1 /* Subroutines used for code generation on the DEC Alpha.
2 Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
3 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
4 Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
6 This file is part of GNU CC.
8 GNU CC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 2, or (at your option)
13 GNU CC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GNU CC; see the file COPYING. If not, write to
20 the Free Software Foundation, 59 Temple Place - Suite 330,
21 Boston, MA 02111-1307, USA. */
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 /* Specify which cpu to schedule for. */
58 enum processor_type alpha_cpu;
59 static const char * const alpha_cpu_name[] =
64 /* Specify how accurate floating-point traps need to be. */
66 enum alpha_trap_precision alpha_tp;
68 /* Specify the floating-point rounding mode. */
70 enum alpha_fp_rounding_mode alpha_fprm;
72 /* Specify which things cause traps. */
74 enum alpha_fp_trap_mode alpha_fptm;
76 /* Specify bit size of immediate TLS offsets. */
78 int alpha_tls_size = 32;
80 /* Strings decoded into the above options. */
82 const char *alpha_cpu_string; /* -mcpu= */
83 const char *alpha_tune_string; /* -mtune= */
84 const char *alpha_tp_string; /* -mtrap-precision=[p|s|i] */
85 const char *alpha_fprm_string; /* -mfp-rounding-mode=[n|m|c|d] */
86 const char *alpha_fptm_string; /* -mfp-trap-mode=[n|u|su|sui] */
87 const char *alpha_mlat_string; /* -mmemory-latency= */
88 const char *alpha_tls_size_string; /* -mtls-size=[16|32|64] */
90 /* Save information from a "cmpxx" operation until the branch or scc is
93 struct alpha_compare alpha_compare;
95 /* Nonzero if inside of a function, because the Alpha asm can't
96 handle .files inside of functions. */
98 static int inside_function = FALSE;
100 /* The number of cycles of latency we should assume on memory reads. */
102 int alpha_memory_latency = 3;
104 /* Whether the function needs the GP. */
106 static int alpha_function_needs_gp;
108 /* The alias set for prologue/epilogue register save/restore. */
110 static GTY(()) int alpha_sr_alias_set;
112 /* The assembler name of the current function. */
114 static const char *alpha_fnname;
116 /* The next explicit relocation sequence number. */
117 extern GTY(()) int alpha_next_sequence_number;
118 int alpha_next_sequence_number = 1;
120 /* The literal and gpdisp sequence numbers for this insn, as printed
121 by %# and %* respectively. */
122 extern GTY(()) int alpha_this_literal_sequence_number;
123 extern GTY(()) int alpha_this_gpdisp_sequence_number;
124 int alpha_this_literal_sequence_number;
125 int alpha_this_gpdisp_sequence_number;
127 /* Costs of various operations on the different architectures. */
129 struct alpha_rtx_cost_data
131 unsigned char fp_add;
132 unsigned char fp_mult;
133 unsigned char fp_div_sf;
134 unsigned char fp_div_df;
135 unsigned char int_mult_si;
136 unsigned char int_mult_di;
137 unsigned char int_shift;
138 unsigned char int_cmov;
141 static struct alpha_rtx_cost_data const alpha_rtx_cost_data[PROCESSOR_MAX] =
144 COSTS_N_INSNS (6), /* fp_add */
145 COSTS_N_INSNS (6), /* fp_mult */
146 COSTS_N_INSNS (34), /* fp_div_sf */
147 COSTS_N_INSNS (63), /* fp_div_df */
148 COSTS_N_INSNS (23), /* int_mult_si */
149 COSTS_N_INSNS (23), /* int_mult_di */
150 COSTS_N_INSNS (2), /* int_shift */
151 COSTS_N_INSNS (2), /* int_cmov */
154 COSTS_N_INSNS (4), /* fp_add */
155 COSTS_N_INSNS (4), /* fp_mult */
156 COSTS_N_INSNS (15), /* fp_div_sf */
157 COSTS_N_INSNS (22), /* fp_div_df */
158 COSTS_N_INSNS (8), /* int_mult_si */
159 COSTS_N_INSNS (12), /* int_mult_di */
160 COSTS_N_INSNS (1) + 1, /* int_shift */
161 COSTS_N_INSNS (1), /* int_cmov */
164 COSTS_N_INSNS (4), /* fp_add */
165 COSTS_N_INSNS (4), /* fp_mult */
166 COSTS_N_INSNS (12), /* fp_div_sf */
167 COSTS_N_INSNS (15), /* fp_div_df */
168 COSTS_N_INSNS (7), /* int_mult_si */
169 COSTS_N_INSNS (7), /* int_mult_di */
170 COSTS_N_INSNS (1), /* int_shift */
171 COSTS_N_INSNS (2), /* int_cmov */
175 /* Get the number of args of a function in one of two ways. */
176 #if TARGET_ABI_OPEN_VMS || TARGET_ABI_UNICOSMK
177 #define NUM_ARGS current_function_args_info.num_args
179 #define NUM_ARGS current_function_args_info
185 /* Declarations of static functions. */
186 static struct machine_function *alpha_init_machine_status (void);
187 static rtx alpha_emit_xfloating_compare (enum rtx_code, rtx, rtx);
189 #if TARGET_ABI_OPEN_VMS
190 static void alpha_write_linkage (FILE *, const char *, tree);
193 static void unicosmk_output_deferred_case_vectors (FILE *);
194 static void unicosmk_gen_dsib (unsigned long *);
195 static void unicosmk_output_ssib (FILE *, const char *);
196 static int unicosmk_need_dex (rtx);
198 /* Parse target option strings. */
201 override_options (void)
204 static const struct cpu_table {
205 const char *const name;
206 const enum processor_type processor;
209 #define EV5_MASK (MASK_CPU_EV5)
210 #define EV6_MASK (MASK_CPU_EV6|MASK_BWX|MASK_MAX|MASK_FIX)
211 { "ev4", PROCESSOR_EV4, 0 },
212 { "ev45", PROCESSOR_EV4, 0 },
213 { "21064", PROCESSOR_EV4, 0 },
214 { "ev5", PROCESSOR_EV5, EV5_MASK },
215 { "21164", PROCESSOR_EV5, EV5_MASK },
216 { "ev56", PROCESSOR_EV5, EV5_MASK|MASK_BWX },
217 { "21164a", PROCESSOR_EV5, EV5_MASK|MASK_BWX },
218 { "pca56", PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
219 { "21164PC",PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
220 { "21164pc",PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
221 { "ev6", PROCESSOR_EV6, EV6_MASK },
222 { "21264", PROCESSOR_EV6, EV6_MASK },
223 { "ev67", PROCESSOR_EV6, EV6_MASK|MASK_CIX },
224 { "21264a", PROCESSOR_EV6, EV6_MASK|MASK_CIX },
228 /* Unicos/Mk doesn't have shared libraries. */
229 if (TARGET_ABI_UNICOSMK && flag_pic)
231 warning ("-f%s ignored for Unicos/Mk (not supported)",
232 (flag_pic > 1) ? "PIC" : "pic");
236 /* On Unicos/Mk, the native compiler consistently generates /d suffices for
237 floating-point instructions. Make that the default for this target. */
238 if (TARGET_ABI_UNICOSMK)
239 alpha_fprm = ALPHA_FPRM_DYN;
241 alpha_fprm = ALPHA_FPRM_NORM;
243 alpha_tp = ALPHA_TP_PROG;
244 alpha_fptm = ALPHA_FPTM_N;
246 /* We cannot use su and sui qualifiers for conversion instructions on
247 Unicos/Mk. I'm not sure if this is due to assembler or hardware
248 limitations. Right now, we issue a warning if -mieee is specified
249 and then ignore it; eventually, we should either get it right or
250 disable the option altogether. */
254 if (TARGET_ABI_UNICOSMK)
255 warning ("-mieee not supported on Unicos/Mk");
258 alpha_tp = ALPHA_TP_INSN;
259 alpha_fptm = ALPHA_FPTM_SU;
263 if (TARGET_IEEE_WITH_INEXACT)
265 if (TARGET_ABI_UNICOSMK)
266 warning ("-mieee-with-inexact not supported on Unicos/Mk");
269 alpha_tp = ALPHA_TP_INSN;
270 alpha_fptm = ALPHA_FPTM_SUI;
276 if (! strcmp (alpha_tp_string, "p"))
277 alpha_tp = ALPHA_TP_PROG;
278 else if (! strcmp (alpha_tp_string, "f"))
279 alpha_tp = ALPHA_TP_FUNC;
280 else if (! strcmp (alpha_tp_string, "i"))
281 alpha_tp = ALPHA_TP_INSN;
283 error ("bad value `%s' for -mtrap-precision switch", alpha_tp_string);
286 if (alpha_fprm_string)
288 if (! strcmp (alpha_fprm_string, "n"))
289 alpha_fprm = ALPHA_FPRM_NORM;
290 else if (! strcmp (alpha_fprm_string, "m"))
291 alpha_fprm = ALPHA_FPRM_MINF;
292 else if (! strcmp (alpha_fprm_string, "c"))
293 alpha_fprm = ALPHA_FPRM_CHOP;
294 else if (! strcmp (alpha_fprm_string,"d"))
295 alpha_fprm = ALPHA_FPRM_DYN;
297 error ("bad value `%s' for -mfp-rounding-mode switch",
301 if (alpha_fptm_string)
303 if (strcmp (alpha_fptm_string, "n") == 0)
304 alpha_fptm = ALPHA_FPTM_N;
305 else if (strcmp (alpha_fptm_string, "u") == 0)
306 alpha_fptm = ALPHA_FPTM_U;
307 else if (strcmp (alpha_fptm_string, "su") == 0)
308 alpha_fptm = ALPHA_FPTM_SU;
309 else if (strcmp (alpha_fptm_string, "sui") == 0)
310 alpha_fptm = ALPHA_FPTM_SUI;
312 error ("bad value `%s' for -mfp-trap-mode switch", alpha_fptm_string);
315 if (alpha_tls_size_string)
317 if (strcmp (alpha_tls_size_string, "16") == 0)
319 else if (strcmp (alpha_tls_size_string, "32") == 0)
321 else if (strcmp (alpha_tls_size_string, "64") == 0)
324 error ("bad value `%s' for -mtls-size switch", alpha_tls_size_string);
328 = TARGET_CPU_DEFAULT & MASK_CPU_EV6 ? PROCESSOR_EV6
329 : (TARGET_CPU_DEFAULT & MASK_CPU_EV5 ? PROCESSOR_EV5 : PROCESSOR_EV4);
331 if (alpha_cpu_string)
333 for (i = 0; cpu_table [i].name; i++)
334 if (! strcmp (alpha_cpu_string, cpu_table [i].name))
336 alpha_cpu = cpu_table [i].processor;
337 target_flags &= ~ (MASK_BWX | MASK_MAX | MASK_FIX | MASK_CIX
338 | MASK_CPU_EV5 | MASK_CPU_EV6);
339 target_flags |= cpu_table [i].flags;
342 if (! cpu_table [i].name)
343 error ("bad value `%s' for -mcpu switch", alpha_cpu_string);
346 if (alpha_tune_string)
348 for (i = 0; cpu_table [i].name; i++)
349 if (! strcmp (alpha_tune_string, cpu_table [i].name))
351 alpha_cpu = cpu_table [i].processor;
354 if (! cpu_table [i].name)
355 error ("bad value `%s' for -mcpu switch", alpha_tune_string);
358 /* Do some sanity checks on the above options. */
360 if (TARGET_ABI_UNICOSMK && alpha_fptm != ALPHA_FPTM_N)
362 warning ("trap mode not supported on Unicos/Mk");
363 alpha_fptm = ALPHA_FPTM_N;
366 if ((alpha_fptm == ALPHA_FPTM_SU || alpha_fptm == ALPHA_FPTM_SUI)
367 && alpha_tp != ALPHA_TP_INSN && ! TARGET_CPU_EV6)
369 warning ("fp software completion requires -mtrap-precision=i");
370 alpha_tp = ALPHA_TP_INSN;
375 /* Except for EV6 pass 1 (not released), we always have precise
376 arithmetic traps. Which means we can do software completion
377 without minding trap shadows. */
378 alpha_tp = ALPHA_TP_PROG;
381 if (TARGET_FLOAT_VAX)
383 if (alpha_fprm == ALPHA_FPRM_MINF || alpha_fprm == ALPHA_FPRM_DYN)
385 warning ("rounding mode not supported for VAX floats");
386 alpha_fprm = ALPHA_FPRM_NORM;
388 if (alpha_fptm == ALPHA_FPTM_SUI)
390 warning ("trap mode not supported for VAX floats");
391 alpha_fptm = ALPHA_FPTM_SU;
399 if (!alpha_mlat_string)
400 alpha_mlat_string = "L1";
402 if (ISDIGIT ((unsigned char)alpha_mlat_string[0])
403 && (lat = strtol (alpha_mlat_string, &end, 10), *end == '\0'))
405 else if ((alpha_mlat_string[0] == 'L' || alpha_mlat_string[0] == 'l')
406 && ISDIGIT ((unsigned char)alpha_mlat_string[1])
407 && alpha_mlat_string[2] == '\0')
409 static int const cache_latency[][4] =
411 { 3, 30, -1 }, /* ev4 -- Bcache is a guess */
412 { 2, 12, 38 }, /* ev5 -- Bcache from PC164 LMbench numbers */
413 { 3, 12, 30 }, /* ev6 -- Bcache from DS20 LMbench. */
416 lat = alpha_mlat_string[1] - '0';
417 if (lat <= 0 || lat > 3 || cache_latency[alpha_cpu][lat-1] == -1)
419 warning ("L%d cache latency unknown for %s",
420 lat, alpha_cpu_name[alpha_cpu]);
424 lat = cache_latency[alpha_cpu][lat-1];
426 else if (! strcmp (alpha_mlat_string, "main"))
428 /* Most current memories have about 370ns latency. This is
429 a reasonable guess for a fast cpu. */
434 warning ("bad value `%s' for -mmemory-latency", alpha_mlat_string);
438 alpha_memory_latency = lat;
441 /* Default the definition of "small data" to 8 bytes. */
445 /* Infer TARGET_SMALL_DATA from -fpic/-fPIC. */
447 target_flags |= MASK_SMALL_DATA;
448 else if (flag_pic == 2)
449 target_flags &= ~MASK_SMALL_DATA;
451 /* Align labels and loops for optimal branching. */
452 /* ??? Kludge these by not doing anything if we don't optimize and also if
453 we are writing ECOFF symbols to work around a bug in DEC's assembler. */
454 if (optimize > 0 && write_symbols != SDB_DEBUG)
456 if (align_loops <= 0)
458 if (align_jumps <= 0)
461 if (align_functions <= 0)
462 align_functions = 16;
464 /* Acquire a unique set number for our register saves and restores. */
465 alpha_sr_alias_set = new_alias_set ();
467 /* Register variables and functions with the garbage collector. */
469 /* Set up function hooks. */
470 init_machine_status = alpha_init_machine_status;
472 /* Tell the compiler when we're using VAX floating point. */
473 if (TARGET_FLOAT_VAX)
475 real_format_for_mode[SFmode - QFmode] = &vax_f_format;
476 real_format_for_mode[DFmode - QFmode] = &vax_g_format;
477 real_format_for_mode[TFmode - QFmode] = NULL;
481 /* Returns 1 if VALUE is a mask that contains full bytes of zero or ones. */
484 zap_mask (HOST_WIDE_INT value)
488 for (i = 0; i < HOST_BITS_PER_WIDE_INT / HOST_BITS_PER_CHAR;
490 if ((value & 0xff) != 0 && (value & 0xff) != 0xff)
496 /* Returns 1 if OP is either the constant zero or a register. If a
497 register, it must be in the proper mode unless MODE is VOIDmode. */
500 reg_or_0_operand (rtx op, enum machine_mode mode)
502 return op == CONST0_RTX (mode) || register_operand (op, mode);
505 /* Return 1 if OP is a constant in the range of 0-63 (for a shift) or
509 reg_or_6bit_operand (rtx op, enum machine_mode mode)
511 return ((GET_CODE (op) == CONST_INT
512 && (unsigned HOST_WIDE_INT) INTVAL (op) < 64)
513 || register_operand (op, mode));
517 /* Return 1 if OP is an 8-bit constant or any register. */
520 reg_or_8bit_operand (rtx op, enum machine_mode mode)
522 return ((GET_CODE (op) == CONST_INT
523 && (unsigned HOST_WIDE_INT) INTVAL (op) < 0x100)
524 || register_operand (op, mode));
527 /* Return 1 if OP is a constant or any register. */
530 reg_or_const_int_operand (rtx op, enum machine_mode mode)
532 return GET_CODE (op) == CONST_INT || register_operand (op, mode);
535 /* Return 1 if OP is an 8-bit constant. */
538 cint8_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
540 return ((GET_CODE (op) == CONST_INT
541 && (unsigned HOST_WIDE_INT) INTVAL (op) < 0x100));
544 /* Return 1 if the operand is a valid second operand to an add insn. */
547 add_operand (rtx op, enum machine_mode mode)
549 if (GET_CODE (op) == CONST_INT)
550 /* Constraints I, J, O and P are covered by K. */
551 return (CONST_OK_FOR_LETTER_P (INTVAL (op), 'K')
552 || CONST_OK_FOR_LETTER_P (INTVAL (op), 'L'));
554 return register_operand (op, mode);
557 /* Return 1 if the operand is a valid second operand to a sign-extending
561 sext_add_operand (rtx op, enum machine_mode mode)
563 if (GET_CODE (op) == CONST_INT)
564 return (CONST_OK_FOR_LETTER_P (INTVAL (op), 'I')
565 || CONST_OK_FOR_LETTER_P (INTVAL (op), 'O'));
567 return reg_not_elim_operand (op, mode);
570 /* Return 1 if OP is the constant 4 or 8. */
573 const48_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
575 return (GET_CODE (op) == CONST_INT
576 && (INTVAL (op) == 4 || INTVAL (op) == 8));
579 /* Return 1 if OP is a valid first operand to an AND insn. */
582 and_operand (rtx op, enum machine_mode mode)
584 if (GET_CODE (op) == CONST_DOUBLE && GET_MODE (op) == VOIDmode)
585 return (zap_mask (CONST_DOUBLE_LOW (op))
586 && zap_mask (CONST_DOUBLE_HIGH (op)));
588 if (GET_CODE (op) == CONST_INT)
589 return ((unsigned HOST_WIDE_INT) INTVAL (op) < 0x100
590 || (unsigned HOST_WIDE_INT) ~ INTVAL (op) < 0x100
591 || zap_mask (INTVAL (op)));
593 return register_operand (op, mode);
596 /* Return 1 if OP is a valid first operand to an IOR or XOR insn. */
599 or_operand (rtx op, enum machine_mode mode)
601 if (GET_CODE (op) == CONST_INT)
602 return ((unsigned HOST_WIDE_INT) INTVAL (op) < 0x100
603 || (unsigned HOST_WIDE_INT) ~ INTVAL (op) < 0x100);
605 return register_operand (op, mode);
608 /* Return 1 if OP is a constant that is the width, in bits, of an integral
609 mode smaller than DImode. */
612 mode_width_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
614 return (GET_CODE (op) == CONST_INT
615 && (INTVAL (op) == 8 || INTVAL (op) == 16
616 || INTVAL (op) == 32 || INTVAL (op) == 64));
619 /* Return 1 if OP is a constant that is the width of an integral machine mode
620 smaller than an integer. */
623 mode_mask_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
625 if (GET_CODE (op) == CONST_INT)
627 HOST_WIDE_INT value = INTVAL (op);
633 if (value == 0xffffffff)
638 else if (HOST_BITS_PER_WIDE_INT == 32 && GET_CODE (op) == CONST_DOUBLE)
640 if (CONST_DOUBLE_LOW (op) == 0xffffffff && CONST_DOUBLE_HIGH (op) == 0)
647 /* Return 1 if OP is a multiple of 8 less than 64. */
650 mul8_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
652 return (GET_CODE (op) == CONST_INT
653 && (unsigned HOST_WIDE_INT) INTVAL (op) < 64
654 && (INTVAL (op) & 7) == 0);
657 /* Return 1 if OP is the zero constant for MODE. */
660 const0_operand (rtx op, enum machine_mode mode)
662 return op == CONST0_RTX (mode);
665 /* Return 1 if OP is a hard floating-point register. */
668 hard_fp_register_operand (rtx op, enum machine_mode mode)
670 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
673 if (GET_CODE (op) == SUBREG)
674 op = SUBREG_REG (op);
675 return GET_CODE (op) == REG && REGNO_REG_CLASS (REGNO (op)) == FLOAT_REGS;
678 /* Return 1 if OP is a hard general register. */
681 hard_int_register_operand (rtx op, enum machine_mode mode)
683 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
686 if (GET_CODE (op) == SUBREG)
687 op = SUBREG_REG (op);
688 return GET_CODE (op) == REG && REGNO_REG_CLASS (REGNO (op)) == GENERAL_REGS;
691 /* Return 1 if OP is a register or a constant integer. */
695 reg_or_cint_operand (rtx op, enum machine_mode mode)
697 return (GET_CODE (op) == CONST_INT
698 || register_operand (op, mode));
701 /* Return 1 if OP is something that can be reloaded into a register;
702 if it is a MEM, it need not be valid. */
705 some_operand (rtx op, enum machine_mode mode)
707 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
710 switch (GET_CODE (op))
724 return some_operand (SUBREG_REG (op), VOIDmode);
733 /* Likewise, but don't accept constants. */
736 some_ni_operand (rtx op, enum machine_mode mode)
738 if (GET_MODE (op) != mode && mode != VOIDmode)
741 if (GET_CODE (op) == SUBREG)
742 op = SUBREG_REG (op);
744 return (GET_CODE (op) == REG || GET_CODE (op) == MEM);
747 /* Return 1 if OP is a valid operand for the source of a move insn. */
750 input_operand (rtx op, enum machine_mode mode)
752 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
755 if (GET_MODE_CLASS (mode) == MODE_FLOAT && GET_MODE (op) != mode)
758 switch (GET_CODE (op))
763 if (TARGET_EXPLICIT_RELOCS)
765 /* We don't split symbolic operands into something unintelligable
766 until after reload, but we do not wish non-small, non-global
767 symbolic operands to be reconstructed from their high/lo_sum
769 return (small_symbolic_operand (op, mode)
770 || global_symbolic_operand (op, mode)
771 || gotdtp_symbolic_operand (op, mode)
772 || gottp_symbolic_operand (op, mode));
775 /* This handles both the Windows/NT and OSF cases. */
776 return mode == ptr_mode || mode == DImode;
779 return (TARGET_EXPLICIT_RELOCS
780 && local_symbolic_operand (XEXP (op, 0), mode));
787 if (register_operand (op, mode))
789 /* ... fall through ... */
791 return ((TARGET_BWX || (mode != HImode && mode != QImode))
792 && general_operand (op, mode));
796 return op == CONST0_RTX (mode);
799 return mode == QImode || mode == HImode || add_operand (op, mode);
811 /* Return 1 if OP is a SYMBOL_REF for a function known to be in this
812 file, and in the same section as the current function. */
815 samegp_function_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
817 if (GET_CODE (op) != SYMBOL_REF)
820 /* Easy test for recursion. */
821 if (op == XEXP (DECL_RTL (current_function_decl), 0))
824 /* Functions that are not local can be overridden, and thus may
825 not share the same gp. */
826 if (! SYMBOL_REF_LOCAL_P (op))
829 /* If -msmall-data is in effect, assume that there is only one GP
830 for the module, and so any local symbol has this property. We
831 need explicit relocations to be able to enforce this for symbols
832 not defined in this unit of translation, however. */
833 if (TARGET_EXPLICIT_RELOCS && TARGET_SMALL_DATA)
836 /* Functions that are not external are defined in this UoT,
837 and thus must share the same gp. */
838 return ! SYMBOL_REF_EXTERNAL_P (op);
841 /* Return 1 if OP is a SYMBOL_REF for which we can make a call via bsr. */
844 direct_call_operand (rtx op, enum machine_mode mode)
846 tree op_decl, cfun_sec, op_sec;
848 /* Must share the same GP. */
849 if (!samegp_function_operand (op, mode))
852 /* If profiling is implemented via linker tricks, we can't jump
853 to the nogp alternate entry point. Note that current_function_profile
854 would not be correct, since that doesn't indicate if the target
855 function uses profiling. */
856 /* ??? TARGET_PROFILING_NEEDS_GP isn't really the right test,
857 but is approximately correct for the OSF ABIs. Don't know
858 what to do for VMS, NT, or UMK. */
859 if (!TARGET_PROFILING_NEEDS_GP && profile_flag)
862 /* Must be a function. In some cases folks create thunks in static
863 data structures and then make calls to them. If we allow the
864 direct call, we'll get an error from the linker about !samegp reloc
865 against a symbol without a .prologue directive. */
866 if (!SYMBOL_REF_FUNCTION_P (op))
869 /* Must be "near" so that the branch is assumed to reach. With
870 -msmall-text, this is assumed true of all local symbols. Since
871 we've already checked samegp, locality is already assured. */
872 if (TARGET_SMALL_TEXT)
875 /* Otherwise, a decl is "near" if it is defined in the same section. */
876 if (flag_function_sections)
879 op_decl = SYMBOL_REF_DECL (op);
880 if (DECL_ONE_ONLY (current_function_decl)
881 || (op_decl && DECL_ONE_ONLY (op_decl)))
884 cfun_sec = DECL_SECTION_NAME (current_function_decl);
885 op_sec = op_decl ? DECL_SECTION_NAME (op_decl) : NULL;
886 return ((!cfun_sec && !op_sec)
887 || (cfun_sec && op_sec
888 && strcmp (TREE_STRING_POINTER (cfun_sec),
889 TREE_STRING_POINTER (op_sec)) == 0));
892 /* Return true if OP is a LABEL_REF, or SYMBOL_REF or CONST referencing
893 a (non-tls) variable known to be defined in this file. */
896 local_symbolic_operand (rtx op, enum machine_mode mode)
898 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
901 if (GET_CODE (op) == LABEL_REF)
904 if (GET_CODE (op) == CONST
905 && GET_CODE (XEXP (op, 0)) == PLUS
906 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT)
907 op = XEXP (XEXP (op, 0), 0);
909 if (GET_CODE (op) != SYMBOL_REF)
912 return SYMBOL_REF_LOCAL_P (op) && !SYMBOL_REF_TLS_MODEL (op);
915 /* Return true if OP is a SYMBOL_REF or CONST referencing a variable
916 known to be defined in this file in the small data area. */
919 small_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
921 if (! TARGET_SMALL_DATA)
924 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
927 if (GET_CODE (op) == CONST
928 && GET_CODE (XEXP (op, 0)) == PLUS
929 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT)
930 op = XEXP (XEXP (op, 0), 0);
932 if (GET_CODE (op) != SYMBOL_REF)
935 /* ??? There's no encode_section_info equivalent for the rtl
936 constant pool, so SYMBOL_FLAG_SMALL never gets set. */
937 if (CONSTANT_POOL_ADDRESS_P (op))
938 return GET_MODE_SIZE (get_pool_mode (op)) <= g_switch_value;
940 return (SYMBOL_REF_LOCAL_P (op)
941 && SYMBOL_REF_SMALL_P (op)
942 && SYMBOL_REF_TLS_MODEL (op) == 0);
945 /* Return true if OP is a SYMBOL_REF or CONST referencing a variable
946 not known (or known not) to be defined in this file. */
949 global_symbolic_operand (rtx op, enum machine_mode mode)
951 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
954 if (GET_CODE (op) == CONST
955 && GET_CODE (XEXP (op, 0)) == PLUS
956 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT)
957 op = XEXP (XEXP (op, 0), 0);
959 if (GET_CODE (op) != SYMBOL_REF)
962 return !SYMBOL_REF_LOCAL_P (op) && !SYMBOL_REF_TLS_MODEL (op);
965 /* Return 1 if OP is a valid operand for the MEM of a CALL insn. */
968 call_operand (rtx op, enum machine_mode mode)
973 if (GET_CODE (op) == REG)
977 /* Disallow virtual registers to cope with pathological test cases
978 such as compile/930117-1.c in which the virtual reg decomposes
979 to the frame pointer. Which is a hard reg that is not $27. */
980 return (REGNO (op) == 27 || REGNO (op) > LAST_VIRTUAL_REGISTER);
985 if (TARGET_ABI_UNICOSMK)
987 if (GET_CODE (op) == SYMBOL_REF)
993 /* Returns 1 if OP is a symbolic operand, i.e. a symbol_ref or a label_ref,
994 possibly with an offset. */
997 symbolic_operand (rtx op, enum machine_mode mode)
999 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
1001 if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
1003 if (GET_CODE (op) == CONST
1004 && GET_CODE (XEXP (op,0)) == PLUS
1005 && GET_CODE (XEXP (XEXP (op,0), 0)) == SYMBOL_REF
1006 && GET_CODE (XEXP (XEXP (op,0), 1)) == CONST_INT)
1011 /* Return true if OP is valid for a particular TLS relocation. */
1014 tls_symbolic_operand_1 (rtx op, enum machine_mode mode, int size, int unspec)
1016 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
1019 if (GET_CODE (op) != CONST)
1023 if (GET_CODE (op) != UNSPEC || XINT (op, 1) != unspec)
1025 op = XVECEXP (op, 0, 0);
1027 if (GET_CODE (op) != SYMBOL_REF)
1030 if (SYMBOL_REF_LOCAL_P (op))
1032 if (alpha_tls_size > size)
1041 switch (SYMBOL_REF_TLS_MODEL (op))
1043 case TLS_MODEL_LOCAL_DYNAMIC:
1044 return unspec == UNSPEC_DTPREL;
1045 case TLS_MODEL_INITIAL_EXEC:
1046 return unspec == UNSPEC_TPREL && size == 64;
1047 case TLS_MODEL_LOCAL_EXEC:
1048 return unspec == UNSPEC_TPREL;
1054 /* Return true if OP is valid for 16-bit DTP relative relocations. */
1057 dtp16_symbolic_operand (rtx op, enum machine_mode mode)
1059 return tls_symbolic_operand_1 (op, mode, 16, UNSPEC_DTPREL);
1062 /* Return true if OP is valid for 32-bit DTP relative relocations. */
1065 dtp32_symbolic_operand (rtx op, enum machine_mode mode)
1067 return tls_symbolic_operand_1 (op, mode, 32, UNSPEC_DTPREL);
1070 /* Return true if OP is valid for 64-bit DTP relative relocations. */
1073 gotdtp_symbolic_operand (rtx op, enum machine_mode mode)
1075 return tls_symbolic_operand_1 (op, mode, 64, UNSPEC_DTPREL);
1078 /* Return true if OP is valid for 16-bit TP relative relocations. */
1081 tp16_symbolic_operand (rtx op, enum machine_mode mode)
1083 return tls_symbolic_operand_1 (op, mode, 16, UNSPEC_TPREL);
1086 /* Return true if OP is valid for 32-bit TP relative relocations. */
1089 tp32_symbolic_operand (rtx op, enum machine_mode mode)
1091 return tls_symbolic_operand_1 (op, mode, 32, UNSPEC_TPREL);
1094 /* Return true if OP is valid for 64-bit TP relative relocations. */
1097 gottp_symbolic_operand (rtx op, enum machine_mode mode)
1099 return tls_symbolic_operand_1 (op, mode, 64, UNSPEC_TPREL);
1102 /* Return 1 if OP is a valid Alpha comparison operator. Here we know which
1103 comparisons are valid in which insn. */
1106 alpha_comparison_operator (rtx op, enum machine_mode mode)
1108 enum rtx_code code = GET_CODE (op);
1110 if (mode != GET_MODE (op) && mode != VOIDmode)
1113 return (code == EQ || code == LE || code == LT
1114 || code == LEU || code == LTU);
1117 /* Return 1 if OP is a valid Alpha comparison operator against zero.
1118 Here we know which comparisons are valid in which insn. */
1121 alpha_zero_comparison_operator (rtx op, enum machine_mode mode)
1123 enum rtx_code code = GET_CODE (op);
1125 if (mode != GET_MODE (op) && mode != VOIDmode)
1128 return (code == EQ || code == NE || code == LE || code == LT
1129 || code == LEU || code == LTU);
1132 /* Return 1 if OP is a valid Alpha swapped comparison operator. */
1135 alpha_swapped_comparison_operator (rtx op, enum machine_mode mode)
1137 enum rtx_code code = GET_CODE (op);
1139 if ((mode != GET_MODE (op) && mode != VOIDmode)
1140 || GET_RTX_CLASS (code) != '<')
1143 code = swap_condition (code);
1144 return (code == EQ || code == LE || code == LT
1145 || code == LEU || code == LTU);
1148 /* Return 1 if OP is a signed comparison operation. */
1151 signed_comparison_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1153 enum rtx_code code = GET_CODE (op);
1155 if (mode != GET_MODE (op) && mode != VOIDmode)
1158 return (code == EQ || code == NE
1159 || code == LE || code == LT
1160 || code == GE || code == GT);
1163 /* Return 1 if OP is a valid Alpha floating point comparison operator.
1164 Here we know which comparisons are valid in which insn. */
1167 alpha_fp_comparison_operator (rtx op, enum machine_mode mode)
1169 enum rtx_code code = GET_CODE (op);
1171 if (mode != GET_MODE (op) && mode != VOIDmode)
1174 return (code == EQ || code == LE || code == LT || code == UNORDERED);
1177 /* Return 1 if this is a divide or modulus operator. */
1180 divmod_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1182 switch (GET_CODE (op))
1184 case DIV: case MOD: case UDIV: case UMOD:
1194 /* Return 1 if this memory address is a known aligned register plus
1195 a constant. It must be a valid address. This means that we can do
1196 this as an aligned reference plus some offset.
1198 Take into account what reload will do. */
1201 aligned_memory_operand (rtx op, enum machine_mode mode)
1205 if (reload_in_progress)
1208 if (GET_CODE (tmp) == SUBREG)
1209 tmp = SUBREG_REG (tmp);
1210 if (GET_CODE (tmp) == REG
1211 && REGNO (tmp) >= FIRST_PSEUDO_REGISTER)
1213 op = reg_equiv_memory_loc[REGNO (tmp)];
1219 if (GET_CODE (op) != MEM
1220 || GET_MODE (op) != mode)
1224 /* LEGITIMIZE_RELOAD_ADDRESS creates (plus (plus reg const_hi) const_lo)
1225 sorts of constructs. Dig for the real base register. */
1226 if (reload_in_progress
1227 && GET_CODE (op) == PLUS
1228 && GET_CODE (XEXP (op, 0)) == PLUS)
1229 base = XEXP (XEXP (op, 0), 0);
1232 if (! memory_address_p (mode, op))
1234 base = (GET_CODE (op) == PLUS ? XEXP (op, 0) : op);
1237 return (GET_CODE (base) == REG && REGNO_POINTER_ALIGN (REGNO (base)) >= 32);
1240 /* Similar, but return 1 if OP is a MEM which is not alignable. */
1243 unaligned_memory_operand (rtx op, enum machine_mode mode)
1247 if (reload_in_progress)
1250 if (GET_CODE (tmp) == SUBREG)
1251 tmp = SUBREG_REG (tmp);
1252 if (GET_CODE (tmp) == REG
1253 && REGNO (tmp) >= FIRST_PSEUDO_REGISTER)
1255 op = reg_equiv_memory_loc[REGNO (tmp)];
1261 if (GET_CODE (op) != MEM
1262 || GET_MODE (op) != mode)
1266 /* LEGITIMIZE_RELOAD_ADDRESS creates (plus (plus reg const_hi) const_lo)
1267 sorts of constructs. Dig for the real base register. */
1268 if (reload_in_progress
1269 && GET_CODE (op) == PLUS
1270 && GET_CODE (XEXP (op, 0)) == PLUS)
1271 base = XEXP (XEXP (op, 0), 0);
1274 if (! memory_address_p (mode, op))
1276 base = (GET_CODE (op) == PLUS ? XEXP (op, 0) : op);
1279 return (GET_CODE (base) == REG && REGNO_POINTER_ALIGN (REGNO (base)) < 32);
1282 /* Return 1 if OP is either a register or an unaligned memory location. */
1285 reg_or_unaligned_mem_operand (rtx op, enum machine_mode mode)
1287 return register_operand (op, mode) || unaligned_memory_operand (op, mode);
1290 /* Return 1 if OP is any memory location. During reload a pseudo matches. */
1293 any_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1295 return (GET_CODE (op) == MEM
1296 || (GET_CODE (op) == SUBREG && GET_CODE (SUBREG_REG (op)) == REG)
1297 || (reload_in_progress && GET_CODE (op) == REG
1298 && REGNO (op) >= FIRST_PSEUDO_REGISTER)
1299 || (reload_in_progress && GET_CODE (op) == SUBREG
1300 && GET_CODE (SUBREG_REG (op)) == REG
1301 && REGNO (SUBREG_REG (op)) >= FIRST_PSEUDO_REGISTER));
1304 /* Returns 1 if OP is not an eliminable register.
1306 This exists to cure a pathological abort in the s8addq (et al) patterns,
1308 long foo () { long t; bar(); return (long) &t * 26107; }
1310 which run afoul of a hack in reload to cure a (presumably) similar
1311 problem with lea-type instructions on other targets. But there is
1312 one of us and many of them, so work around the problem by selectively
1313 preventing combine from making the optimization. */
1316 reg_not_elim_operand (rtx op, enum machine_mode mode)
1319 if (GET_CODE (op) == SUBREG)
1320 inner = SUBREG_REG (op);
1321 if (inner == frame_pointer_rtx || inner == arg_pointer_rtx)
1324 return register_operand (op, mode);
1327 /* Return 1 is OP is a memory location that is not a reference (using
1328 an AND) to an unaligned location. Take into account what reload
1332 normal_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1334 if (reload_in_progress)
1337 if (GET_CODE (tmp) == SUBREG)
1338 tmp = SUBREG_REG (tmp);
1339 if (GET_CODE (tmp) == REG
1340 && REGNO (tmp) >= FIRST_PSEUDO_REGISTER)
1342 op = reg_equiv_memory_loc[REGNO (tmp)];
1344 /* This may not have been assigned an equivalent address if it will
1345 be eliminated. In that case, it doesn't matter what we do. */
1351 return GET_CODE (op) == MEM && GET_CODE (XEXP (op, 0)) != AND;
1354 /* Accept a register, but not a subreg of any kind. This allows us to
1355 avoid pathological cases in reload wrt data movement common in
1356 int->fp conversion. */
1359 reg_no_subreg_operand (rtx op, enum machine_mode mode)
1361 if (GET_CODE (op) != REG)
1363 return register_operand (op, mode);
1366 /* Recognize an addition operation that includes a constant. Used to
1367 convince reload to canonize (plus (plus reg c1) c2) during register
1371 addition_operation (rtx op, enum machine_mode mode)
1373 if (GET_MODE (op) != mode && mode != VOIDmode)
1375 if (GET_CODE (op) == PLUS
1376 && register_operand (XEXP (op, 0), mode)
1377 && GET_CODE (XEXP (op, 1)) == CONST_INT
1378 && CONST_OK_FOR_LETTER_P (INTVAL (XEXP (op, 1)), 'K'))
1383 /* Implements CONST_OK_FOR_LETTER_P. Return true if the value matches
1384 the range defined for C in [I-P]. */
1387 alpha_const_ok_for_letter_p (HOST_WIDE_INT value, int c)
1392 /* An unsigned 8 bit constant. */
1393 return (unsigned HOST_WIDE_INT) value < 0x100;
1395 /* The constant zero. */
1398 /* A signed 16 bit constant. */
1399 return (unsigned HOST_WIDE_INT) (value + 0x8000) < 0x10000;
1401 /* A shifted signed 16 bit constant appropriate for LDAH. */
1402 return ((value & 0xffff) == 0
1403 && ((value) >> 31 == -1 || value >> 31 == 0));
1405 /* A constant that can be AND'ed with using a ZAP insn. */
1406 return zap_mask (value);
1408 /* A complemented unsigned 8 bit constant. */
1409 return (unsigned HOST_WIDE_INT) (~ value) < 0x100;
1411 /* A negated unsigned 8 bit constant. */
1412 return (unsigned HOST_WIDE_INT) (- value) < 0x100;
1414 /* The constant 1, 2 or 3. */
1415 return value == 1 || value == 2 || value == 3;
1422 /* Implements CONST_DOUBLE_OK_FOR_LETTER_P. Return true if VALUE
1423 matches for C in [GH]. */
1426 alpha_const_double_ok_for_letter_p (rtx value, int c)
1431 /* The floating point zero constant. */
1432 return (GET_MODE_CLASS (GET_MODE (value)) == MODE_FLOAT
1433 && value == CONST0_RTX (GET_MODE (value)));
1436 /* A valid operand of a ZAP insn. */
1437 return (GET_MODE (value) == VOIDmode
1438 && zap_mask (CONST_DOUBLE_LOW (value))
1439 && zap_mask (CONST_DOUBLE_HIGH (value)));
1446 /* Implements CONST_DOUBLE_OK_FOR_LETTER_P. Return true if VALUE
1450 alpha_extra_constraint (rtx value, int c)
1455 return normal_memory_operand (value, VOIDmode);
1457 return direct_call_operand (value, Pmode);
1459 return (GET_CODE (value) == CONST_INT
1460 && (unsigned HOST_WIDE_INT) INTVAL (value) < 64);
1462 return GET_CODE (value) == HIGH;
1464 return TARGET_ABI_UNICOSMK && symbolic_operand (value, VOIDmode);
1466 return (GET_CODE (value) == CONST_VECTOR
1467 && value == CONST0_RTX (GET_MODE (value)));
1473 /* Return 1 if this function can directly return via $26. */
1476 direct_return (void)
1478 return (! TARGET_ABI_OPEN_VMS && ! TARGET_ABI_UNICOSMK
1480 && alpha_sa_size () == 0
1481 && get_frame_size () == 0
1482 && current_function_outgoing_args_size == 0
1483 && current_function_pretend_args_size == 0);
1486 /* Return the ADDR_VEC associated with a tablejump insn. */
1489 alpha_tablejump_addr_vec (rtx insn)
1493 tmp = JUMP_LABEL (insn);
1496 tmp = NEXT_INSN (tmp);
1499 if (GET_CODE (tmp) == JUMP_INSN
1500 && GET_CODE (PATTERN (tmp)) == ADDR_DIFF_VEC)
1501 return PATTERN (tmp);
1505 /* Return the label of the predicted edge, or CONST0_RTX if we don't know. */
1508 alpha_tablejump_best_label (rtx insn)
1510 rtx jump_table = alpha_tablejump_addr_vec (insn);
1511 rtx best_label = NULL_RTX;
1513 /* ??? Once the CFG doesn't keep getting completely rebuilt, look
1514 there for edge frequency counts from profile data. */
1518 int n_labels = XVECLEN (jump_table, 1);
1519 int best_count = -1;
1522 for (i = 0; i < n_labels; i++)
1526 for (j = i + 1; j < n_labels; j++)
1527 if (XEXP (XVECEXP (jump_table, 1, i), 0)
1528 == XEXP (XVECEXP (jump_table, 1, j), 0))
1531 if (count > best_count)
1532 best_count = count, best_label = XVECEXP (jump_table, 1, i);
1536 return best_label ? best_label : const0_rtx;
1539 /* Return the TLS model to use for SYMBOL. */
1541 static enum tls_model
1542 tls_symbolic_operand_type (rtx symbol)
1544 enum tls_model model;
1546 if (GET_CODE (symbol) != SYMBOL_REF)
1548 model = SYMBOL_REF_TLS_MODEL (symbol);
1550 /* Local-exec with a 64-bit size is the same code as initial-exec. */
1551 if (model == TLS_MODEL_LOCAL_EXEC && alpha_tls_size == 64)
1552 model = TLS_MODEL_INITIAL_EXEC;
1557 /* Return true if the function DECL will share the same GP as any
1558 function in the current unit of translation. */
1561 decl_has_samegp (tree decl)
1563 /* Functions that are not local can be overridden, and thus may
1564 not share the same gp. */
1565 if (!(*targetm.binds_local_p) (decl))
1568 /* If -msmall-data is in effect, assume that there is only one GP
1569 for the module, and so any local symbol has this property. We
1570 need explicit relocations to be able to enforce this for symbols
1571 not defined in this unit of translation, however. */
1572 if (TARGET_EXPLICIT_RELOCS && TARGET_SMALL_DATA)
1575 /* Functions that are not external are defined in this UoT. */
1576 /* ??? Irritatingly, static functions not yet emitted are still
1577 marked "external". Apply this to non-static functions only. */
1578 return !TREE_PUBLIC (decl) || !DECL_EXTERNAL (decl);
1581 /* Return true if EXP should be placed in the small data section. */
1584 alpha_in_small_data_p (tree exp)
1586 /* We want to merge strings, so we never consider them small data. */
1587 if (TREE_CODE (exp) == STRING_CST)
1590 if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
1592 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));
1593 if (strcmp (section, ".sdata") == 0
1594 || strcmp (section, ".sbss") == 0)
1599 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
1601 /* If this is an incomplete type with size 0, then we can't put it
1602 in sdata because it might be too big when completed. */
1603 if (size > 0 && (unsigned HOST_WIDE_INT) size <= g_switch_value)
1610 #if TARGET_ABI_OPEN_VMS
1612 alpha_linkage_symbol_p (const char *symname)
1614 int symlen = strlen (symname);
1617 return strcmp (&symname [symlen - 4], "..lk") == 0;
1622 #define LINKAGE_SYMBOL_REF_P(X) \
1623 ((GET_CODE (X) == SYMBOL_REF \
1624 && alpha_linkage_symbol_p (XSTR (X, 0))) \
1625 || (GET_CODE (X) == CONST \
1626 && GET_CODE (XEXP (X, 0)) == PLUS \
1627 && GET_CODE (XEXP (XEXP (X, 0), 0)) == SYMBOL_REF \
1628 && alpha_linkage_symbol_p (XSTR (XEXP (XEXP (X, 0), 0), 0))))
1631 /* legitimate_address_p recognizes an RTL expression that is a valid
1632 memory address for an instruction. The MODE argument is the
1633 machine mode for the MEM expression that wants to use this address.
1635 For Alpha, we have either a constant address or the sum of a
1636 register and a constant address, or just a register. For DImode,
1637 any of those forms can be surrounded with an AND that clear the
1638 low-order three bits; this is an "unaligned" access. */
1641 alpha_legitimate_address_p (enum machine_mode mode, rtx x, int strict)
1643 /* If this is an ldq_u type address, discard the outer AND. */
1645 && GET_CODE (x) == AND
1646 && GET_CODE (XEXP (x, 1)) == CONST_INT
1647 && INTVAL (XEXP (x, 1)) == -8)
1650 /* Discard non-paradoxical subregs. */
1651 if (GET_CODE (x) == SUBREG
1652 && (GET_MODE_SIZE (GET_MODE (x))
1653 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))))
1656 /* Unadorned general registers are valid. */
1659 ? STRICT_REG_OK_FOR_BASE_P (x)
1660 : NONSTRICT_REG_OK_FOR_BASE_P (x)))
1663 /* Constant addresses (i.e. +/- 32k) are valid. */
1664 if (CONSTANT_ADDRESS_P (x))
1667 #if TARGET_ABI_OPEN_VMS
1668 if (LINKAGE_SYMBOL_REF_P (x))
1672 /* Register plus a small constant offset is valid. */
1673 if (GET_CODE (x) == PLUS)
1675 rtx ofs = XEXP (x, 1);
1678 /* Discard non-paradoxical subregs. */
1679 if (GET_CODE (x) == SUBREG
1680 && (GET_MODE_SIZE (GET_MODE (x))
1681 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))))
1687 && NONSTRICT_REG_OK_FP_BASE_P (x)
1688 && GET_CODE (ofs) == CONST_INT)
1691 ? STRICT_REG_OK_FOR_BASE_P (x)
1692 : NONSTRICT_REG_OK_FOR_BASE_P (x))
1693 && CONSTANT_ADDRESS_P (ofs))
1696 else if (GET_CODE (x) == ADDRESSOF
1697 && GET_CODE (ofs) == CONST_INT)
1701 /* If we're managing explicit relocations, LO_SUM is valid, as
1702 are small data symbols. */
1703 else if (TARGET_EXPLICIT_RELOCS)
1705 if (small_symbolic_operand (x, Pmode))
1708 if (GET_CODE (x) == LO_SUM)
1710 rtx ofs = XEXP (x, 1);
1713 /* Discard non-paradoxical subregs. */
1714 if (GET_CODE (x) == SUBREG
1715 && (GET_MODE_SIZE (GET_MODE (x))
1716 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))))
1719 /* Must have a valid base register. */
1722 ? STRICT_REG_OK_FOR_BASE_P (x)
1723 : NONSTRICT_REG_OK_FOR_BASE_P (x))))
1726 /* The symbol must be local. */
1727 if (local_symbolic_operand (ofs, Pmode)
1728 || dtp32_symbolic_operand (ofs, Pmode)
1729 || tp32_symbolic_operand (ofs, Pmode))
1737 /* Build the SYMBOL_REF for __tls_get_addr. */
1739 static GTY(()) rtx tls_get_addr_libfunc;
1742 get_tls_get_addr (void)
1744 if (!tls_get_addr_libfunc)
1745 tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr");
1746 return tls_get_addr_libfunc;
1749 /* Try machine-dependent ways of modifying an illegitimate address
1750 to be legitimate. If we find one, return the new, valid address. */
1753 alpha_legitimize_address (rtx x, rtx scratch,
1754 enum machine_mode mode ATTRIBUTE_UNUSED)
1756 HOST_WIDE_INT addend;
1758 /* If the address is (plus reg const_int) and the CONST_INT is not a
1759 valid offset, compute the high part of the constant and add it to
1760 the register. Then our address is (plus temp low-part-const). */
1761 if (GET_CODE (x) == PLUS
1762 && GET_CODE (XEXP (x, 0)) == REG
1763 && GET_CODE (XEXP (x, 1)) == CONST_INT
1764 && ! CONSTANT_ADDRESS_P (XEXP (x, 1)))
1766 addend = INTVAL (XEXP (x, 1));
1771 /* If the address is (const (plus FOO const_int)), find the low-order
1772 part of the CONST_INT. Then load FOO plus any high-order part of the
1773 CONST_INT into a register. Our address is (plus reg low-part-const).
1774 This is done to reduce the number of GOT entries. */
1776 && GET_CODE (x) == CONST
1777 && GET_CODE (XEXP (x, 0)) == PLUS
1778 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
1780 addend = INTVAL (XEXP (XEXP (x, 0), 1));
1781 x = force_reg (Pmode, XEXP (XEXP (x, 0), 0));
1785 /* If we have a (plus reg const), emit the load as in (2), then add
1786 the two registers, and finally generate (plus reg low-part-const) as
1789 && GET_CODE (x) == PLUS
1790 && GET_CODE (XEXP (x, 0)) == REG
1791 && GET_CODE (XEXP (x, 1)) == CONST
1792 && GET_CODE (XEXP (XEXP (x, 1), 0)) == PLUS
1793 && GET_CODE (XEXP (XEXP (XEXP (x, 1), 0), 1)) == CONST_INT)
1795 addend = INTVAL (XEXP (XEXP (XEXP (x, 1), 0), 1));
1796 x = expand_simple_binop (Pmode, PLUS, XEXP (x, 0),
1797 XEXP (XEXP (XEXP (x, 1), 0), 0),
1798 NULL_RTX, 1, OPTAB_LIB_WIDEN);
1802 /* If this is a local symbol, split the address into HIGH/LO_SUM parts. */
1803 if (TARGET_EXPLICIT_RELOCS && symbolic_operand (x, Pmode))
1805 rtx r0, r16, eqv, tga, tp, insn, dest, seq;
1807 switch (tls_symbolic_operand_type (x))
1809 case TLS_MODEL_GLOBAL_DYNAMIC:
1812 r0 = gen_rtx_REG (Pmode, 0);
1813 r16 = gen_rtx_REG (Pmode, 16);
1814 tga = get_tls_get_addr ();
1815 dest = gen_reg_rtx (Pmode);
1816 seq = GEN_INT (alpha_next_sequence_number++);
1818 emit_insn (gen_movdi_er_tlsgd (r16, pic_offset_table_rtx, x, seq));
1819 insn = gen_call_value_osf_tlsgd (r0, tga, seq);
1820 insn = emit_call_insn (insn);
1821 CONST_OR_PURE_CALL_P (insn) = 1;
1822 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), r16);
1824 insn = get_insns ();
1827 emit_libcall_block (insn, dest, r0, x);
1830 case TLS_MODEL_LOCAL_DYNAMIC:
1833 r0 = gen_rtx_REG (Pmode, 0);
1834 r16 = gen_rtx_REG (Pmode, 16);
1835 tga = get_tls_get_addr ();
1836 scratch = gen_reg_rtx (Pmode);
1837 seq = GEN_INT (alpha_next_sequence_number++);
1839 emit_insn (gen_movdi_er_tlsldm (r16, pic_offset_table_rtx, seq));
1840 insn = gen_call_value_osf_tlsldm (r0, tga, seq);
1841 insn = emit_call_insn (insn);
1842 CONST_OR_PURE_CALL_P (insn) = 1;
1843 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), r16);
1845 insn = get_insns ();
1848 eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
1849 UNSPEC_TLSLDM_CALL);
1850 emit_libcall_block (insn, scratch, r0, eqv);
1852 eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), UNSPEC_DTPREL);
1853 eqv = gen_rtx_CONST (Pmode, eqv);
1855 if (alpha_tls_size == 64)
1857 dest = gen_reg_rtx (Pmode);
1858 emit_insn (gen_rtx_SET (VOIDmode, dest, eqv));
1859 emit_insn (gen_adddi3 (dest, dest, scratch));
1862 if (alpha_tls_size == 32)
1864 insn = gen_rtx_HIGH (Pmode, eqv);
1865 insn = gen_rtx_PLUS (Pmode, scratch, insn);
1866 scratch = gen_reg_rtx (Pmode);
1867 emit_insn (gen_rtx_SET (VOIDmode, scratch, insn));
1869 return gen_rtx_LO_SUM (Pmode, scratch, eqv);
1871 case TLS_MODEL_INITIAL_EXEC:
1872 eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), UNSPEC_TPREL);
1873 eqv = gen_rtx_CONST (Pmode, eqv);
1874 tp = gen_reg_rtx (Pmode);
1875 scratch = gen_reg_rtx (Pmode);
1876 dest = gen_reg_rtx (Pmode);
1878 emit_insn (gen_load_tp (tp));
1879 emit_insn (gen_rtx_SET (VOIDmode, scratch, eqv));
1880 emit_insn (gen_adddi3 (dest, tp, scratch));
1883 case TLS_MODEL_LOCAL_EXEC:
1884 eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), UNSPEC_TPREL);
1885 eqv = gen_rtx_CONST (Pmode, eqv);
1886 tp = gen_reg_rtx (Pmode);
1888 emit_insn (gen_load_tp (tp));
1889 if (alpha_tls_size == 32)
1891 insn = gen_rtx_HIGH (Pmode, eqv);
1892 insn = gen_rtx_PLUS (Pmode, tp, insn);
1893 tp = gen_reg_rtx (Pmode);
1894 emit_insn (gen_rtx_SET (VOIDmode, tp, insn));
1896 return gen_rtx_LO_SUM (Pmode, tp, eqv);
1899 if (local_symbolic_operand (x, Pmode))
1901 if (small_symbolic_operand (x, Pmode))
1905 if (!no_new_pseudos)
1906 scratch = gen_reg_rtx (Pmode);
1907 emit_insn (gen_rtx_SET (VOIDmode, scratch,
1908 gen_rtx_HIGH (Pmode, x)));
1909 return gen_rtx_LO_SUM (Pmode, scratch, x);
1918 HOST_WIDE_INT low, high;
1920 low = ((addend & 0xffff) ^ 0x8000) - 0x8000;
1922 high = ((addend & 0xffffffff) ^ 0x80000000) - 0x80000000;
1926 x = expand_simple_binop (Pmode, PLUS, x, GEN_INT (addend),
1927 (no_new_pseudos ? scratch : NULL_RTX),
1928 1, OPTAB_LIB_WIDEN);
1930 x = expand_simple_binop (Pmode, PLUS, x, GEN_INT (high),
1931 (no_new_pseudos ? scratch : NULL_RTX),
1932 1, OPTAB_LIB_WIDEN);
1934 return plus_constant (x, low);
1938 /* We do not allow indirect calls to be optimized into sibling calls, nor
1939 can we allow a call to a function with a different GP to be optimized
1943 alpha_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
1945 /* Can't do indirect tail calls, since we don't know if the target
1946 uses the same GP. */
1950 /* Otherwise, we can make a tail call if the target function shares
1952 return decl_has_samegp (decl);
1955 /* For TARGET_EXPLICIT_RELOCS, we don't obfuscate a SYMBOL_REF to a
1956 small symbolic operand until after reload. At which point we need
1957 to replace (mem (symbol_ref)) with (mem (lo_sum $29 symbol_ref))
1958 so that sched2 has the proper dependency information. */
1961 some_small_symbolic_operand_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
1965 /* Don't re-split. */
1966 if (GET_CODE (x) == LO_SUM)
1969 return small_symbolic_operand (x, Pmode) != 0;
1973 some_small_symbolic_operand (rtx x, enum machine_mode mode ATTRIBUTE_UNUSED)
1975 return for_each_rtx (&x, some_small_symbolic_operand_1, NULL);
1979 split_small_symbolic_operand_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
1983 /* Don't re-split. */
1984 if (GET_CODE (x) == LO_SUM)
1987 if (small_symbolic_operand (x, Pmode))
1989 x = gen_rtx_LO_SUM (Pmode, pic_offset_table_rtx, x);
1998 split_small_symbolic_operand (rtx x)
2001 for_each_rtx (&x, split_small_symbolic_operand_1, NULL);
2005 /* Indicate that INSN cannot be duplicated. This is true for any insn
2006 that we've marked with gpdisp relocs, since those have to stay in
2007 1-1 correspondence with one another.
2009 Technically we could copy them if we could set up a mapping from one
2010 sequence number to another, across the set of insns to be duplicated.
2011 This seems overly complicated and error-prone since interblock motion
2012 from sched-ebb could move one of the pair of insns to a different block. */
2015 alpha_cannot_copy_insn_p (rtx insn)
2019 if (!reload_completed || !TARGET_EXPLICIT_RELOCS)
2022 if (GET_CODE (insn) != INSN)
2024 if (asm_noperands (insn) >= 0)
2027 pat = PATTERN (insn);
2028 if (GET_CODE (pat) != SET)
2030 pat = SET_SRC (pat);
2031 if (GET_CODE (pat) == UNSPEC_VOLATILE)
2033 if (XINT (pat, 1) == UNSPECV_LDGP1
2034 || XINT (pat, 1) == UNSPECV_PLDGP2)
2037 else if (GET_CODE (pat) == UNSPEC)
2039 if (XINT (pat, 1) == UNSPEC_LDGP2)
2047 /* Try a machine-dependent way of reloading an illegitimate address
2048 operand. If we find one, push the reload and return the new rtx. */
2051 alpha_legitimize_reload_address (rtx x,
2052 enum machine_mode mode ATTRIBUTE_UNUSED,
2053 int opnum, int type,
2054 int ind_levels ATTRIBUTE_UNUSED)
2056 /* We must recognize output that we have already generated ourselves. */
2057 if (GET_CODE (x) == PLUS
2058 && GET_CODE (XEXP (x, 0)) == PLUS
2059 && GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
2060 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
2061 && GET_CODE (XEXP (x, 1)) == CONST_INT)
2063 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
2064 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
2069 /* We wish to handle large displacements off a base register by
2070 splitting the addend across an ldah and the mem insn. This
2071 cuts number of extra insns needed from 3 to 1. */
2072 if (GET_CODE (x) == PLUS
2073 && GET_CODE (XEXP (x, 0)) == REG
2074 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER
2075 && REGNO_OK_FOR_BASE_P (REGNO (XEXP (x, 0)))
2076 && GET_CODE (XEXP (x, 1)) == CONST_INT)
2078 HOST_WIDE_INT val = INTVAL (XEXP (x, 1));
2079 HOST_WIDE_INT low = ((val & 0xffff) ^ 0x8000) - 0x8000;
2081 = (((val - low) & 0xffffffff) ^ 0x80000000) - 0x80000000;
2083 /* Check for 32-bit overflow. */
2084 if (high + low != val)
2087 /* Reload the high part into a base reg; leave the low part
2088 in the mem directly. */
2089 x = gen_rtx_PLUS (GET_MODE (x),
2090 gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
2094 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
2095 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
2103 /* Compute a (partial) cost for rtx X. Return true if the complete
2104 cost has been computed, and false if subexpressions should be
2105 scanned. In either case, *TOTAL contains the cost result. */
2108 alpha_rtx_costs (rtx x, int code, int outer_code, int *total)
2110 enum machine_mode mode = GET_MODE (x);
2111 bool float_mode_p = FLOAT_MODE_P (mode);
2115 /* If this is an 8-bit constant, return zero since it can be used
2116 nearly anywhere with no cost. If it is a valid operand for an
2117 ADD or AND, likewise return 0 if we know it will be used in that
2118 context. Otherwise, return 2 since it might be used there later.
2119 All other constants take at least two insns. */
2121 if (INTVAL (x) >= 0 && INTVAL (x) < 256)
2129 if (x == CONST0_RTX (mode))
2131 else if ((outer_code == PLUS && add_operand (x, VOIDmode))
2132 || (outer_code == AND && and_operand (x, VOIDmode)))
2134 else if (add_operand (x, VOIDmode) || and_operand (x, VOIDmode))
2137 *total = COSTS_N_INSNS (2);
2143 if (TARGET_EXPLICIT_RELOCS && small_symbolic_operand (x, VOIDmode))
2144 *total = COSTS_N_INSNS (outer_code != MEM);
2145 else if (TARGET_EXPLICIT_RELOCS && local_symbolic_operand (x, VOIDmode))
2146 *total = COSTS_N_INSNS (1 + (outer_code != MEM));
2147 else if (tls_symbolic_operand_type (x))
2148 /* Estimate of cost for call_pal rduniq. */
2149 *total = COSTS_N_INSNS (15);
2151 /* Otherwise we do a load from the GOT. */
2152 *total = COSTS_N_INSNS (alpha_memory_latency);
2158 *total = alpha_rtx_cost_data[alpha_cpu].fp_add;
2159 else if (GET_CODE (XEXP (x, 0)) == MULT
2160 && const48_operand (XEXP (XEXP (x, 0), 1), VOIDmode))
2162 *total = (rtx_cost (XEXP (XEXP (x, 0), 0), outer_code)
2163 + rtx_cost (XEXP (x, 1), outer_code) + 2);
2170 *total = alpha_rtx_cost_data[alpha_cpu].fp_mult;
2171 else if (mode == DImode)
2172 *total = alpha_rtx_cost_data[alpha_cpu].int_mult_di;
2174 *total = alpha_rtx_cost_data[alpha_cpu].int_mult_si;
2178 if (GET_CODE (XEXP (x, 1)) == CONST_INT
2179 && INTVAL (XEXP (x, 1)) <= 3)
2181 *total = COSTS_N_INSNS (1);
2188 *total = alpha_rtx_cost_data[alpha_cpu].int_shift;
2193 *total = alpha_rtx_cost_data[alpha_cpu].fp_add;
2195 *total = alpha_rtx_cost_data[alpha_cpu].int_cmov;
2203 *total = COSTS_N_INSNS (70); /* ??? */
2204 else if (mode == SFmode)
2205 *total = alpha_rtx_cost_data[alpha_cpu].fp_div_sf;
2207 *total = alpha_rtx_cost_data[alpha_cpu].fp_div_df;
2211 *total = COSTS_N_INSNS (alpha_memory_latency);
2217 *total = COSTS_N_INSNS (1);
2225 *total = COSTS_N_INSNS (1) + alpha_rtx_cost_data[alpha_cpu].int_cmov;
2231 case UNSIGNED_FLOAT:
2235 case FLOAT_TRUNCATE:
2236 *total = alpha_rtx_cost_data[alpha_cpu].fp_add;
2244 /* REF is an alignable memory location. Place an aligned SImode
2245 reference into *PALIGNED_MEM and the number of bits to shift into
2246 *PBITNUM. SCRATCH is a free register for use in reloading out
2247 of range stack slots. */
2250 get_aligned_mem (rtx ref, rtx *paligned_mem, rtx *pbitnum)
2253 HOST_WIDE_INT offset = 0;
2255 if (GET_CODE (ref) != MEM)
2258 if (reload_in_progress
2259 && ! memory_address_p (GET_MODE (ref), XEXP (ref, 0)))
2261 base = find_replacement (&XEXP (ref, 0));
2263 if (! memory_address_p (GET_MODE (ref), base))
2268 base = XEXP (ref, 0);
2271 if (GET_CODE (base) == PLUS)
2272 offset += INTVAL (XEXP (base, 1)), base = XEXP (base, 0);
2275 = widen_memory_access (ref, SImode, (offset & ~3) - offset);
2277 if (WORDS_BIG_ENDIAN)
2278 *pbitnum = GEN_INT (32 - (GET_MODE_BITSIZE (GET_MODE (ref))
2279 + (offset & 3) * 8));
2281 *pbitnum = GEN_INT ((offset & 3) * 8);
2284 /* Similar, but just get the address. Handle the two reload cases.
2285 Add EXTRA_OFFSET to the address we return. */
2288 get_unaligned_address (rtx ref, int extra_offset)
2291 HOST_WIDE_INT offset = 0;
2293 if (GET_CODE (ref) != MEM)
2296 if (reload_in_progress
2297 && ! memory_address_p (GET_MODE (ref), XEXP (ref, 0)))
2299 base = find_replacement (&XEXP (ref, 0));
2301 if (! memory_address_p (GET_MODE (ref), base))
2306 base = XEXP (ref, 0);
2309 if (GET_CODE (base) == PLUS)
2310 offset += INTVAL (XEXP (base, 1)), base = XEXP (base, 0);
2312 return plus_constant (base, offset + extra_offset);
2315 /* On the Alpha, all (non-symbolic) constants except zero go into
2316 a floating-point register via memory. Note that we cannot
2317 return anything that is not a subset of CLASS, and that some
2318 symbolic constants cannot be dropped to memory. */
2321 alpha_preferred_reload_class(rtx x, enum reg_class class)
2323 /* Zero is present in any register class. */
2324 if (x == CONST0_RTX (GET_MODE (x)))
2327 /* These sorts of constants we can easily drop to memory. */
2328 if (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE)
2330 if (class == FLOAT_REGS)
2332 if (class == ALL_REGS)
2333 return GENERAL_REGS;
2337 /* All other kinds of constants should not (and in the case of HIGH
2338 cannot) be dropped to memory -- instead we use a GENERAL_REGS
2339 secondary reload. */
2341 return (class == ALL_REGS ? GENERAL_REGS : class);
2346 /* Loading and storing HImode or QImode values to and from memory
2347 usually requires a scratch register. The exceptions are loading
2348 QImode and HImode from an aligned address to a general register
2349 unless byte instructions are permitted.
2351 We also cannot load an unaligned address or a paradoxical SUBREG
2352 into an FP register.
2354 We also cannot do integral arithmetic into FP regs, as might result
2355 from register elimination into a DImode fp register. */
2358 secondary_reload_class (enum reg_class class, enum machine_mode mode,
2361 if ((mode == QImode || mode == HImode) && ! TARGET_BWX)
2363 if (GET_CODE (x) == MEM
2364 || (GET_CODE (x) == REG && REGNO (x) >= FIRST_PSEUDO_REGISTER)
2365 || (GET_CODE (x) == SUBREG
2366 && (GET_CODE (SUBREG_REG (x)) == MEM
2367 || (GET_CODE (SUBREG_REG (x)) == REG
2368 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER))))
2370 if (!in || !aligned_memory_operand(x, mode))
2371 return GENERAL_REGS;
2375 if (class == FLOAT_REGS)
2377 if (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == AND)
2378 return GENERAL_REGS;
2380 if (GET_CODE (x) == SUBREG
2381 && (GET_MODE_SIZE (GET_MODE (x))
2382 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))))
2383 return GENERAL_REGS;
2385 if (in && INTEGRAL_MODE_P (mode)
2386 && ! (memory_operand (x, mode) || x == const0_rtx))
2387 return GENERAL_REGS;
2393 /* Subfunction of the following function. Update the flags of any MEM
2394 found in part of X. */
2397 alpha_set_memflags_1 (rtx x, int in_struct_p, int volatile_p, int unchanging_p)
2401 switch (GET_CODE (x))
2407 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
2408 alpha_set_memflags_1 (XVECEXP (x, 0, i), in_struct_p, volatile_p,
2413 alpha_set_memflags_1 (PATTERN (x), in_struct_p, volatile_p,
2418 alpha_set_memflags_1 (SET_DEST (x), in_struct_p, volatile_p,
2420 alpha_set_memflags_1 (SET_SRC (x), in_struct_p, volatile_p,
2425 MEM_IN_STRUCT_P (x) = in_struct_p;
2426 MEM_VOLATILE_P (x) = volatile_p;
2427 RTX_UNCHANGING_P (x) = unchanging_p;
2428 /* Sadly, we cannot use alias sets because the extra aliasing
2429 produced by the AND interferes. Given that two-byte quantities
2430 are the only thing we would be able to differentiate anyway,
2431 there does not seem to be any point in convoluting the early
2432 out of the alias check. */
2440 /* Given INSN, which is an INSN list or the PATTERN of a single insn
2441 generated to perform a memory operation, look for any MEMs in either
2442 a SET_DEST or a SET_SRC and copy the in-struct, unchanging, and
2443 volatile flags from REF into each of the MEMs found. If REF is not
2444 a MEM, don't do anything. */
2447 alpha_set_memflags (rtx insn, rtx ref)
2449 int in_struct_p, volatile_p, unchanging_p;
2451 if (GET_CODE (ref) != MEM)
2454 in_struct_p = MEM_IN_STRUCT_P (ref);
2455 volatile_p = MEM_VOLATILE_P (ref);
2456 unchanging_p = RTX_UNCHANGING_P (ref);
2458 /* This is only called from alpha.md, after having had something
2459 generated from one of the insn patterns. So if everything is
2460 zero, the pattern is already up-to-date. */
2461 if (! in_struct_p && ! volatile_p && ! unchanging_p)
2464 alpha_set_memflags_1 (insn, in_struct_p, volatile_p, unchanging_p);
2467 /* Internal routine for alpha_emit_set_const to check for N or below insns. */
2470 alpha_emit_set_const_1 (rtx target, enum machine_mode mode,
2471 HOST_WIDE_INT c, int n)
2475 /* Use a pseudo if highly optimizing and still generating RTL. */
2477 = (flag_expensive_optimizations && !no_new_pseudos ? 0 : target);
2480 /* If this is a sign-extended 32-bit constant, we can do this in at most
2481 three insns, so do it if we have enough insns left. We always have
2482 a sign-extended 32-bit constant when compiling on a narrow machine. */
2484 if (HOST_BITS_PER_WIDE_INT != 64
2485 || c >> 31 == -1 || c >> 31 == 0)
2487 HOST_WIDE_INT low = ((c & 0xffff) ^ 0x8000) - 0x8000;
2488 HOST_WIDE_INT tmp1 = c - low;
2489 HOST_WIDE_INT high = (((tmp1 >> 16) & 0xffff) ^ 0x8000) - 0x8000;
2490 HOST_WIDE_INT extra = 0;
2492 /* If HIGH will be interpreted as negative but the constant is
2493 positive, we must adjust it to do two ldha insns. */
2495 if ((high & 0x8000) != 0 && c >= 0)
2499 high = ((tmp1 >> 16) & 0xffff) - 2 * ((tmp1 >> 16) & 0x8000);
2502 if (c == low || (low == 0 && extra == 0))
2504 /* We used to use copy_to_suggested_reg (GEN_INT (c), target, mode)
2505 but that meant that we can't handle INT_MIN on 32-bit machines
2506 (like NT/Alpha), because we recurse indefinitely through
2507 emit_move_insn to gen_movdi. So instead, since we know exactly
2508 what we want, create it explicitly. */
2511 target = gen_reg_rtx (mode);
2512 emit_insn (gen_rtx_SET (VOIDmode, target, GEN_INT (c)));
2515 else if (n >= 2 + (extra != 0))
2517 temp = copy_to_suggested_reg (GEN_INT (high << 16), subtarget, mode);
2519 /* As of 2002-02-23, addsi3 is only available when not optimizing.
2520 This means that if we go through expand_binop, we'll try to
2521 generate extensions, etc, which will require new pseudos, which
2522 will fail during some split phases. The SImode add patterns
2523 still exist, but are not named. So build the insns by hand. */
2528 subtarget = gen_reg_rtx (mode);
2529 insn = gen_rtx_PLUS (mode, temp, GEN_INT (extra << 16));
2530 insn = gen_rtx_SET (VOIDmode, subtarget, insn);
2536 target = gen_reg_rtx (mode);
2537 insn = gen_rtx_PLUS (mode, temp, GEN_INT (low));
2538 insn = gen_rtx_SET (VOIDmode, target, insn);
2544 /* If we couldn't do it that way, try some other methods. But if we have
2545 no instructions left, don't bother. Likewise, if this is SImode and
2546 we can't make pseudos, we can't do anything since the expand_binop
2547 and expand_unop calls will widen and try to make pseudos. */
2549 if (n == 1 || (mode == SImode && no_new_pseudos))
2552 /* Next, see if we can load a related constant and then shift and possibly
2553 negate it to get the constant we want. Try this once each increasing
2554 numbers of insns. */
2556 for (i = 1; i < n; i++)
2558 /* First, see if minus some low bits, we've an easy load of
2561 new = ((c & 0xffff) ^ 0x8000) - 0x8000;
2563 && (temp = alpha_emit_set_const (subtarget, mode, c - new, i)) != 0)
2564 return expand_binop (mode, add_optab, temp, GEN_INT (new),
2565 target, 0, OPTAB_WIDEN);
2567 /* Next try complementing. */
2568 if ((temp = alpha_emit_set_const (subtarget, mode, ~ c, i)) != 0)
2569 return expand_unop (mode, one_cmpl_optab, temp, target, 0);
2571 /* Next try to form a constant and do a left shift. We can do this
2572 if some low-order bits are zero; the exact_log2 call below tells
2573 us that information. The bits we are shifting out could be any
2574 value, but here we'll just try the 0- and sign-extended forms of
2575 the constant. To try to increase the chance of having the same
2576 constant in more than one insn, start at the highest number of
2577 bits to shift, but try all possibilities in case a ZAPNOT will
2580 if ((bits = exact_log2 (c & - c)) > 0)
2581 for (; bits > 0; bits--)
2582 if ((temp = (alpha_emit_set_const
2583 (subtarget, mode, c >> bits, i))) != 0
2584 || ((temp = (alpha_emit_set_const
2586 ((unsigned HOST_WIDE_INT) c) >> bits, i)))
2588 return expand_binop (mode, ashl_optab, temp, GEN_INT (bits),
2589 target, 0, OPTAB_WIDEN);
2591 /* Now try high-order zero bits. Here we try the shifted-in bits as
2592 all zero and all ones. Be careful to avoid shifting outside the
2593 mode and to avoid shifting outside the host wide int size. */
2594 /* On narrow hosts, don't shift a 1 into the high bit, since we'll
2595 confuse the recursive call and set all of the high 32 bits. */
2597 if ((bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
2598 - floor_log2 (c) - 1 - (HOST_BITS_PER_WIDE_INT < 64))) > 0)
2599 for (; bits > 0; bits--)
2600 if ((temp = alpha_emit_set_const (subtarget, mode,
2602 || ((temp = (alpha_emit_set_const
2604 ((c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1)),
2607 return expand_binop (mode, lshr_optab, temp, GEN_INT (bits),
2608 target, 1, OPTAB_WIDEN);
2610 /* Now try high-order 1 bits. We get that with a sign-extension.
2611 But one bit isn't enough here. Be careful to avoid shifting outside
2612 the mode and to avoid shifting outside the host wide int size. */
2614 if ((bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
2615 - floor_log2 (~ c) - 2)) > 0)
2616 for (; bits > 0; bits--)
2617 if ((temp = alpha_emit_set_const (subtarget, mode,
2619 || ((temp = (alpha_emit_set_const
2621 ((c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1)),
2624 return expand_binop (mode, ashr_optab, temp, GEN_INT (bits),
2625 target, 0, OPTAB_WIDEN);
2628 #if HOST_BITS_PER_WIDE_INT == 64
2629 /* Finally, see if can load a value into the target that is the same as the
2630 constant except that all bytes that are 0 are changed to be 0xff. If we
2631 can, then we can do a ZAPNOT to obtain the desired constant. */
2634 for (i = 0; i < 64; i += 8)
2635 if ((new & ((HOST_WIDE_INT) 0xff << i)) == 0)
2636 new |= (HOST_WIDE_INT) 0xff << i;
2638 /* We are only called for SImode and DImode. If this is SImode, ensure that
2639 we are sign extended to a full word. */
2642 new = ((new & 0xffffffff) ^ 0x80000000) - 0x80000000;
2644 if (new != c && new != -1
2645 && (temp = alpha_emit_set_const (subtarget, mode, new, n - 1)) != 0)
2646 return expand_binop (mode, and_optab, temp, GEN_INT (c | ~ new),
2647 target, 0, OPTAB_WIDEN);
2653 /* Try to output insns to set TARGET equal to the constant C if it can be
2654 done in less than N insns. Do all computations in MODE. Returns the place
2655 where the output has been placed if it can be done and the insns have been
2656 emitted. If it would take more than N insns, zero is returned and no
2657 insns and emitted. */
2660 alpha_emit_set_const (rtx target, enum machine_mode mode,
2661 HOST_WIDE_INT c, int n)
2664 rtx orig_target = target;
2667 /* If we can't make any pseudos, TARGET is an SImode hard register, we
2668 can't load this constant in one insn, do this in DImode. */
2669 if (no_new_pseudos && mode == SImode
2670 && GET_CODE (target) == REG && REGNO (target) < FIRST_PSEUDO_REGISTER
2671 && (result = alpha_emit_set_const_1 (target, mode, c, 1)) == 0)
2673 target = gen_lowpart (DImode, target);
2677 /* Try 1 insn, then 2, then up to N. */
2678 for (i = 1; i <= n; i++)
2680 result = alpha_emit_set_const_1 (target, mode, c, i);
2683 rtx insn = get_last_insn ();
2684 rtx set = single_set (insn);
2685 if (! CONSTANT_P (SET_SRC (set)))
2686 set_unique_reg_note (get_last_insn (), REG_EQUAL, GEN_INT (c));
2691 /* Allow for the case where we changed the mode of TARGET. */
2692 if (result == target)
2693 result = orig_target;
2698 /* Having failed to find a 3 insn sequence in alpha_emit_set_const,
2699 fall back to a straight forward decomposition. We do this to avoid
2700 exponential run times encountered when looking for longer sequences
2701 with alpha_emit_set_const. */
2704 alpha_emit_set_long_const (rtx target, HOST_WIDE_INT c1, HOST_WIDE_INT c2)
2706 HOST_WIDE_INT d1, d2, d3, d4;
2708 /* Decompose the entire word */
2709 #if HOST_BITS_PER_WIDE_INT >= 64
2710 if (c2 != -(c1 < 0))
2712 d1 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
2714 d2 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
2715 c1 = (c1 - d2) >> 32;
2716 d3 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
2718 d4 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
2722 d1 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
2724 d2 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
2728 d3 = ((c2 & 0xffff) ^ 0x8000) - 0x8000;
2730 d4 = ((c2 & 0xffffffff) ^ 0x80000000) - 0x80000000;
2735 /* Construct the high word */
2738 emit_move_insn (target, GEN_INT (d4));
2740 emit_move_insn (target, gen_rtx_PLUS (DImode, target, GEN_INT (d3)));
2743 emit_move_insn (target, GEN_INT (d3));
2745 /* Shift it into place */
2746 emit_move_insn (target, gen_rtx_ASHIFT (DImode, target, GEN_INT (32)));
2748 /* Add in the low bits. */
2750 emit_move_insn (target, gen_rtx_PLUS (DImode, target, GEN_INT (d2)));
2752 emit_move_insn (target, gen_rtx_PLUS (DImode, target, GEN_INT (d1)));
2757 /* Expand a move instruction; return true if all work is done.
2758 We don't handle non-bwx subword loads here. */
2761 alpha_expand_mov (enum machine_mode mode, rtx *operands)
2763 /* If the output is not a register, the input must be. */
2764 if (GET_CODE (operands[0]) == MEM
2765 && ! reg_or_0_operand (operands[1], mode))
2766 operands[1] = force_reg (mode, operands[1]);
2768 /* Allow legitimize_address to perform some simplifications. */
2769 if (mode == Pmode && symbolic_operand (operands[1], mode))
2773 /* With RTL inlining, at -O3, rtl is generated, stored, then actually
2774 compiled at the end of compilation. In the meantime, someone can
2775 re-encode-section-info on some symbol changing it e.g. from global
2776 to local-not-small. If this happens, we'd have emitted a plain
2777 load rather than a high+losum load and not recognize the insn.
2779 So if rtl inlining is in effect, we delay the global/not-global
2780 decision until rest_of_compilation by wrapping it in an
2782 if (TARGET_EXPLICIT_RELOCS && flag_inline_functions
2783 && rtx_equal_function_value_matters
2784 && global_symbolic_operand (operands[1], mode))
2786 emit_insn (gen_movdi_er_maybe_g (operands[0], operands[1]));
2790 tmp = alpha_legitimize_address (operands[1], operands[0], mode);
2793 if (tmp == operands[0])
2800 /* Early out for non-constants and valid constants. */
2801 if (! CONSTANT_P (operands[1]) || input_operand (operands[1], mode))
2804 /* Split large integers. */
2805 if (GET_CODE (operands[1]) == CONST_INT
2806 || GET_CODE (operands[1]) == CONST_DOUBLE)
2808 HOST_WIDE_INT i0, i1;
2809 rtx temp = NULL_RTX;
2811 if (GET_CODE (operands[1]) == CONST_INT)
2813 i0 = INTVAL (operands[1]);
2816 else if (HOST_BITS_PER_WIDE_INT >= 64)
2818 i0 = CONST_DOUBLE_LOW (operands[1]);
2823 i0 = CONST_DOUBLE_LOW (operands[1]);
2824 i1 = CONST_DOUBLE_HIGH (operands[1]);
2827 if (HOST_BITS_PER_WIDE_INT >= 64 || i1 == -(i0 < 0))
2828 temp = alpha_emit_set_const (operands[0], mode, i0, 3);
2830 if (!temp && TARGET_BUILD_CONSTANTS)
2831 temp = alpha_emit_set_long_const (operands[0], i0, i1);
2835 if (rtx_equal_p (operands[0], temp))
2842 /* Otherwise we've nothing left but to drop the thing to memory. */
2843 operands[1] = force_const_mem (mode, operands[1]);
2844 if (reload_in_progress)
2846 emit_move_insn (operands[0], XEXP (operands[1], 0));
2847 operands[1] = copy_rtx (operands[1]);
2848 XEXP (operands[1], 0) = operands[0];
2851 operands[1] = validize_mem (operands[1]);
2855 /* Expand a non-bwx QImode or HImode move instruction;
2856 return true if all work is done. */
2859 alpha_expand_mov_nobwx (enum machine_mode mode, rtx *operands)
2861 /* If the output is not a register, the input must be. */
2862 if (GET_CODE (operands[0]) == MEM)
2863 operands[1] = force_reg (mode, operands[1]);
2865 /* Handle four memory cases, unaligned and aligned for either the input
2866 or the output. The only case where we can be called during reload is
2867 for aligned loads; all other cases require temporaries. */
2869 if (GET_CODE (operands[1]) == MEM
2870 || (GET_CODE (operands[1]) == SUBREG
2871 && GET_CODE (SUBREG_REG (operands[1])) == MEM)
2872 || (reload_in_progress && GET_CODE (operands[1]) == REG
2873 && REGNO (operands[1]) >= FIRST_PSEUDO_REGISTER)
2874 || (reload_in_progress && GET_CODE (operands[1]) == SUBREG
2875 && GET_CODE (SUBREG_REG (operands[1])) == REG
2876 && REGNO (SUBREG_REG (operands[1])) >= FIRST_PSEUDO_REGISTER))
2878 if (aligned_memory_operand (operands[1], mode))
2880 if (reload_in_progress)
2882 emit_insn ((mode == QImode
2883 ? gen_reload_inqi_help
2884 : gen_reload_inhi_help)
2885 (operands[0], operands[1],
2886 gen_rtx_REG (SImode, REGNO (operands[0]))));
2890 rtx aligned_mem, bitnum;
2891 rtx scratch = gen_reg_rtx (SImode);
2893 get_aligned_mem (operands[1], &aligned_mem, &bitnum);
2895 emit_insn ((mode == QImode
2896 ? gen_aligned_loadqi
2897 : gen_aligned_loadhi)
2898 (operands[0], aligned_mem, bitnum, scratch));
2903 /* Don't pass these as parameters since that makes the generated
2904 code depend on parameter evaluation order which will cause
2905 bootstrap failures. */
2907 rtx temp1 = gen_reg_rtx (DImode);
2908 rtx temp2 = gen_reg_rtx (DImode);
2909 rtx seq = ((mode == QImode
2910 ? gen_unaligned_loadqi
2911 : gen_unaligned_loadhi)
2912 (operands[0], get_unaligned_address (operands[1], 0),
2915 alpha_set_memflags (seq, operands[1]);
2921 if (GET_CODE (operands[0]) == MEM
2922 || (GET_CODE (operands[0]) == SUBREG
2923 && GET_CODE (SUBREG_REG (operands[0])) == MEM)
2924 || (reload_in_progress && GET_CODE (operands[0]) == REG
2925 && REGNO (operands[0]) >= FIRST_PSEUDO_REGISTER)
2926 || (reload_in_progress && GET_CODE (operands[0]) == SUBREG
2927 && GET_CODE (SUBREG_REG (operands[0])) == REG
2928 && REGNO (operands[0]) >= FIRST_PSEUDO_REGISTER))
2930 if (aligned_memory_operand (operands[0], mode))
2932 rtx aligned_mem, bitnum;
2933 rtx temp1 = gen_reg_rtx (SImode);
2934 rtx temp2 = gen_reg_rtx (SImode);
2936 get_aligned_mem (operands[0], &aligned_mem, &bitnum);
2938 emit_insn (gen_aligned_store (aligned_mem, operands[1], bitnum,
2943 rtx temp1 = gen_reg_rtx (DImode);
2944 rtx temp2 = gen_reg_rtx (DImode);
2945 rtx temp3 = gen_reg_rtx (DImode);
2946 rtx seq = ((mode == QImode
2947 ? gen_unaligned_storeqi
2948 : gen_unaligned_storehi)
2949 (get_unaligned_address (operands[0], 0),
2950 operands[1], temp1, temp2, temp3));
2952 alpha_set_memflags (seq, operands[0]);
2961 /* Generate an unsigned DImode to FP conversion. This is the same code
2962 optabs would emit if we didn't have TFmode patterns.
2964 For SFmode, this is the only construction I've found that can pass
2965 gcc.c-torture/execute/ieee/rbug.c. No scenario that uses DFmode
2966 intermediates will work, because you'll get intermediate rounding
2967 that ruins the end result. Some of this could be fixed by turning
2968 on round-to-positive-infinity, but that requires diddling the fpsr,
2969 which kills performance. I tried turning this around and converting
2970 to a negative number, so that I could turn on /m, but either I did
2971 it wrong or there's something else cause I wound up with the exact
2972 same single-bit error. There is a branch-less form of this same code:
2983 fcmoveq $f10,$f11,$f0
2985 I'm not using it because it's the same number of instructions as
2986 this branch-full form, and it has more serialized long latency
2987 instructions on the critical path.
2989 For DFmode, we can avoid rounding errors by breaking up the word
2990 into two pieces, converting them separately, and adding them back:
2992 LC0: .long 0,0x5f800000
2997 cpyse $f11,$f31,$f10
2998 cpyse $f31,$f11,$f11
3006 This doesn't seem to be a clear-cut win over the optabs form.
3007 It probably all depends on the distribution of numbers being
3008 converted -- in the optabs form, all but high-bit-set has a
3009 much lower minimum execution time. */
3012 alpha_emit_floatuns (rtx operands[2])
3014 rtx neglab, donelab, i0, i1, f0, in, out;
3015 enum machine_mode mode;
3018 in = force_reg (DImode, operands[1]);
3019 mode = GET_MODE (out);
3020 neglab = gen_label_rtx ();
3021 donelab = gen_label_rtx ();
3022 i0 = gen_reg_rtx (DImode);
3023 i1 = gen_reg_rtx (DImode);
3024 f0 = gen_reg_rtx (mode);
3026 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
3028 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
3029 emit_jump_insn (gen_jump (donelab));
3032 emit_label (neglab);
3034 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
3035 emit_insn (gen_anddi3 (i1, in, const1_rtx));
3036 emit_insn (gen_iordi3 (i0, i0, i1));
3037 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
3038 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
3040 emit_label (donelab);
3043 /* Generate the comparison for a conditional branch. */
3046 alpha_emit_conditional_branch (enum rtx_code code)
3048 enum rtx_code cmp_code, branch_code;
3049 enum machine_mode cmp_mode, branch_mode = VOIDmode;
3050 rtx op0 = alpha_compare.op0, op1 = alpha_compare.op1;
3053 if (alpha_compare.fp_p && GET_MODE (op0) == TFmode)
3055 if (! TARGET_HAS_XFLOATING_LIBS)
3058 /* X_floating library comparison functions return
3062 Convert the compare against the raw return value. */
3084 op0 = alpha_emit_xfloating_compare (cmp_code, op0, op1);
3086 alpha_compare.fp_p = 0;
3089 /* The general case: fold the comparison code to the types of compares
3090 that we have, choosing the branch as necessary. */
3093 case EQ: case LE: case LT: case LEU: case LTU:
3095 /* We have these compares: */
3096 cmp_code = code, branch_code = NE;
3101 /* These must be reversed. */
3102 cmp_code = reverse_condition (code), branch_code = EQ;
3105 case GE: case GT: case GEU: case GTU:
3106 /* For FP, we swap them, for INT, we reverse them. */
3107 if (alpha_compare.fp_p)
3109 cmp_code = swap_condition (code);
3111 tem = op0, op0 = op1, op1 = tem;
3115 cmp_code = reverse_condition (code);
3124 if (alpha_compare.fp_p)
3127 if (flag_unsafe_math_optimizations)
3129 /* When we are not as concerned about non-finite values, and we
3130 are comparing against zero, we can branch directly. */
3131 if (op1 == CONST0_RTX (DFmode))
3132 cmp_code = NIL, branch_code = code;
3133 else if (op0 == CONST0_RTX (DFmode))
3135 /* Undo the swap we probably did just above. */
3136 tem = op0, op0 = op1, op1 = tem;
3137 branch_code = swap_condition (cmp_code);
3143 /* ??? We mark the branch mode to be CCmode to prevent the
3144 compare and branch from being combined, since the compare
3145 insn follows IEEE rules that the branch does not. */
3146 branch_mode = CCmode;
3153 /* The following optimizations are only for signed compares. */
3154 if (code != LEU && code != LTU && code != GEU && code != GTU)
3156 /* Whee. Compare and branch against 0 directly. */
3157 if (op1 == const0_rtx)
3158 cmp_code = NIL, branch_code = code;
3160 /* We want to use cmpcc/bcc when we can, since there is a zero delay
3161 bypass between logicals and br/cmov on EV5. But we don't want to
3162 force valid immediate constants into registers needlessly. */
3163 else if (GET_CODE (op1) == CONST_INT)
3165 HOST_WIDE_INT v = INTVAL (op1), n = -v;
3167 if (! CONST_OK_FOR_LETTER_P (v, 'I')
3168 && (CONST_OK_FOR_LETTER_P (n, 'K')
3169 || CONST_OK_FOR_LETTER_P (n, 'L')))
3171 cmp_code = PLUS, branch_code = code;
3177 if (!reg_or_0_operand (op0, DImode))
3178 op0 = force_reg (DImode, op0);
3179 if (cmp_code != PLUS && !reg_or_8bit_operand (op1, DImode))
3180 op1 = force_reg (DImode, op1);
3183 /* Emit an initial compare instruction, if necessary. */
3185 if (cmp_code != NIL)
3187 tem = gen_reg_rtx (cmp_mode);
3188 emit_move_insn (tem, gen_rtx_fmt_ee (cmp_code, cmp_mode, op0, op1));
3191 /* Zero the operands. */
3192 memset (&alpha_compare, 0, sizeof (alpha_compare));
3194 /* Return the branch comparison. */
3195 return gen_rtx_fmt_ee (branch_code, branch_mode, tem, CONST0_RTX (cmp_mode));
3198 /* Certain simplifications can be done to make invalid setcc operations
3199 valid. Return the final comparison, or NULL if we can't work. */
3202 alpha_emit_setcc (enum rtx_code code)
3204 enum rtx_code cmp_code;
3205 rtx op0 = alpha_compare.op0, op1 = alpha_compare.op1;
3206 int fp_p = alpha_compare.fp_p;
3209 /* Zero the operands. */
3210 memset (&alpha_compare, 0, sizeof (alpha_compare));
3212 if (fp_p && GET_MODE (op0) == TFmode)
3214 if (! TARGET_HAS_XFLOATING_LIBS)
3217 /* X_floating library comparison functions return
3221 Convert the compare against the raw return value. */
3223 if (code == UNORDERED || code == ORDERED)
3228 op0 = alpha_emit_xfloating_compare (cmp_code, op0, op1);
3232 if (code == UNORDERED)
3234 else if (code == ORDERED)
3240 if (fp_p && !TARGET_FIX)
3243 /* The general case: fold the comparison code to the types of compares
3244 that we have, choosing the branch as necessary. */
3249 case EQ: case LE: case LT: case LEU: case LTU:
3251 /* We have these compares. */
3253 cmp_code = code, code = NE;
3257 if (!fp_p && op1 == const0_rtx)
3262 cmp_code = reverse_condition (code);
3266 case GE: case GT: case GEU: case GTU:
3267 /* These normally need swapping, but for integer zero we have
3268 special patterns that recognize swapped operands. */
3269 if (!fp_p && op1 == const0_rtx)
3271 code = swap_condition (code);
3273 cmp_code = code, code = NE;
3274 tmp = op0, op0 = op1, op1 = tmp;
3283 if (!register_operand (op0, DImode))
3284 op0 = force_reg (DImode, op0);
3285 if (!reg_or_8bit_operand (op1, DImode))
3286 op1 = force_reg (DImode, op1);
3289 /* Emit an initial compare instruction, if necessary. */
3290 if (cmp_code != NIL)
3292 enum machine_mode mode = fp_p ? DFmode : DImode;
3294 tmp = gen_reg_rtx (mode);
3295 emit_insn (gen_rtx_SET (VOIDmode, tmp,
3296 gen_rtx_fmt_ee (cmp_code, mode, op0, op1)));
3298 op0 = fp_p ? gen_lowpart (DImode, tmp) : tmp;
3302 /* Return the setcc comparison. */
3303 return gen_rtx_fmt_ee (code, DImode, op0, op1);
3307 /* Rewrite a comparison against zero CMP of the form
3308 (CODE (cc0) (const_int 0)) so it can be written validly in
3309 a conditional move (if_then_else CMP ...).
3310 If both of the operands that set cc0 are nonzero we must emit
3311 an insn to perform the compare (it can't be done within
3312 the conditional move). */
3315 alpha_emit_conditional_move (rtx cmp, enum machine_mode mode)
3317 enum rtx_code code = GET_CODE (cmp);
3318 enum rtx_code cmov_code = NE;
3319 rtx op0 = alpha_compare.op0;
3320 rtx op1 = alpha_compare.op1;
3321 int fp_p = alpha_compare.fp_p;
3322 enum machine_mode cmp_mode
3323 = (GET_MODE (op0) == VOIDmode ? DImode : GET_MODE (op0));
3324 enum machine_mode cmp_op_mode = fp_p ? DFmode : DImode;
3325 enum machine_mode cmov_mode = VOIDmode;
3326 int local_fast_math = flag_unsafe_math_optimizations;
3329 /* Zero the operands. */
3330 memset (&alpha_compare, 0, sizeof (alpha_compare));
3332 if (fp_p != FLOAT_MODE_P (mode))
3334 enum rtx_code cmp_code;
3339 /* If we have fp<->int register move instructions, do a cmov by
3340 performing the comparison in fp registers, and move the
3341 zero/nonzero value to integer registers, where we can then
3342 use a normal cmov, or vice-versa. */
3346 case EQ: case LE: case LT: case LEU: case LTU:
3347 /* We have these compares. */
3348 cmp_code = code, code = NE;
3352 /* This must be reversed. */
3353 cmp_code = EQ, code = EQ;
3356 case GE: case GT: case GEU: case GTU:
3357 /* These normally need swapping, but for integer zero we have
3358 special patterns that recognize swapped operands. */
3359 if (!fp_p && op1 == const0_rtx)
3360 cmp_code = code, code = NE;
3363 cmp_code = swap_condition (code);
3365 tem = op0, op0 = op1, op1 = tem;
3373 tem = gen_reg_rtx (cmp_op_mode);
3374 emit_insn (gen_rtx_SET (VOIDmode, tem,
3375 gen_rtx_fmt_ee (cmp_code, cmp_op_mode,
3378 cmp_mode = cmp_op_mode = fp_p ? DImode : DFmode;
3379 op0 = gen_lowpart (cmp_op_mode, tem);
3380 op1 = CONST0_RTX (cmp_op_mode);
3382 local_fast_math = 1;
3385 /* We may be able to use a conditional move directly.
3386 This avoids emitting spurious compares. */
3387 if (signed_comparison_operator (cmp, VOIDmode)
3388 && (!fp_p || local_fast_math)
3389 && (op0 == CONST0_RTX (cmp_mode) || op1 == CONST0_RTX (cmp_mode)))
3390 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
3392 /* We can't put the comparison inside the conditional move;
3393 emit a compare instruction and put that inside the
3394 conditional move. Make sure we emit only comparisons we have;
3395 swap or reverse as necessary. */
3402 case EQ: case LE: case LT: case LEU: case LTU:
3403 /* We have these compares: */
3407 /* This must be reversed. */
3408 code = reverse_condition (code);
3412 case GE: case GT: case GEU: case GTU:
3413 /* These must be swapped. */
3414 if (op1 != CONST0_RTX (cmp_mode))
3416 code = swap_condition (code);
3417 tem = op0, op0 = op1, op1 = tem;
3427 if (!reg_or_0_operand (op0, DImode))
3428 op0 = force_reg (DImode, op0);
3429 if (!reg_or_8bit_operand (op1, DImode))
3430 op1 = force_reg (DImode, op1);
3433 /* ??? We mark the branch mode to be CCmode to prevent the compare
3434 and cmov from being combined, since the compare insn follows IEEE
3435 rules that the cmov does not. */
3436 if (fp_p && !local_fast_math)
3439 tem = gen_reg_rtx (cmp_op_mode);
3440 emit_move_insn (tem, gen_rtx_fmt_ee (code, cmp_op_mode, op0, op1));
3441 return gen_rtx_fmt_ee (cmov_code, cmov_mode, tem, CONST0_RTX (cmp_op_mode));
3444 /* Simplify a conditional move of two constants into a setcc with
3445 arithmetic. This is done with a splitter since combine would
3446 just undo the work if done during code generation. It also catches
3447 cases we wouldn't have before cse. */
3450 alpha_split_conditional_move (enum rtx_code code, rtx dest, rtx cond,
3451 rtx t_rtx, rtx f_rtx)
3453 HOST_WIDE_INT t, f, diff;
3454 enum machine_mode mode;
3455 rtx target, subtarget, tmp;
3457 mode = GET_MODE (dest);
3462 if (((code == NE || code == EQ) && diff < 0)
3463 || (code == GE || code == GT))
3465 code = reverse_condition (code);
3466 diff = t, t = f, f = diff;
3470 subtarget = target = dest;
3473 target = gen_lowpart (DImode, dest);
3474 if (! no_new_pseudos)
3475 subtarget = gen_reg_rtx (DImode);
3479 /* Below, we must be careful to use copy_rtx on target and subtarget
3480 in intermediate insns, as they may be a subreg rtx, which may not
3483 if (f == 0 && exact_log2 (diff) > 0
3484 /* On EV6, we've got enough shifters to make non-arithmetic shifts
3485 viable over a longer latency cmove. On EV5, the E0 slot is a
3486 scarce resource, and on EV4 shift has the same latency as a cmove. */
3487 && (diff <= 8 || alpha_cpu == PROCESSOR_EV6))
3489 tmp = gen_rtx_fmt_ee (code, DImode, cond, const0_rtx);
3490 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (subtarget), tmp));
3492 tmp = gen_rtx_ASHIFT (DImode, copy_rtx (subtarget),
3493 GEN_INT (exact_log2 (t)));
3494 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
3496 else if (f == 0 && t == -1)
3498 tmp = gen_rtx_fmt_ee (code, DImode, cond, const0_rtx);
3499 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (subtarget), tmp));
3501 emit_insn (gen_negdi2 (target, copy_rtx (subtarget)));
3503 else if (diff == 1 || diff == 4 || diff == 8)
3507 tmp = gen_rtx_fmt_ee (code, DImode, cond, const0_rtx);
3508 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (subtarget), tmp));
3511 emit_insn (gen_adddi3 (target, copy_rtx (subtarget), GEN_INT (f)));
3514 add_op = GEN_INT (f);
3515 if (sext_add_operand (add_op, mode))
3517 tmp = gen_rtx_MULT (DImode, copy_rtx (subtarget),
3519 tmp = gen_rtx_PLUS (DImode, tmp, add_op);
3520 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
3532 /* Look up the function X_floating library function name for the
3536 alpha_lookup_xfloating_lib_func (enum rtx_code code)
3540 const enum rtx_code code;
3541 const char *const func;
3544 static const struct xfloating_op vms_xfloating_ops[] =
3546 { PLUS, "OTS$ADD_X" },
3547 { MINUS, "OTS$SUB_X" },
3548 { MULT, "OTS$MUL_X" },
3549 { DIV, "OTS$DIV_X" },
3550 { EQ, "OTS$EQL_X" },
3551 { NE, "OTS$NEQ_X" },
3552 { LT, "OTS$LSS_X" },
3553 { LE, "OTS$LEQ_X" },
3554 { GT, "OTS$GTR_X" },
3555 { GE, "OTS$GEQ_X" },
3556 { FIX, "OTS$CVTXQ" },
3557 { FLOAT, "OTS$CVTQX" },
3558 { UNSIGNED_FLOAT, "OTS$CVTQUX" },
3559 { FLOAT_EXTEND, "OTS$CVT_FLOAT_T_X" },
3560 { FLOAT_TRUNCATE, "OTS$CVT_FLOAT_X_T" },
3563 static const struct xfloating_op osf_xfloating_ops[] =
3565 { PLUS, "_OtsAddX" },
3566 { MINUS, "_OtsSubX" },
3567 { MULT, "_OtsMulX" },
3568 { DIV, "_OtsDivX" },
3575 { FIX, "_OtsCvtXQ" },
3576 { FLOAT, "_OtsCvtQX" },
3577 { UNSIGNED_FLOAT, "_OtsCvtQUX" },
3578 { FLOAT_EXTEND, "_OtsConvertFloatTX" },
3579 { FLOAT_TRUNCATE, "_OtsConvertFloatXT" },
3582 const struct xfloating_op *ops;
3583 const long n = ARRAY_SIZE (osf_xfloating_ops);
3586 /* How irritating. Nothing to key off for the table. Hardcode
3587 knowledge of the G_floating routines. */
3588 if (TARGET_FLOAT_VAX)
3590 if (TARGET_ABI_OPEN_VMS)
3592 if (code == FLOAT_EXTEND)
3593 return "OTS$CVT_FLOAT_G_X";
3594 if (code == FLOAT_TRUNCATE)
3595 return "OTS$CVT_FLOAT_X_G";
3599 if (code == FLOAT_EXTEND)
3600 return "_OtsConvertFloatGX";
3601 if (code == FLOAT_TRUNCATE)
3602 return "_OtsConvertFloatXG";
3606 if (TARGET_ABI_OPEN_VMS)
3607 ops = vms_xfloating_ops;
3609 ops = osf_xfloating_ops;
3611 for (i = 0; i < n; ++i)
3612 if (ops[i].code == code)
3618 /* Most X_floating operations take the rounding mode as an argument.
3619 Compute that here. */
3622 alpha_compute_xfloating_mode_arg (enum rtx_code code,
3623 enum alpha_fp_rounding_mode round)
3629 case ALPHA_FPRM_NORM:
3632 case ALPHA_FPRM_MINF:
3635 case ALPHA_FPRM_CHOP:
3638 case ALPHA_FPRM_DYN:
3644 /* XXX For reference, round to +inf is mode = 3. */
3647 if (code == FLOAT_TRUNCATE && alpha_fptm == ALPHA_FPTM_N)
3653 /* Emit an X_floating library function call.
3655 Note that these functions do not follow normal calling conventions:
3656 TFmode arguments are passed in two integer registers (as opposed to
3657 indirect); TFmode return values appear in R16+R17.
3659 FUNC is the function name to call.
3660 TARGET is where the output belongs.
3661 OPERANDS are the inputs.
3662 NOPERANDS is the count of inputs.
3663 EQUIV is the expression equivalent for the function.
3667 alpha_emit_xfloating_libcall (const char *func, rtx target, rtx operands[],
3668 int noperands, rtx equiv)
3670 rtx usage = NULL_RTX, tmp, reg;
3675 for (i = 0; i < noperands; ++i)
3677 switch (GET_MODE (operands[i]))
3680 reg = gen_rtx_REG (TFmode, regno);
3685 reg = gen_rtx_REG (DFmode, regno + 32);
3690 if (GET_CODE (operands[i]) != CONST_INT)
3694 reg = gen_rtx_REG (DImode, regno);
3702 emit_move_insn (reg, operands[i]);
3703 usage = alloc_EXPR_LIST (0, gen_rtx_USE (VOIDmode, reg), usage);
3706 switch (GET_MODE (target))
3709 reg = gen_rtx_REG (TFmode, 16);
3712 reg = gen_rtx_REG (DFmode, 32);
3715 reg = gen_rtx_REG (DImode, 0);
3721 tmp = gen_rtx_MEM (QImode, init_one_libfunc (func));
3722 tmp = emit_call_insn (GEN_CALL_VALUE (reg, tmp, const0_rtx,
3723 const0_rtx, const0_rtx));
3724 CALL_INSN_FUNCTION_USAGE (tmp) = usage;
3729 emit_libcall_block (tmp, target, reg, equiv);
3732 /* Emit an X_floating library function call for arithmetic (+,-,*,/). */
3735 alpha_emit_xfloating_arith (enum rtx_code code, rtx operands[])
3739 rtx out_operands[3];
3741 func = alpha_lookup_xfloating_lib_func (code);
3742 mode = alpha_compute_xfloating_mode_arg (code, alpha_fprm);
3744 out_operands[0] = operands[1];
3745 out_operands[1] = operands[2];
3746 out_operands[2] = GEN_INT (mode);
3747 alpha_emit_xfloating_libcall (func, operands[0], out_operands, 3,
3748 gen_rtx_fmt_ee (code, TFmode, operands[1],
3752 /* Emit an X_floating library function call for a comparison. */
3755 alpha_emit_xfloating_compare (enum rtx_code code, rtx op0, rtx op1)
3758 rtx out, operands[2];
3760 func = alpha_lookup_xfloating_lib_func (code);
3764 out = gen_reg_rtx (DImode);
3766 /* ??? Strange mode for equiv because what's actually returned
3767 is -1,0,1, not a proper boolean value. */
3768 alpha_emit_xfloating_libcall (func, out, operands, 2,
3769 gen_rtx_fmt_ee (code, CCmode, op0, op1));
3774 /* Emit an X_floating library function call for a conversion. */
3777 alpha_emit_xfloating_cvt (enum rtx_code code, rtx operands[])
3779 int noperands = 1, mode;
3780 rtx out_operands[2];
3783 func = alpha_lookup_xfloating_lib_func (code);
3785 out_operands[0] = operands[1];
3790 mode = alpha_compute_xfloating_mode_arg (code, ALPHA_FPRM_CHOP);
3791 out_operands[1] = GEN_INT (mode);
3794 case FLOAT_TRUNCATE:
3795 mode = alpha_compute_xfloating_mode_arg (code, alpha_fprm);
3796 out_operands[1] = GEN_INT (mode);
3803 alpha_emit_xfloating_libcall (func, operands[0], out_operands, noperands,
3804 gen_rtx_fmt_e (code, GET_MODE (operands[0]),
3808 /* Split a TFmode OP[1] into DImode OP[2,3] and likewise for
3809 OP[0] into OP[0,1]. Naturally, output operand ordering is
3813 alpha_split_tfmode_pair (rtx operands[4])
3815 if (GET_CODE (operands[1]) == REG)
3817 operands[3] = gen_rtx_REG (DImode, REGNO (operands[1]) + 1);
3818 operands[2] = gen_rtx_REG (DImode, REGNO (operands[1]));
3820 else if (GET_CODE (operands[1]) == MEM)
3822 operands[3] = adjust_address (operands[1], DImode, 8);
3823 operands[2] = adjust_address (operands[1], DImode, 0);
3825 else if (operands[1] == CONST0_RTX (TFmode))
3826 operands[2] = operands[3] = const0_rtx;
3830 if (GET_CODE (operands[0]) == REG)
3832 operands[1] = gen_rtx_REG (DImode, REGNO (operands[0]) + 1);
3833 operands[0] = gen_rtx_REG (DImode, REGNO (operands[0]));
3835 else if (GET_CODE (operands[0]) == MEM)
3837 operands[1] = adjust_address (operands[0], DImode, 8);
3838 operands[0] = adjust_address (operands[0], DImode, 0);
3844 /* Implement negtf2 or abstf2. Op0 is destination, op1 is source,
3845 op2 is a register containing the sign bit, operation is the
3846 logical operation to be performed. */
3849 alpha_split_tfmode_frobsign (rtx operands[3], rtx (*operation) (rtx, rtx, rtx))
3851 rtx high_bit = operands[2];
3855 alpha_split_tfmode_pair (operands);
3857 /* Detect three flavors of operand overlap. */
3859 if (rtx_equal_p (operands[0], operands[2]))
3861 else if (rtx_equal_p (operands[1], operands[2]))
3863 if (rtx_equal_p (operands[0], high_bit))
3870 emit_move_insn (operands[0], operands[2]);
3872 /* ??? If the destination overlaps both source tf and high_bit, then
3873 assume source tf is dead in its entirety and use the other half
3874 for a scratch register. Otherwise "scratch" is just the proper
3875 destination register. */
3876 scratch = operands[move < 2 ? 1 : 3];
3878 emit_insn ((*operation) (scratch, high_bit, operands[3]));
3882 emit_move_insn (operands[0], operands[2]);
3884 emit_move_insn (operands[1], scratch);
3888 /* Use ext[wlq][lh] as the Architecture Handbook describes for extracting
3892 word: ldq_u r1,X(r11) ldq_u r1,X(r11)
3893 ldq_u r2,X+1(r11) ldq_u r2,X+1(r11)
3894 lda r3,X(r11) lda r3,X+2(r11)
3895 extwl r1,r3,r1 extql r1,r3,r1
3896 extwh r2,r3,r2 extqh r2,r3,r2
3897 or r1.r2.r1 or r1,r2,r1
3900 long: ldq_u r1,X(r11) ldq_u r1,X(r11)
3901 ldq_u r2,X+3(r11) ldq_u r2,X+3(r11)
3902 lda r3,X(r11) lda r3,X(r11)
3903 extll r1,r3,r1 extll r1,r3,r1
3904 extlh r2,r3,r2 extlh r2,r3,r2
3905 or r1.r2.r1 addl r1,r2,r1
3907 quad: ldq_u r1,X(r11)
3916 alpha_expand_unaligned_load (rtx tgt, rtx mem, HOST_WIDE_INT size,
3917 HOST_WIDE_INT ofs, int sign)
3919 rtx meml, memh, addr, extl, exth, tmp, mema;
3920 enum machine_mode mode;
3922 meml = gen_reg_rtx (DImode);
3923 memh = gen_reg_rtx (DImode);
3924 addr = gen_reg_rtx (DImode);
3925 extl = gen_reg_rtx (DImode);
3926 exth = gen_reg_rtx (DImode);
3928 mema = XEXP (mem, 0);
3929 if (GET_CODE (mema) == LO_SUM)
3930 mema = force_reg (Pmode, mema);
3932 /* AND addresses cannot be in any alias set, since they may implicitly
3933 alias surrounding code. Ideally we'd have some alias set that
3934 covered all types except those with alignment 8 or higher. */
3936 tmp = change_address (mem, DImode,
3937 gen_rtx_AND (DImode,
3938 plus_constant (mema, ofs),
3940 set_mem_alias_set (tmp, 0);
3941 emit_move_insn (meml, tmp);
3943 tmp = change_address (mem, DImode,
3944 gen_rtx_AND (DImode,
3945 plus_constant (mema, ofs + size - 1),
3947 set_mem_alias_set (tmp, 0);
3948 emit_move_insn (memh, tmp);
3950 if (WORDS_BIG_ENDIAN && sign && (size == 2 || size == 4))
3952 emit_move_insn (addr, plus_constant (mema, -1));
3954 emit_insn (gen_extqh_be (extl, meml, addr));
3955 emit_insn (gen_extxl_be (exth, memh, GEN_INT (64), addr));
3957 addr = expand_binop (DImode, ior_optab, extl, exth, tgt, 1, OPTAB_WIDEN);
3958 addr = expand_binop (DImode, ashr_optab, addr, GEN_INT (64 - size*8),
3959 addr, 1, OPTAB_WIDEN);
3961 else if (sign && size == 2)
3963 emit_move_insn (addr, plus_constant (mema, ofs+2));
3965 emit_insn (gen_extxl_le (extl, meml, GEN_INT (64), addr));
3966 emit_insn (gen_extqh_le (exth, memh, addr));
3968 /* We must use tgt here for the target. Alpha-vms port fails if we use
3969 addr for the target, because addr is marked as a pointer and combine
3970 knows that pointers are always sign-extended 32 bit values. */
3971 addr = expand_binop (DImode, ior_optab, extl, exth, tgt, 1, OPTAB_WIDEN);
3972 addr = expand_binop (DImode, ashr_optab, addr, GEN_INT (48),
3973 addr, 1, OPTAB_WIDEN);
3977 if (WORDS_BIG_ENDIAN)
3979 emit_move_insn (addr, plus_constant (mema, ofs+size-1));
3983 emit_insn (gen_extwh_be (extl, meml, addr));
3988 emit_insn (gen_extlh_be (extl, meml, addr));
3993 emit_insn (gen_extqh_be (extl, meml, addr));
4000 emit_insn (gen_extxl_be (exth, memh, GEN_INT (size*8), addr));
4004 emit_move_insn (addr, plus_constant (mema, ofs));
4005 emit_insn (gen_extxl_le (extl, meml, GEN_INT (size*8), addr));
4009 emit_insn (gen_extwh_le (exth, memh, addr));
4014 emit_insn (gen_extlh_le (exth, memh, addr));
4019 emit_insn (gen_extqh_le (exth, memh, addr));
4028 addr = expand_binop (mode, ior_optab, gen_lowpart (mode, extl),
4029 gen_lowpart (mode, exth), gen_lowpart (mode, tgt),
4034 emit_move_insn (tgt, gen_lowpart(GET_MODE (tgt), addr));
4037 /* Similarly, use ins and msk instructions to perform unaligned stores. */
4040 alpha_expand_unaligned_store (rtx dst, rtx src,
4041 HOST_WIDE_INT size, HOST_WIDE_INT ofs)
4043 rtx dstl, dsth, addr, insl, insh, meml, memh, dsta;
4045 dstl = gen_reg_rtx (DImode);
4046 dsth = gen_reg_rtx (DImode);
4047 insl = gen_reg_rtx (DImode);
4048 insh = gen_reg_rtx (DImode);
4050 dsta = XEXP (dst, 0);
4051 if (GET_CODE (dsta) == LO_SUM)
4052 dsta = force_reg (Pmode, dsta);
4054 /* AND addresses cannot be in any alias set, since they may implicitly
4055 alias surrounding code. Ideally we'd have some alias set that
4056 covered all types except those with alignment 8 or higher. */
4058 meml = change_address (dst, DImode,
4059 gen_rtx_AND (DImode,
4060 plus_constant (dsta, ofs),
4062 set_mem_alias_set (meml, 0);
4064 memh = change_address (dst, DImode,
4065 gen_rtx_AND (DImode,
4066 plus_constant (dsta, ofs + size - 1),
4068 set_mem_alias_set (memh, 0);
4070 emit_move_insn (dsth, memh);
4071 emit_move_insn (dstl, meml);
4072 if (WORDS_BIG_ENDIAN)
4074 addr = copy_addr_to_reg (plus_constant (dsta, ofs+size-1));
4076 if (src != const0_rtx)
4081 emit_insn (gen_inswl_be (insh, gen_lowpart (HImode,src), addr));
4084 emit_insn (gen_insll_be (insh, gen_lowpart (SImode,src), addr));
4087 emit_insn (gen_insql_be (insh, gen_lowpart (DImode,src), addr));
4090 emit_insn (gen_insxh (insl, gen_lowpart (DImode, src),
4091 GEN_INT (size*8), addr));
4097 emit_insn (gen_mskxl_be (dsth, dsth, GEN_INT (0xffff), addr));
4101 rtx msk = immed_double_const (0xffffffff, 0, DImode);
4102 emit_insn (gen_mskxl_be (dsth, dsth, msk, addr));
4106 emit_insn (gen_mskxl_be (dsth, dsth, constm1_rtx, addr));
4110 emit_insn (gen_mskxh (dstl, dstl, GEN_INT (size*8), addr));
4114 addr = copy_addr_to_reg (plus_constant (dsta, ofs));
4116 if (src != const0_rtx)
4118 emit_insn (gen_insxh (insh, gen_lowpart (DImode, src),
4119 GEN_INT (size*8), addr));
4124 emit_insn (gen_inswl_le (insl, gen_lowpart (HImode, src), addr));
4127 emit_insn (gen_insll_le (insl, gen_lowpart (SImode, src), addr));
4130 emit_insn (gen_insql_le (insl, src, addr));
4135 emit_insn (gen_mskxh (dsth, dsth, GEN_INT (size*8), addr));
4140 emit_insn (gen_mskxl_le (dstl, dstl, GEN_INT (0xffff), addr));
4144 rtx msk = immed_double_const (0xffffffff, 0, DImode);
4145 emit_insn (gen_mskxl_le (dstl, dstl, msk, addr));
4149 emit_insn (gen_mskxl_le (dstl, dstl, constm1_rtx, addr));
4154 if (src != const0_rtx)
4156 dsth = expand_binop (DImode, ior_optab, insh, dsth, dsth, 0, OPTAB_WIDEN);
4157 dstl = expand_binop (DImode, ior_optab, insl, dstl, dstl, 0, OPTAB_WIDEN);
4160 if (WORDS_BIG_ENDIAN)
4162 emit_move_insn (meml, dstl);
4163 emit_move_insn (memh, dsth);
4167 /* Must store high before low for degenerate case of aligned. */
4168 emit_move_insn (memh, dsth);
4169 emit_move_insn (meml, dstl);
4173 /* The block move code tries to maximize speed by separating loads and
4174 stores at the expense of register pressure: we load all of the data
4175 before we store it back out. There are two secondary effects worth
4176 mentioning, that this speeds copying to/from aligned and unaligned
4177 buffers, and that it makes the code significantly easier to write. */
4179 #define MAX_MOVE_WORDS 8
4181 /* Load an integral number of consecutive unaligned quadwords. */
4184 alpha_expand_unaligned_load_words (rtx *out_regs, rtx smem,
4185 HOST_WIDE_INT words, HOST_WIDE_INT ofs)
4187 rtx const im8 = GEN_INT (-8);
4188 rtx const i64 = GEN_INT (64);
4189 rtx ext_tmps[MAX_MOVE_WORDS], data_regs[MAX_MOVE_WORDS+1];
4190 rtx sreg, areg, tmp, smema;
4193 smema = XEXP (smem, 0);
4194 if (GET_CODE (smema) == LO_SUM)
4195 smema = force_reg (Pmode, smema);
4197 /* Generate all the tmp registers we need. */
4198 for (i = 0; i < words; ++i)
4200 data_regs[i] = out_regs[i];
4201 ext_tmps[i] = gen_reg_rtx (DImode);
4203 data_regs[words] = gen_reg_rtx (DImode);
4206 smem = adjust_address (smem, GET_MODE (smem), ofs);
4208 /* Load up all of the source data. */
4209 for (i = 0; i < words; ++i)
4211 tmp = change_address (smem, DImode,
4212 gen_rtx_AND (DImode,
4213 plus_constant (smema, 8*i),
4215 set_mem_alias_set (tmp, 0);
4216 emit_move_insn (data_regs[i], tmp);
4219 tmp = change_address (smem, DImode,
4220 gen_rtx_AND (DImode,
4221 plus_constant (smema, 8*words - 1),
4223 set_mem_alias_set (tmp, 0);
4224 emit_move_insn (data_regs[words], tmp);
4226 /* Extract the half-word fragments. Unfortunately DEC decided to make
4227 extxh with offset zero a noop instead of zeroing the register, so
4228 we must take care of that edge condition ourselves with cmov. */
4230 sreg = copy_addr_to_reg (smema);
4231 areg = expand_binop (DImode, and_optab, sreg, GEN_INT (7), NULL,
4233 if (WORDS_BIG_ENDIAN)
4234 emit_move_insn (sreg, plus_constant (sreg, 7));
4235 for (i = 0; i < words; ++i)
4237 if (WORDS_BIG_ENDIAN)
4239 emit_insn (gen_extqh_be (data_regs[i], data_regs[i], sreg));
4240 emit_insn (gen_extxl_be (ext_tmps[i], data_regs[i+1], i64, sreg));
4244 emit_insn (gen_extxl_le (data_regs[i], data_regs[i], i64, sreg));
4245 emit_insn (gen_extqh_le (ext_tmps[i], data_regs[i+1], sreg));
4247 emit_insn (gen_rtx_SET (VOIDmode, ext_tmps[i],
4248 gen_rtx_IF_THEN_ELSE (DImode,
4249 gen_rtx_EQ (DImode, areg,
4251 const0_rtx, ext_tmps[i])));
4254 /* Merge the half-words into whole words. */
4255 for (i = 0; i < words; ++i)
4257 out_regs[i] = expand_binop (DImode, ior_optab, data_regs[i],
4258 ext_tmps[i], data_regs[i], 1, OPTAB_WIDEN);
4262 /* Store an integral number of consecutive unaligned quadwords. DATA_REGS
4263 may be NULL to store zeros. */
4266 alpha_expand_unaligned_store_words (rtx *data_regs, rtx dmem,
4267 HOST_WIDE_INT words, HOST_WIDE_INT ofs)
4269 rtx const im8 = GEN_INT (-8);
4270 rtx const i64 = GEN_INT (64);
4271 rtx ins_tmps[MAX_MOVE_WORDS];
4272 rtx st_tmp_1, st_tmp_2, dreg;
4273 rtx st_addr_1, st_addr_2, dmema;
4276 dmema = XEXP (dmem, 0);
4277 if (GET_CODE (dmema) == LO_SUM)
4278 dmema = force_reg (Pmode, dmema);
4280 /* Generate all the tmp registers we need. */
4281 if (data_regs != NULL)
4282 for (i = 0; i < words; ++i)
4283 ins_tmps[i] = gen_reg_rtx(DImode);
4284 st_tmp_1 = gen_reg_rtx(DImode);
4285 st_tmp_2 = gen_reg_rtx(DImode);
4288 dmem = adjust_address (dmem, GET_MODE (dmem), ofs);
4290 st_addr_2 = change_address (dmem, DImode,
4291 gen_rtx_AND (DImode,
4292 plus_constant (dmema, words*8 - 1),
4294 set_mem_alias_set (st_addr_2, 0);
4296 st_addr_1 = change_address (dmem, DImode,
4297 gen_rtx_AND (DImode, dmema, im8));
4298 set_mem_alias_set (st_addr_1, 0);
4300 /* Load up the destination end bits. */
4301 emit_move_insn (st_tmp_2, st_addr_2);
4302 emit_move_insn (st_tmp_1, st_addr_1);
4304 /* Shift the input data into place. */
4305 dreg = copy_addr_to_reg (dmema);
4306 if (WORDS_BIG_ENDIAN)
4307 emit_move_insn (dreg, plus_constant (dreg, 7));
4308 if (data_regs != NULL)
4310 for (i = words-1; i >= 0; --i)
4312 if (WORDS_BIG_ENDIAN)
4314 emit_insn (gen_insql_be (ins_tmps[i], data_regs[i], dreg));
4315 emit_insn (gen_insxh (data_regs[i], data_regs[i], i64, dreg));
4319 emit_insn (gen_insxh (ins_tmps[i], data_regs[i], i64, dreg));
4320 emit_insn (gen_insql_le (data_regs[i], data_regs[i], dreg));
4323 for (i = words-1; i > 0; --i)
4325 ins_tmps[i-1] = expand_binop (DImode, ior_optab, data_regs[i],
4326 ins_tmps[i-1], ins_tmps[i-1], 1,
4331 /* Split and merge the ends with the destination data. */
4332 if (WORDS_BIG_ENDIAN)
4334 emit_insn (gen_mskxl_be (st_tmp_2, st_tmp_2, constm1_rtx, dreg));
4335 emit_insn (gen_mskxh (st_tmp_1, st_tmp_1, i64, dreg));
4339 emit_insn (gen_mskxh (st_tmp_2, st_tmp_2, i64, dreg));
4340 emit_insn (gen_mskxl_le (st_tmp_1, st_tmp_1, constm1_rtx, dreg));
4343 if (data_regs != NULL)
4345 st_tmp_2 = expand_binop (DImode, ior_optab, st_tmp_2, ins_tmps[words-1],
4346 st_tmp_2, 1, OPTAB_WIDEN);
4347 st_tmp_1 = expand_binop (DImode, ior_optab, st_tmp_1, data_regs[0],
4348 st_tmp_1, 1, OPTAB_WIDEN);
4352 if (WORDS_BIG_ENDIAN)
4353 emit_move_insn (st_addr_1, st_tmp_1);
4355 emit_move_insn (st_addr_2, st_tmp_2);
4356 for (i = words-1; i > 0; --i)
4358 rtx tmp = change_address (dmem, DImode,
4359 gen_rtx_AND (DImode,
4360 plus_constant(dmema,
4361 WORDS_BIG_ENDIAN ? i*8-1 : i*8),
4363 set_mem_alias_set (tmp, 0);
4364 emit_move_insn (tmp, data_regs ? ins_tmps[i-1] : const0_rtx);
4366 if (WORDS_BIG_ENDIAN)
4367 emit_move_insn (st_addr_2, st_tmp_2);
4369 emit_move_insn (st_addr_1, st_tmp_1);
4373 /* Expand string/block move operations.
4375 operands[0] is the pointer to the destination.
4376 operands[1] is the pointer to the source.
4377 operands[2] is the number of bytes to move.
4378 operands[3] is the alignment. */
4381 alpha_expand_block_move (rtx operands[])
4383 rtx bytes_rtx = operands[2];
4384 rtx align_rtx = operands[3];
4385 HOST_WIDE_INT orig_bytes = INTVAL (bytes_rtx);
4386 HOST_WIDE_INT bytes = orig_bytes;
4387 HOST_WIDE_INT src_align = INTVAL (align_rtx) * BITS_PER_UNIT;
4388 HOST_WIDE_INT dst_align = src_align;
4389 rtx orig_src = operands[1];
4390 rtx orig_dst = operands[0];
4391 rtx data_regs[2 * MAX_MOVE_WORDS + 16];
4393 unsigned int i, words, ofs, nregs = 0;
4395 if (orig_bytes <= 0)
4397 else if (orig_bytes > MAX_MOVE_WORDS * UNITS_PER_WORD)
4400 /* Look for additional alignment information from recorded register info. */
4402 tmp = XEXP (orig_src, 0);
4403 if (GET_CODE (tmp) == REG)
4404 src_align = MAX (src_align, REGNO_POINTER_ALIGN (REGNO (tmp)));
4405 else if (GET_CODE (tmp) == PLUS
4406 && GET_CODE (XEXP (tmp, 0)) == REG
4407 && GET_CODE (XEXP (tmp, 1)) == CONST_INT)
4409 unsigned HOST_WIDE_INT c = INTVAL (XEXP (tmp, 1));
4410 unsigned int a = REGNO_POINTER_ALIGN (REGNO (XEXP (tmp, 0)));
4414 if (a >= 64 && c % 8 == 0)
4416 else if (a >= 32 && c % 4 == 0)
4418 else if (a >= 16 && c % 2 == 0)
4423 tmp = XEXP (orig_dst, 0);
4424 if (GET_CODE (tmp) == REG)
4425 dst_align = MAX (dst_align, REGNO_POINTER_ALIGN (REGNO (tmp)));
4426 else if (GET_CODE (tmp) == PLUS
4427 && GET_CODE (XEXP (tmp, 0)) == REG
4428 && GET_CODE (XEXP (tmp, 1)) == CONST_INT)
4430 unsigned HOST_WIDE_INT c = INTVAL (XEXP (tmp, 1));
4431 unsigned int a = REGNO_POINTER_ALIGN (REGNO (XEXP (tmp, 0)));
4435 if (a >= 64 && c % 8 == 0)
4437 else if (a >= 32 && c % 4 == 0)
4439 else if (a >= 16 && c % 2 == 0)
4444 /* Load the entire block into registers. */
4445 if (GET_CODE (XEXP (orig_src, 0)) == ADDRESSOF)
4447 enum machine_mode mode;
4449 tmp = XEXP (XEXP (orig_src, 0), 0);
4451 /* Don't use the existing register if we're reading more than
4452 is held in the register. Nor if there is not a mode that
4453 handles the exact size. */
4454 mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 1);
4456 && GET_MODE_SIZE (GET_MODE (tmp)) >= bytes)
4460 data_regs[nregs] = gen_lowpart (DImode, tmp);
4461 data_regs[nregs + 1] = gen_highpart (DImode, tmp);
4465 data_regs[nregs++] = gen_lowpart (mode, tmp);
4470 /* No appropriate mode; fall back on memory. */
4471 orig_src = replace_equiv_address (orig_src,
4472 copy_addr_to_reg (XEXP (orig_src, 0)));
4473 src_align = GET_MODE_BITSIZE (GET_MODE (tmp));
4477 if (src_align >= 64 && bytes >= 8)
4481 for (i = 0; i < words; ++i)
4482 data_regs[nregs + i] = gen_reg_rtx (DImode);
4484 for (i = 0; i < words; ++i)
4485 emit_move_insn (data_regs[nregs + i],
4486 adjust_address (orig_src, DImode, ofs + i * 8));
4493 if (src_align >= 32 && bytes >= 4)
4497 for (i = 0; i < words; ++i)
4498 data_regs[nregs + i] = gen_reg_rtx (SImode);
4500 for (i = 0; i < words; ++i)
4501 emit_move_insn (data_regs[nregs + i],
4502 adjust_address (orig_src, SImode, ofs + i * 4));
4513 for (i = 0; i < words+1; ++i)
4514 data_regs[nregs + i] = gen_reg_rtx (DImode);
4516 alpha_expand_unaligned_load_words (data_regs + nregs, orig_src,
4524 if (! TARGET_BWX && bytes >= 4)
4526 data_regs[nregs++] = tmp = gen_reg_rtx (SImode);
4527 alpha_expand_unaligned_load (tmp, orig_src, 4, ofs, 0);
4534 if (src_align >= 16)
4537 data_regs[nregs++] = tmp = gen_reg_rtx (HImode);
4538 emit_move_insn (tmp, adjust_address (orig_src, HImode, ofs));
4541 } while (bytes >= 2);
4543 else if (! TARGET_BWX)
4545 data_regs[nregs++] = tmp = gen_reg_rtx (HImode);
4546 alpha_expand_unaligned_load (tmp, orig_src, 2, ofs, 0);
4554 data_regs[nregs++] = tmp = gen_reg_rtx (QImode);
4555 emit_move_insn (tmp, adjust_address (orig_src, QImode, ofs));
4562 if (nregs > ARRAY_SIZE (data_regs))
4565 /* Now save it back out again. */
4569 if (GET_CODE (XEXP (orig_dst, 0)) == ADDRESSOF)
4571 enum machine_mode mode;
4572 tmp = XEXP (XEXP (orig_dst, 0), 0);
4574 mode = mode_for_size (orig_bytes * BITS_PER_UNIT, MODE_INT, 1);
4575 if (GET_MODE (tmp) == mode)
4579 emit_move_insn (tmp, data_regs[0]);
4584 else if (nregs == 2 && mode == TImode)
4586 /* Undo the subregging done above when copying between
4587 two TImode registers. */
4588 if (GET_CODE (data_regs[0]) == SUBREG
4589 && GET_MODE (SUBREG_REG (data_regs[0])) == TImode)
4590 emit_move_insn (tmp, SUBREG_REG (data_regs[0]));
4596 emit_move_insn (gen_lowpart (DImode, tmp), data_regs[0]);
4597 emit_move_insn (gen_highpart (DImode, tmp), data_regs[1]);
4601 emit_no_conflict_block (seq, tmp, data_regs[0],
4602 data_regs[1], NULL_RTX);
4610 /* ??? If nregs > 1, consider reconstructing the word in regs. */
4611 /* ??? Optimize mode < dst_mode with strict_low_part. */
4613 /* No appropriate mode; fall back on memory. We can speed things
4614 up by recognizing extra alignment information. */
4615 orig_dst = replace_equiv_address (orig_dst,
4616 copy_addr_to_reg (XEXP (orig_dst, 0)));
4617 dst_align = GET_MODE_BITSIZE (GET_MODE (tmp));
4620 /* Write out the data in whatever chunks reading the source allowed. */
4621 if (dst_align >= 64)
4623 while (i < nregs && GET_MODE (data_regs[i]) == DImode)
4625 emit_move_insn (adjust_address (orig_dst, DImode, ofs),
4632 if (dst_align >= 32)
4634 /* If the source has remaining DImode regs, write them out in
4636 while (i < nregs && GET_MODE (data_regs[i]) == DImode)
4638 tmp = expand_binop (DImode, lshr_optab, data_regs[i], GEN_INT (32),
4639 NULL_RTX, 1, OPTAB_WIDEN);
4641 emit_move_insn (adjust_address (orig_dst, SImode, ofs),
4642 gen_lowpart (SImode, data_regs[i]));
4643 emit_move_insn (adjust_address (orig_dst, SImode, ofs + 4),
4644 gen_lowpart (SImode, tmp));
4649 while (i < nregs && GET_MODE (data_regs[i]) == SImode)
4651 emit_move_insn (adjust_address (orig_dst, SImode, ofs),
4658 if (i < nregs && GET_MODE (data_regs[i]) == DImode)
4660 /* Write out a remaining block of words using unaligned methods. */
4662 for (words = 1; i + words < nregs; words++)
4663 if (GET_MODE (data_regs[i + words]) != DImode)
4667 alpha_expand_unaligned_store (orig_dst, data_regs[i], 8, ofs);
4669 alpha_expand_unaligned_store_words (data_regs + i, orig_dst,
4676 /* Due to the above, this won't be aligned. */
4677 /* ??? If we have more than one of these, consider constructing full
4678 words in registers and using alpha_expand_unaligned_store_words. */
4679 while (i < nregs && GET_MODE (data_regs[i]) == SImode)
4681 alpha_expand_unaligned_store (orig_dst, data_regs[i], 4, ofs);
4686 if (dst_align >= 16)
4687 while (i < nregs && GET_MODE (data_regs[i]) == HImode)
4689 emit_move_insn (adjust_address (orig_dst, HImode, ofs), data_regs[i]);
4694 while (i < nregs && GET_MODE (data_regs[i]) == HImode)
4696 alpha_expand_unaligned_store (orig_dst, data_regs[i], 2, ofs);
4701 while (i < nregs && GET_MODE (data_regs[i]) == QImode)
4703 emit_move_insn (adjust_address (orig_dst, QImode, ofs), data_regs[i]);
4717 alpha_expand_block_clear (rtx operands[])
4719 rtx bytes_rtx = operands[1];
4720 rtx align_rtx = operands[2];
4721 HOST_WIDE_INT orig_bytes = INTVAL (bytes_rtx);
4722 HOST_WIDE_INT bytes = orig_bytes;
4723 HOST_WIDE_INT align = INTVAL (align_rtx) * BITS_PER_UNIT;
4724 HOST_WIDE_INT alignofs = 0;
4725 rtx orig_dst = operands[0];
4727 int i, words, ofs = 0;
4729 if (orig_bytes <= 0)
4731 if (orig_bytes > MAX_MOVE_WORDS * UNITS_PER_WORD)
4734 /* Look for stricter alignment. */
4735 tmp = XEXP (orig_dst, 0);
4736 if (GET_CODE (tmp) == REG)
4737 align = MAX (align, REGNO_POINTER_ALIGN (REGNO (tmp)));
4738 else if (GET_CODE (tmp) == PLUS
4739 && GET_CODE (XEXP (tmp, 0)) == REG
4740 && GET_CODE (XEXP (tmp, 1)) == CONST_INT)
4742 HOST_WIDE_INT c = INTVAL (XEXP (tmp, 1));
4743 int a = REGNO_POINTER_ALIGN (REGNO (XEXP (tmp, 0)));
4748 align = a, alignofs = 8 - c % 8;
4750 align = a, alignofs = 4 - c % 4;
4752 align = a, alignofs = 2 - c % 2;
4755 else if (GET_CODE (tmp) == ADDRESSOF)
4757 enum machine_mode mode;
4759 mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 1);
4760 if (GET_MODE (XEXP (tmp, 0)) == mode)
4762 emit_move_insn (XEXP (tmp, 0), const0_rtx);
4766 /* No appropriate mode; fall back on memory. */
4767 orig_dst = replace_equiv_address (orig_dst, copy_addr_to_reg (tmp));
4768 align = GET_MODE_BITSIZE (GET_MODE (XEXP (tmp, 0)));
4771 /* Handle an unaligned prefix first. */
4775 #if HOST_BITS_PER_WIDE_INT >= 64
4776 /* Given that alignofs is bounded by align, the only time BWX could
4777 generate three stores is for a 7 byte fill. Prefer two individual
4778 stores over a load/mask/store sequence. */
4779 if ((!TARGET_BWX || alignofs == 7)
4781 && !(alignofs == 4 && bytes >= 4))
4783 enum machine_mode mode = (align >= 64 ? DImode : SImode);
4784 int inv_alignofs = (align >= 64 ? 8 : 4) - alignofs;
4788 mem = adjust_address (orig_dst, mode, ofs - inv_alignofs);
4789 set_mem_alias_set (mem, 0);
4791 mask = ~(~(HOST_WIDE_INT)0 << (inv_alignofs * 8));
4792 if (bytes < alignofs)
4794 mask |= ~(HOST_WIDE_INT)0 << ((inv_alignofs + bytes) * 8);
4805 tmp = expand_binop (mode, and_optab, mem, GEN_INT (mask),
4806 NULL_RTX, 1, OPTAB_WIDEN);
4808 emit_move_insn (mem, tmp);
4812 if (TARGET_BWX && (alignofs & 1) && bytes >= 1)
4814 emit_move_insn (adjust_address (orig_dst, QImode, ofs), const0_rtx);
4819 if (TARGET_BWX && align >= 16 && (alignofs & 3) == 2 && bytes >= 2)
4821 emit_move_insn (adjust_address (orig_dst, HImode, ofs), const0_rtx);
4826 if (alignofs == 4 && bytes >= 4)
4828 emit_move_insn (adjust_address (orig_dst, SImode, ofs), const0_rtx);
4834 /* If we've not used the extra lead alignment information by now,
4835 we won't be able to. Downgrade align to match what's left over. */
4838 alignofs = alignofs & -alignofs;
4839 align = MIN (align, alignofs * BITS_PER_UNIT);
4843 /* Handle a block of contiguous long-words. */
4845 if (align >= 64 && bytes >= 8)
4849 for (i = 0; i < words; ++i)
4850 emit_move_insn (adjust_address (orig_dst, DImode, ofs + i * 8),
4857 /* If the block is large and appropriately aligned, emit a single
4858 store followed by a sequence of stq_u insns. */
4860 if (align >= 32 && bytes > 16)
4864 emit_move_insn (adjust_address (orig_dst, SImode, ofs), const0_rtx);
4868 orig_dsta = XEXP (orig_dst, 0);
4869 if (GET_CODE (orig_dsta) == LO_SUM)
4870 orig_dsta = force_reg (Pmode, orig_dsta);
4873 for (i = 0; i < words; ++i)
4876 = change_address (orig_dst, DImode,
4877 gen_rtx_AND (DImode,
4878 plus_constant (orig_dsta, ofs + i*8),
4880 set_mem_alias_set (mem, 0);
4881 emit_move_insn (mem, const0_rtx);
4884 /* Depending on the alignment, the first stq_u may have overlapped
4885 with the initial stl, which means that the last stq_u didn't
4886 write as much as it would appear. Leave those questionable bytes
4888 bytes -= words * 8 - 4;
4889 ofs += words * 8 - 4;
4892 /* Handle a smaller block of aligned words. */
4894 if ((align >= 64 && bytes == 4)
4895 || (align == 32 && bytes >= 4))
4899 for (i = 0; i < words; ++i)
4900 emit_move_insn (adjust_address (orig_dst, SImode, ofs + i * 4),
4907 /* An unaligned block uses stq_u stores for as many as possible. */
4913 alpha_expand_unaligned_store_words (NULL, orig_dst, words, ofs);
4919 /* Next clean up any trailing pieces. */
4921 #if HOST_BITS_PER_WIDE_INT >= 64
4922 /* Count the number of bits in BYTES for which aligned stores could
4925 for (i = (TARGET_BWX ? 1 : 4); i * BITS_PER_UNIT <= align ; i <<= 1)
4929 /* If we have appropriate alignment (and it wouldn't take too many
4930 instructions otherwise), mask out the bytes we need. */
4931 if (TARGET_BWX ? words > 2 : bytes > 0)
4938 mem = adjust_address (orig_dst, DImode, ofs);
4939 set_mem_alias_set (mem, 0);
4941 mask = ~(HOST_WIDE_INT)0 << (bytes * 8);
4943 tmp = expand_binop (DImode, and_optab, mem, GEN_INT (mask),
4944 NULL_RTX, 1, OPTAB_WIDEN);
4946 emit_move_insn (mem, tmp);
4949 else if (align >= 32 && bytes < 4)
4954 mem = adjust_address (orig_dst, SImode, ofs);
4955 set_mem_alias_set (mem, 0);
4957 mask = ~(HOST_WIDE_INT)0 << (bytes * 8);
4959 tmp = expand_binop (SImode, and_optab, mem, GEN_INT (mask),
4960 NULL_RTX, 1, OPTAB_WIDEN);
4962 emit_move_insn (mem, tmp);
4968 if (!TARGET_BWX && bytes >= 4)
4970 alpha_expand_unaligned_store (orig_dst, const0_rtx, 4, ofs);
4980 emit_move_insn (adjust_address (orig_dst, HImode, ofs),
4984 } while (bytes >= 2);
4986 else if (! TARGET_BWX)
4988 alpha_expand_unaligned_store (orig_dst, const0_rtx, 2, ofs);
4996 emit_move_insn (adjust_address (orig_dst, QImode, ofs), const0_rtx);
5004 /* Returns a mask so that zap(x, value) == x & mask. */
5007 alpha_expand_zap_mask (HOST_WIDE_INT value)
5012 if (HOST_BITS_PER_WIDE_INT >= 64)
5014 HOST_WIDE_INT mask = 0;
5016 for (i = 7; i >= 0; --i)
5019 if (!((value >> i) & 1))
5023 result = gen_int_mode (mask, DImode);
5025 else if (HOST_BITS_PER_WIDE_INT == 32)
5027 HOST_WIDE_INT mask_lo = 0, mask_hi = 0;
5029 for (i = 7; i >= 4; --i)
5032 if (!((value >> i) & 1))
5036 for (i = 3; i >= 0; --i)
5039 if (!((value >> i) & 1))
5043 result = immed_double_const (mask_lo, mask_hi, DImode);
5052 alpha_expand_builtin_vector_binop (rtx (*gen) (rtx, rtx, rtx),
5053 enum machine_mode mode,
5054 rtx op0, rtx op1, rtx op2)
5056 op0 = gen_lowpart (mode, op0);
5058 if (op1 == const0_rtx)
5059 op1 = CONST0_RTX (mode);
5061 op1 = gen_lowpart (mode, op1);
5063 if (op2 == const0_rtx)
5064 op2 = CONST0_RTX (mode);
5066 op2 = gen_lowpart (mode, op2);
5068 emit_insn ((*gen) (op0, op1, op2));
5071 /* Adjust the cost of a scheduling dependency. Return the new cost of
5072 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
5075 alpha_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
5077 enum attr_type insn_type, dep_insn_type;
5079 /* If the dependence is an anti-dependence, there is no cost. For an
5080 output dependence, there is sometimes a cost, but it doesn't seem
5081 worth handling those few cases. */
5082 if (REG_NOTE_KIND (link) != 0)
5085 /* If we can't recognize the insns, we can't really do anything. */
5086 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
5089 insn_type = get_attr_type (insn);
5090 dep_insn_type = get_attr_type (dep_insn);
5092 /* Bring in the user-defined memory latency. */
5093 if (dep_insn_type == TYPE_ILD
5094 || dep_insn_type == TYPE_FLD
5095 || dep_insn_type == TYPE_LDSYM)
5096 cost += alpha_memory_latency-1;
5098 /* Everything else handled in DFA bypasses now. */
5103 /* The number of instructions that can be issued per cycle. */
5106 alpha_issue_rate (void)
5108 return (alpha_cpu == PROCESSOR_EV4 ? 2 : 4);
5112 alpha_use_dfa_pipeline_interface (void)
5117 /* How many alternative schedules to try. This should be as wide as the
5118 scheduling freedom in the DFA, but no wider. Making this value too
5119 large results extra work for the scheduler.
5121 For EV4, loads can be issued to either IB0 or IB1, thus we have 2
5122 alternative schedules. For EV5, we can choose between E0/E1 and
5123 FA/FM. For EV6, an arithmetic insn can be issued to U0/U1/L0/L1. */
5126 alpha_multipass_dfa_lookahead (void)
5128 return (alpha_cpu == PROCESSOR_EV6 ? 4 : 2);
5131 /* Machine-specific function data. */
5133 struct machine_function GTY(())
5136 /* List of call information words for calls from this function. */
5137 struct rtx_def *first_ciw;
5138 struct rtx_def *last_ciw;
5141 /* List of deferred case vectors. */
5142 struct rtx_def *addr_list;
5145 const char *some_ld_name;
5148 /* How to allocate a 'struct machine_function'. */
5150 static struct machine_function *
5151 alpha_init_machine_status (void)
5153 return ((struct machine_function *)
5154 ggc_alloc_cleared (sizeof (struct machine_function)));
5157 /* Functions to save and restore alpha_return_addr_rtx. */
5159 /* Start the ball rolling with RETURN_ADDR_RTX. */
5162 alpha_return_addr (int count, rtx frame ATTRIBUTE_UNUSED)
5167 return get_hard_reg_initial_val (Pmode, REG_RA);
5170 /* Return or create a pseudo containing the gp value for the current
5171 function. Needed only if TARGET_LD_BUGGY_LDGP. */
5174 alpha_gp_save_rtx (void)
5176 rtx r = get_hard_reg_initial_val (DImode, 29);
5177 if (GET_CODE (r) != MEM)
5178 r = gen_mem_addressof (r, NULL_TREE, /*rescan=*/true);
5183 alpha_ra_ever_killed (void)
5187 if (!has_hard_reg_initial_val (Pmode, REG_RA))
5188 return regs_ever_live[REG_RA];
5190 push_topmost_sequence ();
5192 pop_topmost_sequence ();
5194 return reg_set_between_p (gen_rtx_REG (Pmode, REG_RA), top, NULL_RTX);
5198 /* Return the trap mode suffix applicable to the current
5199 instruction, or NULL. */
5202 get_trap_mode_suffix (void)
5204 enum attr_trap_suffix s = get_attr_trap_suffix (current_output_insn);
5208 case TRAP_SUFFIX_NONE:
5211 case TRAP_SUFFIX_SU:
5212 if (alpha_fptm >= ALPHA_FPTM_SU)
5216 case TRAP_SUFFIX_SUI:
5217 if (alpha_fptm >= ALPHA_FPTM_SUI)
5221 case TRAP_SUFFIX_V_SV:
5229 case ALPHA_FPTM_SUI:
5234 case TRAP_SUFFIX_V_SV_SVI:
5243 case ALPHA_FPTM_SUI:
5248 case TRAP_SUFFIX_U_SU_SUI:
5257 case ALPHA_FPTM_SUI:
5265 /* Return the rounding mode suffix applicable to the current
5266 instruction, or NULL. */
5269 get_round_mode_suffix (void)
5271 enum attr_round_suffix s = get_attr_round_suffix (current_output_insn);
5275 case ROUND_SUFFIX_NONE:
5277 case ROUND_SUFFIX_NORMAL:
5280 case ALPHA_FPRM_NORM:
5282 case ALPHA_FPRM_MINF:
5284 case ALPHA_FPRM_CHOP:
5286 case ALPHA_FPRM_DYN:
5291 case ROUND_SUFFIX_C:
5297 /* Locate some local-dynamic symbol still in use by this function
5298 so that we can print its name in some movdi_er_tlsldm pattern. */
5301 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
5305 if (GET_CODE (x) == SYMBOL_REF
5306 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
5308 cfun->machine->some_ld_name = XSTR (x, 0);
5316 get_some_local_dynamic_name (void)
5320 if (cfun->machine->some_ld_name)
5321 return cfun->machine->some_ld_name;
5323 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
5325 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
5326 return cfun->machine->some_ld_name;
5331 /* Print an operand. Recognize special options, documented below. */
5334 print_operand (FILE *file, rtx x, int code)
5341 /* Print the assembler name of the current function. */
5342 assemble_name (file, alpha_fnname);
5346 assemble_name (file, get_some_local_dynamic_name ());
5351 const char *trap = get_trap_mode_suffix ();
5352 const char *round = get_round_mode_suffix ();
5355 fprintf (file, (TARGET_AS_SLASH_BEFORE_SUFFIX ? "/%s%s" : "%s%s"),
5356 (trap ? trap : ""), (round ? round : ""));
5361 /* Generates single precision instruction suffix. */
5362 fputc ((TARGET_FLOAT_VAX ? 'f' : 's'), file);
5366 /* Generates double precision instruction suffix. */
5367 fputc ((TARGET_FLOAT_VAX ? 'g' : 't'), file);
5371 /* Generates a nop after a noreturn call at the very end of the
5373 if (next_real_insn (current_output_insn) == 0)
5374 fprintf (file, "\n\tnop");
5378 if (alpha_this_literal_sequence_number == 0)
5379 alpha_this_literal_sequence_number = alpha_next_sequence_number++;
5380 fprintf (file, "%d", alpha_this_literal_sequence_number);
5384 if (alpha_this_gpdisp_sequence_number == 0)
5385 alpha_this_gpdisp_sequence_number = alpha_next_sequence_number++;
5386 fprintf (file, "%d", alpha_this_gpdisp_sequence_number);
5390 if (GET_CODE (x) == HIGH)
5391 output_addr_const (file, XEXP (x, 0));
5393 output_operand_lossage ("invalid %%H value");
5400 if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLSGD_CALL)
5402 x = XVECEXP (x, 0, 0);
5403 lituse = "lituse_tlsgd";
5405 else if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLSLDM_CALL)
5407 x = XVECEXP (x, 0, 0);
5408 lituse = "lituse_tlsldm";
5410 else if (GET_CODE (x) == CONST_INT)
5411 lituse = "lituse_jsr";
5414 output_operand_lossage ("invalid %%J value");
5418 if (x != const0_rtx)
5419 fprintf (file, "\t\t!%s!%d", lituse, (int) INTVAL (x));
5424 /* If this operand is the constant zero, write it as "$31". */
5425 if (GET_CODE (x) == REG)
5426 fprintf (file, "%s", reg_names[REGNO (x)]);
5427 else if (x == CONST0_RTX (GET_MODE (x)))
5428 fprintf (file, "$31");
5430 output_operand_lossage ("invalid %%r value");
5434 /* Similar, but for floating-point. */
5435 if (GET_CODE (x) == REG)
5436 fprintf (file, "%s", reg_names[REGNO (x)]);
5437 else if (x == CONST0_RTX (GET_MODE (x)))
5438 fprintf (file, "$f31");
5440 output_operand_lossage ("invalid %%R value");
5444 /* Write the 1's complement of a constant. */
5445 if (GET_CODE (x) != CONST_INT)
5446 output_operand_lossage ("invalid %%N value");
5448 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~ INTVAL (x));
5452 /* Write 1 << C, for a constant C. */
5453 if (GET_CODE (x) != CONST_INT)
5454 output_operand_lossage ("invalid %%P value");
5456 fprintf (file, HOST_WIDE_INT_PRINT_DEC, (HOST_WIDE_INT) 1 << INTVAL (x));
5460 /* Write the high-order 16 bits of a constant, sign-extended. */
5461 if (GET_CODE (x) != CONST_INT)
5462 output_operand_lossage ("invalid %%h value");
5464 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x) >> 16);
5468 /* Write the low-order 16 bits of a constant, sign-extended. */
5469 if (GET_CODE (x) != CONST_INT)
5470 output_operand_lossage ("invalid %%L value");
5472 fprintf (file, HOST_WIDE_INT_PRINT_DEC,
5473 (INTVAL (x) & 0xffff) - 2 * (INTVAL (x) & 0x8000));
5477 /* Write mask for ZAP insn. */
5478 if (GET_CODE (x) == CONST_DOUBLE)
5480 HOST_WIDE_INT mask = 0;
5481 HOST_WIDE_INT value;
5483 value = CONST_DOUBLE_LOW (x);
5484 for (i = 0; i < HOST_BITS_PER_WIDE_INT / HOST_BITS_PER_CHAR;
5489 value = CONST_DOUBLE_HIGH (x);
5490 for (i = 0; i < HOST_BITS_PER_WIDE_INT / HOST_BITS_PER_CHAR;
5493 mask |= (1 << (i + sizeof (int)));
5495 fprintf (file, HOST_WIDE_INT_PRINT_DEC, mask & 0xff);
5498 else if (GET_CODE (x) == CONST_INT)
5500 HOST_WIDE_INT mask = 0, value = INTVAL (x);
5502 for (i = 0; i < 8; i++, value >>= 8)
5506 fprintf (file, HOST_WIDE_INT_PRINT_DEC, mask);
5509 output_operand_lossage ("invalid %%m value");
5513 /* 'b', 'w', 'l', or 'q' as the value of the constant. */
5514 if (GET_CODE (x) != CONST_INT
5515 || (INTVAL (x) != 8 && INTVAL (x) != 16
5516 && INTVAL (x) != 32 && INTVAL (x) != 64))
5517 output_operand_lossage ("invalid %%M value");
5519 fprintf (file, "%s",
5520 (INTVAL (x) == 8 ? "b"
5521 : INTVAL (x) == 16 ? "w"
5522 : INTVAL (x) == 32 ? "l"
5527 /* Similar, except do it from the mask. */
5528 if (GET_CODE (x) == CONST_INT)
5530 HOST_WIDE_INT value = INTVAL (x);
5537 if (value == 0xffff)
5542 if (value == 0xffffffff)
5553 else if (HOST_BITS_PER_WIDE_INT == 32
5554 && GET_CODE (x) == CONST_DOUBLE
5555 && CONST_DOUBLE_LOW (x) == 0xffffffff
5556 && CONST_DOUBLE_HIGH (x) == 0)
5561 output_operand_lossage ("invalid %%U value");
5565 /* Write the constant value divided by 8 for little-endian mode or
5566 (56 - value) / 8 for big-endian mode. */
5568 if (GET_CODE (x) != CONST_INT
5569 || (unsigned HOST_WIDE_INT) INTVAL (x) >= (WORDS_BIG_ENDIAN
5572 || (INTVAL (x) & 7) != 0)
5573 output_operand_lossage ("invalid %%s value");
5575 fprintf (file, HOST_WIDE_INT_PRINT_DEC,
5577 ? (56 - INTVAL (x)) / 8
5582 /* Same, except compute (64 - c) / 8 */
5584 if (GET_CODE (x) != CONST_INT
5585 && (unsigned HOST_WIDE_INT) INTVAL (x) >= 64
5586 && (INTVAL (x) & 7) != 8)
5587 output_operand_lossage ("invalid %%s value");
5589 fprintf (file, HOST_WIDE_INT_PRINT_DEC, (64 - INTVAL (x)) / 8);
5594 /* On Unicos/Mk systems: use a DEX expression if the symbol
5595 clashes with a register name. */
5596 int dex = unicosmk_need_dex (x);
5598 fprintf (file, "DEX(%d)", dex);
5600 output_addr_const (file, x);
5604 case 'C': case 'D': case 'c': case 'd':
5605 /* Write out comparison name. */
5607 enum rtx_code c = GET_CODE (x);
5609 if (GET_RTX_CLASS (c) != '<')
5610 output_operand_lossage ("invalid %%C value");
5612 else if (code == 'D')
5613 c = reverse_condition (c);
5614 else if (code == 'c')
5615 c = swap_condition (c);
5616 else if (code == 'd')
5617 c = swap_condition (reverse_condition (c));
5620 fprintf (file, "ule");
5622 fprintf (file, "ult");
5623 else if (c == UNORDERED)
5624 fprintf (file, "un");
5626 fprintf (file, "%s", GET_RTX_NAME (c));
5631 /* Write the divide or modulus operator. */
5632 switch (GET_CODE (x))
5635 fprintf (file, "div%s", GET_MODE (x) == SImode ? "l" : "q");
5638 fprintf (file, "div%su", GET_MODE (x) == SImode ? "l" : "q");
5641 fprintf (file, "rem%s", GET_MODE (x) == SImode ? "l" : "q");
5644 fprintf (file, "rem%su", GET_MODE (x) == SImode ? "l" : "q");
5647 output_operand_lossage ("invalid %%E value");
5653 /* Write "_u" for unaligned access. */
5654 if (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == AND)
5655 fprintf (file, "_u");
5659 if (GET_CODE (x) == REG)
5660 fprintf (file, "%s", reg_names[REGNO (x)]);
5661 else if (GET_CODE (x) == MEM)
5662 output_address (XEXP (x, 0));
5663 else if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == UNSPEC)
5665 switch (XINT (XEXP (x, 0), 1))
5669 output_addr_const (file, XVECEXP (XEXP (x, 0), 0, 0));
5672 output_operand_lossage ("unknown relocation unspec");
5677 output_addr_const (file, x);
5681 output_operand_lossage ("invalid %%xn code");
5686 print_operand_address (FILE *file, rtx addr)
5689 HOST_WIDE_INT offset = 0;
5691 if (GET_CODE (addr) == AND)
5692 addr = XEXP (addr, 0);
5694 if (GET_CODE (addr) == PLUS
5695 && GET_CODE (XEXP (addr, 1)) == CONST_INT)
5697 offset = INTVAL (XEXP (addr, 1));
5698 addr = XEXP (addr, 0);
5701 if (GET_CODE (addr) == LO_SUM)
5703 const char *reloc16, *reloclo;
5704 rtx op1 = XEXP (addr, 1);
5706 if (GET_CODE (op1) == CONST && GET_CODE (XEXP (op1, 0)) == UNSPEC)
5708 op1 = XEXP (op1, 0);
5709 switch (XINT (op1, 1))
5713 reloclo = (alpha_tls_size == 16 ? "dtprel" : "dtprello");
5717 reloclo = (alpha_tls_size == 16 ? "tprel" : "tprello");
5720 output_operand_lossage ("unknown relocation unspec");
5724 output_addr_const (file, XVECEXP (op1, 0, 0));
5729 reloclo = "gprellow";
5730 output_addr_const (file, op1);
5734 fprintf (file, "+" HOST_WIDE_INT_PRINT_DEC, offset);
5736 addr = XEXP (addr, 0);
5737 if (GET_CODE (addr) == REG)
5738 basereg = REGNO (addr);
5739 else if (GET_CODE (addr) == SUBREG
5740 && GET_CODE (SUBREG_REG (addr)) == REG)
5741 basereg = subreg_regno (addr);
5745 fprintf (file, "($%d)\t\t!%s", basereg,
5746 (basereg == 29 ? reloc16 : reloclo));
5750 if (GET_CODE (addr) == REG)
5751 basereg = REGNO (addr);
5752 else if (GET_CODE (addr) == SUBREG
5753 && GET_CODE (SUBREG_REG (addr)) == REG)
5754 basereg = subreg_regno (addr);
5755 else if (GET_CODE (addr) == CONST_INT)
5756 offset = INTVAL (addr);
5758 #if TARGET_ABI_OPEN_VMS
5759 else if (GET_CODE (addr) == SYMBOL_REF)
5761 fprintf (file, "%s", XSTR (addr, 0));
5764 else if (GET_CODE (addr) == CONST
5765 && GET_CODE (XEXP (addr, 0)) == PLUS
5766 && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF)
5768 fprintf (file, "%s+" HOST_WIDE_INT_PRINT_DEC,
5769 XSTR (XEXP (XEXP (addr, 0), 0), 0),
5770 INTVAL (XEXP (XEXP (addr, 0), 1)));
5778 fprintf (file, HOST_WIDE_INT_PRINT_DEC "($%d)", offset, basereg);
5781 /* Emit RTL insns to initialize the variable parts of a trampoline at
5782 TRAMP. FNADDR is an RTX for the address of the function's pure
5783 code. CXT is an RTX for the static chain value for the function.
5785 The three offset parameters are for the individual template's
5786 layout. A JMPOFS < 0 indicates that the trampoline does not
5787 contain instructions at all.
5789 We assume here that a function will be called many more times than
5790 its address is taken (e.g., it might be passed to qsort), so we
5791 take the trouble to initialize the "hint" field in the JMP insn.
5792 Note that the hint field is PC (new) + 4 * bits 13:0. */
5795 alpha_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt,
5796 int fnofs, int cxtofs, int jmpofs)
5798 rtx temp, temp1, addr;
5799 /* VMS really uses DImode pointers in memory at this point. */
5800 enum machine_mode mode = TARGET_ABI_OPEN_VMS ? Pmode : ptr_mode;
5802 #ifdef POINTERS_EXTEND_UNSIGNED
5803 fnaddr = convert_memory_address (mode, fnaddr);
5804 cxt = convert_memory_address (mode, cxt);
5807 /* Store function address and CXT. */
5808 addr = memory_address (mode, plus_constant (tramp, fnofs));
5809 emit_move_insn (gen_rtx_MEM (mode, addr), fnaddr);
5810 addr = memory_address (mode, plus_constant (tramp, cxtofs));
5811 emit_move_insn (gen_rtx_MEM (mode, addr), cxt);
5813 /* This has been disabled since the hint only has a 32k range, and in
5814 no existing OS is the stack within 32k of the text segment. */
5815 if (0 && jmpofs >= 0)
5817 /* Compute hint value. */
5818 temp = force_operand (plus_constant (tramp, jmpofs+4), NULL_RTX);
5819 temp = expand_binop (DImode, sub_optab, fnaddr, temp, temp, 1,
5821 temp = expand_shift (RSHIFT_EXPR, Pmode, temp,
5822 build_int_2 (2, 0), NULL_RTX, 1);
5823 temp = expand_and (SImode, gen_lowpart (SImode, temp),
5824 GEN_INT (0x3fff), 0);
5826 /* Merge in the hint. */
5827 addr = memory_address (SImode, plus_constant (tramp, jmpofs));
5828 temp1 = force_reg (SImode, gen_rtx_MEM (SImode, addr));
5829 temp1 = expand_and (SImode, temp1, GEN_INT (0xffffc000), NULL_RTX);
5830 temp1 = expand_binop (SImode, ior_optab, temp1, temp, temp1, 1,
5832 emit_move_insn (gen_rtx_MEM (SImode, addr), temp1);
5835 #ifdef TRANSFER_FROM_TRAMPOLINE
5836 emit_library_call (init_one_libfunc ("__enable_execute_stack"),
5837 0, VOIDmode, 1, tramp, Pmode);
5841 emit_insn (gen_imb ());
5844 /* Determine where to put an argument to a function.
5845 Value is zero to push the argument on the stack,
5846 or a hard register in which to store the argument.
5848 MODE is the argument's machine mode.
5849 TYPE is the data type of the argument (as a tree).
5850 This is null for libcalls where that information may
5852 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5853 the preceding args and about the function being called.
5854 NAMED is nonzero if this argument is a named parameter
5855 (otherwise it is an extra parameter matching an ellipsis).
5857 On Alpha the first 6 words of args are normally in registers
5858 and the rest are pushed. */
5861 function_arg (CUMULATIVE_ARGS cum, enum machine_mode mode, tree type,
5862 int named ATTRIBUTE_UNUSED)
5867 /* Don't get confused and pass small structures in FP registers. */
5868 if (type && AGGREGATE_TYPE_P (type))
5872 #ifdef ENABLE_CHECKING
5873 /* With SPLIT_COMPLEX_ARGS, we shouldn't see any raw complex
5875 if (COMPLEX_MODE_P (mode))
5879 /* Set up defaults for FP operands passed in FP registers, and
5880 integral operands passed in integer registers. */
5881 if (TARGET_FPREGS && GET_MODE_CLASS (mode) == MODE_FLOAT)
5887 /* ??? Irritatingly, the definition of CUMULATIVE_ARGS is different for
5888 the three platforms, so we can't avoid conditional compilation. */
5889 #if TARGET_ABI_OPEN_VMS
5891 if (mode == VOIDmode)
5892 return alpha_arg_info_reg_val (cum);
5894 num_args = cum.num_args;
5895 if (num_args >= 6 || MUST_PASS_IN_STACK (mode, type))
5898 #elif TARGET_ABI_UNICOSMK
5902 /* If this is the last argument, generate the call info word (CIW). */
5903 /* ??? We don't include the caller's line number in the CIW because
5904 I don't know how to determine it if debug infos are turned off. */
5905 if (mode == VOIDmode)
5914 for (i = 0; i < cum.num_reg_words && i < 5; i++)
5915 if (cum.reg_args_type[i])
5916 lo |= (1 << (7 - i));
5918 if (cum.num_reg_words == 6 && cum.reg_args_type[5])
5921 lo |= cum.num_reg_words;
5923 #if HOST_BITS_PER_WIDE_INT == 32
5924 hi = (cum.num_args << 20) | cum.num_arg_words;
5926 lo = lo | ((HOST_WIDE_INT) cum.num_args << 52)
5927 | ((HOST_WIDE_INT) cum.num_arg_words << 32);
5930 ciw = immed_double_const (lo, hi, DImode);
5932 return gen_rtx_UNSPEC (DImode, gen_rtvec (1, ciw),
5933 UNSPEC_UMK_LOAD_CIW);
5936 size = ALPHA_ARG_SIZE (mode, type, named);
5937 num_args = cum.num_reg_words;
5938 if (MUST_PASS_IN_STACK (mode, type)
5939 || cum.num_reg_words + size > 6 || cum.force_stack)
5941 else if (type && TYPE_MODE (type) == BLKmode)
5945 reg1 = gen_rtx_REG (DImode, num_args + 16);
5946 reg1 = gen_rtx_EXPR_LIST (DImode, reg1, const0_rtx);
5948 /* The argument fits in two registers. Note that we still need to
5949 reserve a register for empty structures. */
5953 return gen_rtx_PARALLEL (mode, gen_rtvec (1, reg1));
5956 reg2 = gen_rtx_REG (DImode, num_args + 17);
5957 reg2 = gen_rtx_EXPR_LIST (DImode, reg2, GEN_INT (8));
5958 return gen_rtx_PARALLEL (mode, gen_rtvec (2, reg1, reg2));
5962 #elif TARGET_ABI_OSF
5968 /* VOID is passed as a special flag for "last argument". */
5969 if (type == void_type_node)
5971 else if (MUST_PASS_IN_STACK (mode, type))
5973 else if (FUNCTION_ARG_PASS_BY_REFERENCE (cum, mode, type, named))
5977 #error Unhandled ABI
5980 return gen_rtx_REG (mode, num_args + basereg);
5983 /* Return true if TYPE must be returned in memory, instead of in registers. */
5986 return_in_memory (tree type, enum machine_mode mode)
5992 mode = TYPE_MODE (type);
5994 /* All aggregates are returned in memory. */
5995 if (AGGREGATE_TYPE_P (type))
5999 size = GET_MODE_SIZE (mode);
6000 switch (GET_MODE_CLASS (mode))
6002 case MODE_VECTOR_FLOAT:
6003 /* Pass all float vectors in memory, like an aggregate. */
6006 case MODE_COMPLEX_FLOAT:
6007 /* We judge complex floats on the size of their element,
6008 not the size of the whole type. */
6009 size = GET_MODE_UNIT_SIZE (mode);
6014 case MODE_COMPLEX_INT:
6015 case MODE_VECTOR_INT:
6019 /* ??? We get called on all sorts of random stuff from
6020 aggregate_value_p. We can't abort, but it's not clear
6021 what's safe to return. Pretend it's a struct I guess. */
6025 /* Otherwise types must fit in one register. */
6026 return size > UNITS_PER_WORD;
6029 /* Define how to find the value returned by a function. VALTYPE is the
6030 data type of the value (as a tree). If the precise function being
6031 called is known, FUNC is its FUNCTION_DECL; otherwise, FUNC is 0.
6032 MODE is set instead of VALTYPE for libcalls.
6034 On Alpha the value is found in $0 for integer functions and
6035 $f0 for floating-point functions. */
6038 function_value (tree valtype, tree func ATTRIBUTE_UNUSED,
6039 enum machine_mode mode)
6041 unsigned int regnum;
6042 enum mode_class class;
6044 #ifdef ENABLE_CHECKING
6045 if (return_in_memory (valtype, mode))
6050 mode = TYPE_MODE (valtype);
6052 class = GET_MODE_CLASS (mode);
6056 /* Do the same thing as PROMOTE_MODE. */
6060 case MODE_COMPLEX_INT:
6061 case MODE_VECTOR_INT:
6069 case MODE_COMPLEX_FLOAT:
6071 enum machine_mode cmode = GET_MODE_INNER (mode);
6073 return gen_rtx_PARALLEL
6076 gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (cmode, 32),
6078 gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (cmode, 33),
6079 GEN_INT (GET_MODE_SIZE (cmode)))));
6086 return gen_rtx_REG (mode, regnum);
6090 alpha_build_va_list (void)
6092 tree base, ofs, record, type_decl;
6094 if (TARGET_ABI_OPEN_VMS || TARGET_ABI_UNICOSMK)
6095 return ptr_type_node;
6097 record = (*lang_hooks.types.make_type) (RECORD_TYPE);
6098 type_decl = build_decl (TYPE_DECL, get_identifier ("__va_list_tag"), record);
6099 TREE_CHAIN (record) = type_decl;
6100 TYPE_NAME (record) = type_decl;
6102 /* C++? SET_IS_AGGR_TYPE (record, 1); */
6104 ofs = build_decl (FIELD_DECL, get_identifier ("__offset"),
6106 DECL_FIELD_CONTEXT (ofs) = record;
6108 base = build_decl (FIELD_DECL, get_identifier ("__base"),
6110 DECL_FIELD_CONTEXT (base) = record;
6111 TREE_CHAIN (base) = ofs;
6113 TYPE_FIELDS (record) = base;
6114 layout_type (record);
6119 /* Perform any needed actions needed for a function that is receiving a
6120 variable number of arguments.
6122 On the Alpha, we allocate space for all 12 arg registers, but only
6123 push those that are remaining. However, if NO registers need to be
6124 saved, don't allocate any space. This is not only because we won't
6125 need the space, but because AP includes the current_pretend_args_size
6126 and we don't want to mess up any ap-relative addresses already made.
6128 If we are not to use the floating-point registers, save the integer
6129 registers where we would put the floating-point registers. This is
6130 not the most efficient way to implement varargs with just one register
6131 class, but it isn't worth doing anything more efficient in this rare
6136 alpha_setup_incoming_varargs(CUMULATIVE_ARGS cum,
6137 enum machine_mode mode ATTRIBUTE_UNUSED,
6138 tree type ATTRIBUTE_UNUSED,
6139 int *pretend_size, int no_rtl)
6146 int set = get_varargs_alias_set ();
6149 tmp = gen_rtx_MEM (BLKmode,
6150 plus_constant (virtual_incoming_args_rtx,
6151 (cum + 6) * UNITS_PER_WORD));
6152 set_mem_alias_set (tmp, set);
6153 move_block_from_reg (16 + cum, tmp, 6 - cum);
6155 tmp = gen_rtx_MEM (BLKmode,
6156 plus_constant (virtual_incoming_args_rtx,
6157 cum * UNITS_PER_WORD));
6158 set_mem_alias_set (tmp, set);
6159 move_block_from_reg (16 + (TARGET_FPREGS ? 32 : 0) + cum, tmp,
6162 *pretend_size = 12 * UNITS_PER_WORD;
6167 alpha_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED)
6169 HOST_WIDE_INT offset;
6170 tree t, offset_field, base_field;
6172 if (TREE_CODE (TREE_TYPE (valist)) == ERROR_MARK)
6175 if (TARGET_ABI_UNICOSMK)
6176 std_expand_builtin_va_start (valist, nextarg);
6178 /* For Unix, SETUP_INCOMING_VARARGS moves the starting address base
6179 up by 48, storing fp arg registers in the first 48 bytes, and the
6180 integer arg registers in the next 48 bytes. This is only done,
6181 however, if any integer registers need to be stored.
6183 If no integer registers need be stored, then we must subtract 48
6184 in order to account for the integer arg registers which are counted
6185 in argsize above, but which are not actually stored on the stack.
6186 Must further be careful here about structures straddling the last
6187 integer argument register; that futzes with pretend_args_size,
6188 which changes the meaning of AP. */
6191 offset = TARGET_ABI_OPEN_VMS ? UNITS_PER_WORD : 6 * UNITS_PER_WORD;
6193 offset = -6 * UNITS_PER_WORD + current_function_pretend_args_size;
6195 if (TARGET_ABI_OPEN_VMS)
6197 nextarg = plus_constant (nextarg, offset);
6198 nextarg = plus_constant (nextarg, NUM_ARGS * UNITS_PER_WORD);
6199 t = build (MODIFY_EXPR, TREE_TYPE (valist), valist,
6200 make_tree (ptr_type_node, nextarg));
6201 TREE_SIDE_EFFECTS (t) = 1;
6203 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6207 base_field = TYPE_FIELDS (TREE_TYPE (valist));
6208 offset_field = TREE_CHAIN (base_field);
6210 base_field = build (COMPONENT_REF, TREE_TYPE (base_field),
6211 valist, base_field);
6212 offset_field = build (COMPONENT_REF, TREE_TYPE (offset_field),
6213 valist, offset_field);
6215 t = make_tree (ptr_type_node, virtual_incoming_args_rtx);
6216 t = build (PLUS_EXPR, ptr_type_node, t, build_int_2 (offset, 0));
6217 t = build (MODIFY_EXPR, TREE_TYPE (base_field), base_field, t);
6218 TREE_SIDE_EFFECTS (t) = 1;
6219 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6221 t = build_int_2 (NUM_ARGS * UNITS_PER_WORD, 0);
6222 t = build (MODIFY_EXPR, TREE_TYPE (offset_field), offset_field, t);
6223 TREE_SIDE_EFFECTS (t) = 1;
6224 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6229 alpha_va_arg (tree valist, tree type)
6232 tree t, type_size, rounded_size;
6233 tree offset_field, base_field, addr_tree, addend;
6234 tree wide_type, wide_ofs;
6237 if (TARGET_ABI_OPEN_VMS || TARGET_ABI_UNICOSMK)
6238 return std_expand_builtin_va_arg (valist, type);
6240 if (type == error_mark_node
6241 || (type_size = TYPE_SIZE_UNIT (TYPE_MAIN_VARIANT (type))) == NULL
6242 || TREE_OVERFLOW (type_size))
6243 rounded_size = size_zero_node;
6245 rounded_size = fold (build (MULT_EXPR, sizetype,
6246 fold (build (TRUNC_DIV_EXPR, sizetype,
6247 fold (build (PLUS_EXPR, sizetype,
6253 base_field = TYPE_FIELDS (TREE_TYPE (valist));
6254 offset_field = TREE_CHAIN (base_field);
6256 base_field = build (COMPONENT_REF, TREE_TYPE (base_field),
6257 valist, base_field);
6258 offset_field = build (COMPONENT_REF, TREE_TYPE (offset_field),
6259 valist, offset_field);
6261 /* If the type could not be passed in registers, skip the block
6262 reserved for the registers. */
6263 if (MUST_PASS_IN_STACK (TYPE_MODE (type), type))
6265 t = build (MODIFY_EXPR, TREE_TYPE (offset_field), offset_field,
6266 build (MAX_EXPR, TREE_TYPE (offset_field),
6267 offset_field, build_int_2 (6*8, 0)));
6268 TREE_SIDE_EFFECTS (t) = 1;
6269 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6272 wide_type = make_signed_type (64);
6273 wide_ofs = save_expr (build1 (CONVERT_EXPR, wide_type, offset_field));
6277 if (TYPE_MODE (type) == TFmode || TYPE_MODE (type) == TCmode)
6280 rounded_size = size_int (UNITS_PER_WORD);
6282 else if (TREE_CODE (type) == COMPLEX_TYPE)
6284 rtx real_part, imag_part, value, tmp;
6286 real_part = alpha_va_arg (valist, TREE_TYPE (type));
6287 imag_part = alpha_va_arg (valist, TREE_TYPE (type));
6289 /* ??? Most irritatingly, we're not returning the value here,
6290 but the address. Since real_part and imag_part are not
6291 necessarily contiguous, we must copy to local storage. */
6293 real_part = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (type)), real_part);
6294 imag_part = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (type)), imag_part);
6295 value = gen_rtx_CONCAT (TYPE_MODE (type), real_part, imag_part);
6297 tmp = assign_temp (type, 0, 1, 0);
6298 emit_move_insn (tmp, value);
6300 return XEXP (tmp, 0);
6302 else if (TREE_CODE (type) == REAL_TYPE)
6304 tree fpaddend, cond;
6306 fpaddend = fold (build (PLUS_EXPR, TREE_TYPE (addend),
6307 addend, build_int_2 (-6*8, 0)));
6309 cond = fold (build (LT_EXPR, integer_type_node,
6310 wide_ofs, build_int_2 (6*8, 0)));
6312 addend = fold (build (COND_EXPR, TREE_TYPE (addend), cond,
6316 addr_tree = build (PLUS_EXPR, TREE_TYPE (base_field),
6317 base_field, addend);
6319 addr = expand_expr (addr_tree, NULL_RTX, Pmode, EXPAND_NORMAL);
6320 addr = copy_to_reg (addr);
6322 t = build (MODIFY_EXPR, TREE_TYPE (offset_field), offset_field,
6323 build (PLUS_EXPR, TREE_TYPE (offset_field),
6324 offset_field, rounded_size));
6325 TREE_SIDE_EFFECTS (t) = 1;
6326 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6330 addr = force_reg (Pmode, addr);
6331 addr = gen_rtx_MEM (Pmode, addr);
6341 ALPHA_BUILTIN_CMPBGE,
6342 ALPHA_BUILTIN_EXTBL,
6343 ALPHA_BUILTIN_EXTWL,
6344 ALPHA_BUILTIN_EXTLL,
6345 ALPHA_BUILTIN_EXTQL,
6346 ALPHA_BUILTIN_EXTWH,
6347 ALPHA_BUILTIN_EXTLH,
6348 ALPHA_BUILTIN_EXTQH,
6349 ALPHA_BUILTIN_INSBL,
6350 ALPHA_BUILTIN_INSWL,
6351 ALPHA_BUILTIN_INSLL,
6352 ALPHA_BUILTIN_INSQL,
6353 ALPHA_BUILTIN_INSWH,
6354 ALPHA_BUILTIN_INSLH,
6355 ALPHA_BUILTIN_INSQH,
6356 ALPHA_BUILTIN_MSKBL,
6357 ALPHA_BUILTIN_MSKWL,
6358 ALPHA_BUILTIN_MSKLL,
6359 ALPHA_BUILTIN_MSKQL,
6360 ALPHA_BUILTIN_MSKWH,
6361 ALPHA_BUILTIN_MSKLH,
6362 ALPHA_BUILTIN_MSKQH,
6363 ALPHA_BUILTIN_UMULH,
6365 ALPHA_BUILTIN_ZAPNOT,
6366 ALPHA_BUILTIN_AMASK,
6367 ALPHA_BUILTIN_IMPLVER,
6369 ALPHA_BUILTIN_THREAD_POINTER,
6370 ALPHA_BUILTIN_SET_THREAD_POINTER,
6373 ALPHA_BUILTIN_MINUB8,
6374 ALPHA_BUILTIN_MINSB8,
6375 ALPHA_BUILTIN_MINUW4,
6376 ALPHA_BUILTIN_MINSW4,
6377 ALPHA_BUILTIN_MAXUB8,
6378 ALPHA_BUILTIN_MAXSB8,
6379 ALPHA_BUILTIN_MAXUW4,
6380 ALPHA_BUILTIN_MAXSW4,
6384 ALPHA_BUILTIN_UNPKBL,
6385 ALPHA_BUILTIN_UNPKBW,
6390 ALPHA_BUILTIN_CTPOP,
6395 static unsigned int const code_for_builtin[ALPHA_BUILTIN_max] = {
6396 CODE_FOR_builtin_cmpbge,
6397 CODE_FOR_builtin_extbl,
6398 CODE_FOR_builtin_extwl,
6399 CODE_FOR_builtin_extll,
6400 CODE_FOR_builtin_extql,
6401 CODE_FOR_builtin_extwh,
6402 CODE_FOR_builtin_extlh,
6403 CODE_FOR_builtin_extqh,
6404 CODE_FOR_builtin_insbl,
6405 CODE_FOR_builtin_inswl,
6406 CODE_FOR_builtin_insll,
6407 CODE_FOR_builtin_insql,
6408 CODE_FOR_builtin_inswh,
6409 CODE_FOR_builtin_inslh,
6410 CODE_FOR_builtin_insqh,
6411 CODE_FOR_builtin_mskbl,
6412 CODE_FOR_builtin_mskwl,
6413 CODE_FOR_builtin_mskll,
6414 CODE_FOR_builtin_mskql,
6415 CODE_FOR_builtin_mskwh,
6416 CODE_FOR_builtin_msklh,
6417 CODE_FOR_builtin_mskqh,
6418 CODE_FOR_umuldi3_highpart,
6419 CODE_FOR_builtin_zap,
6420 CODE_FOR_builtin_zapnot,
6421 CODE_FOR_builtin_amask,
6422 CODE_FOR_builtin_implver,
6423 CODE_FOR_builtin_rpcc,
6428 CODE_FOR_builtin_minub8,
6429 CODE_FOR_builtin_minsb8,
6430 CODE_FOR_builtin_minuw4,
6431 CODE_FOR_builtin_minsw4,
6432 CODE_FOR_builtin_maxub8,
6433 CODE_FOR_builtin_maxsb8,
6434 CODE_FOR_builtin_maxuw4,
6435 CODE_FOR_builtin_maxsw4,
6436 CODE_FOR_builtin_perr,
6437 CODE_FOR_builtin_pklb,
6438 CODE_FOR_builtin_pkwb,
6439 CODE_FOR_builtin_unpkbl,
6440 CODE_FOR_builtin_unpkbw,
6443 CODE_FOR_builtin_cttz,
6444 CODE_FOR_builtin_ctlz,
6445 CODE_FOR_builtin_ctpop
6448 struct alpha_builtin_def
6451 enum alpha_builtin code;
6452 unsigned int target_mask;
6455 static struct alpha_builtin_def const zero_arg_builtins[] = {
6456 { "__builtin_alpha_implver", ALPHA_BUILTIN_IMPLVER, 0 },
6457 { "__builtin_alpha_rpcc", ALPHA_BUILTIN_RPCC, 0 }
6460 static struct alpha_builtin_def const one_arg_builtins[] = {
6461 { "__builtin_alpha_amask", ALPHA_BUILTIN_AMASK, 0 },
6462 { "__builtin_alpha_pklb", ALPHA_BUILTIN_PKLB, MASK_MAX },
6463 { "__builtin_alpha_pkwb", ALPHA_BUILTIN_PKWB, MASK_MAX },
6464 { "__builtin_alpha_unpkbl", ALPHA_BUILTIN_UNPKBL, MASK_MAX },
6465 { "__builtin_alpha_unpkbw", ALPHA_BUILTIN_UNPKBW, MASK_MAX },
6466 { "__builtin_alpha_cttz", ALPHA_BUILTIN_CTTZ, MASK_CIX },
6467 { "__builtin_alpha_ctlz", ALPHA_BUILTIN_CTLZ, MASK_CIX },
6468 { "__builtin_alpha_ctpop", ALPHA_BUILTIN_CTPOP, MASK_CIX }
6471 static struct alpha_builtin_def const two_arg_builtins[] = {
6472 { "__builtin_alpha_cmpbge", ALPHA_BUILTIN_CMPBGE, 0 },
6473 { "__builtin_alpha_extbl", ALPHA_BUILTIN_EXTBL, 0 },
6474 { "__builtin_alpha_extwl", ALPHA_BUILTIN_EXTWL, 0 },
6475 { "__builtin_alpha_extll", ALPHA_BUILTIN_EXTLL, 0 },
6476 { "__builtin_alpha_extql", ALPHA_BUILTIN_EXTQL, 0 },
6477 { "__builtin_alpha_extwh", ALPHA_BUILTIN_EXTWH, 0 },
6478 { "__builtin_alpha_extlh", ALPHA_BUILTIN_EXTLH, 0 },
6479 { "__builtin_alpha_extqh", ALPHA_BUILTIN_EXTQH, 0 },
6480 { "__builtin_alpha_insbl", ALPHA_BUILTIN_INSBL, 0 },
6481 { "__builtin_alpha_inswl", ALPHA_BUILTIN_INSWL, 0 },
6482 { "__builtin_alpha_insll", ALPHA_BUILTIN_INSLL, 0 },
6483 { "__builtin_alpha_insql", ALPHA_BUILTIN_INSQL, 0 },
6484 { "__builtin_alpha_inswh", ALPHA_BUILTIN_INSWH, 0 },
6485 { "__builtin_alpha_inslh", ALPHA_BUILTIN_INSLH, 0 },
6486 { "__builtin_alpha_insqh", ALPHA_BUILTIN_INSQH, 0 },
6487 { "__builtin_alpha_mskbl", ALPHA_BUILTIN_MSKBL, 0 },
6488 { "__builtin_alpha_mskwl", ALPHA_BUILTIN_MSKWL, 0 },
6489 { "__builtin_alpha_mskll", ALPHA_BUILTIN_MSKLL, 0 },
6490 { "__builtin_alpha_mskql", ALPHA_BUILTIN_MSKQL, 0 },
6491 { "__builtin_alpha_mskwh", ALPHA_BUILTIN_MSKWH, 0 },
6492 { "__builtin_alpha_msklh", ALPHA_BUILTIN_MSKLH, 0 },
6493 { "__builtin_alpha_mskqh", ALPHA_BUILTIN_MSKQH, 0 },
6494 { "__builtin_alpha_umulh", ALPHA_BUILTIN_UMULH, 0 },
6495 { "__builtin_alpha_zap", ALPHA_BUILTIN_ZAP, 0 },
6496 { "__builtin_alpha_zapnot", ALPHA_BUILTIN_ZAPNOT, 0 },
6497 { "__builtin_alpha_minub8", ALPHA_BUILTIN_MINUB8, MASK_MAX },
6498 { "__builtin_alpha_minsb8", ALPHA_BUILTIN_MINSB8, MASK_MAX },
6499 { "__builtin_alpha_minuw4", ALPHA_BUILTIN_MINUW4, MASK_MAX },
6500 { "__builtin_alpha_minsw4", ALPHA_BUILTIN_MINSW4, MASK_MAX },
6501 { "__builtin_alpha_maxub8", ALPHA_BUILTIN_MAXUB8, MASK_MAX },
6502 { "__builtin_alpha_maxsb8", ALPHA_BUILTIN_MAXSB8, MASK_MAX },
6503 { "__builtin_alpha_maxuw4", ALPHA_BUILTIN_MAXUW4, MASK_MAX },
6504 { "__builtin_alpha_maxsw4", ALPHA_BUILTIN_MAXSW4, MASK_MAX },
6505 { "__builtin_alpha_perr", ALPHA_BUILTIN_PERR, MASK_MAX }
6509 alpha_init_builtins (void)
6511 const struct alpha_builtin_def *p;
6515 ftype = build_function_type (long_integer_type_node, void_list_node);
6517 p = zero_arg_builtins;
6518 for (i = 0; i < ARRAY_SIZE (zero_arg_builtins); ++i, ++p)
6519 if ((target_flags & p->target_mask) == p->target_mask)
6520 builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
6523 ftype = build_function_type_list (long_integer_type_node,
6524 long_integer_type_node, NULL_TREE);
6526 p = one_arg_builtins;
6527 for (i = 0; i < ARRAY_SIZE (one_arg_builtins); ++i, ++p)
6528 if ((target_flags & p->target_mask) == p->target_mask)
6529 builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
6532 ftype = build_function_type_list (long_integer_type_node,
6533 long_integer_type_node,
6534 long_integer_type_node, NULL_TREE);
6536 p = two_arg_builtins;
6537 for (i = 0; i < ARRAY_SIZE (two_arg_builtins); ++i, ++p)
6538 if ((target_flags & p->target_mask) == p->target_mask)
6539 builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
6542 ftype = build_function_type (ptr_type_node, void_list_node);
6543 builtin_function ("__builtin_thread_pointer", ftype,
6544 ALPHA_BUILTIN_THREAD_POINTER, BUILT_IN_MD,
6547 ftype = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
6548 builtin_function ("__builtin_set_thread_pointer", ftype,
6549 ALPHA_BUILTIN_SET_THREAD_POINTER, BUILT_IN_MD,
6553 /* Expand an expression EXP that calls a built-in function,
6554 with result going to TARGET if that's convenient
6555 (and in mode MODE if that's convenient).
6556 SUBTARGET may be used as the target for computing one of EXP's operands.
6557 IGNORE is nonzero if the value is to be ignored. */
6560 alpha_expand_builtin (tree exp, rtx target,
6561 rtx subtarget ATTRIBUTE_UNUSED,
6562 enum machine_mode mode ATTRIBUTE_UNUSED,
6563 int ignore ATTRIBUTE_UNUSED)
6567 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
6568 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
6569 tree arglist = TREE_OPERAND (exp, 1);
6570 enum insn_code icode;
6571 rtx op[MAX_ARGS], pat;
6575 if (fcode >= ALPHA_BUILTIN_max)
6576 internal_error ("bad builtin fcode");
6577 icode = code_for_builtin[fcode];
6579 internal_error ("bad builtin fcode");
6581 nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
6583 for (arglist = TREE_OPERAND (exp, 1), arity = 0;
6585 arglist = TREE_CHAIN (arglist), arity++)
6587 const struct insn_operand_data *insn_op;
6589 tree arg = TREE_VALUE (arglist);
6590 if (arg == error_mark_node)
6592 if (arity > MAX_ARGS)
6595 insn_op = &insn_data[icode].operand[arity + nonvoid];
6597 op[arity] = expand_expr (arg, NULL_RTX, insn_op->mode, 0);
6599 if (!(*insn_op->predicate) (op[arity], insn_op->mode))
6600 op[arity] = copy_to_mode_reg (insn_op->mode, op[arity]);
6605 enum machine_mode tmode = insn_data[icode].operand[0].mode;
6607 || GET_MODE (target) != tmode
6608 || !(*insn_data[icode].operand[0].predicate) (target, tmode))
6609 target = gen_reg_rtx (tmode);
6615 pat = GEN_FCN (icode) (target);
6619 pat = GEN_FCN (icode) (target, op[0]);
6621 pat = GEN_FCN (icode) (op[0]);
6624 pat = GEN_FCN (icode) (target, op[0], op[1]);
6639 /* This page contains routines that are used to determine what the function
6640 prologue and epilogue code will do and write them out. */
6642 /* Compute the size of the save area in the stack. */
6644 /* These variables are used for communication between the following functions.
6645 They indicate various things about the current function being compiled
6646 that are used to tell what kind of prologue, epilogue and procedure
6647 descriptior to generate. */
6649 /* Nonzero if we need a stack procedure. */
6650 enum alpha_procedure_types {PT_NULL = 0, PT_REGISTER = 1, PT_STACK = 2};
6651 static enum alpha_procedure_types alpha_procedure_type;
6653 /* Register number (either FP or SP) that is used to unwind the frame. */
6654 static int vms_unwind_regno;
6656 /* Register number used to save FP. We need not have one for RA since
6657 we don't modify it for register procedures. This is only defined
6658 for register frame procedures. */
6659 static int vms_save_fp_regno;
6661 /* Register number used to reference objects off our PV. */
6662 static int vms_base_regno;
6664 /* Compute register masks for saved registers. */
6667 alpha_sa_mask (unsigned long *imaskP, unsigned long *fmaskP)
6669 unsigned long imask = 0;
6670 unsigned long fmask = 0;
6673 /* Irritatingly, there are two kinds of thunks -- those created with
6674 TARGET_ASM_OUTPUT_MI_THUNK and those with DECL_THUNK_P that go
6675 through the regular part of the compiler. In the
6676 TARGET_ASM_OUTPUT_MI_THUNK case we don't have valid register life
6677 info, but assemble_start_function wants to output .frame and
6678 .mask directives. */
6679 if (current_function_is_thunk && !no_new_pseudos)
6686 if (TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_STACK)
6687 imask |= (1UL << HARD_FRAME_POINTER_REGNUM);
6689 /* One for every register we have to save. */
6690 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
6691 if (! fixed_regs[i] && ! call_used_regs[i]
6692 && regs_ever_live[i] && i != REG_RA
6693 && (!TARGET_ABI_UNICOSMK || i != HARD_FRAME_POINTER_REGNUM))
6696 imask |= (1UL << i);
6698 fmask |= (1UL << (i - 32));
6701 /* We need to restore these for the handler. */
6702 if (current_function_calls_eh_return)
6705 unsigned regno = EH_RETURN_DATA_REGNO (i);
6706 if (regno == INVALID_REGNUM)
6708 imask |= 1UL << regno;
6711 /* If any register spilled, then spill the return address also. */
6712 /* ??? This is required by the Digital stack unwind specification
6713 and isn't needed if we're doing Dwarf2 unwinding. */
6714 if (imask || fmask || alpha_ra_ever_killed ())
6715 imask |= (1UL << REG_RA);
6722 alpha_sa_size (void)
6724 unsigned long mask[2];
6728 alpha_sa_mask (&mask[0], &mask[1]);
6730 if (TARGET_ABI_UNICOSMK)
6732 if (mask[0] || mask[1])
6737 for (j = 0; j < 2; ++j)
6738 for (i = 0; i < 32; ++i)
6739 if ((mask[j] >> i) & 1)
6743 if (TARGET_ABI_UNICOSMK)
6745 /* We might not need to generate a frame if we don't make any calls
6746 (including calls to __T3E_MISMATCH if this is a vararg function),
6747 don't have any local variables which require stack slots, don't
6748 use alloca and have not determined that we need a frame for other
6751 alpha_procedure_type
6752 = (sa_size || get_frame_size() != 0
6753 || current_function_outgoing_args_size
6754 || current_function_stdarg || current_function_calls_alloca
6755 || frame_pointer_needed)
6756 ? PT_STACK : PT_REGISTER;
6758 /* Always reserve space for saving callee-saved registers if we
6759 need a frame as required by the calling convention. */
6760 if (alpha_procedure_type == PT_STACK)
6763 else if (TARGET_ABI_OPEN_VMS)
6765 /* Start by assuming we can use a register procedure if we don't
6766 make any calls (REG_RA not used) or need to save any
6767 registers and a stack procedure if we do. */
6768 if ((mask[0] >> REG_RA) & 1)
6769 alpha_procedure_type = PT_STACK;
6770 else if (get_frame_size() != 0)
6771 alpha_procedure_type = PT_REGISTER;
6773 alpha_procedure_type = PT_NULL;
6775 /* Don't reserve space for saving FP & RA yet. Do that later after we've
6776 made the final decision on stack procedure vs register procedure. */
6777 if (alpha_procedure_type == PT_STACK)
6780 /* Decide whether to refer to objects off our PV via FP or PV.
6781 If we need FP for something else or if we receive a nonlocal
6782 goto (which expects PV to contain the value), we must use PV.
6783 Otherwise, start by assuming we can use FP. */
6786 = (frame_pointer_needed
6787 || current_function_has_nonlocal_label
6788 || alpha_procedure_type == PT_STACK
6789 || current_function_outgoing_args_size)
6790 ? REG_PV : HARD_FRAME_POINTER_REGNUM;
6792 /* If we want to copy PV into FP, we need to find some register
6793 in which to save FP. */
6795 vms_save_fp_regno = -1;
6796 if (vms_base_regno == HARD_FRAME_POINTER_REGNUM)
6797 for (i = 0; i < 32; i++)
6798 if (! fixed_regs[i] && call_used_regs[i] && ! regs_ever_live[i])
6799 vms_save_fp_regno = i;
6801 if (vms_save_fp_regno == -1 && alpha_procedure_type == PT_REGISTER)
6802 vms_base_regno = REG_PV, alpha_procedure_type = PT_STACK;
6803 else if (alpha_procedure_type == PT_NULL)
6804 vms_base_regno = REG_PV;
6806 /* Stack unwinding should be done via FP unless we use it for PV. */
6807 vms_unwind_regno = (vms_base_regno == REG_PV
6808 ? HARD_FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM);
6810 /* If this is a stack procedure, allow space for saving FP and RA. */
6811 if (alpha_procedure_type == PT_STACK)
6816 /* Our size must be even (multiple of 16 bytes). */
6824 /* Define the offset between two registers, one to be eliminated,
6825 and the other its replacement, at the start of a routine. */
6828 alpha_initial_elimination_offset (unsigned int from,
6829 unsigned int to ATTRIBUTE_UNUSED)
6833 ret = alpha_sa_size ();
6834 ret += ALPHA_ROUND (current_function_outgoing_args_size);
6836 if (from == FRAME_POINTER_REGNUM)
6838 else if (from == ARG_POINTER_REGNUM)
6839 ret += (ALPHA_ROUND (get_frame_size ()
6840 + current_function_pretend_args_size)
6841 - current_function_pretend_args_size);
6849 alpha_pv_save_size (void)
6852 return alpha_procedure_type == PT_STACK ? 8 : 0;
6856 alpha_using_fp (void)
6859 return vms_unwind_regno == HARD_FRAME_POINTER_REGNUM;
6862 #if TARGET_ABI_OPEN_VMS
6864 const struct attribute_spec vms_attribute_table[] =
6866 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
6867 { "overlaid", 0, 0, true, false, false, NULL },
6868 { "global", 0, 0, true, false, false, NULL },
6869 { "initialize", 0, 0, true, false, false, NULL },
6870 { NULL, 0, 0, false, false, false, NULL }
6876 find_lo_sum_using_gp (rtx *px, void *data ATTRIBUTE_UNUSED)
6878 return GET_CODE (*px) == LO_SUM && XEXP (*px, 0) == pic_offset_table_rtx;
6882 alpha_find_lo_sum_using_gp (rtx insn)
6884 return for_each_rtx (&PATTERN (insn), find_lo_sum_using_gp, NULL) > 0;
6888 alpha_does_function_need_gp (void)
6892 /* The GP being variable is an OSF abi thing. */
6893 if (! TARGET_ABI_OSF)
6896 if (TARGET_PROFILING_NEEDS_GP && current_function_profile)
6899 if (current_function_is_thunk)
6902 /* If we need a GP (we have a LDSYM insn or a CALL_INSN), load it first.
6903 Even if we are a static function, we still need to do this in case
6904 our address is taken and passed to something like qsort. */
6906 push_topmost_sequence ();
6907 insn = get_insns ();
6908 pop_topmost_sequence ();
6910 for (; insn; insn = NEXT_INSN (insn))
6912 && GET_CODE (PATTERN (insn)) != USE
6913 && GET_CODE (PATTERN (insn)) != CLOBBER
6914 && get_attr_usegp (insn))
6921 /* Helper function to set RTX_FRAME_RELATED_P on instructions, including
6925 set_frame_related_p (void)
6927 rtx seq = get_insns ();
6938 while (insn != NULL_RTX)
6940 RTX_FRAME_RELATED_P (insn) = 1;
6941 insn = NEXT_INSN (insn);
6943 seq = emit_insn (seq);
6947 seq = emit_insn (seq);
6948 RTX_FRAME_RELATED_P (seq) = 1;
6953 #define FRP(exp) (start_sequence (), exp, set_frame_related_p ())
6955 /* Write function prologue. */
6957 /* On vms we have two kinds of functions:
6959 - stack frame (PROC_STACK)
6960 these are 'normal' functions with local vars and which are
6961 calling other functions
6962 - register frame (PROC_REGISTER)
6963 keeps all data in registers, needs no stack
6965 We must pass this to the assembler so it can generate the
6966 proper pdsc (procedure descriptor)
6967 This is done with the '.pdesc' command.
6969 On not-vms, we don't really differentiate between the two, as we can
6970 simply allocate stack without saving registers. */
6973 alpha_expand_prologue (void)
6975 /* Registers to save. */
6976 unsigned long imask = 0;
6977 unsigned long fmask = 0;
6978 /* Stack space needed for pushing registers clobbered by us. */
6979 HOST_WIDE_INT sa_size;
6980 /* Complete stack size needed. */
6981 HOST_WIDE_INT frame_size;
6982 /* Offset from base reg to register save area. */
6983 HOST_WIDE_INT reg_offset;
6987 sa_size = alpha_sa_size ();
6989 frame_size = get_frame_size ();
6990 if (TARGET_ABI_OPEN_VMS)
6991 frame_size = ALPHA_ROUND (sa_size
6992 + (alpha_procedure_type == PT_STACK ? 8 : 0)
6994 + current_function_pretend_args_size);
6995 else if (TARGET_ABI_UNICOSMK)
6996 /* We have to allocate space for the DSIB if we generate a frame. */
6997 frame_size = ALPHA_ROUND (sa_size
6998 + (alpha_procedure_type == PT_STACK ? 48 : 0))
6999 + ALPHA_ROUND (frame_size
7000 + current_function_outgoing_args_size);
7002 frame_size = (ALPHA_ROUND (current_function_outgoing_args_size)
7004 + ALPHA_ROUND (frame_size
7005 + current_function_pretend_args_size));
7007 if (TARGET_ABI_OPEN_VMS)
7010 reg_offset = ALPHA_ROUND (current_function_outgoing_args_size);
7012 alpha_sa_mask (&imask, &fmask);
7014 /* Emit an insn to reload GP, if needed. */
7017 alpha_function_needs_gp = alpha_does_function_need_gp ();
7018 if (alpha_function_needs_gp)
7019 emit_insn (gen_prologue_ldgp ());
7022 /* TARGET_PROFILING_NEEDS_GP actually implies that we need to insert
7023 the call to mcount ourselves, rather than having the linker do it
7024 magically in response to -pg. Since _mcount has special linkage,
7025 don't represent the call as a call. */
7026 if (TARGET_PROFILING_NEEDS_GP && current_function_profile)
7027 emit_insn (gen_prologue_mcount ());
7029 if (TARGET_ABI_UNICOSMK)
7030 unicosmk_gen_dsib (&imask);
7032 /* Adjust the stack by the frame size. If the frame size is > 4096
7033 bytes, we need to be sure we probe somewhere in the first and last
7034 4096 bytes (we can probably get away without the latter test) and
7035 every 8192 bytes in between. If the frame size is > 32768, we
7036 do this in a loop. Otherwise, we generate the explicit probe
7039 Note that we are only allowed to adjust sp once in the prologue. */
7041 if (frame_size <= 32768)
7043 if (frame_size > 4096)
7048 emit_insn (gen_probe_stack (GEN_INT (TARGET_ABI_UNICOSMK
7051 while ((probed += 8192) < frame_size);
7053 /* We only have to do this probe if we aren't saving registers. */
7054 if (sa_size == 0 && probed + 4096 < frame_size)
7055 emit_insn (gen_probe_stack (GEN_INT (-frame_size)));
7058 if (frame_size != 0)
7059 FRP (emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx,
7060 GEN_INT (TARGET_ABI_UNICOSMK
7066 /* Here we generate code to set R22 to SP + 4096 and set R23 to the
7067 number of 8192 byte blocks to probe. We then probe each block
7068 in the loop and then set SP to the proper location. If the
7069 amount remaining is > 4096, we have to do one more probe if we
7070 are not saving any registers. */
7072 HOST_WIDE_INT blocks = (frame_size + 4096) / 8192;
7073 HOST_WIDE_INT leftover = frame_size + 4096 - blocks * 8192;
7074 rtx ptr = gen_rtx_REG (DImode, 22);
7075 rtx count = gen_rtx_REG (DImode, 23);
7078 emit_move_insn (count, GEN_INT (blocks));
7079 emit_insn (gen_adddi3 (ptr, stack_pointer_rtx,
7080 GEN_INT (TARGET_ABI_UNICOSMK ? 4096 - 64 : 4096)));
7082 /* Because of the difficulty in emitting a new basic block this
7083 late in the compilation, generate the loop as a single insn. */
7084 emit_insn (gen_prologue_stack_probe_loop (count, ptr));
7086 if (leftover > 4096 && sa_size == 0)
7088 rtx last = gen_rtx_MEM (DImode, plus_constant (ptr, -leftover));
7089 MEM_VOLATILE_P (last) = 1;
7090 emit_move_insn (last, const0_rtx);
7093 if (TARGET_ABI_WINDOWS_NT)
7095 /* For NT stack unwind (done by 'reverse execution'), it's
7096 not OK to take the result of a loop, even though the value
7097 is already in ptr, so we reload it via a single operation
7098 and subtract it to sp.
7100 Yes, that's correct -- we have to reload the whole constant
7101 into a temporary via ldah+lda then subtract from sp. To
7102 ensure we get ldah+lda, we use a special pattern. */
7104 HOST_WIDE_INT lo, hi;
7105 lo = ((frame_size & 0xffff) ^ 0x8000) - 0x8000;
7106 hi = frame_size - lo;
7108 emit_move_insn (ptr, GEN_INT (hi));
7109 emit_insn (gen_nt_lda (ptr, GEN_INT (lo)));
7110 seq = emit_insn (gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx,
7115 seq = emit_insn (gen_adddi3 (stack_pointer_rtx, ptr,
7116 GEN_INT (-leftover)));
7119 /* This alternative is special, because the DWARF code cannot
7120 possibly intuit through the loop above. So we invent this
7121 note it looks at instead. */
7122 RTX_FRAME_RELATED_P (seq) = 1;
7124 = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
7125 gen_rtx_SET (VOIDmode, stack_pointer_rtx,
7126 gen_rtx_PLUS (Pmode, stack_pointer_rtx,
7127 GEN_INT (TARGET_ABI_UNICOSMK
7133 if (!TARGET_ABI_UNICOSMK)
7135 /* Cope with very large offsets to the register save area. */
7136 sa_reg = stack_pointer_rtx;
7137 if (reg_offset + sa_size > 0x8000)
7139 int low = ((reg_offset & 0xffff) ^ 0x8000) - 0x8000;
7142 if (low + sa_size <= 0x8000)
7143 bias = reg_offset - low, reg_offset = low;
7145 bias = reg_offset, reg_offset = 0;
7147 sa_reg = gen_rtx_REG (DImode, 24);
7148 FRP (emit_insn (gen_adddi3 (sa_reg, stack_pointer_rtx,
7152 /* Save regs in stack order. Beginning with VMS PV. */
7153 if (TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_STACK)
7155 mem = gen_rtx_MEM (DImode, stack_pointer_rtx);
7156 set_mem_alias_set (mem, alpha_sr_alias_set);
7157 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, REG_PV)));
7160 /* Save register RA next. */
7161 if (imask & (1UL << REG_RA))
7163 mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
7164 set_mem_alias_set (mem, alpha_sr_alias_set);
7165 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, REG_RA)));
7166 imask &= ~(1UL << REG_RA);
7170 /* Now save any other registers required to be saved. */
7171 for (i = 0; i < 32; i++)
7172 if (imask & (1UL << i))
7174 mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
7175 set_mem_alias_set (mem, alpha_sr_alias_set);
7176 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, i)));
7180 for (i = 0; i < 32; i++)
7181 if (fmask & (1UL << i))
7183 mem = gen_rtx_MEM (DFmode, plus_constant (sa_reg, reg_offset));
7184 set_mem_alias_set (mem, alpha_sr_alias_set);
7185 FRP (emit_move_insn (mem, gen_rtx_REG (DFmode, i+32)));
7189 else if (TARGET_ABI_UNICOSMK && alpha_procedure_type == PT_STACK)
7191 /* The standard frame on the T3E includes space for saving registers.
7192 We just have to use it. We don't have to save the return address and
7193 the old frame pointer here - they are saved in the DSIB. */
7196 for (i = 9; i < 15; i++)
7197 if (imask & (1UL << i))
7199 mem = gen_rtx_MEM (DImode, plus_constant(hard_frame_pointer_rtx,
7201 set_mem_alias_set (mem, alpha_sr_alias_set);
7202 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, i)));
7205 for (i = 2; i < 10; i++)
7206 if (fmask & (1UL << i))
7208 mem = gen_rtx_MEM (DFmode, plus_constant (hard_frame_pointer_rtx,
7210 set_mem_alias_set (mem, alpha_sr_alias_set);
7211 FRP (emit_move_insn (mem, gen_rtx_REG (DFmode, i+32)));
7216 if (TARGET_ABI_OPEN_VMS)
7218 if (alpha_procedure_type == PT_REGISTER)
7219 /* Register frame procedures save the fp.
7220 ?? Ought to have a dwarf2 save for this. */
7221 emit_move_insn (gen_rtx_REG (DImode, vms_save_fp_regno),
7222 hard_frame_pointer_rtx);
7224 if (alpha_procedure_type != PT_NULL && vms_base_regno != REG_PV)
7225 emit_insn (gen_force_movdi (gen_rtx_REG (DImode, vms_base_regno),
7226 gen_rtx_REG (DImode, REG_PV)));
7228 if (alpha_procedure_type != PT_NULL
7229 && vms_unwind_regno == HARD_FRAME_POINTER_REGNUM)
7230 FRP (emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx));
7232 /* If we have to allocate space for outgoing args, do it now. */
7233 if (current_function_outgoing_args_size != 0)
7236 = emit_move_insn (stack_pointer_rtx,
7238 (hard_frame_pointer_rtx,
7240 (current_function_outgoing_args_size))));
7242 /* Only set FRAME_RELATED_P on the stack adjustment we just emitted
7243 if ! frame_pointer_needed. Setting the bit will change the CFA
7244 computation rule to use sp again, which would be wrong if we had
7245 frame_pointer_needed, as this means sp might move unpredictably
7249 frame_pointer_needed
7250 => vms_unwind_regno == HARD_FRAME_POINTER_REGNUM
7252 current_function_outgoing_args_size != 0
7253 => alpha_procedure_type != PT_NULL,
7255 so when we are not setting the bit here, we are guaranteed to
7256 have emitted an FRP frame pointer update just before. */
7257 RTX_FRAME_RELATED_P (seq) = ! frame_pointer_needed;
7260 else if (!TARGET_ABI_UNICOSMK)
7262 /* If we need a frame pointer, set it from the stack pointer. */
7263 if (frame_pointer_needed)
7265 if (TARGET_CAN_FAULT_IN_PROLOGUE)
7266 FRP (emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx));
7268 /* This must always be the last instruction in the
7269 prologue, thus we emit a special move + clobber. */
7270 FRP (emit_insn (gen_init_fp (hard_frame_pointer_rtx,
7271 stack_pointer_rtx, sa_reg)));
7275 /* The ABIs for VMS and OSF/1 say that while we can schedule insns into
7276 the prologue, for exception handling reasons, we cannot do this for
7277 any insn that might fault. We could prevent this for mems with a
7278 (clobber:BLK (scratch)), but this doesn't work for fp insns. So we
7279 have to prevent all such scheduling with a blockage.
7281 Linux, on the other hand, never bothered to implement OSF/1's
7282 exception handling, and so doesn't care about such things. Anyone
7283 planning to use dwarf2 frame-unwind info can also omit the blockage. */
7285 if (! TARGET_CAN_FAULT_IN_PROLOGUE)
7286 emit_insn (gen_blockage ());
7289 /* Output the textual info surrounding the prologue. */
7292 alpha_start_function (FILE *file, const char *fnname,
7293 tree decl ATTRIBUTE_UNUSED)
7295 unsigned long imask = 0;
7296 unsigned long fmask = 0;
7297 /* Stack space needed for pushing registers clobbered by us. */
7298 HOST_WIDE_INT sa_size;
7299 /* Complete stack size needed. */
7300 unsigned HOST_WIDE_INT frame_size;
7301 /* Offset from base reg to register save area. */
7302 HOST_WIDE_INT reg_offset;
7303 char *entry_label = (char *) alloca (strlen (fnname) + 6);
7306 /* Don't emit an extern directive for functions defined in the same file. */
7307 if (TARGET_ABI_UNICOSMK)
7310 name_tree = get_identifier (fnname);
7311 TREE_ASM_WRITTEN (name_tree) = 1;
7314 alpha_fnname = fnname;
7315 sa_size = alpha_sa_size ();
7317 frame_size = get_frame_size ();
7318 if (TARGET_ABI_OPEN_VMS)
7319 frame_size = ALPHA_ROUND (sa_size
7320 + (alpha_procedure_type == PT_STACK ? 8 : 0)
7322 + current_function_pretend_args_size);
7323 else if (TARGET_ABI_UNICOSMK)
7324 frame_size = ALPHA_ROUND (sa_size
7325 + (alpha_procedure_type == PT_STACK ? 48 : 0))
7326 + ALPHA_ROUND (frame_size
7327 + current_function_outgoing_args_size);
7329 frame_size = (ALPHA_ROUND (current_function_outgoing_args_size)
7331 + ALPHA_ROUND (frame_size
7332 + current_function_pretend_args_size));
7334 if (TARGET_ABI_OPEN_VMS)
7337 reg_offset = ALPHA_ROUND (current_function_outgoing_args_size);
7339 alpha_sa_mask (&imask, &fmask);
7341 /* Ecoff can handle multiple .file directives, so put out file and lineno.
7342 We have to do that before the .ent directive as we cannot switch
7343 files within procedures with native ecoff because line numbers are
7344 linked to procedure descriptors.
7345 Outputting the lineno helps debugging of one line functions as they
7346 would otherwise get no line number at all. Please note that we would
7347 like to put out last_linenum from final.c, but it is not accessible. */
7349 if (write_symbols == SDB_DEBUG)
7351 #ifdef ASM_OUTPUT_SOURCE_FILENAME
7352 ASM_OUTPUT_SOURCE_FILENAME (file, TREE_FILENAME (current_function_decl));
7354 #ifdef ASM_OUTPUT_SOURCE_LINE
7355 if (debug_info_level != DINFO_LEVEL_TERSE)
7356 ASM_OUTPUT_SOURCE_LINE (file, TREE_LINENO (current_function_decl), 0);
7360 /* Issue function start and label. */
7361 if (TARGET_ABI_OPEN_VMS
7362 || (!TARGET_ABI_UNICOSMK && !flag_inhibit_size_directive))
7364 fputs ("\t.ent ", file);
7365 assemble_name (file, fnname);
7368 /* If the function needs GP, we'll write the "..ng" label there.
7369 Otherwise, do it here. */
7371 && ! alpha_function_needs_gp
7372 && ! current_function_is_thunk)
7375 assemble_name (file, fnname);
7376 fputs ("..ng:\n", file);
7380 strcpy (entry_label, fnname);
7381 if (TARGET_ABI_OPEN_VMS)
7382 strcat (entry_label, "..en");
7384 /* For public functions, the label must be globalized by appending an
7385 additional colon. */
7386 if (TARGET_ABI_UNICOSMK && TREE_PUBLIC (decl))
7387 strcat (entry_label, ":");
7389 ASM_OUTPUT_LABEL (file, entry_label);
7390 inside_function = TRUE;
7392 if (TARGET_ABI_OPEN_VMS)
7393 fprintf (file, "\t.base $%d\n", vms_base_regno);
7395 if (!TARGET_ABI_OPEN_VMS && !TARGET_ABI_UNICOSMK && TARGET_IEEE_CONFORMANT
7396 && !flag_inhibit_size_directive)
7398 /* Set flags in procedure descriptor to request IEEE-conformant
7399 math-library routines. The value we set it to is PDSC_EXC_IEEE
7400 (/usr/include/pdsc.h). */
7401 fputs ("\t.eflag 48\n", file);
7404 /* Set up offsets to alpha virtual arg/local debugging pointer. */
7405 alpha_auto_offset = -frame_size + current_function_pretend_args_size;
7406 alpha_arg_offset = -frame_size + 48;
7408 /* Describe our frame. If the frame size is larger than an integer,
7409 print it as zero to avoid an assembler error. We won't be
7410 properly describing such a frame, but that's the best we can do. */
7411 if (TARGET_ABI_UNICOSMK)
7413 else if (TARGET_ABI_OPEN_VMS)
7414 fprintf (file, "\t.frame $%d," HOST_WIDE_INT_PRINT_DEC ",$26,"
7415 HOST_WIDE_INT_PRINT_DEC "\n",
7417 frame_size >= (1UL << 31) ? 0 : frame_size,
7419 else if (!flag_inhibit_size_directive)
7420 fprintf (file, "\t.frame $%d," HOST_WIDE_INT_PRINT_DEC ",$26,%d\n",
7421 (frame_pointer_needed
7422 ? HARD_FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM),
7423 frame_size >= (1UL << 31) ? 0 : frame_size,
7424 current_function_pretend_args_size);
7426 /* Describe which registers were spilled. */
7427 if (TARGET_ABI_UNICOSMK)
7429 else if (TARGET_ABI_OPEN_VMS)
7432 /* ??? Does VMS care if mask contains ra? The old code didn't
7433 set it, so I don't here. */
7434 fprintf (file, "\t.mask 0x%lx,0\n", imask & ~(1UL << REG_RA));
7436 fprintf (file, "\t.fmask 0x%lx,0\n", fmask);
7437 if (alpha_procedure_type == PT_REGISTER)
7438 fprintf (file, "\t.fp_save $%d\n", vms_save_fp_regno);
7440 else if (!flag_inhibit_size_directive)
7444 fprintf (file, "\t.mask 0x%lx," HOST_WIDE_INT_PRINT_DEC "\n", imask,
7445 frame_size >= (1UL << 31) ? 0 : reg_offset - frame_size);
7447 for (i = 0; i < 32; ++i)
7448 if (imask & (1UL << i))
7453 fprintf (file, "\t.fmask 0x%lx," HOST_WIDE_INT_PRINT_DEC "\n", fmask,
7454 frame_size >= (1UL << 31) ? 0 : reg_offset - frame_size);
7457 #if TARGET_ABI_OPEN_VMS
7458 /* Ifdef'ed cause link_section are only available then. */
7459 readonly_data_section ();
7460 fprintf (file, "\t.align 3\n");
7461 assemble_name (file, fnname); fputs ("..na:\n", file);
7462 fputs ("\t.ascii \"", file);
7463 assemble_name (file, fnname);
7464 fputs ("\\0\"\n", file);
7465 alpha_need_linkage (fnname, 1);
7470 /* Emit the .prologue note at the scheduled end of the prologue. */
7473 alpha_output_function_end_prologue (FILE *file)
7475 if (TARGET_ABI_UNICOSMK)
7477 else if (TARGET_ABI_OPEN_VMS)
7478 fputs ("\t.prologue\n", file);
7479 else if (TARGET_ABI_WINDOWS_NT)
7480 fputs ("\t.prologue 0\n", file);
7481 else if (!flag_inhibit_size_directive)
7482 fprintf (file, "\t.prologue %d\n",
7483 alpha_function_needs_gp || current_function_is_thunk);
7486 /* Write function epilogue. */
7488 /* ??? At some point we will want to support full unwind, and so will
7489 need to mark the epilogue as well. At the moment, we just confuse
7492 #define FRP(exp) exp
7495 alpha_expand_epilogue (void)
7497 /* Registers to save. */
7498 unsigned long imask = 0;
7499 unsigned long fmask = 0;
7500 /* Stack space needed for pushing registers clobbered by us. */
7501 HOST_WIDE_INT sa_size;
7502 /* Complete stack size needed. */
7503 HOST_WIDE_INT frame_size;
7504 /* Offset from base reg to register save area. */
7505 HOST_WIDE_INT reg_offset;
7506 int fp_is_frame_pointer, fp_offset;
7507 rtx sa_reg, sa_reg_exp = NULL;
7508 rtx sp_adj1, sp_adj2, mem;
7512 sa_size = alpha_sa_size ();
7514 frame_size = get_frame_size ();
7515 if (TARGET_ABI_OPEN_VMS)
7516 frame_size = ALPHA_ROUND (sa_size
7517 + (alpha_procedure_type == PT_STACK ? 8 : 0)
7519 + current_function_pretend_args_size);
7520 else if (TARGET_ABI_UNICOSMK)
7521 frame_size = ALPHA_ROUND (sa_size
7522 + (alpha_procedure_type == PT_STACK ? 48 : 0))
7523 + ALPHA_ROUND (frame_size
7524 + current_function_outgoing_args_size);
7526 frame_size = (ALPHA_ROUND (current_function_outgoing_args_size)
7528 + ALPHA_ROUND (frame_size
7529 + current_function_pretend_args_size));
7531 if (TARGET_ABI_OPEN_VMS)
7533 if (alpha_procedure_type == PT_STACK)
7539 reg_offset = ALPHA_ROUND (current_function_outgoing_args_size);
7541 alpha_sa_mask (&imask, &fmask);
7544 = ((TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_STACK)
7545 || (!TARGET_ABI_OPEN_VMS && frame_pointer_needed));
7547 sa_reg = stack_pointer_rtx;
7549 if (current_function_calls_eh_return)
7550 eh_ofs = EH_RETURN_STACKADJ_RTX;
7554 if (!TARGET_ABI_UNICOSMK && sa_size)
7556 /* If we have a frame pointer, restore SP from it. */
7557 if ((TARGET_ABI_OPEN_VMS
7558 && vms_unwind_regno == HARD_FRAME_POINTER_REGNUM)
7559 || (!TARGET_ABI_OPEN_VMS && frame_pointer_needed))
7560 FRP (emit_move_insn (stack_pointer_rtx, hard_frame_pointer_rtx));
7562 /* Cope with very large offsets to the register save area. */
7563 if (reg_offset + sa_size > 0x8000)
7565 int low = ((reg_offset & 0xffff) ^ 0x8000) - 0x8000;
7568 if (low + sa_size <= 0x8000)
7569 bias = reg_offset - low, reg_offset = low;
7571 bias = reg_offset, reg_offset = 0;
7573 sa_reg = gen_rtx_REG (DImode, 22);
7574 sa_reg_exp = plus_constant (stack_pointer_rtx, bias);
7576 FRP (emit_move_insn (sa_reg, sa_reg_exp));
7579 /* Restore registers in order, excepting a true frame pointer. */
7581 mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
7583 set_mem_alias_set (mem, alpha_sr_alias_set);
7584 FRP (emit_move_insn (gen_rtx_REG (DImode, REG_RA), mem));
7587 imask &= ~(1UL << REG_RA);
7589 for (i = 0; i < 32; ++i)
7590 if (imask & (1UL << i))
7592 if (i == HARD_FRAME_POINTER_REGNUM && fp_is_frame_pointer)
7593 fp_offset = reg_offset;
7596 mem = gen_rtx_MEM (DImode, plus_constant(sa_reg, reg_offset));
7597 set_mem_alias_set (mem, alpha_sr_alias_set);
7598 FRP (emit_move_insn (gen_rtx_REG (DImode, i), mem));
7603 for (i = 0; i < 32; ++i)
7604 if (fmask & (1UL << i))
7606 mem = gen_rtx_MEM (DFmode, plus_constant(sa_reg, reg_offset));
7607 set_mem_alias_set (mem, alpha_sr_alias_set);
7608 FRP (emit_move_insn (gen_rtx_REG (DFmode, i+32), mem));
7612 else if (TARGET_ABI_UNICOSMK && alpha_procedure_type == PT_STACK)
7614 /* Restore callee-saved general-purpose registers. */
7618 for (i = 9; i < 15; i++)
7619 if (imask & (1UL << i))
7621 mem = gen_rtx_MEM (DImode, plus_constant(hard_frame_pointer_rtx,
7623 set_mem_alias_set (mem, alpha_sr_alias_set);
7624 FRP (emit_move_insn (gen_rtx_REG (DImode, i), mem));
7628 for (i = 2; i < 10; i++)
7629 if (fmask & (1UL << i))
7631 mem = gen_rtx_MEM (DFmode, plus_constant(hard_frame_pointer_rtx,
7633 set_mem_alias_set (mem, alpha_sr_alias_set);
7634 FRP (emit_move_insn (gen_rtx_REG (DFmode, i+32), mem));
7638 /* Restore the return address from the DSIB. */
7640 mem = gen_rtx_MEM (DImode, plus_constant(hard_frame_pointer_rtx, -8));
7641 set_mem_alias_set (mem, alpha_sr_alias_set);
7642 FRP (emit_move_insn (gen_rtx_REG (DImode, REG_RA), mem));
7645 if (frame_size || eh_ofs)
7647 sp_adj1 = stack_pointer_rtx;
7651 sp_adj1 = gen_rtx_REG (DImode, 23);
7652 emit_move_insn (sp_adj1,
7653 gen_rtx_PLUS (Pmode, stack_pointer_rtx, eh_ofs));
7656 /* If the stack size is large, begin computation into a temporary
7657 register so as not to interfere with a potential fp restore,
7658 which must be consecutive with an SP restore. */
7659 if (frame_size < 32768
7660 && ! (TARGET_ABI_UNICOSMK && current_function_calls_alloca))
7661 sp_adj2 = GEN_INT (frame_size);
7662 else if (TARGET_ABI_UNICOSMK)
7664 sp_adj1 = gen_rtx_REG (DImode, 23);
7665 FRP (emit_move_insn (sp_adj1, hard_frame_pointer_rtx));
7666 sp_adj2 = const0_rtx;
7668 else if (frame_size < 0x40007fffL)
7670 int low = ((frame_size & 0xffff) ^ 0x8000) - 0x8000;
7672 sp_adj2 = plus_constant (sp_adj1, frame_size - low);
7673 if (sa_reg_exp && rtx_equal_p (sa_reg_exp, sp_adj2))
7677 sp_adj1 = gen_rtx_REG (DImode, 23);
7678 FRP (emit_move_insn (sp_adj1, sp_adj2));
7680 sp_adj2 = GEN_INT (low);
7684 rtx tmp = gen_rtx_REG (DImode, 23);
7685 FRP (sp_adj2 = alpha_emit_set_const (tmp, DImode, frame_size, 3));
7688 /* We can't drop new things to memory this late, afaik,
7689 so build it up by pieces. */
7690 FRP (sp_adj2 = alpha_emit_set_long_const (tmp, frame_size,
7691 -(frame_size < 0)));
7697 /* From now on, things must be in order. So emit blockages. */
7699 /* Restore the frame pointer. */
7700 if (TARGET_ABI_UNICOSMK)
7702 emit_insn (gen_blockage ());
7703 mem = gen_rtx_MEM (DImode,
7704 plus_constant (hard_frame_pointer_rtx, -16));
7705 set_mem_alias_set (mem, alpha_sr_alias_set);
7706 FRP (emit_move_insn (hard_frame_pointer_rtx, mem));
7708 else if (fp_is_frame_pointer)
7710 emit_insn (gen_blockage ());
7711 mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, fp_offset));
7712 set_mem_alias_set (mem, alpha_sr_alias_set);
7713 FRP (emit_move_insn (hard_frame_pointer_rtx, mem));
7715 else if (TARGET_ABI_OPEN_VMS)
7717 emit_insn (gen_blockage ());
7718 FRP (emit_move_insn (hard_frame_pointer_rtx,
7719 gen_rtx_REG (DImode, vms_save_fp_regno)));
7722 /* Restore the stack pointer. */
7723 emit_insn (gen_blockage ());
7724 if (sp_adj2 == const0_rtx)
7725 FRP (emit_move_insn (stack_pointer_rtx, sp_adj1));
7727 FRP (emit_move_insn (stack_pointer_rtx,
7728 gen_rtx_PLUS (DImode, sp_adj1, sp_adj2)));
7732 if (TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_REGISTER)
7734 emit_insn (gen_blockage ());
7735 FRP (emit_move_insn (hard_frame_pointer_rtx,
7736 gen_rtx_REG (DImode, vms_save_fp_regno)));
7738 else if (TARGET_ABI_UNICOSMK && alpha_procedure_type != PT_STACK)
7740 /* Decrement the frame pointer if the function does not have a
7743 emit_insn (gen_blockage ());
7744 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx,
7745 hard_frame_pointer_rtx, GEN_INT (-1))));
7750 /* Output the rest of the textual info surrounding the epilogue. */
7753 alpha_end_function (FILE *file, const char *fnname, tree decl ATTRIBUTE_UNUSED)
7755 /* End the function. */
7756 if (!TARGET_ABI_UNICOSMK && !flag_inhibit_size_directive)
7758 fputs ("\t.end ", file);
7759 assemble_name (file, fnname);
7762 inside_function = FALSE;
7764 #if TARGET_ABI_OPEN_VMS
7765 alpha_write_linkage (file, fnname, decl);
7768 /* Output jump tables and the static subroutine information block. */
7769 if (TARGET_ABI_UNICOSMK)
7771 unicosmk_output_ssib (file, fnname);
7772 unicosmk_output_deferred_case_vectors (file);
7777 /* Emit a tail call to FUNCTION after adjusting THIS by DELTA.
7779 In order to avoid the hordes of differences between generated code
7780 with and without TARGET_EXPLICIT_RELOCS, and to avoid duplicating
7781 lots of code loading up large constants, generate rtl and emit it
7782 instead of going straight to text.
7784 Not sure why this idea hasn't been explored before... */
7787 alpha_output_mi_thunk_osf (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
7788 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
7791 HOST_WIDE_INT hi, lo;
7792 rtx this, insn, funexp;
7794 /* We always require a valid GP. */
7795 emit_insn (gen_prologue_ldgp ());
7796 emit_note (NOTE_INSN_PROLOGUE_END);
7798 /* Find the "this" pointer. If the function returns a structure,
7799 the structure return pointer is in $16. */
7800 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
7801 this = gen_rtx_REG (Pmode, 17);
7803 this = gen_rtx_REG (Pmode, 16);
7805 /* Add DELTA. When possible we use ldah+lda. Otherwise load the
7806 entire constant for the add. */
7807 lo = ((delta & 0xffff) ^ 0x8000) - 0x8000;
7808 hi = (((delta - lo) & 0xffffffff) ^ 0x80000000) - 0x80000000;
7809 if (hi + lo == delta)
7812 emit_insn (gen_adddi3 (this, this, GEN_INT (hi)));
7814 emit_insn (gen_adddi3 (this, this, GEN_INT (lo)));
7818 rtx tmp = alpha_emit_set_long_const (gen_rtx_REG (Pmode, 0),
7819 delta, -(delta < 0));
7820 emit_insn (gen_adddi3 (this, this, tmp));
7823 /* Add a delta stored in the vtable at VCALL_OFFSET. */
7828 tmp = gen_rtx_REG (Pmode, 0);
7829 emit_move_insn (tmp, gen_rtx_MEM (Pmode, this));
7831 lo = ((vcall_offset & 0xffff) ^ 0x8000) - 0x8000;
7832 hi = (((vcall_offset - lo) & 0xffffffff) ^ 0x80000000) - 0x80000000;
7833 if (hi + lo == vcall_offset)
7836 emit_insn (gen_adddi3 (tmp, tmp, GEN_INT (hi)));
7840 tmp2 = alpha_emit_set_long_const (gen_rtx_REG (Pmode, 1),
7841 vcall_offset, -(vcall_offset < 0));
7842 emit_insn (gen_adddi3 (tmp, tmp, tmp2));
7846 tmp2 = gen_rtx_PLUS (Pmode, tmp, GEN_INT (lo));
7849 emit_move_insn (tmp, gen_rtx_MEM (Pmode, tmp2));
7851 emit_insn (gen_adddi3 (this, this, tmp));
7854 /* Generate a tail call to the target function. */
7855 if (! TREE_USED (function))
7857 assemble_external (function);
7858 TREE_USED (function) = 1;
7860 funexp = XEXP (DECL_RTL (function), 0);
7861 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
7862 insn = emit_call_insn (gen_sibcall (funexp, const0_rtx));
7863 SIBLING_CALL_P (insn) = 1;
7865 /* Run just enough of rest_of_compilation to get the insns emitted.
7866 There's not really enough bulk here to make other passes such as
7867 instruction scheduling worth while. Note that use_thunk calls
7868 assemble_start_function and assemble_end_function. */
7869 insn = get_insns ();
7870 insn_locators_initialize ();
7871 shorten_branches (insn);
7872 final_start_function (insn, file, 1);
7873 final (insn, file, 1, 0);
7874 final_end_function ();
7876 #endif /* TARGET_ABI_OSF */
7878 /* Debugging support. */
7882 /* Count the number of sdb related labels are generated (to find block
7883 start and end boundaries). */
7885 int sdb_label_count = 0;
7887 /* Next label # for each statement. */
7889 static int sym_lineno = 0;
7891 /* Count the number of .file directives, so that .loc is up to date. */
7893 static int num_source_filenames = 0;
7895 /* Name of the file containing the current function. */
7897 static const char *current_function_file = "";
7899 /* Offsets to alpha virtual arg/local debugging pointers. */
7901 long alpha_arg_offset;
7902 long alpha_auto_offset;
7904 /* Emit a new filename to a stream. */
7907 alpha_output_filename (FILE *stream, const char *name)
7909 static int first_time = TRUE;
7910 char ltext_label_name[100];
7915 ++num_source_filenames;
7916 current_function_file = name;
7917 fprintf (stream, "\t.file\t%d ", num_source_filenames);
7918 output_quoted_string (stream, name);
7919 fprintf (stream, "\n");
7920 if (!TARGET_GAS && write_symbols == DBX_DEBUG)
7921 fprintf (stream, "\t#@stabs\n");
7924 else if (write_symbols == DBX_DEBUG)
7926 ASM_GENERATE_INTERNAL_LABEL (ltext_label_name, "Ltext", 0);
7927 fprintf (stream, "%s", ASM_STABS_OP);
7928 output_quoted_string (stream, name);
7929 fprintf (stream, ",%d,0,0,%s\n", N_SOL, <ext_label_name[1]);
7932 else if (name != current_function_file
7933 && strcmp (name, current_function_file) != 0)
7935 if (inside_function && ! TARGET_GAS)
7936 fprintf (stream, "\t#.file\t%d ", num_source_filenames);
7939 ++num_source_filenames;
7940 current_function_file = name;
7941 fprintf (stream, "\t.file\t%d ", num_source_filenames);
7944 output_quoted_string (stream, name);
7945 fprintf (stream, "\n");
7949 /* Emit a linenumber to a stream. */
7952 alpha_output_lineno (FILE *stream, int line)
7954 if (write_symbols == DBX_DEBUG)
7956 /* mips-tfile doesn't understand .stabd directives. */
7958 fprintf (stream, "$LM%d:\n%s%d,0,%d,$LM%d\n",
7959 sym_lineno, ASM_STABN_OP, N_SLINE, line, sym_lineno);
7962 fprintf (stream, "\n\t.loc\t%d %d\n", num_source_filenames, line);
7965 /* Structure to show the current status of registers and memory. */
7967 struct shadow_summary
7970 unsigned int i : 31; /* Mask of int regs */
7971 unsigned int fp : 31; /* Mask of fp regs */
7972 unsigned int mem : 1; /* mem == imem | fpmem */
7976 /* Summary the effects of expression X on the machine. Update SUM, a pointer
7977 to the summary structure. SET is nonzero if the insn is setting the
7978 object, otherwise zero. */
7981 summarize_insn (rtx x, struct shadow_summary *sum, int set)
7983 const char *format_ptr;
7989 switch (GET_CODE (x))
7991 /* ??? Note that this case would be incorrect if the Alpha had a
7992 ZERO_EXTRACT in SET_DEST. */
7994 summarize_insn (SET_SRC (x), sum, 0);
7995 summarize_insn (SET_DEST (x), sum, 1);
7999 summarize_insn (XEXP (x, 0), sum, 1);
8003 summarize_insn (XEXP (x, 0), sum, 0);
8007 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
8008 summarize_insn (ASM_OPERANDS_INPUT (x, i), sum, 0);
8012 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
8013 summarize_insn (XVECEXP (x, 0, i), sum, 0);
8017 summarize_insn (SUBREG_REG (x), sum, 0);
8022 int regno = REGNO (x);
8023 unsigned long mask = ((unsigned long) 1) << (regno % 32);
8025 if (regno == 31 || regno == 63)
8031 sum->defd.i |= mask;
8033 sum->defd.fp |= mask;
8038 sum->used.i |= mask;
8040 sum->used.fp |= mask;
8051 /* Find the regs used in memory address computation: */
8052 summarize_insn (XEXP (x, 0), sum, 0);
8055 case CONST_INT: case CONST_DOUBLE:
8056 case SYMBOL_REF: case LABEL_REF: case CONST:
8057 case SCRATCH: case ASM_INPUT:
8060 /* Handle common unary and binary ops for efficiency. */
8061 case COMPARE: case PLUS: case MINUS: case MULT: case DIV:
8062 case MOD: case UDIV: case UMOD: case AND: case IOR:
8063 case XOR: case ASHIFT: case ROTATE: case ASHIFTRT: case LSHIFTRT:
8064 case ROTATERT: case SMIN: case SMAX: case UMIN: case UMAX:
8065 case NE: case EQ: case GE: case GT: case LE:
8066 case LT: case GEU: case GTU: case LEU: case LTU:
8067 summarize_insn (XEXP (x, 0), sum, 0);
8068 summarize_insn (XEXP (x, 1), sum, 0);
8071 case NEG: case NOT: case SIGN_EXTEND: case ZERO_EXTEND:
8072 case TRUNCATE: case FLOAT_EXTEND: case FLOAT_TRUNCATE: case FLOAT:
8073 case FIX: case UNSIGNED_FLOAT: case UNSIGNED_FIX: case ABS:
8074 case SQRT: case FFS:
8075 summarize_insn (XEXP (x, 0), sum, 0);
8079 format_ptr = GET_RTX_FORMAT (GET_CODE (x));
8080 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8081 switch (format_ptr[i])
8084 summarize_insn (XEXP (x, i), sum, 0);
8088 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8089 summarize_insn (XVECEXP (x, i, j), sum, 0);
8101 /* Ensure a sufficient number of `trapb' insns are in the code when
8102 the user requests code with a trap precision of functions or
8105 In naive mode, when the user requests a trap-precision of
8106 "instruction", a trapb is needed after every instruction that may
8107 generate a trap. This ensures that the code is resumption safe but
8110 When optimizations are turned on, we delay issuing a trapb as long
8111 as possible. In this context, a trap shadow is the sequence of
8112 instructions that starts with a (potentially) trap generating
8113 instruction and extends to the next trapb or call_pal instruction
8114 (but GCC never generates call_pal by itself). We can delay (and
8115 therefore sometimes omit) a trapb subject to the following
8118 (a) On entry to the trap shadow, if any Alpha register or memory
8119 location contains a value that is used as an operand value by some
8120 instruction in the trap shadow (live on entry), then no instruction
8121 in the trap shadow may modify the register or memory location.
8123 (b) Within the trap shadow, the computation of the base register
8124 for a memory load or store instruction may not involve using the
8125 result of an instruction that might generate an UNPREDICTABLE
8128 (c) Within the trap shadow, no register may be used more than once
8129 as a destination register. (This is to make life easier for the
8132 (d) The trap shadow may not include any branch instructions. */
8135 alpha_handle_trap_shadows (void)
8137 struct shadow_summary shadow;
8138 int trap_pending, exception_nesting;
8142 exception_nesting = 0;
8145 shadow.used.mem = 0;
8146 shadow.defd = shadow.used;
8148 for (i = get_insns (); i ; i = NEXT_INSN (i))
8150 if (GET_CODE (i) == NOTE)
8152 switch (NOTE_LINE_NUMBER (i))
8154 case NOTE_INSN_EH_REGION_BEG:
8155 exception_nesting++;
8160 case NOTE_INSN_EH_REGION_END:
8161 exception_nesting--;
8166 case NOTE_INSN_EPILOGUE_BEG:
8167 if (trap_pending && alpha_tp >= ALPHA_TP_FUNC)
8172 else if (trap_pending)
8174 if (alpha_tp == ALPHA_TP_FUNC)
8176 if (GET_CODE (i) == JUMP_INSN
8177 && GET_CODE (PATTERN (i)) == RETURN)
8180 else if (alpha_tp == ALPHA_TP_INSN)
8184 struct shadow_summary sum;
8189 sum.defd = sum.used;
8191 switch (GET_CODE (i))
8194 /* Annoyingly, get_attr_trap will abort on these. */
8195 if (GET_CODE (PATTERN (i)) == USE
8196 || GET_CODE (PATTERN (i)) == CLOBBER)
8199 summarize_insn (PATTERN (i), &sum, 0);
8201 if ((sum.defd.i & shadow.defd.i)
8202 || (sum.defd.fp & shadow.defd.fp))
8204 /* (c) would be violated */
8208 /* Combine shadow with summary of current insn: */
8209 shadow.used.i |= sum.used.i;
8210 shadow.used.fp |= sum.used.fp;
8211 shadow.used.mem |= sum.used.mem;
8212 shadow.defd.i |= sum.defd.i;
8213 shadow.defd.fp |= sum.defd.fp;
8214 shadow.defd.mem |= sum.defd.mem;
8216 if ((sum.defd.i & shadow.used.i)
8217 || (sum.defd.fp & shadow.used.fp)
8218 || (sum.defd.mem & shadow.used.mem))
8220 /* (a) would be violated (also takes care of (b)) */
8221 if (get_attr_trap (i) == TRAP_YES
8222 && ((sum.defd.i & sum.used.i)
8223 || (sum.defd.fp & sum.used.fp)))
8242 n = emit_insn_before (gen_trapb (), i);
8243 PUT_MODE (n, TImode);
8244 PUT_MODE (i, TImode);
8248 shadow.used.mem = 0;
8249 shadow.defd = shadow.used;
8254 if ((exception_nesting > 0 || alpha_tp >= ALPHA_TP_FUNC)
8255 && GET_CODE (i) == INSN
8256 && GET_CODE (PATTERN (i)) != USE
8257 && GET_CODE (PATTERN (i)) != CLOBBER
8258 && get_attr_trap (i) == TRAP_YES)
8260 if (optimize && !trap_pending)
8261 summarize_insn (PATTERN (i), &shadow, 0);
8267 /* Alpha can only issue instruction groups simultaneously if they are
8268 suitably aligned. This is very processor-specific. */
8270 enum alphaev4_pipe {
8277 enum alphaev5_pipe {
8288 static enum alphaev4_pipe
8289 alphaev4_insn_pipe (rtx insn)
8291 if (recog_memoized (insn) < 0)
8293 if (get_attr_length (insn) != 4)
8296 switch (get_attr_type (insn))
8330 static enum alphaev5_pipe
8331 alphaev5_insn_pipe (rtx insn)
8333 if (recog_memoized (insn) < 0)
8335 if (get_attr_length (insn) != 4)
8338 switch (get_attr_type (insn))
8379 /* IN_USE is a mask of the slots currently filled within the insn group.
8380 The mask bits come from alphaev4_pipe above. If EV4_IBX is set, then
8381 the insn in EV4_IB0 can be swapped by the hardware into EV4_IB1.
8383 LEN is, of course, the length of the group in bytes. */
8386 alphaev4_next_group (rtx insn, int *pin_use, int *plen)
8393 || GET_CODE (PATTERN (insn)) == CLOBBER
8394 || GET_CODE (PATTERN (insn)) == USE)
8399 enum alphaev4_pipe pipe;
8401 pipe = alphaev4_insn_pipe (insn);
8405 /* Force complex instructions to start new groups. */
8409 /* If this is a completely unrecognized insn, its an asm.
8410 We don't know how long it is, so record length as -1 to
8411 signal a needed realignment. */
8412 if (recog_memoized (insn) < 0)
8415 len = get_attr_length (insn);
8419 if (in_use & EV4_IB0)
8421 if (in_use & EV4_IB1)
8426 in_use |= EV4_IB0 | EV4_IBX;
8430 if (in_use & EV4_IB0)
8432 if (!(in_use & EV4_IBX) || (in_use & EV4_IB1))
8440 if (in_use & EV4_IB1)
8450 /* Haifa doesn't do well scheduling branches. */
8451 if (GET_CODE (insn) == JUMP_INSN)
8455 insn = next_nonnote_insn (insn);
8457 if (!insn || ! INSN_P (insn))
8460 /* Let Haifa tell us where it thinks insn group boundaries are. */
8461 if (GET_MODE (insn) == TImode)
8464 if (GET_CODE (insn) == CLOBBER || GET_CODE (insn) == USE)
8469 insn = next_nonnote_insn (insn);
8477 /* IN_USE is a mask of the slots currently filled within the insn group.
8478 The mask bits come from alphaev5_pipe above. If EV5_E01 is set, then
8479 the insn in EV5_E0 can be swapped by the hardware into EV5_E1.
8481 LEN is, of course, the length of the group in bytes. */
8484 alphaev5_next_group (rtx insn, int *pin_use, int *plen)
8491 || GET_CODE (PATTERN (insn)) == CLOBBER
8492 || GET_CODE (PATTERN (insn)) == USE)
8497 enum alphaev5_pipe pipe;
8499 pipe = alphaev5_insn_pipe (insn);
8503 /* Force complex instructions to start new groups. */
8507 /* If this is a completely unrecognized insn, its an asm.
8508 We don't know how long it is, so record length as -1 to
8509 signal a needed realignment. */
8510 if (recog_memoized (insn) < 0)
8513 len = get_attr_length (insn);
8516 /* ??? Most of the places below, we would like to abort, as
8517 it would indicate an error either in Haifa, or in the
8518 scheduling description. Unfortunately, Haifa never
8519 schedules the last instruction of the BB, so we don't
8520 have an accurate TI bit to go off. */
8522 if (in_use & EV5_E0)
8524 if (in_use & EV5_E1)
8529 in_use |= EV5_E0 | EV5_E01;
8533 if (in_use & EV5_E0)
8535 if (!(in_use & EV5_E01) || (in_use & EV5_E1))
8543 if (in_use & EV5_E1)
8549 if (in_use & EV5_FA)
8551 if (in_use & EV5_FM)
8556 in_use |= EV5_FA | EV5_FAM;
8560 if (in_use & EV5_FA)
8566 if (in_use & EV5_FM)
8579 /* Haifa doesn't do well scheduling branches. */
8580 /* ??? If this is predicted not-taken, slotting continues, except
8581 that no more IBR, FBR, or JSR insns may be slotted. */
8582 if (GET_CODE (insn) == JUMP_INSN)
8586 insn = next_nonnote_insn (insn);
8588 if (!insn || ! INSN_P (insn))
8591 /* Let Haifa tell us where it thinks insn group boundaries are. */
8592 if (GET_MODE (insn) == TImode)
8595 if (GET_CODE (insn) == CLOBBER || GET_CODE (insn) == USE)
8600 insn = next_nonnote_insn (insn);
8609 alphaev4_next_nop (int *pin_use)
8611 int in_use = *pin_use;
8614 if (!(in_use & EV4_IB0))
8619 else if ((in_use & (EV4_IBX|EV4_IB1)) == EV4_IBX)
8624 else if (TARGET_FP && !(in_use & EV4_IB1))
8637 alphaev5_next_nop (int *pin_use)
8639 int in_use = *pin_use;
8642 if (!(in_use & EV5_E1))
8647 else if (TARGET_FP && !(in_use & EV5_FA))
8652 else if (TARGET_FP && !(in_use & EV5_FM))
8664 /* The instruction group alignment main loop. */
8667 alpha_align_insns (unsigned int max_align,
8668 rtx (*next_group) (rtx, int *, int *),
8669 rtx (*next_nop) (int *))
8671 /* ALIGN is the known alignment for the insn group. */
8673 /* OFS is the offset of the current insn in the insn group. */
8675 int prev_in_use, in_use, len;
8678 /* Let shorten branches care for assigning alignments to code labels. */
8679 shorten_branches (get_insns ());
8681 if (align_functions < 4)
8683 else if ((unsigned int) align_functions < max_align)
8684 align = align_functions;
8688 ofs = prev_in_use = 0;
8690 if (GET_CODE (i) == NOTE)
8691 i = next_nonnote_insn (i);
8695 next = (*next_group) (i, &in_use, &len);
8697 /* When we see a label, resync alignment etc. */
8698 if (GET_CODE (i) == CODE_LABEL)
8700 unsigned int new_align = 1 << label_to_alignment (i);
8702 if (new_align >= align)
8704 align = new_align < max_align ? new_align : max_align;
8708 else if (ofs & (new_align-1))
8709 ofs = (ofs | (new_align-1)) + 1;
8714 /* Handle complex instructions special. */
8715 else if (in_use == 0)
8717 /* Asms will have length < 0. This is a signal that we have
8718 lost alignment knowledge. Assume, however, that the asm
8719 will not mis-align instructions. */
8728 /* If the known alignment is smaller than the recognized insn group,
8729 realign the output. */
8730 else if ((int) align < len)
8732 unsigned int new_log_align = len > 8 ? 4 : 3;
8735 where = prev = prev_nonnote_insn (i);
8736 if (!where || GET_CODE (where) != CODE_LABEL)
8739 /* Can't realign between a call and its gp reload. */
8740 if (! (TARGET_EXPLICIT_RELOCS
8741 && prev && GET_CODE (prev) == CALL_INSN))
8743 emit_insn_before (gen_realign (GEN_INT (new_log_align)), where);
8744 align = 1 << new_log_align;
8749 /* If the group won't fit in the same INT16 as the previous,
8750 we need to add padding to keep the group together. Rather
8751 than simply leaving the insn filling to the assembler, we
8752 can make use of the knowledge of what sorts of instructions
8753 were issued in the previous group to make sure that all of
8754 the added nops are really free. */
8755 else if (ofs + len > (int) align)
8757 int nop_count = (align - ofs) / 4;
8760 /* Insert nops before labels, branches, and calls to truely merge
8761 the execution of the nops with the previous instruction group. */
8762 where = prev_nonnote_insn (i);
8765 if (GET_CODE (where) == CODE_LABEL)
8767 rtx where2 = prev_nonnote_insn (where);
8768 if (where2 && GET_CODE (where2) == JUMP_INSN)
8771 else if (GET_CODE (where) == INSN)
8778 emit_insn_before ((*next_nop)(&prev_in_use), where);
8779 while (--nop_count);
8783 ofs = (ofs + len) & (align - 1);
8784 prev_in_use = in_use;
8789 /* Machine dependent reorg pass. */
8794 if (alpha_tp != ALPHA_TP_PROG || flag_exceptions)
8795 alpha_handle_trap_shadows ();
8797 /* Due to the number of extra trapb insns, don't bother fixing up
8798 alignment when trap precision is instruction. Moreover, we can
8799 only do our job when sched2 is run. */
8800 if (optimize && !optimize_size
8801 && alpha_tp != ALPHA_TP_INSN
8802 && flag_schedule_insns_after_reload)
8804 if (alpha_cpu == PROCESSOR_EV4)
8805 alpha_align_insns (8, alphaev4_next_group, alphaev4_next_nop);
8806 else if (alpha_cpu == PROCESSOR_EV5)
8807 alpha_align_insns (16, alphaev5_next_group, alphaev5_next_nop);
8811 #if !TARGET_ABI_UNICOSMK
8818 alpha_file_start (void)
8820 #ifdef OBJECT_FORMAT_ELF
8821 /* If emitting dwarf2 debug information, we cannot generate a .file
8822 directive to start the file, as it will conflict with dwarf2out
8823 file numbers. So it's only useful when emitting mdebug output. */
8824 targetm.file_start_file_directive = (write_symbols == DBX_DEBUG);
8827 default_file_start ();
8829 fprintf (asm_out_file, "\t.verstamp %d %d\n", MS_STAMP, LS_STAMP);
8832 fputs ("\t.set noreorder\n", asm_out_file);
8833 fputs ("\t.set volatile\n", asm_out_file);
8834 if (!TARGET_ABI_OPEN_VMS)
8835 fputs ("\t.set noat\n", asm_out_file);
8836 if (TARGET_EXPLICIT_RELOCS)
8837 fputs ("\t.set nomacro\n", asm_out_file);
8838 if (TARGET_SUPPORT_ARCH | TARGET_BWX | TARGET_MAX | TARGET_FIX | TARGET_CIX)
8839 fprintf (asm_out_file,
8841 TARGET_CPU_EV6 ? "ev6"
8843 ? (TARGET_MAX ? "pca56" : TARGET_BWX ? "ev56" : "ev5")
8848 #ifdef OBJECT_FORMAT_ELF
8850 /* Switch to the section to which we should output X. The only thing
8851 special we do here is to honor small data. */
8854 alpha_elf_select_rtx_section (enum machine_mode mode, rtx x,
8855 unsigned HOST_WIDE_INT align)
8857 if (TARGET_SMALL_DATA && GET_MODE_SIZE (mode) <= g_switch_value)
8858 /* ??? Consider using mergeable sdata sections. */
8861 default_elf_select_rtx_section (mode, x, align);
8864 #endif /* OBJECT_FORMAT_ELF */
8866 /* Structure to collect function names for final output in link section. */
8867 /* Note that items marked with GTY can't be ifdef'ed out. */
8869 enum links_kind {KIND_UNUSED, KIND_LOCAL, KIND_EXTERN};
8870 enum reloc_kind {KIND_LINKAGE, KIND_CODEADDR};
8872 struct alpha_links GTY(())
8876 enum links_kind lkind;
8877 enum reloc_kind rkind;
8880 struct alpha_funcs GTY(())
8883 splay_tree GTY ((param1_is (char *), param2_is (struct alpha_links *)))
8887 static GTY ((param1_is (char *), param2_is (struct alpha_links *)))
8888 splay_tree alpha_links_tree;
8889 static GTY ((param1_is (tree), param2_is (struct alpha_funcs *)))
8890 splay_tree alpha_funcs_tree;
8892 static GTY(()) int alpha_funcs_num;
8894 #if TARGET_ABI_OPEN_VMS
8896 /* Return the VMS argument type corresponding to MODE. */
8899 alpha_arg_type (enum machine_mode mode)
8904 return TARGET_FLOAT_VAX ? FF : FS;
8906 return TARGET_FLOAT_VAX ? FD : FT;
8912 /* Return an rtx for an integer representing the VMS Argument Information
8916 alpha_arg_info_reg_val (CUMULATIVE_ARGS cum)
8918 unsigned HOST_WIDE_INT regval = cum.num_args;
8921 for (i = 0; i < 6; i++)
8922 regval |= ((int) cum.atypes[i]) << (i * 3 + 8);
8924 return GEN_INT (regval);
8927 /* Make (or fake) .linkage entry for function call.
8929 IS_LOCAL is 0 if name is used in call, 1 if name is used in definition.
8931 Return an SYMBOL_REF rtx for the linkage. */
8934 alpha_need_linkage (const char *name, int is_local)
8936 splay_tree_node node;
8937 struct alpha_links *al;
8944 struct alpha_funcs *cfaf;
8946 if (!alpha_funcs_tree)
8947 alpha_funcs_tree = splay_tree_new_ggc ((splay_tree_compare_fn)
8948 splay_tree_compare_pointers);
8950 cfaf = (struct alpha_funcs *) ggc_alloc (sizeof (struct alpha_funcs));
8953 cfaf->num = ++alpha_funcs_num;
8955 splay_tree_insert (alpha_funcs_tree,
8956 (splay_tree_key) current_function_decl,
8957 (splay_tree_value) cfaf);
8960 if (alpha_links_tree)
8962 /* Is this name already defined? */
8964 node = splay_tree_lookup (alpha_links_tree, (splay_tree_key) name);
8967 al = (struct alpha_links *) node->value;
8970 /* Defined here but external assumed. */
8971 if (al->lkind == KIND_EXTERN)
8972 al->lkind = KIND_LOCAL;
8976 /* Used here but unused assumed. */
8977 if (al->lkind == KIND_UNUSED)
8978 al->lkind = KIND_LOCAL;
8984 alpha_links_tree = splay_tree_new_ggc ((splay_tree_compare_fn) strcmp);
8986 al = (struct alpha_links *) ggc_alloc (sizeof (struct alpha_links));
8987 name = ggc_strdup (name);
8989 /* Assume external if no definition. */
8990 al->lkind = (is_local ? KIND_UNUSED : KIND_EXTERN);
8992 /* Ensure we have an IDENTIFIER so assemble_name can mark it used. */
8993 get_identifier (name);
8995 /* Construct a SYMBOL_REF for us to call. */
8997 size_t name_len = strlen (name);
8998 char *linksym = alloca (name_len + 6);
9000 memcpy (linksym + 1, name, name_len);
9001 memcpy (linksym + 1 + name_len, "..lk", 5);
9002 al->linkage = gen_rtx_SYMBOL_REF (Pmode,
9003 ggc_alloc_string (linksym, name_len + 5));
9006 splay_tree_insert (alpha_links_tree, (splay_tree_key) name,
9007 (splay_tree_value) al);
9013 alpha_use_linkage (rtx linkage, tree cfundecl, int lflag, int rflag)
9015 splay_tree_node cfunnode;
9016 struct alpha_funcs *cfaf;
9017 struct alpha_links *al;
9018 const char *name = XSTR (linkage, 0);
9020 cfaf = (struct alpha_funcs *) 0;
9021 al = (struct alpha_links *) 0;
9023 cfunnode = splay_tree_lookup (alpha_funcs_tree, (splay_tree_key) cfundecl);
9024 cfaf = (struct alpha_funcs *) cfunnode->value;
9028 splay_tree_node lnode;
9030 /* Is this name already defined? */
9032 lnode = splay_tree_lookup (cfaf->links, (splay_tree_key) name);
9034 al = (struct alpha_links *) lnode->value;
9037 cfaf->links = splay_tree_new_ggc ((splay_tree_compare_fn) strcmp);
9045 splay_tree_node node = 0;
9046 struct alpha_links *anl;
9051 name_len = strlen (name);
9053 al = (struct alpha_links *) ggc_alloc (sizeof (struct alpha_links));
9054 al->num = cfaf->num;
9056 node = splay_tree_lookup (alpha_links_tree, (splay_tree_key) name);
9059 anl = (struct alpha_links *) node->value;
9060 al->lkind = anl->lkind;
9063 sprintf (buf, "$%d..%s..lk", cfaf->num, name);
9064 buflen = strlen (buf);
9065 linksym = alloca (buflen + 1);
9066 memcpy (linksym, buf, buflen + 1);
9068 al->linkage = gen_rtx_SYMBOL_REF
9069 (Pmode, ggc_alloc_string (linksym, buflen + 1));
9071 splay_tree_insert (cfaf->links, (splay_tree_key) name,
9072 (splay_tree_value) al);
9076 al->rkind = KIND_CODEADDR;
9078 al->rkind = KIND_LINKAGE;
9081 return gen_rtx_MEM (Pmode, plus_constant (al->linkage, 8));
9087 alpha_write_one_linkage (splay_tree_node node, void *data)
9089 const char *const name = (const char *) node->key;
9090 struct alpha_links *link = (struct alpha_links *) node->value;
9091 FILE *stream = (FILE *) data;
9093 fprintf (stream, "$%d..%s..lk:\n", link->num, name);
9094 if (link->rkind == KIND_CODEADDR)
9096 if (link->lkind == KIND_LOCAL)
9098 /* Local and used */
9099 fprintf (stream, "\t.quad %s..en\n", name);
9103 /* External and used, request code address. */
9104 fprintf (stream, "\t.code_address %s\n", name);
9109 if (link->lkind == KIND_LOCAL)
9111 /* Local and used, build linkage pair. */
9112 fprintf (stream, "\t.quad %s..en\n", name);
9113 fprintf (stream, "\t.quad %s\n", name);
9117 /* External and used, request linkage pair. */
9118 fprintf (stream, "\t.linkage %s\n", name);
9126 alpha_write_linkage (FILE *stream, const char *funname, tree fundecl)
9128 splay_tree_node node;
9129 struct alpha_funcs *func;
9132 fprintf (stream, "\t.align 3\n");
9133 node = splay_tree_lookup (alpha_funcs_tree, (splay_tree_key) fundecl);
9134 func = (struct alpha_funcs *) node->value;
9136 fputs ("\t.name ", stream);
9137 assemble_name (stream, funname);
9138 fputs ("..na\n", stream);
9139 ASM_OUTPUT_LABEL (stream, funname);
9140 fprintf (stream, "\t.pdesc ");
9141 assemble_name (stream, funname);
9142 fprintf (stream, "..en,%s\n",
9143 alpha_procedure_type == PT_STACK ? "stack"
9144 : alpha_procedure_type == PT_REGISTER ? "reg" : "null");
9148 splay_tree_foreach (func->links, alpha_write_one_linkage, stream);
9149 /* splay_tree_delete (func->links); */
9153 /* Given a decl, a section name, and whether the decl initializer
9154 has relocs, choose attributes for the section. */
9156 #define SECTION_VMS_OVERLAY SECTION_FORGET
9157 #define SECTION_VMS_GLOBAL SECTION_MACH_DEP
9158 #define SECTION_VMS_INITIALIZE (SECTION_VMS_GLOBAL << 1)
9161 vms_section_type_flags (tree decl, const char *name, int reloc)
9163 unsigned int flags = default_section_type_flags (decl, name, reloc);
9165 if (decl && DECL_ATTRIBUTES (decl)
9166 && lookup_attribute ("overlaid", DECL_ATTRIBUTES (decl)))
9167 flags |= SECTION_VMS_OVERLAY;
9168 if (decl && DECL_ATTRIBUTES (decl)
9169 && lookup_attribute ("global", DECL_ATTRIBUTES (decl)))
9170 flags |= SECTION_VMS_GLOBAL;
9171 if (decl && DECL_ATTRIBUTES (decl)
9172 && lookup_attribute ("initialize", DECL_ATTRIBUTES (decl)))
9173 flags |= SECTION_VMS_INITIALIZE;
9178 /* Switch to an arbitrary section NAME with attributes as specified
9179 by FLAGS. ALIGN specifies any known alignment requirements for
9180 the section; 0 if the default should be used. */
9183 vms_asm_named_section (const char *name, unsigned int flags)
9185 fputc ('\n', asm_out_file);
9186 fprintf (asm_out_file, ".section\t%s", name);
9188 if (flags & SECTION_VMS_OVERLAY)
9189 fprintf (asm_out_file, ",OVR");
9190 if (flags & SECTION_VMS_GLOBAL)
9191 fprintf (asm_out_file, ",GBL");
9192 if (flags & SECTION_VMS_INITIALIZE)
9193 fprintf (asm_out_file, ",NOMOD");
9194 if (flags & SECTION_DEBUG)
9195 fprintf (asm_out_file, ",NOWRT");
9197 fputc ('\n', asm_out_file);
9200 /* Record an element in the table of global constructors. SYMBOL is
9201 a SYMBOL_REF of the function to be called; PRIORITY is a number
9202 between 0 and MAX_INIT_PRIORITY.
9204 Differs from default_ctors_section_asm_out_constructor in that the
9205 width of the .ctors entry is always 64 bits, rather than the 32 bits
9206 used by a normal pointer. */
9209 vms_asm_out_constructor (rtx symbol, int priority ATTRIBUTE_UNUSED)
9212 assemble_align (BITS_PER_WORD);
9213 assemble_integer (symbol, UNITS_PER_WORD, BITS_PER_WORD, 1);
9217 vms_asm_out_destructor (rtx symbol, int priority ATTRIBUTE_UNUSED)
9220 assemble_align (BITS_PER_WORD);
9221 assemble_integer (symbol, UNITS_PER_WORD, BITS_PER_WORD, 1);
9226 alpha_need_linkage (const char *name ATTRIBUTE_UNUSED,
9227 int is_local ATTRIBUTE_UNUSED)
9233 alpha_use_linkage (rtx linkage ATTRIBUTE_UNUSED,
9234 tree cfundecl ATTRIBUTE_UNUSED,
9235 int lflag ATTRIBUTE_UNUSED,
9236 int rflag ATTRIBUTE_UNUSED)
9241 #endif /* TARGET_ABI_OPEN_VMS */
9243 #if TARGET_ABI_UNICOSMK
9245 /* Define the offset between two registers, one to be eliminated, and the
9246 other its replacement, at the start of a routine. */
9249 unicosmk_initial_elimination_offset (int from, int to)
9253 fixed_size = alpha_sa_size();
9254 if (fixed_size != 0)
9257 if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
9259 else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
9261 else if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
9262 return (ALPHA_ROUND (current_function_outgoing_args_size)
9263 + ALPHA_ROUND (get_frame_size()));
9264 else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
9265 return (ALPHA_ROUND (fixed_size)
9266 + ALPHA_ROUND (get_frame_size()
9267 + current_function_outgoing_args_size));
9272 /* Output the module name for .ident and .end directives. We have to strip
9273 directories and add make sure that the module name starts with a letter
9277 unicosmk_output_module_name (FILE *file)
9279 const char *name = lbasename (main_input_filename);
9280 unsigned len = strlen (name);
9281 char *clean_name = alloca (len + 2);
9282 char *ptr = clean_name;
9284 /* CAM only accepts module names that start with a letter or '$'. We
9285 prefix the module name with a '$' if necessary. */
9287 if (!ISALPHA (*name))
9289 memcpy (ptr, name, len + 1);
9290 clean_symbol_name (clean_name);
9291 fputs (clean_name, file);
9294 /* Output the definition of a common variable. */
9297 unicosmk_output_common (FILE *file, const char *name, int size, int align)
9300 printf ("T3E__: common %s\n", name);
9303 fputs("\t.endp\n\n\t.psect ", file);
9304 assemble_name(file, name);
9305 fprintf(file, ",%d,common\n", floor_log2 (align / BITS_PER_UNIT));
9306 fprintf(file, "\t.byte\t0:%d\n", size);
9308 /* Mark the symbol as defined in this module. */
9309 name_tree = get_identifier (name);
9310 TREE_ASM_WRITTEN (name_tree) = 1;
9313 #define SECTION_PUBLIC SECTION_MACH_DEP
9314 #define SECTION_MAIN (SECTION_PUBLIC << 1)
9315 static int current_section_align;
9318 unicosmk_section_type_flags (tree decl, const char *name,
9319 int reloc ATTRIBUTE_UNUSED)
9321 unsigned int flags = default_section_type_flags (decl, name, reloc);
9326 if (TREE_CODE (decl) == FUNCTION_DECL)
9328 current_section_align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9329 if (align_functions_log > current_section_align)
9330 current_section_align = align_functions_log;
9332 if (! strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)), "main"))
9333 flags |= SECTION_MAIN;
9336 current_section_align = floor_log2 (DECL_ALIGN (decl) / BITS_PER_UNIT);
9338 if (TREE_PUBLIC (decl))
9339 flags |= SECTION_PUBLIC;
9344 /* Generate a section name for decl and associate it with the
9348 unicosmk_unique_section (tree decl, int reloc ATTRIBUTE_UNUSED)
9356 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
9357 name = default_strip_name_encoding (name);
9358 len = strlen (name);
9360 if (TREE_CODE (decl) == FUNCTION_DECL)
9364 /* It is essential that we prefix the section name here because
9365 otherwise the section names generated for constructors and
9366 destructors confuse collect2. */
9368 string = alloca (len + 6);
9369 sprintf (string, "code@%s", name);
9370 DECL_SECTION_NAME (decl) = build_string (len + 5, string);
9372 else if (TREE_PUBLIC (decl))
9373 DECL_SECTION_NAME (decl) = build_string (len, name);
9378 string = alloca (len + 6);
9379 sprintf (string, "data@%s", name);
9380 DECL_SECTION_NAME (decl) = build_string (len + 5, string);
9384 /* Switch to an arbitrary section NAME with attributes as specified
9385 by FLAGS. ALIGN specifies any known alignment requirements for
9386 the section; 0 if the default should be used. */
9389 unicosmk_asm_named_section (const char *name, unsigned int flags)
9393 /* Close the previous section. */
9395 fputs ("\t.endp\n\n", asm_out_file);
9397 /* Find out what kind of section we are opening. */
9399 if (flags & SECTION_MAIN)
9400 fputs ("\t.start\tmain\n", asm_out_file);
9402 if (flags & SECTION_CODE)
9404 else if (flags & SECTION_PUBLIC)
9409 if (current_section_align != 0)
9410 fprintf (asm_out_file, "\t.psect\t%s,%d,%s\n", name,
9411 current_section_align, kind);
9413 fprintf (asm_out_file, "\t.psect\t%s,%s\n", name, kind);
9417 unicosmk_insert_attributes (tree decl, tree *attr_ptr ATTRIBUTE_UNUSED)
9420 && (TREE_PUBLIC (decl) || TREE_CODE (decl) == FUNCTION_DECL))
9421 unicosmk_unique_section (decl, 0);
9424 /* Output an alignment directive. We have to use the macro 'gcc@code@align'
9425 in code sections because .align fill unused space with zeroes. */
9428 unicosmk_output_align (FILE *file, int align)
9430 if (inside_function)
9431 fprintf (file, "\tgcc@code@align\t%d\n", align);
9433 fprintf (file, "\t.align\t%d\n", align);
9436 /* Add a case vector to the current function's list of deferred case
9437 vectors. Case vectors have to be put into a separate section because CAM
9438 does not allow data definitions in code sections. */
9441 unicosmk_defer_case_vector (rtx lab, rtx vec)
9443 struct machine_function *machine = cfun->machine;
9445 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
9446 machine->addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec,
9447 machine->addr_list);
9450 /* Output a case vector. */
9453 unicosmk_output_addr_vec (FILE *file, rtx vec)
9455 rtx lab = XEXP (vec, 0);
9456 rtx body = XEXP (vec, 1);
9457 int vlen = XVECLEN (body, 0);
9460 (*targetm.asm_out.internal_label) (file, "L", CODE_LABEL_NUMBER (lab));
9462 for (idx = 0; idx < vlen; idx++)
9464 ASM_OUTPUT_ADDR_VEC_ELT
9465 (file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
9469 /* Output current function's deferred case vectors. */
9472 unicosmk_output_deferred_case_vectors (FILE *file)
9474 struct machine_function *machine = cfun->machine;
9477 if (machine->addr_list == NULL_RTX)
9481 for (t = machine->addr_list; t; t = XEXP (t, 1))
9482 unicosmk_output_addr_vec (file, XEXP (t, 0));
9485 /* Generate the name of the SSIB section for the current function. */
9487 #define SSIB_PREFIX "__SSIB_"
9488 #define SSIB_PREFIX_LEN 7
9491 unicosmk_ssib_name (void)
9493 /* This is ok since CAM won't be able to deal with names longer than that
9496 static char name[256];
9502 x = DECL_RTL (cfun->decl);
9503 if (GET_CODE (x) != MEM)
9506 if (GET_CODE (x) != SYMBOL_REF)
9508 fnname = XSTR (x, 0);
9510 len = strlen (fnname);
9511 if (len + SSIB_PREFIX_LEN > 255)
9512 len = 255 - SSIB_PREFIX_LEN;
9514 strcpy (name, SSIB_PREFIX);
9515 strncpy (name + SSIB_PREFIX_LEN, fnname, len);
9516 name[len + SSIB_PREFIX_LEN] = 0;
9521 /* Set up the dynamic subprogram information block (DSIB) and update the
9522 frame pointer register ($15) for subroutines which have a frame. If the
9523 subroutine doesn't have a frame, simply increment $15. */
9526 unicosmk_gen_dsib (unsigned long *imaskP)
9528 if (alpha_procedure_type == PT_STACK)
9530 const char *ssib_name;
9533 /* Allocate 64 bytes for the DSIB. */
9535 FRP (emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx,
9537 emit_insn (gen_blockage ());
9539 /* Save the return address. */
9541 mem = gen_rtx_MEM (DImode, plus_constant (stack_pointer_rtx, 56));
9542 set_mem_alias_set (mem, alpha_sr_alias_set);
9543 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, REG_RA)));
9544 (*imaskP) &= ~(1UL << REG_RA);
9546 /* Save the old frame pointer. */
9548 mem = gen_rtx_MEM (DImode, plus_constant (stack_pointer_rtx, 48));
9549 set_mem_alias_set (mem, alpha_sr_alias_set);
9550 FRP (emit_move_insn (mem, hard_frame_pointer_rtx));
9551 (*imaskP) &= ~(1UL << HARD_FRAME_POINTER_REGNUM);
9553 emit_insn (gen_blockage ());
9555 /* Store the SSIB pointer. */
9557 ssib_name = ggc_strdup (unicosmk_ssib_name ());
9558 mem = gen_rtx_MEM (DImode, plus_constant (stack_pointer_rtx, 32));
9559 set_mem_alias_set (mem, alpha_sr_alias_set);
9561 FRP (emit_move_insn (gen_rtx_REG (DImode, 5),
9562 gen_rtx_SYMBOL_REF (Pmode, ssib_name)));
9563 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, 5)));
9565 /* Save the CIW index. */
9567 mem = gen_rtx_MEM (DImode, plus_constant (stack_pointer_rtx, 24));
9568 set_mem_alias_set (mem, alpha_sr_alias_set);
9569 FRP (emit_move_insn (mem, gen_rtx_REG (DImode, 25)));
9571 emit_insn (gen_blockage ());
9573 /* Set the new frame pointer. */
9575 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx,
9576 stack_pointer_rtx, GEN_INT (64))));
9581 /* Increment the frame pointer register to indicate that we do not
9584 FRP (emit_insn (gen_adddi3 (hard_frame_pointer_rtx,
9585 hard_frame_pointer_rtx, GEN_INT (1))));
9589 /* Output the static subroutine information block for the current
9593 unicosmk_output_ssib (FILE *file, const char *fnname)
9599 struct machine_function *machine = cfun->machine;
9602 fprintf (file, "\t.endp\n\n\t.psect\t%s%s,data\n", user_label_prefix,
9603 unicosmk_ssib_name ());
9605 /* Some required stuff and the function name length. */
9607 len = strlen (fnname);
9608 fprintf (file, "\t.quad\t^X20008%2.2X28\n", len);
9611 ??? We don't do that yet. */
9613 fputs ("\t.quad\t0\n", file);
9615 /* Function address. */
9617 fputs ("\t.quad\t", file);
9618 assemble_name (file, fnname);
9621 fputs ("\t.quad\t0\n", file);
9622 fputs ("\t.quad\t0\n", file);
9625 ??? We do it the same way Cray CC does it but this could be
9628 for( i = 0; i < len; i++ )
9629 fprintf (file, "\t.byte\t%d\n", (int)(fnname[i]));
9630 if( (len % 8) == 0 )
9631 fputs ("\t.quad\t0\n", file);
9633 fprintf (file, "\t.bits\t%d : 0\n", (8 - (len % 8))*8);
9635 /* All call information words used in the function. */
9637 for (x = machine->first_ciw; x; x = XEXP (x, 1))
9640 #if HOST_BITS_PER_WIDE_INT == 32
9641 fprintf (file, "\t.quad\t" HOST_WIDE_INT_PRINT_DOUBLE_HEX "\n",
9642 CONST_DOUBLE_HIGH (ciw), CONST_DOUBLE_LOW (ciw));
9644 fprintf (file, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX "\n", INTVAL (ciw));
9649 /* Add a call information word (CIW) to the list of the current function's
9650 CIWs and return its index.
9652 X is a CONST_INT or CONST_DOUBLE representing the CIW. */
9655 unicosmk_add_call_info_word (rtx x)
9658 struct machine_function *machine = cfun->machine;
9660 node = gen_rtx_EXPR_LIST (VOIDmode, x, NULL_RTX);
9661 if (machine->first_ciw == NULL_RTX)
9662 machine->first_ciw = node;
9664 XEXP (machine->last_ciw, 1) = node;
9666 machine->last_ciw = node;
9667 ++machine->ciw_count;
9669 return GEN_INT (machine->ciw_count
9670 + strlen (current_function_name)/8 + 5);
9673 static char unicosmk_section_buf[100];
9676 unicosmk_text_section (void)
9678 static int count = 0;
9679 sprintf (unicosmk_section_buf, "\t.endp\n\n\t.psect\tgcc@text___%d,code",
9681 return unicosmk_section_buf;
9685 unicosmk_data_section (void)
9687 static int count = 1;
9688 sprintf (unicosmk_section_buf, "\t.endp\n\n\t.psect\tgcc@data___%d,data",
9690 return unicosmk_section_buf;
9693 /* The Cray assembler doesn't accept extern declarations for symbols which
9694 are defined in the same file. We have to keep track of all global
9695 symbols which are referenced and/or defined in a source file and output
9696 extern declarations for those which are referenced but not defined at
9699 /* List of identifiers for which an extern declaration might have to be
9701 /* FIXME: needs to use GC, so it can be saved and restored for PCH. */
9703 struct unicosmk_extern_list
9705 struct unicosmk_extern_list *next;
9709 static struct unicosmk_extern_list *unicosmk_extern_head = 0;
9711 /* Output extern declarations which are required for every asm file. */
9714 unicosmk_output_default_externs (FILE *file)
9716 static const char *const externs[] =
9717 { "__T3E_MISMATCH" };
9722 n = ARRAY_SIZE (externs);
9724 for (i = 0; i < n; i++)
9725 fprintf (file, "\t.extern\t%s\n", externs[i]);
9728 /* Output extern declarations for global symbols which are have been
9729 referenced but not defined. */
9732 unicosmk_output_externs (FILE *file)
9734 struct unicosmk_extern_list *p;
9735 const char *real_name;
9739 len = strlen (user_label_prefix);
9740 for (p = unicosmk_extern_head; p != 0; p = p->next)
9742 /* We have to strip the encoding and possibly remove user_label_prefix
9743 from the identifier in order to handle -fleading-underscore and
9744 explicit asm names correctly (cf. gcc.dg/asm-names-1.c). */
9745 real_name = default_strip_name_encoding (p->name);
9746 if (len && p->name[0] == '*'
9747 && !memcmp (real_name, user_label_prefix, len))
9750 name_tree = get_identifier (real_name);
9751 if (! TREE_ASM_WRITTEN (name_tree))
9753 TREE_ASM_WRITTEN (name_tree) = 1;
9754 fputs ("\t.extern\t", file);
9755 assemble_name (file, p->name);
9761 /* Record an extern. */
9764 unicosmk_add_extern (const char *name)
9766 struct unicosmk_extern_list *p;
9768 p = (struct unicosmk_extern_list *)
9769 xmalloc (sizeof (struct unicosmk_extern_list));
9770 p->next = unicosmk_extern_head;
9772 unicosmk_extern_head = p;
9775 /* The Cray assembler generates incorrect code if identifiers which
9776 conflict with register names are used as instruction operands. We have
9777 to replace such identifiers with DEX expressions. */
9779 /* Structure to collect identifiers which have been replaced by DEX
9781 /* FIXME: needs to use GC, so it can be saved and restored for PCH. */
9783 struct unicosmk_dex {
9784 struct unicosmk_dex *next;
9788 /* List of identifiers which have been replaced by DEX expressions. The DEX
9789 number is determined by the position in the list. */
9791 static struct unicosmk_dex *unicosmk_dex_list = NULL;
9793 /* The number of elements in the DEX list. */
9795 static int unicosmk_dex_count = 0;
9797 /* Check if NAME must be replaced by a DEX expression. */
9800 unicosmk_special_name (const char *name)
9808 if (name[0] != 'r' && name[0] != 'f' && name[0] != 'R' && name[0] != 'F')
9814 return (name[2] == '\0' || (ISDIGIT (name[2]) && name[3] == '\0'));
9817 return (name[2] == '\0'
9818 || ((name[2] == '0' || name[2] == '1') && name[3] == '\0'));
9821 return (ISDIGIT (name[1]) && name[2] == '\0');
9825 /* Return the DEX number if X must be replaced by a DEX expression and 0
9829 unicosmk_need_dex (rtx x)
9831 struct unicosmk_dex *dex;
9835 if (GET_CODE (x) != SYMBOL_REF)
9839 if (! unicosmk_special_name (name))
9842 i = unicosmk_dex_count;
9843 for (dex = unicosmk_dex_list; dex; dex = dex->next)
9845 if (! strcmp (name, dex->name))
9850 dex = (struct unicosmk_dex *) xmalloc (sizeof (struct unicosmk_dex));
9852 dex->next = unicosmk_dex_list;
9853 unicosmk_dex_list = dex;
9855 ++unicosmk_dex_count;
9856 return unicosmk_dex_count;
9859 /* Output the DEX definitions for this file. */
9862 unicosmk_output_dex (FILE *file)
9864 struct unicosmk_dex *dex;
9867 if (unicosmk_dex_list == NULL)
9870 fprintf (file, "\t.dexstart\n");
9872 i = unicosmk_dex_count;
9873 for (dex = unicosmk_dex_list; dex; dex = dex->next)
9875 fprintf (file, "\tDEX (%d) = ", i);
9876 assemble_name (file, dex->name);
9881 fprintf (file, "\t.dexend\n");
9884 /* Output text that to appear at the beginning of an assembler file. */
9887 unicosmk_file_start (void)
9891 fputs ("\t.ident\t", asm_out_file);
9892 unicosmk_output_module_name (asm_out_file);
9893 fputs ("\n\n", asm_out_file);
9895 /* The Unicos/Mk assembler uses different register names. Instead of trying
9896 to support them, we simply use micro definitions. */
9898 /* CAM has different register names: rN for the integer register N and fN
9899 for the floating-point register N. Instead of trying to use these in
9900 alpha.md, we define the symbols $N and $fN to refer to the appropriate
9903 for (i = 0; i < 32; ++i)
9904 fprintf (asm_out_file, "$%d <- r%d\n", i, i);
9906 for (i = 0; i < 32; ++i)
9907 fprintf (asm_out_file, "$f%d <- f%d\n", i, i);
9909 putc ('\n', asm_out_file);
9911 /* The .align directive fill unused space with zeroes which does not work
9912 in code sections. We define the macro 'gcc@code@align' which uses nops
9913 instead. Note that it assumes that code sections always have the
9914 biggest possible alignment since . refers to the current offset from
9915 the beginning of the section. */
9917 fputs ("\t.macro gcc@code@align n\n", asm_out_file);
9918 fputs ("gcc@n@bytes = 1 << n\n", asm_out_file);
9919 fputs ("gcc@here = . % gcc@n@bytes\n", asm_out_file);
9920 fputs ("\t.if ne, gcc@here, 0\n", asm_out_file);
9921 fputs ("\t.repeat (gcc@n@bytes - gcc@here) / 4\n", asm_out_file);
9922 fputs ("\tbis r31,r31,r31\n", asm_out_file);
9923 fputs ("\t.endr\n", asm_out_file);
9924 fputs ("\t.endif\n", asm_out_file);
9925 fputs ("\t.endm gcc@code@align\n\n", asm_out_file);
9927 /* Output extern declarations which should always be visible. */
9928 unicosmk_output_default_externs (asm_out_file);
9930 /* Open a dummy section. We always need to be inside a section for the
9931 section-switching code to work correctly.
9932 ??? This should be a module id or something like that. I still have to
9933 figure out what the rules for those are. */
9934 fputs ("\n\t.psect\t$SG00000,data\n", asm_out_file);
9937 /* Output text to appear at the end of an assembler file. This includes all
9938 pending extern declarations and DEX expressions. */
9941 unicosmk_file_end (void)
9943 fputs ("\t.endp\n\n", asm_out_file);
9945 /* Output all pending externs. */
9947 unicosmk_output_externs (asm_out_file);
9949 /* Output dex definitions used for functions whose names conflict with
9952 unicosmk_output_dex (asm_out_file);
9954 fputs ("\t.end\t", asm_out_file);
9955 unicosmk_output_module_name (asm_out_file);
9956 putc ('\n', asm_out_file);
9962 unicosmk_output_deferred_case_vectors (FILE *file ATTRIBUTE_UNUSED)
9966 unicosmk_gen_dsib (unsigned long *imaskP ATTRIBUTE_UNUSED)
9970 unicosmk_output_ssib (FILE * file ATTRIBUTE_UNUSED,
9971 const char * fnname ATTRIBUTE_UNUSED)
9975 unicosmk_add_call_info_word (rtx x ATTRIBUTE_UNUSED)
9981 unicosmk_need_dex (rtx x ATTRIBUTE_UNUSED)
9986 #endif /* TARGET_ABI_UNICOSMK */
9989 /* Initialize the GCC target structure. */
9990 #if TARGET_ABI_OPEN_VMS
9991 # undef TARGET_ATTRIBUTE_TABLE
9992 # define TARGET_ATTRIBUTE_TABLE vms_attribute_table
9993 # undef TARGET_SECTION_TYPE_FLAGS
9994 # define TARGET_SECTION_TYPE_FLAGS vms_section_type_flags
9997 #undef TARGET_IN_SMALL_DATA_P
9998 #define TARGET_IN_SMALL_DATA_P alpha_in_small_data_p
10000 #if TARGET_ABI_UNICOSMK
10001 # undef TARGET_INSERT_ATTRIBUTES
10002 # define TARGET_INSERT_ATTRIBUTES unicosmk_insert_attributes
10003 # undef TARGET_SECTION_TYPE_FLAGS
10004 # define TARGET_SECTION_TYPE_FLAGS unicosmk_section_type_flags
10005 # undef TARGET_ASM_UNIQUE_SECTION
10006 # define TARGET_ASM_UNIQUE_SECTION unicosmk_unique_section
10007 # undef TARGET_ASM_GLOBALIZE_LABEL
10008 # define TARGET_ASM_GLOBALIZE_LABEL hook_void_FILEptr_constcharptr
10011 #undef TARGET_ASM_ALIGNED_HI_OP
10012 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10013 #undef TARGET_ASM_ALIGNED_DI_OP
10014 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10016 /* Default unaligned ops are provided for ELF systems. To get unaligned
10017 data for non-ELF systems, we have to turn off auto alignment. */
10018 #ifndef OBJECT_FORMAT_ELF
10019 #undef TARGET_ASM_UNALIGNED_HI_OP
10020 #define TARGET_ASM_UNALIGNED_HI_OP "\t.align 0\n\t.word\t"
10021 #undef TARGET_ASM_UNALIGNED_SI_OP
10022 #define TARGET_ASM_UNALIGNED_SI_OP "\t.align 0\n\t.long\t"
10023 #undef TARGET_ASM_UNALIGNED_DI_OP
10024 #define TARGET_ASM_UNALIGNED_DI_OP "\t.align 0\n\t.quad\t"
10027 #ifdef OBJECT_FORMAT_ELF
10028 #undef TARGET_ASM_SELECT_RTX_SECTION
10029 #define TARGET_ASM_SELECT_RTX_SECTION alpha_elf_select_rtx_section
10032 #undef TARGET_ASM_FUNCTION_END_PROLOGUE
10033 #define TARGET_ASM_FUNCTION_END_PROLOGUE alpha_output_function_end_prologue
10035 #if TARGET_ABI_UNICOSMK
10036 #undef TARGET_ASM_FILE_START
10037 #define TARGET_ASM_FILE_START unicosmk_file_start
10038 #undef TARGET_ASM_FILE_END
10039 #define TARGET_ASM_FILE_END unicosmk_file_end
10041 #undef TARGET_ASM_FILE_START
10042 #define TARGET_ASM_FILE_START alpha_file_start
10043 #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE
10044 #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true
10047 #undef TARGET_SCHED_ADJUST_COST
10048 #define TARGET_SCHED_ADJUST_COST alpha_adjust_cost
10049 #undef TARGET_SCHED_ISSUE_RATE
10050 #define TARGET_SCHED_ISSUE_RATE alpha_issue_rate
10051 #undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE
10052 #define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE \
10053 alpha_use_dfa_pipeline_interface
10054 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10055 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
10056 alpha_multipass_dfa_lookahead
10058 #undef TARGET_HAVE_TLS
10059 #define TARGET_HAVE_TLS HAVE_AS_TLS
10061 #undef TARGET_INIT_BUILTINS
10062 #define TARGET_INIT_BUILTINS alpha_init_builtins
10063 #undef TARGET_EXPAND_BUILTIN
10064 #define TARGET_EXPAND_BUILTIN alpha_expand_builtin
10066 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10067 #define TARGET_FUNCTION_OK_FOR_SIBCALL alpha_function_ok_for_sibcall
10068 #undef TARGET_CANNOT_COPY_INSN_P
10069 #define TARGET_CANNOT_COPY_INSN_P alpha_cannot_copy_insn_p
10072 #undef TARGET_ASM_OUTPUT_MI_THUNK
10073 #define TARGET_ASM_OUTPUT_MI_THUNK alpha_output_mi_thunk_osf
10074 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10075 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
10078 #undef TARGET_RTX_COSTS
10079 #define TARGET_RTX_COSTS alpha_rtx_costs
10080 #undef TARGET_ADDRESS_COST
10081 #define TARGET_ADDRESS_COST hook_int_rtx_0
10083 #undef TARGET_MACHINE_DEPENDENT_REORG
10084 #define TARGET_MACHINE_DEPENDENT_REORG alpha_reorg
10086 struct gcc_target targetm = TARGET_INITIALIZER;
10089 #include "gt-alpha.h"