]>
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 | ||
6 | This file is part of GNU CC. | |
7 | ||
8 | GNU CC is free software; you can redistribute it and/or modify | |
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 | ||
13 | GNU CC is distributed in the hope that it will be useful, | |
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 | |
19 | along with GNU CC; 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" |
40 | #include "libfuncs.h" | |
41 | #include "reload.h" | |
87ad11b0 | 42 | #include "c-tree.h" |
d7e2f694 | 43 | #include "integrate.h" |
0a893c29 | 44 | #include "function.h" |
d6f01525 | 45 | #include "obstack.h" |
b1ca791d | 46 | #include "toplev.h" |
5cb4669a | 47 | #include "ggc.h" |
611a88e1 | 48 | #include "recog.h" |
a584fe8a | 49 | #include "predict.h" |
611a88e1 | 50 | #include "tm_p.h" |
a767736d | 51 | #include "target.h" |
52 | #include "target-def.h" | |
87ad11b0 | 53 | |
bea4bad2 | 54 | static int hppa_use_dfa_pipeline_interface PARAMS ((void)); |
55 | ||
56 | #undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE | |
57 | #define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE hppa_use_dfa_pipeline_interface | |
58 | ||
59 | static int | |
60 | hppa_use_dfa_pipeline_interface () | |
61 | { | |
62 | return 1; | |
63 | } | |
64 | ||
cde3e16c | 65 | /* Return nonzero if there is a bypass for the output of |
66 | OUT_INSN and the fp store IN_INSN. */ | |
67 | int | |
68 | hppa_fpstore_bypass_p (out_insn, in_insn) | |
69 | rtx out_insn, in_insn; | |
70 | { | |
71 | enum machine_mode store_mode; | |
72 | enum machine_mode other_mode; | |
73 | rtx set; | |
74 | ||
75 | if (recog_memoized (in_insn) < 0 | |
76 | || get_attr_type (in_insn) != TYPE_FPSTORE | |
77 | || recog_memoized (out_insn) < 0) | |
78 | return 0; | |
79 | ||
80 | store_mode = GET_MODE (SET_SRC (PATTERN (in_insn))); | |
81 | ||
82 | set = single_set (out_insn); | |
83 | if (!set) | |
84 | return 0; | |
85 | ||
86 | other_mode = GET_MODE (SET_SRC (set)); | |
87 | ||
88 | return (GET_MODE_SIZE (store_mode) == GET_MODE_SIZE (other_mode)); | |
89 | } | |
90 | ||
91 | ||
cc858176 | 92 | #ifndef DO_FRAME_NOTES |
93 | #ifdef INCOMING_RETURN_ADDR_RTX | |
94 | #define DO_FRAME_NOTES 1 | |
95 | #else | |
96 | #define DO_FRAME_NOTES 0 | |
97 | #endif | |
98 | #endif | |
99 | ||
ec0457a8 | 100 | static int hppa_address_cost PARAMS ((rtx)); |
fab7adbf | 101 | static bool hppa_rtx_costs PARAMS ((rtx, int, int, int *)); |
9840d99d | 102 | static inline rtx force_mode PARAMS ((enum machine_mode, rtx)); |
cc858176 | 103 | static void pa_combine_instructions PARAMS ((rtx)); |
104 | static int pa_can_combine_p PARAMS ((rtx, rtx, rtx, int, rtx, rtx, rtx)); | |
105 | static int forward_branch_p PARAMS ((rtx)); | |
106 | static int shadd_constant_p PARAMS ((int)); | |
cc858176 | 107 | static void compute_zdepwi_operands PARAMS ((unsigned HOST_WIDE_INT, unsigned *)); |
108 | static int compute_movstrsi_length PARAMS ((rtx)); | |
58356836 | 109 | static bool pa_assemble_integer PARAMS ((rtx, unsigned int, int)); |
cc858176 | 110 | static void remove_useless_addtr_insns PARAMS ((rtx, int)); |
6a2c16d6 | 111 | static void store_reg PARAMS ((int, int, int)); |
a584fe8a | 112 | static void store_reg_modify PARAMS ((int, int, int)); |
6a2c16d6 | 113 | static void load_reg PARAMS ((int, int, int)); |
a584fe8a | 114 | static void set_reg_plus_d PARAMS ((int, int, int, int)); |
6988553d | 115 | static void pa_output_function_prologue PARAMS ((FILE *, HOST_WIDE_INT)); |
17d9b0c3 | 116 | static void pa_output_function_epilogue PARAMS ((FILE *, HOST_WIDE_INT)); |
747af5e7 | 117 | static int pa_adjust_cost PARAMS ((rtx, rtx, rtx, int)); |
118 | static int pa_adjust_priority PARAMS ((rtx, int)); | |
119 | static int pa_issue_rate PARAMS ((void)); | |
52470889 | 120 | static void pa_select_section PARAMS ((tree, int, unsigned HOST_WIDE_INT)) |
121 | ATTRIBUTE_UNUSED; | |
7811991d | 122 | static void pa_encode_section_info PARAMS ((tree, int)); |
7b4a38a6 | 123 | static const char *pa_strip_name_encoding PARAMS ((const char *)); |
805e22b2 | 124 | static bool pa_function_ok_for_sibcall PARAMS ((tree, tree)); |
63b8cd48 | 125 | static void pa_globalize_label PARAMS ((FILE *, const char *)) |
126 | ATTRIBUTE_UNUSED; | |
eb344f43 | 127 | static void pa_asm_output_mi_thunk PARAMS ((FILE *, tree, HOST_WIDE_INT, |
128 | HOST_WIDE_INT, tree)); | |
de419443 | 129 | #if !defined(USE_COLLECT2) |
130 | static void pa_asm_out_constructor PARAMS ((rtx, int)); | |
131 | static void pa_asm_out_destructor PARAMS ((rtx, int)); | |
132 | #endif | |
ece88821 | 133 | static void copy_fp_args PARAMS ((rtx)) ATTRIBUTE_UNUSED; |
134 | static int length_fp_args PARAMS ((rtx)) ATTRIBUTE_UNUSED; | |
135 | static struct deferred_plabel *get_plabel PARAMS ((const char *)) | |
136 | ATTRIBUTE_UNUSED; | |
9d3ddb8f | 137 | |
87ad11b0 | 138 | /* Save the operands last given to a compare for use when we |
139 | generate a scc or bcc insn. */ | |
87ad11b0 | 140 | rtx hppa_compare_op0, hppa_compare_op1; |
141 | enum cmp_type hppa_branch_type; | |
142 | ||
134b4858 | 143 | /* Which cpu we are scheduling for. */ |
144 | enum processor_type pa_cpu; | |
145 | ||
146 | /* String to hold which cpu we are scheduling for. */ | |
611a88e1 | 147 | const char *pa_cpu_string; |
134b4858 | 148 | |
4dc42288 | 149 | /* Which architecture we are generating code for. */ |
150 | enum architecture_type pa_arch; | |
151 | ||
152 | /* String to hold which architecture we are generating code for. */ | |
611a88e1 | 153 | const char *pa_arch_string; |
4dc42288 | 154 | |
a9960cdc | 155 | /* Counts for the number of callee-saved general and floating point |
156 | registers which were saved by the current function's prologue. */ | |
157 | static int gr_saved, fr_saved; | |
158 | ||
611a88e1 | 159 | static rtx find_addr_reg PARAMS ((rtx)); |
87ad11b0 | 160 | |
06ddb6f8 | 161 | /* Keep track of the number of bytes we have output in the CODE subspaces |
162 | during this compilation so we'll know when to emit inline long-calls. */ | |
ece88821 | 163 | unsigned long total_code_bytes; |
06ddb6f8 | 164 | |
e3f53689 | 165 | /* Variables to handle plabels that we discover are necessary at assembly |
01cc3b75 | 166 | output time. They are output after the current function. */ |
1f3233d1 | 167 | struct deferred_plabel GTY(()) |
e3f53689 | 168 | { |
169 | rtx internal_label; | |
8d0cec1a | 170 | const char *name; |
1f3233d1 | 171 | }; |
172 | static GTY((length ("n_deferred_plabels"))) struct deferred_plabel * | |
173 | deferred_plabels; | |
e11bd7e5 | 174 | static size_t n_deferred_plabels = 0; |
a767736d | 175 | \f |
176 | /* Initialize the GCC target structure. */ | |
58356836 | 177 | |
178 | #undef TARGET_ASM_ALIGNED_HI_OP | |
179 | #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t" | |
180 | #undef TARGET_ASM_ALIGNED_SI_OP | |
181 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
182 | #undef TARGET_ASM_ALIGNED_DI_OP | |
183 | #define TARGET_ASM_ALIGNED_DI_OP "\t.dword\t" | |
184 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
185 | #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP | |
186 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
187 | #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP | |
188 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
189 | #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP | |
190 | #undef TARGET_ASM_INTEGER | |
191 | #define TARGET_ASM_INTEGER pa_assemble_integer | |
192 | ||
17d9b0c3 | 193 | #undef TARGET_ASM_FUNCTION_PROLOGUE |
194 | #define TARGET_ASM_FUNCTION_PROLOGUE pa_output_function_prologue | |
195 | #undef TARGET_ASM_FUNCTION_EPILOGUE | |
196 | #define TARGET_ASM_FUNCTION_EPILOGUE pa_output_function_epilogue | |
e3f53689 | 197 | |
747af5e7 | 198 | #undef TARGET_SCHED_ADJUST_COST |
199 | #define TARGET_SCHED_ADJUST_COST pa_adjust_cost | |
200 | #undef TARGET_SCHED_ADJUST_PRIORITY | |
201 | #define TARGET_SCHED_ADJUST_PRIORITY pa_adjust_priority | |
202 | #undef TARGET_SCHED_ISSUE_RATE | |
203 | #define TARGET_SCHED_ISSUE_RATE pa_issue_rate | |
204 | ||
7811991d | 205 | #undef TARGET_ENCODE_SECTION_INFO |
206 | #define TARGET_ENCODE_SECTION_INFO pa_encode_section_info | |
7b4a38a6 | 207 | #undef TARGET_STRIP_NAME_ENCODING |
208 | #define TARGET_STRIP_NAME_ENCODING pa_strip_name_encoding | |
7811991d | 209 | |
805e22b2 | 210 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
211 | #define TARGET_FUNCTION_OK_FOR_SIBCALL pa_function_ok_for_sibcall | |
212 | ||
6988553d | 213 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
214 | #define TARGET_ASM_OUTPUT_MI_THUNK pa_asm_output_mi_thunk | |
eb344f43 | 215 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK |
216 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall | |
6988553d | 217 | |
de419443 | 218 | #if !defined(USE_COLLECT2) |
219 | #undef TARGET_ASM_CONSTRUCTOR | |
220 | #define TARGET_ASM_CONSTRUCTOR pa_asm_out_constructor | |
221 | #undef TARGET_ASM_DESTRUCTOR | |
222 | #define TARGET_ASM_DESTRUCTOR pa_asm_out_destructor | |
223 | #endif | |
224 | ||
fab7adbf | 225 | #undef TARGET_RTX_COSTS |
226 | #define TARGET_RTX_COSTS hppa_rtx_costs | |
ec0457a8 | 227 | #undef TARGET_ADDRESS_COST |
228 | #define TARGET_ADDRESS_COST hppa_address_cost | |
fab7adbf | 229 | |
57e4bbfb | 230 | struct gcc_target targetm = TARGET_INITIALIZER; |
a767736d | 231 | \f |
134b4858 | 232 | void |
233 | override_options () | |
234 | { | |
b3b62da8 | 235 | if (pa_cpu_string == NULL) |
236 | pa_cpu_string = TARGET_SCHED_DEFAULT; | |
237 | ||
238 | if (! strcmp (pa_cpu_string, "8000")) | |
239 | { | |
240 | pa_cpu_string = "8000"; | |
241 | pa_cpu = PROCESSOR_8000; | |
242 | } | |
243 | else if (! strcmp (pa_cpu_string, "7100")) | |
134b4858 | 244 | { |
245 | pa_cpu_string = "7100"; | |
246 | pa_cpu = PROCESSOR_7100; | |
247 | } | |
b3b62da8 | 248 | else if (! strcmp (pa_cpu_string, "700")) |
a87e1e15 | 249 | { |
250 | pa_cpu_string = "700"; | |
251 | pa_cpu = PROCESSOR_700; | |
252 | } | |
b3b62da8 | 253 | else if (! strcmp (pa_cpu_string, "7100LC")) |
134b4858 | 254 | { |
255 | pa_cpu_string = "7100LC"; | |
256 | pa_cpu = PROCESSOR_7100LC; | |
257 | } | |
b3b62da8 | 258 | else if (! strcmp (pa_cpu_string, "7200")) |
b4e0d953 | 259 | { |
260 | pa_cpu_string = "7200"; | |
261 | pa_cpu = PROCESSOR_7200; | |
262 | } | |
b3b62da8 | 263 | else if (! strcmp (pa_cpu_string, "7300")) |
bea4bad2 | 264 | { |
265 | pa_cpu_string = "7300"; | |
266 | pa_cpu = PROCESSOR_7300; | |
267 | } | |
134b4858 | 268 | else |
269 | { | |
bea4bad2 | 270 | warning ("unknown -mschedule= option (%s).\nValid options are 700, 7100, 7100LC, 7200, 7300, and 8000\n", pa_cpu_string); |
134b4858 | 271 | } |
c7a4e712 | 272 | |
4dc42288 | 273 | /* Set the instruction set architecture. */ |
274 | if (pa_arch_string && ! strcmp (pa_arch_string, "1.0")) | |
275 | { | |
276 | pa_arch_string = "1.0"; | |
277 | pa_arch = ARCHITECTURE_10; | |
278 | target_flags &= ~(MASK_PA_11 | MASK_PA_20); | |
279 | } | |
280 | else if (pa_arch_string && ! strcmp (pa_arch_string, "1.1")) | |
281 | { | |
282 | pa_arch_string = "1.1"; | |
283 | pa_arch = ARCHITECTURE_11; | |
284 | target_flags &= ~MASK_PA_20; | |
285 | target_flags |= MASK_PA_11; | |
286 | } | |
287 | else if (pa_arch_string && ! strcmp (pa_arch_string, "2.0")) | |
288 | { | |
289 | pa_arch_string = "2.0"; | |
290 | pa_arch = ARCHITECTURE_20; | |
291 | target_flags |= MASK_PA_11 | MASK_PA_20; | |
292 | } | |
293 | else if (pa_arch_string) | |
294 | { | |
68435912 | 295 | warning ("unknown -march= option (%s).\nValid options are 1.0, 1.1, and 2.0\n", pa_arch_string); |
4dc42288 | 296 | } |
297 | ||
7c5101fc | 298 | /* Unconditional branches in the delay slot are not compatible with dwarf2 |
299 | call frame information. There is no benefit in using this optimization | |
300 | on PA8000 and later processors. */ | |
301 | if (pa_cpu >= PROCESSOR_8000 | |
302 | || (! USING_SJLJ_EXCEPTIONS && flag_exceptions) | |
303 | || flag_unwind_tables) | |
304 | target_flags &= ~MASK_JUMP_IN_DELAY; | |
305 | ||
c7a4e712 | 306 | if (flag_pic && TARGET_PORTABLE_RUNTIME) |
307 | { | |
308 | warning ("PIC code generation is not supported in the portable runtime model\n"); | |
309 | } | |
310 | ||
b29897dd | 311 | if (flag_pic && TARGET_FAST_INDIRECT_CALLS) |
c7a4e712 | 312 | { |
ad87de1e | 313 | warning ("PIC code generation is not compatible with fast indirect calls\n"); |
c7a4e712 | 314 | } |
751e64a1 | 315 | |
5bd7b548 | 316 | if (! TARGET_GAS && write_symbols != NO_DEBUG) |
317 | { | |
318 | warning ("-g is only supported when using GAS on this processor,"); | |
68435912 | 319 | warning ("-g option disabled"); |
5bd7b548 | 320 | write_symbols = NO_DEBUG; |
321 | } | |
5cb4669a | 322 | |
fc44315f | 323 | /* We only support the "big PIC" model now. And we always generate PIC |
324 | code when in 64bit mode. */ | |
325 | if (flag_pic == 1 || TARGET_64BIT) | |
5e3c5739 | 326 | flag_pic = 2; |
327 | ||
58356836 | 328 | /* We can't guarantee that .dword is available for 32-bit targets. */ |
329 | if (UNITS_PER_WORD == 4) | |
330 | targetm.asm_out.aligned_op.di = NULL; | |
331 | ||
332 | /* The unaligned ops are only available when using GAS. */ | |
333 | if (!TARGET_GAS) | |
334 | { | |
335 | targetm.asm_out.unaligned_op.hi = NULL; | |
336 | targetm.asm_out.unaligned_op.si = NULL; | |
337 | targetm.asm_out.unaligned_op.di = NULL; | |
338 | } | |
134b4858 | 339 | } |
340 | ||
e911aedf | 341 | /* Return nonzero only if OP is a register of mode MODE, |
891b55b4 | 342 | or CONST0_RTX. */ |
87ad11b0 | 343 | int |
344 | reg_or_0_operand (op, mode) | |
345 | rtx op; | |
346 | enum machine_mode mode; | |
347 | { | |
891b55b4 | 348 | return (op == CONST0_RTX (mode) || register_operand (op, mode)); |
87ad11b0 | 349 | } |
350 | ||
e911aedf | 351 | /* Return nonzero if OP is suitable for use in a call to a named |
575e0eb4 | 352 | function. |
353 | ||
9dc4cbf1 | 354 | For 2.5 try to eliminate either call_operand_address or |
575e0eb4 | 355 | function_label_operand, they perform very similar functions. */ |
87ad11b0 | 356 | int |
357 | call_operand_address (op, mode) | |
358 | rtx op; | |
b1ca791d | 359 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 360 | { |
2db59d96 | 361 | return (GET_MODE (op) == word_mode |
362 | && CONSTANT_P (op) && ! TARGET_PORTABLE_RUNTIME); | |
87ad11b0 | 363 | } |
364 | ||
6d36483b | 365 | /* Return 1 if X contains a symbolic expression. We know these |
366 | expressions will have one of a few well defined forms, so | |
347b5848 | 367 | we need only check those forms. */ |
368 | int | |
369 | symbolic_expression_p (x) | |
370 | register rtx x; | |
371 | { | |
372 | ||
6dc3b0d9 | 373 | /* Strip off any HIGH. */ |
347b5848 | 374 | if (GET_CODE (x) == HIGH) |
375 | x = XEXP (x, 0); | |
376 | ||
377 | return (symbolic_operand (x, VOIDmode)); | |
378 | } | |
379 | ||
87ad11b0 | 380 | int |
381 | symbolic_operand (op, mode) | |
382 | register rtx op; | |
b1ca791d | 383 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 384 | { |
385 | switch (GET_CODE (op)) | |
386 | { | |
387 | case SYMBOL_REF: | |
388 | case LABEL_REF: | |
389 | return 1; | |
390 | case CONST: | |
391 | op = XEXP (op, 0); | |
392 | return ((GET_CODE (XEXP (op, 0)) == SYMBOL_REF | |
393 | || GET_CODE (XEXP (op, 0)) == LABEL_REF) | |
394 | && GET_CODE (XEXP (op, 1)) == CONST_INT); | |
395 | default: | |
396 | return 0; | |
397 | } | |
398 | } | |
399 | ||
400 | /* Return truth value of statement that OP is a symbolic memory | |
401 | operand of mode MODE. */ | |
402 | ||
403 | int | |
404 | symbolic_memory_operand (op, mode) | |
405 | rtx op; | |
b1ca791d | 406 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 407 | { |
408 | if (GET_CODE (op) == SUBREG) | |
409 | op = SUBREG_REG (op); | |
410 | if (GET_CODE (op) != MEM) | |
411 | return 0; | |
412 | op = XEXP (op, 0); | |
413 | return (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == CONST | |
414 | || GET_CODE (op) == HIGH || GET_CODE (op) == LABEL_REF); | |
415 | } | |
416 | ||
417 | /* Return 1 if the operand is either a register or a memory operand that is | |
418 | not symbolic. */ | |
419 | ||
420 | int | |
421 | reg_or_nonsymb_mem_operand (op, mode) | |
422 | register rtx op; | |
423 | enum machine_mode mode; | |
424 | { | |
425 | if (register_operand (op, mode)) | |
426 | return 1; | |
427 | ||
428 | if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode)) | |
429 | return 1; | |
430 | ||
431 | return 0; | |
432 | } | |
433 | ||
6d36483b | 434 | /* Return 1 if the operand is either a register, zero, or a memory operand |
891b55b4 | 435 | that is not symbolic. */ |
436 | ||
437 | int | |
438 | reg_or_0_or_nonsymb_mem_operand (op, mode) | |
439 | register rtx op; | |
440 | enum machine_mode mode; | |
441 | { | |
442 | if (register_operand (op, mode)) | |
443 | return 1; | |
444 | ||
445 | if (op == CONST0_RTX (mode)) | |
446 | return 1; | |
447 | ||
448 | if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode)) | |
449 | return 1; | |
450 | ||
451 | return 0; | |
452 | } | |
453 | ||
d3138877 | 454 | /* Return 1 if the operand is a register operand or a non-symbolic memory |
455 | operand after reload. This predicate is used for branch patterns that | |
456 | internally handle register reloading. We need to accept non-symbolic | |
457 | memory operands after reload to ensure that the pattern is still valid | |
458 | if reload didn't find a hard register for the operand. */ | |
459 | ||
460 | int | |
461 | reg_before_reload_operand (op, mode) | |
462 | register rtx op; | |
463 | enum machine_mode mode; | |
464 | { | |
6df54749 | 465 | /* Don't accept a SUBREG since it will need a reload. */ |
466 | if (GET_CODE (op) == SUBREG) | |
467 | return 0; | |
468 | ||
d3138877 | 469 | if (register_operand (op, mode)) |
470 | return 1; | |
471 | ||
472 | if (reload_completed | |
473 | && memory_operand (op, mode) | |
474 | && ! symbolic_memory_operand (op, mode)) | |
475 | return 1; | |
476 | ||
477 | return 0; | |
478 | } | |
479 | ||
7c4d3047 | 480 | /* Accept any constant that can be moved in one instruction into a |
d9d7c968 | 481 | general register. */ |
6d36483b | 482 | int |
d9d7c968 | 483 | cint_ok_for_move (intval) |
6d36483b | 484 | HOST_WIDE_INT intval; |
d9d7c968 | 485 | { |
486 | /* OK if ldo, ldil, or zdepi, can be used. */ | |
d8d5c4ff | 487 | return (CONST_OK_FOR_LETTER_P (intval, 'J') |
488 | || CONST_OK_FOR_LETTER_P (intval, 'N') | |
489 | || CONST_OK_FOR_LETTER_P (intval, 'K')); | |
d9d7c968 | 490 | } |
491 | ||
6ecdbaa1 | 492 | /* Accept anything that can be moved in one instruction into a general |
493 | register. */ | |
87ad11b0 | 494 | int |
495 | move_operand (op, mode) | |
496 | rtx op; | |
497 | enum machine_mode mode; | |
498 | { | |
499 | if (register_operand (op, mode)) | |
500 | return 1; | |
501 | ||
cc9205b8 | 502 | if (GET_CODE (op) == CONSTANT_P_RTX) |
503 | return 1; | |
504 | ||
42faba01 | 505 | if (GET_CODE (op) == CONST_INT) |
d9d7c968 | 506 | return cint_ok_for_move (INTVAL (op)); |
87ad11b0 | 507 | |
87ad11b0 | 508 | if (GET_CODE (op) == SUBREG) |
509 | op = SUBREG_REG (op); | |
510 | if (GET_CODE (op) != MEM) | |
511 | return 0; | |
512 | ||
513 | op = XEXP (op, 0); | |
27ef382d | 514 | |
2d14b1f0 | 515 | /* We consider a LO_SUM DLT reference a move_operand now since it has |
516 | been merged into the normal movsi/movdi patterns. */ | |
5d7431a3 | 517 | if (GET_CODE (op) == LO_SUM |
518 | && GET_CODE (XEXP (op, 0)) == REG | |
519 | && REG_OK_FOR_BASE_P (XEXP (op, 0)) | |
2d14b1f0 | 520 | && GET_CODE (XEXP (op, 1)) == UNSPEC |
521 | && GET_MODE (op) == Pmode) | |
522 | return 1; | |
5d7431a3 | 523 | |
27ef382d | 524 | /* Since move_operand is only used for source operands, we can always |
525 | allow scaled indexing! */ | |
822c8a6c | 526 | if (! TARGET_DISABLE_INDEXING |
527 | && GET_CODE (op) == PLUS | |
27ef382d | 528 | && ((GET_CODE (XEXP (op, 0)) == MULT |
529 | && GET_CODE (XEXP (XEXP (op, 0), 0)) == REG | |
530 | && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT | |
4ad2b0ae | 531 | && INTVAL (XEXP (XEXP (op, 0), 1)) |
532 | == (HOST_WIDE_INT) GET_MODE_SIZE (mode) | |
27ef382d | 533 | && GET_CODE (XEXP (op, 1)) == REG) |
534 | || (GET_CODE (XEXP (op, 1)) == MULT | |
535 | &&GET_CODE (XEXP (XEXP (op, 1), 0)) == REG | |
536 | && GET_CODE (XEXP (XEXP (op, 1), 1)) == CONST_INT | |
4ad2b0ae | 537 | && INTVAL (XEXP (XEXP (op, 1), 1)) |
538 | == (HOST_WIDE_INT) GET_MODE_SIZE (mode) | |
27ef382d | 539 | && GET_CODE (XEXP (op, 0)) == REG))) |
540 | return 1; | |
541 | ||
87ad11b0 | 542 | return memory_address_p (mode, op); |
543 | } | |
544 | ||
6ecdbaa1 | 545 | /* Accept REG and any CONST_INT that can be moved in one instruction into a |
546 | general register. */ | |
547 | int | |
548 | reg_or_cint_move_operand (op, mode) | |
549 | rtx op; | |
550 | enum machine_mode mode; | |
551 | { | |
552 | if (register_operand (op, mode)) | |
553 | return 1; | |
554 | ||
555 | if (GET_CODE (op) == CONST_INT) | |
686b848d | 556 | return cint_ok_for_move (INTVAL (op)); |
557 | ||
6ecdbaa1 | 558 | return 0; |
559 | } | |
560 | ||
87ad11b0 | 561 | int |
b4a7bf10 | 562 | pic_label_operand (op, mode) |
87ad11b0 | 563 | rtx op; |
b1ca791d | 564 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 565 | { |
b4a7bf10 | 566 | if (!flag_pic) |
567 | return 0; | |
568 | ||
569 | switch (GET_CODE (op)) | |
570 | { | |
571 | case LABEL_REF: | |
572 | return 1; | |
b4a7bf10 | 573 | case CONST: |
574 | op = XEXP (op, 0); | |
3c69dc97 | 575 | return (GET_CODE (XEXP (op, 0)) == LABEL_REF |
b4a7bf10 | 576 | && GET_CODE (XEXP (op, 1)) == CONST_INT); |
577 | default: | |
578 | return 0; | |
579 | } | |
87ad11b0 | 580 | } |
581 | ||
87ad11b0 | 582 | int |
583 | fp_reg_operand (op, mode) | |
584 | rtx op; | |
b1ca791d | 585 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 586 | { |
587 | return reg_renumber && FP_REG_P (op); | |
588 | } | |
d6f01525 | 589 | |
87ad11b0 | 590 | \f |
87ad11b0 | 591 | |
87ad11b0 | 592 | /* Return truth value of whether OP can be used as an operand in a |
593 | three operand arithmetic insn that accepts registers of mode MODE | |
594 | or 14-bit signed integers. */ | |
595 | int | |
596 | arith_operand (op, mode) | |
597 | rtx op; | |
598 | enum machine_mode mode; | |
599 | { | |
600 | return (register_operand (op, mode) | |
601 | || (GET_CODE (op) == CONST_INT && INT_14_BITS (op))); | |
602 | } | |
603 | ||
604 | /* Return truth value of whether OP can be used as an operand in a | |
605 | three operand arithmetic insn that accepts registers of mode MODE | |
606 | or 11-bit signed integers. */ | |
607 | int | |
608 | arith11_operand (op, mode) | |
609 | rtx op; | |
610 | enum machine_mode mode; | |
611 | { | |
612 | return (register_operand (op, mode) | |
613 | || (GET_CODE (op) == CONST_INT && INT_11_BITS (op))); | |
614 | } | |
615 | ||
2a91e628 | 616 | /* Return truth value of whether OP can be used as an operand in a |
617 | adddi3 insn. */ | |
618 | int | |
619 | adddi3_operand (op, mode) | |
620 | rtx op; | |
621 | enum machine_mode mode; | |
622 | { | |
623 | return (register_operand (op, mode) | |
624 | || (GET_CODE (op) == CONST_INT | |
625 | && (TARGET_64BIT ? INT_14_BITS (op) : INT_11_BITS (op)))); | |
626 | } | |
627 | ||
6d36483b | 628 | /* A constant integer suitable for use in a PRE_MODIFY memory |
757d4970 | 629 | reference. */ |
42faba01 | 630 | int |
631 | pre_cint_operand (op, mode) | |
632 | rtx op; | |
b1ca791d | 633 | enum machine_mode mode ATTRIBUTE_UNUSED; |
42faba01 | 634 | { |
635 | return (GET_CODE (op) == CONST_INT | |
636 | && INTVAL (op) >= -0x2000 && INTVAL (op) < 0x10); | |
637 | } | |
638 | ||
6d36483b | 639 | /* A constant integer suitable for use in a POST_MODIFY memory |
757d4970 | 640 | reference. */ |
641 | int | |
642 | post_cint_operand (op, mode) | |
643 | rtx op; | |
b1ca791d | 644 | enum machine_mode mode ATTRIBUTE_UNUSED; |
757d4970 | 645 | { |
646 | return (GET_CODE (op) == CONST_INT | |
647 | && INTVAL (op) < 0x2000 && INTVAL (op) >= -0x10); | |
648 | } | |
649 | ||
87ad11b0 | 650 | int |
651 | arith_double_operand (op, mode) | |
652 | rtx op; | |
653 | enum machine_mode mode; | |
654 | { | |
655 | return (register_operand (op, mode) | |
656 | || (GET_CODE (op) == CONST_DOUBLE | |
657 | && GET_MODE (op) == mode | |
658 | && VAL_14_BITS_P (CONST_DOUBLE_LOW (op)) | |
1fe9d310 | 659 | && ((CONST_DOUBLE_HIGH (op) >= 0) |
87ad11b0 | 660 | == ((CONST_DOUBLE_LOW (op) & 0x1000) == 0)))); |
661 | } | |
662 | ||
45981c0a | 663 | /* Return truth value of whether OP is an integer which fits the |
546a40bd | 664 | range constraining immediate operands in three-address insns, or |
665 | is an integer register. */ | |
666 | ||
667 | int | |
668 | ireg_or_int5_operand (op, mode) | |
669 | rtx op; | |
b1ca791d | 670 | enum machine_mode mode ATTRIBUTE_UNUSED; |
546a40bd | 671 | { |
672 | return ((GET_CODE (op) == CONST_INT && INT_5_BITS (op)) | |
673 | || (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32)); | |
674 | } | |
675 | ||
c7ae93c8 | 676 | /* Return nonzero if OP is an integer register, else return zero. */ |
677 | int | |
678 | ireg_operand (op, mode) | |
679 | rtx op; | |
680 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
681 | { | |
682 | return (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32); | |
683 | } | |
684 | ||
45981c0a | 685 | /* Return truth value of whether OP is an integer which fits the |
87ad11b0 | 686 | range constraining immediate operands in three-address insns. */ |
687 | ||
688 | int | |
689 | int5_operand (op, mode) | |
690 | rtx op; | |
b1ca791d | 691 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 692 | { |
693 | return (GET_CODE (op) == CONST_INT && INT_5_BITS (op)); | |
694 | } | |
695 | ||
696 | int | |
697 | uint5_operand (op, mode) | |
698 | rtx op; | |
b1ca791d | 699 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 700 | { |
701 | return (GET_CODE (op) == CONST_INT && INT_U5_BITS (op)); | |
702 | } | |
703 | ||
87ad11b0 | 704 | int |
705 | int11_operand (op, mode) | |
706 | rtx op; | |
b1ca791d | 707 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 708 | { |
6d36483b | 709 | return (GET_CODE (op) == CONST_INT && INT_11_BITS (op)); |
710 | } | |
711 | ||
712 | int | |
713 | uint32_operand (op, mode) | |
714 | rtx op; | |
b1ca791d | 715 | enum machine_mode mode ATTRIBUTE_UNUSED; |
6d36483b | 716 | { |
717 | #if HOST_BITS_PER_WIDE_INT > 32 | |
718 | /* All allowed constants will fit a CONST_INT. */ | |
719 | return (GET_CODE (op) == CONST_INT | |
b7d86581 | 720 | && (INTVAL (op) >= 0 && INTVAL (op) < (HOST_WIDE_INT) 1 << 32)); |
6d36483b | 721 | #else |
722 | return (GET_CODE (op) == CONST_INT | |
723 | || (GET_CODE (op) == CONST_DOUBLE | |
724 | && CONST_DOUBLE_HIGH (op) == 0)); | |
725 | #endif | |
87ad11b0 | 726 | } |
727 | ||
728 | int | |
729 | arith5_operand (op, mode) | |
730 | rtx op; | |
731 | enum machine_mode mode; | |
732 | { | |
733 | return register_operand (op, mode) || int5_operand (op, mode); | |
734 | } | |
735 | ||
ea52c577 | 736 | /* True iff zdepi can be used to generate this CONST_INT. |
737 | zdepi first sign extends a 5 bit signed number to a given field | |
738 | length, then places this field anywhere in a zero. */ | |
e057641f | 739 | int |
42faba01 | 740 | zdepi_cint_p (x) |
6d36483b | 741 | unsigned HOST_WIDE_INT x; |
fad0b60f | 742 | { |
3745c59b | 743 | unsigned HOST_WIDE_INT lsb_mask, t; |
fad0b60f | 744 | |
745 | /* This might not be obvious, but it's at least fast. | |
01cc3b75 | 746 | This function is critical; we don't have the time loops would take. */ |
42faba01 | 747 | lsb_mask = x & -x; |
748 | t = ((x >> 4) + lsb_mask) & ~(lsb_mask - 1); | |
749 | /* Return true iff t is a power of two. */ | |
fad0b60f | 750 | return ((t & (t - 1)) == 0); |
751 | } | |
752 | ||
6d36483b | 753 | /* True iff depi or extru can be used to compute (reg & mask). |
754 | Accept bit pattern like these: | |
755 | 0....01....1 | |
756 | 1....10....0 | |
757 | 1..10..01..1 */ | |
e057641f | 758 | int |
42faba01 | 759 | and_mask_p (mask) |
6d36483b | 760 | unsigned HOST_WIDE_INT mask; |
e057641f | 761 | { |
762 | mask = ~mask; | |
763 | mask += mask & -mask; | |
764 | return (mask & (mask - 1)) == 0; | |
765 | } | |
766 | ||
767 | /* True iff depi or extru can be used to compute (reg & OP). */ | |
768 | int | |
769 | and_operand (op, mode) | |
770 | rtx op; | |
771 | enum machine_mode mode; | |
772 | { | |
773 | return (register_operand (op, mode) | |
42faba01 | 774 | || (GET_CODE (op) == CONST_INT && and_mask_p (INTVAL (op)))); |
e057641f | 775 | } |
776 | ||
777 | /* True iff depi can be used to compute (reg | MASK). */ | |
778 | int | |
779 | ior_mask_p (mask) | |
6d36483b | 780 | unsigned HOST_WIDE_INT mask; |
e057641f | 781 | { |
782 | mask += mask & -mask; | |
783 | return (mask & (mask - 1)) == 0; | |
784 | } | |
785 | ||
786 | /* True iff depi can be used to compute (reg | OP). */ | |
787 | int | |
788 | ior_operand (op, mode) | |
789 | rtx op; | |
b1ca791d | 790 | enum machine_mode mode ATTRIBUTE_UNUSED; |
e057641f | 791 | { |
b744c8cb | 792 | return (GET_CODE (op) == CONST_INT && ior_mask_p (INTVAL (op))); |
e057641f | 793 | } |
794 | ||
e5965947 | 795 | int |
796 | lhs_lshift_operand (op, mode) | |
797 | rtx op; | |
798 | enum machine_mode mode; | |
799 | { | |
800 | return register_operand (op, mode) || lhs_lshift_cint_operand (op, mode); | |
801 | } | |
802 | ||
803 | /* True iff OP is a CONST_INT of the forms 0...0xxxx or 0...01...1xxxx. | |
804 | Such values can be the left hand side x in (x << r), using the zvdepi | |
805 | instruction. */ | |
806 | int | |
807 | lhs_lshift_cint_operand (op, mode) | |
808 | rtx op; | |
b1ca791d | 809 | enum machine_mode mode ATTRIBUTE_UNUSED; |
e5965947 | 810 | { |
3745c59b | 811 | unsigned HOST_WIDE_INT x; |
e5965947 | 812 | if (GET_CODE (op) != CONST_INT) |
813 | return 0; | |
814 | x = INTVAL (op) >> 4; | |
815 | return (x & (x + 1)) == 0; | |
816 | } | |
817 | ||
9c6d4825 | 818 | int |
819 | arith32_operand (op, mode) | |
820 | rtx op; | |
821 | enum machine_mode mode; | |
822 | { | |
823 | return register_operand (op, mode) || GET_CODE (op) == CONST_INT; | |
824 | } | |
ead9285f | 825 | |
826 | int | |
827 | pc_or_label_operand (op, mode) | |
828 | rtx op; | |
b1ca791d | 829 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ead9285f | 830 | { |
831 | return (GET_CODE (op) == PC || GET_CODE (op) == LABEL_REF); | |
832 | } | |
87ad11b0 | 833 | \f |
834 | /* Legitimize PIC addresses. If the address is already | |
835 | position-independent, we return ORIG. Newly generated | |
836 | position-independent addresses go to REG. If we need more | |
837 | than one register, we lose. */ | |
838 | ||
839 | rtx | |
840 | legitimize_pic_address (orig, mode, reg) | |
841 | rtx orig, reg; | |
611a88e1 | 842 | enum machine_mode mode; |
87ad11b0 | 843 | { |
844 | rtx pic_ref = orig; | |
845 | ||
b090827b | 846 | /* Labels need special handling. */ |
611a88e1 | 847 | if (pic_label_operand (orig, mode)) |
b4a7bf10 | 848 | { |
2536cc16 | 849 | /* We do not want to go through the movXX expanders here since that |
850 | would create recursion. | |
851 | ||
852 | Nor do we really want to call a generator for a named pattern | |
853 | since that requires multiple patterns if we want to support | |
854 | multiple word sizes. | |
855 | ||
856 | So instead we just emit the raw set, which avoids the movXX | |
857 | expanders completely. */ | |
858 | emit_insn (gen_rtx_SET (VOIDmode, reg, orig)); | |
b4a7bf10 | 859 | current_function_uses_pic_offset_table = 1; |
860 | return reg; | |
861 | } | |
87ad11b0 | 862 | if (GET_CODE (orig) == SYMBOL_REF) |
863 | { | |
864 | if (reg == 0) | |
865 | abort (); | |
866 | ||
fc44315f | 867 | emit_move_insn (reg, |
868 | gen_rtx_PLUS (word_mode, pic_offset_table_rtx, | |
869 | gen_rtx_HIGH (word_mode, orig))); | |
870 | pic_ref | |
871 | = gen_rtx_MEM (Pmode, | |
872 | gen_rtx_LO_SUM (Pmode, reg, | |
873 | gen_rtx_UNSPEC (Pmode, | |
874 | gen_rtvec (1, orig), | |
875 | 0))); | |
7014838c | 876 | |
87ad11b0 | 877 | current_function_uses_pic_offset_table = 1; |
878 | RTX_UNCHANGING_P (pic_ref) = 1; | |
879 | emit_move_insn (reg, pic_ref); | |
880 | return reg; | |
881 | } | |
882 | else if (GET_CODE (orig) == CONST) | |
883 | { | |
57ed30e5 | 884 | rtx base; |
87ad11b0 | 885 | |
886 | if (GET_CODE (XEXP (orig, 0)) == PLUS | |
887 | && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx) | |
888 | return orig; | |
889 | ||
890 | if (reg == 0) | |
891 | abort (); | |
892 | ||
893 | if (GET_CODE (XEXP (orig, 0)) == PLUS) | |
894 | { | |
895 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg); | |
896 | orig = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode, | |
897 | base == reg ? 0 : reg); | |
898 | } | |
899 | else abort (); | |
900 | if (GET_CODE (orig) == CONST_INT) | |
901 | { | |
42faba01 | 902 | if (INT_14_BITS (orig)) |
b244d4c7 | 903 | return plus_constant (base, INTVAL (orig)); |
87ad11b0 | 904 | orig = force_reg (Pmode, orig); |
905 | } | |
ad851752 | 906 | pic_ref = gen_rtx_PLUS (Pmode, base, orig); |
87ad11b0 | 907 | /* Likewise, should we set special REG_NOTEs here? */ |
908 | } | |
909 | return pic_ref; | |
910 | } | |
911 | ||
347b5848 | 912 | /* Try machine-dependent ways of modifying an illegitimate address |
913 | to be legitimate. If we find one, return the new, valid address. | |
914 | This macro is used in only one place: `memory_address' in explow.c. | |
915 | ||
916 | OLDX is the address as it was before break_out_memory_refs was called. | |
917 | In some cases it is useful to look at this to decide what needs to be done. | |
918 | ||
919 | MODE and WIN are passed so that this macro can use | |
920 | GO_IF_LEGITIMATE_ADDRESS. | |
921 | ||
922 | It is always safe for this macro to do nothing. It exists to recognize | |
6d36483b | 923 | opportunities to optimize the output. |
347b5848 | 924 | |
925 | For the PA, transform: | |
926 | ||
927 | memory(X + <large int>) | |
928 | ||
929 | into: | |
930 | ||
931 | if (<large int> & mask) >= 16 | |
932 | Y = (<large int> & ~mask) + mask + 1 Round up. | |
933 | else | |
934 | Y = (<large int> & ~mask) Round down. | |
935 | Z = X + Y | |
936 | memory (Z + (<large int> - Y)); | |
937 | ||
6d36483b | 938 | This is for CSE to find several similar references, and only use one Z. |
347b5848 | 939 | |
940 | X can either be a SYMBOL_REF or REG, but because combine can not | |
941 | perform a 4->2 combination we do nothing for SYMBOL_REF + D where | |
942 | D will not fit in 14 bits. | |
943 | ||
944 | MODE_FLOAT references allow displacements which fit in 5 bits, so use | |
6d36483b | 945 | 0x1f as the mask. |
347b5848 | 946 | |
947 | MODE_INT references allow displacements which fit in 14 bits, so use | |
6d36483b | 948 | 0x3fff as the mask. |
347b5848 | 949 | |
950 | This relies on the fact that most mode MODE_FLOAT references will use FP | |
951 | registers and most mode MODE_INT references will use integer registers. | |
952 | (In the rare case of an FP register used in an integer MODE, we depend | |
953 | on secondary reloads to clean things up.) | |
954 | ||
955 | ||
956 | It is also beneficial to handle (plus (mult (X) (Y)) (Z)) in a special | |
957 | manner if Y is 2, 4, or 8. (allows more shadd insns and shifted indexed | |
01cc3b75 | 958 | addressing modes to be used). |
347b5848 | 959 | |
960 | Put X and Z into registers. Then put the entire expression into | |
961 | a register. */ | |
962 | ||
963 | rtx | |
964 | hppa_legitimize_address (x, oldx, mode) | |
b1ca791d | 965 | rtx x, oldx ATTRIBUTE_UNUSED; |
347b5848 | 966 | enum machine_mode mode; |
967 | { | |
347b5848 | 968 | rtx orig = x; |
969 | ||
b4a7bf10 | 970 | if (flag_pic) |
971 | return legitimize_pic_address (x, mode, gen_reg_rtx (Pmode)); | |
972 | ||
6dc3b0d9 | 973 | /* Strip off CONST. */ |
347b5848 | 974 | if (GET_CODE (x) == CONST) |
975 | x = XEXP (x, 0); | |
976 | ||
42819d4e | 977 | /* Special case. Get the SYMBOL_REF into a register and use indexing. |
978 | That should always be safe. */ | |
979 | if (GET_CODE (x) == PLUS | |
980 | && GET_CODE (XEXP (x, 0)) == REG | |
981 | && GET_CODE (XEXP (x, 1)) == SYMBOL_REF) | |
982 | { | |
440c23df | 983 | rtx reg = force_reg (Pmode, XEXP (x, 1)); |
984 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, reg, XEXP (x, 0))); | |
42819d4e | 985 | } |
986 | ||
166bf021 | 987 | /* Note we must reject symbols which represent function addresses |
988 | since the assembler/linker can't handle arithmetic on plabels. */ | |
347b5848 | 989 | if (GET_CODE (x) == PLUS |
990 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
166bf021 | 991 | && ((GET_CODE (XEXP (x, 0)) == SYMBOL_REF |
992 | && !FUNCTION_NAME_P (XSTR (XEXP (x, 0), 0))) | |
347b5848 | 993 | || GET_CODE (XEXP (x, 0)) == REG)) |
994 | { | |
995 | rtx int_part, ptr_reg; | |
996 | int newoffset; | |
997 | int offset = INTVAL (XEXP (x, 1)); | |
1b6f11e2 | 998 | int mask; |
999 | ||
1000 | mask = (GET_MODE_CLASS (mode) == MODE_FLOAT | |
1001 | ? (TARGET_PA_20 ? 0x3fff : 0x1f) : 0x3fff); | |
347b5848 | 1002 | |
6d36483b | 1003 | /* Choose which way to round the offset. Round up if we |
347b5848 | 1004 | are >= halfway to the next boundary. */ |
1005 | if ((offset & mask) >= ((mask + 1) / 2)) | |
1006 | newoffset = (offset & ~ mask) + mask + 1; | |
1007 | else | |
1008 | newoffset = (offset & ~ mask); | |
1009 | ||
1010 | /* If the newoffset will not fit in 14 bits (ldo), then | |
1011 | handling this would take 4 or 5 instructions (2 to load | |
1012 | the SYMBOL_REF + 1 or 2 to load the newoffset + 1 to | |
1013 | add the new offset and the SYMBOL_REF.) Combine can | |
1014 | not handle 4->2 or 5->2 combinations, so do not create | |
1015 | them. */ | |
1016 | if (! VAL_14_BITS_P (newoffset) | |
1017 | && GET_CODE (XEXP (x, 0)) == SYMBOL_REF) | |
1018 | { | |
7014838c | 1019 | rtx const_part = plus_constant (XEXP (x, 0), newoffset); |
347b5848 | 1020 | rtx tmp_reg |
339613b4 | 1021 | = force_reg (Pmode, |
ad851752 | 1022 | gen_rtx_HIGH (Pmode, const_part)); |
347b5848 | 1023 | ptr_reg |
339613b4 | 1024 | = force_reg (Pmode, |
7014838c | 1025 | gen_rtx_LO_SUM (Pmode, |
1026 | tmp_reg, const_part)); | |
347b5848 | 1027 | } |
1028 | else | |
1029 | { | |
1030 | if (! VAL_14_BITS_P (newoffset)) | |
339613b4 | 1031 | int_part = force_reg (Pmode, GEN_INT (newoffset)); |
347b5848 | 1032 | else |
1033 | int_part = GEN_INT (newoffset); | |
1034 | ||
339613b4 | 1035 | ptr_reg = force_reg (Pmode, |
ad851752 | 1036 | gen_rtx_PLUS (Pmode, |
1037 | force_reg (Pmode, XEXP (x, 0)), | |
1038 | int_part)); | |
347b5848 | 1039 | } |
1040 | return plus_constant (ptr_reg, offset - newoffset); | |
1041 | } | |
45f1285a | 1042 | |
5115683e | 1043 | /* Handle (plus (mult (a) (shadd_constant)) (b)). */ |
45f1285a | 1044 | |
347b5848 | 1045 | if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT |
1046 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
45f1285a | 1047 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1))) |
1048 | && (GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) == 'o' | |
1049 | || GET_CODE (XEXP (x, 1)) == SUBREG) | |
1050 | && GET_CODE (XEXP (x, 1)) != CONST) | |
347b5848 | 1051 | { |
1052 | int val = INTVAL (XEXP (XEXP (x, 0), 1)); | |
1053 | rtx reg1, reg2; | |
5115683e | 1054 | |
1055 | reg1 = XEXP (x, 1); | |
1056 | if (GET_CODE (reg1) != REG) | |
1057 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1058 | ||
1059 | reg2 = XEXP (XEXP (x, 0), 0); | |
1060 | if (GET_CODE (reg2) != REG) | |
1061 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1062 | ||
ad851752 | 1063 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, |
7014838c | 1064 | gen_rtx_MULT (Pmode, |
1065 | reg2, | |
1066 | GEN_INT (val)), | |
ad851752 | 1067 | reg1)); |
347b5848 | 1068 | } |
45f1285a | 1069 | |
00a87639 | 1070 | /* Similarly for (plus (plus (mult (a) (shadd_constant)) (b)) (c)). |
1071 | ||
1072 | Only do so for floating point modes since this is more speculative | |
1073 | and we lose if it's an integer store. */ | |
5115683e | 1074 | if (GET_CODE (x) == PLUS |
00a87639 | 1075 | && GET_CODE (XEXP (x, 0)) == PLUS |
1076 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
1077 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT | |
5115683e | 1078 | && shadd_constant_p (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))) |
1079 | && (mode == SFmode || mode == DFmode)) | |
00a87639 | 1080 | { |
5115683e | 1081 | |
1082 | /* First, try and figure out what to use as a base register. */ | |
1083 | rtx reg1, reg2, base, idx, orig_base; | |
1084 | ||
1085 | reg1 = XEXP (XEXP (x, 0), 1); | |
1086 | reg2 = XEXP (x, 1); | |
1087 | base = NULL_RTX; | |
1088 | idx = NULL_RTX; | |
1089 | ||
1090 | /* Make sure they're both regs. If one was a SYMBOL_REF [+ const], | |
e61a0a7f | 1091 | then emit_move_sequence will turn on REG_POINTER so we'll know |
1092 | it's a base register below. */ | |
5115683e | 1093 | if (GET_CODE (reg1) != REG) |
1094 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1095 | ||
1096 | if (GET_CODE (reg2) != REG) | |
1097 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1098 | ||
1099 | /* Figure out what the base and index are. */ | |
9840d99d | 1100 | |
5115683e | 1101 | if (GET_CODE (reg1) == REG |
e61a0a7f | 1102 | && REG_POINTER (reg1)) |
5115683e | 1103 | { |
1104 | base = reg1; | |
1105 | orig_base = XEXP (XEXP (x, 0), 1); | |
ad851752 | 1106 | idx = gen_rtx_PLUS (Pmode, |
1107 | gen_rtx_MULT (Pmode, | |
1108 | XEXP (XEXP (XEXP (x, 0), 0), 0), | |
1109 | XEXP (XEXP (XEXP (x, 0), 0), 1)), | |
1110 | XEXP (x, 1)); | |
5115683e | 1111 | } |
1112 | else if (GET_CODE (reg2) == REG | |
e61a0a7f | 1113 | && REG_POINTER (reg2)) |
5115683e | 1114 | { |
1115 | base = reg2; | |
1116 | orig_base = XEXP (x, 1); | |
1117 | idx = XEXP (x, 0); | |
1118 | } | |
1119 | ||
1120 | if (base == 0) | |
21f3ee9c | 1121 | return orig; |
5115683e | 1122 | |
1123 | /* If the index adds a large constant, try to scale the | |
1124 | constant so that it can be loaded with only one insn. */ | |
1125 | if (GET_CODE (XEXP (idx, 1)) == CONST_INT | |
1126 | && VAL_14_BITS_P (INTVAL (XEXP (idx, 1)) | |
1127 | / INTVAL (XEXP (XEXP (idx, 0), 1))) | |
1128 | && INTVAL (XEXP (idx, 1)) % INTVAL (XEXP (XEXP (idx, 0), 1)) == 0) | |
1129 | { | |
1130 | /* Divide the CONST_INT by the scale factor, then add it to A. */ | |
1131 | int val = INTVAL (XEXP (idx, 1)); | |
1132 | ||
1133 | val /= INTVAL (XEXP (XEXP (idx, 0), 1)); | |
1134 | reg1 = XEXP (XEXP (idx, 0), 0); | |
1135 | if (GET_CODE (reg1) != REG) | |
1136 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1137 | ||
ad851752 | 1138 | reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, reg1, GEN_INT (val))); |
5115683e | 1139 | |
1140 | /* We can now generate a simple scaled indexed address. */ | |
7014838c | 1141 | return |
1142 | force_reg | |
1143 | (Pmode, gen_rtx_PLUS (Pmode, | |
1144 | gen_rtx_MULT (Pmode, reg1, | |
1145 | XEXP (XEXP (idx, 0), 1)), | |
1146 | base)); | |
5115683e | 1147 | } |
1148 | ||
1149 | /* If B + C is still a valid base register, then add them. */ | |
1150 | if (GET_CODE (XEXP (idx, 1)) == CONST_INT | |
1151 | && INTVAL (XEXP (idx, 1)) <= 4096 | |
1152 | && INTVAL (XEXP (idx, 1)) >= -4096) | |
1153 | { | |
1154 | int val = INTVAL (XEXP (XEXP (idx, 0), 1)); | |
1155 | rtx reg1, reg2; | |
1156 | ||
ad851752 | 1157 | reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, base, XEXP (idx, 1))); |
5115683e | 1158 | |
1159 | reg2 = XEXP (XEXP (idx, 0), 0); | |
1160 | if (GET_CODE (reg2) != CONST_INT) | |
1161 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1162 | ||
ad851752 | 1163 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, |
7014838c | 1164 | gen_rtx_MULT (Pmode, |
1165 | reg2, | |
ad851752 | 1166 | GEN_INT (val)), |
1167 | reg1)); | |
5115683e | 1168 | } |
1169 | ||
1170 | /* Get the index into a register, then add the base + index and | |
1171 | return a register holding the result. */ | |
1172 | ||
1173 | /* First get A into a register. */ | |
1174 | reg1 = XEXP (XEXP (idx, 0), 0); | |
1175 | if (GET_CODE (reg1) != REG) | |
1176 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1177 | ||
1178 | /* And get B into a register. */ | |
1179 | reg2 = XEXP (idx, 1); | |
1180 | if (GET_CODE (reg2) != REG) | |
1181 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1182 | ||
ad851752 | 1183 | reg1 = force_reg (Pmode, |
1184 | gen_rtx_PLUS (Pmode, | |
1185 | gen_rtx_MULT (Pmode, reg1, | |
1186 | XEXP (XEXP (idx, 0), 1)), | |
1187 | reg2)); | |
5115683e | 1188 | |
1189 | /* Add the result to our base register and return. */ | |
ad851752 | 1190 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, base, reg1)); |
9840d99d | 1191 | |
00a87639 | 1192 | } |
1193 | ||
6d36483b | 1194 | /* Uh-oh. We might have an address for x[n-100000]. This needs |
fb5390c1 | 1195 | special handling to avoid creating an indexed memory address |
1196 | with x-100000 as the base. | |
9840d99d | 1197 | |
fb5390c1 | 1198 | If the constant part is small enough, then it's still safe because |
1199 | there is a guard page at the beginning and end of the data segment. | |
1200 | ||
1201 | Scaled references are common enough that we want to try and rearrange the | |
1202 | terms so that we can use indexing for these addresses too. Only | |
00a87639 | 1203 | do the optimization for floatint point modes. */ |
45f1285a | 1204 | |
fb5390c1 | 1205 | if (GET_CODE (x) == PLUS |
1206 | && symbolic_expression_p (XEXP (x, 1))) | |
45f1285a | 1207 | { |
1208 | /* Ugly. We modify things here so that the address offset specified | |
1209 | by the index expression is computed first, then added to x to form | |
fb5390c1 | 1210 | the entire address. */ |
45f1285a | 1211 | |
00a87639 | 1212 | rtx regx1, regx2, regy1, regy2, y; |
45f1285a | 1213 | |
1214 | /* Strip off any CONST. */ | |
1215 | y = XEXP (x, 1); | |
1216 | if (GET_CODE (y) == CONST) | |
1217 | y = XEXP (y, 0); | |
1218 | ||
7ee96d6e | 1219 | if (GET_CODE (y) == PLUS || GET_CODE (y) == MINUS) |
1220 | { | |
00a87639 | 1221 | /* See if this looks like |
1222 | (plus (mult (reg) (shadd_const)) | |
1223 | (const (plus (symbol_ref) (const_int)))) | |
1224 | ||
5115683e | 1225 | Where const_int is small. In that case the const |
9840d99d | 1226 | expression is a valid pointer for indexing. |
5115683e | 1227 | |
1228 | If const_int is big, but can be divided evenly by shadd_const | |
1229 | and added to (reg). This allows more scaled indexed addresses. */ | |
1230 | if (GET_CODE (XEXP (y, 0)) == SYMBOL_REF | |
1231 | && GET_CODE (XEXP (x, 0)) == MULT | |
00a87639 | 1232 | && GET_CODE (XEXP (y, 1)) == CONST_INT |
5115683e | 1233 | && INTVAL (XEXP (y, 1)) >= -4096 |
1234 | && INTVAL (XEXP (y, 1)) <= 4095 | |
1235 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
1236 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))) | |
1237 | { | |
1238 | int val = INTVAL (XEXP (XEXP (x, 0), 1)); | |
1239 | rtx reg1, reg2; | |
1240 | ||
1241 | reg1 = XEXP (x, 1); | |
1242 | if (GET_CODE (reg1) != REG) | |
1243 | reg1 = force_reg (Pmode, force_operand (reg1, 0)); | |
1244 | ||
1245 | reg2 = XEXP (XEXP (x, 0), 0); | |
1246 | if (GET_CODE (reg2) != REG) | |
1247 | reg2 = force_reg (Pmode, force_operand (reg2, 0)); | |
1248 | ||
ad851752 | 1249 | return force_reg (Pmode, |
1250 | gen_rtx_PLUS (Pmode, | |
7014838c | 1251 | gen_rtx_MULT (Pmode, |
1252 | reg2, | |
ad851752 | 1253 | GEN_INT (val)), |
7014838c | 1254 | reg1)); |
5115683e | 1255 | } |
1256 | else if ((mode == DFmode || mode == SFmode) | |
1257 | && GET_CODE (XEXP (y, 0)) == SYMBOL_REF | |
1258 | && GET_CODE (XEXP (x, 0)) == MULT | |
1259 | && GET_CODE (XEXP (y, 1)) == CONST_INT | |
1260 | && INTVAL (XEXP (y, 1)) % INTVAL (XEXP (XEXP (x, 0), 1)) == 0 | |
1261 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
1262 | && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))) | |
00a87639 | 1263 | { |
1264 | regx1 | |
1265 | = force_reg (Pmode, GEN_INT (INTVAL (XEXP (y, 1)) | |
1266 | / INTVAL (XEXP (XEXP (x, 0), 1)))); | |
1267 | regx2 = XEXP (XEXP (x, 0), 0); | |
1268 | if (GET_CODE (regx2) != REG) | |
1269 | regx2 = force_reg (Pmode, force_operand (regx2, 0)); | |
ad851752 | 1270 | regx2 = force_reg (Pmode, gen_rtx_fmt_ee (GET_CODE (y), Pmode, |
1271 | regx2, regx1)); | |
7014838c | 1272 | return |
1273 | force_reg (Pmode, | |
1274 | gen_rtx_PLUS (Pmode, | |
1275 | gen_rtx_MULT (Pmode, regx2, | |
1276 | XEXP (XEXP (x, 0), 1)), | |
1277 | force_reg (Pmode, XEXP (y, 0)))); | |
00a87639 | 1278 | } |
fb5390c1 | 1279 | else if (GET_CODE (XEXP (y, 1)) == CONST_INT |
1280 | && INTVAL (XEXP (y, 1)) >= -4096 | |
1281 | && INTVAL (XEXP (y, 1)) <= 4095) | |
1282 | { | |
1283 | /* This is safe because of the guard page at the | |
1284 | beginning and end of the data space. Just | |
1285 | return the original address. */ | |
1286 | return orig; | |
1287 | } | |
00a87639 | 1288 | else |
1289 | { | |
1290 | /* Doesn't look like one we can optimize. */ | |
1291 | regx1 = force_reg (Pmode, force_operand (XEXP (x, 0), 0)); | |
1292 | regy1 = force_reg (Pmode, force_operand (XEXP (y, 0), 0)); | |
1293 | regy2 = force_reg (Pmode, force_operand (XEXP (y, 1), 0)); | |
1294 | regx1 = force_reg (Pmode, | |
ad851752 | 1295 | gen_rtx_fmt_ee (GET_CODE (y), Pmode, |
1296 | regx1, regy2)); | |
1297 | return force_reg (Pmode, gen_rtx_PLUS (Pmode, regx1, regy1)); | |
00a87639 | 1298 | } |
7ee96d6e | 1299 | } |
45f1285a | 1300 | } |
1301 | ||
347b5848 | 1302 | return orig; |
1303 | } | |
1304 | ||
87ad11b0 | 1305 | /* For the HPPA, REG and REG+CONST is cost 0 |
1306 | and addresses involving symbolic constants are cost 2. | |
1307 | ||
1308 | PIC addresses are very expensive. | |
1309 | ||
1310 | It is no coincidence that this has the same structure | |
1311 | as GO_IF_LEGITIMATE_ADDRESS. */ | |
ec0457a8 | 1312 | |
1313 | static int | |
87ad11b0 | 1314 | hppa_address_cost (X) |
1315 | rtx X; | |
1316 | { | |
ec0457a8 | 1317 | switch (GET_CODE (X)) |
1318 | { | |
1319 | case REG: | |
1320 | case PLUS: | |
1321 | case LO_SUM: | |
87ad11b0 | 1322 | return 1; |
ec0457a8 | 1323 | case HIGH: |
1324 | return 2; | |
1325 | default: | |
1326 | return 4; | |
1327 | } | |
87ad11b0 | 1328 | } |
1329 | ||
fab7adbf | 1330 | /* Compute a (partial) cost for rtx X. Return true if the complete |
1331 | cost has been computed, and false if subexpressions should be | |
1332 | scanned. In either case, *TOTAL contains the cost result. */ | |
1333 | ||
1334 | static bool | |
1335 | hppa_rtx_costs (x, code, outer_code, total) | |
1336 | rtx x; | |
1337 | int code, outer_code; | |
1338 | int *total; | |
1339 | { | |
1340 | switch (code) | |
1341 | { | |
1342 | case CONST_INT: | |
1343 | if (INTVAL (x) == 0) | |
1344 | *total = 0; | |
1345 | else if (INT_14_BITS (x)) | |
1346 | *total = 1; | |
1347 | else | |
1348 | *total = 2; | |
1349 | return true; | |
1350 | ||
1351 | case HIGH: | |
1352 | *total = 2; | |
1353 | return true; | |
1354 | ||
1355 | case CONST: | |
1356 | case LABEL_REF: | |
1357 | case SYMBOL_REF: | |
1358 | *total = 4; | |
1359 | return true; | |
1360 | ||
1361 | case CONST_DOUBLE: | |
1362 | if ((x == CONST0_RTX (DFmode) || x == CONST0_RTX (SFmode)) | |
1363 | && outer_code != SET) | |
1364 | *total = 0; | |
1365 | else | |
1366 | *total = 8; | |
1367 | return true; | |
1368 | ||
1369 | case MULT: | |
1370 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1371 | *total = COSTS_N_INSNS (3); | |
1372 | else if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT) | |
1373 | *total = COSTS_N_INSNS (8); | |
1374 | else | |
1375 | *total = COSTS_N_INSNS (20); | |
1376 | return true; | |
1377 | ||
1378 | case DIV: | |
1379 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1380 | { | |
1381 | *total = COSTS_N_INSNS (14); | |
1382 | return true; | |
1383 | } | |
1384 | /* FALLTHRU */ | |
1385 | ||
1386 | case UDIV: | |
1387 | case MOD: | |
1388 | case UMOD: | |
1389 | *total = COSTS_N_INSNS (60); | |
1390 | return true; | |
1391 | ||
1392 | case PLUS: /* this includes shNadd insns */ | |
1393 | case MINUS: | |
1394 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
1395 | *total = COSTS_N_INSNS (3); | |
1396 | else | |
1397 | *total = COSTS_N_INSNS (1); | |
1398 | return true; | |
1399 | ||
1400 | case ASHIFT: | |
1401 | case ASHIFTRT: | |
1402 | case LSHIFTRT: | |
1403 | *total = COSTS_N_INSNS (1); | |
1404 | return true; | |
1405 | ||
1406 | default: | |
1407 | return false; | |
1408 | } | |
1409 | } | |
1410 | ||
9840d99d | 1411 | /* Ensure mode of ORIG, a REG rtx, is MODE. Returns either ORIG or a |
1412 | new rtx with the correct mode. */ | |
1413 | static inline rtx | |
1414 | force_mode (mode, orig) | |
1415 | enum machine_mode mode; | |
1416 | rtx orig; | |
1417 | { | |
1418 | if (mode == GET_MODE (orig)) | |
1419 | return orig; | |
1420 | ||
1421 | if (REGNO (orig) >= FIRST_PSEUDO_REGISTER) | |
1422 | abort (); | |
1423 | ||
1424 | return gen_rtx_REG (mode, REGNO (orig)); | |
1425 | } | |
1426 | ||
87ad11b0 | 1427 | /* Emit insns to move operands[1] into operands[0]. |
1428 | ||
1429 | Return 1 if we have written out everything that needs to be done to | |
1430 | do the move. Otherwise, return 0 and the caller will emit the move | |
9840d99d | 1431 | normally. |
f756078b | 1432 | |
1433 | Note SCRATCH_REG may not be in the proper mode depending on how it | |
1434 | will be used. This routine is resposible for creating a new copy | |
1435 | of SCRATCH_REG in the proper mode. */ | |
87ad11b0 | 1436 | |
1437 | int | |
d6f01525 | 1438 | emit_move_sequence (operands, mode, scratch_reg) |
87ad11b0 | 1439 | rtx *operands; |
1440 | enum machine_mode mode; | |
d6f01525 | 1441 | rtx scratch_reg; |
87ad11b0 | 1442 | { |
1443 | register rtx operand0 = operands[0]; | |
1444 | register rtx operand1 = operands[1]; | |
4a155b0f | 1445 | register rtx tem; |
87ad11b0 | 1446 | |
2d4dc8d2 | 1447 | if (scratch_reg |
1448 | && reload_in_progress && GET_CODE (operand0) == REG | |
d1e2bb73 | 1449 | && REGNO (operand0) >= FIRST_PSEUDO_REGISTER) |
1450 | operand0 = reg_equiv_mem[REGNO (operand0)]; | |
2d4dc8d2 | 1451 | else if (scratch_reg |
1452 | && reload_in_progress && GET_CODE (operand0) == SUBREG | |
d1e2bb73 | 1453 | && GET_CODE (SUBREG_REG (operand0)) == REG |
1454 | && REGNO (SUBREG_REG (operand0)) >= FIRST_PSEUDO_REGISTER) | |
a3afad75 | 1455 | { |
701e46d0 | 1456 | /* We must not alter SUBREG_BYTE (operand0) since that would confuse |
d3c3f88d | 1457 | the code which tracks sets/uses for delete_output_reload. */ |
1458 | rtx temp = gen_rtx_SUBREG (GET_MODE (operand0), | |
1459 | reg_equiv_mem [REGNO (SUBREG_REG (operand0))], | |
701e46d0 | 1460 | SUBREG_BYTE (operand0)); |
e00b80b2 | 1461 | operand0 = alter_subreg (&temp); |
a3afad75 | 1462 | } |
d1e2bb73 | 1463 | |
2d4dc8d2 | 1464 | if (scratch_reg |
1465 | && reload_in_progress && GET_CODE (operand1) == REG | |
d1e2bb73 | 1466 | && REGNO (operand1) >= FIRST_PSEUDO_REGISTER) |
1467 | operand1 = reg_equiv_mem[REGNO (operand1)]; | |
2d4dc8d2 | 1468 | else if (scratch_reg |
1469 | && reload_in_progress && GET_CODE (operand1) == SUBREG | |
d1e2bb73 | 1470 | && GET_CODE (SUBREG_REG (operand1)) == REG |
1471 | && REGNO (SUBREG_REG (operand1)) >= FIRST_PSEUDO_REGISTER) | |
a3afad75 | 1472 | { |
701e46d0 | 1473 | /* We must not alter SUBREG_BYTE (operand0) since that would confuse |
d3c3f88d | 1474 | the code which tracks sets/uses for delete_output_reload. */ |
1475 | rtx temp = gen_rtx_SUBREG (GET_MODE (operand1), | |
1476 | reg_equiv_mem [REGNO (SUBREG_REG (operand1))], | |
701e46d0 | 1477 | SUBREG_BYTE (operand1)); |
e00b80b2 | 1478 | operand1 = alter_subreg (&temp); |
a3afad75 | 1479 | } |
d1e2bb73 | 1480 | |
2d4dc8d2 | 1481 | if (scratch_reg && reload_in_progress && GET_CODE (operand0) == MEM |
4a155b0f | 1482 | && ((tem = find_replacement (&XEXP (operand0, 0))) |
1483 | != XEXP (operand0, 0))) | |
ad851752 | 1484 | operand0 = gen_rtx_MEM (GET_MODE (operand0), tem); |
2d4dc8d2 | 1485 | if (scratch_reg && reload_in_progress && GET_CODE (operand1) == MEM |
4a155b0f | 1486 | && ((tem = find_replacement (&XEXP (operand1, 0))) |
1487 | != XEXP (operand1, 0))) | |
ad851752 | 1488 | operand1 = gen_rtx_MEM (GET_MODE (operand1), tem); |
4a155b0f | 1489 | |
e8fdbafa | 1490 | /* Handle secondary reloads for loads/stores of FP registers from |
9840d99d | 1491 | REG+D addresses where D does not fit in 5 bits, including |
42819d4e | 1492 | (subreg (mem (addr))) cases. */ |
d6f01525 | 1493 | if (fp_reg_operand (operand0, mode) |
6b1c36c2 | 1494 | && ((GET_CODE (operand1) == MEM |
1495 | && ! memory_address_p (DFmode, XEXP (operand1, 0))) | |
1496 | || ((GET_CODE (operand1) == SUBREG | |
1497 | && GET_CODE (XEXP (operand1, 0)) == MEM | |
1498 | && !memory_address_p (DFmode, XEXP (XEXP (operand1, 0), 0))))) | |
d6f01525 | 1499 | && scratch_reg) |
1500 | { | |
6b1c36c2 | 1501 | if (GET_CODE (operand1) == SUBREG) |
1502 | operand1 = XEXP (operand1, 0); | |
1503 | ||
f756078b | 1504 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1505 | it in WORD_MODE regardless of what mode it was originally given | |
1506 | to us. */ | |
9840d99d | 1507 | scratch_reg = force_mode (word_mode, scratch_reg); |
7ee39d73 | 1508 | |
1509 | /* D might not fit in 14 bits either; for such cases load D into | |
1510 | scratch reg. */ | |
440c23df | 1511 | if (!memory_address_p (Pmode, XEXP (operand1, 0))) |
7ee39d73 | 1512 | { |
1513 | emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1)); | |
ad851752 | 1514 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, 0)), |
440c23df | 1515 | Pmode, |
ad851752 | 1516 | XEXP (XEXP (operand1, 0), 0), |
1517 | scratch_reg)); | |
7ee39d73 | 1518 | } |
1519 | else | |
1520 | emit_move_insn (scratch_reg, XEXP (operand1, 0)); | |
7014838c | 1521 | emit_insn (gen_rtx_SET (VOIDmode, operand0, |
1522 | gen_rtx_MEM (mode, scratch_reg))); | |
d6f01525 | 1523 | return 1; |
1524 | } | |
1525 | else if (fp_reg_operand (operand1, mode) | |
6b1c36c2 | 1526 | && ((GET_CODE (operand0) == MEM |
1527 | && ! memory_address_p (DFmode, XEXP (operand0, 0))) | |
1528 | || ((GET_CODE (operand0) == SUBREG) | |
1529 | && GET_CODE (XEXP (operand0, 0)) == MEM | |
1530 | && !memory_address_p (DFmode, XEXP (XEXP (operand0, 0), 0)))) | |
d6f01525 | 1531 | && scratch_reg) |
1532 | { | |
6b1c36c2 | 1533 | if (GET_CODE (operand0) == SUBREG) |
1534 | operand0 = XEXP (operand0, 0); | |
1535 | ||
f756078b | 1536 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1537 | it in WORD_MODE regardless of what mode it was originally given | |
1538 | to us. */ | |
9840d99d | 1539 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1540 | |
7ee39d73 | 1541 | /* D might not fit in 14 bits either; for such cases load D into |
1542 | scratch reg. */ | |
440c23df | 1543 | if (!memory_address_p (Pmode, XEXP (operand0, 0))) |
7ee39d73 | 1544 | { |
1545 | emit_move_insn (scratch_reg, XEXP (XEXP (operand0, 0), 1)); | |
ad851752 | 1546 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand0, |
1547 | 0)), | |
440c23df | 1548 | Pmode, |
ad851752 | 1549 | XEXP (XEXP (operand0, 0), |
1550 | 0), | |
1551 | scratch_reg)); | |
7ee39d73 | 1552 | } |
1553 | else | |
1554 | emit_move_insn (scratch_reg, XEXP (operand0, 0)); | |
ad851752 | 1555 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_MEM (mode, scratch_reg), |
1556 | operand1)); | |
d6f01525 | 1557 | return 1; |
1558 | } | |
753bd06a | 1559 | /* Handle secondary reloads for loads of FP registers from constant |
1560 | expressions by forcing the constant into memory. | |
1561 | ||
6d36483b | 1562 | use scratch_reg to hold the address of the memory location. |
753bd06a | 1563 | |
9dc4cbf1 | 1564 | The proper fix is to change PREFERRED_RELOAD_CLASS to return |
87fcb603 | 1565 | NO_REGS when presented with a const_int and a register class |
753bd06a | 1566 | containing only FP registers. Doing so unfortunately creates |
1567 | more problems than it solves. Fix this for 2.5. */ | |
1568 | else if (fp_reg_operand (operand0, mode) | |
1569 | && CONSTANT_P (operand1) | |
1570 | && scratch_reg) | |
1571 | { | |
1572 | rtx xoperands[2]; | |
1573 | ||
f756078b | 1574 | /* SCRATCH_REG will hold an address and maybe the actual data. We want |
1575 | it in WORD_MODE regardless of what mode it was originally given | |
1576 | to us. */ | |
9840d99d | 1577 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1578 | |
753bd06a | 1579 | /* Force the constant into memory and put the address of the |
1580 | memory location into scratch_reg. */ | |
1581 | xoperands[0] = scratch_reg; | |
1582 | xoperands[1] = XEXP (force_const_mem (mode, operand1), 0); | |
8f258b49 | 1583 | emit_move_sequence (xoperands, Pmode, 0); |
753bd06a | 1584 | |
1585 | /* Now load the destination register. */ | |
7014838c | 1586 | emit_insn (gen_rtx_SET (mode, operand0, |
1587 | gen_rtx_MEM (mode, scratch_reg))); | |
753bd06a | 1588 | return 1; |
1589 | } | |
e8fdbafa | 1590 | /* Handle secondary reloads for SAR. These occur when trying to load |
2a170404 | 1591 | the SAR from memory, FP register, or with a constant. */ |
e8fdbafa | 1592 | else if (GET_CODE (operand0) == REG |
2a170404 | 1593 | && REGNO (operand0) < FIRST_PSEUDO_REGISTER |
e8fdbafa | 1594 | && REGNO_REG_CLASS (REGNO (operand0)) == SHIFT_REGS |
1595 | && (GET_CODE (operand1) == MEM | |
7d43e0f7 | 1596 | || GET_CODE (operand1) == CONST_INT |
e8fdbafa | 1597 | || (GET_CODE (operand1) == REG |
1598 | && FP_REG_CLASS_P (REGNO_REG_CLASS (REGNO (operand1))))) | |
1599 | && scratch_reg) | |
1600 | { | |
7eac600c | 1601 | /* D might not fit in 14 bits either; for such cases load D into |
1602 | scratch reg. */ | |
1603 | if (GET_CODE (operand1) == MEM | |
440c23df | 1604 | && !memory_address_p (Pmode, XEXP (operand1, 0))) |
7eac600c | 1605 | { |
fc4d127d | 1606 | /* We are reloading the address into the scratch register, so we |
1607 | want to make sure the scratch register is a full register. */ | |
9840d99d | 1608 | scratch_reg = force_mode (word_mode, scratch_reg); |
fc4d127d | 1609 | |
9840d99d | 1610 | emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1)); |
ad851752 | 1611 | emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, |
1612 | 0)), | |
440c23df | 1613 | Pmode, |
ad851752 | 1614 | XEXP (XEXP (operand1, 0), |
1615 | 0), | |
1616 | scratch_reg)); | |
fc4d127d | 1617 | |
1618 | /* Now we are going to load the scratch register from memory, | |
1619 | we want to load it in the same width as the original MEM, | |
1620 | which must be the same as the width of the ultimate destination, | |
1621 | OPERAND0. */ | |
9840d99d | 1622 | scratch_reg = force_mode (GET_MODE (operand0), scratch_reg); |
1623 | ||
fc4d127d | 1624 | emit_move_insn (scratch_reg, gen_rtx_MEM (GET_MODE (operand0), |
ad851752 | 1625 | scratch_reg)); |
7eac600c | 1626 | } |
1627 | else | |
fc4d127d | 1628 | { |
1629 | /* We want to load the scratch register using the same mode as | |
1630 | the ultimate destination. */ | |
9840d99d | 1631 | scratch_reg = force_mode (GET_MODE (operand0), scratch_reg); |
1632 | ||
fc4d127d | 1633 | emit_move_insn (scratch_reg, operand1); |
1634 | } | |
1635 | ||
1636 | /* And emit the insn to set the ultimate destination. We know that | |
1637 | the scratch register has the same mode as the destination at this | |
1638 | point. */ | |
e8fdbafa | 1639 | emit_move_insn (operand0, scratch_reg); |
1640 | return 1; | |
1641 | } | |
d6f01525 | 1642 | /* Handle most common case: storing into a register. */ |
1643 | else if (register_operand (operand0, mode)) | |
87ad11b0 | 1644 | { |
1645 | if (register_operand (operand1, mode) | |
93f6e1c7 | 1646 | || (GET_CODE (operand1) == CONST_INT |
1647 | && cint_ok_for_move (INTVAL (operand1))) | |
891b55b4 | 1648 | || (operand1 == CONST0_RTX (mode)) |
87ad11b0 | 1649 | || (GET_CODE (operand1) == HIGH |
df0651dc | 1650 | && !symbolic_operand (XEXP (operand1, 0), VOIDmode)) |
87ad11b0 | 1651 | /* Only `general_operands' can come here, so MEM is ok. */ |
1652 | || GET_CODE (operand1) == MEM) | |
1653 | { | |
1654 | /* Run this case quickly. */ | |
ad851752 | 1655 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1)); |
87ad11b0 | 1656 | return 1; |
1657 | } | |
1658 | } | |
1659 | else if (GET_CODE (operand0) == MEM) | |
1660 | { | |
85eb4c6e | 1661 | if (mode == DFmode && operand1 == CONST0_RTX (mode) |
1662 | && !(reload_in_progress || reload_completed)) | |
1663 | { | |
1664 | rtx temp = gen_reg_rtx (DFmode); | |
1665 | ||
ad851752 | 1666 | emit_insn (gen_rtx_SET (VOIDmode, temp, operand1)); |
1667 | emit_insn (gen_rtx_SET (VOIDmode, operand0, temp)); | |
85eb4c6e | 1668 | return 1; |
1669 | } | |
891b55b4 | 1670 | if (register_operand (operand1, mode) || operand1 == CONST0_RTX (mode)) |
87ad11b0 | 1671 | { |
1672 | /* Run this case quickly. */ | |
ad851752 | 1673 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1)); |
87ad11b0 | 1674 | return 1; |
1675 | } | |
2ff4bf8d | 1676 | if (! (reload_in_progress || reload_completed)) |
87ad11b0 | 1677 | { |
1678 | operands[0] = validize_mem (operand0); | |
1679 | operands[1] = operand1 = force_reg (mode, operand1); | |
1680 | } | |
1681 | } | |
1682 | ||
37a75d53 | 1683 | /* Simplify the source if we need to. |
1684 | Note we do have to handle function labels here, even though we do | |
1685 | not consider them legitimate constants. Loop optimizations can | |
bea60f66 | 1686 | call the emit_move_xxx with one as a source. */ |
57ed30e5 | 1687 | if ((GET_CODE (operand1) != HIGH && immediate_operand (operand1, mode)) |
37a75d53 | 1688 | || function_label_operand (operand1, mode) |
2ee034bc | 1689 | || (GET_CODE (operand1) == HIGH |
63882853 | 1690 | && symbolic_operand (XEXP (operand1, 0), mode))) |
87ad11b0 | 1691 | { |
2ee034bc | 1692 | int ishighonly = 0; |
1693 | ||
1694 | if (GET_CODE (operand1) == HIGH) | |
1695 | { | |
1696 | ishighonly = 1; | |
1697 | operand1 = XEXP (operand1, 0); | |
1698 | } | |
87ad11b0 | 1699 | if (symbolic_operand (operand1, mode)) |
1700 | { | |
005a7dd0 | 1701 | /* Argh. The assembler and linker can't handle arithmetic |
81653f9b | 1702 | involving plabels. |
005a7dd0 | 1703 | |
81653f9b | 1704 | So we force the plabel into memory, load operand0 from |
1705 | the memory location, then add in the constant part. */ | |
37a75d53 | 1706 | if ((GET_CODE (operand1) == CONST |
1707 | && GET_CODE (XEXP (operand1, 0)) == PLUS | |
1708 | && function_label_operand (XEXP (XEXP (operand1, 0), 0), Pmode)) | |
1709 | || function_label_operand (operand1, mode)) | |
005a7dd0 | 1710 | { |
b3d569a0 | 1711 | rtx temp, const_part; |
81653f9b | 1712 | |
1713 | /* Figure out what (if any) scratch register to use. */ | |
1714 | if (reload_in_progress || reload_completed) | |
f756078b | 1715 | { |
1716 | scratch_reg = scratch_reg ? scratch_reg : operand0; | |
1717 | /* SCRATCH_REG will hold an address and maybe the actual | |
1718 | data. We want it in WORD_MODE regardless of what mode it | |
1719 | was originally given to us. */ | |
9840d99d | 1720 | scratch_reg = force_mode (word_mode, scratch_reg); |
f756078b | 1721 | } |
81653f9b | 1722 | else if (flag_pic) |
1723 | scratch_reg = gen_reg_rtx (Pmode); | |
1724 | ||
37a75d53 | 1725 | if (GET_CODE (operand1) == CONST) |
1726 | { | |
1727 | /* Save away the constant part of the expression. */ | |
1728 | const_part = XEXP (XEXP (operand1, 0), 1); | |
1729 | if (GET_CODE (const_part) != CONST_INT) | |
1730 | abort (); | |
1731 | ||
1732 | /* Force the function label into memory. */ | |
1733 | temp = force_const_mem (mode, XEXP (XEXP (operand1, 0), 0)); | |
1734 | } | |
1735 | else | |
1736 | { | |
1737 | /* No constant part. */ | |
1738 | const_part = NULL_RTX; | |
005a7dd0 | 1739 | |
37a75d53 | 1740 | /* Force the function label into memory. */ |
1741 | temp = force_const_mem (mode, operand1); | |
1742 | } | |
9840d99d | 1743 | |
81653f9b | 1744 | |
1745 | /* Get the address of the memory location. PIC-ify it if | |
1746 | necessary. */ | |
1747 | temp = XEXP (temp, 0); | |
1748 | if (flag_pic) | |
1749 | temp = legitimize_pic_address (temp, mode, scratch_reg); | |
1750 | ||
1751 | /* Put the address of the memory location into our destination | |
1752 | register. */ | |
1753 | operands[1] = temp; | |
1754 | emit_move_sequence (operands, mode, scratch_reg); | |
1755 | ||
1756 | /* Now load from the memory location into our destination | |
1757 | register. */ | |
ad851752 | 1758 | operands[1] = gen_rtx_MEM (Pmode, operands[0]); |
81653f9b | 1759 | emit_move_sequence (operands, mode, scratch_reg); |
1760 | ||
1761 | /* And add back in the constant part. */ | |
37a75d53 | 1762 | if (const_part != NULL_RTX) |
1763 | expand_inc (operand0, const_part); | |
81653f9b | 1764 | |
1765 | return 1; | |
005a7dd0 | 1766 | } |
1767 | ||
87ad11b0 | 1768 | if (flag_pic) |
1769 | { | |
2ff4bf8d | 1770 | rtx temp; |
1771 | ||
1772 | if (reload_in_progress || reload_completed) | |
f756078b | 1773 | { |
1774 | temp = scratch_reg ? scratch_reg : operand0; | |
1775 | /* TEMP will hold an address and maybe the actual | |
1776 | data. We want it in WORD_MODE regardless of what mode it | |
1777 | was originally given to us. */ | |
9840d99d | 1778 | temp = force_mode (word_mode, temp); |
f756078b | 1779 | } |
2ff4bf8d | 1780 | else |
1781 | temp = gen_reg_rtx (Pmode); | |
6d36483b | 1782 | |
81653f9b | 1783 | /* (const (plus (symbol) (const_int))) must be forced to |
1784 | memory during/after reload if the const_int will not fit | |
1785 | in 14 bits. */ | |
1786 | if (GET_CODE (operand1) == CONST | |
96b86ab6 | 1787 | && GET_CODE (XEXP (operand1, 0)) == PLUS |
1788 | && GET_CODE (XEXP (XEXP (operand1, 0), 1)) == CONST_INT | |
1789 | && !INT_14_BITS (XEXP (XEXP (operand1, 0), 1)) | |
1790 | && (reload_completed || reload_in_progress) | |
1791 | && flag_pic) | |
1792 | { | |
1793 | operands[1] = force_const_mem (mode, operand1); | |
1794 | operands[1] = legitimize_pic_address (XEXP (operands[1], 0), | |
1795 | mode, temp); | |
1796 | emit_move_sequence (operands, mode, temp); | |
1797 | } | |
005a7dd0 | 1798 | else |
1799 | { | |
1800 | operands[1] = legitimize_pic_address (operand1, mode, temp); | |
ad851752 | 1801 | emit_insn (gen_rtx_SET (VOIDmode, operand0, operands[1])); |
005a7dd0 | 1802 | } |
87ad11b0 | 1803 | } |
b4a7bf10 | 1804 | /* On the HPPA, references to data space are supposed to use dp, |
1805 | register 27, but showing it in the RTL inhibits various cse | |
1806 | and loop optimizations. */ | |
6d36483b | 1807 | else |
87ad11b0 | 1808 | { |
005a7dd0 | 1809 | rtx temp, set; |
2ee034bc | 1810 | |
6d36483b | 1811 | if (reload_in_progress || reload_completed) |
f756078b | 1812 | { |
1813 | temp = scratch_reg ? scratch_reg : operand0; | |
1814 | /* TEMP will hold an address and maybe the actual | |
1815 | data. We want it in WORD_MODE regardless of what mode it | |
1816 | was originally given to us. */ | |
9840d99d | 1817 | temp = force_mode (word_mode, temp); |
f756078b | 1818 | } |
2ee034bc | 1819 | else |
1820 | temp = gen_reg_rtx (mode); | |
1821 | ||
42819d4e | 1822 | /* Loading a SYMBOL_REF into a register makes that register |
9840d99d | 1823 | safe to be used as the base in an indexed address. |
42819d4e | 1824 | |
1825 | Don't mark hard registers though. That loses. */ | |
47a61b79 | 1826 | if (GET_CODE (operand0) == REG |
1827 | && REGNO (operand0) >= FIRST_PSEUDO_REGISTER) | |
e61a0a7f | 1828 | REG_POINTER (operand0) = 1; |
42819d4e | 1829 | if (REGNO (temp) >= FIRST_PSEUDO_REGISTER) |
e61a0a7f | 1830 | REG_POINTER (temp) = 1; |
2ee034bc | 1831 | if (ishighonly) |
ad851752 | 1832 | set = gen_rtx_SET (mode, operand0, temp); |
2ee034bc | 1833 | else |
7014838c | 1834 | set = gen_rtx_SET (VOIDmode, |
1835 | operand0, | |
ad851752 | 1836 | gen_rtx_LO_SUM (mode, temp, operand1)); |
6d36483b | 1837 | |
ad851752 | 1838 | emit_insn (gen_rtx_SET (VOIDmode, |
1839 | temp, | |
1840 | gen_rtx_HIGH (mode, operand1))); | |
d2498717 | 1841 | emit_insn (set); |
166bf021 | 1842 | |
87ad11b0 | 1843 | } |
2ee034bc | 1844 | return 1; |
87ad11b0 | 1845 | } |
42faba01 | 1846 | else if (GET_CODE (operand1) != CONST_INT |
8c8ec4de | 1847 | || ! cint_ok_for_move (INTVAL (operand1))) |
87ad11b0 | 1848 | { |
93f6e1c7 | 1849 | rtx extend = NULL_RTX; |
2ff4bf8d | 1850 | rtx temp; |
5e3c5739 | 1851 | |
1852 | if (TARGET_64BIT && GET_CODE (operand1) == CONST_INT | |
b7d86581 | 1853 | && HOST_BITS_PER_WIDE_INT > 32 |
5e3c5739 | 1854 | && GET_MODE_BITSIZE (GET_MODE (operand0)) > 32) |
1855 | { | |
1856 | HOST_WIDE_INT val = INTVAL (operand1); | |
b7d86581 | 1857 | HOST_WIDE_INT nval; |
5e3c5739 | 1858 | |
93f6e1c7 | 1859 | /* Extract the low order 32 bits of the value and sign extend. |
1860 | If the new value is the same as the original value, we can | |
1861 | can use the original value as-is. If the new value is | |
1862 | different, we use it and insert the most-significant 32-bits | |
1863 | of the original value into the final result. */ | |
b7d86581 | 1864 | nval = ((val & (((HOST_WIDE_INT) 2 << 31) - 1)) |
1865 | ^ ((HOST_WIDE_INT) 1 << 31)) - ((HOST_WIDE_INT) 1 << 31); | |
5e3c5739 | 1866 | if (val != nval) |
1867 | { | |
93f6e1c7 | 1868 | #if HOST_BITS_PER_WIDE_INT > 32 |
1869 | extend = GEN_INT (val >> 32); | |
1870 | #endif | |
5e3c5739 | 1871 | operand1 = GEN_INT (nval); |
1872 | } | |
1873 | } | |
2ff4bf8d | 1874 | |
1875 | if (reload_in_progress || reload_completed) | |
1876 | temp = operand0; | |
1877 | else | |
1878 | temp = gen_reg_rtx (mode); | |
1879 | ||
7c4d3047 | 1880 | /* We don't directly split DImode constants on 32-bit targets |
1881 | because PLUS uses an 11-bit immediate and the insn sequence | |
1882 | generated is not as efficient as the one using HIGH/LO_SUM. */ | |
1883 | if (GET_CODE (operand1) == CONST_INT | |
1884 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
93f6e1c7 | 1885 | { |
7c4d3047 | 1886 | /* Directly break constant into high and low parts. This |
93f6e1c7 | 1887 | provides better optimization opportunities because various |
1888 | passes recognize constants split with PLUS but not LO_SUM. | |
1889 | We use a 14-bit signed low part except when the addition | |
1890 | of 0x4000 to the high part might change the sign of the | |
1891 | high part. */ | |
1892 | HOST_WIDE_INT value = INTVAL (operand1); | |
1893 | HOST_WIDE_INT low = value & 0x3fff; | |
1894 | HOST_WIDE_INT high = value & ~ 0x3fff; | |
1895 | ||
1896 | if (low >= 0x2000) | |
1897 | { | |
1898 | if (high == 0x7fffc000 || (mode == HImode && high == 0x4000)) | |
1899 | high += 0x2000; | |
1900 | else | |
1901 | high += 0x4000; | |
1902 | } | |
1903 | ||
1904 | low = value - high; | |
5e3c5739 | 1905 | |
93f6e1c7 | 1906 | emit_insn (gen_rtx_SET (VOIDmode, temp, GEN_INT (high))); |
1907 | operands[1] = gen_rtx_PLUS (mode, temp, GEN_INT (low)); | |
1908 | } | |
1909 | else | |
5e3c5739 | 1910 | { |
93f6e1c7 | 1911 | emit_insn (gen_rtx_SET (VOIDmode, temp, |
1912 | gen_rtx_HIGH (mode, operand1))); | |
1913 | operands[1] = gen_rtx_LO_SUM (mode, temp, operand1); | |
5e3c5739 | 1914 | } |
9840d99d | 1915 | |
93f6e1c7 | 1916 | emit_move_insn (operands[0], operands[1]); |
1917 | ||
1918 | if (extend != NULL_RTX) | |
1919 | emit_insn (gen_insv (operands[0], GEN_INT (32), const0_rtx, | |
1920 | extend)); | |
1921 | ||
5e3c5739 | 1922 | return 1; |
87ad11b0 | 1923 | } |
1924 | } | |
1925 | /* Now have insn-emit do whatever it normally does. */ | |
1926 | return 0; | |
1927 | } | |
1928 | ||
1946138e | 1929 | /* Examine EXP and return nonzero if it contains an ADDR_EXPR (meaning |
bd49d362 | 1930 | it will need a link/runtime reloc). */ |
1946138e | 1931 | |
1932 | int | |
1933 | reloc_needed (exp) | |
1934 | tree exp; | |
1935 | { | |
1936 | int reloc = 0; | |
1937 | ||
1938 | switch (TREE_CODE (exp)) | |
1939 | { | |
1940 | case ADDR_EXPR: | |
1941 | return 1; | |
1942 | ||
1943 | case PLUS_EXPR: | |
1944 | case MINUS_EXPR: | |
1945 | reloc = reloc_needed (TREE_OPERAND (exp, 0)); | |
1946 | reloc |= reloc_needed (TREE_OPERAND (exp, 1)); | |
1947 | break; | |
1948 | ||
1949 | case NOP_EXPR: | |
1950 | case CONVERT_EXPR: | |
1951 | case NON_LVALUE_EXPR: | |
1952 | reloc = reloc_needed (TREE_OPERAND (exp, 0)); | |
1953 | break; | |
1954 | ||
1955 | case CONSTRUCTOR: | |
1956 | { | |
1957 | register tree link; | |
1958 | for (link = CONSTRUCTOR_ELTS (exp); link; link = TREE_CHAIN (link)) | |
1959 | if (TREE_VALUE (link) != 0) | |
1960 | reloc |= reloc_needed (TREE_VALUE (link)); | |
1961 | } | |
1962 | break; | |
1963 | ||
1964 | case ERROR_MARK: | |
1965 | break; | |
7d27e4c9 | 1966 | |
1967 | default: | |
1968 | break; | |
1946138e | 1969 | } |
1970 | return reloc; | |
1971 | } | |
1972 | ||
7811991d | 1973 | /* Does operand (which is a symbolic_operand) live in text space? |
1974 | If so, SYMBOL_REF_FLAG, which is set by pa_encode_section_info, | |
1975 | will be true. */ | |
87ad11b0 | 1976 | |
1977 | int | |
611a88e1 | 1978 | read_only_operand (operand, mode) |
87ad11b0 | 1979 | rtx operand; |
611a88e1 | 1980 | enum machine_mode mode ATTRIBUTE_UNUSED; |
87ad11b0 | 1981 | { |
1982 | if (GET_CODE (operand) == CONST) | |
1983 | operand = XEXP (XEXP (operand, 0), 0); | |
b4a7bf10 | 1984 | if (flag_pic) |
1985 | { | |
1986 | if (GET_CODE (operand) == SYMBOL_REF) | |
1987 | return SYMBOL_REF_FLAG (operand) && !CONSTANT_POOL_ADDRESS_P (operand); | |
1988 | } | |
1989 | else | |
1990 | { | |
1991 | if (GET_CODE (operand) == SYMBOL_REF) | |
1992 | return SYMBOL_REF_FLAG (operand) || CONSTANT_POOL_ADDRESS_P (operand); | |
1993 | } | |
87ad11b0 | 1994 | return 1; |
1995 | } | |
6d36483b | 1996 | |
87ad11b0 | 1997 | \f |
1998 | /* Return the best assembler insn template | |
f54b1341 | 1999 | for moving operands[1] into operands[0] as a fullword. */ |
611a88e1 | 2000 | const char * |
87ad11b0 | 2001 | singlemove_string (operands) |
2002 | rtx *operands; | |
2003 | { | |
3745c59b | 2004 | HOST_WIDE_INT intval; |
2005 | ||
87ad11b0 | 2006 | if (GET_CODE (operands[0]) == MEM) |
2007 | return "stw %r1,%0"; | |
3745c59b | 2008 | if (GET_CODE (operands[1]) == MEM) |
87ad11b0 | 2009 | return "ldw %1,%0"; |
3745c59b | 2010 | if (GET_CODE (operands[1]) == CONST_DOUBLE) |
9d5108ea | 2011 | { |
3745c59b | 2012 | long i; |
2013 | REAL_VALUE_TYPE d; | |
9d5108ea | 2014 | |
3745c59b | 2015 | if (GET_MODE (operands[1]) != SFmode) |
2016 | abort (); | |
9d5108ea | 2017 | |
3745c59b | 2018 | /* Translate the CONST_DOUBLE to a CONST_INT with the same target |
2019 | bit pattern. */ | |
2020 | REAL_VALUE_FROM_CONST_DOUBLE (d, operands[1]); | |
2021 | REAL_VALUE_TO_TARGET_SINGLE (d, i); | |
9d5108ea | 2022 | |
3745c59b | 2023 | operands[1] = GEN_INT (i); |
2024 | /* Fall through to CONST_INT case. */ | |
2025 | } | |
2026 | if (GET_CODE (operands[1]) == CONST_INT) | |
9d5108ea | 2027 | { |
3745c59b | 2028 | intval = INTVAL (operands[1]); |
2029 | ||
2030 | if (VAL_14_BITS_P (intval)) | |
2031 | return "ldi %1,%0"; | |
2032 | else if ((intval & 0x7ff) == 0) | |
2033 | return "ldil L'%1,%0"; | |
2034 | else if (zdepi_cint_p (intval)) | |
e4065f95 | 2035 | return "{zdepi %Z1,%0|depwi,z %Z1,%0}"; |
9d5108ea | 2036 | else |
2037 | return "ldil L'%1,%0\n\tldo R'%1(%0),%0"; | |
2038 | } | |
87ad11b0 | 2039 | return "copy %1,%0"; |
2040 | } | |
2041 | \f | |
2042 | ||
201f01e9 | 2043 | /* Compute position (in OP[1]) and width (in OP[2]) |
2044 | useful for copying IMM to a register using the zdepi | |
2045 | instructions. Store the immediate value to insert in OP[0]. */ | |
611a88e1 | 2046 | static void |
fb22aedc | 2047 | compute_zdepwi_operands (imm, op) |
6d36483b | 2048 | unsigned HOST_WIDE_INT imm; |
42faba01 | 2049 | unsigned *op; |
7e10ba53 | 2050 | { |
e057641f | 2051 | int lsb, len; |
7e10ba53 | 2052 | |
e057641f | 2053 | /* Find the least significant set bit in IMM. */ |
2054 | for (lsb = 0; lsb < 32; lsb++) | |
7e10ba53 | 2055 | { |
e057641f | 2056 | if ((imm & 1) != 0) |
7e10ba53 | 2057 | break; |
e057641f | 2058 | imm >>= 1; |
7e10ba53 | 2059 | } |
2060 | ||
e057641f | 2061 | /* Choose variants based on *sign* of the 5-bit field. */ |
2062 | if ((imm & 0x10) == 0) | |
2063 | len = (lsb <= 28) ? 4 : 32 - lsb; | |
7e10ba53 | 2064 | else |
2065 | { | |
e057641f | 2066 | /* Find the width of the bitstring in IMM. */ |
2067 | for (len = 5; len < 32; len++) | |
7e10ba53 | 2068 | { |
e057641f | 2069 | if ((imm & (1 << len)) == 0) |
7e10ba53 | 2070 | break; |
7e10ba53 | 2071 | } |
2072 | ||
e057641f | 2073 | /* Sign extend IMM as a 5-bit value. */ |
2074 | imm = (imm & 0xf) - 0x10; | |
7e10ba53 | 2075 | } |
2076 | ||
42faba01 | 2077 | op[0] = imm; |
2078 | op[1] = 31 - lsb; | |
2079 | op[2] = len; | |
7e10ba53 | 2080 | } |
2081 | ||
5e3c5739 | 2082 | /* Compute position (in OP[1]) and width (in OP[2]) |
2083 | useful for copying IMM to a register using the depdi,z | |
2084 | instructions. Store the immediate value to insert in OP[0]. */ | |
2085 | void | |
2086 | compute_zdepdi_operands (imm, op) | |
2087 | unsigned HOST_WIDE_INT imm; | |
2088 | unsigned *op; | |
2089 | { | |
2090 | HOST_WIDE_INT lsb, len; | |
2091 | ||
2092 | /* Find the least significant set bit in IMM. */ | |
2093 | for (lsb = 0; lsb < HOST_BITS_PER_WIDE_INT; lsb++) | |
2094 | { | |
2095 | if ((imm & 1) != 0) | |
2096 | break; | |
2097 | imm >>= 1; | |
2098 | } | |
2099 | ||
2100 | /* Choose variants based on *sign* of the 5-bit field. */ | |
2101 | if ((imm & 0x10) == 0) | |
2102 | len = ((lsb <= HOST_BITS_PER_WIDE_INT - 4) | |
2103 | ? 4 : HOST_BITS_PER_WIDE_INT - lsb); | |
2104 | else | |
2105 | { | |
2106 | /* Find the width of the bitstring in IMM. */ | |
2107 | for (len = 5; len < HOST_BITS_PER_WIDE_INT; len++) | |
2108 | { | |
ea52c577 | 2109 | if ((imm & ((unsigned HOST_WIDE_INT) 1 << len)) == 0) |
5e3c5739 | 2110 | break; |
2111 | } | |
2112 | ||
2113 | /* Sign extend IMM as a 5-bit value. */ | |
2114 | imm = (imm & 0xf) - 0x10; | |
2115 | } | |
2116 | ||
2117 | op[0] = imm; | |
2118 | op[1] = 63 - lsb; | |
2119 | op[2] = len; | |
2120 | } | |
2121 | ||
87ad11b0 | 2122 | /* Output assembler code to perform a doubleword move insn |
2123 | with operands OPERANDS. */ | |
2124 | ||
611a88e1 | 2125 | const char * |
87ad11b0 | 2126 | output_move_double (operands) |
2127 | rtx *operands; | |
2128 | { | |
2129 | enum { REGOP, OFFSOP, MEMOP, CNSTOP, RNDOP } optype0, optype1; | |
2130 | rtx latehalf[2]; | |
2131 | rtx addreg0 = 0, addreg1 = 0; | |
2132 | ||
2133 | /* First classify both operands. */ | |
2134 | ||
2135 | if (REG_P (operands[0])) | |
2136 | optype0 = REGOP; | |
2137 | else if (offsettable_memref_p (operands[0])) | |
2138 | optype0 = OFFSOP; | |
2139 | else if (GET_CODE (operands[0]) == MEM) | |
2140 | optype0 = MEMOP; | |
2141 | else | |
2142 | optype0 = RNDOP; | |
2143 | ||
2144 | if (REG_P (operands[1])) | |
2145 | optype1 = REGOP; | |
2146 | else if (CONSTANT_P (operands[1])) | |
2147 | optype1 = CNSTOP; | |
2148 | else if (offsettable_memref_p (operands[1])) | |
2149 | optype1 = OFFSOP; | |
2150 | else if (GET_CODE (operands[1]) == MEM) | |
2151 | optype1 = MEMOP; | |
2152 | else | |
2153 | optype1 = RNDOP; | |
2154 | ||
2155 | /* Check for the cases that the operand constraints are not | |
2156 | supposed to allow to happen. Abort if we get one, | |
2157 | because generating code for these cases is painful. */ | |
2158 | ||
2159 | if (optype0 != REGOP && optype1 != REGOP) | |
2160 | abort (); | |
2161 | ||
2162 | /* Handle auto decrementing and incrementing loads and stores | |
2163 | specifically, since the structure of the function doesn't work | |
2164 | for them without major modification. Do it better when we learn | |
2165 | this port about the general inc/dec addressing of PA. | |
2166 | (This was written by tege. Chide him if it doesn't work.) */ | |
2167 | ||
2168 | if (optype0 == MEMOP) | |
2169 | { | |
1df0058a | 2170 | /* We have to output the address syntax ourselves, since print_operand |
2171 | doesn't deal with the addresses we want to use. Fix this later. */ | |
2172 | ||
87ad11b0 | 2173 | rtx addr = XEXP (operands[0], 0); |
1df0058a | 2174 | if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC) |
87ad11b0 | 2175 | { |
ad851752 | 2176 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0); |
1df0058a | 2177 | |
2178 | operands[0] = XEXP (addr, 0); | |
2179 | if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG) | |
2180 | abort (); | |
2181 | ||
2182 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2183 | { | |
2184 | /* No overlap between high target register and address | |
2185 | register. (We do this in a non-obvious way to | |
2186 | save a register file writeback) */ | |
2187 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2188 | return "{stws|stw},ma %1,8(%0)\n\tstw %R1,-4(%0)"; |
2189 | return "{stws|stw},ma %1,-8(%0)\n\tstw %R1,12(%0)"; | |
1df0058a | 2190 | } |
2191 | else | |
ea52c577 | 2192 | abort (); |
a3217f65 | 2193 | } |
1df0058a | 2194 | else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC) |
a3217f65 | 2195 | { |
ad851752 | 2196 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0); |
1df0058a | 2197 | |
2198 | operands[0] = XEXP (addr, 0); | |
2199 | if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG) | |
2200 | abort (); | |
2201 | ||
2202 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2203 | { | |
2204 | /* No overlap between high target register and address | |
2205 | register. (We do this in a non-obvious way to | |
2206 | save a register file writeback) */ | |
2207 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2208 | return "{stws|stw},mb %1,8(%0)\n\tstw %R1,4(%0)"; |
2209 | return "{stws|stw},mb %1,-8(%0)\n\tstw %R1,4(%0)"; | |
1df0058a | 2210 | } |
2211 | else | |
ea52c577 | 2212 | abort (); |
87ad11b0 | 2213 | } |
2214 | } | |
2215 | if (optype1 == MEMOP) | |
2216 | { | |
2217 | /* We have to output the address syntax ourselves, since print_operand | |
2218 | doesn't deal with the addresses we want to use. Fix this later. */ | |
2219 | ||
2220 | rtx addr = XEXP (operands[1], 0); | |
2221 | if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC) | |
2222 | { | |
ad851752 | 2223 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
87ad11b0 | 2224 | |
2225 | operands[1] = XEXP (addr, 0); | |
2226 | if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG) | |
2227 | abort (); | |
2228 | ||
2229 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2230 | { | |
2231 | /* No overlap between high target register and address | |
3857fa62 | 2232 | register. (We do this in a non-obvious way to |
87ad11b0 | 2233 | save a register file writeback) */ |
2234 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2235 | return "{ldws|ldw},ma 8(%1),%0\n\tldw -4(%1),%R0"; |
01fd4b49 | 2236 | return "{ldws|ldw},ma -8(%1),%0\n\tldw 12(%1),%R0"; |
87ad11b0 | 2237 | } |
2238 | else | |
2239 | { | |
2240 | /* This is an undefined situation. We should load into the | |
2241 | address register *and* update that register. Probably | |
2242 | we don't need to handle this at all. */ | |
2243 | if (GET_CODE (addr) == POST_INC) | |
e4065f95 | 2244 | return "ldw 4(%1),%R0\n\t{ldws|ldw},ma 8(%1),%0"; |
2245 | return "ldw 4(%1),%R0\n\t{ldws|ldw},ma -8(%1),%0"; | |
87ad11b0 | 2246 | } |
2247 | } | |
2248 | else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC) | |
2249 | { | |
ad851752 | 2250 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
87ad11b0 | 2251 | |
2252 | operands[1] = XEXP (addr, 0); | |
2253 | if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG) | |
2254 | abort (); | |
2255 | ||
2256 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2257 | { | |
2258 | /* No overlap between high target register and address | |
3857fa62 | 2259 | register. (We do this in a non-obvious way to |
87ad11b0 | 2260 | save a register file writeback) */ |
2261 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2262 | return "{ldws|ldw},mb 8(%1),%0\n\tldw 4(%1),%R0"; |
2263 | return "{ldws|ldw},mb -8(%1),%0\n\tldw 4(%1),%R0"; | |
87ad11b0 | 2264 | } |
2265 | else | |
2266 | { | |
2267 | /* This is an undefined situation. We should load into the | |
2268 | address register *and* update that register. Probably | |
2269 | we don't need to handle this at all. */ | |
2270 | if (GET_CODE (addr) == PRE_INC) | |
e4065f95 | 2271 | return "ldw 12(%1),%R0\n\t{ldws|ldw},mb 8(%1),%0"; |
2272 | return "ldw -4(%1),%R0\n\t{ldws|ldw},mb -8(%1),%0"; | |
87ad11b0 | 2273 | } |
2274 | } | |
12b02046 | 2275 | else if (GET_CODE (addr) == PLUS |
2276 | && GET_CODE (XEXP (addr, 0)) == MULT) | |
2277 | { | |
ad851752 | 2278 | rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0); |
12b02046 | 2279 | |
2280 | if (!reg_overlap_mentioned_p (high_reg, addr)) | |
2281 | { | |
2282 | rtx xoperands[3]; | |
2283 | ||
2284 | xoperands[0] = high_reg; | |
2285 | xoperands[1] = XEXP (addr, 1); | |
2286 | xoperands[2] = XEXP (XEXP (addr, 0), 0); | |
2287 | xoperands[3] = XEXP (XEXP (addr, 0), 1); | |
e4065f95 | 2288 | output_asm_insn ("{sh%O3addl %2,%1,%0|shladd,l %2,%O3,%1,%0}", |
2289 | xoperands); | |
34940871 | 2290 | return "ldw 4(%0),%R0\n\tldw 0(%0),%0"; |
12b02046 | 2291 | } |
2292 | else | |
2293 | { | |
2294 | rtx xoperands[3]; | |
2295 | ||
2296 | xoperands[0] = high_reg; | |
2297 | xoperands[1] = XEXP (addr, 1); | |
2298 | xoperands[2] = XEXP (XEXP (addr, 0), 0); | |
2299 | xoperands[3] = XEXP (XEXP (addr, 0), 1); | |
e4065f95 | 2300 | output_asm_insn ("{sh%O3addl %2,%1,%R0|shladd,l %2,%O3,%1,%R0}", |
2301 | xoperands); | |
34940871 | 2302 | return "ldw 0(%R0),%0\n\tldw 4(%R0),%R0"; |
12b02046 | 2303 | } |
12b02046 | 2304 | } |
87ad11b0 | 2305 | } |
2306 | ||
2307 | /* If an operand is an unoffsettable memory ref, find a register | |
2308 | we can increment temporarily to make it refer to the second word. */ | |
2309 | ||
2310 | if (optype0 == MEMOP) | |
2311 | addreg0 = find_addr_reg (XEXP (operands[0], 0)); | |
2312 | ||
2313 | if (optype1 == MEMOP) | |
2314 | addreg1 = find_addr_reg (XEXP (operands[1], 0)); | |
2315 | ||
2316 | /* Ok, we can do one word at a time. | |
2317 | Normally we do the low-numbered word first. | |
2318 | ||
2319 | In either case, set up in LATEHALF the operands to use | |
2320 | for the high-numbered word and in some cases alter the | |
2321 | operands in OPERANDS to be suitable for the low-numbered word. */ | |
2322 | ||
2323 | if (optype0 == REGOP) | |
ad851752 | 2324 | latehalf[0] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); |
87ad11b0 | 2325 | else if (optype0 == OFFSOP) |
eafc6604 | 2326 | latehalf[0] = adjust_address (operands[0], SImode, 4); |
87ad11b0 | 2327 | else |
2328 | latehalf[0] = operands[0]; | |
2329 | ||
2330 | if (optype1 == REGOP) | |
ad851752 | 2331 | latehalf[1] = gen_rtx_REG (SImode, REGNO (operands[1]) + 1); |
87ad11b0 | 2332 | else if (optype1 == OFFSOP) |
eafc6604 | 2333 | latehalf[1] = adjust_address (operands[1], SImode, 4); |
87ad11b0 | 2334 | else if (optype1 == CNSTOP) |
2335 | split_double (operands[1], &operands[1], &latehalf[1]); | |
2336 | else | |
2337 | latehalf[1] = operands[1]; | |
2338 | ||
2339 | /* If the first move would clobber the source of the second one, | |
2340 | do them in the other order. | |
2341 | ||
cf489d53 | 2342 | This can happen in two cases: |
87ad11b0 | 2343 | |
cf489d53 | 2344 | mem -> register where the first half of the destination register |
2345 | is the same register used in the memory's address. Reload | |
2346 | can create such insns. | |
87ad11b0 | 2347 | |
cf489d53 | 2348 | mem in this case will be either register indirect or register |
9840d99d | 2349 | indirect plus a valid offset. |
cf489d53 | 2350 | |
2351 | register -> register move where REGNO(dst) == REGNO(src + 1) | |
9840d99d | 2352 | someone (Tim/Tege?) claimed this can happen for parameter loads. |
cf489d53 | 2353 | |
2354 | Handle mem -> register case first. */ | |
2355 | if (optype0 == REGOP | |
2356 | && (optype1 == MEMOP || optype1 == OFFSOP) | |
2357 | && refers_to_regno_p (REGNO (operands[0]), REGNO (operands[0]) + 1, | |
2358 | operands[1], 0)) | |
87ad11b0 | 2359 | { |
87ad11b0 | 2360 | /* Do the late half first. */ |
2361 | if (addreg1) | |
6a5d085a | 2362 | output_asm_insn ("ldo 4(%0),%0", &addreg1); |
87ad11b0 | 2363 | output_asm_insn (singlemove_string (latehalf), latehalf); |
cf489d53 | 2364 | |
2365 | /* Then clobber. */ | |
87ad11b0 | 2366 | if (addreg1) |
6a5d085a | 2367 | output_asm_insn ("ldo -4(%0),%0", &addreg1); |
87ad11b0 | 2368 | return singlemove_string (operands); |
2369 | } | |
2370 | ||
cf489d53 | 2371 | /* Now handle register -> register case. */ |
c4fa5937 | 2372 | if (optype0 == REGOP && optype1 == REGOP |
2373 | && REGNO (operands[0]) == REGNO (operands[1]) + 1) | |
2374 | { | |
2375 | output_asm_insn (singlemove_string (latehalf), latehalf); | |
2376 | return singlemove_string (operands); | |
2377 | } | |
2378 | ||
87ad11b0 | 2379 | /* Normal case: do the two words, low-numbered first. */ |
2380 | ||
2381 | output_asm_insn (singlemove_string (operands), operands); | |
2382 | ||
2383 | /* Make any unoffsettable addresses point at high-numbered word. */ | |
2384 | if (addreg0) | |
6a5d085a | 2385 | output_asm_insn ("ldo 4(%0),%0", &addreg0); |
87ad11b0 | 2386 | if (addreg1) |
6a5d085a | 2387 | output_asm_insn ("ldo 4(%0),%0", &addreg1); |
87ad11b0 | 2388 | |
2389 | /* Do that word. */ | |
2390 | output_asm_insn (singlemove_string (latehalf), latehalf); | |
2391 | ||
2392 | /* Undo the adds we just did. */ | |
2393 | if (addreg0) | |
6a5d085a | 2394 | output_asm_insn ("ldo -4(%0),%0", &addreg0); |
87ad11b0 | 2395 | if (addreg1) |
6a5d085a | 2396 | output_asm_insn ("ldo -4(%0),%0", &addreg1); |
87ad11b0 | 2397 | |
2398 | return ""; | |
2399 | } | |
2400 | \f | |
611a88e1 | 2401 | const char * |
87ad11b0 | 2402 | output_fp_move_double (operands) |
2403 | rtx *operands; | |
2404 | { | |
2405 | if (FP_REG_P (operands[0])) | |
2406 | { | |
6d36483b | 2407 | if (FP_REG_P (operands[1]) |
891b55b4 | 2408 | || operands[1] == CONST0_RTX (GET_MODE (operands[0]))) |
c6ae275c | 2409 | output_asm_insn ("fcpy,dbl %f1,%0", operands); |
6d36483b | 2410 | else |
27ef382d | 2411 | output_asm_insn ("fldd%F1 %1,%0", operands); |
87ad11b0 | 2412 | } |
2413 | else if (FP_REG_P (operands[1])) | |
2414 | { | |
27ef382d | 2415 | output_asm_insn ("fstd%F0 %1,%0", operands); |
87ad11b0 | 2416 | } |
891b55b4 | 2417 | else if (operands[1] == CONST0_RTX (GET_MODE (operands[0]))) |
2418 | { | |
2419 | if (GET_CODE (operands[0]) == REG) | |
2420 | { | |
2421 | rtx xoperands[2]; | |
ad851752 | 2422 | xoperands[1] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1); |
891b55b4 | 2423 | xoperands[0] = operands[0]; |
2424 | output_asm_insn ("copy %%r0,%0\n\tcopy %%r0,%1", xoperands); | |
2425 | } | |
6d36483b | 2426 | /* This is a pain. You have to be prepared to deal with an |
01cc3b75 | 2427 | arbitrary address here including pre/post increment/decrement. |
891b55b4 | 2428 | |
2429 | so avoid this in the MD. */ | |
2430 | else | |
2431 | abort (); | |
2432 | } | |
87ad11b0 | 2433 | else abort (); |
2434 | return ""; | |
2435 | } | |
2436 | \f | |
2437 | /* Return a REG that occurs in ADDR with coefficient 1. | |
2438 | ADDR can be effectively incremented by incrementing REG. */ | |
2439 | ||
2440 | static rtx | |
2441 | find_addr_reg (addr) | |
2442 | rtx addr; | |
2443 | { | |
2444 | while (GET_CODE (addr) == PLUS) | |
2445 | { | |
2446 | if (GET_CODE (XEXP (addr, 0)) == REG) | |
2447 | addr = XEXP (addr, 0); | |
2448 | else if (GET_CODE (XEXP (addr, 1)) == REG) | |
2449 | addr = XEXP (addr, 1); | |
2450 | else if (CONSTANT_P (XEXP (addr, 0))) | |
2451 | addr = XEXP (addr, 1); | |
2452 | else if (CONSTANT_P (XEXP (addr, 1))) | |
2453 | addr = XEXP (addr, 0); | |
2454 | else | |
2455 | abort (); | |
2456 | } | |
2457 | if (GET_CODE (addr) == REG) | |
2458 | return addr; | |
2459 | abort (); | |
2460 | } | |
2461 | ||
87ad11b0 | 2462 | /* Emit code to perform a block move. |
2463 | ||
87ad11b0 | 2464 | OPERANDS[0] is the destination pointer as a REG, clobbered. |
2465 | OPERANDS[1] is the source pointer as a REG, clobbered. | |
42819d4e | 2466 | OPERANDS[2] is a register for temporary storage. |
2467 | OPERANDS[4] is the size as a CONST_INT | |
87ad11b0 | 2468 | OPERANDS[3] is a register for temporary storage. |
9840d99d | 2469 | OPERANDS[5] is the alignment safe to use, as a CONST_INT. |
ad87de1e | 2470 | OPERANDS[6] is another temporary register. */ |
87ad11b0 | 2471 | |
611a88e1 | 2472 | const char * |
87ad11b0 | 2473 | output_block_move (operands, size_is_constant) |
2474 | rtx *operands; | |
b1ca791d | 2475 | int size_is_constant ATTRIBUTE_UNUSED; |
87ad11b0 | 2476 | { |
2477 | int align = INTVAL (operands[5]); | |
42819d4e | 2478 | unsigned long n_bytes = INTVAL (operands[4]); |
87ad11b0 | 2479 | |
2480 | /* We can't move more than four bytes at a time because the PA | |
2481 | has no longer integer move insns. (Could use fp mem ops?) */ | |
2482 | if (align > 4) | |
2483 | align = 4; | |
2484 | ||
42819d4e | 2485 | /* Note that we know each loop below will execute at least twice |
2486 | (else we would have open-coded the copy). */ | |
2487 | switch (align) | |
87ad11b0 | 2488 | { |
42819d4e | 2489 | case 4: |
2490 | /* Pre-adjust the loop counter. */ | |
2491 | operands[4] = GEN_INT (n_bytes - 8); | |
2492 | output_asm_insn ("ldi %4,%2", operands); | |
2493 | ||
2494 | /* Copying loop. */ | |
e4065f95 | 2495 | output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands); |
2496 | output_asm_insn ("{ldws|ldw},ma 4(%1),%6", operands); | |
2497 | output_asm_insn ("{stws|stw},ma %3,4(%0)", operands); | |
42819d4e | 2498 | output_asm_insn ("addib,>= -8,%2,.-12", operands); |
e4065f95 | 2499 | output_asm_insn ("{stws|stw},ma %6,4(%0)", operands); |
42819d4e | 2500 | |
2501 | /* Handle the residual. There could be up to 7 bytes of | |
2502 | residual to copy! */ | |
2503 | if (n_bytes % 8 != 0) | |
2504 | { | |
2505 | operands[4] = GEN_INT (n_bytes % 4); | |
2506 | if (n_bytes % 8 >= 4) | |
e4065f95 | 2507 | output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands); |
42819d4e | 2508 | if (n_bytes % 4 != 0) |
34940871 | 2509 | output_asm_insn ("ldw 0(%1),%6", operands); |
42819d4e | 2510 | if (n_bytes % 8 >= 4) |
e4065f95 | 2511 | output_asm_insn ("{stws|stw},ma %3,4(%0)", operands); |
42819d4e | 2512 | if (n_bytes % 4 != 0) |
e4065f95 | 2513 | output_asm_insn ("{stbys|stby},e %6,%4(%0)", operands); |
42819d4e | 2514 | } |
2515 | return ""; | |
87ad11b0 | 2516 | |
42819d4e | 2517 | case 2: |
2518 | /* Pre-adjust the loop counter. */ | |
2519 | operands[4] = GEN_INT (n_bytes - 4); | |
2520 | output_asm_insn ("ldi %4,%2", operands); | |
87ad11b0 | 2521 | |
42819d4e | 2522 | /* Copying loop. */ |
e4065f95 | 2523 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands); |
2524 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%6", operands); | |
2525 | output_asm_insn ("{sths|sth},ma %3,2(%0)", operands); | |
42819d4e | 2526 | output_asm_insn ("addib,>= -4,%2,.-12", operands); |
e4065f95 | 2527 | output_asm_insn ("{sths|sth},ma %6,2(%0)", operands); |
87ad11b0 | 2528 | |
42819d4e | 2529 | /* Handle the residual. */ |
2530 | if (n_bytes % 4 != 0) | |
2531 | { | |
2532 | if (n_bytes % 4 >= 2) | |
e4065f95 | 2533 | output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands); |
42819d4e | 2534 | if (n_bytes % 2 != 0) |
34940871 | 2535 | output_asm_insn ("ldb 0(%1),%6", operands); |
42819d4e | 2536 | if (n_bytes % 4 >= 2) |
e4065f95 | 2537 | output_asm_insn ("{sths|sth},ma %3,2(%0)", operands); |
42819d4e | 2538 | if (n_bytes % 2 != 0) |
34940871 | 2539 | output_asm_insn ("stb %6,0(%0)", operands); |
42819d4e | 2540 | } |
2541 | return ""; | |
87ad11b0 | 2542 | |
42819d4e | 2543 | case 1: |
2544 | /* Pre-adjust the loop counter. */ | |
2545 | operands[4] = GEN_INT (n_bytes - 2); | |
2546 | output_asm_insn ("ldi %4,%2", operands); | |
87ad11b0 | 2547 | |
42819d4e | 2548 | /* Copying loop. */ |
e4065f95 | 2549 | output_asm_insn ("{ldbs|ldb},ma 1(%1),%3", operands); |
2550 | output_asm_insn ("{ldbs|ldb},ma 1(%1),%6", operands); | |
2551 | output_asm_insn ("{stbs|stb},ma %3,1(%0)", operands); | |
42819d4e | 2552 | output_asm_insn ("addib,>= -2,%2,.-12", operands); |
e4065f95 | 2553 | output_asm_insn ("{stbs|stb},ma %6,1(%0)", operands); |
87ad11b0 | 2554 | |
42819d4e | 2555 | /* Handle the residual. */ |
2556 | if (n_bytes % 2 != 0) | |
2557 | { | |
34940871 | 2558 | output_asm_insn ("ldb 0(%1),%3", operands); |
2559 | output_asm_insn ("stb %3,0(%0)", operands); | |
42819d4e | 2560 | } |
2561 | return ""; | |
87ad11b0 | 2562 | |
42819d4e | 2563 | default: |
2564 | abort (); | |
87ad11b0 | 2565 | } |
87ad11b0 | 2566 | } |
58e17b0b | 2567 | |
2568 | /* Count the number of insns necessary to handle this block move. | |
2569 | ||
2570 | Basic structure is the same as emit_block_move, except that we | |
2571 | count insns rather than emit them. */ | |
2572 | ||
611a88e1 | 2573 | static int |
58e17b0b | 2574 | compute_movstrsi_length (insn) |
2575 | rtx insn; | |
2576 | { | |
2577 | rtx pat = PATTERN (insn); | |
fc1fb057 | 2578 | unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 7), 0)); |
2579 | unsigned long n_bytes = INTVAL (XEXP (XVECEXP (pat, 0, 6), 0)); | |
42819d4e | 2580 | unsigned int n_insns = 0; |
58e17b0b | 2581 | |
2582 | /* We can't move more than four bytes at a time because the PA | |
2583 | has no longer integer move insns. (Could use fp mem ops?) */ | |
2584 | if (align > 4) | |
2585 | align = 4; | |
2586 | ||
79bfe6ae | 2587 | /* The basic copying loop. */ |
42819d4e | 2588 | n_insns = 6; |
58e17b0b | 2589 | |
42819d4e | 2590 | /* Residuals. */ |
2591 | if (n_bytes % (2 * align) != 0) | |
58e17b0b | 2592 | { |
79bfe6ae | 2593 | if ((n_bytes % (2 * align)) >= align) |
2594 | n_insns += 2; | |
2595 | ||
2596 | if ((n_bytes % align) != 0) | |
2597 | n_insns += 2; | |
58e17b0b | 2598 | } |
42819d4e | 2599 | |
2600 | /* Lengths are expressed in bytes now; each insn is 4 bytes. */ | |
2601 | return n_insns * 4; | |
58e17b0b | 2602 | } |
87ad11b0 | 2603 | \f |
2604 | ||
611a88e1 | 2605 | const char * |
e057641f | 2606 | output_and (operands) |
2607 | rtx *operands; | |
2608 | { | |
d6f01525 | 2609 | if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0) |
e057641f | 2610 | { |
3745c59b | 2611 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); |
e057641f | 2612 | int ls0, ls1, ms0, p, len; |
2613 | ||
2614 | for (ls0 = 0; ls0 < 32; ls0++) | |
2615 | if ((mask & (1 << ls0)) == 0) | |
2616 | break; | |
2617 | ||
2618 | for (ls1 = ls0; ls1 < 32; ls1++) | |
2619 | if ((mask & (1 << ls1)) != 0) | |
2620 | break; | |
2621 | ||
2622 | for (ms0 = ls1; ms0 < 32; ms0++) | |
2623 | if ((mask & (1 << ms0)) == 0) | |
2624 | break; | |
2625 | ||
2626 | if (ms0 != 32) | |
ea52c577 | 2627 | abort (); |
e057641f | 2628 | |
2629 | if (ls1 == 32) | |
2630 | { | |
2631 | len = ls0; | |
2632 | ||
2633 | if (len == 0) | |
2634 | abort (); | |
2635 | ||
ef618fe4 | 2636 | operands[2] = GEN_INT (len); |
e4065f95 | 2637 | return "{extru|extrw,u} %1,31,%2,%0"; |
e057641f | 2638 | } |
2639 | else | |
2640 | { | |
2641 | /* We could use this `depi' for the case above as well, but `depi' | |
2642 | requires one more register file access than an `extru'. */ | |
2643 | ||
2644 | p = 31 - ls0; | |
2645 | len = ls1 - ls0; | |
2646 | ||
ef618fe4 | 2647 | operands[2] = GEN_INT (p); |
2648 | operands[3] = GEN_INT (len); | |
e4065f95 | 2649 | return "{depi|depwi} 0,%2,%3,%0"; |
e057641f | 2650 | } |
2651 | } | |
2652 | else | |
2653 | return "and %1,%2,%0"; | |
2654 | } | |
2655 | ||
5e3c5739 | 2656 | /* Return a string to perform a bitwise-and of operands[1] with operands[2] |
2657 | storing the result in operands[0]. */ | |
9aadea62 | 2658 | const char * |
5e3c5739 | 2659 | output_64bit_and (operands) |
2660 | rtx *operands; | |
2661 | { | |
2662 | if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0) | |
2663 | { | |
2664 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); | |
b7d86581 | 2665 | int ls0, ls1, ms0, p, len; |
5e3c5739 | 2666 | |
2667 | for (ls0 = 0; ls0 < HOST_BITS_PER_WIDE_INT; ls0++) | |
b7d86581 | 2668 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls0)) == 0) |
5e3c5739 | 2669 | break; |
2670 | ||
2671 | for (ls1 = ls0; ls1 < HOST_BITS_PER_WIDE_INT; ls1++) | |
b7d86581 | 2672 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls1)) != 0) |
5e3c5739 | 2673 | break; |
2674 | ||
2675 | for (ms0 = ls1; ms0 < HOST_BITS_PER_WIDE_INT; ms0++) | |
b7d86581 | 2676 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << ms0)) == 0) |
5e3c5739 | 2677 | break; |
2678 | ||
2679 | if (ms0 != HOST_BITS_PER_WIDE_INT) | |
b7d86581 | 2680 | abort (); |
5e3c5739 | 2681 | |
2682 | if (ls1 == HOST_BITS_PER_WIDE_INT) | |
2683 | { | |
2684 | len = ls0; | |
2685 | ||
2686 | if (len == 0) | |
2687 | abort (); | |
2688 | ||
2689 | operands[2] = GEN_INT (len); | |
2690 | return "extrd,u %1,63,%2,%0"; | |
2691 | } | |
2692 | else | |
2693 | { | |
2694 | /* We could use this `depi' for the case above as well, but `depi' | |
2695 | requires one more register file access than an `extru'. */ | |
2696 | ||
2697 | p = 63 - ls0; | |
2698 | len = ls1 - ls0; | |
2699 | ||
2700 | operands[2] = GEN_INT (p); | |
2701 | operands[3] = GEN_INT (len); | |
2702 | return "depdi 0,%2,%3,%0"; | |
2703 | } | |
2704 | } | |
2705 | else | |
2706 | return "and %1,%2,%0"; | |
2707 | } | |
2708 | ||
611a88e1 | 2709 | const char * |
e057641f | 2710 | output_ior (operands) |
2711 | rtx *operands; | |
2712 | { | |
3745c59b | 2713 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); |
57ed30e5 | 2714 | int bs0, bs1, p, len; |
6d36483b | 2715 | |
c9da5f4d | 2716 | if (INTVAL (operands[2]) == 0) |
2717 | return "copy %1,%0"; | |
e057641f | 2718 | |
c9da5f4d | 2719 | for (bs0 = 0; bs0 < 32; bs0++) |
2720 | if ((mask & (1 << bs0)) != 0) | |
2721 | break; | |
e057641f | 2722 | |
c9da5f4d | 2723 | for (bs1 = bs0; bs1 < 32; bs1++) |
2724 | if ((mask & (1 << bs1)) == 0) | |
2725 | break; | |
e057641f | 2726 | |
3745c59b | 2727 | if (bs1 != 32 && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask) |
ea52c577 | 2728 | abort (); |
e057641f | 2729 | |
c9da5f4d | 2730 | p = 31 - bs0; |
2731 | len = bs1 - bs0; | |
e057641f | 2732 | |
ef618fe4 | 2733 | operands[2] = GEN_INT (p); |
2734 | operands[3] = GEN_INT (len); | |
e4065f95 | 2735 | return "{depi|depwi} -1,%2,%3,%0"; |
e057641f | 2736 | } |
5e3c5739 | 2737 | |
2738 | /* Return a string to perform a bitwise-and of operands[1] with operands[2] | |
2739 | storing the result in operands[0]. */ | |
9aadea62 | 2740 | const char * |
5e3c5739 | 2741 | output_64bit_ior (operands) |
2742 | rtx *operands; | |
2743 | { | |
2744 | unsigned HOST_WIDE_INT mask = INTVAL (operands[2]); | |
b7d86581 | 2745 | int bs0, bs1, p, len; |
5e3c5739 | 2746 | |
2747 | if (INTVAL (operands[2]) == 0) | |
2748 | return "copy %1,%0"; | |
2749 | ||
2750 | for (bs0 = 0; bs0 < HOST_BITS_PER_WIDE_INT; bs0++) | |
b7d86581 | 2751 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs0)) != 0) |
5e3c5739 | 2752 | break; |
2753 | ||
2754 | for (bs1 = bs0; bs1 < HOST_BITS_PER_WIDE_INT; bs1++) | |
b7d86581 | 2755 | if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs1)) == 0) |
5e3c5739 | 2756 | break; |
2757 | ||
2758 | if (bs1 != HOST_BITS_PER_WIDE_INT | |
2759 | && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask) | |
b7d86581 | 2760 | abort (); |
5e3c5739 | 2761 | |
2762 | p = 63 - bs0; | |
2763 | len = bs1 - bs0; | |
2764 | ||
2765 | operands[2] = GEN_INT (p); | |
2766 | operands[3] = GEN_INT (len); | |
2767 | return "depdi -1,%2,%3,%0"; | |
2768 | } | |
e057641f | 2769 | \f |
58356836 | 2770 | /* Target hook for assembling integer objects. This code handles |
2771 | aligned SI and DI integers specially, since function references must | |
2772 | be preceded by P%. */ | |
2773 | ||
2774 | static bool | |
2775 | pa_assemble_integer (x, size, aligned_p) | |
2776 | rtx x; | |
2777 | unsigned int size; | |
2778 | int aligned_p; | |
2779 | { | |
2780 | if (size == UNITS_PER_WORD && aligned_p | |
2781 | && function_label_operand (x, VOIDmode)) | |
2782 | { | |
2783 | fputs (size == 8? "\t.dword\tP%" : "\t.word\tP%", asm_out_file); | |
2784 | output_addr_const (asm_out_file, x); | |
2785 | fputc ('\n', asm_out_file); | |
2786 | return true; | |
2787 | } | |
2788 | return default_assemble_integer (x, size, aligned_p); | |
2789 | } | |
2790 | \f | |
87ad11b0 | 2791 | /* Output an ascii string. */ |
57ed30e5 | 2792 | void |
87ad11b0 | 2793 | output_ascii (file, p, size) |
2794 | FILE *file; | |
feb9af9f | 2795 | const char *p; |
87ad11b0 | 2796 | int size; |
2797 | { | |
2798 | int i; | |
2799 | int chars_output; | |
2800 | unsigned char partial_output[16]; /* Max space 4 chars can occupy. */ | |
2801 | ||
2802 | /* The HP assembler can only take strings of 256 characters at one | |
2803 | time. This is a limitation on input line length, *not* the | |
2804 | length of the string. Sigh. Even worse, it seems that the | |
2805 | restriction is in number of input characters (see \xnn & | |
2806 | \whatever). So we have to do this very carefully. */ | |
2807 | ||
9c0ac0fd | 2808 | fputs ("\t.STRING \"", file); |
87ad11b0 | 2809 | |
2810 | chars_output = 0; | |
2811 | for (i = 0; i < size; i += 4) | |
2812 | { | |
2813 | int co = 0; | |
2814 | int io = 0; | |
2815 | for (io = 0, co = 0; io < MIN (4, size - i); io++) | |
2816 | { | |
bf8aac3e | 2817 | register unsigned int c = (unsigned char) p[i + io]; |
87ad11b0 | 2818 | |
2819 | if (c == '\"' || c == '\\') | |
2820 | partial_output[co++] = '\\'; | |
2821 | if (c >= ' ' && c < 0177) | |
2822 | partial_output[co++] = c; | |
2823 | else | |
2824 | { | |
2825 | unsigned int hexd; | |
2826 | partial_output[co++] = '\\'; | |
2827 | partial_output[co++] = 'x'; | |
2828 | hexd = c / 16 - 0 + '0'; | |
2829 | if (hexd > '9') | |
2830 | hexd -= '9' - 'a' + 1; | |
2831 | partial_output[co++] = hexd; | |
2832 | hexd = c % 16 - 0 + '0'; | |
2833 | if (hexd > '9') | |
2834 | hexd -= '9' - 'a' + 1; | |
2835 | partial_output[co++] = hexd; | |
2836 | } | |
2837 | } | |
2838 | if (chars_output + co > 243) | |
2839 | { | |
9c0ac0fd | 2840 | fputs ("\"\n\t.STRING \"", file); |
87ad11b0 | 2841 | chars_output = 0; |
2842 | } | |
a584fe8a | 2843 | fwrite (partial_output, 1, (size_t) co, file); |
87ad11b0 | 2844 | chars_output += co; |
2845 | co = 0; | |
2846 | } | |
9c0ac0fd | 2847 | fputs ("\"\n", file); |
87ad11b0 | 2848 | } |
c533da59 | 2849 | |
2850 | /* Try to rewrite floating point comparisons & branches to avoid | |
2851 | useless add,tr insns. | |
2852 | ||
2853 | CHECK_NOTES is nonzero if we should examine REG_DEAD notes | |
2854 | to see if FPCC is dead. CHECK_NOTES is nonzero for the | |
2855 | first attempt to remove useless add,tr insns. It is zero | |
2856 | for the second pass as reorg sometimes leaves bogus REG_DEAD | |
2857 | notes lying around. | |
2858 | ||
2859 | When CHECK_NOTES is zero we can only eliminate add,tr insns | |
2860 | when there's a 1:1 correspondence between fcmp and ftest/fbranch | |
2861 | instructions. */ | |
611a88e1 | 2862 | static void |
c533da59 | 2863 | remove_useless_addtr_insns (insns, check_notes) |
2864 | rtx insns; | |
2865 | int check_notes; | |
2866 | { | |
2867 | rtx insn; | |
c533da59 | 2868 | static int pass = 0; |
2869 | ||
2870 | /* This is fairly cheap, so always run it when optimizing. */ | |
2871 | if (optimize > 0) | |
2872 | { | |
2873 | int fcmp_count = 0; | |
2874 | int fbranch_count = 0; | |
2875 | ||
2876 | /* Walk all the insns in this function looking for fcmp & fbranch | |
2877 | instructions. Keep track of how many of each we find. */ | |
2878 | insns = get_insns (); | |
2879 | for (insn = insns; insn; insn = next_insn (insn)) | |
2880 | { | |
2881 | rtx tmp; | |
2882 | ||
2883 | /* Ignore anything that isn't an INSN or a JUMP_INSN. */ | |
2884 | if (GET_CODE (insn) != INSN && GET_CODE (insn) != JUMP_INSN) | |
2885 | continue; | |
2886 | ||
2887 | tmp = PATTERN (insn); | |
2888 | ||
2889 | /* It must be a set. */ | |
2890 | if (GET_CODE (tmp) != SET) | |
2891 | continue; | |
2892 | ||
2893 | /* If the destination is CCFP, then we've found an fcmp insn. */ | |
2894 | tmp = SET_DEST (tmp); | |
2895 | if (GET_CODE (tmp) == REG && REGNO (tmp) == 0) | |
2896 | { | |
2897 | fcmp_count++; | |
2898 | continue; | |
2899 | } | |
9840d99d | 2900 | |
c533da59 | 2901 | tmp = PATTERN (insn); |
2902 | /* If this is an fbranch instruction, bump the fbranch counter. */ | |
2903 | if (GET_CODE (tmp) == SET | |
2904 | && SET_DEST (tmp) == pc_rtx | |
2905 | && GET_CODE (SET_SRC (tmp)) == IF_THEN_ELSE | |
2906 | && GET_CODE (XEXP (SET_SRC (tmp), 0)) == NE | |
2907 | && GET_CODE (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == REG | |
2908 | && REGNO (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == 0) | |
2909 | { | |
2910 | fbranch_count++; | |
2911 | continue; | |
2912 | } | |
2913 | } | |
2914 | ||
2915 | ||
2916 | /* Find all floating point compare + branch insns. If possible, | |
2917 | reverse the comparison & the branch to avoid add,tr insns. */ | |
2918 | for (insn = insns; insn; insn = next_insn (insn)) | |
2919 | { | |
2920 | rtx tmp, next; | |
2921 | ||
2922 | /* Ignore anything that isn't an INSN. */ | |
2923 | if (GET_CODE (insn) != INSN) | |
2924 | continue; | |
2925 | ||
2926 | tmp = PATTERN (insn); | |
2927 | ||
2928 | /* It must be a set. */ | |
2929 | if (GET_CODE (tmp) != SET) | |
2930 | continue; | |
2931 | ||
2932 | /* The destination must be CCFP, which is register zero. */ | |
2933 | tmp = SET_DEST (tmp); | |
2934 | if (GET_CODE (tmp) != REG || REGNO (tmp) != 0) | |
2935 | continue; | |
2936 | ||
2937 | /* INSN should be a set of CCFP. | |
2938 | ||
2939 | See if the result of this insn is used in a reversed FP | |
2940 | conditional branch. If so, reverse our condition and | |
2941 | the branch. Doing so avoids useless add,tr insns. */ | |
2942 | next = next_insn (insn); | |
2943 | while (next) | |
2944 | { | |
2945 | /* Jumps, calls and labels stop our search. */ | |
2946 | if (GET_CODE (next) == JUMP_INSN | |
2947 | || GET_CODE (next) == CALL_INSN | |
2948 | || GET_CODE (next) == CODE_LABEL) | |
2949 | break; | |
2950 | ||
2951 | /* As does another fcmp insn. */ | |
2952 | if (GET_CODE (next) == INSN | |
2953 | && GET_CODE (PATTERN (next)) == SET | |
2954 | && GET_CODE (SET_DEST (PATTERN (next))) == REG | |
2955 | && REGNO (SET_DEST (PATTERN (next))) == 0) | |
2956 | break; | |
2957 | ||
2958 | next = next_insn (next); | |
2959 | } | |
2960 | ||
2961 | /* Is NEXT_INSN a branch? */ | |
2962 | if (next | |
2963 | && GET_CODE (next) == JUMP_INSN) | |
2964 | { | |
2965 | rtx pattern = PATTERN (next); | |
2966 | ||
2967 | /* If it a reversed fp conditional branch (eg uses add,tr) | |
2968 | and CCFP dies, then reverse our conditional and the branch | |
2969 | to avoid the add,tr. */ | |
2970 | if (GET_CODE (pattern) == SET | |
2971 | && SET_DEST (pattern) == pc_rtx | |
2972 | && GET_CODE (SET_SRC (pattern)) == IF_THEN_ELSE | |
2973 | && GET_CODE (XEXP (SET_SRC (pattern), 0)) == NE | |
2974 | && GET_CODE (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == REG | |
2975 | && REGNO (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == 0 | |
2976 | && GET_CODE (XEXP (SET_SRC (pattern), 1)) == PC | |
2977 | && (fcmp_count == fbranch_count | |
2978 | || (check_notes | |
2979 | && find_regno_note (next, REG_DEAD, 0)))) | |
2980 | { | |
2981 | /* Reverse the branch. */ | |
2982 | tmp = XEXP (SET_SRC (pattern), 1); | |
2983 | XEXP (SET_SRC (pattern), 1) = XEXP (SET_SRC (pattern), 2); | |
2984 | XEXP (SET_SRC (pattern), 2) = tmp; | |
2985 | INSN_CODE (next) = -1; | |
2986 | ||
2987 | /* Reverse our condition. */ | |
2988 | tmp = PATTERN (insn); | |
2989 | PUT_CODE (XEXP (tmp, 1), | |
ea52c577 | 2990 | (reverse_condition_maybe_unordered |
2991 | (GET_CODE (XEXP (tmp, 1))))); | |
c533da59 | 2992 | } |
2993 | } | |
2994 | } | |
2995 | } | |
2996 | ||
2997 | pass = !pass; | |
2998 | ||
2999 | } | |
87ad11b0 | 3000 | \f |
ea52c577 | 3001 | /* You may have trouble believing this, but this is the 32 bit HP-PA |
3002 | stack layout. Wow. | |
87ad11b0 | 3003 | |
3004 | Offset Contents | |
3005 | ||
3006 | Variable arguments (optional; any number may be allocated) | |
3007 | ||
3008 | SP-(4*(N+9)) arg word N | |
3009 | : : | |
3010 | SP-56 arg word 5 | |
3011 | SP-52 arg word 4 | |
3012 | ||
3013 | Fixed arguments (must be allocated; may remain unused) | |
3014 | ||
3015 | SP-48 arg word 3 | |
3016 | SP-44 arg word 2 | |
3017 | SP-40 arg word 1 | |
3018 | SP-36 arg word 0 | |
3019 | ||
3020 | Frame Marker | |
3021 | ||
3022 | SP-32 External Data Pointer (DP) | |
3023 | SP-28 External sr4 | |
3024 | SP-24 External/stub RP (RP') | |
3025 | SP-20 Current RP | |
3026 | SP-16 Static Link | |
3027 | SP-12 Clean up | |
3028 | SP-8 Calling Stub RP (RP'') | |
3029 | SP-4 Previous SP | |
3030 | ||
3031 | Top of Frame | |
3032 | ||
3033 | SP-0 Stack Pointer (points to next available address) | |
3034 | ||
3035 | */ | |
3036 | ||
3037 | /* This function saves registers as follows. Registers marked with ' are | |
3038 | this function's registers (as opposed to the previous function's). | |
3039 | If a frame_pointer isn't needed, r4 is saved as a general register; | |
3040 | the space for the frame pointer is still allocated, though, to keep | |
3041 | things simple. | |
3042 | ||
3043 | ||
3044 | Top of Frame | |
3045 | ||
3046 | SP (FP') Previous FP | |
3047 | SP + 4 Alignment filler (sigh) | |
3048 | SP + 8 Space for locals reserved here. | |
3049 | . | |
3050 | . | |
3051 | . | |
3052 | SP + n All call saved register used. | |
3053 | . | |
3054 | . | |
3055 | . | |
3056 | SP + o All call saved fp registers used. | |
3057 | . | |
3058 | . | |
3059 | . | |
3060 | SP + p (SP') points to next available address. | |
6d36483b | 3061 | |
87ad11b0 | 3062 | */ |
3063 | ||
17d9b0c3 | 3064 | /* Global variables set by output_function_prologue(). */ |
cc858176 | 3065 | /* Size of frame. Need to know this to emit return insns from |
3066 | leaf procedures. */ | |
3067 | static int actual_fsize; | |
3068 | static int local_fsize, save_fregs; | |
3069 | ||
daee63dd | 3070 | /* Emit RTL to store REG at the memory location specified by BASE+DISP. |
359a0be8 | 3071 | Handle case where DISP > 8k by using the add_high_const patterns. |
daee63dd | 3072 | |
3073 | Note in DISP > 8k case, we will leave the high part of the address | |
3074 | in %r1. There is code in expand_hppa_{prologue,epilogue} that knows this.*/ | |
7014838c | 3075 | |
6a2c16d6 | 3076 | static void |
daee63dd | 3077 | store_reg (reg, disp, base) |
3078 | int reg, disp, base; | |
87ad11b0 | 3079 | { |
6a2c16d6 | 3080 | rtx insn, dest, src, basereg; |
cc858176 | 3081 | |
3082 | src = gen_rtx_REG (word_mode, reg); | |
3083 | basereg = gen_rtx_REG (Pmode, base); | |
87ad11b0 | 3084 | if (VAL_14_BITS_P (disp)) |
daee63dd | 3085 | { |
cc858176 | 3086 | dest = gen_rtx_MEM (word_mode, plus_constant (basereg, disp)); |
6a2c16d6 | 3087 | insn = emit_move_insn (dest, src); |
daee63dd | 3088 | } |
daee63dd | 3089 | else |
3090 | { | |
cc858176 | 3091 | rtx delta = GEN_INT (disp); |
3092 | rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta)); | |
3093 | rtx tmpreg = gen_rtx_REG (Pmode, 1); | |
3094 | emit_move_insn (tmpreg, high); | |
3095 | dest = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta)); | |
6a2c16d6 | 3096 | insn = emit_move_insn (dest, src); |
3097 | if (DO_FRAME_NOTES) | |
3098 | { | |
3099 | REG_NOTES (insn) | |
3100 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3101 | gen_rtx_SET (VOIDmode, | |
3102 | gen_rtx_MEM (word_mode, | |
3103 | gen_rtx_PLUS (word_mode, basereg, | |
3104 | delta)), | |
3105 | src), | |
3106 | REG_NOTES (insn)); | |
3107 | } | |
daee63dd | 3108 | } |
6a2c16d6 | 3109 | |
3110 | if (DO_FRAME_NOTES) | |
3111 | RTX_FRAME_RELATED_P (insn) = 1; | |
daee63dd | 3112 | } |
3113 | ||
a584fe8a | 3114 | /* Emit RTL to store REG at the memory location specified by BASE and then |
3115 | add MOD to BASE. MOD must be <= 8k. */ | |
daee63dd | 3116 | |
a584fe8a | 3117 | static void |
3118 | store_reg_modify (base, reg, mod) | |
3119 | int base, reg, mod; | |
3120 | { | |
3121 | rtx insn, basereg, srcreg, delta; | |
3122 | ||
3123 | if (! VAL_14_BITS_P (mod)) | |
3124 | abort (); | |
3125 | ||
3126 | basereg = gen_rtx_REG (Pmode, base); | |
3127 | srcreg = gen_rtx_REG (word_mode, reg); | |
3128 | delta = GEN_INT (mod); | |
3129 | ||
3130 | insn = emit_insn (gen_post_store (basereg, srcreg, delta)); | |
3131 | if (DO_FRAME_NOTES) | |
3132 | { | |
3133 | RTX_FRAME_RELATED_P (insn) = 1; | |
3134 | ||
3135 | /* RTX_FRAME_RELATED_P must be set on each frame related set | |
3136 | in a parallel with more than one element. Don't set | |
3137 | RTX_FRAME_RELATED_P in the first set if reg is temporary | |
3138 | register 1. The effect of this operation is recorded in | |
3139 | the initial copy. */ | |
3140 | if (reg != 1) | |
3141 | { | |
3142 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1; | |
3143 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
3144 | } | |
3145 | else | |
3146 | { | |
3147 | /* The first element of a PARALLEL is always processed if it is | |
3148 | a SET. Thus, we need an expression list for this case. */ | |
3149 | REG_NOTES (insn) | |
3150 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3151 | gen_rtx_SET (VOIDmode, basereg, | |
3152 | gen_rtx_PLUS (word_mode, basereg, delta)), | |
3153 | REG_NOTES (insn)); | |
3154 | } | |
3155 | } | |
3156 | } | |
3157 | ||
3158 | /* Emit RTL to set REG to the value specified by BASE+DISP. Handle case | |
3159 | where DISP > 8k by using the add_high_const patterns. NOTE indicates | |
3160 | whether to add a frame note or not. | |
3161 | ||
3162 | In the DISP > 8k case, we leave the high part of the address in %r1. | |
3163 | There is code in expand_hppa_{prologue,epilogue} that knows about this. */ | |
7014838c | 3164 | |
6a2c16d6 | 3165 | static void |
a584fe8a | 3166 | set_reg_plus_d (reg, base, disp, note) |
3167 | int reg, base, disp, note; | |
87ad11b0 | 3168 | { |
6a2c16d6 | 3169 | rtx insn; |
cc858176 | 3170 | |
87ad11b0 | 3171 | if (VAL_14_BITS_P (disp)) |
cc858176 | 3172 | { |
6a2c16d6 | 3173 | insn = emit_move_insn (gen_rtx_REG (Pmode, reg), |
3174 | plus_constant (gen_rtx_REG (Pmode, base), disp)); | |
cc858176 | 3175 | } |
87ad11b0 | 3176 | else |
daee63dd | 3177 | { |
6a2c16d6 | 3178 | rtx basereg = gen_rtx_REG (Pmode, base); |
cc858176 | 3179 | rtx delta = GEN_INT (disp); |
6a2c16d6 | 3180 | |
359a0be8 | 3181 | emit_move_insn (gen_rtx_REG (Pmode, 1), |
6a2c16d6 | 3182 | gen_rtx_PLUS (Pmode, basereg, |
cc858176 | 3183 | gen_rtx_HIGH (Pmode, delta))); |
6a2c16d6 | 3184 | insn = emit_move_insn (gen_rtx_REG (Pmode, reg), |
3185 | gen_rtx_LO_SUM (Pmode, gen_rtx_REG (Pmode, 1), | |
3186 | delta)); | |
daee63dd | 3187 | } |
6a2c16d6 | 3188 | |
a584fe8a | 3189 | if (DO_FRAME_NOTES && note) |
6a2c16d6 | 3190 | RTX_FRAME_RELATED_P (insn) = 1; |
87ad11b0 | 3191 | } |
3192 | ||
3193 | int | |
a1ab4fa3 | 3194 | compute_frame_size (size, fregs_live) |
87ad11b0 | 3195 | int size; |
3ddcbb9d | 3196 | int *fregs_live; |
87ad11b0 | 3197 | { |
256f9b65 | 3198 | int freg_saved = 0; |
3199 | int i, j; | |
3200 | ||
3201 | /* The code in hppa_expand_prologue and hppa_expand_epilogue must | |
3202 | be consistent with the rounding and size calculation done here. | |
3203 | Change them at the same time. */ | |
3204 | ||
3205 | /* We do our own stack alignment. First, round the size of the | |
3206 | stack locals up to a word boundary. */ | |
3207 | size = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1); | |
3208 | ||
3209 | /* Space for previous frame pointer + filler. If any frame is | |
3210 | allocated, we need to add in the STARTING_FRAME_OFFSET. We | |
3211 | waste some space here for the sake of HP compatibility. The | |
3212 | first slot is only used when the frame pointer is needed. */ | |
3213 | if (size || frame_pointer_needed) | |
3214 | size += STARTING_FRAME_OFFSET; | |
3215 | ||
a584fe8a | 3216 | /* If the current function calls __builtin_eh_return, then we need |
3217 | to allocate stack space for registers that will hold data for | |
3218 | the exception handler. */ | |
3219 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3220 | { | |
3221 | unsigned int i; | |
3222 | ||
3223 | for (i = 0; EH_RETURN_DATA_REGNO (i) != INVALID_REGNUM; ++i) | |
3224 | continue; | |
256f9b65 | 3225 | size += i * UNITS_PER_WORD; |
a584fe8a | 3226 | } |
3227 | ||
3a51bad9 | 3228 | /* Account for space used by the callee general register saves. */ |
256f9b65 | 3229 | for (i = 18, j = frame_pointer_needed ? 4 : 3; i >= j; i--) |
7f7c4869 | 3230 | if (regs_ever_live[i]) |
256f9b65 | 3231 | size += UNITS_PER_WORD; |
df0651dc | 3232 | |
3a51bad9 | 3233 | /* Account for space used by the callee floating point register saves. */ |
bac38c40 | 3234 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
5e3c5739 | 3235 | if (regs_ever_live[i] |
256f9b65 | 3236 | || (!TARGET_64BIT && regs_ever_live[i + 1])) |
df0651dc | 3237 | { |
256f9b65 | 3238 | freg_saved = 1; |
002fc5f7 | 3239 | |
3a51bad9 | 3240 | /* We always save both halves of the FP register, so always |
3241 | increment the frame size by 8 bytes. */ | |
256f9b65 | 3242 | size += 8; |
df0651dc | 3243 | } |
3244 | ||
256f9b65 | 3245 | /* If any of the floating registers are saved, account for the |
3246 | alignment needed for the floating point register save block. */ | |
3247 | if (freg_saved) | |
3248 | { | |
3249 | size = (size + 7) & ~7; | |
3250 | if (fregs_live) | |
3251 | *fregs_live = 1; | |
3252 | } | |
3253 | ||
3a51bad9 | 3254 | /* The various ABIs include space for the outgoing parameters in the |
256f9b65 | 3255 | size of the current function's stack frame. We don't need to align |
3256 | for the outgoing arguments as their alignment is set by the final | |
3257 | rounding for the frame as a whole. */ | |
3258 | size += current_function_outgoing_args_size; | |
3a51bad9 | 3259 | |
3260 | /* Allocate space for the fixed frame marker. This space must be | |
3261 | allocated for any function that makes calls or otherwise allocates | |
3262 | stack space. */ | |
256f9b65 | 3263 | if (!current_function_is_leaf || size) |
3264 | size += TARGET_64BIT ? 16 : 32; | |
5e3c5739 | 3265 | |
256f9b65 | 3266 | /* Finally, round to the preferred stack boundary. */ |
3267 | return ((size + PREFERRED_STACK_BOUNDARY / 8 - 1) | |
81d1e315 | 3268 | & ~(PREFERRED_STACK_BOUNDARY / 8 - 1)); |
87ad11b0 | 3269 | } |
6d36483b | 3270 | |
17d9b0c3 | 3271 | /* Generate the assembly code for function entry. FILE is a stdio |
3272 | stream to output the code to. SIZE is an int: how many units of | |
3273 | temporary storage to allocate. | |
3274 | ||
3275 | Refer to the array `regs_ever_live' to determine which registers to | |
3276 | save; `regs_ever_live[I]' is nonzero if register number I is ever | |
3277 | used in the function. This function is responsible for knowing | |
3278 | which registers should not be saved even if used. */ | |
3279 | ||
3280 | /* On HP-PA, move-double insns between fpu and cpu need an 8-byte block | |
3281 | of memory. If any fpu reg is used in the function, we allocate | |
3282 | such a block here, at the bottom of the frame, just in case it's needed. | |
3283 | ||
3284 | If this function is a leaf procedure, then we may choose not | |
3285 | to do a "save" insn. The decision about whether or not | |
3286 | to do this is made in regclass.c. */ | |
3287 | ||
6988553d | 3288 | static void |
17d9b0c3 | 3289 | pa_output_function_prologue (file, size) |
87ad11b0 | 3290 | FILE *file; |
17d9b0c3 | 3291 | HOST_WIDE_INT size ATTRIBUTE_UNUSED; |
87ad11b0 | 3292 | { |
d151162a | 3293 | /* The function's label and associated .PROC must never be |
3294 | separated and must be output *after* any profiling declarations | |
3295 | to avoid changing spaces/subspaces within a procedure. */ | |
3296 | ASM_OUTPUT_LABEL (file, XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0)); | |
3297 | fputs ("\t.PROC\n", file); | |
3298 | ||
daee63dd | 3299 | /* hppa_expand_prologue does the dirty work now. We just need |
3300 | to output the assembler directives which denote the start | |
3301 | of a function. */ | |
2acd4f33 | 3302 | fprintf (file, "\t.CALLINFO FRAME=%d", actual_fsize); |
df6edefa | 3303 | if (regs_ever_live[2]) |
9c0ac0fd | 3304 | fputs (",CALLS,SAVE_RP", file); |
daee63dd | 3305 | else |
9c0ac0fd | 3306 | fputs (",NO_CALLS", file); |
f3ba7709 | 3307 | |
3308 | if (frame_pointer_needed) | |
9c0ac0fd | 3309 | fputs (",SAVE_SP", file); |
f3ba7709 | 3310 | |
a9960cdc | 3311 | /* Pass on information about the number of callee register saves |
9b0c95be | 3312 | performed in the prologue. |
3313 | ||
3314 | The compiler is supposed to pass the highest register number | |
6d36483b | 3315 | saved, the assembler then has to adjust that number before |
9b0c95be | 3316 | entering it into the unwind descriptor (to account for any |
6d36483b | 3317 | caller saved registers with lower register numbers than the |
9b0c95be | 3318 | first callee saved register). */ |
3319 | if (gr_saved) | |
3320 | fprintf (file, ",ENTRY_GR=%d", gr_saved + 2); | |
3321 | ||
3322 | if (fr_saved) | |
3323 | fprintf (file, ",ENTRY_FR=%d", fr_saved + 11); | |
a9960cdc | 3324 | |
9c0ac0fd | 3325 | fputs ("\n\t.ENTRY\n", file); |
daee63dd | 3326 | |
c533da59 | 3327 | remove_useless_addtr_insns (get_insns (), 0); |
daee63dd | 3328 | } |
3329 | ||
57ed30e5 | 3330 | void |
cc858176 | 3331 | hppa_expand_prologue () |
daee63dd | 3332 | { |
afd7b680 | 3333 | int merge_sp_adjust_with_store = 0; |
256f9b65 | 3334 | int size = get_frame_size (); |
daee63dd | 3335 | int i, offset; |
a584fe8a | 3336 | rtx insn, tmpreg; |
daee63dd | 3337 | |
a9960cdc | 3338 | gr_saved = 0; |
3339 | fr_saved = 0; | |
3ddcbb9d | 3340 | save_fregs = 0; |
3a51bad9 | 3341 | |
256f9b65 | 3342 | /* Compute total size for frame pointer, filler, locals and rounding to |
3343 | the next word boundary. Similar code appears in compute_frame_size | |
3344 | and must be changed in tandem with this code. */ | |
3345 | local_fsize = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1); | |
3346 | if (local_fsize || frame_pointer_needed) | |
3347 | local_fsize += STARTING_FRAME_OFFSET; | |
3a51bad9 | 3348 | |
a1ab4fa3 | 3349 | actual_fsize = compute_frame_size (size, &save_fregs); |
87ad11b0 | 3350 | |
daee63dd | 3351 | /* Compute a few things we will use often. */ |
440c23df | 3352 | tmpreg = gen_rtx_REG (word_mode, 1); |
87ad11b0 | 3353 | |
6d36483b | 3354 | /* Save RP first. The calling conventions manual states RP will |
cc858176 | 3355 | always be stored into the caller's frame at sp - 20 or sp - 16 |
5e3c5739 | 3356 | depending on which ABI is in use. */ |
a584fe8a | 3357 | if (regs_ever_live[2] || current_function_calls_eh_return) |
6a2c16d6 | 3358 | store_reg (2, TARGET_64BIT ? -16 : -20, STACK_POINTER_REGNUM); |
6d36483b | 3359 | |
daee63dd | 3360 | /* Allocate the local frame and set up the frame pointer if needed. */ |
58361f39 | 3361 | if (actual_fsize != 0) |
3362 | { | |
3363 | if (frame_pointer_needed) | |
3364 | { | |
3365 | /* Copy the old frame pointer temporarily into %r1. Set up the | |
3366 | new stack pointer, then store away the saved old frame pointer | |
a584fe8a | 3367 | into the stack at sp and at the same time update the stack |
3368 | pointer by actual_fsize bytes. Two versions, first | |
58361f39 | 3369 | handles small (<8k) frames. The second handles large (>=8k) |
3370 | frames. */ | |
a584fe8a | 3371 | insn = emit_move_insn (tmpreg, frame_pointer_rtx); |
3372 | if (DO_FRAME_NOTES) | |
cc858176 | 3373 | { |
a584fe8a | 3374 | /* We need to record the frame pointer save here since the |
3375 | new frame pointer is set in the following insn. */ | |
3376 | RTX_FRAME_RELATED_P (insn) = 1; | |
3377 | REG_NOTES (insn) | |
3378 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3379 | gen_rtx_SET (VOIDmode, | |
3380 | gen_rtx_MEM (word_mode, stack_pointer_rtx), | |
3381 | frame_pointer_rtx), | |
3382 | REG_NOTES (insn)); | |
cc858176 | 3383 | } |
a584fe8a | 3384 | |
3385 | insn = emit_move_insn (frame_pointer_rtx, stack_pointer_rtx); | |
3386 | if (DO_FRAME_NOTES) | |
3387 | RTX_FRAME_RELATED_P (insn) = 1; | |
3388 | ||
3389 | if (VAL_14_BITS_P (actual_fsize)) | |
3390 | store_reg_modify (STACK_POINTER_REGNUM, 1, actual_fsize); | |
58361f39 | 3391 | else |
3392 | { | |
3393 | /* It is incorrect to store the saved frame pointer at *sp, | |
3394 | then increment sp (writes beyond the current stack boundary). | |
3395 | ||
3396 | So instead use stwm to store at *sp and post-increment the | |
3397 | stack pointer as an atomic operation. Then increment sp to | |
3398 | finish allocating the new frame. */ | |
3399 | int adjust1 = 8192 - 64; | |
3400 | int adjust2 = actual_fsize - adjust1; | |
cc858176 | 3401 | |
a584fe8a | 3402 | store_reg_modify (STACK_POINTER_REGNUM, 1, adjust1); |
6a2c16d6 | 3403 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3404 | adjust2, 1); |
58361f39 | 3405 | } |
a584fe8a | 3406 | |
58361f39 | 3407 | /* Prevent register spills from being scheduled before the |
3408 | stack pointer is raised. Necessary as we will be storing | |
3409 | registers using the frame pointer as a base register, and | |
9840d99d | 3410 | we happen to set fp before raising sp. */ |
58361f39 | 3411 | emit_insn (gen_blockage ()); |
3412 | } | |
3413 | /* no frame pointer needed. */ | |
3414 | else | |
3415 | { | |
3416 | /* In some cases we can perform the first callee register save | |
3417 | and allocating the stack frame at the same time. If so, just | |
3418 | make a note of it and defer allocating the frame until saving | |
3419 | the callee registers. */ | |
df6edefa | 3420 | if (VAL_14_BITS_P (actual_fsize) && local_fsize == 0) |
58361f39 | 3421 | merge_sp_adjust_with_store = 1; |
3422 | /* Can not optimize. Adjust the stack frame by actual_fsize | |
3423 | bytes. */ | |
3424 | else | |
6a2c16d6 | 3425 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3426 | actual_fsize, 1); |
58361f39 | 3427 | } |
372ef038 | 3428 | } |
3429 | ||
6d36483b | 3430 | /* Normal register save. |
daee63dd | 3431 | |
3432 | Do not save the frame pointer in the frame_pointer_needed case. It | |
3433 | was done earlier. */ | |
87ad11b0 | 3434 | if (frame_pointer_needed) |
3435 | { | |
a584fe8a | 3436 | offset = local_fsize; |
3437 | ||
3438 | /* Saving the EH return data registers in the frame is the simplest | |
3439 | way to get the frame unwind information emitted. We put them | |
3440 | just before the general registers. */ | |
3441 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3442 | { | |
3443 | unsigned int i, regno; | |
3444 | ||
3445 | for (i = 0; ; ++i) | |
3446 | { | |
3447 | regno = EH_RETURN_DATA_REGNO (i); | |
3448 | if (regno == INVALID_REGNUM) | |
3449 | break; | |
3450 | ||
3451 | store_reg (regno, offset, FRAME_POINTER_REGNUM); | |
3452 | offset += UNITS_PER_WORD; | |
3453 | } | |
3454 | } | |
3455 | ||
3456 | for (i = 18; i >= 4; i--) | |
98328a39 | 3457 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3458 | { |
6a2c16d6 | 3459 | store_reg (i, offset, FRAME_POINTER_REGNUM); |
6ec4380b | 3460 | offset += UNITS_PER_WORD; |
a9960cdc | 3461 | gr_saved++; |
87ad11b0 | 3462 | } |
7f7c4869 | 3463 | /* Account for %r3 which is saved in a special place. */ |
9b0c95be | 3464 | gr_saved++; |
87ad11b0 | 3465 | } |
daee63dd | 3466 | /* No frame pointer needed. */ |
87ad11b0 | 3467 | else |
3468 | { | |
a584fe8a | 3469 | offset = local_fsize - actual_fsize; |
3470 | ||
3471 | /* Saving the EH return data registers in the frame is the simplest | |
3472 | way to get the frame unwind information emitted. */ | |
3473 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3474 | { | |
3475 | unsigned int i, regno; | |
3476 | ||
3477 | for (i = 0; ; ++i) | |
3478 | { | |
3479 | regno = EH_RETURN_DATA_REGNO (i); | |
3480 | if (regno == INVALID_REGNUM) | |
3481 | break; | |
3482 | ||
3483 | /* If merge_sp_adjust_with_store is nonzero, then we can | |
3484 | optimize the first save. */ | |
3485 | if (merge_sp_adjust_with_store) | |
3486 | { | |
3487 | store_reg_modify (STACK_POINTER_REGNUM, regno, -offset); | |
3488 | merge_sp_adjust_with_store = 0; | |
3489 | } | |
3490 | else | |
3491 | store_reg (regno, offset, STACK_POINTER_REGNUM); | |
3492 | offset += UNITS_PER_WORD; | |
3493 | } | |
3494 | } | |
3495 | ||
3496 | for (i = 18; i >= 3; i--) | |
98328a39 | 3497 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3498 | { |
6d36483b | 3499 | /* If merge_sp_adjust_with_store is nonzero, then we can |
afd7b680 | 3500 | optimize the first GR save. */ |
201f01e9 | 3501 | if (merge_sp_adjust_with_store) |
afd7b680 | 3502 | { |
a584fe8a | 3503 | store_reg_modify (STACK_POINTER_REGNUM, i, -offset); |
afd7b680 | 3504 | merge_sp_adjust_with_store = 0; |
afd7b680 | 3505 | } |
3506 | else | |
6a2c16d6 | 3507 | store_reg (i, offset, STACK_POINTER_REGNUM); |
6ec4380b | 3508 | offset += UNITS_PER_WORD; |
a9960cdc | 3509 | gr_saved++; |
87ad11b0 | 3510 | } |
daee63dd | 3511 | |
afd7b680 | 3512 | /* If we wanted to merge the SP adjustment with a GR save, but we never |
daee63dd | 3513 | did any GR saves, then just emit the adjustment here. */ |
201f01e9 | 3514 | if (merge_sp_adjust_with_store) |
6a2c16d6 | 3515 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
a584fe8a | 3516 | actual_fsize, 1); |
87ad11b0 | 3517 | } |
6d36483b | 3518 | |
df6edefa | 3519 | /* The hppa calling conventions say that %r19, the pic offset |
3520 | register, is saved at sp - 32 (in this function's frame) | |
3521 | when generating PIC code. FIXME: What is the correct thing | |
3522 | to do for functions which make no calls and allocate no | |
3523 | frame? Do we need to allocate a frame, or can we just omit | |
3524 | the save? For now we'll just omit the save. */ | |
3525 | if (flag_pic && actual_fsize != 0 && !TARGET_64BIT) | |
3526 | store_reg (PIC_OFFSET_TABLE_REGNUM, -32, STACK_POINTER_REGNUM); | |
3527 | ||
87ad11b0 | 3528 | /* Align pointer properly (doubleword boundary). */ |
3529 | offset = (offset + 7) & ~7; | |
3530 | ||
3531 | /* Floating point register store. */ | |
3532 | if (save_fregs) | |
87ad11b0 | 3533 | { |
a584fe8a | 3534 | rtx base; |
3535 | ||
daee63dd | 3536 | /* First get the frame or stack pointer to the start of the FP register |
3537 | save area. */ | |
a1ab4fa3 | 3538 | if (frame_pointer_needed) |
a584fe8a | 3539 | { |
3540 | set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0); | |
3541 | base = frame_pointer_rtx; | |
3542 | } | |
a1ab4fa3 | 3543 | else |
a584fe8a | 3544 | { |
3545 | set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0); | |
3546 | base = stack_pointer_rtx; | |
3547 | } | |
daee63dd | 3548 | |
3549 | /* Now actually save the FP registers. */ | |
bac38c40 | 3550 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
7f7c4869 | 3551 | { |
5e3c5739 | 3552 | if (regs_ever_live[i] |
3553 | || (! TARGET_64BIT && regs_ever_live[i + 1])) | |
7f7c4869 | 3554 | { |
6a2c16d6 | 3555 | rtx addr, insn, reg; |
cc858176 | 3556 | addr = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg)); |
3557 | reg = gen_rtx_REG (DFmode, i); | |
6a2c16d6 | 3558 | insn = emit_move_insn (addr, reg); |
3559 | if (DO_FRAME_NOTES) | |
3560 | { | |
3561 | RTX_FRAME_RELATED_P (insn) = 1; | |
a584fe8a | 3562 | if (TARGET_64BIT) |
3563 | { | |
3564 | rtx mem = gen_rtx_MEM (DFmode, | |
3565 | plus_constant (base, offset)); | |
3566 | REG_NOTES (insn) | |
3567 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3568 | gen_rtx_SET (VOIDmode, mem, reg), | |
3569 | REG_NOTES (insn)); | |
3570 | } | |
3571 | else | |
3572 | { | |
3573 | rtx meml = gen_rtx_MEM (SFmode, | |
3574 | plus_constant (base, offset)); | |
3575 | rtx memr = gen_rtx_MEM (SFmode, | |
3576 | plus_constant (base, offset + 4)); | |
3577 | rtx regl = gen_rtx_REG (SFmode, i); | |
3578 | rtx regr = gen_rtx_REG (SFmode, i + 1); | |
3579 | rtx setl = gen_rtx_SET (VOIDmode, meml, regl); | |
3580 | rtx setr = gen_rtx_SET (VOIDmode, memr, regr); | |
3581 | rtvec vec; | |
3582 | ||
3583 | RTX_FRAME_RELATED_P (setl) = 1; | |
3584 | RTX_FRAME_RELATED_P (setr) = 1; | |
3585 | vec = gen_rtvec (2, setl, setr); | |
3586 | REG_NOTES (insn) | |
3587 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
3588 | gen_rtx_SEQUENCE (VOIDmode, vec), | |
3589 | REG_NOTES (insn)); | |
3590 | } | |
6a2c16d6 | 3591 | } |
3592 | offset += GET_MODE_SIZE (DFmode); | |
7f7c4869 | 3593 | fr_saved++; |
3594 | } | |
3595 | } | |
87ad11b0 | 3596 | } |
a584fe8a | 3597 | |
3598 | /* FIXME: expand_call and expand_millicode_call need to be fixed to | |
3599 | prevent insns with frame notes being scheduled in the delay slot | |
3600 | of calls. This causes problems because the dwarf2 output code | |
3601 | processes the insn list serially. For now, limit the migration | |
3602 | of prologue insns with a blockage. */ | |
3603 | if (DO_FRAME_NOTES) | |
3604 | emit_insn (gen_blockage ()); | |
87ad11b0 | 3605 | } |
3606 | ||
cc858176 | 3607 | /* Emit RTL to load REG from the memory location specified by BASE+DISP. |
3608 | Handle case where DISP > 8k by using the add_high_const patterns. */ | |
3609 | ||
6a2c16d6 | 3610 | static void |
cc858176 | 3611 | load_reg (reg, disp, base) |
3612 | int reg, disp, base; | |
3613 | { | |
6a2c16d6 | 3614 | rtx src, dest, basereg; |
cc858176 | 3615 | |
3616 | dest = gen_rtx_REG (word_mode, reg); | |
3617 | basereg = gen_rtx_REG (Pmode, base); | |
3618 | if (VAL_14_BITS_P (disp)) | |
3619 | { | |
3620 | src = gen_rtx_MEM (word_mode, plus_constant (basereg, disp)); | |
6a2c16d6 | 3621 | emit_move_insn (dest, src); |
cc858176 | 3622 | } |
3623 | else | |
3624 | { | |
3625 | rtx delta = GEN_INT (disp); | |
3626 | rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta)); | |
3627 | rtx tmpreg = gen_rtx_REG (Pmode, 1); | |
3628 | emit_move_insn (tmpreg, high); | |
3629 | src = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta)); | |
6a2c16d6 | 3630 | emit_move_insn (dest, src); |
cc858176 | 3631 | } |
cc858176 | 3632 | } |
daee63dd | 3633 | |
17d9b0c3 | 3634 | /* This function generates the assembly code for function exit. |
3635 | Args are as for output_function_prologue (). | |
3636 | ||
3637 | The function epilogue should not depend on the current stack | |
3638 | pointer! It should use the frame pointer only. This is mandatory | |
3639 | because of alloca; we also take advantage of it to omit stack | |
6dc3b0d9 | 3640 | adjustments before returning. */ |
17d9b0c3 | 3641 | |
3642 | static void | |
3643 | pa_output_function_epilogue (file, size) | |
87ad11b0 | 3644 | FILE *file; |
17d9b0c3 | 3645 | HOST_WIDE_INT size ATTRIBUTE_UNUSED; |
87ad11b0 | 3646 | { |
90c41894 | 3647 | int last_address = 0; |
3695c664 | 3648 | rtx insn = get_last_insn (); |
3649 | ||
daee63dd | 3650 | /* hppa_expand_epilogue does the dirty work now. We just need |
3651 | to output the assembler directives which denote the end | |
3695c664 | 3652 | of a function. |
3653 | ||
3654 | To make debuggers happy, emit a nop if the epilogue was completely | |
3655 | eliminated due to a volatile call as the last insn in the | |
6d36483b | 3656 | current function. That way the return address (in %r2) will |
3695c664 | 3657 | always point to a valid instruction in the current function. */ |
3658 | ||
3659 | /* Get the last real insn. */ | |
3660 | if (GET_CODE (insn) == NOTE) | |
3661 | insn = prev_real_insn (insn); | |
3662 | ||
3663 | /* If it is a sequence, then look inside. */ | |
3664 | if (insn && GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) | |
3665 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
3666 | ||
6d36483b | 3667 | /* If insn is a CALL_INSN, then it must be a call to a volatile |
3695c664 | 3668 | function (otherwise there would be epilogue insns). */ |
3669 | if (insn && GET_CODE (insn) == CALL_INSN) | |
90c41894 | 3670 | { |
3671 | fputs ("\tnop\n", file); | |
3672 | last_address += 4; | |
3673 | } | |
6d36483b | 3674 | |
9c0ac0fd | 3675 | fputs ("\t.EXIT\n\t.PROCEND\n", file); |
90c41894 | 3676 | |
3677 | /* Finally, update the total number of code bytes output so far. */ | |
3678 | if ((TARGET_PORTABLE_RUNTIME || !TARGET_GAS || !TARGET_SOM) | |
3679 | && !flag_function_sections) | |
3680 | { | |
3681 | if (INSN_ADDRESSES_SET_P ()) | |
3682 | { | |
3683 | unsigned long old_total = total_code_bytes; | |
3684 | ||
3685 | insn = get_last_nonnote_insn (); | |
3686 | last_address += INSN_ADDRESSES (INSN_UID (insn)); | |
3687 | if (INSN_P (insn)) | |
3688 | last_address += insn_default_length (insn); | |
3689 | ||
3690 | total_code_bytes += last_address; | |
3691 | total_code_bytes += FUNCTION_BOUNDARY / BITS_PER_UNIT; | |
3692 | ||
3693 | /* Be prepared to handle overflows. */ | |
3694 | if (old_total > total_code_bytes) | |
3695 | total_code_bytes = -1; | |
3696 | } | |
3697 | else | |
3698 | total_code_bytes = -1; | |
3699 | } | |
daee63dd | 3700 | } |
afd7b680 | 3701 | |
daee63dd | 3702 | void |
3695c664 | 3703 | hppa_expand_epilogue () |
daee63dd | 3704 | { |
6d36483b | 3705 | rtx tmpreg; |
58361f39 | 3706 | int offset, i; |
3707 | int merge_sp_adjust_with_load = 0; | |
3708 | int ret_off = 0; | |
daee63dd | 3709 | |
3710 | /* We will use this often. */ | |
440c23df | 3711 | tmpreg = gen_rtx_REG (word_mode, 1); |
daee63dd | 3712 | |
3713 | /* Try to restore RP early to avoid load/use interlocks when | |
3714 | RP gets used in the return (bv) instruction. This appears to still | |
6dc3b0d9 | 3715 | be necessary even when we schedule the prologue and epilogue. */ |
a584fe8a | 3716 | if (regs_ever_live [2] || current_function_calls_eh_return) |
58361f39 | 3717 | { |
3718 | ret_off = TARGET_64BIT ? -16 : -20; | |
3719 | if (frame_pointer_needed) | |
3720 | { | |
6a2c16d6 | 3721 | load_reg (2, ret_off, FRAME_POINTER_REGNUM); |
58361f39 | 3722 | ret_off = 0; |
3723 | } | |
3724 | else | |
3725 | { | |
3726 | /* No frame pointer, and stack is smaller than 8k. */ | |
3727 | if (VAL_14_BITS_P (ret_off - actual_fsize)) | |
3728 | { | |
6a2c16d6 | 3729 | load_reg (2, ret_off - actual_fsize, STACK_POINTER_REGNUM); |
58361f39 | 3730 | ret_off = 0; |
3731 | } | |
3732 | } | |
3733 | } | |
daee63dd | 3734 | |
3735 | /* General register restores. */ | |
87ad11b0 | 3736 | if (frame_pointer_needed) |
3737 | { | |
a584fe8a | 3738 | offset = local_fsize; |
3739 | ||
3740 | /* If the current function calls __builtin_eh_return, then we need | |
3741 | to restore the saved EH data registers. */ | |
3742 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3743 | { | |
3744 | unsigned int i, regno; | |
3745 | ||
3746 | for (i = 0; ; ++i) | |
3747 | { | |
3748 | regno = EH_RETURN_DATA_REGNO (i); | |
3749 | if (regno == INVALID_REGNUM) | |
3750 | break; | |
3751 | ||
3752 | load_reg (regno, offset, FRAME_POINTER_REGNUM); | |
3753 | offset += UNITS_PER_WORD; | |
3754 | } | |
3755 | } | |
3756 | ||
3757 | for (i = 18; i >= 4; i--) | |
98328a39 | 3758 | if (regs_ever_live[i] && ! call_used_regs[i]) |
87ad11b0 | 3759 | { |
6a2c16d6 | 3760 | load_reg (i, offset, FRAME_POINTER_REGNUM); |
6ec4380b | 3761 | offset += UNITS_PER_WORD; |
87ad11b0 | 3762 | } |
87ad11b0 | 3763 | } |
3764 | else | |
3765 | { | |
a584fe8a | 3766 | offset = local_fsize - actual_fsize; |
3767 | ||
3768 | /* If the current function calls __builtin_eh_return, then we need | |
3769 | to restore the saved EH data registers. */ | |
3770 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3771 | { | |
3772 | unsigned int i, regno; | |
3773 | ||
3774 | for (i = 0; ; ++i) | |
3775 | { | |
3776 | regno = EH_RETURN_DATA_REGNO (i); | |
3777 | if (regno == INVALID_REGNUM) | |
3778 | break; | |
3779 | ||
3780 | /* Only for the first load. | |
3781 | merge_sp_adjust_with_load holds the register load | |
3782 | with which we will merge the sp adjustment. */ | |
3783 | if (merge_sp_adjust_with_load == 0 | |
3784 | && local_fsize == 0 | |
3785 | && VAL_14_BITS_P (-actual_fsize)) | |
3786 | merge_sp_adjust_with_load = regno; | |
3787 | else | |
3788 | load_reg (regno, offset, STACK_POINTER_REGNUM); | |
3789 | offset += UNITS_PER_WORD; | |
3790 | } | |
3791 | } | |
3792 | ||
3793 | for (i = 18; i >= 3; i--) | |
7f7c4869 | 3794 | { |
98328a39 | 3795 | if (regs_ever_live[i] && ! call_used_regs[i]) |
7f7c4869 | 3796 | { |
7f7c4869 | 3797 | /* Only for the first load. |
3798 | merge_sp_adjust_with_load holds the register load | |
3799 | with which we will merge the sp adjustment. */ | |
58361f39 | 3800 | if (merge_sp_adjust_with_load == 0 |
7f7c4869 | 3801 | && local_fsize == 0 |
58361f39 | 3802 | && VAL_14_BITS_P (-actual_fsize)) |
7f7c4869 | 3803 | merge_sp_adjust_with_load = i; |
3804 | else | |
6a2c16d6 | 3805 | load_reg (i, offset, STACK_POINTER_REGNUM); |
6ec4380b | 3806 | offset += UNITS_PER_WORD; |
7f7c4869 | 3807 | } |
3808 | } | |
87ad11b0 | 3809 | } |
daee63dd | 3810 | |
87ad11b0 | 3811 | /* Align pointer properly (doubleword boundary). */ |
3812 | offset = (offset + 7) & ~7; | |
3813 | ||
daee63dd | 3814 | /* FP register restores. */ |
87ad11b0 | 3815 | if (save_fregs) |
87ad11b0 | 3816 | { |
daee63dd | 3817 | /* Adjust the register to index off of. */ |
a1ab4fa3 | 3818 | if (frame_pointer_needed) |
a584fe8a | 3819 | set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0); |
a1ab4fa3 | 3820 | else |
a584fe8a | 3821 | set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0); |
daee63dd | 3822 | |
3823 | /* Actually do the restores now. */ | |
bac38c40 | 3824 | for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP) |
cc858176 | 3825 | if (regs_ever_live[i] |
3826 | || (! TARGET_64BIT && regs_ever_live[i + 1])) | |
3827 | { | |
3828 | rtx src = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg)); | |
3829 | rtx dest = gen_rtx_REG (DFmode, i); | |
6a2c16d6 | 3830 | emit_move_insn (dest, src); |
cc858176 | 3831 | } |
87ad11b0 | 3832 | } |
daee63dd | 3833 | |
14660146 | 3834 | /* Emit a blockage insn here to keep these insns from being moved to |
3835 | an earlier spot in the epilogue, or into the main instruction stream. | |
3836 | ||
3837 | This is necessary as we must not cut the stack back before all the | |
3838 | restores are finished. */ | |
3839 | emit_insn (gen_blockage ()); | |
daee63dd | 3840 | |
9840d99d | 3841 | /* Reset stack pointer (and possibly frame pointer). The stack |
42819d4e | 3842 | pointer is initially set to fp + 64 to avoid a race condition. */ |
58361f39 | 3843 | if (frame_pointer_needed) |
87ad11b0 | 3844 | { |
cc858176 | 3845 | rtx delta = GEN_INT (-64); |
a584fe8a | 3846 | |
3847 | set_reg_plus_d (STACK_POINTER_REGNUM, FRAME_POINTER_REGNUM, 64, 0); | |
3848 | emit_insn (gen_pre_load (frame_pointer_rtx, stack_pointer_rtx, delta)); | |
87ad11b0 | 3849 | } |
daee63dd | 3850 | /* If we were deferring a callee register restore, do it now. */ |
58361f39 | 3851 | else if (merge_sp_adjust_with_load) |
3852 | { | |
3853 | rtx delta = GEN_INT (-actual_fsize); | |
cc858176 | 3854 | rtx dest = gen_rtx_REG (word_mode, merge_sp_adjust_with_load); |
a584fe8a | 3855 | |
3856 | emit_insn (gen_pre_load (dest, stack_pointer_rtx, delta)); | |
58361f39 | 3857 | } |
daee63dd | 3858 | else if (actual_fsize != 0) |
a584fe8a | 3859 | set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM, |
3860 | - actual_fsize, 0); | |
58361f39 | 3861 | |
3862 | /* If we haven't restored %r2 yet (no frame pointer, and a stack | |
3863 | frame greater than 8k), do so now. */ | |
3864 | if (ret_off != 0) | |
6a2c16d6 | 3865 | load_reg (2, ret_off, STACK_POINTER_REGNUM); |
a584fe8a | 3866 | |
3867 | if (DO_FRAME_NOTES && current_function_calls_eh_return) | |
3868 | { | |
3869 | rtx sa = EH_RETURN_STACKADJ_RTX; | |
3870 | ||
3871 | emit_insn (gen_blockage ()); | |
3872 | emit_insn (TARGET_64BIT | |
3873 | ? gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, sa) | |
3874 | : gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, sa)); | |
3875 | } | |
87ad11b0 | 3876 | } |
3877 | ||
d7e2f694 | 3878 | rtx |
3879 | hppa_pic_save_rtx () | |
cf3de5bb | 3880 | { |
d7e2f694 | 3881 | return get_hard_reg_initial_val (word_mode, PIC_OFFSET_TABLE_REGNUM); |
df6edefa | 3882 | } |
3883 | ||
3884 | void | |
3885 | hppa_profile_hook (label_no) | |
b8a21949 | 3886 | int label_no; |
df6edefa | 3887 | { |
a9ac13e4 | 3888 | rtx begin_label_rtx, call_insn; |
3889 | char begin_label_name[16]; | |
df6edefa | 3890 | |
a9ac13e4 | 3891 | ASM_GENERATE_INTERNAL_LABEL (begin_label_name, FUNC_BEGIN_PROLOG_LABEL, |
b8a21949 | 3892 | label_no); |
a9ac13e4 | 3893 | begin_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (begin_label_name)); |
df6edefa | 3894 | |
3895 | if (TARGET_64BIT) | |
3896 | emit_move_insn (arg_pointer_rtx, | |
3897 | gen_rtx_PLUS (word_mode, virtual_outgoing_args_rtx, | |
3898 | GEN_INT (64))); | |
3899 | ||
df6edefa | 3900 | emit_move_insn (gen_rtx_REG (word_mode, 26), gen_rtx_REG (word_mode, 2)); |
3901 | ||
3902 | #ifndef NO_PROFILE_COUNTERS | |
3903 | { | |
3904 | rtx count_label_rtx, addr, r24; | |
a9ac13e4 | 3905 | char count_label_name[16]; |
df6edefa | 3906 | |
a9ac13e4 | 3907 | ASM_GENERATE_INTERNAL_LABEL (count_label_name, "LP", label_no); |
3908 | count_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (count_label_name)); | |
df6edefa | 3909 | |
831a12d9 | 3910 | addr = force_reg (Pmode, count_label_rtx); |
df6edefa | 3911 | r24 = gen_rtx_REG (Pmode, 24); |
3912 | emit_move_insn (r24, addr); | |
3913 | ||
3914 | /* %r25 is set from within the output pattern. */ | |
3915 | call_insn = | |
3916 | emit_call_insn (gen_call_profiler (gen_rtx_SYMBOL_REF (Pmode, "_mcount"), | |
3917 | GEN_INT (TARGET_64BIT ? 24 : 12), | |
a9ac13e4 | 3918 | begin_label_rtx)); |
df6edefa | 3919 | |
3920 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), r24); | |
3921 | } | |
3922 | #else | |
3923 | /* %r25 is set from within the output pattern. */ | |
3924 | call_insn = | |
3925 | emit_call_insn (gen_call_profiler (gen_rtx_SYMBOL_REF (Pmode, "_mcount"), | |
3926 | GEN_INT (TARGET_64BIT ? 16 : 8), | |
a9ac13e4 | 3927 | begin_label_rtx)); |
df6edefa | 3928 | #endif |
3929 | ||
3930 | /* Indicate the _mcount call cannot throw, nor will it execute a | |
3931 | non-local goto. */ | |
3932 | REG_NOTES (call_insn) | |
3933 | = gen_rtx_EXPR_LIST (REG_EH_REGION, constm1_rtx, REG_NOTES (call_insn)); | |
3934 | ||
3935 | if (flag_pic) | |
3936 | { | |
3937 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), pic_offset_table_rtx); | |
3938 | if (TARGET_64BIT) | |
3939 | use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), arg_pointer_rtx); | |
3940 | ||
d7e2f694 | 3941 | emit_move_insn (pic_offset_table_rtx, hppa_pic_save_rtx ()); |
df6edefa | 3942 | } |
cf3de5bb | 3943 | } |
3944 | ||
e07ff380 | 3945 | /* Fetch the return address for the frame COUNT steps up from |
3946 | the current frame, after the prologue. FRAMEADDR is the | |
3947 | frame pointer of the COUNT frame. | |
3948 | ||
f49b2e77 | 3949 | We want to ignore any export stub remnants here. To handle this, |
3950 | we examine the code at the return address, and if it is an export | |
3951 | stub, we return a memory rtx for the stub return address stored | |
3952 | at frame-24. | |
a6c6fd6c | 3953 | |
3954 | The value returned is used in two different ways: | |
3955 | ||
3956 | 1. To find a function's caller. | |
3957 | ||
3958 | 2. To change the return address for a function. | |
3959 | ||
3960 | This function handles most instances of case 1; however, it will | |
3961 | fail if there are two levels of stubs to execute on the return | |
3962 | path. The only way I believe that can happen is if the return value | |
3963 | needs a parameter relocation, which never happens for C code. | |
3964 | ||
3965 | This function handles most instances of case 2; however, it will | |
3966 | fail if we did not originally have stub code on the return path | |
f49b2e77 | 3967 | but will need stub code on the new return path. This can happen if |
a6c6fd6c | 3968 | the caller & callee are both in the main program, but the new |
f49b2e77 | 3969 | return location is in a shared library. */ |
e07ff380 | 3970 | |
3971 | rtx | |
3972 | return_addr_rtx (count, frameaddr) | |
f49b2e77 | 3973 | int count; |
e07ff380 | 3974 | rtx frameaddr; |
3975 | { | |
3976 | rtx label; | |
f49b2e77 | 3977 | rtx rp; |
e07ff380 | 3978 | rtx saved_rp; |
3979 | rtx ins; | |
3980 | ||
f49b2e77 | 3981 | if (count != 0) |
3982 | return NULL_RTX; | |
b29897dd | 3983 | |
f49b2e77 | 3984 | rp = get_hard_reg_initial_val (Pmode, 2); |
e07ff380 | 3985 | |
f49b2e77 | 3986 | if (TARGET_64BIT || TARGET_NO_SPACE_REGS) |
3987 | return rp; | |
e07ff380 | 3988 | |
b29897dd | 3989 | saved_rp = gen_reg_rtx (Pmode); |
f49b2e77 | 3990 | emit_move_insn (saved_rp, rp); |
e07ff380 | 3991 | |
3992 | /* Get pointer to the instruction stream. We have to mask out the | |
3993 | privilege level from the two low order bits of the return address | |
3994 | pointer here so that ins will point to the start of the first | |
3995 | instruction that would have been executed if we returned. */ | |
f49b2e77 | 3996 | ins = copy_to_reg (gen_rtx_AND (Pmode, rp, MASK_RETURN_ADDR)); |
e07ff380 | 3997 | label = gen_label_rtx (); |
3998 | ||
3999 | /* Check the instruction stream at the normal return address for the | |
4000 | export stub: | |
4001 | ||
4002 | 0x4bc23fd1 | stub+8: ldw -18(sr0,sp),rp | |
4003 | 0x004010a1 | stub+12: ldsid (sr0,rp),r1 | |
4004 | 0x00011820 | stub+16: mtsp r1,sr0 | |
4005 | 0xe0400002 | stub+20: be,n 0(sr0,rp) | |
4006 | ||
4007 | If it is an export stub, than our return address is really in | |
4008 | -24[frameaddr]. */ | |
4009 | ||
e00b80b2 | 4010 | emit_cmp_insn (gen_rtx_MEM (SImode, ins), GEN_INT (0x4bc23fd1), NE, |
4011 | NULL_RTX, SImode, 1); | |
e07ff380 | 4012 | emit_jump_insn (gen_bne (label)); |
4013 | ||
ad851752 | 4014 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 4)), |
e00b80b2 | 4015 | GEN_INT (0x004010a1), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4016 | emit_jump_insn (gen_bne (label)); |
4017 | ||
ad851752 | 4018 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 8)), |
e00b80b2 | 4019 | GEN_INT (0x00011820), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4020 | emit_jump_insn (gen_bne (label)); |
4021 | ||
ad851752 | 4022 | emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 12)), |
e00b80b2 | 4023 | GEN_INT (0xe0400002), NE, NULL_RTX, SImode, 1); |
e07ff380 | 4024 | |
f49b2e77 | 4025 | /* If there is no export stub then just use the value saved from |
4026 | the return pointer register. */ | |
e07ff380 | 4027 | |
4028 | emit_jump_insn (gen_bne (label)); | |
4029 | ||
f49b2e77 | 4030 | /* Here we know that our return address points to an export |
e07ff380 | 4031 | stub. We don't want to return the address of the export stub, |
f49b2e77 | 4032 | but rather the return address of the export stub. That return |
4033 | address is stored at -24[frameaddr]. */ | |
e07ff380 | 4034 | |
f49b2e77 | 4035 | emit_move_insn (saved_rp, |
4036 | gen_rtx_MEM (Pmode, | |
4037 | memory_address (Pmode, | |
4038 | plus_constant (frameaddr, | |
4039 | -24)))); | |
e07ff380 | 4040 | |
4041 | emit_label (label); | |
f49b2e77 | 4042 | return saved_rp; |
e07ff380 | 4043 | } |
4044 | ||
757d4970 | 4045 | /* This is only valid once reload has completed because it depends on |
4046 | knowing exactly how much (if any) frame there is and... | |
4047 | ||
4048 | It's only valid if there is no frame marker to de-allocate and... | |
4049 | ||
4050 | It's only valid if %r2 hasn't been saved into the caller's frame | |
4051 | (we're not profiling and %r2 isn't live anywhere). */ | |
4052 | int | |
4053 | hppa_can_use_return_insn_p () | |
4054 | { | |
4055 | return (reload_completed | |
4056 | && (compute_frame_size (get_frame_size (), 0) ? 0 : 1) | |
757d4970 | 4057 | && ! regs_ever_live[2] |
4058 | && ! frame_pointer_needed); | |
4059 | } | |
4060 | ||
87ad11b0 | 4061 | void |
4062 | emit_bcond_fp (code, operand0) | |
4063 | enum rtx_code code; | |
4064 | rtx operand0; | |
4065 | { | |
ad851752 | 4066 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, |
4067 | gen_rtx_IF_THEN_ELSE (VOIDmode, | |
4068 | gen_rtx_fmt_ee (code, | |
4069 | VOIDmode, | |
4070 | gen_rtx_REG (CCFPmode, 0), | |
4071 | const0_rtx), | |
4072 | gen_rtx_LABEL_REF (VOIDmode, operand0), | |
4073 | pc_rtx))); | |
87ad11b0 | 4074 | |
4075 | } | |
4076 | ||
4077 | rtx | |
4078 | gen_cmp_fp (code, operand0, operand1) | |
4079 | enum rtx_code code; | |
4080 | rtx operand0, operand1; | |
4081 | { | |
ad851752 | 4082 | return gen_rtx_SET (VOIDmode, gen_rtx_REG (CCFPmode, 0), |
4083 | gen_rtx_fmt_ee (code, CCFPmode, operand0, operand1)); | |
87ad11b0 | 4084 | } |
4085 | ||
8b49b3c7 | 4086 | /* Adjust the cost of a scheduling dependency. Return the new cost of |
4087 | a dependency LINK or INSN on DEP_INSN. COST is the current cost. */ | |
4088 | ||
747af5e7 | 4089 | static int |
8b49b3c7 | 4090 | pa_adjust_cost (insn, link, dep_insn, cost) |
4091 | rtx insn; | |
4092 | rtx link; | |
4093 | rtx dep_insn; | |
4094 | int cost; | |
4095 | { | |
43048d2c | 4096 | enum attr_type attr_type; |
4097 | ||
cde3e16c | 4098 | /* Don't adjust costs for a pa8000 chip, also do not adjust any |
4099 | true dependencies as they are described with bypasses now. */ | |
4100 | if (pa_cpu >= PROCESSOR_8000 || REG_NOTE_KIND (link) == 0) | |
342aabd9 | 4101 | return cost; |
4102 | ||
d402da4b | 4103 | if (! recog_memoized (insn)) |
4104 | return 0; | |
8b49b3c7 | 4105 | |
43048d2c | 4106 | attr_type = get_attr_type (insn); |
4107 | ||
cde3e16c | 4108 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI) |
8b49b3c7 | 4109 | { |
4110 | /* Anti dependency; DEP_INSN reads a register that INSN writes some | |
4111 | cycles later. */ | |
4112 | ||
43048d2c | 4113 | if (attr_type == TYPE_FPLOAD) |
8b49b3c7 | 4114 | { |
d402da4b | 4115 | rtx pat = PATTERN (insn); |
4116 | rtx dep_pat = PATTERN (dep_insn); | |
4117 | if (GET_CODE (pat) == PARALLEL) | |
4118 | { | |
4119 | /* This happens for the fldXs,mb patterns. */ | |
4120 | pat = XVECEXP (pat, 0, 0); | |
4121 | } | |
4122 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
8b49b3c7 | 4123 | /* If this happens, we have to extend this to schedule |
d402da4b | 4124 | optimally. Return 0 for now. */ |
4125 | return 0; | |
8b49b3c7 | 4126 | |
d402da4b | 4127 | if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat))) |
8b49b3c7 | 4128 | { |
d402da4b | 4129 | if (! recog_memoized (dep_insn)) |
4130 | return 0; | |
8b49b3c7 | 4131 | switch (get_attr_type (dep_insn)) |
4132 | { | |
4133 | case TYPE_FPALU: | |
134b4858 | 4134 | case TYPE_FPMULSGL: |
4135 | case TYPE_FPMULDBL: | |
8b49b3c7 | 4136 | case TYPE_FPDIVSGL: |
4137 | case TYPE_FPDIVDBL: | |
4138 | case TYPE_FPSQRTSGL: | |
4139 | case TYPE_FPSQRTDBL: | |
d402da4b | 4140 | /* A fpload can't be issued until one cycle before a |
01cc3b75 | 4141 | preceding arithmetic operation has finished if |
d402da4b | 4142 | the target of the fpload is any of the sources |
4143 | (or destination) of the arithmetic operation. */ | |
cde3e16c | 4144 | return insn_default_latency (dep_insn) - 1; |
134b4858 | 4145 | |
4146 | default: | |
4147 | return 0; | |
4148 | } | |
4149 | } | |
4150 | } | |
43048d2c | 4151 | else if (attr_type == TYPE_FPALU) |
134b4858 | 4152 | { |
4153 | rtx pat = PATTERN (insn); | |
4154 | rtx dep_pat = PATTERN (dep_insn); | |
4155 | if (GET_CODE (pat) == PARALLEL) | |
4156 | { | |
4157 | /* This happens for the fldXs,mb patterns. */ | |
4158 | pat = XVECEXP (pat, 0, 0); | |
4159 | } | |
4160 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4161 | /* If this happens, we have to extend this to schedule | |
4162 | optimally. Return 0 for now. */ | |
4163 | return 0; | |
4164 | ||
4165 | if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat))) | |
4166 | { | |
4167 | if (! recog_memoized (dep_insn)) | |
4168 | return 0; | |
4169 | switch (get_attr_type (dep_insn)) | |
4170 | { | |
4171 | case TYPE_FPDIVSGL: | |
4172 | case TYPE_FPDIVDBL: | |
4173 | case TYPE_FPSQRTSGL: | |
4174 | case TYPE_FPSQRTDBL: | |
4175 | /* An ALU flop can't be issued until two cycles before a | |
01cc3b75 | 4176 | preceding divide or sqrt operation has finished if |
134b4858 | 4177 | the target of the ALU flop is any of the sources |
4178 | (or destination) of the divide or sqrt operation. */ | |
cde3e16c | 4179 | return insn_default_latency (dep_insn) - 2; |
8b49b3c7 | 4180 | |
4181 | default: | |
4182 | return 0; | |
4183 | } | |
4184 | } | |
4185 | } | |
4186 | ||
4187 | /* For other anti dependencies, the cost is 0. */ | |
4188 | return 0; | |
4189 | } | |
134b4858 | 4190 | else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT) |
4191 | { | |
4192 | /* Output dependency; DEP_INSN writes a register that INSN writes some | |
4193 | cycles later. */ | |
43048d2c | 4194 | if (attr_type == TYPE_FPLOAD) |
134b4858 | 4195 | { |
4196 | rtx pat = PATTERN (insn); | |
4197 | rtx dep_pat = PATTERN (dep_insn); | |
4198 | if (GET_CODE (pat) == PARALLEL) | |
4199 | { | |
4200 | /* This happens for the fldXs,mb patterns. */ | |
4201 | pat = XVECEXP (pat, 0, 0); | |
4202 | } | |
4203 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4204 | /* If this happens, we have to extend this to schedule | |
4205 | optimally. Return 0 for now. */ | |
4206 | return 0; | |
4207 | ||
4208 | if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat))) | |
4209 | { | |
4210 | if (! recog_memoized (dep_insn)) | |
4211 | return 0; | |
4212 | switch (get_attr_type (dep_insn)) | |
4213 | { | |
4214 | case TYPE_FPALU: | |
4215 | case TYPE_FPMULSGL: | |
4216 | case TYPE_FPMULDBL: | |
4217 | case TYPE_FPDIVSGL: | |
4218 | case TYPE_FPDIVDBL: | |
4219 | case TYPE_FPSQRTSGL: | |
4220 | case TYPE_FPSQRTDBL: | |
4221 | /* A fpload can't be issued until one cycle before a | |
01cc3b75 | 4222 | preceding arithmetic operation has finished if |
134b4858 | 4223 | the target of the fpload is the destination of the |
bea4bad2 | 4224 | arithmetic operation. |
4225 | ||
4226 | Exception: For PA7100LC, PA7200 and PA7300, the cost | |
4227 | is 3 cycles, unless they bundle together. We also | |
4228 | pay the penalty if the second insn is a fpload. */ | |
cde3e16c | 4229 | return insn_default_latency (dep_insn) - 1; |
8b49b3c7 | 4230 | |
134b4858 | 4231 | default: |
4232 | return 0; | |
4233 | } | |
4234 | } | |
4235 | } | |
43048d2c | 4236 | else if (attr_type == TYPE_FPALU) |
134b4858 | 4237 | { |
4238 | rtx pat = PATTERN (insn); | |
4239 | rtx dep_pat = PATTERN (dep_insn); | |
4240 | if (GET_CODE (pat) == PARALLEL) | |
4241 | { | |
4242 | /* This happens for the fldXs,mb patterns. */ | |
4243 | pat = XVECEXP (pat, 0, 0); | |
4244 | } | |
4245 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
4246 | /* If this happens, we have to extend this to schedule | |
4247 | optimally. Return 0 for now. */ | |
4248 | return 0; | |
4249 | ||
4250 | if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat))) | |
4251 | { | |
4252 | if (! recog_memoized (dep_insn)) | |
4253 | return 0; | |
4254 | switch (get_attr_type (dep_insn)) | |
4255 | { | |
4256 | case TYPE_FPDIVSGL: | |
4257 | case TYPE_FPDIVDBL: | |
4258 | case TYPE_FPSQRTSGL: | |
4259 | case TYPE_FPSQRTDBL: | |
4260 | /* An ALU flop can't be issued until two cycles before a | |
01cc3b75 | 4261 | preceding divide or sqrt operation has finished if |
134b4858 | 4262 | the target of the ALU flop is also the target of |
3398e91d | 4263 | the divide or sqrt operation. */ |
cde3e16c | 4264 | return insn_default_latency (dep_insn) - 2; |
134b4858 | 4265 | |
4266 | default: | |
4267 | return 0; | |
4268 | } | |
4269 | } | |
4270 | } | |
4271 | ||
4272 | /* For other output dependencies, the cost is 0. */ | |
4273 | return 0; | |
4274 | } | |
4275 | else | |
4276 | abort (); | |
8b49b3c7 | 4277 | } |
87ad11b0 | 4278 | |
747af5e7 | 4279 | /* Adjust scheduling priorities. We use this to try and keep addil |
4280 | and the next use of %r1 close together. */ | |
4281 | static int | |
4282 | pa_adjust_priority (insn, priority) | |
4283 | rtx insn; | |
4284 | int priority; | |
4285 | { | |
4286 | rtx set = single_set (insn); | |
4287 | rtx src, dest; | |
4288 | if (set) | |
4289 | { | |
4290 | src = SET_SRC (set); | |
4291 | dest = SET_DEST (set); | |
4292 | if (GET_CODE (src) == LO_SUM | |
4293 | && symbolic_operand (XEXP (src, 1), VOIDmode) | |
4294 | && ! read_only_operand (XEXP (src, 1), VOIDmode)) | |
4295 | priority >>= 3; | |
4296 | ||
4297 | else if (GET_CODE (src) == MEM | |
4298 | && GET_CODE (XEXP (src, 0)) == LO_SUM | |
4299 | && symbolic_operand (XEXP (XEXP (src, 0), 1), VOIDmode) | |
4300 | && ! read_only_operand (XEXP (XEXP (src, 0), 1), VOIDmode)) | |
4301 | priority >>= 1; | |
4302 | ||
4303 | else if (GET_CODE (dest) == MEM | |
4304 | && GET_CODE (XEXP (dest, 0)) == LO_SUM | |
4305 | && symbolic_operand (XEXP (XEXP (dest, 0), 1), VOIDmode) | |
4306 | && ! read_only_operand (XEXP (XEXP (dest, 0), 1), VOIDmode)) | |
4307 | priority >>= 3; | |
4308 | } | |
4309 | return priority; | |
4310 | } | |
4311 | ||
4312 | /* The 700 can only issue a single insn at a time. | |
4313 | The 7XXX processors can issue two insns at a time. | |
4314 | The 8000 can issue 4 insns at a time. */ | |
4315 | static int | |
4316 | pa_issue_rate () | |
4317 | { | |
4318 | switch (pa_cpu) | |
4319 | { | |
4320 | case PROCESSOR_700: return 1; | |
4321 | case PROCESSOR_7100: return 2; | |
4322 | case PROCESSOR_7100LC: return 2; | |
4323 | case PROCESSOR_7200: return 2; | |
bea4bad2 | 4324 | case PROCESSOR_7300: return 2; |
747af5e7 | 4325 | case PROCESSOR_8000: return 4; |
4326 | ||
4327 | default: | |
4328 | abort (); | |
4329 | } | |
4330 | } | |
4331 | ||
4332 | ||
4333 | ||
58e17b0b | 4334 | /* Return any length adjustment needed by INSN which already has its length |
6d36483b | 4335 | computed as LENGTH. Return zero if no adjustment is necessary. |
58e17b0b | 4336 | |
5fbd5940 | 4337 | For the PA: function calls, millicode calls, and backwards short |
6d36483b | 4338 | conditional branches with unfilled delay slots need an adjustment by +1 |
5fbd5940 | 4339 | (to account for the NOP which will be inserted into the instruction stream). |
58e17b0b | 4340 | |
4341 | Also compute the length of an inline block move here as it is too | |
5fbd5940 | 4342 | complicated to express as a length attribute in pa.md. */ |
58e17b0b | 4343 | int |
4344 | pa_adjust_insn_length (insn, length) | |
4345 | rtx insn; | |
4346 | int length; | |
4347 | { | |
4348 | rtx pat = PATTERN (insn); | |
4349 | ||
5fbd5940 | 4350 | /* Call insns which are *not* indirect and have unfilled delay slots. */ |
58e17b0b | 4351 | if (GET_CODE (insn) == CALL_INSN) |
5fbd5940 | 4352 | { |
4353 | ||
4354 | if (GET_CODE (XVECEXP (pat, 0, 0)) == CALL | |
4355 | && GET_CODE (XEXP (XEXP (XVECEXP (pat, 0, 0), 0), 0)) == SYMBOL_REF) | |
5a1231ef | 4356 | return 4; |
5fbd5940 | 4357 | else if (GET_CODE (XVECEXP (pat, 0, 0)) == SET |
4358 | && GET_CODE (XEXP (XEXP (XEXP (XVECEXP (pat, 0, 0), 1), 0), 0)) | |
4359 | == SYMBOL_REF) | |
5a1231ef | 4360 | return 4; |
5fbd5940 | 4361 | else |
4362 | return 0; | |
4363 | } | |
9840d99d | 4364 | /* Jumps inside switch tables which have unfilled delay slots |
3b1e673e | 4365 | also need adjustment. */ |
4366 | else if (GET_CODE (insn) == JUMP_INSN | |
4367 | && simplejump_p (insn) | |
0708e381 | 4368 | && GET_MODE (insn) == SImode) |
3b1e673e | 4369 | return 4; |
58e17b0b | 4370 | /* Millicode insn with an unfilled delay slot. */ |
4371 | else if (GET_CODE (insn) == INSN | |
4372 | && GET_CODE (pat) != SEQUENCE | |
4373 | && GET_CODE (pat) != USE | |
4374 | && GET_CODE (pat) != CLOBBER | |
4375 | && get_attr_type (insn) == TYPE_MILLI) | |
5a1231ef | 4376 | return 4; |
58e17b0b | 4377 | /* Block move pattern. */ |
4378 | else if (GET_CODE (insn) == INSN | |
4379 | && GET_CODE (pat) == PARALLEL | |
f2ebcf32 | 4380 | && GET_CODE (XVECEXP (pat, 0, 0)) == SET |
58e17b0b | 4381 | && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM |
4382 | && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 1)) == MEM | |
4383 | && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode | |
4384 | && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 1)) == BLKmode) | |
5a1231ef | 4385 | return compute_movstrsi_length (insn) - 4; |
58e17b0b | 4386 | /* Conditional branch with an unfilled delay slot. */ |
5fbd5940 | 4387 | else if (GET_CODE (insn) == JUMP_INSN && ! simplejump_p (insn)) |
4388 | { | |
4389 | /* Adjust a short backwards conditional with an unfilled delay slot. */ | |
4390 | if (GET_CODE (pat) == SET | |
5a1231ef | 4391 | && length == 4 |
5fbd5940 | 4392 | && ! forward_branch_p (insn)) |
5a1231ef | 4393 | return 4; |
546a40bd | 4394 | else if (GET_CODE (pat) == PARALLEL |
4395 | && get_attr_type (insn) == TYPE_PARALLEL_BRANCH | |
4396 | && length == 4) | |
4397 | return 4; | |
5fbd5940 | 4398 | /* Adjust dbra insn with short backwards conditional branch with |
6d36483b | 4399 | unfilled delay slot -- only for case where counter is in a |
6dc3b0d9 | 4400 | general register register. */ |
5fbd5940 | 4401 | else if (GET_CODE (pat) == PARALLEL |
4402 | && GET_CODE (XVECEXP (pat, 0, 1)) == SET | |
4403 | && GET_CODE (XEXP (XVECEXP (pat, 0, 1), 0)) == REG | |
6d36483b | 4404 | && ! FP_REG_P (XEXP (XVECEXP (pat, 0, 1), 0)) |
5a1231ef | 4405 | && length == 4 |
5fbd5940 | 4406 | && ! forward_branch_p (insn)) |
5a1231ef | 4407 | return 4; |
5fbd5940 | 4408 | else |
4409 | return 0; | |
4410 | } | |
546a40bd | 4411 | return 0; |
58e17b0b | 4412 | } |
4413 | ||
87ad11b0 | 4414 | /* Print operand X (an rtx) in assembler syntax to file FILE. |
4415 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
4416 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
4417 | ||
4418 | void | |
4419 | print_operand (file, x, code) | |
4420 | FILE *file; | |
4421 | rtx x; | |
4422 | int code; | |
4423 | { | |
4424 | switch (code) | |
4425 | { | |
4426 | case '#': | |
4427 | /* Output a 'nop' if there's nothing for the delay slot. */ | |
4428 | if (dbr_sequence_length () == 0) | |
4429 | fputs ("\n\tnop", file); | |
4430 | return; | |
4431 | case '*': | |
87fcb603 | 4432 | /* Output a nullification completer if there's nothing for the */ |
6d36483b | 4433 | /* delay slot or nullification is requested. */ |
87ad11b0 | 4434 | if (dbr_sequence_length () == 0 || |
4435 | (final_sequence && | |
4436 | INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)))) | |
4437 | fputs (",n", file); | |
4438 | return; | |
4439 | case 'R': | |
4440 | /* Print out the second register name of a register pair. | |
4441 | I.e., R (6) => 7. */ | |
ea52c577 | 4442 | fputs (reg_names[REGNO (x) + 1], file); |
87ad11b0 | 4443 | return; |
4444 | case 'r': | |
6dc3b0d9 | 4445 | /* A register or zero. */ |
891b55b4 | 4446 | if (x == const0_rtx |
4447 | || (x == CONST0_RTX (DFmode)) | |
4448 | || (x == CONST0_RTX (SFmode))) | |
87ad11b0 | 4449 | { |
c6ae275c | 4450 | fputs ("%r0", file); |
4451 | return; | |
4452 | } | |
4453 | else | |
4454 | break; | |
4455 | case 'f': | |
6dc3b0d9 | 4456 | /* A register or zero (floating point). */ |
c6ae275c | 4457 | if (x == const0_rtx |
4458 | || (x == CONST0_RTX (DFmode)) | |
4459 | || (x == CONST0_RTX (SFmode))) | |
4460 | { | |
4461 | fputs ("%fr0", file); | |
87ad11b0 | 4462 | return; |
4463 | } | |
4464 | else | |
4465 | break; | |
2d14b1f0 | 4466 | case 'A': |
4467 | { | |
4468 | rtx xoperands[2]; | |
4469 | ||
4470 | xoperands[0] = XEXP (XEXP (x, 0), 0); | |
4471 | xoperands[1] = XVECEXP (XEXP (XEXP (x, 0), 1), 0, 0); | |
4472 | output_global_address (file, xoperands[1], 0); | |
4473 | fprintf (file, "(%s)", reg_names [REGNO (xoperands[0])]); | |
4474 | return; | |
4475 | } | |
4476 | ||
c8975385 | 4477 | case 'C': /* Plain (C)ondition */ |
87ad11b0 | 4478 | case 'X': |
4479 | switch (GET_CODE (x)) | |
6d36483b | 4480 | { |
87ad11b0 | 4481 | case EQ: |
9c0ac0fd | 4482 | fputs ("=", file); break; |
87ad11b0 | 4483 | case NE: |
9c0ac0fd | 4484 | fputs ("<>", file); break; |
87ad11b0 | 4485 | case GT: |
9c0ac0fd | 4486 | fputs (">", file); break; |
87ad11b0 | 4487 | case GE: |
9c0ac0fd | 4488 | fputs (">=", file); break; |
87ad11b0 | 4489 | case GEU: |
9c0ac0fd | 4490 | fputs (">>=", file); break; |
87ad11b0 | 4491 | case GTU: |
9c0ac0fd | 4492 | fputs (">>", file); break; |
87ad11b0 | 4493 | case LT: |
9c0ac0fd | 4494 | fputs ("<", file); break; |
87ad11b0 | 4495 | case LE: |
9c0ac0fd | 4496 | fputs ("<=", file); break; |
87ad11b0 | 4497 | case LEU: |
9c0ac0fd | 4498 | fputs ("<<=", file); break; |
87ad11b0 | 4499 | case LTU: |
9c0ac0fd | 4500 | fputs ("<<", file); break; |
87ad11b0 | 4501 | default: |
87ad11b0 | 4502 | abort (); |
4503 | } | |
4504 | return; | |
c8975385 | 4505 | case 'N': /* Condition, (N)egated */ |
87ad11b0 | 4506 | switch (GET_CODE (x)) |
4507 | { | |
4508 | case EQ: | |
9c0ac0fd | 4509 | fputs ("<>", file); break; |
87ad11b0 | 4510 | case NE: |
9c0ac0fd | 4511 | fputs ("=", file); break; |
87ad11b0 | 4512 | case GT: |
9c0ac0fd | 4513 | fputs ("<=", file); break; |
87ad11b0 | 4514 | case GE: |
9c0ac0fd | 4515 | fputs ("<", file); break; |
87ad11b0 | 4516 | case GEU: |
9c0ac0fd | 4517 | fputs ("<<", file); break; |
87ad11b0 | 4518 | case GTU: |
9c0ac0fd | 4519 | fputs ("<<=", file); break; |
87ad11b0 | 4520 | case LT: |
9c0ac0fd | 4521 | fputs (">=", file); break; |
87ad11b0 | 4522 | case LE: |
9c0ac0fd | 4523 | fputs (">", file); break; |
87ad11b0 | 4524 | case LEU: |
9c0ac0fd | 4525 | fputs (">>", file); break; |
87ad11b0 | 4526 | case LTU: |
9c0ac0fd | 4527 | fputs (">>=", file); break; |
87ad11b0 | 4528 | default: |
87ad11b0 | 4529 | abort (); |
4530 | } | |
4531 | return; | |
ea52c577 | 4532 | /* For floating point comparisons. Note that the output |
4533 | predicates are the complement of the desired mode. */ | |
61230bc9 | 4534 | case 'Y': |
4535 | switch (GET_CODE (x)) | |
4536 | { | |
4537 | case EQ: | |
9c0ac0fd | 4538 | fputs ("!=", file); break; |
61230bc9 | 4539 | case NE: |
9c0ac0fd | 4540 | fputs ("=", file); break; |
61230bc9 | 4541 | case GT: |
32509e56 | 4542 | fputs ("!>", file); break; |
61230bc9 | 4543 | case GE: |
32509e56 | 4544 | fputs ("!>=", file); break; |
61230bc9 | 4545 | case LT: |
32509e56 | 4546 | fputs ("!<", file); break; |
61230bc9 | 4547 | case LE: |
32509e56 | 4548 | fputs ("!<=", file); break; |
4549 | case LTGT: | |
4550 | fputs ("!<>", file); break; | |
4551 | case UNLE: | |
85837adc | 4552 | fputs (">", file); break; |
32509e56 | 4553 | case UNLT: |
4554 | fputs (">=", file); break; | |
4555 | case UNGE: | |
4556 | fputs ("<", file); break; | |
4557 | case UNGT: | |
4558 | fputs ("<=", file); break; | |
4559 | case UNEQ: | |
4560 | fputs ("<>", file); break; | |
4561 | case UNORDERED: | |
4562 | fputs ("<=>", file); break; | |
4563 | case ORDERED: | |
4564 | fputs ("!<=>", file); break; | |
61230bc9 | 4565 | default: |
61230bc9 | 4566 | abort (); |
4567 | } | |
4568 | return; | |
c8975385 | 4569 | case 'S': /* Condition, operands are (S)wapped. */ |
4570 | switch (GET_CODE (x)) | |
4571 | { | |
4572 | case EQ: | |
9c0ac0fd | 4573 | fputs ("=", file); break; |
c8975385 | 4574 | case NE: |
9c0ac0fd | 4575 | fputs ("<>", file); break; |
c8975385 | 4576 | case GT: |
9c0ac0fd | 4577 | fputs ("<", file); break; |
c8975385 | 4578 | case GE: |
9c0ac0fd | 4579 | fputs ("<=", file); break; |
c8975385 | 4580 | case GEU: |
9c0ac0fd | 4581 | fputs ("<<=", file); break; |
c8975385 | 4582 | case GTU: |
9c0ac0fd | 4583 | fputs ("<<", file); break; |
c8975385 | 4584 | case LT: |
9c0ac0fd | 4585 | fputs (">", file); break; |
c8975385 | 4586 | case LE: |
9c0ac0fd | 4587 | fputs (">=", file); break; |
c8975385 | 4588 | case LEU: |
9c0ac0fd | 4589 | fputs (">>=", file); break; |
c8975385 | 4590 | case LTU: |
9c0ac0fd | 4591 | fputs (">>", file); break; |
c8975385 | 4592 | default: |
c8975385 | 4593 | abort (); |
6d36483b | 4594 | } |
c8975385 | 4595 | return; |
4596 | case 'B': /* Condition, (B)oth swapped and negate. */ | |
4597 | switch (GET_CODE (x)) | |
4598 | { | |
4599 | case EQ: | |
9c0ac0fd | 4600 | fputs ("<>", file); break; |
c8975385 | 4601 | case NE: |
9c0ac0fd | 4602 | fputs ("=", file); break; |
c8975385 | 4603 | case GT: |
9c0ac0fd | 4604 | fputs (">=", file); break; |
c8975385 | 4605 | case GE: |
9c0ac0fd | 4606 | fputs (">", file); break; |
c8975385 | 4607 | case GEU: |
9c0ac0fd | 4608 | fputs (">>", file); break; |
c8975385 | 4609 | case GTU: |
9c0ac0fd | 4610 | fputs (">>=", file); break; |
c8975385 | 4611 | case LT: |
9c0ac0fd | 4612 | fputs ("<=", file); break; |
c8975385 | 4613 | case LE: |
9c0ac0fd | 4614 | fputs ("<", file); break; |
c8975385 | 4615 | case LEU: |
9c0ac0fd | 4616 | fputs ("<<", file); break; |
c8975385 | 4617 | case LTU: |
9c0ac0fd | 4618 | fputs ("<<=", file); break; |
c8975385 | 4619 | default: |
c8975385 | 4620 | abort (); |
6d36483b | 4621 | } |
c8975385 | 4622 | return; |
4623 | case 'k': | |
4624 | if (GET_CODE (x) == CONST_INT) | |
4625 | { | |
96b529f2 | 4626 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~INTVAL (x)); |
c8975385 | 4627 | return; |
4628 | } | |
ea52c577 | 4629 | abort (); |
5e3c5739 | 4630 | case 'Q': |
4631 | if (GET_CODE (x) == CONST_INT) | |
4632 | { | |
96b529f2 | 4633 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - (INTVAL (x) & 63)); |
5e3c5739 | 4634 | return; |
4635 | } | |
ea52c577 | 4636 | abort (); |
e5965947 | 4637 | case 'L': |
4638 | if (GET_CODE (x) == CONST_INT) | |
4639 | { | |
96b529f2 | 4640 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - (INTVAL (x) & 31)); |
e5965947 | 4641 | return; |
4642 | } | |
ea52c577 | 4643 | abort (); |
3a16146d | 4644 | case 'O': |
4645 | if (GET_CODE (x) == CONST_INT && exact_log2 (INTVAL (x)) >= 0) | |
4646 | { | |
4647 | fprintf (file, "%d", exact_log2 (INTVAL (x))); | |
4648 | return; | |
4649 | } | |
ea52c577 | 4650 | abort (); |
5e3c5739 | 4651 | case 'p': |
4652 | if (GET_CODE (x) == CONST_INT) | |
4653 | { | |
96b529f2 | 4654 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 63 - (INTVAL (x) & 63)); |
5e3c5739 | 4655 | return; |
4656 | } | |
ea52c577 | 4657 | abort (); |
e5965947 | 4658 | case 'P': |
4659 | if (GET_CODE (x) == CONST_INT) | |
4660 | { | |
96b529f2 | 4661 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, 31 - (INTVAL (x) & 31)); |
e5965947 | 4662 | return; |
4663 | } | |
ea52c577 | 4664 | abort (); |
c8975385 | 4665 | case 'I': |
4666 | if (GET_CODE (x) == CONST_INT) | |
4667 | fputs ("i", file); | |
4668 | return; | |
87ad11b0 | 4669 | case 'M': |
27ef382d | 4670 | case 'F': |
87ad11b0 | 4671 | switch (GET_CODE (XEXP (x, 0))) |
4672 | { | |
4673 | case PRE_DEC: | |
4674 | case PRE_INC: | |
e4065f95 | 4675 | if (ASSEMBLER_DIALECT == 0) |
4676 | fputs ("s,mb", file); | |
4677 | else | |
4678 | fputs (",mb", file); | |
87ad11b0 | 4679 | break; |
4680 | case POST_DEC: | |
4681 | case POST_INC: | |
e4065f95 | 4682 | if (ASSEMBLER_DIALECT == 0) |
4683 | fputs ("s,ma", file); | |
4684 | else | |
4685 | fputs (",ma", file); | |
87ad11b0 | 4686 | break; |
27ef382d | 4687 | case PLUS: |
4688 | if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
4689 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT) | |
e4065f95 | 4690 | { |
4691 | if (ASSEMBLER_DIALECT == 0) | |
4692 | fputs ("x,s", file); | |
4693 | else | |
4694 | fputs (",s", file); | |
4695 | } | |
4696 | else if (code == 'F' && ASSEMBLER_DIALECT == 0) | |
27ef382d | 4697 | fputs ("s", file); |
87ad11b0 | 4698 | break; |
4699 | default: | |
e4065f95 | 4700 | if (code == 'F' && ASSEMBLER_DIALECT == 0) |
27ef382d | 4701 | fputs ("s", file); |
87ad11b0 | 4702 | break; |
4703 | } | |
4704 | return; | |
4705 | case 'G': | |
f9333726 | 4706 | output_global_address (file, x, 0); |
4707 | return; | |
4708 | case 'H': | |
4709 | output_global_address (file, x, 1); | |
87ad11b0 | 4710 | return; |
4711 | case 0: /* Don't do anything special */ | |
4712 | break; | |
42faba01 | 4713 | case 'Z': |
4714 | { | |
4715 | unsigned op[3]; | |
fb22aedc | 4716 | compute_zdepwi_operands (INTVAL (x), op); |
42faba01 | 4717 | fprintf (file, "%d,%d,%d", op[0], op[1], op[2]); |
4718 | return; | |
4719 | } | |
5e3c5739 | 4720 | case 'z': |
4721 | { | |
4722 | unsigned op[3]; | |
4723 | compute_zdepdi_operands (INTVAL (x), op); | |
4724 | fprintf (file, "%d,%d,%d", op[0], op[1], op[2]); | |
4725 | return; | |
4726 | } | |
c9cc98e1 | 4727 | case 'c': |
4728 | /* We can get here from a .vtable_inherit due to our | |
4729 | CONSTANT_ADDRESS_P rejecting perfectly good constant | |
4730 | addresses. */ | |
4731 | break; | |
87ad11b0 | 4732 | default: |
4733 | abort (); | |
4734 | } | |
4735 | if (GET_CODE (x) == REG) | |
df0651dc | 4736 | { |
35661368 | 4737 | fputs (reg_names [REGNO (x)], file); |
5e3c5739 | 4738 | if (TARGET_64BIT && FP_REG_P (x) && GET_MODE_SIZE (GET_MODE (x)) <= 4) |
4739 | { | |
4740 | fputs ("R", file); | |
4741 | return; | |
4742 | } | |
4743 | if (FP_REG_P (x) | |
4744 | && GET_MODE_SIZE (GET_MODE (x)) <= 4 | |
4745 | && (REGNO (x) & 1) == 0) | |
35661368 | 4746 | fputs ("L", file); |
df0651dc | 4747 | } |
87ad11b0 | 4748 | else if (GET_CODE (x) == MEM) |
4749 | { | |
4750 | int size = GET_MODE_SIZE (GET_MODE (x)); | |
f7dff90d | 4751 | rtx base = NULL_RTX; |
87ad11b0 | 4752 | switch (GET_CODE (XEXP (x, 0))) |
4753 | { | |
4754 | case PRE_DEC: | |
4755 | case POST_DEC: | |
5e3c5739 | 4756 | base = XEXP (XEXP (x, 0), 0); |
34940871 | 4757 | fprintf (file, "-%d(%s)", size, reg_names [REGNO (base)]); |
87ad11b0 | 4758 | break; |
4759 | case PRE_INC: | |
4760 | case POST_INC: | |
5e3c5739 | 4761 | base = XEXP (XEXP (x, 0), 0); |
34940871 | 4762 | fprintf (file, "%d(%s)", size, reg_names [REGNO (base)]); |
87ad11b0 | 4763 | break; |
4764 | default: | |
27ef382d | 4765 | if (GET_CODE (XEXP (x, 0)) == PLUS |
4766 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT) | |
34940871 | 4767 | fprintf (file, "%s(%s)", |
27ef382d | 4768 | reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 0), 0))], |
4769 | reg_names [REGNO (XEXP (XEXP (x, 0), 1))]); | |
4770 | else if (GET_CODE (XEXP (x, 0)) == PLUS | |
4771 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT) | |
34940871 | 4772 | fprintf (file, "%s(%s)", |
27ef382d | 4773 | reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 1), 0))], |
4774 | reg_names [REGNO (XEXP (XEXP (x, 0), 0))]); | |
4775 | else | |
4776 | output_address (XEXP (x, 0)); | |
87ad11b0 | 4777 | break; |
4778 | } | |
4779 | } | |
87ad11b0 | 4780 | else |
4781 | output_addr_const (file, x); | |
4782 | } | |
4783 | ||
6dc3b0d9 | 4784 | /* output a SYMBOL_REF or a CONST expression involving a SYMBOL_REF. */ |
87ad11b0 | 4785 | |
4786 | void | |
f9333726 | 4787 | output_global_address (file, x, round_constant) |
87ad11b0 | 4788 | FILE *file; |
4789 | rtx x; | |
f9333726 | 4790 | int round_constant; |
87ad11b0 | 4791 | { |
2ee034bc | 4792 | |
4793 | /* Imagine (high (const (plus ...))). */ | |
4794 | if (GET_CODE (x) == HIGH) | |
4795 | x = XEXP (x, 0); | |
4796 | ||
611a88e1 | 4797 | if (GET_CODE (x) == SYMBOL_REF && read_only_operand (x, VOIDmode)) |
87ad11b0 | 4798 | assemble_name (file, XSTR (x, 0)); |
b4a7bf10 | 4799 | else if (GET_CODE (x) == SYMBOL_REF && !flag_pic) |
87ad11b0 | 4800 | { |
4801 | assemble_name (file, XSTR (x, 0)); | |
9c0ac0fd | 4802 | fputs ("-$global$", file); |
87ad11b0 | 4803 | } |
4804 | else if (GET_CODE (x) == CONST) | |
4805 | { | |
611a88e1 | 4806 | const char *sep = ""; |
87ad11b0 | 4807 | int offset = 0; /* assembler wants -$global$ at end */ |
33ae0dba | 4808 | rtx base = NULL_RTX; |
6d36483b | 4809 | |
87ad11b0 | 4810 | if (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) |
4811 | { | |
4812 | base = XEXP (XEXP (x, 0), 0); | |
4813 | output_addr_const (file, base); | |
4814 | } | |
4815 | else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == CONST_INT) | |
4816 | offset = INTVAL (XEXP (XEXP (x, 0), 0)); | |
4817 | else abort (); | |
4818 | ||
4819 | if (GET_CODE (XEXP (XEXP (x, 0), 1)) == SYMBOL_REF) | |
4820 | { | |
4821 | base = XEXP (XEXP (x, 0), 1); | |
4822 | output_addr_const (file, base); | |
4823 | } | |
4824 | else if (GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT) | |
ea52c577 | 4825 | offset = INTVAL (XEXP (XEXP (x, 0), 1)); |
87ad11b0 | 4826 | else abort (); |
4827 | ||
f9333726 | 4828 | /* How bogus. The compiler is apparently responsible for |
4829 | rounding the constant if it uses an LR field selector. | |
4830 | ||
4831 | The linker and/or assembler seem a better place since | |
4832 | they have to do this kind of thing already. | |
4833 | ||
4834 | If we fail to do this, HP's optimizing linker may eliminate | |
4835 | an addil, but not update the ldw/stw/ldo instruction that | |
4836 | uses the result of the addil. */ | |
4837 | if (round_constant) | |
4838 | offset = ((offset + 0x1000) & ~0x1fff); | |
4839 | ||
87ad11b0 | 4840 | if (GET_CODE (XEXP (x, 0)) == PLUS) |
4841 | { | |
4842 | if (offset < 0) | |
4843 | { | |
4844 | offset = -offset; | |
4845 | sep = "-"; | |
4846 | } | |
4847 | else | |
4848 | sep = "+"; | |
4849 | } | |
4850 | else if (GET_CODE (XEXP (x, 0)) == MINUS | |
4851 | && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
4852 | sep = "-"; | |
4853 | else abort (); | |
4854 | ||
611a88e1 | 4855 | if (!read_only_operand (base, VOIDmode) && !flag_pic) |
9c0ac0fd | 4856 | fputs ("-$global$", file); |
f9333726 | 4857 | if (offset) |
ea52c577 | 4858 | fprintf (file, "%s%d", sep, offset); |
87ad11b0 | 4859 | } |
4860 | else | |
4861 | output_addr_const (file, x); | |
4862 | } | |
4863 | ||
ece88821 | 4864 | static struct deferred_plabel * |
4865 | get_plabel (fname) | |
4866 | const char *fname; | |
4867 | { | |
4868 | size_t i; | |
4869 | ||
4870 | /* See if we have already put this function on the list of deferred | |
4871 | plabels. This list is generally small, so a liner search is not | |
4872 | too ugly. If it proves too slow replace it with something faster. */ | |
4873 | for (i = 0; i < n_deferred_plabels; i++) | |
4874 | if (strcmp (fname, deferred_plabels[i].name) == 0) | |
4875 | break; | |
4876 | ||
4877 | /* If the deferred plabel list is empty, or this entry was not found | |
4878 | on the list, create a new entry on the list. */ | |
4879 | if (deferred_plabels == NULL || i == n_deferred_plabels) | |
4880 | { | |
4881 | const char *real_name; | |
4882 | ||
4883 | if (deferred_plabels == 0) | |
4884 | deferred_plabels = (struct deferred_plabel *) | |
4885 | ggc_alloc (sizeof (struct deferred_plabel)); | |
4886 | else | |
4887 | deferred_plabels = (struct deferred_plabel *) | |
4888 | ggc_realloc (deferred_plabels, | |
4889 | ((n_deferred_plabels + 1) | |
4890 | * sizeof (struct deferred_plabel))); | |
4891 | ||
4892 | i = n_deferred_plabels++; | |
4893 | deferred_plabels[i].internal_label = gen_label_rtx (); | |
4894 | deferred_plabels[i].name = ggc_strdup (fname); | |
4895 | ||
4896 | /* Gross. We have just implicitly taken the address of this function, | |
4897 | mark it as such. */ | |
4898 | real_name = (*targetm.strip_name_encoding) (fname); | |
4899 | TREE_SYMBOL_REFERENCED (get_identifier (real_name)) = 1; | |
4900 | } | |
4901 | ||
4902 | return &deferred_plabels[i]; | |
4903 | } | |
4904 | ||
5cc6b2bc | 4905 | void |
4906 | output_deferred_plabels (file) | |
4907 | FILE *file; | |
4908 | { | |
e11bd7e5 | 4909 | size_t i; |
5cc6b2bc | 4910 | /* If we have deferred plabels, then we need to switch into the data |
4911 | section and align it to a 4 byte boundary before we output the | |
4912 | deferred plabels. */ | |
4913 | if (n_deferred_plabels) | |
4914 | { | |
4915 | data_section (); | |
ece88821 | 4916 | ASM_OUTPUT_ALIGN (file, TARGET_64BIT ? 3 : 2); |
5cc6b2bc | 4917 | } |
4918 | ||
4919 | /* Now output the deferred plabels. */ | |
4920 | for (i = 0; i < n_deferred_plabels; i++) | |
4921 | { | |
805e22b2 | 4922 | (*targetm.asm_out.internal_label) (file, "L", CODE_LABEL_NUMBER (deferred_plabels[i].internal_label)); |
09d688ff | 4923 | assemble_integer (gen_rtx_SYMBOL_REF (Pmode, deferred_plabels[i].name), |
b70ea764 | 4924 | TARGET_64BIT ? 8 : 4, TARGET_64BIT ? 64 : 32, 1); |
5cc6b2bc | 4925 | } |
4926 | } | |
4927 | ||
87ad11b0 | 4928 | /* HP's millicode routines mean something special to the assembler. |
4929 | Keep track of which ones we have used. */ | |
4930 | ||
c2271c34 | 4931 | enum millicodes { remI, remU, divI, divU, mulI, end1000 }; |
e750712b | 4932 | static void import_milli PARAMS ((enum millicodes)); |
ea52c577 | 4933 | static char imported[(int) end1000]; |
c2271c34 | 4934 | static const char * const milli_names[] = {"remI", "remU", "divI", "divU", "mulI"}; |
e99c3a1d | 4935 | static const char import_string[] = ".IMPORT $$....,MILLICODE"; |
87ad11b0 | 4936 | #define MILLI_START 10 |
4937 | ||
57ed30e5 | 4938 | static void |
87ad11b0 | 4939 | import_milli (code) |
4940 | enum millicodes code; | |
4941 | { | |
4942 | char str[sizeof (import_string)]; | |
6d36483b | 4943 | |
ea52c577 | 4944 | if (!imported[(int) code]) |
87ad11b0 | 4945 | { |
ea52c577 | 4946 | imported[(int) code] = 1; |
87ad11b0 | 4947 | strcpy (str, import_string); |
ea52c577 | 4948 | strncpy (str + MILLI_START, milli_names[(int) code], 4); |
87ad11b0 | 4949 | output_asm_insn (str, 0); |
4950 | } | |
4951 | } | |
4952 | ||
6d36483b | 4953 | /* The register constraints have put the operands and return value in |
6dc3b0d9 | 4954 | the proper registers. */ |
87ad11b0 | 4955 | |
611a88e1 | 4956 | const char * |
d6686e21 | 4957 | output_mul_insn (unsignedp, insn) |
b1ca791d | 4958 | int unsignedp ATTRIBUTE_UNUSED; |
d6686e21 | 4959 | rtx insn; |
87ad11b0 | 4960 | { |
d178f670 | 4961 | import_milli (mulI); |
440c23df | 4962 | return output_millicode_call (insn, gen_rtx_SYMBOL_REF (Pmode, "$$mulI")); |
87ad11b0 | 4963 | } |
4964 | ||
6dc3b0d9 | 4965 | /* Emit the rtl for doing a division by a constant. */ |
87ad11b0 | 4966 | |
d178f670 | 4967 | /* Do magic division millicodes exist for this value? */ |
e99c3a1d | 4968 | static const int magic_milli[]= {0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, |
4969 | 1, 1}; | |
87ad11b0 | 4970 | |
6d36483b | 4971 | /* We'll use an array to keep track of the magic millicodes and |
87ad11b0 | 4972 | whether or not we've used them already. [n][0] is signed, [n][1] is |
6dc3b0d9 | 4973 | unsigned. */ |
87ad11b0 | 4974 | |
87ad11b0 | 4975 | static int div_milli[16][2]; |
4976 | ||
4977 | int | |
4978 | div_operand (op, mode) | |
4979 | rtx op; | |
4980 | enum machine_mode mode; | |
4981 | { | |
4982 | return (mode == SImode | |
4983 | && ((GET_CODE (op) == REG && REGNO (op) == 25) | |
4984 | || (GET_CODE (op) == CONST_INT && INTVAL (op) > 0 | |
4985 | && INTVAL (op) < 16 && magic_milli[INTVAL (op)]))); | |
4986 | } | |
4987 | ||
4988 | int | |
4989 | emit_hpdiv_const (operands, unsignedp) | |
4990 | rtx *operands; | |
4991 | int unsignedp; | |
4992 | { | |
4993 | if (GET_CODE (operands[2]) == CONST_INT | |
4994 | && INTVAL (operands[2]) > 0 | |
4995 | && INTVAL (operands[2]) < 16 | |
4996 | && magic_milli[INTVAL (operands[2])]) | |
4997 | { | |
2013ddf6 | 4998 | rtx ret = gen_rtx_REG (SImode, TARGET_64BIT ? 2 : 31); |
4999 | ||
ad851752 | 5000 | emit_move_insn (gen_rtx_REG (SImode, 26), operands[1]); |
87ad11b0 | 5001 | emit |
5002 | (gen_rtx | |
5003 | (PARALLEL, VOIDmode, | |
c2078db8 | 5004 | gen_rtvec (6, gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, 29), |
ad851752 | 5005 | gen_rtx_fmt_ee (unsignedp ? UDIV : DIV, |
5006 | SImode, | |
5007 | gen_rtx_REG (SImode, 26), | |
5008 | operands[2])), | |
c2078db8 | 5009 | gen_rtx_CLOBBER (VOIDmode, operands[4]), |
ad851752 | 5010 | gen_rtx_CLOBBER (VOIDmode, operands[3]), |
5011 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 26)), | |
5012 | gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 25)), | |
2013ddf6 | 5013 | gen_rtx_CLOBBER (VOIDmode, ret)))); |
ad851752 | 5014 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 29)); |
87ad11b0 | 5015 | return 1; |
5016 | } | |
5017 | return 0; | |
5018 | } | |
5019 | ||
611a88e1 | 5020 | const char * |
d6686e21 | 5021 | output_div_insn (operands, unsignedp, insn) |
87ad11b0 | 5022 | rtx *operands; |
5023 | int unsignedp; | |
d6686e21 | 5024 | rtx insn; |
87ad11b0 | 5025 | { |
5026 | int divisor; | |
6d36483b | 5027 | |
5028 | /* If the divisor is a constant, try to use one of the special | |
87ad11b0 | 5029 | opcodes .*/ |
5030 | if (GET_CODE (operands[0]) == CONST_INT) | |
5031 | { | |
d6686e21 | 5032 | static char buf[100]; |
87ad11b0 | 5033 | divisor = INTVAL (operands[0]); |
5034 | if (!div_milli[divisor][unsignedp]) | |
5035 | { | |
d6686e21 | 5036 | div_milli[divisor][unsignedp] = 1; |
87ad11b0 | 5037 | if (unsignedp) |
5038 | output_asm_insn (".IMPORT $$divU_%0,MILLICODE", operands); | |
5039 | else | |
5040 | output_asm_insn (".IMPORT $$divI_%0,MILLICODE", operands); | |
87ad11b0 | 5041 | } |
5042 | if (unsignedp) | |
d6686e21 | 5043 | { |
96b529f2 | 5044 | sprintf (buf, "$$divU_"); |
5045 | sprintf (buf + 7, HOST_WIDE_INT_PRINT_DEC, INTVAL (operands[0])); | |
c7a4e712 | 5046 | return output_millicode_call (insn, |
ad851752 | 5047 | gen_rtx_SYMBOL_REF (SImode, buf)); |
d6686e21 | 5048 | } |
5049 | else | |
5050 | { | |
96b529f2 | 5051 | sprintf (buf, "$$divI_"); |
5052 | sprintf (buf + 7, HOST_WIDE_INT_PRINT_DEC, INTVAL (operands[0])); | |
c7a4e712 | 5053 | return output_millicode_call (insn, |
ad851752 | 5054 | gen_rtx_SYMBOL_REF (SImode, buf)); |
d6686e21 | 5055 | } |
87ad11b0 | 5056 | } |
6dc3b0d9 | 5057 | /* Divisor isn't a special constant. */ |
87ad11b0 | 5058 | else |
5059 | { | |
5060 | if (unsignedp) | |
5061 | { | |
5062 | import_milli (divU); | |
c7a4e712 | 5063 | return output_millicode_call (insn, |
ad851752 | 5064 | gen_rtx_SYMBOL_REF (SImode, "$$divU")); |
87ad11b0 | 5065 | } |
5066 | else | |
5067 | { | |
5068 | import_milli (divI); | |
c7a4e712 | 5069 | return output_millicode_call (insn, |
ad851752 | 5070 | gen_rtx_SYMBOL_REF (SImode, "$$divI")); |
87ad11b0 | 5071 | } |
5072 | } | |
5073 | } | |
5074 | ||
6dc3b0d9 | 5075 | /* Output a $$rem millicode to do mod. */ |
87ad11b0 | 5076 | |
611a88e1 | 5077 | const char * |
d6686e21 | 5078 | output_mod_insn (unsignedp, insn) |
87ad11b0 | 5079 | int unsignedp; |
d6686e21 | 5080 | rtx insn; |
87ad11b0 | 5081 | { |
5082 | if (unsignedp) | |
5083 | { | |
5084 | import_milli (remU); | |
c7a4e712 | 5085 | return output_millicode_call (insn, |
ad851752 | 5086 | gen_rtx_SYMBOL_REF (SImode, "$$remU")); |
87ad11b0 | 5087 | } |
5088 | else | |
5089 | { | |
5090 | import_milli (remI); | |
c7a4e712 | 5091 | return output_millicode_call (insn, |
ad851752 | 5092 | gen_rtx_SYMBOL_REF (SImode, "$$remI")); |
87ad11b0 | 5093 | } |
5094 | } | |
5095 | ||
5096 | void | |
df0651dc | 5097 | output_arg_descriptor (call_insn) |
5098 | rtx call_insn; | |
87ad11b0 | 5099 | { |
611a88e1 | 5100 | const char *arg_regs[4]; |
87ad11b0 | 5101 | enum machine_mode arg_mode; |
df0651dc | 5102 | rtx link; |
87ad11b0 | 5103 | int i, output_flag = 0; |
5104 | int regno; | |
6d36483b | 5105 | |
5e3c5739 | 5106 | /* We neither need nor want argument location descriptors for the |
5d0619cc | 5107 | 64bit runtime environment or the ELF32 environment. */ |
5108 | if (TARGET_64BIT || TARGET_ELF32) | |
5e3c5739 | 5109 | return; |
5110 | ||
87ad11b0 | 5111 | for (i = 0; i < 4; i++) |
5112 | arg_regs[i] = 0; | |
5113 | ||
738176ab | 5114 | /* Specify explicitly that no argument relocations should take place |
5115 | if using the portable runtime calling conventions. */ | |
5116 | if (TARGET_PORTABLE_RUNTIME) | |
5117 | { | |
9c0ac0fd | 5118 | fputs ("\t.CALL ARGW0=NO,ARGW1=NO,ARGW2=NO,ARGW3=NO,RETVAL=NO\n", |
5119 | asm_out_file); | |
738176ab | 5120 | return; |
5121 | } | |
5122 | ||
df0651dc | 5123 | if (GET_CODE (call_insn) != CALL_INSN) |
5124 | abort (); | |
5125 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); link; link = XEXP (link, 1)) | |
87ad11b0 | 5126 | { |
df0651dc | 5127 | rtx use = XEXP (link, 0); |
c12afafd | 5128 | |
df0651dc | 5129 | if (! (GET_CODE (use) == USE |
5130 | && GET_CODE (XEXP (use, 0)) == REG | |
5131 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
c12afafd | 5132 | continue; |
5133 | ||
df0651dc | 5134 | arg_mode = GET_MODE (XEXP (use, 0)); |
5135 | regno = REGNO (XEXP (use, 0)); | |
87ad11b0 | 5136 | if (regno >= 23 && regno <= 26) |
372ef038 | 5137 | { |
5138 | arg_regs[26 - regno] = "GR"; | |
5139 | if (arg_mode == DImode) | |
5140 | arg_regs[25 - regno] = "GR"; | |
5141 | } | |
df0651dc | 5142 | else if (regno >= 32 && regno <= 39) |
87ad11b0 | 5143 | { |
5144 | if (arg_mode == SFmode) | |
df0651dc | 5145 | arg_regs[(regno - 32) / 2] = "FR"; |
e6ba640e | 5146 | else |
87ad11b0 | 5147 | { |
eeec72c0 | 5148 | #ifndef HP_FP_ARG_DESCRIPTOR_REVERSED |
df0651dc | 5149 | arg_regs[(regno - 34) / 2] = "FR"; |
5150 | arg_regs[(regno - 34) / 2 + 1] = "FU"; | |
87ad11b0 | 5151 | #else |
df0651dc | 5152 | arg_regs[(regno - 34) / 2] = "FU"; |
5153 | arg_regs[(regno - 34) / 2 + 1] = "FR"; | |
87ad11b0 | 5154 | #endif |
5155 | } | |
87ad11b0 | 5156 | } |
5157 | } | |
5158 | fputs ("\t.CALL ", asm_out_file); | |
5159 | for (i = 0; i < 4; i++) | |
5160 | { | |
5161 | if (arg_regs[i]) | |
5162 | { | |
5163 | if (output_flag++) | |
5164 | fputc (',', asm_out_file); | |
5165 | fprintf (asm_out_file, "ARGW%d=%s", i, arg_regs[i]); | |
5166 | } | |
5167 | } | |
5168 | fputc ('\n', asm_out_file); | |
5169 | } | |
5170 | \f | |
9c0ac0fd | 5171 | /* Return the class of any secondary reload register that is needed to |
5172 | move IN into a register in class CLASS using mode MODE. | |
5173 | ||
5174 | Profiling has showed this routine and its descendants account for | |
5175 | a significant amount of compile time (~7%). So it has been | |
5176 | optimized to reduce redundant computations and eliminate useless | |
5177 | function calls. | |
5178 | ||
5179 | It might be worthwhile to try and make this a leaf function too. */ | |
87ad11b0 | 5180 | |
5181 | enum reg_class | |
5182 | secondary_reload_class (class, mode, in) | |
5183 | enum reg_class class; | |
5184 | enum machine_mode mode; | |
5185 | rtx in; | |
5186 | { | |
9c0ac0fd | 5187 | int regno, is_symbolic; |
87ad11b0 | 5188 | |
b4a7bf10 | 5189 | /* Trying to load a constant into a FP register during PIC code |
5190 | generation will require %r1 as a scratch register. */ | |
fc44315f | 5191 | if (flag_pic |
b4a7bf10 | 5192 | && GET_MODE_CLASS (mode) == MODE_INT |
5193 | && FP_REG_CLASS_P (class) | |
5194 | && (GET_CODE (in) == CONST_INT || GET_CODE (in) == CONST_DOUBLE)) | |
5195 | return R1_REGS; | |
5196 | ||
9c0ac0fd | 5197 | /* Profiling showed the PA port spends about 1.3% of its compilation |
5198 | time in true_regnum from calls inside secondary_reload_class. */ | |
5199 | ||
5200 | if (GET_CODE (in) == REG) | |
5201 | { | |
5202 | regno = REGNO (in); | |
5203 | if (regno >= FIRST_PSEUDO_REGISTER) | |
5204 | regno = true_regnum (in); | |
5205 | } | |
5206 | else if (GET_CODE (in) == SUBREG) | |
5207 | regno = true_regnum (in); | |
9c0ac0fd | 5208 | else |
5209 | regno = -1; | |
5210 | ||
76a0ced5 | 5211 | /* If we have something like (mem (mem (...)), we can safely assume the |
5212 | inner MEM will end up in a general register after reloading, so there's | |
5213 | no need for a secondary reload. */ | |
5214 | if (GET_CODE (in) == MEM | |
5215 | && GET_CODE (XEXP (in, 0)) == MEM) | |
5216 | return NO_REGS; | |
5217 | ||
5218 | /* Handle out of range displacement for integer mode loads/stores of | |
5219 | FP registers. */ | |
d2498717 | 5220 | if (((regno >= FIRST_PSEUDO_REGISTER || regno == -1) |
6d36483b | 5221 | && GET_MODE_CLASS (mode) == MODE_INT |
5222 | && FP_REG_CLASS_P (class)) | |
d6f01525 | 5223 | || (class == SHIFT_REGS && (regno <= 0 || regno >= 32))) |
9c6d4825 | 5224 | return GENERAL_REGS; |
d2c1d63d | 5225 | |
7c4b32f3 | 5226 | /* A SAR<->FP register copy requires a secondary register (GPR) as |
5227 | well as secondary memory. */ | |
5228 | if (regno >= 0 && regno < FIRST_PSEUDO_REGISTER | |
5229 | && ((REGNO_REG_CLASS (regno) == SHIFT_REGS && FP_REG_CLASS_P (class)) | |
5230 | || (class == SHIFT_REGS && FP_REG_CLASS_P (REGNO_REG_CLASS (regno))))) | |
5231 | return GENERAL_REGS; | |
5232 | ||
2ee034bc | 5233 | if (GET_CODE (in) == HIGH) |
5234 | in = XEXP (in, 0); | |
5235 | ||
9c0ac0fd | 5236 | /* Profiling has showed GCC spends about 2.6% of its compilation |
5237 | time in symbolic_operand from calls inside secondary_reload_class. | |
5238 | ||
5239 | We use an inline copy and only compute its return value once to avoid | |
5240 | useless work. */ | |
5241 | switch (GET_CODE (in)) | |
5242 | { | |
5243 | rtx tmp; | |
5244 | ||
5245 | case SYMBOL_REF: | |
5246 | case LABEL_REF: | |
5247 | is_symbolic = 1; | |
5248 | break; | |
5249 | case CONST: | |
5250 | tmp = XEXP (in, 0); | |
5251 | is_symbolic = ((GET_CODE (XEXP (tmp, 0)) == SYMBOL_REF | |
5252 | || GET_CODE (XEXP (tmp, 0)) == LABEL_REF) | |
5253 | && GET_CODE (XEXP (tmp, 1)) == CONST_INT); | |
5254 | break; | |
76a0ced5 | 5255 | |
9c0ac0fd | 5256 | default: |
5257 | is_symbolic = 0; | |
5258 | break; | |
5259 | } | |
9840d99d | 5260 | |
b4a7bf10 | 5261 | if (!flag_pic |
9c0ac0fd | 5262 | && is_symbolic |
611a88e1 | 5263 | && read_only_operand (in, VOIDmode)) |
b4a7bf10 | 5264 | return NO_REGS; |
5265 | ||
9c0ac0fd | 5266 | if (class != R1_REGS && is_symbolic) |
2ee034bc | 5267 | return R1_REGS; |
5268 | ||
d2c1d63d | 5269 | return NO_REGS; |
87ad11b0 | 5270 | } |
5271 | ||
5272 | enum direction | |
5273 | function_arg_padding (mode, type) | |
5274 | enum machine_mode mode; | |
5275 | tree type; | |
5276 | { | |
ac965869 | 5277 | if (mode == BLKmode |
5278 | || (TARGET_64BIT && type && AGGREGATE_TYPE_P (type))) | |
5279 | { | |
5280 | /* Return none if justification is not required. */ | |
5281 | if (type | |
5282 | && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST | |
5283 | && (int_size_in_bytes (type) * BITS_PER_UNIT) % PARM_BOUNDARY == 0) | |
5284 | return none; | |
5285 | ||
5286 | /* The directions set here are ignored when a BLKmode argument larger | |
5287 | than a word is placed in a register. Different code is used for | |
5288 | the stack and registers. This makes it difficult to have a | |
5289 | consistent data representation for both the stack and registers. | |
5290 | For both runtimes, the justification and padding for arguments on | |
5291 | the stack and in registers should be identical. */ | |
5292 | if (TARGET_64BIT) | |
5293 | /* The 64-bit runtime specifies left justification for aggregates. */ | |
5294 | return upward; | |
87ad11b0 | 5295 | else |
ac965869 | 5296 | /* The 32-bit runtime architecture specifies right justification. |
5297 | When the argument is passed on the stack, the argument is padded | |
5298 | with garbage on the left. The HP compiler pads with zeros. */ | |
5299 | return downward; | |
87ad11b0 | 5300 | } |
ac965869 | 5301 | |
5302 | if (GET_MODE_BITSIZE (mode) < PARM_BOUNDARY) | |
87ad11b0 | 5303 | return downward; |
87ad11b0 | 5304 | else |
5305 | return none; | |
5306 | } | |
5307 | ||
87ad11b0 | 5308 | \f |
55f54832 | 5309 | /* Do what is necessary for `va_start'. We look at the current function |
5310 | to determine if stdargs or varargs is used and fill in an initial | |
5311 | va_list. A pointer to this constructor is returned. */ | |
87ad11b0 | 5312 | |
5313 | struct rtx_def * | |
55f54832 | 5314 | hppa_builtin_saveregs () |
87ad11b0 | 5315 | { |
01251cbc | 5316 | rtx offset, dest; |
87ad11b0 | 5317 | tree fntype = TREE_TYPE (current_function_decl); |
5318 | int argadj = ((!(TYPE_ARG_TYPES (fntype) != 0 | |
5319 | && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype))) | |
5320 | != void_type_node))) | |
5321 | ? UNITS_PER_WORD : 0); | |
5322 | ||
5323 | if (argadj) | |
5324 | offset = plus_constant (current_function_arg_offset_rtx, argadj); | |
5325 | else | |
5326 | offset = current_function_arg_offset_rtx; | |
9c6d4825 | 5327 | |
5e3c5739 | 5328 | if (TARGET_64BIT) |
5329 | { | |
5330 | int i, off; | |
9840d99d | 5331 | |
5e3c5739 | 5332 | /* Adjust for varargs/stdarg differences. */ |
5333 | if (argadj) | |
5334 | offset = plus_constant (current_function_arg_offset_rtx, -argadj); | |
5335 | else | |
5336 | offset = current_function_arg_offset_rtx; | |
5337 | ||
5338 | /* We need to save %r26 .. %r19 inclusive starting at offset -64 | |
5339 | from the incoming arg pointer and growing to larger addresses. */ | |
5340 | for (i = 26, off = -64; i >= 19; i--, off += 8) | |
5341 | emit_move_insn (gen_rtx_MEM (word_mode, | |
5342 | plus_constant (arg_pointer_rtx, off)), | |
5343 | gen_rtx_REG (word_mode, i)); | |
5344 | ||
5345 | /* The incoming args pointer points just beyond the flushback area; | |
8ef587dc | 5346 | normally this is not a serious concern. However, when we are doing |
5e3c5739 | 5347 | varargs/stdargs we want to make the arg pointer point to the start |
5348 | of the incoming argument area. */ | |
5349 | emit_move_insn (virtual_incoming_args_rtx, | |
5350 | plus_constant (arg_pointer_rtx, -64)); | |
5351 | ||
5352 | /* Now return a pointer to the first anonymous argument. */ | |
5353 | return copy_to_reg (expand_binop (Pmode, add_optab, | |
5354 | virtual_incoming_args_rtx, | |
5355 | offset, 0, 0, OPTAB_LIB_WIDEN)); | |
5356 | } | |
5357 | ||
6dc3b0d9 | 5358 | /* Store general registers on the stack. */ |
ad851752 | 5359 | dest = gen_rtx_MEM (BLKmode, |
5360 | plus_constant (current_function_internal_arg_pointer, | |
5361 | -16)); | |
ab6ab77e | 5362 | set_mem_alias_set (dest, get_varargs_alias_set ()); |
2a631e19 | 5363 | set_mem_align (dest, BITS_PER_WORD); |
01251cbc | 5364 | move_block_from_reg (23, dest, 4, 4 * UNITS_PER_WORD); |
5365 | ||
76a0ced5 | 5366 | /* move_block_from_reg will emit code to store the argument registers |
5367 | individually as scalar stores. | |
5368 | ||
5369 | However, other insns may later load from the same addresses for | |
ad87de1e | 5370 | a structure load (passing a struct to a varargs routine). |
76a0ced5 | 5371 | |
5372 | The alias code assumes that such aliasing can never happen, so we | |
5373 | have to keep memory referencing insns from moving up beyond the | |
5374 | last argument register store. So we emit a blockage insn here. */ | |
5375 | emit_insn (gen_blockage ()); | |
5376 | ||
9c6d4825 | 5377 | return copy_to_reg (expand_binop (Pmode, add_optab, |
5378 | current_function_internal_arg_pointer, | |
5379 | offset, 0, 0, OPTAB_LIB_WIDEN)); | |
87ad11b0 | 5380 | } |
d6f01525 | 5381 | |
72899a61 | 5382 | void |
7df226a2 | 5383 | hppa_va_start (valist, nextarg) |
72899a61 | 5384 | tree valist; |
5385 | rtx nextarg; | |
5386 | { | |
5387 | nextarg = expand_builtin_saveregs (); | |
7df226a2 | 5388 | std_expand_builtin_va_start (valist, nextarg); |
72899a61 | 5389 | } |
5390 | ||
5391 | rtx | |
5392 | hppa_va_arg (valist, type) | |
5393 | tree valist, type; | |
5394 | { | |
ac965869 | 5395 | HOST_WIDE_INT size = int_size_in_bytes (type); |
5396 | HOST_WIDE_INT ofs; | |
72899a61 | 5397 | tree t, ptr, pptr; |
5398 | ||
5e3c5739 | 5399 | if (TARGET_64BIT) |
5400 | { | |
ac965869 | 5401 | /* Every argument in PA64 is supposed to be passed by value |
5402 | (including large structs). However, as a GCC extension, we | |
5403 | pass zero and variable sized arguments by reference. Empty | |
5404 | structures are a GCC extension not supported by the HP | |
5405 | compilers. Thus, passing them by reference isn't likely | |
5406 | to conflict with the ABI. For variable sized arguments, | |
5407 | GCC doesn't have the infrastructure to allocate these to | |
5408 | registers. */ | |
5409 | ||
5410 | /* Arguments with a size greater than 8 must be aligned 0 MOD 16. */ | |
5e3c5739 | 5411 | |
5e3c5739 | 5412 | if (size > UNITS_PER_WORD) |
5413 | { | |
5414 | t = build (PLUS_EXPR, TREE_TYPE (valist), valist, | |
5415 | build_int_2 (2 * UNITS_PER_WORD - 1, 0)); | |
9840d99d | 5416 | t = build (BIT_AND_EXPR, TREE_TYPE (t), t, |
5e3c5739 | 5417 | build_int_2 (-2 * UNITS_PER_WORD, -1)); |
5418 | t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t); | |
5419 | TREE_SIDE_EFFECTS (t) = 1; | |
ac965869 | 5420 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); |
5e3c5739 | 5421 | } |
72899a61 | 5422 | |
ac965869 | 5423 | if (size > 0) |
5424 | return std_expand_builtin_va_arg (valist, type); | |
5425 | else | |
5426 | { | |
5427 | ptr = build_pointer_type (type); | |
72899a61 | 5428 | |
ac965869 | 5429 | /* Args grow upward. */ |
5430 | t = build (POSTINCREMENT_EXPR, TREE_TYPE (valist), valist, | |
5431 | build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0)); | |
5432 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5433 | |
ac965869 | 5434 | pptr = build_pointer_type (ptr); |
5435 | t = build1 (NOP_EXPR, pptr, t); | |
5436 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5437 | |
ac965869 | 5438 | t = build1 (INDIRECT_REF, ptr, t); |
5439 | TREE_SIDE_EFFECTS (t) = 1; | |
5440 | } | |
72899a61 | 5441 | } |
ac965869 | 5442 | else /* !TARGET_64BIT */ |
72899a61 | 5443 | { |
ac965869 | 5444 | ptr = build_pointer_type (type); |
72899a61 | 5445 | |
ac965869 | 5446 | /* "Large" and variable sized types are passed by reference. */ |
5447 | if (size > 8 || size <= 0) | |
5448 | { | |
5449 | /* Args grow downward. */ | |
5450 | t = build (PREDECREMENT_EXPR, TREE_TYPE (valist), valist, | |
5451 | build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0)); | |
5452 | TREE_SIDE_EFFECTS (t) = 1; | |
72899a61 | 5453 | |
ac965869 | 5454 | pptr = build_pointer_type (ptr); |
5455 | t = build1 (NOP_EXPR, pptr, t); | |
5456 | TREE_SIDE_EFFECTS (t) = 1; | |
9840d99d | 5457 | |
ac965869 | 5458 | t = build1 (INDIRECT_REF, ptr, t); |
72899a61 | 5459 | TREE_SIDE_EFFECTS (t) = 1; |
5460 | } | |
ac965869 | 5461 | else |
5462 | { | |
5463 | t = build (PLUS_EXPR, TREE_TYPE (valist), valist, | |
5464 | build_int_2 (-size, -1)); | |
5465 | ||
5466 | /* Copied from va-pa.h, but we probably don't need to align to | |
5467 | word size, since we generate and preserve that invariant. */ | |
5468 | t = build (BIT_AND_EXPR, TREE_TYPE (valist), t, | |
5469 | build_int_2 ((size > 4 ? -8 : -4), -1)); | |
5470 | ||
5471 | t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t); | |
5472 | TREE_SIDE_EFFECTS (t) = 1; | |
5473 | ||
5474 | ofs = (8 - size) % 4; | |
5475 | if (ofs) | |
5476 | { | |
5477 | t = build (PLUS_EXPR, TREE_TYPE (valist), t, | |
5478 | build_int_2 (ofs, 0)); | |
5479 | TREE_SIDE_EFFECTS (t) = 1; | |
5480 | } | |
72899a61 | 5481 | |
ac965869 | 5482 | t = build1 (NOP_EXPR, ptr, t); |
5483 | TREE_SIDE_EFFECTS (t) = 1; | |
5484 | } | |
72899a61 | 5485 | } |
5486 | ||
5487 | /* Calculate! */ | |
ac965869 | 5488 | return expand_expr (t, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
72899a61 | 5489 | } |
5490 | ||
5491 | ||
5492 | ||
6d36483b | 5493 | /* This routine handles all the normal conditional branch sequences we |
5494 | might need to generate. It handles compare immediate vs compare | |
5495 | register, nullification of delay slots, varying length branches, | |
0d986529 | 5496 | negated branches, and all combinations of the above. It returns the |
6d36483b | 5497 | output appropriate to emit the branch corresponding to all given |
0d986529 | 5498 | parameters. */ |
5499 | ||
611a88e1 | 5500 | const char * |
0d986529 | 5501 | output_cbranch (operands, nullify, length, negated, insn) |
ece88821 | 5502 | rtx *operands; |
5503 | int nullify, length, negated; | |
5504 | rtx insn; | |
29a4502c | 5505 | { |
0d986529 | 5506 | static char buf[100]; |
5507 | int useskip = 0; | |
5508 | ||
29a4502c | 5509 | /* A conditional branch to the following instruction (eg the delay slot) is |
5510 | asking for a disaster. This can happen when not optimizing. | |
5511 | ||
5512 | In such cases it is safe to emit nothing. */ | |
5513 | ||
db3da815 | 5514 | if (next_active_insn (JUMP_LABEL (insn)) == next_active_insn (insn)) |
29a4502c | 5515 | return ""; |
6d36483b | 5516 | |
5fbd5940 | 5517 | /* If this is a long branch with its delay slot unfilled, set `nullify' |
5518 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 5519 | if (length == 8 && dbr_sequence_length () == 0) |
5fbd5940 | 5520 | nullify = 1; |
5521 | ||
5522 | /* If this is a short forward conditional branch which did not get | |
5523 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 5524 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
5fbd5940 | 5525 | nullify = forward_branch_p (insn); |
5526 | ||
6d36483b | 5527 | /* A forward branch over a single nullified insn can be done with a |
0d986529 | 5528 | comclr instruction. This avoids a single cycle penalty due to |
5529 | mis-predicted branch if we fall through (branch not taken). */ | |
5a1231ef | 5530 | if (length == 4 |
5fbd5940 | 5531 | && next_real_insn (insn) != 0 |
5a1231ef | 5532 | && get_attr_length (next_real_insn (insn)) == 4 |
5fbd5940 | 5533 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) |
0d986529 | 5534 | && nullify) |
5535 | useskip = 1; | |
5536 | ||
5537 | switch (length) | |
5538 | { | |
5fbd5940 | 5539 | /* All short conditional branches except backwards with an unfilled |
5540 | delay slot. */ | |
5a1231ef | 5541 | case 4: |
0d986529 | 5542 | if (useskip) |
e4065f95 | 5543 | strcpy (buf, "{com%I2clr,|cmp%I2clr,}"); |
0d986529 | 5544 | else |
e4065f95 | 5545 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5546 | if (GET_MODE (operands[1]) == DImode) |
5547 | strcat (buf, "*"); | |
0d986529 | 5548 | if (negated) |
5549 | strcat (buf, "%B3"); | |
5550 | else | |
5551 | strcat (buf, "%S3"); | |
5552 | if (useskip) | |
5a811d43 | 5553 | strcat (buf, " %2,%r1,%%r0"); |
0d986529 | 5554 | else if (nullify) |
9e49be0e | 5555 | strcat (buf, ",n %2,%r1,%0"); |
6d36483b | 5556 | else |
9e49be0e | 5557 | strcat (buf, " %2,%r1,%0"); |
0d986529 | 5558 | break; |
5559 | ||
87fcb603 | 5560 | /* All long conditionals. Note a short backward branch with an |
5fbd5940 | 5561 | unfilled delay slot is treated just like a long backward branch |
5562 | with an unfilled delay slot. */ | |
5a1231ef | 5563 | case 8: |
5fbd5940 | 5564 | /* Handle weird backwards branch with a filled delay slot |
5565 | with is nullified. */ | |
5566 | if (dbr_sequence_length () != 0 | |
5567 | && ! forward_branch_p (insn) | |
5568 | && nullify) | |
5569 | { | |
e4065f95 | 5570 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5571 | if (GET_MODE (operands[1]) == DImode) |
5572 | strcat (buf, "*"); | |
5fbd5940 | 5573 | if (negated) |
5574 | strcat (buf, "%S3"); | |
5575 | else | |
5576 | strcat (buf, "%B3"); | |
5a811d43 | 5577 | strcat (buf, ",n %2,%r1,.+12\n\tb %0"); |
5fbd5940 | 5578 | } |
43f0c1f2 | 5579 | /* Handle short backwards branch with an unfilled delay slot. |
5580 | Using a comb;nop rather than comiclr;bl saves 1 cycle for both | |
5581 | taken and untaken branches. */ | |
5582 | else if (dbr_sequence_length () == 0 | |
5583 | && ! forward_branch_p (insn) | |
47fc0706 | 5584 | && INSN_ADDRESSES_SET_P () |
5585 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
5586 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 5587 | { |
e4065f95 | 5588 | strcpy (buf, "{com%I2b,|cmp%I2b,}"); |
5e3c5739 | 5589 | if (GET_MODE (operands[1]) == DImode) |
5590 | strcat (buf, "*"); | |
43f0c1f2 | 5591 | if (negated) |
9e49be0e | 5592 | strcat (buf, "%B3 %2,%r1,%0%#"); |
43f0c1f2 | 5593 | else |
9e49be0e | 5594 | strcat (buf, "%S3 %2,%r1,%0%#"); |
43f0c1f2 | 5595 | } |
0d986529 | 5596 | else |
5fbd5940 | 5597 | { |
e4065f95 | 5598 | strcpy (buf, "{com%I2clr,|cmp%I2clr,}"); |
5e3c5739 | 5599 | if (GET_MODE (operands[1]) == DImode) |
5600 | strcat (buf, "*"); | |
5fbd5940 | 5601 | if (negated) |
5602 | strcat (buf, "%S3"); | |
5603 | else | |
5604 | strcat (buf, "%B3"); | |
5605 | if (nullify) | |
5a811d43 | 5606 | strcat (buf, " %2,%r1,%%r0\n\tb,n %0"); |
5fbd5940 | 5607 | else |
5a811d43 | 5608 | strcat (buf, " %2,%r1,%%r0\n\tb %0"); |
5fbd5940 | 5609 | } |
0d986529 | 5610 | break; |
5611 | ||
c8a0e52b | 5612 | case 20: |
5613 | /* Very long branch. Right now we only handle these when not | |
5614 | optimizing. See "jump" pattern in pa.md for details. */ | |
5615 | if (optimize) | |
5616 | abort (); | |
5617 | ||
5618 | /* Create a reversed conditional branch which branches around | |
5619 | the following insns. */ | |
5620 | if (negated) | |
e4065f95 | 5621 | strcpy (buf, "{com%I2b,%S3,n %2,%r1,.+20|cmp%I2b,%S3,n %2,%r1,.+20}"); |
c8a0e52b | 5622 | else |
e4065f95 | 5623 | strcpy (buf, "{com%I2b,%B3,n %2,%r1,.+20|cmp%I2b,%B3,n %2,%r1,.+20}"); |
5e3c5739 | 5624 | if (GET_MODE (operands[1]) == DImode) |
5625 | { | |
5626 | if (negated) | |
5627 | strcpy (buf, | |
5628 | "{com%I2b,*%S3,n %2,%r1,.+20|cmp%I2b,*%S3,n %2,%r1,.+20}"); | |
5629 | else | |
5630 | strcpy (buf, | |
5631 | "{com%I2b,*%B3,n %2,%r1,.+20|cmp%I2b,*%B3,n %2,%r1,.+20}"); | |
5632 | } | |
c8a0e52b | 5633 | output_asm_insn (buf, operands); |
5634 | ||
5635 | /* Output an insn to save %r1. */ | |
5636 | output_asm_insn ("stw %%r1,-16(%%r30)", operands); | |
5637 | ||
5638 | /* Now output a very long branch to the original target. */ | |
5639 | output_asm_insn ("ldil L'%l0,%%r1\n\tbe R'%l0(%%sr4,%%r1)", operands); | |
5640 | ||
5641 | /* Now restore the value of %r1 in the delay slot. We're not | |
5642 | optimizing so we know nothing else can be in the delay slot. */ | |
5643 | return "ldw -16(%%r30),%%r1"; | |
5644 | ||
5645 | case 28: | |
5646 | /* Very long branch when generating PIC code. Right now we only | |
5647 | handle these when not optimizing. See "jump" pattern in pa.md | |
5648 | for details. */ | |
5649 | if (optimize) | |
5650 | abort (); | |
5651 | ||
5652 | /* Create a reversed conditional branch which branches around | |
5653 | the following insns. */ | |
5654 | if (negated) | |
e4065f95 | 5655 | strcpy (buf, "{com%I2b,%S3,n %2,%r1,.+28|cmp%I2b,%S3,n %2,%r1,.+28}"); |
c8a0e52b | 5656 | else |
e4065f95 | 5657 | strcpy (buf, "{com%I2b,%B3,n %2,%r1,.+28|cmp%I2b,%B3,n %2,%r1,.+28}"); |
5e3c5739 | 5658 | if (GET_MODE (operands[1]) == DImode) |
5659 | { | |
5660 | if (negated) | |
5661 | strcpy (buf, "{com%I2b,*%S3,n %2,%r1,.+28|cmp%I2b,*%S3,n %2,%r1,.+28}"); | |
5662 | else | |
5663 | strcpy (buf, "{com%I2b,*%B3,n %2,%r1,.+28|cmp%I2b,*%B3,n %2,%r1,.+28}"); | |
5664 | } | |
c8a0e52b | 5665 | output_asm_insn (buf, operands); |
5666 | ||
5667 | /* Output an insn to save %r1. */ | |
5668 | output_asm_insn ("stw %%r1,-16(%%r30)", operands); | |
5669 | ||
5670 | /* Now output a very long PIC branch to the original target. */ | |
5671 | { | |
5672 | rtx xoperands[5]; | |
5673 | ||
5674 | xoperands[0] = operands[0]; | |
5675 | xoperands[1] = operands[1]; | |
5676 | xoperands[2] = operands[2]; | |
5677 | xoperands[3] = operands[3]; | |
b70ea764 | 5678 | |
5679 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
ece88821 | 5680 | if (TARGET_SOM || !TARGET_GAS) |
b70ea764 | 5681 | { |
ece88821 | 5682 | xoperands[4] = gen_label_rtx (); |
b70ea764 | 5683 | output_asm_insn ("addil L'%l0-%l4,%%r1", xoperands); |
805e22b2 | 5684 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
b70ea764 | 5685 | CODE_LABEL_NUMBER (xoperands[4])); |
5686 | output_asm_insn ("ldo R'%l0-%l4(%%r1),%%r1", xoperands); | |
5687 | } | |
5688 | else | |
5689 | { | |
5690 | output_asm_insn ("addil L'%l0-$PIC_pcrel$0+4,%%r1", xoperands); | |
5691 | output_asm_insn ("ldo R'%l0-$PIC_pcrel$0+8(%%r1),%%r1", | |
5692 | xoperands); | |
5693 | } | |
5694 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
c8a0e52b | 5695 | } |
5696 | ||
5697 | /* Now restore the value of %r1 in the delay slot. We're not | |
5698 | optimizing so we know nothing else can be in the delay slot. */ | |
5699 | return "ldw -16(%%r30),%%r1"; | |
9840d99d | 5700 | |
0d986529 | 5701 | default: |
ea52c577 | 5702 | abort (); |
5fbd5940 | 5703 | } |
0d986529 | 5704 | return buf; |
5705 | } | |
5706 | ||
6d36483b | 5707 | /* This routine handles all the branch-on-bit conditional branch sequences we |
0d986529 | 5708 | might need to generate. It handles nullification of delay slots, |
5709 | varying length branches, negated branches and all combinations of the | |
5710 | above. it returns the appropriate output template to emit the branch. */ | |
5711 | ||
611a88e1 | 5712 | const char * |
0d986529 | 5713 | output_bb (operands, nullify, length, negated, insn, which) |
ece88821 | 5714 | rtx *operands ATTRIBUTE_UNUSED; |
5715 | int nullify, length, negated; | |
5716 | rtx insn; | |
5717 | int which; | |
29a4502c | 5718 | { |
0d986529 | 5719 | static char buf[100]; |
5720 | int useskip = 0; | |
5721 | ||
29a4502c | 5722 | /* A conditional branch to the following instruction (eg the delay slot) is |
5723 | asking for a disaster. I do not think this can happen as this pattern | |
6d36483b | 5724 | is only used when optimizing; jump optimization should eliminate the |
29a4502c | 5725 | jump. But be prepared just in case. */ |
6d36483b | 5726 | |
db3da815 | 5727 | if (next_active_insn (JUMP_LABEL (insn)) == next_active_insn (insn)) |
29a4502c | 5728 | return ""; |
6d36483b | 5729 | |
5fbd5940 | 5730 | /* If this is a long branch with its delay slot unfilled, set `nullify' |
5731 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 5732 | if (length == 8 && dbr_sequence_length () == 0) |
5fbd5940 | 5733 | nullify = 1; |
5734 | ||
5735 | /* If this is a short forward conditional branch which did not get | |
5736 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 5737 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
5fbd5940 | 5738 | nullify = forward_branch_p (insn); |
5739 | ||
6d36483b | 5740 | /* A forward branch over a single nullified insn can be done with a |
0d986529 | 5741 | extrs instruction. This avoids a single cycle penalty due to |
5742 | mis-predicted branch if we fall through (branch not taken). */ | |
5743 | ||
5a1231ef | 5744 | if (length == 4 |
5fbd5940 | 5745 | && next_real_insn (insn) != 0 |
5a1231ef | 5746 | && get_attr_length (next_real_insn (insn)) == 4 |
5fbd5940 | 5747 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) |
0d986529 | 5748 | && nullify) |
5749 | useskip = 1; | |
5750 | ||
5751 | switch (length) | |
5752 | { | |
5753 | ||
5fbd5940 | 5754 | /* All short conditional branches except backwards with an unfilled |
5755 | delay slot. */ | |
5a1231ef | 5756 | case 4: |
0d986529 | 5757 | if (useskip) |
e4065f95 | 5758 | strcpy (buf, "{extrs,|extrw,s,}"); |
6d36483b | 5759 | else |
0d986529 | 5760 | strcpy (buf, "bb,"); |
5e3c5739 | 5761 | if (useskip && GET_MODE (operands[0]) == DImode) |
5762 | strcpy (buf, "extrd,s,*"); | |
5763 | else if (GET_MODE (operands[0]) == DImode) | |
5764 | strcpy (buf, "bb,*"); | |
0d986529 | 5765 | if ((which == 0 && negated) |
5766 | || (which == 1 && ! negated)) | |
5767 | strcat (buf, ">="); | |
5768 | else | |
5769 | strcat (buf, "<"); | |
5770 | if (useskip) | |
5a811d43 | 5771 | strcat (buf, " %0,%1,1,%%r0"); |
0d986529 | 5772 | else if (nullify && negated) |
5773 | strcat (buf, ",n %0,%1,%3"); | |
5774 | else if (nullify && ! negated) | |
5775 | strcat (buf, ",n %0,%1,%2"); | |
5776 | else if (! nullify && negated) | |
5fbd5940 | 5777 | strcat (buf, "%0,%1,%3"); |
0d986529 | 5778 | else if (! nullify && ! negated) |
5fbd5940 | 5779 | strcat (buf, " %0,%1,%2"); |
0d986529 | 5780 | break; |
5781 | ||
87fcb603 | 5782 | /* All long conditionals. Note a short backward branch with an |
5fbd5940 | 5783 | unfilled delay slot is treated just like a long backward branch |
5784 | with an unfilled delay slot. */ | |
5a1231ef | 5785 | case 8: |
5fbd5940 | 5786 | /* Handle weird backwards branch with a filled delay slot |
5787 | with is nullified. */ | |
5788 | if (dbr_sequence_length () != 0 | |
5789 | && ! forward_branch_p (insn) | |
5790 | && nullify) | |
5791 | { | |
5792 | strcpy (buf, "bb,"); | |
5e3c5739 | 5793 | if (GET_MODE (operands[0]) == DImode) |
5794 | strcat (buf, "*"); | |
5fbd5940 | 5795 | if ((which == 0 && negated) |
5796 | || (which == 1 && ! negated)) | |
5797 | strcat (buf, "<"); | |
5798 | else | |
5799 | strcat (buf, ">="); | |
5800 | if (negated) | |
5a811d43 | 5801 | strcat (buf, ",n %0,%1,.+12\n\tb %3"); |
5fbd5940 | 5802 | else |
5a811d43 | 5803 | strcat (buf, ",n %0,%1,.+12\n\tb %2"); |
5fbd5940 | 5804 | } |
43f0c1f2 | 5805 | /* Handle short backwards branch with an unfilled delay slot. |
5806 | Using a bb;nop rather than extrs;bl saves 1 cycle for both | |
5807 | taken and untaken branches. */ | |
5808 | else if (dbr_sequence_length () == 0 | |
5809 | && ! forward_branch_p (insn) | |
47fc0706 | 5810 | && INSN_ADDRESSES_SET_P () |
5811 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
5812 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 5813 | { |
5814 | strcpy (buf, "bb,"); | |
5e3c5739 | 5815 | if (GET_MODE (operands[0]) == DImode) |
5816 | strcat (buf, "*"); | |
43f0c1f2 | 5817 | if ((which == 0 && negated) |
5818 | || (which == 1 && ! negated)) | |
5819 | strcat (buf, ">="); | |
5820 | else | |
5821 | strcat (buf, "<"); | |
5822 | if (negated) | |
5823 | strcat (buf, " %0,%1,%3%#"); | |
5824 | else | |
5825 | strcat (buf, " %0,%1,%2%#"); | |
5826 | } | |
0d986529 | 5827 | else |
5fbd5940 | 5828 | { |
e4065f95 | 5829 | strcpy (buf, "{extrs,|extrw,s,}"); |
5e3c5739 | 5830 | if (GET_MODE (operands[0]) == DImode) |
5831 | strcpy (buf, "extrd,s,*"); | |
5fbd5940 | 5832 | if ((which == 0 && negated) |
5833 | || (which == 1 && ! negated)) | |
5834 | strcat (buf, "<"); | |
5835 | else | |
5836 | strcat (buf, ">="); | |
5837 | if (nullify && negated) | |
c6ae275c | 5838 | strcat (buf, " %0,%1,1,%%r0\n\tb,n %3"); |
5fbd5940 | 5839 | else if (nullify && ! negated) |
c6ae275c | 5840 | strcat (buf, " %0,%1,1,%%r0\n\tb,n %2"); |
5fbd5940 | 5841 | else if (negated) |
5a811d43 | 5842 | strcat (buf, " %0,%1,1,%%r0\n\tb %3"); |
6d36483b | 5843 | else |
5a811d43 | 5844 | strcat (buf, " %0,%1,1,%%r0\n\tb %2"); |
5fbd5940 | 5845 | } |
0d986529 | 5846 | break; |
5847 | ||
5848 | default: | |
ea52c577 | 5849 | abort (); |
5fbd5940 | 5850 | } |
0d986529 | 5851 | return buf; |
5852 | } | |
5853 | ||
c7a4e712 | 5854 | /* This routine handles all the branch-on-variable-bit conditional branch |
5855 | sequences we might need to generate. It handles nullification of delay | |
5856 | slots, varying length branches, negated branches and all combinations | |
5857 | of the above. it returns the appropriate output template to emit the | |
5858 | branch. */ | |
5859 | ||
611a88e1 | 5860 | const char * |
c7a4e712 | 5861 | output_bvb (operands, nullify, length, negated, insn, which) |
ece88821 | 5862 | rtx *operands ATTRIBUTE_UNUSED; |
5863 | int nullify, length, negated; | |
5864 | rtx insn; | |
5865 | int which; | |
c7a4e712 | 5866 | { |
5867 | static char buf[100]; | |
5868 | int useskip = 0; | |
5869 | ||
5870 | /* A conditional branch to the following instruction (eg the delay slot) is | |
5871 | asking for a disaster. I do not think this can happen as this pattern | |
5872 | is only used when optimizing; jump optimization should eliminate the | |
5873 | jump. But be prepared just in case. */ | |
5874 | ||
5875 | if (next_active_insn (JUMP_LABEL (insn)) == next_active_insn (insn)) | |
5876 | return ""; | |
5877 | ||
5878 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
5879 | as it can nullify the delay slot and save a nop. */ | |
5880 | if (length == 8 && dbr_sequence_length () == 0) | |
5881 | nullify = 1; | |
5882 | ||
5883 | /* If this is a short forward conditional branch which did not get | |
5884 | its delay slot filled, the delay slot can still be nullified. */ | |
5885 | if (! nullify && length == 4 && dbr_sequence_length () == 0) | |
5886 | nullify = forward_branch_p (insn); | |
5887 | ||
5888 | /* A forward branch over a single nullified insn can be done with a | |
5889 | extrs instruction. This avoids a single cycle penalty due to | |
5890 | mis-predicted branch if we fall through (branch not taken). */ | |
5891 | ||
5892 | if (length == 4 | |
5893 | && next_real_insn (insn) != 0 | |
5894 | && get_attr_length (next_real_insn (insn)) == 4 | |
5895 | && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn)) | |
5896 | && nullify) | |
5897 | useskip = 1; | |
5898 | ||
5899 | switch (length) | |
5900 | { | |
5901 | ||
5902 | /* All short conditional branches except backwards with an unfilled | |
5903 | delay slot. */ | |
5904 | case 4: | |
5905 | if (useskip) | |
e4065f95 | 5906 | strcpy (buf, "{vextrs,|extrw,s,}"); |
c7a4e712 | 5907 | else |
e4065f95 | 5908 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 5909 | if (useskip && GET_MODE (operands[0]) == DImode) |
5910 | strcpy (buf, "extrd,s,*}"); | |
5911 | else if (GET_MODE (operands[0]) == DImode) | |
5912 | strcpy (buf, "bb,*"); | |
c7a4e712 | 5913 | if ((which == 0 && negated) |
5914 | || (which == 1 && ! negated)) | |
5915 | strcat (buf, ">="); | |
5916 | else | |
5917 | strcat (buf, "<"); | |
5918 | if (useskip) | |
e4065f95 | 5919 | strcat (buf, "{ %0,1,%%r0| %0,%%sar,1,%%r0}"); |
c7a4e712 | 5920 | else if (nullify && negated) |
e4065f95 | 5921 | strcat (buf, "{,n %0,%3|,n %0,%%sar,%3}"); |
c7a4e712 | 5922 | else if (nullify && ! negated) |
e4065f95 | 5923 | strcat (buf, "{,n %0,%2|,n %0,%%sar,%2}"); |
c7a4e712 | 5924 | else if (! nullify && negated) |
e4065f95 | 5925 | strcat (buf, "{%0,%3|%0,%%sar,%3}"); |
c7a4e712 | 5926 | else if (! nullify && ! negated) |
e4065f95 | 5927 | strcat (buf, "{ %0,%2| %0,%%sar,%2}"); |
c7a4e712 | 5928 | break; |
5929 | ||
87fcb603 | 5930 | /* All long conditionals. Note a short backward branch with an |
c7a4e712 | 5931 | unfilled delay slot is treated just like a long backward branch |
5932 | with an unfilled delay slot. */ | |
5933 | case 8: | |
5934 | /* Handle weird backwards branch with a filled delay slot | |
5935 | with is nullified. */ | |
5936 | if (dbr_sequence_length () != 0 | |
5937 | && ! forward_branch_p (insn) | |
5938 | && nullify) | |
5939 | { | |
e4065f95 | 5940 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 5941 | if (GET_MODE (operands[0]) == DImode) |
5942 | strcat (buf, "*"); | |
c7a4e712 | 5943 | if ((which == 0 && negated) |
5944 | || (which == 1 && ! negated)) | |
5945 | strcat (buf, "<"); | |
5946 | else | |
5947 | strcat (buf, ">="); | |
5948 | if (negated) | |
e4065f95 | 5949 | strcat (buf, "{,n %0,.+12\n\tb %3|,n %0,%%sar,.+12\n\tb %3}"); |
c7a4e712 | 5950 | else |
e4065f95 | 5951 | strcat (buf, "{,n %0,.+12\n\tb %2|,n %0,%%sar,.+12\n\tb %2}"); |
c7a4e712 | 5952 | } |
5953 | /* Handle short backwards branch with an unfilled delay slot. | |
5954 | Using a bb;nop rather than extrs;bl saves 1 cycle for both | |
5955 | taken and untaken branches. */ | |
5956 | else if (dbr_sequence_length () == 0 | |
5957 | && ! forward_branch_p (insn) | |
47fc0706 | 5958 | && INSN_ADDRESSES_SET_P () |
5959 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
5960 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
c7a4e712 | 5961 | { |
e4065f95 | 5962 | strcpy (buf, "{bvb,|bb,}"); |
5e3c5739 | 5963 | if (GET_MODE (operands[0]) == DImode) |
5964 | strcat (buf, "*"); | |
c7a4e712 | 5965 | if ((which == 0 && negated) |
5966 | || (which == 1 && ! negated)) | |
5967 | strcat (buf, ">="); | |
5968 | else | |
5969 | strcat (buf, "<"); | |
5970 | if (negated) | |
e4065f95 | 5971 | strcat (buf, "{ %0,%3%#| %0,%%sar,%3%#}"); |
c7a4e712 | 5972 | else |
e4065f95 | 5973 | strcat (buf, "{ %0,%2%#| %0,%%sar,%2%#}"); |
c7a4e712 | 5974 | } |
5975 | else | |
5976 | { | |
e4065f95 | 5977 | strcpy (buf, "{vextrs,|extrw,s,}"); |
5e3c5739 | 5978 | if (GET_MODE (operands[0]) == DImode) |
5979 | strcpy (buf, "extrd,s,*"); | |
c7a4e712 | 5980 | if ((which == 0 && negated) |
5981 | || (which == 1 && ! negated)) | |
5982 | strcat (buf, "<"); | |
5983 | else | |
5984 | strcat (buf, ">="); | |
5985 | if (nullify && negated) | |
e4065f95 | 5986 | strcat (buf, "{ %0,1,%%r0\n\tb,n %3| %0,%%sar,1,%%r0\n\tb,n %3}"); |
c7a4e712 | 5987 | else if (nullify && ! negated) |
e4065f95 | 5988 | strcat (buf, "{ %0,1,%%r0\n\tb,n %2| %0,%%sar,1,%%r0\n\tb,n %2}"); |
c7a4e712 | 5989 | else if (negated) |
e4065f95 | 5990 | strcat (buf, "{ %0,1,%%r0\n\tb %3| %0,%%sar,1,%%r0\n\tb %3}"); |
c7a4e712 | 5991 | else |
e4065f95 | 5992 | strcat (buf, "{ %0,1,%%r0\n\tb %2| %0,%%sar,1,%%r0\n\tb %2}"); |
c7a4e712 | 5993 | } |
5994 | break; | |
5995 | ||
5996 | default: | |
ea52c577 | 5997 | abort (); |
c7a4e712 | 5998 | } |
5999 | return buf; | |
6000 | } | |
6001 | ||
29a4502c | 6002 | /* Return the output template for emitting a dbra type insn. |
6003 | ||
6004 | Note it may perform some output operations on its own before | |
6005 | returning the final output string. */ | |
611a88e1 | 6006 | const char * |
29a4502c | 6007 | output_dbra (operands, insn, which_alternative) |
6008 | rtx *operands; | |
6009 | rtx insn; | |
6010 | int which_alternative; | |
6011 | { | |
6012 | ||
6013 | /* A conditional branch to the following instruction (eg the delay slot) is | |
6014 | asking for a disaster. Be prepared! */ | |
6015 | ||
db3da815 | 6016 | if (next_active_insn (JUMP_LABEL (insn)) == next_active_insn (insn)) |
29a4502c | 6017 | { |
6018 | if (which_alternative == 0) | |
6019 | return "ldo %1(%0),%0"; | |
6020 | else if (which_alternative == 1) | |
6021 | { | |
ea52c577 | 6022 | output_asm_insn ("{fstws|fstw} %0,-16(%%r30)", operands); |
6023 | output_asm_insn ("ldw -16(%%r30),%4", operands); | |
34940871 | 6024 | output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands); |
e4065f95 | 6025 | return "{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6026 | } |
6027 | else | |
6028 | { | |
6029 | output_asm_insn ("ldw %0,%4", operands); | |
6030 | return "ldo %1(%4),%4\n\tstw %4,%0"; | |
6031 | } | |
6032 | } | |
6033 | ||
6034 | if (which_alternative == 0) | |
6035 | { | |
6036 | int nullify = INSN_ANNULLED_BRANCH_P (insn); | |
6037 | int length = get_attr_length (insn); | |
6038 | ||
6039 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
6040 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 6041 | if (length == 8 && dbr_sequence_length () == 0) |
29a4502c | 6042 | nullify = 1; |
6043 | ||
6044 | /* If this is a short forward conditional branch which did not get | |
6045 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 6046 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
29a4502c | 6047 | nullify = forward_branch_p (insn); |
6048 | ||
6049 | /* Handle short versions first. */ | |
5a1231ef | 6050 | if (length == 4 && nullify) |
29a4502c | 6051 | return "addib,%C2,n %1,%0,%3"; |
5a1231ef | 6052 | else if (length == 4 && ! nullify) |
29a4502c | 6053 | return "addib,%C2 %1,%0,%3"; |
5a1231ef | 6054 | else if (length == 8) |
29a4502c | 6055 | { |
6d36483b | 6056 | /* Handle weird backwards branch with a fulled delay slot |
29a4502c | 6057 | which is nullified. */ |
6058 | if (dbr_sequence_length () != 0 | |
6059 | && ! forward_branch_p (insn) | |
6060 | && nullify) | |
5a811d43 | 6061 | return "addib,%N2,n %1,%0,.+12\n\tb %3"; |
43f0c1f2 | 6062 | /* Handle short backwards branch with an unfilled delay slot. |
6063 | Using a addb;nop rather than addi;bl saves 1 cycle for both | |
6064 | taken and untaken branches. */ | |
6065 | else if (dbr_sequence_length () == 0 | |
6066 | && ! forward_branch_p (insn) | |
47fc0706 | 6067 | && INSN_ADDRESSES_SET_P () |
6068 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6069 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 6070 | return "addib,%C2 %1,%0,%3%#"; |
6d36483b | 6071 | |
6072 | /* Handle normal cases. */ | |
29a4502c | 6073 | if (nullify) |
5a811d43 | 6074 | return "addi,%N2 %1,%0,%0\n\tb,n %3"; |
29a4502c | 6075 | else |
5a811d43 | 6076 | return "addi,%N2 %1,%0,%0\n\tb %3"; |
29a4502c | 6077 | } |
6078 | else | |
ea52c577 | 6079 | abort (); |
29a4502c | 6080 | } |
6081 | /* Deal with gross reload from FP register case. */ | |
6082 | else if (which_alternative == 1) | |
6083 | { | |
6084 | /* Move loop counter from FP register to MEM then into a GR, | |
6085 | increment the GR, store the GR into MEM, and finally reload | |
6d36483b | 6086 | the FP register from MEM from within the branch's delay slot. */ |
ea52c577 | 6087 | output_asm_insn ("{fstws|fstw} %0,-16(%%r30)\n\tldw -16(%%r30),%4", |
6088 | operands); | |
34940871 | 6089 | output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands); |
5a1231ef | 6090 | if (get_attr_length (insn) == 24) |
e4065f95 | 6091 | return "{comb|cmpb},%S2 %%r0,%4,%3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6092 | else |
e4065f95 | 6093 | return "{comclr|cmpclr},%B2 %%r0,%4,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6094 | } |
6095 | /* Deal with gross reload from memory case. */ | |
6096 | else | |
6097 | { | |
6098 | /* Reload loop counter from memory, the store back to memory | |
6099 | happens in the branch's delay slot. */ | |
6100 | output_asm_insn ("ldw %0,%4", operands); | |
5a1231ef | 6101 | if (get_attr_length (insn) == 12) |
29a4502c | 6102 | return "addib,%C2 %1,%4,%3\n\tstw %4,%0"; |
6103 | else | |
5a811d43 | 6104 | return "addi,%N2 %1,%4,%4\n\tb %3\n\tstw %4,%0"; |
29a4502c | 6105 | } |
6106 | } | |
6107 | ||
6108 | /* Return the output template for emitting a dbra type insn. | |
6109 | ||
6110 | Note it may perform some output operations on its own before | |
6111 | returning the final output string. */ | |
611a88e1 | 6112 | const char * |
29a4502c | 6113 | output_movb (operands, insn, which_alternative, reverse_comparison) |
6114 | rtx *operands; | |
6115 | rtx insn; | |
6116 | int which_alternative; | |
6117 | int reverse_comparison; | |
6118 | { | |
6119 | ||
6120 | /* A conditional branch to the following instruction (eg the delay slot) is | |
6121 | asking for a disaster. Be prepared! */ | |
6122 | ||
db3da815 | 6123 | if (next_active_insn (JUMP_LABEL (insn)) == next_active_insn (insn)) |
29a4502c | 6124 | { |
6125 | if (which_alternative == 0) | |
6126 | return "copy %1,%0"; | |
6127 | else if (which_alternative == 1) | |
6128 | { | |
ea52c577 | 6129 | output_asm_insn ("stw %1,-16(%%r30)", operands); |
e4065f95 | 6130 | return "{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6131 | } |
546a40bd | 6132 | else if (which_alternative == 2) |
29a4502c | 6133 | return "stw %1,%0"; |
546a40bd | 6134 | else |
6135 | return "mtsar %r1"; | |
29a4502c | 6136 | } |
6137 | ||
6138 | /* Support the second variant. */ | |
6139 | if (reverse_comparison) | |
6140 | PUT_CODE (operands[2], reverse_condition (GET_CODE (operands[2]))); | |
6141 | ||
6142 | if (which_alternative == 0) | |
6143 | { | |
6144 | int nullify = INSN_ANNULLED_BRANCH_P (insn); | |
6145 | int length = get_attr_length (insn); | |
6146 | ||
6147 | /* If this is a long branch with its delay slot unfilled, set `nullify' | |
6148 | as it can nullify the delay slot and save a nop. */ | |
5a1231ef | 6149 | if (length == 8 && dbr_sequence_length () == 0) |
29a4502c | 6150 | nullify = 1; |
6151 | ||
6152 | /* If this is a short forward conditional branch which did not get | |
6153 | its delay slot filled, the delay slot can still be nullified. */ | |
5a1231ef | 6154 | if (! nullify && length == 4 && dbr_sequence_length () == 0) |
29a4502c | 6155 | nullify = forward_branch_p (insn); |
6156 | ||
6157 | /* Handle short versions first. */ | |
5a1231ef | 6158 | if (length == 4 && nullify) |
29a4502c | 6159 | return "movb,%C2,n %1,%0,%3"; |
5a1231ef | 6160 | else if (length == 4 && ! nullify) |
29a4502c | 6161 | return "movb,%C2 %1,%0,%3"; |
5a1231ef | 6162 | else if (length == 8) |
29a4502c | 6163 | { |
6d36483b | 6164 | /* Handle weird backwards branch with a filled delay slot |
29a4502c | 6165 | which is nullified. */ |
6166 | if (dbr_sequence_length () != 0 | |
6167 | && ! forward_branch_p (insn) | |
6168 | && nullify) | |
5a811d43 | 6169 | return "movb,%N2,n %1,%0,.+12\n\tb %3"; |
6d36483b | 6170 | |
43f0c1f2 | 6171 | /* Handle short backwards branch with an unfilled delay slot. |
6172 | Using a movb;nop rather than or;bl saves 1 cycle for both | |
6173 | taken and untaken branches. */ | |
6174 | else if (dbr_sequence_length () == 0 | |
6175 | && ! forward_branch_p (insn) | |
47fc0706 | 6176 | && INSN_ADDRESSES_SET_P () |
6177 | && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn))) | |
6178 | - INSN_ADDRESSES (INSN_UID (insn)) - 8)) | |
43f0c1f2 | 6179 | return "movb,%C2 %1,%0,%3%#"; |
6d36483b | 6180 | /* Handle normal cases. */ |
29a4502c | 6181 | if (nullify) |
5a811d43 | 6182 | return "or,%N2 %1,%%r0,%0\n\tb,n %3"; |
29a4502c | 6183 | else |
5a811d43 | 6184 | return "or,%N2 %1,%%r0,%0\n\tb %3"; |
29a4502c | 6185 | } |
6186 | else | |
ea52c577 | 6187 | abort (); |
29a4502c | 6188 | } |
6189 | /* Deal with gross reload from FP register case. */ | |
6190 | else if (which_alternative == 1) | |
6191 | { | |
6192 | /* Move loop counter from FP register to MEM then into a GR, | |
6193 | increment the GR, store the GR into MEM, and finally reload | |
6d36483b | 6194 | the FP register from MEM from within the branch's delay slot. */ |
ea52c577 | 6195 | output_asm_insn ("stw %1,-16(%%r30)", operands); |
5a1231ef | 6196 | if (get_attr_length (insn) == 12) |
e4065f95 | 6197 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6198 | else |
e4065f95 | 6199 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0"; |
29a4502c | 6200 | } |
6201 | /* Deal with gross reload from memory case. */ | |
546a40bd | 6202 | else if (which_alternative == 2) |
29a4502c | 6203 | { |
6204 | /* Reload loop counter from memory, the store back to memory | |
6205 | happens in the branch's delay slot. */ | |
5a1231ef | 6206 | if (get_attr_length (insn) == 8) |
e4065f95 | 6207 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\tstw %1,%0"; |
29a4502c | 6208 | else |
e4065f95 | 6209 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\tstw %1,%0"; |
29a4502c | 6210 | } |
546a40bd | 6211 | /* Handle SAR as a destination. */ |
6212 | else | |
6213 | { | |
6214 | if (get_attr_length (insn) == 8) | |
e4065f95 | 6215 | return "{comb|cmpb},%S2 %%r0,%1,%3\n\tmtsar %r1"; |
546a40bd | 6216 | else |
e4065f95 | 6217 | return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tbl %3\n\tmtsar %r1"; |
546a40bd | 6218 | } |
29a4502c | 6219 | } |
6220 | ||
ece88821 | 6221 | /* Copy any FP arguments in INSN into integer registers. */ |
6222 | static void | |
6223 | copy_fp_args (insn) | |
6224 | rtx insn; | |
6225 | { | |
6226 | rtx link; | |
6227 | rtx xoperands[2]; | |
29a4502c | 6228 | |
ece88821 | 6229 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) |
6230 | { | |
6231 | int arg_mode, regno; | |
6232 | rtx use = XEXP (link, 0); | |
3683f840 | 6233 | |
ece88821 | 6234 | if (! (GET_CODE (use) == USE |
6235 | && GET_CODE (XEXP (use, 0)) == REG | |
6236 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
6237 | continue; | |
d6686e21 | 6238 | |
ece88821 | 6239 | arg_mode = GET_MODE (XEXP (use, 0)); |
6240 | regno = REGNO (XEXP (use, 0)); | |
5e3c5739 | 6241 | |
ece88821 | 6242 | /* Is it a floating point register? */ |
6243 | if (regno >= 32 && regno <= 39) | |
6244 | { | |
6245 | /* Copy the FP register into an integer register via memory. */ | |
6246 | if (arg_mode == SFmode) | |
6247 | { | |
6248 | xoperands[0] = XEXP (use, 0); | |
6249 | xoperands[1] = gen_rtx_REG (SImode, 26 - (regno - 32) / 2); | |
6250 | output_asm_insn ("{fstws|fstw} %0,-16(%%sr0,%%r30)", xoperands); | |
6251 | output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands); | |
6252 | } | |
6253 | else | |
6254 | { | |
6255 | xoperands[0] = XEXP (use, 0); | |
6256 | xoperands[1] = gen_rtx_REG (DImode, 25 - (regno - 34) / 2); | |
6257 | output_asm_insn ("{fstds|fstd} %0,-16(%%sr0,%%r30)", xoperands); | |
6258 | output_asm_insn ("ldw -12(%%sr0,%%r30),%R1", xoperands); | |
6259 | output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands); | |
6260 | } | |
6261 | } | |
06ddb6f8 | 6262 | } |
ece88821 | 6263 | } |
6264 | ||
6265 | /* Compute length of the FP argument copy sequence for INSN. */ | |
6266 | static int | |
6267 | length_fp_args (insn) | |
6268 | rtx insn; | |
6269 | { | |
6270 | int length = 0; | |
6271 | rtx link; | |
06ddb6f8 | 6272 | |
ece88821 | 6273 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) |
c7a4e712 | 6274 | { |
ece88821 | 6275 | int arg_mode, regno; |
6276 | rtx use = XEXP (link, 0); | |
6277 | ||
6278 | if (! (GET_CODE (use) == USE | |
6279 | && GET_CODE (XEXP (use, 0)) == REG | |
6280 | && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0))))) | |
6281 | continue; | |
c7a4e712 | 6282 | |
ece88821 | 6283 | arg_mode = GET_MODE (XEXP (use, 0)); |
6284 | regno = REGNO (XEXP (use, 0)); | |
6285 | ||
6286 | /* Is it a floating point register? */ | |
6287 | if (regno >= 32 && regno <= 39) | |
c7a4e712 | 6288 | { |
ece88821 | 6289 | if (arg_mode == SFmode) |
6290 | length += 8; | |
6291 | else | |
6292 | length += 12; | |
c7a4e712 | 6293 | } |
ece88821 | 6294 | } |
c7a4e712 | 6295 | |
ece88821 | 6296 | return length; |
6297 | } | |
b70ea764 | 6298 | |
cd0dfcc5 | 6299 | /* Return the attribute length for the millicode call instruction INSN. |
6300 | The length must match the code generated by output_millicode_call. | |
6301 | We include the delay slot in the returned length as it is better to | |
ece88821 | 6302 | over estimate the length than to under estimate it. */ |
b29897dd | 6303 | |
ece88821 | 6304 | int |
cd0dfcc5 | 6305 | attr_length_millicode_call (insn) |
ece88821 | 6306 | rtx insn; |
ece88821 | 6307 | { |
cd0dfcc5 | 6308 | unsigned long distance = -1; |
ece88821 | 6309 | |
cd0dfcc5 | 6310 | if (INSN_ADDRESSES_SET_P ()) |
6311 | { | |
6312 | distance = (total_code_bytes + insn_current_reference_address (insn)); | |
6313 | if (distance < total_code_bytes) | |
6314 | distance = -1; | |
6315 | } | |
ece88821 | 6316 | |
6317 | if (TARGET_64BIT) | |
6318 | { | |
6319 | if (!TARGET_LONG_CALLS && distance < 7600000) | |
cd0dfcc5 | 6320 | return 8; |
ece88821 | 6321 | |
cd0dfcc5 | 6322 | return 20; |
ece88821 | 6323 | } |
6324 | else if (TARGET_PORTABLE_RUNTIME) | |
cd0dfcc5 | 6325 | return 24; |
ece88821 | 6326 | else |
6327 | { | |
6328 | if (!TARGET_LONG_CALLS && distance < 240000) | |
cd0dfcc5 | 6329 | return 8; |
ece88821 | 6330 | |
6331 | if (TARGET_LONG_ABS_CALL && !flag_pic) | |
cd0dfcc5 | 6332 | return 12; |
ece88821 | 6333 | |
cd0dfcc5 | 6334 | return 24; |
ece88821 | 6335 | } |
6336 | } | |
6337 | ||
6338 | /* INSN is a function call. It may have an unconditional jump | |
6339 | in its delay slot. | |
b29897dd | 6340 | |
ece88821 | 6341 | CALL_DEST is the routine we are calling. */ |
b29897dd | 6342 | |
ece88821 | 6343 | const char * |
6344 | output_millicode_call (insn, call_dest) | |
6345 | rtx insn; | |
6346 | rtx call_dest; | |
6347 | { | |
6348 | int attr_length = get_attr_length (insn); | |
6349 | int seq_length = dbr_sequence_length (); | |
6350 | int distance; | |
6351 | rtx seq_insn; | |
6352 | rtx xoperands[3]; | |
b29897dd | 6353 | |
ece88821 | 6354 | xoperands[0] = call_dest; |
6355 | xoperands[2] = gen_rtx_REG (Pmode, TARGET_64BIT ? 2 : 31); | |
6356 | ||
6357 | /* Handle the common case where we are sure that the branch will | |
6358 | reach the beginning of the $CODE$ subspace. The within reach | |
6359 | form of the $$sh_func_adrs call has a length of 28. Because | |
a8b24921 | 6360 | it has an attribute type of multi, it never has a nonzero |
ece88821 | 6361 | sequence length. The length of the $$sh_func_adrs is the same |
6362 | as certain out of reach PIC calls to other routines. */ | |
6363 | if (!TARGET_LONG_CALLS | |
6364 | && ((seq_length == 0 | |
6365 | && (attr_length == 12 | |
6366 | || (attr_length == 28 && get_attr_type (insn) == TYPE_MULTI))) | |
6367 | || (seq_length != 0 && attr_length == 8))) | |
6368 | { | |
6369 | output_asm_insn ("{bl|b,l} %0,%2", xoperands); | |
6370 | } | |
6371 | else | |
6372 | { | |
6373 | if (TARGET_64BIT) | |
6374 | { | |
6375 | /* It might seem that one insn could be saved by accessing | |
6376 | the millicode function using the linkage table. However, | |
6377 | this doesn't work in shared libraries and other dynamically | |
6378 | loaded objects. Using a pc-relative sequence also avoids | |
6379 | problems related to the implicit use of the gp register. */ | |
6380 | output_asm_insn ("b,l .+8,%%r1", xoperands); | |
9bd9af5d | 6381 | |
6382 | if (TARGET_GAS) | |
6383 | { | |
6384 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands); | |
6385 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands); | |
6386 | } | |
6387 | else | |
6388 | { | |
6389 | xoperands[1] = gen_label_rtx (); | |
6390 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
805e22b2 | 6391 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
9bd9af5d | 6392 | CODE_LABEL_NUMBER (xoperands[1])); |
6393 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); | |
6394 | } | |
6395 | ||
ece88821 | 6396 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); |
c7a4e712 | 6397 | } |
c7a4e712 | 6398 | else if (TARGET_PORTABLE_RUNTIME) |
6399 | { | |
ece88821 | 6400 | /* Pure portable runtime doesn't allow be/ble; we also don't |
6401 | have PIC support in the assembler/linker, so this sequence | |
6402 | is needed. */ | |
c7a4e712 | 6403 | |
ece88821 | 6404 | /* Get the address of our target into %r1. */ |
6405 | output_asm_insn ("ldil L'%0,%%r1", xoperands); | |
6406 | output_asm_insn ("ldo R'%0(%%r1),%%r1", xoperands); | |
c7a4e712 | 6407 | |
ece88821 | 6408 | /* Get our return address into %r31. */ |
6409 | output_asm_insn ("{bl|b,l} .+8,%%r31", xoperands); | |
6410 | output_asm_insn ("addi 8,%%r31,%%r31", xoperands); | |
c7a4e712 | 6411 | |
ece88821 | 6412 | /* Jump to our target address in %r1. */ |
6413 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
c7a4e712 | 6414 | } |
ece88821 | 6415 | else if (!flag_pic) |
c7a4e712 | 6416 | { |
ece88821 | 6417 | output_asm_insn ("ldil L'%0,%%r1", xoperands); |
356267e0 | 6418 | if (TARGET_PA_20) |
ece88821 | 6419 | output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31", xoperands); |
356267e0 | 6420 | else |
ece88821 | 6421 | output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands); |
c7a4e712 | 6422 | } |
ece88821 | 6423 | else |
c7a4e712 | 6424 | { |
9bd9af5d | 6425 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); |
6426 | output_asm_insn ("addi 16,%%r1,%%r31", xoperands); | |
6427 | ||
ece88821 | 6428 | if (TARGET_SOM || !TARGET_GAS) |
6429 | { | |
6430 | /* The HP assembler can generate relocations for the | |
6431 | difference of two symbols. GAS can do this for a | |
6432 | millicode symbol but not an arbitrary external | |
6433 | symbol when generating SOM output. */ | |
6434 | xoperands[1] = gen_label_rtx (); | |
805e22b2 | 6435 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
ece88821 | 6436 | CODE_LABEL_NUMBER (xoperands[1])); |
6437 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
6438 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); | |
6439 | } | |
6440 | else | |
6441 | { | |
ece88821 | 6442 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+8,%%r1", xoperands); |
6443 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+12(%%r1),%%r1", | |
6444 | xoperands); | |
6445 | } | |
c7a4e712 | 6446 | |
ece88821 | 6447 | /* Jump to our target address in %r1. */ |
6448 | output_asm_insn ("bv %%r0(%%r1)", xoperands); | |
c7a4e712 | 6449 | } |
c7a4e712 | 6450 | } |
6451 | ||
ece88821 | 6452 | if (seq_length == 0) |
6453 | output_asm_insn ("nop", xoperands); | |
c7a4e712 | 6454 | |
ece88821 | 6455 | /* We are done if there isn't a jump in the delay slot. */ |
6456 | if (seq_length == 0 || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN) | |
6457 | return ""; | |
c7a4e712 | 6458 | |
ece88821 | 6459 | /* This call has an unconditional jump in its delay slot. */ |
6460 | xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1); | |
c7a4e712 | 6461 | |
ece88821 | 6462 | /* See if the return address can be adjusted. Use the containing |
6463 | sequence insn's address. */ | |
cd0dfcc5 | 6464 | if (INSN_ADDRESSES_SET_P ()) |
c7a4e712 | 6465 | { |
cd0dfcc5 | 6466 | seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0))); |
6467 | distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn)))) | |
6468 | - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8); | |
6469 | ||
6470 | if (VAL_14_BITS_P (distance)) | |
6471 | { | |
6472 | xoperands[1] = gen_label_rtx (); | |
6473 | output_asm_insn ("ldo %0-%1(%2),%2", xoperands); | |
6474 | (*targetm.asm_out.internal_label) (asm_out_file, "L", | |
6475 | CODE_LABEL_NUMBER (xoperands[1])); | |
6476 | } | |
6477 | else | |
6478 | /* ??? This branch may not reach its target. */ | |
6479 | output_asm_insn ("nop\n\tb,n %0", xoperands); | |
c7a4e712 | 6480 | } |
ece88821 | 6481 | else |
6482 | /* ??? This branch may not reach its target. */ | |
6483 | output_asm_insn ("nop\n\tb,n %0", xoperands); | |
c7a4e712 | 6484 | |
6485 | /* Delete the jump. */ | |
6486 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6487 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6488 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
ece88821 | 6489 | |
c7a4e712 | 6490 | return ""; |
6491 | } | |
6492 | ||
cd0dfcc5 | 6493 | /* Return the attribute length of the call instruction INSN. The SIBCALL |
6494 | flag indicates whether INSN is a regular call or a sibling call. The | |
6495 | length must match the code generated by output_call. We include the delay | |
6496 | slot in the returned length as it is better to over estimate the length | |
6497 | than to under estimate it. */ | |
ece88821 | 6498 | |
6499 | int | |
6500 | attr_length_call (insn, sibcall) | |
6501 | rtx insn; | |
6502 | int sibcall; | |
6503 | { | |
cd0dfcc5 | 6504 | unsigned long distance = -1; |
ece88821 | 6505 | |
cd0dfcc5 | 6506 | if (INSN_ADDRESSES_SET_P ()) |
6507 | { | |
6508 | distance = (total_code_bytes + insn_current_reference_address (insn)); | |
6509 | if (distance < total_code_bytes) | |
6510 | distance = -1; | |
6511 | } | |
ece88821 | 6512 | |
6513 | if (TARGET_64BIT) | |
6514 | { | |
6515 | if (!TARGET_LONG_CALLS | |
6516 | && ((!sibcall && distance < 7600000) || distance < 240000)) | |
6517 | return 8; | |
6518 | ||
6519 | return (sibcall ? 28 : 24); | |
6520 | } | |
6521 | else | |
6522 | { | |
6523 | if (!TARGET_LONG_CALLS | |
6524 | && ((TARGET_PA_20 && !sibcall && distance < 7600000) | |
6525 | || distance < 240000)) | |
6526 | return 8; | |
6527 | ||
6528 | if (TARGET_LONG_ABS_CALL && !flag_pic) | |
6529 | return 12; | |
6530 | ||
6531 | if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) | |
6532 | || (TARGET_GAS && TARGET_LONG_PIC_PCREL_CALL)) | |
6533 | { | |
6534 | if (TARGET_PA_20) | |
6535 | return 20; | |
6536 | ||
6537 | return 28; | |
6538 | } | |
6539 | else | |
6540 | { | |
6541 | int length = 0; | |
6542 | ||
6543 | if (TARGET_SOM) | |
6544 | length += length_fp_args (insn); | |
6545 | ||
6546 | if (flag_pic) | |
6547 | length += 4; | |
6548 | ||
6549 | if (TARGET_PA_20) | |
6550 | return (length + 32); | |
6551 | ||
ee376abe | 6552 | if (!TARGET_NO_SPACE_REGS) |
6553 | length += 8; | |
6554 | ||
ece88821 | 6555 | if (!sibcall) |
6556 | length += 8; | |
6557 | ||
ee376abe | 6558 | return (length + 32); |
ece88821 | 6559 | } |
6560 | } | |
6561 | } | |
6562 | ||
6563 | /* INSN is a function call. It may have an unconditional jump | |
c7a4e712 | 6564 | in its delay slot. |
6565 | ||
6566 | CALL_DEST is the routine we are calling. */ | |
6567 | ||
611a88e1 | 6568 | const char * |
5e3c5739 | 6569 | output_call (insn, call_dest, sibcall) |
ece88821 | 6570 | rtx insn; |
6571 | rtx call_dest; | |
6572 | int sibcall; | |
c7a4e712 | 6573 | { |
ece88821 | 6574 | int delay_insn_deleted = 0; |
6575 | int delay_slot_filled = 0; | |
b70ea764 | 6576 | int seq_length = dbr_sequence_length (); |
ece88821 | 6577 | rtx xoperands[2]; |
6578 | ||
6579 | xoperands[0] = call_dest; | |
c7a4e712 | 6580 | |
ece88821 | 6581 | /* Handle the common case where we're sure that the branch will reach |
6582 | the beginning of the $CODE$ subspace. */ | |
cd0dfcc5 | 6583 | if (!TARGET_LONG_CALLS && attr_length_call (insn, sibcall) == 8) |
d6686e21 | 6584 | { |
5e3c5739 | 6585 | xoperands[1] = gen_rtx_REG (word_mode, sibcall ? 0 : 2); |
ece88821 | 6586 | output_asm_insn ("{bl|b,l} %0,%1", xoperands); |
06ddb6f8 | 6587 | } |
ece88821 | 6588 | else |
06ddb6f8 | 6589 | { |
ece88821 | 6590 | if (TARGET_64BIT) |
3683f840 | 6591 | { |
ece88821 | 6592 | /* ??? As far as I can tell, the HP linker doesn't support the |
6593 | long pc-relative sequence described in the 64-bit runtime | |
6594 | architecture. So, we use a slightly longer indirect call. */ | |
6595 | struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0)); | |
6596 | ||
6597 | xoperands[0] = p->internal_label; | |
6598 | xoperands[1] = gen_label_rtx (); | |
6599 | ||
6600 | /* If this isn't a sibcall, we put the load of %r27 into the | |
6601 | delay slot. We can't do this in a sibcall as we don't | |
6602 | have a second call-clobbered scratch register available. */ | |
6603 | if (seq_length != 0 | |
6604 | && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN | |
6605 | && !sibcall) | |
6606 | { | |
6607 | final_scan_insn (NEXT_INSN (insn), asm_out_file, | |
6608 | optimize, 0, 0); | |
6609 | ||
6610 | /* Now delete the delay insn. */ | |
6611 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6612 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6613 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
6614 | delay_insn_deleted = 1; | |
6615 | } | |
06ddb6f8 | 6616 | |
ece88821 | 6617 | output_asm_insn ("addil LT'%0,%%r27", xoperands); |
6618 | output_asm_insn ("ldd RT'%0(%%r1),%%r1", xoperands); | |
6619 | output_asm_insn ("ldd 0(%%r1),%%r1", xoperands); | |
06ddb6f8 | 6620 | |
ece88821 | 6621 | if (sibcall) |
06ddb6f8 | 6622 | { |
ece88821 | 6623 | output_asm_insn ("ldd 24(%%r1),%%r27", xoperands); |
6624 | output_asm_insn ("ldd 16(%%r1),%%r1", xoperands); | |
6625 | output_asm_insn ("bve (%%r1)", xoperands); | |
6626 | } | |
6627 | else | |
6628 | { | |
6629 | output_asm_insn ("ldd 16(%%r1),%%r2", xoperands); | |
6630 | output_asm_insn ("bve,l (%%r2),%%r2", xoperands); | |
6631 | output_asm_insn ("ldd 24(%%r1),%%r27", xoperands); | |
6632 | delay_slot_filled = 1; | |
06ddb6f8 | 6633 | } |
6634 | } | |
ece88821 | 6635 | else |
e3f53689 | 6636 | { |
ece88821 | 6637 | int indirect_call = 0; |
6638 | ||
6639 | /* Emit a long call. There are several different sequences | |
6640 | of increasing length and complexity. In most cases, | |
6641 | they don't allow an instruction in the delay slot. */ | |
6642 | if (!(TARGET_LONG_ABS_CALL && !flag_pic) | |
6643 | && !(TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) | |
6644 | && !(TARGET_GAS && TARGET_LONG_PIC_PCREL_CALL)) | |
6645 | indirect_call = 1; | |
6646 | ||
6647 | if (seq_length != 0 | |
6648 | && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN | |
6649 | && !sibcall | |
6650 | && (!TARGET_PA_20 || indirect_call)) | |
5cc6b2bc | 6651 | { |
ece88821 | 6652 | /* A non-jump insn in the delay slot. By definition we can |
6653 | emit this insn before the call (and in fact before argument | |
6654 | relocating. */ | |
6655 | final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0, 0); | |
6656 | ||
6657 | /* Now delete the delay insn. */ | |
6658 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6659 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6660 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
6661 | delay_insn_deleted = 1; | |
5cc6b2bc | 6662 | } |
e3f53689 | 6663 | |
ece88821 | 6664 | if (TARGET_LONG_ABS_CALL && !flag_pic) |
5cc6b2bc | 6665 | { |
ece88821 | 6666 | /* This is the best sequence for making long calls in |
6667 | non-pic code. Unfortunately, GNU ld doesn't provide | |
6668 | the stub needed for external calls, and GAS's support | |
6669 | for this with the SOM linker is buggy. */ | |
6670 | output_asm_insn ("ldil L'%0,%%r1", xoperands); | |
6671 | if (sibcall) | |
6672 | output_asm_insn ("be R'%0(%%sr4,%%r1)", xoperands); | |
6673 | else | |
6674 | { | |
6675 | if (TARGET_PA_20) | |
6676 | output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31", | |
6677 | xoperands); | |
6678 | else | |
6679 | output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands); | |
c7a4e712 | 6680 | |
ece88821 | 6681 | output_asm_insn ("copy %%r31,%%r2", xoperands); |
6682 | delay_slot_filled = 1; | |
6683 | } | |
6684 | } | |
6685 | else | |
6686 | { | |
6687 | if (TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL) | |
b70ea764 | 6688 | { |
ece88821 | 6689 | /* The HP assembler and linker can handle relocations |
6690 | for the difference of two symbols. GAS and the HP | |
6691 | linker can't do this when one of the symbols is | |
6692 | external. */ | |
6693 | xoperands[1] = gen_label_rtx (); | |
6694 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
6695 | output_asm_insn ("addil L'%0-%l1,%%r1", xoperands); | |
805e22b2 | 6696 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
b70ea764 | 6697 | CODE_LABEL_NUMBER (xoperands[1])); |
ece88821 | 6698 | output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands); |
6699 | } | |
6700 | else if (TARGET_GAS && TARGET_LONG_PIC_PCREL_CALL) | |
b70ea764 | 6701 | { |
ece88821 | 6702 | /* GAS currently can't generate the relocations that |
6703 | are needed for the SOM linker under HP-UX using this | |
6704 | sequence. The GNU linker doesn't generate the stubs | |
6705 | that are needed for external calls on TARGET_ELF32 | |
6706 | with this sequence. For now, we have to use a | |
6707 | longer plabel sequence when using GAS. */ | |
6708 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
6709 | output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", | |
b70ea764 | 6710 | xoperands); |
ece88821 | 6711 | output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", |
b70ea764 | 6712 | xoperands); |
6713 | } | |
5e3c5739 | 6714 | else |
6715 | { | |
ece88821 | 6716 | /* Emit a long plabel-based call sequence. This is |
6717 | essentially an inline implementation of $$dyncall. | |
6718 | We don't actually try to call $$dyncall as this is | |
6719 | as difficult as calling the function itself. */ | |
6720 | struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0)); | |
6721 | ||
6722 | xoperands[0] = p->internal_label; | |
6723 | xoperands[1] = gen_label_rtx (); | |
6724 | ||
6725 | /* Since the call is indirect, FP arguments in registers | |
6726 | need to be copied to the general registers. Then, the | |
6727 | argument relocation stub will copy them back. */ | |
6728 | if (TARGET_SOM) | |
6729 | copy_fp_args (insn); | |
6730 | ||
6731 | if (flag_pic) | |
6732 | { | |
6733 | output_asm_insn ("addil LT'%0,%%r19", xoperands); | |
6734 | output_asm_insn ("ldw RT'%0(%%r1),%%r1", xoperands); | |
6735 | output_asm_insn ("ldw 0(%%r1),%%r1", xoperands); | |
6736 | } | |
6737 | else | |
6738 | { | |
6739 | output_asm_insn ("addil LR'%0-$global$,%%r27", | |
6740 | xoperands); | |
6741 | output_asm_insn ("ldw RR'%0-$global$(%%r1),%%r1", | |
6742 | xoperands); | |
6743 | } | |
06ddb6f8 | 6744 | |
ece88821 | 6745 | output_asm_insn ("bb,>=,n %%r1,30,.+16", xoperands); |
6746 | output_asm_insn ("depi 0,31,2,%%r1", xoperands); | |
6747 | output_asm_insn ("ldw 4(%%sr0,%%r1),%%r19", xoperands); | |
6748 | output_asm_insn ("ldw 0(%%sr0,%%r1),%%r1", xoperands); | |
c7a4e712 | 6749 | |
ece88821 | 6750 | if (!sibcall && !TARGET_PA_20) |
6751 | { | |
6752 | output_asm_insn ("{bl|b,l} .+8,%%r2", xoperands); | |
ee376abe | 6753 | if (TARGET_NO_SPACE_REGS) |
6754 | output_asm_insn ("addi 8,%%r2,%%r2", xoperands); | |
6755 | else | |
6756 | output_asm_insn ("addi 16,%%r2,%%r2", xoperands); | |
ece88821 | 6757 | } |
6758 | } | |
c7a4e712 | 6759 | |
ece88821 | 6760 | if (TARGET_PA_20) |
5e3c5739 | 6761 | { |
ece88821 | 6762 | if (sibcall) |
6763 | output_asm_insn ("bve (%%r1)", xoperands); | |
6764 | else | |
6765 | { | |
6766 | if (indirect_call) | |
6767 | { | |
6768 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); | |
6769 | output_asm_insn ("stw %%r2,-24(%%sp)", xoperands); | |
6770 | delay_slot_filled = 1; | |
6771 | } | |
6772 | else | |
6773 | output_asm_insn ("bve,l (%%r1),%%r2", xoperands); | |
6774 | } | |
5e3c5739 | 6775 | } |
6776 | else | |
6777 | { | |
ee376abe | 6778 | if (!TARGET_NO_SPACE_REGS) |
6779 | output_asm_insn ("ldsid (%%r1),%%r31\n\tmtsp %%r31,%%sr0", | |
6780 | xoperands); | |
06ddb6f8 | 6781 | |
ece88821 | 6782 | if (sibcall) |
ee376abe | 6783 | { |
6784 | if (TARGET_NO_SPACE_REGS) | |
6785 | output_asm_insn ("be 0(%%sr4,%%r1)", xoperands); | |
6786 | else | |
6787 | output_asm_insn ("be 0(%%sr0,%%r1)", xoperands); | |
6788 | } | |
ece88821 | 6789 | else |
6790 | { | |
ee376abe | 6791 | if (TARGET_NO_SPACE_REGS) |
6792 | output_asm_insn ("ble 0(%%sr4,%%r1)", xoperands); | |
6793 | else | |
6794 | output_asm_insn ("ble 0(%%sr0,%%r1)", xoperands); | |
06ddb6f8 | 6795 | |
ece88821 | 6796 | if (indirect_call) |
6797 | output_asm_insn ("stw %%r31,-24(%%sp)", xoperands); | |
6798 | else | |
6799 | output_asm_insn ("copy %%r31,%%r2", xoperands); | |
6800 | delay_slot_filled = 1; | |
6801 | } | |
6802 | } | |
6803 | } | |
06ddb6f8 | 6804 | } |
d6686e21 | 6805 | } |
6d36483b | 6806 | |
ece88821 | 6807 | if (seq_length == 0 || (delay_insn_deleted && !delay_slot_filled)) |
6808 | output_asm_insn ("nop", xoperands); | |
d6686e21 | 6809 | |
ece88821 | 6810 | /* We are done if there isn't a jump in the delay slot. */ |
6811 | if (seq_length == 0 | |
6812 | || delay_insn_deleted | |
6813 | || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN) | |
6814 | return ""; | |
d6686e21 | 6815 | |
ece88821 | 6816 | /* A sibcall should never have a branch in the delay slot. */ |
6817 | if (sibcall) | |
6818 | abort (); | |
d6686e21 | 6819 | |
ece88821 | 6820 | /* This call has an unconditional jump in its delay slot. */ |
6821 | xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1); | |
d6686e21 | 6822 | |
cd0dfcc5 | 6823 | if (!delay_slot_filled && INSN_ADDRESSES_SET_P ()) |
d6686e21 | 6824 | { |
ece88821 | 6825 | /* See if the return address can be adjusted. Use the containing |
6826 | sequence insn's address. */ | |
6827 | rtx seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0))); | |
6828 | int distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn)))) | |
6829 | - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8); | |
6830 | ||
6831 | if (VAL_14_BITS_P (distance)) | |
6832 | { | |
6833 | xoperands[1] = gen_label_rtx (); | |
6834 | output_asm_insn ("ldo %0-%1(%%r2),%%r2", xoperands); | |
805e22b2 | 6835 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
6836 | CODE_LABEL_NUMBER (xoperands[1])); | |
ece88821 | 6837 | } |
6838 | else | |
6839 | /* ??? This branch may not reach its target. */ | |
6840 | output_asm_insn ("nop\n\tb,n %0", xoperands); | |
d6686e21 | 6841 | } |
ece88821 | 6842 | else |
6843 | /* ??? This branch may not reach its target. */ | |
6844 | output_asm_insn ("b,n %0", xoperands); | |
d6686e21 | 6845 | |
6846 | /* Delete the jump. */ | |
6847 | PUT_CODE (NEXT_INSN (insn), NOTE); | |
6848 | NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED; | |
6849 | NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0; | |
ece88821 | 6850 | |
d6686e21 | 6851 | return ""; |
6852 | } | |
6853 | ||
cd0dfcc5 | 6854 | /* Return the attribute length of the indirect call instruction INSN. |
6855 | The length must match the code generated by output_indirect call. | |
6856 | The returned length includes the delay slot. Currently, the delay | |
6857 | slot of an indirect call sequence is not exposed and it is used by | |
6858 | the sequence itself. */ | |
6859 | ||
6860 | int | |
6861 | attr_length_indirect_call (insn) | |
6862 | rtx insn; | |
6863 | { | |
6864 | unsigned long distance = -1; | |
6865 | ||
6866 | if (INSN_ADDRESSES_SET_P ()) | |
6867 | { | |
6868 | distance = (total_code_bytes + insn_current_reference_address (insn)); | |
6869 | if (distance < total_code_bytes) | |
6870 | distance = -1; | |
6871 | } | |
6872 | ||
6873 | if (TARGET_64BIT) | |
6874 | return 12; | |
6875 | ||
6876 | if (TARGET_FAST_INDIRECT_CALLS | |
6877 | || (!TARGET_PORTABLE_RUNTIME | |
6878 | && ((TARGET_PA_20 && distance < 7600000) || distance < 240000))) | |
6879 | return 8; | |
6880 | ||
6881 | if (flag_pic) | |
6882 | return 24; | |
6883 | ||
6884 | if (TARGET_PORTABLE_RUNTIME) | |
6885 | return 20; | |
6886 | ||
6887 | /* Out of reach, can use ble. */ | |
6888 | return 12; | |
6889 | } | |
6890 | ||
6891 | const char * | |
6892 | output_indirect_call (insn, call_dest) | |
6893 | rtx insn; | |
6894 | rtx call_dest; | |
6895 | { | |
6896 | rtx xoperands[1]; | |
6897 | ||
6898 | if (TARGET_64BIT) | |
6899 | { | |
6900 | xoperands[0] = call_dest; | |
6901 | output_asm_insn ("ldd 16(%0),%%r2", xoperands); | |
6902 | output_asm_insn ("bve,l (%%r2),%%r2\n\tldd 24(%0),%%r27", xoperands); | |
6903 | return ""; | |
6904 | } | |
6905 | ||
6906 | /* First the special case for kernels, level 0 systems, etc. */ | |
6907 | if (TARGET_FAST_INDIRECT_CALLS) | |
6908 | return "ble 0(%%sr4,%%r22)\n\tcopy %%r31,%%r2"; | |
6909 | ||
6910 | /* Now the normal case -- we can reach $$dyncall directly or | |
6911 | we're sure that we can get there via a long-branch stub. | |
6912 | ||
6913 | No need to check target flags as the length uniquely identifies | |
6914 | the remaining cases. */ | |
6915 | if (attr_length_indirect_call (insn) == 8) | |
6916 | return ".CALL\tARGW0=GR\n\t{bl|b,l} $$dyncall,%%r31\n\tcopy %%r31,%%r2"; | |
6917 | ||
6918 | /* Long millicode call, but we are not generating PIC or portable runtime | |
6919 | code. */ | |
6920 | if (attr_length_indirect_call (insn) == 12) | |
6921 | return ".CALL\tARGW0=GR\n\tldil L'$$dyncall,%%r2\n\tble R'$$dyncall(%%sr4,%%r2)\n\tcopy %%r31,%%r2"; | |
6922 | ||
6923 | /* Long millicode call for portable runtime. */ | |
6924 | if (attr_length_indirect_call (insn) == 20) | |
6925 | return "ldil L'$$dyncall,%%r31\n\tldo R'$$dyncall(%%r31),%%r31\n\tblr %%r0,%%r2\n\tbv,n %%r0(%%r31)\n\tnop"; | |
6926 | ||
6927 | /* We need a long PIC call to $$dyncall. */ | |
6928 | xoperands[0] = NULL_RTX; | |
6929 | output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands); | |
6930 | if (TARGET_SOM || !TARGET_GAS) | |
6931 | { | |
6932 | xoperands[0] = gen_label_rtx (); | |
6933 | output_asm_insn ("addil L'$$dyncall-%0,%%r1", xoperands); | |
6934 | (*targetm.asm_out.internal_label) (asm_out_file, "L", | |
6935 | CODE_LABEL_NUMBER (xoperands[0])); | |
6936 | output_asm_insn ("ldo R'$$dyncall-%0(%%r1),%%r1", xoperands); | |
6937 | } | |
6938 | else | |
6939 | { | |
6940 | output_asm_insn ("addil L'$$dyncall-$PIC_pcrel$0+4,%%r1", xoperands); | |
6941 | output_asm_insn ("ldo R'$$dyncall-$PIC_pcrel$0+8(%%r1),%%r1", | |
6942 | xoperands); | |
6943 | } | |
6944 | output_asm_insn ("blr %%r0,%%r2", xoperands); | |
6945 | output_asm_insn ("bv,n %%r0(%%r1)\n\tnop", xoperands); | |
6946 | return ""; | |
6947 | } | |
6948 | ||
6949 | /* Return the total length of the save and restore instructions needed for | |
6950 | the data linkage table pointer (i.e., the PIC register) across the call | |
6951 | instruction INSN. No-return calls do not require a save and restore. | |
6952 | In addition, we may be able to avoid the save and restore for calls | |
6953 | within the same translation unit. */ | |
6954 | ||
6955 | int | |
6956 | attr_length_save_restore_dltp (insn) | |
6957 | rtx insn; | |
6958 | { | |
6959 | if (find_reg_note (insn, REG_NORETURN, NULL_RTX)) | |
6960 | return 0; | |
6961 | ||
6962 | return 8; | |
6963 | } | |
6964 | ||
d6f01525 | 6965 | /* In HPUX 8.0's shared library scheme, special relocations are needed |
6d36483b | 6966 | for function labels if they might be passed to a function |
d6f01525 | 6967 | in a shared library (because shared libraries don't live in code |
44acf429 | 6968 | space), and special magic is needed to construct their address. */ |
d6f01525 | 6969 | |
6970 | void | |
44acf429 | 6971 | hppa_encode_label (sym) |
d6f01525 | 6972 | rtx sym; |
6973 | { | |
611a88e1 | 6974 | const char *str = XSTR (sym, 0); |
cccfb31e | 6975 | int len = strlen (str) + 1; |
6976 | char *newstr, *p; | |
d6f01525 | 6977 | |
cccfb31e | 6978 | p = newstr = alloca (len + 1); |
cccfb31e | 6979 | *p++ = '@'; |
6980 | strcpy (p, str); | |
74d80a9a | 6981 | |
ea52c577 | 6982 | XSTR (sym, 0) = ggc_alloc_string (newstr, len); |
d6f01525 | 6983 | } |
6d36483b | 6984 | |
7811991d | 6985 | static void |
6986 | pa_encode_section_info (decl, first) | |
6987 | tree decl; | |
6988 | int first; | |
6989 | { | |
6990 | if (first && TEXT_SPACE_P (decl)) | |
6991 | { | |
6992 | rtx rtl; | |
6993 | if (TREE_CODE (decl) == FUNCTION_DECL | |
6994 | || TREE_CODE (decl) == VAR_DECL) | |
6995 | rtl = DECL_RTL (decl); | |
6996 | else | |
6997 | rtl = TREE_CST_RTL (decl); | |
6998 | SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1; | |
6999 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
7000 | hppa_encode_label (XEXP (DECL_RTL (decl), 0)); | |
7001 | } | |
7002 | } | |
7003 | ||
7b4a38a6 | 7004 | /* This is sort of inverse to pa_encode_section_info. */ |
7005 | ||
7006 | static const char * | |
7007 | pa_strip_name_encoding (str) | |
7008 | const char *str; | |
7009 | { | |
c0264367 | 7010 | str += (*str == '@'); |
7011 | str += (*str == '*'); | |
7012 | return str; | |
7b4a38a6 | 7013 | } |
7014 | ||
d6f01525 | 7015 | int |
166bf021 | 7016 | function_label_operand (op, mode) |
d6f01525 | 7017 | rtx op; |
b1ca791d | 7018 | enum machine_mode mode ATTRIBUTE_UNUSED; |
d6f01525 | 7019 | { |
c1b3411e | 7020 | return GET_CODE (op) == SYMBOL_REF && FUNCTION_NAME_P (XSTR (op, 0)); |
d6f01525 | 7021 | } |
f33e3942 | 7022 | |
166bf021 | 7023 | /* Returns 1 if OP is a function label involved in a simple addition |
7024 | with a constant. Used to keep certain patterns from matching | |
7025 | during instruction combination. */ | |
7026 | int | |
7027 | is_function_label_plus_const (op) | |
7028 | rtx op; | |
7029 | { | |
7030 | /* Strip off any CONST. */ | |
7031 | if (GET_CODE (op) == CONST) | |
7032 | op = XEXP (op, 0); | |
7033 | ||
7034 | return (GET_CODE (op) == PLUS | |
7035 | && function_label_operand (XEXP (op, 0), Pmode) | |
7036 | && GET_CODE (XEXP (op, 1)) == CONST_INT); | |
7037 | } | |
7038 | ||
f1752b7e | 7039 | /* Output assembly code for a thunk to FUNCTION. */ |
7040 | ||
6988553d | 7041 | static void |
eb344f43 | 7042 | pa_asm_output_mi_thunk (file, thunk_fndecl, delta, vcall_offset, function) |
f1752b7e | 7043 | FILE *file; |
7044 | tree thunk_fndecl; | |
e11bd7e5 | 7045 | HOST_WIDE_INT delta; |
eb344f43 | 7046 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED; |
f1752b7e | 7047 | tree function; |
7048 | { | |
7049 | const char *target_name = XSTR (XEXP (DECL_RTL (function), 0), 0); | |
7050 | static unsigned int current_thunk_number; | |
7051 | char label[16]; | |
9d2d8bd6 | 7052 | const char *lab; |
f1752b7e | 7053 | ASM_GENERATE_INTERNAL_LABEL (label, "LTHN", current_thunk_number); |
7054 | lab = (*targetm.strip_name_encoding) (label); | |
7055 | target_name = (*targetm.strip_name_encoding) (target_name); | |
7056 | /* FIXME: total_code_bytes is not handled correctly in files with | |
7057 | mi thunks. */ | |
7058 | pa_output_function_prologue (file, 0); | |
7059 | if (VAL_14_BITS_P (delta)) | |
7060 | { | |
ee376abe | 7061 | if (!TARGET_64BIT && !TARGET_PORTABLE_RUNTIME && flag_pic) |
f1752b7e | 7062 | { |
ece88821 | 7063 | fprintf (file, "\taddil LT'%s,%%r19\n", lab); |
7064 | fprintf (file, "\tldw RT'%s(%%r1),%%r22\n", lab); | |
f1752b7e | 7065 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); |
7066 | fprintf (file, "\tbb,>=,n %%r22,30,.+16\n"); | |
7067 | fprintf (file, "\tdepi 0,31,2,%%r22\n"); | |
7068 | fprintf (file, "\tldw 4(%%sr0,%%r22),%%r19\n"); | |
7069 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); | |
ee376abe | 7070 | if (TARGET_NO_SPACE_REGS) |
7071 | fprintf (file, "\tbe 0(%%sr4,%%r22)\n\tldo "); | |
7072 | else | |
7073 | { | |
7074 | fprintf (file, "\tldsid (%%sr0,%%r22),%%r1\n"); | |
7075 | fprintf (file, "\tmtsp %%r1,%%sr0\n"); | |
7076 | fprintf (file, "\tbe 0(%%sr0,%%r22)\n\tldo "); | |
7077 | } | |
f1752b7e | 7078 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
7079 | fprintf (file, "(%%r26),%%r26\n"); | |
7080 | } | |
7081 | else | |
7082 | { | |
7083 | fprintf (file, "\tb %s\n\tldo ", target_name); | |
7084 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); | |
7085 | fprintf (file, "(%%r26),%%r26\n"); | |
7086 | } | |
7087 | } | |
7088 | else | |
7089 | { | |
ee376abe | 7090 | if (!TARGET_64BIT && !TARGET_PORTABLE_RUNTIME && flag_pic) |
f1752b7e | 7091 | { |
ece88821 | 7092 | fprintf (file, "\taddil L'"); |
f1752b7e | 7093 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
ece88821 | 7094 | fprintf (file, ",%%r26\n\tldo R'"); |
f1752b7e | 7095 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
7096 | fprintf (file, "(%%r1),%%r26\n"); | |
ece88821 | 7097 | fprintf (file, "\taddil LT'%s,%%r19\n", lab); |
7098 | fprintf (file, "\tldw RT'%s(%%r1),%%r22\n", lab); | |
f1752b7e | 7099 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); |
7100 | fprintf (file, "\tbb,>=,n %%r22,30,.+16\n"); | |
7101 | fprintf (file, "\tdepi 0,31,2,%%r22\n"); | |
7102 | fprintf (file, "\tldw 4(%%sr0,%%r22),%%r19\n"); | |
7103 | fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n"); | |
ee376abe | 7104 | if (TARGET_NO_SPACE_REGS) |
7105 | fprintf (file, "\tbe 0(%%sr4,%%r22)"); | |
7106 | else | |
7107 | { | |
7108 | fprintf (file, "\tldsid (%%sr0,%%r22),%%r1\n"); | |
7109 | fprintf (file, "\tmtsp %%r1,%%sr0\n"); | |
7110 | fprintf (file, "\tbe,n 0(%%sr0,%%r22)\n"); | |
7111 | } | |
f1752b7e | 7112 | } |
7113 | else | |
7114 | { | |
ece88821 | 7115 | fprintf (file, "\taddil L'"); |
f1752b7e | 7116 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
ece88821 | 7117 | fprintf (file, ",%%r26\n\tb %s\n\tldo R'", target_name); |
f1752b7e | 7118 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
7119 | fprintf (file, "(%%r1),%%r26\n"); | |
7120 | } | |
7121 | } | |
7122 | ||
7123 | fprintf (file, "\t.EXIT\n\t.PROCEND\n"); | |
7124 | if (! TARGET_64BIT && ! TARGET_PORTABLE_RUNTIME && flag_pic) | |
7125 | { | |
7126 | data_section (); | |
7127 | fprintf (file, "\t.align 4\n"); | |
805e22b2 | 7128 | (*targetm.asm_out.internal_label) (file, "LTHN", current_thunk_number); |
ece88821 | 7129 | fprintf (file, "\t.word P'%s\n", target_name); |
f1752b7e | 7130 | function_section (thunk_fndecl); |
7131 | } | |
7132 | current_thunk_number++; | |
7133 | } | |
7134 | ||
805e22b2 | 7135 | /* Only direct calls to static functions are allowed to be sibling (tail) |
7136 | call optimized. | |
7137 | ||
7138 | This restriction is necessary because some linker generated stubs will | |
7139 | store return pointers into rp' in some cases which might clobber a | |
7140 | live value already in rp'. | |
7141 | ||
7142 | In a sibcall the current function and the target function share stack | |
7143 | space. Thus if the path to the current function and the path to the | |
7144 | target function save a value in rp', they save the value into the | |
7145 | same stack slot, which has undesirable consequences. | |
7146 | ||
7147 | Because of the deferred binding nature of shared libraries any function | |
7148 | with external scope could be in a different load module and thus require | |
7149 | rp' to be saved when calling that function. So sibcall optimizations | |
7150 | can only be safe for static function. | |
7151 | ||
7152 | Note that GCC never needs return value relocations, so we don't have to | |
7153 | worry about static calls with return value relocations (which require | |
7154 | saving rp'). | |
7155 | ||
7156 | It is safe to perform a sibcall optimization when the target function | |
7157 | will never return. */ | |
7158 | static bool | |
7159 | pa_function_ok_for_sibcall (decl, exp) | |
7160 | tree decl; | |
7161 | tree exp ATTRIBUTE_UNUSED; | |
7162 | { | |
f62b73b6 | 7163 | /* Sibcalls are ok for TARGET_ELF32 as along as the linker is used in |
7164 | single subspace mode and the call is not indirect. As far as I know, | |
7165 | there is no operating system support for the multiple subspace mode. | |
7166 | It might be possible to support indirect calls if we didn't use | |
7167 | $$dyncall (see the indirect sequence generated in output_call). */ | |
7168 | if (TARGET_ELF32) | |
7169 | return (decl != NULL_TREE); | |
7170 | ||
7171 | /* Sibcalls are not ok because the arg pointer register is not a fixed | |
7172 | register. This prevents the sibcall optimization from occuring. In | |
7173 | addition, there are problems with stub placement using GNU ld. This | |
7174 | is because a normal sibcall branch uses a 17-bit relocation while | |
7175 | a regular call branch uses a 22-bit relocation. As a result, more | |
7176 | care needs to be taken in the placement of long-branch stubs. */ | |
7177 | if (TARGET_64BIT) | |
7178 | return false; | |
7179 | ||
805e22b2 | 7180 | return (decl |
f62b73b6 | 7181 | && !TARGET_PORTABLE_RUNTIME |
7182 | && !TREE_PUBLIC (decl)); | |
805e22b2 | 7183 | } |
7184 | ||
37580c80 | 7185 | /* Returns 1 if the 6 operands specified in OPERANDS are suitable for |
7186 | use in fmpyadd instructions. */ | |
4ed6ee50 | 7187 | int |
df0651dc | 7188 | fmpyaddoperands (operands) |
4ed6ee50 | 7189 | rtx *operands; |
7190 | { | |
201f01e9 | 7191 | enum machine_mode mode = GET_MODE (operands[0]); |
4ed6ee50 | 7192 | |
ab449421 | 7193 | /* Must be a floating point mode. */ |
7194 | if (mode != SFmode && mode != DFmode) | |
7195 | return 0; | |
7196 | ||
4ed6ee50 | 7197 | /* All modes must be the same. */ |
201f01e9 | 7198 | if (! (mode == GET_MODE (operands[1]) |
7199 | && mode == GET_MODE (operands[2]) | |
7200 | && mode == GET_MODE (operands[3]) | |
7201 | && mode == GET_MODE (operands[4]) | |
7202 | && mode == GET_MODE (operands[5]))) | |
4ed6ee50 | 7203 | return 0; |
7204 | ||
ab449421 | 7205 | /* All operands must be registers. */ |
7206 | if (! (GET_CODE (operands[1]) == REG | |
7207 | && GET_CODE (operands[2]) == REG | |
7208 | && GET_CODE (operands[3]) == REG | |
7209 | && GET_CODE (operands[4]) == REG | |
7210 | && GET_CODE (operands[5]) == REG)) | |
4ed6ee50 | 7211 | return 0; |
7212 | ||
37580c80 | 7213 | /* Only 2 real operands to the addition. One of the input operands must |
7214 | be the same as the output operand. */ | |
4ed6ee50 | 7215 | if (! rtx_equal_p (operands[3], operands[4]) |
7216 | && ! rtx_equal_p (operands[3], operands[5])) | |
7217 | return 0; | |
7218 | ||
7219 | /* Inout operand of add can not conflict with any operands from multiply. */ | |
7220 | if (rtx_equal_p (operands[3], operands[0]) | |
7221 | || rtx_equal_p (operands[3], operands[1]) | |
7222 | || rtx_equal_p (operands[3], operands[2])) | |
7223 | return 0; | |
7224 | ||
7225 | /* multiply can not feed into addition operands. */ | |
7226 | if (rtx_equal_p (operands[4], operands[0]) | |
7227 | || rtx_equal_p (operands[5], operands[0])) | |
7228 | return 0; | |
7229 | ||
ab449421 | 7230 | /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */ |
7231 | if (mode == SFmode | |
bac38c40 | 7232 | && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS |
7233 | || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS | |
7234 | || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS | |
7235 | || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS | |
7236 | || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS | |
7237 | || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS)) | |
ab449421 | 7238 | return 0; |
7239 | ||
4ed6ee50 | 7240 | /* Passed. Operands are suitable for fmpyadd. */ |
7241 | return 1; | |
7242 | } | |
7243 | ||
de419443 | 7244 | #if !defined(USE_COLLECT2) |
7245 | static void | |
7246 | pa_asm_out_constructor (symbol, priority) | |
7247 | rtx symbol; | |
7248 | int priority; | |
7249 | { | |
7250 | if (!function_label_operand (symbol, VOIDmode)) | |
7251 | hppa_encode_label (symbol); | |
7252 | ||
7253 | #ifdef CTORS_SECTION_ASM_OP | |
7254 | default_ctor_section_asm_out_constructor (symbol, priority); | |
7255 | #else | |
7256 | # ifdef TARGET_ASM_NAMED_SECTION | |
7257 | default_named_section_asm_out_constructor (symbol, priority); | |
7258 | # else | |
7259 | default_stabs_asm_out_constructor (symbol, priority); | |
7260 | # endif | |
7261 | #endif | |
7262 | } | |
7263 | ||
7264 | static void | |
7265 | pa_asm_out_destructor (symbol, priority) | |
7266 | rtx symbol; | |
7267 | int priority; | |
7268 | { | |
7269 | if (!function_label_operand (symbol, VOIDmode)) | |
7270 | hppa_encode_label (symbol); | |
7271 | ||
7272 | #ifdef DTORS_SECTION_ASM_OP | |
7273 | default_dtor_section_asm_out_destructor (symbol, priority); | |
7274 | #else | |
7275 | # ifdef TARGET_ASM_NAMED_SECTION | |
7276 | default_named_section_asm_out_destructor (symbol, priority); | |
7277 | # else | |
7278 | default_stabs_asm_out_destructor (symbol, priority); | |
7279 | # endif | |
7280 | #endif | |
7281 | } | |
7282 | #endif | |
7283 | ||
37580c80 | 7284 | /* Returns 1 if the 6 operands specified in OPERANDS are suitable for |
7285 | use in fmpysub instructions. */ | |
4ed6ee50 | 7286 | int |
df0651dc | 7287 | fmpysuboperands (operands) |
4ed6ee50 | 7288 | rtx *operands; |
7289 | { | |
201f01e9 | 7290 | enum machine_mode mode = GET_MODE (operands[0]); |
4ed6ee50 | 7291 | |
ab449421 | 7292 | /* Must be a floating point mode. */ |
7293 | if (mode != SFmode && mode != DFmode) | |
7294 | return 0; | |
7295 | ||
4ed6ee50 | 7296 | /* All modes must be the same. */ |
201f01e9 | 7297 | if (! (mode == GET_MODE (operands[1]) |
7298 | && mode == GET_MODE (operands[2]) | |
7299 | && mode == GET_MODE (operands[3]) | |
7300 | && mode == GET_MODE (operands[4]) | |
7301 | && mode == GET_MODE (operands[5]))) | |
4ed6ee50 | 7302 | return 0; |
7303 | ||
ab449421 | 7304 | /* All operands must be registers. */ |
7305 | if (! (GET_CODE (operands[1]) == REG | |
7306 | && GET_CODE (operands[2]) == REG | |
7307 | && GET_CODE (operands[3]) == REG | |
7308 | && GET_CODE (operands[4]) == REG | |
7309 | && GET_CODE (operands[5]) == REG)) | |
4ed6ee50 | 7310 | return 0; |
7311 | ||
37580c80 | 7312 | /* Only 2 real operands to the subtraction. Subtraction is not a commutative |
7313 | operation, so operands[4] must be the same as operand[3]. */ | |
4ed6ee50 | 7314 | if (! rtx_equal_p (operands[3], operands[4])) |
7315 | return 0; | |
7316 | ||
7317 | /* multiply can not feed into subtraction. */ | |
37580c80 | 7318 | if (rtx_equal_p (operands[5], operands[0])) |
4ed6ee50 | 7319 | return 0; |
7320 | ||
37580c80 | 7321 | /* Inout operand of sub can not conflict with any operands from multiply. */ |
4ed6ee50 | 7322 | if (rtx_equal_p (operands[3], operands[0]) |
7323 | || rtx_equal_p (operands[3], operands[1]) | |
7324 | || rtx_equal_p (operands[3], operands[2])) | |
7325 | return 0; | |
7326 | ||
ab449421 | 7327 | /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */ |
7328 | if (mode == SFmode | |
bac38c40 | 7329 | && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS |
7330 | || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS | |
7331 | || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS | |
7332 | || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS | |
7333 | || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS | |
7334 | || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS)) | |
ab449421 | 7335 | return 0; |
7336 | ||
4ed6ee50 | 7337 | /* Passed. Operands are suitable for fmpysub. */ |
7338 | return 1; | |
7339 | } | |
7340 | ||
89f29d73 | 7341 | int |
7342 | plus_xor_ior_operator (op, mode) | |
7343 | rtx op; | |
b1ca791d | 7344 | enum machine_mode mode ATTRIBUTE_UNUSED; |
89f29d73 | 7345 | { |
7346 | return (GET_CODE (op) == PLUS || GET_CODE (op) == XOR | |
7347 | || GET_CODE (op) == IOR); | |
7348 | } | |
6720f95e | 7349 | |
7350 | /* Return 1 if the given constant is 2, 4, or 8. These are the valid | |
7351 | constants for shadd instructions. */ | |
7d27e4c9 | 7352 | static int |
6720f95e | 7353 | shadd_constant_p (val) |
7354 | int val; | |
7355 | { | |
7356 | if (val == 2 || val == 4 || val == 8) | |
7357 | return 1; | |
7358 | else | |
7359 | return 0; | |
7360 | } | |
3a16146d | 7361 | |
7362 | /* Return 1 if OP is a CONST_INT with the value 2, 4, or 8. These are | |
7363 | the valid constant for shadd instructions. */ | |
7364 | int | |
7365 | shadd_operand (op, mode) | |
7366 | rtx op; | |
b1ca791d | 7367 | enum machine_mode mode ATTRIBUTE_UNUSED; |
3a16146d | 7368 | { |
7369 | return (GET_CODE (op) == CONST_INT && shadd_constant_p (INTVAL (op))); | |
7370 | } | |
5fbd5940 | 7371 | |
42819d4e | 7372 | /* Return 1 if OP is valid as a base register in a reg + reg address. */ |
7373 | ||
7374 | int | |
7375 | basereg_operand (op, mode) | |
7376 | rtx op; | |
7377 | enum machine_mode mode; | |
7378 | { | |
21f3ee9c | 7379 | /* cse will create some unscaled indexed addresses, however; it |
7380 | generally isn't a win on the PA, so avoid creating unscaled | |
7381 | indexed addresses until after cse is finished. */ | |
7382 | if (!cse_not_expected) | |
7383 | return 0; | |
7384 | ||
e61a0a7f | 7385 | /* Allow any register when TARGET_NO_SPACE_REGS is in effect since |
7386 | we don't have to worry about the braindamaged implicit space | |
7387 | register selection from the basereg. */ | |
7388 | if (TARGET_NO_SPACE_REGS) | |
23a667d0 | 7389 | return (GET_CODE (op) == REG); |
42819d4e | 7390 | |
21f3ee9c | 7391 | /* While it's always safe to index off the frame pointer, it's not |
7392 | always profitable, particularly when the frame pointer is being | |
7393 | eliminated. */ | |
7394 | if (! flag_omit_frame_pointer && op == frame_pointer_rtx) | |
42819d4e | 7395 | return 1; |
7396 | ||
e61a0a7f | 7397 | return (GET_CODE (op) == REG |
90ae5d76 | 7398 | && REG_POINTER (op) |
7399 | && register_operand (op, mode)); | |
42819d4e | 7400 | } |
7401 | ||
51987f90 | 7402 | /* Return 1 if this operand is anything other than a hard register. */ |
7403 | ||
7404 | int | |
7405 | non_hard_reg_operand (op, mode) | |
7406 | rtx op; | |
b1ca791d | 7407 | enum machine_mode mode ATTRIBUTE_UNUSED; |
51987f90 | 7408 | { |
7409 | return ! (GET_CODE (op) == REG && REGNO (op) < FIRST_PSEUDO_REGISTER); | |
7410 | } | |
7411 | ||
5fbd5940 | 7412 | /* Return 1 if INSN branches forward. Should be using insn_addresses |
6dc3b0d9 | 7413 | to avoid walking through all the insns... */ |
7d27e4c9 | 7414 | static int |
5fbd5940 | 7415 | forward_branch_p (insn) |
7416 | rtx insn; | |
7417 | { | |
7418 | rtx label = JUMP_LABEL (insn); | |
7419 | ||
7420 | while (insn) | |
7421 | { | |
7422 | if (insn == label) | |
7423 | break; | |
7424 | else | |
7425 | insn = NEXT_INSN (insn); | |
7426 | } | |
7427 | ||
7428 | return (insn == label); | |
7429 | } | |
7430 | ||
29a4502c | 7431 | /* Return 1 if OP is an equality comparison, else return 0. */ |
7432 | int | |
7433 | eq_neq_comparison_operator (op, mode) | |
7434 | rtx op; | |
b1ca791d | 7435 | enum machine_mode mode ATTRIBUTE_UNUSED; |
29a4502c | 7436 | { |
7437 | return (GET_CODE (op) == EQ || GET_CODE (op) == NE); | |
7438 | } | |
7439 | ||
7440 | /* Return 1 if OP is an operator suitable for use in a movb instruction. */ | |
7441 | int | |
7442 | movb_comparison_operator (op, mode) | |
7443 | rtx op; | |
b1ca791d | 7444 | enum machine_mode mode ATTRIBUTE_UNUSED; |
29a4502c | 7445 | { |
7446 | return (GET_CODE (op) == EQ || GET_CODE (op) == NE | |
7447 | || GET_CODE (op) == LT || GET_CODE (op) == GE); | |
7448 | } | |
7449 | ||
d6686e21 | 7450 | /* Return 1 if INSN is in the delay slot of a call instruction. */ |
7451 | int | |
7452 | jump_in_call_delay (insn) | |
7453 | rtx insn; | |
7454 | { | |
7455 | ||
7456 | if (GET_CODE (insn) != JUMP_INSN) | |
7457 | return 0; | |
7458 | ||
7459 | if (PREV_INSN (insn) | |
7460 | && PREV_INSN (PREV_INSN (insn)) | |
7461 | && GET_CODE (next_active_insn (PREV_INSN (PREV_INSN (insn)))) == INSN) | |
7462 | { | |
7463 | rtx test_insn = next_active_insn (PREV_INSN (PREV_INSN (insn))); | |
7464 | ||
7465 | return (GET_CODE (PATTERN (test_insn)) == SEQUENCE | |
7466 | && XVECEXP (PATTERN (test_insn), 0, 1) == insn); | |
7467 | ||
7468 | } | |
7469 | else | |
7470 | return 0; | |
7471 | } | |
3b1e673e | 7472 | |
546a40bd | 7473 | /* Output an unconditional move and branch insn. */ |
7474 | ||
611a88e1 | 7475 | const char * |
546a40bd | 7476 | output_parallel_movb (operands, length) |
7477 | rtx *operands; | |
7478 | int length; | |
7479 | { | |
7480 | /* These are the cases in which we win. */ | |
7481 | if (length == 4) | |
7482 | return "mov%I1b,tr %1,%0,%2"; | |
7483 | ||
7484 | /* None of these cases wins, but they don't lose either. */ | |
7485 | if (dbr_sequence_length () == 0) | |
7486 | { | |
7487 | /* Nothing in the delay slot, fake it by putting the combined | |
7488 | insn (the copy or add) in the delay slot of a bl. */ | |
7489 | if (GET_CODE (operands[1]) == CONST_INT) | |
5a811d43 | 7490 | return "b %2\n\tldi %1,%0"; |
546a40bd | 7491 | else |
5a811d43 | 7492 | return "b %2\n\tcopy %1,%0"; |
546a40bd | 7493 | } |
7494 | else | |
7495 | { | |
7496 | /* Something in the delay slot, but we've got a long branch. */ | |
7497 | if (GET_CODE (operands[1]) == CONST_INT) | |
5a811d43 | 7498 | return "ldi %1,%0\n\tb %2"; |
546a40bd | 7499 | else |
5a811d43 | 7500 | return "copy %1,%0\n\tb %2"; |
546a40bd | 7501 | } |
7502 | } | |
7503 | ||
7504 | /* Output an unconditional add and branch insn. */ | |
7505 | ||
611a88e1 | 7506 | const char * |
546a40bd | 7507 | output_parallel_addb (operands, length) |
7508 | rtx *operands; | |
7509 | int length; | |
7510 | { | |
7511 | /* To make life easy we want operand0 to be the shared input/output | |
7512 | operand and operand1 to be the readonly operand. */ | |
7513 | if (operands[0] == operands[1]) | |
7514 | operands[1] = operands[2]; | |
7515 | ||
7516 | /* These are the cases in which we win. */ | |
7517 | if (length == 4) | |
7518 | return "add%I1b,tr %1,%0,%3"; | |
7519 | ||
7520 | /* None of these cases win, but they don't lose either. */ | |
7521 | if (dbr_sequence_length () == 0) | |
7522 | { | |
7523 | /* Nothing in the delay slot, fake it by putting the combined | |
7524 | insn (the copy or add) in the delay slot of a bl. */ | |
5a811d43 | 7525 | return "b %3\n\tadd%I1 %1,%0,%0"; |
546a40bd | 7526 | } |
7527 | else | |
7528 | { | |
7529 | /* Something in the delay slot, but we've got a long branch. */ | |
5a811d43 | 7530 | return "add%I1 %1,%0,%0\n\tb %3"; |
546a40bd | 7531 | } |
7532 | } | |
7533 | ||
7c5101fc | 7534 | /* Return nonzero if INSN (a jump insn) immediately follows a call |
7535 | to a named function. This is used to avoid filling the delay slot | |
7536 | of the jump since it can usually be eliminated by modifying RP in | |
7537 | the delay slot of the call. */ | |
9840d99d | 7538 | |
7d27e4c9 | 7539 | int |
546a40bd | 7540 | following_call (insn) |
7541 | rtx insn; | |
7542 | { | |
ed1b0769 | 7543 | if (! TARGET_JUMP_IN_DELAY) |
1b6f11e2 | 7544 | return 0; |
7545 | ||
546a40bd | 7546 | /* Find the previous real insn, skipping NOTEs. */ |
7547 | insn = PREV_INSN (insn); | |
7548 | while (insn && GET_CODE (insn) == NOTE) | |
7549 | insn = PREV_INSN (insn); | |
7550 | ||
7551 | /* Check for CALL_INSNs and millicode calls. */ | |
7552 | if (insn | |
1d2e016c | 7553 | && ((GET_CODE (insn) == CALL_INSN |
7554 | && get_attr_type (insn) != TYPE_DYNCALL) | |
546a40bd | 7555 | || (GET_CODE (insn) == INSN |
7556 | && GET_CODE (PATTERN (insn)) != SEQUENCE | |
7557 | && GET_CODE (PATTERN (insn)) != USE | |
7558 | && GET_CODE (PATTERN (insn)) != CLOBBER | |
7559 | && get_attr_type (insn) == TYPE_MILLI))) | |
7560 | return 1; | |
7561 | ||
7562 | return 0; | |
7563 | } | |
7564 | ||
3b1e673e | 7565 | /* We use this hook to perform a PA specific optimization which is difficult |
7566 | to do in earlier passes. | |
7567 | ||
7568 | We want the delay slots of branches within jump tables to be filled. | |
7569 | None of the compiler passes at the moment even has the notion that a | |
7570 | PA jump table doesn't contain addresses, but instead contains actual | |
7571 | instructions! | |
7572 | ||
7573 | Because we actually jump into the table, the addresses of each entry | |
01cc3b75 | 7574 | must stay constant in relation to the beginning of the table (which |
3b1e673e | 7575 | itself must stay constant relative to the instruction to jump into |
7576 | it). I don't believe we can guarantee earlier passes of the compiler | |
7577 | will adhere to those rules. | |
7578 | ||
7579 | So, late in the compilation process we find all the jump tables, and | |
7580 | expand them into real code -- eg each entry in the jump table vector | |
7581 | will get an appropriate label followed by a jump to the final target. | |
7582 | ||
7583 | Reorg and the final jump pass can then optimize these branches and | |
7584 | fill their delay slots. We end up with smaller, more efficient code. | |
7585 | ||
9840d99d | 7586 | The jump instructions within the table are special; we must be able |
3b1e673e | 7587 | to identify them during assembly output (if the jumps don't get filled |
7588 | we need to emit a nop rather than nullifying the delay slot)). We | |
9840d99d | 7589 | identify jumps in switch tables by marking the SET with DImode. |
9239127b | 7590 | |
7591 | We also surround the jump table itself with BEGIN_BRTAB and END_BRTAB | |
7592 | insns. This serves two purposes, first it prevents jump.c from | |
7593 | noticing that the last N entries in the table jump to the instruction | |
7594 | immediately after the table and deleting the jumps. Second, those | |
7595 | insns mark where we should emit .begin_brtab and .end_brtab directives | |
7596 | when using GAS (allows for better link time optimizations). */ | |
3b1e673e | 7597 | |
7d27e4c9 | 7598 | void |
3b1e673e | 7599 | pa_reorg (insns) |
7600 | rtx insns; | |
7601 | { | |
7602 | rtx insn; | |
7603 | ||
c533da59 | 7604 | remove_useless_addtr_insns (insns, 1); |
3d457930 | 7605 | |
342aabd9 | 7606 | if (pa_cpu < PROCESSOR_8000) |
7607 | pa_combine_instructions (get_insns ()); | |
7608 | ||
bd49d362 | 7609 | |
3d457930 | 7610 | /* This is fairly cheap, so always run it if optimizing. */ |
a66555b2 | 7611 | if (optimize > 0 && !TARGET_BIG_SWITCH) |
3b1e673e | 7612 | { |
b41266d4 | 7613 | /* Find and explode all ADDR_VEC or ADDR_DIFF_VEC insns. */ |
3b1e673e | 7614 | insns = get_insns (); |
7615 | for (insn = insns; insn; insn = NEXT_INSN (insn)) | |
7616 | { | |
7617 | rtx pattern, tmp, location; | |
7618 | unsigned int length, i; | |
7619 | ||
b41266d4 | 7620 | /* Find an ADDR_VEC or ADDR_DIFF_VEC insn to explode. */ |
3b1e673e | 7621 | if (GET_CODE (insn) != JUMP_INSN |
b41266d4 | 7622 | || (GET_CODE (PATTERN (insn)) != ADDR_VEC |
7623 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)) | |
3b1e673e | 7624 | continue; |
7625 | ||
9239127b | 7626 | /* Emit marker for the beginning of the branch table. */ |
7627 | emit_insn_before (gen_begin_brtab (), insn); | |
f9333726 | 7628 | |
3b1e673e | 7629 | pattern = PATTERN (insn); |
7630 | location = PREV_INSN (insn); | |
b41266d4 | 7631 | length = XVECLEN (pattern, GET_CODE (pattern) == ADDR_DIFF_VEC); |
f9333726 | 7632 | |
3b1e673e | 7633 | for (i = 0; i < length; i++) |
7634 | { | |
a66555b2 | 7635 | /* Emit a label before each jump to keep jump.c from |
7636 | removing this code. */ | |
7637 | tmp = gen_label_rtx (); | |
7638 | LABEL_NUSES (tmp) = 1; | |
7639 | emit_label_after (tmp, location); | |
7640 | location = NEXT_INSN (location); | |
7641 | ||
b41266d4 | 7642 | if (GET_CODE (pattern) == ADDR_VEC) |
7643 | { | |
7644 | /* Emit the jump itself. */ | |
0708e381 | 7645 | tmp = gen_jump (XEXP (XVECEXP (pattern, 0, i), 0)); |
b41266d4 | 7646 | tmp = emit_jump_insn_after (tmp, location); |
7647 | JUMP_LABEL (tmp) = XEXP (XVECEXP (pattern, 0, i), 0); | |
0708e381 | 7648 | /* It is easy to rely on the branch table markers |
7649 | during assembly output to trigger the correct code | |
7650 | for a switch table jump with an unfilled delay slot, | |
7651 | ||
7652 | However, that requires state and assumes that we look | |
7653 | at insns in order. | |
7654 | ||
7655 | We can't make such assumptions when computing the length | |
7656 | of instructions. Ugh. We could walk the insn chain to | |
7657 | determine if this instruction is in a branch table, but | |
7658 | that can get rather expensive, particularly during the | |
7659 | branch shortening phase of the compiler. | |
7660 | ||
7661 | So instead we mark this jump as being special. This is | |
7662 | far from ideal and knows that no code after this will | |
7663 | muck around with the mode of the JUMP_INSN itself. */ | |
7664 | PUT_MODE (tmp, SImode); | |
b41266d4 | 7665 | LABEL_NUSES (JUMP_LABEL (tmp))++; |
7666 | location = NEXT_INSN (location); | |
7667 | } | |
7668 | else | |
7669 | { | |
7670 | /* Emit the jump itself. */ | |
0708e381 | 7671 | tmp = gen_jump (XEXP (XVECEXP (pattern, 1, i), 0)); |
b41266d4 | 7672 | tmp = emit_jump_insn_after (tmp, location); |
7673 | JUMP_LABEL (tmp) = XEXP (XVECEXP (pattern, 1, i), 0); | |
0708e381 | 7674 | /* It is easy to rely on the branch table markers |
7675 | during assembly output to trigger the correct code | |
7676 | for a switch table jump with an unfilled delay slot, | |
7677 | ||
7678 | However, that requires state and assumes that we look | |
7679 | at insns in order. | |
7680 | ||
7681 | We can't make such assumptions when computing the length | |
7682 | of instructions. Ugh. We could walk the insn chain to | |
7683 | determine if this instruction is in a branch table, but | |
7684 | that can get rather expensive, particularly during the | |
7685 | branch shortening phase of the compiler. | |
7686 | ||
7687 | So instead we mark this jump as being special. This is | |
7688 | far from ideal and knows that no code after this will | |
7689 | muck around with the mode of the JUMP_INSN itself. */ | |
7690 | PUT_MODE (tmp, SImode); | |
b41266d4 | 7691 | LABEL_NUSES (JUMP_LABEL (tmp))++; |
7692 | location = NEXT_INSN (location); | |
7693 | } | |
3b1e673e | 7694 | |
7695 | /* Emit a BARRIER after the jump. */ | |
3b1e673e | 7696 | emit_barrier_after (location); |
3b1e673e | 7697 | location = NEXT_INSN (location); |
7698 | } | |
f9333726 | 7699 | |
9239127b | 7700 | /* Emit marker for the end of the branch table. */ |
7701 | emit_insn_before (gen_end_brtab (), location); | |
7702 | location = NEXT_INSN (location); | |
7703 | emit_barrier_after (location); | |
a66555b2 | 7704 | |
b41266d4 | 7705 | /* Delete the ADDR_VEC or ADDR_DIFF_VEC. */ |
3b1e673e | 7706 | delete_insn (insn); |
7707 | } | |
7708 | } | |
9239127b | 7709 | else |
f9333726 | 7710 | { |
7711 | /* Sill need an end_brtab insn. */ | |
7712 | insns = get_insns (); | |
7713 | for (insn = insns; insn; insn = NEXT_INSN (insn)) | |
7714 | { | |
7715 | /* Find an ADDR_VEC insn. */ | |
7716 | if (GET_CODE (insn) != JUMP_INSN | |
b41266d4 | 7717 | || (GET_CODE (PATTERN (insn)) != ADDR_VEC |
7718 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)) | |
f9333726 | 7719 | continue; |
7720 | ||
7721 | /* Now generate markers for the beginning and end of the | |
ad87de1e | 7722 | branch table. */ |
f9333726 | 7723 | emit_insn_before (gen_begin_brtab (), insn); |
7724 | emit_insn_after (gen_end_brtab (), insn); | |
7725 | } | |
7726 | } | |
d3287673 | 7727 | } |
bd49d362 | 7728 | |
7729 | /* The PA has a number of odd instructions which can perform multiple | |
7730 | tasks at once. On first generation PA machines (PA1.0 and PA1.1) | |
7731 | it may be profitable to combine two instructions into one instruction | |
7732 | with two outputs. It's not profitable PA2.0 machines because the | |
7733 | two outputs would take two slots in the reorder buffers. | |
7734 | ||
7735 | This routine finds instructions which can be combined and combines | |
7736 | them. We only support some of the potential combinations, and we | |
7737 | only try common ways to find suitable instructions. | |
7738 | ||
7739 | * addb can add two registers or a register and a small integer | |
7740 | and jump to a nearby (+-8k) location. Normally the jump to the | |
7741 | nearby location is conditional on the result of the add, but by | |
7742 | using the "true" condition we can make the jump unconditional. | |
7743 | Thus addb can perform two independent operations in one insn. | |
7744 | ||
7745 | * movb is similar to addb in that it can perform a reg->reg | |
7746 | or small immediate->reg copy and jump to a nearby (+-8k location). | |
7747 | ||
7748 | * fmpyadd and fmpysub can perform a FP multiply and either an | |
7749 | FP add or FP sub if the operands of the multiply and add/sub are | |
7750 | independent (there are other minor restrictions). Note both | |
7751 | the fmpy and fadd/fsub can in theory move to better spots according | |
7752 | to data dependencies, but for now we require the fmpy stay at a | |
7753 | fixed location. | |
7754 | ||
7755 | * Many of the memory operations can perform pre & post updates | |
7756 | of index registers. GCC's pre/post increment/decrement addressing | |
7757 | is far too simple to take advantage of all the possibilities. This | |
7758 | pass may not be suitable since those insns may not be independent. | |
7759 | ||
7760 | * comclr can compare two ints or an int and a register, nullify | |
7761 | the following instruction and zero some other register. This | |
7762 | is more difficult to use as it's harder to find an insn which | |
7763 | will generate a comclr than finding something like an unconditional | |
7764 | branch. (conditional moves & long branches create comclr insns). | |
7765 | ||
7766 | * Most arithmetic operations can conditionally skip the next | |
7767 | instruction. They can be viewed as "perform this operation | |
7768 | and conditionally jump to this nearby location" (where nearby | |
7769 | is an insns away). These are difficult to use due to the | |
7770 | branch length restrictions. */ | |
7771 | ||
7d27e4c9 | 7772 | static void |
bd49d362 | 7773 | pa_combine_instructions (insns) |
b1ca791d | 7774 | rtx insns ATTRIBUTE_UNUSED; |
bd49d362 | 7775 | { |
7776 | rtx anchor, new; | |
7777 | ||
7778 | /* This can get expensive since the basic algorithm is on the | |
7779 | order of O(n^2) (or worse). Only do it for -O2 or higher | |
ad87de1e | 7780 | levels of optimization. */ |
bd49d362 | 7781 | if (optimize < 2) |
7782 | return; | |
7783 | ||
7784 | /* Walk down the list of insns looking for "anchor" insns which | |
7785 | may be combined with "floating" insns. As the name implies, | |
7786 | "anchor" instructions don't move, while "floating" insns may | |
7787 | move around. */ | |
ad851752 | 7788 | new = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, NULL_RTX, NULL_RTX)); |
bd49d362 | 7789 | new = make_insn_raw (new); |
7790 | ||
7791 | for (anchor = get_insns (); anchor; anchor = NEXT_INSN (anchor)) | |
7792 | { | |
7793 | enum attr_pa_combine_type anchor_attr; | |
7794 | enum attr_pa_combine_type floater_attr; | |
7795 | ||
7796 | /* We only care about INSNs, JUMP_INSNs, and CALL_INSNs. | |
7797 | Also ignore any special USE insns. */ | |
7d27e4c9 | 7798 | if ((GET_CODE (anchor) != INSN |
bd49d362 | 7799 | && GET_CODE (anchor) != JUMP_INSN |
7d27e4c9 | 7800 | && GET_CODE (anchor) != CALL_INSN) |
bd49d362 | 7801 | || GET_CODE (PATTERN (anchor)) == USE |
7802 | || GET_CODE (PATTERN (anchor)) == CLOBBER | |
7803 | || GET_CODE (PATTERN (anchor)) == ADDR_VEC | |
7804 | || GET_CODE (PATTERN (anchor)) == ADDR_DIFF_VEC) | |
7805 | continue; | |
7806 | ||
7807 | anchor_attr = get_attr_pa_combine_type (anchor); | |
7808 | /* See if anchor is an insn suitable for combination. */ | |
7809 | if (anchor_attr == PA_COMBINE_TYPE_FMPY | |
7810 | || anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
7811 | || (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH | |
7812 | && ! forward_branch_p (anchor))) | |
7813 | { | |
7814 | rtx floater; | |
7815 | ||
7816 | for (floater = PREV_INSN (anchor); | |
7817 | floater; | |
7818 | floater = PREV_INSN (floater)) | |
7819 | { | |
7820 | if (GET_CODE (floater) == NOTE | |
7821 | || (GET_CODE (floater) == INSN | |
7822 | && (GET_CODE (PATTERN (floater)) == USE | |
7823 | || GET_CODE (PATTERN (floater)) == CLOBBER))) | |
7824 | continue; | |
7825 | ||
7826 | /* Anything except a regular INSN will stop our search. */ | |
7827 | if (GET_CODE (floater) != INSN | |
7828 | || GET_CODE (PATTERN (floater)) == ADDR_VEC | |
7829 | || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC) | |
7830 | { | |
7831 | floater = NULL_RTX; | |
7832 | break; | |
7833 | } | |
7834 | ||
7835 | /* See if FLOATER is suitable for combination with the | |
7836 | anchor. */ | |
7837 | floater_attr = get_attr_pa_combine_type (floater); | |
7838 | if ((anchor_attr == PA_COMBINE_TYPE_FMPY | |
7839 | && floater_attr == PA_COMBINE_TYPE_FADDSUB) | |
7840 | || (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
7841 | && floater_attr == PA_COMBINE_TYPE_FMPY)) | |
7842 | { | |
7843 | /* If ANCHOR and FLOATER can be combined, then we're | |
7844 | done with this pass. */ | |
7845 | if (pa_can_combine_p (new, anchor, floater, 0, | |
7846 | SET_DEST (PATTERN (floater)), | |
7847 | XEXP (SET_SRC (PATTERN (floater)), 0), | |
7848 | XEXP (SET_SRC (PATTERN (floater)), 1))) | |
7849 | break; | |
7850 | } | |
7851 | ||
7852 | else if (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH | |
7853 | && floater_attr == PA_COMBINE_TYPE_ADDMOVE) | |
7854 | { | |
7855 | if (GET_CODE (SET_SRC (PATTERN (floater))) == PLUS) | |
7856 | { | |
7857 | if (pa_can_combine_p (new, anchor, floater, 0, | |
7858 | SET_DEST (PATTERN (floater)), | |
7859 | XEXP (SET_SRC (PATTERN (floater)), 0), | |
7860 | XEXP (SET_SRC (PATTERN (floater)), 1))) | |
7861 | break; | |
7862 | } | |
7863 | else | |
7864 | { | |
7865 | if (pa_can_combine_p (new, anchor, floater, 0, | |
7866 | SET_DEST (PATTERN (floater)), | |
7867 | SET_SRC (PATTERN (floater)), | |
7868 | SET_SRC (PATTERN (floater)))) | |
7869 | break; | |
7870 | } | |
7871 | } | |
7872 | } | |
7873 | ||
7874 | /* If we didn't find anything on the backwards scan try forwards. */ | |
7875 | if (!floater | |
7876 | && (anchor_attr == PA_COMBINE_TYPE_FMPY | |
7877 | || anchor_attr == PA_COMBINE_TYPE_FADDSUB)) | |
7878 | { | |
7879 | for (floater = anchor; floater; floater = NEXT_INSN (floater)) | |
7880 | { | |
7881 | if (GET_CODE (floater) == NOTE | |
7882 | || (GET_CODE (floater) == INSN | |
7883 | && (GET_CODE (PATTERN (floater)) == USE | |
7884 | || GET_CODE (PATTERN (floater)) == CLOBBER))) | |
9840d99d | 7885 | |
bd49d362 | 7886 | continue; |
7887 | ||
7888 | /* Anything except a regular INSN will stop our search. */ | |
7889 | if (GET_CODE (floater) != INSN | |
7890 | || GET_CODE (PATTERN (floater)) == ADDR_VEC | |
7891 | || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC) | |
7892 | { | |
7893 | floater = NULL_RTX; | |
7894 | break; | |
7895 | } | |
7896 | ||
7897 | /* See if FLOATER is suitable for combination with the | |
7898 | anchor. */ | |
7899 | floater_attr = get_attr_pa_combine_type (floater); | |
7900 | if ((anchor_attr == PA_COMBINE_TYPE_FMPY | |
7901 | && floater_attr == PA_COMBINE_TYPE_FADDSUB) | |
7902 | || (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
7903 | && floater_attr == PA_COMBINE_TYPE_FMPY)) | |
7904 | { | |
7905 | /* If ANCHOR and FLOATER can be combined, then we're | |
7906 | done with this pass. */ | |
7907 | if (pa_can_combine_p (new, anchor, floater, 1, | |
7908 | SET_DEST (PATTERN (floater)), | |
ea52c577 | 7909 | XEXP (SET_SRC (PATTERN (floater)), |
7910 | 0), | |
7911 | XEXP (SET_SRC (PATTERN (floater)), | |
7912 | 1))) | |
bd49d362 | 7913 | break; |
7914 | } | |
7915 | } | |
7916 | } | |
7917 | ||
7918 | /* FLOATER will be nonzero if we found a suitable floating | |
7919 | insn for combination with ANCHOR. */ | |
7920 | if (floater | |
7921 | && (anchor_attr == PA_COMBINE_TYPE_FADDSUB | |
7922 | || anchor_attr == PA_COMBINE_TYPE_FMPY)) | |
7923 | { | |
7924 | /* Emit the new instruction and delete the old anchor. */ | |
7014838c | 7925 | emit_insn_before (gen_rtx_PARALLEL |
7926 | (VOIDmode, | |
7927 | gen_rtvec (2, PATTERN (anchor), | |
7928 | PATTERN (floater))), | |
7929 | anchor); | |
7930 | ||
bd49d362 | 7931 | PUT_CODE (anchor, NOTE); |
7932 | NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED; | |
7933 | NOTE_SOURCE_FILE (anchor) = 0; | |
7934 | ||
7935 | /* Emit a special USE insn for FLOATER, then delete | |
7936 | the floating insn. */ | |
ad851752 | 7937 | emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater); |
bd49d362 | 7938 | delete_insn (floater); |
7939 | ||
7940 | continue; | |
7941 | } | |
7942 | else if (floater | |
7943 | && anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH) | |
7944 | { | |
7945 | rtx temp; | |
7946 | /* Emit the new_jump instruction and delete the old anchor. */ | |
7014838c | 7947 | temp |
7948 | = emit_jump_insn_before (gen_rtx_PARALLEL | |
7949 | (VOIDmode, | |
7950 | gen_rtvec (2, PATTERN (anchor), | |
7951 | PATTERN (floater))), | |
7952 | anchor); | |
7953 | ||
bd49d362 | 7954 | JUMP_LABEL (temp) = JUMP_LABEL (anchor); |
7955 | PUT_CODE (anchor, NOTE); | |
7956 | NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED; | |
7957 | NOTE_SOURCE_FILE (anchor) = 0; | |
7958 | ||
7959 | /* Emit a special USE insn for FLOATER, then delete | |
7960 | the floating insn. */ | |
ad851752 | 7961 | emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater); |
bd49d362 | 7962 | delete_insn (floater); |
7963 | continue; | |
7964 | } | |
7965 | } | |
7966 | } | |
7967 | } | |
7968 | ||
9aadea62 | 7969 | static int |
bd49d362 | 7970 | pa_can_combine_p (new, anchor, floater, reversed, dest, src1, src2) |
7971 | rtx new, anchor, floater; | |
7972 | int reversed; | |
7973 | rtx dest, src1, src2; | |
7974 | { | |
7975 | int insn_code_number; | |
7976 | rtx start, end; | |
7977 | ||
7978 | /* Create a PARALLEL with the patterns of ANCHOR and | |
7979 | FLOATER, try to recognize it, then test constraints | |
7980 | for the resulting pattern. | |
7981 | ||
7982 | If the pattern doesn't match or the constraints | |
7983 | aren't met keep searching for a suitable floater | |
7984 | insn. */ | |
7985 | XVECEXP (PATTERN (new), 0, 0) = PATTERN (anchor); | |
7986 | XVECEXP (PATTERN (new), 0, 1) = PATTERN (floater); | |
7987 | INSN_CODE (new) = -1; | |
7988 | insn_code_number = recog_memoized (new); | |
7989 | if (insn_code_number < 0 | |
bf9fd117 | 7990 | || (extract_insn (new), ! constrain_operands (1))) |
bd49d362 | 7991 | return 0; |
7992 | ||
7993 | if (reversed) | |
7994 | { | |
7995 | start = anchor; | |
7996 | end = floater; | |
7997 | } | |
7998 | else | |
7999 | { | |
8000 | start = floater; | |
8001 | end = anchor; | |
8002 | } | |
8003 | ||
8004 | /* There's up to three operands to consider. One | |
8005 | output and two inputs. | |
8006 | ||
8007 | The output must not be used between FLOATER & ANCHOR | |
8008 | exclusive. The inputs must not be set between | |
8009 | FLOATER and ANCHOR exclusive. */ | |
8010 | ||
8011 | if (reg_used_between_p (dest, start, end)) | |
8012 | return 0; | |
8013 | ||
8014 | if (reg_set_between_p (src1, start, end)) | |
8015 | return 0; | |
8016 | ||
8017 | if (reg_set_between_p (src2, start, end)) | |
8018 | return 0; | |
8019 | ||
8020 | /* If we get here, then everything is good. */ | |
8021 | return 1; | |
8022 | } | |
14d18de3 | 8023 | |
a6582a53 | 8024 | /* Return nonzero if references for INSN are delayed. |
14d18de3 | 8025 | |
8026 | Millicode insns are actually function calls with some special | |
8027 | constraints on arguments and register usage. | |
8028 | ||
8029 | Millicode calls always expect their arguments in the integer argument | |
8030 | registers, and always return their result in %r29 (ret1). They | |
2013ddf6 | 8031 | are expected to clobber their arguments, %r1, %r29, and the return |
8032 | pointer which is %r31 on 32-bit and %r2 on 64-bit, and nothing else. | |
8033 | ||
8034 | This function tells reorg that the references to arguments and | |
8035 | millicode calls do not appear to happen until after the millicode call. | |
8036 | This allows reorg to put insns which set the argument registers into the | |
8037 | delay slot of the millicode call -- thus they act more like traditional | |
8038 | CALL_INSNs. | |
8039 | ||
8040 | Note we can not consider side effects of the insn to be delayed because | |
8041 | the branch and link insn will clobber the return pointer. If we happened | |
8042 | to use the return pointer in the delay slot of the call, then we lose. | |
14d18de3 | 8043 | |
8044 | get_attr_type will try to recognize the given insn, so make sure to | |
8045 | filter out things it will not accept -- SEQUENCE, USE and CLOBBER insns | |
8046 | in particular. */ | |
8047 | int | |
a6582a53 | 8048 | insn_refs_are_delayed (insn) |
14d18de3 | 8049 | rtx insn; |
8050 | { | |
9840d99d | 8051 | return ((GET_CODE (insn) == INSN |
14d18de3 | 8052 | && GET_CODE (PATTERN (insn)) != SEQUENCE |
8053 | && GET_CODE (PATTERN (insn)) != USE | |
8054 | && GET_CODE (PATTERN (insn)) != CLOBBER | |
8055 | && get_attr_type (insn) == TYPE_MILLI)); | |
8056 | } | |
5cb4669a | 8057 | |
58a72cce | 8058 | /* On the HP-PA the value is found in register(s) 28(-29), unless |
8059 | the mode is SF or DF. Then the value is returned in fr4 (32). | |
8060 | ||
8061 | This must perform the same promotions as PROMOTE_MODE, else | |
8062 | PROMOTE_FUNCTION_RETURN will not work correctly. | |
8063 | ||
8064 | Small structures must be returned in a PARALLEL on PA64 in order | |
8065 | to match the HP Compiler ABI. */ | |
8066 | ||
8067 | rtx | |
8068 | function_value (valtype, func) | |
8069 | tree valtype; | |
8070 | tree func ATTRIBUTE_UNUSED; | |
8071 | { | |
8072 | enum machine_mode valmode; | |
8073 | ||
8074 | /* Aggregates with a size less than or equal to 128 bits are returned | |
8075 | in GR 28(-29). They are left justified. The pad bits are undefined. | |
8076 | Larger aggregates are returned in memory. */ | |
8077 | if (TARGET_64BIT && AGGREGATE_TYPE_P (valtype)) | |
8078 | { | |
8079 | rtx loc[2]; | |
8080 | int i, offset = 0; | |
8081 | int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2; | |
8082 | ||
8083 | for (i = 0; i < ub; i++) | |
8084 | { | |
8085 | loc[i] = gen_rtx_EXPR_LIST (VOIDmode, | |
8086 | gen_rtx_REG (DImode, 28 + i), | |
8087 | GEN_INT (offset)); | |
8088 | offset += 8; | |
8089 | } | |
8090 | ||
8091 | return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc)); | |
8092 | } | |
8093 | ||
8094 | if ((INTEGRAL_TYPE_P (valtype) | |
8095 | && TYPE_PRECISION (valtype) < BITS_PER_WORD) | |
8096 | || POINTER_TYPE_P (valtype)) | |
8097 | valmode = word_mode; | |
8098 | else | |
8099 | valmode = TYPE_MODE (valtype); | |
8100 | ||
8101 | if (TREE_CODE (valtype) == REAL_TYPE | |
8102 | && TYPE_MODE (valtype) != TFmode | |
8103 | && !TARGET_SOFT_FLOAT) | |
8104 | return gen_rtx_REG (valmode, 32); | |
8105 | ||
8106 | return gen_rtx_REG (valmode, 28); | |
8107 | } | |
8108 | ||
5e3c5739 | 8109 | /* Return the location of a parameter that is passed in a register or NULL |
8110 | if the parameter has any component that is passed in memory. | |
8111 | ||
8112 | This is new code and will be pushed to into the net sources after | |
9840d99d | 8113 | further testing. |
5e3c5739 | 8114 | |
8115 | ??? We might want to restructure this so that it looks more like other | |
8116 | ports. */ | |
8117 | rtx | |
8118 | function_arg (cum, mode, type, named, incoming) | |
8119 | CUMULATIVE_ARGS *cum; | |
8120 | enum machine_mode mode; | |
8121 | tree type; | |
9aadea62 | 8122 | int named ATTRIBUTE_UNUSED; |
5e3c5739 | 8123 | int incoming; |
8124 | { | |
8125 | int max_arg_words = (TARGET_64BIT ? 8 : 4); | |
2a075f91 | 8126 | int alignment = 0; |
ac965869 | 8127 | int arg_size; |
5e3c5739 | 8128 | int fpr_reg_base; |
8129 | int gpr_reg_base; | |
8130 | rtx retval; | |
8131 | ||
ac965869 | 8132 | if (mode == VOIDmode) |
8133 | return NULL_RTX; | |
8134 | ||
8135 | arg_size = FUNCTION_ARG_SIZE (mode, type); | |
8136 | ||
8137 | /* If this arg would be passed partially or totally on the stack, then | |
8138 | this routine should return zero. FUNCTION_ARG_PARTIAL_NREGS will | |
8139 | handle arguments which are split between regs and stack slots if | |
8140 | the ABI mandates split arguments. */ | |
5e3c5739 | 8141 | if (! TARGET_64BIT) |
8142 | { | |
ac965869 | 8143 | /* The 32-bit ABI does not split arguments. */ |
8144 | if (cum->words + arg_size > max_arg_words) | |
5e3c5739 | 8145 | return NULL_RTX; |
8146 | } | |
8147 | else | |
8148 | { | |
2a075f91 | 8149 | if (arg_size > 1) |
8150 | alignment = cum->words & 1; | |
ac965869 | 8151 | if (cum->words + alignment >= max_arg_words) |
5e3c5739 | 8152 | return NULL_RTX; |
8153 | } | |
8154 | ||
8155 | /* The 32bit ABIs and the 64bit ABIs are rather different, | |
8156 | particularly in their handling of FP registers. We might | |
8157 | be able to cleverly share code between them, but I'm not | |
9aadea62 | 8158 | going to bother in the hope that splitting them up results |
2a075f91 | 8159 | in code that is more easily understood. */ |
5e3c5739 | 8160 | |
5e3c5739 | 8161 | if (TARGET_64BIT) |
8162 | { | |
8163 | /* Advance the base registers to their current locations. | |
8164 | ||
8165 | Remember, gprs grow towards smaller register numbers while | |
2a075f91 | 8166 | fprs grow to higher register numbers. Also remember that |
8167 | although FP regs are 32-bit addressable, we pretend that | |
8168 | the registers are 64-bits wide. */ | |
5e3c5739 | 8169 | gpr_reg_base = 26 - cum->words; |
8170 | fpr_reg_base = 32 + cum->words; | |
9840d99d | 8171 | |
ac965869 | 8172 | /* Arguments wider than one word and small aggregates need special |
8173 | treatment. */ | |
8174 | if (arg_size > 1 | |
8175 | || mode == BLKmode | |
8176 | || (type && AGGREGATE_TYPE_P (type))) | |
5e3c5739 | 8177 | { |
2a075f91 | 8178 | /* Double-extended precision (80-bit), quad-precision (128-bit) |
8179 | and aggregates including complex numbers are aligned on | |
8180 | 128-bit boundaries. The first eight 64-bit argument slots | |
8181 | are associated one-to-one, with general registers r26 | |
8182 | through r19, and also with floating-point registers fr4 | |
8183 | through fr11. Arguments larger than one word are always | |
ac965869 | 8184 | passed in general registers. |
8185 | ||
8186 | Using a PARALLEL with a word mode register results in left | |
8187 | justified data on a big-endian target. */ | |
2a075f91 | 8188 | |
8189 | rtx loc[8]; | |
8190 | int i, offset = 0, ub = arg_size; | |
8191 | ||
8192 | /* Align the base register. */ | |
8193 | gpr_reg_base -= alignment; | |
8194 | ||
8195 | ub = MIN (ub, max_arg_words - cum->words - alignment); | |
8196 | for (i = 0; i < ub; i++) | |
5e3c5739 | 8197 | { |
2a075f91 | 8198 | loc[i] = gen_rtx_EXPR_LIST (VOIDmode, |
8199 | gen_rtx_REG (DImode, gpr_reg_base), | |
8200 | GEN_INT (offset)); | |
8201 | gpr_reg_base -= 1; | |
8202 | offset += 8; | |
5e3c5739 | 8203 | } |
2a075f91 | 8204 | |
8205 | return gen_rtx_PARALLEL (mode, gen_rtvec_v (ub, loc)); | |
5e3c5739 | 8206 | } |
ac965869 | 8207 | } |
5e3c5739 | 8208 | else |
8209 | { | |
8210 | /* If the argument is larger than a word, then we know precisely | |
8211 | which registers we must use. */ | |
2a075f91 | 8212 | if (arg_size > 1) |
5e3c5739 | 8213 | { |
8214 | if (cum->words) | |
8215 | { | |
8216 | gpr_reg_base = 23; | |
8217 | fpr_reg_base = 38; | |
8218 | } | |
8219 | else | |
8220 | { | |
8221 | gpr_reg_base = 25; | |
8222 | fpr_reg_base = 34; | |
8223 | } | |
ac965869 | 8224 | |
8225 | /* Structures 5 to 8 bytes in size are passed in the general | |
8226 | registers in the same manner as other non floating-point | |
8227 | objects. The data is right-justified and zero-extended | |
8228 | to 64 bits. | |
8229 | ||
8230 | This is magic. Normally, using a PARALLEL results in left | |
8231 | justified data on a big-endian target. However, using a | |
8232 | single double-word register provides the required right | |
8233 | justication for 5 to 8 byte structures. This has nothing | |
8234 | to do with the direction of padding specified for the argument. | |
8235 | It has to do with how the data is widened and shifted into | |
8236 | and from the register. | |
8237 | ||
8238 | Aside from adding load_multiple and store_multiple patterns, | |
8239 | this is the only way that I have found to obtain right | |
8240 | justification of BLKmode data when it has a size greater | |
8241 | than one word. Splitting the operation into two SImode loads | |
8242 | or returning a DImode REG results in left justified data. */ | |
8243 | if (mode == BLKmode) | |
8244 | { | |
58a72cce | 8245 | rtx loc = gen_rtx_EXPR_LIST (VOIDmode, |
8246 | gen_rtx_REG (DImode, gpr_reg_base), | |
8247 | const0_rtx); | |
8248 | return gen_rtx_PARALLEL (mode, gen_rtvec (1, loc)); | |
ac965869 | 8249 | } |
5e3c5739 | 8250 | } |
8251 | else | |
8252 | { | |
8253 | /* We have a single word (32 bits). A simple computation | |
8254 | will get us the register #s we need. */ | |
8255 | gpr_reg_base = 26 - cum->words; | |
8256 | fpr_reg_base = 32 + 2 * cum->words; | |
8257 | } | |
8258 | } | |
8259 | ||
9503b0f7 | 8260 | /* Determine if the argument needs to be passed in both general and |
5e3c5739 | 8261 | floating point registers. */ |
9503b0f7 | 8262 | if (((TARGET_PORTABLE_RUNTIME || TARGET_64BIT || TARGET_ELF32) |
8263 | /* If we are doing soft-float with portable runtime, then there | |
8264 | is no need to worry about FP regs. */ | |
8265 | && ! TARGET_SOFT_FLOAT | |
8266 | /* The parameter must be some kind of float, else we can just | |
8267 | pass it in integer registers. */ | |
8268 | && FLOAT_MODE_P (mode) | |
8269 | /* The target function must not have a prototype. */ | |
8270 | && cum->nargs_prototype <= 0 | |
8271 | /* libcalls do not need to pass items in both FP and general | |
8272 | registers. */ | |
8273 | && type != NULL_TREE | |
8274 | /* All this hair applies to outgoing args only. */ | |
8275 | && ! incoming) | |
8276 | /* Also pass outgoing floating arguments in both registers in indirect | |
8277 | calls with the 32 bit ABI and the HP assembler since there is no | |
8278 | way to the specify argument locations in static functions. */ | |
8279 | || (! TARGET_64BIT | |
8280 | && ! TARGET_GAS | |
8281 | && ! incoming | |
8282 | && cum->indirect | |
8283 | && FLOAT_MODE_P (mode))) | |
5e3c5739 | 8284 | { |
8285 | retval | |
8286 | = gen_rtx_PARALLEL | |
8287 | (mode, | |
8288 | gen_rtvec (2, | |
8289 | gen_rtx_EXPR_LIST (VOIDmode, | |
8290 | gen_rtx_REG (mode, fpr_reg_base), | |
8291 | const0_rtx), | |
8292 | gen_rtx_EXPR_LIST (VOIDmode, | |
8293 | gen_rtx_REG (mode, gpr_reg_base), | |
8294 | const0_rtx))); | |
8295 | } | |
8296 | else | |
8297 | { | |
8298 | /* See if we should pass this parameter in a general register. */ | |
8299 | if (TARGET_SOFT_FLOAT | |
8300 | /* Indirect calls in the normal 32bit ABI require all arguments | |
8301 | to be passed in general registers. */ | |
8302 | || (!TARGET_PORTABLE_RUNTIME | |
8303 | && !TARGET_64BIT | |
a052da6f | 8304 | && !TARGET_ELF32 |
5e3c5739 | 8305 | && cum->indirect) |
8306 | /* If the parameter is not a floating point parameter, then | |
8307 | it belongs in GPRs. */ | |
8308 | || !FLOAT_MODE_P (mode)) | |
8309 | retval = gen_rtx_REG (mode, gpr_reg_base); | |
8310 | else | |
8311 | retval = gen_rtx_REG (mode, fpr_reg_base); | |
8312 | } | |
8313 | return retval; | |
8314 | } | |
8315 | ||
8316 | ||
8317 | /* If this arg would be passed totally in registers or totally on the stack, | |
8318 | then this routine should return zero. It is currently called only for | |
6dc3b0d9 | 8319 | the 64-bit target. */ |
5e3c5739 | 8320 | int |
8321 | function_arg_partial_nregs (cum, mode, type, named) | |
8322 | CUMULATIVE_ARGS *cum; | |
8323 | enum machine_mode mode; | |
8324 | tree type; | |
9aadea62 | 8325 | int named ATTRIBUTE_UNUSED; |
5e3c5739 | 8326 | { |
b7d86581 | 8327 | unsigned int max_arg_words = 8; |
8328 | unsigned int offset = 0; | |
5e3c5739 | 8329 | |
b7d86581 | 8330 | if (FUNCTION_ARG_SIZE (mode, type) > 1 && (cum->words & 1)) |
5e3c5739 | 8331 | offset = 1; |
8332 | ||
b7d86581 | 8333 | if (cum->words + offset + FUNCTION_ARG_SIZE (mode, type) <= max_arg_words) |
6dc3b0d9 | 8334 | /* Arg fits fully into registers. */ |
5e3c5739 | 8335 | return 0; |
9840d99d | 8336 | else if (cum->words + offset >= max_arg_words) |
6dc3b0d9 | 8337 | /* Arg fully on the stack. */ |
5e3c5739 | 8338 | return 0; |
8339 | else | |
6dc3b0d9 | 8340 | /* Arg is split. */ |
5e3c5739 | 8341 | return max_arg_words - cum->words - offset; |
5e3c5739 | 8342 | } |
8343 | ||
8344 | ||
8345 | /* Return 1 if this is a comparison operator. This allows the use of | |
8346 | MATCH_OPERATOR to recognize all the branch insns. */ | |
8347 | ||
8348 | int | |
8349 | cmpib_comparison_operator (op, mode) | |
8350 | register rtx op; | |
8351 | enum machine_mode mode; | |
8352 | { | |
8353 | return ((mode == VOIDmode || GET_MODE (op) == mode) | |
8354 | && (GET_CODE (op) == EQ | |
8355 | || GET_CODE (op) == NE | |
8356 | || GET_CODE (op) == GT | |
5e3c5739 | 8357 | || GET_CODE (op) == GTU |
32509e56 | 8358 | || GET_CODE (op) == GE |
5e3c5739 | 8359 | || GET_CODE (op) == LT |
8360 | || GET_CODE (op) == LE | |
8361 | || GET_CODE (op) == LEU)); | |
8362 | } | |
8363 | ||
52470889 | 8364 | /* On hpux10, the linker will give an error if we have a reference |
8365 | in the read-only data section to a symbol defined in a shared | |
8366 | library. Therefore, expressions that might require a reloc can | |
8367 | not be placed in the read-only data section. */ | |
8368 | ||
8369 | static void | |
8370 | pa_select_section (exp, reloc, align) | |
8371 | tree exp; | |
8372 | int reloc; | |
8373 | unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED; | |
8374 | { | |
8375 | if (TREE_CODE (exp) == VAR_DECL | |
8376 | && TREE_READONLY (exp) | |
8377 | && !TREE_THIS_VOLATILE (exp) | |
8378 | && DECL_INITIAL (exp) | |
8379 | && (DECL_INITIAL (exp) == error_mark_node | |
8380 | || TREE_CONSTANT (DECL_INITIAL (exp))) | |
8381 | && !reloc) | |
8382 | readonly_data_section (); | |
8383 | else if (TREE_CODE_CLASS (TREE_CODE (exp)) == 'c' | |
8384 | && !(TREE_CODE (exp) == STRING_CST && flag_writable_strings) | |
8385 | && !reloc) | |
8386 | readonly_data_section (); | |
8387 | else | |
8388 | data_section (); | |
8389 | } | |
1f3233d1 | 8390 | |
67c1e638 | 8391 | static void |
8392 | pa_globalize_label (stream, name) | |
8393 | FILE *stream; | |
8394 | const char *name; | |
8395 | { | |
8396 | /* We only handle DATA objects here, functions are globalized in | |
8397 | ASM_DECLARE_FUNCTION_NAME. */ | |
8398 | if (! FUNCTION_NAME_P (name)) | |
8399 | { | |
8400 | fputs ("\t.EXPORT ", stream); | |
8401 | assemble_name (stream, name); | |
8402 | fputs (",DATA\n", stream); | |
8403 | } | |
8404 | } | |
1f3233d1 | 8405 | #include "gt-pa.h" |