]>
Commit | Line | Data |
---|---|---|
87ad11b0 | 1 | /* Subroutines for insn-output.c for HPPA. |
a9ac13e4 | 2 | Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, |
cd0dfcc5 | 3 | 2002, 2003 Free Software Foundation, Inc. |
87ad11b0 | 4 | Contributed by Tim Moore (moore@cs.utah.edu), based on sparc.c |
5 | ||
5c1d8983 | 6 | This file is part of GCC. |
87ad11b0 | 7 | |
5c1d8983 | 8 | GCC is free software; you can redistribute it and/or modify |
87ad11b0 | 9 | it under the terms of the GNU General Public License as published by |
10 | the Free Software Foundation; either version 2, or (at your option) | |
11 | any later version. | |
12 | ||
5c1d8983 | 13 | GCC is distributed in the hope that it will be useful, |
87ad11b0 | 14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | GNU General Public License for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
5c1d8983 | 19 | along with GCC; see the file COPYING. If not, write to |
01e379d0 | 20 | the Free Software Foundation, 59 Temple Place - Suite 330, |
21 | Boston, MA 02111-1307, USA. */ | |
87ad11b0 | 22 | |
87ad11b0 | 23 | #include "config.h" |
b1ca791d | 24 | #include "system.h" |
805e22b2 | 25 | #include "coretypes.h" |
26 | #include "tm.h" | |
87ad11b0 | 27 | #include "rtl.h" |
28 | #include "regs.h" | |
29 | #include "hard-reg-set.h" | |
30 | #include "real.h" | |
31 | #include "insn-config.h" | |
32 | #include "conditions.h" | |
87ad11b0 | 33 | #include "insn-attr.h" |
34 | #include "flags.h" | |
35 | #include "tree.h" | |
9d2d8bd6 | 36 | #include "output.h" |
a584fe8a | 37 | #include "except.h" |
32509e56 | 38 | #include "expr.h" |
d8fc4d0b | 39 | #include "optabs.h" |
d8fc4d0b | 40 | #include "reload.h" |
d7e2f694 | 41 | #include "integrate.h" |
0a893c29 | 42 | #include "function.h" |
d6f01525 | 43 | #include "obstack.h" |
b1ca791d | 44 | #include "toplev.h" |
5cb4669a | 45 | #include "ggc.h" |
611a88e1 | 46 | #include "recog.h" |
a584fe8a | 47 | #include "predict.h" |
611a88e1 | 48 | #include "tm_p.h" |
a767736d | 49 | #include "target.h" |
50 | #include "target-def.h" | |
87ad11b0 | 51 | |
5c1d8983 | 52 | static int hppa_use_dfa_pipeline_interface (void); |
bea4bad2 | 53 | |
54 | #undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE | |
55 | #define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE hppa_use_dfa_pipeline_interface | |
56 | ||
57 | static int | |
5c1d8983 | 58 | hppa_use_dfa_pipeline_interface (void) |
bea4bad2 | 59 | { |
60 | return 1; | |
61 | } | |
62 | ||
cde3e16c | 63 | /* Return nonzero if there is a bypass for the output of |
64 | OUT_INSN and the fp store IN_INSN. */ | |
65 | int | |
5c1d8983 | 66 | hppa_fpstore_bypass_p (rtx out_insn, rtx in_insn) |
cde3e16c | 67 | { |
68 | enum machine_mode store_mode; | |
69 | enum machine_mode other_mode; | |
70 | rtx set; | |
71 | ||
72 | if (recog_memoized (in_insn) < 0 | |
73 | || get_attr_type (in_insn) != TYPE_FPSTORE | |
74 | || recog_memoized (out_insn) < 0) | |
75 | return 0; | |
76 | ||
77 | store_mode = GET_MODE (SET_SRC (PATTERN (in_insn))); | |
78 | ||
79 | set = single_set (out_insn); | |
80 | if (!set) | |
81 | return 0; | |
82 | ||
83 | other_mode = GET_MODE (SET_SRC (set)); | |
84 | ||
85 | return (GET_MODE_SIZE (store_mode) == GET_MODE_SIZE (other_mode)); | |
86 | } | |
87 | ||
88 | ||
cc858176 | 89 | #ifndef DO_FRAME_NOTES |
90 | #ifdef INCOMING_RETURN_ADDR_RTX | |
91 | #define DO_FRAME_NOTES 1 | |
92 | #else | |
93 | #define DO_FRAME_NOTES 0 | |
94 | #endif | |
95 | #endif | |
96 | ||
5c1d8983 | 97 | static int hppa_address_cost (rtx); |
98 | static bool hppa_rtx_costs (rtx, int, int, int *); | |
99 | static inline rtx force_mode (enum machine_mode, rtx); | |
100 | static void pa_reorg (void); | |
101 | static void pa_combine_instructions (void); | |
102 | static int pa_can_combine_p (rtx, rtx, rtx, int, rtx, rtx, rtx); | |
103 | static int forward_branch_p (rtx); | |
104 | static int shadd_constant_p (int); | |
105 | static void compute_zdepwi_operands (unsigned HOST_WIDE_INT, unsigned *); | |
a7e1bb24 | 106 | static int compute_movstr_length (rtx); |
107 | static int compute_clrstr_length (rtx); | |
5c1d8983 | 108 | static bool pa_assemble_integer (rtx, unsigned int, int); |
109 | static void remove_useless_addtr_insns (int); | |
110 | static void store_reg (int, int, int); | |
111 | static void store_reg_modify (int, int, int); | |
112 | static void load_reg (int, int, int); | |
113 | static void set_reg_plus_d (int, int, int, int); | |
114 | static void pa_output_function_prologue (FILE *, HOST_WIDE_INT); | |
115 | static void update_total_code_bytes (int); | |
116 | static void pa_output_function_epilogue (FILE *, HOST_WIDE_INT); | |
117 | static int pa_adjust_cost (rtx, rtx, rtx, int); | |
118 | static int pa_adjust_priority (rtx, int); | |
119 | static int pa_issue_rate (void); | |
120 | static void pa_select_section (tree, int, unsigned HOST_WIDE_INT) | |
52470889 | 121 | ATTRIBUTE_UNUSED; |
5c1d8983 | 122 | static void pa_encode_section_info (tree, rtx, int); |
123 | static const char *pa_strip_name_encoding (const char *); | |
124 | static bool pa_function_ok_for_sibcall (tree, tree); | |
125 | static void pa_globalize_label (FILE *, const char *) | |
63b8cd48 | 126 | ATTRIBUTE_UNUSED; |
5c1d8983 | 127 | static void pa_asm_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, |
128 | HOST_WIDE_INT, tree); | |
de419443 | 129 | #if !defined(USE_COLLECT2) |
5c1d8983 | 130 | static void pa_asm_out_constructor (rtx, int); |
131 | static void pa_asm_out_destructor (rtx, int); | |
de419443 | 132 | #endif |
5c1d8983 | 133 | static void pa_init_builtins (void); |
134 | static void copy_fp_args (rtx) ATTRIBUTE_UNUSED; | |
135 | static int length_fp_args (rtx) ATTRIBUTE_UNUSED; | |
136 | static struct deferred_plabel *get_plabel (const char *) | |
ece88821 | 137 | ATTRIBUTE_UNUSED; |
5c1d8983 | 138 | static inline void pa_file_start_level (void) ATTRIBUTE_UNUSED; |
139 | static inline void pa_file_start_space (int) ATTRIBUTE_UNUSED; | |
140 | static inline void pa_file_start_file (int) ATTRIBUTE_UNUSED; | |
141 | static inline void pa_file_start_mcount (const char*) ATTRIBUTE_UNUSED; | |
142 | static void pa_elf_file_start (void) ATTRIBUTE_UNUSED; | |
143 | static void pa_som_file_start (void) ATTRIBUTE_UNUSED; | |
144 | static void pa_linux_file_start (void) ATTRIBUTE_UNUSED; | |
145 | static void pa_hpux64_gas_file_start (void) ATTRIBUTE_UNUSED; | |
146 | static void pa_hpux64_hpas_file_start (void) ATTRIBUTE_UNUSED; | |
147 | static void output_deferred_plabels (void); | |
f2f543a3 | 148 | #ifdef HPUX_LONG_DOUBLE_LIBRARY |
149 | static void pa_hpux_init_libfuncs (void); | |
150 | #endif | |
9d3ddb8f | 151 | |
87ad11b0 | 152 | /* Save the operands last given to a compare for use when we |
153 | generate a scc or bcc insn. */ | |
87ad11b0 | 154 | rtx hppa_compare_op0, hppa_compare_op1; |
155 | enum cmp_type hppa_branch_type; | |
156 | ||
134b4858 | 157 | /* Which cpu we are scheduling for. */ |
158 | enum processor_type pa_cpu; | |
159 | ||
160 | /* String to hold which cpu we are scheduling for. */ | |
611a88e1 | 161 | const char *pa_cpu_string; |
134b4858 | 162 | |
4dc42288 | 163 | /* Which architecture we are generating code for. */ |
164 | enum architecture_type pa_arch; | |
165 | ||
166 | /* String to hold which architecture we are generating code for. */ | |
611a88e1 | 167 | const char *pa_arch_string; |
4dc42288 | 168 | |
a9960cdc | 169 | /* Counts for the number of callee-saved general and floating point |
170 | registers which were saved by the current function's prologue. */ | |
171 | static int gr_saved, fr_saved; | |
172 | ||
5c1d8983 | 173 | static rtx find_addr_reg (rtx); |
87ad11b0 | 174 | |
2247cc5f | 175 | /* Keep track of the number of bytes we have output in the CODE subspace |
06ddb6f8 | 176 | during this compilation so we'll know when to emit inline long-calls. */ |
ece88821 | 177 | unsigned long total_code_bytes; |
06ddb6f8 | 178 | |
2247cc5f | 179 | /* The last address of the previous function plus the number of bytes in |
180 | associated thunks that have been output. This is used to determine if | |
181 | a thunk can use an IA-relative branch to reach its target function. */ | |
182 | static int last_address; | |
183 | ||
e3f53689 | 184 | /* Variables to handle plabels that we discover are necessary at assembly |
01cc3b75 | 185 | output time. They are output after the current function. */ |
1f3233d1 | 186 | struct deferred_plabel GTY(()) |
e3f53689 | 187 | { |
188 | rtx internal_label; | |
8d0cec1a | 189 | const char *name; |
1f3233d1 | 190 | }; |
191 | static GTY((length ("n_deferred_plabels"))) struct deferred_plabel * | |
192 | deferred_plabels; | |
e11bd7e5 | 193 | static size_t n_deferred_plabels = 0; |
f6940372 | 194 | |
a767736d | 195 | \f |
196 | /* Initialize the GCC target structure. */ | |
58356836 | 197 | |
198 | #undef TARGET_ASM_ALIGNED_HI_OP | |
199 | #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t" | |
200 | #undef TARGET_ASM_ALIGNED_SI_OP | |
201 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
202 | #undef TARGET_ASM_ALIGNED_DI_OP | |
203 | #define TARGET_ASM_ALIGNED_DI_OP "\t.dword\t" | |
204 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
205 | #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP | |
206 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
207 | #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP | |
208 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
209 | #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP | |
210 | #undef TARGET_ASM_INTEGER | |
211 | #define TARGET_ASM_INTEGER pa_assemble_integer | |
212 | ||
17d9b0c3 | 213 | #undef TARGET_ASM_FUNCTION_PROLOGUE |
214 | #define TARGET_ASM_FUNCTION_PROLOGUE pa_output_function_prologue | |
215 | #undef TARGET_ASM_FUNCTION_EPILOGUE | |
216 | #define TARGET_ASM_FUNCTION_EPILOGUE pa_output_function_epilogue | |
e3f53689 | 217 | |
747af5e7 | 218 | #undef TARGET_SCHED_ADJUST_COST |
219 | #define TARGET_SCHED_ADJUST_COST pa_adjust_cost | |
220 | #undef TARGET_SCHED_ADJUST_PRIORITY | |
221 | #define TARGET_SCHED_ADJUST_PRIORITY pa_adjust_priority | |
222 | #undef TARGET_SCHED_ISSUE_RATE | |
223 | #define TARGET_SCHED_ISSUE_RATE pa_issue_rate | |
224 | ||
7811991d | 225 | #undef TARGET_ENCODE_SECTION_INFO |
226 | #define TARGET_ENCODE_SECTION_INFO pa_encode_section_info | |
7b4a38a6 | 227 | #undef TARGET_STRIP_NAME_ENCODING |
228 | #define TARGET_STRIP_NAME_ENCODING pa_strip_name_encoding | |
7811991d | 229 | |
805e22b2 | 230 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
231 | #define TARGET_FUNCTION_OK_FOR_SIBCALL pa_function_ok_for_sibcall | |
232 | ||
6988553d | 233 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
234 | #define TARGET_ASM_OUTPUT_MI_THUNK pa_asm_output_mi_thunk | |
eb344f43 | 235 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK |
236 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall | |
6988553d | 237 | |
f6940372 | 238 | #undef TARGET_ASM_FILE_END |
239 | #define TARGET_ASM_FILE_END output_deferred_plabels | |
240 | ||
de419443 | 241 | #if !defined(USE_COLLECT2) |
242 | #undef TARGET_ASM_CONSTRUCTOR | |
243 | #define TARGET_ASM_CONSTRUCTOR pa_asm_out_constructor | |
244 | #undef TARGET_ASM_DESTRUCTOR | |
245 | #define TARGET_ASM_DESTRUCTOR pa_asm_out_destructor | |
246 | #endif | |
247 | ||
ffa8918b | 248 | #undef TARGET_INIT_BUILTINS |
249 | #define TARGET_INIT_BUILTINS pa_init_builtins | |
250 | ||
fab7adbf | 251 | #undef TARGET_RTX_COSTS |
252 | #define TARGET_RTX_COSTS hppa_rtx_costs | |
ec0457a8 | 253 | #undef TARGET_ADDRESS_COST |
254 | #define TARGET_ADDRESS_COST hppa_address_cost | |
fab7adbf | 255 | |
2efea8c0 | 256 | #undef TARGET_MACHINE_DEPENDENT_REORG |
257 | #define TARGET_MACHINE_DEPENDENT_REORG pa_reorg | |
258 | ||
f2f543a3 | 259 | #ifdef HPUX_LONG_DOUBLE_LIBRARY |
260 | #undef TARGET_INIT_LIBFUNCS | |
261 | #define TARGET_INIT_LIBFUNCS pa_hpux_init_libfuncs | |
262 | #endif | |
263 | ||
57e4bbfb | 264 | struct gcc_target targetm = TARGET_INITIALIZER; |
a767736d | 265 | \f |
134b4858 | 266 | void |
5c1d8983 | 267 | override_options (void) |
134b4858 | 268 | { |
b3b62da8 | 269 | if (pa_cpu_string == NULL) |
270 | pa_cpu_string = TARGET_SCHED_DEFAULT; | |
271 | ||
272 | if (! strcmp (pa_cpu_string, "8000")) | |
273 | { | |
274 | pa_cpu_string = "8000"; | |
275 | pa_cpu = PROCESSOR_8000; | |
276 | } | |
277 | else if (! strcmp (pa_cpu_string, "7100")) | |
134b4858 | 278 | { |
279 | pa_cpu_string = "7100"; | |
280 | pa_cpu = PROCESSOR_7100; | |
281 | } | |
b3b62da8 | 282 | else if (! strcmp (pa_cpu_string, "700")) |
a87e1e15 | 283 | { |
284 | pa_cpu_string = "700"; | |
285 | pa_cpu = PROCESSOR_700; | |
286 | } | |
b3b62da8 | 287 | else if (! strcmp (pa_cpu_string, "7100LC")) |
134b4858 | 288 | { |
289 | pa_cpu_string = "7100LC"; | |
290 | pa_cpu = PROCESSOR_7100LC; | |
291 | } | |
b3b62da8 | 292 | else if (! strcmp (pa_cpu_string, "7200")) |
b4e0d953 | 293 | { |
294 | pa_cpu_string = "7200"; | |
295 | pa_cpu = PROCESSOR_7200; | |
296 | } | |
b3b62da8 | 297 | else if (! strcmp (pa_cpu_string, "7300")) |
bea4bad2 | 298 | { |
299 | pa_cpu_string = "7300"; | |
300 | pa_cpu = PROCESSOR_7300; | |
301 | } | |
134b4858 | 302 | else |
303 | { | |
bea4bad2 | 304 | warning ("unknown -mschedule= option (%s).\nValid options are 700, 7100, 7100LC, 7200, 7300, and 8000\n", pa_cpu_string); |
134b4858 | 305 | } |
c7a4e712 | 306 | |
4dc42288 | 307 | /* Set the instruction set architecture. */ |
308 | if (pa_arch_string && ! strcmp (pa_arch_string, "1.0")) | |
309 | { | |
310 | pa_arch_string = "1.0"; | |
311 | pa_arch = ARCHITECTURE_10; | |
312 | target_flags &= ~(MASK_PA_11 | MASK_PA_20); | |
313 | } | |
314 | else if (pa_arch_string && ! strcmp (pa_arch_string, "1.1")) | |
315 | { | |
316 | pa_arch_string = "1.1"; | |
317 | pa_arch = ARCHITECTURE_11; | |
318 | target_flags &= ~MASK_PA_20; | |
319 | target_flags |= MASK_PA_11; | |
320 | } | |
321 | else if (pa_arch_string && ! strcmp (pa_arch_string, "2.0")) | |
322 | { | |
323 | pa_arch_string = "2.0"; | |
324 | pa_arch = ARCHITECTURE_20; | |
325 | target_flags |= MASK_PA_11 | MASK_PA_20; | |
326 | } | |
327 | else if (pa_arch_string) | |
328 | { | |
68435912 | 329 | warning ("unknown -march= option (%s).\nValid options are 1.0, 1.1, and 2.0\n", pa_arch_string); |
4dc42288 | 330 | } |
331 | ||
7c5101fc | 332 | /* Unconditional branches in the delay slot are not compatible with dwarf2 |
333 | call frame information. There is no benefit in using this optimization | |
334 | on PA8000 and later processors. */ | |
335 | if (pa_cpu >= PROCESSOR_8000 | |
336 | || (! USING_SJLJ_EXCEPTIONS && flag_exceptions) | |
337 | || flag_unwind_tables) | |
338 | target_flags &= ~MASK_JUMP_IN_DELAY; | |
339 | ||
c7a4e712 | 340 | if (flag_pic && TARGET_PORTABLE_RUNTIME) |
341 | { | |
342 | warning ("PIC code generation is not supported in the portable runtime model\n"); | |
343 | } | |
344 | ||
b29897dd | 345 | if (flag_pic && TARGET_FAST_INDIRECT_CALLS) |
c7a4e712 | 346 | { |
ad87de1e | 347 | warning ("PIC code generation is not compatible with fast indirect calls\n"); |
c7a4e712 | 348 | } |
751e64a1 | 349 | |
5bd7b548 | 350 | if (! TARGET_GAS && write_symbols != NO_DEBUG) |
351 | { | |
352 | warning ("-g is only supported when using GAS on this processor,"); | |
68435912 | 353 | warning ("-g option disabled"); |
5bd7b548 | 354 | write_symbols = NO_DEBUG; |
355 | } | |
5cb4669a | 356 | |
fc44315f | 357 | /* We only support the "big PIC" model now. And we always generate PIC |
358 | code when in 64bit mode. */ | |
359 | if (flag_pic == 1 || TARGET_64BIT) | |
5e3c5739 | 360 | flag_pic = 2; |
361 | ||
58356836 | 362 | /* We can't guarantee that .dword is available for 32-bit targets. */ |
363 | if (UNITS_PER_WORD == 4) | |
364 | targetm.asm_out.aligned_op.di = NULL; | |
365 | ||
366 | /* The unaligned ops are only available when using GAS. */ | |
367 | if (!TARGET_GAS) | |
368 | { | |
369 | targetm.asm_out.unaligned_op.hi = NULL; | |
370 | targetm.asm_out.unaligned_op.si = NULL; | |
371 | targetm.asm_out.unaligned_op.di = NULL; | |
372 | } | |
134b4858 | 373 | } |
374 | ||
066397a3 | 375 | static void |
5c1d8983 | 376 | pa_init_builtins (void) |
ffa8918b | 377 | { |
378 | #ifdef DONT_HAVE_FPUTC_UNLOCKED | |
379 | built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE; | |
5ddc8fb8 | 380 | implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE; |
ffa8918b | 381 | #endif |
382 | } | |
383 | ||
e911aedf | 384 | /* Return nonzero only if OP is a register of mode MODE, |
891b55b4 | 385 | or CONST0_RTX. */ |
87ad11b0 | 386 | int |
5c1d8983 | 387 | reg_or_0_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 388 | { |
891b55b4 | 389 | return (op == CONST0_RTX (mode) || register_operand (op, mode)); |
87ad11b0 | 390 | } |
391 | ||
e911aedf | 392 | /* Return nonzero if OP is suitable for use in a call to a named |
575e0eb4 | 393 | function. |
394 | ||
9dc4cbf1 | 395 | For 2.5 try to eliminate either call_operand_address or |
575e0eb4 | 396 | function_label_operand, they perform very similar functions. */ |
87ad11b0 | 397 | int |
5c1d8983 | 398 | call_operand_address (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 399 | { |
2db59d96 | 400 | return (GET_MODE (op) == word_mode |
401 | && CONSTANT_P (op) && ! TARGET_PORTABLE_RUNTIME); | |
87ad11b0 | 402 | } |
403 | ||
6d36483b | 404 | /* Return 1 if X contains a symbolic expression. We know these |
405 | expressions will have one of a few well defined forms, so | |
347b5848 | 406 | we need only check those forms. */ |
407 | int | |
5c1d8983 | 408 | symbolic_expression_p (rtx x) |
347b5848 | 409 | { |
410 | ||
6dc3b0d9 | 411 | /* Strip off any HIGH. */ |
347b5848 | 412 | if (GET_CODE (x) == HIGH) |
413 | x = XEXP (x, 0); | |
414 | ||
415 | return (symbolic_operand (x, VOIDmode)); | |
416 | } | |
417 | ||
87ad11b0 | 418 | int |
5c1d8983 | 419 | symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 420 | { |
421 | switch (GET_CODE (op)) | |
422 | { | |
423 | case SYMBOL_REF: | |
424 | case LABEL_REF: | |
425 | return 1; | |
426 | case CONST: | |
427 | op = XEXP (op, 0); | |
428 | return ((GET_CODE (XEXP (op, 0)) == SYMBOL_REF | |
429 | || GET_CODE (XEXP (op, 0)) == LABEL_REF) | |
430 | && GET_CODE (XEXP (op, 1)) == CONST_INT); | |
431 | default: | |
432 | return 0; | |
433 | } | |
434 | } | |
435 | ||
436 | /* Return truth value of statement that OP is a symbolic memory | |
437 | operand of mode MODE. */ | |
438 | ||
439 | int | |
5c1d8983 | 440 | symbolic_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 441 | { |
442 | if (GET_CODE (op) == SUBREG) | |
443 | op = SUBREG_REG (op); | |
444 | if (GET_CODE (op) != MEM) | |
445 | return 0; | |
446 | op = XEXP (op, 0); | |
447 | return (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == CONST | |
448 | || GET_CODE (op) == HIGH || GET_CODE (op) == LABEL_REF); | |
449 | } | |
450 | ||
451 | /* Return 1 if the operand is either a register or a memory operand that is | |
452 | not symbolic. */ | |
453 | ||
454 | int | |
5c1d8983 | 455 | reg_or_nonsymb_mem_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 456 | { |
457 | if (register_operand (op, mode)) | |
458 | return 1; | |
459 | ||
460 | if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode)) | |
461 | return 1; | |
462 | ||
463 | return 0; | |
464 | } | |
465 | ||
6d36483b | 466 | /* Return 1 if the operand is either a register, zero, or a memory operand |
891b55b4 | 467 | that is not symbolic. */ |
468 | ||
469 | int | |
5c1d8983 | 470 | reg_or_0_or_nonsymb_mem_operand (rtx op, enum machine_mode mode) |
891b55b4 | 471 | { |
472 | if (register_operand (op, mode)) | |
473 | return 1; | |
474 | ||
475 | if (op == CONST0_RTX (mode)) | |
476 | return 1; | |
477 | ||
478 | if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode)) | |
479 | return 1; | |
480 | ||
481 | return 0; | |
482 | } | |
483 | ||
d3138877 | 484 | /* Return 1 if the operand is a register operand or a non-symbolic memory |
485 | operand after reload. This predicate is used for branch patterns that | |
486 | internally handle register reloading. We need to accept non-symbolic | |
487 | memory operands after reload to ensure that the pattern is still valid | |
488 | if reload didn't find a hard register for the operand. */ | |
489 | ||
490 | int | |
5c1d8983 | 491 | reg_before_reload_operand (rtx op, enum machine_mode mode) |
d3138877 | 492 | { |
6df54749 | 493 | /* Don't accept a SUBREG since it will need a reload. */ |
494 | if (GET_CODE (op) == SUBREG) | |
495 | return 0; | |
496 | ||
d3138877 | 497 | if (register_operand (op, mode)) |
498 | return 1; | |
499 | ||
500 | if (reload_completed | |
501 | && memory_operand (op, mode) | |
502 | && ! symbolic_memory_operand (op, mode)) | |
503 | return 1; | |
504 | ||
505 | return 0; | |
506 | } | |
507 | ||
7c4d3047 | 508 | /* Accept any constant that can be moved in one instruction into a |
d9d7c968 | 509 | general register. */ |
6d36483b | 510 | int |
5c1d8983 | 511 | cint_ok_for_move (HOST_WIDE_INT intval) |
d9d7c968 | 512 | { |
513 | /* OK if ldo, ldil, or zdepi, can be used. */ | |
d8d5c4ff | 514 | return (CONST_OK_FOR_LETTER_P (intval, 'J') |
515 | || CONST_OK_FOR_LETTER_P (intval, 'N') | |
516 | || CONST_OK_FOR_LETTER_P (intval, 'K')); | |
d9d7c968 | 517 | } |
518 | ||
6ecdbaa1 | 519 | /* Accept anything that can be moved in one instruction into a general |
520 | register. */ | |
87ad11b0 | 521 | int |
5c1d8983 | 522 | move_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 523 | { |
524 | if (register_operand (op, mode)) | |
525 | return 1; | |
526 | ||
cc9205b8 | 527 | if (GET_CODE (op) == CONSTANT_P_RTX) |
528 | return 1; | |
529 | ||
42faba01 | 530 | if (GET_CODE (op) == CONST_INT) |
d9d7c968 | 531 | return cint_ok_for_move (INTVAL (op)); |
87ad11b0 | 532 | |
87ad11b0 | 533 | if (GET_CODE (op) == SUBREG) |
534 | op = SUBREG_REG (op); | |
535 | if (GET_CODE (op) != MEM) | |
536 | return 0; | |
537 | ||
538 | op = XEXP (op, 0); | |
27ef382d | 539 | |
2d14b1f0 | 540 | /* We consider a LO_SUM DLT reference a move_operand now since it has |
541 | been merged into the normal movsi/movdi patterns. */ | |
5d7431a3 | 542 | if (GET_CODE (op) == LO_SUM |
543 | && GET_CODE (XEXP (op, 0)) == REG | |
544 | && REG_OK_FOR_BASE_P (XEXP (op, 0)) | |
2d14b1f0 | 545 | && GET_CODE (XEXP (op, 1)) == UNSPEC |
546 | && GET_MODE (op) == Pmode) | |
547 | return 1; | |
5d7431a3 | 548 | |
27ef382d | 549 | /* Since move_operand is only used for source operands, we can always |
550 | allow scaled indexing! */ | |
822c8a6c | 551 | if (! TARGET_DISABLE_INDEXING |
552 | && GET_CODE (op) == PLUS | |
27ef382d | 553 | && ((GET_CODE (XEXP (op, 0)) == MULT |
554 | && GET_CODE (XEXP (XEXP (op, 0), 0)) == REG | |
555 | && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT | |
4ad2b0ae | 556 | && INTVAL (XEXP (XEXP (op, 0), 1)) |
557 | == (HOST_WIDE_INT) GET_MODE_SIZE (mode) | |
27ef382d | 558 | && GET_CODE (XEXP (op, 1)) == REG) |
559 | || (GET_CODE (XEXP (op, 1)) == MULT | |
560 | &&GET_CODE (XEXP (XEXP (op, 1), 0)) == REG | |
561 | && GET_CODE (XEXP (XEXP (op, 1), 1)) == CONST_INT | |
4ad2b0ae | 562 | && INTVAL (XEXP (XEXP (op, 1), 1)) |
563 | == (HOST_WIDE_INT) GET_MODE_SIZE (mode) | |
27ef382d | 564 | && GET_CODE (XEXP (op, 0)) == REG))) |
565 | return 1; | |
566 | ||
87ad11b0 | 567 | return memory_address_p (mode, op); |
568 | } | |
569 | ||
6ecdbaa1 | 570 | /* Accept REG and any CONST_INT that can be moved in one instruction into a |
571 | general register. */ | |
572 | int | |
5c1d8983 | 573 | reg_or_cint_move_operand (rtx op, enum machine_mode mode) |
6ecdbaa1 | 574 | { |
575 | if (register_operand (op, mode)) | |
576 | return 1; | |
577 | ||
578 | if (GET_CODE (op) == CONST_INT) | |
686b848d | 579 | return cint_ok_for_move (INTVAL (op)); |
580 | ||
6ecdbaa1 | 581 | return 0; |
582 | } | |
583 | ||
87ad11b0 | 584 | int |
5c1d8983 | 585 | pic_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 586 | { |
b4a7bf10 | 587 | if (!flag_pic) |
588 | return 0; | |
589 | ||
590 | switch (GET_CODE (op)) | |
591 | { | |
592 | case LABEL_REF: | |
593 | return 1; | |
b4a7bf10 | 594 | case CONST: |
595 | op = XEXP (op, 0); | |
3c69dc97 | 596 | return (GET_CODE (XEXP (op, 0)) == LABEL_REF |
b4a7bf10 | 597 | && GET_CODE (XEXP (op, 1)) == CONST_INT); |
598 | default: | |
599 | return 0; | |
600 | } | |
87ad11b0 | 601 | } |
602 | ||
87ad11b0 | 603 | int |
5c1d8983 | 604 | fp_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 605 | { |
606 | return reg_renumber && FP_REG_P (op); | |
607 | } | |
d6f01525 | 608 | |
87ad11b0 | 609 | \f |
87ad11b0 | 610 | |
87ad11b0 | 611 | /* Return truth value of whether OP can be used as an operand in a |
612 | three operand arithmetic insn that accepts registers of mode MODE | |
613 | or 14-bit signed integers. */ | |
614 | int | |
5c1d8983 | 615 | arith_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 616 | { |
617 | return (register_operand (op, mode) | |
618 | || (GET_CODE (op) == CONST_INT && INT_14_BITS (op))); | |
619 | } | |
620 | ||
621 | /* Return truth value of whether OP can be used as an operand in a | |
622 | three operand arithmetic insn that accepts registers of mode MODE | |
623 | or 11-bit signed integers. */ | |
624 | int | |
5c1d8983 | 625 | arith11_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 626 | { |
627 | return (register_operand (op, mode) | |
628 | || (GET_CODE (op) == CONST_INT && INT_11_BITS (op))); | |
629 | } | |
630 | ||
2a91e628 | 631 | /* Return truth value of whether OP can be used as an operand in a |
632 | adddi3 insn. */ | |
633 | int | |
5c1d8983 | 634 | adddi3_operand (rtx op, enum machine_mode mode) |
2a91e628 | 635 | { |
636 | return (register_operand (op, mode) | |
637 | || (GET_CODE (op) == CONST_INT | |
638 | && (TARGET_64BIT ? INT_14_BITS (op) : INT_11_BITS (op)))); | |
639 | } | |
640 | ||
6d36483b | 641 | /* A constant integer suitable for use in a PRE_MODIFY memory |
757d4970 | 642 | reference. */ |
42faba01 | 643 | int |
5c1d8983 | 644 | pre_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
42faba01 | 645 | { |
646 | return (GET_CODE (op) == CONST_INT | |
647 | && INTVAL (op) >= -0x2000 && INTVAL (op) < 0x10); | |
648 | } | |
649 | ||
6d36483b | 650 | /* A constant integer suitable for use in a POST_MODIFY memory |
757d4970 | 651 | reference. */ |
652 | int | |
5c1d8983 | 653 | post_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
757d4970 | 654 | { |
655 | return (GET_CODE (op) == CONST_INT | |
656 | && INTVAL (op) < 0x2000 && INTVAL (op) >= -0x10); | |
657 | } | |
658 | ||
87ad11b0 | 659 | int |
5c1d8983 | 660 | arith_double_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 661 | { |
662 | return (register_operand (op, mode) | |
663 | || (GET_CODE (op) == CONST_DOUBLE | |
664 | && GET_MODE (op) == mode | |
665 | && VAL_14_BITS_P (CONST_DOUBLE_LOW (op)) | |
1fe9d310 | 666 | && ((CONST_DOUBLE_HIGH (op) >= 0) |
87ad11b0 | 667 | == ((CONST_DOUBLE_LOW (op) & 0x1000) == 0)))); |
668 | } | |
669 | ||
45981c0a | 670 | /* Return truth value of whether OP is an integer which fits the |
546a40bd | 671 | range constraining immediate operands in three-address insns, or |
672 | is an integer register. */ | |
673 | ||
674 | int | |
5c1d8983 | 675 | ireg_or_int5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
546a40bd | 676 | { |
677 | return ((GET_CODE (op) == CONST_INT && INT_5_BITS (op)) | |
678 | || (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32)); | |
679 | } | |
680 | ||
c7ae93c8 | 681 | /* Return nonzero if OP is an integer register, else return zero. */ |
682 | int | |
5c1d8983 | 683 | ireg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
c7ae93c8 | 684 | { |
685 | return (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32); | |
686 | } | |
687 | ||
45981c0a | 688 | /* Return truth value of whether OP is an integer which fits the |
87ad11b0 | 689 | range constraining immediate operands in three-address insns. */ |
690 | ||
691 | int | |
5c1d8983 | 692 | int5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 693 | { |
694 | return (GET_CODE (op) == CONST_INT && INT_5_BITS (op)); | |
695 | } | |
696 | ||
697 | int | |
5c1d8983 | 698 | uint5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 699 | { |
700 | return (GET_CODE (op) == CONST_INT && INT_U5_BITS (op)); | |
701 | } | |
702 | ||
87ad11b0 | 703 | int |
5c1d8983 | 704 | int11_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 705 | { |
6d36483b | 706 | return (GET_CODE (op) == CONST_INT && INT_11_BITS (op)); |
707 | } | |
708 | ||
709 | int | |
5c1d8983 | 710 | uint32_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
6d36483b | 711 | { |
712 | #if HOST_BITS_PER_WIDE_INT > 32 | |
713 | /* All allowed constants will fit a CONST_INT. */ | |
714 | return (GET_CODE (op) == CONST_INT | |
b7d86581 | 715 | && (INTVAL (op) >= 0 && INTVAL (op) < (HOST_WIDE_INT) 1 << 32)); |
6d36483b | 716 | #else |
717 | return (GET_CODE (op) == CONST_INT | |
718 | || (GET_CODE (op) == CONST_DOUBLE | |
719 | && CONST_DOUBLE_HIGH (op) == 0)); | |
720 | #endif | |
87ad11b0 | 721 | } |
722 | ||
723 | int | |
5c1d8983 | 724 | arith5_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 725 | { |
726 | return register_operand (op, mode) || int5_operand (op, mode); | |
727 | } | |
728 | ||
ea52c577 | 729 | /* True iff zdepi can be used to generate this CONST_INT. |
730 | zdepi first sign extends a 5 bit signed number to a given field | |
731 | length, then places this field anywhere in a zero. */ | |
e057641f | 732 | int |
5c1d8983 | 733 | zdepi_cint_p (unsigned HOST_WIDE_INT x) |
fad0b60f | 734 | { |
3745c59b | 735 | unsigned HOST_WIDE_INT lsb_mask, t; |
fad0b60f | 736 | |
737 | /* This might not be obvious, but it's at least fast. | |
01cc3b75 | 738 | This function is critical; we don't have the time loops would take. */ |
42faba01 | 739 | lsb_mask = x & -x; |
740 | t = ((x >> 4) + lsb_mask) & ~(lsb_mask - 1); | |
741 | /* Return true iff t is a power of two. */ | |
fad0b60f | 742 | return ((t & (t - 1)) == 0); |
743 | } | |
744 | ||
6d36483b | 745 | /* True iff depi or extru can be used to compute (reg & mask). |
746 | Accept bit pattern like these: | |
747 | 0....01....1 | |
748 | 1....10....0 | |
749 | 1..10..01..1 */ | |
e057641f | 750 | int |
5c1d8983 | 751 | and_mask_p (unsigned HOST_WIDE_INT mask) |
e057641f | 752 | { |
753 | mask = ~mask; | |
754 | mask += mask & -mask; | |
755 | return (mask & (mask - 1)) == 0; | |
756 | } | |
757 | ||
758 | /* True iff depi or extru can be used to compute (reg & OP). */ | |
759 | int | |
5c1d8983 | 760 | and_operand (rtx op, enum machine_mode mode) |
e057641f | 761 | { |
762 | return (register_operand (op, mode) | |
42faba01 | 763 | || (GET_CODE (op) == CONST_INT && and_mask_p (INTVAL (op)))); |
e057641f | 764 | } |
765 | ||
766 | /* True iff depi can be used to compute (reg | MASK). */ | |
767 | int | |
5c1d8983 | 768 | ior_mask_p (unsigned HOST_WIDE_INT mask) |
e057641f | 769 | { |
770 | mask += mask & -mask; | |
771 | return (mask & (mask - 1)) == 0; | |
772 | } | |
773 | ||
774 | /* True iff depi can be used to compute (reg | OP). */ | |
775 | int | |
5c1d8983 | 776 | ior_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
e057641f | 777 | { |
b744c8cb | 778 | return (GET_CODE (op) == CONST_INT && ior_mask_p (INTVAL (op))); |
e057641f | 779 | } |
780 | ||
e5965947 | 781 | int |
5c1d8983 | 782 | lhs_lshift_operand (rtx op, enum machine_mode mode) |
e5965947 | 783 | { |
784 | return register_operand (op, mode) || lhs_lshift_cint_operand (op, mode); | |
785 | } | |
786 | ||
787 | /* True iff OP is a CONST_INT of the forms 0...0xxxx or 0...01...1xxxx. | |
788 | Such values can be the left hand side x in (x << r), using the zvdepi | |
789 | instruction. */ | |
790 | int | |
5c1d8983 | 791 | lhs_lshift_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
e5965947 | 792 | { |
3745c59b | 793 | unsigned HOST_WIDE_INT x; |
e5965947 | 794 | if (GET_CODE (op) != CONST_INT) |
795 | return 0; | |
796 | x = INTVAL (op) >> 4; | |
797 | return (x & (x + 1)) == 0; | |
798 | } | |
799 | ||
9c6d4825 | 800 | int |
5c1d8983 | 801 | arith32_operand (rtx op, enum machine_mode mode) |
9c6d4825 | 802 | { |
803 | return register_operand (op, mode) || GET_CODE (op) == CONST_INT; | |
804 | } | |
ead9285f | 805 | |
806 | int | |
5c1d8983 | 807 | pc_or_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
ead9285f | 808 | { |
809 | return (GET_CODE (op) == PC || GET_CODE (op) == LABEL_REF); | |
810 | } | |
87ad11b0 | 811 | \f |
812 | /* Legitimize PIC addresses. If the address is already | |
813 | position-independent, we return ORIG. Newly generated | |
814 | position-independent addresses go to REG. If we need more | |
815 | than one register, we lose. */ | |
816 | ||
817 | rtx | |
5c1d8983 | 818 | legitimize_pic_address (rtx orig, enum machine_mode mode, rtx reg) |
87ad11b0 | 819 | { |
820 | rtx pic_ref = orig; | |
821 | ||
b090827b | 822 | /* Labels need special handling. */ |
611a88e1 | 823 | if (pic_label_operand (orig, mode)) |
b4a7bf10 | 824 | { |
2536cc16 | 825 | /* We do not want to go through the movXX expanders here since that |
826 | would create recursion. | |
827 | ||
828 | Nor do we really want to call a generator for a named pattern | |
829 | since that requires multiple patterns if we want to support | |
830 | multiple word sizes. | |
831 | ||
832 | So instead we just emit the raw set, which avoids the movXX | |
833 | expanders completely. */ | |
834 | emit_insn (gen_rtx_SET (VOIDmode, reg, orig)); | |
b4a7bf10 | 835 | current_function_uses_pic_offset_table = 1; |
836 | return reg; | |
837 | } | |
87ad11b0 | 838 | if (GET_CODE (orig) == SYMBOL_REF) |
839 | { | |
f9c1ba9d | 840 | rtx insn, tmp_reg; |
841 | ||
87ad11b0 | 842 | if (reg == 0) |
843 | abort (); | |
844 | ||
f9c1ba9d | 845 | /* Before reload, allocate a temporary register for the intermediate |
846 | result. This allows the sequence to be deleted when the final | |
847 | result is unused and the insns are trivially dead. */ | |
848 | tmp_reg = ((reload_in_progress || reload_completed) | |
849 | ? reg : gen_reg_rtx (Pmode)); | |
850 | ||
851 | emit_move_insn (tmp_reg, | |
fc44315f | 852 | gen_rtx_PLUS (word_mode, pic_offset_table_rtx, |
853 | gen_rtx_HIGH (word_mode, orig))); | |
854 | pic_ref | |
855 | = gen_rtx_MEM (Pmode, | |
f9c1ba9d | 856 | gen_rtx_LO_SUM (Pmode, tmp_reg, |
fc44315f | 857 | gen_rtx_UNSPEC (Pmode, |
858 | gen_rtvec (1, orig), | |
859 | 0))); | |
7014838c | 860 | |
87ad11b0 | 861 | current_function_uses_pic_offset_table = 1; |
f9c1ba9d | 862 | MEM_NOTRAP_P (pic_ref) = 1; |
87ad11b0 | 863 | RTX_UNCHANGING_P (pic_ref) = 1; |
f9c1ba9d | 864 | insn = emit_move_insn (reg, pic_ref); |
865 | ||
866 | /* Put a REG_EQUAL note on this insn, so that it can be optimized. */ | |
867 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig, REG_NOTES (insn)); | |
868 | ||
87ad11b0 | 869 | return reg; |
870 | } | |
871 | else if (GET_CODE (orig) == CONST) | |
872 | { | |
57ed30e5 | 873 | rtx base; |
87ad11b0 | 874 | |
875 | if (GET_CODE (XEXP (orig, 0)) == PLUS | |
876 | && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx) | |
877 | return orig; | |
878 | ||
879 | if (reg == 0) | |
880 | abort (); | |
881 | ||
882 | if (GET_CODE (XEXP (orig, 0)) == PLUS) | |
883 | { | |
884 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg); | |
885 | orig = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode, | |
886 | base == reg ? 0 : reg); | |
887 | } | |
888 | else abort (); | |
889 | if (GET_CODE (orig) == CONST_INT) | |
890 | { | |
42faba01 | 891 | if (INT_14_BITS (orig)) |
b244d4c7 | 892 | return plus_constant (base, INTVAL (orig)); |
87ad11b0 | 893 | orig = force_reg (Pmode, orig); |
894 | } | |
ad851752 | 895 | pic_ref = gen_rtx_PLUS (Pmode, base, orig); |
87ad11b0 | 896 | /* Likewise, should we set special REG_NOTEs here? */ |
897 | } | |
898 | return pic_ref; | |
899 | } | |
900 | ||
347b5848 | 901 | /* Try machine-dependent ways of modifying an illegitimate address |
902 | to be legitimate. If we find one, return the new, valid address. | |
903 | This macro is used in only one place: `memory_address' in explow.c. | |
904 | ||
905 | OLDX is the address as it was before break_out_memory_refs was called. | |
906 | In some cases it is useful to look at this to decide what needs to be done. | |
907 | ||
908 | MODE and WIN are passed so that this macro can use | |
909 | GO_IF_LEGITIMATE_ADDRESS. | |
910 | ||
911 | It is always safe for this macro to do nothing. It exists to recognize | |
6d36483b | 912 | opportunities to optimize the output. |
347b5848 | 913 | |
914 | For the PA, transform: | |
915 | ||
916 | memory(X + <large int>) | |
917 | ||
918 | into: | |
919 | ||
920 | if (<large int> & mask) >= 16 | |
921 | Y = (<large int> & ~mask) + mask + 1 Round up. | |
922 | else | |
923 | Y = (<large int> & ~mask) Round down. | |
924 | Z = X + Y | |
925 | memory (Z + (<large int> - Y)); | |
926 | ||
6d36483b | 927 | This is for CSE to find several similar references, and only use one Z. |
347b5848 | 928 | |
929 | X can either be a SYMBOL_REF or REG, but because combine can not | |
930 | perform a 4->2 combination we do nothing for SYMBOL_REF + D where | |
931 | D will not fit in 14 bits. | |
932 | ||
933 | MODE_FLOAT references allow displacements which fit in 5 bits, so use | |
6d36483b | 934 | 0x1f as the mask. |
347b5848 | 935 | |
936 | MODE_INT references allow displacements which fit in 14 bits, so use | |
6d36483b | 937 | 0x3fff as the mask. |
347b5848 | 938 | |
939 | This relies on the fact that most mode MODE_FLOAT references will use FP | |
940 | registers and most mode MODE_INT references will use integer registers. | |
941 | (In the rare case of an FP register used in an integer MODE, we depend | |
942 | on secondary reloads to clean things up.) | |
943 | ||
944 | ||
945 | It is also beneficial to handle (plus (mult (X) (Y)) (Z)) in a special | |
946 | manner if Y is 2, 4, or 8. (allows more shadd insns and shifted indexed | |
01cc3b75 | 947 | addressing modes to be used). |
347b5848 | 948 | |
949 | Put X and Z into registers. Then put the entire expression into | |
950 | a register. */ | |
951 | ||
952 | rtx | |
5c1d8983 | 953 | hppa_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, |
954 | enum machine_mode mode) | |
347b5848 | 955 | { |
347b5848 | 956 | rtx orig = x; |
957 | ||
b4a7bf10 | 958 | if (flag_pic) |
959 | return legitimize_pic_address (x, mode, gen_reg_rtx (Pmode)); | |
960 | ||
6dc3b0d9 | 961 | /* Strip off CONST. */ |
347b5848 | 962 | if (GET_CODE (x) == CONST) |
963 | x = XEXP (x, 0); | |
964 | ||
42819d4e | 965 | /* Special case. Get the SYMBOL_REF into a register and use indexing. |
966 | That should always be safe. */ | |
967 | if (GET_CODE (x) == PLUS | |
968 | && GET_CODE (XEXP (x, 0)) == REG | |
969 | && GET_CODE (XEXP (x, 1)) == SYMBOL_REF) | |
970 | { | |
440c23df | 971 | rtx reg = force_reg (Pmode, XEXP (x, 1)); |
972 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, reg, XEXP (x, 0))); | |
42819d4e | 973 | } |
974 | ||
166bf021 | 975 | /* Note we must reject symbols which represent function addresses |
976 | since the assembler/linker can't handle arithmetic on plabels. */ | |
347b5848 | 977 | if (GET_CODE (x) == PLUS |
978 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
166bf021 | 979 | && ((GET_CODE (XEXP (x, 0)) == SYMBOL_REF |
980 | && !FUNCTION_NAME_P (XSTR (XEXP (x, 0), 0))) | |
347b5848 | 981 | || GET_CODE (XEXP (x, 0)) == REG)) |
982 | { | |
983 | rtx int_part, ptr_reg; | |
984 | int newoffset; | |
985 | int offset = INTVAL (XEXP (x, 1)); | |
1b6f11e2 | 986 | int mask; |
987 | ||
988 | mask = (GET_MODE_CLASS (mode) == MODE_FLOAT | |
989 | ? (TARGET_PA_20 ? 0x3fff : 0x1f) : 0x3fff); | |
347b5848 | 990 | |
6d36483b | 991 | /* Choose which way to round the offset. Round up if we |
347b5848 | 992 | are >= halfway to the next boundary. */ |
993 | if ((offset & mask) >= ((mask + 1) / 2)) | |
994 | newoffset = (offset & ~ mask) + mask + 1; | |
995 | else | |
996 | newoffset = (offset & ~ mask); | |
997 | ||
998 | /* If the newoffset will not fit in 14 bits (ldo), then | |
999 | handling this would take 4 or 5 instructions (2 to load | |
1000 | the SYMBOL_REF + 1 or 2 to load the newoffset + 1 to | |
1001 | add the new offset and the SYMBOL_REF.) Combine can | |
1002 | not handle 4->2 or 5->2 combinations, so do not create | |
1003 | them. */ | |
1004 | if (! VAL_14_BITS_P (newoffset) | |
1005 | && GET_CODE (XEXP (x, 0)) == SYMBOL_REF) | |
1006 | { | |
7014838c | 1007 | rtx const_part = plus_constant (XEXP (x, 0), newoffset); |
347b5848 | 1008 | rtx tmp_reg |
339613b4 | 1009 | = force_reg (Pmode, |
ad851752 | 1010 | gen_rtx_HIGH (Pmode, const_part)); |
347b5848 | 1011 | ptr_reg |
339613b4 | 1012 | = force_reg (Pmode, |
7014838c | 1013 | gen_rtx_LO_SUM (Pmode, |
1014 | tmp_reg, const_part)); | |
347b5848 | 1015 | } |
1016 | else | |
1017 | { | |
1018 | if (! VAL_14_BITS_P (newoffset)) | |
339613b4 | 1019 | int_part = force_reg (Pmode, GEN_INT (newoffset)); |
347b5848 | 1020 | else |
1021 | int_part = GEN_INT (newoffset); | |
1022 | ||
339613b4 | 1023 | ptr_reg = force_reg (Pmode, |
ad851752 | 1024 | gen_rtx_PLUS (Pmode, |
1025 | force_reg (Pmode, XEXP (x, 0)), | |
1026 | int_part)); | |
347b5848 | 1027 | } |
1028 | return plus_constant (ptr_reg, offset - newoffset); | |
1029 | } | |
45f1285a | 1030 | |
5115683e | 1031 | /* Handle (plus (mult (a) (shadd_constant)) (b)). */ |
45f1285a | 1032 | |
347b5848 | 1033 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT |
1034 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
45f1285a | 1035 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1))) |
1036 | && (GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) == 'o' | |
1037 | || GET_CODE (XEXP (x, 1)) == SUBREG) | |
1038 | && GET_CODE (XEXP (x, 1)) != CONST) | |
347b5848 | 1039 | { |
1040 | int val = INTVAL (XEXP (XEXP (x, 0), 1)); | |
1041 | rtx reg1, reg2; | |
5115683e | 1042 | |
1043 | reg1 = XEXP (x, 1); | |
1044 | if (GET_CODE (reg1) != REG) | |
1045 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1046 | ||
1047 | reg2 = XEXP (XEXP (x, 0), 0); | |
1048 | if (GET_CODE (reg2) != REG) | |
1049 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1050 | ||
ad851752 | 1051 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, |
7014838c | 1052 | gen_rtx_MULT (Pmode, |
1053 | reg2, | |
1054 | GEN_INT (val)), | |
ad851752 | 1055 | reg1)); |
347b5848 | 1056 | } |
45f1285a | 1057 | |
00a87639 | 1058 | /* Similarly for (plus (plus (mult (a) (shadd_constant)) (b)) (c)). |
1059 | ||
1060 | Only do so for floating point modes since this is more speculative | |
1061 | and we lose if it's an integer store. */ | |
5115683e | 1062 | if (GET_CODE (x) == PLUS |
00a87639 | 1063 | && GET_CODE (XEXP (x, 0)) == PLUS |
1064 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
1065 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
5115683e | 1066 | && shadd_constant_p (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))) |
1067 | && (mode == SFmode || mode == DFmode)) | |
00a87639 | 1068 | { |
5115683e | 1069 | |
1070 | /* First, try and figure out what to use as a base register. */ | |
1071 | rtx reg1, reg2, base, idx, orig_base; | |
1072 | ||
1073 | reg1 = XEXP (XEXP (x, 0), 1); | |
1074 | reg2 = XEXP (x, 1); | |
1075 | base = NULL_RTX; | |
1076 | idx = NULL_RTX; | |
1077 | ||
1078 | /* Make sure they're both regs. If one was a SYMBOL_REF [+ const], | |
e61a0a7f | 1079 | then emit_move_sequence will turn on REG_POINTER so we'll know |
1080 | it's a base register below. */ | |
5115683e | 1081 | if (GET_CODE (reg1) != REG) |
1082 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1083 | ||
1084 | if (GET_CODE (reg2) != REG) | |
1085 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1086 | ||
1087 | /* Figure out what the base and index are. */ | |
9840d99d | 1088 | |
5115683e | 1089 | if (GET_CODE (reg1) == REG |
e61a0a7f | 1090 | && REG_POINTER (reg1)) |
5115683e | 1091 | { |
1092 | base = reg1; | |
1093 | orig_base = XEXP (XEXP (x, 0), 1); | |
ad851752 | 1094 | idx = gen_rtx_PLUS (Pmode, |
1095 | gen_rtx_MULT (Pmode, | |
1096 | XEXP (XEXP (XEXP (x, 0), 0), 0), | |
1097 | XEXP (XEXP (XEXP (x, 0), 0), 1)), | |
1098 | XEXP (x, 1)); | |
5115683e | 1099 | } |
1100 | else if (GET_CODE (reg2) == REG | |
e61a0a7f | 1101 | && REG_POINTER (reg2)) |
5115683e | 1102 | { |
1103 | base = reg2; | |
1104 | orig_base = XEXP (x, 1); | |
1105 | idx = XEXP (x, 0); | |
1106 | } | |
1107 | ||
1108 | if (base == 0) | |
21f3ee9c | 1109 | return orig; |
5115683e | 1110 | |
1111 | /* If the index adds a large constant, try to scale the | |
1112 | constant so that it can be loaded with only one insn. */ | |
1113 | if (GET_CODE (XEXP (idx, 1)) == CONST_INT | |
1114 | && VAL_14_BITS_P (INTVAL (XEXP (idx, 1)) | |
1115 | / INTVAL (XEXP (XEXP (idx, 0), 1))) | |
1116 | && INTVAL (XEXP (idx, 1)) % INTVAL (XEXP (XEXP (idx, 0), 1)) == 0) | |
1117 | { | |
1118 | /* Divide the CONST_INT by the scale factor, then add it to A. */ | |
1119 | int val = INTVAL (XEXP (idx, 1)); | |
1120 | ||
1121 | val /= INTVAL (XEXP (XEXP (idx, 0), 1)); | |
1122 | reg1 = XEXP (XEXP (idx, 0), 0); | |
1123 | if (GET_CODE (reg1) != REG) | |
1124 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1125 | ||
ad851752 | 1126 | reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, reg1, GEN_INT (val))); |
5115683e | 1127 | |
1128 | /* We can now generate a simple scaled indexed address. */ | |
7014838c | 1129 | return |
1130 | force_reg | |
1131 | (Pmode, gen_rtx_PLUS (Pmode, | |
1132 | gen_rtx_MULT (Pmode, reg1, | |
1133 | XEXP (XEXP (idx, 0), 1)), | |
1134 | base)); | |
5115683e | 1135 | } |
1136 | ||
1137 | /* If B + C is still a valid base register, then add them. */ | |
1138 | if (GET_CODE (XEXP (idx, 1)) == CONST_INT | |
1139 | && INTVAL (XEXP (idx, 1)) <= 4096 | |
1140 | && INTVAL (XEXP (idx, 1)) >= -4096) | |
1141 | { | |
1142 | int val = INTVAL (XEXP (XEXP (idx, 0), 1)); | |
1143 | rtx reg1, reg2; | |
1144 | ||
ad851752 | 1145 | reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, base, XEXP (idx, 1))); |
5115683e | 1146 | |
1147 | reg2 = XEXP (XEXP (idx, 0), 0); | |
1148 | if (GET_CODE (reg2) != CONST_INT) | |
1149 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1150 | ||
ad851752 | 1151 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, |
7014838c | 1152 | gen_rtx_MULT (Pmode, |
1153 | reg2, | |
ad851752 | 1154 | GEN_INT (val)), |
1155 | reg1)); | |
5115683e | 1156 | } |
1157 | ||
1158 | /* Get the index into a register, then add the base + index and | |
1159 | return a register holding the result. */ | |
1160 | ||
1161 | /* First get A into a register. */ | |
1162 | reg1 = XEXP (XEXP (idx, 0), 0); | |
1163 | if (GET_CODE (reg1) != REG) | |
1164 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1165 | ||
1166 | /* And get B into a register. */ | |
1167 | reg2 = XEXP (idx, 1); | |
1168 | if (GET_CODE (reg2) != REG) | |
1169 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1170 | ||
ad851752 | 1171 | reg1 = force_reg (Pmode, |
1172 | gen_rtx_PLUS (Pmode, | |
1173 | gen_rtx_MULT (Pmode, reg1, | |
1174 | XEXP (XEXP (idx, 0), 1)), | |
1175 | reg2)); | |
5115683e | 1176 | |
1177 | /* Add the result to our base register and return. */ | |
ad851752 | 1178 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, base, reg1)); |
9840d99d | 1179 | |
00a87639 | 1180 | } |
1181 | ||
6d36483b | 1182 | /* Uh-oh. We might have an address for x[n-100000]. This needs |
fb5390c1 | 1183 | special handling to avoid creating an indexed memory address |
1184 | with x-100000 as the base. | |
9840d99d | 1185 | |
fb5390c1 | 1186 | If the constant part is small enough, then it's still safe because |
1187 | there is a guard page at the beginning and end of the data segment. | |
1188 | ||
1189 | Scaled references are common enough that we want to try and rearrange the | |
1190 | terms so that we can use indexing for these addresses too. Only | |
00a87639 | 1191 | do the optimization for floatint point modes. */ |
45f1285a | 1192 | |
fb5390c1 | 1193 | if (GET_CODE (x) == PLUS |
1194 | && symbolic_expression_p (XEXP (x, 1))) | |
45f1285a | 1195 | { |
1196 | /* Ugly. We modify things here so that the address offset specified | |
1197 | by the index expression is computed first, then added to x to form | |
fb5390c1 | 1198 | the entire address. */ |
45f1285a | 1199 | |
00a87639 | 1200 | rtx regx1, regx2, regy1, regy2, y; |
45f1285a | 1201 | |
1202 | /* Strip off any CONST. */ | |
1203 | y = XEXP (x, 1); | |
1204 | if (GET_CODE (y) == CONST) | |
1205 | y = XEXP (y, 0); | |
1206 | ||
7ee96d6e | 1207 | if (GET_CODE (y) == PLUS || GET_CODE (y) == MINUS) |
1208 | { | |
00a87639 | 1209 | /* See if this looks like |
1210 | (plus (mult (reg) (shadd_const)) | |
1211 | (const (plus (symbol_ref) (const_int)))) | |
1212 | ||
5115683e | 1213 | Where const_int is small. In that case the const |
9840d99d | 1214 | expression is a valid pointer for indexing. |
5115683e | 1215 | |
1216 | If const_int is big, but can be divided evenly by shadd_const | |
1217 | and added to (reg). This allows more scaled indexed addresses. */ | |
1218 | if (GET_CODE (XEXP (y, 0)) == SYMBOL_REF | |
1219 | && GET_CODE (XEXP (x, 0)) == MULT | |
00a87639 | 1220 | && GET_CODE (XEXP (y, 1)) == CONST_INT |
5115683e | 1221 | && INTVAL (XEXP (y, 1)) >= -4096 |
1222 | && INTVAL (XEXP (y, 1)) <= 4095 | |
1223 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
1224 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))) | |
1225 | { | |
1226 | int val = INTVAL (XEXP (XEXP (x, 0), 1)); | |
1227 | rtx reg1, reg2; | |
1228 | ||
1229 | reg1 = XEXP (x, 1); | |
1230 | if (GET_CODE (reg1) != REG) | |
1231 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1232 | ||
1233 | reg2 = XEXP (XEXP (x, 0), 0); | |
1234 | if (GET_CODE (reg2) != REG) | |
1235 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1236 | ||
ad851752 | 1237 | return force_reg (Pmode, |
1238 | gen_rtx_PLUS (Pmode, | |
7014838c | 1239 | gen_rtx_MULT (Pmode, |
1240 | reg2, | |
ad851752 | 1241 | GEN_INT (val)), |
7014838c | 1242 | reg1)); |
5115683e | 1243 | } |
1244 | else if ((mode == DFmode || mode == SFmode) | |
1245 | && GET_CODE (XEXP (y, 0)) == SYMBOL_REF | |
1246 | && GET_CODE (XEXP (x, 0)) == MULT | |
1247 | && GET_CODE (XEXP (y, 1)) == CONST_INT | |
1248 | && INTVAL (XEXP (y, 1)) % INTVAL (XEXP (XEXP (x, 0), 1)) == 0 | |
1249 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
1250 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))) | |
00a87639 | 1251 | { |
1252 | regx1 | |
1253 | = force_reg (Pmode, GEN_INT (INTVAL (XEXP (y, 1)) | |
1254 | / INTVAL (XEXP (XEXP (x, 0), 1)))); | |
1255 | regx2 = XEXP (XEXP (x, 0), 0); | |
1256 | if (GET_CODE (regx2) != REG) | |
1257 | regx2 = force_reg (Pmode, force_operand (regx2, 0)); | |
ad851752 | 1258 | regx2 = force_reg (Pmode, gen_rtx_fmt_ee (GET_CODE (y), Pmode, |
1259 | regx2, regx1)); | |
7014838c | 1260 | return |
1261 | force_reg (Pmode, | |
1262 | gen_rtx_PLUS (Pmode, | |
1263 | gen_rtx_MULT (Pmode, regx2, | |
1264 | XEXP (XEXP (x, 0), 1)), | |
1265 | force_reg (Pmode, XEXP (y, 0)))); | |
00a87639 | 1266 | } |
fb5390c1 | 1267 | else if (GET_CODE (XEXP (y, 1)) == CONST_INT |
1268 | && INTVAL (XEXP (y, 1)) >= -4096 | |
1269 | && INTVAL (XEXP (y, 1)) <= 4095) | |
1270 | { | |
1271 | /* This is safe because of the guard page at the | |
1272 | beginning and end of the data space. Just | |
1273 | return the original address. */ | |
1274 | return orig; | |
1275 | } | |
00a87639 | 1276 | else |
1277 | { | |
1278 | /* Doesn't look like one we can optimize. */ | |
1279 | regx1 = force_reg (Pmode, force_operand (XEXP (x, 0), 0)); | |
1280 | regy1 = force_reg (Pmode, force_operand (XEXP (y, 0), 0)); | |
1281 | regy2 = force_reg (Pmode, force_operand (XEXP (y, 1), 0)); | |
1282 | regx1 = force_reg (Pmode, | |
ad851752 | 1283 | gen_rtx_fmt_ee (GET_CODE (y), Pmode, |
1284 | regx1, regy2)); | |
1285 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, regx1, regy1)); | |
00a87639 | 1286 | } |
7ee96d6e | 1287 | } |
45f1285a | 1288 | } |
1289 | ||
347b5848 | 1290 | return orig; |
1291 | } | |
1292 | ||
87ad11b0 | 1293 | /* For the HPPA, REG and REG+CONST is cost 0 |
1294 | and addresses involving symbolic constants are cost 2. | |
1295 | ||
1296 | PIC addresses are very expensive. | |
1297 | ||
1298 | It is no coincidence that this has the same structure | |
1299 | as GO_IF_LEGITIMATE_ADDRESS. */ | |
ec0457a8 | 1300 | |
1301 | static int | |
5c1d8983 | 1302 | hppa_address_cost (rtx X) |
87ad11b0 | 1303 | { |
ec0457a8 | 1304 | switch (GET_CODE (X)) |
1305 | { | |
1306 | case REG: | |
1307 | case PLUS: | |
1308 | case LO_SUM: | |
87ad11b0 | 1309 | return 1; |
ec0457a8 | 1310 | case HIGH: |
1311 | return 2; | |
1312 | default: | |
1313 | return 4; | |
1314 | } | |
87ad11b0 | 1315 | } |
1316 | ||
fab7adbf | 1317 | /* Compute a (partial) cost for rtx X. Return true if the complete |
1318 | cost has been computed, and false if subexpressions should be | |
1319 | scanned. In either case, *TOTAL contains the cost result. */ | |
1320 | ||
1321 | static bool | |
5c1d8983 | 1322 | hppa_rtx_costs (rtx x, int code, int outer_code, int *total) |
fab7adbf | 1323 | { |
1324 | switch (code) | |
1325 | { | |
1326 | case CONST_INT: | |
1327 | if (INTVAL (x) == 0) | |
1328 | *total = 0; | |
1329 | else if (INT_14_BITS (x)) | |
1330 | *total = 1; | |
1331 | else | |
1332 | *total = 2; | |
1333 | return true; | |
1334 | ||
1335 | case HIGH: | |
1336 | *total = 2; | |
1337 | return true; | |
1338 | ||
1339 | case CONST: | |
1340 | case LABEL_REF: | |
1341 | case SYMBOL_REF: | |
1342 | *total = 4; | |
1343 | return true; | |
1344 | ||
1345 | case CONST_DOUBLE: | |
1346 | if ((x == CONST0_RTX (DFmode) || x == CONST0_RTX (SFmode)) | |
1347 | && outer_code != SET) | |
1348 | *total = 0; | |
1349 | else | |
1350 | *total = 8; | |
1351 | return true; | |
1352 | ||
1353 | case MULT: | |
1354 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1355 | *total = COSTS_N_INSNS (3); | |
1356 | else if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT) | |
1357 | *total = COSTS_N_INSNS (8); | |
1358 | else | |
1359 | *total = COSTS_N_INSNS (20); | |
1360 | return true; | |
1361 | ||
1362 | case DIV: | |
1363 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1364 | { | |
1365 | *total = COSTS_N_INSNS (14); | |
1366 | return true; | |
1367 | } | |
1368 | /* FALLTHRU */ | |
1369 | ||
1370 | case UDIV: | |
1371 | case MOD: | |
1372 | case UMOD: | |
1373 | *total = COSTS_N_INSNS (60); | |
1374 | return true; | |
1375 | ||
1376 | case PLUS: /* this includes shNadd insns */ | |
1377 | case MINUS: | |
1378 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1379 | *total = COSTS_N_INSNS (3); | |
1380 | else | |
1381 | *total = COSTS_N_INSNS (1); | |
1382 | return true; | |
1383 | ||
1384 | case ASHIFT: | |
1385 | case ASHIFTRT: | |
1386 | case LSHIFTRT: | |
1387 | *total = COSTS_N_INSNS (1); | |
1388 | return true; | |
1389 | ||
1390 | default: | |
1391 | return false; | |
1392 | } | |
1393 | } | |
1394 | ||
9840d99d | 1395 | /* Ensure mode of ORIG, a REG rtx, is MODE. Returns either ORIG or a |
1396 | new rtx with the correct mode. */ | |
1397 | static inline rtx | |
5c1d8983 | 1398 | force_mode (enum machine_mode mode, rtx orig) |
9840d99d | 1399 | { |
1400 | if (mode == GET_MODE (orig)) | |
1401 | return orig; | |
1402 | ||
1403 | if (REGNO (orig) >= FIRST_PSEUDO_REGISTER) | |
1404 | abort (); | |
1405 | ||
1406 | return gen_rtx_REG (mode, REGNO (orig)); | |
1407 | } | |
1408 | ||
87ad11b0 | 1409 | /* Emit insns to move operands[1] into operands[0]. |
1410 | ||
1411 | Return 1 if we have written out everything that needs to be done to | |
1412 | do the move. Otherwise, return 0 and the caller will emit the move | |
9840d99d | 1413 | normally. |
f756078b | 1414 | |
1415 | Note SCRATCH_REG may not be in the proper mode depending on how it | |
2cecd772 | 1416 | will be used. This routine is responsible for creating a new copy |
f756078b | 1417 | of SCRATCH_REG in the proper mode. */ |
87ad11b0 | 1418 | |
1419 | int | |
5c1d8983 | 1420 | emit_move_sequence (rtx *operands, enum machine_mode mode, rtx scratch_reg) |
87ad11b0 | 1421 | { |
1422 | register rtx operand0 = operands[0]; | |
1423 | register rtx operand1 = operands[1]; | |
4a155b0f | 1424 | register rtx tem; |
87ad11b0 | 1425 | |
2d4dc8d2 | 1426 | if (scratch_reg |
1427 | && reload_in_progress && GET_CODE (operand0) == REG | |
d1e2bb73 | 1428 | && REGNO (operand0) >= FIRST_PSEUDO_REGISTER) |
1429 | operand0 = reg_equiv_mem[REGNO (operand0)]; | |
2d4dc8d2 | 1430 | else if (scratch_reg |
1431 | && reload_in_progress && GET_CODE (operand0) == SUBREG | |
d1e2bb73 | 1432 | && GET_CODE (SUBREG_REG (operand0)) == REG |
1433 | && REGNO (SUBREG_REG (operand0)) >= FIRST_PSEUDO_REGISTER) | |
a3afad75 | 1434 | { |
701e46d0 | 1435 | /* We must not alter SUBREG_BYTE (operand0) since that would confuse |
d3c3f88d | 1436 | the code which tracks sets/uses for delete_output_reload. */ |
1437 | rtx temp = gen_rtx_SUBREG (GET_MODE (operand0), | |
1438 | reg_equiv_mem [REGNO (SUBREG_REG (operand0))], | |
701e46d0 | 1439 | SUBREG_BYTE (operand0)); |
e00b80b2 | 1440 | operand0 = alter_subreg (&temp); |
a3afad75 | 1441 | } |
d1e2bb73 | 1442 | |
2d4dc8d2 | 1443 | if (scratch_reg |
1444 | && reload_in_progress && GET_CODE (operand1) == REG | |
d1e2bb73 | 1445 | && REGNO (operand1) >= FIRST_PSEUDO_REGISTER) |
1446 | operand1 = reg_equiv_mem[REGNO (operand1)]; | |
2d4dc8d2 | 1447 | else if (scratch_reg |
1448 | && reload_in_progress && GET_CODE (operand1) == SUBREG | |
d1e2bb73 | 1449 | && GET_CODE (SUBREG_REG (operand1)) == REG |
1450 | && REGNO (SUBREG_REG (operand1)) >= FIRST_PSEUDO_REGISTER) | |
a3afad75 | 1451 | { |
701e46d0 | 1452 | /* We must not alter SUBREG_BYTE (operand0) since that would confuse |
d3c3f88d | 1453 | the code which tracks sets/uses for delete_output_reload. */ |
1454 | rtx temp = gen_rtx_SUBREG (GET_MODE (operand1), | |
1455 | reg_equiv_mem [REGNO (SUBREG_REG (operand1))], | |
701e46d0 | 1456 | SUBREG_BYTE (operand1)); |
e00b80b2 | 1457 | operand1 = alter_subreg (&temp); |
a3afad75 | 1458 | } |
d1e2bb73 | 1459 | |
2d4dc8d2 | 1460 | if (scratch_reg && reload_in_progress && GET_CODE (operand0) == MEM |
4a155b0f | 1461 | && ((tem = find_replacement (&XEXP (operand0, 0))) |
1462 | != XEXP (operand0, 0))) | |
ad851752 | 1463 | operand0 = gen_rtx_MEM (GET_MODE (operand0), tem); |
2d4dc8d2 | 1464 | if (scratch_reg && reload_in_progress && GET_CODE (operand1) == MEM |
4a155b0f | 1465 | && ((tem = find_replacement (&XEXP (operand1, 0))) |
1466 | != XEXP (operand1, 0))) | |
ad851752 | 1467 | operand1 = gen_rtx_MEM (GET_MODE (operand1), tem); |
4a155b0f | 1468 | |
e8fdbafa | 1469 | /* Handle secondary reloads for loads/stores of FP registers from |
9840d99d | 1470 | REG+D addresses where D does not fit in 5 bits, including |
42819d4e | 1471 | (subreg (mem (addr))) cases. */ |
d6f01525 | 1472 | if (fp_reg_operand (operand0, mode) |
6b1c36c2 | 1473 | && ((GET_CODE (operand1) == MEM |
1474 | && ! memory_address_p (DFmode, XEXP (operand1, 0))) | |
1475 | || ((GET_CODE (operand1) == SUBREG | |
1476 | && GET_CODE (XEXP (operand1, 0)) == MEM | |
1477 | && !memory_address_p (DFmode, XEXP (XEXP (operand1, 0), 0))))) | |
d6f01525 | 1478 | && scratch_reg) |
1479 | { | |
6b1c36c2 | 1480 | if (GET_CODE (operand1) == SUBREG) |
1481 | operand1 = XEXP (operand1, 0); | |
1482 | ||
f756078b | 1483 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1484 | it in WORD_MODE regardless of what mode it was originally given | |
1485 | to us. */ | |
9840d99d | 1486 | scratch_reg = force_mode (word_mode, scratch_reg); |
7ee39d73 | 1487 | |
1488 | /* D might not fit in 14 bits either; for such cases load D into | |
1489 | scratch reg. */ | |
440c23df | 1490 | if (!memory_address_p (Pmode, XEXP (operand1, 0))) |
7ee39d73 | 1491 | { |
1492 | emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1)); | |
ad851752 | 1493 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, 0)), |
440c23df | 1494 | Pmode, |
ad851752 | 1495 | XEXP (XEXP (operand1, 0), 0), |
1496 | scratch_reg)); | |
7ee39d73 | 1497 | } |
1498 | else | |
1499 | emit_move_insn (scratch_reg, XEXP (operand1, 0)); | |
7014838c | 1500 | emit_insn (gen_rtx_SET (VOIDmode, operand0, |
1501 | gen_rtx_MEM (mode, scratch_reg))); | |
d6f01525 | 1502 | return 1; |
1503 | } | |
1504 | else if (fp_reg_operand (operand1, mode) | |
6b1c36c2 | 1505 | && ((GET_CODE (operand0) == MEM |
1506 | && ! memory_address_p (DFmode, XEXP (operand0, 0))) | |
1507 | || ((GET_CODE (operand0) == SUBREG) | |
1508 | && GET_CODE (XEXP (operand0, 0)) == MEM | |
1509 | && !memory_address_p (DFmode, XEXP (XEXP (operand0, 0), 0)))) | |
d6f01525 | 1510 | && scratch_reg) |
1511 | { | |
6b1c36c2 | 1512 | if (GET_CODE (operand0) == SUBREG) |
1513 | operand0 = XEXP (operand0, 0); | |
1514 | ||
f756078b | 1515 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1516 | it in WORD_MODE regardless of what mode it was originally given | |
1517 | to us. */ | |
9840d99d | 1518 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1519 | |
7ee39d73 | 1520 | /* D might not fit in 14 bits either; for such cases load D into |
1521 | scratch reg. */ | |
440c23df | 1522 | if (!memory_address_p (Pmode, XEXP (operand0, 0))) |
7ee39d73 | 1523 | { |
1524 | emit_move_insn (scratch_reg, XEXP (XEXP (operand0, 0), 1)); | |
ad851752 | 1525 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand0, |
1526 | 0)), | |
440c23df | 1527 | Pmode, |
ad851752 | 1528 | XEXP (XEXP (operand0, 0), |
1529 | 0), | |
1530 | scratch_reg)); | |
7ee39d73 | 1531 | } |
1532 | else | |
1533 | emit_move_insn (scratch_reg, XEXP (operand0, 0)); | |
ad851752 | 1534 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_MEM (mode, scratch_reg), |
1535 | operand1)); | |
d6f01525 | 1536 | return 1; |
1537 | } | |
753bd06a | 1538 | /* Handle secondary reloads for loads of FP registers from constant |
1539 | expressions by forcing the constant into memory. | |
1540 | ||
6d36483b | 1541 | use scratch_reg to hold the address of the memory location. |
753bd06a | 1542 | |
9dc4cbf1 | 1543 | The proper fix is to change PREFERRED_RELOAD_CLASS to return |
87fcb603 | 1544 | NO_REGS when presented with a const_int and a register class |
753bd06a | 1545 | containing only FP registers. Doing so unfortunately creates |
1546 | more problems than it solves. Fix this for 2.5. */ | |
1547 | else if (fp_reg_operand (operand0, mode) | |
1548 | && CONSTANT_P (operand1) | |
1549 | && scratch_reg) | |
1550 | { | |
1551 | rtx xoperands[2]; | |
1552 | ||
f756078b | 1553 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1554 | it in WORD_MODE regardless of what mode it was originally given | |
1555 | to us. */ | |
9840d99d | 1556 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1557 | |
753bd06a | 1558 | /* Force the constant into memory and put the address of the |
1559 | memory location into scratch_reg. */ | |
1560 | xoperands[0] = scratch_reg; | |
1561 | xoperands[1] = XEXP (force_const_mem (mode, operand1), 0); | |
8f258b49 | 1562 | emit_move_sequence (xoperands, Pmode, 0); |
753bd06a | 1563 | |
1564 | /* Now load the destination register. */ | |
7014838c | 1565 | emit_insn (gen_rtx_SET (mode, operand0, |
1566 | gen_rtx_MEM (mode, scratch_reg))); | |
753bd06a | 1567 | return 1; |
1568 | } | |
e8fdbafa | 1569 | /* Handle secondary reloads for SAR. These occur when trying to load |
2a170404 | 1570 | the SAR from memory, FP register, or with a constant. */ |
e8fdbafa | 1571 | else if (GET_CODE (operand0) == REG |
2a170404 | 1572 | && REGNO (operand0) < FIRST_PSEUDO_REGISTER |
e8fdbafa | 1573 | && REGNO_REG_CLASS (REGNO (operand0)) == SHIFT_REGS |
1574 | && (GET_CODE (operand1) == MEM | |
7d43e0f7 | 1575 | || GET_CODE (operand1) == CONST_INT |
e8fdbafa | 1576 | || (GET_CODE (operand1) == REG |
1577 | && FP_REG_CLASS_P (REGNO_REG_CLASS (REGNO (operand1))))) | |
1578 | && scratch_reg) | |
1579 | { | |
7eac600c | 1580 | /* D might not fit in 14 bits either; for such cases load D into |
1581 | scratch reg. */ | |
1582 | if (GET_CODE (operand1) == MEM | |
440c23df | 1583 | && !memory_address_p (Pmode, XEXP (operand1, 0))) |
7eac600c | 1584 | { |
fc4d127d | 1585 | /* We are reloading the address into the scratch register, so we |
1586 | want to make sure the scratch register is a full register. */ | |
9840d99d | 1587 | scratch_reg = force_mode (word_mode, scratch_reg); |
fc4d127d | 1588 | |
9840d99d | 1589 | emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1)); |
ad851752 | 1590 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, |
1591 | 0)), | |
440c23df | 1592 | Pmode, |
ad851752 | 1593 | XEXP (XEXP (operand1, 0), |
1594 | 0), | |
1595 | scratch_reg)); | |
fc4d127d | 1596 | |
1597 | /* Now we are going to load the scratch register from memory, | |
1598 | we want to load it in the same width as the original MEM, | |
1599 | which must be the same as the width of the ultimate destination, | |
1600 | OPERAND0. */ | |
9840d99d | 1601 | scratch_reg = force_mode (GET_MODE (operand0), scratch_reg); |
1602 | ||
fc4d127d | 1603 | emit_move_insn (scratch_reg, gen_rtx_MEM (GET_MODE (operand0), |
ad851752 | 1604 | scratch_reg)); |
7eac600c | 1605 | } |
1606 | else | |
fc4d127d | 1607 | { |
1608 | /* We want to load the scratch register using the same mode as | |
1609 | the ultimate destination. */ | |
9840d99d | 1610 | scratch_reg = force_mode (GET_MODE (operand0), scratch_reg); |
1611 | ||
fc4d127d | 1612 | emit_move_insn (scratch_reg, operand1); |
1613 | } | |
1614 | ||
1615 | /* And emit the insn to set the ultimate destination. We know that | |
1616 | the scratch register has the same mode as the destination at this | |
1617 | point. */ | |
e8fdbafa | 1618 | emit_move_insn (operand0, scratch_reg); |
1619 | return 1; | |
1620 | } | |
d6f01525 | 1621 | /* Handle most common case: storing into a register. */ |
1622 | else if (register_operand (operand0, mode)) | |
87ad11b0 | 1623 | { |
1624 | if (register_operand (operand1, mode) | |
93f6e1c7 | 1625 | || (GET_CODE (operand1) == CONST_INT |
1626 | && cint_ok_for_move (INTVAL (operand1))) | |
891b55b4 | 1627 | || (operand1 == CONST0_RTX (mode)) |
87ad11b0 | 1628 | || (GET_CODE (operand1) == HIGH |
df0651dc | 1629 | && !symbolic_operand (XEXP (operand1, 0), VOIDmode)) |
87ad11b0 | 1630 | /* Only `general_operands' can come here, so MEM is ok. */ |
1631 | || GET_CODE (operand1) == MEM) | |
1632 | { | |
1633 | /* Run this case quickly. */ | |
ad851752 | 1634 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1)); |
87ad11b0 | 1635 | return 1; |
1636 | } | |
1637 | } | |
1638 | else if (GET_CODE (operand0) == MEM) | |
1639 | { | |
85eb4c6e | 1640 | if (mode == DFmode && operand1 == CONST0_RTX (mode) |
1641 | && !(reload_in_progress || reload_completed)) | |
1642 | { | |
1643 | rtx temp = gen_reg_rtx (DFmode); | |
1644 | ||
ad851752 | 1645 | emit_insn (gen_rtx_SET (VOIDmode, temp, operand1)); |
1646 | emit_insn (gen_rtx_SET (VOIDmode, operand0, temp)); | |
85eb4c6e | 1647 | return 1; |
1648 | } | |
891b55b4 | 1649 | if (register_operand (operand1, mode) || operand1 == CONST0_RTX (mode)) |
87ad11b0 | 1650 | { |
1651 | /* Run this case quickly. */ | |
ad851752 | 1652 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1)); |
87ad11b0 | 1653 | return 1; |
1654 | } | |
2ff4bf8d | 1655 | if (! (reload_in_progress || reload_completed)) |
87ad11b0 | 1656 | { |
1657 | operands[0] = validize_mem (operand0); | |
1658 | operands[1] = operand1 = force_reg (mode, operand1); | |
1659 | } | |
1660 | } | |
1661 | ||
37a75d53 | 1662 | /* Simplify the source if we need to. |
1663 | Note we do have to handle function labels here, even though we do | |
1664 | not consider them legitimate constants. Loop optimizations can | |
bea60f66 | 1665 | call the emit_move_xxx with one as a source. */ |
57ed30e5 | 1666 | if ((GET_CODE (operand1) != HIGH && immediate_operand (operand1, mode)) |
37a75d53 | 1667 | || function_label_operand (operand1, mode) |
2ee034bc | 1668 | || (GET_CODE (operand1) == HIGH |
63882853 | 1669 | && symbolic_operand (XEXP (operand1, 0), mode))) |
87ad11b0 | 1670 | { |
2ee034bc | 1671 | int ishighonly = 0; |
1672 | ||
1673 | if (GET_CODE (operand1) == HIGH) | |
1674 | { | |
1675 | ishighonly = 1; | |
1676 | operand1 = XEXP (operand1, 0); | |
1677 | } | |
87ad11b0 | 1678 | if (symbolic_operand (operand1, mode)) |
1679 | { | |
005a7dd0 | 1680 | /* Argh. The assembler and linker can't handle arithmetic |
81653f9b | 1681 | involving plabels. |
005a7dd0 | 1682 | |
81653f9b | 1683 | So we force the plabel into memory, load operand0 from |
1684 | the memory location, then add in the constant part. */ | |
37a75d53 | 1685 | if ((GET_CODE (operand1) == CONST |
1686 | && GET_CODE (XEXP (operand1, 0)) == PLUS | |
1687 | && function_label_operand (XEXP (XEXP (operand1, 0), 0), Pmode)) | |
1688 | || function_label_operand (operand1, mode)) | |
005a7dd0 | 1689 | { |
b3d569a0 | 1690 | rtx temp, const_part; |
81653f9b | 1691 | |
1692 | /* Figure out what (if any) scratch register to use. */ | |
1693 | if (reload_in_progress || reload_completed) | |
f756078b | 1694 | { |
1695 | scratch_reg = scratch_reg ? scratch_reg : operand0; | |
1696 | /* SCRATCH_REG will hold an address and maybe the actual | |
1697 | data. We want it in WORD_MODE regardless of what mode it | |
1698 | was originally given to us. */ | |
9840d99d | 1699 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1700 | } |
81653f9b | 1701 | else if (flag_pic) |
1702 | scratch_reg = gen_reg_rtx (Pmode); | |
1703 | ||
37a75d53 | 1704 | if (GET_CODE (operand1) == CONST) |
1705 | { | |
1706 | /* Save away the constant part of the expression. */ | |
1707 | const_part = XEXP (XEXP (operand1, 0), 1); | |
1708 | if (GET_CODE (const_part) != CONST_INT) | |
1709 | abort (); | |
1710 | ||
1711 | /* Force the function label into memory. */ | |
1712 | temp = force_const_mem (mode, XEXP (XEXP (operand1, 0), 0)); | |
1713 | } | |
1714 | else | |
1715 | { | |
1716 | /* No constant part. */ | |
1717 | const_part = NULL_RTX; | |
005a7dd0 | 1718 | |
37a75d53 | 1719 | /* Force the function label into memory. */ |
1720 | temp = force_const_mem (mode, operand1); | |
1721 | } | |
9840d99d | 1722 | |
81653f9b | 1723 | |
1724 | /* Get the address of the memory location. PIC-ify it if | |
1725 | necessary. */ | |
1726 | temp = XEXP (temp, 0); | |
1727 | if (flag_pic) | |
1728 | temp = legitimize_pic_address (temp, mode, scratch_reg); | |
1729 | ||
1730 | /* Put the address of the memory location into our destination | |
1731 | register. */ | |
1732 | operands[1] = temp; | |
1733 | emit_move_sequence (operands, mode, scratch_reg); | |
1734 | ||
1735 | /* Now load from the memory location into our destination | |
1736 | register. */ | |
ad851752 | 1737 | operands[1] = gen_rtx_MEM (Pmode, operands[0]); |
81653f9b | 1738 | emit_move_sequence (operands, mode, scratch_reg); |
1739 | ||
1740 | /* And add back in the constant part. */ | |
37a75d53 | 1741 | if (const_part != NULL_RTX) |
1742 | expand_inc (operand0, const_part); | |
81653f9b | 1743 | |
1744 | return 1; | |
005a7dd0 | 1745 | } |
1746 | ||
87ad11b0 | 1747 | if (flag_pic) |
1748 | { | |
2ff4bf8d | 1749 | rtx temp; |
1750 | ||
1751 | if (reload_in_progress || reload_completed) | |
f756078b | 1752 | { |
1753 | temp = scratch_reg ? scratch_reg : operand0; | |
1754 | /* TEMP will hold an address and maybe the actual | |
1755 | data. We want it in WORD_MODE regardless of what mode it | |
1756 | was originally given to us. */ | |
9840d99d | 1757 | temp = force_mode (word_mode, temp); |
f756078b | 1758 | } |
2ff4bf8d | 1759 | else |
1760 | temp = gen_reg_rtx (Pmode); | |
6d36483b | 1761 | |
81653f9b | 1762 | /* (const (plus (symbol) (const_int))) must be forced to |
1763 | memory during/after reload if the const_int will not fit | |
1764 | in 14 bits. */ | |
1765 | if (GET_CODE (operand1) == CONST | |
96b86ab6 | 1766 | && GET_CODE (XEXP (operand1, 0)) == PLUS |
1767 | && GET_CODE (XEXP (XEXP (operand1, 0), 1)) == CONST_INT | |
1768 | && !INT_14_BITS (XEXP (XEXP (operand1, 0), 1)) | |
1769 | && (reload_completed || reload_in_progress) | |
1770 | && flag_pic) | |
1771 | { | |
1772 | operands[1] = force_const_mem (mode, operand1); | |
1773 | operands[1] = legitimize_pic_address (XEXP (operands[1], 0), | |
1774 | mode, temp); | |
1775 | emit_move_sequence (operands, mode, temp); | |
1776 | } | |
005a7dd0 | 1777 | else |
1778 | { | |
1779 | operands[1] = legitimize_pic_address (operand1, mode, temp); | |
ad851752 | 1780 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operands[1])); |
005a7dd0 | 1781 | } |
87ad11b0 | 1782 | } |
b4a7bf10 | 1783 | /* On the HPPA, references to data space are supposed to use dp, |
1784 | register 27, but showing it in the RTL inhibits various cse | |
1785 | and loop optimizations. */ | |
6d36483b | 1786 | else |
87ad11b0 | 1787 | { |
005a7dd0 | 1788 | rtx temp, set; |
2ee034bc | 1789 | |
6d36483b | 1790 | if (reload_in_progress || reload_completed) |
f756078b | 1791 | { |
1792 | temp = scratch_reg ? scratch_reg : operand0; | |
1793 | /* TEMP will hold an address and maybe the actual | |
1794 | data. We want it in WORD_MODE regardless of what mode it | |
1795 | was originally given to us. */ | |
9840d99d | 1796 | temp = force_mode (word_mode, temp); |
f756078b | 1797 | } |
2ee034bc | 1798 | else |
1799 | temp = gen_reg_rtx (mode); | |
1800 | ||
42819d4e | 1801 | /* Loading a SYMBOL_REF into a register makes that register |
9840d99d | 1802 | safe to be used as the base in an indexed address. |
42819d4e | 1803 | |
1804 | Don't mark hard registers though. That loses. */ | |
47a61b79 | 1805 | if (GET_CODE (operand0) == REG |
1806 | && REGNO (operand0) >= FIRST_PSEUDO_REGISTER) | |
e61a0a7f | 1807 | REG_POINTER (operand0) = 1; |
42819d4e | 1808 | if (REGNO (temp) >= FIRST_PSEUDO_REGISTER) |
e61a0a7f | 1809 | REG_POINTER (temp) = 1; |
2ee034bc | 1810 | if (ishighonly) |
ad851752 | 1811 | set = gen_rtx_SET (mode, operand0, temp); |
2ee034bc | 1812 | else |
7014838c | 1813 | set = gen_rtx_SET (VOIDmode, |
1814 | operand0, | |
ad851752 | 1815 | gen_rtx_LO_SUM (mode, temp, operand1)); |
6d36483b | 1816 | |
ad851752 | 1817 | emit_insn (gen_rtx_SET (VOIDmode, |
1818 | temp, | |
1819 | gen_rtx_HIGH (mode, operand1))); | |
d2498717 | 1820 | emit_insn (set); |
166bf021 | 1821 | |
87ad11b0 | 1822 | } |
2ee034bc | 1823 | return 1; |
87ad11b0 | 1824 | } |
42faba01 | 1825 | else if (GET_CODE (operand1) != CONST_INT |
8c8ec4de | 1826 | || ! cint_ok_for_move (INTVAL (operand1))) |
87ad11b0 | 1827 | { |
93f6e1c7 | 1828 | rtx extend = NULL_RTX; |
2ff4bf8d | 1829 | rtx temp; |
5e3c5739 | 1830 | |
1831 | if (TARGET_64BIT && GET_CODE (operand1) == CONST_INT | |
b7d86581 | 1832 | && HOST_BITS_PER_WIDE_INT > 32 |
5e3c5739 | 1833 | && GET_MODE_BITSIZE (GET_MODE (operand0)) > 32) |
1834 | { | |
1835 | HOST_WIDE_INT val = INTVAL (operand1); | |
b7d86581 | 1836 | HOST_WIDE_INT nval; |
5e3c5739 | 1837 | |
93f6e1c7 | 1838 | /* Extract the low order 32 bits of the value and sign extend. |
1839 | If the new value is the same as the original value, we can | |
1840 | can use the original value as-is. If the new value is | |
1841 | different, we use it and insert the most-significant 32-bits | |
1842 | of the original value into the final result. */ | |
b7d86581 | 1843 | nval = ((val & (((HOST_WIDE_INT) 2 << 31) - 1)) |
1844 | ^ ((HOST_WIDE_INT) 1 << 31)) - ((HOST_WIDE_INT) 1 << 31); | |
5e3c5739 | 1845 | if (val != nval) |
1846 | { | |
93f6e1c7 | 1847 | #if HOST_BITS_PER_WIDE_INT > 32 |
1848 | extend = GEN_INT (val >> 32); | |
1849 | #endif | |
5e3c5739 | 1850 | operand1 = GEN_INT (nval); |
1851 | } | |
1852 | } | |
2ff4bf8d | 1853 | |
1854 | if (reload_in_progress || reload_completed) | |
1855 | temp = operand0; | |
1856 | else | |
1857 | temp = gen_reg_rtx (mode); | |
1858 | ||
7c4d3047 | 1859 | /* We don't directly split DImode constants on 32-bit targets |
1860 | because PLUS uses an 11-bit immediate and the insn sequence | |
1861 | generated is not as efficient as the one using HIGH/LO_SUM. */ | |
1862 | if (GET_CODE (operand1) == CONST_INT | |
1863 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
93f6e1c7 | 1864 | { |
7c4d3047 | 1865 | /* Directly break constant into high and low parts. This |
93f6e1c7 | 1866 | provides better optimization opportunities because various |
1867 | passes recognize constants split with PLUS but not LO_SUM. | |
1868 | We use a 14-bit signed low part except when the addition | |
1869 | of 0x4000 to the high part might change the sign of the | |
1870 | high part. */ | |
1871 | HOST_WIDE_INT value = INTVAL (operand1); | |
1872 | HOST_WIDE_INT low = value & 0x3fff; | |
1873 | HOST_WIDE_INT high = value & ~ 0x3fff; | |
1874 | ||
1875 | if (low >= 0x2000) | |
1876 | { | |
1877 | if (high == 0x7fffc000 || (mode == HImode && high == 0x4000)) | |
1878 | high += 0x2000; | |
1879 | else | |
1880 | high += 0x4000; | |
1881 | } | |
1882 | ||
1883 | low = value - high; | |
5e3c5739 | 1884 | |
93f6e1c7 | 1885 | emit_insn (gen_rtx_SET (VOIDmode, temp, GEN_INT (high))); |
1886 | operands[1] = gen_rtx_PLUS (mode, temp, GEN_INT (low)); | |
1887 | } | |
1888 | else | |
5e3c5739 | 1889 | { |
93f6e1c7 | 1890 | emit_insn (gen_rtx_SET (VOIDmode, temp, |
1891 | gen_rtx_HIGH (mode, operand1))); | |
1892 | operands[1] = gen_rtx_LO_SUM (mode, temp, operand1); | |
5e3c5739 | 1893 | } |
9840d99d | 1894 | |
93f6e1c7 | 1895 | emit_move_insn (operands[0], operands[1]); |
1896 | ||
1897 | if (extend != NULL_RTX) | |
1898 | emit_insn (gen_insv (operands[0], GEN_INT (32), const0_rtx, | |
1899 | extend)); | |
1900 | ||
5e3c5739 | 1901 | return 1; |
87ad11b0 | 1902 | } |
1903 | } | |
1904 | /* Now have insn-emit do whatever it normally does. */ | |
1905 | return 0; | |
1906 | } | |
1907 | ||
1946138e | 1908 | /* Examine EXP and return nonzero if it contains an ADDR_EXPR (meaning |
bd49d362 | 1909 | it will need a link/runtime reloc). */ |
1946138e | 1910 | |
1911 | int | |
5c1d8983 | 1912 | reloc_needed (tree exp) |
1946138e | 1913 | { |
1914 | int reloc = 0; | |
1915 | ||
1916 | switch (TREE_CODE (exp)) | |
1917 | { | |
1918 | case ADDR_EXPR: | |
1919 | return 1; | |
1920 | ||
1921 | case PLUS_EXPR: | |
1922 | case MINUS_EXPR: | |
1923 | reloc = reloc_needed (TREE_OPERAND (exp, 0)); | |
1924 | reloc |= reloc_needed (TREE_OPERAND (exp, 1)); | |
1925 | break; | |
1926 | ||
1927 | case NOP_EXPR: | |
1928 | case CONVERT_EXPR: | |
1929 | case NON_LVALUE_EXPR: | |
1930 | reloc = reloc_needed (TREE_OPERAND (exp, 0)); | |
1931 | break; | |
1932 | ||
1933 | case CONSTRUCTOR: | |
1934 | { | |
1935 | register tree link; | |
1936 | for (link = CONSTRUCTOR_ELTS (exp); link; link = TREE_CHAIN (link)) | |
1937 | if (TREE_VALUE (link) != 0) | |
1938 | reloc |= reloc_needed (TREE_VALUE (link)); | |
1939 | } | |
1940 | break; | |
1941 | ||
1942 | case ERROR_MARK: | |
1943 | break; | |
7d27e4c9 | 1944 | |
1945 | default: | |
1946 | break; | |
1946138e | 1947 | } |
1948 | return reloc; | |
1949 | } | |
1950 | ||
7811991d | 1951 | /* Does operand (which is a symbolic_operand) live in text space? |
1952 | If so, SYMBOL_REF_FLAG, which is set by pa_encode_section_info, | |
1953 | will be true. */ | |
87ad11b0 | 1954 | |
1955 | int | |
5c1d8983 | 1956 | read_only_operand (rtx operand, enum machine_mode mode ATTRIBUTE_UNUSED) |
87ad11b0 | 1957 | { |
1958 | if (GET_CODE (operand) == CONST) | |
1959 | operand = XEXP (XEXP (operand, 0), 0); | |
b4a7bf10 | 1960 | if (flag_pic) |
1961 | { | |
1962 | if (GET_CODE (operand) == SYMBOL_REF) | |
1963 | return SYMBOL_REF_FLAG (operand) && !CONSTANT_POOL_ADDRESS_P (operand); | |
1964 | } | |
1965 | else | |
1966 | { | |
1967 | if (GET_CODE (operand) == SYMBOL_REF) | |
1968 | return SYMBOL_REF_FLAG (operand) || CONSTANT_POOL_ADDRESS_P (operand); | |
1969 | } | |
87ad11b0 | 1970 | return 1; |
1971 | } | |
6d36483b | 1972 | |
87ad11b0 | 1973 | \f |
1974 | /* Return the best assembler insn template | |
f54b1341 | 1975 | for moving operands[1] into operands[0] as a fullword. */ |
611a88e1 | 1976 | const char * |
5c1d8983 | 1977 | singlemove_string (rtx *operands) |
87ad11b0 | 1978 | { |
3745c59b | 1979 | HOST_WIDE_INT intval; |
1980 | ||
87ad11b0 | 1981 | if (GET_CODE (operands[0]) == MEM) |
1982 | return "stw %r1,%0"; | |
3745c59b | 1983 | if (GET_CODE (operands[1]) == MEM) |
87ad11b0 | 1984 | return "ldw %1,%0"; |
3745c59b | 1985 | if (GET_CODE (operands[1]) == CONST_DOUBLE) |
9d5108ea | 1986 | { |
3745c59b | 1987 | long i; |
1988 | REAL_VALUE_TYPE d; | |
9d5108ea | 1989 | |
3745c59b | 1990 | if (GET_MODE (operands[1]) != SFmode) |
1991 | abort (); | |
9d5108ea | 1992 | |
3745c59b | 1993 | /* Translate the CONST_DOUBLE to a CONST_INT with the same target |
1994 | bit pattern. */ | |
1995 | REAL_VALUE_FROM_CONST_DOUBLE (d, operands[1]); | |
1996 | REAL_VALUE_TO_TARGET_SINGLE (d, i); | |
9d5108ea | 1997 | |
3745c59b | 1998 | operands[1] = GEN_INT (i); |
1999 | /* Fall through to CONST_INT case. */ | |
2000 | } | |
2001 | if (GET_CODE (operands[1]) == CONST_INT) | |
9d5108ea | 2002 | { |
3745c59b | 2003 | intval = INTVAL (operands[1]); |
2004 | ||
2005 | if (VAL_14_BITS_P (intval)) | |
2006 | return "ldi %1,%0"; | |
2007 | else if ((intval & 0x7ff) == 0) | |
2008 | return "ldil L'%1,%0"; | |
2009 | else if (zdepi_cint_p (intval)) | |
e4065f95 | 2010 | return "{zdepi %Z1,%0|depwi,z %Z1,%0}"; |
9d5108ea | 2011 | else |
2012 | return "ldil L'%1,%0\n\tldo R'%1(%0),%0"; | |
2013 | } | |
87ad11b0 | 2014 | return "copy %1,%0"; |
2015 | } | |
2016 | \f | |
2017 | ||
201f01e9 | 2018 | /* Compute position (in OP[1]) and width (in OP[2]) |
2019 | useful for copying IMM to a register using the zdepi | |
2020 | instructions. Store the immediate value to insert in OP[0]. */ | |
611a88e1 | 2021 | static void |
5c1d8983 | 2022 | compute_zdepwi_operands (unsigned HOST_WIDE_INT imm, unsigned *op) |
7e10ba53 | 2023 | { |
e057641f | 2024 | int lsb, len; |
7e10ba53 | 2025 | |
e057641f | 2026 | /* Find the least significant set bit in IMM. */ |
2027 | for (lsb = 0; lsb < 32; lsb++) | |
7e10ba53 | 2028 | { |
e057641f | 2029 | if ((imm & 1) != 0) |
7e10ba53 | 2030 | break; |
e057641f | 2031 | imm >>= 1; |
7e10ba53 | 2032 | } |
2033 | ||
e057641f | 2034 | /* Choose variants based on *sign* of the 5-bit field. */ |
2035 | if ((imm & 0x10) == 0) | |
2036 | len = (lsb <= 28) ? 4 : 32 - lsb; | |
7e10ba53 | 2037 | else |
2038 | { | |
e057641f | 2039 | /* Find the width of the bitstring in IMM. */ |
2040 | for (len = 5; len < 32; len++) | |
7e10ba53 | 2041 | { |
e057641f | 2042 | if ((imm & (1 << len)) == 0) |
7e10ba53 | 2043 | break; |
7e10ba53 | 2044 | } |
2045 | ||
e057641f | 2046 | /* Sign extend IMM as a 5-bit value. */ |
2047 | imm = (imm & 0xf) - 0x10; | |
7e10ba53 | 2048 | } |
2049 | ||
42faba01 | 2050 | op[0] = imm; |
2051 | op[1] = 31 - lsb; | |
2052 | op[2] = len; | |
7e10ba53 | 2053 | } |
2054 | ||
5e3c5739 | 2055 | /* Compute position (in OP[1]) and width (in OP[2]) |
2056 | useful for copying IMM to a register using the depdi,z | |
2057 | instructions. Store the immediate value to insert in OP[0]. */ | |
2058 | void | |
5c1d8983 | 2059 | compute_zdepdi_operands (unsigned HOST_WIDE_INT imm, unsigned *op) |
5e3c5739 | 2060 | { |
2061 | HOST_WIDE_INT lsb, len; | |
2062 | ||
2063 | /* Find the least significant set bit in IMM. */ | |
2064 | for (lsb = 0; lsb < HOST_BITS_PER_WIDE_INT; lsb++) | |
2065 | { | |
2066 | if ((imm & 1) != 0) | |
2067 | break; | |
2068 | imm >>= 1; | |
2069 | } | |
2070 | ||
2071 | /* Choose variants based on *sign* of the 5-bit field. */ | |
2072 | if ((imm & 0x10) == 0) | |
2073 | len = ((lsb <= HOST_BITS_PER_WIDE_INT - 4) | |
2074 | ? 4 : HOST_BITS_PER_WIDE_INT - lsb); | |
2075 | else | |
2076 | { | |
2077 | /* Find the width of the bitstring in IMM. */ | |
2078 | for (len = 5; len < HOST_BITS_PER_WIDE_INT; len++) | |
2079 | { | |
ea52c577 | 2080 | if ((imm & ((unsigned HOST_WIDE_INT) 1 << len)) == 0) |
5e3c5739 | 2081 | break; |
2082 | } | |
2083 | ||
2084 | /* Sign extend IMM as a 5-bit value. */ | |
2085 | imm = (imm & 0xf) - 0x10; | |
2086 | } | |
2087 | ||
2088 | op[0] = imm; | |
2089 | op[1] = 63 - lsb; | |
2090 | op[2] = len; | |
2091 | } | |
2092 | ||
87ad11b0 | 2093 | /* Output assembler code to perform a doubleword move insn |
2094 | with operands OPERANDS. */ | |
2095 | ||
611a88e1 | 2096 | const char * |
5c1d8983 | 2097 | output_move_double (rtx *operands) |
87ad11b0 | 2098 | { |
2099 | enum { REGOP, OFFSOP, MEMOP, CNSTOP, RNDOP } optype0, optype1; | |
2100 | rtx latehalf[2]; | |
2101 | rtx addreg0 = 0, addreg1 = 0; | |
2102 | ||
2103 | /* First classify both operands. */ | |
2104 | ||
2105 | if (REG_P (operands[0])) | |
2106 | optype0 = REGOP; | |
2107 | else if (offsettable_memref_p (operands[0])) | |
2108 | optype0 = OFFSOP; | |
2109 | else if (GET_CODE (operands[0]) == MEM) | |
2110 | optype0 = MEMOP; | |
2111 | else | |
2112 | optype0 = RNDOP; | |
2113 | ||
2114 | if (REG_P (operands[1])) | |
2115 | optype1 = REGOP; | |
2116 | else if (CONSTANT_P (operands[1])) | |
2117 | optype1 = CNSTOP; | |
2118 | else if (offsettable_memref_p (operands[1])) | |
2119 | optype1 = OFFSOP; | |
2120 | else if (GET_CODE (operands[1]) == MEM) | |
2121 | optype1 = MEMOP; | |
2122 | else | |
2123 | optype1 = RNDOP; | |
2124 | ||
2125 | /* Check for the cases that the operand constraints are not | |
2126 | supposed to allow to happen. Abort if we get one, | |
2127 | because generating code for these cases is painful. */ | |
2128 | ||
2129 | if (optype0 != REGOP && optype1 != REGOP) | |
2130 | abort (); | |
2131 | ||
2132 | /* Handle auto decrementing and incrementing loads and stores | |
2133 | specifically, since the structure of the function doesn't work | |
2134 | for them without major modification. Do it better when we learn | |
2135 | this port about the general inc/dec addressing of PA. | |
2136 | (This was written by tege. Chide him if it doesn't work.) */ | |
2137 | ||
2138 | if (optype0 == MEMOP) | |
2139 | { | |
1df0058a | 2140 | /* We have to output the address syntax ourselves, since print_operand |
2141 | doesn't deal with the addresses we want to use. Fix this later. */ | |
2142 | ||
87ad11b0 | 2143 | rtx addr = XEXP (operands[0], 0); |
1df0058a | 2144 | if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC) |
87ad11b0 | 2145 | { |
ad851752 | 2146 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0); |
1df0058a | 2147 | |
2148 | operands[0] = XEXP (addr, 0); | |
2149 | if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG) | |
2150 | abort (); | |
2151 | ||
2152 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2153 | { | |
2154 | /* No overlap between high target register and address | |
2155 | register. (We do this in a non-obvious way to | |
2156 | save a register file writeback) */ | |
2157 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2158 | return "{stws|stw},ma %1,8(%0)\n\tstw %R1,-4(%0)"; |
2159 | return "{stws|stw},ma %1,-8(%0)\n\tstw %R1,12(%0)"; | |
1df0058a | 2160 | } |
2161 | else | |
ea52c577 | 2162 | abort (); |
a3217f65 | 2163 | } |
1df0058a | 2164 | else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC) |
a3217f65 | 2165 | { |
ad851752 | 2166 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0); |
1df0058a | 2167 | |
2168 | operands[0] = XEXP (addr, 0); | |
2169 | if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG) | |
2170 | abort (); | |
2171 | ||
2172 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2173 | { | |
2174 | /* No overlap between high target register and address | |
2175 | register. (We do this in a non-obvious way to | |
2176 | save a register file writeback) */ | |
2177 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2178 | return "{stws|stw},mb %1,8(%0)\n\tstw %R1,4(%0)"; |
2179 | return "{stws|stw},mb %1,-8(%0)\n\tstw %R1,4(%0)"; | |
1df0058a | 2180 | } |
2181 | else | |
ea52c577 | 2182 | abort (); |
87ad11b0 | 2183 | } |
2184 | } | |
2185 | if (optype1 == MEMOP) | |
2186 | { | |
2187 | /* We have to output the address syntax ourselves, since print_operand | |
2188 | doesn't deal with the addresses we want to use. Fix this later. */ | |
2189 | ||
2190 | rtx addr = XEXP (operands[1], 0); | |
2191 | if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC) | |
2192 | { | |
ad851752 | 2193 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
87ad11b0 | 2194 | |
2195 | operands[1] = XEXP (addr, 0); | |
2196 | if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG) | |
2197 | abort (); | |
2198 | ||
2199 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2200 | { | |
2201 | /* No overlap between high target register and address | |
3857fa62 | 2202 | register. (We do this in a non-obvious way to |
87ad11b0 | 2203 | save a register file writeback) */ |
2204 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2205 | return "{ldws|ldw},ma 8(%1),%0\n\tldw -4(%1),%R0"; |
01fd4b49 | 2206 | return "{ldws|ldw},ma -8(%1),%0\n\tldw 12(%1),%R0"; |
87ad11b0 | 2207 | } |
2208 | else | |
2209 | { | |
2210 | /* This is an undefined situation. We should load into the | |
2211 | address register *and* update that register. Probably | |
2212 | we don't need to handle this at all. */ | |
2213 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2214 | return "ldw 4(%1),%R0\n\t{ldws|ldw},ma 8(%1),%0"; |
2215 | return "ldw 4(%1),%R0\n\t{ldws|ldw},ma -8(%1),%0"; | |
87ad11b0 | 2216 | } |
2217 | } | |
2218 | else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC) | |
2219 | { | |
ad851752 | 2220 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
87ad11b0 | 2221 | |
2222 | operands[1] = XEXP (addr, 0); | |
2223 | if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG) | |
2224 | abort (); | |
2225 | ||
2226 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2227 | { | |
2228 | /* No overlap between high target register and address | |
3857fa62 | 2229 | register. (We do this in a non-obvious way to |
87ad11b0 | 2230 | save a register file writeback) */ |
2231 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2232 | return "{ldws|ldw},mb 8(%1),%0\n\tldw 4(%1),%R0"; |
2233 | return "{ldws|ldw},mb -8(%1),%0\n\tldw 4(%1),%R0"; | |
87ad11b0 | 2234 | } |
2235 | else | |
2236 | { | |
2237 | /* This is an undefined situation. We should load into the | |
2238 | address register *and* update that register. Probably | |
2239 | we don't need to handle this at all. */ | |
2240 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2241 | return "ldw 12(%1),%R0\n\t{ldws|ldw},mb 8(%1),%0"; |
2242 | return "ldw -4(%1),%R0\n\t{ldws|ldw},mb -8(%1),%0"; | |
87ad11b0 | 2243 | } |
2244 | } | |
12b02046 | 2245 | else if (GET_CODE (addr) == PLUS |
2246 | && GET_CODE (XEXP (addr, 0)) == MULT) | |
2247 | { | |
ad851752 | 2248 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
12b02046 | 2249 | |
2250 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2251 | { | |
2252 | rtx xoperands[3]; | |
2253 | ||
2254 | xoperands[0] = high_reg; | |
2255 | xoperands[1] = XEXP (addr, 1); | |
2256 | xoperands[2] = XEXP (XEXP (addr, 0), 0); | |
2257 | xoperands[3] = XEXP (XEXP (addr, 0), 1); | |
e4065f95 | 2258 | output_asm_insn ("{sh%O3addl %2,%1,%0|shladd,l %2,%O3,%1,%0}", |
2259 | xoperands); | |
34940871 | 2260 | return "ldw 4(%0),%R0\n\tldw 0(%0),%0"; |
12b02046 | 2261 | } |
2262 | else | |
2263 | { | |
2264 | rtx xoperands[3]; | |
2265 | ||
2266 | xoperands[0] = high_reg; | |
2267 | xoperands[1] = XEXP (addr, 1); | |
2268 | xoperands[2] = XEXP (XEXP (addr, 0), 0); | |
2269 | xoperands[3] = XEXP (XEXP (addr, 0), 1); | |
e4065f95 | 2270 | output_asm_insn ("{sh%O3addl %2,%1,%R0|shladd,l %2,%O3,%1,%R0}", |
2271 | xoperands); | |
34940871 | 2272 | return "ldw 0(%R0),%0\n\tldw 4(%R0),%R0"; |
12b02046 | 2273 | } |
12b02046 | 2274 | } |
87ad11b0 | 2275 | } |
2276 | ||
2277 | /* If an operand is an unoffsettable memory ref, find a register | |
2278 | we can increment temporarily to make it refer to the second word. */ | |
2279 | ||
2280 | if (optype0 == MEMOP) | |
2281 | addreg0 = find_addr_reg (XEXP (operands[0], 0)); | |
2282 | ||
2283 | if (optype1 == MEMOP) | |
2284 | addreg1 = find_addr_reg (XEXP (operands[1], 0)); | |
2285 | ||
2286 | /* Ok, we can do one word at a time. | |
2287 | Normally we do the low-numbered word first. | |
2288 | ||
2289 | In either case, set up in LATEHALF the operands to use | |
2290 | for the high-numbered word and in some cases alter the | |
2291 | operands in OPERANDS to be suitable for the low-numbered word. */ | |
2292 | ||
2293 | if (optype0 == REGOP) | |
ad851752 | 2294 | latehalf[0] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); |
87ad11b0 | 2295 | else if (optype0 == OFFSOP) |
eafc6604 | 2296 | latehalf[0] = adjust_address (operands[0], SImode, 4); |
87ad11b0 | 2297 | else |
2298 | latehalf[0] = operands[0]; | |
2299 | ||
2300 | if (optype1 == REGOP) | |
ad851752 | 2301 | latehalf[1] = gen_rtx_REG (SImode, REGNO (operands[1]) + 1); |
87ad11b0 | 2302 | else if (optype1 == OFFSOP) |
eafc6604 | 2303 | latehalf[1] = adjust_address (operands[1], SImode, 4); |
87ad11b0 | 2304 | else if (optype1 == CNSTOP) |
2305 | split_double (operands[1], &operands[1], &latehalf[1]); | |
2306 | else | |
2307 | latehalf[1] = operands[1]; | |
2308 | ||
2309 | /* If the first move would clobber the source of the second one, | |
2310 | do them in the other order. | |
2311 | ||
cf489d53 | 2312 | This can happen in two cases: |
87ad11b0 | 2313 | |
cf489d53 | 2314 | mem -> register where the first half of the destination register |
2315 | is the same register used in the memory's address. Reload | |
2316 | can create such insns. | |
87ad11b0 | 2317 | |
cf489d53 | 2318 | mem in this case will be either register indirect or register |
9840d99d | 2319 | indirect plus a valid offset. |
cf489d53 | 2320 | |
2321 | register -> register move where REGNO(dst) == REGNO(src + 1) | |
9840d99d | 2322 | someone (Tim/Tege?) claimed this can happen for parameter loads. |
cf489d53 | 2323 | |
2324 | Handle mem -> register case first. */ | |
2325 | if (optype0 == REGOP | |
2326 | && (optype1 == MEMOP || optype1 == OFFSOP) | |
2327 | && refers_to_regno_p (REGNO (operands[0]), REGNO (operands[0]) + 1, | |
2328 | operands[1], 0)) | |
87ad11b0 | 2329 | { |
87ad11b0 | 2330 | /* Do the late half first. */ |
2331 | if (addreg1) | |
6a5d085a | 2332 | output_asm_insn ("ldo 4(%0),%0", &addreg1); |
87ad11b0 | 2333 | output_asm_insn (singlemove_string (latehalf), latehalf); |
cf489d53 | 2334 | |
2335 | /* Then clobber. */ | |
87ad11b0 | 2336 | if (addreg1) |
6a5d085a | 2337 | output_asm_insn ("ldo -4(%0),%0", &addreg1); |
87ad11b0 | 2338 | return singlemove_string (operands); |
2339 | } | |
2340 | ||
cf489d53 | 2341 | /* Now handle register -> register case. */ |
c4fa5937 | 2342 | if (optype0 == REGOP && optype1 == REGOP |
2343 | && REGNO (operands[0]) == REGNO (operands[1]) + 1) | |
2344 | { | |
2345 | output_asm_insn (singlemove_string (latehalf), latehalf); | |
2346 | return singlemove_string (operands); | |
2347 | } | |
2348 | ||
87ad11b0 | 2349 | /* Normal case: do the two words, low-numbered first. */ |
2350 | ||
2351 | output_asm_insn (singlemove_string (operands), operands); | |
2352 | ||
2353 | /* Make any unoffsettable addresses point at high-numbered word. */ | |
2354 | if (addreg0) | |
6a5d085a | 2355 | output_asm_insn ("ldo 4(%0),%0", &addreg0); |
87ad11b0 | 2356 | if (addreg1) |
6a5d085a | 2357 | output_asm_insn ("ldo 4(%0),%0", &addreg1); |
87ad11b0 | 2358 | |
2359 | /* Do that word. */ | |
2360 | output_asm_insn (singlemove_string (latehalf), latehalf); | |
2361 | ||
2362 | /* Undo the adds we just did. */ | |
2363 | if (addreg0) | |
6a5d085a | 2364 | output_asm_insn ("ldo -4(%0),%0", &addreg0); |
87ad11b0 | 2365 | if (addreg1) |
6a5d085a | 2366 | output_asm_insn ("ldo -4(%0),%0", &addreg1); |
87ad11b0 | 2367 | |
2368 | return ""; | |
2369 | } | |
2370 | \f | |
611a88e1 | 2371 | const char * |
5c1d8983 | 2372 | output_fp_move_double (rtx *operands) |
87ad11b0 | 2373 | { |
2374 | if (FP_REG_P (operands[0])) | |
2375 | { | |
6d36483b | 2376 | if (FP_REG_P (operands[1]) |
891b55b4 | 2377 | || operands[1] == CONST0_RTX (GET_MODE (operands[0]))) |
c6ae275c | 2378 | output_asm_insn ("fcpy,dbl %f1,%0", operands); |
6d36483b | 2379 | else |
27ef382d | 2380 | output_asm_insn ("fldd%F1 %1,%0", operands); |
87ad11b0 | 2381 | } |
2382 | else if (FP_REG_P (operands[1])) | |
2383 | { | |
27ef382d | 2384 | output_asm_insn ("fstd%F0 %1,%0", operands); |
87ad11b0 | 2385 | } |
891b55b4 | 2386 | else if (operands[1] == CONST0_RTX (GET_MODE (operands[0]))) |
2387 | { | |
2388 | if (GET_CODE (operands[0]) == REG) | |
2389 | { | |
2390 | rtx xoperands[2]; | |
ad851752 | 2391 | xoperands[1] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); |
891b55b4 | 2392 | xoperands[0] = operands[0]; |
2393 | output_asm_insn ("copy %%r0,%0\n\tcopy %%r0,%1", xoperands); | |
2394 | } | |
6d36483b | 2395 | /* This is a pain. You have to be prepared to deal with an |
01cc3b75 | 2396 | arbitrary address here including pre/post increment/decrement. |
891b55b4 | 2397 | |
2398 | so avoid this in the MD. */ | |
2399 | else | |
2400 | abort (); | |
2401 | } | |
87ad11b0 | 2402 | else abort (); |
2403 | return ""; | |
2404 | } | |
2405 | \f | |
2406 | /* Return a REG that occurs in ADDR with coefficient 1. | |
2407 | ADDR can be effectively incremented by incrementing REG. */ | |
2408 | ||
2409 | static rtx | |
5c1d8983 | 2410 | find_addr_reg (rtx addr) |
87ad11b0 | 2411 | { |
2412 | while (GET_CODE (addr) == PLUS) | |
2413 | { | |
2414 | if (GET_CODE (XEXP (addr, 0)) == REG) | |
2415 | addr = XEXP (addr, 0); | |
2416 | else if (GET_CODE (XEXP (addr, 1)) == REG) | |
2417 | addr = XEXP (addr, 1); | |
2418 | else if (CONSTANT_P (XEXP (addr, 0))) | |
2419 | addr = XEXP (addr, 1); | |
2420 | else if (CONSTANT_P (XEXP (addr, 1))) | |
2421 | addr = XEXP (addr, 0); | |
2422 | else | |
2423 | abort (); | |
2424 | } | |
2425 | if (GET_CODE (addr) == REG) | |
2426 | return addr; | |
2427 | abort (); | |
2428 | } | |
2429 | ||
87ad11b0 | 2430 | /* Emit code to perform a block move. |
2431 | ||
87ad11b0 | 2432 | OPERANDS[0] is the destination pointer as a REG, clobbered. |
2433 | OPERANDS[1] is the source pointer as a REG, clobbered. | |
42819d4e | 2434 | OPERANDS[2] is a register for temporary storage. |
87ad11b0 | 2435 | OPERANDS[3] is a register for temporary storage. |
a7e1bb24 | 2436 | OPERANDS[4] is the size as a CONST_INT |
9840d99d | 2437 | OPERANDS[5] is the alignment safe to use, as a CONST_INT. |
ad87de1e | 2438 | OPERANDS[6] is another temporary register. */ |
87ad11b0 | 2439 | |
611a88e1 | 2440 | const char * |
5c1d8983 | 2441 | output_block_move (rtx *operands, int size_is_constant ATTRIBUTE_UNUSED) |
87ad11b0 | 2442 | { |
2443 | int align = INTVAL (operands[5]); | |
42819d4e | 2444 | unsigned long n_bytes = INTVAL (operands[4]); |
87ad11b0 | 2445 | |
a7e1bb24 | 2446 | /* We can't move more than a word at a time because the PA |
87ad11b0 | 2447 | has no longer integer move insns. (Could use fp mem ops?) */ |
a7e1bb24 | 2448 | if (align > (TARGET_64BIT ? 8 : 4)) |
2449 | align = (TARGET_64BIT ? 8 : 4); | |
87ad11b0 | 2450 | |
42819d4e | 2451 | /* Note that we know each loop below will execute at least twice |
2452 | (else we would have open-coded the copy). */ | |
2453 | switch (align) | |
87ad11b0 | 2454 | { |
a7e1bb24 | 2455 | case 8: |
2456 | /* Pre-adjust the loop counter. */ | |
2457 | operands[4] = GEN_INT (n_bytes - 16); | |
2458 | output_asm_insn ("ldi %4,%2", operands); | |
2459 | ||
2460 | /* Copying loop. */ | |
2461 | output_asm_insn ("ldd,ma 8(%1),%3", operands); | |
2462 | output_asm_insn ("ldd,ma 8(%1),%6", operands); | |
2463 | output_asm_insn ("std,ma %3,8(%0)", operands); | |
2464 | output_asm_insn ("addib,>= -16,%2,.-12", operands); | |
2465 | output_asm_insn ("std,ma %6,8(%0)", operands); | |
2466 | ||
2467 | /* Handle the residual. There could be up to 7 bytes of | |
2468 | residual to copy! */ | |
2469 | if (n_bytes % 16 != 0) | |
2470 | { | |
2471 | operands[4] = GEN_INT (n_bytes % 8); | |
2472 | if (n_bytes % 16 >= 8) | |
2473 | output_asm_insn ("ldd,ma 8(%1),%3", operands); | |
2474 | if (n_bytes % 8 != 0) | |
2475 | output_asm_insn ("ldd 0(%1),%6", operands); | |
2476 | if (n_bytes % 16 >= 8) | |
2477 | output_asm_insn ("std,ma %3,8(%0)", operands); | |
2478 | if (n_bytes % 8 != 0) | |
2479 | output_asm_insn ("stdby,e %6,%4(%0)", operands); | |
2480 | } | |
2481 | return ""; | |
2482 | ||
42819d4e | 2483 | case 4: |
2484 | /* Pre-adjust the loop counter. */ | |
2485 | operands[4] = GEN_INT (n_bytes - 8); | |
2486 | output_asm_insn ("ldi %4,%2", operands); | |
2487 | ||
2488 | /* Copying loop. */ | |
e4065f95 | 2489 | output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands); |
2490 | output_asm_insn ("{ldws|ldw},ma 4(%1),%6", operands); | |
2491 | output_asm_insn ("{stws|stw},ma %3,4(%0)", operands); | |
42819d4e | 2492 | output_asm_insn ("addib,>= -8,%2,.-12", operands); |
e4065f95 | 2493 | output_asm_insn ("{stws|stw},ma %6,4(%0)", operands); |
42819d4e | 2494 | |
2495 | /* Handle the residual. There could be up to 7 bytes of | |
2496 | residual to copy! */ | |
2497 | if (n_bytes % 8 != 0) | |
2498 | { | |
2499 | operands[4] = GEN_INT (n_bytes % 4); | |
2500 | if (n_bytes % 8 >= 4) | |
e4065f95 | 2501 | output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands); |
42819d4e | 2502 | if (n_bytes % 4 != 0) |
34940871 | 2503 | output_asm_insn ("ldw 0(%1),%6", operands); |
42819d4e | 2504 | if (n_bytes % 8 >= 4) |
e4065f95 | 2505 | output_asm_insn ("{stws|stw},ma %3,4(%0)", operands); |
42819d4e | 2506 | if (n_bytes % 4 != 0) |
e4065f95 | 2507 | output_asm_insn ("{stbys|stby},e %6,%4(%0)", operands); |
42819d4e | 2508 | } |
2509 | return ""; | |
87ad11b0 | 2510 | |
42819d4e | 2511 | case 2: |
2512 | /* Pre-adjust the loop counter. */ | |
2513 | operands[4] = GEN_INT (n_bytes - 4); | |
2514 | output_asm_insn ("ldi %4,%2", operands); | |
87ad11b0 | 2515 | |
42819d4e | 2516 | /* Copying loop. */ |
e4065f95 | 2517 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands); |
2518 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%6", operands); | |
2519 | output_asm_insn ("{sths|sth},ma %3,2(%0)", operands); | |
42819d4e | 2520 | output_asm_insn ("addib,>= -4,%2,.-12", operands); |
e4065f95 | 2521 | output_asm_insn ("{sths|sth},ma %6,2(%0)", operands); |
87ad11b0 | 2522 | |
42819d4e | 2523 | /* Handle the residual. */ |
2524 | if (n_bytes % 4 != 0) | |
2525 | { | |
2526 | if (n_bytes % 4 >= 2) | |
e4065f95 | 2527 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands); |
42819d4e | 2528 | if (n_bytes % 2 != 0) |
34940871 | 2529 | output_asm_insn ("ldb 0(%1),%6", operands); |
42819d4e | 2530 | if (n_bytes % 4 >= 2) |
e4065f95 | 2531 | output_asm_insn ("{sths|sth},ma %3,2(%0)", operands); |
42819d4e | 2532 | if (n_bytes % 2 != 0) |
34940871 | 2533 | output_asm_insn ("stb %6,0(%0)", operands); |
42819d4e | 2534 | } |
2535 | return ""; | |
87ad11b0 | 2536 | |
42819d4e | 2537 | case 1: |
2538 | /* Pre-adjust the loop counter. */ | |
2539 | operands[4] = GEN_INT (n_bytes - 2); | |
2540 | output_asm_insn ("ldi %4,%2", operands); | |
87ad11b0 | 2541 | |
42819d4e | 2542 | /* Copying loop. */ |
e4065f95 | 2543 | output_asm_insn ("{ldbs|ldb},ma 1(%1),%3", operands); |
2544 | output_asm_insn ("{ldbs|ldb},ma 1(%1),%6", operands); | |
2545 | output_asm_insn ("{stbs|stb},ma %3,1(%0)", operands); | |
42819d4e | 2546 | output_asm_insn ("addib,>= -2,%2,.-12", operands); |
e4065f95 | 2547 | output_asm_insn ("{stbs|stb},ma %6,1(%0)", operands); |
87ad11b0 | 2548 | |
42819d4e | 2549 | /* Handle the residual. */ |
2550 | if (n_bytes % 2 != 0) | |
2551 | { | |
34940871 | 2552 | output_asm_insn ("ldb 0(%1),%3", operands); |
2553 | output_asm_insn ("stb %3,0(%0)", operands); | |
42819d4e | 2554 | } |
2555 | return ""; | |
87ad11b0 | 2556 | |
42819d4e | 2557 | default: |
2558 | abort (); | |
87ad11b0 | 2559 | } |
87ad11b0 | 2560 | } |
58e17b0b | 2561 | |
2562 | /* Count the number of insns necessary to handle this block move. | |
2563 | ||
2564 | Basic structure is the same as emit_block_move, except that we | |
2565 | count insns rather than emit them. */ | |
2566 | ||
611a88e1 | 2567 | static int |
a7e1bb24 | 2568 | compute_movstr_length (rtx insn) |
58e17b0b | 2569 | { |
2570 | rtx pat = PATTERN (insn); | |
fc1fb057 | 2571 | unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 7), 0)); |
2572 | unsigned long n_bytes = INTVAL (XEXP (XVECEXP (pat, 0, 6), 0)); | |
42819d4e | 2573 | unsigned int n_insns = 0; |
58e17b0b | 2574 | |
2575 | /* We can't move more than four bytes at a time because the PA | |
2576 | has no longer integer move insns. (Could use fp mem ops?) */ | |
a7e1bb24 | 2577 | if (align > (TARGET_64BIT ? 8 : 4)) |
2578 | align = (TARGET_64BIT ? 8 : 4); | |
58e17b0b | 2579 | |
79bfe6ae | 2580 | /* The basic copying loop. */ |
42819d4e | 2581 | n_insns = 6; |
58e17b0b | 2582 | |
42819d4e | 2583 | /* Residuals. */ |
2584 | if (n_bytes % (2 * align) != 0) | |
58e17b0b | 2585 | { |
79bfe6ae | 2586 | if ((n_bytes % (2 * align)) >= align) |
2587 | n_insns += 2; | |
2588 | ||
2589 | if ((n_bytes % align) != 0) | |
2590 | n_insns += 2; | |
58e17b0b | 2591 | } |
42819d4e | 2592 | |
2593 | /* Lengths are expressed in bytes now; each insn is 4 bytes. */ | |
2594 | return n_insns * 4; | |
58e17b0b | 2595 | } |
a7e1bb24 | 2596 | |
2597 | /* Emit code to perform a block clear. | |
2598 | ||
2599 | OPERANDS[0] is the destination pointer as a REG, clobbered. | |
2600 | OPERANDS[1] is a register for temporary storage. | |
2601 | OPERANDS[2] is the size as a CONST_INT | |
2602 | OPERANDS[3] is the alignment safe to use, as a CONST_INT. */ | |
2603 | ||
2604 | const char * | |
2605 | output_block_clear (rtx *operands, int size_is_constant ATTRIBUTE_UNUSED) | |
2606 | { | |
2607 | int align = INTVAL (operands[3]); | |
2608 | unsigned long n_bytes = INTVAL (operands[2]); | |
2609 | ||
2610 | /* We can't clear more than a word at a time because the PA | |
2611 | has no longer integer move insns. */ | |
2612 | if (align > (TARGET_64BIT ? 8 : 4)) | |
2613 | align = (TARGET_64BIT ? 8 : 4); | |
2614 | ||
2615 | /* Note that we know each loop below will execute at least twice | |
2616 | (else we would have open-coded the copy). */ | |
2617 | switch (align) | |
2618 | { | |
2619 | case 8: | |
2620 | /* Pre-adjust the loop counter. */ | |
2621 | operands[2] = GEN_INT (n_bytes - 16); | |
2622 | output_asm_insn ("ldi %2,%1", operands); | |
2623 | ||
2624 | /* Loop. */ | |
2625 | output_asm_insn ("std,ma %%r0,8(%0)", operands); | |
2626 | output_asm_insn ("addib,>= -16,%1,.-4", operands); | |
2627 | output_asm_insn ("std,ma %%r0,8(%0)", operands); | |
2628 | ||
2629 | /* Handle the residual. There could be up to 7 bytes of | |
2630 | residual to copy! */ | |
2631 | if (n_bytes % 16 != 0) | |
2632 | { | |
2633 | operands[2] = GEN_INT (n_bytes % 8); | |
2634 | if (n_bytes % 16 >= 8) | |
2635 | output_asm_insn ("std,ma %%r0,8(%0)", operands); | |
2636 | if (n_bytes % 8 != 0) | |
2637 | output_asm_insn ("stdby,e %%r0,%2(%0)", operands); | |
2638 | } | |
2639 | return ""; | |
2640 | ||
2641 | case 4: | |
2642 | /* Pre-adjust the loop counter. */ | |
2643 | operands[2] = GEN_INT (n_bytes - 8); | |
2644 | output_asm_insn ("ldi %2,%1", operands); | |
2645 | ||
2646 | /* Loop. */ | |
2647 | output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands); | |
2648 | output_asm_insn ("addib,>= -8,%1,.-4", operands); | |
2649 | output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands); | |
2650 | ||
2651 | /* Handle the residual. There could be up to 7 bytes of | |
2652 | residual to copy! */ | |
2653 | if (n_bytes % 8 != 0) | |
2654 | { | |
2655 | operands[2] = GEN_INT (n_bytes % 4); | |
2656 | if (n_bytes % 8 >= 4) | |
2657 | output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands); | |
2658 | if (n_bytes % 4 != 0) | |
2659 | output_asm_insn ("{stbys|stby},e %%r0,%2(%0)", operands); | |
2660 | } | |
2661 | return ""; | |
2662 | ||
2663 | case 2: | |
2664 | /* Pre-adjust the loop counter. */ | |
2665 | operands[2] = GEN_INT (n_bytes - 4); | |
2666 | output_asm_insn ("ldi %2,%1", operands); | |
2667 | ||
2668 | /* Loop. */ | |
2669 | output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands); | |
2670 | output_asm_insn ("addib,>= -4,%1,.-4", operands); | |
2671 | output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands); | |
2672 | ||
2673 | /* Handle the residual. */ | |
2674 | if (n_bytes % 4 != 0) | |
2675 | { | |
2676 | if (n_bytes % 4 >= 2) | |
2677 | output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands); | |
2678 | if (n_bytes % 2 != 0) | |
2679 | output_asm_insn ("stb %%r0,0(%0)", operands); | |
2680 | } | |
2681 | return ""; | |
2682 | ||
2683 | case 1: | |
2684 | /* Pre-adjust the loop counter. */ | |
2685 | operands[2] = GEN_INT (n_bytes - 2); | |
2686 | output_asm_insn ("ldi %2,%1", operands); | |
2687 | ||
2688 | /* Loop. */ | |
2689 | output_asm_insn ("{stbs|stb},ma %%r0,1(%0)", operands); | |
2690 | output_asm_insn ("addib,>= -2,%1,.-4", operands); | |
2691 | output_asm_insn ("{stbs|stb},ma %%r0,1(%0)", operands); | |
2692 | ||
2693 | /* Handle the residual. */ | |
2694 | if (n_bytes % 2 != 0) | |
2695 | output_asm_insn ("stb %%r0,0(%0)", operands); | |
2696 | ||
2697 | return ""; | |
2698 | ||
2699 | default: | |
2700 | abort (); | |
2701 | } | |
2702 | } | |
2703 | ||
2704 | /* Count the number of insns necessary to handle this block move. | |
2705 | ||
2706 | Basic structure is the same as emit_block_move, except that we | |
2707 | count insns rather than emit them. */ | |
2708 | ||
2709 | static int | |
2710 | compute_clrstr_length (rtx insn) | |
2711 | { | |
2712 | rtx pat = PATTERN (insn); | |
2713 | unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 4), 0)); | |
2714 | unsigned long n_bytes = INTVAL (XEXP (XVECEXP (pat, 0, 3), 0)); | |
2715 | unsigned int n_insns = 0; | |
2716 | ||
2717 | /* We can't clear more than a word at a time because the PA | |
2718 | has no longer integer move insns. */ | |
2719 | if (align > (TARGET_64BIT ? 8 : 4)) | |
2720 | align = (TARGET_64BIT ? 8 : 4); | |
2721 | ||
2722 | /* The basic loop. */ | |
2723 | n_insns = 4; | |
2724 | ||
2725 | /* Residuals. */ | |
2726 | if (n_bytes % (2 * align) != 0) | |
2727 | { | |
2728 | if ((n_bytes % (2 * align)) >= align) | |
2729 | n_insns++; | |
2730 | ||
2731 | if ((n_bytes % align) != 0) | |
2732 | n_insns++; | |
2733 | } | |
2734 | ||
2735 | /* Lengths are expressed in bytes now; each insn is 4 bytes. */ | |
2736 | return n_insns * 4; | |
2737 | } | |
87ad11b0 | 2738 | \f |
2739 | ||
611a88e1 | 2740 | const char * |
5c1d8983 | 2741 | output_and (rtx *operands) |
e057641f | 2742 | { |
d6f01525 | 2743 | if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0) |
e057641f | 2744 | { |
3745c59b | 2745 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); |
e057641f | 2746 | int ls0, ls1, ms0, p, len; |
2747 | ||
2748 | for (ls0 = 0; ls0 < 32; ls0++) | |
2749 | if ((mask & (1 << ls0)) == 0) | |
2750 | break; | |
2751 | ||
2752 | for (ls1 = ls0; ls1 < 32; ls1++) | |
2753 | if ((mask & (1 << ls1)) != 0) | |
2754 | break; | |
2755 | ||
2756 | for (ms0 = ls1; ms0 < 32; ms0++) | |
2757 | if ((mask & (1 << ms0)) == 0) | |
2758 | break; | |
2759 | ||
2760 | if (ms0 != 32) | |
ea52c577 | 2761 | abort (); |
e057641f | 2762 | |
2763 | if (ls1 == 32) | |
2764 | { | |
2765 | len = ls0; | |
2766 | ||
2767 | if (len == 0) | |
2768 | abort (); | |
2769 | ||
ef618fe4 | 2770 | operands[2] = GEN_INT (len); |
e4065f95 | 2771 | return "{extru|extrw,u} %1,31,%2,%0"; |
e057641f | 2772 | } |
2773 | else | |
2774 | { | |
2775 | /* We could use this `depi' for the case above as well, but `depi' | |
2776 | requires one more register file access than an `extru'. */ | |
2777 | ||
2778 | p = 31 - ls0; | |
2779 | len = ls1 - ls0; | |
2780 | ||
ef618fe4 | 2781 | operands[2] = GEN_INT (p); |
2782 | operands[3] = GEN_INT (len); | |
e4065f95 | 2783 | return "{depi|depwi} 0,%2,%3,%0"; |
e057641f | 2784 | } |
2785 | } | |
2786 | else | |
2787 | return "and %1,%2,%0"; | |
2788 | } | |
2789 | ||
5e3c5739 | 2790 | /* Return a string to perform a bitwise-and of operands[1] with operands[2] |
2791 | storing the result in operands[0]. */ | |
9aadea62 | 2792 | const char * |
5c1d8983 | 2793 | output_64bit_and (rtx *operands) |
5e3c5739 | 2794 | { |
2795 | if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0) | |
2796 | { | |
2797 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); | |
b7d86581 | 2798 | int ls0, ls1, ms0, p, len; |
5e3c5739 | 2799 | |
2800 | for (ls0 = 0; ls0 < HOST_BITS_PER_WIDE_INT; ls0++) | |
b7d86581 | 2801 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls0)) == 0) |
5e3c5739 | 2802 | break; |
2803 | ||
2804 | for (ls1 = ls0; ls1 < HOST_BITS_PER_WIDE_INT; ls1++) | |
b7d86581 | 2805 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls1)) != 0) |
5e3c5739 | 2806 | break; |
2807 | ||
2808 | for (ms0 = ls1; ms0 < HOST_BITS_PER_WIDE_INT; ms0++) | |
b7d86581 | 2809 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ms0)) == 0) |
5e3c5739 | 2810 | break; |
2811 | ||
2812 | if (ms0 != HOST_BITS_PER_WIDE_INT) | |
b7d86581 | 2813 | abort (); |
5e3c5739 | 2814 | |
2815 | if (ls1 == HOST_BITS_PER_WIDE_INT) | |
2816 | { | |
2817 | len = ls0; | |
2818 | ||
2819 | if (len == 0) | |
2820 | abort (); | |
2821 | ||
2822 | operands[2] = GEN_INT (len); | |
2823 | return "extrd,u %1,63,%2,%0"; | |
2824 | } | |
2825 | else | |
2826 | { | |
2827 | /* We could use this `depi' for the case above as well, but `depi' | |
2828 | requires one more register file access than an `extru'. */ | |
2829 | ||
2830 | p = 63 - ls0; | |
2831 | len = ls1 - ls0; | |
2832 | ||
2833 | operands[2] = GEN_INT (p); | |
2834 | operands[3] = GEN_INT (len); | |
2835 | return "depdi 0,%2,%3,%0"; | |
2836 | } | |
2837 | } | |
2838 | else | |
2839 | return "and %1,%2,%0"; | |
2840 | } | |
2841 | ||
611a88e1 | 2842 | const char * |
5c1d8983 | 2843 | output_ior (rtx *operands) |
e057641f | 2844 | { |
3745c59b | 2845 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); |
57ed30e5 | 2846 | int bs0, bs1, p, len; |
6d36483b | 2847 | |
c9da5f4d | 2848 | if (INTVAL (operands[2]) == 0) |
2849 | return "copy %1,%0"; | |
e057641f | 2850 | |
c9da5f4d | 2851 | for (bs0 = 0; bs0 < 32; bs0++) |
2852 | if ((mask & (1 << bs0)) != 0) | |
2853 | break; | |
e057641f | 2854 | |
c9da5f4d | 2855 | for (bs1 = bs0; bs1 < 32; bs1++) |
2856 | if ((mask & (1 << bs1)) == 0) | |
2857 | break; | |
e057641f | 2858 | |
3745c59b | 2859 | if (bs1 != 32 && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask) |
ea52c577 | 2860 | abort (); |
e057641f | 2861 | |
c9da5f4d | 2862 | p = 31 - bs0; |
2863 | len = bs1 - bs0; | |
e057641f | 2864 | |
ef618fe4 | 2865 | operands[2] = GEN_INT (p); |
2866 | operands[3] = GEN_INT (len); | |
e4065f95 | 2867 | return "{depi|depwi} -1,%2,%3,%0"; |
e057641f | 2868 | } |
5e3c5739 | 2869 | |
2870 | /* Return a string to perform a bitwise-and of operands[1] with operands[2] | |
2871 | storing the result in operands[0]. */ | |
9aadea62 | 2872 | const char * |
5c1d8983 | 2873 | output_64bit_ior (rtx *operands) |
5e3c5739 | 2874 | { |
2875 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); | |
b7d86581 | 2876 | int bs0, bs1, p, len; |
5e3c5739 | 2877 | |
2878 | if (INTVAL (operands[2]) == 0) | |
2879 | return "copy %1,%0"; | |
2880 | ||
2881 | for (bs0 = 0; bs0 < HOST_BITS_PER_WIDE_INT; bs0++) | |
b7d86581 | 2882 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs0)) != 0) |
5e3c5739 | 2883 | break; |
2884 | ||
2885 | for (bs1 = bs0; bs1 < HOST_BITS_PER_WIDE_INT; bs1++) | |
b7d86581 | 2886 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs1)) == 0) |
5e3c5739 | 2887 | break; |
2888 | ||
2889 | if (bs1 != HOST_BITS_PER_WIDE_INT | |
2890 | && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask) | |
b7d86581 | 2891 | abort (); |
5e3c5739 | 2892 | |
2893 | p = 63 - bs0; | |
2894 | len = bs1 - bs0; | |
2895 | ||
2896 | operands[2] = GEN_INT (p); | |
2897 | operands[3] = GEN_INT (len); | |
2898 | return "depdi -1,%2,%3,%0"; | |
2899 | } | |
e057641f | 2900 | \f |
58356836 | 2901 | /* Target hook for assembling integer objects. This code handles |
2902 | aligned SI and DI integers specially, since function references must | |
2903 | be preceded by P%. */ | |
2904 | ||
2905 | static bool | |
5c1d8983 | 2906 | pa_assemble_integer (rtx x, unsigned int size, int aligned_p) |
58356836 | 2907 | { |
2908 | if (size == UNITS_PER_WORD && aligned_p | |
2909 | && function_label_operand (x, VOIDmode)) | |
2910 | { | |
2911 | fputs (size == 8? "\t.dword\tP%" : "\t.word\tP%", asm_out_file); | |
2912 | output_addr_const (asm_out_file, x); | |
2913 | fputc ('\n', asm_out_file); | |
2914 | return true; | |
2915 | } | |
2916 | return default_assemble_integer (x, size, aligned_p); | |
2917 | } | |
2918 | \f | |
87ad11b0 | 2919 | /* Output an ascii string. */ |
57ed30e5 | 2920 | void |
5c1d8983 | 2921 | output_ascii (FILE *file, const char *p, int size) |
87ad11b0 | 2922 | { |
2923 | int i; | |
2924 | int chars_output; | |
2925 | unsigned char partial_output[16]; /* Max space 4 chars can occupy. */ | |
2926 | ||
2927 | /* The HP assembler can only take strings of 256 characters at one | |
2928 | time. This is a limitation on input line length, *not* the | |
2929 | length of the string. Sigh. Even worse, it seems that the | |
2930 | restriction is in number of input characters (see \xnn & | |
2931 | \whatever). So we have to do this very carefully. */ | |
2932 | ||
9c0ac0fd | 2933 | fputs ("\t.STRING \"", file); |
87ad11b0 | 2934 | |
2935 | chars_output = 0; | |
2936 | for (i = 0; i < size; i += 4) | |
2937 | { | |
2938 | int co = 0; | |
2939 | int io = 0; | |
2940 | for (io = 0, co = 0; io < MIN (4, size - i); io++) | |
2941 | { | |
bf8aac3e | 2942 | register unsigned int c = (unsigned char) p[i + io]; |
87ad11b0 | 2943 | |
2944 | if (c == '\"' || c == '\\') | |
2945 | partial_output[co++] = '\\'; | |
2946 | if (c >= ' ' && c < 0177) | |
2947 | partial_output[co++] = c; | |
2948 | else | |
2949 | { | |
2950 | unsigned int hexd; | |
2951 | partial_output[co++] = '\\'; | |
2952 | partial_output[co++] = 'x'; | |
2953 | hexd = c / 16 - 0 + '0'; | |
2954 | if (hexd > '9') | |
2955 | hexd -= '9' - 'a' + 1; | |
2956 | partial_output[co++] = hexd; | |
2957 | hexd = c % 16 - 0 + '0'; | |
2958 | if (hexd > '9') | |
2959 | hexd -= '9' - 'a' + 1; | |
2960 | partial_output[co++] = hexd; | |
2961 | } | |
2962 | } | |
2963 | if (chars_output + co > 243) | |
2964 | { | |
9c0ac0fd | 2965 | fputs ("\"\n\t.STRING \"", file); |
87ad11b0 | 2966 | chars_output = 0; |
2967 | } | |
a584fe8a | 2968 | fwrite (partial_output, 1, (size_t) co, file); |
87ad11b0 | 2969 | chars_output += co; |
2970 | co = 0; | |
2971 | } | |
9c0ac0fd | 2972 | fputs ("\"\n", file); |
87ad11b0 | 2973 | } |
c533da59 | 2974 | |
2975 | /* Try to rewrite floating point comparisons & branches to avoid | |
2976 | useless add,tr insns. | |
2977 | ||
2978 | CHECK_NOTES is nonzero if we should examine REG_DEAD notes | |
2979 | to see if FPCC is dead. CHECK_NOTES is nonzero for the | |
2980 | first attempt to remove useless add,tr insns. It is zero | |
2981 | for the second pass as reorg sometimes leaves bogus REG_DEAD | |
2982 | notes lying around. | |
2983 | ||
2984 | When CHECK_NOTES is zero we can only eliminate add,tr insns | |
2985 | when there's a 1:1 correspondence between fcmp and ftest/fbranch | |
2986 | instructions. */ | |
611a88e1 | 2987 | static void |
5c1d8983 | 2988 | remove_useless_addtr_insns (int check_notes) |
c533da59 | 2989 | { |
2990 | rtx insn; | |
c533da59 | 2991 | static int pass = 0; |
2992 | ||
2993 | /* This is fairly cheap, so always run it when optimizing. */ | |
2994 | if (optimize > 0) | |
2995 | { | |
2996 | int fcmp_count = 0; | |
2997 | int fbranch_count = 0; | |
2998 | ||
2999 | /* Walk all the insns in this function looking for fcmp & fbranch | |
3000 | instructions. Keep track of how many of each we find. */ | |
2efea8c0 | 3001 | for (insn = get_insns (); insn; insn = next_insn (insn)) |
c533da59 | 3002 | { |
3003 | rtx tmp; | |
3004 | ||
3005 | /* Ignore anything that isn't an INSN or a JUMP_INSN. */ | |
3006 | if (GET_CODE (insn) != INSN && GET_CODE (insn) != JUMP_INSN) | |
3007 | continue; | |
3008 | ||
3009 | tmp = PATTERN (insn); | |
3010 | ||
3011 | /* It must be a set. */ | |
3012 | if (GET_CODE (tmp) != SET) | |
3013 | continue; | |
3014 | ||
3015 | /* If the destination is CCFP, then we've found an fcmp insn. */ | |
3016 | tmp = SET_DEST (tmp); | |
3017 | if (GET_CODE (tmp) == REG && REGNO (tmp) == 0) | |
3018 | { | |
3019 | fcmp_count++; | |
3020 | continue; | |
3021 | } | |
9840d99d | 3022 | |
c533da59 | 3023 | tmp = PATTERN (insn); |
3024 | /* If this is an fbranch instruction, bump the fbranch counter. */ | |
3025 | if (GET_CODE (tmp) == SET | |
3026 | && SET_DEST (tmp) == pc_rtx | |
3027 | && GET_CODE (SET_SRC (tmp)) == IF_THEN_ELSE | |
3028 | && GET_CODE (XEXP (SET_SRC (tmp), 0)) == NE | |
3029 | && GET_CODE (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == REG | |
3030 | && REGNO (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == 0) | |
3031 | { | |
3032 | fbranch_count++; | |
3033 | continue; | |
3034 | } | |
3035 | } | |
3036 | ||
3037 | ||
3038 | /* Find all floating point compare + branch insns. If possible, | |
3039 | reverse the comparison & the branch to avoid add,tr insns. */ | |
2efea8c0 | 3040 | for (insn = get_insns (); insn; insn = next_insn (insn)) |
c533da59 | 3041 | { |
3042 | rtx tmp, next; | |
3043 | ||
3044 | /* Ignore anything that isn't an INSN. */ | |
3045 | if (GET_CODE (insn) != INSN) | |
3046 | continue; | |
3047 | ||
3048 | tmp = PATTERN (insn); | |
3049 | ||
3050 | /* It must be a set. */ | |
3051 | if (GET_CODE (tmp) != SET) | |
3052 | continue; | |
3053 | ||
3054 | /* The destination must be CCFP, which is register zero. */ | |
3055 | tmp = SET_DEST (tmp); | |
3056 | if (GET_CODE (tmp) != REG || REGNO (tmp) != 0) | |
3057 | continue; | |
3058 | ||
3059 | /* INSN should be a set of CCFP. | |
3060 | ||
3061 | See if the result of this insn is used in a reversed FP | |
3062 | conditional branch. If so, reverse our condition and | |
3063 | the branch. Doing so avoids useless add,tr insns. */ | |
3064 | next = next_insn (insn); | |
3065 | while (next) | |
3066 | { | |
3067 | /* Jumps, calls and labels stop our search. */ | |
3068 | if (GET_CODE (next) == JUMP_INSN | |
3069 | || GET_CODE (next) == CALL_INSN | |
3070 | || GET_CODE (next) == CODE_LABEL) | |
3071 | break; | |
3072 | ||
3073 | /* As does another fcmp insn. */ | |
3074 | if (GET_CODE (next) == INSN | |
3075 | && GET_CODE (PATTERN (next)) == SET | |
3076 | && GET_CODE (SET_DEST (PATTERN (next))) == REG | |
3077 | && REGNO (SET_DEST (PATTERN (next))) == 0) | |
3078 | break; | |
3079 | ||
3080 | next = next_insn (next); | |
3081 | } | |
3082 | ||
3083 | /* Is NEXT_INSN a branch? */ | |
3084 | if (next | |
3085 | && GET_CODE (next) == JUMP_INSN) | |
3086 | { | |
3087 | rtx pattern = PATTERN (next); | |
3088 | ||
3089 | /* If it a reversed fp conditional branch (eg uses add,tr) | |
3090 | and CCFP dies, then reverse our conditional and the branch | |
3091 | to avoid the add,tr. */ | |
3092 | if (GET_CODE (pattern) == SET | |
3093 | && SET_DEST (pattern) == pc_rtx | |
3094 | && GET_CODE (SET_SRC (pattern)) == IF_THEN_ELSE | |
3095 | && GET_CODE (XEXP (SET_SRC (pattern), 0)) == NE | |
3096 | && GET_CODE (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == REG | |
3097 | && REGNO (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == 0 | |
3098 | && GET_CODE (XEXP (SET_SRC (pattern), 1)) == PC | |
3099 | && (fcmp_count == fbranch_count | |
3100 | || (check_notes | |
3101 | && find_regno_note (next, REG_DEAD, 0)))) | |
3102 | { | |
3103 | /* Reverse the branch. */ | |
3104 | tmp = XEXP (SET_SRC (pattern), 1); | |
3105 | XEXP (SET_SRC (pattern), 1) = XEXP (SET_SRC (pattern), 2); | |
3106 | XEXP (SET_SRC (pattern), 2) = tmp; | |
3107 | INSN_CODE (next) = -1; | |
3108 | ||
3109 | /* Reverse our condition. */ | |
3110 | tmp = PATTERN (insn); | |
3111 | PUT_CODE (XEXP (tmp, 1), | |
ea52c577 | 3112 | (reverse_condition_maybe_unordered |
3113 | (GET_CODE (XEXP (tmp, 1))))); | |
c533da59 | 3114 | } |
3115 | } | |
3116 | } | |
3117 | } | |
3118 | ||
3119 | pass = !pass; | |
3120 | ||
3121 | } | |
87ad11b0 | 3122 | \f |
ea52c577 | 3123 | /* You may have trouble believing this, but this is the 32 bit HP-PA |
3124 | stack layout. Wow. | |
87ad11b0 | 3125 | |
3126 | Offset Contents | |
3127 | ||
3128 | Variable arguments (optional; any number may be allocated) | |
3129 | ||
3130 | SP-(4*(N+9)) arg word N | |
3131 | : : | |
3132 | SP-56 arg word 5 | |
3133 | SP-52 arg word 4 | |
3134 | ||
3135 | Fixed arguments (must be allocated; may remain unused) | |
3136 | ||
3137 | SP-48 arg word 3 | |
3138 | SP-44 arg word 2 | |
3139 | SP-40 arg word 1 | |
3140 | SP-36 arg word 0 | |
3141 | ||
3142 | Frame Marker | |
3143 | ||
3144 | SP-32 External Data Pointer (DP) | |
3145 | SP-28 External sr4 | |
3146 | SP-24 External/stub RP (RP') | |
3147 | SP-20 Current RP | |
3148 | SP-16 Static Link | |
3149 | SP-12 Clean up | |
3150 | SP-8 Calling Stub RP (RP'') | |
3151 | SP-4 Previous SP | |
3152 | ||
3153 | Top of Frame | |
3154 | ||
3155 | SP-0 Stack Pointer (points to next available address) | |
3156 | ||
3157 | */ | |
3158 | ||
3159 | /* This function saves registers as follows. Registers marked with ' are | |
3160 | this function's registers (as opposed to the previous function's). | |
3161 | If a frame_pointer isn't needed, r4 is saved as a general register; | |
3162 | the space for the frame pointer is still allocated, though, to keep | |
3163 | things simple. | |
3164 | ||
3165 | ||
3166 | Top of Frame | |
3167 | ||
3168 | SP (FP') Previous FP | |
3169 | SP + 4 Alignment filler (sigh) | |
3170 | SP + 8 Space for locals reserved here. | |
3171 | . | |
3172 | . | |
3173 | . | |
3174 | SP + n All call saved register used. | |
3175 | . | |
3176 | . | |
3177 | . | |
3178 | SP + o All call saved fp registers used. | |
3179 | . | |
3180 | . | |
3181 | . | |
3182 | SP + p (SP') points to next available address. | |
6d36483b | 3183 | |
87ad11b0 | 3184 | */ |
3185 | ||
17d9b0c3 | 3186 | /* Global variables set by output_function_prologue(). */ |
cc858176 | 3187 | /* Size of frame. Need to know this to emit return insns from |
3188 | leaf procedures. */ | |
3189 | static int actual_fsize; | |
3190 | static int local_fsize, save_fregs; | |
3191 | ||
daee63dd | 3192 | /* Emit RTL to store REG at the memory location specified by BASE+DISP. |
359a0be8 | 3193 | Handle case where DISP > 8k by using the add_high_const patterns. |
daee63dd | 3194 | |
3195 | Note in DISP > 8k case, we will leave the high part of the address | |
3196 | in %r1. There is code in expand_hppa_{prologue,epilogue} that knows this.*/ | |
7014838c | 3197 | |
6a2c16d6 | 3198 | static void |
5c1d8983 | 3199 | store_reg (int reg, int disp, int base) |
87ad11b0 | 3200 | { |
6a2c16d6 | 3201 | rtx insn, dest, src, basereg; |
cc858176 | 3202 | |
3203 | src = gen_rtx_REG (word_mode, reg); | |
3204 | basereg = gen_rtx_REG (Pmode, base); | |
87ad11b0 | 3205 | if (VAL_14_BITS_P (disp)) |
daee63dd | 3206 | { |
cc858176 | 3207 | dest = gen_rtx_MEM (word_mode, plus_constant (basereg, disp)); |
6a2c16d6 | 3208 | insn = emit_move_insn (dest, src); |
daee63dd | 3209 | } |
daee63dd | 3210 | else |
3211 | { | |
cc858176 | 3212 | rtx delta = GEN_INT (disp); |
3213 | rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta)); | |
3214 | rtx tmpreg = gen_rtx_REG (Pmode, 1); | |
3215 | emit_move_insn (tmpreg, high); | |
3216 | dest = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta)); | |
6a2c16d6 | 3217 | insn = emit_move_insn (dest, src); |
3218 | if (DO_FRAME_NOTES) | |
3219 | { | |
3220 | REG_NOTES (insn) | |
3221 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3222 | gen_rtx_SET (VOIDmode, | |
3223 | gen_rtx_MEM (word_mode, | |
3224 | gen_rtx_PLUS (word_mode, basereg, | |
3225 | delta)), | |
3226 | src), | |
3227 | REG_NOTES (insn)); | |
3228 | } | |
daee63dd | 3229 | } |
6a2c16d6 | 3230 | |
3231 | if (DO_FRAME_NOTES) | |
3232 | RTX_FRAME_RELATED_P (insn) = 1; | |
daee63dd | 3233 | } |
3234 | ||
a584fe8a | 3235 | /* Emit RTL to store REG at the memory location specified by BASE and then |
3236 | add MOD to BASE. MOD must be <= 8k. */ | |
daee63dd | 3237 | |
a584fe8a | 3238 | static void |
5c1d8983 | 3239 | store_reg_modify (int base, int reg, int mod) |
a584fe8a | 3240 | { |
3241 | rtx insn, basereg, srcreg, delta; | |
3242 | ||
3243 | if (! VAL_14_BITS_P (mod)) | |
3244 | abort (); | |
3245 | ||
3246 | basereg = gen_rtx_REG (Pmode, base); | |
3247 | srcreg = gen_rtx_REG (word_mode, reg); | |
3248 | delta = GEN_INT (mod); | |
3249 | ||
3250 | insn = emit_insn (gen_post_store (basereg, srcreg, delta)); | |
3251 | if (DO_FRAME_NOTES) | |
3252 | { | |
3253 | RTX_FRAME_RELATED_P (insn) = 1; | |
3254 | ||
3255 | /* RTX_FRAME_RELATED_P must be set on each frame related set | |
3256 | in a parallel with more than one element. Don't set | |
3257 | RTX_FRAME_RELATED_P in the first set if reg is temporary | |
3258 | register 1. The effect of this operation is recorded in | |
3259 | the initial copy. */ | |
3260 | if (reg != 1) | |
3261 | { | |
3262 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1; | |
3263 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
3264 | } | |
3265 | else | |
3266 | { | |
3267 | /* The first element of a PARALLEL is always processed if it is | |
3268 | a SET. Thus, we need an expression list for this case. */ | |
3269 | REG_NOTES (insn) | |
3270 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3271 | gen_rtx_SET (VOIDmode, basereg, | |
3272 | gen_rtx_PLUS (word_mode, basereg, delta)), | |
3273 | REG_NOTES (insn)); | |
3274 | } | |
3275 | } | |
3276 | } | |
3277 | ||
3278 | /* Emit RTL to set REG to the value specified by BASE+DISP. Handle case | |
3279 | where DISP > 8k by using the add_high_const patterns. NOTE indicates | |
3280 | whether to add a frame note or not. | |
3281 | ||
3282 | In the DISP > 8k case, we leave the high part of the address in %r1. | |
3283 | There is code in expand_hppa_{prologue,epilogue} that knows about this. */ | |
7014838c | 3284 | |
6a2c16d6 | 3285 | static void |
5c1d8983 | 3286 | set_reg_plus_d (int reg, int base, int disp, int note) |
87ad11b0 | 3287 | { |
6a2c16d6 | 3288 | rtx insn; |
cc858176 | 3289 | |
87ad11b0 | 3290 | if (VAL_14_BITS_P (disp)) |
cc858176 | 3291 | { |
6a2c16d6 | 3292 | insn = emit_move_insn (gen_rtx_REG (Pmode, reg), |
3293 | plus_constant (gen_rtx_REG (Pmode, base), disp)); | |
cc858176 | 3294 | } |
87ad11b0 | 3295 | else |
daee63dd | 3296 | { |
6a2c16d6 | 3297 | rtx basereg = gen_rtx_REG (Pmode, base); |
cc858176 | 3298 | rtx delta = GEN_INT (disp); |
6a2c16d6 | 3299 | |
359a0be8 | 3300 | emit_move_insn (gen_rtx_REG (Pmode, 1), |
6a2c16d6 | 3301 | gen_rtx_PLUS (Pmode, basereg, |
cc858176 | 3302 | gen_rtx_HIGH (Pmode, delta))); |
6a2c16d6 | 3303 | insn = emit_move_insn (gen_rtx_REG (Pmode, reg), |
3304 | gen_rtx_LO_SUM (Pmode, gen_rtx_REG (Pmode, 1), | |
3305 | delta)); | |
daee63dd | 3306 | } |
6a2c16d6 | 3307 | |
a584fe8a | 3308 | if (DO_FRAME_NOTES && note) |
6a2c16d6 | 3309 | RTX_FRAME_RELATED_P (insn) = 1; |
87ad11b0 | 3310 | } |
3311 | ||
3312 | int | |
5c1d8983 | 3313 | compute_frame_size (int size, int *fregs_live) |
87ad11b0 | 3314 | { |
256f9b65 | 3315 | int freg_saved = 0; |
3316 | int i, j; | |
3317 | ||
3318 | /* The code in hppa_expand_prologue and hppa_expand_epilogue must | |
3319 | be consistent with the rounding and size calculation done here. | |
3320 | Change them at the same time. */ | |
3321 | ||
3322 | /* We do our own stack alignment. First, round the size of the | |
3323 | stack locals up to a word boundary. */ | |
3324 | size = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1); | |
3325 | ||
3326 | /* Space for previous frame pointer + filler. If any frame is | |
3327 | allocated, we need to add in the STARTING_FRAME_OFFSET. We | |
3328 | waste some space here for the sake of HP compatibility. The | |
3329 | first slot is only used when the frame pointer is needed. */ | |
3330 | if (size || frame_pointer_needed) | |
3331 | size += STARTING_FRAME_OFFSET; | |
3332 | ||
a584fe8a | 3333 | /* If the current function calls __builtin_eh_return, then we need |
3334 | to allocate stack space for registers that will hold data for | |
3335 | the exception handler. */ | |
3336 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3337 | { | |
3338 | unsigned int i; | |
3339 | ||
3340 | for (i = 0; EH_RETURN_DATA_REGNO (i) != INVALID_REGNUM; ++i) | |
3341 | continue; | |
256f9b65 | 3342 | size += i * UNITS_PER_WORD; |
a584fe8a | 3343 | } |
3344 | ||
3a51bad9 | 3345 | /* Account for space used by the callee general register saves. */ |
256f9b65 | 3346 | for (i = 18, j = frame_pointer_needed ? 4 : 3; i >= j; i--) |
7f7c4869 | 3347 | if (regs_ever_live[i]) |
256f9b65 | 3348 | size += UNITS_PER_WORD; |
df0651dc | 3349 | |
3a51bad9 | 3350 | /* Account for space used by the callee floating point register saves. */ |
bac38c40 | 3351 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
5e3c5739 | 3352 | if (regs_ever_live[i] |
256f9b65 | 3353 | || (!TARGET_64BIT && regs_ever_live[i + 1])) |
df0651dc | 3354 | { |
256f9b65 | 3355 | freg_saved = 1; |
002fc5f7 | 3356 | |
3a51bad9 | 3357 | /* We always save both halves of the FP register, so always |
3358 | increment the frame size by 8 bytes. */ | |
256f9b65 | 3359 | size += 8; |
df0651dc | 3360 | } |
3361 | ||
256f9b65 | 3362 | /* If any of the floating registers are saved, account for the |
3363 | alignment needed for the floating point register save block. */ | |
3364 | if (freg_saved) | |
3365 | { | |
3366 | size = (size + 7) & ~7; | |
3367 | if (fregs_live) | |
3368 | *fregs_live = 1; | |
3369 | } | |
3370 | ||
3a51bad9 | 3371 | /* The various ABIs include space for the outgoing parameters in the |
256f9b65 | 3372 | size of the current function's stack frame. We don't need to align |
3373 | for the outgoing arguments as their alignment is set by the final | |
3374 | rounding for the frame as a whole. */ | |
3375 | size += current_function_outgoing_args_size; | |
3a51bad9 | 3376 | |
3377 | /* Allocate space for the fixed frame marker. This space must be | |
e9ec370e | 3378 | allocated for any function that makes calls or allocates |
3a51bad9 | 3379 | stack space. */ |
256f9b65 | 3380 | if (!current_function_is_leaf || size) |
e9ec370e | 3381 | size += TARGET_64BIT ? 48 : 32; |
5e3c5739 | 3382 | |
256f9b65 | 3383 | /* Finally, round to the preferred stack boundary. */ |
2247cc5f | 3384 | return ((size + PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1) |
3385 | & ~(PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)); | |
87ad11b0 | 3386 | } |
6d36483b | 3387 | |
17d9b0c3 | 3388 | /* Generate the assembly code for function entry. FILE is a stdio |
3389 | stream to output the code to. SIZE is an int: how many units of | |
3390 | temporary storage to allocate. | |
3391 | ||
3392 | Refer to the array `regs_ever_live' to determine which registers to | |
3393 | save; `regs_ever_live[I]' is nonzero if register number I is ever | |
3394 | used in the function. This function is responsible for knowing | |
3395 | which registers should not be saved even if used. */ | |
3396 | ||
3397 | /* On HP-PA, move-double insns between fpu and cpu need an 8-byte block | |
3398 | of memory. If any fpu reg is used in the function, we allocate | |
3399 | such a block here, at the bottom of the frame, just in case it's needed. | |
3400 | ||
3401 | If this function is a leaf procedure, then we may choose not | |
3402 | to do a "save" insn. The decision about whether or not | |
3403 | to do this is made in regclass.c. */ | |
3404 | ||
6988553d | 3405 | static void |
5c1d8983 | 3406 | pa_output_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED) |
87ad11b0 | 3407 | { |
d151162a | 3408 | /* The function's label and associated .PROC must never be |
3409 | separated and must be output *after* any profiling declarations | |
3410 | to avoid changing spaces/subspaces within a procedure. */ | |
3411 | ASM_OUTPUT_LABEL (file, XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0)); | |
3412 | fputs ("\t.PROC\n", file); | |
3413 | ||
daee63dd | 3414 | /* hppa_expand_prologue does the dirty work now. We just need |
3415 | to output the assembler directives which denote the start | |
3416 | of a function. */ | |
2acd4f33 | 3417 | fprintf (file, "\t.CALLINFO FRAME=%d", actual_fsize); |
df6edefa | 3418 | if (regs_ever_live[2]) |
9c0ac0fd | 3419 | fputs (",CALLS,SAVE_RP", file); |
daee63dd | 3420 | else |
9c0ac0fd | 3421 | fputs (",NO_CALLS", file); |
f3ba7709 | 3422 | |
e9ec370e | 3423 | /* The SAVE_SP flag is used to indicate that register %r3 is stored |
3424 | at the beginning of the frame and that it is used as the frame | |
3425 | pointer for the frame. We do this because our current frame | |
3426 | layout doesn't conform to that specified in the the HP runtime | |
3427 | documentation and we need a way to indicate to programs such as | |
3428 | GDB where %r3 is saved. The SAVE_SP flag was chosen because it | |
3429 | isn't used by HP compilers but is supported by the assembler. | |
3430 | However, SAVE_SP is supposed to indicate that the previous stack | |
3431 | pointer has been saved in the frame marker. */ | |
f3ba7709 | 3432 | if (frame_pointer_needed) |
9c0ac0fd | 3433 | fputs (",SAVE_SP", file); |
f3ba7709 | 3434 | |
a9960cdc | 3435 | /* Pass on information about the number of callee register saves |
9b0c95be | 3436 | performed in the prologue. |
3437 | ||
3438 | The compiler is supposed to pass the highest register number | |
6d36483b | 3439 | saved, the assembler then has to adjust that number before |
9b0c95be | 3440 | entering it into the unwind descriptor (to account for any |
6d36483b | 3441 | caller saved registers with lower register numbers than the |
9b0c95be | 3442 | first callee saved register). */ |
3443 | if (gr_saved) | |
3444 | fprintf (file, ",ENTRY_GR=%d", gr_saved + 2); | |
3445 | ||
3446 | if (fr_saved) | |
3447 | fprintf (file, ",ENTRY_FR=%d", fr_saved + 11); | |
a9960cdc | 3448 | |
9c0ac0fd | 3449 | fputs ("\n\t.ENTRY\n", file); |
daee63dd | 3450 | |
2efea8c0 | 3451 | remove_useless_addtr_insns (0); |
daee63dd | 3452 | } |
3453 | ||
57ed30e5 | 3454 | void |
5c1d8983 | 3455 | hppa_expand_prologue (void) |
daee63dd | 3456 | { |
afd7b680 | 3457 | int merge_sp_adjust_with_store = 0; |
256f9b65 | 3458 | int size = get_frame_size (); |
daee63dd | 3459 | int i, offset; |
a584fe8a | 3460 | rtx insn, tmpreg; |
daee63dd | 3461 | |
a9960cdc | 3462 | gr_saved = 0; |
3463 | fr_saved = 0; | |
3ddcbb9d | 3464 | save_fregs = 0; |
3a51bad9 | 3465 | |
256f9b65 | 3466 | /* Compute total size for frame pointer, filler, locals and rounding to |
3467 | the next word boundary. Similar code appears in compute_frame_size | |
3468 | and must be changed in tandem with this code. */ | |
3469 | local_fsize = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1); | |
3470 | if (local_fsize || frame_pointer_needed) | |
3471 | local_fsize += STARTING_FRAME_OFFSET; | |
3a51bad9 | 3472 | |
a1ab4fa3 | 3473 | actual_fsize = compute_frame_size (size, &save_fregs); |
87ad11b0 | 3474 | |
daee63dd | 3475 | /* Compute a few things we will use often. */ |
440c23df | 3476 | tmpreg = gen_rtx_REG (word_mode, 1); |
87ad11b0 | 3477 | |
6d36483b | 3478 | /* Save RP first. The calling conventions manual states RP will |
cc858176 | 3479 | always be stored into the caller's frame at sp - 20 or sp - 16 |
5e3c5739 | 3480 | depending on which ABI is in use. */ |
a584fe8a | 3481 | if (regs_ever_live[2] || current_function_calls_eh_return) |
6a2c16d6 | 3482 | store_reg (2, TARGET_64BIT ? -16 : -20, STACK_POINTER_REGNUM); |
6d36483b | 3483 | |
daee63dd | 3484 | /* Allocate the local frame and set up the frame pointer if needed. */ |
58361f39 | 3485 | if (actual_fsize != 0) |
3486 | { | |
3487 | if (frame_pointer_needed) | |
3488 | { | |
3489 | /* Copy the old frame pointer temporarily into %r1. Set up the | |
3490 | new stack pointer, then store away the saved old frame pointer | |
a584fe8a | 3491 | into the stack at sp and at the same time update the stack |
3492 | pointer by actual_fsize bytes. Two versions, first | |
58361f39 | 3493 | handles small (<8k) frames. The second handles large (>=8k) |
3494 | frames. */ | |
a584fe8a | 3495 | insn = emit_move_insn (tmpreg, frame_pointer_rtx); |
3496 | if (DO_FRAME_NOTES) | |
cc858176 | 3497 | { |
a584fe8a | 3498 | /* We need to record the frame pointer save here since the |
3499 | new frame pointer is set in the following insn. */ | |
3500 | RTX_FRAME_RELATED_P (insn) = 1; | |
3501 | REG_NOTES (insn) | |
3502 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3503 | gen_rtx_SET (VOIDmode, | |
3504 | gen_rtx_MEM (word_mode, stack_pointer_rtx), | |
3505 | frame_pointer_rtx), | |
3506 | REG_NOTES (insn)); | |
cc858176 | 3507 | } |
a584fe8a | 3508 | |
3509 | insn = emit_move_insn (frame_pointer_rtx, stack_pointer_rtx); | |
3510 | if (DO_FRAME_NOTES) | |
3511 | RTX_FRAME_RELATED_P (insn) = 1; | |
3512 | ||
3513 | if (VAL_14_BITS_P (actual_fsize)) | |
3514 | store_reg_modify (STACK_POINTER_REGNUM, 1, actual_fsize); | |
58361f39 | 3515 | else |
3516 | { | |
3517 | /* It is incorrect to store the saved frame pointer at *sp, | |
3518 | then increment sp (writes beyond the current stack boundary). | |
3519 | ||
3520 | So instead use stwm to store at *sp and post-increment the | |
3521 | stack pointer as an atomic operation. Then increment sp to | |
3522 | finish allocating the new frame. */ | |
3523 | int adjust1 = 8192 - 64; | |
3524 | int adjust2 = actual_fsize - adjust1; | |
cc858176 | 3525 | |
a584fe8a | 3526 | store_reg_modify (STACK_POINTER_REGNUM, 1, adjust1); |
6a2c16d6 | 3527 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3528 | adjust2, 1); |
58361f39 | 3529 | } |
a584fe8a | 3530 | |
e9ec370e | 3531 | /* We set SAVE_SP in frames that need a frame pointer. Thus, |
3532 | we need to store the previous stack pointer (frame pointer) | |
3533 | into the frame marker on targets that use the HP unwind | |
3534 | library. This allows the HP unwind library to be used to | |
3535 | unwind GCC frames. However, we are not fully compatible | |
3536 | with the HP library because our frame layout differs from | |
3537 | that specified in the HP runtime specification. | |
3538 | ||
3539 | We don't want a frame note on this instruction as the frame | |
3540 | marker moves during dynamic stack allocation. | |
3541 | ||
3542 | This instruction also serves as a blockage to prevent | |
3543 | register spills from being scheduled before the stack | |
3544 | pointer is raised. This is necessary as we store | |
3545 | registers using the frame pointer as a base register, | |
3546 | and the frame pointer is set before sp is raised. */ | |
3547 | if (TARGET_HPUX_UNWIND_LIBRARY) | |
3548 | { | |
3549 | rtx addr = gen_rtx_PLUS (word_mode, stack_pointer_rtx, | |
3550 | GEN_INT (TARGET_64BIT ? -8 : -4)); | |
3551 | ||
3552 | emit_move_insn (gen_rtx_MEM (word_mode, addr), | |
3553 | frame_pointer_rtx); | |
3554 | } | |
3555 | else | |
3556 | emit_insn (gen_blockage ()); | |
58361f39 | 3557 | } |
3558 | /* no frame pointer needed. */ | |
3559 | else | |
3560 | { | |
3561 | /* In some cases we can perform the first callee register save | |
3562 | and allocating the stack frame at the same time. If so, just | |
3563 | make a note of it and defer allocating the frame until saving | |
3564 | the callee registers. */ | |
df6edefa | 3565 | if (VAL_14_BITS_P (actual_fsize) && local_fsize == 0) |
58361f39 | 3566 | merge_sp_adjust_with_store = 1; |
3567 | /* Can not optimize. Adjust the stack frame by actual_fsize | |
3568 | bytes. */ | |
3569 | else | |
6a2c16d6 | 3570 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3571 | actual_fsize, 1); |
58361f39 | 3572 | } |
372ef038 | 3573 | } |
3574 | ||
6d36483b | 3575 | /* Normal register save. |
daee63dd | 3576 | |
3577 | Do not save the frame pointer in the frame_pointer_needed case. It | |
3578 | was done earlier. */ | |
87ad11b0 | 3579 | if (frame_pointer_needed) |
3580 | { | |
a584fe8a | 3581 | offset = local_fsize; |
3582 | ||
3583 | /* Saving the EH return data registers in the frame is the simplest | |
3584 | way to get the frame unwind information emitted. We put them | |
3585 | just before the general registers. */ | |
3586 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3587 | { | |
3588 | unsigned int i, regno; | |
3589 | ||
3590 | for (i = 0; ; ++i) | |
3591 | { | |
3592 | regno = EH_RETURN_DATA_REGNO (i); | |
3593 | if (regno == INVALID_REGNUM) | |
3594 | break; | |
3595 | ||
3596 | store_reg (regno, offset, FRAME_POINTER_REGNUM); | |
3597 | offset += UNITS_PER_WORD; | |
3598 | } | |
3599 | } | |
3600 | ||
3601 | for (i = 18; i >= 4; i--) | |
98328a39 | 3602 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3603 | { |
6a2c16d6 | 3604 | store_reg (i, offset, FRAME_POINTER_REGNUM); |
6ec4380b | 3605 | offset += UNITS_PER_WORD; |
a9960cdc | 3606 | gr_saved++; |
87ad11b0 | 3607 | } |
7f7c4869 | 3608 | /* Account for %r3 which is saved in a special place. */ |
9b0c95be | 3609 | gr_saved++; |
87ad11b0 | 3610 | } |
daee63dd | 3611 | /* No frame pointer needed. */ |
87ad11b0 | 3612 | else |
3613 | { | |
a584fe8a | 3614 | offset = local_fsize - actual_fsize; |
3615 | ||
3616 | /* Saving the EH return data registers in the frame is the simplest | |
3617 | way to get the frame unwind information emitted. */ | |
3618 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3619 | { | |
3620 | unsigned int i, regno; | |
3621 | ||
3622 | for (i = 0; ; ++i) | |
3623 | { | |
3624 | regno = EH_RETURN_DATA_REGNO (i); | |
3625 | if (regno == INVALID_REGNUM) | |
3626 | break; | |
3627 | ||
3628 | /* If merge_sp_adjust_with_store is nonzero, then we can | |
3629 | optimize the first save. */ | |
3630 | if (merge_sp_adjust_with_store) | |
3631 | { | |
3632 | store_reg_modify (STACK_POINTER_REGNUM, regno, -offset); | |
3633 | merge_sp_adjust_with_store = 0; | |
3634 | } | |
3635 | else | |
3636 | store_reg (regno, offset, STACK_POINTER_REGNUM); | |
3637 | offset += UNITS_PER_WORD; | |
3638 | } | |
3639 | } | |
3640 | ||
3641 | for (i = 18; i >= 3; i--) | |
98328a39 | 3642 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3643 | { |
6d36483b | 3644 | /* If merge_sp_adjust_with_store is nonzero, then we can |
afd7b680 | 3645 | optimize the first GR save. */ |
201f01e9 | 3646 | if (merge_sp_adjust_with_store) |
afd7b680 | 3647 | { |
a584fe8a | 3648 | store_reg_modify (STACK_POINTER_REGNUM, i, -offset); |
afd7b680 | 3649 | merge_sp_adjust_with_store = 0; |
afd7b680 | 3650 | } |
3651 | else | |
6a2c16d6 | 3652 | store_reg (i, offset, STACK_POINTER_REGNUM); |
6ec4380b | 3653 | offset += UNITS_PER_WORD; |
a9960cdc | 3654 | gr_saved++; |
87ad11b0 | 3655 | } |
daee63dd | 3656 | |
afd7b680 | 3657 | /* If we wanted to merge the SP adjustment with a GR save, but we never |
daee63dd | 3658 | did any GR saves, then just emit the adjustment here. */ |
201f01e9 | 3659 | if (merge_sp_adjust_with_store) |
6a2c16d6 | 3660 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3661 | actual_fsize, 1); |
87ad11b0 | 3662 | } |
6d36483b | 3663 | |
df6edefa | 3664 | /* The hppa calling conventions say that %r19, the pic offset |
3665 | register, is saved at sp - 32 (in this function's frame) | |
3666 | when generating PIC code. FIXME: What is the correct thing | |
3667 | to do for functions which make no calls and allocate no | |
3668 | frame? Do we need to allocate a frame, or can we just omit | |
8f177faf | 3669 | the save? For now we'll just omit the save. |
3670 | ||
3671 | We don't want a note on this insn as the frame marker can | |
3672 | move if there is a dynamic stack allocation. */ | |
df6edefa | 3673 | if (flag_pic && actual_fsize != 0 && !TARGET_64BIT) |
8f177faf | 3674 | { |
3675 | rtx addr = gen_rtx_PLUS (word_mode, stack_pointer_rtx, GEN_INT (-32)); | |
3676 | ||
3677 | emit_move_insn (gen_rtx_MEM (word_mode, addr), pic_offset_table_rtx); | |
3678 | ||
3679 | } | |
df6edefa | 3680 | |
87ad11b0 | 3681 | /* Align pointer properly (doubleword boundary). */ |
3682 | offset = (offset + 7) & ~7; | |
3683 | ||
3684 | /* Floating point register store. */ | |
3685 | if (save_fregs) | |
87ad11b0 | 3686 | { |
a584fe8a | 3687 | rtx base; |
3688 | ||
daee63dd | 3689 | /* First get the frame or stack pointer to the start of the FP register |
3690 | save area. */ | |
a1ab4fa3 | 3691 | if (frame_pointer_needed) |
a584fe8a | 3692 | { |
3693 | set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0); | |
3694 | base = frame_pointer_rtx; | |
3695 | } | |
a1ab4fa3 | 3696 | else |
a584fe8a | 3697 | { |
3698 | set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0); | |
3699 | base = stack_pointer_rtx; | |
3700 | } | |
daee63dd | 3701 | |
3702 | /* Now actually save the FP registers. */ | |
bac38c40 | 3703 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
7f7c4869 | 3704 | { |
5e3c5739 | 3705 | if (regs_ever_live[i] |
3706 | || (! TARGET_64BIT && regs_ever_live[i + 1])) | |
7f7c4869 | 3707 | { |
6a2c16d6 | 3708 | rtx addr, insn, reg; |
cc858176 | 3709 | addr = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg)); |
3710 | reg = gen_rtx_REG (DFmode, i); | |
6a2c16d6 | 3711 | insn = emit_move_insn (addr, reg); |
3712 | if (DO_FRAME_NOTES) | |
3713 | { | |
3714 | RTX_FRAME_RELATED_P (insn) = 1; | |
a584fe8a | 3715 | if (TARGET_64BIT) |
3716 | { | |
3717 | rtx mem = gen_rtx_MEM (DFmode, | |
3718 | plus_constant (base, offset)); | |
3719 | REG_NOTES (insn) | |
3720 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3721 | gen_rtx_SET (VOIDmode, mem, reg), | |
3722 | REG_NOTES (insn)); | |
3723 | } | |
3724 | else | |
3725 | { | |
3726 | rtx meml = gen_rtx_MEM (SFmode, | |
3727 | plus_constant (base, offset)); | |
3728 | rtx memr = gen_rtx_MEM (SFmode, | |
3729 | plus_constant (base, offset + 4)); | |
3730 | rtx regl = gen_rtx_REG (SFmode, i); | |
3731 | rtx regr = gen_rtx_REG (SFmode, i + 1); | |
3732 | rtx setl = gen_rtx_SET (VOIDmode, meml, regl); | |
3733 | rtx setr = gen_rtx_SET (VOIDmode, memr, regr); | |
3734 | rtvec vec; | |
3735 | ||
3736 | RTX_FRAME_RELATED_P (setl) = 1; | |
3737 | RTX_FRAME_RELATED_P (setr) = 1; | |
3738 | vec = gen_rtvec (2, setl, setr); | |
3739 | REG_NOTES (insn) | |
3740 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3741 | gen_rtx_SEQUENCE (VOIDmode, vec), | |
3742 | REG_NOTES (insn)); | |
3743 | } | |
6a2c16d6 | 3744 | } |
3745 | offset += GET_MODE_SIZE (DFmode); | |
7f7c4869 | 3746 | fr_saved++; |
3747 | } | |
3748 | } | |
87ad11b0 | 3749 | } |
3750 | } | |
3751 | ||
cc858176 | 3752 | /* Emit RTL to load REG from the memory location specified by BASE+DISP. |
3753 | Handle case where DISP > 8k by using the add_high_const patterns. */ | |
3754 | ||
6a2c16d6 | 3755 | static void |
5c1d8983 | 3756 | load_reg (int reg, int disp, int base) |
cc858176 | 3757 | { |
6a2c16d6 | 3758 | rtx src, dest, basereg; |
cc858176 | 3759 | |
3760 | dest = gen_rtx_REG (word_mode, reg); | |
3761 | basereg = gen_rtx_REG (Pmode, base); | |
3762 | if (VAL_14_BITS_P (disp)) | |
3763 | { | |
3764 | src = gen_rtx_MEM (word_mode, plus_constant (basereg, disp)); | |
6a2c16d6 | 3765 | emit_move_insn (dest, src); |
cc858176 | 3766 | } |
3767 | else | |
3768 | { | |
3769 | rtx delta = GEN_INT (disp); | |
3770 | rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta)); | |
3771 | rtx tmpreg = gen_rtx_REG (Pmode, 1); | |
3772 | emit_move_insn (tmpreg, high); | |
3773 | src = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta)); | |
6a2c16d6 | 3774 | emit_move_insn (dest, src); |
cc858176 | 3775 | } |
cc858176 | 3776 | } |
daee63dd | 3777 | |
2247cc5f | 3778 | /* Update the total code bytes output to the text section. */ |
3779 | ||
3780 | static void | |
5c1d8983 | 3781 | update_total_code_bytes (int nbytes) |
2247cc5f | 3782 | { |
3783 | if ((TARGET_PORTABLE_RUNTIME || !TARGET_GAS || !TARGET_SOM) | |
8a05c3c2 | 3784 | && !IN_NAMED_SECTION_P (cfun->decl)) |
2247cc5f | 3785 | { |
3786 | if (INSN_ADDRESSES_SET_P ()) | |
3787 | { | |
3788 | unsigned long old_total = total_code_bytes; | |
3789 | ||
3790 | total_code_bytes += nbytes; | |
3791 | ||
3792 | /* Be prepared to handle overflows. */ | |
3793 | if (old_total > total_code_bytes) | |
3794 | total_code_bytes = -1; | |
3795 | } | |
3796 | else | |
3797 | total_code_bytes = -1; | |
3798 | } | |
3799 | } | |
3800 | ||
17d9b0c3 | 3801 | /* This function generates the assembly code for function exit. |
3802 | Args are as for output_function_prologue (). | |
3803 | ||
3804 | The function epilogue should not depend on the current stack | |
3805 | pointer! It should use the frame pointer only. This is mandatory | |
3806 | because of alloca; we also take advantage of it to omit stack | |
6dc3b0d9 | 3807 | adjustments before returning. */ |
17d9b0c3 | 3808 | |
3809 | static void | |
5c1d8983 | 3810 | pa_output_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED) |
87ad11b0 | 3811 | { |
3695c664 | 3812 | rtx insn = get_last_insn (); |
3813 | ||
2247cc5f | 3814 | last_address = 0; |
3815 | ||
daee63dd | 3816 | /* hppa_expand_epilogue does the dirty work now. We just need |
3817 | to output the assembler directives which denote the end | |
3695c664 | 3818 | of a function. |
3819 | ||
3820 | To make debuggers happy, emit a nop if the epilogue was completely | |
3821 | eliminated due to a volatile call as the last insn in the | |
6d36483b | 3822 | current function. That way the return address (in %r2) will |
3695c664 | 3823 | always point to a valid instruction in the current function. */ |
3824 | ||
3825 | /* Get the last real insn. */ | |
3826 | if (GET_CODE (insn) == NOTE) | |
3827 | insn = prev_real_insn (insn); | |
3828 | ||
3829 | /* If it is a sequence, then look inside. */ | |
3830 | if (insn && GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) | |
3831 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
3832 | ||
6d36483b | 3833 | /* If insn is a CALL_INSN, then it must be a call to a volatile |
3695c664 | 3834 | function (otherwise there would be epilogue insns). */ |
3835 | if (insn && GET_CODE (insn) == CALL_INSN) | |
90c41894 | 3836 | { |
3837 | fputs ("\tnop\n", file); | |
3838 | last_address += 4; | |
3839 | } | |
6d36483b | 3840 | |
9c0ac0fd | 3841 | fputs ("\t.EXIT\n\t.PROCEND\n", file); |
90c41894 | 3842 | |
2247cc5f | 3843 | if (INSN_ADDRESSES_SET_P ()) |
90c41894 | 3844 | { |
2247cc5f | 3845 | insn = get_last_nonnote_insn (); |
3846 | last_address += INSN_ADDRESSES (INSN_UID (insn)); | |
3847 | if (INSN_P (insn)) | |
3848 | last_address += insn_default_length (insn); | |
3849 | last_address = ((last_address + FUNCTION_BOUNDARY / BITS_PER_UNIT - 1) | |
3850 | & ~(FUNCTION_BOUNDARY / BITS_PER_UNIT - 1)); | |
90c41894 | 3851 | } |
2247cc5f | 3852 | |
3853 | /* Finally, update the total number of code bytes output so far. */ | |
3854 | update_total_code_bytes (last_address); | |
daee63dd | 3855 | } |
afd7b680 | 3856 | |
daee63dd | 3857 | void |
5c1d8983 | 3858 | hppa_expand_epilogue (void) |
daee63dd | 3859 | { |
6d36483b | 3860 | rtx tmpreg; |
58361f39 | 3861 | int offset, i; |
3862 | int merge_sp_adjust_with_load = 0; | |
3863 | int ret_off = 0; | |
daee63dd | 3864 | |
3865 | /* We will use this often. */ | |
440c23df | 3866 | tmpreg = gen_rtx_REG (word_mode, 1); |
daee63dd | 3867 | |
3868 | /* Try to restore RP early to avoid load/use interlocks when | |
3869 | RP gets used in the return (bv) instruction. This appears to still | |
6dc3b0d9 | 3870 | be necessary even when we schedule the prologue and epilogue. */ |
a584fe8a | 3871 | if (regs_ever_live [2] || current_function_calls_eh_return) |
58361f39 | 3872 | { |
3873 | ret_off = TARGET_64BIT ? -16 : -20; | |
3874 | if (frame_pointer_needed) | |
3875 | { | |
6a2c16d6 | 3876 | load_reg (2, ret_off, FRAME_POINTER_REGNUM); |
58361f39 | 3877 | ret_off = 0; |
3878 | } | |
3879 | else | |
3880 | { | |
3881 | /* No frame pointer, and stack is smaller than 8k. */ | |
3882 | if (VAL_14_BITS_P (ret_off - actual_fsize)) | |
3883 | { | |
6a2c16d6 | 3884 | load_reg (2, ret_off - actual_fsize, STACK_POINTER_REGNUM); |
58361f39 | 3885 | ret_off = 0; |
3886 | } | |
3887 | } | |
3888 | } | |
daee63dd | 3889 | |
3890 | /* General register restores. */ | |
87ad11b0 | 3891 | if (frame_pointer_needed) |
3892 | { | |
a584fe8a | 3893 | offset = local_fsize; |
3894 | ||
3895 | /* If the current function calls __builtin_eh_return, then we need | |
3896 | to restore the saved EH data registers. */ | |
3897 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3898 | { | |
3899 | unsigned int i, regno; | |
3900 | ||
3901 | for (i = 0; ; ++i) | |
3902 | { | |
3903 | regno = EH_RETURN_DATA_REGNO (i); | |
3904 | if (regno == INVALID_REGNUM) | |
3905 | break; | |
3906 | ||
3907 | load_reg (regno, offset, FRAME_POINTER_REGNUM); | |
3908 | offset += UNITS_PER_WORD; | |
3909 | } | |
3910 | } | |
3911 | ||
3912 | for (i = 18; i >= 4; i--) | |
98328a39 | 3913 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3914 | { |
6a2c16d6 | 3915 | load_reg (i, offset, FRAME_POINTER_REGNUM); |
6ec4380b | 3916 | offset += UNITS_PER_WORD; |
87ad11b0 | 3917 | } |
87ad11b0 | 3918 | } |
3919 | else | |
3920 | { | |
a584fe8a | 3921 | offset = local_fsize - actual_fsize; |
3922 | ||
3923 | /* If the current function calls __builtin_eh_return, then we need | |
3924 | to restore the saved EH data registers. */ | |
3925 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3926 | { | |
3927 | unsigned int i, regno; | |
3928 | ||
3929 | for (i = 0; ; ++i) | |
3930 | { | |
3931 | regno = EH_RETURN_DATA_REGNO (i); | |
3932 | if (regno == INVALID_REGNUM) | |
3933 | break; | |
3934 | ||
3935 | /* Only for the first load. | |
3936 | merge_sp_adjust_with_load holds the register load | |
3937 | with which we will merge the sp adjustment. */ | |
3938 | if (merge_sp_adjust_with_load == 0 | |
3939 | && local_fsize == 0 | |
3940 | && VAL_14_BITS_P (-actual_fsize)) | |
3941 | merge_sp_adjust_with_load = regno; | |
3942 | else | |
3943 | load_reg (regno, offset, STACK_POINTER_REGNUM); | |
3944 | offset += UNITS_PER_WORD; | |
3945 | } | |
3946 | } | |
3947 | ||
3948 | for (i = 18; i >= 3; i--) | |
7f7c4869 | 3949 | { |
98328a39 | 3950 | if (regs_ever_live[i] && ! call_used_regs[i]) |
7f7c4869 | 3951 | { |
7f7c4869 | 3952 | /* Only for the first load. |
3953 | merge_sp_adjust_with_load holds the register load | |
3954 | with which we will merge the sp adjustment. */ | |
58361f39 | 3955 | if (merge_sp_adjust_with_load == 0 |
7f7c4869 | 3956 | && local_fsize == 0 |
58361f39 | 3957 | && VAL_14_BITS_P (-actual_fsize)) |
7f7c4869 | 3958 | merge_sp_adjust_with_load = i; |
3959 | else | |
6a2c16d6 | 3960 | load_reg (i, offset, STACK_POINTER_REGNUM); |
6ec4380b | 3961 | offset += UNITS_PER_WORD; |
7f7c4869 | 3962 | } |
3963 | } | |
87ad11b0 | 3964 | } |
daee63dd | 3965 | |
87ad11b0 | 3966 | /* Align pointer properly (doubleword boundary). */ |
3967 | offset = (offset + 7) & ~7; | |
3968 | ||
daee63dd | 3969 | /* FP register restores. */ |
87ad11b0 | 3970 | if (save_fregs) |
87ad11b0 | 3971 | { |
daee63dd | 3972 | /* Adjust the register to index off of. */ |
a1ab4fa3 | 3973 | if (frame_pointer_needed) |
a584fe8a | 3974 | set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0); |
a1ab4fa3 | 3975 | else |
a584fe8a | 3976 | set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0); |
daee63dd | 3977 | |
3978 | /* Actually do the restores now. */ | |
bac38c40 | 3979 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
cc858176 | 3980 | if (regs_ever_live[i] |
3981 | || (! TARGET_64BIT && regs_ever_live[i + 1])) | |
3982 | { | |
3983 | rtx src = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg)); | |
3984 | rtx dest = gen_rtx_REG (DFmode, i); | |
6a2c16d6 | 3985 | emit_move_insn (dest, src); |
cc858176 | 3986 | } |
87ad11b0 | 3987 | } |
daee63dd | 3988 | |
14660146 | 3989 | /* Emit a blockage insn here to keep these insns from being moved to |
3990 | an earlier spot in the epilogue, or into the main instruction stream. | |
3991 | ||
3992 | This is necessary as we must not cut the stack back before all the | |
3993 | restores are finished. */ | |
3994 | emit_insn (gen_blockage ()); | |
daee63dd | 3995 | |
9840d99d | 3996 | /* Reset stack pointer (and possibly frame pointer). The stack |
42819d4e | 3997 | pointer is initially set to fp + 64 to avoid a race condition. */ |
58361f39 | 3998 | if (frame_pointer_needed) |
87ad11b0 | 3999 | { |
cc858176 | 4000 | rtx delta = GEN_INT (-64); |
a584fe8a | 4001 | |
4002 | set_reg_plus_d (STACK_POINTER_REGNUM, FRAME_POINTER_REGNUM, 64, 0); | |
4003 | emit_insn (gen_pre_load (frame_pointer_rtx, stack_pointer_rtx, delta)); | |
87ad11b0 | 4004 | } |
daee63dd | 4005 | /* If we were deferring a callee register restore, do it now. */ |
58361f39 | 4006 | else if (merge_sp_adjust_with_load) |
4007 | { | |
4008 | rtx delta = GEN_INT (-actual_fsize); | |
cc858176 | 4009 | rtx dest = gen_rtx_REG (word_mode, merge_sp_adjust_with_load); |
a584fe8a | 4010 | |
4011 | emit_insn (gen_pre_load (dest, stack_pointer_rtx, delta)); | |
58361f39 | 4012 | } |
daee63dd | 4013 | else if (actual_fsize != 0) |
a584fe8a | 4014 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
4015 | - actual_fsize, 0); | |
58361f39 | 4016 | |
4017 | /* If we haven't restored %r2 yet (no frame pointer, and a stack | |
4018 | frame greater than 8k), do so now. */ | |
4019 | if (ret_off != 0) | |
6a2c16d6 | 4020 | load_reg (2, ret_off, STACK_POINTER_REGNUM); |
a584fe8a | 4021 | |
4022 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
4023 | { | |
4024 | rtx sa = EH_RETURN_STACKADJ_RTX; | |
4025 | ||
4026 | emit_insn (gen_blockage ()); | |
4027 | emit_insn (TARGET_64BIT | |
4028 | ? gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, sa) | |
4029 | : gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, sa)); | |
4030 | } | |
87ad11b0 | 4031 | } |
4032 | ||
d7e2f694 | 4033 | rtx |
5c1d8983 | 4034 | hppa_pic_save_rtx (void) |
cf3de5bb | 4035 | { |
d7e2f694 | 4036 | return get_hard_reg_initial_val (word_mode, PIC_OFFSET_TABLE_REGNUM); |
df6edefa | 4037 | } |
4038 | ||
4039 | void | |
5c1d8983 | 4040 | hppa_profile_hook (int label_no) |
df6edefa | 4041 | { |
4da37e2f | 4042 | /* We use SImode for the address of the function in both 32 and |
4043 | 64-bit code to avoid having to provide DImode versions of the | |
4044 | lcla2 and load_offset_label_address insn patterns. */ | |
4045 | rtx reg = gen_reg_rtx (SImode); | |
4046 | rtx label_rtx = gen_label_rtx (); | |
a9ac13e4 | 4047 | rtx begin_label_rtx, call_insn; |
4048 | char begin_label_name[16]; | |
df6edefa | 4049 | |
a9ac13e4 | 4050 | ASM_GENERATE_INTERNAL_LABEL (begin_label_name, FUNC_BEGIN_PROLOG_LABEL, |
b8a21949 | 4051 | label_no); |
4da37e2f | 4052 | begin_label_rtx = gen_rtx_SYMBOL_REF (SImode, ggc_strdup (begin_label_name)); |
df6edefa | 4053 | |
4054 | if (TARGET_64BIT) | |
4055 | emit_move_insn (arg_pointer_rtx, | |
4056 | gen_rtx_PLUS (word_mode, virtual_outgoing_args_rtx, | |
4057 | GEN_INT (64))); | |
4058 | ||
df6edefa | 4059 | emit_move_insn (gen_rtx_REG (word_mode, 26), gen_rtx_REG (word_mode, 2)); |
4060 | ||
4da37e2f | 4061 | /* The address of the function is loaded into %r25 with a instruction- |
4062 | relative sequence that avoids the use of relocations. The sequence | |
4063 | is split so that the load_offset_label_address instruction can | |
4064 | occupy the delay slot of the call to _mcount. */ | |
4065 | if (TARGET_PA_20) | |
4066 | emit_insn (gen_lcla2 (reg, label_rtx)); | |
4067 | else | |
4068 | emit_insn (gen_lcla1 (reg, label_rtx)); | |
4069 | ||
4070 | emit_insn (gen_load_offset_label_address (gen_rtx_REG (SImode, 25), | |
4071 | reg, begin_label_rtx, label_rtx)); | |
4072 | ||
df6edefa | 4073 | #ifndef NO_PROFILE_COUNTERS |
4074 | { | |
4075 | rtx count_label_rtx, addr, r24; | |
a9ac13e4 | 4076 | char count_label_name[16]; |
df6edefa | 4077 | |
a9ac13e4 | 4078 | ASM_GENERATE_INTERNAL_LABEL (count_label_name, "LP", label_no); |
4079 | count_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (count_label_name)); | |
df6edefa | 4080 | |
831a12d9 | 4081 | addr = force_reg (Pmode, count_label_rtx); |
df6edefa | 4082 | r24 = gen_rtx_REG (Pmode, 24); |
4083 | emit_move_insn (r24, addr); | |
4084 | ||
df6edefa | 4085 | call_insn = |
4da37e2f | 4086 | emit_call_insn (gen_call (gen_rtx_MEM (Pmode, |
4087 | gen_rtx_SYMBOL_REF (Pmode, | |
4088 | "_mcount")), | |
4089 | GEN_INT (TARGET_64BIT ? 24 : 12))); | |
df6edefa | 4090 | |
4091 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), r24); | |
4092 | } | |
4093 | #else | |
4da37e2f | 4094 | |
df6edefa | 4095 | call_insn = |
4da37e2f | 4096 | emit_call_insn (gen_call (gen_rtx_MEM (Pmode, |
4097 | gen_rtx_SYMBOL_REF (Pmode, | |
4098 | "_mcount")), | |
4099 | GEN_INT (TARGET_64BIT ? 16 : 8))); | |
4100 | ||
df6edefa | 4101 | #endif |
4102 | ||
4da37e2f | 4103 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 25)); |
4104 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 26)); | |
4105 | ||
df6edefa | 4106 | /* Indicate the _mcount call cannot throw, nor will it execute a |
4107 | non-local goto. */ | |
4108 | REG_NOTES (call_insn) | |
4109 | = gen_rtx_EXPR_LIST (REG_EH_REGION, constm1_rtx, REG_NOTES (call_insn)); | |
cf3de5bb | 4110 | } |
4111 | ||
e07ff380 | 4112 | /* Fetch the return address for the frame COUNT steps up from |
4113 | the current frame, after the prologue. FRAMEADDR is the | |
4114 | frame pointer of the COUNT frame. | |
4115 | ||
f49b2e77 | 4116 | We want to ignore any export stub remnants here. To handle this, |
4117 | we examine the code at the return address, and if it is an export | |
4118 | stub, we return a memory rtx for the stub return address stored | |
4119 | at frame-24. | |
a6c6fd6c | 4120 | |
4121 | The value returned is used in two different ways: | |
4122 | ||
4123 | 1. To find a function's caller. | |
4124 | ||
4125 | 2. To change the return address for a function. | |
4126 | ||
4127 | This function handles most instances of case 1; however, it will | |
4128 | fail if there are two levels of stubs to execute on the return | |
4129 | path. The only way I believe that can happen is if the return value | |
4130 | needs a parameter relocation, which never happens for C code. | |
4131 | ||
4132 | This function handles most instances of case 2; however, it will | |
4133 | fail if we did not originally have stub code on the return path | |
f49b2e77 | 4134 | but will need stub code on the new return path. This can happen if |
a6c6fd6c | 4135 | the caller & callee are both in the main program, but the new |
f49b2e77 | 4136 | return location is in a shared library. */ |
e07ff380 | 4137 | |
4138 | rtx | |
5c1d8983 | 4139 | return_addr_rtx (int count, rtx frameaddr) |
e07ff380 | 4140 | { |
4141 | rtx label; | |
f49b2e77 | 4142 | rtx rp; |
e07ff380 | 4143 | rtx saved_rp; |
4144 | rtx ins; | |
4145 | ||
f49b2e77 | 4146 | if (count != 0) |
4147 | return NULL_RTX; | |
b29897dd | 4148 | |
f49b2e77 | 4149 | rp = get_hard_reg_initial_val (Pmode, 2); |
e07ff380 | 4150 | |
f49b2e77 | 4151 | if (TARGET_64BIT || TARGET_NO_SPACE_REGS) |
4152 | return rp; | |
e07ff380 | 4153 | |
b29897dd | 4154 | saved_rp = gen_reg_rtx (Pmode); |
f49b2e77 | 4155 | emit_move_insn (saved_rp, rp); |
e07ff380 | 4156 | |
4157 | /* Get pointer to the instruction stream. We have to mask out the | |
4158 | privilege level from the two low order bits of the return address | |
4159 | pointer here so that ins will point to the start of the first | |
4160 | instruction that would have been executed if we returned. */ | |
f49b2e77 | 4161 | ins = copy_to_reg (gen_rtx_AND (Pmode, rp, MASK_RETURN_ADDR)); |
e07ff380 | 4162 | label = gen_label_rtx (); |
4163 | ||
4164 | /* Check the instruction stream at the normal return address for the | |
4165 | export stub: | |
4166 | ||
4167 | 0x4bc23fd1 | stub+8: ldw -18(sr0,sp),rp | |
4168 | 0x004010a1 | stub+12: ldsid (sr0,rp),r1 | |
4169 | 0x00011820 | stub+16: mtsp r1,sr0 | |
4170 | 0xe0400002 | stub+20: be,n 0(sr0,rp) | |
4171 | ||
4172 | If it is an export stub, than our return address is really in | |
4173 | -24[frameaddr]. */ | |
4174 | ||
e00b80b2 | 4175 | emit_cmp_insn (gen_rtx_MEM (SImode, ins), GEN_INT (0x4bc23fd1), NE, |
4176 | NULL_RTX, SImode, 1); | |
e07ff380 | 4177 | emit_jump_insn (gen_bne (label)); |
4178 | ||
ad851752 | 4179 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 4)), |
e00b80b2 | 4180 | GEN_INT (0x004010a1), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4181 | emit_jump_insn (gen_bne (label)); |
4182 | ||
ad851752 | 4183 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 8)), |
e00b80b2 | 4184 | GEN_INT (0x00011820), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4185 | emit_jump_insn (gen_bne (label)); |
4186 | ||
ad851752 | 4187 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 12)), |
e00b80b2 | 4188 | GEN_INT (0xe0400002), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4189 | |
f49b2e77 | 4190 | /* If there is no export stub then just use the value saved from |
4191 | the return pointer register. */ | |
e07ff380 | 4192 | |
4193 | emit_jump_insn (gen_bne (label)); | |
4194 | ||
f49b2e77 | 4195 | /* Here we know that our return address points to an export |
e07ff380 | 4196 | stub. We don't want to return the address of the export stub, |
f49b2e77 | 4197 | but rather the return address of the export stub. That return |
4198 | address is stored at -24[frameaddr]. */ | |
e07ff380 | 4199 | |
f49b2e77 | 4200 | emit_move_insn (saved_rp, |
4201 | gen_rtx_MEM (Pmode, | |
4202 | memory_address (Pmode, | |
4203 | plus_constant (frameaddr, | |
4204 | -24)))); | |
e07ff380 | 4205 | |
4206 | emit_label (label); | |
f49b2e77 | 4207 | return saved_rp; |
e07ff380 | 4208 | } |
4209 | ||
757d4970 | 4210 | /* This is only valid once reload has completed because it depends on |
4211 | knowing exactly how much (if any) frame there is and... | |
4212 | ||
4213 | It's only valid if there is no frame marker to de-allocate and... | |
4214 | ||
4215 | It's only valid if %r2 hasn't been saved into the caller's frame | |
4216 | (we're not profiling and %r2 isn't live anywhere). */ | |
4217 | int | |
5c1d8983 | 4218 | hppa_can_use_return_insn_p (void) |
757d4970 | 4219 | { |
4220 | return (reload_completed | |
4221 | && (compute_frame_size (get_frame_size (), 0) ? 0 : 1) | |
757d4970 | 4222 | && ! regs_ever_live[2] |
4223 | && ! frame_pointer_needed); | |
4224 | } | |
4225 | ||
87ad11b0 | 4226 | void |
5c1d8983 | 4227 | emit_bcond_fp (enum rtx_code code, rtx operand0) |
87ad11b0 | 4228 | { |
ad851752 | 4229 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, |
4230 | gen_rtx_IF_THEN_ELSE (VOIDmode, | |
4231 | gen_rtx_fmt_ee (code, | |
4232 | VOIDmode, | |
4233 | gen_rtx_REG (CCFPmode, 0), | |
4234 | const0_rtx), | |
4235 | gen_rtx_LABEL_REF (VOIDmode, operand0), | |
4236 | pc_rtx))); | |
87ad11b0 | 4237 | |
4238 | } | |
4239 | ||
4240 | rtx | |
5c1d8983 | 4241 | gen_cmp_fp (enum rtx_code code, rtx operand0, rtx operand1) |
87ad11b0 | 4242 | { |
ad851752 | 4243 | return gen_rtx_SET (VOIDmode, gen_rtx_REG (CCFPmode, 0), |
4244 | gen_rtx_fmt_ee (code, CCFPmode, operand0, operand1)); | |
87ad11b0 | 4245 | } |
4246 | ||
8b49b3c7 | 4247 | /* Adjust the cost of a scheduling dependency. Return the new cost of |
4248 | a dependency LINK or INSN on DEP_INSN. COST is the current cost. */ | |
4249 | ||
747af5e7 | 4250 | static int |
5c1d8983 | 4251 | pa_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost) |
8b49b3c7 | 4252 | { |
43048d2c | 4253 | enum attr_type attr_type; |
4254 | ||
cde3e16c | 4255 | /* Don't adjust costs for a pa8000 chip, also do not adjust any |
4256 | true dependencies as they are described with bypasses now. */ | |
4257 | if (pa_cpu >= PROCESSOR_8000 || REG_NOTE_KIND (link) == 0) | |
342aabd9 | 4258 | return cost; |
4259 | ||
d402da4b | 4260 | if (! recog_memoized (insn)) |
4261 | return 0; | |
8b49b3c7 | 4262 | |
43048d2c | 4263 | attr_type = get_attr_type (insn); |
4264 | ||
cde3e16c | 4265 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI) |
8b49b3c7 | 4266 | { |
4267 | /* Anti dependency; DEP_INSN reads a register that INSN writes some | |
4268 | cycles later. */ | |
4269 | ||
43048d2c | 4270 | if (attr_type == TYPE_FPLOAD) |
8b49b3c7 | 4271 | { |
d402da4b | 4272 | rtx pat = PATTERN (insn); |
4273 | rtx dep_pat = PATTERN (dep_insn); | |
4274 | if (GET_CODE (pat) == PARALLEL) | |
4275 | { | |
4276 | /* This happens for the fldXs,mb patterns. */ | |
4277 | pat = XVECEXP (pat, 0, 0); | |
4278 | } | |
4279 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
8b49b3c7 | 4280 | /* If this happens, we have to extend this to schedule |
d402da4b | 4281 | optimally. Return 0 for now. */ |
4282 | return 0; | |
8b49b3c7 | 4283 | |
d402da4b | 4284 | if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat))) |
8b49b3c7 | 4285 | { |
d402da4b | 4286 | if (! recog_memoized (dep_insn)) |
4287 | return 0; | |
8b49b3c7 | 4288 | switch (get_attr_type (dep_insn)) |
4289 | { | |
4290 | case TYPE_FPALU: | |
134b4858 | 4291 | case TYPE_FPMULSGL: |
4292 | case TYPE_FPMULDBL: | |
8b49b3c7 | 4293 | case TYPE_FPDIVSGL: |
4294 | case TYPE_FPDIVDBL: | |
4295 | case TYPE_FPSQRTSGL: | |
4296 | case TYPE_FPSQRTDBL: | |
d402da4b | 4297 | /* A fpload can't be issued until one cycle before a |
01cc3b75 | 4298 | preceding arithmetic operation has finished if |
d402da4b | 4299 | the target of the fpload is any of the sources |
4300 | (or destination) of the arithmetic operation. */ | |
cde3e16c | 4301 | return insn_default_latency (dep_insn) - 1; |
134b4858 | 4302 | |
4303 | default: | |
4304 | return 0; | |
4305 | } | |
4306 | } | |
4307 | } | |
43048d2c | 4308 | else if (attr_type == TYPE_FPALU) |
134b4858 | 4309 | { |
4310 | rtx pat = PATTERN (insn); | |
4311 | rtx dep_pat = PATTERN (dep_insn); | |
4312 | if (GET_CODE (pat) == PARALLEL) | |
4313 | { | |
4314 | /* This happens for the fldXs,mb patterns. */ | |
4315 | pat = XVECEXP (pat, 0, 0); | |
4316 | } | |
4317 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4318 | /* If this happens, we have to extend this to schedule | |
4319 | optimally. Return 0 for now. */ | |
4320 | return 0; | |
4321 | ||
4322 | if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat))) | |
4323 | { | |
4324 | if (! recog_memoized (dep_insn)) | |
4325 | return 0; | |
4326 | switch (get_attr_type (dep_insn)) | |
4327 | { | |
4328 | case TYPE_FPDIVSGL: | |
4329 | case TYPE_FPDIVDBL: | |
4330 | case TYPE_FPSQRTSGL: | |
4331 | case TYPE_FPSQRTDBL: | |
4332 | /* An ALU flop can't be issued until two cycles before a | |
01cc3b75 | 4333 | preceding divide or sqrt operation has finished if |
134b4858 | 4334 | the target of the ALU flop is any of the sources |
4335 | (or destination) of the divide or sqrt operation. */ | |
cde3e16c | 4336 | return insn_default_latency (dep_insn) - 2; |
8b49b3c7 | 4337 | |
4338 | default: | |
4339 | return 0; | |
4340 | } | |
4341 | } | |
4342 | } | |
4343 | ||
4344 | /* For other anti dependencies, the cost is 0. */ | |
4345 | return 0; | |
4346 | } | |
134b4858 | 4347 | else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT) |
4348 | { | |
4349 | /* Output dependency; DEP_INSN writes a register that INSN writes some | |
4350 | cycles later. */ | |
43048d2c | 4351 | if (attr_type == TYPE_FPLOAD) |
134b4858 | 4352 | { |
4353 | rtx pat = PATTERN (insn); | |
4354 | rtx dep_pat = PATTERN (dep_insn); | |
4355 | if (GET_CODE (pat) == PARALLEL) | |
4356 | { | |
4357 | /* This happens for the fldXs,mb patterns. */ | |
4358 | pat = XVECEXP (pat, 0, 0); | |
4359 | } | |
4360 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4361 | /* If this happens, we have to extend this to schedule | |
4362 | optimally. Return 0 for now. */ | |
4363 | return 0; | |
4364 | ||
4365 | if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat))) | |
4366 | { | |
4367 | if (! recog_memoized (dep_insn)) | |
4368 | return 0; | |
4369 | switch (get_attr_type (dep_insn)) | |
4370 | { | |
4371 | case TYPE_FPALU: | |
4372 | case TYPE_FPMULSGL: | |
4373 | case TYPE_FPMULDBL: | |
4374 | case TYPE_FPDIVSGL: | |
4375 | case TYPE_FPDIVDBL: | |
4376 | case TYPE_FPSQRTSGL: | |
4377 | case TYPE_FPSQRTDBL: | |
4378 | /* A fpload can't be issued until one cycle before a | |
01cc3b75 | 4379 | preceding arithmetic operation has finished if |
134b4858 | 4380 | the target of the fpload is the destination of the |
bea4bad2 | 4381 | arithmetic operation. |
4382 | ||
4383 | Exception: For PA7100LC, PA7200 and PA7300, the cost | |
4384 | is 3 cycles, unless they bundle together. We also | |
4385 | pay the penalty if the second insn is a fpload. */ | |
cde3e16c | 4386 | return insn_default_latency (dep_insn) - 1; |
8b49b3c7 | 4387 | |
134b4858 | 4388 | default: |
4389 | return 0; | |
4390 | } | |
4391 | } | |
4392 | } | |
43048d2c | 4393 | else if (attr_type == TYPE_FPALU) |
134b4858 | 4394 | { |
4395 | rtx pat = PATTERN (insn); | |
4396 | rtx dep_pat = PATTERN (dep_insn); | |
4397 | if (GET_CODE (pat) == PARALLEL) | |
4398 | { | |
4399 | /* This happens for the fldXs,mb patterns. */ | |
4400 | pat = XVECEXP (pat, 0, 0); | |
4401 | } | |
4402 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4403 | /* If this happens, we have to extend this to schedule | |
4404 | optimally. Return 0 for now. */ | |
4405 | return 0; | |
4406 | ||
4407 | if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat))) | |
4408 | { | |
4409 | if (! recog_memoized (dep_insn)) | |
4410 | return 0; | |
4411 | switch (get_attr_type (dep_insn)) | |
4412 | { | |
4413 | case TYPE_FPDIVSGL: | |
4414 | case TYPE_FPDIVDBL: | |
4415 | case TYPE_FPSQRTSGL: | |
4416 | case TYPE_FPSQRTDBL: | |
4417 | /* An ALU flop can't be issued until two cycles before a | |
01cc3b75 | 4418 | preceding divide or sqrt operation has finished if |
134b4858 | 4419 | the target of the ALU flop is also the target of |
3398e91d | 4420 | the divide or sqrt operation. */ |
cde3e16c | 4421 | return insn_default_latency (dep_insn) - 2; |
134b4858 | 4422 | |
4423 | default: | |
4424 | return 0; | |
4425 | } | |
4426 | } | |
4427 | } | |
4428 | ||
4429 | /* For other output dependencies, the cost is 0. */ | |
4430 | return 0; | |
4431 | } | |
4432 | else | |
4433 | abort (); | |
8b49b3c7 | 4434 | } |
87ad11b0 | 4435 | |
747af5e7 | 4436 | /* Adjust scheduling priorities. We use this to try and keep addil |
4437 | and the next use of %r1 close together. */ | |
4438 | static int | |
5c1d8983 | 4439 | pa_adjust_priority (rtx insn, int priority) |
747af5e7 | 4440 | { |
4441 | rtx set = single_set (insn); | |
4442 | rtx src, dest; | |
4443 | if (set) | |
4444 | { | |
4445 | src = SET_SRC (set); | |
4446 | dest = SET_DEST (set); | |
4447 | if (GET_CODE (src) == LO_SUM | |
4448 | && symbolic_operand (XEXP (src, 1), VOIDmode) | |
4449 | && ! read_only_operand (XEXP (src, 1), VOIDmode)) | |
4450 | priority >>= 3; | |
4451 | ||
4452 | else if (GET_CODE (src) == MEM | |
4453 | && GET_CODE (XEXP (src, 0)) == LO_SUM | |
4454 | && symbolic_operand (XEXP (XEXP (src, 0), 1), VOIDmode) | |
4455 | && ! read_only_operand (XEXP (XEXP (src, 0), 1), VOIDmode)) | |
4456 | priority >>= 1; | |
4457 | ||
4458 | else if (GET_CODE (dest) == MEM | |
4459 | && GET_CODE (XEXP (dest, 0)) == LO_SUM | |
4460 | && symbolic_operand (XEXP (XEXP (dest, 0), 1), VOIDmode) | |
4461 | && ! read_only_operand (XEXP (XEXP (dest, 0), 1), VOIDmode)) | |
4462 | priority >>= 3; | |
4463 | } | |
4464 | return priority; | |
4465 | } | |
4466 | ||
4467 | /* The 700 can only issue a single insn at a time. | |
4468 | The 7XXX processors can issue two insns at a time. | |
4469 | The 8000 can issue 4 insns at a time. */ | |
4470 | static int | |
5c1d8983 | 4471 | pa_issue_rate (void) |
747af5e7 | 4472 | { |
4473 | switch (pa_cpu) | |
4474 | { | |
4475 | case PROCESSOR_700: return 1; | |
4476 | case PROCESSOR_7100: return 2; | |
4477 | case PROCESSOR_7100LC: return 2; | |
4478 | case PROCESSOR_7200: return 2; | |
bea4bad2 | 4479 | case PROCESSOR_7300: return 2; |
747af5e7 | 4480 | case PROCESSOR_8000: return 4; |
4481 | ||
4482 | default: | |
4483 | abort (); | |
4484 | } | |
4485 | } | |
4486 | ||
4487 | ||
4488 | ||
58e17b0b | 4489 | /* Return any length adjustment needed by INSN which already has its length |
6d36483b | 4490 | computed as LENGTH. Return zero if no adjustment is necessary. |
58e17b0b | 4491 | |
5fbd5940 | 4492 | For the PA: function calls, millicode calls, and backwards short |
6d36483b | 4493 | conditional branches with unfilled delay slots need an adjustment by +1 |
5fbd5940 | 4494 | (to account for the NOP which will be inserted into the instruction stream). |
58e17b0b | 4495 | |
4496 | Also compute the length of an inline block move here as it is too | |
5fbd5940 | 4497 | complicated to express as a length attribute in pa.md. */ |
58e17b0b | 4498 | int |
5c1d8983 | 4499 | pa_adjust_insn_length (rtx insn, int length) |
58e17b0b | 4500 | { |
4501 | rtx pat = PATTERN (insn); | |
4502 | ||
faf3f8c1 | 4503 | /* Jumps inside switch tables which have unfilled delay slots need |
4504 | adjustment. */ | |
4505 | if (GET_CODE (insn) == JUMP_INSN | |
b932f66c | 4506 | && GET_CODE (pat) == PARALLEL |
4507 | && get_attr_type (insn) == TYPE_BTABLE_BRANCH) | |
3b1e673e | 4508 | return 4; |
58e17b0b | 4509 | /* Millicode insn with an unfilled delay slot. */ |
4510 | else if (GET_CODE (insn) == INSN | |
4511 | && GET_CODE (pat) != SEQUENCE | |
4512 | && GET_CODE (pat) != USE | |
4513 | && GET_CODE (pat) != CLOBBER | |
4514 | && get_attr_type (insn) == TYPE_MILLI) | |
5a1231ef | 4515 | return 4; |
58e17b0b | 4516 | /* Block move pattern. */ |
4517 | else if (GET_CODE (insn) == INSN | |
4518 | && GET_CODE (pat) == PARALLEL | |
f2ebcf32 | 4519 | && GET_CODE (XVECEXP (pat, 0, 0)) == SET |
58e17b0b | 4520 | && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM |
4521 | && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 1)) == MEM | |
4522 | && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode | |
4523 | && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 1)) == BLKmode) | |
a7e1bb24 | 4524 | return compute_movstr_length (insn) - 4; |
4525 | /* Block clear pattern. */ | |
4526 | else if (GET_CODE (insn) == INSN | |
4527 | && GET_CODE (pat) == PARALLEL | |
4528 | && GET_CODE (XVECEXP (pat, 0, 0)) == SET | |
4529 | && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM | |
4530 | && XEXP (XVECEXP (pat, 0, 0), 1) == const0_rtx | |
4531 | && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode) | |
4532 | return compute_clrstr_length (insn) - 4; | |
58e17b0b | 4533 | /* Conditional branch with an unfilled delay slot. */ |
5fbd5940 | 4534 | else if (GET_CODE (insn) == JUMP_INSN && ! simplejump_p (insn)) |
4535 | { | |
4536 | /* Adjust a short backwards conditional with an unfilled delay slot. */ | |
4537 | if (GET_CODE (pat) == SET | |
5a1231ef | 4538 | && length == 4 |
5fbd5940 | 4539 | && ! forward_branch_p (insn)) |
5a1231ef | 4540 | return 4; |
546a40bd | 4541 | else if (GET_CODE (pat) == PARALLEL |
4542 | && get_attr_type (insn) == TYPE_PARALLEL_BRANCH | |
4543 | && length == 4) | |
4544 | return 4; | |
5fbd5940 | 4545 | /* Adjust dbra insn with short backwards conditional branch with |
6d36483b | 4546 | unfilled delay slot -- only for case where counter is in a |
6dc3b0d9 | 4547 | general register register. */ |
5fbd5940 | 4548 | else if (GET_CODE (pat) == PARALLEL |
4549 | && GET_CODE (XVECEXP (pat, 0, 1)) == SET | |
4550 | && GET_CODE (XEXP (XVECEXP (pat, 0, 1), 0)) == REG | |
6d36483b | 4551 | && ! FP_REG_P (XEXP (XVECEXP (pat, 0, 1), 0)) |
5a1231ef | 4552 | && length == 4 |
5fbd5940 | 4553 | && ! forward_branch_p (insn)) |
5a1231ef | 4554 | return 4; |
5fbd5940 | 4555 | else |
4556 | return 0; | |
4557 | } | |
546a40bd | 4558 | return 0; |
58e17b0b | 4559 | } |
4560 | ||
87ad11b0 | 4561 | /* Print operand X (an rtx) in assembler syntax to file FILE. |
4562 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
4563 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
4564 | ||
4565 | void | |
5c1d8983 | 4566 | print_operand (FILE *file, rtx x, int code) |
87ad11b0 | 4567 | { |
4568 | switch (code) | |
4569 | { | |
4570 | case '#': | |
4571 | /* Output a 'nop' if there's nothing for the delay slot. */ | |
4572 | if (dbr_sequence_length () == 0) | |
4573 | fputs ("\n\tnop", file); | |
4574 | return; | |
4575 | case '*': | |
87fcb603 | 4576 | /* Output a nullification completer if there's nothing for the */ |
6d36483b | 4577 | /* delay slot or nullification is requested. */ |
87ad11b0 | 4578 | if (dbr_sequence_length () == 0 || |
4579 | (final_sequence && | |
4580 | INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)))) | |
4581 | fputs (",n", file); | |
4582 | return; | |
4583 | case 'R': | |
4584 | /* Print out the second register name of a register pair. | |
4585 | I.e., R (6) => 7. */ | |
ea52c577 | 4586 | fputs (reg_names[REGNO (x) + 1], file); |
87ad11b0 | 4587 | return; |
4588 | case 'r': | |
6dc3b0d9 | 4589 | /* A register or zero. */ |
891b55b4 | 4590 | if (x == const0_rtx |
4591 | || (x == CONST0_RTX (DFmode)) | |
4592 | || (x == CONST0_RTX (SFmode))) | |
87ad11b0 | 4593 | { |
c6ae275c | 4594 | fputs ("%r0", file); |
4595 | return; | |
4596 | } | |
4597 | else | |
4598 | break; | |
4599 | case 'f': | |
6dc3b0d9 | 4600 | /* A register or zero (floating point). */ |
c6ae275c | 4601 | if (x == const0_rtx |
4602 | || (x == CONST0_RTX (DFmode)) | |
4603 | || (x == CONST0_RTX (SFmode))) | |
4604 | { | |
4605 | fputs ("%fr0", file); | |
87ad11b0 | 4606 | return; |
4607 | } | |
4608 | else | |
4609 | break; | |
2d14b1f0 | 4610 | case 'A': |
4611 | { | |
4612 | rtx xoperands[2]; | |
4613 | ||
4614 | xoperands[0] = XEXP (XEXP (x, 0), 0); | |
4615 | xoperands[1] = XVECEXP (XEXP (XEXP (x, 0), 1), 0, 0); | |
4616 | output_global_address (file, xoperands[1], 0); | |
4617 | fprintf (file, "(%s)", reg_names [REGNO (xoperands[0])]); | |
4618 | return; | |
4619 | } | |
4620 | ||
c8975385 | 4621 | case 'C': /* Plain (C)ondition */ |
87ad11b0 | 4622 | case 'X': |
4623 | switch (GET_CODE (x)) | |
6d36483b | 4624 | { |
87ad11b0 | 4625 | case EQ: |
9c0ac0fd | 4626 | fputs ("=", file); break; |
87ad11b0 | 4627 | case NE: |
9c0ac0fd | 4628 | fputs ("<>", file); break; |
87ad11b0 | 4629 | case GT: |
9c0ac0fd | 4630 | fputs (">", file); break; |
87ad11b0 | 4631 | case GE: |
9c0ac0fd | 4632 | fputs (">=", file); break; |
87ad11b0 | 4633 | case GEU: |
9c0ac0fd | 4634 | fputs (">>=", file); break; |
87ad11b0 | 4635 | case GTU: |
9c0ac0fd | 4636 | fputs (">>", file); break; |
87ad11b0 | 4637 | case LT: |
9c0ac0fd | 4638 | fputs ("<", file); break; |
87ad11b0 | 4639 | case LE: |
9c0ac0fd | 4640 | fputs ("<=", file); break; |
87ad11b0 | 4641 | case LEU: |
9c0ac0fd | 4642 | fputs ("<<=", file); break; |
87ad11b0 | 4643 | case LTU: |
9c0ac0fd | 4644 | fputs ("<<", file); break; |
87ad11b0 | 4645 | default: |
87ad11b0 | 4646 | abort (); |
4647 | } | |
4648 | return; | |
c8975385 | 4649 | case 'N': /* Condition, (N)egated */ |
87ad11b0 | 4650 | switch (GET_CODE (x)) |
4651 | { | |
4652 | case EQ: | |
9c0ac0fd | 4653 | fputs ("<>", file); break; |
87ad11b0 | 4654 | case NE: |
9c0ac0fd | 4655 | fputs ("=", file); break; |
87ad11b0 | 4656 | case GT: |
9c0ac0fd | 4657 | fputs ("<=", file); break; |
87ad11b0 | 4658 | case GE: |
9c0ac0fd | 4659 | fputs ("<", file); break; |
87ad11b0 | 4660 | case GEU: |
9c0ac0fd | 4661 | fputs ("<<", file); break; |
87ad11b0 | 4662 | case GTU: |
9c0ac0fd | 4663 | fputs ("<<=", file); break; |
87ad11b0 | 4664 | case LT: |
9c0ac0fd | 4665 | fputs (">=", file); break; |
87ad11b0 | 4666 | case LE: |
9c0ac0fd | 4667 | fputs (">", file); break; |
87ad11b0 | 4668 | case LEU: |
9c0ac0fd | 4669 | fputs (">>", file); break; |
87ad11b0 | 4670 | case LTU: |
9c0ac0fd | 4671 | fputs (">>=", file); break; |
87ad11b0 | 4672 | default: |
87ad11b0 | 4673 | abort (); |
4674 | } | |
4675 | return; | |
ea52c577 | 4676 | /* For floating point comparisons. Note that the output |
4677 | predicates are the complement of the desired mode. */ | |
61230bc9 | 4678 | case 'Y': |
4679 | switch (GET_CODE (x)) | |
4680 | { | |
4681 | case EQ: | |
9c0ac0fd | 4682 | fputs ("!=", file); break; |
61230bc9 | 4683 | case NE: |
9c0ac0fd | 4684 | fputs ("=", file); break; |
61230bc9 | 4685 | case GT: |
32509e56 | 4686 | fputs ("!>", file); break; |
61230bc9 | 4687 | case GE: |
32509e56 | 4688 | fputs ("!>=", file); break; |
61230bc9 | 4689 | case LT: |
32509e56 | 4690 | fputs ("!<", file); break; |
61230bc9 | 4691 | case LE: |
32509e56 | 4692 | fputs ("!<=", file); break; |
4693 | case LTGT: | |
4694 | fputs ("!<>", file); break; | |
4695 | case UNLE: | |
85837adc | 4696 | fputs (">", file); break; |
32509e56 | 4697 | case UNLT: |
4698 | fputs (">=", file); break; | |
4699 | case UNGE: | |
4700 | fputs ("<", file); break; | |
4701 | case UNGT: | |
4702 | fputs ("<=", file); break; | |
4703 | case UNEQ: | |
4704 | fputs ("<>", file); break; | |
4705 | case UNORDERED: | |
4706 | fputs ("<=>", file); break; | |
4707 | case ORDERED: | |
4708 | fputs ("!<=>", file); break; | |
61230bc9 | 4709 | default: |
61230bc9 | 4710 | abort (); |
4711 | } | |
4712 | return; | |
c8975385 | 4713 | case 'S': /* Condition, operands are (S)wapped. */ |
4714 | switch (GET_CODE (x)) | |
4715 | { | |
4716 | case EQ: | |
9c0ac0fd | 4717 | fputs ("=", file); break; |
c8975385 | 4718 | case NE: |
9c0ac0fd | 4719 | fputs ("<>", file); break; |
c8975385 | 4720 | case GT: |
9c0ac0fd | 4721 | fputs ("<", file); break; |
c8975385 | 4722 | case GE: |
9c0ac0fd | 4723 | fputs ("<=", file); break; |
c8975385 | 4724 | case GEU: |
9c0ac0fd | 4725 | fputs ("<<=", file); break; |
c8975385 | 4726 | case GTU: |
9c0ac0fd | 4727 | fputs ("<<", file); break; |
c8975385 | 4728 | case LT: |
9c0ac0fd | 4729 | fputs (">", file); break; |
c8975385 | 4730 | case LE: |
9c0ac0fd | 4731 | fputs (">=", file); break; |
c8975385 | 4732 | case LEU: |
9c0ac0fd | 4733 | fputs (">>=", file); break; |
c8975385 | 4734 | case LTU: |
9c0ac0fd | 4735 | fputs (">>", file); break; |
c8975385 | 4736 | default: |
c8975385 | 4737 | abort (); |
6d36483b | 4738 | } |
c8975385 | 4739 | return; |
4740 | case 'B': /* Condition, (B)oth swapped and negate. */ | |
4741 | switch (GET_CODE (x)) | |
4742 | { | |
4743 | case EQ: | |
9c0ac0fd | 4744 | fputs ("<>", file); break; |
c8975385 | 4745 | case NE: |
9c0ac0fd | 4746 | fputs ("=", file); break; |
c8975385 | 4747 | case GT: |
9c0ac0fd | 4748 | fputs (">=", file); break; |
c8975385 | 4749 | case GE: |
9c0ac0fd | 4750 | fputs (">", file); break; |
c8975385 | 4751 | case GEU: |
9c0ac0fd | 4752 | fputs (">>", file); break; |
c8975385 | 4753 | case GTU: |
9c0ac0fd | 4754 | fputs (">>=", file); break; |
c8975385 | 4755 | case LT: |
9c0ac0fd | 4756 | fputs ("<=", file); break; |
c8975385 | 4757 | case LE: |
9c0ac0fd | 4758 | fputs ("<", file); break; |
c8975385 | 4759 | case LEU: |
9c0ac0fd | 4760 | fputs ("<<", file); break; |
c8975385 | 4761 | case LTU: |
9c0ac0fd | 4762 | fputs ("<<=", file); break; |
c8975385 | 4763 | default: |
c8975385 | 4764 | abort (); |
6d36483b | 4765 | } |
c8975385 | 4766 | return; |
4767 | case 'k': | |
4768 | if (GET_CODE (x) == CONST_INT) | |
4769 | { | |
96b529f2 | 4770 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~INTVAL (x)); |
c8975385 | 4771 | return; |
4772 | } | |
ea52c577 | 4773 | abort (); |
5e3c5739 | 4774 | case 'Q': |
4775 | if (GET_CODE (x) == CONST_INT) | |
4776 | { | |
96b529f2 | 4777 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - (INTVAL (x) & 63)); |
5e3c5739 | 4778 | return; |
4779 | } | |
ea52c577 | 4780 | abort (); |
e5965947 | 4781 | case 'L': |
4782 | if (GET_CODE (x) == CONST_INT) | |
4783 | { | |
96b529f2 | 4784 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - (INTVAL (x) & 31)); |
e5965947 | 4785 | return; |
4786 | } | |
ea52c577 | 4787 | abort (); |
3a16146d | 4788 | case 'O': |
4789 | if (GET_CODE (x) == CONST_INT && exact_log2 (INTVAL (x)) >= 0) | |
4790 | { | |
4791 | fprintf (file, "%d", exact_log2 (INTVAL (x))); | |
4792 | return; | |
4793 | } | |
ea52c577 | 4794 | abort (); |
5e3c5739 | 4795 | case 'p': |
4796 | if (GET_CODE (x) == CONST_INT) | |
4797 | { | |
96b529f2 | 4798 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 63 - (INTVAL (x) & 63)); |
5e3c5739 | 4799 | return; |
4800 | } | |
ea52c577 | 4801 | abort (); |
e5965947 | 4802 | case 'P': |
4803 | if (GET_CODE (x) == CONST_INT) | |
4804 | { | |
96b529f2 | 4805 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 31 - (INTVAL (x) & 31)); |
e5965947 | 4806 | return; |
4807 | } | |
ea52c577 | 4808 | abort (); |
c8975385 | 4809 | case 'I': |
4810 | if (GET_CODE (x) == CONST_INT) | |
4811 | fputs ("i", file); | |
4812 | return; | |
87ad11b0 | 4813 | case 'M': |
27ef382d | 4814 | case 'F': |
87ad11b0 | 4815 | switch (GET_CODE (XEXP (x, 0))) |
4816 | { | |
4817 | case PRE_DEC: | |
4818 | case PRE_INC: | |
e4065f95 | 4819 | if (ASSEMBLER_DIALECT == 0) |
4820 | fputs ("s,mb", file); | |
4821 | else | |
4822 | fputs (",mb", file); | |
87ad11b0 | 4823 | break; |
4824 | case POST_DEC: | |
4825 | case POST_INC: | |
e4065f95 | 4826 | if (ASSEMBLER_DIALECT == 0) |
4827 | fputs ("s,ma", file); | |
4828 | else | |
4829 | fputs (",ma", file); | |
87ad11b0 | 4830 | break; |
27ef382d | 4831 | case PLUS: |
4832 | if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
4833 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT) | |
e4065f95 | 4834 | { |
4835 | if (ASSEMBLER_DIALECT == 0) | |
4836 | fputs ("x,s", file); | |
4837 | else | |
4838 | fputs (",s", file); | |
4839 | } | |
4840 | else if (code == 'F' && ASSEMBLER_DIALECT == 0) | |
27ef382d | 4841 | fputs ("s", file); |
87ad11b0 | 4842 | break; |
4843 | default: | |
e4065f95 | 4844 | if (code == 'F' && ASSEMBLER_DIALECT == 0) |
27ef382d | 4845 | fputs ("s", file); |
87ad11b0 | 4846 | break; |
4847 | } | |
4848 | return; | |
4849 | case 'G': | |
f9333726 | 4850 | output_global_address (file, x, 0); |
4851 | return; | |
4852 | case 'H': | |
4853 | output_global_address (file, x, 1); | |
87ad11b0 | 4854 | return; |
4855 | case 0: /* Don't do anything special */ | |
4856 | break; | |
42faba01 | 4857 | case 'Z': |
4858 | { | |
4859 | unsigned op[3]; | |
fb22aedc | 4860 | compute_zdepwi_operands (INTVAL (x), op); |
42faba01 | 4861 | fprintf (file, "%d,%d,%d", op[0], op[1], op[2]); |
4862 | return; | |
4863 | } | |
5e3c5739 | 4864 | case 'z': |
4865 | { | |
4866 | unsigned op[3]; | |
4867 | compute_zdepdi_operands (INTVAL (x), op); | |
4868 | fprintf (file, "%d,%d,%d", op[0], op[1], op[2]); | |
4869 | return; | |
4870 | } | |
c9cc98e1 | 4871 | case 'c': |
4872 | /* We can get here from a .vtable_inherit due to our | |
4873 | CONSTANT_ADDRESS_P rejecting perfectly good constant | |
4874 | addresses. */ | |
4875 | break; | |
87ad11b0 | 4876 | default: |
4877 | abort (); | |
4878 | } | |
4879 | if (GET_CODE (x) == REG) | |
df0651dc | 4880 | { |
35661368 | 4881 | fputs (reg_names [REGNO (x)], file); |
5e3c5739 | 4882 | if (TARGET_64BIT && FP_REG_P (x) && GET_MODE_SIZE (GET_MODE (x)) <= 4) |
4883 | { | |
4884 | fputs ("R", file); | |
4885 | return; | |
4886 | } | |
4887 | if (FP_REG_P (x) | |
4888 | && GET_MODE_SIZE (GET_MODE (x)) <= 4 | |
4889 | && (REGNO (x) & 1) == 0) | |
35661368 | 4890 | fputs ("L", file); |
df0651dc | 4891 | } |
87ad11b0 | 4892 | else if (GET_CODE (x) == MEM) |
4893 | { | |
4894 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
f7dff90d | 4895 | rtx base = NULL_RTX; |
87ad11b0 | 4896 | switch (GET_CODE (XEXP (x, 0))) |
4897 | { | |
4898 | case PRE_DEC: | |
4899 | case POST_DEC: | |
5e3c5739 | 4900 | base = XEXP (XEXP (x, 0), 0); |
34940871 | 4901 | fprintf (file, "-%d(%s)", size, reg_names [REGNO (base)]); |
87ad11b0 | 4902 | break; |
4903 | case PRE_INC: | |
4904 | case POST_INC: | |
5e3c5739 | 4905 | base = XEXP (XEXP (x, 0), 0); |
34940871 | 4906 | fprintf (file, "%d(%s)", size, reg_names [REGNO (base)]); |
87ad11b0 | 4907 | break; |
4908 | default: | |
27ef382d | 4909 | if (GET_CODE (XEXP (x, 0)) == PLUS |
4910 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT) | |
34940871 | 4911 | fprintf (file, "%s(%s)", |
27ef382d | 4912 | reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 0), 0))], |
4913 | reg_names [REGNO (XEXP (XEXP (x, 0), 1))]); | |
4914 | else if (GET_CODE (XEXP (x, 0)) == PLUS | |
4915 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT) | |
34940871 | 4916 | fprintf (file, "%s(%s)", |
27ef382d | 4917 | reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 1), 0))], |
4918 | reg_names [REGNO (XEXP (XEXP (x, 0), 0))]); | |
4919 | else | |
4920 | output_address (XEXP (x, 0)); | |
87ad11b0 | 4921 | break; |
4922 | } | |
4923 | } | |
87ad11b0 | 4924 | else |
4925 | output_addr_const (file, x); | |
4926 | } | |
4927 | ||
6dc3b0d9 | 4928 | /* output a SYMBOL_REF or a CONST expression involving a SYMBOL_REF. */ |
87ad11b0 | 4929 | |
4930 | void | |
5c1d8983 | 4931 | output_global_address (FILE *file, rtx x, int round_constant) |
87ad11b0 | 4932 | { |
2ee034bc | 4933 | |
4934 | /* Imagine (high (const (plus ...))). */ | |
4935 | if (GET_CODE (x) == HIGH) | |
4936 | x = XEXP (x, 0); | |
4937 | ||
611a88e1 | 4938 | if (GET_CODE (x) == SYMBOL_REF && read_only_operand (x, VOIDmode)) |
87ad11b0 | 4939 | assemble_name (file, XSTR (x, 0)); |
b4a7bf10 | 4940 | else if (GET_CODE (x) == SYMBOL_REF && !flag_pic) |
87ad11b0 | 4941 | { |
4942 | assemble_name (file, XSTR (x, 0)); | |
9c0ac0fd | 4943 | fputs ("-$global$", file); |
87ad11b0 | 4944 | } |
4945 | else if (GET_CODE (x) == CONST) | |
4946 | { | |
611a88e1 | 4947 | const char *sep = ""; |
87ad11b0 | 4948 | int offset = 0; /* assembler wants -$global$ at end */ |
33ae0dba | 4949 | rtx base = NULL_RTX; |
6d36483b | 4950 | |
87ad11b0 | 4951 | if (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) |
4952 | { | |
4953 | base = XEXP (XEXP (x, 0), 0); | |
4954 | output_addr_const (file, base); | |
4955 | } | |
4956 | else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == CONST_INT) | |
4957 | offset = INTVAL (XEXP (XEXP (x, 0), 0)); | |
4958 | else abort (); | |
4959 | ||
4960 | if (GET_CODE (XEXP (XEXP (x, 0), 1)) == SYMBOL_REF) | |
4961 | { | |
4962 | base = XEXP (XEXP (x, 0), 1); | |
4963 | output_addr_const (file, base); | |
4964 | } | |
4965 | else if (GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT) | |
ea52c577 | 4966 | offset = INTVAL (XEXP (XEXP (x, 0), 1)); |
87ad11b0 | 4967 | else abort (); |
4968 | ||
f9333726 | 4969 | /* How bogus. The compiler is apparently responsible for |
4970 | rounding the constant if it uses an LR field selector. | |
4971 | ||
4972 | The linker and/or assembler seem a better place since | |
4973 | they have to do this kind of thing already. | |
4974 | ||
4975 | If we fail to do this, HP's optimizing linker may eliminate | |
4976 | an addil, but not update the ldw/stw/ldo instruction that | |
4977 | uses the result of the addil. */ | |
4978 | if (round_constant) | |
4979 | offset = ((offset + 0x1000) & ~0x1fff); | |
4980 | ||
87ad11b0 | 4981 | if (GET_CODE (XEXP (x, 0)) == PLUS) |
4982 | { | |
4983 | if (offset < 0) | |
4984 | { | |
4985 | offset = -offset; | |
4986 | sep = "-"; | |
4987 | } | |
4988 | else | |
4989 | sep = "+"; | |
4990 | } | |
4991 | else if (GET_CODE (XEXP (x, 0)) == MINUS | |
4992 | && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
4993 | sep = "-"; | |
4994 | else abort (); | |
4995 | ||
611a88e1 | 4996 | if (!read_only_operand (base, VOIDmode) && !flag_pic) |
9c0ac0fd | 4997 | fputs ("-$global$", file); |
f9333726 | 4998 | if (offset) |
ea52c577 | 4999 | fprintf (file, "%s%d", sep, offset); |
87ad11b0 | 5000 | } |
5001 | else | |
5002 | output_addr_const (file, x); | |
5003 | } | |
5004 | ||
92c473b8 | 5005 | /* Output boilerplate text to appear at the beginning of the file. |
5006 | There are several possible versions. */ | |
5007 | #define aputs(x) fputs(x, asm_out_file) | |
5008 | static inline void | |
5c1d8983 | 5009 | pa_file_start_level (void) |
92c473b8 | 5010 | { |
5011 | if (TARGET_64BIT) | |
5012 | aputs ("\t.LEVEL 2.0w\n"); | |
5013 | else if (TARGET_PA_20) | |
5014 | aputs ("\t.LEVEL 2.0\n"); | |
5015 | else if (TARGET_PA_11) | |
5016 | aputs ("\t.LEVEL 1.1\n"); | |
5017 | else | |
5018 | aputs ("\t.LEVEL 1.0\n"); | |
5019 | } | |
5020 | ||
5021 | static inline void | |
5c1d8983 | 5022 | pa_file_start_space (int sortspace) |
92c473b8 | 5023 | { |
5024 | aputs ("\t.SPACE $PRIVATE$"); | |
5025 | if (sortspace) | |
5026 | aputs (",SORT=16"); | |
5027 | aputs ("\n\t.SUBSPA $DATA$,QUAD=1,ALIGN=8,ACCESS=31" | |
5028 | "\n\t.SUBSPA $BSS$,QUAD=1,ALIGN=8,ACCESS=31,ZERO,SORT=82" | |
5029 | "\n\t.SPACE $TEXT$"); | |
5030 | if (sortspace) | |
5031 | aputs (",SORT=8"); | |
5032 | aputs ("\n\t.SUBSPA $LIT$,QUAD=0,ALIGN=8,ACCESS=44" | |
5033 | "\n\t.SUBSPA $CODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY\n"); | |
5034 | } | |
5035 | ||
5036 | static inline void | |
5c1d8983 | 5037 | pa_file_start_file (int want_version) |
92c473b8 | 5038 | { |
5039 | if (write_symbols != NO_DEBUG) | |
5040 | { | |
5041 | output_file_directive (asm_out_file, main_input_filename); | |
5042 | if (want_version) | |
5043 | aputs ("\t.version\t\"01.01\"\n"); | |
5044 | } | |
5045 | } | |
5046 | ||
5047 | static inline void | |
5c1d8983 | 5048 | pa_file_start_mcount (const char *aswhat) |
92c473b8 | 5049 | { |
5050 | if (profile_flag) | |
5051 | fprintf (asm_out_file, "\t.IMPORT _mcount,%s\n", aswhat); | |
5052 | } | |
5053 | ||
5054 | static void | |
5c1d8983 | 5055 | pa_elf_file_start (void) |
92c473b8 | 5056 | { |
5057 | pa_file_start_level (); | |
5058 | pa_file_start_mcount ("ENTRY"); | |
5059 | pa_file_start_file (0); | |
5060 | } | |
5061 | ||
5062 | static void | |
5c1d8983 | 5063 | pa_som_file_start (void) |
92c473b8 | 5064 | { |
5065 | pa_file_start_level (); | |
5066 | pa_file_start_space (0); | |
5067 | aputs ("\t.IMPORT $global$,DATA\n" | |
5068 | "\t.IMPORT $$dyncall,MILLICODE\n"); | |
5069 | pa_file_start_mcount ("CODE"); | |
5070 | pa_file_start_file (0); | |
5071 | } | |
5072 | ||
5073 | static void | |
5c1d8983 | 5074 | pa_linux_file_start (void) |
92c473b8 | 5075 | { |
5076 | pa_file_start_file (1); | |
5077 | pa_file_start_level (); | |
5078 | pa_file_start_mcount ("CODE"); | |
5079 | } | |
5080 | ||
5081 | static void | |
5c1d8983 | 5082 | pa_hpux64_gas_file_start (void) |
92c473b8 | 5083 | { |
5084 | pa_file_start_level (); | |
5085 | #ifdef ASM_OUTPUT_TYPE_DIRECTIVE | |
5086 | if (profile_flag) | |
5087 | ASM_OUTPUT_TYPE_DIRECTIVE (asm_out_file, "_mcount", "function"); | |
5088 | #endif | |
5089 | pa_file_start_file (1); | |
5090 | } | |
5091 | ||
5092 | static void | |
5c1d8983 | 5093 | pa_hpux64_hpas_file_start (void) |
92c473b8 | 5094 | { |
5095 | pa_file_start_level (); | |
5096 | pa_file_start_space (1); | |
5097 | pa_file_start_mcount ("CODE"); | |
5098 | pa_file_start_file (0); | |
5099 | } | |
5100 | #undef aputs | |
5101 | ||
ece88821 | 5102 | static struct deferred_plabel * |
5c1d8983 | 5103 | get_plabel (const char *fname) |
ece88821 | 5104 | { |
5105 | size_t i; | |
5106 | ||
5107 | /* See if we have already put this function on the list of deferred | |
5108 | plabels. This list is generally small, so a liner search is not | |
5109 | too ugly. If it proves too slow replace it with something faster. */ | |
5110 | for (i = 0; i < n_deferred_plabels; i++) | |
5111 | if (strcmp (fname, deferred_plabels[i].name) == 0) | |
5112 | break; | |
5113 | ||
5114 | /* If the deferred plabel list is empty, or this entry was not found | |
5115 | on the list, create a new entry on the list. */ | |
5116 | if (deferred_plabels == NULL || i == n_deferred_plabels) | |
5117 | { | |
5118 | const char *real_name; | |
5119 | ||
5120 | if (deferred_plabels == 0) | |
5121 | deferred_plabels = (struct deferred_plabel *) | |
5122 | ggc_alloc (sizeof (struct deferred_plabel)); | |
5123 | else | |
5124 | deferred_plabels = (struct deferred_plabel *) | |
5125 | ggc_realloc (deferred_plabels, | |
5126 | ((n_deferred_plabels + 1) | |
5127 | * sizeof (struct deferred_plabel))); | |
5128 | ||
5129 | i = n_deferred_plabels++; | |
5130 | deferred_plabels[i].internal_label = gen_label_rtx (); | |
5131 | deferred_plabels[i].name = ggc_strdup (fname); | |
5132 | ||
5133 | /* Gross. We have just implicitly taken the address of this function, | |
5134 | mark it as such. */ | |
5135 | real_name = (*targetm.strip_name_encoding) (fname); | |
5136 | TREE_SYMBOL_REFERENCED (get_identifier (real_name)) = 1; | |
5137 | } | |
5138 | ||
5139 | return &deferred_plabels[i]; | |
5140 | } | |
5141 | ||
f6940372 | 5142 | static void |
5c1d8983 | 5143 | output_deferred_plabels (void) |
5cc6b2bc | 5144 | { |
e11bd7e5 | 5145 | size_t i; |
5cc6b2bc | 5146 | /* If we have deferred plabels, then we need to switch into the data |
5147 | section and align it to a 4 byte boundary before we output the | |
5148 | deferred plabels. */ | |
5149 | if (n_deferred_plabels) | |
5150 | { | |
5151 | data_section (); | |
f6940372 | 5152 | ASM_OUTPUT_ALIGN (asm_out_file, TARGET_64BIT ? 3 : 2); |
5cc6b2bc | 5153 | } |
5154 | ||
5155 | /* Now output the deferred plabels. */ | |
5156 | for (i = 0; i < n_deferred_plabels; i++) | |
5157 | { | |
f6940372 | 5158 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
5159 | CODE_LABEL_NUMBER (deferred_plabels[i].internal_label)); | |
09d688ff | 5160 | assemble_integer (gen_rtx_SYMBOL_REF (Pmode, deferred_plabels[i].name), |
b70ea764 | 5161 | TARGET_64BIT ? 8 : 4, TARGET_64BIT ? 64 : 32, 1); |
5cc6b2bc | 5162 | } |
5163 | } | |
5164 | ||
f2f543a3 | 5165 | #ifdef HPUX_LONG_DOUBLE_LIBRARY |
5166 | /* Initialize optabs to point to HPUX long double emulation routines. */ | |
5167 | static void | |
5168 | pa_hpux_init_libfuncs (void) | |
5169 | { | |
5170 | set_optab_libfunc (add_optab, TFmode, "_U_Qfadd"); | |
5171 | set_optab_libfunc (sub_optab, TFmode, "_U_Qfsub"); | |
5172 | set_optab_libfunc (smul_optab, TFmode, "_U_Qfmpy"); | |
5173 | set_optab_libfunc (sdiv_optab, TFmode, "_U_Qfdiv"); | |
5174 | set_optab_libfunc (smin_optab, TFmode, "_U_Qmin"); | |
5175 | set_optab_libfunc (smax_optab, TFmode, "_U_Qfmax"); | |
5176 | set_optab_libfunc (sqrt_optab, TFmode, "_U_Qfsqrt"); | |
5177 | set_optab_libfunc (abs_optab, TFmode, "_U_Qfabs"); | |
5178 | set_optab_libfunc (neg_optab, TFmode, "_U_Qfneg"); | |
5179 | ||
c88a6ebd | 5180 | set_optab_libfunc (eq_optab, TFmode, "_U_Qfeq"); |
5181 | set_optab_libfunc (ne_optab, TFmode, "_U_Qfne"); | |
5182 | set_optab_libfunc (gt_optab, TFmode, "_U_Qfgt"); | |
5183 | set_optab_libfunc (ge_optab, TFmode, "_U_Qfge"); | |
5184 | set_optab_libfunc (lt_optab, TFmode, "_U_Qflt"); | |
5185 | set_optab_libfunc (le_optab, TFmode, "_U_Qfle"); | |
f2f543a3 | 5186 | |
a7cc195f | 5187 | set_conv_libfunc (sext_optab, TFmode, SFmode, "_U_Qfcnvff_sgl_to_quad"); |
5188 | set_conv_libfunc (sext_optab, TFmode, DFmode, "_U_Qfcnvff_dbl_to_quad"); | |
5189 | set_conv_libfunc (trunc_optab, SFmode, TFmode, "_U_Qfcnvff_quad_to_sgl"); | |
5190 | set_conv_libfunc (trunc_optab, DFmode, TFmode, "_U_Qfcnvff_quad_to_dbl"); | |
5191 | ||
5192 | set_conv_libfunc (sfix_optab, SImode, TFmode, TARGET_64BIT | |
5193 | ? "__U_Qfcnvfxt_quad_to_sgl" | |
5194 | : "_U_Qfcnvfxt_quad_to_sgl"); | |
5195 | set_conv_libfunc (sfix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_dbl"); | |
5196 | set_conv_libfunc (ufix_optab, SImode, TFmode, "_U_Qfcnvfxt_quad_to_usgl"); | |
5197 | set_conv_libfunc (ufix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_udbl"); | |
5198 | ||
5199 | set_conv_libfunc (sfloat_optab, TFmode, SImode, "_U_Qfcnvxf_sgl_to_quad"); | |
5200 | set_conv_libfunc (sfloat_optab, TFmode, DImode, "_U_Qfcnvxf_dbl_to_quad"); | |
f2f543a3 | 5201 | } |
5202 | #endif | |
5203 | ||
87ad11b0 | 5204 | /* HP's millicode routines mean something special to the assembler. |
5205 | Keep track of which ones we have used. */ | |
5206 | ||
c2271c34 | 5207 | enum millicodes { remI, remU, divI, divU, mulI, end1000 }; |
5c1d8983 | 5208 | static void import_milli (enum millicodes); |
ea52c577 | 5209 | static char imported[(int) end1000]; |
c2271c34 | 5210 | static const char * const milli_names[] = {"remI", "remU", "divI", "divU", "mulI"}; |
e99c3a1d | 5211 | static const char import_string[] = ".IMPORT $$....,MILLICODE"; |
87ad11b0 | 5212 | #define MILLI_START 10 |
5213 | ||
57ed30e5 | 5214 | static void |
5c1d8983 | 5215 | import_milli (enum millicodes code) |
87ad11b0 | 5216 | { |
5217 | char str[sizeof (import_string)]; | |
6d36483b | 5218 | |
ea52c577 | 5219 | if (!imported[(int) code]) |
87ad11b0 | 5220 | { |
ea52c577 | 5221 | imported[(int) code] = 1; |
87ad11b0 | 5222 | strcpy (str, import_string); |
ea52c577 | 5223 | strncpy (str + MILLI_START, milli_names[(int) code], 4); |
87ad11b0 | 5224 | output_asm_insn (str, 0); |
5225 | } | |
5226 | } | |
5227 | ||
6d36483b | 5228 | /* The register constraints have put the operands and return value in |
6dc3b0d9 | 5229 | the proper registers. */ |
87ad11b0 | 5230 | |
611a88e1 | 5231 | const char * |
5c1d8983 | 5232 | output_mul_insn (int unsignedp ATTRIBUTE_UNUSED, rtx insn) |
87ad11b0 | 5233 | { |
d178f670 | 5234 | import_milli (mulI); |
440c23df | 5235 | return output_millicode_call (insn, gen_rtx_SYMBOL_REF (Pmode, "$$mulI")); |
87ad11b0 | 5236 | } |
5237 | ||
6dc3b0d9 | 5238 | /* Emit the rtl for doing a division by a constant. */ |
87ad11b0 | 5239 | |
d178f670 | 5240 | /* Do magic division millicodes exist for this value? */ |
e99c3a1d | 5241 | static const int magic_milli[]= {0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, |
5242 | 1, 1}; | |
87ad11b0 | 5243 | |
6d36483b | 5244 | /* We'll use an array to keep track of the magic millicodes and |
87ad11b0 | 5245 | whether or not we've used them already. [n][0] is signed, [n][1] is |
6dc3b0d9 | 5246 | unsigned. */ |
87ad11b0 | 5247 | |
87ad11b0 | 5248 | static int div_milli[16][2]; |
5249 | ||
5250 | int | |
5c1d8983 | 5251 | div_operand (rtx op, enum machine_mode mode) |
87ad11b0 | 5252 | { |
5253 | return (mode == SImode | |
5254 | && ((GET_CODE (op) == REG && REGNO (op) == 25) | |
5255 | || (GET_CODE (op) == CONST_INT && INTVAL (op) > 0 | |
5256 | && INTVAL (op) < 16 && magic_milli[INTVAL (op)]))); | |
5257 | } | |
5258 | ||
5259 | int | |
5c1d8983 | 5260 | emit_hpdiv_const (rtx *operands, int unsignedp) |
87ad11b0 | 5261 | { |
5262 | if (GET_CODE (operands[2]) == CONST_INT | |
5263 | && INTVAL (operands[2]) > 0 | |
5264 | && INTVAL (operands[2]) < 16 | |
5265 | && magic_milli[INTVAL (operands[2])]) | |
5266 | { | |
2013ddf6 | 5267 | rtx ret = gen_rtx_REG (SImode, TARGET_64BIT ? 2 : 31); |
5268 | ||
ad851752 | 5269 | emit_move_insn (gen_rtx_REG (SImode, 26), operands[1]); |
87ad11b0 | 5270 | emit |
5271 | (gen_rtx | |
5272 | (PARALLEL, VOIDmode, | |
c2078db8 | 5273 | gen_rtvec (6, gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, 29), |
ad851752 | 5274 | gen_rtx_fmt_ee (unsignedp ? UDIV : DIV, |
5275 | SImode, | |
5276 | gen_rtx_REG (SImode, 26), | |
5277 | operands[2])), | |
c2078db8 | 5278 | gen_rtx_CLOBBER (VOIDmode, operands[4]), |
ad851752 | 5279 | gen_rtx_CLOBBER (VOIDmode, operands[3]), |
5280 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 26)), | |
5281 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 25)), | |
2013ddf6 | 5282 | gen_rtx_CLOBBER (VOIDmode, ret)))); |
ad851752 | 5283 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 29)); |
87ad11b0 | 5284 | return 1; |
5285 | } | |
5286 | return 0; | |
5287 | } | |
5288 | ||
611a88e1 | 5289 | const char * |
5c1d8983 | 5290 | output_div_insn (rtx *operands, int unsignedp, rtx insn) |
87ad11b0 | 5291 | { |
5292 | int divisor; | |
6d36483b | 5293 | |
5294 | /* If the divisor is a constant, try to use one of the special | |
87ad11b0 | 5295 | opcodes .*/ |
5296 | if (GET_CODE (operands[0]) == CONST_INT) | |
5297 | { | |
d6686e21 | 5298 | static char buf[100]; |
87ad11b0 | 5299 | divisor = INTVAL (operands[0]); |
5300 | if (!div_milli[divisor][unsignedp]) | |
5301 | { | |
d6686e21 | 5302 | div_milli[divisor][unsignedp] = 1; |
87ad11b0 | 5303 | if (unsignedp) |
5304 | output_asm_insn (".IMPORT $$divU_%0,MILLICODE", operands); | |
5305 | else | |
5306 | output_asm_insn (".IMPORT $$divI_%0,MILLICODE", operands); | |
87ad11b0 | 5307 | } |
5308 | if (unsignedp) | |
d6686e21 | 5309 | { |
4840a03a | 5310 | sprintf (buf, "$$divU_" HOST_WIDE_INT_PRINT_DEC, |
5311 | INTVAL (operands[0])); | |
c7a4e712 | 5312 | return output_millicode_call (insn, |
ad851752 | 5313 | gen_rtx_SYMBOL_REF (SImode, buf)); |
d6686e21 | 5314 | } |
5315 | else | |
5316 | { | |
4840a03a | 5317 | sprintf (buf, "$$divI_" HOST_WIDE_INT_PRINT_DEC, |
5318 | INTVAL (operands[0])); | |
c7a4e712 | 5319 | return output_millicode_call (insn, |
ad851752 | 5320 | gen_rtx_SYMBOL_REF (SImode, buf)); |
d6686e21 | 5321 | } |
87ad11b0 | 5322 | } |
6dc3b0d9 | 5323 | /* Divisor isn't a special constant. */ |
87ad11b0 | 5324 | else |
5325 | { | |
5326 | if (unsignedp) | |
5327 | { | |
5328 | import_milli (divU); | |
c7a4e712 | 5329 | return output_millicode_call (insn, |
ad851752 | 5330 | gen_rtx_SYMBOL_REF (SImode, "$$divU")); |
87ad11b0 | 5331 | } |
5332 | else | |
5333 | { | |
5334 | import_milli (divI); | |
c7a4e712 | 5335 | return output_millicode_call (insn, |
ad851752 | 5336 | gen_rtx_SYMBOL_REF (SImode, "$$divI")); |
87ad11b0 | 5337 | } |
5338 | } | |
5339 | } | |
5340 | ||
6dc3b0d9 | 5341 | /* Output a $$rem millicode to do mod. */ |
87ad11b0 | 5342 | |
611a88e1 | 5343 | const char * |
5c1d8983 | 5344 | output_mod_insn (int unsignedp, rtx insn) |
87ad11b0 | 5345 | { |
5346 | if (unsignedp) | |
5347 | { | |
5348 | import_milli (remU); | |
c7a4e712 | 5349 | return output_millicode_call (insn, |
ad851752 | 5350 | gen_rtx_SYMBOL_REF (SImode, "$$remU")); |
87ad11b0 | 5351 | } |
5352 | else | |
5353 | { | |
5354 | import_milli (remI); | |
c7a4e712 | 5355 | return output_millicode_call (insn, |
ad851752 | 5356 | gen_rtx_SYMBOL_REF (SImode, "$$remI")); |
87ad11b0 | 5357 | } |
5358 | } | |
5359 | ||
5360 | void | |
5c1d8983 | 5361 | output_arg_descriptor (rtx call_insn) |
87ad11b0 | 5362 | { |
611a88e1 | 5363 | const char *arg_regs[4]; |
87ad11b0 | 5364 | enum machine_mode arg_mode; |
df0651dc | 5365 | rtx link; |
87ad11b0 | 5366 | int i, output_flag = 0; |
5367 | int regno; | |
6d36483b | 5368 | |
5e3c5739 | 5369 | /* We neither need nor want argument location descriptors for the |
5d0619cc | 5370 | 64bit runtime environment or the ELF32 environment. */ |
5371 | if (TARGET_64BIT || TARGET_ELF32) | |
5e3c5739 | 5372 | return; |
5373 | ||
87ad11b0 | 5374 | for (i = 0; i < 4; i++) |
5375 | arg_regs[i] = 0; | |
5376 | ||
738176ab | 5377 | /* Specify explicitly that no argument relocations should take place |
5378 | if using the portable runtime calling conventions. */ | |
5379 | if (TARGET_PORTABLE_RUNTIME) | |
5380 | { | |
9c0ac0fd | 5381 | fputs ("\t.CALL ARGW0=NO,ARGW1=NO,ARGW2=NO,ARGW3=NO,RETVAL=NO\n", |
5382 | asm_out_file); | |
738176ab | 5383 | return; |
5384 | } | |
5385 | ||
df0651dc | 5386 | if (GET_CODE (call_insn) != CALL_INSN) |
5387 | abort (); | |
5388 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); link; link = XEXP (link, 1)) | |
87ad11b0 | 5389 | { |
df0651dc | 5390 | rtx use = XEXP (link, 0); |
c12afafd | 5391 | |
df0651dc | 5392 | if (! (GET_CODE (use) == USE |
5393 | && GET_CODE (XEXP (use, 0)) == REG | |
5394 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
c12afafd | 5395 | continue; |
5396 | ||
df0651dc | 5397 | arg_mode = GET_MODE (XEXP (use, 0)); |
5398 | regno = REGNO (XEXP (use, 0)); | |
87ad11b0 | 5399 | if (regno >= 23 && regno <= 26) |
372ef038 | 5400 | { |
5401 | arg_regs[26 - regno] = "GR"; | |
5402 | if (arg_mode == DImode) | |
5403 | arg_regs[25 - regno] = "GR"; | |
5404 | } | |
df0651dc | 5405 | else if (regno >= 32 && regno <= 39) |
87ad11b0 | 5406 | { |
5407 | if (arg_mode == SFmode) | |
df0651dc | 5408 | arg_regs[(regno - 32) / 2] = "FR"; |
e6ba640e | 5409 | else |
87ad11b0 | 5410 | { |
eeec72c0 | 5411 | #ifndef HP_FP_ARG_DESCRIPTOR_REVERSED |
df0651dc | 5412 | arg_regs[(regno - 34) / 2] = "FR"; |
5413 | arg_regs[(regno - 34) / 2 + 1] = "FU"; | |
87ad11b0 | 5414 | #else |
df0651dc | 5415 | arg_regs[(regno - 34) / 2] = "FU"; |
5416 | arg_regs[(regno - 34) / 2 + 1] = "FR"; | |
87ad11b0 | 5417 | #endif |
5418 | } | |
87ad11b0 | 5419 | } |
5420 | } | |
5421 | fputs ("\t.CALL ", asm_out_file); | |
5422 | for (i = 0; i < 4; i++) | |
5423 | { | |
5424 | if (arg_regs[i]) | |
5425 | { | |
5426 | if (output_flag++) | |
5427 | fputc (',', asm_out_file); | |
5428 | fprintf (asm_out_file, "ARGW%d=%s", i, arg_regs[i]); | |
5429 | } | |
5430 | } | |
5431 | fputc ('\n', asm_out_file); | |
5432 | } | |
5433 | \f | |
9c0ac0fd | 5434 | /* Return the class of any secondary reload register that is needed to |
5435 | move IN into a register in class CLASS using mode MODE. | |
5436 | ||
5437 | Profiling has showed this routine and its descendants account for | |
5438 | a significant amount of compile time (~7%). So it has been | |
5439 | optimized to reduce redundant computations and eliminate useless | |
5440 | function calls. | |
5441 | ||
5442 | It might be worthwhile to try and make this a leaf function too. */ | |
87ad11b0 | 5443 | |
5444 | enum reg_class | |
5c1d8983 | 5445 | secondary_reload_class (enum reg_class class, enum machine_mode mode, rtx in) |
87ad11b0 | 5446 | { |
9c0ac0fd | 5447 | int regno, is_symbolic; |
87ad11b0 | 5448 | |
b4a7bf10 | 5449 | /* Trying to load a constant into a FP register during PIC code |
5450 | generation will require %r1 as a scratch register. */ | |
fc44315f | 5451 | if (flag_pic |
b4a7bf10 | 5452 | && GET_MODE_CLASS (mode) == MODE_INT |
5453 | && FP_REG_CLASS_P (class) | |
5454 | && (GET_CODE (in) == CONST_INT || GET_CODE (in) == CONST_DOUBLE)) | |
5455 | return R1_REGS; | |
5456 | ||
9c0ac0fd | 5457 | /* Profiling showed the PA port spends about 1.3% of its compilation |
5458 | time in true_regnum from calls inside secondary_reload_class. */ | |
5459 | ||
5460 | if (GET_CODE (in) == REG) | |
5461 | { | |
5462 | regno = REGNO (in); | |
5463 | if (regno >= FIRST_PSEUDO_REGISTER) | |
5464 | regno = true_regnum (in); | |
5465 | } | |
5466 | else if (GET_CODE (in) == SUBREG) | |
5467 | regno = true_regnum (in); | |
9c0ac0fd | 5468 | else |
5469 | regno = -1; | |
5470 | ||
76a0ced5 | 5471 | /* If we have something like (mem (mem (...)), we can safely assume the |
5472 | inner MEM will end up in a general register after reloading, so there's | |
5473 | no need for a secondary reload. */ | |
5474 | if (GET_CODE (in) == MEM | |
5475 | && GET_CODE (XEXP (in, 0)) == MEM) | |
5476 | return NO_REGS; | |
5477 | ||
5478 | /* Handle out of range displacement for integer mode loads/stores of | |
5479 | FP registers. */ | |
d2498717 | 5480 | if (((regno >= FIRST_PSEUDO_REGISTER || regno == -1) |
6d36483b | 5481 | && GET_MODE_CLASS (mode) == MODE_INT |
5482 | && FP_REG_CLASS_P (class)) | |
d6f01525 | 5483 | || (class == SHIFT_REGS && (regno <= 0 || regno >= 32))) |
9c6d4825 | 5484 | return GENERAL_REGS; |
d2c1d63d | 5485 | |
7c4b32f3 | 5486 | /* A SAR<->FP register copy requires a secondary register (GPR) as |
5487 | well as secondary memory. */ | |
5488 | if (regno >= 0 && regno < FIRST_PSEUDO_REGISTER | |
5489 | && ((REGNO_REG_CLASS (regno) == SHIFT_REGS && FP_REG_CLASS_P (class)) | |
5490 | || (class == SHIFT_REGS && FP_REG_CLASS_P (REGNO_REG_CLASS (regno))))) | |
5491 | return GENERAL_REGS; | |
5492 | ||
2ee034bc | 5493 | if (GET_CODE (in) == HIGH) |
5494 | in = XEXP (in, 0); | |
5495 | ||
9c0ac0fd | 5496 | /* Profiling has showed GCC spends about 2.6% of its compilation |
5497 | time in symbolic_operand from calls inside secondary_reload_class. | |
5498 | ||
5499 | We use an inline copy and only compute its return value once to avoid | |
5500 | useless work. */ | |
5501 | switch (GET_CODE (in)) | |
5502 | { | |
5503 | rtx tmp; | |
5504 | ||
5505 | case SYMBOL_REF: | |
5506 | case LABEL_REF: | |
5507 | is_symbolic = 1; | |
5508 | break; | |
5509 | case CONST: | |
5510 | tmp = XEXP (in, 0); | |
5511 | is_symbolic = ((GET_CODE (XEXP (tmp, 0)) == SYMBOL_REF | |
5512 | || GET_CODE (XEXP (tmp, 0)) == LABEL_REF) | |
5513 | && GET_CODE (XEXP (tmp, 1)) == CONST_INT); | |
5514 | break; | |
76a0ced5 | 5515 | |
9c0ac0fd | 5516 | default: |
5517 | is_symbolic = 0; | |
5518 | break; | |
5519 | } | |
9840d99d | 5520 | |
b4a7bf10 | 5521 | if (!flag_pic |
9c0ac0fd | 5522 | && is_symbolic |
611a88e1 | 5523 | && read_only_operand (in, VOIDmode)) |
b4a7bf10 | 5524 | return NO_REGS; |
5525 | ||
9c0ac0fd | 5526 | if (class != R1_REGS && is_symbolic) |
2ee034bc | 5527 | return R1_REGS; |
5528 | ||
d2c1d63d | 5529 | return NO_REGS; |
87ad11b0 | 5530 | } |
5531 | ||
5532 | enum direction | |
5c1d8983 | 5533 | function_arg_padding (enum machine_mode mode, tree type) |
87ad11b0 | 5534 | { |
ac965869 | 5535 | if (mode == BLKmode |
5536 | || (TARGET_64BIT && type && AGGREGATE_TYPE_P (type))) | |
5537 | { | |
5538 | /* Return none if justification is not required. */ | |
5539 | if (type | |
5540 | && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST | |
5541 | && (int_size_in_bytes (type) * BITS_PER_UNIT) % PARM_BOUNDARY == 0) | |
5542 | return none; | |
5543 | ||
5544 | /* The directions set here are ignored when a BLKmode argument larger | |
5545 | than a word is placed in a register. Different code is used for | |
5546 | the stack and registers. This makes it difficult to have a | |
5547 | consistent data representation for both the stack and registers. | |
5548 | For both runtimes, the justification and padding for arguments on | |
5549 | the stack and in registers should be identical. */ | |
5550 | if (TARGET_64BIT) | |
5551 | /* The 64-bit runtime specifies left justification for aggregates. */ | |
5552 | return upward; | |
87ad11b0 | 5553 | else |
ac965869 | 5554 | /* The 32-bit runtime architecture specifies right justification. |
5555 | When the argument is passed on the stack, the argument is padded | |
5556 | with garbage on the left. The HP compiler pads with zeros. */ | |
5557 | return downward; | |
87ad11b0 | 5558 | } |
ac965869 | 5559 | |
5560 | if (GET_MODE_BITSIZE (mode) < PARM_BOUNDARY) | |
87ad11b0 | 5561 | return downward; |
87ad11b0 | 5562 | else |
5563 | return none; | |
5564 | } | |
5565 | ||
87ad11b0 | 5566 | \f |
55f54832 | 5567 | /* Do what is necessary for `va_start'. We look at the current function |
5568 | to determine if stdargs or varargs is used and fill in an initial | |
5569 | va_list. A pointer to this constructor is returned. */ | |
87ad11b0 | 5570 | |
5571 | struct rtx_def * | |
5c1d8983 | 5572 | hppa_builtin_saveregs (void) |
87ad11b0 | 5573 | { |
01251cbc | 5574 | rtx offset, dest; |
87ad11b0 | 5575 | tree fntype = TREE_TYPE (current_function_decl); |
5576 | int argadj = ((!(TYPE_ARG_TYPES (fntype) != 0 | |
5577 | && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype))) | |
5578 | != void_type_node))) | |
5579 | ? UNITS_PER_WORD : 0); | |
5580 | ||
5581 | if (argadj) | |
5582 | offset = plus_constant (current_function_arg_offset_rtx, argadj); | |
5583 | else | |
5584 | offset = current_function_arg_offset_rtx; | |
9c6d4825 | 5585 | |
5e3c5739 | 5586 | if (TARGET_64BIT) |
5587 | { | |
5588 | int i, off; | |
9840d99d | 5589 | |
5e3c5739 | 5590 | /* Adjust for varargs/stdarg differences. */ |
5591 | if (argadj) | |
5592 | offset = plus_constant (current_function_arg_offset_rtx, -argadj); | |
5593 | else | |
5594 | offset = current_function_arg_offset_rtx; | |
5595 | ||
5596 | /* We need to save %r26 .. %r19 inclusive starting at offset -64 | |
5597 | from the incoming arg pointer and growing to larger addresses. */ | |
5598 | for (i = 26, off = -64; i >= 19; i--, off += 8) | |
5599 | emit_move_insn (gen_rtx_MEM (word_mode, | |
5600 | plus_constant (arg_pointer_rtx, off)), | |
5601 | gen_rtx_REG (word_mode, i)); | |
5602 | ||
5603 | /* The incoming args pointer points just beyond the flushback area; | |
8ef587dc | 5604 | normally this is not a serious concern. However, when we are doing |
5e3c5739 | 5605 | varargs/stdargs we want to make the arg pointer point to the start |
5606 | of the incoming argument area. */ | |
5607 | emit_move_insn (virtual_incoming_args_rtx, | |
5608 | plus_constant (arg_pointer_rtx, -64)); | |
5609 | ||
5610 | /* Now return a pointer to the first anonymous argument. */ | |
5611 | return copy_to_reg (expand_binop (Pmode, add_optab, | |
5612 | virtual_incoming_args_rtx, | |
5613 | offset, 0, 0, OPTAB_LIB_WIDEN)); | |
5614 | } | |
5615 | ||
6dc3b0d9 | 5616 | /* Store general registers on the stack. */ |
ad851752 | 5617 | dest = gen_rtx_MEM (BLKmode, |
5618 | plus_constant (current_function_internal_arg_pointer, | |
5619 | -16)); | |
ab6ab77e | 5620 | set_mem_alias_set (dest, get_varargs_alias_set ()); |
2a631e19 | 5621 | set_mem_align (dest, BITS_PER_WORD); |
530178a9 | 5622 | move_block_from_reg (23, dest, 4); |
01251cbc | 5623 | |
76a0ced5 | 5624 | /* move_block_from_reg will emit code to store the argument registers |
5625 | individually as scalar stores. | |
5626 | ||
5627 | However, other insns may later load from the same addresses for | |
ad87de1e | 5628 | a structure load (passing a struct to a varargs routine). |
76a0ced5 | 5629 | |
5630 | The alias code assumes that such aliasing can never happen, so we | |
5631 | have to keep memory referencing insns from moving up beyond the | |
5632 | last argument register store. So we emit a blockage insn here. */ | |
5633 | emit_insn (gen_blockage ()); | |
5634 | ||
9c6d4825 | 5635 | return copy_to_reg (expand_binop (Pmode, add_optab, |
5636 | current_function_internal_arg_pointer, | |
5637 | offset, 0, 0, OPTAB_LIB_WIDEN)); | |
87ad11b0 | 5638 | } |
d6f01525 | 5639 | |
72899a61 | 5640 | void |
5c1d8983 | 5641 | hppa_va_start (tree valist, rtx nextarg) |
72899a61 | 5642 | { |
5643 | nextarg = expand_builtin_saveregs (); | |
7df226a2 | 5644 | std_expand_builtin_va_start (valist, nextarg); |
72899a61 | 5645 | } |
5646 | ||
5647 | rtx | |
5c1d8983 | 5648 | hppa_va_arg (tree valist, tree type) |
72899a61 | 5649 | { |
ac965869 | 5650 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5651 | HOST_WIDE_INT ofs; | |
72899a61 | 5652 | tree t, ptr, pptr; |
5653 | ||
5e3c5739 | 5654 | if (TARGET_64BIT) |
5655 | { | |
ac965869 | 5656 | /* Every argument in PA64 is supposed to be passed by value |
5657 | (including large structs). However, as a GCC extension, we | |
5658 | pass zero and variable sized arguments by reference. Empty | |
5659 | structures are a GCC extension not supported by the HP | |
5660 | compilers. Thus, passing them by reference isn't likely | |
5661 | to conflict with the ABI. For variable sized arguments, | |
5662 | GCC doesn't have the infrastructure to allocate these to | |
5663 | registers. */ | |
5664 | ||
5665 | /* Arguments with a size greater than 8 must be aligned 0 MOD 16. */ | |
5e3c5739 | 5666 | |
5e3c5739 | 5667 | if (size > UNITS_PER_WORD) |
5668 | { | |
5669 | t = build (PLUS_EXPR, TREE_TYPE (valist), valist, | |
5670 | build_int_2 (2 * UNITS_PER_WORD - 1, 0)); | |
9840d99d | 5671 | t = build (BIT_AND_EXPR, TREE_TYPE (t), t, |
5e3c5739 | 5672 | build_int_2 (-2 * UNITS_PER_WORD, -1)); |
5673 | t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t); | |
5674 | TREE_SIDE_EFFECTS (t) = 1; | |
ac965869 | 5675 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); |
5e3c5739 | 5676 | } |
72899a61 | 5677 | |
ac965869 | 5678 | if (size > 0) |
5679 | return std_expand_builtin_va_arg (valist, type); | |
5680 | else | |
5681 | { | |
5682 | ptr = build_pointer_type (type); | |
72899a61 | 5683 | |
ac965869 | 5684 | /* Args grow upward. */ |
5685 | t = build (POSTINCREMENT_EXPR, TREE_TYPE (valist), valist, | |
5686 | build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0)); | |
5687 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5688 | |
ac965869 | 5689 | pptr = build_pointer_type (ptr); |
5690 | t = build1 (NOP_EXPR, pptr, t); | |
5691 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5692 | |
ac965869 | 5693 | t = build1 (INDIRECT_REF, ptr, t); |
5694 | TREE_SIDE_EFFECTS (t) = 1; | |
5695 | } | |
72899a61 | 5696 | } |
ac965869 | 5697 | else /* !TARGET_64BIT */ |
72899a61 | 5698 | { |
ac965869 | 5699 | ptr = build_pointer_type (type); |
72899a61 | 5700 | |
ac965869 | 5701 | /* "Large" and variable sized types are passed by reference. */ |
5702 | if (size > 8 || size <= 0) | |
5703 | { | |
5704 | /* Args grow downward. */ | |
5705 | t = build (PREDECREMENT_EXPR, TREE_TYPE (valist), valist, | |
5706 | build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0)); | |
5707 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5708 | |
ac965869 | 5709 | pptr = build_pointer_type (ptr); |
5710 | t = build1 (NOP_EXPR, pptr, t); | |
5711 | TREE_SIDE_EFFECTS (t) = 1; | |
9840d99d | 5712 | |
ac965869 | 5713 | t = build1 (INDIRECT_REF, ptr, t); |
72899a61 | 5714 | TREE_SIDE_EFFECTS (t) = 1; |
5715 | } | |
ac965869 | 5716 | else |
5717 | { | |
5718 | t = build (PLUS_EXPR, TREE_TYPE (valist), valist, | |
5719 | build_int_2 (-size, -1)); | |
5720 | ||
5721 | /* Copied from va-pa.h, but we probably don't need to align to | |
5722 | word size, since we generate and preserve that invariant. */ | |
5723 | t = build (BIT_AND_EXPR, TREE_TYPE (valist), t, | |
5724 | build_int_2 ((size > 4 ? -8 : -4), -1)); | |
5725 | ||
5726 | t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t); | |
5727 | TREE_SIDE_EFFECTS (t) = 1; | |
5728 | ||
5729 | ofs = (8 - size) % 4; | |
5730 | if (ofs) | |
5731 | { | |
5732 | t = build (PLUS_EXPR, TREE_TYPE (valist), t, | |
5733 | build_int_2 (ofs, 0)); | |
5734 | TREE_SIDE_EFFECTS (t) = 1; | |
5735 | } | |
72899a61 | 5736 | |
ac965869 | 5737 | t = build1 (NOP_EXPR, ptr, t); |
5738 | TREE_SIDE_EFFECTS (t) = 1; | |
5739 | } | |
72899a61 | 5740 | } |
5741 | ||
5742 | /* Calculate! */ | |
ac965869 | 5743 | return expand_expr (t, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
72899a61 | 5744 | } |
5745 | ||
5746 | ||
5747 | ||
6d36483b | 5748 | /* This routine handles all the normal conditional branch sequences we |
5749 | might need to generate. It handles compare immediate vs compare | |
5750 | register, nullification of delay slots, varying length branches, | |
0d986529 | 5751 | negated branches, and all combinations of the above. It returns the |
6d36483b | 5752 | output appropriate to emit the branch corresponding to all given |
0d986529 | 5753 | parameters. */ |
5754 | ||
611a88e1 | 5755 | const char * |
5c1d8983 | 5756 | output_cbranch (rtx *operands, int nullify, int length, int negated, rtx insn) |
29a4502c | 5757 | { |
0d986529 | 5758 | static char buf[100]; |
5759 | int useskip = 0; | |
e9ec370e | 5760 | rtx xoperands[5]; |
0d986529 | 5761 | |
ece0fa59 | 5762 | /* A conditional branch to the following instruction (eg the delay slot) |
5763 | is asking for a disaster. This can happen when not optimizing and | |
5764 | when jump optimization fails. | |
29a4502c | 5765 | |
38efba6c | 5766 | While it is usually safe to emit nothing, this can fail if the |
5767 | preceding instruction is a nullified branch with an empty delay | |
5768 | slot and the same branch target as this branch. We could check | |
5769 | for this but jump optimization should eliminate nop jumps. It | |
5770 | is always safe to emit a nop. */ | |
ece0fa59 | 5771 | if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn)) |
5772 | return "nop"; | |
6d36483b | 5773 | |
22699a7e | 5774 | /* The doubleword form of the cmpib instruction doesn't have the LEU |
5775 | and GTU conditions while the cmpb instruction does. Since we accept | |
5776 | zero for cmpb, we must ensure that we use cmpb for the comparison. */ | |
5777 | if (GET_MODE (operands[1]) == DImode && operands[2] == const0_rtx) | |
5778 | operands[2] = gen_rtx_REG (DImode, 0); | |
5779 | ||
5fbd5940 | 5780 | /* If this is a long branch with its delay slot unfilled, set `nullify' |
5781 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 5782 | if (length == 8 && dbr_sequence_length () == 0) |
5fbd5940 | 5783 | nullify = 1; |
5784 | ||
5785 | /* If this is a short forward conditional branch which did not get | |
5786 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 5787 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
5fbd5940 | 5788 | nullify = forward_branch_p (insn); |
5789 | ||
6d36483b | 5790 | /* A forward branch over a single nullified insn can be done with a |
0d986529 | 5791 | comclr instruction. This avoids a single cycle penalty due to |
5792 | mis-predicted branch if we fall through (branch not taken). */ | |
5a1231ef | 5793 | if (length == 4 |
5fbd5940 | 5794 | && next_real_insn (insn) != 0 |
5a1231ef | 5795 | && get_attr_length (next_real_insn (insn)) == 4 |
5fbd5940 | 5796 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) |
0d986529 | 5797 | && nullify) |
5798 | useskip = 1; | |
5799 | ||
5800 | switch (length) | |
5801 | { | |
5fbd5940 | 5802 | /* All short conditional branches except backwards with an unfilled |
5803 | delay slot. */ | |
5a1231ef | 5804 | case 4: |
0d986529 | 5805 | if (useskip) |
e4065f95 | 5806 | strcpy (buf, "{com%I2clr,|cmp%I2clr,}"); |
0d986529 | 5807 | else |
e4065f95 | 5808 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5809 | if (GET_MODE (operands[1]) == DImode) |
5810 | strcat (buf, "*"); | |
0d986529 | 5811 | if (negated) |
5812 | strcat (buf, "%B3"); | |
5813 | else | |
5814 | strcat (buf, "%S3"); | |
5815 | if (useskip) | |
5a811d43 | 5816 | strcat (buf, " %2,%r1,%%r0"); |
0d986529 | 5817 | else if (nullify) |
9e49be0e | 5818 | strcat (buf, ",n %2,%r1,%0"); |
6d36483b | 5819 | else |
9e49be0e | 5820 | strcat (buf, " %2,%r1,%0"); |
0d986529 | 5821 | break; |
5822 | ||
87fcb603 | 5823 | /* All long conditionals. Note a short backward branch with an |
5fbd5940 | 5824 | unfilled delay slot is treated just like a long backward branch |
5825 | with an unfilled delay slot. */ | |
5a1231ef | 5826 | case 8: |
5fbd5940 | 5827 | /* Handle weird backwards branch with a filled delay slot |
5828 | with is nullified. */ | |
5829 | if (dbr_sequence_length () != 0 | |
5830 | && ! forward_branch_p (insn) | |
5831 | && nullify) | |
5832 | { | |
e4065f95 | 5833 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5834 | if (GET_MODE (operands[1]) == DImode) |
5835 | strcat (buf, "*"); | |
5fbd5940 | 5836 | if (negated) |
5837 | strcat (buf, "%S3"); | |
5838 | else | |
5839 | strcat (buf, "%B3"); | |
5a811d43 | 5840 | strcat (buf, ",n %2,%r1,.+12\n\tb %0"); |
5fbd5940 | 5841 | } |
43f0c1f2 | 5842 | /* Handle short backwards branch with an unfilled delay slot. |
5843 | Using a comb;nop rather than comiclr;bl saves 1 cycle for both | |
5844 | taken and untaken branches. */ | |
5845 | else if (dbr_sequence_length () == 0 | |
5846 | && ! forward_branch_p (insn) | |
47fc0706 | 5847 | && INSN_ADDRESSES_SET_P () |
5848 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
5849 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 5850 | { |
e4065f95 | 5851 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5852 | if (GET_MODE (operands[1]) == DImode) |
5853 | strcat (buf, "*"); | |
43f0c1f2 | 5854 | if (negated) |
9e49be0e | 5855 | strcat (buf, "%B3 %2,%r1,%0%#"); |
43f0c1f2 | 5856 | else |
9e49be0e | 5857 | strcat (buf, "%S3 %2,%r1,%0%#"); |
43f0c1f2 | 5858 | } |
0d986529 | 5859 | else |
5fbd5940 | 5860 | { |
e4065f95 | 5861 | strcpy (buf, "{com%I2clr,|cmp%I2clr,}"); |
5e3c5739 | 5862 | if (GET_MODE (operands[1]) == DImode) |
5863 | strcat (buf, "*"); | |
5fbd5940 | 5864 | if (negated) |
5865 | strcat (buf, "%S3"); | |
5866 | else | |
5867 | strcat (buf, "%B3"); | |
5868 | if (nullify) | |
5a811d43 | 5869 | strcat (buf, " %2,%r1,%%r0\n\tb,n %0"); |
5fbd5940 | 5870 | else |
5a811d43 | 5871 | strcat (buf, " %2,%r1,%%r0\n\tb %0"); |
5fbd5940 | 5872 | } |
0d986529 | 5873 | break; |
5874 | ||
c8a0e52b | 5875 | case 20: |
e9ec370e | 5876 | case 28: |
5877 | xoperands[0] = operands[0]; | |
5878 | xoperands[1] = operands[1]; | |
5879 | xoperands[2] = operands[2]; | |
5880 | xoperands[3] = operands[3]; | |
5881 | ||
5882 | /* The reversed conditional branch must branch over one additional | |
5883 | instruction if the delay slot is filled. If the delay slot | |
5884 | is empty, the instruction after the reversed condition branch | |
5885 | must be nullified. */ | |
5886 | nullify = dbr_sequence_length () == 0; | |
5887 | xoperands[4] = nullify ? GEN_INT (length) : GEN_INT (length + 4); | |
c8a0e52b | 5888 | |
5889 | /* Create a reversed conditional branch which branches around | |
5890 | the following insns. */ | |
e9ec370e | 5891 | if (GET_MODE (operands[1]) != DImode) |
5892 | { | |
5893 | if (nullify) | |
5894 | { | |
5895 | if (negated) | |
5896 | strcpy (buf, | |
5897 | "{com%I2b,%S3,n %2,%r1,.+%4|cmp%I2b,%S3,n %2,%r1,.+%4}"); | |
5898 | else | |
5899 | strcpy (buf, | |
5900 | "{com%I2b,%B3,n %2,%r1,.+%4|cmp%I2b,%B3,n %2,%r1,.+%4}"); | |
5901 | } | |
5902 | else | |
5903 | { | |
5904 | if (negated) | |
5905 | strcpy (buf, | |
5906 | "{com%I2b,%S3 %2,%r1,.+%4|cmp%I2b,%S3 %2,%r1,.+%4}"); | |
5907 | else | |
5908 | strcpy (buf, | |
5909 | "{com%I2b,%B3 %2,%r1,.+%4|cmp%I2b,%B3 %2,%r1,.+%4}"); | |
5910 | } | |
5911 | } | |
c8a0e52b | 5912 | else |
5e3c5739 | 5913 | { |
e9ec370e | 5914 | if (nullify) |
5915 | { | |
5916 | if (negated) | |
5917 | strcpy (buf, | |
5918 | "{com%I2b,*%S3,n %2,%r1,.+%4|cmp%I2b,*%S3,n %2,%r1,.+%4}"); | |
5919 | else | |
5920 | strcpy (buf, | |
5921 | "{com%I2b,*%B3,n %2,%r1,.+%4|cmp%I2b,*%B3,n %2,%r1,.+%4}"); | |
5922 | } | |
5e3c5739 | 5923 | else |
e9ec370e | 5924 | { |
5925 | if (negated) | |
5926 | strcpy (buf, | |
5927 | "{com%I2b,*%S3 %2,%r1,.+%4|cmp%I2b,*%S3 %2,%r1,.+%4}"); | |
5928 | else | |
5929 | strcpy (buf, | |
5930 | "{com%I2b,*%B3 %2,%r1,.+%4|cmp%I2b,*%B3 %2,%r1,.+%4}"); | |
5931 | } | |
5e3c5739 | 5932 | } |
c8a0e52b | 5933 | |
e9ec370e | 5934 | output_asm_insn (buf, xoperands); |
5935 | return output_lbranch (operands[0], insn); | |
c8a0e52b | 5936 | |
e9ec370e | 5937 | default: |
5938 | abort (); | |
5939 | } | |
5940 | return buf; | |
5941 | } | |
c8a0e52b | 5942 | |
e9ec370e | 5943 | /* This routine handles long unconditional branches that exceed the |
5944 | maximum range of a simple branch instruction. */ | |
c8a0e52b | 5945 | |
e9ec370e | 5946 | const char * |
5c1d8983 | 5947 | output_lbranch (rtx dest, rtx insn) |
e9ec370e | 5948 | { |
5949 | rtx xoperands[2]; | |
5950 | ||
5951 | xoperands[0] = dest; | |
c8a0e52b | 5952 | |
e9ec370e | 5953 | /* First, free up the delay slot. */ |
5954 | if (dbr_sequence_length () != 0) | |
5955 | { | |
5956 | /* We can't handle a jump in the delay slot. */ | |
5957 | if (GET_CODE (NEXT_INSN (insn)) == JUMP_INSN) | |
5958 | abort (); | |
c8a0e52b | 5959 | |
e9ec370e | 5960 | final_scan_insn (NEXT_INSN (insn), asm_out_file, |
5961 | optimize, 0, 0); | |
c8a0e52b | 5962 | |
e9ec370e | 5963 | /* Now delete the delay insn. */ |
5964 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
5965 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
5966 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
5967 | } | |
c8a0e52b | 5968 | |
e9ec370e | 5969 | /* Output an insn to save %r1. The runtime documentation doesn't |
5970 | specify whether the "Clean Up" slot in the callers frame can | |
5971 | be clobbered by the callee. It isn't copied by HP's builtin | |
5972 | alloca, so this suggests that it can be clobbered if necessary. | |
5973 | The "Static Link" location is copied by HP builtin alloca, so | |
5974 | we avoid using it. Using the cleanup slot might be a problem | |
5975 | if we have to interoperate with languages that pass cleanup | |
5976 | information. However, it should be possible to handle these | |
5977 | situations with GCC's asm feature. | |
5978 | ||
5979 | The "Current RP" slot is reserved for the called procedure, so | |
5980 | we try to use it when we don't have a frame of our own. It's | |
5981 | rather unlikely that we won't have a frame when we need to emit | |
5982 | a very long branch. | |
5983 | ||
5984 | Really the way to go long term is a register scavenger; goto | |
5985 | the target of the jump and find a register which we can use | |
5986 | as a scratch to hold the value in %r1. Then, we wouldn't have | |
5987 | to free up the delay slot or clobber a slot that may be needed | |
5988 | for other purposes. */ | |
5989 | if (TARGET_64BIT) | |
5990 | { | |
5991 | if (actual_fsize == 0 && !regs_ever_live[2]) | |
5992 | /* Use the return pointer slot in the frame marker. */ | |
5993 | output_asm_insn ("std %%r1,-16(%%r30)", xoperands); | |
5994 | else | |
5995 | /* Use the slot at -40 in the frame marker since HP builtin | |
5996 | alloca doesn't copy it. */ | |
5997 | output_asm_insn ("std %%r1,-40(%%r30)", xoperands); | |
5998 | } | |
5999 | else | |
6000 | { | |
6001 | if (actual_fsize == 0 && !regs_ever_live[2]) | |
6002 | /* Use the return pointer slot in the frame marker. */ | |
6003 | output_asm_insn ("stw %%r1,-20(%%r30)", xoperands); | |
6004 | else | |
6005 | /* Use the "Clean Up" slot in the frame marker. In GCC, | |
6006 | the only other use of this location is for copying a | |
6007 | floating point double argument from a floating-point | |
6008 | register to two general registers. The copy is done | |
19ee40ed | 6009 | as an "atomic" operation when outputting a call, so it |
e9ec370e | 6010 | won't interfere with our using the location here. */ |
6011 | output_asm_insn ("stw %%r1,-12(%%r30)", xoperands); | |
6012 | } | |
b70ea764 | 6013 | |
2247cc5f | 6014 | if (TARGET_PORTABLE_RUNTIME) |
6015 | { | |
6016 | output_asm_insn ("ldil L'%0,%%r1", xoperands); | |
6017 | output_asm_insn ("ldo R'%0(%%r1),%%r1", xoperands); | |
6018 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
6019 | } | |
6020 | else if (flag_pic) | |
e9ec370e | 6021 | { |
6022 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
6023 | if (TARGET_SOM || !TARGET_GAS) | |
6024 | { | |
6025 | xoperands[1] = gen_label_rtx (); | |
6026 | output_asm_insn ("addil L'%l0-%l1,%%r1", xoperands); | |
6027 | (*targetm.asm_out.internal_label) (asm_out_file, "L", | |
6028 | CODE_LABEL_NUMBER (xoperands[1])); | |
6029 | output_asm_insn ("ldo R'%l0-%l1(%%r1),%%r1", xoperands); | |
c8a0e52b | 6030 | } |
e9ec370e | 6031 | else |
6032 | { | |
6033 | output_asm_insn ("addil L'%l0-$PIC_pcrel$0+4,%%r1", xoperands); | |
6034 | output_asm_insn ("ldo R'%l0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands); | |
6035 | } | |
6036 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
6037 | } | |
6038 | else | |
6039 | /* Now output a very long branch to the original target. */ | |
6040 | output_asm_insn ("ldil L'%l0,%%r1\n\tbe R'%l0(%%sr4,%%r1)", xoperands); | |
c8a0e52b | 6041 | |
e9ec370e | 6042 | /* Now restore the value of %r1 in the delay slot. */ |
6043 | if (TARGET_64BIT) | |
6044 | { | |
6045 | if (actual_fsize == 0 && !regs_ever_live[2]) | |
6046 | return "ldd -16(%%r30),%%r1"; | |
6047 | else | |
6048 | return "ldd -40(%%r30),%%r1"; | |
6049 | } | |
6050 | else | |
6051 | { | |
6052 | if (actual_fsize == 0 && !regs_ever_live[2]) | |
6053 | return "ldw -20(%%r30),%%r1"; | |
6054 | else | |
6055 | return "ldw -12(%%r30),%%r1"; | |
5fbd5940 | 6056 | } |
0d986529 | 6057 | } |
6058 | ||
6d36483b | 6059 | /* This routine handles all the branch-on-bit conditional branch sequences we |
0d986529 | 6060 | might need to generate. It handles nullification of delay slots, |
6061 | varying length branches, negated branches and all combinations of the | |
6062 | above. it returns the appropriate output template to emit the branch. */ | |
6063 | ||
611a88e1 | 6064 | const char * |
5c1d8983 | 6065 | output_bb (rtx *operands ATTRIBUTE_UNUSED, int nullify, int length, |
6066 | int negated, rtx insn, int which) | |
29a4502c | 6067 | { |
0d986529 | 6068 | static char buf[100]; |
6069 | int useskip = 0; | |
6070 | ||
29a4502c | 6071 | /* A conditional branch to the following instruction (eg the delay slot) is |
6072 | asking for a disaster. I do not think this can happen as this pattern | |
6d36483b | 6073 | is only used when optimizing; jump optimization should eliminate the |
29a4502c | 6074 | jump. But be prepared just in case. */ |
6d36483b | 6075 | |
ece0fa59 | 6076 | if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn)) |
6077 | return "nop"; | |
6d36483b | 6078 | |
5fbd5940 | 6079 | /* If this is a long branch with its delay slot unfilled, set `nullify' |
6080 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 6081 | if (length == 8 && dbr_sequence_length () == 0) |
5fbd5940 | 6082 | nullify = 1; |
6083 | ||
6084 | /* If this is a short forward conditional branch which did not get | |
6085 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 6086 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
5fbd5940 | 6087 | nullify = forward_branch_p (insn); |
6088 | ||
6d36483b | 6089 | /* A forward branch over a single nullified insn can be done with a |
0d986529 | 6090 | extrs instruction. This avoids a single cycle penalty due to |
6091 | mis-predicted branch if we fall through (branch not taken). */ | |
6092 | ||
5a1231ef | 6093 | if (length == 4 |
5fbd5940 | 6094 | && next_real_insn (insn) != 0 |
5a1231ef | 6095 | && get_attr_length (next_real_insn (insn)) == 4 |
5fbd5940 | 6096 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) |
0d986529 | 6097 | && nullify) |
6098 | useskip = 1; | |
6099 | ||
6100 | switch (length) | |
6101 | { | |
6102 | ||
5fbd5940 | 6103 | /* All short conditional branches except backwards with an unfilled |
6104 | delay slot. */ | |
5a1231ef | 6105 | case 4: |
0d986529 | 6106 | if (useskip) |
e4065f95 | 6107 | strcpy (buf, "{extrs,|extrw,s,}"); |
6d36483b | 6108 | else |
0d986529 | 6109 | strcpy (buf, "bb,"); |
5e3c5739 | 6110 | if (useskip && GET_MODE (operands[0]) == DImode) |
6111 | strcpy (buf, "extrd,s,*"); | |
6112 | else if (GET_MODE (operands[0]) == DImode) | |
6113 | strcpy (buf, "bb,*"); | |
0d986529 | 6114 | if ((which == 0 && negated) |
6115 | || (which == 1 && ! negated)) | |
6116 | strcat (buf, ">="); | |
6117 | else | |
6118 | strcat (buf, "<"); | |
6119 | if (useskip) | |
5a811d43 | 6120 | strcat (buf, " %0,%1,1,%%r0"); |
0d986529 | 6121 | else if (nullify && negated) |
6122 | strcat (buf, ",n %0,%1,%3"); | |
6123 | else if (nullify && ! negated) | |
6124 | strcat (buf, ",n %0,%1,%2"); | |
6125 | else if (! nullify && negated) | |
5fbd5940 | 6126 | strcat (buf, "%0,%1,%3"); |
0d986529 | 6127 | else if (! nullify && ! negated) |
5fbd5940 | 6128 | strcat (buf, " %0,%1,%2"); |
0d986529 | 6129 | break; |
6130 | ||
87fcb603 | 6131 | /* All long conditionals. Note a short backward branch with an |
5fbd5940 | 6132 | unfilled delay slot is treated just like a long backward branch |
6133 | with an unfilled delay slot. */ | |
5a1231ef | 6134 | case 8: |
5fbd5940 | 6135 | /* Handle weird backwards branch with a filled delay slot |
6136 | with is nullified. */ | |
6137 | if (dbr_sequence_length () != 0 | |
6138 | && ! forward_branch_p (insn) | |
6139 | && nullify) | |
6140 | { | |
6141 | strcpy (buf, "bb,"); | |
5e3c5739 | 6142 | if (GET_MODE (operands[0]) == DImode) |
6143 | strcat (buf, "*"); | |
5fbd5940 | 6144 | if ((which == 0 && negated) |
6145 | || (which == 1 && ! negated)) | |
6146 | strcat (buf, "<"); | |
6147 | else | |
6148 | strcat (buf, ">="); | |
6149 | if (negated) | |
5a811d43 | 6150 | strcat (buf, ",n %0,%1,.+12\n\tb %3"); |
5fbd5940 | 6151 | else |
5a811d43 | 6152 | strcat (buf, ",n %0,%1,.+12\n\tb %2"); |
5fbd5940 | 6153 | } |
43f0c1f2 | 6154 | /* Handle short backwards branch with an unfilled delay slot. |
6155 | Using a bb;nop rather than extrs;bl saves 1 cycle for both | |
6156 | taken and untaken branches. */ | |
6157 | else if (dbr_sequence_length () == 0 | |
6158 | && ! forward_branch_p (insn) | |
47fc0706 | 6159 | && INSN_ADDRESSES_SET_P () |
6160 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6161 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 6162 | { |
6163 | strcpy (buf, "bb,"); | |
5e3c5739 | 6164 | if (GET_MODE (operands[0]) == DImode) |
6165 | strcat (buf, "*"); | |
43f0c1f2 | 6166 | if ((which == 0 && negated) |
6167 | || (which == 1 && ! negated)) | |
6168 | strcat (buf, ">="); | |
6169 | else | |
6170 | strcat (buf, "<"); | |
6171 | if (negated) | |
6172 | strcat (buf, " %0,%1,%3%#"); | |
6173 | else | |
6174 | strcat (buf, " %0,%1,%2%#"); | |
6175 | } | |
0d986529 | 6176 | else |
5fbd5940 | 6177 | { |
e4065f95 | 6178 | strcpy (buf, "{extrs,|extrw,s,}"); |
5e3c5739 | 6179 | if (GET_MODE (operands[0]) == DImode) |
6180 | strcpy (buf, "extrd,s,*"); | |
5fbd5940 | 6181 | if ((which == 0 && negated) |
6182 | || (which == 1 && ! negated)) | |
6183 | strcat (buf, "<"); | |
6184 | else | |
6185 | strcat (buf, ">="); | |
6186 | if (nullify && negated) | |
c6ae275c | 6187 | strcat (buf, " %0,%1,1,%%r0\n\tb,n %3"); |
5fbd5940 | 6188 | else if (nullify && ! negated) |
c6ae275c | 6189 | strcat (buf, " %0,%1,1,%%r0\n\tb,n %2"); |
5fbd5940 | 6190 | else if (negated) |
5a811d43 | 6191 | strcat (buf, " %0,%1,1,%%r0\n\tb %3"); |
6d36483b | 6192 | else |
5a811d43 | 6193 | strcat (buf, " %0,%1,1,%%r0\n\tb %2"); |
5fbd5940 | 6194 | } |
0d986529 | 6195 | break; |
6196 | ||
6197 | default: | |
ea52c577 | 6198 | abort (); |
5fbd5940 | 6199 | } |
0d986529 | 6200 | return buf; |
6201 | } | |
6202 | ||
c7a4e712 | 6203 | /* This routine handles all the branch-on-variable-bit conditional branch |
6204 | sequences we might need to generate. It handles nullification of delay | |
6205 | slots, varying length branches, negated branches and all combinations | |
6206 | of the above. it returns the appropriate output template to emit the | |
6207 | branch. */ | |
6208 | ||
611a88e1 | 6209 | const char * |
5c1d8983 | 6210 | output_bvb (rtx *operands ATTRIBUTE_UNUSED, int nullify, int length, |
6211 | int negated, rtx insn, int which) | |
c7a4e712 | 6212 | { |
6213 | static char buf[100]; | |
6214 | int useskip = 0; | |
6215 | ||
6216 | /* A conditional branch to the following instruction (eg the delay slot) is | |
6217 | asking for a disaster. I do not think this can happen as this pattern | |
6218 | is only used when optimizing; jump optimization should eliminate the | |
6219 | jump. But be prepared just in case. */ | |
6220 | ||
ece0fa59 | 6221 | if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn)) |
6222 | return "nop"; | |
c7a4e712 | 6223 | |
6224 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
6225 | as it can nullify the delay slot and save a nop. */ | |
6226 | if (length == 8 && dbr_sequence_length () == 0) | |
6227 | nullify = 1; | |
6228 | ||
6229 | /* If this is a short forward conditional branch which did not get | |
6230 | its delay slot filled, the delay slot can still be nullified. */ | |
6231 | if (! nullify && length == 4 && dbr_sequence_length () == 0) | |
6232 | nullify = forward_branch_p (insn); | |
6233 | ||
6234 | /* A forward branch over a single nullified insn can be done with a | |
6235 | extrs instruction. This avoids a single cycle penalty due to | |
6236 | mis-predicted branch if we fall through (branch not taken). */ | |
6237 | ||
6238 | if (length == 4 | |
6239 | && next_real_insn (insn) != 0 | |
6240 | && get_attr_length (next_real_insn (insn)) == 4 | |
6241 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) | |
6242 | && nullify) | |
6243 | useskip = 1; | |
6244 | ||
6245 | switch (length) | |
6246 | { | |
6247 | ||
6248 | /* All short conditional branches except backwards with an unfilled | |
6249 | delay slot. */ | |
6250 | case 4: | |
6251 | if (useskip) | |
e4065f95 | 6252 | strcpy (buf, "{vextrs,|extrw,s,}"); |
c7a4e712 | 6253 | else |
e4065f95 | 6254 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 6255 | if (useskip && GET_MODE (operands[0]) == DImode) |
e75269fd | 6256 | strcpy (buf, "extrd,s,*"); |
5e3c5739 | 6257 | else if (GET_MODE (operands[0]) == DImode) |
6258 | strcpy (buf, "bb,*"); | |
c7a4e712 | 6259 | if ((which == 0 && negated) |
6260 | || (which == 1 && ! negated)) | |
6261 | strcat (buf, ">="); | |
6262 | else | |
6263 | strcat (buf, "<"); | |
6264 | if (useskip) | |
e4065f95 | 6265 | strcat (buf, "{ %0,1,%%r0| %0,%%sar,1,%%r0}"); |
c7a4e712 | 6266 | else if (nullify && negated) |
e4065f95 | 6267 | strcat (buf, "{,n %0,%3|,n %0,%%sar,%3}"); |
c7a4e712 | 6268 | else if (nullify && ! negated) |
e4065f95 | 6269 | strcat (buf, "{,n %0,%2|,n %0,%%sar,%2}"); |
c7a4e712 | 6270 | else if (! nullify && negated) |
e4065f95 | 6271 | strcat (buf, "{%0,%3|%0,%%sar,%3}"); |
c7a4e712 | 6272 | else if (! nullify && ! negated) |
e4065f95 | 6273 | strcat (buf, "{ %0,%2| %0,%%sar,%2}"); |
c7a4e712 | 6274 | break; |
6275 | ||
87fcb603 | 6276 | /* All long conditionals. Note a short backward branch with an |
c7a4e712 | 6277 | unfilled delay slot is treated just like a long backward branch |
6278 | with an unfilled delay slot. */ | |
6279 | case 8: | |
6280 | /* Handle weird backwards branch with a filled delay slot | |
6281 | with is nullified. */ | |
6282 | if (dbr_sequence_length () != 0 | |
6283 | && ! forward_branch_p (insn) | |
6284 | && nullify) | |
6285 | { | |
e4065f95 | 6286 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 6287 | if (GET_MODE (operands[0]) == DImode) |
6288 | strcat (buf, "*"); | |
c7a4e712 | 6289 | if ((which == 0 && negated) |
6290 | || (which == 1 && ! negated)) | |
6291 | strcat (buf, "<"); | |
6292 | else | |
6293 | strcat (buf, ">="); | |
6294 | if (negated) | |
e4065f95 | 6295 | strcat (buf, "{,n %0,.+12\n\tb %3|,n %0,%%sar,.+12\n\tb %3}"); |
c7a4e712 | 6296 | else |
e4065f95 | 6297 | strcat (buf, "{,n %0,.+12\n\tb %2|,n %0,%%sar,.+12\n\tb %2}"); |
c7a4e712 | 6298 | } |
6299 | /* Handle short backwards branch with an unfilled delay slot. | |
6300 | Using a bb;nop rather than extrs;bl saves 1 cycle for both | |
6301 | taken and untaken branches. */ | |
6302 | else if (dbr_sequence_length () == 0 | |
6303 | && ! forward_branch_p (insn) | |
47fc0706 | 6304 | && INSN_ADDRESSES_SET_P () |
6305 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6306 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
c7a4e712 | 6307 | { |
e4065f95 | 6308 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 6309 | if (GET_MODE (operands[0]) == DImode) |
6310 | strcat (buf, "*"); | |
c7a4e712 | 6311 | if ((which == 0 && negated) |
6312 | || (which == 1 && ! negated)) | |
6313 | strcat (buf, ">="); | |
6314 | else | |
6315 | strcat (buf, "<"); | |
6316 | if (negated) | |
e4065f95 | 6317 | strcat (buf, "{ %0,%3%#| %0,%%sar,%3%#}"); |
c7a4e712 | 6318 | else |
e4065f95 | 6319 | strcat (buf, "{ %0,%2%#| %0,%%sar,%2%#}"); |
c7a4e712 | 6320 | } |
6321 | else | |
6322 | { | |
e4065f95 | 6323 | strcpy (buf, "{vextrs,|extrw,s,}"); |
5e3c5739 | 6324 | if (GET_MODE (operands[0]) == DImode) |
6325 | strcpy (buf, "extrd,s,*"); | |
c7a4e712 | 6326 | if ((which == 0 && negated) |
6327 | || (which == 1 && ! negated)) | |
6328 | strcat (buf, "<"); | |
6329 | else | |
6330 | strcat (buf, ">="); | |
6331 | if (nullify && negated) | |
e4065f95 | 6332 | strcat (buf, "{ %0,1,%%r0\n\tb,n %3| %0,%%sar,1,%%r0\n\tb,n %3}"); |
c7a4e712 | 6333 | else if (nullify && ! negated) |
e4065f95 | 6334 | strcat (buf, "{ %0,1,%%r0\n\tb,n %2| %0,%%sar,1,%%r0\n\tb,n %2}"); |
c7a4e712 | 6335 | else if (negated) |
e4065f95 | 6336 | strcat (buf, "{ %0,1,%%r0\n\tb %3| %0,%%sar,1,%%r0\n\tb %3}"); |
c7a4e712 | 6337 | else |
e4065f95 | 6338 | strcat (buf, "{ %0,1,%%r0\n\tb %2| %0,%%sar,1,%%r0\n\tb %2}"); |
c7a4e712 | 6339 | } |
6340 | break; | |
6341 | ||
6342 | default: | |
ea52c577 | 6343 | abort (); |
c7a4e712 | 6344 | } |
6345 | return buf; | |
6346 | } | |
6347 | ||
29a4502c | 6348 | /* Return the output template for emitting a dbra type insn. |
6349 | ||
6350 | Note it may perform some output operations on its own before | |
6351 | returning the final output string. */ | |
611a88e1 | 6352 | const char * |
5c1d8983 | 6353 | output_dbra (rtx *operands, rtx insn, int which_alternative) |
29a4502c | 6354 | { |
6355 | ||
6356 | /* A conditional branch to the following instruction (eg the delay slot) is | |
6357 | asking for a disaster. Be prepared! */ | |
6358 | ||
ece0fa59 | 6359 | if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn)) |
29a4502c | 6360 | { |
6361 | if (which_alternative == 0) | |
6362 | return "ldo %1(%0),%0"; | |
6363 | else if (which_alternative == 1) | |
6364 | { | |
ea52c577 | 6365 | output_asm_insn ("{fstws|fstw} %0,-16(%%r30)", operands); |
6366 | output_asm_insn ("ldw -16(%%r30),%4", operands); | |
34940871 | 6367 | output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands); |
e4065f95 | 6368 | return "{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6369 | } |
6370 | else | |
6371 | { | |
6372 | output_asm_insn ("ldw %0,%4", operands); | |
6373 | return "ldo %1(%4),%4\n\tstw %4,%0"; | |
6374 | } | |
6375 | } | |
6376 | ||
6377 | if (which_alternative == 0) | |
6378 | { | |
6379 | int nullify = INSN_ANNULLED_BRANCH_P (insn); | |
6380 | int length = get_attr_length (insn); | |
6381 | ||
6382 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
6383 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 6384 | if (length == 8 && dbr_sequence_length () == 0) |
29a4502c | 6385 | nullify = 1; |
6386 | ||
6387 | /* If this is a short forward conditional branch which did not get | |
6388 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 6389 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
29a4502c | 6390 | nullify = forward_branch_p (insn); |
6391 | ||
6392 | /* Handle short versions first. */ | |
5a1231ef | 6393 | if (length == 4 && nullify) |
29a4502c | 6394 | return "addib,%C2,n %1,%0,%3"; |
5a1231ef | 6395 | else if (length == 4 && ! nullify) |
29a4502c | 6396 | return "addib,%C2 %1,%0,%3"; |
5a1231ef | 6397 | else if (length == 8) |
29a4502c | 6398 | { |
6d36483b | 6399 | /* Handle weird backwards branch with a fulled delay slot |
29a4502c | 6400 | which is nullified. */ |
6401 | if (dbr_sequence_length () != 0 | |
6402 | && ! forward_branch_p (insn) | |
6403 | && nullify) | |
5a811d43 | 6404 | return "addib,%N2,n %1,%0,.+12\n\tb %3"; |
43f0c1f2 | 6405 | /* Handle short backwards branch with an unfilled delay slot. |
6406 | Using a addb;nop rather than addi;bl saves 1 cycle for both | |
6407 | taken and untaken branches. */ | |
6408 | else if (dbr_sequence_length () == 0 | |
6409 | && ! forward_branch_p (insn) | |
47fc0706 | 6410 | && INSN_ADDRESSES_SET_P () |
6411 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6412 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 6413 | return "addib,%C2 %1,%0,%3%#"; |
6d36483b | 6414 | |
6415 | /* Handle normal cases. */ | |
29a4502c | 6416 | if (nullify) |
5a811d43 | 6417 | return "addi,%N2 %1,%0,%0\n\tb,n %3"; |
29a4502c | 6418 | else |
5a811d43 | 6419 | return "addi,%N2 %1,%0,%0\n\tb %3"; |
29a4502c | 6420 | } |
6421 | else | |
ea52c577 | 6422 | abort (); |
29a4502c | 6423 | } |
6424 | /* Deal with gross reload from FP register case. */ | |
6425 | else if (which_alternative == 1) | |
6426 | { | |
6427 | /* Move loop counter from FP register to MEM then into a GR, | |
6428 | increment the GR, store the GR into MEM, and finally reload | |
6d36483b | 6429 | the FP register from MEM from within the branch's delay slot. */ |
ea52c577 | 6430 | output_asm_insn ("{fstws|fstw} %0,-16(%%r30)\n\tldw -16(%%r30),%4", |
6431 | operands); | |
34940871 | 6432 | output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands); |
5a1231ef | 6433 | if (get_attr_length (insn) == 24) |
e4065f95 | 6434 | return "{comb|cmpb},%S2 %%r0,%4,%3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6435 | else |
e4065f95 | 6436 | return "{comclr|cmpclr},%B2 %%r0,%4,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6437 | } |
6438 | /* Deal with gross reload from memory case. */ | |
6439 | else | |
6440 | { | |
6441 | /* Reload loop counter from memory, the store back to memory | |
6442 | happens in the branch's delay slot. */ | |
6443 | output_asm_insn ("ldw %0,%4", operands); | |
5a1231ef | 6444 | if (get_attr_length (insn) == 12) |
29a4502c | 6445 | return "addib,%C2 %1,%4,%3\n\tstw %4,%0"; |
6446 | else | |
5a811d43 | 6447 | return "addi,%N2 %1,%4,%4\n\tb %3\n\tstw %4,%0"; |
29a4502c | 6448 | } |
6449 | } | |
6450 | ||
6451 | /* Return the output template for emitting a dbra type insn. | |
6452 | ||
6453 | Note it may perform some output operations on its own before | |
6454 | returning the final output string. */ | |
611a88e1 | 6455 | const char * |
5c1d8983 | 6456 | output_movb (rtx *operands, rtx insn, int which_alternative, |
6457 | int reverse_comparison) | |
29a4502c | 6458 | { |
6459 | ||
6460 | /* A conditional branch to the following instruction (eg the delay slot) is | |
6461 | asking for a disaster. Be prepared! */ | |
6462 | ||
ece0fa59 | 6463 | if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn)) |
29a4502c | 6464 | { |
6465 | if (which_alternative == 0) | |
6466 | return "copy %1,%0"; | |
6467 | else if (which_alternative == 1) | |
6468 | { | |
ea52c577 | 6469 | output_asm_insn ("stw %1,-16(%%r30)", operands); |
e4065f95 | 6470 | return "{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6471 | } |
546a40bd | 6472 | else if (which_alternative == 2) |
29a4502c | 6473 | return "stw %1,%0"; |
546a40bd | 6474 | else |
6475 | return "mtsar %r1"; | |
29a4502c | 6476 | } |
6477 | ||
6478 | /* Support the second variant. */ | |
6479 | if (reverse_comparison) | |
6480 | PUT_CODE (operands[2], reverse_condition (GET_CODE (operands[2]))); | |
6481 | ||
6482 | if (which_alternative == 0) | |
6483 | { | |
6484 | int nullify = INSN_ANNULLED_BRANCH_P (insn); | |
6485 | int length = get_attr_length (insn); | |
6486 | ||
6487 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
6488 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 6489 | if (length == 8 && dbr_sequence_length () == 0) |
29a4502c | 6490 | nullify = 1; |
6491 | ||
6492 | /* If this is a short forward conditional branch which did not get | |
6493 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 6494 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
29a4502c | 6495 | nullify = forward_branch_p (insn); |
6496 | ||
6497 | /* Handle short versions first. */ | |
5a1231ef | 6498 | if (length == 4 && nullify) |
29a4502c | 6499 | return "movb,%C2,n %1,%0,%3"; |
5a1231ef | 6500 | else if (length == 4 && ! nullify) |
29a4502c | 6501 | return "movb,%C2 %1,%0,%3"; |
5a1231ef | 6502 | else if (length == 8) |
29a4502c | 6503 | { |
6d36483b | 6504 | /* Handle weird backwards branch with a filled delay slot |
29a4502c | 6505 | which is nullified. */ |
6506 | if (dbr_sequence_length () != 0 | |
6507 | && ! forward_branch_p (insn) | |
6508 | && nullify) | |
5a811d43 | 6509 | return "movb,%N2,n %1,%0,.+12\n\tb %3"; |
6d36483b | 6510 | |
43f0c1f2 | 6511 | /* Handle short backwards branch with an unfilled delay slot. |
6512 | Using a movb;nop rather than or;bl saves 1 cycle for both | |
6513 | taken and untaken branches. */ | |
6514 | else if (dbr_sequence_length () == 0 | |
6515 | && ! forward_branch_p (insn) | |
47fc0706 | 6516 | && INSN_ADDRESSES_SET_P () |
6517 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6518 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 6519 | return "movb,%C2 %1,%0,%3%#"; |
6d36483b | 6520 | /* Handle normal cases. */ |
29a4502c | 6521 | if (nullify) |
5a811d43 | 6522 | return "or,%N2 %1,%%r0,%0\n\tb,n %3"; |
29a4502c | 6523 | else |
5a811d43 | 6524 | return "or,%N2 %1,%%r0,%0\n\tb %3"; |
29a4502c | 6525 | } |
6526 | else | |
ea52c577 | 6527 | abort (); |
29a4502c | 6528 | } |
6529 | /* Deal with gross reload from FP register case. */ | |
6530 | else if (which_alternative == 1) | |
6531 | { | |
6532 | /* Move loop counter from FP register to MEM then into a GR, | |
6533 | increment the GR, store the GR into MEM, and finally reload | |
6d36483b | 6534 | the FP register from MEM from within the branch's delay slot. */ |
ea52c577 | 6535 | output_asm_insn ("stw %1,-16(%%r30)", operands); |
5a1231ef | 6536 | if (get_attr_length (insn) == 12) |
e4065f95 | 6537 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6538 | else |
e4065f95 | 6539 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6540 | } |
6541 | /* Deal with gross reload from memory case. */ | |
546a40bd | 6542 | else if (which_alternative == 2) |
29a4502c | 6543 | { |
6544 | /* Reload loop counter from memory, the store back to memory | |
6545 | happens in the branch's delay slot. */ | |
5a1231ef | 6546 | if (get_attr_length (insn) == 8) |
e4065f95 | 6547 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\tstw %1,%0"; |
29a4502c | 6548 | else |
e4065f95 | 6549 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\tstw %1,%0"; |
29a4502c | 6550 | } |
546a40bd | 6551 | /* Handle SAR as a destination. */ |
6552 | else | |
6553 | { | |
6554 | if (get_attr_length (insn) == 8) | |
e4065f95 | 6555 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\tmtsar %r1"; |
546a40bd | 6556 | else |
e4065f95 | 6557 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tbl %3\n\tmtsar %r1"; |
546a40bd | 6558 | } |
29a4502c | 6559 | } |
6560 | ||
ece88821 | 6561 | /* Copy any FP arguments in INSN into integer registers. */ |
6562 | static void | |
5c1d8983 | 6563 | copy_fp_args (rtx insn) |
ece88821 | 6564 | { |
6565 | rtx link; | |
6566 | rtx xoperands[2]; | |
29a4502c | 6567 | |
ece88821 | 6568 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) |
6569 | { | |
6570 | int arg_mode, regno; | |
6571 | rtx use = XEXP (link, 0); | |
3683f840 | 6572 | |
ece88821 | 6573 | if (! (GET_CODE (use) == USE |
6574 | && GET_CODE (XEXP (use, 0)) == REG | |
6575 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
6576 | continue; | |
d6686e21 | 6577 | |
ece88821 | 6578 | arg_mode = GET_MODE (XEXP (use, 0)); |
6579 | regno = REGNO (XEXP (use, 0)); | |
5e3c5739 | 6580 | |
ece88821 | 6581 | /* Is it a floating point register? */ |
6582 | if (regno >= 32 && regno <= 39) | |
6583 | { | |
6584 | /* Copy the FP register into an integer register via memory. */ | |
6585 | if (arg_mode == SFmode) | |
6586 | { | |
6587 | xoperands[0] = XEXP (use, 0); | |
6588 | xoperands[1] = gen_rtx_REG (SImode, 26 - (regno - 32) / 2); | |
6589 | output_asm_insn ("{fstws|fstw} %0,-16(%%sr0,%%r30)", xoperands); | |
6590 | output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands); | |
6591 | } | |
6592 | else | |
6593 | { | |
6594 | xoperands[0] = XEXP (use, 0); | |
6595 | xoperands[1] = gen_rtx_REG (DImode, 25 - (regno - 34) / 2); | |
6596 | output_asm_insn ("{fstds|fstd} %0,-16(%%sr0,%%r30)", xoperands); | |
6597 | output_asm_insn ("ldw -12(%%sr0,%%r30),%R1", xoperands); | |
6598 | output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands); | |
6599 | } | |
6600 | } | |
06ddb6f8 | 6601 | } |
ece88821 | 6602 | } |
6603 | ||
6604 | /* Compute length of the FP argument copy sequence for INSN. */ | |
6605 | static int | |
5c1d8983 | 6606 | length_fp_args (rtx insn) |
ece88821 | 6607 | { |
6608 | int length = 0; | |
6609 | rtx link; | |
06ddb6f8 | 6610 | |
ece88821 | 6611 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) |
c7a4e712 | 6612 | { |
ece88821 | 6613 | int arg_mode, regno; |
6614 | rtx use = XEXP (link, 0); | |
6615 | ||
6616 | if (! (GET_CODE (use) == USE | |
6617 | && GET_CODE (XEXP (use, 0)) == REG | |
6618 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
6619 | continue; | |
c7a4e712 | 6620 | |
ece88821 | 6621 | arg_mode = GET_MODE (XEXP (use, 0)); |
6622 | regno = REGNO (XEXP (use, 0)); | |
6623 | ||
6624 | /* Is it a floating point register? */ | |
6625 | if (regno >= 32 && regno <= 39) | |
c7a4e712 | 6626 | { |
ece88821 | 6627 | if (arg_mode == SFmode) |
6628 | length += 8; | |
6629 | else | |
6630 | length += 12; | |
c7a4e712 | 6631 | } |
ece88821 | 6632 | } |
c7a4e712 | 6633 | |
ece88821 | 6634 | return length; |
6635 | } | |
b70ea764 | 6636 | |
cd0dfcc5 | 6637 | /* Return the attribute length for the millicode call instruction INSN. |
6638 | The length must match the code generated by output_millicode_call. | |
6639 | We include the delay slot in the returned length as it is better to | |
ece88821 | 6640 | over estimate the length than to under estimate it. */ |
b29897dd | 6641 | |
ece88821 | 6642 | int |
5c1d8983 | 6643 | attr_length_millicode_call (rtx insn) |
ece88821 | 6644 | { |
cd0dfcc5 | 6645 | unsigned long distance = -1; |
8a05c3c2 | 6646 | unsigned long total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes; |
ece88821 | 6647 | |
cd0dfcc5 | 6648 | if (INSN_ADDRESSES_SET_P ()) |
6649 | { | |
2247cc5f | 6650 | distance = (total + insn_current_reference_address (insn)); |
6651 | if (distance < total) | |
cd0dfcc5 | 6652 | distance = -1; |
6653 | } | |
ece88821 | 6654 | |
6655 | if (TARGET_64BIT) | |
6656 | { | |
6657 | if (!TARGET_LONG_CALLS && distance < 7600000) | |
cd0dfcc5 | 6658 | return 8; |
ece88821 | 6659 | |
cd0dfcc5 | 6660 | return 20; |
ece88821 | 6661 | } |
6662 | else if (TARGET_PORTABLE_RUNTIME) | |
cd0dfcc5 | 6663 | return 24; |
ece88821 | 6664 | else |
6665 | { | |
6666 | if (!TARGET_LONG_CALLS && distance < 240000) | |
cd0dfcc5 | 6667 | return 8; |
ece88821 | 6668 | |
6669 | if (TARGET_LONG_ABS_CALL && !flag_pic) | |
cd0dfcc5 | 6670 | return 12; |
ece88821 | 6671 | |
cd0dfcc5 | 6672 | return 24; |
ece88821 | 6673 | } |
6674 | } | |
6675 | ||
6676 | /* INSN is a function call. It may have an unconditional jump | |
6677 | in its delay slot. | |
b29897dd | 6678 | |
ece88821 | 6679 | CALL_DEST is the routine we are calling. */ |
b29897dd | 6680 | |
ece88821 | 6681 | const char * |
5c1d8983 | 6682 | output_millicode_call (rtx insn, rtx call_dest) |
ece88821 | 6683 | { |
6684 | int attr_length = get_attr_length (insn); | |
6685 | int seq_length = dbr_sequence_length (); | |
6686 | int distance; | |
6687 | rtx seq_insn; | |
6688 | rtx xoperands[3]; | |
b29897dd | 6689 | |
ece88821 | 6690 | xoperands[0] = call_dest; |
6691 | xoperands[2] = gen_rtx_REG (Pmode, TARGET_64BIT ? 2 : 31); | |
6692 | ||
6693 | /* Handle the common case where we are sure that the branch will | |
6694 | reach the beginning of the $CODE$ subspace. The within reach | |
6695 | form of the $$sh_func_adrs call has a length of 28. Because | |
a8b24921 | 6696 | it has an attribute type of multi, it never has a nonzero |
ece88821 | 6697 | sequence length. The length of the $$sh_func_adrs is the same |
6698 | as certain out of reach PIC calls to other routines. */ | |
6699 | if (!TARGET_LONG_CALLS | |
6700 | && ((seq_length == 0 | |
6701 | && (attr_length == 12 | |
6702 | || (attr_length == 28 && get_attr_type (insn) == TYPE_MULTI))) | |
6703 | || (seq_length != 0 && attr_length == 8))) | |
6704 | { | |
6705 | output_asm_insn ("{bl|b,l} %0,%2", xoperands); | |
6706 | } | |
6707 | else | |
6708 | { | |
6709 | if (TARGET_64BIT) | |
6710 | { | |
6711 | /* It might seem that one insn could be saved by accessing | |
6712 | the millicode function using the linkage table. However, | |
6713 | this doesn't work in shared libraries and other dynamically | |
6714 | loaded objects. Using a pc-relative sequence also avoids | |
6715 | problems related to the implicit use of the gp register. */ | |
6716 | output_asm_insn ("b,l .+8,%%r1", xoperands); | |
9bd9af5d | 6717 | |
6718 | if (TARGET_GAS) | |
6719 | { | |
6720 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands); | |
6721 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands); | |
6722 | } | |
6723 | else | |
6724 | { | |
6725 | xoperands[1] = gen_label_rtx (); | |
6726 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
805e22b2 | 6727 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
9bd9af5d | 6728 | CODE_LABEL_NUMBER (xoperands[1])); |
6729 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); | |
6730 | } | |
6731 | ||
ece88821 | 6732 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); |
c7a4e712 | 6733 | } |
c7a4e712 | 6734 | else if (TARGET_PORTABLE_RUNTIME) |
6735 | { | |
ece88821 | 6736 | /* Pure portable runtime doesn't allow be/ble; we also don't |
6737 | have PIC support in the assembler/linker, so this sequence | |
6738 | is needed. */ | |
c7a4e712 | 6739 | |
ece88821 | 6740 | /* Get the address of our target into %r1. */ |
6741 | output_asm_insn ("ldil L'%0,%%r1", xoperands); | |
6742 | output_asm_insn ("ldo R'%0(%%r1),%%r1", xoperands); | |
c7a4e712 | 6743 | |
ece88821 | 6744 | /* Get our return address into %r31. */ |
6745 | output_asm_insn ("{bl|b,l} .+8,%%r31", xoperands); | |
6746 | output_asm_insn ("addi 8,%%r31,%%r31", xoperands); | |
c7a4e712 | 6747 | |
ece88821 | 6748 | /* Jump to our target address in %r1. */ |
6749 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
c7a4e712 | 6750 | } |
ece88821 | 6751 | else if (!flag_pic) |
c7a4e712 | 6752 | { |
ece88821 | 6753 | output_asm_insn ("ldil L'%0,%%r1", xoperands); |
356267e0 | 6754 | if (TARGET_PA_20) |
ece88821 | 6755 | output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31", xoperands); |
356267e0 | 6756 | else |
ece88821 | 6757 | output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands); |
c7a4e712 | 6758 | } |
ece88821 | 6759 | else |
c7a4e712 | 6760 | { |
9bd9af5d | 6761 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); |
6762 | output_asm_insn ("addi 16,%%r1,%%r31", xoperands); | |
6763 | ||
ece88821 | 6764 | if (TARGET_SOM || !TARGET_GAS) |
6765 | { | |
6766 | /* The HP assembler can generate relocations for the | |
6767 | difference of two symbols. GAS can do this for a | |
6768 | millicode symbol but not an arbitrary external | |
6769 | symbol when generating SOM output. */ | |
6770 | xoperands[1] = gen_label_rtx (); | |
805e22b2 | 6771 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
ece88821 | 6772 | CODE_LABEL_NUMBER (xoperands[1])); |
6773 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
6774 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); | |
6775 | } | |
6776 | else | |
6777 | { | |
ece88821 | 6778 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+8,%%r1", xoperands); |
6779 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+12(%%r1),%%r1", | |
6780 | xoperands); | |
6781 | } | |
c7a4e712 | 6782 | |
ece88821 | 6783 | /* Jump to our target address in %r1. */ |
6784 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
c7a4e712 | 6785 | } |
c7a4e712 | 6786 | } |
6787 | ||
ece88821 | 6788 | if (seq_length == 0) |
6789 | output_asm_insn ("nop", xoperands); | |
c7a4e712 | 6790 | |
ece88821 | 6791 | /* We are done if there isn't a jump in the delay slot. */ |
6792 | if (seq_length == 0 || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN) | |
6793 | return ""; | |
c7a4e712 | 6794 | |
ece88821 | 6795 | /* This call has an unconditional jump in its delay slot. */ |
6796 | xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1); | |
c7a4e712 | 6797 | |
ece88821 | 6798 | /* See if the return address can be adjusted. Use the containing |
6799 | sequence insn's address. */ | |
cd0dfcc5 | 6800 | if (INSN_ADDRESSES_SET_P ()) |
c7a4e712 | 6801 | { |
cd0dfcc5 | 6802 | seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0))); |
6803 | distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn)))) | |
6804 | - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8); | |
6805 | ||
6806 | if (VAL_14_BITS_P (distance)) | |
6807 | { | |
6808 | xoperands[1] = gen_label_rtx (); | |
6809 | output_asm_insn ("ldo %0-%1(%2),%2", xoperands); | |
6810 | (*targetm.asm_out.internal_label) (asm_out_file, "L", | |
6811 | CODE_LABEL_NUMBER (xoperands[1])); | |
6812 | } | |
6813 | else | |
6814 | /* ??? This branch may not reach its target. */ | |
6815 | output_asm_insn ("nop\n\tb,n %0", xoperands); | |
c7a4e712 | 6816 | } |
ece88821 | 6817 | else |
6818 | /* ??? This branch may not reach its target. */ | |
6819 | output_asm_insn ("nop\n\tb,n %0", xoperands); | |
c7a4e712 | 6820 | |
6821 | /* Delete the jump. */ | |
6822 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6823 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6824 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
ece88821 | 6825 | |
c7a4e712 | 6826 | return ""; |
6827 | } | |
6828 | ||
cd0dfcc5 | 6829 | /* Return the attribute length of the call instruction INSN. The SIBCALL |
6830 | flag indicates whether INSN is a regular call or a sibling call. The | |
faf3f8c1 | 6831 | length returned must be longer than the code actually generated by |
6832 | output_call. Since branch shortening is done before delay branch | |
6833 | sequencing, there is no way to determine whether or not the delay | |
6834 | slot will be filled during branch shortening. Even when the delay | |
6835 | slot is filled, we may have to add a nop if the delay slot contains | |
6836 | a branch that can't reach its target. Thus, we always have to include | |
6837 | the delay slot in the length estimate. This used to be done in | |
6838 | pa_adjust_insn_length but we do it here now as some sequences always | |
6839 | fill the delay slot and we can save four bytes in the estimate for | |
6840 | these sequences. */ | |
ece88821 | 6841 | |
6842 | int | |
5c1d8983 | 6843 | attr_length_call (rtx insn, int sibcall) |
ece88821 | 6844 | { |
faf3f8c1 | 6845 | int local_call; |
6846 | rtx call_dest; | |
6847 | tree call_decl; | |
6848 | int length = 0; | |
6849 | rtx pat = PATTERN (insn); | |
cd0dfcc5 | 6850 | unsigned long distance = -1; |
ece88821 | 6851 | |
cd0dfcc5 | 6852 | if (INSN_ADDRESSES_SET_P ()) |
6853 | { | |
faf3f8c1 | 6854 | unsigned long total; |
6855 | ||
6856 | total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes; | |
2247cc5f | 6857 | distance = (total + insn_current_reference_address (insn)); |
6858 | if (distance < total) | |
cd0dfcc5 | 6859 | distance = -1; |
6860 | } | |
ece88821 | 6861 | |
faf3f8c1 | 6862 | /* Determine if this is a local call. */ |
6863 | if (GET_CODE (XVECEXP (pat, 0, 0)) == CALL) | |
6864 | call_dest = XEXP (XEXP (XVECEXP (pat, 0, 0), 0), 0); | |
ece88821 | 6865 | else |
faf3f8c1 | 6866 | call_dest = XEXP (XEXP (XEXP (XVECEXP (pat, 0, 0), 1), 0), 0); |
ece88821 | 6867 | |
faf3f8c1 | 6868 | call_decl = SYMBOL_REF_DECL (call_dest); |
6869 | local_call = call_decl && (*targetm.binds_local_p) (call_decl); | |
ece88821 | 6870 | |
faf3f8c1 | 6871 | /* pc-relative branch. */ |
6872 | if (!TARGET_LONG_CALLS | |
6873 | && ((TARGET_PA_20 && !sibcall && distance < 7600000) | |
6874 | || distance < 240000)) | |
6875 | length += 8; | |
ece88821 | 6876 | |
faf3f8c1 | 6877 | /* 64-bit plabel sequence. */ |
6878 | else if (TARGET_64BIT && !local_call) | |
6879 | length += sibcall ? 28 : 24; | |
ece88821 | 6880 | |
faf3f8c1 | 6881 | /* non-pic long absolute branch sequence. */ |
6882 | else if ((TARGET_LONG_ABS_CALL || local_call) && !flag_pic) | |
6883 | length += 12; | |
ece88821 | 6884 | |
faf3f8c1 | 6885 | /* long pc-relative branch sequence. */ |
6886 | else if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) | |
6887 | || (TARGET_64BIT && !TARGET_GAS) | |
6888 | || (TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call))) | |
6889 | { | |
6890 | length += 20; | |
ece88821 | 6891 | |
faf3f8c1 | 6892 | if (!TARGET_PA_20 && !TARGET_NO_SPACE_REGS) |
6893 | length += 8; | |
6894 | } | |
8a05c3c2 | 6895 | |
faf3f8c1 | 6896 | /* 32-bit plabel sequence. */ |
6897 | else | |
6898 | { | |
6899 | length += 32; | |
ece88821 | 6900 | |
faf3f8c1 | 6901 | if (TARGET_SOM) |
6902 | length += length_fp_args (insn); | |
6903 | ||
6904 | if (flag_pic) | |
6905 | length += 4; | |
ee376abe | 6906 | |
faf3f8c1 | 6907 | if (!TARGET_PA_20) |
6908 | { | |
ece88821 | 6909 | if (!sibcall) |
6910 | length += 8; | |
6911 | ||
faf3f8c1 | 6912 | if (!TARGET_NO_SPACE_REGS) |
6913 | length += 8; | |
ece88821 | 6914 | } |
6915 | } | |
faf3f8c1 | 6916 | |
6917 | return length; | |
ece88821 | 6918 | } |
6919 | ||
6920 | /* INSN is a function call. It may have an unconditional jump | |
c7a4e712 | 6921 | in its delay slot. |
6922 | ||
6923 | CALL_DEST is the routine we are calling. */ | |
6924 | ||
611a88e1 | 6925 | const char * |
5c1d8983 | 6926 | output_call (rtx insn, rtx call_dest, int sibcall) |
c7a4e712 | 6927 | { |
ece88821 | 6928 | int delay_insn_deleted = 0; |
6929 | int delay_slot_filled = 0; | |
b70ea764 | 6930 | int seq_length = dbr_sequence_length (); |
2247cc5f | 6931 | tree call_decl = SYMBOL_REF_DECL (call_dest); |
faf3f8c1 | 6932 | int local_call = call_decl && (*targetm.binds_local_p) (call_decl); |
ece88821 | 6933 | rtx xoperands[2]; |
6934 | ||
6935 | xoperands[0] = call_dest; | |
c7a4e712 | 6936 | |
ece88821 | 6937 | /* Handle the common case where we're sure that the branch will reach |
2247cc5f | 6938 | the beginning of the "$CODE$" subspace. This is the beginning of |
6939 | the current function if we are in a named section. */ | |
cd0dfcc5 | 6940 | if (!TARGET_LONG_CALLS && attr_length_call (insn, sibcall) == 8) |
d6686e21 | 6941 | { |
5e3c5739 | 6942 | xoperands[1] = gen_rtx_REG (word_mode, sibcall ? 0 : 2); |
ece88821 | 6943 | output_asm_insn ("{bl|b,l} %0,%1", xoperands); |
06ddb6f8 | 6944 | } |
ece88821 | 6945 | else |
06ddb6f8 | 6946 | { |
2247cc5f | 6947 | if (TARGET_64BIT && !local_call) |
3683f840 | 6948 | { |
ece88821 | 6949 | /* ??? As far as I can tell, the HP linker doesn't support the |
6950 | long pc-relative sequence described in the 64-bit runtime | |
6951 | architecture. So, we use a slightly longer indirect call. */ | |
6952 | struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0)); | |
6953 | ||
6954 | xoperands[0] = p->internal_label; | |
6955 | xoperands[1] = gen_label_rtx (); | |
6956 | ||
6957 | /* If this isn't a sibcall, we put the load of %r27 into the | |
6958 | delay slot. We can't do this in a sibcall as we don't | |
6959 | have a second call-clobbered scratch register available. */ | |
6960 | if (seq_length != 0 | |
6961 | && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN | |
6962 | && !sibcall) | |
6963 | { | |
6964 | final_scan_insn (NEXT_INSN (insn), asm_out_file, | |
6965 | optimize, 0, 0); | |
6966 | ||
6967 | /* Now delete the delay insn. */ | |
6968 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6969 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6970 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
6971 | delay_insn_deleted = 1; | |
6972 | } | |
06ddb6f8 | 6973 | |
ece88821 | 6974 | output_asm_insn ("addil LT'%0,%%r27", xoperands); |
6975 | output_asm_insn ("ldd RT'%0(%%r1),%%r1", xoperands); | |
6976 | output_asm_insn ("ldd 0(%%r1),%%r1", xoperands); | |
06ddb6f8 | 6977 | |
ece88821 | 6978 | if (sibcall) |
06ddb6f8 | 6979 | { |
ece88821 | 6980 | output_asm_insn ("ldd 24(%%r1),%%r27", xoperands); |
6981 | output_asm_insn ("ldd 16(%%r1),%%r1", xoperands); | |
6982 | output_asm_insn ("bve (%%r1)", xoperands); | |
6983 | } | |
6984 | else | |
6985 | { | |
6986 | output_asm_insn ("ldd 16(%%r1),%%r2", xoperands); | |
6987 | output_asm_insn ("bve,l (%%r2),%%r2", xoperands); | |
6988 | output_asm_insn ("ldd 24(%%r1),%%r27", xoperands); | |
6989 | delay_slot_filled = 1; | |
06ddb6f8 | 6990 | } |
6991 | } | |
ece88821 | 6992 | else |
e3f53689 | 6993 | { |
ece88821 | 6994 | int indirect_call = 0; |
6995 | ||
6996 | /* Emit a long call. There are several different sequences | |
6997 | of increasing length and complexity. In most cases, | |
6998 | they don't allow an instruction in the delay slot. */ | |
2247cc5f | 6999 | if (!((TARGET_LONG_ABS_CALL || local_call) && !flag_pic) |
ece88821 | 7000 | && !(TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) |
2247cc5f | 7001 | && !(TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call)) |
7002 | && !TARGET_64BIT) | |
ece88821 | 7003 | indirect_call = 1; |
7004 | ||
7005 | if (seq_length != 0 | |
7006 | && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN | |
7007 | && !sibcall | |
7008 | && (!TARGET_PA_20 || indirect_call)) | |
5cc6b2bc | 7009 | { |
ece88821 | 7010 | /* A non-jump insn in the delay slot. By definition we can |
7011 | emit this insn before the call (and in fact before argument | |
7012 | relocating. */ | |
7013 | final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0, 0); | |
7014 | ||
7015 | /* Now delete the delay insn. */ | |
7016 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
7017 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
7018 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
7019 | delay_insn_deleted = 1; | |
5cc6b2bc | 7020 | } |
e3f53689 | 7021 | |
2247cc5f | 7022 | if ((TARGET_LONG_ABS_CALL || local_call) && !flag_pic) |
5cc6b2bc | 7023 | { |
ece88821 | 7024 | /* This is the best sequence for making long calls in |
7025 | non-pic code. Unfortunately, GNU ld doesn't provide | |
7026 | the stub needed for external calls, and GAS's support | |
2247cc5f | 7027 | for this with the SOM linker is buggy. It is safe |
7028 | to use this for local calls. */ | |
ece88821 | 7029 | output_asm_insn ("ldil L'%0,%%r1", xoperands); |
7030 | if (sibcall) | |
7031 | output_asm_insn ("be R'%0(%%sr4,%%r1)", xoperands); | |
7032 | else | |
7033 | { | |
7034 | if (TARGET_PA_20) | |
7035 | output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31", | |
7036 | xoperands); | |
7037 | else | |
7038 | output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands); | |
c7a4e712 | 7039 | |
ece88821 | 7040 | output_asm_insn ("copy %%r31,%%r2", xoperands); |
7041 | delay_slot_filled = 1; | |
7042 | } | |
7043 | } | |
7044 | else | |
7045 | { | |
2247cc5f | 7046 | if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) |
7047 | || (TARGET_64BIT && !TARGET_GAS)) | |
b70ea764 | 7048 | { |
ece88821 | 7049 | /* The HP assembler and linker can handle relocations |
7050 | for the difference of two symbols. GAS and the HP | |
7051 | linker can't do this when one of the symbols is | |
7052 | external. */ | |
7053 | xoperands[1] = gen_label_rtx (); | |
7054 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
7055 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
805e22b2 | 7056 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
b70ea764 | 7057 | CODE_LABEL_NUMBER (xoperands[1])); |
ece88821 | 7058 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); |
7059 | } | |
2247cc5f | 7060 | else if (TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call)) |
b70ea764 | 7061 | { |
ece88821 | 7062 | /* GAS currently can't generate the relocations that |
7063 | are needed for the SOM linker under HP-UX using this | |
7064 | sequence. The GNU linker doesn't generate the stubs | |
7065 | that are needed for external calls on TARGET_ELF32 | |
7066 | with this sequence. For now, we have to use a | |
7067 | longer plabel sequence when using GAS. */ | |
7068 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
7069 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", | |
b70ea764 | 7070 | xoperands); |
ece88821 | 7071 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", |
b70ea764 | 7072 | xoperands); |
7073 | } | |
5e3c5739 | 7074 | else |
7075 | { | |
ece88821 | 7076 | /* Emit a long plabel-based call sequence. This is |
7077 | essentially an inline implementation of $$dyncall. | |
7078 | We don't actually try to call $$dyncall as this is | |
7079 | as difficult as calling the function itself. */ | |
7080 | struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0)); | |
7081 | ||
7082 | xoperands[0] = p->internal_label; | |
7083 | xoperands[1] = gen_label_rtx (); | |
7084 | ||
7085 | /* Since the call is indirect, FP arguments in registers | |
7086 | need to be copied to the general registers. Then, the | |
7087 | argument relocation stub will copy them back. */ | |
7088 | if (TARGET_SOM) | |
7089 | copy_fp_args (insn); | |
7090 | ||
7091 | if (flag_pic) | |
7092 | { | |
7093 | output_asm_insn ("addil LT'%0,%%r19", xoperands); | |
7094 | output_asm_insn ("ldw RT'%0(%%r1),%%r1", xoperands); | |
7095 | output_asm_insn ("ldw 0(%%r1),%%r1", xoperands); | |
7096 | } | |
7097 | else | |
7098 | { | |
7099 | output_asm_insn ("addil LR'%0-$global$,%%r27", | |
7100 | xoperands); | |
7101 | output_asm_insn ("ldw RR'%0-$global$(%%r1),%%r1", | |
7102 | xoperands); | |
7103 | } | |
06ddb6f8 | 7104 | |
ece88821 | 7105 | output_asm_insn ("bb,>=,n %%r1,30,.+16", xoperands); |
7106 | output_asm_insn ("depi 0,31,2,%%r1", xoperands); | |
7107 | output_asm_insn ("ldw 4(%%sr0,%%r1),%%r19", xoperands); | |
7108 | output_asm_insn ("ldw 0(%%sr0,%%r1),%%r1", xoperands); | |
c7a4e712 | 7109 | |
ece88821 | 7110 | if (!sibcall && !TARGET_PA_20) |
7111 | { | |
7112 | output_asm_insn ("{bl|b,l} .+8,%%r2", xoperands); | |
ee376abe | 7113 | if (TARGET_NO_SPACE_REGS) |
7114 | output_asm_insn ("addi 8,%%r2,%%r2", xoperands); | |
7115 | else | |
7116 | output_asm_insn ("addi 16,%%r2,%%r2", xoperands); | |
ece88821 | 7117 | } |
7118 | } | |
c7a4e712 | 7119 | |
ece88821 | 7120 | if (TARGET_PA_20) |
5e3c5739 | 7121 | { |
ece88821 | 7122 | if (sibcall) |
7123 | output_asm_insn ("bve (%%r1)", xoperands); | |
7124 | else | |
7125 | { | |
7126 | if (indirect_call) | |
7127 | { | |
7128 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); | |
7129 | output_asm_insn ("stw %%r2,-24(%%sp)", xoperands); | |
7130 | delay_slot_filled = 1; | |
7131 | } | |
7132 | else | |
7133 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); | |
7134 | } | |
5e3c5739 | 7135 | } |
7136 | else | |
7137 | { | |
ee376abe | 7138 | if (!TARGET_NO_SPACE_REGS) |
7139 | output_asm_insn ("ldsid (%%r1),%%r31\n\tmtsp %%r31,%%sr0", | |
7140 | xoperands); | |
06ddb6f8 | 7141 | |
ece88821 | 7142 | if (sibcall) |
ee376abe | 7143 | { |
7144 | if (TARGET_NO_SPACE_REGS) | |
7145 | output_asm_insn ("be 0(%%sr4,%%r1)", xoperands); | |
7146 | else | |
7147 | output_asm_insn ("be 0(%%sr0,%%r1)", xoperands); | |
7148 | } | |
ece88821 | 7149 | else |
7150 | { | |
ee376abe | 7151 | if (TARGET_NO_SPACE_REGS) |
7152 | output_asm_insn ("ble 0(%%sr4,%%r1)", xoperands); | |
7153 | else | |
7154 | output_asm_insn ("ble 0(%%sr0,%%r1)", xoperands); | |
06ddb6f8 | 7155 | |
ece88821 | 7156 | if (indirect_call) |
7157 | output_asm_insn ("stw %%r31,-24(%%sp)", xoperands); | |
7158 | else | |
7159 | output_asm_insn ("copy %%r31,%%r2", xoperands); | |
7160 | delay_slot_filled = 1; | |
7161 | } | |
7162 | } | |
7163 | } | |
06ddb6f8 | 7164 | } |
d6686e21 | 7165 | } |
6d36483b | 7166 | |
8a05c3c2 | 7167 | if (!delay_slot_filled && (seq_length == 0 || delay_insn_deleted)) |
ece88821 | 7168 | output_asm_insn ("nop", xoperands); |
d6686e21 | 7169 | |
ece88821 | 7170 | /* We are done if there isn't a jump in the delay slot. */ |
7171 | if (seq_length == 0 | |
7172 | || delay_insn_deleted | |
7173 | || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN) | |
7174 | return ""; | |
d6686e21 | 7175 | |
ece88821 | 7176 | /* A sibcall should never have a branch in the delay slot. */ |
7177 | if (sibcall) | |
7178 | abort (); | |
d6686e21 | 7179 | |
ece88821 | 7180 | /* This call has an unconditional jump in its delay slot. */ |
7181 | xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1); | |
d6686e21 | 7182 | |
cd0dfcc5 | 7183 | if (!delay_slot_filled && INSN_ADDRESSES_SET_P ()) |
d6686e21 | 7184 | { |
ece88821 | 7185 | /* See if the return address can be adjusted. Use the containing |
7186 | sequence insn's address. */ | |
7187 | rtx seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0))); | |
7188 | int distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn)))) | |
7189 | - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8); | |
7190 | ||
7191 | if (VAL_14_BITS_P (distance)) | |
7192 | { | |
7193 | xoperands[1] = gen_label_rtx (); | |
7194 | output_asm_insn ("ldo %0-%1(%%r2),%%r2", xoperands); | |
805e22b2 | 7195 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
7196 | CODE_LABEL_NUMBER (xoperands[1])); | |
ece88821 | 7197 | } |
7198 | else | |
ece88821 | 7199 | output_asm_insn ("nop\n\tb,n %0", xoperands); |
d6686e21 | 7200 | } |
ece88821 | 7201 | else |
ece88821 | 7202 | output_asm_insn ("b,n %0", xoperands); |
d6686e21 | 7203 | |
7204 | /* Delete the jump. */ | |
7205 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
7206 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
7207 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
ece88821 | 7208 | |
d6686e21 | 7209 | return ""; |
7210 | } | |
7211 | ||
cd0dfcc5 | 7212 | /* Return the attribute length of the indirect call instruction INSN. |
7213 | The length must match the code generated by output_indirect call. | |
7214 | The returned length includes the delay slot. Currently, the delay | |
7215 | slot of an indirect call sequence is not exposed and it is used by | |
7216 | the sequence itself. */ | |
7217 | ||
7218 | int | |
5c1d8983 | 7219 | attr_length_indirect_call (rtx insn) |
cd0dfcc5 | 7220 | { |
7221 | unsigned long distance = -1; | |
8a05c3c2 | 7222 | unsigned long total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes; |
cd0dfcc5 | 7223 | |
7224 | if (INSN_ADDRESSES_SET_P ()) | |
7225 | { | |
2247cc5f | 7226 | distance = (total + insn_current_reference_address (insn)); |
7227 | if (distance < total) | |
cd0dfcc5 | 7228 | distance = -1; |
7229 | } | |
7230 | ||
7231 | if (TARGET_64BIT) | |
7232 | return 12; | |
7233 | ||
7234 | if (TARGET_FAST_INDIRECT_CALLS | |
7235 | || (!TARGET_PORTABLE_RUNTIME | |
7236 | && ((TARGET_PA_20 && distance < 7600000) || distance < 240000))) | |
7237 | return 8; | |
7238 | ||
7239 | if (flag_pic) | |
7240 | return 24; | |
7241 | ||
7242 | if (TARGET_PORTABLE_RUNTIME) | |
7243 | return 20; | |
7244 | ||
7245 | /* Out of reach, can use ble. */ | |
7246 | return 12; | |
7247 | } | |
7248 | ||
7249 | const char * | |
5c1d8983 | 7250 | output_indirect_call (rtx insn, rtx call_dest) |
cd0dfcc5 | 7251 | { |
7252 | rtx xoperands[1]; | |
7253 | ||
7254 | if (TARGET_64BIT) | |
7255 | { | |
7256 | xoperands[0] = call_dest; | |
7257 | output_asm_insn ("ldd 16(%0),%%r2", xoperands); | |
7258 | output_asm_insn ("bve,l (%%r2),%%r2\n\tldd 24(%0),%%r27", xoperands); | |
7259 | return ""; | |
7260 | } | |
7261 | ||
7262 | /* First the special case for kernels, level 0 systems, etc. */ | |
7263 | if (TARGET_FAST_INDIRECT_CALLS) | |
7264 | return "ble 0(%%sr4,%%r22)\n\tcopy %%r31,%%r2"; | |
7265 | ||
7266 | /* Now the normal case -- we can reach $$dyncall directly or | |
7267 | we're sure that we can get there via a long-branch stub. | |
7268 | ||
7269 | No need to check target flags as the length uniquely identifies | |
7270 | the remaining cases. */ | |
7271 | if (attr_length_indirect_call (insn) == 8) | |
7272 | return ".CALL\tARGW0=GR\n\t{bl|b,l} $$dyncall,%%r31\n\tcopy %%r31,%%r2"; | |
7273 | ||
7274 | /* Long millicode call, but we are not generating PIC or portable runtime | |
7275 | code. */ | |
7276 | if (attr_length_indirect_call (insn) == 12) | |
7277 | return ".CALL\tARGW0=GR\n\tldil L'$$dyncall,%%r2\n\tble R'$$dyncall(%%sr4,%%r2)\n\tcopy %%r31,%%r2"; | |
7278 | ||
7279 | /* Long millicode call for portable runtime. */ | |
7280 | if (attr_length_indirect_call (insn) == 20) | |
7281 | return "ldil L'$$dyncall,%%r31\n\tldo R'$$dyncall(%%r31),%%r31\n\tblr %%r0,%%r2\n\tbv,n %%r0(%%r31)\n\tnop"; | |
7282 | ||
7283 | /* We need a long PIC call to $$dyncall. */ | |
7284 | xoperands[0] = NULL_RTX; | |
7285 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
7286 | if (TARGET_SOM || !TARGET_GAS) | |
7287 | { | |
7288 | xoperands[0] = gen_label_rtx (); | |
7289 | output_asm_insn ("addil L'$$dyncall-%0,%%r1", xoperands); | |
7290 | (*targetm.asm_out.internal_label) (asm_out_file, "L", | |
7291 | CODE_LABEL_NUMBER (xoperands[0])); | |
7292 | output_asm_insn ("ldo R'$$dyncall-%0(%%r1),%%r1", xoperands); | |
7293 | } | |
7294 | else | |
7295 | { | |
7296 | output_asm_insn ("addil L'$$dyncall-$PIC_pcrel$0+4,%%r1", xoperands); | |
7297 | output_asm_insn ("ldo R'$$dyncall-$PIC_pcrel$0+8(%%r1),%%r1", | |
7298 | xoperands); | |
7299 | } | |
7300 | output_asm_insn ("blr %%r0,%%r2", xoperands); | |
7301 | output_asm_insn ("bv,n %%r0(%%r1)\n\tnop", xoperands); | |
7302 | return ""; | |
7303 | } | |
7304 | ||
7305 | /* Return the total length of the save and restore instructions needed for | |
7306 | the data linkage table pointer (i.e., the PIC register) across the call | |
7307 | instruction INSN. No-return calls do not require a save and restore. | |
7308 | In addition, we may be able to avoid the save and restore for calls | |
7309 | within the same translation unit. */ | |
7310 | ||
7311 | int | |
5c1d8983 | 7312 | attr_length_save_restore_dltp (rtx insn) |
cd0dfcc5 | 7313 | { |
7314 | if (find_reg_note (insn, REG_NORETURN, NULL_RTX)) | |
7315 | return 0; | |
7316 | ||
7317 | return 8; | |
7318 | } | |
7319 | ||
d6f01525 | 7320 | /* In HPUX 8.0's shared library scheme, special relocations are needed |
6d36483b | 7321 | for function labels if they might be passed to a function |
d6f01525 | 7322 | in a shared library (because shared libraries don't live in code |
44acf429 | 7323 | space), and special magic is needed to construct their address. */ |
d6f01525 | 7324 | |
7325 | void | |
5c1d8983 | 7326 | hppa_encode_label (rtx sym) |
d6f01525 | 7327 | { |
611a88e1 | 7328 | const char *str = XSTR (sym, 0); |
cccfb31e | 7329 | int len = strlen (str) + 1; |
7330 | char *newstr, *p; | |
d6f01525 | 7331 | |
cccfb31e | 7332 | p = newstr = alloca (len + 1); |
cccfb31e | 7333 | *p++ = '@'; |
7334 | strcpy (p, str); | |
74d80a9a | 7335 | |
ea52c577 | 7336 | XSTR (sym, 0) = ggc_alloc_string (newstr, len); |
d6f01525 | 7337 | } |
6d36483b | 7338 | |
7811991d | 7339 | static void |
5c1d8983 | 7340 | pa_encode_section_info (tree decl, rtx rtl, int first) |
7811991d | 7341 | { |
7342 | if (first && TEXT_SPACE_P (decl)) | |
7343 | { | |
7811991d | 7344 | SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1; |
7345 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
2c129d70 | 7346 | hppa_encode_label (XEXP (rtl, 0)); |
7811991d | 7347 | } |
7348 | } | |
7349 | ||
7b4a38a6 | 7350 | /* This is sort of inverse to pa_encode_section_info. */ |
7351 | ||
7352 | static const char * | |
5c1d8983 | 7353 | pa_strip_name_encoding (const char *str) |
7b4a38a6 | 7354 | { |
c0264367 | 7355 | str += (*str == '@'); |
7356 | str += (*str == '*'); | |
7357 | return str; | |
7b4a38a6 | 7358 | } |
7359 | ||
d6f01525 | 7360 | int |
5c1d8983 | 7361 | function_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
d6f01525 | 7362 | { |
c1b3411e | 7363 | return GET_CODE (op) == SYMBOL_REF && FUNCTION_NAME_P (XSTR (op, 0)); |
d6f01525 | 7364 | } |
f33e3942 | 7365 | |
166bf021 | 7366 | /* Returns 1 if OP is a function label involved in a simple addition |
7367 | with a constant. Used to keep certain patterns from matching | |
7368 | during instruction combination. */ | |
7369 | int | |
5c1d8983 | 7370 | is_function_label_plus_const (rtx op) |
166bf021 | 7371 | { |
7372 | /* Strip off any CONST. */ | |
7373 | if (GET_CODE (op) == CONST) | |
7374 | op = XEXP (op, 0); | |
7375 | ||
7376 | return (GET_CODE (op) == PLUS | |
7377 | && function_label_operand (XEXP (op, 0), Pmode) | |
7378 | && GET_CODE (XEXP (op, 1)) == CONST_INT); | |
7379 | } | |
7380 | ||
f1752b7e | 7381 | /* Output assembly code for a thunk to FUNCTION. */ |
7382 | ||
6988553d | 7383 | static void |
5c1d8983 | 7384 | pa_asm_output_mi_thunk (FILE *file, tree thunk_fndecl, HOST_WIDE_INT delta, |
7385 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED, | |
7386 | tree function) | |
f1752b7e | 7387 | { |
2247cc5f | 7388 | const char *fname = XSTR (XEXP (DECL_RTL (function), 0), 0); |
7389 | const char *tname = XSTR (XEXP (DECL_RTL (thunk_fndecl), 0), 0); | |
7390 | int val_14 = VAL_14_BITS_P (delta); | |
7391 | int nbytes = 0; | |
f1752b7e | 7392 | static unsigned int current_thunk_number; |
7393 | char label[16]; | |
2247cc5f | 7394 | |
7395 | ASM_OUTPUT_LABEL (file, tname); | |
7396 | fprintf (file, "\t.PROC\n\t.CALLINFO FRAME=0,NO_CALLS\n\t.ENTRY\n"); | |
7397 | ||
7398 | fname = (*targetm.strip_name_encoding) (fname); | |
7399 | tname = (*targetm.strip_name_encoding) (tname); | |
7400 | ||
7401 | /* Output the thunk. We know that the function is in the same | |
7402 | translation unit (i.e., the same space) as the thunk, and that | |
7403 | thunks are output after their method. Thus, we don't need an | |
7404 | external branch to reach the function. With SOM and GAS, | |
7405 | functions and thunks are effectively in different sections. | |
7406 | Thus, we can always use a IA-relative branch and the linker | |
7407 | will add a long branch stub if necessary. | |
7408 | ||
7409 | However, we have to be careful when generating PIC code on the | |
7410 | SOM port to ensure that the sequence does not transfer to an | |
7411 | import stub for the target function as this could clobber the | |
7412 | return value saved at SP-24. This would also apply to the | |
7413 | 32-bit linux port if the multi-space model is implemented. */ | |
7414 | if ((!TARGET_LONG_CALLS && TARGET_SOM && !TARGET_PORTABLE_RUNTIME | |
7415 | && !(flag_pic && TREE_PUBLIC (function)) | |
7416 | && (TARGET_GAS || last_address < 262132)) | |
7417 | || (!TARGET_LONG_CALLS && !TARGET_SOM && !TARGET_PORTABLE_RUNTIME | |
7418 | && ((targetm.have_named_sections | |
7419 | && DECL_SECTION_NAME (thunk_fndecl) != NULL | |
7420 | /* The GNU 64-bit linker has rather poor stub management. | |
7421 | So, we use a long branch from thunks that aren't in | |
7422 | the same section as the target function. */ | |
7423 | && ((!TARGET_64BIT | |
7424 | && (DECL_SECTION_NAME (thunk_fndecl) | |
7425 | != DECL_SECTION_NAME (function))) | |
7426 | || ((DECL_SECTION_NAME (thunk_fndecl) | |
7427 | == DECL_SECTION_NAME (function)) | |
7428 | && last_address < 262132))) | |
7429 | || (!targetm.have_named_sections && last_address < 262132)))) | |
7430 | { | |
7431 | if (val_14) | |
7432 | { | |
7433 | fprintf (file, "\tb %s\n\tldo " HOST_WIDE_INT_PRINT_DEC | |
7434 | "(%%r26),%%r26\n", fname, delta); | |
7435 | nbytes += 8; | |
7436 | } | |
7437 | else | |
7438 | { | |
7439 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC | |
7440 | ",%%r26\n", delta); | |
7441 | fprintf (file, "\tb %s\n\tldo R'" HOST_WIDE_INT_PRINT_DEC | |
7442 | "(%%r1),%%r26\n", fname, delta); | |
7443 | nbytes += 12; | |
7444 | } | |
7445 | } | |
7446 | else if (TARGET_64BIT) | |
7447 | { | |
7448 | /* We only have one call-clobbered scratch register, so we can't | |
7449 | make use of the delay slot if delta doesn't fit in 14 bits. */ | |
7450 | if (!val_14) | |
7451 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC | |
7452 | ",%%r26\n\tldo R'" HOST_WIDE_INT_PRINT_DEC | |
7453 | "(%%r1),%%r26\n", delta, delta); | |
7454 | ||
7455 | fprintf (file, "\tb,l .+8,%%r1\n"); | |
7456 | ||
7457 | if (TARGET_GAS) | |
7458 | { | |
7459 | fprintf (file, "\taddil L'%s-$PIC_pcrel$0+4,%%r1\n", fname); | |
7460 | fprintf (file, "\tldo R'%s-$PIC_pcrel$0+8(%%r1),%%r1\n", fname); | |
7461 | } | |
7462 | else | |
7463 | { | |
7464 | int off = val_14 ? 8 : 16; | |
7465 | fprintf (file, "\taddil L'%s-%s-%d,%%r1\n", fname, tname, off); | |
7466 | fprintf (file, "\tldo R'%s-%s-%d(%%r1),%%r1\n", fname, tname, off); | |
7467 | } | |
7468 | ||
7469 | if (val_14) | |
7470 | { | |
7471 | fprintf (file, "\tbv %%r0(%%r1)\n\tldo "); | |
7472 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta); | |
7473 | nbytes += 20; | |
7474 | } | |
7475 | else | |
7476 | { | |
7477 | fprintf (file, "\tbv,n %%r0(%%r1)\n"); | |
7478 | nbytes += 24; | |
7479 | } | |
7480 | } | |
7481 | else if (TARGET_PORTABLE_RUNTIME) | |
7482 | { | |
7483 | fprintf (file, "\tldil L'%s,%%r1\n", fname); | |
7484 | fprintf (file, "\tldo R'%s(%%r1),%%r22\n", fname); | |
7485 | ||
7486 | if (val_14) | |
7487 | { | |
7488 | fprintf (file, "\tbv %%r0(%%r22)\n\tldo "); | |
7489 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta); | |
7490 | nbytes += 16; | |
7491 | } | |
7492 | else | |
7493 | { | |
7494 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC | |
7495 | ",%%r26\n", delta); | |
7496 | fprintf (file, "\tbv %%r0(%%r22)\n\tldo "); | |
7497 | fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta); | |
7498 | nbytes += 20; | |
7499 | } | |
7500 | } | |
7501 | else if (TARGET_SOM && flag_pic && TREE_PUBLIC (function)) | |
7502 | { | |
7503 | /* The function is accessible from outside this module. The only | |
7504 | way to avoid an import stub between the thunk and function is to | |
7505 | call the function directly with an indirect sequence similar to | |
7506 | that used by $$dyncall. This is possible because $$dyncall acts | |
7507 | as the import stub in an indirect call. */ | |
7508 | const char *lab; | |
7509 | ||
7510 | ASM_GENERATE_INTERNAL_LABEL (label, "LTHN", current_thunk_number); | |
7511 | lab = (*targetm.strip_name_encoding) (label); | |
7512 | ||
7513 | fprintf (file, "\taddil LT'%s,%%r19\n", lab); | |
7514 | fprintf (file, "\tldw RT'%s(%%r1),%%r22\n", lab); | |
7515 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); | |
7516 | fprintf (file, "\tbb,>=,n %%r22,30,.+16\n"); | |
7517 | fprintf (file, "\tdepi 0,31,2,%%r22\n"); | |
7518 | fprintf (file, "\tldw 4(%%sr0,%%r22),%%r19\n"); | |
7519 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); | |
7520 | if (!val_14) | |
7521 | { | |
7522 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC | |
7523 | ",%%r26\n", delta); | |
7524 | nbytes += 4; | |
7525 | } | |
7526 | if (TARGET_PA_20) | |
7527 | { | |
7528 | fprintf (file, "\tbve (%%r22)\n\tldo "); | |
7529 | nbytes += 36; | |
7530 | } | |
7531 | else | |
f1752b7e | 7532 | { |
ee376abe | 7533 | if (TARGET_NO_SPACE_REGS) |
2247cc5f | 7534 | { |
7535 | fprintf (file, "\tbe 0(%%sr4,%%r22)\n\tldo "); | |
7536 | nbytes += 36; | |
7537 | } | |
ee376abe | 7538 | else |
7539 | { | |
6454616d | 7540 | fprintf (file, "\tldsid (%%sr0,%%r22),%%r21\n"); |
2247cc5f | 7541 | fprintf (file, "\tmtsp %%r21,%%sr0\n"); |
ee376abe | 7542 | fprintf (file, "\tbe 0(%%sr0,%%r22)\n\tldo "); |
2247cc5f | 7543 | nbytes += 44; |
ee376abe | 7544 | } |
2247cc5f | 7545 | } |
7546 | ||
7547 | if (val_14) | |
7548 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta); | |
7549 | else | |
7550 | fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta); | |
7551 | } | |
7552 | else if (flag_pic) | |
7553 | { | |
7554 | if (TARGET_PA_20) | |
7555 | fprintf (file, "\tb,l .+8,%%r1\n"); | |
7556 | else | |
7557 | fprintf (file, "\tbl .+8,%%r1\n"); | |
7558 | ||
7559 | if (TARGET_SOM || !TARGET_GAS) | |
7560 | { | |
7561 | fprintf (file, "\taddil L'%s-%s-8,%%r1\n", fname, tname); | |
7562 | fprintf (file, "\tldo R'%s-%s-8(%%r1),%%r22\n", fname, tname); | |
7563 | } | |
7564 | else | |
7565 | { | |
7566 | fprintf (file, "\taddil L'%s-$PIC_pcrel$0+4,%%r1\n", fname); | |
7567 | fprintf (file, "\tldo R'%s-$PIC_pcrel$0+8(%%r1),%%r22\n", fname); | |
7568 | } | |
7569 | ||
7570 | if (val_14) | |
7571 | { | |
7572 | fprintf (file, "\tbv %%r0(%%r22)\n\tldo "); | |
4840a03a | 7573 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta); |
2247cc5f | 7574 | nbytes += 20; |
f1752b7e | 7575 | } |
7576 | else | |
2247cc5f | 7577 | { |
7578 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC | |
7579 | ",%%r26\n", delta); | |
7580 | fprintf (file, "\tbv %%r0(%%r22)\n\tldo "); | |
7581 | fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta); | |
7582 | nbytes += 24; | |
7583 | } | |
f1752b7e | 7584 | } |
7585 | else | |
7586 | { | |
2247cc5f | 7587 | if (!val_14) |
7588 | fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC ",%%r26\n", delta); | |
7589 | ||
7590 | fprintf (file, "\tldil L'%s,%%r22\n", fname); | |
7591 | fprintf (file, "\tbe R'%s(%%sr4,%%r22)\n\tldo ", fname); | |
7592 | ||
7593 | if (val_14) | |
f1752b7e | 7594 | { |
2247cc5f | 7595 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta); |
7596 | nbytes += 12; | |
f1752b7e | 7597 | } |
7598 | else | |
2247cc5f | 7599 | { |
7600 | fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta); | |
7601 | nbytes += 16; | |
7602 | } | |
f1752b7e | 7603 | } |
2247cc5f | 7604 | |
f1752b7e | 7605 | fprintf (file, "\t.EXIT\n\t.PROCEND\n"); |
2247cc5f | 7606 | |
7607 | if (TARGET_SOM && flag_pic && TREE_PUBLIC (function)) | |
f1752b7e | 7608 | { |
7609 | data_section (); | |
7610 | fprintf (file, "\t.align 4\n"); | |
2247cc5f | 7611 | ASM_OUTPUT_LABEL (file, label); |
7612 | fprintf (file, "\t.word P'%s\n", fname); | |
f1752b7e | 7613 | function_section (thunk_fndecl); |
7614 | } | |
2247cc5f | 7615 | |
f1752b7e | 7616 | current_thunk_number++; |
2247cc5f | 7617 | nbytes = ((nbytes + FUNCTION_BOUNDARY / BITS_PER_UNIT - 1) |
7618 | & ~(FUNCTION_BOUNDARY / BITS_PER_UNIT - 1)); | |
7619 | last_address += nbytes; | |
7620 | update_total_code_bytes (nbytes); | |
f1752b7e | 7621 | } |
7622 | ||
805e22b2 | 7623 | /* Only direct calls to static functions are allowed to be sibling (tail) |
7624 | call optimized. | |
7625 | ||
7626 | This restriction is necessary because some linker generated stubs will | |
7627 | store return pointers into rp' in some cases which might clobber a | |
7628 | live value already in rp'. | |
7629 | ||
7630 | In a sibcall the current function and the target function share stack | |
7631 | space. Thus if the path to the current function and the path to the | |
7632 | target function save a value in rp', they save the value into the | |
7633 | same stack slot, which has undesirable consequences. | |
7634 | ||
7635 | Because of the deferred binding nature of shared libraries any function | |
7636 | with external scope could be in a different load module and thus require | |
7637 | rp' to be saved when calling that function. So sibcall optimizations | |
7638 | can only be safe for static function. | |
7639 | ||
7640 | Note that GCC never needs return value relocations, so we don't have to | |
7641 | worry about static calls with return value relocations (which require | |
7642 | saving rp'). | |
7643 | ||
7644 | It is safe to perform a sibcall optimization when the target function | |
7645 | will never return. */ | |
7646 | static bool | |
5c1d8983 | 7647 | pa_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
805e22b2 | 7648 | { |
f62b73b6 | 7649 | /* Sibcalls are ok for TARGET_ELF32 as along as the linker is used in |
7650 | single subspace mode and the call is not indirect. As far as I know, | |
7651 | there is no operating system support for the multiple subspace mode. | |
7652 | It might be possible to support indirect calls if we didn't use | |
7653 | $$dyncall (see the indirect sequence generated in output_call). */ | |
7654 | if (TARGET_ELF32) | |
7655 | return (decl != NULL_TREE); | |
7656 | ||
7657 | /* Sibcalls are not ok because the arg pointer register is not a fixed | |
2cecd772 | 7658 | register. This prevents the sibcall optimization from occurring. In |
f62b73b6 | 7659 | addition, there are problems with stub placement using GNU ld. This |
7660 | is because a normal sibcall branch uses a 17-bit relocation while | |
7661 | a regular call branch uses a 22-bit relocation. As a result, more | |
7662 | care needs to be taken in the placement of long-branch stubs. */ | |
7663 | if (TARGET_64BIT) | |
7664 | return false; | |
7665 | ||
805e22b2 | 7666 | return (decl |
f62b73b6 | 7667 | && !TARGET_PORTABLE_RUNTIME |
7668 | && !TREE_PUBLIC (decl)); | |
805e22b2 | 7669 | } |
7670 | ||
37580c80 | 7671 | /* Returns 1 if the 6 operands specified in OPERANDS are suitable for |
7672 | use in fmpyadd instructions. */ | |
4ed6ee50 | 7673 | int |
5c1d8983 | 7674 | fmpyaddoperands (rtx *operands) |
4ed6ee50 | 7675 | { |
201f01e9 | 7676 | enum machine_mode mode = GET_MODE (operands[0]); |
4ed6ee50 | 7677 | |
ab449421 | 7678 | /* Must be a floating point mode. */ |
7679 | if (mode != SFmode && mode != DFmode) | |
7680 | return 0; | |
7681 | ||
4ed6ee50 | 7682 | /* All modes must be the same. */ |
201f01e9 | 7683 | if (! (mode == GET_MODE (operands[1]) |
7684 | && mode == GET_MODE (operands[2]) | |
7685 | && mode == GET_MODE (operands[3]) | |
7686 | && mode == GET_MODE (operands[4]) | |
7687 | && mode == GET_MODE (operands[5]))) | |
4ed6ee50 | 7688 | return 0; |
7689 | ||
ab449421 | 7690 | /* All operands must be registers. */ |
7691 | if (! (GET_CODE (operands[1]) == REG | |
7692 | && GET_CODE (operands[2]) == REG | |
7693 | && GET_CODE (operands[3]) == REG | |
7694 | && GET_CODE (operands[4]) == REG | |
7695 | && GET_CODE (operands[5]) == REG)) | |
4ed6ee50 | 7696 | return 0; |
7697 | ||
37580c80 | 7698 | /* Only 2 real operands to the addition. One of the input operands must |
7699 | be the same as the output operand. */ | |
4ed6ee50 | 7700 | if (! rtx_equal_p (operands[3], operands[4]) |
7701 | && ! rtx_equal_p (operands[3], operands[5])) | |
7702 | return 0; | |
7703 | ||
7704 | /* Inout operand of add can not conflict with any operands from multiply. */ | |
7705 | if (rtx_equal_p (operands[3], operands[0]) | |
7706 | || rtx_equal_p (operands[3], operands[1]) | |
7707 | || rtx_equal_p (operands[3], operands[2])) | |
7708 | return 0; | |
7709 | ||
7710 | /* multiply can not feed into addition operands. */ | |
7711 | if (rtx_equal_p (operands[4], operands[0]) | |
7712 | || rtx_equal_p (operands[5], operands[0])) | |
7713 | return 0; | |
7714 | ||
ab449421 | 7715 | /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */ |
7716 | if (mode == SFmode | |
bac38c40 | 7717 | && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS |
7718 | || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS | |
7719 | || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS | |
7720 | || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS | |
7721 | || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS | |
7722 | || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS)) | |
ab449421 | 7723 | return 0; |
7724 | ||
4ed6ee50 | 7725 | /* Passed. Operands are suitable for fmpyadd. */ |
7726 | return 1; | |
7727 | } | |
7728 | ||
de419443 | 7729 | #if !defined(USE_COLLECT2) |
7730 | static void | |
5c1d8983 | 7731 | pa_asm_out_constructor (rtx symbol, int priority) |
de419443 | 7732 | { |
7733 | if (!function_label_operand (symbol, VOIDmode)) | |
7734 | hppa_encode_label (symbol); | |
7735 | ||
7736 | #ifdef CTORS_SECTION_ASM_OP | |
7737 | default_ctor_section_asm_out_constructor (symbol, priority); | |
7738 | #else | |
7739 | # ifdef TARGET_ASM_NAMED_SECTION | |
7740 | default_named_section_asm_out_constructor (symbol, priority); | |
7741 | # else | |
7742 | default_stabs_asm_out_constructor (symbol, priority); | |
7743 | # endif | |
7744 | #endif | |
7745 | } | |
7746 | ||
7747 | static void | |
5c1d8983 | 7748 | pa_asm_out_destructor (rtx symbol, int priority) |
de419443 | 7749 | { |
7750 | if (!function_label_operand (symbol, VOIDmode)) | |
7751 | hppa_encode_label (symbol); | |
7752 | ||
7753 | #ifdef DTORS_SECTION_ASM_OP | |
7754 | default_dtor_section_asm_out_destructor (symbol, priority); | |
7755 | #else | |
7756 | # ifdef TARGET_ASM_NAMED_SECTION | |
7757 | default_named_section_asm_out_destructor (symbol, priority); | |
7758 | # else | |
7759 | default_stabs_asm_out_destructor (symbol, priority); | |
7760 | # endif | |
7761 | #endif | |
7762 | } | |
7763 | #endif | |
7764 | ||
37580c80 | 7765 | /* Returns 1 if the 6 operands specified in OPERANDS are suitable for |
7766 | use in fmpysub instructions. */ | |
4ed6ee50 | 7767 | int |
5c1d8983 | 7768 | fmpysuboperands (rtx *operands) |
4ed6ee50 | 7769 | { |
201f01e9 | 7770 | enum machine_mode mode = GET_MODE (operands[0]); |
4ed6ee50 | 7771 | |
ab449421 | 7772 | /* Must be a floating point mode. */ |
7773 | if (mode != SFmode && mode != DFmode) | |
7774 | return 0; | |
7775 | ||
4ed6ee50 | 7776 | /* All modes must be the same. */ |
201f01e9 | 7777 | if (! (mode == GET_MODE (operands[1]) |
7778 | && mode == GET_MODE (operands[2]) | |
7779 | && mode == GET_MODE (operands[3]) | |
7780 | && mode == GET_MODE (operands[4]) | |
7781 | && mode == GET_MODE (operands[5]))) | |
4ed6ee50 | 7782 | return 0; |
7783 | ||
ab449421 | 7784 | /* All operands must be registers. */ |
7785 | if (! (GET_CODE (operands[1]) == REG | |
7786 | && GET_CODE (operands[2]) == REG | |
7787 | && GET_CODE (operands[3]) == REG | |
7788 | && GET_CODE (operands[4]) == REG | |
7789 | && GET_CODE (operands[5]) == REG)) | |
4ed6ee50 | 7790 | return 0; |
7791 | ||
37580c80 | 7792 | /* Only 2 real operands to the subtraction. Subtraction is not a commutative |
7793 | operation, so operands[4] must be the same as operand[3]. */ | |
4ed6ee50 | 7794 | if (! rtx_equal_p (operands[3], operands[4])) |
7795 | return 0; | |
7796 | ||
7797 | /* multiply can not feed into subtraction. */ | |
37580c80 | 7798 | if (rtx_equal_p (operands[5], operands[0])) |
4ed6ee50 | 7799 | return 0; |
7800 | ||
37580c80 | 7801 | /* Inout operand of sub can not conflict with any operands from multiply. */ |
4ed6ee50 | 7802 | if (rtx_equal_p (operands[3], operands[0]) |
7803 | || rtx_equal_p (operands[3], operands[1]) | |
7804 | || rtx_equal_p (operands[3], operands[2])) | |
7805 | return 0; | |
7806 | ||
ab449421 | 7807 | /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */ |
7808 | if (mode == SFmode | |
bac38c40 | 7809 | && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS |
7810 | || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS | |
7811 | || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS | |
7812 | || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS | |
7813 | || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS | |
7814 | || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS)) | |
ab449421 | 7815 | return 0; |
7816 | ||
4ed6ee50 | 7817 | /* Passed. Operands are suitable for fmpysub. */ |
7818 | return 1; | |
7819 | } | |
7820 | ||
89f29d73 | 7821 | int |
5c1d8983 | 7822 | plus_xor_ior_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
89f29d73 | 7823 | { |
7824 | return (GET_CODE (op) == PLUS || GET_CODE (op) == XOR | |
7825 | || GET_CODE (op) == IOR); | |
7826 | } | |
6720f95e | 7827 | |
7828 | /* Return 1 if the given constant is 2, 4, or 8. These are the valid | |
7829 | constants for shadd instructions. */ | |
7d27e4c9 | 7830 | static int |
5c1d8983 | 7831 | shadd_constant_p (int val) |
6720f95e | 7832 | { |
7833 | if (val == 2 || val == 4 || val == 8) | |
7834 | return 1; | |
7835 | else | |
7836 | return 0; | |
7837 | } | |
3a16146d | 7838 | |
7839 | /* Return 1 if OP is a CONST_INT with the value 2, 4, or 8. These are | |
7840 | the valid constant for shadd instructions. */ | |
7841 | int | |
5c1d8983 | 7842 | shadd_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
3a16146d | 7843 | { |
7844 | return (GET_CODE (op) == CONST_INT && shadd_constant_p (INTVAL (op))); | |
7845 | } | |
5fbd5940 | 7846 | |
42819d4e | 7847 | /* Return 1 if OP is valid as a base register in a reg + reg address. */ |
7848 | ||
7849 | int | |
5c1d8983 | 7850 | basereg_operand (rtx op, enum machine_mode mode) |
42819d4e | 7851 | { |
21f3ee9c | 7852 | /* cse will create some unscaled indexed addresses, however; it |
7853 | generally isn't a win on the PA, so avoid creating unscaled | |
7854 | indexed addresses until after cse is finished. */ | |
7855 | if (!cse_not_expected) | |
7856 | return 0; | |
7857 | ||
e61a0a7f | 7858 | /* Allow any register when TARGET_NO_SPACE_REGS is in effect since |
7859 | we don't have to worry about the braindamaged implicit space | |
7860 | register selection from the basereg. */ | |
7861 | if (TARGET_NO_SPACE_REGS) | |
23a667d0 | 7862 | return (GET_CODE (op) == REG); |
42819d4e | 7863 | |
21f3ee9c | 7864 | /* While it's always safe to index off the frame pointer, it's not |
7865 | always profitable, particularly when the frame pointer is being | |
7866 | eliminated. */ | |
7867 | if (! flag_omit_frame_pointer && op == frame_pointer_rtx) | |
42819d4e | 7868 | return 1; |
7869 | ||
e61a0a7f | 7870 | return (GET_CODE (op) == REG |
90ae5d76 | 7871 | && REG_POINTER (op) |
7872 | && register_operand (op, mode)); | |
42819d4e | 7873 | } |
7874 | ||
51987f90 | 7875 | /* Return 1 if this operand is anything other than a hard register. */ |
7876 | ||
7877 | int | |
5c1d8983 | 7878 | non_hard_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
51987f90 | 7879 | { |
7880 | return ! (GET_CODE (op) == REG && REGNO (op) < FIRST_PSEUDO_REGISTER); | |
7881 | } | |
7882 | ||
5fbd5940 | 7883 | /* Return 1 if INSN branches forward. Should be using insn_addresses |
6dc3b0d9 | 7884 | to avoid walking through all the insns... */ |
7d27e4c9 | 7885 | static int |
5c1d8983 | 7886 | forward_branch_p (rtx insn) |
5fbd5940 | 7887 | { |
7888 | rtx label = JUMP_LABEL (insn); | |
7889 | ||
7890 | while (insn) | |
7891 | { | |
7892 | if (insn == label) | |
7893 | break; | |
7894 | else | |
7895 | insn = NEXT_INSN (insn); | |
7896 | } | |
7897 | ||
7898 | return (insn == label); | |
7899 | } | |
7900 | ||
29a4502c | 7901 | /* Return 1 if OP is an equality comparison, else return 0. */ |
7902 | int | |
5c1d8983 | 7903 | eq_neq_comparison_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
29a4502c | 7904 | { |
7905 | return (GET_CODE (op) == EQ || GET_CODE (op) == NE); | |
7906 | } | |
7907 | ||
7908 | /* Return 1 if OP is an operator suitable for use in a movb instruction. */ | |
7909 | int | |
5c1d8983 | 7910 | movb_comparison_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
29a4502c | 7911 | { |
7912 | return (GET_CODE (op) == EQ || GET_CODE (op) == NE | |
7913 | || GET_CODE (op) == LT || GET_CODE (op) == GE); | |
7914 | } | |
7915 | ||
d6686e21 | 7916 | /* Return 1 if INSN is in the delay slot of a call instruction. */ |
7917 | int | |
5c1d8983 | 7918 | jump_in_call_delay (rtx insn) |
d6686e21 | 7919 | { |
7920 | ||
7921 | if (GET_CODE (insn) != JUMP_INSN) | |
7922 | return 0; | |
7923 | ||
7924 | if (PREV_INSN (insn) | |
7925 | && PREV_INSN (PREV_INSN (insn)) | |
ece0fa59 | 7926 | && GET_CODE (next_real_insn (PREV_INSN (PREV_INSN (insn)))) == INSN) |
d6686e21 | 7927 | { |
ece0fa59 | 7928 | rtx test_insn = next_real_insn (PREV_INSN (PREV_INSN (insn))); |
d6686e21 | 7929 | |
7930 | return (GET_CODE (PATTERN (test_insn)) == SEQUENCE | |
7931 | && XVECEXP (PATTERN (test_insn), 0, 1) == insn); | |
7932 | ||
7933 | } | |
7934 | else | |
7935 | return 0; | |
7936 | } | |
3b1e673e | 7937 | |
546a40bd | 7938 | /* Output an unconditional move and branch insn. */ |
7939 | ||
611a88e1 | 7940 | const char * |
5c1d8983 | 7941 | output_parallel_movb (rtx *operands, int length) |
546a40bd | 7942 | { |
7943 | /* These are the cases in which we win. */ | |
7944 | if (length == 4) | |
7945 | return "mov%I1b,tr %1,%0,%2"; | |
7946 | ||
7947 | /* None of these cases wins, but they don't lose either. */ | |
7948 | if (dbr_sequence_length () == 0) | |
7949 | { | |
7950 | /* Nothing in the delay slot, fake it by putting the combined | |
7951 | insn (the copy or add) in the delay slot of a bl. */ | |
7952 | if (GET_CODE (operands[1]) == CONST_INT) | |
5a811d43 | 7953 | return "b %2\n\tldi %1,%0"; |
546a40bd | 7954 | else |
5a811d43 | 7955 | return "b %2\n\tcopy %1,%0"; |
546a40bd | 7956 | } |
7957 | else | |
7958 | { | |
7959 | /* Something in the delay slot, but we've got a long branch. */ | |
7960 | if (GET_CODE (operands[1]) == CONST_INT) | |
5a811d43 | 7961 | return "ldi %1,%0\n\tb %2"; |
546a40bd | 7962 | else |
5a811d43 | 7963 | return "copy %1,%0\n\tb %2"; |
546a40bd | 7964 | } |
7965 | } | |
7966 | ||
7967 | /* Output an unconditional add and branch insn. */ | |
7968 | ||
611a88e1 | 7969 | const char * |
5c1d8983 | 7970 | output_parallel_addb (rtx *operands, int length) |
546a40bd | 7971 | { |
7972 | /* To make life easy we want operand0 to be the shared input/output | |
7973 | operand and operand1 to be the readonly operand. */ | |
7974 | if (operands[0] == operands[1]) | |
7975 | operands[1] = operands[2]; | |
7976 | ||
7977 | /* These are the cases in which we win. */ | |
7978 | if (length == 4) | |
7979 | return "add%I1b,tr %1,%0,%3"; | |
7980 | ||
7981 | /* None of these cases win, but they don't lose either. */ | |
7982 | if (dbr_sequence_length () == 0) | |
7983 | { | |
7984 | /* Nothing in the delay slot, fake it by putting the combined | |
7985 | insn (the copy or add) in the delay slot of a bl. */ | |
5a811d43 | 7986 | return "b %3\n\tadd%I1 %1,%0,%0"; |
546a40bd | 7987 | } |
7988 | else | |
7989 | { | |
7990 | /* Something in the delay slot, but we've got a long branch. */ | |
5a811d43 | 7991 | return "add%I1 %1,%0,%0\n\tb %3"; |
546a40bd | 7992 | } |
7993 | } | |
7994 | ||
7c5101fc | 7995 | /* Return nonzero if INSN (a jump insn) immediately follows a call |
7996 | to a named function. This is used to avoid filling the delay slot | |
7997 | of the jump since it can usually be eliminated by modifying RP in | |
7998 | the delay slot of the call. */ | |
9840d99d | 7999 | |
7d27e4c9 | 8000 | int |
5c1d8983 | 8001 | following_call (rtx insn) |
546a40bd | 8002 | { |
ed1b0769 | 8003 | if (! TARGET_JUMP_IN_DELAY) |
1b6f11e2 | 8004 | return 0; |
8005 | ||
546a40bd | 8006 | /* Find the previous real insn, skipping NOTEs. */ |
8007 | insn = PREV_INSN (insn); | |
8008 | while (insn && GET_CODE (insn) == NOTE) | |
8009 | insn = PREV_INSN (insn); | |
8010 | ||
8011 | /* Check for CALL_INSNs and millicode calls. */ | |
8012 | if (insn | |
1d2e016c | 8013 | && ((GET_CODE (insn) == CALL_INSN |
8014 | && get_attr_type (insn) != TYPE_DYNCALL) | |
546a40bd | 8015 | || (GET_CODE (insn) == INSN |
8016 | && GET_CODE (PATTERN (insn)) != SEQUENCE | |
8017 | && GET_CODE (PATTERN (insn)) != USE | |
8018 | && GET_CODE (PATTERN (insn)) != CLOBBER | |
8019 | && get_attr_type (insn) == TYPE_MILLI))) | |
8020 | return 1; | |
8021 | ||
8022 | return 0; | |
8023 | } | |
8024 | ||
3b1e673e | 8025 | /* We use this hook to perform a PA specific optimization which is difficult |
8026 | to do in earlier passes. | |
8027 | ||
8028 | We want the delay slots of branches within jump tables to be filled. | |
8029 | None of the compiler passes at the moment even has the notion that a | |
8030 | PA jump table doesn't contain addresses, but instead contains actual | |
8031 | instructions! | |
8032 | ||
8033 | Because we actually jump into the table, the addresses of each entry | |
01cc3b75 | 8034 | must stay constant in relation to the beginning of the table (which |
3b1e673e | 8035 | itself must stay constant relative to the instruction to jump into |
8036 | it). I don't believe we can guarantee earlier passes of the compiler | |
8037 | will adhere to those rules. | |
8038 | ||
8039 | So, late in the compilation process we find all the jump tables, and | |
8040 | expand them into real code -- eg each entry in the jump table vector | |
8041 | will get an appropriate label followed by a jump to the final target. | |
8042 | ||
8043 | Reorg and the final jump pass can then optimize these branches and | |
8044 | fill their delay slots. We end up with smaller, more efficient code. | |
8045 | ||
9840d99d | 8046 | The jump instructions within the table are special; we must be able |
3b1e673e | 8047 | to identify them during assembly output (if the jumps don't get filled |
8048 | we need to emit a nop rather than nullifying the delay slot)). We | |
b932f66c | 8049 | identify jumps in switch tables by using insns with the attribute |
8050 | type TYPE_BTABLE_BRANCH. | |
9239127b | 8051 | |
8052 | We also surround the jump table itself with BEGIN_BRTAB and END_BRTAB | |
8053 | insns. This serves two purposes, first it prevents jump.c from | |
8054 | noticing that the last N entries in the table jump to the instruction | |
8055 | immediately after the table and deleting the jumps. Second, those | |
8056 | insns mark where we should emit .begin_brtab and .end_brtab directives | |
8057 | when using GAS (allows for better link time optimizations). */ | |
3b1e673e | 8058 | |
2efea8c0 | 8059 | static void |
5c1d8983 | 8060 | pa_reorg (void) |
3b1e673e | 8061 | { |
8062 | rtx insn; | |
8063 | ||
2efea8c0 | 8064 | remove_useless_addtr_insns (1); |
3d457930 | 8065 | |
342aabd9 | 8066 | if (pa_cpu < PROCESSOR_8000) |
2efea8c0 | 8067 | pa_combine_instructions (); |
342aabd9 | 8068 | |
bd49d362 | 8069 | |
3d457930 | 8070 | /* This is fairly cheap, so always run it if optimizing. */ |
a66555b2 | 8071 | if (optimize > 0 && !TARGET_BIG_SWITCH) |
3b1e673e | 8072 | { |
b41266d4 | 8073 | /* Find and explode all ADDR_VEC or ADDR_DIFF_VEC insns. */ |
2efea8c0 | 8074 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
3b1e673e | 8075 | { |
b932f66c | 8076 | rtx pattern, tmp, location, label; |
3b1e673e | 8077 | unsigned int length, i; |
8078 | ||
b41266d4 | 8079 | /* Find an ADDR_VEC or ADDR_DIFF_VEC insn to explode. */ |
3b1e673e | 8080 | if (GET_CODE (insn) != JUMP_INSN |
b41266d4 | 8081 | || (GET_CODE (PATTERN (insn)) != ADDR_VEC |
8082 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)) | |
3b1e673e | 8083 | continue; |
8084 | ||
9239127b | 8085 | /* Emit marker for the beginning of the branch table. */ |
8086 | emit_insn_before (gen_begin_brtab (), insn); | |
f9333726 | 8087 | |
3b1e673e | 8088 | pattern = PATTERN (insn); |
8089 | location = PREV_INSN (insn); | |
b41266d4 | 8090 | length = XVECLEN (pattern, GET_CODE (pattern) == ADDR_DIFF_VEC); |
f9333726 | 8091 | |
3b1e673e | 8092 | for (i = 0; i < length; i++) |
8093 | { | |
a66555b2 | 8094 | /* Emit a label before each jump to keep jump.c from |
8095 | removing this code. */ | |
8096 | tmp = gen_label_rtx (); | |
8097 | LABEL_NUSES (tmp) = 1; | |
8098 | emit_label_after (tmp, location); | |
8099 | location = NEXT_INSN (location); | |
8100 | ||
b41266d4 | 8101 | if (GET_CODE (pattern) == ADDR_VEC) |
b932f66c | 8102 | label = XEXP (XVECEXP (pattern, 0, i), 0); |
b41266d4 | 8103 | else |
b932f66c | 8104 | label = XEXP (XVECEXP (pattern, 1, i), 0); |
8105 | ||
8106 | tmp = gen_short_jump (label); | |
8107 | ||
8108 | /* Emit the jump itself. */ | |
8109 | tmp = emit_jump_insn_after (tmp, location); | |
8110 | JUMP_LABEL (tmp) = label; | |
8111 | LABEL_NUSES (label)++; | |
8112 | location = NEXT_INSN (location); | |
3b1e673e | 8113 | |
8114 | /* Emit a BARRIER after the jump. */ | |
3b1e673e | 8115 | emit_barrier_after (location); |
3b1e673e | 8116 | location = NEXT_INSN (location); |
8117 | } | |
f9333726 | 8118 | |
9239127b | 8119 | /* Emit marker for the end of the branch table. */ |
8120 | emit_insn_before (gen_end_brtab (), location); | |
8121 | location = NEXT_INSN (location); | |
8122 | emit_barrier_after (location); | |
a66555b2 | 8123 | |
b41266d4 | 8124 | /* Delete the ADDR_VEC or ADDR_DIFF_VEC. */ |
3b1e673e | 8125 | delete_insn (insn); |
8126 | } | |
8127 | } | |
9239127b | 8128 | else |
f9333726 | 8129 | { |
b932f66c | 8130 | /* Still need brtab marker insns. FIXME: the presence of these |
8131 | markers disables output of the branch table to readonly memory, | |
8132 | and any alignment directives that might be needed. Possibly, | |
8133 | the begin_brtab insn should be output before the label for the | |
8134 | table. This doesn matter at the moment since the tables are | |
8135 | always output in the text section. */ | |
2efea8c0 | 8136 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
f9333726 | 8137 | { |
8138 | /* Find an ADDR_VEC insn. */ | |
8139 | if (GET_CODE (insn) != JUMP_INSN | |
b41266d4 | 8140 | || (GET_CODE (PATTERN (insn)) != ADDR_VEC |
8141 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)) | |
f9333726 | 8142 | continue; |
8143 | ||
8144 | /* Now generate markers for the beginning and end of the | |
ad87de1e | 8145 | branch table. */ |
f9333726 | 8146 | emit_insn_before (gen_begin_brtab (), insn); |
8147 | emit_insn_after (gen_end_brtab (), insn); | |
8148 | } | |
8149 | } | |
d3287673 | 8150 | } |
bd49d362 | 8151 | |
8152 | /* The PA has a number of odd instructions which can perform multiple | |
8153 | tasks at once. On first generation PA machines (PA1.0 and PA1.1) | |
8154 | it may be profitable to combine two instructions into one instruction | |
8155 | with two outputs. It's not profitable PA2.0 machines because the | |
8156 | two outputs would take two slots in the reorder buffers. | |
8157 | ||
8158 | This routine finds instructions which can be combined and combines | |
8159 | them. We only support some of the potential combinations, and we | |
8160 | only try common ways to find suitable instructions. | |
8161 | ||
8162 | * addb can add two registers or a register and a small integer | |
8163 | and jump to a nearby (+-8k) location. Normally the jump to the | |
8164 | nearby location is conditional on the result of the add, but by | |
8165 | using the "true" condition we can make the jump unconditional. | |
8166 | Thus addb can perform two independent operations in one insn. | |
8167 | ||
8168 | * movb is similar to addb in that it can perform a reg->reg | |
8169 | or small immediate->reg copy and jump to a nearby (+-8k location). | |
8170 | ||
8171 | * fmpyadd and fmpysub can perform a FP multiply and either an | |
8172 | FP add or FP sub if the operands of the multiply and add/sub are | |
8173 | independent (there are other minor restrictions). Note both | |
8174 | the fmpy and fadd/fsub can in theory move to better spots according | |
8175 | to data dependencies, but for now we require the fmpy stay at a | |
8176 | fixed location. | |
8177 | ||
8178 | * Many of the memory operations can perform pre & post updates | |
8179 | of index registers. GCC's pre/post increment/decrement addressing | |
8180 | is far too simple to take advantage of all the possibilities. This | |
8181 | pass may not be suitable since those insns may not be independent. | |
8182 | ||
8183 | * comclr can compare two ints or an int and a register, nullify | |
8184 | the following instruction and zero some other register. This | |
8185 | is more difficult to use as it's harder to find an insn which | |
8186 | will generate a comclr than finding something like an unconditional | |
8187 | branch. (conditional moves & long branches create comclr insns). | |
8188 | ||
8189 | * Most arithmetic operations can conditionally skip the next | |
8190 | instruction. They can be viewed as "perform this operation | |
8191 | and conditionally jump to this nearby location" (where nearby | |
8192 | is an insns away). These are difficult to use due to the | |
8193 | branch length restrictions. */ | |
8194 | ||
7d27e4c9 | 8195 | static void |
5c1d8983 | 8196 | pa_combine_instructions (void) |
bd49d362 | 8197 | { |
8198 | rtx anchor, new; | |
8199 | ||
8200 | /* This can get expensive since the basic algorithm is on the | |
8201 | order of O(n^2) (or worse). Only do it for -O2 or higher | |
ad87de1e | 8202 | levels of optimization. */ |
bd49d362 | 8203 | if (optimize < 2) |
8204 | return; | |
8205 | ||
8206 | /* Walk down the list of insns looking for "anchor" insns which | |
8207 | may be combined with "floating" insns. As the name implies, | |
8208 | "anchor" instructions don't move, while "floating" insns may | |
8209 | move around. */ | |
ad851752 | 8210 | new = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, NULL_RTX, NULL_RTX)); |
bd49d362 | 8211 | new = make_insn_raw (new); |
8212 | ||
8213 | for (anchor = get_insns (); anchor; anchor = NEXT_INSN (anchor)) | |
8214 | { | |
8215 | enum attr_pa_combine_type anchor_attr; | |
8216 | enum attr_pa_combine_type floater_attr; | |
8217 | ||
8218 | /* We only care about INSNs, JUMP_INSNs, and CALL_INSNs. | |
8219 | Also ignore any special USE insns. */ | |
7d27e4c9 | 8220 | if ((GET_CODE (anchor) != INSN |
bd49d362 | 8221 | && GET_CODE (anchor) != JUMP_INSN |
7d27e4c9 | 8222 | && GET_CODE (anchor) != CALL_INSN) |
bd49d362 | 8223 | || GET_CODE (PATTERN (anchor)) == USE |
8224 | || GET_CODE (PATTERN (anchor)) == CLOBBER | |
8225 | || GET_CODE (PATTERN (anchor)) == ADDR_VEC | |
8226 | || GET_CODE (PATTERN (anchor)) == ADDR_DIFF_VEC) | |
8227 | continue; | |
8228 | ||
8229 | anchor_attr = get_attr_pa_combine_type (anchor); | |
8230 | /* See if anchor is an insn suitable for combination. */ | |
8231 | if (anchor_attr == PA_COMBINE_TYPE_FMPY | |
8232 | || anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
8233 | || (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH | |
8234 | && ! forward_branch_p (anchor))) | |
8235 | { | |
8236 | rtx floater; | |
8237 | ||
8238 | for (floater = PREV_INSN (anchor); | |
8239 | floater; | |
8240 | floater = PREV_INSN (floater)) | |
8241 | { | |
8242 | if (GET_CODE (floater) == NOTE | |
8243 | || (GET_CODE (floater) == INSN | |
8244 | && (GET_CODE (PATTERN (floater)) == USE | |
8245 | || GET_CODE (PATTERN (floater)) == CLOBBER))) | |
8246 | continue; | |
8247 | ||
8248 | /* Anything except a regular INSN will stop our search. */ | |
8249 | if (GET_CODE (floater) != INSN | |
8250 | || GET_CODE (PATTERN (floater)) == ADDR_VEC | |
8251 | || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC) | |
8252 | { | |
8253 | floater = NULL_RTX; | |
8254 | break; | |
8255 | } | |
8256 | ||
8257 | /* See if FLOATER is suitable for combination with the | |
8258 | anchor. */ | |
8259 | floater_attr = get_attr_pa_combine_type (floater); | |
8260 | if ((anchor_attr == PA_COMBINE_TYPE_FMPY | |
8261 | && floater_attr == PA_COMBINE_TYPE_FADDSUB) | |
8262 | || (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
8263 | && floater_attr == PA_COMBINE_TYPE_FMPY)) | |
8264 | { | |
8265 | /* If ANCHOR and FLOATER can be combined, then we're | |
8266 | done with this pass. */ | |
8267 | if (pa_can_combine_p (new, anchor, floater, 0, | |
8268 | SET_DEST (PATTERN (floater)), | |
8269 | XEXP (SET_SRC (PATTERN (floater)), 0), | |
8270 | XEXP (SET_SRC (PATTERN (floater)), 1))) | |
8271 | break; | |
8272 | } | |
8273 | ||
8274 | else if (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH | |
8275 | && floater_attr == PA_COMBINE_TYPE_ADDMOVE) | |
8276 | { | |
8277 | if (GET_CODE (SET_SRC (PATTERN (floater))) == PLUS) | |
8278 | { | |
8279 | if (pa_can_combine_p (new, anchor, floater, 0, | |
8280 | SET_DEST (PATTERN (floater)), | |
8281 | XEXP (SET_SRC (PATTERN (floater)), 0), | |
8282 | XEXP (SET_SRC (PATTERN (floater)), 1))) | |
8283 | break; | |
8284 | } | |
8285 | else | |
8286 | { | |
8287 | if (pa_can_combine_p (new, anchor, floater, 0, | |
8288 | SET_DEST (PATTERN (floater)), | |
8289 | SET_SRC (PATTERN (floater)), | |
8290 | SET_SRC (PATTERN (floater)))) | |
8291 | break; | |
8292 | } | |
8293 | } | |
8294 | } | |
8295 | ||
8296 | /* If we didn't find anything on the backwards scan try forwards. */ | |
8297 | if (!floater | |
8298 | && (anchor_attr == PA_COMBINE_TYPE_FMPY | |
8299 | || anchor_attr == PA_COMBINE_TYPE_FADDSUB)) | |
8300 | { | |
8301 | for (floater = anchor; floater; floater = NEXT_INSN (floater)) | |
8302 | { | |
8303 | if (GET_CODE (floater) == NOTE | |
8304 | || (GET_CODE (floater) == INSN | |
8305 | && (GET_CODE (PATTERN (floater)) == USE | |
8306 | || GET_CODE (PATTERN (floater)) == CLOBBER))) | |
9840d99d | 8307 | |
bd49d362 | 8308 | continue; |
8309 | ||
8310 | /* Anything except a regular INSN will stop our search. */ | |
8311 | if (GET_CODE (floater) != INSN | |
8312 | || GET_CODE (PATTERN (floater)) == ADDR_VEC | |
8313 | || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC) | |
8314 | { | |
8315 | floater = NULL_RTX; | |
8316 | break; | |
8317 | } | |
8318 | ||
8319 | /* See if FLOATER is suitable for combination with the | |
8320 | anchor. */ | |
8321 | floater_attr = get_attr_pa_combine_type (floater); | |
8322 | if ((anchor_attr == PA_COMBINE_TYPE_FMPY | |
8323 | && floater_attr == PA_COMBINE_TYPE_FADDSUB) | |
8324 | || (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
8325 | && floater_attr == PA_COMBINE_TYPE_FMPY)) | |
8326 | { | |
8327 | /* If ANCHOR and FLOATER can be combined, then we're | |
8328 | done with this pass. */ | |
8329 | if (pa_can_combine_p (new, anchor, floater, 1, | |
8330 | SET_DEST (PATTERN (floater)), | |
ea52c577 | 8331 | XEXP (SET_SRC (PATTERN (floater)), |
8332 | 0), | |
8333 | XEXP (SET_SRC (PATTERN (floater)), | |
8334 | 1))) | |
bd49d362 | 8335 | break; |
8336 | } | |
8337 | } | |
8338 | } | |
8339 | ||
8340 | /* FLOATER will be nonzero if we found a suitable floating | |
8341 | insn for combination with ANCHOR. */ | |
8342 | if (floater | |
8343 | && (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
8344 | || anchor_attr == PA_COMBINE_TYPE_FMPY)) | |
8345 | { | |
8346 | /* Emit the new instruction and delete the old anchor. */ | |
7014838c | 8347 | emit_insn_before (gen_rtx_PARALLEL |
8348 | (VOIDmode, | |
8349 | gen_rtvec (2, PATTERN (anchor), | |
8350 | PATTERN (floater))), | |
8351 | anchor); | |
8352 | ||
bd49d362 | 8353 | PUT_CODE (anchor, NOTE); |
8354 | NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED; | |
8355 | NOTE_SOURCE_FILE (anchor) = 0; | |
8356 | ||
8357 | /* Emit a special USE insn for FLOATER, then delete | |
8358 | the floating insn. */ | |
ad851752 | 8359 | emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater); |
bd49d362 | 8360 | delete_insn (floater); |
8361 | ||
8362 | continue; | |
8363 | } | |
8364 | else if (floater | |
8365 | && anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH) | |
8366 | { | |
8367 | rtx temp; | |
8368 | /* Emit the new_jump instruction and delete the old anchor. */ | |
7014838c | 8369 | temp |
8370 | = emit_jump_insn_before (gen_rtx_PARALLEL | |
8371 | (VOIDmode, | |
8372 | gen_rtvec (2, PATTERN (anchor), | |
8373 | PATTERN (floater))), | |
8374 | anchor); | |
8375 | ||
bd49d362 | 8376 | JUMP_LABEL (temp) = JUMP_LABEL (anchor); |
8377 | PUT_CODE (anchor, NOTE); | |
8378 | NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED; | |
8379 | NOTE_SOURCE_FILE (anchor) = 0; | |
8380 | ||
8381 | /* Emit a special USE insn for FLOATER, then delete | |
8382 | the floating insn. */ | |
ad851752 | 8383 | emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater); |
bd49d362 | 8384 | delete_insn (floater); |
8385 | continue; | |
8386 | } | |
8387 | } | |
8388 | } | |
8389 | } | |
8390 | ||
9aadea62 | 8391 | static int |
5c1d8983 | 8392 | pa_can_combine_p (rtx new, rtx anchor, rtx floater, int reversed, rtx dest, |
8393 | rtx src1, rtx src2) | |
bd49d362 | 8394 | { |
8395 | int insn_code_number; | |
8396 | rtx start, end; | |
8397 | ||
8398 | /* Create a PARALLEL with the patterns of ANCHOR and | |
8399 | FLOATER, try to recognize it, then test constraints | |
8400 | for the resulting pattern. | |
8401 | ||
8402 | If the pattern doesn't match or the constraints | |
8403 | aren't met keep searching for a suitable floater | |
8404 | insn. */ | |
8405 | XVECEXP (PATTERN (new), 0, 0) = PATTERN (anchor); | |
8406 | XVECEXP (PATTERN (new), 0, 1) = PATTERN (floater); | |
8407 | INSN_CODE (new) = -1; | |
8408 | insn_code_number = recog_memoized (new); | |
8409 | if (insn_code_number < 0 | |
bf9fd117 | 8410 | || (extract_insn (new), ! constrain_operands (1))) |
bd49d362 | 8411 | return 0; |
8412 | ||
8413 | if (reversed) | |
8414 | { | |
8415 | start = anchor; | |
8416 | end = floater; | |
8417 | } | |
8418 | else | |
8419 | { | |
8420 | start = floater; | |
8421 | end = anchor; | |
8422 | } | |
8423 | ||
8424 | /* There's up to three operands to consider. One | |
8425 | output and two inputs. | |
8426 | ||
8427 | The output must not be used between FLOATER & ANCHOR | |
8428 | exclusive. The inputs must not be set between | |
8429 | FLOATER and ANCHOR exclusive. */ | |
8430 | ||
8431 | if (reg_used_between_p (dest, start, end)) | |
8432 | return 0; | |
8433 | ||
8434 | if (reg_set_between_p (src1, start, end)) | |
8435 | return 0; | |
8436 | ||
8437 | if (reg_set_between_p (src2, start, end)) | |
8438 | return 0; | |
8439 | ||
8440 | /* If we get here, then everything is good. */ | |
8441 | return 1; | |
8442 | } | |
14d18de3 | 8443 | |
a6582a53 | 8444 | /* Return nonzero if references for INSN are delayed. |
14d18de3 | 8445 | |
8446 | Millicode insns are actually function calls with some special | |
8447 | constraints on arguments and register usage. | |
8448 | ||
8449 | Millicode calls always expect their arguments in the integer argument | |
8450 | registers, and always return their result in %r29 (ret1). They | |
2013ddf6 | 8451 | are expected to clobber their arguments, %r1, %r29, and the return |
8452 | pointer which is %r31 on 32-bit and %r2 on 64-bit, and nothing else. | |
8453 | ||
8454 | This function tells reorg that the references to arguments and | |
8455 | millicode calls do not appear to happen until after the millicode call. | |
8456 | This allows reorg to put insns which set the argument registers into the | |
8457 | delay slot of the millicode call -- thus they act more like traditional | |
8458 | CALL_INSNs. | |
8459 | ||
8460 | Note we can not consider side effects of the insn to be delayed because | |
8461 | the branch and link insn will clobber the return pointer. If we happened | |
8462 | to use the return pointer in the delay slot of the call, then we lose. | |
14d18de3 | 8463 | |
8464 | get_attr_type will try to recognize the given insn, so make sure to | |
8465 | filter out things it will not accept -- SEQUENCE, USE and CLOBBER insns | |
8466 | in particular. */ | |
8467 | int | |
5c1d8983 | 8468 | insn_refs_are_delayed (rtx insn) |
14d18de3 | 8469 | { |
9840d99d | 8470 | return ((GET_CODE (insn) == INSN |
14d18de3 | 8471 | && GET_CODE (PATTERN (insn)) != SEQUENCE |
8472 | && GET_CODE (PATTERN (insn)) != USE | |
8473 | && GET_CODE (PATTERN (insn)) != CLOBBER | |
8474 | && get_attr_type (insn) == TYPE_MILLI)); | |
8475 | } | |
5cb4669a | 8476 | |
58a72cce | 8477 | /* On the HP-PA the value is found in register(s) 28(-29), unless |
8478 | the mode is SF or DF. Then the value is returned in fr4 (32). | |
8479 | ||
8480 | This must perform the same promotions as PROMOTE_MODE, else | |
8481 | PROMOTE_FUNCTION_RETURN will not work correctly. | |
8482 | ||
8483 | Small structures must be returned in a PARALLEL on PA64 in order | |
8484 | to match the HP Compiler ABI. */ | |
8485 | ||
8486 | rtx | |
5c1d8983 | 8487 | function_value (tree valtype, tree func ATTRIBUTE_UNUSED) |
58a72cce | 8488 | { |
8489 | enum machine_mode valmode; | |
8490 | ||
8491 | /* Aggregates with a size less than or equal to 128 bits are returned | |
8492 | in GR 28(-29). They are left justified. The pad bits are undefined. | |
8493 | Larger aggregates are returned in memory. */ | |
8494 | if (TARGET_64BIT && AGGREGATE_TYPE_P (valtype)) | |
8495 | { | |
8496 | rtx loc[2]; | |
8497 | int i, offset = 0; | |
8498 | int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2; | |
8499 | ||
8500 | for (i = 0; i < ub; i++) | |
8501 | { | |
8502 | loc[i] = gen_rtx_EXPR_LIST (VOIDmode, | |
8503 | gen_rtx_REG (DImode, 28 + i), | |
8504 | GEN_INT (offset)); | |
8505 | offset += 8; | |
8506 | } | |
8507 | ||
8508 | return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc)); | |
8509 | } | |
8510 | ||
8511 | if ((INTEGRAL_TYPE_P (valtype) | |
8512 | && TYPE_PRECISION (valtype) < BITS_PER_WORD) | |
8513 | || POINTER_TYPE_P (valtype)) | |
8514 | valmode = word_mode; | |
8515 | else | |
8516 | valmode = TYPE_MODE (valtype); | |
8517 | ||
8518 | if (TREE_CODE (valtype) == REAL_TYPE | |
8519 | && TYPE_MODE (valtype) != TFmode | |
8520 | && !TARGET_SOFT_FLOAT) | |
8521 | return gen_rtx_REG (valmode, 32); | |
8522 | ||
8523 | return gen_rtx_REG (valmode, 28); | |
8524 | } | |
8525 | ||
5e3c5739 | 8526 | /* Return the location of a parameter that is passed in a register or NULL |
8527 | if the parameter has any component that is passed in memory. | |
8528 | ||
8529 | This is new code and will be pushed to into the net sources after | |
9840d99d | 8530 | further testing. |
5e3c5739 | 8531 | |
8532 | ??? We might want to restructure this so that it looks more like other | |
8533 | ports. */ | |
8534 | rtx | |
5c1d8983 | 8535 | function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type, |
8536 | int named ATTRIBUTE_UNUSED) | |
5e3c5739 | 8537 | { |
8538 | int max_arg_words = (TARGET_64BIT ? 8 : 4); | |
2a075f91 | 8539 | int alignment = 0; |
ac965869 | 8540 | int arg_size; |
5e3c5739 | 8541 | int fpr_reg_base; |
8542 | int gpr_reg_base; | |
8543 | rtx retval; | |
8544 | ||
ac965869 | 8545 | if (mode == VOIDmode) |
8546 | return NULL_RTX; | |
8547 | ||
8548 | arg_size = FUNCTION_ARG_SIZE (mode, type); | |
8549 | ||
8550 | /* If this arg would be passed partially or totally on the stack, then | |
8551 | this routine should return zero. FUNCTION_ARG_PARTIAL_NREGS will | |
8552 | handle arguments which are split between regs and stack slots if | |
8553 | the ABI mandates split arguments. */ | |
5e3c5739 | 8554 | if (! TARGET_64BIT) |
8555 | { | |
ac965869 | 8556 | /* The 32-bit ABI does not split arguments. */ |
8557 | if (cum->words + arg_size > max_arg_words) | |
5e3c5739 | 8558 | return NULL_RTX; |
8559 | } | |
8560 | else | |
8561 | { | |
2a075f91 | 8562 | if (arg_size > 1) |
8563 | alignment = cum->words & 1; | |
ac965869 | 8564 | if (cum->words + alignment >= max_arg_words) |
5e3c5739 | 8565 | return NULL_RTX; |
8566 | } | |
8567 | ||
8568 | /* The 32bit ABIs and the 64bit ABIs are rather different, | |
8569 | particularly in their handling of FP registers. We might | |
8570 | be able to cleverly share code between them, but I'm not | |
9aadea62 | 8571 | going to bother in the hope that splitting them up results |
2a075f91 | 8572 | in code that is more easily understood. */ |
5e3c5739 | 8573 | |
5e3c5739 | 8574 | if (TARGET_64BIT) |
8575 | { | |
8576 | /* Advance the base registers to their current locations. | |
8577 | ||
8578 | Remember, gprs grow towards smaller register numbers while | |
2a075f91 | 8579 | fprs grow to higher register numbers. Also remember that |
8580 | although FP regs are 32-bit addressable, we pretend that | |
8581 | the registers are 64-bits wide. */ | |
5e3c5739 | 8582 | gpr_reg_base = 26 - cum->words; |
8583 | fpr_reg_base = 32 + cum->words; | |
9840d99d | 8584 | |
ac965869 | 8585 | /* Arguments wider than one word and small aggregates need special |
8586 | treatment. */ | |
8587 | if (arg_size > 1 | |
8588 | || mode == BLKmode | |
8589 | || (type && AGGREGATE_TYPE_P (type))) | |
5e3c5739 | 8590 | { |
2a075f91 | 8591 | /* Double-extended precision (80-bit), quad-precision (128-bit) |
8592 | and aggregates including complex numbers are aligned on | |
8593 | 128-bit boundaries. The first eight 64-bit argument slots | |
8594 | are associated one-to-one, with general registers r26 | |
8595 | through r19, and also with floating-point registers fr4 | |
8596 | through fr11. Arguments larger than one word are always | |
ac965869 | 8597 | passed in general registers. |
8598 | ||
8599 | Using a PARALLEL with a word mode register results in left | |
8600 | justified data on a big-endian target. */ | |
2a075f91 | 8601 | |
8602 | rtx loc[8]; | |
8603 | int i, offset = 0, ub = arg_size; | |
8604 | ||
8605 | /* Align the base register. */ | |
8606 | gpr_reg_base -= alignment; | |
8607 | ||
8608 | ub = MIN (ub, max_arg_words - cum->words - alignment); | |
8609 | for (i = 0; i < ub; i++) | |
5e3c5739 | 8610 | { |
2a075f91 | 8611 | loc[i] = gen_rtx_EXPR_LIST (VOIDmode, |
8612 | gen_rtx_REG (DImode, gpr_reg_base), | |
8613 | GEN_INT (offset)); | |
8614 | gpr_reg_base -= 1; | |
8615 | offset += 8; | |
5e3c5739 | 8616 | } |
2a075f91 | 8617 | |
8618 | return gen_rtx_PARALLEL (mode, gen_rtvec_v (ub, loc)); | |
5e3c5739 | 8619 | } |
ac965869 | 8620 | } |
5e3c5739 | 8621 | else |
8622 | { | |
8623 | /* If the argument is larger than a word, then we know precisely | |
8624 | which registers we must use. */ | |
2a075f91 | 8625 | if (arg_size > 1) |
5e3c5739 | 8626 | { |
8627 | if (cum->words) | |
8628 | { | |
8629 | gpr_reg_base = 23; | |
8630 | fpr_reg_base = 38; | |
8631 | } | |
8632 | else | |
8633 | { | |
8634 | gpr_reg_base = 25; | |
8635 | fpr_reg_base = 34; | |
8636 | } | |
ac965869 | 8637 | |
8638 | /* Structures 5 to 8 bytes in size are passed in the general | |
8639 | registers in the same manner as other non floating-point | |
8640 | objects. The data is right-justified and zero-extended | |
8641 | to 64 bits. | |
8642 | ||
8643 | This is magic. Normally, using a PARALLEL results in left | |
8644 | justified data on a big-endian target. However, using a | |
8645 | single double-word register provides the required right | |
2cecd772 | 8646 | justification for 5 to 8 byte structures. This has nothing |
ac965869 | 8647 | to do with the direction of padding specified for the argument. |
8648 | It has to do with how the data is widened and shifted into | |
8649 | and from the register. | |
8650 | ||
8651 | Aside from adding load_multiple and store_multiple patterns, | |
8652 | this is the only way that I have found to obtain right | |
8653 | justification of BLKmode data when it has a size greater | |
8654 | than one word. Splitting the operation into two SImode loads | |
8655 | or returning a DImode REG results in left justified data. */ | |
8656 | if (mode == BLKmode) | |
8657 | { | |
58a72cce | 8658 | rtx loc = gen_rtx_EXPR_LIST (VOIDmode, |
8659 | gen_rtx_REG (DImode, gpr_reg_base), | |
8660 | const0_rtx); | |
8661 | return gen_rtx_PARALLEL (mode, gen_rtvec (1, loc)); | |
ac965869 | 8662 | } |
5e3c5739 | 8663 | } |
8664 | else | |
8665 | { | |
8666 | /* We have a single word (32 bits). A simple computation | |
8667 | will get us the register #s we need. */ | |
8668 | gpr_reg_base = 26 - cum->words; | |
8669 | fpr_reg_base = 32 + 2 * cum->words; | |
8670 | } | |
8671 | } | |
8672 | ||
9503b0f7 | 8673 | /* Determine if the argument needs to be passed in both general and |
5e3c5739 | 8674 | floating point registers. */ |
9503b0f7 | 8675 | if (((TARGET_PORTABLE_RUNTIME || TARGET_64BIT || TARGET_ELF32) |
8676 | /* If we are doing soft-float with portable runtime, then there | |
8677 | is no need to worry about FP regs. */ | |
f336e0bc | 8678 | && !TARGET_SOFT_FLOAT |
9503b0f7 | 8679 | /* The parameter must be some kind of float, else we can just |
8680 | pass it in integer registers. */ | |
8681 | && FLOAT_MODE_P (mode) | |
8682 | /* The target function must not have a prototype. */ | |
8683 | && cum->nargs_prototype <= 0 | |
8684 | /* libcalls do not need to pass items in both FP and general | |
8685 | registers. */ | |
8686 | && type != NULL_TREE | |
f336e0bc | 8687 | /* All this hair applies to "outgoing" args only. This includes |
8688 | sibcall arguments setup with FUNCTION_INCOMING_ARG. */ | |
8689 | && !cum->incoming) | |
9503b0f7 | 8690 | /* Also pass outgoing floating arguments in both registers in indirect |
8691 | calls with the 32 bit ABI and the HP assembler since there is no | |
8692 | way to the specify argument locations in static functions. */ | |
f336e0bc | 8693 | || (!TARGET_64BIT |
8694 | && !TARGET_GAS | |
8695 | && !cum->incoming | |
9503b0f7 | 8696 | && cum->indirect |
8697 | && FLOAT_MODE_P (mode))) | |
5e3c5739 | 8698 | { |
8699 | retval | |
8700 | = gen_rtx_PARALLEL | |
8701 | (mode, | |
8702 | gen_rtvec (2, | |
8703 | gen_rtx_EXPR_LIST (VOIDmode, | |
8704 | gen_rtx_REG (mode, fpr_reg_base), | |
8705 | const0_rtx), | |
8706 | gen_rtx_EXPR_LIST (VOIDmode, | |
8707 | gen_rtx_REG (mode, gpr_reg_base), | |
8708 | const0_rtx))); | |
8709 | } | |
8710 | else | |
8711 | { | |
8712 | /* See if we should pass this parameter in a general register. */ | |
8713 | if (TARGET_SOFT_FLOAT | |
8714 | /* Indirect calls in the normal 32bit ABI require all arguments | |
8715 | to be passed in general registers. */ | |
8716 | || (!TARGET_PORTABLE_RUNTIME | |
8717 | && !TARGET_64BIT | |
a052da6f | 8718 | && !TARGET_ELF32 |
5e3c5739 | 8719 | && cum->indirect) |
8720 | /* If the parameter is not a floating point parameter, then | |
8721 | it belongs in GPRs. */ | |
8722 | || !FLOAT_MODE_P (mode)) | |
8723 | retval = gen_rtx_REG (mode, gpr_reg_base); | |
8724 | else | |
8725 | retval = gen_rtx_REG (mode, fpr_reg_base); | |
8726 | } | |
8727 | return retval; | |
8728 | } | |
8729 | ||
8730 | ||
8731 | /* If this arg would be passed totally in registers or totally on the stack, | |
8732 | then this routine should return zero. It is currently called only for | |
6dc3b0d9 | 8733 | the 64-bit target. */ |
5e3c5739 | 8734 | int |
5c1d8983 | 8735 | function_arg_partial_nregs (CUMULATIVE_ARGS *cum, enum machine_mode mode, |
8736 | tree type, int named ATTRIBUTE_UNUSED) | |
5e3c5739 | 8737 | { |
b7d86581 | 8738 | unsigned int max_arg_words = 8; |
8739 | unsigned int offset = 0; | |
5e3c5739 | 8740 | |
b7d86581 | 8741 | if (FUNCTION_ARG_SIZE (mode, type) > 1 && (cum->words & 1)) |
5e3c5739 | 8742 | offset = 1; |
8743 | ||
b7d86581 | 8744 | if (cum->words + offset + FUNCTION_ARG_SIZE (mode, type) <= max_arg_words) |
6dc3b0d9 | 8745 | /* Arg fits fully into registers. */ |
5e3c5739 | 8746 | return 0; |
9840d99d | 8747 | else if (cum->words + offset >= max_arg_words) |
6dc3b0d9 | 8748 | /* Arg fully on the stack. */ |
5e3c5739 | 8749 | return 0; |
8750 | else | |
6dc3b0d9 | 8751 | /* Arg is split. */ |
5e3c5739 | 8752 | return max_arg_words - cum->words - offset; |
5e3c5739 | 8753 | } |
8754 | ||
8755 | ||
8756 | /* Return 1 if this is a comparison operator. This allows the use of | |
8757 | MATCH_OPERATOR to recognize all the branch insns. */ | |
8758 | ||
8759 | int | |
5c1d8983 | 8760 | cmpib_comparison_operator (rtx op, enum machine_mode mode) |
5e3c5739 | 8761 | { |
8762 | return ((mode == VOIDmode || GET_MODE (op) == mode) | |
8763 | && (GET_CODE (op) == EQ | |
8764 | || GET_CODE (op) == NE | |
8765 | || GET_CODE (op) == GT | |
5e3c5739 | 8766 | || GET_CODE (op) == GTU |
32509e56 | 8767 | || GET_CODE (op) == GE |
5e3c5739 | 8768 | || GET_CODE (op) == LT |
8769 | || GET_CODE (op) == LE | |
8770 | || GET_CODE (op) == LEU)); | |
8771 | } | |
8772 | ||
52470889 | 8773 | /* On hpux10, the linker will give an error if we have a reference |
8774 | in the read-only data section to a symbol defined in a shared | |
8775 | library. Therefore, expressions that might require a reloc can | |
8776 | not be placed in the read-only data section. */ | |
8777 | ||
8778 | static void | |
b572d1a5 | 8779 | pa_select_section (tree exp, int reloc, |
8780 | unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED) | |
52470889 | 8781 | { |
8782 | if (TREE_CODE (exp) == VAR_DECL | |
8783 | && TREE_READONLY (exp) | |
8784 | && !TREE_THIS_VOLATILE (exp) | |
8785 | && DECL_INITIAL (exp) | |
8786 | && (DECL_INITIAL (exp) == error_mark_node | |
8787 | || TREE_CONSTANT (DECL_INITIAL (exp))) | |
8788 | && !reloc) | |
8789 | readonly_data_section (); | |
8790 | else if (TREE_CODE_CLASS (TREE_CODE (exp)) == 'c' | |
8791 | && !(TREE_CODE (exp) == STRING_CST && flag_writable_strings) | |
8792 | && !reloc) | |
8793 | readonly_data_section (); | |
8794 | else | |
8795 | data_section (); | |
8796 | } | |
1f3233d1 | 8797 | |
67c1e638 | 8798 | static void |
5c1d8983 | 8799 | pa_globalize_label (FILE *stream, const char *name) |
67c1e638 | 8800 | { |
8801 | /* We only handle DATA objects here, functions are globalized in | |
8802 | ASM_DECLARE_FUNCTION_NAME. */ | |
8803 | if (! FUNCTION_NAME_P (name)) | |
8804 | { | |
8805 | fputs ("\t.EXPORT ", stream); | |
8806 | assemble_name (stream, name); | |
8807 | fputs (",DATA\n", stream); | |
8808 | } | |
8809 | } | |
1f3233d1 | 8810 | #include "gt-pa.h" |