 gen_thread_pointer (void)
786 if (!thread_pointer_rtx)
787 thread_pointer_rtx = gen_rtx_REG (Pmode, 13);
788 return thread_pointer_rtx;
791 static rtx
792 ia64_expand_tls_address (enum tls_model tls_kind, rtx op0, rtx op1)
794 rtx tga_op1, tga_op2, tga_ret, tga_eqv, tmp, insns;
795 rtx orig_op0 = op0;
797 switch (tls_kind)
799 case TLS_MODEL_GLOBAL_DYNAMIC:
800 start_sequence ();
802 tga_op1 = gen_reg_rtx (Pmode);
803 emit_insn (gen_load_ltoff_dtpmod (tga_op1, op1));
804 tga_op1 = gen_const_mem (Pmode, tga_op1);
806 tga_op2 = gen_reg_rtx (Pmode);
807 emit_insn (gen_load_ltoff_dtprel (tga_op2, op1));
808 tga_op2 = gen_const_mem (Pmode, tga_op2);
810 tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
811 LCT_CONST, Pmode, 2, tga_op1,
812 Pmode, tga_op2, Pmode);
814 insns = get_insns ();
815 end_sequence ();
817 if (GET_MODE (op0) != Pmode)
818 op0 = tga_ret;
819 emit_libcall_block (insns, op0, tga_ret, op1);
820 break;
822 case TLS_MODEL_LOCAL_DYNAMIC:
823 /* ??? This isn't the completely proper way to do local-dynamic
824 If the call to __tls_get_addr is used only by a single symbol,
825 then we should (somehow) move the dtprel to the second arg
826 to avoid the extra add. */
827 start_sequence ();
829 tga_op1 = gen_reg_rtx (Pmode);
830 emit_insn (gen_load_ltoff_dtpmod (tga_op1, op1));
831 tga_op1 = gen_const_mem (Pmode, tga_op1);
833 tga_op2 = const0_rtx;
835 tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
836 LCT_CONST, Pmode, 2, tga_op1,
837 Pmode, tga_op2, Pmode);
839 insns = get_insns ();
840 end_sequence ();
842 tga_eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
843 UNSPEC_LD_BASE);
844 tmp = gen_reg_rtx (Pmode);
845 emit_libcall_block (insns, tmp, tga_ret, tga_eqv);
847 if (!register_operand (op0, Pmode))
848 op0 = gen_reg_rtx (Pmode);
849 if (TARGET_TLS64)
851 emit_insn (gen_load_dtprel (op0, op1));
852 emit_insn (gen_adddi3 (op0, tmp, op0));
854 else
855 emit_insn (gen_add_dtprel (op0, tmp, op1));
856 break;
858 case TLS_MODEL_INITIAL_EXEC:
859 tmp = gen_reg_rtx (Pmode);
860 emit_insn (gen_load_ltoff_tprel (tmp, op1));
861 tmp = gen_const_mem (Pmode, tmp);
862 tmp = force_reg (Pmode, tmp);
864 if (!register_operand (op0, Pmode))
865 op0 = gen_reg_rtx (Pmode);
866 emit_insn (gen_adddi3 (op0, tmp, gen_thread_pointer ()));
867 break;
869 case TLS_MODEL_LOCAL_EXEC:
870 if (!register_operand (op0, Pmode))
871 op0 = gen_reg_rtx (Pmode);
872 if (TARGET_TLS64)
874 emit_insn (gen_load_tprel (op0, op1));
875 emit_insn (gen_adddi3 (op0, gen_thread_pointer (), op0));
877 else
878 emit_insn (gen_add_tprel (op0, gen_thread_pointer (), op1));
879 break;
881 default:
882 abort ();
885 if (orig_op0 == op0)
886 return NULL_RTX;
887 if (GET_MODE (orig_op0) == Pmode)
888 return op0;
889 return gen_lowpart (GET_MODE (orig_op0), op0);
893 ia64_expand_move (rtx op0, rtx op1)
895 enum machine_mode mode = GET_MODE (op0);
897 if (!reload_in_progress && !reload_completed && !ia64_move_ok (op0, op1))
898 op1 = force_reg (mode, op1);
900 if ((mode == Pmode || mode == ptr_mode) && symbolic_operand (op1, VOIDmode))
902 enum tls_model tls_kind;
903 if (GET_CODE (op1) == SYMBOL_REF
904 && (tls_kind = SYMBOL_REF_TLS_MODEL (op1)))
905 return ia64_expand_tls_address (tls_kind, op0, op1);
907 if (!TARGET_NO_PIC && reload_completed)
909 ia64_expand_load_address (op0, op1);
910 return NULL_RTX;
914 return op1;
917 /* Split a move from OP1 to OP0 conditional on COND. */
919 void
920 ia64_emit_cond_move (rtx op0, rtx op1, rtx cond)
922 rtx insn, first = get_last_insn ();
924 emit_move_insn (op0, op1);
926 for (insn = get_last_insn (); insn != first; insn = PREV_INSN (insn))
927 if (INSN_P (insn))
928 PATTERN (insn) = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (cond),
929 PATTERN (insn));
932 /* Split a post-reload TImode or TFmode reference into two DImode
933 components. This is made extra difficult by the fact that we do
934 not get any scratch registers to work with, because reload cannot
935 be prevented from giving us a scratch that overlaps the register
936 pair involved. So instead, when addressing memory, we tweak the
937 pointer register up and back down with POST_INCs. Or up and not
938 back down when we can get away with it.
940 REVERSED is true when the loads must be done in reversed order
941 (high word first) for correctness. DEAD is true when the pointer
942 dies with the second insn we generate and therefore the second
943 address must not carry a postmodify.
945 May return an insn which is to be emitted after the moves. */
947 static rtx
948 ia64_split_tmode (rtx out[2], rtx in, bool reversed, bool dead)
950 rtx fixup = 0;
952 switch (GET_CODE (in))
954 case REG:
955 out[reversed] = gen_rtx_REG (DImode, REGNO (in));
956 out[!reversed] = gen_rtx_REG (DImode, REGNO (in) + 1);
957 break;
959 case CONST_INT:
960 case CONST_DOUBLE:
961 /* Cannot occur reversed. */
962 if (reversed) abort ();
964 if (GET_MODE (in) != TFmode)
965 split_double (in, &out[0], &out[1]);
966 else
967 /* split_double does not understand how to split a TFmode
968 quantity into a pair of DImode constants. */
970 REAL_VALUE_TYPE r;
971 unsigned HOST_WIDE_INT p[2];
972 long l[4]; /* TFmode is 128 bits */
974 REAL_VALUE_FROM_CONST_DOUBLE (r, in);
975 real_to_target (l, &r, TFmode);
977 if (FLOAT_WORDS_BIG_ENDIAN)
979 p[0] = (((unsigned HOST_WIDE_INT) l[0]) << 32) + l[1];
980 p[1] = (((unsigned HOST_WIDE_INT) l[2]) << 32) + l[3];
982 else
984 p[0] = (((unsigned HOST_WIDE_INT) l[3]) << 32) + l[2];
985 p[1] = (((unsigned HOST_WIDE_INT) l[1]) << 32) + l[0];
987 out[0] = GEN_INT (p[0]);
988 out[1] = GEN_INT (p[1]);
990 break;
992 case MEM:
994 rtx base = XEXP (in, 0);
995 rtx offset;
997 switch (GET_CODE (base))
999 case REG:
1000 if (!reversed)
1002 out[0] = adjust_automodify_address
1003 (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
1004 out[1] = adjust_automodify_address
1005 (in, DImode, dead ? 0 : gen_rtx_POST_DEC (Pmode, base), 8);
1007 else
1009 /* Reversal requires a pre-increment, which can only
1010 be done as a separate insn. */
1011 emit_insn (gen_adddi3 (base, base, GEN_INT (8)));
1012 out[0] = adjust_automodify_address
1013 (in, DImode, gen_rtx_POST_DEC (Pmode, base), 8);
1014 out[1] = adjust_address (in, DImode, 0);
1016 break;
1018 case POST_INC:
1019 if (reversed || dead) abort ();
1020 /* Just do the increment in two steps. */
1021 out[0] = adjust_automodify_address (in, DImode, 0, 0);
1022 out[1] = adjust_automodify_address (in, DImode, 0, 8);
1023 break;
1025 case POST_DEC:
1026 if (reversed || dead) abort ();
1027 /* Add 8, subtract 24. */
1028 base = XEXP (base, 0);
1029 out[0] = adjust_automodify_address
1030 (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
1031 out[1] = adjust_automodify_address
1032 (in, DImode,
1033 gen_rtx_POST_MODIFY (Pmode, base, plus_constant (base, -24)),
1035 break;
1037 case POST_MODIFY:
1038 if (reversed || dead) abort ();
1039 /* Extract and adjust the modification. This case is
1040 trickier than the others, because we might have an
1041 index register, or we might have a combined offset that
1042 doesn't fit a signed 9-bit displacement field. We can
1043 assume the incoming expression is already legitimate. */
1044 offset = XEXP (base, 1);
1045 base = XEXP (base, 0);
1047 out[0] = adjust_automodify_address
1048 (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
1050 if (GET_CODE (XEXP (offset, 1)) == REG)
1052 /* Can't adjust the postmodify to match. Emit the
1053 original, then a separate addition insn. */
1054 out[1] = adjust_automodify_address (in, DImode, 0, 8);
1055 fixup = gen_adddi3 (base, base, GEN_INT (-8));
1057 else if (GET_CODE (XEXP (offset, 1)) != CONST_INT)
1058 abort ();
1059 else if (INTVAL (XEXP (offset, 1)) < -256 + 8)
1061 /* Again the postmodify cannot be made to match, but
1062 in this case it's more efficient to get rid of the
1063 postmodify entirely and fix up with an add insn. */
1064 out[1] = adjust_automodify_address (in, DImode, base, 8);
1065 fixup = gen_adddi3 (base, base,
1066 GEN_INT (INTVAL (XEXP (offset, 1)) - 8));
1068 else
1070 /* Combined offset still fits in the displacement field.
1071 (We cannot overflow it at the high end.) */
1072 out[1] = adjust_automodify_address
1073 (in, DImode,
1074 gen_rtx_POST_MODIFY (Pmode, base,
1075 gen_rtx_PLUS (Pmode, base,
1076 GEN_INT (INTVAL (XEXP (offset, 1)) - 8))),
1079 break;
1081 default:
1082 abort ();
1084 break;
1087 default:
1088 abort ();
1091 return fixup;
1094 /* Split a TImode or TFmode move instruction after reload.
1095 This is used by *movtf_internal and *movti_internal. */
1096 void
1097 ia64_split_tmode_move (rtx operands[])
1099 rtx in[2], out[2], insn;
1100 rtx fixup[2];
1101 bool dead = false;
1102 bool reversed = false;
1104 /* It is possible for reload to decide to overwrite a pointer with
1105 the value it points to. In that case we have to do the loads in
1106 the appropriate order so that the pointer is not destroyed too
1107 early. Also we must not generate a postmodify for that second
1108 load, or rws_access_regno will abort. */
1109 if (GET_CODE (operands[1]) == MEM
1110 && reg_overlap_mentioned_p (operands[0], operands[1]))
1112 rtx base = XEXP (operands[1], 0);
1113 while (GET_CODE (base) != REG)
1114 base = XEXP (base, 0);
1116 if (REGNO (base) == REGNO (operands[0]))
1117 reversed = true;
1118 dead = true;
1120 /* Another reason to do the moves in reversed order is if the first
1121 element of the target register pair is also the second element of
1122 the source register pair. */
1123 if (GET_CODE (operands[0]) == REG && GET_CODE (operands[1]) == REG
1124 && REGNO (operands[0]) == REGNO (operands[1]) + 1)
1125 reversed = true;
1127 fixup[0] = ia64_split_tmode (in, operands[1], reversed, dead);
1128 fixup[1] = ia64_split_tmode (out, operands[0], reversed, dead);
1130 #define MAYBE_ADD_REG_INC_NOTE(INSN, EXP) \
1131 if (GET_CODE (EXP) == MEM \
1132 && (GET_CODE (XEXP (EXP, 0)) == POST_MODIFY \
1133 || GET_CODE (XEXP (EXP, 0)) == POST_INC \
1134 || GET_CODE (XEXP (EXP, 0)) == POST_DEC)) \
1135 REG_NOTES (INSN) = gen_rtx_EXPR_LIST (REG_INC, \
1136 XEXP (XEXP (EXP, 0), 0), \
1137 REG_NOTES (INSN))
1139 insn = emit_insn (gen_rtx_SET (VOIDmode, out[0], in[0]));
1140 MAYBE_ADD_REG_INC_NOTE (insn, in[0]);
1141 MAYBE_ADD_REG_INC_NOTE (insn, out[0]);
1143 insn = emit_insn (gen_rtx_SET (VOIDmode, out[1], in[1]));
1144 MAYBE_ADD_REG_INC_NOTE (insn, in[1]);
1145 MAYBE_ADD_REG_INC_NOTE (insn, out[1]);
1147 if (fixup[0])
1148 emit_insn (fixup[0]);
1149 if (fixup[1])
1150 emit_insn (fixup[1]);
1152 #undef MAYBE_ADD_REG_INC_NOTE
1155 /* ??? Fixing GR->FR XFmode moves during reload is hard. You need to go
1156 through memory plus an extra GR scratch register. Except that you can
1157 either get the first from SECONDARY_MEMORY_NEEDED or the second from
1158 SECONDARY_RELOAD_CLASS, but not both.
1160 We got into problems in the first place by allowing a construct like
1161 (subreg:XF (reg:TI)), which we got from a union containing a long double.
1162 This solution attempts to prevent this situation from occurring. When
1163 we see something like the above, we spill the inner register to memory. */
1166 spill_xfmode_operand (rtx in, int force)
1168 if (GET_CODE (in) == SUBREG
1169 && GET_MODE (SUBREG_REG (in)) == TImode
1170 && GET_CODE (SUBREG_REG (in)) == REG)
1172 rtx memt = assign_stack_temp (TImode, 16, 0);
1173 emit_move_insn (memt, SUBREG_REG (in));
1174 return adjust_address (memt, XFmode, 0);
1176 else if (force && GET_CODE (in) == REG)
1178 rtx memx = assign_stack_temp (XFmode, 16, 0);
1179 emit_move_insn (memx, in);
1180 return memx;
1182 else
1183 return in;
1186 /* Emit comparison instruction if necessary, returning the expression
1187 that holds the compare result in the proper mode. */
1189 static GTY(()) rtx cmptf_libfunc;
1192 ia64_expand_compare (enum rtx_code code, enum machine_mode mode)
1194 rtx op0 = ia64_compare_op0, op1 = ia64_compare_op1;
1195 rtx cmp;
1197 /* If we have a BImode input, then we already have a compare result, and
1198 do not need to emit another comparison. */
1199 if (GET_MODE (op0) == BImode)
1201 if ((code == NE || code == EQ) && op1 == const0_rtx)
1202 cmp = op0;
1203 else
1204 abort ();
1206 /* HPUX TFmode compare requires a library call to _U_Qfcmp, which takes a
1207 magic number as its third argument, that indicates what to do.
1208 The return value is an integer to be compared against zero. */
1209 else if (GET_MODE (op0) == TFmode)
1211 enum qfcmp_magic {
1212 QCMP_INV = 1, /* Raise FP_INVALID on SNaN as a side effect. */
1213 QCMP_UNORD = 2,
1214 QCMP_EQ = 4,
1215 QCMP_LT = 8,
1216 QCMP_GT = 16
1217 } magic;
1218 enum rtx_code ncode;
1219 rtx ret, insns;
1220 if (!cmptf_libfunc || GET_MODE (op1) != TFmode)
1221 abort ();
1222 switch (code)
1224 /* 1 = equal, 0 = not equal. Equality operators do
1225 not raise FP_INVALID when given an SNaN operand. */
1226 case EQ: magic = QCMP_EQ; ncode = NE; break;
1227 case NE: magic = QCMP_EQ; ncode = EQ; break;
1228 /* isunordered() from C99. */
1229 case UNORDERED: magic = QCMP_UNORD; ncode = NE; break;
1230 case ORDERED: magic = QCMP_UNORD; ncode = EQ; break;
1231 /* Relational operators raise FP_INVALID when given
1232 an SNaN operand. */
1233 case LT: magic = QCMP_LT |QCMP_INV; ncode = NE; break;
1234 case LE: magic = QCMP_LT|QCMP_EQ|QCMP_INV; ncode = NE; break;
1235 case GT: magic = QCMP_GT |QCMP_INV; ncode = NE; break;
1236 case GE: magic = QCMP_GT|QCMP_EQ|QCMP_INV; ncode = NE; break;
1237 /* FUTURE: Implement UNEQ, UNLT, UNLE, UNGT, UNGE, LTGT.
1238 Expanders for buneq etc. weuld have to be added to ia64.md
1239 for this to be useful. */
1240 default: abort ();
1243 start_sequence ();
1245 ret = emit_library_call_value (cmptf_libfunc, 0, LCT_CONST, DImode, 3,
1246 op0, TFmode, op1, TFmode,
1247 GEN_INT (magic), DImode);
1248 cmp = gen_reg_rtx (BImode);
1249 emit_insn (gen_rtx_SET (VOIDmode, cmp,
1250 gen_rtx_fmt_ee (ncode, BImode,
1251 ret, const0_rtx)));
1253 insns = get_insns ();
1254 end_sequence ();
1256 emit_libcall_block (insns, cmp, cmp,
1257 gen_rtx_fmt_ee (code, BImode, op0, op1));
1258 code = NE;
1260 else
1262 cmp = gen_reg_rtx (BImode);
1263 emit_insn (gen_rtx_SET (VOIDmode, cmp,
1264 gen_rtx_fmt_ee (code, BImode, op0, op1)));
1265 code = NE;
1268 return gen_rtx_fmt_ee (code, mode, cmp, const0_rtx);
1271 /* Generate an integral vector comparison. */
1273 static bool
1274 ia64_expand_vecint_compare (enum rtx_code code, enum machine_mode mode,
1275 rtx dest, rtx op0, rtx op1)
1277 bool negate = false;
1278 rtx x;
1280 switch (code)
1282 case EQ:
1283 case GT:
1284 break;
1286 case NE:
1287 code = EQ;
1288 negate = true;
1289 break;
1291 case LE:
1292 code = GT;
1293 negate = true;
1294 break;
1296 case GE:
1297 negate = true;
1298 /* FALLTHRU */
1300 case LT:
1301 x = op0;
1302 op0 = op1;
1303 op1 = x;
1304 code = GT;
1305 break;
1307 case GTU:
1308 case GEU:
1309 case LTU:
1310 case LEU:
1312 rtx w0h, w0l, w1h, w1l, ch, cl;
1313 enum machine_mode wmode;
1314 rtx (*unpack_l) (rtx, rtx, rtx);
1315 rtx (*unpack_h) (rtx, rtx, rtx);
1316 rtx (*pack) (rtx, rtx, rtx);
1318 /* We don't have native unsigned comparisons, but we can generate
1319 them better than generic code can. */
1321 if (mode == V2SImode)
1322 abort ();
1323 else if (mode == V8QImode)
1325 wmode = V4HImode;
1326 pack = gen_pack2_sss;
1327 unpack_l = gen_unpack1_l;
1328 unpack_h = gen_unpack1_h;
1330 else if (mode == V4HImode)
1332 wmode = V2SImode;
1333 pack = gen_pack4_sss;
1334 unpack_l = gen_unpack2_l;
1335 unpack_h = gen_unpack2_h;
1337 else
1338 abort ();
1340 /* Unpack into wider vectors, zero extending the elements. */
1342 w0l = gen_reg_rtx (wmode);
1343 w0h = gen_reg_rtx (wmode);
1344 w1l = gen_reg_rtx (wmode);
1345 w1h = gen_reg_rtx (wmode);
1346 emit_insn (unpack_l (gen_lowpart (mode, w0l), op0, CONST0_RTX (mode)));
1347 emit_insn (unpack_h (gen_lowpart (mode, w0h), op0, CONST0_RTX (mode)));
1348 emit_insn (unpack_l (gen_lowpart (mode, w1l), op1, CONST0_RTX (mode)));
1349 emit_insn (unpack_h (gen_lowpart (mode, w1h), op1, CONST0_RTX (mode)));
1351 /* Compare in the wider mode. */
1353 cl = gen_reg_rtx (wmode);
1354 ch = gen_reg_rtx (wmode);
1355 code = signed_condition (code);
1356 ia64_expand_vecint_compare (code, wmode, cl, w0l, w1l);
1357 negate = ia64_expand_vecint_compare (code, wmode, ch, w0h, w1h);
1359 /* Repack into a single narrower vector. */
1361 emit_insn (pack (dest, cl, ch));
1363 return negate;
1365 default:
1366 abort ();
1369 x = gen_rtx_fmt_ee (code, mode, op0, op1);
1370 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
1372 return negate;
1375 static void
1376 ia64_expand_vcondu_v2si (enum rtx_code code, rtx operands[])
1378 rtx dl, dh, bl, bh, op1l, op1h, op2l, op2h, op4l, op4h, op5l, op5h, x;
1380 /* In this case, we extract the two SImode quantities and generate
1381 normal comparisons for each of them. */
1383 op1l = gen_lowpart (SImode, operands[1]);
1384 op2l = gen_lowpart (SImode, operands[2]);
1385 op4l = gen_lowpart (SImode, operands[4]);
1386 op5l = gen_lowpart (SImode, operands[5]);
1388 op1h = gen_reg_rtx (SImode);
1389 op2h = gen_reg_rtx (SImode);
1390 op4h = gen_reg_rtx (SImode);
1391 op5h = gen_reg_rtx (SImode);
1393 emit_insn (gen_lshrdi3 (gen_lowpart (DImode, op1h),
1394 gen_lowpart (DImode, operands[1]), GEN_INT (32)));
1395 emit_insn (gen_lshrdi3 (gen_lowpart (DImode, op2h),
1396 gen_lowpart (DImode, operands[2]), GEN_INT (32)));
1397 emit_insn (gen_lshrdi3 (gen_lowpart (DImode, op4h),
1398 gen_lowpart (DImode, operands[4]), GEN_INT (32)));
1399 emit_insn (gen_lshrdi3 (gen_lowpart (DImode, op5h),
1400 gen_lowpart (DImode, operands[5]), GEN_INT (32)));
1402 bl = gen_reg_rtx (BImode);
1403 x = gen_rtx_fmt_ee (code, BImode, op4l, op5l);
1404 emit_insn (gen_rtx_SET (VOIDmode, bl, x));
1406 bh = gen_reg_rtx (BImode);
1407 x = gen_rtx_fmt_ee (code, BImode, op4h, op5h);
1408 emit_insn (gen_rtx_SET (VOIDmode, bh, x));
1410 /* With the results of the comparisons, emit conditional moves. */
1412 dl = gen_reg_rtx (SImode);
1413 x = gen_rtx_IF_THEN_ELSE (SImode, bl, op1l, op2l);
1414 emit_insn (gen_rtx_SET (VOIDmode, dl, x));
1416 dh = gen_reg_rtx (SImode);
1417 x = gen_rtx_IF_THEN_ELSE (SImode, bh, op1h, op2h);
1418 emit_insn (gen_rtx_SET (VOIDmode, dh, x));
1420 /* Merge the two partial results back into a vector. */
1422 x = gen_rtx_VEC_CONCAT (V2SImode, dl, dh);
1423 emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
1426 /* Emit an integral vector conditional move. */
1428 void
1429 ia64_expand_vecint_cmov (rtx operands[])
1431 enum machine_mode mode = GET_MODE (operands[0]);
1432 enum rtx_code code = GET_CODE (operands[3]);
1433 bool negate;
1434 rtx cmp, x, ot, of;
1436 /* Since we don't have unsigned V2SImode comparisons, it's more efficient
1437 to special-case them entirely. */
1438 if (mode == V2SImode
1439 && (code == GTU || code == GEU || code == LEU || code == LTU))
1441 ia64_expand_vcondu_v2si (code, operands);
1442 return;
1445 cmp = gen_reg_rtx (mode);
1446 negate = ia64_expand_vecint_compare (code, mode, cmp,
1447 operands[4], operands[5]);
1449 ot = operands[1+negate];
1450 of = operands[2-negate];
1452 if (ot == CONST0_RTX (mode))
1454 if (of == CONST0_RTX (mode))
1456 emit_move_insn (operands[0], ot);
1457 return;
1460 x = gen_rtx_NOT (mode, cmp);
1461 x = gen_rtx_AND (mode, x, of);
1462 emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
1464 else if (of == CONST0_RTX (mode))
1466 x = gen_rtx_AND (mode, cmp, ot);
1467 emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
1469 else
1471 rtx t, f;
1473 t = gen_reg_rtx (mode);
1474 x = gen_rtx_AND (mode, cmp, operands[1+negate]);
1475 emit_insn (gen_rtx_SET (VOIDmode, t, x));
1477 f = gen_reg_rtx (mode);
1478 x = gen_rtx_NOT (mode, cmp);
1479 x = gen_rtx_AND (mode, x, operands[2-negate]);
1480 emit_insn (gen_rtx_SET (VOIDmode, f, x));
1482 x = gen_rtx_IOR (mode, t, f);
1483 emit_insn (gen_rtx_SET (VOIDmode, operands[0], x));
1487 /* Emit an integral vector min or max operation. Return true if all done. */
1489 bool
1490 ia64_expand_vecint_minmax (enum rtx_code code, enum machine_mode mode,
1491 rtx operands[])
1493 rtx xops[5];
1495 /* These four combinations are supported directly. */
1496 if (mode == V8QImode && (code == UMIN || code == UMAX))
1497 return false;
1498 if (mode == V4HImode && (code == SMIN || code == SMAX))
1499 return false;
1501 /* Everything else implemented via vector comparisons. */
1502 xops[0] = operands[0];
1503 xops[4] = xops[1] = operands[1];
1504 xops[5] = xops[2] = operands[2];
1506 switch (code)
1508 case UMIN:
1509 code = LTU;
1510 break;
1511 case UMAX:
1512 code = GTU;
1513 break;
1514 case SMIN:
1515 code = LT;
1516 break;
1517 case SMAX:
1518 code = GT;
1519 break;
1520 default:
1521 abort ();
1523 xops[3] = gen_rtx_fmt_ee (code, VOIDmode, operands[1], operands[2]);
1525 ia64_expand_vecint_cmov (xops);
1526 return true;
1529 /* Emit the appropriate sequence for a call. */
1531 void
1532 ia64_expand_call (rtx retval, rtx addr, rtx nextarg ATTRIBUTE_UNUSED,
1533 int sibcall_p)
1535 rtx insn, b0;
1537 addr = XEXP (addr, 0);
1538 addr = convert_memory_address (DImode, addr);
1539 b0 = gen_rtx_REG (DImode, R_BR (0));
1541 /* ??? Should do this for functions known to bind local too. */
1542 if (TARGET_NO_PIC || TARGET_AUTO_PIC)
1544 if (sibcall_p)
1545 insn = gen_sibcall_nogp (addr);
1546 else if (! retval)
1547 insn = gen_call_nogp (addr, b0);
1548 else
1549 insn = gen_call_value_nogp (retval, addr, b0);
1550 insn = emit_call_insn (insn);
1552 else
1554 if (sibcall_p)
1555 insn = gen_sibcall_gp (addr);
1556 else if (! retval)
1557 insn = gen_call_gp (addr, b0);
1558 else
1559 insn = gen_call_value_gp (retval, addr, b0);
1560 insn = emit_call_insn (insn);
1562 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
1565 if (sibcall_p)
1566 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), b0);
1569 void
1570 ia64_reload_gp (void)
1572 rtx tmp;
1574 if (current_frame_info.reg_save_gp)
1575 tmp = gen_rtx_REG (DImode, current_frame_info.reg_save_gp);
1576 else
1578 HOST_WIDE_INT offset;
1580 offset = (current_frame_info.spill_cfa_off
1581 + current_frame_info.spill_size);
1582 if (frame_pointer_needed)
1584 tmp = hard_frame_pointer_rtx;
1585 offset = -offset;
1587 else
1589 tmp = stack_pointer_rtx;
1590 offset = current_frame_info.total_size - offset;
1593 if (CONST_OK_FOR_I (offset))
1594 emit_insn (gen_adddi3 (pic_offset_table_rtx,
1595 tmp, GEN_INT (offset)));
1596 else
1598 emit_move_insn (pic_offset_table_rtx, GEN_INT (offset));
1599 emit_insn (gen_adddi3 (pic_offset_table_rtx,
1600 pic_offset_table_rtx, tmp));
1603 tmp = gen_rtx_MEM (DImode, pic_offset_table_rtx);
1606 emit_move_insn (pic_offset_table_rtx, tmp);
1609 void
1610 ia64_split_call (rtx retval, rtx addr, rtx retaddr, rtx scratch_r,
1611 rtx scratch_b, int noreturn_p, int sibcall_p)
1613 rtx insn;
1614 bool is_desc = false;
1616 /* If we find we're calling through a register, then we're actually
1617 calling through a descriptor, so load up the values. */
1618 if (REG_P (addr) && GR_REGNO_P (REGNO (addr)))
1620 rtx tmp;
1621 bool addr_dead_p;
1623 /* ??? We are currently constrained to *not* use peep2, because
1624 we can legitimately change the global lifetime of the GP
1625 (in the form of killing where previously live). This is
1626 because a call through a descriptor doesn't use the previous
1627 value of the GP, while a direct call does, and we do not
1628 commit to either form until the split here.
1630 That said, this means that we lack precise life info for
1631 whether ADDR is dead after this call. This is not terribly
1632 important, since we can fix things up essentially for free
1633 with the POST_DEC below, but it's nice to not use it when we
1634 can immediately tell it's not necessary. */
1635 addr_dead_p = ((noreturn_p || sibcall_p
1636 || TEST_HARD_REG_BIT (regs_invalidated_by_call,
1637 REGNO (addr)))
1638 && !FUNCTION_ARG_REGNO_P (REGNO (addr)));
1640 /* Load the code address into scratch_b. */
1641 tmp = gen_rtx_POST_INC (Pmode, addr);
1642 tmp = gen_rtx_MEM (Pmode, tmp);
1643 emit_move_insn (scratch_r, tmp);
1644 emit_move_insn (scratch_b, scratch_r);
1646 /* Load the GP address. If ADDR is not dead here, then we must
1647 revert the change made above via the POST_INCREMENT. */
1648 if (!addr_dead_p)
1649 tmp = gen_rtx_POST_DEC (Pmode, addr);
1650 else
1651 tmp = addr;
1652 tmp = gen_rtx_MEM (Pmode, tmp);
1653 emit_move_insn (pic_offset_table_rtx, tmp);
1655 is_desc = true;
1656 addr = scratch_b;
1659 if (sibcall_p)
1660 insn = gen_sibcall_nogp (addr);
1661 else if (retval)
1662 insn = gen_call_value_nogp (retval, addr, retaddr);
1663 else
1664 insn = gen_call_nogp (addr, retaddr);
1665 emit_call_insn (insn);
1667 if ((!TARGET_CONST_GP || is_desc) && !noreturn_p && !sibcall_p)
1668 ia64_reload_gp ();
1671 /* Begin the assembly file. */
1673 static void
1674 ia64_file_start (void)
1676 default_file_start ();
1677 emit_safe_across_calls ();
1680 void
1681 emit_safe_across_calls (void)
1683 unsigned int rs, re;
1684 int out_state;
1686 rs = 1;
1687 out_state = 0;
1688 while (1)
1690 while (rs < 64 && call_used_regs[PR_REG (rs)])
1691 rs++;
1692 if (rs >= 64)
1693 break;
1694 for (re = rs + 1; re < 64 && ! call_used_regs[PR_REG (re)]; re++)
1695 continue;
1696 if (out_state == 0)
1698 fputs ("\t.pred.safe_across_calls ", asm_out_file);
1699 out_state = 1;
1701 else
1702 fputc (',', asm_out_file);
1703 if (re == rs + 1)
1704 fprintf (asm_out_file, "p%u", rs);
1705 else
1706 fprintf (asm_out_file, "p%u-p%u", rs, re - 1);
1707 rs = re + 1;
1709 if (out_state)
1710 fputc ('\n', asm_out_file);
1713 /* Helper function for ia64_compute_frame_size: find an appropriate general
1714 register to spill some special register to. SPECIAL_SPILL_MASK contains
1715 bits in GR0 to GR31 that have already been allocated by this routine.
1716 TRY_LOCALS is true if we should attempt to locate a local regnum. */
1718 static int
1719 find_gr_spill (int try_locals)
1721 int regno;
1723 /* If this is a leaf function, first try an otherwise unused
1724 call-clobbered register. */
1725 if (current_function_is_leaf)
1727 for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
1728 if (! regs_ever_live[regno]
1729 && call_used_regs[regno]
1730 && ! fixed_regs[regno]
1731 && ! global_regs[regno]
1732 && ((current_frame_info.gr_used_mask >> regno) & 1) == 0)
1734 current_frame_info.gr_used_mask |= 1 << regno;
1735 return regno;
1739 if (try_locals)
1741 regno = current_frame_info.n_local_regs;
1742 /* If there is a frame pointer, then we can't use loc79, because
1743 that is HARD_FRAME_POINTER_REGNUM. In particular, see the
1744 reg_name switching code in ia64_expand_prologue. */
1745 if (regno < (80 - frame_pointer_needed))
1747 current_frame_info.n_local_regs = regno + 1;
1748 return LOC_REG (0) + regno;
1752 /* Failed to find a general register to spill to. Must use stack. */
1753 return 0;
1756 /* In order to make for nice schedules, we try to allocate every temporary
1757 to a different register. We must of course stay away from call-saved,
1758 fixed, and global registers. We must also stay away from registers
1759 allocated in current_frame_info.gr_used_mask, since those include regs
1760 used all through the prologue.
1762 Any register allocated here must be used immediately. The idea is to
1763 aid scheduling, not to solve data flow problems. */
1765 static int last_scratch_gr_reg;
1767 static int
1768 next_scratch_gr_reg (void)
1770 int i, regno;
1772 for (i = 0; i < 32; ++i)
1774 regno = (last_scratch_gr_reg + i + 1) & 31;
1775 if (call_used_regs[regno]
1776 && ! fixed_regs[regno]
1777 && ! global_regs[regno]
1778 && ((current_frame_info.gr_used_mask >> regno) & 1) == 0)
1780 last_scratch_gr_reg = regno;
1781 return regno;
1785 /* There must be _something_ available. */
1786 abort ();
1789 /* Helper function for ia64_compute_frame_size, called through
1790 diddle_return_value. Mark REG in current_frame_info.gr_used_mask. */
1792 static void
1793 mark_reg_gr_used_mask (rtx reg, void *data ATTRIBUTE_UNUSED)
1795 unsigned int regno = REGNO (reg);
1796 if (regno < 32)
1798 unsigned int i, n = HARD_REGNO_NREGS (regno, GET_MODE (reg));
1799 for (i = 0; i < n; ++i)
1800 current_frame_info.gr_used_mask |= 1 << (regno + i);
1804 /* Returns the number of bytes offset between the frame pointer and the stack
1805 pointer for the current function. SIZE is the number of bytes of space
1806 needed for local variables. */
1808 static void
1809 ia64_compute_frame_size (HOST_WIDE_INT size)
1811 HOST_WIDE_INT total_size;
1812 HOST_WIDE_INT spill_size = 0;
1813 HOST_WIDE_INT extra_spill_size = 0;
1814 HOST_WIDE_INT pretend_args_size;
1815 HARD_REG_SET mask;
1816 int n_spilled = 0;
1817 int spilled_gr_p = 0;
1818 int spilled_fr_p = 0;
1819 unsigned int regno;
1820 int i;
1822 if (current_frame_info.initialized)
1823 return;
1825 memset (&current_frame_info, 0, sizeof current_frame_info);
1826 CLEAR_HARD_REG_SET (mask);
1828 /* Don't allocate scratches to the return register. */
1829 diddle_return_value (mark_reg_gr_used_mask, NULL);
1831 /* Don't allocate scratches to the EH scratch registers. */
1832 if (cfun->machine->ia64_eh_epilogue_sp)
1833 mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_sp, NULL);
1834 if (cfun->machine->ia64_eh_epilogue_bsp)
1835 mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_bsp, NULL);
1837 /* Find the size of the register stack frame. We have only 80 local
1838 registers, because we reserve 8 for the inputs and 8 for the
1839 outputs. */
1841 /* Skip HARD_FRAME_POINTER_REGNUM (loc79) when frame_pointer_needed,
1842 since we'll be adjusting that down later. */
1843 regno = LOC_REG (78) + ! frame_pointer_needed;
1844 for (; regno >= LOC_REG (0); regno--)
1845 if (regs_ever_live[regno])
1846 break;
1847 current_frame_info.n_local_regs = regno - LOC_REG (0) + 1;
1849 /* For functions marked with the syscall_linkage attribute, we must mark
1850 all eight input registers as in use, so that locals aren't visible to
1851 the caller. */
1853 if (cfun->machine->n_varargs > 0
1854 || lookup_attribute ("syscall_linkage",
1855 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
1856 current_frame_info.n_input_regs = 8;
1857 else
1859 for (regno = IN_REG (7); regno >= IN_REG (0); regno--)
1860 if (regs_ever_live[regno])
1861 break;
1862 current_frame_info.n_input_regs = regno - IN_REG (0) + 1;
1865 for (regno = OUT_REG (7); regno >= OUT_REG (0); regno--)
1866 if (regs_ever_live[regno])
1867 break;
1868 i = regno - OUT_REG (0) + 1;
1870 /* When -p profiling, we need one output register for the mcount argument.
1871 Likewise for -a profiling for the bb_init_func argument. For -ax
1872 profiling, we need two output registers for the two bb_init_trace_func
1873 arguments. */
1874 if (current_function_profile)
1875 i = MAX (i, 1);
1876 current_frame_info.n_output_regs = i;
1878 /* ??? No rotating register support yet. */
1879 current_frame_info.n_rotate_regs = 0;
1881 /* Discover which registers need spilling, and how much room that
1882 will take. Begin with floating point and general registers,
1883 which will always wind up on the stack. */
1885 for (regno = FR_REG (2); regno <= FR_REG (127); regno++)
1886 if (regs_ever_live[regno] && ! call_used_regs[regno])
1888 SET_HARD_REG_BIT (mask, regno);
1889 spill_size += 16;
1890 n_spilled += 1;
1891 spilled_fr_p = 1;
1894 for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
1895 if (regs_ever_live[regno] && ! call_used_regs[regno])
1897 SET_HARD_REG_BIT (mask, regno);
1898 spill_size += 8;
1899 n_spilled += 1;
1900 spilled_gr_p = 1;
1903 for (regno = BR_REG (1); regno <= BR_REG (7); regno++)
1904 if (regs_ever_live[regno] && ! call_used_regs[regno])
1906 SET_HARD_REG_BIT (mask, regno);
1907 spill_size += 8;
1908 n_spilled += 1;
1911 /* Now come all special registers that might get saved in other
1912 general registers. */
1914 if (frame_pointer_needed)
1916 current_frame_info.reg_fp = find_gr_spill (1);
1917 /* If we did not get a register, then we take LOC79. This is guaranteed
1918 to be free, even if regs_ever_live is already set, because this is
1919 HARD_FRAME_POINTER_REGNUM. This requires incrementing n_local_regs,
1920 as we don't count loc79 above. */
1921 if (current_frame_info.reg_fp == 0)
1923 current_frame_info.reg_fp = LOC_REG (79);
1924 current_frame_info.n_local_regs++;
1928 if (! current_function_is_leaf)
1930 /* Emit a save of BR0 if we call other functions. Do this even
1931 if this function doesn't return, as EH depends on this to be
1932 able to unwind the stack. */
1933 SET_HARD_REG_BIT (mask, BR_REG (0));
1935 current_frame_info.reg_save_b0 = find_gr_spill (1);
1936 if (current_frame_info.reg_save_b0 == 0)
1938 spill_size += 8;
1939 n_spilled += 1;
1942 /* Similarly for ar.pfs. */
1943 SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
1944 current_frame_info.reg_save_ar_pfs = find_gr_spill (1);
1945 if (current_frame_info.reg_save_ar_pfs == 0)
1947 extra_spill_size += 8;
1948 n_spilled += 1;
1951 /* Similarly for gp. Note that if we're calling setjmp, the stacked
1952 registers are clobbered, so we fall back to the stack. */
1953 current_frame_info.reg_save_gp
1954 = (current_function_calls_setjmp ? 0 : find_gr_spill (1));
1955 if (current_frame_info.reg_save_gp == 0)
1957 SET_HARD_REG_BIT (mask, GR_REG (1));
1958 spill_size += 8;
1959 n_spilled += 1;
1962 else
1964 if (regs_ever_live[BR_REG (0)] && ! call_used_regs[BR_REG (0)])
1966 SET_HARD_REG_BIT (mask, BR_REG (0));
1967 spill_size += 8;
1968 n_spilled += 1;
1971 if (regs_ever_live[AR_PFS_REGNUM])
1973 SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
1974 current_frame_info.reg_save_ar_pfs = find_gr_spill (1);
1975 if (current_frame_info.reg_save_ar_pfs == 0)
1977 extra_spill_size += 8;
1978 n_spilled += 1;
1983 /* Unwind descriptor hackery: things are most efficient if we allocate
1984 consecutive GR save registers for RP, PFS, FP in that order. However,
1985 it is absolutely critical that FP get the only hard register that's
1986 guaranteed to be free, so we allocated it first. If all three did
1987 happen to be allocated hard regs, and are consecutive, rearrange them
1988 into the preferred order now. */
1989 if (current_frame_info.reg_fp != 0
1990 && current_frame_info.reg_save_b0 == current_frame_info.reg_fp + 1
1991 && current_frame_info.reg_save_ar_pfs == current_frame_info.reg_fp + 2)
1993 current_frame_info.reg_save_b0 = current_frame_info.reg_fp;
1994 current_frame_info.reg_save_ar_pfs = current_frame_info.reg_fp + 1;
1995 current_frame_info.reg_fp = current_frame_info.reg_fp + 2;
1998 /* See if we need to store the predicate register block. */
1999 for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
2000 if (regs_ever_live[regno] && ! call_used_regs[regno])
2001 break;
2002 if (regno <= PR_REG (63))
2004 SET_HARD_REG_BIT (mask, PR_REG (0));
2005 current_frame_info.reg_save_pr = find_gr_spill (1);
2006 if (current_frame_info.reg_save_pr == 0)
2008 extra_spill_size += 8;
2009 n_spilled += 1;
2012 /* ??? Mark them all as used so that register renaming and such
2013 are free to use them. */
2014 for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
2015 regs_ever_live[regno] = 1;
2018 /* If we're forced to use st8.spill, we're forced to save and restore
2019 ar.unat as well. The check for existing liveness allows inline asm
2020 to touch ar.unat. */
2021 if (spilled_gr_p || cfun->machine->n_varargs
2022 || regs_ever_live[AR_UNAT_REGNUM])
2024 regs_ever_live[AR_UNAT_REGNUM] = 1;
2025 SET_HARD_REG_BIT (mask, AR_UNAT_REGNUM);
2026 current_frame_info.reg_save_ar_unat = find_gr_spill (spill_size == 0);
2027 if (current_frame_info.reg_save_ar_unat == 0)
2029 extra_spill_size += 8;
2030 n_spilled += 1;
2034 if (regs_ever_live[AR_LC_REGNUM])
2036 SET_HARD_REG_BIT (mask, AR_LC_REGNUM);
2037 current_frame_info.reg_save_ar_lc = find_gr_spill (spill_size == 0);
2038 if (current_frame_info.reg_save_ar_lc == 0)
2040 extra_spill_size += 8;
2041 n_spilled += 1;
2045 /* If we have an odd number of words of pretend arguments written to
2046 the stack, then the FR save area will be unaligned. We round the
2047 size of this area up to keep things 16 byte aligned. */
2048 if (spilled_fr_p)
2049 pretend_args_size = IA64_STACK_ALIGN (current_function_pretend_args_size);
2050 else
2051 pretend_args_size = current_function_pretend_args_size;
2053 total_size = (spill_size + extra_spill_size + size + pretend_args_size
2054 + current_function_outgoing_args_size);
2055 total_size = IA64_STACK_ALIGN (total_size);
2057 /* We always use the 16-byte scratch area provided by the caller, but
2058 if we are a leaf function, there's no one to which we need to provide
2059 a scratch area. */
2060 if (current_function_is_leaf)
2061 total_size = MAX (0, total_size - 16);
2063 current_frame_info.total_size = total_size;
2064 current_frame_info.spill_cfa_off = pretend_args_size - 16;
2065 current_frame_info.spill_size = spill_size;
2066 current_frame_info.extra_spill_size = extra_spill_size;
2067 COPY_HARD_REG_SET (current_frame_info.mask, mask);
2068 current_frame_info.n_spilled = n_spilled;
2069 current_frame_info.initialized = reload_completed;
2072 /* Compute the initial difference between the specified pair of registers. */
2074 HOST_WIDE_INT
2075 ia64_initial_elimination_offset (int from, int to)
2077 HOST_WIDE_INT offset;
2079 ia64_compute_frame_size (get_frame_size ());
2080 switch (from)
2082 case FRAME_POINTER_REGNUM:
2083 if (to == HARD_FRAME_POINTER_REGNUM)
2085 if (current_function_is_leaf)
2086 offset = -current_frame_info.total_size;
2087 else
2088 offset = -(current_frame_info.total_size
2089 - current_function_outgoing_args_size - 16);
2091 else if (to == STACK_POINTER_REGNUM)
2093 if (current_function_is_leaf)
2094 offset = 0;
2095 else
2096 offset = 16 + current_function_outgoing_args_size;
2098 else
2099 abort ();
2100 break;
2102 case ARG_POINTER_REGNUM:
2103 /* Arguments start above the 16 byte save area, unless stdarg
2104 in which case we store through the 16 byte save area. */
2105 if (to == HARD_FRAME_POINTER_REGNUM)
2106 offset = 16 - current_function_pretend_args_size;
2107 else if (to == STACK_POINTER_REGNUM)
2108 offset = (current_frame_info.total_size
2109 + 16 - current_function_pretend_args_size);
2110 else
2111 abort ();
2112 break;
2114 default:
2115 abort ();
2118 return offset;
2121 /* If there are more than a trivial number of register spills, we use
2122 two interleaved iterators so that we can get two memory references
2123 per insn group.
2125 In order to simplify things in the prologue and epilogue expanders,
2126 we use helper functions to fix up the memory references after the
2127 fact with the appropriate offsets to a POST_MODIFY memory mode.
2128 The following data structure tracks the state of the two iterators
2129 while insns are being emitted. */
2131 struct spill_fill_data
2133 rtx init_after; /* point at which to emit initializations */
2134 rtx init_reg[2]; /* initial base register */
2135 rtx iter_reg[2]; /* the iterator registers */
2136 rtx *prev_addr[2]; /* address of last memory use */
2137 rtx prev_insn[2]; /* the insn corresponding to prev_addr */
2138 HOST_WIDE_INT prev_off[2]; /* last offset */
2139 int n_iter; /* number of iterators in use */
2140 int next_iter; /* next iterator to use */
2141 unsigned int save_gr_used_mask;
2144 static struct spill_fill_data spill_fill_data;
2146 static void
2147 setup_spill_pointers (int n_spills, rtx init_reg, HOST_WIDE_INT cfa_off)
2149 int i;
2151 spill_fill_data.init_after = get_last_insn ();
2152 spill_fill_data.init_reg[0] = init_reg;
2153 spill_fill_data.init_reg[1] = init_reg;
2154 spill_fill_data.prev_addr[0] = NULL;
2155 spill_fill_data.prev_addr[1] = NULL;
2156 spill_fill_data.prev_insn[0] = NULL;
2157 spill_fill_data.prev_insn[1] = NULL;
2158 spill_fill_data.prev_off[0] = cfa_off;
2159 spill_fill_data.prev_off[1] = cfa_off;
2160 spill_fill_data.next_iter = 0;
2161 spill_fill_data.save_gr_used_mask = current_frame_info.gr_used_mask;
2163 spill_fill_data.n_iter = 1 + (n_spills > 2);
2164 for (i = 0; i < spill_fill_data.n_iter; ++i)
2166 int regno = next_scratch_gr_reg ();
2167 spill_fill_data.iter_reg[i] = gen_rtx_REG (DImode, regno);
2168 current_frame_info.gr_used_mask |= 1 << regno;
2172 static void
2173 finish_spill_pointers (void)
2175 current_frame_info.gr_used_mask = spill_fill_data.save_gr_used_mask;
2178 static rtx
2179 spill_restore_mem (rtx reg, HOST_WIDE_INT cfa_off)
2181 int iter = spill_fill_data.next_iter;
2182 HOST_WIDE_INT disp = spill_fill_data.prev_off[iter] - cfa_off;
2183 rtx disp_rtx = GEN_INT (disp);
2184 rtx mem;
2186 if (spill_fill_data.prev_addr[iter])
2188 if (CONST_OK_FOR_N (disp))
2190 *spill_fill_data.prev_addr[iter]
2191 = gen_rtx_POST_MODIFY (DImode, spill_fill_data.iter_reg[iter],
2192 gen_rtx_PLUS (DImode,
2193 spill_fill_data.iter_reg[iter],
2194 disp_rtx));
2195 REG_NOTES (spill_fill_data.prev_insn[iter])
2196 = gen_rtx_EXPR_LIST (REG_INC, spill_fill_data.iter_reg[iter],
2197 REG_NOTES (spill_fill_data.prev_insn[iter]));
2199 else
2201 /* ??? Could use register post_modify for loads. */
2202 if (! CONST_OK_FOR_I (disp))
2204 rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
2205 emit_move_insn (tmp, disp_rtx);
2206 disp_rtx = tmp;
2208 emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
2209 spill_fill_data.iter_reg[iter], disp_rtx));
2212 /* Micro-optimization: if we've created a frame pointer, it's at
2213 CFA 0, which may allow the real iterator to be initialized lower,
2214 slightly increasing parallelism. Also, if there are few saves
2215 it may eliminate the iterator entirely. */
2216 else if (disp == 0
2217 && spill_fill_data.init_reg[iter] == stack_pointer_rtx
2218 && frame_pointer_needed)
2220 mem = gen_rtx_MEM (GET_MODE (reg), hard_frame_pointer_rtx);
2221 set_mem_alias_set (mem, get_varargs_alias_set ());
2222 return mem;
2224 else
2226 rtx seq, insn;
2228 if (disp == 0)
2229 seq = gen_movdi (spill_fill_data.iter_reg[iter],
2230 spill_fill_data.init_reg[iter]);
2231 else
2233 start_sequence ();
2235 if (! CONST_OK_FOR_I (disp))
2237 rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
2238 emit_move_insn (tmp, disp_rtx);
2239 disp_rtx = tmp;
2242 emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
2243 spill_fill_data.init_reg[iter],
2244 disp_rtx));
2246 seq = get_insns ();
2247 end_sequence ();
2250 /* Careful for being the first insn in a sequence. */
2251 if (spill_fill_data.init_after)
2252 insn = emit_insn_after (seq, spill_fill_data.init_after);
2253 else
2255 rtx first = get_insns ();
2256 if (first)
2257 insn = emit_insn_before (seq, first);
2258 else
2259 insn = emit_insn (seq);
2261 spill_fill_data.init_after = insn;
2263 /* If DISP is 0, we may or may not have a further adjustment
2264 afterward. If we do, then the load/store insn may be modified
2265 to be a post-modify. If we don't, then this copy may be
2266 eliminated by copyprop_hardreg_forward, which makes this
2267 insn garbage, which runs afoul of the sanity check in
2268 propagate_one_insn. So mark this insn as legal to delete. */
2269 if (disp == 0)
2270 REG_NOTES(insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
2271 REG_NOTES (insn));
2274 mem = gen_rtx_MEM (GET_MODE (reg), spill_fill_data.iter_reg[iter]);
2276 /* ??? Not all of the spills are for varargs, but some of them are.
2277 The rest of the spills belong in an alias set of their own. But
2278 it doesn't actually hurt to include them here. */
2279 set_mem_alias_set (mem, get_varargs_alias_set ());
2281 spill_fill_data.prev_addr[iter] = &XEXP (mem, 0);
2282 spill_fill_data.prev_off[iter] = cfa_off;
2284 if (++iter >= spill_fill_data.n_iter)
2285 iter = 0;
2286 spill_fill_data.next_iter = iter;
2288 return mem;
2291 static void
2292 do_spill (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off,
2293 rtx frame_reg)
2295 int iter = spill_fill_data.next_iter;
2296 rtx mem, insn;
2298 mem = spill_restore_mem (reg, cfa_off);
2299 insn = emit_insn ((*move_fn) (mem, reg, GEN_INT (cfa_off)));
2300 spill_fill_data.prev_insn[iter] = insn;
2302 if (frame_reg)
2304 rtx base;
2305 HOST_WIDE_INT off;
2307 RTX_FRAME_RELATED_P (insn) = 1;
2309 /* Don't even pretend that the unwind code can intuit its way
2310 through a pair of interleaved post_modify iterators. Just
2311 provide the correct answer. */
2313 if (frame_pointer_needed)
2315 base = hard_frame_pointer_rtx;
2316 off = - cfa_off;
2318 else
2320 base = stack_pointer_rtx;
2321 off = current_frame_info.total_size - cfa_off;
2324 REG_NOTES (insn)
2325 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
2326 gen_rtx_SET (VOIDmode,
2327 gen_rtx_MEM (GET_MODE (reg),
2328 plus_constant (base, off)),
2329 frame_reg),
2330 REG_NOTES (insn));
2334 static void
2335 do_restore (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off)
2337 int iter = spill_fill_data.next_iter;
2338 rtx insn;
2340 insn = emit_insn ((*move_fn) (reg, spill_restore_mem (reg, cfa_off),
2341 GEN_INT (cfa_off)));
2342 spill_fill_data.prev_insn[iter] = insn;
2345 /* Wrapper functions that discards the CONST_INT spill offset. These
2346 exist so that we can give gr_spill/gr_fill the offset they need and
2347 use a consistent function interface. */
2349 static rtx
2350 gen_movdi_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
2352 return gen_movdi (dest, src);
2355 static rtx
2356 gen_fr_spill_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
2358 return gen_fr_spill (dest, src);
2361 static rtx
2362 gen_fr_restore_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
2364 return gen_fr_restore (dest, src);
2367 /* Called after register allocation to add any instructions needed for the
2368 prologue. Using a prologue insn is favored compared to putting all of the
2369 instructions in output_function_prologue(), since it allows the scheduler
2370 to intermix instructions with the saves of the caller saved registers. In
2371 some cases, it might be necessary to emit a barrier instruction as the last
2372 insn to prevent such scheduling.
2374 Also any insns generated here should have RTX_FRAME_RELATED_P(insn) = 1
2375 so that the debug info generation code can handle them properly.
2377 The register save area is layed out like so:
2378 cfa+16
2379 [ varargs spill area ]
2380 [ fr register spill area ]
2381 [ br register spill area ]
2382 [ ar register spill area ]
2383 [ pr register spill area ]
2384 [ gr register spill area ] */
2386 /* ??? Get inefficient code when the frame size is larger than can fit in an
2387 adds instruction. */
2389 void
2390 ia64_expand_prologue (void)
2392 rtx insn, ar_pfs_save_reg, ar_unat_save_reg;
2393 int i, epilogue_p, regno, alt_regno, cfa_off, n_varargs;
2394 rtx reg, alt_reg;
2396 ia64_compute_frame_size (get_frame_size ());
2397 last_scratch_gr_reg = 15;
2399 /* If there is no epilogue, then we don't need some prologue insns.
2400 We need to avoid emitting the dead prologue insns, because flow
2401 will complain about them. */
2402 if (optimize)
2404 edge e;
2405 edge_iterator ei;
2407 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
2408 if ((e->flags & EDGE_FAKE) == 0
2409 && (e->flags & EDGE_FALLTHRU) != 0)
2410 break;
2411 epilogue_p = (e != NULL);
2413 else
2414 epilogue_p = 1;
2416 /* Set the local, input, and output register names. We need to do this
2417 for GNU libc, which creates crti.S/crtn.S by splitting initfini.c in
2418 half. If we use in/loc/out register names, then we get assembler errors
2419 in crtn.S because there is no alloc insn or regstk directive in there. */
2420 if (! TARGET_REG_NAMES)
2422 int inputs = current_frame_info.n_input_regs;
2423 int locals = current_frame_info.n_local_regs;
2424 int outputs = current_frame_info.n_output_regs;
2426 for (i = 0; i < inputs; i++)
2427 reg_names[IN_REG (i)] = ia64_reg_numbers[i];
2428 for (i = 0; i < locals; i++)
2429 reg_names[LOC_REG (i)] = ia64_reg_numbers[inputs + i];
2430 for (i = 0; i < outputs; i++)
2431 reg_names[OUT_REG (i)] = ia64_reg_numbers[inputs + locals + i];
2434 /* Set the frame pointer register name. The regnum is logically loc79,
2435 but of course we'll not have allocated that many locals. Rather than
2436 worrying about renumbering the existing rtxs, we adjust the name. */
2437 /* ??? This code means that we can never use one local register when
2438 there is a frame pointer. loc79 gets wasted in this case, as it is
2439 renamed to a register that will never be used. See also the try_locals
2440 code in find_gr_spill. */
2441 if (current_frame_info.reg_fp)
2443 const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
2444 reg_names[HARD_FRAME_POINTER_REGNUM]
2445 = reg_names[current_frame_info.reg_fp];
2446 reg_names[current_frame_info.reg_fp] = tmp;
2449 /* We don't need an alloc instruction if we've used no outputs or locals. */
2450 if (current_frame_info.n_local_regs == 0
2451 && current_frame_info.n_output_regs == 0
2452 && current_frame_info.n_input_regs <= current_function_args_info.int_regs
2453 && !TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
2455 /* If there is no alloc, but there are input registers used, then we
2456 need a .regstk directive. */
2457 current_frame_info.need_regstk = (TARGET_REG_NAMES != 0);
2458 ar_pfs_save_reg = NULL_RTX;
2460 else
2462 current_frame_info.need_regstk = 0;
2464 if (current_frame_info.reg_save_ar_pfs)
2465 regno = current_frame_info.reg_save_ar_pfs;
2466 else
2467 regno = next_scratch_gr_reg ();
2468 ar_pfs_save_reg = gen_rtx_REG (DImode, regno);
2470 insn = emit_insn (gen_alloc (ar_pfs_save_reg,
2471 GEN_INT (current_frame_info.n_input_regs),
2472 GEN_INT (current_frame_info.n_local_regs),
2473 GEN_INT (current_frame_info.n_output_regs),
2474 GEN_INT (current_frame_info.n_rotate_regs)));
2475 RTX_FRAME_RELATED_P (insn) = (current_frame_info.reg_save_ar_pfs != 0);
2478 /* Set up frame pointer, stack pointer, and spill iterators. */
2480 n_varargs = cfun->machine->n_varargs;
2481 setup_spill_pointers (current_frame_info.n_spilled + n_varargs,
2482 stack_pointer_rtx, 0);
2484 if (frame_pointer_needed)
2486 insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);
2487 RTX_FRAME_RELATED_P (insn) = 1;
2490 if (current_frame_info.total_size != 0)
2492 rtx frame_size_rtx = GEN_INT (- current_frame_info.total_size);
2493 rtx offset;
2495 if (CONST_OK_FOR_I (- current_frame_info.total_size))
2496 offset = frame_size_rtx;
2497 else
2499 regno = next_scratch_gr_reg ();
2500 offset = gen_rtx_REG (DImode, regno);
2501 emit_move_insn (offset, frame_size_rtx);
2504 insn = emit_insn (gen_adddi3 (stack_pointer_rtx,
2505 stack_pointer_rtx, offset));
2507 if (! frame_pointer_needed)
2509 RTX_FRAME_RELATED_P (insn) = 1;
2510 if (GET_CODE (offset) != CONST_INT)
2512 REG_NOTES (insn)
2513 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
2514 gen_rtx_SET (VOIDmode,
2515 stack_pointer_rtx,
2516 gen_rtx_PLUS (DImode,
2517 stack_pointer_rtx,
2518 frame_size_rtx)),
2519 REG_NOTES (insn));
2523 /* ??? At this point we must generate a magic insn that appears to
2524 modify the stack pointer, the frame pointer, and all spill
2525 iterators. This would allow the most scheduling freedom. For
2526 now, just hard stop. */
2527 emit_insn (gen_blockage ());
2530 /* Must copy out ar.unat before doing any integer spills. */
2531 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
2533 if (current_frame_info.reg_save_ar_unat)
2534 ar_unat_save_reg
2535 = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_unat);
2536 else
2538 alt_regno = next_scratch_gr_reg ();
2539 ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
2540 current_frame_info.gr_used_mask |= 1 << alt_regno;
2543 reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
2544 insn = emit_move_insn (ar_unat_save_reg, reg);
2545 RTX_FRAME_RELATED_P (insn) = (current_frame_info.reg_save_ar_unat != 0);
2547 /* Even if we're not going to generate an epilogue, we still
2548 need to save the register so that EH works. */
2549 if (! epilogue_p && current_frame_info.reg_save_ar_unat)
2550 emit_insn (gen_prologue_use (ar_unat_save_reg));
2552 else
2553 ar_unat_save_reg = NULL_RTX;
2555 /* Spill all varargs registers. Do this before spilling any GR registers,
2556 since we want the UNAT bits for the GR registers to override the UNAT
2557 bits from varargs, which we don't care about. */
2559 cfa_off = -16;
2560 for (regno = GR_ARG_FIRST + 7; n_varargs > 0; --n_varargs, --regno)
2562 reg = gen_rtx_REG (DImode, regno);
2563 do_spill (gen_gr_spill, reg, cfa_off += 8, NULL_RTX);
2566 /* Locate the bottom of the register save area. */
2567 cfa_off = (current_frame_info.spill_cfa_off
2568 + current_frame_info.spill_size
2569 + current_frame_info.extra_spill_size);
2571 /* Save the predicate register block either in a register or in memory. */
2572 if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
2574 reg = gen_rtx_REG (DImode, PR_REG (0));
2575 if (current_frame_info.reg_save_pr != 0)
2577 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_pr);
2578 insn = emit_move_insn (alt_reg, reg);
2580 /* ??? Denote pr spill/fill by a DImode move that modifies all
2581 64 hard registers. */
2582 RTX_FRAME_RELATED_P (insn) = 1;
2583 REG_NOTES (insn)
2584 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
2585 gen_rtx_SET (VOIDmode, alt_reg, reg),
2586 REG_NOTES (insn));
2588 /* Even if we're not going to generate an epilogue, we still
2589 need to save the register so that EH works. */
2590 if (! epilogue_p)
2591 emit_insn (gen_prologue_use (alt_reg));
2593 else
2595 alt_regno = next_scratch_gr_reg ();
2596 alt_reg = gen_rtx_REG (DImode, alt_regno);
2597 insn = emit_move_insn (alt_reg, reg);
2598 do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
2599 cfa_off -= 8;
2603 /* Handle AR regs in numerical order. All of them get special handling. */
2604 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM)
2605 && current_frame_info.reg_save_ar_unat == 0)
2607 reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
2608 do_spill (gen_movdi_x, ar_unat_save_reg, cfa_off, reg);
2609 cfa_off -= 8;
2612 /* The alloc insn already copied ar.pfs into a general register. The
2613 only thing we have to do now is copy that register to a stack slot
2614 if we'd not allocated a local register for the job. */
2615 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM)
2616 && current_frame_info.reg_save_ar_pfs == 0)
2618 reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
2619 do_spill (gen_movdi_x, ar_pfs_save_reg, cfa_off, reg);
2620 cfa_off -= 8;
2623 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
2625 reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
2626 if (current_frame_info.reg_save_ar_lc != 0)
2628 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_lc);
2629 insn = emit_move_insn (alt_reg, reg);
2630 RTX_FRAME_RELATED_P (insn) = 1;
2632 /* Even if we're not going to generate an epilogue, we still
2633 need to save the register so that EH works. */
2634 if (! epilogue_p)
2635 emit_insn (gen_prologue_use (alt_reg));
2637 else
2639 alt_regno = next_scratch_gr_reg ();
2640 alt_reg = gen_rtx_REG (DImode, alt_regno);
2641 emit_move_insn (alt_reg, reg);
2642 do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
2643 cfa_off -= 8;
2647 if (current_frame_info.reg_save_gp)
2649 insn = emit_move_insn (gen_rtx_REG (DImode,
2650 current_frame_info.reg_save_gp),
2651 pic_offset_table_rtx);
2652 /* We don't know for sure yet if this is actually needed, since
2653 we've not split the PIC call patterns. If all of the calls
2654 are indirect, and not followed by any uses of the gp, then
2655 this save is dead. Allow it to go away. */
2656 REG_NOTES (insn)
2657 = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx, REG_NOTES (insn));
2660 /* We should now be at the base of the gr/br/fr spill area. */
2661 if (cfa_off != (current_frame_info.spill_cfa_off
2662 + current_frame_info.spill_size))
2663 abort ();
2665 /* Spill all general registers. */
2666 for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
2667 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2669 reg = gen_rtx_REG (DImode, regno);
2670 do_spill (gen_gr_spill, reg, cfa_off, reg);
2671 cfa_off -= 8;
2674 /* Handle BR0 specially -- it may be getting stored permanently in
2675 some GR register. */
2676 if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
2678 reg = gen_rtx_REG (DImode, BR_REG (0));
2679 if (current_frame_info.reg_save_b0 != 0)
2681 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
2682 insn = emit_move_insn (alt_reg, reg);
2683 RTX_FRAME_RELATED_P (insn) = 1;
2685 /* Even if we're not going to generate an epilogue, we still
2686 need to save the register so that EH works. */
2687 if (! epilogue_p)
2688 emit_insn (gen_prologue_use (alt_reg));
2690 else
2692 alt_regno = next_scratch_gr_reg ();
2693 alt_reg = gen_rtx_REG (DImode, alt_regno);
2694 emit_move_insn (alt_reg, reg);
2695 do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
2696 cfa_off -= 8;
2700 /* Spill the rest of the BR registers. */
2701 for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
2702 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2704 alt_regno = next_scratch_gr_reg ();
2705 alt_reg = gen_rtx_REG (DImode, alt_regno);
2706 reg = gen_rtx_REG (DImode, regno);
2707 emit_move_insn (alt_reg, reg);
2708 do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
2709 cfa_off -= 8;
2712 /* Align the frame and spill all FR registers. */
2713 for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
2714 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2716 if (cfa_off & 15)
2717 abort ();
2718 reg = gen_rtx_REG (XFmode, regno);
2719 do_spill (gen_fr_spill_x, reg, cfa_off, reg);
2720 cfa_off -= 16;
2723 if (cfa_off != current_frame_info.spill_cfa_off)
2724 abort ();
2726 finish_spill_pointers ();
2729 /* Called after register allocation to add any instructions needed for the
2730 epilogue. Using an epilogue insn is favored compared to putting all of the
2731 instructions in output_function_prologue(), since it allows the scheduler
2732 to intermix instructions with the saves of the caller saved registers. In
2733 some cases, it might be necessary to emit a barrier instruction as the last
2734 insn to prevent such scheduling. */
2736 void
2737 ia64_expand_epilogue (int sibcall_p)
2739 rtx insn, reg, alt_reg, ar_unat_save_reg;
2740 int regno, alt_regno, cfa_off;
2742 ia64_compute_frame_size (get_frame_size ());
2744 /* If there is a frame pointer, then we use it instead of the stack
2745 pointer, so that the stack pointer does not need to be valid when
2746 the epilogue starts. See EXIT_IGNORE_STACK. */
2747 if (frame_pointer_needed)
2748 setup_spill_pointers (current_frame_info.n_spilled,
2749 hard_frame_pointer_rtx, 0);
2750 else
2751 setup_spill_pointers (current_frame_info.n_spilled, stack_pointer_rtx,
2752 current_frame_info.total_size);
2754 if (current_frame_info.total_size != 0)
2756 /* ??? At this point we must generate a magic insn that appears to
2757 modify the spill iterators and the frame pointer. This would
2758 allow the most scheduling freedom. For now, just hard stop. */
2759 emit_insn (gen_blockage ());
2762 /* Locate the bottom of the register save area. */
2763 cfa_off = (current_frame_info.spill_cfa_off
2764 + current_frame_info.spill_size
2765 + current_frame_info.extra_spill_size);
2767 /* Restore the predicate registers. */
2768 if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
2770 if (current_frame_info.reg_save_pr != 0)
2771 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_pr);
2772 else
2774 alt_regno = next_scratch_gr_reg ();
2775 alt_reg = gen_rtx_REG (DImode, alt_regno);
2776 do_restore (gen_movdi_x, alt_reg, cfa_off);
2777 cfa_off -= 8;
2779 reg = gen_rtx_REG (DImode, PR_REG (0));
2780 emit_move_insn (reg, alt_reg);
2783 /* Restore the application registers. */
2785 /* Load the saved unat from the stack, but do not restore it until
2786 after the GRs have been restored. */
2787 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
2789 if (current_frame_info.reg_save_ar_unat != 0)
2790 ar_unat_save_reg
2791 = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_unat);
2792 else
2794 alt_regno = next_scratch_gr_reg ();
2795 ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
2796 current_frame_info.gr_used_mask |= 1 << alt_regno;
2797 do_restore (gen_movdi_x, ar_unat_save_reg, cfa_off);
2798 cfa_off -= 8;
2801 else
2802 ar_unat_save_reg = NULL_RTX;
2804 if (current_frame_info.reg_save_ar_pfs != 0)
2806 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_pfs);
2807 reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
2808 emit_move_insn (reg, alt_reg);
2810 else if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
2812 alt_regno = next_scratch_gr_reg ();
2813 alt_reg = gen_rtx_REG (DImode, alt_regno);
2814 do_restore (gen_movdi_x, alt_reg, cfa_off);
2815 cfa_off -= 8;
2816 reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
2817 emit_move_insn (reg, alt_reg);
2820 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
2822 if (current_frame_info.reg_save_ar_lc != 0)
2823 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_lc);
2824 else
2826 alt_regno = next_scratch_gr_reg ();
2827 alt_reg = gen_rtx_REG (DImode, alt_regno);
2828 do_restore (gen_movdi_x, alt_reg, cfa_off);
2829 cfa_off -= 8;
2831 reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
2832 emit_move_insn (reg, alt_reg);
2835 /* We should now be at the base of the gr/br/fr spill area. */
2836 if (cfa_off != (current_frame_info.spill_cfa_off
2837 + current_frame_info.spill_size))
2838 abort ();
2840 /* The GP may be stored on the stack in the prologue, but it's
2841 never restored in the epilogue. Skip the stack slot. */
2842 if (TEST_HARD_REG_BIT (current_frame_info.mask, GR_REG (1)))
2843 cfa_off -= 8;
2845 /* Restore all general registers. */
2846 for (regno = GR_REG (2); regno <= GR_REG (31); ++regno)
2847 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2849 reg = gen_rtx_REG (DImode, regno);
2850 do_restore (gen_gr_restore, reg, cfa_off);
2851 cfa_off -= 8;
2854 /* Restore the branch registers. Handle B0 specially, as it may
2855 have gotten stored in some GR register. */
2856 if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
2858 if (current_frame_info.reg_save_b0 != 0)
2859 alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
2860 else
2862 alt_regno = next_scratch_gr_reg ();
2863 alt_reg = gen_rtx_REG (DImode, alt_regno);
2864 do_restore (gen_movdi_x, alt_reg, cfa_off);
2865 cfa_off -= 8;
2867 reg = gen_rtx_REG (DImode, BR_REG (0));
2868 emit_move_insn (reg, alt_reg);
2871 for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
2872 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2874 alt_regno = next_scratch_gr_reg ();
2875 alt_reg = gen_rtx_REG (DImode, alt_regno);
2876 do_restore (gen_movdi_x, alt_reg, cfa_off);
2877 cfa_off -= 8;
2878 reg = gen_rtx_REG (DImode, regno);
2879 emit_move_insn (reg, alt_reg);
2882 /* Restore floating point registers. */
2883 for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
2884 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
2886 if (cfa_off & 15)
2887 abort ();
2888 reg = gen_rtx_REG (XFmode, regno);
2889 do_restore (gen_fr_restore_x, reg, cfa_off);
2890 cfa_off -= 16;
2893 /* Restore ar.unat for real. */
2894 if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
2896 reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
2897 emit_move_insn (reg, ar_unat_save_reg);
2900 if (cfa_off != current_frame_info.spill_cfa_off)
2901 abort ();
2903 finish_spill_pointers ();
2905 if (current_frame_info.total_size || cfun->machine->ia64_eh_epilogue_sp)
2907 /* ??? At this point we must generate a magic insn that appears to
2908 modify the spill iterators, the stack pointer, and the frame
2909 pointer. This would allow the most scheduling freedom. For now,
2910 just hard stop. */
2911 emit_insn (gen_blockage ());
2914 if (cfun->machine->ia64_eh_epilogue_sp)
2915 emit_move_insn (stack_pointer_rtx, cfun->machine->ia64_eh_epilogue_sp);
2916 else if (frame_pointer_needed)
2918 insn = emit_move_insn (stack_pointer_rtx, hard_frame_pointer_rtx);
2919 RTX_FRAME_RELATED_P (insn) = 1;
2921 else if (current_frame_info.total_size)
2923 rtx offset, frame_size_rtx;
2925 frame_size_rtx = GEN_INT (current_frame_info.total_size);
2926 if (CONST_OK_FOR_I (current_frame_info.total_size))
2927 offset = frame_size_rtx;
2928 else
2930 regno = next_scratch_gr_reg ();
2931 offset = gen_rtx_REG (DImode, regno);
2932 emit_move_insn (offset, frame_size_rtx);
2935 insn = emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx,
2936 offset));
2938 RTX_FRAME_RELATED_P (insn) = 1;
2939 if (GET_CODE (offset) != CONST_INT)
2941 REG_NOTES (insn)
2942 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
2943 gen_rtx_SET (VOIDmode,
2944 stack_pointer_rtx,
2945 gen_rtx_PLUS (DImode,
2946 stack_pointer_rtx,
2947 frame_size_rtx)),
2948 REG_NOTES (insn));
2952 if (cfun->machine->ia64_eh_epilogue_bsp)
2953 emit_insn (gen_set_bsp (cfun->machine->ia64_eh_epilogue_bsp));
2955 if (! sibcall_p)
2956 emit_jump_insn (gen_return_internal (gen_rtx_REG (DImode, BR_REG (0))));
2957 else
2959 int fp = GR_REG (2);
2960 /* We need a throw away register here, r0 and r1 are reserved, so r2 is the
2961 first available call clobbered register. If there was a frame_pointer
2962 register, we may have swapped the names of r2 and HARD_FRAME_POINTER_REGNUM,
2963 so we have to make sure we're using the string "r2" when emitting
2964 the register name for the assembler. */
2965 if (current_frame_info.reg_fp && current_frame_info.reg_fp == GR_REG (2))
2966 fp = HARD_FRAME_POINTER_REGNUM;
2968 /* We must emit an alloc to force the input registers to become output
2969 registers. Otherwise, if the callee tries to pass its parameters
2970 through to another call without an intervening alloc, then these
2971 values get lost. */
2972 /* ??? We don't need to preserve all input registers. We only need to
2973 preserve those input registers used as arguments to the sibling call.
2974 It is unclear how to compute that number here. */
2975 if (current_frame_info.n_input_regs != 0)
2977 rtx n_inputs = GEN_INT (current_frame_info.n_input_regs);
2978 insn = emit_insn (gen_alloc (gen_rtx_REG (DImode, fp),
2979 const0_rtx, const0_rtx,
2980 n_inputs, const0_rtx));
2981 RTX_FRAME_RELATED_P (insn) = 1;
2986 /* Return 1 if br.ret can do all the work required to return from a
2987 function. */
2990 ia64_direct_return (void)
2992 if (reload_completed && ! frame_pointer_needed)
2994 ia64_compute_frame_size (get_frame_size ());
2996 return (current_frame_info.total_size == 0
2997 && current_frame_info.n_spilled == 0
2998 && current_frame_info.reg_save_b0 == 0
2999 && current_frame_info.reg_save_pr == 0
3000 && current_frame_info.reg_save_ar_pfs == 0
3001 && current_frame_info.reg_save_ar_unat == 0
3002 && current_frame_info.reg_save_ar_lc == 0);
3004 return 0;
3007 /* Return the magic cookie that we use to hold the return address
3008 during early compilation. */
3011 ia64_return_addr_rtx (HOST_WIDE_INT count, rtx frame ATTRIBUTE_UNUSED)
3013 if (count != 0)
3014 return NULL;
3015 return gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), UNSPEC_RET_ADDR);
3018 /* Split this value after reload, now that we know where the return
3019 address is saved. */
3021 void
3022 ia64_split_return_addr_rtx (rtx dest)
3024 rtx src;
3026 if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
3028 if (current_frame_info.reg_save_b0 != 0)
3029 src = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
3030 else
3032 HOST_WIDE_INT off;
3033 unsigned int regno;
3035 /* Compute offset from CFA for BR0. */
3036 /* ??? Must be kept in sync with ia64_expand_prologue. */
3037 off = (current_frame_info.spill_cfa_off
3038 + current_frame_info.spill_size);
3039 for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
3040 if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
3041 off -= 8;
3043 /* Convert CFA offset to a register based offset. */
3044 if (frame_pointer_needed)
3045 src = hard_frame_pointer_rtx;
3046 else
3048 src = stack_pointer_rtx;
3049 off += current_frame_info.total_size;
3052 /* Load address into scratch register. */
3053 if (CONST_OK_FOR_I (off))
3054 emit_insn (gen_adddi3 (dest, src, GEN_INT (off)));
3055 else
3057 emit_move_insn (dest, GEN_INT (off));
3058 emit_insn (gen_adddi3 (dest, src, dest));
3061 src = gen_rtx_MEM (Pmode, dest);
3064 else
3065 src = gen_rtx_REG (DImode, BR_REG (0));
3067 emit_move_insn (dest, src);
3071 ia64_hard_regno_rename_ok (int from, int to)
3073 /* Don't clobber any of the registers we reserved for the prologue. */
3074 if (to == current_frame_info.reg_fp
3075 || to == current_frame_info.reg_save_b0
3076 || to == current_frame_info.reg_save_pr
3077 || to == current_frame_info.reg_save_ar_pfs
3078 || to == current_frame_info.reg_save_ar_unat
3079 || to == current_frame_info.reg_save_ar_lc)
3080 return 0;
3082 if (from == current_frame_info.reg_fp
3083 || from == current_frame_info.reg_save_b0
3084 || from == current_frame_info.reg_save_pr
3085 || from == current_frame_info.reg_save_ar_pfs
3086 || from == current_frame_info.reg_save_ar_unat
3087 || from == current_frame_info.reg_save_ar_lc)
3088 return 0;
3090 /* Don't use output registers outside the register frame. */
3091 if (OUT_REGNO_P (to) && to >= OUT_REG (current_frame_info.n_output_regs))
3092 return 0;
3094 /* Retain even/oddness on predicate register pairs. */
3095 if (PR_REGNO_P (from) && PR_REGNO_P (to))
3096 return (from & 1) == (to & 1);
3098 return 1;
3101 /* Target hook for assembling integer objects. Handle word-sized
3102 aligned objects and detect the cases when @fptr is needed. */
3104 static bool
3105 ia64_assemble_integer (rtx x, unsigned int size, int aligned_p)
3107 if (size == POINTER_SIZE / BITS_PER_UNIT
3108 && !(TARGET_NO_PIC || TARGET_AUTO_PIC)
3109 && GET_CODE (x) == SYMBOL_REF
3110 && SYMBOL_REF_FUNCTION_P (x))
3112 static const char * const directive[2][2] = {
3113 /* 64-bit pointer */ /* 32-bit pointer */
3114 { "\tdata8.ua\t@fptr(", "\tdata4.ua\t@fptr("}, /* unaligned */
3115 { "\tdata8\t@fptr(", "\tdata4\t@fptr("} /* aligned */
3117 fputs (directive[(aligned_p != 0)][POINTER_SIZE == 32], asm_out_file);
3118 output_addr_const (asm_out_file, x);
3119 fputs (")\n", asm_out_file);
3120 return true;
3122 return default_assemble_integer (x, size, aligned_p);
3125 /* Emit the function prologue. */
3127 static void
3128 ia64_output_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
3130 int mask, grsave, grsave_prev;
3132 if (current_frame_info.need_regstk)
3133 fprintf (file, "\t.regstk %d, %d, %d, %d\n",
3134 current_frame_info.n_input_regs,
3135 current_frame_info.n_local_regs,
3136 current_frame_info.n_output_regs,
3137 current_frame_info.n_rotate_regs);
3139 if (!flag_unwind_tables && (!flag_exceptions || USING_SJLJ_EXCEPTIONS))
3140 return;
3142 /* Emit the .prologue directive. */
3144 mask = 0;
3145 grsave = grsave_prev = 0;
3146 if (current_frame_info.reg_save_b0 != 0)
3148 mask |= 8;
3149 grsave = grsave_prev = current_frame_info.reg_save_b0;
3151 if (current_frame_info.reg_save_ar_pfs != 0
3152 && (grsave_prev == 0
3153 || current_frame_info.reg_save_ar_pfs == grsave_prev + 1))
3155 mask |= 4;
3156 if (grsave_prev == 0)
3157 grsave = current_frame_info.reg_save_ar_pfs;
3158 grsave_prev = current_frame_info.reg_save_ar_pfs;
3160 if (current_frame_info.reg_fp != 0
3161 && (grsave_prev == 0
3162 || current_frame_info.reg_fp == grsave_prev + 1))
3164 mask |= 2;
3165 if (grsave_prev == 0)
3166 grsave = HARD_FRAME_POINTER_REGNUM;
3167 grsave_prev = current_frame_info.reg_fp;
3169 if (current_frame_info.reg_save_pr != 0
3170 && (grsave_prev == 0
3171 || current_frame_info.reg_save_pr == grsave_prev + 1))
3173 mask |= 1;
3174 if (grsave_prev == 0)
3175 grsave = current_frame_info.reg_save_pr;
3178 if (mask && TARGET_GNU_AS)
3179 fprintf (file, "\t.prologue %d, %d\n", mask,
3180 ia64_dbx_register_number (grsave));
3181 else
3182 fputs ("\t.prologue\n", file);
3184 /* Emit a .spill directive, if necessary, to relocate the base of
3185 the register spill area. */
3186 if (current_frame_info.spill_cfa_off != -16)
3187 fprintf (file, "\t.spill %ld\n",
3188 (long) (current_frame_info.spill_cfa_off
3189 + current_frame_info.spill_size));
3192 /* Emit the .body directive at the scheduled end of the prologue. */
3194 static void
3195 ia64_output_function_end_prologue (FILE *file)
3197 if (!flag_unwind_tables && (!flag_exceptions || USING_SJLJ_EXCEPTIONS))
3198 return;
3200 fputs ("\t.body\n", file);
3203 /* Emit the function epilogue. */
3205 static void
3206 ia64_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
3207 HOST_WIDE_INT size ATTRIBUTE_UNUSED)
3209 int i;
3211 if (current_frame_info.reg_fp)
3213 const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
3214 reg_names[HARD_FRAME_POINTER_REGNUM]
3215 = reg_names[current_frame_info.reg_fp];
3216 reg_names[current_frame_info.reg_fp] = tmp;
3218 if (! TARGET_REG_NAMES)
3220 for (i = 0; i < current_frame_info.n_input_regs; i++)
3221 reg_names[IN_REG (i)] = ia64_input_reg_names[i];
3222 for (i = 0; i < current_frame_info.n_local_regs; i++)
3223 reg_names[LOC_REG (i)] = ia64_local_reg_names[i];
3224 for (i = 0; i < current_frame_info.n_output_regs; i++)
3225 reg_names[OUT_REG (i)] = ia64_output_reg_names[i];
3228 current_frame_info.initialized = 0;
3232 ia64_dbx_register_number (int regno)
3234 /* In ia64_expand_prologue we quite literally renamed the frame pointer
3235 from its home at loc79 to something inside the register frame. We
3236 must perform the same renumbering here for the debug info. */
3237 if (current_frame_info.reg_fp)
3239 if (regno == HARD_FRAME_POINTER_REGNUM)
3240 regno = current_frame_info.reg_fp;
3241 else if (regno == current_frame_info.reg_fp)
3242 regno = HARD_FRAME_POINTER_REGNUM;
3245 if (IN_REGNO_P (regno))
3246 return 32 + regno - IN_REG (0);
3247 else if (LOC_REGNO_P (regno))
3248 return 32 + current_frame_info.n_input_regs + regno - LOC_REG (0);
3249 else if (OUT_REGNO_P (regno))
3250 return (32 + current_frame_info.n_input_regs
3251 + current_frame_info.n_local_regs + regno - OUT_REG (0));
3252 else
3253 return regno;
3256 void
3257 ia64_initialize_trampoline (rtx addr, rtx fnaddr, rtx static_chain)
3259 rtx addr_reg, eight = GEN_INT (8);
3261 /* The Intel assembler requires that the global __ia64_trampoline symbol
3262 be declared explicitly */
3263 if (!TARGET_GNU_AS)
3265 static bool declared_ia64_trampoline = false;
3267 if (!declared_ia64_trampoline)
3269 declared_ia64_trampoline = true;
3270 (*targetm.asm_out.globalize_label) (asm_out_file,
3271 "__ia64_trampoline");
3275 /* Make sure addresses are Pmode even if we are in ILP32 mode. */
3276 addr = convert_memory_address (Pmode, addr);
3277 fnaddr = convert_memory_address (Pmode, fnaddr);
3278 static_chain = convert_memory_address (Pmode, static_chain);
3280 /* Load up our iterator. */
3281 addr_reg = gen_reg_rtx (Pmode);
3282 emit_move_insn (addr_reg, addr);
3284 /* The first two words are the fake descriptor:
3285 __ia64_trampoline, ADDR+16. */
3286 emit_move_insn (gen_rtx_MEM (Pmode, addr_reg),
3287 gen_rtx_SYMBOL_REF (Pmode, "__ia64_trampoline"));
3288 emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
3290 emit_move_insn (gen_rtx_MEM (Pmode, addr_reg),
3291 copy_to_reg (plus_constant (addr, 16)));
3292 emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
3294 /* The third word is the target descriptor. */
3295 emit_move_insn (gen_rtx_MEM (Pmode, addr_reg), fnaddr);
3296 emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));
3298 /* The fourth word is the static chain. */
3299 emit_move_insn (gen_rtx_MEM (Pmode, addr_reg), static_chain);
3302 /* Do any needed setup for a variadic function. CUM has not been updated
3303 for the last named argument which has type TYPE and mode MODE.
3305 We generate the actual spill instructions during prologue generation. */
3307 static void
3308 ia64_setup_incoming_varargs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
3309 tree type, int * pretend_size,
3310 int second_time ATTRIBUTE_UNUSED)
3312 CUMULATIVE_ARGS next_cum = *cum;
3314 /* Skip the current argument. */
3315 ia64_function_arg_advance (&next_cum, mode, type, 1);
3317 if (next_cum.words < MAX_ARGUMENT_SLOTS)
3319 int n = MAX_ARGUMENT_SLOTS - next_cum.words;
3320 *pretend_size = n * UNITS_PER_WORD;
3321 cfun->machine->n_varargs = n;
3325 /* Check whether TYPE is a homogeneous floating point aggregate. If
3326 it is, return the mode of the floating point type that appears
3327 in all leafs. If it is not, return VOIDmode.
3329 An aggregate is a homogeneous floating point aggregate is if all
3330 fields/elements in it have the same floating point type (e.g,
3331 SFmode). 128-bit quad-precision floats are excluded.
3333 Variable sized aggregates should never arrive here, since we should
3334 have already decided to pass them by reference. Top-level zero-sized
3335 aggregates are excluded because our parallels crash the middle-end. */
3337 static enum machine_mode
3338 hfa_element_mode (tree type, bool nested)
3340 enum machine_mode element_mode = VOIDmode;
3341 enum machine_mode mode;
3342 enum tree_code code = TREE_CODE (type);
3343 int know_element_mode = 0;
3344 tree t;
3346 if (!nested && (!TYPE_SIZE (type) || integer_zerop (TYPE_SIZE (type))))
3347 return VOIDmode;
3349 switch (code)
3351 case VOID_TYPE: case INTEGER_TYPE: case ENUMERAL_TYPE:
3352 case BOOLEAN_TYPE: case CHAR_TYPE: case POINTER_TYPE:
3353 case OFFSET_TYPE: case REFERENCE_TYPE: case METHOD_TYPE:
3354 case FILE_TYPE: case LANG_TYPE: case FUNCTION_TYPE:
3355 return VOIDmode;
3357 /* Fortran complex types are supposed to be HFAs, so we need to handle
3358 gcc's COMPLEX_TYPEs as HFAs. We need to exclude the integral complex
3359 types though. */
3360 case COMPLEX_TYPE:
3361 if (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_COMPLEX_FLOAT
3362 && TYPE_MODE (type) != TCmode)
3363 return GET_MODE_INNER (TYPE_MODE (type));
3364 else
3365 return VOIDmode;
3367 case REAL_TYPE:
3368 /* We want to return VOIDmode for raw REAL_TYPEs, but the actual
3369 mode if this is contained within an aggregate. */
3370 if (nested && TYPE_MODE (type) != TFmode)
3371 return TYPE_MODE (type);
3372 else
3373 return VOIDmode;
3375 case ARRAY_TYPE:
3376 return hfa_element_mode (TREE_TYPE (type), 1);
3378 case RECORD_TYPE:
3379 case UNION_TYPE:
3380 case QUAL_UNION_TYPE:
3381 for (t = TYPE_FIELDS (type); t; t = TREE_CHAIN (t))
3383 if (TREE_CODE (t) != FIELD_DECL)
3384 continue;
3386 mode = hfa_element_mode (TREE_TYPE (t), 1);
3387 if (know_element_mode)
3389 if (mode != element_mode)
3390 return VOIDmode;
3392 else if (GET_MODE_CLASS (mode) != MODE_FLOAT)
3393 return VOIDmode;
3394 else
3396 know_element_mode = 1;
3397 element_mode = mode;
3400 return element_mode;
3402 default:
3403 /* If we reach here, we probably have some front-end specific type
3404 that the backend doesn't know about. This can happen via the
3405 aggregate_value_p call in init_function_start. All we can do is
3406 ignore unknown tree types. */
3407 return VOIDmode;
3410 return VOIDmode;
3413 /* Return the number of words required to hold a quantity of TYPE and MODE
3414 when passed as an argument. */
3415 static int
3416 ia64_function_arg_words (tree type, enum machine_mode mode)
3418 int words;
3420 if (mode == BLKmode)
3421 words = int_size_in_bytes (type);
3422 else
3423 words = GET_MODE_SIZE (mode);
3425 return (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD; /* round up */
3428 /* Return the number of registers that should be skipped so the current
3429 argument (described by TYPE and WORDS) will be properly aligned.
3431 Integer and float arguments larger than 8 bytes start at the next
3432 even boundary. Aggregates larger than 8 bytes start at the next
3433 even boundary if the aggregate has 16 byte alignment. Note that
3434 in the 32-bit ABI, TImode and TFmode have only 8-byte alignment
3435 but are still to be aligned in registers.
3437 ??? The ABI does not specify how to handle aggregates with
3438 alignment from 9 to 15 bytes, or greater than 16. We handle them
3439 all as if they had 16 byte alignment. Such aggregates can occur
3440 only if gcc extensions are used. */
3441 static int
3442 ia64_function_arg_offset (CUMULATIVE_ARGS *cum, tree type, int words)
3444 if ((cum->words & 1) == 0)
3445 return 0;
3447 if (type
3448 && TREE_CODE (type) != INTEGER_TYPE
3449 && TREE_CODE (type) != REAL_TYPE)
3450 return TYPE_ALIGN (type) > 8 * BITS_PER_UNIT;
3451 else
3452 return words > 1;
3455 /* Return rtx for register where argument is passed, or zero if it is passed
3456 on the stack. */
3457 /* ??? 128-bit quad-precision floats are always passed in general
3458 registers. */
3461 ia64_function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
3462 int named, int incoming)
3464 int basereg = (incoming ? GR_ARG_FIRST : AR_ARG_FIRST);
3465 int words = ia64_function_arg_words (type, mode);
3466 int offset = ia64_function_arg_offset (cum, type, words);
3467 enum machine_mode hfa_mode = VOIDmode;
3469 /* If all argument slots are used, then it must go on the stack. */
3470 if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
3471 return 0;
3473 /* Check for and handle homogeneous FP aggregates. */
3474 if (type)
3475 hfa_mode = hfa_element_mode (type, 0);
3477 /* Unnamed prototyped hfas are passed as usual. Named prototyped hfas
3478 and unprototyped hfas are passed specially. */
3479 if (hfa_mode != VOIDmode && (! cum->prototype || named))
3481 rtx loc[16];
3482 int i = 0;
3483 int fp_regs = cum->fp_regs;
3484 int int_regs = cum->words + offset;
3485 int hfa_size = GET_MODE_SIZE (hfa_mode);
3486 int byte_size;
3487 int args_byte_size;
3489 /* If prototyped, pass it in FR regs then GR regs.
3490 If not prototyped, pass it in both FR and GR regs.
3492 If this is an SFmode aggregate, then it is possible to run out of
3493 FR regs while GR regs are still left. In that case, we pass the
3494 remaining part in the GR regs. */
3496 /* Fill the FP regs. We do this always. We stop if we reach the end
3497 of the argument, the last FP register, or the last argument slot. */
3499 byte_size = ((mode == BLKmode)
3500 ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
3501 args_byte_size = int_regs * UNITS_PER_WORD;
3502 offset = 0;
3503 for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
3504 && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD)); i++)
3506 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
3507 gen_rtx_REG (hfa_mode, (FR_ARG_FIRST
3508 + fp_regs)),
3509 GEN_INT (offset));
3510 offset += hfa_size;
3511 args_byte_size += hfa_size;
3512 fp_regs++;
3515 /* If no prototype, then the whole thing must go in GR regs. */
3516 if (! cum->prototype)
3517 offset = 0;
3518 /* If this is an SFmode aggregate, then we might have some left over
3519 that needs to go in GR regs. */
3520 else if (byte_size != offset)
3521 int_regs += offset / UNITS_PER_WORD;
3523 /* Fill in the GR regs. We must use DImode here, not the hfa mode. */
3525 for (; offset < byte_size && int_regs < MAX_ARGUMENT_SLOTS; i++)
3527 enum machine_mode gr_mode = DImode;
3528 unsigned int gr_size;
3530 /* If we have an odd 4 byte hunk because we ran out of FR regs,
3531 then this goes in a GR reg left adjusted/little endian, right
3532 adjusted/big endian. */
3533 /* ??? Currently this is handled wrong, because 4-byte hunks are
3534 always right adjusted/little endian. */
3535 if (offset & 0x4)
3536 gr_mode = SImode;
3537 /* If we have an even 4 byte hunk because the aggregate is a
3538 multiple of 4 bytes in size, then this goes in a GR reg right
3539 adjusted/little endian. */
3540 else if (byte_size - offset == 4)
3541 gr_mode = SImode;
3543 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
3544 gen_rtx_REG (gr_mode, (basereg
3545 + int_regs)),
3546 GEN_INT (offset));
3548 gr_size = GET_MODE_SIZE (gr_mode);
3549 offset += gr_size;
3550 if (gr_size == UNITS_PER_WORD
3551 || (gr_size < UNITS_PER_WORD && offset % UNITS_PER_WORD == 0))
3552 int_regs++;
3553 else if (gr_size > UNITS_PER_WORD)
3554 int_regs += gr_size / UNITS_PER_WORD;
3556 return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
3559 /* Integral and aggregates go in general registers. If we have run out of
3560 FR registers, then FP values must also go in general registers. This can
3561 happen when we have a SFmode HFA. */
3562 else if (mode == TFmode || mode == TCmode
3563 || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
3565 int byte_size = ((mode == BLKmode)
3566 ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
3567 if (BYTES_BIG_ENDIAN
3568 && (mode == BLKmode || (type && AGGREGATE_TYPE_P (type)))
3569 && byte_size < UNITS_PER_WORD
3570 && byte_size > 0)
3572 rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
3573 gen_rtx_REG (DImode,
3574 (basereg + cum->words
3575 + offset)),
3576 const0_rtx);
3577 return gen_rtx_PARALLEL (mode, gen_rtvec (1, gr_reg));
3579 else
3580 return gen_rtx_REG (mode, basereg + cum->words + offset);
3584 /* If there is a prototype, then FP values go in a FR register when
3585 named, and in a GR register when unnamed. */
3586 else if (cum->prototype)
3588 if (named)
3589 return gen_rtx_REG (mode, FR_ARG_FIRST + cum->fp_regs);
3590 /* In big-endian mode, an anonymous SFmode value must be represented
3591 as (parallel:SF [(expr_list (reg:DI n) (const_int 0))]) to force
3592 the value into the high half of the general register. */
3593 else if (BYTES_BIG_ENDIAN && mode == SFmode)
3594 return gen_rtx_PARALLEL (mode,
3595 gen_rtvec (1,
3596 gen_rtx_EXPR_LIST (VOIDmode,
3597 gen_rtx_REG (DImode, basereg + cum->words + offset),
3598 const0_rtx)));
3599 else
3600 return gen_rtx_REG (mode, basereg + cum->words + offset);
3602 /* If there is no prototype, then FP values go in both FR and GR
3603 registers. */
3604 else
3606 /* See comment above. */
3607 enum machine_mode inner_mode =
3608 (BYTES_BIG_ENDIAN && mode == SFmode) ? DImode : mode;
3610 rtx fp_reg = gen_rtx_EXPR_LIST (VOIDmode,
3611 gen_rtx_REG (mode, (FR_ARG_FIRST
3612 + cum->fp_regs)),
3613 const0_rtx);
3614 rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
3615 gen_rtx_REG (inner_mode,
3616 (basereg + cum->words
3617 + offset)),
3618 const0_rtx);
3620 return gen_rtx_PARALLEL (mode, gen_rtvec (2, fp_reg, gr_reg));
3624 /* Return number of bytes, at the beginning of the argument, that must be
3625 put in registers. 0 is the argument is entirely in registers or entirely
3626 in memory. */
3628 static int
3629 ia64_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
3630 tree type, bool named ATTRIBUTE_UNUSED)
3632 int words = ia64_function_arg_words (type, mode);
3633 int offset = ia64_function_arg_offset (cum, type, words);
3635 /* If all argument slots are used, then it must go on the stack. */
3636 if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
3637 return 0;
3639 /* It doesn't matter whether the argument goes in FR or GR regs. If
3640 it fits within the 8 argument slots, then it goes entirely in
3641 registers. If it extends past the last argument slot, then the rest
3642 goes on the stack. */
3644 if (words + cum->words + offset <= MAX_ARGUMENT_SLOTS)
3645 return 0;
3647 return (MAX_ARGUMENT_SLOTS - cum->words - offset) * UNITS_PER_WORD;
3650 /* Update CUM to point after this argument. This is patterned after
3651 ia64_function_arg. */
3653 void
3654 ia64_function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode,
3655 tree type, int named)
3657 int words = ia64_function_arg_words (type, mode);
3658 int offset = ia64_function_arg_offset (cum, type, words);
3659 enum machine_mode hfa_mode = VOIDmode;
3661 /* If all arg slots are already full, then there is nothing to do. */
3662 if (cum->words >= MAX_ARGUMENT_SLOTS)
3663 return;
3665 cum->words += words + offset;
3667 /* Check for and handle homogeneous FP aggregates. */
3668 if (type)
3669 hfa_mode = hfa_element_mode (type, 0);
3671 /* Unnamed prototyped hfas are passed as usual. Named prototyped hfas
3672 and unprototyped hfas are passed specially. */
3673 if (hfa_mode != VOIDmode && (! cum->prototype || named))
3675 int fp_regs = cum->fp_regs;
3676 /* This is the original value of cum->words + offset. */
3677 int int_regs = cum->words - words;
3678 int hfa_size = GET_MODE_SIZE (hfa_mode);
3679 int byte_size;
3680 int args_byte_size;
3682 /* If prototyped, pass it in FR regs then GR regs.
3683 If not prototyped, pass it in both FR and GR regs.
3685 If this is an SFmode aggregate, then it is possible to run out of
3686 FR regs while GR regs are still left. In that case, we pass the
3687 remaining part in the GR regs. */
3689 /* Fill the FP regs. We do this always. We stop if we reach the end
3690 of the argument, the last FP register, or the last argument slot. */
3692 byte_size = ((mode == BLKmode)
3693 ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
3694 args_byte_size = int_regs * UNITS_PER_WORD;
3695 offset = 0;
3696 for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
3697 && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD));)
3699 offset += hfa_size;
3700 args_byte_size += hfa_size;
3701 fp_regs++;
3704 cum->fp_regs = fp_regs;
3707 /* Integral and aggregates go in general registers. So do TFmode FP values.
3708 If we have run out of FR registers, then other FP values must also go in
3709 general registers. This can happen when we have a SFmode HFA. */
3710 else if (mode == TFmode || mode == TCmode
3711 || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
3712 cum->int_regs = cum->words;
3714 /* If there is a prototype, then FP values go in a FR register when
3715 named, and in a GR register when unnamed. */
3716 else if (cum->prototype)
3718 if (! named)
3719 cum->int_regs = cum->words;
3720 else
3721 /* ??? Complex types should not reach here. */
3722 cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
3724 /* If there is no prototype, then FP values go in both FR and GR
3725 registers. */
3726 else
3728 /* ??? Complex types should not reach here. */
3729 cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
3730 cum->int_regs = cum->words;
3734 /* Arguments with alignment larger than 8 bytes start at the next even
3735 boundary. On ILP32 HPUX, TFmode arguments start on next even boundary
3736 even though their normal alignment is 8 bytes. See ia64_function_arg. */
3739 ia64_function_arg_boundary (enum machine_mode mode, tree type)
3742 if (mode == TFmode && TARGET_HPUX && TARGET_ILP32)
3743 return PARM_BOUNDARY * 2;
3745 if (type)
3747 if (TYPE_ALIGN (type) > PARM_BOUNDARY)
3748 return PARM_BOUNDARY * 2;
3749 else
3750 return PARM_BOUNDARY;
3753 if (GET_MODE_BITSIZE (mode) > PARM_BOUNDARY)
3754 return PARM_BOUNDARY * 2;
3755 else
3756 return PARM_BOUNDARY;
3759 /* Variable sized types are passed by reference. */
3760 /* ??? At present this is a GCC extension to the IA-64 ABI. */
3762 static bool
3763 ia64_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
3764 enum machine_mode mode ATTRIBUTE_UNUSED,
3765 tree type, bool named ATTRIBUTE_UNUSED)
3767 return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
3770 /* True if it is OK to do sibling call optimization for the specified
3771 call expression EXP. DECL will be the called function, or NULL if
3772 this is an indirect call. */
3773 static bool
3774 ia64_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
3776 /* We can't perform a sibcall if the current function has the syscall_linkage
3777 attribute. */
3778 if (lookup_attribute ("syscall_linkage",
3779 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
3780 return false;
3782 /* We must always return with our current GP. This means we can
3783 only sibcall to functions defined in the current module. */
3784 return decl && (*targetm.binds_local_p) (decl);
3788 /* Implement va_arg. */
3790 static tree
3791 ia64_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
3793 /* Variable sized types are passed by reference. */
3794 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
3796 tree ptrtype = build_pointer_type (type);
3797 tree addr = std_gimplify_va_arg_expr (valist, ptrtype, pre_p, post_p);
3798 return build_va_arg_indirect_ref (addr);
3801 /* Aggregate arguments with alignment larger than 8 bytes start at
3802 the next even boundary. Integer and floating point arguments
3803 do so if they are larger than 8 bytes, whether or not they are
3804 also aligned larger than 8 bytes. */
3805 if ((TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == INTEGER_TYPE)
3806 ? int_size_in_bytes (type) > 8 : TYPE_ALIGN (type) > 8 * BITS_PER_UNIT)
3808 tree t = build (PLUS_EXPR, TREE_TYPE (valist), valist,
3809 build_int_cst (NULL_TREE, 2 * UNITS_PER_WORD - 1));
3810 t = build (BIT_AND_EXPR, TREE_TYPE (t), t,
3811 build_int_cst (NULL_TREE, -2 * UNITS_PER_WORD));
3812 t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t);
3813 gimplify_and_add (t, pre_p);
3816 return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
3819 /* Return 1 if function return value returned in memory. Return 0 if it is
3820 in a register. */
3822 static bool
3823 ia64_return_in_memory (tree valtype, tree fntype ATTRIBUTE_UNUSED)
3825 enum machine_mode mode;
3826 enum machine_mode hfa_mode;
3827 HOST_WIDE_INT byte_size;
3829 mode = TYPE_MODE (valtype);
3830 byte_size = GET_MODE_SIZE (mode);
3831 if (mode == BLKmode)
3833 byte_size = int_size_in_bytes (valtype);
3834 if (byte_size < 0)
3835 return true;
3838 /* Hfa's with up to 8 elements are returned in the FP argument registers. */
3840 hfa_mode = hfa_element_mode (valtype, 0);
3841 if (hfa_mode != VOIDmode)
3843 int hfa_size = GET_MODE_SIZE (hfa_mode);
3845 if (byte_size / hfa_size > MAX_ARGUMENT_SLOTS)
3846 return true;
3847 else
3848 return false;
3850 else if (byte_size > UNITS_PER_WORD * MAX_INT_RETURN_SLOTS)
3851 return true;
3852 else
3853 return false;
3856 /* Return rtx for register that holds the function return value. */
3859 ia64_function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
3861 enum machine_mode mode;
3862 enum machine_mode hfa_mode;
3864 mode = TYPE_MODE (valtype);
3865 hfa_mode = hfa_element_mode (valtype, 0);
3867 if (hfa_mode != VOIDmode)
3869 rtx loc[8];
3870 int i;
3871 int hfa_size;
3872 int byte_size;
3873 int offset;
3875 hfa_size = GET_MODE_SIZE (hfa_mode);
3876 byte_size = ((mode == BLKmode)
3877 ? int_size_in_bytes (valtype) : GET_MODE_SIZE (mode));
3878 offset = 0;
3879 for (i = 0; offset < byte_size; i++)
3881 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
3882 gen_rtx_REG (hfa_mode, FR_ARG_FIRST + i),
3883 GEN_INT (offset));
3884 offset += hfa_size;
3886 return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
3888 else if (FLOAT_TYPE_P (valtype) && mode != TFmode && mode != TCmode)
3889 return gen_rtx_REG (mode, FR_ARG_FIRST);
3890 else
3892 bool need_parallel = false;
3894 /* In big-endian mode, we need to manage the layout of aggregates
3895 in the registers so that we get the bits properly aligned in
3896 the highpart of the registers. */
3897 if (BYTES_BIG_ENDIAN
3898 && (mode == BLKmode || (valtype && AGGREGATE_TYPE_P (valtype))))
3899 need_parallel = true;
3901 /* Something like struct S { long double x; char a[0] } is not an
3902 HFA structure, and therefore doesn't go in fp registers. But
3903 the middle-end will give it XFmode anyway, and XFmode values
3904 don't normally fit in integer registers. So we need to smuggle
3905 the value inside a parallel. */
3906 else if (mode == XFmode || mode == XCmode)
3907 need_parallel = true;
3909 if (need_parallel)
3911 rtx loc[8];
3912 int offset;
3913 int bytesize;
3914 int i;
3916 offset = 0;
3917 bytesize = int_size_in_bytes (valtype);
3918 for (i = 0; offset < bytesize; i++)
3920 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
3921 gen_rtx_REG (DImode,
3922 GR_RET_FIRST + i),
3923 GEN_INT (offset));
3924 offset += UNITS_PER_WORD;
3926 return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
3929 return gen_rtx_REG (mode, GR_RET_FIRST);
3933 /* This is called from dwarf2out.c via ASM_OUTPUT_DWARF_DTPREL.
3934 We need to emit DTP-relative relocations. */
3936 void
3937 ia64_output_dwarf_dtprel (FILE *file, int size, rtx x)
3939 if (size != 8)
3940 abort ();
3941 fputs ("\tdata8.ua\t@dtprel(", file);
3942 output_addr_const (file, x);
3943 fputs (")", file);
3946 /* Print a memory address as an operand to reference that memory location. */
3948 /* ??? Do we need this? It gets used only for 'a' operands. We could perhaps
3949 also call this from ia64_print_operand for memory addresses. */
3951 void
3952 ia64_print_operand_address (FILE * stream ATTRIBUTE_UNUSED,
3953 rtx address ATTRIBUTE_UNUSED)
3957 /* Print an operand to an assembler instruction.
3958 C Swap and print a comparison operator.
3959 D Print an FP comparison operator.
3960 E Print 32 - constant, for SImode shifts as extract.
3961 e Print 64 - constant, for DImode rotates.
3962 F A floating point constant 0.0 emitted as f0, or 1.0 emitted as f1, or
3963 a floating point register emitted normally.
3964 I Invert a predicate register by adding 1.
3965 J Select the proper predicate register for a condition.
3966 j Select the inverse predicate register for a condition.
3967 O Append .acq for volatile load.
3968 P Postincrement of a MEM.
3969 Q Append .rel for volatile store.
3970 S Shift amount for shladd instruction.
3971 T Print an 8-bit sign extended number (K) as a 32-bit unsigned number
3972 for Intel assembler.
3973 U Print an 8-bit sign extended number (K) as a 64-bit unsigned number
3974 for Intel assembler.
3975 r Print register name, or constant 0 as r0. HP compatibility for
3976 Linux kernel.
3977 v Print vector constant value as an 8-byte integer value. */
3979 void
3980 ia64_print_operand (FILE * file, rtx x, int code)
3982 const char *str;
3984 switch (code)
3986 case 0:
3987 /* Handled below. */
3988 break;
3990 case 'C':
3992 enum rtx_code c = swap_condition (GET_CODE (x));
3993 fputs (GET_RTX_NAME (c), file);
3994 return;
3997 case 'D':
3998 switch (GET_CODE (x))
4000 case NE:
4001 str = "neq";
4002 break;
4003 case UNORDERED:
4004 str = "unord";
4005 break;
4006 case ORDERED:
4007 str = "ord";
4008 break;
4009 default:
4010 str = GET_RTX_NAME (GET_CODE (x));
4011 break;
4013 fputs (str, file);
4014 return;
4016 case 'E':
4017 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - INTVAL (x));
4018 return;
4020 case 'e':
4021 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - INTVAL (x));
4022 return;
4024 case 'F':
4025 if (x == CONST0_RTX (GET_MODE (x)))
4026 str = reg_names [FR_REG (0)];
4027 else if (x == CONST1_RTX (GET_MODE (x)))
4028 str = reg_names [FR_REG (1)];
4029 else if (GET_CODE (x) == REG)
4030 str = reg_names [REGNO (x)];
4031 else
4032 abort ();
4033 fputs (str, file);
4034 return;
4036 case 'I':
4037 fputs (reg_names [REGNO (x) + 1], file);
4038 return;
4040 case 'J':
4041 case 'j':
4043 unsigned int regno = REGNO (XEXP (x, 0));
4044 if (GET_CODE (x) == EQ)
4045 regno += 1;
4046 if (code == 'j')
4047 regno ^= 1;
4048 fputs (reg_names [regno], file);
4050 return;
4052 case 'O':
4053 if (MEM_VOLATILE_P (x))
4054 fputs(".acq", file);
4055 return;
4057 case 'P':
4059 HOST_WIDE_INT value;
4061 switch (GET_CODE (XEXP (x, 0)))
4063 default:
4064 return;
4066 case POST_MODIFY:
4067 x = XEXP (XEXP (XEXP (x, 0), 1), 1);
4068 if (GET_CODE (x) == CONST_INT)
4069 value = INTVAL (x);
4070 else if (GET_CODE (x) == REG)
4072 fprintf (file, ", %s", reg_names[REGNO (x)]);
4073 return;
4075 else
4076 abort ();
4077 break;
4079 case POST_INC:
4080 value = GET_MODE_SIZE (GET_MODE (x));
4081 break;
4083 case POST_DEC:
4084 value = - (HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (x));
4085 break;
4088 fprintf (file, ", " HOST_WIDE_INT_PRINT_DEC, value);
4089 return;
4092 case 'Q':
4093 if (MEM_VOLATILE_P (x))
4094 fputs(".rel", file);
4095 return;
4097 case 'S':
4098 fprintf (file, "%d", exact_log2 (INTVAL (x)));
4099 return;
4101 case 'T':
4102 if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
4104 fprintf (file, "0x%x", (int) INTVAL (x) & 0xffffffff);
4105 return;
4107 break;
4109 case 'U':
4110 if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
4112 const char *prefix = "0x";
4113 if (INTVAL (x) & 0x80000000)
4115 fprintf (file, "0xffffffff");
4116 prefix = "";
4118 fprintf (file, "%s%x", prefix, (int) INTVAL (x) & 0xffffffff);
4119 return;
4121 break;
4123 case 'r':
4124 /* If this operand is the constant zero, write it as register zero.
4125 Any register, zero, or CONST_INT value is OK here. */
4126 if (GET_CODE (x) == REG)
4127 fputs (reg_names[REGNO (x)], file);
4128 else if (x == CONST0_RTX (GET_MODE (x)))
4129 fputs ("r0", file);
4130 else if (GET_CODE (x) == CONST_INT)
4131 output_addr_const (file, x);
4132 else
4133 output_operand_lossage ("invalid %%r value");
4134 return;
4136 case 'v':
4137 gcc_assert (GET_CODE (x) == CONST_VECTOR);
4138 x = simplify_subreg (DImode, x, GET_MODE (x), 0);
4139 break;
4141 case '+':
4143 const char *which;
4145 /* For conditional branches, returns or calls, substitute
4146 sptk, dptk, dpnt, or spnt for %s. */
4147 x = find_reg_note (current_output_insn, REG_BR_PROB, 0);
4148 if (x)
4150 int pred_val = INTVAL (XEXP (x, 0));
4152 /* Guess top and bottom 10% statically predicted. */
4153 if (pred_val < REG_BR_PROB_BASE / 50)
4154 which = ".spnt";
4155 else if (pred_val < REG_BR_PROB_BASE / 2)
4156 which = ".dpnt";
4157 else if (pred_val < REG_BR_PROB_BASE / 100 * 98)
4158 which = ".dptk";
4159 else
4160 which = ".sptk";
4162 else if (GET_CODE (current_output_insn) == CALL_INSN)
4163 which = ".sptk";
4164 else
4165 which = ".dptk";
4167 fputs (which, file);
4168 return;
4171 case ',':
4172 x = current_insn_predicate;
4173 if (x)
4175 unsigned int regno = REGNO (XEXP (x, 0));
4176 if (GET_CODE (x) == EQ)
4177 regno += 1;
4178 fprintf (file, "(%s) ", reg_names [regno]);
4180 return;
4182 default:
4183 output_operand_lossage ("ia64_print_operand: unknown code");
4184 return;
4187 switch (GET_CODE (x))
4189 /* This happens for the spill/restore instructions. */
4190 case POST_INC:
4191 case POST_DEC:
4192 case POST_MODIFY:
4193 x = XEXP (x, 0);
4194 /* ... fall through ... */
4196 case REG:
4197 fputs (reg_names [REGNO (x)], file);
4198 break;
4200 case MEM:
4202 rtx addr = XEXP (x, 0);
4203 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC)
4204 addr = XEXP (addr, 0);
4205 fprintf (file, "[%s]", reg_names [REGNO (addr)]);
4206 break;
4209 default:
4210 output_addr_const (file, x);
4211 break;
4214 return;
4217 /* Compute a (partial) cost for rtx X. Return true if the complete
4218 cost has been computed, and false if subexpressions should be
4219 scanned. In either case, *TOTAL contains the cost result. */
4220 /* ??? This is incomplete. */
4222 static bool
4223 ia64_rtx_costs (rtx x, int code, int outer_code, int *total)
4225 switch (code)
4227 case CONST_INT:
4228 switch (outer_code)
4230 case SET:
4231 *total = CONST_OK_FOR_J (INTVAL (x)) ? 0 : COSTS_N_INSNS (1);
4232 return true;
4233 case PLUS:
4234 if (CONST_OK_FOR_I (INTVAL (x)))
4235 *total = 0;
4236 else if (CONST_OK_FOR_J (INTVAL (x)))
4237 *total = 1;
4238 else
4239 *total = COSTS_N_INSNS (1);
4240 return true;
4241 default:
4242 if (CONST_OK_FOR_K (INTVAL (x)) || CONST_OK_FOR_L (INTVAL (x)))
4243 *total = 0;
4244 else
4245 *total = COSTS_N_INSNS (1);
4246 return true;
4249 case CONST_DOUBLE:
4250 *total = COSTS_N_INSNS (1);
4251 return true;
4253 case CONST:
4254 case SYMBOL_REF:
4255 case LABEL_REF:
4256 *total = COSTS_N_INSNS (3);
4257 return true;
4259 case MULT:
4260 /* For multiplies wider than HImode, we have to go to the FPU,
4261 which normally involves copies. Plus there's the latency
4262 of the multiply itself, and the latency of the instructions to
4263 transfer integer regs to FP regs. */
4264 /* ??? Check for FP mode. */
4265 if (GET_MODE_SIZE (GET_MODE (x)) > 2)
4266 *total = COSTS_N_INSNS (10);
4267 else
4268 *total = COSTS_N_INSNS (2);
4269 return true;
4271 case PLUS:
4272 case MINUS:
4273 case ASHIFT:
4274 case ASHIFTRT:
4275 case LSHIFTRT:
4276 *total = COSTS_N_INSNS (1);
4277 return true;
4279 case DIV:
4280 case UDIV:
4281 case MOD:
4282 case UMOD:
4283 /* We make divide expensive, so that divide-by-constant will be
4284 optimized to a multiply. */
4285 *total = COSTS_N_INSNS (60);
4286 return true;
4288 default:
4289 return false;
4293 /* Calculate the cost of moving data from a register in class FROM to
4294 one in class TO, using MODE. */
4297 ia64_register_move_cost (enum machine_mode mode, enum reg_class from,
4298 enum reg_class to)
4300 /* ADDL_REGS is the same as GR_REGS for movement purposes. */
4301 if (to == ADDL_REGS)
4302 to = GR_REGS;
4303 if (from == ADDL_REGS)
4304 from = GR_REGS;
4306 /* All costs are symmetric, so reduce cases by putting the
4307 lower number class as the destination. */
4308 if (from < to)
4310 enum reg_class tmp = to;
4311 to = from, from = tmp;
4314 /* Moving from FR<->GR in XFmode must be more expensive than 2,
4315 so that we get secondary memory reloads. Between FR_REGS,
4316 we have to make this at least as expensive as MEMORY_MOVE_COST
4317 to avoid spectacularly poor register class preferencing. */
4318 if (mode == XFmode)
4320 if (to != GR_REGS || from != GR_REGS)
4321 return MEMORY_MOVE_COST (mode, to, 0);
4322 else
4323 return 3;
4326 switch (to)
4328 case PR_REGS:
4329 /* Moving between PR registers takes two insns. */
4330 if (from == PR_REGS)
4331 return 3;
4332 /* Moving between PR and anything but GR is impossible. */
4333 if (from != GR_REGS)
4334 return MEMORY_MOVE_COST (mode, to, 0);
4335 break;
4337 case BR_REGS:
4338 /* Moving between BR and anything but GR is impossible. */
4339 if (from != GR_REGS && from != GR_AND_BR_REGS)
4340 return MEMORY_MOVE_COST (mode, to, 0);
4341 break;
4343 case AR_I_REGS:
4344 case AR_M_REGS:
4345 /* Moving between AR and anything but GR is impossible. */
4346 if (from != GR_REGS)
4347 return MEMORY_MOVE_COST (mode, to, 0);
4348 break;
4350 case GR_REGS:
4351 case FR_REGS:
4352 case GR_AND_FR_REGS:
4353 case GR_AND_BR_REGS:
4354 case ALL_REGS:
4355 break;
4357 default:
4358 abort ();
4361 return 2;
4364 /* Implement PREFERRED_RELOAD_CLASS. Place additional restrictions on CLASS
4365 to use when copying X into that class. */
4367 enum reg_class
4368 ia64_preferred_reload_class (rtx x, enum reg_class class)
4370 switch (class)
4372 case FR_REGS:
4373 /* Don't allow volatile mem reloads into floating point registers.
4374 This is defined to force reload to choose the r/m case instead
4375 of the f/f case when reloading (set (reg fX) (mem/v)). */
4376 if (MEM_P (x) && MEM_VOLATILE_P (x))
4377 return NO_REGS;
4379 /* Force all unrecognized constants into the constant pool. */
4380 if (CONSTANT_P (x))
4381 return NO_REGS;
4382 break;
4384 case AR_M_REGS:
4385 case AR_I_REGS:
4386 if (!OBJECT_P (x))
4387 return NO_REGS;
4388 break;
4390 default:
4391 break;
4394 return class;
4397 /* This function returns the register class required for a secondary
4398 register when copying between one of the registers in CLASS, and X,
4399 using MODE. A return value of NO_REGS means that no secondary register
4400 is required. */
4402 enum reg_class
4403 ia64_secondary_reload_class (enum reg_class class,
4404 enum machine_mode mode ATTRIBUTE_UNUSED, rtx x)
4406 int regno = -1;
4408 if (GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
4409 regno = true_regnum (x);
4411 switch (class)
4413 case BR_REGS:
4414 case AR_M_REGS:
4415 case AR_I_REGS:
4416 /* ??? BR<->BR register copies can happen due to a bad gcse/cse/global
4417 interaction. We end up with two pseudos with overlapping lifetimes
4418 both of which are equiv to the same constant, and both which need
4419 to be in BR_REGS. This seems to be a cse bug. cse_basic_block_end
4420 changes depending on the path length, which means the qty_first_reg
4421 check in make_regs_eqv can give different answers at different times.
4422 At some point I'll probably need a reload_indi pattern to handle
4423 this.
4425 We can also get GR_AND_FR_REGS to BR_REGS/AR_REGS copies, where we
4426 wound up with a FP register from GR_AND_FR_REGS. Extend that to all
4427 non-general registers for good measure. */
4428 if (regno >= 0 && ! GENERAL_REGNO_P (regno))
4429 return GR_REGS;
4431 /* This is needed if a pseudo used as a call_operand gets spilled to a
4432 stack slot. */
4433 if (GET_CODE (x) == MEM)
4434 return GR_REGS;
4435 break;
4437 case FR_REGS:
4438 /* Need to go through general registers to get to other class regs. */
4439 if (regno >= 0 && ! (FR_REGNO_P (regno) || GENERAL_REGNO_P (regno)))
4440 return GR_REGS;
4442 /* This can happen when a paradoxical subreg is an operand to the
4443 muldi3 pattern. */
4444 /* ??? This shouldn't be necessary after instruction scheduling is
4445 enabled, because paradoxical subregs are not accepted by
4446 register_operand when INSN_SCHEDULING is defined. Or alternatively,
4447 stop the paradoxical subreg stupidity in the *_operand functions
4448 in recog.c. */
4449 if (GET_CODE (x) == MEM
4450 && (GET_MODE (x) == SImode || GET_MODE (x) == HImode
4451 || GET_MODE (x) == QImode))
4452 return GR_REGS;
4454 /* This can happen because of the ior/and/etc patterns that accept FP
4455 registers as operands. If the third operand is a constant, then it
4456 needs to be reloaded into a FP register. */
4457 if (GET_CODE (x) == CONST_INT)
4458 return GR_REGS;
4460 /* This can happen because of register elimination in a muldi3 insn.
4461 E.g. `26107 * (unsigned long)&u'. */
4462 if (GET_CODE (x) == PLUS)
4463 return GR_REGS;
4464 break;
4466 case PR_REGS:
4467 /* ??? This happens if we cse/gcse a BImode value across a call,
4468 and the function has a nonlocal goto. This is because global
4469 does not allocate call crossing pseudos to hard registers when
4470 current_function_has_nonlocal_goto is true. This is relatively
4471 common for C++ programs that use exceptions. To reproduce,
4472 return NO_REGS and compile libstdc++. */
4473 if (GET_CODE (x) == MEM)
4474 return GR_REGS;
4476 /* This can happen when we take a BImode subreg of a DImode value,
4477 and that DImode value winds up in some non-GR register. */
4478 if (regno >= 0 && ! GENERAL_REGNO_P (regno) && ! PR_REGNO_P (regno))
4479 return GR_REGS;
4480 break;
4482 default:
4483 break;
4486 return NO_REGS;
4490 /* Emit text to declare externally defined variables and functions, because
4491 the Intel assembler does not support undefined externals. */
4493 void
4494 ia64_asm_output_external (FILE *file, tree decl, const char *name)
4496 int save_referenced;
4498 /* GNU as does not need anything here, but the HP linker does need
4499 something for external functions. */
4501 if (TARGET_GNU_AS
4502 && (!TARGET_HPUX_LD
4503 || TREE_CODE (decl) != FUNCTION_DECL
4504 || strstr (name, "__builtin_") == name))
4505 return;
4507 /* ??? The Intel assembler creates a reference that needs to be satisfied by
4508 the linker when we do this, so we need to be careful not to do this for
4509 builtin functions which have no library equivalent. Unfortunately, we
4510 can't tell here whether or not a function will actually be called by
4511 expand_expr, so we pull in library functions even if we may not need
4512 them later. */
4513 if (! strcmp (name, "__builtin_next_arg")
4514 || ! strcmp (name, "alloca")
4515 || ! strcmp (name, "__builtin_constant_p")
4516 || ! strcmp (name, "__builtin_args_info"))
4517 return;
4519 if (TARGET_HPUX_LD)
4520 ia64_hpux_add_extern_decl (decl);
4521 else
4523 /* assemble_name will set TREE_SYMBOL_REFERENCED, so we must save and
4524 restore it. */
4525 save_referenced = TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl));
4526 if (TREE_CODE (decl) == FUNCTION_DECL)
4527 ASM_OUTPUT_TYPE_DIRECTIVE (file, name, "function");
4528 (*targetm.asm_out.globalize_label) (file, name);
4529 TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)) = save_referenced;
4533 /* Parse the -mfixed-range= option string. */
4535 static void
4536 fix_range (const char *const_str)
4538 int i, first, last;
4539 char *str, *dash, *comma;
4541 /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and
4542 REG2 are either register names or register numbers. The effect
4543 of this option is to mark the registers in the range from REG1 to
4544 REG2 as ``fixed'' so they won't be used by the compiler. This is
4545 used, e.g., to ensure that kernel mode code doesn't use f32-f127. */
4547 i = strlen (const_str);
4548 str = (char *) alloca (i + 1);
4549 memcpy (str, const_str, i + 1);
4551 while (1)
4553 dash = strchr (str, '-');
4554 if (!dash)
4556 warning ("value of -mfixed-range must have form REG1-REG2");
4557 return;
4559 *dash = '\0';
4561 comma = strchr (dash + 1, ',');
4562 if (comma)
4563 *comma = '\0';
4565 first = decode_reg_name (str);
4566 if (first < 0)
4568 warning ("unknown register name: %s", str);
4569 return;
4572 last = decode_reg_name (dash + 1);
4573 if (last < 0)
4575 warning ("unknown register name: %s", dash + 1);
4576 return;
4579 *dash = '-';
4581 if (first > last)
4583 warning ("%s-%s is an empty range", str, dash + 1);
4584 return;
4587 for (i = first; i <= last; ++i)
4588 fixed_regs[i] = call_used_regs[i] = 1;
4590 if (!comma)
4591 break;
4593 *comma = ',';
4594 str = comma + 1;
4598 static struct machine_function *
4599 ia64_init_machine_status (void)
4601 return ggc_alloc_cleared (sizeof (struct machine_function));
4604 /* Handle TARGET_OPTIONS switches. */
4606 void
4607 ia64_override_options (void)
4609 static struct pta
4611 const char *const name; /* processor name or nickname. */
4612 const enum processor_type processor;
4614 const processor_alias_table[] =
4616 {"itanium", PROCESSOR_ITANIUM},
4617 {"itanium1", PROCESSOR_ITANIUM},
4618 {"merced", PROCESSOR_ITANIUM},
4619 {"itanium2", PROCESSOR_ITANIUM2},
4620 {"mckinley", PROCESSOR_ITANIUM2},
4623 int const pta_size = ARRAY_SIZE (processor_alias_table);
4624 int i;
4626 if (TARGET_AUTO_PIC)
4627 target_flags |= MASK_CONST_GP;
4629 if (TARGET_INLINE_FLOAT_DIV_LAT && TARGET_INLINE_FLOAT_DIV_THR)
4631 if ((target_flags_explicit & MASK_INLINE_FLOAT_DIV_LAT)
4632 && (target_flags_explicit & MASK_INLINE_FLOAT_DIV_THR))
4634 warning ("cannot optimize floating point division for both latency and throughput");
4635 target_flags &= ~MASK_INLINE_FLOAT_DIV_THR;
4637 else
4639 if (target_flags_explicit & MASK_INLINE_FLOAT_DIV_THR)
4640 target_flags &= ~MASK_INLINE_FLOAT_DIV_LAT;
4641 else
4642 target_flags &= ~MASK_INLINE_FLOAT_DIV_THR;
4646 if (TARGET_INLINE_INT_DIV_LAT && TARGET_INLINE_INT_DIV_THR)
4648 if ((target_flags_explicit & MASK_INLINE_INT_DIV_LAT)
4649 && (target_flags_explicit & MASK_INLINE_INT_DIV_THR))
4651 warning ("cannot optimize integer division for both latency and throughput");
4652 target_flags &= ~MASK_INLINE_INT_DIV_THR;
4654 else
4656 if (target_flags_explicit & MASK_INLINE_INT_DIV_THR)
4657 target_flags &= ~MASK_INLINE_INT_DIV_LAT;
4658 else
4659 target_flags &= ~MASK_INLINE_INT_DIV_THR;
4663 if (TARGET_INLINE_SQRT_LAT && TARGET_INLINE_SQRT_THR)
4665 if ((target_flags_explicit & MASK_INLINE_SQRT_LAT)
4666 && (target_flags_explicit & MASK_INLINE_SQRT_THR))
4668 warning ("cannot optimize square root for both latency and throughput");
4669 target_flags &= ~MASK_INLINE_SQRT_THR;
4671 else
4673 if (target_flags_explicit & MASK_INLINE_SQRT_THR)
4674 target_flags &= ~MASK_INLINE_SQRT_LAT;
4675 else
4676 target_flags &= ~MASK_INLINE_SQRT_THR;
4680 if (TARGET_INLINE_SQRT_LAT)
4682 warning ("not yet implemented: latency-optimized inline square root");
4683 target_flags &= ~MASK_INLINE_SQRT_LAT;
4686 if (ia64_fixed_range_string)
4687 fix_range (ia64_fixed_range_string);
4689 if (ia64_tls_size_string)
4691 char *end;
4692 unsigned long tmp = strtoul (ia64_tls_size_string, &end, 10);
4693 if (*end || (tmp != 14 && tmp != 22 && tmp != 64))
4694 error ("bad value (%s) for -mtls-size= switch", ia64_tls_size_string);
4695 else
4696 ia64_tls_size = tmp;
4699 if (!ia64_tune_string)
4700 ia64_tune_string = "itanium2";
4702 for (i = 0; i < pta_size; i++)
4703 if (! strcmp (ia64_tune_string, processor_alias_table[i].name))
4705 ia64_tune = processor_alias_table[i].processor;
4706 break;
4709 if (i == pta_size)
4710 error ("bad value (%s) for -tune= switch", ia64_tune_string);
4712 ia64_flag_schedule_insns2 = flag_schedule_insns_after_reload;
4713 flag_schedule_insns_after_reload = 0;
4715 /* Variable tracking should be run after all optimizations which change order
4716 of insns. It also needs a valid CFG. */
4717 ia64_flag_var_tracking = flag_var_tracking;
4718 flag_var_tracking = 0;
4720 ia64_section_threshold = g_switch_set ? g_switch_value : IA64_DEFAULT_GVALUE;
4722 init_machine_status = ia64_init_machine_status;
4725 static enum attr_itanium_class ia64_safe_itanium_class (rtx);
4726 static enum attr_type ia64_safe_type (rtx);
4728 static enum attr_itanium_class
4729 ia64_safe_itanium_class (rtx insn)
4731 if (recog_memoized (insn) >= 0)
4732 return get_attr_itanium_class (insn);
4733 else
4734 return ITANIUM_CLASS_UNKNOWN;
4737 static enum attr_type
4738 ia64_safe_type (rtx insn)
4740 if (recog_memoized (insn) >= 0)
4741 return get_attr_type (insn);
4742 else
4743 return TYPE_UNKNOWN;
4746 /* The following collection of routines emit instruction group stop bits as
4747 necessary to avoid dependencies. */
4749 /* Need to track some additional registers as far as serialization is
4750 concerned so we can properly handle br.call and br.ret. We could
4751 make these registers visible to gcc, but since these registers are
4752 never explicitly used in gcc generated code, it seems wasteful to
4753 do so (plus it would make the call and return patterns needlessly
4754 complex). */
4755 #define REG_RP (BR_REG (0))
4756 #define REG_AR_CFM (FIRST_PSEUDO_REGISTER + 1)
4757 /* This is used for volatile asms which may require a stop bit immediately
4758 before and after them. */
4759 #define REG_VOLATILE (FIRST_PSEUDO_REGISTER + 2)
4760 #define AR_UNAT_BIT_0 (FIRST_PSEUDO_REGISTER + 3)
4761 #define NUM_REGS (AR_UNAT_BIT_0 + 64)
4763 /* For each register, we keep track of how it has been written in the
4764 current instruction group.
4766 If a register is written unconditionally (no qualifying predicate),
4767 WRITE_COUNT is set to 2 and FIRST_PRED is ignored.
4769 If a register is written if its qualifying predicate P is true, we
4770 set WRITE_COUNT to 1 and FIRST_PRED to P. Later on, the same register
4771 may be written again by the complement of P (P^1) and when this happens,
4772 WRITE_COUNT gets set to 2.
4774 The result of this is that whenever an insn attempts to write a register
4775 whose WRITE_COUNT is two, we need to issue an insn group barrier first.
4777 If a predicate register is written by a floating-point insn, we set
4778 WRITTEN_BY_FP to true.
4780 If a predicate register is written by an AND.ORCM we set WRITTEN_BY_AND
4781 to true; if it was written by an OR.ANDCM we set WRITTEN_BY_OR to true. */
4783 struct reg_write_state
4785 unsigned int write_count : 2;
4786 unsigned int first_pred : 16;
4787 unsigned int written_by_fp : 1;
4788 unsigned int written_by_and : 1;
4789 unsigned int written_by_or : 1;
4792 /* Cumulative info for the current instruction group. */
4793 struct reg_write_state rws_sum[NUM_REGS];
4794 /* Info for the current instruction. This gets copied to rws_sum after a
4795 stop bit is emitted. */
4796 struct reg_write_state rws_insn[NUM_REGS];
4798 /* Indicates whether this is the first instruction after a stop bit,
4799 in which case we don't need another stop bit. Without this, we hit
4800 the abort in ia64_variable_issue when scheduling an alloc. */
4801 static int first_instruction;
4803 /* Misc flags needed to compute RAW/WAW dependencies while we are traversing
4804 RTL for one instruction. */
4805 struct reg_flags
4807 unsigned int is_write : 1; /* Is register being written? */
4808 unsigned int is_fp : 1; /* Is register used as part of an fp op? */
4809 unsigned int is_branch : 1; /* Is register used as part of a branch? */
4810 unsigned int is_and : 1; /* Is register used as part of and.orcm? */
4811 unsigned int is_or : 1; /* Is register used as part of or.andcm? */
4812 unsigned int is_sibcall : 1; /* Is this a sibling or normal call? */
4815 static void rws_update (struct reg_write_state *, int, struct reg_flags, int);
4816 static int rws_access_regno (int, struct reg_flags, int);
4817 static int rws_access_reg (rtx, struct reg_flags, int);
4818 static void update_set_flags (rtx, struct reg_flags *, int *, rtx *);
4819 static int set_src_needs_barrier (rtx, struct reg_flags, int, rtx);
4820 static int rtx_needs_barrier (rtx, struct reg_flags, int);
4821 static void init_insn_group_barriers (void);
4822 static int group_barrier_needed_p (rtx);
4823 static int safe_group_barrier_needed_p (rtx);
4825 /* Update *RWS for REGNO, which is being written by the current instruction,
4826 with predicate PRED, and associated register flags in FLAGS. */
4828 static void
4829 rws_update (struct reg_write_state *rws, int regno, struct reg_flags flags, int pred)
4831 if (pred)
4832 rws[regno].write_count++;
4833 else
4834 rws[regno].write_count = 2;
4835 rws[regno].written_by_fp |= flags.is_fp;
4836 /* ??? Not tracking and/or across differing predicates. */
4837 rws[regno].written_by_and = flags.is_and;
4838 rws[regno].written_by_or = flags.is_or;
4839 rws[regno].first_pred = pred;
4842 /* Handle an access to register REGNO of type FLAGS using predicate register
4843 PRED. Update rws_insn and rws_sum arrays. Return 1 if this access creates
4844 a dependency with an earlier instruction in the same group. */
4846 static int
4847 rws_access_regno (int regno, struct reg_flags flags, int pred)
4849 int need_barrier = 0;
4851 if (regno >= NUM_REGS)
4852 abort ();
4854 if (! PR_REGNO_P (regno))
4855 flags.is_and = flags.is_or = 0;
4857 if (flags.is_write)
4859 int write_count;
4861 /* One insn writes same reg multiple times? */
4862 if (rws_insn[regno].write_count > 0)
4863 abort ();
4865 /* Update info for current instruction. */
4866 rws_update (rws_insn, regno, flags, pred);
4867 write_count = rws_sum[regno].write_count;
4869 switch (write_count)
4871 case 0:
4872 /* The register has not been written yet. */
4873 rws_update (rws_sum, regno, flags, pred);
4874 break;
4876 case 1:
4877 /* The register has been written via a predicate. If this is
4878 not a complementary predicate, then we need a barrier. */
4879 /* ??? This assumes that P and P+1 are always complementary
4880 predicates for P even. */
4881 if (flags.is_and && rws_sum[regno].written_by_and)
4883 else if (flags.is_or && rws_sum[regno].written_by_or)
4885 else if ((rws_sum[regno].first_pred ^ 1) != pred)
4886 need_barrier = 1;
4887 rws_update (rws_sum, regno, flags, pred);
4888 break;
4890 case 2:
4891 /* The register has been unconditionally written already. We
4892 need a barrier. */
4893 if (flags.is_and && rws_sum[regno].written_by_and)
4895 else if (flags.is_or && rws_sum[regno].written_by_or)
4897 else
4898 need_barrier = 1;
4899 rws_sum[regno].written_by_and = flags.is_and;
4900 rws_sum[regno].written_by_or = flags.is_or;
4901 break;
4903 default:
4904 abort ();
4907 else
4909 if (flags.is_branch)
4911 /* Branches have several RAW exceptions that allow to avoid
4912 barriers. */
4914 if (REGNO_REG_CLASS (regno) == BR_REGS || regno == AR_PFS_REGNUM)
4915 /* RAW dependencies on branch regs are permissible as long
4916 as the writer is a non-branch instruction. Since we
4917 never generate code that uses a branch register written
4918 by a branch instruction, handling this case is
4919 easy. */
4920 return 0;
4922 if (REGNO_REG_CLASS (regno) == PR_REGS
4923 && ! rws_sum[regno].written_by_fp)
4924 /* The predicates of a branch are available within the
4925 same insn group as long as the predicate was written by
4926 something other than a floating-point instruction. */
4927 return 0;
4930 if (flags.is_and && rws_sum[regno].written_by_and)
4931 return 0;
4932 if (flags.is_or && rws_sum[regno].written_by_or)
4933 return 0;
4935 switch (rws_sum[regno].write_count)
4937 case 0:
4938 /* The register has not been written yet. */
4939 break;
4941 case 1:
4942 /* The register has been written via a predicate. If this is
4943 not a complementary predicate, then we need a barrier. */
4944 /* ??? This assumes that P and P+1 are always complementary
4945 predicates for P even. */
4946 if ((rws_sum[regno].first_pred ^ 1) != pred)
4947 need_barrier = 1;
4948 break;
4950 case 2:
4951 /* The register has been unconditionally written already. We
4952 need a barrier. */
4953 need_barrier = 1;
4954 break;
4956 default:
4957 abort ();
4961 return need_barrier;
4964 static int
4965 rws_access_reg (rtx reg, struct reg_flags flags, int pred)
4967 int regno = REGNO (reg);
4968 int n = HARD_REGNO_NREGS (REGNO (reg), GET_MODE (reg));
4970 if (n == 1)
4971 return rws_access_regno (regno, flags, pred);
4972 else
4974 int need_barrier = 0;
4975 while (--n >= 0)
4976 need_barrier |= rws_access_regno (regno + n, flags, pred);
4977 return need_barrier;
4981 /* Examine X, which is a SET rtx, and update the flags, the predicate, and
4982 the condition, stored in *PFLAGS, *PPRED and *PCOND. */
4984 static void
4985 update_set_flags (rtx x, struct reg_flags *pflags, int *ppred, rtx *pcond)
4987 rtx src = SET_SRC (x);
4989 *pcond = 0;
4991 switch (GET_CODE (src))
4993 case CALL:
4994 return;
4996 case IF_THEN_ELSE:
4997 if (SET_DEST (x) == pc_rtx)
4998 /* X is a conditional branch. */
4999 return;
5000 else
5002 int is_complemented = 0;
5004 /* X is a conditional move. */
5005 rtx cond = XEXP (src, 0);
5006 if (GET_CODE (cond) == EQ)
5007 is_complemented = 1;
5008 cond = XEXP (cond, 0);
5009 if (GET_CODE (cond) != REG
5010 && REGNO_REG_CLASS (REGNO (cond)) != PR_REGS)
5011 abort ();
5012 *pcond = cond;
5013 if (XEXP (src, 1) == SET_DEST (x)
5014 || XEXP (src, 2) == SET_DEST (x))
5016 /* X is a conditional move that conditionally writes the
5017 destination. */
5019 /* We need another complement in this case. */
5020 if (XEXP (src, 1) == SET_DEST (x))
5021 is_complemented = ! is_complemented;
5023 *ppred = REGNO (cond);
5024 if (is_complemented)
5025 ++*ppred;
5028 /* ??? If this is a conditional write to the dest, then this
5029 instruction does not actually read one source. This probably
5030 doesn't matter, because that source is also the dest. */
5031 /* ??? Multiple writes to predicate registers are allowed
5032 if they are all AND type compares, or if they are all OR
5033 type compares. We do not generate such instructions
5034 currently. */
5036 /* ... fall through ... */
5038 default:
5039 if (COMPARISON_P (src)
5040 && GET_MODE_CLASS (GET_MODE (XEXP (src, 0))) == MODE_FLOAT)
5041 /* Set pflags->is_fp to 1 so that we know we're dealing
5042 with a floating point comparison when processing the
5043 destination of the SET. */
5044 pflags->is_fp = 1;
5046 /* Discover if this is a parallel comparison. We only handle
5047 and.orcm and or.andcm at present, since we must retain a
5048 strict inverse on the predicate pair. */
5049 else if (GET_CODE (src) == AND)
5050 pflags->is_and = 1;
5051 else if (GET_CODE (src) == IOR)
5052 pflags->is_or = 1;
5054 break;
5058 /* Subroutine of rtx_needs_barrier; this function determines whether the
5059 source of a given SET rtx found in X needs a barrier. FLAGS and PRED
5060 are as in rtx_needs_barrier. COND is an rtx that holds the condition
5061 for this insn. */
5063 static int
5064 set_src_needs_barrier (rtx x, struct reg_flags flags, int pred, rtx cond)
5066 int need_barrier = 0;
5067 rtx dst;
5068 rtx src = SET_SRC (x);
5070 if (GET_CODE (src) == CALL)
5071 /* We don't need to worry about the result registers that
5072 get written by subroutine call. */
5073 return rtx_needs_barrier (src, flags, pred);
5074 else if (SET_DEST (x) == pc_rtx)
5076 /* X is a conditional branch. */
5077 /* ??? This seems redundant, as the caller sets this bit for
5078 all JUMP_INSNs. */
5079 flags.is_branch = 1;
5080 return rtx_needs_barrier (src, flags, pred);
5083 need_barrier = rtx_needs_barrier (src, flags, pred);
5085 /* This instruction unconditionally uses a predicate register. */
5086 if (cond)
5087 need_barrier |= rws_access_reg (cond, flags, 0);
5089 dst = SET_DEST (x);
5090 if (GET_CODE (dst) == ZERO_EXTRACT)
5092 need_barrier |= rtx_needs_barrier (XEXP (dst, 1), flags, pred);
5093 need_barrier |= rtx_needs_barrier (XEXP (dst, 2), flags, pred);
5094 dst = XEXP (dst, 0);
5096 return need_barrier;
5099 /* Handle an access to rtx X of type FLAGS using predicate register
5100 PRED. Return 1 if this access creates a dependency with an earlier
5101 instruction in the same group. */
5103 static int
5104 rtx_needs_barrier (rtx x, struct reg_flags flags, int pred)
5106 int i, j;
5107 int is_complemented = 0;
5108 int need_barrier = 0;
5109 const char *format_ptr;
5110 struct reg_flags new_flags;
5111 rtx cond = 0;
5113 if (! x)
5114 return 0;
5116 new_flags = flags;
5118 switch (GET_CODE (x))
5120 case SET:
5121 update_set_flags (x, &new_flags, &pred, &cond);
5122 need_barrier = set_src_needs_barrier (x, new_flags, pred, cond);
5123 if (GET_CODE (SET_SRC (x)) != CALL)
5125 new_flags.is_write = 1;
5126 need_barrier |= rtx_needs_barrier (SET_DEST (x), new_flags, pred);
5128 break;
5130 case CALL:
5131 new_flags.is_write = 0;
5132 need_barrier |= rws_access_regno (AR_EC_REGNUM, new_flags, pred);
5134 /* Avoid multiple register writes, in case this is a pattern with
5135 multiple CALL rtx. This avoids an abort in rws_access_reg. */
5136 if (! flags.is_sibcall && ! rws_insn[REG_AR_CFM].write_count)
5138 new_flags.is_write = 1;
5139 need_barrier |= rws_access_regno (REG_RP, new_flags, pred);
5140 need_barrier |= rws_access_regno (AR_PFS_REGNUM, new_flags, pred);
5141 need_barrier |= rws_access_regno (REG_AR_CFM, new_flags, pred);
5143 break;
5145 case COND_EXEC:
5146 /* X is a predicated instruction. */
5148 cond = COND_EXEC_TEST (x);
5149 if (pred)
5150 abort ();
5151 need_barrier = rtx_needs_barrier (cond, flags, 0);
5153 if (GET_CODE (cond) == EQ)
5154 is_complemented = 1;
5155 cond = XEXP (cond, 0);
5156 if (GET_CODE (cond) != REG
5157 && REGNO_REG_CLASS (REGNO (cond)) != PR_REGS)
5158 abort ();
5159 pred = REGNO (cond);
5160 if (is_complemented)
5161 ++pred;
5163 need_barrier |= rtx_needs_barrier (COND_EXEC_CODE (x), flags, pred);
5164 return need_barrier;
5166 case CLOBBER:
5167 case USE:
5168 /* Clobber & use are for earlier compiler-phases only. */
5169 break;
5171 case ASM_OPERANDS:
5172 case ASM_INPUT:
5173 /* We always emit stop bits for traditional asms. We emit stop bits
5174 for volatile extended asms if TARGET_VOL_ASM_STOP is true. */
5175 if (GET_CODE (x) != ASM_OPERANDS
5176 || (MEM_VOLATILE_P (x) && TARGET_VOL_ASM_STOP))
5178 /* Avoid writing the register multiple times if we have multiple
5179 asm outputs. This avoids an abort in rws_access_reg. */
5180 if (! rws_insn[REG_VOLATILE].write_count)
5182 new_flags.is_write = 1;
5183 rws_access_regno (REG_VOLATILE, new_flags, pred);
5185 return 1;
5188 /* For all ASM_OPERANDS, we must traverse the vector of input operands.
5189 We cannot just fall through here since then we would be confused
5190 by the ASM_INPUT rtx inside ASM_OPERANDS, which do not indicate
5191 traditional asms unlike their normal usage. */
5193 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; --i)
5194 if (rtx_needs_barrier (ASM_OPERANDS_INPUT (x, i), flags, pred))
5195 need_barrier = 1;
5196 break;
5198 case PARALLEL:
5199 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
5201 rtx pat = XVECEXP (x, 0, i);
5202 switch (GET_CODE (pat))
5204 case SET:
5205 update_set_flags (pat, &new_flags, &pred, &cond);
5206 need_barrier |= set_src_needs_barrier (pat, new_flags,
5207 pred, cond);
5208 break;
5210 case USE:
5211 case CALL:
5212 case ASM_OPERANDS:
5213 need_barrier |= rtx_needs_barrier (pat, flags, pred);
5214 break;
5216 case CLOBBER:
5217 case RETURN:
5218 break;
5220 default:
5221 gcc_unreachable ();
5224 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
5226 rtx pat = XVECEXP (x, 0, i);
5227 if (GET_CODE (pat) == SET)
5229 if (GET_CODE (SET_SRC (pat)) != CALL)
5231 new_flags.is_write = 1;
5232 need_barrier |= rtx_needs_barrier (SET_DEST (pat), new_flags,
5233 pred);
5236 else if (GET_CODE (pat) == CLOBBER || GET_CODE (pat) == RETURN)
5237 need_barrier |= rtx_needs_barrier (pat, flags, pred);
5239 break;
5241 case SUBREG:
5242 x = SUBREG_REG (x);
5243 /* FALLTHRU */
5244 case REG:
5245 if (REGNO (x) == AR_UNAT_REGNUM)
5247 for (i = 0; i < 64; ++i)
5248 need_barrier |= rws_access_regno (AR_UNAT_BIT_0 + i, flags, pred);
5250 else
5251 need_barrier = rws_access_reg (x, flags, pred);
5252 break;
5254 case MEM:
5255 /* Find the regs used in memory address computation. */
5256 new_flags.is_write = 0;
5257 need_barrier = rtx_needs_barrier (XEXP (x, 0), new_flags, pred);
5258 break;
5260 case CONST_INT: case CONST_DOUBLE: case CONST_VECTOR:
5261 case SYMBOL_REF: case LABEL_REF: case CONST:
5262 break;
5264 /* Operators with side-effects. */
5265 case POST_INC: case POST_DEC:
5266 if (GET_CODE (XEXP (x, 0)) != REG)
5267 abort ();
5269 new_flags.is_write = 0;
5270 need_barrier = rws_access_reg (XEXP (x, 0), new_flags, pred);
5271 new_flags.is_write = 1;
5272 need_barrier |= rws_access_reg (XEXP (x, 0), new_flags, pred);
5273 break;
5275 case POST_MODIFY:
5276 if (GET_CODE (XEXP (x, 0)) != REG)
5277 abort ();
5279 new_flags.is_write = 0;
5280 need_barrier = rws_access_reg (XEXP (x, 0), new_flags, pred);
5281 need_barrier |= rtx_needs_barrier (XEXP (x, 1), new_flags, pred);
5282 new_flags.is_write = 1;
5283 need_barrier |= rws_access_reg (XEXP (x, 0), new_flags, pred);
5284 break;
5286 /* Handle common unary and binary ops for efficiency. */
5287 case COMPARE: case PLUS: case MINUS: case MULT: case DIV:
5288 case MOD: case UDIV: case UMOD: case AND: case IOR:
5289 case XOR: case ASHIFT: case ROTATE: case ASHIFTRT: case LSHIFTRT:
5290 case ROTATERT: case SMIN: case SMAX: case UMIN: case UMAX:
5291 case NE: case EQ: case GE: case GT: case LE:
5292 case LT: case GEU: case GTU: case LEU: case LTU:
5293 need_barrier = rtx_needs_barrier (XEXP (x, 0), new_flags, pred);
5294 need_barrier |= rtx_needs_barrier (XEXP (x, 1), new_flags, pred);
5295 break;
5297 case NEG: case NOT: case SIGN_EXTEND: case ZERO_EXTEND:
5298 case TRUNCATE: case FLOAT_EXTEND: case FLOAT_TRUNCATE: case FLOAT:
5299 case FIX: case UNSIGNED_FLOAT: case UNSIGNED_FIX: case ABS:
5300 case SQRT: case FFS: case POPCOUNT:
5301 need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
5302 break;
5304 case VEC_SELECT:
5305 /* VEC_SELECT's second argument is a PARALLEL with integers that
5306 describe the elements selected. On ia64, those integers are
5307 always constants. Avoid walking the PARALLEL so that we don't
5308 get confused with "normal" parallels and abort. */
5309 need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
5310 break;
5312 case UNSPEC:
5313 switch (XINT (x, 1))
5315 case UNSPEC_LTOFF_DTPMOD:
5316 case UNSPEC_LTOFF_DTPREL:
5317 case UNSPEC_DTPREL:
5318 case UNSPEC_LTOFF_TPREL:
5319 case UNSPEC_TPREL:
5320 case UNSPEC_PRED_REL_MUTEX:
5321 case UNSPEC_PIC_CALL:
5322 case UNSPEC_MF:
5323 case UNSPEC_FETCHADD_ACQ:
5324 case UNSPEC_BSP_VALUE:
5325 case UNSPEC_FLUSHRS:
5326 case UNSPEC_BUNDLE_SELECTOR:
5327 break;
5329 case UNSPEC_GR_SPILL:
5330 case UNSPEC_GR_RESTORE:
5332 HOST_WIDE_INT offset = INTVAL (XVECEXP (x, 0, 1));
5333 HOST_WIDE_INT bit = (offset >> 3) & 63;
5335 need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
5336 new_flags.is_write = (XINT (x, 1) == UNSPEC_GR_SPILL);
5337 need_barrier |= rws_access_regno (AR_UNAT_BIT_0 + bit,
5338 new_flags, pred);
5339 break;
5342 case UNSPEC_FR_SPILL:
5343 case UNSPEC_FR_RESTORE:
5344 case UNSPEC_GETF_EXP:
5345 case UNSPEC_SETF_EXP:
5346 case UNSPEC_ADDP4:
5347 case UNSPEC_FR_SQRT_RECIP_APPROX:
5348 need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
5349 break;
5351 case UNSPEC_FR_RECIP_APPROX:
5352 case UNSPEC_SHRP:
5353 need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
5354 need_barrier |= rtx_needs_barrier (XVECEXP (x, 0, 1), flags, pred);
5355 break;
5357 case UNSPEC_CMPXCHG_ACQ:
5358 need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 1), flags, pred);
5359 need_barrier |= rtx_needs_barrier (XVECEXP (x, 0, 2), flags, pred);
5360 break;
5362 default:
5363 abort ();
5365 break;
5367 case UNSPEC_VOLATILE:
5368 switch (XINT (x, 1))
5370 case UNSPECV_ALLOC:
5371 /* Alloc must always be the first instruction of a group.
5372 We force this by always returning true. */
5373 /* ??? We might get better scheduling if we explicitly check for
5374 input/local/output register dependencies, and modify the
5375 scheduler so that alloc is always reordered to the start of
5376 the current group. We could then eliminate all of the
5377 first_instruction code. */
5378 rws_access_regno (AR_PFS_REGNUM, flags, pred);
5380 new_flags.is_write = 1;
5381 rws_access_regno (REG_AR_CFM, new_flags, pred);
5382 return 1;
5384 case UNSPECV_SET_BSP:
5385 need_barrier = 1;
5386 break;
5388 case UNSPECV_BLOCKAGE:
5389 case UNSPECV_INSN_GROUP_BARRIER:
5390 case UNSPECV_BREAK:
5391 case UNSPECV_PSAC_ALL:
5392 case UNSPECV_PSAC_NORMAL:
5393 return 0;
5395 default:
5396 abort ();
5398 break;
5400 case RETURN:
5401 new_flags.is_write = 0;
5402 need_barrier = rws_access_regno (REG_RP, flags, pred);
5403 need_barrier |= rws_access_regno (AR_PFS_REGNUM, flags, pred);
5405 new_flags.is_write = 1;
5406 need_barrier |= rws_access_regno (AR_EC_REGNUM, new_flags, pred);
5407 need_barrier |= rws_access_regno (REG_AR_CFM, new_flags, pred);
5408 break;
5410 default:
5411 format_ptr = GET_RTX_FORMAT (GET_CODE (x));
5412 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
5413 switch (format_ptr[i])
5415 case '0': /* unused field */
5416 case 'i': /* integer */
5417 case 'n': /* note */
5418 case 'w': /* wide integer */
5419 case 's': /* pointer to string */
5420 case 'S': /* optional pointer to string */
5421 break;
5423 case 'e':
5424 if (rtx_needs_barrier (XEXP (x, i), flags, pred))
5425 need_barrier = 1;
5426 break;
5428 case 'E':
5429 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
5430 if (rtx_needs_barrier (XVECEXP (x, i, j), flags, pred))
5431 need_barrier = 1;
5432 break;
5434 default:
5435 abort ();
5437 break;
5439 return need_barrier;
5442 /* Clear out the state for group_barrier_needed_p at the start of a
5443 sequence of insns. */
5445 static void
5446 init_insn_group_barriers (void)
5448 memset (rws_sum, 0, sizeof (rws_sum));
5449 first_instruction = 1;
5452 /* Given the current state, recorded by previous calls to this function,
5453 determine whether a group barrier (a stop bit) is necessary before INSN.
5454 Return nonzero if so. */
5456 static int
5457 group_barrier_needed_p (rtx insn)
5459 rtx pat;
5460 int need_barrier = 0;
5461 struct reg_flags flags;
5463 memset (&flags, 0, sizeof (flags));
5464 switch (GET_CODE (insn))
5466 case NOTE:
5467 break;
5469 case BARRIER:
5470 /* A barrier doesn't imply an instruction group boundary. */
5471 break;
5473 case CODE_LABEL:
5474 memset (rws_insn, 0, sizeof (rws_insn));
5475 return 1;
5477 case CALL_INSN:
5478 flags.is_branch = 1;
5479 flags.is_sibcall = SIBLING_CALL_P (insn);
5480 memset (rws_insn, 0, sizeof (rws_insn));
5482 /* Don't bundle a call following another call. */
5483 if ((pat = prev_active_insn (insn))
5484 && GET_CODE (pat) == CALL_INSN)
5486 need_barrier = 1;
5487 break;
5490 need_barrier = rtx_needs_barrier (PATTERN (insn), flags, 0);
5491 break;
5493 case JUMP_INSN:
5494 flags.is_branch = 1;
5496 /* Don't bundle a jump following a call. */
5497 if ((pat = prev_active_insn (insn))
5498 && GET_CODE (pat) == CALL_INSN)
5500 need_barrier = 1;
5501 break;
5503 /* FALLTHRU */
5505 case INSN:
5506 if (GET_CODE (PATTERN (insn)) == USE
5507 || GET_CODE (PATTERN (insn)) == CLOBBER)
5508 /* Don't care about USE and CLOBBER "insns"---those are used to
5509 indicate to the optimizer that it shouldn't get rid of
5510 certain operations. */
5511 break;
5513 pat = PATTERN (insn);
5515 /* Ug. Hack hacks hacked elsewhere. */
5516 switch (recog_memoized (insn))
5518 /* We play dependency tricks with the epilogue in order
5519 to get proper schedules. Undo this for dv analysis. */
5520 case CODE_FOR_epilogue_deallocate_stack:
5521 case CODE_FOR_prologue_allocate_stack:
5522 pat = XVECEXP (pat, 0, 0);
5523 break;
5525 /* The pattern we use for br.cloop confuses the code above.
5526 The second element of the vector is representative. */
5527 case CODE_FOR_doloop_end_internal:
5528 pat = XVECEXP (pat, 0, 1);
5529 break;
5531 /* Doesn't generate code. */
5532 case CODE_FOR_pred_rel_mutex:
5533 case CODE_FOR_prologue_use:
5534 return 0;
5536 default:
5537 break;
5540 memset (rws_insn, 0, sizeof (rws_insn));
5541 need_barrier = rtx_needs_barrier (pat, flags, 0);
5543 /* Check to see if the previous instruction was a volatile
5544 asm. */
5545 if (! need_barrier)
5546 need_barrier = rws_access_regno (REG_VOLATILE, flags, 0);
5547 break;
5549 default:
5550 abort ();
5553 if (first_instruction && INSN_P (insn)
5554 && ia64_safe_itanium_class (insn) != ITANIUM_CLASS_IGNORE
5555 && GET_CODE (PATTERN (insn)) != USE
5556 && GET_CODE (PATTERN (insn)) != CLOBBER)
5558 need_barrier = 0;
5559 first_instruction = 0;
5562 return need_barrier;
5565 /* Like group_barrier_needed_p, but do not clobber the current state. */
5567 static int
5568 safe_group_barrier_needed_p (rtx insn)
5570 struct reg_write_state rws_saved[NUM_REGS];
5571 int saved_first_instruction;
5572 int t;
5574 memcpy (rws_saved, rws_sum, NUM_REGS * sizeof *rws_saved);
5575 saved_first_instruction = first_instruction;
5577 t = group_barrier_needed_p (insn);
5579 memcpy (rws_sum, rws_saved, NUM_REGS * sizeof *rws_saved);
5580 first_instruction = saved_first_instruction;
5582 return t;
5585 /* Scan the current function and insert stop bits as necessary to
5586 eliminate dependencies. This function assumes that a final
5587 instruction scheduling pass has been run which has already
5588 inserted most of the necessary stop bits. This function only
5589 inserts new ones at basic block boundaries, since these are
5590 invisible to the scheduler. */
5592 static void
5593 emit_insn_group_barriers (FILE *dump)
5595 rtx insn;
5596 rtx last_label = 0;
5597 int insns_since_last_label = 0;
5599 init_insn_group_barriers ();
5601 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
5603 if (GET_CODE (insn) == CODE_LABEL)
5605 if (insns_since_last_label)
5606 last_label = insn;
5607 insns_since_last_label = 0;
5609 else if (GET_CODE (insn) == NOTE
5610 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_BASIC_BLOCK)
5612 if (insns_since_last_label)
5613 last_label = insn;
5614 insns_since_last_label = 0;
5616 else if (GET_CODE (insn) == INSN
5617 && GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
5618 && XINT (PATTERN (insn), 1) == UNSPECV_INSN_GROUP_BARRIER)
5620 init_insn_group_barriers ();
5621 last_label = 0;
5623 else if (INSN_P (insn))
5625 insns_since_last_label = 1;
5627 if (group_barrier_needed_p (insn))
5629 if (last_label)
5631 if (dump)
5632 fprintf (dump, "Emitting stop before label %d\n",
5633 INSN_UID (last_label));
5634 emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), last_label);
5635 insn = last_label;
5637 init_insn_group_barriers ();
5638 last_label = 0;
5645 /* Like emit_insn_group_barriers, but run if no final scheduling pass was run.
5646 This function has to emit all necessary group barriers. */
5648 static void
5649 emit_all_insn_group_barriers (FILE *dump ATTRIBUTE_UNUSED)
5651 rtx insn;
5653 init_insn_group_barriers ();
5655 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
5657 if (GET_CODE (insn) == BARRIER)
5659 rtx last = prev_active_insn (insn);
5661 if (! last)
5662 continue;
5663 if (GET_CODE (last) == JUMP_INSN
5664 && GET_CODE (PATTERN (last)) == ADDR_DIFF_VEC)
5665 last = prev_active_insn (last);
5666 if (recog_memoized (last) != CODE_FOR_insn_group_barrier)
5667 emit_insn_after (gen_insn_group_barrier (GEN_INT (3)), last);
5669 init_insn_group_barriers ();
5671 else if (INSN_P (insn))
5673 if (recog_memoized (insn) == CODE_FOR_insn_group_barrier)
5674 init_insn_group_barriers ();
5675 else if (group_barrier_needed_p (insn))
5677 emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), insn);
5678 init_insn_group_barriers ();
5679 group_barrier_needed_p (insn);
5686 static int errata_find_address_regs (rtx *, void *);
5687 static void errata_emit_nops (rtx);
5688 static void fixup_errata (void);
5690 /* This structure is used to track some details about the previous insns
5691 groups so we can determine if it may be necessary to insert NOPs to
5692 workaround hardware errata. */
5693 static struct group
5695 HARD_REG_SET p_reg_set;
5696 HARD_REG_SET gr_reg_conditionally_set;
5697 } last_group[2];
5699 /* Index into the last_group array. */
5700 static int group_idx;
5702 /* Called through for_each_rtx; determines if a hard register that was
5703 conditionally set in the previous group is used as an address register.
5704 It ensures that for_each_rtx returns 1 in that case. */
5705 static int
5706 errata_find_address_regs (rtx *xp, void *data ATTRIBUTE_UNUSED)
5708 rtx x = *xp;
5709 if (GET_CODE (x) != MEM)
5710 return 0;
5711 x = XEXP (x, 0);
5712 if (GET_CODE (x) == POST_MODIFY)
5713 x = XEXP (x, 0);
5714 if (GET_CODE (x) == REG)
5716 struct group *prev_group = last_group + (group_idx ^ 1);
5717 if (TEST_HARD_REG_BIT (prev_group->gr_reg_conditionally_set,
5718 REGNO (x)))
5719 return 1;
5720 return -1;
5722 return 0;
5725 /* Called for each insn; this function keeps track of the state in
5726 last_group and emits additional NOPs if necessary to work around
5727 an Itanium A/B step erratum. */
5728 static void
5729 errata_emit_nops (rtx insn)
5731 struct group *this_group = last_group + group_idx;
5732 struct group *prev_group = last_group + (group_idx ^ 1);
5733 rtx pat = PATTERN (insn);
5734 rtx cond = GET_CODE (pat) == COND_EXEC ? COND_EXEC_TEST (pat) : 0;
5735 rtx real_pat = cond ? COND_EXEC_CODE (pat) : pat;
5736 enum attr_type type;
5737 rtx set = real_pat;
5739 if (GET_CODE (real_pat) == USE
5740 || GET_CODE (real_pat) == CLOBBER
5741 || GET_CODE (real_pat) == ASM_INPUT
5742 || GET_CODE (real_pat) == ADDR_VEC
5743 || GET_CODE (real_pat) == ADDR_DIFF_VEC
5744 || asm_noperands (PATTERN (insn)) >= 0)
5745 return;
5747 /* single_set doesn't work for COND_EXEC insns, so we have to duplicate
5748 parts of it. */
5750 if (GET_CODE (set) == PARALLEL)
5752 int i;
5753 set = XVECEXP (real_pat, 0, 0);
5754 for (i = 1; i < XVECLEN (real_pat, 0); i++)
5755 if (GET_CODE (XVECEXP (real_pat, 0, i)) != USE
5756 && GET_CODE (XVECEXP (real_pat, 0, i)) != CLOBBER)
5758 set = 0;
5759 break;
5763 if (set && GET_CODE (set) != SET)
5764 set = 0;
5766 type = get_attr_type (insn);
5768 if (type == TYPE_F
5769 && set && REG_P (SET_DEST (set)) && PR_REGNO_P (REGNO (SET_DEST (set))))
5770 SET_HARD_REG_BIT (this_group->p_reg_set, REGNO (SET_DEST (set)));
5772 if ((type == TYPE_M || type == TYPE_A) && cond && set
5773 && REG_P (SET_DEST (set))
5774 && GET_CODE (SET_SRC (set)) != PLUS
5775 && GET_CODE (SET_SRC (set)) != MINUS
5776 && (GET_CODE (SET_SRC (set)) != ASHIFT
5777 || !shladd_operand (XEXP (SET_SRC (set), 1), VOIDmode))
5778 && (GET_CODE (SET_SRC (set)) != MEM
5779 || GET_CODE (XEXP (SET_SRC (set), 0)) != POST_MODIFY)
5780 && GENERAL_REGNO_P (REGNO (SET_DEST (set))))
5782 if (!COMPARISON_P (cond)
5783 || !REG_P (XEXP (cond, 0)))
5784 abort ();
5786 if (TEST_HARD_REG_BIT (prev_group->p_reg_set, REGNO (XEXP (cond, 0))))
5787 SET_HARD_REG_BIT (this_group->gr_reg_conditionally_set, REGNO (SET_DEST (set)));
5789 if (for_each_rtx (&real_pat, errata_find_address_regs, NULL))
5791 emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), insn);
5792 emit_insn_before (gen_nop (), insn);
5793 emit_insn_before (gen_insn_group_barrier (GEN_INT (3)), insn);
5794 group_idx = 0;
5795 memset (last_group, 0, sizeof last_group);
5799 /* Emit extra nops if they are required to work around hardware errata. */
5801 static void
5802 fixup_errata (void)
5804 rtx insn;
5806 if (! TARGET_B_STEP)
5807 return;
5809 group_idx = 0;
5810 memset (last_group, 0, sizeof last_group);
5812 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
5814 if (!INSN_P (insn))
5815 continue;
5817 if (ia64_safe_type (insn) == TYPE_S)
5819 group_idx ^= 1;
5820 memset (last_group + group_idx, 0, sizeof last_group[group_idx]);
5822 else
5823 errata_emit_nops (insn);
5828 /* Instruction scheduling support. */
5830 #define NR_BUNDLES 10
5832 /* A list of names of all available bundles. */
5834 static const char *bundle_name [NR_BUNDLES] =
5836 ".mii",
5837 ".mmi",
5838 ".mfi",
5839 ".mmf",
5840 #if NR_BUNDLES == 10
5841 ".bbb",
5842 ".mbb",
5843 #endif
5844 ".mib",
5845 ".mmb",
5846 ".mfb",
5847 ".mlx"
5850 /* Nonzero if we should insert stop bits into the schedule. */
5852 int ia64_final_schedule = 0;
5854 /* Codes of the corresponding queried units: */
5856 static int _0mii_, _0mmi_, _0mfi_, _0mmf_;
5857 static int _0bbb_, _0mbb_, _0mib_, _0mmb_, _0mfb_, _0mlx_;
5859 static int _1mii_, _1mmi_, _1mfi_, _1mmf_;
5860 static int _1bbb_, _1mbb_, _1mib_, _1mmb_, _1mfb_, _1mlx_;
5862 static int pos_1, pos_2, pos_3, pos_4, pos_5, pos_6;
5864 /* The following variable value is an insn group barrier. */
5866 static rtx dfa_stop_insn;
5868 /* The following variable value is the last issued insn. */
5870 static rtx last_scheduled_insn;
5872 /* The following variable value is size of the DFA state. */
5874 static size_t dfa_state_size;
5876 /* The following variable value is pointer to a DFA state used as
5877 temporary variable. */
5879 static state_t temp_dfa_state = NULL;
5881 /* The following variable value is DFA state after issuing the last
5882 insn. */
5884 static state_t prev_cycle_state = NULL;
5886 /* The following array element values are TRUE if the corresponding
5887 insn requires to add stop bits before it. */
5889 static char *stops_p;
5891 /* The following variable is used to set up the mentioned above array. */
5893 static int stop_before_p = 0;
5895 /* The following variable value is length of the arrays `clocks' and
5896 `add_cycles'. */
5898 static int clocks_length;
5900 /* The following array element values are cycles on which the
5901 corresponding insn will be issued. The array is used only for
5902 Itanium1. */
5904 static int *clocks;
5906 /* The following array element values are numbers of cycles should be
5907 added to improve insn scheduling for MM_insns for Itanium1. */
5909 static int *add_cycles;
5911 static rtx ia64_single_set (rtx);
5912 static void ia64_emit_insn_before (rtx, rtx);
5914 /* Map a bundle number to its pseudo-op. */
5916 const char *
5917 get_bundle_name (int b)
5919 return bundle_name[b];
5923 /* Return the maximum number of instructions a cpu can issue. */
5925 static int
5926 ia64_issue_rate (void)
5928 return 6;
5931 /* Helper function - like single_set, but look inside COND_EXEC. */
5933 static rtx
5934 ia64_single_set (rtx insn)
5936 rtx x = PATTERN (insn), ret;
5937 if (GET_CODE (x) == COND_EXEC)
5938 x = COND_EXEC_CODE (x);
5939 if (GET_CODE (x) == SET)
5940 return x;
5942 /* Special case here prologue_allocate_stack and epilogue_deallocate_stack.
5943 Although they are not classical single set, the second set is there just
5944 to protect it from moving past FP-relative stack accesses. */
5945 switch (recog_memoized (insn))
5947 case CODE_FOR_prologue_allocate_stack:
5948 case CODE_FOR_epilogue_deallocate_stack:
5949 ret = XVECEXP (x, 0, 0);
5950 break;
5952 default:
5953 ret = single_set_2 (insn, x);
5954 break;
5957 return ret;
5960 /* Adjust the cost of a scheduling dependency. Return the new cost of
5961 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
5963 static int
5964 ia64_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
5966 enum attr_itanium_class dep_class;
5967 enum attr_itanium_class insn_class;
5969 if (REG_NOTE_KIND (link) != REG_DEP_OUTPUT)
5970 return cost;
5972 insn_class = ia64_safe_itanium_class (insn);
5973 dep_class = ia64_safe_itanium_class (dep_insn);
5974 if (dep_class == ITANIUM_CLASS_ST || dep_class == ITANIUM_CLASS_STF
5975 || insn_class == ITANIUM_CLASS_ST || insn_class == ITANIUM_CLASS_STF)
5976 return 0;
5978 return cost;
5981 /* Like emit_insn_before, but skip cycle_display notes.
5982 ??? When cycle display notes are implemented, update this. */
5984 static void
5985 ia64_emit_insn_before (rtx insn, rtx before)
5987 emit_insn_before (insn, before);
5990 /* The following function marks insns who produce addresses for load
5991 and store insns. Such insns will be placed into M slots because it
5992 decrease latency time for Itanium1 (see function
5993 `ia64_produce_address_p' and the DFA descriptions). */
5995 static void
5996 ia64_dependencies_evaluation_hook (rtx head, rtx tail)
5998 rtx insn, link, next, next_tail;
6000 next_tail = NEXT_INSN (tail);
6001 for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
6002 if (INSN_P (insn))
6003 insn->call = 0;
6004 for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
6005 if (INSN_P (insn)
6006 && ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IALU)
6008 for (link = INSN_DEPEND (insn); link != 0; link = XEXP (link, 1))
6010 next = XEXP (link, 0);
6011 if ((ia64_safe_itanium_class (next) == ITANIUM_CLASS_ST
6012 || ia64_safe_itanium_class (next) == ITANIUM_CLASS_STF)
6013 && ia64_st_address_bypass_p (insn, next))
6014 break;
6015 else if ((ia64_safe_itanium_class (next) == ITANIUM_CLASS_LD
6016 || ia64_safe_itanium_class (next)
6017 == ITANIUM_CLASS_FLD)
6018 && ia64_ld_address_bypass_p (insn, next))
6019 break;
6021 insn->call = link != 0;
6025 /* We're beginning a new block. Initialize data structures as necessary. */
6027 static void
6028 ia64_sched_init (FILE *dump ATTRIBUTE_UNUSED,
6029 int sched_verbose ATTRIBUTE_UNUSED,
6030 int max_ready ATTRIBUTE_UNUSED)
6032 #ifdef ENABLE_CHECKING
6033 rtx insn;
6035 if (reload_completed)
6036 for (insn = NEXT_INSN (current_sched_info->prev_head);
6037 insn != current_sched_info->next_tail;
6038 insn = NEXT_INSN (insn))
6039 if (SCHED_GROUP_P (insn))
6040 abort ();
6041 #endif
6042 last_scheduled_insn = NULL_RTX;
6043 init_insn_group_barriers ();
6046 /* We are about to being issuing insns for this clock cycle.
6047 Override the default sort algorithm to better slot instructions. */
6049 static int
6050 ia64_dfa_sched_reorder (FILE *dump, int sched_verbose, rtx *ready,
6051 int *pn_ready, int clock_var ATTRIBUTE_UNUSED,
6052 int reorder_type)
6054 int n_asms;
6055 int n_ready = *pn_ready;
6056 rtx *e_ready = ready + n_ready;
6057 rtx *insnp;
6059 if (sched_verbose)
6060 fprintf (dump, "// ia64_dfa_sched_reorder (type %d):\n", reorder_type);
6062 if (reorder_type == 0)
6064 /* First, move all USEs, CLOBBERs and other crud out of the way. */
6065 n_asms = 0;
6066 for (insnp = ready; insnp < e_ready; insnp++)
6067 if (insnp < e_ready)
6069 rtx insn = *insnp;
6070 enum attr_type t = ia64_safe_type (insn);
6071 if (t == TYPE_UNKNOWN)
6073 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
6074 || asm_noperands (PATTERN (insn)) >= 0)
6076 rtx lowest = ready[n_asms];
6077 ready[n_asms] = insn;
6078 *insnp = lowest;
6079 n_asms++;
6081 else
6083 rtx highest = ready[n_ready - 1];
6084 ready[n_ready - 1] = insn;
6085 *insnp = highest;
6086 return 1;
6091 if (n_asms < n_ready)
6093 /* Some normal insns to process. Skip the asms. */
6094 ready += n_asms;
6095 n_ready -= n_asms;
6097 else if (n_ready > 0)
6098 return 1;
6101 if (ia64_final_schedule)
6103 int deleted = 0;
6104 int nr_need_stop = 0;
6106 for (insnp = ready; insnp < e_ready; insnp++)
6107 if (safe_group_barrier_needed_p (*insnp))
6108 nr_need_stop++;
6110 if (reorder_type == 1 && n_ready == nr_need_stop)
6111 return 0;
6112 if (reorder_type == 0)
6113 return 1;
6114 insnp = e_ready;
6115 /* Move down everything that needs a stop bit, preserving
6116 relative order. */
6117 while (insnp-- > ready + deleted)
6118 while (insnp >= ready + deleted)
6120 rtx insn = *insnp;
6121 if (! safe_group_barrier_needed_p (insn))
6122 break;
6123 memmove (ready + 1, ready, (insnp - ready) * sizeof (rtx));
6124 *ready = insn;
6125 deleted++;
6127 n_ready -= deleted;
6128 ready += deleted;
6131 return 1;
6134 /* We are about to being issuing insns for this clock cycle. Override
6135 the default sort algorithm to better slot instructions. */
6137 static int
6138 ia64_sched_reorder (FILE *dump, int sched_verbose, rtx *ready, int *pn_ready,
6139 int clock_var)
6141 return ia64_dfa_sched_reorder (dump, sched_verbose, ready,
6142 pn_ready, clock_var, 0);
6145 /* Like ia64_sched_reorder, but called after issuing each insn.
6146 Override the default sort algorithm to better slot instructions. */
6148 static int
6149 ia64_sched_reorder2 (FILE *dump ATTRIBUTE_UNUSED,
6150 int sched_verbose ATTRIBUTE_UNUSED, rtx *ready,
6151 int *pn_ready, int clock_var)
6153 if (ia64_tune == PROCESSOR_ITANIUM && reload_completed && last_scheduled_insn)
6154 clocks [INSN_UID (last_scheduled_insn)] = clock_var;
6155 return ia64_dfa_sched_reorder (dump, sched_verbose, ready, pn_ready,
6156 clock_var, 1);
6159 /* We are about to issue INSN. Return the number of insns left on the
6160 ready queue that can be issued this cycle. */
6162 static int
6163 ia64_variable_issue (FILE *dump ATTRIBUTE_UNUSED,
6164 int sched_verbose ATTRIBUTE_UNUSED,
6165 rtx insn ATTRIBUTE_UNUSED,
6166 int can_issue_more ATTRIBUTE_UNUSED)
6168 last_scheduled_insn = insn;
6169 memcpy (prev_cycle_state, curr_state, dfa_state_size);
6170 if (reload_completed)
6172 if (group_barrier_needed_p (insn))
6173 abort ();
6174 if (GET_CODE (insn) == CALL_INSN)
6175 init_insn_group_barriers ();
6176 stops_p [INSN_UID (insn)] = stop_before_p;
6177 stop_before_p = 0;
6179 return 1;
6182 /* We are choosing insn from the ready queue. Return nonzero if INSN
6183 can be chosen. */
6185 static int
6186 ia64_first_cycle_multipass_dfa_lookahead_guard (rtx insn)
6188 if (insn == NULL_RTX || !INSN_P (insn))
6189 abort ();
6190 return (!reload_completed
6191 || !safe_group_barrier_needed_p (insn));
6194 /* The following variable value is pseudo-insn used by the DFA insn
6195 scheduler to change the DFA state when the simulated clock is
6196 increased. */
6198 static rtx dfa_pre_cycle_insn;
6200 /* We are about to being issuing INSN. Return nonzero if we cannot
6201 issue it on given cycle CLOCK and return zero if we should not sort
6202 the ready queue on the next clock start. */
6204 static int
6205 ia64_dfa_new_cycle (FILE *dump, int verbose, rtx insn, int last_clock,
6206 int clock, int *sort_p)
6208 int setup_clocks_p = FALSE;
6210 if (insn == NULL_RTX || !INSN_P (insn))
6211 abort ();
6212 if ((reload_completed && safe_group_barrier_needed_p (insn))
6213 || (last_scheduled_insn
6214 && (GET_CODE (last_scheduled_insn) == CALL_INSN
6215 || GET_CODE (PATTERN (last_scheduled_insn)) == ASM_INPUT
6216 || asm_noperands (PATTERN (last_scheduled_insn)) >= 0)))
6218 init_insn_group_barriers ();
6219 if (verbose && dump)
6220 fprintf (dump, "// Stop should be before %d%s\n", INSN_UID (insn),
6221 last_clock == clock ? " + cycle advance" : "");
6222 stop_before_p = 1;
6223 if (last_clock == clock)
6225 state_transition (curr_state, dfa_stop_insn);
6226 if (TARGET_EARLY_STOP_BITS)
6227 *sort_p = (last_scheduled_insn == NULL_RTX
6228 || GET_CODE (last_scheduled_insn) != CALL_INSN);
6229 else
6230 *sort_p = 0;
6231 return 1;
6233 else if (reload_completed)
6234 setup_clocks_p = TRUE;
6235 if (GET_CODE (PATTERN (last_scheduled_insn)) == ASM_INPUT
6236 || asm_noperands (PATTERN (last_scheduled_insn)) >= 0)
6237 state_reset (curr_state);
6238 else
6240 memcpy (curr_state, prev_cycle_state, dfa_state_size);
6241 state_transition (curr_state, dfa_stop_insn);
6242 state_transition (curr_state, dfa_pre_cycle_insn);
6243 state_transition (curr_state, NULL);
6246 else if (reload_completed)
6247 setup_clocks_p = TRUE;
6248 if (setup_clocks_p && ia64_tune == PROCESSOR_ITANIUM
6249 && GET_CODE (PATTERN (insn)) != ASM_INPUT
6250 && asm_noperands (PATTERN (insn)) < 0)
6252 enum attr_itanium_class c = ia64_safe_itanium_class (insn);
6254 if (c != ITANIUM_CLASS_MMMUL && c != ITANIUM_CLASS_MMSHF)
6256 rtx link;
6257 int d = -1;
6259 for (link = LOG_LINKS (insn); link; link = XEXP (link, 1))
6260 if (REG_NOTE_KIND (link) == 0)
6262 enum attr_itanium_class dep_class;
6263 rtx dep_insn = XEXP (link, 0);
6265 dep_class = ia64_safe_itanium_class (dep_insn);
6266 if ((dep_class == ITANIUM_CLASS_MMMUL
6267 || dep_class == ITANIUM_CLASS_MMSHF)
6268 && last_clock - clocks [INSN_UID (dep_insn)] < 4
6269 && (d < 0
6270 || last_clock - clocks [INSN_UID (dep_insn)] < d))
6271 d = last_clock - clocks [INSN_UID (dep_insn)];
6273 if (d >= 0)
6274 add_cycles [INSN_UID (insn)] = 3 - d;
6277 return 0;
6282 /* The following page contains abstract data `bundle states' which are
6283 used for bundling insns (inserting nops and template generation). */
6285 /* The following describes state of insn bundling. */
6287 struct bundle_state
6289 /* Unique bundle state number to identify them in the debugging
6290 output */
6291 int unique_num;
6292 rtx insn; /* corresponding insn, NULL for the 1st and the last state */
6293 /* number nops before and after the insn */
6294 short before_nops_num, after_nops_num;
6295 int insn_num; /* insn number (0 - for initial state, 1 - for the 1st
6296 insn */
6297 int cost; /* cost of the state in cycles */
6298 int accumulated_insns_num; /* number of all previous insns including
6299 nops. L is considered as 2 insns */
6300 int branch_deviation; /* deviation of previous branches from 3rd slots */
6301 struct bundle_state *next; /* next state with the same insn_num */
6302 struct bundle_state *originator; /* originator (previous insn state) */
6303 /* All bundle states are in the following chain. */
6304 struct bundle_state *allocated_states_chain;
6305 /* The DFA State after issuing the insn and the nops. */
6306 state_t dfa_state;
6309 /* The following is map insn number to the corresponding bundle state. */
6311 static struct bundle_state **index_to_bundle_states;
6313 /* The unique number of next bundle state. */
6315 static int bundle_states_num;
6317 /* All allocated bundle states are in the following chain. */
6319 static struct bundle_state *allocated_bundle_states_chain;
6321 /* All allocated but not used bundle states are in the following
6322 chain. */
6324 static struct bundle_state *free_bundle_state_chain;
6327 /* The following function returns a free bundle state. */
6329 static struct bundle_state *
6330 get_free_bundle_state (void)
6332 struct bundle_state *result;
6334 if (free_bundle_state_chain != NULL)
6336 result = free_bundle_state_chain;
6337 free_bundle_state_chain = result->next;
6339 else
6341 result = xmalloc (sizeof (struct bundle_state));
6342 result->dfa_state = xmalloc (dfa_state_size);
6343 result->allocated_states_chain = allocated_bundle_states_chain;
6344 allocated_bundle_states_chain = result;
6346 result->unique_num = bundle_states_num++;
6347 return result;
6351 /* The following function frees given bundle state. */
6353 static void
6354 free_bundle_state (struct bundle_state *state)
6356 state->next = free_bundle_state_chain;
6357 free_bundle_state_chain = state;
6360 /* Start work with abstract data `bundle states'. */
6362 static void
6363 initiate_bundle_states (void)
6365 bundle_states_num = 0;
6366 free_bundle_state_chain = NULL;
6367 allocated_bundle_states_chain = NULL;
6370 /* Finish work with abstract data `bundle states'. */
6372 static void
6373 finish_bundle_states (void)
6375 struct bundle_state *curr_state, *next_state;
6377 for (curr_state = allocated_bundle_states_chain;
6378 curr_state != NULL;
6379 curr_state = next_state)
6381 next_state = curr_state->allocated_states_chain;
6382 free (curr_state->dfa_state);
6383 free (curr_state);
6387 /* Hash table of the bundle states. The key is dfa_state and insn_num
6388 of the bundle states. */
6390 static htab_t bundle_state_table;
6392 /* The function returns hash of BUNDLE_STATE. */
6394 static unsigned
6395 bundle_state_hash (const void *bundle_state)
6397 const struct bundle_state *state = (struct bundle_state *) bundle_state;
6398 unsigned result, i;
6400 for (result = i = 0; i < dfa_state_size; i++)
6401 result += (((unsigned char *) state->dfa_state) [i]
6402 << ((i % CHAR_BIT) * 3 + CHAR_BIT));
6403 return result + state->insn_num;
6406 /* The function returns nonzero if the bundle state keys are equal. */
6408 static int
6409 bundle_state_eq_p (const void *bundle_state_1, const void *bundle_state_2)
6411 const struct bundle_state * state1 = (struct bundle_state *) bundle_state_1;
6412 const struct bundle_state * state2 = (struct bundle_state *) bundle_state_2;
6414 return (state1->insn_num == state2->insn_num
6415 && memcmp (state1->dfa_state, state2->dfa_state,
6416 dfa_state_size) == 0);
6419 /* The function inserts the BUNDLE_STATE into the hash table. The
6420 function returns nonzero if the bundle has been inserted into the
6421 table. The table contains the best bundle state with given key. */
6423 static int
6424 insert_bundle_state (struct bundle_state *bundle_state)
6426 void **entry_ptr;
6428 entry_ptr = htab_find_slot (bundle_state_table, bundle_state, 1);
6429 if (*entry_ptr == NULL)
6431 bundle_state->next = index_to_bundle_states [bundle_state->insn_num];
6432 index_to_bundle_states [bundle_state->insn_num] = bundle_state;
6433 *entry_ptr = (void *) bundle_state;
6434 return TRUE;
6436 else if (bundle_state->cost < ((struct bundle_state *) *entry_ptr)->cost
6437 || (bundle_state->cost == ((struct bundle_state *) *entry_ptr)->cost
6438 && (((struct bundle_state *)*entry_ptr)->accumulated_insns_num
6439 > bundle_state->accumulated_insns_num
6440 || (((struct bundle_state *)
6441 *entry_ptr)->accumulated_insns_num
6442 == bundle_state->accumulated_insns_num
6443 && ((struct bundle_state *)
6444 *entry_ptr)->branch_deviation
6445 > bundle_state->branch_deviation))))
6448 struct bundle_state temp;
6450 temp = *(struct bundle_state *) *entry_ptr;
6451 *(struct bundle_state *) *entry_ptr = *bundle_state;
6452 ((struct bundle_state *) *entry_ptr)->next = temp.next;
6453 *bundle_state = temp;
6455 return FALSE;
6458 /* Start work with the hash table. */
6460 static void
6461 initiate_bundle_state_table (void)
6463 bundle_state_table = htab_create (50, bundle_state_hash, bundle_state_eq_p,
6464 (htab_del) 0);
6467 /* Finish work with the hash table. */
6469 static void
6470 finish_bundle_state_table (void)
6472 htab_delete (bundle_state_table);
6477 /* The following variable is a insn `nop' used to check bundle states
6478 with different number of inserted nops. */
6480 static rtx ia64_nop;
6482 /* The following function tries to issue NOPS_NUM nops for the current
6483 state without advancing processor cycle. If it failed, the
6484 function returns FALSE and frees the current state. */
6486 static int
6487 try_issue_nops (struct bundle_state *curr_state, int nops_num)
6489 int i;
6491 for (i = 0; i < nops_num; i++)
6492 if (state_transition (curr_state->dfa_state, ia64_nop) >= 0)
6494 free_bundle_state (curr_state);
6495 return FALSE;
6497 return TRUE;
6500 /* The following function tries to issue INSN for the current
6501 state without advancing processor cycle. If it failed, the
6502 function returns FALSE and frees the current state. */
6504 static int
6505 try_issue_insn (struct bundle_state *curr_state, rtx insn)
6507 if (insn && state_transition (curr_state->dfa_state, insn) >= 0)
6509 free_bundle_state (curr_state);
6510 return FALSE;
6512 return TRUE;
6515 /* The following function tries to issue BEFORE_NOPS_NUM nops and INSN
6516 starting with ORIGINATOR without advancing processor cycle. If
6517 TRY_BUNDLE_END_P is TRUE, the function also/only (if
6518 ONLY_BUNDLE_END_P is TRUE) tries to issue nops to fill all bundle.
6519 If it was successful, the function creates new bundle state and
6520 insert into the hash table and into `index_to_bundle_states'. */
6522 static void
6523 issue_nops_and_insn (struct bundle_state *originator, int before_nops_num,
6524 rtx insn, int try_bundle_end_p, int only_bundle_end_p)
6526 struct bundle_state *curr_state;
6528 curr_state = get_free_bundle_state ();
6529 memcpy (curr_state->dfa_state, originator->dfa_state, dfa_state_size);
6530 curr_state->insn = insn;
6531 curr_state->insn_num = originator->insn_num + 1;
6532 curr_state->cost = originator->cost;
6533 curr_state->originator = originator;
6534 curr_state->before_nops_num = before_nops_num;
6535 curr_state->after_nops_num = 0;
6536 curr_state->accumulated_insns_num
6537 = originator->accumulated_insns_num + before_nops_num;
6538 curr_state->branch_deviation = originator->branch_deviation;
6539 if (insn == NULL_RTX)
6540 abort ();
6541 else if (INSN_CODE (insn) == CODE_FOR_insn_group_barrier)
6543 if (GET_MODE (insn) == TImode)
6544 abort ();
6545 if (!try_issue_nops (curr_state, before_nops_num))
6546 return;
6547 if (!try_issue_insn (curr_state, insn))
6548 return;
6549 memcpy (temp_dfa_state, curr_state->dfa_state, dfa_state_size);
6550 if (state_transition (temp_dfa_state, dfa_pre_cycle_insn) >= 0
6551 && curr_state->accumulated_insns_num % 3 != 0)
6553 free_bundle_state (curr_state);
6554 return;
6557 else if (GET_MODE (insn) != TImode)
6559 if (!try_issue_nops (curr_state, before_nops_num))
6560 return;
6561 if (!try_issue_insn (curr_state, insn))
6562 return;
6563 curr_state->accumulated_insns_num++;
6564 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
6565 || asm_noperands (PATTERN (insn)) >= 0)
6566 abort ();
6567 if (ia64_safe_type (insn) == TYPE_L)
6568 curr_state->accumulated_insns_num++;
6570 else
6572 state_transition (curr_state->dfa_state, dfa_pre_cycle_insn);
6573 state_transition (curr_state->dfa_state, NULL);
6574 curr_state->cost++;
6575 if (!try_issue_nops (curr_state, before_nops_num))
6576 return;
6577 if (!try_issue_insn (curr_state, insn))
6578 return;
6579 curr_state->accumulated_insns_num++;
6580 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
6581 || asm_noperands (PATTERN (insn)) >= 0)
6583 /* Finish bundle containing asm insn. */
6584 curr_state->after_nops_num
6585 = 3 - curr_state->accumulated_insns_num % 3;
6586 curr_state->accumulated_insns_num
6587 += 3 - curr_state->accumulated_insns_num % 3;
6589 else if (ia64_safe_type (insn) == TYPE_L)
6590 curr_state->accumulated_insns_num++;
6592 if (ia64_safe_type (insn) == TYPE_B)
6593 curr_state->branch_deviation
6594 += 2 - (curr_state->accumulated_insns_num - 1) % 3;
6595 if (try_bundle_end_p && curr_state->accumulated_insns_num % 3 != 0)
6597 if (!only_bundle_end_p && insert_bundle_state (curr_state))
6599 state_t dfa_state;
6600 struct bundle_state *curr_state1;
6601 struct bundle_state *allocated_states_chain;
6603 curr_state1 = get_free_bundle_state ();
6604 dfa_state = curr_state1->dfa_state;
6605 allocated_states_chain = curr_state1->allocated_states_chain;
6606 *curr_state1 = *curr_state;
6607 curr_state1->dfa_state = dfa_state;
6608 curr_state1->allocated_states_chain = allocated_states_chain;
6609 memcpy (curr_state1->dfa_state, curr_state->dfa_state,
6610 dfa_state_size);
6611 curr_state = curr_state1;
6613 if (!try_issue_nops (curr_state,
6614 3 - curr_state->accumulated_insns_num % 3))
6615 return;
6616 curr_state->after_nops_num
6617 = 3 - curr_state->accumulated_insns_num % 3;
6618 curr_state->accumulated_insns_num
6619 += 3 - curr_state->accumulated_insns_num % 3;
6621 if (!insert_bundle_state (curr_state))
6622 free_bundle_state (curr_state);
6623 return;
6626 /* The following function returns position in the two window bundle
6627 for given STATE. */
6629 static int
6630 get_max_pos (state_t state)
6632 if (cpu_unit_reservation_p (state, pos_6))
6633 return 6;
6634 else if (cpu_unit_reservation_p (state, pos_5))
6635 return 5;
6636 else if (cpu_unit_reservation_p (state, pos_4))
6637 return 4;
6638 else if (cpu_unit_reservation_p (state, pos_3))
6639 return 3;
6640 else if (cpu_unit_reservation_p (state, pos_2))
6641 return 2;
6642 else if (cpu_unit_reservation_p (state, pos_1))
6643 return 1;
6644 else
6645 return 0;
6648 /* The function returns code of a possible template for given position
6649 and state. The function should be called only with 2 values of
6650 position equal to 3 or 6. */
6652 static int
6653 get_template (state_t state, int pos)
6655 switch (pos)
6657 case 3:
6658 if (cpu_unit_reservation_p (state, _0mii_))
6659 return 0;
6660 else if (cpu_unit_reservation_p (state, _0mmi_))
6661 return 1;
6662 else if (cpu_unit_reservation_p (state, _0mfi_))
6663 return 2;
6664 else if (cpu_unit_reservation_p (state, _0mmf_))
6665 return 3;
6666 else if (cpu_unit_reservation_p (state, _0bbb_))
6667 return 4;
6668 else if (cpu_unit_reservation_p (state, _0mbb_))
6669 return 5;
6670 else if (cpu_unit_reservation_p (state, _0mib_))
6671 return 6;
6672 else if (cpu_unit_reservation_p (state, _0mmb_))
6673 return 7;
6674 else if (cpu_unit_reservation_p (state, _0mfb_))
6675 return 8;
6676 else if (cpu_unit_reservation_p (state, _0mlx_))
6677 return 9;
6678 else
6679 abort ();
6680 case 6:
6681 if (cpu_unit_reservation_p (state, _1mii_))
6682 return 0;
6683 else if (cpu_unit_reservation_p (state, _1mmi_))
6684 return 1;
6685 else if (cpu_unit_reservation_p (state, _1mfi_))
6686 return 2;
6687 else if (_1mmf_ >= 0 && cpu_unit_reservation_p (state, _1mmf_))
6688 return 3;
6689 else if (cpu_unit_reservation_p (state, _1bbb_))
6690 return 4;
6691 else if (cpu_unit_reservation_p (state, _1mbb_))
6692 return 5;
6693 else if (cpu_unit_reservation_p (state, _1mib_))
6694 return 6;
6695 else if (cpu_unit_reservation_p (state, _1mmb_))
6696 return 7;
6697 else if (cpu_unit_reservation_p (state, _1mfb_))
6698 return 8;
6699 else if (cpu_unit_reservation_p (state, _1mlx_))
6700 return 9;
6701 else
6702 abort ();
6703 default:
6704 abort ();
6708 /* The following function returns an insn important for insn bundling
6709 followed by INSN and before TAIL. */
6711 static rtx
6712 get_next_important_insn (rtx insn, rtx tail)
6714 for (; insn && insn != tail; insn = NEXT_INSN (insn))
6715 if (INSN_P (insn)
6716 && ia64_safe_itanium_class (insn) != ITANIUM_CLASS_IGNORE
6717 && GET_CODE (PATTERN (insn)) != USE
6718 && GET_CODE (PATTERN (insn)) != CLOBBER)
6719 return insn;
6720 return NULL_RTX;
6723 /* The following function does insn bundling. Bundling means
6724 inserting templates and nop insns to fit insn groups into permitted
6725 templates. Instruction scheduling uses NDFA (non-deterministic
6726 finite automata) encoding informations about the templates and the
6727 inserted nops. Nondeterminism of the automata permits follows
6728 all possible insn sequences very fast.
6730 Unfortunately it is not possible to get information about inserting
6731 nop insns and used templates from the automata states. The
6732 automata only says that we can issue an insn possibly inserting
6733 some nops before it and using some template. Therefore insn
6734 bundling in this function is implemented by using DFA
6735 (deterministic finite automata). We follows all possible insn
6736 sequences by inserting 0-2 nops (that is what the NDFA describe for
6737 insn scheduling) before/after each insn being bundled. We know the
6738 start of simulated processor cycle from insn scheduling (insn
6739 starting a new cycle has TImode).
6741 Simple implementation of insn bundling would create enormous
6742 number of possible insn sequences satisfying information about new
6743 cycle ticks taken from the insn scheduling. To make the algorithm
6744 practical we use dynamic programming. Each decision (about
6745 inserting nops and implicitly about previous decisions) is described
6746 by structure bundle_state (see above). If we generate the same
6747 bundle state (key is automaton state after issuing the insns and
6748 nops for it), we reuse already generated one. As consequence we
6749 reject some decisions which cannot improve the solution and
6750 reduce memory for the algorithm.
6752 When we reach the end of EBB (extended basic block), we choose the
6753 best sequence and then, moving back in EBB, insert templates for
6754 the best alternative. The templates are taken from querying
6755 automaton state for each insn in chosen bundle states.
6757 So the algorithm makes two (forward and backward) passes through
6758 EBB. There is an additional forward pass through EBB for Itanium1
6759 processor. This pass inserts more nops to make dependency between
6760 a producer insn and MMMUL/MMSHF at least 4 cycles long. */
6762 static void
6763 bundling (FILE *dump, int verbose, rtx prev_head_insn, rtx tail)
6765 struct bundle_state *curr_state, *next_state, *best_state;
6766 rtx insn, next_insn;
6767 int insn_num;
6768 int i, bundle_end_p, only_bundle_end_p, asm_p;
6769 int pos = 0, max_pos, template0, template1;
6770 rtx b;
6771 rtx nop;
6772 enum attr_type type;
6774 insn_num = 0;
6775 /* Count insns in the EBB. */
6776 for (insn = NEXT_INSN (prev_head_insn);
6777 insn && insn != tail;
6778 insn = NEXT_INSN (insn))
6779 if (INSN_P (insn))
6780 insn_num++;
6781 if (insn_num == 0)
6782 return;
6783 bundling_p = 1;
6784 dfa_clean_insn_cache ();
6785 initiate_bundle_state_table ();
6786 index_to_bundle_states = xmalloc ((insn_num + 2)
6787 * sizeof (struct bundle_state *));
6788 /* First (forward) pass -- generation of bundle states. */
6789 curr_state = get_free_bundle_state ();
6790 curr_state->insn = NULL;
6791 curr_state->before_nops_num = 0;
6792 curr_state->after_nops_num = 0;
6793 curr_state->insn_num = 0;
6794 curr_state->cost = 0;
6795 curr_state->accumulated_insns_num = 0;
6796 curr_state->branch_deviation = 0;
6797 curr_state->next = NULL;
6798 curr_state->originator = NULL;
6799 state_reset (curr_state->dfa_state);
6800 index_to_bundle_states [0] = curr_state;
6801 insn_num = 0;
6802 /* Shift cycle mark if it is put on insn which could be ignored. */
6803 for (insn = NEXT_INSN (prev_head_insn);
6804 insn != tail;
6805 insn = NEXT_INSN (insn))
6806 if (INSN_P (insn)
6807 && (ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IGNORE
6808 || GET_CODE (PATTERN (insn)) == USE
6809 || GET_CODE (PATTERN (insn)) == CLOBBER)
6810 && GET_MODE (insn) == TImode)
6812 PUT_MODE (insn, VOIDmode);
6813 for (next_insn = NEXT_INSN (insn);
6814 next_insn != tail;
6815 next_insn = NEXT_INSN (next_insn))
6816 if (INSN_P (next_insn)
6817 && ia64_safe_itanium_class (next_insn) != ITANIUM_CLASS_IGNORE
6818 && GET_CODE (PATTERN (next_insn)) != USE
6819 && GET_CODE (PATTERN (next_insn)) != CLOBBER)
6821 PUT_MODE (next_insn, TImode);
6822 break;
6825 /* Froward pass: generation of bundle states. */
6826 for (insn = get_next_important_insn (NEXT_INSN (prev_head_insn), tail);
6827 insn != NULL_RTX;
6828 insn = next_insn)
6830 if (!INSN_P (insn)
6831 || ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IGNORE
6832 || GET_CODE (PATTERN (insn)) == USE
6833 || GET_CODE (PATTERN (insn)) == CLOBBER)
6834 abort ();
6835 type = ia64_safe_type (insn);
6836 next_insn = get_next_important_insn (NEXT_INSN (insn), tail);
6837 insn_num++;
6838 index_to_bundle_states [insn_num] = NULL;
6839 for (curr_state = index_to_bundle_states [insn_num - 1];
6840 curr_state != NULL;
6841 curr_state = next_state)
6843 pos = curr_state->accumulated_insns_num % 3;
6844 next_state = curr_state->next;
6845 /* We must fill up the current bundle in order to start a
6846 subsequent asm insn in a new bundle. Asm insn is always
6847 placed in a separate bundle. */
6848 only_bundle_end_p
6849 = (next_insn != NULL_RTX
6850 && INSN_CODE (insn) == CODE_FOR_insn_group_barrier
6851 && ia64_safe_type (next_insn) == TYPE_UNKNOWN);
6852 /* We may fill up the current bundle if it is the cycle end
6853 without a group barrier. */
6854 bundle_end_p
6855 = (only_bundle_end_p || next_insn == NULL_RTX
6856 || (GET_MODE (next_insn) == TImode
6857 && INSN_CODE (insn) != CODE_FOR_insn_group_barrier));
6858 if (type == TYPE_F || type == TYPE_B || type == TYPE_L
6859 || type == TYPE_S
6860 /* We need to insert 2 nops for cases like M_MII. To
6861 guarantee issuing all insns on the same cycle for
6862 Itanium 1, we need to issue 2 nops after the first M
6863 insn (MnnMII where n is a nop insn). */
6864 || ((type == TYPE_M || type == TYPE_A)
6865 && ia64_tune == PROCESSOR_ITANIUM
6866 && !bundle_end_p && pos == 1))
6867 issue_nops_and_insn (curr_state, 2, insn, bundle_end_p,
6868 only_bundle_end_p);
6869 issue_nops_and_insn (curr_state, 1, insn, bundle_end_p,
6870 only_bundle_end_p);
6871 issue_nops_and_insn (curr_state, 0, insn, bundle_end_p,
6872 only_bundle_end_p);
6874 if (index_to_bundle_states [insn_num] == NULL)
6875 abort ();
6876 for (curr_state = index_to_bundle_states [insn_num];
6877 curr_state != NULL;
6878 curr_state = curr_state->next)
6879 if (verbose >= 2 && dump)
6881 /* This structure is taken from generated code of the
6882 pipeline hazard recognizer (see file insn-attrtab.c).
6883 Please don't forget to change the structure if a new
6884 automaton is added to .md file. */
6885 struct DFA_chip
6887 unsigned short one_automaton_state;
6888 unsigned short oneb_automaton_state;
6889 unsigned short two_automaton_state;
6890 unsigned short twob_automaton_state;
6893 fprintf
6894 (dump,
6895 "// Bundle state %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, state %d) for %d\n",
6896 curr_state->unique_num,
6897 (curr_state->originator == NULL
6898 ? -1 : curr_state->originator->unique_num),
6899 curr_state->cost,
6900 curr_state->before_nops_num, curr_state->after_nops_num,
6901 curr_state->accumulated_insns_num, curr_state->branch_deviation,
6902 (ia64_tune == PROCESSOR_ITANIUM
6903 ? ((struct DFA_chip *) curr_state->dfa_state)->oneb_automaton_state
6904 : ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state),
6905 INSN_UID (insn));
6908 if (index_to_bundle_states [insn_num] == NULL)
6909 /* We should find a solution because the 2nd insn scheduling has
6910 found one. */
6911 abort ();
6912 /* Find a state corresponding to the best insn sequence. */
6913 best_state = NULL;
6914 for (curr_state = index_to_bundle_states [insn_num];
6915 curr_state != NULL;
6916 curr_state = curr_state->next)
6917 /* We are just looking at the states with fully filled up last
6918 bundle. The first we prefer insn sequences with minimal cost
6919 then with minimal inserted nops and finally with branch insns
6920 placed in the 3rd slots. */
6921 if (curr_state->accumulated_insns_num % 3 == 0
6922 && (best_state == NULL || best_state->cost > curr_state->cost
6923 || (best_state->cost == curr_state->cost
6924 && (curr_state->accumulated_insns_num
6925 < best_state->accumulated_insns_num
6926 || (curr_state->accumulated_insns_num
6927 == best_state->accumulated_insns_num
6928 && curr_state->branch_deviation
6929 < best_state->branch_deviation)))))
6930 best_state = curr_state;
6931 /* Second (backward) pass: adding nops and templates. */
6932 insn_num = best_state->before_nops_num;
6933 template0 = template1 = -1;
6934 for (curr_state = best_state;
6935 curr_state->originator != NULL;
6936 curr_state = curr_state->originator)
6938 insn = curr_state->insn;
6939 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
6940 || asm_noperands (PATTERN (insn)) >= 0);
6941 insn_num++;
6942 if (verbose >= 2 && dump)
6944 struct DFA_chip
6946 unsigned short one_automaton_state;
6947 unsigned short oneb_automaton_state;
6948 unsigned short two_automaton_state;
6949 unsigned short twob_automaton_state;
6952 fprintf
6953 (dump,
6954 "// Best %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, state %d) for %d\n",
6955 curr_state->unique_num,
6956 (curr_state->originator == NULL
6957 ? -1 : curr_state->originator->unique_num),
6958 curr_state->cost,
6959 curr_state->before_nops_num, curr_state->after_nops_num,
6960 curr_state->accumulated_insns_num, curr_state->branch_deviation,
6961 (ia64_tune == PROCESSOR_ITANIUM
6962 ? ((struct DFA_chip *) curr_state->dfa_state)->oneb_automaton_state
6963 : ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state),
6964 INSN_UID (insn));
6966 /* Find the position in the current bundle window. The window can
6967 contain at most two bundles. Two bundle window means that
6968 the processor will make two bundle rotation. */
6969 max_pos = get_max_pos (curr_state->dfa_state);
6970 if (max_pos == 6
6971 /* The following (negative template number) means that the
6972 processor did one bundle rotation. */
6973 || (max_pos == 3 && template0 < 0))
6975 /* We are at the end of the window -- find template(s) for
6976 its bundle(s). */
6977 pos = max_pos;
6978 if (max_pos == 3)
6979 template0 = get_template (curr_state->dfa_state, 3);
6980 else
6982 template1 = get_template (curr_state->dfa_state, 3);
6983 template0 = get_template (curr_state->dfa_state, 6);
6986 if (max_pos > 3 && template1 < 0)
6987 /* It may happen when we have the stop inside a bundle. */
6989 if (pos > 3)
6990 abort ();
6991 template1 = get_template (curr_state->dfa_state, 3);
6992 pos += 3;
6994 if (!asm_p)
6995 /* Emit nops after the current insn. */
6996 for (i = 0; i < curr_state->after_nops_num; i++)
6998 nop = gen_nop ();
6999 emit_insn_after (nop, insn);
7000 pos--;
7001 if (pos < 0)
7002 abort ();
7003 if (pos % 3 == 0)
7005 /* We are at the start of a bundle: emit the template
7006 (it should be defined). */
7007 if (template0 < 0)
7008 abort ();
7009 b = gen_bundle_selector (GEN_INT (template0));
7010 ia64_emit_insn_before (b, nop);
7011 /* If we have two bundle window, we make one bundle
7012 rotation. Otherwise template0 will be undefined
7013 (negative value). */
7014 template0 = template1;
7015 template1 = -1;
7018 /* Move the position backward in the window. Group barrier has
7019 no slot. Asm insn takes all bundle. */
7020 if (INSN_CODE (insn) != CODE_FOR_insn_group_barrier
7021 && GET_CODE (PATTERN (insn)) != ASM_INPUT
7022 && asm_noperands (PATTERN (insn)) < 0)
7023 pos--;
7024 /* Long insn takes 2 slots. */
7025 if (ia64_safe_type (insn) == TYPE_L)
7026 pos--;
7027 if (pos < 0)
7028 abort ();
7029 if (pos % 3 == 0
7030 && INSN_CODE (insn) != CODE_FOR_insn_group_barrier
7031 && GET_CODE (PATTERN (insn)) != ASM_INPUT
7032 && asm_noperands (PATTERN (insn)) < 0)
7034 /* The current insn is at the bundle start: emit the
7035 template. */
7036 if (template0 < 0)
7037 abort ();
7038 b = gen_bundle_selector (GEN_INT (template0));
7039 ia64_emit_insn_before (b, insn);
7040 b = PREV_INSN (insn);
7041 insn = b;
7042 /* See comment above in analogous place for emitting nops
7043 after the insn. */
7044 template0 = template1;
7045 template1 = -1;
7047 /* Emit nops after the current insn. */
7048 for (i = 0; i < curr_state->before_nops_num; i++)
7050 nop = gen_nop ();
7051 ia64_emit_insn_before (nop, insn);
7052 nop = PREV_INSN (insn);
7053 insn = nop;
7054 pos--;
7055 if (pos < 0)
7056 abort ();
7057 if (pos % 3 == 0)
7059 /* See comment above in analogous place for emitting nops
7060 after the insn. */
7061 if (template0 < 0)
7062 abort ();
7063 b = gen_bundle_selector (GEN_INT (template0));
7064 ia64_emit_insn_before (b, insn);
7065 b = PREV_INSN (insn);
7066 insn = b;
7067 template0 = template1;
7068 template1 = -1;
7072 if (ia64_tune == PROCESSOR_ITANIUM)
7073 /* Insert additional cycles for MM-insns (MMMUL and MMSHF).
7074 Itanium1 has a strange design, if the distance between an insn
7075 and dependent MM-insn is less 4 then we have a 6 additional
7076 cycles stall. So we make the distance equal to 4 cycles if it
7077 is less. */
7078 for (insn = get_next_important_insn (NEXT_INSN (prev_head_insn), tail);
7079 insn != NULL_RTX;
7080 insn = next_insn)
7082 if (!INSN_P (insn)
7083 || ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IGNORE
7084 || GET_CODE (PATTERN (insn)) == USE
7085 || GET_CODE (PATTERN (insn)) == CLOBBER)
7086 abort ();
7087 next_insn = get_next_important_insn (NEXT_INSN (insn), tail);
7088 if (INSN_UID (insn) < clocks_length && add_cycles [INSN_UID (insn)])
7089 /* We found a MM-insn which needs additional cycles. */
7091 rtx last;
7092 int i, j, n;
7093 int pred_stop_p;
7095 /* Now we are searching for a template of the bundle in
7096 which the MM-insn is placed and the position of the
7097 insn in the bundle (0, 1, 2). Also we are searching
7098 for that there is a stop before the insn. */
7099 last = prev_active_insn (insn);
7100 pred_stop_p = recog_memoized (last) == CODE_FOR_insn_group_barrier;
7101 if (pred_stop_p)
7102 last = prev_active_insn (last);
7103 n = 0;
7104 for (;; last = prev_active_insn (last))
7105 if (recog_memoized (last) == CODE_FOR_bundle_selector)
7107 template0 = XINT (XVECEXP (PATTERN (last), 0, 0), 0);
7108 if (template0 == 9)
7109 /* The insn is in MLX bundle. Change the template
7110 onto MFI because we will add nops before the
7111 insn. It simplifies subsequent code a lot. */
7112 PATTERN (last)
7113 = gen_bundle_selector (const2_rtx); /* -> MFI */
7114 break;
7116 else if (recog_memoized (last) != CODE_FOR_insn_group_barrier
7117 && (ia64_safe_itanium_class (last)
7118 != ITANIUM_CLASS_IGNORE))
7119 n++;
7120 /* Some check of correctness: the stop is not at the
7121 bundle start, there are no more 3 insns in the bundle,
7122 and the MM-insn is not at the start of bundle with
7123 template MLX. */
7124 if ((pred_stop_p && n == 0) || n > 2
7125 || (template0 == 9 && n != 0))
7126 abort ();
7127 /* Put nops after the insn in the bundle. */
7128 for (j = 3 - n; j > 0; j --)
7129 ia64_emit_insn_before (gen_nop (), insn);
7130 /* It takes into account that we will add more N nops
7131 before the insn lately -- please see code below. */
7132 add_cycles [INSN_UID (insn)]--;
7133 if (!pred_stop_p || add_cycles [INSN_UID (insn)])
7134 ia64_emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
7135 insn);
7136 if (pred_stop_p)
7137 add_cycles [INSN_UID (insn)]--;
7138 for (i = add_cycles [INSN_UID (insn)]; i > 0; i--)
7140 /* Insert "MII;" template. */
7141 ia64_emit_insn_before (gen_bundle_selector (const0_rtx),
7142 insn);
7143 ia64_emit_insn_before (gen_nop (), insn);
7144 ia64_emit_insn_before (gen_nop (), insn);
7145 if (i > 1)
7147 /* To decrease code size, we use "MI;I;"
7148 template. */
7149 ia64_emit_insn_before
7150 (gen_insn_group_barrier (GEN_INT (3)), insn);
7151 i--;
7153 ia64_emit_insn_before (gen_nop (), insn);
7154 ia64_emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
7155 insn);
7157 /* Put the MM-insn in the same slot of a bundle with the
7158 same template as the original one. */
7159 ia64_emit_insn_before (gen_bundle_selector (GEN_INT (template0)),
7160 insn);
7161 /* To put the insn in the same slot, add necessary number
7162 of nops. */
7163 for (j = n; j > 0; j --)
7164 ia64_emit_insn_before (gen_nop (), insn);
7165 /* Put the stop if the original bundle had it. */
7166 if (pred_stop_p)
7167 ia64_emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
7168 insn);
7171 free (index_to_bundle_states);
7172 finish_bundle_state_table ();
7173 bundling_p = 0;
7174 dfa_clean_insn_cache ();
7177 /* The following function is called at the end of scheduling BB or
7178 EBB. After reload, it inserts stop bits and does insn bundling. */
7180 static void
7181 ia64_sched_finish (FILE *dump, int sched_verbose)
7183 if (sched_verbose)
7184 fprintf (dump, "// Finishing schedule.\n");
7185 if (!reload_completed)
7186 return;
7187 if (reload_completed)
7189 final_emit_insn_group_barriers (dump);
7190 bundling (dump, sched_verbose, current_sched_info->prev_head,
7191 current_sched_info->next_tail);
7192 if (sched_verbose && dump)
7193 fprintf (dump, "// finishing %d-%d\n",
7194 INSN_UID (NEXT_INSN (current_sched_info->prev_head)),
7195 INSN_UID (PREV_INSN (current_sched_info->next_tail)));
7197 return;
7201 /* The following function inserts stop bits in scheduled BB or EBB. */
7203 static void
7204 final_emit_insn_group_barriers (FILE *dump ATTRIBUTE_UNUSED)
7206 rtx insn;
7207 int need_barrier_p = 0;
7208 rtx prev_insn = NULL_RTX;
7210 init_insn_group_barriers ();
7212 for (insn = NEXT_INSN (current_sched_info->prev_head);
7213 insn != current_sched_info->next_tail;
7214 insn = NEXT_INSN (insn))
7216 if (GET_CODE (insn) == BARRIER)
7218 rtx last = prev_active_insn (insn);
7220 if (! last)
7221 continue;
7222 if (GET_CODE (last) == JUMP_INSN
7223 && GET_CODE (PATTERN (last)) == ADDR_DIFF_VEC)
7224 last = prev_active_insn (last);
7225 if (recog_memoized (last) != CODE_FOR_insn_group_barrier)
7226 emit_insn_after (gen_insn_group_barrier (GEN_INT (3)), last);
7228 init_insn_group_barriers ();
7229 need_barrier_p = 0;
7230 prev_insn = NULL_RTX;
7232 else if (INSN_P (insn))
7234 if (recog_memoized (insn) == CODE_FOR_insn_group_barrier)
7236 init_insn_group_barriers ();
7237 need_barrier_p = 0;
7238 prev_insn = NULL_RTX;
7240 else if (need_barrier_p || group_barrier_needed_p (insn))
7242 if (TARGET_EARLY_STOP_BITS)
7244 rtx last;
7246 for (last = insn;
7247 last != current_sched_info->prev_head;
7248 last = PREV_INSN (last))
7249 if (INSN_P (last) && GET_MODE (last) == TImode
7250 && stops_p [INSN_UID (last)])
7251 break;
7252 if (last == current_sched_info->prev_head)
7253 last = insn;
7254 last = prev_active_insn (last);
7255 if (last
7256 && recog_memoized (last) != CODE_FOR_insn_group_barrier)
7257 emit_insn_after (gen_insn_group_barrier (GEN_INT (3)),
7258 last);
7259 init_insn_group_barriers ();
7260 for (last = NEXT_INSN (last);
7261 last != insn;
7262 last = NEXT_INSN (last))
7263 if (INSN_P (last))
7264 group_barrier_needed_p (last);
7266 else
7268 emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
7269 insn);
7270 init_insn_group_barriers ();
7272 group_barrier_needed_p (insn);
7273 prev_insn = NULL_RTX;
7275 else if (recog_memoized (insn) >= 0)
7276 prev_insn = insn;
7277 need_barrier_p = (GET_CODE (insn) == CALL_INSN
7278 || GET_CODE (PATTERN (insn)) == ASM_INPUT
7279 || asm_noperands (PATTERN (insn)) >= 0);
7286 /* If the following function returns TRUE, we will use the the DFA
7287 insn scheduler. */
7289 static int
7290 ia64_first_cycle_multipass_dfa_lookahead (void)
7292 return (reload_completed ? 6 : 4);
7295 /* The following function initiates variable `dfa_pre_cycle_insn'. */
7297 static void
7298 ia64_init_dfa_pre_cycle_insn (void)
7300 if (temp_dfa_state == NULL)
7302 dfa_state_size = state_size ();
7303 temp_dfa_state = xmalloc (dfa_state_size);
7304 prev_cycle_state = xmalloc (dfa_state_size);
7306 dfa_pre_cycle_insn = make_insn_raw (gen_pre_cycle ());
7307 PREV_INSN (dfa_pre_cycle_insn) = NEXT_INSN (dfa_pre_cycle_insn) = NULL_RTX;
7308 recog_memoized (dfa_pre_cycle_insn);
7309 dfa_stop_insn = make_insn_raw (gen_insn_group_barrier (GEN_INT (3)));
7310 PREV_INSN (dfa_stop_insn) = NEXT_INSN (dfa_stop_insn) = NULL_RTX;
7311 recog_memoized (dfa_stop_insn);
7314 /* The following function returns the pseudo insn DFA_PRE_CYCLE_INSN
7315 used by the DFA insn scheduler. */
7317 static rtx
7318 ia64_dfa_pre_cycle_insn (void)
7320 return dfa_pre_cycle_insn;
7323 /* The following function returns TRUE if PRODUCER (of type ilog or
7324 ld) produces address for CONSUMER (of type st or stf). */
7327 ia64_st_address_bypass_p (rtx producer, rtx consumer)
7329 rtx dest, reg, mem;
7331 if (producer == NULL_RTX || consumer == NULL_RTX)
7332 abort ();
7333 dest = ia64_single_set (producer);
7334 if (dest == NULL_RTX || (reg = SET_DEST (dest)) == NULL_RTX
7335 || (GET_CODE (reg) != REG && GET_CODE (reg) != SUBREG))
7336 abort ();
7337 if (GET_CODE (reg) == SUBREG)
7338 reg = SUBREG_REG (reg);
7339 dest = ia64_single_set (consumer);
7340 if (dest == NULL_RTX || (mem = SET_DEST (dest)) == NULL_RTX
7341 || GET_CODE (mem) != MEM)
7342 abort ();
7343 return reg_mentioned_p (reg, mem);
7346 /* The following function returns TRUE if PRODUCER (of type ilog or
7347 ld) produces address for CONSUMER (of type ld or fld). */
7350 ia64_ld_address_bypass_p (rtx producer, rtx consumer)
7352 rtx dest, src, reg, mem;
7354 if (producer == NULL_RTX || consumer == NULL_RTX)
7355 abort ();
7356 dest = ia64_single_set (producer);
7357 if (dest == NULL_RTX || (reg = SET_DEST (dest)) == NULL_RTX
7358 || (GET_CODE (reg) != REG && GET_CODE (reg) != SUBREG))
7359 abort ();
7360 if (GET_CODE (reg) == SUBREG)
7361 reg = SUBREG_REG (reg);
7362 src = ia64_single_set (consumer);
7363 if (src == NULL_RTX || (mem = SET_SRC (src)) == NULL_RTX)
7364 abort ();
7365 if (GET_CODE (mem) == UNSPEC && XVECLEN (mem, 0) > 0)
7366 mem = XVECEXP (mem, 0, 0);
7367 while (GET_CODE (mem) == SUBREG || GET_CODE (mem) == ZERO_EXTEND)
7368 mem = XEXP (mem, 0);
7370 /* Note that LO_SUM is used for GOT loads. */
7371 if (GET_CODE (mem) != LO_SUM && GET_CODE (mem) != MEM)
7372 abort ();
7374 return reg_mentioned_p (reg, mem);
7377 /* The following function returns TRUE if INSN produces address for a
7378 load/store insn. We will place such insns into M slot because it
7379 decreases its latency time. */
7382 ia64_produce_address_p (rtx insn)
7384 return insn->call;
7388 /* Emit pseudo-ops for the assembler to describe predicate relations.
7389 At present this assumes that we only consider predicate pairs to
7390 be mutex, and that the assembler can deduce proper values from
7391 straight-line code. */
7393 static void
7394 emit_predicate_relation_info (void)
7396 basic_block bb;
7398 FOR_EACH_BB_REVERSE (bb)
7400 int r;
7401 rtx head = BB_HEAD (bb);
7403 /* We only need such notes at code labels. */
7404 if (GET_CODE (head) != CODE_LABEL)
7405 continue;
7406 if (GET_CODE (NEXT_INSN (head)) == NOTE
7407 && NOTE_LINE_NUMBER (NEXT_INSN (head)) == NOTE_INSN_BASIC_BLOCK)
7408 head = NEXT_INSN (head);
7410 for (r = PR_REG (0); r < PR_REG (64); r += 2)
7411 if (REGNO_REG_SET_P (bb->global_live_at_start, r))
7413 rtx p = gen_rtx_REG (BImode, r);
7414 rtx n = emit_insn_after (gen_pred_rel_mutex (p), head);
7415 if (head == BB_END (bb))
7416 BB_END (bb) = n;
7417 head = n;
7421 /* Look for conditional calls that do not return, and protect predicate
7422 relations around them. Otherwise the assembler will assume the call
7423 returns, and complain about uses of call-clobbered predicates after
7424 the call. */
7425 FOR_EACH_BB_REVERSE (bb)
7427 rtx insn = BB_HEAD (bb);
7429 while (1)
7431 if (GET_CODE (insn) == CALL_INSN
7432 && GET_CODE (PATTERN (insn)) == COND_EXEC
7433 && find_reg_note (insn, REG_NORETURN, NULL_RTX))
7435 rtx b = emit_insn_before (gen_safe_across_calls_all (), insn);
7436 rtx a = emit_insn_after (gen_safe_across_calls_normal (), insn);
7437 if (BB_HEAD (bb) == insn)
7438 BB_HEAD (bb) = b;
7439 if (BB_END (bb) == insn)
7440 BB_END (bb) = a;
7443 if (insn == BB_END (bb))
7444 break;
7445 insn = NEXT_INSN (insn);
7450 /* Perform machine dependent operations on the rtl chain INSNS. */
7452 static void
7453 ia64_reorg (void)
7455 /* We are freeing block_for_insn in the toplev to keep compatibility
7456 with old MDEP_REORGS that are not CFG based. Recompute it now. */
7457 compute_bb_for_insn ();
7459 /* If optimizing, we'll have split before scheduling. */
7460 if (optimize == 0)
7461 split_all_insns (0);
7463 /* ??? update_life_info_in_dirty_blocks fails to terminate during
7464 non-optimizing bootstrap. */
7465 update_life_info (NULL, UPDATE_LIFE_GLOBAL_RM_NOTES, PROP_DEATH_NOTES);
7467 if (ia64_flag_schedule_insns2)
7469 timevar_push (TV_SCHED2);
7470 ia64_final_schedule = 1;
7472 initiate_bundle_states ();
7473 ia64_nop = make_insn_raw (gen_nop ());
7474 PREV_INSN (ia64_nop) = NEXT_INSN (ia64_nop) = NULL_RTX;
7475 recog_memoized (ia64_nop);
7476 clocks_length = get_max_uid () + 1;
7477 stops_p = xcalloc (1, clocks_length);
7478 if (ia64_tune == PROCESSOR_ITANIUM)
7480 clocks = xcalloc (clocks_length, sizeof (int));
7481 add_cycles = xcalloc (clocks_length, sizeof (int));
7483 if (ia64_tune == PROCESSOR_ITANIUM2)
7485 pos_1 = get_cpu_unit_code ("2_1");
7486 pos_2 = get_cpu_unit_code ("2_2");
7487 pos_3 = get_cpu_unit_code ("2_3");
7488 pos_4 = get_cpu_unit_code ("2_4");
7489 pos_5 = get_cpu_unit_code ("2_5");
7490 pos_6 = get_cpu_unit_code ("2_6");
7491 _0mii_ = get_cpu_unit_code ("2b_0mii.");
7492 _0mmi_ = get_cpu_unit_code ("2b_0mmi.");
7493 _0mfi_ = get_cpu_unit_code ("2b_0mfi.");
7494 _0mmf_ = get_cpu_unit_code ("2b_0mmf.");
7495 _0bbb_ = get_cpu_unit_code ("2b_0bbb.");
7496 _0mbb_ = get_cpu_unit_code ("2b_0mbb.");
7497 _0mib_ = get_cpu_unit_code ("2b_0mib.");
7498 _0mmb_ = get_cpu_unit_code ("2b_0mmb.");
7499 _0mfb_ = get_cpu_unit_code ("2b_0mfb.");
7500 _0mlx_ = get_cpu_unit_code ("2b_0mlx.");
7501 _1mii_ = get_cpu_unit_code ("2b_1mii.");
7502 _1mmi_ = get_cpu_unit_code ("2b_1mmi.");
7503 _1mfi_ = get_cpu_unit_code ("2b_1mfi.");
7504 _1mmf_ = get_cpu_unit_code ("2b_1mmf.");
7505 _1bbb_ = get_cpu_unit_code ("2b_1bbb.");
7506 _1mbb_ = get_cpu_unit_code ("2b_1mbb.");
7507 _1mib_ = get_cpu_unit_code ("2b_1mib.");
7508 _1mmb_ = get_cpu_unit_code ("2b_1mmb.");
7509 _1mfb_ = get_cpu_unit_code ("2b_1mfb.");
7510 _1mlx_ = get_cpu_unit_code ("2b_1mlx.");
7512 else
7514 pos_1 = get_cpu_unit_code ("1_1");
7515 pos_2 = get_cpu_unit_code ("1_2");
7516 pos_3 = get_cpu_unit_code ("1_3");
7517 pos_4 = get_cpu_unit_code ("1_4");
7518 pos_5 = get_cpu_unit_code ("1_5");
7519 pos_6 = get_cpu_unit_code ("1_6");
7520 _0mii_ = get_cpu_unit_code ("1b_0mii.");
7521 _0mmi_ = get_cpu_unit_code ("1b_0mmi.");
7522 _0mfi_ = get_cpu_unit_code ("1b_0mfi.");
7523 _0mmf_ = get_cpu_unit_code ("1b_0mmf.");
7524 _0bbb_ = get_cpu_unit_code ("1b_0bbb.");
7525 _0mbb_ = get_cpu_unit_code ("1b_0mbb.");
7526 _0mib_ = get_cpu_unit_code ("1b_0mib.");
7527 _0mmb_ = get_cpu_unit_code ("1b_0mmb.");
7528 _0mfb_ = get_cpu_unit_code ("1b_0mfb.");
7529 _0mlx_ = get_cpu_unit_code ("1b_0mlx.");
7530 _1mii_ = get_cpu_unit_code ("1b_1mii.");
7531 _1mmi_ = get_cpu_unit_code ("1b_1mmi.");
7532 _1mfi_ = get_cpu_unit_code ("1b_1mfi.");
7533 _1mmf_ = get_cpu_unit_code ("1b_1mmf.");
7534 _1bbb_ = get_cpu_unit_code ("1b_1bbb.");
7535 _1mbb_ = get_cpu_unit_code ("1b_1mbb.");
7536 _1mib_ = get_cpu_unit_code ("1b_1mib.");
7537 _1mmb_ = get_cpu_unit_code ("1b_1mmb.");
7538 _1mfb_ = get_cpu_unit_code ("1b_1mfb.");
7539 _1mlx_ = get_cpu_unit_code ("1b_1mlx.");
7541 schedule_ebbs (dump_file);
7542 finish_bundle_states ();
7543 if (ia64_tune == PROCESSOR_ITANIUM)
7545 free (add_cycles);
7546 free (clocks);
7548 free (stops_p);
7549 emit_insn_group_barriers (dump_file);
7551 ia64_final_schedule = 0;
7552 timevar_pop (TV_SCHED2);
7554 else
7555 emit_all_insn_group_barriers (dump_file);
7557 /* A call must not be the last instruction in a function, so that the
7558 return address is still within the function, so that unwinding works
7559 properly. Note that IA-64 differs from dwarf2 on this point. */
7560 if (flag_unwind_tables || (flag_exceptions && !USING_SJLJ_EXCEPTIONS))
7562 rtx insn;
7563 int saw_stop = 0;
7565 insn = get_last_insn ();
7566 if (! INSN_P (insn))
7567 insn = prev_active_insn (insn);
7568 /* Skip over insns that expand to nothing. */
7569 while (GET_CODE (insn) == INSN && get_attr_empty (insn) == EMPTY_YES)
7571 if (GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
7572 && XINT (PATTERN (insn), 1) == UNSPECV_INSN_GROUP_BARRIER)
7573 saw_stop = 1;
7574 insn = prev_active_insn (insn);
7576 if (GET_CODE (insn) == CALL_INSN)
7578 if (! saw_stop)
7579 emit_insn (gen_insn_group_barrier (GEN_INT (3)));
7580 emit_insn (gen_break_f ());
7581 emit_insn (gen_insn_group_barrier (GEN_INT (3)));
7585 fixup_errata ();
7586 emit_predicate_relation_info ();
7588 if (ia64_flag_var_tracking)
7590 timevar_push (TV_VAR_TRACKING);
7591 variable_tracking_main ();
7592 timevar_pop (TV_VAR_TRACKING);
7596 /* Return true if REGNO is used by the epilogue. */
7599 ia64_epilogue_uses (int regno)
7601 switch (regno)
7603 case R_GR (1):
7604 /* With a call to a function in another module, we will write a new
7605 value to "gp". After returning from such a call, we need to make
7606 sure the function restores the original gp-value, even if the
7607 function itself does not use the gp anymore. */
7608 return !(TARGET_AUTO_PIC || TARGET_NO_PIC);
7610 case IN_REG (0): case IN_REG (1): case IN_REG (2): case IN_REG (3):
7611 case IN_REG (4): case IN_REG (5): case IN_REG (6): case IN_REG (7):
7612 /* For functions defined with the syscall_linkage attribute, all
7613 input registers are marked as live at all function exits. This
7614 prevents the register allocator from using the input registers,
7615 which in turn makes it possible to restart a system call after
7616 an interrupt without having to save/restore the input registers.
7617 This also prevents kernel data from leaking to application code. */
7618 return lookup_attribute ("syscall_linkage",
7619 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))) != NULL;
7621 case R_BR (0):
7622 /* Conditional return patterns can't represent the use of `b0' as
7623 the return address, so we force the value live this way. */
7624 return 1;
7626 case AR_PFS_REGNUM:
7627 /* Likewise for ar.pfs, which is used by br.ret. */
7628 return 1;
7630 default:
7631 return 0;
7635 /* Return true if REGNO is used by the frame unwinder. */
7638 ia64_eh_uses (int regno)
7640 if (! reload_completed)
7641 return 0;
7643 if (current_frame_info.reg_save_b0
7644 && regno == current_frame_info.reg_save_b0)
7645 return 1;
7646 if (current_frame_info.reg_save_pr
7647 && regno == current_frame_info.reg_save_pr)
7648 return 1;
7649 if (current_frame_info.reg_save_ar_pfs
7650 && regno == current_frame_info.reg_save_ar_pfs)
7651 return 1;
7652 if (current_frame_info.reg_save_ar_unat
7653 && regno == current_frame_info.reg_save_ar_unat)
7654 return 1;
7655 if (current_frame_info.reg_save_ar_lc
7656 && regno == current_frame_info.reg_save_ar_lc)
7657 return 1;
7659 return 0;
7662 /* Return true if this goes in small data/bss. */
7664 /* ??? We could also support own long data here. Generating movl/add/ld8
7665 instead of addl,ld8/ld8. This makes the code bigger, but should make the
7666 code faster because there is one less load. This also includes incomplete
7667 types which can't go in sdata/sbss. */
7669 static bool
7670 ia64_in_small_data_p (tree exp)
7672 if (TARGET_NO_SDATA)
7673 return false;
7675 /* We want to merge strings, so we never consider them small data. */
7676 if (TREE_CODE (exp) == STRING_CST)
7677 return false;
7679 /* Functions are never small data. */
7680 if (TREE_CODE (exp) == FUNCTION_DECL)
7681 return false;
7683 if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
7685 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));
7687 if (strcmp (section, ".sdata") == 0
7688 || strncmp (section, ".sdata.", 7) == 0
7689 || strncmp (section, ".gnu.linkonce.s.", 16) == 0
7690 || strcmp (section, ".sbss") == 0
7691 || strncmp (section, ".sbss.", 6) == 0
7692 || strncmp (section, ".gnu.linkonce.sb.", 17) == 0)
7693 return true;
7695 else
7697 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
7699 /* If this is an incomplete type with size 0, then we can't put it
7700 in sdata because it might be too big when completed. */
7701 if (size > 0 && size <= ia64_section_threshold)
7702 return true;
7705 return false;
7708 /* Output assembly directives for prologue regions. */
7710 /* The current basic block number. */
7712 static bool last_block;
7714 /* True if we need a copy_state command at the start of the next block. */
7716 static bool need_copy_state;
7718 /* The function emits unwind directives for the start of an epilogue. */
7720 static void
7721 process_epilogue (void)
7723 /* If this isn't the last block of the function, then we need to label the
7724 current state, and copy it back in at the start of the next block. */
7726 if (!last_block)
7728 fprintf (asm_out_file, "\t.label_state 1\n");
7729 need_copy_state = true;
7732 fprintf (asm_out_file, "\t.restore sp\n");
7735 /* This function processes a SET pattern looking for specific patterns
7736 which result in emitting an assembly directive required for unwinding. */
7738 static int
7739 process_set (FILE *asm_out_file, rtx pat)
7741 rtx src = SET_SRC (pat);
7742 rtx dest = SET_DEST (pat);
7743 int src_regno, dest_regno;
7745 /* Look for the ALLOC insn. */
7746 if (GET_CODE (src) == UNSPEC_VOLATILE
7747 && XINT (src, 1) == UNSPECV_ALLOC
7748 && GET_CODE (dest) == REG)
7750 dest_regno = REGNO (dest);
7752 /* If this is the final destination for ar.pfs, then this must
7753 be the alloc in the prologue. */
7754 if (dest_regno == current_frame_info.reg_save_ar_pfs)
7755 fprintf (asm_out_file, "\t.save ar.pfs, r%d\n",
7756 ia64_dbx_register_number (dest_regno));
7757 else
7759 /* This must be an alloc before a sibcall. We must drop the
7760 old frame info. The easiest way to drop the old frame
7761 info is to ensure we had a ".restore sp" directive
7762 followed by a new prologue. If the procedure doesn't
7763 have a memory-stack frame, we'll issue a dummy ".restore
7764 sp" now. */
7765 if (current_frame_info.total_size == 0 && !frame_pointer_needed)
7766 /* if haven't done process_epilogue() yet, do it now */
7767 process_epilogue ();
7768 fprintf (asm_out_file, "\t.prologue\n");
7770 return 1;
7773 /* Look for SP = .... */
7774 if (GET_CODE (dest) == REG && REGNO (dest) == STACK_POINTER_REGNUM)
7776 if (GET_CODE (src) == PLUS)
7778 rtx op0 = XEXP (src, 0);
7779 rtx op1 = XEXP (src, 1);
7780 if (op0 == dest && GET_CODE (op1) == CONST_INT)
7782 if (INTVAL (op1) < 0)
7783 fprintf (asm_out_file, "\t.fframe "HOST_WIDE_INT_PRINT_DEC"\n",
7784 -INTVAL (op1));
7785 else
7786 process_epilogue ();
7788 else
7789 abort ();
7791 else if (GET_CODE (src) == REG
7792 && REGNO (src) == HARD_FRAME_POINTER_REGNUM)
7793 process_epilogue ();
7794 else
7795 abort ();
7797 return 1;
7800 /* Register move we need to look at. */
7801 if (GET_CODE (dest) == REG && GET_CODE (src) == REG)
7803 src_regno = REGNO (src);
7804 dest_regno = REGNO (dest);
7806 switch (src_regno)
7808 case BR_REG (0):
7809 /* Saving return address pointer. */
7810 if (dest_regno != current_frame_info.reg_save_b0)
7811 abort ();
7812 fprintf (asm_out_file, "\t.save rp, r%d\n",
7813 ia64_dbx_register_number (dest_regno));
7814 return 1;
7816 case PR_REG (0):
7817 if (dest_regno != current_frame_info.reg_save_pr)
7818 abort ();
7819 fprintf (asm_out_file, "\t.save pr, r%d\n",
7820 ia64_dbx_register_number (dest_regno));
7821 return 1;
7823 case AR_UNAT_REGNUM:
7824 if (dest_regno != current_frame_info.reg_save_ar_unat)
7825 abort ();
7826 fprintf (asm_out_file, "\t.save ar.unat, r%d\n",
7827 ia64_dbx_register_number (dest_regno));
7828 return 1;
7830 case AR_LC_REGNUM:
7831 if (dest_regno != current_frame_info.reg_save_ar_lc)
7832 abort ();
7833 fprintf (asm_out_file, "\t.save ar.lc, r%d\n",
7834 ia64_dbx_register_number (dest_regno));
7835 return 1;
7837 case STACK_POINTER_REGNUM:
7838 if (dest_regno != HARD_FRAME_POINTER_REGNUM
7839 || ! frame_pointer_needed)
7840 abort ();
7841 fprintf (asm_out_file, "\t.vframe r%d\n",
7842 ia64_dbx_register_number (dest_regno));
7843 return 1;
7845 default:
7846 /* Everything else should indicate being stored to memory. */
7847 abort ();
7851 /* Memory store we need to look at. */
7852 if (GET_CODE (dest) == MEM && GET_CODE (src) == REG)
7854 long off;
7855 rtx base;
7856 const char *saveop;
7858 if (GET_CODE (XEXP (dest, 0)) == REG)
7860 base = XEXP (dest, 0);
7861 off = 0;
7863 else if (GET_CODE (XEXP (dest, 0)) == PLUS
7864 && GET_CODE (XEXP (XEXP (dest, 0), 1)) == CONST_INT)
7866 base = XEXP (XEXP (dest, 0), 0);
7867 off = INTVAL (XEXP (XEXP (dest, 0), 1));
7869 else
7870 abort ();
7872 if (base == hard_frame_pointer_rtx)
7874 saveop = ".savepsp";
7875 off = - off;
7877 else if (base == stack_pointer_rtx)
7878 saveop = ".savesp";
7879 else
7880 abort ();
7882 src_regno = REGNO (src);
7883 switch (src_regno)
7885 case BR_REG (0):
7886 if (current_frame_info.reg_save_b0 != 0)
7887 abort ();
7888 fprintf (asm_out_file, "\t%s rp, %ld\n", saveop, off);
7889 return 1;
7891 case PR_REG (0):
7892 if (current_frame_info.reg_save_pr != 0)
7893 abort ();
7894 fprintf (asm_out_file, "\t%s pr, %ld\n", saveop, off);
7895 return 1;
7897 case AR_LC_REGNUM:
7898 if (current_frame_info.reg_save_ar_lc != 0)
7899 abort ();
7900 fprintf (asm_out_file, "\t%s ar.lc, %ld\n", saveop, off);
7901 return 1;
7903 case AR_PFS_REGNUM:
7904 if (current_frame_info.reg_save_ar_pfs != 0)
7905 abort ();
7906 fprintf (asm_out_file, "\t%s ar.pfs, %ld\n", saveop, off);
7907 return 1;
7909 case AR_UNAT_REGNUM:
7910 if (current_frame_info.reg_save_ar_unat != 0)
7911 abort ();
7912 fprintf (asm_out_file, "\t%s ar.unat, %ld\n", saveop, off);
7913 return 1;
7915 case GR_REG (4):
7916 case GR_REG (5):
7917 case GR_REG (6):
7918 case GR_REG (7):
7919 fprintf (asm_out_file, "\t.save.g 0x%x\n",
7920 1 << (src_regno - GR_REG (4)));
7921 return 1;
7923 case BR_REG (1):
7924 case BR_REG (2):
7925 case BR_REG (3):
7926 case BR_REG (4):
7927 case BR_REG (5):
7928 fprintf (asm_out_file, "\t.save.b 0x%x\n",
7929 1 << (src_regno - BR_REG (1)));
7930 return 1;
7932 case FR_REG (2):
7933 case FR_REG (3):
7934 case FR_REG (4):
7935 case FR_REG (5):
7936 fprintf (asm_out_file, "\t.save.f 0x%x\n",
7937 1 << (src_regno - FR_REG (2)));
7938 return 1;
7940 case FR_REG (16): case FR_REG (17): case FR_REG (18): case FR_REG (19):
7941 case FR_REG (20): case FR_REG (21): case FR_REG (22): case FR_REG (23):
7942 case FR_REG (24): case FR_REG (25): case FR_REG (26): case FR_REG (27):
7943 case FR_REG (28): case FR_REG (29): case FR_REG (30): case FR_REG (31):
7944 fprintf (asm_out_file, "\t.save.gf 0x0, 0x%x\n",
7945 1 << (src_regno - FR_REG (12)));
7946 return 1;
7948 default:
7949 return 0;
7953 return 0;
7957 /* This function looks at a single insn and emits any directives
7958 required to unwind this insn. */
7959 void
7960 process_for_unwind_directive (FILE *asm_out_file, rtx insn)
7962 if (flag_unwind_tables
7963 || (flag_exceptions && !USING_SJLJ_EXCEPTIONS))
7965 rtx pat;
7967 if (GET_CODE (insn) == NOTE
7968 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_BASIC_BLOCK)
7970 last_block = NOTE_BASIC_BLOCK (insn)->next_bb == EXIT_BLOCK_PTR;
7972 /* Restore unwind state from immediately before the epilogue. */
7973 if (need_copy_state)
7975 fprintf (asm_out_file, "\t.body\n");
7976 fprintf (asm_out_file, "\t.copy_state 1\n");
7977 need_copy_state = false;
7981 if (GET_CODE (insn) == NOTE || ! RTX_FRAME_RELATED_P (insn))
7982 return;
7984 pat = find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX);
7985 if (pat)
7986 pat = XEXP (pat, 0);
7987 else
7988 pat = PATTERN (insn);
7990 switch (GET_CODE (pat))
7992 case SET:
7993 process_set (asm_out_file, pat);
7994 break;
7996 case PARALLEL:
7998 int par_index;
7999 int limit = XVECLEN (pat, 0);
8000 for (par_index = 0; par_index < limit; par_index++)
8002 rtx x = XVECEXP (pat, 0, par_index);
8003 if (GET_CODE (x) == SET)
8004 process_set (asm_out_file, x);
8006 break;
8009 default:
8010 abort ();
8016 void
8017 ia64_init_builtins (void)
8019 tree psi_type_node = build_pointer_type (integer_type_node);
8020 tree pdi_type_node = build_pointer_type (long_integer_type_node);
8022 /* __sync_val_compare_and_swap_si, __sync_bool_compare_and_swap_si */
8023 tree si_ftype_psi_si_si
8024 = build_function_type_list (integer_type_node,
8025 psi_type_node, integer_type_node,
8026 integer_type_node, NULL_TREE);
8028 /* __sync_val_compare_and_swap_di */
8029 tree di_ftype_pdi_di_di
8030 = build_function_type_list (long_integer_type_node,
8031 pdi_type_node, long_integer_type_node,
8032 long_integer_type_node, NULL_TREE);
8033 /* __sync_bool_compare_and_swap_di */
8034 tree si_ftype_pdi_di_di
8035 = build_function_type_list (integer_type_node,
8036 pdi_type_node, long_integer_type_node,
8037 long_integer_type_node, NULL_TREE);
8038 /* __sync_synchronize */
8039 tree void_ftype_void
8040 = build_function_type (void_type_node, void_list_node);
8042 /* __sync_lock_test_and_set_si */
8043 tree si_ftype_psi_si
8044 = build_function_type_list (integer_type_node,
8045 psi_type_node, integer_type_node, NULL_TREE);
8047 /* __sync_lock_test_and_set_di */
8048 tree di_ftype_pdi_di
8049 = build_function_type_list (long_integer_type_node,
8050 pdi_type_node, long_integer_type_node,
8051 NULL_TREE);
8053 /* __sync_lock_release_si */
8054 tree void_ftype_psi
8055 = build_function_type_list (void_type_node, psi_type_node, NULL_TREE);
8057 /* __sync_lock_release_di */
8058 tree void_ftype_pdi
8059 = build_function_type_list (void_type_node, pdi_type_node, NULL_TREE);
8061 tree fpreg_type;
8062 tree float80_type;
8064 /* The __fpreg type. */
8065 fpreg_type = make_node (REAL_TYPE);
8066 /* ??? The back end should know to load/save __fpreg variables using
8067 the ldf.fill and stf.spill instructions. */
8068 TYPE_PRECISION (fpreg_type) = 80;
8069 layout_type (fpreg_type);
8070 (*lang_hooks.types.register_builtin_type) (fpreg_type, "__fpreg");
8072 /* The __float80 type. */
8073 float80_type = make_node (REAL_TYPE);
8074 TYPE_PRECISION (float80_type) = 80;
8075 layout_type (float80_type);
8076 (*lang_hooks.types.register_builtin_type) (float80_type, "__float80");
8078 /* The __float128 type. */
8079 if (!TARGET_HPUX)
8081 tree float128_type = make_node (REAL_TYPE);
8082 TYPE_PRECISION (float128_type) = 128;
8083 layout_type (float128_type);
8084 (*lang_hooks.types.register_builtin_type) (float128_type, "__float128");
8086 else
8087 /* Under HPUX, this is a synonym for "long double". */
8088 (*lang_hooks.types.register_builtin_type) (long_double_type_node,
8089 "__float128");
8091 #define def_builtin(name, type, code) \
8092 lang_hooks.builtin_function ((name), (type), (code), BUILT_IN_MD, \
8093 NULL, NULL_TREE)
8095 def_builtin ("__sync_val_compare_and_swap_si", si_ftype_psi_si_si,
8096 IA64_BUILTIN_VAL_COMPARE_AND_SWAP_SI);
8097 def_builtin ("__sync_val_compare_and_swap_di", di_ftype_pdi_di_di,
8098 IA64_BUILTIN_VAL_COMPARE_AND_SWAP_DI);
8099 def_builtin ("__sync_bool_compare_and_swap_si", si_ftype_psi_si_si,
8100 IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_SI);
8101 def_builtin ("__sync_bool_compare_and_swap_di", si_ftype_pdi_di_di,
8102 IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_DI);
8104 def_builtin ("__sync_synchronize", void_ftype_void,
8105 IA64_BUILTIN_SYNCHRONIZE);
8107 def_builtin ("__sync_lock_test_and_set_si", si_ftype_psi_si,
8108 IA64_BUILTIN_LOCK_TEST_AND_SET_SI);
8109 def_builtin ("__sync_lock_test_and_set_di", di_ftype_pdi_di,
8110 IA64_BUILTIN_LOCK_TEST_AND_SET_DI);
8111 def_builtin ("__sync_lock_release_si", void_ftype_psi,
8112 IA64_BUILTIN_LOCK_RELEASE_SI);
8113 def_builtin ("__sync_lock_release_di", void_ftype_pdi,
8114 IA64_BUILTIN_LOCK_RELEASE_DI);
8116 def_builtin ("__builtin_ia64_bsp",
8117 build_function_type (ptr_type_node, void_list_node),
8118 IA64_BUILTIN_BSP);
8120 def_builtin ("__builtin_ia64_flushrs",
8121 build_function_type (void_type_node, void_list_node),
8122 IA64_BUILTIN_FLUSHRS);
8124 def_builtin ("__sync_fetch_and_add_si", si_ftype_psi_si,
8125 IA64_BUILTIN_FETCH_AND_ADD_SI);
8126 def_builtin ("__sync_fetch_and_sub_si", si_ftype_psi_si,
8127 IA64_BUILTIN_FETCH_AND_SUB_SI);
8128 def_builtin ("__sync_fetch_and_or_si", si_ftype_psi_si,
8129 IA64_BUILTIN_FETCH_AND_OR_SI);
8130 def_builtin ("__sync_fetch_and_and_si", si_ftype_psi_si,
8131 IA64_BUILTIN_FETCH_AND_AND_SI);
8132 def_builtin ("__sync_fetch_and_xor_si", si_ftype_psi_si,
8133 IA64_BUILTIN_FETCH_AND_XOR_SI);
8134 def_builtin ("__sync_fetch_and_nand_si", si_ftype_psi_si,
8135 IA64_BUILTIN_FETCH_AND_NAND_SI);
8137 def_builtin ("__sync_add_and_fetch_si", si_ftype_psi_si,
8138 IA64_BUILTIN_ADD_AND_FETCH_SI);
8139 def_builtin ("__sync_sub_and_fetch_si", si_ftype_psi_si,
8140 IA64_BUILTIN_SUB_AND_FETCH_SI);
8141 def_builtin ("__sync_or_and_fetch_si", si_ftype_psi_si,
8142 IA64_BUILTIN_OR_AND_FETCH_SI);
8143 def_builtin ("__sync_and_and_fetch_si", si_ftype_psi_si,
8144 IA64_BUILTIN_AND_AND_FETCH_SI);
8145 def_builtin ("__sync_xor_and_fetch_si", si_ftype_psi_si,
8146 IA64_BUILTIN_XOR_AND_FETCH_SI);
8147 def_builtin ("__sync_nand_and_fetch_si", si_ftype_psi_si,
8148 IA64_BUILTIN_NAND_AND_FETCH_SI);
8150 def_builtin ("__sync_fetch_and_add_di", di_ftype_pdi_di,
8151 IA64_BUILTIN_FETCH_AND_ADD_DI);
8152 def_builtin ("__sync_fetch_and_sub_di", di_ftype_pdi_di,
8153 IA64_BUILTIN_FETCH_AND_SUB_DI);
8154 def_builtin ("__sync_fetch_and_or_di", di_ftype_pdi_di,
8155 IA64_BUILTIN_FETCH_AND_OR_DI);
8156 def_builtin ("__sync_fetch_and_and_di", di_ftype_pdi_di,
8157 IA64_BUILTIN_FETCH_AND_AND_DI);
8158 def_builtin ("__sync_fetch_and_xor_di", di_ftype_pdi_di,
8159 IA64_BUILTIN_FETCH_AND_XOR_DI);
8160 def_builtin ("__sync_fetch_and_nand_di", di_ftype_pdi_di,
8161 IA64_BUILTIN_FETCH_AND_NAND_DI);
8163 def_builtin ("__sync_add_and_fetch_di", di_ftype_pdi_di,
8164 IA64_BUILTIN_ADD_AND_FETCH_DI);
8165 def_builtin ("__sync_sub_and_fetch_di", di_ftype_pdi_di,
8166 IA64_BUILTIN_SUB_AND_FETCH_DI);
8167 def_builtin ("__sync_or_and_fetch_di", di_ftype_pdi_di,
8168 IA64_BUILTIN_OR_AND_FETCH_DI);
8169 def_builtin ("__sync_and_and_fetch_di", di_ftype_pdi_di,
8170 IA64_BUILTIN_AND_AND_FETCH_DI);
8171 def_builtin ("__sync_xor_and_fetch_di", di_ftype_pdi_di,
8172 IA64_BUILTIN_XOR_AND_FETCH_DI);
8173 def_builtin ("__sync_nand_and_fetch_di", di_ftype_pdi_di,
8174 IA64_BUILTIN_NAND_AND_FETCH_DI);
8176 #undef def_builtin
8179 /* Expand fetch_and_op intrinsics. The basic code sequence is:
8182 tmp = [ptr];
8183 do {
8184 ret = tmp;
8185 ar.ccv = tmp;
8186 tmp <op>= value;
8187 cmpxchgsz.acq tmp = [ptr], tmp
8188 } while (tmp != ret)
8191 static rtx
8192 ia64_expand_fetch_and_op (optab binoptab, enum machine_mode mode,
8193 tree arglist, rtx target)
8195 rtx ret, label, tmp, ccv, insn, mem, value;
8196 tree arg0, arg1;
8198 arg0 = TREE_VALUE (arglist);
8199 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8200 mem = expand_expr (arg0, NULL_RTX, Pmode, 0);
8201 #ifdef POINTERS_EXTEND_UNSIGNED
8202 if (GET_MODE(mem) != Pmode)
8203 mem = convert_memory_address (Pmode, mem);
8204 #endif
8205 value = expand_expr (arg1, NULL_RTX, mode, 0);
8207 mem = gen_rtx_MEM (mode, force_reg (Pmode, mem));
8208 MEM_VOLATILE_P (mem) = 1;
8210 if (target && register_operand (target, mode))
8211 ret = target;
8212 else
8213 ret = gen_reg_rtx (mode);
8215 emit_insn (gen_mf ());
8217 /* Special case for fetchadd instructions. */
8218 if (binoptab == add_optab && fetchadd_operand (value, VOIDmode))
8220 if (mode == SImode)
8221 insn = gen_fetchadd_acq_si (ret, mem, value);
8222 else
8223 insn = gen_fetchadd_acq_di (ret, mem, value);
8224 emit_insn (insn);
8225 return ret;
8228 tmp = gen_reg_rtx (mode);
8229 /* ar.ccv must always be loaded with a zero-extended DImode value. */
8230 ccv = gen_rtx_REG (DImode, AR_CCV_REGNUM);
8231 emit_move_insn (tmp, mem);
8233 label = gen_label_rtx ();
8234 emit_label (label);
8235 emit_move_insn (ret, tmp);
8236 convert_move (ccv, tmp, /*unsignedp=*/1);
8238 /* Perform the specific operation. Special case NAND by noticing
8239 one_cmpl_optab instead. */
8240 if (binoptab == one_cmpl_optab)
8242 tmp = expand_unop (mode, binoptab, tmp, NULL, OPTAB_WIDEN);
8243 binoptab = and_optab;
8245 tmp = expand_binop (mode, binoptab, tmp, value, tmp, 1, OPTAB_WIDEN);
8247 if (mode == SImode)
8248 insn = gen_cmpxchg_acq_si (tmp, mem, tmp, ccv);
8249 else
8250 insn = gen_cmpxchg_acq_di (tmp, mem, tmp, ccv);
8251 emit_insn (insn);
8253 emit_cmp_and_jump_insns (tmp, ret, NE, 0, mode, 1, label);
8255 return ret;
8258 /* Expand op_and_fetch intrinsics. The basic code sequence is:
8261 tmp = [ptr];
8262 do {
8263 old = tmp;
8264 ar.ccv = tmp;
8265 ret = tmp <op> value;
8266 cmpxchgsz.acq tmp = [ptr], ret
8267 } while (tmp != old)
8270 static rtx
8271 ia64_expand_op_and_fetch (optab binoptab, enum machine_mode mode,
8272 tree arglist, rtx target)
8274 rtx old, label, tmp, ret, ccv, insn, mem, value;
8275 tree arg0, arg1;
8277 arg0 = TREE_VALUE (arglist);
8278 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8279 mem = expand_expr (arg0, NULL_RTX, Pmode, 0);
8280 #ifdef POINTERS_EXTEND_UNSIGNED
8281 if (GET_MODE(mem) != Pmode)
8282 mem = convert_memory_address (Pmode, mem);
8283 #endif
8285 value = expand_expr (arg1, NULL_RTX, mode, 0);
8287 mem = gen_rtx_MEM (mode, force_reg (Pmode, mem));
8288 MEM_VOLATILE_P (mem) = 1;
8290 if (target && ! register_operand (target, mode))
8291 target = NULL_RTX;
8293 emit_insn (gen_mf ());
8294 tmp = gen_reg_rtx (mode);
8295 old = gen_reg_rtx (mode);
8296 /* ar.ccv must always be loaded with a zero-extended DImode value. */
8297 ccv = gen_rtx_REG (DImode, AR_CCV_REGNUM);
8299 emit_move_insn (tmp, mem);
8301 label = gen_label_rtx ();
8302 emit_label (label);
8303 emit_move_insn (old, tmp);
8304 convert_move (ccv, tmp, /*unsignedp=*/1);
8306 /* Perform the specific operation. Special case NAND by noticing
8307 one_cmpl_optab instead. */
8308 if (binoptab == one_cmpl_optab)
8310 tmp = expand_unop (mode, binoptab, tmp, NULL, OPTAB_WIDEN);
8311 binoptab = and_optab;
8313 ret = expand_binop (mode, binoptab, tmp, value, target, 1, OPTAB_WIDEN);
8315 if (mode == SImode)
8316 insn = gen_cmpxchg_acq_si (tmp, mem, ret, ccv);
8317 else
8318 insn = gen_cmpxchg_acq_di (tmp, mem, ret, ccv);
8319 emit_insn (insn);
8321 emit_cmp_and_jump_insns (tmp, old, NE, 0, mode, 1, label);
8323 return ret;
8326 /* Expand val_ and bool_compare_and_swap. For val_ we want:
8328 ar.ccv = oldval
8330 cmpxchgsz.acq ret = [ptr], newval, ar.ccv
8331 return ret
8333 For bool_ it's the same except return ret == oldval.
8336 static rtx
8337 ia64_expand_compare_and_swap (enum machine_mode rmode, enum machine_mode mode,
8338 int boolp, tree arglist, rtx target)
8340 tree arg0, arg1, arg2;
8341 rtx mem, old, new, ccv, tmp, insn;
8343 arg0 = TREE_VALUE (arglist);
8344 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8345 arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
8346 mem = expand_expr (arg0, NULL_RTX, ptr_mode, 0);
8347 old = expand_expr (arg1, NULL_RTX, mode, 0);
8348 new = expand_expr (arg2, NULL_RTX, mode, 0);
8350 mem = gen_rtx_MEM (mode, force_reg (ptr_mode, mem));
8351 MEM_VOLATILE_P (mem) = 1;
8353 if (GET_MODE (old) != mode)
8354 old = convert_to_mode (mode, old, /*unsignedp=*/1);
8355 if (GET_MODE (new) != mode)
8356 new = convert_to_mode (mode, new, /*unsignedp=*/1);
8358 if (! register_operand (old, mode))
8359 old = copy_to_mode_reg (mode, old);
8360 if (! register_operand (new, mode))
8361 new = copy_to_mode_reg (mode, new);
8363 if (! boolp && target && register_operand (target, mode))
8364 tmp = target;
8365 else
8366 tmp = gen_reg_rtx (mode);
8368 ccv = gen_rtx_REG (DImode, AR_CCV_REGNUM);
8369 convert_move (ccv, old, /*unsignedp=*/1);
8370 emit_insn (gen_mf ());
8371 if (mode == SImode)
8372 insn = gen_cmpxchg_acq_si (tmp, mem, new, ccv);
8373 else
8374 insn = gen_cmpxchg_acq_di (tmp, mem, new, ccv);
8375 emit_insn (insn);
8377 if (boolp)
8379 if (! target)
8380 target = gen_reg_rtx (rmode);
8381 return emit_store_flag_force (target, EQ, tmp, old, mode, 1, 1);
8383 else
8384 return tmp;
8387 /* Expand lock_test_and_set. I.e. `xchgsz ret = [ptr], new'. */
8389 static rtx
8390 ia64_expand_lock_test_and_set (enum machine_mode mode, tree arglist,
8391 rtx target)
8393 tree arg0, arg1;
8394 rtx mem, new, ret, insn;
8396 arg0 = TREE_VALUE (arglist);
8397 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8398 mem = expand_expr (arg0, NULL_RTX, ptr_mode, 0);
8399 new = expand_expr (arg1, NULL_RTX, mode, 0);
8401 mem = gen_rtx_MEM (mode, force_reg (ptr_mode, mem));
8402 MEM_VOLATILE_P (mem) = 1;
8403 if (! register_operand (new, mode))
8404 new = copy_to_mode_reg (mode, new);
8406 if (target && register_operand (target, mode))
8407 ret = target;
8408 else
8409 ret = gen_reg_rtx (mode);
8411 if (mode == SImode)
8412 insn = gen_xchgsi (ret, mem, new);
8413 else
8414 insn = gen_xchgdi (ret, mem, new);
8415 emit_insn (insn);
8417 return ret;
8420 /* Expand lock_release. I.e. `stsz.rel [ptr] = r0'. */
8422 static rtx
8423 ia64_expand_lock_release (enum machine_mode mode, tree arglist,
8424 rtx target ATTRIBUTE_UNUSED)
8426 tree arg0;
8427 rtx mem;
8429 arg0 = TREE_VALUE (arglist);
8430 mem = expand_expr (arg0, NULL_RTX, ptr_mode, 0);
8432 mem = gen_rtx_MEM (mode, force_reg (ptr_mode, mem));
8433 MEM_VOLATILE_P (mem) = 1;
8435 emit_move_insn (mem, const0_rtx);
8437 return const0_rtx;
8441 ia64_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
8442 enum machine_mode mode ATTRIBUTE_UNUSED,
8443 int ignore ATTRIBUTE_UNUSED)
8445 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
8446 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
8447 tree arglist = TREE_OPERAND (exp, 1);
8448 enum machine_mode rmode = VOIDmode;
8450 switch (fcode)
8452 case IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_SI:
8453 case IA64_BUILTIN_VAL_COMPARE_AND_SWAP_SI:
8454 mode = SImode;
8455 rmode = SImode;
8456 break;
8458 case IA64_BUILTIN_LOCK_TEST_AND_SET_SI:
8459 case IA64_BUILTIN_LOCK_RELEASE_SI:
8460 case IA64_BUILTIN_FETCH_AND_ADD_SI:
8461 case IA64_BUILTIN_FETCH_AND_SUB_SI:
8462 case IA64_BUILTIN_FETCH_AND_OR_SI:
8463 case IA64_BUILTIN_FETCH_AND_AND_SI:
8464 case IA64_BUILTIN_FETCH_AND_XOR_SI:
8465 case IA64_BUILTIN_FETCH_AND_NAND_SI:
8466 case IA64_BUILTIN_ADD_AND_FETCH_SI:
8467 case IA64_BUILTIN_SUB_AND_FETCH_SI:
8468 case IA64_BUILTIN_OR_AND_FETCH_SI:
8469 case IA64_BUILTIN_AND_AND_FETCH_SI:
8470 case IA64_BUILTIN_XOR_AND_FETCH_SI:
8471 case IA64_BUILTIN_NAND_AND_FETCH_SI:
8472 mode = SImode;
8473 break;
8475 case IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_DI:
8476 mode = DImode;
8477 rmode = SImode;
8478 break;
8480 case IA64_BUILTIN_VAL_COMPARE_AND_SWAP_DI:
8481 mode = DImode;
8482 rmode = DImode;
8483 break;
8485 case IA64_BUILTIN_LOCK_TEST_AND_SET_DI:
8486 case IA64_BUILTIN_LOCK_RELEASE_DI:
8487 case IA64_BUILTIN_FETCH_AND_ADD_DI:
8488 case IA64_BUILTIN_FETCH_AND_SUB_DI:
8489 case IA64_BUILTIN_FETCH_AND_OR_DI:
8490 case IA64_BUILTIN_FETCH_AND_AND_DI:
8491 case IA64_BUILTIN_FETCH_AND_XOR_DI:
8492 case IA64_BUILTIN_FETCH_AND_NAND_DI:
8493 case IA64_BUILTIN_ADD_AND_FETCH_DI:
8494 case IA64_BUILTIN_SUB_AND_FETCH_DI:
8495 case IA64_BUILTIN_OR_AND_FETCH_DI:
8496 case IA64_BUILTIN_AND_AND_FETCH_DI:
8497 case IA64_BUILTIN_XOR_AND_FETCH_DI:
8498 case IA64_BUILTIN_NAND_AND_FETCH_DI:
8499 mode = DImode;
8500 break;
8502 default:
8503 break;
8506 switch (fcode)
8508 case IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_SI:
8509 case IA64_BUILTIN_BOOL_COMPARE_AND_SWAP_DI:
8510 return ia64_expand_compare_and_swap (rmode, mode, 1, arglist,
8511 target);
8513 case IA64_BUILTIN_VAL_COMPARE_AND_SWAP_SI:
8514 case IA64_BUILTIN_VAL_COMPARE_AND_SWAP_DI:
8515 return ia64_expand_compare_and_swap (rmode, mode, 0, arglist,
8516 target);
8518 case IA64_BUILTIN_SYNCHRONIZE:
8519 emit_insn (gen_mf ());
8520 return const0_rtx;
8522 case IA64_BUILTIN_LOCK_TEST_AND_SET_SI:
8523 case IA64_BUILTIN_LOCK_TEST_AND_SET_DI:
8524 return ia64_expand_lock_test_and_set (mode, arglist, target);
8526 case IA64_BUILTIN_LOCK_RELEASE_SI:
8527 case IA64_BUILTIN_LOCK_RELEASE_DI:
8528 return ia64_expand_lock_release (mode, arglist, target);
8530 case IA64_BUILTIN_BSP:
8531 if (! target || ! register_operand (target, DImode))
8532 target = gen_reg_rtx (DImode);
8533 emit_insn (gen_bsp_value (target));
8534 #ifdef POINTERS_EXTEND_UNSIGNED
8535 target = convert_memory_address (ptr_mode, target);
8536 #endif
8537 return target;
8539 case IA64_BUILTIN_FLUSHRS:
8540 emit_insn (gen_flushrs ());
8541 return const0_rtx;
8543 case IA64_BUILTIN_FETCH_AND_ADD_SI:
8544 case IA64_BUILTIN_FETCH_AND_ADD_DI:
8545 return ia64_expand_fetch_and_op (add_optab, mode, arglist, target);
8547 case IA64_BUILTIN_FETCH_AND_SUB_SI:
8548 case IA64_BUILTIN_FETCH_AND_SUB_DI:
8549 return ia64_expand_fetch_and_op (sub_optab, mode, arglist, target);
8551 case IA64_BUILTIN_FETCH_AND_OR_SI:
8552 case IA64_BUILTIN_FETCH_AND_OR_DI:
8553 return ia64_expand_fetch_and_op (ior_optab, mode, arglist, target);
8555 case IA64_BUILTIN_FETCH_AND_AND_SI:
8556 case IA64_BUILTIN_FETCH_AND_AND_DI:
8557 return ia64_expand_fetch_and_op (and_optab, mode, arglist, target);
8559 case IA64_BUILTIN_FETCH_AND_XOR_SI:
8560 case IA64_BUILTIN_FETCH_AND_XOR_DI:
8561 return ia64_expand_fetch_and_op (xor_optab, mode, arglist, target);
8563 case IA64_BUILTIN_FETCH_AND_NAND_SI:
8564 case IA64_BUILTIN_FETCH_AND_NAND_DI:
8565 return ia64_expand_fetch_and_op (one_cmpl_optab, mode, arglist, target);
8567 case IA64_BUILTIN_ADD_AND_FETCH_SI:
8568 case IA64_BUILTIN_ADD_AND_FETCH_DI:
8569 return ia64_expand_op_and_fetch (add_optab, mode, arglist, target);
8571 case IA64_BUILTIN_SUB_AND_FETCH_SI:
8572 case IA64_BUILTIN_SUB_AND_FETCH_DI:
8573 return ia64_expand_op_and_fetch (sub_optab, mode, arglist, target);
8575 case IA64_BUILTIN_OR_AND_FETCH_SI:
8576 case IA64_BUILTIN_OR_AND_FETCH_DI:
8577 return ia64_expand_op_and_fetch (ior_optab, mode, arglist, target);
8579 case IA64_BUILTIN_AND_AND_FETCH_SI:
8580 case IA64_BUILTIN_AND_AND_FETCH_DI:
8581 return ia64_expand_op_and_fetch (and_optab, mode, arglist, target);
8583 case IA64_BUILTIN_XOR_AND_FETCH_SI:
8584 case IA64_BUILTIN_XOR_AND_FETCH_DI:
8585 return ia64_expand_op_and_fetch (xor_optab, mode, arglist, target);
8587 case IA64_BUILTIN_NAND_AND_FETCH_SI:
8588 case IA64_BUILTIN_NAND_AND_FETCH_DI:
8589 return ia64_expand_op_and_fetch (one_cmpl_optab, mode, arglist, target);
8591 default:
8592 break;
8595 return NULL_RTX;
8598 /* For the HP-UX IA64 aggregate parameters are passed stored in the
8599 most significant bits of the stack slot. */
8601 enum direction
8602 ia64_hpux_function_arg_padding (enum machine_mode mode, tree type)
8604 /* Exception to normal case for structures/unions/etc. */
8606 if (type && AGGREGATE_TYPE_P (type)
8607 && int_size_in_bytes (type) < UNITS_PER_WORD)
8608 return upward;
8610 /* Fall back to the default. */
8611 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
8614 /* Linked list of all external functions that are to be emitted by GCC.
8615 We output the name if and only if TREE_SYMBOL_REFERENCED is set in
8616 order to avoid putting out names that are never really used. */
8618 struct extern_func_list GTY(())
8620 struct extern_func_list *next;
8621 tree decl;
8624 static GTY(()) struct extern_func_list *extern_func_head;
8626 static void
8627 ia64_hpux_add_extern_decl (tree decl)
8629 struct extern_func_list *p = ggc_alloc (sizeof (struct extern_func_list));
8631 p->decl = decl;
8632 p->next = extern_func_head;
8633 extern_func_head = p;
8636 /* Print out the list of used global functions. */
8638 static void
8639 ia64_hpux_file_end (void)
8641 struct extern_func_list *p;
8643 for (p = extern_func_head; p; p = p->next)
8645 tree decl = p->decl;
8646 tree id = DECL_ASSEMBLER_NAME (decl);
8648 if (!id)
8649 abort ();
8651 if (!TREE_ASM_WRITTEN (decl) && TREE_SYMBOL_REFERENCED (id))
8653 const char *name = XSTR (XEXP (DECL_RTL (decl), 0), 0);
8655 TREE_ASM_WRITTEN (decl) = 1;
8656 (*targetm.asm_out.globalize_label) (asm_out_file, name);
8657 fputs (TYPE_ASM_OP, asm_out_file);
8658 assemble_name (asm_out_file, name);
8659 fprintf (asm_out_file, "," TYPE_OPERAND_FMT "\n", "function");
8663 extern_func_head = 0;
8666 /* Set SImode div/mod functions, init_integral_libfuncs only initializes
8667 modes of word_mode and larger. Rename the TFmode libfuncs using the
8668 HPUX conventions. __divtf3 is used for XFmode. We need to keep it for
8669 backward compatibility. */
8671 static void
8672 ia64_init_libfuncs (void)
8674 set_optab_libfunc (sdiv_optab, SImode, "__divsi3");
8675 set_optab_libfunc (udiv_optab, SImode, "__udivsi3");
8676 set_optab_libfunc (smod_optab, SImode, "__modsi3");
8677 set_optab_libfunc (umod_optab, SImode, "__umodsi3");
8679 set_optab_libfunc (add_optab, TFmode, "_U_Qfadd");
8680 set_optab_libfunc (sub_optab, TFmode, "_U_Qfsub");
8681 set_optab_libfunc (smul_optab, TFmode, "_U_Qfmpy");
8682 set_optab_libfunc (sdiv_optab, TFmode, "_U_Qfdiv");
8683 set_optab_libfunc (neg_optab, TFmode, "_U_Qfneg");
8685 set_conv_libfunc (sext_optab, TFmode, SFmode, "_U_Qfcnvff_sgl_to_quad");
8686 set_conv_libfunc (sext_optab, TFmode, DFmode, "_U_Qfcnvff_dbl_to_quad");
8687 set_conv_libfunc (sext_optab, TFmode, XFmode, "_U_Qfcnvff_f80_to_quad");
8688 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_U_Qfcnvff_quad_to_sgl");
8689 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_U_Qfcnvff_quad_to_dbl");
8690 set_conv_libfunc (trunc_optab, XFmode, TFmode, "_U_Qfcnvff_quad_to_f80");
8692 set_conv_libfunc (sfix_optab, SImode, TFmode, "_U_Qfcnvfxt_quad_to_sgl");
8693 set_conv_libfunc (sfix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_dbl");
8694 set_conv_libfunc (ufix_optab, SImode, TFmode, "_U_Qfcnvfxut_quad_to_sgl");
8695 set_conv_libfunc (ufix_optab, DImode, TFmode, "_U_Qfcnvfxut_quad_to_dbl");
8697 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_U_Qfcnvxf_sgl_to_quad");
8698 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_U_Qfcnvxf_dbl_to_quad");
8701 /* Rename all the TFmode libfuncs using the HPUX conventions. */
8703 static void
8704 ia64_hpux_init_libfuncs (void)
8706 ia64_init_libfuncs ();
8708 set_optab_libfunc (smin_optab, TFmode, "_U_Qfmin");
8709 set_optab_libfunc (smax_optab, TFmode, "_U_Qfmax");
8710 set_optab_libfunc (abs_optab, TFmode, "_U_Qfabs");
8712 /* ia64_expand_compare uses this. */
8713 cmptf_libfunc = init_one_libfunc ("_U_Qfcmp");
8715 /* These should never be used. */
8716 set_optab_libfunc (eq_optab, TFmode, 0);
8717 set_optab_libfunc (ne_optab, TFmode, 0);
8718 set_optab_libfunc (gt_optab, TFmode, 0);
8719 set_optab_libfunc (ge_optab, TFmode, 0);
8720 set_optab_libfunc (lt_optab, TFmode, 0);
8721 set_optab_libfunc (le_optab, TFmode, 0);
8724 /* Rename the division and modulus functions in VMS. */
8726 static void
8727 ia64_vms_init_libfuncs (void)
8729 set_optab_libfunc (sdiv_optab, SImode, "OTS$DIV_I");
8730 set_optab_libfunc (sdiv_optab, DImode, "OTS$DIV_L");
8731 set_optab_libfunc (udiv_optab, SImode, "OTS$DIV_UI");
8732 set_optab_libfunc (udiv_optab, DImode, "OTS$DIV_UL");
8733 set_optab_libfunc (smod_optab, SImode, "OTS$REM_I");
8734 set_optab_libfunc (smod_optab, DImode, "OTS$REM_L");
8735 set_optab_libfunc (umod_optab, SImode, "OTS$REM_UI");
8736 set_optab_libfunc (umod_optab, DImode, "OTS$REM_UL");
8739 /* Rename the TFmode libfuncs available from soft-fp in glibc using
8740 the HPUX conventions. */
8742 static void
8743 ia64_sysv4_init_libfuncs (void)
8745 ia64_init_libfuncs ();
8747 /* These functions are not part of the HPUX TFmode interface. We
8748 use them instead of _U_Qfcmp, which doesn't work the way we
8749 expect. */
8750 set_optab_libfunc (eq_optab, TFmode, "_U_Qfeq");
8751 set_optab_libfunc (ne_optab, TFmode, "_U_Qfne");
8752 set_optab_libfunc (gt_optab, TFmode, "_U_Qfgt");
8753 set_optab_libfunc (ge_optab, TFmode, "_U_Qfge");
8754 set_optab_libfunc (lt_optab, TFmode, "_U_Qflt");
8755 set_optab_libfunc (le_optab, TFmode, "_U_Qfle");
8757 /* We leave out _U_Qfmin, _U_Qfmax and _U_Qfabs since soft-fp in
8758 glibc doesn't have them. */
8761 /* Switch to the section to which we should output X. The only thing
8762 special we do here is to honor small data. */
8764 static void
8765 ia64_select_rtx_section (enum machine_mode mode, rtx x,
8766 unsigned HOST_WIDE_INT align)
8768 if (GET_MODE_SIZE (mode) > 0
8769 && GET_MODE_SIZE (mode) <= ia64_section_threshold)
8770 sdata_section ();
8771 else
8772 default_elf_select_rtx_section (mode, x, align);
8775 /* It is illegal to have relocations in shared segments on AIX and HPUX.
8776 Pretend flag_pic is always set. */
8778 static void
8779 ia64_rwreloc_select_section (tree exp, int reloc, unsigned HOST_WIDE_INT align)
8781 default_elf_select_section_1 (exp, reloc, align, true);
8784 static void
8785 ia64_rwreloc_unique_section (tree decl, int reloc)
8787 default_unique_section_1 (decl, reloc, true);
8790 static void
8791 ia64_rwreloc_select_rtx_section (enum machine_mode mode, rtx x,
8792 unsigned HOST_WIDE_INT align)
8794 int save_pic = flag_pic;
8795 flag_pic = 1;
8796 ia64_select_rtx_section (mode, x, align);
8797 flag_pic = save_pic;
8800 #ifndef TARGET_RWRELOC
8801 #define TARGET_RWRELOC flag_pic
8802 #endif
8804 static unsigned int
8805 ia64_section_type_flags (tree decl, const char *name, int reloc)
8807 unsigned int flags = 0;
8809 if (strcmp (name, ".sdata") == 0
8810 || strncmp (name, ".sdata.", 7) == 0
8811 || strncmp (name, ".gnu.linkonce.s.", 16) == 0
8812 || strncmp (name, ".sdata2.", 8) == 0
8813 || strncmp (name, ".gnu.linkonce.s2.", 17) == 0
8814 || strcmp (name, ".sbss") == 0
8815 || strncmp (name, ".sbss.", 6) == 0
8816 || strncmp (name, ".gnu.linkonce.sb.", 17) == 0)
8817 flags = SECTION_SMALL;
8819 flags |= default_section_type_flags_1 (decl, name, reloc, TARGET_RWRELOC);
8820 return flags;
8823 /* Returns true if FNTYPE (a FUNCTION_TYPE or a METHOD_TYPE) returns a
8824 structure type and that the address of that type should be passed
8825 in out0, rather than in r8. */
8827 static bool
8828 ia64_struct_retval_addr_is_first_parm_p (tree fntype)
8830 tree ret_type = TREE_TYPE (fntype);
8832 /* The Itanium C++ ABI requires that out0, rather than r8, be used
8833 as the structure return address parameter, if the return value
8834 type has a non-trivial copy constructor or destructor. It is not
8835 clear if this same convention should be used for other
8836 programming languages. Until G++ 3.4, we incorrectly used r8 for
8837 these return values. */
8838 return (abi_version_at_least (2)
8839 && ret_type
8840 && TYPE_MODE (ret_type) == BLKmode
8841 && TREE_ADDRESSABLE (ret_type)
8842 && strcmp (lang_hooks.name, "GNU C++") == 0);
8845 /* Output the assembler code for a thunk function. THUNK_DECL is the
8846 declaration for the thunk function itself, FUNCTION is the decl for
8847 the target function. DELTA is an immediate constant offset to be
8848 added to THIS. If VCALL_OFFSET is nonzero, the word at
8849 *(*this + vcall_offset) should be added to THIS. */
8851 static void
8852 ia64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED,
8853 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
8854 tree function)
8856 rtx this, insn, funexp;
8857 unsigned int this_parmno;
8858 unsigned int this_regno;
8860 reload_completed = 1;
8861 epilogue_completed = 1;
8862 no_new_pseudos = 1;
8863 reset_block_changes ();
8865 /* Set things up as ia64_expand_prologue might. */
8866 last_scratch_gr_reg = 15;
8868 memset (&current_frame_info, 0, sizeof (current_frame_info));
8869 current_frame_info.spill_cfa_off = -16;
8870 current_frame_info.n_input_regs = 1;
8871 current_frame_info.need_regstk = (TARGET_REG_NAMES != 0);
8873 /* Mark the end of the (empty) prologue. */
8874 emit_note (NOTE_INSN_PROLOGUE_END);
8876 /* Figure out whether "this" will be the first parameter (the
8877 typical case) or the second parameter (as happens when the
8878 virtual function returns certain class objects). */
8879 this_parmno
8880 = (ia64_struct_retval_addr_is_first_parm_p (TREE_TYPE (thunk))
8881 ? 1 : 0);
8882 this_regno = IN_REG (this_parmno);
8883 if (!TARGET_REG_NAMES)
8884 reg_names[this_regno] = ia64_reg_numbers[this_parmno];
8886 this = gen_rtx_REG (Pmode, this_regno);
8887 if (TARGET_ILP32)
8889 rtx tmp = gen_rtx_REG (ptr_mode, this_regno);
8890 REG_POINTER (tmp) = 1;
8891 if (delta && CONST_OK_FOR_I (delta))
8893 emit_insn (gen_ptr_extend_plus_imm (this, tmp, GEN_INT (delta)));
8894 delta = 0;
8896 else
8897 emit_insn (gen_ptr_extend (this, tmp));
8900 /* Apply the constant offset, if required. */
8901 if (delta)
8903 rtx delta_rtx = GEN_INT (delta);
8905 if (!CONST_OK_FOR_I (delta))
8907 rtx tmp = gen_rtx_REG (Pmode, 2);
8908 emit_move_insn (tmp, delta_rtx);
8909 delta_rtx = tmp;
8911 emit_insn (gen_adddi3 (this, this, delta_rtx));
8914 /* Apply the offset from the vtable, if required. */
8915 if (vcall_offset)
8917 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
8918 rtx tmp = gen_rtx_REG (Pmode, 2);
8920 if (TARGET_ILP32)
8922 rtx t = gen_rtx_REG (ptr_mode, 2);
8923 REG_POINTER (t) = 1;
8924 emit_move_insn (t, gen_rtx_MEM (ptr_mode, this));
8925 if (CONST_OK_FOR_I (vcall_offset))
8927 emit_insn (gen_ptr_extend_plus_imm (tmp, t,
8928 vcall_offset_rtx));
8929 vcall_offset = 0;
8931 else
8932 emit_insn (gen_ptr_extend (tmp, t));
8934 else
8935 emit_move_insn (tmp, gen_rtx_MEM (Pmode, this));
8937 if (vcall_offset)
8939 if (!CONST_OK_FOR_J (vcall_offset))
8941 rtx tmp2 = gen_rtx_REG (Pmode, next_scratch_gr_reg ());
8942 emit_move_insn (tmp2, vcall_offset_rtx);
8943 vcall_offset_rtx = tmp2;
8945 emit_insn (gen_adddi3 (tmp, tmp, vcall_offset_rtx));
8948 if (TARGET_ILP32)
8949 emit_move_insn (gen_rtx_REG (ptr_mode, 2),
8950 gen_rtx_MEM (ptr_mode, tmp));
8951 else
8952 emit_move_insn (tmp, gen_rtx_MEM (Pmode, tmp));
8954 emit_insn (gen_adddi3 (this, this, tmp));
8957 /* Generate a tail call to the target function. */
8958 if (! TREE_USED (function))
8960 assemble_external (function);
8961 TREE_USED (function) = 1;
8963 funexp = XEXP (DECL_RTL (function), 0);
8964 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8965 ia64_expand_call (NULL_RTX, funexp, NULL_RTX, 1);
8966 insn = get_last_insn ();
8967 SIBLING_CALL_P (insn) = 1;
8969 /* Code generation for calls relies on splitting. */
8970 reload_completed = 1;
8971 epilogue_completed = 1;
8972 try_split (PATTERN (insn), insn, 0);
8974 emit_barrier ();
8976 /* Run just enough of rest_of_compilation to get the insns emitted.
8977 There's not really enough bulk here to make other passes such as
8978 instruction scheduling worth while. Note that use_thunk calls
8979 assemble_start_function and assemble_end_function. */
8981 insn_locators_initialize ();
8982 emit_all_insn_group_barriers (NULL);
8983 insn = get_insns ();
8984 shorten_branches (insn);
8985 final_start_function (insn, file, 1);
8986 final (insn, file, 1, 0);
8987 final_end_function ();
8989 reload_completed = 0;
8990 epilogue_completed = 0;
8991 no_new_pseudos = 0;
8994 /* Worker function for TARGET_STRUCT_VALUE_RTX. */
8996 static rtx
8997 ia64_struct_value_rtx (tree fntype,
8998 int incoming ATTRIBUTE_UNUSED)
9000 if (fntype && ia64_struct_retval_addr_is_first_parm_p (fntype))
9001 return NULL_RTX;
9002 return gen_rtx_REG (Pmode, GR_REG (8));
9005 static bool
9006 ia64_scalar_mode_supported_p (enum machine_mode mode)
9008 switch (mode)
9010 case QImode:
9011 case HImode:
9012 case SImode:
9013 case DImode:
9014 case TImode:
9015 return true;
9017 case SFmode:
9018 case DFmode:
9019 case XFmode:
9020 return true;
9022 case TFmode:
9023 return TARGET_HPUX;
9025 default:
9026 return false;
9030 static bool
9031 ia64_vector_mode_supported_p (enum machine_mode mode)
9033 switch (mode)
9035 case V8QImode:
9036 case V4HImode:
9037 case V2SImode:
9038 return true;
9040 case V2SFmode:
9041 return true;
9043 default:
9044 return false;
9048 #include "gt-ia64.h"