2 * Tiny Code Generator for QEMU
4 * Copyright (c) 2008 Fabrice Bellard
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
25 #include "../tcg-pool.c.inc"
27 #ifdef CONFIG_DEBUG_TCG
28 static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
29 #if TCG_TARGET_REG_BITS == 64
30 "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
32 "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34 "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
35 "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
36 #if TCG_TARGET_REG_BITS == 64
37 "%xmm8", "%xmm9", "%xmm10", "%xmm11",
38 "%xmm12", "%xmm13", "%xmm14", "%xmm15",
43 static const int tcg_target_reg_alloc_order[] = {
44 #if TCG_TARGET_REG_BITS == 64
76 /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
77 any of them. Therefore only allow xmm0-xmm5 to be allocated. */
80 #if TCG_TARGET_REG_BITS == 64
93 static const int tcg_target_call_iarg_regs[] = {
94 #if TCG_TARGET_REG_BITS == 64
107 /* 32 bit mode uses stack based calling convention (GCC default). */
111 static const int tcg_target_call_oarg_regs[] = {
113 #if TCG_TARGET_REG_BITS == 32
118 /* Constants we accept. */
119 #define TCG_CT_CONST_S32 0x100
120 #define TCG_CT_CONST_U32 0x200
121 #define TCG_CT_CONST_I32 0x400
122 #define TCG_CT_CONST_WSZ 0x800
124 /* Registers used with L constraint, which are the first argument
125 registers on x86_64, and two random call clobbered registers on
127 #if TCG_TARGET_REG_BITS == 64
128 # define TCG_REG_L0 tcg_target_call_iarg_regs[0]
129 # define TCG_REG_L1 tcg_target_call_iarg_regs[1]
131 # define TCG_REG_L0 TCG_REG_EAX
132 # define TCG_REG_L1 TCG_REG_EDX
135 #define ALL_BYTEH_REGS 0x0000000fu
136 #if TCG_TARGET_REG_BITS == 64
137 # define ALL_GENERAL_REGS 0x0000ffffu
138 # define ALL_VECTOR_REGS 0xffff0000u
139 # define ALL_BYTEL_REGS ALL_GENERAL_REGS
141 # define ALL_GENERAL_REGS 0x000000ffu
142 # define ALL_VECTOR_REGS 0x00ff0000u
143 # define ALL_BYTEL_REGS ALL_BYTEH_REGS
145 #ifdef CONFIG_SOFTMMU
146 # define SOFTMMU_RESERVE_REGS ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
148 # define SOFTMMU_RESERVE_REGS 0
151 /* The host compiler should supply <cpuid.h> to enable runtime features
152 detection, as we're not going to go so far as our own inline assembly.
153 If not available, default values will be assumed. */
154 #if defined(CONFIG_CPUID_H)
155 #include "qemu/cpuid.h"
158 /* For 64-bit, we always know that CMOV is available. */
159 #if TCG_TARGET_REG_BITS == 64
161 #elif defined(CONFIG_CPUID_H)
162 static bool have_cmov;
167 /* We need these symbols in tcg-target.h, and we can't properly conditionalize
168 it there. Therefore we always define the variable. */
175 #ifdef CONFIG_CPUID_H
176 static bool have_bmi2;
177 static bool have_lzcnt;
180 # define have_lzcnt 0
183 static const tcg_insn_unit *tb_ret_addr;
185 static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
186 intptr_t value, intptr_t addend)
191 value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
192 if (value != (int32_t)value) {
197 tcg_patch32(code_ptr, value);
200 value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
201 if (value != (int8_t)value) {
204 tcg_patch8(code_ptr, value);
212 /* test if a constant matches the constraint */
213 static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
214 const TCGArgConstraint *arg_ct)
217 if (ct & TCG_CT_CONST) {
220 if (type == TCG_TYPE_I32) {
221 if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
225 if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
228 if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
231 if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
235 if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
241 # define LOWREGMASK(x) ((x) & 7)
243 #define P_EXT 0x100 /* 0x0f opcode prefix */
244 #define P_EXT38 0x200 /* 0x0f 0x38 opcode prefix */
245 #define P_DATA16 0x400 /* 0x66 opcode prefix */
246 #if TCG_TARGET_REG_BITS == 64
247 # define P_REXW 0x1000 /* Set REX.W = 1 */
248 # define P_REXB_R 0x2000 /* REG field as byte register */
249 # define P_REXB_RM 0x4000 /* R/M field as byte register */
250 # define P_GS 0x8000 /* gs segment override */
257 #define P_EXT3A 0x10000 /* 0x0f 0x3a opcode prefix */
258 #define P_SIMDF3 0x20000 /* 0xf3 opcode prefix */
259 #define P_SIMDF2 0x40000 /* 0xf2 opcode prefix */
260 #define P_VEXL 0x80000 /* Set VEX.L = 1 */
262 #define OPC_ARITH_EvIz (0x81)
263 #define OPC_ARITH_EvIb (0x83)
264 #define OPC_ARITH_GvEv (0x03) /* ... plus (ARITH_FOO << 3) */
265 #define OPC_ANDN (0xf2 | P_EXT38)
266 #define OPC_ADD_GvEv (OPC_ARITH_GvEv | (ARITH_ADD << 3))
267 #define OPC_AND_GvEv (OPC_ARITH_GvEv | (ARITH_AND << 3))
268 #define OPC_BLENDPS (0x0c | P_EXT3A | P_DATA16)
269 #define OPC_BSF (0xbc | P_EXT)
270 #define OPC_BSR (0xbd | P_EXT)
271 #define OPC_BSWAP (0xc8 | P_EXT)
272 #define OPC_CALL_Jz (0xe8)
273 #define OPC_CMOVCC (0x40 | P_EXT) /* ... plus condition code */
274 #define OPC_CMP_GvEv (OPC_ARITH_GvEv | (ARITH_CMP << 3))
275 #define OPC_DEC_r32 (0x48)
276 #define OPC_IMUL_GvEv (0xaf | P_EXT)
277 #define OPC_IMUL_GvEvIb (0x6b)
278 #define OPC_IMUL_GvEvIz (0x69)
279 #define OPC_INC_r32 (0x40)
280 #define OPC_JCC_long (0x80 | P_EXT) /* ... plus condition code */
281 #define OPC_JCC_short (0x70) /* ... plus condition code */
282 #define OPC_JMP_long (0xe9)
283 #define OPC_JMP_short (0xeb)
284 #define OPC_LEA (0x8d)
285 #define OPC_LZCNT (0xbd | P_EXT | P_SIMDF3)
286 #define OPC_MOVB_EvGv (0x88) /* stores, more or less */
287 #define OPC_MOVL_EvGv (0x89) /* stores, more or less */
288 #define OPC_MOVL_GvEv (0x8b) /* loads, more or less */
289 #define OPC_MOVB_EvIz (0xc6)
290 #define OPC_MOVL_EvIz (0xc7)
291 #define OPC_MOVL_Iv (0xb8)
292 #define OPC_MOVBE_GyMy (0xf0 | P_EXT38)
293 #define OPC_MOVBE_MyGy (0xf1 | P_EXT38)
294 #define OPC_MOVD_VyEy (0x6e | P_EXT | P_DATA16)
295 #define OPC_MOVD_EyVy (0x7e | P_EXT | P_DATA16)
296 #define OPC_MOVDDUP (0x12 | P_EXT | P_SIMDF2)
297 #define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
298 #define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
299 #define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
300 #define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
301 #define OPC_MOVQ_VqWq (0x7e | P_EXT | P_SIMDF3)
302 #define OPC_MOVQ_WqVq (0xd6 | P_EXT | P_DATA16)
303 #define OPC_MOVSBL (0xbe | P_EXT)
304 #define OPC_MOVSWL (0xbf | P_EXT)
305 #define OPC_MOVSLQ (0x63 | P_REXW)
306 #define OPC_MOVZBL (0xb6 | P_EXT)
307 #define OPC_MOVZWL (0xb7 | P_EXT)
308 #define OPC_PABSB (0x1c | P_EXT38 | P_DATA16)
309 #define OPC_PABSW (0x1d | P_EXT38 | P_DATA16)
310 #define OPC_PABSD (0x1e | P_EXT38 | P_DATA16)
311 #define OPC_PACKSSDW (0x6b | P_EXT | P_DATA16)
312 #define OPC_PACKSSWB (0x63 | P_EXT | P_DATA16)
313 #define OPC_PACKUSDW (0x2b | P_EXT38 | P_DATA16)
314 #define OPC_PACKUSWB (0x67 | P_EXT | P_DATA16)
315 #define OPC_PADDB (0xfc | P_EXT | P_DATA16)
316 #define OPC_PADDW (0xfd | P_EXT | P_DATA16)
317 #define OPC_PADDD (0xfe | P_EXT | P_DATA16)
318 #define OPC_PADDQ (0xd4 | P_EXT | P_DATA16)
319 #define OPC_PADDSB (0xec | P_EXT | P_DATA16)
320 #define OPC_PADDSW (0xed | P_EXT | P_DATA16)
321 #define OPC_PADDUB (0xdc | P_EXT | P_DATA16)
322 #define OPC_PADDUW (0xdd | P_EXT | P_DATA16)
323 #define OPC_PAND (0xdb | P_EXT | P_DATA16)
324 #define OPC_PANDN (0xdf | P_EXT | P_DATA16)
325 #define OPC_PBLENDW (0x0e | P_EXT3A | P_DATA16)
326 #define OPC_PCMPEQB (0x74 | P_EXT | P_DATA16)
327 #define OPC_PCMPEQW (0x75 | P_EXT | P_DATA16)
328 #define OPC_PCMPEQD (0x76 | P_EXT | P_DATA16)
329 #define OPC_PCMPEQQ (0x29 | P_EXT38 | P_DATA16)
330 #define OPC_PCMPGTB (0x64 | P_EXT | P_DATA16)
331 #define OPC_PCMPGTW (0x65 | P_EXT | P_DATA16)
332 #define OPC_PCMPGTD (0x66 | P_EXT | P_DATA16)
333 #define OPC_PCMPGTQ (0x37 | P_EXT38 | P_DATA16)
334 #define OPC_PMAXSB (0x3c | P_EXT38 | P_DATA16)
335 #define OPC_PMAXSW (0xee | P_EXT | P_DATA16)
336 #define OPC_PMAXSD (0x3d | P_EXT38 | P_DATA16)
337 #define OPC_PMAXUB (0xde | P_EXT | P_DATA16)
338 #define OPC_PMAXUW (0x3e | P_EXT38 | P_DATA16)
339 #define OPC_PMAXUD (0x3f | P_EXT38 | P_DATA16)
340 #define OPC_PMINSB (0x38 | P_EXT38 | P_DATA16)
341 #define OPC_PMINSW (0xea | P_EXT | P_DATA16)
342 #define OPC_PMINSD (0x39 | P_EXT38 | P_DATA16)
343 #define OPC_PMINUB (0xda | P_EXT | P_DATA16)
344 #define OPC_PMINUW (0x3a | P_EXT38 | P_DATA16)
345 #define OPC_PMINUD (0x3b | P_EXT38 | P_DATA16)
346 #define OPC_PMOVSXBW (0x20 | P_EXT38 | P_DATA16)
347 #define OPC_PMOVSXWD (0x23 | P_EXT38 | P_DATA16)
348 #define OPC_PMOVSXDQ (0x25 | P_EXT38 | P_DATA16)
349 #define OPC_PMOVZXBW (0x30 | P_EXT38 | P_DATA16)
350 #define OPC_PMOVZXWD (0x33 | P_EXT38 | P_DATA16)
351 #define OPC_PMOVZXDQ (0x35 | P_EXT38 | P_DATA16)
352 #define OPC_PMULLW (0xd5 | P_EXT | P_DATA16)
353 #define OPC_PMULLD (0x40 | P_EXT38 | P_DATA16)
354 #define OPC_POR (0xeb | P_EXT | P_DATA16)
355 #define OPC_PSHUFB (0x00 | P_EXT38 | P_DATA16)
356 #define OPC_PSHUFD (0x70 | P_EXT | P_DATA16)
357 #define OPC_PSHUFLW (0x70 | P_EXT | P_SIMDF2)
358 #define OPC_PSHUFHW (0x70 | P_EXT | P_SIMDF3)
359 #define OPC_PSHIFTW_Ib (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
360 #define OPC_PSHIFTD_Ib (0x72 | P_EXT | P_DATA16) /* /2 /6 /4 */
361 #define OPC_PSHIFTQ_Ib (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
362 #define OPC_PSLLW (0xf1 | P_EXT | P_DATA16)
363 #define OPC_PSLLD (0xf2 | P_EXT | P_DATA16)
364 #define OPC_PSLLQ (0xf3 | P_EXT | P_DATA16)
365 #define OPC_PSRAW (0xe1 | P_EXT | P_DATA16)
366 #define OPC_PSRAD (0xe2 | P_EXT | P_DATA16)
367 #define OPC_PSRLW (0xd1 | P_EXT | P_DATA16)
368 #define OPC_PSRLD (0xd2 | P_EXT | P_DATA16)
369 #define OPC_PSRLQ (0xd3 | P_EXT | P_DATA16)
370 #define OPC_PSUBB (0xf8 | P_EXT | P_DATA16)
371 #define OPC_PSUBW (0xf9 | P_EXT | P_DATA16)
372 #define OPC_PSUBD (0xfa | P_EXT | P_DATA16)
373 #define OPC_PSUBQ (0xfb | P_EXT | P_DATA16)
374 #define OPC_PSUBSB (0xe8 | P_EXT | P_DATA16)
375 #define OPC_PSUBSW (0xe9 | P_EXT | P_DATA16)
376 #define OPC_PSUBUB (0xd8 | P_EXT | P_DATA16)
377 #define OPC_PSUBUW (0xd9 | P_EXT | P_DATA16)
378 #define OPC_PUNPCKLBW (0x60 | P_EXT | P_DATA16)
379 #define OPC_PUNPCKLWD (0x61 | P_EXT | P_DATA16)
380 #define OPC_PUNPCKLDQ (0x62 | P_EXT | P_DATA16)
381 #define OPC_PUNPCKLQDQ (0x6c | P_EXT | P_DATA16)
382 #define OPC_PUNPCKHBW (0x68 | P_EXT | P_DATA16)
383 #define OPC_PUNPCKHWD (0x69 | P_EXT | P_DATA16)
384 #define OPC_PUNPCKHDQ (0x6a | P_EXT | P_DATA16)
385 #define OPC_PUNPCKHQDQ (0x6d | P_EXT | P_DATA16)
386 #define OPC_PXOR (0xef | P_EXT | P_DATA16)
387 #define OPC_POP_r32 (0x58)
388 #define OPC_POPCNT (0xb8 | P_EXT | P_SIMDF3)
389 #define OPC_PUSH_r32 (0x50)
390 #define OPC_PUSH_Iv (0x68)
391 #define OPC_PUSH_Ib (0x6a)
392 #define OPC_RET (0xc3)
393 #define OPC_SETCC (0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
394 #define OPC_SHIFT_1 (0xd1)
395 #define OPC_SHIFT_Ib (0xc1)
396 #define OPC_SHIFT_cl (0xd3)
397 #define OPC_SARX (0xf7 | P_EXT38 | P_SIMDF3)
398 #define OPC_SHUFPS (0xc6 | P_EXT)
399 #define OPC_SHLX (0xf7 | P_EXT38 | P_DATA16)
400 #define OPC_SHRX (0xf7 | P_EXT38 | P_SIMDF2)
401 #define OPC_SHRD_Ib (0xac | P_EXT)
402 #define OPC_TESTL (0x85)
403 #define OPC_TZCNT (0xbc | P_EXT | P_SIMDF3)
404 #define OPC_UD2 (0x0b | P_EXT)
405 #define OPC_VPBLENDD (0x02 | P_EXT3A | P_DATA16)
406 #define OPC_VPBLENDVB (0x4c | P_EXT3A | P_DATA16)
407 #define OPC_VPINSRB (0x20 | P_EXT3A | P_DATA16)
408 #define OPC_VPINSRW (0xc4 | P_EXT | P_DATA16)
409 #define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
410 #define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
411 #define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
412 #define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
413 #define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
414 #define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
415 #define OPC_VPERMQ (0x00 | P_EXT3A | P_DATA16 | P_REXW)
416 #define OPC_VPERM2I128 (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
417 #define OPC_VPSLLVD (0x47 | P_EXT38 | P_DATA16)
418 #define OPC_VPSLLVQ (0x47 | P_EXT38 | P_DATA16 | P_REXW)
419 #define OPC_VPSRAVD (0x46 | P_EXT38 | P_DATA16)
420 #define OPC_VPSRLVD (0x45 | P_EXT38 | P_DATA16)
421 #define OPC_VPSRLVQ (0x45 | P_EXT38 | P_DATA16 | P_REXW)
422 #define OPC_VZEROUPPER (0x77 | P_EXT)
423 #define OPC_XCHG_ax_r32 (0x90)
425 #define OPC_GRP3_Ev (0xf7)
426 #define OPC_GRP5 (0xff)
427 #define OPC_GRP14 (0x73 | P_EXT | P_DATA16)
429 /* Group 1 opcode extensions for 0x80-0x83.
430 These are also used as modifiers for OPC_ARITH. */
440 /* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3. */
447 /* Group 3 opcode extensions for 0xf6, 0xf7. To be used with OPC_GRP3. */
455 /* Group 5 opcode extensions for 0xff. To be used with OPC_GRP5. */
456 #define EXT5_INC_Ev 0
457 #define EXT5_DEC_Ev 1
458 #define EXT5_CALLN_Ev 2
459 #define EXT5_JMPN_Ev 4
461 /* Condition codes to be added to OPC_JCC_{long,short}. */
480 static const uint8_t tcg_cond_to_jcc[] = {
481 [TCG_COND_EQ] = JCC_JE,
482 [TCG_COND_NE] = JCC_JNE,
483 [TCG_COND_LT] = JCC_JL,
484 [TCG_COND_GE] = JCC_JGE,
485 [TCG_COND_LE] = JCC_JLE,
486 [TCG_COND_GT] = JCC_JG,
487 [TCG_COND_LTU] = JCC_JB,
488 [TCG_COND_GEU] = JCC_JAE,
489 [TCG_COND_LEU] = JCC_JBE,
490 [TCG_COND_GTU] = JCC_JA,
493 #if TCG_TARGET_REG_BITS == 64
494 static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
501 if (opc & P_DATA16) {
502 /* We should never be asking for both 16 and 64-bit operation. */
503 tcg_debug_assert((opc & P_REXW) == 0);
506 if (opc & P_SIMDF3) {
508 } else if (opc & P_SIMDF2) {
513 rex |= (opc & P_REXW) ? 0x8 : 0x0; /* REX.W */
514 rex |= (r & 8) >> 1; /* REX.R */
515 rex |= (x & 8) >> 2; /* REX.X */
516 rex |= (rm & 8) >> 3; /* REX.B */
518 /* P_REXB_{R,RM} indicates that the given register is the low byte.
519 For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
520 as otherwise the encoding indicates %[abcd]h. Note that the values
521 that are ORed in merely indicate that the REX byte must be present;
522 those bits get discarded in output. */
523 rex |= opc & (r >= 4 ? P_REXB_R : 0);
524 rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
527 tcg_out8(s, (uint8_t)(rex | 0x40));
530 if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
534 } else if (opc & P_EXT3A) {
542 static void tcg_out_opc(TCGContext *s, int opc)
544 if (opc & P_DATA16) {
547 if (opc & P_SIMDF3) {
549 } else if (opc & P_SIMDF2) {
552 if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
556 } else if (opc & P_EXT3A) {
562 /* Discard the register arguments to tcg_out_opc early, so as not to penalize
563 the 32-bit compilation paths. This method works with all versions of gcc,
564 whereas relying on optimization may not be able to exclude them. */
565 #define tcg_out_opc(s, opc, r, rm, x) (tcg_out_opc)(s, opc)
568 static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
570 tcg_out_opc(s, opc, r, rm, 0);
571 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
574 static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
579 /* Use the two byte form if possible, which cannot encode
580 VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT. */
581 if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_REXW)) == P_EXT
582 && ((rm | index) & 8) == 0) {
583 /* Two byte VEX prefix. */
586 tmp = (r & 8 ? 0 : 0x80); /* VEX.R */
588 /* Three byte VEX prefix. */
594 } else if (opc & P_EXT38) {
596 } else if (opc & P_EXT) {
599 g_assert_not_reached();
601 tmp |= (r & 8 ? 0 : 0x80); /* VEX.R */
602 tmp |= (index & 8 ? 0 : 0x40); /* VEX.X */
603 tmp |= (rm & 8 ? 0 : 0x20); /* VEX.B */
606 tmp = (opc & P_REXW ? 0x80 : 0); /* VEX.W */
609 tmp |= (opc & P_VEXL ? 0x04 : 0); /* VEX.L */
611 if (opc & P_DATA16) {
613 } else if (opc & P_SIMDF3) {
615 } else if (opc & P_SIMDF2) {
618 tmp |= (~v & 15) << 3; /* VEX.vvvv */
623 static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
625 tcg_out_vex_opc(s, opc, r, v, rm, 0);
626 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
629 /* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
630 We handle either RM and INDEX missing with a negative value. In 64-bit
631 mode for absolute addresses, ~RM is the size of the immediate operand
632 that will follow the instruction. */
634 static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
635 int shift, intptr_t offset)
639 if (index < 0 && rm < 0) {
640 if (TCG_TARGET_REG_BITS == 64) {
641 /* Try for a rip-relative addressing mode. This has replaced
642 the 32-bit-mode absolute addressing encoding. */
643 intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
644 intptr_t disp = offset - pc;
645 if (disp == (int32_t)disp) {
646 tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
651 /* Try for an absolute address encoding. This requires the
652 use of the MODRM+SIB encoding and is therefore larger than
653 rip-relative addressing. */
654 if (offset == (int32_t)offset) {
655 tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
656 tcg_out8(s, (4 << 3) | 5);
657 tcg_out32(s, offset);
661 /* ??? The memory isn't directly addressable. */
662 g_assert_not_reached();
664 /* Absolute address. */
665 tcg_out8(s, (r << 3) | 5);
666 tcg_out32(s, offset);
671 /* Find the length of the immediate addend. Note that the encoding
672 that would be used for (%ebp) indicates absolute addressing. */
674 mod = 0, len = 4, rm = 5;
675 } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
677 } else if (offset == (int8_t)offset) {
683 /* Use a single byte MODRM format if possible. Note that the encoding
684 that would be used for %esp is the escape to the two byte form. */
685 if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
686 /* Single byte MODRM format. */
687 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
689 /* Two byte MODRM+SIB format. */
691 /* Note that the encoding that would place %esp into the index
692 field indicates no index register. In 64-bit mode, the REX.X
693 bit counts, so %r12 can be used as the index. */
697 tcg_debug_assert(index != TCG_REG_ESP);
700 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
701 tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
706 } else if (len == 4) {
707 tcg_out32(s, offset);
711 static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
712 int index, int shift, intptr_t offset)
714 tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
715 tcg_out_sib_offset(s, r, rm, index, shift, offset);
718 static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
719 int rm, int index, int shift,
722 tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
723 tcg_out_sib_offset(s, r, rm, index, shift, offset);
726 /* A simplification of the above with no index or shift. */
727 static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
728 int rm, intptr_t offset)
730 tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
733 static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
734 int v, int rm, intptr_t offset)
736 tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
739 /* Output an opcode with an expected reference to the constant pool. */
740 static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
742 tcg_out_opc(s, opc, r, 0, 0);
743 /* Absolute for 32-bit, pc-relative for 64-bit. */
744 tcg_out8(s, LOWREGMASK(r) << 3 | 5);
748 /* Output an opcode with an expected reference to the constant pool. */
749 static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
751 tcg_out_vex_opc(s, opc, r, 0, 0, 0);
752 /* Absolute for 32-bit, pc-relative for 64-bit. */
753 tcg_out8(s, LOWREGMASK(r) << 3 | 5);
757 /* Generate dest op= src. Uses the same ARITH_* codes as tgen_arithi. */
758 static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
760 /* Propagate an opcode prefix, such as P_REXW. */
761 int ext = subop & ~0x7;
764 tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
767 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
781 tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
783 tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
787 tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
789 tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
795 tcg_debug_assert(ret >= 16 && arg >= 16);
796 tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
799 tcg_debug_assert(ret >= 16 && arg >= 16);
800 tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
803 tcg_debug_assert(ret >= 16 && arg >= 16);
804 tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
808 g_assert_not_reached();
813 static const int avx2_dup_insn[4] = {
814 OPC_VPBROADCASTB, OPC_VPBROADCASTW,
815 OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
818 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
822 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
823 tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
827 /* ??? With zero in a register, use PSHUFB. */
828 tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
832 tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
836 tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
837 /* imm8 operand: all output lanes selected from input lane 0. */
841 tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
844 g_assert_not_reached();
850 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
851 TCGReg r, TCGReg base, intptr_t offset)
854 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
855 tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
860 tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
863 tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
866 tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
867 tcg_out8(s, 0); /* imm8 */
868 tcg_out_dup_vec(s, type, vece, r, r);
871 tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
872 tcg_out8(s, 0); /* imm8 */
873 tcg_out_dup_vec(s, type, vece, r, r);
876 g_assert_not_reached();
882 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
883 TCGReg ret, int64_t arg)
885 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
888 tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
892 tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
896 if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
898 tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
900 tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
902 new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
904 if (type == TCG_TYPE_V64) {
905 tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
906 } else if (have_avx2) {
907 tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
909 tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
911 if (TCG_TARGET_REG_BITS == 64) {
912 new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
914 new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
919 static void tcg_out_movi_vec(TCGContext *s, TCGType type,
920 TCGReg ret, tcg_target_long arg)
923 tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
927 tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
931 int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
932 tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
933 if (TCG_TARGET_REG_BITS == 64) {
934 new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
936 new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
940 static void tcg_out_movi_int(TCGContext *s, TCGType type,
941 TCGReg ret, tcg_target_long arg)
943 tcg_target_long diff;
946 tgen_arithr(s, ARITH_XOR, ret, ret);
949 if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
950 tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
954 if (arg == (int32_t)arg) {
955 tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
960 /* Try a 7 byte pc-relative lea before the 10 byte movq. */
961 diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
962 if (diff == (int32_t)diff) {
963 tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
964 tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
969 tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
973 static void tcg_out_movi(TCGContext *s, TCGType type,
974 TCGReg ret, tcg_target_long arg)
978 #if TCG_TARGET_REG_BITS == 64
982 tcg_out_movi_int(s, type, ret, arg);
984 tcg_out_movi_vec(s, type, ret, arg);
988 g_assert_not_reached();
992 static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
994 if (val == (int8_t)val) {
995 tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
997 } else if (val == (int32_t)val) {
998 tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1005 static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1007 /* Given the strength of x86 memory ordering, we only need care for
1008 store-load ordering. Experimentally, "lock orl $0,0(%esp)" is
1009 faster than "mfence", so don't bother with the sse insn. */
1010 if (a0 & TCG_MO_ST_LD) {
1012 tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1017 static inline void tcg_out_push(TCGContext *s, int reg)
1019 tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1022 static inline void tcg_out_pop(TCGContext *s, int reg)
1024 tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1027 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1028 TCGReg arg1, intptr_t arg2)
1033 tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1035 tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1040 tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1045 /* There is no instruction that can validate 8-byte alignment. */
1046 tcg_debug_assert(ret >= 16);
1047 tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1051 * The gvec infrastructure is asserts that v128 vector loads
1052 * and stores use a 16-byte aligned offset. Validate that the
1053 * final pointer is aligned by using an insn that will SIGSEGV.
1055 tcg_debug_assert(ret >= 16);
1056 tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1060 * The gvec infrastructure only requires 16-byte alignment,
1061 * so here we must use an unaligned load.
1063 tcg_debug_assert(ret >= 16);
1064 tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1065 ret, 0, arg1, arg2);
1068 g_assert_not_reached();
1072 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1073 TCGReg arg1, intptr_t arg2)
1078 tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1080 tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1085 tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1090 /* There is no instruction that can validate 8-byte alignment. */
1091 tcg_debug_assert(arg >= 16);
1092 tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1096 * The gvec infrastructure is asserts that v128 vector loads
1097 * and stores use a 16-byte aligned offset. Validate that the
1098 * final pointer is aligned by using an insn that will SIGSEGV.
1100 tcg_debug_assert(arg >= 16);
1101 tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1105 * The gvec infrastructure only requires 16-byte alignment,
1106 * so here we must use an unaligned store.
1108 tcg_debug_assert(arg >= 16);
1109 tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1110 arg, 0, arg1, arg2);
1113 g_assert_not_reached();
1117 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1118 TCGReg base, intptr_t ofs)
1121 if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1122 if (val != (int32_t)val) {
1126 } else if (type != TCG_TYPE_I32) {
1129 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1134 static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1136 /* Propagate an opcode prefix, such as P_DATA16. */
1137 int ext = subopc & ~0x7;
1141 tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1143 tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1148 static inline void tcg_out_bswap32(TCGContext *s, int reg)
1150 tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1153 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1155 tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1158 static inline void tcg_out_ext8u(TCGContext *s, int dest, int src)
1161 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1162 tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1165 static void tcg_out_ext8s(TCGContext *s, int dest, int src, int rexw)
1168 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1169 tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1172 static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
1175 tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1178 static inline void tcg_out_ext16s(TCGContext *s, int dest, int src, int rexw)
1181 tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1184 static inline void tcg_out_ext32u(TCGContext *s, int dest, int src)
1186 /* 32-bit mov zero extends. */
1187 tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1190 static inline void tcg_out_ext32s(TCGContext *s, int dest, int src)
1192 tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1195 static inline void tcg_out_bswap64(TCGContext *s, int reg)
1197 tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1200 static void tgen_arithi(TCGContext *s, int c, int r0,
1201 tcg_target_long val, int cf)
1205 if (TCG_TARGET_REG_BITS == 64) {
1210 /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1211 partial flags update stalls on Pentium4 and are not recommended
1212 by current Intel optimization manuals. */
1213 if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1214 int is_inc = (c == ARITH_ADD) ^ (val < 0);
1215 if (TCG_TARGET_REG_BITS == 64) {
1216 /* The single-byte increment encodings are re-tasked as the
1217 REX prefixes. Use the MODRM encoding. */
1218 tcg_out_modrm(s, OPC_GRP5 + rexw,
1219 (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1221 tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1226 if (c == ARITH_AND) {
1227 if (TCG_TARGET_REG_BITS == 64) {
1228 if (val == 0xffffffffu) {
1229 tcg_out_ext32u(s, r0, r0);
1232 if (val == (uint32_t)val) {
1233 /* AND with no high bits set can use a 32-bit operation. */
1237 if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1238 tcg_out_ext8u(s, r0, r0);
1241 if (val == 0xffffu) {
1242 tcg_out_ext16u(s, r0, r0);
1247 if (val == (int8_t)val) {
1248 tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1252 if (rexw == 0 || val == (int32_t)val) {
1253 tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1261 static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1264 tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1268 /* Use SMALL != 0 to force a short forward branch. */
1269 static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, int small)
1274 val = tcg_pcrel_diff(s, l->u.value_ptr);
1276 if ((int8_t)val1 == val1) {
1278 tcg_out8(s, OPC_JMP_short);
1280 tcg_out8(s, OPC_JCC_short + opc);
1288 tcg_out8(s, OPC_JMP_long);
1289 tcg_out32(s, val - 5);
1291 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1292 tcg_out32(s, val - 6);
1297 tcg_out8(s, OPC_JMP_short);
1299 tcg_out8(s, OPC_JCC_short + opc);
1301 tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1305 tcg_out8(s, OPC_JMP_long);
1307 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1309 tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1314 static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1315 int const_arg2, int rexw)
1320 tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1322 tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1325 tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1329 static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
1330 TCGArg arg1, TCGArg arg2, int const_arg2,
1331 TCGLabel *label, int small)
1333 tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1334 tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1337 #if TCG_TARGET_REG_BITS == 64
1338 static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
1339 TCGArg arg1, TCGArg arg2, int const_arg2,
1340 TCGLabel *label, int small)
1342 tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1343 tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1346 /* XXX: we implement it at the target level to avoid having to
1347 handle cross basic blocks temporaries */
1348 static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1349 const int *const_args, int small)
1351 TCGLabel *label_next = gen_new_label();
1352 TCGLabel *label_this = arg_label(args[5]);
1356 tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1358 tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
1362 tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1364 tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
1368 tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1370 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1371 tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1375 tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1377 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1378 tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1382 tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1384 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1385 tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1389 tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1391 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1392 tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1396 tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1398 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1399 tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1403 tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1405 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1406 tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1410 tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1412 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1413 tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1417 tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1419 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1420 tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1426 tcg_out_label(s, label_next);
1430 static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1431 TCGArg arg1, TCGArg arg2, int const_arg2)
1433 tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1434 tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1435 tcg_out_ext8u(s, dest, dest);
1438 #if TCG_TARGET_REG_BITS == 64
1439 static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1440 TCGArg arg1, TCGArg arg2, int const_arg2)
1442 tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1443 tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1444 tcg_out_ext8u(s, dest, dest);
1447 static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1448 const int *const_args)
1451 TCGLabel *label_true, *label_over;
1453 memcpy(new_args, args+1, 5*sizeof(TCGArg));
1455 if (args[0] == args[1] || args[0] == args[2]
1456 || (!const_args[3] && args[0] == args[3])
1457 || (!const_args[4] && args[0] == args[4])) {
1458 /* When the destination overlaps with one of the argument
1459 registers, don't do anything tricky. */
1460 label_true = gen_new_label();
1461 label_over = gen_new_label();
1463 new_args[5] = label_arg(label_true);
1464 tcg_out_brcond2(s, new_args, const_args+1, 1);
1466 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1467 tcg_out_jxx(s, JCC_JMP, label_over, 1);
1468 tcg_out_label(s, label_true);
1470 tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1471 tcg_out_label(s, label_over);
1473 /* When the destination does not overlap one of the arguments,
1474 clear the destination first, jump if cond false, and emit an
1475 increment in the true case. This results in smaller code. */
1477 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1479 label_over = gen_new_label();
1480 new_args[4] = tcg_invert_cond(new_args[4]);
1481 new_args[5] = label_arg(label_over);
1482 tcg_out_brcond2(s, new_args, const_args+1, 1);
1484 tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1485 tcg_out_label(s, label_over);
1490 static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1491 TCGReg dest, TCGReg v1)
1494 tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1496 TCGLabel *over = gen_new_label();
1497 tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1498 tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1499 tcg_out_label(s, over);
1503 static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1504 TCGReg c1, TCGArg c2, int const_c2,
1507 tcg_out_cmp(s, c1, c2, const_c2, 0);
1508 tcg_out_cmov(s, cond, 0, dest, v1);
1511 #if TCG_TARGET_REG_BITS == 64
1512 static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1513 TCGReg c1, TCGArg c2, int const_c2,
1516 tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1517 tcg_out_cmov(s, cond, P_REXW, dest, v1);
1521 static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1522 TCGArg arg2, bool const_a2)
1525 tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1527 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1529 tcg_debug_assert(dest != arg2);
1530 tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1533 tcg_debug_assert(dest != arg2);
1534 tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1535 tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1539 static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1540 TCGArg arg2, bool const_a2)
1543 tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1545 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1547 tcg_debug_assert(dest != arg2);
1548 tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1551 tcg_debug_assert(!const_a2);
1552 tcg_debug_assert(dest != arg1);
1553 tcg_debug_assert(dest != arg2);
1555 /* Recall that the output of BSR is the index not the count. */
1556 tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1557 tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1559 /* Since we have destroyed the flags from BSR, we have to re-test. */
1560 tcg_out_cmp(s, arg1, 0, 1, rexw);
1561 tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1565 static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1567 intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1569 if (disp == (int32_t)disp) {
1570 tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1573 /* rip-relative addressing into the constant pool.
1574 This is 6 + 8 = 14 bytes, as compared to using an
1575 an immediate load 10 + 6 = 16 bytes, plus we may
1576 be able to re-use the pool constant for more calls. */
1577 tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1578 tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1579 new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1584 static inline void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest)
1586 tcg_out_branch(s, 1, dest);
1589 static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1591 tcg_out_branch(s, 0, dest);
1594 static void tcg_out_nopn(TCGContext *s, int n)
1597 /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1598 * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1599 * duplicate prefix, and all of the interesting recent cores can
1600 * decode and discard the duplicates in a single cycle.
1602 tcg_debug_assert(n >= 1);
1603 for (i = 1; i < n; ++i) {
1609 #if defined(CONFIG_SOFTMMU)
1610 #include "../tcg-ldst.c.inc"
1612 /* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1613 * int mmu_idx, uintptr_t ra)
1615 static void * const qemu_ld_helpers[16] = {
1616 [MO_UB] = helper_ret_ldub_mmu,
1617 [MO_LEUW] = helper_le_lduw_mmu,
1618 [MO_LEUL] = helper_le_ldul_mmu,
1619 [MO_LEQ] = helper_le_ldq_mmu,
1620 [MO_BEUW] = helper_be_lduw_mmu,
1621 [MO_BEUL] = helper_be_ldul_mmu,
1622 [MO_BEQ] = helper_be_ldq_mmu,
1625 /* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1626 * uintxx_t val, int mmu_idx, uintptr_t ra)
1628 static void * const qemu_st_helpers[16] = {
1629 [MO_UB] = helper_ret_stb_mmu,
1630 [MO_LEUW] = helper_le_stw_mmu,
1631 [MO_LEUL] = helper_le_stl_mmu,
1632 [MO_LEQ] = helper_le_stq_mmu,
1633 [MO_BEUW] = helper_be_stw_mmu,
1634 [MO_BEUL] = helper_be_stl_mmu,
1635 [MO_BEQ] = helper_be_stq_mmu,
1638 /* Perform the TLB load and compare.
1641 ADDRLO and ADDRHI contain the low and high part of the address.
1643 MEM_INDEX and S_BITS are the memory context and log2 size of the load.
1645 WHICH is the offset into the CPUTLBEntry structure of the slot to read.
1646 This should be offsetof addr_read or addr_write.
1649 LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
1650 positions of the displacements of forward jumps to the TLB miss case.
1652 Second argument register is loaded with the low part of the address.
1653 In the TLB hit case, it has been adjusted as indicated by the TLB
1654 and so is a host address. In the TLB miss case, it continues to
1655 hold a guest address.
1657 First argument register is clobbered. */
1659 static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1660 int mem_index, MemOp opc,
1661 tcg_insn_unit **label_ptr, int which)
1663 const TCGReg r0 = TCG_REG_L0;
1664 const TCGReg r1 = TCG_REG_L1;
1665 TCGType ttype = TCG_TYPE_I32;
1666 TCGType tlbtype = TCG_TYPE_I32;
1667 int trexw = 0, hrexw = 0, tlbrexw = 0;
1668 unsigned a_bits = get_alignment_bits(opc);
1669 unsigned s_bits = opc & MO_SIZE;
1670 unsigned a_mask = (1 << a_bits) - 1;
1671 unsigned s_mask = (1 << s_bits) - 1;
1672 target_ulong tlb_mask;
1674 if (TCG_TARGET_REG_BITS == 64) {
1675 if (TARGET_LONG_BITS == 64) {
1676 ttype = TCG_TYPE_I64;
1679 if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1681 if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
1682 tlbtype = TCG_TYPE_I64;
1688 tcg_out_mov(s, tlbtype, r0, addrlo);
1689 tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
1690 TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
1692 tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
1693 TLB_MASK_TABLE_OFS(mem_index) +
1694 offsetof(CPUTLBDescFast, mask));
1696 tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
1697 TLB_MASK_TABLE_OFS(mem_index) +
1698 offsetof(CPUTLBDescFast, table));
1700 /* If the required alignment is at least as large as the access, simply
1701 copy the address and mask. For lesser alignments, check that we don't
1702 cross pages for the complete access. */
1703 if (a_bits >= s_bits) {
1704 tcg_out_mov(s, ttype, r1, addrlo);
1706 tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
1708 tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
1709 tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
1712 tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
1714 /* Prepare for both the fast path add of the tlb addend, and the slow
1715 path function argument setup. */
1716 tcg_out_mov(s, ttype, r1, addrlo);
1719 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1720 label_ptr[0] = s->code_ptr;
1723 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1724 /* cmp 4(r0), addrhi */
1725 tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
1728 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1729 label_ptr[1] = s->code_ptr;
1735 /* add addend(r0), r1 */
1736 tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
1737 offsetof(CPUTLBEntry, addend));
1741 * Record the context of a call to the out of line helper code for the slow path
1742 * for a load or store, so that we can later generate the correct helper code
1744 static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
1746 TCGReg datalo, TCGReg datahi,
1747 TCGReg addrlo, TCGReg addrhi,
1748 tcg_insn_unit *raddr,
1749 tcg_insn_unit **label_ptr)
1751 TCGLabelQemuLdst *label = new_ldst_label(s);
1753 label->is_ld = is_ld;
1755 label->type = is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32;
1756 label->datalo_reg = datalo;
1757 label->datahi_reg = datahi;
1758 label->addrlo_reg = addrlo;
1759 label->addrhi_reg = addrhi;
1760 label->raddr = tcg_splitwx_to_rx(raddr);
1761 label->label_ptr[0] = label_ptr[0];
1762 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1763 label->label_ptr[1] = label_ptr[1];
1768 * Generate code for the slow path for a load at the end of block
1770 static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1772 TCGMemOpIdx oi = l->oi;
1773 MemOp opc = get_memop(oi);
1775 tcg_insn_unit **label_ptr = &l->label_ptr[0];
1776 int rexw = (l->type == TCG_TYPE_I64 ? P_REXW : 0);
1778 /* resolve label address */
1779 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1780 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1781 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1784 if (TCG_TARGET_REG_BITS == 32) {
1787 tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1790 tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1793 if (TARGET_LONG_BITS == 64) {
1794 tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1798 tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1801 tcg_out_sti(s, TCG_TYPE_PTR, (uintptr_t)l->raddr, TCG_REG_ESP, ofs);
1803 tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1804 /* The second argument is already loaded with addrlo. */
1805 tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2], oi);
1806 tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
1807 (uintptr_t)l->raddr);
1810 tcg_out_call(s, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1812 data_reg = l->datalo_reg;
1813 switch (opc & MO_SSIZE) {
1815 tcg_out_ext8s(s, data_reg, TCG_REG_EAX, rexw);
1818 tcg_out_ext16s(s, data_reg, TCG_REG_EAX, rexw);
1820 #if TCG_TARGET_REG_BITS == 64
1822 tcg_out_ext32s(s, data_reg, TCG_REG_EAX);
1827 /* Note that the helpers have zero-extended to tcg_target_long. */
1829 tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1832 if (TCG_TARGET_REG_BITS == 64) {
1833 tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
1834 } else if (data_reg == TCG_REG_EDX) {
1835 /* xchg %edx, %eax */
1836 tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
1837 tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
1839 tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1840 tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
1847 /* Jump to the code corresponding to next IR of qemu_st */
1848 tcg_out_jmp(s, l->raddr);
1853 * Generate code for the slow path for a store at the end of block
1855 static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1857 TCGMemOpIdx oi = l->oi;
1858 MemOp opc = get_memop(oi);
1859 MemOp s_bits = opc & MO_SIZE;
1860 tcg_insn_unit **label_ptr = &l->label_ptr[0];
1863 /* resolve label address */
1864 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1865 if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1866 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1869 if (TCG_TARGET_REG_BITS == 32) {
1872 tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1875 tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1878 if (TARGET_LONG_BITS == 64) {
1879 tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1883 tcg_out_st(s, TCG_TYPE_I32, l->datalo_reg, TCG_REG_ESP, ofs);
1886 if (s_bits == MO_64) {
1887 tcg_out_st(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_ESP, ofs);
1891 tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1894 retaddr = TCG_REG_EAX;
1895 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1896 tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
1898 tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1899 /* The second argument is already loaded with addrlo. */
1900 tcg_out_mov(s, (s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
1901 tcg_target_call_iarg_regs[2], l->datalo_reg);
1902 tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
1904 if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
1905 retaddr = tcg_target_call_iarg_regs[4];
1906 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1908 retaddr = TCG_REG_RAX;
1909 tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1910 tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP,
1911 TCG_TARGET_CALL_STACK_OFFSET);
1915 /* "Tail call" to the helper, with the return address back inline. */
1916 tcg_out_push(s, retaddr);
1917 tcg_out_jmp(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1920 #elif TCG_TARGET_REG_BITS == 32
1921 # define x86_guest_base_seg 0
1922 # define x86_guest_base_index -1
1923 # define x86_guest_base_offset guest_base
1925 static int x86_guest_base_seg;
1926 static int x86_guest_base_index = -1;
1927 static int32_t x86_guest_base_offset;
1928 # if defined(__x86_64__) && defined(__linux__)
1929 # include <asm/prctl.h>
1930 # include <sys/prctl.h>
1931 int arch_prctl(int code, unsigned long addr);
1932 static inline int setup_guest_base_seg(void)
1934 if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
1939 # elif defined (__FreeBSD__) || defined (__FreeBSD_kernel__)
1940 # include <machine/sysarch.h>
1941 static inline int setup_guest_base_seg(void)
1943 if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
1949 static inline int setup_guest_base_seg(void)
1954 #endif /* SOFTMMU */
1956 static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
1957 TCGReg base, int index, intptr_t ofs,
1958 int seg, bool is64, MemOp memop)
1960 bool use_movbe = false;
1961 int rexw = is64 * P_REXW;
1962 int movop = OPC_MOVL_GvEv;
1964 /* Do big-endian loads with movbe. */
1965 if (memop & MO_BSWAP) {
1966 tcg_debug_assert(have_movbe);
1968 movop = OPC_MOVBE_GyMy;
1971 switch (memop & MO_SSIZE) {
1973 tcg_out_modrm_sib_offset(s, OPC_MOVZBL + seg, datalo,
1974 base, index, 0, ofs);
1977 tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + seg, datalo,
1978 base, index, 0, ofs);
1982 /* There is no extending movbe; only low 16-bits are modified. */
1983 if (datalo != base && datalo != index) {
1984 /* XOR breaks dependency chains. */
1985 tgen_arithr(s, ARITH_XOR, datalo, datalo);
1986 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
1987 datalo, base, index, 0, ofs);
1989 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
1990 datalo, base, index, 0, ofs);
1991 tcg_out_ext16u(s, datalo, datalo);
1994 tcg_out_modrm_sib_offset(s, OPC_MOVZWL + seg, datalo,
1995 base, index, 0, ofs);
2000 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2001 datalo, base, index, 0, ofs);
2002 tcg_out_ext16s(s, datalo, datalo, rexw);
2004 tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + seg,
2005 datalo, base, index, 0, ofs);
2009 tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2011 #if TCG_TARGET_REG_BITS == 64
2014 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + seg, datalo,
2015 base, index, 0, ofs);
2016 tcg_out_ext32s(s, datalo, datalo);
2018 tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + seg, datalo,
2019 base, index, 0, ofs);
2024 if (TCG_TARGET_REG_BITS == 64) {
2025 tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2026 base, index, 0, ofs);
2033 if (base != datalo) {
2034 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2035 base, index, 0, ofs);
2036 tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2037 base, index, 0, ofs + 4);
2039 tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2040 base, index, 0, ofs + 4);
2041 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2042 base, index, 0, ofs);
2047 g_assert_not_reached();
2051 /* XXX: qemu_ld and qemu_st could be modified to clobber only EDX and
2052 EAX. It will be useful once fixed registers globals are less
2054 static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
2056 TCGReg datalo, datahi, addrlo;
2057 TCGReg addrhi __attribute__((unused));
2060 #if defined(CONFIG_SOFTMMU)
2062 tcg_insn_unit *label_ptr[2];
2066 datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2068 addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2070 opc = get_memop(oi);
2072 #if defined(CONFIG_SOFTMMU)
2073 mem_index = get_mmuidx(oi);
2075 tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2076 label_ptr, offsetof(CPUTLBEntry, addr_read));
2079 tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, is64, opc);
2081 /* Record the current context of a load into ldst label */
2082 add_qemu_ldst_label(s, true, is64, oi, datalo, datahi, addrlo, addrhi,
2083 s->code_ptr, label_ptr);
2085 tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2086 x86_guest_base_offset, x86_guest_base_seg,
2091 static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2092 TCGReg base, int index, intptr_t ofs,
2093 int seg, MemOp memop)
2095 bool use_movbe = false;
2096 int movop = OPC_MOVL_EvGv;
2099 * Do big-endian stores with movbe or softmmu.
2100 * User-only without movbe will have its swapping done generically.
2102 if (memop & MO_BSWAP) {
2103 tcg_debug_assert(have_movbe);
2105 movop = OPC_MOVBE_MyGy;
2108 switch (memop & MO_SIZE) {
2110 /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2111 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2112 tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + seg,
2113 datalo, base, index, 0, ofs);
2116 tcg_out_modrm_sib_offset(s, movop + P_DATA16 + seg, datalo,
2117 base, index, 0, ofs);
2120 tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2123 if (TCG_TARGET_REG_BITS == 64) {
2124 tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2125 base, index, 0, ofs);
2132 tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2133 base, index, 0, ofs);
2134 tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2135 base, index, 0, ofs + 4);
2139 g_assert_not_reached();
2143 static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
2145 TCGReg datalo, datahi, addrlo;
2146 TCGReg addrhi __attribute__((unused));
2149 #if defined(CONFIG_SOFTMMU)
2151 tcg_insn_unit *label_ptr[2];
2155 datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2157 addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2159 opc = get_memop(oi);
2161 #if defined(CONFIG_SOFTMMU)
2162 mem_index = get_mmuidx(oi);
2164 tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2165 label_ptr, offsetof(CPUTLBEntry, addr_write));
2168 tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
2170 /* Record the current context of a store into ldst label */
2171 add_qemu_ldst_label(s, false, is64, oi, datalo, datahi, addrlo, addrhi,
2172 s->code_ptr, label_ptr);
2174 tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2175 x86_guest_base_offset, x86_guest_base_seg, opc);
2179 static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2180 const TCGArg *args, const int *const_args)
2183 int c, const_a2, vexop, rexw = 0;
2185 #if TCG_TARGET_REG_BITS == 64
2186 # define OP_32_64(x) \
2187 case glue(glue(INDEX_op_, x), _i64): \
2188 rexw = P_REXW; /* FALLTHRU */ \
2189 case glue(glue(INDEX_op_, x), _i32)
2191 # define OP_32_64(x) \
2192 case glue(glue(INDEX_op_, x), _i32)
2195 /* Hoist the loads of the most common arguments. */
2199 const_a2 = const_args[2];
2202 case INDEX_op_exit_tb:
2203 /* Reuse the zeroing that exists for goto_ptr. */
2205 tcg_out_jmp(s, tcg_code_gen_epilogue);
2207 tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2208 tcg_out_jmp(s, tb_ret_addr);
2211 case INDEX_op_goto_tb:
2212 if (s->tb_jmp_insn_offset) {
2213 /* direct jump method */
2215 /* jump displacement must be aligned for atomic patching;
2216 * see if we need to add extra nops before jump
2218 gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2220 tcg_out_nopn(s, gap - 1);
2222 tcg_out8(s, OPC_JMP_long); /* jmp im */
2223 s->tb_jmp_insn_offset[a0] = tcg_current_code_size(s);
2226 /* indirect jump method */
2227 tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, -1,
2228 (intptr_t)(s->tb_jmp_target_addr + a0));
2230 set_jmp_reset_offset(s, a0);
2232 case INDEX_op_goto_ptr:
2233 /* jmp to the given host address (could be epilogue) */
2234 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2237 tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2240 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
2241 tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2244 tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2247 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
2248 tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2251 tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2253 #if TCG_TARGET_REG_BITS == 64
2254 case INDEX_op_ld32u_i64:
2256 case INDEX_op_ld_i32:
2257 tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2261 if (const_args[0]) {
2262 tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2265 tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2269 if (const_args[0]) {
2270 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2273 tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2276 #if TCG_TARGET_REG_BITS == 64
2277 case INDEX_op_st32_i64:
2279 case INDEX_op_st_i32:
2280 if (const_args[0]) {
2281 tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2284 tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2289 /* For 3-operand addition, use LEA. */
2294 } else if (a0 == a2) {
2295 /* Watch out for dest = src + dest, since we've removed
2296 the matching constraint on the add. */
2297 tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2301 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2320 tgen_arithi(s, c + rexw, a0, a2, 0);
2322 tgen_arithr(s, c + rexw, a0, a2);
2328 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2329 tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2331 tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2339 if (val == (int8_t)val) {
2340 tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2343 tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2347 tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2352 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2355 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2359 /* For small constant 3-operand shift, use LEA. */
2360 if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2362 /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2363 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2365 /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2366 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2372 goto gen_shift_maybe_vex;
2376 goto gen_shift_maybe_vex;
2380 goto gen_shift_maybe_vex;
2387 gen_shift_maybe_vex:
2390 tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2393 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2398 tcg_out_shifti(s, c + rexw, a0, a2);
2400 tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2405 tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2408 tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2411 tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2414 case INDEX_op_brcond_i32:
2415 tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2417 case INDEX_op_setcond_i32:
2418 tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2420 case INDEX_op_movcond_i32:
2421 tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2425 tcg_out_rolw_8(s, a0);
2428 tcg_out_bswap32(s, a0);
2432 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2435 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2439 tcg_out_ext8s(s, a0, a1, rexw);
2442 tcg_out_ext16s(s, a0, a1, rexw);
2445 tcg_out_ext8u(s, a0, a1);
2448 tcg_out_ext16u(s, a0, a1);
2451 case INDEX_op_qemu_ld_i32:
2452 tcg_out_qemu_ld(s, args, 0);
2454 case INDEX_op_qemu_ld_i64:
2455 tcg_out_qemu_ld(s, args, 1);
2457 case INDEX_op_qemu_st_i32:
2458 case INDEX_op_qemu_st8_i32:
2459 tcg_out_qemu_st(s, args, 0);
2461 case INDEX_op_qemu_st_i64:
2462 tcg_out_qemu_st(s, args, 1);
2466 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2469 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2472 if (const_args[4]) {
2473 tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2475 tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2477 if (const_args[5]) {
2478 tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2480 tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2484 if (const_args[4]) {
2485 tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2487 tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2489 if (const_args[5]) {
2490 tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2492 tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2496 #if TCG_TARGET_REG_BITS == 32
2497 case INDEX_op_brcond2_i32:
2498 tcg_out_brcond2(s, args, const_args, 0);
2500 case INDEX_op_setcond2_i32:
2501 tcg_out_setcond2(s, args, const_args);
2503 #else /* TCG_TARGET_REG_BITS == 64 */
2504 case INDEX_op_ld32s_i64:
2505 tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2507 case INDEX_op_ld_i64:
2508 tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2510 case INDEX_op_st_i64:
2511 if (const_args[0]) {
2512 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2515 tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2519 case INDEX_op_brcond_i64:
2520 tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2522 case INDEX_op_setcond_i64:
2523 tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2525 case INDEX_op_movcond_i64:
2526 tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2529 case INDEX_op_bswap64_i64:
2530 tcg_out_bswap64(s, a0);
2532 case INDEX_op_extu_i32_i64:
2533 case INDEX_op_ext32u_i64:
2534 case INDEX_op_extrl_i64_i32:
2535 tcg_out_ext32u(s, a0, a1);
2537 case INDEX_op_ext_i32_i64:
2538 case INDEX_op_ext32s_i64:
2539 tcg_out_ext32s(s, a0, a1);
2541 case INDEX_op_extrh_i64_i32:
2542 tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2547 if (args[3] == 0 && args[4] == 8) {
2548 /* load bits 0..7 */
2549 tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2550 } else if (args[3] == 8 && args[4] == 8) {
2551 /* load bits 8..15 */
2552 tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2553 } else if (args[3] == 0 && args[4] == 16) {
2554 /* load bits 0..15 */
2555 tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2561 case INDEX_op_extract_i64:
2562 if (a2 + args[3] == 32) {
2563 /* This is a 32-bit zero-extending right shift. */
2564 tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2565 tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2569 case INDEX_op_extract_i32:
2570 /* On the off-chance that we can use the high-byte registers.
2571 Otherwise we emit the same ext16 + shift pattern that we
2572 would have gotten from the normal tcg-op.c expansion. */
2573 tcg_debug_assert(a2 == 8 && args[3] == 8);
2574 if (a1 < 4 && a0 < 8) {
2575 tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2577 tcg_out_ext16u(s, a0, a1);
2578 tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2582 case INDEX_op_sextract_i32:
2583 /* We don't implement sextract_i64, as we cannot sign-extend to
2584 64-bits without using the REX prefix that explicitly excludes
2585 access to the high-byte registers. */
2586 tcg_debug_assert(a2 == 8 && args[3] == 8);
2587 if (a1 < 4 && a0 < 8) {
2588 tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2590 tcg_out_ext16s(s, a0, a1, 0);
2591 tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2596 /* Note that SHRD outputs to the r/m operand. */
2597 tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2598 tcg_out8(s, args[3]);
2604 case INDEX_op_mov_i32: /* Always emitted via tcg_out_mov. */
2605 case INDEX_op_mov_i64:
2606 case INDEX_op_call: /* Always emitted via tcg_out_call. */
2614 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2615 unsigned vecl, unsigned vece,
2616 const TCGArg *args, const int *const_args)
2618 static int const add_insn[4] = {
2619 OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2621 static int const ssadd_insn[4] = {
2622 OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2624 static int const usadd_insn[4] = {
2625 OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2627 static int const sub_insn[4] = {
2628 OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2630 static int const sssub_insn[4] = {
2631 OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2633 static int const ussub_insn[4] = {
2634 OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2636 static int const mul_insn[4] = {
2637 OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_UD2
2639 static int const shift_imm_insn[4] = {
2640 OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2642 static int const cmpeq_insn[4] = {
2643 OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2645 static int const cmpgt_insn[4] = {
2646 OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2648 static int const punpckl_insn[4] = {
2649 OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2651 static int const punpckh_insn[4] = {
2652 OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2654 static int const packss_insn[4] = {
2655 OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2657 static int const packus_insn[4] = {
2658 OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2660 static int const smin_insn[4] = {
2661 OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_UD2
2663 static int const smax_insn[4] = {
2664 OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_UD2
2666 static int const umin_insn[4] = {
2667 OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_UD2
2669 static int const umax_insn[4] = {
2670 OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_UD2
2672 static int const shlv_insn[4] = {
2673 /* TODO: AVX512 adds support for MO_16. */
2674 OPC_UD2, OPC_UD2, OPC_VPSLLVD, OPC_VPSLLVQ
2676 static int const shrv_insn[4] = {
2677 /* TODO: AVX512 adds support for MO_16. */
2678 OPC_UD2, OPC_UD2, OPC_VPSRLVD, OPC_VPSRLVQ
2680 static int const sarv_insn[4] = {
2681 /* TODO: AVX512 adds support for MO_16, MO_64. */
2682 OPC_UD2, OPC_UD2, OPC_VPSRAVD, OPC_UD2
2684 static int const shls_insn[4] = {
2685 OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2687 static int const shrs_insn[4] = {
2688 OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2690 static int const sars_insn[4] = {
2691 OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_UD2
2693 static int const abs_insn[4] = {
2694 /* TODO: AVX512 adds support for MO_64. */
2695 OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_UD2
2698 TCGType type = vecl + TCG_TYPE_V64;
2707 case INDEX_op_add_vec:
2708 insn = add_insn[vece];
2710 case INDEX_op_ssadd_vec:
2711 insn = ssadd_insn[vece];
2713 case INDEX_op_usadd_vec:
2714 insn = usadd_insn[vece];
2716 case INDEX_op_sub_vec:
2717 insn = sub_insn[vece];
2719 case INDEX_op_sssub_vec:
2720 insn = sssub_insn[vece];
2722 case INDEX_op_ussub_vec:
2723 insn = ussub_insn[vece];
2725 case INDEX_op_mul_vec:
2726 insn = mul_insn[vece];
2728 case INDEX_op_and_vec:
2731 case INDEX_op_or_vec:
2734 case INDEX_op_xor_vec:
2737 case INDEX_op_smin_vec:
2738 insn = smin_insn[vece];
2740 case INDEX_op_umin_vec:
2741 insn = umin_insn[vece];
2743 case INDEX_op_smax_vec:
2744 insn = smax_insn[vece];
2746 case INDEX_op_umax_vec:
2747 insn = umax_insn[vece];
2749 case INDEX_op_shlv_vec:
2750 insn = shlv_insn[vece];
2752 case INDEX_op_shrv_vec:
2753 insn = shrv_insn[vece];
2755 case INDEX_op_sarv_vec:
2756 insn = sarv_insn[vece];
2758 case INDEX_op_shls_vec:
2759 insn = shls_insn[vece];
2761 case INDEX_op_shrs_vec:
2762 insn = shrs_insn[vece];
2764 case INDEX_op_sars_vec:
2765 insn = sars_insn[vece];
2767 case INDEX_op_x86_punpckl_vec:
2768 insn = punpckl_insn[vece];
2770 case INDEX_op_x86_punpckh_vec:
2771 insn = punpckh_insn[vece];
2773 case INDEX_op_x86_packss_vec:
2774 insn = packss_insn[vece];
2776 case INDEX_op_x86_packus_vec:
2777 insn = packus_insn[vece];
2779 #if TCG_TARGET_REG_BITS == 32
2780 case INDEX_op_dup2_vec:
2781 /* First merge the two 32-bit inputs to a single 64-bit element. */
2782 tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
2783 /* Then replicate the 64-bit elements across the rest of the vector. */
2784 if (type != TCG_TYPE_V64) {
2785 tcg_out_dup_vec(s, type, MO_64, a0, a0);
2789 case INDEX_op_abs_vec:
2790 insn = abs_insn[vece];
2795 tcg_debug_assert(insn != OPC_UD2);
2796 if (type == TCG_TYPE_V256) {
2799 tcg_out_vex_modrm(s, insn, a0, a1, a2);
2802 case INDEX_op_cmp_vec:
2804 if (sub == TCG_COND_EQ) {
2805 insn = cmpeq_insn[vece];
2806 } else if (sub == TCG_COND_GT) {
2807 insn = cmpgt_insn[vece];
2809 g_assert_not_reached();
2813 case INDEX_op_andc_vec:
2815 if (type == TCG_TYPE_V256) {
2818 tcg_out_vex_modrm(s, insn, a0, a2, a1);
2821 case INDEX_op_shli_vec:
2824 case INDEX_op_shri_vec:
2827 case INDEX_op_sari_vec:
2828 tcg_debug_assert(vece != MO_64);
2831 tcg_debug_assert(vece != MO_8);
2832 insn = shift_imm_insn[vece];
2833 if (type == TCG_TYPE_V256) {
2836 tcg_out_vex_modrm(s, insn, sub, a0, a1);
2840 case INDEX_op_ld_vec:
2841 tcg_out_ld(s, type, a0, a1, a2);
2843 case INDEX_op_st_vec:
2844 tcg_out_st(s, type, a0, a1, a2);
2846 case INDEX_op_dupm_vec:
2847 tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2850 case INDEX_op_x86_shufps_vec:
2854 case INDEX_op_x86_blend_vec:
2855 if (vece == MO_16) {
2857 } else if (vece == MO_32) {
2858 insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
2860 g_assert_not_reached();
2864 case INDEX_op_x86_vperm2i128_vec:
2865 insn = OPC_VPERM2I128;
2869 if (type == TCG_TYPE_V256) {
2872 tcg_out_vex_modrm(s, insn, a0, a1, a2);
2876 case INDEX_op_x86_vpblendvb_vec:
2877 insn = OPC_VPBLENDVB;
2878 if (type == TCG_TYPE_V256) {
2881 tcg_out_vex_modrm(s, insn, a0, a1, a2);
2882 tcg_out8(s, args[3] << 4);
2885 case INDEX_op_x86_psrldq_vec:
2886 tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
2890 case INDEX_op_mov_vec: /* Always emitted via tcg_out_mov. */
2891 case INDEX_op_dup_vec: /* Always emitted via tcg_out_dup_vec. */
2893 g_assert_not_reached();
2897 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2900 case INDEX_op_goto_ptr:
2903 case INDEX_op_ld8u_i32:
2904 case INDEX_op_ld8u_i64:
2905 case INDEX_op_ld8s_i32:
2906 case INDEX_op_ld8s_i64:
2907 case INDEX_op_ld16u_i32:
2908 case INDEX_op_ld16u_i64:
2909 case INDEX_op_ld16s_i32:
2910 case INDEX_op_ld16s_i64:
2911 case INDEX_op_ld_i32:
2912 case INDEX_op_ld32u_i64:
2913 case INDEX_op_ld32s_i64:
2914 case INDEX_op_ld_i64:
2915 return C_O1_I1(r, r);
2917 case INDEX_op_st8_i32:
2918 case INDEX_op_st8_i64:
2919 return C_O0_I2(qi, r);
2921 case INDEX_op_st16_i32:
2922 case INDEX_op_st16_i64:
2923 case INDEX_op_st_i32:
2924 case INDEX_op_st32_i64:
2925 return C_O0_I2(ri, r);
2927 case INDEX_op_st_i64:
2928 return C_O0_I2(re, r);
2930 case INDEX_op_add_i32:
2931 case INDEX_op_add_i64:
2932 return C_O1_I2(r, r, re);
2934 case INDEX_op_sub_i32:
2935 case INDEX_op_sub_i64:
2936 case INDEX_op_mul_i32:
2937 case INDEX_op_mul_i64:
2938 case INDEX_op_or_i32:
2939 case INDEX_op_or_i64:
2940 case INDEX_op_xor_i32:
2941 case INDEX_op_xor_i64:
2942 return C_O1_I2(r, 0, re);
2944 case INDEX_op_and_i32:
2945 case INDEX_op_and_i64:
2946 return C_O1_I2(r, 0, reZ);
2948 case INDEX_op_andc_i32:
2949 case INDEX_op_andc_i64:
2950 return C_O1_I2(r, r, rI);
2952 case INDEX_op_shl_i32:
2953 case INDEX_op_shl_i64:
2954 case INDEX_op_shr_i32:
2955 case INDEX_op_shr_i64:
2956 case INDEX_op_sar_i32:
2957 case INDEX_op_sar_i64:
2958 return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
2960 case INDEX_op_rotl_i32:
2961 case INDEX_op_rotl_i64:
2962 case INDEX_op_rotr_i32:
2963 case INDEX_op_rotr_i64:
2964 return C_O1_I2(r, 0, ci);
2966 case INDEX_op_brcond_i32:
2967 case INDEX_op_brcond_i64:
2968 return C_O0_I2(r, re);
2970 case INDEX_op_bswap16_i32:
2971 case INDEX_op_bswap16_i64:
2972 case INDEX_op_bswap32_i32:
2973 case INDEX_op_bswap32_i64:
2974 case INDEX_op_bswap64_i64:
2975 case INDEX_op_neg_i32:
2976 case INDEX_op_neg_i64:
2977 case INDEX_op_not_i32:
2978 case INDEX_op_not_i64:
2979 case INDEX_op_extrh_i64_i32:
2980 return C_O1_I1(r, 0);
2982 case INDEX_op_ext8s_i32:
2983 case INDEX_op_ext8s_i64:
2984 case INDEX_op_ext8u_i32:
2985 case INDEX_op_ext8u_i64:
2986 return C_O1_I1(r, q);
2988 case INDEX_op_ext16s_i32:
2989 case INDEX_op_ext16s_i64:
2990 case INDEX_op_ext16u_i32:
2991 case INDEX_op_ext16u_i64:
2992 case INDEX_op_ext32s_i64:
2993 case INDEX_op_ext32u_i64:
2994 case INDEX_op_ext_i32_i64:
2995 case INDEX_op_extu_i32_i64:
2996 case INDEX_op_extrl_i64_i32:
2997 case INDEX_op_extract_i32:
2998 case INDEX_op_extract_i64:
2999 case INDEX_op_sextract_i32:
3000 case INDEX_op_ctpop_i32:
3001 case INDEX_op_ctpop_i64:
3002 return C_O1_I1(r, r);
3004 case INDEX_op_extract2_i32:
3005 case INDEX_op_extract2_i64:
3006 return C_O1_I2(r, 0, r);
3008 case INDEX_op_deposit_i32:
3009 case INDEX_op_deposit_i64:
3010 return C_O1_I2(Q, 0, Q);
3012 case INDEX_op_setcond_i32:
3013 case INDEX_op_setcond_i64:
3014 return C_O1_I2(q, r, re);
3016 case INDEX_op_movcond_i32:
3017 case INDEX_op_movcond_i64:
3018 return C_O1_I4(r, r, re, r, 0);
3020 case INDEX_op_div2_i32:
3021 case INDEX_op_div2_i64:
3022 case INDEX_op_divu2_i32:
3023 case INDEX_op_divu2_i64:
3024 return C_O2_I3(a, d, 0, 1, r);
3026 case INDEX_op_mulu2_i32:
3027 case INDEX_op_mulu2_i64:
3028 case INDEX_op_muls2_i32:
3029 case INDEX_op_muls2_i64:
3030 return C_O2_I2(a, d, a, r);
3032 case INDEX_op_add2_i32:
3033 case INDEX_op_add2_i64:
3034 case INDEX_op_sub2_i32:
3035 case INDEX_op_sub2_i64:
3036 return C_O2_I4(r, r, 0, 1, re, re);
3038 case INDEX_op_ctz_i32:
3039 case INDEX_op_ctz_i64:
3040 return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3042 case INDEX_op_clz_i32:
3043 case INDEX_op_clz_i64:
3044 return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3046 case INDEX_op_qemu_ld_i32:
3047 return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3048 ? C_O1_I1(r, L) : C_O1_I2(r, L, L));
3050 case INDEX_op_qemu_st_i32:
3051 return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3052 ? C_O0_I2(L, L) : C_O0_I3(L, L, L));
3053 case INDEX_op_qemu_st8_i32:
3054 return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3055 ? C_O0_I2(s, L) : C_O0_I3(s, L, L));
3057 case INDEX_op_qemu_ld_i64:
3058 return (TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L)
3059 : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O2_I1(r, r, L)
3060 : C_O2_I2(r, r, L, L));
3062 case INDEX_op_qemu_st_i64:
3063 return (TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L)
3064 : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O0_I3(L, L, L)
3065 : C_O0_I4(L, L, L, L));
3067 case INDEX_op_brcond2_i32:
3068 return C_O0_I4(r, r, ri, ri);
3070 case INDEX_op_setcond2_i32:
3071 return C_O1_I4(r, r, r, ri, ri);
3073 case INDEX_op_ld_vec:
3074 case INDEX_op_dupm_vec:
3075 return C_O1_I1(x, r);
3077 case INDEX_op_st_vec:
3078 return C_O0_I2(x, r);
3080 case INDEX_op_add_vec:
3081 case INDEX_op_sub_vec:
3082 case INDEX_op_mul_vec:
3083 case INDEX_op_and_vec:
3084 case INDEX_op_or_vec:
3085 case INDEX_op_xor_vec:
3086 case INDEX_op_andc_vec:
3087 case INDEX_op_ssadd_vec:
3088 case INDEX_op_usadd_vec:
3089 case INDEX_op_sssub_vec:
3090 case INDEX_op_ussub_vec:
3091 case INDEX_op_smin_vec:
3092 case INDEX_op_umin_vec:
3093 case INDEX_op_smax_vec:
3094 case INDEX_op_umax_vec:
3095 case INDEX_op_shlv_vec:
3096 case INDEX_op_shrv_vec:
3097 case INDEX_op_sarv_vec:
3098 case INDEX_op_shls_vec:
3099 case INDEX_op_shrs_vec:
3100 case INDEX_op_sars_vec:
3101 case INDEX_op_rotls_vec:
3102 case INDEX_op_cmp_vec:
3103 case INDEX_op_x86_shufps_vec:
3104 case INDEX_op_x86_blend_vec:
3105 case INDEX_op_x86_packss_vec:
3106 case INDEX_op_x86_packus_vec:
3107 case INDEX_op_x86_vperm2i128_vec:
3108 case INDEX_op_x86_punpckl_vec:
3109 case INDEX_op_x86_punpckh_vec:
3110 #if TCG_TARGET_REG_BITS == 32
3111 case INDEX_op_dup2_vec:
3113 return C_O1_I2(x, x, x);
3115 case INDEX_op_abs_vec:
3116 case INDEX_op_dup_vec:
3117 case INDEX_op_shli_vec:
3118 case INDEX_op_shri_vec:
3119 case INDEX_op_sari_vec:
3120 case INDEX_op_x86_psrldq_vec:
3121 return C_O1_I1(x, x);
3123 case INDEX_op_x86_vpblendvb_vec:
3124 return C_O1_I3(x, x, x, x);
3127 g_assert_not_reached();
3131 int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3134 case INDEX_op_add_vec:
3135 case INDEX_op_sub_vec:
3136 case INDEX_op_and_vec:
3137 case INDEX_op_or_vec:
3138 case INDEX_op_xor_vec:
3139 case INDEX_op_andc_vec:
3141 case INDEX_op_rotli_vec:
3142 case INDEX_op_cmp_vec:
3143 case INDEX_op_cmpsel_vec:
3146 case INDEX_op_shli_vec:
3147 case INDEX_op_shri_vec:
3148 /* We must expand the operation for MO_8. */
3149 return vece == MO_8 ? -1 : 1;
3151 case INDEX_op_sari_vec:
3152 /* We must expand the operation for MO_8. */
3156 /* We can emulate this for MO_64, but it does not pay off
3157 unless we're producing at least 4 values. */
3158 if (vece == MO_64) {
3159 return type >= TCG_TYPE_V256 ? -1 : 0;
3163 case INDEX_op_shls_vec:
3164 case INDEX_op_shrs_vec:
3165 return vece >= MO_16;
3166 case INDEX_op_sars_vec:
3167 return vece >= MO_16 && vece <= MO_32;
3168 case INDEX_op_rotls_vec:
3169 return vece >= MO_16 ? -1 : 0;
3171 case INDEX_op_shlv_vec:
3172 case INDEX_op_shrv_vec:
3173 return have_avx2 && vece >= MO_32;
3174 case INDEX_op_sarv_vec:
3175 return have_avx2 && vece == MO_32;
3176 case INDEX_op_rotlv_vec:
3177 case INDEX_op_rotrv_vec:
3178 return have_avx2 && vece >= MO_32 ? -1 : 0;
3180 case INDEX_op_mul_vec:
3182 /* We can expand the operation for MO_8. */
3185 if (vece == MO_64) {
3190 case INDEX_op_ssadd_vec:
3191 case INDEX_op_usadd_vec:
3192 case INDEX_op_sssub_vec:
3193 case INDEX_op_ussub_vec:
3194 return vece <= MO_16;
3195 case INDEX_op_smin_vec:
3196 case INDEX_op_smax_vec:
3197 case INDEX_op_umin_vec:
3198 case INDEX_op_umax_vec:
3199 case INDEX_op_abs_vec:
3200 return vece <= MO_32;
3207 static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3208 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3212 tcg_debug_assert(vece == MO_8);
3214 t1 = tcg_temp_new_vec(type);
3215 t2 = tcg_temp_new_vec(type);
3218 * Unpack to W, shift, and repack. Tricky bits:
3219 * (1) Use punpck*bw x,x to produce DDCCBBAA,
3220 * i.e. duplicate in other half of the 16-bit lane.
3221 * (2) For right-shift, add 8 so that the high half of the lane
3222 * becomes zero. For left-shift, and left-rotate, we must
3223 * shift up and down again.
3224 * (3) Step 2 leaves high half zero such that PACKUSWB
3225 * (pack with unsigned saturation) does not modify
3228 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3229 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3230 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3231 tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3233 if (opc != INDEX_op_rotli_vec) {
3236 if (opc == INDEX_op_shri_vec) {
3237 tcg_gen_shri_vec(MO_16, t1, t1, imm);
3238 tcg_gen_shri_vec(MO_16, t2, t2, imm);
3240 tcg_gen_shli_vec(MO_16, t1, t1, imm);
3241 tcg_gen_shli_vec(MO_16, t2, t2, imm);
3242 tcg_gen_shri_vec(MO_16, t1, t1, 8);
3243 tcg_gen_shri_vec(MO_16, t2, t2, 8);
3246 vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3247 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3248 tcg_temp_free_vec(t1);
3249 tcg_temp_free_vec(t2);
3252 static void expand_vec_sari(TCGType type, unsigned vece,
3253 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3259 /* Unpack to W, shift, and repack, as in expand_vec_shi. */
3260 t1 = tcg_temp_new_vec(type);
3261 t2 = tcg_temp_new_vec(type);
3262 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3263 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3264 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3265 tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3266 tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3267 tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3268 vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3269 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3270 tcg_temp_free_vec(t1);
3271 tcg_temp_free_vec(t2);
3277 * We can emulate a small sign extend by performing an arithmetic
3278 * 32-bit shift and overwriting the high half of a 64-bit logical
3279 * shift. Note that the ISA says shift of 32 is valid, but TCG
3280 * does not, so we have to bound the smaller shift -- we get the
3281 * same result in the high half either way.
3283 t1 = tcg_temp_new_vec(type);
3284 tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3285 tcg_gen_shri_vec(MO_64, v0, v1, imm);
3286 vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3287 tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3288 tcgv_vec_arg(t1), 0xaa);
3289 tcg_temp_free_vec(t1);
3291 /* Otherwise we will need to use a compare vs 0 to produce
3292 * the sign-extend, shift and merge.
3294 t1 = tcg_const_zeros_vec(type);
3295 tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1, t1, v1);
3296 tcg_gen_shri_vec(MO_64, v0, v1, imm);
3297 tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3298 tcg_gen_or_vec(MO_64, v0, v0, t1);
3299 tcg_temp_free_vec(t1);
3304 g_assert_not_reached();
3308 static void expand_vec_rotli(TCGType type, unsigned vece,
3309 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3314 expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3318 t = tcg_temp_new_vec(type);
3319 tcg_gen_shli_vec(vece, t, v1, imm);
3320 tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3321 tcg_gen_or_vec(vece, v0, v0, t);
3322 tcg_temp_free_vec(t);
3325 static void expand_vec_rotls(TCGType type, unsigned vece,
3326 TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3331 tcg_debug_assert(vece != MO_8);
3333 t = tcg_temp_new_vec(type);
3334 rsh = tcg_temp_new_i32();
3336 tcg_gen_neg_i32(rsh, lsh);
3337 tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3338 tcg_gen_shls_vec(vece, t, v1, lsh);
3339 tcg_gen_shrs_vec(vece, v0, v1, rsh);
3340 tcg_gen_or_vec(vece, v0, v0, t);
3341 tcg_temp_free_vec(t);
3342 tcg_temp_free_i32(rsh);
3345 static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3346 TCGv_vec v1, TCGv_vec sh, bool right)
3348 TCGv_vec t = tcg_temp_new_vec(type);
3350 tcg_gen_dupi_vec(vece, t, 8 << vece);
3351 tcg_gen_sub_vec(vece, t, t, sh);
3353 tcg_gen_shlv_vec(vece, t, v1, t);
3354 tcg_gen_shrv_vec(vece, v0, v1, sh);
3356 tcg_gen_shrv_vec(vece, t, v1, t);
3357 tcg_gen_shlv_vec(vece, v0, v1, sh);
3359 tcg_gen_or_vec(vece, v0, v0, t);
3360 tcg_temp_free_vec(t);
3363 static void expand_vec_mul(TCGType type, unsigned vece,
3364 TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3366 TCGv_vec t1, t2, t3, t4, zero;
3368 tcg_debug_assert(vece == MO_8);
3371 * Unpack v1 bytes to words, 0 | x.
3372 * Unpack v2 bytes to words, y | 0.
3373 * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3374 * Shift logical right by 8 bits to clear the high 8 bytes before
3375 * using an unsigned saturated pack.
3377 * The difference between the V64, V128 and V256 cases is merely how
3378 * we distribute the expansion between temporaries.
3382 t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3383 t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3384 zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3385 vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3386 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3387 vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3388 tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3389 tcg_gen_mul_vec(MO_16, t1, t1, t2);
3390 tcg_gen_shri_vec(MO_16, t1, t1, 8);
3391 vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3392 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3393 tcg_temp_free_vec(t1);
3394 tcg_temp_free_vec(t2);
3399 t1 = tcg_temp_new_vec(type);
3400 t2 = tcg_temp_new_vec(type);
3401 t3 = tcg_temp_new_vec(type);
3402 t4 = tcg_temp_new_vec(type);
3403 zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3404 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3405 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3406 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3407 tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3408 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3409 tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3410 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3411 tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3412 tcg_gen_mul_vec(MO_16, t1, t1, t2);
3413 tcg_gen_mul_vec(MO_16, t3, t3, t4);
3414 tcg_gen_shri_vec(MO_16, t1, t1, 8);
3415 tcg_gen_shri_vec(MO_16, t3, t3, 8);
3416 vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3417 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3418 tcg_temp_free_vec(t1);
3419 tcg_temp_free_vec(t2);
3420 tcg_temp_free_vec(t3);
3421 tcg_temp_free_vec(t4);
3425 g_assert_not_reached();
3429 static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3430 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3439 TCGv_vec t1, t2, t3;
3455 fixup = NEED_SWAP | NEED_INV;
3458 if (vece <= MO_32) {
3461 fixup = NEED_BIAS | NEED_INV;
3465 if (vece <= MO_32) {
3466 fixup = NEED_UMIN | NEED_INV;
3472 if (vece <= MO_32) {
3475 fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3479 if (vece <= MO_32) {
3480 fixup = NEED_UMAX | NEED_INV;
3482 fixup = NEED_BIAS | NEED_SWAP;
3486 g_assert_not_reached();
3489 if (fixup & NEED_INV) {
3490 cond = tcg_invert_cond(cond);
3492 if (fixup & NEED_SWAP) {
3493 t1 = v1, v1 = v2, v2 = t1;
3494 cond = tcg_swap_cond(cond);
3498 if (fixup & (NEED_UMIN | NEED_UMAX)) {
3499 t1 = tcg_temp_new_vec(type);
3500 if (fixup & NEED_UMIN) {
3501 tcg_gen_umin_vec(vece, t1, v1, v2);
3503 tcg_gen_umax_vec(vece, t1, v1, v2);
3507 } else if (fixup & NEED_BIAS) {
3508 t1 = tcg_temp_new_vec(type);
3509 t2 = tcg_temp_new_vec(type);
3510 t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3511 tcg_gen_sub_vec(vece, t1, v1, t3);
3512 tcg_gen_sub_vec(vece, t2, v2, t3);
3515 cond = tcg_signed_cond(cond);
3518 tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3519 /* Expand directly; do not recurse. */
3520 vec_gen_4(INDEX_op_cmp_vec, type, vece,
3521 tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3524 tcg_temp_free_vec(t1);
3526 tcg_temp_free_vec(t2);
3529 return fixup & NEED_INV;
3532 static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
3533 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3535 if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
3536 tcg_gen_not_vec(vece, v0, v0);
3540 static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
3541 TCGv_vec c1, TCGv_vec c2,
3542 TCGv_vec v3, TCGv_vec v4, TCGCond cond)
3544 TCGv_vec t = tcg_temp_new_vec(type);
3546 if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
3547 /* Invert the sense of the compare by swapping arguments. */
3549 x = v3, v3 = v4, v4 = x;
3551 vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
3552 tcgv_vec_arg(v0), tcgv_vec_arg(v4),
3553 tcgv_vec_arg(v3), tcgv_vec_arg(t));
3554 tcg_temp_free_vec(t);
3557 void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3562 TCGv_vec v0, v1, v2, v3, v4;
3565 v0 = temp_tcgv_vec(arg_temp(a0));
3566 v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3567 a2 = va_arg(va, TCGArg);
3570 case INDEX_op_shli_vec:
3571 case INDEX_op_shri_vec:
3572 expand_vec_shi(type, vece, opc, v0, v1, a2);
3575 case INDEX_op_sari_vec:
3576 expand_vec_sari(type, vece, v0, v1, a2);
3579 case INDEX_op_rotli_vec:
3580 expand_vec_rotli(type, vece, v0, v1, a2);
3583 case INDEX_op_rotls_vec:
3584 expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
3587 case INDEX_op_rotlv_vec:
3588 v2 = temp_tcgv_vec(arg_temp(a2));
3589 expand_vec_rotv(type, vece, v0, v1, v2, false);
3591 case INDEX_op_rotrv_vec:
3592 v2 = temp_tcgv_vec(arg_temp(a2));
3593 expand_vec_rotv(type, vece, v0, v1, v2, true);
3596 case INDEX_op_mul_vec:
3597 v2 = temp_tcgv_vec(arg_temp(a2));
3598 expand_vec_mul(type, vece, v0, v1, v2);
3601 case INDEX_op_cmp_vec:
3602 v2 = temp_tcgv_vec(arg_temp(a2));
3603 expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
3606 case INDEX_op_cmpsel_vec:
3607 v2 = temp_tcgv_vec(arg_temp(a2));
3608 v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3609 v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3610 expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
3620 static const int tcg_target_callee_save_regs[] = {
3621 #if TCG_TARGET_REG_BITS == 64
3630 TCG_REG_R14, /* Currently used for the global env. */
3633 TCG_REG_EBP, /* Currently used for the global env. */
3640 /* Compute frame size via macros, to share between tcg_target_qemu_prologue
3641 and tcg_register_jit. */
3644 ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
3645 * (TCG_TARGET_REG_BITS / 8))
3647 #define FRAME_SIZE \
3649 + TCG_STATIC_CALL_ARGS_SIZE \
3650 + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3651 + TCG_TARGET_STACK_ALIGN - 1) \
3652 & ~(TCG_TARGET_STACK_ALIGN - 1))
3654 /* Generate global QEMU prologue and epilogue code */
3655 static void tcg_target_qemu_prologue(TCGContext *s)
3657 int i, stack_addend;
3661 /* Reserve some stack space, also for TCG temps. */
3662 stack_addend = FRAME_SIZE - PUSH_SIZE;
3663 tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3664 CPU_TEMP_BUF_NLONGS * sizeof(long));
3666 /* Save all callee saved registers. */
3667 for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
3668 tcg_out_push(s, tcg_target_callee_save_regs[i]);
3671 #if TCG_TARGET_REG_BITS == 32
3672 tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
3673 (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
3674 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
3676 tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
3677 (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
3680 # if !defined(CONFIG_SOFTMMU) && TCG_TARGET_REG_BITS == 64
3682 int seg = setup_guest_base_seg();
3684 x86_guest_base_seg = seg;
3685 } else if (guest_base == (int32_t)guest_base) {
3686 x86_guest_base_offset = guest_base;
3688 /* Choose R12 because, as a base, it requires a SIB byte. */
3689 x86_guest_base_index = TCG_REG_R12;
3690 tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base_index, guest_base);
3691 tcg_regset_set_reg(s->reserved_regs, x86_guest_base_index);
3695 tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3696 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
3698 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
3702 * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3703 * and fall through to the rest of the epilogue.
3705 tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3706 tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
3709 tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
3711 tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
3714 tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
3716 for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
3717 tcg_out_pop(s, tcg_target_callee_save_regs[i]);
3719 tcg_out_opc(s, OPC_RET, 0, 0, 0);
3722 static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3724 memset(p, 0x90, count);
3727 static void tcg_target_init(TCGContext *s)
3729 #ifdef CONFIG_CPUID_H
3730 unsigned a, b, c, d, b7 = 0;
3731 int max = __get_cpuid_max(0, 0);
3734 /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs. */
3735 __cpuid_count(7, 0, a, b7, c, d);
3736 have_bmi1 = (b7 & bit_BMI) != 0;
3737 have_bmi2 = (b7 & bit_BMI2) != 0;
3741 __cpuid(1, a, b, c, d);
3743 /* For 32-bit, 99% certainty that we're running on hardware that
3744 supports cmov, but we still need to check. In case cmov is not
3745 available, we'll use a small forward branch. */
3746 have_cmov = (d & bit_CMOV) != 0;
3749 /* MOVBE is only available on Intel Atom and Haswell CPUs, so we
3750 need to probe for it. */
3751 have_movbe = (c & bit_MOVBE) != 0;
3752 have_popcnt = (c & bit_POPCNT) != 0;
3754 /* There are a number of things we must check before we can be
3755 sure of not hitting invalid opcode. */
3756 if (c & bit_OSXSAVE) {
3757 unsigned xcrl, xcrh;
3758 /* The xgetbv instruction is not available to older versions of
3759 * the assembler, so we encode the instruction manually.
3761 asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcrl), "=d" (xcrh) : "c" (0));
3762 if ((xcrl & 6) == 6) {
3763 have_avx1 = (c & bit_AVX) != 0;
3764 have_avx2 = (b7 & bit_AVX2) != 0;
3769 max = __get_cpuid_max(0x8000000, 0);
3771 __cpuid(0x80000001, a, b, c, d);
3772 /* LZCNT was introduced with AMD Barcelona and Intel Haswell CPUs. */
3773 have_lzcnt = (c & bit_LZCNT) != 0;
3775 #endif /* CONFIG_CPUID_H */
3777 tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
3778 if (TCG_TARGET_REG_BITS == 64) {
3779 tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
3782 tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
3783 tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
3786 tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
3789 tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
3790 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
3791 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
3792 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
3793 if (TCG_TARGET_REG_BITS == 64) {
3794 #if !defined(_WIN64)
3795 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
3796 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
3798 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
3799 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
3800 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
3801 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
3804 s->reserved_regs = 0;
3805 tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
3810 uint8_t fde_def_cfa[4];
3811 uint8_t fde_reg_ofs[14];
3814 /* We're expecting a 2 byte uleb128 encoded value. */
3815 QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3817 #if !defined(__ELF__)
3818 /* Host machine without ELF. */
3819 #elif TCG_TARGET_REG_BITS == 64
3820 #define ELF_HOST_MACHINE EM_X86_64
3821 static const DebugFrame debug_frame = {
3822 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3825 .h.cie.code_align = 1,
3826 .h.cie.data_align = 0x78, /* sleb128 -8 */
3827 .h.cie.return_column = 16,
3829 /* Total FDE size does not include the "len" member. */
3830 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3833 12, 7, /* DW_CFA_def_cfa %rsp, ... */
3834 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
3838 0x90, 1, /* DW_CFA_offset, %rip, -8 */
3839 /* The following ordering must match tcg_target_callee_save_regs. */
3840 0x86, 2, /* DW_CFA_offset, %rbp, -16 */
3841 0x83, 3, /* DW_CFA_offset, %rbx, -24 */
3842 0x8c, 4, /* DW_CFA_offset, %r12, -32 */
3843 0x8d, 5, /* DW_CFA_offset, %r13, -40 */
3844 0x8e, 6, /* DW_CFA_offset, %r14, -48 */
3845 0x8f, 7, /* DW_CFA_offset, %r15, -56 */
3849 #define ELF_HOST_MACHINE EM_386
3850 static const DebugFrame debug_frame = {
3851 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3854 .h.cie.code_align = 1,
3855 .h.cie.data_align = 0x7c, /* sleb128 -4 */
3856 .h.cie.return_column = 8,
3858 /* Total FDE size does not include the "len" member. */
3859 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3862 12, 4, /* DW_CFA_def_cfa %esp, ... */
3863 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
3867 0x88, 1, /* DW_CFA_offset, %eip, -4 */
3868 /* The following ordering must match tcg_target_callee_save_regs. */
3869 0x85, 2, /* DW_CFA_offset, %ebp, -8 */
3870 0x83, 3, /* DW_CFA_offset, %ebx, -12 */
3871 0x86, 4, /* DW_CFA_offset, %esi, -16 */
3872 0x87, 5, /* DW_CFA_offset, %edi, -20 */
3877 #if defined(ELF_HOST_MACHINE)
3878 void tcg_register_jit(const void *buf, size_t buf_size)
3880 tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));