1 /* Subroutines for insn-output.c for HPPA.
2 Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
3 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
4 Contributed by Tim Moore (moore@cs.utah.edu), based on sparc.c
6 This file is part of GCC.
8 GCC 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)
13 GCC 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.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to
20 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
21 Boston, MA 02110-1301, USA. */
25 #include "coretypes.h"
29 #include "hard-reg-set.h"
31 #include "insn-config.h"
32 #include "conditions.h"
33 #include "insn-attr.h"
41 #include "integrate.h"
49 #include "target-def.h"
51 /* Return nonzero if there is a bypass for the output of
52 OUT_INSN and the fp store IN_INSN. */
54 hppa_fpstore_bypass_p (rtx out_insn, rtx in_insn)
56 enum machine_mode store_mode;
57 enum machine_mode other_mode;
60 if (recog_memoized (in_insn) < 0
61 || get_attr_type (in_insn) != TYPE_FPSTORE
62 || recog_memoized (out_insn) < 0)
65 store_mode = GET_MODE (SET_SRC (PATTERN (in_insn)));
67 set = single_set (out_insn);
71 other_mode = GET_MODE (SET_SRC (set));
73 return (GET_MODE_SIZE (store_mode) == GET_MODE_SIZE (other_mode));
77 #ifndef DO_FRAME_NOTES
78 #ifdef INCOMING_RETURN_ADDR_RTX
79 #define DO_FRAME_NOTES 1
81 #define DO_FRAME_NOTES 0
85 static void copy_reg_pointer (rtx, rtx);
86 static void fix_range (const char *);
87 static bool pa_handle_option (size_t, const char *, int);
88 static int hppa_address_cost (rtx);
89 static bool hppa_rtx_costs (rtx, int, int, int *);
90 static inline rtx force_mode (enum machine_mode, rtx);
91 static void pa_reorg (void);
92 static void pa_combine_instructions (void);
93 static int pa_can_combine_p (rtx, rtx, rtx, int, rtx, rtx, rtx);
94 static int forward_branch_p (rtx);
95 static void compute_zdepwi_operands (unsigned HOST_WIDE_INT, unsigned *);
96 static int compute_movmem_length (rtx);
97 static int compute_clrmem_length (rtx);
98 static bool pa_assemble_integer (rtx, unsigned int, int);
99 static void remove_useless_addtr_insns (int);
100 static void store_reg (int, HOST_WIDE_INT, int);
101 static void store_reg_modify (int, int, HOST_WIDE_INT);
102 static void load_reg (int, HOST_WIDE_INT, int);
103 static void set_reg_plus_d (int, int, HOST_WIDE_INT, int);
104 static void pa_output_function_prologue (FILE *, HOST_WIDE_INT);
105 static void update_total_code_bytes (int);
106 static void pa_output_function_epilogue (FILE *, HOST_WIDE_INT);
107 static int pa_adjust_cost (rtx, rtx, rtx, int);
108 static int pa_adjust_priority (rtx, int);
109 static int pa_issue_rate (void);
110 static void pa_som_asm_init_sections (void) ATTRIBUTE_UNUSED;
111 static section *pa_select_section (tree, int, unsigned HOST_WIDE_INT)
113 static void pa_encode_section_info (tree, rtx, int);
114 static const char *pa_strip_name_encoding (const char *);
115 static bool pa_function_ok_for_sibcall (tree, tree);
116 static void pa_globalize_label (FILE *, const char *)
118 static void pa_asm_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
119 HOST_WIDE_INT, tree);
120 #if !defined(USE_COLLECT2)
121 static void pa_asm_out_constructor (rtx, int);
122 static void pa_asm_out_destructor (rtx, int);
124 static void pa_init_builtins (void);
125 static rtx hppa_builtin_saveregs (void);
126 static tree hppa_gimplify_va_arg_expr (tree, tree, tree *, tree *);
127 static bool pa_scalar_mode_supported_p (enum machine_mode);
128 static bool pa_commutative_p (rtx x, int outer_code);
129 static void copy_fp_args (rtx) ATTRIBUTE_UNUSED;
130 static int length_fp_args (rtx) ATTRIBUTE_UNUSED;
131 static inline void pa_file_start_level (void) ATTRIBUTE_UNUSED;
132 static inline void pa_file_start_space (int) ATTRIBUTE_UNUSED;
133 static inline void pa_file_start_file (int) ATTRIBUTE_UNUSED;
134 static inline void pa_file_start_mcount (const char*) ATTRIBUTE_UNUSED;
135 static void pa_elf_file_start (void) ATTRIBUTE_UNUSED;
136 static void pa_som_file_start (void) ATTRIBUTE_UNUSED;
137 static void pa_linux_file_start (void) ATTRIBUTE_UNUSED;
138 static void pa_hpux64_gas_file_start (void) ATTRIBUTE_UNUSED;
139 static void pa_hpux64_hpas_file_start (void) ATTRIBUTE_UNUSED;
140 static void output_deferred_plabels (void);
141 static void output_deferred_profile_counters (void) ATTRIBUTE_UNUSED;
142 #ifdef ASM_OUTPUT_EXTERNAL_REAL
143 static void pa_hpux_file_end (void);
145 #ifdef HPUX_LONG_DOUBLE_LIBRARY
146 static void pa_hpux_init_libfuncs (void);
148 static rtx pa_struct_value_rtx (tree, int);
149 static bool pa_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
151 static int pa_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
153 static struct machine_function * pa_init_machine_status (void);
154 static enum reg_class pa_secondary_reload (bool, rtx, enum reg_class,
156 secondary_reload_info *);
159 /* The following extra sections are only used for SOM. */
160 static GTY(()) section *som_readonly_data_section;
161 static GTY(()) section *som_one_only_readonly_data_section;
162 static GTY(()) section *som_one_only_data_section;
164 /* Save the operands last given to a compare for use when we
165 generate a scc or bcc insn. */
166 rtx hppa_compare_op0, hppa_compare_op1;
167 enum cmp_type hppa_branch_type;
169 /* Which cpu we are scheduling for. */
170 enum processor_type pa_cpu = TARGET_SCHED_DEFAULT;
172 /* The UNIX standard to use for predefines and linking. */
173 int flag_pa_unix = TARGET_HPUX_11_11 ? 1998 : TARGET_HPUX_10_10 ? 1995 : 1993;
175 /* Counts for the number of callee-saved general and floating point
176 registers which were saved by the current function's prologue. */
177 static int gr_saved, fr_saved;
179 static rtx find_addr_reg (rtx);
181 /* Keep track of the number of bytes we have output in the CODE subspace
182 during this compilation so we'll know when to emit inline long-calls. */
183 unsigned long total_code_bytes;
185 /* The last address of the previous function plus the number of bytes in
186 associated thunks that have been output. This is used to determine if
187 a thunk can use an IA-relative branch to reach its target function. */
188 static int last_address;
190 /* Variables to handle plabels that we discover are necessary at assembly
191 output time. They are output after the current function. */
192 struct deferred_plabel GTY(())
197 static GTY((length ("n_deferred_plabels"))) struct deferred_plabel *
199 static size_t n_deferred_plabels = 0;
202 /* Initialize the GCC target structure. */
204 #undef TARGET_ASM_ALIGNED_HI_OP
205 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
206 #undef TARGET_ASM_ALIGNED_SI_OP
207 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
208 #undef TARGET_ASM_ALIGNED_DI_OP
209 #define TARGET_ASM_ALIGNED_DI_OP "\t.dword\t"
210 #undef TARGET_ASM_UNALIGNED_HI_OP
211 #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP
212 #undef TARGET_ASM_UNALIGNED_SI_OP
213 #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP
214 #undef TARGET_ASM_UNALIGNED_DI_OP
215 #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP
216 #undef TARGET_ASM_INTEGER
217 #define TARGET_ASM_INTEGER pa_assemble_integer
219 #undef TARGET_ASM_FUNCTION_PROLOGUE
220 #define TARGET_ASM_FUNCTION_PROLOGUE pa_output_function_prologue
221 #undef TARGET_ASM_FUNCTION_EPILOGUE
222 #define TARGET_ASM_FUNCTION_EPILOGUE pa_output_function_epilogue
224 #undef TARGET_SCHED_ADJUST_COST
225 #define TARGET_SCHED_ADJUST_COST pa_adjust_cost
226 #undef TARGET_SCHED_ADJUST_PRIORITY
227 #define TARGET_SCHED_ADJUST_PRIORITY pa_adjust_priority
228 #undef TARGET_SCHED_ISSUE_RATE
229 #define TARGET_SCHED_ISSUE_RATE pa_issue_rate
231 #undef TARGET_ENCODE_SECTION_INFO
232 #define TARGET_ENCODE_SECTION_INFO pa_encode_section_info
233 #undef TARGET_STRIP_NAME_ENCODING
234 #define TARGET_STRIP_NAME_ENCODING pa_strip_name_encoding
236 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
237 #define TARGET_FUNCTION_OK_FOR_SIBCALL pa_function_ok_for_sibcall
239 #undef TARGET_COMMUTATIVE_P
240 #define TARGET_COMMUTATIVE_P pa_commutative_p
242 #undef TARGET_ASM_OUTPUT_MI_THUNK
243 #define TARGET_ASM_OUTPUT_MI_THUNK pa_asm_output_mi_thunk
244 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
245 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall
247 #undef TARGET_ASM_FILE_END
248 #ifdef ASM_OUTPUT_EXTERNAL_REAL
249 #define TARGET_ASM_FILE_END pa_hpux_file_end
251 #define TARGET_ASM_FILE_END output_deferred_plabels
254 #if !defined(USE_COLLECT2)
255 #undef TARGET_ASM_CONSTRUCTOR
256 #define TARGET_ASM_CONSTRUCTOR pa_asm_out_constructor
257 #undef TARGET_ASM_DESTRUCTOR
258 #define TARGET_ASM_DESTRUCTOR pa_asm_out_destructor
261 #undef TARGET_DEFAULT_TARGET_FLAGS
262 #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT | TARGET_CPU_DEFAULT)
263 #undef TARGET_HANDLE_OPTION
264 #define TARGET_HANDLE_OPTION pa_handle_option
266 #undef TARGET_INIT_BUILTINS
267 #define TARGET_INIT_BUILTINS pa_init_builtins
269 #undef TARGET_RTX_COSTS
270 #define TARGET_RTX_COSTS hppa_rtx_costs
271 #undef TARGET_ADDRESS_COST
272 #define TARGET_ADDRESS_COST hppa_address_cost
274 #undef TARGET_MACHINE_DEPENDENT_REORG
275 #define TARGET_MACHINE_DEPENDENT_REORG pa_reorg
277 #ifdef HPUX_LONG_DOUBLE_LIBRARY
278 #undef TARGET_INIT_LIBFUNCS
279 #define TARGET_INIT_LIBFUNCS pa_hpux_init_libfuncs
282 #undef TARGET_PROMOTE_FUNCTION_RETURN
283 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
284 #undef TARGET_PROMOTE_PROTOTYPES
285 #define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
287 #undef TARGET_STRUCT_VALUE_RTX
288 #define TARGET_STRUCT_VALUE_RTX pa_struct_value_rtx
289 #undef TARGET_RETURN_IN_MEMORY
290 #define TARGET_RETURN_IN_MEMORY pa_return_in_memory
291 #undef TARGET_MUST_PASS_IN_STACK
292 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
293 #undef TARGET_PASS_BY_REFERENCE
294 #define TARGET_PASS_BY_REFERENCE pa_pass_by_reference
295 #undef TARGET_CALLEE_COPIES
296 #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_true
297 #undef TARGET_ARG_PARTIAL_BYTES
298 #define TARGET_ARG_PARTIAL_BYTES pa_arg_partial_bytes
300 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
301 #define TARGET_EXPAND_BUILTIN_SAVEREGS hppa_builtin_saveregs
302 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
303 #define TARGET_GIMPLIFY_VA_ARG_EXPR hppa_gimplify_va_arg_expr
305 #undef TARGET_SCALAR_MODE_SUPPORTED_P
306 #define TARGET_SCALAR_MODE_SUPPORTED_P pa_scalar_mode_supported_p
308 #undef TARGET_CANNOT_FORCE_CONST_MEM
309 #define TARGET_CANNOT_FORCE_CONST_MEM pa_tls_referenced_p
311 #undef TARGET_SECONDARY_RELOAD
312 #define TARGET_SECONDARY_RELOAD pa_secondary_reload
314 struct gcc_target targetm = TARGET_INITIALIZER;
316 /* Parse the -mfixed-range= option string. */
319 fix_range (const char *const_str)
322 char *str, *dash, *comma;
324 /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and
325 REG2 are either register names or register numbers. The effect
326 of this option is to mark the registers in the range from REG1 to
327 REG2 as ``fixed'' so they won't be used by the compiler. This is
328 used, e.g., to ensure that kernel mode code doesn't use fr4-fr31. */
330 i = strlen (const_str);
331 str = (char *) alloca (i + 1);
332 memcpy (str, const_str, i + 1);
336 dash = strchr (str, '-');
339 warning (0, "value of -mfixed-range must have form REG1-REG2");
344 comma = strchr (dash + 1, ',');
348 first = decode_reg_name (str);
351 warning (0, "unknown register name: %s", str);
355 last = decode_reg_name (dash + 1);
358 warning (0, "unknown register name: %s", dash + 1);
366 warning (0, "%s-%s is an empty range", str, dash + 1);
370 for (i = first; i <= last; ++i)
371 fixed_regs[i] = call_used_regs[i] = 1;
380 /* Check if all floating point registers have been fixed. */
381 for (i = FP_REG_FIRST; i <= FP_REG_LAST; i++)
386 target_flags |= MASK_DISABLE_FPREGS;
389 /* Implement TARGET_HANDLE_OPTION. */
392 pa_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED)
397 case OPT_mpa_risc_1_0:
399 target_flags &= ~(MASK_PA_11 | MASK_PA_20);
403 case OPT_mpa_risc_1_1:
405 target_flags &= ~MASK_PA_20;
406 target_flags |= MASK_PA_11;
409 case OPT_mpa_risc_2_0:
411 target_flags |= MASK_PA_11 | MASK_PA_20;
415 if (strcmp (arg, "8000") == 0)
416 pa_cpu = PROCESSOR_8000;
417 else if (strcmp (arg, "7100") == 0)
418 pa_cpu = PROCESSOR_7100;
419 else if (strcmp (arg, "700") == 0)
420 pa_cpu = PROCESSOR_700;
421 else if (strcmp (arg, "7100LC") == 0)
422 pa_cpu = PROCESSOR_7100LC;
423 else if (strcmp (arg, "7200") == 0)
424 pa_cpu = PROCESSOR_7200;
425 else if (strcmp (arg, "7300") == 0)
426 pa_cpu = PROCESSOR_7300;
431 case OPT_mfixed_range_:
441 #if TARGET_HPUX_10_10
447 #if TARGET_HPUX_11_11
459 override_options (void)
461 /* Unconditional branches in the delay slot are not compatible with dwarf2
462 call frame information. There is no benefit in using this optimization
463 on PA8000 and later processors. */
464 if (pa_cpu >= PROCESSOR_8000
465 || (! USING_SJLJ_EXCEPTIONS && flag_exceptions)
466 || flag_unwind_tables)
467 target_flags &= ~MASK_JUMP_IN_DELAY;
469 if (flag_pic && TARGET_PORTABLE_RUNTIME)
471 warning (0, "PIC code generation is not supported in the portable runtime model");
474 if (flag_pic && TARGET_FAST_INDIRECT_CALLS)
476 warning (0, "PIC code generation is not compatible with fast indirect calls");
479 if (! TARGET_GAS && write_symbols != NO_DEBUG)
481 warning (0, "-g is only supported when using GAS on this processor,");
482 warning (0, "-g option disabled");
483 write_symbols = NO_DEBUG;
486 /* We only support the "big PIC" model now. And we always generate PIC
487 code when in 64bit mode. */
488 if (flag_pic == 1 || TARGET_64BIT)
491 /* We can't guarantee that .dword is available for 32-bit targets. */
492 if (UNITS_PER_WORD == 4)
493 targetm.asm_out.aligned_op.di = NULL;
495 /* The unaligned ops are only available when using GAS. */
498 targetm.asm_out.unaligned_op.hi = NULL;
499 targetm.asm_out.unaligned_op.si = NULL;
500 targetm.asm_out.unaligned_op.di = NULL;
503 init_machine_status = pa_init_machine_status;
507 pa_init_builtins (void)
509 #ifdef DONT_HAVE_FPUTC_UNLOCKED
510 built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] =
511 built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
512 implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED]
513 = implicit_built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
517 /* Function to init struct machine_function.
518 This will be called, via a pointer variable,
519 from push_function_context. */
521 static struct machine_function *
522 pa_init_machine_status (void)
524 return ggc_alloc_cleared (sizeof (machine_function));
527 /* If FROM is a probable pointer register, mark TO as a probable
528 pointer register with the same pointer alignment as FROM. */
531 copy_reg_pointer (rtx to, rtx from)
533 if (REG_POINTER (from))
534 mark_reg_pointer (to, REGNO_POINTER_ALIGN (REGNO (from)));
537 /* Return 1 if X contains a symbolic expression. We know these
538 expressions will have one of a few well defined forms, so
539 we need only check those forms. */
541 symbolic_expression_p (rtx x)
544 /* Strip off any HIGH. */
545 if (GET_CODE (x) == HIGH)
548 return (symbolic_operand (x, VOIDmode));
551 /* Accept any constant that can be moved in one instruction into a
554 cint_ok_for_move (HOST_WIDE_INT intval)
556 /* OK if ldo, ldil, or zdepi, can be used. */
557 return (CONST_OK_FOR_LETTER_P (intval, 'J')
558 || CONST_OK_FOR_LETTER_P (intval, 'N')
559 || CONST_OK_FOR_LETTER_P (intval, 'K'));
562 /* Return truth value of whether OP can be used as an operand in a
565 adddi3_operand (rtx op, enum machine_mode mode)
567 return (register_operand (op, mode)
568 || (GET_CODE (op) == CONST_INT
569 && (TARGET_64BIT ? INT_14_BITS (op) : INT_11_BITS (op))));
572 /* True iff zdepi can be used to generate this CONST_INT.
573 zdepi first sign extends a 5 bit signed number to a given field
574 length, then places this field anywhere in a zero. */
576 zdepi_cint_p (unsigned HOST_WIDE_INT x)
578 unsigned HOST_WIDE_INT lsb_mask, t;
580 /* This might not be obvious, but it's at least fast.
581 This function is critical; we don't have the time loops would take. */
583 t = ((x >> 4) + lsb_mask) & ~(lsb_mask - 1);
584 /* Return true iff t is a power of two. */
585 return ((t & (t - 1)) == 0);
588 /* True iff depi or extru can be used to compute (reg & mask).
589 Accept bit pattern like these:
594 and_mask_p (unsigned HOST_WIDE_INT mask)
597 mask += mask & -mask;
598 return (mask & (mask - 1)) == 0;
601 /* True iff depi can be used to compute (reg | MASK). */
603 ior_mask_p (unsigned HOST_WIDE_INT mask)
605 mask += mask & -mask;
606 return (mask & (mask - 1)) == 0;
609 /* Legitimize PIC addresses. If the address is already
610 position-independent, we return ORIG. Newly generated
611 position-independent addresses go to REG. If we need more
612 than one register, we lose. */
615 legitimize_pic_address (rtx orig, enum machine_mode mode, rtx reg)
619 gcc_assert (!PA_SYMBOL_REF_TLS_P (orig));
621 /* Labels need special handling. */
622 if (pic_label_operand (orig, mode))
624 /* We do not want to go through the movXX expanders here since that
625 would create recursion.
627 Nor do we really want to call a generator for a named pattern
628 since that requires multiple patterns if we want to support
631 So instead we just emit the raw set, which avoids the movXX
632 expanders completely. */
633 mark_reg_pointer (reg, BITS_PER_UNIT);
634 emit_insn (gen_rtx_SET (VOIDmode, reg, orig));
635 current_function_uses_pic_offset_table = 1;
638 if (GET_CODE (orig) == SYMBOL_REF)
644 /* Before reload, allocate a temporary register for the intermediate
645 result. This allows the sequence to be deleted when the final
646 result is unused and the insns are trivially dead. */
647 tmp_reg = ((reload_in_progress || reload_completed)
648 ? reg : gen_reg_rtx (Pmode));
650 emit_move_insn (tmp_reg,
651 gen_rtx_PLUS (word_mode, pic_offset_table_rtx,
652 gen_rtx_HIGH (word_mode, orig)));
654 = gen_const_mem (Pmode,
655 gen_rtx_LO_SUM (Pmode, tmp_reg,
656 gen_rtx_UNSPEC (Pmode,
660 current_function_uses_pic_offset_table = 1;
661 mark_reg_pointer (reg, BITS_PER_UNIT);
662 insn = emit_move_insn (reg, pic_ref);
664 /* Put a REG_EQUAL note on this insn, so that it can be optimized. */
665 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig, REG_NOTES (insn));
669 else if (GET_CODE (orig) == CONST)
673 if (GET_CODE (XEXP (orig, 0)) == PLUS
674 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
678 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
680 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
681 orig = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
682 base == reg ? 0 : reg);
684 if (GET_CODE (orig) == CONST_INT)
686 if (INT_14_BITS (orig))
687 return plus_constant (base, INTVAL (orig));
688 orig = force_reg (Pmode, orig);
690 pic_ref = gen_rtx_PLUS (Pmode, base, orig);
691 /* Likewise, should we set special REG_NOTEs here? */
697 static GTY(()) rtx gen_tls_tga;
700 gen_tls_get_addr (void)
703 gen_tls_tga = init_one_libfunc ("__tls_get_addr");
708 hppa_tls_call (rtx arg)
712 ret = gen_reg_rtx (Pmode);
713 emit_library_call_value (gen_tls_get_addr (), ret,
714 LCT_CONST, Pmode, 1, arg, Pmode);
720 legitimize_tls_address (rtx addr)
722 rtx ret, insn, tmp, t1, t2, tp;
723 enum tls_model model = SYMBOL_REF_TLS_MODEL (addr);
727 case TLS_MODEL_GLOBAL_DYNAMIC:
728 tmp = gen_reg_rtx (Pmode);
729 emit_insn (gen_tgd_load (tmp, addr));
730 ret = hppa_tls_call (tmp);
733 case TLS_MODEL_LOCAL_DYNAMIC:
734 ret = gen_reg_rtx (Pmode);
735 tmp = gen_reg_rtx (Pmode);
737 emit_insn (gen_tld_load (tmp, addr));
738 t1 = hppa_tls_call (tmp);
741 t2 = gen_reg_rtx (Pmode);
742 emit_libcall_block (insn, t2, t1,
743 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
745 emit_insn (gen_tld_offset_load (ret, addr, t2));
748 case TLS_MODEL_INITIAL_EXEC:
749 tp = gen_reg_rtx (Pmode);
750 tmp = gen_reg_rtx (Pmode);
751 ret = gen_reg_rtx (Pmode);
752 emit_insn (gen_tp_load (tp));
753 emit_insn (gen_tie_load (tmp, addr));
754 emit_move_insn (ret, gen_rtx_PLUS (Pmode, tp, tmp));
757 case TLS_MODEL_LOCAL_EXEC:
758 tp = gen_reg_rtx (Pmode);
759 ret = gen_reg_rtx (Pmode);
760 emit_insn (gen_tp_load (tp));
761 emit_insn (gen_tle_load (ret, addr, tp));
771 /* Try machine-dependent ways of modifying an illegitimate address
772 to be legitimate. If we find one, return the new, valid address.
773 This macro is used in only one place: `memory_address' in explow.c.
775 OLDX is the address as it was before break_out_memory_refs was called.
776 In some cases it is useful to look at this to decide what needs to be done.
778 MODE and WIN are passed so that this macro can use
779 GO_IF_LEGITIMATE_ADDRESS.
781 It is always safe for this macro to do nothing. It exists to recognize
782 opportunities to optimize the output.
784 For the PA, transform:
786 memory(X + <large int>)
790 if (<large int> & mask) >= 16
791 Y = (<large int> & ~mask) + mask + 1 Round up.
793 Y = (<large int> & ~mask) Round down.
795 memory (Z + (<large int> - Y));
797 This is for CSE to find several similar references, and only use one Z.
799 X can either be a SYMBOL_REF or REG, but because combine cannot
800 perform a 4->2 combination we do nothing for SYMBOL_REF + D where
801 D will not fit in 14 bits.
803 MODE_FLOAT references allow displacements which fit in 5 bits, so use
806 MODE_INT references allow displacements which fit in 14 bits, so use
809 This relies on the fact that most mode MODE_FLOAT references will use FP
810 registers and most mode MODE_INT references will use integer registers.
811 (In the rare case of an FP register used in an integer MODE, we depend
812 on secondary reloads to clean things up.)
815 It is also beneficial to handle (plus (mult (X) (Y)) (Z)) in a special
816 manner if Y is 2, 4, or 8. (allows more shadd insns and shifted indexed
817 addressing modes to be used).
819 Put X and Z into registers. Then put the entire expression into
823 hppa_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
824 enum machine_mode mode)
828 /* We need to canonicalize the order of operands in unscaled indexed
829 addresses since the code that checks if an address is valid doesn't
830 always try both orders. */
831 if (!TARGET_NO_SPACE_REGS
832 && GET_CODE (x) == PLUS
833 && GET_MODE (x) == Pmode
834 && REG_P (XEXP (x, 0))
835 && REG_P (XEXP (x, 1))
836 && REG_POINTER (XEXP (x, 0))
837 && !REG_POINTER (XEXP (x, 1)))
838 return gen_rtx_PLUS (Pmode, XEXP (x, 1), XEXP (x, 0));
840 if (PA_SYMBOL_REF_TLS_P (x))
841 return legitimize_tls_address (x);
843 return legitimize_pic_address (x, mode, gen_reg_rtx (Pmode));
845 /* Strip off CONST. */
846 if (GET_CODE (x) == CONST)
849 /* Special case. Get the SYMBOL_REF into a register and use indexing.
850 That should always be safe. */
851 if (GET_CODE (x) == PLUS
852 && GET_CODE (XEXP (x, 0)) == REG
853 && GET_CODE (XEXP (x, 1)) == SYMBOL_REF)
855 rtx reg = force_reg (Pmode, XEXP (x, 1));
856 return force_reg (Pmode, gen_rtx_PLUS (Pmode, reg, XEXP (x, 0)));
859 /* Note we must reject symbols which represent function addresses
860 since the assembler/linker can't handle arithmetic on plabels. */
861 if (GET_CODE (x) == PLUS
862 && GET_CODE (XEXP (x, 1)) == CONST_INT
863 && ((GET_CODE (XEXP (x, 0)) == SYMBOL_REF
864 && !FUNCTION_NAME_P (XSTR (XEXP (x, 0), 0)))
865 || GET_CODE (XEXP (x, 0)) == REG))
867 rtx int_part, ptr_reg;
869 int offset = INTVAL (XEXP (x, 1));
872 mask = (GET_MODE_CLASS (mode) == MODE_FLOAT
873 ? (TARGET_PA_20 ? 0x3fff : 0x1f) : 0x3fff);
875 /* Choose which way to round the offset. Round up if we
876 are >= halfway to the next boundary. */
877 if ((offset & mask) >= ((mask + 1) / 2))
878 newoffset = (offset & ~ mask) + mask + 1;
880 newoffset = (offset & ~ mask);
882 /* If the newoffset will not fit in 14 bits (ldo), then
883 handling this would take 4 or 5 instructions (2 to load
884 the SYMBOL_REF + 1 or 2 to load the newoffset + 1 to
885 add the new offset and the SYMBOL_REF.) Combine can
886 not handle 4->2 or 5->2 combinations, so do not create
888 if (! VAL_14_BITS_P (newoffset)
889 && GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
891 rtx const_part = plus_constant (XEXP (x, 0), newoffset);
894 gen_rtx_HIGH (Pmode, const_part));
897 gen_rtx_LO_SUM (Pmode,
898 tmp_reg, const_part));
902 if (! VAL_14_BITS_P (newoffset))
903 int_part = force_reg (Pmode, GEN_INT (newoffset));
905 int_part = GEN_INT (newoffset);
907 ptr_reg = force_reg (Pmode,
909 force_reg (Pmode, XEXP (x, 0)),
912 return plus_constant (ptr_reg, offset - newoffset);
915 /* Handle (plus (mult (a) (shadd_constant)) (b)). */
917 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT
918 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
919 && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))
920 && (OBJECT_P (XEXP (x, 1))
921 || GET_CODE (XEXP (x, 1)) == SUBREG)
922 && GET_CODE (XEXP (x, 1)) != CONST)
924 int val = INTVAL (XEXP (XEXP (x, 0), 1));
928 if (GET_CODE (reg1) != REG)
929 reg1 = force_reg (Pmode, force_operand (reg1, 0));
931 reg2 = XEXP (XEXP (x, 0), 0);
932 if (GET_CODE (reg2) != REG)
933 reg2 = force_reg (Pmode, force_operand (reg2, 0));
935 return force_reg (Pmode, gen_rtx_PLUS (Pmode,
942 /* Similarly for (plus (plus (mult (a) (shadd_constant)) (b)) (c)).
944 Only do so for floating point modes since this is more speculative
945 and we lose if it's an integer store. */
946 if (GET_CODE (x) == PLUS
947 && GET_CODE (XEXP (x, 0)) == PLUS
948 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
949 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
950 && shadd_constant_p (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)))
951 && (mode == SFmode || mode == DFmode))
954 /* First, try and figure out what to use as a base register. */
955 rtx reg1, reg2, base, idx, orig_base;
957 reg1 = XEXP (XEXP (x, 0), 1);
962 /* Make sure they're both regs. If one was a SYMBOL_REF [+ const],
963 then emit_move_sequence will turn on REG_POINTER so we'll know
964 it's a base register below. */
965 if (GET_CODE (reg1) != REG)
966 reg1 = force_reg (Pmode, force_operand (reg1, 0));
968 if (GET_CODE (reg2) != REG)
969 reg2 = force_reg (Pmode, force_operand (reg2, 0));
971 /* Figure out what the base and index are. */
973 if (GET_CODE (reg1) == REG
974 && REG_POINTER (reg1))
977 orig_base = XEXP (XEXP (x, 0), 1);
978 idx = gen_rtx_PLUS (Pmode,
980 XEXP (XEXP (XEXP (x, 0), 0), 0),
981 XEXP (XEXP (XEXP (x, 0), 0), 1)),
984 else if (GET_CODE (reg2) == REG
985 && REG_POINTER (reg2))
988 orig_base = XEXP (x, 1);
995 /* If the index adds a large constant, try to scale the
996 constant so that it can be loaded with only one insn. */
997 if (GET_CODE (XEXP (idx, 1)) == CONST_INT
998 && VAL_14_BITS_P (INTVAL (XEXP (idx, 1))
999 / INTVAL (XEXP (XEXP (idx, 0), 1)))
1000 && INTVAL (XEXP (idx, 1)) % INTVAL (XEXP (XEXP (idx, 0), 1)) == 0)
1002 /* Divide the CONST_INT by the scale factor, then add it to A. */
1003 int val = INTVAL (XEXP (idx, 1));
1005 val /= INTVAL (XEXP (XEXP (idx, 0), 1));
1006 reg1 = XEXP (XEXP (idx, 0), 0);
1007 if (GET_CODE (reg1) != REG)
1008 reg1 = force_reg (Pmode, force_operand (reg1, 0));
1010 reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, reg1, GEN_INT (val)));
1012 /* We can now generate a simple scaled indexed address. */
1015 (Pmode, gen_rtx_PLUS (Pmode,
1016 gen_rtx_MULT (Pmode, reg1,
1017 XEXP (XEXP (idx, 0), 1)),
1021 /* If B + C is still a valid base register, then add them. */
1022 if (GET_CODE (XEXP (idx, 1)) == CONST_INT
1023 && INTVAL (XEXP (idx, 1)) <= 4096
1024 && INTVAL (XEXP (idx, 1)) >= -4096)
1026 int val = INTVAL (XEXP (XEXP (idx, 0), 1));
1029 reg1 = force_reg (Pmode, gen_rtx_PLUS (Pmode, base, XEXP (idx, 1)));
1031 reg2 = XEXP (XEXP (idx, 0), 0);
1032 if (GET_CODE (reg2) != CONST_INT)
1033 reg2 = force_reg (Pmode, force_operand (reg2, 0));
1035 return force_reg (Pmode, gen_rtx_PLUS (Pmode,
1036 gen_rtx_MULT (Pmode,
1042 /* Get the index into a register, then add the base + index and
1043 return a register holding the result. */
1045 /* First get A into a register. */
1046 reg1 = XEXP (XEXP (idx, 0), 0);
1047 if (GET_CODE (reg1) != REG)
1048 reg1 = force_reg (Pmode, force_operand (reg1, 0));
1050 /* And get B into a register. */
1051 reg2 = XEXP (idx, 1);
1052 if (GET_CODE (reg2) != REG)
1053 reg2 = force_reg (Pmode, force_operand (reg2, 0));
1055 reg1 = force_reg (Pmode,
1056 gen_rtx_PLUS (Pmode,
1057 gen_rtx_MULT (Pmode, reg1,
1058 XEXP (XEXP (idx, 0), 1)),
1061 /* Add the result to our base register and return. */
1062 return force_reg (Pmode, gen_rtx_PLUS (Pmode, base, reg1));
1066 /* Uh-oh. We might have an address for x[n-100000]. This needs
1067 special handling to avoid creating an indexed memory address
1068 with x-100000 as the base.
1070 If the constant part is small enough, then it's still safe because
1071 there is a guard page at the beginning and end of the data segment.
1073 Scaled references are common enough that we want to try and rearrange the
1074 terms so that we can use indexing for these addresses too. Only
1075 do the optimization for floatint point modes. */
1077 if (GET_CODE (x) == PLUS
1078 && symbolic_expression_p (XEXP (x, 1)))
1080 /* Ugly. We modify things here so that the address offset specified
1081 by the index expression is computed first, then added to x to form
1082 the entire address. */
1084 rtx regx1, regx2, regy1, regy2, y;
1086 /* Strip off any CONST. */
1088 if (GET_CODE (y) == CONST)
1091 if (GET_CODE (y) == PLUS || GET_CODE (y) == MINUS)
1093 /* See if this looks like
1094 (plus (mult (reg) (shadd_const))
1095 (const (plus (symbol_ref) (const_int))))
1097 Where const_int is small. In that case the const
1098 expression is a valid pointer for indexing.
1100 If const_int is big, but can be divided evenly by shadd_const
1101 and added to (reg). This allows more scaled indexed addresses. */
1102 if (GET_CODE (XEXP (y, 0)) == SYMBOL_REF
1103 && GET_CODE (XEXP (x, 0)) == MULT
1104 && GET_CODE (XEXP (y, 1)) == CONST_INT
1105 && INTVAL (XEXP (y, 1)) >= -4096
1106 && INTVAL (XEXP (y, 1)) <= 4095
1107 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
1108 && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1))))
1110 int val = INTVAL (XEXP (XEXP (x, 0), 1));
1114 if (GET_CODE (reg1) != REG)
1115 reg1 = force_reg (Pmode, force_operand (reg1, 0));
1117 reg2 = XEXP (XEXP (x, 0), 0);
1118 if (GET_CODE (reg2) != REG)
1119 reg2 = force_reg (Pmode, force_operand (reg2, 0));
1121 return force_reg (Pmode,
1122 gen_rtx_PLUS (Pmode,
1123 gen_rtx_MULT (Pmode,
1128 else if ((mode == DFmode || mode == SFmode)
1129 && GET_CODE (XEXP (y, 0)) == SYMBOL_REF
1130 && GET_CODE (XEXP (x, 0)) == MULT
1131 && GET_CODE (XEXP (y, 1)) == CONST_INT
1132 && INTVAL (XEXP (y, 1)) % INTVAL (XEXP (XEXP (x, 0), 1)) == 0
1133 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
1134 && shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1))))
1137 = force_reg (Pmode, GEN_INT (INTVAL (XEXP (y, 1))
1138 / INTVAL (XEXP (XEXP (x, 0), 1))));
1139 regx2 = XEXP (XEXP (x, 0), 0);
1140 if (GET_CODE (regx2) != REG)
1141 regx2 = force_reg (Pmode, force_operand (regx2, 0));
1142 regx2 = force_reg (Pmode, gen_rtx_fmt_ee (GET_CODE (y), Pmode,
1146 gen_rtx_PLUS (Pmode,
1147 gen_rtx_MULT (Pmode, regx2,
1148 XEXP (XEXP (x, 0), 1)),
1149 force_reg (Pmode, XEXP (y, 0))));
1151 else if (GET_CODE (XEXP (y, 1)) == CONST_INT
1152 && INTVAL (XEXP (y, 1)) >= -4096
1153 && INTVAL (XEXP (y, 1)) <= 4095)
1155 /* This is safe because of the guard page at the
1156 beginning and end of the data space. Just
1157 return the original address. */
1162 /* Doesn't look like one we can optimize. */
1163 regx1 = force_reg (Pmode, force_operand (XEXP (x, 0), 0));
1164 regy1 = force_reg (Pmode, force_operand (XEXP (y, 0), 0));
1165 regy2 = force_reg (Pmode, force_operand (XEXP (y, 1), 0));
1166 regx1 = force_reg (Pmode,
1167 gen_rtx_fmt_ee (GET_CODE (y), Pmode,
1169 return force_reg (Pmode, gen_rtx_PLUS (Pmode, regx1, regy1));
1177 /* For the HPPA, REG and REG+CONST is cost 0
1178 and addresses involving symbolic constants are cost 2.
1180 PIC addresses are very expensive.
1182 It is no coincidence that this has the same structure
1183 as GO_IF_LEGITIMATE_ADDRESS. */
1186 hppa_address_cost (rtx X)
1188 switch (GET_CODE (X))
1201 /* Compute a (partial) cost for rtx X. Return true if the complete
1202 cost has been computed, and false if subexpressions should be
1203 scanned. In either case, *TOTAL contains the cost result. */
1206 hppa_rtx_costs (rtx x, int code, int outer_code, int *total)
1211 if (INTVAL (x) == 0)
1213 else if (INT_14_BITS (x))
1230 if ((x == CONST0_RTX (DFmode) || x == CONST0_RTX (SFmode))
1231 && outer_code != SET)
1238 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
1239 *total = COSTS_N_INSNS (3);
1240 else if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT)
1241 *total = COSTS_N_INSNS (8);
1243 *total = COSTS_N_INSNS (20);
1247 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
1249 *total = COSTS_N_INSNS (14);
1257 *total = COSTS_N_INSNS (60);
1260 case PLUS: /* this includes shNadd insns */
1262 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
1263 *total = COSTS_N_INSNS (3);
1265 *total = COSTS_N_INSNS (1);
1271 *total = COSTS_N_INSNS (1);
1279 /* Ensure mode of ORIG, a REG rtx, is MODE. Returns either ORIG or a
1280 new rtx with the correct mode. */
1282 force_mode (enum machine_mode mode, rtx orig)
1284 if (mode == GET_MODE (orig))
1287 gcc_assert (REGNO (orig) < FIRST_PSEUDO_REGISTER);
1289 return gen_rtx_REG (mode, REGNO (orig));
1292 /* Return 1 if *X is a thread-local symbol. */
1295 pa_tls_symbol_ref_1 (rtx *x, void *data ATTRIBUTE_UNUSED)
1297 return PA_SYMBOL_REF_TLS_P (*x);
1300 /* Return 1 if X contains a thread-local symbol. */
1303 pa_tls_referenced_p (rtx x)
1305 if (!TARGET_HAVE_TLS)
1308 return for_each_rtx (&x, &pa_tls_symbol_ref_1, 0);
1311 /* Emit insns to move operands[1] into operands[0].
1313 Return 1 if we have written out everything that needs to be done to
1314 do the move. Otherwise, return 0 and the caller will emit the move
1317 Note SCRATCH_REG may not be in the proper mode depending on how it
1318 will be used. This routine is responsible for creating a new copy
1319 of SCRATCH_REG in the proper mode. */
1322 emit_move_sequence (rtx *operands, enum machine_mode mode, rtx scratch_reg)
1324 register rtx operand0 = operands[0];
1325 register rtx operand1 = operands[1];
1328 /* We can only handle indexed addresses in the destination operand
1329 of floating point stores. Thus, we need to break out indexed
1330 addresses from the destination operand. */
1331 if (GET_CODE (operand0) == MEM && IS_INDEX_ADDR_P (XEXP (operand0, 0)))
1333 /* This is only safe up to the beginning of life analysis. */
1334 gcc_assert (!no_new_pseudos);
1336 tem = copy_to_mode_reg (Pmode, XEXP (operand0, 0));
1337 operand0 = replace_equiv_address (operand0, tem);
1340 /* On targets with non-equivalent space registers, break out unscaled
1341 indexed addresses from the source operand before the final CSE.
1342 We have to do this because the REG_POINTER flag is not correctly
1343 carried through various optimization passes and CSE may substitute
1344 a pseudo without the pointer set for one with the pointer set. As
1345 a result, we loose various opportunities to create insns with
1346 unscaled indexed addresses. */
1347 if (!TARGET_NO_SPACE_REGS
1348 && !cse_not_expected
1349 && GET_CODE (operand1) == MEM
1350 && GET_CODE (XEXP (operand1, 0)) == PLUS
1351 && REG_P (XEXP (XEXP (operand1, 0), 0))
1352 && REG_P (XEXP (XEXP (operand1, 0), 1)))
1354 = replace_equiv_address (operand1,
1355 copy_to_mode_reg (Pmode, XEXP (operand1, 0)));
1358 && reload_in_progress && GET_CODE (operand0) == REG
1359 && REGNO (operand0) >= FIRST_PSEUDO_REGISTER)
1360 operand0 = reg_equiv_mem[REGNO (operand0)];
1361 else if (scratch_reg
1362 && reload_in_progress && GET_CODE (operand0) == SUBREG
1363 && GET_CODE (SUBREG_REG (operand0)) == REG
1364 && REGNO (SUBREG_REG (operand0)) >= FIRST_PSEUDO_REGISTER)
1366 /* We must not alter SUBREG_BYTE (operand0) since that would confuse
1367 the code which tracks sets/uses for delete_output_reload. */
1368 rtx temp = gen_rtx_SUBREG (GET_MODE (operand0),
1369 reg_equiv_mem [REGNO (SUBREG_REG (operand0))],
1370 SUBREG_BYTE (operand0));
1371 operand0 = alter_subreg (&temp);
1375 && reload_in_progress && GET_CODE (operand1) == REG
1376 && REGNO (operand1) >= FIRST_PSEUDO_REGISTER)
1377 operand1 = reg_equiv_mem[REGNO (operand1)];
1378 else if (scratch_reg
1379 && reload_in_progress && GET_CODE (operand1) == SUBREG
1380 && GET_CODE (SUBREG_REG (operand1)) == REG
1381 && REGNO (SUBREG_REG (operand1)) >= FIRST_PSEUDO_REGISTER)
1383 /* We must not alter SUBREG_BYTE (operand0) since that would confuse
1384 the code which tracks sets/uses for delete_output_reload. */
1385 rtx temp = gen_rtx_SUBREG (GET_MODE (operand1),
1386 reg_equiv_mem [REGNO (SUBREG_REG (operand1))],
1387 SUBREG_BYTE (operand1));
1388 operand1 = alter_subreg (&temp);
1391 if (scratch_reg && reload_in_progress && GET_CODE (operand0) == MEM
1392 && ((tem = find_replacement (&XEXP (operand0, 0)))
1393 != XEXP (operand0, 0)))
1394 operand0 = replace_equiv_address (operand0, tem);
1396 if (scratch_reg && reload_in_progress && GET_CODE (operand1) == MEM
1397 && ((tem = find_replacement (&XEXP (operand1, 0)))
1398 != XEXP (operand1, 0)))
1399 operand1 = replace_equiv_address (operand1, tem);
1401 /* Handle secondary reloads for loads/stores of FP registers from
1402 REG+D addresses where D does not fit in 5 or 14 bits, including
1403 (subreg (mem (addr))) cases. */
1405 && fp_reg_operand (operand0, mode)
1406 && ((GET_CODE (operand1) == MEM
1407 && !memory_address_p ((GET_MODE_SIZE (mode) == 4 ? SFmode : DFmode),
1408 XEXP (operand1, 0)))
1409 || ((GET_CODE (operand1) == SUBREG
1410 && GET_CODE (XEXP (operand1, 0)) == MEM
1411 && !memory_address_p ((GET_MODE_SIZE (mode) == 4
1413 XEXP (XEXP (operand1, 0), 0))))))
1415 if (GET_CODE (operand1) == SUBREG)
1416 operand1 = XEXP (operand1, 0);
1418 /* SCRATCH_REG will hold an address and maybe the actual data. We want
1419 it in WORD_MODE regardless of what mode it was originally given
1421 scratch_reg = force_mode (word_mode, scratch_reg);
1423 /* D might not fit in 14 bits either; for such cases load D into
1425 if (!memory_address_p (Pmode, XEXP (operand1, 0)))
1427 emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1));
1428 emit_move_insn (scratch_reg,
1429 gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, 0)),
1431 XEXP (XEXP (operand1, 0), 0),
1435 emit_move_insn (scratch_reg, XEXP (operand1, 0));
1436 emit_insn (gen_rtx_SET (VOIDmode, operand0,
1437 replace_equiv_address (operand1, scratch_reg)));
1440 else if (scratch_reg
1441 && fp_reg_operand (operand1, mode)
1442 && ((GET_CODE (operand0) == MEM
1443 && !memory_address_p ((GET_MODE_SIZE (mode) == 4
1445 XEXP (operand0, 0)))
1446 || ((GET_CODE (operand0) == SUBREG)
1447 && GET_CODE (XEXP (operand0, 0)) == MEM
1448 && !memory_address_p ((GET_MODE_SIZE (mode) == 4
1450 XEXP (XEXP (operand0, 0), 0)))))
1452 if (GET_CODE (operand0) == SUBREG)
1453 operand0 = XEXP (operand0, 0);
1455 /* SCRATCH_REG will hold an address and maybe the actual data. We want
1456 it in WORD_MODE regardless of what mode it was originally given
1458 scratch_reg = force_mode (word_mode, scratch_reg);
1460 /* D might not fit in 14 bits either; for such cases load D into
1462 if (!memory_address_p (Pmode, XEXP (operand0, 0)))
1464 emit_move_insn (scratch_reg, XEXP (XEXP (operand0, 0), 1));
1465 emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand0,
1468 XEXP (XEXP (operand0, 0),
1473 emit_move_insn (scratch_reg, XEXP (operand0, 0));
1474 emit_insn (gen_rtx_SET (VOIDmode,
1475 replace_equiv_address (operand0, scratch_reg),
1479 /* Handle secondary reloads for loads of FP registers from constant
1480 expressions by forcing the constant into memory.
1482 Use scratch_reg to hold the address of the memory location.
1484 The proper fix is to change PREFERRED_RELOAD_CLASS to return
1485 NO_REGS when presented with a const_int and a register class
1486 containing only FP registers. Doing so unfortunately creates
1487 more problems than it solves. Fix this for 2.5. */
1488 else if (scratch_reg
1489 && CONSTANT_P (operand1)
1490 && fp_reg_operand (operand0, mode))
1492 rtx const_mem, xoperands[2];
1494 /* SCRATCH_REG will hold an address and maybe the actual data. We want
1495 it in WORD_MODE regardless of what mode it was originally given
1497 scratch_reg = force_mode (word_mode, scratch_reg);
1499 /* Force the constant into memory and put the address of the
1500 memory location into scratch_reg. */
1501 const_mem = force_const_mem (mode, operand1);
1502 xoperands[0] = scratch_reg;
1503 xoperands[1] = XEXP (const_mem, 0);
1504 emit_move_sequence (xoperands, Pmode, 0);
1506 /* Now load the destination register. */
1507 emit_insn (gen_rtx_SET (mode, operand0,
1508 replace_equiv_address (const_mem, scratch_reg)));
1511 /* Handle secondary reloads for SAR. These occur when trying to load
1512 the SAR from memory, FP register, or with a constant. */
1513 else if (scratch_reg
1514 && GET_CODE (operand0) == REG
1515 && REGNO (operand0) < FIRST_PSEUDO_REGISTER
1516 && REGNO_REG_CLASS (REGNO (operand0)) == SHIFT_REGS
1517 && (GET_CODE (operand1) == MEM
1518 || GET_CODE (operand1) == CONST_INT
1519 || (GET_CODE (operand1) == REG
1520 && FP_REG_CLASS_P (REGNO_REG_CLASS (REGNO (operand1))))))
1522 /* D might not fit in 14 bits either; for such cases load D into
1524 if (GET_CODE (operand1) == MEM
1525 && !memory_address_p (Pmode, XEXP (operand1, 0)))
1527 /* We are reloading the address into the scratch register, so we
1528 want to make sure the scratch register is a full register. */
1529 scratch_reg = force_mode (word_mode, scratch_reg);
1531 emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1));
1532 emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1,
1535 XEXP (XEXP (operand1, 0),
1539 /* Now we are going to load the scratch register from memory,
1540 we want to load it in the same width as the original MEM,
1541 which must be the same as the width of the ultimate destination,
1543 scratch_reg = force_mode (GET_MODE (operand0), scratch_reg);
1545 emit_move_insn (scratch_reg,
1546 replace_equiv_address (operand1, scratch_reg));
1550 /* We want to load the scratch register using the same mode as
1551 the ultimate destination. */
1552 scratch_reg = force_mode (GET_MODE (operand0), scratch_reg);
1554 emit_move_insn (scratch_reg, operand1);
1557 /* And emit the insn to set the ultimate destination. We know that
1558 the scratch register has the same mode as the destination at this
1560 emit_move_insn (operand0, scratch_reg);
1563 /* Handle the most common case: storing into a register. */
1564 else if (register_operand (operand0, mode))
1566 if (register_operand (operand1, mode)
1567 || (GET_CODE (operand1) == CONST_INT
1568 && cint_ok_for_move (INTVAL (operand1)))
1569 || (operand1 == CONST0_RTX (mode))
1570 || (GET_CODE (operand1) == HIGH
1571 && !symbolic_operand (XEXP (operand1, 0), VOIDmode))
1572 /* Only `general_operands' can come here, so MEM is ok. */
1573 || GET_CODE (operand1) == MEM)
1575 /* Various sets are created during RTL generation which don't
1576 have the REG_POINTER flag correctly set. After the CSE pass,
1577 instruction recognition can fail if we don't consistently
1578 set this flag when performing register copies. This should
1579 also improve the opportunities for creating insns that use
1580 unscaled indexing. */
1581 if (REG_P (operand0) && REG_P (operand1))
1583 if (REG_POINTER (operand1)
1584 && !REG_POINTER (operand0)
1585 && !HARD_REGISTER_P (operand0))
1586 copy_reg_pointer (operand0, operand1);
1587 else if (REG_POINTER (operand0)
1588 && !REG_POINTER (operand1)
1589 && !HARD_REGISTER_P (operand1))
1590 copy_reg_pointer (operand1, operand0);
1593 /* When MEMs are broken out, the REG_POINTER flag doesn't
1594 get set. In some cases, we can set the REG_POINTER flag
1595 from the declaration for the MEM. */
1596 if (REG_P (operand0)
1597 && GET_CODE (operand1) == MEM
1598 && !REG_POINTER (operand0))
1600 tree decl = MEM_EXPR (operand1);
1602 /* Set the register pointer flag and register alignment
1603 if the declaration for this memory reference is a
1604 pointer type. Fortran indirect argument references
1607 && !(flag_argument_noalias > 1
1608 && TREE_CODE (decl) == INDIRECT_REF
1609 && TREE_CODE (TREE_OPERAND (decl, 0)) == PARM_DECL))
1613 /* If this is a COMPONENT_REF, use the FIELD_DECL from
1615 if (TREE_CODE (decl) == COMPONENT_REF)
1616 decl = TREE_OPERAND (decl, 1);
1618 type = TREE_TYPE (decl);
1619 if (TREE_CODE (type) == ARRAY_TYPE)
1620 type = get_inner_array_type (type);
1622 if (POINTER_TYPE_P (type))
1626 type = TREE_TYPE (type);
1627 /* Using TYPE_ALIGN_OK is rather conservative as
1628 only the ada frontend actually sets it. */
1629 align = (TYPE_ALIGN_OK (type) ? TYPE_ALIGN (type)
1631 mark_reg_pointer (operand0, align);
1636 emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1));
1640 else if (GET_CODE (operand0) == MEM)
1642 if (mode == DFmode && operand1 == CONST0_RTX (mode)
1643 && !(reload_in_progress || reload_completed))
1645 rtx temp = gen_reg_rtx (DFmode);
1647 emit_insn (gen_rtx_SET (VOIDmode, temp, operand1));
1648 emit_insn (gen_rtx_SET (VOIDmode, operand0, temp));
1651 if (register_operand (operand1, mode) || operand1 == CONST0_RTX (mode))
1653 /* Run this case quickly. */
1654 emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1));
1657 if (! (reload_in_progress || reload_completed))
1659 operands[0] = validize_mem (operand0);
1660 operands[1] = operand1 = force_reg (mode, operand1);
1664 /* Simplify the source if we need to.
1665 Note we do have to handle function labels here, even though we do
1666 not consider them legitimate constants. Loop optimizations can
1667 call the emit_move_xxx with one as a source. */
1668 if ((GET_CODE (operand1) != HIGH && immediate_operand (operand1, mode))
1669 || function_label_operand (operand1, mode)
1670 || (GET_CODE (operand1) == HIGH
1671 && symbolic_operand (XEXP (operand1, 0), mode)))
1675 if (GET_CODE (operand1) == HIGH)
1678 operand1 = XEXP (operand1, 0);
1680 if (symbolic_operand (operand1, mode))
1682 /* Argh. The assembler and linker can't handle arithmetic
1685 So we force the plabel into memory, load operand0 from
1686 the memory location, then add in the constant part. */
1687 if ((GET_CODE (operand1) == CONST
1688 && GET_CODE (XEXP (operand1, 0)) == PLUS
1689 && function_label_operand (XEXP (XEXP (operand1, 0), 0), Pmode))
1690 || function_label_operand (operand1, mode))
1692 rtx temp, const_part;
1694 /* Figure out what (if any) scratch register to use. */
1695 if (reload_in_progress || reload_completed)
1697 scratch_reg = scratch_reg ? scratch_reg : operand0;
1698 /* SCRATCH_REG will hold an address and maybe the actual
1699 data. We want it in WORD_MODE regardless of what mode it
1700 was originally given to us. */
1701 scratch_reg = force_mode (word_mode, scratch_reg);
1704 scratch_reg = gen_reg_rtx (Pmode);
1706 if (GET_CODE (operand1) == CONST)
1708 /* Save away the constant part of the expression. */
1709 const_part = XEXP (XEXP (operand1, 0), 1);
1710 gcc_assert (GET_CODE (const_part) == CONST_INT);
1712 /* Force the function label into memory. */
1713 temp = force_const_mem (mode, XEXP (XEXP (operand1, 0), 0));
1717 /* No constant part. */
1718 const_part = NULL_RTX;
1720 /* Force the function label into memory. */
1721 temp = force_const_mem (mode, operand1);
1725 /* Get the address of the memory location. PIC-ify it if
1727 temp = XEXP (temp, 0);
1729 temp = legitimize_pic_address (temp, mode, scratch_reg);
1731 /* Put the address of the memory location into our destination
1734 emit_move_sequence (operands, mode, scratch_reg);
1736 /* Now load from the memory location into our destination
1738 operands[1] = gen_rtx_MEM (Pmode, operands[0]);
1739 emit_move_sequence (operands, mode, scratch_reg);
1741 /* And add back in the constant part. */
1742 if (const_part != NULL_RTX)
1743 expand_inc (operand0, const_part);
1752 if (reload_in_progress || reload_completed)
1754 temp = scratch_reg ? scratch_reg : operand0;
1755 /* TEMP will hold an address and maybe the actual
1756 data. We want it in WORD_MODE regardless of what mode it
1757 was originally given to us. */
1758 temp = force_mode (word_mode, temp);
1761 temp = gen_reg_rtx (Pmode);
1763 /* (const (plus (symbol) (const_int))) must be forced to
1764 memory during/after reload if the const_int will not fit
1766 if (GET_CODE (operand1) == CONST
1767 && GET_CODE (XEXP (operand1, 0)) == PLUS
1768 && GET_CODE (XEXP (XEXP (operand1, 0), 1)) == CONST_INT
1769 && !INT_14_BITS (XEXP (XEXP (operand1, 0), 1))
1770 && (reload_completed || reload_in_progress)
1773 rtx const_mem = force_const_mem (mode, operand1);
1774 operands[1] = legitimize_pic_address (XEXP (const_mem, 0),
1776 operands[1] = replace_equiv_address (const_mem, operands[1]);
1777 emit_move_sequence (operands, mode, temp);
1781 operands[1] = legitimize_pic_address (operand1, mode, temp);
1782 if (REG_P (operand0) && REG_P (operands[1]))
1783 copy_reg_pointer (operand0, operands[1]);
1784 emit_insn (gen_rtx_SET (VOIDmode, operand0, operands[1]));
1787 /* On the HPPA, references to data space are supposed to use dp,
1788 register 27, but showing it in the RTL inhibits various cse
1789 and loop optimizations. */
1794 if (reload_in_progress || reload_completed)
1796 temp = scratch_reg ? scratch_reg : operand0;
1797 /* TEMP will hold an address and maybe the actual
1798 data. We want it in WORD_MODE regardless of what mode it
1799 was originally given to us. */
1800 temp = force_mode (word_mode, temp);
1803 temp = gen_reg_rtx (mode);
1805 /* Loading a SYMBOL_REF into a register makes that register
1806 safe to be used as the base in an indexed address.
1808 Don't mark hard registers though. That loses. */
1809 if (GET_CODE (operand0) == REG
1810 && REGNO (operand0) >= FIRST_PSEUDO_REGISTER)
1811 mark_reg_pointer (operand0, BITS_PER_UNIT);
1812 if (REGNO (temp) >= FIRST_PSEUDO_REGISTER)
1813 mark_reg_pointer (temp, BITS_PER_UNIT);
1816 set = gen_rtx_SET (mode, operand0, temp);
1818 set = gen_rtx_SET (VOIDmode,
1820 gen_rtx_LO_SUM (mode, temp, operand1));
1822 emit_insn (gen_rtx_SET (VOIDmode,
1824 gen_rtx_HIGH (mode, operand1)));
1830 else if (pa_tls_referenced_p (operand1))
1835 if (GET_CODE (tmp) == CONST && GET_CODE (XEXP (tmp, 0)) == PLUS)
1837 addend = XEXP (XEXP (tmp, 0), 1);
1838 tmp = XEXP (XEXP (tmp, 0), 0);
1841 gcc_assert (GET_CODE (tmp) == SYMBOL_REF);
1842 tmp = legitimize_tls_address (tmp);
1845 tmp = gen_rtx_PLUS (mode, tmp, addend);
1846 tmp = force_operand (tmp, operands[0]);
1850 else if (GET_CODE (operand1) != CONST_INT
1851 || !cint_ok_for_move (INTVAL (operand1)))
1855 HOST_WIDE_INT value = 0;
1856 HOST_WIDE_INT insv = 0;
1859 if (GET_CODE (operand1) == CONST_INT)
1860 value = INTVAL (operand1);
1863 && GET_CODE (operand1) == CONST_INT
1864 && HOST_BITS_PER_WIDE_INT > 32
1865 && GET_MODE_BITSIZE (GET_MODE (operand0)) > 32)
1869 /* Extract the low order 32 bits of the value and sign extend.
1870 If the new value is the same as the original value, we can
1871 can use the original value as-is. If the new value is
1872 different, we use it and insert the most-significant 32-bits
1873 of the original value into the final result. */
1874 nval = ((value & (((HOST_WIDE_INT) 2 << 31) - 1))
1875 ^ ((HOST_WIDE_INT) 1 << 31)) - ((HOST_WIDE_INT) 1 << 31);
1878 #if HOST_BITS_PER_WIDE_INT > 32
1879 insv = value >= 0 ? value >> 32 : ~(~value >> 32);
1883 operand1 = GEN_INT (nval);
1887 if (reload_in_progress || reload_completed)
1888 temp = scratch_reg ? scratch_reg : operand0;
1890 temp = gen_reg_rtx (mode);
1892 /* We don't directly split DImode constants on 32-bit targets
1893 because PLUS uses an 11-bit immediate and the insn sequence
1894 generated is not as efficient as the one using HIGH/LO_SUM. */
1895 if (GET_CODE (operand1) == CONST_INT
1896 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
1899 /* Directly break constant into high and low parts. This
1900 provides better optimization opportunities because various
1901 passes recognize constants split with PLUS but not LO_SUM.
1902 We use a 14-bit signed low part except when the addition
1903 of 0x4000 to the high part might change the sign of the
1905 HOST_WIDE_INT low = value & 0x3fff;
1906 HOST_WIDE_INT high = value & ~ 0x3fff;
1910 if (high == 0x7fffc000 || (mode == HImode && high == 0x4000))
1918 emit_insn (gen_rtx_SET (VOIDmode, temp, GEN_INT (high)));
1919 operands[1] = gen_rtx_PLUS (mode, temp, GEN_INT (low));
1923 emit_insn (gen_rtx_SET (VOIDmode, temp,
1924 gen_rtx_HIGH (mode, operand1)));
1925 operands[1] = gen_rtx_LO_SUM (mode, temp, operand1);
1928 insn = emit_move_insn (operands[0], operands[1]);
1930 /* Now insert the most significant 32 bits of the value
1931 into the register. When we don't have a second register
1932 available, it could take up to nine instructions to load
1933 a 64-bit integer constant. Prior to reload, we force
1934 constants that would take more than three instructions
1935 to load to the constant pool. During and after reload,
1936 we have to handle all possible values. */
1939 /* Use a HIGH/LO_SUM/INSV sequence if we have a second
1940 register and the value to be inserted is outside the
1941 range that can be loaded with three depdi instructions. */
1942 if (temp != operand0 && (insv >= 16384 || insv < -16384))
1944 operand1 = GEN_INT (insv);
1946 emit_insn (gen_rtx_SET (VOIDmode, temp,
1947 gen_rtx_HIGH (mode, operand1)));
1948 emit_move_insn (temp, gen_rtx_LO_SUM (mode, temp, operand1));
1949 emit_insn (gen_insv (operand0, GEN_INT (32),
1954 int len = 5, pos = 27;
1956 /* Insert the bits using the depdi instruction. */
1959 HOST_WIDE_INT v5 = ((insv & 31) ^ 16) - 16;
1960 HOST_WIDE_INT sign = v5 < 0;
1962 /* Left extend the insertion. */
1963 insv = (insv >= 0 ? insv >> len : ~(~insv >> len));
1964 while (pos > 0 && (insv & 1) == sign)
1966 insv = (insv >= 0 ? insv >> 1 : ~(~insv >> 1));
1971 emit_insn (gen_insv (operand0, GEN_INT (len),
1972 GEN_INT (pos), GEN_INT (v5)));
1974 len = pos > 0 && pos < 5 ? pos : 5;
1981 = gen_rtx_EXPR_LIST (REG_EQUAL, op1, REG_NOTES (insn));
1986 /* Now have insn-emit do whatever it normally does. */
1990 /* Examine EXP and return nonzero if it contains an ADDR_EXPR (meaning
1991 it will need a link/runtime reloc). */
1994 reloc_needed (tree exp)
1998 switch (TREE_CODE (exp))
2005 reloc = reloc_needed (TREE_OPERAND (exp, 0));
2006 reloc |= reloc_needed (TREE_OPERAND (exp, 1));
2011 case NON_LVALUE_EXPR:
2012 reloc = reloc_needed (TREE_OPERAND (exp, 0));
2018 unsigned HOST_WIDE_INT ix;
2020 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), ix, value)
2022 reloc |= reloc_needed (value);
2035 /* Does operand (which is a symbolic_operand) live in text space?
2036 If so, SYMBOL_REF_FLAG, which is set by pa_encode_section_info,
2040 read_only_operand (rtx operand, enum machine_mode mode ATTRIBUTE_UNUSED)
2042 if (GET_CODE (operand) == CONST)
2043 operand = XEXP (XEXP (operand, 0), 0);
2046 if (GET_CODE (operand) == SYMBOL_REF)
2047 return SYMBOL_REF_FLAG (operand) && !CONSTANT_POOL_ADDRESS_P (operand);
2051 if (GET_CODE (operand) == SYMBOL_REF)
2052 return SYMBOL_REF_FLAG (operand) || CONSTANT_POOL_ADDRESS_P (operand);
2058 /* Return the best assembler insn template
2059 for moving operands[1] into operands[0] as a fullword. */
2061 singlemove_string (rtx *operands)
2063 HOST_WIDE_INT intval;
2065 if (GET_CODE (operands[0]) == MEM)
2066 return "stw %r1,%0";
2067 if (GET_CODE (operands[1]) == MEM)
2069 if (GET_CODE (operands[1]) == CONST_DOUBLE)
2074 gcc_assert (GET_MODE (operands[1]) == SFmode);
2076 /* Translate the CONST_DOUBLE to a CONST_INT with the same target
2078 REAL_VALUE_FROM_CONST_DOUBLE (d, operands[1]);
2079 REAL_VALUE_TO_TARGET_SINGLE (d, i);
2081 operands[1] = GEN_INT (i);
2082 /* Fall through to CONST_INT case. */
2084 if (GET_CODE (operands[1]) == CONST_INT)
2086 intval = INTVAL (operands[1]);
2088 if (VAL_14_BITS_P (intval))
2090 else if ((intval & 0x7ff) == 0)
2091 return "ldil L'%1,%0";
2092 else if (zdepi_cint_p (intval))
2093 return "{zdepi %Z1,%0|depwi,z %Z1,%0}";
2095 return "ldil L'%1,%0\n\tldo R'%1(%0),%0";
2097 return "copy %1,%0";
2101 /* Compute position (in OP[1]) and width (in OP[2])
2102 useful for copying IMM to a register using the zdepi
2103 instructions. Store the immediate value to insert in OP[0]. */
2105 compute_zdepwi_operands (unsigned HOST_WIDE_INT imm, unsigned *op)
2109 /* Find the least significant set bit in IMM. */
2110 for (lsb = 0; lsb < 32; lsb++)
2117 /* Choose variants based on *sign* of the 5-bit field. */
2118 if ((imm & 0x10) == 0)
2119 len = (lsb <= 28) ? 4 : 32 - lsb;
2122 /* Find the width of the bitstring in IMM. */
2123 for (len = 5; len < 32; len++)
2125 if ((imm & (1 << len)) == 0)
2129 /* Sign extend IMM as a 5-bit value. */
2130 imm = (imm & 0xf) - 0x10;
2138 /* Compute position (in OP[1]) and width (in OP[2])
2139 useful for copying IMM to a register using the depdi,z
2140 instructions. Store the immediate value to insert in OP[0]. */
2142 compute_zdepdi_operands (unsigned HOST_WIDE_INT imm, unsigned *op)
2144 HOST_WIDE_INT lsb, len;
2146 /* Find the least significant set bit in IMM. */
2147 for (lsb = 0; lsb < HOST_BITS_PER_WIDE_INT; lsb++)
2154 /* Choose variants based on *sign* of the 5-bit field. */
2155 if ((imm & 0x10) == 0)
2156 len = ((lsb <= HOST_BITS_PER_WIDE_INT - 4)
2157 ? 4 : HOST_BITS_PER_WIDE_INT - lsb);
2160 /* Find the width of the bitstring in IMM. */
2161 for (len = 5; len < HOST_BITS_PER_WIDE_INT; len++)
2163 if ((imm & ((unsigned HOST_WIDE_INT) 1 << len)) == 0)
2167 /* Sign extend IMM as a 5-bit value. */
2168 imm = (imm & 0xf) - 0x10;
2176 /* Output assembler code to perform a doubleword move insn
2177 with operands OPERANDS. */
2180 output_move_double (rtx *operands)
2182 enum { REGOP, OFFSOP, MEMOP, CNSTOP, RNDOP } optype0, optype1;
2184 rtx addreg0 = 0, addreg1 = 0;
2186 /* First classify both operands. */
2188 if (REG_P (operands[0]))
2190 else if (offsettable_memref_p (operands[0]))
2192 else if (GET_CODE (operands[0]) == MEM)
2197 if (REG_P (operands[1]))
2199 else if (CONSTANT_P (operands[1]))
2201 else if (offsettable_memref_p (operands[1]))
2203 else if (GET_CODE (operands[1]) == MEM)
2208 /* Check for the cases that the operand constraints are not
2209 supposed to allow to happen. */
2210 gcc_assert (optype0 == REGOP || optype1 == REGOP);
2212 /* Handle auto decrementing and incrementing loads and stores
2213 specifically, since the structure of the function doesn't work
2214 for them without major modification. Do it better when we learn
2215 this port about the general inc/dec addressing of PA.
2216 (This was written by tege. Chide him if it doesn't work.) */
2218 if (optype0 == MEMOP)
2220 /* We have to output the address syntax ourselves, since print_operand
2221 doesn't deal with the addresses we want to use. Fix this later. */
2223 rtx addr = XEXP (operands[0], 0);
2224 if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC)
2226 rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0);
2228 operands[0] = XEXP (addr, 0);
2229 gcc_assert (GET_CODE (operands[1]) == REG
2230 && GET_CODE (operands[0]) == REG);
2232 gcc_assert (!reg_overlap_mentioned_p (high_reg, addr));
2234 /* No overlap between high target register and address
2235 register. (We do this in a non-obvious way to
2236 save a register file writeback) */
2237 if (GET_CODE (addr) == POST_INC)
2238 return "{stws|stw},ma %1,8(%0)\n\tstw %R1,-4(%0)";
2239 return "{stws|stw},ma %1,-8(%0)\n\tstw %R1,12(%0)";
2241 else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC)
2243 rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0);
2245 operands[0] = XEXP (addr, 0);
2246 gcc_assert (GET_CODE (operands[1]) == REG
2247 && GET_CODE (operands[0]) == REG);
2249 gcc_assert (!reg_overlap_mentioned_p (high_reg, addr));
2250 /* No overlap between high target register and address
2251 register. (We do this in a non-obvious way to save a
2252 register file writeback) */
2253 if (GET_CODE (addr) == PRE_INC)
2254 return "{stws|stw},mb %1,8(%0)\n\tstw %R1,4(%0)";
2255 return "{stws|stw},mb %1,-8(%0)\n\tstw %R1,4(%0)";
2258 if (optype1 == MEMOP)
2260 /* We have to output the address syntax ourselves, since print_operand
2261 doesn't deal with the addresses we want to use. Fix this later. */
2263 rtx addr = XEXP (operands[1], 0);
2264 if (GET_CODE (addr) == POST_INC || GET_CODE (addr) == POST_DEC)
2266 rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0);
2268 operands[1] = XEXP (addr, 0);
2269 gcc_assert (GET_CODE (operands[0]) == REG
2270 && GET_CODE (operands[1]) == REG);
2272 if (!reg_overlap_mentioned_p (high_reg, addr))
2274 /* No overlap between high target register and address
2275 register. (We do this in a non-obvious way to
2276 save a register file writeback) */
2277 if (GET_CODE (addr) == POST_INC)
2278 return "{ldws|ldw},ma 8(%1),%0\n\tldw -4(%1),%R0";
2279 return "{ldws|ldw},ma -8(%1),%0\n\tldw 12(%1),%R0";
2283 /* This is an undefined situation. We should load into the
2284 address register *and* update that register. Probably
2285 we don't need to handle this at all. */
2286 if (GET_CODE (addr) == POST_INC)
2287 return "ldw 4(%1),%R0\n\t{ldws|ldw},ma 8(%1),%0";
2288 return "ldw 4(%1),%R0\n\t{ldws|ldw},ma -8(%1),%0";
2291 else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC)
2293 rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0);
2295 operands[1] = XEXP (addr, 0);
2296 gcc_assert (GET_CODE (operands[0]) == REG
2297 && GET_CODE (operands[1]) == REG);
2299 if (!reg_overlap_mentioned_p (high_reg, addr))
2301 /* No overlap between high target register and address
2302 register. (We do this in a non-obvious way to
2303 save a register file writeback) */
2304 if (GET_CODE (addr) == PRE_INC)
2305 return "{ldws|ldw},mb 8(%1),%0\n\tldw 4(%1),%R0";
2306 return "{ldws|ldw},mb -8(%1),%0\n\tldw 4(%1),%R0";
2310 /* This is an undefined situation. We should load into the
2311 address register *and* update that register. Probably
2312 we don't need to handle this at all. */
2313 if (GET_CODE (addr) == PRE_INC)
2314 return "ldw 12(%1),%R0\n\t{ldws|ldw},mb 8(%1),%0";
2315 return "ldw -4(%1),%R0\n\t{ldws|ldw},mb -8(%1),%0";
2318 else if (GET_CODE (addr) == PLUS
2319 && GET_CODE (XEXP (addr, 0)) == MULT)
2321 rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0);
2323 if (!reg_overlap_mentioned_p (high_reg, addr))
2327 xoperands[0] = high_reg;
2328 xoperands[1] = XEXP (addr, 1);
2329 xoperands[2] = XEXP (XEXP (addr, 0), 0);
2330 xoperands[3] = XEXP (XEXP (addr, 0), 1);
2331 output_asm_insn ("{sh%O3addl %2,%1,%0|shladd,l %2,%O3,%1,%0}",
2333 return "ldw 4(%0),%R0\n\tldw 0(%0),%0";
2339 xoperands[0] = high_reg;
2340 xoperands[1] = XEXP (addr, 1);
2341 xoperands[2] = XEXP (XEXP (addr, 0), 0);
2342 xoperands[3] = XEXP (XEXP (addr, 0), 1);
2343 output_asm_insn ("{sh%O3addl %2,%1,%R0|shladd,l %2,%O3,%1,%R0}",
2345 return "ldw 0(%R0),%0\n\tldw 4(%R0),%R0";
2350 /* If an operand is an unoffsettable memory ref, find a register
2351 we can increment temporarily to make it refer to the second word. */
2353 if (optype0 == MEMOP)
2354 addreg0 = find_addr_reg (XEXP (operands[0], 0));
2356 if (optype1 == MEMOP)
2357 addreg1 = find_addr_reg (XEXP (operands[1], 0));
2359 /* Ok, we can do one word at a time.
2360 Normally we do the low-numbered word first.
2362 In either case, set up in LATEHALF the operands to use
2363 for the high-numbered word and in some cases alter the
2364 operands in OPERANDS to be suitable for the low-numbered word. */
2366 if (optype0 == REGOP)
2367 latehalf[0] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1);
2368 else if (optype0 == OFFSOP)
2369 latehalf[0] = adjust_address (operands[0], SImode, 4);
2371 latehalf[0] = operands[0];
2373 if (optype1 == REGOP)
2374 latehalf[1] = gen_rtx_REG (SImode, REGNO (operands[1]) + 1);
2375 else if (optype1 == OFFSOP)
2376 latehalf[1] = adjust_address (operands[1], SImode, 4);
2377 else if (optype1 == CNSTOP)
2378 split_double (operands[1], &operands[1], &latehalf[1]);
2380 latehalf[1] = operands[1];
2382 /* If the first move would clobber the source of the second one,
2383 do them in the other order.
2385 This can happen in two cases:
2387 mem -> register where the first half of the destination register
2388 is the same register used in the memory's address. Reload
2389 can create such insns.
2391 mem in this case will be either register indirect or register
2392 indirect plus a valid offset.
2394 register -> register move where REGNO(dst) == REGNO(src + 1)
2395 someone (Tim/Tege?) claimed this can happen for parameter loads.
2397 Handle mem -> register case first. */
2398 if (optype0 == REGOP
2399 && (optype1 == MEMOP || optype1 == OFFSOP)
2400 && refers_to_regno_p (REGNO (operands[0]), REGNO (operands[0]) + 1,
2403 /* Do the late half first. */
2405 output_asm_insn ("ldo 4(%0),%0", &addreg1);
2406 output_asm_insn (singlemove_string (latehalf), latehalf);
2410 output_asm_insn ("ldo -4(%0),%0", &addreg1);
2411 return singlemove_string (operands);
2414 /* Now handle register -> register case. */
2415 if (optype0 == REGOP && optype1 == REGOP
2416 && REGNO (operands[0]) == REGNO (operands[1]) + 1)
2418 output_asm_insn (singlemove_string (latehalf), latehalf);
2419 return singlemove_string (operands);
2422 /* Normal case: do the two words, low-numbered first. */
2424 output_asm_insn (singlemove_string (operands), operands);
2426 /* Make any unoffsettable addresses point at high-numbered word. */
2428 output_asm_insn ("ldo 4(%0),%0", &addreg0);
2430 output_asm_insn ("ldo 4(%0),%0", &addreg1);
2433 output_asm_insn (singlemove_string (latehalf), latehalf);
2435 /* Undo the adds we just did. */
2437 output_asm_insn ("ldo -4(%0),%0", &addreg0);
2439 output_asm_insn ("ldo -4(%0),%0", &addreg1);
2445 output_fp_move_double (rtx *operands)
2447 if (FP_REG_P (operands[0]))
2449 if (FP_REG_P (operands[1])
2450 || operands[1] == CONST0_RTX (GET_MODE (operands[0])))
2451 output_asm_insn ("fcpy,dbl %f1,%0", operands);
2453 output_asm_insn ("fldd%F1 %1,%0", operands);
2455 else if (FP_REG_P (operands[1]))
2457 output_asm_insn ("fstd%F0 %1,%0", operands);
2463 gcc_assert (operands[1] == CONST0_RTX (GET_MODE (operands[0])));
2465 /* This is a pain. You have to be prepared to deal with an
2466 arbitrary address here including pre/post increment/decrement.
2468 so avoid this in the MD. */
2469 gcc_assert (GET_CODE (operands[0]) == REG);
2471 xoperands[1] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1);
2472 xoperands[0] = operands[0];
2473 output_asm_insn ("copy %%r0,%0\n\tcopy %%r0,%1", xoperands);
2478 /* Return a REG that occurs in ADDR with coefficient 1.
2479 ADDR can be effectively incremented by incrementing REG. */
2482 find_addr_reg (rtx addr)
2484 while (GET_CODE (addr) == PLUS)
2486 if (GET_CODE (XEXP (addr, 0)) == REG)
2487 addr = XEXP (addr, 0);
2488 else if (GET_CODE (XEXP (addr, 1)) == REG)
2489 addr = XEXP (addr, 1);
2490 else if (CONSTANT_P (XEXP (addr, 0)))
2491 addr = XEXP (addr, 1);
2492 else if (CONSTANT_P (XEXP (addr, 1)))
2493 addr = XEXP (addr, 0);
2497 gcc_assert (GET_CODE (addr) == REG);
2501 /* Emit code to perform a block move.
2503 OPERANDS[0] is the destination pointer as a REG, clobbered.
2504 OPERANDS[1] is the source pointer as a REG, clobbered.
2505 OPERANDS[2] is a register for temporary storage.
2506 OPERANDS[3] is a register for temporary storage.
2507 OPERANDS[4] is the size as a CONST_INT
2508 OPERANDS[5] is the alignment safe to use, as a CONST_INT.
2509 OPERANDS[6] is another temporary register. */
2512 output_block_move (rtx *operands, int size_is_constant ATTRIBUTE_UNUSED)
2514 int align = INTVAL (operands[5]);
2515 unsigned long n_bytes = INTVAL (operands[4]);
2517 /* We can't move more than a word at a time because the PA
2518 has no longer integer move insns. (Could use fp mem ops?) */
2519 if (align > (TARGET_64BIT ? 8 : 4))
2520 align = (TARGET_64BIT ? 8 : 4);
2522 /* Note that we know each loop below will execute at least twice
2523 (else we would have open-coded the copy). */
2527 /* Pre-adjust the loop counter. */
2528 operands[4] = GEN_INT (n_bytes - 16);
2529 output_asm_insn ("ldi %4,%2", operands);
2532 output_asm_insn ("ldd,ma 8(%1),%3", operands);
2533 output_asm_insn ("ldd,ma 8(%1),%6", operands);
2534 output_asm_insn ("std,ma %3,8(%0)", operands);
2535 output_asm_insn ("addib,>= -16,%2,.-12", operands);
2536 output_asm_insn ("std,ma %6,8(%0)", operands);
2538 /* Handle the residual. There could be up to 7 bytes of
2539 residual to copy! */
2540 if (n_bytes % 16 != 0)
2542 operands[4] = GEN_INT (n_bytes % 8);
2543 if (n_bytes % 16 >= 8)
2544 output_asm_insn ("ldd,ma 8(%1),%3", operands);
2545 if (n_bytes % 8 != 0)
2546 output_asm_insn ("ldd 0(%1),%6", operands);
2547 if (n_bytes % 16 >= 8)
2548 output_asm_insn ("std,ma %3,8(%0)", operands);
2549 if (n_bytes % 8 != 0)
2550 output_asm_insn ("stdby,e %6,%4(%0)", operands);
2555 /* Pre-adjust the loop counter. */
2556 operands[4] = GEN_INT (n_bytes - 8);
2557 output_asm_insn ("ldi %4,%2", operands);
2560 output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands);
2561 output_asm_insn ("{ldws|ldw},ma 4(%1),%6", operands);
2562 output_asm_insn ("{stws|stw},ma %3,4(%0)", operands);
2563 output_asm_insn ("addib,>= -8,%2,.-12", operands);
2564 output_asm_insn ("{stws|stw},ma %6,4(%0)", operands);
2566 /* Handle the residual. There could be up to 7 bytes of
2567 residual to copy! */
2568 if (n_bytes % 8 != 0)
2570 operands[4] = GEN_INT (n_bytes % 4);
2571 if (n_bytes % 8 >= 4)
2572 output_asm_insn ("{ldws|ldw},ma 4(%1),%3", operands);
2573 if (n_bytes % 4 != 0)
2574 output_asm_insn ("ldw 0(%1),%6", operands);
2575 if (n_bytes % 8 >= 4)
2576 output_asm_insn ("{stws|stw},ma %3,4(%0)", operands);
2577 if (n_bytes % 4 != 0)
2578 output_asm_insn ("{stbys|stby},e %6,%4(%0)", operands);
2583 /* Pre-adjust the loop counter. */
2584 operands[4] = GEN_INT (n_bytes - 4);
2585 output_asm_insn ("ldi %4,%2", operands);
2588 output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands);
2589 output_asm_insn ("{ldhs|ldh},ma 2(%1),%6", operands);
2590 output_asm_insn ("{sths|sth},ma %3,2(%0)", operands);
2591 output_asm_insn ("addib,>= -4,%2,.-12", operands);
2592 output_asm_insn ("{sths|sth},ma %6,2(%0)", operands);
2594 /* Handle the residual. */
2595 if (n_bytes % 4 != 0)
2597 if (n_bytes % 4 >= 2)
2598 output_asm_insn ("{ldhs|ldh},ma 2(%1),%3", operands);
2599 if (n_bytes % 2 != 0)
2600 output_asm_insn ("ldb 0(%1),%6", operands);
2601 if (n_bytes % 4 >= 2)
2602 output_asm_insn ("{sths|sth},ma %3,2(%0)", operands);
2603 if (n_bytes % 2 != 0)
2604 output_asm_insn ("stb %6,0(%0)", operands);
2609 /* Pre-adjust the loop counter. */
2610 operands[4] = GEN_INT (n_bytes - 2);
2611 output_asm_insn ("ldi %4,%2", operands);
2614 output_asm_insn ("{ldbs|ldb},ma 1(%1),%3", operands);
2615 output_asm_insn ("{ldbs|ldb},ma 1(%1),%6", operands);
2616 output_asm_insn ("{stbs|stb},ma %3,1(%0)", operands);
2617 output_asm_insn ("addib,>= -2,%2,.-12", operands);
2618 output_asm_insn ("{stbs|stb},ma %6,1(%0)", operands);
2620 /* Handle the residual. */
2621 if (n_bytes % 2 != 0)
2623 output_asm_insn ("ldb 0(%1),%3", operands);
2624 output_asm_insn ("stb %3,0(%0)", operands);
2633 /* Count the number of insns necessary to handle this block move.
2635 Basic structure is the same as emit_block_move, except that we
2636 count insns rather than emit them. */
2639 compute_movmem_length (rtx insn)
2641 rtx pat = PATTERN (insn);
2642 unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 7), 0));
2643 unsigned long n_bytes = INTVAL (XEXP (XVECEXP (pat, 0, 6), 0));
2644 unsigned int n_insns = 0;
2646 /* We can't move more than four bytes at a time because the PA
2647 has no longer integer move insns. (Could use fp mem ops?) */
2648 if (align > (TARGET_64BIT ? 8 : 4))
2649 align = (TARGET_64BIT ? 8 : 4);
2651 /* The basic copying loop. */
2655 if (n_bytes % (2 * align) != 0)
2657 if ((n_bytes % (2 * align)) >= align)
2660 if ((n_bytes % align) != 0)
2664 /* Lengths are expressed in bytes now; each insn is 4 bytes. */
2668 /* Emit code to perform a block clear.
2670 OPERANDS[0] is the destination pointer as a REG, clobbered.
2671 OPERANDS[1] is a register for temporary storage.
2672 OPERANDS[2] is the size as a CONST_INT
2673 OPERANDS[3] is the alignment safe to use, as a CONST_INT. */
2676 output_block_clear (rtx *operands, int size_is_constant ATTRIBUTE_UNUSED)
2678 int align = INTVAL (operands[3]);
2679 unsigned long n_bytes = INTVAL (operands[2]);
2681 /* We can't clear more than a word at a time because the PA
2682 has no longer integer move insns. */
2683 if (align > (TARGET_64BIT ? 8 : 4))
2684 align = (TARGET_64BIT ? 8 : 4);
2686 /* Note that we know each loop below will execute at least twice
2687 (else we would have open-coded the copy). */
2691 /* Pre-adjust the loop counter. */
2692 operands[2] = GEN_INT (n_bytes - 16);
2693 output_asm_insn ("ldi %2,%1", operands);
2696 output_asm_insn ("std,ma %%r0,8(%0)", operands);
2697 output_asm_insn ("addib,>= -16,%1,.-4", operands);
2698 output_asm_insn ("std,ma %%r0,8(%0)", operands);
2700 /* Handle the residual. There could be up to 7 bytes of
2701 residual to copy! */
2702 if (n_bytes % 16 != 0)
2704 operands[2] = GEN_INT (n_bytes % 8);
2705 if (n_bytes % 16 >= 8)
2706 output_asm_insn ("std,ma %%r0,8(%0)", operands);
2707 if (n_bytes % 8 != 0)
2708 output_asm_insn ("stdby,e %%r0,%2(%0)", operands);
2713 /* Pre-adjust the loop counter. */
2714 operands[2] = GEN_INT (n_bytes - 8);
2715 output_asm_insn ("ldi %2,%1", operands);
2718 output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands);
2719 output_asm_insn ("addib,>= -8,%1,.-4", operands);
2720 output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands);
2722 /* Handle the residual. There could be up to 7 bytes of
2723 residual to copy! */
2724 if (n_bytes % 8 != 0)
2726 operands[2] = GEN_INT (n_bytes % 4);
2727 if (n_bytes % 8 >= 4)
2728 output_asm_insn ("{stws|stw},ma %%r0,4(%0)", operands);
2729 if (n_bytes % 4 != 0)
2730 output_asm_insn ("{stbys|stby},e %%r0,%2(%0)", operands);
2735 /* Pre-adjust the loop counter. */
2736 operands[2] = GEN_INT (n_bytes - 4);
2737 output_asm_insn ("ldi %2,%1", operands);
2740 output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands);
2741 output_asm_insn ("addib,>= -4,%1,.-4", operands);
2742 output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands);
2744 /* Handle the residual. */
2745 if (n_bytes % 4 != 0)
2747 if (n_bytes % 4 >= 2)
2748 output_asm_insn ("{sths|sth},ma %%r0,2(%0)", operands);
2749 if (n_bytes % 2 != 0)
2750 output_asm_insn ("stb %%r0,0(%0)", operands);
2755 /* Pre-adjust the loop counter. */
2756 operands[2] = GEN_INT (n_bytes - 2);
2757 output_asm_insn ("ldi %2,%1", operands);
2760 output_asm_insn ("{stbs|stb},ma %%r0,1(%0)", operands);
2761 output_asm_insn ("addib,>= -2,%1,.-4", operands);
2762 output_asm_insn ("{stbs|stb},ma %%r0,1(%0)", operands);
2764 /* Handle the residual. */
2765 if (n_bytes % 2 != 0)
2766 output_asm_insn ("stb %%r0,0(%0)", operands);
2775 /* Count the number of insns necessary to handle this block move.
2777 Basic structure is the same as emit_block_move, except that we
2778 count insns rather than emit them. */
2781 compute_clrmem_length (rtx insn)
2783 rtx pat = PATTERN (insn);
2784 unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 4), 0));
2785 unsigned long n_bytes = INTVAL (XEXP (XVECEXP (pat, 0, 3), 0));
2786 unsigned int n_insns = 0;
2788 /* We can't clear more than a word at a time because the PA
2789 has no longer integer move insns. */
2790 if (align > (TARGET_64BIT ? 8 : 4))
2791 align = (TARGET_64BIT ? 8 : 4);
2793 /* The basic loop. */
2797 if (n_bytes % (2 * align) != 0)
2799 if ((n_bytes % (2 * align)) >= align)
2802 if ((n_bytes % align) != 0)
2806 /* Lengths are expressed in bytes now; each insn is 4 bytes. */
2812 output_and (rtx *operands)
2814 if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0)
2816 unsigned HOST_WIDE_INT mask = INTVAL (operands[2]);
2817 int ls0, ls1, ms0, p, len;
2819 for (ls0 = 0; ls0 < 32; ls0++)
2820 if ((mask & (1 << ls0)) == 0)
2823 for (ls1 = ls0; ls1 < 32; ls1++)
2824 if ((mask & (1 << ls1)) != 0)
2827 for (ms0 = ls1; ms0 < 32; ms0++)
2828 if ((mask & (1 << ms0)) == 0)
2831 gcc_assert (ms0 == 32);
2839 operands[2] = GEN_INT (len);
2840 return "{extru|extrw,u} %1,31,%2,%0";
2844 /* We could use this `depi' for the case above as well, but `depi'
2845 requires one more register file access than an `extru'. */
2850 operands[2] = GEN_INT (p);
2851 operands[3] = GEN_INT (len);
2852 return "{depi|depwi} 0,%2,%3,%0";
2856 return "and %1,%2,%0";
2859 /* Return a string to perform a bitwise-and of operands[1] with operands[2]
2860 storing the result in operands[0]. */
2862 output_64bit_and (rtx *operands)
2864 if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0)
2866 unsigned HOST_WIDE_INT mask = INTVAL (operands[2]);
2867 int ls0, ls1, ms0, p, len;
2869 for (ls0 = 0; ls0 < HOST_BITS_PER_WIDE_INT; ls0++)
2870 if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls0)) == 0)
2873 for (ls1 = ls0; ls1 < HOST_BITS_PER_WIDE_INT; ls1++)
2874 if ((mask & ((unsigned HOST_WIDE_INT) 1 << ls1)) != 0)
2877 for (ms0 = ls1; ms0 < HOST_BITS_PER_WIDE_INT; ms0++)
2878 if ((mask & ((unsigned HOST_WIDE_INT) 1 << ms0)) == 0)
2881 gcc_assert (ms0 == HOST_BITS_PER_WIDE_INT);
2883 if (ls1 == HOST_BITS_PER_WIDE_INT)
2889 operands[2] = GEN_INT (len);
2890 return "extrd,u %1,63,%2,%0";
2894 /* We could use this `depi' for the case above as well, but `depi'
2895 requires one more register file access than an `extru'. */
2900 operands[2] = GEN_INT (p);
2901 operands[3] = GEN_INT (len);
2902 return "depdi 0,%2,%3,%0";
2906 return "and %1,%2,%0";
2910 output_ior (rtx *operands)
2912 unsigned HOST_WIDE_INT mask = INTVAL (operands[2]);
2913 int bs0, bs1, p, len;
2915 if (INTVAL (operands[2]) == 0)
2916 return "copy %1,%0";
2918 for (bs0 = 0; bs0 < 32; bs0++)
2919 if ((mask & (1 << bs0)) != 0)
2922 for (bs1 = bs0; bs1 < 32; bs1++)
2923 if ((mask & (1 << bs1)) == 0)
2926 gcc_assert (bs1 == 32 || ((unsigned HOST_WIDE_INT) 1 << bs1) > mask);
2931 operands[2] = GEN_INT (p);
2932 operands[3] = GEN_INT (len);
2933 return "{depi|depwi} -1,%2,%3,%0";
2936 /* Return a string to perform a bitwise-and of operands[1] with operands[2]
2937 storing the result in operands[0]. */
2939 output_64bit_ior (rtx *operands)
2941 unsigned HOST_WIDE_INT mask = INTVAL (operands[2]);
2942 int bs0, bs1, p, len;
2944 if (INTVAL (operands[2]) == 0)
2945 return "copy %1,%0";
2947 for (bs0 = 0; bs0 < HOST_BITS_PER_WIDE_INT; bs0++)
2948 if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs0)) != 0)
2951 for (bs1 = bs0; bs1 < HOST_BITS_PER_WIDE_INT; bs1++)
2952 if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs1)) == 0)
2955 gcc_assert (bs1 == HOST_BITS_PER_WIDE_INT
2956 || ((unsigned HOST_WIDE_INT) 1 << bs1) > mask);
2961 operands[2] = GEN_INT (p);
2962 operands[3] = GEN_INT (len);
2963 return "depdi -1,%2,%3,%0";
2966 /* Target hook for assembling integer objects. This code handles
2967 aligned SI and DI integers specially since function references
2968 must be preceded by P%. */
2971 pa_assemble_integer (rtx x, unsigned int size, int aligned_p)
2973 if (size == UNITS_PER_WORD
2975 && function_label_operand (x, VOIDmode))
2977 fputs (size == 8? "\t.dword\tP%" : "\t.word\tP%", asm_out_file);
2978 output_addr_const (asm_out_file, x);
2979 fputc ('\n', asm_out_file);
2982 return default_assemble_integer (x, size, aligned_p);
2985 /* Output an ascii string. */
2987 output_ascii (FILE *file, const char *p, int size)
2991 unsigned char partial_output[16]; /* Max space 4 chars can occupy. */
2993 /* The HP assembler can only take strings of 256 characters at one
2994 time. This is a limitation on input line length, *not* the
2995 length of the string. Sigh. Even worse, it seems that the
2996 restriction is in number of input characters (see \xnn &
2997 \whatever). So we have to do this very carefully. */
2999 fputs ("\t.STRING \"", file);
3002 for (i = 0; i < size; i += 4)
3006 for (io = 0, co = 0; io < MIN (4, size - i); io++)
3008 register unsigned int c = (unsigned char) p[i + io];
3010 if (c == '\"' || c == '\\')
3011 partial_output[co++] = '\\';
3012 if (c >= ' ' && c < 0177)
3013 partial_output[co++] = c;
3017 partial_output[co++] = '\\';
3018 partial_output[co++] = 'x';
3019 hexd = c / 16 - 0 + '0';
3021 hexd -= '9' - 'a' + 1;
3022 partial_output[co++] = hexd;
3023 hexd = c % 16 - 0 + '0';
3025 hexd -= '9' - 'a' + 1;
3026 partial_output[co++] = hexd;
3029 if (chars_output + co > 243)
3031 fputs ("\"\n\t.STRING \"", file);
3034 fwrite (partial_output, 1, (size_t) co, file);
3038 fputs ("\"\n", file);
3041 /* Try to rewrite floating point comparisons & branches to avoid
3042 useless add,tr insns.
3044 CHECK_NOTES is nonzero if we should examine REG_DEAD notes
3045 to see if FPCC is dead. CHECK_NOTES is nonzero for the
3046 first attempt to remove useless add,tr insns. It is zero
3047 for the second pass as reorg sometimes leaves bogus REG_DEAD
3050 When CHECK_NOTES is zero we can only eliminate add,tr insns
3051 when there's a 1:1 correspondence between fcmp and ftest/fbranch
3054 remove_useless_addtr_insns (int check_notes)
3057 static int pass = 0;
3059 /* This is fairly cheap, so always run it when optimizing. */
3063 int fbranch_count = 0;
3065 /* Walk all the insns in this function looking for fcmp & fbranch
3066 instructions. Keep track of how many of each we find. */
3067 for (insn = get_insns (); insn; insn = next_insn (insn))
3071 /* Ignore anything that isn't an INSN or a JUMP_INSN. */
3072 if (GET_CODE (insn) != INSN && GET_CODE (insn) != JUMP_INSN)
3075 tmp = PATTERN (insn);
3077 /* It must be a set. */
3078 if (GET_CODE (tmp) != SET)
3081 /* If the destination is CCFP, then we've found an fcmp insn. */
3082 tmp = SET_DEST (tmp);
3083 if (GET_CODE (tmp) == REG && REGNO (tmp) == 0)
3089 tmp = PATTERN (insn);
3090 /* If this is an fbranch instruction, bump the fbranch counter. */
3091 if (GET_CODE (tmp) == SET
3092 && SET_DEST (tmp) == pc_rtx
3093 && GET_CODE (SET_SRC (tmp)) == IF_THEN_ELSE
3094 && GET_CODE (XEXP (SET_SRC (tmp), 0)) == NE
3095 && GET_CODE (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == REG
3096 && REGNO (XEXP (XEXP (SET_SRC (tmp), 0), 0)) == 0)
3104 /* Find all floating point compare + branch insns. If possible,
3105 reverse the comparison & the branch to avoid add,tr insns. */
3106 for (insn = get_insns (); insn; insn = next_insn (insn))
3110 /* Ignore anything that isn't an INSN. */
3111 if (GET_CODE (insn) != INSN)
3114 tmp = PATTERN (insn);
3116 /* It must be a set. */
3117 if (GET_CODE (tmp) != SET)
3120 /* The destination must be CCFP, which is register zero. */
3121 tmp = SET_DEST (tmp);
3122 if (GET_CODE (tmp) != REG || REGNO (tmp) != 0)
3125 /* INSN should be a set of CCFP.
3127 See if the result of this insn is used in a reversed FP
3128 conditional branch. If so, reverse our condition and
3129 the branch. Doing so avoids useless add,tr insns. */
3130 next = next_insn (insn);
3133 /* Jumps, calls and labels stop our search. */
3134 if (GET_CODE (next) == JUMP_INSN
3135 || GET_CODE (next) == CALL_INSN
3136 || GET_CODE (next) == CODE_LABEL)
3139 /* As does another fcmp insn. */
3140 if (GET_CODE (next) == INSN
3141 && GET_CODE (PATTERN (next)) == SET
3142 && GET_CODE (SET_DEST (PATTERN (next))) == REG
3143 && REGNO (SET_DEST (PATTERN (next))) == 0)
3146 next = next_insn (next);
3149 /* Is NEXT_INSN a branch? */
3151 && GET_CODE (next) == JUMP_INSN)
3153 rtx pattern = PATTERN (next);
3155 /* If it a reversed fp conditional branch (e.g. uses add,tr)
3156 and CCFP dies, then reverse our conditional and the branch
3157 to avoid the add,tr. */
3158 if (GET_CODE (pattern) == SET
3159 && SET_DEST (pattern) == pc_rtx
3160 && GET_CODE (SET_SRC (pattern)) == IF_THEN_ELSE
3161 && GET_CODE (XEXP (SET_SRC (pattern), 0)) == NE
3162 && GET_CODE (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == REG
3163 && REGNO (XEXP (XEXP (SET_SRC (pattern), 0), 0)) == 0
3164 && GET_CODE (XEXP (SET_SRC (pattern), 1)) == PC
3165 && (fcmp_count == fbranch_count
3167 && find_regno_note (next, REG_DEAD, 0))))
3169 /* Reverse the branch. */
3170 tmp = XEXP (SET_SRC (pattern), 1);
3171 XEXP (SET_SRC (pattern), 1) = XEXP (SET_SRC (pattern), 2);
3172 XEXP (SET_SRC (pattern), 2) = tmp;
3173 INSN_CODE (next) = -1;
3175 /* Reverse our condition. */
3176 tmp = PATTERN (insn);
3177 PUT_CODE (XEXP (tmp, 1),
3178 (reverse_condition_maybe_unordered
3179 (GET_CODE (XEXP (tmp, 1)))));
3189 /* You may have trouble believing this, but this is the 32 bit HP-PA
3194 Variable arguments (optional; any number may be allocated)
3196 SP-(4*(N+9)) arg word N
3201 Fixed arguments (must be allocated; may remain unused)
3210 SP-32 External Data Pointer (DP)
3212 SP-24 External/stub RP (RP')
3216 SP-8 Calling Stub RP (RP'')
3221 SP-0 Stack Pointer (points to next available address)
3225 /* This function saves registers as follows. Registers marked with ' are
3226 this function's registers (as opposed to the previous function's).
3227 If a frame_pointer isn't needed, r4 is saved as a general register;
3228 the space for the frame pointer is still allocated, though, to keep
3234 SP (FP') Previous FP
3235 SP + 4 Alignment filler (sigh)
3236 SP + 8 Space for locals reserved here.
3240 SP + n All call saved register used.
3244 SP + o All call saved fp registers used.
3248 SP + p (SP') points to next available address.
3252 /* Global variables set by output_function_prologue(). */
3253 /* Size of frame. Need to know this to emit return insns from
3255 static HOST_WIDE_INT actual_fsize, local_fsize;
3256 static int save_fregs;
3258 /* Emit RTL to store REG at the memory location specified by BASE+DISP.
3259 Handle case where DISP > 8k by using the add_high_const patterns.
3261 Note in DISP > 8k case, we will leave the high part of the address
3262 in %r1. There is code in expand_hppa_{prologue,epilogue} that knows this.*/
3265 store_reg (int reg, HOST_WIDE_INT disp, int base)
3267 rtx insn, dest, src, basereg;
3269 src = gen_rtx_REG (word_mode, reg);
3270 basereg = gen_rtx_REG (Pmode, base);
3271 if (VAL_14_BITS_P (disp))
3273 dest = gen_rtx_MEM (word_mode, plus_constant (basereg, disp));
3274 insn = emit_move_insn (dest, src);
3276 else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
3278 rtx delta = GEN_INT (disp);
3279 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3281 emit_move_insn (tmpreg, delta);
3282 insn = emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
3286 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
3287 gen_rtx_SET (VOIDmode, tmpreg,
3288 gen_rtx_PLUS (Pmode, basereg, delta)),
3290 RTX_FRAME_RELATED_P (insn) = 1;
3292 dest = gen_rtx_MEM (word_mode, tmpreg);
3293 insn = emit_move_insn (dest, src);
3297 rtx delta = GEN_INT (disp);
3298 rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta));
3299 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3301 emit_move_insn (tmpreg, high);
3302 dest = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta));
3303 insn = emit_move_insn (dest, src);
3307 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
3308 gen_rtx_SET (VOIDmode,
3309 gen_rtx_MEM (word_mode,
3310 gen_rtx_PLUS (word_mode, basereg,
3318 RTX_FRAME_RELATED_P (insn) = 1;
3321 /* Emit RTL to store REG at the memory location specified by BASE and then
3322 add MOD to BASE. MOD must be <= 8k. */
3325 store_reg_modify (int base, int reg, HOST_WIDE_INT mod)
3327 rtx insn, basereg, srcreg, delta;
3329 gcc_assert (VAL_14_BITS_P (mod));
3331 basereg = gen_rtx_REG (Pmode, base);
3332 srcreg = gen_rtx_REG (word_mode, reg);
3333 delta = GEN_INT (mod);
3335 insn = emit_insn (gen_post_store (basereg, srcreg, delta));
3338 RTX_FRAME_RELATED_P (insn) = 1;
3340 /* RTX_FRAME_RELATED_P must be set on each frame related set
3341 in a parallel with more than one element. */
3342 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1;
3343 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
3347 /* Emit RTL to set REG to the value specified by BASE+DISP. Handle case
3348 where DISP > 8k by using the add_high_const patterns. NOTE indicates
3349 whether to add a frame note or not.
3351 In the DISP > 8k case, we leave the high part of the address in %r1.
3352 There is code in expand_hppa_{prologue,epilogue} that knows about this. */
3355 set_reg_plus_d (int reg, int base, HOST_WIDE_INT disp, int note)
3359 if (VAL_14_BITS_P (disp))
3361 insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
3362 plus_constant (gen_rtx_REG (Pmode, base), disp));
3364 else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
3366 rtx basereg = gen_rtx_REG (Pmode, base);
3367 rtx delta = GEN_INT (disp);
3368 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3370 emit_move_insn (tmpreg, delta);
3371 insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
3372 gen_rtx_PLUS (Pmode, tmpreg, basereg));
3375 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
3376 gen_rtx_SET (VOIDmode, tmpreg,
3377 gen_rtx_PLUS (Pmode, basereg, delta)),
3382 rtx basereg = gen_rtx_REG (Pmode, base);
3383 rtx delta = GEN_INT (disp);
3384 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3386 emit_move_insn (tmpreg,
3387 gen_rtx_PLUS (Pmode, basereg,
3388 gen_rtx_HIGH (Pmode, delta)));
3389 insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
3390 gen_rtx_LO_SUM (Pmode, tmpreg, delta));
3393 if (DO_FRAME_NOTES && note)
3394 RTX_FRAME_RELATED_P (insn) = 1;
3398 compute_frame_size (HOST_WIDE_INT size, int *fregs_live)
3403 /* The code in hppa_expand_prologue and hppa_expand_epilogue must
3404 be consistent with the rounding and size calculation done here.
3405 Change them at the same time. */
3407 /* We do our own stack alignment. First, round the size of the
3408 stack locals up to a word boundary. */
3409 size = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1);
3411 /* Space for previous frame pointer + filler. If any frame is
3412 allocated, we need to add in the STARTING_FRAME_OFFSET. We
3413 waste some space here for the sake of HP compatibility. The
3414 first slot is only used when the frame pointer is needed. */
3415 if (size || frame_pointer_needed)
3416 size += STARTING_FRAME_OFFSET;
3418 /* If the current function calls __builtin_eh_return, then we need
3419 to allocate stack space for registers that will hold data for
3420 the exception handler. */
3421 if (DO_FRAME_NOTES && current_function_calls_eh_return)
3425 for (i = 0; EH_RETURN_DATA_REGNO (i) != INVALID_REGNUM; ++i)
3427 size += i * UNITS_PER_WORD;
3430 /* Account for space used by the callee general register saves. */
3431 for (i = 18, j = frame_pointer_needed ? 4 : 3; i >= j; i--)
3432 if (regs_ever_live[i])
3433 size += UNITS_PER_WORD;
3435 /* Account for space used by the callee floating point register saves. */
3436 for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP)
3437 if (regs_ever_live[i]
3438 || (!TARGET_64BIT && regs_ever_live[i + 1]))
3442 /* We always save both halves of the FP register, so always
3443 increment the frame size by 8 bytes. */
3447 /* If any of the floating registers are saved, account for the
3448 alignment needed for the floating point register save block. */
3451 size = (size + 7) & ~7;
3456 /* The various ABIs include space for the outgoing parameters in the
3457 size of the current function's stack frame. We don't need to align
3458 for the outgoing arguments as their alignment is set by the final
3459 rounding for the frame as a whole. */
3460 size += current_function_outgoing_args_size;
3462 /* Allocate space for the fixed frame marker. This space must be
3463 allocated for any function that makes calls or allocates
3465 if (!current_function_is_leaf || size)
3466 size += TARGET_64BIT ? 48 : 32;
3468 /* Finally, round to the preferred stack boundary. */
3469 return ((size + PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)
3470 & ~(PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1));
3473 /* Generate the assembly code for function entry. FILE is a stdio
3474 stream to output the code to. SIZE is an int: how many units of
3475 temporary storage to allocate.
3477 Refer to the array `regs_ever_live' to determine which registers to
3478 save; `regs_ever_live[I]' is nonzero if register number I is ever
3479 used in the function. This function is responsible for knowing
3480 which registers should not be saved even if used. */
3482 /* On HP-PA, move-double insns between fpu and cpu need an 8-byte block
3483 of memory. If any fpu reg is used in the function, we allocate
3484 such a block here, at the bottom of the frame, just in case it's needed.
3486 If this function is a leaf procedure, then we may choose not
3487 to do a "save" insn. The decision about whether or not
3488 to do this is made in regclass.c. */
3491 pa_output_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
3493 /* The function's label and associated .PROC must never be
3494 separated and must be output *after* any profiling declarations
3495 to avoid changing spaces/subspaces within a procedure. */
3496 ASM_OUTPUT_LABEL (file, XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0));
3497 fputs ("\t.PROC\n", file);
3499 /* hppa_expand_prologue does the dirty work now. We just need
3500 to output the assembler directives which denote the start
3502 fprintf (file, "\t.CALLINFO FRAME=" HOST_WIDE_INT_PRINT_DEC, actual_fsize);
3503 if (regs_ever_live[2])
3504 fputs (",CALLS,SAVE_RP", file);
3506 fputs (",NO_CALLS", file);
3508 /* The SAVE_SP flag is used to indicate that register %r3 is stored
3509 at the beginning of the frame and that it is used as the frame
3510 pointer for the frame. We do this because our current frame
3511 layout doesn't conform to that specified in the HP runtime
3512 documentation and we need a way to indicate to programs such as
3513 GDB where %r3 is saved. The SAVE_SP flag was chosen because it
3514 isn't used by HP compilers but is supported by the assembler.
3515 However, SAVE_SP is supposed to indicate that the previous stack
3516 pointer has been saved in the frame marker. */
3517 if (frame_pointer_needed)
3518 fputs (",SAVE_SP", file);
3520 /* Pass on information about the number of callee register saves
3521 performed in the prologue.
3523 The compiler is supposed to pass the highest register number
3524 saved, the assembler then has to adjust that number before
3525 entering it into the unwind descriptor (to account for any
3526 caller saved registers with lower register numbers than the
3527 first callee saved register). */
3529 fprintf (file, ",ENTRY_GR=%d", gr_saved + 2);
3532 fprintf (file, ",ENTRY_FR=%d", fr_saved + 11);
3534 fputs ("\n\t.ENTRY\n", file);
3536 remove_useless_addtr_insns (0);
3540 hppa_expand_prologue (void)
3542 int merge_sp_adjust_with_store = 0;
3543 HOST_WIDE_INT size = get_frame_size ();
3544 HOST_WIDE_INT offset;
3552 /* Compute total size for frame pointer, filler, locals and rounding to
3553 the next word boundary. Similar code appears in compute_frame_size
3554 and must be changed in tandem with this code. */
3555 local_fsize = (size + UNITS_PER_WORD - 1) & ~(UNITS_PER_WORD - 1);
3556 if (local_fsize || frame_pointer_needed)
3557 local_fsize += STARTING_FRAME_OFFSET;
3559 actual_fsize = compute_frame_size (size, &save_fregs);
3561 /* Compute a few things we will use often. */
3562 tmpreg = gen_rtx_REG (word_mode, 1);
3564 /* Save RP first. The calling conventions manual states RP will
3565 always be stored into the caller's frame at sp - 20 or sp - 16
3566 depending on which ABI is in use. */
3567 if (regs_ever_live[2] || current_function_calls_eh_return)
3568 store_reg (2, TARGET_64BIT ? -16 : -20, STACK_POINTER_REGNUM);
3570 /* Allocate the local frame and set up the frame pointer if needed. */
3571 if (actual_fsize != 0)
3573 if (frame_pointer_needed)
3575 /* Copy the old frame pointer temporarily into %r1. Set up the
3576 new stack pointer, then store away the saved old frame pointer
3577 into the stack at sp and at the same time update the stack
3578 pointer by actual_fsize bytes. Two versions, first
3579 handles small (<8k) frames. The second handles large (>=8k)
3581 insn = emit_move_insn (tmpreg, frame_pointer_rtx);
3583 RTX_FRAME_RELATED_P (insn) = 1;
3585 insn = emit_move_insn (frame_pointer_rtx, stack_pointer_rtx);
3587 RTX_FRAME_RELATED_P (insn) = 1;
3589 if (VAL_14_BITS_P (actual_fsize))
3590 store_reg_modify (STACK_POINTER_REGNUM, 1, actual_fsize);
3593 /* It is incorrect to store the saved frame pointer at *sp,
3594 then increment sp (writes beyond the current stack boundary).
3596 So instead use stwm to store at *sp and post-increment the
3597 stack pointer as an atomic operation. Then increment sp to
3598 finish allocating the new frame. */
3599 HOST_WIDE_INT adjust1 = 8192 - 64;
3600 HOST_WIDE_INT adjust2 = actual_fsize - adjust1;
3602 store_reg_modify (STACK_POINTER_REGNUM, 1, adjust1);
3603 set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM,
3607 /* We set SAVE_SP in frames that need a frame pointer. Thus,
3608 we need to store the previous stack pointer (frame pointer)
3609 into the frame marker on targets that use the HP unwind
3610 library. This allows the HP unwind library to be used to
3611 unwind GCC frames. However, we are not fully compatible
3612 with the HP library because our frame layout differs from
3613 that specified in the HP runtime specification.
3615 We don't want a frame note on this instruction as the frame
3616 marker moves during dynamic stack allocation.
3618 This instruction also serves as a blockage to prevent
3619 register spills from being scheduled before the stack
3620 pointer is raised. This is necessary as we store
3621 registers using the frame pointer as a base register,
3622 and the frame pointer is set before sp is raised. */
3623 if (TARGET_HPUX_UNWIND_LIBRARY)
3625 rtx addr = gen_rtx_PLUS (word_mode, stack_pointer_rtx,
3626 GEN_INT (TARGET_64BIT ? -8 : -4));
3628 emit_move_insn (gen_rtx_MEM (word_mode, addr),
3632 emit_insn (gen_blockage ());
3634 /* no frame pointer needed. */
3637 /* In some cases we can perform the first callee register save
3638 and allocating the stack frame at the same time. If so, just
3639 make a note of it and defer allocating the frame until saving
3640 the callee registers. */
3641 if (VAL_14_BITS_P (actual_fsize) && local_fsize == 0)
3642 merge_sp_adjust_with_store = 1;
3643 /* Can not optimize. Adjust the stack frame by actual_fsize
3646 set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM,
3651 /* Normal register save.
3653 Do not save the frame pointer in the frame_pointer_needed case. It
3654 was done earlier. */
3655 if (frame_pointer_needed)
3657 offset = local_fsize;
3659 /* Saving the EH return data registers in the frame is the simplest
3660 way to get the frame unwind information emitted. We put them
3661 just before the general registers. */
3662 if (DO_FRAME_NOTES && current_function_calls_eh_return)
3664 unsigned int i, regno;
3668 regno = EH_RETURN_DATA_REGNO (i);
3669 if (regno == INVALID_REGNUM)
3672 store_reg (regno, offset, FRAME_POINTER_REGNUM);
3673 offset += UNITS_PER_WORD;
3677 for (i = 18; i >= 4; i--)
3678 if (regs_ever_live[i] && ! call_used_regs[i])
3680 store_reg (i, offset, FRAME_POINTER_REGNUM);
3681 offset += UNITS_PER_WORD;
3684 /* Account for %r3 which is saved in a special place. */
3687 /* No frame pointer needed. */
3690 offset = local_fsize - actual_fsize;
3692 /* Saving the EH return data registers in the frame is the simplest
3693 way to get the frame unwind information emitted. */
3694 if (DO_FRAME_NOTES && current_function_calls_eh_return)
3696 unsigned int i, regno;
3700 regno = EH_RETURN_DATA_REGNO (i);
3701 if (regno == INVALID_REGNUM)
3704 /* If merge_sp_adjust_with_store is nonzero, then we can
3705 optimize the first save. */
3706 if (merge_sp_adjust_with_store)
3708 store_reg_modify (STACK_POINTER_REGNUM, regno, -offset);
3709 merge_sp_adjust_with_store = 0;
3712 store_reg (regno, offset, STACK_POINTER_REGNUM);
3713 offset += UNITS_PER_WORD;
3717 for (i = 18; i >= 3; i--)
3718 if (regs_ever_live[i] && ! call_used_regs[i])
3720 /* If merge_sp_adjust_with_store is nonzero, then we can
3721 optimize the first GR save. */
3722 if (merge_sp_adjust_with_store)
3724 store_reg_modify (STACK_POINTER_REGNUM, i, -offset);
3725 merge_sp_adjust_with_store = 0;
3728 store_reg (i, offset, STACK_POINTER_REGNUM);
3729 offset += UNITS_PER_WORD;
3733 /* If we wanted to merge the SP adjustment with a GR save, but we never
3734 did any GR saves, then just emit the adjustment here. */
3735 if (merge_sp_adjust_with_store)
3736 set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM,
3740 /* The hppa calling conventions say that %r19, the pic offset
3741 register, is saved at sp - 32 (in this function's frame)
3742 when generating PIC code. FIXME: What is the correct thing
3743 to do for functions which make no calls and allocate no
3744 frame? Do we need to allocate a frame, or can we just omit
3745 the save? For now we'll just omit the save.
3747 We don't want a note on this insn as the frame marker can
3748 move if there is a dynamic stack allocation. */
3749 if (flag_pic && actual_fsize != 0 && !TARGET_64BIT)
3751 rtx addr = gen_rtx_PLUS (word_mode, stack_pointer_rtx, GEN_INT (-32));
3753 emit_move_insn (gen_rtx_MEM (word_mode, addr), pic_offset_table_rtx);
3757 /* Align pointer properly (doubleword boundary). */
3758 offset = (offset + 7) & ~7;
3760 /* Floating point register store. */
3765 /* First get the frame or stack pointer to the start of the FP register
3767 if (frame_pointer_needed)
3769 set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0);
3770 base = frame_pointer_rtx;
3774 set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0);
3775 base = stack_pointer_rtx;
3778 /* Now actually save the FP registers. */
3779 for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP)
3781 if (regs_ever_live[i]
3782 || (! TARGET_64BIT && regs_ever_live[i + 1]))
3784 rtx addr, insn, reg;
3785 addr = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg));
3786 reg = gen_rtx_REG (DFmode, i);
3787 insn = emit_move_insn (addr, reg);
3790 RTX_FRAME_RELATED_P (insn) = 1;
3793 rtx mem = gen_rtx_MEM (DFmode,
3794 plus_constant (base, offset));
3796 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
3797 gen_rtx_SET (VOIDmode, mem, reg),
3802 rtx meml = gen_rtx_MEM (SFmode,
3803 plus_constant (base, offset));
3804 rtx memr = gen_rtx_MEM (SFmode,
3805 plus_constant (base, offset + 4));
3806 rtx regl = gen_rtx_REG (SFmode, i);
3807 rtx regr = gen_rtx_REG (SFmode, i + 1);
3808 rtx setl = gen_rtx_SET (VOIDmode, meml, regl);
3809 rtx setr = gen_rtx_SET (VOIDmode, memr, regr);
3812 RTX_FRAME_RELATED_P (setl) = 1;
3813 RTX_FRAME_RELATED_P (setr) = 1;
3814 vec = gen_rtvec (2, setl, setr);
3816 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
3817 gen_rtx_SEQUENCE (VOIDmode, vec),
3821 offset += GET_MODE_SIZE (DFmode);
3828 /* Emit RTL to load REG from the memory location specified by BASE+DISP.
3829 Handle case where DISP > 8k by using the add_high_const patterns. */
3832 load_reg (int reg, HOST_WIDE_INT disp, int base)
3834 rtx dest = gen_rtx_REG (word_mode, reg);
3835 rtx basereg = gen_rtx_REG (Pmode, base);
3838 if (VAL_14_BITS_P (disp))
3839 src = gen_rtx_MEM (word_mode, plus_constant (basereg, disp));
3840 else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
3842 rtx delta = GEN_INT (disp);
3843 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3845 emit_move_insn (tmpreg, delta);
3846 if (TARGET_DISABLE_INDEXING)
3848 emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
3849 src = gen_rtx_MEM (word_mode, tmpreg);
3852 src = gen_rtx_MEM (word_mode, gen_rtx_PLUS (Pmode, tmpreg, basereg));
3856 rtx delta = GEN_INT (disp);
3857 rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta));
3858 rtx tmpreg = gen_rtx_REG (Pmode, 1);
3860 emit_move_insn (tmpreg, high);
3861 src = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta));
3864 emit_move_insn (dest, src);
3867 /* Update the total code bytes output to the text section. */
3870 update_total_code_bytes (int nbytes)
3872 if ((TARGET_PORTABLE_RUNTIME || !TARGET_GAS || !TARGET_SOM)
3873 && !IN_NAMED_SECTION_P (cfun->decl))
3875 if (INSN_ADDRESSES_SET_P ())
3877 unsigned long old_total = total_code_bytes;
3879 total_code_bytes += nbytes;
3881 /* Be prepared to handle overflows. */
3882 if (old_total > total_code_bytes)
3883 total_code_bytes = -1;
3886 total_code_bytes = -1;
3890 /* This function generates the assembly code for function exit.
3891 Args are as for output_function_prologue ().
3893 The function epilogue should not depend on the current stack
3894 pointer! It should use the frame pointer only. This is mandatory
3895 because of alloca; we also take advantage of it to omit stack
3896 adjustments before returning. */
3899 pa_output_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
3901 rtx insn = get_last_insn ();
3905 /* hppa_expand_epilogue does the dirty work now. We just need
3906 to output the assembler directives which denote the end
3909 To make debuggers happy, emit a nop if the epilogue was completely
3910 eliminated due to a volatile call as the last insn in the
3911 current function. That way the return address (in %r2) will
3912 always point to a valid instruction in the current function. */
3914 /* Get the last real insn. */
3915 if (GET_CODE (insn) == NOTE)
3916 insn = prev_real_insn (insn);
3918 /* If it is a sequence, then look inside. */
3919 if (insn && GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE)
3920 insn = XVECEXP (PATTERN (insn), 0, 0);
3922 /* If insn is a CALL_INSN, then it must be a call to a volatile
3923 function (otherwise there would be epilogue insns). */
3924 if (insn && GET_CODE (insn) == CALL_INSN)
3926 fputs ("\tnop\n", file);
3930 fputs ("\t.EXIT\n\t.PROCEND\n", file);
3932 if (TARGET_SOM && TARGET_GAS)
3934 /* We done with this subspace except possibly for some additional
3935 debug information. Forget that we are in this subspace to ensure
3936 that the next function is output in its own subspace. */
3938 cfun->machine->in_nsubspa = 2;
3941 if (INSN_ADDRESSES_SET_P ())
3943 insn = get_last_nonnote_insn ();
3944 last_address += INSN_ADDRESSES (INSN_UID (insn));
3946 last_address += insn_default_length (insn);
3947 last_address = ((last_address + FUNCTION_BOUNDARY / BITS_PER_UNIT - 1)
3948 & ~(FUNCTION_BOUNDARY / BITS_PER_UNIT - 1));
3951 /* Finally, update the total number of code bytes output so far. */
3952 update_total_code_bytes (last_address);
3956 hppa_expand_epilogue (void)
3959 HOST_WIDE_INT offset;
3960 HOST_WIDE_INT ret_off = 0;
3962 int merge_sp_adjust_with_load = 0;
3964 /* We will use this often. */
3965 tmpreg = gen_rtx_REG (word_mode, 1);
3967 /* Try to restore RP early to avoid load/use interlocks when
3968 RP gets used in the return (bv) instruction. This appears to still
3969 be necessary even when we schedule the prologue and epilogue. */
3970 if (regs_ever_live [2] || current_function_calls_eh_return)
3972 ret_off = TARGET_64BIT ? -16 : -20;
3973 if (frame_pointer_needed)
3975 load_reg (2, ret_off, FRAME_POINTER_REGNUM);
3980 /* No frame pointer, and stack is smaller than 8k. */
3981 if (VAL_14_BITS_P (ret_off - actual_fsize))
3983 load_reg (2, ret_off - actual_fsize, STACK_POINTER_REGNUM);
3989 /* General register restores. */
3990 if (frame_pointer_needed)
3992 offset = local_fsize;
3994 /* If the current function calls __builtin_eh_return, then we need
3995 to restore the saved EH data registers. */
3996 if (DO_FRAME_NOTES && current_function_calls_eh_return)
3998 unsigned int i, regno;
4002 regno = EH_RETURN_DATA_REGNO (i);
4003 if (regno == INVALID_REGNUM)
4006 load_reg (regno, offset, FRAME_POINTER_REGNUM);
4007 offset += UNITS_PER_WORD;
4011 for (i = 18; i >= 4; i--)
4012 if (regs_ever_live[i] && ! call_used_regs[i])
4014 load_reg (i, offset, FRAME_POINTER_REGNUM);
4015 offset += UNITS_PER_WORD;
4020 offset = local_fsize - actual_fsize;
4022 /* If the current function calls __builtin_eh_return, then we need
4023 to restore the saved EH data registers. */
4024 if (DO_FRAME_NOTES && current_function_calls_eh_return)
4026 unsigned int i, regno;
4030 regno = EH_RETURN_DATA_REGNO (i);
4031 if (regno == INVALID_REGNUM)
4034 /* Only for the first load.
4035 merge_sp_adjust_with_load holds the register load
4036 with which we will merge the sp adjustment. */
4037 if (merge_sp_adjust_with_load == 0
4039 && VAL_14_BITS_P (-actual_fsize))
4040 merge_sp_adjust_with_load = regno;
4042 load_reg (regno, offset, STACK_POINTER_REGNUM);
4043 offset += UNITS_PER_WORD;
4047 for (i = 18; i >= 3; i--)
4049 if (regs_ever_live[i] && ! call_used_regs[i])
4051 /* Only for the first load.
4052 merge_sp_adjust_with_load holds the register load
4053 with which we will merge the sp adjustment. */
4054 if (merge_sp_adjust_with_load == 0
4056 && VAL_14_BITS_P (-actual_fsize))
4057 merge_sp_adjust_with_load = i;
4059 load_reg (i, offset, STACK_POINTER_REGNUM);
4060 offset += UNITS_PER_WORD;
4065 /* Align pointer properly (doubleword boundary). */
4066 offset = (offset + 7) & ~7;
4068 /* FP register restores. */
4071 /* Adjust the register to index off of. */
4072 if (frame_pointer_needed)
4073 set_reg_plus_d (1, FRAME_POINTER_REGNUM, offset, 0);
4075 set_reg_plus_d (1, STACK_POINTER_REGNUM, offset, 0);
4077 /* Actually do the restores now. */
4078 for (i = FP_SAVED_REG_LAST; i >= FP_SAVED_REG_FIRST; i -= FP_REG_STEP)
4079 if (regs_ever_live[i]
4080 || (! TARGET_64BIT && regs_ever_live[i + 1]))
4082 rtx src = gen_rtx_MEM (DFmode, gen_rtx_POST_INC (DFmode, tmpreg));
4083 rtx dest = gen_rtx_REG (DFmode, i);
4084 emit_move_insn (dest, src);
4088 /* Emit a blockage insn here to keep these insns from being moved to
4089 an earlier spot in the epilogue, or into the main instruction stream.
4091 This is necessary as we must not cut the stack back before all the
4092 restores are finished. */
4093 emit_insn (gen_blockage ());
4095 /* Reset stack pointer (and possibly frame pointer). The stack
4096 pointer is initially set to fp + 64 to avoid a race condition. */
4097 if (frame_pointer_needed)
4099 rtx delta = GEN_INT (-64);
4101 set_reg_plus_d (STACK_POINTER_REGNUM, FRAME_POINTER_REGNUM, 64, 0);
4102 emit_insn (gen_pre_load (frame_pointer_rtx, stack_pointer_rtx, delta));
4104 /* If we were deferring a callee register restore, do it now. */
4105 else if (merge_sp_adjust_with_load)
4107 rtx delta = GEN_INT (-actual_fsize);
4108 rtx dest = gen_rtx_REG (word_mode, merge_sp_adjust_with_load);
4110 emit_insn (gen_pre_load (dest, stack_pointer_rtx, delta));
4112 else if (actual_fsize != 0)
4113 set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM,
4116 /* If we haven't restored %r2 yet (no frame pointer, and a stack
4117 frame greater than 8k), do so now. */
4119 load_reg (2, ret_off, STACK_POINTER_REGNUM);
4121 if (DO_FRAME_NOTES && current_function_calls_eh_return)
4123 rtx sa = EH_RETURN_STACKADJ_RTX;
4125 emit_insn (gen_blockage ());
4126 emit_insn (TARGET_64BIT
4127 ? gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, sa)
4128 : gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, sa));
4133 hppa_pic_save_rtx (void)
4135 return get_hard_reg_initial_val (word_mode, PIC_OFFSET_TABLE_REGNUM);
4138 #ifndef NO_DEFERRED_PROFILE_COUNTERS
4139 #define NO_DEFERRED_PROFILE_COUNTERS 0
4142 /* Define heap vector type for funcdef numbers. */
4144 DEF_VEC_ALLOC_I(int,heap);
4146 /* Vector of funcdef numbers. */
4147 static VEC(int,heap) *funcdef_nos;
4149 /* Output deferred profile counters. */
4151 output_deferred_profile_counters (void)
4156 if (VEC_empty (int, funcdef_nos))
4159 switch_to_section (data_section);
4160 align = MIN (BIGGEST_ALIGNMENT, LONG_TYPE_SIZE);
4161 ASM_OUTPUT_ALIGN (asm_out_file, floor_log2 (align / BITS_PER_UNIT));
4163 for (i = 0; VEC_iterate (int, funcdef_nos, i, n); i++)
4165 targetm.asm_out.internal_label (asm_out_file, "LP", n);
4166 assemble_integer (const0_rtx, LONG_TYPE_SIZE / BITS_PER_UNIT, align, 1);
4169 VEC_free (int, heap, funcdef_nos);
4173 hppa_profile_hook (int label_no)
4175 /* We use SImode for the address of the function in both 32 and
4176 64-bit code to avoid having to provide DImode versions of the
4177 lcla2 and load_offset_label_address insn patterns. */
4178 rtx reg = gen_reg_rtx (SImode);
4179 rtx label_rtx = gen_label_rtx ();
4180 rtx begin_label_rtx, call_insn;
4181 char begin_label_name[16];
4183 ASM_GENERATE_INTERNAL_LABEL (begin_label_name, FUNC_BEGIN_PROLOG_LABEL,
4185 begin_label_rtx = gen_rtx_SYMBOL_REF (SImode, ggc_strdup (begin_label_name));
4188 emit_move_insn (arg_pointer_rtx,
4189 gen_rtx_PLUS (word_mode, virtual_outgoing_args_rtx,
4192 emit_move_insn (gen_rtx_REG (word_mode, 26), gen_rtx_REG (word_mode, 2));
4194 /* The address of the function is loaded into %r25 with a instruction-
4195 relative sequence that avoids the use of relocations. The sequence
4196 is split so that the load_offset_label_address instruction can
4197 occupy the delay slot of the call to _mcount. */
4199 emit_insn (gen_lcla2 (reg, label_rtx));
4201 emit_insn (gen_lcla1 (reg, label_rtx));
4203 emit_insn (gen_load_offset_label_address (gen_rtx_REG (SImode, 25),
4204 reg, begin_label_rtx, label_rtx));
4206 #if !NO_DEFERRED_PROFILE_COUNTERS
4208 rtx count_label_rtx, addr, r24;
4209 char count_label_name[16];
4211 VEC_safe_push (int, heap, funcdef_nos, label_no);
4212 ASM_GENERATE_INTERNAL_LABEL (count_label_name, "LP", label_no);
4213 count_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (count_label_name));
4215 addr = force_reg (Pmode, count_label_rtx);
4216 r24 = gen_rtx_REG (Pmode, 24);
4217 emit_move_insn (r24, addr);
4220 emit_call_insn (gen_call (gen_rtx_MEM (Pmode,
4221 gen_rtx_SYMBOL_REF (Pmode,
4223 GEN_INT (TARGET_64BIT ? 24 : 12)));
4225 use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), r24);
4230 emit_call_insn (gen_call (gen_rtx_MEM (Pmode,
4231 gen_rtx_SYMBOL_REF (Pmode,
4233 GEN_INT (TARGET_64BIT ? 16 : 8)));
4237 use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 25));
4238 use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 26));
4240 /* Indicate the _mcount call cannot throw, nor will it execute a
4242 REG_NOTES (call_insn)
4243 = gen_rtx_EXPR_LIST (REG_EH_REGION, constm1_rtx, REG_NOTES (call_insn));
4246 /* Fetch the return address for the frame COUNT steps up from
4247 the current frame, after the prologue. FRAMEADDR is the
4248 frame pointer of the COUNT frame.
4250 We want to ignore any export stub remnants here. To handle this,
4251 we examine the code at the return address, and if it is an export
4252 stub, we return a memory rtx for the stub return address stored
4255 The value returned is used in two different ways:
4257 1. To find a function's caller.
4259 2. To change the return address for a function.
4261 This function handles most instances of case 1; however, it will
4262 fail if there are two levels of stubs to execute on the return
4263 path. The only way I believe that can happen is if the return value
4264 needs a parameter relocation, which never happens for C code.
4266 This function handles most instances of case 2; however, it will
4267 fail if we did not originally have stub code on the return path
4268 but will need stub code on the new return path. This can happen if
4269 the caller & callee are both in the main program, but the new
4270 return location is in a shared library. */
4273 return_addr_rtx (int count, rtx frameaddr)
4283 rp = get_hard_reg_initial_val (Pmode, 2);
4285 if (TARGET_64BIT || TARGET_NO_SPACE_REGS)
4288 saved_rp = gen_reg_rtx (Pmode);
4289 emit_move_insn (saved_rp, rp);
4291 /* Get pointer to the instruction stream. We have to mask out the
4292 privilege level from the two low order bits of the return address
4293 pointer here so that ins will point to the start of the first
4294 instruction that would have been executed if we returned. */
4295 ins = copy_to_reg (gen_rtx_AND (Pmode, rp, MASK_RETURN_ADDR));
4296 label = gen_label_rtx ();
4298 /* Check the instruction stream at the normal return address for the
4301 0x4bc23fd1 | stub+8: ldw -18(sr0,sp),rp
4302 0x004010a1 | stub+12: ldsid (sr0,rp),r1
4303 0x00011820 | stub+16: mtsp r1,sr0
4304 0xe0400002 | stub+20: be,n 0(sr0,rp)
4306 If it is an export stub, than our return address is really in
4309 emit_cmp_insn (gen_rtx_MEM (SImode, ins), GEN_INT (0x4bc23fd1), NE,
4310 NULL_RTX, SImode, 1);
4311 emit_jump_insn (gen_bne (label));
4313 emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 4)),
4314 GEN_INT (0x004010a1), NE, NULL_RTX, SImode, 1);
4315 emit_jump_insn (gen_bne (label));
4317 emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 8)),
4318 GEN_INT (0x00011820), NE, NULL_RTX, SImode, 1);
4319 emit_jump_insn (gen_bne (label));
4321 emit_cmp_insn (gen_rtx_MEM (SImode, plus_constant (ins, 12)),
4322 GEN_INT (0xe0400002), NE, NULL_RTX, SImode, 1);
4324 /* If there is no export stub then just use the value saved from
4325 the return pointer register. */
4327 emit_jump_insn (gen_bne (label));
4329 /* Here we know that our return address points to an export
4330 stub. We don't want to return the address of the export stub,
4331 but rather the return address of the export stub. That return
4332 address is stored at -24[frameaddr]. */
4334 emit_move_insn (saved_rp,
4336 memory_address (Pmode,
4337 plus_constant (frameaddr,
4344 /* This is only valid once reload has completed because it depends on
4345 knowing exactly how much (if any) frame there is and...
4347 It's only valid if there is no frame marker to de-allocate and...
4349 It's only valid if %r2 hasn't been saved into the caller's frame
4350 (we're not profiling and %r2 isn't live anywhere). */
4352 hppa_can_use_return_insn_p (void)
4354 return (reload_completed
4355 && (compute_frame_size (get_frame_size (), 0) ? 0 : 1)
4356 && ! regs_ever_live[2]
4357 && ! frame_pointer_needed);
4361 emit_bcond_fp (enum rtx_code code, rtx operand0)
4363 emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
4364 gen_rtx_IF_THEN_ELSE (VOIDmode,
4365 gen_rtx_fmt_ee (code,
4367 gen_rtx_REG (CCFPmode, 0),
4369 gen_rtx_LABEL_REF (VOIDmode, operand0),
4375 gen_cmp_fp (enum rtx_code code, rtx operand0, rtx operand1)
4377 return gen_rtx_SET (VOIDmode, gen_rtx_REG (CCFPmode, 0),
4378 gen_rtx_fmt_ee (code, CCFPmode, operand0, operand1));
4381 /* Adjust the cost of a scheduling dependency. Return the new cost of
4382 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
4385 pa_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
4387 enum attr_type attr_type;
4389 /* Don't adjust costs for a pa8000 chip, also do not adjust any
4390 true dependencies as they are described with bypasses now. */
4391 if (pa_cpu >= PROCESSOR_8000 || REG_NOTE_KIND (link) == 0)
4394 if (! recog_memoized (insn))
4397 attr_type = get_attr_type (insn);
4399 switch (REG_NOTE_KIND (link))
4402 /* Anti dependency; DEP_INSN reads a register that INSN writes some
4405 if (attr_type == TYPE_FPLOAD)
4407 rtx pat = PATTERN (insn);
4408 rtx dep_pat = PATTERN (dep_insn);
4409 if (GET_CODE (pat) == PARALLEL)
4411 /* This happens for the fldXs,mb patterns. */
4412 pat = XVECEXP (pat, 0, 0);
4414 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
4415 /* If this happens, we have to extend this to schedule
4416 optimally. Return 0 for now. */
4419 if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat)))
4421 if (! recog_memoized (dep_insn))
4423 switch (get_attr_type (dep_insn))
4430 case TYPE_FPSQRTSGL:
4431 case TYPE_FPSQRTDBL:
4432 /* A fpload can't be issued until one cycle before a
4433 preceding arithmetic operation has finished if
4434 the target of the fpload is any of the sources
4435 (or destination) of the arithmetic operation. */
4436 return insn_default_latency (dep_insn) - 1;
4443 else if (attr_type == TYPE_FPALU)
4445 rtx pat = PATTERN (insn);
4446 rtx dep_pat = PATTERN (dep_insn);
4447 if (GET_CODE (pat) == PARALLEL)
4449 /* This happens for the fldXs,mb patterns. */
4450 pat = XVECEXP (pat, 0, 0);
4452 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
4453 /* If this happens, we have to extend this to schedule
4454 optimally. Return 0 for now. */
4457 if (reg_mentioned_p (SET_DEST (pat), SET_SRC (dep_pat)))
4459 if (! recog_memoized (dep_insn))
4461 switch (get_attr_type (dep_insn))
4465 case TYPE_FPSQRTSGL:
4466 case TYPE_FPSQRTDBL:
4467 /* An ALU flop can't be issued until two cycles before a
4468 preceding divide or sqrt operation has finished if
4469 the target of the ALU flop is any of the sources
4470 (or destination) of the divide or sqrt operation. */
4471 return insn_default_latency (dep_insn) - 2;
4479 /* For other anti dependencies, the cost is 0. */
4482 case REG_DEP_OUTPUT:
4483 /* Output dependency; DEP_INSN writes a register that INSN writes some
4485 if (attr_type == TYPE_FPLOAD)
4487 rtx pat = PATTERN (insn);
4488 rtx dep_pat = PATTERN (dep_insn);
4489 if (GET_CODE (pat) == PARALLEL)
4491 /* This happens for the fldXs,mb patterns. */
4492 pat = XVECEXP (pat, 0, 0);
4494 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
4495 /* If this happens, we have to extend this to schedule
4496 optimally. Return 0 for now. */
4499 if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat)))
4501 if (! recog_memoized (dep_insn))
4503 switch (get_attr_type (dep_insn))
4510 case TYPE_FPSQRTSGL:
4511 case TYPE_FPSQRTDBL:
4512 /* A fpload can't be issued until one cycle before a
4513 preceding arithmetic operation has finished if
4514 the target of the fpload is the destination of the
4515 arithmetic operation.
4517 Exception: For PA7100LC, PA7200 and PA7300, the cost
4518 is 3 cycles, unless they bundle together. We also
4519 pay the penalty if the second insn is a fpload. */
4520 return insn_default_latency (dep_insn) - 1;
4527 else if (attr_type == TYPE_FPALU)
4529 rtx pat = PATTERN (insn);
4530 rtx dep_pat = PATTERN (dep_insn);
4531 if (GET_CODE (pat) == PARALLEL)
4533 /* This happens for the fldXs,mb patterns. */
4534 pat = XVECEXP (pat, 0, 0);
4536 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
4537 /* If this happens, we have to extend this to schedule
4538 optimally. Return 0 for now. */
4541 if (reg_mentioned_p (SET_DEST (pat), SET_DEST (dep_pat)))
4543 if (! recog_memoized (dep_insn))
4545 switch (get_attr_type (dep_insn))
4549 case TYPE_FPSQRTSGL:
4550 case TYPE_FPSQRTDBL:
4551 /* An ALU flop can't be issued until two cycles before a
4552 preceding divide or sqrt operation has finished if
4553 the target of the ALU flop is also the target of
4554 the divide or sqrt operation. */
4555 return insn_default_latency (dep_insn) - 2;
4563 /* For other output dependencies, the cost is 0. */
4571 /* Adjust scheduling priorities. We use this to try and keep addil
4572 and the next use of %r1 close together. */
4574 pa_adjust_priority (rtx insn, int priority)
4576 rtx set = single_set (insn);
4580 src = SET_SRC (set);
4581 dest = SET_DEST (set);
4582 if (GET_CODE (src) == LO_SUM
4583 && symbolic_operand (XEXP (src, 1), VOIDmode)
4584 && ! read_only_operand (XEXP (src, 1), VOIDmode))
4587 else if (GET_CODE (src) == MEM
4588 && GET_CODE (XEXP (src, 0)) == LO_SUM
4589 && symbolic_operand (XEXP (XEXP (src, 0), 1), VOIDmode)
4590 && ! read_only_operand (XEXP (XEXP (src, 0), 1), VOIDmode))
4593 else if (GET_CODE (dest) == MEM
4594 && GET_CODE (XEXP (dest, 0)) == LO_SUM
4595 && symbolic_operand (XEXP (XEXP (dest, 0), 1), VOIDmode)
4596 && ! read_only_operand (XEXP (XEXP (dest, 0), 1), VOIDmode))
4602 /* The 700 can only issue a single insn at a time.
4603 The 7XXX processors can issue two insns at a time.
4604 The 8000 can issue 4 insns at a time. */
4606 pa_issue_rate (void)
4610 case PROCESSOR_700: return 1;
4611 case PROCESSOR_7100: return 2;
4612 case PROCESSOR_7100LC: return 2;
4613 case PROCESSOR_7200: return 2;
4614 case PROCESSOR_7300: return 2;
4615 case PROCESSOR_8000: return 4;
4624 /* Return any length adjustment needed by INSN which already has its length
4625 computed as LENGTH. Return zero if no adjustment is necessary.
4627 For the PA: function calls, millicode calls, and backwards short
4628 conditional branches with unfilled delay slots need an adjustment by +1
4629 (to account for the NOP which will be inserted into the instruction stream).
4631 Also compute the length of an inline block move here as it is too
4632 complicated to express as a length attribute in pa.md. */
4634 pa_adjust_insn_length (rtx insn, int length)
4636 rtx pat = PATTERN (insn);
4638 /* Jumps inside switch tables which have unfilled delay slots need
4640 if (GET_CODE (insn) == JUMP_INSN
4641 && GET_CODE (pat) == PARALLEL
4642 && get_attr_type (insn) == TYPE_BTABLE_BRANCH)
4644 /* Millicode insn with an unfilled delay slot. */
4645 else if (GET_CODE (insn) == INSN
4646 && GET_CODE (pat) != SEQUENCE
4647 && GET_CODE (pat) != USE
4648 && GET_CODE (pat) != CLOBBER
4649 && get_attr_type (insn) == TYPE_MILLI)
4651 /* Block move pattern. */
4652 else if (GET_CODE (insn) == INSN
4653 && GET_CODE (pat) == PARALLEL
4654 && GET_CODE (XVECEXP (pat, 0, 0)) == SET
4655 && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM
4656 && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 1)) == MEM
4657 && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode
4658 && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 1)) == BLKmode)
4659 return compute_movmem_length (insn) - 4;
4660 /* Block clear pattern. */
4661 else if (GET_CODE (insn) == INSN
4662 && GET_CODE (pat) == PARALLEL
4663 && GET_CODE (XVECEXP (pat, 0, 0)) == SET
4664 && GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM
4665 && XEXP (XVECEXP (pat, 0, 0), 1) == const0_rtx
4666 && GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode)
4667 return compute_clrmem_length (insn) - 4;
4668 /* Conditional branch with an unfilled delay slot. */
4669 else if (GET_CODE (insn) == JUMP_INSN && ! simplejump_p (insn))
4671 /* Adjust a short backwards conditional with an unfilled delay slot. */
4672 if (GET_CODE (pat) == SET
4674 && ! forward_branch_p (insn))
4676 else if (GET_CODE (pat) == PARALLEL
4677 && get_attr_type (insn) == TYPE_PARALLEL_BRANCH
4680 /* Adjust dbra insn with short backwards conditional branch with
4681 unfilled delay slot -- only for case where counter is in a
4682 general register register. */
4683 else if (GET_CODE (pat) == PARALLEL
4684 && GET_CODE (XVECEXP (pat, 0, 1)) == SET
4685 && GET_CODE (XEXP (XVECEXP (pat, 0, 1), 0)) == REG
4686 && ! FP_REG_P (XEXP (XVECEXP (pat, 0, 1), 0))
4688 && ! forward_branch_p (insn))
4696 /* Print operand X (an rtx) in assembler syntax to file FILE.
4697 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
4698 For `%' followed by punctuation, CODE is the punctuation and X is null. */
4701 print_operand (FILE *file, rtx x, int code)
4706 /* Output a 'nop' if there's nothing for the delay slot. */
4707 if (dbr_sequence_length () == 0)
4708 fputs ("\n\tnop", file);
4711 /* Output a nullification completer if there's nothing for the */
4712 /* delay slot or nullification is requested. */
4713 if (dbr_sequence_length () == 0 ||
4715 INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0))))
4719 /* Print out the second register name of a register pair.
4720 I.e., R (6) => 7. */
4721 fputs (reg_names[REGNO (x) + 1], file);
4724 /* A register or zero. */
4726 || (x == CONST0_RTX (DFmode))
4727 || (x == CONST0_RTX (SFmode)))
4729 fputs ("%r0", file);
4735 /* A register or zero (floating point). */
4737 || (x == CONST0_RTX (DFmode))
4738 || (x == CONST0_RTX (SFmode)))
4740 fputs ("%fr0", file);
4749 xoperands[0] = XEXP (XEXP (x, 0), 0);
4750 xoperands[1] = XVECEXP (XEXP (XEXP (x, 0), 1), 0, 0);
4751 output_global_address (file, xoperands[1], 0);
4752 fprintf (file, "(%s)", reg_names [REGNO (xoperands[0])]);
4756 case 'C': /* Plain (C)ondition */
4758 switch (GET_CODE (x))
4761 fputs ("=", file); break;
4763 fputs ("<>", file); break;
4765 fputs (">", file); break;
4767 fputs (">=", file); break;
4769 fputs (">>=", file); break;
4771 fputs (">>", file); break;
4773 fputs ("<", file); break;
4775 fputs ("<=", file); break;
4777 fputs ("<<=", file); break;
4779 fputs ("<<", file); break;
4784 case 'N': /* Condition, (N)egated */
4785 switch (GET_CODE (x))
4788 fputs ("<>", file); break;
4790 fputs ("=", file); break;
4792 fputs ("<=", file); break;
4794 fputs ("<", file); break;
4796 fputs ("<<", file); break;
4798 fputs ("<<=", file); break;
4800 fputs (">=", file); break;
4802 fputs (">", file); break;
4804 fputs (">>", file); break;
4806 fputs (">>=", file); break;
4811 /* For floating point comparisons. Note that the output
4812 predicates are the complement of the desired mode. The
4813 conditions for GT, GE, LT, LE and LTGT cause an invalid
4814 operation exception if the result is unordered and this
4815 exception is enabled in the floating-point status register. */
4817 switch (GET_CODE (x))
4820 fputs ("!=", file); break;
4822 fputs ("=", file); break;
4824 fputs ("!>", file); break;
4826 fputs ("!>=", file); break;
4828 fputs ("!<", file); break;
4830 fputs ("!<=", file); break;
4832 fputs ("!<>", file); break;
4834 fputs ("!?<=", file); break;
4836 fputs ("!?<", file); break;
4838 fputs ("!?>=", file); break;
4840 fputs ("!?>", file); break;
4842 fputs ("!?=", file); break;
4844 fputs ("!?", file); break;
4846 fputs ("?", file); break;
4851 case 'S': /* Condition, operands are (S)wapped. */
4852 switch (GET_CODE (x))
4855 fputs ("=", file); break;
4857 fputs ("<>", file); break;
4859 fputs ("<", file); break;
4861 fputs ("<=", file); break;
4863 fputs ("<<=", file); break;
4865 fputs ("<<", file); break;
4867 fputs (">", file); break;
4869 fputs (">=", file); break;
4871 fputs (">>=", file); break;
4873 fputs (">>", file); break;
4878 case 'B': /* Condition, (B)oth swapped and negate. */
4879 switch (GET_CODE (x))
4882 fputs ("<>", file); break;
4884 fputs ("=", file); break;
4886 fputs (">=", file); break;
4888 fputs (">", file); break;
4890 fputs (">>", file); break;
4892 fputs (">>=", file); break;
4894 fputs ("<=", file); break;
4896 fputs ("<", file); break;
4898 fputs ("<<", file); break;
4900 fputs ("<<=", file); break;
4906 gcc_assert (GET_CODE (x) == CONST_INT);
4907 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~INTVAL (x));
4910 gcc_assert (GET_CODE (x) == CONST_INT);
4911 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - (INTVAL (x) & 63));
4914 gcc_assert (GET_CODE (x) == CONST_INT);
4915 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - (INTVAL (x) & 31));
4918 gcc_assert (GET_CODE (x) == CONST_INT && exact_log2 (INTVAL (x)) >= 0);
4919 fprintf (file, "%d", exact_log2 (INTVAL (x)));
4922 gcc_assert (GET_CODE (x) == CONST_INT);
4923 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 63 - (INTVAL (x) & 63));
4926 gcc_assert (GET_CODE (x) == CONST_INT);
4927 fprintf (file, HOST_WIDE_INT_PRINT_DEC, 31 - (INTVAL (x) & 31));
4930 if (GET_CODE (x) == CONST_INT)
4935 switch (GET_CODE (XEXP (x, 0)))
4939 if (ASSEMBLER_DIALECT == 0)
4940 fputs ("s,mb", file);
4942 fputs (",mb", file);
4946 if (ASSEMBLER_DIALECT == 0)
4947 fputs ("s,ma", file);
4949 fputs (",ma", file);
4952 if (GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
4953 && GET_CODE (XEXP (XEXP (x, 0), 1)) == REG)
4955 if (ASSEMBLER_DIALECT == 0)
4958 else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
4959 || GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
4961 if (ASSEMBLER_DIALECT == 0)
4962 fputs ("x,s", file);
4966 else if (code == 'F' && ASSEMBLER_DIALECT == 0)
4970 if (code == 'F' && ASSEMBLER_DIALECT == 0)
4976 output_global_address (file, x, 0);
4979 output_global_address (file, x, 1);
4981 case 0: /* Don't do anything special */
4986 compute_zdepwi_operands (INTVAL (x), op);
4987 fprintf (file, "%d,%d,%d", op[0], op[1], op[2]);
4993 compute_zdepdi_operands (INTVAL (x), op);
4994 fprintf (file, "%d,%d,%d", op[0], op[1], op[2]);
4998 /* We can get here from a .vtable_inherit due to our
4999 CONSTANT_ADDRESS_P rejecting perfectly good constant
5005 if (GET_CODE (x) == REG)
5007 fputs (reg_names [REGNO (x)], file);
5008 if (TARGET_64BIT && FP_REG_P (x) && GET_MODE_SIZE (GET_MODE (x)) <= 4)
5014 && GET_MODE_SIZE (GET_MODE (x)) <= 4
5015 && (REGNO (x) & 1) == 0)
5018 else if (GET_CODE (x) == MEM)
5020 int size = GET_MODE_SIZE (GET_MODE (x));
5021 rtx base = NULL_RTX;
5022 switch (GET_CODE (XEXP (x, 0)))
5026 base = XEXP (XEXP (x, 0), 0);
5027 fprintf (file, "-%d(%s)", size, reg_names [REGNO (base)]);
5031 base = XEXP (XEXP (x, 0), 0);
5032 fprintf (file, "%d(%s)", size, reg_names [REGNO (base)]);
5035 if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT)
5036 fprintf (file, "%s(%s)",
5037 reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 0), 0))],
5038 reg_names [REGNO (XEXP (XEXP (x, 0), 1))]);
5039 else if (GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
5040 fprintf (file, "%s(%s)",
5041 reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 1), 0))],
5042 reg_names [REGNO (XEXP (XEXP (x, 0), 0))]);
5043 else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
5044 && GET_CODE (XEXP (XEXP (x, 0), 1)) == REG)
5046 /* Because the REG_POINTER flag can get lost during reload,
5047 GO_IF_LEGITIMATE_ADDRESS canonicalizes the order of the
5048 index and base registers in the combined move patterns. */
5049 rtx base = XEXP (XEXP (x, 0), 1);
5050 rtx index = XEXP (XEXP (x, 0), 0);
5052 fprintf (file, "%s(%s)",
5053 reg_names [REGNO (index)], reg_names [REGNO (base)]);
5056 output_address (XEXP (x, 0));
5059 output_address (XEXP (x, 0));
5064 output_addr_const (file, x);
5067 /* output a SYMBOL_REF or a CONST expression involving a SYMBOL_REF. */
5070 output_global_address (FILE *file, rtx x, int round_constant)
5073 /* Imagine (high (const (plus ...))). */
5074 if (GET_CODE (x) == HIGH)
5077 if (GET_CODE (x) == SYMBOL_REF && read_only_operand (x, VOIDmode))
5078 output_addr_const (file, x);
5079 else if (GET_CODE (x) == SYMBOL_REF && !flag_pic)
5081 output_addr_const (file, x);
5082 fputs ("-$global$", file);
5084 else if (GET_CODE (x) == CONST)
5086 const char *sep = "";
5087 int offset = 0; /* assembler wants -$global$ at end */
5088 rtx base = NULL_RTX;
5090 switch (GET_CODE (XEXP (XEXP (x, 0), 0)))
5093 base = XEXP (XEXP (x, 0), 0);
5094 output_addr_const (file, base);
5097 offset = INTVAL (XEXP (XEXP (x, 0), 0));
5103 switch (GET_CODE (XEXP (XEXP (x, 0), 1)))
5106 base = XEXP (XEXP (x, 0), 1);
5107 output_addr_const (file, base);
5110 offset = INTVAL (XEXP (XEXP (x, 0), 1));
5116 /* How bogus. The compiler is apparently responsible for
5117 rounding the constant if it uses an LR field selector.
5119 The linker and/or assembler seem a better place since
5120 they have to do this kind of thing already.
5122 If we fail to do this, HP's optimizing linker may eliminate
5123 an addil, but not update the ldw/stw/ldo instruction that
5124 uses the result of the addil. */
5126 offset = ((offset + 0x1000) & ~0x1fff);
5128 switch (GET_CODE (XEXP (x, 0)))
5141 gcc_assert (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF);
5149 if (!read_only_operand (base, VOIDmode) && !flag_pic)
5150 fputs ("-$global$", file);
5152 fprintf (file, "%s%d", sep, offset);
5155 output_addr_const (file, x);
5158 /* Output boilerplate text to appear at the beginning of the file.
5159 There are several possible versions. */
5160 #define aputs(x) fputs(x, asm_out_file)
5162 pa_file_start_level (void)
5165 aputs ("\t.LEVEL 2.0w\n");
5166 else if (TARGET_PA_20)
5167 aputs ("\t.LEVEL 2.0\n");
5168 else if (TARGET_PA_11)
5169 aputs ("\t.LEVEL 1.1\n");
5171 aputs ("\t.LEVEL 1.0\n");
5175 pa_file_start_space (int sortspace)
5177 aputs ("\t.SPACE $PRIVATE$");
5180 aputs ("\n\t.SUBSPA $DATA$,QUAD=1,ALIGN=8,ACCESS=31"
5181 "\n\t.SUBSPA $BSS$,QUAD=1,ALIGN=8,ACCESS=31,ZERO,SORT=82"
5182 "\n\t.SPACE $TEXT$");
5185 aputs ("\n\t.SUBSPA $LIT$,QUAD=0,ALIGN=8,ACCESS=44"
5186 "\n\t.SUBSPA $CODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY\n");
5190 pa_file_start_file (int want_version)
5192 if (write_symbols != NO_DEBUG)
5194 output_file_directive (asm_out_file, main_input_filename);
5196 aputs ("\t.version\t\"01.01\"\n");
5201 pa_file_start_mcount (const char *aswhat)
5204 fprintf (asm_out_file, "\t.IMPORT _mcount,%s\n", aswhat);
5208 pa_elf_file_start (void)
5210 pa_file_start_level ();
5211 pa_file_start_mcount ("ENTRY");
5212 pa_file_start_file (0);
5216 pa_som_file_start (void)
5218 pa_file_start_level ();
5219 pa_file_start_space (0);
5220 aputs ("\t.IMPORT $global$,DATA\n"
5221 "\t.IMPORT $$dyncall,MILLICODE\n");
5222 pa_file_start_mcount ("CODE");
5223 pa_file_start_file (0);
5227 pa_linux_file_start (void)
5229 pa_file_start_file (1);
5230 pa_file_start_level ();
5231 pa_file_start_mcount ("CODE");
5235 pa_hpux64_gas_file_start (void)
5237 pa_file_start_level ();
5238 #ifdef ASM_OUTPUT_TYPE_DIRECTIVE
5240 ASM_OUTPUT_TYPE_DIRECTIVE (asm_out_file, "_mcount", "function");
5242 pa_file_start_file (1);
5246 pa_hpux64_hpas_file_start (void)
5248 pa_file_start_level ();
5249 pa_file_start_space (1);
5250 pa_file_start_mcount ("CODE");
5251 pa_file_start_file (0);
5255 /* Search the deferred plabel list for SYMBOL and return its internal
5256 label. If an entry for SYMBOL is not found, a new entry is created. */
5259 get_deferred_plabel (rtx symbol)
5261 const char *fname = XSTR (symbol, 0);
5264 /* See if we have already put this function on the list of deferred
5265 plabels. This list is generally small, so a liner search is not
5266 too ugly. If it proves too slow replace it with something faster. */
5267 for (i = 0; i < n_deferred_plabels; i++)
5268 if (strcmp (fname, XSTR (deferred_plabels[i].symbol, 0)) == 0)
5271 /* If the deferred plabel list is empty, or this entry was not found
5272 on the list, create a new entry on the list. */
5273 if (deferred_plabels == NULL || i == n_deferred_plabels)
5277 if (deferred_plabels == 0)
5278 deferred_plabels = (struct deferred_plabel *)
5279 ggc_alloc (sizeof (struct deferred_plabel));
5281 deferred_plabels = (struct deferred_plabel *)
5282 ggc_realloc (deferred_plabels,
5283 ((n_deferred_plabels + 1)
5284 * sizeof (struct deferred_plabel)));
5286 i = n_deferred_plabels++;
5287 deferred_plabels[i].internal_label = gen_label_rtx ();
5288 deferred_plabels[i].symbol = symbol;
5290 /* Gross. We have just implicitly taken the address of this
5291 function. Mark it in the same manner as assemble_name. */
5292 id = maybe_get_identifier (targetm.strip_name_encoding (fname));
5294 mark_referenced (id);
5297 return deferred_plabels[i].internal_label;
5301 output_deferred_plabels (void)
5305 /* If we have some deferred plabels, then we need to switch into the
5306 data or readonly data section, and align it to a 4 byte boundary
5307 before outputting the deferred plabels. */
5308 if (n_deferred_plabels)
5310 switch_to_section (flag_pic ? data_section : readonly_data_section);
5311 ASM_OUTPUT_ALIGN (asm_out_file, TARGET_64BIT ? 3 : 2);
5314 /* Now output the deferred plabels. */
5315 for (i = 0; i < n_deferred_plabels; i++)
5317 (*targetm.asm_out.internal_label) (asm_out_file, "L",
5318 CODE_LABEL_NUMBER (deferred_plabels[i].internal_label));
5319 assemble_integer (deferred_plabels[i].symbol,
5320 TARGET_64BIT ? 8 : 4, TARGET_64BIT ? 64 : 32, 1);
5324 #ifdef HPUX_LONG_DOUBLE_LIBRARY
5325 /* Initialize optabs to point to HPUX long double emulation routines. */
5327 pa_hpux_init_libfuncs (void)
5329 set_optab_libfunc (add_optab, TFmode, "_U_Qfadd");
5330 set_optab_libfunc (sub_optab, TFmode, "_U_Qfsub");
5331 set_optab_libfunc (smul_optab, TFmode, "_U_Qfmpy");
5332 set_optab_libfunc (sdiv_optab, TFmode, "_U_Qfdiv");
5333 set_optab_libfunc (smin_optab, TFmode, "_U_Qmin");
5334 set_optab_libfunc (smax_optab, TFmode, "_U_Qfmax");
5335 set_optab_libfunc (sqrt_optab, TFmode, "_U_Qfsqrt");
5336 set_optab_libfunc (abs_optab, TFmode, "_U_Qfabs");
5337 set_optab_libfunc (neg_optab, TFmode, "_U_Qfneg");
5339 set_optab_libfunc (eq_optab, TFmode, "_U_Qfeq");
5340 set_optab_libfunc (ne_optab, TFmode, "_U_Qfne");
5341 set_optab_libfunc (gt_optab, TFmode, "_U_Qfgt");
5342 set_optab_libfunc (ge_optab, TFmode, "_U_Qfge");
5343 set_optab_libfunc (lt_optab, TFmode, "_U_Qflt");
5344 set_optab_libfunc (le_optab, TFmode, "_U_Qfle");
5345 set_optab_libfunc (unord_optab, TFmode, "_U_Qfunord");
5347 set_conv_libfunc (sext_optab, TFmode, SFmode, "_U_Qfcnvff_sgl_to_quad");
5348 set_conv_libfunc (sext_optab, TFmode, DFmode, "_U_Qfcnvff_dbl_to_quad");
5349 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_U_Qfcnvff_quad_to_sgl");
5350 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_U_Qfcnvff_quad_to_dbl");
5352 set_conv_libfunc (sfix_optab, SImode, TFmode, TARGET_64BIT
5353 ? "__U_Qfcnvfxt_quad_to_sgl"
5354 : "_U_Qfcnvfxt_quad_to_sgl");
5355 set_conv_libfunc (sfix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_dbl");
5356 set_conv_libfunc (ufix_optab, SImode, TFmode, "_U_Qfcnvfxt_quad_to_usgl");
5357 set_conv_libfunc (ufix_optab, DImode, TFmode, "_U_Qfcnvfxt_quad_to_udbl");
5359 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_U_Qfcnvxf_sgl_to_quad");
5360 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_U_Qfcnvxf_dbl_to_quad");
5361 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_U_Qfcnvxf_usgl_to_quad");
5362 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_U_Qfcnvxf_udbl_to_quad");
5366 /* HP's millicode routines mean something special to the assembler.
5367 Keep track of which ones we have used. */
5369 enum millicodes { remI, remU, divI, divU, mulI, end1000 };
5370 static void import_milli (enum millicodes);
5371 static char imported[(int) end1000];
5372 static const char * const milli_names[] = {"remI", "remU", "divI", "divU", "mulI"};
5373 static const char import_string[] = ".IMPORT $$....,MILLICODE";
5374 #define MILLI_START 10
5377 import_milli (enum millicodes code)
5379 char str[sizeof (import_string)];
5381 if (!imported[(int) code])
5383 imported[(int) code] = 1;
5384 strcpy (str, import_string);
5385 strncpy (str + MILLI_START, milli_names[(int) code], 4);
5386 output_asm_insn (str, 0);
5390 /* The register constraints have put the operands and return value in
5391 the proper registers. */
5394 output_mul_insn (int unsignedp ATTRIBUTE_UNUSED, rtx insn)
5396 import_milli (mulI);
5397 return output_millicode_call (insn, gen_rtx_SYMBOL_REF (Pmode, "$$mulI"));
5400 /* Emit the rtl for doing a division by a constant. */
5402 /* Do magic division millicodes exist for this value? */
5403 const int magic_milli[]= {0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, 1, 1};
5405 /* We'll use an array to keep track of the magic millicodes and
5406 whether or not we've used them already. [n][0] is signed, [n][1] is
5409 static int div_milli[16][2];
5412 emit_hpdiv_const (rtx *operands, int unsignedp)
5414 if (GET_CODE (operands[2]) == CONST_INT
5415 && INTVAL (operands[2]) > 0
5416 && INTVAL (operands[2]) < 16
5417 && magic_milli[INTVAL (operands[2])])
5419 rtx ret = gen_rtx_REG (SImode, TARGET_64BIT ? 2 : 31);
5421 emit_move_insn (gen_rtx_REG (SImode, 26), operands[1]);
5425 gen_rtvec (6, gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, 29),
5426 gen_rtx_fmt_ee (unsignedp ? UDIV : DIV,
5428 gen_rtx_REG (SImode, 26),
5430 gen_rtx_CLOBBER (VOIDmode, operands[4]),
5431 gen_rtx_CLOBBER (VOIDmode, operands[3]),
5432 gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 26)),
5433 gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 25)),
5434 gen_rtx_CLOBBER (VOIDmode, ret))));
5435 emit_move_insn (operands[0], gen_rtx_REG (SImode, 29));
5442 output_div_insn (rtx *operands, int unsignedp, rtx insn)
5446 /* If the divisor is a constant, try to use one of the special
5448 if (GET_CODE (operands[0]) == CONST_INT)
5450 static char buf[100];
5451 divisor = INTVAL (operands[0]);
5452 if (!div_milli[divisor][unsignedp])
5454 div_milli[divisor][unsignedp] = 1;
5456 output_asm_insn (".IMPORT $$divU_%0,MILLICODE", operands);
5458 output_asm_insn (".IMPORT $$divI_%0,MILLICODE", operands);
5462 sprintf (buf, "$$divU_" HOST_WIDE_INT_PRINT_DEC,
5463 INTVAL (operands[0]));
5464 return output_millicode_call (insn,
5465 gen_rtx_SYMBOL_REF (SImode, buf));
5469 sprintf (buf, "$$divI_" HOST_WIDE_INT_PRINT_DEC,
5470 INTVAL (operands[0]));
5471 return output_millicode_call (insn,
5472 gen_rtx_SYMBOL_REF (SImode, buf));
5475 /* Divisor isn't a special constant. */
5480 import_milli (divU);
5481 return output_millicode_call (insn,
5482 gen_rtx_SYMBOL_REF (SImode, "$$divU"));
5486 import_milli (divI);
5487 return output_millicode_call (insn,
5488 gen_rtx_SYMBOL_REF (SImode, "$$divI"));
5493 /* Output a $$rem millicode to do mod. */
5496 output_mod_insn (int unsignedp, rtx insn)
5500 import_milli (remU);
5501 return output_millicode_call (insn,
5502 gen_rtx_SYMBOL_REF (SImode, "$$remU"));
5506 import_milli (remI);
5507 return output_millicode_call (insn,
5508 gen_rtx_SYMBOL_REF (SImode, "$$remI"));
5513 output_arg_descriptor (rtx call_insn)
5515 const char *arg_regs[4];
5516 enum machine_mode arg_mode;
5518 int i, output_flag = 0;
5521 /* We neither need nor want argument location descriptors for the
5522 64bit runtime environment or the ELF32 environment. */
5523 if (TARGET_64BIT || TARGET_ELF32)
5526 for (i = 0; i < 4; i++)
5529 /* Specify explicitly that no argument relocations should take place
5530 if using the portable runtime calling conventions. */
5531 if (TARGET_PORTABLE_RUNTIME)
5533 fputs ("\t.CALL ARGW0=NO,ARGW1=NO,ARGW2=NO,ARGW3=NO,RETVAL=NO\n",
5538 gcc_assert (GET_CODE (call_insn) == CALL_INSN);
5539 for (link = CALL_INSN_FUNCTION_USAGE (call_insn);
5540 link; link = XEXP (link, 1))
5542 rtx use = XEXP (link, 0);
5544 if (! (GET_CODE (use) == USE
5545 && GET_CODE (XEXP (use, 0)) == REG
5546 && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0)))))
5549 arg_mode = GET_MODE (XEXP (use, 0));
5550 regno = REGNO (XEXP (use, 0));
5551 if (regno >= 23 && regno <= 26)
5553 arg_regs[26 - regno] = "GR";
5554 if (arg_mode == DImode)
5555 arg_regs[25 - regno] = "GR";
5557 else if (regno >= 32 && regno <= 39)
5559 if (arg_mode == SFmode)
5560 arg_regs[(regno - 32) / 2] = "FR";
5563 #ifndef HP_FP_ARG_DESCRIPTOR_REVERSED
5564 arg_regs[(regno - 34) / 2] = "FR";
5565 arg_regs[(regno - 34) / 2 + 1] = "FU";
5567 arg_regs[(regno - 34) / 2] = "FU";
5568 arg_regs[(regno - 34) / 2 + 1] = "FR";
5573 fputs ("\t.CALL ", asm_out_file);
5574 for (i = 0; i < 4; i++)
5579 fputc (',', asm_out_file);
5580 fprintf (asm_out_file, "ARGW%d=%s", i, arg_regs[i]);
5583 fputc ('\n', asm_out_file);
5586 static enum reg_class
5587 pa_secondary_reload (bool in_p, rtx x, enum reg_class class,
5588 enum machine_mode mode, secondary_reload_info *sri)
5590 int is_symbolic, regno;
5592 /* Handle the easy stuff first. */
5593 if (class == R1_REGS)
5599 if (class == BASE_REG_CLASS && regno < FIRST_PSEUDO_REGISTER)
5605 /* If we have something like (mem (mem (...)), we can safely assume the
5606 inner MEM will end up in a general register after reloading, so there's
5607 no need for a secondary reload. */
5608 if (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == MEM)
5611 /* Trying to load a constant into a FP register during PIC code
5612 generation requires %r1 as a scratch register. */
5614 && GET_MODE_CLASS (mode) == MODE_INT
5615 && FP_REG_CLASS_P (class)
5616 && (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE))
5618 gcc_assert (mode == SImode || mode == DImode);
5619 sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
5620 : CODE_FOR_reload_indi_r1);
5624 /* Profiling showed the PA port spends about 1.3% of its compilation
5625 time in true_regnum from calls inside pa_secondary_reload_class. */
5626 if (regno >= FIRST_PSEUDO_REGISTER || GET_CODE (x) == SUBREG)
5627 regno = true_regnum (x);
5629 /* Handle out of range displacement for integer mode loads/stores of
5631 if (((regno >= FIRST_PSEUDO_REGISTER || regno == -1)
5632 && GET_MODE_CLASS (mode) == MODE_INT
5633 && FP_REG_CLASS_P (class))
5634 || (class == SHIFT_REGS && (regno <= 0 || regno >= 32)))
5636 sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
5640 /* A SAR<->FP register copy requires a secondary register (GPR) as
5641 well as secondary memory. */
5642 if (regno >= 0 && regno < FIRST_PSEUDO_REGISTER
5643 && ((REGNO_REG_CLASS (regno) == SHIFT_REGS && FP_REG_CLASS_P (class))
5644 || (class == SHIFT_REGS
5645 && FP_REG_CLASS_P (REGNO_REG_CLASS (regno)))))
5647 sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
5651 /* Secondary reloads of symbolic operands require %r1 as a scratch
5652 register when we're generating PIC code and the operand isn't
5654 if (GET_CODE (x) == HIGH)
5657 /* Profiling has showed GCC spends about 2.6% of its compilation
5658 time in symbolic_operand from calls inside pa_secondary_reload_class.
5659 So, we use an inline copy to avoid useless work. */
5660 switch (GET_CODE (x))
5665 is_symbolic = !SYMBOL_REF_TLS_MODEL (x);
5672 is_symbolic = (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
5673 && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
5674 || GET_CODE (XEXP (op, 0)) == LABEL_REF)
5675 && GET_CODE (XEXP (op, 1)) == CONST_INT);
5682 if (is_symbolic && (flag_pic || !read_only_operand (x, VOIDmode)))
5684 gcc_assert (mode == SImode || mode == DImode);
5685 sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
5686 : CODE_FOR_reload_indi_r1);
5692 /* In the 32-bit runtime, arguments larger than eight bytes are passed
5693 by invisible reference. As a GCC extension, we also pass anything
5694 with a zero or variable size by reference.
5696 The 64-bit runtime does not describe passing any types by invisible
5697 reference. The internals of GCC can't currently handle passing
5698 empty structures, and zero or variable length arrays when they are
5699 not passed entirely on the stack or by reference. Thus, as a GCC
5700 extension, we pass these types by reference. The HP compiler doesn't
5701 support these types, so hopefully there shouldn't be any compatibility
5702 issues. This may have to be revisited when HP releases a C99 compiler
5703 or updates the ABI. */
5706 pa_pass_by_reference (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED,
5707 enum machine_mode mode, tree type,
5708 bool named ATTRIBUTE_UNUSED)
5713 size = int_size_in_bytes (type);
5715 size = GET_MODE_SIZE (mode);
5720 return size <= 0 || size > 8;
5724 function_arg_padding (enum machine_mode mode, tree type)
5727 || (TARGET_64BIT && type && AGGREGATE_TYPE_P (type)))
5729 /* Return none if justification is not required. */
5731 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
5732 && (int_size_in_bytes (type) * BITS_PER_UNIT) % PARM_BOUNDARY == 0)
5735 /* The directions set here are ignored when a BLKmode argument larger
5736 than a word is placed in a register. Different code is used for
5737 the stack and registers. This makes it difficult to have a
5738 consistent data representation for both the stack and registers.
5739 For both runtimes, the justification and padding for arguments on
5740 the stack and in registers should be identical. */
5742 /* The 64-bit runtime specifies left justification for aggregates. */
5745 /* The 32-bit runtime architecture specifies right justification.
5746 When the argument is passed on the stack, the argument is padded
5747 with garbage on the left. The HP compiler pads with zeros. */
5751 if (GET_MODE_BITSIZE (mode) < PARM_BOUNDARY)
5758 /* Do what is necessary for `va_start'. We look at the current function
5759 to determine if stdargs or varargs is used and fill in an initial
5760 va_list. A pointer to this constructor is returned. */
5763 hppa_builtin_saveregs (void)
5766 tree fntype = TREE_TYPE (current_function_decl);
5767 int argadj = ((!(TYPE_ARG_TYPES (fntype) != 0
5768 && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype)))
5769 != void_type_node)))
5770 ? UNITS_PER_WORD : 0);
5773 offset = plus_constant (current_function_arg_offset_rtx, argadj);
5775 offset = current_function_arg_offset_rtx;
5781 /* Adjust for varargs/stdarg differences. */
5783 offset = plus_constant (current_function_arg_offset_rtx, -argadj);
5785 offset = current_function_arg_offset_rtx;
5787 /* We need to save %r26 .. %r19 inclusive starting at offset -64
5788 from the incoming arg pointer and growing to larger addresses. */
5789 for (i = 26, off = -64; i >= 19; i--, off += 8)
5790 emit_move_insn (gen_rtx_MEM (word_mode,
5791 plus_constant (arg_pointer_rtx, off)),
5792 gen_rtx_REG (word_mode, i));
5794 /* The incoming args pointer points just beyond the flushback area;
5795 normally this is not a serious concern. However, when we are doing
5796 varargs/stdargs we want to make the arg pointer point to the start
5797 of the incoming argument area. */
5798 emit_move_insn (virtual_incoming_args_rtx,
5799 plus_constant (arg_pointer_rtx, -64));
5801 /* Now return a pointer to the first anonymous argument. */
5802 return copy_to_reg (expand_binop (Pmode, add_optab,
5803 virtual_incoming_args_rtx,
5804 offset, 0, 0, OPTAB_LIB_WIDEN));
5807 /* Store general registers on the stack. */
5808 dest = gen_rtx_MEM (BLKmode,
5809 plus_constant (current_function_internal_arg_pointer,
5811 set_mem_alias_set (dest, get_varargs_alias_set ());
5812 set_mem_align (dest, BITS_PER_WORD);
5813 move_block_from_reg (23, dest, 4);
5815 /* move_block_from_reg will emit code to store the argument registers
5816 individually as scalar stores.
5818 However, other insns may later load from the same addresses for
5819 a structure load (passing a struct to a varargs routine).
5821 The alias code assumes that such aliasing can never happen, so we
5822 have to keep memory referencing insns from moving up beyond the
5823 last argument register store. So we emit a blockage insn here. */
5824 emit_insn (gen_blockage ());
5826 return copy_to_reg (expand_binop (Pmode, add_optab,
5827 current_function_internal_arg_pointer,
5828 offset, 0, 0, OPTAB_LIB_WIDEN));
5832 hppa_va_start (tree valist, rtx nextarg)
5834 nextarg = expand_builtin_saveregs ();
5835 std_expand_builtin_va_start (valist, nextarg);
5839 hppa_gimplify_va_arg_expr (tree valist, tree type, tree *pre_p, tree *post_p)
5843 /* Args grow upward. We can use the generic routines. */
5844 return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
5846 else /* !TARGET_64BIT */
5848 tree ptr = build_pointer_type (type);
5851 unsigned int size, ofs;
5854 indirect = pass_by_reference (NULL, TYPE_MODE (type), type, 0);
5858 ptr = build_pointer_type (type);
5860 size = int_size_in_bytes (type);
5861 valist_type = TREE_TYPE (valist);
5863 /* Args grow down. Not handled by generic routines. */
5865 u = fold_convert (valist_type, size_in_bytes (type));
5866 t = build2 (MINUS_EXPR, valist_type, valist, u);
5868 /* Copied from va-pa.h, but we probably don't need to align to
5869 word size, since we generate and preserve that invariant. */
5870 u = build_int_cst (valist_type, (size > 4 ? -8 : -4));
5871 t = build2 (BIT_AND_EXPR, valist_type, t, u);
5873 t = build2 (MODIFY_EXPR, valist_type, valist, t);
5875 ofs = (8 - size) % 4;
5878 u = fold_convert (valist_type, size_int (ofs));
5879 t = build2 (PLUS_EXPR, valist_type, t, u);
5882 t = fold_convert (ptr, t);
5883 t = build_va_arg_indirect_ref (t);
5886 t = build_va_arg_indirect_ref (t);
5892 /* True if MODE is valid for the target. By "valid", we mean able to
5893 be manipulated in non-trivial ways. In particular, this means all
5894 the arithmetic is supported.
5896 Currently, TImode is not valid as the HP 64-bit runtime documentation
5897 doesn't document the alignment and calling conventions for this type.
5898 Thus, we return false when PRECISION is 2 * BITS_PER_WORD and
5899 2 * BITS_PER_WORD isn't equal LONG_LONG_TYPE_SIZE. */
5902 pa_scalar_mode_supported_p (enum machine_mode mode)
5904 int precision = GET_MODE_PRECISION (mode);
5906 switch (GET_MODE_CLASS (mode))
5908 case MODE_PARTIAL_INT:
5910 if (precision == CHAR_TYPE_SIZE)
5912 if (precision == SHORT_TYPE_SIZE)
5914 if (precision == INT_TYPE_SIZE)
5916 if (precision == LONG_TYPE_SIZE)
5918 if (precision == LONG_LONG_TYPE_SIZE)
5923 if (precision == FLOAT_TYPE_SIZE)
5925 if (precision == DOUBLE_TYPE_SIZE)
5927 if (precision == LONG_DOUBLE_TYPE_SIZE)
5931 case MODE_DECIMAL_FLOAT:
5939 /* This routine handles all the normal conditional branch sequences we
5940 might need to generate. It handles compare immediate vs compare
5941 register, nullification of delay slots, varying length branches,
5942 negated branches, and all combinations of the above. It returns the
5943 output appropriate to emit the branch corresponding to all given
5947 output_cbranch (rtx *operands, int nullify, int length, int negated, rtx insn)
5949 static char buf[100];
5953 /* A conditional branch to the following instruction (e.g. the delay slot)
5954 is asking for a disaster. This can happen when not optimizing and
5955 when jump optimization fails.
5957 While it is usually safe to emit nothing, this can fail if the
5958 preceding instruction is a nullified branch with an empty delay
5959 slot and the same branch target as this branch. We could check
5960 for this but jump optimization should eliminate nop jumps. It
5961 is always safe to emit a nop. */
5962 if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
5965 /* The doubleword form of the cmpib instruction doesn't have the LEU
5966 and GTU conditions while the cmpb instruction does. Since we accept
5967 zero for cmpb, we must ensure that we use cmpb for the comparison. */
5968 if (GET_MODE (operands[1]) == DImode && operands[2] == const0_rtx)
5969 operands[2] = gen_rtx_REG (DImode, 0);
5970 if (GET_MODE (operands[2]) == DImode && operands[1] == const0_rtx)
5971 operands[1] = gen_rtx_REG (DImode, 0);
5973 /* If this is a long branch with its delay slot unfilled, set `nullify'
5974 as it can nullify the delay slot and save a nop. */
5975 if (length == 8 && dbr_sequence_length () == 0)
5978 /* If this is a short forward conditional branch which did not get
5979 its delay slot filled, the delay slot can still be nullified. */
5980 if (! nullify && length == 4 && dbr_sequence_length () == 0)
5981 nullify = forward_branch_p (insn);
5983 /* A forward branch over a single nullified insn can be done with a
5984 comclr instruction. This avoids a single cycle penalty due to
5985 mis-predicted branch if we fall through (branch not taken). */
5987 && next_real_insn (insn) != 0
5988 && get_attr_length (next_real_insn (insn)) == 4
5989 && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn))
5995 /* All short conditional branches except backwards with an unfilled
5999 strcpy (buf, "{com%I2clr,|cmp%I2clr,}");
6001 strcpy (buf, "{com%I2b,|cmp%I2b,}");
6002 if (GET_MODE (operands[1]) == DImode)
6005 strcat (buf, "%B3");
6007 strcat (buf, "%S3");
6009 strcat (buf, " %2,%r1,%%r0");
6011 strcat (buf, ",n %2,%r1,%0");
6013 strcat (buf, " %2,%r1,%0");
6016 /* All long conditionals. Note a short backward branch with an
6017 unfilled delay slot is treated just like a long backward branch
6018 with an unfilled delay slot. */
6020 /* Handle weird backwards branch with a filled delay slot
6021 with is nullified. */
6022 if (dbr_sequence_length () != 0
6023 && ! forward_branch_p (insn)
6026 strcpy (buf, "{com%I2b,|cmp%I2b,}");
6027 if (GET_MODE (operands[1]) == DImode)
6030 strcat (buf, "%S3");
6032 strcat (buf, "%B3");
6033 strcat (buf, ",n %2,%r1,.+12\n\tb %0");
6035 /* Handle short backwards branch with an unfilled delay slot.
6036 Using a comb;nop rather than comiclr;bl saves 1 cycle for both
6037 taken and untaken branches. */
6038 else if (dbr_sequence_length () == 0
6039 && ! forward_branch_p (insn)
6040 && INSN_ADDRESSES_SET_P ()
6041 && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn)))
6042 - INSN_ADDRESSES (INSN_UID (insn)) - 8))
6044 strcpy (buf, "{com%I2b,|cmp%I2b,}");
6045 if (GET_MODE (operands[1]) == DImode)
6048 strcat (buf, "%B3 %2,%r1,%0%#");
6050 strcat (buf, "%S3 %2,%r1,%0%#");
6054 strcpy (buf, "{com%I2clr,|cmp%I2clr,}");
6055 if (GET_MODE (operands[1]) == DImode)
6058 strcat (buf, "%S3");
6060 strcat (buf, "%B3");
6062 strcat (buf, " %2,%r1,%%r0\n\tb,n %0");
6064 strcat (buf, " %2,%r1,%%r0\n\tb %0");
6070 xoperands[0] = operands[0];
6071 xoperands[1] = operands[1];
6072 xoperands[2] = operands[2];
6073 xoperands[3] = operands[3];
6075 /* The reversed conditional branch must branch over one additional
6076 instruction if the delay slot is filled. If the delay slot
6077 is empty, the instruction after the reversed condition branch
6078 must be nullified. */
6079 nullify = dbr_sequence_length () == 0;
6080 xoperands[4] = nullify ? GEN_INT (length) : GEN_INT (length + 4);
6082 /* Create a reversed conditional branch which branches around
6083 the following insns. */
6084 if (GET_MODE (operands[1]) != DImode)
6090 "{com%I2b,%S3,n %2,%r1,.+%4|cmp%I2b,%S3,n %2,%r1,.+%4}");
6093 "{com%I2b,%B3,n %2,%r1,.+%4|cmp%I2b,%B3,n %2,%r1,.+%4}");
6099 "{com%I2b,%S3 %2,%r1,.+%4|cmp%I2b,%S3 %2,%r1,.+%4}");
6102 "{com%I2b,%B3 %2,%r1,.+%4|cmp%I2b,%B3 %2,%r1,.+%4}");
6111 "{com%I2b,*%S3,n %2,%r1,.+%4|cmp%I2b,*%S3,n %2,%r1,.+%4}");
6114 "{com%I2b,*%B3,n %2,%r1,.+%4|cmp%I2b,*%B3,n %2,%r1,.+%4}");
6120 "{com%I2b,*%S3 %2,%r1,.+%4|cmp%I2b,*%S3 %2,%r1,.+%4}");
6123 "{com%I2b,*%B3 %2,%r1,.+%4|cmp%I2b,*%B3 %2,%r1,.+%4}");
6127 output_asm_insn (buf, xoperands);
6128 return output_lbranch (operands[0], insn);
6136 /* This routine handles long unconditional branches that exceed the
6137 maximum range of a simple branch instruction. */
6140 output_lbranch (rtx dest, rtx insn)
6144 xoperands[0] = dest;
6146 /* First, free up the delay slot. */
6147 if (dbr_sequence_length () != 0)
6149 /* We can't handle a jump in the delay slot. */
6150 gcc_assert (GET_CODE (NEXT_INSN (insn)) != JUMP_INSN);
6152 final_scan_insn (NEXT_INSN (insn), asm_out_file,
6155 /* Now delete the delay insn. */
6156 PUT_CODE (NEXT_INSN (insn), NOTE);
6157 NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
6158 NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
6161 /* Output an insn to save %r1. The runtime documentation doesn't
6162 specify whether the "Clean Up" slot in the callers frame can
6163 be clobbered by the callee. It isn't copied by HP's builtin
6164 alloca, so this suggests that it can be clobbered if necessary.
6165 The "Static Link" location is copied by HP builtin alloca, so
6166 we avoid using it. Using the cleanup slot might be a problem
6167 if we have to interoperate with languages that pass cleanup
6168 information. However, it should be possible to handle these
6169 situations with GCC's asm feature.
6171 The "Current RP" slot is reserved for the called procedure, so
6172 we try to use it when we don't have a frame of our own. It's
6173 rather unlikely that we won't have a frame when we need to emit
6176 Really the way to go long term is a register scavenger; goto
6177 the target of the jump and find a register which we can use
6178 as a scratch to hold the value in %r1. Then, we wouldn't have
6179 to free up the delay slot or clobber a slot that may be needed
6180 for other purposes. */
6183 if (actual_fsize == 0 && !regs_ever_live[2])
6184 /* Use the return pointer slot in the frame marker. */
6185 output_asm_insn ("std %%r1,-16(%%r30)", xoperands);
6187 /* Use the slot at -40 in the frame marker since HP builtin
6188 alloca doesn't copy it. */
6189 output_asm_insn ("std %%r1,-40(%%r30)", xoperands);
6193 if (actual_fsize == 0 && !regs_ever_live[2])
6194 /* Use the return pointer slot in the frame marker. */
6195 output_asm_insn ("stw %%r1,-20(%%r30)", xoperands);
6197 /* Use the "Clean Up" slot in the frame marker. In GCC,
6198 the only other use of this location is for copying a
6199 floating point double argument from a floating-point
6200 register to two general registers. The copy is done
6201 as an "atomic" operation when outputting a call, so it
6202 won't interfere with our using the location here. */
6203 output_asm_insn ("stw %%r1,-12(%%r30)", xoperands);
6206 if (TARGET_PORTABLE_RUNTIME)
6208 output_asm_insn ("ldil L'%0,%%r1", xoperands);
6209 output_asm_insn ("ldo R'%0(%%r1),%%r1", xoperands);
6210 output_asm_insn ("bv %%r0(%%r1)", xoperands);
6214 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
6215 if (TARGET_SOM || !TARGET_GAS)
6217 xoperands[1] = gen_label_rtx ();
6218 output_asm_insn ("addil L'%l0-%l1,%%r1", xoperands);
6219 (*targetm.asm_out.internal_label) (asm_out_file, "L",
6220 CODE_LABEL_NUMBER (xoperands[1]));
6221 output_asm_insn ("ldo R'%l0-%l1(%%r1),%%r1", xoperands);
6225 output_asm_insn ("addil L'%l0-$PIC_pcrel$0+4,%%r1", xoperands);
6226 output_asm_insn ("ldo R'%l0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands);
6228 output_asm_insn ("bv %%r0(%%r1)", xoperands);
6231 /* Now output a very long branch to the original target. */
6232 output_asm_insn ("ldil L'%l0,%%r1\n\tbe R'%l0(%%sr4,%%r1)", xoperands);
6234 /* Now restore the value of %r1 in the delay slot. */
6237 if (actual_fsize == 0 && !regs_ever_live[2])
6238 return "ldd -16(%%r30),%%r1";
6240 return "ldd -40(%%r30),%%r1";
6244 if (actual_fsize == 0 && !regs_ever_live[2])
6245 return "ldw -20(%%r30),%%r1";
6247 return "ldw -12(%%r30),%%r1";
6251 /* This routine handles all the branch-on-bit conditional branch sequences we
6252 might need to generate. It handles nullification of delay slots,
6253 varying length branches, negated branches and all combinations of the
6254 above. it returns the appropriate output template to emit the branch. */
6257 output_bb (rtx *operands ATTRIBUTE_UNUSED, int nullify, int length,
6258 int negated, rtx insn, int which)
6260 static char buf[100];
6263 /* A conditional branch to the following instruction (e.g. the delay slot) is
6264 asking for a disaster. I do not think this can happen as this pattern
6265 is only used when optimizing; jump optimization should eliminate the
6266 jump. But be prepared just in case. */
6268 if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
6271 /* If this is a long branch with its delay slot unfilled, set `nullify'
6272 as it can nullify the delay slot and save a nop. */
6273 if (length == 8 && dbr_sequence_length () == 0)
6276 /* If this is a short forward conditional branch which did not get
6277 its delay slot filled, the delay slot can still be nullified. */
6278 if (! nullify && length == 4 && dbr_sequence_length () == 0)
6279 nullify = forward_branch_p (insn);
6281 /* A forward branch over a single nullified insn can be done with a
6282 extrs instruction. This avoids a single cycle penalty due to
6283 mis-predicted branch if we fall through (branch not taken). */
6286 && next_real_insn (insn) != 0
6287 && get_attr_length (next_real_insn (insn)) == 4
6288 && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn))
6295 /* All short conditional branches except backwards with an unfilled
6299 strcpy (buf, "{extrs,|extrw,s,}");
6301 strcpy (buf, "bb,");
6302 if (useskip && GET_MODE (operands[0]) == DImode)
6303 strcpy (buf, "extrd,s,*");
6304 else if (GET_MODE (operands[0]) == DImode)
6305 strcpy (buf, "bb,*");
6306 if ((which == 0 && negated)
6307 || (which == 1 && ! negated))
6312 strcat (buf, " %0,%1,1,%%r0");
6313 else if (nullify && negated)
6314 strcat (buf, ",n %0,%1,%3");
6315 else if (nullify && ! negated)
6316 strcat (buf, ",n %0,%1,%2");
6317 else if (! nullify && negated)
6318 strcat (buf, "%0,%1,%3");
6319 else if (! nullify && ! negated)
6320 strcat (buf, " %0,%1,%2");
6323 /* All long conditionals. Note a short backward branch with an
6324 unfilled delay slot is treated just like a long backward branch
6325 with an unfilled delay slot. */
6327 /* Handle weird backwards branch with a filled delay slot
6328 with is nullified. */
6329 if (dbr_sequence_length () != 0
6330 && ! forward_branch_p (insn)
6333 strcpy (buf, "bb,");
6334 if (GET_MODE (operands[0]) == DImode)
6336 if ((which == 0 && negated)
6337 || (which == 1 && ! negated))
6342 strcat (buf, ",n %0,%1,.+12\n\tb %3");
6344 strcat (buf, ",n %0,%1,.+12\n\tb %2");
6346 /* Handle short backwards branch with an unfilled delay slot.
6347 Using a bb;nop rather than extrs;bl saves 1 cycle for both
6348 taken and untaken branches. */
6349 else if (dbr_sequence_length () == 0
6350 && ! forward_branch_p (insn)
6351 && INSN_ADDRESSES_SET_P ()
6352 && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn)))
6353 - INSN_ADDRESSES (INSN_UID (insn)) - 8))
6355 strcpy (buf, "bb,");
6356 if (GET_MODE (operands[0]) == DImode)
6358 if ((which == 0 && negated)
6359 || (which == 1 && ! negated))
6364 strcat (buf, " %0,%1,%3%#");
6366 strcat (buf, " %0,%1,%2%#");
6370 strcpy (buf, "{extrs,|extrw,s,}");
6371 if (GET_MODE (operands[0]) == DImode)
6372 strcpy (buf, "extrd,s,*");
6373 if ((which == 0 && negated)
6374 || (which == 1 && ! negated))
6378 if (nullify && negated)
6379 strcat (buf, " %0,%1,1,%%r0\n\tb,n %3");
6380 else if (nullify && ! negated)
6381 strcat (buf, " %0,%1,1,%%r0\n\tb,n %2");
6383 strcat (buf, " %0,%1,1,%%r0\n\tb %3");
6385 strcat (buf, " %0,%1,1,%%r0\n\tb %2");
6395 /* This routine handles all the branch-on-variable-bit conditional branch
6396 sequences we might need to generate. It handles nullification of delay
6397 slots, varying length branches, negated branches and all combinations
6398 of the above. it returns the appropriate output template to emit the
6402 output_bvb (rtx *operands ATTRIBUTE_UNUSED, int nullify, int length,
6403 int negated, rtx insn, int which)
6405 static char buf[100];
6408 /* A conditional branch to the following instruction (e.g. the delay slot) is
6409 asking for a disaster. I do not think this can happen as this pattern
6410 is only used when optimizing; jump optimization should eliminate the
6411 jump. But be prepared just in case. */
6413 if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
6416 /* If this is a long branch with its delay slot unfilled, set `nullify'
6417 as it can nullify the delay slot and save a nop. */
6418 if (length == 8 && dbr_sequence_length () == 0)
6421 /* If this is a short forward conditional branch which did not get
6422 its delay slot filled, the delay slot can still be nullified. */
6423 if (! nullify && length == 4 && dbr_sequence_length () == 0)
6424 nullify = forward_branch_p (insn);
6426 /* A forward branch over a single nullified insn can be done with a
6427 extrs instruction. This avoids a single cycle penalty due to
6428 mis-predicted branch if we fall through (branch not taken). */
6431 && next_real_insn (insn) != 0
6432 && get_attr_length (next_real_insn (insn)) == 4
6433 && JUMP_LABEL (insn) == next_nonnote_insn (next_real_insn (insn))
6440 /* All short conditional branches except backwards with an unfilled
6444 strcpy (buf, "{vextrs,|extrw,s,}");
6446 strcpy (buf, "{bvb,|bb,}");
6447 if (useskip && GET_MODE (operands[0]) == DImode)
6448 strcpy (buf, "extrd,s,*");
6449 else if (GET_MODE (operands[0]) == DImode)
6450 strcpy (buf, "bb,*");
6451 if ((which == 0 && negated)
6452 || (which == 1 && ! negated))
6457 strcat (buf, "{ %0,1,%%r0| %0,%%sar,1,%%r0}");
6458 else if (nullify && negated)
6459 strcat (buf, "{,n %0,%3|,n %0,%%sar,%3}");
6460 else if (nullify && ! negated)
6461 strcat (buf, "{,n %0,%2|,n %0,%%sar,%2}");
6462 else if (! nullify && negated)
6463 strcat (buf, "{%0,%3|%0,%%sar,%3}");
6464 else if (! nullify && ! negated)
6465 strcat (buf, "{ %0,%2| %0,%%sar,%2}");
6468 /* All long conditionals. Note a short backward branch with an
6469 unfilled delay slot is treated just like a long backward branch
6470 with an unfilled delay slot. */
6472 /* Handle weird backwards branch with a filled delay slot
6473 with is nullified. */
6474 if (dbr_sequence_length () != 0
6475 && ! forward_branch_p (insn)
6478 strcpy (buf, "{bvb,|bb,}");
6479 if (GET_MODE (operands[0]) == DImode)
6481 if ((which == 0 && negated)
6482 || (which == 1 && ! negated))
6487 strcat (buf, "{,n %0,.+12\n\tb %3|,n %0,%%sar,.+12\n\tb %3}");
6489 strcat (buf, "{,n %0,.+12\n\tb %2|,n %0,%%sar,.+12\n\tb %2}");
6491 /* Handle short backwards branch with an unfilled delay slot.
6492 Using a bb;nop rather than extrs;bl saves 1 cycle for both
6493 taken and untaken branches. */
6494 else if (dbr_sequence_length () == 0
6495 && ! forward_branch_p (insn)
6496 && INSN_ADDRESSES_SET_P ()
6497 && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn)))
6498 - INSN_ADDRESSES (INSN_UID (insn)) - 8))
6500 strcpy (buf, "{bvb,|bb,}");
6501 if (GET_MODE (operands[0]) == DImode)
6503 if ((which == 0 && negated)
6504 || (which == 1 && ! negated))
6509 strcat (buf, "{ %0,%3%#| %0,%%sar,%3%#}");
6511 strcat (buf, "{ %0,%2%#| %0,%%sar,%2%#}");
6515 strcpy (buf, "{vextrs,|extrw,s,}");
6516 if (GET_MODE (operands[0]) == DImode)
6517 strcpy (buf, "extrd,s,*");
6518 if ((which == 0 && negated)
6519 || (which == 1 && ! negated))
6523 if (nullify && negated)
6524 strcat (buf, "{ %0,1,%%r0\n\tb,n %3| %0,%%sar,1,%%r0\n\tb,n %3}");
6525 else if (nullify && ! negated)
6526 strcat (buf, "{ %0,1,%%r0\n\tb,n %2| %0,%%sar,1,%%r0\n\tb,n %2}");
6528 strcat (buf, "{ %0,1,%%r0\n\tb %3| %0,%%sar,1,%%r0\n\tb %3}");
6530 strcat (buf, "{ %0,1,%%r0\n\tb %2| %0,%%sar,1,%%r0\n\tb %2}");
6540 /* Return the output template for emitting a dbra type insn.
6542 Note it may perform some output operations on its own before
6543 returning the final output string. */
6545 output_dbra (rtx *operands, rtx insn, int which_alternative)
6548 /* A conditional branch to the following instruction (e.g. the delay slot) is
6549 asking for a disaster. Be prepared! */
6551 if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
6553 if (which_alternative == 0)
6554 return "ldo %1(%0),%0";
6555 else if (which_alternative == 1)
6557 output_asm_insn ("{fstws|fstw} %0,-16(%%r30)", operands);
6558 output_asm_insn ("ldw -16(%%r30),%4", operands);
6559 output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands);
6560 return "{fldws|fldw} -16(%%r30),%0";
6564 output_asm_insn ("ldw %0,%4", operands);
6565 return "ldo %1(%4),%4\n\tstw %4,%0";
6569 if (which_alternative == 0)
6571 int nullify = INSN_ANNULLED_BRANCH_P (insn);
6572 int length = get_attr_length (insn);
6574 /* If this is a long branch with its delay slot unfilled, set `nullify'
6575 as it can nullify the delay slot and save a nop. */
6576 if (length == 8 && dbr_sequence_length () == 0)
6579 /* If this is a short forward conditional branch which did not get
6580 its delay slot filled, the delay slot can still be nullified. */
6581 if (! nullify && length == 4 && dbr_sequence_length () == 0)
6582 nullify = forward_branch_p (insn);
6588 return "addib,%C2,n %1,%0,%3";
6590 return "addib,%C2 %1,%0,%3";
6593 /* Handle weird backwards branch with a fulled delay slot
6594 which is nullified. */
6595 if (dbr_sequence_length () != 0
6596 && ! forward_branch_p (insn)
6598 return "addib,%N2,n %1,%0,.+12\n\tb %3";
6599 /* Handle short backwards branch with an unfilled delay slot.
6600 Using a addb;nop rather than addi;bl saves 1 cycle for both
6601 taken and untaken branches. */
6602 else if (dbr_sequence_length () == 0
6603 && ! forward_branch_p (insn)
6604 && INSN_ADDRESSES_SET_P ()
6605 && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn)))
6606 - INSN_ADDRESSES (INSN_UID (insn)) - 8))
6607 return "addib,%C2 %1,%0,%3%#";
6609 /* Handle normal cases. */
6611 return "addi,%N2 %1,%0,%0\n\tb,n %3";
6613 return "addi,%N2 %1,%0,%0\n\tb %3";
6620 /* Deal with gross reload from FP register case. */
6621 else if (which_alternative == 1)
6623 /* Move loop counter from FP register to MEM then into a GR,
6624 increment the GR, store the GR into MEM, and finally reload
6625 the FP register from MEM from within the branch's delay slot. */
6626 output_asm_insn ("{fstws|fstw} %0,-16(%%r30)\n\tldw -16(%%r30),%4",
6628 output_asm_insn ("ldo %1(%4),%4\n\tstw %4,-16(%%r30)", operands);
6629 if (get_attr_length (insn) == 24)
6630 return "{comb|cmpb},%S2 %%r0,%4,%3\n\t{fldws|fldw} -16(%%r30),%0";
6632 return "{comclr|cmpclr},%B2 %%r0,%4,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0";
6634 /* Deal with gross reload from memory case. */
6637 /* Reload loop counter from memory, the store back to memory
6638 happens in the branch's delay slot. */
6639 output_asm_insn ("ldw %0,%4", operands);
6640 if (get_attr_length (insn) == 12)
6641 return "addib,%C2 %1,%4,%3\n\tstw %4,%0";
6643 return "addi,%N2 %1,%4,%4\n\tb %3\n\tstw %4,%0";
6647 /* Return the output template for emitting a dbra type insn.
6649 Note it may perform some output operations on its own before
6650 returning the final output string. */
6652 output_movb (rtx *operands, rtx insn, int which_alternative,
6653 int reverse_comparison)
6656 /* A conditional branch to the following instruction (e.g. the delay slot) is
6657 asking for a disaster. Be prepared! */
6659 if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
6661 if (which_alternative == 0)
6662 return "copy %1,%0";
6663 else if (which_alternative == 1)
6665 output_asm_insn ("stw %1,-16(%%r30)", operands);
6666 return "{fldws|fldw} -16(%%r30),%0";
6668 else if (which_alternative == 2)
6674 /* Support the second variant. */
6675 if (reverse_comparison)
6676 PUT_CODE (operands[2], reverse_condition (GET_CODE (operands[2])));
6678 if (which_alternative == 0)
6680 int nullify = INSN_ANNULLED_BRANCH_P (insn);
6681 int length = get_attr_length (insn);
6683 /* If this is a long branch with its delay slot unfilled, set `nullify'
6684 as it can nullify the delay slot and save a nop. */
6685 if (length == 8 && dbr_sequence_length () == 0)
6688 /* If this is a short forward conditional branch which did not get
6689 its delay slot filled, the delay slot can still be nullified. */
6690 if (! nullify && length == 4 && dbr_sequence_length () == 0)
6691 nullify = forward_branch_p (insn);
6697 return "movb,%C2,n %1,%0,%3";
6699 return "movb,%C2 %1,%0,%3";
6702 /* Handle weird backwards branch with a filled delay slot
6703 which is nullified. */
6704 if (dbr_sequence_length () != 0
6705 && ! forward_branch_p (insn)
6707 return "movb,%N2,n %1,%0,.+12\n\tb %3";
6709 /* Handle short backwards branch with an unfilled delay slot.
6710 Using a movb;nop rather than or;bl saves 1 cycle for both
6711 taken and untaken branches. */
6712 else if (dbr_sequence_length () == 0
6713 && ! forward_branch_p (insn)
6714 && INSN_ADDRESSES_SET_P ()
6715 && VAL_14_BITS_P (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (insn)))
6716 - INSN_ADDRESSES (INSN_UID (insn)) - 8))
6717 return "movb,%C2 %1,%0,%3%#";
6718 /* Handle normal cases. */
6720 return "or,%N2 %1,%%r0,%0\n\tb,n %3";
6722 return "or,%N2 %1,%%r0,%0\n\tb %3";
6728 /* Deal with gross reload from FP register case. */
6729 else if (which_alternative == 1)
6731 /* Move loop counter from FP register to MEM then into a GR,
6732 increment the GR, store the GR into MEM, and finally reload
6733 the FP register from MEM from within the branch's delay slot. */
6734 output_asm_insn ("stw %1,-16(%%r30)", operands);
6735 if (get_attr_length (insn) == 12)
6736 return "{comb|cmpb},%S2 %%r0,%1,%3\n\t{fldws|fldw} -16(%%r30),%0";
6738 return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\t{fldws|fldw} -16(%%r30),%0";
6740 /* Deal with gross reload from memory case. */
6741 else if (which_alternative == 2)
6743 /* Reload loop counter from memory, the store back to memory
6744 happens in the branch's delay slot. */
6745 if (get_attr_length (insn) == 8)
6746 return "{comb|cmpb},%S2 %%r0,%1,%3\n\tstw %1,%0";
6748 return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\tstw %1,%0";
6750 /* Handle SAR as a destination. */
6753 if (get_attr_length (insn) == 8)
6754 return "{comb|cmpb},%S2 %%r0,%1,%3\n\tmtsar %r1";
6756 return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\tmtsar %r1";
6760 /* Copy any FP arguments in INSN into integer registers. */
6762 copy_fp_args (rtx insn)
6767 for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1))
6769 int arg_mode, regno;
6770 rtx use = XEXP (link, 0);
6772 if (! (GET_CODE (use) == USE
6773 && GET_CODE (XEXP (use, 0)) == REG
6774 && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0)))))
6777 arg_mode = GET_MODE (XEXP (use, 0));
6778 regno = REGNO (XEXP (use, 0));
6780 /* Is it a floating point register? */
6781 if (regno >= 32 && regno <= 39)
6783 /* Copy the FP register into an integer register via memory. */
6784 if (arg_mode == SFmode)
6786 xoperands[0] = XEXP (use, 0);
6787 xoperands[1] = gen_rtx_REG (SImode, 26 - (regno - 32) / 2);
6788 output_asm_insn ("{fstws|fstw} %0,-16(%%sr0,%%r30)", xoperands);
6789 output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands);
6793 xoperands[0] = XEXP (use, 0);
6794 xoperands[1] = gen_rtx_REG (DImode, 25 - (regno - 34) / 2);
6795 output_asm_insn ("{fstds|fstd} %0,-16(%%sr0,%%r30)", xoperands);
6796 output_asm_insn ("ldw -12(%%sr0,%%r30),%R1", xoperands);
6797 output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands);
6803 /* Compute length of the FP argument copy sequence for INSN. */
6805 length_fp_args (rtx insn)
6810 for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1))
6812 int arg_mode, regno;
6813 rtx use = XEXP (link, 0);
6815 if (! (GET_CODE (use) == USE
6816 && GET_CODE (XEXP (use, 0)) == REG
6817 && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0)))))
6820 arg_mode = GET_MODE (XEXP (use, 0));
6821 regno = REGNO (XEXP (use, 0));
6823 /* Is it a floating point register? */
6824 if (regno >= 32 && regno <= 39)
6826 if (arg_mode == SFmode)
6836 /* Return the attribute length for the millicode call instruction INSN.
6837 The length must match the code generated by output_millicode_call.
6838 We include the delay slot in the returned length as it is better to
6839 over estimate the length than to under estimate it. */
6842 attr_length_millicode_call (rtx insn)
6844 unsigned long distance = -1;
6845 unsigned long total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes;
6847 if (INSN_ADDRESSES_SET_P ())
6849 distance = (total + insn_current_reference_address (insn));
6850 if (distance < total)
6856 if (!TARGET_LONG_CALLS && distance < 7600000)
6861 else if (TARGET_PORTABLE_RUNTIME)
6865 if (!TARGET_LONG_CALLS && distance < 240000)
6868 if (TARGET_LONG_ABS_CALL && !flag_pic)
6875 /* INSN is a function call. It may have an unconditional jump
6878 CALL_DEST is the routine we are calling. */
6881 output_millicode_call (rtx insn, rtx call_dest)
6883 int attr_length = get_attr_length (insn);
6884 int seq_length = dbr_sequence_length ();
6889 xoperands[0] = call_dest;
6890 xoperands[2] = gen_rtx_REG (Pmode, TARGET_64BIT ? 2 : 31);
6892 /* Handle the common case where we are sure that the branch will
6893 reach the beginning of the $CODE$ subspace. The within reach
6894 form of the $$sh_func_adrs call has a length of 28. Because
6895 it has an attribute type of multi, it never has a nonzero
6896 sequence length. The length of the $$sh_func_adrs is the same
6897 as certain out of reach PIC calls to other routines. */
6898 if (!TARGET_LONG_CALLS
6899 && ((seq_length == 0
6900 && (attr_length == 12
6901 || (attr_length == 28 && get_attr_type (insn) == TYPE_MULTI)))
6902 || (seq_length != 0 && attr_length == 8)))
6904 output_asm_insn ("{bl|b,l} %0,%2", xoperands);
6910 /* It might seem that one insn could be saved by accessing
6911 the millicode function using the linkage table. However,
6912 this doesn't work in shared libraries and other dynamically
6913 loaded objects. Using a pc-relative sequence also avoids
6914 problems related to the implicit use of the gp register. */
6915 output_asm_insn ("b,l .+8,%%r1", xoperands);
6919 output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands);
6920 output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands);
6924 xoperands[1] = gen_label_rtx ();
6925 output_asm_insn ("addil L'%0-%l1,%%r1", xoperands);
6926 (*targetm.asm_out.internal_label) (asm_out_file, "L",
6927 CODE_LABEL_NUMBER (xoperands[1]));
6928 output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands);
6931 output_asm_insn ("bve,l (%%r1),%%r2", xoperands);
6933 else if (TARGET_PORTABLE_RUNTIME)
6935 /* Pure portable runtime doesn't allow be/ble; we also don't
6936 have PIC support in the assembler/linker, so this sequence
6939 /* Get the address of our target into %r1. */
6940 output_asm_insn ("ldil L'%0,%%r1", xoperands);
6941 output_asm_insn ("ldo R'%0(%%r1),%%r1", xoperands);
6943 /* Get our return address into %r31. */
6944 output_asm_insn ("{bl|b,l} .+8,%%r31", xoperands);
6945 output_asm_insn ("addi 8,%%r31,%%r31", xoperands);
6947 /* Jump to our target address in %r1. */
6948 output_asm_insn ("bv %%r0(%%r1)", xoperands);
6952 output_asm_insn ("ldil L'%0,%%r1", xoperands);
6954 output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31", xoperands);
6956 output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands);
6960 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
6961 output_asm_insn ("addi 16,%%r1,%%r31", xoperands);
6963 if (TARGET_SOM || !TARGET_GAS)
6965 /* The HP assembler can generate relocations for the
6966 difference of two symbols. GAS can do this for a
6967 millicode symbol but not an arbitrary external
6968 symbol when generating SOM output. */
6969 xoperands[1] = gen_label_rtx ();
6970 (*targetm.asm_out.internal_label) (asm_out_file, "L",
6971 CODE_LABEL_NUMBER (xoperands[1]));
6972 output_asm_insn ("addil L'%0-%l1,%%r1", xoperands);
6973 output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands);
6977 output_asm_insn ("addil L'%0-$PIC_pcrel$0+8,%%r1", xoperands);
6978 output_asm_insn ("ldo R'%0-$PIC_pcrel$0+12(%%r1),%%r1",
6982 /* Jump to our target address in %r1. */
6983 output_asm_insn ("bv %%r0(%%r1)", xoperands);
6987 if (seq_length == 0)
6988 output_asm_insn ("nop", xoperands);
6990 /* We are done if there isn't a jump in the delay slot. */
6991 if (seq_length == 0 || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN)
6994 /* This call has an unconditional jump in its delay slot. */
6995 xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1);
6997 /* See if the return address can be adjusted. Use the containing
6998 sequence insn's address. */
6999 if (INSN_ADDRESSES_SET_P ())
7001 seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0)));
7002 distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn))))
7003 - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8);
7005 if (VAL_14_BITS_P (distance))
7007 xoperands[1] = gen_label_rtx ();
7008 output_asm_insn ("ldo %0-%1(%2),%2", xoperands);
7009 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7010 CODE_LABEL_NUMBER (xoperands[1]));
7013 /* ??? This branch may not reach its target. */
7014 output_asm_insn ("nop\n\tb,n %0", xoperands);
7017 /* ??? This branch may not reach its target. */
7018 output_asm_insn ("nop\n\tb,n %0", xoperands);
7020 /* Delete the jump. */
7021 PUT_CODE (NEXT_INSN (insn), NOTE);
7022 NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
7023 NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
7028 /* Return the attribute length of the call instruction INSN. The SIBCALL
7029 flag indicates whether INSN is a regular call or a sibling call. The
7030 length returned must be longer than the code actually generated by
7031 output_call. Since branch shortening is done before delay branch
7032 sequencing, there is no way to determine whether or not the delay
7033 slot will be filled during branch shortening. Even when the delay
7034 slot is filled, we may have to add a nop if the delay slot contains
7035 a branch that can't reach its target. Thus, we always have to include
7036 the delay slot in the length estimate. This used to be done in
7037 pa_adjust_insn_length but we do it here now as some sequences always
7038 fill the delay slot and we can save four bytes in the estimate for
7042 attr_length_call (rtx insn, int sibcall)
7048 rtx pat = PATTERN (insn);
7049 unsigned long distance = -1;
7051 if (INSN_ADDRESSES_SET_P ())
7053 unsigned long total;
7055 total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes;
7056 distance = (total + insn_current_reference_address (insn));
7057 if (distance < total)
7061 /* Determine if this is a local call. */
7062 if (GET_CODE (XVECEXP (pat, 0, 0)) == CALL)
7063 call_dest = XEXP (XEXP (XVECEXP (pat, 0, 0), 0), 0);
7065 call_dest = XEXP (XEXP (XEXP (XVECEXP (pat, 0, 0), 1), 0), 0);
7067 call_decl = SYMBOL_REF_DECL (call_dest);
7068 local_call = call_decl && (*targetm.binds_local_p) (call_decl);
7070 /* pc-relative branch. */
7071 if (!TARGET_LONG_CALLS
7072 && ((TARGET_PA_20 && !sibcall && distance < 7600000)
7073 || distance < 240000))
7076 /* 64-bit plabel sequence. */
7077 else if (TARGET_64BIT && !local_call)
7078 length += sibcall ? 28 : 24;
7080 /* non-pic long absolute branch sequence. */
7081 else if ((TARGET_LONG_ABS_CALL || local_call) && !flag_pic)
7084 /* long pc-relative branch sequence. */
7085 else if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL)
7086 || (TARGET_64BIT && !TARGET_GAS)
7087 || (TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call)))
7091 if (!TARGET_PA_20 && !TARGET_NO_SPACE_REGS)
7095 /* 32-bit plabel sequence. */
7101 length += length_fp_args (insn);
7111 if (!TARGET_NO_SPACE_REGS)
7119 /* INSN is a function call. It may have an unconditional jump
7122 CALL_DEST is the routine we are calling. */
7125 output_call (rtx insn, rtx call_dest, int sibcall)
7127 int delay_insn_deleted = 0;
7128 int delay_slot_filled = 0;
7129 int seq_length = dbr_sequence_length ();
7130 tree call_decl = SYMBOL_REF_DECL (call_dest);
7131 int local_call = call_decl && (*targetm.binds_local_p) (call_decl);
7134 xoperands[0] = call_dest;
7136 /* Handle the common case where we're sure that the branch will reach
7137 the beginning of the "$CODE$" subspace. This is the beginning of
7138 the current function if we are in a named section. */
7139 if (!TARGET_LONG_CALLS && attr_length_call (insn, sibcall) == 8)
7141 xoperands[1] = gen_rtx_REG (word_mode, sibcall ? 0 : 2);
7142 output_asm_insn ("{bl|b,l} %0,%1", xoperands);
7146 if (TARGET_64BIT && !local_call)
7148 /* ??? As far as I can tell, the HP linker doesn't support the
7149 long pc-relative sequence described in the 64-bit runtime
7150 architecture. So, we use a slightly longer indirect call. */
7151 xoperands[0] = get_deferred_plabel (call_dest);
7152 xoperands[1] = gen_label_rtx ();
7154 /* If this isn't a sibcall, we put the load of %r27 into the
7155 delay slot. We can't do this in a sibcall as we don't
7156 have a second call-clobbered scratch register available. */
7158 && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN
7161 final_scan_insn (NEXT_INSN (insn), asm_out_file,
7164 /* Now delete the delay insn. */
7165 PUT_CODE (NEXT_INSN (insn), NOTE);
7166 NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
7167 NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
7168 delay_insn_deleted = 1;
7171 output_asm_insn ("addil LT'%0,%%r27", xoperands);
7172 output_asm_insn ("ldd RT'%0(%%r1),%%r1", xoperands);
7173 output_asm_insn ("ldd 0(%%r1),%%r1", xoperands);
7177 output_asm_insn ("ldd 24(%%r1),%%r27", xoperands);
7178 output_asm_insn ("ldd 16(%%r1),%%r1", xoperands);
7179 output_asm_insn ("bve (%%r1)", xoperands);
7183 output_asm_insn ("ldd 16(%%r1),%%r2", xoperands);
7184 output_asm_insn ("bve,l (%%r2),%%r2", xoperands);
7185 output_asm_insn ("ldd 24(%%r1),%%r27", xoperands);
7186 delay_slot_filled = 1;
7191 int indirect_call = 0;
7193 /* Emit a long call. There are several different sequences
7194 of increasing length and complexity. In most cases,
7195 they don't allow an instruction in the delay slot. */
7196 if (!((TARGET_LONG_ABS_CALL || local_call) && !flag_pic)
7197 && !(TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL)
7198 && !(TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call))
7203 && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN
7205 && (!TARGET_PA_20 || indirect_call))
7207 /* A non-jump insn in the delay slot. By definition we can
7208 emit this insn before the call (and in fact before argument
7210 final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0,
7213 /* Now delete the delay insn. */
7214 PUT_CODE (NEXT_INSN (insn), NOTE);
7215 NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
7216 NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
7217 delay_insn_deleted = 1;
7220 if ((TARGET_LONG_ABS_CALL || local_call) && !flag_pic)
7222 /* This is the best sequence for making long calls in
7223 non-pic code. Unfortunately, GNU ld doesn't provide
7224 the stub needed for external calls, and GAS's support
7225 for this with the SOM linker is buggy. It is safe
7226 to use this for local calls. */
7227 output_asm_insn ("ldil L'%0,%%r1", xoperands);
7229 output_asm_insn ("be R'%0(%%sr4,%%r1)", xoperands);
7233 output_asm_insn ("be,l R'%0(%%sr4,%%r1),%%sr0,%%r31",
7236 output_asm_insn ("ble R'%0(%%sr4,%%r1)", xoperands);
7238 output_asm_insn ("copy %%r31,%%r2", xoperands);
7239 delay_slot_filled = 1;
7244 if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL)
7245 || (TARGET_64BIT && !TARGET_GAS))
7247 /* The HP assembler and linker can handle relocations
7248 for the difference of two symbols. GAS and the HP
7249 linker can't do this when one of the symbols is
7251 xoperands[1] = gen_label_rtx ();
7252 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
7253 output_asm_insn ("addil L'%0-%l1,%%r1", xoperands);
7254 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7255 CODE_LABEL_NUMBER (xoperands[1]));
7256 output_asm_insn ("ldo R'%0-%l1(%%r1),%%r1", xoperands);
7258 else if (TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call))
7260 /* GAS currently can't generate the relocations that
7261 are needed for the SOM linker under HP-UX using this
7262 sequence. The GNU linker doesn't generate the stubs
7263 that are needed for external calls on TARGET_ELF32
7264 with this sequence. For now, we have to use a
7265 longer plabel sequence when using GAS. */
7266 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
7267 output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1",
7269 output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1",
7274 /* Emit a long plabel-based call sequence. This is
7275 essentially an inline implementation of $$dyncall.
7276 We don't actually try to call $$dyncall as this is
7277 as difficult as calling the function itself. */
7278 xoperands[0] = get_deferred_plabel (call_dest);
7279 xoperands[1] = gen_label_rtx ();
7281 /* Since the call is indirect, FP arguments in registers
7282 need to be copied to the general registers. Then, the
7283 argument relocation stub will copy them back. */
7285 copy_fp_args (insn);
7289 output_asm_insn ("addil LT'%0,%%r19", xoperands);
7290 output_asm_insn ("ldw RT'%0(%%r1),%%r1", xoperands);
7291 output_asm_insn ("ldw 0(%%r1),%%r1", xoperands);
7295 output_asm_insn ("addil LR'%0-$global$,%%r27",
7297 output_asm_insn ("ldw RR'%0-$global$(%%r1),%%r1",
7301 output_asm_insn ("bb,>=,n %%r1,30,.+16", xoperands);
7302 output_asm_insn ("depi 0,31,2,%%r1", xoperands);
7303 output_asm_insn ("ldw 4(%%sr0,%%r1),%%r19", xoperands);
7304 output_asm_insn ("ldw 0(%%sr0,%%r1),%%r1", xoperands);
7306 if (!sibcall && !TARGET_PA_20)
7308 output_asm_insn ("{bl|b,l} .+8,%%r2", xoperands);
7309 if (TARGET_NO_SPACE_REGS)
7310 output_asm_insn ("addi 8,%%r2,%%r2", xoperands);
7312 output_asm_insn ("addi 16,%%r2,%%r2", xoperands);
7319 output_asm_insn ("bve (%%r1)", xoperands);
7324 output_asm_insn ("bve,l (%%r1),%%r2", xoperands);
7325 output_asm_insn ("stw %%r2,-24(%%sp)", xoperands);
7326 delay_slot_filled = 1;
7329 output_asm_insn ("bve,l (%%r1),%%r2", xoperands);
7334 if (!TARGET_NO_SPACE_REGS)
7335 output_asm_insn ("ldsid (%%r1),%%r31\n\tmtsp %%r31,%%sr0",
7340 if (TARGET_NO_SPACE_REGS)
7341 output_asm_insn ("be 0(%%sr4,%%r1)", xoperands);
7343 output_asm_insn ("be 0(%%sr0,%%r1)", xoperands);
7347 if (TARGET_NO_SPACE_REGS)
7348 output_asm_insn ("ble 0(%%sr4,%%r1)", xoperands);
7350 output_asm_insn ("ble 0(%%sr0,%%r1)", xoperands);
7353 output_asm_insn ("stw %%r31,-24(%%sp)", xoperands);
7355 output_asm_insn ("copy %%r31,%%r2", xoperands);
7356 delay_slot_filled = 1;
7363 if (!delay_slot_filled && (seq_length == 0 || delay_insn_deleted))
7364 output_asm_insn ("nop", xoperands);
7366 /* We are done if there isn't a jump in the delay slot. */
7368 || delay_insn_deleted
7369 || GET_CODE (NEXT_INSN (insn)) != JUMP_INSN)
7372 /* A sibcall should never have a branch in the delay slot. */
7373 gcc_assert (!sibcall);
7375 /* This call has an unconditional jump in its delay slot. */
7376 xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1);
7378 if (!delay_slot_filled && INSN_ADDRESSES_SET_P ())
7380 /* See if the return address can be adjusted. Use the containing
7381 sequence insn's address. */
7382 rtx seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0)));
7383 int distance = (INSN_ADDRESSES (INSN_UID (JUMP_LABEL (NEXT_INSN (insn))))
7384 - INSN_ADDRESSES (INSN_UID (seq_insn)) - 8);
7386 if (VAL_14_BITS_P (distance))
7388 xoperands[1] = gen_label_rtx ();
7389 output_asm_insn ("ldo %0-%1(%%r2),%%r2", xoperands);
7390 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7391 CODE_LABEL_NUMBER (xoperands[1]));
7394 output_asm_insn ("nop\n\tb,n %0", xoperands);
7397 output_asm_insn ("b,n %0", xoperands);
7399 /* Delete the jump. */
7400 PUT_CODE (NEXT_INSN (insn), NOTE);
7401 NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
7402 NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
7407 /* Return the attribute length of the indirect call instruction INSN.
7408 The length must match the code generated by output_indirect call.
7409 The returned length includes the delay slot. Currently, the delay
7410 slot of an indirect call sequence is not exposed and it is used by
7411 the sequence itself. */
7414 attr_length_indirect_call (rtx insn)
7416 unsigned long distance = -1;
7417 unsigned long total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes;
7419 if (INSN_ADDRESSES_SET_P ())
7421 distance = (total + insn_current_reference_address (insn));
7422 if (distance < total)
7429 if (TARGET_FAST_INDIRECT_CALLS
7430 || (!TARGET_PORTABLE_RUNTIME
7431 && ((TARGET_PA_20 && distance < 7600000) || distance < 240000)))
7437 if (TARGET_PORTABLE_RUNTIME)
7440 /* Out of reach, can use ble. */
7445 output_indirect_call (rtx insn, rtx call_dest)
7451 xoperands[0] = call_dest;
7452 output_asm_insn ("ldd 16(%0),%%r2", xoperands);
7453 output_asm_insn ("bve,l (%%r2),%%r2\n\tldd 24(%0),%%r27", xoperands);
7457 /* First the special case for kernels, level 0 systems, etc. */
7458 if (TARGET_FAST_INDIRECT_CALLS)
7459 return "ble 0(%%sr4,%%r22)\n\tcopy %%r31,%%r2";
7461 /* Now the normal case -- we can reach $$dyncall directly or
7462 we're sure that we can get there via a long-branch stub.
7464 No need to check target flags as the length uniquely identifies
7465 the remaining cases. */
7466 if (attr_length_indirect_call (insn) == 8)
7468 /* The HP linker substitutes a BLE for millicode calls using
7469 the short PIC PCREL form. Thus, we must use %r31 as the
7470 link register when generating PA 1.x code. */
7472 return ".CALL\tARGW0=GR\n\tb,l $$dyncall,%%r2\n\tcopy %%r2,%%r31";
7474 return ".CALL\tARGW0=GR\n\tbl $$dyncall,%%r31\n\tcopy %%r31,%%r2";
7477 /* Long millicode call, but we are not generating PIC or portable runtime
7479 if (attr_length_indirect_call (insn) == 12)
7480 return ".CALL\tARGW0=GR\n\tldil L'$$dyncall,%%r2\n\tble R'$$dyncall(%%sr4,%%r2)\n\tcopy %%r31,%%r2";
7482 /* Long millicode call for portable runtime. */
7483 if (attr_length_indirect_call (insn) == 20)
7484 return "ldil L'$$dyncall,%%r31\n\tldo R'$$dyncall(%%r31),%%r31\n\tblr %%r0,%%r2\n\tbv,n %%r0(%%r31)\n\tnop";
7486 /* We need a long PIC call to $$dyncall. */
7487 xoperands[0] = NULL_RTX;
7488 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
7489 if (TARGET_SOM || !TARGET_GAS)
7491 xoperands[0] = gen_label_rtx ();
7492 output_asm_insn ("addil L'$$dyncall-%0,%%r1", xoperands);
7493 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7494 CODE_LABEL_NUMBER (xoperands[0]));
7495 output_asm_insn ("ldo R'$$dyncall-%0(%%r1),%%r1", xoperands);
7499 output_asm_insn ("addil L'$$dyncall-$PIC_pcrel$0+4,%%r1", xoperands);
7500 output_asm_insn ("ldo R'$$dyncall-$PIC_pcrel$0+8(%%r1),%%r1",
7503 output_asm_insn ("blr %%r0,%%r2", xoperands);
7504 output_asm_insn ("bv,n %%r0(%%r1)\n\tnop", xoperands);
7508 /* Return the total length of the save and restore instructions needed for
7509 the data linkage table pointer (i.e., the PIC register) across the call
7510 instruction INSN. No-return calls do not require a save and restore.
7511 In addition, we may be able to avoid the save and restore for calls
7512 within the same translation unit. */
7515 attr_length_save_restore_dltp (rtx insn)
7517 if (find_reg_note (insn, REG_NORETURN, NULL_RTX))
7523 /* In HPUX 8.0's shared library scheme, special relocations are needed
7524 for function labels if they might be passed to a function
7525 in a shared library (because shared libraries don't live in code
7526 space), and special magic is needed to construct their address. */
7529 hppa_encode_label (rtx sym)
7531 const char *str = XSTR (sym, 0);
7532 int len = strlen (str) + 1;
7535 p = newstr = alloca (len + 1);
7539 XSTR (sym, 0) = ggc_alloc_string (newstr, len);
7543 pa_encode_section_info (tree decl, rtx rtl, int first)
7545 default_encode_section_info (decl, rtl, first);
7547 if (first && TEXT_SPACE_P (decl))
7549 SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1;
7550 if (TREE_CODE (decl) == FUNCTION_DECL)
7551 hppa_encode_label (XEXP (rtl, 0));
7555 /* This is sort of inverse to pa_encode_section_info. */
7558 pa_strip_name_encoding (const char *str)
7560 str += (*str == '@');
7561 str += (*str == '*');
7566 function_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
7568 return GET_CODE (op) == SYMBOL_REF && FUNCTION_NAME_P (XSTR (op, 0));
7571 /* Returns 1 if OP is a function label involved in a simple addition
7572 with a constant. Used to keep certain patterns from matching
7573 during instruction combination. */
7575 is_function_label_plus_const (rtx op)
7577 /* Strip off any CONST. */
7578 if (GET_CODE (op) == CONST)
7581 return (GET_CODE (op) == PLUS
7582 && function_label_operand (XEXP (op, 0), Pmode)
7583 && GET_CODE (XEXP (op, 1)) == CONST_INT);
7586 /* Output assembly code for a thunk to FUNCTION. */
7589 pa_asm_output_mi_thunk (FILE *file, tree thunk_fndecl, HOST_WIDE_INT delta,
7590 HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED,
7593 static unsigned int current_thunk_number;
7594 int val_14 = VAL_14_BITS_P (delta);
7599 xoperands[0] = XEXP (DECL_RTL (function), 0);
7600 xoperands[1] = XEXP (DECL_RTL (thunk_fndecl), 0);
7601 xoperands[2] = GEN_INT (delta);
7603 ASM_OUTPUT_LABEL (file, XSTR (xoperands[1], 0));
7604 fprintf (file, "\t.PROC\n\t.CALLINFO FRAME=0,NO_CALLS\n\t.ENTRY\n");
7606 /* Output the thunk. We know that the function is in the same
7607 translation unit (i.e., the same space) as the thunk, and that
7608 thunks are output after their method. Thus, we don't need an
7609 external branch to reach the function. With SOM and GAS,
7610 functions and thunks are effectively in different sections.
7611 Thus, we can always use a IA-relative branch and the linker
7612 will add a long branch stub if necessary.
7614 However, we have to be careful when generating PIC code on the
7615 SOM port to ensure that the sequence does not transfer to an
7616 import stub for the target function as this could clobber the
7617 return value saved at SP-24. This would also apply to the
7618 32-bit linux port if the multi-space model is implemented. */
7619 if ((!TARGET_LONG_CALLS && TARGET_SOM && !TARGET_PORTABLE_RUNTIME
7620 && !(flag_pic && TREE_PUBLIC (function))
7621 && (TARGET_GAS || last_address < 262132))
7622 || (!TARGET_LONG_CALLS && !TARGET_SOM && !TARGET_PORTABLE_RUNTIME
7623 && ((targetm.have_named_sections
7624 && DECL_SECTION_NAME (thunk_fndecl) != NULL
7625 /* The GNU 64-bit linker has rather poor stub management.
7626 So, we use a long branch from thunks that aren't in
7627 the same section as the target function. */
7629 && (DECL_SECTION_NAME (thunk_fndecl)
7630 != DECL_SECTION_NAME (function)))
7631 || ((DECL_SECTION_NAME (thunk_fndecl)
7632 == DECL_SECTION_NAME (function))
7633 && last_address < 262132)))
7634 || (!targetm.have_named_sections && last_address < 262132))))
7637 output_asm_insn ("addil L'%2,%%r26", xoperands);
7639 output_asm_insn ("b %0", xoperands);
7643 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7648 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7652 else if (TARGET_64BIT)
7654 /* We only have one call-clobbered scratch register, so we can't
7655 make use of the delay slot if delta doesn't fit in 14 bits. */
7658 output_asm_insn ("addil L'%2,%%r26", xoperands);
7659 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7662 output_asm_insn ("b,l .+8,%%r1", xoperands);
7666 output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands);
7667 output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands);
7671 xoperands[3] = GEN_INT (val_14 ? 8 : 16);
7672 output_asm_insn ("addil L'%0-%1-%3,%%r1", xoperands);
7677 output_asm_insn ("bv %%r0(%%r1)", xoperands);
7678 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7683 output_asm_insn ("bv,n %%r0(%%r1)", xoperands);
7687 else if (TARGET_PORTABLE_RUNTIME)
7689 output_asm_insn ("ldil L'%0,%%r1", xoperands);
7690 output_asm_insn ("ldo R'%0(%%r1),%%r22", xoperands);
7693 output_asm_insn ("addil L'%2,%%r26", xoperands);
7695 output_asm_insn ("bv %%r0(%%r22)", xoperands);
7699 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7704 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7708 else if (TARGET_SOM && flag_pic && TREE_PUBLIC (function))
7710 /* The function is accessible from outside this module. The only
7711 way to avoid an import stub between the thunk and function is to
7712 call the function directly with an indirect sequence similar to
7713 that used by $$dyncall. This is possible because $$dyncall acts
7714 as the import stub in an indirect call. */
7715 ASM_GENERATE_INTERNAL_LABEL (label, "LTHN", current_thunk_number);
7716 xoperands[3] = gen_rtx_SYMBOL_REF (Pmode, label);
7717 output_asm_insn ("addil LT'%3,%%r19", xoperands);
7718 output_asm_insn ("ldw RT'%3(%%r1),%%r22", xoperands);
7719 output_asm_insn ("ldw 0(%%sr0,%%r22),%%r22", xoperands);
7720 output_asm_insn ("bb,>=,n %%r22,30,.+16", xoperands);
7721 output_asm_insn ("depi 0,31,2,%%r22", xoperands);
7722 output_asm_insn ("ldw 4(%%sr0,%%r22),%%r19", xoperands);
7723 output_asm_insn ("ldw 0(%%sr0,%%r22),%%r22", xoperands);
7727 output_asm_insn ("addil L'%2,%%r26", xoperands);
7733 output_asm_insn ("bve (%%r22)", xoperands);
7736 else if (TARGET_NO_SPACE_REGS)
7738 output_asm_insn ("be 0(%%sr4,%%r22)", xoperands);
7743 output_asm_insn ("ldsid (%%sr0,%%r22),%%r21", xoperands);
7744 output_asm_insn ("mtsp %%r21,%%sr0", xoperands);
7745 output_asm_insn ("be 0(%%sr0,%%r22)", xoperands);
7750 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7752 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7756 output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
7758 if (TARGET_SOM || !TARGET_GAS)
7760 output_asm_insn ("addil L'%0-%1-8,%%r1", xoperands);
7761 output_asm_insn ("ldo R'%0-%1-8(%%r1),%%r22", xoperands);
7765 output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands);
7766 output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r22", xoperands);
7770 output_asm_insn ("addil L'%2,%%r26", xoperands);
7772 output_asm_insn ("bv %%r0(%%r22)", xoperands);
7776 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7781 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7788 output_asm_insn ("addil L'%2,%%r26", xoperands);
7790 output_asm_insn ("ldil L'%0,%%r22", xoperands);
7791 output_asm_insn ("be R'%0(%%sr4,%%r22)", xoperands);
7795 output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
7800 output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
7805 fprintf (file, "\t.EXIT\n\t.PROCEND\n");
7807 if (TARGET_SOM && TARGET_GAS)
7809 /* We done with this subspace except possibly for some additional
7810 debug information. Forget that we are in this subspace to ensure
7811 that the next function is output in its own subspace. */
7813 cfun->machine->in_nsubspa = 2;
7816 if (TARGET_SOM && flag_pic && TREE_PUBLIC (function))
7818 switch_to_section (data_section);
7819 output_asm_insn (".align 4", xoperands);
7820 ASM_OUTPUT_LABEL (file, label);
7821 output_asm_insn (".word P'%0", xoperands);
7824 current_thunk_number++;
7825 nbytes = ((nbytes + FUNCTION_BOUNDARY / BITS_PER_UNIT - 1)
7826 & ~(FUNCTION_BOUNDARY / BITS_PER_UNIT - 1));
7827 last_address += nbytes;
7828 update_total_code_bytes (nbytes);
7831 /* Only direct calls to static functions are allowed to be sibling (tail)
7834 This restriction is necessary because some linker generated stubs will
7835 store return pointers into rp' in some cases which might clobber a
7836 live value already in rp'.
7838 In a sibcall the current function and the target function share stack
7839 space. Thus if the path to the current function and the path to the
7840 target function save a value in rp', they save the value into the
7841 same stack slot, which has undesirable consequences.
7843 Because of the deferred binding nature of shared libraries any function
7844 with external scope could be in a different load module and thus require
7845 rp' to be saved when calling that function. So sibcall optimizations
7846 can only be safe for static function.
7848 Note that GCC never needs return value relocations, so we don't have to
7849 worry about static calls with return value relocations (which require
7852 It is safe to perform a sibcall optimization when the target function
7853 will never return. */
7855 pa_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
7857 if (TARGET_PORTABLE_RUNTIME)
7860 /* Sibcalls are ok for TARGET_ELF32 as along as the linker is used in
7861 single subspace mode and the call is not indirect. As far as I know,
7862 there is no operating system support for the multiple subspace mode.
7863 It might be possible to support indirect calls if we didn't use
7864 $$dyncall (see the indirect sequence generated in output_call). */
7866 return (decl != NULL_TREE);
7868 /* Sibcalls are not ok because the arg pointer register is not a fixed
7869 register. This prevents the sibcall optimization from occurring. In
7870 addition, there are problems with stub placement using GNU ld. This
7871 is because a normal sibcall branch uses a 17-bit relocation while
7872 a regular call branch uses a 22-bit relocation. As a result, more
7873 care needs to be taken in the placement of long-branch stubs. */
7877 /* Sibcalls are only ok within a translation unit. */
7878 return (decl && !TREE_PUBLIC (decl));
7881 /* ??? Addition is not commutative on the PA due to the weird implicit
7882 space register selection rules for memory addresses. Therefore, we
7883 don't consider a + b == b + a, as this might be inside a MEM. */
7885 pa_commutative_p (rtx x, int outer_code)
7887 return (COMMUTATIVE_P (x)
7888 && (TARGET_NO_SPACE_REGS
7889 || (outer_code != UNKNOWN && outer_code != MEM)
7890 || GET_CODE (x) != PLUS));
7893 /* Returns 1 if the 6 operands specified in OPERANDS are suitable for
7894 use in fmpyadd instructions. */
7896 fmpyaddoperands (rtx *operands)
7898 enum machine_mode mode = GET_MODE (operands[0]);
7900 /* Must be a floating point mode. */
7901 if (mode != SFmode && mode != DFmode)
7904 /* All modes must be the same. */
7905 if (! (mode == GET_MODE (operands[1])
7906 && mode == GET_MODE (operands[2])
7907 && mode == GET_MODE (operands[3])
7908 && mode == GET_MODE (operands[4])
7909 && mode == GET_MODE (operands[5])))
7912 /* All operands must be registers. */
7913 if (! (GET_CODE (operands[1]) == REG
7914 && GET_CODE (operands[2]) == REG
7915 && GET_CODE (operands[3]) == REG
7916 && GET_CODE (operands[4]) == REG
7917 && GET_CODE (operands[5]) == REG))
7920 /* Only 2 real operands to the addition. One of the input operands must
7921 be the same as the output operand. */
7922 if (! rtx_equal_p (operands[3], operands[4])
7923 && ! rtx_equal_p (operands[3], operands[5]))
7926 /* Inout operand of add cannot conflict with any operands from multiply. */
7927 if (rtx_equal_p (operands[3], operands[0])
7928 || rtx_equal_p (operands[3], operands[1])
7929 || rtx_equal_p (operands[3], operands[2]))
7932 /* multiply cannot feed into addition operands. */
7933 if (rtx_equal_p (operands[4], operands[0])
7934 || rtx_equal_p (operands[5], operands[0]))
7937 /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */
7939 && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS
7940 || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS
7941 || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS
7942 || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS
7943 || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS
7944 || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS))
7947 /* Passed. Operands are suitable for fmpyadd. */
7951 #if !defined(USE_COLLECT2)
7953 pa_asm_out_constructor (rtx symbol, int priority)
7955 if (!function_label_operand (symbol, VOIDmode))
7956 hppa_encode_label (symbol);
7958 #ifdef CTORS_SECTION_ASM_OP
7959 default_ctor_section_asm_out_constructor (symbol, priority);
7961 # ifdef TARGET_ASM_NAMED_SECTION
7962 default_named_section_asm_out_constructor (symbol, priority);
7964 default_stabs_asm_out_constructor (symbol, priority);
7970 pa_asm_out_destructor (rtx symbol, int priority)
7972 if (!function_label_operand (symbol, VOIDmode))
7973 hppa_encode_label (symbol);
7975 #ifdef DTORS_SECTION_ASM_OP
7976 default_dtor_section_asm_out_destructor (symbol, priority);
7978 # ifdef TARGET_ASM_NAMED_SECTION
7979 default_named_section_asm_out_destructor (symbol, priority);
7981 default_stabs_asm_out_destructor (symbol, priority);
7987 /* This function places uninitialized global data in the bss section.
7988 The ASM_OUTPUT_ALIGNED_BSS macro needs to be defined to call this
7989 function on the SOM port to prevent uninitialized global data from
7990 being placed in the data section. */
7993 pa_asm_output_aligned_bss (FILE *stream,
7995 unsigned HOST_WIDE_INT size,
7998 switch_to_section (bss_section);
7999 fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
8001 #ifdef ASM_OUTPUT_TYPE_DIRECTIVE
8002 ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "object");
8005 #ifdef ASM_OUTPUT_SIZE_DIRECTIVE
8006 ASM_OUTPUT_SIZE_DIRECTIVE (stream, name, size);
8009 fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
8010 ASM_OUTPUT_LABEL (stream, name);
8011 fprintf (stream, "\t.block "HOST_WIDE_INT_PRINT_UNSIGNED"\n", size);
8014 /* Both the HP and GNU assemblers under HP-UX provide a .comm directive
8015 that doesn't allow the alignment of global common storage to be directly
8016 specified. The SOM linker aligns common storage based on the rounded
8017 value of the NUM_BYTES parameter in the .comm directive. It's not
8018 possible to use the .align directive as it doesn't affect the alignment
8019 of the label associated with a .comm directive. */
8022 pa_asm_output_aligned_common (FILE *stream,
8024 unsigned HOST_WIDE_INT size,
8027 unsigned int max_common_align;
8029 max_common_align = TARGET_64BIT ? 128 : (size >= 4096 ? 256 : 64);
8030 if (align > max_common_align)
8032 warning (0, "alignment (%u) for %s exceeds maximum alignment "
8033 "for global common data. Using %u",
8034 align / BITS_PER_UNIT, name, max_common_align / BITS_PER_UNIT);
8035 align = max_common_align;
8038 switch_to_section (bss_section);
8040 assemble_name (stream, name);
8041 fprintf (stream, "\t.comm "HOST_WIDE_INT_PRINT_UNSIGNED"\n",
8042 MAX (size, align / BITS_PER_UNIT));
8045 /* We can't use .comm for local common storage as the SOM linker effectively
8046 treats the symbol as universal and uses the same storage for local symbols
8047 with the same name in different object files. The .block directive
8048 reserves an uninitialized block of storage. However, it's not common
8049 storage. Fortunately, GCC never requests common storage with the same
8050 name in any given translation unit. */
8053 pa_asm_output_aligned_local (FILE *stream,
8055 unsigned HOST_WIDE_INT size,
8058 switch_to_section (bss_section);
8059 fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
8062 fprintf (stream, "%s", LOCAL_ASM_OP);
8063 assemble_name (stream, name);
8064 fprintf (stream, "\n");
8067 ASM_OUTPUT_LABEL (stream, name);
8068 fprintf (stream, "\t.block "HOST_WIDE_INT_PRINT_UNSIGNED"\n", size);
8071 /* Returns 1 if the 6 operands specified in OPERANDS are suitable for
8072 use in fmpysub instructions. */
8074 fmpysuboperands (rtx *operands)
8076 enum machine_mode mode = GET_MODE (operands[0]);
8078 /* Must be a floating point mode. */
8079 if (mode != SFmode && mode != DFmode)
8082 /* All modes must be the same. */
8083 if (! (mode == GET_MODE (operands[1])
8084 && mode == GET_MODE (operands[2])
8085 && mode == GET_MODE (operands[3])
8086 && mode == GET_MODE (operands[4])
8087 && mode == GET_MODE (operands[5])))
8090 /* All operands must be registers. */
8091 if (! (GET_CODE (operands[1]) == REG
8092 && GET_CODE (operands[2]) == REG
8093 && GET_CODE (operands[3]) == REG
8094 && GET_CODE (operands[4]) == REG
8095 && GET_CODE (operands[5]) == REG))
8098 /* Only 2 real operands to the subtraction. Subtraction is not a commutative
8099 operation, so operands[4] must be the same as operand[3]. */
8100 if (! rtx_equal_p (operands[3], operands[4]))
8103 /* multiply cannot feed into subtraction. */
8104 if (rtx_equal_p (operands[5], operands[0]))
8107 /* Inout operand of sub cannot conflict with any operands from multiply. */
8108 if (rtx_equal_p (operands[3], operands[0])
8109 || rtx_equal_p (operands[3], operands[1])
8110 || rtx_equal_p (operands[3], operands[2]))
8113 /* SFmode limits the registers to the upper 32 of the 32bit FP regs. */
8115 && (REGNO_REG_CLASS (REGNO (operands[0])) != FPUPPER_REGS
8116 || REGNO_REG_CLASS (REGNO (operands[1])) != FPUPPER_REGS
8117 || REGNO_REG_CLASS (REGNO (operands[2])) != FPUPPER_REGS
8118 || REGNO_REG_CLASS (REGNO (operands[3])) != FPUPPER_REGS
8119 || REGNO_REG_CLASS (REGNO (operands[4])) != FPUPPER_REGS
8120 || REGNO_REG_CLASS (REGNO (operands[5])) != FPUPPER_REGS))
8123 /* Passed. Operands are suitable for fmpysub. */
8127 /* Return 1 if the given constant is 2, 4, or 8. These are the valid
8128 constants for shadd instructions. */
8130 shadd_constant_p (int val)
8132 if (val == 2 || val == 4 || val == 8)
8138 /* Return 1 if OP is valid as a base or index register in a
8142 borx_reg_operand (rtx op, enum machine_mode mode)
8144 if (GET_CODE (op) != REG)
8147 /* We must reject virtual registers as the only expressions that
8148 can be instantiated are REG and REG+CONST. */
8149 if (op == virtual_incoming_args_rtx
8150 || op == virtual_stack_vars_rtx
8151 || op == virtual_stack_dynamic_rtx
8152 || op == virtual_outgoing_args_rtx
8153 || op == virtual_cfa_rtx)
8156 /* While it's always safe to index off the frame pointer, it's not
8157 profitable to do so when the frame pointer is being eliminated. */
8158 if (!reload_completed
8159 && flag_omit_frame_pointer
8160 && !current_function_calls_alloca
8161 && op == frame_pointer_rtx)
8164 return register_operand (op, mode);
8167 /* Return 1 if this operand is anything other than a hard register. */
8170 non_hard_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
8172 return ! (GET_CODE (op) == REG && REGNO (op) < FIRST_PSEUDO_REGISTER);
8175 /* Return 1 if INSN branches forward. Should be using insn_addresses
8176 to avoid walking through all the insns... */
8178 forward_branch_p (rtx insn)
8180 rtx label = JUMP_LABEL (insn);
8187 insn = NEXT_INSN (insn);
8190 return (insn == label);
8193 /* Return 1 if OP is an equality comparison, else return 0. */
8195 eq_neq_comparison_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
8197 return (GET_CODE (op) == EQ || GET_CODE (op) == NE);
8200 /* Return 1 if INSN is in the delay slot of a call instruction. */
8202 jump_in_call_delay (rtx insn)
8205 if (GET_CODE (insn) != JUMP_INSN)
8208 if (PREV_INSN (insn)
8209 && PREV_INSN (PREV_INSN (insn))
8210 && GET_CODE (next_real_insn (PREV_INSN (PREV_INSN (insn)))) == INSN)
8212 rtx test_insn = next_real_insn (PREV_INSN (PREV_INSN (insn)));
8214 return (GET_CODE (PATTERN (test_insn)) == SEQUENCE
8215 && XVECEXP (PATTERN (test_insn), 0, 1) == insn);
8222 /* Output an unconditional move and branch insn. */
8225 output_parallel_movb (rtx *operands, int length)
8227 /* These are the cases in which we win. */
8229 return "mov%I1b,tr %1,%0,%2";
8231 /* None of these cases wins, but they don't lose either. */
8232 if (dbr_sequence_length () == 0)
8234 /* Nothing in the delay slot, fake it by putting the combined
8235 insn (the copy or add) in the delay slot of a bl. */
8236 if (GET_CODE (operands[1]) == CONST_INT)
8237 return "b %2\n\tldi %1,%0";
8239 return "b %2\n\tcopy %1,%0";
8243 /* Something in the delay slot, but we've got a long branch. */
8244 if (GET_CODE (operands[1]) == CONST_INT)
8245 return "ldi %1,%0\n\tb %2";
8247 return "copy %1,%0\n\tb %2";
8251 /* Output an unconditional add and branch insn. */
8254 output_parallel_addb (rtx *operands, int length)
8256 /* To make life easy we want operand0 to be the shared input/output
8257 operand and operand1 to be the readonly operand. */
8258 if (operands[0] == operands[1])
8259 operands[1] = operands[2];
8261 /* These are the cases in which we win. */
8263 return "add%I1b,tr %1,%0,%3";
8265 /* None of these cases win, but they don't lose either. */
8266 if (dbr_sequence_length () == 0)
8268 /* Nothing in the delay slot, fake it by putting the combined
8269 insn (the copy or add) in the delay slot of a bl. */
8270 return "b %3\n\tadd%I1 %1,%0,%0";
8274 /* Something in the delay slot, but we've got a long branch. */
8275 return "add%I1 %1,%0,%0\n\tb %3";
8279 /* Return nonzero if INSN (a jump insn) immediately follows a call
8280 to a named function. This is used to avoid filling the delay slot
8281 of the jump since it can usually be eliminated by modifying RP in
8282 the delay slot of the call. */
8285 following_call (rtx insn)
8287 if (! TARGET_JUMP_IN_DELAY)
8290 /* Find the previous real insn, skipping NOTEs. */
8291 insn = PREV_INSN (insn);
8292 while (insn && GET_CODE (insn) == NOTE)
8293 insn = PREV_INSN (insn);
8295 /* Check for CALL_INSNs and millicode calls. */
8297 && ((GET_CODE (insn) == CALL_INSN
8298 && get_attr_type (insn) != TYPE_DYNCALL)
8299 || (GET_CODE (insn) == INSN
8300 && GET_CODE (PATTERN (insn)) != SEQUENCE
8301 && GET_CODE (PATTERN (insn)) != USE
8302 && GET_CODE (PATTERN (insn)) != CLOBBER
8303 && get_attr_type (insn) == TYPE_MILLI)))
8309 /* We use this hook to perform a PA specific optimization which is difficult
8310 to do in earlier passes.
8312 We want the delay slots of branches within jump tables to be filled.
8313 None of the compiler passes at the moment even has the notion that a
8314 PA jump table doesn't contain addresses, but instead contains actual
8317 Because we actually jump into the table, the addresses of each entry
8318 must stay constant in relation to the beginning of the table (which
8319 itself must stay constant relative to the instruction to jump into
8320 it). I don't believe we can guarantee earlier passes of the compiler
8321 will adhere to those rules.
8323 So, late in the compilation process we find all the jump tables, and
8324 expand them into real code -- e.g. each entry in the jump table vector
8325 will get an appropriate label followed by a jump to the final target.
8327 Reorg and the final jump pass can then optimize these branches and
8328 fill their delay slots. We end up with smaller, more efficient code.
8330 The jump instructions within the table are special; we must be able
8331 to identify them during assembly output (if the jumps don't get filled
8332 we need to emit a nop rather than nullifying the delay slot)). We
8333 identify jumps in switch tables by using insns with the attribute
8334 type TYPE_BTABLE_BRANCH.
8336 We also surround the jump table itself with BEGIN_BRTAB and END_BRTAB
8337 insns. This serves two purposes, first it prevents jump.c from
8338 noticing that the last N entries in the table jump to the instruction
8339 immediately after the table and deleting the jumps. Second, those
8340 insns mark where we should emit .begin_brtab and .end_brtab directives
8341 when using GAS (allows for better link time optimizations). */
8348 remove_useless_addtr_insns (1);
8350 if (pa_cpu < PROCESSOR_8000)
8351 pa_combine_instructions ();
8354 /* This is fairly cheap, so always run it if optimizing. */
8355 if (optimize > 0 && !TARGET_BIG_SWITCH)
8357 /* Find and explode all ADDR_VEC or ADDR_DIFF_VEC insns. */
8358 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
8360 rtx pattern, tmp, location, label;
8361 unsigned int length, i;
8363 /* Find an ADDR_VEC or ADDR_DIFF_VEC insn to explode. */
8364 if (GET_CODE (insn) != JUMP_INSN
8365 || (GET_CODE (PATTERN (insn)) != ADDR_VEC
8366 && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC))
8369 /* Emit marker for the beginning of the branch table. */
8370 emit_insn_before (gen_begin_brtab (), insn);
8372 pattern = PATTERN (insn);
8373 location = PREV_INSN (insn);
8374 length = XVECLEN (pattern, GET_CODE (pattern) == ADDR_DIFF_VEC);
8376 for (i = 0; i < length; i++)
8378 /* Emit a label before each jump to keep jump.c from
8379 removing this code. */
8380 tmp = gen_label_rtx ();
8381 LABEL_NUSES (tmp) = 1;
8382 emit_label_after (tmp, location);
8383 location = NEXT_INSN (location);
8385 if (GET_CODE (pattern) == ADDR_VEC)
8386 label = XEXP (XVECEXP (pattern, 0, i), 0);
8388 label = XEXP (XVECEXP (pattern, 1, i), 0);
8390 tmp = gen_short_jump (label);
8392 /* Emit the jump itself. */
8393 tmp = emit_jump_insn_after (tmp, location);
8394 JUMP_LABEL (tmp) = label;
8395 LABEL_NUSES (label)++;
8396 location = NEXT_INSN (location);
8398 /* Emit a BARRIER after the jump. */
8399 emit_barrier_after (location);
8400 location = NEXT_INSN (location);
8403 /* Emit marker for the end of the branch table. */
8404 emit_insn_before (gen_end_brtab (), location);
8405 location = NEXT_INSN (location);
8406 emit_barrier_after (location);
8408 /* Delete the ADDR_VEC or ADDR_DIFF_VEC. */
8414 /* Still need brtab marker insns. FIXME: the presence of these
8415 markers disables output of the branch table to readonly memory,
8416 and any alignment directives that might be needed. Possibly,
8417 the begin_brtab insn should be output before the label for the
8418 table. This doesn't matter at the moment since the tables are
8419 always output in the text section. */
8420 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
8422 /* Find an ADDR_VEC insn. */
8423 if (GET_CODE (insn) != JUMP_INSN
8424 || (GET_CODE (PATTERN (insn)) != ADDR_VEC
8425 && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC))
8428 /* Now generate markers for the beginning and end of the
8430 emit_insn_before (gen_begin_brtab (), insn);
8431 emit_insn_after (gen_end_brtab (), insn);
8436 /* The PA has a number of odd instructions which can perform multiple
8437 tasks at once. On first generation PA machines (PA1.0 and PA1.1)
8438 it may be profitable to combine two instructions into one instruction
8439 with two outputs. It's not profitable PA2.0 machines because the
8440 two outputs would take two slots in the reorder buffers.
8442 This routine finds instructions which can be combined and combines
8443 them. We only support some of the potential combinations, and we
8444 only try common ways to find suitable instructions.
8446 * addb can add two registers or a register and a small integer
8447 and jump to a nearby (+-8k) location. Normally the jump to the
8448 nearby location is conditional on the result of the add, but by
8449 using the "true" condition we can make the jump unconditional.
8450 Thus addb can perform two independent operations in one insn.
8452 * movb is similar to addb in that it can perform a reg->reg
8453 or small immediate->reg copy and jump to a nearby (+-8k location).
8455 * fmpyadd and fmpysub can perform a FP multiply and either an
8456 FP add or FP sub if the operands of the multiply and add/sub are
8457 independent (there are other minor restrictions). Note both
8458 the fmpy and fadd/fsub can in theory move to better spots according
8459 to data dependencies, but for now we require the fmpy stay at a
8462 * Many of the memory operations can perform pre & post updates
8463 of index registers. GCC's pre/post increment/decrement addressing
8464 is far too simple to take advantage of all the possibilities. This
8465 pass may not be suitable since those insns may not be independent.
8467 * comclr can compare two ints or an int and a register, nullify
8468 the following instruction and zero some other register. This
8469 is more difficult to use as it's harder to find an insn which
8470 will generate a comclr than finding something like an unconditional
8471 branch. (conditional moves & long branches create comclr insns).
8473 * Most arithmetic operations can conditionally skip the next
8474 instruction. They can be viewed as "perform this operation
8475 and conditionally jump to this nearby location" (where nearby
8476 is an insns away). These are difficult to use due to the
8477 branch length restrictions. */
8480 pa_combine_instructions (void)
8484 /* This can get expensive since the basic algorithm is on the
8485 order of O(n^2) (or worse). Only do it for -O2 or higher
8486 levels of optimization. */
8490 /* Walk down the list of insns looking for "anchor" insns which
8491 may be combined with "floating" insns. As the name implies,
8492 "anchor" instructions don't move, while "floating" insns may
8494 new = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, NULL_RTX, NULL_RTX));
8495 new = make_insn_raw (new);
8497 for (anchor = get_insns (); anchor; anchor = NEXT_INSN (anchor))
8499 enum attr_pa_combine_type anchor_attr;
8500 enum attr_pa_combine_type floater_attr;
8502 /* We only care about INSNs, JUMP_INSNs, and CALL_INSNs.
8503 Also ignore any special USE insns. */
8504 if ((GET_CODE (anchor) != INSN
8505 && GET_CODE (anchor) != JUMP_INSN
8506 && GET_CODE (anchor) != CALL_INSN)
8507 || GET_CODE (PATTERN (anchor)) == USE
8508 || GET_CODE (PATTERN (anchor)) == CLOBBER
8509 || GET_CODE (PATTERN (anchor)) == ADDR_VEC
8510 || GET_CODE (PATTERN (anchor)) == ADDR_DIFF_VEC)
8513 anchor_attr = get_attr_pa_combine_type (anchor);
8514 /* See if anchor is an insn suitable for combination. */
8515 if (anchor_attr == PA_COMBINE_TYPE_FMPY
8516 || anchor_attr == PA_COMBINE_TYPE_FADDSUB
8517 || (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH
8518 && ! forward_branch_p (anchor)))
8522 for (floater = PREV_INSN (anchor);
8524 floater = PREV_INSN (floater))
8526 if (GET_CODE (floater) == NOTE
8527 || (GET_CODE (floater) == INSN
8528 && (GET_CODE (PATTERN (floater)) == USE
8529 || GET_CODE (PATTERN (floater)) == CLOBBER)))
8532 /* Anything except a regular INSN will stop our search. */
8533 if (GET_CODE (floater) != INSN
8534 || GET_CODE (PATTERN (floater)) == ADDR_VEC
8535 || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC)
8541 /* See if FLOATER is suitable for combination with the
8543 floater_attr = get_attr_pa_combine_type (floater);
8544 if ((anchor_attr == PA_COMBINE_TYPE_FMPY
8545 && floater_attr == PA_COMBINE_TYPE_FADDSUB)
8546 || (anchor_attr == PA_COMBINE_TYPE_FADDSUB
8547 && floater_attr == PA_COMBINE_TYPE_FMPY))
8549 /* If ANCHOR and FLOATER can be combined, then we're
8550 done with this pass. */
8551 if (pa_can_combine_p (new, anchor, floater, 0,
8552 SET_DEST (PATTERN (floater)),
8553 XEXP (SET_SRC (PATTERN (floater)), 0),
8554 XEXP (SET_SRC (PATTERN (floater)), 1)))
8558 else if (anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH
8559 && floater_attr == PA_COMBINE_TYPE_ADDMOVE)
8561 if (GET_CODE (SET_SRC (PATTERN (floater))) == PLUS)
8563 if (pa_can_combine_p (new, anchor, floater, 0,
8564 SET_DEST (PATTERN (floater)),
8565 XEXP (SET_SRC (PATTERN (floater)), 0),
8566 XEXP (SET_SRC (PATTERN (floater)), 1)))
8571 if (pa_can_combine_p (new, anchor, floater, 0,
8572 SET_DEST (PATTERN (floater)),
8573 SET_SRC (PATTERN (floater)),
8574 SET_SRC (PATTERN (floater))))
8580 /* If we didn't find anything on the backwards scan try forwards. */
8582 && (anchor_attr == PA_COMBINE_TYPE_FMPY
8583 || anchor_attr == PA_COMBINE_TYPE_FADDSUB))
8585 for (floater = anchor; floater; floater = NEXT_INSN (floater))
8587 if (GET_CODE (floater) == NOTE
8588 || (GET_CODE (floater) == INSN
8589 && (GET_CODE (PATTERN (floater)) == USE
8590 || GET_CODE (PATTERN (floater)) == CLOBBER)))
8594 /* Anything except a regular INSN will stop our search. */
8595 if (GET_CODE (floater) != INSN
8596 || GET_CODE (PATTERN (floater)) == ADDR_VEC
8597 || GET_CODE (PATTERN (floater)) == ADDR_DIFF_VEC)
8603 /* See if FLOATER is suitable for combination with the
8605 floater_attr = get_attr_pa_combine_type (floater);
8606 if ((anchor_attr == PA_COMBINE_TYPE_FMPY
8607 && floater_attr == PA_COMBINE_TYPE_FADDSUB)
8608 || (anchor_attr == PA_COMBINE_TYPE_FADDSUB
8609 && floater_attr == PA_COMBINE_TYPE_FMPY))
8611 /* If ANCHOR and FLOATER can be combined, then we're
8612 done with this pass. */
8613 if (pa_can_combine_p (new, anchor, floater, 1,
8614 SET_DEST (PATTERN (floater)),
8615 XEXP (SET_SRC (PATTERN (floater)),
8617 XEXP (SET_SRC (PATTERN (floater)),
8624 /* FLOATER will be nonzero if we found a suitable floating
8625 insn for combination with ANCHOR. */
8627 && (anchor_attr == PA_COMBINE_TYPE_FADDSUB
8628 || anchor_attr == PA_COMBINE_TYPE_FMPY))
8630 /* Emit the new instruction and delete the old anchor. */
8631 emit_insn_before (gen_rtx_PARALLEL
8633 gen_rtvec (2, PATTERN (anchor),
8634 PATTERN (floater))),
8637 PUT_CODE (anchor, NOTE);
8638 NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED;
8639 NOTE_SOURCE_FILE (anchor) = 0;
8641 /* Emit a special USE insn for FLOATER, then delete
8642 the floating insn. */
8643 emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater);
8644 delete_insn (floater);
8649 && anchor_attr == PA_COMBINE_TYPE_UNCOND_BRANCH)
8652 /* Emit the new_jump instruction and delete the old anchor. */
8654 = emit_jump_insn_before (gen_rtx_PARALLEL
8656 gen_rtvec (2, PATTERN (anchor),
8657 PATTERN (floater))),
8660 JUMP_LABEL (temp) = JUMP_LABEL (anchor);
8661 PUT_CODE (anchor, NOTE);
8662 NOTE_LINE_NUMBER (anchor) = NOTE_INSN_DELETED;
8663 NOTE_SOURCE_FILE (anchor) = 0;
8665 /* Emit a special USE insn for FLOATER, then delete
8666 the floating insn. */
8667 emit_insn_before (gen_rtx_USE (VOIDmode, floater), floater);
8668 delete_insn (floater);
8676 pa_can_combine_p (rtx new, rtx anchor, rtx floater, int reversed, rtx dest,
8679 int insn_code_number;
8682 /* Create a PARALLEL with the patterns of ANCHOR and
8683 FLOATER, try to recognize it, then test constraints
8684 for the resulting pattern.
8686 If the pattern doesn't match or the constraints
8687 aren't met keep searching for a suitable floater
8689 XVECEXP (PATTERN (new), 0, 0) = PATTERN (anchor);
8690 XVECEXP (PATTERN (new), 0, 1) = PATTERN (floater);
8691 INSN_CODE (new) = -1;
8692 insn_code_number = recog_memoized (new);
8693 if (insn_code_number < 0
8694 || (extract_insn (new), ! constrain_operands (1)))
8708 /* There's up to three operands to consider. One
8709 output and two inputs.
8711 The output must not be used between FLOATER & ANCHOR
8712 exclusive. The inputs must not be set between
8713 FLOATER and ANCHOR exclusive. */
8715 if (reg_used_between_p (dest, start, end))
8718 if (reg_set_between_p (src1, start, end))
8721 if (reg_set_between_p (src2, start, end))
8724 /* If we get here, then everything is good. */
8728 /* Return nonzero if references for INSN are delayed.
8730 Millicode insns are actually function calls with some special
8731 constraints on arguments and register usage.
8733 Millicode calls always expect their arguments in the integer argument
8734 registers, and always return their result in %r29 (ret1). They
8735 are expected to clobber their arguments, %r1, %r29, and the return
8736 pointer which is %r31 on 32-bit and %r2 on 64-bit, and nothing else.
8738 This function tells reorg that the references to arguments and
8739 millicode calls do not appear to happen until after the millicode call.
8740 This allows reorg to put insns which set the argument registers into the
8741 delay slot of the millicode call -- thus they act more like traditional
8744 Note we cannot consider side effects of the insn to be delayed because
8745 the branch and link insn will clobber the return pointer. If we happened
8746 to use the return pointer in the delay slot of the call, then we lose.
8748 get_attr_type will try to recognize the given insn, so make sure to
8749 filter out things it will not accept -- SEQUENCE, USE and CLOBBER insns
8752 insn_refs_are_delayed (rtx insn)
8754 return ((GET_CODE (insn) == INSN
8755 && GET_CODE (PATTERN (insn)) != SEQUENCE
8756 && GET_CODE (PATTERN (insn)) != USE
8757 && GET_CODE (PATTERN (insn)) != CLOBBER
8758 && get_attr_type (insn) == TYPE_MILLI));
8761 /* On the HP-PA the value is found in register(s) 28(-29), unless
8762 the mode is SF or DF. Then the value is returned in fr4 (32).
8764 This must perform the same promotions as PROMOTE_MODE, else
8765 TARGET_PROMOTE_FUNCTION_RETURN will not work correctly.
8767 Small structures must be returned in a PARALLEL on PA64 in order
8768 to match the HP Compiler ABI. */
8771 function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
8773 enum machine_mode valmode;
8775 if (AGGREGATE_TYPE_P (valtype))
8779 /* Aggregates with a size less than or equal to 128 bits are
8780 returned in GR 28(-29). They are left justified. The pad
8781 bits are undefined. Larger aggregates are returned in
8785 int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2;
8787 for (i = 0; i < ub; i++)
8789 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
8790 gen_rtx_REG (DImode, 28 + i),
8795 return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc));
8797 else if (int_size_in_bytes (valtype) > UNITS_PER_WORD)
8799 /* Aggregates 5 to 8 bytes in size are returned in general
8800 registers r28-r29 in the same manner as other non
8801 floating-point objects. The data is right-justified and
8802 zero-extended to 64 bits. This is opposite to the normal
8803 justification used on big endian targets and requires
8804 special treatment. */
8805 rtx loc = gen_rtx_EXPR_LIST (VOIDmode,
8806 gen_rtx_REG (DImode, 28), const0_rtx);
8807 return gen_rtx_PARALLEL (BLKmode, gen_rtvec (1, loc));
8811 if ((INTEGRAL_TYPE_P (valtype)
8812 && TYPE_PRECISION (valtype) < BITS_PER_WORD)
8813 || POINTER_TYPE_P (valtype))
8814 valmode = word_mode;
8816 valmode = TYPE_MODE (valtype);
8818 if (TREE_CODE (valtype) == REAL_TYPE
8819 && !AGGREGATE_TYPE_P (valtype)
8820 && TYPE_MODE (valtype) != TFmode
8821 && !TARGET_SOFT_FLOAT)
8822 return gen_rtx_REG (valmode, 32);
8824 return gen_rtx_REG (valmode, 28);
8827 /* Return the location of a parameter that is passed in a register or NULL
8828 if the parameter has any component that is passed in memory.
8830 This is new code and will be pushed to into the net sources after
8833 ??? We might want to restructure this so that it looks more like other
8836 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
8837 int named ATTRIBUTE_UNUSED)
8839 int max_arg_words = (TARGET_64BIT ? 8 : 4);
8846 if (mode == VOIDmode)
8849 arg_size = FUNCTION_ARG_SIZE (mode, type);
8851 /* If this arg would be passed partially or totally on the stack, then
8852 this routine should return zero. pa_arg_partial_bytes will
8853 handle arguments which are split between regs and stack slots if
8854 the ABI mandates split arguments. */
8857 /* The 32-bit ABI does not split arguments. */
8858 if (cum->words + arg_size > max_arg_words)
8864 alignment = cum->words & 1;
8865 if (cum->words + alignment >= max_arg_words)
8869 /* The 32bit ABIs and the 64bit ABIs are rather different,
8870 particularly in their handling of FP registers. We might
8871 be able to cleverly share code between them, but I'm not
8872 going to bother in the hope that splitting them up results
8873 in code that is more easily understood. */
8877 /* Advance the base registers to their current locations.
8879 Remember, gprs grow towards smaller register numbers while
8880 fprs grow to higher register numbers. Also remember that
8881 although FP regs are 32-bit addressable, we pretend that
8882 the registers are 64-bits wide. */
8883 gpr_reg_base = 26 - cum->words;
8884 fpr_reg_base = 32 + cum->words;
8886 /* Arguments wider than one word and small aggregates need special
8890 || (type && AGGREGATE_TYPE_P (type)))
8892 /* Double-extended precision (80-bit), quad-precision (128-bit)
8893 and aggregates including complex numbers are aligned on
8894 128-bit boundaries. The first eight 64-bit argument slots
8895 are associated one-to-one, with general registers r26
8896 through r19, and also with floating-point registers fr4
8897 through fr11. Arguments larger than one word are always
8898 passed in general registers.
8900 Using a PARALLEL with a word mode register results in left
8901 justified data on a big-endian target. */
8904 int i, offset = 0, ub = arg_size;
8906 /* Align the base register. */
8907 gpr_reg_base -= alignment;
8909 ub = MIN (ub, max_arg_words - cum->words - alignment);
8910 for (i = 0; i < ub; i++)
8912 loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
8913 gen_rtx_REG (DImode, gpr_reg_base),
8919 return gen_rtx_PARALLEL (mode, gen_rtvec_v (ub, loc));
8924 /* If the argument is larger than a word, then we know precisely
8925 which registers we must use. */
8939 /* Structures 5 to 8 bytes in size are passed in the general
8940 registers in the same manner as other non floating-point
8941 objects. The data is right-justified and zero-extended
8942 to 64 bits. This is opposite to the normal justification
8943 used on big endian targets and requires special treatment.
8944 We now define BLOCK_REG_PADDING to pad these objects. */
8945 if (mode == BLKmode || (type && AGGREGATE_TYPE_P (type)))
8947 rtx loc = gen_rtx_EXPR_LIST (VOIDmode,
8948 gen_rtx_REG (DImode, gpr_reg_base),
8950 return gen_rtx_PARALLEL (BLKmode, gen_rtvec (1, loc));
8955 /* We have a single word (32 bits). A simple computation
8956 will get us the register #s we need. */
8957 gpr_reg_base = 26 - cum->words;
8958 fpr_reg_base = 32 + 2 * cum->words;
8962 /* Determine if the argument needs to be passed in both general and
8963 floating point registers. */
8964 if (((TARGET_PORTABLE_RUNTIME || TARGET_64BIT || TARGET_ELF32)
8965 /* If we are doing soft-float with portable runtime, then there
8966 is no need to worry about FP regs. */
8967 && !TARGET_SOFT_FLOAT
8968 /* The parameter must be some kind of float, else we can just
8969 pass it in integer registers. */
8970 && FLOAT_MODE_P (mode)
8971 /* The target function must not have a prototype. */
8972 && cum->nargs_prototype <= 0
8973 /* libcalls do not need to pass items in both FP and general
8975 && type != NULL_TREE
8976 /* All this hair applies to "outgoing" args only. This includes
8977 sibcall arguments setup with FUNCTION_INCOMING_ARG. */
8979 /* Also pass outgoing floating arguments in both registers in indirect
8980 calls with the 32 bit ABI and the HP assembler since there is no
8981 way to the specify argument locations in static functions. */
8986 && FLOAT_MODE_P (mode)))
8992 gen_rtx_EXPR_LIST (VOIDmode,
8993 gen_rtx_REG (mode, fpr_reg_base),
8995 gen_rtx_EXPR_LIST (VOIDmode,
8996 gen_rtx_REG (mode, gpr_reg_base),
9001 /* See if we should pass this parameter in a general register. */
9002 if (TARGET_SOFT_FLOAT
9003 /* Indirect calls in the normal 32bit ABI require all arguments
9004 to be passed in general registers. */
9005 || (!TARGET_PORTABLE_RUNTIME
9009 /* If the parameter is not a floating point parameter, then
9010 it belongs in GPRs. */
9011 || !FLOAT_MODE_P (mode)
9012 /* Structure with single SFmode field belongs in GPR. */
9013 || (type && AGGREGATE_TYPE_P (type)))
9014 retval = gen_rtx_REG (mode, gpr_reg_base);
9016 retval = gen_rtx_REG (mode, fpr_reg_base);
9022 /* If this arg would be passed totally in registers or totally on the stack,
9023 then this routine should return zero. */
9026 pa_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
9027 tree type, bool named ATTRIBUTE_UNUSED)
9029 unsigned int max_arg_words = 8;
9030 unsigned int offset = 0;
9035 if (FUNCTION_ARG_SIZE (mode, type) > 1 && (cum->words & 1))
9038 if (cum->words + offset + FUNCTION_ARG_SIZE (mode, type) <= max_arg_words)
9039 /* Arg fits fully into registers. */
9041 else if (cum->words + offset >= max_arg_words)
9042 /* Arg fully on the stack. */
9046 return (max_arg_words - cum->words - offset) * UNITS_PER_WORD;
9050 /* A get_unnamed_section callback for switching to the text section.
9052 This function is only used with SOM. Because we don't support
9053 named subspaces, we can only create a new subspace or switch back
9054 to the default text subspace. */
9057 som_output_text_section_asm_op (const void *data ATTRIBUTE_UNUSED)
9059 gcc_assert (TARGET_SOM);
9062 if (cfun && !cfun->machine->in_nsubspa)
9064 /* We only want to emit a .nsubspa directive once at the
9065 start of the function. */
9066 cfun->machine->in_nsubspa = 1;
9068 /* Create a new subspace for the text. This provides
9069 better stub placement and one-only functions. */
9071 && DECL_ONE_ONLY (cfun->decl)
9072 && !DECL_WEAK (cfun->decl))
9074 output_section_asm_op ("\t.SPACE $TEXT$\n"
9075 "\t.NSUBSPA $CODE$,QUAD=0,ALIGN=8,"
9076 "ACCESS=44,SORT=24,COMDAT");
9082 /* There isn't a current function or the body of the current
9083 function has been completed. So, we are changing to the
9084 text section to output debugging information. Thus, we
9085 need to forget that we are in the text section so that
9086 varasm.c will call us when text_section is selected again. */
9087 gcc_assert (!cfun || cfun->machine->in_nsubspa == 2);
9090 output_section_asm_op ("\t.SPACE $TEXT$\n\t.NSUBSPA $CODE$");
9093 output_section_asm_op ("\t.SPACE $TEXT$\n\t.SUBSPA $CODE$");
9096 /* A get_unnamed_section callback for switching to comdat data
9097 sections. This function is only used with SOM. */
9100 som_output_comdat_data_section_asm_op (const void *data)
9103 output_section_asm_op (data);
9106 /* Implement TARGET_ASM_INITIALIZE_SECTIONS */
9109 pa_som_asm_init_sections (void)
9112 = get_unnamed_section (0, som_output_text_section_asm_op, NULL);
9114 /* SOM puts readonly data in the default $LIT$ subspace when PIC code
9115 is not being generated. */
9116 som_readonly_data_section
9117 = get_unnamed_section (0, output_section_asm_op,
9118 "\t.SPACE $TEXT$\n\t.SUBSPA $LIT$");
9120 /* When secondary definitions are not supported, SOM makes readonly
9121 data one-only by creating a new $LIT$ subspace in $TEXT$ with
9123 som_one_only_readonly_data_section
9124 = get_unnamed_section (0, som_output_comdat_data_section_asm_op,
9126 "\t.NSUBSPA $LIT$,QUAD=0,ALIGN=8,"
9127 "ACCESS=0x2c,SORT=16,COMDAT");
9130 /* When secondary definitions are not supported, SOM makes data one-only
9131 by creating a new $DATA$ subspace in $PRIVATE$ with the comdat flag. */
9132 som_one_only_data_section
9133 = get_unnamed_section (SECTION_WRITE,
9134 som_output_comdat_data_section_asm_op,
9135 "\t.SPACE $PRIVATE$\n"
9136 "\t.NSUBSPA $DATA$,QUAD=1,ALIGN=8,"
9137 "ACCESS=31,SORT=24,COMDAT");
9139 /* FIXME: HPUX ld generates incorrect GOT entries for "T" fixups
9140 which reference data within the $TEXT$ space (for example constant
9141 strings in the $LIT$ subspace).
9143 The assemblers (GAS and HP as) both have problems with handling
9144 the difference of two symbols which is the other correct way to
9145 reference constant data during PIC code generation.
9147 So, there's no way to reference constant data which is in the
9148 $TEXT$ space during PIC generation. Instead place all constant
9149 data into the $PRIVATE$ subspace (this reduces sharing, but it
9150 works correctly). */
9151 readonly_data_section = flag_pic ? data_section : som_readonly_data_section;
9153 /* We must not have a reference to an external symbol defined in a
9154 shared library in a readonly section, else the SOM linker will
9157 So, we force exception information into the data section. */
9158 exception_section = data_section;
9161 /* On hpux10, the linker will give an error if we have a reference
9162 in the read-only data section to a symbol defined in a shared
9163 library. Therefore, expressions that might require a reloc can
9164 not be placed in the read-only data section. */
9167 pa_select_section (tree exp, int reloc,
9168 unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED)
9170 if (TREE_CODE (exp) == VAR_DECL
9171 && TREE_READONLY (exp)
9172 && !TREE_THIS_VOLATILE (exp)
9173 && DECL_INITIAL (exp)
9174 && (DECL_INITIAL (exp) == error_mark_node
9175 || TREE_CONSTANT (DECL_INITIAL (exp)))
9179 && DECL_ONE_ONLY (exp)
9180 && !DECL_WEAK (exp))
9181 return som_one_only_readonly_data_section;
9183 return readonly_data_section;
9185 else if (CONSTANT_CLASS_P (exp) && !reloc)
9186 return readonly_data_section;
9188 && TREE_CODE (exp) == VAR_DECL
9189 && DECL_ONE_ONLY (exp)
9190 && !DECL_WEAK (exp))
9191 return som_one_only_data_section;
9193 return data_section;
9197 pa_globalize_label (FILE *stream, const char *name)
9199 /* We only handle DATA objects here, functions are globalized in
9200 ASM_DECLARE_FUNCTION_NAME. */
9201 if (! FUNCTION_NAME_P (name))
9203 fputs ("\t.EXPORT ", stream);
9204 assemble_name (stream, name);
9205 fputs (",DATA\n", stream);
9209 /* Worker function for TARGET_STRUCT_VALUE_RTX. */
9212 pa_struct_value_rtx (tree fntype ATTRIBUTE_UNUSED,
9213 int incoming ATTRIBUTE_UNUSED)
9215 return gen_rtx_REG (Pmode, PA_STRUCT_VALUE_REGNUM);
9218 /* Worker function for TARGET_RETURN_IN_MEMORY. */
9221 pa_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
9223 /* SOM ABI says that objects larger than 64 bits are returned in memory.
9224 PA64 ABI says that objects larger than 128 bits are returned in memory.
9225 Note, int_size_in_bytes can return -1 if the size of the object is
9226 variable or larger than the maximum value that can be expressed as
9227 a HOST_WIDE_INT. It can also return zero for an empty type. The
9228 simplest way to handle variable and empty types is to pass them in
9229 memory. This avoids problems in defining the boundaries of argument
9230 slots, allocating registers, etc. */
9231 return (int_size_in_bytes (type) > (TARGET_64BIT ? 16 : 8)
9232 || int_size_in_bytes (type) <= 0);
9235 /* Structure to hold declaration and name of external symbols that are
9236 emitted by GCC. We generate a vector of these symbols and output them
9237 at the end of the file if and only if SYMBOL_REF_REFERENCED_P is true.
9238 This avoids putting out names that are never really used. */
9240 typedef struct extern_symbol GTY(())
9246 /* Define gc'd vector type for extern_symbol. */
9247 DEF_VEC_O(extern_symbol);
9248 DEF_VEC_ALLOC_O(extern_symbol,gc);
9250 /* Vector of extern_symbol pointers. */
9251 static GTY(()) VEC(extern_symbol,gc) *extern_symbols;
9253 #ifdef ASM_OUTPUT_EXTERNAL_REAL
9254 /* Mark DECL (name NAME) as an external reference (assembler output
9255 file FILE). This saves the names to output at the end of the file
9256 if actually referenced. */
9259 pa_hpux_asm_output_external (FILE *file, tree decl, const char *name)
9261 extern_symbol * p = VEC_safe_push (extern_symbol, gc, extern_symbols, NULL);
9263 gcc_assert (file == asm_out_file);
9268 /* Output text required at the end of an assembler file.
9269 This includes deferred plabels and .import directives for
9270 all external symbols that were actually referenced. */
9273 pa_hpux_file_end (void)
9278 if (!NO_DEFERRED_PROFILE_COUNTERS)
9279 output_deferred_profile_counters ();
9281 output_deferred_plabels ();
9283 for (i = 0; VEC_iterate (extern_symbol, extern_symbols, i, p); i++)
9285 tree decl = p->decl;
9287 if (!TREE_ASM_WRITTEN (decl)
9288 && SYMBOL_REF_REFERENCED_P (XEXP (DECL_RTL (decl), 0)))
9289 ASM_OUTPUT_EXTERNAL_REAL (asm_out_file, decl, p->name);
9292 VEC_free (extern_symbol, gc, extern_symbols);