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-ldst.c.inc"
26 #include "../tcg-pool.c.inc"
28 #ifdef CONFIG_DEBUG_TCG
29 static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
30 #if TCG_TARGET_REG_BITS == 64
31 "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
33 "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
35 "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
36 "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
37 #if TCG_TARGET_REG_BITS == 64
38 "%xmm8", "%xmm9", "%xmm10", "%xmm11",
39 "%xmm12", "%xmm13", "%xmm14", "%xmm15",
44 static const int tcg_target_reg_alloc_order[] = {
45 #if TCG_TARGET_REG_BITS == 64
77 /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
78 any of them. Therefore only allow xmm0-xmm5 to be allocated. */
81 #if TCG_TARGET_REG_BITS == 64
94 #define TCG_TMP_VEC TCG_REG_XMM5
96 static const int tcg_target_call_iarg_regs[] = {
97 #if TCG_TARGET_REG_BITS == 64
110 /* 32 bit mode uses stack based calling convention (GCC default). */
114 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
117 case TCG_CALL_RET_NORMAL:
118 tcg_debug_assert(slot >= 0 && slot <= 1);
119 return slot ? TCG_REG_EDX : TCG_REG_EAX;
121 case TCG_CALL_RET_BY_VEC:
122 tcg_debug_assert(slot == 0);
126 g_assert_not_reached();
130 /* Constants we accept. */
131 #define TCG_CT_CONST_S32 0x100
132 #define TCG_CT_CONST_U32 0x200
133 #define TCG_CT_CONST_I32 0x400
134 #define TCG_CT_CONST_WSZ 0x800
135 #define TCG_CT_CONST_TST 0x1000
137 /* Registers used with L constraint, which are the first argument
138 registers on x86_64, and two random call clobbered registers on
140 #if TCG_TARGET_REG_BITS == 64
141 # define TCG_REG_L0 tcg_target_call_iarg_regs[0]
142 # define TCG_REG_L1 tcg_target_call_iarg_regs[1]
144 # define TCG_REG_L0 TCG_REG_EAX
145 # define TCG_REG_L1 TCG_REG_EDX
148 #if TCG_TARGET_REG_BITS == 64
149 # define ALL_GENERAL_REGS 0x0000ffffu
150 # define ALL_VECTOR_REGS 0xffff0000u
151 # define ALL_BYTEL_REGS ALL_GENERAL_REGS
153 # define ALL_GENERAL_REGS 0x000000ffu
154 # define ALL_VECTOR_REGS 0x00ff0000u
155 # define ALL_BYTEL_REGS 0x0000000fu
157 #define SOFTMMU_RESERVE_REGS \
158 (tcg_use_softmmu ? (1 << TCG_REG_L0) | (1 << TCG_REG_L1) : 0)
160 /* For 64-bit, we always know that CMOV is available. */
161 #if TCG_TARGET_REG_BITS == 64
162 # define have_cmov true
164 # define have_cmov (cpuinfo & CPUINFO_CMOV)
166 #define have_bmi2 (cpuinfo & CPUINFO_BMI2)
167 #define have_lzcnt (cpuinfo & CPUINFO_LZCNT)
169 static const tcg_insn_unit *tb_ret_addr;
171 static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
172 intptr_t value, intptr_t addend)
177 value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
178 if (value != (int32_t)value) {
183 tcg_patch32(code_ptr, value);
186 value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
187 if (value != (int8_t)value) {
190 tcg_patch8(code_ptr, value);
193 g_assert_not_reached();
198 /* test if a constant matches the constraint */
199 static bool tcg_target_const_match(int64_t val, int ct,
200 TCGType type, TCGCond cond, int vece)
202 if (ct & TCG_CT_CONST) {
205 if (type == TCG_TYPE_I32) {
206 if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 |
207 TCG_CT_CONST_I32 | TCG_CT_CONST_TST)) {
211 if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
214 if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
217 if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
221 * This will be used in combination with TCG_CT_CONST_S32,
222 * so "normal" TESTQ is already matched. Also accept:
223 * TESTQ -> TESTL (uint32_t)
224 * TESTQ -> BT (is_power_of_2)
226 if ((ct & TCG_CT_CONST_TST)
228 && (val == (uint32_t)val || is_power_of_2(val))) {
232 if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
238 # define LOWREGMASK(x) ((x) & 7)
240 #define P_EXT 0x100 /* 0x0f opcode prefix */
241 #define P_EXT38 0x200 /* 0x0f 0x38 opcode prefix */
242 #define P_DATA16 0x400 /* 0x66 opcode prefix */
243 #define P_VEXW 0x1000 /* Set VEX.W = 1 */
244 #if TCG_TARGET_REG_BITS == 64
245 # define P_REXW P_VEXW /* Set REX.W = 1; match VEXW */
246 # define P_REXB_R 0x2000 /* REG field as byte register */
247 # define P_REXB_RM 0x4000 /* R/M field as byte register */
248 # define P_GS 0x8000 /* gs segment override */
255 #define P_EXT3A 0x10000 /* 0x0f 0x3a opcode prefix */
256 #define P_SIMDF3 0x20000 /* 0xf3 opcode prefix */
257 #define P_SIMDF2 0x40000 /* 0xf2 opcode prefix */
258 #define P_VEXL 0x80000 /* Set VEX.L = 1 */
259 #define P_EVEX 0x100000 /* Requires EVEX encoding */
261 #define OPC_ARITH_EbIb (0x80)
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_MOVB_Ib (0xb0)
292 #define OPC_MOVL_Iv (0xb8)
293 #define OPC_MOVBE_GyMy (0xf0 | P_EXT38)
294 #define OPC_MOVBE_MyGy (0xf1 | P_EXT38)
295 #define OPC_MOVD_VyEy (0x6e | P_EXT | P_DATA16)
296 #define OPC_MOVD_EyVy (0x7e | P_EXT | P_DATA16)
297 #define OPC_MOVDDUP (0x12 | P_EXT | P_SIMDF2)
298 #define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
299 #define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
300 #define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
301 #define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
302 #define OPC_MOVQ_VqWq (0x7e | P_EXT | P_SIMDF3)
303 #define OPC_MOVQ_WqVq (0xd6 | P_EXT | P_DATA16)
304 #define OPC_MOVSBL (0xbe | P_EXT)
305 #define OPC_MOVSWL (0xbf | P_EXT)
306 #define OPC_MOVSLQ (0x63 | P_REXW)
307 #define OPC_MOVZBL (0xb6 | P_EXT)
308 #define OPC_MOVZWL (0xb7 | P_EXT)
309 #define OPC_PABSB (0x1c | P_EXT38 | P_DATA16)
310 #define OPC_PABSW (0x1d | P_EXT38 | P_DATA16)
311 #define OPC_PABSD (0x1e | P_EXT38 | P_DATA16)
312 #define OPC_VPABSQ (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
313 #define OPC_PACKSSDW (0x6b | P_EXT | P_DATA16)
314 #define OPC_PACKSSWB (0x63 | P_EXT | P_DATA16)
315 #define OPC_PACKUSDW (0x2b | P_EXT38 | P_DATA16)
316 #define OPC_PACKUSWB (0x67 | P_EXT | P_DATA16)
317 #define OPC_PADDB (0xfc | P_EXT | P_DATA16)
318 #define OPC_PADDW (0xfd | P_EXT | P_DATA16)
319 #define OPC_PADDD (0xfe | P_EXT | P_DATA16)
320 #define OPC_PADDQ (0xd4 | P_EXT | P_DATA16)
321 #define OPC_PADDSB (0xec | P_EXT | P_DATA16)
322 #define OPC_PADDSW (0xed | P_EXT | P_DATA16)
323 #define OPC_PADDUB (0xdc | P_EXT | P_DATA16)
324 #define OPC_PADDUW (0xdd | P_EXT | P_DATA16)
325 #define OPC_PAND (0xdb | P_EXT | P_DATA16)
326 #define OPC_PANDN (0xdf | P_EXT | P_DATA16)
327 #define OPC_PBLENDW (0x0e | P_EXT3A | P_DATA16)
328 #define OPC_PCMPEQB (0x74 | P_EXT | P_DATA16)
329 #define OPC_PCMPEQW (0x75 | P_EXT | P_DATA16)
330 #define OPC_PCMPEQD (0x76 | P_EXT | P_DATA16)
331 #define OPC_PCMPEQQ (0x29 | P_EXT38 | P_DATA16)
332 #define OPC_PCMPGTB (0x64 | P_EXT | P_DATA16)
333 #define OPC_PCMPGTW (0x65 | P_EXT | P_DATA16)
334 #define OPC_PCMPGTD (0x66 | P_EXT | P_DATA16)
335 #define OPC_PCMPGTQ (0x37 | P_EXT38 | P_DATA16)
336 #define OPC_PEXTRD (0x16 | P_EXT3A | P_DATA16)
337 #define OPC_PINSRD (0x22 | P_EXT3A | P_DATA16)
338 #define OPC_PMAXSB (0x3c | P_EXT38 | P_DATA16)
339 #define OPC_PMAXSW (0xee | P_EXT | P_DATA16)
340 #define OPC_PMAXSD (0x3d | P_EXT38 | P_DATA16)
341 #define OPC_VPMAXSQ (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
342 #define OPC_PMAXUB (0xde | P_EXT | P_DATA16)
343 #define OPC_PMAXUW (0x3e | P_EXT38 | P_DATA16)
344 #define OPC_PMAXUD (0x3f | P_EXT38 | P_DATA16)
345 #define OPC_VPMAXUQ (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
346 #define OPC_PMINSB (0x38 | P_EXT38 | P_DATA16)
347 #define OPC_PMINSW (0xea | P_EXT | P_DATA16)
348 #define OPC_PMINSD (0x39 | P_EXT38 | P_DATA16)
349 #define OPC_VPMINSQ (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
350 #define OPC_PMINUB (0xda | P_EXT | P_DATA16)
351 #define OPC_PMINUW (0x3a | P_EXT38 | P_DATA16)
352 #define OPC_PMINUD (0x3b | P_EXT38 | P_DATA16)
353 #define OPC_VPMINUQ (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
354 #define OPC_PMOVSXBW (0x20 | P_EXT38 | P_DATA16)
355 #define OPC_PMOVSXWD (0x23 | P_EXT38 | P_DATA16)
356 #define OPC_PMOVSXDQ (0x25 | P_EXT38 | P_DATA16)
357 #define OPC_PMOVZXBW (0x30 | P_EXT38 | P_DATA16)
358 #define OPC_PMOVZXWD (0x33 | P_EXT38 | P_DATA16)
359 #define OPC_PMOVZXDQ (0x35 | P_EXT38 | P_DATA16)
360 #define OPC_PMULLW (0xd5 | P_EXT | P_DATA16)
361 #define OPC_PMULLD (0x40 | P_EXT38 | P_DATA16)
362 #define OPC_VPMULLQ (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
363 #define OPC_POR (0xeb | P_EXT | P_DATA16)
364 #define OPC_PSHUFB (0x00 | P_EXT38 | P_DATA16)
365 #define OPC_PSHUFD (0x70 | P_EXT | P_DATA16)
366 #define OPC_PSHUFLW (0x70 | P_EXT | P_SIMDF2)
367 #define OPC_PSHUFHW (0x70 | P_EXT | P_SIMDF3)
368 #define OPC_PSHIFTW_Ib (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
369 #define OPC_PSHIFTD_Ib (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
370 #define OPC_PSHIFTQ_Ib (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
371 #define OPC_PSLLW (0xf1 | P_EXT | P_DATA16)
372 #define OPC_PSLLD (0xf2 | P_EXT | P_DATA16)
373 #define OPC_PSLLQ (0xf3 | P_EXT | P_DATA16)
374 #define OPC_PSRAW (0xe1 | P_EXT | P_DATA16)
375 #define OPC_PSRAD (0xe2 | P_EXT | P_DATA16)
376 #define OPC_VPSRAQ (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
377 #define OPC_PSRLW (0xd1 | P_EXT | P_DATA16)
378 #define OPC_PSRLD (0xd2 | P_EXT | P_DATA16)
379 #define OPC_PSRLQ (0xd3 | P_EXT | P_DATA16)
380 #define OPC_PSUBB (0xf8 | P_EXT | P_DATA16)
381 #define OPC_PSUBW (0xf9 | P_EXT | P_DATA16)
382 #define OPC_PSUBD (0xfa | P_EXT | P_DATA16)
383 #define OPC_PSUBQ (0xfb | P_EXT | P_DATA16)
384 #define OPC_PSUBSB (0xe8 | P_EXT | P_DATA16)
385 #define OPC_PSUBSW (0xe9 | P_EXT | P_DATA16)
386 #define OPC_PSUBUB (0xd8 | P_EXT | P_DATA16)
387 #define OPC_PSUBUW (0xd9 | P_EXT | P_DATA16)
388 #define OPC_PUNPCKLBW (0x60 | P_EXT | P_DATA16)
389 #define OPC_PUNPCKLWD (0x61 | P_EXT | P_DATA16)
390 #define OPC_PUNPCKLDQ (0x62 | P_EXT | P_DATA16)
391 #define OPC_PUNPCKLQDQ (0x6c | P_EXT | P_DATA16)
392 #define OPC_PUNPCKHBW (0x68 | P_EXT | P_DATA16)
393 #define OPC_PUNPCKHWD (0x69 | P_EXT | P_DATA16)
394 #define OPC_PUNPCKHDQ (0x6a | P_EXT | P_DATA16)
395 #define OPC_PUNPCKHQDQ (0x6d | P_EXT | P_DATA16)
396 #define OPC_PXOR (0xef | P_EXT | P_DATA16)
397 #define OPC_POP_r32 (0x58)
398 #define OPC_POPCNT (0xb8 | P_EXT | P_SIMDF3)
399 #define OPC_PUSH_r32 (0x50)
400 #define OPC_PUSH_Iv (0x68)
401 #define OPC_PUSH_Ib (0x6a)
402 #define OPC_RET (0xc3)
403 #define OPC_SETCC (0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
404 #define OPC_SHIFT_1 (0xd1)
405 #define OPC_SHIFT_Ib (0xc1)
406 #define OPC_SHIFT_cl (0xd3)
407 #define OPC_SARX (0xf7 | P_EXT38 | P_SIMDF3)
408 #define OPC_SHUFPS (0xc6 | P_EXT)
409 #define OPC_SHLX (0xf7 | P_EXT38 | P_DATA16)
410 #define OPC_SHRX (0xf7 | P_EXT38 | P_SIMDF2)
411 #define OPC_SHRD_Ib (0xac | P_EXT)
412 #define OPC_TESTB (0x84)
413 #define OPC_TESTL (0x85)
414 #define OPC_TZCNT (0xbc | P_EXT | P_SIMDF3)
415 #define OPC_UD2 (0x0b | P_EXT)
416 #define OPC_VPBLENDD (0x02 | P_EXT3A | P_DATA16)
417 #define OPC_VPBLENDVB (0x4c | P_EXT3A | P_DATA16)
418 #define OPC_VPINSRB (0x20 | P_EXT3A | P_DATA16)
419 #define OPC_VPINSRW (0xc4 | P_EXT | P_DATA16)
420 #define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
421 #define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
422 #define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
423 #define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
424 #define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
425 #define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
426 #define OPC_VPERMQ (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
427 #define OPC_VPERM2I128 (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
428 #define OPC_VPROLVD (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
429 #define OPC_VPROLVQ (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
430 #define OPC_VPRORVD (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
431 #define OPC_VPRORVQ (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
432 #define OPC_VPSHLDW (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
433 #define OPC_VPSHLDD (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
434 #define OPC_VPSHLDQ (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
435 #define OPC_VPSHLDVW (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
436 #define OPC_VPSHLDVD (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
437 #define OPC_VPSHLDVQ (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
438 #define OPC_VPSHRDVW (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
439 #define OPC_VPSHRDVD (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
440 #define OPC_VPSHRDVQ (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
441 #define OPC_VPSLLVW (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
442 #define OPC_VPSLLVD (0x47 | P_EXT38 | P_DATA16)
443 #define OPC_VPSLLVQ (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
444 #define OPC_VPSRAVW (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
445 #define OPC_VPSRAVD (0x46 | P_EXT38 | P_DATA16)
446 #define OPC_VPSRAVQ (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
447 #define OPC_VPSRLVW (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
448 #define OPC_VPSRLVD (0x45 | P_EXT38 | P_DATA16)
449 #define OPC_VPSRLVQ (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
450 #define OPC_VPTERNLOGQ (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
451 #define OPC_VZEROUPPER (0x77 | P_EXT)
452 #define OPC_XCHG_ax_r32 (0x90)
453 #define OPC_XCHG_EvGv (0x87)
455 #define OPC_GRP3_Eb (0xf6)
456 #define OPC_GRP3_Ev (0xf7)
457 #define OPC_GRP5 (0xff)
458 #define OPC_GRP14 (0x73 | P_EXT | P_DATA16)
459 #define OPC_GRPBT (0xba | P_EXT)
461 #define OPC_GRPBT_BT 4
462 #define OPC_GRPBT_BTS 5
463 #define OPC_GRPBT_BTR 6
464 #define OPC_GRPBT_BTC 7
466 /* Group 1 opcode extensions for 0x80-0x83.
467 These are also used as modifiers for OPC_ARITH. */
477 /* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3. */
484 /* Group 3 opcode extensions for 0xf6, 0xf7. To be used with OPC_GRP3. */
493 /* Group 5 opcode extensions for 0xff. To be used with OPC_GRP5. */
494 #define EXT5_INC_Ev 0
495 #define EXT5_DEC_Ev 1
496 #define EXT5_CALLN_Ev 2
497 #define EXT5_JMPN_Ev 4
499 /* Condition codes to be added to OPC_JCC_{long,short}. */
518 static const uint8_t tcg_cond_to_jcc[] = {
519 [TCG_COND_EQ] = JCC_JE,
520 [TCG_COND_NE] = JCC_JNE,
521 [TCG_COND_LT] = JCC_JL,
522 [TCG_COND_GE] = JCC_JGE,
523 [TCG_COND_LE] = JCC_JLE,
524 [TCG_COND_GT] = JCC_JG,
525 [TCG_COND_LTU] = JCC_JB,
526 [TCG_COND_GEU] = JCC_JAE,
527 [TCG_COND_LEU] = JCC_JBE,
528 [TCG_COND_GTU] = JCC_JA,
529 [TCG_COND_TSTEQ] = JCC_JE,
530 [TCG_COND_TSTNE] = JCC_JNE,
533 #if TCG_TARGET_REG_BITS == 64
534 static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
541 if (opc & P_DATA16) {
542 /* We should never be asking for both 16 and 64-bit operation. */
543 tcg_debug_assert((opc & P_REXW) == 0);
546 if (opc & P_SIMDF3) {
548 } else if (opc & P_SIMDF2) {
553 rex |= (opc & P_REXW) ? 0x8 : 0x0; /* REX.W */
554 rex |= (r & 8) >> 1; /* REX.R */
555 rex |= (x & 8) >> 2; /* REX.X */
556 rex |= (rm & 8) >> 3; /* REX.B */
558 /* P_REXB_{R,RM} indicates that the given register is the low byte.
559 For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
560 as otherwise the encoding indicates %[abcd]h. Note that the values
561 that are ORed in merely indicate that the REX byte must be present;
562 those bits get discarded in output. */
563 rex |= opc & (r >= 4 ? P_REXB_R : 0);
564 rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
567 tcg_out8(s, (uint8_t)(rex | 0x40));
570 if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
574 } else if (opc & P_EXT3A) {
582 static void tcg_out_opc(TCGContext *s, int opc)
584 if (opc & P_DATA16) {
587 if (opc & P_SIMDF3) {
589 } else if (opc & P_SIMDF2) {
592 if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
596 } else if (opc & P_EXT3A) {
602 /* Discard the register arguments to tcg_out_opc early, so as not to penalize
603 the 32-bit compilation paths. This method works with all versions of gcc,
604 whereas relying on optimization may not be able to exclude them. */
605 #define tcg_out_opc(s, opc, r, rm, x) (tcg_out_opc)(s, opc)
608 static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
610 tcg_out_opc(s, opc, r, rm, 0);
611 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
614 static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
622 /* Use the two byte form if possible, which cannot encode
623 VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT. */
624 if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
625 && ((rm | index) & 8) == 0) {
626 /* Two byte VEX prefix. */
629 tmp = (r & 8 ? 0 : 0x80); /* VEX.R */
631 /* Three byte VEX prefix. */
637 } else if (opc & P_EXT38) {
639 } else if (opc & P_EXT) {
642 g_assert_not_reached();
644 tmp |= (r & 8 ? 0 : 0x80); /* VEX.R */
645 tmp |= (index & 8 ? 0 : 0x40); /* VEX.X */
646 tmp |= (rm & 8 ? 0 : 0x20); /* VEX.B */
649 tmp = (opc & P_VEXW ? 0x80 : 0); /* VEX.W */
652 tmp |= (opc & P_VEXL ? 0x04 : 0); /* VEX.L */
654 if (opc & P_DATA16) {
656 } else if (opc & P_SIMDF3) {
658 } else if (opc & P_SIMDF2) {
661 tmp |= (~v & 15) << 3; /* VEX.vvvv */
666 static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
669 /* The entire 4-byte evex prefix; with R' and V' set. */
670 uint32_t p = 0x08041062;
673 tcg_debug_assert(have_avx512vl);
678 } else if (opc & P_EXT38) {
680 } else if (opc & P_EXT) {
683 g_assert_not_reached();
687 if (opc & P_DATA16) {
689 } else if (opc & P_SIMDF3) {
691 } else if (opc & P_SIMDF2) {
697 p = deposit32(p, 8, 2, mm);
698 p = deposit32(p, 13, 1, (rm & 8) == 0); /* EVEX.RXB.B */
699 p = deposit32(p, 14, 1, (index & 8) == 0); /* EVEX.RXB.X */
700 p = deposit32(p, 15, 1, (r & 8) == 0); /* EVEX.RXB.R */
701 p = deposit32(p, 16, 2, pp);
702 p = deposit32(p, 19, 4, ~v);
703 p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
704 p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
710 static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
713 tcg_out_evex_opc(s, opc, r, v, rm, 0);
715 tcg_out_vex_opc(s, opc, r, v, rm, 0);
717 tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
720 /* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
721 We handle either RM and INDEX missing with a negative value. In 64-bit
722 mode for absolute addresses, ~RM is the size of the immediate operand
723 that will follow the instruction. */
725 static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
726 int shift, intptr_t offset)
730 if (index < 0 && rm < 0) {
731 if (TCG_TARGET_REG_BITS == 64) {
732 /* Try for a rip-relative addressing mode. This has replaced
733 the 32-bit-mode absolute addressing encoding. */
734 intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
735 intptr_t disp = offset - pc;
736 if (disp == (int32_t)disp) {
737 tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
742 /* Try for an absolute address encoding. This requires the
743 use of the MODRM+SIB encoding and is therefore larger than
744 rip-relative addressing. */
745 if (offset == (int32_t)offset) {
746 tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
747 tcg_out8(s, (4 << 3) | 5);
748 tcg_out32(s, offset);
752 /* ??? The memory isn't directly addressable. */
753 g_assert_not_reached();
755 /* Absolute address. */
756 tcg_out8(s, (r << 3) | 5);
757 tcg_out32(s, offset);
762 /* Find the length of the immediate addend. Note that the encoding
763 that would be used for (%ebp) indicates absolute addressing. */
765 mod = 0, len = 4, rm = 5;
766 } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
768 } else if (offset == (int8_t)offset) {
774 /* Use a single byte MODRM format if possible. Note that the encoding
775 that would be used for %esp is the escape to the two byte form. */
776 if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
777 /* Single byte MODRM format. */
778 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
780 /* Two byte MODRM+SIB format. */
782 /* Note that the encoding that would place %esp into the index
783 field indicates no index register. In 64-bit mode, the REX.X
784 bit counts, so %r12 can be used as the index. */
788 tcg_debug_assert(index != TCG_REG_ESP);
791 tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
792 tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
797 } else if (len == 4) {
798 tcg_out32(s, offset);
802 static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
803 int index, int shift, intptr_t offset)
805 tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
806 tcg_out_sib_offset(s, r, rm, index, shift, offset);
809 static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
810 int rm, int index, int shift,
813 tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
814 tcg_out_sib_offset(s, r, rm, index, shift, offset);
817 /* A simplification of the above with no index or shift. */
818 static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
819 int rm, intptr_t offset)
821 tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
824 static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
825 int v, int rm, intptr_t offset)
827 tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
830 /* Output an opcode with an expected reference to the constant pool. */
831 static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
833 tcg_out_opc(s, opc, r, 0, 0);
834 /* Absolute for 32-bit, pc-relative for 64-bit. */
835 tcg_out8(s, LOWREGMASK(r) << 3 | 5);
839 /* Output an opcode with an expected reference to the constant pool. */
840 static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
842 tcg_out_vex_opc(s, opc, r, 0, 0, 0);
843 /* Absolute for 32-bit, pc-relative for 64-bit. */
844 tcg_out8(s, LOWREGMASK(r) << 3 | 5);
848 /* Generate dest op= src. Uses the same ARITH_* codes as tgen_arithi. */
849 static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
851 /* Propagate an opcode prefix, such as P_REXW. */
852 int ext = subop & ~0x7;
855 tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
858 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
872 tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
874 tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
878 tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
880 tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
886 tcg_debug_assert(ret >= 16 && arg >= 16);
887 tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
890 tcg_debug_assert(ret >= 16 && arg >= 16);
891 tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
894 tcg_debug_assert(ret >= 16 && arg >= 16);
895 tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
899 g_assert_not_reached();
904 static const int avx2_dup_insn[4] = {
905 OPC_VPBROADCASTB, OPC_VPBROADCASTW,
906 OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
909 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
913 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
914 tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
918 /* ??? With zero in a register, use PSHUFB. */
919 tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
923 tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
927 tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
928 /* imm8 operand: all output lanes selected from input lane 0. */
932 tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
935 g_assert_not_reached();
941 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
942 TCGReg r, TCGReg base, intptr_t offset)
945 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
946 tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
951 tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
954 tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
957 tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
958 tcg_out8(s, 0); /* imm8 */
959 tcg_out_dup_vec(s, type, vece, r, r);
962 tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
963 tcg_out8(s, 0); /* imm8 */
964 tcg_out_dup_vec(s, type, vece, r, r);
967 g_assert_not_reached();
973 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
974 TCGReg ret, int64_t arg)
976 int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
979 tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
983 tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
987 if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
989 tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
991 tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
993 new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
995 if (type == TCG_TYPE_V64) {
996 tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
997 } else if (have_avx2) {
998 tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1000 tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1002 if (TCG_TARGET_REG_BITS == 64) {
1003 new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1005 new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1010 static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1011 TCGReg ret, tcg_target_long arg)
1014 tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1018 tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1022 int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1023 tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1024 if (TCG_TARGET_REG_BITS == 64) {
1025 new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1027 new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1031 static void tcg_out_movi_int(TCGContext *s, TCGType type,
1032 TCGReg ret, tcg_target_long arg)
1034 tcg_target_long diff;
1037 tgen_arithr(s, ARITH_XOR, ret, ret);
1040 if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1041 tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1045 if (arg == (int32_t)arg) {
1046 tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1051 /* Try a 7 byte pc-relative lea before the 10 byte movq. */
1052 diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1053 if (diff == (int32_t)diff) {
1054 tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1055 tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1060 tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1064 static void tcg_out_movi(TCGContext *s, TCGType type,
1065 TCGReg ret, tcg_target_long arg)
1069 #if TCG_TARGET_REG_BITS == 64
1073 tcg_out_movi_int(s, type, ret, arg);
1075 tcg_out_movi_vec(s, type, ret, arg);
1079 g_assert_not_reached();
1083 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1085 int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1086 tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1090 static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1091 tcg_target_long imm)
1093 /* This function is only used for passing structs by reference. */
1094 tcg_debug_assert(imm == (int32_t)imm);
1095 tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1098 static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1100 if (val == (int8_t)val) {
1101 tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1103 } else if (val == (int32_t)val) {
1104 tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1107 g_assert_not_reached();
1111 static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1113 /* Given the strength of x86 memory ordering, we only need care for
1114 store-load ordering. Experimentally, "lock orl $0,0(%esp)" is
1115 faster than "mfence", so don't bother with the sse insn. */
1116 if (a0 & TCG_MO_ST_LD) {
1118 tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1123 static inline void tcg_out_push(TCGContext *s, int reg)
1125 tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1128 static inline void tcg_out_pop(TCGContext *s, int reg)
1130 tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1133 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1134 TCGReg arg1, intptr_t arg2)
1139 tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1141 tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1146 tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1151 /* There is no instruction that can validate 8-byte alignment. */
1152 tcg_debug_assert(ret >= 16);
1153 tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1157 * The gvec infrastructure is asserts that v128 vector loads
1158 * and stores use a 16-byte aligned offset. Validate that the
1159 * final pointer is aligned by using an insn that will SIGSEGV.
1161 tcg_debug_assert(ret >= 16);
1162 tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1166 * The gvec infrastructure only requires 16-byte alignment,
1167 * so here we must use an unaligned load.
1169 tcg_debug_assert(ret >= 16);
1170 tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1171 ret, 0, arg1, arg2);
1174 g_assert_not_reached();
1178 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1179 TCGReg arg1, intptr_t arg2)
1184 tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1186 tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1191 tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1196 /* There is no instruction that can validate 8-byte alignment. */
1197 tcg_debug_assert(arg >= 16);
1198 tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1202 * The gvec infrastructure is asserts that v128 vector loads
1203 * and stores use a 16-byte aligned offset. Validate that the
1204 * final pointer is aligned by using an insn that will SIGSEGV.
1206 * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1207 * for _WIN64, which must have SSE2 but may not have AVX.
1209 tcg_debug_assert(arg >= 16);
1211 tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1213 tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1218 * The gvec infrastructure only requires 16-byte alignment,
1219 * so here we must use an unaligned store.
1221 tcg_debug_assert(arg >= 16);
1222 tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1223 arg, 0, arg1, arg2);
1226 g_assert_not_reached();
1230 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1231 TCGReg base, intptr_t ofs)
1234 if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1235 if (val != (int32_t)val) {
1239 } else if (type != TCG_TYPE_I32) {
1242 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1247 static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1249 /* Propagate an opcode prefix, such as P_DATA16. */
1250 int ext = subopc & ~0x7;
1254 tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1256 tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1261 static inline void tcg_out_bswap32(TCGContext *s, int reg)
1263 tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1266 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1268 tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1271 static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1274 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1275 tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1278 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1280 int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1282 tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1283 tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1286 static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1289 tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1292 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1294 int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1296 tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1299 static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1301 /* 32-bit mov zero extends. */
1302 tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1305 static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1307 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1308 tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1311 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1313 tcg_out_ext32s(s, dest, src);
1316 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1319 tcg_out_ext32u(s, dest, src);
1323 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1325 tcg_out_ext32u(s, dest, src);
1328 static inline void tcg_out_bswap64(TCGContext *s, int reg)
1330 tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1333 static void tgen_arithi(TCGContext *s, int c, int r0,
1334 tcg_target_long val, int cf)
1338 if (TCG_TARGET_REG_BITS == 64) {
1348 * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1349 * partial flags update stalls on Pentium4 and are not recommended
1350 * by current Intel optimization manuals.
1352 if (val == 1 || val == -1) {
1353 int is_inc = (c == ARITH_ADD) ^ (val < 0);
1354 if (TCG_TARGET_REG_BITS == 64) {
1356 * The single-byte increment encodings are re-tasked
1357 * as the REX prefixes. Use the MODRM encoding.
1359 tcg_out_modrm(s, OPC_GRP5 + rexw,
1360 (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1362 tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1368 * Facilitate using an 8-bit immediate. Carry is inverted
1369 * by this transformation, so do it only if cf == 0.
1371 c ^= ARITH_ADD ^ ARITH_SUB;
1378 if (TCG_TARGET_REG_BITS == 64) {
1379 if (val == 0xffffffffu) {
1380 tcg_out_ext32u(s, r0, r0);
1383 if (val == (uint32_t)val) {
1384 /* AND with no high bits set can use a 32-bit operation. */
1388 if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1389 tcg_out_ext8u(s, r0, r0);
1392 if (val == 0xffffu) {
1393 tcg_out_ext16u(s, r0, r0);
1400 if (val >= 0x80 && val <= 0xff
1401 && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1402 tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1409 if (val == (int8_t)val) {
1410 tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1414 if (rexw == 0 || val == (int32_t)val) {
1415 tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1420 g_assert_not_reached();
1423 static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1426 tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1430 /* Set SMALL to force a short forward branch. */
1431 static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1436 val = tcg_pcrel_diff(s, l->u.value_ptr);
1438 if ((int8_t)val1 == val1) {
1440 tcg_out8(s, OPC_JMP_short);
1442 tcg_out8(s, OPC_JCC_short + opc);
1446 tcg_debug_assert(!small);
1448 tcg_out8(s, OPC_JMP_long);
1449 tcg_out32(s, val - 5);
1451 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1452 tcg_out32(s, val - 6);
1457 tcg_out8(s, OPC_JMP_short);
1459 tcg_out8(s, OPC_JCC_short + opc);
1461 tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1465 tcg_out8(s, OPC_JMP_long);
1467 tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1469 tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1474 static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1475 TCGArg arg2, int const_arg2, int rexw)
1479 if (!is_tst_cond(cond)) {
1481 tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1482 } else if (arg2 == 0) {
1483 tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1485 tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1486 tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1488 return tcg_cond_to_jcc[cond];
1491 jz = tcg_cond_to_jcc[cond];
1492 js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1495 tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1499 if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1501 tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1505 tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1508 tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1513 if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1514 if (arg2 == 0x8000) {
1515 tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1518 if (arg2 == 0xff00) {
1519 tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1522 tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1523 tcg_out8(s, arg2 >> 8);
1527 if (arg2 == 0xffff) {
1528 tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1531 if (arg2 == 0xffffffffu) {
1532 tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1536 if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1537 int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1538 int sh = ctz64(arg2);
1540 rexw = (sh & 32 ? P_REXW : 0);
1541 if ((sh & 31) == 31) {
1542 tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1545 tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1552 if (arg2 == (uint32_t)arg2) {
1555 tcg_debug_assert(arg2 == (int32_t)arg2);
1558 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1563 static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1564 TCGArg arg1, TCGArg arg2, int const_arg2,
1565 TCGLabel *label, bool small)
1567 int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1568 tcg_out_jxx(s, jcc, label, small);
1571 #if TCG_TARGET_REG_BITS == 32
1572 static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1573 const int *const_args, bool small)
1575 TCGLabel *label_next = gen_new_label();
1576 TCGLabel *label_this = arg_label(args[5]);
1577 TCGCond cond = args[4];
1581 case TCG_COND_TSTEQ:
1582 tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1583 args[0], args[2], const_args[2], label_next, 1);
1584 tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1588 case TCG_COND_TSTNE:
1589 tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1591 tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1595 tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1597 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1598 tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1602 tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1604 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1605 tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1609 tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1611 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1612 tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1616 tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1618 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1619 tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1623 tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1625 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1626 tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1630 tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1632 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1633 tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1637 tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1639 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1640 tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1644 tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1646 tcg_out_jxx(s, JCC_JNE, label_next, 1);
1647 tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1651 g_assert_not_reached();
1653 tcg_out_label(s, label_next);
1657 static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1658 TCGArg dest, TCGArg arg1, TCGArg arg2,
1659 int const_arg2, bool neg)
1670 /* If arg2 is 0, convert to LTU/GEU vs 1. */
1671 if (const_arg2 && arg2 == 0) {
1681 /* If arg2 is a register, swap for LTU/GEU. */
1696 * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1697 * We can then use NEG or INC to produce the desired result.
1698 * This is always smaller than the SETCC expansion.
1700 tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, rexw);
1702 /* X - X - C = -C = (C ? -1 : 0) */
1703 tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1705 /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1706 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1708 /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1709 tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1711 /* -(C ? -1 : 0) = (C ? 1 : 0) */
1712 tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1720 /* If arg2 is 0, extract the sign bit. */
1721 if (const_arg2 && arg2 == 0) {
1722 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1724 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1726 tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1727 dest, rexw ? 63 : 31);
1737 * If dest does not overlap the inputs, clearing it first is preferred.
1738 * The XOR breaks any false dependency for the low-byte write to dest,
1739 * and is also one byte smaller than MOVZBL.
1742 if (dest != arg1 && (const_arg2 || dest != arg2)) {
1743 tgen_arithr(s, ARITH_XOR, dest, dest);
1747 jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1748 tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1751 tcg_out_ext8u(s, dest, dest);
1754 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1758 #if TCG_TARGET_REG_BITS == 32
1759 static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1760 const int *const_args)
1763 TCGLabel *label_true, *label_over;
1765 memcpy(new_args, args+1, 5*sizeof(TCGArg));
1767 if (args[0] == args[1] || args[0] == args[2]
1768 || (!const_args[3] && args[0] == args[3])
1769 || (!const_args[4] && args[0] == args[4])) {
1770 /* When the destination overlaps with one of the argument
1771 registers, don't do anything tricky. */
1772 label_true = gen_new_label();
1773 label_over = gen_new_label();
1775 new_args[5] = label_arg(label_true);
1776 tcg_out_brcond2(s, new_args, const_args+1, 1);
1778 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1779 tcg_out_jxx(s, JCC_JMP, label_over, 1);
1780 tcg_out_label(s, label_true);
1782 tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1783 tcg_out_label(s, label_over);
1785 /* When the destination does not overlap one of the arguments,
1786 clear the destination first, jump if cond false, and emit an
1787 increment in the true case. This results in smaller code. */
1789 tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1791 label_over = gen_new_label();
1792 new_args[4] = tcg_invert_cond(new_args[4]);
1793 new_args[5] = label_arg(label_over);
1794 tcg_out_brcond2(s, new_args, const_args+1, 1);
1796 tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1797 tcg_out_label(s, label_over);
1802 static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1803 TCGReg dest, TCGReg v1)
1806 tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1808 TCGLabel *over = gen_new_label();
1809 tcg_out_jxx(s, jcc ^ 1, over, 1);
1810 tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1811 tcg_out_label(s, over);
1815 static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1816 TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1819 int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1820 tcg_out_cmov(s, jcc, rexw, dest, v1);
1823 static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1824 TCGArg arg2, bool const_a2)
1827 tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1829 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1831 tcg_debug_assert(dest != arg2);
1832 tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1835 tcg_debug_assert(dest != arg2);
1836 tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1837 tcg_out_cmov(s, JCC_JE, rexw, dest, arg2);
1841 static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1842 TCGArg arg2, bool const_a2)
1845 tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1847 tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1849 tcg_debug_assert(dest != arg2);
1850 tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1853 tcg_debug_assert(!const_a2);
1854 tcg_debug_assert(dest != arg1);
1855 tcg_debug_assert(dest != arg2);
1857 /* Recall that the output of BSR is the index not the count. */
1858 tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1859 tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1861 /* Since we have destroyed the flags from BSR, we have to re-test. */
1862 int jcc = tcg_out_cmp(s, TCG_COND_EQ, arg1, 0, 1, rexw);
1863 tcg_out_cmov(s, jcc, rexw, dest, arg2);
1867 static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1869 intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1871 if (disp == (int32_t)disp) {
1872 tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1875 /* rip-relative addressing into the constant pool.
1876 This is 6 + 8 = 14 bytes, as compared to using an
1877 immediate load 10 + 6 = 16 bytes, plus we may
1878 be able to re-use the pool constant for more calls. */
1879 tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1880 tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1881 new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1886 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1887 const TCGHelperInfo *info)
1889 tcg_out_branch(s, 1, dest);
1892 if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1894 * The sysv i386 abi for struct return places a reference as the
1895 * first argument of the stack, and pops that argument with the
1896 * return statement. Since we want to retain the aligned stack
1897 * pointer for the callee, we do not want to actually push that
1898 * argument before the call but rely on the normal store to the
1899 * stack slot. But we do need to compensate for the pop in order
1900 * to reset our correct stack pointer value.
1901 * Pushing a garbage value back onto the stack is quickest.
1903 tcg_out_push(s, TCG_REG_EAX);
1908 static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1910 tcg_out_branch(s, 0, dest);
1913 static void tcg_out_nopn(TCGContext *s, int n)
1916 /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1917 * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1918 * duplicate prefix, and all of the interesting recent cores can
1919 * decode and discard the duplicates in a single cycle.
1921 tcg_debug_assert(n >= 1);
1922 for (i = 1; i < n; ++i) {
1936 bool tcg_target_has_memory_bswap(MemOp memop)
1943 if ((memop & MO_SIZE) < MO_128) {
1948 * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1949 * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1951 aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1952 return aa.atom < MO_128;
1956 * Because i686 has no register parameters and because x86_64 has xchg
1957 * to handle addr/data register overlap, we have placed all input arguments
1958 * before we need might need a scratch reg.
1960 * Even then, a scratch is only needed for l->raddr. Rather than expose
1961 * a general-purpose scratch when we don't actually know it's available,
1962 * use the ra_gen hook to load into RAX if needed.
1964 #if TCG_TARGET_REG_BITS == 64
1965 static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1970 tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
1973 static const TCGLdstHelperParam ldst_helper_param = {
1974 .ra_gen = ldst_ra_gen
1977 static const TCGLdstHelperParam ldst_helper_param = { };
1980 static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
1981 TCGReg l, TCGReg h, TCGReg v)
1983 int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1985 /* vpmov{d,q} %v, %l */
1986 tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
1987 /* vpextr{d,q} $1, %v, %h */
1988 tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
1992 static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
1993 TCGReg v, TCGReg l, TCGReg h)
1995 int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1997 /* vmov{d,q} %l, %v */
1998 tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
1999 /* vpinsr{d,q} $1, %h, %v, %v */
2000 tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2005 * Generate code for the slow path for a load at the end of block
2007 static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2009 MemOp opc = get_memop(l->oi);
2010 tcg_insn_unit **label_ptr = &l->label_ptr[0];
2012 /* resolve label address */
2013 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2015 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2018 tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2019 tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2020 tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2022 tcg_out_jmp(s, l->raddr);
2027 * Generate code for the slow path for a store at the end of block
2029 static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2031 MemOp opc = get_memop(l->oi);
2032 tcg_insn_unit **label_ptr = &l->label_ptr[0];
2034 /* resolve label address */
2035 tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2037 tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2040 tcg_out_st_helper_args(s, l, &ldst_helper_param);
2041 tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2043 tcg_out_jmp(s, l->raddr);
2047 #ifdef CONFIG_USER_ONLY
2048 static HostAddress x86_guest_base = {
2052 #if defined(__x86_64__) && defined(__linux__)
2053 # include <asm/prctl.h>
2054 # include <sys/prctl.h>
2055 int arch_prctl(int code, unsigned long addr);
2056 static inline int setup_guest_base_seg(void)
2058 if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2063 #define setup_guest_base_seg setup_guest_base_seg
2064 #elif defined(__x86_64__) && \
2065 (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2066 # include <machine/sysarch.h>
2067 static inline int setup_guest_base_seg(void)
2069 if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2074 #define setup_guest_base_seg setup_guest_base_seg
2077 # define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2078 #endif /* CONFIG_USER_ONLY */
2079 #ifndef setup_guest_base_seg
2080 # define setup_guest_base_seg() 0
2083 #define MIN_TLB_MASK_TABLE_OFS INT_MIN
2086 * For softmmu, perform the TLB load and compare.
2087 * For useronly, perform any required alignment tests.
2088 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2089 * is required and fill in @h with the host address for the fast path.
2091 static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2092 TCGReg addrlo, TCGReg addrhi,
2093 MemOpIdx oi, bool is_ld)
2095 TCGLabelQemuLdst *ldst = NULL;
2096 MemOp opc = get_memop(oi);
2097 MemOp s_bits = opc & MO_SIZE;
2100 if (tcg_use_softmmu) {
2101 h->index = TCG_REG_L0;
2105 *h = x86_guest_base;
2108 h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2109 a_mask = (1 << h->aa.align) - 1;
2111 if (tcg_use_softmmu) {
2112 int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2113 : offsetof(CPUTLBEntry, addr_write);
2114 TCGType ttype = TCG_TYPE_I32;
2115 TCGType tlbtype = TCG_TYPE_I32;
2116 int trexw = 0, hrexw = 0, tlbrexw = 0;
2117 unsigned mem_index = get_mmuidx(oi);
2118 unsigned s_mask = (1 << s_bits) - 1;
2119 int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2122 ldst = new_ldst_label(s);
2123 ldst->is_ld = is_ld;
2125 ldst->addrlo_reg = addrlo;
2126 ldst->addrhi_reg = addrhi;
2128 if (TCG_TARGET_REG_BITS == 64) {
2129 ttype = s->addr_type;
2130 trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2131 if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2133 if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2134 tlbtype = TCG_TYPE_I64;
2140 tcg_out_mov(s, tlbtype, TCG_REG_L0, addrlo);
2141 tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2142 s->page_bits - CPU_TLB_ENTRY_BITS);
2144 tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2145 fast_ofs + offsetof(CPUTLBDescFast, mask));
2147 tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2148 fast_ofs + offsetof(CPUTLBDescFast, table));
2151 * If the required alignment is at least as large as the access,
2152 * simply copy the address and mask. For lesser alignments,
2153 * check that we don't cross pages for the complete access.
2155 if (a_mask >= s_mask) {
2156 tcg_out_mov(s, ttype, TCG_REG_L1, addrlo);
2158 tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2159 addrlo, s_mask - a_mask);
2161 tlb_mask = s->page_mask | a_mask;
2162 tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2164 /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2165 tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2166 TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2169 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2170 ldst->label_ptr[0] = s->code_ptr;
2173 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I64) {
2174 /* cmp 4(TCG_REG_L0), addrhi */
2175 tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi,
2176 TCG_REG_L0, cmp_ofs + 4);
2179 tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2180 ldst->label_ptr[1] = s->code_ptr;
2185 tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2186 offsetof(CPUTLBEntry, addend));
2187 } else if (a_mask) {
2190 ldst = new_ldst_label(s);
2191 ldst->is_ld = is_ld;
2193 ldst->addrlo_reg = addrlo;
2194 ldst->addrhi_reg = addrhi;
2197 jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addrlo, a_mask, true, false);
2198 tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2199 ldst->label_ptr[0] = s->code_ptr;
2206 static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2207 HostAddress h, TCGType type, MemOp memop)
2209 bool use_movbe = false;
2210 int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2211 int movop = OPC_MOVL_GvEv;
2213 /* Do big-endian loads with movbe. */
2214 if (memop & MO_BSWAP) {
2215 tcg_debug_assert(have_movbe);
2217 movop = OPC_MOVBE_GyMy;
2220 switch (memop & MO_SSIZE) {
2222 tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2223 h.base, h.index, 0, h.ofs);
2226 tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2227 h.base, h.index, 0, h.ofs);
2231 /* There is no extending movbe; only low 16-bits are modified. */
2232 if (datalo != h.base && datalo != h.index) {
2233 /* XOR breaks dependency chains. */
2234 tgen_arithr(s, ARITH_XOR, datalo, datalo);
2235 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2236 datalo, h.base, h.index, 0, h.ofs);
2238 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2239 datalo, h.base, h.index, 0, h.ofs);
2240 tcg_out_ext16u(s, datalo, datalo);
2243 tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2244 h.base, h.index, 0, h.ofs);
2249 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2250 datalo, h.base, h.index, 0, h.ofs);
2251 tcg_out_ext16s(s, type, datalo, datalo);
2253 tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2254 datalo, h.base, h.index, 0, h.ofs);
2258 tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2259 h.base, h.index, 0, h.ofs);
2261 #if TCG_TARGET_REG_BITS == 64
2264 tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2265 h.base, h.index, 0, h.ofs);
2266 tcg_out_ext32s(s, datalo, datalo);
2268 tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2269 h.base, h.index, 0, h.ofs);
2274 if (TCG_TARGET_REG_BITS == 64) {
2275 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2276 h.base, h.index, 0, h.ofs);
2284 if (h.base == datalo || h.index == datalo) {
2285 tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2286 h.base, h.index, 0, h.ofs);
2287 tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2288 tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2290 tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2291 h.base, h.index, 0, h.ofs);
2292 tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2293 h.base, h.index, 0, h.ofs + 4);
2298 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2301 * Without 16-byte atomicity, use integer regs.
2302 * That is where we want the data, and it allows bswaps.
2304 if (h.aa.atom < MO_128) {
2310 if (h.base == datalo || h.index == datalo) {
2311 tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2312 h.base, h.index, 0, h.ofs);
2313 tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2315 tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2318 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2319 h.base, h.index, 0, h.ofs);
2320 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2321 h.base, h.index, 0, h.ofs + 8);
2327 * With 16-byte atomicity, a vector load is required.
2328 * If we already have 16-byte alignment, then VMOVDQA always works.
2329 * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2330 * Else use we require a runtime test for alignment for VMOVDQA;
2331 * use VMOVDQU on the unaligned nonatomic path for simplicity.
2333 if (h.aa.align >= MO_128) {
2334 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2336 h.base, h.index, 0, h.ofs);
2337 } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2338 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2340 h.base, h.index, 0, h.ofs);
2342 TCGLabel *l1 = gen_new_label();
2343 TCGLabel *l2 = gen_new_label();
2346 jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2347 tcg_out_jxx(s, jcc, l1, true);
2349 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2351 h.base, h.index, 0, h.ofs);
2352 tcg_out_jxx(s, JCC_JMP, l2, true);
2354 tcg_out_label(s, l1);
2355 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2357 h.base, h.index, 0, h.ofs);
2358 tcg_out_label(s, l2);
2360 tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2364 g_assert_not_reached();
2368 static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2369 TCGReg addrlo, TCGReg addrhi,
2370 MemOpIdx oi, TCGType data_type)
2372 TCGLabelQemuLdst *ldst;
2375 ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
2376 tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2379 ldst->type = data_type;
2380 ldst->datalo_reg = datalo;
2381 ldst->datahi_reg = datahi;
2382 ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2386 static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2387 HostAddress h, MemOp memop)
2389 bool use_movbe = false;
2390 int movop = OPC_MOVL_EvGv;
2393 * Do big-endian stores with movbe or system-mode.
2394 * User-only without movbe will have its swapping done generically.
2396 if (memop & MO_BSWAP) {
2397 tcg_debug_assert(have_movbe);
2399 movop = OPC_MOVBE_MyGy;
2402 switch (memop & MO_SIZE) {
2404 /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2405 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2406 tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2407 datalo, h.base, h.index, 0, h.ofs);
2410 tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2411 h.base, h.index, 0, h.ofs);
2414 tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2415 h.base, h.index, 0, h.ofs);
2418 if (TCG_TARGET_REG_BITS == 64) {
2419 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2420 h.base, h.index, 0, h.ofs);
2427 tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2428 h.base, h.index, 0, h.ofs);
2429 tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2430 h.base, h.index, 0, h.ofs + 4);
2435 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2438 * Without 16-byte atomicity, use integer regs.
2439 * That is where we have the data, and it allows bswaps.
2441 if (h.aa.atom < MO_128) {
2447 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2448 h.base, h.index, 0, h.ofs);
2449 tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2450 h.base, h.index, 0, h.ofs + 8);
2455 * With 16-byte atomicity, a vector store is required.
2456 * If we already have 16-byte alignment, then VMOVDQA always works.
2457 * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2458 * Else use we require a runtime test for alignment for VMOVDQA;
2459 * use VMOVDQU on the unaligned nonatomic path for simplicity.
2461 tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2462 if (h.aa.align >= MO_128) {
2463 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2465 h.base, h.index, 0, h.ofs);
2466 } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2467 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2469 h.base, h.index, 0, h.ofs);
2471 TCGLabel *l1 = gen_new_label();
2472 TCGLabel *l2 = gen_new_label();
2475 jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2476 tcg_out_jxx(s, jcc, l1, true);
2478 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2480 h.base, h.index, 0, h.ofs);
2481 tcg_out_jxx(s, JCC_JMP, l2, true);
2483 tcg_out_label(s, l1);
2484 tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2486 h.base, h.index, 0, h.ofs);
2487 tcg_out_label(s, l2);
2492 g_assert_not_reached();
2496 static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2497 TCGReg addrlo, TCGReg addrhi,
2498 MemOpIdx oi, TCGType data_type)
2500 TCGLabelQemuLdst *ldst;
2503 ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
2504 tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2507 ldst->type = data_type;
2508 ldst->datalo_reg = datalo;
2509 ldst->datahi_reg = datahi;
2510 ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2514 static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2516 /* Reuse the zeroing that exists for goto_ptr. */
2518 tcg_out_jmp(s, tcg_code_gen_epilogue);
2520 tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2521 tcg_out_jmp(s, tb_ret_addr);
2525 static void tcg_out_goto_tb(TCGContext *s, int which)
2528 * Jump displacement must be aligned for atomic patching;
2529 * see if we need to add extra nops before jump
2531 int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2533 tcg_out_nopn(s, gap - 1);
2535 tcg_out8(s, OPC_JMP_long); /* jmp im */
2536 set_jmp_insn_offset(s, which);
2538 set_jmp_reset_offset(s, which);
2541 void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2542 uintptr_t jmp_rx, uintptr_t jmp_rw)
2544 /* patch the branch destination */
2545 uintptr_t addr = tb->jmp_target_addr[n];
2546 qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2547 /* no need to flush icache explicitly */
2550 static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2551 const TCGArg args[TCG_MAX_OP_ARGS],
2552 const int const_args[TCG_MAX_OP_ARGS])
2555 int c, const_a2, vexop, rexw = 0;
2557 #if TCG_TARGET_REG_BITS == 64
2558 # define OP_32_64(x) \
2559 case glue(glue(INDEX_op_, x), _i64): \
2560 rexw = P_REXW; /* FALLTHRU */ \
2561 case glue(glue(INDEX_op_, x), _i32)
2563 # define OP_32_64(x) \
2564 case glue(glue(INDEX_op_, x), _i32)
2567 /* Hoist the loads of the most common arguments. */
2571 const_a2 = const_args[2];
2574 case INDEX_op_goto_ptr:
2575 /* jmp to the given host address (could be epilogue) */
2576 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2579 tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2582 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
2583 tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2586 tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2589 /* Note that we can ignore REXW for the zero-extend to 64-bit. */
2590 tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2593 tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2595 #if TCG_TARGET_REG_BITS == 64
2596 case INDEX_op_ld32u_i64:
2598 case INDEX_op_ld_i32:
2599 tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2603 if (const_args[0]) {
2604 tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2607 tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2611 if (const_args[0]) {
2612 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2615 tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2618 #if TCG_TARGET_REG_BITS == 64
2619 case INDEX_op_st32_i64:
2621 case INDEX_op_st_i32:
2622 if (const_args[0]) {
2623 tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2626 tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2631 /* For 3-operand addition, use LEA. */
2636 } else if (a0 == a2) {
2637 /* Watch out for dest = src + dest, since we've removed
2638 the matching constraint on the add. */
2639 tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2643 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2662 tgen_arithi(s, c + rexw, a0, a2, 0);
2664 tgen_arithr(s, c + rexw, a0, a2);
2670 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2671 tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2673 tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2681 if (val == (int8_t)val) {
2682 tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2685 tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2689 tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2694 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2697 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2701 /* For small constant 3-operand shift, use LEA. */
2702 if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2704 /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2705 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2707 /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2708 tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2714 goto gen_shift_maybe_vex;
2718 goto gen_shift_maybe_vex;
2722 goto gen_shift_maybe_vex;
2729 gen_shift_maybe_vex:
2732 tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2735 tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2740 tcg_out_shifti(s, c + rexw, a0, a2);
2742 tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2747 tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2750 tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2753 tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2757 tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
2758 arg_label(args[3]), 0);
2761 tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
2763 OP_32_64(negsetcond):
2764 tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
2767 tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
2771 if (a2 & TCG_BSWAP_OS) {
2772 /* Output must be sign-extended. */
2774 tcg_out_bswap64(s, a0);
2775 tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2777 tcg_out_bswap32(s, a0);
2778 tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2780 } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2781 /* Output must be zero-extended, but input isn't. */
2782 tcg_out_bswap32(s, a0);
2783 tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2785 tcg_out_rolw_8(s, a0);
2789 tcg_out_bswap32(s, a0);
2790 if (rexw && (a2 & TCG_BSWAP_OS)) {
2791 tcg_out_ext32s(s, a0, a0);
2796 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2799 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2802 case INDEX_op_qemu_ld_a64_i32:
2803 if (TCG_TARGET_REG_BITS == 32) {
2804 tcg_out_qemu_ld(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2808 case INDEX_op_qemu_ld_a32_i32:
2809 tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2811 case INDEX_op_qemu_ld_a32_i64:
2812 if (TCG_TARGET_REG_BITS == 64) {
2813 tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2815 tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2818 case INDEX_op_qemu_ld_a64_i64:
2819 if (TCG_TARGET_REG_BITS == 64) {
2820 tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2822 tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2825 case INDEX_op_qemu_ld_a32_i128:
2826 case INDEX_op_qemu_ld_a64_i128:
2827 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2828 tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2831 case INDEX_op_qemu_st_a64_i32:
2832 case INDEX_op_qemu_st8_a64_i32:
2833 if (TCG_TARGET_REG_BITS == 32) {
2834 tcg_out_qemu_st(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2838 case INDEX_op_qemu_st_a32_i32:
2839 case INDEX_op_qemu_st8_a32_i32:
2840 tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2842 case INDEX_op_qemu_st_a32_i64:
2843 if (TCG_TARGET_REG_BITS == 64) {
2844 tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2846 tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2849 case INDEX_op_qemu_st_a64_i64:
2850 if (TCG_TARGET_REG_BITS == 64) {
2851 tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2853 tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2856 case INDEX_op_qemu_st_a32_i128:
2857 case INDEX_op_qemu_st_a64_i128:
2858 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2859 tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2863 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2866 tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2869 if (const_args[4]) {
2870 tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2872 tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2874 if (const_args[5]) {
2875 tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2877 tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2881 if (const_args[4]) {
2882 tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2884 tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2886 if (const_args[5]) {
2887 tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2889 tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2893 #if TCG_TARGET_REG_BITS == 32
2894 case INDEX_op_brcond2_i32:
2895 tcg_out_brcond2(s, args, const_args, 0);
2897 case INDEX_op_setcond2_i32:
2898 tcg_out_setcond2(s, args, const_args);
2900 #else /* TCG_TARGET_REG_BITS == 64 */
2901 case INDEX_op_ld32s_i64:
2902 tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2904 case INDEX_op_ld_i64:
2905 tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2907 case INDEX_op_st_i64:
2908 if (const_args[0]) {
2909 tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2912 tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2916 case INDEX_op_bswap64_i64:
2917 tcg_out_bswap64(s, a0);
2919 case INDEX_op_extrh_i64_i32:
2920 tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2925 if (args[3] == 0 && args[4] == 8) {
2926 /* load bits 0..7 */
2928 tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
2932 tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2934 } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
2935 /* load bits 8..15 */
2937 tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
2940 tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2942 } else if (args[3] == 0 && args[4] == 16) {
2943 /* load bits 0..15 */
2945 tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
2949 tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2952 g_assert_not_reached();
2956 case INDEX_op_extract_i64:
2957 if (a2 + args[3] == 32) {
2958 /* This is a 32-bit zero-extending right shift. */
2959 tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2960 tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2964 case INDEX_op_extract_i32:
2965 /* On the off-chance that we can use the high-byte registers.
2966 Otherwise we emit the same ext16 + shift pattern that we
2967 would have gotten from the normal tcg-op.c expansion. */
2968 tcg_debug_assert(a2 == 8 && args[3] == 8);
2969 if (a1 < 4 && a0 < 8) {
2970 tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2972 tcg_out_ext16u(s, a0, a1);
2973 tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2977 case INDEX_op_sextract_i32:
2978 /* We don't implement sextract_i64, as we cannot sign-extend to
2979 64-bits without using the REX prefix that explicitly excludes
2980 access to the high-byte registers. */
2981 tcg_debug_assert(a2 == 8 && args[3] == 8);
2982 if (a1 < 4 && a0 < 8) {
2983 tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2985 tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
2986 tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2991 /* Note that SHRD outputs to the r/m operand. */
2992 tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2993 tcg_out8(s, args[3]);
2999 case INDEX_op_mov_i32: /* Always emitted via tcg_out_mov. */
3000 case INDEX_op_mov_i64:
3001 case INDEX_op_call: /* Always emitted via tcg_out_call. */
3002 case INDEX_op_exit_tb: /* Always emitted via tcg_out_exit_tb. */
3003 case INDEX_op_goto_tb: /* Always emitted via tcg_out_goto_tb. */
3004 case INDEX_op_ext8s_i32: /* Always emitted via tcg_reg_alloc_op. */
3005 case INDEX_op_ext8s_i64:
3006 case INDEX_op_ext8u_i32:
3007 case INDEX_op_ext8u_i64:
3008 case INDEX_op_ext16s_i32:
3009 case INDEX_op_ext16s_i64:
3010 case INDEX_op_ext16u_i32:
3011 case INDEX_op_ext16u_i64:
3012 case INDEX_op_ext32s_i64:
3013 case INDEX_op_ext32u_i64:
3014 case INDEX_op_ext_i32_i64:
3015 case INDEX_op_extu_i32_i64:
3016 case INDEX_op_extrl_i64_i32:
3018 g_assert_not_reached();
3024 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3025 unsigned vecl, unsigned vece,
3026 const TCGArg args[TCG_MAX_OP_ARGS],
3027 const int const_args[TCG_MAX_OP_ARGS])
3029 static int const add_insn[4] = {
3030 OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3032 static int const ssadd_insn[4] = {
3033 OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3035 static int const usadd_insn[4] = {
3036 OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3038 static int const sub_insn[4] = {
3039 OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3041 static int const sssub_insn[4] = {
3042 OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3044 static int const ussub_insn[4] = {
3045 OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3047 static int const mul_insn[4] = {
3048 OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3050 static int const shift_imm_insn[4] = {
3051 OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3053 static int const cmpeq_insn[4] = {
3054 OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3056 static int const cmpgt_insn[4] = {
3057 OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3059 static int const punpckl_insn[4] = {
3060 OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3062 static int const punpckh_insn[4] = {
3063 OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3065 static int const packss_insn[4] = {
3066 OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3068 static int const packus_insn[4] = {
3069 OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3071 static int const smin_insn[4] = {
3072 OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3074 static int const smax_insn[4] = {
3075 OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3077 static int const umin_insn[4] = {
3078 OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3080 static int const umax_insn[4] = {
3081 OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3083 static int const rotlv_insn[4] = {
3084 OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3086 static int const rotrv_insn[4] = {
3087 OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3089 static int const shlv_insn[4] = {
3090 OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3092 static int const shrv_insn[4] = {
3093 OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3095 static int const sarv_insn[4] = {
3096 OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3098 static int const shls_insn[4] = {
3099 OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3101 static int const shrs_insn[4] = {
3102 OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3104 static int const sars_insn[4] = {
3105 OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3107 static int const vpshldi_insn[4] = {
3108 OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3110 static int const vpshldv_insn[4] = {
3111 OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3113 static int const vpshrdv_insn[4] = {
3114 OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3116 static int const abs_insn[4] = {
3117 OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3120 TCGType type = vecl + TCG_TYPE_V64;
3122 TCGArg a0, a1, a2, a3;
3129 case INDEX_op_add_vec:
3130 insn = add_insn[vece];
3132 case INDEX_op_ssadd_vec:
3133 insn = ssadd_insn[vece];
3135 case INDEX_op_usadd_vec:
3136 insn = usadd_insn[vece];
3138 case INDEX_op_sub_vec:
3139 insn = sub_insn[vece];
3141 case INDEX_op_sssub_vec:
3142 insn = sssub_insn[vece];
3144 case INDEX_op_ussub_vec:
3145 insn = ussub_insn[vece];
3147 case INDEX_op_mul_vec:
3148 insn = mul_insn[vece];
3150 case INDEX_op_and_vec:
3153 case INDEX_op_or_vec:
3156 case INDEX_op_xor_vec:
3159 case INDEX_op_smin_vec:
3160 insn = smin_insn[vece];
3162 case INDEX_op_umin_vec:
3163 insn = umin_insn[vece];
3165 case INDEX_op_smax_vec:
3166 insn = smax_insn[vece];
3168 case INDEX_op_umax_vec:
3169 insn = umax_insn[vece];
3171 case INDEX_op_shlv_vec:
3172 insn = shlv_insn[vece];
3174 case INDEX_op_shrv_vec:
3175 insn = shrv_insn[vece];
3177 case INDEX_op_sarv_vec:
3178 insn = sarv_insn[vece];
3180 case INDEX_op_rotlv_vec:
3181 insn = rotlv_insn[vece];
3183 case INDEX_op_rotrv_vec:
3184 insn = rotrv_insn[vece];
3186 case INDEX_op_shls_vec:
3187 insn = shls_insn[vece];
3189 case INDEX_op_shrs_vec:
3190 insn = shrs_insn[vece];
3192 case INDEX_op_sars_vec:
3193 insn = sars_insn[vece];
3195 case INDEX_op_x86_punpckl_vec:
3196 insn = punpckl_insn[vece];
3198 case INDEX_op_x86_punpckh_vec:
3199 insn = punpckh_insn[vece];
3201 case INDEX_op_x86_packss_vec:
3202 insn = packss_insn[vece];
3204 case INDEX_op_x86_packus_vec:
3205 insn = packus_insn[vece];
3207 case INDEX_op_x86_vpshldv_vec:
3208 insn = vpshldv_insn[vece];
3212 case INDEX_op_x86_vpshrdv_vec:
3213 insn = vpshrdv_insn[vece];
3217 #if TCG_TARGET_REG_BITS == 32
3218 case INDEX_op_dup2_vec:
3219 /* First merge the two 32-bit inputs to a single 64-bit element. */
3220 tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3221 /* Then replicate the 64-bit elements across the rest of the vector. */
3222 if (type != TCG_TYPE_V64) {
3223 tcg_out_dup_vec(s, type, MO_64, a0, a0);
3227 case INDEX_op_abs_vec:
3228 insn = abs_insn[vece];
3233 tcg_debug_assert(insn != OPC_UD2);
3234 if (type == TCG_TYPE_V256) {
3237 tcg_out_vex_modrm(s, insn, a0, a1, a2);
3240 case INDEX_op_cmp_vec:
3242 if (sub == TCG_COND_EQ) {
3243 insn = cmpeq_insn[vece];
3244 } else if (sub == TCG_COND_GT) {
3245 insn = cmpgt_insn[vece];
3247 g_assert_not_reached();
3251 case INDEX_op_andc_vec:
3253 if (type == TCG_TYPE_V256) {
3256 tcg_out_vex_modrm(s, insn, a0, a2, a1);
3259 case INDEX_op_shli_vec:
3260 insn = shift_imm_insn[vece];
3263 case INDEX_op_shri_vec:
3264 insn = shift_imm_insn[vece];
3267 case INDEX_op_sari_vec:
3268 if (vece == MO_64) {
3269 insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3271 insn = shift_imm_insn[vece];
3275 case INDEX_op_rotli_vec:
3276 insn = OPC_PSHIFTD_Ib | P_EVEX; /* VPROL[DQ] */
3277 if (vece == MO_64) {
3283 tcg_debug_assert(vece != MO_8);
3284 if (type == TCG_TYPE_V256) {
3287 tcg_out_vex_modrm(s, insn, sub, a0, a1);
3291 case INDEX_op_ld_vec:
3292 tcg_out_ld(s, type, a0, a1, a2);
3294 case INDEX_op_st_vec:
3295 tcg_out_st(s, type, a0, a1, a2);
3297 case INDEX_op_dupm_vec:
3298 tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3301 case INDEX_op_x86_shufps_vec:
3305 case INDEX_op_x86_blend_vec:
3306 if (vece == MO_16) {
3308 } else if (vece == MO_32) {
3309 insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3311 g_assert_not_reached();
3315 case INDEX_op_x86_vperm2i128_vec:
3316 insn = OPC_VPERM2I128;
3319 case INDEX_op_x86_vpshldi_vec:
3320 insn = vpshldi_insn[vece];
3324 case INDEX_op_not_vec:
3325 insn = OPC_VPTERNLOGQ;
3327 sub = 0x33; /* !B */
3329 case INDEX_op_nor_vec:
3330 insn = OPC_VPTERNLOGQ;
3331 sub = 0x11; /* norCB */
3333 case INDEX_op_nand_vec:
3334 insn = OPC_VPTERNLOGQ;
3335 sub = 0x77; /* nandCB */
3337 case INDEX_op_eqv_vec:
3338 insn = OPC_VPTERNLOGQ;
3339 sub = 0x99; /* xnorCB */
3341 case INDEX_op_orc_vec:
3342 insn = OPC_VPTERNLOGQ;
3343 sub = 0xdd; /* orB!C */
3346 case INDEX_op_bitsel_vec:
3347 insn = OPC_VPTERNLOGQ;
3352 sub = 0xca; /* A?B:C */
3353 } else if (a0 == a2) {
3355 sub = 0xe2; /* B?A:C */
3357 tcg_out_mov(s, type, a0, a3);
3358 sub = 0xb8; /* B?C:A */
3363 tcg_debug_assert(insn != OPC_UD2);
3364 if (type == TCG_TYPE_V256) {
3367 tcg_out_vex_modrm(s, insn, a0, a1, a2);
3371 case INDEX_op_x86_vpblendvb_vec:
3372 insn = OPC_VPBLENDVB;
3373 if (type == TCG_TYPE_V256) {
3376 tcg_out_vex_modrm(s, insn, a0, a1, a2);
3377 tcg_out8(s, args[3] << 4);
3380 case INDEX_op_x86_psrldq_vec:
3381 tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3385 case INDEX_op_mov_vec: /* Always emitted via tcg_out_mov. */
3386 case INDEX_op_dup_vec: /* Always emitted via tcg_out_dup_vec. */
3388 g_assert_not_reached();
3392 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3395 case INDEX_op_goto_ptr:
3398 case INDEX_op_ld8u_i32:
3399 case INDEX_op_ld8u_i64:
3400 case INDEX_op_ld8s_i32:
3401 case INDEX_op_ld8s_i64:
3402 case INDEX_op_ld16u_i32:
3403 case INDEX_op_ld16u_i64:
3404 case INDEX_op_ld16s_i32:
3405 case INDEX_op_ld16s_i64:
3406 case INDEX_op_ld_i32:
3407 case INDEX_op_ld32u_i64:
3408 case INDEX_op_ld32s_i64:
3409 case INDEX_op_ld_i64:
3410 return C_O1_I1(r, r);
3412 case INDEX_op_st8_i32:
3413 case INDEX_op_st8_i64:
3414 return C_O0_I2(qi, r);
3416 case INDEX_op_st16_i32:
3417 case INDEX_op_st16_i64:
3418 case INDEX_op_st_i32:
3419 case INDEX_op_st32_i64:
3420 return C_O0_I2(ri, r);
3422 case INDEX_op_st_i64:
3423 return C_O0_I2(re, r);
3425 case INDEX_op_add_i32:
3426 case INDEX_op_add_i64:
3427 return C_O1_I2(r, r, re);
3429 case INDEX_op_sub_i32:
3430 case INDEX_op_sub_i64:
3431 case INDEX_op_mul_i32:
3432 case INDEX_op_mul_i64:
3433 case INDEX_op_or_i32:
3434 case INDEX_op_or_i64:
3435 case INDEX_op_xor_i32:
3436 case INDEX_op_xor_i64:
3437 return C_O1_I2(r, 0, re);
3439 case INDEX_op_and_i32:
3440 case INDEX_op_and_i64:
3441 return C_O1_I2(r, 0, reZ);
3443 case INDEX_op_andc_i32:
3444 case INDEX_op_andc_i64:
3445 return C_O1_I2(r, r, rI);
3447 case INDEX_op_shl_i32:
3448 case INDEX_op_shl_i64:
3449 case INDEX_op_shr_i32:
3450 case INDEX_op_shr_i64:
3451 case INDEX_op_sar_i32:
3452 case INDEX_op_sar_i64:
3453 return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3455 case INDEX_op_rotl_i32:
3456 case INDEX_op_rotl_i64:
3457 case INDEX_op_rotr_i32:
3458 case INDEX_op_rotr_i64:
3459 return C_O1_I2(r, 0, ci);
3461 case INDEX_op_brcond_i32:
3462 case INDEX_op_brcond_i64:
3463 return C_O0_I2(r, reT);
3465 case INDEX_op_bswap16_i32:
3466 case INDEX_op_bswap16_i64:
3467 case INDEX_op_bswap32_i32:
3468 case INDEX_op_bswap32_i64:
3469 case INDEX_op_bswap64_i64:
3470 case INDEX_op_neg_i32:
3471 case INDEX_op_neg_i64:
3472 case INDEX_op_not_i32:
3473 case INDEX_op_not_i64:
3474 case INDEX_op_extrh_i64_i32:
3475 return C_O1_I1(r, 0);
3477 case INDEX_op_ext8s_i32:
3478 case INDEX_op_ext8s_i64:
3479 case INDEX_op_ext8u_i32:
3480 case INDEX_op_ext8u_i64:
3481 return C_O1_I1(r, q);
3483 case INDEX_op_ext16s_i32:
3484 case INDEX_op_ext16s_i64:
3485 case INDEX_op_ext16u_i32:
3486 case INDEX_op_ext16u_i64:
3487 case INDEX_op_ext32s_i64:
3488 case INDEX_op_ext32u_i64:
3489 case INDEX_op_ext_i32_i64:
3490 case INDEX_op_extu_i32_i64:
3491 case INDEX_op_extrl_i64_i32:
3492 case INDEX_op_extract_i32:
3493 case INDEX_op_extract_i64:
3494 case INDEX_op_sextract_i32:
3495 case INDEX_op_ctpop_i32:
3496 case INDEX_op_ctpop_i64:
3497 return C_O1_I1(r, r);
3499 case INDEX_op_extract2_i32:
3500 case INDEX_op_extract2_i64:
3501 return C_O1_I2(r, 0, r);
3503 case INDEX_op_deposit_i32:
3504 case INDEX_op_deposit_i64:
3505 return C_O1_I2(q, 0, qi);
3507 case INDEX_op_setcond_i32:
3508 case INDEX_op_setcond_i64:
3509 case INDEX_op_negsetcond_i32:
3510 case INDEX_op_negsetcond_i64:
3511 return C_O1_I2(q, r, reT);
3513 case INDEX_op_movcond_i32:
3514 case INDEX_op_movcond_i64:
3515 return C_O1_I4(r, r, reT, r, 0);
3517 case INDEX_op_div2_i32:
3518 case INDEX_op_div2_i64:
3519 case INDEX_op_divu2_i32:
3520 case INDEX_op_divu2_i64:
3521 return C_O2_I3(a, d, 0, 1, r);
3523 case INDEX_op_mulu2_i32:
3524 case INDEX_op_mulu2_i64:
3525 case INDEX_op_muls2_i32:
3526 case INDEX_op_muls2_i64:
3527 return C_O2_I2(a, d, a, r);
3529 case INDEX_op_add2_i32:
3530 case INDEX_op_add2_i64:
3531 case INDEX_op_sub2_i32:
3532 case INDEX_op_sub2_i64:
3533 return C_N1_O1_I4(r, r, 0, 1, re, re);
3535 case INDEX_op_ctz_i32:
3536 case INDEX_op_ctz_i64:
3537 return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3539 case INDEX_op_clz_i32:
3540 case INDEX_op_clz_i64:
3541 return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3543 case INDEX_op_qemu_ld_a32_i32:
3544 return C_O1_I1(r, L);
3545 case INDEX_op_qemu_ld_a64_i32:
3546 return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O1_I2(r, L, L);
3548 case INDEX_op_qemu_st_a32_i32:
3549 return C_O0_I2(L, L);
3550 case INDEX_op_qemu_st_a64_i32:
3551 return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3552 case INDEX_op_qemu_st8_a32_i32:
3553 return C_O0_I2(s, L);
3554 case INDEX_op_qemu_st8_a64_i32:
3555 return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(s, L) : C_O0_I3(s, L, L);
3557 case INDEX_op_qemu_ld_a32_i64:
3558 return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3559 case INDEX_op_qemu_ld_a64_i64:
3560 return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I2(r, r, L, L);
3562 case INDEX_op_qemu_st_a32_i64:
3563 return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3564 case INDEX_op_qemu_st_a64_i64:
3565 return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
3567 case INDEX_op_qemu_ld_a32_i128:
3568 case INDEX_op_qemu_ld_a64_i128:
3569 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3570 return C_O2_I1(r, r, L);
3571 case INDEX_op_qemu_st_a32_i128:
3572 case INDEX_op_qemu_st_a64_i128:
3573 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3574 return C_O0_I3(L, L, L);
3576 case INDEX_op_brcond2_i32:
3577 return C_O0_I4(r, r, ri, ri);
3579 case INDEX_op_setcond2_i32:
3580 return C_O1_I4(r, r, r, ri, ri);
3582 case INDEX_op_ld_vec:
3583 case INDEX_op_dupm_vec:
3584 return C_O1_I1(x, r);
3586 case INDEX_op_st_vec:
3587 return C_O0_I2(x, r);
3589 case INDEX_op_add_vec:
3590 case INDEX_op_sub_vec:
3591 case INDEX_op_mul_vec:
3592 case INDEX_op_and_vec:
3593 case INDEX_op_or_vec:
3594 case INDEX_op_xor_vec:
3595 case INDEX_op_andc_vec:
3596 case INDEX_op_orc_vec:
3597 case INDEX_op_nand_vec:
3598 case INDEX_op_nor_vec:
3599 case INDEX_op_eqv_vec:
3600 case INDEX_op_ssadd_vec:
3601 case INDEX_op_usadd_vec:
3602 case INDEX_op_sssub_vec:
3603 case INDEX_op_ussub_vec:
3604 case INDEX_op_smin_vec:
3605 case INDEX_op_umin_vec:
3606 case INDEX_op_smax_vec:
3607 case INDEX_op_umax_vec:
3608 case INDEX_op_shlv_vec:
3609 case INDEX_op_shrv_vec:
3610 case INDEX_op_sarv_vec:
3611 case INDEX_op_rotlv_vec:
3612 case INDEX_op_rotrv_vec:
3613 case INDEX_op_shls_vec:
3614 case INDEX_op_shrs_vec:
3615 case INDEX_op_sars_vec:
3616 case INDEX_op_cmp_vec:
3617 case INDEX_op_x86_shufps_vec:
3618 case INDEX_op_x86_blend_vec:
3619 case INDEX_op_x86_packss_vec:
3620 case INDEX_op_x86_packus_vec:
3621 case INDEX_op_x86_vperm2i128_vec:
3622 case INDEX_op_x86_punpckl_vec:
3623 case INDEX_op_x86_punpckh_vec:
3624 case INDEX_op_x86_vpshldi_vec:
3625 #if TCG_TARGET_REG_BITS == 32
3626 case INDEX_op_dup2_vec:
3628 return C_O1_I2(x, x, x);
3630 case INDEX_op_abs_vec:
3631 case INDEX_op_dup_vec:
3632 case INDEX_op_not_vec:
3633 case INDEX_op_shli_vec:
3634 case INDEX_op_shri_vec:
3635 case INDEX_op_sari_vec:
3636 case INDEX_op_rotli_vec:
3637 case INDEX_op_x86_psrldq_vec:
3638 return C_O1_I1(x, x);
3640 case INDEX_op_x86_vpshldv_vec:
3641 case INDEX_op_x86_vpshrdv_vec:
3642 return C_O1_I3(x, 0, x, x);
3644 case INDEX_op_bitsel_vec:
3645 case INDEX_op_x86_vpblendvb_vec:
3646 return C_O1_I3(x, x, x, x);
3649 g_assert_not_reached();
3653 int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3656 case INDEX_op_add_vec:
3657 case INDEX_op_sub_vec:
3658 case INDEX_op_and_vec:
3659 case INDEX_op_or_vec:
3660 case INDEX_op_xor_vec:
3661 case INDEX_op_andc_vec:
3662 case INDEX_op_orc_vec:
3663 case INDEX_op_nand_vec:
3664 case INDEX_op_nor_vec:
3665 case INDEX_op_eqv_vec:
3666 case INDEX_op_not_vec:
3667 case INDEX_op_bitsel_vec:
3669 case INDEX_op_cmp_vec:
3670 case INDEX_op_cmpsel_vec:
3673 case INDEX_op_rotli_vec:
3674 return have_avx512vl && vece >= MO_32 ? 1 : -1;
3676 case INDEX_op_shli_vec:
3677 case INDEX_op_shri_vec:
3678 /* We must expand the operation for MO_8. */
3679 return vece == MO_8 ? -1 : 1;
3681 case INDEX_op_sari_vec:
3689 if (have_avx512vl) {
3693 * We can emulate this for MO_64, but it does not pay off
3694 * unless we're producing at least 4 values.
3696 return type >= TCG_TYPE_V256 ? -1 : 0;
3700 case INDEX_op_shls_vec:
3701 case INDEX_op_shrs_vec:
3702 return vece >= MO_16;
3703 case INDEX_op_sars_vec:
3709 return have_avx512vl;
3712 case INDEX_op_rotls_vec:
3713 return vece >= MO_16 ? -1 : 0;
3715 case INDEX_op_shlv_vec:
3716 case INDEX_op_shrv_vec:
3719 return have_avx512bw;
3725 case INDEX_op_sarv_vec:
3728 return have_avx512bw;
3732 return have_avx512vl;
3735 case INDEX_op_rotlv_vec:
3736 case INDEX_op_rotrv_vec:
3739 return have_avx512vbmi2 ? -1 : 0;
3742 return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3746 case INDEX_op_mul_vec:
3751 return have_avx512dq;
3755 case INDEX_op_ssadd_vec:
3756 case INDEX_op_usadd_vec:
3757 case INDEX_op_sssub_vec:
3758 case INDEX_op_ussub_vec:
3759 return vece <= MO_16;
3760 case INDEX_op_smin_vec:
3761 case INDEX_op_smax_vec:
3762 case INDEX_op_umin_vec:
3763 case INDEX_op_umax_vec:
3764 case INDEX_op_abs_vec:
3765 return vece <= MO_32 || have_avx512vl;
3772 static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3773 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3777 tcg_debug_assert(vece == MO_8);
3779 t1 = tcg_temp_new_vec(type);
3780 t2 = tcg_temp_new_vec(type);
3783 * Unpack to W, shift, and repack. Tricky bits:
3784 * (1) Use punpck*bw x,x to produce DDCCBBAA,
3785 * i.e. duplicate in other half of the 16-bit lane.
3786 * (2) For right-shift, add 8 so that the high half of the lane
3787 * becomes zero. For left-shift, and left-rotate, we must
3788 * shift up and down again.
3789 * (3) Step 2 leaves high half zero such that PACKUSWB
3790 * (pack with unsigned saturation) does not modify
3793 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3794 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3795 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3796 tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3798 if (opc != INDEX_op_rotli_vec) {
3801 if (opc == INDEX_op_shri_vec) {
3802 tcg_gen_shri_vec(MO_16, t1, t1, imm);
3803 tcg_gen_shri_vec(MO_16, t2, t2, imm);
3805 tcg_gen_shli_vec(MO_16, t1, t1, imm);
3806 tcg_gen_shli_vec(MO_16, t2, t2, imm);
3807 tcg_gen_shri_vec(MO_16, t1, t1, 8);
3808 tcg_gen_shri_vec(MO_16, t2, t2, 8);
3811 vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3812 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3813 tcg_temp_free_vec(t1);
3814 tcg_temp_free_vec(t2);
3817 static void expand_vec_sari(TCGType type, unsigned vece,
3818 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3824 /* Unpack to W, shift, and repack, as in expand_vec_shi. */
3825 t1 = tcg_temp_new_vec(type);
3826 t2 = tcg_temp_new_vec(type);
3827 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3828 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3829 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3830 tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3831 tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3832 tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3833 vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3834 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3835 tcg_temp_free_vec(t1);
3836 tcg_temp_free_vec(t2);
3840 t1 = tcg_temp_new_vec(type);
3843 * We can emulate a small sign extend by performing an arithmetic
3844 * 32-bit shift and overwriting the high half of a 64-bit logical
3845 * shift. Note that the ISA says shift of 32 is valid, but TCG
3846 * does not, so we have to bound the smaller shift -- we get the
3847 * same result in the high half either way.
3849 tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3850 tcg_gen_shri_vec(MO_64, v0, v1, imm);
3851 vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3852 tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3853 tcgv_vec_arg(t1), 0xaa);
3855 /* Otherwise we will need to use a compare vs 0 to produce
3856 * the sign-extend, shift and merge.
3858 tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
3859 tcg_constant_vec(type, MO_64, 0), v1);
3860 tcg_gen_shri_vec(MO_64, v0, v1, imm);
3861 tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3862 tcg_gen_or_vec(MO_64, v0, v0, t1);
3864 tcg_temp_free_vec(t1);
3868 g_assert_not_reached();
3872 static void expand_vec_rotli(TCGType type, unsigned vece,
3873 TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3878 expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3882 if (have_avx512vbmi2) {
3883 vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3884 tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3888 t = tcg_temp_new_vec(type);
3889 tcg_gen_shli_vec(vece, t, v1, imm);
3890 tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3891 tcg_gen_or_vec(vece, v0, v0, t);
3892 tcg_temp_free_vec(t);
3895 static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3896 TCGv_vec v1, TCGv_vec sh, bool right)
3900 if (have_avx512vbmi2) {
3901 vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3902 type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3903 tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3907 t = tcg_temp_new_vec(type);
3908 tcg_gen_dupi_vec(vece, t, 8 << vece);
3909 tcg_gen_sub_vec(vece, t, t, sh);
3911 tcg_gen_shlv_vec(vece, t, v1, t);
3912 tcg_gen_shrv_vec(vece, v0, v1, sh);
3914 tcg_gen_shrv_vec(vece, t, v1, t);
3915 tcg_gen_shlv_vec(vece, v0, v1, sh);
3917 tcg_gen_or_vec(vece, v0, v0, t);
3918 tcg_temp_free_vec(t);
3921 static void expand_vec_rotls(TCGType type, unsigned vece,
3922 TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3924 TCGv_vec t = tcg_temp_new_vec(type);
3926 tcg_debug_assert(vece != MO_8);
3928 if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3929 tcg_gen_dup_i32_vec(vece, t, lsh);
3930 if (vece >= MO_32) {
3931 tcg_gen_rotlv_vec(vece, v0, v1, t);
3933 expand_vec_rotv(type, vece, v0, v1, t, false);
3936 TCGv_i32 rsh = tcg_temp_new_i32();
3938 tcg_gen_neg_i32(rsh, lsh);
3939 tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3940 tcg_gen_shls_vec(vece, t, v1, lsh);
3941 tcg_gen_shrs_vec(vece, v0, v1, rsh);
3942 tcg_gen_or_vec(vece, v0, v0, t);
3944 tcg_temp_free_i32(rsh);
3947 tcg_temp_free_vec(t);
3950 static void expand_vec_mul(TCGType type, unsigned vece,
3951 TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3953 TCGv_vec t1, t2, t3, t4, zero;
3955 tcg_debug_assert(vece == MO_8);
3958 * Unpack v1 bytes to words, 0 | x.
3959 * Unpack v2 bytes to words, y | 0.
3960 * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3961 * Shift logical right by 8 bits to clear the high 8 bytes before
3962 * using an unsigned saturated pack.
3964 * The difference between the V64, V128 and V256 cases is merely how
3965 * we distribute the expansion between temporaries.
3969 t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3970 t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3971 zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3972 vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3973 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3974 vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3975 tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3976 tcg_gen_mul_vec(MO_16, t1, t1, t2);
3977 tcg_gen_shri_vec(MO_16, t1, t1, 8);
3978 vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3979 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3980 tcg_temp_free_vec(t1);
3981 tcg_temp_free_vec(t2);
3986 t1 = tcg_temp_new_vec(type);
3987 t2 = tcg_temp_new_vec(type);
3988 t3 = tcg_temp_new_vec(type);
3989 t4 = tcg_temp_new_vec(type);
3990 zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3991 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3992 tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3993 vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3994 tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3995 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3996 tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3997 vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3998 tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3999 tcg_gen_mul_vec(MO_16, t1, t1, t2);
4000 tcg_gen_mul_vec(MO_16, t3, t3, t4);
4001 tcg_gen_shri_vec(MO_16, t1, t1, 8);
4002 tcg_gen_shri_vec(MO_16, t3, t3, 8);
4003 vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4004 tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4005 tcg_temp_free_vec(t1);
4006 tcg_temp_free_vec(t2);
4007 tcg_temp_free_vec(t3);
4008 tcg_temp_free_vec(t4);
4012 g_assert_not_reached();
4016 static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
4017 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
4026 TCGv_vec t1, t2, t3;
4042 fixup = NEED_SWAP | NEED_INV;
4045 if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
4048 fixup = NEED_BIAS | NEED_INV;
4052 if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
4053 fixup = NEED_UMIN | NEED_INV;
4059 if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
4062 fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
4066 if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
4067 fixup = NEED_UMAX | NEED_INV;
4069 fixup = NEED_BIAS | NEED_SWAP;
4073 g_assert_not_reached();
4076 if (fixup & NEED_INV) {
4077 cond = tcg_invert_cond(cond);
4079 if (fixup & NEED_SWAP) {
4080 t1 = v1, v1 = v2, v2 = t1;
4081 cond = tcg_swap_cond(cond);
4085 if (fixup & (NEED_UMIN | NEED_UMAX)) {
4086 t1 = tcg_temp_new_vec(type);
4087 if (fixup & NEED_UMIN) {
4088 tcg_gen_umin_vec(vece, t1, v1, v2);
4090 tcg_gen_umax_vec(vece, t1, v1, v2);
4094 } else if (fixup & NEED_BIAS) {
4095 t1 = tcg_temp_new_vec(type);
4096 t2 = tcg_temp_new_vec(type);
4097 t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4098 tcg_gen_sub_vec(vece, t1, v1, t3);
4099 tcg_gen_sub_vec(vece, t2, v2, t3);
4102 cond = tcg_signed_cond(cond);
4105 tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
4106 /* Expand directly; do not recurse. */
4107 vec_gen_4(INDEX_op_cmp_vec, type, vece,
4108 tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
4111 tcg_temp_free_vec(t1);
4113 tcg_temp_free_vec(t2);
4116 return fixup & NEED_INV;
4119 static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
4120 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
4122 if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
4123 tcg_gen_not_vec(vece, v0, v0);
4127 static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
4128 TCGv_vec c1, TCGv_vec c2,
4129 TCGv_vec v3, TCGv_vec v4, TCGCond cond)
4131 TCGv_vec t = tcg_temp_new_vec(type);
4133 if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
4134 /* Invert the sense of the compare by swapping arguments. */
4136 x = v3, v3 = v4, v4 = x;
4138 vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
4139 tcgv_vec_arg(v0), tcgv_vec_arg(v4),
4140 tcgv_vec_arg(v3), tcgv_vec_arg(t));
4141 tcg_temp_free_vec(t);
4144 void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4149 TCGv_vec v0, v1, v2, v3, v4;
4152 v0 = temp_tcgv_vec(arg_temp(a0));
4153 v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4154 a2 = va_arg(va, TCGArg);
4157 case INDEX_op_shli_vec:
4158 case INDEX_op_shri_vec:
4159 expand_vec_shi(type, vece, opc, v0, v1, a2);
4162 case INDEX_op_sari_vec:
4163 expand_vec_sari(type, vece, v0, v1, a2);
4166 case INDEX_op_rotli_vec:
4167 expand_vec_rotli(type, vece, v0, v1, a2);
4170 case INDEX_op_rotls_vec:
4171 expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4174 case INDEX_op_rotlv_vec:
4175 v2 = temp_tcgv_vec(arg_temp(a2));
4176 expand_vec_rotv(type, vece, v0, v1, v2, false);
4178 case INDEX_op_rotrv_vec:
4179 v2 = temp_tcgv_vec(arg_temp(a2));
4180 expand_vec_rotv(type, vece, v0, v1, v2, true);
4183 case INDEX_op_mul_vec:
4184 v2 = temp_tcgv_vec(arg_temp(a2));
4185 expand_vec_mul(type, vece, v0, v1, v2);
4188 case INDEX_op_cmp_vec:
4189 v2 = temp_tcgv_vec(arg_temp(a2));
4190 expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
4193 case INDEX_op_cmpsel_vec:
4194 v2 = temp_tcgv_vec(arg_temp(a2));
4195 v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4196 v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4197 expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
4207 static const int tcg_target_callee_save_regs[] = {
4208 #if TCG_TARGET_REG_BITS == 64
4217 TCG_REG_R14, /* Currently used for the global env. */
4224 TCG_REG_EBP, /* Currently used for the global env. */
4231 /* Compute frame size via macros, to share between tcg_target_qemu_prologue
4232 and tcg_register_jit. */
4235 ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4236 * (TCG_TARGET_REG_BITS / 8))
4238 #define FRAME_SIZE \
4240 + TCG_STATIC_CALL_ARGS_SIZE \
4241 + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4242 + TCG_TARGET_STACK_ALIGN - 1) \
4243 & ~(TCG_TARGET_STACK_ALIGN - 1))
4245 /* Generate global QEMU prologue and epilogue code */
4246 static void tcg_target_qemu_prologue(TCGContext *s)
4248 int i, stack_addend;
4252 /* Reserve some stack space, also for TCG temps. */
4253 stack_addend = FRAME_SIZE - PUSH_SIZE;
4254 tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4255 CPU_TEMP_BUF_NLONGS * sizeof(long));
4257 /* Save all callee saved registers. */
4258 for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4259 tcg_out_push(s, tcg_target_callee_save_regs[i]);
4262 if (!tcg_use_softmmu && guest_base) {
4263 int seg = setup_guest_base_seg();
4265 x86_guest_base.seg = seg;
4266 } else if (guest_base == (int32_t)guest_base) {
4267 x86_guest_base.ofs = guest_base;
4269 assert(TCG_TARGET_REG_BITS == 64);
4270 /* Choose R12 because, as a base, it requires a SIB byte. */
4271 x86_guest_base.index = TCG_REG_R12;
4272 tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4273 tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4277 if (TCG_TARGET_REG_BITS == 32) {
4278 tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4279 (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4280 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4282 tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4283 (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4286 tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4287 tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4289 tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4293 * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4294 * and fall through to the rest of the epilogue.
4296 tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4297 tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4300 tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4302 tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4305 tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4307 for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4308 tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4310 tcg_out_opc(s, OPC_RET, 0, 0, 0);
4313 static void tcg_out_tb_start(TCGContext *s)
4318 static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4320 memset(p, 0x90, count);
4323 static void tcg_target_init(TCGContext *s)
4325 tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4326 if (TCG_TARGET_REG_BITS == 64) {
4327 tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4330 tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4331 tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4334 tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4337 tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4338 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4339 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4340 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4341 if (TCG_TARGET_REG_BITS == 64) {
4342 #if !defined(_WIN64)
4343 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4344 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4346 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4347 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4348 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4349 tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4352 s->reserved_regs = 0;
4353 tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4354 tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4356 /* These are call saved, and we don't save them, so don't use them. */
4357 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4358 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4359 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4360 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4361 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4362 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4363 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4364 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4365 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4366 tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4372 uint8_t fde_def_cfa[4];
4373 uint8_t fde_reg_ofs[14];
4376 /* We're expecting a 2 byte uleb128 encoded value. */
4377 QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4379 #if !defined(__ELF__)
4380 /* Host machine without ELF. */
4381 #elif TCG_TARGET_REG_BITS == 64
4382 #define ELF_HOST_MACHINE EM_X86_64
4383 static const DebugFrame debug_frame = {
4384 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4387 .h.cie.code_align = 1,
4388 .h.cie.data_align = 0x78, /* sleb128 -8 */
4389 .h.cie.return_column = 16,
4391 /* Total FDE size does not include the "len" member. */
4392 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4395 12, 7, /* DW_CFA_def_cfa %rsp, ... */
4396 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
4400 0x90, 1, /* DW_CFA_offset, %rip, -8 */
4401 /* The following ordering must match tcg_target_callee_save_regs. */
4402 0x86, 2, /* DW_CFA_offset, %rbp, -16 */
4403 0x83, 3, /* DW_CFA_offset, %rbx, -24 */
4404 0x8c, 4, /* DW_CFA_offset, %r12, -32 */
4405 0x8d, 5, /* DW_CFA_offset, %r13, -40 */
4406 0x8e, 6, /* DW_CFA_offset, %r14, -48 */
4407 0x8f, 7, /* DW_CFA_offset, %r15, -56 */
4411 #define ELF_HOST_MACHINE EM_386
4412 static const DebugFrame debug_frame = {
4413 .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4416 .h.cie.code_align = 1,
4417 .h.cie.data_align = 0x7c, /* sleb128 -4 */
4418 .h.cie.return_column = 8,
4420 /* Total FDE size does not include the "len" member. */
4421 .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4424 12, 4, /* DW_CFA_def_cfa %esp, ... */
4425 (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
4429 0x88, 1, /* DW_CFA_offset, %eip, -4 */
4430 /* The following ordering must match tcg_target_callee_save_regs. */
4431 0x85, 2, /* DW_CFA_offset, %ebp, -8 */
4432 0x83, 3, /* DW_CFA_offset, %ebx, -12 */
4433 0x86, 4, /* DW_CFA_offset, %esi, -16 */
4434 0x87, 5, /* DW_CFA_offset, %edi, -20 */
4439 #if defined(ELF_HOST_MACHINE)
4440 void tcg_register_jit(const void *buf, size_t buf_size)
4442 tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));