xref: /netbsd-src/external/gpl3/gcc.old/dist/gcc/config/i386/i386.h (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55)
1 /* Definitions of target machine for GCC for IA-32.
2    Copyright (C) 1988, 1992, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
3    2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
4    Free Software Foundation, Inc.
5 
6 This file is part of GCC.
7 
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
12 
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16 GNU General Public License for more details.
17 
18 Under Section 7 of GPL version 3, you are granted additional
19 permissions described in the GCC Runtime Library Exception, version
20 3.1, as published by the Free Software Foundation.
21 
22 You should have received a copy of the GNU General Public License and
23 a copy of the GCC Runtime Library Exception along with this program;
24 see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
25 <http://www.gnu.org/licenses/>.  */
26 
27 /* The purpose of this file is to define the characteristics of the i386,
28    independent of assembler syntax or operating system.
29 
30    Three other files build on this one to describe a specific assembler syntax:
31    bsd386.h, att386.h, and sun386.h.
32 
33    The actual tm.h file for a particular system should include
34    this file, and then the file for the appropriate assembler syntax.
35 
36    Many macros that specify assembler syntax are omitted entirely from
37    this file because they really belong in the files for particular
38    assemblers.  These include RP, IP, LPREFIX, PUT_OP_SIZE, USE_STAR,
39    ADDR_BEG, ADDR_END, PRINT_IREG, PRINT_SCALE, PRINT_B_I_S, and many
40    that start with ASM_ or end in ASM_OP.  */
41 
42 /* Redefines for option macros.  */
43 
44 #define TARGET_64BIT	OPTION_ISA_64BIT
45 #define TARGET_MMX	OPTION_ISA_MMX
46 #define TARGET_3DNOW	OPTION_ISA_3DNOW
47 #define TARGET_3DNOW_A	OPTION_ISA_3DNOW_A
48 #define TARGET_SSE	OPTION_ISA_SSE
49 #define TARGET_SSE2	OPTION_ISA_SSE2
50 #define TARGET_SSE3	OPTION_ISA_SSE3
51 #define TARGET_SSSE3	OPTION_ISA_SSSE3
52 #define TARGET_SSE4_1	OPTION_ISA_SSE4_1
53 #define TARGET_SSE4_2	OPTION_ISA_SSE4_2
54 #define TARGET_AVX	OPTION_ISA_AVX
55 #define TARGET_FMA	OPTION_ISA_FMA
56 #define TARGET_SSE4A	OPTION_ISA_SSE4A
57 #define TARGET_FMA4	OPTION_ISA_FMA4
58 #define TARGET_XOP	OPTION_ISA_XOP
59 #define TARGET_LWP	OPTION_ISA_LWP
60 #define TARGET_ROUND	OPTION_ISA_ROUND
61 #define TARGET_ABM	OPTION_ISA_ABM
62 #define TARGET_POPCNT	OPTION_ISA_POPCNT
63 #define TARGET_SAHF	OPTION_ISA_SAHF
64 #define TARGET_MOVBE	OPTION_ISA_MOVBE
65 #define TARGET_CRC32	OPTION_ISA_CRC32
66 #define TARGET_AES	OPTION_ISA_AES
67 #define TARGET_PCLMUL	OPTION_ISA_PCLMUL
68 #define TARGET_CMPXCHG16B OPTION_ISA_CX16
69 
70 
71 /* SSE4.1 defines round instructions */
72 #define	OPTION_MASK_ISA_ROUND	OPTION_MASK_ISA_SSE4_1
73 #define	OPTION_ISA_ROUND	((ix86_isa_flags & OPTION_MASK_ISA_ROUND) != 0)
74 
75 #include "config/vxworks-dummy.h"
76 
77 /* Algorithm to expand string function with.  */
78 enum stringop_alg
79 {
80    no_stringop,
81    libcall,
82    rep_prefix_1_byte,
83    rep_prefix_4_byte,
84    rep_prefix_8_byte,
85    loop_1_byte,
86    loop,
87    unrolled_loop
88 };
89 
90 #define NAX_STRINGOP_ALGS 4
91 
92 /* Specify what algorithm to use for stringops on known size.
93    When size is unknown, the UNKNOWN_SIZE alg is used.  When size is
94    known at compile time or estimated via feedback, the SIZE array
95    is walked in order until MAX is greater then the estimate (or -1
96    means infinity).  Corresponding ALG is used then.
97    For example initializer:
98     {{256, loop}, {-1, rep_prefix_4_byte}}
99    will use loop for blocks smaller or equal to 256 bytes, rep prefix will
100    be used otherwise.  */
101 struct stringop_algs
102 {
103   const enum stringop_alg unknown_size;
104   const struct stringop_strategy {
105     const int max;
106     const enum stringop_alg alg;
107   } size [NAX_STRINGOP_ALGS];
108 };
109 
110 /* Define the specific costs for a given cpu */
111 
112 struct processor_costs {
113   const int add;		/* cost of an add instruction */
114   const int lea;		/* cost of a lea instruction */
115   const int shift_var;		/* variable shift costs */
116   const int shift_const;	/* constant shift costs */
117   const int mult_init[5];	/* cost of starting a multiply
118 				   in QImode, HImode, SImode, DImode, TImode*/
119   const int mult_bit;		/* cost of multiply per each bit set */
120   const int divide[5];		/* cost of a divide/mod
121 				   in QImode, HImode, SImode, DImode, TImode*/
122   int movsx;			/* The cost of movsx operation.  */
123   int movzx;			/* The cost of movzx operation.  */
124   const int large_insn;		/* insns larger than this cost more */
125   const int move_ratio;		/* The threshold of number of scalar
126 				   memory-to-memory move insns.  */
127   const int movzbl_load;	/* cost of loading using movzbl */
128   const int int_load[3];	/* cost of loading integer registers
129 				   in QImode, HImode and SImode relative
130 				   to reg-reg move (2).  */
131   const int int_store[3];	/* cost of storing integer register
132 				   in QImode, HImode and SImode */
133   const int fp_move;		/* cost of reg,reg fld/fst */
134   const int fp_load[3];		/* cost of loading FP register
135 				   in SFmode, DFmode and XFmode */
136   const int fp_store[3];	/* cost of storing FP register
137 				   in SFmode, DFmode and XFmode */
138   const int mmx_move;		/* cost of moving MMX register.  */
139   const int mmx_load[2];	/* cost of loading MMX register
140 				   in SImode and DImode */
141   const int mmx_store[2];	/* cost of storing MMX register
142 				   in SImode and DImode */
143   const int sse_move;		/* cost of moving SSE register.  */
144   const int sse_load[3];	/* cost of loading SSE register
145 				   in SImode, DImode and TImode*/
146   const int sse_store[3];	/* cost of storing SSE register
147 				   in SImode, DImode and TImode*/
148   const int mmxsse_to_integer;	/* cost of moving mmxsse register to
149 				   integer and vice versa.  */
150   const int l1_cache_size;	/* size of l1 cache, in kilobytes.  */
151   const int l2_cache_size;	/* size of l2 cache, in kilobytes.  */
152   const int prefetch_block;	/* bytes moved to cache for prefetch.  */
153   const int simultaneous_prefetches; /* number of parallel prefetch
154 				   operations.  */
155   const int branch_cost;	/* Default value for BRANCH_COST.  */
156   const int fadd;		/* cost of FADD and FSUB instructions.  */
157   const int fmul;		/* cost of FMUL instruction.  */
158   const int fdiv;		/* cost of FDIV instruction.  */
159   const int fabs;		/* cost of FABS instruction.  */
160   const int fchs;		/* cost of FCHS instruction.  */
161   const int fsqrt;		/* cost of FSQRT instruction.  */
162 				/* Specify what algorithm
163 				   to use for stringops on unknown size.  */
164   struct stringop_algs memcpy[2], memset[2];
165   const int scalar_stmt_cost;   /* Cost of any scalar operation, excluding
166 				   load and store.  */
167   const int scalar_load_cost;   /* Cost of scalar load.  */
168   const int scalar_store_cost;  /* Cost of scalar store.  */
169   const int vec_stmt_cost;      /* Cost of any vector operation, excluding
170                                    load, store, vector-to-scalar and
171                                    scalar-to-vector operation.  */
172   const int vec_to_scalar_cost;    /* Cost of vect-to-scalar operation.  */
173   const int scalar_to_vec_cost;    /* Cost of scalar-to-vector operation.  */
174   const int vec_align_load_cost;   /* Cost of aligned vector load.  */
175   const int vec_unalign_load_cost; /* Cost of unaligned vector load.  */
176   const int vec_store_cost;        /* Cost of vector store.  */
177   const int cond_taken_branch_cost;    /* Cost of taken branch for vectorizer
178 					  cost model.  */
179   const int cond_not_taken_branch_cost;/* Cost of not taken branch for
180 					  vectorizer cost model.  */
181 };
182 
183 extern const struct processor_costs *ix86_cost;
184 extern const struct processor_costs ix86_size_cost;
185 
186 #define ix86_cur_cost() \
187   (optimize_insn_for_size_p () ? &ix86_size_cost: ix86_cost)
188 
189 /* Macros used in the machine description to test the flags.  */
190 
191 /* configure can arrange to make this 2, to force a 486.  */
192 
193 #ifndef TARGET_CPU_DEFAULT
194 #define TARGET_CPU_DEFAULT TARGET_CPU_DEFAULT_generic
195 #endif
196 
197 #ifndef TARGET_FPMATH_DEFAULT
198 #define TARGET_FPMATH_DEFAULT \
199   (TARGET_64BIT && TARGET_SSE ? FPMATH_SSE : FPMATH_387)
200 #endif
201 
202 #define TARGET_FLOAT_RETURNS_IN_80387 TARGET_FLOAT_RETURNS
203 
204 /* 64bit Sledgehammer mode.  For libgcc2 we make sure this is a
205    compile-time constant.  */
206 #ifdef IN_LIBGCC2
207 #undef TARGET_64BIT
208 #ifdef __x86_64__
209 #define TARGET_64BIT 1
210 #else
211 #define TARGET_64BIT 0
212 #endif
213 #else
214 #ifndef TARGET_BI_ARCH
215 #undef TARGET_64BIT
216 #if TARGET_64BIT_DEFAULT
217 #define TARGET_64BIT 1
218 #else
219 #define TARGET_64BIT 0
220 #endif
221 #endif
222 #endif
223 
224 #define HAS_LONG_COND_BRANCH 1
225 #define HAS_LONG_UNCOND_BRANCH 1
226 
227 #define TARGET_386 (ix86_tune == PROCESSOR_I386)
228 #define TARGET_486 (ix86_tune == PROCESSOR_I486)
229 #define TARGET_PENTIUM (ix86_tune == PROCESSOR_PENTIUM)
230 #define TARGET_PENTIUMPRO (ix86_tune == PROCESSOR_PENTIUMPRO)
231 #define TARGET_GEODE (ix86_tune == PROCESSOR_GEODE)
232 #define TARGET_K6 (ix86_tune == PROCESSOR_K6)
233 #define TARGET_ATHLON (ix86_tune == PROCESSOR_ATHLON)
234 #define TARGET_PENTIUM4 (ix86_tune == PROCESSOR_PENTIUM4)
235 #define TARGET_K8 (ix86_tune == PROCESSOR_K8)
236 #define TARGET_ATHLON_K8 (TARGET_K8 || TARGET_ATHLON)
237 #define TARGET_NOCONA (ix86_tune == PROCESSOR_NOCONA)
238 #define TARGET_CORE2 (ix86_tune == PROCESSOR_CORE2)
239 #define TARGET_GENERIC32 (ix86_tune == PROCESSOR_GENERIC32)
240 #define TARGET_GENERIC64 (ix86_tune == PROCESSOR_GENERIC64)
241 #define TARGET_GENERIC (TARGET_GENERIC32 || TARGET_GENERIC64)
242 #define TARGET_AMDFAM10 (ix86_tune == PROCESSOR_AMDFAM10)
243 #define TARGET_ATOM (ix86_tune == PROCESSOR_ATOM)
244 
245 /* Feature tests against the various tunings.  */
246 enum ix86_tune_indices {
247   X86_TUNE_USE_LEAVE,
248   X86_TUNE_PUSH_MEMORY,
249   X86_TUNE_ZERO_EXTEND_WITH_AND,
250   X86_TUNE_UNROLL_STRLEN,
251   X86_TUNE_DEEP_BRANCH_PREDICTION,
252   X86_TUNE_BRANCH_PREDICTION_HINTS,
253   X86_TUNE_DOUBLE_WITH_ADD,
254   X86_TUNE_USE_SAHF,
255   X86_TUNE_MOVX,
256   X86_TUNE_PARTIAL_REG_STALL,
257   X86_TUNE_PARTIAL_FLAG_REG_STALL,
258   X86_TUNE_USE_HIMODE_FIOP,
259   X86_TUNE_USE_SIMODE_FIOP,
260   X86_TUNE_USE_MOV0,
261   X86_TUNE_USE_CLTD,
262   X86_TUNE_USE_XCHGB,
263   X86_TUNE_SPLIT_LONG_MOVES,
264   X86_TUNE_READ_MODIFY_WRITE,
265   X86_TUNE_READ_MODIFY,
266   X86_TUNE_PROMOTE_QIMODE,
267   X86_TUNE_FAST_PREFIX,
268   X86_TUNE_SINGLE_STRINGOP,
269   X86_TUNE_QIMODE_MATH,
270   X86_TUNE_HIMODE_MATH,
271   X86_TUNE_PROMOTE_QI_REGS,
272   X86_TUNE_PROMOTE_HI_REGS,
273   X86_TUNE_ADD_ESP_4,
274   X86_TUNE_ADD_ESP_8,
275   X86_TUNE_SUB_ESP_4,
276   X86_TUNE_SUB_ESP_8,
277   X86_TUNE_INTEGER_DFMODE_MOVES,
278   X86_TUNE_PARTIAL_REG_DEPENDENCY,
279   X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY,
280   X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL,
281   X86_TUNE_SSE_SPLIT_REGS,
282   X86_TUNE_SSE_TYPELESS_STORES,
283   X86_TUNE_SSE_LOAD0_BY_PXOR,
284   X86_TUNE_MEMORY_MISMATCH_STALL,
285   X86_TUNE_PROLOGUE_USING_MOVE,
286   X86_TUNE_EPILOGUE_USING_MOVE,
287   X86_TUNE_SHIFT1,
288   X86_TUNE_USE_FFREEP,
289   X86_TUNE_INTER_UNIT_MOVES,
290   X86_TUNE_INTER_UNIT_CONVERSIONS,
291   X86_TUNE_FOUR_JUMP_LIMIT,
292   X86_TUNE_SCHEDULE,
293   X86_TUNE_USE_BT,
294   X86_TUNE_USE_INCDEC,
295   X86_TUNE_PAD_RETURNS,
296   X86_TUNE_EXT_80387_CONSTANTS,
297   X86_TUNE_SHORTEN_X87_SSE,
298   X86_TUNE_AVOID_VECTOR_DECODE,
299   X86_TUNE_PROMOTE_HIMODE_IMUL,
300   X86_TUNE_SLOW_IMUL_IMM32_MEM,
301   X86_TUNE_SLOW_IMUL_IMM8,
302   X86_TUNE_MOVE_M1_VIA_OR,
303   X86_TUNE_NOT_UNPAIRABLE,
304   X86_TUNE_NOT_VECTORMODE,
305   X86_TUNE_USE_VECTOR_FP_CONVERTS,
306   X86_TUNE_USE_VECTOR_CONVERTS,
307   X86_TUNE_FUSE_CMP_AND_BRANCH,
308   X86_TUNE_OPT_AGU,
309 
310   X86_TUNE_LAST
311 };
312 
313 extern unsigned char ix86_tune_features[X86_TUNE_LAST];
314 
315 #define TARGET_USE_LEAVE	ix86_tune_features[X86_TUNE_USE_LEAVE]
316 #define TARGET_PUSH_MEMORY	ix86_tune_features[X86_TUNE_PUSH_MEMORY]
317 #define TARGET_ZERO_EXTEND_WITH_AND \
318 	ix86_tune_features[X86_TUNE_ZERO_EXTEND_WITH_AND]
319 #define TARGET_UNROLL_STRLEN	ix86_tune_features[X86_TUNE_UNROLL_STRLEN]
320 #define TARGET_DEEP_BRANCH_PREDICTION \
321 	ix86_tune_features[X86_TUNE_DEEP_BRANCH_PREDICTION]
322 #define TARGET_BRANCH_PREDICTION_HINTS \
323 	ix86_tune_features[X86_TUNE_BRANCH_PREDICTION_HINTS]
324 #define TARGET_DOUBLE_WITH_ADD	ix86_tune_features[X86_TUNE_DOUBLE_WITH_ADD]
325 #define TARGET_USE_SAHF		ix86_tune_features[X86_TUNE_USE_SAHF]
326 #define TARGET_MOVX		ix86_tune_features[X86_TUNE_MOVX]
327 #define TARGET_PARTIAL_REG_STALL ix86_tune_features[X86_TUNE_PARTIAL_REG_STALL]
328 #define TARGET_PARTIAL_FLAG_REG_STALL \
329 	ix86_tune_features[X86_TUNE_PARTIAL_FLAG_REG_STALL]
330 #define TARGET_USE_HIMODE_FIOP	ix86_tune_features[X86_TUNE_USE_HIMODE_FIOP]
331 #define TARGET_USE_SIMODE_FIOP	ix86_tune_features[X86_TUNE_USE_SIMODE_FIOP]
332 #define TARGET_USE_MOV0		ix86_tune_features[X86_TUNE_USE_MOV0]
333 #define TARGET_USE_CLTD		ix86_tune_features[X86_TUNE_USE_CLTD]
334 #define TARGET_USE_XCHGB	ix86_tune_features[X86_TUNE_USE_XCHGB]
335 #define TARGET_SPLIT_LONG_MOVES	ix86_tune_features[X86_TUNE_SPLIT_LONG_MOVES]
336 #define TARGET_READ_MODIFY_WRITE ix86_tune_features[X86_TUNE_READ_MODIFY_WRITE]
337 #define TARGET_READ_MODIFY	ix86_tune_features[X86_TUNE_READ_MODIFY]
338 #define TARGET_PROMOTE_QImode	ix86_tune_features[X86_TUNE_PROMOTE_QIMODE]
339 #define TARGET_FAST_PREFIX	ix86_tune_features[X86_TUNE_FAST_PREFIX]
340 #define TARGET_SINGLE_STRINGOP	ix86_tune_features[X86_TUNE_SINGLE_STRINGOP]
341 #define TARGET_QIMODE_MATH	ix86_tune_features[X86_TUNE_QIMODE_MATH]
342 #define TARGET_HIMODE_MATH	ix86_tune_features[X86_TUNE_HIMODE_MATH]
343 #define TARGET_PROMOTE_QI_REGS	ix86_tune_features[X86_TUNE_PROMOTE_QI_REGS]
344 #define TARGET_PROMOTE_HI_REGS	ix86_tune_features[X86_TUNE_PROMOTE_HI_REGS]
345 #define TARGET_ADD_ESP_4	ix86_tune_features[X86_TUNE_ADD_ESP_4]
346 #define TARGET_ADD_ESP_8	ix86_tune_features[X86_TUNE_ADD_ESP_8]
347 #define TARGET_SUB_ESP_4	ix86_tune_features[X86_TUNE_SUB_ESP_4]
348 #define TARGET_SUB_ESP_8	ix86_tune_features[X86_TUNE_SUB_ESP_8]
349 #define TARGET_INTEGER_DFMODE_MOVES \
350 	ix86_tune_features[X86_TUNE_INTEGER_DFMODE_MOVES]
351 #define TARGET_PARTIAL_REG_DEPENDENCY \
352 	ix86_tune_features[X86_TUNE_PARTIAL_REG_DEPENDENCY]
353 #define TARGET_SSE_PARTIAL_REG_DEPENDENCY \
354 	ix86_tune_features[X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY]
355 #define TARGET_SSE_UNALIGNED_MOVE_OPTIMAL \
356 	ix86_tune_features[X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL]
357 #define TARGET_SSE_SPLIT_REGS	ix86_tune_features[X86_TUNE_SSE_SPLIT_REGS]
358 #define TARGET_SSE_TYPELESS_STORES \
359 	ix86_tune_features[X86_TUNE_SSE_TYPELESS_STORES]
360 #define TARGET_SSE_LOAD0_BY_PXOR ix86_tune_features[X86_TUNE_SSE_LOAD0_BY_PXOR]
361 #define TARGET_MEMORY_MISMATCH_STALL \
362 	ix86_tune_features[X86_TUNE_MEMORY_MISMATCH_STALL]
363 #define TARGET_PROLOGUE_USING_MOVE \
364 	ix86_tune_features[X86_TUNE_PROLOGUE_USING_MOVE]
365 #define TARGET_EPILOGUE_USING_MOVE \
366 	ix86_tune_features[X86_TUNE_EPILOGUE_USING_MOVE]
367 #define TARGET_SHIFT1		ix86_tune_features[X86_TUNE_SHIFT1]
368 #define TARGET_USE_FFREEP	ix86_tune_features[X86_TUNE_USE_FFREEP]
369 #define TARGET_INTER_UNIT_MOVES	ix86_tune_features[X86_TUNE_INTER_UNIT_MOVES]
370 #define TARGET_INTER_UNIT_CONVERSIONS\
371 	ix86_tune_features[X86_TUNE_INTER_UNIT_CONVERSIONS]
372 #define TARGET_FOUR_JUMP_LIMIT	ix86_tune_features[X86_TUNE_FOUR_JUMP_LIMIT]
373 #define TARGET_SCHEDULE		ix86_tune_features[X86_TUNE_SCHEDULE]
374 #define TARGET_USE_BT		ix86_tune_features[X86_TUNE_USE_BT]
375 #define TARGET_USE_INCDEC	ix86_tune_features[X86_TUNE_USE_INCDEC]
376 #define TARGET_PAD_RETURNS	ix86_tune_features[X86_TUNE_PAD_RETURNS]
377 #define TARGET_EXT_80387_CONSTANTS \
378 	ix86_tune_features[X86_TUNE_EXT_80387_CONSTANTS]
379 #define TARGET_SHORTEN_X87_SSE	ix86_tune_features[X86_TUNE_SHORTEN_X87_SSE]
380 #define TARGET_AVOID_VECTOR_DECODE \
381 	ix86_tune_features[X86_TUNE_AVOID_VECTOR_DECODE]
382 #define TARGET_TUNE_PROMOTE_HIMODE_IMUL \
383 	ix86_tune_features[X86_TUNE_PROMOTE_HIMODE_IMUL]
384 #define TARGET_SLOW_IMUL_IMM32_MEM \
385 	ix86_tune_features[X86_TUNE_SLOW_IMUL_IMM32_MEM]
386 #define TARGET_SLOW_IMUL_IMM8	ix86_tune_features[X86_TUNE_SLOW_IMUL_IMM8]
387 #define	TARGET_MOVE_M1_VIA_OR	ix86_tune_features[X86_TUNE_MOVE_M1_VIA_OR]
388 #define TARGET_NOT_UNPAIRABLE	ix86_tune_features[X86_TUNE_NOT_UNPAIRABLE]
389 #define TARGET_NOT_VECTORMODE	ix86_tune_features[X86_TUNE_NOT_VECTORMODE]
390 #define TARGET_USE_VECTOR_FP_CONVERTS \
391 	ix86_tune_features[X86_TUNE_USE_VECTOR_FP_CONVERTS]
392 #define TARGET_USE_VECTOR_CONVERTS \
393 	ix86_tune_features[X86_TUNE_USE_VECTOR_CONVERTS]
394 #define TARGET_FUSE_CMP_AND_BRANCH \
395 	ix86_tune_features[X86_TUNE_FUSE_CMP_AND_BRANCH]
396 #define TARGET_OPT_AGU ix86_tune_features[X86_TUNE_OPT_AGU]
397 
398 /* Feature tests against the various architecture variations.  */
399 enum ix86_arch_indices {
400   X86_ARCH_CMOV,
401   X86_ARCH_CMPXCHG,
402   X86_ARCH_CMPXCHG8B,
403   X86_ARCH_XADD,
404   X86_ARCH_BSWAP,
405 
406   X86_ARCH_LAST
407 };
408 
409 extern unsigned char ix86_arch_features[X86_ARCH_LAST];
410 
411 #define TARGET_CMOV		ix86_arch_features[X86_ARCH_CMOV]
412 #define TARGET_CMPXCHG		ix86_arch_features[X86_ARCH_CMPXCHG]
413 #define TARGET_CMPXCHG8B	ix86_arch_features[X86_ARCH_CMPXCHG8B]
414 #define TARGET_XADD		ix86_arch_features[X86_ARCH_XADD]
415 #define TARGET_BSWAP		ix86_arch_features[X86_ARCH_BSWAP]
416 
417 /* For sane SSE instruction set generation we need fcomi instruction.
418    It is safe to enable all CMOVE instructions.  */
419 #define TARGET_CMOVE		(TARGET_CMOV || TARGET_SSE)
420 
421 #define TARGET_FISTTP		(TARGET_SSE3 && TARGET_80387)
422 
423 extern int x86_prefetch_sse;
424 
425 #define TARGET_PREFETCH_SSE	x86_prefetch_sse
426 
427 #define ASSEMBLER_DIALECT	(ix86_asm_dialect)
428 
429 #define TARGET_SSE_MATH		((ix86_fpmath & FPMATH_SSE) != 0)
430 #define TARGET_MIX_SSE_I387 \
431  ((ix86_fpmath & (FPMATH_SSE | FPMATH_387)) == (FPMATH_SSE | FPMATH_387))
432 
433 #define TARGET_GNU_TLS		(ix86_tls_dialect == TLS_DIALECT_GNU)
434 #define TARGET_GNU2_TLS		(ix86_tls_dialect == TLS_DIALECT_GNU2)
435 #define TARGET_ANY_GNU_TLS	(TARGET_GNU_TLS || TARGET_GNU2_TLS)
436 #define TARGET_SUN_TLS		0
437 
438 extern int ix86_isa_flags;
439 
440 #ifndef TARGET_64BIT_DEFAULT
441 #define TARGET_64BIT_DEFAULT 0
442 #endif
443 #ifndef TARGET_TLS_DIRECT_SEG_REFS_DEFAULT
444 #define TARGET_TLS_DIRECT_SEG_REFS_DEFAULT 0
445 #endif
446 
447 /* Fence to use after loop using storent.  */
448 
449 extern tree x86_mfence;
450 #define FENCE_FOLLOWING_MOVNT x86_mfence
451 
452 /* Once GDB has been enhanced to deal with functions without frame
453    pointers, we can change this to allow for elimination of
454    the frame pointer in leaf functions.  */
455 #define TARGET_DEFAULT 0
456 
457 /* Extra bits to force.  */
458 #define TARGET_SUBTARGET_DEFAULT 0
459 #define TARGET_SUBTARGET_ISA_DEFAULT 0
460 
461 /* Extra bits to force on w/ 32-bit mode.  */
462 #define TARGET_SUBTARGET32_DEFAULT 0
463 #define TARGET_SUBTARGET32_ISA_DEFAULT 0
464 
465 /* Extra bits to force on w/ 64-bit mode.  */
466 #define TARGET_SUBTARGET64_DEFAULT 0
467 #define TARGET_SUBTARGET64_ISA_DEFAULT 0
468 
469 /* This is not really a target flag, but is done this way so that
470    it's analogous to similar code for Mach-O on PowerPC.  darwin.h
471    redefines this to 1.  */
472 #define TARGET_MACHO 0
473 
474 /* Likewise, for the Windows 64-bit ABI.  */
475 #define TARGET_64BIT_MS_ABI (TARGET_64BIT && ix86_cfun_abi () == MS_ABI)
476 
477 /* Available call abi.  */
478 enum calling_abi
479 {
480   SYSV_ABI = 0,
481   MS_ABI = 1
482 };
483 
484 /* The abi used by target.  */
485 extern enum calling_abi ix86_abi;
486 
487 /* The default abi used by target.  */
488 #define DEFAULT_ABI SYSV_ABI
489 
490 /* Subtargets may reset this to 1 in order to enable 96-bit long double
491    with the rounding mode forced to 53 bits.  */
492 #define TARGET_96_ROUND_53_LONG_DOUBLE 0
493 
494 /* Sometimes certain combinations of command options do not make
495    sense on a particular target machine.  You can define a macro
496    `OVERRIDE_OPTIONS' to take account of this.  This macro, if
497    defined, is executed once just after all the command options have
498    been parsed.
499 
500    Don't use this macro to turn on various extra optimizations for
501    `-O'.  That is what `OPTIMIZATION_OPTIONS' is for.  */
502 
503 #define OVERRIDE_OPTIONS override_options (true)
504 
505 /* Define this to change the optimizations performed by default.  */
506 #define OPTIMIZATION_OPTIONS(LEVEL, SIZE) \
507   optimization_options ((LEVEL), (SIZE))
508 
509 /* -march=native handling only makes sense with compiler running on
510    an x86 or x86_64 chip.  If changing this condition, also change
511    the condition in driver-i386.c.  */
512 #if defined(__i386__) || defined(__x86_64__)
513 /* In driver-i386.c.  */
514 extern const char *host_detect_local_cpu (int argc, const char **argv);
515 #define EXTRA_SPEC_FUNCTIONS \
516   { "local_cpu_detect", host_detect_local_cpu },
517 #define HAVE_LOCAL_CPU_DETECT
518 #endif
519 
520 #if TARGET_64BIT_DEFAULT
521 #define OPT_ARCH64 "!m32"
522 #define OPT_ARCH32 "m32"
523 #else
524 #define OPT_ARCH64 "m64"
525 #define OPT_ARCH32 "!m64"
526 #endif
527 
528 /* Support for configure-time defaults of some command line options.
529    The order here is important so that -march doesn't squash the
530    tune or cpu values.  */
531 #define OPTION_DEFAULT_SPECS					   \
532   {"tune", "%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}" }, \
533   {"tune_32", "%{" OPT_ARCH32 ":%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}}" }, \
534   {"tune_64", "%{" OPT_ARCH64 ":%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}}" }, \
535   {"cpu", "%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}" },  \
536   {"cpu_32", "%{" OPT_ARCH32 ":%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}}" }, \
537   {"cpu_64", "%{" OPT_ARCH64 ":%{!mtune=*:%{!mcpu=*:%{!march=*:-mtune=%(VALUE)}}}}" }, \
538   {"arch", "%{!march=*:-march=%(VALUE)}"},			   \
539   {"arch_32", "%{" OPT_ARCH32 ":%{!march=*:-march=%(VALUE)}}"},	   \
540   {"arch_64", "%{" OPT_ARCH64 ":%{!march=*:-march=%(VALUE)}}"},
541 
542 /* Specs for the compiler proper */
543 
544 #ifndef CC1_CPU_SPEC
545 #define CC1_CPU_SPEC_1 "\
546 %{mcpu=*:-mtune=%* \
547 %n`-mcpu=' is deprecated. Use `-mtune=' or '-march=' instead.\n} \
548 %<mcpu=* \
549 %{mintel-syntax:-masm=intel \
550 %n`-mintel-syntax' is deprecated. Use `-masm=intel' instead.\n} \
551 %{msse5:-mavx \
552 %n'-msse5' was removed.\n} \
553 %{mno-intel-syntax:-masm=att \
554 %n`-mno-intel-syntax' is deprecated. Use `-masm=att' instead.\n}"
555 
556 #ifndef HAVE_LOCAL_CPU_DETECT
557 #define CC1_CPU_SPEC CC1_CPU_SPEC_1
558 #else
559 #define CC1_CPU_SPEC CC1_CPU_SPEC_1 \
560 "%{march=native:%<march=native %:local_cpu_detect(arch) \
561   %{!mtune=*:%<mtune=native %:local_cpu_detect(tune)}} \
562 %{mtune=native:%<mtune=native %:local_cpu_detect(tune)}"
563 #endif
564 #endif
565 
566 /* Target CPU builtins.  */
567 #define TARGET_CPU_CPP_BUILTINS() ix86_target_macros ()
568 
569 /* Target Pragmas.  */
570 #define REGISTER_TARGET_PRAGMAS() ix86_register_pragmas ()
571 
572 enum target_cpu_default
573 {
574   TARGET_CPU_DEFAULT_generic = 0,
575 
576   TARGET_CPU_DEFAULT_i386,
577   TARGET_CPU_DEFAULT_i486,
578   TARGET_CPU_DEFAULT_pentium,
579   TARGET_CPU_DEFAULT_pentium_mmx,
580   TARGET_CPU_DEFAULT_pentiumpro,
581   TARGET_CPU_DEFAULT_pentium2,
582   TARGET_CPU_DEFAULT_pentium3,
583   TARGET_CPU_DEFAULT_pentium4,
584   TARGET_CPU_DEFAULT_pentium_m,
585   TARGET_CPU_DEFAULT_prescott,
586   TARGET_CPU_DEFAULT_nocona,
587   TARGET_CPU_DEFAULT_core2,
588   TARGET_CPU_DEFAULT_atom,
589 
590   TARGET_CPU_DEFAULT_geode,
591   TARGET_CPU_DEFAULT_k6,
592   TARGET_CPU_DEFAULT_k6_2,
593   TARGET_CPU_DEFAULT_k6_3,
594   TARGET_CPU_DEFAULT_athlon,
595   TARGET_CPU_DEFAULT_athlon_sse,
596   TARGET_CPU_DEFAULT_k8,
597   TARGET_CPU_DEFAULT_amdfam10,
598 
599   TARGET_CPU_DEFAULT_max
600 };
601 
602 #ifndef CC1_SPEC
603 #define CC1_SPEC "%(cc1_cpu) "
604 #endif
605 
606 /* This macro defines names of additional specifications to put in the
607    specs that can be used in various specifications like CC1_SPEC.  Its
608    definition is an initializer with a subgrouping for each command option.
609 
610    Each subgrouping contains a string constant, that defines the
611    specification name, and a string constant that used by the GCC driver
612    program.
613 
614    Do not define this macro if it does not need to do anything.  */
615 
616 #ifndef SUBTARGET_EXTRA_SPECS
617 #define SUBTARGET_EXTRA_SPECS
618 #endif
619 
620 #define EXTRA_SPECS							\
621   { "cc1_cpu",  CC1_CPU_SPEC },						\
622   SUBTARGET_EXTRA_SPECS
623 
624 
625 /* Set the value of FLT_EVAL_METHOD in float.h.  When using only the
626    FPU, assume that the fpcw is set to extended precision; when using
627    only SSE, rounding is correct; when using both SSE and the FPU,
628    the rounding precision is indeterminate, since either may be chosen
629    apparently at random.  */
630 #define TARGET_FLT_EVAL_METHOD \
631   (TARGET_MIX_SSE_I387 ? -1 : TARGET_SSE_MATH ? 0 : 2)
632 
633 /* Whether to allow x87 floating-point arithmetic on MODE (one of
634    SFmode, DFmode and XFmode) in the current excess precision
635    configuration.  */
636 #define X87_ENABLE_ARITH(MODE) \
637   (flag_excess_precision == EXCESS_PRECISION_FAST || (MODE) == XFmode)
638 
639 /* Likewise, whether to allow direct conversions from integer mode
640    IMODE (HImode, SImode or DImode) to MODE.  */
641 #define X87_ENABLE_FLOAT(MODE, IMODE)			\
642   (flag_excess_precision == EXCESS_PRECISION_FAST	\
643    || (MODE) == XFmode					\
644    || ((MODE) == DFmode && (IMODE) == SImode)		\
645    || (IMODE) == HImode)
646 
647 /* target machine storage layout */
648 
649 #define SHORT_TYPE_SIZE 16
650 #define INT_TYPE_SIZE 32
651 #define FLOAT_TYPE_SIZE 32
652 #define LONG_TYPE_SIZE BITS_PER_WORD
653 #define DOUBLE_TYPE_SIZE 64
654 #define LONG_LONG_TYPE_SIZE 64
655 #define LONG_DOUBLE_TYPE_SIZE 80
656 
657 #define WIDEST_HARDWARE_FP_SIZE LONG_DOUBLE_TYPE_SIZE
658 
659 #if defined (TARGET_BI_ARCH) || TARGET_64BIT_DEFAULT
660 #define MAX_BITS_PER_WORD 64
661 #else
662 #define MAX_BITS_PER_WORD 32
663 #endif
664 
665 /* Define this if most significant byte of a word is the lowest numbered.  */
666 /* That is true on the 80386.  */
667 
668 #define BITS_BIG_ENDIAN 0
669 
670 /* Define this if most significant byte of a word is the lowest numbered.  */
671 /* That is not true on the 80386.  */
672 #define BYTES_BIG_ENDIAN 0
673 
674 /* Define this if most significant word of a multiword number is the lowest
675    numbered.  */
676 /* Not true for 80386 */
677 #define WORDS_BIG_ENDIAN 0
678 
679 /* Width of a word, in units (bytes).  */
680 #define UNITS_PER_WORD		(TARGET_64BIT ? 8 : 4)
681 #ifdef IN_LIBGCC2
682 #define MIN_UNITS_PER_WORD	(TARGET_64BIT ? 8 : 4)
683 #else
684 #define MIN_UNITS_PER_WORD	4
685 #endif
686 
687 /* Allocation boundary (in *bits*) for storing arguments in argument list.  */
688 #define PARM_BOUNDARY BITS_PER_WORD
689 
690 /* Boundary (in *bits*) on which stack pointer should be aligned.  */
691 #define STACK_BOUNDARY \
692  (TARGET_64BIT && ix86_abi == MS_ABI ? 128 : BITS_PER_WORD)
693 
694 /* Stack boundary of the main function guaranteed by OS.  */
695 #define MAIN_STACK_BOUNDARY (TARGET_64BIT ? 128 : 32)
696 
697 /* Minimum stack boundary.  */
698 #define MIN_STACK_BOUNDARY (TARGET_64BIT ? 128 : 32)
699 
700 /* Boundary (in *bits*) on which the stack pointer prefers to be
701    aligned; the compiler cannot rely on having this alignment.  */
702 #define PREFERRED_STACK_BOUNDARY ix86_preferred_stack_boundary
703 
704 /* It should be MIN_STACK_BOUNDARY.  But we set it to 128 bits for
705    both 32bit and 64bit, to support codes that need 128 bit stack
706    alignment for SSE instructions, but can't realign the stack.  */
707 #define PREFERRED_STACK_BOUNDARY_DEFAULT 128
708 
709 /* 1 if -mstackrealign should be turned on by default.  It will
710    generate an alternate prologue and epilogue that realigns the
711    runtime stack if nessary.  This supports mixing codes that keep a
712    4-byte aligned stack, as specified by i386 psABI, with codes that
713    need a 16-byte aligned stack, as required by SSE instructions.  */
714 #define STACK_REALIGN_DEFAULT 0
715 
716 /* Boundary (in *bits*) on which the incoming stack is aligned.  */
717 #define INCOMING_STACK_BOUNDARY ix86_incoming_stack_boundary
718 
719 /* Target OS keeps a vector-aligned (128-bit, 16-byte) stack.  This is
720    mandatory for the 64-bit ABI, and may or may not be true for other
721    operating systems.  */
722 #define TARGET_KEEPS_VECTOR_ALIGNED_STACK TARGET_64BIT
723 
724 /* Minimum allocation boundary for the code of a function.  */
725 #define FUNCTION_BOUNDARY 8
726 
727 /* C++ stores the virtual bit in the lowest bit of function pointers.  */
728 #define TARGET_PTRMEMFUNC_VBIT_LOCATION ptrmemfunc_vbit_in_pfn
729 
730 /* Alignment of field after `int : 0' in a structure.  */
731 
732 #define EMPTY_FIELD_BOUNDARY BITS_PER_WORD
733 
734 /* Minimum size in bits of the largest boundary to which any
735    and all fundamental data types supported by the hardware
736    might need to be aligned. No data type wants to be aligned
737    rounder than this.
738 
739    Pentium+ prefers DFmode values to be aligned to 64 bit boundary
740    and Pentium Pro XFmode values at 128 bit boundaries.  */
741 
742 #define BIGGEST_ALIGNMENT (TARGET_AVX ? 256: 128)
743 
744 /* Maximum stack alignment.  */
745 #define MAX_STACK_ALIGNMENT MAX_OFILE_ALIGNMENT
746 
747 /* Alignment value for attribute ((aligned)).  It is a constant since
748    it is the part of the ABI.  We shouldn't change it with -mavx.  */
749 #define ATTRIBUTE_ALIGNED_VALUE 128
750 
751 /* Decide whether a variable of mode MODE should be 128 bit aligned.  */
752 #define ALIGN_MODE_128(MODE) \
753  ((MODE) == XFmode || SSE_REG_MODE_P (MODE))
754 
755 /* The published ABIs say that doubles should be aligned on word
756    boundaries, so lower the alignment for structure fields unless
757    -malign-double is set.  */
758 
759 /* ??? Blah -- this macro is used directly by libobjc.  Since it
760    supports no vector modes, cut out the complexity and fall back
761    on BIGGEST_FIELD_ALIGNMENT.  */
762 #ifdef IN_TARGET_LIBS
763 #ifdef __x86_64__
764 #define BIGGEST_FIELD_ALIGNMENT 128
765 #else
766 #define BIGGEST_FIELD_ALIGNMENT 32
767 #endif
768 #else
769 #define ADJUST_FIELD_ALIGN(FIELD, COMPUTED) \
770    x86_field_alignment (FIELD, COMPUTED)
771 #endif
772 
773 /* If defined, a C expression to compute the alignment given to a
774    constant that is being placed in memory.  EXP is the constant
775    and ALIGN is the alignment that the object would ordinarily have.
776    The value of this macro is used instead of that alignment to align
777    the object.
778 
779    If this macro is not defined, then ALIGN is used.
780 
781    The typical use of this macro is to increase alignment for string
782    constants to be word aligned so that `strcpy' calls that copy
783    constants can be done inline.  */
784 
785 #define CONSTANT_ALIGNMENT(EXP, ALIGN) ix86_constant_alignment ((EXP), (ALIGN))
786 
787 /* If defined, a C expression to compute the alignment for a static
788    variable.  TYPE is the data type, and ALIGN is the alignment that
789    the object would ordinarily have.  The value of this macro is used
790    instead of that alignment to align the object.
791 
792    If this macro is not defined, then ALIGN is used.
793 
794    One use of this macro is to increase alignment of medium-size
795    data to make it all fit in fewer cache lines.  Another is to
796    cause character arrays to be word-aligned so that `strcpy' calls
797    that copy constants to character arrays can be done inline.  */
798 
799 #define DATA_ALIGNMENT(TYPE, ALIGN) ix86_data_alignment ((TYPE), (ALIGN))
800 
801 /* If defined, a C expression to compute the alignment for a local
802    variable.  TYPE is the data type, and ALIGN is the alignment that
803    the object would ordinarily have.  The value of this macro is used
804    instead of that alignment to align the object.
805 
806    If this macro is not defined, then ALIGN is used.
807 
808    One use of this macro is to increase alignment of medium-size
809    data to make it all fit in fewer cache lines.  */
810 
811 #define LOCAL_ALIGNMENT(TYPE, ALIGN) \
812   ix86_local_alignment ((TYPE), VOIDmode, (ALIGN))
813 
814 /* If defined, a C expression to compute the alignment for stack slot.
815    TYPE is the data type, MODE is the widest mode available, and ALIGN
816    is the alignment that the slot would ordinarily have.  The value of
817    this macro is used instead of that alignment to align the slot.
818 
819    If this macro is not defined, then ALIGN is used when TYPE is NULL,
820    Otherwise, LOCAL_ALIGNMENT will be used.
821 
822    One use of this macro is to set alignment of stack slot to the
823    maximum alignment of all possible modes which the slot may have.  */
824 
825 #define STACK_SLOT_ALIGNMENT(TYPE, MODE, ALIGN) \
826   ix86_local_alignment ((TYPE), (MODE), (ALIGN))
827 
828 /* If defined, a C expression to compute the alignment for a local
829    variable DECL.
830 
831    If this macro is not defined, then
832    LOCAL_ALIGNMENT (TREE_TYPE (DECL), DECL_ALIGN (DECL)) will be used.
833 
834    One use of this macro is to increase alignment of medium-size
835    data to make it all fit in fewer cache lines.  */
836 
837 #define LOCAL_DECL_ALIGNMENT(DECL) \
838   ix86_local_alignment ((DECL), VOIDmode, DECL_ALIGN (DECL))
839 
840 /* If defined, a C expression to compute the minimum required alignment
841    for dynamic stack realignment purposes for EXP (a TYPE or DECL),
842    MODE, assuming normal alignment ALIGN.
843 
844    If this macro is not defined, then (ALIGN) will be used.  */
845 
846 #define MINIMUM_ALIGNMENT(EXP, MODE, ALIGN) \
847   ix86_minimum_alignment (EXP, MODE, ALIGN)
848 
849 
850 /* If defined, a C expression that gives the alignment boundary, in
851    bits, of an argument with the specified mode and type.  If it is
852    not defined, `PARM_BOUNDARY' is used for all arguments.  */
853 
854 #define FUNCTION_ARG_BOUNDARY(MODE, TYPE) \
855   ix86_function_arg_boundary ((MODE), (TYPE))
856 
857 /* Set this nonzero if move instructions will actually fail to work
858    when given unaligned data.  */
859 #define STRICT_ALIGNMENT 0
860 
861 /* If bit field type is int, don't let it cross an int,
862    and give entire struct the alignment of an int.  */
863 /* Required on the 386 since it doesn't have bit-field insns.  */
864 #define PCC_BITFIELD_TYPE_MATTERS 1
865 
866 /* Standard register usage.  */
867 
868 /* This processor has special stack-like registers.  See reg-stack.c
869    for details.  */
870 
871 #define STACK_REGS
872 
873 #define IS_STACK_MODE(MODE)					\
874   (((MODE) == SFmode && (!TARGET_SSE || !TARGET_SSE_MATH))	\
875    || ((MODE) == DFmode && (!TARGET_SSE2 || !TARGET_SSE_MATH))  \
876    || (MODE) == XFmode)
877 
878 /* Cover class containing the stack registers.  */
879 #define STACK_REG_COVER_CLASS FLOAT_REGS
880 
881 /* Number of actual hardware registers.
882    The hardware registers are assigned numbers for the compiler
883    from 0 to just below FIRST_PSEUDO_REGISTER.
884    All registers that the compiler knows about must be given numbers,
885    even those that are not normally considered general registers.
886 
887    In the 80386 we give the 8 general purpose registers the numbers 0-7.
888    We number the floating point registers 8-15.
889    Note that registers 0-7 can be accessed as a  short or int,
890    while only 0-3 may be used with byte `mov' instructions.
891 
892    Reg 16 does not correspond to any hardware register, but instead
893    appears in the RTL as an argument pointer prior to reload, and is
894    eliminated during reloading in favor of either the stack or frame
895    pointer.  */
896 
897 #define FIRST_PSEUDO_REGISTER 53
898 
899 /* Number of hardware registers that go into the DWARF-2 unwind info.
900    If not defined, equals FIRST_PSEUDO_REGISTER.  */
901 
902 #define DWARF_FRAME_REGISTERS 17
903 
904 /* 1 for registers that have pervasive standard uses
905    and are not available for the register allocator.
906    On the 80386, the stack pointer is such, as is the arg pointer.
907 
908    The value is zero if the register is not fixed on either 32 or
909    64 bit targets, one if the register if fixed on both 32 and 64
910    bit targets, two if it is only fixed on 32bit targets and three
911    if its only fixed on 64bit targets.
912    Proper values are computed in the CONDITIONAL_REGISTER_USAGE.
913  */
914 #define FIXED_REGISTERS						\
915 /*ax,dx,cx,bx,si,di,bp,sp,st,st1,st2,st3,st4,st5,st6,st7*/	\
916 {  0, 0, 0, 0, 0, 0, 0, 1, 0,  0,  0,  0,  0,  0,  0,  0,	\
917 /*arg,flags,fpsr,fpcr,frame*/					\
918     1,    1,   1,   1,    1,					\
919 /*xmm0,xmm1,xmm2,xmm3,xmm4,xmm5,xmm6,xmm7*/			\
920      0,   0,   0,   0,   0,   0,   0,   0,			\
921 /* mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7*/			\
922      0,   0,   0,   0,   0,   0,   0,   0,			\
923 /*  r8,  r9, r10, r11, r12, r13, r14, r15*/			\
924      2,   2,   2,   2,   2,   2,   2,   2,			\
925 /*xmm8,xmm9,xmm10,xmm11,xmm12,xmm13,xmm14,xmm15*/		\
926      2,   2,    2,    2,    2,    2,    2,    2 }
927 
928 
929 /* 1 for registers not available across function calls.
930    These must include the FIXED_REGISTERS and also any
931    registers that can be used without being saved.
932    The latter must include the registers where values are returned
933    and the register where structure-value addresses are passed.
934    Aside from that, you can include as many other registers as you like.
935 
936    The value is zero if the register is not call used on either 32 or
937    64 bit targets, one if the register if call used on both 32 and 64
938    bit targets, two if it is only call used on 32bit targets and three
939    if its only call used on 64bit targets.
940    Proper values are computed in the CONDITIONAL_REGISTER_USAGE.
941 */
942 #define CALL_USED_REGISTERS					\
943 /*ax,dx,cx,bx,si,di,bp,sp,st,st1,st2,st3,st4,st5,st6,st7*/	\
944 {  1, 1, 1, 0, 3, 3, 0, 1, 1,  1,  1,  1,  1,  1,  1,  1,	\
945 /*arg,flags,fpsr,fpcr,frame*/					\
946     1,   1,    1,   1,    1,					\
947 /*xmm0,xmm1,xmm2,xmm3,xmm4,xmm5,xmm6,xmm7*/			\
948      1,   1,   1,   1,   1,   1,   1,   1,			\
949 /* mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7*/			\
950      1,   1,   1,   1,   1,   1,   1,   1,			\
951 /*  r8,  r9, r10, r11, r12, r13, r14, r15*/			\
952      1,   1,   1,   1,   2,   2,   2,   2,			\
953 /*xmm8,xmm9,xmm10,xmm11,xmm12,xmm13,xmm14,xmm15*/		\
954      1,   1,    1,    1,    1,    1,    1,    1 }
955 
956 /* Order in which to allocate registers.  Each register must be
957    listed once, even those in FIXED_REGISTERS.  List frame pointer
958    late and fixed registers last.  Note that, in general, we prefer
959    registers listed in CALL_USED_REGISTERS, keeping the others
960    available for storage of persistent values.
961 
962    The ORDER_REGS_FOR_LOCAL_ALLOC actually overwrite the order,
963    so this is just empty initializer for array.  */
964 
965 #define REG_ALLOC_ORDER 					\
966 {  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,\
967    18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32,	\
968    33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,  \
969    48, 49, 50, 51, 52 }
970 
971 /* ORDER_REGS_FOR_LOCAL_ALLOC is a macro which permits reg_alloc_order
972    to be rearranged based on a particular function.  When using sse math,
973    we want to allocate SSE before x87 registers and vice versa.  */
974 
975 #define ORDER_REGS_FOR_LOCAL_ALLOC x86_order_regs_for_local_alloc ()
976 
977 
978 #define OVERRIDE_ABI_FORMAT(FNDECL) ix86_call_abi_override (FNDECL)
979 
980 /* Macro to conditionally modify fixed_regs/call_used_regs.  */
981 #define CONDITIONAL_REGISTER_USAGE  ix86_conditional_register_usage ()
982 
983 /* Return number of consecutive hard regs needed starting at reg REGNO
984    to hold something of mode MODE.
985    This is ordinarily the length in words of a value of mode MODE
986    but can be less for certain modes in special long registers.
987 
988    Actually there are no two word move instructions for consecutive
989    registers.  And only registers 0-3 may have mov byte instructions
990    applied to them.
991    */
992 
993 #define HARD_REGNO_NREGS(REGNO, MODE)					\
994   (FP_REGNO_P (REGNO) || SSE_REGNO_P (REGNO) || MMX_REGNO_P (REGNO)	\
995    ? (COMPLEX_MODE_P (MODE) ? 2 : 1)					\
996    : ((MODE) == XFmode							\
997       ? (TARGET_64BIT ? 2 : 3)						\
998       : (MODE) == XCmode						\
999       ? (TARGET_64BIT ? 4 : 6)						\
1000       : ((GET_MODE_SIZE (MODE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)))
1001 
1002 #define HARD_REGNO_NREGS_HAS_PADDING(REGNO, MODE)			\
1003   ((TARGET_128BIT_LONG_DOUBLE && !TARGET_64BIT)				\
1004    ? (FP_REGNO_P (REGNO) || SSE_REGNO_P (REGNO) || MMX_REGNO_P (REGNO)	\
1005       ? 0								\
1006       : ((MODE) == XFmode || (MODE) == XCmode))				\
1007    : 0)
1008 
1009 #define HARD_REGNO_NREGS_WITH_PADDING(REGNO, MODE) ((MODE) == XFmode ? 4 : 8)
1010 
1011 #define VALID_AVX256_REG_MODE(MODE)					\
1012   ((MODE) == V32QImode || (MODE) == V16HImode || (MODE) == V8SImode	\
1013    || (MODE) == V4DImode || (MODE) == V8SFmode || (MODE) == V4DFmode)
1014 
1015 #define VALID_SSE2_REG_MODE(MODE)					\
1016   ((MODE) == V16QImode || (MODE) == V8HImode || (MODE) == V2DFmode	\
1017    || (MODE) == V2DImode || (MODE) == DFmode)
1018 
1019 #define VALID_SSE_REG_MODE(MODE)					\
1020   ((MODE) == V1TImode || (MODE) == TImode				\
1021    || (MODE) == V4SFmode || (MODE) == V4SImode				\
1022    || (MODE) == SFmode || (MODE) == TFmode)
1023 
1024 #define VALID_MMX_REG_MODE_3DNOW(MODE) \
1025   ((MODE) == V2SFmode || (MODE) == SFmode)
1026 
1027 #define VALID_MMX_REG_MODE(MODE)					\
1028   ((MODE == V1DImode) || (MODE) == DImode				\
1029    || (MODE) == V2SImode || (MODE) == SImode				\
1030    || (MODE) == V4HImode || (MODE) == V8QImode)
1031 
1032 /* ??? No autovectorization into MMX or 3DNOW until we can reliably
1033    place emms and femms instructions.
1034    FIXME: AVX has 32byte floating point vector operations and 16byte
1035    integer vector operations.  But vectorizer doesn't support
1036    different sizes for integer and floating point vectors.  We limit
1037    vector size to 16byte.  */
1038 #define UNITS_PER_SIMD_WORD(MODE)					\
1039   (TARGET_AVX ? (((MODE) == DFmode || (MODE) == SFmode) ? 16 : 16)	\
1040    	      : (TARGET_SSE ? 16 : UNITS_PER_WORD))
1041 
1042 #define VALID_DFP_MODE_P(MODE) \
1043   ((MODE) == SDmode || (MODE) == DDmode || (MODE) == TDmode)
1044 
1045 #define VALID_FP_MODE_P(MODE)						\
1046   ((MODE) == SFmode || (MODE) == DFmode || (MODE) == XFmode		\
1047    || (MODE) == SCmode || (MODE) == DCmode || (MODE) == XCmode)		\
1048 
1049 #define VALID_INT_MODE_P(MODE)						\
1050   ((MODE) == QImode || (MODE) == HImode || (MODE) == SImode		\
1051    || (MODE) == DImode							\
1052    || (MODE) == CQImode || (MODE) == CHImode || (MODE) == CSImode	\
1053    || (MODE) == CDImode							\
1054    || (TARGET_64BIT && ((MODE) == TImode || (MODE) == CTImode		\
1055 			|| (MODE) == TFmode || (MODE) == TCmode)))
1056 
1057 /* Return true for modes passed in SSE registers.  */
1058 #define SSE_REG_MODE_P(MODE)						\
1059   ((MODE) == V1TImode || (MODE) == TImode || (MODE) == V16QImode	\
1060    || (MODE) == TFmode || (MODE) == V8HImode || (MODE) == V2DFmode	\
1061    || (MODE) == V2DImode || (MODE) == V4SFmode || (MODE) == V4SImode	\
1062    || (MODE) == V32QImode || (MODE) == V16HImode || (MODE) == V8SImode	\
1063    || (MODE) == V4DImode || (MODE) == V8SFmode || (MODE) == V4DFmode)
1064 
1065 /* Value is 1 if hard register REGNO can hold a value of machine-mode MODE.  */
1066 
1067 #define HARD_REGNO_MODE_OK(REGNO, MODE)	\
1068    ix86_hard_regno_mode_ok ((REGNO), (MODE))
1069 
1070 /* Value is 1 if it is a good idea to tie two pseudo registers
1071    when one has mode MODE1 and one has mode MODE2.
1072    If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
1073    for any hard reg, then this must be 0 for correct output.  */
1074 
1075 #define MODES_TIEABLE_P(MODE1, MODE2)  ix86_modes_tieable_p (MODE1, MODE2)
1076 
1077 /* It is possible to write patterns to move flags; but until someone
1078    does it,  */
1079 #define AVOID_CCMODE_COPIES
1080 
1081 /* Specify the modes required to caller save a given hard regno.
1082    We do this on i386 to prevent flags from being saved at all.
1083 
1084    Kill any attempts to combine saving of modes.  */
1085 
1086 #define HARD_REGNO_CALLER_SAVE_MODE(REGNO, NREGS, MODE)			\
1087   (CC_REGNO_P (REGNO) ? VOIDmode					\
1088    : MMX_REGNO_P (REGNO) ? V8QImode					\
1089    : (MODE) == VOIDmode && (NREGS) != 1 ? VOIDmode			\
1090    : (MODE) == VOIDmode ? choose_hard_reg_mode ((REGNO), (NREGS), false) \
1091    : (MODE) == HImode && !TARGET_PARTIAL_REG_STALL ? SImode		\
1092    : (MODE) == QImode && (REGNO) > BX_REG && !TARGET_64BIT ? SImode 	\
1093    : (MODE))
1094 
1095 /* Specify the registers used for certain standard purposes.
1096    The values of these macros are register numbers.  */
1097 
1098 /* on the 386 the pc register is %eip, and is not usable as a general
1099    register.  The ordinary mov instructions won't work */
1100 /* #define PC_REGNUM  */
1101 
1102 /* Register to use for pushing function arguments.  */
1103 #define STACK_POINTER_REGNUM 7
1104 
1105 /* Base register for access to local variables of the function.  */
1106 #define HARD_FRAME_POINTER_REGNUM 6
1107 
1108 /* Base register for access to local variables of the function.  */
1109 #define FRAME_POINTER_REGNUM 20
1110 
1111 /* First floating point reg */
1112 #define FIRST_FLOAT_REG 8
1113 
1114 /* First & last stack-like regs */
1115 #define FIRST_STACK_REG FIRST_FLOAT_REG
1116 #define LAST_STACK_REG (FIRST_FLOAT_REG + 7)
1117 
1118 #define FIRST_SSE_REG (FRAME_POINTER_REGNUM + 1)
1119 #define LAST_SSE_REG  (FIRST_SSE_REG + 7)
1120 
1121 #define FIRST_MMX_REG  (LAST_SSE_REG + 1)
1122 #define LAST_MMX_REG   (FIRST_MMX_REG + 7)
1123 
1124 #define FIRST_REX_INT_REG  (LAST_MMX_REG + 1)
1125 #define LAST_REX_INT_REG   (FIRST_REX_INT_REG + 7)
1126 
1127 #define FIRST_REX_SSE_REG  (LAST_REX_INT_REG + 1)
1128 #define LAST_REX_SSE_REG   (FIRST_REX_SSE_REG + 7)
1129 
1130 /* Override this in other tm.h files to cope with various OS lossage
1131    requiring a frame pointer.  */
1132 #ifndef SUBTARGET_FRAME_POINTER_REQUIRED
1133 #define SUBTARGET_FRAME_POINTER_REQUIRED 0
1134 #endif
1135 
1136 /* Make sure we can access arbitrary call frames.  */
1137 #define SETUP_FRAME_ADDRESSES()  ix86_setup_frame_addresses ()
1138 
1139 /* Base register for access to arguments of the function.  */
1140 #define ARG_POINTER_REGNUM 16
1141 
1142 /* Register to hold the addressing base for position independent
1143    code access to data items.  We don't use PIC pointer for 64bit
1144    mode.  Define the regnum to dummy value to prevent gcc from
1145    pessimizing code dealing with EBX.
1146 
1147    To avoid clobbering a call-saved register unnecessarily, we renumber
1148    the pic register when possible.  The change is visible after the
1149    prologue has been emitted.  */
1150 
1151 #define REAL_PIC_OFFSET_TABLE_REGNUM  BX_REG
1152 
1153 #define PIC_OFFSET_TABLE_REGNUM				\
1154   ((TARGET_64BIT && ix86_cmodel == CM_SMALL_PIC)	\
1155    || !flag_pic ? INVALID_REGNUM			\
1156    : reload_completed ? REGNO (pic_offset_table_rtx)	\
1157    : REAL_PIC_OFFSET_TABLE_REGNUM)
1158 
1159 #define GOT_SYMBOL_NAME "_GLOBAL_OFFSET_TABLE_"
1160 
1161 /* This is overridden by <cygwin.h>.  */
1162 #define MS_AGGREGATE_RETURN 0
1163 
1164 /* This is overridden by <netware.h>.  */
1165 #define KEEP_AGGREGATE_RETURN_POINTER 0
1166 
1167 /* Define the classes of registers for register constraints in the
1168    machine description.  Also define ranges of constants.
1169 
1170    One of the classes must always be named ALL_REGS and include all hard regs.
1171    If there is more than one class, another class must be named NO_REGS
1172    and contain no registers.
1173 
1174    The name GENERAL_REGS must be the name of a class (or an alias for
1175    another name such as ALL_REGS).  This is the class of registers
1176    that is allowed by "g" or "r" in a register constraint.
1177    Also, registers outside this class are allocated only when
1178    instructions express preferences for them.
1179 
1180    The classes must be numbered in nondecreasing order; that is,
1181    a larger-numbered class must never be contained completely
1182    in a smaller-numbered class.
1183 
1184    For any two classes, it is very desirable that there be another
1185    class that represents their union.
1186 
1187    It might seem that class BREG is unnecessary, since no useful 386
1188    opcode needs reg %ebx.  But some systems pass args to the OS in ebx,
1189    and the "b" register constraint is useful in asms for syscalls.
1190 
1191    The flags, fpsr and fpcr registers are in no class.  */
1192 
1193 enum reg_class
1194 {
1195   NO_REGS,
1196   AREG, DREG, CREG, BREG, SIREG, DIREG,
1197   AD_REGS,			/* %eax/%edx for DImode */
1198   CLOBBERED_REGS,		/* call-clobbered integers */
1199   Q_REGS,			/* %eax %ebx %ecx %edx */
1200   NON_Q_REGS,			/* %esi %edi %ebp %esp */
1201   INDEX_REGS,			/* %eax %ebx %ecx %edx %esi %edi %ebp */
1202   LEGACY_REGS,			/* %eax %ebx %ecx %edx %esi %edi %ebp %esp */
1203   GENERAL_REGS,			/* %eax %ebx %ecx %edx %esi %edi %ebp %esp %r8 - %r15*/
1204   FP_TOP_REG, FP_SECOND_REG,	/* %st(0) %st(1) */
1205   FLOAT_REGS,
1206   SSE_FIRST_REG,
1207   SSE_REGS,
1208   MMX_REGS,
1209   FP_TOP_SSE_REGS,
1210   FP_SECOND_SSE_REGS,
1211   FLOAT_SSE_REGS,
1212   FLOAT_INT_REGS,
1213   INT_SSE_REGS,
1214   FLOAT_INT_SSE_REGS,
1215   ALL_REGS, LIM_REG_CLASSES
1216 };
1217 
1218 #define N_REG_CLASSES ((int) LIM_REG_CLASSES)
1219 
1220 #define INTEGER_CLASS_P(CLASS) \
1221   reg_class_subset_p ((CLASS), GENERAL_REGS)
1222 #define FLOAT_CLASS_P(CLASS) \
1223   reg_class_subset_p ((CLASS), FLOAT_REGS)
1224 #define SSE_CLASS_P(CLASS) \
1225   reg_class_subset_p ((CLASS), SSE_REGS)
1226 #define MMX_CLASS_P(CLASS) \
1227   ((CLASS) == MMX_REGS)
1228 #define MAYBE_INTEGER_CLASS_P(CLASS) \
1229   reg_classes_intersect_p ((CLASS), GENERAL_REGS)
1230 #define MAYBE_FLOAT_CLASS_P(CLASS) \
1231   reg_classes_intersect_p ((CLASS), FLOAT_REGS)
1232 #define MAYBE_SSE_CLASS_P(CLASS) \
1233   reg_classes_intersect_p (SSE_REGS, (CLASS))
1234 #define MAYBE_MMX_CLASS_P(CLASS) \
1235   reg_classes_intersect_p (MMX_REGS, (CLASS))
1236 
1237 #define Q_CLASS_P(CLASS) \
1238   reg_class_subset_p ((CLASS), Q_REGS)
1239 
1240 /* Give names of register classes as strings for dump file.  */
1241 
1242 #define REG_CLASS_NAMES \
1243 {  "NO_REGS",				\
1244    "AREG", "DREG", "CREG", "BREG",	\
1245    "SIREG", "DIREG",			\
1246    "AD_REGS",				\
1247    "CLOBBERED_REGS",			\
1248    "Q_REGS", "NON_Q_REGS",		\
1249    "INDEX_REGS",			\
1250    "LEGACY_REGS",			\
1251    "GENERAL_REGS",			\
1252    "FP_TOP_REG", "FP_SECOND_REG",	\
1253    "FLOAT_REGS",			\
1254    "SSE_FIRST_REG",			\
1255    "SSE_REGS",				\
1256    "MMX_REGS",				\
1257    "FP_TOP_SSE_REGS",			\
1258    "FP_SECOND_SSE_REGS",		\
1259    "FLOAT_SSE_REGS",			\
1260    "FLOAT_INT_REGS",			\
1261    "INT_SSE_REGS",			\
1262    "FLOAT_INT_SSE_REGS",		\
1263    "ALL_REGS" }
1264 
1265 /* Define which registers fit in which classes.  This is an initializer
1266    for a vector of HARD_REG_SET of length N_REG_CLASSES.
1267 
1268    Note that the default setting of CLOBBERED_REGS is for 32-bit; this
1269    is adjusted by CONDITIONAL_REGISTER_USAGE for the 64-bit ABI in effect.  */
1270 
1271 #define REG_CLASS_CONTENTS						\
1272 {     { 0x00,     0x0 },						\
1273       { 0x01,     0x0 }, { 0x02, 0x0 },	/* AREG, DREG */		\
1274       { 0x04,     0x0 }, { 0x08, 0x0 },	/* CREG, BREG */		\
1275       { 0x10,     0x0 }, { 0x20, 0x0 },	/* SIREG, DIREG */		\
1276       { 0x03,     0x0 },		/* AD_REGS */			\
1277       { 0x07,     0x0 },		/* CLOBBERED_REGS */		\
1278       { 0x0f,     0x0 },		/* Q_REGS */			\
1279   { 0x1100f0,  0x1fe0 },		/* NON_Q_REGS */		\
1280       { 0x7f,  0x1fe0 },		/* INDEX_REGS */		\
1281   { 0x1100ff,     0x0 },		/* LEGACY_REGS */		\
1282   { 0x1100ff,  0x1fe0 },		/* GENERAL_REGS */		\
1283      { 0x100,     0x0 }, { 0x0200, 0x0 },/* FP_TOP_REG, FP_SECOND_REG */\
1284     { 0xff00,     0x0 },		/* FLOAT_REGS */		\
1285   { 0x200000,     0x0 },		/* SSE_FIRST_REG */		\
1286 { 0x1fe00000,0x1fe000 },		/* SSE_REGS */			\
1287 { 0xe0000000,    0x1f },		/* MMX_REGS */			\
1288 { 0x1fe00100,0x1fe000 },		/* FP_TOP_SSE_REG */		\
1289 { 0x1fe00200,0x1fe000 },		/* FP_SECOND_SSE_REG */		\
1290 { 0x1fe0ff00,0x1fe000 },		/* FLOAT_SSE_REGS */		\
1291    { 0x1ffff,  0x1fe0 },		/* FLOAT_INT_REGS */		\
1292 { 0x1fe100ff,0x1fffe0 },		/* INT_SSE_REGS */		\
1293 { 0x1fe1ffff,0x1fffe0 },		/* FLOAT_INT_SSE_REGS */	\
1294 { 0xffffffff,0x1fffff }							\
1295 }
1296 
1297 /* The same information, inverted:
1298    Return the class number of the smallest class containing
1299    reg number REGNO.  This could be a conditional expression
1300    or could index an array.  */
1301 
1302 #define REGNO_REG_CLASS(REGNO) (regclass_map[REGNO])
1303 
1304 /* When defined, the compiler allows registers explicitly used in the
1305    rtl to be used as spill registers but prevents the compiler from
1306    extending the lifetime of these registers.  */
1307 
1308 #define SMALL_REGISTER_CLASSES 1
1309 
1310 #define QI_REG_P(X) (REG_P (X) && REGNO (X) <= BX_REG)
1311 
1312 #define GENERAL_REGNO_P(N) \
1313   ((N) <= STACK_POINTER_REGNUM || REX_INT_REGNO_P (N))
1314 
1315 #define GENERAL_REG_P(X) \
1316   (REG_P (X) && GENERAL_REGNO_P (REGNO (X)))
1317 
1318 #define ANY_QI_REG_P(X) (TARGET_64BIT ? GENERAL_REG_P(X) : QI_REG_P (X))
1319 
1320 #define REX_INT_REGNO_P(N) \
1321   IN_RANGE ((N), FIRST_REX_INT_REG, LAST_REX_INT_REG)
1322 #define REX_INT_REG_P(X) (REG_P (X) && REX_INT_REGNO_P (REGNO (X)))
1323 
1324 #define FP_REG_P(X) (REG_P (X) && FP_REGNO_P (REGNO (X)))
1325 #define FP_REGNO_P(N) IN_RANGE ((N), FIRST_STACK_REG, LAST_STACK_REG)
1326 #define ANY_FP_REG_P(X) (REG_P (X) && ANY_FP_REGNO_P (REGNO (X)))
1327 #define ANY_FP_REGNO_P(N) (FP_REGNO_P (N) || SSE_REGNO_P (N))
1328 
1329 #define X87_FLOAT_MODE_P(MODE)	\
1330   (TARGET_80387 && ((MODE) == SFmode || (MODE) == DFmode || (MODE) == XFmode))
1331 
1332 #define SSE_REG_P(N) (REG_P (N) && SSE_REGNO_P (REGNO (N)))
1333 #define SSE_REGNO_P(N)						\
1334   (IN_RANGE ((N), FIRST_SSE_REG, LAST_SSE_REG)			\
1335    || REX_SSE_REGNO_P (N))
1336 
1337 #define REX_SSE_REGNO_P(N) \
1338   IN_RANGE ((N), FIRST_REX_SSE_REG, LAST_REX_SSE_REG)
1339 
1340 #define SSE_REGNO(N) \
1341   ((N) < 8 ? FIRST_SSE_REG + (N) : FIRST_REX_SSE_REG + (N) - 8)
1342 
1343 #define SSE_FLOAT_MODE_P(MODE) \
1344   ((TARGET_SSE && (MODE) == SFmode) || (TARGET_SSE2 && (MODE) == DFmode))
1345 
1346 #define SSE_VEC_FLOAT_MODE_P(MODE) \
1347   ((TARGET_SSE && (MODE) == V4SFmode) || (TARGET_SSE2 && (MODE) == V2DFmode))
1348 
1349 #define AVX_FLOAT_MODE_P(MODE) \
1350   (TARGET_AVX && ((MODE) == SFmode || (MODE) == DFmode))
1351 
1352 #define AVX128_VEC_FLOAT_MODE_P(MODE) \
1353   (TARGET_AVX && ((MODE) == V4SFmode || (MODE) == V2DFmode))
1354 
1355 #define AVX256_VEC_FLOAT_MODE_P(MODE) \
1356   (TARGET_AVX && ((MODE) == V8SFmode || (MODE) == V4DFmode))
1357 
1358 #define AVX_VEC_FLOAT_MODE_P(MODE) \
1359   (TARGET_AVX && ((MODE) == V4SFmode || (MODE) == V2DFmode \
1360 		  || (MODE) == V8SFmode || (MODE) == V4DFmode))
1361 
1362 #define FMA4_VEC_FLOAT_MODE_P(MODE) \
1363   (TARGET_FMA4 && ((MODE) == V4SFmode || (MODE) == V2DFmode \
1364 		  || (MODE) == V8SFmode || (MODE) == V4DFmode))
1365 
1366 #define MMX_REG_P(XOP) (REG_P (XOP) && MMX_REGNO_P (REGNO (XOP)))
1367 #define MMX_REGNO_P(N) IN_RANGE ((N), FIRST_MMX_REG, LAST_MMX_REG)
1368 
1369 #define STACK_REG_P(XOP) (REG_P (XOP) && STACK_REGNO_P (REGNO (XOP)))
1370 #define STACK_REGNO_P(N) IN_RANGE ((N), FIRST_STACK_REG, LAST_STACK_REG)
1371 
1372 #define STACK_TOP_P(XOP) (REG_P (XOP) && REGNO (XOP) == FIRST_STACK_REG)
1373 
1374 #define CC_REG_P(X) (REG_P (X) && CC_REGNO_P (REGNO (X)))
1375 #define CC_REGNO_P(X) ((X) == FLAGS_REG || (X) == FPSR_REG)
1376 
1377 /* The class value for index registers, and the one for base regs.  */
1378 
1379 #define INDEX_REG_CLASS INDEX_REGS
1380 #define BASE_REG_CLASS GENERAL_REGS
1381 
1382 /* Place additional restrictions on the register class to use when it
1383    is necessary to be able to hold a value of mode MODE in a reload
1384    register for which class CLASS would ordinarily be used.  */
1385 
1386 #define LIMIT_RELOAD_CLASS(MODE, CLASS) 			\
1387   ((MODE) == QImode && !TARGET_64BIT				\
1388    && ((CLASS) == ALL_REGS || (CLASS) == GENERAL_REGS		\
1389        || (CLASS) == LEGACY_REGS || (CLASS) == INDEX_REGS)	\
1390    ? Q_REGS : (CLASS))
1391 
1392 /* Given an rtx X being reloaded into a reg required to be
1393    in class CLASS, return the class of reg to actually use.
1394    In general this is just CLASS; but on some machines
1395    in some cases it is preferable to use a more restrictive class.
1396    On the 80386 series, we prevent floating constants from being
1397    reloaded into floating registers (since no move-insn can do that)
1398    and we ensure that QImodes aren't reloaded into the esi or edi reg.  */
1399 
1400 /* Put float CONST_DOUBLE in the constant pool instead of fp regs.
1401    QImode must go into class Q_REGS.
1402    Narrow ALL_REGS to GENERAL_REGS.  This supports allowing movsf and
1403    movdf to do mem-to-mem moves through integer regs.  */
1404 
1405 #define PREFERRED_RELOAD_CLASS(X, CLASS) \
1406    ix86_preferred_reload_class ((X), (CLASS))
1407 
1408 /* Discourage putting floating-point values in SSE registers unless
1409    SSE math is being used, and likewise for the 387 registers.  */
1410 
1411 #define PREFERRED_OUTPUT_RELOAD_CLASS(X, CLASS) \
1412    ix86_preferred_output_reload_class ((X), (CLASS))
1413 
1414 /* If we are copying between general and FP registers, we need a memory
1415    location. The same is true for SSE and MMX registers.  */
1416 #define SECONDARY_MEMORY_NEEDED(CLASS1, CLASS2, MODE) \
1417   ix86_secondary_memory_needed ((CLASS1), (CLASS2), (MODE), 1)
1418 
1419 /* Get_secondary_mem widens integral modes to BITS_PER_WORD.
1420    There is no need to emit full 64 bit move on 64 bit targets
1421    for integral modes that can be moved using 32 bit move.  */
1422 #define SECONDARY_MEMORY_NEEDED_MODE(MODE)			\
1423   (GET_MODE_BITSIZE (MODE) < 32 && INTEGRAL_MODE_P (MODE)	\
1424    ? mode_for_size (32, GET_MODE_CLASS (MODE), 0)		\
1425    : MODE)
1426 
1427 /* Return the maximum number of consecutive registers
1428    needed to represent mode MODE in a register of class CLASS.  */
1429 /* On the 80386, this is the size of MODE in words,
1430    except in the FP regs, where a single reg is always enough.  */
1431 #define CLASS_MAX_NREGS(CLASS, MODE)					\
1432  (!MAYBE_INTEGER_CLASS_P (CLASS)					\
1433   ? (COMPLEX_MODE_P (MODE) ? 2 : 1)					\
1434   : (((((MODE) == XFmode ? 12 : GET_MODE_SIZE (MODE)))			\
1435       + UNITS_PER_WORD - 1) / UNITS_PER_WORD))
1436 
1437 /* A C expression whose value is nonzero if pseudos that have been
1438    assigned to registers of class CLASS would likely be spilled
1439    because registers of CLASS are needed for spill registers.
1440 
1441    The default value of this macro returns 1 if CLASS has exactly one
1442    register and zero otherwise.  On most machines, this default
1443    should be used.  Only define this macro to some other expression
1444    if pseudo allocated by `local-alloc.c' end up in memory because
1445    their hard registers were needed for spill registers.  If this
1446    macro returns nonzero for those classes, those pseudos will only
1447    be allocated by `global.c', which knows how to reallocate the
1448    pseudo to another register.  If there would not be another
1449    register available for reallocation, you should not change the
1450    definition of this macro since the only effect of such a
1451    definition would be to slow down register allocation.  */
1452 
1453 #define CLASS_LIKELY_SPILLED_P(CLASS)					\
1454   (((CLASS) == AREG)							\
1455    || ((CLASS) == DREG)							\
1456    || ((CLASS) == CREG)							\
1457    || ((CLASS) == BREG)							\
1458    || ((CLASS) == AD_REGS)						\
1459    || ((CLASS) == SIREG)						\
1460    || ((CLASS) == DIREG)						\
1461    || ((CLASS) == SSE_FIRST_REG)					\
1462    || ((CLASS) == FP_TOP_REG)						\
1463    || ((CLASS) == FP_SECOND_REG))
1464 
1465 /* Return a class of registers that cannot change FROM mode to TO mode.  */
1466 
1467 #define CANNOT_CHANGE_MODE_CLASS(FROM, TO, CLASS) \
1468   ix86_cannot_change_mode_class (FROM, TO, CLASS)
1469 
1470 /* Stack layout; function entry, exit and calling.  */
1471 
1472 /* Define this if pushing a word on the stack
1473    makes the stack pointer a smaller address.  */
1474 #define STACK_GROWS_DOWNWARD
1475 
1476 /* Define this to nonzero if the nominal address of the stack frame
1477    is at the high-address end of the local variables;
1478    that is, each additional local variable allocated
1479    goes at a more negative offset in the frame.  */
1480 #define FRAME_GROWS_DOWNWARD 1
1481 
1482 /* Offset within stack frame to start allocating local variables at.
1483    If FRAME_GROWS_DOWNWARD, this is the offset to the END of the
1484    first local allocated.  Otherwise, it is the offset to the BEGINNING
1485    of the first local allocated.  */
1486 #define STARTING_FRAME_OFFSET 0
1487 
1488 /* If we generate an insn to push BYTES bytes,
1489    this says how many the stack pointer really advances by.
1490    On 386, we have pushw instruction that decrements by exactly 2 no
1491    matter what the position was, there is no pushb.
1492    But as CIE data alignment factor on this arch is -4, we need to make
1493    sure all stack pointer adjustments are in multiple of 4.
1494 
1495    For 64bit ABI we round up to 8 bytes.
1496  */
1497 
1498 #define PUSH_ROUNDING(BYTES) \
1499   (TARGET_64BIT		     \
1500    ? (((BYTES) + 7) & (-8))  \
1501    : (((BYTES) + 3) & (-4)))
1502 
1503 /* If defined, the maximum amount of space required for outgoing arguments will
1504    be computed and placed into the variable
1505    `crtl->outgoing_args_size'.  No space will be pushed onto the
1506    stack for each call; instead, the function prologue should increase the stack
1507    frame size by this amount.
1508 
1509    MS ABI seem to require 16 byte alignment everywhere except for function
1510    prologue and apilogue.  This is not possible without
1511    ACCUMULATE_OUTGOING_ARGS.  */
1512 
1513 #define ACCUMULATE_OUTGOING_ARGS \
1514   (TARGET_ACCUMULATE_OUTGOING_ARGS || ix86_cfun_abi () == MS_ABI)
1515 
1516 /* If defined, a C expression whose value is nonzero when we want to use PUSH
1517    instructions to pass outgoing arguments.  */
1518 
1519 #define PUSH_ARGS (TARGET_PUSH_ARGS && !ACCUMULATE_OUTGOING_ARGS)
1520 
1521 /* We want the stack and args grow in opposite directions, even if
1522    PUSH_ARGS is 0.  */
1523 #define PUSH_ARGS_REVERSED 1
1524 
1525 /* Offset of first parameter from the argument pointer register value.  */
1526 #define FIRST_PARM_OFFSET(FNDECL) 0
1527 
1528 /* Define this macro if functions should assume that stack space has been
1529    allocated for arguments even when their values are passed in registers.
1530 
1531    The value of this macro is the size, in bytes, of the area reserved for
1532    arguments passed in registers for the function represented by FNDECL.
1533 
1534    This space can be allocated by the caller, or be a part of the
1535    machine-dependent stack frame: `OUTGOING_REG_PARM_STACK_SPACE' says
1536    which.  */
1537 #define REG_PARM_STACK_SPACE(FNDECL) ix86_reg_parm_stack_space (FNDECL)
1538 
1539 #define OUTGOING_REG_PARM_STACK_SPACE(FNTYPE) \
1540   (ix86_function_type_abi (FNTYPE) == MS_ABI)
1541 
1542 /* Value is the number of bytes of arguments automatically
1543    popped when returning from a subroutine call.
1544    FUNDECL is the declaration node of the function (as a tree),
1545    FUNTYPE is the data type of the function (as a tree),
1546    or for a library call it is an identifier node for the subroutine name.
1547    SIZE is the number of bytes of arguments passed on the stack.
1548 
1549    On the 80386, the RTD insn may be used to pop them if the number
1550      of args is fixed, but if the number is variable then the caller
1551      must pop them all.  RTD can't be used for library calls now
1552      because the library is compiled with the Unix compiler.
1553    Use of RTD is a selectable option, since it is incompatible with
1554    standard Unix calling sequences.  If the option is not selected,
1555    the caller must always pop the args.
1556 
1557    The attribute stdcall is equivalent to RTD on a per module basis.  */
1558 
1559 #define RETURN_POPS_ARGS(FUNDECL, FUNTYPE, SIZE) \
1560   ix86_return_pops_args ((FUNDECL), (FUNTYPE), (SIZE))
1561 
1562 #define FUNCTION_VALUE_REGNO_P(N) ix86_function_value_regno_p (N)
1563 
1564 /* Define how to find the value returned by a library function
1565    assuming the value has mode MODE.  */
1566 
1567 #define LIBCALL_VALUE(MODE) ix86_libcall_value (MODE)
1568 
1569 /* Define the size of the result block used for communication between
1570    untyped_call and untyped_return.  The block contains a DImode value
1571    followed by the block used by fnsave and frstor.  */
1572 
1573 #define APPLY_RESULT_SIZE (8+108)
1574 
1575 /* 1 if N is a possible register number for function argument passing.  */
1576 #define FUNCTION_ARG_REGNO_P(N) ix86_function_arg_regno_p (N)
1577 
1578 /* Define a data type for recording info about an argument list
1579    during the scan of that argument list.  This data type should
1580    hold all necessary information about the function itself
1581    and about the args processed so far, enough to enable macros
1582    such as FUNCTION_ARG to determine where the next arg should go.  */
1583 
1584 typedef struct ix86_args {
1585   int words;			/* # words passed so far */
1586   int nregs;			/* # registers available for passing */
1587   int regno;			/* next available register number */
1588   int fastcall;			/* fastcall calling convention is used */
1589   int sse_words;		/* # sse words passed so far */
1590   int sse_nregs;		/* # sse registers available for passing */
1591   int warn_avx;			/* True when we want to warn about AVX ABI.  */
1592   int warn_sse;			/* True when we want to warn about SSE ABI.  */
1593   int warn_mmx;			/* True when we want to warn about MMX ABI.  */
1594   int sse_regno;		/* next available sse register number */
1595   int mmx_words;		/* # mmx words passed so far */
1596   int mmx_nregs;		/* # mmx registers available for passing */
1597   int mmx_regno;		/* next available mmx register number */
1598   int maybe_vaarg;		/* true for calls to possibly vardic fncts.  */
1599   int float_in_sse;		/* 1 if in 32-bit mode SFmode (2 for DFmode) should
1600 				   be passed in SSE registers.  Otherwise 0.  */
1601   enum calling_abi call_abi;	/* Set to SYSV_ABI for sysv abi. Otherwise
1602  				   MS_ABI for ms abi.  */
1603 } CUMULATIVE_ARGS;
1604 
1605 /* Initialize a variable CUM of type CUMULATIVE_ARGS
1606    for a call to a function whose data type is FNTYPE.
1607    For a library call, FNTYPE is 0.  */
1608 
1609 #define INIT_CUMULATIVE_ARGS(CUM, FNTYPE, LIBNAME, FNDECL, N_NAMED_ARGS) \
1610   init_cumulative_args (&(CUM), (FNTYPE), (LIBNAME), (FNDECL))
1611 
1612 /* Update the data in CUM to advance over an argument
1613    of mode MODE and data type TYPE.
1614    (TYPE is null for libcalls where that information may not be available.)  */
1615 
1616 #define FUNCTION_ARG_ADVANCE(CUM, MODE, TYPE, NAMED) \
1617   function_arg_advance (&(CUM), (MODE), (TYPE), (NAMED))
1618 
1619 /* Define where to put the arguments to a function.
1620    Value is zero to push the argument on the stack,
1621    or a hard register in which to store the argument.
1622 
1623    MODE is the argument's machine mode.
1624    TYPE is the data type of the argument (as a tree).
1625     This is null for libcalls where that information may
1626     not be available.
1627    CUM is a variable of type CUMULATIVE_ARGS which gives info about
1628     the preceding args and about the function being called.
1629    NAMED is nonzero if this argument is a named parameter
1630     (otherwise it is an extra parameter matching an ellipsis).  */
1631 
1632 #define FUNCTION_ARG(CUM, MODE, TYPE, NAMED) \
1633   function_arg (&(CUM), (MODE), (TYPE), (NAMED))
1634 
1635 /* Output assembler code to FILE to increment profiler label # LABELNO
1636    for profiling a function entry.  */
1637 
1638 #define FUNCTION_PROFILER(FILE, LABELNO) x86_function_profiler (FILE, LABELNO)
1639 
1640 #define MCOUNT_NAME "_mcount"
1641 
1642 #define PROFILE_COUNT_REGISTER "edx"
1643 
1644 /* EXIT_IGNORE_STACK should be nonzero if, when returning from a function,
1645    the stack pointer does not matter.  The value is tested only in
1646    functions that have frame pointers.
1647    No definition is equivalent to always zero.  */
1648 /* Note on the 386 it might be more efficient not to define this since
1649    we have to restore it ourselves from the frame pointer, in order to
1650    use pop */
1651 
1652 #define EXIT_IGNORE_STACK 1
1653 
1654 /* Output assembler code for a block containing the constant parts
1655    of a trampoline, leaving space for the variable parts.  */
1656 
1657 /* On the 386, the trampoline contains two instructions:
1658      mov #STATIC,ecx
1659      jmp FUNCTION
1660    The trampoline is generated entirely at runtime.  The operand of JMP
1661    is the address of FUNCTION relative to the instruction following the
1662    JMP (which is 5 bytes long).  */
1663 
1664 /* Length in units of the trampoline for entering a nested function.  */
1665 
1666 #define TRAMPOLINE_SIZE (TARGET_64BIT ? 24 : 10)
1667 
1668 /* Definitions for register eliminations.
1669 
1670    This is an array of structures.  Each structure initializes one pair
1671    of eliminable registers.  The "from" register number is given first,
1672    followed by "to".  Eliminations of the same "from" register are listed
1673    in order of preference.
1674 
1675    There are two registers that can always be eliminated on the i386.
1676    The frame pointer and the arg pointer can be replaced by either the
1677    hard frame pointer or to the stack pointer, depending upon the
1678    circumstances.  The hard frame pointer is not used before reload and
1679    so it is not eligible for elimination.  */
1680 
1681 #define ELIMINABLE_REGS					\
1682 {{ ARG_POINTER_REGNUM, STACK_POINTER_REGNUM},		\
1683  { ARG_POINTER_REGNUM, HARD_FRAME_POINTER_REGNUM},	\
1684  { FRAME_POINTER_REGNUM, STACK_POINTER_REGNUM},		\
1685  { FRAME_POINTER_REGNUM, HARD_FRAME_POINTER_REGNUM}}	\
1686 
1687 /* Define the offset between two registers, one to be eliminated, and the other
1688    its replacement, at the start of a routine.  */
1689 
1690 #define INITIAL_ELIMINATION_OFFSET(FROM, TO, OFFSET) \
1691   ((OFFSET) = ix86_initial_elimination_offset ((FROM), (TO)))
1692 
1693 /* Addressing modes, and classification of registers for them.  */
1694 
1695 /* Macros to check register numbers against specific register classes.  */
1696 
1697 /* These assume that REGNO is a hard or pseudo reg number.
1698    They give nonzero only if REGNO is a hard reg of the suitable class
1699    or a pseudo reg currently allocated to a suitable hard reg.
1700    Since they use reg_renumber, they are safe only once reg_renumber
1701    has been allocated, which happens in local-alloc.c.  */
1702 
1703 #define REGNO_OK_FOR_INDEX_P(REGNO) 					\
1704   ((REGNO) < STACK_POINTER_REGNUM 					\
1705    || REX_INT_REGNO_P (REGNO)						\
1706    || (unsigned) reg_renumber[(REGNO)] < STACK_POINTER_REGNUM		\
1707    || REX_INT_REGNO_P ((unsigned) reg_renumber[(REGNO)]))
1708 
1709 #define REGNO_OK_FOR_BASE_P(REGNO) 					\
1710   (GENERAL_REGNO_P (REGNO)						\
1711    || (REGNO) == ARG_POINTER_REGNUM 					\
1712    || (REGNO) == FRAME_POINTER_REGNUM 					\
1713    || GENERAL_REGNO_P ((unsigned) reg_renumber[(REGNO)]))
1714 
1715 /* The macros REG_OK_FOR..._P assume that the arg is a REG rtx
1716    and check its validity for a certain class.
1717    We have two alternate definitions for each of them.
1718    The usual definition accepts all pseudo regs; the other rejects
1719    them unless they have been allocated suitable hard regs.
1720    The symbol REG_OK_STRICT causes the latter definition to be used.
1721 
1722    Most source files want to accept pseudo regs in the hope that
1723    they will get allocated to the class that the insn wants them to be in.
1724    Source files for reload pass need to be strict.
1725    After reload, it makes no difference, since pseudo regs have
1726    been eliminated by then.  */
1727 
1728 
1729 /* Non strict versions, pseudos are ok.  */
1730 #define REG_OK_FOR_INDEX_NONSTRICT_P(X)					\
1731   (REGNO (X) < STACK_POINTER_REGNUM					\
1732    || REX_INT_REGNO_P (REGNO (X))					\
1733    || REGNO (X) >= FIRST_PSEUDO_REGISTER)
1734 
1735 #define REG_OK_FOR_BASE_NONSTRICT_P(X)					\
1736   (GENERAL_REGNO_P (REGNO (X))						\
1737    || REGNO (X) == ARG_POINTER_REGNUM					\
1738    || REGNO (X) == FRAME_POINTER_REGNUM 				\
1739    || REGNO (X) >= FIRST_PSEUDO_REGISTER)
1740 
1741 /* Strict versions, hard registers only */
1742 #define REG_OK_FOR_INDEX_STRICT_P(X) REGNO_OK_FOR_INDEX_P (REGNO (X))
1743 #define REG_OK_FOR_BASE_STRICT_P(X)  REGNO_OK_FOR_BASE_P (REGNO (X))
1744 
1745 #ifndef REG_OK_STRICT
1746 #define REG_OK_FOR_INDEX_P(X)  REG_OK_FOR_INDEX_NONSTRICT_P (X)
1747 #define REG_OK_FOR_BASE_P(X)   REG_OK_FOR_BASE_NONSTRICT_P (X)
1748 
1749 #else
1750 #define REG_OK_FOR_INDEX_P(X)  REG_OK_FOR_INDEX_STRICT_P (X)
1751 #define REG_OK_FOR_BASE_P(X)   REG_OK_FOR_BASE_STRICT_P (X)
1752 #endif
1753 
1754 /* TARGET_LEGITIMATE_ADDRESS_P recognizes an RTL expression
1755    that is a valid memory address for an instruction.
1756    The MODE argument is the machine mode for the MEM expression
1757    that wants to use this address.
1758 
1759    The other macros defined here are used only in TARGET_LEGITIMATE_ADDRESS_P,
1760    except for CONSTANT_ADDRESS_P which is usually machine-independent.
1761 
1762    See legitimize_pic_address in i386.c for details as to what
1763    constitutes a legitimate address when -fpic is used.  */
1764 
1765 #define MAX_REGS_PER_ADDRESS 2
1766 
1767 #define CONSTANT_ADDRESS_P(X)  constant_address_p (X)
1768 
1769 /* Nonzero if the constant value X is a legitimate general operand.
1770    It is given that X satisfies CONSTANT_P or is a CONST_DOUBLE.  */
1771 
1772 #define LEGITIMATE_CONSTANT_P(X)  legitimate_constant_p (X)
1773 
1774 /* Try a machine-dependent way of reloading an illegitimate address
1775    operand.  If we find one, push the reload and jump to WIN.  This
1776    macro is used in only one place: `find_reloads_address' in reload.c.  */
1777 
1778 #define LEGITIMIZE_RELOAD_ADDRESS(X, MODE, OPNUM, TYPE, INDL, WIN)	\
1779 do {									\
1780   if (ix86_legitimize_reload_address ((X), (MODE), (OPNUM),		\
1781 				      (int)(TYPE), (INDL)))		\
1782     goto WIN;								\
1783 } while (0)
1784 
1785 /* If defined, a C expression to determine the base term of address X.
1786    This macro is used in only one place: `find_base_term' in alias.c.
1787 
1788    It is always safe for this macro to not be defined.  It exists so
1789    that alias analysis can understand machine-dependent addresses.
1790 
1791    The typical use of this macro is to handle addresses containing
1792    a label_ref or symbol_ref within an UNSPEC.  */
1793 
1794 #define FIND_BASE_TERM(X) ix86_find_base_term (X)
1795 
1796 /* Nonzero if the constant value X is a legitimate general operand
1797    when generating PIC code.  It is given that flag_pic is on and
1798    that X satisfies CONSTANT_P or is a CONST_DOUBLE.  */
1799 
1800 #define LEGITIMATE_PIC_OPERAND_P(X) legitimate_pic_operand_p (X)
1801 
1802 #define SYMBOLIC_CONST(X)	\
1803   (GET_CODE (X) == SYMBOL_REF						\
1804    || GET_CODE (X) == LABEL_REF						\
1805    || (GET_CODE (X) == CONST && symbolic_reference_mentioned_p (X)))
1806 
1807 /* Max number of args passed in registers.  If this is more than 3, we will
1808    have problems with ebx (register #4), since it is a caller save register and
1809    is also used as the pic register in ELF.  So for now, don't allow more than
1810    3 registers to be passed in registers.  */
1811 
1812 /* Abi specific values for REGPARM_MAX and SSE_REGPARM_MAX */
1813 #define X86_64_REGPARM_MAX 6
1814 #define X86_64_MS_REGPARM_MAX 4
1815 
1816 #define X86_32_REGPARM_MAX 3
1817 
1818 #define REGPARM_MAX							\
1819   (TARGET_64BIT ? (TARGET_64BIT_MS_ABI ? X86_64_MS_REGPARM_MAX		\
1820 		   : X86_64_REGPARM_MAX)				\
1821    : X86_32_REGPARM_MAX)
1822 
1823 #define X86_64_SSE_REGPARM_MAX 8
1824 #define X86_64_MS_SSE_REGPARM_MAX 4
1825 
1826 #define X86_32_SSE_REGPARM_MAX (TARGET_SSE ? (TARGET_MACHO ? 4 : 3) : 0)
1827 
1828 #define SSE_REGPARM_MAX							\
1829   (TARGET_64BIT ? (TARGET_64BIT_MS_ABI ? X86_64_MS_SSE_REGPARM_MAX	\
1830 		   : X86_64_SSE_REGPARM_MAX)				\
1831    : X86_32_SSE_REGPARM_MAX)
1832 
1833 #define MMX_REGPARM_MAX (TARGET_64BIT ? 0 : (TARGET_MMX ? 3 : 0))
1834 
1835 
1836 /* Specify the machine mode that this machine uses
1837    for the index in the tablejump instruction.  */
1838 #define CASE_VECTOR_MODE \
1839  (!TARGET_64BIT || (flag_pic && ix86_cmodel != CM_LARGE_PIC) ? SImode : DImode)
1840 
1841 /* Define this as 1 if `char' should by default be signed; else as 0.  */
1842 #define DEFAULT_SIGNED_CHAR 1
1843 
1844 /* Max number of bytes we can move from memory to memory
1845    in one reasonably fast instruction.  */
1846 #define MOVE_MAX 16
1847 
1848 /* MOVE_MAX_PIECES is the number of bytes at a time which we can
1849    move efficiently, as opposed to  MOVE_MAX which is the maximum
1850    number of bytes we can move with a single instruction.  */
1851 #define MOVE_MAX_PIECES (TARGET_64BIT ? 8 : 4)
1852 
1853 /* If a memory-to-memory move would take MOVE_RATIO or more simple
1854    move-instruction pairs, we will do a movmem or libcall instead.
1855    Increasing the value will always make code faster, but eventually
1856    incurs high cost in increased code size.
1857 
1858    If you don't define this, a reasonable default is used.  */
1859 
1860 #define MOVE_RATIO(speed) ((speed) ? ix86_cost->move_ratio : 3)
1861 
1862 /* If a clear memory operation would take CLEAR_RATIO or more simple
1863    move-instruction sequences, we will do a clrmem or libcall instead.  */
1864 
1865 #define CLEAR_RATIO(speed) ((speed) ? MIN (6, ix86_cost->move_ratio) : 2)
1866 
1867 /* Define if shifts truncate the shift count
1868    which implies one can omit a sign-extension or zero-extension
1869    of a shift count.  */
1870 /* On i386, shifts do truncate the count.  But bit opcodes don't.  */
1871 
1872 /* #define SHIFT_COUNT_TRUNCATED */
1873 
1874 /* Value is 1 if truncating an integer of INPREC bits to OUTPREC bits
1875    is done just by pretending it is already truncated.  */
1876 #define TRULY_NOOP_TRUNCATION(OUTPREC, INPREC) 1
1877 
1878 /* A macro to update M and UNSIGNEDP when an object whose type is
1879    TYPE and which has the specified mode and signedness is to be
1880    stored in a register.  This macro is only called when TYPE is a
1881    scalar type.
1882 
1883    On i386 it is sometimes useful to promote HImode and QImode
1884    quantities to SImode.  The choice depends on target type.  */
1885 
1886 #define PROMOTE_MODE(MODE, UNSIGNEDP, TYPE) 		\
1887 do {							\
1888   if (((MODE) == HImode && TARGET_PROMOTE_HI_REGS)	\
1889       || ((MODE) == QImode && TARGET_PROMOTE_QI_REGS))	\
1890     (MODE) = SImode;					\
1891 } while (0)
1892 
1893 /* Specify the machine mode that pointers have.
1894    After generation of rtl, the compiler makes no further distinction
1895    between pointers and any other objects of this machine mode.  */
1896 #define Pmode (TARGET_64BIT ? DImode : SImode)
1897 
1898 /* A function address in a call instruction
1899    is a byte address (for indexing purposes)
1900    so give the MEM rtx a byte's mode.  */
1901 #define FUNCTION_MODE QImode
1902 
1903 /* A C expression for the cost of moving data from a register in class FROM to
1904    one in class TO.  The classes are expressed using the enumeration values
1905    such as `GENERAL_REGS'.  A value of 2 is the default; other values are
1906    interpreted relative to that.
1907 
1908    It is not required that the cost always equal 2 when FROM is the same as TO;
1909    on some machines it is expensive to move between registers if they are not
1910    general registers.  */
1911 
1912 #define REGISTER_MOVE_COST(MODE, CLASS1, CLASS2) \
1913    ix86_register_move_cost ((MODE), (CLASS1), (CLASS2))
1914 
1915 /* A C expression for the cost of moving data of mode M between a
1916    register and memory.  A value of 2 is the default; this cost is
1917    relative to those in `REGISTER_MOVE_COST'.
1918 
1919    If moving between registers and memory is more expensive than
1920    between two registers, you should define this macro to express the
1921    relative cost.  */
1922 
1923 #define MEMORY_MOVE_COST(MODE, CLASS, IN)	\
1924   ix86_memory_move_cost ((MODE), (CLASS), (IN))
1925 
1926 /* A C expression for the cost of a branch instruction.  A value of 1
1927    is the default; other values are interpreted relative to that.  */
1928 
1929 #define BRANCH_COST(speed_p, predictable_p) \
1930   (!(speed_p) ? 2 : (predictable_p) ? 0 : ix86_branch_cost)
1931 
1932 /* Define this macro as a C expression which is nonzero if accessing
1933    less than a word of memory (i.e. a `char' or a `short') is no
1934    faster than accessing a word of memory, i.e., if such access
1935    require more than one instruction or if there is no difference in
1936    cost between byte and (aligned) word loads.
1937 
1938    When this macro is not defined, the compiler will access a field by
1939    finding the smallest containing object; when it is defined, a
1940    fullword load will be used if alignment permits.  Unless bytes
1941    accesses are faster than word accesses, using word accesses is
1942    preferable since it may eliminate subsequent memory access if
1943    subsequent accesses occur to other fields in the same word of the
1944    structure, but to different bytes.  */
1945 
1946 #define SLOW_BYTE_ACCESS 0
1947 
1948 /* Nonzero if access to memory by shorts is slow and undesirable.  */
1949 #define SLOW_SHORT_ACCESS 0
1950 
1951 /* Define this macro to be the value 1 if unaligned accesses have a
1952    cost many times greater than aligned accesses, for example if they
1953    are emulated in a trap handler.
1954 
1955    When this macro is nonzero, the compiler will act as if
1956    `STRICT_ALIGNMENT' were nonzero when generating code for block
1957    moves.  This can cause significantly more instructions to be
1958    produced.  Therefore, do not set this macro nonzero if unaligned
1959    accesses only add a cycle or two to the time for a memory access.
1960 
1961    If the value of this macro is always zero, it need not be defined.  */
1962 
1963 /* #define SLOW_UNALIGNED_ACCESS(MODE, ALIGN) 0 */
1964 
1965 /* Define this macro if it is as good or better to call a constant
1966    function address than to call an address kept in a register.
1967 
1968    Desirable on the 386 because a CALL with a constant address is
1969    faster than one with a register address.  */
1970 
1971 #define NO_FUNCTION_CSE
1972 
1973 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
1974    return the mode to be used for the comparison.
1975 
1976    For floating-point equality comparisons, CCFPEQmode should be used.
1977    VOIDmode should be used in all other cases.
1978 
1979    For integer comparisons against zero, reduce to CCNOmode or CCZmode if
1980    possible, to allow for more combinations.  */
1981 
1982 #define SELECT_CC_MODE(OP, X, Y) ix86_cc_mode ((OP), (X), (Y))
1983 
1984 /* Return nonzero if MODE implies a floating point inequality can be
1985    reversed.  */
1986 
1987 #define REVERSIBLE_CC_MODE(MODE) 1
1988 
1989 /* A C expression whose value is reversed condition code of the CODE for
1990    comparison done in CC_MODE mode.  */
1991 #define REVERSE_CONDITION(CODE, MODE) ix86_reverse_condition ((CODE), (MODE))
1992 
1993 
1994 /* Control the assembler format that we output, to the extent
1995    this does not vary between assemblers.  */
1996 
1997 /* How to refer to registers in assembler output.
1998    This sequence is indexed by compiler's hard-register-number (see above).  */
1999 
2000 /* In order to refer to the first 8 regs as 32-bit regs, prefix an "e".
2001    For non floating point regs, the following are the HImode names.
2002 
2003    For float regs, the stack top is sometimes referred to as "%st(0)"
2004    instead of just "%st".  PRINT_OPERAND handles this with the "y" code.  */
2005 
2006 #define HI_REGISTER_NAMES						\
2007 {"ax","dx","cx","bx","si","di","bp","sp",				\
2008  "st","st(1)","st(2)","st(3)","st(4)","st(5)","st(6)","st(7)",		\
2009  "argp", "flags", "fpsr", "fpcr", "frame",				\
2010  "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7",		\
2011  "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",		\
2012  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",			\
2013  "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15"}
2014 
2015 #define REGISTER_NAMES HI_REGISTER_NAMES
2016 
2017 /* Table of additional register names to use in user input.  */
2018 
2019 #define ADDITIONAL_REGISTER_NAMES \
2020 { { "eax", 0 }, { "edx", 1 }, { "ecx", 2 }, { "ebx", 3 },	\
2021   { "esi", 4 }, { "edi", 5 }, { "ebp", 6 }, { "esp", 7 },	\
2022   { "rax", 0 }, { "rdx", 1 }, { "rcx", 2 }, { "rbx", 3 },	\
2023   { "rsi", 4 }, { "rdi", 5 }, { "rbp", 6 }, { "rsp", 7 },	\
2024   { "al", 0 }, { "dl", 1 }, { "cl", 2 }, { "bl", 3 },		\
2025   { "ah", 0 }, { "dh", 1 }, { "ch", 2 }, { "bh", 3 } }
2026 
2027 /* Note we are omitting these since currently I don't know how
2028 to get gcc to use these, since they want the same but different
2029 number as al, and ax.
2030 */
2031 
2032 #define QI_REGISTER_NAMES \
2033 {"al", "dl", "cl", "bl", "sil", "dil", "bpl", "spl",}
2034 
2035 /* These parallel the array above, and can be used to access bits 8:15
2036    of regs 0 through 3.  */
2037 
2038 #define QI_HIGH_REGISTER_NAMES \
2039 {"ah", "dh", "ch", "bh", }
2040 
2041 /* How to renumber registers for dbx and gdb.  */
2042 
2043 #define DBX_REGISTER_NUMBER(N) \
2044   (TARGET_64BIT ? dbx64_register_map[(N)] : dbx_register_map[(N)])
2045 
2046 extern int const dbx_register_map[FIRST_PSEUDO_REGISTER];
2047 extern int const dbx64_register_map[FIRST_PSEUDO_REGISTER];
2048 extern int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER];
2049 
2050 /* Before the prologue, RA is at 0(%esp).  */
2051 #define INCOMING_RETURN_ADDR_RTX \
2052   gen_rtx_MEM (VOIDmode, gen_rtx_REG (VOIDmode, STACK_POINTER_REGNUM))
2053 
2054 /* After the prologue, RA is at -4(AP) in the current frame.  */
2055 #define RETURN_ADDR_RTX(COUNT, FRAME)					   \
2056   ((COUNT) == 0								   \
2057    ? gen_rtx_MEM (Pmode, plus_constant (arg_pointer_rtx, -UNITS_PER_WORD)) \
2058    : gen_rtx_MEM (Pmode, plus_constant (FRAME, UNITS_PER_WORD)))
2059 
2060 /* PC is dbx register 8; let's use that column for RA.  */
2061 #define DWARF_FRAME_RETURN_COLUMN 	(TARGET_64BIT ? 16 : 8)
2062 
2063 /* Before the prologue, the top of the frame is at 4(%esp).  */
2064 #define INCOMING_FRAME_SP_OFFSET UNITS_PER_WORD
2065 
2066 /* Describe how we implement __builtin_eh_return.  */
2067 #define EH_RETURN_DATA_REGNO(N)	((N) < 2 ? (N) : INVALID_REGNUM)
2068 #define EH_RETURN_STACKADJ_RTX	gen_rtx_REG (Pmode, 2)
2069 
2070 
2071 /* Select a format to encode pointers in exception handling data.  CODE
2072    is 0 for data, 1 for code labels, 2 for function pointers.  GLOBAL is
2073    true if the symbol may be affected by dynamic relocations.
2074 
2075    ??? All x86 object file formats are capable of representing this.
2076    After all, the relocation needed is the same as for the call insn.
2077    Whether or not a particular assembler allows us to enter such, I
2078    guess we'll have to see.  */
2079 #define ASM_PREFERRED_EH_DATA_FORMAT(CODE, GLOBAL)       		\
2080   asm_preferred_eh_data_format ((CODE), (GLOBAL))
2081 
2082 /* This is how to output an insn to push a register on the stack.
2083    It need not be very fast code.  */
2084 
2085 #define ASM_OUTPUT_REG_PUSH(FILE, REGNO)  \
2086 do {									\
2087   if (TARGET_64BIT)							\
2088     asm_fprintf ((FILE), "\tpush{q}\t%%r%s\n",				\
2089 		 reg_names[(REGNO)] + (REX_INT_REGNO_P (REGNO) != 0));	\
2090   else									\
2091     asm_fprintf ((FILE), "\tpush{l}\t%%e%s\n", reg_names[(REGNO)]);	\
2092 } while (0)
2093 
2094 /* This is how to output an insn to pop a register from the stack.
2095    It need not be very fast code.  */
2096 
2097 #define ASM_OUTPUT_REG_POP(FILE, REGNO)  \
2098 do {									\
2099   if (TARGET_64BIT)							\
2100     asm_fprintf ((FILE), "\tpop{q}\t%%r%s\n",				\
2101 		 reg_names[(REGNO)] + (REX_INT_REGNO_P (REGNO) != 0));	\
2102   else									\
2103     asm_fprintf ((FILE), "\tpop{l}\t%%e%s\n", reg_names[(REGNO)]);	\
2104 } while (0)
2105 
2106 /* This is how to output an element of a case-vector that is absolute.  */
2107 
2108 #define ASM_OUTPUT_ADDR_VEC_ELT(FILE, VALUE)  \
2109   ix86_output_addr_vec_elt ((FILE), (VALUE))
2110 
2111 /* This is how to output an element of a case-vector that is relative.  */
2112 
2113 #define ASM_OUTPUT_ADDR_DIFF_ELT(FILE, BODY, VALUE, REL) \
2114   ix86_output_addr_diff_elt ((FILE), (VALUE), (REL))
2115 
2116 /* When we see %v, we will print the 'v' prefix if TARGET_AVX is
2117    true.  */
2118 
2119 #define ASM_OUTPUT_AVX_PREFIX(STREAM, PTR)	\
2120 {						\
2121   if ((PTR)[0] == '%' && (PTR)[1] == 'v')	\
2122     {						\
2123       if (TARGET_AVX)				\
2124 	(PTR) += 1;				\
2125       else					\
2126 	(PTR) += 2;				\
2127     }						\
2128 }
2129 
2130 /* A C statement or statements which output an assembler instruction
2131    opcode to the stdio stream STREAM.  The macro-operand PTR is a
2132    variable of type `char *' which points to the opcode name in
2133    its "internal" form--the form that is written in the machine
2134    description.  */
2135 
2136 #define ASM_OUTPUT_OPCODE(STREAM, PTR) \
2137   ASM_OUTPUT_AVX_PREFIX ((STREAM), (PTR))
2138 
2139 /* A C statement to output to the stdio stream FILE an assembler
2140    command to pad the location counter to a multiple of 1<<LOG
2141    bytes if it is within MAX_SKIP bytes.  */
2142 
2143 #ifdef HAVE_GAS_MAX_SKIP_P2ALIGN
2144 #undef  ASM_OUTPUT_MAX_SKIP_PAD
2145 #define ASM_OUTPUT_MAX_SKIP_PAD(FILE, LOG, MAX_SKIP)			\
2146   if ((LOG) != 0)							\
2147     {									\
2148       if ((MAX_SKIP) == 0)						\
2149         fprintf ((FILE), "\t.p2align %d\n", (LOG));			\
2150       else								\
2151         fprintf ((FILE), "\t.p2align %d,,%d\n", (LOG), (MAX_SKIP));	\
2152     }
2153 #endif
2154 
2155 /* Under some conditions we need jump tables in the text section,
2156    because the assembler cannot handle label differences between
2157    sections.  This is the case for x86_64 on Mach-O for example.  */
2158 
2159 #define JUMP_TABLES_IN_TEXT_SECTION \
2160   (flag_pic && ((TARGET_MACHO && TARGET_64BIT) \
2161    || (!TARGET_64BIT && !HAVE_AS_GOTOFF_IN_DATA)))
2162 
2163 /* Switch to init or fini section via SECTION_OP, emit a call to FUNC,
2164    and switch back.  For x86 we do this only to save a few bytes that
2165    would otherwise be unused in the text section.  */
2166 #define CRT_MKSTR2(VAL) #VAL
2167 #define CRT_MKSTR(x) CRT_MKSTR2(x)
2168 
2169 #define CRT_CALL_STATIC_FUNCTION(SECTION_OP, FUNC)		\
2170    asm (SECTION_OP "\n\t"					\
2171 	"call " CRT_MKSTR(__USER_LABEL_PREFIX__) #FUNC "\n"	\
2172 	TEXT_SECTION_ASM_OP);
2173 
2174 /* Print operand X (an rtx) in assembler syntax to file FILE.
2175    CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
2176    Effect of various CODE letters is described in i386.c near
2177    print_operand function.  */
2178 
2179 #define PRINT_OPERAND_PUNCT_VALID_P(CODE) \
2180   ((CODE) == '*' || (CODE) == '+' || (CODE) == '&' || (CODE) == ';')
2181 
2182 #define PRINT_OPERAND(FILE, X, CODE)  \
2183   print_operand ((FILE), (X), (CODE))
2184 
2185 #define PRINT_OPERAND_ADDRESS(FILE, ADDR)  \
2186   print_operand_address ((FILE), (ADDR))
2187 
2188 #define OUTPUT_ADDR_CONST_EXTRA(FILE, X, FAIL)	\
2189 do {						\
2190   if (! output_addr_const_extra (FILE, (X)))	\
2191     goto FAIL;					\
2192 } while (0);
2193 
2194 /* Which processor to schedule for. The cpu attribute defines a list that
2195    mirrors this list, so changes to i386.md must be made at the same time.  */
2196 
2197 enum processor_type
2198 {
2199   PROCESSOR_I386 = 0,			/* 80386 */
2200   PROCESSOR_I486,			/* 80486DX, 80486SX, 80486DX[24] */
2201   PROCESSOR_PENTIUM,
2202   PROCESSOR_PENTIUMPRO,
2203   PROCESSOR_GEODE,
2204   PROCESSOR_K6,
2205   PROCESSOR_ATHLON,
2206   PROCESSOR_PENTIUM4,
2207   PROCESSOR_K8,
2208   PROCESSOR_NOCONA,
2209   PROCESSOR_CORE2,
2210   PROCESSOR_GENERIC32,
2211   PROCESSOR_GENERIC64,
2212   PROCESSOR_AMDFAM10,
2213   PROCESSOR_ATOM,
2214   PROCESSOR_max
2215 };
2216 
2217 extern enum processor_type ix86_tune;
2218 extern enum processor_type ix86_arch;
2219 
2220 enum fpmath_unit
2221 {
2222   FPMATH_387 = 1,
2223   FPMATH_SSE = 2
2224 };
2225 
2226 extern enum fpmath_unit ix86_fpmath;
2227 
2228 enum tls_dialect
2229 {
2230   TLS_DIALECT_GNU,
2231   TLS_DIALECT_GNU2,
2232   TLS_DIALECT_SUN
2233 };
2234 
2235 extern enum tls_dialect ix86_tls_dialect;
2236 
2237 enum cmodel {
2238   CM_32,	/* The traditional 32-bit ABI.  */
2239   CM_SMALL,	/* Assumes all code and data fits in the low 31 bits.  */
2240   CM_KERNEL,	/* Assumes all code and data fits in the high 31 bits.  */
2241   CM_MEDIUM,	/* Assumes code fits in the low 31 bits; data unlimited.  */
2242   CM_LARGE,	/* No assumptions.  */
2243   CM_SMALL_PIC,	/* Assumes code+data+got/plt fits in a 31 bit region.  */
2244   CM_MEDIUM_PIC,/* Assumes code+got/plt fits in a 31 bit region.  */
2245   CM_LARGE_PIC	/* No assumptions.  */
2246 };
2247 
2248 extern enum cmodel ix86_cmodel;
2249 
2250 /* Size of the RED_ZONE area.  */
2251 #define RED_ZONE_SIZE 128
2252 /* Reserved area of the red zone for temporaries.  */
2253 #define RED_ZONE_RESERVE 8
2254 
2255 enum asm_dialect {
2256   ASM_ATT,
2257   ASM_INTEL
2258 };
2259 
2260 extern enum asm_dialect ix86_asm_dialect;
2261 extern unsigned int ix86_preferred_stack_boundary;
2262 extern unsigned int ix86_incoming_stack_boundary;
2263 extern int ix86_branch_cost, ix86_section_threshold;
2264 
2265 /* Smallest class containing REGNO.  */
2266 extern enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER];
2267 
2268 extern rtx ix86_compare_op0;	/* operand 0 for comparisons */
2269 extern rtx ix86_compare_op1;	/* operand 1 for comparisons */
2270 
2271 enum ix86_fpcmp_strategy {
2272   IX86_FPCMP_SAHF,
2273   IX86_FPCMP_COMI,
2274   IX86_FPCMP_ARITH
2275 };
2276 
2277 /* To properly truncate FP values into integers, we need to set i387 control
2278    word.  We can't emit proper mode switching code before reload, as spills
2279    generated by reload may truncate values incorrectly, but we still can avoid
2280    redundant computation of new control word by the mode switching pass.
2281    The fldcw instructions are still emitted redundantly, but this is probably
2282    not going to be noticeable problem, as most CPUs do have fast path for
2283    the sequence.
2284 
2285    The machinery is to emit simple truncation instructions and split them
2286    before reload to instructions having USEs of two memory locations that
2287    are filled by this code to old and new control word.
2288 
2289    Post-reload pass may be later used to eliminate the redundant fildcw if
2290    needed.  */
2291 
2292 enum ix86_entity
2293 {
2294   I387_TRUNC = 0,
2295   I387_FLOOR,
2296   I387_CEIL,
2297   I387_MASK_PM,
2298   MAX_386_ENTITIES
2299 };
2300 
2301 enum ix86_stack_slot
2302 {
2303   SLOT_VIRTUAL = 0,
2304   SLOT_TEMP,
2305   SLOT_CW_STORED,
2306   SLOT_CW_TRUNC,
2307   SLOT_CW_FLOOR,
2308   SLOT_CW_CEIL,
2309   SLOT_CW_MASK_PM,
2310   MAX_386_STACK_LOCALS
2311 };
2312 
2313 /* Define this macro if the port needs extra instructions inserted
2314    for mode switching in an optimizing compilation.  */
2315 
2316 #define OPTIMIZE_MODE_SWITCHING(ENTITY) \
2317    ix86_optimize_mode_switching[(ENTITY)]
2318 
2319 /* If you define `OPTIMIZE_MODE_SWITCHING', you have to define this as
2320    initializer for an array of integers.  Each initializer element N
2321    refers to an entity that needs mode switching, and specifies the
2322    number of different modes that might need to be set for this
2323    entity.  The position of the initializer in the initializer -
2324    starting counting at zero - determines the integer that is used to
2325    refer to the mode-switched entity in question.  */
2326 
2327 #define NUM_MODES_FOR_MODE_SWITCHING \
2328    { I387_CW_ANY, I387_CW_ANY, I387_CW_ANY, I387_CW_ANY }
2329 
2330 /* ENTITY is an integer specifying a mode-switched entity.  If
2331    `OPTIMIZE_MODE_SWITCHING' is defined, you must define this macro to
2332    return an integer value not larger than the corresponding element
2333    in `NUM_MODES_FOR_MODE_SWITCHING', to denote the mode that ENTITY
2334    must be switched into prior to the execution of INSN. */
2335 
2336 #define MODE_NEEDED(ENTITY, I) ix86_mode_needed ((ENTITY), (I))
2337 
2338 /* This macro specifies the order in which modes for ENTITY are
2339    processed.  0 is the highest priority.  */
2340 
2341 #define MODE_PRIORITY_TO_MODE(ENTITY, N) (N)
2342 
2343 /* Generate one or more insns to set ENTITY to MODE.  HARD_REG_LIVE
2344    is the set of hard registers live at the point where the insn(s)
2345    are to be inserted.  */
2346 
2347 #define EMIT_MODE_SET(ENTITY, MODE, HARD_REGS_LIVE) 			\
2348   ((MODE) != I387_CW_ANY && (MODE) != I387_CW_UNINITIALIZED		\
2349    ? emit_i387_cw_initialization (MODE), 0				\
2350    : 0)
2351 
2352 
2353 /* Avoid renaming of stack registers, as doing so in combination with
2354    scheduling just increases amount of live registers at time and in
2355    the turn amount of fxch instructions needed.
2356 
2357    ??? Maybe Pentium chips benefits from renaming, someone can try....  */
2358 
2359 #define HARD_REGNO_RENAME_OK(SRC, TARGET)  \
2360   (! IN_RANGE ((SRC), FIRST_STACK_REG, LAST_STACK_REG))
2361 
2362 
2363 #define FASTCALL_PREFIX '@'
2364 
2365 /* Machine specific CFA tracking during prologue/epilogue generation.  */
2366 
2367 #ifndef USED_FOR_TARGET
2368 struct GTY(()) machine_cfa_state
2369 {
2370   rtx reg;
2371   HOST_WIDE_INT offset;
2372 };
2373 
2374 struct GTY(()) machine_function {
2375   struct stack_local_entry *stack_locals;
2376   const char *some_ld_name;
2377   int varargs_gpr_size;
2378   int varargs_fpr_size;
2379   int optimize_mode_switching[MAX_386_ENTITIES];
2380 
2381   /* Number of saved registers USE_FAST_PROLOGUE_EPILOGUE
2382      has been computed for.  */
2383   int use_fast_prologue_epilogue_nregs;
2384 
2385   /* The CFA state at the end of the prologue.  */
2386   struct machine_cfa_state cfa;
2387 
2388   /* This value is used for amd64 targets and specifies the current abi
2389      to be used. MS_ABI means ms abi. Otherwise SYSV_ABI means sysv abi.  */
2390   enum calling_abi call_abi;
2391 
2392   /* Nonzero if the function accesses a previous frame.  */
2393   BOOL_BITFIELD accesses_prev_frame : 1;
2394 
2395   /* Nonzero if the function requires a CLD in the prologue.  */
2396   BOOL_BITFIELD needs_cld : 1;
2397 
2398   /* Set by ix86_compute_frame_layout and used by prologue/epilogue
2399      expander to determine the style used.  */
2400   BOOL_BITFIELD use_fast_prologue_epilogue : 1;
2401 
2402   /* If true, the current function needs the default PIC register, not
2403      an alternate register (on x86) and must not use the red zone (on
2404      x86_64), even if it's a leaf function.  We don't want the
2405      function to be regarded as non-leaf because TLS calls need not
2406      affect register allocation.  This flag is set when a TLS call
2407      instruction is expanded within a function, and never reset, even
2408      if all such instructions are optimized away.  Use the
2409      ix86_current_function_calls_tls_descriptor macro for a better
2410      approximation.  */
2411   BOOL_BITFIELD tls_descriptor_call_expanded_p : 1;
2412 
2413   /* If true, the current function has a STATIC_CHAIN is placed on the
2414      stack below the return address.  */
2415   BOOL_BITFIELD static_chain_on_stack : 1;
2416 };
2417 #endif
2418 
2419 #define ix86_stack_locals (cfun->machine->stack_locals)
2420 #define ix86_varargs_gpr_size (cfun->machine->varargs_gpr_size)
2421 #define ix86_varargs_fpr_size (cfun->machine->varargs_fpr_size)
2422 #define ix86_optimize_mode_switching (cfun->machine->optimize_mode_switching)
2423 #define ix86_current_function_needs_cld (cfun->machine->needs_cld)
2424 #define ix86_tls_descriptor_calls_expanded_in_cfun \
2425   (cfun->machine->tls_descriptor_call_expanded_p)
2426 /* Since tls_descriptor_call_expanded is not cleared, even if all TLS
2427    calls are optimized away, we try to detect cases in which it was
2428    optimized away.  Since such instructions (use (reg REG_SP)), we can
2429    verify whether there's any such instruction live by testing that
2430    REG_SP is live.  */
2431 #define ix86_current_function_calls_tls_descriptor \
2432   (ix86_tls_descriptor_calls_expanded_in_cfun && df_regs_ever_live_p (SP_REG))
2433 #define ix86_cfa_state (&cfun->machine->cfa)
2434 #define ix86_static_chain_on_stack (cfun->machine->static_chain_on_stack)
2435 
2436 /* Control behavior of x86_file_start.  */
2437 #define X86_FILE_START_VERSION_DIRECTIVE false
2438 #define X86_FILE_START_FLTUSED false
2439 
2440 /* Flag to mark data that is in the large address area.  */
2441 #define SYMBOL_FLAG_FAR_ADDR		(SYMBOL_FLAG_MACH_DEP << 0)
2442 #define SYMBOL_REF_FAR_ADDR_P(X)	\
2443 	((SYMBOL_REF_FLAGS (X) & SYMBOL_FLAG_FAR_ADDR) != 0)
2444 
2445 /* Flags to mark dllimport/dllexport.  Used by PE ports, but handy to
2446    have defined always, to avoid ifdefing.  */
2447 #define SYMBOL_FLAG_DLLIMPORT		(SYMBOL_FLAG_MACH_DEP << 1)
2448 #define SYMBOL_REF_DLLIMPORT_P(X) \
2449 	((SYMBOL_REF_FLAGS (X) & SYMBOL_FLAG_DLLIMPORT) != 0)
2450 
2451 #define SYMBOL_FLAG_DLLEXPORT		(SYMBOL_FLAG_MACH_DEP << 2)
2452 #define SYMBOL_REF_DLLEXPORT_P(X) \
2453 	((SYMBOL_REF_FLAGS (X) & SYMBOL_FLAG_DLLEXPORT) != 0)
2454 
2455 /* Model costs for vectorizer.  */
2456 
2457 /* Cost of conditional branch.  */
2458 #undef TARG_COND_BRANCH_COST
2459 #define TARG_COND_BRANCH_COST           ix86_cost->branch_cost
2460 
2461 /* Enum through the target specific extra va_list types.
2462    Please, do not iterate the base va_list type name.  */
2463 #define TARGET_ENUM_VA_LIST(IDX, PNAME, PTYPE) \
2464   (TARGET_64BIT ? ix86_enum_va_list (IDX, PNAME, PTYPE) : 0)
2465 
2466 /* Cost of any scalar operation, excluding load and store.  */
2467 #undef TARG_SCALAR_STMT_COST
2468 #define TARG_SCALAR_STMT_COST           ix86_cost->scalar_stmt_cost
2469 
2470 /* Cost of scalar load.  */
2471 #undef TARG_SCALAR_LOAD_COST
2472 #define TARG_SCALAR_LOAD_COST           ix86_cost->scalar_load_cost
2473 
2474 /* Cost of scalar store.  */
2475 #undef TARG_SCALAR_STORE_COST
2476 #define TARG_SCALAR_STORE_COST          ix86_cost->scalar_store_cost
2477 
2478 /* Cost of any vector operation, excluding load, store or vector to scalar
2479    operation.  */
2480 #undef TARG_VEC_STMT_COST
2481 #define TARG_VEC_STMT_COST              ix86_cost->vec_stmt_cost
2482 
2483 /* Cost of vector to scalar operation.  */
2484 #undef TARG_VEC_TO_SCALAR_COST
2485 #define TARG_VEC_TO_SCALAR_COST         ix86_cost->vec_to_scalar_cost
2486 
2487 /* Cost of scalar to vector operation.  */
2488 #undef TARG_SCALAR_TO_VEC_COST
2489 #define TARG_SCALAR_TO_VEC_COST         ix86_cost->scalar_to_vec_cost
2490 
2491 /* Cost of aligned vector load.  */
2492 #undef TARG_VEC_LOAD_COST
2493 #define TARG_VEC_LOAD_COST              ix86_cost->vec_align_load_cost
2494 
2495 /* Cost of misaligned vector load.  */
2496 #undef TARG_VEC_UNALIGNED_LOAD_COST
2497 #define TARG_VEC_UNALIGNED_LOAD_COST    ix86_cost->vec_unalign_load_cost
2498 
2499 /* Cost of vector store.  */
2500 #undef TARG_VEC_STORE_COST
2501 #define TARG_VEC_STORE_COST             ix86_cost->vec_store_cost
2502 
2503 /* Cost of conditional taken branch for vectorizer cost model.  */
2504 #undef TARG_COND_TAKEN_BRANCH_COST
2505 #define TARG_COND_TAKEN_BRANCH_COST     ix86_cost->cond_taken_branch_cost
2506 
2507 /* Cost of conditional not taken branch for vectorizer cost model.  */
2508 #undef TARG_COND_NOT_TAKEN_BRANCH_COST
2509 #define TARG_COND_NOT_TAKEN_BRANCH_COST ix86_cost->cond_not_taken_branch_cost
2510 
2511 /*
2512 Local variables:
2513 version-control: t
2514 End:
2515 */
2516