1 /* Subroutines used for code generation on the DEC Alpha.
2 Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
3 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
4 Free Software Foundation, Inc.
5 Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
26 #include "coretypes.h"
31 #include "hard-reg-set.h"
33 #include "insn-config.h"
34 #include "conditions.h"
36 #include "insn-attr.h"
47 #include "integrate.h"
50 #include "target-def.h"
52 #include "langhooks.h"
53 #include <splay-tree.h>
54 #include "cfglayout.h"
56 #include "tree-flow.h"
57 #include "tree-stdarg.h"
58 #include "tm-constrs.h"
61 /* Specify which cpu to schedule for. */
62 enum processor_type alpha_tune
;
64 /* Which cpu we're generating code for. */
65 enum processor_type alpha_cpu
;
67 static const char * const alpha_cpu_name
[] =
72 /* Specify how accurate floating-point traps need to be. */
74 enum alpha_trap_precision alpha_tp
;
76 /* Specify the floating-point rounding mode. */
78 enum alpha_fp_rounding_mode alpha_fprm
;
80 /* Specify which things cause traps. */
82 enum alpha_fp_trap_mode alpha_fptm
;
84 /* Nonzero if inside of a function, because the Alpha asm can't
85 handle .files inside of functions. */
87 static int inside_function
= FALSE
;
89 /* The number of cycles of latency we should assume on memory reads. */
91 int alpha_memory_latency
= 3;
93 /* Whether the function needs the GP. */
95 static int alpha_function_needs_gp
;
97 /* The alias set for prologue/epilogue register save/restore. */
99 static GTY(()) alias_set_type alpha_sr_alias_set
;
101 /* The assembler name of the current function. */
103 static const char *alpha_fnname
;
105 /* The next explicit relocation sequence number. */
106 extern GTY(()) int alpha_next_sequence_number
;
107 int alpha_next_sequence_number
= 1;
109 /* The literal and gpdisp sequence numbers for this insn, as printed
110 by %# and %* respectively. */
111 extern GTY(()) int alpha_this_literal_sequence_number
;
112 extern GTY(()) int alpha_this_gpdisp_sequence_number
;
113 int alpha_this_literal_sequence_number
;
114 int alpha_this_gpdisp_sequence_number
;
116 /* Costs of various operations on the different architectures. */
118 struct alpha_rtx_cost_data
120 unsigned char fp_add
;
121 unsigned char fp_mult
;
122 unsigned char fp_div_sf
;
123 unsigned char fp_div_df
;
124 unsigned char int_mult_si
;
125 unsigned char int_mult_di
;
126 unsigned char int_shift
;
127 unsigned char int_cmov
;
128 unsigned short int_div
;
131 static struct alpha_rtx_cost_data
const alpha_rtx_cost_data
[PROCESSOR_MAX
] =
134 COSTS_N_INSNS (6), /* fp_add */
135 COSTS_N_INSNS (6), /* fp_mult */
136 COSTS_N_INSNS (34), /* fp_div_sf */
137 COSTS_N_INSNS (63), /* fp_div_df */
138 COSTS_N_INSNS (23), /* int_mult_si */
139 COSTS_N_INSNS (23), /* int_mult_di */
140 COSTS_N_INSNS (2), /* int_shift */
141 COSTS_N_INSNS (2), /* int_cmov */
142 COSTS_N_INSNS (97), /* int_div */
145 COSTS_N_INSNS (4), /* fp_add */
146 COSTS_N_INSNS (4), /* fp_mult */
147 COSTS_N_INSNS (15), /* fp_div_sf */
148 COSTS_N_INSNS (22), /* fp_div_df */
149 COSTS_N_INSNS (8), /* int_mult_si */
150 COSTS_N_INSNS (12), /* int_mult_di */
151 COSTS_N_INSNS (1) + 1, /* int_shift */
152 COSTS_N_INSNS (1), /* int_cmov */
153 COSTS_N_INSNS (83), /* int_div */
156 COSTS_N_INSNS (4), /* fp_add */
157 COSTS_N_INSNS (4), /* fp_mult */
158 COSTS_N_INSNS (12), /* fp_div_sf */
159 COSTS_N_INSNS (15), /* fp_div_df */
160 COSTS_N_INSNS (7), /* int_mult_si */
161 COSTS_N_INSNS (7), /* int_mult_di */
162 COSTS_N_INSNS (1), /* int_shift */
163 COSTS_N_INSNS (2), /* int_cmov */
164 COSTS_N_INSNS (86), /* int_div */
168 /* Similar but tuned for code size instead of execution latency. The
169 extra +N is fractional cost tuning based on latency. It's used to
170 encourage use of cheaper insns like shift, but only if there's just
173 static struct alpha_rtx_cost_data
const alpha_rtx_cost_size
=
175 COSTS_N_INSNS (1), /* fp_add */
176 COSTS_N_INSNS (1), /* fp_mult */
177 COSTS_N_INSNS (1), /* fp_div_sf */
178 COSTS_N_INSNS (1) + 1, /* fp_div_df */
179 COSTS_N_INSNS (1) + 1, /* int_mult_si */
180 COSTS_N_INSNS (1) + 2, /* int_mult_di */
181 COSTS_N_INSNS (1), /* int_shift */
182 COSTS_N_INSNS (1), /* int_cmov */
183 COSTS_N_INSNS (6), /* int_div */
186 /* Get the number of args of a function in one of two ways. */
187 #if TARGET_ABI_OPEN_VMS || TARGET_ABI_UNICOSMK
188 #define NUM_ARGS crtl->args.info.num_args
190 #define NUM_ARGS crtl->args.info
196 /* Declarations of static functions. */
197 static struct machine_function
*alpha_init_machine_status (void);
198 static rtx
alpha_emit_xfloating_compare (enum rtx_code
*, rtx
, rtx
);
200 #if TARGET_ABI_OPEN_VMS
201 static void alpha_write_linkage (FILE *, const char *, tree
);
204 static void unicosmk_output_deferred_case_vectors (FILE *);
205 static void unicosmk_gen_dsib (unsigned long *);
206 static void unicosmk_output_ssib (FILE *, const char *);
207 static int unicosmk_need_dex (rtx
);
209 /* Implement TARGET_HANDLE_OPTION. */
212 alpha_handle_option (size_t code
, const char *arg
, int value
)
218 target_flags
|= MASK_SOFT_FP
;
222 case OPT_mieee_with_inexact
:
223 target_flags
|= MASK_IEEE_CONFORMANT
;
227 if (value
!= 16 && value
!= 32 && value
!= 64)
228 error ("bad value %qs for -mtls-size switch", arg
);
235 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
236 /* Implement TARGET_MANGLE_TYPE. */
239 alpha_mangle_type (const_tree type
)
241 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
242 && TARGET_LONG_DOUBLE_128
)
245 /* For all other types, use normal C++ mangling. */
250 /* Parse target option strings. */
253 override_options (void)
255 static const struct cpu_table
{
256 const char *const name
;
257 const enum processor_type processor
;
260 { "ev4", PROCESSOR_EV4
, 0 },
261 { "ev45", PROCESSOR_EV4
, 0 },
262 { "21064", PROCESSOR_EV4
, 0 },
263 { "ev5", PROCESSOR_EV5
, 0 },
264 { "21164", PROCESSOR_EV5
, 0 },
265 { "ev56", PROCESSOR_EV5
, MASK_BWX
},
266 { "21164a", PROCESSOR_EV5
, MASK_BWX
},
267 { "pca56", PROCESSOR_EV5
, MASK_BWX
|MASK_MAX
},
268 { "21164PC",PROCESSOR_EV5
, MASK_BWX
|MASK_MAX
},
269 { "21164pc",PROCESSOR_EV5
, MASK_BWX
|MASK_MAX
},
270 { "ev6", PROCESSOR_EV6
, MASK_BWX
|MASK_MAX
|MASK_FIX
},
271 { "21264", PROCESSOR_EV6
, MASK_BWX
|MASK_MAX
|MASK_FIX
},
272 { "ev67", PROCESSOR_EV6
, MASK_BWX
|MASK_MAX
|MASK_FIX
|MASK_CIX
},
273 { "21264a", PROCESSOR_EV6
, MASK_BWX
|MASK_MAX
|MASK_FIX
|MASK_CIX
}
276 int const ct_size
= ARRAY_SIZE (cpu_table
);
279 /* Unicos/Mk doesn't have shared libraries. */
280 if (TARGET_ABI_UNICOSMK
&& flag_pic
)
282 warning (0, "-f%s ignored for Unicos/Mk (not supported)",
283 (flag_pic
> 1) ? "PIC" : "pic");
287 /* On Unicos/Mk, the native compiler consistently generates /d suffices for
288 floating-point instructions. Make that the default for this target. */
289 if (TARGET_ABI_UNICOSMK
)
290 alpha_fprm
= ALPHA_FPRM_DYN
;
292 alpha_fprm
= ALPHA_FPRM_NORM
;
294 alpha_tp
= ALPHA_TP_PROG
;
295 alpha_fptm
= ALPHA_FPTM_N
;
297 /* We cannot use su and sui qualifiers for conversion instructions on
298 Unicos/Mk. I'm not sure if this is due to assembler or hardware
299 limitations. Right now, we issue a warning if -mieee is specified
300 and then ignore it; eventually, we should either get it right or
301 disable the option altogether. */
305 if (TARGET_ABI_UNICOSMK
)
306 warning (0, "-mieee not supported on Unicos/Mk");
309 alpha_tp
= ALPHA_TP_INSN
;
310 alpha_fptm
= ALPHA_FPTM_SU
;
314 if (TARGET_IEEE_WITH_INEXACT
)
316 if (TARGET_ABI_UNICOSMK
)
317 warning (0, "-mieee-with-inexact not supported on Unicos/Mk");
320 alpha_tp
= ALPHA_TP_INSN
;
321 alpha_fptm
= ALPHA_FPTM_SUI
;
327 if (! strcmp (alpha_tp_string
, "p"))
328 alpha_tp
= ALPHA_TP_PROG
;
329 else if (! strcmp (alpha_tp_string
, "f"))
330 alpha_tp
= ALPHA_TP_FUNC
;
331 else if (! strcmp (alpha_tp_string
, "i"))
332 alpha_tp
= ALPHA_TP_INSN
;
334 error ("bad value %qs for -mtrap-precision switch", alpha_tp_string
);
337 if (alpha_fprm_string
)
339 if (! strcmp (alpha_fprm_string
, "n"))
340 alpha_fprm
= ALPHA_FPRM_NORM
;
341 else if (! strcmp (alpha_fprm_string
, "m"))
342 alpha_fprm
= ALPHA_FPRM_MINF
;
343 else if (! strcmp (alpha_fprm_string
, "c"))
344 alpha_fprm
= ALPHA_FPRM_CHOP
;
345 else if (! strcmp (alpha_fprm_string
,"d"))
346 alpha_fprm
= ALPHA_FPRM_DYN
;
348 error ("bad value %qs for -mfp-rounding-mode switch",
352 if (alpha_fptm_string
)
354 if (strcmp (alpha_fptm_string
, "n") == 0)
355 alpha_fptm
= ALPHA_FPTM_N
;
356 else if (strcmp (alpha_fptm_string
, "u") == 0)
357 alpha_fptm
= ALPHA_FPTM_U
;
358 else if (strcmp (alpha_fptm_string
, "su") == 0)
359 alpha_fptm
= ALPHA_FPTM_SU
;
360 else if (strcmp (alpha_fptm_string
, "sui") == 0)
361 alpha_fptm
= ALPHA_FPTM_SUI
;
363 error ("bad value %qs for -mfp-trap-mode switch", alpha_fptm_string
);
366 if (alpha_cpu_string
)
368 for (i
= 0; i
< ct_size
; i
++)
369 if (! strcmp (alpha_cpu_string
, cpu_table
[i
].name
))
371 alpha_tune
= alpha_cpu
= cpu_table
[i
].processor
;
372 target_flags
&= ~ (MASK_BWX
| MASK_MAX
| MASK_FIX
| MASK_CIX
);
373 target_flags
|= cpu_table
[i
].flags
;
377 error ("bad value %qs for -mcpu switch", alpha_cpu_string
);
380 if (alpha_tune_string
)
382 for (i
= 0; i
< ct_size
; i
++)
383 if (! strcmp (alpha_tune_string
, cpu_table
[i
].name
))
385 alpha_tune
= cpu_table
[i
].processor
;
389 error ("bad value %qs for -mcpu switch", alpha_tune_string
);
392 /* Do some sanity checks on the above options. */
394 if (TARGET_ABI_UNICOSMK
&& alpha_fptm
!= ALPHA_FPTM_N
)
396 warning (0, "trap mode not supported on Unicos/Mk");
397 alpha_fptm
= ALPHA_FPTM_N
;
400 if ((alpha_fptm
== ALPHA_FPTM_SU
|| alpha_fptm
== ALPHA_FPTM_SUI
)
401 && alpha_tp
!= ALPHA_TP_INSN
&& alpha_cpu
!= PROCESSOR_EV6
)
403 warning (0, "fp software completion requires -mtrap-precision=i");
404 alpha_tp
= ALPHA_TP_INSN
;
407 if (alpha_cpu
== PROCESSOR_EV6
)
409 /* Except for EV6 pass 1 (not released), we always have precise
410 arithmetic traps. Which means we can do software completion
411 without minding trap shadows. */
412 alpha_tp
= ALPHA_TP_PROG
;
415 if (TARGET_FLOAT_VAX
)
417 if (alpha_fprm
== ALPHA_FPRM_MINF
|| alpha_fprm
== ALPHA_FPRM_DYN
)
419 warning (0, "rounding mode not supported for VAX floats");
420 alpha_fprm
= ALPHA_FPRM_NORM
;
422 if (alpha_fptm
== ALPHA_FPTM_SUI
)
424 warning (0, "trap mode not supported for VAX floats");
425 alpha_fptm
= ALPHA_FPTM_SU
;
427 if (target_flags_explicit
& MASK_LONG_DOUBLE_128
)
428 warning (0, "128-bit long double not supported for VAX floats");
429 target_flags
&= ~MASK_LONG_DOUBLE_128
;
436 if (!alpha_mlat_string
)
437 alpha_mlat_string
= "L1";
439 if (ISDIGIT ((unsigned char)alpha_mlat_string
[0])
440 && (lat
= strtol (alpha_mlat_string
, &end
, 10), *end
== '\0'))
442 else if ((alpha_mlat_string
[0] == 'L' || alpha_mlat_string
[0] == 'l')
443 && ISDIGIT ((unsigned char)alpha_mlat_string
[1])
444 && alpha_mlat_string
[2] == '\0')
446 static int const cache_latency
[][4] =
448 { 3, 30, -1 }, /* ev4 -- Bcache is a guess */
449 { 2, 12, 38 }, /* ev5 -- Bcache from PC164 LMbench numbers */
450 { 3, 12, 30 }, /* ev6 -- Bcache from DS20 LMbench. */
453 lat
= alpha_mlat_string
[1] - '0';
454 if (lat
<= 0 || lat
> 3 || cache_latency
[alpha_tune
][lat
-1] == -1)
456 warning (0, "L%d cache latency unknown for %s",
457 lat
, alpha_cpu_name
[alpha_tune
]);
461 lat
= cache_latency
[alpha_tune
][lat
-1];
463 else if (! strcmp (alpha_mlat_string
, "main"))
465 /* Most current memories have about 370ns latency. This is
466 a reasonable guess for a fast cpu. */
471 warning (0, "bad value %qs for -mmemory-latency", alpha_mlat_string
);
475 alpha_memory_latency
= lat
;
478 /* Default the definition of "small data" to 8 bytes. */
482 /* Infer TARGET_SMALL_DATA from -fpic/-fPIC. */
484 target_flags
|= MASK_SMALL_DATA
;
485 else if (flag_pic
== 2)
486 target_flags
&= ~MASK_SMALL_DATA
;
488 /* Align labels and loops for optimal branching. */
489 /* ??? Kludge these by not doing anything if we don't optimize and also if
490 we are writing ECOFF symbols to work around a bug in DEC's assembler. */
491 if (optimize
> 0 && write_symbols
!= SDB_DEBUG
)
493 if (align_loops
<= 0)
495 if (align_jumps
<= 0)
498 if (align_functions
<= 0)
499 align_functions
= 16;
501 /* Acquire a unique set number for our register saves and restores. */
502 alpha_sr_alias_set
= new_alias_set ();
504 /* Register variables and functions with the garbage collector. */
506 /* Set up function hooks. */
507 init_machine_status
= alpha_init_machine_status
;
509 /* Tell the compiler when we're using VAX floating point. */
510 if (TARGET_FLOAT_VAX
)
512 REAL_MODE_FORMAT (SFmode
) = &vax_f_format
;
513 REAL_MODE_FORMAT (DFmode
) = &vax_g_format
;
514 REAL_MODE_FORMAT (TFmode
) = NULL
;
517 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
518 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
519 target_flags
|= MASK_LONG_DOUBLE_128
;
522 /* If using typedef char *va_list, signal that __builtin_va_start (&ap, 0)
523 can be optimized to ap = __builtin_next_arg (0). */
524 if (TARGET_ABI_UNICOSMK
)
525 targetm
.expand_builtin_va_start
= NULL
;
528 /* Returns 1 if VALUE is a mask that contains full bytes of zero or ones. */
531 zap_mask (HOST_WIDE_INT value
)
535 for (i
= 0; i
< HOST_BITS_PER_WIDE_INT
/ HOST_BITS_PER_CHAR
;
537 if ((value
& 0xff) != 0 && (value
& 0xff) != 0xff)
543 /* Return true if OP is valid for a particular TLS relocation.
544 We are already guaranteed that OP is a CONST. */
547 tls_symbolic_operand_1 (rtx op
, int size
, int unspec
)
551 if (GET_CODE (op
) != UNSPEC
|| XINT (op
, 1) != unspec
)
553 op
= XVECEXP (op
, 0, 0);
555 if (GET_CODE (op
) != SYMBOL_REF
)
558 switch (SYMBOL_REF_TLS_MODEL (op
))
560 case TLS_MODEL_LOCAL_DYNAMIC
:
561 return unspec
== UNSPEC_DTPREL
&& size
== alpha_tls_size
;
562 case TLS_MODEL_INITIAL_EXEC
:
563 return unspec
== UNSPEC_TPREL
&& size
== 64;
564 case TLS_MODEL_LOCAL_EXEC
:
565 return unspec
== UNSPEC_TPREL
&& size
== alpha_tls_size
;
571 /* Used by aligned_memory_operand and unaligned_memory_operand to
572 resolve what reload is going to do with OP if it's a register. */
575 resolve_reload_operand (rtx op
)
577 if (reload_in_progress
)
580 if (GET_CODE (tmp
) == SUBREG
)
581 tmp
= SUBREG_REG (tmp
);
583 && REGNO (tmp
) >= FIRST_PSEUDO_REGISTER
)
585 op
= reg_equiv_memory_loc
[REGNO (tmp
)];
593 /* The scalar modes supported differs from the default check-what-c-supports
594 version in that sometimes TFmode is available even when long double
595 indicates only DFmode. On unicosmk, we have the situation that HImode
596 doesn't map to any C type, but of course we still support that. */
599 alpha_scalar_mode_supported_p (enum machine_mode mode
)
607 case TImode
: /* via optabs.c */
615 return TARGET_HAS_XFLOATING_LIBS
;
622 /* Alpha implements a couple of integer vector mode operations when
623 TARGET_MAX is enabled. We do not check TARGET_MAX here, however,
624 which allows the vectorizer to operate on e.g. move instructions,
625 or when expand_vector_operations can do something useful. */
628 alpha_vector_mode_supported_p (enum machine_mode mode
)
630 return mode
== V8QImode
|| mode
== V4HImode
|| mode
== V2SImode
;
633 /* Return 1 if this function can directly return via $26. */
638 return (! TARGET_ABI_OPEN_VMS
&& ! TARGET_ABI_UNICOSMK
640 && alpha_sa_size () == 0
641 && get_frame_size () == 0
642 && crtl
->outgoing_args_size
== 0
643 && crtl
->args
.pretend_args_size
== 0);
646 /* Return the ADDR_VEC associated with a tablejump insn. */
649 alpha_tablejump_addr_vec (rtx insn
)
653 tmp
= JUMP_LABEL (insn
);
656 tmp
= NEXT_INSN (tmp
);
660 && GET_CODE (PATTERN (tmp
)) == ADDR_DIFF_VEC
)
661 return PATTERN (tmp
);
665 /* Return the label of the predicted edge, or CONST0_RTX if we don't know. */
668 alpha_tablejump_best_label (rtx insn
)
670 rtx jump_table
= alpha_tablejump_addr_vec (insn
);
671 rtx best_label
= NULL_RTX
;
673 /* ??? Once the CFG doesn't keep getting completely rebuilt, look
674 there for edge frequency counts from profile data. */
678 int n_labels
= XVECLEN (jump_table
, 1);
682 for (i
= 0; i
< n_labels
; i
++)
686 for (j
= i
+ 1; j
< n_labels
; j
++)
687 if (XEXP (XVECEXP (jump_table
, 1, i
), 0)
688 == XEXP (XVECEXP (jump_table
, 1, j
), 0))
691 if (count
> best_count
)
692 best_count
= count
, best_label
= XVECEXP (jump_table
, 1, i
);
696 return best_label
? best_label
: const0_rtx
;
699 /* Return the TLS model to use for SYMBOL. */
701 static enum tls_model
702 tls_symbolic_operand_type (rtx symbol
)
704 enum tls_model model
;
706 if (GET_CODE (symbol
) != SYMBOL_REF
)
707 return TLS_MODEL_NONE
;
708 model
= SYMBOL_REF_TLS_MODEL (symbol
);
710 /* Local-exec with a 64-bit size is the same code as initial-exec. */
711 if (model
== TLS_MODEL_LOCAL_EXEC
&& alpha_tls_size
== 64)
712 model
= TLS_MODEL_INITIAL_EXEC
;
717 /* Return true if the function DECL will share the same GP as any
718 function in the current unit of translation. */
721 decl_has_samegp (const_tree decl
)
723 /* Functions that are not local can be overridden, and thus may
724 not share the same gp. */
725 if (!(*targetm
.binds_local_p
) (decl
))
728 /* If -msmall-data is in effect, assume that there is only one GP
729 for the module, and so any local symbol has this property. We
730 need explicit relocations to be able to enforce this for symbols
731 not defined in this unit of translation, however. */
732 if (TARGET_EXPLICIT_RELOCS
&& TARGET_SMALL_DATA
)
735 /* Functions that are not external are defined in this UoT. */
736 /* ??? Irritatingly, static functions not yet emitted are still
737 marked "external". Apply this to non-static functions only. */
738 return !TREE_PUBLIC (decl
) || !DECL_EXTERNAL (decl
);
741 /* Return true if EXP should be placed in the small data section. */
744 alpha_in_small_data_p (const_tree exp
)
746 /* We want to merge strings, so we never consider them small data. */
747 if (TREE_CODE (exp
) == STRING_CST
)
750 /* Functions are never in the small data area. Duh. */
751 if (TREE_CODE (exp
) == FUNCTION_DECL
)
754 if (TREE_CODE (exp
) == VAR_DECL
&& DECL_SECTION_NAME (exp
))
756 const char *section
= TREE_STRING_POINTER (DECL_SECTION_NAME (exp
));
757 if (strcmp (section
, ".sdata") == 0
758 || strcmp (section
, ".sbss") == 0)
763 HOST_WIDE_INT size
= int_size_in_bytes (TREE_TYPE (exp
));
765 /* If this is an incomplete type with size 0, then we can't put it
766 in sdata because it might be too big when completed. */
767 if (size
> 0 && (unsigned HOST_WIDE_INT
) size
<= g_switch_value
)
774 #if TARGET_ABI_OPEN_VMS
776 alpha_linkage_symbol_p (const char *symname
)
778 int symlen
= strlen (symname
);
781 return strcmp (&symname
[symlen
- 4], "..lk") == 0;
786 #define LINKAGE_SYMBOL_REF_P(X) \
787 ((GET_CODE (X) == SYMBOL_REF \
788 && alpha_linkage_symbol_p (XSTR (X, 0))) \
789 || (GET_CODE (X) == CONST \
790 && GET_CODE (XEXP (X, 0)) == PLUS \
791 && GET_CODE (XEXP (XEXP (X, 0), 0)) == SYMBOL_REF \
792 && alpha_linkage_symbol_p (XSTR (XEXP (XEXP (X, 0), 0), 0))))
795 /* legitimate_address_p recognizes an RTL expression that is a valid
796 memory address for an instruction. The MODE argument is the
797 machine mode for the MEM expression that wants to use this address.
799 For Alpha, we have either a constant address or the sum of a
800 register and a constant address, or just a register. For DImode,
801 any of those forms can be surrounded with an AND that clear the
802 low-order three bits; this is an "unaligned" access. */
805 alpha_legitimate_address_p (enum machine_mode mode
, rtx x
, bool strict
)
807 /* If this is an ldq_u type address, discard the outer AND. */
809 && GET_CODE (x
) == AND
810 && CONST_INT_P (XEXP (x
, 1))
811 && INTVAL (XEXP (x
, 1)) == -8)
814 /* Discard non-paradoxical subregs. */
815 if (GET_CODE (x
) == SUBREG
816 && (GET_MODE_SIZE (GET_MODE (x
))
817 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x
)))))
820 /* Unadorned general registers are valid. */
823 ? STRICT_REG_OK_FOR_BASE_P (x
)
824 : NONSTRICT_REG_OK_FOR_BASE_P (x
)))
827 /* Constant addresses (i.e. +/- 32k) are valid. */
828 if (CONSTANT_ADDRESS_P (x
))
831 #if TARGET_ABI_OPEN_VMS
832 if (LINKAGE_SYMBOL_REF_P (x
))
836 /* Register plus a small constant offset is valid. */
837 if (GET_CODE (x
) == PLUS
)
839 rtx ofs
= XEXP (x
, 1);
842 /* Discard non-paradoxical subregs. */
843 if (GET_CODE (x
) == SUBREG
844 && (GET_MODE_SIZE (GET_MODE (x
))
845 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x
)))))
851 && NONSTRICT_REG_OK_FP_BASE_P (x
)
852 && CONST_INT_P (ofs
))
855 ? STRICT_REG_OK_FOR_BASE_P (x
)
856 : NONSTRICT_REG_OK_FOR_BASE_P (x
))
857 && CONSTANT_ADDRESS_P (ofs
))
862 /* If we're managing explicit relocations, LO_SUM is valid, as are small
863 data symbols. Avoid explicit relocations of modes larger than word
864 mode since i.e. $LC0+8($1) can fold around +/- 32k offset. */
865 else if (TARGET_EXPLICIT_RELOCS
866 && GET_MODE_SIZE (mode
) <= UNITS_PER_WORD
)
868 if (small_symbolic_operand (x
, Pmode
))
871 if (GET_CODE (x
) == LO_SUM
)
873 rtx ofs
= XEXP (x
, 1);
876 /* Discard non-paradoxical subregs. */
877 if (GET_CODE (x
) == SUBREG
878 && (GET_MODE_SIZE (GET_MODE (x
))
879 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x
)))))
882 /* Must have a valid base register. */
885 ? STRICT_REG_OK_FOR_BASE_P (x
)
886 : NONSTRICT_REG_OK_FOR_BASE_P (x
))))
889 /* The symbol must be local. */
890 if (local_symbolic_operand (ofs
, Pmode
)
891 || dtp32_symbolic_operand (ofs
, Pmode
)
892 || tp32_symbolic_operand (ofs
, Pmode
))
900 /* Build the SYMBOL_REF for __tls_get_addr. */
902 static GTY(()) rtx tls_get_addr_libfunc
;
905 get_tls_get_addr (void)
907 if (!tls_get_addr_libfunc
)
908 tls_get_addr_libfunc
= init_one_libfunc ("__tls_get_addr");
909 return tls_get_addr_libfunc
;
912 /* Try machine-dependent ways of modifying an illegitimate address
913 to be legitimate. If we find one, return the new, valid address. */
916 alpha_legitimize_address_1 (rtx x
, rtx scratch
, enum machine_mode mode
)
918 HOST_WIDE_INT addend
;
920 /* If the address is (plus reg const_int) and the CONST_INT is not a
921 valid offset, compute the high part of the constant and add it to
922 the register. Then our address is (plus temp low-part-const). */
923 if (GET_CODE (x
) == PLUS
924 && REG_P (XEXP (x
, 0))
925 && CONST_INT_P (XEXP (x
, 1))
926 && ! CONSTANT_ADDRESS_P (XEXP (x
, 1)))
928 addend
= INTVAL (XEXP (x
, 1));
933 /* If the address is (const (plus FOO const_int)), find the low-order
934 part of the CONST_INT. Then load FOO plus any high-order part of the
935 CONST_INT into a register. Our address is (plus reg low-part-const).
936 This is done to reduce the number of GOT entries. */
937 if (can_create_pseudo_p ()
938 && GET_CODE (x
) == CONST
939 && GET_CODE (XEXP (x
, 0)) == PLUS
940 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))
942 addend
= INTVAL (XEXP (XEXP (x
, 0), 1));
943 x
= force_reg (Pmode
, XEXP (XEXP (x
, 0), 0));
947 /* If we have a (plus reg const), emit the load as in (2), then add
948 the two registers, and finally generate (plus reg low-part-const) as
950 if (can_create_pseudo_p ()
951 && GET_CODE (x
) == PLUS
952 && REG_P (XEXP (x
, 0))
953 && GET_CODE (XEXP (x
, 1)) == CONST
954 && GET_CODE (XEXP (XEXP (x
, 1), 0)) == PLUS
955 && CONST_INT_P (XEXP (XEXP (XEXP (x
, 1), 0), 1)))
957 addend
= INTVAL (XEXP (XEXP (XEXP (x
, 1), 0), 1));
958 x
= expand_simple_binop (Pmode
, PLUS
, XEXP (x
, 0),
959 XEXP (XEXP (XEXP (x
, 1), 0), 0),
960 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
964 /* If this is a local symbol, split the address into HIGH/LO_SUM parts.
965 Avoid modes larger than word mode since i.e. $LC0+8($1) can fold
966 around +/- 32k offset. */
967 if (TARGET_EXPLICIT_RELOCS
968 && GET_MODE_SIZE (mode
) <= UNITS_PER_WORD
969 && symbolic_operand (x
, Pmode
))
971 rtx r0
, r16
, eqv
, tga
, tp
, insn
, dest
, seq
;
973 switch (tls_symbolic_operand_type (x
))
978 case TLS_MODEL_GLOBAL_DYNAMIC
:
981 r0
= gen_rtx_REG (Pmode
, 0);
982 r16
= gen_rtx_REG (Pmode
, 16);
983 tga
= get_tls_get_addr ();
984 dest
= gen_reg_rtx (Pmode
);
985 seq
= GEN_INT (alpha_next_sequence_number
++);
987 emit_insn (gen_movdi_er_tlsgd (r16
, pic_offset_table_rtx
, x
, seq
));
988 insn
= gen_call_value_osf_tlsgd (r0
, tga
, seq
);
989 insn
= emit_call_insn (insn
);
990 RTL_CONST_CALL_P (insn
) = 1;
991 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), r16
);
996 emit_libcall_block (insn
, dest
, r0
, x
);
999 case TLS_MODEL_LOCAL_DYNAMIC
:
1002 r0
= gen_rtx_REG (Pmode
, 0);
1003 r16
= gen_rtx_REG (Pmode
, 16);
1004 tga
= get_tls_get_addr ();
1005 scratch
= gen_reg_rtx (Pmode
);
1006 seq
= GEN_INT (alpha_next_sequence_number
++);
1008 emit_insn (gen_movdi_er_tlsldm (r16
, pic_offset_table_rtx
, seq
));
1009 insn
= gen_call_value_osf_tlsldm (r0
, tga
, seq
);
1010 insn
= emit_call_insn (insn
);
1011 RTL_CONST_CALL_P (insn
) = 1;
1012 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), r16
);
1014 insn
= get_insns ();
1017 eqv
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
),
1018 UNSPEC_TLSLDM_CALL
);
1019 emit_libcall_block (insn
, scratch
, r0
, eqv
);
1021 eqv
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
), UNSPEC_DTPREL
);
1022 eqv
= gen_rtx_CONST (Pmode
, eqv
);
1024 if (alpha_tls_size
== 64)
1026 dest
= gen_reg_rtx (Pmode
);
1027 emit_insn (gen_rtx_SET (VOIDmode
, dest
, eqv
));
1028 emit_insn (gen_adddi3 (dest
, dest
, scratch
));
1031 if (alpha_tls_size
== 32)
1033 insn
= gen_rtx_HIGH (Pmode
, eqv
);
1034 insn
= gen_rtx_PLUS (Pmode
, scratch
, insn
);
1035 scratch
= gen_reg_rtx (Pmode
);
1036 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, insn
));
1038 return gen_rtx_LO_SUM (Pmode
, scratch
, eqv
);
1040 case TLS_MODEL_INITIAL_EXEC
:
1041 eqv
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
), UNSPEC_TPREL
);
1042 eqv
= gen_rtx_CONST (Pmode
, eqv
);
1043 tp
= gen_reg_rtx (Pmode
);
1044 scratch
= gen_reg_rtx (Pmode
);
1045 dest
= gen_reg_rtx (Pmode
);
1047 emit_insn (gen_load_tp (tp
));
1048 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, eqv
));
1049 emit_insn (gen_adddi3 (dest
, tp
, scratch
));
1052 case TLS_MODEL_LOCAL_EXEC
:
1053 eqv
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
), UNSPEC_TPREL
);
1054 eqv
= gen_rtx_CONST (Pmode
, eqv
);
1055 tp
= gen_reg_rtx (Pmode
);
1057 emit_insn (gen_load_tp (tp
));
1058 if (alpha_tls_size
== 32)
1060 insn
= gen_rtx_HIGH (Pmode
, eqv
);
1061 insn
= gen_rtx_PLUS (Pmode
, tp
, insn
);
1062 tp
= gen_reg_rtx (Pmode
);
1063 emit_insn (gen_rtx_SET (VOIDmode
, tp
, insn
));
1065 return gen_rtx_LO_SUM (Pmode
, tp
, eqv
);
1071 if (local_symbolic_operand (x
, Pmode
))
1073 if (small_symbolic_operand (x
, Pmode
))
1077 if (can_create_pseudo_p ())
1078 scratch
= gen_reg_rtx (Pmode
);
1079 emit_insn (gen_rtx_SET (VOIDmode
, scratch
,
1080 gen_rtx_HIGH (Pmode
, x
)));
1081 return gen_rtx_LO_SUM (Pmode
, scratch
, x
);
1090 HOST_WIDE_INT low
, high
;
1092 low
= ((addend
& 0xffff) ^ 0x8000) - 0x8000;
1094 high
= ((addend
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1098 x
= expand_simple_binop (Pmode
, PLUS
, x
, GEN_INT (addend
),
1099 (!can_create_pseudo_p () ? scratch
: NULL_RTX
),
1100 1, OPTAB_LIB_WIDEN
);
1102 x
= expand_simple_binop (Pmode
, PLUS
, x
, GEN_INT (high
),
1103 (!can_create_pseudo_p () ? scratch
: NULL_RTX
),
1104 1, OPTAB_LIB_WIDEN
);
1106 return plus_constant (x
, low
);
1111 /* Try machine-dependent ways of modifying an illegitimate address
1112 to be legitimate. Return X or the new, valid address. */
1115 alpha_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
1116 enum machine_mode mode
)
1118 rtx new_x
= alpha_legitimize_address_1 (x
, NULL_RTX
, mode
);
1119 return new_x
? new_x
: x
;
1122 /* Primarily this is required for TLS symbols, but given that our move
1123 patterns *ought* to be able to handle any symbol at any time, we
1124 should never be spilling symbolic operands to the constant pool, ever. */
1127 alpha_cannot_force_const_mem (rtx x
)
1129 enum rtx_code code
= GET_CODE (x
);
1130 return code
== SYMBOL_REF
|| code
== LABEL_REF
|| code
== CONST
;
1133 /* We do not allow indirect calls to be optimized into sibling calls, nor
1134 can we allow a call to a function with a different GP to be optimized
1138 alpha_function_ok_for_sibcall (tree decl
, tree exp ATTRIBUTE_UNUSED
)
1140 /* Can't do indirect tail calls, since we don't know if the target
1141 uses the same GP. */
1145 /* Otherwise, we can make a tail call if the target function shares
1147 return decl_has_samegp (decl
);
1151 some_small_symbolic_operand_int (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
1155 /* Don't re-split. */
1156 if (GET_CODE (x
) == LO_SUM
)
1159 return small_symbolic_operand (x
, Pmode
) != 0;
1163 split_small_symbolic_operand_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
1167 /* Don't re-split. */
1168 if (GET_CODE (x
) == LO_SUM
)
1171 if (small_symbolic_operand (x
, Pmode
))
1173 x
= gen_rtx_LO_SUM (Pmode
, pic_offset_table_rtx
, x
);
1182 split_small_symbolic_operand (rtx x
)
1185 for_each_rtx (&x
, split_small_symbolic_operand_1
, NULL
);
1189 /* Indicate that INSN cannot be duplicated. This is true for any insn
1190 that we've marked with gpdisp relocs, since those have to stay in
1191 1-1 correspondence with one another.
1193 Technically we could copy them if we could set up a mapping from one
1194 sequence number to another, across the set of insns to be duplicated.
1195 This seems overly complicated and error-prone since interblock motion
1196 from sched-ebb could move one of the pair of insns to a different block.
1198 Also cannot allow jsr insns to be duplicated. If they throw exceptions,
1199 then they'll be in a different block from their ldgp. Which could lead
1200 the bb reorder code to think that it would be ok to copy just the block
1201 containing the call and branch to the block containing the ldgp. */
1204 alpha_cannot_copy_insn_p (rtx insn
)
1206 if (!reload_completed
|| !TARGET_EXPLICIT_RELOCS
)
1208 if (recog_memoized (insn
) >= 0)
1209 return get_attr_cannot_copy (insn
);
1215 /* Try a machine-dependent way of reloading an illegitimate address
1216 operand. If we find one, push the reload and return the new rtx. */
1219 alpha_legitimize_reload_address (rtx x
,
1220 enum machine_mode mode ATTRIBUTE_UNUSED
,
1221 int opnum
, int type
,
1222 int ind_levels ATTRIBUTE_UNUSED
)
1224 /* We must recognize output that we have already generated ourselves. */
1225 if (GET_CODE (x
) == PLUS
1226 && GET_CODE (XEXP (x
, 0)) == PLUS
1227 && REG_P (XEXP (XEXP (x
, 0), 0))
1228 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
1229 && CONST_INT_P (XEXP (x
, 1)))
1231 push_reload (XEXP (x
, 0), NULL_RTX
, &XEXP (x
, 0), NULL
,
1232 BASE_REG_CLASS
, GET_MODE (x
), VOIDmode
, 0, 0,
1233 opnum
, (enum reload_type
) type
);
1237 /* We wish to handle large displacements off a base register by
1238 splitting the addend across an ldah and the mem insn. This
1239 cuts number of extra insns needed from 3 to 1. */
1240 if (GET_CODE (x
) == PLUS
1241 && REG_P (XEXP (x
, 0))
1242 && REGNO (XEXP (x
, 0)) < FIRST_PSEUDO_REGISTER
1243 && REGNO_OK_FOR_BASE_P (REGNO (XEXP (x
, 0)))
1244 && GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1246 HOST_WIDE_INT val
= INTVAL (XEXP (x
, 1));
1247 HOST_WIDE_INT low
= ((val
& 0xffff) ^ 0x8000) - 0x8000;
1249 = (((val
- low
) & 0xffffffff) ^ 0x80000000) - 0x80000000;
1251 /* Check for 32-bit overflow. */
1252 if (high
+ low
!= val
)
1255 /* Reload the high part into a base reg; leave the low part
1256 in the mem directly. */
1257 x
= gen_rtx_PLUS (GET_MODE (x
),
1258 gen_rtx_PLUS (GET_MODE (x
), XEXP (x
, 0),
1262 push_reload (XEXP (x
, 0), NULL_RTX
, &XEXP (x
, 0), NULL
,
1263 BASE_REG_CLASS
, GET_MODE (x
), VOIDmode
, 0, 0,
1264 opnum
, (enum reload_type
) type
);
1271 /* Compute a (partial) cost for rtx X. Return true if the complete
1272 cost has been computed, and false if subexpressions should be
1273 scanned. In either case, *TOTAL contains the cost result. */
1276 alpha_rtx_costs (rtx x
, int code
, int outer_code
, int *total
,
1279 enum machine_mode mode
= GET_MODE (x
);
1280 bool float_mode_p
= FLOAT_MODE_P (mode
);
1281 const struct alpha_rtx_cost_data
*cost_data
;
1284 cost_data
= &alpha_rtx_cost_size
;
1286 cost_data
= &alpha_rtx_cost_data
[alpha_tune
];
1291 /* If this is an 8-bit constant, return zero since it can be used
1292 nearly anywhere with no cost. If it is a valid operand for an
1293 ADD or AND, likewise return 0 if we know it will be used in that
1294 context. Otherwise, return 2 since it might be used there later.
1295 All other constants take at least two insns. */
1296 if (INTVAL (x
) >= 0 && INTVAL (x
) < 256)
1304 if (x
== CONST0_RTX (mode
))
1306 else if ((outer_code
== PLUS
&& add_operand (x
, VOIDmode
))
1307 || (outer_code
== AND
&& and_operand (x
, VOIDmode
)))
1309 else if (add_operand (x
, VOIDmode
) || and_operand (x
, VOIDmode
))
1312 *total
= COSTS_N_INSNS (2);
1318 if (TARGET_EXPLICIT_RELOCS
&& small_symbolic_operand (x
, VOIDmode
))
1319 *total
= COSTS_N_INSNS (outer_code
!= MEM
);
1320 else if (TARGET_EXPLICIT_RELOCS
&& local_symbolic_operand (x
, VOIDmode
))
1321 *total
= COSTS_N_INSNS (1 + (outer_code
!= MEM
));
1322 else if (tls_symbolic_operand_type (x
))
1323 /* Estimate of cost for call_pal rduniq. */
1324 /* ??? How many insns do we emit here? More than one... */
1325 *total
= COSTS_N_INSNS (15);
1327 /* Otherwise we do a load from the GOT. */
1328 *total
= COSTS_N_INSNS (!speed
? 1 : alpha_memory_latency
);
1332 /* This is effectively an add_operand. */
1339 *total
= cost_data
->fp_add
;
1340 else if (GET_CODE (XEXP (x
, 0)) == MULT
1341 && const48_operand (XEXP (XEXP (x
, 0), 1), VOIDmode
))
1343 *total
= (rtx_cost (XEXP (XEXP (x
, 0), 0),
1344 (enum rtx_code
) outer_code
, speed
)
1345 + rtx_cost (XEXP (x
, 1),
1346 (enum rtx_code
) outer_code
, speed
)
1347 + COSTS_N_INSNS (1));
1354 *total
= cost_data
->fp_mult
;
1355 else if (mode
== DImode
)
1356 *total
= cost_data
->int_mult_di
;
1358 *total
= cost_data
->int_mult_si
;
1362 if (CONST_INT_P (XEXP (x
, 1))
1363 && INTVAL (XEXP (x
, 1)) <= 3)
1365 *total
= COSTS_N_INSNS (1);
1372 *total
= cost_data
->int_shift
;
1377 *total
= cost_data
->fp_add
;
1379 *total
= cost_data
->int_cmov
;
1387 *total
= cost_data
->int_div
;
1388 else if (mode
== SFmode
)
1389 *total
= cost_data
->fp_div_sf
;
1391 *total
= cost_data
->fp_div_df
;
1395 *total
= COSTS_N_INSNS (!speed
? 1 : alpha_memory_latency
);
1401 *total
= COSTS_N_INSNS (1);
1409 *total
= COSTS_N_INSNS (1) + cost_data
->int_cmov
;
1415 case UNSIGNED_FLOAT
:
1418 case FLOAT_TRUNCATE
:
1419 *total
= cost_data
->fp_add
;
1423 if (MEM_P (XEXP (x
, 0)))
1426 *total
= cost_data
->fp_add
;
1434 /* REF is an alignable memory location. Place an aligned SImode
1435 reference into *PALIGNED_MEM and the number of bits to shift into
1436 *PBITNUM. SCRATCH is a free register for use in reloading out
1437 of range stack slots. */
1440 get_aligned_mem (rtx ref
, rtx
*paligned_mem
, rtx
*pbitnum
)
1443 HOST_WIDE_INT disp
, offset
;
1445 gcc_assert (MEM_P (ref
));
1447 if (reload_in_progress
1448 && ! memory_address_p (GET_MODE (ref
), XEXP (ref
, 0)))
1450 base
= find_replacement (&XEXP (ref
, 0));
1451 gcc_assert (memory_address_p (GET_MODE (ref
), base
));
1454 base
= XEXP (ref
, 0);
1456 if (GET_CODE (base
) == PLUS
)
1457 disp
= INTVAL (XEXP (base
, 1)), base
= XEXP (base
, 0);
1461 /* Find the byte offset within an aligned word. If the memory itself is
1462 claimed to be aligned, believe it. Otherwise, aligned_memory_operand
1463 will have examined the base register and determined it is aligned, and
1464 thus displacements from it are naturally alignable. */
1465 if (MEM_ALIGN (ref
) >= 32)
1470 /* Access the entire aligned word. */
1471 *paligned_mem
= widen_memory_access (ref
, SImode
, -offset
);
1473 /* Convert the byte offset within the word to a bit offset. */
1474 if (WORDS_BIG_ENDIAN
)
1475 offset
= 32 - (GET_MODE_BITSIZE (GET_MODE (ref
)) + offset
* 8);
1478 *pbitnum
= GEN_INT (offset
);
1481 /* Similar, but just get the address. Handle the two reload cases.
1482 Add EXTRA_OFFSET to the address we return. */
1485 get_unaligned_address (rtx ref
)
1488 HOST_WIDE_INT offset
= 0;
1490 gcc_assert (MEM_P (ref
));
1492 if (reload_in_progress
1493 && ! memory_address_p (GET_MODE (ref
), XEXP (ref
, 0)))
1495 base
= find_replacement (&XEXP (ref
, 0));
1497 gcc_assert (memory_address_p (GET_MODE (ref
), base
));
1500 base
= XEXP (ref
, 0);
1502 if (GET_CODE (base
) == PLUS
)
1503 offset
+= INTVAL (XEXP (base
, 1)), base
= XEXP (base
, 0);
1505 return plus_constant (base
, offset
);
1508 /* Compute a value X, such that X & 7 == (ADDR + OFS) & 7.
1509 X is always returned in a register. */
1512 get_unaligned_offset (rtx addr
, HOST_WIDE_INT ofs
)
1514 if (GET_CODE (addr
) == PLUS
)
1516 ofs
+= INTVAL (XEXP (addr
, 1));
1517 addr
= XEXP (addr
, 0);
1520 return expand_simple_binop (Pmode
, PLUS
, addr
, GEN_INT (ofs
& 7),
1521 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
1524 /* On the Alpha, all (non-symbolic) constants except zero go into
1525 a floating-point register via memory. Note that we cannot
1526 return anything that is not a subset of RCLASS, and that some
1527 symbolic constants cannot be dropped to memory. */
1530 alpha_preferred_reload_class(rtx x
, enum reg_class rclass
)
1532 /* Zero is present in any register class. */
1533 if (x
== CONST0_RTX (GET_MODE (x
)))
1536 /* These sorts of constants we can easily drop to memory. */
1538 || GET_CODE (x
) == CONST_DOUBLE
1539 || GET_CODE (x
) == CONST_VECTOR
)
1541 if (rclass
== FLOAT_REGS
)
1543 if (rclass
== ALL_REGS
)
1544 return GENERAL_REGS
;
1548 /* All other kinds of constants should not (and in the case of HIGH
1549 cannot) be dropped to memory -- instead we use a GENERAL_REGS
1550 secondary reload. */
1552 return (rclass
== ALL_REGS
? GENERAL_REGS
: rclass
);
1557 /* Inform reload about cases where moving X with a mode MODE to a register in
1558 RCLASS requires an extra scratch or immediate register. Return the class
1559 needed for the immediate register. */
1561 static enum reg_class
1562 alpha_secondary_reload (bool in_p
, rtx x
, enum reg_class rclass
,
1563 enum machine_mode mode
, secondary_reload_info
*sri
)
1565 /* Loading and storing HImode or QImode values to and from memory
1566 usually requires a scratch register. */
1567 if (!TARGET_BWX
&& (mode
== QImode
|| mode
== HImode
|| mode
== CQImode
))
1569 if (any_memory_operand (x
, mode
))
1573 if (!aligned_memory_operand (x
, mode
))
1574 sri
->icode
= reload_in_optab
[mode
];
1577 sri
->icode
= reload_out_optab
[mode
];
1582 /* We also cannot do integral arithmetic into FP regs, as might result
1583 from register elimination into a DImode fp register. */
1584 if (rclass
== FLOAT_REGS
)
1586 if (MEM_P (x
) && GET_CODE (XEXP (x
, 0)) == AND
)
1587 return GENERAL_REGS
;
1588 if (in_p
&& INTEGRAL_MODE_P (mode
)
1589 && !MEM_P (x
) && !REG_P (x
) && !CONST_INT_P (x
))
1590 return GENERAL_REGS
;
1596 /* Subfunction of the following function. Update the flags of any MEM
1597 found in part of X. */
1600 alpha_set_memflags_1 (rtx
*xp
, void *data
)
1602 rtx x
= *xp
, orig
= (rtx
) data
;
1607 MEM_VOLATILE_P (x
) = MEM_VOLATILE_P (orig
);
1608 MEM_IN_STRUCT_P (x
) = MEM_IN_STRUCT_P (orig
);
1609 MEM_SCALAR_P (x
) = MEM_SCALAR_P (orig
);
1610 MEM_NOTRAP_P (x
) = MEM_NOTRAP_P (orig
);
1611 MEM_READONLY_P (x
) = MEM_READONLY_P (orig
);
1613 /* Sadly, we cannot use alias sets because the extra aliasing
1614 produced by the AND interferes. Given that two-byte quantities
1615 are the only thing we would be able to differentiate anyway,
1616 there does not seem to be any point in convoluting the early
1617 out of the alias check. */
1622 /* Given SEQ, which is an INSN list, look for any MEMs in either
1623 a SET_DEST or a SET_SRC and copy the in-struct, unchanging, and
1624 volatile flags from REF into each of the MEMs found. If REF is not
1625 a MEM, don't do anything. */
1628 alpha_set_memflags (rtx seq
, rtx ref
)
1635 /* This is only called from alpha.md, after having had something
1636 generated from one of the insn patterns. So if everything is
1637 zero, the pattern is already up-to-date. */
1638 if (!MEM_VOLATILE_P (ref
)
1639 && !MEM_IN_STRUCT_P (ref
)
1640 && !MEM_SCALAR_P (ref
)
1641 && !MEM_NOTRAP_P (ref
)
1642 && !MEM_READONLY_P (ref
))
1645 for (insn
= seq
; insn
; insn
= NEXT_INSN (insn
))
1647 for_each_rtx (&PATTERN (insn
), alpha_set_memflags_1
, (void *) ref
);
1652 static rtx
alpha_emit_set_const (rtx
, enum machine_mode
, HOST_WIDE_INT
,
1655 /* Internal routine for alpha_emit_set_const to check for N or below insns.
1656 If NO_OUTPUT is true, then we only check to see if N insns are possible,
1657 and return pc_rtx if successful. */
1660 alpha_emit_set_const_1 (rtx target
, enum machine_mode mode
,
1661 HOST_WIDE_INT c
, int n
, bool no_output
)
1663 HOST_WIDE_INT new_const
;
1665 /* Use a pseudo if highly optimizing and still generating RTL. */
1667 = (flag_expensive_optimizations
&& can_create_pseudo_p () ? 0 : target
);
1670 /* If this is a sign-extended 32-bit constant, we can do this in at most
1671 three insns, so do it if we have enough insns left. We always have
1672 a sign-extended 32-bit constant when compiling on a narrow machine. */
1674 if (HOST_BITS_PER_WIDE_INT
!= 64
1675 || c
>> 31 == -1 || c
>> 31 == 0)
1677 HOST_WIDE_INT low
= ((c
& 0xffff) ^ 0x8000) - 0x8000;
1678 HOST_WIDE_INT tmp1
= c
- low
;
1679 HOST_WIDE_INT high
= (((tmp1
>> 16) & 0xffff) ^ 0x8000) - 0x8000;
1680 HOST_WIDE_INT extra
= 0;
1682 /* If HIGH will be interpreted as negative but the constant is
1683 positive, we must adjust it to do two ldha insns. */
1685 if ((high
& 0x8000) != 0 && c
>= 0)
1689 high
= ((tmp1
>> 16) & 0xffff) - 2 * ((tmp1
>> 16) & 0x8000);
1692 if (c
== low
|| (low
== 0 && extra
== 0))
1694 /* We used to use copy_to_suggested_reg (GEN_INT (c), target, mode)
1695 but that meant that we can't handle INT_MIN on 32-bit machines
1696 (like NT/Alpha), because we recurse indefinitely through
1697 emit_move_insn to gen_movdi. So instead, since we know exactly
1698 what we want, create it explicitly. */
1703 target
= gen_reg_rtx (mode
);
1704 emit_insn (gen_rtx_SET (VOIDmode
, target
, GEN_INT (c
)));
1707 else if (n
>= 2 + (extra
!= 0))
1711 if (!can_create_pseudo_p ())
1713 emit_insn (gen_rtx_SET (VOIDmode
, target
, GEN_INT (high
<< 16)));
1717 temp
= copy_to_suggested_reg (GEN_INT (high
<< 16),
1720 /* As of 2002-02-23, addsi3 is only available when not optimizing.
1721 This means that if we go through expand_binop, we'll try to
1722 generate extensions, etc, which will require new pseudos, which
1723 will fail during some split phases. The SImode add patterns
1724 still exist, but are not named. So build the insns by hand. */
1729 subtarget
= gen_reg_rtx (mode
);
1730 insn
= gen_rtx_PLUS (mode
, temp
, GEN_INT (extra
<< 16));
1731 insn
= gen_rtx_SET (VOIDmode
, subtarget
, insn
);
1737 target
= gen_reg_rtx (mode
);
1738 insn
= gen_rtx_PLUS (mode
, temp
, GEN_INT (low
));
1739 insn
= gen_rtx_SET (VOIDmode
, target
, insn
);
1745 /* If we couldn't do it that way, try some other methods. But if we have
1746 no instructions left, don't bother. Likewise, if this is SImode and
1747 we can't make pseudos, we can't do anything since the expand_binop
1748 and expand_unop calls will widen and try to make pseudos. */
1750 if (n
== 1 || (mode
== SImode
&& !can_create_pseudo_p ()))
1753 /* Next, see if we can load a related constant and then shift and possibly
1754 negate it to get the constant we want. Try this once each increasing
1755 numbers of insns. */
1757 for (i
= 1; i
< n
; i
++)
1759 /* First, see if minus some low bits, we've an easy load of
1762 new_const
= ((c
& 0xffff) ^ 0x8000) - 0x8000;
1765 temp
= alpha_emit_set_const (subtarget
, mode
, c
- new_const
, i
, no_output
);
1770 return expand_binop (mode
, add_optab
, temp
, GEN_INT (new_const
),
1771 target
, 0, OPTAB_WIDEN
);
1775 /* Next try complementing. */
1776 temp
= alpha_emit_set_const (subtarget
, mode
, ~c
, i
, no_output
);
1781 return expand_unop (mode
, one_cmpl_optab
, temp
, target
, 0);
1784 /* Next try to form a constant and do a left shift. We can do this
1785 if some low-order bits are zero; the exact_log2 call below tells
1786 us that information. The bits we are shifting out could be any
1787 value, but here we'll just try the 0- and sign-extended forms of
1788 the constant. To try to increase the chance of having the same
1789 constant in more than one insn, start at the highest number of
1790 bits to shift, but try all possibilities in case a ZAPNOT will
1793 bits
= exact_log2 (c
& -c
);
1795 for (; bits
> 0; bits
--)
1797 new_const
= c
>> bits
;
1798 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
, i
, no_output
);
1801 new_const
= (unsigned HOST_WIDE_INT
)c
>> bits
;
1802 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
,
1809 return expand_binop (mode
, ashl_optab
, temp
, GEN_INT (bits
),
1810 target
, 0, OPTAB_WIDEN
);
1814 /* Now try high-order zero bits. Here we try the shifted-in bits as
1815 all zero and all ones. Be careful to avoid shifting outside the
1816 mode and to avoid shifting outside the host wide int size. */
1817 /* On narrow hosts, don't shift a 1 into the high bit, since we'll
1818 confuse the recursive call and set all of the high 32 bits. */
1820 bits
= (MIN (HOST_BITS_PER_WIDE_INT
, GET_MODE_SIZE (mode
) * 8)
1821 - floor_log2 (c
) - 1 - (HOST_BITS_PER_WIDE_INT
< 64));
1823 for (; bits
> 0; bits
--)
1825 new_const
= c
<< bits
;
1826 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
, i
, no_output
);
1829 new_const
= (c
<< bits
) | (((HOST_WIDE_INT
) 1 << bits
) - 1);
1830 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
,
1837 return expand_binop (mode
, lshr_optab
, temp
, GEN_INT (bits
),
1838 target
, 1, OPTAB_WIDEN
);
1842 /* Now try high-order 1 bits. We get that with a sign-extension.
1843 But one bit isn't enough here. Be careful to avoid shifting outside
1844 the mode and to avoid shifting outside the host wide int size. */
1846 bits
= (MIN (HOST_BITS_PER_WIDE_INT
, GET_MODE_SIZE (mode
) * 8)
1847 - floor_log2 (~ c
) - 2);
1849 for (; bits
> 0; bits
--)
1851 new_const
= c
<< bits
;
1852 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
, i
, no_output
);
1855 new_const
= (c
<< bits
) | (((HOST_WIDE_INT
) 1 << bits
) - 1);
1856 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
,
1863 return expand_binop (mode
, ashr_optab
, temp
, GEN_INT (bits
),
1864 target
, 0, OPTAB_WIDEN
);
1869 #if HOST_BITS_PER_WIDE_INT == 64
1870 /* Finally, see if can load a value into the target that is the same as the
1871 constant except that all bytes that are 0 are changed to be 0xff. If we
1872 can, then we can do a ZAPNOT to obtain the desired constant. */
1875 for (i
= 0; i
< 64; i
+= 8)
1876 if ((new_const
& ((HOST_WIDE_INT
) 0xff << i
)) == 0)
1877 new_const
|= (HOST_WIDE_INT
) 0xff << i
;
1879 /* We are only called for SImode and DImode. If this is SImode, ensure that
1880 we are sign extended to a full word. */
1883 new_const
= ((new_const
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1887 temp
= alpha_emit_set_const (subtarget
, mode
, new_const
, n
- 1, no_output
);
1892 return expand_binop (mode
, and_optab
, temp
, GEN_INT (c
| ~ new_const
),
1893 target
, 0, OPTAB_WIDEN
);
1901 /* Try to output insns to set TARGET equal to the constant C if it can be
1902 done in less than N insns. Do all computations in MODE. Returns the place
1903 where the output has been placed if it can be done and the insns have been
1904 emitted. If it would take more than N insns, zero is returned and no
1905 insns and emitted. */
1908 alpha_emit_set_const (rtx target
, enum machine_mode mode
,
1909 HOST_WIDE_INT c
, int n
, bool no_output
)
1911 enum machine_mode orig_mode
= mode
;
1912 rtx orig_target
= target
;
1916 /* If we can't make any pseudos, TARGET is an SImode hard register, we
1917 can't load this constant in one insn, do this in DImode. */
1918 if (!can_create_pseudo_p () && mode
== SImode
1919 && REG_P (target
) && REGNO (target
) < FIRST_PSEUDO_REGISTER
)
1921 result
= alpha_emit_set_const_1 (target
, mode
, c
, 1, no_output
);
1925 target
= no_output
? NULL
: gen_lowpart (DImode
, target
);
1928 else if (mode
== V8QImode
|| mode
== V4HImode
|| mode
== V2SImode
)
1930 target
= no_output
? NULL
: gen_lowpart (DImode
, target
);
1934 /* Try 1 insn, then 2, then up to N. */
1935 for (i
= 1; i
<= n
; i
++)
1937 result
= alpha_emit_set_const_1 (target
, mode
, c
, i
, no_output
);
1945 insn
= get_last_insn ();
1946 set
= single_set (insn
);
1947 if (! CONSTANT_P (SET_SRC (set
)))
1948 set_unique_reg_note (get_last_insn (), REG_EQUAL
, GEN_INT (c
));
1953 /* Allow for the case where we changed the mode of TARGET. */
1956 if (result
== target
)
1957 result
= orig_target
;
1958 else if (mode
!= orig_mode
)
1959 result
= gen_lowpart (orig_mode
, result
);
1965 /* Having failed to find a 3 insn sequence in alpha_emit_set_const,
1966 fall back to a straight forward decomposition. We do this to avoid
1967 exponential run times encountered when looking for longer sequences
1968 with alpha_emit_set_const. */
1971 alpha_emit_set_long_const (rtx target
, HOST_WIDE_INT c1
, HOST_WIDE_INT c2
)
1973 HOST_WIDE_INT d1
, d2
, d3
, d4
;
1975 /* Decompose the entire word */
1976 #if HOST_BITS_PER_WIDE_INT >= 64
1977 gcc_assert (c2
== -(c1
< 0));
1978 d1
= ((c1
& 0xffff) ^ 0x8000) - 0x8000;
1980 d2
= ((c1
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1981 c1
= (c1
- d2
) >> 32;
1982 d3
= ((c1
& 0xffff) ^ 0x8000) - 0x8000;
1984 d4
= ((c1
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1985 gcc_assert (c1
== d4
);
1987 d1
= ((c1
& 0xffff) ^ 0x8000) - 0x8000;
1989 d2
= ((c1
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1990 gcc_assert (c1
== d2
);
1992 d3
= ((c2
& 0xffff) ^ 0x8000) - 0x8000;
1994 d4
= ((c2
& 0xffffffff) ^ 0x80000000) - 0x80000000;
1995 gcc_assert (c2
== d4
);
1998 /* Construct the high word */
2001 emit_move_insn (target
, GEN_INT (d4
));
2003 emit_move_insn (target
, gen_rtx_PLUS (DImode
, target
, GEN_INT (d3
)));
2006 emit_move_insn (target
, GEN_INT (d3
));
2008 /* Shift it into place */
2009 emit_move_insn (target
, gen_rtx_ASHIFT (DImode
, target
, GEN_INT (32)));
2011 /* Add in the low bits. */
2013 emit_move_insn (target
, gen_rtx_PLUS (DImode
, target
, GEN_INT (d2
)));
2015 emit_move_insn (target
, gen_rtx_PLUS (DImode
, target
, GEN_INT (d1
)));
2020 /* Given an integral CONST_INT, CONST_DOUBLE, or CONST_VECTOR, return
2024 alpha_extract_integer (rtx x
, HOST_WIDE_INT
*p0
, HOST_WIDE_INT
*p1
)
2026 HOST_WIDE_INT i0
, i1
;
2028 if (GET_CODE (x
) == CONST_VECTOR
)
2029 x
= simplify_subreg (DImode
, x
, GET_MODE (x
), 0);
2032 if (CONST_INT_P (x
))
2037 else if (HOST_BITS_PER_WIDE_INT
>= 64)
2039 i0
= CONST_DOUBLE_LOW (x
);
2044 i0
= CONST_DOUBLE_LOW (x
);
2045 i1
= CONST_DOUBLE_HIGH (x
);
2052 /* Implement LEGITIMATE_CONSTANT_P. This is all constants for which we
2053 are willing to load the value into a register via a move pattern.
2054 Normally this is all symbolic constants, integral constants that
2055 take three or fewer instructions, and floating-point zero. */
2058 alpha_legitimate_constant_p (rtx x
)
2060 enum machine_mode mode
= GET_MODE (x
);
2061 HOST_WIDE_INT i0
, i1
;
2063 switch (GET_CODE (x
))
2071 /* TLS symbols are never valid. */
2072 return SYMBOL_REF_TLS_MODEL (x
) == 0;
2075 if (x
== CONST0_RTX (mode
))
2077 if (FLOAT_MODE_P (mode
))
2082 if (x
== CONST0_RTX (mode
))
2084 if (GET_MODE_CLASS (mode
) != MODE_VECTOR_INT
)
2086 if (GET_MODE_SIZE (mode
) != 8)
2092 if (TARGET_BUILD_CONSTANTS
)
2094 alpha_extract_integer (x
, &i0
, &i1
);
2095 if (HOST_BITS_PER_WIDE_INT
>= 64 || i1
== (-i0
< 0))
2096 return alpha_emit_set_const_1 (x
, mode
, i0
, 3, true) != NULL
;
2104 /* Operand 1 is known to be a constant, and should require more than one
2105 instruction to load. Emit that multi-part load. */
2108 alpha_split_const_mov (enum machine_mode mode
, rtx
*operands
)
2110 HOST_WIDE_INT i0
, i1
;
2111 rtx temp
= NULL_RTX
;
2113 alpha_extract_integer (operands
[1], &i0
, &i1
);
2115 if (HOST_BITS_PER_WIDE_INT
>= 64 || i1
== -(i0
< 0))
2116 temp
= alpha_emit_set_const (operands
[0], mode
, i0
, 3, false);
2118 if (!temp
&& TARGET_BUILD_CONSTANTS
)
2119 temp
= alpha_emit_set_long_const (operands
[0], i0
, i1
);
2123 if (!rtx_equal_p (operands
[0], temp
))
2124 emit_move_insn (operands
[0], temp
);
2131 /* Expand a move instruction; return true if all work is done.
2132 We don't handle non-bwx subword loads here. */
2135 alpha_expand_mov (enum machine_mode mode
, rtx
*operands
)
2139 /* If the output is not a register, the input must be. */
2140 if (MEM_P (operands
[0])
2141 && ! reg_or_0_operand (operands
[1], mode
))
2142 operands
[1] = force_reg (mode
, operands
[1]);
2144 /* Allow legitimize_address to perform some simplifications. */
2145 if (mode
== Pmode
&& symbolic_operand (operands
[1], mode
))
2147 tmp
= alpha_legitimize_address_1 (operands
[1], operands
[0], mode
);
2150 if (tmp
== operands
[0])
2157 /* Early out for non-constants and valid constants. */
2158 if (! CONSTANT_P (operands
[1]) || input_operand (operands
[1], mode
))
2161 /* Split large integers. */
2162 if (CONST_INT_P (operands
[1])
2163 || GET_CODE (operands
[1]) == CONST_DOUBLE
2164 || GET_CODE (operands
[1]) == CONST_VECTOR
)
2166 if (alpha_split_const_mov (mode
, operands
))
2170 /* Otherwise we've nothing left but to drop the thing to memory. */
2171 tmp
= force_const_mem (mode
, operands
[1]);
2173 if (tmp
== NULL_RTX
)
2176 if (reload_in_progress
)
2178 emit_move_insn (operands
[0], XEXP (tmp
, 0));
2179 operands
[1] = replace_equiv_address (tmp
, operands
[0]);
2182 operands
[1] = validize_mem (tmp
);
2186 /* Expand a non-bwx QImode or HImode move instruction;
2187 return true if all work is done. */
2190 alpha_expand_mov_nobwx (enum machine_mode mode
, rtx
*operands
)
2194 /* If the output is not a register, the input must be. */
2195 if (MEM_P (operands
[0]))
2196 operands
[1] = force_reg (mode
, operands
[1]);
2198 /* Handle four memory cases, unaligned and aligned for either the input
2199 or the output. The only case where we can be called during reload is
2200 for aligned loads; all other cases require temporaries. */
2202 if (any_memory_operand (operands
[1], mode
))
2204 if (aligned_memory_operand (operands
[1], mode
))
2206 if (reload_in_progress
)
2209 seq
= gen_reload_inqi_aligned (operands
[0], operands
[1]);
2211 seq
= gen_reload_inhi_aligned (operands
[0], operands
[1]);
2216 rtx aligned_mem
, bitnum
;
2217 rtx scratch
= gen_reg_rtx (SImode
);
2221 get_aligned_mem (operands
[1], &aligned_mem
, &bitnum
);
2223 subtarget
= operands
[0];
2224 if (REG_P (subtarget
))
2225 subtarget
= gen_lowpart (DImode
, subtarget
), copyout
= false;
2227 subtarget
= gen_reg_rtx (DImode
), copyout
= true;
2230 seq
= gen_aligned_loadqi (subtarget
, aligned_mem
,
2233 seq
= gen_aligned_loadhi (subtarget
, aligned_mem
,
2238 emit_move_insn (operands
[0], gen_lowpart (mode
, subtarget
));
2243 /* Don't pass these as parameters since that makes the generated
2244 code depend on parameter evaluation order which will cause
2245 bootstrap failures. */
2247 rtx temp1
, temp2
, subtarget
, ua
;
2250 temp1
= gen_reg_rtx (DImode
);
2251 temp2
= gen_reg_rtx (DImode
);
2253 subtarget
= operands
[0];
2254 if (REG_P (subtarget
))
2255 subtarget
= gen_lowpart (DImode
, subtarget
), copyout
= false;
2257 subtarget
= gen_reg_rtx (DImode
), copyout
= true;
2259 ua
= get_unaligned_address (operands
[1]);
2261 seq
= gen_unaligned_loadqi (subtarget
, ua
, temp1
, temp2
);
2263 seq
= gen_unaligned_loadhi (subtarget
, ua
, temp1
, temp2
);
2265 alpha_set_memflags (seq
, operands
[1]);
2269 emit_move_insn (operands
[0], gen_lowpart (mode
, subtarget
));
2274 if (any_memory_operand (operands
[0], mode
))
2276 if (aligned_memory_operand (operands
[0], mode
))
2278 rtx aligned_mem
, bitnum
;
2279 rtx temp1
= gen_reg_rtx (SImode
);
2280 rtx temp2
= gen_reg_rtx (SImode
);
2282 get_aligned_mem (operands
[0], &aligned_mem
, &bitnum
);
2284 emit_insn (gen_aligned_store (aligned_mem
, operands
[1], bitnum
,
2289 rtx temp1
= gen_reg_rtx (DImode
);
2290 rtx temp2
= gen_reg_rtx (DImode
);
2291 rtx temp3
= gen_reg_rtx (DImode
);
2292 rtx ua
= get_unaligned_address (operands
[0]);
2295 seq
= gen_unaligned_storeqi (ua
, operands
[1], temp1
, temp2
, temp3
);
2297 seq
= gen_unaligned_storehi (ua
, operands
[1], temp1
, temp2
, temp3
);
2299 alpha_set_memflags (seq
, operands
[0]);
2308 /* Implement the movmisalign patterns. One of the operands is a memory
2309 that is not naturally aligned. Emit instructions to load it. */
2312 alpha_expand_movmisalign (enum machine_mode mode
, rtx
*operands
)
2314 /* Honor misaligned loads, for those we promised to do so. */
2315 if (MEM_P (operands
[1]))
2319 if (register_operand (operands
[0], mode
))
2322 tmp
= gen_reg_rtx (mode
);
2324 alpha_expand_unaligned_load (tmp
, operands
[1], 8, 0, 0);
2325 if (tmp
!= operands
[0])
2326 emit_move_insn (operands
[0], tmp
);
2328 else if (MEM_P (operands
[0]))
2330 if (!reg_or_0_operand (operands
[1], mode
))
2331 operands
[1] = force_reg (mode
, operands
[1]);
2332 alpha_expand_unaligned_store (operands
[0], operands
[1], 8, 0);
2338 /* Generate an unsigned DImode to FP conversion. This is the same code
2339 optabs would emit if we didn't have TFmode patterns.
2341 For SFmode, this is the only construction I've found that can pass
2342 gcc.c-torture/execute/ieee/rbug.c. No scenario that uses DFmode
2343 intermediates will work, because you'll get intermediate rounding
2344 that ruins the end result. Some of this could be fixed by turning
2345 on round-to-positive-infinity, but that requires diddling the fpsr,
2346 which kills performance. I tried turning this around and converting
2347 to a negative number, so that I could turn on /m, but either I did
2348 it wrong or there's something else cause I wound up with the exact
2349 same single-bit error. There is a branch-less form of this same code:
2360 fcmoveq $f10,$f11,$f0
2362 I'm not using it because it's the same number of instructions as
2363 this branch-full form, and it has more serialized long latency
2364 instructions on the critical path.
2366 For DFmode, we can avoid rounding errors by breaking up the word
2367 into two pieces, converting them separately, and adding them back:
2369 LC0: .long 0,0x5f800000
2374 cpyse $f11,$f31,$f10
2375 cpyse $f31,$f11,$f11
2383 This doesn't seem to be a clear-cut win over the optabs form.
2384 It probably all depends on the distribution of numbers being
2385 converted -- in the optabs form, all but high-bit-set has a
2386 much lower minimum execution time. */
2389 alpha_emit_floatuns (rtx operands
[2])
2391 rtx neglab
, donelab
, i0
, i1
, f0
, in
, out
;
2392 enum machine_mode mode
;
2395 in
= force_reg (DImode
, operands
[1]);
2396 mode
= GET_MODE (out
);
2397 neglab
= gen_label_rtx ();
2398 donelab
= gen_label_rtx ();
2399 i0
= gen_reg_rtx (DImode
);
2400 i1
= gen_reg_rtx (DImode
);
2401 f0
= gen_reg_rtx (mode
);
2403 emit_cmp_and_jump_insns (in
, const0_rtx
, LT
, const0_rtx
, DImode
, 0, neglab
);
2405 emit_insn (gen_rtx_SET (VOIDmode
, out
, gen_rtx_FLOAT (mode
, in
)));
2406 emit_jump_insn (gen_jump (donelab
));
2409 emit_label (neglab
);
2411 emit_insn (gen_lshrdi3 (i0
, in
, const1_rtx
));
2412 emit_insn (gen_anddi3 (i1
, in
, const1_rtx
));
2413 emit_insn (gen_iordi3 (i0
, i0
, i1
));
2414 emit_insn (gen_rtx_SET (VOIDmode
, f0
, gen_rtx_FLOAT (mode
, i0
)));
2415 emit_insn (gen_rtx_SET (VOIDmode
, out
, gen_rtx_PLUS (mode
, f0
, f0
)));
2417 emit_label (donelab
);
2420 /* Generate the comparison for a conditional branch. */
2423 alpha_emit_conditional_branch (rtx operands
[], enum machine_mode cmp_mode
)
2425 enum rtx_code cmp_code
, branch_code
;
2426 enum machine_mode branch_mode
= VOIDmode
;
2427 enum rtx_code code
= GET_CODE (operands
[0]);
2428 rtx op0
= operands
[1], op1
= operands
[2];
2431 if (cmp_mode
== TFmode
)
2433 op0
= alpha_emit_xfloating_compare (&code
, op0
, op1
);
2438 /* The general case: fold the comparison code to the types of compares
2439 that we have, choosing the branch as necessary. */
2442 case EQ
: case LE
: case LT
: case LEU
: case LTU
:
2444 /* We have these compares: */
2445 cmp_code
= code
, branch_code
= NE
;
2450 /* These must be reversed. */
2451 cmp_code
= reverse_condition (code
), branch_code
= EQ
;
2454 case GE
: case GT
: case GEU
: case GTU
:
2455 /* For FP, we swap them, for INT, we reverse them. */
2456 if (cmp_mode
== DFmode
)
2458 cmp_code
= swap_condition (code
);
2460 tem
= op0
, op0
= op1
, op1
= tem
;
2464 cmp_code
= reverse_condition (code
);
2473 if (cmp_mode
== DFmode
)
2475 if (flag_unsafe_math_optimizations
&& cmp_code
!= UNORDERED
)
2477 /* When we are not as concerned about non-finite values, and we
2478 are comparing against zero, we can branch directly. */
2479 if (op1
== CONST0_RTX (DFmode
))
2480 cmp_code
= UNKNOWN
, branch_code
= code
;
2481 else if (op0
== CONST0_RTX (DFmode
))
2483 /* Undo the swap we probably did just above. */
2484 tem
= op0
, op0
= op1
, op1
= tem
;
2485 branch_code
= swap_condition (cmp_code
);
2491 /* ??? We mark the branch mode to be CCmode to prevent the
2492 compare and branch from being combined, since the compare
2493 insn follows IEEE rules that the branch does not. */
2494 branch_mode
= CCmode
;
2499 /* The following optimizations are only for signed compares. */
2500 if (code
!= LEU
&& code
!= LTU
&& code
!= GEU
&& code
!= GTU
)
2502 /* Whee. Compare and branch against 0 directly. */
2503 if (op1
== const0_rtx
)
2504 cmp_code
= UNKNOWN
, branch_code
= code
;
2506 /* If the constants doesn't fit into an immediate, but can
2507 be generated by lda/ldah, we adjust the argument and
2508 compare against zero, so we can use beq/bne directly. */
2509 /* ??? Don't do this when comparing against symbols, otherwise
2510 we'll reduce (&x == 0x1234) to (&x-0x1234 == 0), which will
2511 be declared false out of hand (at least for non-weak). */
2512 else if (CONST_INT_P (op1
)
2513 && (code
== EQ
|| code
== NE
)
2514 && !(symbolic_operand (op0
, VOIDmode
)
2515 || (REG_P (op0
) && REG_POINTER (op0
))))
2517 rtx n_op1
= GEN_INT (-INTVAL (op1
));
2519 if (! satisfies_constraint_I (op1
)
2520 && (satisfies_constraint_K (n_op1
)
2521 || satisfies_constraint_L (n_op1
)))
2522 cmp_code
= PLUS
, branch_code
= code
, op1
= n_op1
;
2526 if (!reg_or_0_operand (op0
, DImode
))
2527 op0
= force_reg (DImode
, op0
);
2528 if (cmp_code
!= PLUS
&& !reg_or_8bit_operand (op1
, DImode
))
2529 op1
= force_reg (DImode
, op1
);
2532 /* Emit an initial compare instruction, if necessary. */
2534 if (cmp_code
!= UNKNOWN
)
2536 tem
= gen_reg_rtx (cmp_mode
);
2537 emit_move_insn (tem
, gen_rtx_fmt_ee (cmp_code
, cmp_mode
, op0
, op1
));
2540 /* Emit the branch instruction. */
2541 tem
= gen_rtx_SET (VOIDmode
, pc_rtx
,
2542 gen_rtx_IF_THEN_ELSE (VOIDmode
,
2543 gen_rtx_fmt_ee (branch_code
,
2545 CONST0_RTX (cmp_mode
)),
2546 gen_rtx_LABEL_REF (VOIDmode
,
2549 emit_jump_insn (tem
);
2552 /* Certain simplifications can be done to make invalid setcc operations
2553 valid. Return the final comparison, or NULL if we can't work. */
2556 alpha_emit_setcc (rtx operands
[], enum machine_mode cmp_mode
)
2558 enum rtx_code cmp_code
;
2559 enum rtx_code code
= GET_CODE (operands
[1]);
2560 rtx op0
= operands
[2], op1
= operands
[3];
2563 if (cmp_mode
== TFmode
)
2565 op0
= alpha_emit_xfloating_compare (&code
, op0
, op1
);
2570 if (cmp_mode
== DFmode
&& !TARGET_FIX
)
2573 /* The general case: fold the comparison code to the types of compares
2574 that we have, choosing the branch as necessary. */
2579 case EQ
: case LE
: case LT
: case LEU
: case LTU
:
2581 /* We have these compares. */
2582 if (cmp_mode
== DFmode
)
2583 cmp_code
= code
, code
= NE
;
2587 if (cmp_mode
== DImode
&& op1
== const0_rtx
)
2592 cmp_code
= reverse_condition (code
);
2596 case GE
: case GT
: case GEU
: case GTU
:
2597 /* These normally need swapping, but for integer zero we have
2598 special patterns that recognize swapped operands. */
2599 if (cmp_mode
== DImode
&& op1
== const0_rtx
)
2601 code
= swap_condition (code
);
2602 if (cmp_mode
== DFmode
)
2603 cmp_code
= code
, code
= NE
;
2604 tmp
= op0
, op0
= op1
, op1
= tmp
;
2611 if (cmp_mode
== DImode
)
2613 if (!register_operand (op0
, DImode
))
2614 op0
= force_reg (DImode
, op0
);
2615 if (!reg_or_8bit_operand (op1
, DImode
))
2616 op1
= force_reg (DImode
, op1
);
2619 /* Emit an initial compare instruction, if necessary. */
2620 if (cmp_code
!= UNKNOWN
)
2622 tmp
= gen_reg_rtx (cmp_mode
);
2623 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
2624 gen_rtx_fmt_ee (cmp_code
, cmp_mode
, op0
, op1
)));
2626 op0
= cmp_mode
== DImode
? gen_lowpart (DImode
, tmp
) : tmp
;
2630 /* Emit the setcc instruction. */
2631 emit_insn (gen_rtx_SET (VOIDmode
, operands
[0],
2632 gen_rtx_fmt_ee (code
, DImode
, op0
, op1
)));
2637 /* Rewrite a comparison against zero CMP of the form
2638 (CODE (cc0) (const_int 0)) so it can be written validly in
2639 a conditional move (if_then_else CMP ...).
2640 If both of the operands that set cc0 are nonzero we must emit
2641 an insn to perform the compare (it can't be done within
2642 the conditional move). */
2645 alpha_emit_conditional_move (rtx cmp
, enum machine_mode mode
)
2647 enum rtx_code code
= GET_CODE (cmp
);
2648 enum rtx_code cmov_code
= NE
;
2649 rtx op0
= XEXP (cmp
, 0);
2650 rtx op1
= XEXP (cmp
, 1);
2651 enum machine_mode cmp_mode
2652 = (GET_MODE (op0
) == VOIDmode
? DImode
: GET_MODE (op0
));
2653 enum machine_mode cmov_mode
= VOIDmode
;
2654 int local_fast_math
= flag_unsafe_math_optimizations
;
2657 gcc_assert (cmp_mode
== DFmode
|| cmp_mode
== DImode
);
2659 if (FLOAT_MODE_P (cmp_mode
) != FLOAT_MODE_P (mode
))
2661 enum rtx_code cmp_code
;
2666 /* If we have fp<->int register move instructions, do a cmov by
2667 performing the comparison in fp registers, and move the
2668 zero/nonzero value to integer registers, where we can then
2669 use a normal cmov, or vice-versa. */
2673 case EQ
: case LE
: case LT
: case LEU
: case LTU
:
2674 /* We have these compares. */
2675 cmp_code
= code
, code
= NE
;
2679 /* This must be reversed. */
2680 cmp_code
= EQ
, code
= EQ
;
2683 case GE
: case GT
: case GEU
: case GTU
:
2684 /* These normally need swapping, but for integer zero we have
2685 special patterns that recognize swapped operands. */
2686 if (cmp_mode
== DImode
&& op1
== const0_rtx
)
2687 cmp_code
= code
, code
= NE
;
2690 cmp_code
= swap_condition (code
);
2692 tem
= op0
, op0
= op1
, op1
= tem
;
2700 tem
= gen_reg_rtx (cmp_mode
);
2701 emit_insn (gen_rtx_SET (VOIDmode
, tem
,
2702 gen_rtx_fmt_ee (cmp_code
, cmp_mode
,
2705 cmp_mode
= cmp_mode
== DImode
? DFmode
: DImode
;
2706 op0
= gen_lowpart (cmp_mode
, tem
);
2707 op1
= CONST0_RTX (cmp_mode
);
2708 local_fast_math
= 1;
2711 /* We may be able to use a conditional move directly.
2712 This avoids emitting spurious compares. */
2713 if (signed_comparison_operator (cmp
, VOIDmode
)
2714 && (cmp_mode
== DImode
|| local_fast_math
)
2715 && (op0
== CONST0_RTX (cmp_mode
) || op1
== CONST0_RTX (cmp_mode
)))
2716 return gen_rtx_fmt_ee (code
, VOIDmode
, op0
, op1
);
2718 /* We can't put the comparison inside the conditional move;
2719 emit a compare instruction and put that inside the
2720 conditional move. Make sure we emit only comparisons we have;
2721 swap or reverse as necessary. */
2723 if (!can_create_pseudo_p ())
2728 case EQ
: case LE
: case LT
: case LEU
: case LTU
:
2729 /* We have these compares: */
2733 /* This must be reversed. */
2734 code
= reverse_condition (code
);
2738 case GE
: case GT
: case GEU
: case GTU
:
2739 /* These must be swapped. */
2740 if (op1
!= CONST0_RTX (cmp_mode
))
2742 code
= swap_condition (code
);
2743 tem
= op0
, op0
= op1
, op1
= tem
;
2751 if (cmp_mode
== DImode
)
2753 if (!reg_or_0_operand (op0
, DImode
))
2754 op0
= force_reg (DImode
, op0
);
2755 if (!reg_or_8bit_operand (op1
, DImode
))
2756 op1
= force_reg (DImode
, op1
);
2759 /* ??? We mark the branch mode to be CCmode to prevent the compare
2760 and cmov from being combined, since the compare insn follows IEEE
2761 rules that the cmov does not. */
2762 if (cmp_mode
== DFmode
&& !local_fast_math
)
2765 tem
= gen_reg_rtx (cmp_mode
);
2766 emit_move_insn (tem
, gen_rtx_fmt_ee (code
, cmp_mode
, op0
, op1
));
2767 return gen_rtx_fmt_ee (cmov_code
, cmov_mode
, tem
, CONST0_RTX (cmp_mode
));
2770 /* Simplify a conditional move of two constants into a setcc with
2771 arithmetic. This is done with a splitter since combine would
2772 just undo the work if done during code generation. It also catches
2773 cases we wouldn't have before cse. */
2776 alpha_split_conditional_move (enum rtx_code code
, rtx dest
, rtx cond
,
2777 rtx t_rtx
, rtx f_rtx
)
2779 HOST_WIDE_INT t
, f
, diff
;
2780 enum machine_mode mode
;
2781 rtx target
, subtarget
, tmp
;
2783 mode
= GET_MODE (dest
);
2788 if (((code
== NE
|| code
== EQ
) && diff
< 0)
2789 || (code
== GE
|| code
== GT
))
2791 code
= reverse_condition (code
);
2792 diff
= t
, t
= f
, f
= diff
;
2796 subtarget
= target
= dest
;
2799 target
= gen_lowpart (DImode
, dest
);
2800 if (can_create_pseudo_p ())
2801 subtarget
= gen_reg_rtx (DImode
);
2805 /* Below, we must be careful to use copy_rtx on target and subtarget
2806 in intermediate insns, as they may be a subreg rtx, which may not
2809 if (f
== 0 && exact_log2 (diff
) > 0
2810 /* On EV6, we've got enough shifters to make non-arithmetic shifts
2811 viable over a longer latency cmove. On EV5, the E0 slot is a
2812 scarce resource, and on EV4 shift has the same latency as a cmove. */
2813 && (diff
<= 8 || alpha_tune
== PROCESSOR_EV6
))
2815 tmp
= gen_rtx_fmt_ee (code
, DImode
, cond
, const0_rtx
);
2816 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (subtarget
), tmp
));
2818 tmp
= gen_rtx_ASHIFT (DImode
, copy_rtx (subtarget
),
2819 GEN_INT (exact_log2 (t
)));
2820 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
2822 else if (f
== 0 && t
== -1)
2824 tmp
= gen_rtx_fmt_ee (code
, DImode
, cond
, const0_rtx
);
2825 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (subtarget
), tmp
));
2827 emit_insn (gen_negdi2 (target
, copy_rtx (subtarget
)));
2829 else if (diff
== 1 || diff
== 4 || diff
== 8)
2833 tmp
= gen_rtx_fmt_ee (code
, DImode
, cond
, const0_rtx
);
2834 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (subtarget
), tmp
));
2837 emit_insn (gen_adddi3 (target
, copy_rtx (subtarget
), GEN_INT (f
)));
2840 add_op
= GEN_INT (f
);
2841 if (sext_add_operand (add_op
, mode
))
2843 tmp
= gen_rtx_MULT (DImode
, copy_rtx (subtarget
),
2845 tmp
= gen_rtx_PLUS (DImode
, tmp
, add_op
);
2846 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
2858 /* Look up the function X_floating library function name for the
2861 struct GTY(()) xfloating_op
2863 const enum rtx_code code
;
2864 const char *const GTY((skip
)) osf_func
;
2865 const char *const GTY((skip
)) vms_func
;
2869 static GTY(()) struct xfloating_op xfloating_ops
[] =
2871 { PLUS
, "_OtsAddX", "OTS$ADD_X", 0 },
2872 { MINUS
, "_OtsSubX", "OTS$SUB_X", 0 },
2873 { MULT
, "_OtsMulX", "OTS$MUL_X", 0 },
2874 { DIV
, "_OtsDivX", "OTS$DIV_X", 0 },
2875 { EQ
, "_OtsEqlX", "OTS$EQL_X", 0 },
2876 { NE
, "_OtsNeqX", "OTS$NEQ_X", 0 },
2877 { LT
, "_OtsLssX", "OTS$LSS_X", 0 },
2878 { LE
, "_OtsLeqX", "OTS$LEQ_X", 0 },
2879 { GT
, "_OtsGtrX", "OTS$GTR_X", 0 },
2880 { GE
, "_OtsGeqX", "OTS$GEQ_X", 0 },
2881 { FIX
, "_OtsCvtXQ", "OTS$CVTXQ", 0 },
2882 { FLOAT
, "_OtsCvtQX", "OTS$CVTQX", 0 },
2883 { UNSIGNED_FLOAT
, "_OtsCvtQUX", "OTS$CVTQUX", 0 },
2884 { FLOAT_EXTEND
, "_OtsConvertFloatTX", "OTS$CVT_FLOAT_T_X", 0 },
2885 { FLOAT_TRUNCATE
, "_OtsConvertFloatXT", "OTS$CVT_FLOAT_X_T", 0 }
2888 static GTY(()) struct xfloating_op vax_cvt_ops
[] =
2890 { FLOAT_EXTEND
, "_OtsConvertFloatGX", "OTS$CVT_FLOAT_G_X", 0 },
2891 { FLOAT_TRUNCATE
, "_OtsConvertFloatXG", "OTS$CVT_FLOAT_X_G", 0 }
2895 alpha_lookup_xfloating_lib_func (enum rtx_code code
)
2897 struct xfloating_op
*ops
= xfloating_ops
;
2898 long n
= ARRAY_SIZE (xfloating_ops
);
2901 gcc_assert (TARGET_HAS_XFLOATING_LIBS
);
2903 /* How irritating. Nothing to key off for the main table. */
2904 if (TARGET_FLOAT_VAX
&& (code
== FLOAT_EXTEND
|| code
== FLOAT_TRUNCATE
))
2907 n
= ARRAY_SIZE (vax_cvt_ops
);
2910 for (i
= 0; i
< n
; ++i
, ++ops
)
2911 if (ops
->code
== code
)
2913 rtx func
= ops
->libcall
;
2916 func
= init_one_libfunc (TARGET_ABI_OPEN_VMS
2917 ? ops
->vms_func
: ops
->osf_func
);
2918 ops
->libcall
= func
;
2926 /* Most X_floating operations take the rounding mode as an argument.
2927 Compute that here. */
2930 alpha_compute_xfloating_mode_arg (enum rtx_code code
,
2931 enum alpha_fp_rounding_mode round
)
2937 case ALPHA_FPRM_NORM
:
2940 case ALPHA_FPRM_MINF
:
2943 case ALPHA_FPRM_CHOP
:
2946 case ALPHA_FPRM_DYN
:
2952 /* XXX For reference, round to +inf is mode = 3. */
2955 if (code
== FLOAT_TRUNCATE
&& alpha_fptm
== ALPHA_FPTM_N
)
2961 /* Emit an X_floating library function call.
2963 Note that these functions do not follow normal calling conventions:
2964 TFmode arguments are passed in two integer registers (as opposed to
2965 indirect); TFmode return values appear in R16+R17.
2967 FUNC is the function to call.
2968 TARGET is where the output belongs.
2969 OPERANDS are the inputs.
2970 NOPERANDS is the count of inputs.
2971 EQUIV is the expression equivalent for the function.
2975 alpha_emit_xfloating_libcall (rtx func
, rtx target
, rtx operands
[],
2976 int noperands
, rtx equiv
)
2978 rtx usage
= NULL_RTX
, tmp
, reg
;
2983 for (i
= 0; i
< noperands
; ++i
)
2985 switch (GET_MODE (operands
[i
]))
2988 reg
= gen_rtx_REG (TFmode
, regno
);
2993 reg
= gen_rtx_REG (DFmode
, regno
+ 32);
2998 gcc_assert (CONST_INT_P (operands
[i
]));
3001 reg
= gen_rtx_REG (DImode
, regno
);
3009 emit_move_insn (reg
, operands
[i
]);
3010 usage
= alloc_EXPR_LIST (0, gen_rtx_USE (VOIDmode
, reg
), usage
);
3013 switch (GET_MODE (target
))
3016 reg
= gen_rtx_REG (TFmode
, 16);
3019 reg
= gen_rtx_REG (DFmode
, 32);
3022 reg
= gen_rtx_REG (DImode
, 0);
3028 tmp
= gen_rtx_MEM (QImode
, func
);
3029 tmp
= emit_call_insn (GEN_CALL_VALUE (reg
, tmp
, const0_rtx
,
3030 const0_rtx
, const0_rtx
));
3031 CALL_INSN_FUNCTION_USAGE (tmp
) = usage
;
3032 RTL_CONST_CALL_P (tmp
) = 1;
3037 emit_libcall_block (tmp
, target
, reg
, equiv
);
3040 /* Emit an X_floating library function call for arithmetic (+,-,*,/). */
3043 alpha_emit_xfloating_arith (enum rtx_code code
, rtx operands
[])
3047 rtx out_operands
[3];
3049 func
= alpha_lookup_xfloating_lib_func (code
);
3050 mode
= alpha_compute_xfloating_mode_arg (code
, alpha_fprm
);
3052 out_operands
[0] = operands
[1];
3053 out_operands
[1] = operands
[2];
3054 out_operands
[2] = GEN_INT (mode
);
3055 alpha_emit_xfloating_libcall (func
, operands
[0], out_operands
, 3,
3056 gen_rtx_fmt_ee (code
, TFmode
, operands
[1],
3060 /* Emit an X_floating library function call for a comparison. */
3063 alpha_emit_xfloating_compare (enum rtx_code
*pcode
, rtx op0
, rtx op1
)
3065 enum rtx_code cmp_code
, res_code
;
3066 rtx func
, out
, operands
[2], note
;
3068 /* X_floating library comparison functions return
3072 Convert the compare against the raw return value. */
3100 func
= alpha_lookup_xfloating_lib_func (cmp_code
);
3104 out
= gen_reg_rtx (DImode
);
3106 /* What's actually returned is -1,0,1, not a proper boolean value,
3107 so use an EXPR_LIST as with a generic libcall instead of a
3108 comparison type expression. */
3109 note
= gen_rtx_EXPR_LIST (VOIDmode
, op1
, NULL_RTX
);
3110 note
= gen_rtx_EXPR_LIST (VOIDmode
, op0
, note
);
3111 note
= gen_rtx_EXPR_LIST (VOIDmode
, func
, note
);
3112 alpha_emit_xfloating_libcall (func
, out
, operands
, 2, note
);
3117 /* Emit an X_floating library function call for a conversion. */
3120 alpha_emit_xfloating_cvt (enum rtx_code orig_code
, rtx operands
[])
3122 int noperands
= 1, mode
;
3123 rtx out_operands
[2];
3125 enum rtx_code code
= orig_code
;
3127 if (code
== UNSIGNED_FIX
)
3130 func
= alpha_lookup_xfloating_lib_func (code
);
3132 out_operands
[0] = operands
[1];
3137 mode
= alpha_compute_xfloating_mode_arg (code
, ALPHA_FPRM_CHOP
);
3138 out_operands
[1] = GEN_INT (mode
);
3141 case FLOAT_TRUNCATE
:
3142 mode
= alpha_compute_xfloating_mode_arg (code
, alpha_fprm
);
3143 out_operands
[1] = GEN_INT (mode
);
3150 alpha_emit_xfloating_libcall (func
, operands
[0], out_operands
, noperands
,
3151 gen_rtx_fmt_e (orig_code
,
3152 GET_MODE (operands
[0]),
3156 /* Split a TImode or TFmode move from OP[1] to OP[0] into a pair of
3157 DImode moves from OP[2,3] to OP[0,1]. If FIXUP_OVERLAP is true,
3158 guarantee that the sequence
3161 is valid. Naturally, output operand ordering is little-endian.
3162 This is used by *movtf_internal and *movti_internal. */
3165 alpha_split_tmode_pair (rtx operands
[4], enum machine_mode mode
,
3168 switch (GET_CODE (operands
[1]))
3171 operands
[3] = gen_rtx_REG (DImode
, REGNO (operands
[1]) + 1);
3172 operands
[2] = gen_rtx_REG (DImode
, REGNO (operands
[1]));
3176 operands
[3] = adjust_address (operands
[1], DImode
, 8);
3177 operands
[2] = adjust_address (operands
[1], DImode
, 0);
3182 gcc_assert (operands
[1] == CONST0_RTX (mode
));
3183 operands
[2] = operands
[3] = const0_rtx
;
3190 switch (GET_CODE (operands
[0]))
3193 operands
[1] = gen_rtx_REG (DImode
, REGNO (operands
[0]) + 1);
3194 operands
[0] = gen_rtx_REG (DImode
, REGNO (operands
[0]));
3198 operands
[1] = adjust_address (operands
[0], DImode
, 8);
3199 operands
[0] = adjust_address (operands
[0], DImode
, 0);
3206 if (fixup_overlap
&& reg_overlap_mentioned_p (operands
[0], operands
[3]))
3209 tmp
= operands
[0], operands
[0] = operands
[1], operands
[1] = tmp
;
3210 tmp
= operands
[2], operands
[2] = operands
[3], operands
[3] = tmp
;
3214 /* Implement negtf2 or abstf2. Op0 is destination, op1 is source,
3215 op2 is a register containing the sign bit, operation is the
3216 logical operation to be performed. */
3219 alpha_split_tfmode_frobsign (rtx operands
[3], rtx (*operation
) (rtx
, rtx
, rtx
))
3221 rtx high_bit
= operands
[2];
3225 alpha_split_tmode_pair (operands
, TFmode
, false);
3227 /* Detect three flavors of operand overlap. */
3229 if (rtx_equal_p (operands
[0], operands
[2]))
3231 else if (rtx_equal_p (operands
[1], operands
[2]))
3233 if (rtx_equal_p (operands
[0], high_bit
))
3240 emit_move_insn (operands
[0], operands
[2]);
3242 /* ??? If the destination overlaps both source tf and high_bit, then
3243 assume source tf is dead in its entirety and use the other half
3244 for a scratch register. Otherwise "scratch" is just the proper
3245 destination register. */
3246 scratch
= operands
[move
< 2 ? 1 : 3];
3248 emit_insn ((*operation
) (scratch
, high_bit
, operands
[3]));
3252 emit_move_insn (operands
[0], operands
[2]);
3254 emit_move_insn (operands
[1], scratch
);
3258 /* Use ext[wlq][lh] as the Architecture Handbook describes for extracting
3262 word: ldq_u r1,X(r11) ldq_u r1,X(r11)
3263 ldq_u r2,X+1(r11) ldq_u r2,X+1(r11)
3264 lda r3,X(r11) lda r3,X+2(r11)
3265 extwl r1,r3,r1 extql r1,r3,r1
3266 extwh r2,r3,r2 extqh r2,r3,r2
3267 or r1.r2.r1 or r1,r2,r1
3270 long: ldq_u r1,X(r11) ldq_u r1,X(r11)
3271 ldq_u r2,X+3(r11) ldq_u r2,X+3(r11)
3272 lda r3,X(r11) lda r3,X(r11)
3273 extll r1,r3,r1 extll r1,r3,r1
3274 extlh r2,r3,r2 extlh r2,r3,r2
3275 or r1.r2.r1 addl r1,r2,r1
3277 quad: ldq_u r1,X(r11)
3286 alpha_expand_unaligned_load (rtx tgt
, rtx mem
, HOST_WIDE_INT size
,
3287 HOST_WIDE_INT ofs
, int sign
)
3289 rtx meml
, memh
, addr
, extl
, exth
, tmp
, mema
;
3290 enum machine_mode mode
;
3292 if (TARGET_BWX
&& size
== 2)
3294 meml
= adjust_address (mem
, QImode
, ofs
);
3295 memh
= adjust_address (mem
, QImode
, ofs
+1);
3296 if (BYTES_BIG_ENDIAN
)
3297 tmp
= meml
, meml
= memh
, memh
= tmp
;
3298 extl
= gen_reg_rtx (DImode
);
3299 exth
= gen_reg_rtx (DImode
);
3300 emit_insn (gen_zero_extendqidi2 (extl
, meml
));
3301 emit_insn (gen_zero_extendqidi2 (exth
, memh
));
3302 exth
= expand_simple_binop (DImode
, ASHIFT
, exth
, GEN_INT (8),
3303 NULL
, 1, OPTAB_LIB_WIDEN
);
3304 addr
= expand_simple_binop (DImode
, IOR
, extl
, exth
,
3305 NULL
, 1, OPTAB_LIB_WIDEN
);
3307 if (sign
&& GET_MODE (tgt
) != HImode
)
3309 addr
= gen_lowpart (HImode
, addr
);
3310 emit_insn (gen_extend_insn (tgt
, addr
, GET_MODE (tgt
), HImode
, 0));
3314 if (GET_MODE (tgt
) != DImode
)
3315 addr
= gen_lowpart (GET_MODE (tgt
), addr
);
3316 emit_move_insn (tgt
, addr
);
3321 meml
= gen_reg_rtx (DImode
);
3322 memh
= gen_reg_rtx (DImode
);
3323 addr
= gen_reg_rtx (DImode
);
3324 extl
= gen_reg_rtx (DImode
);
3325 exth
= gen_reg_rtx (DImode
);
3327 mema
= XEXP (mem
, 0);
3328 if (GET_CODE (mema
) == LO_SUM
)
3329 mema
= force_reg (Pmode
, mema
);
3331 /* AND addresses cannot be in any alias set, since they may implicitly
3332 alias surrounding code. Ideally we'd have some alias set that
3333 covered all types except those with alignment 8 or higher. */
3335 tmp
= change_address (mem
, DImode
,
3336 gen_rtx_AND (DImode
,
3337 plus_constant (mema
, ofs
),
3339 set_mem_alias_set (tmp
, 0);
3340 emit_move_insn (meml
, tmp
);
3342 tmp
= change_address (mem
, DImode
,
3343 gen_rtx_AND (DImode
,
3344 plus_constant (mema
, ofs
+ size
- 1),
3346 set_mem_alias_set (tmp
, 0);
3347 emit_move_insn (memh
, tmp
);
3349 if (WORDS_BIG_ENDIAN
&& sign
&& (size
== 2 || size
== 4))
3351 emit_move_insn (addr
, plus_constant (mema
, -1));
3353 emit_insn (gen_extqh_be (extl
, meml
, addr
));
3354 emit_insn (gen_extxl_be (exth
, memh
, GEN_INT (64), addr
));
3356 addr
= expand_binop (DImode
, ior_optab
, extl
, exth
, tgt
, 1, OPTAB_WIDEN
);
3357 addr
= expand_binop (DImode
, ashr_optab
, addr
, GEN_INT (64 - size
*8),
3358 addr
, 1, OPTAB_WIDEN
);
3360 else if (sign
&& size
== 2)
3362 emit_move_insn (addr
, plus_constant (mema
, ofs
+2));
3364 emit_insn (gen_extxl_le (extl
, meml
, GEN_INT (64), addr
));
3365 emit_insn (gen_extqh_le (exth
, memh
, addr
));
3367 /* We must use tgt here for the target. Alpha-vms port fails if we use
3368 addr for the target, because addr is marked as a pointer and combine
3369 knows that pointers are always sign-extended 32-bit values. */
3370 addr
= expand_binop (DImode
, ior_optab
, extl
, exth
, tgt
, 1, OPTAB_WIDEN
);
3371 addr
= expand_binop (DImode
, ashr_optab
, addr
, GEN_INT (48),
3372 addr
, 1, OPTAB_WIDEN
);
3376 if (WORDS_BIG_ENDIAN
)
3378 emit_move_insn (addr
, plus_constant (mema
, ofs
+size
-1));
3382 emit_insn (gen_extwh_be (extl
, meml
, addr
));
3387 emit_insn (gen_extlh_be (extl
, meml
, addr
));
3392 emit_insn (gen_extqh_be (extl
, meml
, addr
));
3399 emit_insn (gen_extxl_be (exth
, memh
, GEN_INT (size
*8), addr
));
3403 emit_move_insn (addr
, plus_constant (mema
, ofs
));
3404 emit_insn (gen_extxl_le (extl
, meml
, GEN_INT (size
*8), addr
));
3408 emit_insn (gen_extwh_le (exth
, memh
, addr
));
3413 emit_insn (gen_extlh_le (exth
, memh
, addr
));
3418 emit_insn (gen_extqh_le (exth
, memh
, addr
));
3427 addr
= expand_binop (mode
, ior_optab
, gen_lowpart (mode
, extl
),
3428 gen_lowpart (mode
, exth
), gen_lowpart (mode
, tgt
),
3433 emit_move_insn (tgt
, gen_lowpart (GET_MODE (tgt
), addr
));
3436 /* Similarly, use ins and msk instructions to perform unaligned stores. */
3439 alpha_expand_unaligned_store (rtx dst
, rtx src
,
3440 HOST_WIDE_INT size
, HOST_WIDE_INT ofs
)
3442 rtx dstl
, dsth
, addr
, insl
, insh
, meml
, memh
, dsta
;
3444 if (TARGET_BWX
&& size
== 2)
3446 if (src
!= const0_rtx
)
3448 dstl
= gen_lowpart (QImode
, src
);
3449 dsth
= expand_simple_binop (DImode
, LSHIFTRT
, src
, GEN_INT (8),
3450 NULL
, 1, OPTAB_LIB_WIDEN
);
3451 dsth
= gen_lowpart (QImode
, dsth
);
3454 dstl
= dsth
= const0_rtx
;
3456 meml
= adjust_address (dst
, QImode
, ofs
);
3457 memh
= adjust_address (dst
, QImode
, ofs
+1);
3458 if (BYTES_BIG_ENDIAN
)
3459 addr
= meml
, meml
= memh
, memh
= addr
;
3461 emit_move_insn (meml
, dstl
);
3462 emit_move_insn (memh
, dsth
);
3466 dstl
= gen_reg_rtx (DImode
);
3467 dsth
= gen_reg_rtx (DImode
);
3468 insl
= gen_reg_rtx (DImode
);
3469 insh
= gen_reg_rtx (DImode
);
3471 dsta
= XEXP (dst
, 0);
3472 if (GET_CODE (dsta
) == LO_SUM
)
3473 dsta
= force_reg (Pmode
, dsta
);
3475 /* AND addresses cannot be in any alias set, since they may implicitly
3476 alias surrounding code. Ideally we'd have some alias set that
3477 covered all types except those with alignment 8 or higher. */
3479 meml
= change_address (dst
, DImode
,
3480 gen_rtx_AND (DImode
,
3481 plus_constant (dsta
, ofs
),
3483 set_mem_alias_set (meml
, 0);
3485 memh
= change_address (dst
, DImode
,
3486 gen_rtx_AND (DImode
,
3487 plus_constant (dsta
, ofs
+ size
- 1),
3489 set_mem_alias_set (memh
, 0);
3491 emit_move_insn (dsth
, memh
);
3492 emit_move_insn (dstl
, meml
);
3493 if (WORDS_BIG_ENDIAN
)
3495 addr
= copy_addr_to_reg (plus_constant (dsta
, ofs
+size
-1));
3497 if (src
!= const0_rtx
)
3502 emit_insn (gen_inswl_be (insh
, gen_lowpart (HImode
,src
), addr
));
3505 emit_insn (gen_insll_be (insh
, gen_lowpart (SImode
,src
), addr
));
3508 emit_insn (gen_insql_be (insh
, gen_lowpart (DImode
,src
), addr
));
3511 emit_insn (gen_insxh (insl
, gen_lowpart (DImode
, src
),
3512 GEN_INT (size
*8), addr
));
3518 emit_insn (gen_mskxl_be (dsth
, dsth
, GEN_INT (0xffff), addr
));
3522 rtx msk
= immed_double_const (0xffffffff, 0, DImode
);
3523 emit_insn (gen_mskxl_be (dsth
, dsth
, msk
, addr
));
3527 emit_insn (gen_mskxl_be (dsth
, dsth
, constm1_rtx
, addr
));
3531 emit_insn (gen_mskxh (dstl
, dstl
, GEN_INT (size
*8), addr
));
3535 addr
= copy_addr_to_reg (plus_constant (dsta
, ofs
));
3537 if (src
!= CONST0_RTX (GET_MODE (src
)))
3539 emit_insn (gen_insxh (insh
, gen_lowpart (DImode
, src
),
3540 GEN_INT (size
*8), addr
));
3545 emit_insn (gen_inswl_le (insl
, gen_lowpart (HImode
, src
), addr
));
3548 emit_insn (gen_insll_le (insl
, gen_lowpart (SImode
, src
), addr
));
3551 emit_insn (gen_insql_le (insl
, src
, addr
));
3556 emit_insn (gen_mskxh (dsth
, dsth
, GEN_INT (size
*8), addr
));
3561 emit_insn (gen_mskxl_le (dstl
, dstl
, GEN_INT (0xffff), addr
));
3565 rtx msk
= immed_double_const (0xffffffff, 0, DImode
);
3566 emit_insn (gen_mskxl_le (dstl
, dstl
, msk
, addr
));
3570 emit_insn (gen_mskxl_le (dstl
, dstl
, constm1_rtx
, addr
));
3575 if (src
!= CONST0_RTX (GET_MODE (src
)))
3577 dsth
= expand_binop (DImode
, ior_optab
, insh
, dsth
, dsth
, 0, OPTAB_WIDEN
);
3578 dstl
= expand_binop (DImode
, ior_optab
, insl
, dstl
, dstl
, 0, OPTAB_WIDEN
);
3581 if (WORDS_BIG_ENDIAN
)
3583 emit_move_insn (meml
, dstl
);
3584 emit_move_insn (memh
, dsth
);
3588 /* Must store high before low for degenerate case of aligned. */
3589 emit_move_insn (memh
, dsth
);
3590 emit_move_insn (meml
, dstl
);
3594 /* The block move code tries to maximize speed by separating loads and
3595 stores at the expense of register pressure: we load all of the data
3596 before we store it back out. There are two secondary effects worth
3597 mentioning, that this speeds copying to/from aligned and unaligned
3598 buffers, and that it makes the code significantly easier to write. */
3600 #define MAX_MOVE_WORDS 8
3602 /* Load an integral number of consecutive unaligned quadwords. */
3605 alpha_expand_unaligned_load_words (rtx
*out_regs
, rtx smem
,
3606 HOST_WIDE_INT words
, HOST_WIDE_INT ofs
)
3608 rtx
const im8
= GEN_INT (-8);
3609 rtx
const i64
= GEN_INT (64);
3610 rtx ext_tmps
[MAX_MOVE_WORDS
], data_regs
[MAX_MOVE_WORDS
+1];
3611 rtx sreg
, areg
, tmp
, smema
;
3614 smema
= XEXP (smem
, 0);
3615 if (GET_CODE (smema
) == LO_SUM
)
3616 smema
= force_reg (Pmode
, smema
);
3618 /* Generate all the tmp registers we need. */
3619 for (i
= 0; i
< words
; ++i
)
3621 data_regs
[i
] = out_regs
[i
];
3622 ext_tmps
[i
] = gen_reg_rtx (DImode
);
3624 data_regs
[words
] = gen_reg_rtx (DImode
);
3627 smem
= adjust_address (smem
, GET_MODE (smem
), ofs
);
3629 /* Load up all of the source data. */
3630 for (i
= 0; i
< words
; ++i
)
3632 tmp
= change_address (smem
, DImode
,
3633 gen_rtx_AND (DImode
,
3634 plus_constant (smema
, 8*i
),
3636 set_mem_alias_set (tmp
, 0);
3637 emit_move_insn (data_regs
[i
], tmp
);
3640 tmp
= change_address (smem
, DImode
,
3641 gen_rtx_AND (DImode
,
3642 plus_constant (smema
, 8*words
- 1),
3644 set_mem_alias_set (tmp
, 0);
3645 emit_move_insn (data_regs
[words
], tmp
);
3647 /* Extract the half-word fragments. Unfortunately DEC decided to make
3648 extxh with offset zero a noop instead of zeroing the register, so
3649 we must take care of that edge condition ourselves with cmov. */
3651 sreg
= copy_addr_to_reg (smema
);
3652 areg
= expand_binop (DImode
, and_optab
, sreg
, GEN_INT (7), NULL
,
3654 if (WORDS_BIG_ENDIAN
)
3655 emit_move_insn (sreg
, plus_constant (sreg
, 7));
3656 for (i
= 0; i
< words
; ++i
)
3658 if (WORDS_BIG_ENDIAN
)
3660 emit_insn (gen_extqh_be (data_regs
[i
], data_regs
[i
], sreg
));
3661 emit_insn (gen_extxl_be (ext_tmps
[i
], data_regs
[i
+1], i64
, sreg
));
3665 emit_insn (gen_extxl_le (data_regs
[i
], data_regs
[i
], i64
, sreg
));
3666 emit_insn (gen_extqh_le (ext_tmps
[i
], data_regs
[i
+1], sreg
));
3668 emit_insn (gen_rtx_SET (VOIDmode
, ext_tmps
[i
],
3669 gen_rtx_IF_THEN_ELSE (DImode
,
3670 gen_rtx_EQ (DImode
, areg
,
3672 const0_rtx
, ext_tmps
[i
])));
3675 /* Merge the half-words into whole words. */
3676 for (i
= 0; i
< words
; ++i
)
3678 out_regs
[i
] = expand_binop (DImode
, ior_optab
, data_regs
[i
],
3679 ext_tmps
[i
], data_regs
[i
], 1, OPTAB_WIDEN
);
3683 /* Store an integral number of consecutive unaligned quadwords. DATA_REGS
3684 may be NULL to store zeros. */
3687 alpha_expand_unaligned_store_words (rtx
*data_regs
, rtx dmem
,
3688 HOST_WIDE_INT words
, HOST_WIDE_INT ofs
)
3690 rtx
const im8
= GEN_INT (-8);
3691 rtx
const i64
= GEN_INT (64);
3692 rtx ins_tmps
[MAX_MOVE_WORDS
];
3693 rtx st_tmp_1
, st_tmp_2
, dreg
;
3694 rtx st_addr_1
, st_addr_2
, dmema
;
3697 dmema
= XEXP (dmem
, 0);
3698 if (GET_CODE (dmema
) == LO_SUM
)
3699 dmema
= force_reg (Pmode
, dmema
);
3701 /* Generate all the tmp registers we need. */
3702 if (data_regs
!= NULL
)
3703 for (i
= 0; i
< words
; ++i
)
3704 ins_tmps
[i
] = gen_reg_rtx(DImode
);
3705 st_tmp_1
= gen_reg_rtx(DImode
);
3706 st_tmp_2
= gen_reg_rtx(DImode
);
3709 dmem
= adjust_address (dmem
, GET_MODE (dmem
), ofs
);
3711 st_addr_2
= change_address (dmem
, DImode
,
3712 gen_rtx_AND (DImode
,
3713 plus_constant (dmema
, words
*8 - 1),
3715 set_mem_alias_set (st_addr_2
, 0);
3717 st_addr_1
= change_address (dmem
, DImode
,
3718 gen_rtx_AND (DImode
, dmema
, im8
));
3719 set_mem_alias_set (st_addr_1
, 0);
3721 /* Load up the destination end bits. */
3722 emit_move_insn (st_tmp_2
, st_addr_2
);
3723 emit_move_insn (st_tmp_1
, st_addr_1
);
3725 /* Shift the input data into place. */
3726 dreg
= copy_addr_to_reg (dmema
);
3727 if (WORDS_BIG_ENDIAN
)
3728 emit_move_insn (dreg
, plus_constant (dreg
, 7));
3729 if (data_regs
!= NULL
)
3731 for (i
= words
-1; i
>= 0; --i
)
3733 if (WORDS_BIG_ENDIAN
)
3735 emit_insn (gen_insql_be (ins_tmps
[i
], data_regs
[i
], dreg
));
3736 emit_insn (gen_insxh (data_regs
[i
], data_regs
[i
], i64
, dreg
));
3740 emit_insn (gen_insxh (ins_tmps
[i
], data_regs
[i
], i64
, dreg
));
3741 emit_insn (gen_insql_le (data_regs
[i
], data_regs
[i
], dreg
));
3744 for (i
= words
-1; i
> 0; --i
)
3746 ins_tmps
[i
-1] = expand_binop (DImode
, ior_optab
, data_regs
[i
],
3747 ins_tmps
[i
-1], ins_tmps
[i
-1], 1,
3752 /* Split and merge the ends with the destination data. */
3753 if (WORDS_BIG_ENDIAN
)
3755 emit_insn (gen_mskxl_be (st_tmp_2
, st_tmp_2
, constm1_rtx
, dreg
));
3756 emit_insn (gen_mskxh (st_tmp_1
, st_tmp_1
, i64
, dreg
));
3760 emit_insn (gen_mskxh (st_tmp_2
, st_tmp_2
, i64
, dreg
));
3761 emit_insn (gen_mskxl_le (st_tmp_1
, st_tmp_1
, constm1_rtx
, dreg
));
3764 if (data_regs
!= NULL
)
3766 st_tmp_2
= expand_binop (DImode
, ior_optab
, st_tmp_2
, ins_tmps
[words
-1],
3767 st_tmp_2
, 1, OPTAB_WIDEN
);
3768 st_tmp_1
= expand_binop (DImode
, ior_optab
, st_tmp_1
, data_regs
[0],
3769 st_tmp_1
, 1, OPTAB_WIDEN
);
3773 if (WORDS_BIG_ENDIAN
)
3774 emit_move_insn (st_addr_1
, st_tmp_1
);
3776 emit_move_insn (st_addr_2
, st_tmp_2
);
3777 for (i
= words
-1; i
> 0; --i
)
3779 rtx tmp
= change_address (dmem
, DImode
,
3780 gen_rtx_AND (DImode
,
3781 plus_constant(dmema
,
3782 WORDS_BIG_ENDIAN
? i
*8-1 : i
*8),
3784 set_mem_alias_set (tmp
, 0);
3785 emit_move_insn (tmp
, data_regs
? ins_tmps
[i
-1] : const0_rtx
);
3787 if (WORDS_BIG_ENDIAN
)
3788 emit_move_insn (st_addr_2
, st_tmp_2
);
3790 emit_move_insn (st_addr_1
, st_tmp_1
);
3794 /* Expand string/block move operations.
3796 operands[0] is the pointer to the destination.
3797 operands[1] is the pointer to the source.
3798 operands[2] is the number of bytes to move.
3799 operands[3] is the alignment. */
3802 alpha_expand_block_move (rtx operands
[])
3804 rtx bytes_rtx
= operands
[2];
3805 rtx align_rtx
= operands
[3];
3806 HOST_WIDE_INT orig_bytes
= INTVAL (bytes_rtx
);
3807 HOST_WIDE_INT bytes
= orig_bytes
;
3808 HOST_WIDE_INT src_align
= INTVAL (align_rtx
) * BITS_PER_UNIT
;
3809 HOST_WIDE_INT dst_align
= src_align
;
3810 rtx orig_src
= operands
[1];
3811 rtx orig_dst
= operands
[0];
3812 rtx data_regs
[2 * MAX_MOVE_WORDS
+ 16];
3814 unsigned int i
, words
, ofs
, nregs
= 0;
3816 if (orig_bytes
<= 0)
3818 else if (orig_bytes
> MAX_MOVE_WORDS
* UNITS_PER_WORD
)
3821 /* Look for additional alignment information from recorded register info. */
3823 tmp
= XEXP (orig_src
, 0);
3825 src_align
= MAX (src_align
, REGNO_POINTER_ALIGN (REGNO (tmp
)));
3826 else if (GET_CODE (tmp
) == PLUS
3827 && REG_P (XEXP (tmp
, 0))
3828 && CONST_INT_P (XEXP (tmp
, 1)))
3830 unsigned HOST_WIDE_INT c
= INTVAL (XEXP (tmp
, 1));
3831 unsigned int a
= REGNO_POINTER_ALIGN (REGNO (XEXP (tmp
, 0)));
3835 if (a
>= 64 && c
% 8 == 0)
3837 else if (a
>= 32 && c
% 4 == 0)
3839 else if (a
>= 16 && c
% 2 == 0)
3844 tmp
= XEXP (orig_dst
, 0);
3846 dst_align
= MAX (dst_align
, REGNO_POINTER_ALIGN (REGNO (tmp
)));
3847 else if (GET_CODE (tmp
) == PLUS
3848 && REG_P (XEXP (tmp
, 0))
3849 && CONST_INT_P (XEXP (tmp
, 1)))
3851 unsigned HOST_WIDE_INT c
= INTVAL (XEXP (tmp
, 1));
3852 unsigned int a
= REGNO_POINTER_ALIGN (REGNO (XEXP (tmp
, 0)));
3856 if (a
>= 64 && c
% 8 == 0)
3858 else if (a
>= 32 && c
% 4 == 0)
3860 else if (a
>= 16 && c
% 2 == 0)
3866 if (src_align
>= 64 && bytes
>= 8)
3870 for (i
= 0; i
< words
; ++i
)
3871 data_regs
[nregs
+ i
] = gen_reg_rtx (DImode
);
3873 for (i
= 0; i
< words
; ++i
)
3874 emit_move_insn (data_regs
[nregs
+ i
],
3875 adjust_address (orig_src
, DImode
, ofs
+ i
* 8));
3882 if (src_align
>= 32 && bytes
>= 4)
3886 for (i
= 0; i
< words
; ++i
)
3887 data_regs
[nregs
+ i
] = gen_reg_rtx (SImode
);
3889 for (i
= 0; i
< words
; ++i
)
3890 emit_move_insn (data_regs
[nregs
+ i
],
3891 adjust_address (orig_src
, SImode
, ofs
+ i
* 4));
3902 for (i
= 0; i
< words
+1; ++i
)
3903 data_regs
[nregs
+ i
] = gen_reg_rtx (DImode
);
3905 alpha_expand_unaligned_load_words (data_regs
+ nregs
, orig_src
,
3913 if (! TARGET_BWX
&& bytes
>= 4)
3915 data_regs
[nregs
++] = tmp
= gen_reg_rtx (SImode
);
3916 alpha_expand_unaligned_load (tmp
, orig_src
, 4, ofs
, 0);
3923 if (src_align
>= 16)
3926 data_regs
[nregs
++] = tmp
= gen_reg_rtx (HImode
);
3927 emit_move_insn (tmp
, adjust_address (orig_src
, HImode
, ofs
));
3930 } while (bytes
>= 2);
3932 else if (! TARGET_BWX
)
3934 data_regs
[nregs
++] = tmp
= gen_reg_rtx (HImode
);
3935 alpha_expand_unaligned_load (tmp
, orig_src
, 2, ofs
, 0);
3943 data_regs
[nregs
++] = tmp
= gen_reg_rtx (QImode
);
3944 emit_move_insn (tmp
, adjust_address (orig_src
, QImode
, ofs
));
3949 gcc_assert (nregs
<= ARRAY_SIZE (data_regs
));
3951 /* Now save it back out again. */
3955 /* Write out the data in whatever chunks reading the source allowed. */
3956 if (dst_align
>= 64)
3958 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == DImode
)
3960 emit_move_insn (adjust_address (orig_dst
, DImode
, ofs
),
3967 if (dst_align
>= 32)
3969 /* If the source has remaining DImode regs, write them out in
3971 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == DImode
)
3973 tmp
= expand_binop (DImode
, lshr_optab
, data_regs
[i
], GEN_INT (32),
3974 NULL_RTX
, 1, OPTAB_WIDEN
);
3976 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
),
3977 gen_lowpart (SImode
, data_regs
[i
]));
3978 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
+ 4),
3979 gen_lowpart (SImode
, tmp
));
3984 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == SImode
)
3986 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
),
3993 if (i
< nregs
&& GET_MODE (data_regs
[i
]) == DImode
)
3995 /* Write out a remaining block of words using unaligned methods. */
3997 for (words
= 1; i
+ words
< nregs
; words
++)
3998 if (GET_MODE (data_regs
[i
+ words
]) != DImode
)
4002 alpha_expand_unaligned_store (orig_dst
, data_regs
[i
], 8, ofs
);
4004 alpha_expand_unaligned_store_words (data_regs
+ i
, orig_dst
,
4011 /* Due to the above, this won't be aligned. */
4012 /* ??? If we have more than one of these, consider constructing full
4013 words in registers and using alpha_expand_unaligned_store_words. */
4014 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == SImode
)
4016 alpha_expand_unaligned_store (orig_dst
, data_regs
[i
], 4, ofs
);
4021 if (dst_align
>= 16)
4022 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == HImode
)
4024 emit_move_insn (adjust_address (orig_dst
, HImode
, ofs
), data_regs
[i
]);
4029 while (i
< nregs
&& GET_MODE (data_regs
[i
]) == HImode
)
4031 alpha_expand_unaligned_store (orig_dst
, data_regs
[i
], 2, ofs
);
4036 /* The remainder must be byte copies. */
4039 gcc_assert (GET_MODE (data_regs
[i
]) == QImode
);
4040 emit_move_insn (adjust_address (orig_dst
, QImode
, ofs
), data_regs
[i
]);
4049 alpha_expand_block_clear (rtx operands
[])
4051 rtx bytes_rtx
= operands
[1];
4052 rtx align_rtx
= operands
[3];
4053 HOST_WIDE_INT orig_bytes
= INTVAL (bytes_rtx
);
4054 HOST_WIDE_INT bytes
= orig_bytes
;
4055 HOST_WIDE_INT align
= INTVAL (align_rtx
) * BITS_PER_UNIT
;
4056 HOST_WIDE_INT alignofs
= 0;
4057 rtx orig_dst
= operands
[0];
4059 int i
, words
, ofs
= 0;
4061 if (orig_bytes
<= 0)
4063 if (orig_bytes
> MAX_MOVE_WORDS
* UNITS_PER_WORD
)
4066 /* Look for stricter alignment. */
4067 tmp
= XEXP (orig_dst
, 0);
4069 align
= MAX (align
, REGNO_POINTER_ALIGN (REGNO (tmp
)));
4070 else if (GET_CODE (tmp
) == PLUS
4071 && REG_P (XEXP (tmp
, 0))
4072 && CONST_INT_P (XEXP (tmp
, 1)))
4074 HOST_WIDE_INT c
= INTVAL (XEXP (tmp
, 1));
4075 int a
= REGNO_POINTER_ALIGN (REGNO (XEXP (tmp
, 0)));
4080 align
= a
, alignofs
= 8 - c
% 8;
4082 align
= a
, alignofs
= 4 - c
% 4;
4084 align
= a
, alignofs
= 2 - c
% 2;
4088 /* Handle an unaligned prefix first. */
4092 #if HOST_BITS_PER_WIDE_INT >= 64
4093 /* Given that alignofs is bounded by align, the only time BWX could
4094 generate three stores is for a 7 byte fill. Prefer two individual
4095 stores over a load/mask/store sequence. */
4096 if ((!TARGET_BWX
|| alignofs
== 7)
4098 && !(alignofs
== 4 && bytes
>= 4))
4100 enum machine_mode mode
= (align
>= 64 ? DImode
: SImode
);
4101 int inv_alignofs
= (align
>= 64 ? 8 : 4) - alignofs
;
4105 mem
= adjust_address (orig_dst
, mode
, ofs
- inv_alignofs
);
4106 set_mem_alias_set (mem
, 0);
4108 mask
= ~(~(HOST_WIDE_INT
)0 << (inv_alignofs
* 8));
4109 if (bytes
< alignofs
)
4111 mask
|= ~(HOST_WIDE_INT
)0 << ((inv_alignofs
+ bytes
) * 8);
4122 tmp
= expand_binop (mode
, and_optab
, mem
, GEN_INT (mask
),
4123 NULL_RTX
, 1, OPTAB_WIDEN
);
4125 emit_move_insn (mem
, tmp
);
4129 if (TARGET_BWX
&& (alignofs
& 1) && bytes
>= 1)
4131 emit_move_insn (adjust_address (orig_dst
, QImode
, ofs
), const0_rtx
);
4136 if (TARGET_BWX
&& align
>= 16 && (alignofs
& 3) == 2 && bytes
>= 2)
4138 emit_move_insn (adjust_address (orig_dst
, HImode
, ofs
), const0_rtx
);
4143 if (alignofs
== 4 && bytes
>= 4)
4145 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
), const0_rtx
);
4151 /* If we've not used the extra lead alignment information by now,
4152 we won't be able to. Downgrade align to match what's left over. */
4155 alignofs
= alignofs
& -alignofs
;
4156 align
= MIN (align
, alignofs
* BITS_PER_UNIT
);
4160 /* Handle a block of contiguous long-words. */
4162 if (align
>= 64 && bytes
>= 8)
4166 for (i
= 0; i
< words
; ++i
)
4167 emit_move_insn (adjust_address (orig_dst
, DImode
, ofs
+ i
* 8),
4174 /* If the block is large and appropriately aligned, emit a single
4175 store followed by a sequence of stq_u insns. */
4177 if (align
>= 32 && bytes
> 16)
4181 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
), const0_rtx
);
4185 orig_dsta
= XEXP (orig_dst
, 0);
4186 if (GET_CODE (orig_dsta
) == LO_SUM
)
4187 orig_dsta
= force_reg (Pmode
, orig_dsta
);
4190 for (i
= 0; i
< words
; ++i
)
4193 = change_address (orig_dst
, DImode
,
4194 gen_rtx_AND (DImode
,
4195 plus_constant (orig_dsta
, ofs
+ i
*8),
4197 set_mem_alias_set (mem
, 0);
4198 emit_move_insn (mem
, const0_rtx
);
4201 /* Depending on the alignment, the first stq_u may have overlapped
4202 with the initial stl, which means that the last stq_u didn't
4203 write as much as it would appear. Leave those questionable bytes
4205 bytes
-= words
* 8 - 4;
4206 ofs
+= words
* 8 - 4;
4209 /* Handle a smaller block of aligned words. */
4211 if ((align
>= 64 && bytes
== 4)
4212 || (align
== 32 && bytes
>= 4))
4216 for (i
= 0; i
< words
; ++i
)
4217 emit_move_insn (adjust_address (orig_dst
, SImode
, ofs
+ i
* 4),
4224 /* An unaligned block uses stq_u stores for as many as possible. */
4230 alpha_expand_unaligned_store_words (NULL
, orig_dst
, words
, ofs
);
4236 /* Next clean up any trailing pieces. */
4238 #if HOST_BITS_PER_WIDE_INT >= 64
4239 /* Count the number of bits in BYTES for which aligned stores could
4242 for (i
= (TARGET_BWX
? 1 : 4); i
* BITS_PER_UNIT
<= align
; i
<<= 1)
4246 /* If we have appropriate alignment (and it wouldn't take too many
4247 instructions otherwise), mask out the bytes we need. */
4248 if (TARGET_BWX
? words
> 2 : bytes
> 0)
4255 mem
= adjust_address (orig_dst
, DImode
, ofs
);
4256 set_mem_alias_set (mem
, 0);
4258 mask
= ~(HOST_WIDE_INT
)0 << (bytes
* 8);
4260 tmp
= expand_binop (DImode
, and_optab
, mem
, GEN_INT (mask
),
4261 NULL_RTX
, 1, OPTAB_WIDEN
);
4263 emit_move_insn (mem
, tmp
);
4266 else if (align
>= 32 && bytes
< 4)
4271 mem
= adjust_address (orig_dst
, SImode
, ofs
);
4272 set_mem_alias_set (mem
, 0);
4274 mask
= ~(HOST_WIDE_INT
)0 << (bytes
* 8);
4276 tmp
= expand_binop (SImode
, and_optab
, mem
, GEN_INT (mask
),
4277 NULL_RTX
, 1, OPTAB_WIDEN
);
4279 emit_move_insn (mem
, tmp
);
4285 if (!TARGET_BWX
&& bytes
>= 4)
4287 alpha_expand_unaligned_store (orig_dst
, const0_rtx
, 4, ofs
);
4297 emit_move_insn (adjust_address (orig_dst
, HImode
, ofs
),
4301 } while (bytes
>= 2);
4303 else if (! TARGET_BWX
)
4305 alpha_expand_unaligned_store (orig_dst
, const0_rtx
, 2, ofs
);
4313 emit_move_insn (adjust_address (orig_dst
, QImode
, ofs
), const0_rtx
);
4321 /* Returns a mask so that zap(x, value) == x & mask. */
4324 alpha_expand_zap_mask (HOST_WIDE_INT value
)
4329 if (HOST_BITS_PER_WIDE_INT
>= 64)
4331 HOST_WIDE_INT mask
= 0;
4333 for (i
= 7; i
>= 0; --i
)
4336 if (!((value
>> i
) & 1))
4340 result
= gen_int_mode (mask
, DImode
);
4344 HOST_WIDE_INT mask_lo
= 0, mask_hi
= 0;
4346 gcc_assert (HOST_BITS_PER_WIDE_INT
== 32);
4348 for (i
= 7; i
>= 4; --i
)
4351 if (!((value
>> i
) & 1))
4355 for (i
= 3; i
>= 0; --i
)
4358 if (!((value
>> i
) & 1))
4362 result
= immed_double_const (mask_lo
, mask_hi
, DImode
);
4369 alpha_expand_builtin_vector_binop (rtx (*gen
) (rtx
, rtx
, rtx
),
4370 enum machine_mode mode
,
4371 rtx op0
, rtx op1
, rtx op2
)
4373 op0
= gen_lowpart (mode
, op0
);
4375 if (op1
== const0_rtx
)
4376 op1
= CONST0_RTX (mode
);
4378 op1
= gen_lowpart (mode
, op1
);
4380 if (op2
== const0_rtx
)
4381 op2
= CONST0_RTX (mode
);
4383 op2
= gen_lowpart (mode
, op2
);
4385 emit_insn ((*gen
) (op0
, op1
, op2
));
4388 /* A subroutine of the atomic operation splitters. Jump to LABEL if
4389 COND is true. Mark the jump as unlikely to be taken. */
4392 emit_unlikely_jump (rtx cond
, rtx label
)
4394 rtx very_unlikely
= GEN_INT (REG_BR_PROB_BASE
/ 100 - 1);
4397 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, label
, pc_rtx
);
4398 x
= emit_jump_insn (gen_rtx_SET (VOIDmode
, pc_rtx
, x
));
4399 add_reg_note (x
, REG_BR_PROB
, very_unlikely
);
4402 /* A subroutine of the atomic operation splitters. Emit a load-locked
4403 instruction in MODE. */
4406 emit_load_locked (enum machine_mode mode
, rtx reg
, rtx mem
)
4408 rtx (*fn
) (rtx
, rtx
) = NULL
;
4410 fn
= gen_load_locked_si
;
4411 else if (mode
== DImode
)
4412 fn
= gen_load_locked_di
;
4413 emit_insn (fn (reg
, mem
));
4416 /* A subroutine of the atomic operation splitters. Emit a store-conditional
4417 instruction in MODE. */
4420 emit_store_conditional (enum machine_mode mode
, rtx res
, rtx mem
, rtx val
)
4422 rtx (*fn
) (rtx
, rtx
, rtx
) = NULL
;
4424 fn
= gen_store_conditional_si
;
4425 else if (mode
== DImode
)
4426 fn
= gen_store_conditional_di
;
4427 emit_insn (fn (res
, mem
, val
));
4430 /* A subroutine of the atomic operation splitters. Emit an insxl
4431 instruction in MODE. */
4434 emit_insxl (enum machine_mode mode
, rtx op1
, rtx op2
)
4436 rtx ret
= gen_reg_rtx (DImode
);
4437 rtx (*fn
) (rtx
, rtx
, rtx
);
4439 if (WORDS_BIG_ENDIAN
)
4453 /* The insbl and inswl patterns require a register operand. */
4454 op1
= force_reg (mode
, op1
);
4455 emit_insn (fn (ret
, op1
, op2
));
4460 /* Expand an atomic fetch-and-operate pattern. CODE is the binary operation
4461 to perform. MEM is the memory on which to operate. VAL is the second
4462 operand of the binary operator. BEFORE and AFTER are optional locations to
4463 return the value of MEM either before of after the operation. SCRATCH is
4464 a scratch register. */
4467 alpha_split_atomic_op (enum rtx_code code
, rtx mem
, rtx val
,
4468 rtx before
, rtx after
, rtx scratch
)
4470 enum machine_mode mode
= GET_MODE (mem
);
4471 rtx label
, x
, cond
= gen_rtx_REG (DImode
, REGNO (scratch
));
4473 emit_insn (gen_memory_barrier ());
4475 label
= gen_label_rtx ();
4477 label
= gen_rtx_LABEL_REF (DImode
, label
);
4481 emit_load_locked (mode
, before
, mem
);
4485 x
= gen_rtx_AND (mode
, before
, val
);
4486 emit_insn (gen_rtx_SET (VOIDmode
, val
, x
));
4488 x
= gen_rtx_NOT (mode
, val
);
4491 x
= gen_rtx_fmt_ee (code
, mode
, before
, val
);
4493 emit_insn (gen_rtx_SET (VOIDmode
, after
, copy_rtx (x
)));
4494 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, x
));
4496 emit_store_conditional (mode
, cond
, mem
, scratch
);
4498 x
= gen_rtx_EQ (DImode
, cond
, const0_rtx
);
4499 emit_unlikely_jump (x
, label
);
4501 emit_insn (gen_memory_barrier ());
4504 /* Expand a compare and swap operation. */
4507 alpha_split_compare_and_swap (rtx retval
, rtx mem
, rtx oldval
, rtx newval
,
4510 enum machine_mode mode
= GET_MODE (mem
);
4511 rtx label1
, label2
, x
, cond
= gen_lowpart (DImode
, scratch
);
4513 emit_insn (gen_memory_barrier ());
4515 label1
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4516 label2
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4517 emit_label (XEXP (label1
, 0));
4519 emit_load_locked (mode
, retval
, mem
);
4521 x
= gen_lowpart (DImode
, retval
);
4522 if (oldval
== const0_rtx
)
4523 x
= gen_rtx_NE (DImode
, x
, const0_rtx
);
4526 x
= gen_rtx_EQ (DImode
, x
, oldval
);
4527 emit_insn (gen_rtx_SET (VOIDmode
, cond
, x
));
4528 x
= gen_rtx_EQ (DImode
, cond
, const0_rtx
);
4530 emit_unlikely_jump (x
, label2
);
4532 emit_move_insn (scratch
, newval
);
4533 emit_store_conditional (mode
, cond
, mem
, scratch
);
4535 x
= gen_rtx_EQ (DImode
, cond
, const0_rtx
);
4536 emit_unlikely_jump (x
, label1
);
4538 emit_insn (gen_memory_barrier ());
4539 emit_label (XEXP (label2
, 0));
4543 alpha_expand_compare_and_swap_12 (rtx dst
, rtx mem
, rtx oldval
, rtx newval
)
4545 enum machine_mode mode
= GET_MODE (mem
);
4546 rtx addr
, align
, wdst
;
4547 rtx (*fn5
) (rtx
, rtx
, rtx
, rtx
, rtx
);
4549 addr
= force_reg (DImode
, XEXP (mem
, 0));
4550 align
= expand_simple_binop (Pmode
, AND
, addr
, GEN_INT (-8),
4551 NULL_RTX
, 1, OPTAB_DIRECT
);
4553 oldval
= convert_modes (DImode
, mode
, oldval
, 1);
4554 newval
= emit_insxl (mode
, newval
, addr
);
4556 wdst
= gen_reg_rtx (DImode
);
4558 fn5
= gen_sync_compare_and_swapqi_1
;
4560 fn5
= gen_sync_compare_and_swaphi_1
;
4561 emit_insn (fn5 (wdst
, addr
, oldval
, newval
, align
));
4563 emit_move_insn (dst
, gen_lowpart (mode
, wdst
));
4567 alpha_split_compare_and_swap_12 (enum machine_mode mode
, rtx dest
, rtx addr
,
4568 rtx oldval
, rtx newval
, rtx align
,
4569 rtx scratch
, rtx cond
)
4571 rtx label1
, label2
, mem
, width
, mask
, x
;
4573 mem
= gen_rtx_MEM (DImode
, align
);
4574 MEM_VOLATILE_P (mem
) = 1;
4576 emit_insn (gen_memory_barrier ());
4577 label1
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4578 label2
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4579 emit_label (XEXP (label1
, 0));
4581 emit_load_locked (DImode
, scratch
, mem
);
4583 width
= GEN_INT (GET_MODE_BITSIZE (mode
));
4584 mask
= GEN_INT (mode
== QImode
? 0xff : 0xffff);
4585 if (WORDS_BIG_ENDIAN
)
4586 emit_insn (gen_extxl_be (dest
, scratch
, width
, addr
));
4588 emit_insn (gen_extxl_le (dest
, scratch
, width
, addr
));
4590 if (oldval
== const0_rtx
)
4591 x
= gen_rtx_NE (DImode
, dest
, const0_rtx
);
4594 x
= gen_rtx_EQ (DImode
, dest
, oldval
);
4595 emit_insn (gen_rtx_SET (VOIDmode
, cond
, x
));
4596 x
= gen_rtx_EQ (DImode
, cond
, const0_rtx
);
4598 emit_unlikely_jump (x
, label2
);
4600 if (WORDS_BIG_ENDIAN
)
4601 emit_insn (gen_mskxl_be (scratch
, scratch
, mask
, addr
));
4603 emit_insn (gen_mskxl_le (scratch
, scratch
, mask
, addr
));
4604 emit_insn (gen_iordi3 (scratch
, scratch
, newval
));
4606 emit_store_conditional (DImode
, scratch
, mem
, scratch
);
4608 x
= gen_rtx_EQ (DImode
, scratch
, const0_rtx
);
4609 emit_unlikely_jump (x
, label1
);
4611 emit_insn (gen_memory_barrier ());
4612 emit_label (XEXP (label2
, 0));
4615 /* Expand an atomic exchange operation. */
4618 alpha_split_lock_test_and_set (rtx retval
, rtx mem
, rtx val
, rtx scratch
)
4620 enum machine_mode mode
= GET_MODE (mem
);
4621 rtx label
, x
, cond
= gen_lowpart (DImode
, scratch
);
4623 label
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4624 emit_label (XEXP (label
, 0));
4626 emit_load_locked (mode
, retval
, mem
);
4627 emit_move_insn (scratch
, val
);
4628 emit_store_conditional (mode
, cond
, mem
, scratch
);
4630 x
= gen_rtx_EQ (DImode
, cond
, const0_rtx
);
4631 emit_unlikely_jump (x
, label
);
4633 emit_insn (gen_memory_barrier ());
4637 alpha_expand_lock_test_and_set_12 (rtx dst
, rtx mem
, rtx val
)
4639 enum machine_mode mode
= GET_MODE (mem
);
4640 rtx addr
, align
, wdst
;
4641 rtx (*fn4
) (rtx
, rtx
, rtx
, rtx
);
4643 /* Force the address into a register. */
4644 addr
= force_reg (DImode
, XEXP (mem
, 0));
4646 /* Align it to a multiple of 8. */
4647 align
= expand_simple_binop (Pmode
, AND
, addr
, GEN_INT (-8),
4648 NULL_RTX
, 1, OPTAB_DIRECT
);
4650 /* Insert val into the correct byte location within the word. */
4651 val
= emit_insxl (mode
, val
, addr
);
4653 wdst
= gen_reg_rtx (DImode
);
4655 fn4
= gen_sync_lock_test_and_setqi_1
;
4657 fn4
= gen_sync_lock_test_and_sethi_1
;
4658 emit_insn (fn4 (wdst
, addr
, val
, align
));
4660 emit_move_insn (dst
, gen_lowpart (mode
, wdst
));
4664 alpha_split_lock_test_and_set_12 (enum machine_mode mode
, rtx dest
, rtx addr
,
4665 rtx val
, rtx align
, rtx scratch
)
4667 rtx label
, mem
, width
, mask
, x
;
4669 mem
= gen_rtx_MEM (DImode
, align
);
4670 MEM_VOLATILE_P (mem
) = 1;
4672 label
= gen_rtx_LABEL_REF (DImode
, gen_label_rtx ());
4673 emit_label (XEXP (label
, 0));
4675 emit_load_locked (DImode
, scratch
, mem
);
4677 width
= GEN_INT (GET_MODE_BITSIZE (mode
));
4678 mask
= GEN_INT (mode
== QImode
? 0xff : 0xffff);
4679 if (WORDS_BIG_ENDIAN
)
4681 emit_insn (gen_extxl_be (dest
, scratch
, width
, addr
));
4682 emit_insn (gen_mskxl_be (scratch
, scratch
, mask
, addr
));
4686 emit_insn (gen_extxl_le (dest
, scratch
, width
, addr
));
4687 emit_insn (gen_mskxl_le (scratch
, scratch
, mask
, addr
));
4689 emit_insn (gen_iordi3 (scratch
, scratch
, val
));
4691 emit_store_conditional (DImode
, scratch
, mem
, scratch
);
4693 x
= gen_rtx_EQ (DImode
, scratch
, const0_rtx
);
4694 emit_unlikely_jump (x
, label
);
4696 emit_insn (gen_memory_barrier ());
4699 /* Adjust the cost of a scheduling dependency. Return the new cost of
4700 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
4703 alpha_adjust_cost (rtx insn
, rtx link
, rtx dep_insn
, int cost
)
4705 enum attr_type insn_type
, dep_insn_type
;
4707 /* If the dependence is an anti-dependence, there is no cost. For an
4708 output dependence, there is sometimes a cost, but it doesn't seem
4709 worth handling those few cases. */
4710 if (REG_NOTE_KIND (link
) != 0)
4713 /* If we can't recognize the insns, we can't really do anything. */
4714 if (recog_memoized (insn
) < 0 || recog_memoized (dep_insn
) < 0)
4717 insn_type
= get_attr_type (insn
);
4718 dep_insn_type
= get_attr_type (dep_insn
);
4720 /* Bring in the user-defined memory latency. */
4721 if (dep_insn_type
== TYPE_ILD
4722 || dep_insn_type
== TYPE_FLD
4723 || dep_insn_type
== TYPE_LDSYM
)
4724 cost
+= alpha_memory_latency
-1;
4726 /* Everything else handled in DFA bypasses now. */
4731 /* The number of instructions that can be issued per cycle. */
4734 alpha_issue_rate (void)
4736 return (alpha_tune
== PROCESSOR_EV4
? 2 : 4);
4739 /* How many alternative schedules to try. This should be as wide as the
4740 scheduling freedom in the DFA, but no wider. Making this value too
4741 large results extra work for the scheduler.
4743 For EV4, loads can be issued to either IB0 or IB1, thus we have 2
4744 alternative schedules. For EV5, we can choose between E0/E1 and
4745 FA/FM. For EV6, an arithmetic insn can be issued to U0/U1/L0/L1. */
4748 alpha_multipass_dfa_lookahead (void)
4750 return (alpha_tune
== PROCESSOR_EV6
? 4 : 2);
4753 /* Machine-specific function data. */
4755 struct GTY(()) machine_function
4758 /* List of call information words for calls from this function. */
4759 struct rtx_def
*first_ciw
;
4760 struct rtx_def
*last_ciw
;
4763 /* List of deferred case vectors. */
4764 struct rtx_def
*addr_list
;
4767 const char *some_ld_name
;
4769 /* For TARGET_LD_BUGGY_LDGP. */
4770 struct rtx_def
*gp_save_rtx
;
4773 /* How to allocate a 'struct machine_function'. */
4775 static struct machine_function
*
4776 alpha_init_machine_status (void)
4778 return ((struct machine_function
*)
4779 ggc_alloc_cleared (sizeof (struct machine_function
)));
4782 /* Functions to save and restore alpha_return_addr_rtx. */
4784 /* Start the ball rolling with RETURN_ADDR_RTX. */
4787 alpha_return_addr (int count
, rtx frame ATTRIBUTE_UNUSED
)
4792 return get_hard_reg_initial_val (Pmode
, REG_RA
);
4795 /* Return or create a memory slot containing the gp value for the current
4796 function. Needed only if TARGET_LD_BUGGY_LDGP. */
4799 alpha_gp_save_rtx (void)
4801 rtx seq
, m
= cfun
->machine
->gp_save_rtx
;
4807 m
= assign_stack_local (DImode
, UNITS_PER_WORD
, BITS_PER_WORD
);
4808 m
= validize_mem (m
);
4809 emit_move_insn (m
, pic_offset_table_rtx
);
4814 /* We used to simply emit the sequence after entry_of_function.
4815 However this breaks the CFG if the first instruction in the
4816 first block is not the NOTE_INSN_BASIC_BLOCK, for example a
4817 label. Emit the sequence properly on the edge. We are only
4818 invoked from dw2_build_landing_pads and finish_eh_generation
4819 will call commit_edge_insertions thanks to a kludge. */
4820 insert_insn_on_edge (seq
, single_succ_edge (ENTRY_BLOCK_PTR
));
4822 cfun
->machine
->gp_save_rtx
= m
;
4829 alpha_ra_ever_killed (void)
4833 if (!has_hard_reg_initial_val (Pmode
, REG_RA
))
4834 return (int)df_regs_ever_live_p (REG_RA
);
4836 push_topmost_sequence ();
4838 pop_topmost_sequence ();
4840 return reg_set_between_p (gen_rtx_REG (Pmode
, REG_RA
), top
, NULL_RTX
);
4844 /* Return the trap mode suffix applicable to the current
4845 instruction, or NULL. */
4848 get_trap_mode_suffix (void)
4850 enum attr_trap_suffix s
= get_attr_trap_suffix (current_output_insn
);
4854 case TRAP_SUFFIX_NONE
:
4857 case TRAP_SUFFIX_SU
:
4858 if (alpha_fptm
>= ALPHA_FPTM_SU
)
4862 case TRAP_SUFFIX_SUI
:
4863 if (alpha_fptm
>= ALPHA_FPTM_SUI
)
4867 case TRAP_SUFFIX_V_SV
:
4875 case ALPHA_FPTM_SUI
:
4881 case TRAP_SUFFIX_V_SV_SVI
:
4890 case ALPHA_FPTM_SUI
:
4897 case TRAP_SUFFIX_U_SU_SUI
:
4906 case ALPHA_FPTM_SUI
:
4919 /* Return the rounding mode suffix applicable to the current
4920 instruction, or NULL. */
4923 get_round_mode_suffix (void)
4925 enum attr_round_suffix s
= get_attr_round_suffix (current_output_insn
);
4929 case ROUND_SUFFIX_NONE
:
4931 case ROUND_SUFFIX_NORMAL
:
4934 case ALPHA_FPRM_NORM
:
4936 case ALPHA_FPRM_MINF
:
4938 case ALPHA_FPRM_CHOP
:
4940 case ALPHA_FPRM_DYN
:
4947 case ROUND_SUFFIX_C
:
4956 /* Locate some local-dynamic symbol still in use by this function
4957 so that we can print its name in some movdi_er_tlsldm pattern. */
4960 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
4964 if (GET_CODE (x
) == SYMBOL_REF
4965 && SYMBOL_REF_TLS_MODEL (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
4967 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
4975 get_some_local_dynamic_name (void)
4979 if (cfun
->machine
->some_ld_name
)
4980 return cfun
->machine
->some_ld_name
;
4982 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
4984 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
4985 return cfun
->machine
->some_ld_name
;
4990 /* Print an operand. Recognize special options, documented below. */
4993 print_operand (FILE *file
, rtx x
, int code
)
5000 /* Print the assembler name of the current function. */
5001 assemble_name (file
, alpha_fnname
);
5005 assemble_name (file
, get_some_local_dynamic_name ());
5010 const char *trap
= get_trap_mode_suffix ();
5011 const char *round
= get_round_mode_suffix ();
5014 fprintf (file
, (TARGET_AS_SLASH_BEFORE_SUFFIX
? "/%s%s" : "%s%s"),
5015 (trap
? trap
: ""), (round
? round
: ""));
5020 /* Generates single precision instruction suffix. */
5021 fputc ((TARGET_FLOAT_VAX
? 'f' : 's'), file
);
5025 /* Generates double precision instruction suffix. */
5026 fputc ((TARGET_FLOAT_VAX
? 'g' : 't'), file
);
5030 if (alpha_this_literal_sequence_number
== 0)
5031 alpha_this_literal_sequence_number
= alpha_next_sequence_number
++;
5032 fprintf (file
, "%d", alpha_this_literal_sequence_number
);
5036 if (alpha_this_gpdisp_sequence_number
== 0)
5037 alpha_this_gpdisp_sequence_number
= alpha_next_sequence_number
++;
5038 fprintf (file
, "%d", alpha_this_gpdisp_sequence_number
);
5042 if (GET_CODE (x
) == HIGH
)
5043 output_addr_const (file
, XEXP (x
, 0));
5045 output_operand_lossage ("invalid %%H value");
5052 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD_CALL
)
5054 x
= XVECEXP (x
, 0, 0);
5055 lituse
= "lituse_tlsgd";
5057 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM_CALL
)
5059 x
= XVECEXP (x
, 0, 0);
5060 lituse
= "lituse_tlsldm";
5062 else if (CONST_INT_P (x
))
5063 lituse
= "lituse_jsr";
5066 output_operand_lossage ("invalid %%J value");
5070 if (x
!= const0_rtx
)
5071 fprintf (file
, "\t\t!%s!%d", lituse
, (int) INTVAL (x
));
5079 #ifdef HAVE_AS_JSRDIRECT_RELOCS
5080 lituse
= "lituse_jsrdirect";
5082 lituse
= "lituse_jsr";
5085 gcc_assert (INTVAL (x
) != 0);
5086 fprintf (file
, "\t\t!%s!%d", lituse
, (int) INTVAL (x
));
5090 /* If this operand is the constant zero, write it as "$31". */
5092 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5093 else if (x
== CONST0_RTX (GET_MODE (x
)))
5094 fprintf (file
, "$31");
5096 output_operand_lossage ("invalid %%r value");
5100 /* Similar, but for floating-point. */
5102 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5103 else if (x
== CONST0_RTX (GET_MODE (x
)))
5104 fprintf (file
, "$f31");
5106 output_operand_lossage ("invalid %%R value");
5110 /* Write the 1's complement of a constant. */
5111 if (!CONST_INT_P (x
))
5112 output_operand_lossage ("invalid %%N value");
5114 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ~ INTVAL (x
));
5118 /* Write 1 << C, for a constant C. */
5119 if (!CONST_INT_P (x
))
5120 output_operand_lossage ("invalid %%P value");
5122 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, (HOST_WIDE_INT
) 1 << INTVAL (x
));
5126 /* Write the high-order 16 bits of a constant, sign-extended. */
5127 if (!CONST_INT_P (x
))
5128 output_operand_lossage ("invalid %%h value");
5130 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) >> 16);
5134 /* Write the low-order 16 bits of a constant, sign-extended. */
5135 if (!CONST_INT_P (x
))
5136 output_operand_lossage ("invalid %%L value");
5138 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5139 (INTVAL (x
) & 0xffff) - 2 * (INTVAL (x
) & 0x8000));
5143 /* Write mask for ZAP insn. */
5144 if (GET_CODE (x
) == CONST_DOUBLE
)
5146 HOST_WIDE_INT mask
= 0;
5147 HOST_WIDE_INT value
;
5149 value
= CONST_DOUBLE_LOW (x
);
5150 for (i
= 0; i
< HOST_BITS_PER_WIDE_INT
/ HOST_BITS_PER_CHAR
;
5155 value
= CONST_DOUBLE_HIGH (x
);
5156 for (i
= 0; i
< HOST_BITS_PER_WIDE_INT
/ HOST_BITS_PER_CHAR
;
5159 mask
|= (1 << (i
+ sizeof (int)));
5161 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, mask
& 0xff);
5164 else if (CONST_INT_P (x
))
5166 HOST_WIDE_INT mask
= 0, value
= INTVAL (x
);
5168 for (i
= 0; i
< 8; i
++, value
>>= 8)
5172 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, mask
);
5175 output_operand_lossage ("invalid %%m value");
5179 /* 'b', 'w', 'l', or 'q' as the value of the constant. */
5180 if (!CONST_INT_P (x
)
5181 || (INTVAL (x
) != 8 && INTVAL (x
) != 16
5182 && INTVAL (x
) != 32 && INTVAL (x
) != 64))
5183 output_operand_lossage ("invalid %%M value");
5185 fprintf (file
, "%s",
5186 (INTVAL (x
) == 8 ? "b"
5187 : INTVAL (x
) == 16 ? "w"
5188 : INTVAL (x
) == 32 ? "l"
5193 /* Similar, except do it from the mask. */
5194 if (CONST_INT_P (x
))
5196 HOST_WIDE_INT value
= INTVAL (x
);
5203 if (value
== 0xffff)
5208 if (value
== 0xffffffff)
5219 else if (HOST_BITS_PER_WIDE_INT
== 32
5220 && GET_CODE (x
) == CONST_DOUBLE
5221 && CONST_DOUBLE_LOW (x
) == 0xffffffff
5222 && CONST_DOUBLE_HIGH (x
) == 0)
5227 output_operand_lossage ("invalid %%U value");
5231 /* Write the constant value divided by 8 for little-endian mode or
5232 (56 - value) / 8 for big-endian mode. */
5234 if (!CONST_INT_P (x
)
5235 || (unsigned HOST_WIDE_INT
) INTVAL (x
) >= (WORDS_BIG_ENDIAN
5238 || (INTVAL (x
) & 7) != 0)
5239 output_operand_lossage ("invalid %%s value");
5241 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5243 ? (56 - INTVAL (x
)) / 8
5248 /* Same, except compute (64 - c) / 8 */
5250 if (!CONST_INT_P (x
)
5251 && (unsigned HOST_WIDE_INT
) INTVAL (x
) >= 64
5252 && (INTVAL (x
) & 7) != 8)
5253 output_operand_lossage ("invalid %%s value");
5255 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, (64 - INTVAL (x
)) / 8);
5260 /* On Unicos/Mk systems: use a DEX expression if the symbol
5261 clashes with a register name. */
5262 int dex
= unicosmk_need_dex (x
);
5264 fprintf (file
, "DEX(%d)", dex
);
5266 output_addr_const (file
, x
);
5270 case 'C': case 'D': case 'c': case 'd':
5271 /* Write out comparison name. */
5273 enum rtx_code c
= GET_CODE (x
);
5275 if (!COMPARISON_P (x
))
5276 output_operand_lossage ("invalid %%C value");
5278 else if (code
== 'D')
5279 c
= reverse_condition (c
);
5280 else if (code
== 'c')
5281 c
= swap_condition (c
);
5282 else if (code
== 'd')
5283 c
= swap_condition (reverse_condition (c
));
5286 fprintf (file
, "ule");
5288 fprintf (file
, "ult");
5289 else if (c
== UNORDERED
)
5290 fprintf (file
, "un");
5292 fprintf (file
, "%s", GET_RTX_NAME (c
));
5297 /* Write the divide or modulus operator. */
5298 switch (GET_CODE (x
))
5301 fprintf (file
, "div%s", GET_MODE (x
) == SImode
? "l" : "q");
5304 fprintf (file
, "div%su", GET_MODE (x
) == SImode
? "l" : "q");
5307 fprintf (file
, "rem%s", GET_MODE (x
) == SImode
? "l" : "q");
5310 fprintf (file
, "rem%su", GET_MODE (x
) == SImode
? "l" : "q");
5313 output_operand_lossage ("invalid %%E value");
5319 /* Write "_u" for unaligned access. */
5320 if (MEM_P (x
) && GET_CODE (XEXP (x
, 0)) == AND
)
5321 fprintf (file
, "_u");
5326 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5328 output_address (XEXP (x
, 0));
5329 else if (GET_CODE (x
) == CONST
&& GET_CODE (XEXP (x
, 0)) == UNSPEC
)
5331 switch (XINT (XEXP (x
, 0), 1))
5335 output_addr_const (file
, XVECEXP (XEXP (x
, 0), 0, 0));
5338 output_operand_lossage ("unknown relocation unspec");
5343 output_addr_const (file
, x
);
5347 output_operand_lossage ("invalid %%xn code");
5352 print_operand_address (FILE *file
, rtx addr
)
5355 HOST_WIDE_INT offset
= 0;
5357 if (GET_CODE (addr
) == AND
)
5358 addr
= XEXP (addr
, 0);
5360 if (GET_CODE (addr
) == PLUS
5361 && CONST_INT_P (XEXP (addr
, 1)))
5363 offset
= INTVAL (XEXP (addr
, 1));
5364 addr
= XEXP (addr
, 0);
5367 if (GET_CODE (addr
) == LO_SUM
)
5369 const char *reloc16
, *reloclo
;
5370 rtx op1
= XEXP (addr
, 1);
5372 if (GET_CODE (op1
) == CONST
&& GET_CODE (XEXP (op1
, 0)) == UNSPEC
)
5374 op1
= XEXP (op1
, 0);
5375 switch (XINT (op1
, 1))
5379 reloclo
= (alpha_tls_size
== 16 ? "dtprel" : "dtprello");
5383 reloclo
= (alpha_tls_size
== 16 ? "tprel" : "tprello");
5386 output_operand_lossage ("unknown relocation unspec");
5390 output_addr_const (file
, XVECEXP (op1
, 0, 0));
5395 reloclo
= "gprellow";
5396 output_addr_const (file
, op1
);
5400 fprintf (file
, "+" HOST_WIDE_INT_PRINT_DEC
, offset
);
5402 addr
= XEXP (addr
, 0);
5403 switch (GET_CODE (addr
))
5406 basereg
= REGNO (addr
);
5410 basereg
= subreg_regno (addr
);
5417 fprintf (file
, "($%d)\t\t!%s", basereg
,
5418 (basereg
== 29 ? reloc16
: reloclo
));
5422 switch (GET_CODE (addr
))
5425 basereg
= REGNO (addr
);
5429 basereg
= subreg_regno (addr
);
5433 offset
= INTVAL (addr
);
5436 #if TARGET_ABI_OPEN_VMS
5438 fprintf (file
, "%s", XSTR (addr
, 0));
5442 gcc_assert (GET_CODE (XEXP (addr
, 0)) == PLUS
5443 && GET_CODE (XEXP (XEXP (addr
, 0), 0)) == SYMBOL_REF
);
5444 fprintf (file
, "%s+" HOST_WIDE_INT_PRINT_DEC
,
5445 XSTR (XEXP (XEXP (addr
, 0), 0), 0),
5446 INTVAL (XEXP (XEXP (addr
, 0), 1)));
5454 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
"($%d)", offset
, basereg
);
5457 /* Emit RTL insns to initialize the variable parts of a trampoline at
5458 TRAMP. FNADDR is an RTX for the address of the function's pure
5459 code. CXT is an RTX for the static chain value for the function.
5461 The three offset parameters are for the individual template's
5462 layout. A JMPOFS < 0 indicates that the trampoline does not
5463 contain instructions at all.
5465 We assume here that a function will be called many more times than
5466 its address is taken (e.g., it might be passed to qsort), so we
5467 take the trouble to initialize the "hint" field in the JMP insn.
5468 Note that the hint field is PC (new) + 4 * bits 13:0. */
5471 alpha_initialize_trampoline (rtx tramp
, rtx fnaddr
, rtx cxt
,
5472 int fnofs
, int cxtofs
, int jmpofs
)
5475 /* VMS really uses DImode pointers in memory at this point. */
5476 enum machine_mode mode
= TARGET_ABI_OPEN_VMS
? Pmode
: ptr_mode
;
5478 #ifdef POINTERS_EXTEND_UNSIGNED
5479 fnaddr
= convert_memory_address (mode
, fnaddr
);
5480 cxt
= convert_memory_address (mode
, cxt
);
5483 /* Store function address and CXT. */
5484 addr
= memory_address (mode
, plus_constant (tramp
, fnofs
));
5485 emit_move_insn (gen_rtx_MEM (mode
, addr
), fnaddr
);
5486 addr
= memory_address (mode
, plus_constant (tramp
, cxtofs
));
5487 emit_move_insn (gen_rtx_MEM (mode
, addr
), cxt
);
5489 #ifdef ENABLE_EXECUTE_STACK
5490 emit_library_call (init_one_libfunc ("__enable_execute_stack"),
5491 LCT_NORMAL
, VOIDmode
, 1, tramp
, Pmode
);
5495 emit_insn (gen_imb ());
5498 /* Determine where to put an argument to a function.
5499 Value is zero to push the argument on the stack,
5500 or a hard register in which to store the argument.
5502 MODE is the argument's machine mode.
5503 TYPE is the data type of the argument (as a tree).
5504 This is null for libcalls where that information may
5506 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5507 the preceding args and about the function being called.
5508 NAMED is nonzero if this argument is a named parameter
5509 (otherwise it is an extra parameter matching an ellipsis).
5511 On Alpha the first 6 words of args are normally in registers
5512 and the rest are pushed. */
5515 function_arg (CUMULATIVE_ARGS cum
, enum machine_mode mode
, tree type
,
5516 int named ATTRIBUTE_UNUSED
)
5521 /* Don't get confused and pass small structures in FP registers. */
5522 if (type
&& AGGREGATE_TYPE_P (type
))
5526 #ifdef ENABLE_CHECKING
5527 /* With alpha_split_complex_arg, we shouldn't see any raw complex
5529 gcc_assert (!COMPLEX_MODE_P (mode
));
5532 /* Set up defaults for FP operands passed in FP registers, and
5533 integral operands passed in integer registers. */
5534 if (TARGET_FPREGS
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
)
5540 /* ??? Irritatingly, the definition of CUMULATIVE_ARGS is different for
5541 the three platforms, so we can't avoid conditional compilation. */
5542 #if TARGET_ABI_OPEN_VMS
5544 if (mode
== VOIDmode
)
5545 return alpha_arg_info_reg_val (cum
);
5547 num_args
= cum
.num_args
;
5549 || targetm
.calls
.must_pass_in_stack (mode
, type
))
5552 #elif TARGET_ABI_UNICOSMK
5556 /* If this is the last argument, generate the call info word (CIW). */
5557 /* ??? We don't include the caller's line number in the CIW because
5558 I don't know how to determine it if debug infos are turned off. */
5559 if (mode
== VOIDmode
)
5568 for (i
= 0; i
< cum
.num_reg_words
&& i
< 5; i
++)
5569 if (cum
.reg_args_type
[i
])
5570 lo
|= (1 << (7 - i
));
5572 if (cum
.num_reg_words
== 6 && cum
.reg_args_type
[5])
5575 lo
|= cum
.num_reg_words
;
5577 #if HOST_BITS_PER_WIDE_INT == 32
5578 hi
= (cum
.num_args
<< 20) | cum
.num_arg_words
;
5580 lo
= lo
| ((HOST_WIDE_INT
) cum
.num_args
<< 52)
5581 | ((HOST_WIDE_INT
) cum
.num_arg_words
<< 32);
5584 ciw
= immed_double_const (lo
, hi
, DImode
);
5586 return gen_rtx_UNSPEC (DImode
, gen_rtvec (1, ciw
),
5587 UNSPEC_UMK_LOAD_CIW
);
5590 size
= ALPHA_ARG_SIZE (mode
, type
, named
);
5591 num_args
= cum
.num_reg_words
;
5593 || cum
.num_reg_words
+ size
> 6
5594 || targetm
.calls
.must_pass_in_stack (mode
, type
))
5596 else if (type
&& TYPE_MODE (type
) == BLKmode
)
5600 reg1
= gen_rtx_REG (DImode
, num_args
+ 16);
5601 reg1
= gen_rtx_EXPR_LIST (DImode
, reg1
, const0_rtx
);
5603 /* The argument fits in two registers. Note that we still need to
5604 reserve a register for empty structures. */
5608 return gen_rtx_PARALLEL (mode
, gen_rtvec (1, reg1
));
5611 reg2
= gen_rtx_REG (DImode
, num_args
+ 17);
5612 reg2
= gen_rtx_EXPR_LIST (DImode
, reg2
, GEN_INT (8));
5613 return gen_rtx_PARALLEL (mode
, gen_rtvec (2, reg1
, reg2
));
5617 #elif TARGET_ABI_OSF
5623 /* VOID is passed as a special flag for "last argument". */
5624 if (type
== void_type_node
)
5626 else if (targetm
.calls
.must_pass_in_stack (mode
, type
))
5630 #error Unhandled ABI
5633 return gen_rtx_REG (mode
, num_args
+ basereg
);
5637 alpha_arg_partial_bytes (CUMULATIVE_ARGS
*cum ATTRIBUTE_UNUSED
,
5638 enum machine_mode mode ATTRIBUTE_UNUSED
,
5639 tree type ATTRIBUTE_UNUSED
,
5640 bool named ATTRIBUTE_UNUSED
)
5644 #if TARGET_ABI_OPEN_VMS
5645 if (cum
->num_args
< 6
5646 && 6 < cum
->num_args
+ ALPHA_ARG_SIZE (mode
, type
, named
))
5647 words
= 6 - cum
->num_args
;
5648 #elif TARGET_ABI_UNICOSMK
5649 /* Never any split arguments. */
5650 #elif TARGET_ABI_OSF
5651 if (*cum
< 6 && 6 < *cum
+ ALPHA_ARG_SIZE (mode
, type
, named
))
5654 #error Unhandled ABI
5657 return words
* UNITS_PER_WORD
;
5661 /* Return true if TYPE must be returned in memory, instead of in registers. */
5664 alpha_return_in_memory (const_tree type
, const_tree fndecl ATTRIBUTE_UNUSED
)
5666 enum machine_mode mode
= VOIDmode
;
5671 mode
= TYPE_MODE (type
);
5673 /* All aggregates are returned in memory. */
5674 if (AGGREGATE_TYPE_P (type
))
5678 size
= GET_MODE_SIZE (mode
);
5679 switch (GET_MODE_CLASS (mode
))
5681 case MODE_VECTOR_FLOAT
:
5682 /* Pass all float vectors in memory, like an aggregate. */
5685 case MODE_COMPLEX_FLOAT
:
5686 /* We judge complex floats on the size of their element,
5687 not the size of the whole type. */
5688 size
= GET_MODE_UNIT_SIZE (mode
);
5693 case MODE_COMPLEX_INT
:
5694 case MODE_VECTOR_INT
:
5698 /* ??? We get called on all sorts of random stuff from
5699 aggregate_value_p. We must return something, but it's not
5700 clear what's safe to return. Pretend it's a struct I
5705 /* Otherwise types must fit in one register. */
5706 return size
> UNITS_PER_WORD
;
5709 /* Return true if TYPE should be passed by invisible reference. */
5712 alpha_pass_by_reference (CUMULATIVE_ARGS
*ca ATTRIBUTE_UNUSED
,
5713 enum machine_mode mode
,
5714 const_tree type ATTRIBUTE_UNUSED
,
5715 bool named ATTRIBUTE_UNUSED
)
5717 return mode
== TFmode
|| mode
== TCmode
;
5720 /* Define how to find the value returned by a function. VALTYPE is the
5721 data type of the value (as a tree). If the precise function being
5722 called is known, FUNC is its FUNCTION_DECL; otherwise, FUNC is 0.
5723 MODE is set instead of VALTYPE for libcalls.
5725 On Alpha the value is found in $0 for integer functions and
5726 $f0 for floating-point functions. */
5729 function_value (const_tree valtype
, const_tree func ATTRIBUTE_UNUSED
,
5730 enum machine_mode mode
)
5732 unsigned int regnum
, dummy
;
5733 enum mode_class mclass
;
5735 gcc_assert (!valtype
|| !alpha_return_in_memory (valtype
, func
));
5738 mode
= TYPE_MODE (valtype
);
5740 mclass
= GET_MODE_CLASS (mode
);
5744 PROMOTE_MODE (mode
, dummy
, valtype
);
5747 case MODE_COMPLEX_INT
:
5748 case MODE_VECTOR_INT
:
5756 case MODE_COMPLEX_FLOAT
:
5758 enum machine_mode cmode
= GET_MODE_INNER (mode
);
5760 return gen_rtx_PARALLEL
5763 gen_rtx_EXPR_LIST (VOIDmode
, gen_rtx_REG (cmode
, 32),
5765 gen_rtx_EXPR_LIST (VOIDmode
, gen_rtx_REG (cmode
, 33),
5766 GEN_INT (GET_MODE_SIZE (cmode
)))));
5773 return gen_rtx_REG (mode
, regnum
);
5776 /* TCmode complex values are passed by invisible reference. We
5777 should not split these values. */
5780 alpha_split_complex_arg (const_tree type
)
5782 return TYPE_MODE (type
) != TCmode
;
5786 alpha_build_builtin_va_list (void)
5788 tree base
, ofs
, space
, record
, type_decl
;
5790 if (TARGET_ABI_OPEN_VMS
|| TARGET_ABI_UNICOSMK
)
5791 return ptr_type_node
;
5793 record
= (*lang_hooks
.types
.make_type
) (RECORD_TYPE
);
5794 type_decl
= build_decl (TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
5795 TREE_CHAIN (record
) = type_decl
;
5796 TYPE_NAME (record
) = type_decl
;
5798 /* C++? SET_IS_AGGR_TYPE (record, 1); */
5800 /* Dummy field to prevent alignment warnings. */
5801 space
= build_decl (FIELD_DECL
, NULL_TREE
, integer_type_node
);
5802 DECL_FIELD_CONTEXT (space
) = record
;
5803 DECL_ARTIFICIAL (space
) = 1;
5804 DECL_IGNORED_P (space
) = 1;
5806 ofs
= build_decl (FIELD_DECL
, get_identifier ("__offset"),
5808 DECL_FIELD_CONTEXT (ofs
) = record
;
5809 TREE_CHAIN (ofs
) = space
;
5811 base
= build_decl (FIELD_DECL
, get_identifier ("__base"),
5813 DECL_FIELD_CONTEXT (base
) = record
;
5814 TREE_CHAIN (base
) = ofs
;
5816 TYPE_FIELDS (record
) = base
;
5817 layout_type (record
);
5819 va_list_gpr_counter_field
= ofs
;
5824 /* Helper function for alpha_stdarg_optimize_hook. Skip over casts
5825 and constant additions. */
5828 va_list_skip_additions (tree lhs
)
5834 enum tree_code code
;
5836 stmt
= SSA_NAME_DEF_STMT (lhs
);
5838 if (gimple_code (stmt
) == GIMPLE_PHI
)
5841 if (!is_gimple_assign (stmt
)
5842 || gimple_assign_lhs (stmt
) != lhs
)
5845 if (TREE_CODE (gimple_assign_rhs1 (stmt
)) != SSA_NAME
)
5847 code
= gimple_assign_rhs_code (stmt
);
5848 if (!CONVERT_EXPR_CODE_P (code
)
5849 && ((code
!= PLUS_EXPR
&& code
!= POINTER_PLUS_EXPR
)
5850 || TREE_CODE (gimple_assign_rhs2 (stmt
)) != INTEGER_CST
5851 || !host_integerp (gimple_assign_rhs2 (stmt
), 1)))
5854 lhs
= gimple_assign_rhs1 (stmt
);
5858 /* Check if LHS = RHS statement is
5859 LHS = *(ap.__base + ap.__offset + cst)
5862 + ((ap.__offset + cst <= 47)
5863 ? ap.__offset + cst - 48 : ap.__offset + cst) + cst2).
5864 If the former, indicate that GPR registers are needed,
5865 if the latter, indicate that FPR registers are needed.
5867 Also look for LHS = (*ptr).field, where ptr is one of the forms
5870 On alpha, cfun->va_list_gpr_size is used as size of the needed
5871 regs and cfun->va_list_fpr_size is a bitmask, bit 0 set if GPR
5872 registers are needed and bit 1 set if FPR registers are needed.
5873 Return true if va_list references should not be scanned for the
5874 current statement. */
5877 alpha_stdarg_optimize_hook (struct stdarg_info
*si
, const_gimple stmt
)
5879 tree base
, offset
, rhs
;
5883 if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
))
5884 != GIMPLE_SINGLE_RHS
)
5887 rhs
= gimple_assign_rhs1 (stmt
);
5888 while (handled_component_p (rhs
))
5889 rhs
= TREE_OPERAND (rhs
, 0);
5890 if (TREE_CODE (rhs
) != INDIRECT_REF
5891 || TREE_CODE (TREE_OPERAND (rhs
, 0)) != SSA_NAME
)
5894 stmt
= va_list_skip_additions (TREE_OPERAND (rhs
, 0));
5896 || !is_gimple_assign (stmt
)
5897 || gimple_assign_rhs_code (stmt
) != POINTER_PLUS_EXPR
)
5900 base
= gimple_assign_rhs1 (stmt
);
5901 if (TREE_CODE (base
) == SSA_NAME
)
5903 base_stmt
= va_list_skip_additions (base
);
5905 && is_gimple_assign (base_stmt
)
5906 && gimple_assign_rhs_code (base_stmt
) == COMPONENT_REF
)
5907 base
= gimple_assign_rhs1 (base_stmt
);
5910 if (TREE_CODE (base
) != COMPONENT_REF
5911 || TREE_OPERAND (base
, 1) != TYPE_FIELDS (va_list_type_node
))
5913 base
= gimple_assign_rhs2 (stmt
);
5914 if (TREE_CODE (base
) == SSA_NAME
)
5916 base_stmt
= va_list_skip_additions (base
);
5918 && is_gimple_assign (base_stmt
)
5919 && gimple_assign_rhs_code (base_stmt
) == COMPONENT_REF
)
5920 base
= gimple_assign_rhs1 (base_stmt
);
5923 if (TREE_CODE (base
) != COMPONENT_REF
5924 || TREE_OPERAND (base
, 1) != TYPE_FIELDS (va_list_type_node
))
5930 base
= get_base_address (base
);
5931 if (TREE_CODE (base
) != VAR_DECL
5932 || !bitmap_bit_p (si
->va_list_vars
, DECL_UID (base
)))
5935 offset
= gimple_op (stmt
, 1 + offset_arg
);
5936 if (TREE_CODE (offset
) == SSA_NAME
)
5938 gimple offset_stmt
= va_list_skip_additions (offset
);
5941 && gimple_code (offset_stmt
) == GIMPLE_PHI
)
5944 gimple arg1_stmt
, arg2_stmt
;
5946 enum tree_code code1
, code2
;
5948 if (gimple_phi_num_args (offset_stmt
) != 2)
5952 = va_list_skip_additions (gimple_phi_arg_def (offset_stmt
, 0));
5954 = va_list_skip_additions (gimple_phi_arg_def (offset_stmt
, 1));
5955 if (arg1_stmt
== NULL
5956 || !is_gimple_assign (arg1_stmt
)
5957 || arg2_stmt
== NULL
5958 || !is_gimple_assign (arg2_stmt
))
5961 code1
= gimple_assign_rhs_code (arg1_stmt
);
5962 code2
= gimple_assign_rhs_code (arg2_stmt
);
5963 if (code1
== COMPONENT_REF
5964 && (code2
== MINUS_EXPR
|| code2
== PLUS_EXPR
))
5966 else if (code2
== COMPONENT_REF
5967 && (code1
== MINUS_EXPR
|| code1
== PLUS_EXPR
))
5969 gimple tem
= arg1_stmt
;
5971 arg1_stmt
= arg2_stmt
;
5977 if (!host_integerp (gimple_assign_rhs2 (arg2_stmt
), 0))
5980 sub
= tree_low_cst (gimple_assign_rhs2 (arg2_stmt
), 0);
5981 if (code2
== MINUS_EXPR
)
5983 if (sub
< -48 || sub
> -32)
5986 arg1
= gimple_assign_rhs1 (arg1_stmt
);
5987 arg2
= gimple_assign_rhs1 (arg2_stmt
);
5988 if (TREE_CODE (arg2
) == SSA_NAME
)
5990 arg2_stmt
= va_list_skip_additions (arg2
);
5991 if (arg2_stmt
== NULL
5992 || !is_gimple_assign (arg2_stmt
)
5993 || gimple_assign_rhs_code (arg2_stmt
) != COMPONENT_REF
)
5995 arg2
= gimple_assign_rhs1 (arg2_stmt
);
6000 if (TREE_CODE (arg1
) != COMPONENT_REF
6001 || TREE_OPERAND (arg1
, 1) != va_list_gpr_counter_field
6002 || get_base_address (arg1
) != base
)
6005 /* Need floating point regs. */
6006 cfun
->va_list_fpr_size
|= 2;
6010 && is_gimple_assign (offset_stmt
)
6011 && gimple_assign_rhs_code (offset_stmt
) == COMPONENT_REF
)
6012 offset
= gimple_assign_rhs1 (offset_stmt
);
6014 if (TREE_CODE (offset
) != COMPONENT_REF
6015 || TREE_OPERAND (offset
, 1) != va_list_gpr_counter_field
6016 || get_base_address (offset
) != base
)
6019 /* Need general regs. */
6020 cfun
->va_list_fpr_size
|= 1;
6024 si
->va_list_escapes
= true;
6029 /* Perform any needed actions needed for a function that is receiving a
6030 variable number of arguments. */
6033 alpha_setup_incoming_varargs (CUMULATIVE_ARGS
*pcum
, enum machine_mode mode
,
6034 tree type
, int *pretend_size
, int no_rtl
)
6036 CUMULATIVE_ARGS cum
= *pcum
;
6038 /* Skip the current argument. */
6039 FUNCTION_ARG_ADVANCE (cum
, mode
, type
, 1);
6041 #if TARGET_ABI_UNICOSMK
6042 /* On Unicos/Mk, the standard subroutine __T3E_MISMATCH stores all register
6043 arguments on the stack. Unfortunately, it doesn't always store the first
6044 one (i.e. the one that arrives in $16 or $f16). This is not a problem
6045 with stdargs as we always have at least one named argument there. */
6046 if (cum
.num_reg_words
< 6)
6050 emit_insn (gen_umk_mismatch_args (GEN_INT (cum
.num_reg_words
)));
6051 emit_insn (gen_arg_home_umk ());
6055 #elif TARGET_ABI_OPEN_VMS
6056 /* For VMS, we allocate space for all 6 arg registers plus a count.
6058 However, if NO registers need to be saved, don't allocate any space.
6059 This is not only because we won't need the space, but because AP
6060 includes the current_pretend_args_size and we don't want to mess up
6061 any ap-relative addresses already made. */
6062 if (cum
.num_args
< 6)
6066 emit_move_insn (gen_rtx_REG (DImode
, 1), virtual_incoming_args_rtx
);
6067 emit_insn (gen_arg_home ());
6069 *pretend_size
= 7 * UNITS_PER_WORD
;
6072 /* On OSF/1 and friends, we allocate space for all 12 arg registers, but
6073 only push those that are remaining. However, if NO registers need to
6074 be saved, don't allocate any space. This is not only because we won't
6075 need the space, but because AP includes the current_pretend_args_size
6076 and we don't want to mess up any ap-relative addresses already made.
6078 If we are not to use the floating-point registers, save the integer
6079 registers where we would put the floating-point registers. This is
6080 not the most efficient way to implement varargs with just one register
6081 class, but it isn't worth doing anything more efficient in this rare
6089 alias_set_type set
= get_varargs_alias_set ();
6092 count
= cfun
->va_list_gpr_size
/ UNITS_PER_WORD
;
6093 if (count
> 6 - cum
)
6096 /* Detect whether integer registers or floating-point registers
6097 are needed by the detected va_arg statements. See above for
6098 how these values are computed. Note that the "escape" value
6099 is VA_LIST_MAX_FPR_SIZE, which is 255, which has both of
6101 gcc_assert ((VA_LIST_MAX_FPR_SIZE
& 3) == 3);
6103 if (cfun
->va_list_fpr_size
& 1)
6105 tmp
= gen_rtx_MEM (BLKmode
,
6106 plus_constant (virtual_incoming_args_rtx
,
6107 (cum
+ 6) * UNITS_PER_WORD
));
6108 MEM_NOTRAP_P (tmp
) = 1;
6109 set_mem_alias_set (tmp
, set
);
6110 move_block_from_reg (16 + cum
, tmp
, count
);
6113 if (cfun
->va_list_fpr_size
& 2)
6115 tmp
= gen_rtx_MEM (BLKmode
,
6116 plus_constant (virtual_incoming_args_rtx
,
6117 cum
* UNITS_PER_WORD
));
6118 MEM_NOTRAP_P (tmp
) = 1;
6119 set_mem_alias_set (tmp
, set
);
6120 move_block_from_reg (16 + cum
+ TARGET_FPREGS
*32, tmp
, count
);
6123 *pretend_size
= 12 * UNITS_PER_WORD
;
6128 alpha_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
6130 HOST_WIDE_INT offset
;
6131 tree t
, offset_field
, base_field
;
6133 if (TREE_CODE (TREE_TYPE (valist
)) == ERROR_MARK
)
6136 if (TARGET_ABI_UNICOSMK
)
6137 std_expand_builtin_va_start (valist
, nextarg
);
6139 /* For Unix, TARGET_SETUP_INCOMING_VARARGS moves the starting address base
6140 up by 48, storing fp arg registers in the first 48 bytes, and the
6141 integer arg registers in the next 48 bytes. This is only done,
6142 however, if any integer registers need to be stored.
6144 If no integer registers need be stored, then we must subtract 48
6145 in order to account for the integer arg registers which are counted
6146 in argsize above, but which are not actually stored on the stack.
6147 Must further be careful here about structures straddling the last
6148 integer argument register; that futzes with pretend_args_size,
6149 which changes the meaning of AP. */
6152 offset
= TARGET_ABI_OPEN_VMS
? UNITS_PER_WORD
: 6 * UNITS_PER_WORD
;
6154 offset
= -6 * UNITS_PER_WORD
+ crtl
->args
.pretend_args_size
;
6156 if (TARGET_ABI_OPEN_VMS
)
6158 nextarg
= plus_constant (nextarg
, offset
);
6159 nextarg
= plus_constant (nextarg
, NUM_ARGS
* UNITS_PER_WORD
);
6160 t
= build2 (MODIFY_EXPR
, TREE_TYPE (valist
), valist
,
6161 make_tree (ptr_type_node
, nextarg
));
6162 TREE_SIDE_EFFECTS (t
) = 1;
6164 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
6168 base_field
= TYPE_FIELDS (TREE_TYPE (valist
));
6169 offset_field
= TREE_CHAIN (base_field
);
6171 base_field
= build3 (COMPONENT_REF
, TREE_TYPE (base_field
),
6172 valist
, base_field
, NULL_TREE
);
6173 offset_field
= build3 (COMPONENT_REF
, TREE_TYPE (offset_field
),
6174 valist
, offset_field
, NULL_TREE
);
6176 t
= make_tree (ptr_type_node
, virtual_incoming_args_rtx
);
6177 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
,
6179 t
= build2 (MODIFY_EXPR
, TREE_TYPE (base_field
), base_field
, t
);
6180 TREE_SIDE_EFFECTS (t
) = 1;
6181 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
6183 t
= build_int_cst (NULL_TREE
, NUM_ARGS
* UNITS_PER_WORD
);
6184 t
= build2 (MODIFY_EXPR
, TREE_TYPE (offset_field
), offset_field
, t
);
6185 TREE_SIDE_EFFECTS (t
) = 1;
6186 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
6191 alpha_gimplify_va_arg_1 (tree type
, tree base
, tree offset
,
6194 tree type_size
, ptr_type
, addend
, t
, addr
;
6195 gimple_seq internal_post
;
6197 /* If the type could not be passed in registers, skip the block
6198 reserved for the registers. */
6199 if (targetm
.calls
.must_pass_in_stack (TYPE_MODE (type
), type
))
6201 t
= build_int_cst (TREE_TYPE (offset
), 6*8);
6202 gimplify_assign (offset
,
6203 build2 (MAX_EXPR
, TREE_TYPE (offset
), offset
, t
),
6208 ptr_type
= build_pointer_type (type
);
6210 if (TREE_CODE (type
) == COMPLEX_TYPE
)
6212 tree real_part
, imag_part
, real_temp
;
6214 real_part
= alpha_gimplify_va_arg_1 (TREE_TYPE (type
), base
,
6217 /* Copy the value into a new temporary, lest the formal temporary
6218 be reused out from under us. */
6219 real_temp
= get_initialized_tmp_var (real_part
, pre_p
, NULL
);
6221 imag_part
= alpha_gimplify_va_arg_1 (TREE_TYPE (type
), base
,
6224 return build2 (COMPLEX_EXPR
, type
, real_temp
, imag_part
);
6226 else if (TREE_CODE (type
) == REAL_TYPE
)
6228 tree fpaddend
, cond
, fourtyeight
;
6230 fourtyeight
= build_int_cst (TREE_TYPE (addend
), 6*8);
6231 fpaddend
= fold_build2 (MINUS_EXPR
, TREE_TYPE (addend
),
6232 addend
, fourtyeight
);
6233 cond
= fold_build2 (LT_EXPR
, boolean_type_node
, addend
, fourtyeight
);
6234 addend
= fold_build3 (COND_EXPR
, TREE_TYPE (addend
), cond
,
6238 /* Build the final address and force that value into a temporary. */
6239 addr
= build2 (POINTER_PLUS_EXPR
, ptr_type
, fold_convert (ptr_type
, base
),
6240 fold_convert (sizetype
, addend
));
6241 internal_post
= NULL
;
6242 gimplify_expr (&addr
, pre_p
, &internal_post
, is_gimple_val
, fb_rvalue
);
6243 gimple_seq_add_seq (pre_p
, internal_post
);
6245 /* Update the offset field. */
6246 type_size
= TYPE_SIZE_UNIT (TYPE_MAIN_VARIANT (type
));
6247 if (type_size
== NULL
|| TREE_OVERFLOW (type_size
))
6251 t
= size_binop (PLUS_EXPR
, type_size
, size_int (7));
6252 t
= size_binop (TRUNC_DIV_EXPR
, t
, size_int (8));
6253 t
= size_binop (MULT_EXPR
, t
, size_int (8));
6255 t
= fold_convert (TREE_TYPE (offset
), t
);
6256 gimplify_assign (offset
, build2 (PLUS_EXPR
, TREE_TYPE (offset
), offset
, t
),
6259 return build_va_arg_indirect_ref (addr
);
6263 alpha_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
6266 tree offset_field
, base_field
, offset
, base
, t
, r
;
6269 if (TARGET_ABI_OPEN_VMS
|| TARGET_ABI_UNICOSMK
)
6270 return std_gimplify_va_arg_expr (valist
, type
, pre_p
, post_p
);
6272 base_field
= TYPE_FIELDS (va_list_type_node
);
6273 offset_field
= TREE_CHAIN (base_field
);
6274 base_field
= build3 (COMPONENT_REF
, TREE_TYPE (base_field
),
6275 valist
, base_field
, NULL_TREE
);
6276 offset_field
= build3 (COMPONENT_REF
, TREE_TYPE (offset_field
),
6277 valist
, offset_field
, NULL_TREE
);
6279 /* Pull the fields of the structure out into temporaries. Since we never
6280 modify the base field, we can use a formal temporary. Sign-extend the
6281 offset field so that it's the proper width for pointer arithmetic. */
6282 base
= get_formal_tmp_var (base_field
, pre_p
);
6284 t
= fold_convert (lang_hooks
.types
.type_for_size (64, 0), offset_field
);
6285 offset
= get_initialized_tmp_var (t
, pre_p
, NULL
);
6287 indirect
= pass_by_reference (NULL
, TYPE_MODE (type
), type
, false);
6289 type
= build_pointer_type (type
);
6291 /* Find the value. Note that this will be a stable indirection, or
6292 a composite of stable indirections in the case of complex. */
6293 r
= alpha_gimplify_va_arg_1 (type
, base
, offset
, pre_p
);
6295 /* Stuff the offset temporary back into its field. */
6296 gimplify_assign (unshare_expr (offset_field
),
6297 fold_convert (TREE_TYPE (offset_field
), offset
), pre_p
);
6300 r
= build_va_arg_indirect_ref (r
);
6309 ALPHA_BUILTIN_CMPBGE
,
6310 ALPHA_BUILTIN_EXTBL
,
6311 ALPHA_BUILTIN_EXTWL
,
6312 ALPHA_BUILTIN_EXTLL
,
6313 ALPHA_BUILTIN_EXTQL
,
6314 ALPHA_BUILTIN_EXTWH
,
6315 ALPHA_BUILTIN_EXTLH
,
6316 ALPHA_BUILTIN_EXTQH
,
6317 ALPHA_BUILTIN_INSBL
,
6318 ALPHA_BUILTIN_INSWL
,
6319 ALPHA_BUILTIN_INSLL
,
6320 ALPHA_BUILTIN_INSQL
,
6321 ALPHA_BUILTIN_INSWH
,
6322 ALPHA_BUILTIN_INSLH
,
6323 ALPHA_BUILTIN_INSQH
,
6324 ALPHA_BUILTIN_MSKBL
,
6325 ALPHA_BUILTIN_MSKWL
,
6326 ALPHA_BUILTIN_MSKLL
,
6327 ALPHA_BUILTIN_MSKQL
,
6328 ALPHA_BUILTIN_MSKWH
,
6329 ALPHA_BUILTIN_MSKLH
,
6330 ALPHA_BUILTIN_MSKQH
,
6331 ALPHA_BUILTIN_UMULH
,
6333 ALPHA_BUILTIN_ZAPNOT
,
6334 ALPHA_BUILTIN_AMASK
,
6335 ALPHA_BUILTIN_IMPLVER
,
6337 ALPHA_BUILTIN_THREAD_POINTER
,
6338 ALPHA_BUILTIN_SET_THREAD_POINTER
,
6341 ALPHA_BUILTIN_MINUB8
,
6342 ALPHA_BUILTIN_MINSB8
,
6343 ALPHA_BUILTIN_MINUW4
,
6344 ALPHA_BUILTIN_MINSW4
,
6345 ALPHA_BUILTIN_MAXUB8
,
6346 ALPHA_BUILTIN_MAXSB8
,
6347 ALPHA_BUILTIN_MAXUW4
,
6348 ALPHA_BUILTIN_MAXSW4
,
6352 ALPHA_BUILTIN_UNPKBL
,
6353 ALPHA_BUILTIN_UNPKBW
,
6358 ALPHA_BUILTIN_CTPOP
,
6363 static enum insn_code
const code_for_builtin
[ALPHA_BUILTIN_max
] = {
6364 CODE_FOR_builtin_cmpbge
,
6365 CODE_FOR_builtin_extbl
,
6366 CODE_FOR_builtin_extwl
,
6367 CODE_FOR_builtin_extll
,
6368 CODE_FOR_builtin_extql
,
6369 CODE_FOR_builtin_extwh
,
6370 CODE_FOR_builtin_extlh
,
6371 CODE_FOR_builtin_extqh
,
6372 CODE_FOR_builtin_insbl
,
6373 CODE_FOR_builtin_inswl
,
6374 CODE_FOR_builtin_insll
,
6375 CODE_FOR_builtin_insql
,
6376 CODE_FOR_builtin_inswh
,
6377 CODE_FOR_builtin_inslh
,
6378 CODE_FOR_builtin_insqh
,
6379 CODE_FOR_builtin_mskbl
,
6380 CODE_FOR_builtin_mskwl
,
6381 CODE_FOR_builtin_mskll
,
6382 CODE_FOR_builtin_mskql
,
6383 CODE_FOR_builtin_mskwh
,
6384 CODE_FOR_builtin_msklh
,
6385 CODE_FOR_builtin_mskqh
,
6386 CODE_FOR_umuldi3_highpart
,
6387 CODE_FOR_builtin_zap
,
6388 CODE_FOR_builtin_zapnot
,
6389 CODE_FOR_builtin_amask
,
6390 CODE_FOR_builtin_implver
,
6391 CODE_FOR_builtin_rpcc
,
6396 CODE_FOR_builtin_minub8
,
6397 CODE_FOR_builtin_minsb8
,
6398 CODE_FOR_builtin_minuw4
,
6399 CODE_FOR_builtin_minsw4
,
6400 CODE_FOR_builtin_maxub8
,
6401 CODE_FOR_builtin_maxsb8
,
6402 CODE_FOR_builtin_maxuw4
,
6403 CODE_FOR_builtin_maxsw4
,
6404 CODE_FOR_builtin_perr
,
6405 CODE_FOR_builtin_pklb
,
6406 CODE_FOR_builtin_pkwb
,
6407 CODE_FOR_builtin_unpkbl
,
6408 CODE_FOR_builtin_unpkbw
,
6413 CODE_FOR_popcountdi2
6416 struct alpha_builtin_def
6419 enum alpha_builtin code
;
6420 unsigned int target_mask
;
6424 static struct alpha_builtin_def
const zero_arg_builtins
[] = {
6425 { "__builtin_alpha_implver", ALPHA_BUILTIN_IMPLVER
, 0, true },
6426 { "__builtin_alpha_rpcc", ALPHA_BUILTIN_RPCC
, 0, false }
6429 static struct alpha_builtin_def
const one_arg_builtins
[] = {
6430 { "__builtin_alpha_amask", ALPHA_BUILTIN_AMASK
, 0, true },
6431 { "__builtin_alpha_pklb", ALPHA_BUILTIN_PKLB
, MASK_MAX
, true },
6432 { "__builtin_alpha_pkwb", ALPHA_BUILTIN_PKWB
, MASK_MAX
, true },
6433 { "__builtin_alpha_unpkbl", ALPHA_BUILTIN_UNPKBL
, MASK_MAX
, true },
6434 { "__builtin_alpha_unpkbw", ALPHA_BUILTIN_UNPKBW
, MASK_MAX
, true },
6435 { "__builtin_alpha_cttz", ALPHA_BUILTIN_CTTZ
, MASK_CIX
, true },
6436 { "__builtin_alpha_ctlz", ALPHA_BUILTIN_CTLZ
, MASK_CIX
, true },
6437 { "__builtin_alpha_ctpop", ALPHA_BUILTIN_CTPOP
, MASK_CIX
, true }
6440 static struct alpha_builtin_def
const two_arg_builtins
[] = {
6441 { "__builtin_alpha_cmpbge", ALPHA_BUILTIN_CMPBGE
, 0, true },
6442 { "__builtin_alpha_extbl", ALPHA_BUILTIN_EXTBL
, 0, true },
6443 { "__builtin_alpha_extwl", ALPHA_BUILTIN_EXTWL
, 0, true },
6444 { "__builtin_alpha_extll", ALPHA_BUILTIN_EXTLL
, 0, true },
6445 { "__builtin_alpha_extql", ALPHA_BUILTIN_EXTQL
, 0, true },
6446 { "__builtin_alpha_extwh", ALPHA_BUILTIN_EXTWH
, 0, true },
6447 { "__builtin_alpha_extlh", ALPHA_BUILTIN_EXTLH
, 0, true },
6448 { "__builtin_alpha_extqh", ALPHA_BUILTIN_EXTQH
, 0, true },
6449 { "__builtin_alpha_insbl", ALPHA_BUILTIN_INSBL
, 0, true },
6450 { "__builtin_alpha_inswl", ALPHA_BUILTIN_INSWL
, 0, true },
6451 { "__builtin_alpha_insll", ALPHA_BUILTIN_INSLL
, 0, true },
6452 { "__builtin_alpha_insql", ALPHA_BUILTIN_INSQL
, 0, true },
6453 { "__builtin_alpha_inswh", ALPHA_BUILTIN_INSWH
, 0, true },
6454 { "__builtin_alpha_inslh", ALPHA_BUILTIN_INSLH
, 0, true },
6455 { "__builtin_alpha_insqh", ALPHA_BUILTIN_INSQH
, 0, true },
6456 { "__builtin_alpha_mskbl", ALPHA_BUILTIN_MSKBL
, 0, true },
6457 { "__builtin_alpha_mskwl", ALPHA_BUILTIN_MSKWL
, 0, true },
6458 { "__builtin_alpha_mskll", ALPHA_BUILTIN_MSKLL
, 0, true },
6459 { "__builtin_alpha_mskql", ALPHA_BUILTIN_MSKQL
, 0, true },
6460 { "__builtin_alpha_mskwh", ALPHA_BUILTIN_MSKWH
, 0, true },
6461 { "__builtin_alpha_msklh", ALPHA_BUILTIN_MSKLH
, 0, true },
6462 { "__builtin_alpha_mskqh", ALPHA_BUILTIN_MSKQH
, 0, true },
6463 { "__builtin_alpha_umulh", ALPHA_BUILTIN_UMULH
, 0, true },
6464 { "__builtin_alpha_zap", ALPHA_BUILTIN_ZAP
, 0, true },
6465 { "__builtin_alpha_zapnot", ALPHA_BUILTIN_ZAPNOT
, 0, true },
6466 { "__builtin_alpha_minub8", ALPHA_BUILTIN_MINUB8
, MASK_MAX
, true },
6467 { "__builtin_alpha_minsb8", ALPHA_BUILTIN_MINSB8
, MASK_MAX
, true },
6468 { "__builtin_alpha_minuw4", ALPHA_BUILTIN_MINUW4
, MASK_MAX
, true },
6469 { "__builtin_alpha_minsw4", ALPHA_BUILTIN_MINSW4
, MASK_MAX
, true },
6470 { "__builtin_alpha_maxub8", ALPHA_BUILTIN_MAXUB8
, MASK_MAX
, true },
6471 { "__builtin_alpha_maxsb8", ALPHA_BUILTIN_MAXSB8
, MASK_MAX
, true },
6472 { "__builtin_alpha_maxuw4", ALPHA_BUILTIN_MAXUW4
, MASK_MAX
, true },
6473 { "__builtin_alpha_maxsw4", ALPHA_BUILTIN_MAXSW4
, MASK_MAX
, true },
6474 { "__builtin_alpha_perr", ALPHA_BUILTIN_PERR
, MASK_MAX
, true }
6477 static GTY(()) tree alpha_v8qi_u
;
6478 static GTY(()) tree alpha_v8qi_s
;
6479 static GTY(()) tree alpha_v4hi_u
;
6480 static GTY(()) tree alpha_v4hi_s
;
6482 /* Helper function of alpha_init_builtins. Add the COUNT built-in
6483 functions pointed to by P, with function type FTYPE. */
6486 alpha_add_builtins (const struct alpha_builtin_def
*p
, size_t count
,
6492 for (i
= 0; i
< count
; ++i
, ++p
)
6493 if ((target_flags
& p
->target_mask
) == p
->target_mask
)
6495 decl
= add_builtin_function (p
->name
, ftype
, p
->code
, BUILT_IN_MD
,
6498 TREE_READONLY (decl
) = 1;
6499 TREE_NOTHROW (decl
) = 1;
6505 alpha_init_builtins (void)
6507 tree dimode_integer_type_node
;
6510 dimode_integer_type_node
= lang_hooks
.types
.type_for_mode (DImode
, 0);
6512 ftype
= build_function_type (dimode_integer_type_node
, void_list_node
);
6513 alpha_add_builtins (zero_arg_builtins
, ARRAY_SIZE (zero_arg_builtins
),
6516 ftype
= build_function_type_list (dimode_integer_type_node
,
6517 dimode_integer_type_node
, NULL_TREE
);
6518 alpha_add_builtins (one_arg_builtins
, ARRAY_SIZE (one_arg_builtins
),
6521 ftype
= build_function_type_list (dimode_integer_type_node
,
6522 dimode_integer_type_node
,
6523 dimode_integer_type_node
, NULL_TREE
);
6524 alpha_add_builtins (two_arg_builtins
, ARRAY_SIZE (two_arg_builtins
),
6527 ftype
= build_function_type (ptr_type_node
, void_list_node
);
6528 decl
= add_builtin_function ("__builtin_thread_pointer", ftype
,
6529 ALPHA_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
6531 TREE_NOTHROW (decl
) = 1;
6533 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
6534 decl
= add_builtin_function ("__builtin_set_thread_pointer", ftype
,
6535 ALPHA_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
6537 TREE_NOTHROW (decl
) = 1;
6539 alpha_v8qi_u
= build_vector_type (unsigned_intQI_type_node
, 8);
6540 alpha_v8qi_s
= build_vector_type (intQI_type_node
, 8);
6541 alpha_v4hi_u
= build_vector_type (unsigned_intHI_type_node
, 4);
6542 alpha_v4hi_s
= build_vector_type (intHI_type_node
, 4);
6545 /* Expand an expression EXP that calls a built-in function,
6546 with result going to TARGET if that's convenient
6547 (and in mode MODE if that's convenient).
6548 SUBTARGET may be used as the target for computing one of EXP's operands.
6549 IGNORE is nonzero if the value is to be ignored. */
6552 alpha_expand_builtin (tree exp
, rtx target
,
6553 rtx subtarget ATTRIBUTE_UNUSED
,
6554 enum machine_mode mode ATTRIBUTE_UNUSED
,
6555 int ignore ATTRIBUTE_UNUSED
)
6559 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
6560 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
6562 call_expr_arg_iterator iter
;
6563 enum insn_code icode
;
6564 rtx op
[MAX_ARGS
], pat
;
6568 if (fcode
>= ALPHA_BUILTIN_max
)
6569 internal_error ("bad builtin fcode");
6570 icode
= code_for_builtin
[fcode
];
6572 internal_error ("bad builtin fcode");
6574 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
6577 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
6579 const struct insn_operand_data
*insn_op
;
6581 if (arg
== error_mark_node
)
6583 if (arity
> MAX_ARGS
)
6586 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
6588 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
6590 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
6591 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
6597 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
6599 || GET_MODE (target
) != tmode
6600 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
6601 target
= gen_reg_rtx (tmode
);
6607 pat
= GEN_FCN (icode
) (target
);
6611 pat
= GEN_FCN (icode
) (target
, op
[0]);
6613 pat
= GEN_FCN (icode
) (op
[0]);
6616 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
6632 /* Several bits below assume HWI >= 64 bits. This should be enforced
6634 #if HOST_BITS_PER_WIDE_INT < 64
6635 # error "HOST_WIDE_INT too small"
6638 /* Fold the builtin for the CMPBGE instruction. This is a vector comparison
6639 with an 8-bit output vector. OPINT contains the integer operands; bit N
6640 of OP_CONST is set if OPINT[N] is valid. */
6643 alpha_fold_builtin_cmpbge (unsigned HOST_WIDE_INT opint
[], long op_const
)
6648 for (i
= 0, val
= 0; i
< 8; ++i
)
6650 unsigned HOST_WIDE_INT c0
= (opint
[0] >> (i
* 8)) & 0xff;
6651 unsigned HOST_WIDE_INT c1
= (opint
[1] >> (i
* 8)) & 0xff;
6655 return build_int_cst (long_integer_type_node
, val
);
6657 else if (op_const
== 2 && opint
[1] == 0)
6658 return build_int_cst (long_integer_type_node
, 0xff);
6662 /* Fold the builtin for the ZAPNOT instruction. This is essentially a
6663 specialized form of an AND operation. Other byte manipulation instructions
6664 are defined in terms of this instruction, so this is also used as a
6665 subroutine for other builtins.
6667 OP contains the tree operands; OPINT contains the extracted integer values.
6668 Bit N of OP_CONST it set if OPINT[N] is valid. OP may be null if only
6669 OPINT may be considered. */
6672 alpha_fold_builtin_zapnot (tree
*op
, unsigned HOST_WIDE_INT opint
[],
6677 unsigned HOST_WIDE_INT mask
= 0;
6680 for (i
= 0; i
< 8; ++i
)
6681 if ((opint
[1] >> i
) & 1)
6682 mask
|= (unsigned HOST_WIDE_INT
)0xff << (i
* 8);
6685 return build_int_cst (long_integer_type_node
, opint
[0] & mask
);
6688 return fold_build2 (BIT_AND_EXPR
, long_integer_type_node
, op
[0],
6689 build_int_cst (long_integer_type_node
, mask
));
6691 else if ((op_const
& 1) && opint
[0] == 0)
6692 return build_int_cst (long_integer_type_node
, 0);
6696 /* Fold the builtins for the EXT family of instructions. */
6699 alpha_fold_builtin_extxx (tree op
[], unsigned HOST_WIDE_INT opint
[],
6700 long op_const
, unsigned HOST_WIDE_INT bytemask
,
6704 tree
*zap_op
= NULL
;
6708 unsigned HOST_WIDE_INT loc
;
6711 if (BYTES_BIG_ENDIAN
)
6719 unsigned HOST_WIDE_INT temp
= opint
[0];
6732 opint
[1] = bytemask
;
6733 return alpha_fold_builtin_zapnot (zap_op
, opint
, zap_const
);
6736 /* Fold the builtins for the INS family of instructions. */
6739 alpha_fold_builtin_insxx (tree op
[], unsigned HOST_WIDE_INT opint
[],
6740 long op_const
, unsigned HOST_WIDE_INT bytemask
,
6743 if ((op_const
& 1) && opint
[0] == 0)
6744 return build_int_cst (long_integer_type_node
, 0);
6748 unsigned HOST_WIDE_INT temp
, loc
, byteloc
;
6749 tree
*zap_op
= NULL
;
6752 if (BYTES_BIG_ENDIAN
)
6759 byteloc
= (64 - (loc
* 8)) & 0x3f;
6776 opint
[1] = bytemask
;
6777 return alpha_fold_builtin_zapnot (zap_op
, opint
, op_const
);
6784 alpha_fold_builtin_mskxx (tree op
[], unsigned HOST_WIDE_INT opint
[],
6785 long op_const
, unsigned HOST_WIDE_INT bytemask
,
6790 unsigned HOST_WIDE_INT loc
;
6793 if (BYTES_BIG_ENDIAN
)
6800 opint
[1] = bytemask
^ 0xff;
6803 return alpha_fold_builtin_zapnot (op
, opint
, op_const
);
6807 alpha_fold_builtin_umulh (unsigned HOST_WIDE_INT opint
[], long op_const
)
6813 unsigned HOST_WIDE_INT l
;
6816 mul_double (opint
[0], 0, opint
[1], 0, &l
, &h
);
6818 #if HOST_BITS_PER_WIDE_INT > 64
6822 return build_int_cst (long_integer_type_node
, h
);
6826 opint
[1] = opint
[0];
6829 /* Note that (X*1) >> 64 == 0. */
6830 if (opint
[1] == 0 || opint
[1] == 1)
6831 return build_int_cst (long_integer_type_node
, 0);
6838 alpha_fold_vector_minmax (enum tree_code code
, tree op
[], tree vtype
)
6840 tree op0
= fold_convert (vtype
, op
[0]);
6841 tree op1
= fold_convert (vtype
, op
[1]);
6842 tree val
= fold_build2 (code
, vtype
, op0
, op1
);
6843 return fold_build1 (VIEW_CONVERT_EXPR
, long_integer_type_node
, val
);
6847 alpha_fold_builtin_perr (unsigned HOST_WIDE_INT opint
[], long op_const
)
6849 unsigned HOST_WIDE_INT temp
= 0;
6855 for (i
= 0; i
< 8; ++i
)
6857 unsigned HOST_WIDE_INT a
= (opint
[0] >> (i
* 8)) & 0xff;
6858 unsigned HOST_WIDE_INT b
= (opint
[1] >> (i
* 8)) & 0xff;
6865 return build_int_cst (long_integer_type_node
, temp
);
6869 alpha_fold_builtin_pklb (unsigned HOST_WIDE_INT opint
[], long op_const
)
6871 unsigned HOST_WIDE_INT temp
;
6876 temp
= opint
[0] & 0xff;
6877 temp
|= (opint
[0] >> 24) & 0xff00;
6879 return build_int_cst (long_integer_type_node
, temp
);
6883 alpha_fold_builtin_pkwb (unsigned HOST_WIDE_INT opint
[], long op_const
)
6885 unsigned HOST_WIDE_INT temp
;
6890 temp
= opint
[0] & 0xff;
6891 temp
|= (opint
[0] >> 8) & 0xff00;
6892 temp
|= (opint
[0] >> 16) & 0xff0000;
6893 temp
|= (opint
[0] >> 24) & 0xff000000;
6895 return build_int_cst (long_integer_type_node
, temp
);
6899 alpha_fold_builtin_unpkbl (unsigned HOST_WIDE_INT opint
[], long op_const
)
6901 unsigned HOST_WIDE_INT temp
;
6906 temp
= opint
[0] & 0xff;
6907 temp
|= (opint
[0] & 0xff00) << 24;
6909 return build_int_cst (long_integer_type_node
, temp
);
6913 alpha_fold_builtin_unpkbw (unsigned HOST_WIDE_INT opint
[], long op_const
)
6915 unsigned HOST_WIDE_INT temp
;
6920 temp
= opint
[0] & 0xff;
6921 temp
|= (opint
[0] & 0x0000ff00) << 8;
6922 temp
|= (opint
[0] & 0x00ff0000) << 16;
6923 temp
|= (opint
[0] & 0xff000000) << 24;
6925 return build_int_cst (long_integer_type_node
, temp
);
6929 alpha_fold_builtin_cttz (unsigned HOST_WIDE_INT opint
[], long op_const
)
6931 unsigned HOST_WIDE_INT temp
;
6939 temp
= exact_log2 (opint
[0] & -opint
[0]);
6941 return build_int_cst (long_integer_type_node
, temp
);
6945 alpha_fold_builtin_ctlz (unsigned HOST_WIDE_INT opint
[], long op_const
)
6947 unsigned HOST_WIDE_INT temp
;
6955 temp
= 64 - floor_log2 (opint
[0]) - 1;
6957 return build_int_cst (long_integer_type_node
, temp
);
6961 alpha_fold_builtin_ctpop (unsigned HOST_WIDE_INT opint
[], long op_const
)
6963 unsigned HOST_WIDE_INT temp
, op
;
6971 temp
++, op
&= op
- 1;
6973 return build_int_cst (long_integer_type_node
, temp
);
6976 /* Fold one of our builtin functions. */
6979 alpha_fold_builtin (tree fndecl
, tree arglist
, bool ignore ATTRIBUTE_UNUSED
)
6981 tree op
[MAX_ARGS
], t
;
6982 unsigned HOST_WIDE_INT opint
[MAX_ARGS
];
6983 long op_const
= 0, arity
= 0;
6985 for (t
= arglist
; t
; t
= TREE_CHAIN (t
), ++arity
)
6987 tree arg
= TREE_VALUE (t
);
6988 if (arg
== error_mark_node
)
6990 if (arity
>= MAX_ARGS
)
6995 if (TREE_CODE (arg
) == INTEGER_CST
)
6997 op_const
|= 1L << arity
;
6998 opint
[arity
] = int_cst_value (arg
);
7002 switch (DECL_FUNCTION_CODE (fndecl
))
7004 case ALPHA_BUILTIN_CMPBGE
:
7005 return alpha_fold_builtin_cmpbge (opint
, op_const
);
7007 case ALPHA_BUILTIN_EXTBL
:
7008 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0x01, false);
7009 case ALPHA_BUILTIN_EXTWL
:
7010 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0x03, false);
7011 case ALPHA_BUILTIN_EXTLL
:
7012 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0x0f, false);
7013 case ALPHA_BUILTIN_EXTQL
:
7014 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0xff, false);
7015 case ALPHA_BUILTIN_EXTWH
:
7016 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0x03, true);
7017 case ALPHA_BUILTIN_EXTLH
:
7018 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0x0f, true);
7019 case ALPHA_BUILTIN_EXTQH
:
7020 return alpha_fold_builtin_extxx (op
, opint
, op_const
, 0xff, true);
7022 case ALPHA_BUILTIN_INSBL
:
7023 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0x01, false);
7024 case ALPHA_BUILTIN_INSWL
:
7025 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0x03, false);
7026 case ALPHA_BUILTIN_INSLL
:
7027 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0x0f, false);
7028 case ALPHA_BUILTIN_INSQL
:
7029 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0xff, false);
7030 case ALPHA_BUILTIN_INSWH
:
7031 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0x03, true);
7032 case ALPHA_BUILTIN_INSLH
:
7033 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0x0f, true);
7034 case ALPHA_BUILTIN_INSQH
:
7035 return alpha_fold_builtin_insxx (op
, opint
, op_const
, 0xff, true);
7037 case ALPHA_BUILTIN_MSKBL
:
7038 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0x01, false);
7039 case ALPHA_BUILTIN_MSKWL
:
7040 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0x03, false);
7041 case ALPHA_BUILTIN_MSKLL
:
7042 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0x0f, false);
7043 case ALPHA_BUILTIN_MSKQL
:
7044 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0xff, false);
7045 case ALPHA_BUILTIN_MSKWH
:
7046 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0x03, true);
7047 case ALPHA_BUILTIN_MSKLH
:
7048 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0x0f, true);
7049 case ALPHA_BUILTIN_MSKQH
:
7050 return alpha_fold_builtin_mskxx (op
, opint
, op_const
, 0xff, true);
7052 case ALPHA_BUILTIN_UMULH
:
7053 return alpha_fold_builtin_umulh (opint
, op_const
);
7055 case ALPHA_BUILTIN_ZAP
:
7058 case ALPHA_BUILTIN_ZAPNOT
:
7059 return alpha_fold_builtin_zapnot (op
, opint
, op_const
);
7061 case ALPHA_BUILTIN_MINUB8
:
7062 return alpha_fold_vector_minmax (MIN_EXPR
, op
, alpha_v8qi_u
);
7063 case ALPHA_BUILTIN_MINSB8
:
7064 return alpha_fold_vector_minmax (MIN_EXPR
, op
, alpha_v8qi_s
);
7065 case ALPHA_BUILTIN_MINUW4
:
7066 return alpha_fold_vector_minmax (MIN_EXPR
, op
, alpha_v4hi_u
);
7067 case ALPHA_BUILTIN_MINSW4
:
7068 return alpha_fold_vector_minmax (MIN_EXPR
, op
, alpha_v4hi_s
);
7069 case ALPHA_BUILTIN_MAXUB8
:
7070 return alpha_fold_vector_minmax (MAX_EXPR
, op
, alpha_v8qi_u
);
7071 case ALPHA_BUILTIN_MAXSB8
:
7072 return alpha_fold_vector_minmax (MAX_EXPR
, op
, alpha_v8qi_s
);
7073 case ALPHA_BUILTIN_MAXUW4
:
7074 return alpha_fold_vector_minmax (MAX_EXPR
, op
, alpha_v4hi_u
);
7075 case ALPHA_BUILTIN_MAXSW4
:
7076 return alpha_fold_vector_minmax (MAX_EXPR
, op
, alpha_v4hi_s
);
7078 case ALPHA_BUILTIN_PERR
:
7079 return alpha_fold_builtin_perr (opint
, op_const
);
7080 case ALPHA_BUILTIN_PKLB
:
7081 return alpha_fold_builtin_pklb (opint
, op_const
);
7082 case ALPHA_BUILTIN_PKWB
:
7083 return alpha_fold_builtin_pkwb (opint
, op_const
);
7084 case ALPHA_BUILTIN_UNPKBL
:
7085 return alpha_fold_builtin_unpkbl (opint
, op_const
);
7086 case ALPHA_BUILTIN_UNPKBW
:
7087 return alpha_fold_builtin_unpkbw (opint
, op_const
);
7089 case ALPHA_BUILTIN_CTTZ
:
7090 return alpha_fold_builtin_cttz (opint
, op_const
);
7091 case ALPHA_BUILTIN_CTLZ
:
7092 return alpha_fold_builtin_ctlz (opint
, op_const
);
7093 case ALPHA_BUILTIN_CTPOP
:
7094 return alpha_fold_builtin_ctpop (opint
, op_const
);
7096 case ALPHA_BUILTIN_AMASK
:
7097 case ALPHA_BUILTIN_IMPLVER
:
7098 case ALPHA_BUILTIN_RPCC
:
7099 case ALPHA_BUILTIN_THREAD_POINTER
:
7100 case ALPHA_BUILTIN_SET_THREAD_POINTER
:
7101 /* None of these are foldable at compile-time. */
7107 /* This page contains routines that are used to determine what the function
7108 prologue and epilogue code will do and write them out. */
7110 /* Compute the size of the save area in the stack. */
7112 /* These variables are used for communication between the following functions.
7113 They indicate various things about the current function being compiled
7114 that are used to tell what kind of prologue, epilogue and procedure
7115 descriptor to generate. */
7117 /* Nonzero if we need a stack procedure. */
7118 enum alpha_procedure_types
{PT_NULL
= 0, PT_REGISTER
= 1, PT_STACK
= 2};
7119 static enum alpha_procedure_types alpha_procedure_type
;
7121 /* Register number (either FP or SP) that is used to unwind the frame. */
7122 static int vms_unwind_regno
;
7124 /* Register number used to save FP. We need not have one for RA since
7125 we don't modify it for register procedures. This is only defined
7126 for register frame procedures. */
7127 static int vms_save_fp_regno
;
7129 /* Register number used to reference objects off our PV. */
7130 static int vms_base_regno
;
7132 /* Compute register masks for saved registers. */
7135 alpha_sa_mask (unsigned long *imaskP
, unsigned long *fmaskP
)
7137 unsigned long imask
= 0;
7138 unsigned long fmask
= 0;
7141 /* When outputting a thunk, we don't have valid register life info,
7142 but assemble_start_function wants to output .frame and .mask
7151 if (TARGET_ABI_OPEN_VMS
&& alpha_procedure_type
== PT_STACK
)
7152 imask
|= (1UL << HARD_FRAME_POINTER_REGNUM
);
7154 /* One for every register we have to save. */
7155 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
7156 if (! fixed_regs
[i
] && ! call_used_regs
[i
]
7157 && df_regs_ever_live_p (i
) && i
!= REG_RA
7158 && (!TARGET_ABI_UNICOSMK
|| i
!= HARD_FRAME_POINTER_REGNUM
))
7161 imask
|= (1UL << i
);
7163 fmask
|= (1UL << (i
- 32));
7166 /* We need to restore these for the handler. */
7167 if (crtl
->calls_eh_return
)
7171 unsigned regno
= EH_RETURN_DATA_REGNO (i
);
7172 if (regno
== INVALID_REGNUM
)
7174 imask
|= 1UL << regno
;
7178 /* If any register spilled, then spill the return address also. */
7179 /* ??? This is required by the Digital stack unwind specification
7180 and isn't needed if we're doing Dwarf2 unwinding. */
7181 if (imask
|| fmask
|| alpha_ra_ever_killed ())
7182 imask
|= (1UL << REG_RA
);
7189 alpha_sa_size (void)
7191 unsigned long mask
[2];
7195 alpha_sa_mask (&mask
[0], &mask
[1]);
7197 if (TARGET_ABI_UNICOSMK
)
7199 if (mask
[0] || mask
[1])
7204 for (j
= 0; j
< 2; ++j
)
7205 for (i
= 0; i
< 32; ++i
)
7206 if ((mask
[j
] >> i
) & 1)
7210 if (TARGET_ABI_UNICOSMK
)
7212 /* We might not need to generate a frame if we don't make any calls
7213 (including calls to __T3E_MISMATCH if this is a vararg function),
7214 don't have any local variables which require stack slots, don't
7215 use alloca and have not determined that we need a frame for other
7218 alpha_procedure_type
7219 = (sa_size
|| get_frame_size() != 0
7220 || crtl
->outgoing_args_size
7221 || cfun
->stdarg
|| cfun
->calls_alloca
7222 || frame_pointer_needed
)
7223 ? PT_STACK
: PT_REGISTER
;
7225 /* Always reserve space for saving callee-saved registers if we
7226 need a frame as required by the calling convention. */
7227 if (alpha_procedure_type
== PT_STACK
)
7230 else if (TARGET_ABI_OPEN_VMS
)
7232 /* Start by assuming we can use a register procedure if we don't
7233 make any calls (REG_RA not used) or need to save any
7234 registers and a stack procedure if we do. */
7235 if ((mask
[0] >> REG_RA
) & 1)
7236 alpha_procedure_type
= PT_STACK
;
7237 else if (get_frame_size() != 0)
7238 alpha_procedure_type
= PT_REGISTER
;
7240 alpha_procedure_type
= PT_NULL
;
7242 /* Don't reserve space for saving FP & RA yet. Do that later after we've
7243 made the final decision on stack procedure vs register procedure. */
7244 if (alpha_procedure_type
== PT_STACK
)
7247 /* Decide whether to refer to objects off our PV via FP or PV.
7248 If we need FP for something else or if we receive a nonlocal
7249 goto (which expects PV to contain the value), we must use PV.
7250 Otherwise, start by assuming we can use FP. */
7253 = (frame_pointer_needed
7254 || cfun
->has_nonlocal_label
7255 || alpha_procedure_type
== PT_STACK
7256 || crtl
->outgoing_args_size
)
7257 ? REG_PV
: HARD_FRAME_POINTER_REGNUM
;
7259 /* If we want to copy PV into FP, we need to find some register
7260 in which to save FP. */
7262 vms_save_fp_regno
= -1;
7263 if (vms_base_regno
== HARD_FRAME_POINTER_REGNUM
)
7264 for (i
= 0; i
< 32; i
++)
7265 if (! fixed_regs
[i
] && call_used_regs
[i
] && ! df_regs_ever_live_p (i
))
7266 vms_save_fp_regno
= i
;
7268 if (vms_save_fp_regno
== -1 && alpha_procedure_type
== PT_REGISTER
)
7269 vms_base_regno
= REG_PV
, alpha_procedure_type
= PT_STACK
;
7270 else if (alpha_procedure_type
== PT_NULL
)
7271 vms_base_regno
= REG_PV
;
7273 /* Stack unwinding should be done via FP unless we use it for PV. */
7274 vms_unwind_regno
= (vms_base_regno
== REG_PV
7275 ? HARD_FRAME_POINTER_REGNUM
: STACK_POINTER_REGNUM
);
7277 /* If this is a stack procedure, allow space for saving FP and RA. */
7278 if (alpha_procedure_type
== PT_STACK
)
7283 /* Our size must be even (multiple of 16 bytes). */
7291 /* Define the offset between two registers, one to be eliminated,
7292 and the other its replacement, at the start of a routine. */
7295 alpha_initial_elimination_offset (unsigned int from
,
7296 unsigned int to ATTRIBUTE_UNUSED
)
7300 ret
= alpha_sa_size ();
7301 ret
+= ALPHA_ROUND (crtl
->outgoing_args_size
);
7305 case FRAME_POINTER_REGNUM
:
7308 case ARG_POINTER_REGNUM
:
7309 ret
+= (ALPHA_ROUND (get_frame_size ()
7310 + crtl
->args
.pretend_args_size
)
7311 - crtl
->args
.pretend_args_size
);
7322 alpha_pv_save_size (void)
7325 return alpha_procedure_type
== PT_STACK
? 8 : 0;
7329 alpha_using_fp (void)
7332 return vms_unwind_regno
== HARD_FRAME_POINTER_REGNUM
;
7335 #if TARGET_ABI_OPEN_VMS
7337 static const struct attribute_spec vms_attribute_table
[] =
7339 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
7340 { "overlaid", 0, 0, true, false, false, NULL
},
7341 { "global", 0, 0, true, false, false, NULL
},
7342 { "initialize", 0, 0, true, false, false, NULL
},
7343 { NULL
, 0, 0, false, false, false, NULL
}
7349 find_lo_sum_using_gp (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
7351 return GET_CODE (*px
) == LO_SUM
&& XEXP (*px
, 0) == pic_offset_table_rtx
;
7355 alpha_find_lo_sum_using_gp (rtx insn
)
7357 return for_each_rtx (&PATTERN (insn
), find_lo_sum_using_gp
, NULL
) > 0;
7361 alpha_does_function_need_gp (void)
7365 /* The GP being variable is an OSF abi thing. */
7366 if (! TARGET_ABI_OSF
)
7369 /* We need the gp to load the address of __mcount. */
7370 if (TARGET_PROFILING_NEEDS_GP
&& crtl
->profile
)
7373 /* The code emitted by alpha_output_mi_thunk_osf uses the gp. */
7377 /* The nonlocal receiver pattern assumes that the gp is valid for
7378 the nested function. Reasonable because it's almost always set
7379 correctly already. For the cases where that's wrong, make sure
7380 the nested function loads its gp on entry. */
7381 if (crtl
->has_nonlocal_goto
)
7384 /* If we need a GP (we have a LDSYM insn or a CALL_INSN), load it first.
7385 Even if we are a static function, we still need to do this in case
7386 our address is taken and passed to something like qsort. */
7388 push_topmost_sequence ();
7389 insn
= get_insns ();
7390 pop_topmost_sequence ();
7392 for (; insn
; insn
= NEXT_INSN (insn
))
7394 && ! JUMP_TABLE_DATA_P (insn
)
7395 && GET_CODE (PATTERN (insn
)) != USE
7396 && GET_CODE (PATTERN (insn
)) != CLOBBER
7397 && get_attr_usegp (insn
))
7404 /* Helper function to set RTX_FRAME_RELATED_P on instructions, including
7408 set_frame_related_p (void)
7410 rtx seq
= get_insns ();
7421 while (insn
!= NULL_RTX
)
7423 RTX_FRAME_RELATED_P (insn
) = 1;
7424 insn
= NEXT_INSN (insn
);
7426 seq
= emit_insn (seq
);
7430 seq
= emit_insn (seq
);
7431 RTX_FRAME_RELATED_P (seq
) = 1;
7436 #define FRP(exp) (start_sequence (), exp, set_frame_related_p ())
7438 /* Generates a store with the proper unwind info attached. VALUE is
7439 stored at BASE_REG+BASE_OFS. If FRAME_BIAS is nonzero, then BASE_REG
7440 contains SP+FRAME_BIAS, and that is the unwind info that should be
7441 generated. If FRAME_REG != VALUE, then VALUE is being stored on
7442 behalf of FRAME_REG, and FRAME_REG should be present in the unwind. */
7445 emit_frame_store_1 (rtx value
, rtx base_reg
, HOST_WIDE_INT frame_bias
,
7446 HOST_WIDE_INT base_ofs
, rtx frame_reg
)
7448 rtx addr
, mem
, insn
;
7450 addr
= plus_constant (base_reg
, base_ofs
);
7451 mem
= gen_rtx_MEM (DImode
, addr
);
7452 set_mem_alias_set (mem
, alpha_sr_alias_set
);
7454 insn
= emit_move_insn (mem
, value
);
7455 RTX_FRAME_RELATED_P (insn
) = 1;
7457 if (frame_bias
|| value
!= frame_reg
)
7461 addr
= plus_constant (stack_pointer_rtx
, frame_bias
+ base_ofs
);
7462 mem
= gen_rtx_MEM (DImode
, addr
);
7465 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
7466 gen_rtx_SET (VOIDmode
, mem
, frame_reg
));
7471 emit_frame_store (unsigned int regno
, rtx base_reg
,
7472 HOST_WIDE_INT frame_bias
, HOST_WIDE_INT base_ofs
)
7474 rtx reg
= gen_rtx_REG (DImode
, regno
);
7475 emit_frame_store_1 (reg
, base_reg
, frame_bias
, base_ofs
, reg
);
7478 /* Write function prologue. */
7480 /* On vms we have two kinds of functions:
7482 - stack frame (PROC_STACK)
7483 these are 'normal' functions with local vars and which are
7484 calling other functions
7485 - register frame (PROC_REGISTER)
7486 keeps all data in registers, needs no stack
7488 We must pass this to the assembler so it can generate the
7489 proper pdsc (procedure descriptor)
7490 This is done with the '.pdesc' command.
7492 On not-vms, we don't really differentiate between the two, as we can
7493 simply allocate stack without saving registers. */
7496 alpha_expand_prologue (void)
7498 /* Registers to save. */
7499 unsigned long imask
= 0;
7500 unsigned long fmask
= 0;
7501 /* Stack space needed for pushing registers clobbered by us. */
7502 HOST_WIDE_INT sa_size
;
7503 /* Complete stack size needed. */
7504 HOST_WIDE_INT frame_size
;
7505 /* Offset from base reg to register save area. */
7506 HOST_WIDE_INT reg_offset
;
7510 sa_size
= alpha_sa_size ();
7512 frame_size
= get_frame_size ();
7513 if (TARGET_ABI_OPEN_VMS
)
7514 frame_size
= ALPHA_ROUND (sa_size
7515 + (alpha_procedure_type
== PT_STACK
? 8 : 0)
7517 + crtl
->args
.pretend_args_size
);
7518 else if (TARGET_ABI_UNICOSMK
)
7519 /* We have to allocate space for the DSIB if we generate a frame. */
7520 frame_size
= ALPHA_ROUND (sa_size
7521 + (alpha_procedure_type
== PT_STACK
? 48 : 0))
7522 + ALPHA_ROUND (frame_size
7523 + crtl
->outgoing_args_size
);
7525 frame_size
= (ALPHA_ROUND (crtl
->outgoing_args_size
)
7527 + ALPHA_ROUND (frame_size
7528 + crtl
->args
.pretend_args_size
));
7530 if (TARGET_ABI_OPEN_VMS
)
7533 reg_offset
= ALPHA_ROUND (crtl
->outgoing_args_size
);
7535 alpha_sa_mask (&imask
, &fmask
);
7537 /* Emit an insn to reload GP, if needed. */
7540 alpha_function_needs_gp
= alpha_does_function_need_gp ();
7541 if (alpha_function_needs_gp
)
7542 emit_insn (gen_prologue_ldgp ());
7545 /* TARGET_PROFILING_NEEDS_GP actually implies that we need to insert
7546 the call to mcount ourselves, rather than having the linker do it
7547 magically in response to -pg. Since _mcount has special linkage,
7548 don't represent the call as a call. */
7549 if (TARGET_PROFILING_NEEDS_GP
&& crtl
->profile
)
7550 emit_insn (gen_prologue_mcount ());
7552 if (TARGET_ABI_UNICOSMK
)
7553 unicosmk_gen_dsib (&imask
);
7555 /* Adjust the stack by the frame size. If the frame size is > 4096
7556 bytes, we need to be sure we probe somewhere in the first and last
7557 4096 bytes (we can probably get away without the latter test) and
7558 every 8192 bytes in between. If the frame size is > 32768, we
7559 do this in a loop. Otherwise, we generate the explicit probe
7562 Note that we are only allowed to adjust sp once in the prologue. */
7564 if (frame_size
<= 32768)
7566 if (frame_size
> 4096)
7570 for (probed
= 4096; probed
< frame_size
; probed
+= 8192)
7571 emit_insn (gen_probe_stack (GEN_INT (TARGET_ABI_UNICOSMK
7575 /* We only have to do this probe if we aren't saving registers. */
7576 if (sa_size
== 0 && frame_size
> probed
- 4096)
7577 emit_insn (gen_probe_stack (GEN_INT (-frame_size
)));
7580 if (frame_size
!= 0)
7581 FRP (emit_insn (gen_adddi3 (stack_pointer_rtx
, stack_pointer_rtx
,
7582 GEN_INT (TARGET_ABI_UNICOSMK
7588 /* Here we generate code to set R22 to SP + 4096 and set R23 to the
7589 number of 8192 byte blocks to probe. We then probe each block
7590 in the loop and then set SP to the proper location. If the
7591 amount remaining is > 4096, we have to do one more probe if we
7592 are not saving any registers. */
7594 HOST_WIDE_INT blocks
= (frame_size
+ 4096) / 8192;
7595 HOST_WIDE_INT leftover
= frame_size
+ 4096 - blocks
* 8192;
7596 rtx ptr
= gen_rtx_REG (DImode
, 22);
7597 rtx count
= gen_rtx_REG (DImode
, 23);
7600 emit_move_insn (count
, GEN_INT (blocks
));
7601 emit_insn (gen_adddi3 (ptr
, stack_pointer_rtx
,
7602 GEN_INT (TARGET_ABI_UNICOSMK
? 4096 - 64 : 4096)));
7604 /* Because of the difficulty in emitting a new basic block this
7605 late in the compilation, generate the loop as a single insn. */
7606 emit_insn (gen_prologue_stack_probe_loop (count
, ptr
));
7608 if (leftover
> 4096 && sa_size
== 0)
7610 rtx last
= gen_rtx_MEM (DImode
, plus_constant (ptr
, -leftover
));
7611 MEM_VOLATILE_P (last
) = 1;
7612 emit_move_insn (last
, const0_rtx
);
7615 if (TARGET_ABI_WINDOWS_NT
)
7617 /* For NT stack unwind (done by 'reverse execution'), it's
7618 not OK to take the result of a loop, even though the value
7619 is already in ptr, so we reload it via a single operation
7620 and subtract it to sp.
7622 Yes, that's correct -- we have to reload the whole constant
7623 into a temporary via ldah+lda then subtract from sp. */
7625 HOST_WIDE_INT lo
, hi
;
7626 lo
= ((frame_size
& 0xffff) ^ 0x8000) - 0x8000;
7627 hi
= frame_size
- lo
;
7629 emit_move_insn (ptr
, GEN_INT (hi
));
7630 emit_insn (gen_adddi3 (ptr
, ptr
, GEN_INT (lo
)));
7631 seq
= emit_insn (gen_subdi3 (stack_pointer_rtx
, stack_pointer_rtx
,
7636 seq
= emit_insn (gen_adddi3 (stack_pointer_rtx
, ptr
,
7637 GEN_INT (-leftover
)));
7640 /* This alternative is special, because the DWARF code cannot
7641 possibly intuit through the loop above. So we invent this
7642 note it looks at instead. */
7643 RTX_FRAME_RELATED_P (seq
) = 1;
7644 add_reg_note (seq
, REG_FRAME_RELATED_EXPR
,
7645 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
7646 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
7647 GEN_INT (TARGET_ABI_UNICOSMK
7652 if (!TARGET_ABI_UNICOSMK
)
7654 HOST_WIDE_INT sa_bias
= 0;
7656 /* Cope with very large offsets to the register save area. */
7657 sa_reg
= stack_pointer_rtx
;
7658 if (reg_offset
+ sa_size
> 0x8000)
7660 int low
= ((reg_offset
& 0xffff) ^ 0x8000) - 0x8000;
7663 if (low
+ sa_size
<= 0x8000)
7664 sa_bias
= reg_offset
- low
, reg_offset
= low
;
7666 sa_bias
= reg_offset
, reg_offset
= 0;
7668 sa_reg
= gen_rtx_REG (DImode
, 24);
7669 sa_bias_rtx
= GEN_INT (sa_bias
);
7671 if (add_operand (sa_bias_rtx
, DImode
))
7672 emit_insn (gen_adddi3 (sa_reg
, stack_pointer_rtx
, sa_bias_rtx
));
7675 emit_move_insn (sa_reg
, sa_bias_rtx
);
7676 emit_insn (gen_adddi3 (sa_reg
, stack_pointer_rtx
, sa_reg
));
7680 /* Save regs in stack order. Beginning with VMS PV. */
7681 if (TARGET_ABI_OPEN_VMS
&& alpha_procedure_type
== PT_STACK
)
7682 emit_frame_store (REG_PV
, stack_pointer_rtx
, 0, 0);
7684 /* Save register RA next. */
7685 if (imask
& (1UL << REG_RA
))
7687 emit_frame_store (REG_RA
, sa_reg
, sa_bias
, reg_offset
);
7688 imask
&= ~(1UL << REG_RA
);
7692 /* Now save any other registers required to be saved. */
7693 for (i
= 0; i
< 31; i
++)
7694 if (imask
& (1UL << i
))
7696 emit_frame_store (i
, sa_reg
, sa_bias
, reg_offset
);
7700 for (i
= 0; i
< 31; i
++)
7701 if (fmask
& (1UL << i
))
7703 emit_frame_store (i
+32, sa_reg
, sa_bias
, reg_offset
);
7707 else if (TARGET_ABI_UNICOSMK
&& alpha_procedure_type
== PT_STACK
)
7709 /* The standard frame on the T3E includes space for saving registers.
7710 We just have to use it. We don't have to save the return address and
7711 the old frame pointer here - they are saved in the DSIB. */
7714 for (i
= 9; i
< 15; i
++)
7715 if (imask
& (1UL << i
))
7717 emit_frame_store (i
, hard_frame_pointer_rtx
, 0, reg_offset
);
7720 for (i
= 2; i
< 10; i
++)
7721 if (fmask
& (1UL << i
))
7723 emit_frame_store (i
+32, hard_frame_pointer_rtx
, 0, reg_offset
);
7728 if (TARGET_ABI_OPEN_VMS
)
7730 if (alpha_procedure_type
== PT_REGISTER
)
7731 /* Register frame procedures save the fp.
7732 ?? Ought to have a dwarf2 save for this. */
7733 emit_move_insn (gen_rtx_REG (DImode
, vms_save_fp_regno
),
7734 hard_frame_pointer_rtx
);
7736 if (alpha_procedure_type
!= PT_NULL
&& vms_base_regno
!= REG_PV
)
7737 emit_insn (gen_force_movdi (gen_rtx_REG (DImode
, vms_base_regno
),
7738 gen_rtx_REG (DImode
, REG_PV
)));
7740 if (alpha_procedure_type
!= PT_NULL
7741 && vms_unwind_regno
== HARD_FRAME_POINTER_REGNUM
)
7742 FRP (emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
));
7744 /* If we have to allocate space for outgoing args, do it now. */
7745 if (crtl
->outgoing_args_size
!= 0)
7748 = emit_move_insn (stack_pointer_rtx
,
7750 (hard_frame_pointer_rtx
,
7752 (crtl
->outgoing_args_size
))));
7754 /* Only set FRAME_RELATED_P on the stack adjustment we just emitted
7755 if ! frame_pointer_needed. Setting the bit will change the CFA
7756 computation rule to use sp again, which would be wrong if we had
7757 frame_pointer_needed, as this means sp might move unpredictably
7761 frame_pointer_needed
7762 => vms_unwind_regno == HARD_FRAME_POINTER_REGNUM
7764 crtl->outgoing_args_size != 0
7765 => alpha_procedure_type != PT_NULL,
7767 so when we are not setting the bit here, we are guaranteed to
7768 have emitted an FRP frame pointer update just before. */
7769 RTX_FRAME_RELATED_P (seq
) = ! frame_pointer_needed
;
7772 else if (!TARGET_ABI_UNICOSMK
)
7774 /* If we need a frame pointer, set it from the stack pointer. */
7775 if (frame_pointer_needed
)
7777 if (TARGET_CAN_FAULT_IN_PROLOGUE
)
7778 FRP (emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
));
7780 /* This must always be the last instruction in the
7781 prologue, thus we emit a special move + clobber. */
7782 FRP (emit_insn (gen_init_fp (hard_frame_pointer_rtx
,
7783 stack_pointer_rtx
, sa_reg
)));
7787 /* The ABIs for VMS and OSF/1 say that while we can schedule insns into
7788 the prologue, for exception handling reasons, we cannot do this for
7789 any insn that might fault. We could prevent this for mems with a
7790 (clobber:BLK (scratch)), but this doesn't work for fp insns. So we
7791 have to prevent all such scheduling with a blockage.
7793 Linux, on the other hand, never bothered to implement OSF/1's
7794 exception handling, and so doesn't care about such things. Anyone
7795 planning to use dwarf2 frame-unwind info can also omit the blockage. */
7797 if (! TARGET_CAN_FAULT_IN_PROLOGUE
)
7798 emit_insn (gen_blockage ());
7801 /* Count the number of .file directives, so that .loc is up to date. */
7802 int num_source_filenames
= 0;
7804 /* Output the textual info surrounding the prologue. */
7807 alpha_start_function (FILE *file
, const char *fnname
,
7808 tree decl ATTRIBUTE_UNUSED
)
7810 unsigned long imask
= 0;
7811 unsigned long fmask
= 0;
7812 /* Stack space needed for pushing registers clobbered by us. */
7813 HOST_WIDE_INT sa_size
;
7814 /* Complete stack size needed. */
7815 unsigned HOST_WIDE_INT frame_size
;
7816 /* The maximum debuggable frame size (512 Kbytes using Tru64 as). */
7817 unsigned HOST_WIDE_INT max_frame_size
= TARGET_ABI_OSF
&& !TARGET_GAS
7820 /* Offset from base reg to register save area. */
7821 HOST_WIDE_INT reg_offset
;
7822 char *entry_label
= (char *) alloca (strlen (fnname
) + 6);
7825 /* Don't emit an extern directive for functions defined in the same file. */
7826 if (TARGET_ABI_UNICOSMK
)
7829 name_tree
= get_identifier (fnname
);
7830 TREE_ASM_WRITTEN (name_tree
) = 1;
7833 alpha_fnname
= fnname
;
7834 sa_size
= alpha_sa_size ();
7836 frame_size
= get_frame_size ();
7837 if (TARGET_ABI_OPEN_VMS
)
7838 frame_size
= ALPHA_ROUND (sa_size
7839 + (alpha_procedure_type
== PT_STACK
? 8 : 0)
7841 + crtl
->args
.pretend_args_size
);
7842 else if (TARGET_ABI_UNICOSMK
)
7843 frame_size
= ALPHA_ROUND (sa_size
7844 + (alpha_procedure_type
== PT_STACK
? 48 : 0))
7845 + ALPHA_ROUND (frame_size
7846 + crtl
->outgoing_args_size
);
7848 frame_size
= (ALPHA_ROUND (crtl
->outgoing_args_size
)
7850 + ALPHA_ROUND (frame_size
7851 + crtl
->args
.pretend_args_size
));
7853 if (TARGET_ABI_OPEN_VMS
)
7856 reg_offset
= ALPHA_ROUND (crtl
->outgoing_args_size
);
7858 alpha_sa_mask (&imask
, &fmask
);
7860 /* Ecoff can handle multiple .file directives, so put out file and lineno.
7861 We have to do that before the .ent directive as we cannot switch
7862 files within procedures with native ecoff because line numbers are
7863 linked to procedure descriptors.
7864 Outputting the lineno helps debugging of one line functions as they
7865 would otherwise get no line number at all. Please note that we would
7866 like to put out last_linenum from final.c, but it is not accessible. */
7868 if (write_symbols
== SDB_DEBUG
)
7870 #ifdef ASM_OUTPUT_SOURCE_FILENAME
7871 ASM_OUTPUT_SOURCE_FILENAME (file
,
7872 DECL_SOURCE_FILE (current_function_decl
));
7874 #ifdef SDB_OUTPUT_SOURCE_LINE
7875 if (debug_info_level
!= DINFO_LEVEL_TERSE
)
7876 SDB_OUTPUT_SOURCE_LINE (file
,
7877 DECL_SOURCE_LINE (current_function_decl
));
7881 /* Issue function start and label. */
7882 if (TARGET_ABI_OPEN_VMS
7883 || (!TARGET_ABI_UNICOSMK
&& !flag_inhibit_size_directive
))
7885 fputs ("\t.ent ", file
);
7886 assemble_name (file
, fnname
);
7889 /* If the function needs GP, we'll write the "..ng" label there.
7890 Otherwise, do it here. */
7892 && ! alpha_function_needs_gp
7893 && ! cfun
->is_thunk
)
7896 assemble_name (file
, fnname
);
7897 fputs ("..ng:\n", file
);
7901 strcpy (entry_label
, fnname
);
7902 if (TARGET_ABI_OPEN_VMS
)
7903 strcat (entry_label
, "..en");
7905 /* For public functions, the label must be globalized by appending an
7906 additional colon. */
7907 if (TARGET_ABI_UNICOSMK
&& TREE_PUBLIC (decl
))
7908 strcat (entry_label
, ":");
7910 ASM_OUTPUT_LABEL (file
, entry_label
);
7911 inside_function
= TRUE
;
7913 if (TARGET_ABI_OPEN_VMS
)
7914 fprintf (file
, "\t.base $%d\n", vms_base_regno
);
7916 if (!TARGET_ABI_OPEN_VMS
&& !TARGET_ABI_UNICOSMK
&& TARGET_IEEE_CONFORMANT
7917 && !flag_inhibit_size_directive
)
7919 /* Set flags in procedure descriptor to request IEEE-conformant
7920 math-library routines. The value we set it to is PDSC_EXC_IEEE
7921 (/usr/include/pdsc.h). */
7922 fputs ("\t.eflag 48\n", file
);
7925 /* Set up offsets to alpha virtual arg/local debugging pointer. */
7926 alpha_auto_offset
= -frame_size
+ crtl
->args
.pretend_args_size
;
7927 alpha_arg_offset
= -frame_size
+ 48;
7929 /* Describe our frame. If the frame size is larger than an integer,
7930 print it as zero to avoid an assembler error. We won't be
7931 properly describing such a frame, but that's the best we can do. */
7932 if (TARGET_ABI_UNICOSMK
)
7934 else if (TARGET_ABI_OPEN_VMS
)
7935 fprintf (file
, "\t.frame $%d," HOST_WIDE_INT_PRINT_DEC
",$26,"
7936 HOST_WIDE_INT_PRINT_DEC
"\n",
7938 frame_size
>= (1UL << 31) ? 0 : frame_size
,
7940 else if (!flag_inhibit_size_directive
)
7941 fprintf (file
, "\t.frame $%d," HOST_WIDE_INT_PRINT_DEC
",$26,%d\n",
7942 (frame_pointer_needed
7943 ? HARD_FRAME_POINTER_REGNUM
: STACK_POINTER_REGNUM
),
7944 frame_size
>= max_frame_size
? 0 : frame_size
,
7945 crtl
->args
.pretend_args_size
);
7947 /* Describe which registers were spilled. */
7948 if (TARGET_ABI_UNICOSMK
)
7950 else if (TARGET_ABI_OPEN_VMS
)
7953 /* ??? Does VMS care if mask contains ra? The old code didn't
7954 set it, so I don't here. */
7955 fprintf (file
, "\t.mask 0x%lx,0\n", imask
& ~(1UL << REG_RA
));
7957 fprintf (file
, "\t.fmask 0x%lx,0\n", fmask
);
7958 if (alpha_procedure_type
== PT_REGISTER
)
7959 fprintf (file
, "\t.fp_save $%d\n", vms_save_fp_regno
);
7961 else if (!flag_inhibit_size_directive
)
7965 fprintf (file
, "\t.mask 0x%lx," HOST_WIDE_INT_PRINT_DEC
"\n", imask
,
7966 frame_size
>= max_frame_size
? 0 : reg_offset
- frame_size
);
7968 for (i
= 0; i
< 32; ++i
)
7969 if (imask
& (1UL << i
))
7974 fprintf (file
, "\t.fmask 0x%lx," HOST_WIDE_INT_PRINT_DEC
"\n", fmask
,
7975 frame_size
>= max_frame_size
? 0 : reg_offset
- frame_size
);
7978 #if TARGET_ABI_OPEN_VMS
7979 /* Ifdef'ed cause link_section are only available then. */
7980 switch_to_section (readonly_data_section
);
7981 fprintf (file
, "\t.align 3\n");
7982 assemble_name (file
, fnname
); fputs ("..na:\n", file
);
7983 fputs ("\t.ascii \"", file
);
7984 assemble_name (file
, fnname
);
7985 fputs ("\\0\"\n", file
);
7986 alpha_need_linkage (fnname
, 1);
7987 switch_to_section (text_section
);
7991 /* Emit the .prologue note at the scheduled end of the prologue. */
7994 alpha_output_function_end_prologue (FILE *file
)
7996 if (TARGET_ABI_UNICOSMK
)
7998 else if (TARGET_ABI_OPEN_VMS
)
7999 fputs ("\t.prologue\n", file
);
8000 else if (TARGET_ABI_WINDOWS_NT
)
8001 fputs ("\t.prologue 0\n", file
);
8002 else if (!flag_inhibit_size_directive
)
8003 fprintf (file
, "\t.prologue %d\n",
8004 alpha_function_needs_gp
|| cfun
->is_thunk
);
8007 /* Write function epilogue. */
8009 /* ??? At some point we will want to support full unwind, and so will
8010 need to mark the epilogue as well. At the moment, we just confuse
8013 #define FRP(exp) exp
8016 alpha_expand_epilogue (void)
8018 /* Registers to save. */
8019 unsigned long imask
= 0;
8020 unsigned long fmask
= 0;
8021 /* Stack space needed for pushing registers clobbered by us. */
8022 HOST_WIDE_INT sa_size
;
8023 /* Complete stack size needed. */
8024 HOST_WIDE_INT frame_size
;
8025 /* Offset from base reg to register save area. */
8026 HOST_WIDE_INT reg_offset
;
8027 int fp_is_frame_pointer
, fp_offset
;
8028 rtx sa_reg
, sa_reg_exp
= NULL
;
8029 rtx sp_adj1
, sp_adj2
, mem
;
8033 sa_size
= alpha_sa_size ();
8035 frame_size
= get_frame_size ();
8036 if (TARGET_ABI_OPEN_VMS
)
8037 frame_size
= ALPHA_ROUND (sa_size
8038 + (alpha_procedure_type
== PT_STACK
? 8 : 0)
8040 + crtl
->args
.pretend_args_size
);
8041 else if (TARGET_ABI_UNICOSMK
)
8042 frame_size
= ALPHA_ROUND (sa_size
8043 + (alpha_procedure_type
== PT_STACK
? 48 : 0))
8044 + ALPHA_ROUND (frame_size
8045 + crtl
->outgoing_args_size
);
8047 frame_size
= (ALPHA_ROUND (crtl
->outgoing_args_size
)
8049 + ALPHA_ROUND (frame_size
8050 + crtl
->args
.pretend_args_size
));
8052 if (TARGET_ABI_OPEN_VMS
)
8054 if (alpha_procedure_type
== PT_STACK
)
8060 reg_offset
= ALPHA_ROUND (crtl
->outgoing_args_size
);
8062 alpha_sa_mask (&imask
, &fmask
);
8065 = ((TARGET_ABI_OPEN_VMS
&& alpha_procedure_type
== PT_STACK
)
8066 || (!TARGET_ABI_OPEN_VMS
&& frame_pointer_needed
));
8068 sa_reg
= stack_pointer_rtx
;
8070 if (crtl
->calls_eh_return
)
8071 eh_ofs
= EH_RETURN_STACKADJ_RTX
;
8075 if (!TARGET_ABI_UNICOSMK
&& sa_size
)
8077 /* If we have a frame pointer, restore SP from it. */
8078 if ((TARGET_ABI_OPEN_VMS
8079 && vms_unwind_regno
== HARD_FRAME_POINTER_REGNUM
)
8080 || (!TARGET_ABI_OPEN_VMS
&& frame_pointer_needed
))
8081 FRP (emit_move_insn (stack_pointer_rtx
, hard_frame_pointer_rtx
));
8083 /* Cope with very large offsets to the register save area. */
8084 if (reg_offset
+ sa_size
> 0x8000)
8086 int low
= ((reg_offset
& 0xffff) ^ 0x8000) - 0x8000;
8089 if (low
+ sa_size
<= 0x8000)
8090 bias
= reg_offset
- low
, reg_offset
= low
;
8092 bias
= reg_offset
, reg_offset
= 0;
8094 sa_reg
= gen_rtx_REG (DImode
, 22);
8095 sa_reg_exp
= plus_constant (stack_pointer_rtx
, bias
);
8097 FRP (emit_move_insn (sa_reg
, sa_reg_exp
));
8100 /* Restore registers in order, excepting a true frame pointer. */
8102 mem
= gen_rtx_MEM (DImode
, plus_constant (sa_reg
, reg_offset
));
8104 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8105 FRP (emit_move_insn (gen_rtx_REG (DImode
, REG_RA
), mem
));
8108 imask
&= ~(1UL << REG_RA
);
8110 for (i
= 0; i
< 31; ++i
)
8111 if (imask
& (1UL << i
))
8113 if (i
== HARD_FRAME_POINTER_REGNUM
&& fp_is_frame_pointer
)
8114 fp_offset
= reg_offset
;
8117 mem
= gen_rtx_MEM (DImode
, plus_constant(sa_reg
, reg_offset
));
8118 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8119 FRP (emit_move_insn (gen_rtx_REG (DImode
, i
), mem
));
8124 for (i
= 0; i
< 31; ++i
)
8125 if (fmask
& (1UL << i
))
8127 mem
= gen_rtx_MEM (DFmode
, plus_constant(sa_reg
, reg_offset
));
8128 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8129 FRP (emit_move_insn (gen_rtx_REG (DFmode
, i
+32), mem
));
8133 else if (TARGET_ABI_UNICOSMK
&& alpha_procedure_type
== PT_STACK
)
8135 /* Restore callee-saved general-purpose registers. */
8139 for (i
= 9; i
< 15; i
++)
8140 if (imask
& (1UL << i
))
8142 mem
= gen_rtx_MEM (DImode
, plus_constant(hard_frame_pointer_rtx
,
8144 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8145 FRP (emit_move_insn (gen_rtx_REG (DImode
, i
), mem
));
8149 for (i
= 2; i
< 10; i
++)
8150 if (fmask
& (1UL << i
))
8152 mem
= gen_rtx_MEM (DFmode
, plus_constant(hard_frame_pointer_rtx
,
8154 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8155 FRP (emit_move_insn (gen_rtx_REG (DFmode
, i
+32), mem
));
8159 /* Restore the return address from the DSIB. */
8161 mem
= gen_rtx_MEM (DImode
, plus_constant(hard_frame_pointer_rtx
, -8));
8162 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8163 FRP (emit_move_insn (gen_rtx_REG (DImode
, REG_RA
), mem
));
8166 if (frame_size
|| eh_ofs
)
8168 sp_adj1
= stack_pointer_rtx
;
8172 sp_adj1
= gen_rtx_REG (DImode
, 23);
8173 emit_move_insn (sp_adj1
,
8174 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
, eh_ofs
));
8177 /* If the stack size is large, begin computation into a temporary
8178 register so as not to interfere with a potential fp restore,
8179 which must be consecutive with an SP restore. */
8180 if (frame_size
< 32768
8181 && ! (TARGET_ABI_UNICOSMK
&& cfun
->calls_alloca
))
8182 sp_adj2
= GEN_INT (frame_size
);
8183 else if (TARGET_ABI_UNICOSMK
)
8185 sp_adj1
= gen_rtx_REG (DImode
, 23);
8186 FRP (emit_move_insn (sp_adj1
, hard_frame_pointer_rtx
));
8187 sp_adj2
= const0_rtx
;
8189 else if (frame_size
< 0x40007fffL
)
8191 int low
= ((frame_size
& 0xffff) ^ 0x8000) - 0x8000;
8193 sp_adj2
= plus_constant (sp_adj1
, frame_size
- low
);
8194 if (sa_reg_exp
&& rtx_equal_p (sa_reg_exp
, sp_adj2
))
8198 sp_adj1
= gen_rtx_REG (DImode
, 23);
8199 FRP (emit_move_insn (sp_adj1
, sp_adj2
));
8201 sp_adj2
= GEN_INT (low
);
8205 rtx tmp
= gen_rtx_REG (DImode
, 23);
8206 FRP (sp_adj2
= alpha_emit_set_const (tmp
, DImode
, frame_size
,
8210 /* We can't drop new things to memory this late, afaik,
8211 so build it up by pieces. */
8212 FRP (sp_adj2
= alpha_emit_set_long_const (tmp
, frame_size
,
8213 -(frame_size
< 0)));
8214 gcc_assert (sp_adj2
);
8218 /* From now on, things must be in order. So emit blockages. */
8220 /* Restore the frame pointer. */
8221 if (TARGET_ABI_UNICOSMK
)
8223 emit_insn (gen_blockage ());
8224 mem
= gen_rtx_MEM (DImode
,
8225 plus_constant (hard_frame_pointer_rtx
, -16));
8226 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8227 FRP (emit_move_insn (hard_frame_pointer_rtx
, mem
));
8229 else if (fp_is_frame_pointer
)
8231 emit_insn (gen_blockage ());
8232 mem
= gen_rtx_MEM (DImode
, plus_constant (sa_reg
, fp_offset
));
8233 set_mem_alias_set (mem
, alpha_sr_alias_set
);
8234 FRP (emit_move_insn (hard_frame_pointer_rtx
, mem
));
8236 else if (TARGET_ABI_OPEN_VMS
)
8238 emit_insn (gen_blockage ());
8239 FRP (emit_move_insn (hard_frame_pointer_rtx
,
8240 gen_rtx_REG (DImode
, vms_save_fp_regno
)));
8243 /* Restore the stack pointer. */
8244 emit_insn (gen_blockage ());
8245 if (sp_adj2
== const0_rtx
)
8246 FRP (emit_move_insn (stack_pointer_rtx
, sp_adj1
));
8248 FRP (emit_move_insn (stack_pointer_rtx
,
8249 gen_rtx_PLUS (DImode
, sp_adj1
, sp_adj2
)));
8253 if (TARGET_ABI_OPEN_VMS
&& alpha_procedure_type
== PT_REGISTER
)
8255 emit_insn (gen_blockage ());
8256 FRP (emit_move_insn (hard_frame_pointer_rtx
,
8257 gen_rtx_REG (DImode
, vms_save_fp_regno
)));
8259 else if (TARGET_ABI_UNICOSMK
&& alpha_procedure_type
!= PT_STACK
)
8261 /* Decrement the frame pointer if the function does not have a
8264 emit_insn (gen_blockage ());
8265 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx
,
8266 hard_frame_pointer_rtx
, constm1_rtx
)));
8271 /* Output the rest of the textual info surrounding the epilogue. */
8274 alpha_end_function (FILE *file
, const char *fnname
, tree decl ATTRIBUTE_UNUSED
)
8278 /* We output a nop after noreturn calls at the very end of the function to
8279 ensure that the return address always remains in the caller's code range,
8280 as not doing so might confuse unwinding engines. */
8281 insn
= get_last_insn ();
8283 insn
= prev_active_insn (insn
);
8285 output_asm_insn (get_insn_template (CODE_FOR_nop
, NULL
), NULL
);
8289 free_after_compilation (cfun
);
8292 #if TARGET_ABI_OPEN_VMS
8293 alpha_write_linkage (file
, fnname
, decl
);
8296 /* End the function. */
8297 if (!TARGET_ABI_UNICOSMK
&& !flag_inhibit_size_directive
)
8299 fputs ("\t.end ", file
);
8300 assemble_name (file
, fnname
);
8303 inside_function
= FALSE
;
8305 /* Output jump tables and the static subroutine information block. */
8306 if (TARGET_ABI_UNICOSMK
)
8308 unicosmk_output_ssib (file
, fnname
);
8309 unicosmk_output_deferred_case_vectors (file
);
8314 /* Emit a tail call to FUNCTION after adjusting THIS by DELTA.
8316 In order to avoid the hordes of differences between generated code
8317 with and without TARGET_EXPLICIT_RELOCS, and to avoid duplicating
8318 lots of code loading up large constants, generate rtl and emit it
8319 instead of going straight to text.
8321 Not sure why this idea hasn't been explored before... */
8324 alpha_output_mi_thunk_osf (FILE *file
, tree thunk_fndecl ATTRIBUTE_UNUSED
,
8325 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
8328 HOST_WIDE_INT hi
, lo
;
8329 rtx this_rtx
, insn
, funexp
;
8331 gcc_assert (cfun
->is_thunk
);
8333 /* We always require a valid GP. */
8334 emit_insn (gen_prologue_ldgp ());
8335 emit_note (NOTE_INSN_PROLOGUE_END
);
8337 /* Find the "this" pointer. If the function returns a structure,
8338 the structure return pointer is in $16. */
8339 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
8340 this_rtx
= gen_rtx_REG (Pmode
, 17);
8342 this_rtx
= gen_rtx_REG (Pmode
, 16);
8344 /* Add DELTA. When possible we use ldah+lda. Otherwise load the
8345 entire constant for the add. */
8346 lo
= ((delta
& 0xffff) ^ 0x8000) - 0x8000;
8347 hi
= (((delta
- lo
) & 0xffffffff) ^ 0x80000000) - 0x80000000;
8348 if (hi
+ lo
== delta
)
8351 emit_insn (gen_adddi3 (this_rtx
, this_rtx
, GEN_INT (hi
)));
8353 emit_insn (gen_adddi3 (this_rtx
, this_rtx
, GEN_INT (lo
)));
8357 rtx tmp
= alpha_emit_set_long_const (gen_rtx_REG (Pmode
, 0),
8358 delta
, -(delta
< 0));
8359 emit_insn (gen_adddi3 (this_rtx
, this_rtx
, tmp
));
8362 /* Add a delta stored in the vtable at VCALL_OFFSET. */
8367 tmp
= gen_rtx_REG (Pmode
, 0);
8368 emit_move_insn (tmp
, gen_rtx_MEM (Pmode
, this_rtx
));
8370 lo
= ((vcall_offset
& 0xffff) ^ 0x8000) - 0x8000;
8371 hi
= (((vcall_offset
- lo
) & 0xffffffff) ^ 0x80000000) - 0x80000000;
8372 if (hi
+ lo
== vcall_offset
)
8375 emit_insn (gen_adddi3 (tmp
, tmp
, GEN_INT (hi
)));
8379 tmp2
= alpha_emit_set_long_const (gen_rtx_REG (Pmode
, 1),
8380 vcall_offset
, -(vcall_offset
< 0));
8381 emit_insn (gen_adddi3 (tmp
, tmp
, tmp2
));
8385 tmp2
= gen_rtx_PLUS (Pmode
, tmp
, GEN_INT (lo
));
8388 emit_move_insn (tmp
, gen_rtx_MEM (Pmode
, tmp2
));
8390 emit_insn (gen_adddi3 (this_rtx
, this_rtx
, tmp
));
8393 /* Generate a tail call to the target function. */
8394 if (! TREE_USED (function
))
8396 assemble_external (function
);
8397 TREE_USED (function
) = 1;
8399 funexp
= XEXP (DECL_RTL (function
), 0);
8400 funexp
= gen_rtx_MEM (FUNCTION_MODE
, funexp
);
8401 insn
= emit_call_insn (gen_sibcall (funexp
, const0_rtx
));
8402 SIBLING_CALL_P (insn
) = 1;
8404 /* Run just enough of rest_of_compilation to get the insns emitted.
8405 There's not really enough bulk here to make other passes such as
8406 instruction scheduling worth while. Note that use_thunk calls
8407 assemble_start_function and assemble_end_function. */
8408 insn
= get_insns ();
8409 insn_locators_alloc ();
8410 shorten_branches (insn
);
8411 final_start_function (insn
, file
, 1);
8412 final (insn
, file
, 1);
8413 final_end_function ();
8415 #endif /* TARGET_ABI_OSF */
8417 /* Debugging support. */
8421 /* Count the number of sdb related labels are generated (to find block
8422 start and end boundaries). */
8424 int sdb_label_count
= 0;
8426 /* Name of the file containing the current function. */
8428 static const char *current_function_file
= "";
8430 /* Offsets to alpha virtual arg/local debugging pointers. */
8432 long alpha_arg_offset
;
8433 long alpha_auto_offset
;
8435 /* Emit a new filename to a stream. */
8438 alpha_output_filename (FILE *stream
, const char *name
)
8440 static int first_time
= TRUE
;
8445 ++num_source_filenames
;
8446 current_function_file
= name
;
8447 fprintf (stream
, "\t.file\t%d ", num_source_filenames
);
8448 output_quoted_string (stream
, name
);
8449 fprintf (stream
, "\n");
8450 if (!TARGET_GAS
&& write_symbols
== DBX_DEBUG
)
8451 fprintf (stream
, "\t#@stabs\n");
8454 else if (write_symbols
== DBX_DEBUG
)
8455 /* dbxout.c will emit an appropriate .stabs directive. */
8458 else if (name
!= current_function_file
8459 && strcmp (name
, current_function_file
) != 0)
8461 if (inside_function
&& ! TARGET_GAS
)
8462 fprintf (stream
, "\t#.file\t%d ", num_source_filenames
);
8465 ++num_source_filenames
;
8466 current_function_file
= name
;
8467 fprintf (stream
, "\t.file\t%d ", num_source_filenames
);
8470 output_quoted_string (stream
, name
);
8471 fprintf (stream
, "\n");
8475 /* Structure to show the current status of registers and memory. */
8477 struct shadow_summary
8480 unsigned int i
: 31; /* Mask of int regs */
8481 unsigned int fp
: 31; /* Mask of fp regs */
8482 unsigned int mem
: 1; /* mem == imem | fpmem */
8486 /* Summary the effects of expression X on the machine. Update SUM, a pointer
8487 to the summary structure. SET is nonzero if the insn is setting the
8488 object, otherwise zero. */
8491 summarize_insn (rtx x
, struct shadow_summary
*sum
, int set
)
8493 const char *format_ptr
;
8499 switch (GET_CODE (x
))
8501 /* ??? Note that this case would be incorrect if the Alpha had a
8502 ZERO_EXTRACT in SET_DEST. */
8504 summarize_insn (SET_SRC (x
), sum
, 0);
8505 summarize_insn (SET_DEST (x
), sum
, 1);
8509 summarize_insn (XEXP (x
, 0), sum
, 1);
8513 summarize_insn (XEXP (x
, 0), sum
, 0);
8517 for (i
= ASM_OPERANDS_INPUT_LENGTH (x
) - 1; i
>= 0; i
--)
8518 summarize_insn (ASM_OPERANDS_INPUT (x
, i
), sum
, 0);
8522 for (i
= XVECLEN (x
, 0) - 1; i
>= 0; i
--)
8523 summarize_insn (XVECEXP (x
, 0, i
), sum
, 0);
8527 summarize_insn (SUBREG_REG (x
), sum
, 0);
8532 int regno
= REGNO (x
);
8533 unsigned long mask
= ((unsigned long) 1) << (regno
% 32);
8535 if (regno
== 31 || regno
== 63)
8541 sum
->defd
.i
|= mask
;
8543 sum
->defd
.fp
|= mask
;
8548 sum
->used
.i
|= mask
;
8550 sum
->used
.fp
|= mask
;
8561 /* Find the regs used in memory address computation: */
8562 summarize_insn (XEXP (x
, 0), sum
, 0);
8565 case CONST_INT
: case CONST_DOUBLE
:
8566 case SYMBOL_REF
: case LABEL_REF
: case CONST
:
8567 case SCRATCH
: case ASM_INPUT
:
8570 /* Handle common unary and binary ops for efficiency. */
8571 case COMPARE
: case PLUS
: case MINUS
: case MULT
: case DIV
:
8572 case MOD
: case UDIV
: case UMOD
: case AND
: case IOR
:
8573 case XOR
: case ASHIFT
: case ROTATE
: case ASHIFTRT
: case LSHIFTRT
:
8574 case ROTATERT
: case SMIN
: case SMAX
: case UMIN
: case UMAX
:
8575 case NE
: case EQ
: case GE
: case GT
: case LE
:
8576 case LT
: case GEU
: case GTU
: case LEU
: case LTU
:
8577 summarize_insn (XEXP (x
, 0), sum
, 0);
8578 summarize_insn (XEXP (x
, 1), sum
, 0);
8581 case NEG
: case NOT
: case SIGN_EXTEND
: case ZERO_EXTEND
:
8582 case TRUNCATE
: case FLOAT_EXTEND
: case FLOAT_TRUNCATE
: case FLOAT
:
8583 case FIX
: case UNSIGNED_FLOAT
: case UNSIGNED_FIX
: case ABS
:
8584 case SQRT
: case FFS
:
8585 summarize_insn (XEXP (x
, 0), sum
, 0);
8589 format_ptr
= GET_RTX_FORMAT (GET_CODE (x
));
8590 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
8591 switch (format_ptr
[i
])
8594 summarize_insn (XEXP (x
, i
), sum
, 0);
8598 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
8599 summarize_insn (XVECEXP (x
, i
, j
), sum
, 0);
8611 /* Ensure a sufficient number of `trapb' insns are in the code when
8612 the user requests code with a trap precision of functions or
8615 In naive mode, when the user requests a trap-precision of
8616 "instruction", a trapb is needed after every instruction that may
8617 generate a trap. This ensures that the code is resumption safe but
8620 When optimizations are turned on, we delay issuing a trapb as long
8621 as possible. In this context, a trap shadow is the sequence of
8622 instructions that starts with a (potentially) trap generating
8623 instruction and extends to the next trapb or call_pal instruction
8624 (but GCC never generates call_pal by itself). We can delay (and
8625 therefore sometimes omit) a trapb subject to the following
8628 (a) On entry to the trap shadow, if any Alpha register or memory
8629 location contains a value that is used as an operand value by some
8630 instruction in the trap shadow (live on entry), then no instruction
8631 in the trap shadow may modify the register or memory location.
8633 (b) Within the trap shadow, the computation of the base register
8634 for a memory load or store instruction may not involve using the
8635 result of an instruction that might generate an UNPREDICTABLE
8638 (c) Within the trap shadow, no register may be used more than once
8639 as a destination register. (This is to make life easier for the
8642 (d) The trap shadow may not include any branch instructions. */
8645 alpha_handle_trap_shadows (void)
8647 struct shadow_summary shadow
;
8648 int trap_pending
, exception_nesting
;
8652 exception_nesting
= 0;
8655 shadow
.used
.mem
= 0;
8656 shadow
.defd
= shadow
.used
;
8658 for (i
= get_insns (); i
; i
= NEXT_INSN (i
))
8662 switch (NOTE_KIND (i
))
8664 case NOTE_INSN_EH_REGION_BEG
:
8665 exception_nesting
++;
8670 case NOTE_INSN_EH_REGION_END
:
8671 exception_nesting
--;
8676 case NOTE_INSN_EPILOGUE_BEG
:
8677 if (trap_pending
&& alpha_tp
>= ALPHA_TP_FUNC
)
8682 else if (trap_pending
)
8684 if (alpha_tp
== ALPHA_TP_FUNC
)
8687 && GET_CODE (PATTERN (i
)) == RETURN
)
8690 else if (alpha_tp
== ALPHA_TP_INSN
)
8694 struct shadow_summary sum
;
8699 sum
.defd
= sum
.used
;
8701 switch (GET_CODE (i
))
8704 /* Annoyingly, get_attr_trap will die on these. */
8705 if (GET_CODE (PATTERN (i
)) == USE
8706 || GET_CODE (PATTERN (i
)) == CLOBBER
)
8709 summarize_insn (PATTERN (i
), &sum
, 0);
8711 if ((sum
.defd
.i
& shadow
.defd
.i
)
8712 || (sum
.defd
.fp
& shadow
.defd
.fp
))
8714 /* (c) would be violated */
8718 /* Combine shadow with summary of current insn: */
8719 shadow
.used
.i
|= sum
.used
.i
;
8720 shadow
.used
.fp
|= sum
.used
.fp
;
8721 shadow
.used
.mem
|= sum
.used
.mem
;
8722 shadow
.defd
.i
|= sum
.defd
.i
;
8723 shadow
.defd
.fp
|= sum
.defd
.fp
;
8724 shadow
.defd
.mem
|= sum
.defd
.mem
;
8726 if ((sum
.defd
.i
& shadow
.used
.i
)
8727 || (sum
.defd
.fp
& shadow
.used
.fp
)
8728 || (sum
.defd
.mem
& shadow
.used
.mem
))
8730 /* (a) would be violated (also takes care of (b)) */
8731 gcc_assert (get_attr_trap (i
) != TRAP_YES
8732 || (!(sum
.defd
.i
& sum
.used
.i
)
8733 && !(sum
.defd
.fp
& sum
.used
.fp
)));
8751 n
= emit_insn_before (gen_trapb (), i
);
8752 PUT_MODE (n
, TImode
);
8753 PUT_MODE (i
, TImode
);
8757 shadow
.used
.mem
= 0;
8758 shadow
.defd
= shadow
.used
;
8763 if ((exception_nesting
> 0 || alpha_tp
>= ALPHA_TP_FUNC
)
8764 && NONJUMP_INSN_P (i
)
8765 && GET_CODE (PATTERN (i
)) != USE
8766 && GET_CODE (PATTERN (i
)) != CLOBBER
8767 && get_attr_trap (i
) == TRAP_YES
)
8769 if (optimize
&& !trap_pending
)
8770 summarize_insn (PATTERN (i
), &shadow
, 0);
8776 /* Alpha can only issue instruction groups simultaneously if they are
8777 suitably aligned. This is very processor-specific. */
8778 /* There are a number of entries in alphaev4_insn_pipe and alphaev5_insn_pipe
8779 that are marked "fake". These instructions do not exist on that target,
8780 but it is possible to see these insns with deranged combinations of
8781 command-line options, such as "-mtune=ev4 -mmax". Instead of aborting,
8782 choose a result at random. */
8784 enum alphaev4_pipe
{
8791 enum alphaev5_pipe
{
8802 static enum alphaev4_pipe
8803 alphaev4_insn_pipe (rtx insn
)
8805 if (recog_memoized (insn
) < 0)
8807 if (get_attr_length (insn
) != 4)
8810 switch (get_attr_type (insn
))
8826 case TYPE_MVI
: /* fake */
8841 case TYPE_FSQRT
: /* fake */
8842 case TYPE_FTOI
: /* fake */
8843 case TYPE_ITOF
: /* fake */
8851 static enum alphaev5_pipe
8852 alphaev5_insn_pipe (rtx insn
)
8854 if (recog_memoized (insn
) < 0)
8856 if (get_attr_length (insn
) != 4)
8859 switch (get_attr_type (insn
))
8879 case TYPE_FTOI
: /* fake */
8880 case TYPE_ITOF
: /* fake */
8895 case TYPE_FSQRT
: /* fake */
8906 /* IN_USE is a mask of the slots currently filled within the insn group.
8907 The mask bits come from alphaev4_pipe above. If EV4_IBX is set, then
8908 the insn in EV4_IB0 can be swapped by the hardware into EV4_IB1.
8910 LEN is, of course, the length of the group in bytes. */
8913 alphaev4_next_group (rtx insn
, int *pin_use
, int *plen
)
8920 || GET_CODE (PATTERN (insn
)) == CLOBBER
8921 || GET_CODE (PATTERN (insn
)) == USE
)
8926 enum alphaev4_pipe pipe
;
8928 pipe
= alphaev4_insn_pipe (insn
);
8932 /* Force complex instructions to start new groups. */
8936 /* If this is a completely unrecognized insn, it's an asm.
8937 We don't know how long it is, so record length as -1 to
8938 signal a needed realignment. */
8939 if (recog_memoized (insn
) < 0)
8942 len
= get_attr_length (insn
);
8946 if (in_use
& EV4_IB0
)
8948 if (in_use
& EV4_IB1
)
8953 in_use
|= EV4_IB0
| EV4_IBX
;
8957 if (in_use
& EV4_IB0
)
8959 if (!(in_use
& EV4_IBX
) || (in_use
& EV4_IB1
))
8967 if (in_use
& EV4_IB1
)
8977 /* Haifa doesn't do well scheduling branches. */
8982 insn
= next_nonnote_insn (insn
);
8984 if (!insn
|| ! INSN_P (insn
))
8987 /* Let Haifa tell us where it thinks insn group boundaries are. */
8988 if (GET_MODE (insn
) == TImode
)
8991 if (GET_CODE (insn
) == CLOBBER
|| GET_CODE (insn
) == USE
)
8996 insn
= next_nonnote_insn (insn
);
9004 /* IN_USE is a mask of the slots currently filled within the insn group.
9005 The mask bits come from alphaev5_pipe above. If EV5_E01 is set, then
9006 the insn in EV5_E0 can be swapped by the hardware into EV5_E1.
9008 LEN is, of course, the length of the group in bytes. */
9011 alphaev5_next_group (rtx insn
, int *pin_use
, int *plen
)
9018 || GET_CODE (PATTERN (insn
)) == CLOBBER
9019 || GET_CODE (PATTERN (insn
)) == USE
)
9024 enum alphaev5_pipe pipe
;
9026 pipe
= alphaev5_insn_pipe (insn
);
9030 /* Force complex instructions to start new groups. */
9034 /* If this is a completely unrecognized insn, it's an asm.
9035 We don't know how long it is, so record length as -1 to
9036 signal a needed realignment. */
9037 if (recog_memoized (insn
) < 0)
9040 len
= get_attr_length (insn
);
9043 /* ??? Most of the places below, we would like to assert never
9044 happen, as it would indicate an error either in Haifa, or
9045 in the scheduling description. Unfortunately, Haifa never
9046 schedules the last instruction of the BB, so we don't have
9047 an accurate TI bit to go off. */
9049 if (in_use
& EV5_E0
)
9051 if (in_use
& EV5_E1
)
9056 in_use
|= EV5_E0
| EV5_E01
;
9060 if (in_use
& EV5_E0
)
9062 if (!(in_use
& EV5_E01
) || (in_use
& EV5_E1
))
9070 if (in_use
& EV5_E1
)
9076 if (in_use
& EV5_FA
)
9078 if (in_use
& EV5_FM
)
9083 in_use
|= EV5_FA
| EV5_FAM
;
9087 if (in_use
& EV5_FA
)
9093 if (in_use
& EV5_FM
)
9106 /* Haifa doesn't do well scheduling branches. */
9107 /* ??? If this is predicted not-taken, slotting continues, except
9108 that no more IBR, FBR, or JSR insns may be slotted. */
9113 insn
= next_nonnote_insn (insn
);
9115 if (!insn
|| ! INSN_P (insn
))
9118 /* Let Haifa tell us where it thinks insn group boundaries are. */
9119 if (GET_MODE (insn
) == TImode
)
9122 if (GET_CODE (insn
) == CLOBBER
|| GET_CODE (insn
) == USE
)
9127 insn
= next_nonnote_insn (insn
);
9136 alphaev4_next_nop (int *pin_use
)
9138 int in_use
= *pin_use
;
9141 if (!(in_use
& EV4_IB0
))
9146 else if ((in_use
& (EV4_IBX
|EV4_IB1
)) == EV4_IBX
)
9151 else if (TARGET_FP
&& !(in_use
& EV4_IB1
))
9164 alphaev5_next_nop (int *pin_use
)
9166 int in_use
= *pin_use
;
9169 if (!(in_use
& EV5_E1
))
9174 else if (TARGET_FP
&& !(in_use
& EV5_FA
))
9179 else if (TARGET_FP
&& !(in_use
& EV5_FM
))
9191 /* The instruction group alignment main loop. */
9194 alpha_align_insns (unsigned int max_align
,
9195 rtx (*next_group
) (rtx
, int *, int *),
9196 rtx (*next_nop
) (int *))
9198 /* ALIGN is the known alignment for the insn group. */
9200 /* OFS is the offset of the current insn in the insn group. */
9202 int prev_in_use
, in_use
, len
, ldgp
;
9205 /* Let shorten branches care for assigning alignments to code labels. */
9206 shorten_branches (get_insns ());
9208 if (align_functions
< 4)
9210 else if ((unsigned int) align_functions
< max_align
)
9211 align
= align_functions
;
9215 ofs
= prev_in_use
= 0;
9218 i
= next_nonnote_insn (i
);
9220 ldgp
= alpha_function_needs_gp
? 8 : 0;
9224 next
= (*next_group
) (i
, &in_use
, &len
);
9226 /* When we see a label, resync alignment etc. */
9229 unsigned int new_align
= 1 << label_to_alignment (i
);
9231 if (new_align
>= align
)
9233 align
= new_align
< max_align
? new_align
: max_align
;
9237 else if (ofs
& (new_align
-1))
9238 ofs
= (ofs
| (new_align
-1)) + 1;
9242 /* Handle complex instructions special. */
9243 else if (in_use
== 0)
9245 /* Asms will have length < 0. This is a signal that we have
9246 lost alignment knowledge. Assume, however, that the asm
9247 will not mis-align instructions. */
9256 /* If the known alignment is smaller than the recognized insn group,
9257 realign the output. */
9258 else if ((int) align
< len
)
9260 unsigned int new_log_align
= len
> 8 ? 4 : 3;
9263 where
= prev
= prev_nonnote_insn (i
);
9264 if (!where
|| !LABEL_P (where
))
9267 /* Can't realign between a call and its gp reload. */
9268 if (! (TARGET_EXPLICIT_RELOCS
9269 && prev
&& CALL_P (prev
)))
9271 emit_insn_before (gen_realign (GEN_INT (new_log_align
)), where
);
9272 align
= 1 << new_log_align
;
9277 /* We may not insert padding inside the initial ldgp sequence. */
9281 /* If the group won't fit in the same INT16 as the previous,
9282 we need to add padding to keep the group together. Rather
9283 than simply leaving the insn filling to the assembler, we
9284 can make use of the knowledge of what sorts of instructions
9285 were issued in the previous group to make sure that all of
9286 the added nops are really free. */
9287 else if (ofs
+ len
> (int) align
)
9289 int nop_count
= (align
- ofs
) / 4;
9292 /* Insert nops before labels, branches, and calls to truly merge
9293 the execution of the nops with the previous instruction group. */
9294 where
= prev_nonnote_insn (i
);
9297 if (LABEL_P (where
))
9299 rtx where2
= prev_nonnote_insn (where
);
9300 if (where2
&& JUMP_P (where2
))
9303 else if (NONJUMP_INSN_P (where
))
9310 emit_insn_before ((*next_nop
)(&prev_in_use
), where
);
9311 while (--nop_count
);
9315 ofs
= (ofs
+ len
) & (align
- 1);
9316 prev_in_use
= in_use
;
9321 /* Insert an unop between a noreturn function call and GP load. */
9324 alpha_pad_noreturn (void)
9328 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
9331 || !find_reg_note (insn
, REG_NORETURN
, NULL_RTX
))
9334 next
= next_active_insn (insn
);
9338 rtx pat
= PATTERN (next
);
9340 if (GET_CODE (pat
) == SET
9341 && GET_CODE (SET_SRC (pat
)) == UNSPEC_VOLATILE
9342 && XINT (SET_SRC (pat
), 1) == UNSPECV_LDGP1
)
9343 emit_insn_after (gen_unop (), insn
);
9348 /* Machine dependent reorg pass. */
9353 /* Workaround for a linker error that triggers when an
9354 exception handler immediatelly follows a noreturn function.
9356 The instruction stream from an object file:
9358 54: 00 40 5b 6b jsr ra,(t12),58 <__func+0x58>
9359 58: 00 00 ba 27 ldah gp,0(ra)
9360 5c: 00 00 bd 23 lda gp,0(gp)
9361 60: 00 00 7d a7 ldq t12,0(gp)
9362 64: 00 40 5b 6b jsr ra,(t12),68 <__func+0x68>
9364 was converted in the final link pass to:
9366 fdb24: a0 03 40 d3 bsr ra,fe9a8 <_called_func+0x8>
9367 fdb28: 00 00 fe 2f unop
9368 fdb2c: 00 00 fe 2f unop
9369 fdb30: 30 82 7d a7 ldq t12,-32208(gp)
9370 fdb34: 00 40 5b 6b jsr ra,(t12),fdb38 <__func+0x68>
9372 GP load instructions were wrongly cleared by the linker relaxation
9373 pass. This workaround prevents removal of GP loads by inserting
9374 an unop instruction between a noreturn function call and
9375 exception handler prologue. */
9377 if (current_function_has_exception_handlers ())
9378 alpha_pad_noreturn ();
9380 if (alpha_tp
!= ALPHA_TP_PROG
|| flag_exceptions
)
9381 alpha_handle_trap_shadows ();
9383 /* Due to the number of extra trapb insns, don't bother fixing up
9384 alignment when trap precision is instruction. Moreover, we can
9385 only do our job when sched2 is run. */
9386 if (optimize
&& !optimize_size
9387 && alpha_tp
!= ALPHA_TP_INSN
9388 && flag_schedule_insns_after_reload
)
9390 if (alpha_tune
== PROCESSOR_EV4
)
9391 alpha_align_insns (8, alphaev4_next_group
, alphaev4_next_nop
);
9392 else if (alpha_tune
== PROCESSOR_EV5
)
9393 alpha_align_insns (16, alphaev5_next_group
, alphaev5_next_nop
);
9397 #if !TARGET_ABI_UNICOSMK
9404 alpha_file_start (void)
9406 #ifdef OBJECT_FORMAT_ELF
9407 /* If emitting dwarf2 debug information, we cannot generate a .file
9408 directive to start the file, as it will conflict with dwarf2out
9409 file numbers. So it's only useful when emitting mdebug output. */
9410 targetm
.file_start_file_directive
= (write_symbols
== DBX_DEBUG
);
9413 default_file_start ();
9415 fprintf (asm_out_file
, "\t.verstamp %d %d\n", MS_STAMP
, LS_STAMP
);
9418 fputs ("\t.set noreorder\n", asm_out_file
);
9419 fputs ("\t.set volatile\n", asm_out_file
);
9420 if (!TARGET_ABI_OPEN_VMS
)
9421 fputs ("\t.set noat\n", asm_out_file
);
9422 if (TARGET_EXPLICIT_RELOCS
)
9423 fputs ("\t.set nomacro\n", asm_out_file
);
9424 if (TARGET_SUPPORT_ARCH
| TARGET_BWX
| TARGET_MAX
| TARGET_FIX
| TARGET_CIX
)
9428 if (alpha_cpu
== PROCESSOR_EV6
|| TARGET_FIX
|| TARGET_CIX
)
9430 else if (TARGET_MAX
)
9432 else if (TARGET_BWX
)
9434 else if (alpha_cpu
== PROCESSOR_EV5
)
9439 fprintf (asm_out_file
, "\t.arch %s\n", arch
);
9444 #ifdef OBJECT_FORMAT_ELF
9445 /* Since we don't have a .dynbss section, we should not allow global
9446 relocations in the .rodata section. */
9449 alpha_elf_reloc_rw_mask (void)
9451 return flag_pic
? 3 : 2;
9454 /* Return a section for X. The only special thing we do here is to
9455 honor small data. */
9458 alpha_elf_select_rtx_section (enum machine_mode mode
, rtx x
,
9459 unsigned HOST_WIDE_INT align
)
9461 if (TARGET_SMALL_DATA
&& GET_MODE_SIZE (mode
) <= g_switch_value
)
9462 /* ??? Consider using mergeable sdata sections. */
9463 return sdata_section
;
9465 return default_elf_select_rtx_section (mode
, x
, align
);
9469 alpha_elf_section_type_flags (tree decl
, const char *name
, int reloc
)
9471 unsigned int flags
= 0;
9473 if (strcmp (name
, ".sdata") == 0
9474 || strncmp (name
, ".sdata.", 7) == 0
9475 || strncmp (name
, ".gnu.linkonce.s.", 16) == 0
9476 || strcmp (name
, ".sbss") == 0
9477 || strncmp (name
, ".sbss.", 6) == 0
9478 || strncmp (name
, ".gnu.linkonce.sb.", 17) == 0)
9479 flags
= SECTION_SMALL
;
9481 flags
|= default_section_type_flags (decl
, name
, reloc
);
9484 #endif /* OBJECT_FORMAT_ELF */
9486 /* Structure to collect function names for final output in link section. */
9487 /* Note that items marked with GTY can't be ifdef'ed out. */
9489 enum links_kind
{KIND_UNUSED
, KIND_LOCAL
, KIND_EXTERN
};
9490 enum reloc_kind
{KIND_LINKAGE
, KIND_CODEADDR
};
9492 struct GTY(()) alpha_links
9496 enum links_kind lkind
;
9497 enum reloc_kind rkind
;
9500 struct GTY(()) alpha_funcs
9503 splay_tree
GTY ((param1_is (char *), param2_is (struct alpha_links
*)))
9507 static GTY ((param1_is (char *), param2_is (struct alpha_links
*)))
9508 splay_tree alpha_links_tree
;
9509 static GTY ((param1_is (tree
), param2_is (struct alpha_funcs
*)))
9510 splay_tree alpha_funcs_tree
;
9512 static GTY(()) int alpha_funcs_num
;
9514 #if TARGET_ABI_OPEN_VMS
9516 /* Return the VMS argument type corresponding to MODE. */
9519 alpha_arg_type (enum machine_mode mode
)
9524 return TARGET_FLOAT_VAX
? FF
: FS
;
9526 return TARGET_FLOAT_VAX
? FD
: FT
;
9532 /* Return an rtx for an integer representing the VMS Argument Information
9536 alpha_arg_info_reg_val (CUMULATIVE_ARGS cum
)
9538 unsigned HOST_WIDE_INT regval
= cum
.num_args
;
9541 for (i
= 0; i
< 6; i
++)
9542 regval
|= ((int) cum
.atypes
[i
]) << (i
* 3 + 8);
9544 return GEN_INT (regval
);
9547 /* Make (or fake) .linkage entry for function call.
9549 IS_LOCAL is 0 if name is used in call, 1 if name is used in definition.
9551 Return an SYMBOL_REF rtx for the linkage. */
9554 alpha_need_linkage (const char *name
, int is_local
)
9556 splay_tree_node node
;
9557 struct alpha_links
*al
;
9564 struct alpha_funcs
*cfaf
;
9566 if (!alpha_funcs_tree
)
9567 alpha_funcs_tree
= splay_tree_new_ggc ((splay_tree_compare_fn
)
9568 splay_tree_compare_pointers
);
9570 cfaf
= (struct alpha_funcs
*) ggc_alloc (sizeof (struct alpha_funcs
));
9573 cfaf
->num
= ++alpha_funcs_num
;
9575 splay_tree_insert (alpha_funcs_tree
,
9576 (splay_tree_key
) current_function_decl
,
9577 (splay_tree_value
) cfaf
);
9580 if (alpha_links_tree
)
9582 /* Is this name already defined? */
9584 node
= splay_tree_lookup (alpha_links_tree
, (splay_tree_key
) name
);
9587 al
= (struct alpha_links
*) node
->value
;
9590 /* Defined here but external assumed. */
9591 if (al
->lkind
== KIND_EXTERN
)
9592 al
->lkind
= KIND_LOCAL
;
9596 /* Used here but unused assumed. */
9597 if (al
->lkind
== KIND_UNUSED
)
9598 al
->lkind
= KIND_LOCAL
;
9604 alpha_links_tree
= splay_tree_new_ggc ((splay_tree_compare_fn
) strcmp
);
9606 al
= (struct alpha_links
*) ggc_alloc (sizeof (struct alpha_links
));
9607 name
= ggc_strdup (name
);
9609 /* Assume external if no definition. */
9610 al
->lkind
= (is_local
? KIND_UNUSED
: KIND_EXTERN
);
9612 /* Ensure we have an IDENTIFIER so assemble_name can mark it used. */
9613 get_identifier (name
);
9615 /* Construct a SYMBOL_REF for us to call. */
9617 size_t name_len
= strlen (name
);
9618 char *linksym
= XALLOCAVEC (char, name_len
+ 6);
9620 memcpy (linksym
+ 1, name
, name_len
);
9621 memcpy (linksym
+ 1 + name_len
, "..lk", 5);
9622 al
->linkage
= gen_rtx_SYMBOL_REF (Pmode
,
9623 ggc_alloc_string (linksym
, name_len
+ 5));
9626 splay_tree_insert (alpha_links_tree
, (splay_tree_key
) name
,
9627 (splay_tree_value
) al
);
9633 alpha_use_linkage (rtx linkage
, tree cfundecl
, int lflag
, int rflag
)
9635 splay_tree_node cfunnode
;
9636 struct alpha_funcs
*cfaf
;
9637 struct alpha_links
*al
;
9638 const char *name
= XSTR (linkage
, 0);
9640 cfaf
= (struct alpha_funcs
*) 0;
9641 al
= (struct alpha_links
*) 0;
9643 cfunnode
= splay_tree_lookup (alpha_funcs_tree
, (splay_tree_key
) cfundecl
);
9644 cfaf
= (struct alpha_funcs
*) cfunnode
->value
;
9648 splay_tree_node lnode
;
9650 /* Is this name already defined? */
9652 lnode
= splay_tree_lookup (cfaf
->links
, (splay_tree_key
) name
);
9654 al
= (struct alpha_links
*) lnode
->value
;
9657 cfaf
->links
= splay_tree_new_ggc ((splay_tree_compare_fn
) strcmp
);
9665 splay_tree_node node
= 0;
9666 struct alpha_links
*anl
;
9671 name_len
= strlen (name
);
9673 al
= (struct alpha_links
*) ggc_alloc (sizeof (struct alpha_links
));
9674 al
->num
= cfaf
->num
;
9676 node
= splay_tree_lookup (alpha_links_tree
, (splay_tree_key
) name
);
9679 anl
= (struct alpha_links
*) node
->value
;
9680 al
->lkind
= anl
->lkind
;
9683 sprintf (buf
, "$%d..%s..lk", cfaf
->num
, name
);
9684 buflen
= strlen (buf
);
9685 linksym
= XALLOCAVEC (char, buflen
+ 1);
9686 memcpy (linksym
, buf
, buflen
+ 1);
9688 al
->linkage
= gen_rtx_SYMBOL_REF
9689 (Pmode
, ggc_alloc_string (linksym
, buflen
+ 1));
9691 splay_tree_insert (cfaf
->links
, (splay_tree_key
) name
,
9692 (splay_tree_value
) al
);
9696 al
->rkind
= KIND_CODEADDR
;
9698 al
->rkind
= KIND_LINKAGE
;
9701 return gen_rtx_MEM (Pmode
, plus_constant (al
->linkage
, 8));
9707 alpha_write_one_linkage (splay_tree_node node
, void *data
)
9709 const char *const name
= (const char *) node
->key
;
9710 struct alpha_links
*link
= (struct alpha_links
*) node
->value
;
9711 FILE *stream
= (FILE *) data
;
9713 fprintf (stream
, "$%d..%s..lk:\n", link
->num
, name
);
9714 if (link
->rkind
== KIND_CODEADDR
)
9716 if (link
->lkind
== KIND_LOCAL
)
9718 /* Local and used */
9719 fprintf (stream
, "\t.quad %s..en\n", name
);
9723 /* External and used, request code address. */
9724 fprintf (stream
, "\t.code_address %s\n", name
);
9729 if (link
->lkind
== KIND_LOCAL
)
9731 /* Local and used, build linkage pair. */
9732 fprintf (stream
, "\t.quad %s..en\n", name
);
9733 fprintf (stream
, "\t.quad %s\n", name
);
9737 /* External and used, request linkage pair. */
9738 fprintf (stream
, "\t.linkage %s\n", name
);
9746 alpha_write_linkage (FILE *stream
, const char *funname
, tree fundecl
)
9748 splay_tree_node node
;
9749 struct alpha_funcs
*func
;
9751 fprintf (stream
, "\t.link\n");
9752 fprintf (stream
, "\t.align 3\n");
9755 node
= splay_tree_lookup (alpha_funcs_tree
, (splay_tree_key
) fundecl
);
9756 func
= (struct alpha_funcs
*) node
->value
;
9758 fputs ("\t.name ", stream
);
9759 assemble_name (stream
, funname
);
9760 fputs ("..na\n", stream
);
9761 ASM_OUTPUT_LABEL (stream
, funname
);
9762 fprintf (stream
, "\t.pdesc ");
9763 assemble_name (stream
, funname
);
9764 fprintf (stream
, "..en,%s\n",
9765 alpha_procedure_type
== PT_STACK
? "stack"
9766 : alpha_procedure_type
== PT_REGISTER
? "reg" : "null");
9770 splay_tree_foreach (func
->links
, alpha_write_one_linkage
, stream
);
9771 /* splay_tree_delete (func->links); */
9775 /* Given a decl, a section name, and whether the decl initializer
9776 has relocs, choose attributes for the section. */
9778 #define SECTION_VMS_OVERLAY SECTION_FORGET
9779 #define SECTION_VMS_GLOBAL SECTION_MACH_DEP
9780 #define SECTION_VMS_INITIALIZE (SECTION_VMS_GLOBAL << 1)
9783 vms_section_type_flags (tree decl
, const char *name
, int reloc
)
9785 unsigned int flags
= default_section_type_flags (decl
, name
, reloc
);
9787 if (decl
&& DECL_ATTRIBUTES (decl
)
9788 && lookup_attribute ("overlaid", DECL_ATTRIBUTES (decl
)))
9789 flags
|= SECTION_VMS_OVERLAY
;
9790 if (decl
&& DECL_ATTRIBUTES (decl
)
9791 && lookup_attribute ("global", DECL_ATTRIBUTES (decl
)))
9792 flags
|= SECTION_VMS_GLOBAL
;
9793 if (decl
&& DECL_ATTRIBUTES (decl
)
9794 && lookup_attribute ("initialize", DECL_ATTRIBUTES (decl
)))
9795 flags
|= SECTION_VMS_INITIALIZE
;
9800 /* Switch to an arbitrary section NAME with attributes as specified
9801 by FLAGS. ALIGN specifies any known alignment requirements for
9802 the section; 0 if the default should be used. */
9805 vms_asm_named_section (const char *name
, unsigned int flags
,
9806 tree decl ATTRIBUTE_UNUSED
)
9808 fputc ('\n', asm_out_file
);
9809 fprintf (asm_out_file
, ".section\t%s", name
);
9811 if (flags
& SECTION_VMS_OVERLAY
)
9812 fprintf (asm_out_file
, ",OVR");
9813 if (flags
& SECTION_VMS_GLOBAL
)
9814 fprintf (asm_out_file
, ",GBL");
9815 if (flags
& SECTION_VMS_INITIALIZE
)
9816 fprintf (asm_out_file
, ",NOMOD");
9817 if (flags
& SECTION_DEBUG
)
9818 fprintf (asm_out_file
, ",NOWRT");
9820 fputc ('\n', asm_out_file
);
9823 /* Record an element in the table of global constructors. SYMBOL is
9824 a SYMBOL_REF of the function to be called; PRIORITY is a number
9825 between 0 and MAX_INIT_PRIORITY.
9827 Differs from default_ctors_section_asm_out_constructor in that the
9828 width of the .ctors entry is always 64 bits, rather than the 32 bits
9829 used by a normal pointer. */
9832 vms_asm_out_constructor (rtx symbol
, int priority ATTRIBUTE_UNUSED
)
9834 switch_to_section (ctors_section
);
9835 assemble_align (BITS_PER_WORD
);
9836 assemble_integer (symbol
, UNITS_PER_WORD
, BITS_PER_WORD
, 1);
9840 vms_asm_out_destructor (rtx symbol
, int priority ATTRIBUTE_UNUSED
)
9842 switch_to_section (dtors_section
);
9843 assemble_align (BITS_PER_WORD
);
9844 assemble_integer (symbol
, UNITS_PER_WORD
, BITS_PER_WORD
, 1);
9849 alpha_need_linkage (const char *name ATTRIBUTE_UNUSED
,
9850 int is_local ATTRIBUTE_UNUSED
)
9856 alpha_use_linkage (rtx linkage ATTRIBUTE_UNUSED
,
9857 tree cfundecl ATTRIBUTE_UNUSED
,
9858 int lflag ATTRIBUTE_UNUSED
,
9859 int rflag ATTRIBUTE_UNUSED
)
9864 #endif /* TARGET_ABI_OPEN_VMS */
9866 #if TARGET_ABI_UNICOSMK
9868 /* This evaluates to true if we do not know how to pass TYPE solely in
9869 registers. This is the case for all arguments that do not fit in two
9873 unicosmk_must_pass_in_stack (enum machine_mode mode
, const_tree type
)
9878 if (TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
9880 if (TREE_ADDRESSABLE (type
))
9883 return ALPHA_ARG_SIZE (mode
, type
, 0) > 2;
9886 /* Define the offset between two registers, one to be eliminated, and the
9887 other its replacement, at the start of a routine. */
9890 unicosmk_initial_elimination_offset (int from
, int to
)
9894 fixed_size
= alpha_sa_size();
9895 if (fixed_size
!= 0)
9898 if (from
== FRAME_POINTER_REGNUM
&& to
== HARD_FRAME_POINTER_REGNUM
)
9900 else if (from
== ARG_POINTER_REGNUM
&& to
== HARD_FRAME_POINTER_REGNUM
)
9902 else if (from
== FRAME_POINTER_REGNUM
&& to
== STACK_POINTER_REGNUM
)
9903 return (ALPHA_ROUND (crtl
->outgoing_args_size
)
9904 + ALPHA_ROUND (get_frame_size()));
9905 else if (from
== ARG_POINTER_REGNUM
&& to
== STACK_POINTER_REGNUM
)
9906 return (ALPHA_ROUND (fixed_size
)
9907 + ALPHA_ROUND (get_frame_size()
9908 + crtl
->outgoing_args_size
));
9913 /* Output the module name for .ident and .end directives. We have to strip
9914 directories and add make sure that the module name starts with a letter
9918 unicosmk_output_module_name (FILE *file
)
9920 const char *name
= lbasename (main_input_filename
);
9921 unsigned len
= strlen (name
);
9922 char *clean_name
= alloca (len
+ 2);
9923 char *ptr
= clean_name
;
9925 /* CAM only accepts module names that start with a letter or '$'. We
9926 prefix the module name with a '$' if necessary. */
9928 if (!ISALPHA (*name
))
9930 memcpy (ptr
, name
, len
+ 1);
9931 clean_symbol_name (clean_name
);
9932 fputs (clean_name
, file
);
9935 /* Output the definition of a common variable. */
9938 unicosmk_output_common (FILE *file
, const char *name
, int size
, int align
)
9941 printf ("T3E__: common %s\n", name
);
9944 fputs("\t.endp\n\n\t.psect ", file
);
9945 assemble_name(file
, name
);
9946 fprintf(file
, ",%d,common\n", floor_log2 (align
/ BITS_PER_UNIT
));
9947 fprintf(file
, "\t.byte\t0:%d\n", size
);
9949 /* Mark the symbol as defined in this module. */
9950 name_tree
= get_identifier (name
);
9951 TREE_ASM_WRITTEN (name_tree
) = 1;
9954 #define SECTION_PUBLIC SECTION_MACH_DEP
9955 #define SECTION_MAIN (SECTION_PUBLIC << 1)
9956 static int current_section_align
;
9958 /* A get_unnamed_section callback for switching to the text section. */
9961 unicosmk_output_text_section_asm_op (const void *data ATTRIBUTE_UNUSED
)
9963 static int count
= 0;
9964 fprintf (asm_out_file
, "\t.endp\n\n\t.psect\tgcc@text___%d,code\n", count
++);
9967 /* A get_unnamed_section callback for switching to the data section. */
9970 unicosmk_output_data_section_asm_op (const void *data ATTRIBUTE_UNUSED
)
9972 static int count
= 1;
9973 fprintf (asm_out_file
, "\t.endp\n\n\t.psect\tgcc@data___%d,data\n", count
++);
9976 /* Implement TARGET_ASM_INIT_SECTIONS.
9978 The Cray assembler is really weird with respect to sections. It has only
9979 named sections and you can't reopen a section once it has been closed.
9980 This means that we have to generate unique names whenever we want to
9981 reenter the text or the data section. */
9984 unicosmk_init_sections (void)
9986 text_section
= get_unnamed_section (SECTION_CODE
,
9987 unicosmk_output_text_section_asm_op
,
9989 data_section
= get_unnamed_section (SECTION_WRITE
,
9990 unicosmk_output_data_section_asm_op
,
9992 readonly_data_section
= data_section
;
9996 unicosmk_section_type_flags (tree decl
, const char *name
,
9997 int reloc ATTRIBUTE_UNUSED
)
9999 unsigned int flags
= default_section_type_flags (decl
, name
, reloc
);
10004 if (TREE_CODE (decl
) == FUNCTION_DECL
)
10006 current_section_align
= floor_log2 (FUNCTION_BOUNDARY
/ BITS_PER_UNIT
);
10007 if (align_functions_log
> current_section_align
)
10008 current_section_align
= align_functions_log
;
10010 if (! strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl
)), "main"))
10011 flags
|= SECTION_MAIN
;
10014 current_section_align
= floor_log2 (DECL_ALIGN (decl
) / BITS_PER_UNIT
);
10016 if (TREE_PUBLIC (decl
))
10017 flags
|= SECTION_PUBLIC
;
10022 /* Generate a section name for decl and associate it with the
10026 unicosmk_unique_section (tree decl
, int reloc ATTRIBUTE_UNUSED
)
10033 name
= IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl
));
10034 name
= default_strip_name_encoding (name
);
10035 len
= strlen (name
);
10037 if (TREE_CODE (decl
) == FUNCTION_DECL
)
10041 /* It is essential that we prefix the section name here because
10042 otherwise the section names generated for constructors and
10043 destructors confuse collect2. */
10045 string
= alloca (len
+ 6);
10046 sprintf (string
, "code@%s", name
);
10047 DECL_SECTION_NAME (decl
) = build_string (len
+ 5, string
);
10049 else if (TREE_PUBLIC (decl
))
10050 DECL_SECTION_NAME (decl
) = build_string (len
, name
);
10055 string
= alloca (len
+ 6);
10056 sprintf (string
, "data@%s", name
);
10057 DECL_SECTION_NAME (decl
) = build_string (len
+ 5, string
);
10061 /* Switch to an arbitrary section NAME with attributes as specified
10062 by FLAGS. ALIGN specifies any known alignment requirements for
10063 the section; 0 if the default should be used. */
10066 unicosmk_asm_named_section (const char *name
, unsigned int flags
,
10067 tree decl ATTRIBUTE_UNUSED
)
10071 /* Close the previous section. */
10073 fputs ("\t.endp\n\n", asm_out_file
);
10075 /* Find out what kind of section we are opening. */
10077 if (flags
& SECTION_MAIN
)
10078 fputs ("\t.start\tmain\n", asm_out_file
);
10080 if (flags
& SECTION_CODE
)
10082 else if (flags
& SECTION_PUBLIC
)
10087 if (current_section_align
!= 0)
10088 fprintf (asm_out_file
, "\t.psect\t%s,%d,%s\n", name
,
10089 current_section_align
, kind
);
10091 fprintf (asm_out_file
, "\t.psect\t%s,%s\n", name
, kind
);
10095 unicosmk_insert_attributes (tree decl
, tree
*attr_ptr ATTRIBUTE_UNUSED
)
10098 && (TREE_PUBLIC (decl
) || TREE_CODE (decl
) == FUNCTION_DECL
))
10099 unicosmk_unique_section (decl
, 0);
10102 /* Output an alignment directive. We have to use the macro 'gcc@code@align'
10103 in code sections because .align fill unused space with zeroes. */
10106 unicosmk_output_align (FILE *file
, int align
)
10108 if (inside_function
)
10109 fprintf (file
, "\tgcc@code@align\t%d\n", align
);
10111 fprintf (file
, "\t.align\t%d\n", align
);
10114 /* Add a case vector to the current function's list of deferred case
10115 vectors. Case vectors have to be put into a separate section because CAM
10116 does not allow data definitions in code sections. */
10119 unicosmk_defer_case_vector (rtx lab
, rtx vec
)
10121 struct machine_function
*machine
= cfun
->machine
;
10123 vec
= gen_rtx_EXPR_LIST (VOIDmode
, lab
, vec
);
10124 machine
->addr_list
= gen_rtx_EXPR_LIST (VOIDmode
, vec
,
10125 machine
->addr_list
);
10128 /* Output a case vector. */
10131 unicosmk_output_addr_vec (FILE *file
, rtx vec
)
10133 rtx lab
= XEXP (vec
, 0);
10134 rtx body
= XEXP (vec
, 1);
10135 int vlen
= XVECLEN (body
, 0);
10138 (*targetm
.asm_out
.internal_label
) (file
, "L", CODE_LABEL_NUMBER (lab
));
10140 for (idx
= 0; idx
< vlen
; idx
++)
10142 ASM_OUTPUT_ADDR_VEC_ELT
10143 (file
, CODE_LABEL_NUMBER (XEXP (XVECEXP (body
, 0, idx
), 0)));
10147 /* Output current function's deferred case vectors. */
10150 unicosmk_output_deferred_case_vectors (FILE *file
)
10152 struct machine_function
*machine
= cfun
->machine
;
10155 if (machine
->addr_list
== NULL_RTX
)
10158 switch_to_section (data_section
);
10159 for (t
= machine
->addr_list
; t
; t
= XEXP (t
, 1))
10160 unicosmk_output_addr_vec (file
, XEXP (t
, 0));
10163 /* Generate the name of the SSIB section for the current function. */
10165 #define SSIB_PREFIX "__SSIB_"
10166 #define SSIB_PREFIX_LEN 7
10168 static const char *
10169 unicosmk_ssib_name (void)
10171 /* This is ok since CAM won't be able to deal with names longer than that
10174 static char name
[256];
10177 const char *fnname
;
10180 x
= DECL_RTL (cfun
->decl
);
10181 gcc_assert (MEM_P (x
));
10183 gcc_assert (GET_CODE (x
) == SYMBOL_REF
);
10184 fnname
= XSTR (x
, 0);
10186 len
= strlen (fnname
);
10187 if (len
+ SSIB_PREFIX_LEN
> 255)
10188 len
= 255 - SSIB_PREFIX_LEN
;
10190 strcpy (name
, SSIB_PREFIX
);
10191 strncpy (name
+ SSIB_PREFIX_LEN
, fnname
, len
);
10192 name
[len
+ SSIB_PREFIX_LEN
] = 0;
10197 /* Set up the dynamic subprogram information block (DSIB) and update the
10198 frame pointer register ($15) for subroutines which have a frame. If the
10199 subroutine doesn't have a frame, simply increment $15. */
10202 unicosmk_gen_dsib (unsigned long *imaskP
)
10204 if (alpha_procedure_type
== PT_STACK
)
10206 const char *ssib_name
;
10209 /* Allocate 64 bytes for the DSIB. */
10211 FRP (emit_insn (gen_adddi3 (stack_pointer_rtx
, stack_pointer_rtx
,
10213 emit_insn (gen_blockage ());
10215 /* Save the return address. */
10217 mem
= gen_rtx_MEM (DImode
, plus_constant (stack_pointer_rtx
, 56));
10218 set_mem_alias_set (mem
, alpha_sr_alias_set
);
10219 FRP (emit_move_insn (mem
, gen_rtx_REG (DImode
, REG_RA
)));
10220 (*imaskP
) &= ~(1UL << REG_RA
);
10222 /* Save the old frame pointer. */
10224 mem
= gen_rtx_MEM (DImode
, plus_constant (stack_pointer_rtx
, 48));
10225 set_mem_alias_set (mem
, alpha_sr_alias_set
);
10226 FRP (emit_move_insn (mem
, hard_frame_pointer_rtx
));
10227 (*imaskP
) &= ~(1UL << HARD_FRAME_POINTER_REGNUM
);
10229 emit_insn (gen_blockage ());
10231 /* Store the SSIB pointer. */
10233 ssib_name
= ggc_strdup (unicosmk_ssib_name ());
10234 mem
= gen_rtx_MEM (DImode
, plus_constant (stack_pointer_rtx
, 32));
10235 set_mem_alias_set (mem
, alpha_sr_alias_set
);
10237 FRP (emit_move_insn (gen_rtx_REG (DImode
, 5),
10238 gen_rtx_SYMBOL_REF (Pmode
, ssib_name
)));
10239 FRP (emit_move_insn (mem
, gen_rtx_REG (DImode
, 5)));
10241 /* Save the CIW index. */
10243 mem
= gen_rtx_MEM (DImode
, plus_constant (stack_pointer_rtx
, 24));
10244 set_mem_alias_set (mem
, alpha_sr_alias_set
);
10245 FRP (emit_move_insn (mem
, gen_rtx_REG (DImode
, 25)));
10247 emit_insn (gen_blockage ());
10249 /* Set the new frame pointer. */
10251 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx
,
10252 stack_pointer_rtx
, GEN_INT (64))));
10257 /* Increment the frame pointer register to indicate that we do not
10260 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx
,
10261 hard_frame_pointer_rtx
, const1_rtx
)));
10265 /* Output the static subroutine information block for the current
10269 unicosmk_output_ssib (FILE *file
, const char *fnname
)
10275 struct machine_function
*machine
= cfun
->machine
;
10278 fprintf (file
, "\t.endp\n\n\t.psect\t%s%s,data\n", user_label_prefix
,
10279 unicosmk_ssib_name ());
10281 /* Some required stuff and the function name length. */
10283 len
= strlen (fnname
);
10284 fprintf (file
, "\t.quad\t^X20008%2.2X28\n", len
);
10287 ??? We don't do that yet. */
10289 fputs ("\t.quad\t0\n", file
);
10291 /* Function address. */
10293 fputs ("\t.quad\t", file
);
10294 assemble_name (file
, fnname
);
10297 fputs ("\t.quad\t0\n", file
);
10298 fputs ("\t.quad\t0\n", file
);
10301 ??? We do it the same way Cray CC does it but this could be
10304 for( i
= 0; i
< len
; i
++ )
10305 fprintf (file
, "\t.byte\t%d\n", (int)(fnname
[i
]));
10306 if( (len
% 8) == 0 )
10307 fputs ("\t.quad\t0\n", file
);
10309 fprintf (file
, "\t.bits\t%d : 0\n", (8 - (len
% 8))*8);
10311 /* All call information words used in the function. */
10313 for (x
= machine
->first_ciw
; x
; x
= XEXP (x
, 1))
10316 #if HOST_BITS_PER_WIDE_INT == 32
10317 fprintf (file
, "\t.quad\t" HOST_WIDE_INT_PRINT_DOUBLE_HEX
"\n",
10318 CONST_DOUBLE_HIGH (ciw
), CONST_DOUBLE_LOW (ciw
));
10320 fprintf (file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n", INTVAL (ciw
));
10325 /* Add a call information word (CIW) to the list of the current function's
10326 CIWs and return its index.
10328 X is a CONST_INT or CONST_DOUBLE representing the CIW. */
10331 unicosmk_add_call_info_word (rtx x
)
10334 struct machine_function
*machine
= cfun
->machine
;
10336 node
= gen_rtx_EXPR_LIST (VOIDmode
, x
, NULL_RTX
);
10337 if (machine
->first_ciw
== NULL_RTX
)
10338 machine
->first_ciw
= node
;
10340 XEXP (machine
->last_ciw
, 1) = node
;
10342 machine
->last_ciw
= node
;
10343 ++machine
->ciw_count
;
10345 return GEN_INT (machine
->ciw_count
10346 + strlen (current_function_name ())/8 + 5);
10349 /* The Cray assembler doesn't accept extern declarations for symbols which
10350 are defined in the same file. We have to keep track of all global
10351 symbols which are referenced and/or defined in a source file and output
10352 extern declarations for those which are referenced but not defined at
10353 the end of file. */
10355 /* List of identifiers for which an extern declaration might have to be
10357 /* FIXME: needs to use GC, so it can be saved and restored for PCH. */
10359 struct unicosmk_extern_list
10361 struct unicosmk_extern_list
*next
;
10365 static struct unicosmk_extern_list
*unicosmk_extern_head
= 0;
10367 /* Output extern declarations which are required for every asm file. */
10370 unicosmk_output_default_externs (FILE *file
)
10372 static const char *const externs
[] =
10373 { "__T3E_MISMATCH" };
10378 n
= ARRAY_SIZE (externs
);
10380 for (i
= 0; i
< n
; i
++)
10381 fprintf (file
, "\t.extern\t%s\n", externs
[i
]);
10384 /* Output extern declarations for global symbols which are have been
10385 referenced but not defined. */
10388 unicosmk_output_externs (FILE *file
)
10390 struct unicosmk_extern_list
*p
;
10391 const char *real_name
;
10395 len
= strlen (user_label_prefix
);
10396 for (p
= unicosmk_extern_head
; p
!= 0; p
= p
->next
)
10398 /* We have to strip the encoding and possibly remove user_label_prefix
10399 from the identifier in order to handle -fleading-underscore and
10400 explicit asm names correctly (cf. gcc.dg/asm-names-1.c). */
10401 real_name
= default_strip_name_encoding (p
->name
);
10402 if (len
&& p
->name
[0] == '*'
10403 && !memcmp (real_name
, user_label_prefix
, len
))
10406 name_tree
= get_identifier (real_name
);
10407 if (! TREE_ASM_WRITTEN (name_tree
))
10409 TREE_ASM_WRITTEN (name_tree
) = 1;
10410 fputs ("\t.extern\t", file
);
10411 assemble_name (file
, p
->name
);
10417 /* Record an extern. */
10420 unicosmk_add_extern (const char *name
)
10422 struct unicosmk_extern_list
*p
;
10424 p
= (struct unicosmk_extern_list
*)
10425 xmalloc (sizeof (struct unicosmk_extern_list
));
10426 p
->next
= unicosmk_extern_head
;
10428 unicosmk_extern_head
= p
;
10431 /* The Cray assembler generates incorrect code if identifiers which
10432 conflict with register names are used as instruction operands. We have
10433 to replace such identifiers with DEX expressions. */
10435 /* Structure to collect identifiers which have been replaced by DEX
10437 /* FIXME: needs to use GC, so it can be saved and restored for PCH. */
10439 struct unicosmk_dex
{
10440 struct unicosmk_dex
*next
;
10444 /* List of identifiers which have been replaced by DEX expressions. The DEX
10445 number is determined by the position in the list. */
10447 static struct unicosmk_dex
*unicosmk_dex_list
= NULL
;
10449 /* The number of elements in the DEX list. */
10451 static int unicosmk_dex_count
= 0;
10453 /* Check if NAME must be replaced by a DEX expression. */
10456 unicosmk_special_name (const char *name
)
10458 if (name
[0] == '*')
10461 if (name
[0] == '$')
10464 if (name
[0] != 'r' && name
[0] != 'f' && name
[0] != 'R' && name
[0] != 'F')
10469 case '1': case '2':
10470 return (name
[2] == '\0' || (ISDIGIT (name
[2]) && name
[3] == '\0'));
10473 return (name
[2] == '\0'
10474 || ((name
[2] == '0' || name
[2] == '1') && name
[3] == '\0'));
10477 return (ISDIGIT (name
[1]) && name
[2] == '\0');
10481 /* Return the DEX number if X must be replaced by a DEX expression and 0
10485 unicosmk_need_dex (rtx x
)
10487 struct unicosmk_dex
*dex
;
10491 if (GET_CODE (x
) != SYMBOL_REF
)
10495 if (! unicosmk_special_name (name
))
10498 i
= unicosmk_dex_count
;
10499 for (dex
= unicosmk_dex_list
; dex
; dex
= dex
->next
)
10501 if (! strcmp (name
, dex
->name
))
10506 dex
= (struct unicosmk_dex
*) xmalloc (sizeof (struct unicosmk_dex
));
10508 dex
->next
= unicosmk_dex_list
;
10509 unicosmk_dex_list
= dex
;
10511 ++unicosmk_dex_count
;
10512 return unicosmk_dex_count
;
10515 /* Output the DEX definitions for this file. */
10518 unicosmk_output_dex (FILE *file
)
10520 struct unicosmk_dex
*dex
;
10523 if (unicosmk_dex_list
== NULL
)
10526 fprintf (file
, "\t.dexstart\n");
10528 i
= unicosmk_dex_count
;
10529 for (dex
= unicosmk_dex_list
; dex
; dex
= dex
->next
)
10531 fprintf (file
, "\tDEX (%d) = ", i
);
10532 assemble_name (file
, dex
->name
);
10537 fprintf (file
, "\t.dexend\n");
10540 /* Output text that to appear at the beginning of an assembler file. */
10543 unicosmk_file_start (void)
10547 fputs ("\t.ident\t", asm_out_file
);
10548 unicosmk_output_module_name (asm_out_file
);
10549 fputs ("\n\n", asm_out_file
);
10551 /* The Unicos/Mk assembler uses different register names. Instead of trying
10552 to support them, we simply use micro definitions. */
10554 /* CAM has different register names: rN for the integer register N and fN
10555 for the floating-point register N. Instead of trying to use these in
10556 alpha.md, we define the symbols $N and $fN to refer to the appropriate
10559 for (i
= 0; i
< 32; ++i
)
10560 fprintf (asm_out_file
, "$%d <- r%d\n", i
, i
);
10562 for (i
= 0; i
< 32; ++i
)
10563 fprintf (asm_out_file
, "$f%d <- f%d\n", i
, i
);
10565 putc ('\n', asm_out_file
);
10567 /* The .align directive fill unused space with zeroes which does not work
10568 in code sections. We define the macro 'gcc@code@align' which uses nops
10569 instead. Note that it assumes that code sections always have the
10570 biggest possible alignment since . refers to the current offset from
10571 the beginning of the section. */
10573 fputs ("\t.macro gcc@code@align n\n", asm_out_file
);
10574 fputs ("gcc@n@bytes = 1 << n\n", asm_out_file
);
10575 fputs ("gcc@here = . % gcc@n@bytes\n", asm_out_file
);
10576 fputs ("\t.if ne, gcc@here, 0\n", asm_out_file
);
10577 fputs ("\t.repeat (gcc@n@bytes - gcc@here) / 4\n", asm_out_file
);
10578 fputs ("\tbis r31,r31,r31\n", asm_out_file
);
10579 fputs ("\t.endr\n", asm_out_file
);
10580 fputs ("\t.endif\n", asm_out_file
);
10581 fputs ("\t.endm gcc@code@align\n\n", asm_out_file
);
10583 /* Output extern declarations which should always be visible. */
10584 unicosmk_output_default_externs (asm_out_file
);
10586 /* Open a dummy section. We always need to be inside a section for the
10587 section-switching code to work correctly.
10588 ??? This should be a module id or something like that. I still have to
10589 figure out what the rules for those are. */
10590 fputs ("\n\t.psect\t$SG00000,data\n", asm_out_file
);
10593 /* Output text to appear at the end of an assembler file. This includes all
10594 pending extern declarations and DEX expressions. */
10597 unicosmk_file_end (void)
10599 fputs ("\t.endp\n\n", asm_out_file
);
10601 /* Output all pending externs. */
10603 unicosmk_output_externs (asm_out_file
);
10605 /* Output dex definitions used for functions whose names conflict with
10608 unicosmk_output_dex (asm_out_file
);
10610 fputs ("\t.end\t", asm_out_file
);
10611 unicosmk_output_module_name (asm_out_file
);
10612 putc ('\n', asm_out_file
);
10618 unicosmk_output_deferred_case_vectors (FILE *file ATTRIBUTE_UNUSED
)
10622 unicosmk_gen_dsib (unsigned long *imaskP ATTRIBUTE_UNUSED
)
10626 unicosmk_output_ssib (FILE * file ATTRIBUTE_UNUSED
,
10627 const char * fnname ATTRIBUTE_UNUSED
)
10631 unicosmk_add_call_info_word (rtx x ATTRIBUTE_UNUSED
)
10637 unicosmk_need_dex (rtx x ATTRIBUTE_UNUSED
)
10642 #endif /* TARGET_ABI_UNICOSMK */
10645 alpha_init_libfuncs (void)
10647 if (TARGET_ABI_UNICOSMK
)
10649 /* Prevent gcc from generating calls to __divsi3. */
10650 set_optab_libfunc (sdiv_optab
, SImode
, 0);
10651 set_optab_libfunc (udiv_optab
, SImode
, 0);
10653 /* Use the functions provided by the system library
10654 for DImode integer division. */
10655 set_optab_libfunc (sdiv_optab
, DImode
, "$sldiv");
10656 set_optab_libfunc (udiv_optab
, DImode
, "$uldiv");
10658 else if (TARGET_ABI_OPEN_VMS
)
10660 /* Use the VMS runtime library functions for division and
10662 set_optab_libfunc (sdiv_optab
, SImode
, "OTS$DIV_I");
10663 set_optab_libfunc (sdiv_optab
, DImode
, "OTS$DIV_L");
10664 set_optab_libfunc (udiv_optab
, SImode
, "OTS$DIV_UI");
10665 set_optab_libfunc (udiv_optab
, DImode
, "OTS$DIV_UL");
10666 set_optab_libfunc (smod_optab
, SImode
, "OTS$REM_I");
10667 set_optab_libfunc (smod_optab
, DImode
, "OTS$REM_L");
10668 set_optab_libfunc (umod_optab
, SImode
, "OTS$REM_UI");
10669 set_optab_libfunc (umod_optab
, DImode
, "OTS$REM_UL");
10674 /* Initialize the GCC target structure. */
10675 #if TARGET_ABI_OPEN_VMS
10676 # undef TARGET_ATTRIBUTE_TABLE
10677 # define TARGET_ATTRIBUTE_TABLE vms_attribute_table
10678 # undef TARGET_SECTION_TYPE_FLAGS
10679 # define TARGET_SECTION_TYPE_FLAGS vms_section_type_flags
10682 #undef TARGET_IN_SMALL_DATA_P
10683 #define TARGET_IN_SMALL_DATA_P alpha_in_small_data_p
10685 #if TARGET_ABI_UNICOSMK
10686 # undef TARGET_INSERT_ATTRIBUTES
10687 # define TARGET_INSERT_ATTRIBUTES unicosmk_insert_attributes
10688 # undef TARGET_SECTION_TYPE_FLAGS
10689 # define TARGET_SECTION_TYPE_FLAGS unicosmk_section_type_flags
10690 # undef TARGET_ASM_UNIQUE_SECTION
10691 # define TARGET_ASM_UNIQUE_SECTION unicosmk_unique_section
10692 #undef TARGET_ASM_FUNCTION_RODATA_SECTION
10693 #define TARGET_ASM_FUNCTION_RODATA_SECTION default_no_function_rodata_section
10694 # undef TARGET_ASM_GLOBALIZE_LABEL
10695 # define TARGET_ASM_GLOBALIZE_LABEL hook_void_FILEptr_constcharptr
10696 # undef TARGET_MUST_PASS_IN_STACK
10697 # define TARGET_MUST_PASS_IN_STACK unicosmk_must_pass_in_stack
10700 #undef TARGET_ASM_ALIGNED_HI_OP
10701 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10702 #undef TARGET_ASM_ALIGNED_DI_OP
10703 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10705 /* Default unaligned ops are provided for ELF systems. To get unaligned
10706 data for non-ELF systems, we have to turn off auto alignment. */
10707 #ifndef OBJECT_FORMAT_ELF
10708 #undef TARGET_ASM_UNALIGNED_HI_OP
10709 #define TARGET_ASM_UNALIGNED_HI_OP "\t.align 0\n\t.word\t"
10710 #undef TARGET_ASM_UNALIGNED_SI_OP
10711 #define TARGET_ASM_UNALIGNED_SI_OP "\t.align 0\n\t.long\t"
10712 #undef TARGET_ASM_UNALIGNED_DI_OP
10713 #define TARGET_ASM_UNALIGNED_DI_OP "\t.align 0\n\t.quad\t"
10716 #ifdef OBJECT_FORMAT_ELF
10717 #undef TARGET_ASM_RELOC_RW_MASK
10718 #define TARGET_ASM_RELOC_RW_MASK alpha_elf_reloc_rw_mask
10719 #undef TARGET_ASM_SELECT_RTX_SECTION
10720 #define TARGET_ASM_SELECT_RTX_SECTION alpha_elf_select_rtx_section
10721 #undef TARGET_SECTION_TYPE_FLAGS
10722 #define TARGET_SECTION_TYPE_FLAGS alpha_elf_section_type_flags
10725 #undef TARGET_ASM_FUNCTION_END_PROLOGUE
10726 #define TARGET_ASM_FUNCTION_END_PROLOGUE alpha_output_function_end_prologue
10728 #undef TARGET_INIT_LIBFUNCS
10729 #define TARGET_INIT_LIBFUNCS alpha_init_libfuncs
10731 #undef TARGET_LEGITIMIZE_ADDRESS
10732 #define TARGET_LEGITIMIZE_ADDRESS alpha_legitimize_address
10734 #if TARGET_ABI_UNICOSMK
10735 #undef TARGET_ASM_FILE_START
10736 #define TARGET_ASM_FILE_START unicosmk_file_start
10737 #undef TARGET_ASM_FILE_END
10738 #define TARGET_ASM_FILE_END unicosmk_file_end
10740 #undef TARGET_ASM_FILE_START
10741 #define TARGET_ASM_FILE_START alpha_file_start
10742 #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE
10743 #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true
10746 #undef TARGET_SCHED_ADJUST_COST
10747 #define TARGET_SCHED_ADJUST_COST alpha_adjust_cost
10748 #undef TARGET_SCHED_ISSUE_RATE
10749 #define TARGET_SCHED_ISSUE_RATE alpha_issue_rate
10750 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10751 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
10752 alpha_multipass_dfa_lookahead
10754 #undef TARGET_HAVE_TLS
10755 #define TARGET_HAVE_TLS HAVE_AS_TLS
10757 #undef TARGET_INIT_BUILTINS
10758 #define TARGET_INIT_BUILTINS alpha_init_builtins
10759 #undef TARGET_EXPAND_BUILTIN
10760 #define TARGET_EXPAND_BUILTIN alpha_expand_builtin
10761 #undef TARGET_FOLD_BUILTIN
10762 #define TARGET_FOLD_BUILTIN alpha_fold_builtin
10764 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10765 #define TARGET_FUNCTION_OK_FOR_SIBCALL alpha_function_ok_for_sibcall
10766 #undef TARGET_CANNOT_COPY_INSN_P
10767 #define TARGET_CANNOT_COPY_INSN_P alpha_cannot_copy_insn_p
10768 #undef TARGET_CANNOT_FORCE_CONST_MEM
10769 #define TARGET_CANNOT_FORCE_CONST_MEM alpha_cannot_force_const_mem
10772 #undef TARGET_ASM_OUTPUT_MI_THUNK
10773 #define TARGET_ASM_OUTPUT_MI_THUNK alpha_output_mi_thunk_osf
10774 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10775 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10776 #undef TARGET_STDARG_OPTIMIZE_HOOK
10777 #define TARGET_STDARG_OPTIMIZE_HOOK alpha_stdarg_optimize_hook
10780 #undef TARGET_RTX_COSTS
10781 #define TARGET_RTX_COSTS alpha_rtx_costs
10782 #undef TARGET_ADDRESS_COST
10783 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0
10785 #undef TARGET_MACHINE_DEPENDENT_REORG
10786 #define TARGET_MACHINE_DEPENDENT_REORG alpha_reorg
10788 #undef TARGET_PROMOTE_FUNCTION_ARGS
10789 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
10790 #undef TARGET_PROMOTE_FUNCTION_RETURN
10791 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
10792 #undef TARGET_PROMOTE_PROTOTYPES
10793 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_false
10794 #undef TARGET_RETURN_IN_MEMORY
10795 #define TARGET_RETURN_IN_MEMORY alpha_return_in_memory
10796 #undef TARGET_PASS_BY_REFERENCE
10797 #define TARGET_PASS_BY_REFERENCE alpha_pass_by_reference
10798 #undef TARGET_SETUP_INCOMING_VARARGS
10799 #define TARGET_SETUP_INCOMING_VARARGS alpha_setup_incoming_varargs
10800 #undef TARGET_STRICT_ARGUMENT_NAMING
10801 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
10802 #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED
10803 #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED hook_bool_CUMULATIVE_ARGS_true
10804 #undef TARGET_SPLIT_COMPLEX_ARG
10805 #define TARGET_SPLIT_COMPLEX_ARG alpha_split_complex_arg
10806 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10807 #define TARGET_GIMPLIFY_VA_ARG_EXPR alpha_gimplify_va_arg
10808 #undef TARGET_ARG_PARTIAL_BYTES
10809 #define TARGET_ARG_PARTIAL_BYTES alpha_arg_partial_bytes
10811 #undef TARGET_SECONDARY_RELOAD
10812 #define TARGET_SECONDARY_RELOAD alpha_secondary_reload
10814 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10815 #define TARGET_SCALAR_MODE_SUPPORTED_P alpha_scalar_mode_supported_p
10816 #undef TARGET_VECTOR_MODE_SUPPORTED_P
10817 #define TARGET_VECTOR_MODE_SUPPORTED_P alpha_vector_mode_supported_p
10819 #undef TARGET_BUILD_BUILTIN_VA_LIST
10820 #define TARGET_BUILD_BUILTIN_VA_LIST alpha_build_builtin_va_list
10822 #undef TARGET_EXPAND_BUILTIN_VA_START
10823 #define TARGET_EXPAND_BUILTIN_VA_START alpha_va_start
10825 /* The Alpha architecture does not require sequential consistency. See
10826 http://www.cs.umd.edu/~pugh/java/memoryModel/AlphaReordering.html
10827 for an example of how it can be violated in practice. */
10828 #undef TARGET_RELAXED_ORDERING
10829 #define TARGET_RELAXED_ORDERING true
10831 #undef TARGET_DEFAULT_TARGET_FLAGS
10832 #define TARGET_DEFAULT_TARGET_FLAGS \
10833 (TARGET_DEFAULT | TARGET_CPU_DEFAULT | TARGET_DEFAULT_EXPLICIT_RELOCS)
10834 #undef TARGET_HANDLE_OPTION
10835 #define TARGET_HANDLE_OPTION alpha_handle_option
10837 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10838 #undef TARGET_MANGLE_TYPE
10839 #define TARGET_MANGLE_TYPE alpha_mangle_type
10842 #undef TARGET_LEGITIMATE_ADDRESS_P
10843 #define TARGET_LEGITIMATE_ADDRESS_P alpha_legitimate_address_p
10845 struct gcc_target targetm
= TARGET_INITIALIZER
;
10848 #include "gt-alpha.h"