xref: /netbsd-src/external/gpl3/gcc.old/dist/gcc/config/sparc/sparc.h (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55)
1 /* Definitions of target machine for GNU compiler, for Sun SPARC.
2    Copyright (C) 1987, 1988, 1989, 1992, 1994, 1995, 1996, 1997, 1998, 1999
3    2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
4    Free Software Foundation, Inc.
5    Contributed by Michael Tiemann (tiemann@cygnus.com).
6    64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans,
7    at Cygnus Support.
8 
9 This file is part of GCC.
10 
11 GCC is free software; you can redistribute it and/or modify
12 it under the terms of the GNU General Public License as published by
13 the Free Software Foundation; either version 3, or (at your option)
14 any later version.
15 
16 GCC is distributed in the hope that it will be useful,
17 but WITHOUT ANY WARRANTY; without even the implied warranty of
18 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
19 GNU General Public License for more details.
20 
21 You should have received a copy of the GNU General Public License
22 along with GCC; see the file COPYING3.  If not see
23 <http://www.gnu.org/licenses/>.  */
24 
25 #include "config/vxworks-dummy.h"
26 
27 /* Note that some other tm.h files include this one and then override
28    whatever definitions are necessary.  */
29 
30 /* Define the specific costs for a given cpu */
31 
32 struct processor_costs {
33   /* Integer load */
34   const int int_load;
35 
36   /* Integer signed load */
37   const int int_sload;
38 
39   /* Integer zeroed load */
40   const int int_zload;
41 
42   /* Float load */
43   const int float_load;
44 
45   /* fmov, fneg, fabs */
46   const int float_move;
47 
48   /* fadd, fsub */
49   const int float_plusminus;
50 
51   /* fcmp */
52   const int float_cmp;
53 
54   /* fmov, fmovr */
55   const int float_cmove;
56 
57   /* fmul */
58   const int float_mul;
59 
60   /* fdivs */
61   const int float_div_sf;
62 
63   /* fdivd */
64   const int float_div_df;
65 
66   /* fsqrts */
67   const int float_sqrt_sf;
68 
69   /* fsqrtd */
70   const int float_sqrt_df;
71 
72   /* umul/smul */
73   const int int_mul;
74 
75   /* mulX */
76   const int int_mulX;
77 
78   /* integer multiply cost for each bit set past the most
79      significant 3, so the formula for multiply cost becomes:
80 
81 	if (rs1 < 0)
82 	  highest_bit = highest_clear_bit(rs1);
83 	else
84 	  highest_bit = highest_set_bit(rs1);
85 	if (highest_bit < 3)
86 	  highest_bit = 3;
87 	cost = int_mul{,X} + ((highest_bit - 3) / int_mul_bit_factor);
88 
89      A value of zero indicates that the multiply costs is fixed,
90      and not variable.  */
91   const int int_mul_bit_factor;
92 
93   /* udiv/sdiv */
94   const int int_div;
95 
96   /* divX */
97   const int int_divX;
98 
99   /* movcc, movr */
100   const int int_cmove;
101 
102   /* penalty for shifts, due to scheduling rules etc. */
103   const int shift_penalty;
104 };
105 
106 extern const struct processor_costs *sparc_costs;
107 
108 /* Target CPU builtins.  FIXME: Defining sparc is for the benefit of
109    Solaris only; otherwise just define __sparc__.  Sadly the headers
110    are such a mess there is no Solaris-specific header.  */
111 #define TARGET_CPU_CPP_BUILTINS()		\
112   do						\
113     {						\
114 	builtin_define_std ("sparc");		\
115 	if (TARGET_64BIT)			\
116 	  { 					\
117 	    builtin_assert ("cpu=sparc64");	\
118 	    builtin_assert ("machine=sparc64");	\
119 	  }					\
120 	else					\
121 	  { 					\
122 	    builtin_assert ("cpu=sparc");	\
123 	    builtin_assert ("machine=sparc");	\
124 	  }					\
125     }						\
126   while (0)
127 
128 /* Specify this in a cover file to provide bi-architecture (32/64) support.  */
129 /* #define SPARC_BI_ARCH */
130 
131 /* Macro used later in this file to determine default architecture.  */
132 #define DEFAULT_ARCH32_P ((TARGET_DEFAULT & MASK_64BIT) == 0)
133 
134 /* TARGET_ARCH{32,64} are the main macros to decide which of the two
135    architectures to compile for.  We allow targets to choose compile time or
136    runtime selection.  */
137 #ifdef IN_LIBGCC2
138 #if defined(__sparcv9) || defined(__arch64__)
139 #define TARGET_ARCH32 0
140 #else
141 #define TARGET_ARCH32 1
142 #endif /* sparc64 */
143 #else
144 #ifdef SPARC_BI_ARCH
145 #define TARGET_ARCH32 (! TARGET_64BIT)
146 #else
147 #define TARGET_ARCH32 (DEFAULT_ARCH32_P)
148 #endif /* SPARC_BI_ARCH */
149 #endif /* IN_LIBGCC2 */
150 #define TARGET_ARCH64 (! TARGET_ARCH32)
151 
152 /* Code model selection in 64-bit environment.
153 
154    The machine mode used for addresses is 32-bit wide:
155 
156    TARGET_CM_32:     32-bit address space.
157                      It is the code model used when generating 32-bit code.
158 
159    The machine mode used for addresses is 64-bit wide:
160 
161    TARGET_CM_MEDLOW: 32-bit address space.
162                      The executable must be in the low 32 bits of memory.
163                      This avoids generating %uhi and %ulo terms.  Programs
164                      can be statically or dynamically linked.
165 
166    TARGET_CM_MEDMID: 44-bit address space.
167                      The executable must be in the low 44 bits of memory,
168                      and the %[hml]44 terms are used.  The text and data
169                      segments have a maximum size of 2GB (31-bit span).
170                      The maximum offset from any instruction to the label
171                      _GLOBAL_OFFSET_TABLE_ is 2GB (31-bit span).
172 
173    TARGET_CM_MEDANY: 64-bit address space.
174                      The text and data segments have a maximum size of 2GB
175                      (31-bit span) and may be located anywhere in memory.
176                      The maximum offset from any instruction to the label
177                      _GLOBAL_OFFSET_TABLE_ is 2GB (31-bit span).
178 
179    TARGET_CM_EMBMEDANY: 64-bit address space.
180                      The text and data segments have a maximum size of 2GB
181                      (31-bit span) and may be located anywhere in memory.
182                      The global register %g4 contains the start address of
183                      the data segment.  Programs are statically linked and
184                      PIC is not supported.
185 
186    Different code models are not supported in 32-bit environment.  */
187 
188 enum cmodel {
189   CM_32,
190   CM_MEDLOW,
191   CM_MEDMID,
192   CM_MEDANY,
193   CM_EMBMEDANY
194 };
195 
196 /* One of CM_FOO.  */
197 extern enum cmodel sparc_cmodel;
198 
199 /* V9 code model selection.  */
200 #define TARGET_CM_MEDLOW    (sparc_cmodel == CM_MEDLOW)
201 #define TARGET_CM_MEDMID    (sparc_cmodel == CM_MEDMID)
202 #define TARGET_CM_MEDANY    (sparc_cmodel == CM_MEDANY)
203 #define TARGET_CM_EMBMEDANY (sparc_cmodel == CM_EMBMEDANY)
204 
205 #define SPARC_DEFAULT_CMODEL CM_32
206 
207 /* The SPARC-V9 architecture defines a relaxed memory ordering model (RMO)
208    which requires the following macro to be true if enabled.  Prior to V9,
209    there are no instructions to even talk about memory synchronization.
210    Note that the UltraSPARC III processors don't implement RMO, unlike the
211    UltraSPARC II processors.  Niagara and Niagara-2 do not implement RMO
212    either.
213 
214    Default to false; for example, Solaris never enables RMO, only ever uses
215    total memory ordering (TMO).  */
216 #define SPARC_RELAXED_ORDERING false
217 
218 /* Do not use the .note.GNU-stack convention by default.  */
219 #define NEED_INDICATE_EXEC_STACK 0
220 
221 /* This is call-clobbered in the normal ABI, but is reserved in the
222    home grown (aka upward compatible) embedded ABI.  */
223 #define EMBMEDANY_BASE_REG "%g4"
224 
225 /* Values of TARGET_CPU_DEFAULT, set via -D in the Makefile,
226    and specified by the user via --with-cpu=foo.
227    This specifies the cpu implementation, not the architecture size.  */
228 /* Note that TARGET_CPU_v9 is assumed to start the list of 64-bit
229    capable cpu's.  */
230 #define TARGET_CPU_sparc	0
231 #define TARGET_CPU_v7		0	/* alias for previous */
232 #define TARGET_CPU_sparclet	1
233 #define TARGET_CPU_sparclite	2
234 #define TARGET_CPU_v8		3	/* generic v8 implementation */
235 #define TARGET_CPU_supersparc	4
236 #define TARGET_CPU_hypersparc   5
237 #define TARGET_CPU_sparc86x	6
238 #define TARGET_CPU_sparclite86x	6
239 #define TARGET_CPU_v9		7	/* generic v9 implementation */
240 #define TARGET_CPU_sparcv9	7	/* alias */
241 #define TARGET_CPU_sparc64	7	/* alias */
242 #define TARGET_CPU_ultrasparc	8
243 #define TARGET_CPU_ultrasparc3	9
244 #define TARGET_CPU_niagara	10
245 #define TARGET_CPU_niagara2	11
246 
247 #if TARGET_CPU_DEFAULT == TARGET_CPU_v9 \
248  || TARGET_CPU_DEFAULT == TARGET_CPU_ultrasparc \
249  || TARGET_CPU_DEFAULT == TARGET_CPU_ultrasparc3 \
250  || TARGET_CPU_DEFAULT == TARGET_CPU_niagara \
251  || TARGET_CPU_DEFAULT == TARGET_CPU_niagara2
252 
253 #define CPP_CPU32_DEFAULT_SPEC ""
254 #define ASM_CPU32_DEFAULT_SPEC ""
255 
256 #if TARGET_CPU_DEFAULT == TARGET_CPU_v9
257 /* ??? What does Sun's CC pass?  */
258 #define CPP_CPU64_DEFAULT_SPEC "-D__sparc_v9__"
259 /* ??? It's not clear how other assemblers will handle this, so by default
260    use GAS.  Sun's Solaris assembler recognizes -xarch=v8plus, but this case
261    is handled in sol2.h.  */
262 #define ASM_CPU64_DEFAULT_SPEC "-Av9"
263 #endif
264 #if TARGET_CPU_DEFAULT == TARGET_CPU_ultrasparc
265 #define CPP_CPU64_DEFAULT_SPEC "-D__sparc_v9__"
266 #define ASM_CPU64_DEFAULT_SPEC "-Av9a"
267 #endif
268 #if TARGET_CPU_DEFAULT == TARGET_CPU_ultrasparc3
269 #define CPP_CPU64_DEFAULT_SPEC "-D__sparc_v9__"
270 #define ASM_CPU64_DEFAULT_SPEC "-Av9b"
271 #endif
272 #if TARGET_CPU_DEFAULT == TARGET_CPU_niagara
273 #define CPP_CPU64_DEFAULT_SPEC "-D__sparc_v9__"
274 #define ASM_CPU64_DEFAULT_SPEC "-Av9b"
275 #endif
276 #if TARGET_CPU_DEFAULT == TARGET_CPU_niagara2
277 #define CPP_CPU64_DEFAULT_SPEC "-D__sparc_v9__"
278 #define ASM_CPU64_DEFAULT_SPEC "-Av9b"
279 #endif
280 
281 #else
282 
283 #define CPP_CPU64_DEFAULT_SPEC ""
284 #define ASM_CPU64_DEFAULT_SPEC ""
285 
286 #if TARGET_CPU_DEFAULT == TARGET_CPU_sparc \
287  || TARGET_CPU_DEFAULT == TARGET_CPU_v8
288 #define CPP_CPU32_DEFAULT_SPEC ""
289 #define ASM_CPU32_DEFAULT_SPEC ""
290 #endif
291 
292 #if TARGET_CPU_DEFAULT == TARGET_CPU_sparclet
293 #define CPP_CPU32_DEFAULT_SPEC "-D__sparclet__"
294 #define ASM_CPU32_DEFAULT_SPEC "-Asparclet"
295 #endif
296 
297 #if TARGET_CPU_DEFAULT == TARGET_CPU_sparclite
298 #define CPP_CPU32_DEFAULT_SPEC "-D__sparclite__"
299 #define ASM_CPU32_DEFAULT_SPEC "-Asparclite"
300 #endif
301 
302 #if TARGET_CPU_DEFAULT == TARGET_CPU_supersparc
303 #define CPP_CPU32_DEFAULT_SPEC "-D__supersparc__ -D__sparc_v8__"
304 #define ASM_CPU32_DEFAULT_SPEC ""
305 #endif
306 
307 #if TARGET_CPU_DEFAULT == TARGET_CPU_hypersparc
308 #define CPP_CPU32_DEFAULT_SPEC "-D__hypersparc__ -D__sparc_v8__"
309 #define ASM_CPU32_DEFAULT_SPEC ""
310 #endif
311 
312 #if TARGET_CPU_DEFAULT == TARGET_CPU_sparclite86x
313 #define CPP_CPU32_DEFAULT_SPEC "-D__sparclite86x__"
314 #define ASM_CPU32_DEFAULT_SPEC "-Asparclite"
315 #endif
316 
317 #endif
318 
319 #if !defined(CPP_CPU32_DEFAULT_SPEC) || !defined(CPP_CPU64_DEFAULT_SPEC)
320  #error Unrecognized value in TARGET_CPU_DEFAULT.
321 #endif
322 
323 #ifdef SPARC_BI_ARCH
324 
325 #define CPP_CPU_DEFAULT_SPEC \
326 (DEFAULT_ARCH32_P ? "\
327 %{m64:" CPP_CPU64_DEFAULT_SPEC "} \
328 %{!m64:" CPP_CPU32_DEFAULT_SPEC "} \
329 " : "\
330 %{m32:" CPP_CPU32_DEFAULT_SPEC "} \
331 %{!m32:" CPP_CPU64_DEFAULT_SPEC "} \
332 ")
333 #define ASM_CPU_DEFAULT_SPEC \
334 (DEFAULT_ARCH32_P ? "\
335 %{m64:" ASM_CPU64_DEFAULT_SPEC "} \
336 %{!m64:" ASM_CPU32_DEFAULT_SPEC "} \
337 " : "\
338 %{m32:" ASM_CPU32_DEFAULT_SPEC "} \
339 %{!m32:" ASM_CPU64_DEFAULT_SPEC "} \
340 ")
341 
342 #else /* !SPARC_BI_ARCH */
343 
344 #define CPP_CPU_DEFAULT_SPEC (DEFAULT_ARCH32_P ? CPP_CPU32_DEFAULT_SPEC : CPP_CPU64_DEFAULT_SPEC)
345 #define ASM_CPU_DEFAULT_SPEC (DEFAULT_ARCH32_P ? ASM_CPU32_DEFAULT_SPEC : ASM_CPU64_DEFAULT_SPEC)
346 
347 #endif /* !SPARC_BI_ARCH */
348 
349 /* Define macros to distinguish architectures.  */
350 
351 /* Common CPP definitions used by CPP_SPEC amongst the various targets
352    for handling -mcpu=xxx switches.  */
353 #define CPP_CPU_SPEC "\
354 %{msoft-float:-D_SOFT_FLOAT} \
355 %{mcypress:} \
356 %{msparclite:-D__sparclite__} \
357 %{mf930:-D__sparclite__} %{mf934:-D__sparclite__} \
358 %{mv8:-D__sparc_v8__} \
359 %{msupersparc:-D__supersparc__ -D__sparc_v8__} \
360 %{mcpu=sparclet:-D__sparclet__} %{mcpu=tsc701:-D__sparclet__} \
361 %{mcpu=sparclite:-D__sparclite__} \
362 %{mcpu=f930:-D__sparclite__} %{mcpu=f934:-D__sparclite__} \
363 %{mcpu=v8:-D__sparc_v8__} \
364 %{mcpu=supersparc:-D__supersparc__ -D__sparc_v8__} \
365 %{mcpu=hypersparc:-D__hypersparc__ -D__sparc_v8__} \
366 %{mcpu=sparclite86x:-D__sparclite86x__} \
367 %{mcpu=v9:-D__sparc_v9__} \
368 %{mcpu=ultrasparc:-D__sparc_v9__} \
369 %{mcpu=ultrasparc3:-D__sparc_v9__} \
370 %{mcpu=niagara:-D__sparc_v9__} \
371 %{mcpu=niagara2:-D__sparc_v9__} \
372 %{!mcpu*:%{!mcypress:%{!msparclite:%{!mf930:%{!mf934:%{!mv8:%{!msupersparc:%(cpp_cpu_default)}}}}}}} \
373 "
374 #define CPP_ARCH32_SPEC ""
375 #define CPP_ARCH64_SPEC "-D__arch64__"
376 
377 #define CPP_ARCH_DEFAULT_SPEC \
378 (DEFAULT_ARCH32_P ? CPP_ARCH32_SPEC : CPP_ARCH64_SPEC)
379 
380 #define CPP_ARCH_SPEC "\
381 %{m32:%(cpp_arch32)} \
382 %{m64:%(cpp_arch64)} \
383 %{!m32:%{!m64:%(cpp_arch_default)}} \
384 "
385 
386 /* Macros to distinguish endianness.  */
387 #define CPP_ENDIAN_SPEC "\
388 %{mlittle-endian:-D__LITTLE_ENDIAN__} \
389 %{mlittle-endian-data:-D__LITTLE_ENDIAN_DATA__}"
390 
391 /* Macros to distinguish the particular subtarget.  */
392 #define CPP_SUBTARGET_SPEC ""
393 
394 #define CPP_SPEC "%(cpp_cpu) %(cpp_arch) %(cpp_endian) %(cpp_subtarget)"
395 
396 /* Prevent error on `-sun4' and `-target sun4' options.  */
397 /* This used to translate -dalign to -malign, but that is no good
398    because it can't turn off the usual meaning of making debugging dumps.  */
399 /* Translate old style -m<cpu> into new style -mcpu=<cpu>.
400    ??? Delete support for -m<cpu> for 2.9.  */
401 
402 #define CC1_SPEC "\
403 %{sun4:} %{target:} \
404 %{mcypress:-mcpu=cypress} \
405 %{msparclite:-mcpu=sparclite} %{mf930:-mcpu=f930} %{mf934:-mcpu=f934} \
406 %{mv8:-mcpu=v8} %{msupersparc:-mcpu=supersparc} \
407 "
408 
409 /* Override in target specific files.  */
410 #define ASM_CPU_SPEC "\
411 %{mcpu=sparclet:-Asparclet} %{mcpu=tsc701:-Asparclet} \
412 %{msparclite:-Asparclite} \
413 %{mf930:-Asparclite} %{mf934:-Asparclite} \
414 %{mcpu=sparclite:-Asparclite} \
415 %{mcpu=sparclite86x:-Asparclite} \
416 %{mcpu=f930:-Asparclite} %{mcpu=f934:-Asparclite} \
417 %{mcpu=v8:-Av8} \
418 %{mv8plus:-Av8plus} \
419 %{mcpu=v9:-Av9} \
420 %{mcpu=ultrasparc:%{!mv8plus:-Av9a}} \
421 %{mcpu=ultrasparc3:%{!mv8plus:-Av9b}} \
422 %{mcpu=niagara:%{!mv8plus:-Av9b}} \
423 %{mcpu=niagara2:%{!mv8plus:-Av9b}} \
424 %{!mcpu*:%{!mcypress:%{!msparclite:%{!mf930:%{!mf934:%{!mv8:%{!msupersparc:%(asm_cpu_default)}}}}}}} \
425 "
426 
427 /* Word size selection, among other things.
428    This is what GAS uses.  Add %(asm_arch) to ASM_SPEC to enable.  */
429 
430 #define ASM_ARCH32_SPEC "-32"
431 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
432 #define ASM_ARCH64_SPEC "-64 -no-undeclared-regs"
433 #else
434 #define ASM_ARCH64_SPEC "-64"
435 #endif
436 #define ASM_ARCH_DEFAULT_SPEC \
437 (DEFAULT_ARCH32_P ? ASM_ARCH32_SPEC : ASM_ARCH64_SPEC)
438 
439 #define ASM_ARCH_SPEC "\
440 %{m32:%(asm_arch32)} \
441 %{m64:%(asm_arch64)} \
442 %{!m32:%{!m64:%(asm_arch_default)}} \
443 "
444 
445 #ifdef HAVE_AS_RELAX_OPTION
446 #define ASM_RELAX_SPEC "%{!mno-relax:-relax}"
447 #else
448 #define ASM_RELAX_SPEC ""
449 #endif
450 
451 /* Special flags to the Sun-4 assembler when using pipe for input.  */
452 
453 #define ASM_SPEC "\
454 %{R} %{!pg:%{!p:%{fpic|fPIC|fpie|fPIE:-k}}} %{keep-local-as-symbols:-L} \
455 %(asm_cpu) %(asm_relax)"
456 
457 #define AS_NEEDS_DASH_FOR_PIPED_INPUT
458 
459 /* This macro defines names of additional specifications to put in the specs
460    that can be used in various specifications like CC1_SPEC.  Its definition
461    is an initializer with a subgrouping for each command option.
462 
463    Each subgrouping contains a string constant, that defines the
464    specification name, and a string constant that used by the GCC driver
465    program.
466 
467    Do not define this macro if it does not need to do anything.  */
468 
469 #define EXTRA_SPECS \
470   { "cpp_cpu",		CPP_CPU_SPEC },		\
471   { "cpp_cpu_default",	CPP_CPU_DEFAULT_SPEC },	\
472   { "cpp_arch32",	CPP_ARCH32_SPEC },	\
473   { "cpp_arch64",	CPP_ARCH64_SPEC },	\
474   { "cpp_arch_default",	CPP_ARCH_DEFAULT_SPEC },\
475   { "cpp_arch",		CPP_ARCH_SPEC },	\
476   { "cpp_endian",	CPP_ENDIAN_SPEC },	\
477   { "cpp_subtarget",	CPP_SUBTARGET_SPEC },	\
478   { "asm_cpu",		ASM_CPU_SPEC },		\
479   { "asm_cpu_default",	ASM_CPU_DEFAULT_SPEC },	\
480   { "asm_arch32",	ASM_ARCH32_SPEC },	\
481   { "asm_arch64",	ASM_ARCH64_SPEC },	\
482   { "asm_relax",	ASM_RELAX_SPEC },	\
483   { "asm_arch_default",	ASM_ARCH_DEFAULT_SPEC },\
484   { "asm_arch",		ASM_ARCH_SPEC },	\
485   SUBTARGET_EXTRA_SPECS
486 
487 #define SUBTARGET_EXTRA_SPECS
488 
489 /* Because libgcc can generate references back to libc (via .umul etc.) we have
490    to list libc again after the second libgcc.  */
491 #define LINK_GCC_C_SEQUENCE_SPEC "%G %L %G %L"
492 
493 
494 #define PTRDIFF_TYPE (TARGET_ARCH64 ? "long int" : "int")
495 #define SIZE_TYPE (TARGET_ARCH64 ? "long unsigned int" : "unsigned int")
496 
497 /* ??? This should be 32 bits for v9 but what can we do?  */
498 #define WCHAR_TYPE "short unsigned int"
499 #define WCHAR_TYPE_SIZE 16
500 
501 /* Show we can debug even without a frame pointer.  */
502 #define CAN_DEBUG_WITHOUT_FP
503 
504 /* Option handling.  */
505 #define OPTIMIZATION_OPTIONS(LEVEL, SIZE) \
506   sparc_optimization_options ((LEVEL), (SIZE))
507 #define OVERRIDE_OPTIONS  sparc_override_options ()
508 
509 /* Mask of all CPU selection flags.  */
510 #define MASK_ISA \
511 (MASK_V8 + MASK_SPARCLITE + MASK_SPARCLET + MASK_V9 + MASK_DEPRECATED_V8_INSNS)
512 
513 /* TARGET_HARD_MUL: Use hardware multiply instructions but not %y.
514    TARGET_HARD_MUL32: Use hardware multiply instructions with rd %y
515    to get high 32 bits.  False in V8+ or V9 because multiply stores
516    a 64-bit result in a register.  */
517 
518 #define TARGET_HARD_MUL32				\
519   ((TARGET_V8 || TARGET_SPARCLITE			\
520     || TARGET_SPARCLET || TARGET_DEPRECATED_V8_INSNS)	\
521    && ! TARGET_V8PLUS && TARGET_ARCH32)
522 
523 #define TARGET_HARD_MUL					\
524   (TARGET_V8 || TARGET_SPARCLITE || TARGET_SPARCLET	\
525    || TARGET_DEPRECATED_V8_INSNS || TARGET_V8PLUS)
526 
527 /* MASK_APP_REGS must always be the default because that's what
528    FIXED_REGISTERS is set to and -ffixed- is processed before
529    CONDITIONAL_REGISTER_USAGE is called (where we process -mno-app-regs).  */
530 #define TARGET_DEFAULT (MASK_APP_REGS + MASK_FPU)
531 
532 /* Processor type.
533    These must match the values for the cpu attribute in sparc.md.  */
534 enum processor_type {
535   PROCESSOR_V7,
536   PROCESSOR_CYPRESS,
537   PROCESSOR_V8,
538   PROCESSOR_SUPERSPARC,
539   PROCESSOR_SPARCLITE,
540   PROCESSOR_F930,
541   PROCESSOR_F934,
542   PROCESSOR_HYPERSPARC,
543   PROCESSOR_SPARCLITE86X,
544   PROCESSOR_SPARCLET,
545   PROCESSOR_TSC701,
546   PROCESSOR_V9,
547   PROCESSOR_ULTRASPARC,
548   PROCESSOR_ULTRASPARC3,
549   PROCESSOR_NIAGARA,
550   PROCESSOR_NIAGARA2
551 };
552 
553 /* This is set from -m{cpu,tune}=xxx.  */
554 extern enum processor_type sparc_cpu;
555 
556 /* Recast the cpu class to be the cpu attribute.
557    Every file includes us, but not every file includes insn-attr.h.  */
558 #define sparc_cpu_attr ((enum attr_cpu) sparc_cpu)
559 
560 /* Support for a compile-time default CPU, et cetera.  The rules are:
561    --with-cpu is ignored if -mcpu is specified.
562    --with-tune is ignored if -mtune is specified.
563    --with-float is ignored if -mhard-float, -msoft-float, -mfpu, or -mno-fpu
564      are specified.  */
565 #define OPTION_DEFAULT_SPECS \
566   {"cpu", "%{!mcpu=*:-mcpu=%(VALUE)}" }, \
567   {"tune", "%{!mtune=*:-mtune=%(VALUE)}" }, \
568   {"float", "%{!msoft-float:%{!mhard-float:%{!fpu:%{!no-fpu:-m%(VALUE)-float}}}}" }
569 
570 /* sparc_select[0] is reserved for the default cpu.  */
571 struct sparc_cpu_select
572 {
573   const char *string;
574   const char *const name;
575   const int set_tune_p;
576   const int set_arch_p;
577 };
578 
579 extern struct sparc_cpu_select sparc_select[];
580 
581 /* target machine storage layout */
582 
583 /* Define this if most significant bit is lowest numbered
584    in instructions that operate on numbered bit-fields.  */
585 #define BITS_BIG_ENDIAN 1
586 
587 /* Define this if most significant byte of a word is the lowest numbered.  */
588 #define BYTES_BIG_ENDIAN 1
589 
590 /* Define this if most significant word of a multiword number is the lowest
591    numbered.  */
592 #define WORDS_BIG_ENDIAN 1
593 
594 /* Define this to set the endianness to use in libgcc2.c, which can
595    not depend on target_flags.  */
596 #if defined (__LITTLE_ENDIAN__) || defined(__LITTLE_ENDIAN_DATA__)
597 #define LIBGCC2_WORDS_BIG_ENDIAN 0
598 #else
599 #define LIBGCC2_WORDS_BIG_ENDIAN 1
600 #endif
601 
602 #define MAX_BITS_PER_WORD	64
603 
604 /* Width of a word, in units (bytes).  */
605 #define UNITS_PER_WORD		(TARGET_ARCH64 ? 8 : 4)
606 #ifdef IN_LIBGCC2
607 #define MIN_UNITS_PER_WORD	UNITS_PER_WORD
608 #else
609 #define MIN_UNITS_PER_WORD	4
610 #endif
611 
612 #define UNITS_PER_SIMD_WORD(MODE) (TARGET_VIS ? 8 : UNITS_PER_WORD)
613 
614 /* Now define the sizes of the C data types.  */
615 
616 #define SHORT_TYPE_SIZE		16
617 #define INT_TYPE_SIZE		32
618 #define LONG_TYPE_SIZE		(TARGET_ARCH64 ? 64 : 32)
619 #define LONG_LONG_TYPE_SIZE	64
620 #define FLOAT_TYPE_SIZE		32
621 #define DOUBLE_TYPE_SIZE	64
622 
623 /* LONG_DOUBLE_TYPE_SIZE is defined per OS even though the
624    SPARC ABI says that it is 128-bit wide.  */
625 /* #define LONG_DOUBLE_TYPE_SIZE	128 */
626 
627 /* The widest floating-point format really supported by the hardware.  */
628 #define WIDEST_HARDWARE_FP_SIZE 64
629 
630 /* Width in bits of a pointer.  This is the size of ptr_mode.  */
631 #define POINTER_SIZE (TARGET_PTR64 ? 64 : 32)
632 
633 /* This is the machine mode used for addresses.  */
634 #define Pmode (TARGET_ARCH64 ? DImode : SImode)
635 
636 /* If we have to extend pointers (only when TARGET_ARCH64 and not
637    TARGET_PTR64), we want to do it unsigned.   This macro does nothing
638    if ptr_mode and Pmode are the same.  */
639 #define POINTERS_EXTEND_UNSIGNED 1
640 
641 /* Allocation boundary (in *bits*) for storing arguments in argument list.  */
642 #define PARM_BOUNDARY (TARGET_ARCH64 ? 64 : 32)
643 
644 /* Boundary (in *bits*) on which stack pointer should be aligned.  */
645 /* FIXME, this is wrong when TARGET_ARCH64 and TARGET_STACK_BIAS, because
646    then %sp+2047 is 128-bit aligned so %sp is really only byte-aligned.  */
647 #define STACK_BOUNDARY (TARGET_ARCH64 ? 128 : 64)
648 /* Temporary hack until the FIXME above is fixed.  */
649 #define SPARC_STACK_BOUNDARY_HACK (TARGET_ARCH64 && TARGET_STACK_BIAS)
650 
651 /* ALIGN FRAMES on double word boundaries */
652 
653 #define SPARC_STACK_ALIGN(LOC) \
654   (TARGET_ARCH64 ? (((LOC)+15) & ~15) : (((LOC)+7) & ~7))
655 
656 /* Allocation boundary (in *bits*) for the code of a function.  */
657 #define FUNCTION_BOUNDARY 32
658 
659 /* Alignment of field after `int : 0' in a structure.  */
660 #define EMPTY_FIELD_BOUNDARY (TARGET_ARCH64 ? 64 : 32)
661 
662 /* Every structure's size must be a multiple of this.  */
663 #define STRUCTURE_SIZE_BOUNDARY 8
664 
665 /* A bit-field declared as `int' forces `int' alignment for the struct.  */
666 #define PCC_BITFIELD_TYPE_MATTERS 1
667 
668 /* No data type wants to be aligned rounder than this.  */
669 #define BIGGEST_ALIGNMENT (TARGET_ARCH64 ? 128 : 64)
670 
671 /* The best alignment to use in cases where we have a choice.  */
672 #define FASTEST_ALIGNMENT 64
673 
674 /* Define this macro as an expression for the alignment of a structure
675    (given by STRUCT as a tree node) if the alignment computed in the
676    usual way is COMPUTED and the alignment explicitly specified was
677    SPECIFIED.
678 
679    The default is to use SPECIFIED if it is larger; otherwise, use
680    the smaller of COMPUTED and `BIGGEST_ALIGNMENT' */
681 #define ROUND_TYPE_ALIGN(STRUCT, COMPUTED, SPECIFIED)	\
682  (TARGET_FASTER_STRUCTS ?				\
683   ((TREE_CODE (STRUCT) == RECORD_TYPE			\
684     || TREE_CODE (STRUCT) == UNION_TYPE                 \
685     || TREE_CODE (STRUCT) == QUAL_UNION_TYPE)           \
686    && TYPE_FIELDS (STRUCT) != 0                         \
687      ? MAX (MAX ((COMPUTED), (SPECIFIED)), BIGGEST_ALIGNMENT) \
688      : MAX ((COMPUTED), (SPECIFIED)))			\
689    :  MAX ((COMPUTED), (SPECIFIED)))
690 
691 /* Make strings word-aligned so strcpy from constants will be faster.  */
692 #define CONSTANT_ALIGNMENT(EXP, ALIGN)  \
693   ((TREE_CODE (EXP) == STRING_CST	\
694     && (ALIGN) < FASTEST_ALIGNMENT)	\
695    ? FASTEST_ALIGNMENT : (ALIGN))
696 
697 /* Make arrays of chars word-aligned for the same reasons.  */
698 #define DATA_ALIGNMENT(TYPE, ALIGN)		\
699   (TREE_CODE (TYPE) == ARRAY_TYPE		\
700    && TYPE_MODE (TREE_TYPE (TYPE)) == QImode	\
701    && (ALIGN) < FASTEST_ALIGNMENT ? FASTEST_ALIGNMENT : (ALIGN))
702 
703 /* Make local arrays of chars word-aligned for the same reasons.  */
704 #define LOCAL_ALIGNMENT(TYPE, ALIGN) DATA_ALIGNMENT (TYPE, ALIGN)
705 
706 /* Set this nonzero if move instructions will actually fail to work
707    when given unaligned data.  */
708 #define STRICT_ALIGNMENT 1
709 
710 /* Things that must be doubleword aligned cannot go in the text section,
711    because the linker fails to align the text section enough!
712    Put them in the data section.  This macro is only used in this file.  */
713 #define MAX_TEXT_ALIGN 32
714 
715 /* Standard register usage.  */
716 
717 /* Number of actual hardware registers.
718    The hardware registers are assigned numbers for the compiler
719    from 0 to just below FIRST_PSEUDO_REGISTER.
720    All registers that the compiler knows about must be given numbers,
721    even those that are not normally considered general registers.
722 
723    SPARC has 32 integer registers and 32 floating point registers.
724    64-bit SPARC has 32 additional fp regs, but the odd numbered ones are not
725    accessible.  We still account for them to simplify register computations
726    (e.g.: in CLASS_MAX_NREGS).  There are also 4 fp condition code registers, so
727    32+32+32+4 == 100.
728    Register 100 is used as the integer condition code register.
729    Register 101 is used as the soft frame pointer register.  */
730 
731 #define FIRST_PSEUDO_REGISTER 102
732 
733 #define SPARC_FIRST_FP_REG     32
734 /* Additional V9 fp regs.  */
735 #define SPARC_FIRST_V9_FP_REG  64
736 #define SPARC_LAST_V9_FP_REG   95
737 /* V9 %fcc[0123].  V8 uses (figuratively) %fcc0.  */
738 #define SPARC_FIRST_V9_FCC_REG 96
739 #define SPARC_LAST_V9_FCC_REG  99
740 /* V8 fcc reg.  */
741 #define SPARC_FCC_REG 96
742 /* Integer CC reg.  We don't distinguish %icc from %xcc.  */
743 #define SPARC_ICC_REG 100
744 
745 /* Nonzero if REGNO is an fp reg.  */
746 #define SPARC_FP_REG_P(REGNO) \
747 ((REGNO) >= SPARC_FIRST_FP_REG && (REGNO) <= SPARC_LAST_V9_FP_REG)
748 
749 /* Argument passing regs.  */
750 #define SPARC_OUTGOING_INT_ARG_FIRST 8
751 #define SPARC_INCOMING_INT_ARG_FIRST 24
752 #define SPARC_FP_ARG_FIRST           32
753 
754 /* 1 for registers that have pervasive standard uses
755    and are not available for the register allocator.
756 
757    On non-v9 systems:
758    g1 is free to use as temporary.
759    g2-g4 are reserved for applications.  Gcc normally uses them as
760    temporaries, but this can be disabled via the -mno-app-regs option.
761    g5 through g7 are reserved for the operating system.
762 
763    On v9 systems:
764    g1,g5 are free to use as temporaries, and are free to use between calls
765    if the call is to an external function via the PLT.
766    g4 is free to use as a temporary in the non-embedded case.
767    g4 is reserved in the embedded case.
768    g2-g3 are reserved for applications.  Gcc normally uses them as
769    temporaries, but this can be disabled via the -mno-app-regs option.
770    g6-g7 are reserved for the operating system (or application in
771    embedded case).
772    ??? Register 1 is used as a temporary by the 64 bit sethi pattern, so must
773    currently be a fixed register until this pattern is rewritten.
774    Register 1 is also used when restoring call-preserved registers in large
775    stack frames.
776 
777    Registers fixed in arch32 and not arch64 (or vice-versa) are marked in
778    CONDITIONAL_REGISTER_USAGE in order to properly handle -ffixed-.
779 */
780 
781 #define FIXED_REGISTERS  \
782  {1, 0, 2, 2, 2, 2, 1, 1,	\
783   0, 0, 0, 0, 0, 0, 1, 0,	\
784   0, 0, 0, 0, 0, 0, 0, 0,	\
785   0, 0, 0, 0, 0, 0, 1, 1,	\
786 				\
787   0, 0, 0, 0, 0, 0, 0, 0,	\
788   0, 0, 0, 0, 0, 0, 0, 0,	\
789   0, 0, 0, 0, 0, 0, 0, 0,	\
790   0, 0, 0, 0, 0, 0, 0, 0,	\
791 				\
792   0, 0, 0, 0, 0, 0, 0, 0,	\
793   0, 0, 0, 0, 0, 0, 0, 0,	\
794   0, 0, 0, 0, 0, 0, 0, 0,	\
795   0, 0, 0, 0, 0, 0, 0, 0,	\
796 				\
797   0, 0, 0, 0, 0, 1}
798 
799 /* 1 for registers not available across function calls.
800    These must include the FIXED_REGISTERS and also any
801    registers that can be used without being saved.
802    The latter must include the registers where values are returned
803    and the register where structure-value addresses are passed.
804    Aside from that, you can include as many other registers as you like.  */
805 
806 #define CALL_USED_REGISTERS  \
807  {1, 1, 1, 1, 1, 1, 1, 1,	\
808   1, 1, 1, 1, 1, 1, 1, 1,	\
809   0, 0, 0, 0, 0, 0, 0, 0,	\
810   0, 0, 0, 0, 0, 0, 1, 1,	\
811 				\
812   1, 1, 1, 1, 1, 1, 1, 1,	\
813   1, 1, 1, 1, 1, 1, 1, 1,	\
814   1, 1, 1, 1, 1, 1, 1, 1,	\
815   1, 1, 1, 1, 1, 1, 1, 1,	\
816 				\
817   1, 1, 1, 1, 1, 1, 1, 1,	\
818   1, 1, 1, 1, 1, 1, 1, 1,	\
819   1, 1, 1, 1, 1, 1, 1, 1,	\
820   1, 1, 1, 1, 1, 1, 1, 1,	\
821 				\
822   1, 1, 1, 1, 1, 1}
823 
824 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
825    they won't be allocated.  */
826 
827 #define CONDITIONAL_REGISTER_USAGE				\
828 do								\
829   {								\
830     if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)		\
831       {								\
832 	fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1;		\
833 	call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1;		\
834       }								\
835     /* If the user has passed -f{fixed,call-{used,saved}}-g5 */	\
836     /* then honor it.  */					\
837     if (TARGET_ARCH32 && fixed_regs[5])				\
838       fixed_regs[5] = 1;					\
839     else if (TARGET_ARCH64 && fixed_regs[5] == 2)		\
840       fixed_regs[5] = 0;					\
841     if (! TARGET_V9)						\
842       {								\
843 	int regno;						\
844 	for (regno = SPARC_FIRST_V9_FP_REG;			\
845 	     regno <= SPARC_LAST_V9_FP_REG;			\
846 	     regno++)						\
847 	  fixed_regs[regno] = 1;				\
848 	/* %fcc0 is used by v8 and v9.  */			\
849 	for (regno = SPARC_FIRST_V9_FCC_REG + 1;		\
850 	     regno <= SPARC_LAST_V9_FCC_REG;			\
851 	     regno++)						\
852 	  fixed_regs[regno] = 1;				\
853       }								\
854     if (! TARGET_FPU)						\
855       {								\
856 	int regno;						\
857 	for (regno = 32; regno < SPARC_LAST_V9_FCC_REG; regno++) \
858 	  fixed_regs[regno] = 1;				\
859       }								\
860     /* If the user has passed -f{fixed,call-{used,saved}}-g2 */	\
861     /* then honor it.  Likewise with g3 and g4.  */		\
862     if (fixed_regs[2] == 2)					\
863       fixed_regs[2] = ! TARGET_APP_REGS;			\
864     if (fixed_regs[3] == 2)					\
865       fixed_regs[3] = ! TARGET_APP_REGS;			\
866     if (TARGET_ARCH32 && fixed_regs[4] == 2)			\
867       fixed_regs[4] = ! TARGET_APP_REGS;			\
868     else if (TARGET_CM_EMBMEDANY)				\
869       fixed_regs[4] = 1;					\
870     else if (fixed_regs[4] == 2)				\
871       fixed_regs[4] = 0;					\
872   }								\
873 while (0)
874 
875 /* Return number of consecutive hard regs needed starting at reg REGNO
876    to hold something of mode MODE.
877    This is ordinarily the length in words of a value of mode MODE
878    but can be less for certain modes in special long registers.
879 
880    On SPARC, ordinary registers hold 32 bits worth;
881    this means both integer and floating point registers.
882    On v9, integer regs hold 64 bits worth; floating point regs hold
883    32 bits worth (this includes the new fp regs as even the odd ones are
884    included in the hard register count).  */
885 
886 #define HARD_REGNO_NREGS(REGNO, MODE) \
887   (TARGET_ARCH64							\
888    ? ((REGNO) < 32 || (REGNO) == FRAME_POINTER_REGNUM			\
889       ? (GET_MODE_SIZE (MODE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD	\
890       : (GET_MODE_SIZE (MODE) + 3) / 4)					\
891    : ((GET_MODE_SIZE (MODE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD))
892 
893 /* Due to the ARCH64 discrepancy above we must override this next
894    macro too.  */
895 #define REGMODE_NATURAL_SIZE(MODE) \
896   ((TARGET_ARCH64 && FLOAT_MODE_P (MODE)) ? 4 : UNITS_PER_WORD)
897 
898 /* Value is 1 if hard register REGNO can hold a value of machine-mode MODE.
899    See sparc.c for how we initialize this.  */
900 extern const int *hard_regno_mode_classes;
901 extern int sparc_mode_class[];
902 
903 /* ??? Because of the funny way we pass parameters we should allow certain
904    ??? types of float/complex values to be in integer registers during
905    ??? RTL generation.  This only matters on arch32.  */
906 #define HARD_REGNO_MODE_OK(REGNO, MODE) \
907   ((hard_regno_mode_classes[REGNO] & sparc_mode_class[MODE]) != 0)
908 
909 /* Value is 1 if it is OK to rename a hard register FROM to another hard
910    register TO.  We cannot rename %g1 as it may be used before the save
911    register window instruction in the prologue.  */
912 #define HARD_REGNO_RENAME_OK(FROM, TO) ((FROM) != 1)
913 
914 /* Value is 1 if it is a good idea to tie two pseudo registers
915    when one has mode MODE1 and one has mode MODE2.
916    If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
917    for any hard reg, then this must be 0 for correct output.
918 
919    For V9: SFmode can't be combined with other float modes, because they can't
920    be allocated to the %d registers.  Also, DFmode won't fit in odd %f
921    registers, but SFmode will.  */
922 #define MODES_TIEABLE_P(MODE1, MODE2) \
923   ((MODE1) == (MODE2)						\
924    || (GET_MODE_CLASS (MODE1) == GET_MODE_CLASS (MODE2)		\
925        && (! TARGET_V9						\
926 	   || (GET_MODE_CLASS (MODE1) != MODE_FLOAT		\
927 	       || (MODE1 != SFmode && MODE2 != SFmode)))))
928 
929 /* Specify the registers used for certain standard purposes.
930    The values of these macros are register numbers.  */
931 
932 /* Register to use for pushing function arguments.  */
933 #define STACK_POINTER_REGNUM 14
934 
935 /* The stack bias (amount by which the hardware register is offset by).  */
936 #define SPARC_STACK_BIAS ((TARGET_ARCH64 && TARGET_STACK_BIAS) ? 2047 : 0)
937 
938 /* Actual top-of-stack address is 92/176 greater than the contents of the
939    stack pointer register for !v9/v9.  That is:
940    - !v9: 64 bytes for the in and local registers, 4 bytes for structure return
941      address, and 6*4 bytes for the 6 register parameters.
942    - v9: 128 bytes for the in and local registers + 6*8 bytes for the integer
943      parameter regs.  */
944 #define STACK_POINTER_OFFSET (FIRST_PARM_OFFSET(0) + SPARC_STACK_BIAS)
945 
946 /* Base register for access to local variables of the function.  */
947 #define HARD_FRAME_POINTER_REGNUM 30
948 
949 /* The soft frame pointer does not have the stack bias applied.  */
950 #define FRAME_POINTER_REGNUM 101
951 
952 /* Given the stack bias, the stack pointer isn't actually aligned.  */
953 #define INIT_EXPANDERS							 \
954   do {									 \
955     if (crtl->emit.regno_pointer_align && SPARC_STACK_BIAS)	 \
956       {									 \
957 	REGNO_POINTER_ALIGN (STACK_POINTER_REGNUM) = BITS_PER_UNIT;	 \
958 	REGNO_POINTER_ALIGN (HARD_FRAME_POINTER_REGNUM) = BITS_PER_UNIT; \
959       }									 \
960   } while (0)
961 
962 /* Base register for access to arguments of the function.  */
963 #define ARG_POINTER_REGNUM FRAME_POINTER_REGNUM
964 
965 /* Register in which static-chain is passed to a function.  This must
966    not be a register used by the prologue.  */
967 #define STATIC_CHAIN_REGNUM (TARGET_ARCH64 ? 5 : 2)
968 
969 /* Register which holds the global offset table, if any.  */
970 
971 #define GLOBAL_OFFSET_TABLE_REGNUM 23
972 
973 /* Register which holds offset table for position-independent
974    data references.  */
975 
976 #define PIC_OFFSET_TABLE_REGNUM \
977   (flag_pic ? GLOBAL_OFFSET_TABLE_REGNUM : INVALID_REGNUM)
978 
979 /* Pick a default value we can notice from override_options:
980    !v9: Default is on.
981    v9: Default is off.
982    Originally it was -1, but later on the container of options changed to
983    unsigned byte, so we decided to pick 127 as default value, which does
984    reflect an undefined default value in case of 0/1.  */
985 
986 #define DEFAULT_PCC_STRUCT_RETURN 127
987 
988 /* Functions which return large structures get the address
989    to place the wanted value at offset 64 from the frame.
990    Must reserve 64 bytes for the in and local registers.
991    v9: Functions which return large structures get the address to place the
992    wanted value from an invisible first argument.  */
993 #define STRUCT_VALUE_OFFSET 64
994 
995 /* Define the classes of registers for register constraints in the
996    machine description.  Also define ranges of constants.
997 
998    One of the classes must always be named ALL_REGS and include all hard regs.
999    If there is more than one class, another class must be named NO_REGS
1000    and contain no registers.
1001 
1002    The name GENERAL_REGS must be the name of a class (or an alias for
1003    another name such as ALL_REGS).  This is the class of registers
1004    that is allowed by "g" or "r" in a register constraint.
1005    Also, registers outside this class are allocated only when
1006    instructions express preferences for them.
1007 
1008    The classes must be numbered in nondecreasing order; that is,
1009    a larger-numbered class must never be contained completely
1010    in a smaller-numbered class.
1011 
1012    For any two classes, it is very desirable that there be another
1013    class that represents their union.  */
1014 
1015 /* The SPARC has various kinds of registers: general, floating point,
1016    and condition codes [well, it has others as well, but none that we
1017    care directly about].
1018 
1019    For v9 we must distinguish between the upper and lower floating point
1020    registers because the upper ones can't hold SFmode values.
1021    HARD_REGNO_MODE_OK won't help here because reload assumes that register(s)
1022    satisfying a group need for a class will also satisfy a single need for
1023    that class.  EXTRA_FP_REGS is a bit of a misnomer as it covers all 64 fp
1024    regs.
1025 
1026    It is important that one class contains all the general and all the standard
1027    fp regs.  Otherwise find_reg() won't properly allocate int regs for moves,
1028    because reg_class_record() will bias the selection in favor of fp regs,
1029    because reg_class_subunion[GENERAL_REGS][FP_REGS] will yield FP_REGS,
1030    because FP_REGS > GENERAL_REGS.
1031 
1032    It is also important that one class contain all the general and all
1033    the fp regs.  Otherwise when spilling a DFmode reg, it may be from
1034    EXTRA_FP_REGS but find_reloads() may use class
1035    GENERAL_OR_FP_REGS. This will cause allocate_reload_reg() to die
1036    because the compiler thinks it doesn't have a spill reg when in
1037    fact it does.
1038 
1039    v9 also has 4 floating point condition code registers.  Since we don't
1040    have a class that is the union of FPCC_REGS with either of the others,
1041    it is important that it appear first.  Otherwise the compiler will die
1042    trying to compile _fixunsdfsi because fix_truncdfsi2 won't match its
1043    constraints.
1044 
1045    It is important that SPARC_ICC_REG have class NO_REGS.  Otherwise combine
1046    may try to use it to hold an SImode value.  See register_operand.
1047    ??? Should %fcc[0123] be handled similarly?
1048 */
1049 
1050 enum reg_class { NO_REGS, FPCC_REGS, I64_REGS, GENERAL_REGS, FP_REGS,
1051 		 EXTRA_FP_REGS, GENERAL_OR_FP_REGS, GENERAL_OR_EXTRA_FP_REGS,
1052 		 ALL_REGS, LIM_REG_CLASSES };
1053 
1054 #define N_REG_CLASSES (int) LIM_REG_CLASSES
1055 
1056 /* Give names of register classes as strings for dump file.  */
1057 
1058 #define REG_CLASS_NAMES \
1059   { "NO_REGS", "FPCC_REGS", "I64_REGS", "GENERAL_REGS", "FP_REGS",	\
1060      "EXTRA_FP_REGS", "GENERAL_OR_FP_REGS", "GENERAL_OR_EXTRA_FP_REGS",	\
1061      "ALL_REGS" }
1062 
1063 /* Define which registers fit in which classes.
1064    This is an initializer for a vector of HARD_REG_SET
1065    of length N_REG_CLASSES.  */
1066 
1067 #define REG_CLASS_CONTENTS				\
1068   {{0, 0, 0, 0},	/* NO_REGS */			\
1069    {0, 0, 0, 0xf},	/* FPCC_REGS */			\
1070    {0xffff, 0, 0, 0},	/* I64_REGS */			\
1071    {-1, 0, 0, 0x20},	/* GENERAL_REGS */		\
1072    {0, -1, 0, 0},	/* FP_REGS */			\
1073    {0, -1, -1, 0},	/* EXTRA_FP_REGS */		\
1074    {-1, -1, 0, 0x20},	/* GENERAL_OR_FP_REGS */	\
1075    {-1, -1, -1, 0x20},	/* GENERAL_OR_EXTRA_FP_REGS */	\
1076    {-1, -1, -1, 0x3f}}	/* ALL_REGS */
1077 
1078 /* The same information, inverted:
1079    Return the class number of the smallest class containing
1080    reg number REGNO.  This could be a conditional expression
1081    or could index an array.  */
1082 
1083 extern enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
1084 
1085 #define REGNO_REG_CLASS(REGNO) sparc_regno_reg_class[(REGNO)]
1086 
1087 /* The following macro defines cover classes for Integrated Register
1088    Allocator.  Cover classes is a set of non-intersected register
1089    classes covering all hard registers used for register allocation
1090    purpose.  Any move between two registers of a cover class should be
1091    cheaper than load or store of the registers.  The macro value is
1092    array of register classes with LIM_REG_CLASSES used as the end
1093    marker.  */
1094 
1095 #define IRA_COVER_CLASSES						     \
1096 {									     \
1097   GENERAL_REGS, EXTRA_FP_REGS, FPCC_REGS, LIM_REG_CLASSES		     \
1098 }
1099 
1100 /* Defines invalid mode changes.  Borrowed from pa64-regs.h.
1101 
1102    SImode loads to floating-point registers are not zero-extended.
1103    The definition for LOAD_EXTEND_OP specifies that integer loads
1104    narrower than BITS_PER_WORD will be zero-extended.  As a result,
1105    we inhibit changes from SImode unless they are to a mode that is
1106    identical in size.  */
1107 
1108 #define CANNOT_CHANGE_MODE_CLASS(FROM, TO, CLASS)		\
1109   (TARGET_ARCH64						\
1110    && (FROM) == SImode						\
1111    && GET_MODE_SIZE (FROM) != GET_MODE_SIZE (TO)		\
1112    ? reg_classes_intersect_p (CLASS, FP_REGS) : 0)
1113 
1114 /* This is the order in which to allocate registers normally.
1115 
1116    We put %f0-%f7 last among the float registers, so as to make it more
1117    likely that a pseudo-register which dies in the float return register
1118    area will get allocated to the float return register, thus saving a move
1119    instruction at the end of the function.
1120 
1121    Similarly for integer return value registers.
1122 
1123    We know in this case that we will not end up with a leaf function.
1124 
1125    The register allocator is given the global and out registers first
1126    because these registers are call clobbered and thus less useful to
1127    global register allocation.
1128 
1129    Next we list the local and in registers.  They are not call clobbered
1130    and thus very useful for global register allocation.  We list the input
1131    registers before the locals so that it is more likely the incoming
1132    arguments received in those registers can just stay there and not be
1133    reloaded.  */
1134 
1135 #define REG_ALLOC_ORDER \
1136 { 1, 2, 3, 4, 5, 6, 7,			/* %g1-%g7 */	\
1137   13, 12, 11, 10, 9, 8, 		/* %o5-%o0 */	\
1138   15,					/* %o7 */	\
1139   16, 17, 18, 19, 20, 21, 22, 23,	/* %l0-%l7 */ 	\
1140   29, 28, 27, 26, 25, 24, 31,		/* %i5-%i0,%i7 */\
1141   40, 41, 42, 43, 44, 45, 46, 47,	/* %f8-%f15 */  \
1142   48, 49, 50, 51, 52, 53, 54, 55,	/* %f16-%f23 */ \
1143   56, 57, 58, 59, 60, 61, 62, 63,	/* %f24-%f31 */ \
1144   64, 65, 66, 67, 68, 69, 70, 71,	/* %f32-%f39 */ \
1145   72, 73, 74, 75, 76, 77, 78, 79,	/* %f40-%f47 */ \
1146   80, 81, 82, 83, 84, 85, 86, 87,	/* %f48-%f55 */ \
1147   88, 89, 90, 91, 92, 93, 94, 95,	/* %f56-%f63 */ \
1148   39, 38, 37, 36, 35, 34, 33, 32,	/* %f7-%f0 */   \
1149   96, 97, 98, 99,			/* %fcc0-3 */   \
1150   100, 0, 14, 30, 101}			/* %icc, %g0, %o6, %i6, %sfp */
1151 
1152 /* This is the order in which to allocate registers for
1153    leaf functions.  If all registers can fit in the global and
1154    output registers, then we have the possibility of having a leaf
1155    function.
1156 
1157    The macro actually mentioned the input registers first,
1158    because they get renumbered into the output registers once
1159    we know really do have a leaf function.
1160 
1161    To be more precise, this register allocation order is used
1162    when %o7 is found to not be clobbered right before register
1163    allocation.  Normally, the reason %o7 would be clobbered is
1164    due to a call which could not be transformed into a sibling
1165    call.
1166 
1167    As a consequence, it is possible to use the leaf register
1168    allocation order and not end up with a leaf function.  We will
1169    not get suboptimal register allocation in that case because by
1170    definition of being potentially leaf, there were no function
1171    calls.  Therefore, allocation order within the local register
1172    window is not critical like it is when we do have function calls.  */
1173 
1174 #define REG_LEAF_ALLOC_ORDER \
1175 { 1, 2, 3, 4, 5, 6, 7, 			/* %g1-%g7 */	\
1176   29, 28, 27, 26, 25, 24,		/* %i5-%i0 */	\
1177   15,					/* %o7 */	\
1178   13, 12, 11, 10, 9, 8,			/* %o5-%o0 */	\
1179   16, 17, 18, 19, 20, 21, 22, 23,	/* %l0-%l7 */	\
1180   40, 41, 42, 43, 44, 45, 46, 47,	/* %f8-%f15 */	\
1181   48, 49, 50, 51, 52, 53, 54, 55,	/* %f16-%f23 */	\
1182   56, 57, 58, 59, 60, 61, 62, 63,	/* %f24-%f31 */	\
1183   64, 65, 66, 67, 68, 69, 70, 71,	/* %f32-%f39 */	\
1184   72, 73, 74, 75, 76, 77, 78, 79,	/* %f40-%f47 */	\
1185   80, 81, 82, 83, 84, 85, 86, 87,	/* %f48-%f55 */	\
1186   88, 89, 90, 91, 92, 93, 94, 95,	/* %f56-%f63 */	\
1187   39, 38, 37, 36, 35, 34, 33, 32,	/* %f7-%f0 */	\
1188   96, 97, 98, 99,			/* %fcc0-3 */	\
1189   100, 0, 14, 30, 31, 101}		/* %icc, %g0, %o6, %i6, %i7, %sfp */
1190 
1191 #define ORDER_REGS_FOR_LOCAL_ALLOC order_regs_for_local_alloc ()
1192 
1193 extern char sparc_leaf_regs[];
1194 #define LEAF_REGISTERS sparc_leaf_regs
1195 
1196 extern char leaf_reg_remap[];
1197 #define LEAF_REG_REMAP(REGNO) (leaf_reg_remap[REGNO])
1198 
1199 /* The class value for index registers, and the one for base regs.  */
1200 #define INDEX_REG_CLASS GENERAL_REGS
1201 #define BASE_REG_CLASS GENERAL_REGS
1202 
1203 /* Local macro to handle the two v9 classes of FP regs.  */
1204 #define FP_REG_CLASS_P(CLASS) ((CLASS) == FP_REGS || (CLASS) == EXTRA_FP_REGS)
1205 
1206 /* Predicates for 10-bit, 11-bit and 13-bit signed constants.  */
1207 #define SPARC_SIMM10_P(X) ((unsigned HOST_WIDE_INT) (X) + 0x200 < 0x400)
1208 #define SPARC_SIMM11_P(X) ((unsigned HOST_WIDE_INT) (X) + 0x400 < 0x800)
1209 #define SPARC_SIMM13_P(X) ((unsigned HOST_WIDE_INT) (X) + 0x1000 < 0x2000)
1210 
1211 /* 10- and 11-bit immediates are only used for a few specific insns.
1212    SMALL_INT is used throughout the port so we continue to use it.  */
1213 #define SMALL_INT(X) (SPARC_SIMM13_P (INTVAL (X)))
1214 
1215 /* Predicate for constants that can be loaded with a sethi instruction.
1216    This is the general, 64-bit aware, bitwise version that ensures that
1217    only constants whose representation fits in the mask
1218 
1219      0x00000000fffffc00
1220 
1221    are accepted.  It will reject, for example, negative SImode constants
1222    on 64-bit hosts, so correct handling is to mask the value beforehand
1223    according to the mode of the instruction.  */
1224 #define SPARC_SETHI_P(X) \
1225   (((unsigned HOST_WIDE_INT) (X) \
1226     & ((unsigned HOST_WIDE_INT) 0x3ff - GET_MODE_MASK (SImode) - 1)) == 0)
1227 
1228 /* Version of the above predicate for SImode constants and below.  */
1229 #define SPARC_SETHI32_P(X) \
1230   (SPARC_SETHI_P ((unsigned HOST_WIDE_INT) (X) & GET_MODE_MASK (SImode)))
1231 
1232 /* Given an rtx X being reloaded into a reg required to be
1233    in class CLASS, return the class of reg to actually use.
1234    In general this is just CLASS; but on some machines
1235    in some cases it is preferable to use a more restrictive class.  */
1236 /* - We can't load constants into FP registers.
1237    - We can't load FP constants into integer registers when soft-float,
1238      because there is no soft-float pattern with a r/F constraint.
1239    - We can't load FP constants into integer registers for TFmode unless
1240      it is 0.0L, because there is no movtf pattern with a r/F constraint.
1241    - Try and reload integer constants (symbolic or otherwise) back into
1242      registers directly, rather than having them dumped to memory.  */
1243 
1244 #define PREFERRED_RELOAD_CLASS(X,CLASS)			\
1245   (CONSTANT_P (X)					\
1246    ? ((FP_REG_CLASS_P (CLASS)				\
1247        || (CLASS) == GENERAL_OR_FP_REGS			\
1248        || (CLASS) == GENERAL_OR_EXTRA_FP_REGS		\
1249        || (GET_MODE_CLASS (GET_MODE (X)) == MODE_FLOAT	\
1250 	   && ! TARGET_FPU)				\
1251        || (GET_MODE (X) == TFmode			\
1252 	   && ! const_zero_operand (X, TFmode)))	\
1253       ? NO_REGS						\
1254       : (!FP_REG_CLASS_P (CLASS)			\
1255          && GET_MODE_CLASS (GET_MODE (X)) == MODE_INT)	\
1256       ? GENERAL_REGS					\
1257       : (CLASS))					\
1258    : (CLASS))
1259 
1260 /* Return the register class of a scratch register needed to load IN into
1261    a register of class CLASS in MODE.
1262 
1263    We need a temporary when loading/storing a HImode/QImode value
1264    between memory and the FPU registers.  This can happen when combine puts
1265    a paradoxical subreg in a float/fix conversion insn.
1266 
1267    We need a temporary when loading/storing a DFmode value between
1268    unaligned memory and the upper FPU registers.  */
1269 
1270 #define SECONDARY_INPUT_RELOAD_CLASS(CLASS, MODE, IN)		\
1271   ((FP_REG_CLASS_P (CLASS)					\
1272     && ((MODE) == HImode || (MODE) == QImode)			\
1273     && (GET_CODE (IN) == MEM					\
1274         || ((GET_CODE (IN) == REG || GET_CODE (IN) == SUBREG)	\
1275             && true_regnum (IN) == -1)))			\
1276    ? GENERAL_REGS						\
1277    : ((CLASS) == EXTRA_FP_REGS && (MODE) == DFmode		\
1278       && GET_CODE (IN) == MEM && TARGET_ARCH32			\
1279       && ! mem_min_alignment ((IN), 8))				\
1280      ? FP_REGS							\
1281      : (((TARGET_CM_MEDANY					\
1282 	  && symbolic_operand ((IN), (MODE)))			\
1283 	 || (TARGET_CM_EMBMEDANY				\
1284 	     && text_segment_operand ((IN), (MODE))))		\
1285 	&& !flag_pic)						\
1286        ? GENERAL_REGS						\
1287        : NO_REGS)
1288 
1289 #define SECONDARY_OUTPUT_RELOAD_CLASS(CLASS, MODE, IN)		\
1290   ((FP_REG_CLASS_P (CLASS)					\
1291      && ((MODE) == HImode || (MODE) == QImode)			\
1292      && (GET_CODE (IN) == MEM					\
1293          || ((GET_CODE (IN) == REG || GET_CODE (IN) == SUBREG)	\
1294              && true_regnum (IN) == -1)))			\
1295    ? GENERAL_REGS						\
1296    : ((CLASS) == EXTRA_FP_REGS && (MODE) == DFmode		\
1297       && GET_CODE (IN) == MEM && TARGET_ARCH32			\
1298       && ! mem_min_alignment ((IN), 8))				\
1299      ? FP_REGS							\
1300      : (((TARGET_CM_MEDANY					\
1301 	  && symbolic_operand ((IN), (MODE)))			\
1302 	 || (TARGET_CM_EMBMEDANY				\
1303 	     && text_segment_operand ((IN), (MODE))))		\
1304 	&& !flag_pic)						\
1305        ? GENERAL_REGS						\
1306        : NO_REGS)
1307 
1308 /* On SPARC it is not possible to directly move data between
1309    GENERAL_REGS and FP_REGS.  */
1310 #define SECONDARY_MEMORY_NEEDED(CLASS1, CLASS2, MODE) \
1311   (FP_REG_CLASS_P (CLASS1) != FP_REG_CLASS_P (CLASS2))
1312 
1313 /* Get_secondary_mem widens its argument to BITS_PER_WORD which loses on v9
1314    because the movsi and movsf patterns don't handle r/f moves.
1315    For v8 we copy the default definition.  */
1316 #define SECONDARY_MEMORY_NEEDED_MODE(MODE) \
1317   (TARGET_ARCH64						\
1318    ? (GET_MODE_BITSIZE (MODE) < 32				\
1319       ? mode_for_size (32, GET_MODE_CLASS (MODE), 0)		\
1320       : MODE)							\
1321    : (GET_MODE_BITSIZE (MODE) < BITS_PER_WORD			\
1322       ? mode_for_size (BITS_PER_WORD, GET_MODE_CLASS (MODE), 0)	\
1323       : MODE))
1324 
1325 /* Return the maximum number of consecutive registers
1326    needed to represent mode MODE in a register of class CLASS.  */
1327 /* On SPARC, this is the size of MODE in words.  */
1328 #define CLASS_MAX_NREGS(CLASS, MODE)	\
1329   (FP_REG_CLASS_P (CLASS) ? (GET_MODE_SIZE (MODE) + 3) / 4 \
1330    : (GET_MODE_SIZE (MODE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
1331 
1332 /* Stack layout; function entry, exit and calling.  */
1333 
1334 /* Define this if pushing a word on the stack
1335    makes the stack pointer a smaller address.  */
1336 #define STACK_GROWS_DOWNWARD
1337 
1338 /* Define this to nonzero if the nominal address of the stack frame
1339    is at the high-address end of the local variables;
1340    that is, each additional local variable allocated
1341    goes at a more negative offset in the frame.  */
1342 #define FRAME_GROWS_DOWNWARD 1
1343 
1344 /* Offset within stack frame to start allocating local variables at.
1345    If FRAME_GROWS_DOWNWARD, this is the offset to the END of the
1346    first local allocated.  Otherwise, it is the offset to the BEGINNING
1347    of the first local allocated.  */
1348 #define STARTING_FRAME_OFFSET 0
1349 
1350 /* Offset of first parameter from the argument pointer register value.
1351    !v9: This is 64 for the ins and locals, plus 4 for the struct-return reg
1352    even if this function isn't going to use it.
1353    v9: This is 128 for the ins and locals.  */
1354 #define FIRST_PARM_OFFSET(FNDECL) \
1355   (TARGET_ARCH64 ? 16 * UNITS_PER_WORD : STRUCT_VALUE_OFFSET + UNITS_PER_WORD)
1356 
1357 /* Offset from the argument pointer register value to the CFA.
1358    This is different from FIRST_PARM_OFFSET because the register window
1359    comes between the CFA and the arguments.  */
1360 #define ARG_POINTER_CFA_OFFSET(FNDECL)  0
1361 
1362 /* When a parameter is passed in a register, stack space is still
1363    allocated for it.
1364    !v9: All 6 possible integer registers have backing store allocated.
1365    v9: Only space for the arguments passed is allocated.  */
1366 /* ??? Ideally, we'd use zero here (as the minimum), but zero has special
1367    meaning to the backend.  Further, we need to be able to detect if a
1368    varargs/unprototyped function is called, as they may want to spill more
1369    registers than we've provided space.  Ugly, ugly.  So for now we retain
1370    all 6 slots even for v9.  */
1371 #define REG_PARM_STACK_SPACE(DECL) (6 * UNITS_PER_WORD)
1372 
1373 /* Definitions for register elimination.  */
1374 
1375 #define ELIMINABLE_REGS \
1376   {{ FRAME_POINTER_REGNUM, STACK_POINTER_REGNUM}, \
1377    { FRAME_POINTER_REGNUM, HARD_FRAME_POINTER_REGNUM} }
1378 
1379 /* We always pretend that this is a leaf function because if it's not,
1380    there's no point in trying to eliminate the frame pointer.  If it
1381    is a leaf function, we guessed right!  */
1382 #define INITIAL_ELIMINATION_OFFSET(FROM, TO, OFFSET) 			\
1383   do {									\
1384     if ((TO) == STACK_POINTER_REGNUM)					\
1385       (OFFSET) = sparc_compute_frame_size (get_frame_size (), 1);	\
1386     else								\
1387       (OFFSET) = 0;							\
1388     (OFFSET) += SPARC_STACK_BIAS;					\
1389   } while (0)
1390 
1391 /* Keep the stack pointer constant throughout the function.
1392    This is both an optimization and a necessity: longjmp
1393    doesn't behave itself when the stack pointer moves within
1394    the function!  */
1395 #define ACCUMULATE_OUTGOING_ARGS 1
1396 
1397 /* Value is the number of bytes of arguments automatically
1398    popped when returning from a subroutine call.
1399    FUNDECL is the declaration node of the function (as a tree),
1400    FUNTYPE is the data type of the function (as a tree),
1401    or for a library call it is an identifier node for the subroutine name.
1402    SIZE is the number of bytes of arguments passed on the stack.  */
1403 
1404 #define RETURN_POPS_ARGS(FUNDECL,FUNTYPE,SIZE) 0
1405 
1406 /* Define this macro if the target machine has "register windows".  This
1407    C expression returns the register number as seen by the called function
1408    corresponding to register number OUT as seen by the calling function.
1409    Return OUT if register number OUT is not an outbound register.  */
1410 
1411 #define INCOMING_REGNO(OUT) \
1412  (((OUT) < 8 || (OUT) > 15) ? (OUT) : (OUT) + 16)
1413 
1414 /* Define this macro if the target machine has "register windows".  This
1415    C expression returns the register number as seen by the calling function
1416    corresponding to register number IN as seen by the called function.
1417    Return IN if register number IN is not an inbound register.  */
1418 
1419 #define OUTGOING_REGNO(IN) \
1420  (((IN) < 24 || (IN) > 31) ? (IN) : (IN) - 16)
1421 
1422 /* Define this macro if the target machine has register windows.  This
1423    C expression returns true if the register is call-saved but is in the
1424    register window.  */
1425 
1426 #define LOCAL_REGNO(REGNO) \
1427   ((REGNO) >= 16 && (REGNO) <= 31)
1428 
1429 /* Define how to find the value returned by a function.
1430    VALTYPE is the data type of the value (as a tree).
1431    If the precise function being called is known, FUNC is its FUNCTION_DECL;
1432    otherwise, FUNC is 0.  */
1433 
1434 /* On SPARC the value is found in the first "output" register.  */
1435 
1436 #define FUNCTION_VALUE(VALTYPE, FUNC) \
1437   function_value ((VALTYPE), TYPE_MODE (VALTYPE), 1)
1438 
1439 /* But the called function leaves it in the first "input" register.  */
1440 
1441 #define FUNCTION_OUTGOING_VALUE(VALTYPE, FUNC) \
1442   function_value ((VALTYPE), TYPE_MODE (VALTYPE), 0)
1443 
1444 /* Define how to find the value returned by a library function
1445    assuming the value has mode MODE.  */
1446 
1447 #define LIBCALL_VALUE(MODE) \
1448   function_value (NULL_TREE, (MODE), 1)
1449 
1450 /* 1 if N is a possible register number for a function value
1451    as seen by the caller.
1452    On SPARC, the first "output" reg is used for integer values,
1453    and the first floating point register is used for floating point values.  */
1454 
1455 #define FUNCTION_VALUE_REGNO_P(N) ((N) == 8 || (N) == 32)
1456 
1457 /* Define the size of space to allocate for the return value of an
1458    untyped_call.  */
1459 
1460 #define APPLY_RESULT_SIZE (TARGET_ARCH64 ? 24 : 16)
1461 
1462 /* 1 if N is a possible register number for function argument passing.
1463    On SPARC, these are the "output" registers.  v9 also uses %f0-%f31.  */
1464 
1465 #define FUNCTION_ARG_REGNO_P(N) \
1466 (TARGET_ARCH64 \
1467  ? (((N) >= 8 && (N) <= 13) || ((N) >= 32 && (N) <= 63)) \
1468  : ((N) >= 8 && (N) <= 13))
1469 
1470 /* Define a data type for recording info about an argument list
1471    during the scan of that argument list.  This data type should
1472    hold all necessary information about the function itself
1473    and about the args processed so far, enough to enable macros
1474    such as FUNCTION_ARG to determine where the next arg should go.
1475 
1476    On SPARC (!v9), this is a single integer, which is a number of words
1477    of arguments scanned so far (including the invisible argument,
1478    if any, which holds the structure-value-address).
1479    Thus 7 or more means all following args should go on the stack.
1480 
1481    For v9, we also need to know whether a prototype is present.  */
1482 
1483 struct sparc_args {
1484   int words;       /* number of words passed so far */
1485   int prototype_p; /* nonzero if a prototype is present */
1486   int libcall_p;   /* nonzero if a library call */
1487 };
1488 #define CUMULATIVE_ARGS struct sparc_args
1489 
1490 /* Initialize a variable CUM of type CUMULATIVE_ARGS
1491    for a call to a function whose data type is FNTYPE.
1492    For a library call, FNTYPE is 0.  */
1493 
1494 #define INIT_CUMULATIVE_ARGS(CUM, FNTYPE, LIBNAME, FNDECL, N_NAMED_ARGS) \
1495 init_cumulative_args (& (CUM), (FNTYPE), (LIBNAME), (FNDECL));
1496 
1497 /* Update the data in CUM to advance over an argument
1498    of mode MODE and data type TYPE.
1499    TYPE is null for libcalls where that information may not be available.  */
1500 
1501 #define FUNCTION_ARG_ADVANCE(CUM, MODE, TYPE, NAMED) \
1502 function_arg_advance (& (CUM), (MODE), (TYPE), (NAMED))
1503 
1504 /* Determine where to put an argument to a function.
1505    Value is zero to push the argument on the stack,
1506    or a hard register in which to store the argument.
1507 
1508    MODE is the argument's machine mode.
1509    TYPE is the data type of the argument (as a tree).
1510     This is null for libcalls where that information may
1511     not be available.
1512    CUM is a variable of type CUMULATIVE_ARGS which gives info about
1513     the preceding args and about the function being called.
1514    NAMED is nonzero if this argument is a named parameter
1515     (otherwise it is an extra parameter matching an ellipsis).  */
1516 
1517 #define FUNCTION_ARG(CUM, MODE, TYPE, NAMED) \
1518 function_arg (& (CUM), (MODE), (TYPE), (NAMED), 0)
1519 
1520 /* Define where a function finds its arguments.
1521    This is different from FUNCTION_ARG because of register windows.  */
1522 
1523 #define FUNCTION_INCOMING_ARG(CUM, MODE, TYPE, NAMED) \
1524 function_arg (& (CUM), (MODE), (TYPE), (NAMED), 1)
1525 
1526 /* If defined, a C expression which determines whether, and in which direction,
1527    to pad out an argument with extra space.  The value should be of type
1528    `enum direction': either `upward' to pad above the argument,
1529    `downward' to pad below, or `none' to inhibit padding.  */
1530 
1531 #define FUNCTION_ARG_PADDING(MODE, TYPE) \
1532 function_arg_padding ((MODE), (TYPE))
1533 
1534 /* If defined, a C expression that gives the alignment boundary, in bits,
1535    of an argument with the specified mode and type.  If it is not defined,
1536    PARM_BOUNDARY is used for all arguments.
1537    For sparc64, objects requiring 16 byte alignment are passed that way.  */
1538 
1539 #define FUNCTION_ARG_BOUNDARY(MODE, TYPE) \
1540 ((TARGET_ARCH64					\
1541   && (GET_MODE_ALIGNMENT (MODE) == 128		\
1542       || ((TYPE) && TYPE_ALIGN (TYPE) == 128)))	\
1543  ? 128 : PARM_BOUNDARY)
1544 
1545 
1546 /* Generate the special assembly code needed to tell the assembler whatever
1547    it might need to know about the return value of a function.
1548 
1549    For SPARC assemblers, we need to output a .proc pseudo-op which conveys
1550    information to the assembler relating to peephole optimization (done in
1551    the assembler).  */
1552 
1553 #define ASM_DECLARE_RESULT(FILE, RESULT) \
1554   fprintf ((FILE), "\t.proc\t0%lo\n", sparc_type_code (TREE_TYPE (RESULT)))
1555 
1556 /* Output the special assembly code needed to tell the assembler some
1557    register is used as global register variable.
1558 
1559    SPARC 64bit psABI declares registers %g2 and %g3 as application
1560    registers and %g6 and %g7 as OS registers.  Any object using them
1561    should declare (for %g2/%g3 has to, for %g6/%g7 can) that it uses them
1562    and how they are used (scratch or some global variable).
1563    Linker will then refuse to link together objects which use those
1564    registers incompatibly.
1565 
1566    Unless the registers are used for scratch, two different global
1567    registers cannot be declared to the same name, so in the unlikely
1568    case of a global register variable occupying more than one register
1569    we prefix the second and following registers with .gnu.part1. etc.  */
1570 
1571 extern GTY(()) char sparc_hard_reg_printed[8];
1572 
1573 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
1574 #define ASM_DECLARE_REGISTER_GLOBAL(FILE, DECL, REGNO, NAME)		\
1575 do {									\
1576   if (TARGET_ARCH64)							\
1577     {									\
1578       int end = HARD_REGNO_NREGS ((REGNO), DECL_MODE (decl)) + (REGNO); \
1579       int reg;								\
1580       for (reg = (REGNO); reg < 8 && reg < end; reg++)			\
1581 	if ((reg & ~1) == 2 || (reg & ~1) == 6)				\
1582 	  {								\
1583 	    if (reg == (REGNO))						\
1584 	      fprintf ((FILE), "\t.register\t%%g%d, %s\n", reg, (NAME)); \
1585 	    else							\
1586 	      fprintf ((FILE), "\t.register\t%%g%d, .gnu.part%d.%s\n",	\
1587 		       reg, reg - (REGNO), (NAME));			\
1588 	    sparc_hard_reg_printed[reg] = 1;				\
1589 	  }								\
1590     }									\
1591 } while (0)
1592 #endif
1593 
1594 
1595 /* Emit rtl for profiling.  */
1596 #define PROFILE_HOOK(LABEL)   sparc_profile_hook (LABEL)
1597 
1598 /* All the work done in PROFILE_HOOK, but still required.  */
1599 #define FUNCTION_PROFILER(FILE, LABELNO) do { } while (0)
1600 
1601 /* Set the name of the mcount function for the system.  */
1602 #define MCOUNT_FUNCTION "*mcount"
1603 
1604 /* EXIT_IGNORE_STACK should be nonzero if, when returning from a function,
1605    the stack pointer does not matter.  The value is tested only in
1606    functions that have frame pointers.
1607    No definition is equivalent to always zero.  */
1608 
1609 #define EXIT_IGNORE_STACK	\
1610  (get_frame_size () != 0	\
1611   || cfun->calls_alloca || crtl->outgoing_args_size)
1612 
1613 /* Define registers used by the epilogue and return instruction.  */
1614 #define EPILOGUE_USES(REGNO) ((REGNO) == 31 \
1615   || (crtl->calls_eh_return && (REGNO) == 1))
1616 
1617 /* Length in units of the trampoline for entering a nested function.  */
1618 
1619 #define TRAMPOLINE_SIZE (TARGET_ARCH64 ? 32 : 16)
1620 
1621 #define TRAMPOLINE_ALIGNMENT 128 /* 16 bytes */
1622 
1623 /* Generate RTL to flush the register windows so as to make arbitrary frames
1624    available.  */
1625 #define SETUP_FRAME_ADDRESSES()		\
1626   emit_insn (gen_flush_register_windows ())
1627 
1628 /* Given an rtx for the address of a frame,
1629    return an rtx for the address of the word in the frame
1630    that holds the dynamic chain--the previous frame's address.  */
1631 #define DYNAMIC_CHAIN_ADDRESS(frame)	\
1632   plus_constant (frame, 14 * UNITS_PER_WORD + SPARC_STACK_BIAS)
1633 
1634 /* Given an rtx for the frame pointer,
1635    return an rtx for the address of the frame.  */
1636 #define FRAME_ADDR_RTX(frame) plus_constant (frame, SPARC_STACK_BIAS)
1637 
1638 /* The return address isn't on the stack, it is in a register, so we can't
1639    access it from the current frame pointer.  We can access it from the
1640    previous frame pointer though by reading a value from the register window
1641    save area.  */
1642 #define RETURN_ADDR_IN_PREVIOUS_FRAME
1643 
1644 /* This is the offset of the return address to the true next instruction to be
1645    executed for the current function.  */
1646 #define RETURN_ADDR_OFFSET \
1647   (8 + 4 * (! TARGET_ARCH64 && cfun->returns_struct))
1648 
1649 /* The current return address is in %i7.  The return address of anything
1650    farther back is in the register window save area at [%fp+60].  */
1651 /* ??? This ignores the fact that the actual return address is +8 for normal
1652    returns, and +12 for structure returns.  */
1653 #define RETURN_ADDR_RTX(count, frame)		\
1654   ((count == -1)				\
1655    ? gen_rtx_REG (Pmode, 31)			\
1656    : gen_rtx_MEM (Pmode,			\
1657 		  memory_address (Pmode, plus_constant (frame, \
1658 							15 * UNITS_PER_WORD \
1659 							+ SPARC_STACK_BIAS))))
1660 
1661 /* Before the prologue, the return address is %o7 + 8.  OK, sometimes it's
1662    +12, but always using +8 is close enough for frame unwind purposes.
1663    Actually, just using %o7 is close enough for unwinding, but %o7+8
1664    is something you can return to.  */
1665 #define INCOMING_RETURN_ADDR_RTX \
1666   plus_constant (gen_rtx_REG (word_mode, 15), 8)
1667 #define DWARF_FRAME_RETURN_COLUMN	DWARF_FRAME_REGNUM (15)
1668 
1669 /* The offset from the incoming value of %sp to the top of the stack frame
1670    for the current function.  On sparc64, we have to account for the stack
1671    bias if present.  */
1672 #define INCOMING_FRAME_SP_OFFSET SPARC_STACK_BIAS
1673 
1674 /* Describe how we implement __builtin_eh_return.  */
1675 #define EH_RETURN_DATA_REGNO(N) ((N) < 4 ? (N) + 24 : INVALID_REGNUM)
1676 #define EH_RETURN_STACKADJ_RTX	gen_rtx_REG (Pmode, 1)	/* %g1 */
1677 #define EH_RETURN_HANDLER_RTX	gen_rtx_REG (Pmode, 31)	/* %i7 */
1678 
1679 /* Select a format to encode pointers in exception handling data.  CODE
1680    is 0 for data, 1 for code labels, 2 for function pointers.  GLOBAL is
1681    true if the symbol may be affected by dynamic relocations.
1682 
1683    If assembler and linker properly support .uaword %r_disp32(foo),
1684    then use PC relative 32-bit relocations instead of absolute relocs
1685    for shared libraries.  On sparc64, use pc relative 32-bit relocs even
1686    for binaries, to save memory.
1687 
1688    binutils 2.12 would emit a R_SPARC_DISP32 dynamic relocation if the
1689    symbol %r_disp32() is against was not local, but .hidden.  In that
1690    case, we have to use DW_EH_PE_absptr for pic personality.  */
1691 #ifdef HAVE_AS_SPARC_UA_PCREL
1692 #ifdef HAVE_AS_SPARC_UA_PCREL_HIDDEN
1693 #define ASM_PREFERRED_EH_DATA_FORMAT(CODE,GLOBAL)			\
1694   (flag_pic								\
1695    ? (GLOBAL ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | DW_EH_PE_sdata4\
1696    : ((TARGET_ARCH64 && ! GLOBAL)					\
1697       ? (DW_EH_PE_pcrel | DW_EH_PE_sdata4)				\
1698       : DW_EH_PE_absptr))
1699 #else
1700 #define ASM_PREFERRED_EH_DATA_FORMAT(CODE,GLOBAL)			\
1701   (flag_pic								\
1702    ? (GLOBAL ? DW_EH_PE_absptr : (DW_EH_PE_pcrel | DW_EH_PE_sdata4))	\
1703    : ((TARGET_ARCH64 && ! GLOBAL)					\
1704       ? (DW_EH_PE_pcrel | DW_EH_PE_sdata4)				\
1705       : DW_EH_PE_absptr))
1706 #endif
1707 
1708 /* Emit a PC-relative relocation.  */
1709 #define ASM_OUTPUT_DWARF_PCREL(FILE, SIZE, LABEL)	\
1710   do {							\
1711     fputs (integer_asm_op (SIZE, FALSE), FILE);		\
1712     fprintf (FILE, "%%r_disp%d(", SIZE * 8);		\
1713     assemble_name (FILE, LABEL);			\
1714     fputc (')', FILE);					\
1715   } while (0)
1716 #endif
1717 
1718 /* Addressing modes, and classification of registers for them.  */
1719 
1720 /* Macros to check register numbers against specific register classes.  */
1721 
1722 /* These assume that REGNO is a hard or pseudo reg number.
1723    They give nonzero only if REGNO is a hard reg of the suitable class
1724    or a pseudo reg currently allocated to a suitable hard reg.
1725    Since they use reg_renumber, they are safe only once reg_renumber
1726    has been allocated, which happens in local-alloc.c.  */
1727 
1728 #define REGNO_OK_FOR_INDEX_P(REGNO) \
1729 ((REGNO) < 32 || (unsigned) reg_renumber[REGNO] < (unsigned)32	\
1730  || (REGNO) == FRAME_POINTER_REGNUM				\
1731  || reg_renumber[REGNO] == FRAME_POINTER_REGNUM)
1732 
1733 #define REGNO_OK_FOR_BASE_P(REGNO)  REGNO_OK_FOR_INDEX_P (REGNO)
1734 
1735 #define REGNO_OK_FOR_FP_P(REGNO) \
1736   (((unsigned) (REGNO) - 32 < (TARGET_V9 ? (unsigned)64 : (unsigned)32)) \
1737    || ((unsigned) reg_renumber[REGNO] - 32 < (TARGET_V9 ? (unsigned)64 : (unsigned)32)))
1738 #define REGNO_OK_FOR_CCFP_P(REGNO) \
1739  (TARGET_V9 \
1740   && (((unsigned) (REGNO) - 96 < (unsigned)4) \
1741       || ((unsigned) reg_renumber[REGNO] - 96 < (unsigned)4)))
1742 
1743 /* Now macros that check whether X is a register and also,
1744    strictly, whether it is in a specified class.
1745 
1746    These macros are specific to the SPARC, and may be used only
1747    in code for printing assembler insns and in conditions for
1748    define_optimization.  */
1749 
1750 /* 1 if X is an fp register.  */
1751 
1752 #define FP_REG_P(X) (REG_P (X) && REGNO_OK_FOR_FP_P (REGNO (X)))
1753 
1754 /* Is X, a REG, an in or global register?  i.e. is regno 0..7 or 24..31 */
1755 #define IN_OR_GLOBAL_P(X) (REGNO (X) < 8 || (REGNO (X) >= 24 && REGNO (X) <= 31))
1756 
1757 /* Maximum number of registers that can appear in a valid memory address.  */
1758 
1759 #define MAX_REGS_PER_ADDRESS 2
1760 
1761 /* Recognize any constant value that is a valid address.
1762    When PIC, we do not accept an address that would require a scratch reg
1763    to load into a register.  */
1764 
1765 #define CONSTANT_ADDRESS_P(X) constant_address_p (X)
1766 
1767 /* Define this, so that when PIC, reload won't try to reload invalid
1768    addresses which require two reload registers.  */
1769 
1770 #define LEGITIMATE_PIC_OPERAND_P(X) legitimate_pic_operand_p (X)
1771 
1772 /* Nonzero if the constant value X is a legitimate general operand.
1773    Anything can be made to work except floating point constants.
1774    If TARGET_VIS, 0.0 can be made to work as well.  */
1775 
1776 #define LEGITIMATE_CONSTANT_P(X) legitimate_constant_p (X)
1777 
1778 /* The macros REG_OK_FOR..._P assume that the arg is a REG rtx
1779    and check its validity for a certain class.
1780    We have two alternate definitions for each of them.
1781    The usual definition accepts all pseudo regs; the other rejects
1782    them unless they have been allocated suitable hard regs.
1783    The symbol REG_OK_STRICT causes the latter definition to be used.
1784 
1785    Most source files want to accept pseudo regs in the hope that
1786    they will get allocated to the class that the insn wants them to be in.
1787    Source files for reload pass need to be strict.
1788    After reload, it makes no difference, since pseudo regs have
1789    been eliminated by then.  */
1790 
1791 #ifndef REG_OK_STRICT
1792 
1793 /* Nonzero if X is a hard reg that can be used as an index
1794    or if it is a pseudo reg.  */
1795 #define REG_OK_FOR_INDEX_P(X) \
1796   (REGNO (X) < 32				\
1797    || REGNO (X) == FRAME_POINTER_REGNUM		\
1798    || REGNO (X) >= FIRST_PSEUDO_REGISTER)
1799 
1800 /* Nonzero if X is a hard reg that can be used as a base reg
1801    or if it is a pseudo reg.  */
1802 #define REG_OK_FOR_BASE_P(X)  REG_OK_FOR_INDEX_P (X)
1803 
1804 #else
1805 
1806 /* Nonzero if X is a hard reg that can be used as an index.  */
1807 #define REG_OK_FOR_INDEX_P(X) REGNO_OK_FOR_INDEX_P (REGNO (X))
1808 /* Nonzero if X is a hard reg that can be used as a base reg.  */
1809 #define REG_OK_FOR_BASE_P(X) REGNO_OK_FOR_BASE_P (REGNO (X))
1810 
1811 #endif
1812 
1813 /* Should gcc use [%reg+%lo(xx)+offset] addresses?  */
1814 
1815 #ifdef HAVE_AS_OFFSETABLE_LO10
1816 #define USE_AS_OFFSETABLE_LO10 1
1817 #else
1818 #define USE_AS_OFFSETABLE_LO10 0
1819 #endif
1820 
1821 /* On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
1822    ordinarily.  This changes a bit when generating PIC.  The details are
1823    in sparc.c's implementation of TARGET_LEGITIMATE_ADDRESS_P.  */
1824 
1825 #define SYMBOLIC_CONST(X) symbolic_operand (X, VOIDmode)
1826 
1827 #define RTX_OK_FOR_BASE_P(X)						\
1828   ((GET_CODE (X) == REG && REG_OK_FOR_BASE_P (X))			\
1829   || (GET_CODE (X) == SUBREG						\
1830       && GET_CODE (SUBREG_REG (X)) == REG				\
1831       && REG_OK_FOR_BASE_P (SUBREG_REG (X))))
1832 
1833 #define RTX_OK_FOR_INDEX_P(X)						\
1834   ((GET_CODE (X) == REG && REG_OK_FOR_INDEX_P (X))			\
1835   || (GET_CODE (X) == SUBREG						\
1836       && GET_CODE (SUBREG_REG (X)) == REG				\
1837       && REG_OK_FOR_INDEX_P (SUBREG_REG (X))))
1838 
1839 #define RTX_OK_FOR_OFFSET_P(X)						\
1840   (GET_CODE (X) == CONST_INT && INTVAL (X) >= -0x1000 && INTVAL (X) < 0x1000 - 8)
1841 
1842 #define RTX_OK_FOR_OLO10_P(X)						\
1843   (GET_CODE (X) == CONST_INT && INTVAL (X) >= -0x1000 && INTVAL (X) < 0xc00 - 8)
1844 
1845 /* Go to LABEL if ADDR (a legitimate address expression)
1846    has an effect that depends on the machine mode it is used for.
1847 
1848    In PIC mode,
1849 
1850       (mem:HI [%l7+a])
1851 
1852    is not equivalent to
1853 
1854       (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
1855 
1856    because [%l7+a+1] is interpreted as the address of (a+1).  */
1857 
1858 #define GO_IF_MODE_DEPENDENT_ADDRESS(ADDR, LABEL)	\
1859 {							\
1860   if (flag_pic == 1)					\
1861     {							\
1862       if (GET_CODE (ADDR) == PLUS)			\
1863 	{						\
1864 	  rtx op0 = XEXP (ADDR, 0);			\
1865 	  rtx op1 = XEXP (ADDR, 1);			\
1866 	  if (op0 == pic_offset_table_rtx		\
1867 	      && SYMBOLIC_CONST (op1))			\
1868 	    goto LABEL;					\
1869 	}						\
1870     }							\
1871 }
1872 
1873 /* Try a machine-dependent way of reloading an illegitimate address
1874    operand.  If we find one, push the reload and jump to WIN.  This
1875    macro is used in only one place: `find_reloads_address' in reload.c.
1876 
1877    For SPARC 32, we wish to handle addresses by splitting them into
1878    HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
1879    This cuts the number of extra insns by one.
1880 
1881    Do nothing when generating PIC code and the address is a
1882    symbolic operand or requires a scratch register.  */
1883 
1884 #define LEGITIMIZE_RELOAD_ADDRESS(X,MODE,OPNUM,TYPE,IND_LEVELS,WIN)     \
1885 do {                                                                    \
1886   /* Decompose SImode constants into hi+lo_sum.  We do have to 		\
1887      rerecognize what we produce, so be careful.  */			\
1888   if (CONSTANT_P (X)							\
1889       && (MODE != TFmode || TARGET_ARCH64)				\
1890       && GET_MODE (X) == SImode						\
1891       && GET_CODE (X) != LO_SUM && GET_CODE (X) != HIGH			\
1892       && ! (flag_pic							\
1893 	    && (symbolic_operand (X, Pmode)				\
1894 		|| pic_address_needs_scratch (X)))			\
1895       && sparc_cmodel <= CM_MEDLOW)					\
1896     {									\
1897       X = gen_rtx_LO_SUM (GET_MODE (X),					\
1898 			  gen_rtx_HIGH (GET_MODE (X), X), X);		\
1899       push_reload (XEXP (X, 0), NULL_RTX, &XEXP (X, 0), NULL,		\
1900                    BASE_REG_CLASS, GET_MODE (X), VOIDmode, 0, 0,	\
1901                    OPNUM, TYPE);					\
1902       goto WIN;								\
1903     }									\
1904   /* ??? 64-bit reloads.  */						\
1905 } while (0)
1906 
1907 /* Specify the machine mode that this machine uses
1908    for the index in the tablejump instruction.  */
1909 /* If we ever implement any of the full models (such as CM_FULLANY),
1910    this has to be DImode in that case */
1911 #ifdef HAVE_GAS_SUBSECTION_ORDERING
1912 #define CASE_VECTOR_MODE \
1913 (! TARGET_PTR64 ? SImode : flag_pic ? SImode : TARGET_CM_MEDLOW ? SImode : DImode)
1914 #else
1915 /* If assembler does not have working .subsection -1, we use DImode for pic, as otherwise
1916    we have to sign extend which slows things down.  */
1917 #define CASE_VECTOR_MODE \
1918 (! TARGET_PTR64 ? SImode : flag_pic ? DImode : TARGET_CM_MEDLOW ? SImode : DImode)
1919 #endif
1920 
1921 /* Define this as 1 if `char' should by default be signed; else as 0.  */
1922 #define DEFAULT_SIGNED_CHAR 1
1923 
1924 /* Max number of bytes we can move from memory to memory
1925    in one reasonably fast instruction.  */
1926 #define MOVE_MAX 8
1927 
1928 /* If a memory-to-memory move would take MOVE_RATIO or more simple
1929    move-instruction pairs, we will do a movmem or libcall instead.  */
1930 
1931 #define MOVE_RATIO(speed) ((speed) ? 8 : 3)
1932 
1933 /* Define if operations between registers always perform the operation
1934    on the full register even if a narrower mode is specified.  */
1935 #define WORD_REGISTER_OPERATIONS
1936 
1937 /* Define if loading in MODE, an integral mode narrower than BITS_PER_WORD
1938    will either zero-extend or sign-extend.  The value of this macro should
1939    be the code that says which one of the two operations is implicitly
1940    done, UNKNOWN if none.  */
1941 #define LOAD_EXTEND_OP(MODE) ZERO_EXTEND
1942 
1943 /* Nonzero if access to memory by bytes is slow and undesirable.
1944    For RISC chips, it means that access to memory by bytes is no
1945    better than access by words when possible, so grab a whole word
1946    and maybe make use of that.  */
1947 #define SLOW_BYTE_ACCESS 1
1948 
1949 /* Define this to be nonzero if shift instructions ignore all but the low-order
1950    few bits.  */
1951 #define SHIFT_COUNT_TRUNCATED 1
1952 
1953 /* Value is 1 if truncating an integer of INPREC bits to OUTPREC bits
1954    is done just by pretending it is already truncated.  */
1955 #define TRULY_NOOP_TRUNCATION(OUTPREC, INPREC) 1
1956 
1957 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
1958    return the mode to be used for the comparison.  For floating-point,
1959    CCFP[E]mode is used.  CC_NOOVmode should be used when the first operand
1960    is a PLUS, MINUS, NEG, or ASHIFT.  CCmode should be used when no special
1961    processing is needed.  */
1962 #define SELECT_CC_MODE(OP,X,Y)  select_cc_mode ((OP), (X), (Y))
1963 
1964 /* Return nonzero if MODE implies a floating point inequality can be
1965    reversed.  For SPARC this is always true because we have a full
1966    compliment of ordered and unordered comparisons, but until generic
1967    code knows how to reverse it correctly we keep the old definition.  */
1968 #define REVERSIBLE_CC_MODE(MODE) ((MODE) != CCFPEmode && (MODE) != CCFPmode)
1969 
1970 /* A function address in a call instruction for indexing purposes.  */
1971 #define FUNCTION_MODE Pmode
1972 
1973 /* Define this if addresses of constant functions
1974    shouldn't be put through pseudo regs where they can be cse'd.
1975    Desirable on machines where ordinary constants are expensive
1976    but a CALL with constant address is cheap.  */
1977 #define NO_FUNCTION_CSE
1978 
1979 /* alloca should avoid clobbering the old register save area.  */
1980 #define SETJMP_VIA_SAVE_AREA
1981 
1982 /* The _Q_* comparison libcalls return booleans.  */
1983 #define FLOAT_LIB_COMPARE_RETURNS_BOOL(MODE, COMPARISON) ((MODE) == TFmode)
1984 
1985 /* Assume by default that the _Qp_* 64-bit libcalls are implemented such
1986    that the inputs are fully consumed before the output memory is clobbered.  */
1987 
1988 #define TARGET_BUGGY_QP_LIB	0
1989 
1990 /* Assume by default that we do not have the Solaris-specific conversion
1991    routines nor 64-bit integer multiply and divide routines.  */
1992 
1993 #define SUN_CONVERSION_LIBFUNCS 	0
1994 #define DITF_CONVERSION_LIBFUNCS	0
1995 #define SUN_INTEGER_MULTIPLY_64 	0
1996 
1997 /* Compute extra cost of moving data between one register class
1998    and another.  */
1999 #define GENERAL_OR_I64(C) ((C) == GENERAL_REGS || (C) == I64_REGS)
2000 #define REGISTER_MOVE_COST(MODE, CLASS1, CLASS2)		\
2001   (((FP_REG_CLASS_P (CLASS1) && GENERAL_OR_I64 (CLASS2)) \
2002     || (GENERAL_OR_I64 (CLASS1) && FP_REG_CLASS_P (CLASS2)) \
2003     || (CLASS1) == FPCC_REGS || (CLASS2) == FPCC_REGS)		\
2004    ? ((sparc_cpu == PROCESSOR_ULTRASPARC \
2005        || sparc_cpu == PROCESSOR_ULTRASPARC3 \
2006        || sparc_cpu == PROCESSOR_NIAGARA \
2007        || sparc_cpu == PROCESSOR_NIAGARA2) ? 12 : 6) : 2)
2008 
2009 /* Provide the cost of a branch.  For pre-v9 processors we use
2010    a value of 3 to take into account the potential annulling of
2011    the delay slot (which ends up being a bubble in the pipeline slot)
2012    plus a cycle to take into consideration the instruction cache
2013    effects.
2014 
2015    On v9 and later, which have branch prediction facilities, we set
2016    it to the depth of the pipeline as that is the cost of a
2017    mispredicted branch.
2018 
2019    On Niagara, normal branches insert 3 bubbles into the pipe
2020    and annulled branches insert 4 bubbles.
2021 
2022    On Niagara-2, a not-taken branch costs 1 cycle whereas a taken
2023    branch costs 6 cycles.  */
2024 
2025 #define BRANCH_COST(speed_p, predictable_p) \
2026 	((sparc_cpu == PROCESSOR_V9 \
2027 	  || sparc_cpu == PROCESSOR_ULTRASPARC) \
2028 	 ? 7 \
2029          : (sparc_cpu == PROCESSOR_ULTRASPARC3 \
2030             ? 9 \
2031 	 : (sparc_cpu == PROCESSOR_NIAGARA \
2032 	    ? 4 \
2033 	 : (sparc_cpu == PROCESSOR_NIAGARA2 \
2034 	    ? 5 \
2035 	 : 3))))
2036 
2037 /* Control the assembler format that we output.  */
2038 
2039 /* A C string constant describing how to begin a comment in the target
2040    assembler language.  The compiler assumes that the comment will end at
2041    the end of the line.  */
2042 
2043 #define ASM_COMMENT_START "!"
2044 
2045 /* Output to assembler file text saying following lines
2046    may contain character constants, extra white space, comments, etc.  */
2047 
2048 #define ASM_APP_ON ""
2049 
2050 /* Output to assembler file text saying following lines
2051    no longer contain unusual constructs.  */
2052 
2053 #define ASM_APP_OFF ""
2054 
2055 /* How to refer to registers in assembler output.
2056    This sequence is indexed by compiler's hard-register-number (see above).  */
2057 
2058 #define REGISTER_NAMES \
2059 {"%g0", "%g1", "%g2", "%g3", "%g4", "%g5", "%g6", "%g7",		\
2060  "%o0", "%o1", "%o2", "%o3", "%o4", "%o5", "%sp", "%o7",		\
2061  "%l0", "%l1", "%l2", "%l3", "%l4", "%l5", "%l6", "%l7",		\
2062  "%i0", "%i1", "%i2", "%i3", "%i4", "%i5", "%fp", "%i7",		\
2063  "%f0", "%f1", "%f2", "%f3", "%f4", "%f5", "%f6", "%f7",		\
2064  "%f8", "%f9", "%f10", "%f11", "%f12", "%f13", "%f14", "%f15",		\
2065  "%f16", "%f17", "%f18", "%f19", "%f20", "%f21", "%f22", "%f23",	\
2066  "%f24", "%f25", "%f26", "%f27", "%f28", "%f29", "%f30", "%f31",	\
2067  "%f32", "%f33", "%f34", "%f35", "%f36", "%f37", "%f38", "%f39",	\
2068  "%f40", "%f41", "%f42", "%f43", "%f44", "%f45", "%f46", "%f47",	\
2069  "%f48", "%f49", "%f50", "%f51", "%f52", "%f53", "%f54", "%f55",	\
2070  "%f56", "%f57", "%f58", "%f59", "%f60", "%f61", "%f62", "%f63",	\
2071  "%fcc0", "%fcc1", "%fcc2", "%fcc3", "%icc", "%sfp" }
2072 
2073 /* Define additional names for use in asm clobbers and asm declarations.  */
2074 
2075 #define ADDITIONAL_REGISTER_NAMES \
2076 {{"ccr", SPARC_ICC_REG}, {"cc", SPARC_ICC_REG}}
2077 
2078 /* On Sun 4, this limit is 2048.  We use 1000 to be safe, since the length
2079    can run past this up to a continuation point.  Once we used 1500, but
2080    a single entry in C++ can run more than 500 bytes, due to the length of
2081    mangled symbol names.  dbxout.c should really be fixed to do
2082    continuations when they are actually needed instead of trying to
2083    guess...  */
2084 #define DBX_CONTIN_LENGTH 1000
2085 
2086 /* This is how to output a command to make the user-level label named NAME
2087    defined for reference from other files.  */
2088 
2089 /* Globalizing directive for a label.  */
2090 #define GLOBAL_ASM_OP "\t.global "
2091 
2092 /* The prefix to add to user-visible assembler symbols.  */
2093 
2094 #define USER_LABEL_PREFIX "_"
2095 
2096 /* This is how to store into the string LABEL
2097    the symbol_ref name of an internal numbered label where
2098    PREFIX is the class of label and NUM is the number within the class.
2099    This is suitable for output with `assemble_name'.  */
2100 
2101 #define ASM_GENERATE_INTERNAL_LABEL(LABEL,PREFIX,NUM)	\
2102   sprintf ((LABEL), "*%s%ld", (PREFIX), (long)(NUM))
2103 
2104 /* This is how we hook in and defer the case-vector until the end of
2105    the function.  */
2106 #define ASM_OUTPUT_ADDR_VEC(LAB,VEC) \
2107   sparc_defer_case_vector ((LAB),(VEC), 0)
2108 
2109 #define ASM_OUTPUT_ADDR_DIFF_VEC(LAB,VEC) \
2110   sparc_defer_case_vector ((LAB),(VEC), 1)
2111 
2112 /* This is how to output an element of a case-vector that is absolute.  */
2113 
2114 #define ASM_OUTPUT_ADDR_VEC_ELT(FILE, VALUE)  \
2115 do {									\
2116   char label[30];							\
2117   ASM_GENERATE_INTERNAL_LABEL (label, "L", VALUE);			\
2118   if (CASE_VECTOR_MODE == SImode)					\
2119     fprintf (FILE, "\t.word\t");					\
2120   else									\
2121     fprintf (FILE, "\t.xword\t");					\
2122   assemble_name (FILE, label);						\
2123   fputc ('\n', FILE);							\
2124 } while (0)
2125 
2126 /* This is how to output an element of a case-vector that is relative.
2127    (SPARC uses such vectors only when generating PIC.)  */
2128 
2129 #define ASM_OUTPUT_ADDR_DIFF_ELT(FILE, BODY, VALUE, REL)		\
2130 do {									\
2131   char label[30];							\
2132   ASM_GENERATE_INTERNAL_LABEL (label, "L", (VALUE));			\
2133   if (CASE_VECTOR_MODE == SImode)					\
2134     fprintf (FILE, "\t.word\t");					\
2135   else									\
2136     fprintf (FILE, "\t.xword\t");					\
2137   assemble_name (FILE, label);						\
2138   ASM_GENERATE_INTERNAL_LABEL (label, "L", (REL));			\
2139   fputc ('-', FILE);							\
2140   assemble_name (FILE, label);						\
2141   fputc ('\n', FILE);							\
2142 } while (0)
2143 
2144 /* This is what to output before and after case-vector (both
2145    relative and absolute).  If .subsection -1 works, we put case-vectors
2146    at the beginning of the current section.  */
2147 
2148 #ifdef HAVE_GAS_SUBSECTION_ORDERING
2149 
2150 #define ASM_OUTPUT_ADDR_VEC_START(FILE)					\
2151   fprintf(FILE, "\t.subsection\t-1\n")
2152 
2153 #define ASM_OUTPUT_ADDR_VEC_END(FILE)					\
2154   fprintf(FILE, "\t.previous\n")
2155 
2156 #endif
2157 
2158 /* This is how to output an assembler line
2159    that says to advance the location counter
2160    to a multiple of 2**LOG bytes.  */
2161 
2162 #define ASM_OUTPUT_ALIGN(FILE,LOG)	\
2163   if ((LOG) != 0)			\
2164     fprintf (FILE, "\t.align %d\n", (1<<(LOG)))
2165 
2166 #define ASM_OUTPUT_SKIP(FILE,SIZE)  \
2167   fprintf (FILE, "\t.skip "HOST_WIDE_INT_PRINT_UNSIGNED"\n", (SIZE))
2168 
2169 /* This says how to output an assembler line
2170    to define a global common symbol.  */
2171 
2172 #define ASM_OUTPUT_COMMON(FILE, NAME, SIZE, ROUNDED)  \
2173 ( fputs ("\t.common ", (FILE)),		\
2174   assemble_name ((FILE), (NAME)),		\
2175   fprintf ((FILE), ","HOST_WIDE_INT_PRINT_UNSIGNED",\"bss\"\n", (SIZE)))
2176 
2177 /* This says how to output an assembler line to define a local common
2178    symbol.  */
2179 
2180 #define ASM_OUTPUT_ALIGNED_LOCAL(FILE, NAME, SIZE, ALIGNED)		\
2181 ( fputs ("\t.reserve ", (FILE)),					\
2182   assemble_name ((FILE), (NAME)),					\
2183   fprintf ((FILE), ","HOST_WIDE_INT_PRINT_UNSIGNED",\"bss\",%u\n",	\
2184 	   (SIZE), ((ALIGNED) / BITS_PER_UNIT)))
2185 
2186 /* A C statement (sans semicolon) to output to the stdio stream
2187    FILE the assembler definition of uninitialized global DECL named
2188    NAME whose size is SIZE bytes and alignment is ALIGN bytes.
2189    Try to use asm_output_aligned_bss to implement this macro.  */
2190 
2191 #define ASM_OUTPUT_ALIGNED_BSS(FILE, DECL, NAME, SIZE, ALIGN)	\
2192   do {								\
2193     ASM_OUTPUT_ALIGNED_LOCAL (FILE, NAME, SIZE, ALIGN);		\
2194   } while (0)
2195 
2196 #define IDENT_ASM_OP "\t.ident\t"
2197 
2198 /* Output #ident as a .ident.  */
2199 
2200 #define ASM_OUTPUT_IDENT(FILE, NAME) \
2201   fprintf (FILE, "%s\"%s\"\n", IDENT_ASM_OP, NAME);
2202 
2203 /* Prettify the assembly.  */
2204 
2205 extern int sparc_indent_opcode;
2206 
2207 #define ASM_OUTPUT_OPCODE(FILE, PTR)	\
2208   do {					\
2209     if (sparc_indent_opcode)		\
2210       {					\
2211 	putc (' ', FILE);		\
2212 	sparc_indent_opcode = 0;	\
2213       }					\
2214   } while (0)
2215 
2216 #define PRINT_OPERAND_PUNCT_VALID_P(CHAR) \
2217   ((CHAR) == '#' || (CHAR) == '*' || (CHAR) == '('		\
2218    || (CHAR) == ')' || (CHAR) == '_' || (CHAR) == '&')
2219 
2220 /* Print operand X (an rtx) in assembler syntax to file FILE.
2221    CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
2222    For `%' followed by punctuation, CODE is the punctuation and X is null.  */
2223 
2224 #define PRINT_OPERAND(FILE, X, CODE) print_operand (FILE, X, CODE)
2225 
2226 /* Print a memory address as an operand to reference that memory location.  */
2227 
2228 #define PRINT_OPERAND_ADDRESS(FILE, ADDR)  \
2229 { register rtx base, index = 0;					\
2230   int offset = 0;						\
2231   register rtx addr = ADDR;					\
2232   if (GET_CODE (addr) == REG)					\
2233     fputs (reg_names[REGNO (addr)], FILE);			\
2234   else if (GET_CODE (addr) == PLUS)				\
2235     {								\
2236       if (GET_CODE (XEXP (addr, 0)) == CONST_INT)		\
2237 	offset = INTVAL (XEXP (addr, 0)), base = XEXP (addr, 1);\
2238       else if (GET_CODE (XEXP (addr, 1)) == CONST_INT)		\
2239 	offset = INTVAL (XEXP (addr, 1)), base = XEXP (addr, 0);\
2240       else							\
2241 	base = XEXP (addr, 0), index = XEXP (addr, 1);		\
2242       if (GET_CODE (base) == LO_SUM)				\
2243 	{							\
2244 	  gcc_assert (USE_AS_OFFSETABLE_LO10			\
2245 	      	      && TARGET_ARCH64				\
2246 		      && ! TARGET_CM_MEDMID);			\
2247 	  output_operand (XEXP (base, 0), 0);			\
2248 	  fputs ("+%lo(", FILE);				\
2249 	  output_address (XEXP (base, 1));			\
2250 	  fprintf (FILE, ")+%d", offset);			\
2251 	}							\
2252       else							\
2253 	{							\
2254 	  fputs (reg_names[REGNO (base)], FILE);		\
2255 	  if (index == 0)					\
2256 	    fprintf (FILE, "%+d", offset);			\
2257 	  else if (GET_CODE (index) == REG)			\
2258 	    fprintf (FILE, "+%s", reg_names[REGNO (index)]);	\
2259 	  else if (GET_CODE (index) == SYMBOL_REF		\
2260 		   || GET_CODE (index) == LABEL_REF		\
2261 		   || GET_CODE (index) == CONST)		\
2262 	    fputc ('+', FILE), output_addr_const (FILE, index);	\
2263 	  else gcc_unreachable ();				\
2264 	}							\
2265     }								\
2266   else if (GET_CODE (addr) == MINUS				\
2267 	   && GET_CODE (XEXP (addr, 1)) == LABEL_REF)		\
2268     {								\
2269       output_addr_const (FILE, XEXP (addr, 0));			\
2270       fputs ("-(", FILE);					\
2271       output_addr_const (FILE, XEXP (addr, 1));			\
2272       fputs ("-.)", FILE);					\
2273     }								\
2274   else if (GET_CODE (addr) == LO_SUM)				\
2275     {								\
2276       output_operand (XEXP (addr, 0), 0);			\
2277       if (TARGET_CM_MEDMID)					\
2278         fputs ("+%l44(", FILE);					\
2279       else							\
2280         fputs ("+%lo(", FILE);					\
2281       output_address (XEXP (addr, 1));				\
2282       fputc (')', FILE);					\
2283     }								\
2284   else if (flag_pic && GET_CODE (addr) == CONST			\
2285 	   && GET_CODE (XEXP (addr, 0)) == MINUS		\
2286 	   && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST	\
2287 	   && GET_CODE (XEXP (XEXP (XEXP (addr, 0), 1), 0)) == MINUS	\
2288 	   && XEXP (XEXP (XEXP (XEXP (addr, 0), 1), 0), 1) == pc_rtx)	\
2289     {								\
2290       addr = XEXP (addr, 0);					\
2291       output_addr_const (FILE, XEXP (addr, 0));			\
2292       /* Group the args of the second CONST in parenthesis.  */	\
2293       fputs ("-(", FILE);					\
2294       /* Skip past the second CONST--it does nothing for us.  */\
2295       output_addr_const (FILE, XEXP (XEXP (addr, 1), 0));	\
2296       /* Close the parenthesis.  */				\
2297       fputc (')', FILE);					\
2298     }								\
2299   else								\
2300     {								\
2301       output_addr_const (FILE, addr);				\
2302     }								\
2303 }
2304 
2305 /* TLS support defaulting to original Sun flavor.  GNU extensions
2306    must be activated in separate configuration files.  */
2307 #ifdef HAVE_AS_TLS
2308 #define TARGET_TLS 1
2309 #else
2310 #define TARGET_TLS 0
2311 #endif
2312 
2313 #define TARGET_SUN_TLS TARGET_TLS
2314 #define TARGET_GNU_TLS 0
2315 
2316 /* The number of Pmode words for the setjmp buffer.  */
2317 #define JMP_BUF_SIZE 12
2318 
2319 /* We use gcc _mcount for profiling.  */
2320 #define NO_PROFILE_COUNTERS 0
2321