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
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"
34 #include "insn-config.h"
35 #include "insn-codes.h"
36 #include "conditions.h"
38 #include "insn-attr.h"
49 #include "target-def.h"
50 #include "cfglayout.h"
52 #include "langhooks.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 /* Save the operands last given to a compare for use when we
284 generate a scc or bcc insn. */
285 rtx sparc_compare_op0, sparc_compare_op1;
287 /* Vector to say how input registers are mapped to output registers.
288 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
289 eliminate it. You must use -fomit-frame-pointer to get that. */
290 char leaf_reg_remap[] =
291 { 0, 1, 2, 3, 4, 5, 6, 7,
292 -1, -1, -1, -1, -1, -1, 14, -1,
293 -1, -1, -1, -1, -1, -1, -1, -1,
294 8, 9, 10, 11, 12, 13, -1, 15,
296 32, 33, 34, 35, 36, 37, 38, 39,
297 40, 41, 42, 43, 44, 45, 46, 47,
298 48, 49, 50, 51, 52, 53, 54, 55,
299 56, 57, 58, 59, 60, 61, 62, 63,
300 64, 65, 66, 67, 68, 69, 70, 71,
301 72, 73, 74, 75, 76, 77, 78, 79,
302 80, 81, 82, 83, 84, 85, 86, 87,
303 88, 89, 90, 91, 92, 93, 94, 95,
304 96, 97, 98, 99, 100};
306 /* Vector, indexed by hard register number, which contains 1
307 for a register that is allowable in a candidate for leaf
308 function treatment. */
309 char sparc_leaf_regs[] =
310 { 1, 1, 1, 1, 1, 1, 1, 1,
311 0, 0, 0, 0, 0, 0, 1, 0,
312 0, 0, 0, 0, 0, 0, 0, 0,
313 1, 1, 1, 1, 1, 1, 0, 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,
318 1, 1, 1, 1, 1, 1, 1, 1,
319 1, 1, 1, 1, 1, 1, 1, 1,
320 1, 1, 1, 1, 1, 1, 1, 1,
321 1, 1, 1, 1, 1, 1, 1, 1,
324 struct GTY(()) machine_function
326 /* Some local-dynamic TLS symbol name. */
327 const char *some_ld_name;
329 /* True if the current function is leaf and uses only leaf regs,
330 so that the SPARC leaf function optimization can be applied.
331 Private version of current_function_uses_only_leaf_regs, see
332 sparc_expand_prologue for the rationale. */
335 /* True if the data calculated by sparc_expand_prologue are valid. */
336 bool prologue_data_valid_p;
339 #define sparc_leaf_function_p cfun->machine->leaf_function_p
340 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
342 /* Register we pretend to think the frame pointer is allocated to.
343 Normally, this is %fp, but if we are in a leaf procedure, this
344 is %sp+"something". We record "something" separately as it may
345 be too big for reg+constant addressing. */
346 static rtx frame_base_reg;
347 static HOST_WIDE_INT frame_base_offset;
349 /* 1 if the next opcode is to be specially indented. */
350 int sparc_indent_opcode = 0;
352 static bool sparc_handle_option (size_t, const char *, int);
353 static void sparc_init_modes (void);
354 static void scan_record_type (tree, int *, int *, int *);
355 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
356 tree, int, int, int *, int *);
358 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
359 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
361 static void sparc_output_addr_vec (rtx);
362 static void sparc_output_addr_diff_vec (rtx);
363 static void sparc_output_deferred_case_vectors (void);
364 static rtx sparc_builtin_saveregs (void);
365 static int epilogue_renumber (rtx *, int);
366 static bool sparc_assemble_integer (rtx, unsigned int, int);
367 static int set_extends (rtx);
368 static void emit_pic_helper (void);
369 static void load_pic_register (bool);
370 static int save_or_restore_regs (int, int, rtx, int, int);
371 static void emit_save_or_restore_regs (int);
372 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
373 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
374 #ifdef OBJECT_FORMAT_ELF
375 static void sparc_elf_asm_named_section (const char *, unsigned int, tree);
378 static int sparc_adjust_cost (rtx, rtx, rtx, int);
379 static int sparc_issue_rate (void);
380 static void sparc_sched_init (FILE *, int, int);
381 static int sparc_use_sched_lookahead (void);
383 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
384 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
385 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
386 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
387 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
389 static bool sparc_function_ok_for_sibcall (tree, tree);
390 static void sparc_init_libfuncs (void);
391 static void sparc_init_builtins (void);
392 static void sparc_vis_init_builtins (void);
393 static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
394 static tree sparc_fold_builtin (tree, tree, bool);
395 static int sparc_vis_mul8x16 (int, int);
396 static tree sparc_handle_vis_mul8x16 (int, tree, tree, tree);
397 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
398 HOST_WIDE_INT, tree);
399 static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT,
400 HOST_WIDE_INT, const_tree);
401 static struct machine_function * sparc_init_machine_status (void);
402 static bool sparc_cannot_force_const_mem (rtx);
403 static rtx sparc_tls_get_addr (void);
404 static rtx sparc_tls_got (void);
405 static const char *get_some_local_dynamic_name (void);
406 static int get_some_local_dynamic_name_1 (rtx *, void *);
407 static bool sparc_rtx_costs (rtx, int, int, int *, bool);
408 static bool sparc_promote_prototypes (const_tree);
409 static rtx sparc_struct_value_rtx (tree, int);
410 static bool sparc_return_in_memory (const_tree, const_tree);
411 static bool sparc_strict_argument_naming (CUMULATIVE_ARGS *);
412 static void sparc_va_start (tree, rtx);
413 static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
414 static bool sparc_vector_mode_supported_p (enum machine_mode);
415 static bool sparc_pass_by_reference (CUMULATIVE_ARGS *,
416 enum machine_mode, const_tree, bool);
417 static int sparc_arg_partial_bytes (CUMULATIVE_ARGS *,
418 enum machine_mode, tree, bool);
419 static void sparc_dwarf_handle_frame_unspec (const char *, rtx, int);
420 static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
421 static void sparc_file_end (void);
422 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
423 static const char *sparc_mangle_type (const_tree);
425 #ifdef SUBTARGET_ATTRIBUTE_TABLE
426 const struct attribute_spec sparc_attribute_table[];
429 /* Option handling. */
432 enum cmodel sparc_cmodel;
434 char sparc_hard_reg_printed[8];
436 struct sparc_cpu_select sparc_select[] =
438 /* switch name, tune arch */
439 { (char *)0, "default", 1, 1 },
440 { (char *)0, "-mcpu=", 1, 1 },
441 { (char *)0, "-mtune=", 1, 0 },
445 /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */
446 enum processor_type sparc_cpu;
448 /* Whether
\fan FPU option was specified. */
449 static bool fpu_option_set = false;
451 /* Initialize the GCC target structure. */
453 /* The sparc default is to use .half rather than .short for aligned
454 HI objects. Use .word instead of .long on non-ELF systems. */
455 #undef TARGET_ASM_ALIGNED_HI_OP
456 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
457 #ifndef OBJECT_FORMAT_ELF
458 #undef TARGET_ASM_ALIGNED_SI_OP
459 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
462 #undef TARGET_ASM_UNALIGNED_HI_OP
463 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
464 #undef TARGET_ASM_UNALIGNED_SI_OP
465 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
466 #undef TARGET_ASM_UNALIGNED_DI_OP
467 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
469 /* The target hook has to handle DI-mode values. */
470 #undef TARGET_ASM_INTEGER
471 #define TARGET_ASM_INTEGER sparc_assemble_integer
473 #undef TARGET_ASM_FUNCTION_PROLOGUE
474 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
475 #undef TARGET_ASM_FUNCTION_EPILOGUE
476 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
478 #undef TARGET_SCHED_ADJUST_COST
479 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
480 #undef TARGET_SCHED_ISSUE_RATE
481 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
482 #undef TARGET_SCHED_INIT
483 #define TARGET_SCHED_INIT sparc_sched_init
484 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
485 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
487 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
488 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
490 #undef TARGET_INIT_LIBFUNCS
491 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
492 #undef TARGET_INIT_BUILTINS
493 #define TARGET_INIT_BUILTINS sparc_init_builtins
495 #undef TARGET_EXPAND_BUILTIN
496 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
497 #undef TARGET_FOLD_BUILTIN
498 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
501 #undef TARGET_HAVE_TLS
502 #define TARGET_HAVE_TLS true
505 #undef TARGET_CANNOT_FORCE_CONST_MEM
506 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
508 #undef TARGET_ASM_OUTPUT_MI_THUNK
509 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
510 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
511 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
513 #undef TARGET_RTX_COSTS
514 #define TARGET_RTX_COSTS sparc_rtx_costs
515 #undef TARGET_ADDRESS_COST
516 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0
518 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
519 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
520 test for this value. */
521 #undef TARGET_PROMOTE_FUNCTION_ARGS
522 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
524 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
525 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
526 test for this value. */
527 #undef TARGET_PROMOTE_FUNCTION_RETURN
528 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
530 #undef TARGET_PROMOTE_PROTOTYPES
531 #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes
533 #undef TARGET_STRUCT_VALUE_RTX
534 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
535 #undef TARGET_RETURN_IN_MEMORY
536 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
537 #undef TARGET_MUST_PASS_IN_STACK
538 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
539 #undef TARGET_PASS_BY_REFERENCE
540 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
541 #undef TARGET_ARG_PARTIAL_BYTES
542 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
544 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
545 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
546 #undef TARGET_STRICT_ARGUMENT_NAMING
547 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
549 #undef TARGET_EXPAND_BUILTIN_VA_START
550 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
551 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
552 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
554 #undef TARGET_VECTOR_MODE_SUPPORTED_P
555 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
557 #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC
558 #define TARGET_DWARF_HANDLE_FRAME_UNSPEC sparc_dwarf_handle_frame_unspec
560 #ifdef SUBTARGET_INSERT_ATTRIBUTES
561 #undef TARGET_INSERT_ATTRIBUTES
562 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
565 #ifdef SUBTARGET_ATTRIBUTE_TABLE
566 #undef TARGET_ATTRIBUTE_TABLE
567 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
570 #undef TARGET_RELAXED_ORDERING
571 #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING
573 #undef TARGET_DEFAULT_TARGET_FLAGS
574 #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT
575 #undef TARGET_HANDLE_OPTION
576 #define TARGET_HANDLE_OPTION sparc_handle_option
578 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
579 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
580 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
583 #undef TARGET_ASM_FILE_END
584 #define TARGET_ASM_FILE_END sparc_file_end
586 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
587 #undef TARGET_MANGLE_TYPE
588 #define TARGET_MANGLE_TYPE sparc_mangle_type
591 struct gcc_target targetm = TARGET_INITIALIZER;
593 /* Implement TARGET_HANDLE_OPTION. */
596 sparc_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED)
601 case OPT_mhard_float:
602 case OPT_msoft_float:
603 fpu_option_set = true;
607 sparc_select[1].string = arg;
611 sparc_select[2].string = arg;
618 /* Validate and override various options, and do some machine dependent
622 sparc_override_options (void)
624 static struct code_model {
625 const char *const name;
627 } const cmodels[] = {
629 { "medlow", CM_MEDLOW },
630 { "medmid", CM_MEDMID },
631 { "medany", CM_MEDANY },
632 { "embmedany", CM_EMBMEDANY },
635 const struct code_model *cmodel;
636 /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */
637 static struct cpu_default {
639 const char *const name;
640 } const cpu_default[] = {
641 /* There must be one entry here for each TARGET_CPU value. */
642 { TARGET_CPU_sparc, "cypress" },
643 { TARGET_CPU_sparclet, "tsc701" },
644 { TARGET_CPU_sparclite, "f930" },
645 { TARGET_CPU_v8, "v8" },
646 { TARGET_CPU_hypersparc, "hypersparc" },
647 { TARGET_CPU_sparclite86x, "sparclite86x" },
648 { TARGET_CPU_supersparc, "supersparc" },
649 { TARGET_CPU_v9, "v9" },
650 { TARGET_CPU_ultrasparc, "ultrasparc" },
651 { TARGET_CPU_ultrasparc3, "ultrasparc3" },
652 { TARGET_CPU_niagara, "niagara" },
653 { TARGET_CPU_niagara2, "niagara2" },
656 const struct cpu_default *def;
657 /* Table of values for -m{cpu,tune}=. */
658 static struct cpu_table {
659 const char *const name;
660 const enum processor_type processor;
663 } const cpu_table[] = {
664 { "v7", PROCESSOR_V7, MASK_ISA, 0 },
665 { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 },
666 { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 },
667 /* TI TMS390Z55 supersparc */
668 { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 },
669 { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE },
670 /* The Fujitsu MB86930 is the original sparclite chip, with no fpu.
671 The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */
672 { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE },
673 { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU },
674 { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU },
675 { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU,
677 { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET },
679 { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET },
680 { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 },
681 /* TI ultrasparc I, II, IIi */
682 { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9
683 /* Although insns using %y are deprecated, it is a clear win on current
685 |MASK_DEPRECATED_V8_INSNS},
686 /* TI ultrasparc III */
687 /* ??? Check if %y issue still holds true in ultra3. */
688 { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
690 { "niagara", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
691 { "niagara2", PROCESSOR_NIAGARA, MASK_ISA, MASK_V9},
694 const struct cpu_table *cpu;
695 const struct sparc_cpu_select *sel;
698 #ifndef SPARC_BI_ARCH
699 /* Check for unsupported architecture size. */
700 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
701 error ("%s is not supported by this configuration",
702 DEFAULT_ARCH32_P ? "-m64" : "-m32");
705 /* We force all 64bit archs to use 128 bit long double */
706 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
708 error ("-mlong-double-64 not allowed with -m64");
709 target_flags |= MASK_LONG_DOUBLE_128;
712 /* Code model selection. */
713 sparc_cmodel = SPARC_DEFAULT_CMODEL;
717 sparc_cmodel = CM_32;
720 if (sparc_cmodel_string != NULL)
724 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
725 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
727 if (cmodel->name == NULL)
728 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
730 sparc_cmodel = cmodel->value;
733 error ("-mcmodel= is not supported on 32 bit systems");
736 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
738 /* Set the default CPU. */
739 for (def = &cpu_default[0]; def->name; ++def)
740 if (def->cpu == TARGET_CPU_DEFAULT)
742 gcc_assert (def->name);
743 sparc_select[0].string = def->name;
745 for (sel = &sparc_select[0]; sel->name; ++sel)
749 for (cpu = &cpu_table[0]; cpu->name; ++cpu)
750 if (! strcmp (sel->string, cpu->name))
753 sparc_cpu = cpu->processor;
757 target_flags &= ~cpu->disable;
758 target_flags |= cpu->enable;
764 error ("bad value (%s) for %s switch", sel->string, sel->name);
768 /* If -mfpu or -mno-fpu was explicitly used, don't override with
769 the processor default. */
771 target_flags = (target_flags & ~MASK_FPU) | fpu;
773 /* Don't allow -mvis if FPU is disabled. */
775 target_flags &= ~MASK_VIS;
777 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
779 -m64 also implies v9. */
780 if (TARGET_VIS || TARGET_ARCH64)
782 target_flags |= MASK_V9;
783 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
786 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
787 if (TARGET_V9 && TARGET_ARCH32)
788 target_flags |= MASK_DEPRECATED_V8_INSNS;
790 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
791 if (! TARGET_V9 || TARGET_ARCH64)
792 target_flags &= ~MASK_V8PLUS;
794 /* Don't use stack biasing in 32 bit mode. */
796 target_flags &= ~MASK_STACK_BIAS;
798 /* Supply a default value for align_functions. */
799 if (align_functions == 0
800 && (sparc_cpu == PROCESSOR_ULTRASPARC
801 || sparc_cpu == PROCESSOR_ULTRASPARC3
802 || sparc_cpu == PROCESSOR_NIAGARA
803 || sparc_cpu == PROCESSOR_NIAGARA2))
804 align_functions = 32;
806 /* Validate PCC_STRUCT_RETURN. */
807 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
808 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
810 /* Only use .uaxword when compiling for a 64-bit target. */
812 targetm.asm_out.unaligned_op.di = NULL;
814 /* Do various machine dependent initializations. */
817 /* Acquire unique alias sets for our private stuff. */
818 sparc_sr_alias_set = new_alias_set ();
819 struct_value_alias_set = new_alias_set ();
821 /* Set up function hooks. */
822 init_machine_status = sparc_init_machine_status;
827 case PROCESSOR_CYPRESS:
828 sparc_costs = &cypress_costs;
831 case PROCESSOR_SPARCLITE:
832 case PROCESSOR_SUPERSPARC:
833 sparc_costs = &supersparc_costs;
837 case PROCESSOR_HYPERSPARC:
838 case PROCESSOR_SPARCLITE86X:
839 sparc_costs = &hypersparc_costs;
841 case PROCESSOR_SPARCLET:
842 case PROCESSOR_TSC701:
843 sparc_costs = &sparclet_costs;
846 case PROCESSOR_ULTRASPARC:
847 sparc_costs = &ultrasparc_costs;
849 case PROCESSOR_ULTRASPARC3:
850 sparc_costs = &ultrasparc3_costs;
852 case PROCESSOR_NIAGARA:
853 sparc_costs = &niagara_costs;
855 case PROCESSOR_NIAGARA2:
856 sparc_costs = &niagara2_costs;
860 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
861 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
862 target_flags |= MASK_LONG_DOUBLE_128;
865 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES))
866 set_param_value ("simultaneous-prefetches",
867 ((sparc_cpu == PROCESSOR_ULTRASPARC
868 || sparc_cpu == PROCESSOR_NIAGARA
869 || sparc_cpu == PROCESSOR_NIAGARA2)
871 : (sparc_cpu == PROCESSOR_ULTRASPARC3
873 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE))
874 set_param_value ("l1-cache-line-size",
875 ((sparc_cpu == PROCESSOR_ULTRASPARC
876 || sparc_cpu == PROCESSOR_ULTRASPARC3
877 || sparc_cpu == PROCESSOR_NIAGARA
878 || sparc_cpu == PROCESSOR_NIAGARA2)
882 #ifdef SUBTARGET_ATTRIBUTE_TABLE
883 /* Table of valid machine attributes. */
884 const struct attribute_spec sparc_attribute_table[] =
886 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
887 SUBTARGET_ATTRIBUTE_TABLE,
888 { NULL, 0, 0, false, false, false, NULL }
892 /* Miscellaneous utilities. */
894 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
895 or branch on register contents instructions. */
898 v9_regcmp_p (enum rtx_code code)
900 return (code == EQ || code == NE || code == GE || code == LT
901 || code == LE || code == GT);
904 /* Nonzero if OP is a floating point constant which can
905 be loaded into an integer register using a single
906 sethi instruction. */
911 if (GET_CODE (op) == CONST_DOUBLE)
916 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
917 REAL_VALUE_TO_TARGET_SINGLE (r, i);
918 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
924 /* Nonzero if OP is a floating point constant which can
925 be loaded into an integer register using a single
931 if (GET_CODE (op) == CONST_DOUBLE)
936 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
937 REAL_VALUE_TO_TARGET_SINGLE (r, i);
938 return SPARC_SIMM13_P (i);
944 /* Nonzero if OP is a floating point constant which can
945 be loaded into an integer register using a high/losum
946 instruction sequence. */
949 fp_high_losum_p (rtx op)
951 /* The constraints calling this should only be in
952 SFmode move insns, so any constant which cannot
953 be moved using a single insn will do. */
954 if (GET_CODE (op) == CONST_DOUBLE)
959 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
960 REAL_VALUE_TO_TARGET_SINGLE (r, i);
961 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
967 /* Expand a move instruction. Return true if all work is done. */
970 sparc_expand_move (enum machine_mode mode, rtx *operands)
972 /* Handle sets of MEM first. */
973 if (GET_CODE (operands[0]) == MEM)
975 /* 0 is a register (or a pair of registers) on SPARC. */
976 if (register_or_zero_operand (operands[1], mode))
979 if (!reload_in_progress)
981 operands[0] = validize_mem (operands[0]);
982 operands[1] = force_reg (mode, operands[1]);
986 /* Fixup TLS cases. */
988 && CONSTANT_P (operands[1])
989 && GET_CODE (operands[1]) != HIGH
990 && sparc_tls_referenced_p (operands [1]))
992 rtx sym = operands[1];
995 if (GET_CODE (sym) == CONST && GET_CODE (XEXP (sym, 0)) == PLUS)
997 addend = XEXP (XEXP (sym, 0), 1);
998 sym = XEXP (XEXP (sym, 0), 0);
1001 gcc_assert (SPARC_SYMBOL_REF_TLS_P (sym));
1003 sym = legitimize_tls_address (sym);
1006 sym = gen_rtx_PLUS (mode, sym, addend);
1007 sym = force_operand (sym, operands[0]);
1012 /* Fixup PIC cases. */
1013 if (flag_pic && CONSTANT_P (operands[1]))
1015 if (pic_address_needs_scratch (operands[1]))
1016 operands[1] = legitimize_pic_address (operands[1], mode, 0);
1018 /* VxWorks does not impose a fixed gap between segments; the run-time
1019 gap can be different from the object-file gap. We therefore can't
1020 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1021 are absolutely sure that X is in the same segment as the GOT.
1022 Unfortunately, the flexibility of linker scripts means that we
1023 can't be sure of that in general, so assume that _G_O_T_-relative
1024 accesses are never valid on VxWorks. */
1025 if (GET_CODE (operands[1]) == LABEL_REF && !TARGET_VXWORKS_RTP)
1029 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1035 gcc_assert (TARGET_ARCH64);
1036 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1041 if (symbolic_operand (operands[1], mode))
1043 operands[1] = legitimize_pic_address (operands[1],
1045 (reload_in_progress ?
1052 /* If we are trying to toss an integer constant into FP registers,
1053 or loading a FP or vector constant, force it into memory. */
1054 if (CONSTANT_P (operands[1])
1055 && REG_P (operands[0])
1056 && (SPARC_FP_REG_P (REGNO (operands[0]))
1057 || SCALAR_FLOAT_MODE_P (mode)
1058 || VECTOR_MODE_P (mode)))
1060 /* emit_group_store will send such bogosity to us when it is
1061 not storing directly into memory. So fix this up to avoid
1062 crashes in output_constant_pool. */
1063 if (operands [1] == const0_rtx)
1064 operands[1] = CONST0_RTX (mode);
1066 /* We can clear FP registers if TARGET_VIS, and always other regs. */
1067 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1068 && const_zero_operand (operands[1], mode))
1071 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1072 /* We are able to build any SF constant in integer registers
1073 with at most 2 instructions. */
1075 /* And any DF constant in integer registers. */
1077 && (reload_completed || reload_in_progress))))
1080 operands[1] = force_const_mem (mode, operands[1]);
1081 if (!reload_in_progress)
1082 operands[1] = validize_mem (operands[1]);
1086 /* Accept non-constants and valid constants unmodified. */
1087 if (!CONSTANT_P (operands[1])
1088 || GET_CODE (operands[1]) == HIGH
1089 || input_operand (operands[1], mode))
1095 /* All QImode constants require only one insn, so proceed. */
1100 sparc_emit_set_const32 (operands[0], operands[1]);
1104 /* input_operand should have filtered out 32-bit mode. */
1105 sparc_emit_set_const64 (operands[0], operands[1]);
1115 /* Load OP1, a 32-bit constant, into OP0, a register.
1116 We know it can't be done in one insn when we get
1117 here, the move expander guarantees this. */
1120 sparc_emit_set_const32 (rtx op0, rtx op1)
1122 enum machine_mode mode = GET_MODE (op0);
1125 if (reload_in_progress || reload_completed)
1128 temp = gen_reg_rtx (mode);
1130 if (GET_CODE (op1) == CONST_INT)
1132 gcc_assert (!small_int_operand (op1, mode)
1133 && !const_high_operand (op1, mode));
1135 /* Emit them as real moves instead of a HIGH/LO_SUM,
1136 this way CSE can see everything and reuse intermediate
1137 values if it wants. */
1138 emit_insn (gen_rtx_SET (VOIDmode, temp,
1139 GEN_INT (INTVAL (op1)
1140 & ~(HOST_WIDE_INT)0x3ff)));
1142 emit_insn (gen_rtx_SET (VOIDmode,
1144 gen_rtx_IOR (mode, temp,
1145 GEN_INT (INTVAL (op1) & 0x3ff))));
1149 /* A symbol, emit in the traditional way. */
1150 emit_insn (gen_rtx_SET (VOIDmode, temp,
1151 gen_rtx_HIGH (mode, op1)));
1152 emit_insn (gen_rtx_SET (VOIDmode,
1153 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1157 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1158 If TEMP is nonzero, we are forbidden to use any other scratch
1159 registers. Otherwise, we are allowed to generate them as needed.
1161 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1162 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1165 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1167 rtx temp1, temp2, temp3, temp4, temp5;
1170 if (temp && GET_MODE (temp) == TImode)
1173 temp = gen_rtx_REG (DImode, REGNO (temp));
1176 /* SPARC-V9 code-model support. */
1177 switch (sparc_cmodel)
1180 /* The range spanned by all instructions in the object is less
1181 than 2^31 bytes (2GB) and the distance from any instruction
1182 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1183 than 2^31 bytes (2GB).
1185 The executable must be in the low 4TB of the virtual address
1188 sethi %hi(symbol), %temp1
1189 or %temp1, %lo(symbol), %reg */
1191 temp1 = temp; /* op0 is allowed. */
1193 temp1 = gen_reg_rtx (DImode);
1195 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1196 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1200 /* The range spanned by all instructions in the object is less
1201 than 2^31 bytes (2GB) and the distance from any instruction
1202 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1203 than 2^31 bytes (2GB).
1205 The executable must be in the low 16TB of the virtual address
1208 sethi %h44(symbol), %temp1
1209 or %temp1, %m44(symbol), %temp2
1210 sllx %temp2, 12, %temp3
1211 or %temp3, %l44(symbol), %reg */
1216 temp3 = temp; /* op0 is allowed. */
1220 temp1 = gen_reg_rtx (DImode);
1221 temp2 = gen_reg_rtx (DImode);
1222 temp3 = gen_reg_rtx (DImode);
1225 emit_insn (gen_seth44 (temp1, op1));
1226 emit_insn (gen_setm44 (temp2, temp1, op1));
1227 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1228 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1229 emit_insn (gen_setl44 (op0, temp3, op1));
1233 /* The range spanned by all instructions in the object is less
1234 than 2^31 bytes (2GB) and the distance from any instruction
1235 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1236 than 2^31 bytes (2GB).
1238 The executable can be placed anywhere in the virtual address
1241 sethi %hh(symbol), %temp1
1242 sethi %lm(symbol), %temp2
1243 or %temp1, %hm(symbol), %temp3
1244 sllx %temp3, 32, %temp4
1245 or %temp4, %temp2, %temp5
1246 or %temp5, %lo(symbol), %reg */
1249 /* It is possible that one of the registers we got for operands[2]
1250 might coincide with that of operands[0] (which is why we made
1251 it TImode). Pick the other one to use as our scratch. */
1252 if (rtx_equal_p (temp, op0))
1254 gcc_assert (ti_temp);
1255 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1258 temp2 = temp; /* op0 is _not_ allowed, see above. */
1265 temp1 = gen_reg_rtx (DImode);
1266 temp2 = gen_reg_rtx (DImode);
1267 temp3 = gen_reg_rtx (DImode);
1268 temp4 = gen_reg_rtx (DImode);
1269 temp5 = gen_reg_rtx (DImode);
1272 emit_insn (gen_sethh (temp1, op1));
1273 emit_insn (gen_setlm (temp2, op1));
1274 emit_insn (gen_sethm (temp3, temp1, op1));
1275 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1276 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1277 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1278 gen_rtx_PLUS (DImode, temp4, temp2)));
1279 emit_insn (gen_setlo (op0, temp5, op1));
1283 /* Old old old backwards compatibility kruft here.
1284 Essentially it is MEDLOW with a fixed 64-bit
1285 virtual base added to all data segment addresses.
1286 Text-segment stuff is computed like MEDANY, we can't
1287 reuse the code above because the relocation knobs
1290 Data segment: sethi %hi(symbol), %temp1
1291 add %temp1, EMBMEDANY_BASE_REG, %temp2
1292 or %temp2, %lo(symbol), %reg */
1293 if (data_segment_operand (op1, GET_MODE (op1)))
1297 temp1 = temp; /* op0 is allowed. */
1302 temp1 = gen_reg_rtx (DImode);
1303 temp2 = gen_reg_rtx (DImode);
1306 emit_insn (gen_embmedany_sethi (temp1, op1));
1307 emit_insn (gen_embmedany_brsum (temp2, temp1));
1308 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1311 /* Text segment: sethi %uhi(symbol), %temp1
1312 sethi %hi(symbol), %temp2
1313 or %temp1, %ulo(symbol), %temp3
1314 sllx %temp3, 32, %temp4
1315 or %temp4, %temp2, %temp5
1316 or %temp5, %lo(symbol), %reg */
1321 /* It is possible that one of the registers we got for operands[2]
1322 might coincide with that of operands[0] (which is why we made
1323 it TImode). Pick the other one to use as our scratch. */
1324 if (rtx_equal_p (temp, op0))
1326 gcc_assert (ti_temp);
1327 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1330 temp2 = temp; /* op0 is _not_ allowed, see above. */
1337 temp1 = gen_reg_rtx (DImode);
1338 temp2 = gen_reg_rtx (DImode);
1339 temp3 = gen_reg_rtx (DImode);
1340 temp4 = gen_reg_rtx (DImode);
1341 temp5 = gen_reg_rtx (DImode);
1344 emit_insn (gen_embmedany_textuhi (temp1, op1));
1345 emit_insn (gen_embmedany_texthi (temp2, op1));
1346 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1347 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1348 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1349 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1350 gen_rtx_PLUS (DImode, temp4, temp2)));
1351 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1360 #if HOST_BITS_PER_WIDE_INT == 32
1362 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1367 /* These avoid problems when cross compiling. If we do not
1368 go through all this hair then the optimizer will see
1369 invalid REG_EQUAL notes or in some cases none at all. */
1370 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1371 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1372 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1373 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1375 /* The optimizer is not to assume anything about exactly
1376 which bits are set for a HIGH, they are unspecified.
1377 Unfortunately this leads to many missed optimizations
1378 during CSE. We mask out the non-HIGH bits, and matches
1379 a plain movdi, to alleviate this problem. */
1381 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1383 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1387 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1389 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1393 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1395 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1399 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1401 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1404 /* Worker routines for 64-bit constant formation on arch64.
1405 One of the key things to be doing in these emissions is
1406 to create as many temp REGs as possible. This makes it
1407 possible for half-built constants to be used later when
1408 such values are similar to something required later on.
1409 Without doing this, the optimizer cannot see such
1412 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1413 unsigned HOST_WIDE_INT, int);
1416 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1417 unsigned HOST_WIDE_INT low_bits, int is_neg)
1419 unsigned HOST_WIDE_INT high_bits;
1422 high_bits = (~low_bits) & 0xffffffff;
1424 high_bits = low_bits;
1426 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1429 emit_insn (gen_rtx_SET (VOIDmode, op0,
1430 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1434 /* If we are XOR'ing with -1, then we should emit a one's complement
1435 instead. This way the combiner will notice logical operations
1436 such as ANDN later on and substitute. */
1437 if ((low_bits & 0x3ff) == 0x3ff)
1439 emit_insn (gen_rtx_SET (VOIDmode, op0,
1440 gen_rtx_NOT (DImode, temp)));
1444 emit_insn (gen_rtx_SET (VOIDmode, op0,
1445 gen_safe_XOR64 (temp,
1446 (-(HOST_WIDE_INT)0x400
1447 | (low_bits & 0x3ff)))));
1452 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1453 unsigned HOST_WIDE_INT, int);
1456 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1457 unsigned HOST_WIDE_INT high_bits,
1458 unsigned HOST_WIDE_INT low_immediate,
1463 if ((high_bits & 0xfffffc00) != 0)
1465 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1466 if ((high_bits & ~0xfffffc00) != 0)
1467 emit_insn (gen_rtx_SET (VOIDmode, op0,
1468 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1474 emit_insn (gen_safe_SET64 (temp, high_bits));
1478 /* Now shift it up into place. */
1479 emit_insn (gen_rtx_SET (VOIDmode, op0,
1480 gen_rtx_ASHIFT (DImode, temp2,
1481 GEN_INT (shift_count))));
1483 /* If there is a low immediate part piece, finish up by
1484 putting that in as well. */
1485 if (low_immediate != 0)
1486 emit_insn (gen_rtx_SET (VOIDmode, op0,
1487 gen_safe_OR64 (op0, low_immediate)));
1490 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1491 unsigned HOST_WIDE_INT);
1493 /* Full 64-bit constant decomposition. Even though this is the
1494 'worst' case, we still optimize a few things away. */
1496 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1497 unsigned HOST_WIDE_INT high_bits,
1498 unsigned HOST_WIDE_INT low_bits)
1502 if (reload_in_progress || reload_completed)
1505 sub_temp = gen_reg_rtx (DImode);
1507 if ((high_bits & 0xfffffc00) != 0)
1509 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1510 if ((high_bits & ~0xfffffc00) != 0)
1511 emit_insn (gen_rtx_SET (VOIDmode,
1513 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1519 emit_insn (gen_safe_SET64 (temp, high_bits));
1523 if (!reload_in_progress && !reload_completed)
1525 rtx temp2 = gen_reg_rtx (DImode);
1526 rtx temp3 = gen_reg_rtx (DImode);
1527 rtx temp4 = gen_reg_rtx (DImode);
1529 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1530 gen_rtx_ASHIFT (DImode, sub_temp,
1533 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1534 if ((low_bits & ~0xfffffc00) != 0)
1536 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1537 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1538 emit_insn (gen_rtx_SET (VOIDmode, op0,
1539 gen_rtx_PLUS (DImode, temp4, temp3)));
1543 emit_insn (gen_rtx_SET (VOIDmode, op0,
1544 gen_rtx_PLUS (DImode, temp4, temp2)));
1549 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1550 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1551 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1554 /* We are in the middle of reload, so this is really
1555 painful. However we do still make an attempt to
1556 avoid emitting truly stupid code. */
1557 if (low1 != const0_rtx)
1559 emit_insn (gen_rtx_SET (VOIDmode, op0,
1560 gen_rtx_ASHIFT (DImode, sub_temp,
1561 GEN_INT (to_shift))));
1562 emit_insn (gen_rtx_SET (VOIDmode, op0,
1563 gen_rtx_IOR (DImode, op0, low1)));
1571 if (low2 != const0_rtx)
1573 emit_insn (gen_rtx_SET (VOIDmode, op0,
1574 gen_rtx_ASHIFT (DImode, sub_temp,
1575 GEN_INT (to_shift))));
1576 emit_insn (gen_rtx_SET (VOIDmode, op0,
1577 gen_rtx_IOR (DImode, op0, low2)));
1585 emit_insn (gen_rtx_SET (VOIDmode, op0,
1586 gen_rtx_ASHIFT (DImode, sub_temp,
1587 GEN_INT (to_shift))));
1588 if (low3 != const0_rtx)
1589 emit_insn (gen_rtx_SET (VOIDmode, op0,
1590 gen_rtx_IOR (DImode, op0, low3)));
1595 /* Analyze a 64-bit constant for certain properties. */
1596 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1597 unsigned HOST_WIDE_INT,
1598 int *, int *, int *);
1601 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1602 unsigned HOST_WIDE_INT low_bits,
1603 int *hbsp, int *lbsp, int *abbasp)
1605 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1608 lowest_bit_set = highest_bit_set = -1;
1612 if ((lowest_bit_set == -1)
1613 && ((low_bits >> i) & 1))
1615 if ((highest_bit_set == -1)
1616 && ((high_bits >> (32 - i - 1)) & 1))
1617 highest_bit_set = (64 - i - 1);
1620 && ((highest_bit_set == -1)
1621 || (lowest_bit_set == -1)));
1627 if ((lowest_bit_set == -1)
1628 && ((high_bits >> i) & 1))
1629 lowest_bit_set = i + 32;
1630 if ((highest_bit_set == -1)
1631 && ((low_bits >> (32 - i - 1)) & 1))
1632 highest_bit_set = 32 - i - 1;
1635 && ((highest_bit_set == -1)
1636 || (lowest_bit_set == -1)));
1638 /* If there are no bits set this should have gone out
1639 as one instruction! */
1640 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
1641 all_bits_between_are_set = 1;
1642 for (i = lowest_bit_set; i <= highest_bit_set; i++)
1646 if ((low_bits & (1 << i)) != 0)
1651 if ((high_bits & (1 << (i - 32))) != 0)
1654 all_bits_between_are_set = 0;
1657 *hbsp = highest_bit_set;
1658 *lbsp = lowest_bit_set;
1659 *abbasp = all_bits_between_are_set;
1662 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
1665 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
1666 unsigned HOST_WIDE_INT low_bits)
1668 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
1671 || high_bits == 0xffffffff)
1674 analyze_64bit_constant (high_bits, low_bits,
1675 &highest_bit_set, &lowest_bit_set,
1676 &all_bits_between_are_set);
1678 if ((highest_bit_set == 63
1679 || lowest_bit_set == 0)
1680 && all_bits_between_are_set != 0)
1683 if ((highest_bit_set - lowest_bit_set) < 21)
1689 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
1690 unsigned HOST_WIDE_INT,
1693 static unsigned HOST_WIDE_INT
1694 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
1695 unsigned HOST_WIDE_INT low_bits,
1696 int lowest_bit_set, int shift)
1698 HOST_WIDE_INT hi, lo;
1700 if (lowest_bit_set < 32)
1702 lo = (low_bits >> lowest_bit_set) << shift;
1703 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
1708 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
1710 gcc_assert (! (hi & lo));
1714 /* Here we are sure to be arch64 and this is an integer constant
1715 being loaded into a register. Emit the most efficient
1716 insn sequence possible. Detection of all the 1-insn cases
1717 has been done already. */
1719 sparc_emit_set_const64 (rtx op0, rtx op1)
1721 unsigned HOST_WIDE_INT high_bits, low_bits;
1722 int lowest_bit_set, highest_bit_set;
1723 int all_bits_between_are_set;
1726 /* Sanity check that we know what we are working with. */
1727 gcc_assert (TARGET_ARCH64
1728 && (GET_CODE (op0) == SUBREG
1729 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
1731 if (reload_in_progress || reload_completed)
1734 if (GET_CODE (op1) != CONST_INT)
1736 sparc_emit_set_symbolic_const64 (op0, op1, temp);
1741 temp = gen_reg_rtx (DImode);
1743 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
1744 low_bits = (INTVAL (op1) & 0xffffffff);
1746 /* low_bits bits 0 --> 31
1747 high_bits bits 32 --> 63 */
1749 analyze_64bit_constant (high_bits, low_bits,
1750 &highest_bit_set, &lowest_bit_set,
1751 &all_bits_between_are_set);
1753 /* First try for a 2-insn sequence. */
1755 /* These situations are preferred because the optimizer can
1756 * do more things with them:
1758 * sllx %reg, shift, %reg
1760 * srlx %reg, shift, %reg
1761 * 3) mov some_small_const, %reg
1762 * sllx %reg, shift, %reg
1764 if (((highest_bit_set == 63
1765 || lowest_bit_set == 0)
1766 && all_bits_between_are_set != 0)
1767 || ((highest_bit_set - lowest_bit_set) < 12))
1769 HOST_WIDE_INT the_const = -1;
1770 int shift = lowest_bit_set;
1772 if ((highest_bit_set != 63
1773 && lowest_bit_set != 0)
1774 || all_bits_between_are_set == 0)
1777 create_simple_focus_bits (high_bits, low_bits,
1780 else if (lowest_bit_set == 0)
1781 shift = -(63 - highest_bit_set);
1783 gcc_assert (SPARC_SIMM13_P (the_const));
1784 gcc_assert (shift != 0);
1786 emit_insn (gen_safe_SET64 (temp, the_const));
1788 emit_insn (gen_rtx_SET (VOIDmode,
1790 gen_rtx_ASHIFT (DImode,
1794 emit_insn (gen_rtx_SET (VOIDmode,
1796 gen_rtx_LSHIFTRT (DImode,
1798 GEN_INT (-shift))));
1802 /* Now a range of 22 or less bits set somewhere.
1803 * 1) sethi %hi(focus_bits), %reg
1804 * sllx %reg, shift, %reg
1805 * 2) sethi %hi(focus_bits), %reg
1806 * srlx %reg, shift, %reg
1808 if ((highest_bit_set - lowest_bit_set) < 21)
1810 unsigned HOST_WIDE_INT focus_bits =
1811 create_simple_focus_bits (high_bits, low_bits,
1812 lowest_bit_set, 10);
1814 gcc_assert (SPARC_SETHI_P (focus_bits));
1815 gcc_assert (lowest_bit_set != 10);
1817 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
1819 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
1820 if (lowest_bit_set < 10)
1821 emit_insn (gen_rtx_SET (VOIDmode,
1823 gen_rtx_LSHIFTRT (DImode, temp,
1824 GEN_INT (10 - lowest_bit_set))));
1825 else if (lowest_bit_set > 10)
1826 emit_insn (gen_rtx_SET (VOIDmode,
1828 gen_rtx_ASHIFT (DImode, temp,
1829 GEN_INT (lowest_bit_set - 10))));
1833 /* 1) sethi %hi(low_bits), %reg
1834 * or %reg, %lo(low_bits), %reg
1835 * 2) sethi %hi(~low_bits), %reg
1836 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
1839 || high_bits == 0xffffffff)
1841 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
1842 (high_bits == 0xffffffff));
1846 /* Now, try 3-insn sequences. */
1848 /* 1) sethi %hi(high_bits), %reg
1849 * or %reg, %lo(high_bits), %reg
1850 * sllx %reg, 32, %reg
1854 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
1858 /* We may be able to do something quick
1859 when the constant is negated, so try that. */
1860 if (const64_is_2insns ((~high_bits) & 0xffffffff,
1861 (~low_bits) & 0xfffffc00))
1863 /* NOTE: The trailing bits get XOR'd so we need the
1864 non-negated bits, not the negated ones. */
1865 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
1867 if ((((~high_bits) & 0xffffffff) == 0
1868 && ((~low_bits) & 0x80000000) == 0)
1869 || (((~high_bits) & 0xffffffff) == 0xffffffff
1870 && ((~low_bits) & 0x80000000) != 0))
1872 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
1874 if ((SPARC_SETHI_P (fast_int)
1875 && (~high_bits & 0xffffffff) == 0)
1876 || SPARC_SIMM13_P (fast_int))
1877 emit_insn (gen_safe_SET64 (temp, fast_int));
1879 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
1884 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
1885 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
1886 sparc_emit_set_const64 (temp, negated_const);
1889 /* If we are XOR'ing with -1, then we should emit a one's complement
1890 instead. This way the combiner will notice logical operations
1891 such as ANDN later on and substitute. */
1892 if (trailing_bits == 0x3ff)
1894 emit_insn (gen_rtx_SET (VOIDmode, op0,
1895 gen_rtx_NOT (DImode, temp)));
1899 emit_insn (gen_rtx_SET (VOIDmode,
1901 gen_safe_XOR64 (temp,
1902 (-0x400 | trailing_bits))));
1907 /* 1) sethi %hi(xxx), %reg
1908 * or %reg, %lo(xxx), %reg
1909 * sllx %reg, yyy, %reg
1911 * ??? This is just a generalized version of the low_bits==0
1912 * thing above, FIXME...
1914 if ((highest_bit_set - lowest_bit_set) < 32)
1916 unsigned HOST_WIDE_INT focus_bits =
1917 create_simple_focus_bits (high_bits, low_bits,
1920 /* We can't get here in this state. */
1921 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
1923 /* So what we know is that the set bits straddle the
1924 middle of the 64-bit word. */
1925 sparc_emit_set_const64_quick2 (op0, temp,
1931 /* 1) sethi %hi(high_bits), %reg
1932 * or %reg, %lo(high_bits), %reg
1933 * sllx %reg, 32, %reg
1934 * or %reg, low_bits, %reg
1936 if (SPARC_SIMM13_P(low_bits)
1937 && ((int)low_bits > 0))
1939 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
1943 /* The easiest way when all else fails, is full decomposition. */
1945 printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n",
1946 high_bits, low_bits, ~high_bits, ~low_bits);
1948 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
1950 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
1952 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
1953 return the mode to be used for the comparison. For floating-point,
1954 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
1955 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
1956 processing is needed. */
1959 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
1961 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
1987 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
1988 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
1990 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
1991 return CCX_NOOVmode;
1997 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2004 /* Emit the compare insn and return the CC reg for a CODE comparison. */
2007 gen_compare_reg (enum rtx_code code)
2009 enum machine_mode mode;
2012 if (GET_MODE_CLASS (GET_MODE (sparc_compare_op0)) == MODE_CC)
2013 return sparc_compare_op0;
2015 x = sparc_compare_op0;
2016 y = sparc_compare_op1;
2017 mode = SELECT_CC_MODE (code, x, y);
2019 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2020 fcc regs (cse can't tell they're really call clobbered regs and will
2021 remove a duplicate comparison even if there is an intervening function
2022 call - it will then try to reload the cc reg via an int reg which is why
2023 we need the movcc patterns). It is possible to provide the movcc
2024 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2025 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2026 to tell cse that CCFPE mode registers (even pseudos) are call
2029 /* ??? This is an experiment. Rather than making changes to cse which may
2030 or may not be easy/clean, we do our own cse. This is possible because
2031 we will generate hard registers. Cse knows they're call clobbered (it
2032 doesn't know the same thing about pseudos). If we guess wrong, no big
2033 deal, but if we win, great! */
2035 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2036 #if 1 /* experiment */
2039 /* We cycle through the registers to ensure they're all exercised. */
2040 static int next_fcc_reg = 0;
2041 /* Previous x,y for each fcc reg. */
2042 static rtx prev_args[4][2];
2044 /* Scan prev_args for x,y. */
2045 for (reg = 0; reg < 4; reg++)
2046 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2051 prev_args[reg][0] = x;
2052 prev_args[reg][1] = y;
2053 next_fcc_reg = (next_fcc_reg + 1) & 3;
2055 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2058 cc_reg = gen_reg_rtx (mode);
2059 #endif /* ! experiment */
2060 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2061 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2063 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2065 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2066 will only result in an unrecognizable insn so no point in asserting. */
2067 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2072 /* Same as above but return the whole compare operator. */
2075 gen_compare_operator (enum rtx_code code)
2079 if (GET_MODE (sparc_compare_op0) == TFmode && !TARGET_HARD_QUAD)
2081 = sparc_emit_float_lib_cmp (sparc_compare_op0, sparc_compare_op1, code);
2083 cc_reg = gen_compare_reg (code);
2084 return gen_rtx_fmt_ee (code, GET_MODE (cc_reg), cc_reg, const0_rtx);
2087 /* This function is used for v9 only.
2088 CODE is the code for an Scc's comparison.
2089 OPERANDS[0] is the target of the Scc insn.
2090 OPERANDS[1] is the value we compare against const0_rtx (which hasn't
2091 been generated yet).
2093 This function is needed to turn
2096 (gt (reg:CCX 100 %icc)
2100 (gt:DI (reg:CCX 100 %icc)
2103 IE: The instruction recognizer needs to see the mode of the comparison to
2104 find the right instruction. We could use "gt:DI" right in the
2105 define_expand, but leaving it out allows us to handle DI, SI, etc.
2107 We refer to the global sparc compare operands sparc_compare_op0 and
2108 sparc_compare_op1. */
2111 gen_v9_scc (enum rtx_code compare_code, register rtx *operands)
2114 && (GET_MODE (sparc_compare_op0) == DImode
2115 || GET_MODE (operands[0]) == DImode))
2118 /* Try to use the movrCC insns. */
2120 && GET_MODE_CLASS (GET_MODE (sparc_compare_op0)) == MODE_INT
2121 && sparc_compare_op1 == const0_rtx
2122 && v9_regcmp_p (compare_code))
2124 rtx op0 = sparc_compare_op0;
2127 /* Special case for op0 != 0. This can be done with one instruction if
2128 operands[0] == sparc_compare_op0. */
2130 if (compare_code == NE
2131 && GET_MODE (operands[0]) == DImode
2132 && rtx_equal_p (op0, operands[0]))
2134 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2135 gen_rtx_IF_THEN_ELSE (DImode,
2136 gen_rtx_fmt_ee (compare_code, DImode,
2143 if (reg_overlap_mentioned_p (operands[0], op0))
2145 /* Handle the case where operands[0] == sparc_compare_op0.
2146 We "early clobber" the result. */
2147 op0 = gen_reg_rtx (GET_MODE (sparc_compare_op0));
2148 emit_move_insn (op0, sparc_compare_op0);
2151 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2152 if (GET_MODE (op0) != DImode)
2154 temp = gen_reg_rtx (DImode);
2155 convert_move (temp, op0, 0);
2159 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2160 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2161 gen_rtx_fmt_ee (compare_code, DImode,
2169 operands[1] = gen_compare_reg (compare_code);
2171 switch (GET_MODE (operands[1]))
2181 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2182 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2183 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2184 gen_rtx_fmt_ee (compare_code,
2185 GET_MODE (operands[1]),
2186 operands[1], const0_rtx),
2187 const1_rtx, operands[0])));
2192 /* Emit a conditional jump insn for the v9 architecture using comparison code
2193 CODE and jump target LABEL.
2194 This function exists to take advantage of the v9 brxx insns. */
2197 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2199 gcc_assert (GET_MODE_CLASS (GET_MODE (sparc_compare_op0)) != MODE_CC);
2200 emit_jump_insn (gen_rtx_SET (VOIDmode,
2202 gen_rtx_IF_THEN_ELSE (VOIDmode,
2203 gen_rtx_fmt_ee (code, GET_MODE (op0),
2205 gen_rtx_LABEL_REF (VOIDmode, label),
2209 /* Generate a DFmode part of a hard TFmode register.
2210 REG is the TFmode hard register, LOW is 1 for the
2211 low 64bit of the register and 0 otherwise.
2214 gen_df_reg (rtx reg, int low)
2216 int regno = REGNO (reg);
2218 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2219 regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2;
2220 return gen_rtx_REG (DFmode, regno);
2223 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2224 Unlike normal calls, TFmode operands are passed by reference. It is
2225 assumed that no more than 3 operands are required. */
2228 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2230 rtx ret_slot = NULL, arg[3], func_sym;
2233 /* We only expect to be called for conversions, unary, and binary ops. */
2234 gcc_assert (nargs == 2 || nargs == 3);
2236 for (i = 0; i < nargs; ++i)
2238 rtx this_arg = operands[i];
2241 /* TFmode arguments and return values are passed by reference. */
2242 if (GET_MODE (this_arg) == TFmode)
2244 int force_stack_temp;
2246 force_stack_temp = 0;
2247 if (TARGET_BUGGY_QP_LIB && i == 0)
2248 force_stack_temp = 1;
2250 if (GET_CODE (this_arg) == MEM
2251 && ! force_stack_temp)
2252 this_arg = XEXP (this_arg, 0);
2253 else if (CONSTANT_P (this_arg)
2254 && ! force_stack_temp)
2256 this_slot = force_const_mem (TFmode, this_arg);
2257 this_arg = XEXP (this_slot, 0);
2261 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2263 /* Operand 0 is the return value. We'll copy it out later. */
2265 emit_move_insn (this_slot, this_arg);
2267 ret_slot = this_slot;
2269 this_arg = XEXP (this_slot, 0);
2276 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2278 if (GET_MODE (operands[0]) == TFmode)
2281 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2282 arg[0], GET_MODE (arg[0]),
2283 arg[1], GET_MODE (arg[1]));
2285 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2286 arg[0], GET_MODE (arg[0]),
2287 arg[1], GET_MODE (arg[1]),
2288 arg[2], GET_MODE (arg[2]));
2291 emit_move_insn (operands[0], ret_slot);
2297 gcc_assert (nargs == 2);
2299 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2300 GET_MODE (operands[0]), 1,
2301 arg[1], GET_MODE (arg[1]));
2303 if (ret != operands[0])
2304 emit_move_insn (operands[0], ret);
2308 /* Expand soft-float TFmode calls to sparc abi routines. */
2311 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2333 emit_soft_tfmode_libcall (func, 3, operands);
2337 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2341 gcc_assert (code == SQRT);
2344 emit_soft_tfmode_libcall (func, 2, operands);
2348 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2355 switch (GET_MODE (operands[1]))
2368 case FLOAT_TRUNCATE:
2369 switch (GET_MODE (operands[0]))
2383 switch (GET_MODE (operands[1]))
2388 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2398 case UNSIGNED_FLOAT:
2399 switch (GET_MODE (operands[1]))
2404 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2415 switch (GET_MODE (operands[0]))
2429 switch (GET_MODE (operands[0]))
2446 emit_soft_tfmode_libcall (func, 2, operands);
2449 /* Expand a hard-float tfmode operation. All arguments must be in
2453 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2457 if (GET_RTX_CLASS (code) == RTX_UNARY)
2459 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2460 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2464 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2465 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2466 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2467 operands[1], operands[2]);
2470 if (register_operand (operands[0], VOIDmode))
2473 dest = gen_reg_rtx (GET_MODE (operands[0]));
2475 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2477 if (dest != operands[0])
2478 emit_move_insn (operands[0], dest);
2482 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2484 if (TARGET_HARD_QUAD)
2485 emit_hard_tfmode_operation (code, operands);
2487 emit_soft_tfmode_binop (code, operands);
2491 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2493 if (TARGET_HARD_QUAD)
2494 emit_hard_tfmode_operation (code, operands);
2496 emit_soft_tfmode_unop (code, operands);
2500 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2502 if (TARGET_HARD_QUAD)
2503 emit_hard_tfmode_operation (code, operands);
2505 emit_soft_tfmode_cvt (code, operands);
2508 /* Return nonzero if a branch/jump/call instruction will be emitting
2509 nop into its delay slot. */
2512 empty_delay_slot (rtx insn)
2516 /* If no previous instruction (should not happen), return true. */
2517 if (PREV_INSN (insn) == NULL)
2520 seq = NEXT_INSN (PREV_INSN (insn));
2521 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
2527 /* Return nonzero if TRIAL can go into the call delay slot. */
2530 tls_call_delay (rtx trial)
2535 call __tls_get_addr, %tgd_call (foo)
2536 add %l7, %o0, %o0, %tgd_add (foo)
2537 while Sun as/ld does not. */
2538 if (TARGET_GNU_TLS || !TARGET_TLS)
2541 pat = PATTERN (trial);
2543 /* We must reject tgd_add{32|64}, i.e.
2544 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
2545 and tldm_add{32|64}, i.e.
2546 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
2548 if (GET_CODE (pat) == SET
2549 && GET_CODE (SET_SRC (pat)) == PLUS)
2551 rtx unspec = XEXP (SET_SRC (pat), 1);
2553 if (GET_CODE (unspec) == UNSPEC
2554 && (XINT (unspec, 1) == UNSPEC_TLSGD
2555 || XINT (unspec, 1) == UNSPEC_TLSLDM))
2562 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
2563 instruction. RETURN_P is true if the v9 variant 'return' is to be
2564 considered in the test too.
2566 TRIAL must be a SET whose destination is a REG appropriate for the
2567 'restore' instruction or, if RETURN_P is true, for the 'return'
2571 eligible_for_restore_insn (rtx trial, bool return_p)
2573 rtx pat = PATTERN (trial);
2574 rtx src = SET_SRC (pat);
2576 /* The 'restore src,%g0,dest' pattern for word mode and below. */
2577 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
2578 && arith_operand (src, GET_MODE (src)))
2581 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
2583 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
2586 /* The 'restore src,%g0,dest' pattern for double-word mode. */
2587 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
2588 && arith_double_operand (src, GET_MODE (src)))
2589 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
2591 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
2592 else if (! TARGET_FPU && register_operand (src, SFmode))
2595 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
2596 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
2599 /* If we have the 'return' instruction, anything that does not use
2600 local or output registers and can go into a delay slot wins. */
2601 else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1)
2602 && (get_attr_in_uncond_branch_delay (trial)
2603 == IN_UNCOND_BRANCH_DELAY_TRUE))
2606 /* The 'restore src1,src2,dest' pattern for SImode. */
2607 else if (GET_CODE (src) == PLUS
2608 && register_operand (XEXP (src, 0), SImode)
2609 && arith_operand (XEXP (src, 1), SImode))
2612 /* The 'restore src1,src2,dest' pattern for DImode. */
2613 else if (GET_CODE (src) == PLUS
2614 && register_operand (XEXP (src, 0), DImode)
2615 && arith_double_operand (XEXP (src, 1), DImode))
2618 /* The 'restore src1,%lo(src2),dest' pattern. */
2619 else if (GET_CODE (src) == LO_SUM
2620 && ! TARGET_CM_MEDMID
2621 && ((register_operand (XEXP (src, 0), SImode)
2622 && immediate_operand (XEXP (src, 1), SImode))
2624 && register_operand (XEXP (src, 0), DImode)
2625 && immediate_operand (XEXP (src, 1), DImode))))
2628 /* The 'restore src,src,dest' pattern. */
2629 else if (GET_CODE (src) == ASHIFT
2630 && (register_operand (XEXP (src, 0), SImode)
2631 || register_operand (XEXP (src, 0), DImode))
2632 && XEXP (src, 1) == const1_rtx)
2638 /* Return nonzero if TRIAL can go into the function return's
2642 eligible_for_return_delay (rtx trial)
2646 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
2649 if (get_attr_length (trial) != 1)
2652 /* If there are any call-saved registers, we should scan TRIAL if it
2653 does not reference them. For now just make it easy. */
2657 /* If the function uses __builtin_eh_return, the eh_return machinery
2658 occupies the delay slot. */
2659 if (crtl->calls_eh_return)
2662 /* In the case of a true leaf function, anything can go into the slot. */
2663 if (sparc_leaf_function_p)
2664 return get_attr_in_uncond_branch_delay (trial)
2665 == IN_UNCOND_BRANCH_DELAY_TRUE;
2667 pat = PATTERN (trial);
2669 /* Otherwise, only operations which can be done in tandem with
2670 a `restore' or `return' insn can go into the delay slot. */
2671 if (GET_CODE (SET_DEST (pat)) != REG
2672 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24))
2675 /* If this instruction sets up floating point register and we have a return
2676 instruction, it can probably go in. But restore will not work
2678 if (REGNO (SET_DEST (pat)) >= 32)
2680 && ! epilogue_renumber (&pat, 1)
2681 && (get_attr_in_uncond_branch_delay (trial)
2682 == IN_UNCOND_BRANCH_DELAY_TRUE));
2684 return eligible_for_restore_insn (trial, true);
2687 /* Return nonzero if TRIAL can go into the sibling call's
2691 eligible_for_sibcall_delay (rtx trial)
2695 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
2698 if (get_attr_length (trial) != 1)
2701 pat = PATTERN (trial);
2703 if (sparc_leaf_function_p)
2705 /* If the tail call is done using the call instruction,
2706 we have to restore %o7 in the delay slot. */
2707 if (LEAF_SIBCALL_SLOT_RESERVED_P)
2710 /* %g1 is used to build the function address */
2711 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
2717 /* Otherwise, only operations which can be done in tandem with
2718 a `restore' insn can go into the delay slot. */
2719 if (GET_CODE (SET_DEST (pat)) != REG
2720 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
2721 || REGNO (SET_DEST (pat)) >= 32)
2724 /* If it mentions %o7, it can't go in, because sibcall will clobber it
2726 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
2729 return eligible_for_restore_insn (trial, false);
2733 short_branch (int uid1, int uid2)
2735 int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2);
2737 /* Leave a few words of "slop". */
2738 if (delta >= -1023 && delta <= 1022)
2744 /* Return nonzero if REG is not used after INSN.
2745 We assume REG is a reload reg, and therefore does
2746 not live past labels or calls or jumps. */
2748 reg_unused_after (rtx reg, rtx insn)
2750 enum rtx_code code, prev_code = UNKNOWN;
2752 while ((insn = NEXT_INSN (insn)))
2754 if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)])
2757 code = GET_CODE (insn);
2758 if (GET_CODE (insn) == CODE_LABEL)
2763 rtx set = single_set (insn);
2764 int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set));
2767 if (set && reg_overlap_mentioned_p (reg, SET_DEST (set)))
2769 if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
2777 /* Determine if it's legal to put X into the constant pool. This
2778 is not possible if X contains the address of a symbol that is
2779 not constant (TLS) or not known at final link time (PIC). */
2782 sparc_cannot_force_const_mem (rtx x)
2784 switch (GET_CODE (x))
2789 /* Accept all non-symbolic constants. */
2793 /* Labels are OK iff we are non-PIC. */
2794 return flag_pic != 0;
2797 /* 'Naked' TLS symbol references are never OK,
2798 non-TLS symbols are OK iff we are non-PIC. */
2799 if (SYMBOL_REF_TLS_MODEL (x))
2802 return flag_pic != 0;
2805 return sparc_cannot_force_const_mem (XEXP (x, 0));
2808 return sparc_cannot_force_const_mem (XEXP (x, 0))
2809 || sparc_cannot_force_const_mem (XEXP (x, 1));
2818 static GTY(()) char pic_helper_symbol_name[256];
2819 static GTY(()) rtx pic_helper_symbol;
2820 static GTY(()) bool pic_helper_emitted_p = false;
2821 static GTY(()) rtx global_offset_table;
2823 /* Ensure that we are not using patterns that are not OK with PIC. */
2831 gcc_assert (GET_CODE (recog_data.operand[i]) != SYMBOL_REF
2832 && (GET_CODE (recog_data.operand[i]) != CONST
2833 || (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS
2834 && (XEXP (XEXP (recog_data.operand[i], 0), 0)
2835 == global_offset_table)
2836 && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1))
2844 /* Return true if X is an address which needs a temporary register when
2845 reloaded while generating PIC code. */
2848 pic_address_needs_scratch (rtx x)
2850 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
2851 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
2852 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
2853 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
2854 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
2860 /* Determine if a given RTX is a valid constant. We already know this
2861 satisfies CONSTANT_P. */
2864 legitimate_constant_p (rtx x)
2868 switch (GET_CODE (x))
2871 /* TLS symbols are not constant. */
2872 if (SYMBOL_REF_TLS_MODEL (x))
2877 inner = XEXP (x, 0);
2879 /* Offsets of TLS symbols are never valid.
2880 Discourage CSE from creating them. */
2881 if (GET_CODE (inner) == PLUS
2882 && SPARC_SYMBOL_REF_TLS_P (XEXP (inner, 0)))
2887 if (GET_MODE (x) == VOIDmode)
2890 /* Floating point constants are generally not ok.
2891 The only exception is 0.0 in VIS. */
2893 && SCALAR_FLOAT_MODE_P (GET_MODE (x))
2894 && const_zero_operand (x, GET_MODE (x)))
2900 /* Vector constants are generally not ok.
2901 The only exception is 0 in VIS. */
2903 && const_zero_operand (x, GET_MODE (x)))
2915 /* Determine if a given RTX is a valid constant address. */
2918 constant_address_p (rtx x)
2920 switch (GET_CODE (x))
2928 if (flag_pic && pic_address_needs_scratch (x))
2930 return legitimate_constant_p (x);
2933 return !flag_pic && legitimate_constant_p (x);
2940 /* Nonzero if the constant value X is a legitimate general operand
2941 when generating PIC code. It is given that flag_pic is on and
2942 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
2945 legitimate_pic_operand_p (rtx x)
2947 if (pic_address_needs_scratch (x))
2949 if (SPARC_SYMBOL_REF_TLS_P (x)
2950 || (GET_CODE (x) == CONST
2951 && GET_CODE (XEXP (x, 0)) == PLUS
2952 && SPARC_SYMBOL_REF_TLS_P (XEXP (XEXP (x, 0), 0))))
2957 /* Return nonzero if ADDR is a valid memory address.
2958 STRICT specifies whether strict register checking applies. */
2961 legitimate_address_p (enum machine_mode mode, rtx addr, int strict)
2963 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
2965 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
2967 else if (GET_CODE (addr) == PLUS)
2969 rs1 = XEXP (addr, 0);
2970 rs2 = XEXP (addr, 1);
2972 /* Canonicalize. REG comes first, if there are no regs,
2973 LO_SUM comes first. */
2975 && GET_CODE (rs1) != SUBREG
2977 || GET_CODE (rs2) == SUBREG
2978 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
2980 rs1 = XEXP (addr, 1);
2981 rs2 = XEXP (addr, 0);
2985 && rs1 == pic_offset_table_rtx
2987 && GET_CODE (rs2) != SUBREG
2988 && GET_CODE (rs2) != LO_SUM
2989 && GET_CODE (rs2) != MEM
2990 && ! SPARC_SYMBOL_REF_TLS_P (rs2)
2991 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
2992 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
2994 || GET_CODE (rs1) == SUBREG)
2995 && RTX_OK_FOR_OFFSET_P (rs2)))
3000 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3001 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3003 /* We prohibit REG + REG for TFmode when there are no quad move insns
3004 and we consequently need to split. We do this because REG+REG
3005 is not an offsettable address. If we get the situation in reload
3006 where source and destination of a movtf pattern are both MEMs with
3007 REG+REG address, then only one of them gets converted to an
3008 offsettable address. */
3010 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3013 /* We prohibit REG + REG on ARCH32 if not optimizing for
3014 DFmode/DImode because then mem_min_alignment is likely to be zero
3015 after reload and the forced split would lack a matching splitter
3017 if (TARGET_ARCH32 && !optimize
3018 && (mode == DFmode || mode == DImode))
3021 else if (USE_AS_OFFSETABLE_LO10
3022 && GET_CODE (rs1) == LO_SUM
3024 && ! TARGET_CM_MEDMID
3025 && RTX_OK_FOR_OLO10_P (rs2))
3028 imm1 = XEXP (rs1, 1);
3029 rs1 = XEXP (rs1, 0);
3030 if (! CONSTANT_P (imm1) || SPARC_SYMBOL_REF_TLS_P (rs1))
3034 else if (GET_CODE (addr) == LO_SUM)
3036 rs1 = XEXP (addr, 0);
3037 imm1 = XEXP (addr, 1);
3039 if (! CONSTANT_P (imm1) || SPARC_SYMBOL_REF_TLS_P (rs1))
3042 /* We can't allow TFmode in 32-bit mode, because an offset greater
3043 than the alignment (8) may cause the LO_SUM to overflow. */
3044 if (mode == TFmode && TARGET_ARCH32)
3047 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3052 if (GET_CODE (rs1) == SUBREG)
3053 rs1 = SUBREG_REG (rs1);
3059 if (GET_CODE (rs2) == SUBREG)
3060 rs2 = SUBREG_REG (rs2);
3067 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3068 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3073 if ((REGNO (rs1) >= 32
3074 && REGNO (rs1) != FRAME_POINTER_REGNUM
3075 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3077 && (REGNO (rs2) >= 32
3078 && REGNO (rs2) != FRAME_POINTER_REGNUM
3079 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3085 /* Construct the SYMBOL_REF for the tls_get_offset function. */
3087 static GTY(()) rtx sparc_tls_symbol;
3090 sparc_tls_get_addr (void)
3092 if (!sparc_tls_symbol)
3093 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3095 return sparc_tls_symbol;
3099 sparc_tls_got (void)
3104 crtl->uses_pic_offset_table = 1;
3105 return pic_offset_table_rtx;
3108 if (!global_offset_table)
3109 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3110 temp = gen_reg_rtx (Pmode);
3111 emit_move_insn (temp, global_offset_table);
3115 /* Return 1 if *X is a thread-local symbol. */
3118 sparc_tls_symbol_ref_1 (rtx *x, void *data ATTRIBUTE_UNUSED)
3120 return SPARC_SYMBOL_REF_TLS_P (*x);
3123 /* Return 1 if X contains a thread-local symbol. */
3126 sparc_tls_referenced_p (rtx x)
3128 if (!TARGET_HAVE_TLS)
3131 return for_each_rtx (&x, &sparc_tls_symbol_ref_1, 0);
3134 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3135 this (thread-local) address. */
3138 legitimize_tls_address (rtx addr)
3140 rtx temp1, temp2, temp3, ret, o0, got, insn;
3142 gcc_assert (can_create_pseudo_p ());
3144 if (GET_CODE (addr) == SYMBOL_REF)
3145 switch (SYMBOL_REF_TLS_MODEL (addr))
3147 case TLS_MODEL_GLOBAL_DYNAMIC:
3149 temp1 = gen_reg_rtx (SImode);
3150 temp2 = gen_reg_rtx (SImode);
3151 ret = gen_reg_rtx (Pmode);
3152 o0 = gen_rtx_REG (Pmode, 8);
3153 got = sparc_tls_got ();
3154 emit_insn (gen_tgd_hi22 (temp1, addr));
3155 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3158 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3159 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3164 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3165 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3168 CALL_INSN_FUNCTION_USAGE (insn)
3169 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3170 CALL_INSN_FUNCTION_USAGE (insn));
3171 insn = get_insns ();
3173 emit_libcall_block (insn, ret, o0, addr);
3176 case TLS_MODEL_LOCAL_DYNAMIC:
3178 temp1 = gen_reg_rtx (SImode);
3179 temp2 = gen_reg_rtx (SImode);
3180 temp3 = gen_reg_rtx (Pmode);
3181 ret = gen_reg_rtx (Pmode);
3182 o0 = gen_rtx_REG (Pmode, 8);
3183 got = sparc_tls_got ();
3184 emit_insn (gen_tldm_hi22 (temp1));
3185 emit_insn (gen_tldm_lo10 (temp2, temp1));
3188 emit_insn (gen_tldm_add32 (o0, got, temp2));
3189 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3194 emit_insn (gen_tldm_add64 (o0, got, temp2));
3195 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3198 CALL_INSN_FUNCTION_USAGE (insn)
3199 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3200 CALL_INSN_FUNCTION_USAGE (insn));
3201 insn = get_insns ();
3203 emit_libcall_block (insn, temp3, o0,
3204 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3205 UNSPEC_TLSLD_BASE));
3206 temp1 = gen_reg_rtx (SImode);
3207 temp2 = gen_reg_rtx (SImode);
3208 emit_insn (gen_tldo_hix22 (temp1, addr));
3209 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3211 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3213 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3216 case TLS_MODEL_INITIAL_EXEC:
3217 temp1 = gen_reg_rtx (SImode);
3218 temp2 = gen_reg_rtx (SImode);
3219 temp3 = gen_reg_rtx (Pmode);
3220 got = sparc_tls_got ();
3221 emit_insn (gen_tie_hi22 (temp1, addr));
3222 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3224 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3226 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3229 ret = gen_reg_rtx (Pmode);
3231 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3234 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3238 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3241 case TLS_MODEL_LOCAL_EXEC:
3242 temp1 = gen_reg_rtx (Pmode);
3243 temp2 = gen_reg_rtx (Pmode);
3246 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3247 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3251 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3252 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3254 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3262 gcc_unreachable (); /* for now ... */
3268 /* Legitimize PIC addresses. If the address is already position-independent,
3269 we return ORIG. Newly generated position-independent addresses go into a
3270 reg. This is REG if nonzero, otherwise we allocate register(s) as
3274 legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED,
3277 if (GET_CODE (orig) == SYMBOL_REF
3278 /* See the comment in sparc_expand_move. */
3279 || (TARGET_VXWORKS_RTP && GET_CODE (orig) == LABEL_REF))
3281 rtx pic_ref, address;
3286 gcc_assert (! reload_in_progress && ! reload_completed);
3287 reg = gen_reg_rtx (Pmode);
3292 /* If not during reload, allocate another temp reg here for loading
3293 in the address, so that these instructions can be optimized
3295 rtx temp_reg = ((reload_in_progress || reload_completed)
3296 ? reg : gen_reg_rtx (Pmode));
3298 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3299 won't get confused into thinking that these two instructions
3300 are loading in the true address of the symbol. If in the
3301 future a PIC rtx exists, that should be used instead. */
3304 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3305 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3309 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3310 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3317 pic_ref = gen_const_mem (Pmode,
3318 gen_rtx_PLUS (Pmode,
3319 pic_offset_table_rtx, address));
3320 crtl->uses_pic_offset_table = 1;
3321 insn = emit_move_insn (reg, pic_ref);
3322 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3324 set_unique_reg_note (insn, REG_EQUAL, orig);
3327 else if (GET_CODE (orig) == CONST)
3331 if (GET_CODE (XEXP (orig, 0)) == PLUS
3332 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3337 gcc_assert (! reload_in_progress && ! reload_completed);
3338 reg = gen_reg_rtx (Pmode);
3341 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3342 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
3343 offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
3344 base == reg ? 0 : reg);
3346 if (GET_CODE (offset) == CONST_INT)
3348 if (SMALL_INT (offset))
3349 return plus_constant (base, INTVAL (offset));
3350 else if (! reload_in_progress && ! reload_completed)
3351 offset = force_reg (Pmode, offset);
3353 /* If we reach here, then something is seriously wrong. */
3356 return gen_rtx_PLUS (Pmode, base, offset);
3358 else if (GET_CODE (orig) == LABEL_REF)
3359 /* ??? Why do we do this? */
3360 /* Now movsi_pic_label_ref uses it, but we ought to be checking that
3361 the register is live instead, in case it is eliminated. */
3362 crtl->uses_pic_offset_table = 1;
3367 /* Try machine-dependent ways of modifying an illegitimate address X
3368 to be legitimate. If we find one, return the new, valid address.
3370 OLDX is the address as it was before break_out_memory_refs was called.
3371 In some cases it is useful to look at this to decide what needs to be done.
3373 MODE is the mode of the operand pointed to by X. */
3376 legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, enum machine_mode mode)
3380 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3381 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3382 force_operand (XEXP (x, 0), NULL_RTX));
3383 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3384 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3385 force_operand (XEXP (x, 1), NULL_RTX));
3386 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3387 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3389 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3390 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3391 force_operand (XEXP (x, 1), NULL_RTX));
3393 if (x != orig_x && legitimate_address_p (mode, x, FALSE))
3396 if (SPARC_SYMBOL_REF_TLS_P (x))
3397 x = legitimize_tls_address (x);
3399 x = legitimize_pic_address (x, mode, 0);
3400 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3401 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3402 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3403 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3404 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3405 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3406 else if (GET_CODE (x) == SYMBOL_REF
3407 || GET_CODE (x) == CONST
3408 || GET_CODE (x) == LABEL_REF)
3409 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3413 /* Emit the special PIC helper function. */
3416 emit_pic_helper (void)
3418 const char *pic_name = reg_names[REGNO (pic_offset_table_rtx)];
3421 switch_to_section (text_section);
3423 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
3425 ASM_OUTPUT_ALIGN (asm_out_file, align);
3426 ASM_OUTPUT_LABEL (asm_out_file, pic_helper_symbol_name);
3427 if (flag_delayed_branch)
3428 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
3429 pic_name, pic_name);
3431 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
3432 pic_name, pic_name);
3434 pic_helper_emitted_p = true;
3437 /* Emit code to load the PIC register. */
3440 load_pic_register (bool delay_pic_helper)
3442 int orig_flag_pic = flag_pic;
3444 if (TARGET_VXWORKS_RTP)
3446 emit_insn (gen_vxworks_load_got ());
3447 emit_use (pic_offset_table_rtx);
3451 /* If we haven't initialized the special PIC symbols, do so now. */
3452 if (!pic_helper_symbol_name[0])
3454 ASM_GENERATE_INTERNAL_LABEL (pic_helper_symbol_name, "LADDPC", 0);
3455 pic_helper_symbol = gen_rtx_SYMBOL_REF (Pmode, pic_helper_symbol_name);
3456 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3459 /* If we haven't emitted the special PIC helper function, do so now unless
3460 we are requested to delay it. */
3461 if (!delay_pic_helper && !pic_helper_emitted_p)
3466 emit_insn (gen_load_pcrel_symdi (pic_offset_table_rtx, global_offset_table,
3467 pic_helper_symbol));
3469 emit_insn (gen_load_pcrel_symsi (pic_offset_table_rtx, global_offset_table,
3470 pic_helper_symbol));
3471 flag_pic = orig_flag_pic;
3473 /* Need to emit this whether or not we obey regdecls,
3474 since setjmp/longjmp can cause life info to screw up.
3475 ??? In the case where we don't obey regdecls, this is not sufficient
3476 since we may not fall out the bottom. */
3477 emit_use (pic_offset_table_rtx);
3480 /* Emit a call instruction with the pattern given by PAT. ADDR is the
3481 address of the call target. */
3484 sparc_emit_call_insn (rtx pat, rtx addr)
3488 insn = emit_call_insn (pat);
3490 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
3491 if (TARGET_VXWORKS_RTP
3493 && GET_CODE (addr) == SYMBOL_REF
3494 && (SYMBOL_REF_DECL (addr)
3495 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
3496 : !SYMBOL_REF_LOCAL_P (addr)))
3498 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
3499 crtl->uses_pic_offset_table = 1;
3503 /* Return 1 if RTX is a MEM which is known to be aligned to at
3504 least a DESIRED byte boundary. */
3507 mem_min_alignment (rtx mem, int desired)
3509 rtx addr, base, offset;
3511 /* If it's not a MEM we can't accept it. */
3512 if (GET_CODE (mem) != MEM)
3516 if (!TARGET_UNALIGNED_DOUBLES
3517 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
3520 /* ??? The rest of the function predates MEM_ALIGN so
3521 there is probably a bit of redundancy. */
3522 addr = XEXP (mem, 0);
3523 base = offset = NULL_RTX;
3524 if (GET_CODE (addr) == PLUS)
3526 if (GET_CODE (XEXP (addr, 0)) == REG)
3528 base = XEXP (addr, 0);
3530 /* What we are saying here is that if the base
3531 REG is aligned properly, the compiler will make
3532 sure any REG based index upon it will be so
3534 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3535 offset = XEXP (addr, 1);
3537 offset = const0_rtx;
3540 else if (GET_CODE (addr) == REG)
3543 offset = const0_rtx;
3546 if (base != NULL_RTX)
3548 int regno = REGNO (base);
3550 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
3552 /* Check if the compiler has recorded some information
3553 about the alignment of the base REG. If reload has
3554 completed, we already matched with proper alignments.
3555 If not running global_alloc, reload might give us
3556 unaligned pointer to local stack though. */
3558 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
3559 || (optimize && reload_completed))
3560 && (INTVAL (offset) & (desired - 1)) == 0)
3565 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
3569 else if (! TARGET_UNALIGNED_DOUBLES
3570 || CONSTANT_P (addr)
3571 || GET_CODE (addr) == LO_SUM)
3573 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
3574 is true, in which case we can only assume that an access is aligned if
3575 it is to a constant address, or the address involves a LO_SUM. */
3579 /* An obviously unaligned address. */
3584 /* Vectors to keep interesting information about registers where it can easily
3585 be got. We used to use the actual mode value as the bit number, but there
3586 are more than 32 modes now. Instead we use two tables: one indexed by
3587 hard register number, and one indexed by mode. */
3589 /* The purpose of sparc_mode_class is to shrink the range of modes so that
3590 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
3591 mapped into one sparc_mode_class mode. */
3593 enum sparc_mode_class {
3594 S_MODE, D_MODE, T_MODE, O_MODE,
3595 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
3599 /* Modes for single-word and smaller quantities. */
3600 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
3602 /* Modes for double-word and smaller quantities. */
3603 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
3605 /* Modes for quad-word and smaller quantities. */
3606 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
3608 /* Modes for 8-word and smaller quantities. */
3609 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
3611 /* Modes for single-float quantities. We must allow any single word or
3612 smaller quantity. This is because the fix/float conversion instructions
3613 take integer inputs/outputs from the float registers. */
3614 #define SF_MODES (S_MODES)
3616 /* Modes for double-float and smaller quantities. */
3617 #define DF_MODES (S_MODES | D_MODES)
3619 /* Modes for double-float only quantities. */
3620 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
3622 /* Modes for quad-float only quantities. */
3623 #define TF_ONLY_MODES (1 << (int) TF_MODE)
3625 /* Modes for quad-float and smaller quantities. */
3626 #define TF_MODES (DF_MODES | TF_ONLY_MODES)
3628 /* Modes for quad-float and double-float quantities. */
3629 #define TF_MODES_NO_S (DF_MODES_NO_S | TF_ONLY_MODES)
3631 /* Modes for quad-float pair only quantities. */
3632 #define OF_ONLY_MODES (1 << (int) OF_MODE)
3634 /* Modes for quad-float pairs and smaller quantities. */
3635 #define OF_MODES (TF_MODES | OF_ONLY_MODES)
3637 #define OF_MODES_NO_S (TF_MODES_NO_S | OF_ONLY_MODES)
3639 /* Modes for condition codes. */
3640 #define CC_MODES (1 << (int) CC_MODE)
3641 #define CCFP_MODES (1 << (int) CCFP_MODE)
3643 /* Value is 1 if register/mode pair is acceptable on sparc.
3644 The funny mixture of D and T modes is because integer operations
3645 do not specially operate on tetra quantities, so non-quad-aligned
3646 registers can hold quadword quantities (except %o4 and %i4 because
3647 they cross fixed registers). */
3649 /* This points to either the 32 bit or the 64 bit version. */
3650 const int *hard_regno_mode_classes;
3652 static const int hard_32bit_mode_classes[] = {
3653 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
3654 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
3655 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
3656 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
3658 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3659 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3660 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3661 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
3663 /* FP regs f32 to f63. Only the even numbered registers actually exist,
3664 and none can hold SFmode/SImode values. */
3665 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3666 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3667 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3668 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3671 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
3677 static const int hard_64bit_mode_classes[] = {
3678 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3679 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3680 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3681 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
3683 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3684 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3685 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
3686 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
3688 /* FP regs f32 to f63. Only the even numbered registers actually exist,
3689 and none can hold SFmode/SImode values. */
3690 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3691 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3692 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3693 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
3696 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
3702 int sparc_mode_class [NUM_MACHINE_MODES];
3704 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
3707 sparc_init_modes (void)
3711 for (i = 0; i < NUM_MACHINE_MODES; i++)
3713 switch (GET_MODE_CLASS (i))
3716 case MODE_PARTIAL_INT:
3717 case MODE_COMPLEX_INT:
3718 if (GET_MODE_SIZE (i) <= 4)
3719 sparc_mode_class[i] = 1 << (int) S_MODE;
3720 else if (GET_MODE_SIZE (i) == 8)
3721 sparc_mode_class[i] = 1 << (int) D_MODE;
3722 else if (GET_MODE_SIZE (i) == 16)
3723 sparc_mode_class[i] = 1 << (int) T_MODE;
3724 else if (GET_MODE_SIZE (i) == 32)
3725 sparc_mode_class[i] = 1 << (int) O_MODE;
3727 sparc_mode_class[i] = 0;
3729 case MODE_VECTOR_INT:
3730 if (GET_MODE_SIZE (i) <= 4)
3731 sparc_mode_class[i] = 1 << (int)SF_MODE;
3732 else if (GET_MODE_SIZE (i) == 8)
3733 sparc_mode_class[i] = 1 << (int)DF_MODE;
3736 case MODE_COMPLEX_FLOAT:
3737 if (GET_MODE_SIZE (i) <= 4)
3738 sparc_mode_class[i] = 1 << (int) SF_MODE;
3739 else if (GET_MODE_SIZE (i) == 8)
3740 sparc_mode_class[i] = 1 << (int) DF_MODE;
3741 else if (GET_MODE_SIZE (i) == 16)
3742 sparc_mode_class[i] = 1 << (int) TF_MODE;
3743 else if (GET_MODE_SIZE (i) == 32)
3744 sparc_mode_class[i] = 1 << (int) OF_MODE;
3746 sparc_mode_class[i] = 0;
3749 if (i == (int) CCFPmode || i == (int) CCFPEmode)
3750 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
3752 sparc_mode_class[i] = 1 << (int) CC_MODE;
3755 sparc_mode_class[i] = 0;
3761 hard_regno_mode_classes = hard_64bit_mode_classes;
3763 hard_regno_mode_classes = hard_32bit_mode_classes;
3765 /* Initialize the array used by REGNO_REG_CLASS. */
3766 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3768 if (i < 16 && TARGET_V8PLUS)
3769 sparc_regno_reg_class[i] = I64_REGS;
3770 else if (i < 32 || i == FRAME_POINTER_REGNUM)
3771 sparc_regno_reg_class[i] = GENERAL_REGS;
3773 sparc_regno_reg_class[i] = FP_REGS;
3775 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
3777 sparc_regno_reg_class[i] = FPCC_REGS;
3779 sparc_regno_reg_class[i] = NO_REGS;
3783 /* Compute the frame size required by the function. This function is called
3784 during the reload pass and also by sparc_expand_prologue. */
3787 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p)
3789 int outgoing_args_size = (crtl->outgoing_args_size
3790 + REG_PARM_STACK_SPACE (current_function_decl));
3791 int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */
3796 for (i = 0; i < 8; i++)
3797 if (df_regs_ever_live_p (i) && ! call_used_regs[i])
3802 for (i = 0; i < 8; i += 2)
3803 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
3804 || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1]))
3808 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
3809 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
3810 || (df_regs_ever_live_p (i+1) && ! call_used_regs[i+1]))
3813 /* Set up values for use in prologue and epilogue. */
3814 num_gfregs = n_regs;
3819 && crtl->outgoing_args_size == 0)
3820 actual_fsize = apparent_fsize = 0;
3823 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
3824 apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8;
3825 apparent_fsize += n_regs * 4;
3826 actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8);
3829 /* Make sure nothing can clobber our register windows.
3830 If a SAVE must be done, or there is a stack-local variable,
3831 the register window area must be allocated. */
3832 if (! leaf_function_p || size > 0)
3833 actual_fsize += FIRST_PARM_OFFSET (current_function_decl);
3835 return SPARC_STACK_ALIGN (actual_fsize);
3838 /* Output any necessary .register pseudo-ops. */
3841 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
3843 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
3849 /* Check if %g[2367] were used without
3850 .register being printed for them already. */
3851 for (i = 2; i < 8; i++)
3853 if (df_regs_ever_live_p (i)
3854 && ! sparc_hard_reg_printed [i])
3856 sparc_hard_reg_printed [i] = 1;
3857 /* %g7 is used as TLS base register, use #ignore
3858 for it instead of #scratch. */
3859 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
3860 i == 7 ? "ignore" : "scratch");
3867 /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET
3868 as needed. LOW should be double-word aligned for 32-bit registers.
3869 Return the new OFFSET. */
3872 #define SORR_RESTORE 1
3875 save_or_restore_regs (int low, int high, rtx base, int offset, int action)
3880 if (TARGET_ARCH64 && high <= 32)
3882 for (i = low; i < high; i++)
3884 if (df_regs_ever_live_p (i) && ! call_used_regs[i])
3886 mem = gen_rtx_MEM (DImode, plus_constant (base, offset));
3887 set_mem_alias_set (mem, sparc_sr_alias_set);
3888 if (action == SORR_SAVE)
3890 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
3891 RTX_FRAME_RELATED_P (insn) = 1;
3893 else /* action == SORR_RESTORE */
3894 emit_move_insn (gen_rtx_REG (DImode, i), mem);
3901 for (i = low; i < high; i += 2)
3903 bool reg0 = df_regs_ever_live_p (i) && ! call_used_regs[i];
3904 bool reg1 = df_regs_ever_live_p (i+1) && ! call_used_regs[i+1];
3905 enum machine_mode mode;
3910 mode = i < 32 ? DImode : DFmode;
3915 mode = i < 32 ? SImode : SFmode;
3920 mode = i < 32 ? SImode : SFmode;
3927 mem = gen_rtx_MEM (mode, plus_constant (base, offset));
3928 set_mem_alias_set (mem, sparc_sr_alias_set);
3929 if (action == SORR_SAVE)
3931 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
3932 RTX_FRAME_RELATED_P (insn) = 1;
3934 else /* action == SORR_RESTORE */
3935 emit_move_insn (gen_rtx_REG (mode, regno), mem);
3937 /* Always preserve double-word alignment. */
3938 offset = (offset + 7) & -8;
3945 /* Emit code to save call-saved registers. */
3948 emit_save_or_restore_regs (int action)
3950 HOST_WIDE_INT offset;
3953 offset = frame_base_offset - apparent_fsize;
3955 if (offset < -4096 || offset + num_gfregs * 4 > 4095)
3957 /* ??? This might be optimized a little as %g1 might already have a
3958 value close enough that a single add insn will do. */
3959 /* ??? Although, all of this is probably only a temporary fix
3960 because if %g1 can hold a function result, then
3961 sparc_expand_epilogue will lose (the result will be
3963 base = gen_rtx_REG (Pmode, 1);
3964 emit_move_insn (base, GEN_INT (offset));
3965 emit_insn (gen_rtx_SET (VOIDmode,
3967 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
3971 base = frame_base_reg;
3973 offset = save_or_restore_regs (0, 8, base, offset, action);
3974 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, action);
3977 /* Generate a save_register_window insn. */
3980 gen_save_register_window (rtx increment)
3983 return gen_save_register_windowdi (increment);
3985 return gen_save_register_windowsi (increment);
3988 /* Generate an increment for the stack pointer. */
3991 gen_stack_pointer_inc (rtx increment)
3993 return gen_rtx_SET (VOIDmode,
3995 gen_rtx_PLUS (Pmode,
4000 /* Generate a decrement for the stack pointer. */
4003 gen_stack_pointer_dec (rtx decrement)
4005 return gen_rtx_SET (VOIDmode,
4007 gen_rtx_MINUS (Pmode,
4012 /* Expand the function prologue. The prologue is responsible for reserving
4013 storage for the frame, saving the call-saved registers and loading the
4014 PIC register if needed. */
4017 sparc_expand_prologue (void)
4022 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4023 on the final value of the flag means deferring the prologue/epilogue
4024 expansion until just before the second scheduling pass, which is too
4025 late to emit multiple epilogues or return insns.
4027 Of course we are making the assumption that the value of the flag
4028 will not change between now and its final value. Of the three parts
4029 of the formula, only the last one can reasonably vary. Let's take a
4030 closer look, after assuming that the first two ones are set to true
4031 (otherwise the last value is effectively silenced).
4033 If only_leaf_regs_used returns false, the global predicate will also
4034 be false so the actual frame size calculated below will be positive.
4035 As a consequence, the save_register_window insn will be emitted in
4036 the instruction stream; now this insn explicitly references %fp
4037 which is not a leaf register so only_leaf_regs_used will always
4038 return false subsequently.
4040 If only_leaf_regs_used returns true, we hope that the subsequent
4041 optimization passes won't cause non-leaf registers to pop up. For
4042 example, the regrename pass has special provisions to not rename to
4043 non-leaf registers in a leaf function. */
4044 sparc_leaf_function_p
4045 = optimize > 0 && leaf_function_p () && only_leaf_regs_used ();
4047 /* Need to use actual_fsize, since we are also allocating
4048 space for our callee (and our own register save area). */
4050 = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4052 /* Advertise that the data calculated just above are now valid. */
4053 sparc_prologue_data_valid_p = true;
4055 if (sparc_leaf_function_p)
4057 frame_base_reg = stack_pointer_rtx;
4058 frame_base_offset = actual_fsize + SPARC_STACK_BIAS;
4062 frame_base_reg = hard_frame_pointer_rtx;
4063 frame_base_offset = SPARC_STACK_BIAS;
4066 if (actual_fsize == 0)
4068 else if (sparc_leaf_function_p)
4070 if (actual_fsize <= 4096)
4071 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-actual_fsize)));
4072 else if (actual_fsize <= 8192)
4074 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
4075 /* %sp is still the CFA register. */
4076 RTX_FRAME_RELATED_P (insn) = 1;
4078 = emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize)));
4082 rtx reg = gen_rtx_REG (Pmode, 1);
4083 emit_move_insn (reg, GEN_INT (-actual_fsize));
4084 insn = emit_insn (gen_stack_pointer_inc (reg));
4085 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4086 gen_stack_pointer_inc (GEN_INT (-actual_fsize)));
4089 RTX_FRAME_RELATED_P (insn) = 1;
4093 if (actual_fsize <= 4096)
4094 insn = emit_insn (gen_save_register_window (GEN_INT (-actual_fsize)));
4095 else if (actual_fsize <= 8192)
4097 insn = emit_insn (gen_save_register_window (GEN_INT (-4096)));
4098 /* %sp is not the CFA register anymore. */
4099 emit_insn (gen_stack_pointer_inc (GEN_INT (4096-actual_fsize)));
4103 rtx reg = gen_rtx_REG (Pmode, 1);
4104 emit_move_insn (reg, GEN_INT (-actual_fsize));
4105 insn = emit_insn (gen_save_register_window (reg));
4108 RTX_FRAME_RELATED_P (insn) = 1;
4109 for (i=0; i < XVECLEN (PATTERN (insn), 0); i++)
4110 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, i)) = 1;
4114 emit_save_or_restore_regs (SORR_SAVE);
4116 /* Load the PIC register if needed. */
4117 if (flag_pic && crtl->uses_pic_offset_table)
4118 load_pic_register (false);
4121 /* This function generates the assembly code for function entry, which boils
4122 down to emitting the necessary .register directives. */
4125 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4127 /* Check that the assumption we made in sparc_expand_prologue is valid. */
4128 gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs);
4130 sparc_output_scratch_registers (file);
4133 /* Expand the function epilogue, either normal or part of a sibcall.
4134 We emit all the instructions except the return or the call. */
4137 sparc_expand_epilogue (void)
4140 emit_save_or_restore_regs (SORR_RESTORE);
4142 if (actual_fsize == 0)
4144 else if (sparc_leaf_function_p)
4146 if (actual_fsize <= 4096)
4147 emit_insn (gen_stack_pointer_dec (GEN_INT (- actual_fsize)));
4148 else if (actual_fsize <= 8192)
4150 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
4151 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - actual_fsize)));
4155 rtx reg = gen_rtx_REG (Pmode, 1);
4156 emit_move_insn (reg, GEN_INT (-actual_fsize));
4157 emit_insn (gen_stack_pointer_dec (reg));
4162 /* Return true if it is appropriate to emit `return' instructions in the
4163 body of a function. */
4166 sparc_can_use_return_insn_p (void)
4168 return sparc_prologue_data_valid_p
4169 && (actual_fsize == 0 || !sparc_leaf_function_p);
4172 /* This function generates the assembly code for function exit. */
4175 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4177 /* If code does not drop into the epilogue, we have to still output
4178 a dummy nop for the sake of sane backtraces. Otherwise, if the
4179 last two instructions of a function were "call foo; dslot;" this
4180 can make the return PC of foo (i.e. address of call instruction
4181 plus 8) point to the first instruction in the next function. */
4183 rtx insn, last_real_insn;
4185 insn = get_last_insn ();
4187 last_real_insn = prev_real_insn (insn);
4189 && GET_CODE (last_real_insn) == INSN
4190 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
4191 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
4193 if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
4194 fputs("\tnop\n", file);
4196 sparc_output_deferred_case_vectors ();
4199 /* Output a 'restore' instruction. */
4202 output_restore (rtx pat)
4208 fputs ("\t restore\n", asm_out_file);
4212 gcc_assert (GET_CODE (pat) == SET);
4214 operands[0] = SET_DEST (pat);
4215 pat = SET_SRC (pat);
4217 switch (GET_CODE (pat))
4220 operands[1] = XEXP (pat, 0);
4221 operands[2] = XEXP (pat, 1);
4222 output_asm_insn (" restore %r1, %2, %Y0", operands);
4225 operands[1] = XEXP (pat, 0);
4226 operands[2] = XEXP (pat, 1);
4227 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
4230 operands[1] = XEXP (pat, 0);
4231 gcc_assert (XEXP (pat, 1) == const1_rtx);
4232 output_asm_insn (" restore %r1, %r1, %Y0", operands);
4236 output_asm_insn (" restore %%g0, %1, %Y0", operands);
4241 /* Output a return. */
4244 output_return (rtx insn)
4246 if (sparc_leaf_function_p)
4248 /* This is a leaf function so we don't have to bother restoring the
4249 register window, which frees us from dealing with the convoluted
4250 semantics of restore/return. We simply output the jump to the
4251 return address and the insn in the delay slot (if any). */
4253 gcc_assert (! crtl->calls_eh_return);
4255 return "jmp\t%%o7+%)%#";
4259 /* This is a regular function so we have to restore the register window.
4260 We may have a pending insn for the delay slot, which will be either
4261 combined with the 'restore' instruction or put in the delay slot of
4262 the 'return' instruction. */
4264 if (crtl->calls_eh_return)
4266 /* If the function uses __builtin_eh_return, the eh_return
4267 machinery occupies the delay slot. */
4268 gcc_assert (! final_sequence);
4270 if (! flag_delayed_branch)
4271 fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file);
4274 fputs ("\treturn\t%i7+8\n", asm_out_file);
4276 fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file);
4278 if (flag_delayed_branch)
4279 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
4281 fputs ("\t nop\n", asm_out_file);
4283 else if (final_sequence)
4287 delay = NEXT_INSN (insn);
4290 pat = PATTERN (delay);
4292 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
4294 epilogue_renumber (&pat, 0);
4295 return "return\t%%i7+%)%#";
4299 output_asm_insn ("jmp\t%%i7+%)", NULL);
4300 output_restore (pat);
4301 PATTERN (delay) = gen_blockage ();
4302 INSN_CODE (delay) = -1;
4307 /* The delay slot is empty. */
4309 return "return\t%%i7+%)\n\t nop";
4310 else if (flag_delayed_branch)
4311 return "jmp\t%%i7+%)\n\t restore";
4313 return "restore\n\tjmp\t%%o7+%)\n\t nop";
4320 /* Output a sibling call. */
4323 output_sibcall (rtx insn, rtx call_operand)
4327 gcc_assert (flag_delayed_branch);
4329 operands[0] = call_operand;
4331 if (sparc_leaf_function_p)
4333 /* This is a leaf function so we don't have to bother restoring the
4334 register window. We simply output the jump to the function and
4335 the insn in the delay slot (if any). */
4337 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
4340 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
4343 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
4344 it into branch if possible. */
4345 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
4350 /* This is a regular function so we have to restore the register window.
4351 We may have a pending insn for the delay slot, which will be combined
4352 with the 'restore' instruction. */
4354 output_asm_insn ("call\t%a0, 0", operands);
4358 rtx delay = NEXT_INSN (insn);
4361 output_restore (PATTERN (delay));
4363 PATTERN (delay) = gen_blockage ();
4364 INSN_CODE (delay) = -1;
4367 output_restore (NULL_RTX);
4373 /* Functions for handling argument passing.
4375 For 32-bit, the first 6 args are normally in registers and the rest are
4376 pushed. Any arg that starts within the first 6 words is at least
4377 partially passed in a register unless its data type forbids.
4379 For 64-bit, the argument registers are laid out as an array of 16 elements
4380 and arguments are added sequentially. The first 6 int args and up to the
4381 first 16 fp args (depending on size) are passed in regs.
4383 Slot Stack Integral Float Float in structure Double Long Double
4384 ---- ----- -------- ----- ------------------ ------ -----------
4385 15 [SP+248] %f31 %f30,%f31 %d30
4386 14 [SP+240] %f29 %f28,%f29 %d28 %q28
4387 13 [SP+232] %f27 %f26,%f27 %d26
4388 12 [SP+224] %f25 %f24,%f25 %d24 %q24
4389 11 [SP+216] %f23 %f22,%f23 %d22
4390 10 [SP+208] %f21 %f20,%f21 %d20 %q20
4391 9 [SP+200] %f19 %f18,%f19 %d18
4392 8 [SP+192] %f17 %f16,%f17 %d16 %q16
4393 7 [SP+184] %f15 %f14,%f15 %d14
4394 6 [SP+176] %f13 %f12,%f13 %d12 %q12
4395 5 [SP+168] %o5 %f11 %f10,%f11 %d10
4396 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
4397 3 [SP+152] %o3 %f7 %f6,%f7 %d6
4398 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
4399 1 [SP+136] %o1 %f3 %f2,%f3 %d2
4400 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
4402 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
4404 Integral arguments are always passed as 64-bit quantities appropriately
4407 Passing of floating point values is handled as follows.
4408 If a prototype is in scope:
4409 If the value is in a named argument (i.e. not a stdarg function or a
4410 value not part of the `...') then the value is passed in the appropriate
4412 If the value is part of the `...' and is passed in one of the first 6
4413 slots then the value is passed in the appropriate int reg.
4414 If the value is part of the `...' and is not passed in one of the first 6
4415 slots then the value is passed in memory.
4416 If a prototype is not in scope:
4417 If the value is one of the first 6 arguments the value is passed in the
4418 appropriate integer reg and the appropriate fp reg.
4419 If the value is not one of the first 6 arguments the value is passed in
4420 the appropriate fp reg and in memory.
4423 Summary of the calling conventions implemented by GCC on SPARC:
4426 size argument return value
4428 small integer <4 int. reg. int. reg.
4429 word 4 int. reg. int. reg.
4430 double word 8 int. reg. int. reg.
4432 _Complex small integer <8 int. reg. int. reg.
4433 _Complex word 8 int. reg. int. reg.
4434 _Complex double word 16 memory int. reg.
4436 vector integer <=8 int. reg. FP reg.
4437 vector integer >8 memory memory
4439 float 4 int. reg. FP reg.
4440 double 8 int. reg. FP reg.
4441 long double 16 memory memory
4443 _Complex float 8 memory FP reg.
4444 _Complex double 16 memory FP reg.
4445 _Complex long double 32 memory FP reg.
4447 vector float any memory memory
4449 aggregate any memory memory
4454 size argument return value
4456 small integer <8 int. reg. int. reg.
4457 word 8 int. reg. int. reg.
4458 double word 16 int. reg. int. reg.
4460 _Complex small integer <16 int. reg. int. reg.
4461 _Complex word 16 int. reg. int. reg.
4462 _Complex double word 32 memory int. reg.
4464 vector integer <=16 FP reg. FP reg.
4465 vector integer 16<s<=32 memory FP reg.
4466 vector integer >32 memory memory
4468 float 4 FP reg. FP reg.
4469 double 8 FP reg. FP reg.
4470 long double 16 FP reg. FP reg.
4472 _Complex float 8 FP reg. FP reg.
4473 _Complex double 16 FP reg. FP reg.
4474 _Complex long double 32 memory FP reg.
4476 vector float <=16 FP reg. FP reg.
4477 vector float 16<s<=32 memory FP reg.
4478 vector float >32 memory memory
4480 aggregate <=16 reg. reg.
4481 aggregate 16<s<=32 memory reg.
4482 aggregate >32 memory memory
4486 Note #1: complex floating-point types follow the extended SPARC ABIs as
4487 implemented by the Sun compiler.
4489 Note #2: integral vector types follow the scalar floating-point types
4490 conventions to match what is implemented by the Sun VIS SDK.
4492 Note #3: floating-point vector types follow the aggregate types
4496 /* Maximum number of int regs for args. */
4497 #define SPARC_INT_ARG_MAX 6
4498 /* Maximum number of fp regs for args. */
4499 #define SPARC_FP_ARG_MAX 16
4501 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
4503 /* Handle the INIT_CUMULATIVE_ARGS macro.
4504 Initialize a variable CUM of type CUMULATIVE_ARGS
4505 for a call to a function whose data type is FNTYPE.
4506 For a library call, FNTYPE is 0. */
4509 init_cumulative_args (struct sparc_args *cum, tree fntype,
4510 rtx libname ATTRIBUTE_UNUSED,
4511 tree fndecl ATTRIBUTE_UNUSED)
4514 cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype);
4515 cum->libcall_p = fntype == 0;
4518 /* Handle the TARGET_PROMOTE_PROTOTYPES target hook.
4519 When a prototype says `char' or `short', really pass an `int'. */
4522 sparc_promote_prototypes (const_tree fntype ATTRIBUTE_UNUSED)
4524 return TARGET_ARCH32 ? true : false;
4527 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
4530 sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED)
4532 return TARGET_ARCH64 ? true : false;
4535 /* Scan the record type TYPE and return the following predicates:
4536 - INTREGS_P: the record contains at least one field or sub-field
4537 that is eligible for promotion in integer registers.
4538 - FP_REGS_P: the record contains at least one field or sub-field
4539 that is eligible for promotion in floating-point registers.
4540 - PACKED_P: the record contains at least one field that is packed.
4542 Sub-fields are not taken into account for the PACKED_P predicate. */
4545 scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p)
4549 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4551 if (TREE_CODE (field) == FIELD_DECL)
4553 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4554 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
4555 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
4556 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
4562 if (packed_p && DECL_PACKED (field))
4568 /* Compute the slot number to pass an argument in.
4569 Return the slot number or -1 if passing on the stack.
4571 CUM is a variable of type CUMULATIVE_ARGS which gives info about
4572 the preceding args and about the function being called.
4573 MODE is the argument's machine mode.
4574 TYPE is the data type of the argument (as a tree).
4575 This is null for libcalls where that information may
4577 NAMED is nonzero if this argument is a named parameter
4578 (otherwise it is an extra parameter matching an ellipsis).
4579 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
4580 *PREGNO records the register number to use if scalar type.
4581 *PPADDING records the amount of padding needed in words. */
4584 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
4585 tree type, int named, int incoming_p,
4586 int *pregno, int *ppadding)
4588 int regbase = (incoming_p
4589 ? SPARC_INCOMING_INT_ARG_FIRST
4590 : SPARC_OUTGOING_INT_ARG_FIRST);
4591 int slotno = cum->words;
4592 enum mode_class mclass;
4597 if (type && TREE_ADDRESSABLE (type))
4603 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
4606 /* For SPARC64, objects requiring 16-byte alignment get it. */
4608 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
4609 && (slotno & 1) != 0)
4610 slotno++, *ppadding = 1;
4612 mclass = GET_MODE_CLASS (mode);
4613 if (type && TREE_CODE (type) == VECTOR_TYPE)
4615 /* Vector types deserve special treatment because they are
4616 polymorphic wrt their mode, depending upon whether VIS
4617 instructions are enabled. */
4618 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
4620 /* The SPARC port defines no floating-point vector modes. */
4621 gcc_assert (mode == BLKmode);
4625 /* Integral vector types should either have a vector
4626 mode or an integral mode, because we are guaranteed
4627 by pass_by_reference that their size is not greater
4628 than 16 bytes and TImode is 16-byte wide. */
4629 gcc_assert (mode != BLKmode);
4631 /* Vector integers are handled like floats according to
4633 mclass = MODE_FLOAT;
4640 case MODE_COMPLEX_FLOAT:
4641 case MODE_VECTOR_INT:
4642 if (TARGET_ARCH64 && TARGET_FPU && named)
4644 if (slotno >= SPARC_FP_ARG_MAX)
4646 regno = SPARC_FP_ARG_FIRST + slotno * 2;
4647 /* Arguments filling only one single FP register are
4648 right-justified in the outer double FP register. */
4649 if (GET_MODE_SIZE (mode) <= 4)
4656 case MODE_COMPLEX_INT:
4657 if (slotno >= SPARC_INT_ARG_MAX)
4659 regno = regbase + slotno;
4663 if (mode == VOIDmode)
4664 /* MODE is VOIDmode when generating the actual call. */
4667 gcc_assert (mode == BLKmode);
4671 || (TREE_CODE (type) != VECTOR_TYPE
4672 && TREE_CODE (type) != RECORD_TYPE))
4674 if (slotno >= SPARC_INT_ARG_MAX)
4676 regno = regbase + slotno;
4678 else /* TARGET_ARCH64 && type */
4680 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
4682 /* First see what kinds of registers we would need. */
4683 if (TREE_CODE (type) == VECTOR_TYPE)
4686 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
4688 /* The ABI obviously doesn't specify how packed structures
4689 are passed. These are defined to be passed in int regs
4690 if possible, otherwise memory. */
4691 if (packed_p || !named)
4692 fpregs_p = 0, intregs_p = 1;
4694 /* If all arg slots are filled, then must pass on stack. */
4695 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
4698 /* If there are only int args and all int arg slots are filled,
4699 then must pass on stack. */
4700 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
4703 /* Note that even if all int arg slots are filled, fp members may
4704 still be passed in regs if such regs are available.
4705 *PREGNO isn't set because there may be more than one, it's up
4706 to the caller to compute them. */
4719 /* Handle recursive register counting for structure field layout. */
4721 struct function_arg_record_value_parms
4723 rtx ret; /* return expression being built. */
4724 int slotno; /* slot number of the argument. */
4725 int named; /* whether the argument is named. */
4726 int regbase; /* regno of the base register. */
4727 int stack; /* 1 if part of the argument is on the stack. */
4728 int intoffset; /* offset of the first pending integer field. */
4729 unsigned int nregs; /* number of words passed in registers. */
4732 static void function_arg_record_value_3
4733 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
4734 static void function_arg_record_value_2
4735 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
4736 static void function_arg_record_value_1
4737 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
4738 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
4739 static rtx function_arg_union_value (int, enum machine_mode, int, int);
4741 /* A subroutine of function_arg_record_value. Traverse the structure
4742 recursively and determine how many registers will be required. */
4745 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
4746 struct function_arg_record_value_parms *parms,
4751 /* We need to compute how many registers are needed so we can
4752 allocate the PARALLEL but before we can do that we need to know
4753 whether there are any packed fields. The ABI obviously doesn't
4754 specify how structures are passed in this case, so they are
4755 defined to be passed in int regs if possible, otherwise memory,
4756 regardless of whether there are fp values present. */
4759 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4761 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
4768 /* Compute how many registers we need. */
4769 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4771 if (TREE_CODE (field) == FIELD_DECL)
4773 HOST_WIDE_INT bitpos = startbitpos;
4775 if (DECL_SIZE (field) != 0)
4777 if (integer_zerop (DECL_SIZE (field)))
4780 if (host_integerp (bit_position (field), 1))
4781 bitpos += int_bit_position (field);
4784 /* ??? FIXME: else assume zero offset. */
4786 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4787 function_arg_record_value_1 (TREE_TYPE (field),
4791 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
4792 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
4797 if (parms->intoffset != -1)
4799 unsigned int startbit, endbit;
4800 int intslots, this_slotno;
4802 startbit = parms->intoffset & -BITS_PER_WORD;
4803 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
4805 intslots = (endbit - startbit) / BITS_PER_WORD;
4806 this_slotno = parms->slotno + parms->intoffset
4809 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
4811 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
4812 /* We need to pass this field on the stack. */
4816 parms->nregs += intslots;
4817 parms->intoffset = -1;
4820 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
4821 If it wasn't true we wouldn't be here. */
4822 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
4823 && DECL_MODE (field) == BLKmode)
4824 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
4825 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
4832 if (parms->intoffset == -1)
4833 parms->intoffset = bitpos;
4839 /* A subroutine of function_arg_record_value. Assign the bits of the
4840 structure between parms->intoffset and bitpos to integer registers. */
4843 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
4844 struct function_arg_record_value_parms *parms)
4846 enum machine_mode mode;
4848 unsigned int startbit, endbit;
4849 int this_slotno, intslots, intoffset;
4852 if (parms->intoffset == -1)
4855 intoffset = parms->intoffset;
4856 parms->intoffset = -1;
4858 startbit = intoffset & -BITS_PER_WORD;
4859 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
4860 intslots = (endbit - startbit) / BITS_PER_WORD;
4861 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
4863 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
4867 /* If this is the trailing part of a word, only load that much into
4868 the register. Otherwise load the whole register. Note that in
4869 the latter case we may pick up unwanted bits. It's not a problem
4870 at the moment but may wish to revisit. */
4872 if (intoffset % BITS_PER_WORD != 0)
4873 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
4878 intoffset /= BITS_PER_UNIT;
4881 regno = parms->regbase + this_slotno;
4882 reg = gen_rtx_REG (mode, regno);
4883 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
4884 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
4887 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
4892 while (intslots > 0);
4895 /* A subroutine of function_arg_record_value. Traverse the structure
4896 recursively and assign bits to floating point registers. Track which
4897 bits in between need integer registers; invoke function_arg_record_value_3
4898 to make that happen. */
4901 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
4902 struct function_arg_record_value_parms *parms,
4908 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4910 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
4917 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4919 if (TREE_CODE (field) == FIELD_DECL)
4921 HOST_WIDE_INT bitpos = startbitpos;
4923 if (DECL_SIZE (field) != 0)
4925 if (integer_zerop (DECL_SIZE (field)))
4928 if (host_integerp (bit_position (field), 1))
4929 bitpos += int_bit_position (field);
4932 /* ??? FIXME: else assume zero offset. */
4934 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4935 function_arg_record_value_2 (TREE_TYPE (field),
4939 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
4940 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
4945 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
4946 int regno, nregs, pos;
4947 enum machine_mode mode = DECL_MODE (field);
4950 function_arg_record_value_3 (bitpos, parms);
4952 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
4955 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
4956 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
4958 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
4960 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
4966 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
4967 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
4969 reg = gen_rtx_REG (mode, regno);
4970 pos = bitpos / BITS_PER_UNIT;
4971 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
4972 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
4976 regno += GET_MODE_SIZE (mode) / 4;
4977 reg = gen_rtx_REG (mode, regno);
4978 pos += GET_MODE_SIZE (mode);
4979 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
4980 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
4986 if (parms->intoffset == -1)
4987 parms->intoffset = bitpos;
4993 /* Used by function_arg and function_value to implement the complex
4994 conventions of the 64-bit ABI for passing and returning structures.
4995 Return an expression valid as a return value for the two macros
4996 FUNCTION_ARG and FUNCTION_VALUE.
4998 TYPE is the data type of the argument (as a tree).
4999 This is null for libcalls where that information may
5001 MODE is the argument's machine mode.
5002 SLOTNO is the index number of the argument's slot in the parameter array.
5003 NAMED is nonzero if this argument is a named parameter
5004 (otherwise it is an extra parameter matching an ellipsis).
5005 REGBASE is the regno of the base register for the parameter array. */
5008 function_arg_record_value (const_tree type, enum machine_mode mode,
5009 int slotno, int named, int regbase)
5011 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5012 struct function_arg_record_value_parms parms;
5015 parms.ret = NULL_RTX;
5016 parms.slotno = slotno;
5017 parms.named = named;
5018 parms.regbase = regbase;
5021 /* Compute how many registers we need. */
5023 parms.intoffset = 0;
5024 function_arg_record_value_1 (type, 0, &parms, false);
5026 /* Take into account pending integer fields. */
5027 if (parms.intoffset != -1)
5029 unsigned int startbit, endbit;
5030 int intslots, this_slotno;
5032 startbit = parms.intoffset & -BITS_PER_WORD;
5033 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5034 intslots = (endbit - startbit) / BITS_PER_WORD;
5035 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
5037 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5039 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5040 /* We need to pass this field on the stack. */
5044 parms.nregs += intslots;
5046 nregs = parms.nregs;
5048 /* Allocate the vector and handle some annoying special cases. */
5051 /* ??? Empty structure has no value? Duh? */
5054 /* Though there's nothing really to store, return a word register
5055 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
5056 leads to breakage due to the fact that there are zero bytes to
5058 return gen_rtx_REG (mode, regbase);
5062 /* ??? C++ has structures with no fields, and yet a size. Give up
5063 for now and pass everything back in integer registers. */
5064 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5066 if (nregs + slotno > SPARC_INT_ARG_MAX)
5067 nregs = SPARC_INT_ARG_MAX - slotno;
5069 gcc_assert (nregs != 0);
5071 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
5073 /* If at least one field must be passed on the stack, generate
5074 (parallel [(expr_list (nil) ...) ...]) so that all fields will
5075 also be passed on the stack. We can't do much better because the
5076 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
5077 of structures for which the fields passed exclusively in registers
5078 are not at the beginning of the structure. */
5080 XVECEXP (parms.ret, 0, 0)
5081 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5083 /* Fill in the entries. */
5085 parms.intoffset = 0;
5086 function_arg_record_value_2 (type, 0, &parms, false);
5087 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
5089 gcc_assert (parms.nregs == nregs);
5094 /* Used by function_arg and function_value to implement the conventions
5095 of the 64-bit ABI for passing and returning unions.
5096 Return an expression valid as a return value for the two macros
5097 FUNCTION_ARG and FUNCTION_VALUE.
5099 SIZE is the size in bytes of the union.
5100 MODE is the argument's machine mode.
5101 REGNO is the hard register the union will be passed in. */
5104 function_arg_union_value (int size, enum machine_mode mode, int slotno,
5107 int nwords = ROUND_ADVANCE (size), i;
5110 /* See comment in previous function for empty structures. */
5112 return gen_rtx_REG (mode, regno);
5114 if (slotno == SPARC_INT_ARG_MAX - 1)
5117 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
5119 for (i = 0; i < nwords; i++)
5121 /* Unions are passed left-justified. */
5122 XVECEXP (regs, 0, i)
5123 = gen_rtx_EXPR_LIST (VOIDmode,
5124 gen_rtx_REG (word_mode, regno),
5125 GEN_INT (UNITS_PER_WORD * i));
5132 /* Used by function_arg and function_value to implement the conventions
5133 for passing and returning large (BLKmode) vectors.
5134 Return an expression valid as a return value for the two macros
5135 FUNCTION_ARG and FUNCTION_VALUE.
5137 SIZE is the size in bytes of the vector (at least 8 bytes).
5138 REGNO is the FP hard register the vector will be passed in. */
5141 function_arg_vector_value (int size, int regno)
5143 int i, nregs = size / 8;
5146 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
5148 for (i = 0; i < nregs; i++)
5150 XVECEXP (regs, 0, i)
5151 = gen_rtx_EXPR_LIST (VOIDmode,
5152 gen_rtx_REG (DImode, regno + 2*i),
5159 /* Handle the FUNCTION_ARG macro.
5160 Determine where to put an argument to a function.
5161 Value is zero to push the argument on the stack,
5162 or a hard register in which to store the argument.
5164 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5165 the preceding args and about the function being called.
5166 MODE is the argument's machine mode.
5167 TYPE is the data type of the argument (as a tree).
5168 This is null for libcalls where that information may
5170 NAMED is nonzero if this argument is a named parameter
5171 (otherwise it is an extra parameter matching an ellipsis).
5172 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */
5175 function_arg (const struct sparc_args *cum, enum machine_mode mode,
5176 tree type, int named, int incoming_p)
5178 int regbase = (incoming_p
5179 ? SPARC_INCOMING_INT_ARG_FIRST
5180 : SPARC_OUTGOING_INT_ARG_FIRST);
5181 int slotno, regno, padding;
5182 enum mode_class mclass = GET_MODE_CLASS (mode);
5184 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
5189 /* Vector types deserve special treatment because they are polymorphic wrt
5190 their mode, depending upon whether VIS instructions are enabled. */
5191 if (type && TREE_CODE (type) == VECTOR_TYPE)
5193 HOST_WIDE_INT size = int_size_in_bytes (type);
5194 gcc_assert ((TARGET_ARCH32 && size <= 8)
5195 || (TARGET_ARCH64 && size <= 16));
5197 if (mode == BLKmode)
5198 return function_arg_vector_value (size,
5199 SPARC_FP_ARG_FIRST + 2*slotno);
5201 mclass = MODE_FLOAT;
5205 return gen_rtx_REG (mode, regno);
5207 /* Structures up to 16 bytes in size are passed in arg slots on the stack
5208 and are promoted to registers if possible. */
5209 if (type && TREE_CODE (type) == RECORD_TYPE)
5211 HOST_WIDE_INT size = int_size_in_bytes (type);
5212 gcc_assert (size <= 16);
5214 return function_arg_record_value (type, mode, slotno, named, regbase);
5217 /* Unions up to 16 bytes in size are passed in integer registers. */
5218 else if (type && TREE_CODE (type) == UNION_TYPE)
5220 HOST_WIDE_INT size = int_size_in_bytes (type);
5221 gcc_assert (size <= 16);
5223 return function_arg_union_value (size, mode, slotno, regno);
5226 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
5227 but also have the slot allocated for them.
5228 If no prototype is in scope fp values in register slots get passed
5229 in two places, either fp regs and int regs or fp regs and memory. */
5230 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5231 && SPARC_FP_REG_P (regno))
5233 rtx reg = gen_rtx_REG (mode, regno);
5234 if (cum->prototype_p || cum->libcall_p)
5236 /* "* 2" because fp reg numbers are recorded in 4 byte
5239 /* ??? This will cause the value to be passed in the fp reg and
5240 in the stack. When a prototype exists we want to pass the
5241 value in the reg but reserve space on the stack. That's an
5242 optimization, and is deferred [for a bit]. */
5243 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
5244 return gen_rtx_PARALLEL (mode,
5246 gen_rtx_EXPR_LIST (VOIDmode,
5247 NULL_RTX, const0_rtx),
5248 gen_rtx_EXPR_LIST (VOIDmode,
5252 /* ??? It seems that passing back a register even when past
5253 the area declared by REG_PARM_STACK_SPACE will allocate
5254 space appropriately, and will not copy the data onto the
5255 stack, exactly as we desire.
5257 This is due to locate_and_pad_parm being called in
5258 expand_call whenever reg_parm_stack_space > 0, which
5259 while beneficial to our example here, would seem to be
5260 in error from what had been intended. Ho hum... -- r~ */
5268 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
5272 /* On incoming, we don't need to know that the value
5273 is passed in %f0 and %i0, and it confuses other parts
5274 causing needless spillage even on the simplest cases. */
5278 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
5279 + (regno - SPARC_FP_ARG_FIRST) / 2);
5281 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5282 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
5284 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5288 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5289 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5290 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5295 /* All other aggregate types are passed in an integer register in a mode
5296 corresponding to the size of the type. */
5297 else if (type && AGGREGATE_TYPE_P (type))
5299 HOST_WIDE_INT size = int_size_in_bytes (type);
5300 gcc_assert (size <= 16);
5302 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
5305 return gen_rtx_REG (mode, regno);
5308 /* For an arg passed partly in registers and partly in memory,
5309 this is the number of bytes of registers used.
5310 For args passed entirely in registers or entirely in memory, zero.
5312 Any arg that starts in the first 6 regs but won't entirely fit in them
5313 needs partial registers on v8. On v9, structures with integer
5314 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
5315 values that begin in the last fp reg [where "last fp reg" varies with the
5316 mode] will be split between that reg and memory. */
5319 sparc_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5320 tree type, bool named)
5322 int slotno, regno, padding;
5324 /* We pass 0 for incoming_p here, it doesn't matter. */
5325 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5332 if ((slotno + (mode == BLKmode
5333 ? ROUND_ADVANCE (int_size_in_bytes (type))
5334 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
5335 > SPARC_INT_ARG_MAX)
5336 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
5340 /* We are guaranteed by pass_by_reference that the size of the
5341 argument is not greater than 16 bytes, so we only need to return
5342 one word if the argument is partially passed in registers. */
5344 if (type && AGGREGATE_TYPE_P (type))
5346 int size = int_size_in_bytes (type);
5348 if (size > UNITS_PER_WORD
5349 && slotno == SPARC_INT_ARG_MAX - 1)
5350 return UNITS_PER_WORD;
5352 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
5353 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5354 && ! (TARGET_FPU && named)))
5356 /* The complex types are passed as packed types. */
5357 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
5358 && slotno == SPARC_INT_ARG_MAX - 1)
5359 return UNITS_PER_WORD;
5361 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
5363 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
5365 return UNITS_PER_WORD;
5372 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
5373 Specify whether to pass the argument by reference. */
5376 sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5377 enum machine_mode mode, const_tree type,
5378 bool named ATTRIBUTE_UNUSED)
5381 /* Original SPARC 32-bit ABI says that structures and unions,
5382 and quad-precision floats are passed by reference. For Pascal,
5383 also pass arrays by reference. All other base types are passed
5386 Extended ABI (as implemented by the Sun compiler) says that all
5387 complex floats are passed by reference. Pass complex integers
5388 in registers up to 8 bytes. More generally, enforce the 2-word
5389 cap for passing arguments in registers.
5391 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5392 integers are passed like floats of the same size, that is in
5393 registers up to 8 bytes. Pass all vector floats by reference
5394 like structure and unions. */
5395 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
5397 /* Catch CDImode, TFmode, DCmode and TCmode. */
5398 || GET_MODE_SIZE (mode) > 8
5400 && TREE_CODE (type) == VECTOR_TYPE
5401 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5403 /* Original SPARC 64-bit ABI says that structures and unions
5404 smaller than 16 bytes are passed in registers, as well as
5405 all other base types.
5407 Extended ABI (as implemented by the Sun compiler) says that
5408 complex floats are passed in registers up to 16 bytes. Pass
5409 all complex integers in registers up to 16 bytes. More generally,
5410 enforce the 2-word cap for passing arguments in registers.
5412 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5413 integers are passed like floats of the same size, that is in
5414 registers (up to 16 bytes). Pass all vector floats like structure
5417 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
5418 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
5419 /* Catch CTImode and TCmode. */
5420 || GET_MODE_SIZE (mode) > 16);
5423 /* Handle the FUNCTION_ARG_ADVANCE macro.
5424 Update the data in CUM to advance over an argument
5425 of mode MODE and data type TYPE.
5426 TYPE is null for libcalls where that information may not be available. */
5429 function_arg_advance (struct sparc_args *cum, enum machine_mode mode,
5430 tree type, int named)
5432 int slotno, regno, padding;
5434 /* We pass 0 for incoming_p here, it doesn't matter. */
5435 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5437 /* If register required leading padding, add it. */
5439 cum->words += padding;
5443 cum->words += (mode != BLKmode
5444 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5445 : ROUND_ADVANCE (int_size_in_bytes (type)));
5449 if (type && AGGREGATE_TYPE_P (type))
5451 int size = int_size_in_bytes (type);
5455 else if (size <= 16)
5457 else /* passed by reference */
5462 cum->words += (mode != BLKmode
5463 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5464 : ROUND_ADVANCE (int_size_in_bytes (type)));
5469 /* Handle the FUNCTION_ARG_PADDING macro.
5470 For the 64 bit ABI structs are always stored left shifted in their
5474 function_arg_padding (enum machine_mode mode, const_tree type)
5476 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
5479 /* Fall back to the default. */
5480 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
5483 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
5484 Specify whether to return the return value in memory. */
5487 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
5490 /* Original SPARC 32-bit ABI says that structures and unions,
5491 and quad-precision floats are returned in memory. All other
5492 base types are returned in registers.
5494 Extended ABI (as implemented by the Sun compiler) says that
5495 all complex floats are returned in registers (8 FP registers
5496 at most for '_Complex long double'). Return all complex integers
5497 in registers (4 at most for '_Complex long long').
5499 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5500 integers are returned like floats of the same size, that is in
5501 registers up to 8 bytes and in memory otherwise. Return all
5502 vector floats in memory like structure and unions; note that
5503 they always have BLKmode like the latter. */
5504 return (TYPE_MODE (type) == BLKmode
5505 || TYPE_MODE (type) == TFmode
5506 || (TREE_CODE (type) == VECTOR_TYPE
5507 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5509 /* Original SPARC 64-bit ABI says that structures and unions
5510 smaller than 32 bytes are returned in registers, as well as
5511 all other base types.
5513 Extended ABI (as implemented by the Sun compiler) says that all
5514 complex floats are returned in registers (8 FP registers at most
5515 for '_Complex long double'). Return all complex integers in
5516 registers (4 at most for '_Complex TItype').
5518 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5519 integers are returned like floats of the same size, that is in
5520 registers. Return all vector floats like structure and unions;
5521 note that they always have BLKmode like the latter. */
5522 return ((TYPE_MODE (type) == BLKmode
5523 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32));
5526 /* Handle the TARGET_STRUCT_VALUE target hook.
5527 Return where to find the structure return value address. */
5530 sparc_struct_value_rtx (tree fndecl, int incoming)
5539 mem = gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx,
5540 STRUCT_VALUE_OFFSET));
5542 mem = gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx,
5543 STRUCT_VALUE_OFFSET));
5545 /* Only follow the SPARC ABI for fixed-size structure returns.
5546 Variable size structure returns are handled per the normal
5547 procedures in GCC. This is enabled by -mstd-struct-return */
5549 && sparc_std_struct_return
5550 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
5551 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
5553 /* We must check and adjust the return address, as it is
5554 optional as to whether the return object is really
5556 rtx ret_rtx = gen_rtx_REG (Pmode, 31);
5557 rtx scratch = gen_reg_rtx (SImode);
5558 rtx endlab = gen_label_rtx ();
5560 /* Calculate the return object size */
5561 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
5562 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
5563 /* Construct a temporary return value */
5564 rtx temp_val = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
5566 /* Implement SPARC 32-bit psABI callee returns struck checking
5569 Fetch the instruction where we will return to and see if
5570 it's an unimp instruction (the most significant 10 bits
5572 emit_move_insn (scratch, gen_rtx_MEM (SImode,
5573 plus_constant (ret_rtx, 8)));
5574 /* Assume the size is valid and pre-adjust */
5575 emit_insn (gen_add3_insn (ret_rtx, ret_rtx, GEN_INT (4)));
5576 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode, 0, endlab);
5577 emit_insn (gen_sub3_insn (ret_rtx, ret_rtx, GEN_INT (4)));
5578 /* Assign stack temp:
5579 Write the address of the memory pointed to by temp_val into
5580 the memory pointed to by mem */
5581 emit_move_insn (mem, XEXP (temp_val, 0));
5582 emit_label (endlab);
5585 set_mem_alias_set (mem, struct_value_alias_set);
5590 /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros.
5591 For v9, function return values are subject to the same rules as arguments,
5592 except that up to 32 bytes may be returned in registers. */
5595 function_value (const_tree type, enum machine_mode mode, int incoming_p)
5597 /* Beware that the two values are swapped here wrt function_arg. */
5598 int regbase = (incoming_p
5599 ? SPARC_OUTGOING_INT_ARG_FIRST
5600 : SPARC_INCOMING_INT_ARG_FIRST);
5601 enum mode_class mclass = GET_MODE_CLASS (mode);
5604 /* Vector types deserve special treatment because they are polymorphic wrt
5605 their mode, depending upon whether VIS instructions are enabled. */
5606 if (type && TREE_CODE (type) == VECTOR_TYPE)
5608 HOST_WIDE_INT size = int_size_in_bytes (type);
5609 gcc_assert ((TARGET_ARCH32 && size <= 8)
5610 || (TARGET_ARCH64 && size <= 32));
5612 if (mode == BLKmode)
5613 return function_arg_vector_value (size,
5614 SPARC_FP_ARG_FIRST);
5616 mclass = MODE_FLOAT;
5619 if (TARGET_ARCH64 && type)
5621 /* Structures up to 32 bytes in size are returned in registers. */
5622 if (TREE_CODE (type) == RECORD_TYPE)
5624 HOST_WIDE_INT size = int_size_in_bytes (type);
5625 gcc_assert (size <= 32);
5627 return function_arg_record_value (type, mode, 0, 1, regbase);
5630 /* Unions up to 32 bytes in size are returned in integer registers. */
5631 else if (TREE_CODE (type) == UNION_TYPE)
5633 HOST_WIDE_INT size = int_size_in_bytes (type);
5634 gcc_assert (size <= 32);
5636 return function_arg_union_value (size, mode, 0, regbase);
5639 /* Objects that require it are returned in FP registers. */
5640 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5643 /* All other aggregate types are returned in an integer register in a
5644 mode corresponding to the size of the type. */
5645 else if (AGGREGATE_TYPE_P (type))
5647 /* All other aggregate types are passed in an integer register
5648 in a mode corresponding to the size of the type. */
5649 HOST_WIDE_INT size = int_size_in_bytes (type);
5650 gcc_assert (size <= 32);
5652 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
5654 /* ??? We probably should have made the same ABI change in
5655 3.4.0 as the one we made for unions. The latter was
5656 required by the SCD though, while the former is not
5657 specified, so we favored compatibility and efficiency.
5659 Now we're stuck for aggregates larger than 16 bytes,
5660 because OImode vanished in the meantime. Let's not
5661 try to be unduly clever, and simply follow the ABI
5662 for unions in that case. */
5663 if (mode == BLKmode)
5664 return function_arg_union_value (size, mode, 0, regbase);
5669 /* This must match PROMOTE_FUNCTION_MODE. */
5670 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5674 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
5675 regno = SPARC_FP_ARG_FIRST;
5679 return gen_rtx_REG (mode, regno);
5682 /* Do what is necessary for `va_start'. We look at the current function
5683 to determine if stdarg or varargs is used and return the address of
5684 the first unnamed parameter. */
5687 sparc_builtin_saveregs (void)
5689 int first_reg = crtl->args.info.words;
5693 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
5694 emit_move_insn (gen_rtx_MEM (word_mode,
5695 gen_rtx_PLUS (Pmode,
5697 GEN_INT (FIRST_PARM_OFFSET (0)
5700 gen_rtx_REG (word_mode,
5701 SPARC_INCOMING_INT_ARG_FIRST + regno));
5703 address = gen_rtx_PLUS (Pmode,
5705 GEN_INT (FIRST_PARM_OFFSET (0)
5706 + UNITS_PER_WORD * first_reg));
5711 /* Implement `va_start' for stdarg. */
5714 sparc_va_start (tree valist, rtx nextarg)
5716 nextarg = expand_builtin_saveregs ();
5717 std_expand_builtin_va_start (valist, nextarg);
5720 /* Implement `va_arg' for stdarg. */
5723 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
5726 HOST_WIDE_INT size, rsize, align;
5729 tree ptrtype = build_pointer_type (type);
5731 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
5734 size = rsize = UNITS_PER_WORD;
5740 size = int_size_in_bytes (type);
5741 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
5746 /* For SPARC64, objects requiring 16-byte alignment get it. */
5747 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
5748 align = 2 * UNITS_PER_WORD;
5750 /* SPARC-V9 ABI states that structures up to 16 bytes in size
5751 are left-justified in their slots. */
5752 if (AGGREGATE_TYPE_P (type))
5755 size = rsize = UNITS_PER_WORD;
5765 incr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr,
5766 size_int (align - 1));
5767 incr = fold_convert (sizetype, incr);
5768 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
5770 incr = fold_convert (ptr_type_node, incr);
5773 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
5776 if (BYTES_BIG_ENDIAN && size < rsize)
5777 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr,
5778 size_int (rsize - size));
5782 addr = fold_convert (build_pointer_type (ptrtype), addr);
5783 addr = build_va_arg_indirect_ref (addr);
5785 /* If the address isn't aligned properly for the type,
5786 we may need to copy to a temporary.
5787 FIXME: This is inefficient. Usually we can do this
5790 && TYPE_ALIGN (type) > BITS_PER_WORD)
5792 tree tmp = create_tmp_var (type, "va_arg_tmp");
5793 tree dest_addr = build_fold_addr_expr (tmp);
5795 tree copy = build_call_expr (implicit_built_in_decls[BUILT_IN_MEMCPY], 3,
5800 gimplify_and_add (copy, pre_p);
5804 addr = fold_convert (ptrtype, addr);
5806 incr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, incr, size_int (rsize));
5807 gimplify_assign (valist, incr, post_p);
5809 return build_va_arg_indirect_ref (addr);
5812 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
5813 Specify whether the vector mode is supported by the hardware. */
5816 sparc_vector_mode_supported_p (enum machine_mode mode)
5818 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
5821 /* Return the string to output an unconditional branch to LABEL, which is
5822 the operand number of the label.
5824 DEST is the destination insn (i.e. the label), INSN is the source. */
5827 output_ubranch (rtx dest, int label, rtx insn)
5829 static char string[64];
5830 bool v9_form = false;
5833 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
5835 int delta = (INSN_ADDRESSES (INSN_UID (dest))
5836 - INSN_ADDRESSES (INSN_UID (insn)));
5837 /* Leave some instructions for "slop". */
5838 if (delta >= -260000 && delta < 260000)
5843 strcpy (string, "ba%*,pt\t%%xcc, ");
5845 strcpy (string, "b%*\t");
5847 p = strchr (string, '\0');
5858 /* Return the string to output a conditional branch to LABEL, which is
5859 the operand number of the label. OP is the conditional expression.
5860 XEXP (OP, 0) is assumed to be a condition code register (integer or
5861 floating point) and its mode specifies what kind of comparison we made.
5863 DEST is the destination insn (i.e. the label), INSN is the source.
5865 REVERSED is nonzero if we should reverse the sense of the comparison.
5867 ANNUL is nonzero if we should generate an annulling branch. */
5870 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
5873 static char string[64];
5874 enum rtx_code code = GET_CODE (op);
5875 rtx cc_reg = XEXP (op, 0);
5876 enum machine_mode mode = GET_MODE (cc_reg);
5877 const char *labelno, *branch;
5878 int spaces = 8, far;
5881 /* v9 branches are limited to +-1MB. If it is too far away,
5894 fbne,a,pn %fcc2, .LC29
5902 far = TARGET_V9 && (get_attr_length (insn) >= 3);
5905 /* Reversal of FP compares takes care -- an ordered compare
5906 becomes an unordered compare and vice versa. */
5907 if (mode == CCFPmode || mode == CCFPEmode)
5908 code = reverse_condition_maybe_unordered (code);
5910 code = reverse_condition (code);
5913 /* Start by writing the branch condition. */
5914 if (mode == CCFPmode || mode == CCFPEmode)
5965 /* ??? !v9: FP branches cannot be preceded by another floating point
5966 insn. Because there is currently no concept of pre-delay slots,
5967 we can fix this only by always emitting a nop before a floating
5972 strcpy (string, "nop\n\t");
5973 strcat (string, branch);
5986 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
5998 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6019 strcpy (string, branch);
6021 spaces -= strlen (branch);
6022 p = strchr (string, '\0');
6024 /* Now add the annulling, the label, and a possible noop. */
6037 if (! far && insn && INSN_ADDRESSES_SET_P ())
6039 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6040 - INSN_ADDRESSES (INSN_UID (insn)));
6041 /* Leave some instructions for "slop". */
6042 if (delta < -260000 || delta >= 260000)
6046 if (mode == CCFPmode || mode == CCFPEmode)
6048 static char v9_fcc_labelno[] = "%%fccX, ";
6049 /* Set the char indicating the number of the fcc reg to use. */
6050 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
6051 labelno = v9_fcc_labelno;
6054 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
6058 else if (mode == CCXmode || mode == CCX_NOOVmode)
6060 labelno = "%%xcc, ";
6065 labelno = "%%icc, ";
6070 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6073 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6086 strcpy (p, labelno);
6087 p = strchr (p, '\0');
6090 strcpy (p, ".+12\n\t nop\n\tb\t");
6091 /* Skip the next insn if requested or
6092 if we know that it will be a nop. */
6093 if (annul || ! final_sequence)
6107 /* Emit a library call comparison between floating point X and Y.
6108 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
6109 Return the new operator to be used in the comparison sequence.
6111 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
6112 values as arguments instead of the TFmode registers themselves,
6113 that's why we cannot call emit_float_lib_cmp. */
6116 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
6119 rtx slot0, slot1, result, tem, tem2;
6120 enum machine_mode mode;
6121 enum rtx_code new_comparison;
6126 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
6130 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
6134 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
6138 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
6142 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
6146 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
6157 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
6170 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6171 emit_move_insn (slot0, x);
6178 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6179 emit_move_insn (slot1, y);
6182 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6184 XEXP (slot0, 0), Pmode,
6185 XEXP (slot1, 0), Pmode);
6190 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6192 x, TFmode, y, TFmode);
6197 /* Immediately move the result of the libcall into a pseudo
6198 register so reload doesn't clobber the value if it needs
6199 the return register for a spill reg. */
6200 result = gen_reg_rtx (mode);
6201 emit_move_insn (result, hard_libcall_value (mode));
6206 new_comparison = NE;
6207 emit_cmp_insn (result, const0_rtx, new_comparison, NULL_RTX, mode, 0);
6211 new_comparison = (comparison == UNORDERED ? EQ : NE);
6212 emit_cmp_insn (result, GEN_INT(3), new_comparison, NULL_RTX, mode, 0);
6216 new_comparison = (comparison == UNGT ? GT : NE);
6217 emit_cmp_insn (result, const1_rtx, new_comparison, NULL_RTX, mode, 0);
6220 new_comparison = NE;
6221 emit_cmp_insn (result, const2_rtx, new_comparison, NULL_RTX, mode, 0);
6224 tem = gen_reg_rtx (mode);
6226 emit_insn (gen_andsi3 (tem, result, const1_rtx));
6228 emit_insn (gen_anddi3 (tem, result, const1_rtx));
6229 new_comparison = NE;
6230 emit_cmp_insn (tem, const0_rtx, new_comparison, NULL_RTX, mode, 0);
6234 tem = gen_reg_rtx (mode);
6236 emit_insn (gen_addsi3 (tem, result, const1_rtx));
6238 emit_insn (gen_adddi3 (tem, result, const1_rtx));
6239 tem2 = gen_reg_rtx (mode);
6241 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
6243 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
6244 new_comparison = (comparison == UNEQ ? EQ : NE);
6245 emit_cmp_insn (tem2, const0_rtx, new_comparison, NULL_RTX, mode, 0);
6249 return new_comparison;
6252 /* Generate an unsigned DImode to FP conversion. This is the same code
6253 optabs would emit if we didn't have TFmode patterns. */
6256 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
6258 rtx neglab, donelab, i0, i1, f0, in, out;
6261 in = force_reg (DImode, operands[1]);
6262 neglab = gen_label_rtx ();
6263 donelab = gen_label_rtx ();
6264 i0 = gen_reg_rtx (DImode);
6265 i1 = gen_reg_rtx (DImode);
6266 f0 = gen_reg_rtx (mode);
6268 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
6270 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
6271 emit_jump_insn (gen_jump (donelab));
6274 emit_label (neglab);
6276 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
6277 emit_insn (gen_anddi3 (i1, in, const1_rtx));
6278 emit_insn (gen_iordi3 (i0, i0, i1));
6279 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
6280 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
6282 emit_label (donelab);
6285 /* Generate an FP to unsigned DImode conversion. This is the same code
6286 optabs would emit if we didn't have TFmode patterns. */
6289 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
6291 rtx neglab, donelab, i0, i1, f0, in, out, limit;
6294 in = force_reg (mode, operands[1]);
6295 neglab = gen_label_rtx ();
6296 donelab = gen_label_rtx ();
6297 i0 = gen_reg_rtx (DImode);
6298 i1 = gen_reg_rtx (DImode);
6299 limit = gen_reg_rtx (mode);
6300 f0 = gen_reg_rtx (mode);
6302 emit_move_insn (limit,
6303 CONST_DOUBLE_FROM_REAL_VALUE (
6304 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
6305 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
6307 emit_insn (gen_rtx_SET (VOIDmode,
6309 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
6310 emit_jump_insn (gen_jump (donelab));
6313 emit_label (neglab);
6315 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
6316 emit_insn (gen_rtx_SET (VOIDmode,
6318 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
6319 emit_insn (gen_movdi (i1, const1_rtx));
6320 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
6321 emit_insn (gen_xordi3 (out, i0, i1));
6323 emit_label (donelab);
6326 /* Return the string to output a conditional branch to LABEL, testing
6327 register REG. LABEL is the operand number of the label; REG is the
6328 operand number of the reg. OP is the conditional expression. The mode
6329 of REG says what kind of comparison we made.
6331 DEST is the destination insn (i.e. the label), INSN is the source.
6333 REVERSED is nonzero if we should reverse the sense of the comparison.
6335 ANNUL is nonzero if we should generate an annulling branch. */
6338 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
6339 int annul, rtx insn)
6341 static char string[64];
6342 enum rtx_code code = GET_CODE (op);
6343 enum machine_mode mode = GET_MODE (XEXP (op, 0));
6348 /* branch on register are limited to +-128KB. If it is too far away,
6361 brgez,a,pn %o1, .LC29
6367 ba,pt %xcc, .LC29 */
6369 far = get_attr_length (insn) >= 3;
6371 /* If not floating-point or if EQ or NE, we can just reverse the code. */
6373 code = reverse_condition (code);
6375 /* Only 64 bit versions of these instructions exist. */
6376 gcc_assert (mode == DImode);
6378 /* Start by writing the branch condition. */
6383 strcpy (string, "brnz");
6387 strcpy (string, "brz");
6391 strcpy (string, "brgez");
6395 strcpy (string, "brlz");
6399 strcpy (string, "brlez");
6403 strcpy (string, "brgz");
6410 p = strchr (string, '\0');
6412 /* Now add the annulling, reg, label, and nop. */
6419 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6422 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6427 *p = p < string + 8 ? '\t' : ' ';
6435 int veryfar = 1, delta;
6437 if (INSN_ADDRESSES_SET_P ())
6439 delta = (INSN_ADDRESSES (INSN_UID (dest))
6440 - INSN_ADDRESSES (INSN_UID (insn)));
6441 /* Leave some instructions for "slop". */
6442 if (delta >= -260000 && delta < 260000)
6446 strcpy (p, ".+12\n\t nop\n\t");
6447 /* Skip the next insn if requested or
6448 if we know that it will be a nop. */
6449 if (annul || ! final_sequence)
6459 strcpy (p, "ba,pt\t%%xcc, ");
6473 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
6474 Such instructions cannot be used in the delay slot of return insn on v9.
6475 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
6479 epilogue_renumber (register rtx *where, int test)
6481 register const char *fmt;
6483 register enum rtx_code code;
6488 code = GET_CODE (*where);
6493 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
6495 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
6496 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
6504 /* Do not replace the frame pointer with the stack pointer because
6505 it can cause the delayed instruction to load below the stack.
6506 This occurs when instructions like:
6508 (set (reg/i:SI 24 %i0)
6509 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
6510 (const_int -20 [0xffffffec])) 0))
6512 are in the return delayed slot. */
6514 if (GET_CODE (XEXP (*where, 0)) == REG
6515 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
6516 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
6517 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
6522 if (SPARC_STACK_BIAS
6523 && GET_CODE (XEXP (*where, 0)) == REG
6524 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
6532 fmt = GET_RTX_FORMAT (code);
6534 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6539 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
6540 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
6543 else if (fmt[i] == 'e'
6544 && epilogue_renumber (&(XEXP (*where, i)), test))
6550 /* Leaf functions and non-leaf functions have different needs. */
6553 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
6556 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
6558 static const int *const reg_alloc_orders[] = {
6559 reg_leaf_alloc_order,
6560 reg_nonleaf_alloc_order};
6563 order_regs_for_local_alloc (void)
6565 static int last_order_nonleaf = 1;
6567 if (df_regs_ever_live_p (15) != last_order_nonleaf)
6569 last_order_nonleaf = !last_order_nonleaf;
6570 memcpy ((char *) reg_alloc_order,
6571 (const char *) reg_alloc_orders[last_order_nonleaf],
6572 FIRST_PSEUDO_REGISTER * sizeof (int));
6576 /* Return 1 if REG and MEM are legitimate enough to allow the various
6577 mem<-->reg splits to be run. */
6580 sparc_splitdi_legitimate (rtx reg, rtx mem)
6582 /* Punt if we are here by mistake. */
6583 gcc_assert (reload_completed);
6585 /* We must have an offsettable memory reference. */
6586 if (! offsettable_memref_p (mem))
6589 /* If we have legitimate args for ldd/std, we do not want
6590 the split to happen. */
6591 if ((REGNO (reg) % 2) == 0
6592 && mem_min_alignment (mem, 8))
6599 /* Return 1 if x and y are some kind of REG and they refer to
6600 different hard registers. This test is guaranteed to be
6601 run after reload. */
6604 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
6606 if (GET_CODE (x) != REG)
6608 if (GET_CODE (y) != REG)
6610 if (REGNO (x) == REGNO (y))
6615 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
6616 This makes them candidates for using ldd and std insns.
6618 Note reg1 and reg2 *must* be hard registers. */
6621 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
6623 /* We might have been passed a SUBREG. */
6624 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
6627 if (REGNO (reg1) % 2 != 0)
6630 /* Integer ldd is deprecated in SPARC V9 */
6631 if (TARGET_V9 && REGNO (reg1) < 32)
6634 return (REGNO (reg1) == REGNO (reg2) - 1);
6637 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
6640 This can only happen when addr1 and addr2, the addresses in mem1
6641 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
6642 addr1 must also be aligned on a 64-bit boundary.
6644 Also iff dependent_reg_rtx is not null it should not be used to
6645 compute the address for mem1, i.e. we cannot optimize a sequence
6657 But, note that the transformation from:
6662 is perfectly fine. Thus, the peephole2 patterns always pass us
6663 the destination register of the first load, never the second one.
6665 For stores we don't have a similar problem, so dependent_reg_rtx is
6669 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
6673 HOST_WIDE_INT offset1;
6675 /* The mems cannot be volatile. */
6676 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
6679 /* MEM1 should be aligned on a 64-bit boundary. */
6680 if (MEM_ALIGN (mem1) < 64)
6683 addr1 = XEXP (mem1, 0);
6684 addr2 = XEXP (mem2, 0);
6686 /* Extract a register number and offset (if used) from the first addr. */
6687 if (GET_CODE (addr1) == PLUS)
6689 /* If not a REG, return zero. */
6690 if (GET_CODE (XEXP (addr1, 0)) != REG)
6694 reg1 = REGNO (XEXP (addr1, 0));
6695 /* The offset must be constant! */
6696 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
6698 offset1 = INTVAL (XEXP (addr1, 1));
6701 else if (GET_CODE (addr1) != REG)
6705 reg1 = REGNO (addr1);
6706 /* This was a simple (mem (reg)) expression. Offset is 0. */
6710 /* Make sure the second address is a (mem (plus (reg) (const_int). */
6711 if (GET_CODE (addr2) != PLUS)
6714 if (GET_CODE (XEXP (addr2, 0)) != REG
6715 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
6718 if (reg1 != REGNO (XEXP (addr2, 0)))
6721 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
6724 /* The first offset must be evenly divisible by 8 to ensure the
6725 address is 64 bit aligned. */
6726 if (offset1 % 8 != 0)
6729 /* The offset for the second addr must be 4 more than the first addr. */
6730 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
6733 /* All the tests passed. addr1 and addr2 are valid for ldd and std
6738 /* Return 1 if reg is a pseudo, or is the first register in
6739 a hard register pair. This makes it suitable for use in
6740 ldd and std insns. */
6743 register_ok_for_ldd (rtx reg)
6745 /* We might have been passed a SUBREG. */
6749 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
6750 return (REGNO (reg) % 2 == 0);
6755 /* Return 1 if OP is a memory whose address is known to be
6756 aligned to 8-byte boundary, or a pseudo during reload.
6757 This makes it suitable for use in ldd and std insns. */
6760 memory_ok_for_ldd (rtx op)
6764 /* In 64-bit mode, we assume that the address is word-aligned. */
6765 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
6768 if ((reload_in_progress || reload_completed)
6769 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
6772 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
6774 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
6783 /* Print operand X (an rtx) in assembler syntax to file FILE.
6784 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
6785 For `%' followed by punctuation, CODE is the punctuation and X is null. */
6788 print_operand (FILE *file, rtx x, int code)
6793 /* Output an insn in a delay slot. */
6795 sparc_indent_opcode = 1;
6797 fputs ("\n\t nop", file);
6800 /* Output an annul flag if there's nothing for the delay slot and we
6801 are optimizing. This is always used with '(' below.
6802 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
6803 this is a dbx bug. So, we only do this when optimizing.
6804 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
6805 Always emit a nop in case the next instruction is a branch. */
6806 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
6810 /* Output a 'nop' if there's nothing for the delay slot and we are
6811 not optimizing. This is always used with '*' above. */
6812 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
6813 fputs ("\n\t nop", file);
6814 else if (final_sequence)
6815 sparc_indent_opcode = 1;
6818 /* Output the right displacement from the saved PC on function return.
6819 The caller may have placed an "unimp" insn immediately after the call
6820 so we have to account for it. This insn is used in the 32-bit ABI
6821 when calling a function that returns a non zero-sized structure. The
6822 64-bit ABI doesn't have it. Be careful to have this test be the same
6823 as that used on the call. The exception here is that when
6824 sparc_std_struct_return is enabled, the psABI is followed exactly
6825 and the adjustment is made by the code in sparc_struct_value_rtx.
6826 The call emitted is the same when sparc_std_struct_return is
6829 && cfun->returns_struct
6830 && ! sparc_std_struct_return
6831 && (TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
6833 && ! integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
6839 /* Output the Embedded Medium/Anywhere code model base register. */
6840 fputs (EMBMEDANY_BASE_REG, file);
6843 /* Print some local dynamic TLS name. */
6844 assemble_name (file, get_some_local_dynamic_name ());
6848 /* Adjust the operand to take into account a RESTORE operation. */
6849 if (GET_CODE (x) == CONST_INT)
6851 else if (GET_CODE (x) != REG)
6852 output_operand_lossage ("invalid %%Y operand");
6853 else if (REGNO (x) < 8)
6854 fputs (reg_names[REGNO (x)], file);
6855 else if (REGNO (x) >= 24 && REGNO (x) < 32)
6856 fputs (reg_names[REGNO (x)-16], file);
6858 output_operand_lossage ("invalid %%Y operand");
6861 /* Print out the low order register name of a register pair. */
6862 if (WORDS_BIG_ENDIAN)
6863 fputs (reg_names[REGNO (x)+1], file);
6865 fputs (reg_names[REGNO (x)], file);
6868 /* Print out the high order register name of a register pair. */
6869 if (WORDS_BIG_ENDIAN)
6870 fputs (reg_names[REGNO (x)], file);
6872 fputs (reg_names[REGNO (x)+1], file);
6875 /* Print out the second register name of a register pair or quad.
6876 I.e., R (%o0) => %o1. */
6877 fputs (reg_names[REGNO (x)+1], file);
6880 /* Print out the third register name of a register quad.
6881 I.e., S (%o0) => %o2. */
6882 fputs (reg_names[REGNO (x)+2], file);
6885 /* Print out the fourth register name of a register quad.
6886 I.e., T (%o0) => %o3. */
6887 fputs (reg_names[REGNO (x)+3], file);
6890 /* Print a condition code register. */
6891 if (REGNO (x) == SPARC_ICC_REG)
6893 /* We don't handle CC[X]_NOOVmode because they're not supposed
6895 if (GET_MODE (x) == CCmode)
6896 fputs ("%icc", file);
6897 else if (GET_MODE (x) == CCXmode)
6898 fputs ("%xcc", file);
6903 /* %fccN register */
6904 fputs (reg_names[REGNO (x)], file);
6907 /* Print the operand's address only. */
6908 output_address (XEXP (x, 0));
6911 /* In this case we need a register. Use %g0 if the
6912 operand is const0_rtx. */
6914 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
6916 fputs ("%g0", file);
6923 switch (GET_CODE (x))
6925 case IOR: fputs ("or", file); break;
6926 case AND: fputs ("and", file); break;
6927 case XOR: fputs ("xor", file); break;
6928 default: output_operand_lossage ("invalid %%A operand");
6933 switch (GET_CODE (x))
6935 case IOR: fputs ("orn", file); break;
6936 case AND: fputs ("andn", file); break;
6937 case XOR: fputs ("xnor", file); break;
6938 default: output_operand_lossage ("invalid %%B operand");
6942 /* These are used by the conditional move instructions. */
6946 enum rtx_code rc = GET_CODE (x);
6950 enum machine_mode mode = GET_MODE (XEXP (x, 0));
6951 if (mode == CCFPmode || mode == CCFPEmode)
6952 rc = reverse_condition_maybe_unordered (GET_CODE (x));
6954 rc = reverse_condition (GET_CODE (x));
6958 case NE: fputs ("ne", file); break;
6959 case EQ: fputs ("e", file); break;
6960 case GE: fputs ("ge", file); break;
6961 case GT: fputs ("g", file); break;
6962 case LE: fputs ("le", file); break;
6963 case LT: fputs ("l", file); break;
6964 case GEU: fputs ("geu", file); break;
6965 case GTU: fputs ("gu", file); break;
6966 case LEU: fputs ("leu", file); break;
6967 case LTU: fputs ("lu", file); break;
6968 case LTGT: fputs ("lg", file); break;
6969 case UNORDERED: fputs ("u", file); break;
6970 case ORDERED: fputs ("o", file); break;
6971 case UNLT: fputs ("ul", file); break;
6972 case UNLE: fputs ("ule", file); break;
6973 case UNGT: fputs ("ug", file); break;
6974 case UNGE: fputs ("uge", file); break;
6975 case UNEQ: fputs ("ue", file); break;
6976 default: output_operand_lossage (code == 'c'
6977 ? "invalid %%c operand"
6978 : "invalid %%C operand");
6983 /* These are used by the movr instruction pattern. */
6987 enum rtx_code rc = (code == 'd'
6988 ? reverse_condition (GET_CODE (x))
6992 case NE: fputs ("ne", file); break;
6993 case EQ: fputs ("e", file); break;
6994 case GE: fputs ("gez", file); break;
6995 case LT: fputs ("lz", file); break;
6996 case LE: fputs ("lez", file); break;
6997 case GT: fputs ("gz", file); break;
6998 default: output_operand_lossage (code == 'd'
6999 ? "invalid %%d operand"
7000 : "invalid %%D operand");
7007 /* Print a sign-extended character. */
7008 int i = trunc_int_for_mode (INTVAL (x), QImode);
7009 fprintf (file, "%d", i);
7014 /* Operand must be a MEM; write its address. */
7015 if (GET_CODE (x) != MEM)
7016 output_operand_lossage ("invalid %%f operand");
7017 output_address (XEXP (x, 0));
7022 /* Print a sign-extended 32-bit value. */
7024 if (GET_CODE(x) == CONST_INT)
7026 else if (GET_CODE(x) == CONST_DOUBLE)
7027 i = CONST_DOUBLE_LOW (x);
7030 output_operand_lossage ("invalid %%s operand");
7033 i = trunc_int_for_mode (i, SImode);
7034 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
7039 /* Do nothing special. */
7043 /* Undocumented flag. */
7044 output_operand_lossage ("invalid operand output code");
7047 if (GET_CODE (x) == REG)
7048 fputs (reg_names[REGNO (x)], file);
7049 else if (GET_CODE (x) == MEM)
7052 /* Poor Sun assembler doesn't understand absolute addressing. */
7053 if (CONSTANT_P (XEXP (x, 0)))
7054 fputs ("%g0+", file);
7055 output_address (XEXP (x, 0));
7058 else if (GET_CODE (x) == HIGH)
7060 fputs ("%hi(", file);
7061 output_addr_const (file, XEXP (x, 0));
7064 else if (GET_CODE (x) == LO_SUM)
7066 print_operand (file, XEXP (x, 0), 0);
7067 if (TARGET_CM_MEDMID)
7068 fputs ("+%l44(", file);
7070 fputs ("+%lo(", file);
7071 output_addr_const (file, XEXP (x, 1));
7074 else if (GET_CODE (x) == CONST_DOUBLE
7075 && (GET_MODE (x) == VOIDmode
7076 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
7078 if (CONST_DOUBLE_HIGH (x) == 0)
7079 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
7080 else if (CONST_DOUBLE_HIGH (x) == -1
7081 && CONST_DOUBLE_LOW (x) < 0)
7082 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
7084 output_operand_lossage ("long long constant not a valid immediate operand");
7086 else if (GET_CODE (x) == CONST_DOUBLE)
7087 output_operand_lossage ("floating point constant not a valid immediate operand");
7088 else { output_addr_const (file, x); }
7091 /* Target hook for assembling integer objects. The sparc version has
7092 special handling for aligned DI-mode objects. */
7095 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
7097 /* ??? We only output .xword's for symbols and only then in environments
7098 where the assembler can handle them. */
7099 if (aligned_p && size == 8
7100 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
7104 assemble_integer_with_op ("\t.xword\t", x);
7109 assemble_aligned_integer (4, const0_rtx);
7110 assemble_aligned_integer (4, x);
7114 return default_assemble_integer (x, size, aligned_p);
7117 /* Return the value of a code used in the .proc pseudo-op that says
7118 what kind of result this function returns. For non-C types, we pick
7119 the closest C type. */
7121 #ifndef SHORT_TYPE_SIZE
7122 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
7125 #ifndef INT_TYPE_SIZE
7126 #define INT_TYPE_SIZE BITS_PER_WORD
7129 #ifndef LONG_TYPE_SIZE
7130 #define LONG_TYPE_SIZE BITS_PER_WORD
7133 #ifndef LONG_LONG_TYPE_SIZE
7134 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
7137 #ifndef FLOAT_TYPE_SIZE
7138 #define FLOAT_TYPE_SIZE BITS_PER_WORD
7141 #ifndef DOUBLE_TYPE_SIZE
7142 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7145 #ifndef LONG_DOUBLE_TYPE_SIZE
7146 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7150 sparc_type_code (register tree type)
7152 register unsigned long qualifiers = 0;
7153 register unsigned shift;
7155 /* Only the first 30 bits of the qualifier are valid. We must refrain from
7156 setting more, since some assemblers will give an error for this. Also,
7157 we must be careful to avoid shifts of 32 bits or more to avoid getting
7158 unpredictable results. */
7160 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
7162 switch (TREE_CODE (type))
7168 qualifiers |= (3 << shift);
7173 qualifiers |= (2 << shift);
7177 case REFERENCE_TYPE:
7179 qualifiers |= (1 << shift);
7183 return (qualifiers | 8);
7186 case QUAL_UNION_TYPE:
7187 return (qualifiers | 9);
7190 return (qualifiers | 10);
7193 return (qualifiers | 16);
7196 /* If this is a range type, consider it to be the underlying
7198 if (TREE_TYPE (type) != 0)
7201 /* Carefully distinguish all the standard types of C,
7202 without messing up if the language is not C. We do this by
7203 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
7204 look at both the names and the above fields, but that's redundant.
7205 Any type whose size is between two C types will be considered
7206 to be the wider of the two types. Also, we do not have a
7207 special code to use for "long long", so anything wider than
7208 long is treated the same. Note that we can't distinguish
7209 between "int" and "long" in this code if they are the same
7210 size, but that's fine, since neither can the assembler. */
7212 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
7213 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
7215 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
7216 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
7218 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
7219 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
7222 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
7225 /* If this is a range type, consider it to be the underlying
7227 if (TREE_TYPE (type) != 0)
7230 /* Carefully distinguish all the standard types of C,
7231 without messing up if the language is not C. */
7233 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
7234 return (qualifiers | 6);
7237 return (qualifiers | 7);
7239 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
7240 /* ??? We need to distinguish between double and float complex types,
7241 but I don't know how yet because I can't reach this code from
7242 existing front-ends. */
7243 return (qualifiers | 7); /* Who knows? */
7246 case BOOLEAN_TYPE: /* Boolean truth value type. */
7247 case LANG_TYPE: /* ? */
7251 gcc_unreachable (); /* Not a type! */
7258 /* Nested function support. */
7260 /* Emit RTL insns to initialize the variable parts of a trampoline.
7261 FNADDR is an RTX for the address of the function's pure code.
7262 CXT is an RTX for the static chain value for the function.
7264 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
7265 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
7266 (to store insns). This is a bit excessive. Perhaps a different
7267 mechanism would be better here.
7269 Emit enough FLUSH insns to synchronize the data and instruction caches. */
7272 sparc_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7274 /* SPARC 32-bit trampoline:
7277 sethi %hi(static), %g2
7279 or %g2, %lo(static), %g2
7281 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
7282 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
7286 (gen_rtx_MEM (SImode, plus_constant (tramp, 0)),
7287 expand_binop (SImode, ior_optab,
7288 expand_shift (RSHIFT_EXPR, SImode, fnaddr,
7289 size_int (10), 0, 1),
7290 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
7291 NULL_RTX, 1, OPTAB_DIRECT));
7294 (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7295 expand_binop (SImode, ior_optab,
7296 expand_shift (RSHIFT_EXPR, SImode, cxt,
7297 size_int (10), 0, 1),
7298 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
7299 NULL_RTX, 1, OPTAB_DIRECT));
7302 (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7303 expand_binop (SImode, ior_optab,
7304 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
7305 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
7306 NULL_RTX, 1, OPTAB_DIRECT));
7309 (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7310 expand_binop (SImode, ior_optab,
7311 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
7312 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
7313 NULL_RTX, 1, OPTAB_DIRECT));
7315 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
7316 aligned on a 16 byte boundary so one flush clears it all. */
7317 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, tramp))));
7318 if (sparc_cpu != PROCESSOR_ULTRASPARC
7319 && sparc_cpu != PROCESSOR_ULTRASPARC3
7320 && sparc_cpu != PROCESSOR_NIAGARA
7321 && sparc_cpu != PROCESSOR_NIAGARA2)
7322 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode,
7323 plus_constant (tramp, 8)))));
7325 /* Call __enable_execute_stack after writing onto the stack to make sure
7326 the stack address is accessible. */
7327 #ifdef ENABLE_EXECUTE_STACK
7328 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7329 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7334 /* The 64-bit version is simpler because it makes more sense to load the
7335 values as "immediate" data out of the trampoline. It's also easier since
7336 we can read the PC without clobbering a register. */
7339 sparc64_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7341 /* SPARC 64-bit trampoline:
7350 emit_move_insn (gen_rtx_MEM (SImode, tramp),
7351 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
7352 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7353 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
7354 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7355 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
7356 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7357 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
7358 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), cxt);
7359 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 24)), fnaddr);
7360 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, tramp))));
7362 if (sparc_cpu != PROCESSOR_ULTRASPARC
7363 && sparc_cpu != PROCESSOR_ULTRASPARC3
7364 && sparc_cpu != PROCESSOR_NIAGARA
7365 && sparc_cpu != PROCESSOR_NIAGARA2)
7366 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, plus_constant (tramp, 8)))));
7368 /* Call __enable_execute_stack after writing onto the stack to make sure
7369 the stack address is accessible. */
7370 #ifdef ENABLE_EXECUTE_STACK
7371 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7372 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7376 /* Adjust the cost of a scheduling dependency. Return the new cost of
7377 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
7380 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7382 enum attr_type insn_type;
7384 if (! recog_memoized (insn))
7387 insn_type = get_attr_type (insn);
7389 if (REG_NOTE_KIND (link) == 0)
7391 /* Data dependency; DEP_INSN writes a register that INSN reads some
7394 /* if a load, then the dependence must be on the memory address;
7395 add an extra "cycle". Note that the cost could be two cycles
7396 if the reg was written late in an instruction group; we ca not tell
7398 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
7401 /* Get the delay only if the address of the store is the dependence. */
7402 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
7404 rtx pat = PATTERN(insn);
7405 rtx dep_pat = PATTERN (dep_insn);
7407 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7408 return cost; /* This should not happen! */
7410 /* The dependency between the two instructions was on the data that
7411 is being stored. Assume that this implies that the address of the
7412 store is not dependent. */
7413 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7416 return cost + 3; /* An approximation. */
7419 /* A shift instruction cannot receive its data from an instruction
7420 in the same cycle; add a one cycle penalty. */
7421 if (insn_type == TYPE_SHIFT)
7422 return cost + 3; /* Split before cascade into shift. */
7426 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
7427 INSN writes some cycles later. */
7429 /* These are only significant for the fpu unit; writing a fp reg before
7430 the fpu has finished with it stalls the processor. */
7432 /* Reusing an integer register causes no problems. */
7433 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7441 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7443 enum attr_type insn_type, dep_type;
7444 rtx pat = PATTERN(insn);
7445 rtx dep_pat = PATTERN (dep_insn);
7447 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
7450 insn_type = get_attr_type (insn);
7451 dep_type = get_attr_type (dep_insn);
7453 switch (REG_NOTE_KIND (link))
7456 /* Data dependency; DEP_INSN writes a register that INSN reads some
7463 /* Get the delay iff the address of the store is the dependence. */
7464 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7467 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7474 /* If a load, then the dependence must be on the memory address. If
7475 the addresses aren't equal, then it might be a false dependency */
7476 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
7478 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
7479 || GET_CODE (SET_DEST (dep_pat)) != MEM
7480 || GET_CODE (SET_SRC (pat)) != MEM
7481 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
7482 XEXP (SET_SRC (pat), 0)))
7490 /* Compare to branch latency is 0. There is no benefit from
7491 separating compare and branch. */
7492 if (dep_type == TYPE_COMPARE)
7494 /* Floating point compare to branch latency is less than
7495 compare to conditional move. */
7496 if (dep_type == TYPE_FPCMP)
7505 /* Anti-dependencies only penalize the fpu unit. */
7506 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7518 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
7522 case PROCESSOR_SUPERSPARC:
7523 cost = supersparc_adjust_cost (insn, link, dep, cost);
7525 case PROCESSOR_HYPERSPARC:
7526 case PROCESSOR_SPARCLITE86X:
7527 cost = hypersparc_adjust_cost (insn, link, dep, cost);
7536 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
7537 int sched_verbose ATTRIBUTE_UNUSED,
7538 int max_ready ATTRIBUTE_UNUSED)
7543 sparc_use_sched_lookahead (void)
7545 if (sparc_cpu == PROCESSOR_NIAGARA
7546 || sparc_cpu == PROCESSOR_NIAGARA2)
7548 if (sparc_cpu == PROCESSOR_ULTRASPARC
7549 || sparc_cpu == PROCESSOR_ULTRASPARC3)
7551 if ((1 << sparc_cpu) &
7552 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
7553 (1 << PROCESSOR_SPARCLITE86X)))
7559 sparc_issue_rate (void)
7563 case PROCESSOR_NIAGARA:
7564 case PROCESSOR_NIAGARA2:
7568 /* Assume V9 processors are capable of at least dual-issue. */
7570 case PROCESSOR_SUPERSPARC:
7572 case PROCESSOR_HYPERSPARC:
7573 case PROCESSOR_SPARCLITE86X:
7575 case PROCESSOR_ULTRASPARC:
7576 case PROCESSOR_ULTRASPARC3:
7582 set_extends (rtx insn)
7584 register rtx pat = PATTERN (insn);
7586 switch (GET_CODE (SET_SRC (pat)))
7588 /* Load and some shift instructions zero extend. */
7591 /* sethi clears the high bits */
7593 /* LO_SUM is used with sethi. sethi cleared the high
7594 bits and the values used with lo_sum are positive */
7596 /* Store flag stores 0 or 1 */
7606 rtx op0 = XEXP (SET_SRC (pat), 0);
7607 rtx op1 = XEXP (SET_SRC (pat), 1);
7608 if (GET_CODE (op1) == CONST_INT)
7609 return INTVAL (op1) >= 0;
7610 if (GET_CODE (op0) != REG)
7612 if (sparc_check_64 (op0, insn) == 1)
7614 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7619 rtx op0 = XEXP (SET_SRC (pat), 0);
7620 rtx op1 = XEXP (SET_SRC (pat), 1);
7621 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
7623 if (GET_CODE (op1) == CONST_INT)
7624 return INTVAL (op1) >= 0;
7625 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7628 return GET_MODE (SET_SRC (pat)) == SImode;
7629 /* Positive integers leave the high bits zero. */
7631 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
7633 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
7636 return - (GET_MODE (SET_SRC (pat)) == SImode);
7638 return sparc_check_64 (SET_SRC (pat), insn);
7644 /* We _ought_ to have only one kind per function, but... */
7645 static GTY(()) rtx sparc_addr_diff_list;
7646 static GTY(()) rtx sparc_addr_list;
7649 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
7651 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
7653 sparc_addr_diff_list
7654 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
7656 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
7660 sparc_output_addr_vec (rtx vec)
7662 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7663 int idx, vlen = XVECLEN (body, 0);
7665 #ifdef ASM_OUTPUT_ADDR_VEC_START
7666 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7669 #ifdef ASM_OUTPUT_CASE_LABEL
7670 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7673 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7676 for (idx = 0; idx < vlen; idx++)
7678 ASM_OUTPUT_ADDR_VEC_ELT
7679 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
7682 #ifdef ASM_OUTPUT_ADDR_VEC_END
7683 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7688 sparc_output_addr_diff_vec (rtx vec)
7690 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7691 rtx base = XEXP (XEXP (body, 0), 0);
7692 int idx, vlen = XVECLEN (body, 1);
7694 #ifdef ASM_OUTPUT_ADDR_VEC_START
7695 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7698 #ifdef ASM_OUTPUT_CASE_LABEL
7699 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7702 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7705 for (idx = 0; idx < vlen; idx++)
7707 ASM_OUTPUT_ADDR_DIFF_ELT
7710 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
7711 CODE_LABEL_NUMBER (base));
7714 #ifdef ASM_OUTPUT_ADDR_VEC_END
7715 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7720 sparc_output_deferred_case_vectors (void)
7725 if (sparc_addr_list == NULL_RTX
7726 && sparc_addr_diff_list == NULL_RTX)
7729 /* Align to cache line in the function's code section. */
7730 switch_to_section (current_function_section ());
7732 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
7734 ASM_OUTPUT_ALIGN (asm_out_file, align);
7736 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
7737 sparc_output_addr_vec (XEXP (t, 0));
7738 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
7739 sparc_output_addr_diff_vec (XEXP (t, 0));
7741 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
7744 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
7745 unknown. Return 1 if the high bits are zero, -1 if the register is
7748 sparc_check_64 (rtx x, rtx insn)
7750 /* If a register is set only once it is safe to ignore insns this
7751 code does not know how to handle. The loop will either recognize
7752 the single set and return the correct value or fail to recognize
7757 gcc_assert (GET_CODE (x) == REG);
7759 if (GET_MODE (x) == DImode)
7760 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
7762 if (flag_expensive_optimizations
7763 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
7769 insn = get_last_insn_anywhere ();
7774 while ((insn = PREV_INSN (insn)))
7776 switch (GET_CODE (insn))
7789 rtx pat = PATTERN (insn);
7790 if (GET_CODE (pat) != SET)
7792 if (rtx_equal_p (x, SET_DEST (pat)))
7793 return set_extends (insn);
7794 if (y && rtx_equal_p (y, SET_DEST (pat)))
7795 return set_extends (insn);
7796 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
7804 /* Returns assembly code to perform a DImode shift using
7805 a 64-bit global or out register on SPARC-V8+. */
7807 output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
7809 static char asm_code[60];
7811 /* The scratch register is only required when the destination
7812 register is not a 64-bit global or out register. */
7813 if (which_alternative != 2)
7814 operands[3] = operands[0];
7816 /* We can only shift by constants <= 63. */
7817 if (GET_CODE (operands[2]) == CONST_INT)
7818 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
7820 if (GET_CODE (operands[1]) == CONST_INT)
7822 output_asm_insn ("mov\t%1, %3", operands);
7826 output_asm_insn ("sllx\t%H1, 32, %3", operands);
7827 if (sparc_check_64 (operands[1], insn) <= 0)
7828 output_asm_insn ("srl\t%L1, 0, %L1", operands);
7829 output_asm_insn ("or\t%L1, %3, %3", operands);
7832 strcpy(asm_code, opcode);
7834 if (which_alternative != 2)
7835 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
7837 return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
7840 /* Output rtl to increment the profiler label LABELNO
7841 for profiling a function entry. */
7844 sparc_profile_hook (int labelno)
7849 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
7850 if (NO_PROFILE_COUNTERS)
7852 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
7856 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
7857 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
7858 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
7862 #ifdef OBJECT_FORMAT_ELF
7864 sparc_elf_asm_named_section (const char *name, unsigned int flags,
7867 if (flags & SECTION_MERGE)
7869 /* entsize cannot be expressed in this section attributes
7871 default_elf_asm_named_section (name, flags, decl);
7875 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
7877 if (!(flags & SECTION_DEBUG))
7878 fputs (",#alloc", asm_out_file);
7879 if (flags & SECTION_WRITE)
7880 fputs (",#write", asm_out_file);
7881 if (flags & SECTION_TLS)
7882 fputs (",#tls", asm_out_file);
7883 if (flags & SECTION_CODE)
7884 fputs (",#execinstr", asm_out_file);
7886 /* ??? Handle SECTION_BSS. */
7888 fputc ('\n', asm_out_file);
7890 #endif /* OBJECT_FORMAT_ELF */
7892 /* We do not allow indirect calls to be optimized into sibling calls.
7894 We cannot use sibling calls when delayed branches are disabled
7895 because they will likely require the call delay slot to be filled.
7897 Also, on SPARC 32-bit we cannot emit a sibling call when the
7898 current function returns a structure. This is because the "unimp
7899 after call" convention would cause the callee to return to the
7900 wrong place. The generic code already disallows cases where the
7901 function being called returns a structure.
7903 It may seem strange how this last case could occur. Usually there
7904 is code after the call which jumps to epilogue code which dumps the
7905 return value into the struct return area. That ought to invalidate
7906 the sibling call right? Well, in the C++ case we can end up passing
7907 the pointer to the struct return area to a constructor (which returns
7908 void) and then nothing else happens. Such a sibling call would look
7909 valid without the added check here.
7911 VxWorks PIC PLT entries require the global pointer to be initialized
7912 on entry. We therefore can't emit sibling calls to them. */
7914 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
7917 && flag_delayed_branch
7918 && (TARGET_ARCH64 || ! cfun->returns_struct)
7919 && !(TARGET_VXWORKS_RTP
7921 && !targetm.binds_local_p (decl)));
7924 /* libfunc renaming. */
7925 #include "config/gofast.h"
7928 sparc_init_libfuncs (void)
7932 /* Use the subroutines that Sun's library provides for integer
7933 multiply and divide. The `*' prevents an underscore from
7934 being prepended by the compiler. .umul is a little faster
7936 set_optab_libfunc (smul_optab, SImode, "*.umul");
7937 set_optab_libfunc (sdiv_optab, SImode, "*.div");
7938 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
7939 set_optab_libfunc (smod_optab, SImode, "*.rem");
7940 set_optab_libfunc (umod_optab, SImode, "*.urem");
7942 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
7943 set_optab_libfunc (add_optab, TFmode, "_Q_add");
7944 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
7945 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
7946 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
7947 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
7949 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
7950 is because with soft-float, the SFmode and DFmode sqrt
7951 instructions will be absent, and the compiler will notice and
7952 try to use the TFmode sqrt instruction for calls to the
7953 builtin function sqrt, but this fails. */
7955 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
7957 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
7958 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
7959 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
7960 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
7961 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
7962 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
7964 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
7965 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
7966 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
7967 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
7969 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
7970 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
7971 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
7972 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
7974 if (DITF_CONVERSION_LIBFUNCS)
7976 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
7977 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
7978 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
7979 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
7982 if (SUN_CONVERSION_LIBFUNCS)
7984 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
7985 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
7986 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
7987 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
7992 /* In the SPARC 64bit ABI, SImode multiply and divide functions
7993 do not exist in the library. Make sure the compiler does not
7994 emit calls to them by accident. (It should always use the
7995 hardware instructions.) */
7996 set_optab_libfunc (smul_optab, SImode, 0);
7997 set_optab_libfunc (sdiv_optab, SImode, 0);
7998 set_optab_libfunc (udiv_optab, SImode, 0);
7999 set_optab_libfunc (smod_optab, SImode, 0);
8000 set_optab_libfunc (umod_optab, SImode, 0);
8002 if (SUN_INTEGER_MULTIPLY_64)
8004 set_optab_libfunc (smul_optab, DImode, "__mul64");
8005 set_optab_libfunc (sdiv_optab, DImode, "__div64");
8006 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
8007 set_optab_libfunc (smod_optab, DImode, "__rem64");
8008 set_optab_libfunc (umod_optab, DImode, "__urem64");
8011 if (SUN_CONVERSION_LIBFUNCS)
8013 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
8014 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
8015 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
8016 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
8020 gofast_maybe_init_libfuncs ();
8023 #define def_builtin(NAME, CODE, TYPE) \
8024 add_builtin_function((NAME), (TYPE), (CODE), BUILT_IN_MD, NULL, \
8027 /* Implement the TARGET_INIT_BUILTINS target hook.
8028 Create builtin functions for special SPARC instructions. */
8031 sparc_init_builtins (void)
8034 sparc_vis_init_builtins ();
8037 /* Create builtin functions for VIS 1.0 instructions. */
8040 sparc_vis_init_builtins (void)
8042 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
8043 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
8044 tree v4hi = build_vector_type (intHI_type_node, 4);
8045 tree v2hi = build_vector_type (intHI_type_node, 2);
8046 tree v2si = build_vector_type (intSI_type_node, 2);
8048 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
8049 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
8050 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
8051 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
8052 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
8053 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
8054 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
8055 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
8056 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
8057 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
8058 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
8059 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
8060 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
8062 intDI_type_node, 0);
8063 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
8065 intDI_type_node, 0);
8066 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
8068 intSI_type_node, 0);
8069 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
8071 intDI_type_node, 0);
8073 /* Packing and expanding vectors. */
8074 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis, v4qi_ftype_v4hi);
8075 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
8076 v8qi_ftype_v2si_v8qi);
8077 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
8079 def_builtin ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis, v4hi_ftype_v4qi);
8080 def_builtin ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
8081 v8qi_ftype_v4qi_v4qi);
8083 /* Multiplications. */
8084 def_builtin ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
8085 v4hi_ftype_v4qi_v4hi);
8086 def_builtin ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
8087 v4hi_ftype_v4qi_v2hi);
8088 def_builtin ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
8089 v4hi_ftype_v4qi_v2hi);
8090 def_builtin ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
8091 v4hi_ftype_v8qi_v4hi);
8092 def_builtin ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
8093 v4hi_ftype_v8qi_v4hi);
8094 def_builtin ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
8095 v2si_ftype_v4qi_v2hi);
8096 def_builtin ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
8097 v2si_ftype_v4qi_v2hi);
8099 /* Data aligning. */
8100 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
8101 v4hi_ftype_v4hi_v4hi);
8102 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
8103 v8qi_ftype_v8qi_v8qi);
8104 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
8105 v2si_ftype_v2si_v2si);
8106 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatadi_vis,
8109 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
8112 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
8115 /* Pixel distance. */
8116 def_builtin ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
8117 di_ftype_v8qi_v8qi_di);
8120 /* Handle TARGET_EXPAND_BUILTIN target hook.
8121 Expand builtin functions for sparc intrinsics. */
8124 sparc_expand_builtin (tree exp, rtx target,
8125 rtx subtarget ATTRIBUTE_UNUSED,
8126 enum machine_mode tmode ATTRIBUTE_UNUSED,
8127 int ignore ATTRIBUTE_UNUSED)
8130 call_expr_arg_iterator iter;
8131 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
8132 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
8134 enum machine_mode mode[4];
8137 mode[0] = insn_data[icode].operand[0].mode;
8139 || GET_MODE (target) != mode[0]
8140 || ! (*insn_data[icode].operand[0].predicate) (target, mode[0]))
8141 op[0] = gen_reg_rtx (mode[0]);
8145 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
8148 mode[arg_count] = insn_data[icode].operand[arg_count].mode;
8149 op[arg_count] = expand_normal (arg);
8151 if (! (*insn_data[icode].operand[arg_count].predicate) (op[arg_count],
8153 op[arg_count] = copy_to_mode_reg (mode[arg_count], op[arg_count]);
8159 pat = GEN_FCN (icode) (op[0], op[1]);
8162 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
8165 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
8180 sparc_vis_mul8x16 (int e8, int e16)
8182 return (e8 * e16 + 128) / 256;
8185 /* Multiply the vector elements in ELTS0 to the elements in ELTS1 as specified
8186 by FNCODE. All of the elements in ELTS0 and ELTS1 lists must be integer
8187 constants. A tree list with the results of the multiplications is returned,
8188 and each element in the list is of INNER_TYPE. */
8191 sparc_handle_vis_mul8x16 (int fncode, tree inner_type, tree elts0, tree elts1)
8193 tree n_elts = NULL_TREE;
8198 case CODE_FOR_fmul8x16_vis:
8199 for (; elts0 && elts1;
8200 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8203 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8204 TREE_INT_CST_LOW (TREE_VALUE (elts1)));
8205 n_elts = tree_cons (NULL_TREE,
8206 build_int_cst (inner_type, val),
8211 case CODE_FOR_fmul8x16au_vis:
8212 scale = TREE_INT_CST_LOW (TREE_VALUE (elts1));
8214 for (; elts0; elts0 = TREE_CHAIN (elts0))
8217 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8219 n_elts = tree_cons (NULL_TREE,
8220 build_int_cst (inner_type, val),
8225 case CODE_FOR_fmul8x16al_vis:
8226 scale = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (elts1)));
8228 for (; elts0; elts0 = TREE_CHAIN (elts0))
8231 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8233 n_elts = tree_cons (NULL_TREE,
8234 build_int_cst (inner_type, val),
8243 return nreverse (n_elts);
8246 /* Handle TARGET_FOLD_BUILTIN target hook.
8247 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
8248 result of the function call is ignored. NULL_TREE is returned if the
8249 function could not be folded. */
8252 sparc_fold_builtin (tree fndecl, tree arglist, bool ignore)
8254 tree arg0, arg1, arg2;
8255 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
8256 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
8259 && icode != CODE_FOR_alignaddrsi_vis
8260 && icode != CODE_FOR_alignaddrdi_vis)
8261 return fold_convert (rtype, integer_zero_node);
8265 case CODE_FOR_fexpand_vis:
8266 arg0 = TREE_VALUE (arglist);
8269 if (TREE_CODE (arg0) == VECTOR_CST)
8271 tree inner_type = TREE_TYPE (rtype);
8272 tree elts = TREE_VECTOR_CST_ELTS (arg0);
8273 tree n_elts = NULL_TREE;
8275 for (; elts; elts = TREE_CHAIN (elts))
8277 unsigned int val = TREE_INT_CST_LOW (TREE_VALUE (elts)) << 4;
8278 n_elts = tree_cons (NULL_TREE,
8279 build_int_cst (inner_type, val),
8282 return build_vector (rtype, nreverse (n_elts));
8286 case CODE_FOR_fmul8x16_vis:
8287 case CODE_FOR_fmul8x16au_vis:
8288 case CODE_FOR_fmul8x16al_vis:
8289 arg0 = TREE_VALUE (arglist);
8290 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8294 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
8296 tree inner_type = TREE_TYPE (rtype);
8297 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8298 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8299 tree n_elts = sparc_handle_vis_mul8x16 (icode, inner_type, elts0,
8302 return build_vector (rtype, n_elts);
8306 case CODE_FOR_fpmerge_vis:
8307 arg0 = TREE_VALUE (arglist);
8308 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8312 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
8314 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8315 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8316 tree n_elts = NULL_TREE;
8318 for (; elts0 && elts1;
8319 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8321 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts0), n_elts);
8322 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts1), n_elts);
8325 return build_vector (rtype, nreverse (n_elts));
8329 case CODE_FOR_pdist_vis:
8330 arg0 = TREE_VALUE (arglist);
8331 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
8332 arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
8337 if (TREE_CODE (arg0) == VECTOR_CST
8338 && TREE_CODE (arg1) == VECTOR_CST
8339 && TREE_CODE (arg2) == INTEGER_CST)
8342 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
8343 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
8344 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
8345 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
8347 for (; elts0 && elts1;
8348 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
8350 unsigned HOST_WIDE_INT
8351 low0 = TREE_INT_CST_LOW (TREE_VALUE (elts0)),
8352 low1 = TREE_INT_CST_LOW (TREE_VALUE (elts1));
8353 HOST_WIDE_INT high0 = TREE_INT_CST_HIGH (TREE_VALUE (elts0));
8354 HOST_WIDE_INT high1 = TREE_INT_CST_HIGH (TREE_VALUE (elts1));
8356 unsigned HOST_WIDE_INT l;
8359 overflow |= neg_double (low1, high1, &l, &h);
8360 overflow |= add_double (low0, high0, l, h, &l, &h);
8362 overflow |= neg_double (l, h, &l, &h);
8364 overflow |= add_double (low, high, l, h, &low, &high);
8367 gcc_assert (overflow == 0);
8369 return build_int_cst_wide (rtype, low, high);
8379 /* ??? This duplicates information provided to the compiler by the
8380 ??? scheduler description. Some day, teach genautomata to output
8381 ??? the latencies and then CSE will just use that. */
8384 sparc_rtx_costs (rtx x, int code, int outer_code, int *total,
8385 bool speed ATTRIBUTE_UNUSED)
8387 enum machine_mode mode = GET_MODE (x);
8388 bool float_mode_p = FLOAT_MODE_P (mode);
8393 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
8411 if (GET_MODE (x) == VOIDmode
8412 && ((CONST_DOUBLE_HIGH (x) == 0
8413 && CONST_DOUBLE_LOW (x) < 0x1000)
8414 || (CONST_DOUBLE_HIGH (x) == -1
8415 && CONST_DOUBLE_LOW (x) < 0
8416 && CONST_DOUBLE_LOW (x) >= -0x1000)))
8423 /* If outer-code was a sign or zero extension, a cost
8424 of COSTS_N_INSNS (1) was already added in. This is
8425 why we are subtracting it back out. */
8426 if (outer_code == ZERO_EXTEND)
8428 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
8430 else if (outer_code == SIGN_EXTEND)
8432 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
8434 else if (float_mode_p)
8436 *total = sparc_costs->float_load;
8440 *total = sparc_costs->int_load;
8448 *total = sparc_costs->float_plusminus;
8450 *total = COSTS_N_INSNS (1);
8455 *total = sparc_costs->float_mul;
8456 else if (! TARGET_HARD_MUL)
8457 *total = COSTS_N_INSNS (25);
8463 if (sparc_costs->int_mul_bit_factor)
8467 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8469 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
8470 for (nbits = 0; value != 0; value &= value - 1)
8473 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
8474 && GET_MODE (XEXP (x, 1)) == VOIDmode)
8476 rtx x1 = XEXP (x, 1);
8477 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
8478 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
8480 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
8482 for (; value2 != 0; value2 &= value2 - 1)
8490 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
8491 bit_cost = COSTS_N_INSNS (bit_cost);
8495 *total = sparc_costs->int_mulX + bit_cost;
8497 *total = sparc_costs->int_mul + bit_cost;
8504 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
8514 *total = sparc_costs->float_div_df;
8516 *total = sparc_costs->float_div_sf;
8521 *total = sparc_costs->int_divX;
8523 *total = sparc_costs->int_div;
8530 *total = COSTS_N_INSNS (1);
8537 case UNSIGNED_FLOAT:
8541 case FLOAT_TRUNCATE:
8542 *total = sparc_costs->float_move;
8547 *total = sparc_costs->float_sqrt_df;
8549 *total = sparc_costs->float_sqrt_sf;
8554 *total = sparc_costs->float_cmp;
8556 *total = COSTS_N_INSNS (1);
8561 *total = sparc_costs->float_cmove;
8563 *total = sparc_costs->int_cmove;
8567 /* Handle the NAND vector patterns. */
8568 if (sparc_vector_mode_supported_p (GET_MODE (x))
8569 && GET_CODE (XEXP (x, 0)) == NOT
8570 && GET_CODE (XEXP (x, 1)) == NOT)
8572 *total = COSTS_N_INSNS (1);
8583 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
8584 This is achieved by means of a manual dynamic stack space allocation in
8585 the current frame. We make the assumption that SEQ doesn't contain any
8586 function calls, with the possible exception of calls to the PIC helper. */
8589 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
8591 /* We must preserve the lowest 16 words for the register save area. */
8592 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
8593 /* We really need only 2 words of fresh stack space. */
8594 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
8597 = gen_rtx_MEM (word_mode, plus_constant (stack_pointer_rtx,
8598 SPARC_STACK_BIAS + offset));
8600 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
8601 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
8603 emit_insn (gen_rtx_SET (VOIDmode,
8604 adjust_address (slot, word_mode, UNITS_PER_WORD),
8608 emit_insn (gen_rtx_SET (VOIDmode,
8610 adjust_address (slot, word_mode, UNITS_PER_WORD)));
8611 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
8612 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
8615 /* Output the assembler code for a thunk function. THUNK_DECL is the
8616 declaration for the thunk function itself, FUNCTION is the decl for
8617 the target function. DELTA is an immediate constant offset to be
8618 added to THIS. If VCALL_OFFSET is nonzero, the word at address
8619 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
8622 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
8623 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
8626 rtx this_rtx, insn, funexp;
8627 unsigned int int_arg_first;
8629 reload_completed = 1;
8630 epilogue_completed = 1;
8632 emit_note (NOTE_INSN_PROLOGUE_END);
8634 if (flag_delayed_branch)
8636 /* We will emit a regular sibcall below, so we need to instruct
8637 output_sibcall that we are in a leaf function. */
8638 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
8640 /* This will cause final.c to invoke leaf_renumber_regs so we
8641 must behave as if we were in a not-yet-leafified function. */
8642 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
8646 /* We will emit the sibcall manually below, so we will need to
8647 manually spill non-leaf registers. */
8648 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
8650 /* We really are in a leaf function. */
8651 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
8654 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
8655 returns a structure, the structure return pointer is there instead. */
8656 if (TARGET_ARCH64 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
8657 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
8659 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
8661 /* Add DELTA. When possible use a plain add, otherwise load it into
8662 a register first. */
8665 rtx delta_rtx = GEN_INT (delta);
8667 if (! SPARC_SIMM13_P (delta))
8669 rtx scratch = gen_rtx_REG (Pmode, 1);
8670 emit_move_insn (scratch, delta_rtx);
8671 delta_rtx = scratch;
8674 /* THIS_RTX += DELTA. */
8675 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
8678 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
8681 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
8682 rtx scratch = gen_rtx_REG (Pmode, 1);
8684 gcc_assert (vcall_offset < 0);
8686 /* SCRATCH = *THIS_RTX. */
8687 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
8689 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
8690 may not have any available scratch register at this point. */
8691 if (SPARC_SIMM13_P (vcall_offset))
8693 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
8694 else if (! fixed_regs[5]
8695 /* The below sequence is made up of at least 2 insns,
8696 while the default method may need only one. */
8697 && vcall_offset < -8192)
8699 rtx scratch2 = gen_rtx_REG (Pmode, 5);
8700 emit_move_insn (scratch2, vcall_offset_rtx);
8701 vcall_offset_rtx = scratch2;
8705 rtx increment = GEN_INT (-4096);
8707 /* VCALL_OFFSET is a negative number whose typical range can be
8708 estimated as -32768..0 in 32-bit mode. In almost all cases
8709 it is therefore cheaper to emit multiple add insns than
8710 spilling and loading the constant into a register (at least
8712 while (! SPARC_SIMM13_P (vcall_offset))
8714 emit_insn (gen_add2_insn (scratch, increment));
8715 vcall_offset += 4096;
8717 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
8720 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
8721 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
8722 gen_rtx_PLUS (Pmode,
8724 vcall_offset_rtx)));
8726 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
8727 emit_insn (gen_add2_insn (this_rtx, scratch));
8730 /* Generate a tail call to the target function. */
8731 if (! TREE_USED (function))
8733 assemble_external (function);
8734 TREE_USED (function) = 1;
8736 funexp = XEXP (DECL_RTL (function), 0);
8738 if (flag_delayed_branch)
8740 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8741 insn = emit_call_insn (gen_sibcall (funexp));
8742 SIBLING_CALL_P (insn) = 1;
8746 /* The hoops we have to jump through in order to generate a sibcall
8747 without using delay slots... */
8748 rtx spill_reg, spill_reg2, seq, scratch = gen_rtx_REG (Pmode, 1);
8752 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
8753 spill_reg2 = gen_rtx_REG (word_mode, PIC_OFFSET_TABLE_REGNUM);
8755 /* Delay emitting the PIC helper function because it needs to
8756 change the section and we are emitting assembly code. */
8757 load_pic_register (true); /* clobbers %o7 */
8758 scratch = legitimize_pic_address (funexp, Pmode, scratch);
8761 emit_and_preserve (seq, spill_reg, spill_reg2);
8763 else if (TARGET_ARCH32)
8765 emit_insn (gen_rtx_SET (VOIDmode,
8767 gen_rtx_HIGH (SImode, funexp)));
8768 emit_insn (gen_rtx_SET (VOIDmode,
8770 gen_rtx_LO_SUM (SImode, scratch, funexp)));
8772 else /* TARGET_ARCH64 */
8774 switch (sparc_cmodel)
8778 /* The destination can serve as a temporary. */
8779 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
8784 /* The destination cannot serve as a temporary. */
8785 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
8787 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
8790 emit_and_preserve (seq, spill_reg, 0);
8798 emit_jump_insn (gen_indirect_jump (scratch));
8803 /* Run just enough of rest_of_compilation to get the insns emitted.
8804 There's not really enough bulk here to make other passes such as
8805 instruction scheduling worth while. Note that use_thunk calls
8806 assemble_start_function and assemble_end_function. */
8807 insn = get_insns ();
8808 insn_locators_alloc ();
8809 shorten_branches (insn);
8810 final_start_function (insn, file, 1);
8811 final (insn, file, 1);
8812 final_end_function ();
8813 free_after_compilation (cfun);
8815 reload_completed = 0;
8816 epilogue_completed = 0;
8819 /* Return true if sparc_output_mi_thunk would be able to output the
8820 assembler code for the thunk function specified by the arguments
8821 it is passed, and false otherwise. */
8823 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
8824 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
8825 HOST_WIDE_INT vcall_offset,
8826 const_tree function ATTRIBUTE_UNUSED)
8828 /* Bound the loop used in the default method above. */
8829 return (vcall_offset >= -32768 || ! fixed_regs[5]);
8832 /* How to allocate a 'struct machine_function'. */
8834 static struct machine_function *
8835 sparc_init_machine_status (void)
8837 return GGC_CNEW (struct machine_function);
8840 /* Locate some local-dynamic symbol still in use by this function
8841 so that we can print its name in local-dynamic base patterns. */
8844 get_some_local_dynamic_name (void)
8848 if (cfun->machine->some_ld_name)
8849 return cfun->machine->some_ld_name;
8851 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
8853 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
8854 return cfun->machine->some_ld_name;
8860 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
8865 && GET_CODE (x) == SYMBOL_REF
8866 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
8868 cfun->machine->some_ld_name = XSTR (x, 0);
8875 /* Handle the TARGET_DWARF_HANDLE_FRAME_UNSPEC hook.
8876 This is called from dwarf2out.c to emit call frame instructions
8877 for frame-related insns containing UNSPECs and UNSPEC_VOLATILEs. */
8879 sparc_dwarf_handle_frame_unspec (const char *label,
8880 rtx pattern ATTRIBUTE_UNUSED,
8881 int index ATTRIBUTE_UNUSED)
8883 gcc_assert (index == UNSPECV_SAVEW);
8884 dwarf2out_window_save (label);
8887 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
8888 We need to emit DTP-relative relocations. */
8891 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
8896 fputs ("\t.word\t%r_tls_dtpoff32(", file);
8899 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
8904 output_addr_const (file, x);
8908 /* Do whatever processing is required at the end of a file. */
8911 sparc_file_end (void)
8913 /* If we haven't emitted the special PIC helper function, do so now. */
8914 if (pic_helper_symbol_name[0] && !pic_helper_emitted_p)
8917 if (NEED_INDICATE_EXEC_STACK)
8918 file_end_indicate_exec_stack ();
8921 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
8922 /* Implement TARGET_MANGLE_TYPE. */
8925 sparc_mangle_type (const_tree type)
8928 && TYPE_MAIN_VARIANT (type) == long_double_type_node
8929 && TARGET_LONG_DOUBLE_128)
8932 /* For all other types, use normal C++ mangling. */
8937 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
8938 compare and swap on the word containing the byte or half-word. */
8941 sparc_expand_compare_and_swap_12 (rtx result, rtx mem, rtx oldval, rtx newval)
8943 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
8944 rtx addr = gen_reg_rtx (Pmode);
8945 rtx off = gen_reg_rtx (SImode);
8946 rtx oldv = gen_reg_rtx (SImode);
8947 rtx newv = gen_reg_rtx (SImode);
8948 rtx oldvalue = gen_reg_rtx (SImode);
8949 rtx newvalue = gen_reg_rtx (SImode);
8950 rtx res = gen_reg_rtx (SImode);
8951 rtx resv = gen_reg_rtx (SImode);
8952 rtx memsi, val, mask, end_label, loop_label, cc;
8954 emit_insn (gen_rtx_SET (VOIDmode, addr,
8955 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
8957 if (Pmode != SImode)
8958 addr1 = gen_lowpart (SImode, addr1);
8959 emit_insn (gen_rtx_SET (VOIDmode, off,
8960 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
8962 memsi = gen_rtx_MEM (SImode, addr);
8963 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
8964 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
8966 val = force_reg (SImode, memsi);
8968 emit_insn (gen_rtx_SET (VOIDmode, off,
8969 gen_rtx_XOR (SImode, off,
8970 GEN_INT (GET_MODE (mem) == QImode
8973 emit_insn (gen_rtx_SET (VOIDmode, off,
8974 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
8976 if (GET_MODE (mem) == QImode)
8977 mask = force_reg (SImode, GEN_INT (0xff));
8979 mask = force_reg (SImode, GEN_INT (0xffff));
8981 emit_insn (gen_rtx_SET (VOIDmode, mask,
8982 gen_rtx_ASHIFT (SImode, mask, off)));
8984 emit_insn (gen_rtx_SET (VOIDmode, val,
8985 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
8988 oldval = gen_lowpart (SImode, oldval);
8989 emit_insn (gen_rtx_SET (VOIDmode, oldv,
8990 gen_rtx_ASHIFT (SImode, oldval, off)));
8992 newval = gen_lowpart_common (SImode, newval);
8993 emit_insn (gen_rtx_SET (VOIDmode, newv,
8994 gen_rtx_ASHIFT (SImode, newval, off)));
8996 emit_insn (gen_rtx_SET (VOIDmode, oldv,
8997 gen_rtx_AND (SImode, oldv, mask)));
8999 emit_insn (gen_rtx_SET (VOIDmode, newv,
9000 gen_rtx_AND (SImode, newv, mask)));
9002 end_label = gen_label_rtx ();
9003 loop_label = gen_label_rtx ();
9004 emit_label (loop_label);
9006 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
9007 gen_rtx_IOR (SImode, oldv, val)));
9009 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
9010 gen_rtx_IOR (SImode, newv, val)));
9012 emit_insn (gen_sync_compare_and_swapsi (res, memsi, oldvalue, newvalue));
9014 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
9016 emit_insn (gen_rtx_SET (VOIDmode, resv,
9017 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
9020 sparc_compare_op0 = resv;
9021 sparc_compare_op1 = val;
9022 cc = gen_compare_reg (NE);
9024 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
9026 sparc_compare_op0 = cc;
9027 sparc_compare_op1 = const0_rtx;
9028 emit_jump_insn (gen_bne (loop_label));
9030 emit_label (end_label);
9032 emit_insn (gen_rtx_SET (VOIDmode, res,
9033 gen_rtx_AND (SImode, res, mask)));
9035 emit_insn (gen_rtx_SET (VOIDmode, res,
9036 gen_rtx_LSHIFTRT (SImode, res, off)));
9038 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
9041 #include "gt-sparc.h"