1 /*
2  * Copyright (C) 1995-2010 University of Karlsruhe.  All right reserved.
3  *
4  * This file is part of libFirm.
5  *
6  * This file may be distributed and/or modified under the terms of the
7  * GNU General Public License version 2 as published by the Free Software
8  * Foundation and appearing in the file LICENSE.GPL included in the
9  * packaging of this file.
10  *
11  * Licensees holding valid libFirm Professional Edition licenses may use
12  * this file in accordance with the libFirm Commercial License.
13  * Agreement provided with the Software.
14  *
15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
17  * PURPOSE.
18  */
19 
20 /**
21  * @file
22  * @brief       ia32 architecture variants
23  * @author      Michael Beck, Matthias Braun
24  */
25 #include "config.h"
26 
27 #include <stdbool.h>
28 #include "lc_opts.h"
29 #include "lc_opts_enum.h"
30 
31 #include "irtools.h"
32 
33 #include "bearch_ia32_t.h"
34 #include "ia32_architecture.h"
35 
36 #undef NATIVE_X86
37 
38 #ifdef _MSC_VER
39 #if defined(_M_IX86) || defined(_M_X64)
40 #include <intrin.h>
41 #define NATIVE_X86
42 #endif
43 #else
44 #if defined(__i386__) || defined(__x86_64__)
45 #define NATIVE_X86
46 #endif
47 #endif
48 
49 ia32_code_gen_config_t  ia32_cg_config;
50 
51 /**
52  * CPU architectures and features.
53  */
54 typedef enum cpu_arch_features {
55 	arch_generic32        = 0x00000001, /**< no specific architecture */
56 
57 	arch_i386             = 0x00000002, /**< i386 architecture */
58 	arch_i486             = 0x00000004, /**< i486 architecture */
59 	arch_pentium          = 0x00000008, /**< Pentium architecture */
60 	arch_ppro             = 0x00000010, /**< PentiumPro architecture */
61 	arch_netburst         = 0x00000020, /**< Netburst architecture */
62 	arch_nocona           = 0x00000040, /**< Nocona architecture */
63 	arch_core2            = 0x00000080, /**< Core2 architecture */
64 	arch_atom             = 0x00000100, /**< Atom architecture */
65 
66 	arch_k6               = 0x00000200, /**< k6 architecture */
67 	arch_geode            = 0x00000400, /**< Geode architecture */
68 	arch_athlon           = 0x00000800, /**< Athlon architecture */
69 	arch_k8               = 0x00001000, /**< K8/Opteron architecture */
70 	arch_k10              = 0x00002000, /**< K10/Barcelona architecture */
71 
72 	arch_mask             = 0x00003FFF,
73 
74 	arch_athlon_plus      = arch_athlon | arch_k8 | arch_k10,
75 	arch_all_amd          = arch_k6 | arch_geode | arch_athlon_plus,
76 
77 	arch_feature_mmx      = 0x00004000, /**< MMX instructions */
78 	arch_feature_cmov     = 0x00008000, /**< cmov instructions */
79 	arch_feature_p6_insn  = 0x00010000, /**< PentiumPro instructions */
80 	arch_feature_sse1     = 0x00020000, /**< SSE1 instructions */
81 	arch_feature_sse2     = 0x00040000, /**< SSE2 instructions */
82 	arch_feature_sse3     = 0x00080000, /**< SSE3 instructions */
83 	arch_feature_ssse3    = 0x00100000, /**< SSSE3 instructions */
84 	arch_feature_3DNow    = 0x00200000, /**< 3DNow! instructions */
85 	arch_feature_3DNowE   = 0x00400000, /**< Enhanced 3DNow! instructions */
86 	arch_feature_64bit    = 0x00800000, /**< x86_64 support */
87 	arch_feature_sse4_1   = 0x01000000, /**< SSE4.1 instructions */
88 	arch_feature_sse4_2   = 0x02000000, /**< SSE4.2 instructions */
89 	arch_feature_sse4a    = 0x04000000, /**< SSE4a instructions */
90 	arch_feature_popcnt   = 0x08000000, /**< popcnt instruction */
91 
92 	arch_mmx_insn     = arch_feature_mmx,                         /**< MMX instructions */
93 	arch_sse1_insn    = arch_feature_sse1   | arch_mmx_insn,      /**< SSE1 instructions, include MMX */
94 	arch_sse2_insn    = arch_feature_sse2   | arch_sse1_insn,     /**< SSE2 instructions, include SSE1 */
95 	arch_sse3_insn    = arch_feature_sse3   | arch_sse2_insn,     /**< SSE3 instructions, include SSE2 */
96 	arch_ssse3_insn   = arch_feature_ssse3  | arch_sse3_insn,     /**< SSSE3 instructions, include SSE3 */
97 	arch_sse4_1_insn  = arch_feature_sse4_1 | arch_ssse3_insn,    /**< SSE4.1 instructions, include SSSE3 */
98 	arch_sse4_2_insn  = arch_feature_sse4_2 | arch_sse4_1_insn,   /**< SSE4.2 instructions, include SSE4.1 */
99 	arch_sse4a_insn   = arch_feature_sse4a  | arch_ssse3_insn,    /**< SSE4a instructions, include SSSE3 */
100 
101 	arch_3DNow_insn   = arch_feature_3DNow  | arch_feature_mmx,   /**< 3DNow! instructions, including MMX */
102 	arch_3DNowE_insn  = arch_feature_3DNowE | arch_3DNow_insn,    /**< Enhanced 3DNow! instructions */
103 	arch_64bit_insn   = arch_feature_64bit  | arch_sse2_insn,     /**< x86_64 support, includes SSE2 */
104 
105 	cpu_generic             = arch_generic32,
106 
107 	/* intel CPUs */
108 	cpu_i386                = arch_i386,
109 	cpu_i486                = arch_i486,
110 	cpu_pentium             = arch_pentium,
111 	cpu_pentium_mmx         = arch_pentium | arch_mmx_insn,
112 	cpu_pentium_pro_generic = arch_ppro | arch_feature_p6_insn,
113 	cpu_pentium_pro         = arch_ppro | arch_feature_cmov | arch_feature_p6_insn,
114 	cpu_pentium_2           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_mmx_insn,
115 	cpu_pentium_3           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse1_insn,
116 	cpu_pentium_m           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse2_insn,
117 	cpu_netburst_generic    = arch_netburst | arch_feature_p6_insn,
118 	cpu_pentium_4           = arch_netburst | arch_feature_cmov | arch_feature_p6_insn | arch_sse2_insn,
119 	cpu_prescott            = arch_nocona | arch_feature_cmov | arch_feature_p6_insn | arch_sse3_insn,
120 	cpu_nocona              = arch_nocona | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
121 	cpu_core2_generic       = arch_core2 | arch_feature_p6_insn,
122 	cpu_core2               = arch_core2 | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_ssse3_insn,
123 	cpu_penryn              = arch_core2 | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse4_1_insn,
124 	cpu_atom_generic        = arch_atom | arch_feature_p6_insn,
125 	cpu_atom                = arch_atom | arch_feature_cmov | arch_feature_p6_insn | arch_ssse3_insn,
126 
127 	/* AMD CPUs */
128 	cpu_k6_generic     = arch_k6,
129 	cpu_k6             = arch_k6 | arch_mmx_insn,
130 	cpu_k6_PLUS        = arch_k6 | arch_3DNow_insn,
131 	cpu_geode_generic  = arch_geode,
132 	cpu_geode          = arch_geode  | arch_sse1_insn | arch_3DNowE_insn,
133 	cpu_athlon_generic = arch_athlon | arch_feature_p6_insn,
134 	cpu_athlon_old     = arch_athlon | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn,
135 	cpu_athlon         = arch_athlon | arch_sse1_insn | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn,
136 	cpu_athlon64       = arch_athlon | arch_sse2_insn | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn,
137 	cpu_k8_generic     = arch_k8  | arch_feature_p6_insn,
138 	cpu_k8             = arch_k8  | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn,
139 	cpu_k8_sse3        = arch_k8  | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
140 	cpu_k10_generic    = arch_k10 | arch_feature_p6_insn,
141 	cpu_k10            = arch_k10 | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_feature_popcnt | arch_64bit_insn | arch_sse4a_insn,
142 
143 	/* other CPUs */
144 	cpu_winchip_c6  = arch_i486 | arch_feature_mmx,
145 	cpu_winchip2    = arch_i486 | arch_feature_mmx | arch_feature_3DNow,
146 	cpu_c3          = arch_i486 | arch_feature_mmx | arch_feature_3DNow,
147 	cpu_c3_2        = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse1_insn, /* really no 3DNow! */
148 
149 	cpu_autodetect  = 0,
150 } cpu_arch_features;
151 ENUM_BITSET(cpu_arch_features)
152 
153 static int               opt_size             = 0;
154 static int               emit_machcode        = 0;
155 static cpu_arch_features arch                 = cpu_generic;
156 static cpu_arch_features opt_arch             = cpu_generic;
157 static int               fpu_arch             = 0;
158 static int               opt_cc               = 1;
159 static int               opt_unsafe_floatconv = 0;
160 
161 /* instruction set architectures. */
162 static const lc_opt_enum_int_items_t arch_items[] = {
163 	{ "i386",         cpu_i386 },
164 	{ "i486",         cpu_i486 },
165 	{ "i586",         cpu_pentium },
166 	{ "pentium",      cpu_pentium },
167 	{ "pentium-mmx",  cpu_pentium_mmx },
168 	{ "i686",         cpu_pentium_pro },
169 	{ "pentiumpro",   cpu_pentium_pro },
170 	{ "pentium2",     cpu_pentium_2 },
171 	{ "p2",           cpu_pentium_2 },
172 	{ "pentium3",     cpu_pentium_3 },
173 	{ "pentium3m",    cpu_pentium_3 },
174 	{ "p3",           cpu_pentium_3 },
175 	{ "pentium-m",    cpu_pentium_m },
176 	{ "pm",           cpu_pentium_m },
177 	{ "pentium4",     cpu_pentium_4 },
178 	{ "pentium4m",    cpu_pentium_4 },
179 	{ "p4",           cpu_pentium_4 },
180 	{ "prescott",     cpu_prescott },
181 	{ "nocona",       cpu_nocona },
182 	{ "merom",        cpu_core2 },
183 	{ "core2",        cpu_core2 },
184 	{ "penryn",       cpu_penryn },
185 	{ "atom",         cpu_atom },
186 
187 	{ "k6",           cpu_k6 },
188 	{ "k6-2",         cpu_k6_PLUS },
189 	{ "k6-3",         cpu_k6_PLUS },
190 	{ "geode",        cpu_geode },
191 	{ "athlon",       cpu_athlon_old },
192 	{ "athlon-tbird", cpu_athlon },
193 	{ "athlon-4",     cpu_athlon },
194 	{ "athlon-xp",    cpu_athlon },
195 	{ "athlon-mp",    cpu_athlon },
196 	{ "athlon64",     cpu_athlon64 },
197 	{ "k8",           cpu_k8 },
198 	{ "opteron",      cpu_k8 },
199 	{ "athlon-fx",    cpu_k8 },
200 	{ "k8-sse3",      cpu_k8_sse3 },
201 	{ "opteron-sse3", cpu_k8_sse3 },
202 	{ "k10",          cpu_k10 },
203 	{ "barcelona",    cpu_k10 },
204 	{ "amdfam10",     cpu_k10 },
205 
206 	{ "winchip-c6",   cpu_winchip_c6, },
207 	{ "winchip2",     cpu_winchip2 },
208 	{ "c3",           cpu_c3 },
209 	{ "c3-2",         cpu_c3_2 },
210 
211 	{ "generic",      cpu_generic },
212 	{ "generic32",    cpu_generic },
213 
214 #ifdef NATIVE_X86
215 	{ "native",       cpu_autodetect },
216 #endif
217 
218 	{ NULL,           0 }
219 };
220 
221 static lc_opt_enum_int_var_t arch_var = {
222 	(int*) &arch, arch_items
223 };
224 
225 static lc_opt_enum_int_var_t opt_arch_var = {
226 	(int*) &opt_arch, arch_items
227 };
228 
229 static const lc_opt_enum_int_items_t fp_unit_items[] = {
230 	{ "x87" ,      IA32_FPU_ARCH_X87 },
231 	{ "sse2",      IA32_FPU_ARCH_SSE2 },
232 	{ "softfloat", IA32_FPU_ARCH_SOFTFLOAT },
233 	{ NULL,        IA32_FPU_ARCH_NONE }
234 };
235 
236 static lc_opt_enum_int_var_t fp_unit_var = {
237 	&fpu_arch, fp_unit_items
238 };
239 
240 static const lc_opt_table_entry_t ia32_architecture_options[] = {
241 	LC_OPT_ENT_BOOL    ("size",             "optimize for size",                                  &opt_size),
242 	LC_OPT_ENT_ENUM_INT("arch",             "select the instruction architecture",                &arch_var),
243 	LC_OPT_ENT_ENUM_INT("opt",              "optimize for instruction architecture",              &opt_arch_var),
244 	LC_OPT_ENT_ENUM_INT("fpunit",           "select the floating point unit",                     &fp_unit_var),
245 	LC_OPT_ENT_NEGBOOL ("nooptcc",          "do not optimize calling convention",                 &opt_cc),
246 	LC_OPT_ENT_BOOL    ("unsafe_floatconv", "do unsafe floating point controlword optimisations", &opt_unsafe_floatconv),
247 	LC_OPT_ENT_BOOL    ("machcode",         "output machine code instead of assembler",           &emit_machcode),
248 	LC_OPT_LAST
249 };
250 
251 typedef struct insn_const {
252 	int      add_cost;                 /**< cost of an add instruction */
253 	int      lea_cost;                 /**< cost of a lea instruction */
254 	int      const_shf_cost;           /**< cost of a constant shift instruction */
255 	int      cost_mul_start;           /**< starting cost of a multiply instruction */
256 	int      cost_mul_bit;             /**< cost of multiply for every set bit */
257 	unsigned function_alignment;       /**< logarithm for alignment of function labels */
258 	unsigned label_alignment;          /**< logarithm for alignment of loops labels */
259 	unsigned label_alignment_max_skip; /**< maximum skip for alignment of loops labels */
260 } insn_const;
261 
262 /* costs for optimizing for size */
263 static const insn_const size_cost = {
264 	2,   /* cost of an add instruction */
265 	3,   /* cost of a lea instruction */
266 	3,   /* cost of a constant shift instruction */
267 	4,   /* starting cost of a multiply instruction */
268 	0,   /* cost of multiply for every set bit */
269 	0,   /* logarithm for alignment of function labels */
270 	0,   /* logarithm for alignment of loops labels */
271 	0,   /* maximum skip for alignment of loops labels */
272 };
273 
274 /* costs for the i386 */
275 static const insn_const i386_cost = {
276 	1,   /* cost of an add instruction */
277 	1,   /* cost of a lea instruction */
278 	3,   /* cost of a constant shift instruction */
279 	9,   /* starting cost of a multiply instruction */
280 	1,   /* cost of multiply for every set bit */
281 	2,   /* logarithm for alignment of function labels */
282 	2,   /* logarithm for alignment of loops labels */
283 	3,   /* maximum skip for alignment of loops labels */
284 };
285 
286 /* costs for the i486 */
287 static const insn_const i486_cost = {
288 	1,   /* cost of an add instruction */
289 	1,   /* cost of a lea instruction */
290 	2,   /* cost of a constant shift instruction */
291 	12,  /* starting cost of a multiply instruction */
292 	1,   /* cost of multiply for every set bit */
293 	4,   /* logarithm for alignment of function labels */
294 	4,   /* logarithm for alignment of loops labels */
295 	15,  /* maximum skip for alignment of loops labels */
296 };
297 
298 /* costs for the Pentium */
299 static const insn_const pentium_cost = {
300 	1,   /* cost of an add instruction */
301 	1,   /* cost of a lea instruction */
302 	1,   /* cost of a constant shift instruction */
303 	11,  /* starting cost of a multiply instruction */
304 	0,   /* cost of multiply for every set bit */
305 	4,   /* logarithm for alignment of function labels */
306 	4,   /* logarithm for alignment of loops labels */
307 	7,   /* maximum skip for alignment of loops labels */
308 };
309 
310 /* costs for the Pentium Pro */
311 static const insn_const pentiumpro_cost = {
312 	1,   /* cost of an add instruction */
313 	1,   /* cost of a lea instruction */
314 	1,   /* cost of a constant shift instruction */
315 	4,   /* starting cost of a multiply instruction */
316 	0,   /* cost of multiply for every set bit */
317 	4,   /* logarithm for alignment of function labels */
318 	4,   /* logarithm for alignment of loops labels */
319 	10,  /* maximum skip for alignment of loops labels */
320 };
321 
322 /* costs for the K6 */
323 static const insn_const k6_cost = {
324 	1,   /* cost of an add instruction */
325 	2,   /* cost of a lea instruction */
326 	1,   /* cost of a constant shift instruction */
327 	3,   /* starting cost of a multiply instruction */
328 	0,   /* cost of multiply for every set bit */
329 	5,   /* logarithm for alignment of function labels */
330 	5,   /* logarithm for alignment of loops labels */
331 	7,   /* maximum skip for alignment of loops labels */
332 };
333 
334 /* costs for the Geode */
335 static const insn_const geode_cost = {
336 	1,   /* cost of an add instruction */
337 	1,   /* cost of a lea instruction */
338 	1,   /* cost of a constant shift instruction */
339 	7,   /* starting cost of a multiply instruction */
340 	0,   /* cost of multiply for every set bit */
341 	0,   /* logarithm for alignment of function labels */
342 	0,   /* logarithm for alignment of loops labels */
343 	0,   /* maximum skip for alignment of loops labels */
344 };
345 
346 /* costs for the Athlon */
347 static const insn_const athlon_cost = {
348 	1,   /* cost of an add instruction */
349 	2,   /* cost of a lea instruction */
350 	1,   /* cost of a constant shift instruction */
351 	5,   /* starting cost of a multiply instruction */
352 	0,   /* cost of multiply for every set bit */
353 	4,   /* logarithm for alignment of function labels */
354 	4,   /* logarithm for alignment of loops labels */
355 	7,   /* maximum skip for alignment of loops labels */
356 };
357 
358 /* costs for the Opteron/K8 */
359 static const insn_const k8_cost = {
360 	1,   /* cost of an add instruction */
361 	2,   /* cost of a lea instruction */
362 	1,   /* cost of a constant shift instruction */
363 	3,   /* starting cost of a multiply instruction */
364 	0,   /* cost of multiply for every set bit */
365 #if 0 /* TEST */
366 	4,   /* logarithm for alignment of function labels */
367 	4,   /* logarithm for alignment of loops labels */
368 	7,   /* maximum skip for alignment of loops labels */
369 #else
370 	0,
371 	0,
372 	0
373 #endif
374 };
375 
376 /* costs for the K10 */
377 static const insn_const k10_cost = {
378 	1,   /* cost of an add instruction */
379 	2,   /* cost of a lea instruction */
380 	1,   /* cost of a constant shift instruction */
381 	3,   /* starting cost of a multiply instruction */
382 	0,   /* cost of multiply for every set bit */
383 	5,   /* logarithm for alignment of function labels */
384 	5,   /* logarithm for alignment of loops labels */
385 	7,   /* maximum skip for alignment of loops labels */
386 };
387 
388 /* costs for the Pentium 4 */
389 static const insn_const netburst_cost = {
390 	1,   /* cost of an add instruction */
391 	3,   /* cost of a lea instruction */
392 	4,   /* cost of a constant shift instruction */
393 	15,  /* starting cost of a multiply instruction */
394 	0,   /* cost of multiply for every set bit */
395 	4,   /* logarithm for alignment of function labels */
396 	4,   /* logarithm for alignment of loops labels */
397 	7,   /* maximum skip for alignment of loops labels */
398 };
399 
400 /* costs for the Nocona and Core */
401 static const insn_const nocona_cost = {
402 	1,   /* cost of an add instruction */
403 	1,   /* cost of a lea instruction */
404 	1,   /* cost of a constant shift instruction */
405 	10,  /* starting cost of a multiply instruction */
406 	0,   /* cost of multiply for every set bit */
407 	4,   /* logarithm for alignment of function labels */
408 	4,   /* logarithm for alignment of loops labels */
409 	7,   /* maximum skip for alignment of loops labels */
410 };
411 
412 /* costs for the Core2 */
413 static const insn_const core2_cost = {
414 	1,   /* cost of an add instruction */
415 	1,   /* cost of a lea instruction */
416 	1,   /* cost of a constant shift instruction */
417 	3,   /* starting cost of a multiply instruction */
418 	0,   /* cost of multiply for every set bit */
419 	4,   /* logarithm for alignment of function labels */
420 	4,   /* logarithm for alignment of loops labels */
421 	10,  /* maximum skip for alignment of loops labels */
422 };
423 
424 /* costs for the generic32 */
425 static const insn_const generic32_cost = {
426 	1,   /* cost of an add instruction */
427 	2,   /* cost of a lea instruction */
428 	1,   /* cost of a constant shift instruction */
429 	4,   /* starting cost of a multiply instruction */
430 	0,   /* cost of multiply for every set bit */
431 	4,   /* logarithm for alignment of function labels */
432 	4,   /* logarithm for alignment of loops labels */
433 	7,   /* maximum skip for alignment of loops labels */
434 };
435 
436 static const insn_const *arch_costs = &generic32_cost;
437 
set_arch_costs(void)438 static void set_arch_costs(void)
439 {
440 	if (opt_size) {
441 		arch_costs = &size_cost;
442 		return;
443 	}
444 	switch (opt_arch & arch_mask) {
445 	case arch_i386:      arch_costs = &i386_cost;       break;
446 	case arch_i486:      arch_costs = &i486_cost;       break;
447 	case arch_pentium:   arch_costs = &pentium_cost;    break;
448 	case arch_ppro:      arch_costs = &pentiumpro_cost; break;
449 	case arch_netburst:  arch_costs = &netburst_cost;   break;
450 	case arch_nocona:    arch_costs = &nocona_cost;     break;
451 	case arch_core2:     arch_costs = &core2_cost;      break;
452 	case arch_k6:        arch_costs = &k6_cost;         break;
453 	case arch_geode:     arch_costs = &geode_cost;      break;
454 	case arch_athlon:    arch_costs = &athlon_cost;     break;
455 	case arch_k8:        arch_costs = &k8_cost;         break;
456 	case arch_k10:       arch_costs = &k10_cost;        break;
457 	default:
458 	case arch_generic32: arch_costs = &generic32_cost;  break;
459 	}
460 }
461 
462 /* Evaluate the costs of an instruction. */
ia32_evaluate_insn(insn_kind kind,const ir_mode * mode,ir_tarval * tv)463 int ia32_evaluate_insn(insn_kind kind, const ir_mode *mode, ir_tarval *tv)
464 {
465 	int cost;
466 
467 	switch (kind) {
468 	case MUL:
469 		cost = arch_costs->cost_mul_start;
470 		if (arch_costs->cost_mul_bit > 0) {
471 			char *bitstr = get_tarval_bitpattern(tv);
472 			int i;
473 
474 			for (i = 0; bitstr[i] != '\0'; ++i) {
475 				if (bitstr[i] == '1') {
476 					cost += arch_costs->cost_mul_bit;
477 				}
478 			}
479 			free(bitstr);
480 		}
481 		if (get_mode_size_bits(mode) <= 32)
482 			return cost;
483 		/* 64bit mul supported, approx 4times of a 32bit mul*/
484 		return 4 * cost;
485 	case LEA:
486 		/* lea is only supported for 32 bit */
487 		if (get_mode_size_bits(mode) <= 32)
488 			return arch_costs->lea_cost;
489 		/* in 64bit mode, the Lea cost are at wort 2 shifts and one add */
490 		return 2 * arch_costs->add_cost + 2 * (2 * arch_costs->const_shf_cost);
491 	case ADD:
492 	case SUB:
493 		if (get_mode_size_bits(mode) <= 32)
494 			return arch_costs->add_cost;
495 		/* 64bit add/sub supported, double the cost */
496 		return 2 * arch_costs->add_cost;
497 	case SHIFT:
498 		if (get_mode_size_bits(mode) <= 32)
499 			return arch_costs->const_shf_cost;
500 		/* 64bit shift supported, double the cost */
501 		return 2 * arch_costs->const_shf_cost;
502 	case ZERO:
503 		return arch_costs->add_cost;
504 	default:
505 		return 1;
506 	}
507 }
508 
509 /* auto detection code only works if we're on an x86 cpu obviously */
510 #ifdef NATIVE_X86
511 typedef struct x86_cpu_info_t {
512 	unsigned char cpu_stepping;
513 	unsigned char cpu_model;
514 	unsigned char cpu_family;
515 	unsigned char cpu_type;
516 	unsigned char cpu_ext_model;
517 	unsigned char cpu_ext_family;
518 	unsigned      edx_features;
519 	unsigned      ecx_features;
520 	unsigned      add_features;
521 } x86_cpu_info_t;
522 
523 enum {
524 	CPUID_FEAT_ECX_SSE3      = 1 << 0,
525 	CPUID_FEAT_ECX_PCLMUL    = 1 << 1,
526 	CPUID_FEAT_ECX_DTES64    = 1 << 2,
527 	CPUID_FEAT_ECX_MONITOR   = 1 << 3,
528 	CPUID_FEAT_ECX_DS_CPL    = 1 << 4,
529 	CPUID_FEAT_ECX_VMX       = 1 << 5,
530 	CPUID_FEAT_ECX_SMX       = 1 << 6,
531 	CPUID_FEAT_ECX_EST       = 1 << 7,
532 	CPUID_FEAT_ECX_TM2       = 1 << 8,
533 	CPUID_FEAT_ECX_SSSE3     = 1 << 9,
534 	CPUID_FEAT_ECX_CID       = 1 << 10,
535 	CPUID_FEAT_ECX_FMA       = 1 << 12,
536 	CPUID_FEAT_ECX_CX16      = 1 << 13,
537 	CPUID_FEAT_ECX_ETPRD     = 1 << 14,
538 	CPUID_FEAT_ECX_PDCM      = 1 << 15,
539 	CPUID_FEAT_ECX_DCA       = 1 << 18,
540 	CPUID_FEAT_ECX_SSE4_1    = 1 << 19,
541 	CPUID_FEAT_ECX_SSE4_2    = 1 << 20,
542 	CPUID_FEAT_ECX_x2APIC    = 1 << 21,
543 	CPUID_FEAT_ECX_MOVBE     = 1 << 22,
544 	CPUID_FEAT_ECX_POPCNT    = 1 << 23,
545 	CPUID_FEAT_ECX_AES       = 1 << 25,
546 	CPUID_FEAT_ECX_XSAVE     = 1 << 26,
547 	CPUID_FEAT_ECX_OSXSAVE   = 1 << 27,
548 	CPUID_FEAT_ECX_AVX       = 1 << 28,
549 
550 	CPUID_FEAT_EDX_FPU       = 1 << 0,
551 	CPUID_FEAT_EDX_VME       = 1 << 1,
552 	CPUID_FEAT_EDX_DE        = 1 << 2,
553 	CPUID_FEAT_EDX_PSE       = 1 << 3,
554 	CPUID_FEAT_EDX_TSC       = 1 << 4,
555 	CPUID_FEAT_EDX_MSR       = 1 << 5,
556 	CPUID_FEAT_EDX_PAE       = 1 << 6,
557 	CPUID_FEAT_EDX_MCE       = 1 << 7,
558 	CPUID_FEAT_EDX_CX8       = 1 << 8,
559 	CPUID_FEAT_EDX_APIC      = 1 << 9,
560 	CPUID_FEAT_EDX_SEP       = 1 << 11,
561 	CPUID_FEAT_EDX_MTRR      = 1 << 12,
562 	CPUID_FEAT_EDX_PGE       = 1 << 13,
563 	CPUID_FEAT_EDX_MCA       = 1 << 14,
564 	CPUID_FEAT_EDX_CMOV      = 1 << 15,
565 	CPUID_FEAT_EDX_PAT       = 1 << 16,
566 	CPUID_FEAT_EDX_PSE36     = 1 << 17,
567 	CPUID_FEAT_EDX_PSN       = 1 << 18,
568 	CPUID_FEAT_EDX_CLF       = 1 << 19,
569 	CPUID_FEAT_EDX_DTES      = 1 << 21,
570 	CPUID_FEAT_EDX_ACPI      = 1 << 22,
571 	CPUID_FEAT_EDX_MMX       = 1 << 23,
572 	CPUID_FEAT_EDX_FXSR      = 1 << 24,
573 	CPUID_FEAT_EDX_SSE       = 1 << 25,
574 	CPUID_FEAT_EDX_SSE2      = 1 << 26,
575 	CPUID_FEAT_EDX_SS        = 1 << 27,
576 	CPUID_FEAT_EDX_HTT       = 1 << 28,
577 	CPUID_FEAT_EDX_TM1       = 1 << 29,
578 	CPUID_FEAT_EDX_IA64      = 1 << 30,
579 	CPUID_FEAT_EDX_PBE       = 1 << 31
580 };
581 
auto_detect_Intel(x86_cpu_info_t const * info)582 static cpu_arch_features auto_detect_Intel(x86_cpu_info_t const *info)
583 {
584 	cpu_arch_features auto_arch = cpu_generic;
585 
586 	unsigned family = info->cpu_ext_family + info->cpu_family;
587 	unsigned model  = (info->cpu_ext_model << 4) | info->cpu_model;
588 
589 	switch (family) {
590 	case 4:
591 		auto_arch = cpu_i486;
592 		break;
593 	case 5:
594 		auto_arch = cpu_pentium;
595 		break;
596 	case 6:
597 		switch (model) {
598 		case 0x01: /* PentiumPro */
599 		case 0x03: /* Pentium II Model 3 */
600 		case 0x05: /* Pentium II Model 5 */
601 		case 0x06: /* Celeron Model 6 */
602 		case 0x07: /* Pentium III Model 7 */
603 		case 0x08: /* Pentium III Model 8 */
604 		case 0x09: /* Pentium M Model 9 */
605 		case 0x0A: /* Pentium III Model 0A */
606 		case 0x0B: /* Pentium III Model 0B */
607 		case 0x0D: /* Pentium M Model 0D */
608 		case 0x0E: /* Core Model 0E */
609 			auto_arch = cpu_pentium_pro_generic;
610 			break;
611 		case 0x0F: /* Core2 Model 0F */
612 		case 0x15: /* Intel EP80579 */
613 		case 0x16: /* Celeron Model 16 */
614 		case 0x17: /* Core2 Model 17 */
615 			auto_arch = cpu_core2_generic;
616 			break;
617 		default:
618 			/* unknown */
619 			break;
620 		}
621 		break;
622 	case 15:
623 		switch (model) {
624 		case 0x00: /* Pentium 4 Model 00 */
625 		case 0x01: /* Pentium 4 Model 01 */
626 		case 0x02: /* Pentium 4 Model 02 */
627 		case 0x03: /* Pentium 4 Model 03 */
628 		case 0x04: /* Pentium 4 Model 04 */
629 		case 0x06: /* Pentium 4 Model 06 */
630 			auto_arch = cpu_netburst_generic;
631 			break;
632 		case 0x1A: /* Core i7 */
633 			auto_arch = cpu_core2_generic;
634 			break;
635 		case 0x1C: /* Atom */
636 			auto_arch = cpu_atom_generic;
637 			break;
638 		case 0x1D: /* Xeon MP */
639 			auto_arch = cpu_core2_generic;
640 			break;
641 		default:
642 			/* unknown */
643 			break;
644 		}
645 		break;
646 	default:
647 		/* unknown */
648 		break;
649 	}
650 
651 	return auto_arch;
652 }
653 
auto_detect_AMD(x86_cpu_info_t const * info)654 static cpu_arch_features auto_detect_AMD(x86_cpu_info_t const *info)
655 {
656 	cpu_arch_features auto_arch = cpu_generic;
657 
658 	unsigned family, model;
659 
660 	if (info->cpu_family == 0x0F) {
661 		family = info->cpu_ext_family + info->cpu_family;
662 		model  = (info->cpu_ext_model << 4) | info->cpu_model;
663 	} else {
664 		family = info->cpu_family;
665 		model  = info->cpu_model;
666 	}
667 
668 	switch (family) {
669 	case 0x04:
670 		auto_arch = cpu_i486;
671 		break;
672 	case 0x05:
673 		switch (model) {
674 		case 0x00: /* K5 Model 0 */
675 		case 0x01: /* K5 Model 1 */
676 		case 0x02: /* K5 Model 2 */
677 		case 0x03: /* K5 Model 3 */
678 			auto_arch = cpu_pentium;
679 			break;
680 		case 0x06: /* K6 Model 6 */
681 		case 0x07: /* K6 Model 7 */
682 		case 0x08: /* K6-2 Model 8 */
683 		case 0x09: /* K6-III Model 9 */
684 		case 0x0D: /* K6-2+ or K6-III+ */
685 			auto_arch = cpu_k6_generic;
686 			break;
687 		case 0x0A: /* Geode LX */
688 			auto_arch = cpu_geode_generic;
689 			break;
690 		default:
691 			/* unknown K6 */
692 			auto_arch = cpu_k6_generic;
693 			break;
694 		}
695 		break;
696 	case 0x06:
697 		switch (model) {
698 		case 0x01: /* Athlon Model 1 */
699 		case 0x02: /* Athlon Model 2 */
700 		case 0x03: /* Duron Model 3 */
701 		case 0x04: /* Athlon Model 4 */
702 		case 0x06: /* Athlon MP/Mobile Athlon Model 6 */
703 		case 0x07: /* Mobile Duron Model 7 */
704 		case 0x08: /* Athlon (TH/AP core) including Geode NX */
705 		case 0x0A: /* Athlon (BT core) */
706 		default:   /* unknown K7 */
707 			auto_arch = cpu_athlon_generic;
708 			break;
709 		}
710 		break;
711 	case 0x0F:
712 		auto_arch = cpu_k8_generic;
713 		break;
714 	case 0x10:
715 	case 0x11: /* AMD Family 11h */
716 	case 0x12: /* AMD Family 12h */
717 	case 0x14: /* AMD Family 14h */
718 	case 0x15: /* AMD Family 15h */
719 		auto_arch = cpu_k10_generic;
720 		break;
721 	default:
722 		/* unknown */
723 		break;
724 	}
725 
726 	return auto_arch;
727 }
728 
729 typedef union {
730 	struct {
731         unsigned eax;
732         unsigned ebx;
733         unsigned ecx;
734         unsigned edx;
735 	} r;
736 	int bulk[4];
737 } cpuid_registers;
738 
x86_cpuid(cpuid_registers * regs,unsigned level)739 static void x86_cpuid(cpuid_registers *regs, unsigned level)
740 {
741 #if defined(__GNUC__)
742 #	if defined(__PIC__) && !defined(__amd64) // GCC cannot handle EBX in PIC
743 	__asm (
744 		"pushl %%ebx\n\t"
745 		"cpuid\n\t"
746 		"movl %%ebx, %1\n\t"
747 		"popl %%ebx"
748 	: "=a" (regs->r.eax), "=r" (regs->r.ebx), "=c" (regs->r.ecx), "=d" (regs->r.edx)
749 	: "a" (level)
750 	);
751 #	else
752 	__asm ("cpuid\n\t"
753 	: "=a" (regs->r.eax), "=b" (regs->r.ebx), "=c" (regs->r.ecx), "=d" (regs->r.edx)
754 	: "a" (level)
755 	);
756 #	endif
757 #elif defined(_MSC_VER)
758 	__cpuid(regs->bulk, level);
759 #else
760 #	error CPUID is missing
761 #endif
762 }
763 
x86_toogle_cpuid(void)764 static bool x86_toogle_cpuid(void)
765 {
766 	unsigned eflags_before = 0;
767 	unsigned eflags_after = 0;
768 
769 #if defined(__GNUC__)
770 #ifdef __i386__
771 	/* If bit 21 of the EFLAGS register can be changed, the cpuid instruction is available */
772 	__asm__(
773 		"pushf\n\t"
774 		"popl %0\n\t"
775 		"movl %0, %1\n\t"
776 		"xorl $0x00200000, %1\n\t"
777 		"pushl %1\n\t"
778 		"popf\n\t"
779 		"pushf\n\t"
780 		"popl %1"
781 		: "=r" (eflags_before), "=r" (eflags_after) :: "cc"
782 		);
783 #else
784 	eflags_after = 0x00200000;
785 #endif
786 #elif defined(_MSC_VER)
787 #if defined(_M_IX86)
788 	__asm {
789 		pushfd
790 		pop eax
791 		mov eflags_before, eax
792 		xor eax, 0x00200000
793 		push eax
794 		popfd
795 		pushfd
796 		pop eax
797 		mov eflags_after, eax
798 	}
799 #else
800 	eflags_after = 0x00200000;
801 #endif
802 #endif
803 	return (eflags_before ^ eflags_after) & 0x00200000;
804 }
805 
autodetect_arch(void)806 static void autodetect_arch(void)
807 {
808 	cpu_arch_features auto_arch = cpu_generic;
809 
810 	/* We use the cpuid instruction to detect the CPU features */
811 	if (x86_toogle_cpuid()) {
812 		cpuid_registers   regs;
813 		char              vendorid[13];
814 		x86_cpu_info_t    cpu_info;
815 
816 		/* get vendor ID */
817 		x86_cpuid(&regs, 0);
818 		memcpy(&vendorid[0], &regs.r.ebx, 4);
819 		memcpy(&vendorid[4], &regs.r.edx, 4);
820 		memcpy(&vendorid[8], &regs.r.ecx, 4);
821 		vendorid[12] = '\0';
822 
823 		/* get processor info and feature bits */
824 		x86_cpuid(&regs, 1);
825 
826 		cpu_info.cpu_stepping   = (regs.r.eax >>  0) & 0x0F;
827 		cpu_info.cpu_model      = (regs.r.eax >>  4) & 0x0F;
828 		cpu_info.cpu_family     = (regs.r.eax >>  8) & 0x0F;
829 		cpu_info.cpu_type       = (regs.r.eax >> 12) & 0x03;
830 		cpu_info.cpu_ext_model  = (regs.r.eax >> 16) & 0x0F;
831 		cpu_info.cpu_ext_family = (regs.r.eax >> 20) & 0xFF;
832 		cpu_info.edx_features   = regs.r.edx;
833 		cpu_info.ecx_features   = regs.r.ecx;
834 		cpu_info.add_features   = regs.r.ebx;
835 
836 		if        (0 == strcmp(vendorid, "GenuineIntel")) {
837 			auto_arch = auto_detect_Intel(&cpu_info);
838 		} else if (0 == strcmp(vendorid, "AuthenticAMD")) {
839 			auto_arch = auto_detect_AMD(&cpu_info);
840 		} else if (0 == strcmp(vendorid, "Geode by NSC")) {
841 			auto_arch = cpu_geode_generic;
842 		}
843 
844 		if (cpu_info.edx_features & CPUID_FEAT_EDX_CMOV)
845 			auto_arch |= arch_feature_cmov;
846 		if (cpu_info.edx_features & CPUID_FEAT_EDX_MMX)
847 			auto_arch |= arch_feature_mmx;
848 		if (cpu_info.edx_features & CPUID_FEAT_EDX_SSE)
849 			auto_arch |= arch_feature_sse1;
850 		if (cpu_info.edx_features & CPUID_FEAT_EDX_SSE2)
851 			auto_arch |= arch_feature_sse2;
852 
853 		if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE3)
854 			auto_arch |= arch_feature_sse3;
855 		if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSSE3)
856 			auto_arch |= arch_feature_ssse3;
857 		if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE4_1)
858 			auto_arch |= arch_feature_sse4_1;
859 		if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE4_2)
860 			auto_arch |= arch_feature_sse4_2;
861 		if (cpu_info.ecx_features & CPUID_FEAT_ECX_POPCNT)
862 			auto_arch |= arch_feature_popcnt;
863 	}
864 
865 	arch     = auto_arch;
866 	opt_arch = auto_arch;
867 }
868 #endif  /* NATIVE_X86 */
869 
flags(cpu_arch_features features,cpu_arch_features flags)870 static bool flags(cpu_arch_features features, cpu_arch_features flags)
871 {
872 	return (features & flags) != 0;
873 }
874 
ia32_setup_cg_config(void)875 void ia32_setup_cg_config(void)
876 {
877 	ia32_code_gen_config_t *const c = &ia32_cg_config;
878 	memset(c, 0, sizeof(*c));
879 
880 	set_arch_costs();
881 
882 #ifdef NATIVE_X86
883 	if (arch == cpu_autodetect)
884 		autodetect_arch();
885 #endif
886 
887 	c->optimize_size        = opt_size != 0;
888 	/* on newer intel cpus mov, pop is often faster than leave although it has a
889 	 * longer opcode */
890 	c->use_leave            = flags(opt_arch, arch_i386 | arch_all_amd | arch_core2) || opt_size;
891 	/* P4s don't like inc/decs because they only partially write the flags
892 	 * register which produces false dependencies */
893 	c->use_incdec           = !flags(opt_arch, arch_netburst | arch_nocona | arch_core2 | arch_geode) || opt_size;
894 	c->use_softfloat        = (fpu_arch & IA32_FPU_ARCH_SOFTFLOAT) != 0;
895 	c->use_sse2             = (fpu_arch & IA32_FPU_ARCH_SSE2) != 0 && flags(arch, arch_feature_sse2);
896 	c->use_ffreep           = flags(opt_arch, arch_athlon_plus);
897 	/* valgrind can't cope with femms yet and the usefulness of the optimization
898 	 * is questionable anyway */
899 #if 0
900 	c->use_femms            = flags(opt_arch, arch_athlon_plus) &&
901 		flags(arch, arch_feature_mmx | arch_all_amd);
902 #else
903 	c->use_femms            = 0;
904 #endif
905 	c->use_fucomi           = flags(arch, arch_feature_p6_insn);
906 	c->use_cmov             = flags(arch, arch_feature_cmov);
907 	c->use_modeD_moves      = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_geode);
908 	c->use_add_esp_4        = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 |             arch_geode)                         && !opt_size;
909 	c->use_add_esp_8        = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_geode | arch_i386 | arch_i486) && !opt_size;
910 	c->use_sub_esp_4        = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro)                                      && !opt_size;
911 	c->use_sub_esp_8        = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro |              arch_i386 | arch_i486) && !opt_size;
912 	c->use_imul_mem_imm32   = !flags(opt_arch, arch_k8 | arch_k10) || opt_size;
913 	c->use_pxor             = flags(opt_arch, arch_netburst);
914 	c->use_mov_0            = flags(opt_arch, arch_k6) && !opt_size;
915 	c->use_short_sex_eax    = !flags(opt_arch, arch_k6) && !opt_size;
916 	c->use_pad_return       = flags(opt_arch, arch_athlon_plus) && !opt_size;
917 	c->use_bt               = flags(opt_arch, arch_core2 | arch_athlon_plus) || opt_size;
918 	c->use_fisttp           = flags(opt_arch & arch, arch_feature_sse3);
919 	c->use_sse_prefetch     = flags(arch, (arch_feature_3DNowE | arch_feature_sse1));
920 	c->use_3dnow_prefetch   = flags(arch, arch_feature_3DNow);
921 	c->use_popcnt           = flags(arch, arch_feature_popcnt);
922 	c->use_bswap            = (arch & arch_mask) >= arch_i486;
923 	c->optimize_cc          = opt_cc;
924 	c->use_unsafe_floatconv = opt_unsafe_floatconv;
925 	c->emit_machcode        = emit_machcode;
926 
927 	c->function_alignment       = arch_costs->function_alignment;
928 	c->label_alignment          = arch_costs->label_alignment;
929 	c->label_alignment_max_skip = arch_costs->label_alignment_max_skip;
930 
931 	c->label_alignment_factor =
932 		flags(opt_arch, arch_i386 | arch_i486) || opt_size ? 0 :
933 		opt_arch & arch_all_amd                            ? 3 :
934 		2;
935 }
936 
ia32_init_architecture(void)937 void ia32_init_architecture(void)
938 {
939 	lc_opt_entry_t *be_grp, *ia32_grp;
940 
941 	memset(&ia32_cg_config, 0, sizeof(ia32_cg_config));
942 
943 	be_grp   = lc_opt_get_grp(firm_opt_get_root(), "be");
944 	ia32_grp = lc_opt_get_grp(be_grp, "ia32");
945 
946 	lc_opt_add_table(ia32_grp, ia32_architecture_options);
947 }
948