1 /* Subroutines for insn-output.c for SPARC.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
4 Free Software Foundation, Inc.
5 Contributed by Michael Tiemann (tiemann@cygnus.com)
6 64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans,
9 This file is part of GCC.
11 GCC is free software; you can redistribute it and/or modify
12 it under the terms of the GNU General Public License as published by
13 the Free Software Foundation; either version 3, or (at your option)
16 GCC is distributed in the hope that it will be useful,
17 but WITHOUT ANY WARRANTY; without even the implied warranty of
18 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 GNU General Public License for more details.
21 You should have received a copy of the GNU General Public License
22 along with GCC; see the file COPYING3. If not see
23 <http://www.gnu.org/licenses/>. */
27 #include "coretypes.h"
32 #include "hard-reg-set.h"
33 #include "insn-config.h"
34 #include "insn-codes.h"
35 #include "conditions.h"
37 #include "insn-attr.h"
48 #include "target-def.h"
49 #include "cfglayout.h"
51 #include "langhooks.h"
54 #include "dwarf2out.h"
58 struct processor_costs cypress_costs = {
59 COSTS_N_INSNS (2), /* int load */
60 COSTS_N_INSNS (2), /* int signed load */
61 COSTS_N_INSNS (2), /* int zeroed load */
62 COSTS_N_INSNS (2), /* float load */
63 COSTS_N_INSNS (5), /* fmov, fneg, fabs */
64 COSTS_N_INSNS (5), /* fadd, fsub */
65 COSTS_N_INSNS (1), /* fcmp */
66 COSTS_N_INSNS (1), /* fmov, fmovr */
67 COSTS_N_INSNS (7), /* fmul */
68 COSTS_N_INSNS (37), /* fdivs */
69 COSTS_N_INSNS (37), /* fdivd */
70 COSTS_N_INSNS (63), /* fsqrts */
71 COSTS_N_INSNS (63), /* fsqrtd */
72 COSTS_N_INSNS (1), /* imul */
73 COSTS_N_INSNS (1), /* imulX */
74 0, /* imul bit factor */
75 COSTS_N_INSNS (1), /* idiv */
76 COSTS_N_INSNS (1), /* idivX */
77 COSTS_N_INSNS (1), /* movcc/movr */
78 0, /* shift penalty */
82 struct processor_costs supersparc_costs = {
83 COSTS_N_INSNS (1), /* int load */
84 COSTS_N_INSNS (1), /* int signed load */
85 COSTS_N_INSNS (1), /* int zeroed load */
86 COSTS_N_INSNS (0), /* float load */
87 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
88 COSTS_N_INSNS (3), /* fadd, fsub */
89 COSTS_N_INSNS (3), /* fcmp */
90 COSTS_N_INSNS (1), /* fmov, fmovr */
91 COSTS_N_INSNS (3), /* fmul */
92 COSTS_N_INSNS (6), /* fdivs */
93 COSTS_N_INSNS (9), /* fdivd */
94 COSTS_N_INSNS (12), /* fsqrts */
95 COSTS_N_INSNS (12), /* fsqrtd */
96 COSTS_N_INSNS (4), /* imul */
97 COSTS_N_INSNS (4), /* imulX */
98 0, /* imul bit factor */
99 COSTS_N_INSNS (4), /* idiv */
100 COSTS_N_INSNS (4), /* idivX */
101 COSTS_N_INSNS (1), /* movcc/movr */
102 1, /* shift penalty */
106 struct processor_costs hypersparc_costs = {
107 COSTS_N_INSNS (1), /* int load */
108 COSTS_N_INSNS (1), /* int signed load */
109 COSTS_N_INSNS (1), /* int zeroed load */
110 COSTS_N_INSNS (1), /* float load */
111 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
112 COSTS_N_INSNS (1), /* fadd, fsub */
113 COSTS_N_INSNS (1), /* fcmp */
114 COSTS_N_INSNS (1), /* fmov, fmovr */
115 COSTS_N_INSNS (1), /* fmul */
116 COSTS_N_INSNS (8), /* fdivs */
117 COSTS_N_INSNS (12), /* fdivd */
118 COSTS_N_INSNS (17), /* fsqrts */
119 COSTS_N_INSNS (17), /* fsqrtd */
120 COSTS_N_INSNS (17), /* imul */
121 COSTS_N_INSNS (17), /* imulX */
122 0, /* imul bit factor */
123 COSTS_N_INSNS (17), /* idiv */
124 COSTS_N_INSNS (17), /* idivX */
125 COSTS_N_INSNS (1), /* movcc/movr */
126 0, /* shift penalty */
130 struct processor_costs sparclet_costs = {
131 COSTS_N_INSNS (3), /* int load */
132 COSTS_N_INSNS (3), /* int signed load */
133 COSTS_N_INSNS (1), /* int zeroed load */
134 COSTS_N_INSNS (1), /* float load */
135 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
136 COSTS_N_INSNS (1), /* fadd, fsub */
137 COSTS_N_INSNS (1), /* fcmp */
138 COSTS_N_INSNS (1), /* fmov, fmovr */
139 COSTS_N_INSNS (1), /* fmul */
140 COSTS_N_INSNS (1), /* fdivs */
141 COSTS_N_INSNS (1), /* fdivd */
142 COSTS_N_INSNS (1), /* fsqrts */
143 COSTS_N_INSNS (1), /* fsqrtd */
144 COSTS_N_INSNS (5), /* imul */
145 COSTS_N_INSNS (5), /* imulX */
146 0, /* imul bit factor */
147 COSTS_N_INSNS (5), /* idiv */
148 COSTS_N_INSNS (5), /* idivX */
149 COSTS_N_INSNS (1), /* movcc/movr */
150 0, /* shift penalty */
154 struct processor_costs ultrasparc_costs = {
155 COSTS_N_INSNS (2), /* int load */
156 COSTS_N_INSNS (3), /* int signed load */
157 COSTS_N_INSNS (2), /* int zeroed load */
158 COSTS_N_INSNS (2), /* float load */
159 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
160 COSTS_N_INSNS (4), /* fadd, fsub */
161 COSTS_N_INSNS (1), /* fcmp */
162 COSTS_N_INSNS (2), /* fmov, fmovr */
163 COSTS_N_INSNS (4), /* fmul */
164 COSTS_N_INSNS (13), /* fdivs */
165 COSTS_N_INSNS (23), /* fdivd */
166 COSTS_N_INSNS (13), /* fsqrts */
167 COSTS_N_INSNS (23), /* fsqrtd */
168 COSTS_N_INSNS (4), /* imul */
169 COSTS_N_INSNS (4), /* imulX */
170 2, /* imul bit factor */
171 COSTS_N_INSNS (37), /* idiv */
172 COSTS_N_INSNS (68), /* idivX */
173 COSTS_N_INSNS (2), /* movcc/movr */
174 2, /* shift penalty */
178 struct processor_costs ultrasparc3_costs = {
179 COSTS_N_INSNS (2), /* int load */
180 COSTS_N_INSNS (3), /* int signed load */
181 COSTS_N_INSNS (3), /* int zeroed load */
182 COSTS_N_INSNS (2), /* float load */
183 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
184 COSTS_N_INSNS (4), /* fadd, fsub */
185 COSTS_N_INSNS (5), /* fcmp */
186 COSTS_N_INSNS (3), /* fmov, fmovr */
187 COSTS_N_INSNS (4), /* fmul */
188 COSTS_N_INSNS (17), /* fdivs */
189 COSTS_N_INSNS (20), /* fdivd */
190 COSTS_N_INSNS (20), /* fsqrts */
191 COSTS_N_INSNS (29), /* fsqrtd */
192 COSTS_N_INSNS (6), /* imul */
193 COSTS_N_INSNS (6), /* imulX */
194 0, /* imul bit factor */
195 COSTS_N_INSNS (40), /* idiv */
196 COSTS_N_INSNS (71), /* idivX */
197 COSTS_N_INSNS (2), /* movcc/movr */
198 0, /* shift penalty */
202 struct processor_costs niagara_costs = {
203 COSTS_N_INSNS (3), /* int load */
204 COSTS_N_INSNS (3), /* int signed load */
205 COSTS_N_INSNS (3), /* int zeroed load */
206 COSTS_N_INSNS (9), /* float load */
207 COSTS_N_INSNS (8), /* fmov, fneg, fabs */
208 COSTS_N_INSNS (8), /* fadd, fsub */
209 COSTS_N_INSNS (26), /* fcmp */
210 COSTS_N_INSNS (8), /* fmov, fmovr */
211 COSTS_N_INSNS (29), /* fmul */
212 COSTS_N_INSNS (54), /* fdivs */
213 COSTS_N_INSNS (83), /* fdivd */
214 COSTS_N_INSNS (100), /* fsqrts - not implemented in hardware */
215 COSTS_N_INSNS (100), /* fsqrtd - not implemented in hardware */
216 COSTS_N_INSNS (11), /* imul */
217 COSTS_N_INSNS (11), /* imulX */
218 0, /* imul bit factor */
219 COSTS_N_INSNS (72), /* idiv */
220 COSTS_N_INSNS (72), /* idivX */
221 COSTS_N_INSNS (1), /* movcc/movr */
222 0, /* shift penalty */
226 struct processor_costs niagara2_costs = {
227 COSTS_N_INSNS (3), /* int load */
228 COSTS_N_INSNS (3), /* int signed load */
229 COSTS_N_INSNS (3), /* int zeroed load */
230 COSTS_N_INSNS (3), /* float load */
231 COSTS_N_INSNS (6), /* fmov, fneg, fabs */
232 COSTS_N_INSNS (6), /* fadd, fsub */
233 COSTS_N_INSNS (6), /* fcmp */
234 COSTS_N_INSNS (6), /* fmov, fmovr */
235 COSTS_N_INSNS (6), /* fmul */
236 COSTS_N_INSNS (19), /* fdivs */
237 COSTS_N_INSNS (33), /* fdivd */
238 COSTS_N_INSNS (19), /* fsqrts */
239 COSTS_N_INSNS (33), /* fsqrtd */
240 COSTS_N_INSNS (5), /* imul */
241 COSTS_N_INSNS (5), /* imulX */
242 0, /* imul bit factor */
243 COSTS_N_INSNS (31), /* idiv, average of 12 - 41 cycle range */
244 COSTS_N_INSNS (31), /* idivX, average of 12 - 41 cycle range */
245 COSTS_N_INSNS (1), /* movcc/movr */
246 0, /* shift penalty */
249 const struct processor_costs *sparc_costs = &cypress_costs;
251 #ifdef HAVE_AS_RELAX_OPTION
252 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
253 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
254 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
255 somebody does not branch between the sethi and jmp. */
256 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
258 #define LEAF_SIBCALL_SLOT_RESERVED_P \
259 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
262 /* Global variables for machine-dependent things. */
264 /* Size of frame. Need to know this to emit return insns from leaf procedures.
265 ACTUAL_FSIZE is set by sparc_compute_frame_size() which is called during the
266 reload pass. This is important as the value is later used for scheduling
267 (to see what can go in a delay slot).
268 APPARENT_FSIZE is the size of the stack less the register save area and less
269 the outgoing argument area. It is used when saving call preserved regs. */
270 static HOST_WIDE_INT apparent_fsize;
271 static HOST_WIDE_INT actual_fsize;
273 /* Number of live general or floating point registers needed to be
274 saved (as 4-byte quantities). */
275 static int num_gfregs;
277 /* The alias set for prologue/epilogue register save/restore. */
278 static GTY(()) alias_set_type sparc_sr_alias_set;
280 /* The alias set for the structure return value. */
281 static GTY(()) alias_set_type struct_value_alias_set;
283 /* Vector to say how input registers are mapped to output registers.
284 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
285 eliminate it. You must use -fomit-frame-pointer to get that. */
286 char leaf_reg_remap[] =
287 { 0, 1, 2, 3, 4, 5, 6, 7,
288 -1, -1, -1, -1, -1, -1, 14, -1,
289 -1, -1, -1, -1, -1, -1, -1, -1,
290 8, 9, 10, 11, 12, 13, -1, 15,
292 32, 33, 34, 35, 36, 37, 38, 39,
293 40, 41, 42, 43, 44, 45, 46, 47,
294 48, 49, 50, 51, 52, 53, 54, 55,
295 56, 57, 58, 59, 60, 61, 62, 63,
296 64, 65, 66, 67, 68, 69, 70, 71,
297 72, 73, 74, 75, 76, 77, 78, 79,
298 80, 81, 82, 83, 84, 85, 86, 87,
299 88, 89, 90, 91, 92, 93, 94, 95,
300 96, 97, 98, 99, 100};
302 /* Vector, indexed by hard register number, which contains 1
303 for a register that is allowable in a candidate for leaf
304 function treatment. */
305 char sparc_leaf_regs[] =
306 { 1, 1, 1, 1, 1, 1, 1, 1,
307 0, 0, 0, 0, 0, 0, 1, 0,
308 0, 0, 0, 0, 0, 0, 0, 0,
309 1, 1, 1, 1, 1, 1, 0, 1,
310 1, 1, 1, 1, 1, 1, 1, 1,
311 1, 1, 1, 1, 1, 1, 1, 1,
312 1, 1, 1, 1, 1, 1, 1, 1,
313 1, 1, 1, 1, 1, 1, 1, 1,
314 1, 1, 1, 1, 1, 1, 1, 1,
315 1, 1, 1, 1, 1, 1, 1, 1,
316 1, 1, 1, 1, 1, 1, 1, 1,
317 1, 1, 1, 1, 1, 1, 1, 1,
320 struct GTY(()) machine_function
322 /* Some local-dynamic TLS symbol name. */
323 const char *some_ld_name;
325 /* True if the current function is leaf and uses only leaf regs,
326 so that the SPARC leaf function optimization can be applied.
327 Private version of current_function_uses_only_leaf_regs, see
328 sparc_expand_prologue for the rationale. */
331 /* True if the data calculated by sparc_expand_prologue are valid. */
332 bool prologue_data_valid_p;
335 #define sparc_leaf_function_p cfun->machine->leaf_function_p
336 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
338 /* Register we pretend to think the frame pointer is allocated to.
339 Normally, this is %fp, but if we are in a leaf procedure, this
340 is %sp+"something". We record "something" separately as it may
341 be too big for reg+constant addressing. */
342 static rtx frame_base_reg;
343 static HOST_WIDE_INT frame_base_offset;
345 /* 1 if the next opcode is to be specially indented. */
346 int sparc_indent_opcode = 0;
348 static bool sparc_handle_option (size_t, const char *, int);
349 static void sparc_init_modes (void);
350 static void scan_record_type (tree, int *, int *, int *);
351 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
352 tree, int, int, int *, int *);
354 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
355 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
357 static void sparc_emit_set_const32 (rtx, rtx);
358 static void sparc_emit_set_const64 (rtx, rtx);
359 static void sparc_output_addr_vec (rtx);
360 static void sparc_output_addr_diff_vec (rtx);
361 static void sparc_output_deferred_case_vectors (void);
362 static bool sparc_legitimate_address_p (enum machine_mode, rtx, bool);
363 static rtx sparc_builtin_saveregs (void);
364 static int epilogue_renumber (rtx *, int);
365 static bool sparc_assemble_integer (rtx, unsigned int, int);
366 static int set_extends (rtx);
367 static void load_pic_register (void);
368 static int save_or_restore_regs (int, int, rtx, int, int);
369 static void emit_save_or_restore_regs (int);
370 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
371 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
372 static void sparc_solaris_elf_asm_named_section (const char *, unsigned int,
373 tree) ATTRIBUTE_UNUSED;
374 static int sparc_adjust_cost (rtx, rtx, rtx, int);
375 static int sparc_issue_rate (void);
376 static void sparc_sched_init (FILE *, int, int);
377 static int sparc_use_sched_lookahead (void);
379 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
380 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
381 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
382 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
383 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
385 static bool sparc_function_ok_for_sibcall (tree, tree);
386 static void sparc_init_libfuncs (void);
387 static void sparc_init_builtins (void);
388 static void sparc_vis_init_builtins (void);
389 static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
390 static tree sparc_fold_builtin (tree, int, tree *, bool);
391 static int sparc_vis_mul8x16 (int, int);
392 static tree sparc_handle_vis_mul8x16 (int, tree, tree, tree);
393 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
394 HOST_WIDE_INT, tree);
395 static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT,
396 HOST_WIDE_INT, const_tree);
397 static struct machine_function * sparc_init_machine_status (void);
398 static bool sparc_cannot_force_const_mem (rtx);
399 static rtx sparc_tls_get_addr (void);
400 static rtx sparc_tls_got (void);
401 static const char *get_some_local_dynamic_name (void);
402 static int get_some_local_dynamic_name_1 (rtx *, void *);
403 static bool sparc_rtx_costs (rtx, int, int, int *, bool);
404 static bool sparc_promote_prototypes (const_tree);
405 static rtx sparc_function_value (const_tree, const_tree, bool);
406 static rtx sparc_libcall_value (enum machine_mode, const_rtx);
407 static bool sparc_function_value_regno_p (const unsigned int);
408 static rtx sparc_struct_value_rtx (tree, int);
409 static enum machine_mode sparc_promote_function_mode (const_tree, enum machine_mode,
410 int *, const_tree, int);
411 static bool sparc_return_in_memory (const_tree, const_tree);
412 static bool sparc_strict_argument_naming (CUMULATIVE_ARGS *);
413 static void sparc_va_start (tree, rtx);
414 static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
415 static bool sparc_vector_mode_supported_p (enum machine_mode);
416 static bool sparc_tls_referenced_p (rtx);
417 static rtx legitimize_tls_address (rtx);
418 static rtx legitimize_pic_address (rtx, rtx);
419 static rtx sparc_legitimize_address (rtx, rtx, enum machine_mode);
420 static bool sparc_mode_dependent_address_p (const_rtx);
421 static bool sparc_pass_by_reference (CUMULATIVE_ARGS *,
422 enum machine_mode, const_tree, bool);
423 static int sparc_arg_partial_bytes (CUMULATIVE_ARGS *,
424 enum machine_mode, tree, bool);
425 static void sparc_dwarf_handle_frame_unspec (const char *, rtx, int);
426 static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
427 static void sparc_file_end (void);
428 static bool sparc_frame_pointer_required (void);
429 static bool sparc_can_eliminate (const int, const int);
430 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
431 static const char *sparc_mangle_type (const_tree);
433 static void sparc_trampoline_init (rtx, tree, rtx);
435 #ifdef SUBTARGET_ATTRIBUTE_TABLE
436 /* Table of valid machine attributes. */
437 static const struct attribute_spec sparc_attribute_table[] =
439 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
440 SUBTARGET_ATTRIBUTE_TABLE,
441 { NULL, 0, 0, false, false, false, NULL }
445 /* Option handling. */
448 enum cmodel sparc_cmodel;
450 char sparc_hard_reg_printed[8];
452 struct sparc_cpu_select sparc_select[] =
454 /* switch name, tune arch */
455 { (char *)0, "default", 1, 1 },
456 { (char *)0, "-mcpu=", 1, 1 },
457 { (char *)0, "-mtune=", 1, 0 },
461 /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */
462 enum processor_type sparc_cpu;
464 /* Whether
\fan FPU option was specified. */
465 static bool fpu_option_set = false;
467 /* Initialize the GCC target structure. */
469 /* The default is to use .half rather than .short for aligned HI objects. */
470 #undef TARGET_ASM_ALIGNED_HI_OP
471 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
473 #undef TARGET_ASM_UNALIGNED_HI_OP
474 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
475 #undef TARGET_ASM_UNALIGNED_SI_OP
476 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
477 #undef TARGET_ASM_UNALIGNED_DI_OP
478 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
480 /* The target hook has to handle DI-mode values. */
481 #undef TARGET_ASM_INTEGER
482 #define TARGET_ASM_INTEGER sparc_assemble_integer
484 #undef TARGET_ASM_FUNCTION_PROLOGUE
485 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
486 #undef TARGET_ASM_FUNCTION_EPILOGUE
487 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
489 #undef TARGET_SCHED_ADJUST_COST
490 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
491 #undef TARGET_SCHED_ISSUE_RATE
492 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
493 #undef TARGET_SCHED_INIT
494 #define TARGET_SCHED_INIT sparc_sched_init
495 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
496 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
498 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
499 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
501 #undef TARGET_INIT_LIBFUNCS
502 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
503 #undef TARGET_INIT_BUILTINS
504 #define TARGET_INIT_BUILTINS sparc_init_builtins
506 #undef TARGET_LEGITIMIZE_ADDRESS
507 #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address
508 #undef TARGET_MODE_DEPENDENT_ADDRESS_P
509 #define TARGET_MODE_DEPENDENT_ADDRESS_P sparc_mode_dependent_address_p
511 #undef TARGET_EXPAND_BUILTIN
512 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
513 #undef TARGET_FOLD_BUILTIN
514 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
517 #undef TARGET_HAVE_TLS
518 #define TARGET_HAVE_TLS true
521 #undef TARGET_CANNOT_FORCE_CONST_MEM
522 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
524 #undef TARGET_ASM_OUTPUT_MI_THUNK
525 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
526 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
527 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
529 #undef TARGET_RTX_COSTS
530 #define TARGET_RTX_COSTS sparc_rtx_costs
531 #undef TARGET_ADDRESS_COST
532 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0
534 #undef TARGET_PROMOTE_FUNCTION_MODE
535 #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode
537 #undef TARGET_PROMOTE_PROTOTYPES
538 #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes
540 #undef TARGET_FUNCTION_VALUE
541 #define TARGET_FUNCTION_VALUE sparc_function_value
542 #undef TARGET_LIBCALL_VALUE
543 #define TARGET_LIBCALL_VALUE sparc_libcall_value
544 #undef TARGET_FUNCTION_VALUE_REGNO_P
545 #define TARGET_FUNCTION_VALUE_REGNO_P sparc_function_value_regno_p
547 #undef TARGET_STRUCT_VALUE_RTX
548 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
549 #undef TARGET_RETURN_IN_MEMORY
550 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
551 #undef TARGET_MUST_PASS_IN_STACK
552 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
553 #undef TARGET_PASS_BY_REFERENCE
554 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
555 #undef TARGET_ARG_PARTIAL_BYTES
556 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
558 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
559 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
560 #undef TARGET_STRICT_ARGUMENT_NAMING
561 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
563 #undef TARGET_EXPAND_BUILTIN_VA_START
564 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
565 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
566 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
568 #undef TARGET_VECTOR_MODE_SUPPORTED_P
569 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
571 #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC
572 #define TARGET_DWARF_HANDLE_FRAME_UNSPEC sparc_dwarf_handle_frame_unspec
574 #ifdef SUBTARGET_INSERT_ATTRIBUTES
575 #undef TARGET_INSERT_ATTRIBUTES
576 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
579 #ifdef SUBTARGET_ATTRIBUTE_TABLE
580 #undef TARGET_ATTRIBUTE_TABLE
581 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
584 #undef TARGET_RELAXED_ORDERING
585 #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING
587 #undef TARGET_DEFAULT_TARGET_FLAGS
588 #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT
589 #undef TARGET_HANDLE_OPTION
590 #define TARGET_HANDLE_OPTION sparc_handle_option
592 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
593 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
594 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
597 #undef TARGET_ASM_FILE_END
598 #define TARGET_ASM_FILE_END sparc_file_end
600 #undef TARGET_FRAME_POINTER_REQUIRED
601 #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required
603 #undef TARGET_CAN_ELIMINATE
604 #define TARGET_CAN_ELIMINATE sparc_can_eliminate
606 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
607 #undef TARGET_MANGLE_TYPE
608 #define TARGET_MANGLE_TYPE sparc_mangle_type
611 #undef TARGET_LEGITIMATE_ADDRESS_P
612 #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p
614 #undef TARGET_TRAMPOLINE_INIT
615 #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init
617 struct gcc_target targetm = TARGET_INITIALIZER;
619 /* Implement TARGET_HANDLE_OPTION. */
622 sparc_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED)
627 case OPT_mhard_float:
628 case OPT_msoft_float:
629 fpu_option_set = true;
633 sparc_select[1].string = arg;
637 sparc_select[2].string = arg;
644 /* Validate and override various options, and do some machine dependent
648 sparc_override_options (void)
650 static struct code_model {
651 const char *const name;
652 const enum cmodel value;
653 } const cmodels[] = {
655 { "medlow", CM_MEDLOW },
656 { "medmid", CM_MEDMID },
657 { "medany", CM_MEDANY },
658 { "embmedany", CM_EMBMEDANY },
659 { NULL, (enum cmodel) 0 }
661 const struct code_model *cmodel;
662 /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */
663 static struct cpu_default {
665 const char *const name;
666 } const cpu_default[] = {
667 /* There must be one entry here for each TARGET_CPU value. */
668 { TARGET_CPU_sparc, "cypress" },
669 { TARGET_CPU_sparclet, "tsc701" },
670 { TARGET_CPU_sparclite, "f930" },
671 { TARGET_CPU_v8, "v8" },
672 { TARGET_CPU_hypersparc, "hypersparc" },
673 { TARGET_CPU_sparclite86x, "sparclite86x" },
674 { TARGET_CPU_supersparc, "supersparc" },
675 { TARGET_CPU_v9, "v9" },
676 { TARGET_CPU_ultrasparc, "ultrasparc" },
677 { TARGET_CPU_ultrasparc3, "ultrasparc3" },
678 { TARGET_CPU_niagara, "niagara" },
679 { TARGET_CPU_niagara2, "niagara2" },
682 const struct cpu_default *def;
683 /* Table of values for -m{cpu,tune}=. */
684 static struct cpu_table {
685 const char *const name;
686 const enum processor_type processor;
689 } const cpu_table[] = {
690 { "v7", PROCESSOR_V7, MASK_ISA, 0 },
691 { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 },
692 { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 },
693 /* TI TMS390Z55 supersparc */
694 { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 },
695 { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE },
696 /* The Fujitsu MB86930 is the original sparclite chip, with no fpu.
697 The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */
698 { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE },
699 { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU },
700 { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU },
701 { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU,
703 { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET },
705 { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET },
706 { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 },
707 /* TI ultrasparc I, II, IIi */
708 { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9
709 /* Although insns using %y are deprecated, it is a clear win on current
711 |MASK_DEPRECATED_V8_INSNS},
712 /* TI ultrasparc III */
713 /* ??? Check if %y issue still holds true in ultra3. */
714 { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
716 { "niagara", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
717 { "niagara2", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9},
718 { 0, (enum processor_type) 0, 0, 0 }
720 const struct cpu_table *cpu;
721 const struct sparc_cpu_select *sel;
724 #ifndef SPARC_BI_ARCH
725 /* Check for unsupported architecture size. */
726 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
727 error ("%s is not supported by this configuration",
728 DEFAULT_ARCH32_P ? "-m64" : "-m32");
731 /* We force all 64bit archs to use 128 bit long double */
732 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
734 error ("-mlong-double-64 not allowed with -m64");
735 target_flags |= MASK_LONG_DOUBLE_128;
738 /* Code model selection. */
739 sparc_cmodel = SPARC_DEFAULT_CMODEL;
743 sparc_cmodel = CM_32;
746 if (sparc_cmodel_string != NULL)
750 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
751 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
753 if (cmodel->name == NULL)
754 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
756 sparc_cmodel = cmodel->value;
759 error ("-mcmodel= is not supported on 32 bit systems");
762 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
764 /* Set the default CPU. */
765 for (def = &cpu_default[0]; def->name; ++def)
766 if (def->cpu == TARGET_CPU_DEFAULT)
768 gcc_assert (def->name);
769 sparc_select[0].string = def->name;
771 for (sel = &sparc_select[0]; sel->name; ++sel)
775 for (cpu = &cpu_table[0]; cpu->name; ++cpu)
776 if (! strcmp (sel->string, cpu->name))
779 sparc_cpu = cpu->processor;
783 target_flags &= ~cpu->disable;
784 target_flags |= cpu->enable;
790 error ("bad value (%s) for %s switch", sel->string, sel->name);
794 /* If -mfpu or -mno-fpu was explicitly used, don't override with
795 the processor default. */
797 target_flags = (target_flags & ~MASK_FPU) | fpu;
799 /* Don't allow -mvis if FPU is disabled. */
801 target_flags &= ~MASK_VIS;
803 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
805 -m64 also implies v9. */
806 if (TARGET_VIS || TARGET_ARCH64)
808 target_flags |= MASK_V9;
809 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
812 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
813 if (TARGET_V9 && TARGET_ARCH32)
814 target_flags |= MASK_DEPRECATED_V8_INSNS;
816 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
817 if (! TARGET_V9 || TARGET_ARCH64)
818 target_flags &= ~MASK_V8PLUS;
820 /* Don't use stack biasing in 32 bit mode. */
822 target_flags &= ~MASK_STACK_BIAS;
824 /* Supply a default value for align_functions. */
825 if (align_functions == 0
826 && (sparc_cpu == PROCESSOR_ULTRASPARC
827 || sparc_cpu == PROCESSOR_ULTRASPARC3
828 || sparc_cpu == PROCESSOR_NIAGARA
829 || sparc_cpu == PROCESSOR_NIAGARA2))
830 align_functions = 32;
832 /* Validate PCC_STRUCT_RETURN. */
833 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
834 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
836 /* Only use .uaxword when compiling for a 64-bit target. */
838 targetm.asm_out.unaligned_op.di = NULL;
840 /* Do various machine dependent initializations. */
843 /* Acquire unique alias sets for our private stuff. */
844 sparc_sr_alias_set = new_alias_set ();
845 struct_value_alias_set = new_alias_set ();
847 /* Set up function hooks. */
848 init_machine_status = sparc_init_machine_status;
853 case PROCESSOR_CYPRESS:
854 sparc_costs = &cypress_costs;
857 case PROCESSOR_SPARCLITE:
858 case PROCESSOR_SUPERSPARC:
859 sparc_costs = &supersparc_costs;
863 case PROCESSOR_HYPERSPARC:
864 case PROCESSOR_SPARCLITE86X:
865 sparc_costs = &hypersparc_costs;
867 case PROCESSOR_SPARCLET:
868 case PROCESSOR_TSC701:
869 sparc_costs = &sparclet_costs;
872 case PROCESSOR_ULTRASPARC:
873 sparc_costs = &ultrasparc_costs;
875 case PROCESSOR_ULTRASPARC3:
876 sparc_costs = &ultrasparc3_costs;
878 case PROCESSOR_NIAGARA:
879 sparc_costs = &niagara_costs;
881 case PROCESSOR_NIAGARA2:
882 sparc_costs = &niagara2_costs;
886 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
887 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
888 target_flags |= MASK_LONG_DOUBLE_128;
891 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES))
892 set_param_value ("simultaneous-prefetches",
893 ((sparc_cpu == PROCESSOR_ULTRASPARC
894 || sparc_cpu == PROCESSOR_NIAGARA
895 || sparc_cpu == PROCESSOR_NIAGARA2)
897 : (sparc_cpu == PROCESSOR_ULTRASPARC3
899 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE))
900 set_param_value ("l1-cache-line-size",
901 ((sparc_cpu == PROCESSOR_ULTRASPARC
902 || sparc_cpu == PROCESSOR_ULTRASPARC3
903 || sparc_cpu == PROCESSOR_NIAGARA
904 || sparc_cpu == PROCESSOR_NIAGARA2)
908 /* Miscellaneous utilities. */
910 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
911 or branch on register contents instructions. */
914 v9_regcmp_p (enum rtx_code code)
916 return (code == EQ || code == NE || code == GE || code == LT
917 || code == LE || code == GT);
920 /* Nonzero if OP is a floating point constant which can
921 be loaded into an integer register using a single
922 sethi instruction. */
927 if (GET_CODE (op) == CONST_DOUBLE)
932 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
933 REAL_VALUE_TO_TARGET_SINGLE (r, i);
934 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
940 /* Nonzero if OP is a floating point constant which can
941 be loaded into an integer register using a single
947 if (GET_CODE (op) == CONST_DOUBLE)
952 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
953 REAL_VALUE_TO_TARGET_SINGLE (r, i);
954 return SPARC_SIMM13_P (i);
960 /* Nonzero if OP is a floating point constant which can
961 be loaded into an integer register using a high/losum
962 instruction sequence. */
965 fp_high_losum_p (rtx op)
967 /* The constraints calling this should only be in
968 SFmode move insns, so any constant which cannot
969 be moved using a single insn will do. */
970 if (GET_CODE (op) == CONST_DOUBLE)
975 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
976 REAL_VALUE_TO_TARGET_SINGLE (r, i);
977 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
983 /* Expand a move instruction. Return true if all work is done. */
986 sparc_expand_move (enum machine_mode mode, rtx *operands)
988 /* Handle sets of MEM first. */
989 if (GET_CODE (operands[0]) == MEM)
991 /* 0 is a register (or a pair of registers) on SPARC. */
992 if (register_or_zero_operand (operands[1], mode))
995 if (!reload_in_progress)
997 operands[0] = validize_mem (operands[0]);
998 operands[1] = force_reg (mode, operands[1]);
1002 /* Fixup TLS cases. */
1004 && CONSTANT_P (operands[1])
1005 && sparc_tls_referenced_p (operands [1]))
1007 operands[1] = legitimize_tls_address (operands[1]);
1011 /* Fixup PIC cases. */
1012 if (flag_pic && CONSTANT_P (operands[1]))
1014 if (pic_address_needs_scratch (operands[1]))
1015 operands[1] = legitimize_pic_address (operands[1], NULL_RTX);
1017 /* VxWorks does not impose a fixed gap between segments; the run-time
1018 gap can be different from the object-file gap. We therefore can't
1019 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1020 are absolutely sure that X is in the same segment as the GOT.
1021 Unfortunately, the flexibility of linker scripts means that we
1022 can't be sure of that in general, so assume that _G_O_T_-relative
1023 accesses are never valid on VxWorks. */
1024 if (GET_CODE (operands[1]) == LABEL_REF && !TARGET_VXWORKS_RTP)
1028 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1034 gcc_assert (TARGET_ARCH64);
1035 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1040 if (symbolic_operand (operands[1], mode))
1042 operands[1] = legitimize_pic_address (operands[1],
1044 ? operands[0] : NULL_RTX);
1049 /* If we are trying to toss an integer constant into FP registers,
1050 or loading a FP or vector constant, force it into memory. */
1051 if (CONSTANT_P (operands[1])
1052 && REG_P (operands[0])
1053 && (SPARC_FP_REG_P (REGNO (operands[0]))
1054 || SCALAR_FLOAT_MODE_P (mode)
1055 || VECTOR_MODE_P (mode)))
1057 /* emit_group_store will send such bogosity to us when it is
1058 not storing directly into memory. So fix this up to avoid
1059 crashes in output_constant_pool. */
1060 if (operands [1] == const0_rtx)
1061 operands[1] = CONST0_RTX (mode);
1063 /* We can clear FP registers if TARGET_VIS, and always other regs. */
1064 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1065 && const_zero_operand (operands[1], mode))
1068 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1069 /* We are able to build any SF constant in integer registers
1070 with at most 2 instructions. */
1072 /* And any DF constant in integer registers. */
1074 && (reload_completed || reload_in_progress))))
1077 operands[1] = force_const_mem (mode, operands[1]);
1078 if (!reload_in_progress)
1079 operands[1] = validize_mem (operands[1]);
1083 /* Accept non-constants and valid constants unmodified. */
1084 if (!CONSTANT_P (operands[1])
1085 || GET_CODE (operands[1]) == HIGH
1086 || input_operand (operands[1], mode))
1092 /* All QImode constants require only one insn, so proceed. */
1097 sparc_emit_set_const32 (operands[0], operands[1]);
1101 /* input_operand should have filtered out 32-bit mode. */
1102 sparc_emit_set_const64 (operands[0], operands[1]);
1112 /* Load OP1, a 32-bit constant, into OP0, a register.
1113 We know it can't be done in one insn when we get
1114 here, the move expander guarantees this. */
1117 sparc_emit_set_const32 (rtx op0, rtx op1)
1119 enum machine_mode mode = GET_MODE (op0);
1122 if (reload_in_progress || reload_completed)
1125 temp = gen_reg_rtx (mode);
1127 if (GET_CODE (op1) == CONST_INT)
1129 gcc_assert (!small_int_operand (op1, mode)
1130 && !const_high_operand (op1, mode));
1132 /* Emit them as real moves instead of a HIGH/LO_SUM,
1133 this way CSE can see everything and reuse intermediate
1134 values if it wants. */
1135 emit_insn (gen_rtx_SET (VOIDmode, temp,
1136 GEN_INT (INTVAL (op1)
1137 & ~(HOST_WIDE_INT)0x3ff)));
1139 emit_insn (gen_rtx_SET (VOIDmode,
1141 gen_rtx_IOR (mode, temp,
1142 GEN_INT (INTVAL (op1) & 0x3ff))));
1146 /* A symbol, emit in the traditional way. */
1147 emit_insn (gen_rtx_SET (VOIDmode, temp,
1148 gen_rtx_HIGH (mode, op1)));
1149 emit_insn (gen_rtx_SET (VOIDmode,
1150 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1154 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1155 If TEMP is nonzero, we are forbidden to use any other scratch
1156 registers. Otherwise, we are allowed to generate them as needed.
1158 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1159 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1162 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1164 rtx temp1, temp2, temp3, temp4, temp5;
1167 if (temp && GET_MODE (temp) == TImode)
1170 temp = gen_rtx_REG (DImode, REGNO (temp));
1173 /* SPARC-V9 code-model support. */
1174 switch (sparc_cmodel)
1177 /* The range spanned by all instructions in the object is less
1178 than 2^31 bytes (2GB) and the distance from any instruction
1179 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1180 than 2^31 bytes (2GB).
1182 The executable must be in the low 4TB of the virtual address
1185 sethi %hi(symbol), %temp1
1186 or %temp1, %lo(symbol), %reg */
1188 temp1 = temp; /* op0 is allowed. */
1190 temp1 = gen_reg_rtx (DImode);
1192 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1193 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1197 /* The range spanned by all instructions in the object is less
1198 than 2^31 bytes (2GB) and the distance from any instruction
1199 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1200 than 2^31 bytes (2GB).
1202 The executable must be in the low 16TB of the virtual address
1205 sethi %h44(symbol), %temp1
1206 or %temp1, %m44(symbol), %temp2
1207 sllx %temp2, 12, %temp3
1208 or %temp3, %l44(symbol), %reg */
1213 temp3 = temp; /* op0 is allowed. */
1217 temp1 = gen_reg_rtx (DImode);
1218 temp2 = gen_reg_rtx (DImode);
1219 temp3 = gen_reg_rtx (DImode);
1222 emit_insn (gen_seth44 (temp1, op1));
1223 emit_insn (gen_setm44 (temp2, temp1, op1));
1224 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1225 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1226 emit_insn (gen_setl44 (op0, temp3, op1));
1230 /* The range spanned by all instructions in the object is less
1231 than 2^31 bytes (2GB) and the distance from any instruction
1232 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1233 than 2^31 bytes (2GB).
1235 The executable can be placed anywhere in the virtual address
1238 sethi %hh(symbol), %temp1
1239 sethi %lm(symbol), %temp2
1240 or %temp1, %hm(symbol), %temp3
1241 sllx %temp3, 32, %temp4
1242 or %temp4, %temp2, %temp5
1243 or %temp5, %lo(symbol), %reg */
1246 /* It is possible that one of the registers we got for operands[2]
1247 might coincide with that of operands[0] (which is why we made
1248 it TImode). Pick the other one to use as our scratch. */
1249 if (rtx_equal_p (temp, op0))
1251 gcc_assert (ti_temp);
1252 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1255 temp2 = temp; /* op0 is _not_ allowed, see above. */
1262 temp1 = gen_reg_rtx (DImode);
1263 temp2 = gen_reg_rtx (DImode);
1264 temp3 = gen_reg_rtx (DImode);
1265 temp4 = gen_reg_rtx (DImode);
1266 temp5 = gen_reg_rtx (DImode);
1269 emit_insn (gen_sethh (temp1, op1));
1270 emit_insn (gen_setlm (temp2, op1));
1271 emit_insn (gen_sethm (temp3, temp1, op1));
1272 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1273 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1274 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1275 gen_rtx_PLUS (DImode, temp4, temp2)));
1276 emit_insn (gen_setlo (op0, temp5, op1));
1280 /* Old old old backwards compatibility kruft here.
1281 Essentially it is MEDLOW with a fixed 64-bit
1282 virtual base added to all data segment addresses.
1283 Text-segment stuff is computed like MEDANY, we can't
1284 reuse the code above because the relocation knobs
1287 Data segment: sethi %hi(symbol), %temp1
1288 add %temp1, EMBMEDANY_BASE_REG, %temp2
1289 or %temp2, %lo(symbol), %reg */
1290 if (data_segment_operand (op1, GET_MODE (op1)))
1294 temp1 = temp; /* op0 is allowed. */
1299 temp1 = gen_reg_rtx (DImode);
1300 temp2 = gen_reg_rtx (DImode);
1303 emit_insn (gen_embmedany_sethi (temp1, op1));
1304 emit_insn (gen_embmedany_brsum (temp2, temp1));
1305 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1308 /* Text segment: sethi %uhi(symbol), %temp1
1309 sethi %hi(symbol), %temp2
1310 or %temp1, %ulo(symbol), %temp3
1311 sllx %temp3, 32, %temp4
1312 or %temp4, %temp2, %temp5
1313 or %temp5, %lo(symbol), %reg */
1318 /* It is possible that one of the registers we got for operands[2]
1319 might coincide with that of operands[0] (which is why we made
1320 it TImode). Pick the other one to use as our scratch. */
1321 if (rtx_equal_p (temp, op0))
1323 gcc_assert (ti_temp);
1324 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1327 temp2 = temp; /* op0 is _not_ allowed, see above. */
1334 temp1 = gen_reg_rtx (DImode);
1335 temp2 = gen_reg_rtx (DImode);
1336 temp3 = gen_reg_rtx (DImode);
1337 temp4 = gen_reg_rtx (DImode);
1338 temp5 = gen_reg_rtx (DImode);
1341 emit_insn (gen_embmedany_textuhi (temp1, op1));
1342 emit_insn (gen_embmedany_texthi (temp2, op1));
1343 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1344 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1345 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1346 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1347 gen_rtx_PLUS (DImode, temp4, temp2)));
1348 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1357 #if HOST_BITS_PER_WIDE_INT == 32
1359 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1364 /* These avoid problems when cross compiling. If we do not
1365 go through all this hair then the optimizer will see
1366 invalid REG_EQUAL notes or in some cases none at all. */
1367 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1368 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1369 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1370 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1372 /* The optimizer is not to assume anything about exactly
1373 which bits are set for a HIGH, they are unspecified.
1374 Unfortunately this leads to many missed optimizations
1375 during CSE. We mask out the non-HIGH bits, and matches
1376 a plain movdi, to alleviate this problem. */
1378 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1380 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1384 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1386 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1390 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1392 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1396 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1398 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1401 /* Worker routines for 64-bit constant formation on arch64.
1402 One of the key things to be doing in these emissions is
1403 to create as many temp REGs as possible. This makes it
1404 possible for half-built constants to be used later when
1405 such values are similar to something required later on.
1406 Without doing this, the optimizer cannot see such
1409 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1410 unsigned HOST_WIDE_INT, int);
1413 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1414 unsigned HOST_WIDE_INT low_bits, int is_neg)
1416 unsigned HOST_WIDE_INT high_bits;
1419 high_bits = (~low_bits) & 0xffffffff;
1421 high_bits = low_bits;
1423 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1426 emit_insn (gen_rtx_SET (VOIDmode, op0,
1427 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1431 /* If we are XOR'ing with -1, then we should emit a one's complement
1432 instead. This way the combiner will notice logical operations
1433 such as ANDN later on and substitute. */
1434 if ((low_bits & 0x3ff) == 0x3ff)
1436 emit_insn (gen_rtx_SET (VOIDmode, op0,
1437 gen_rtx_NOT (DImode, temp)));
1441 emit_insn (gen_rtx_SET (VOIDmode, op0,
1442 gen_safe_XOR64 (temp,
1443 (-(HOST_WIDE_INT)0x400
1444 | (low_bits & 0x3ff)))));
1449 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1450 unsigned HOST_WIDE_INT, int);
1453 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1454 unsigned HOST_WIDE_INT high_bits,
1455 unsigned HOST_WIDE_INT low_immediate,
1460 if ((high_bits & 0xfffffc00) != 0)
1462 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1463 if ((high_bits & ~0xfffffc00) != 0)
1464 emit_insn (gen_rtx_SET (VOIDmode, op0,
1465 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1471 emit_insn (gen_safe_SET64 (temp, high_bits));
1475 /* Now shift it up into place. */
1476 emit_insn (gen_rtx_SET (VOIDmode, op0,
1477 gen_rtx_ASHIFT (DImode, temp2,
1478 GEN_INT (shift_count))));
1480 /* If there is a low immediate part piece, finish up by
1481 putting that in as well. */
1482 if (low_immediate != 0)
1483 emit_insn (gen_rtx_SET (VOIDmode, op0,
1484 gen_safe_OR64 (op0, low_immediate)));
1487 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1488 unsigned HOST_WIDE_INT);
1490 /* Full 64-bit constant decomposition. Even though this is the
1491 'worst' case, we still optimize a few things away. */
1493 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1494 unsigned HOST_WIDE_INT high_bits,
1495 unsigned HOST_WIDE_INT low_bits)
1499 if (reload_in_progress || reload_completed)
1502 sub_temp = gen_reg_rtx (DImode);
1504 if ((high_bits & 0xfffffc00) != 0)
1506 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1507 if ((high_bits & ~0xfffffc00) != 0)
1508 emit_insn (gen_rtx_SET (VOIDmode,
1510 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1516 emit_insn (gen_safe_SET64 (temp, high_bits));
1520 if (!reload_in_progress && !reload_completed)
1522 rtx temp2 = gen_reg_rtx (DImode);
1523 rtx temp3 = gen_reg_rtx (DImode);
1524 rtx temp4 = gen_reg_rtx (DImode);
1526 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1527 gen_rtx_ASHIFT (DImode, sub_temp,
1530 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1531 if ((low_bits & ~0xfffffc00) != 0)
1533 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1534 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1535 emit_insn (gen_rtx_SET (VOIDmode, op0,
1536 gen_rtx_PLUS (DImode, temp4, temp3)));
1540 emit_insn (gen_rtx_SET (VOIDmode, op0,
1541 gen_rtx_PLUS (DImode, temp4, temp2)));
1546 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1547 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1548 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1551 /* We are in the middle of reload, so this is really
1552 painful. However we do still make an attempt to
1553 avoid emitting truly stupid code. */
1554 if (low1 != const0_rtx)
1556 emit_insn (gen_rtx_SET (VOIDmode, op0,
1557 gen_rtx_ASHIFT (DImode, sub_temp,
1558 GEN_INT (to_shift))));
1559 emit_insn (gen_rtx_SET (VOIDmode, op0,
1560 gen_rtx_IOR (DImode, op0, low1)));
1568 if (low2 != const0_rtx)
1570 emit_insn (gen_rtx_SET (VOIDmode, op0,
1571 gen_rtx_ASHIFT (DImode, sub_temp,
1572 GEN_INT (to_shift))));
1573 emit_insn (gen_rtx_SET (VOIDmode, op0,
1574 gen_rtx_IOR (DImode, op0, low2)));
1582 emit_insn (gen_rtx_SET (VOIDmode, op0,
1583 gen_rtx_ASHIFT (DImode, sub_temp,
1584 GEN_INT (to_shift))));
1585 if (low3 != const0_rtx)
1586 emit_insn (gen_rtx_SET (VOIDmode, op0,
1587 gen_rtx_IOR (DImode, op0, low3)));
1592 /* Analyze a 64-bit constant for certain properties. */
1593 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1594 unsigned HOST_WIDE_INT,
1595 int *, int *, int *);
1598 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1599 unsigned HOST_WIDE_INT low_bits,
1600 int *hbsp, int *lbsp, int *abbasp)
1602 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1605 lowest_bit_set = highest_bit_set = -1;
1609 if ((lowest_bit_set == -1)
1610 && ((low_bits >> i) & 1))
1612 if ((highest_bit_set == -1)
1613 && ((high_bits >> (32 - i - 1)) & 1))
1614 highest_bit_set = (64 - i - 1);
1617 && ((highest_bit_set == -1)
1618 || (lowest_bit_set == -1)));
1624 if ((lowest_bit_set == -1)
1625 && ((high_bits >> i) & 1))
1626 lowest_bit_set = i + 32;
1627 if ((highest_bit_set == -1)
1628 && ((low_bits >> (32 - i - 1)) & 1))
1629 highest_bit_set = 32 - i - 1;
1632 && ((highest_bit_set == -1)
1633 || (lowest_bit_set == -1)));
1635 /* If there are no bits set this should have gone out
1636 as one instruction! */
1637 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
1638 all_bits_between_are_set = 1;
1639 for (i = lowest_bit_set; i <= highest_bit_set; i++)
1643 if ((low_bits & (1 << i)) != 0)
1648 if ((high_bits & (1 << (i - 32))) != 0)
1651 all_bits_between_are_set = 0;
1654 *hbsp = highest_bit_set;
1655 *lbsp = lowest_bit_set;
1656 *abbasp = all_bits_between_are_set;
1659 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
1662 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
1663 unsigned HOST_WIDE_INT low_bits)
1665 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
1668 || high_bits == 0xffffffff)
1671 analyze_64bit_constant (high_bits, low_bits,
1672 &highest_bit_set, &lowest_bit_set,
1673 &all_bits_between_are_set);
1675 if ((highest_bit_set == 63
1676 || lowest_bit_set == 0)
1677 && all_bits_between_are_set != 0)
1680 if ((highest_bit_set - lowest_bit_set) < 21)
1686 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
1687 unsigned HOST_WIDE_INT,
1690 static unsigned HOST_WIDE_INT
1691 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
1692 unsigned HOST_WIDE_INT low_bits,
1693 int lowest_bit_set, int shift)
1695 HOST_WIDE_INT hi, lo;
1697 if (lowest_bit_set < 32)
1699 lo = (low_bits >> lowest_bit_set) << shift;
1700 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
1705 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
1707 gcc_assert (! (hi & lo));
1711 /* Here we are sure to be arch64 and this is an integer constant
1712 being loaded into a register. Emit the most efficient
1713 insn sequence possible. Detection of all the 1-insn cases
1714 has been done already. */
1716 sparc_emit_set_const64 (rtx op0, rtx op1)
1718 unsigned HOST_WIDE_INT high_bits, low_bits;
1719 int lowest_bit_set, highest_bit_set;
1720 int all_bits_between_are_set;
1723 /* Sanity check that we know what we are working with. */
1724 gcc_assert (TARGET_ARCH64
1725 && (GET_CODE (op0) == SUBREG
1726 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
1728 if (reload_in_progress || reload_completed)
1731 if (GET_CODE (op1) != CONST_INT)
1733 sparc_emit_set_symbolic_const64 (op0, op1, temp);
1738 temp = gen_reg_rtx (DImode);
1740 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
1741 low_bits = (INTVAL (op1) & 0xffffffff);
1743 /* low_bits bits 0 --> 31
1744 high_bits bits 32 --> 63 */
1746 analyze_64bit_constant (high_bits, low_bits,
1747 &highest_bit_set, &lowest_bit_set,
1748 &all_bits_between_are_set);
1750 /* First try for a 2-insn sequence. */
1752 /* These situations are preferred because the optimizer can
1753 * do more things with them:
1755 * sllx %reg, shift, %reg
1757 * srlx %reg, shift, %reg
1758 * 3) mov some_small_const, %reg
1759 * sllx %reg, shift, %reg
1761 if (((highest_bit_set == 63
1762 || lowest_bit_set == 0)
1763 && all_bits_between_are_set != 0)
1764 || ((highest_bit_set - lowest_bit_set) < 12))
1766 HOST_WIDE_INT the_const = -1;
1767 int shift = lowest_bit_set;
1769 if ((highest_bit_set != 63
1770 && lowest_bit_set != 0)
1771 || all_bits_between_are_set == 0)
1774 create_simple_focus_bits (high_bits, low_bits,
1777 else if (lowest_bit_set == 0)
1778 shift = -(63 - highest_bit_set);
1780 gcc_assert (SPARC_SIMM13_P (the_const));
1781 gcc_assert (shift != 0);
1783 emit_insn (gen_safe_SET64 (temp, the_const));
1785 emit_insn (gen_rtx_SET (VOIDmode,
1787 gen_rtx_ASHIFT (DImode,
1791 emit_insn (gen_rtx_SET (VOIDmode,
1793 gen_rtx_LSHIFTRT (DImode,
1795 GEN_INT (-shift))));
1799 /* Now a range of 22 or less bits set somewhere.
1800 * 1) sethi %hi(focus_bits), %reg
1801 * sllx %reg, shift, %reg
1802 * 2) sethi %hi(focus_bits), %reg
1803 * srlx %reg, shift, %reg
1805 if ((highest_bit_set - lowest_bit_set) < 21)
1807 unsigned HOST_WIDE_INT focus_bits =
1808 create_simple_focus_bits (high_bits, low_bits,
1809 lowest_bit_set, 10);
1811 gcc_assert (SPARC_SETHI_P (focus_bits));
1812 gcc_assert (lowest_bit_set != 10);
1814 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
1816 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
1817 if (lowest_bit_set < 10)
1818 emit_insn (gen_rtx_SET (VOIDmode,
1820 gen_rtx_LSHIFTRT (DImode, temp,
1821 GEN_INT (10 - lowest_bit_set))));
1822 else if (lowest_bit_set > 10)
1823 emit_insn (gen_rtx_SET (VOIDmode,
1825 gen_rtx_ASHIFT (DImode, temp,
1826 GEN_INT (lowest_bit_set - 10))));
1830 /* 1) sethi %hi(low_bits), %reg
1831 * or %reg, %lo(low_bits), %reg
1832 * 2) sethi %hi(~low_bits), %reg
1833 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
1836 || high_bits == 0xffffffff)
1838 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
1839 (high_bits == 0xffffffff));
1843 /* Now, try 3-insn sequences. */
1845 /* 1) sethi %hi(high_bits), %reg
1846 * or %reg, %lo(high_bits), %reg
1847 * sllx %reg, 32, %reg
1851 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
1855 /* We may be able to do something quick
1856 when the constant is negated, so try that. */
1857 if (const64_is_2insns ((~high_bits) & 0xffffffff,
1858 (~low_bits) & 0xfffffc00))
1860 /* NOTE: The trailing bits get XOR'd so we need the
1861 non-negated bits, not the negated ones. */
1862 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
1864 if ((((~high_bits) & 0xffffffff) == 0
1865 && ((~low_bits) & 0x80000000) == 0)
1866 || (((~high_bits) & 0xffffffff) == 0xffffffff
1867 && ((~low_bits) & 0x80000000) != 0))
1869 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
1871 if ((SPARC_SETHI_P (fast_int)
1872 && (~high_bits & 0xffffffff) == 0)
1873 || SPARC_SIMM13_P (fast_int))
1874 emit_insn (gen_safe_SET64 (temp, fast_int));
1876 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
1881 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
1882 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
1883 sparc_emit_set_const64 (temp, negated_const);
1886 /* If we are XOR'ing with -1, then we should emit a one's complement
1887 instead. This way the combiner will notice logical operations
1888 such as ANDN later on and substitute. */
1889 if (trailing_bits == 0x3ff)
1891 emit_insn (gen_rtx_SET (VOIDmode, op0,
1892 gen_rtx_NOT (DImode, temp)));
1896 emit_insn (gen_rtx_SET (VOIDmode,
1898 gen_safe_XOR64 (temp,
1899 (-0x400 | trailing_bits))));
1904 /* 1) sethi %hi(xxx), %reg
1905 * or %reg, %lo(xxx), %reg
1906 * sllx %reg, yyy, %reg
1908 * ??? This is just a generalized version of the low_bits==0
1909 * thing above, FIXME...
1911 if ((highest_bit_set - lowest_bit_set) < 32)
1913 unsigned HOST_WIDE_INT focus_bits =
1914 create_simple_focus_bits (high_bits, low_bits,
1917 /* We can't get here in this state. */
1918 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
1920 /* So what we know is that the set bits straddle the
1921 middle of the 64-bit word. */
1922 sparc_emit_set_const64_quick2 (op0, temp,
1928 /* 1) sethi %hi(high_bits), %reg
1929 * or %reg, %lo(high_bits), %reg
1930 * sllx %reg, 32, %reg
1931 * or %reg, low_bits, %reg
1933 if (SPARC_SIMM13_P(low_bits)
1934 && ((int)low_bits > 0))
1936 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
1940 /* The easiest way when all else fails, is full decomposition. */
1941 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
1943 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
1945 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
1946 return the mode to be used for the comparison. For floating-point,
1947 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
1948 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
1949 processing is needed. */
1952 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
1954 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
1980 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
1981 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
1983 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
1984 return CCX_NOOVmode;
1990 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
1997 /* Emit the compare insn and return the CC reg for a CODE comparison
1998 with operands X and Y. */
2001 gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y)
2003 enum machine_mode mode;
2006 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2009 mode = SELECT_CC_MODE (code, x, y);
2011 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2012 fcc regs (cse can't tell they're really call clobbered regs and will
2013 remove a duplicate comparison even if there is an intervening function
2014 call - it will then try to reload the cc reg via an int reg which is why
2015 we need the movcc patterns). It is possible to provide the movcc
2016 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2017 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2018 to tell cse that CCFPE mode registers (even pseudos) are call
2021 /* ??? This is an experiment. Rather than making changes to cse which may
2022 or may not be easy/clean, we do our own cse. This is possible because
2023 we will generate hard registers. Cse knows they're call clobbered (it
2024 doesn't know the same thing about pseudos). If we guess wrong, no big
2025 deal, but if we win, great! */
2027 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2028 #if 1 /* experiment */
2031 /* We cycle through the registers to ensure they're all exercised. */
2032 static int next_fcc_reg = 0;
2033 /* Previous x,y for each fcc reg. */
2034 static rtx prev_args[4][2];
2036 /* Scan prev_args for x,y. */
2037 for (reg = 0; reg < 4; reg++)
2038 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2043 prev_args[reg][0] = x;
2044 prev_args[reg][1] = y;
2045 next_fcc_reg = (next_fcc_reg + 1) & 3;
2047 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2050 cc_reg = gen_reg_rtx (mode);
2051 #endif /* ! experiment */
2052 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2053 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2055 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2057 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2058 will only result in an unrecognizable insn so no point in asserting. */
2059 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2065 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2068 gen_compare_reg (rtx cmp)
2070 return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1));
2073 /* This function is used for v9 only.
2074 DEST is the target of the Scc insn.
2075 CODE is the code for an Scc's comparison.
2076 X and Y are the values we compare.
2078 This function is needed to turn
2081 (gt (reg:CCX 100 %icc)
2085 (gt:DI (reg:CCX 100 %icc)
2088 IE: The instruction recognizer needs to see the mode of the comparison to
2089 find the right instruction. We could use "gt:DI" right in the
2090 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2093 gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y)
2096 && (GET_MODE (x) == DImode
2097 || GET_MODE (dest) == DImode))
2100 /* Try to use the movrCC insns. */
2102 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
2104 && v9_regcmp_p (compare_code))
2109 /* Special case for op0 != 0. This can be done with one instruction if
2112 if (compare_code == NE
2113 && GET_MODE (dest) == DImode
2114 && rtx_equal_p (op0, dest))
2116 emit_insn (gen_rtx_SET (VOIDmode, dest,
2117 gen_rtx_IF_THEN_ELSE (DImode,
2118 gen_rtx_fmt_ee (compare_code, DImode,
2125 if (reg_overlap_mentioned_p (dest, op0))
2127 /* Handle the case where dest == x.
2128 We "early clobber" the result. */
2129 op0 = gen_reg_rtx (GET_MODE (x));
2130 emit_move_insn (op0, x);
2133 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2134 if (GET_MODE (op0) != DImode)
2136 temp = gen_reg_rtx (DImode);
2137 convert_move (temp, op0, 0);
2141 emit_insn (gen_rtx_SET (VOIDmode, dest,
2142 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2143 gen_rtx_fmt_ee (compare_code, DImode,
2151 x = gen_compare_reg_1 (compare_code, x, y);
2154 gcc_assert (GET_MODE (x) != CC_NOOVmode
2155 && GET_MODE (x) != CCX_NOOVmode);
2157 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2158 emit_insn (gen_rtx_SET (VOIDmode, dest,
2159 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2160 gen_rtx_fmt_ee (compare_code,
2161 GET_MODE (x), x, y),
2162 const1_rtx, dest)));
2168 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2169 without jumps using the addx/subx instructions. */
2172 emit_scc_insn (rtx operands[])
2179 /* The quad-word fp compare library routines all return nonzero to indicate
2180 true, which is different from the equivalent libgcc routines, so we must
2181 handle them specially here. */
2182 if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD)
2184 operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3],
2185 GET_CODE (operands[1]));
2186 operands[2] = XEXP (operands[1], 0);
2187 operands[3] = XEXP (operands[1], 1);
2190 code = GET_CODE (operands[1]);
2194 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2195 more applications). The exception to this is "reg != 0" which can
2196 be done in one instruction on v9 (so we do it). */
2199 if (GET_MODE (x) == SImode)
2201 rtx pat = gen_seqsi_special (operands[0], x, y);
2205 else if (GET_MODE (x) == DImode)
2207 rtx pat = gen_seqdi_special (operands[0], x, y);
2215 if (GET_MODE (x) == SImode)
2217 rtx pat = gen_snesi_special (operands[0], x, y);
2221 else if (GET_MODE (x) == DImode)
2223 rtx pat = gen_snedi_special (operands[0], x, y);
2229 /* For the rest, on v9 we can use conditional moves. */
2233 if (gen_v9_scc (operands[0], code, x, y))
2237 /* We can do LTU and GEU using the addx/subx instructions too. And
2238 for GTU/LEU, if both operands are registers swap them and fall
2239 back to the easy case. */
2240 if (code == GTU || code == LEU)
2242 if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
2243 && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG))
2248 code = swap_condition (code);
2252 if (code == LTU || code == GEU)
2254 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2255 gen_rtx_fmt_ee (code, SImode,
2256 gen_compare_reg_1 (code, x, y),
2261 /* Nope, do branches. */
2265 /* Emit a conditional jump insn for the v9 architecture using comparison code
2266 CODE and jump target LABEL.
2267 This function exists to take advantage of the v9 brxx insns. */
2270 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2272 emit_jump_insn (gen_rtx_SET (VOIDmode,
2274 gen_rtx_IF_THEN_ELSE (VOIDmode,
2275 gen_rtx_fmt_ee (code, GET_MODE (op0),
2277 gen_rtx_LABEL_REF (VOIDmode, label),
2282 emit_conditional_branch_insn (rtx operands[])
2284 /* The quad-word fp compare library routines all return nonzero to indicate
2285 true, which is different from the equivalent libgcc routines, so we must
2286 handle them specially here. */
2287 if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD)
2289 operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2],
2290 GET_CODE (operands[0]));
2291 operands[1] = XEXP (operands[0], 0);
2292 operands[2] = XEXP (operands[0], 1);
2295 if (TARGET_ARCH64 && operands[2] == const0_rtx
2296 && GET_CODE (operands[1]) == REG
2297 && GET_MODE (operands[1]) == DImode)
2299 emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]);
2303 operands[1] = gen_compare_reg (operands[0]);
2304 operands[2] = const0_rtx;
2305 operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode,
2306 operands[1], operands[2]);
2307 emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2],
2312 /* Generate a DFmode part of a hard TFmode register.
2313 REG is the TFmode hard register, LOW is 1 for the
2314 low 64bit of the register and 0 otherwise.
2317 gen_df_reg (rtx reg, int low)
2319 int regno = REGNO (reg);
2321 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2322 regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2;
2323 return gen_rtx_REG (DFmode, regno);
2326 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2327 Unlike normal calls, TFmode operands are passed by reference. It is
2328 assumed that no more than 3 operands are required. */
2331 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2333 rtx ret_slot = NULL, arg[3], func_sym;
2336 /* We only expect to be called for conversions, unary, and binary ops. */
2337 gcc_assert (nargs == 2 || nargs == 3);
2339 for (i = 0; i < nargs; ++i)
2341 rtx this_arg = operands[i];
2344 /* TFmode arguments and return values are passed by reference. */
2345 if (GET_MODE (this_arg) == TFmode)
2347 int force_stack_temp;
2349 force_stack_temp = 0;
2350 if (TARGET_BUGGY_QP_LIB && i == 0)
2351 force_stack_temp = 1;
2353 if (GET_CODE (this_arg) == MEM
2354 && ! force_stack_temp)
2355 this_arg = XEXP (this_arg, 0);
2356 else if (CONSTANT_P (this_arg)
2357 && ! force_stack_temp)
2359 this_slot = force_const_mem (TFmode, this_arg);
2360 this_arg = XEXP (this_slot, 0);
2364 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2366 /* Operand 0 is the return value. We'll copy it out later. */
2368 emit_move_insn (this_slot, this_arg);
2370 ret_slot = this_slot;
2372 this_arg = XEXP (this_slot, 0);
2379 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2381 if (GET_MODE (operands[0]) == TFmode)
2384 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2385 arg[0], GET_MODE (arg[0]),
2386 arg[1], GET_MODE (arg[1]));
2388 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2389 arg[0], GET_MODE (arg[0]),
2390 arg[1], GET_MODE (arg[1]),
2391 arg[2], GET_MODE (arg[2]));
2394 emit_move_insn (operands[0], ret_slot);
2400 gcc_assert (nargs == 2);
2402 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2403 GET_MODE (operands[0]), 1,
2404 arg[1], GET_MODE (arg[1]));
2406 if (ret != operands[0])
2407 emit_move_insn (operands[0], ret);
2411 /* Expand soft-float TFmode calls to sparc abi routines. */
2414 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2436 emit_soft_tfmode_libcall (func, 3, operands);
2440 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2444 gcc_assert (code == SQRT);
2447 emit_soft_tfmode_libcall (func, 2, operands);
2451 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2458 switch (GET_MODE (operands[1]))
2471 case FLOAT_TRUNCATE:
2472 switch (GET_MODE (operands[0]))
2486 switch (GET_MODE (operands[1]))
2491 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2501 case UNSIGNED_FLOAT:
2502 switch (GET_MODE (operands[1]))
2507 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2518 switch (GET_MODE (operands[0]))
2532 switch (GET_MODE (operands[0]))
2549 emit_soft_tfmode_libcall (func, 2, operands);
2552 /* Expand a hard-float tfmode operation. All arguments must be in
2556 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2560 if (GET_RTX_CLASS (code) == RTX_UNARY)
2562 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2563 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2567 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2568 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2569 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2570 operands[1], operands[2]);
2573 if (register_operand (operands[0], VOIDmode))
2576 dest = gen_reg_rtx (GET_MODE (operands[0]));
2578 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2580 if (dest != operands[0])
2581 emit_move_insn (operands[0], dest);
2585 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2587 if (TARGET_HARD_QUAD)
2588 emit_hard_tfmode_operation (code, operands);
2590 emit_soft_tfmode_binop (code, operands);
2594 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2596 if (TARGET_HARD_QUAD)
2597 emit_hard_tfmode_operation (code, operands);
2599 emit_soft_tfmode_unop (code, operands);
2603 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2605 if (TARGET_HARD_QUAD)
2606 emit_hard_tfmode_operation (code, operands);
2608 emit_soft_tfmode_cvt (code, operands);
2611 /* Return nonzero if a branch/jump/call instruction will be emitting
2612 nop into its delay slot. */
2615 empty_delay_slot (rtx insn)
2619 /* If no previous instruction (should not happen), return true. */
2620 if (PREV_INSN (insn) == NULL)
2623 seq = NEXT_INSN (PREV_INSN (insn));
2624 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
2630 /* Return nonzero if TRIAL can go into the call delay slot. */
2633 tls_call_delay (rtx trial)
2638 call __tls_get_addr, %tgd_call (foo)
2639 add %l7, %o0, %o0, %tgd_add (foo)
2640 while Sun as/ld does not. */
2641 if (TARGET_GNU_TLS || !TARGET_TLS)
2644 pat = PATTERN (trial);
2646 /* We must reject tgd_add{32|64}, i.e.
2647 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
2648 and tldm_add{32|64}, i.e.
2649 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
2651 if (GET_CODE (pat) == SET
2652 && GET_CODE (SET_SRC (pat)) == PLUS)
2654 rtx unspec = XEXP (SET_SRC (pat), 1);
2656 if (GET_CODE (unspec) == UNSPEC
2657 && (XINT (unspec, 1) == UNSPEC_TLSGD
2658 || XINT (unspec, 1) == UNSPEC_TLSLDM))
2665 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
2666 instruction. RETURN_P is true if the v9 variant 'return' is to be
2667 considered in the test too.
2669 TRIAL must be a SET whose destination is a REG appropriate for the
2670 'restore' instruction or, if RETURN_P is true, for the 'return'
2674 eligible_for_restore_insn (rtx trial, bool return_p)
2676 rtx pat = PATTERN (trial);
2677 rtx src = SET_SRC (pat);
2679 /* The 'restore src,%g0,dest' pattern for word mode and below. */
2680 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
2681 && arith_operand (src, GET_MODE (src)))
2684 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
2686 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
2689 /* The 'restore src,%g0,dest' pattern for double-word mode. */
2690 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
2691 && arith_double_operand (src, GET_MODE (src)))
2692 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
2694 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
2695 else if (! TARGET_FPU && register_operand (src, SFmode))
2698 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
2699 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
2702 /* If we have the 'return' instruction, anything that does not use
2703 local or output registers and can go into a delay slot wins. */
2704 else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1)
2705 && (get_attr_in_uncond_branch_delay (trial)
2706 == IN_UNCOND_BRANCH_DELAY_TRUE))
2709 /* The 'restore src1,src2,dest' pattern for SImode. */
2710 else if (GET_CODE (src) == PLUS
2711 && register_operand (XEXP (src, 0), SImode)
2712 && arith_operand (XEXP (src, 1), SImode))
2715 /* The 'restore src1,src2,dest' pattern for DImode. */
2716 else if (GET_CODE (src) == PLUS
2717 && register_operand (XEXP (src, 0), DImode)
2718 && arith_double_operand (XEXP (src, 1), DImode))
2721 /* The 'restore src1,%lo(src2),dest' pattern. */
2722 else if (GET_CODE (src) == LO_SUM
2723 && ! TARGET_CM_MEDMID
2724 && ((register_operand (XEXP (src, 0), SImode)
2725 && immediate_operand (XEXP (src, 1), SImode))
2727 && register_operand (XEXP (src, 0), DImode)
2728 && immediate_operand (XEXP (src, 1), DImode))))
2731 /* The 'restore src,src,dest' pattern. */
2732 else if (GET_CODE (src) == ASHIFT
2733 && (register_operand (XEXP (src, 0), SImode)
2734 || register_operand (XEXP (src, 0), DImode))
2735 && XEXP (src, 1) == const1_rtx)
2741 /* Return nonzero if TRIAL can go into the function return's
2745 eligible_for_return_delay (rtx trial)
2749 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
2752 if (get_attr_length (trial) != 1)
2755 /* If there are any call-saved registers, we should scan TRIAL if it
2756 does not reference them. For now just make it easy. */
2760 /* If the function uses __builtin_eh_return, the eh_return machinery
2761 occupies the delay slot. */
2762 if (crtl->calls_eh_return)
2765 /* In the case of a true leaf function, anything can go into the slot. */
2766 if (sparc_leaf_function_p)
2767 return get_attr_in_uncond_branch_delay (trial)
2768 == IN_UNCOND_BRANCH_DELAY_TRUE;
2770 pat = PATTERN (trial);
2772 /* Otherwise, only operations which can be done in tandem with
2773 a `restore' or `return' insn can go into the delay slot. */
2774 if (GET_CODE (SET_DEST (pat)) != REG
2775 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24))
2778 /* If this instruction sets up floating point register and we have a return
2779 instruction, it can probably go in. But restore will not work
2781 if (REGNO (SET_DEST (pat)) >= 32)
2783 && ! epilogue_renumber (&pat, 1)
2784 && (get_attr_in_uncond_branch_delay (trial)
2785 == IN_UNCOND_BRANCH_DELAY_TRUE));
2787 return eligible_for_restore_insn (trial, true);
2790 /* Return nonzero if TRIAL can go into the sibling call's
2794 eligible_for_sibcall_delay (rtx trial)
2798 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
2801 if (get_attr_length (trial) != 1)
2804 pat = PATTERN (trial);
2806 if (sparc_leaf_function_p)
2808 /* If the tail call is done using the call instruction,
2809 we have to restore %o7 in the delay slot. */
2810 if (LEAF_SIBCALL_SLOT_RESERVED_P)
2813 /* %g1 is used to build the function address */
2814 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
2820 /* Otherwise, only operations which can be done in tandem with
2821 a `restore' insn can go into the delay slot. */
2822 if (GET_CODE (SET_DEST (pat)) != REG
2823 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
2824 || REGNO (SET_DEST (pat)) >= 32)
2827 /* If it mentions %o7, it can't go in, because sibcall will clobber it
2829 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
2832 return eligible_for_restore_insn (trial, false);
2836 short_branch (int uid1, int uid2)
2838 int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2);
2840 /* Leave a few words of "slop". */
2841 if (delta >= -1023 && delta <= 1022)
2847 /* Return nonzero if REG is not used after INSN.
2848 We assume REG is a reload reg, and therefore does
2849 not live past labels or calls or jumps. */
2851 reg_unused_after (rtx reg, rtx insn)
2853 enum rtx_code code, prev_code = UNKNOWN;
2855 while ((insn = NEXT_INSN (insn)))
2857 if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)])
2860 code = GET_CODE (insn);
2861 if (GET_CODE (insn) == CODE_LABEL)
2866 rtx set = single_set (insn);
2867 int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set));
2870 if (set && reg_overlap_mentioned_p (reg, SET_DEST (set)))
2872 if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
2880 /* Determine if it's legal to put X into the constant pool. This
2881 is not possible if X contains the address of a symbol that is
2882 not constant (TLS) or not known at final link time (PIC). */
2885 sparc_cannot_force_const_mem (rtx x)
2887 switch (GET_CODE (x))
2892 /* Accept all non-symbolic constants. */
2896 /* Labels are OK iff we are non-PIC. */
2897 return flag_pic != 0;
2900 /* 'Naked' TLS symbol references are never OK,
2901 non-TLS symbols are OK iff we are non-PIC. */
2902 if (SYMBOL_REF_TLS_MODEL (x))
2905 return flag_pic != 0;
2908 return sparc_cannot_force_const_mem (XEXP (x, 0));
2911 return sparc_cannot_force_const_mem (XEXP (x, 0))
2912 || sparc_cannot_force_const_mem (XEXP (x, 1));
2921 static GTY(()) bool pic_helper_needed = false;
2922 static GTY(()) rtx pic_helper_symbol;
2923 static GTY(()) rtx global_offset_table;
2925 /* Ensure that we are not using patterns that are not OK with PIC. */
2933 gcc_assert (GET_CODE (recog_data.operand[i]) != SYMBOL_REF
2934 && (GET_CODE (recog_data.operand[i]) != CONST
2935 || (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS
2936 && (XEXP (XEXP (recog_data.operand[i], 0), 0)
2937 == global_offset_table)
2938 && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1))
2946 /* Return true if X is an address which needs a temporary register when
2947 reloaded while generating PIC code. */
2950 pic_address_needs_scratch (rtx x)
2952 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
2953 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
2954 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
2955 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
2956 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
2962 /* Determine if a given RTX is a valid constant. We already know this
2963 satisfies CONSTANT_P. */
2966 legitimate_constant_p (rtx x)
2968 switch (GET_CODE (x))
2972 if (sparc_tls_referenced_p (x))
2977 if (GET_MODE (x) == VOIDmode)
2980 /* Floating point constants are generally not ok.
2981 The only exception is 0.0 in VIS. */
2983 && SCALAR_FLOAT_MODE_P (GET_MODE (x))
2984 && const_zero_operand (x, GET_MODE (x)))
2990 /* Vector constants are generally not ok.
2991 The only exception is 0 in VIS. */
2993 && const_zero_operand (x, GET_MODE (x)))
3005 /* Determine if a given RTX is a valid constant address. */
3008 constant_address_p (rtx x)
3010 switch (GET_CODE (x))
3018 if (flag_pic && pic_address_needs_scratch (x))
3020 return legitimate_constant_p (x);
3023 return !flag_pic && legitimate_constant_p (x);
3030 /* Nonzero if the constant value X is a legitimate general operand
3031 when generating PIC code. It is given that flag_pic is on and
3032 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3035 legitimate_pic_operand_p (rtx x)
3037 if (pic_address_needs_scratch (x))
3039 if (sparc_tls_referenced_p (x))
3044 /* Return nonzero if ADDR is a valid memory address.
3045 STRICT specifies whether strict register checking applies. */
3048 sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict)
3050 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
3052 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3054 else if (GET_CODE (addr) == PLUS)
3056 rs1 = XEXP (addr, 0);
3057 rs2 = XEXP (addr, 1);
3059 /* Canonicalize. REG comes first, if there are no regs,
3060 LO_SUM comes first. */
3062 && GET_CODE (rs1) != SUBREG
3064 || GET_CODE (rs2) == SUBREG
3065 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3067 rs1 = XEXP (addr, 1);
3068 rs2 = XEXP (addr, 0);
3072 && rs1 == pic_offset_table_rtx
3074 && GET_CODE (rs2) != SUBREG
3075 && GET_CODE (rs2) != LO_SUM
3076 && GET_CODE (rs2) != MEM
3077 && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2))
3078 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3079 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3081 || GET_CODE (rs1) == SUBREG)
3082 && RTX_OK_FOR_OFFSET_P (rs2)))
3087 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3088 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3090 /* We prohibit REG + REG for TFmode when there are no quad move insns
3091 and we consequently need to split. We do this because REG+REG
3092 is not an offsettable address. If we get the situation in reload
3093 where source and destination of a movtf pattern are both MEMs with
3094 REG+REG address, then only one of them gets converted to an
3095 offsettable address. */
3097 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3100 /* We prohibit REG + REG on ARCH32 if not optimizing for
3101 DFmode/DImode because then mem_min_alignment is likely to be zero
3102 after reload and the forced split would lack a matching splitter
3104 if (TARGET_ARCH32 && !optimize
3105 && (mode == DFmode || mode == DImode))
3108 else if (USE_AS_OFFSETABLE_LO10
3109 && GET_CODE (rs1) == LO_SUM
3111 && ! TARGET_CM_MEDMID
3112 && RTX_OK_FOR_OLO10_P (rs2))
3115 imm1 = XEXP (rs1, 1);
3116 rs1 = XEXP (rs1, 0);
3117 if (!CONSTANT_P (imm1)
3118 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3122 else if (GET_CODE (addr) == LO_SUM)
3124 rs1 = XEXP (addr, 0);
3125 imm1 = XEXP (addr, 1);
3127 if (!CONSTANT_P (imm1)
3128 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3131 /* We can't allow TFmode in 32-bit mode, because an offset greater
3132 than the alignment (8) may cause the LO_SUM to overflow. */
3133 if (mode == TFmode && TARGET_ARCH32)
3136 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3141 if (GET_CODE (rs1) == SUBREG)
3142 rs1 = SUBREG_REG (rs1);
3148 if (GET_CODE (rs2) == SUBREG)
3149 rs2 = SUBREG_REG (rs2);
3156 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3157 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3162 if ((REGNO (rs1) >= 32
3163 && REGNO (rs1) != FRAME_POINTER_REGNUM
3164 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3166 && (REGNO (rs2) >= 32
3167 && REGNO (rs2) != FRAME_POINTER_REGNUM
3168 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3174 /* Construct the SYMBOL_REF for the tls_get_offset function. */
3176 static GTY(()) rtx sparc_tls_symbol;
3179 sparc_tls_get_addr (void)
3181 if (!sparc_tls_symbol)
3182 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3184 return sparc_tls_symbol;
3188 sparc_tls_got (void)
3193 crtl->uses_pic_offset_table = 1;
3194 return pic_offset_table_rtx;
3197 if (!global_offset_table)
3198 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3199 temp = gen_reg_rtx (Pmode);
3200 emit_move_insn (temp, global_offset_table);
3204 /* Return true if X contains a thread-local symbol. */
3207 sparc_tls_referenced_p (rtx x)
3209 if (!TARGET_HAVE_TLS)
3212 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS)
3213 x = XEXP (XEXP (x, 0), 0);
3215 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x))
3218 /* That's all we handle in legitimize_tls_address for now. */
3222 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3223 this (thread-local) address. */
3226 legitimize_tls_address (rtx addr)
3228 rtx temp1, temp2, temp3, ret, o0, got, insn;
3230 gcc_assert (can_create_pseudo_p ());
3232 if (GET_CODE (addr) == SYMBOL_REF)
3233 switch (SYMBOL_REF_TLS_MODEL (addr))
3235 case TLS_MODEL_GLOBAL_DYNAMIC:
3237 temp1 = gen_reg_rtx (SImode);
3238 temp2 = gen_reg_rtx (SImode);
3239 ret = gen_reg_rtx (Pmode);
3240 o0 = gen_rtx_REG (Pmode, 8);
3241 got = sparc_tls_got ();
3242 emit_insn (gen_tgd_hi22 (temp1, addr));
3243 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3246 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3247 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3252 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3253 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3256 CALL_INSN_FUNCTION_USAGE (insn)
3257 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3258 CALL_INSN_FUNCTION_USAGE (insn));
3259 insn = get_insns ();
3261 emit_libcall_block (insn, ret, o0, addr);
3264 case TLS_MODEL_LOCAL_DYNAMIC:
3266 temp1 = gen_reg_rtx (SImode);
3267 temp2 = gen_reg_rtx (SImode);
3268 temp3 = gen_reg_rtx (Pmode);
3269 ret = gen_reg_rtx (Pmode);
3270 o0 = gen_rtx_REG (Pmode, 8);
3271 got = sparc_tls_got ();
3272 emit_insn (gen_tldm_hi22 (temp1));
3273 emit_insn (gen_tldm_lo10 (temp2, temp1));
3276 emit_insn (gen_tldm_add32 (o0, got, temp2));
3277 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3282 emit_insn (gen_tldm_add64 (o0, got, temp2));
3283 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3286 CALL_INSN_FUNCTION_USAGE (insn)
3287 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3288 CALL_INSN_FUNCTION_USAGE (insn));
3289 insn = get_insns ();
3291 emit_libcall_block (insn, temp3, o0,
3292 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3293 UNSPEC_TLSLD_BASE));
3294 temp1 = gen_reg_rtx (SImode);
3295 temp2 = gen_reg_rtx (SImode);
3296 emit_insn (gen_tldo_hix22 (temp1, addr));
3297 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3299 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3301 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3304 case TLS_MODEL_INITIAL_EXEC:
3305 temp1 = gen_reg_rtx (SImode);
3306 temp2 = gen_reg_rtx (SImode);
3307 temp3 = gen_reg_rtx (Pmode);
3308 got = sparc_tls_got ();
3309 emit_insn (gen_tie_hi22 (temp1, addr));
3310 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3312 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3314 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3317 ret = gen_reg_rtx (Pmode);
3319 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3322 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3326 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3329 case TLS_MODEL_LOCAL_EXEC:
3330 temp1 = gen_reg_rtx (Pmode);
3331 temp2 = gen_reg_rtx (Pmode);
3334 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3335 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3339 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3340 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3342 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3349 else if (GET_CODE (addr) == CONST)
3353 gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS);
3355 base = legitimize_tls_address (XEXP (XEXP (addr, 0), 0));
3356 offset = XEXP (XEXP (addr, 0), 1);
3358 base = force_operand (base, NULL_RTX);
3359 if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset)))
3360 offset = force_reg (Pmode, offset);
3361 ret = gen_rtx_PLUS (Pmode, base, offset);
3365 gcc_unreachable (); /* for now ... */
3370 /* Legitimize PIC addresses. If the address is already position-independent,
3371 we return ORIG. Newly generated position-independent addresses go into a
3372 reg. This is REG if nonzero, otherwise we allocate register(s) as
3376 legitimize_pic_address (rtx orig, rtx reg)
3378 bool gotdata_op = false;
3380 if (GET_CODE (orig) == SYMBOL_REF
3381 /* See the comment in sparc_expand_move. */
3382 || (TARGET_VXWORKS_RTP && GET_CODE (orig) == LABEL_REF))
3384 rtx pic_ref, address;
3389 gcc_assert (! reload_in_progress && ! reload_completed);
3390 reg = gen_reg_rtx (Pmode);
3395 /* If not during reload, allocate another temp reg here for loading
3396 in the address, so that these instructions can be optimized
3398 rtx temp_reg = ((reload_in_progress || reload_completed)
3399 ? reg : gen_reg_rtx (Pmode));
3401 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3402 won't get confused into thinking that these two instructions
3403 are loading in the true address of the symbol. If in the
3404 future a PIC rtx exists, that should be used instead. */
3407 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3408 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3412 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3413 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3421 crtl->uses_pic_offset_table = 1;
3425 insn = emit_insn (gen_movdi_pic_gotdata_op (reg, pic_offset_table_rtx,
3428 insn = emit_insn (gen_movsi_pic_gotdata_op (reg, pic_offset_table_rtx,
3433 pic_ref = gen_const_mem (Pmode,
3434 gen_rtx_PLUS (Pmode,
3435 pic_offset_table_rtx, address));
3436 insn = emit_move_insn (reg, pic_ref);
3438 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3440 set_unique_reg_note (insn, REG_EQUAL, orig);
3443 else if (GET_CODE (orig) == CONST)
3447 if (GET_CODE (XEXP (orig, 0)) == PLUS
3448 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3453 gcc_assert (! reload_in_progress && ! reload_completed);
3454 reg = gen_reg_rtx (Pmode);
3457 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3458 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg);
3459 offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
3460 base == reg ? NULL_RTX : reg);
3462 if (GET_CODE (offset) == CONST_INT)
3464 if (SMALL_INT (offset))
3465 return plus_constant (base, INTVAL (offset));
3466 else if (! reload_in_progress && ! reload_completed)
3467 offset = force_reg (Pmode, offset);
3469 /* If we reach here, then something is seriously wrong. */
3472 return gen_rtx_PLUS (Pmode, base, offset);
3474 else if (GET_CODE (orig) == LABEL_REF)
3475 /* ??? Why do we do this? */
3476 /* Now movsi_pic_label_ref uses it, but we ought to be checking that
3477 the register is live instead, in case it is eliminated. */
3478 crtl->uses_pic_offset_table = 1;
3483 /* Try machine-dependent ways of modifying an illegitimate address X
3484 to be legitimate. If we find one, return the new, valid address.
3486 OLDX is the address as it was before break_out_memory_refs was called.
3487 In some cases it is useful to look at this to decide what needs to be done.
3489 MODE is the mode of the operand pointed to by X.
3491 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
3494 sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
3495 enum machine_mode mode)
3499 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3500 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3501 force_operand (XEXP (x, 0), NULL_RTX));
3502 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3503 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3504 force_operand (XEXP (x, 1), NULL_RTX));
3505 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3506 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3508 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3509 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3510 force_operand (XEXP (x, 1), NULL_RTX));
3512 if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE))
3515 if (sparc_tls_referenced_p (x))
3516 x = legitimize_tls_address (x);
3518 x = legitimize_pic_address (x, NULL_RTX);
3519 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3520 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3521 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3522 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3523 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3524 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3525 else if (GET_CODE (x) == SYMBOL_REF
3526 || GET_CODE (x) == CONST
3527 || GET_CODE (x) == LABEL_REF)
3528 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3533 /* Return true if ADDR (a legitimate address expression)
3534 has an effect that depends on the machine mode it is used for.
3540 is not equivalent to
3542 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
3544 because [%l7+a+1] is interpreted as the address of (a+1). */
3548 sparc_mode_dependent_address_p (const_rtx addr)
3550 if (flag_pic && GET_CODE (addr) == PLUS)
3552 rtx op0 = XEXP (addr, 0);
3553 rtx op1 = XEXP (addr, 1);
3554 if (op0 == pic_offset_table_rtx
3555 && SYMBOLIC_CONST (op1))
3562 #ifdef HAVE_GAS_HIDDEN
3563 # define USE_HIDDEN_LINKONCE 1
3565 # define USE_HIDDEN_LINKONCE 0
3569 get_pc_thunk_name (char name[32], unsigned int regno)
3571 const char *pic_name = reg_names[regno];
3573 /* Skip the leading '%' as that cannot be used in a
3577 if (USE_HIDDEN_LINKONCE)
3578 sprintf (name, "__sparc_get_pc_thunk.%s", pic_name);
3580 ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno);
3583 /* Emit code to load the PIC register. */
3586 load_pic_register (void)
3588 int orig_flag_pic = flag_pic;
3590 if (TARGET_VXWORKS_RTP)
3592 emit_insn (gen_vxworks_load_got ());
3593 emit_use (pic_offset_table_rtx);
3597 /* If we haven't initialized the special PIC symbols, do so now. */
3598 if (!pic_helper_needed)
3602 pic_helper_needed = true;
3604 get_pc_thunk_name (name, REGNO (pic_offset_table_rtx));
3605 pic_helper_symbol = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
3607 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3612 emit_insn (gen_load_pcrel_symdi (pic_offset_table_rtx, global_offset_table,
3613 pic_helper_symbol));
3615 emit_insn (gen_load_pcrel_symsi (pic_offset_table_rtx, global_offset_table,
3616 pic_helper_symbol));
3617 flag_pic = orig_flag_pic;
3619 /* Need to emit this whether or not we obey regdecls,
3620 since setjmp/longjmp can cause life info to screw up.
3621 ??? In the case where we don't obey regdecls, this is not sufficient
3622 since we may not fall out the bottom. */
3623 emit_use (pic_offset_table_rtx);
3626 /* Emit a call instruction with the pattern given by PAT. ADDR is the
3627 address of the call target. */
3630 sparc_emit_call_insn (rtx pat, rtx addr)
3634 insn = emit_call_insn (pat);
3636 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
3637 if (TARGET_VXWORKS_RTP
3639 && GET_CODE (addr) == SYMBOL_REF
3640 && (SYMBOL_REF_DECL (addr)
3641 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
3642 : !SYMBOL_REF_LOCAL_P (addr)))
3644 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
3645 crtl->uses_pic_offset_table = 1;
3649 /* Return 1 if RTX is a MEM which is known to be aligned to at
3650 least a DESIRED byte boundary. */
3653 mem_min_alignment (rtx mem, int desired)
3655 rtx addr, base, offset;
3657 /* If it's not a MEM we can't accept it. */
3658 if (GET_CODE (mem) != MEM)
3662 if (!TARGET_UNALIGNED_DOUBLES
3663 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
3666 /* ??? The rest of the function predates MEM_ALIGN so
3667 there is probably a bit of redundancy. */
3668 addr = XEXP (mem, 0);
3669 base = offset = NULL_RTX;
3670 if (GET_CODE (addr) == PLUS)
3672 if (GET_CODE (XEXP (addr, 0)) == REG)
3674 base = XEXP (addr, 0);
3676 /* What we are saying here is that if the base
3677 REG is aligned properly, the compiler will make
3678 sure any REG based index upon it will be so
3680 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3681 offset = XEXP (addr, 1);
3683 offset = const0_rtx;
3686 else if (GET_CODE (addr) == REG)
3689 offset = const0_rtx;
3692 if (base != NULL_RTX)
3694 int regno = REGNO (base);
3696 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
3698 /* Check if the compiler has recorded some information
3699 about the alignment of the base REG. If reload has
3700 completed, we already matched with proper alignments.
3701 If not running global_alloc, reload might give us
3702 unaligned pointer to local stack though. */
3704 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
3705 || (optimize && reload_completed))
3706 && (INTVAL (offset) & (desired - 1)) == 0)
3711 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
3715 else if (! TARGET_UNALIGNED_DOUBLES
3716 || CONSTANT_P (addr)
3717 || GET_CODE (addr) == LO_SUM)
3719 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
3720 is true, in which case we can only assume that an access is aligned if
3721 it is to a constant address, or the address involves a LO_SUM. */
3725 /* An obviously unaligned address. */
3730 /* Vectors to keep interesting information about registers where it can easily
3731 be got. We used to use the actual mode value as the bit number, but there
3732 are more than 32 modes now. Instead we use two tables: one indexed by
3733 hard register number, and one indexed by mode. */
3735 /* The purpose of sparc_mode_class is to shrink the range of modes so that
3736 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
3737 mapped into one sparc_mode_class mode. */
3739 enum sparc_mode_class {
3740 S_MODE, D_MODE, T_MODE, O_MODE,
3741 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
3745 /* Modes for single-word and smaller quantities. */
3746 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
3748 /* Modes for double-word and smaller quantities. */
3749 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
3751 /* Modes for quad-word and smaller quantities. */
3752 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
3754 /* Modes for 8-word and smaller quantities. */
3755 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
3757 /* Modes for single-float quantities. We must allow any single word or
3758 smaller quantity. This is because the fix/float conversion instructions
3759 take integer inputs/outputs from the float registers. */
3760 #define SF_MODES (S_MODES)
3762 /* Modes for double-float and smaller quantities. */
3763 #define DF_MODES (D_MODES)
3765 /* Modes for quad-float and smaller quantities. */
3766 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
3768 /* Modes for quad-float pairs and smaller quantities. */
3769 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
3771 /* Modes for double-float only quantities. */
3772 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
3774 /* Modes for quad-float and double-float only quantities. */
3775 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
3777 /* Modes for quad-float pairs and double-float only quantities. */
3778 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
3780 /* Modes for condition codes. */
3781 #define CC_MODES (1 << (int) CC_MODE)
3782 #define CCFP_MODES (1 << (int) CCFP_MODE)
3784 /* Value is 1 if register/mode pair is acceptable on sparc.
3785 The funny mixture of D and T modes is because integer operations
3786 do not specially operate on tetra quantities, so non-quad-aligned
3787 registers can hold quadword quantities (except %o4 and %i4 because
3788 they cross fixed registers). */
3790 /* This points to either the 32 bit or the 64 bit version. */
3791 const int *hard_regno_mode_classes;
3793 static const int hard_32bit_mode_classes[] = {
3794 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
3795 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
3796 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
3797 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
3799 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3800 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3801 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3802 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
3804 /* FP regs f32 to f63. Only the even numbered registers actually exist,
3805 and none can hold SFmode/SImode values. */
3806 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3807 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3808 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3809 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3812 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
3818 static const int hard_64bit_mode_classes[] = {
3819 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3820 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3821 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3822 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3824 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3825 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3826 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3827 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
3829 /* FP regs f32 to f63. Only the even numbered registers actually exist,
3830 and none can hold SFmode/SImode values. */
3831 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3832 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3833 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3834 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3837 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
3843 int sparc_mode_class [NUM_MACHINE_MODES];
3845 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
3848 sparc_init_modes (void)
3852 for (i = 0; i < NUM_MACHINE_MODES; i++)
3854 switch (GET_MODE_CLASS (i))
3857 case MODE_PARTIAL_INT:
3858 case MODE_COMPLEX_INT:
3859 if (GET_MODE_SIZE (i) <= 4)
3860 sparc_mode_class[i] = 1 << (int) S_MODE;
3861 else if (GET_MODE_SIZE (i) == 8)
3862 sparc_mode_class[i] = 1 << (int) D_MODE;
3863 else if (GET_MODE_SIZE (i) == 16)
3864 sparc_mode_class[i] = 1 << (int) T_MODE;
3865 else if (GET_MODE_SIZE (i) == 32)
3866 sparc_mode_class[i] = 1 << (int) O_MODE;
3868 sparc_mode_class[i] = 0;
3870 case MODE_VECTOR_INT:
3871 if (GET_MODE_SIZE (i) <= 4)
3872 sparc_mode_class[i] = 1 << (int)SF_MODE;
3873 else if (GET_MODE_SIZE (i) == 8)
3874 sparc_mode_class[i] = 1 << (int)DF_MODE;
3877 case MODE_COMPLEX_FLOAT:
3878 if (GET_MODE_SIZE (i) <= 4)
3879 sparc_mode_class[i] = 1 << (int) SF_MODE;
3880 else if (GET_MODE_SIZE (i) == 8)
3881 sparc_mode_class[i] = 1 << (int) DF_MODE;
3882 else if (GET_MODE_SIZE (i) == 16)
3883 sparc_mode_class[i] = 1 << (int) TF_MODE;
3884 else if (GET_MODE_SIZE (i) == 32)
3885 sparc_mode_class[i] = 1 << (int) OF_MODE;
3887 sparc_mode_class[i] = 0;
3890 if (i == (int) CCFPmode || i == (int) CCFPEmode)
3891 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
3893 sparc_mode_class[i] = 1 << (int) CC_MODE;
3896 sparc_mode_class[i] = 0;
3902 hard_regno_mode_classes = hard_64bit_mode_classes;
3904 hard_regno_mode_classes = hard_32bit_mode_classes;
3906 /* Initialize the array used by REGNO_REG_CLASS. */
3907 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3909 if (i < 16 && TARGET_V8PLUS)
3910 sparc_regno_reg_class[i] = I64_REGS;
3911 else if (i < 32 || i == FRAME_POINTER_REGNUM)
3912 sparc_regno_reg_class[i] = GENERAL_REGS;
3914 sparc_regno_reg_class[i] = FP_REGS;
3916 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
3918 sparc_regno_reg_class[i] = FPCC_REGS;
3920 sparc_regno_reg_class[i] = NO_REGS;
3924 /* Compute the frame size required by the function. This function is called
3925 during the reload pass and also by sparc_expand_prologue. */
3928 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p)
3930 int outgoing_args_size = (crtl->outgoing_args_size
3931 + REG_PARM_STACK_SPACE (current_function_decl));
3932 int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */
3937 for (i = 0; i < 8; i++)
3938 if (df_regs_ever_live_p (i) && ! call_used_regs[i])
3943 for (i = 0; i < 8; i += 2)
3944 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
3945 || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1]))
3949 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
3950 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
3951 || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1]))
3954 /* Set up values for use in prologue and epilogue. */
3955 num_gfregs = n_regs;
3960 && crtl->outgoing_args_size == 0)
3961 actual_fsize = apparent_fsize = 0;
3964 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
3965 apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8;
3966 apparent_fsize += n_regs * 4;
3967 actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8);
3970 /* Make sure nothing can clobber our register windows.
3971 If a SAVE must be done, or there is a stack-local variable,
3972 the register window area must be allocated. */
3973 if (! leaf_function_p || size > 0)
3974 actual_fsize += FIRST_PARM_OFFSET (current_function_decl);
3976 return SPARC_STACK_ALIGN (actual_fsize);
3979 /* Output any necessary .register pseudo-ops. */
3982 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
3984 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
3990 /* Check if %g[2367] were used without
3991 .register being printed for them already. */
3992 for (i = 2; i < 8; i++)
3994 if (df_regs_ever_live_p (i)
3995 && ! sparc_hard_reg_printed [i])
3997 sparc_hard_reg_printed [i] = 1;
3998 /* %g7 is used as TLS base register, use #ignore
3999 for it instead of #scratch. */
4000 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
4001 i == 7 ? "ignore" : "scratch");
4008 /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET
4009 as needed. LOW should be double-word aligned for 32-bit registers.
4010 Return the new OFFSET. */
4013 #define SORR_RESTORE 1
4016 save_or_restore_regs (int low, int high, rtx base, int offset, int action)
4021 if (TARGET_ARCH64 && high <= 32)
4023 for (i = low; i < high; i++)
4025 if (df_regs_ever_live_p (i) && ! call_used_regs[i])
4027 mem = gen_rtx_MEM (DImode, plus_constant (base, offset));
4028 set_mem_alias_set (mem, sparc_sr_alias_set);
4029 if (action == SORR_SAVE)
4031 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4032 RTX_FRAME_RELATED_P (insn) = 1;
4034 else /* action == SORR_RESTORE */
4035 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4042 for (i = low; i < high; i += 2)
4044 bool reg0 = df_regs_ever_live_p (i) && ! call_used_regs[i];
4045 bool reg1 = df_regs_ever_live_p (i+1) && ! call_used_regs[i+1];
4046 enum machine_mode mode;
4051 mode = i < 32 ? DImode : DFmode;
4056 mode = i < 32 ? SImode : SFmode;
4061 mode = i < 32 ? SImode : SFmode;
4068 mem = gen_rtx_MEM (mode, plus_constant (base, offset));
4069 set_mem_alias_set (mem, sparc_sr_alias_set);
4070 if (action == SORR_SAVE)
4072 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4073 RTX_FRAME_RELATED_P (insn) = 1;
4075 else /* action == SORR_RESTORE */
4076 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4078 /* Always preserve double-word alignment. */
4079 offset = (offset + 7) & -8;
4086 /* Emit code to save call-saved registers. */
4089 emit_save_or_restore_regs (int action)
4091 HOST_WIDE_INT offset;
4094 offset = frame_base_offset - apparent_fsize;
4096 if (offset < -4096 || offset + num_gfregs * 4 > 4095)
4098 /* ??? This might be optimized a little as %g1 might already have a
4099 value close enough that a single add insn will do. */
4100 /* ??? Although, all of this is probably only a temporary fix
4101 because if %g1 can hold a function result, then
4102 sparc_expand_epilogue will lose (the result will be
4104 base = gen_rtx_REG (Pmode, 1);
4105 emit_move_insn (base, GEN_INT (offset));
4106 emit_insn (gen_rtx_SET (VOIDmode,
4108 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4112 base = frame_base_reg;
4114 offset = save_or_restore_regs (0, 8, base, offset, action);
4115 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, action);
4118 /* Generate a save_register_window insn. */
4121 gen_save_register_window (rtx increment)
4124 return gen_save_register_windowdi (increment);
4126 return gen_save_register_windowsi (increment);
4129 /* Generate an increment for the stack pointer. */
4132 gen_stack_pointer_inc (rtx increment)
4134 return gen_rtx_SET (VOIDmode,
4136 gen_rtx_PLUS (Pmode,
4141 /* Generate a decrement for the stack pointer. */
4144 gen_stack_pointer_dec (rtx decrement)
4146 return gen_rtx_SET (VOIDmode,
4148 gen_rtx_MINUS (Pmode,
4153 /* Expand the function prologue. The prologue is responsible for reserving
4154 storage for the frame, saving the call-saved registers and loading the
4155 PIC register if needed. */
4158 sparc_expand_prologue (void)
4163 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4164 on the final value of the flag means deferring the prologue/epilogue
4165 expansion until just before the second scheduling pass, which is too
4166 late to emit multiple epilogues or return insns.
4168 Of course we are making the assumption that the value of the flag
4169 will not change between now and its final value. Of the three parts
4170 of the formula, only the last one can reasonably vary. Let's take a
4171 closer look, after assuming that the first two ones are set to true
4172 (otherwise the last value is effectively silenced).
4174 If only_leaf_regs_used returns false, the global predicate will also
4175 be false so the actual frame size calculated below will be positive.
4176 As a consequence, the save_register_window insn will be emitted in
4177 the instruction stream; now this insn explicitly references %fp
4178 which is not a leaf register so only_leaf_regs_used will always
4179 return false subsequently.
4181 If only_leaf_regs_used returns true, we hope that the subsequent
4182 optimization passes won't cause non-leaf registers to pop up. For
4183 example, the regrename pass has special provisions to not rename to
4184 non-leaf registers in a leaf function. */
4185 sparc_leaf_function_p
4186 = optimize > 0 && leaf_function_p () && only_leaf_regs_used ();
4188 /* Need to use actual_fsize, since we are also allocating
4189 space for our callee (and our own register save area). */
4191 = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4193 /* Advertise that the data calculated just above are now valid. */
4194 sparc_prologue_data_valid_p = true;
4196 if (sparc_leaf_function_p)
4198 frame_base_reg = stack_pointer_rtx;
4199 frame_base_offset = actual_fsize + SPARC_STACK_BIAS;
4203 frame_base_reg = hard_frame_pointer_rtx;
4204 frame_base_offset = SPARC_STACK_BIAS;
4207 if (actual_fsize == 0)
4209 else if (sparc_leaf_function_p)
4211 if (actual_fsize <= 4096)
4212 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-actual_fsize)));
4213 else if (actual_fsize <= 8192)
4215 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
4216 /* %sp is still the CFA register. */
4217 RTX_FRAME_RELATED_P (insn) = 1;
4219 = emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize)));
4223 rtx reg = gen_rtx_REG (Pmode, 1);
4224 emit_move_insn (reg, GEN_INT (-actual_fsize));
4225 insn = emit_insn (gen_stack_pointer_inc (reg));
4226 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4227 gen_stack_pointer_inc (GEN_INT (-actual_fsize)));
4230 RTX_FRAME_RELATED_P (insn) = 1;
4234 if (actual_fsize <= 4096)
4235 insn = emit_insn (gen_save_register_window (GEN_INT (-actual_fsize)));
4236 else if (actual_fsize <= 8192)
4238 insn = emit_insn (gen_save_register_window (GEN_INT (-4096)));
4239 /* %sp is not the CFA register anymore. */
4240 emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize)));
4244 rtx reg = gen_rtx_REG (Pmode, 1);
4245 emit_move_insn (reg, GEN_INT (-actual_fsize));
4246 insn = emit_insn (gen_save_register_window (reg));
4249 RTX_FRAME_RELATED_P (insn) = 1;
4250 for (i=0; i < XVECLEN (PATTERN (insn), 0); i++)
4251 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, i)) = 1;
4255 emit_save_or_restore_regs (SORR_SAVE);
4257 /* Load the PIC register if needed. */
4258 if (flag_pic && crtl->uses_pic_offset_table)
4259 load_pic_register ();
4262 /* This function generates the assembly code for function entry, which boils
4263 down to emitting the necessary .register directives. */
4266 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4268 /* Check that the assumption we made in sparc_expand_prologue is valid. */
4269 gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs);
4271 sparc_output_scratch_registers (file);
4274 /* Expand the function epilogue, either normal or part of a sibcall.
4275 We emit all the instructions except the return or the call. */
4278 sparc_expand_epilogue (void)
4281 emit_save_or_restore_regs (SORR_RESTORE);
4283 if (actual_fsize == 0)
4285 else if (sparc_leaf_function_p)
4287 if (actual_fsize <= 4096)
4288 emit_insn (gen_stack_pointer_dec (GEN_INT (- actual_fsize)));
4289 else if (actual_fsize <= 8192)
4291 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
4292 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - actual_fsize)));
4296 rtx reg = gen_rtx_REG (Pmode, 1);
4297 emit_move_insn (reg, GEN_INT (-actual_fsize));
4298 emit_insn (gen_stack_pointer_dec (reg));
4303 /* Return true if it is appropriate to emit `return' instructions in the
4304 body of a function. */
4307 sparc_can_use_return_insn_p (void)
4309 return sparc_prologue_data_valid_p
4310 && (actual_fsize == 0 || !sparc_leaf_function_p);
4313 /* This function generates the assembly code for function exit. */
4316 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4318 /* If code does not drop into the epilogue, we have to still output
4319 a dummy nop for the sake of sane backtraces. Otherwise, if the
4320 last two instructions of a function were "call foo; dslot;" this
4321 can make the return PC of foo (i.e. address of call instruction
4322 plus 8) point to the first instruction in the next function. */
4324 rtx insn, last_real_insn;
4326 insn = get_last_insn ();
4328 last_real_insn = prev_real_insn (insn);
4330 && GET_CODE (last_real_insn) == INSN
4331 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
4332 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
4334 if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
4335 fputs("\tnop\n", file);
4337 sparc_output_deferred_case_vectors ();
4340 /* Output a 'restore' instruction. */
4343 output_restore (rtx pat)
4349 fputs ("\t restore\n", asm_out_file);
4353 gcc_assert (GET_CODE (pat) == SET);
4355 operands[0] = SET_DEST (pat);
4356 pat = SET_SRC (pat);
4358 switch (GET_CODE (pat))
4361 operands[1] = XEXP (pat, 0);
4362 operands[2] = XEXP (pat, 1);
4363 output_asm_insn (" restore %r1, %2, %Y0", operands);
4366 operands[1] = XEXP (pat, 0);
4367 operands[2] = XEXP (pat, 1);
4368 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
4371 operands[1] = XEXP (pat, 0);
4372 gcc_assert (XEXP (pat, 1) == const1_rtx);
4373 output_asm_insn (" restore %r1, %r1, %Y0", operands);
4377 output_asm_insn (" restore %%g0, %1, %Y0", operands);
4382 /* Output a return. */
4385 output_return (rtx insn)
4387 if (sparc_leaf_function_p)
4389 /* This is a leaf function so we don't have to bother restoring the
4390 register window, which frees us from dealing with the convoluted
4391 semantics of restore/return. We simply output the jump to the
4392 return address and the insn in the delay slot (if any). */
4394 gcc_assert (! crtl->calls_eh_return);
4396 return "jmp\t%%o7+%)%#";
4400 /* This is a regular function so we have to restore the register window.
4401 We may have a pending insn for the delay slot, which will be either
4402 combined with the 'restore' instruction or put in the delay slot of
4403 the 'return' instruction. */
4405 if (crtl->calls_eh_return)
4407 /* If the function uses __builtin_eh_return, the eh_return
4408 machinery occupies the delay slot. */
4409 gcc_assert (! final_sequence);
4411 if (! flag_delayed_branch)
4412 fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file);
4415 fputs ("\treturn\t%i7+8\n", asm_out_file);
4417 fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file);
4419 if (flag_delayed_branch)
4420 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
4422 fputs ("\t nop\n", asm_out_file);
4424 else if (final_sequence)
4428 delay = NEXT_INSN (insn);
4431 pat = PATTERN (delay);
4433 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
4435 epilogue_renumber (&pat, 0);
4436 return "return\t%%i7+%)%#";
4440 output_asm_insn ("jmp\t%%i7+%)", NULL);
4441 output_restore (pat);
4442 PATTERN (delay) = gen_blockage ();
4443 INSN_CODE (delay) = -1;
4448 /* The delay slot is empty. */
4450 return "return\t%%i7+%)\n\t nop";
4451 else if (flag_delayed_branch)
4452 return "jmp\t%%i7+%)\n\t restore";
4454 return "restore\n\tjmp\t%%o7+%)\n\t nop";
4461 /* Output a sibling call. */
4464 output_sibcall (rtx insn, rtx call_operand)
4468 gcc_assert (flag_delayed_branch);
4470 operands[0] = call_operand;
4472 if (sparc_leaf_function_p)
4474 /* This is a leaf function so we don't have to bother restoring the
4475 register window. We simply output the jump to the function and
4476 the insn in the delay slot (if any). */
4478 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
4481 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
4484 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
4485 it into branch if possible. */
4486 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
4491 /* This is a regular function so we have to restore the register window.
4492 We may have a pending insn for the delay slot, which will be combined
4493 with the 'restore' instruction. */
4495 output_asm_insn ("call\t%a0, 0", operands);
4499 rtx delay = NEXT_INSN (insn);
4502 output_restore (PATTERN (delay));
4504 PATTERN (delay) = gen_blockage ();
4505 INSN_CODE (delay) = -1;
4508 output_restore (NULL_RTX);
4514 /* Functions for handling argument passing.
4516 For 32-bit, the first 6 args are normally in registers and the rest are
4517 pushed. Any arg that starts within the first 6 words is at least
4518 partially passed in a register unless its data type forbids.
4520 For 64-bit, the argument registers are laid out as an array of 16 elements
4521 and arguments are added sequentially. The first 6 int args and up to the
4522 first 16 fp args (depending on size) are passed in regs.
4524 Slot Stack Integral Float Float in structure Double Long Double
4525 ---- ----- -------- ----- ------------------ ------ -----------
4526 15 [SP+248] %f31 %f30,%f31 %d30
4527 14 [SP+240] %f29 %f28,%f29 %d28 %q28
4528 13 [SP+232] %f27 %f26,%f27 %d26
4529 12 [SP+224] %f25 %f24,%f25 %d24 %q24
4530 11 [SP+216] %f23 %f22,%f23 %d22
4531 10 [SP+208] %f21 %f20,%f21 %d20 %q20
4532 9 [SP+200] %f19 %f18,%f19 %d18
4533 8 [SP+192] %f17 %f16,%f17 %d16 %q16
4534 7 [SP+184] %f15 %f14,%f15 %d14
4535 6 [SP+176] %f13 %f12,%f13 %d12 %q12
4536 5 [SP+168] %o5 %f11 %f10,%f11 %d10
4537 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
4538 3 [SP+152] %o3 %f7 %f6,%f7 %d6
4539 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
4540 1 [SP+136] %o1 %f3 %f2,%f3 %d2
4541 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
4543 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
4545 Integral arguments are always passed as 64-bit quantities appropriately
4548 Passing of floating point values is handled as follows.
4549 If a prototype is in scope:
4550 If the value is in a named argument (i.e. not a stdarg function or a
4551 value not part of the `...') then the value is passed in the appropriate
4553 If the value is part of the `...' and is passed in one of the first 6
4554 slots then the value is passed in the appropriate int reg.
4555 If the value is part of the `...' and is not passed in one of the first 6
4556 slots then the value is passed in memory.
4557 If a prototype is not in scope:
4558 If the value is one of the first 6 arguments the value is passed in the
4559 appropriate integer reg and the appropriate fp reg.
4560 If the value is not one of the first 6 arguments the value is passed in
4561 the appropriate fp reg and in memory.
4564 Summary of the calling conventions implemented by GCC on the SPARC:
4567 size argument return value
4569 small integer <4 int. reg. int. reg.
4570 word 4 int. reg. int. reg.
4571 double word 8 int. reg. int. reg.
4573 _Complex small integer <8 int. reg. int. reg.
4574 _Complex word 8 int. reg. int. reg.
4575 _Complex double word 16 memory int. reg.
4577 vector integer <=8 int. reg. FP reg.
4578 vector integer >8 memory memory
4580 float 4 int. reg. FP reg.
4581 double 8 int. reg. FP reg.
4582 long double 16 memory memory
4584 _Complex float 8 memory FP reg.
4585 _Complex double 16 memory FP reg.
4586 _Complex long double 32 memory FP reg.
4588 vector float any memory memory
4590 aggregate any memory memory
4595 size argument return value
4597 small integer <8 int. reg. int. reg.
4598 word 8 int. reg. int. reg.
4599 double word 16 int. reg. int. reg.
4601 _Complex small integer <16 int. reg. int. reg.
4602 _Complex word 16 int. reg. int. reg.
4603 _Complex double word 32 memory int. reg.
4605 vector integer <=16 FP reg. FP reg.
4606 vector integer 16<s<=32 memory FP reg.
4607 vector integer >32 memory memory
4609 float 4 FP reg. FP reg.
4610 double 8 FP reg. FP reg.
4611 long double 16 FP reg. FP reg.
4613 _Complex float 8 FP reg. FP reg.
4614 _Complex double 16 FP reg. FP reg.
4615 _Complex long double 32 memory FP reg.
4617 vector float <=16 FP reg. FP reg.
4618 vector float 16<s<=32 memory FP reg.
4619 vector float >32 memory memory
4621 aggregate <=16 reg. reg.
4622 aggregate 16<s<=32 memory reg.
4623 aggregate >32 memory memory
4627 Note #1: complex floating-point types follow the extended SPARC ABIs as
4628 implemented by the Sun compiler.
4630 Note #2: integral vector types follow the scalar floating-point types
4631 conventions to match what is implemented by the Sun VIS SDK.
4633 Note #3: floating-point vector types follow the aggregate types
4637 /* Maximum number of int regs for args. */
4638 #define SPARC_INT_ARG_MAX 6
4639 /* Maximum number of fp regs for args. */
4640 #define SPARC_FP_ARG_MAX 16
4642 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
4644 /* Handle the INIT_CUMULATIVE_ARGS macro.
4645 Initialize a variable CUM of type CUMULATIVE_ARGS
4646 for a call to a function whose data type is FNTYPE.
4647 For a library call, FNTYPE is 0. */
4650 init_cumulative_args (struct sparc_args *cum, tree fntype,
4651 rtx libname ATTRIBUTE_UNUSED,
4652 tree fndecl ATTRIBUTE_UNUSED)
4655 cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype);
4656 cum->libcall_p = fntype == 0;
4659 /* Handle the TARGET_PROMOTE_PROTOTYPES target hook.
4660 When a prototype says `char' or `short', really pass an `int'. */
4663 sparc_promote_prototypes (const_tree fntype ATTRIBUTE_UNUSED)
4665 return TARGET_ARCH32 ? true : false;
4668 /* Handle promotion of pointer and integer arguments. */
4670 static enum machine_mode
4671 sparc_promote_function_mode (const_tree type ATTRIBUTE_UNUSED,
4672 enum machine_mode mode,
4673 int *punsignedp ATTRIBUTE_UNUSED,
4674 const_tree fntype ATTRIBUTE_UNUSED,
4675 int for_return ATTRIBUTE_UNUSED)
4677 if (POINTER_TYPE_P (type))
4679 *punsignedp = POINTERS_EXTEND_UNSIGNED;
4683 /* For TARGET_ARCH64 we need this, as we don't have instructions
4684 for arithmetic operations which do zero/sign extension at the same time,
4685 so without this we end up with a srl/sra after every assignment to an
4686 user variable, which means very very bad code. */
4688 && GET_MODE_CLASS (mode) == MODE_INT
4689 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4695 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
4698 sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED)
4700 return TARGET_ARCH64 ? true : false;
4703 /* Scan the record type TYPE and return the following predicates:
4704 - INTREGS_P: the record contains at least one field or sub-field
4705 that is eligible for promotion in integer registers.
4706 - FP_REGS_P: the record contains at least one field or sub-field
4707 that is eligible for promotion in floating-point registers.
4708 - PACKED_P: the record contains at least one field that is packed.
4710 Sub-fields are not taken into account for the PACKED_P predicate. */
4713 scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p)
4717 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4719 if (TREE_CODE (field) == FIELD_DECL)
4721 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4722 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
4723 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
4724 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
4730 if (packed_p && DECL_PACKED (field))
4736 /* Compute the slot number to pass an argument in.
4737 Return the slot number or -1 if passing on the stack.
4739 CUM is a variable of type CUMULATIVE_ARGS which gives info about
4740 the preceding args and about the function being called.
4741 MODE is the argument's machine mode.
4742 TYPE is the data type of the argument (as a tree).
4743 This is null for libcalls where that information may
4745 NAMED is nonzero if this argument is a named parameter
4746 (otherwise it is an extra parameter matching an ellipsis).
4747 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
4748 *PREGNO records the register number to use if scalar type.
4749 *PPADDING records the amount of padding needed in words. */
4752 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
4753 tree type, int named, int incoming_p,
4754 int *pregno, int *ppadding)
4756 int regbase = (incoming_p
4757 ? SPARC_INCOMING_INT_ARG_FIRST
4758 : SPARC_OUTGOING_INT_ARG_FIRST);
4759 int slotno = cum->words;
4760 enum mode_class mclass;
4765 if (type && TREE_ADDRESSABLE (type))
4771 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
4774 /* For SPARC64, objects requiring 16-byte alignment get it. */
4776 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
4777 && (slotno & 1) != 0)
4778 slotno++, *ppadding = 1;
4780 mclass = GET_MODE_CLASS (mode);
4781 if (type && TREE_CODE (type) == VECTOR_TYPE)
4783 /* Vector types deserve special treatment because they are
4784 polymorphic wrt their mode, depending upon whether VIS
4785 instructions are enabled. */
4786 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
4788 /* The SPARC port defines no floating-point vector modes. */
4789 gcc_assert (mode == BLKmode);
4793 /* Integral vector types should either have a vector
4794 mode or an integral mode, because we are guaranteed
4795 by pass_by_reference that their size is not greater
4796 than 16 bytes and TImode is 16-byte wide. */
4797 gcc_assert (mode != BLKmode);
4799 /* Vector integers are handled like floats according to
4801 mclass = MODE_FLOAT;
4808 case MODE_COMPLEX_FLOAT:
4809 case MODE_VECTOR_INT:
4810 if (TARGET_ARCH64 && TARGET_FPU && named)
4812 if (slotno >= SPARC_FP_ARG_MAX)
4814 regno = SPARC_FP_ARG_FIRST + slotno * 2;
4815 /* Arguments filling only one single FP register are
4816 right-justified in the outer double FP register. */
4817 if (GET_MODE_SIZE (mode) <= 4)
4824 case MODE_COMPLEX_INT:
4825 if (slotno >= SPARC_INT_ARG_MAX)
4827 regno = regbase + slotno;
4831 if (mode == VOIDmode)
4832 /* MODE is VOIDmode when generating the actual call. */
4835 gcc_assert (mode == BLKmode);
4839 || (TREE_CODE (type) != VECTOR_TYPE
4840 && TREE_CODE (type) != RECORD_TYPE))
4842 if (slotno >= SPARC_INT_ARG_MAX)
4844 regno = regbase + slotno;
4846 else /* TARGET_ARCH64 && type */
4848 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
4850 /* First see what kinds of registers we would need. */
4851 if (TREE_CODE (type) == VECTOR_TYPE)
4854 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
4856 /* The ABI obviously doesn't specify how packed structures
4857 are passed. These are defined to be passed in int regs
4858 if possible, otherwise memory. */
4859 if (packed_p || !named)
4860 fpregs_p = 0, intregs_p = 1;
4862 /* If all arg slots are filled, then must pass on stack. */
4863 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
4866 /* If there are only int args and all int arg slots are filled,
4867 then must pass on stack. */
4868 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
4871 /* Note that even if all int arg slots are filled, fp members may
4872 still be passed in regs if such regs are available.
4873 *PREGNO isn't set because there may be more than one, it's up
4874 to the caller to compute them. */
4887 /* Handle recursive register counting for structure field layout. */
4889 struct function_arg_record_value_parms
4891 rtx ret; /* return expression being built. */
4892 int slotno; /* slot number of the argument. */
4893 int named; /* whether the argument is named. */
4894 int regbase; /* regno of the base register. */
4895 int stack; /* 1 if part of the argument is on the stack. */
4896 int intoffset; /* offset of the first pending integer field. */
4897 unsigned int nregs; /* number of words passed in registers. */
4900 static void function_arg_record_value_3
4901 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
4902 static void function_arg_record_value_2
4903 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
4904 static void function_arg_record_value_1
4905 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
4906 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
4907 static rtx function_arg_union_value (int, enum machine_mode, int, int);
4909 /* A subroutine of function_arg_record_value. Traverse the structure
4910 recursively and determine how many registers will be required. */
4913 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
4914 struct function_arg_record_value_parms *parms,
4919 /* We need to compute how many registers are needed so we can
4920 allocate the PARALLEL but before we can do that we need to know
4921 whether there are any packed fields. The ABI obviously doesn't
4922 specify how structures are passed in this case, so they are
4923 defined to be passed in int regs if possible, otherwise memory,
4924 regardless of whether there are fp values present. */
4927 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4929 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
4936 /* Compute how many registers we need. */
4937 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4939 if (TREE_CODE (field) == FIELD_DECL)
4941 HOST_WIDE_INT bitpos = startbitpos;
4943 if (DECL_SIZE (field) != 0)
4945 if (integer_zerop (DECL_SIZE (field)))
4948 if (host_integerp (bit_position (field), 1))
4949 bitpos += int_bit_position (field);
4952 /* ??? FIXME: else assume zero offset. */
4954 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4955 function_arg_record_value_1 (TREE_TYPE (field),
4959 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
4960 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
4965 if (parms->intoffset != -1)
4967 unsigned int startbit, endbit;
4968 int intslots, this_slotno;
4970 startbit = parms->intoffset & -BITS_PER_WORD;
4971 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
4973 intslots = (endbit - startbit) / BITS_PER_WORD;
4974 this_slotno = parms->slotno + parms->intoffset
4977 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
4979 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
4980 /* We need to pass this field on the stack. */
4984 parms->nregs += intslots;
4985 parms->intoffset = -1;
4988 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
4989 If it wasn't true we wouldn't be here. */
4990 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
4991 && DECL_MODE (field) == BLKmode)
4992 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
4993 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5000 if (parms->intoffset == -1)
5001 parms->intoffset = bitpos;
5007 /* A subroutine of function_arg_record_value. Assign the bits of the
5008 structure between parms->intoffset and bitpos to integer registers. */
5011 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5012 struct function_arg_record_value_parms *parms)
5014 enum machine_mode mode;
5016 unsigned int startbit, endbit;
5017 int this_slotno, intslots, intoffset;
5020 if (parms->intoffset == -1)
5023 intoffset = parms->intoffset;
5024 parms->intoffset = -1;
5026 startbit = intoffset & -BITS_PER_WORD;
5027 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5028 intslots = (endbit - startbit) / BITS_PER_WORD;
5029 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
5031 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
5035 /* If this is the trailing part of a word, only load that much into
5036 the register. Otherwise load the whole register. Note that in
5037 the latter case we may pick up unwanted bits. It's not a problem
5038 at the moment but may wish to revisit. */
5040 if (intoffset % BITS_PER_WORD != 0)
5041 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
5046 intoffset /= BITS_PER_UNIT;
5049 regno = parms->regbase + this_slotno;
5050 reg = gen_rtx_REG (mode, regno);
5051 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5052 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5055 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5060 while (intslots > 0);
5063 /* A subroutine of function_arg_record_value. Traverse the structure
5064 recursively and assign bits to floating point registers. Track which
5065 bits in between need integer registers; invoke function_arg_record_value_3
5066 to make that happen. */
5069 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
5070 struct function_arg_record_value_parms *parms,
5076 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5078 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5085 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5087 if (TREE_CODE (field) == FIELD_DECL)
5089 HOST_WIDE_INT bitpos = startbitpos;
5091 if (DECL_SIZE (field) != 0)
5093 if (integer_zerop (DECL_SIZE (field)))
5096 if (host_integerp (bit_position (field), 1))
5097 bitpos += int_bit_position (field);
5100 /* ??? FIXME: else assume zero offset. */
5102 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5103 function_arg_record_value_2 (TREE_TYPE (field),
5107 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5108 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5113 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
5114 int regno, nregs, pos;
5115 enum machine_mode mode = DECL_MODE (field);
5118 function_arg_record_value_3 (bitpos, parms);
5120 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5123 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5124 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5126 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5128 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5134 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
5135 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
5137 reg = gen_rtx_REG (mode, regno);
5138 pos = bitpos / BITS_PER_UNIT;
5139 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5140 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5144 regno += GET_MODE_SIZE (mode) / 4;
5145 reg = gen_rtx_REG (mode, regno);
5146 pos += GET_MODE_SIZE (mode);
5147 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5148 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5154 if (parms->intoffset == -1)
5155 parms->intoffset = bitpos;
5161 /* Used by function_arg and sparc_function_value_1 to implement the complex
5162 conventions of the 64-bit ABI for passing and returning structures.
5163 Return an expression valid as a return value for the FUNCTION_ARG
5164 and TARGET_FUNCTION_VALUE.
5166 TYPE is the data type of the argument (as a tree).
5167 This is null for libcalls where that information may
5169 MODE is the argument's machine mode.
5170 SLOTNO is the index number of the argument's slot in the parameter array.
5171 NAMED is nonzero if this argument is a named parameter
5172 (otherwise it is an extra parameter matching an ellipsis).
5173 REGBASE is the regno of the base register for the parameter array. */
5176 function_arg_record_value (const_tree type, enum machine_mode mode,
5177 int slotno, int named, int regbase)
5179 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5180 struct function_arg_record_value_parms parms;
5183 parms.ret = NULL_RTX;
5184 parms.slotno = slotno;
5185 parms.named = named;
5186 parms.regbase = regbase;
5189 /* Compute how many registers we need. */
5191 parms.intoffset = 0;
5192 function_arg_record_value_1 (type, 0, &parms, false);
5194 /* Take into account pending integer fields. */
5195 if (parms.intoffset != -1)
5197 unsigned int startbit, endbit;
5198 int intslots, this_slotno;
5200 startbit = parms.intoffset & -BITS_PER_WORD;
5201 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5202 intslots = (endbit - startbit) / BITS_PER_WORD;
5203 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
5205 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5207 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5208 /* We need to pass this field on the stack. */
5212 parms.nregs += intslots;
5214 nregs = parms.nregs;
5216 /* Allocate the vector and handle some annoying special cases. */
5219 /* ??? Empty structure has no value? Duh? */
5222 /* Though there's nothing really to store, return a word register
5223 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
5224 leads to breakage due to the fact that there are zero bytes to
5226 return gen_rtx_REG (mode, regbase);
5230 /* ??? C++ has structures with no fields, and yet a size. Give up
5231 for now and pass everything back in integer registers. */
5232 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5234 if (nregs + slotno > SPARC_INT_ARG_MAX)
5235 nregs = SPARC_INT_ARG_MAX - slotno;
5237 gcc_assert (nregs != 0);
5239 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
5241 /* If at least one field must be passed on the stack, generate
5242 (parallel [(expr_list (nil) ...) ...]) so that all fields will
5243 also be passed on the stack. We can't do much better because the
5244 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
5245 of structures for which the fields passed exclusively in registers
5246 are not at the beginning of the structure. */
5248 XVECEXP (parms.ret, 0, 0)
5249 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5251 /* Fill in the entries. */
5253 parms.intoffset = 0;
5254 function_arg_record_value_2 (type, 0, &parms, false);
5255 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
5257 gcc_assert (parms.nregs == nregs);
5262 /* Used by function_arg and sparc_function_value_1 to implement the conventions
5263 of the 64-bit ABI for passing and returning unions.
5264 Return an expression valid as a return value for the FUNCTION_ARG
5265 and TARGET_FUNCTION_VALUE.
5267 SIZE is the size in bytes of the union.
5268 MODE is the argument's machine mode.
5269 REGNO is the hard register the union will be passed in. */
5272 function_arg_union_value (int size, enum machine_mode mode, int slotno,
5275 int nwords = ROUND_ADVANCE (size), i;
5278 /* See comment in previous function for empty structures. */
5280 return gen_rtx_REG (mode, regno);
5282 if (slotno == SPARC_INT_ARG_MAX - 1)
5285 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
5287 for (i = 0; i < nwords; i++)
5289 /* Unions are passed left-justified. */
5290 XVECEXP (regs, 0, i)
5291 = gen_rtx_EXPR_LIST (VOIDmode,
5292 gen_rtx_REG (word_mode, regno),
5293 GEN_INT (UNITS_PER_WORD * i));
5300 /* Used by function_arg and sparc_function_value_1 to implement the conventions
5301 for passing and returning large (BLKmode) vectors.
5302 Return an expression valid as a return value for the FUNCTION_ARG
5303 and TARGET_FUNCTION_VALUE.
5305 SIZE is the size in bytes of the vector (at least 8 bytes).
5306 REGNO is the FP hard register the vector will be passed in. */
5309 function_arg_vector_value (int size, int regno)
5311 int i, nregs = size / 8;
5314 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
5316 for (i = 0; i < nregs; i++)
5318 XVECEXP (regs, 0, i)
5319 = gen_rtx_EXPR_LIST (VOIDmode,
5320 gen_rtx_REG (DImode, regno + 2*i),
5327 /* Handle the FUNCTION_ARG macro.
5328 Determine where to put an argument to a function.
5329 Value is zero to push the argument on the stack,
5330 or a hard register in which to store the argument.
5332 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5333 the preceding args and about the function being called.
5334 MODE is the argument's machine mode.
5335 TYPE is the data type of the argument (as a tree).
5336 This is null for libcalls where that information may
5338 NAMED is nonzero if this argument is a named parameter
5339 (otherwise it is an extra parameter matching an ellipsis).
5340 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */
5343 function_arg (const struct sparc_args *cum, enum machine_mode mode,
5344 tree type, int named, int incoming_p)
5346 int regbase = (incoming_p
5347 ? SPARC_INCOMING_INT_ARG_FIRST
5348 : SPARC_OUTGOING_INT_ARG_FIRST);
5349 int slotno, regno, padding;
5350 enum mode_class mclass = GET_MODE_CLASS (mode);
5352 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
5357 /* Vector types deserve special treatment because they are polymorphic wrt
5358 their mode, depending upon whether VIS instructions are enabled. */
5359 if (type && TREE_CODE (type) == VECTOR_TYPE)
5361 HOST_WIDE_INT size = int_size_in_bytes (type);
5362 gcc_assert ((TARGET_ARCH32 && size <= 8)
5363 || (TARGET_ARCH64 && size <= 16));
5365 if (mode == BLKmode)
5366 return function_arg_vector_value (size,
5367 SPARC_FP_ARG_FIRST + 2*slotno);
5369 mclass = MODE_FLOAT;
5373 return gen_rtx_REG (mode, regno);
5375 /* Structures up to 16 bytes in size are passed in arg slots on the stack
5376 and are promoted to registers if possible. */
5377 if (type && TREE_CODE (type) == RECORD_TYPE)
5379 HOST_WIDE_INT size = int_size_in_bytes (type);
5380 gcc_assert (size <= 16);
5382 return function_arg_record_value (type, mode, slotno, named, regbase);
5385 /* Unions up to 16 bytes in size are passed in integer registers. */
5386 else if (type && TREE_CODE (type) == UNION_TYPE)
5388 HOST_WIDE_INT size = int_size_in_bytes (type);
5389 gcc_assert (size <= 16);
5391 return function_arg_union_value (size, mode, slotno, regno);
5394 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
5395 but also have the slot allocated for them.
5396 If no prototype is in scope fp values in register slots get passed
5397 in two places, either fp regs and int regs or fp regs and memory. */
5398 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5399 && SPARC_FP_REG_P (regno))
5401 rtx reg = gen_rtx_REG (mode, regno);
5402 if (cum->prototype_p || cum->libcall_p)
5404 /* "* 2" because fp reg numbers are recorded in 4 byte
5407 /* ??? This will cause the value to be passed in the fp reg and
5408 in the stack. When a prototype exists we want to pass the
5409 value in the reg but reserve space on the stack. That's an
5410 optimization, and is deferred [for a bit]. */
5411 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
5412 return gen_rtx_PARALLEL (mode,
5414 gen_rtx_EXPR_LIST (VOIDmode,
5415 NULL_RTX, const0_rtx),
5416 gen_rtx_EXPR_LIST (VOIDmode,
5420 /* ??? It seems that passing back a register even when past
5421 the area declared by REG_PARM_STACK_SPACE will allocate
5422 space appropriately, and will not copy the data onto the
5423 stack, exactly as we desire.
5425 This is due to locate_and_pad_parm being called in
5426 expand_call whenever reg_parm_stack_space > 0, which
5427 while beneficial to our example here, would seem to be
5428 in error from what had been intended. Ho hum... -- r~ */
5436 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
5440 /* On incoming, we don't need to know that the value
5441 is passed in %f0 and %i0, and it confuses other parts
5442 causing needless spillage even on the simplest cases. */
5446 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
5447 + (regno - SPARC_FP_ARG_FIRST) / 2);
5449 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5450 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
5452 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5456 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5457 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5458 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5463 /* All other aggregate types are passed in an integer register in a mode
5464 corresponding to the size of the type. */
5465 else if (type && AGGREGATE_TYPE_P (type))
5467 HOST_WIDE_INT size = int_size_in_bytes (type);
5468 gcc_assert (size <= 16);
5470 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
5473 return gen_rtx_REG (mode, regno);
5476 /* For an arg passed partly in registers and partly in memory,
5477 this is the number of bytes of registers used.
5478 For args passed entirely in registers or entirely in memory, zero.
5480 Any arg that starts in the first 6 regs but won't entirely fit in them
5481 needs partial registers on v8. On v9, structures with integer
5482 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
5483 values that begin in the last fp reg [where "last fp reg" varies with the
5484 mode] will be split between that reg and memory. */
5487 sparc_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5488 tree type, bool named)
5490 int slotno, regno, padding;
5492 /* We pass 0 for incoming_p here, it doesn't matter. */
5493 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5500 if ((slotno + (mode == BLKmode
5501 ? ROUND_ADVANCE (int_size_in_bytes (type))
5502 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
5503 > SPARC_INT_ARG_MAX)
5504 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
5508 /* We are guaranteed by pass_by_reference that the size of the
5509 argument is not greater than 16 bytes, so we only need to return
5510 one word if the argument is partially passed in registers. */
5512 if (type && AGGREGATE_TYPE_P (type))
5514 int size = int_size_in_bytes (type);
5516 if (size > UNITS_PER_WORD
5517 && slotno == SPARC_INT_ARG_MAX - 1)
5518 return UNITS_PER_WORD;
5520 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
5521 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5522 && ! (TARGET_FPU && named)))
5524 /* The complex types are passed as packed types. */
5525 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
5526 && slotno == SPARC_INT_ARG_MAX - 1)
5527 return UNITS_PER_WORD;
5529 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
5531 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
5533 return UNITS_PER_WORD;
5540 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
5541 Specify whether to pass the argument by reference. */
5544 sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5545 enum machine_mode mode, const_tree type,
5546 bool named ATTRIBUTE_UNUSED)
5549 /* Original SPARC 32-bit ABI says that structures and unions,
5550 and quad-precision floats are passed by reference. For Pascal,
5551 also pass arrays by reference. All other base types are passed
5554 Extended ABI (as implemented by the Sun compiler) says that all
5555 complex floats are passed by reference. Pass complex integers
5556 in registers up to 8 bytes. More generally, enforce the 2-word
5557 cap for passing arguments in registers.
5559 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5560 integers are passed like floats of the same size, that is in
5561 registers up to 8 bytes. Pass all vector floats by reference
5562 like structure and unions. */
5563 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
5565 /* Catch CDImode, TFmode, DCmode and TCmode. */
5566 || GET_MODE_SIZE (mode) > 8
5568 && TREE_CODE (type) == VECTOR_TYPE
5569 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5571 /* Original SPARC 64-bit ABI says that structures and unions
5572 smaller than 16 bytes are passed in registers, as well as
5573 all other base types.
5575 Extended ABI (as implemented by the Sun compiler) says that
5576 complex floats are passed in registers up to 16 bytes. Pass
5577 all complex integers in registers up to 16 bytes. More generally,
5578 enforce the 2-word cap for passing arguments in registers.
5580 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5581 integers are passed like floats of the same size, that is in
5582 registers (up to 16 bytes). Pass all vector floats like structure
5585 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
5586 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
5587 /* Catch CTImode and TCmode. */
5588 || GET_MODE_SIZE (mode) > 16);
5591 /* Handle the FUNCTION_ARG_ADVANCE macro.
5592 Update the data in CUM to advance over an argument
5593 of mode MODE and data type TYPE.
5594 TYPE is null for libcalls where that information may not be available. */
5597 function_arg_advance (struct sparc_args *cum, enum machine_mode mode,
5598 tree type, int named)
5600 int slotno, regno, padding;
5602 /* We pass 0 for incoming_p here, it doesn't matter. */
5603 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5605 /* If register required leading padding, add it. */
5607 cum->words += padding;
5611 cum->words += (mode != BLKmode
5612 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5613 : ROUND_ADVANCE (int_size_in_bytes (type)));
5617 if (type && AGGREGATE_TYPE_P (type))
5619 int size = int_size_in_bytes (type);
5623 else if (size <= 16)
5625 else /* passed by reference */
5630 cum->words += (mode != BLKmode
5631 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5632 : ROUND_ADVANCE (int_size_in_bytes (type)));
5637 /* Handle the FUNCTION_ARG_PADDING macro.
5638 For the 64 bit ABI structs are always stored left shifted in their
5642 function_arg_padding (enum machine_mode mode, const_tree type)
5644 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
5647 /* Fall back to the default. */
5648 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
5651 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
5652 Specify whether to return the return value in memory. */
5655 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
5658 /* Original SPARC 32-bit ABI says that structures and unions,
5659 and quad-precision floats are returned in memory. All other
5660 base types are returned in registers.
5662 Extended ABI (as implemented by the Sun compiler) says that
5663 all complex floats are returned in registers (8 FP registers
5664 at most for '_Complex long double'). Return all complex integers
5665 in registers (4 at most for '_Complex long long').
5667 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5668 integers are returned like floats of the same size, that is in
5669 registers up to 8 bytes and in memory otherwise. Return all
5670 vector floats in memory like structure and unions; note that
5671 they always have BLKmode like the latter. */
5672 return (TYPE_MODE (type) == BLKmode
5673 || TYPE_MODE (type) == TFmode
5674 || (TREE_CODE (type) == VECTOR_TYPE
5675 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5677 /* Original SPARC 64-bit ABI says that structures and unions
5678 smaller than 32 bytes are returned in registers, as well as
5679 all other base types.
5681 Extended ABI (as implemented by the Sun compiler) says that all
5682 complex floats are returned in registers (8 FP registers at most
5683 for '_Complex long double'). Return all complex integers in
5684 registers (4 at most for '_Complex TItype').
5686 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5687 integers are returned like floats of the same size, that is in
5688 registers. Return all vector floats like structure and unions;
5689 note that they always have BLKmode like the latter. */
5690 return ((TYPE_MODE (type) == BLKmode
5691 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32));
5694 /* Handle the TARGET_STRUCT_VALUE target hook.
5695 Return where to find the structure return value address. */
5698 sparc_struct_value_rtx (tree fndecl, int incoming)
5707 mem = gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx,
5708 STRUCT_VALUE_OFFSET));
5710 mem = gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx,
5711 STRUCT_VALUE_OFFSET));
5713 /* Only follow the SPARC ABI for fixed-size structure returns.
5714 Variable size structure returns are handled per the normal
5715 procedures in GCC. This is enabled by -mstd-struct-return */
5717 && sparc_std_struct_return
5718 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
5719 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
5721 /* We must check and adjust the return address, as it is
5722 optional as to whether the return object is really
5724 rtx ret_rtx = gen_rtx_REG (Pmode, 31);
5725 rtx scratch = gen_reg_rtx (SImode);
5726 rtx endlab = gen_label_rtx ();
5728 /* Calculate the return object size */
5729 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
5730 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
5731 /* Construct a temporary return value */
5732 rtx temp_val = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
5734 /* Implement SPARC 32-bit psABI callee returns struck checking
5737 Fetch the instruction where we will return to and see if
5738 it's an unimp instruction (the most significant 10 bits
5740 emit_move_insn (scratch, gen_rtx_MEM (SImode,
5741 plus_constant (ret_rtx, 8)));
5742 /* Assume the size is valid and pre-adjust */
5743 emit_insn (gen_add3_insn (ret_rtx, ret_rtx, GEN_INT (4)));
5744 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode, 0, endlab);
5745 emit_insn (gen_sub3_insn (ret_rtx, ret_rtx, GEN_INT (4)));
5746 /* Assign stack temp:
5747 Write the address of the memory pointed to by temp_val into
5748 the memory pointed to by mem */
5749 emit_move_insn (mem, XEXP (temp_val, 0));
5750 emit_label (endlab);
5753 set_mem_alias_set (mem, struct_value_alias_set);
5758 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
5759 For v9, function return values are subject to the same rules as arguments,
5760 except that up to 32 bytes may be returned in registers. */
5763 sparc_function_value_1 (const_tree type, enum machine_mode mode,
5766 /* Beware that the two values are swapped here wrt function_arg. */
5767 int regbase = (outgoing
5768 ? SPARC_INCOMING_INT_ARG_FIRST
5769 : SPARC_OUTGOING_INT_ARG_FIRST);
5770 enum mode_class mclass = GET_MODE_CLASS (mode);
5773 /* Vector types deserve special treatment because they are polymorphic wrt
5774 their mode, depending upon whether VIS instructions are enabled. */
5775 if (type && TREE_CODE (type) == VECTOR_TYPE)
5777 HOST_WIDE_INT size = int_size_in_bytes (type);
5778 gcc_assert ((TARGET_ARCH32 && size <= 8)
5779 || (TARGET_ARCH64 && size <= 32));
5781 if (mode == BLKmode)
5782 return function_arg_vector_value (size,
5783 SPARC_FP_ARG_FIRST);
5785 mclass = MODE_FLOAT;
5788 if (TARGET_ARCH64 && type)
5790 /* Structures up to 32 bytes in size are returned in registers. */
5791 if (TREE_CODE (type) == RECORD_TYPE)
5793 HOST_WIDE_INT size = int_size_in_bytes (type);
5794 gcc_assert (size <= 32);
5796 return function_arg_record_value (type, mode, 0, 1, regbase);
5799 /* Unions up to 32 bytes in size are returned in integer registers. */
5800 else if (TREE_CODE (type) == UNION_TYPE)
5802 HOST_WIDE_INT size = int_size_in_bytes (type);
5803 gcc_assert (size <= 32);
5805 return function_arg_union_value (size, mode, 0, regbase);
5808 /* Objects that require it are returned in FP registers. */
5809 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5812 /* All other aggregate types are returned in an integer register in a
5813 mode corresponding to the size of the type. */
5814 else if (AGGREGATE_TYPE_P (type))
5816 /* All other aggregate types are passed in an integer register
5817 in a mode corresponding to the size of the type. */
5818 HOST_WIDE_INT size = int_size_in_bytes (type);
5819 gcc_assert (size <= 32);
5821 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
5823 /* ??? We probably should have made the same ABI change in
5824 3.4.0 as the one we made for unions. The latter was
5825 required by the SCD though, while the former is not
5826 specified, so we favored compatibility and efficiency.
5828 Now we're stuck for aggregates larger than 16 bytes,
5829 because OImode vanished in the meantime. Let's not
5830 try to be unduly clever, and simply follow the ABI
5831 for unions in that case. */
5832 if (mode == BLKmode)
5833 return function_arg_union_value (size, mode, 0, regbase);
5838 /* This must match sparc_promote_function_mode.
5839 ??? Maybe 32-bit pointers should actually remain in Pmode? */
5840 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5844 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
5845 regno = SPARC_FP_ARG_FIRST;
5849 return gen_rtx_REG (mode, regno);
5852 /* Handle TARGET_FUNCTION_VALUE.
5854 On SPARC the value is found in the first "output" register, but the called
5855 function leaves it in the first "input" register. */
5858 sparc_function_value (const_tree valtype,
5859 const_tree fn_decl_or_type ATTRIBUTE_UNUSED,
5862 return sparc_function_value_1 (valtype, TYPE_MODE (valtype), outgoing);
5865 /* Handle TARGET_LIBCALL_VALUE. */
5868 sparc_libcall_value (enum machine_mode mode,
5869 const_rtx fun ATTRIBUTE_UNUSED)
5871 return sparc_function_value_1 (NULL_TREE, mode, false);
5874 /* Handle FUNCTION_VALUE_REGNO_P.
5875 On SPARC, the first "output" reg is used for integer values, and
5876 the first floating point register is used for floating point values. */
5879 sparc_function_value_regno_p (const unsigned int regno)
5881 return (regno == 8 || regno == 32);
5884 /* Do what is necessary for `va_start'. We look at the current function
5885 to determine if stdarg or varargs is used and return the address of
5886 the first unnamed parameter. */
5889 sparc_builtin_saveregs (void)
5891 int first_reg = crtl->args.info.words;
5895 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
5896 emit_move_insn (gen_rtx_MEM (word_mode,
5897 gen_rtx_PLUS (Pmode,
5899 GEN_INT (FIRST_PARM_OFFSET (0)
5902 gen_rtx_REG (word_mode,
5903 SPARC_INCOMING_INT_ARG_FIRST + regno));
5905 address = gen_rtx_PLUS (Pmode,
5907 GEN_INT (FIRST_PARM_OFFSET (0)
5908 + UNITS_PER_WORD * first_reg));
5913 /* Implement `va_start' for stdarg. */
5916 sparc_va_start (tree valist, rtx nextarg)
5918 nextarg = expand_builtin_saveregs ();
5919 std_expand_builtin_va_start (valist, nextarg);
5922 /* Implement `va_arg' for stdarg. */
5925 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
5928 HOST_WIDE_INT size, rsize, align;
5931 tree ptrtype = build_pointer_type (type);
5933 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
5936 size = rsize = UNITS_PER_WORD;
5942 size = int_size_in_bytes (type);
5943 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
5948 /* For SPARC64, objects requiring 16-byte alignment get it. */
5949 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
5950 align = 2 * UNITS_PER_WORD;
5952 /* SPARC-V9 ABI states that structures up to 16 bytes in size
5953 are left-justified in their slots. */
5954 if (AGGREGATE_TYPE_P (type))
5957 size = rsize = UNITS_PER_WORD;
5967 incr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr,
5968 size_int (align - 1));
5969 incr = fold_convert (sizetype, incr);
5970 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
5972 incr = fold_convert (ptr_type_node, incr);
5975 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
5978 if (BYTES_BIG_ENDIAN && size < rsize)
5979 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr,
5980 size_int (rsize - size));
5984 addr = fold_convert (build_pointer_type (ptrtype), addr);
5985 addr = build_va_arg_indirect_ref (addr);
5988 /* If the address isn't aligned properly for the type, we need a temporary.
5989 FIXME: This is inefficient, usually we can do this in registers. */
5990 else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD)
5992 tree tmp = create_tmp_var (type, "va_arg_tmp");
5993 tree dest_addr = build_fold_addr_expr (tmp);
5994 tree copy = build_call_expr (implicit_built_in_decls[BUILT_IN_MEMCPY],
5995 3, dest_addr, addr, size_int (rsize));
5996 TREE_ADDRESSABLE (tmp) = 1;
5997 gimplify_and_add (copy, pre_p);
6002 addr = fold_convert (ptrtype, addr);
6005 = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr, size_int (rsize));
6006 gimplify_assign (valist, incr, post_p);
6008 return build_va_arg_indirect_ref (addr);
6011 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
6012 Specify whether the vector mode is supported by the hardware. */
6015 sparc_vector_mode_supported_p (enum machine_mode mode)
6017 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
6020 /* Return the string to output an unconditional branch to LABEL, which is
6021 the operand number of the label.
6023 DEST is the destination insn (i.e. the label), INSN is the source. */
6026 output_ubranch (rtx dest, int label, rtx insn)
6028 static char string[64];
6029 bool v9_form = false;
6032 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
6034 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6035 - INSN_ADDRESSES (INSN_UID (insn)));
6036 /* Leave some instructions for "slop". */
6037 if (delta >= -260000 && delta < 260000)
6042 strcpy (string, "ba%*,pt\t%%xcc, ");
6044 strcpy (string, "b%*\t");
6046 p = strchr (string, '\0');
6057 /* Return the string to output a conditional branch to LABEL, which is
6058 the operand number of the label. OP is the conditional expression.
6059 XEXP (OP, 0) is assumed to be a condition code register (integer or
6060 floating point) and its mode specifies what kind of comparison we made.
6062 DEST is the destination insn (i.e. the label), INSN is the source.
6064 REVERSED is nonzero if we should reverse the sense of the comparison.
6066 ANNUL is nonzero if we should generate an annulling branch. */
6069 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
6072 static char string[64];
6073 enum rtx_code code = GET_CODE (op);
6074 rtx cc_reg = XEXP (op, 0);
6075 enum machine_mode mode = GET_MODE (cc_reg);
6076 const char *labelno, *branch;
6077 int spaces = 8, far;
6080 /* v9 branches are limited to +-1MB. If it is too far away,
6093 fbne,a,pn %fcc2, .LC29
6101 far = TARGET_V9 && (get_attr_length (insn) >= 3);
6104 /* Reversal of FP compares takes care -- an ordered compare
6105 becomes an unordered compare and vice versa. */
6106 if (mode == CCFPmode || mode == CCFPEmode)
6107 code = reverse_condition_maybe_unordered (code);
6109 code = reverse_condition (code);
6112 /* Start by writing the branch condition. */
6113 if (mode == CCFPmode || mode == CCFPEmode)
6164 /* ??? !v9: FP branches cannot be preceded by another floating point
6165 insn. Because there is currently no concept of pre-delay slots,
6166 we can fix this only by always emitting a nop before a floating
6171 strcpy (string, "nop\n\t");
6172 strcat (string, branch);
6185 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6197 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6218 strcpy (string, branch);
6220 spaces -= strlen (branch);
6221 p = strchr (string, '\0');
6223 /* Now add the annulling, the label, and a possible noop. */
6236 if (! far && insn && INSN_ADDRESSES_SET_P ())
6238 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6239 - INSN_ADDRESSES (INSN_UID (insn)));
6240 /* Leave some instructions for "slop". */
6241 if (delta < -260000 || delta >= 260000)
6245 if (mode == CCFPmode || mode == CCFPEmode)
6247 static char v9_fcc_labelno[] = "%%fccX, ";
6248 /* Set the char indicating the number of the fcc reg to use. */
6249 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
6250 labelno = v9_fcc_labelno;
6253 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
6257 else if (mode == CCXmode || mode == CCX_NOOVmode)
6259 labelno = "%%xcc, ";
6264 labelno = "%%icc, ";
6269 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6272 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6285 strcpy (p, labelno);
6286 p = strchr (p, '\0');
6289 strcpy (p, ".+12\n\t nop\n\tb\t");
6290 /* Skip the next insn if requested or
6291 if we know that it will be a nop. */
6292 if (annul || ! final_sequence)
6306 /* Emit a library call comparison between floating point X and Y.
6307 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
6308 Return the new operator to be used in the comparison sequence.
6310 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
6311 values as arguments instead of the TFmode registers themselves,
6312 that's why we cannot call emit_float_lib_cmp. */
6315 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
6318 rtx slot0, slot1, result, tem, tem2, libfunc;
6319 enum machine_mode mode;
6320 enum rtx_code new_comparison;
6325 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
6329 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
6333 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
6337 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
6341 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
6345 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
6356 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
6369 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6370 emit_move_insn (slot0, x);
6377 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6378 emit_move_insn (slot1, y);
6381 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
6382 emit_library_call (libfunc, LCT_NORMAL,
6384 XEXP (slot0, 0), Pmode,
6385 XEXP (slot1, 0), Pmode);
6390 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
6391 emit_library_call (libfunc, LCT_NORMAL,
6393 x, TFmode, y, TFmode);
6398 /* Immediately move the result of the libcall into a pseudo
6399 register so reload doesn't clobber the value if it needs
6400 the return register for a spill reg. */
6401 result = gen_reg_rtx (mode);
6402 emit_move_insn (result, hard_libcall_value (mode, libfunc));
6407 return gen_rtx_NE (VOIDmode, result, const0_rtx);
6410 new_comparison = (comparison == UNORDERED ? EQ : NE);
6411 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3));
6414 new_comparison = (comparison == UNGT ? GT : NE);
6415 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx);
6417 return gen_rtx_NE (VOIDmode, result, const2_rtx);
6419 tem = gen_reg_rtx (mode);
6421 emit_insn (gen_andsi3 (tem, result, const1_rtx));
6423 emit_insn (gen_anddi3 (tem, result, const1_rtx));
6424 return gen_rtx_NE (VOIDmode, tem, const0_rtx);
6427 tem = gen_reg_rtx (mode);
6429 emit_insn (gen_addsi3 (tem, result, const1_rtx));
6431 emit_insn (gen_adddi3 (tem, result, const1_rtx));
6432 tem2 = gen_reg_rtx (mode);
6434 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
6436 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
6437 new_comparison = (comparison == UNEQ ? EQ : NE);
6438 return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx);
6444 /* Generate an unsigned DImode to FP conversion. This is the same code
6445 optabs would emit if we didn't have TFmode patterns. */
6448 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
6450 rtx neglab, donelab, i0, i1, f0, in, out;
6453 in = force_reg (DImode, operands[1]);
6454 neglab = gen_label_rtx ();
6455 donelab = gen_label_rtx ();
6456 i0 = gen_reg_rtx (DImode);
6457 i1 = gen_reg_rtx (DImode);
6458 f0 = gen_reg_rtx (mode);
6460 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
6462 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
6463 emit_jump_insn (gen_jump (donelab));
6466 emit_label (neglab);
6468 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
6469 emit_insn (gen_anddi3 (i1, in, const1_rtx));
6470 emit_insn (gen_iordi3 (i0, i0, i1));
6471 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
6472 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
6474 emit_label (donelab);
6477 /* Generate an FP to unsigned DImode conversion. This is the same code
6478 optabs would emit if we didn't have TFmode patterns. */
6481 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
6483 rtx neglab, donelab, i0, i1, f0, in, out, limit;
6486 in = force_reg (mode, operands[1]);
6487 neglab = gen_label_rtx ();
6488 donelab = gen_label_rtx ();
6489 i0 = gen_reg_rtx (DImode);
6490 i1 = gen_reg_rtx (DImode);
6491 limit = gen_reg_rtx (mode);
6492 f0 = gen_reg_rtx (mode);
6494 emit_move_insn (limit,
6495 CONST_DOUBLE_FROM_REAL_VALUE (
6496 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
6497 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
6499 emit_insn (gen_rtx_SET (VOIDmode,
6501 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
6502 emit_jump_insn (gen_jump (donelab));
6505 emit_label (neglab);
6507 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
6508 emit_insn (gen_rtx_SET (VOIDmode,
6510 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
6511 emit_insn (gen_movdi (i1, const1_rtx));
6512 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
6513 emit_insn (gen_xordi3 (out, i0, i1));
6515 emit_label (donelab);
6518 /* Return the string to output a conditional branch to LABEL, testing
6519 register REG. LABEL is the operand number of the label; REG is the
6520 operand number of the reg. OP is the conditional expression. The mode
6521 of REG says what kind of comparison we made.
6523 DEST is the destination insn (i.e. the label), INSN is the source.
6525 REVERSED is nonzero if we should reverse the sense of the comparison.
6527 ANNUL is nonzero if we should generate an annulling branch. */
6530 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
6531 int annul, rtx insn)
6533 static char string[64];
6534 enum rtx_code code = GET_CODE (op);
6535 enum machine_mode mode = GET_MODE (XEXP (op, 0));
6540 /* branch on register are limited to +-128KB. If it is too far away,
6553 brgez,a,pn %o1, .LC29
6559 ba,pt %xcc, .LC29 */
6561 far = get_attr_length (insn) >= 3;
6563 /* If not floating-point or if EQ or NE, we can just reverse the code. */
6565 code = reverse_condition (code);
6567 /* Only 64 bit versions of these instructions exist. */
6568 gcc_assert (mode == DImode);
6570 /* Start by writing the branch condition. */
6575 strcpy (string, "brnz");
6579 strcpy (string, "brz");
6583 strcpy (string, "brgez");
6587 strcpy (string, "brlz");
6591 strcpy (string, "brlez");
6595 strcpy (string, "brgz");
6602 p = strchr (string, '\0');
6604 /* Now add the annulling, reg, label, and nop. */
6611 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6614 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6619 *p = p < string + 8 ? '\t' : ' ';
6627 int veryfar = 1, delta;
6629 if (INSN_ADDRESSES_SET_P ())
6631 delta = (INSN_ADDRESSES (INSN_UID (dest))
6632 - INSN_ADDRESSES (INSN_UID (insn)));
6633 /* Leave some instructions for "slop". */
6634 if (delta >= -260000 && delta < 260000)
6638 strcpy (p, ".+12\n\t nop\n\t");
6639 /* Skip the next insn if requested or
6640 if we know that it will be a nop. */
6641 if (annul || ! final_sequence)
6651 strcpy (p, "ba,pt\t%%xcc, ");
6665 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
6666 Such instructions cannot be used in the delay slot of return insn on v9.
6667 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
6671 epilogue_renumber (register rtx *where, int test)
6673 register const char *fmt;
6675 register enum rtx_code code;
6680 code = GET_CODE (*where);
6685 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
6687 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
6688 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
6696 /* Do not replace the frame pointer with the stack pointer because
6697 it can cause the delayed instruction to load below the stack.
6698 This occurs when instructions like:
6700 (set (reg/i:SI 24 %i0)
6701 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
6702 (const_int -20 [0xffffffec])) 0))
6704 are in the return delayed slot. */
6706 if (GET_CODE (XEXP (*where, 0)) == REG
6707 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
6708 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
6709 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
6714 if (SPARC_STACK_BIAS
6715 && GET_CODE (XEXP (*where, 0)) == REG
6716 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
6724 fmt = GET_RTX_FORMAT (code);
6726 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6731 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
6732 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
6735 else if (fmt[i] == 'e'
6736 && epilogue_renumber (&(XEXP (*where, i)), test))
6742 /* Leaf functions and non-leaf functions have different needs. */
6745 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
6748 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
6750 static const int *const reg_alloc_orders[] = {
6751 reg_leaf_alloc_order,
6752 reg_nonleaf_alloc_order};
6755 order_regs_for_local_alloc (void)
6757 static int last_order_nonleaf = 1;
6759 if (df_regs_ever_live_p (15) != last_order_nonleaf)
6761 last_order_nonleaf = !last_order_nonleaf;
6762 memcpy ((char *) reg_alloc_order,
6763 (const char *) reg_alloc_orders[last_order_nonleaf],
6764 FIRST_PSEUDO_REGISTER * sizeof (int));
6768 /* Return 1 if REG and MEM are legitimate enough to allow the various
6769 mem<-->reg splits to be run. */
6772 sparc_splitdi_legitimate (rtx reg, rtx mem)
6774 /* Punt if we are here by mistake. */
6775 gcc_assert (reload_completed);
6777 /* We must have an offsettable memory reference. */
6778 if (! offsettable_memref_p (mem))
6781 /* If we have legitimate args for ldd/std, we do not want
6782 the split to happen. */
6783 if ((REGNO (reg) % 2) == 0
6784 && mem_min_alignment (mem, 8))
6791 /* Return 1 if x and y are some kind of REG and they refer to
6792 different hard registers. This test is guaranteed to be
6793 run after reload. */
6796 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
6798 if (GET_CODE (x) != REG)
6800 if (GET_CODE (y) != REG)
6802 if (REGNO (x) == REGNO (y))
6807 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
6808 This makes them candidates for using ldd and std insns.
6810 Note reg1 and reg2 *must* be hard registers. */
6813 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
6815 /* We might have been passed a SUBREG. */
6816 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
6819 if (REGNO (reg1) % 2 != 0)
6822 /* Integer ldd is deprecated in SPARC V9 */
6823 if (TARGET_V9 && REGNO (reg1) < 32)
6826 return (REGNO (reg1) == REGNO (reg2) - 1);
6829 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
6832 This can only happen when addr1 and addr2, the addresses in mem1
6833 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
6834 addr1 must also be aligned on a 64-bit boundary.
6836 Also iff dependent_reg_rtx is not null it should not be used to
6837 compute the address for mem1, i.e. we cannot optimize a sequence
6849 But, note that the transformation from:
6854 is perfectly fine. Thus, the peephole2 patterns always pass us
6855 the destination register of the first load, never the second one.
6857 For stores we don't have a similar problem, so dependent_reg_rtx is
6861 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
6865 HOST_WIDE_INT offset1;
6867 /* The mems cannot be volatile. */
6868 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
6871 /* MEM1 should be aligned on a 64-bit boundary. */
6872 if (MEM_ALIGN (mem1) < 64)
6875 addr1 = XEXP (mem1, 0);
6876 addr2 = XEXP (mem2, 0);
6878 /* Extract a register number and offset (if used) from the first addr. */
6879 if (GET_CODE (addr1) == PLUS)
6881 /* If not a REG, return zero. */
6882 if (GET_CODE (XEXP (addr1, 0)) != REG)
6886 reg1 = REGNO (XEXP (addr1, 0));
6887 /* The offset must be constant! */
6888 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
6890 offset1 = INTVAL (XEXP (addr1, 1));
6893 else if (GET_CODE (addr1) != REG)
6897 reg1 = REGNO (addr1);
6898 /* This was a simple (mem (reg)) expression. Offset is 0. */
6902 /* Make sure the second address is a (mem (plus (reg) (const_int). */
6903 if (GET_CODE (addr2) != PLUS)
6906 if (GET_CODE (XEXP (addr2, 0)) != REG
6907 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
6910 if (reg1 != REGNO (XEXP (addr2, 0)))
6913 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
6916 /* The first offset must be evenly divisible by 8 to ensure the
6917 address is 64 bit aligned. */
6918 if (offset1 % 8 != 0)
6921 /* The offset for the second addr must be 4 more than the first addr. */
6922 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
6925 /* All the tests passed. addr1 and addr2 are valid for ldd and std
6930 /* Return 1 if reg is a pseudo, or is the first register in
6931 a hard register pair. This makes it suitable for use in
6932 ldd and std insns. */
6935 register_ok_for_ldd (rtx reg)
6937 /* We might have been passed a SUBREG. */
6941 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
6942 return (REGNO (reg) % 2 == 0);
6947 /* Return 1 if OP is a memory whose address is known to be
6948 aligned to 8-byte boundary, or a pseudo during reload.
6949 This makes it suitable for use in ldd and std insns. */
6952 memory_ok_for_ldd (rtx op)
6956 /* In 64-bit mode, we assume that the address is word-aligned. */
6957 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
6960 if ((reload_in_progress || reload_completed)
6961 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
6964 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
6966 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
6975 /* Print operand X (an rtx) in assembler syntax to file FILE.
6976 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
6977 For `%' followed by punctuation, CODE is the punctuation and X is null. */
6980 print_operand (FILE *file, rtx x, int code)
6985 /* Output an insn in a delay slot. */
6987 sparc_indent_opcode = 1;
6989 fputs ("\n\t nop", file);
6992 /* Output an annul flag if there's nothing for the delay slot and we
6993 are optimizing. This is always used with '(' below.
6994 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
6995 this is a dbx bug. So, we only do this when optimizing.
6996 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
6997 Always emit a nop in case the next instruction is a branch. */
6998 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
7002 /* Output a 'nop' if there's nothing for the delay slot and we are
7003 not optimizing. This is always used with '*' above. */
7004 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
7005 fputs ("\n\t nop", file);
7006 else if (final_sequence)
7007 sparc_indent_opcode = 1;
7010 /* Output the right displacement from the saved PC on function return.
7011 The caller may have placed an "unimp" insn immediately after the call
7012 so we have to account for it. This insn is used in the 32-bit ABI
7013 when calling a function that returns a non zero-sized structure. The
7014 64-bit ABI doesn't have it. Be careful to have this test be the same
7015 as that for the call. The exception is when sparc_std_struct_return
7016 is enabled, the psABI is followed exactly and the adjustment is made
7017 by the code in sparc_struct_value_rtx. The call emitted is the same
7018 when sparc_std_struct_return is enabled. */
7020 && cfun->returns_struct
7021 && !sparc_std_struct_return
7022 && DECL_SIZE (DECL_RESULT (current_function_decl))
7023 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
7025 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
7031 /* Output the Embedded Medium/Anywhere code model base register. */
7032 fputs (EMBMEDANY_BASE_REG, file);
7035 /* Print some local dynamic TLS name. */
7036 assemble_name (file, get_some_local_dynamic_name ());
7040 /* Adjust the operand to take into account a RESTORE operation. */
7041 if (GET_CODE (x) == CONST_INT)
7043 else if (GET_CODE (x) != REG)
7044 output_operand_lossage ("invalid %%Y operand");
7045 else if (REGNO (x) < 8)
7046 fputs (reg_names[REGNO (x)], file);
7047 else if (REGNO (x) >= 24 && REGNO (x) < 32)
7048 fputs (reg_names[REGNO (x)-16], file);
7050 output_operand_lossage ("invalid %%Y operand");
7053 /* Print out the low order register name of a register pair. */
7054 if (WORDS_BIG_ENDIAN)
7055 fputs (reg_names[REGNO (x)+1], file);
7057 fputs (reg_names[REGNO (x)], file);
7060 /* Print out the high order register name of a register pair. */
7061 if (WORDS_BIG_ENDIAN)
7062 fputs (reg_names[REGNO (x)], file);
7064 fputs (reg_names[REGNO (x)+1], file);
7067 /* Print out the second register name of a register pair or quad.
7068 I.e., R (%o0) => %o1. */
7069 fputs (reg_names[REGNO (x)+1], file);
7072 /* Print out the third register name of a register quad.
7073 I.e., S (%o0) => %o2. */
7074 fputs (reg_names[REGNO (x)+2], file);
7077 /* Print out the fourth register name of a register quad.
7078 I.e., T (%o0) => %o3. */
7079 fputs (reg_names[REGNO (x)+3], file);
7082 /* Print a condition code register. */
7083 if (REGNO (x) == SPARC_ICC_REG)
7085 /* We don't handle CC[X]_NOOVmode because they're not supposed
7087 if (GET_MODE (x) == CCmode)
7088 fputs ("%icc", file);
7089 else if (GET_MODE (x) == CCXmode)
7090 fputs ("%xcc", file);
7095 /* %fccN register */
7096 fputs (reg_names[REGNO (x)], file);
7099 /* Print the operand's address only. */
7100 output_address (XEXP (x, 0));
7103 /* In this case we need a register. Use %g0 if the
7104 operand is const0_rtx. */
7106 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
7108 fputs ("%g0", file);
7115 switch (GET_CODE (x))
7117 case IOR: fputs ("or", file); break;
7118 case AND: fputs ("and", file); break;
7119 case XOR: fputs ("xor", file); break;
7120 default: output_operand_lossage ("invalid %%A operand");
7125 switch (GET_CODE (x))
7127 case IOR: fputs ("orn", file); break;
7128 case AND: fputs ("andn", file); break;
7129 case XOR: fputs ("xnor", file); break;
7130 default: output_operand_lossage ("invalid %%B operand");
7134 /* These are used by the conditional move instructions. */
7138 enum rtx_code rc = GET_CODE (x);
7142 enum machine_mode mode = GET_MODE (XEXP (x, 0));
7143 if (mode == CCFPmode || mode == CCFPEmode)
7144 rc = reverse_condition_maybe_unordered (GET_CODE (x));
7146 rc = reverse_condition (GET_CODE (x));
7150 case NE: fputs ("ne", file); break;
7151 case EQ: fputs ("e", file); break;
7152 case GE: fputs ("ge", file); break;
7153 case GT: fputs ("g", file); break;
7154 case LE: fputs ("le", file); break;
7155 case LT: fputs ("l", file); break;
7156 case GEU: fputs ("geu", file); break;
7157 case GTU: fputs ("gu", file); break;
7158 case LEU: fputs ("leu", file); break;
7159 case LTU: fputs ("lu", file); break;
7160 case LTGT: fputs ("lg", file); break;
7161 case UNORDERED: fputs ("u", file); break;
7162 case ORDERED: fputs ("o", file); break;
7163 case UNLT: fputs ("ul", file); break;
7164 case UNLE: fputs ("ule", file); break;
7165 case UNGT: fputs ("ug", file); break;
7166 case UNGE: fputs ("uge", file); break;
7167 case UNEQ: fputs ("ue", file); break;
7168 default: output_operand_lossage (code == 'c'
7169 ? "invalid %%c operand"
7170 : "invalid %%C operand");
7175 /* These are used by the movr instruction pattern. */
7179 enum rtx_code rc = (code == 'd'
7180 ? reverse_condition (GET_CODE (x))
7184 case NE: fputs ("ne", file); break;
7185 case EQ: fputs ("e", file); break;
7186 case GE: fputs ("gez", file); break;
7187 case LT: fputs ("lz", file); break;
7188 case LE: fputs ("lez", file); break;
7189 case GT: fputs ("gz", file); break;
7190 default: output_operand_lossage (code == 'd'
7191 ? "invalid %%d operand"
7192 : "invalid %%D operand");
7199 /* Print a sign-extended character. */
7200 int i = trunc_int_for_mode (INTVAL (x), QImode);
7201 fprintf (file, "%d", i);
7206 /* Operand must be a MEM; write its address. */
7207 if (GET_CODE (x) != MEM)
7208 output_operand_lossage ("invalid %%f operand");
7209 output_address (XEXP (x, 0));
7214 /* Print a sign-extended 32-bit value. */
7216 if (GET_CODE(x) == CONST_INT)
7218 else if (GET_CODE(x) == CONST_DOUBLE)
7219 i = CONST_DOUBLE_LOW (x);
7222 output_operand_lossage ("invalid %%s operand");
7225 i = trunc_int_for_mode (i, SImode);
7226 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
7231 /* Do nothing special. */
7235 /* Undocumented flag. */
7236 output_operand_lossage ("invalid operand output code");
7239 if (GET_CODE (x) == REG)
7240 fputs (reg_names[REGNO (x)], file);
7241 else if (GET_CODE (x) == MEM)
7244 /* Poor Sun assembler doesn't understand absolute addressing. */
7245 if (CONSTANT_P (XEXP (x, 0)))
7246 fputs ("%g0+", file);
7247 output_address (XEXP (x, 0));
7250 else if (GET_CODE (x) == HIGH)
7252 fputs ("%hi(", file);
7253 output_addr_const (file, XEXP (x, 0));
7256 else if (GET_CODE (x) == LO_SUM)
7258 print_operand (file, XEXP (x, 0), 0);
7259 if (TARGET_CM_MEDMID)
7260 fputs ("+%l44(", file);
7262 fputs ("+%lo(", file);
7263 output_addr_const (file, XEXP (x, 1));
7266 else if (GET_CODE (x) == CONST_DOUBLE
7267 && (GET_MODE (x) == VOIDmode
7268 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
7270 if (CONST_DOUBLE_HIGH (x) == 0)
7271 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
7272 else if (CONST_DOUBLE_HIGH (x) == -1
7273 && CONST_DOUBLE_LOW (x) < 0)
7274 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
7276 output_operand_lossage ("long long constant not a valid immediate operand");
7278 else if (GET_CODE (x) == CONST_DOUBLE)
7279 output_operand_lossage ("floating point constant not a valid immediate operand");
7280 else { output_addr_const (file, x); }
7283 /* Target hook for assembling integer objects. The sparc version has
7284 special handling for aligned DI-mode objects. */
7287 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
7289 /* ??? We only output .xword's for symbols and only then in environments
7290 where the assembler can handle them. */
7291 if (aligned_p && size == 8
7292 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
7296 assemble_integer_with_op ("\t.xword\t", x);
7301 assemble_aligned_integer (4, const0_rtx);
7302 assemble_aligned_integer (4, x);
7306 return default_assemble_integer (x, size, aligned_p);
7309 /* Return the value of a code used in the .proc pseudo-op that says
7310 what kind of result this function returns. For non-C types, we pick
7311 the closest C type. */
7313 #ifndef SHORT_TYPE_SIZE
7314 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
7317 #ifndef INT_TYPE_SIZE
7318 #define INT_TYPE_SIZE BITS_PER_WORD
7321 #ifndef LONG_TYPE_SIZE
7322 #define LONG_TYPE_SIZE BITS_PER_WORD
7325 #ifndef LONG_LONG_TYPE_SIZE
7326 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
7329 #ifndef FLOAT_TYPE_SIZE
7330 #define FLOAT_TYPE_SIZE BITS_PER_WORD
7333 #ifndef DOUBLE_TYPE_SIZE
7334 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7337 #ifndef LONG_DOUBLE_TYPE_SIZE
7338 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7342 sparc_type_code (register tree type)
7344 register unsigned long qualifiers = 0;
7345 register unsigned shift;
7347 /* Only the first 30 bits of the qualifier are valid. We must refrain from
7348 setting more, since some assemblers will give an error for this. Also,
7349 we must be careful to avoid shifts of 32 bits or more to avoid getting
7350 unpredictable results. */
7352 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
7354 switch (TREE_CODE (type))
7360 qualifiers |= (3 << shift);
7365 qualifiers |= (2 << shift);
7369 case REFERENCE_TYPE:
7371 qualifiers |= (1 << shift);
7375 return (qualifiers | 8);
7378 case QUAL_UNION_TYPE:
7379 return (qualifiers | 9);
7382 return (qualifiers | 10);
7385 return (qualifiers | 16);
7388 /* If this is a range type, consider it to be the underlying
7390 if (TREE_TYPE (type) != 0)
7393 /* Carefully distinguish all the standard types of C,
7394 without messing up if the language is not C. We do this by
7395 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
7396 look at both the names and the above fields, but that's redundant.
7397 Any type whose size is between two C types will be considered
7398 to be the wider of the two types. Also, we do not have a
7399 special code to use for "long long", so anything wider than
7400 long is treated the same. Note that we can't distinguish
7401 between "int" and "long" in this code if they are the same
7402 size, but that's fine, since neither can the assembler. */
7404 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
7405 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
7407 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
7408 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
7410 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
7411 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
7414 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
7417 /* If this is a range type, consider it to be the underlying
7419 if (TREE_TYPE (type) != 0)
7422 /* Carefully distinguish all the standard types of C,
7423 without messing up if the language is not C. */
7425 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
7426 return (qualifiers | 6);
7429 return (qualifiers | 7);
7431 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
7432 /* ??? We need to distinguish between double and float complex types,
7433 but I don't know how yet because I can't reach this code from
7434 existing front-ends. */
7435 return (qualifiers | 7); /* Who knows? */
7438 case BOOLEAN_TYPE: /* Boolean truth value type. */
7439 case LANG_TYPE: /* ? */
7443 gcc_unreachable (); /* Not a type! */
7450 /* Nested function support. */
7452 /* Emit RTL insns to initialize the variable parts of a trampoline.
7453 FNADDR is an RTX for the address of the function's pure code.
7454 CXT is an RTX for the static chain value for the function.
7456 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
7457 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
7458 (to store insns). This is a bit excessive. Perhaps a different
7459 mechanism would be better here.
7461 Emit enough FLUSH insns to synchronize the data and instruction caches. */
7464 sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
7466 /* SPARC 32-bit trampoline:
7469 sethi %hi(static), %g2
7471 or %g2, %lo(static), %g2
7473 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
7474 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
7478 (adjust_address (m_tramp, SImode, 0),
7479 expand_binop (SImode, ior_optab,
7480 expand_shift (RSHIFT_EXPR, SImode, fnaddr,
7481 size_int (10), 0, 1),
7482 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
7483 NULL_RTX, 1, OPTAB_DIRECT));
7486 (adjust_address (m_tramp, SImode, 4),
7487 expand_binop (SImode, ior_optab,
7488 expand_shift (RSHIFT_EXPR, SImode, cxt,
7489 size_int (10), 0, 1),
7490 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
7491 NULL_RTX, 1, OPTAB_DIRECT));
7494 (adjust_address (m_tramp, SImode, 8),
7495 expand_binop (SImode, ior_optab,
7496 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
7497 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
7498 NULL_RTX, 1, OPTAB_DIRECT));
7501 (adjust_address (m_tramp, SImode, 12),
7502 expand_binop (SImode, ior_optab,
7503 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
7504 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
7505 NULL_RTX, 1, OPTAB_DIRECT));
7507 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
7508 aligned on a 16 byte boundary so one flush clears it all. */
7509 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0))));
7510 if (sparc_cpu != PROCESSOR_ULTRASPARC
7511 && sparc_cpu != PROCESSOR_ULTRASPARC3
7512 && sparc_cpu != PROCESSOR_NIAGARA
7513 && sparc_cpu != PROCESSOR_NIAGARA2)
7514 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8))));
7516 /* Call __enable_execute_stack after writing onto the stack to make sure
7517 the stack address is accessible. */
7518 #ifdef ENABLE_EXECUTE_STACK
7519 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7520 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
7525 /* The 64-bit version is simpler because it makes more sense to load the
7526 values as "immediate" data out of the trampoline. It's also easier since
7527 we can read the PC without clobbering a register. */
7530 sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
7532 /* SPARC 64-bit trampoline:
7541 emit_move_insn (adjust_address (m_tramp, SImode, 0),
7542 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
7543 emit_move_insn (adjust_address (m_tramp, SImode, 4),
7544 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
7545 emit_move_insn (adjust_address (m_tramp, SImode, 8),
7546 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
7547 emit_move_insn (adjust_address (m_tramp, SImode, 12),
7548 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
7549 emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt);
7550 emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr);
7551 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0))));
7553 if (sparc_cpu != PROCESSOR_ULTRASPARC
7554 && sparc_cpu != PROCESSOR_ULTRASPARC3
7555 && sparc_cpu != PROCESSOR_NIAGARA
7556 && sparc_cpu != PROCESSOR_NIAGARA2)
7557 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8))));
7559 /* Call __enable_execute_stack after writing onto the stack to make sure
7560 the stack address is accessible. */
7561 #ifdef ENABLE_EXECUTE_STACK
7562 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7563 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
7567 /* Worker for TARGET_TRAMPOLINE_INIT. */
7570 sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt)
7572 rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0));
7573 cxt = force_reg (Pmode, cxt);
7575 sparc64_initialize_trampoline (m_tramp, fnaddr, cxt);
7577 sparc32_initialize_trampoline (m_tramp, fnaddr, cxt);
7580 /* Adjust the cost of a scheduling dependency. Return the new cost of
7581 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
7584 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7586 enum attr_type insn_type;
7588 if (! recog_memoized (insn))
7591 insn_type = get_attr_type (insn);
7593 if (REG_NOTE_KIND (link) == 0)
7595 /* Data dependency; DEP_INSN writes a register that INSN reads some
7598 /* if a load, then the dependence must be on the memory address;
7599 add an extra "cycle". Note that the cost could be two cycles
7600 if the reg was written late in an instruction group; we ca not tell
7602 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
7605 /* Get the delay only if the address of the store is the dependence. */
7606 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
7608 rtx pat = PATTERN(insn);
7609 rtx dep_pat = PATTERN (dep_insn);
7611 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7612 return cost; /* This should not happen! */
7614 /* The dependency between the two instructions was on the data that
7615 is being stored. Assume that this implies that the address of the
7616 store is not dependent. */
7617 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7620 return cost + 3; /* An approximation. */
7623 /* A shift instruction cannot receive its data from an instruction
7624 in the same cycle; add a one cycle penalty. */
7625 if (insn_type == TYPE_SHIFT)
7626 return cost + 3; /* Split before cascade into shift. */
7630 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
7631 INSN writes some cycles later. */
7633 /* These are only significant for the fpu unit; writing a fp reg before
7634 the fpu has finished with it stalls the processor. */
7636 /* Reusing an integer register causes no problems. */
7637 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7645 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7647 enum attr_type insn_type, dep_type;
7648 rtx pat = PATTERN(insn);
7649 rtx dep_pat = PATTERN (dep_insn);
7651 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
7654 insn_type = get_attr_type (insn);
7655 dep_type = get_attr_type (dep_insn);
7657 switch (REG_NOTE_KIND (link))
7660 /* Data dependency; DEP_INSN writes a register that INSN reads some
7667 /* Get the delay iff the address of the store is the dependence. */
7668 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7671 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7678 /* If a load, then the dependence must be on the memory address. If
7679 the addresses aren't equal, then it might be a false dependency */
7680 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
7682 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
7683 || GET_CODE (SET_DEST (dep_pat)) != MEM
7684 || GET_CODE (SET_SRC (pat)) != MEM
7685 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
7686 XEXP (SET_SRC (pat), 0)))
7694 /* Compare to branch latency is 0. There is no benefit from
7695 separating compare and branch. */
7696 if (dep_type == TYPE_COMPARE)
7698 /* Floating point compare to branch latency is less than
7699 compare to conditional move. */
7700 if (dep_type == TYPE_FPCMP)
7709 /* Anti-dependencies only penalize the fpu unit. */
7710 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7722 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
7726 case PROCESSOR_SUPERSPARC:
7727 cost = supersparc_adjust_cost (insn, link, dep, cost);
7729 case PROCESSOR_HYPERSPARC:
7730 case PROCESSOR_SPARCLITE86X:
7731 cost = hypersparc_adjust_cost (insn, link, dep, cost);
7740 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
7741 int sched_verbose ATTRIBUTE_UNUSED,
7742 int max_ready ATTRIBUTE_UNUSED)
7746 sparc_use_sched_lookahead (void)
7748 if (sparc_cpu == PROCESSOR_NIAGARA
7749 || sparc_cpu == PROCESSOR_NIAGARA2)
7751 if (sparc_cpu == PROCESSOR_ULTRASPARC
7752 || sparc_cpu == PROCESSOR_ULTRASPARC3)
7754 if ((1 << sparc_cpu) &
7755 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
7756 (1 << PROCESSOR_SPARCLITE86X)))
7762 sparc_issue_rate (void)
7766 case PROCESSOR_NIAGARA:
7767 case PROCESSOR_NIAGARA2:
7771 /* Assume V9 processors are capable of at least dual-issue. */
7773 case PROCESSOR_SUPERSPARC:
7775 case PROCESSOR_HYPERSPARC:
7776 case PROCESSOR_SPARCLITE86X:
7778 case PROCESSOR_ULTRASPARC:
7779 case PROCESSOR_ULTRASPARC3:
7785 set_extends (rtx insn)
7787 register rtx pat = PATTERN (insn);
7789 switch (GET_CODE (SET_SRC (pat)))
7791 /* Load and some shift instructions zero extend. */
7794 /* sethi clears the high bits */
7796 /* LO_SUM is used with sethi. sethi cleared the high
7797 bits and the values used with lo_sum are positive */
7799 /* Store flag stores 0 or 1 */
7809 rtx op0 = XEXP (SET_SRC (pat), 0);
7810 rtx op1 = XEXP (SET_SRC (pat), 1);
7811 if (GET_CODE (op1) == CONST_INT)
7812 return INTVAL (op1) >= 0;
7813 if (GET_CODE (op0) != REG)
7815 if (sparc_check_64 (op0, insn) == 1)
7817 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7822 rtx op0 = XEXP (SET_SRC (pat), 0);
7823 rtx op1 = XEXP (SET_SRC (pat), 1);
7824 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
7826 if (GET_CODE (op1) == CONST_INT)
7827 return INTVAL (op1) >= 0;
7828 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7831 return GET_MODE (SET_SRC (pat)) == SImode;
7832 /* Positive integers leave the high bits zero. */
7834 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
7836 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
7839 return - (GET_MODE (SET_SRC (pat)) == SImode);
7841 return sparc_check_64 (SET_SRC (pat), insn);
7847 /* We _ought_ to have only one kind per function, but... */
7848 static GTY(()) rtx sparc_addr_diff_list;
7849 static GTY(()) rtx sparc_addr_list;
7852 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
7854 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
7856 sparc_addr_diff_list
7857 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
7859 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
7863 sparc_output_addr_vec (rtx vec)
7865 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7866 int idx, vlen = XVECLEN (body, 0);
7868 #ifdef ASM_OUTPUT_ADDR_VEC_START
7869 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7872 #ifdef ASM_OUTPUT_CASE_LABEL
7873 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7876 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7879 for (idx = 0; idx < vlen; idx++)
7881 ASM_OUTPUT_ADDR_VEC_ELT
7882 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
7885 #ifdef ASM_OUTPUT_ADDR_VEC_END
7886 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7891 sparc_output_addr_diff_vec (rtx vec)
7893 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7894 rtx base = XEXP (XEXP (body, 0), 0);
7895 int idx, vlen = XVECLEN (body, 1);
7897 #ifdef ASM_OUTPUT_ADDR_VEC_START
7898 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7901 #ifdef ASM_OUTPUT_CASE_LABEL
7902 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7905 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7908 for (idx = 0; idx < vlen; idx++)
7910 ASM_OUTPUT_ADDR_DIFF_ELT
7913 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
7914 CODE_LABEL_NUMBER (base));
7917 #ifdef ASM_OUTPUT_ADDR_VEC_END
7918 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7923 sparc_output_deferred_case_vectors (void)
7928 if (sparc_addr_list == NULL_RTX
7929 && sparc_addr_diff_list == NULL_RTX)
7932 /* Align to cache line in the function's code section. */
7933 switch_to_section (current_function_section ());
7935 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
7937 ASM_OUTPUT_ALIGN (asm_out_file, align);
7939 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
7940 sparc_output_addr_vec (XEXP (t, 0));
7941 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
7942 sparc_output_addr_diff_vec (XEXP (t, 0));
7944 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
7947 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
7948 unknown. Return 1 if the high bits are zero, -1 if the register is
7951 sparc_check_64 (rtx x, rtx insn)
7953 /* If a register is set only once it is safe to ignore insns this
7954 code does not know how to handle. The loop will either recognize
7955 the single set and return the correct value or fail to recognize
7960 gcc_assert (GET_CODE (x) == REG);
7962 if (GET_MODE (x) == DImode)
7963 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
7965 if (flag_expensive_optimizations
7966 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
7972 insn = get_last_insn_anywhere ();
7977 while ((insn = PREV_INSN (insn)))
7979 switch (GET_CODE (insn))
7992 rtx pat = PATTERN (insn);
7993 if (GET_CODE (pat) != SET)
7995 if (rtx_equal_p (x, SET_DEST (pat)))
7996 return set_extends (insn);
7997 if (y && rtx_equal_p (y, SET_DEST (pat)))
7998 return set_extends (insn);
7999 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
8007 /* Returns assembly code to perform a DImode shift using
8008 a 64-bit global or out register on SPARC-V8+. */
8010 output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
8012 static char asm_code[60];
8014 /* The scratch register is only required when the destination
8015 register is not a 64-bit global or out register. */
8016 if (which_alternative != 2)
8017 operands[3] = operands[0];
8019 /* We can only shift by constants <= 63. */
8020 if (GET_CODE (operands[2]) == CONST_INT)
8021 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
8023 if (GET_CODE (operands[1]) == CONST_INT)
8025 output_asm_insn ("mov\t%1, %3", operands);
8029 output_asm_insn ("sllx\t%H1, 32, %3", operands);
8030 if (sparc_check_64 (operands[1], insn) <= 0)
8031 output_asm_insn ("srl\t%L1, 0, %L1", operands);
8032 output_asm_insn ("or\t%L1, %3, %3", operands);
8035 strcpy(asm_code, opcode);
8037 if (which_alternative != 2)
8038 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
8040 return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
8043 /* Output rtl to increment the profiler label LABELNO
8044 for profiling a function entry. */
8047 sparc_profile_hook (int labelno)
8052 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
8053 if (NO_PROFILE_COUNTERS)
8055 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
8059 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
8060 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
8061 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
8065 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
8068 sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags,
8069 tree decl ATTRIBUTE_UNUSED)
8071 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
8073 if (!(flags & SECTION_DEBUG))
8074 fputs (",#alloc", asm_out_file);
8075 if (flags & SECTION_WRITE)
8076 fputs (",#write", asm_out_file);
8077 if (flags & SECTION_TLS)
8078 fputs (",#tls", asm_out_file);
8079 if (flags & SECTION_CODE)
8080 fputs (",#execinstr", asm_out_file);
8082 /* ??? Handle SECTION_BSS. */
8084 fputc ('\n', asm_out_file);
8087 /* We do not allow indirect calls to be optimized into sibling calls.
8089 We cannot use sibling calls when delayed branches are disabled
8090 because they will likely require the call delay slot to be filled.
8092 Also, on SPARC 32-bit we cannot emit a sibling call when the
8093 current function returns a structure. This is because the "unimp
8094 after call" convention would cause the callee to return to the
8095 wrong place. The generic code already disallows cases where the
8096 function being called returns a structure.
8098 It may seem strange how this last case could occur. Usually there
8099 is code after the call which jumps to epilogue code which dumps the
8100 return value into the struct return area. That ought to invalidate
8101 the sibling call right? Well, in the C++ case we can end up passing
8102 the pointer to the struct return area to a constructor (which returns
8103 void) and then nothing else happens. Such a sibling call would look
8104 valid without the added check here.
8106 VxWorks PIC PLT entries require the global pointer to be initialized
8107 on entry. We therefore can't emit sibling calls to them. */
8109 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
8112 && flag_delayed_branch
8113 && (TARGET_ARCH64 || ! cfun->returns_struct)
8114 && !(TARGET_VXWORKS_RTP
8116 && !targetm.binds_local_p (decl)));
8119 /* libfunc renaming. */
8120 #include "config/gofast.h"
8123 sparc_init_libfuncs (void)
8127 /* Use the subroutines that Sun's library provides for integer
8128 multiply and divide. The `*' prevents an underscore from
8129 being prepended by the compiler. .umul is a little faster
8131 set_optab_libfunc (smul_optab, SImode, "*.umul");
8132 set_optab_libfunc (sdiv_optab, SImode, "*.div");
8133 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
8134 set_optab_libfunc (smod_optab, SImode, "*.rem");
8135 set_optab_libfunc (umod_optab, SImode, "*.urem");
8137 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
8138 set_optab_libfunc (add_optab, TFmode, "_Q_add");
8139 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
8140 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
8141 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
8142 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
8144 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
8145 is because with soft-float, the SFmode and DFmode sqrt
8146 instructions will be absent, and the compiler will notice and
8147 try to use the TFmode sqrt instruction for calls to the
8148 builtin function sqrt, but this fails. */
8150 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
8152 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
8153 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
8154 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
8155 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
8156 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
8157 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
8159 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
8160 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
8161 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
8162 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
8164 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
8165 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
8166 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
8167 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
8169 if (DITF_CONVERSION_LIBFUNCS)
8171 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
8172 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
8173 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
8174 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
8177 if (SUN_CONVERSION_LIBFUNCS)
8179 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
8180 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
8181 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
8182 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
8187 /* In the SPARC 64bit ABI, SImode multiply and divide functions
8188 do not exist in the library. Make sure the compiler does not
8189 emit calls to them by accident. (It should always use the
8190 hardware instructions.) */
8191 set_optab_libfunc (smul_optab, SImode, 0);
8192 set_optab_libfunc (sdiv_optab, SImode, 0);
8193 set_optab_libfunc (udiv_optab, SImode, 0);
8194 set_optab_libfunc (smod_optab, SImode, 0);
8195 set_optab_libfunc (umod_optab, SImode, 0);
8197 if (SUN_INTEGER_MULTIPLY_64)
8199 set_optab_libfunc (smul_optab, DImode, "__mul64");
8200 set_optab_libfunc (sdiv_optab, DImode, "__div64");
8201 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
8202 set_optab_libfunc (smod_optab, DImode, "__rem64");
8203 set_optab_libfunc (umod_optab, DImode, "__urem64");
8206 if (SUN_CONVERSION_LIBFUNCS)
8208 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
8209 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
8210 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
8211 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
8215 gofast_maybe_init_libfuncs ();
8218 #define def_builtin(NAME, CODE, TYPE) \
8219 add_builtin_function((NAME), (TYPE), (CODE), BUILT_IN_MD, NULL, \
8222 /* Implement the TARGET_INIT_BUILTINS target hook.
8223 Create builtin functions for special SPARC instructions. */
8226 sparc_init_builtins (void)
8229 sparc_vis_init_builtins ();
8232 /* Create builtin functions for VIS 1.0 instructions. */
8235 sparc_vis_init_builtins (void)
8237 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
8238 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
8239 tree v4hi = build_vector_type (intHI_type_node, 4);
8240 tree v2hi = build_vector_type (intHI_type_node, 2);
8241 tree v2si = build_vector_type (intSI_type_node, 2);
8243 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
8244 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
8245 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
8246 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
8247 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
8248 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
8249 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
8250 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
8251 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
8252 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
8253 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
8254 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
8255 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
8257 intDI_type_node, 0);
8258 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
8260 intDI_type_node, 0);
8261 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
8263 intSI_type_node, 0);
8264 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
8266 intDI_type_node, 0);
8268 /* Packing and expanding vectors. */
8269 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis, v4qi_ftype_v4hi);
8270 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
8271 v8qi_ftype_v2si_v8qi);
8272 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
8274 def_builtin ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis, v4hi_ftype_v4qi);
8275 def_builtin ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
8276 v8qi_ftype_v4qi_v4qi);
8278 /* Multiplications. */
8279 def_builtin ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
8280 v4hi_ftype_v4qi_v4hi);
8281 def_builtin ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
8282 v4hi_ftype_v4qi_v2hi);
8283 def_builtin ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
8284 v4hi_ftype_v4qi_v2hi);
8285 def_builtin ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
8286 v4hi_ftype_v8qi_v4hi);
8287 def_builtin ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
8288 v4hi_ftype_v8qi_v4hi);
8289 def_builtin ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
8290 v2si_ftype_v4qi_v2hi);
8291 def_builtin ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
8292 v2si_ftype_v4qi_v2hi);
8294 /* Data aligning. */
8295 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
8296 v4hi_ftype_v4hi_v4hi);
8297 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
8298 v8qi_ftype_v8qi_v8qi);
8299 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
8300 v2si_ftype_v2si_v2si);
8301 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatadi_vis,
8304 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
8307 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
8310 /* Pixel distance. */
8311 def_builtin ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
8312 di_ftype_v8qi_v8qi_di);
8315 /* Handle TARGET_EXPAND_BUILTIN target hook.
8316 Expand builtin functions for sparc intrinsics. */
8319 sparc_expand_builtin (tree exp, rtx target,
8320 rtx subtarget ATTRIBUTE_UNUSED,
8321 enum machine_mode tmode ATTRIBUTE_UNUSED,
8322 int ignore ATTRIBUTE_UNUSED)
8325 call_expr_arg_iterator iter;
8326 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
8327 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
8329 enum machine_mode mode[4];
8332 mode[0] = insn_data[icode].operand[0].mode;
8334 || GET_MODE (target) != mode[0]
8335 || ! (*insn_data[icode].operand[0].predicate) (target, mode[0]))
8336 op[0] = gen_reg_rtx (mode[0]);
8340 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
8343 mode[arg_count] = insn_data[icode].operand[arg_count].mode;
8344 op[arg_count] = expand_normal (arg);
8346 if (! (*insn_data[icode].operand[arg_count].predicate) (op[arg_count],
8348 op[arg_count] = copy_to_mode_reg (mode[arg_count], op[arg_count]);
8354 pat = GEN_FCN (icode) (op[0], op[1]);
8357 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
8360 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
8375 sparc_vis_mul8x16 (int e8, int e16)
8377 return (e8 * e16 + 128) / 256;
8380 /* Multiply the vector elements in ELTS0 to the elements in ELTS1 as specified
8381 by FNCODE. All of the elements in ELTS0 and ELTS1 lists must be integer
8382 constants. A tree list with the results of the multiplications is returned,
8383 and each element in the list is of INNER_TYPE. */
8386 sparc_handle_vis_mul8x16 (int fncode, tree inner_type, tree elts0, tree elts1)
8388 tree n_elts = NULL_TREE;
8393 case CODE_FOR_fmul8x16_vis:
8394 for (; elts0 && elts1;
8395 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8398 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8399 TREE_INT_CST_LOW (TREE_VALUE (elts1)));
8400 n_elts = tree_cons (NULL_TREE,
8401 build_int_cst (inner_type, val),
8406 case CODE_FOR_fmul8x16au_vis:
8407 scale = TREE_INT_CST_LOW (TREE_VALUE (elts1));
8409 for (; elts0; elts0 = TREE_CHAIN (elts0))
8412 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8414 n_elts = tree_cons (NULL_TREE,
8415 build_int_cst (inner_type, val),
8420 case CODE_FOR_fmul8x16al_vis:
8421 scale = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (elts1)));
8423 for (; elts0; elts0 = TREE_CHAIN (elts0))
8426 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8428 n_elts = tree_cons (NULL_TREE,
8429 build_int_cst (inner_type, val),
8438 return nreverse (n_elts);
8441 /* Handle TARGET_FOLD_BUILTIN target hook.
8442 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
8443 result of the function call is ignored. NULL_TREE is returned if the
8444 function could not be folded. */
8447 sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
8448 tree *args, bool ignore)
8450 tree arg0, arg1, arg2;
8451 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
8452 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
8455 && icode != CODE_FOR_alignaddrsi_vis
8456 && icode != CODE_FOR_alignaddrdi_vis)
8457 return fold_convert (rtype, integer_zero_node);
8461 case CODE_FOR_fexpand_vis:
8465 if (TREE_CODE (arg0) == VECTOR_CST)
8467 tree inner_type = TREE_TYPE (rtype);
8468 tree elts = TREE_VECTOR_CST_ELTS (arg0);
8469 tree n_elts = NULL_TREE;
8471 for (; elts; elts = TREE_CHAIN (elts))
8473 unsigned int val = TREE_INT_CST_LOW (TREE_VALUE (elts)) << 4;
8474 n_elts = tree_cons (NULL_TREE,
8475 build_int_cst (inner_type, val),
8478 return build_vector (rtype, nreverse (n_elts));
8482 case CODE_FOR_fmul8x16_vis:
8483 case CODE_FOR_fmul8x16au_vis:
8484 case CODE_FOR_fmul8x16al_vis:
8490 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
8492 tree inner_type = TREE_TYPE (rtype);
8493 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8494 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8495 tree n_elts = sparc_handle_vis_mul8x16 (icode, inner_type, elts0,
8498 return build_vector (rtype, n_elts);
8502 case CODE_FOR_fpmerge_vis:
8508 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
8510 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8511 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8512 tree n_elts = NULL_TREE;
8514 for (; elts0 && elts1;
8515 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8517 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts0), n_elts);
8518 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts1), n_elts);
8521 return build_vector (rtype, nreverse (n_elts));
8525 case CODE_FOR_pdist_vis:
8533 if (TREE_CODE (arg0) == VECTOR_CST
8534 && TREE_CODE (arg1) == VECTOR_CST
8535 && TREE_CODE (arg2) == INTEGER_CST)
8538 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
8539 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
8540 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8541 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8543 for (; elts0 && elts1;
8544 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8546 unsigned HOST_WIDE_INT
8547 low0 = TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8548 low1 = TREE_INT_CST_LOW (TREE_VALUE (elts1));
8549 HOST_WIDE_INT high0 = TREE_INT_CST_HIGH (TREE_VALUE (elts0));
8550 HOST_WIDE_INT high1 = TREE_INT_CST_HIGH (TREE_VALUE (elts1));
8552 unsigned HOST_WIDE_INT l;
8555 overflow |= neg_double (low1, high1, &l, &h);
8556 overflow |= add_double (low0, high0, l, h, &l, &h);
8558 overflow |= neg_double (l, h, &l, &h);
8560 overflow |= add_double (low, high, l, h, &low, &high);
8563 gcc_assert (overflow == 0);
8565 return build_int_cst_wide (rtype, low, high);
8575 /* ??? This duplicates information provided to the compiler by the
8576 ??? scheduler description. Some day, teach genautomata to output
8577 ??? the latencies and then CSE will just use that. */
8580 sparc_rtx_costs (rtx x, int code, int outer_code, int *total,
8581 bool speed ATTRIBUTE_UNUSED)
8583 enum machine_mode mode = GET_MODE (x);
8584 bool float_mode_p = FLOAT_MODE_P (mode);
8589 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
8607 if (GET_MODE (x) == VOIDmode
8608 && ((CONST_DOUBLE_HIGH (x) == 0
8609 && CONST_DOUBLE_LOW (x) < 0x1000)
8610 || (CONST_DOUBLE_HIGH (x) == -1
8611 && CONST_DOUBLE_LOW (x) < 0
8612 && CONST_DOUBLE_LOW (x) >= -0x1000)))
8619 /* If outer-code was a sign or zero extension, a cost
8620 of COSTS_N_INSNS (1) was already added in. This is
8621 why we are subtracting it back out. */
8622 if (outer_code == ZERO_EXTEND)
8624 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
8626 else if (outer_code == SIGN_EXTEND)
8628 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
8630 else if (float_mode_p)
8632 *total = sparc_costs->float_load;
8636 *total = sparc_costs->int_load;
8644 *total = sparc_costs->float_plusminus;
8646 *total = COSTS_N_INSNS (1);
8651 *total = sparc_costs->float_mul;
8652 else if (! TARGET_HARD_MUL)
8653 *total = COSTS_N_INSNS (25);
8659 if (sparc_costs->int_mul_bit_factor)
8663 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8665 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
8666 for (nbits = 0; value != 0; value &= value - 1)
8669 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
8670 && GET_MODE (XEXP (x, 1)) == VOIDmode)
8672 rtx x1 = XEXP (x, 1);
8673 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
8674 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
8676 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
8678 for (; value2 != 0; value2 &= value2 - 1)
8686 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
8687 bit_cost = COSTS_N_INSNS (bit_cost);
8691 *total = sparc_costs->int_mulX + bit_cost;
8693 *total = sparc_costs->int_mul + bit_cost;
8700 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
8710 *total = sparc_costs->float_div_df;
8712 *total = sparc_costs->float_div_sf;
8717 *total = sparc_costs->int_divX;
8719 *total = sparc_costs->int_div;
8726 *total = COSTS_N_INSNS (1);
8733 case UNSIGNED_FLOAT:
8737 case FLOAT_TRUNCATE:
8738 *total = sparc_costs->float_move;
8743 *total = sparc_costs->float_sqrt_df;
8745 *total = sparc_costs->float_sqrt_sf;
8750 *total = sparc_costs->float_cmp;
8752 *total = COSTS_N_INSNS (1);
8757 *total = sparc_costs->float_cmove;
8759 *total = sparc_costs->int_cmove;
8763 /* Handle the NAND vector patterns. */
8764 if (sparc_vector_mode_supported_p (GET_MODE (x))
8765 && GET_CODE (XEXP (x, 0)) == NOT
8766 && GET_CODE (XEXP (x, 1)) == NOT)
8768 *total = COSTS_N_INSNS (1);
8779 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
8780 This is achieved by means of a manual dynamic stack space allocation in
8781 the current frame. We make the assumption that SEQ doesn't contain any
8782 function calls, with the possible exception of calls to the PIC helper. */
8785 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
8787 /* We must preserve the lowest 16 words for the register save area. */
8788 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
8789 /* We really need only 2 words of fresh stack space. */
8790 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
8793 = gen_rtx_MEM (word_mode, plus_constant (stack_pointer_rtx,
8794 SPARC_STACK_BIAS + offset));
8796 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
8797 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
8799 emit_insn (gen_rtx_SET (VOIDmode,
8800 adjust_address (slot, word_mode, UNITS_PER_WORD),
8804 emit_insn (gen_rtx_SET (VOIDmode,
8806 adjust_address (slot, word_mode, UNITS_PER_WORD)));
8807 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
8808 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
8811 /* Output the assembler code for a thunk function. THUNK_DECL is the
8812 declaration for the thunk function itself, FUNCTION is the decl for
8813 the target function. DELTA is an immediate constant offset to be
8814 added to THIS. If VCALL_OFFSET is nonzero, the word at address
8815 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
8818 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
8819 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
8822 rtx this_rtx, insn, funexp;
8823 unsigned int int_arg_first;
8825 reload_completed = 1;
8826 epilogue_completed = 1;
8828 emit_note (NOTE_INSN_PROLOGUE_END);
8830 if (flag_delayed_branch)
8832 /* We will emit a regular sibcall below, so we need to instruct
8833 output_sibcall that we are in a leaf function. */
8834 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
8836 /* This will cause final.c to invoke leaf_renumber_regs so we
8837 must behave as if we were in a not-yet-leafified function. */
8838 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
8842 /* We will emit the sibcall manually below, so we will need to
8843 manually spill non-leaf registers. */
8844 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
8846 /* We really are in a leaf function. */
8847 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
8850 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
8851 returns a structure, the structure return pointer is there instead. */
8853 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
8854 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
8856 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
8858 /* Add DELTA. When possible use a plain add, otherwise load it into
8859 a register first. */
8862 rtx delta_rtx = GEN_INT (delta);
8864 if (! SPARC_SIMM13_P (delta))
8866 rtx scratch = gen_rtx_REG (Pmode, 1);
8867 emit_move_insn (scratch, delta_rtx);
8868 delta_rtx = scratch;
8871 /* THIS_RTX += DELTA. */
8872 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
8875 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
8878 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
8879 rtx scratch = gen_rtx_REG (Pmode, 1);
8881 gcc_assert (vcall_offset < 0);
8883 /* SCRATCH = *THIS_RTX. */
8884 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
8886 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
8887 may not have any available scratch register at this point. */
8888 if (SPARC_SIMM13_P (vcall_offset))
8890 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
8891 else if (! fixed_regs[5]
8892 /* The below sequence is made up of at least 2 insns,
8893 while the default method may need only one. */
8894 && vcall_offset < -8192)
8896 rtx scratch2 = gen_rtx_REG (Pmode, 5);
8897 emit_move_insn (scratch2, vcall_offset_rtx);
8898 vcall_offset_rtx = scratch2;
8902 rtx increment = GEN_INT (-4096);
8904 /* VCALL_OFFSET is a negative number whose typical range can be
8905 estimated as -32768..0 in 32-bit mode. In almost all cases
8906 it is therefore cheaper to emit multiple add insns than
8907 spilling and loading the constant into a register (at least
8909 while (! SPARC_SIMM13_P (vcall_offset))
8911 emit_insn (gen_add2_insn (scratch, increment));
8912 vcall_offset += 4096;
8914 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
8917 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
8918 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
8919 gen_rtx_PLUS (Pmode,
8921 vcall_offset_rtx)));
8923 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
8924 emit_insn (gen_add2_insn (this_rtx, scratch));
8927 /* Generate a tail call to the target function. */
8928 if (! TREE_USED (function))
8930 assemble_external (function);
8931 TREE_USED (function) = 1;
8933 funexp = XEXP (DECL_RTL (function), 0);
8935 if (flag_delayed_branch)
8937 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8938 insn = emit_call_insn (gen_sibcall (funexp));
8939 SIBLING_CALL_P (insn) = 1;
8943 /* The hoops we have to jump through in order to generate a sibcall
8944 without using delay slots... */
8945 rtx spill_reg, spill_reg2, seq, scratch = gen_rtx_REG (Pmode, 1);
8949 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
8950 spill_reg2 = gen_rtx_REG (word_mode, PIC_OFFSET_TABLE_REGNUM);
8952 /* Delay emitting the PIC helper function because it needs to
8953 change the section and we are emitting assembly code. */
8954 load_pic_register (); /* clobbers %o7 */
8955 scratch = legitimize_pic_address (funexp, scratch);
8958 emit_and_preserve (seq, spill_reg, spill_reg2);
8960 else if (TARGET_ARCH32)
8962 emit_insn (gen_rtx_SET (VOIDmode,
8964 gen_rtx_HIGH (SImode, funexp)));
8965 emit_insn (gen_rtx_SET (VOIDmode,
8967 gen_rtx_LO_SUM (SImode, scratch, funexp)));
8969 else /* TARGET_ARCH64 */
8971 switch (sparc_cmodel)
8975 /* The destination can serve as a temporary. */
8976 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
8981 /* The destination cannot serve as a temporary. */
8982 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
8984 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
8987 emit_and_preserve (seq, spill_reg, 0);
8995 emit_jump_insn (gen_indirect_jump (scratch));
9000 /* Run just enough of rest_of_compilation to get the insns emitted.
9001 There's not really enough bulk here to make other passes such as
9002 instruction scheduling worth while. Note that use_thunk calls
9003 assemble_start_function and assemble_end_function. */
9004 insn = get_insns ();
9005 insn_locators_alloc ();
9006 shorten_branches (insn);
9007 final_start_function (insn, file, 1);
9008 final (insn, file, 1);
9009 final_end_function ();
9011 reload_completed = 0;
9012 epilogue_completed = 0;
9015 /* Return true if sparc_output_mi_thunk would be able to output the
9016 assembler code for the thunk function specified by the arguments
9017 it is passed, and false otherwise. */
9019 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
9020 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
9021 HOST_WIDE_INT vcall_offset,
9022 const_tree function ATTRIBUTE_UNUSED)
9024 /* Bound the loop used in the default method above. */
9025 return (vcall_offset >= -32768 || ! fixed_regs[5]);
9028 /* How to allocate a 'struct machine_function'. */
9030 static struct machine_function *
9031 sparc_init_machine_status (void)
9033 return ggc_alloc_cleared_machine_function ();
9036 /* Locate some local-dynamic symbol still in use by this function
9037 so that we can print its name in local-dynamic base patterns. */
9040 get_some_local_dynamic_name (void)
9044 if (cfun->machine->some_ld_name)
9045 return cfun->machine->some_ld_name;
9047 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
9049 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
9050 return cfun->machine->some_ld_name;
9056 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
9061 && GET_CODE (x) == SYMBOL_REF
9062 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
9064 cfun->machine->some_ld_name = XSTR (x, 0);
9071 /* Handle the TARGET_DWARF_HANDLE_FRAME_UNSPEC hook.
9072 This is called from dwarf2out.c to emit call frame instructions
9073 for frame-related insns containing UNSPECs and UNSPEC_VOLATILEs. */
9075 sparc_dwarf_handle_frame_unspec (const char *label,
9076 rtx pattern ATTRIBUTE_UNUSED,
9077 int index ATTRIBUTE_UNUSED)
9079 gcc_assert (index == UNSPECV_SAVEW);
9080 dwarf2out_window_save (label);
9083 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
9084 We need to emit DTP-relative relocations. */
9087 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
9092 fputs ("\t.word\t%r_tls_dtpoff32(", file);
9095 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
9100 output_addr_const (file, x);
9104 /* Do whatever processing is required at the end of a file. */
9107 sparc_file_end (void)
9109 /* If need to emit the special PIC helper function, do so now. */
9110 if (pic_helper_needed)
9112 unsigned int regno = REGNO (pic_offset_table_rtx);
9113 const char *pic_name = reg_names[regno];
9115 #ifdef DWARF2_UNWIND_INFO
9119 get_pc_thunk_name (name, regno);
9120 if (USE_HIDDEN_LINKONCE)
9122 tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL,
9123 get_identifier (name),
9124 build_function_type (void_type_node,
9126 DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL,
9127 NULL_TREE, void_type_node);
9128 TREE_STATIC (decl) = 1;
9129 make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
9130 DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN;
9131 DECL_VISIBILITY_SPECIFIED (decl) = 1;
9132 allocate_struct_function (decl, true);
9133 current_function_decl = decl;
9134 init_varasm_status ();
9135 assemble_start_function (decl, name);
9139 const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9140 switch_to_section (text_section);
9142 ASM_OUTPUT_ALIGN (asm_out_file, align);
9143 ASM_OUTPUT_LABEL (asm_out_file, name);
9146 #ifdef DWARF2_UNWIND_INFO
9147 do_cfi = dwarf2out_do_cfi_asm ();
9149 fprintf (asm_out_file, "\t.cfi_startproc\n");
9151 if (flag_delayed_branch)
9152 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
9153 pic_name, pic_name);
9155 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
9156 pic_name, pic_name);
9157 #ifdef DWARF2_UNWIND_INFO
9159 fprintf (asm_out_file, "\t.cfi_endproc\n");
9163 if (NEED_INDICATE_EXEC_STACK)
9164 file_end_indicate_exec_stack ();
9167 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
9168 /* Implement TARGET_MANGLE_TYPE. */
9171 sparc_mangle_type (const_tree type)
9174 && TYPE_MAIN_VARIANT (type) == long_double_type_node
9175 && TARGET_LONG_DOUBLE_128)
9178 /* For all other types, use normal C++ mangling. */
9183 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
9184 compare and swap on the word containing the byte or half-word. */
9187 sparc_expand_compare_and_swap_12 (rtx result, rtx mem, rtx oldval, rtx newval)
9189 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
9190 rtx addr = gen_reg_rtx (Pmode);
9191 rtx off = gen_reg_rtx (SImode);
9192 rtx oldv = gen_reg_rtx (SImode);
9193 rtx newv = gen_reg_rtx (SImode);
9194 rtx oldvalue = gen_reg_rtx (SImode);
9195 rtx newvalue = gen_reg_rtx (SImode);
9196 rtx res = gen_reg_rtx (SImode);
9197 rtx resv = gen_reg_rtx (SImode);
9198 rtx memsi, val, mask, end_label, loop_label, cc;
9200 emit_insn (gen_rtx_SET (VOIDmode, addr,
9201 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
9203 if (Pmode != SImode)
9204 addr1 = gen_lowpart (SImode, addr1);
9205 emit_insn (gen_rtx_SET (VOIDmode, off,
9206 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
9208 memsi = gen_rtx_MEM (SImode, addr);
9209 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
9210 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
9212 val = force_reg (SImode, memsi);
9214 emit_insn (gen_rtx_SET (VOIDmode, off,
9215 gen_rtx_XOR (SImode, off,
9216 GEN_INT (GET_MODE (mem) == QImode
9219 emit_insn (gen_rtx_SET (VOIDmode, off,
9220 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
9222 if (GET_MODE (mem) == QImode)
9223 mask = force_reg (SImode, GEN_INT (0xff));
9225 mask = force_reg (SImode, GEN_INT (0xffff));
9227 emit_insn (gen_rtx_SET (VOIDmode, mask,
9228 gen_rtx_ASHIFT (SImode, mask, off)));
9230 emit_insn (gen_rtx_SET (VOIDmode, val,
9231 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
9234 oldval = gen_lowpart (SImode, oldval);
9235 emit_insn (gen_rtx_SET (VOIDmode, oldv,
9236 gen_rtx_ASHIFT (SImode, oldval, off)));
9238 newval = gen_lowpart_common (SImode, newval);
9239 emit_insn (gen_rtx_SET (VOIDmode, newv,
9240 gen_rtx_ASHIFT (SImode, newval, off)));
9242 emit_insn (gen_rtx_SET (VOIDmode, oldv,
9243 gen_rtx_AND (SImode, oldv, mask)));
9245 emit_insn (gen_rtx_SET (VOIDmode, newv,
9246 gen_rtx_AND (SImode, newv, mask)));
9248 end_label = gen_label_rtx ();
9249 loop_label = gen_label_rtx ();
9250 emit_label (loop_label);
9252 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
9253 gen_rtx_IOR (SImode, oldv, val)));
9255 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
9256 gen_rtx_IOR (SImode, newv, val)));
9258 emit_insn (gen_sync_compare_and_swapsi (res, memsi, oldvalue, newvalue));
9260 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
9262 emit_insn (gen_rtx_SET (VOIDmode, resv,
9263 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
9266 cc = gen_compare_reg_1 (NE, resv, val);
9267 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
9269 /* Use cbranchcc4 to separate the compare and branch! */
9270 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx),
9271 cc, const0_rtx, loop_label));
9273 emit_label (end_label);
9275 emit_insn (gen_rtx_SET (VOIDmode, res,
9276 gen_rtx_AND (SImode, res, mask)));
9278 emit_insn (gen_rtx_SET (VOIDmode, res,
9279 gen_rtx_LSHIFTRT (SImode, res, off)));
9281 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
9284 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
9287 sparc_frame_pointer_required (void)
9289 return !(leaf_function_p () && only_leaf_regs_used ());
9292 /* The way this is structured, we can't eliminate SFP in favor of SP
9293 if the frame pointer is required: we want to use the SFP->HFP elimination
9294 in that case. But the test in update_eliminables doesn't know we are
9295 assuming below that we only do the former elimination. */
9298 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
9300 return (to == HARD_FRAME_POINTER_REGNUM
9301 || !targetm.frame_pointer_required ());
9304 #include "gt-sparc.h"