xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision b6d69fcefbd45ca33b896abfbc8e27e0f713bdf0)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
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:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
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
22 * THE SOFTWARE.
23 */
24
25/* Used for function call generation. */
26#define TCG_TARGET_STACK_ALIGN 16
27#if defined(_WIN64)
28#define TCG_TARGET_CALL_STACK_OFFSET 32
29#else
30#define TCG_TARGET_CALL_STACK_OFFSET 0
31#endif
32#define TCG_TARGET_CALL_ARG_I32      TCG_CALL_ARG_NORMAL
33#define TCG_TARGET_CALL_ARG_I64      TCG_CALL_ARG_NORMAL
34#if defined(_WIN64)
35# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_BY_REF
36# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_VEC
37#elif TCG_TARGET_REG_BITS == 64
38# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
39# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_NORMAL
40#else
41# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
42# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_REF
43#endif
44
45#ifdef CONFIG_DEBUG_TCG
46static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
47#if TCG_TARGET_REG_BITS == 64
48    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
49#else
50    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
51#endif
52    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
53    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
54#if TCG_TARGET_REG_BITS == 64
55    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
56    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
57#endif
58};
59#endif
60
61static const int tcg_target_reg_alloc_order[] = {
62#if TCG_TARGET_REG_BITS == 64
63    TCG_REG_RBP,
64    TCG_REG_RBX,
65    TCG_REG_R12,
66    TCG_REG_R13,
67    TCG_REG_R14,
68    TCG_REG_R15,
69    TCG_REG_R10,
70    TCG_REG_R11,
71    TCG_REG_R9,
72    TCG_REG_R8,
73    TCG_REG_RCX,
74    TCG_REG_RDX,
75    TCG_REG_RSI,
76    TCG_REG_RDI,
77    TCG_REG_RAX,
78#else
79    TCG_REG_EBX,
80    TCG_REG_ESI,
81    TCG_REG_EDI,
82    TCG_REG_EBP,
83    TCG_REG_ECX,
84    TCG_REG_EDX,
85    TCG_REG_EAX,
86#endif
87    TCG_REG_XMM0,
88    TCG_REG_XMM1,
89    TCG_REG_XMM2,
90    TCG_REG_XMM3,
91    TCG_REG_XMM4,
92    TCG_REG_XMM5,
93#ifndef _WIN64
94    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
95       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
96    TCG_REG_XMM6,
97    TCG_REG_XMM7,
98#if TCG_TARGET_REG_BITS == 64
99    TCG_REG_XMM8,
100    TCG_REG_XMM9,
101    TCG_REG_XMM10,
102    TCG_REG_XMM11,
103    TCG_REG_XMM12,
104    TCG_REG_XMM13,
105    TCG_REG_XMM14,
106    TCG_REG_XMM15,
107#endif
108#endif
109};
110
111#define TCG_TMP_VEC  TCG_REG_XMM5
112
113static const int tcg_target_call_iarg_regs[] = {
114#if TCG_TARGET_REG_BITS == 64
115#if defined(_WIN64)
116    TCG_REG_RCX,
117    TCG_REG_RDX,
118#else
119    TCG_REG_RDI,
120    TCG_REG_RSI,
121    TCG_REG_RDX,
122    TCG_REG_RCX,
123#endif
124    TCG_REG_R8,
125    TCG_REG_R9,
126#else
127    /* 32 bit mode uses stack based calling convention (GCC default). */
128#endif
129};
130
131static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
132{
133    switch (kind) {
134    case TCG_CALL_RET_NORMAL:
135        tcg_debug_assert(slot >= 0 && slot <= 1);
136        return slot ? TCG_REG_EDX : TCG_REG_EAX;
137#ifdef _WIN64
138    case TCG_CALL_RET_BY_VEC:
139        tcg_debug_assert(slot == 0);
140        return TCG_REG_XMM0;
141#endif
142    default:
143        g_assert_not_reached();
144    }
145}
146
147/* Constants we accept.  */
148#define TCG_CT_CONST_S32 0x100
149#define TCG_CT_CONST_U32 0x200
150#define TCG_CT_CONST_I32 0x400
151#define TCG_CT_CONST_WSZ 0x800
152#define TCG_CT_CONST_TST 0x1000
153#define TCG_CT_CONST_ZERO 0x2000
154
155/* Registers used with L constraint, which are the first argument
156   registers on x86_64, and two random call clobbered registers on
157   i386. */
158#if TCG_TARGET_REG_BITS == 64
159# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
160# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
161#else
162# define TCG_REG_L0 TCG_REG_EAX
163# define TCG_REG_L1 TCG_REG_EDX
164#endif
165
166#if TCG_TARGET_REG_BITS == 64
167# define ALL_GENERAL_REGS      0x0000ffffu
168# define ALL_VECTOR_REGS       0xffff0000u
169# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
170#else
171# define ALL_GENERAL_REGS      0x000000ffu
172# define ALL_VECTOR_REGS       0x00ff0000u
173# define ALL_BYTEL_REGS        0x0000000fu
174#endif
175#define SOFTMMU_RESERVE_REGS \
176    (tcg_use_softmmu ? (1 << TCG_REG_L0) | (1 << TCG_REG_L1) : 0)
177
178#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
179#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
180
181static const tcg_insn_unit *tb_ret_addr;
182
183static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
184                        intptr_t value, intptr_t addend)
185{
186    value += addend;
187    switch(type) {
188    case R_386_PC32:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int32_t)value) {
191            return false;
192        }
193        /* FALLTHRU */
194    case R_386_32:
195        tcg_patch32(code_ptr, value);
196        break;
197    case R_386_PC8:
198        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
199        if (value != (int8_t)value) {
200            return false;
201        }
202        tcg_patch8(code_ptr, value);
203        break;
204    default:
205        g_assert_not_reached();
206    }
207    return true;
208}
209
210/* test if a constant matches the constraint */
211static bool tcg_target_const_match(int64_t val, int ct,
212                                   TCGType type, TCGCond cond, int vece)
213{
214    if (ct & TCG_CT_CONST) {
215        return 1;
216    }
217    if (type == TCG_TYPE_I32) {
218        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 |
219                  TCG_CT_CONST_I32 | TCG_CT_CONST_TST)) {
220            return 1;
221        }
222    } else {
223        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
224            return 1;
225        }
226        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
227            return 1;
228        }
229        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
230            return 1;
231        }
232        /*
233         * This will be used in combination with TCG_CT_CONST_S32,
234         * so "normal" TESTQ is already matched.  Also accept:
235         *    TESTQ -> TESTL   (uint32_t)
236         *    TESTQ -> BT      (is_power_of_2)
237         */
238        if ((ct & TCG_CT_CONST_TST)
239            && is_tst_cond(cond)
240            && (val == (uint32_t)val || is_power_of_2(val))) {
241            return 1;
242        }
243    }
244    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
245        return 1;
246    }
247    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
248        return 1;
249    }
250    return 0;
251}
252
253# define LOWREGMASK(x)	((x) & 7)
254
255#define P_EXT		0x100		/* 0x0f opcode prefix */
256#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
257#define P_DATA16        0x400           /* 0x66 opcode prefix */
258#define P_VEXW          0x1000          /* Set VEX.W = 1 */
259#if TCG_TARGET_REG_BITS == 64
260# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
261# define P_REXB_R       0x2000          /* REG field as byte register */
262# define P_REXB_RM      0x4000          /* R/M field as byte register */
263# define P_GS           0x8000          /* gs segment override */
264#else
265# define P_REXW		0
266# define P_REXB_R	0
267# define P_REXB_RM	0
268# define P_GS           0
269#endif
270#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
271#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
272#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
273#define P_VEXL          0x80000         /* Set VEX.L = 1 */
274#define P_EVEX          0x100000        /* Requires EVEX encoding */
275
276#define OPC_ARITH_EbIb	(0x80)
277#define OPC_ARITH_EvIz	(0x81)
278#define OPC_ARITH_EvIb	(0x83)
279#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
280#define OPC_ANDN        (0xf2 | P_EXT38)
281#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
282#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
283#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
284#define OPC_BSF         (0xbc | P_EXT)
285#define OPC_BSR         (0xbd | P_EXT)
286#define OPC_BSWAP	(0xc8 | P_EXT)
287#define OPC_CALL_Jz	(0xe8)
288#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
289#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
290#define OPC_DEC_r32	(0x48)
291#define OPC_IMUL_GvEv	(0xaf | P_EXT)
292#define OPC_IMUL_GvEvIb	(0x6b)
293#define OPC_IMUL_GvEvIz	(0x69)
294#define OPC_INC_r32	(0x40)
295#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
296#define OPC_JCC_short	(0x70)		/* ... plus condition code */
297#define OPC_JMP_long	(0xe9)
298#define OPC_JMP_short	(0xeb)
299#define OPC_LEA         (0x8d)
300#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
301#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
302#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
303#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
304#define OPC_MOVB_EvIz   (0xc6)
305#define OPC_MOVL_EvIz	(0xc7)
306#define OPC_MOVB_Ib     (0xb0)
307#define OPC_MOVL_Iv     (0xb8)
308#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
309#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
310#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
311#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
312#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
313#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
314#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
315#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
316#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
317#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
318#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
319#define OPC_MOVSBL	(0xbe | P_EXT)
320#define OPC_MOVSWL	(0xbf | P_EXT)
321#define OPC_MOVSLQ	(0x63 | P_REXW)
322#define OPC_MOVZBL	(0xb6 | P_EXT)
323#define OPC_MOVZWL	(0xb7 | P_EXT)
324#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
325#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
326#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
327#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
328#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
329#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
330#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
331#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
332#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
333#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
334#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
335#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
336#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
337#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
338#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
339#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
340#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
341#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
342#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
343#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
344#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
345#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
346#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
347#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
348#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
349#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
350#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
351#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
352#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
353#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
354#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
355#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
356#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
357#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
358#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
359#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
360#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
361#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
362#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
363#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
364#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
366#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
367#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
368#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
369#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
370#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
371#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
372#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
373#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
374#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
375#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
376#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
377#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
378#define OPC_POR         (0xeb | P_EXT | P_DATA16)
379#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
380#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
381#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
382#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
383#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
384#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
385#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
386#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
387#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
388#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
389#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
390#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
391#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
392#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
393#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
394#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
395#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
396#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
397#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
398#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
399#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
400#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
401#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
402#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
403#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
404#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
405#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
406#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
407#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
408#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
409#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
410#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
411#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
412#define OPC_POP_r32	(0x58)
413#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
414#define OPC_PUSH_r32	(0x50)
415#define OPC_PUSH_Iv	(0x68)
416#define OPC_PUSH_Ib	(0x6a)
417#define OPC_RET		(0xc3)
418#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
419#define OPC_SHIFT_1	(0xd1)
420#define OPC_SHIFT_Ib	(0xc1)
421#define OPC_SHIFT_cl	(0xd3)
422#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
423#define OPC_SHUFPS      (0xc6 | P_EXT)
424#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
425#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
426#define OPC_SHRD_Ib     (0xac | P_EXT)
427#define OPC_TESTB	(0x84)
428#define OPC_TESTL	(0x85)
429#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
430#define OPC_UD2         (0x0b | P_EXT)
431#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
432#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
433#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
434#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
436#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
437#define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
438#define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
439#define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPCMPUW     (0x3e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
441#define OPC_VPCMPD      (0x1f | P_EXT3A | P_DATA16 | P_EVEX)
442#define OPC_VPCMPUD     (0x1e | P_EXT3A | P_DATA16 | P_EVEX)
443#define OPC_VPCMPQ      (0x1f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
444#define OPC_VPCMPUQ     (0x1e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
446#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
447#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
448#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
449#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
450#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
451#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
452#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
453#define OPC_VPMOVM2B    (0x28 | P_EXT38 | P_SIMDF3 | P_EVEX)
454#define OPC_VPMOVM2W    (0x28 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
455#define OPC_VPMOVM2D    (0x38 | P_EXT38 | P_SIMDF3 | P_EVEX)
456#define OPC_VPMOVM2Q    (0x38 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
457#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
458#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
459#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
460#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
461#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
462#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
463#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
464#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
465#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
466#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
467#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
468#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
469#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
470#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
471#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
472#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
473#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
474#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
475#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
476#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
477#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
478#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
479#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
480#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
481#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
482#define OPC_VPTESTMB    (0x26 | P_EXT38 | P_DATA16 | P_EVEX)
483#define OPC_VPTESTMW    (0x26 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
484#define OPC_VPTESTMD    (0x27 | P_EXT38 | P_DATA16 | P_EVEX)
485#define OPC_VPTESTMQ    (0x27 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
486#define OPC_VPTESTNMB   (0x26 | P_EXT38 | P_SIMDF3 | P_EVEX)
487#define OPC_VPTESTNMW   (0x26 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
488#define OPC_VPTESTNMD   (0x27 | P_EXT38 | P_SIMDF3 | P_EVEX)
489#define OPC_VPTESTNMQ   (0x27 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
490#define OPC_VZEROUPPER  (0x77 | P_EXT)
491#define OPC_XCHG_ax_r32	(0x90)
492#define OPC_XCHG_EvGv   (0x87)
493
494#define OPC_GRP3_Eb     (0xf6)
495#define OPC_GRP3_Ev     (0xf7)
496#define OPC_GRP5        (0xff)
497#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
498#define OPC_GRPBT       (0xba | P_EXT)
499
500#define OPC_GRPBT_BT    4
501#define OPC_GRPBT_BTS   5
502#define OPC_GRPBT_BTR   6
503#define OPC_GRPBT_BTC   7
504
505/* Group 1 opcode extensions for 0x80-0x83.
506   These are also used as modifiers for OPC_ARITH.  */
507#define ARITH_ADD 0
508#define ARITH_OR  1
509#define ARITH_ADC 2
510#define ARITH_SBB 3
511#define ARITH_AND 4
512#define ARITH_SUB 5
513#define ARITH_XOR 6
514#define ARITH_CMP 7
515
516/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
517#define SHIFT_ROL 0
518#define SHIFT_ROR 1
519#define SHIFT_SHL 4
520#define SHIFT_SHR 5
521#define SHIFT_SAR 7
522
523/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
524#define EXT3_TESTi 0
525#define EXT3_NOT   2
526#define EXT3_NEG   3
527#define EXT3_MUL   4
528#define EXT3_IMUL  5
529#define EXT3_DIV   6
530#define EXT3_IDIV  7
531
532/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
533#define EXT5_INC_Ev	0
534#define EXT5_DEC_Ev	1
535#define EXT5_CALLN_Ev	2
536#define EXT5_JMPN_Ev	4
537
538/* Condition codes to be added to OPC_JCC_{long,short}.  */
539#define JCC_JMP (-1)
540#define JCC_JO  0x0
541#define JCC_JNO 0x1
542#define JCC_JB  0x2
543#define JCC_JAE 0x3
544#define JCC_JE  0x4
545#define JCC_JNE 0x5
546#define JCC_JBE 0x6
547#define JCC_JA  0x7
548#define JCC_JS  0x8
549#define JCC_JNS 0x9
550#define JCC_JP  0xa
551#define JCC_JNP 0xb
552#define JCC_JL  0xc
553#define JCC_JGE 0xd
554#define JCC_JLE 0xe
555#define JCC_JG  0xf
556
557static const uint8_t tcg_cond_to_jcc[] = {
558    [TCG_COND_EQ] = JCC_JE,
559    [TCG_COND_NE] = JCC_JNE,
560    [TCG_COND_LT] = JCC_JL,
561    [TCG_COND_GE] = JCC_JGE,
562    [TCG_COND_LE] = JCC_JLE,
563    [TCG_COND_GT] = JCC_JG,
564    [TCG_COND_LTU] = JCC_JB,
565    [TCG_COND_GEU] = JCC_JAE,
566    [TCG_COND_LEU] = JCC_JBE,
567    [TCG_COND_GTU] = JCC_JA,
568    [TCG_COND_TSTEQ] = JCC_JE,
569    [TCG_COND_TSTNE] = JCC_JNE,
570};
571
572#if TCG_TARGET_REG_BITS == 64
573static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
574{
575    int rex;
576
577    if (opc & P_GS) {
578        tcg_out8(s, 0x65);
579    }
580    if (opc & P_DATA16) {
581        /* We should never be asking for both 16 and 64-bit operation.  */
582        tcg_debug_assert((opc & P_REXW) == 0);
583        tcg_out8(s, 0x66);
584    }
585    if (opc & P_SIMDF3) {
586        tcg_out8(s, 0xf3);
587    } else if (opc & P_SIMDF2) {
588        tcg_out8(s, 0xf2);
589    }
590
591    rex = 0;
592    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
593    rex |= (r & 8) >> 1;                /* REX.R */
594    rex |= (x & 8) >> 2;                /* REX.X */
595    rex |= (rm & 8) >> 3;               /* REX.B */
596
597    /* P_REXB_{R,RM} indicates that the given register is the low byte.
598       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
599       as otherwise the encoding indicates %[abcd]h.  Note that the values
600       that are ORed in merely indicate that the REX byte must be present;
601       those bits get discarded in output.  */
602    rex |= opc & (r >= 4 ? P_REXB_R : 0);
603    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
604
605    if (rex) {
606        tcg_out8(s, (uint8_t)(rex | 0x40));
607    }
608
609    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
610        tcg_out8(s, 0x0f);
611        if (opc & P_EXT38) {
612            tcg_out8(s, 0x38);
613        } else if (opc & P_EXT3A) {
614            tcg_out8(s, 0x3a);
615        }
616    }
617
618    tcg_out8(s, opc);
619}
620#else
621static void tcg_out_opc(TCGContext *s, int opc)
622{
623    if (opc & P_DATA16) {
624        tcg_out8(s, 0x66);
625    }
626    if (opc & P_SIMDF3) {
627        tcg_out8(s, 0xf3);
628    } else if (opc & P_SIMDF2) {
629        tcg_out8(s, 0xf2);
630    }
631    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
632        tcg_out8(s, 0x0f);
633        if (opc & P_EXT38) {
634            tcg_out8(s, 0x38);
635        } else if (opc & P_EXT3A) {
636            tcg_out8(s, 0x3a);
637        }
638    }
639    tcg_out8(s, opc);
640}
641/* Discard the register arguments to tcg_out_opc early, so as not to penalize
642   the 32-bit compilation paths.  This method works with all versions of gcc,
643   whereas relying on optimization may not be able to exclude them.  */
644#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
645#endif
646
647static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
648{
649    tcg_out_opc(s, opc, r, rm, 0);
650    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
651}
652
653static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
654                            int rm, int index)
655{
656    int tmp;
657
658    if (opc & P_GS) {
659        tcg_out8(s, 0x65);
660    }
661    /* Use the two byte form if possible, which cannot encode
662       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
663    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
664        && ((rm | index) & 8) == 0) {
665        /* Two byte VEX prefix.  */
666        tcg_out8(s, 0xc5);
667
668        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
669    } else {
670        /* Three byte VEX prefix.  */
671        tcg_out8(s, 0xc4);
672
673        /* VEX.m-mmmm */
674        if (opc & P_EXT3A) {
675            tmp = 3;
676        } else if (opc & P_EXT38) {
677            tmp = 2;
678        } else if (opc & P_EXT) {
679            tmp = 1;
680        } else {
681            g_assert_not_reached();
682        }
683        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
684        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
685        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
686        tcg_out8(s, tmp);
687
688        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
689    }
690
691    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
692    /* VEX.pp */
693    if (opc & P_DATA16) {
694        tmp |= 1;                          /* 0x66 */
695    } else if (opc & P_SIMDF3) {
696        tmp |= 2;                          /* 0xf3 */
697    } else if (opc & P_SIMDF2) {
698        tmp |= 3;                          /* 0xf2 */
699    }
700    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
701    tcg_out8(s, tmp);
702    tcg_out8(s, opc);
703}
704
705static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
706                             int rm, int index, int aaa, bool z)
707{
708    /* The entire 4-byte evex prefix; with R' and V' set. */
709    uint32_t p = 0x08041062;
710    int mm, pp;
711
712    tcg_debug_assert(have_avx512vl);
713
714    /* EVEX.mm */
715    if (opc & P_EXT3A) {
716        mm = 3;
717    } else if (opc & P_EXT38) {
718        mm = 2;
719    } else if (opc & P_EXT) {
720        mm = 1;
721    } else {
722        g_assert_not_reached();
723    }
724
725    /* EVEX.pp */
726    if (opc & P_DATA16) {
727        pp = 1;                          /* 0x66 */
728    } else if (opc & P_SIMDF3) {
729        pp = 2;                          /* 0xf3 */
730    } else if (opc & P_SIMDF2) {
731        pp = 3;                          /* 0xf2 */
732    } else {
733        pp = 0;
734    }
735
736    p = deposit32(p, 8, 2, mm);
737    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
738    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
739    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
740    p = deposit32(p, 16, 2, pp);
741    p = deposit32(p, 19, 4, ~v);
742    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
743    p = deposit32(p, 24, 3, aaa);
744    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
745    p = deposit32(p, 31, 1, z);
746
747    tcg_out32(s, p);
748    tcg_out8(s, opc);
749}
750
751static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
752{
753    if (opc & P_EVEX) {
754        tcg_out_evex_opc(s, opc, r, v, rm, 0, 0, false);
755    } else {
756        tcg_out_vex_opc(s, opc, r, v, rm, 0);
757    }
758    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
759}
760
761static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
762                                   int r, int v, int rm, TCGType type)
763{
764    if (type == TCG_TYPE_V256) {
765        opc |= P_VEXL;
766    }
767    tcg_out_vex_modrm(s, opc, r, v, rm);
768}
769
770static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
771                                    int rm, int aaa, bool z, TCGType type)
772{
773    if (type == TCG_TYPE_V256) {
774        opc |= P_VEXL;
775    }
776    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
777    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
778}
779
780/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
781   We handle either RM and INDEX missing with a negative value.  In 64-bit
782   mode for absolute addresses, ~RM is the size of the immediate operand
783   that will follow the instruction.  */
784
785static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
786                               int shift, intptr_t offset)
787{
788    int mod, len;
789
790    if (index < 0 && rm < 0) {
791        if (TCG_TARGET_REG_BITS == 64) {
792            /* Try for a rip-relative addressing mode.  This has replaced
793               the 32-bit-mode absolute addressing encoding.  */
794            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
795            intptr_t disp = offset - pc;
796            if (disp == (int32_t)disp) {
797                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
798                tcg_out32(s, disp);
799                return;
800            }
801
802            /* Try for an absolute address encoding.  This requires the
803               use of the MODRM+SIB encoding and is therefore larger than
804               rip-relative addressing.  */
805            if (offset == (int32_t)offset) {
806                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
807                tcg_out8(s, (4 << 3) | 5);
808                tcg_out32(s, offset);
809                return;
810            }
811
812            /* ??? The memory isn't directly addressable.  */
813            g_assert_not_reached();
814        } else {
815            /* Absolute address.  */
816            tcg_out8(s, (r << 3) | 5);
817            tcg_out32(s, offset);
818            return;
819        }
820    }
821
822    /* Find the length of the immediate addend.  Note that the encoding
823       that would be used for (%ebp) indicates absolute addressing.  */
824    if (rm < 0) {
825        mod = 0, len = 4, rm = 5;
826    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
827        mod = 0, len = 0;
828    } else if (offset == (int8_t)offset) {
829        mod = 0x40, len = 1;
830    } else {
831        mod = 0x80, len = 4;
832    }
833
834    /* Use a single byte MODRM format if possible.  Note that the encoding
835       that would be used for %esp is the escape to the two byte form.  */
836    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
837        /* Single byte MODRM format.  */
838        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
839    } else {
840        /* Two byte MODRM+SIB format.  */
841
842        /* Note that the encoding that would place %esp into the index
843           field indicates no index register.  In 64-bit mode, the REX.X
844           bit counts, so %r12 can be used as the index.  */
845        if (index < 0) {
846            index = 4;
847        } else {
848            tcg_debug_assert(index != TCG_REG_ESP);
849        }
850
851        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
852        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
853    }
854
855    if (len == 1) {
856        tcg_out8(s, offset);
857    } else if (len == 4) {
858        tcg_out32(s, offset);
859    }
860}
861
862static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
863                                     int index, int shift, intptr_t offset)
864{
865    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
866    tcg_out_sib_offset(s, r, rm, index, shift, offset);
867}
868
869static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
870                                         int rm, int index, int shift,
871                                         intptr_t offset)
872{
873    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
874    tcg_out_sib_offset(s, r, rm, index, shift, offset);
875}
876
877/* A simplification of the above with no index or shift.  */
878static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
879                                        int rm, intptr_t offset)
880{
881    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
882}
883
884static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
885                                            int v, int rm, intptr_t offset)
886{
887    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
888}
889
890/* Output an opcode with an expected reference to the constant pool.  */
891static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
892{
893    tcg_out_opc(s, opc, r, 0, 0);
894    /* Absolute for 32-bit, pc-relative for 64-bit.  */
895    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
896    tcg_out32(s, 0);
897}
898
899/* Output an opcode with an expected reference to the constant pool.  */
900static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
901{
902    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
903    /* Absolute for 32-bit, pc-relative for 64-bit.  */
904    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
905    tcg_out32(s, 0);
906}
907
908/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
909static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
910{
911    /* Propagate an opcode prefix, such as P_REXW.  */
912    int ext = subop & ~0x7;
913    subop &= 0x7;
914
915    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
916}
917
918static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
919{
920    int rexw = 0;
921
922    if (arg == ret) {
923        return true;
924    }
925    switch (type) {
926    case TCG_TYPE_I64:
927        rexw = P_REXW;
928        /* fallthru */
929    case TCG_TYPE_I32:
930        if (ret < 16) {
931            if (arg < 16) {
932                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
933            } else {
934                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
935            }
936        } else {
937            if (arg < 16) {
938                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
939            } else {
940                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
941            }
942        }
943        break;
944
945    case TCG_TYPE_V64:
946        tcg_debug_assert(ret >= 16 && arg >= 16);
947        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
948        break;
949    case TCG_TYPE_V128:
950        tcg_debug_assert(ret >= 16 && arg >= 16);
951        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
952        break;
953    case TCG_TYPE_V256:
954        tcg_debug_assert(ret >= 16 && arg >= 16);
955        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
956        break;
957
958    default:
959        g_assert_not_reached();
960    }
961    return true;
962}
963
964static const int avx2_dup_insn[4] = {
965    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
966    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
967};
968
969static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
970                            TCGReg r, TCGReg a)
971{
972    if (have_avx2) {
973        tcg_out_vex_modrm_type(s, avx2_dup_insn[vece], r, 0, a, type);
974    } else {
975        switch (vece) {
976        case MO_8:
977            /* ??? With zero in a register, use PSHUFB.  */
978            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
979            a = r;
980            /* FALLTHRU */
981        case MO_16:
982            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
983            a = r;
984            /* FALLTHRU */
985        case MO_32:
986            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
987            /* imm8 operand: all output lanes selected from input lane 0.  */
988            tcg_out8(s, 0);
989            break;
990        case MO_64:
991            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
992            break;
993        default:
994            g_assert_not_reached();
995        }
996    }
997    return true;
998}
999
1000static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1001                             TCGReg r, TCGReg base, intptr_t offset)
1002{
1003    if (have_avx2) {
1004        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1005        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
1006                                 r, 0, base, offset);
1007    } else {
1008        switch (vece) {
1009        case MO_64:
1010            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
1011            break;
1012        case MO_32:
1013            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
1014            break;
1015        case MO_16:
1016            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
1017            tcg_out8(s, 0); /* imm8 */
1018            tcg_out_dup_vec(s, type, vece, r, r);
1019            break;
1020        case MO_8:
1021            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
1022            tcg_out8(s, 0); /* imm8 */
1023            tcg_out_dup_vec(s, type, vece, r, r);
1024            break;
1025        default:
1026            g_assert_not_reached();
1027        }
1028    }
1029    return true;
1030}
1031
1032static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1033                             TCGReg ret, int64_t arg)
1034{
1035    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1036
1037    if (arg == 0) {
1038        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1039        return;
1040    }
1041    if (arg == -1) {
1042        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
1043        return;
1044    }
1045
1046    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
1047        if (have_avx2) {
1048            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
1049        } else {
1050            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
1051        }
1052        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1053    } else {
1054        if (type == TCG_TYPE_V64) {
1055            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
1056        } else if (have_avx2) {
1057            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1058        } else {
1059            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1060        }
1061        if (TCG_TARGET_REG_BITS == 64) {
1062            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1063        } else {
1064            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1065        }
1066    }
1067}
1068
1069static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1070                             TCGReg ret, tcg_target_long arg)
1071{
1072    if (arg == 0) {
1073        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1074        return;
1075    }
1076    if (arg == -1) {
1077        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1078        return;
1079    }
1080
1081    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1082    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1083    if (TCG_TARGET_REG_BITS == 64) {
1084        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1085    } else {
1086        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1087    }
1088}
1089
1090static void tcg_out_movi_int(TCGContext *s, TCGType type,
1091                             TCGReg ret, tcg_target_long arg)
1092{
1093    tcg_target_long diff;
1094
1095    if (arg == 0) {
1096        tgen_arithr(s, ARITH_XOR, ret, ret);
1097        return;
1098    }
1099    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1100        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1101        tcg_out32(s, arg);
1102        return;
1103    }
1104    if (arg == (int32_t)arg) {
1105        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1106        tcg_out32(s, arg);
1107        return;
1108    }
1109
1110    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1111    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1112    if (diff == (int32_t)diff) {
1113        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1114        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1115        tcg_out32(s, diff);
1116        return;
1117    }
1118
1119    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1120    tcg_out64(s, arg);
1121}
1122
1123static void tcg_out_movi(TCGContext *s, TCGType type,
1124                         TCGReg ret, tcg_target_long arg)
1125{
1126    switch (type) {
1127    case TCG_TYPE_I32:
1128#if TCG_TARGET_REG_BITS == 64
1129    case TCG_TYPE_I64:
1130#endif
1131        if (ret < 16) {
1132            tcg_out_movi_int(s, type, ret, arg);
1133        } else {
1134            tcg_out_movi_vec(s, type, ret, arg);
1135        }
1136        break;
1137    default:
1138        g_assert_not_reached();
1139    }
1140}
1141
1142static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1143{
1144    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1145    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1146    return true;
1147}
1148
1149static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1150                             tcg_target_long imm)
1151{
1152    /* This function is only used for passing structs by reference. */
1153    tcg_debug_assert(imm == (int32_t)imm);
1154    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1155}
1156
1157static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1158{
1159    if (val == (int8_t)val) {
1160        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1161        tcg_out8(s, val);
1162    } else if (val == (int32_t)val) {
1163        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1164        tcg_out32(s, val);
1165    } else {
1166        g_assert_not_reached();
1167    }
1168}
1169
1170static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1171{
1172    /* Given the strength of x86 memory ordering, we only need care for
1173       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1174       faster than "mfence", so don't bother with the sse insn.  */
1175    if (a0 & TCG_MO_ST_LD) {
1176        tcg_out8(s, 0xf0);
1177        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1178        tcg_out8(s, 0);
1179    }
1180}
1181
1182static inline void tcg_out_push(TCGContext *s, int reg)
1183{
1184    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1185}
1186
1187static inline void tcg_out_pop(TCGContext *s, int reg)
1188{
1189    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1190}
1191
1192static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1193                       TCGReg arg1, intptr_t arg2)
1194{
1195    switch (type) {
1196    case TCG_TYPE_I32:
1197        if (ret < 16) {
1198            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1199        } else {
1200            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1201        }
1202        break;
1203    case TCG_TYPE_I64:
1204        if (ret < 16) {
1205            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1206            break;
1207        }
1208        /* FALLTHRU */
1209    case TCG_TYPE_V64:
1210        /* There is no instruction that can validate 8-byte alignment.  */
1211        tcg_debug_assert(ret >= 16);
1212        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1213        break;
1214    case TCG_TYPE_V128:
1215        /*
1216         * The gvec infrastructure is asserts that v128 vector loads
1217         * and stores use a 16-byte aligned offset.  Validate that the
1218         * final pointer is aligned by using an insn that will SIGSEGV.
1219         */
1220        tcg_debug_assert(ret >= 16);
1221        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1222        break;
1223    case TCG_TYPE_V256:
1224        /*
1225         * The gvec infrastructure only requires 16-byte alignment,
1226         * so here we must use an unaligned load.
1227         */
1228        tcg_debug_assert(ret >= 16);
1229        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1230                                 ret, 0, arg1, arg2);
1231        break;
1232    default:
1233        g_assert_not_reached();
1234    }
1235}
1236
1237static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1238                       TCGReg arg1, intptr_t arg2)
1239{
1240    switch (type) {
1241    case TCG_TYPE_I32:
1242        if (arg < 16) {
1243            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1244        } else {
1245            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1246        }
1247        break;
1248    case TCG_TYPE_I64:
1249        if (arg < 16) {
1250            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1251            break;
1252        }
1253        /* FALLTHRU */
1254    case TCG_TYPE_V64:
1255        /* There is no instruction that can validate 8-byte alignment.  */
1256        tcg_debug_assert(arg >= 16);
1257        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1258        break;
1259    case TCG_TYPE_V128:
1260        /*
1261         * The gvec infrastructure is asserts that v128 vector loads
1262         * and stores use a 16-byte aligned offset.  Validate that the
1263         * final pointer is aligned by using an insn that will SIGSEGV.
1264         *
1265         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1266         * for _WIN64, which must have SSE2 but may not have AVX.
1267         */
1268        tcg_debug_assert(arg >= 16);
1269        if (have_avx1) {
1270            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1271        } else {
1272            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1273        }
1274        break;
1275    case TCG_TYPE_V256:
1276        /*
1277         * The gvec infrastructure only requires 16-byte alignment,
1278         * so here we must use an unaligned store.
1279         */
1280        tcg_debug_assert(arg >= 16);
1281        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1282                                 arg, 0, arg1, arg2);
1283        break;
1284    default:
1285        g_assert_not_reached();
1286    }
1287}
1288
1289static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1290                        TCGReg base, intptr_t ofs)
1291{
1292    int rexw = 0;
1293    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1294        if (val != (int32_t)val) {
1295            return false;
1296        }
1297        rexw = P_REXW;
1298    } else if (type != TCG_TYPE_I32) {
1299        return false;
1300    }
1301    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1302    tcg_out32(s, val);
1303    return true;
1304}
1305
1306static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1307{
1308    /* Propagate an opcode prefix, such as P_DATA16.  */
1309    int ext = subopc & ~0x7;
1310    subopc &= 0x7;
1311
1312    if (count == 1) {
1313        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1314    } else {
1315        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1316        tcg_out8(s, count);
1317    }
1318}
1319
1320static inline void tcg_out_bswap32(TCGContext *s, int reg)
1321{
1322    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1323}
1324
1325static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1326{
1327    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1328}
1329
1330static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1331{
1332    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1333        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1334        if (dest >= 4) {
1335            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
1336            tcg_out32(s, 0xff);
1337            return;
1338        }
1339        src = dest;
1340    }
1341    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1342}
1343
1344static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1345{
1346    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1347
1348    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1349        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1350        if (dest >= 4) {
1351            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
1352            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
1353            return;
1354        }
1355        src = dest;
1356    }
1357    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1358}
1359
1360static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1361{
1362    /* movzwl */
1363    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1364}
1365
1366static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1367{
1368    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1369    /* movsw[lq] */
1370    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1371}
1372
1373static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1374{
1375    /* 32-bit mov zero extends.  */
1376    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1377}
1378
1379static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1380{
1381    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1382    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1383}
1384
1385static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1386{
1387    tcg_out_ext32s(s, dest, src);
1388}
1389
1390static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1391{
1392    if (dest != src) {
1393        tcg_out_ext32u(s, dest, src);
1394    }
1395}
1396
1397static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1398{
1399    tcg_out_ext32u(s, dest, src);
1400}
1401
1402static inline void tcg_out_bswap64(TCGContext *s, int reg)
1403{
1404    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1405}
1406
1407static void tgen_arithi(TCGContext *s, int c, int r0,
1408                        tcg_target_long val, int cf)
1409{
1410    int rexw = 0;
1411
1412    if (TCG_TARGET_REG_BITS == 64) {
1413        rexw = c & -8;
1414        c &= 7;
1415    }
1416
1417    switch (c) {
1418    case ARITH_ADD:
1419    case ARITH_SUB:
1420        if (!cf) {
1421            /*
1422             * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1423             * partial flags update stalls on Pentium4 and are not recommended
1424             * by current Intel optimization manuals.
1425             */
1426            if (val == 1 || val == -1) {
1427                int is_inc = (c == ARITH_ADD) ^ (val < 0);
1428                if (TCG_TARGET_REG_BITS == 64) {
1429                    /*
1430                     * The single-byte increment encodings are re-tasked
1431                     * as the REX prefixes.  Use the MODRM encoding.
1432                     */
1433                    tcg_out_modrm(s, OPC_GRP5 + rexw,
1434                                  (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1435                } else {
1436                    tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1437                }
1438                return;
1439            }
1440            if (val == 128) {
1441                /*
1442                 * Facilitate using an 8-bit immediate.  Carry is inverted
1443                 * by this transformation, so do it only if cf == 0.
1444                 */
1445                c ^= ARITH_ADD ^ ARITH_SUB;
1446                val = -128;
1447            }
1448        }
1449        break;
1450
1451    case ARITH_AND:
1452        if (TCG_TARGET_REG_BITS == 64) {
1453            if (val == 0xffffffffu) {
1454                tcg_out_ext32u(s, r0, r0);
1455                return;
1456            }
1457            if (val == (uint32_t)val) {
1458                /* AND with no high bits set can use a 32-bit operation.  */
1459                rexw = 0;
1460            }
1461        }
1462        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1463            tcg_out_ext8u(s, r0, r0);
1464            return;
1465        }
1466        if (val == 0xffffu) {
1467            tcg_out_ext16u(s, r0, r0);
1468            return;
1469        }
1470        break;
1471
1472    case ARITH_OR:
1473    case ARITH_XOR:
1474        if (val >= 0x80 && val <= 0xff
1475            && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1476            tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1477            tcg_out8(s, val);
1478            return;
1479        }
1480        break;
1481    }
1482
1483    if (val == (int8_t)val) {
1484        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1485        tcg_out8(s, val);
1486        return;
1487    }
1488    if (rexw == 0 || val == (int32_t)val) {
1489        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1490        tcg_out32(s, val);
1491        return;
1492    }
1493
1494    g_assert_not_reached();
1495}
1496
1497static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1498{
1499    if (val != 0) {
1500        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1501    }
1502}
1503
1504/* Set SMALL to force a short forward branch.  */
1505static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1506{
1507    int32_t val, val1;
1508
1509    if (l->has_value) {
1510        val = tcg_pcrel_diff(s, l->u.value_ptr);
1511        val1 = val - 2;
1512        if ((int8_t)val1 == val1) {
1513            if (opc == -1) {
1514                tcg_out8(s, OPC_JMP_short);
1515            } else {
1516                tcg_out8(s, OPC_JCC_short + opc);
1517            }
1518            tcg_out8(s, val1);
1519        } else {
1520            tcg_debug_assert(!small);
1521            if (opc == -1) {
1522                tcg_out8(s, OPC_JMP_long);
1523                tcg_out32(s, val - 5);
1524            } else {
1525                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1526                tcg_out32(s, val - 6);
1527            }
1528        }
1529    } else if (small) {
1530        if (opc == -1) {
1531            tcg_out8(s, OPC_JMP_short);
1532        } else {
1533            tcg_out8(s, OPC_JCC_short + opc);
1534        }
1535        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1536        s->code_ptr += 1;
1537    } else {
1538        if (opc == -1) {
1539            tcg_out8(s, OPC_JMP_long);
1540        } else {
1541            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1542        }
1543        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1544        s->code_ptr += 4;
1545    }
1546}
1547
1548static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1549                       TCGArg arg2, int const_arg2, int rexw)
1550{
1551    int jz, js;
1552
1553    if (!is_tst_cond(cond)) {
1554        if (!const_arg2) {
1555            tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1556        } else if (arg2 == 0) {
1557            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1558        } else {
1559            tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1560            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1561        }
1562        return tcg_cond_to_jcc[cond];
1563    }
1564
1565    jz = tcg_cond_to_jcc[cond];
1566    js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1567
1568    if (!const_arg2) {
1569        tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1570        return jz;
1571    }
1572
1573    if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1574        if (arg2 == 0x80) {
1575            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1576            return js;
1577        }
1578        if (arg2 == 0xff) {
1579            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1580            return jz;
1581        }
1582        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1583        tcg_out8(s, arg2);
1584        return jz;
1585    }
1586
1587    if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1588        if (arg2 == 0x8000) {
1589            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1590            return js;
1591        }
1592        if (arg2 == 0xff00) {
1593            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1594            return jz;
1595        }
1596        tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1597        tcg_out8(s, arg2 >> 8);
1598        return jz;
1599    }
1600
1601    if (arg2 == 0xffff) {
1602        tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1603        return jz;
1604    }
1605    if (arg2 == 0xffffffffu) {
1606        tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1607        return jz;
1608    }
1609
1610    if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1611        int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1612        int sh = ctz64(arg2);
1613
1614        rexw = (sh & 32 ? P_REXW : 0);
1615        if ((sh & 31) == 31) {
1616            tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1617            return js;
1618        } else {
1619            tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1620            tcg_out8(s, sh);
1621            return jc;
1622        }
1623    }
1624
1625    if (rexw) {
1626        if (arg2 == (uint32_t)arg2) {
1627            rexw = 0;
1628        } else {
1629            tcg_debug_assert(arg2 == (int32_t)arg2);
1630        }
1631    }
1632    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1633    tcg_out32(s, arg2);
1634    return jz;
1635}
1636
1637static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1638                           TCGArg arg1, TCGArg arg2, int const_arg2,
1639                           TCGLabel *label, bool small)
1640{
1641    int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1642    tcg_out_jxx(s, jcc, label, small);
1643}
1644
1645static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
1646                        TCGReg arg1, TCGReg arg2, TCGLabel *label)
1647{
1648    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1649    tcg_out_brcond(s, rexw, cond, arg1, arg2, false, label, false);
1650}
1651
1652static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
1653                         TCGReg arg1, tcg_target_long arg2, TCGLabel *label)
1654{
1655    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1656    tcg_out_brcond(s, rexw, cond, arg1, arg2, true, label, false);
1657}
1658
1659static const TCGOutOpBrcond outop_brcond = {
1660    .base.static_constraint = C_O0_I2(r, reT),
1661    .out_rr = tgen_brcond,
1662    .out_ri = tgen_brcondi,
1663};
1664
1665#if TCG_TARGET_REG_BITS == 32
1666static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1667                            const int *const_args, bool small)
1668{
1669    TCGLabel *label_next = gen_new_label();
1670    TCGLabel *label_this = arg_label(args[5]);
1671    TCGCond cond = args[4];
1672
1673    switch (cond) {
1674    case TCG_COND_EQ:
1675    case TCG_COND_TSTEQ:
1676        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1677                       args[0], args[2], const_args[2], label_next, 1);
1678        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1679                       label_this, small);
1680        break;
1681
1682    case TCG_COND_NE:
1683    case TCG_COND_TSTNE:
1684        tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1685                       label_this, small);
1686        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1687                       label_this, small);
1688        break;
1689
1690    default:
1691        tcg_out_brcond(s, 0, tcg_high_cond(cond), args[1],
1692                       args[3], const_args[3], label_this, small);
1693        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1694        tcg_out_brcond(s, 0, tcg_unsigned_cond(cond), args[0],
1695                       args[2], const_args[2], label_this, small);
1696        break;
1697    }
1698    tcg_out_label(s, label_next);
1699}
1700#endif
1701
1702static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond,
1703                            TCGReg dest, TCGReg arg1, TCGArg arg2,
1704                            bool const_arg2, bool neg)
1705{
1706    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1707    int cmp_rexw = rexw;
1708    bool inv = false;
1709    bool cleared;
1710    int jcc;
1711
1712    switch (cond) {
1713    case TCG_COND_NE:
1714        inv = true;
1715        /* fall through */
1716    case TCG_COND_EQ:
1717        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1718        if (const_arg2 && arg2 == 0) {
1719            arg2 = 1;
1720            goto do_ltu;
1721        }
1722        break;
1723
1724    case TCG_COND_TSTNE:
1725        inv = true;
1726        /* fall through */
1727    case TCG_COND_TSTEQ:
1728        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1729        if (const_arg2 && arg2 == 0xffffffffu) {
1730            arg2 = 1;
1731            cmp_rexw = 0;
1732            goto do_ltu;
1733        }
1734        break;
1735
1736    case TCG_COND_LEU:
1737        inv = true;
1738        /* fall through */
1739    case TCG_COND_GTU:
1740        /* If arg2 is a register, swap for LTU/GEU. */
1741        if (!const_arg2) {
1742            TCGReg t = arg1;
1743            arg1 = arg2;
1744            arg2 = t;
1745            goto do_ltu;
1746        }
1747        break;
1748
1749    case TCG_COND_GEU:
1750        inv = true;
1751        /* fall through */
1752    case TCG_COND_LTU:
1753    do_ltu:
1754        /*
1755         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1756         * We can then use NEG or INC to produce the desired result.
1757         * This is always smaller than the SETCC expansion.
1758         */
1759        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1760
1761        /* X - X - C = -C = (C ? -1 : 0) */
1762        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1763        if (inv && neg) {
1764            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1765            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1766        } else if (inv) {
1767            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1768            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1769        } else if (!neg) {
1770            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1771            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1772        }
1773        return;
1774
1775    case TCG_COND_GE:
1776        inv = true;
1777        /* fall through */
1778    case TCG_COND_LT:
1779        /* If arg2 is 0, extract the sign bit. */
1780        if (const_arg2 && arg2 == 0) {
1781            tcg_out_mov(s, type, dest, arg1);
1782            if (inv) {
1783                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1784            }
1785            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1786                           dest, rexw ? 63 : 31);
1787            return;
1788        }
1789        break;
1790
1791    default:
1792        break;
1793    }
1794
1795    /*
1796     * If dest does not overlap the inputs, clearing it first is preferred.
1797     * The XOR breaks any false dependency for the low-byte write to dest,
1798     * and is also one byte smaller than MOVZBL.
1799     */
1800    cleared = false;
1801    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1802        tgen_arithr(s, ARITH_XOR, dest, dest);
1803        cleared = true;
1804    }
1805
1806    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1807    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1808
1809    if (!cleared) {
1810        tcg_out_ext8u(s, dest, dest);
1811    }
1812    if (neg) {
1813        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1814    }
1815}
1816
1817static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
1818                         TCGReg dest, TCGReg arg1, TCGReg arg2)
1819{
1820    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, false);
1821}
1822
1823static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
1824                          TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1825{
1826    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, false);
1827}
1828
1829static const TCGOutOpSetcond outop_setcond = {
1830    .base.static_constraint = C_O1_I2(q, r, reT),
1831    .out_rrr = tgen_setcond,
1832    .out_rri = tgen_setcondi,
1833};
1834
1835static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
1836                            TCGReg dest, TCGReg arg1, TCGReg arg2)
1837{
1838    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, true);
1839}
1840
1841static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
1842                             TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1843{
1844    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, true);
1845}
1846
1847static const TCGOutOpSetcond outop_negsetcond = {
1848    .base.static_constraint = C_O1_I2(q, r, reT),
1849    .out_rrr = tgen_negsetcond,
1850    .out_rri = tgen_negsetcondi,
1851};
1852
1853#if TCG_TARGET_REG_BITS == 32
1854static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1855                             const int *const_args)
1856{
1857    TCGArg new_args[6];
1858    TCGLabel *label_true, *label_over;
1859
1860    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1861
1862    if (args[0] == args[1] || args[0] == args[2]
1863        || (!const_args[3] && args[0] == args[3])
1864        || (!const_args[4] && args[0] == args[4])) {
1865        /* When the destination overlaps with one of the argument
1866           registers, don't do anything tricky.  */
1867        label_true = gen_new_label();
1868        label_over = gen_new_label();
1869
1870        new_args[5] = label_arg(label_true);
1871        tcg_out_brcond2(s, new_args, const_args+1, 1);
1872
1873        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1874        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1875        tcg_out_label(s, label_true);
1876
1877        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1878        tcg_out_label(s, label_over);
1879    } else {
1880        /* When the destination does not overlap one of the arguments,
1881           clear the destination first, jump if cond false, and emit an
1882           increment in the true case.  This results in smaller code.  */
1883
1884        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1885
1886        label_over = gen_new_label();
1887        new_args[4] = tcg_invert_cond(new_args[4]);
1888        new_args[5] = label_arg(label_over);
1889        tcg_out_brcond2(s, new_args, const_args+1, 1);
1890
1891        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1892        tcg_out_label(s, label_over);
1893    }
1894}
1895#endif
1896
1897static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1898                         TCGReg dest, TCGReg v1)
1899{
1900    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1901}
1902
1903static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1904                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1905                            TCGReg v1)
1906{
1907    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1908    tcg_out_cmov(s, jcc, rexw, dest, v1);
1909}
1910
1911static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1912{
1913    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1914
1915    if (disp == (int32_t)disp) {
1916        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1917        tcg_out32(s, disp);
1918    } else {
1919        /* rip-relative addressing into the constant pool.
1920           This is 6 + 8 = 14 bytes, as compared to using an
1921           immediate load 10 + 6 = 16 bytes, plus we may
1922           be able to re-use the pool constant for more calls.  */
1923        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1924        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1925        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1926        tcg_out32(s, 0);
1927    }
1928}
1929
1930static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1931                         const TCGHelperInfo *info)
1932{
1933    tcg_out_branch(s, 1, dest);
1934
1935#ifndef _WIN32
1936    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1937        /*
1938         * The sysv i386 abi for struct return places a reference as the
1939         * first argument of the stack, and pops that argument with the
1940         * return statement.  Since we want to retain the aligned stack
1941         * pointer for the callee, we do not want to actually push that
1942         * argument before the call but rely on the normal store to the
1943         * stack slot.  But we do need to compensate for the pop in order
1944         * to reset our correct stack pointer value.
1945         * Pushing a garbage value back onto the stack is quickest.
1946         */
1947        tcg_out_push(s, TCG_REG_EAX);
1948    }
1949#endif
1950}
1951
1952static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1953{
1954    tcg_out_branch(s, 0, dest);
1955}
1956
1957static void tcg_out_nopn(TCGContext *s, int n)
1958{
1959    int i;
1960    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1961     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1962     * duplicate prefix, and all of the interesting recent cores can
1963     * decode and discard the duplicates in a single cycle.
1964     */
1965    tcg_debug_assert(n >= 1);
1966    for (i = 1; i < n; ++i) {
1967        tcg_out8(s, 0x66);
1968    }
1969    tcg_out8(s, 0x90);
1970}
1971
1972typedef struct {
1973    TCGReg base;
1974    int index;
1975    int ofs;
1976    int seg;
1977    TCGAtomAlign aa;
1978} HostAddress;
1979
1980bool tcg_target_has_memory_bswap(MemOp memop)
1981{
1982    TCGAtomAlign aa;
1983
1984    if (!have_movbe) {
1985        return false;
1986    }
1987    if ((memop & MO_SIZE) < MO_128) {
1988        return true;
1989    }
1990
1991    /*
1992     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1993     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1994     */
1995    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1996    return aa.atom < MO_128;
1997}
1998
1999/*
2000 * Because i686 has no register parameters and because x86_64 has xchg
2001 * to handle addr/data register overlap, we have placed all input arguments
2002 * before we need might need a scratch reg.
2003 *
2004 * Even then, a scratch is only needed for l->raddr.  Rather than expose
2005 * a general-purpose scratch when we don't actually know it's available,
2006 * use the ra_gen hook to load into RAX if needed.
2007 */
2008#if TCG_TARGET_REG_BITS == 64
2009static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
2010{
2011    if (arg < 0) {
2012        arg = TCG_REG_RAX;
2013    }
2014    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2015    return arg;
2016}
2017static const TCGLdstHelperParam ldst_helper_param = {
2018    .ra_gen = ldst_ra_gen
2019};
2020#else
2021static const TCGLdstHelperParam ldst_helper_param = { };
2022#endif
2023
2024static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2025                                TCGReg l, TCGReg h, TCGReg v)
2026{
2027    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2028
2029    /* vpmov{d,q} %v, %l */
2030    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2031    /* vpextr{d,q} $1, %v, %h */
2032    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2033    tcg_out8(s, 1);
2034}
2035
2036static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2037                                TCGReg v, TCGReg l, TCGReg h)
2038{
2039    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2040
2041    /* vmov{d,q} %l, %v */
2042    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2043    /* vpinsr{d,q} $1, %h, %v, %v */
2044    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2045    tcg_out8(s, 1);
2046}
2047
2048/*
2049 * Generate code for the slow path for a load at the end of block
2050 */
2051static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2052{
2053    MemOp opc = get_memop(l->oi);
2054    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2055
2056    /* resolve label address */
2057    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2058    if (label_ptr[1]) {
2059        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2060    }
2061
2062    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2063    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2064    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2065
2066    tcg_out_jmp(s, l->raddr);
2067    return true;
2068}
2069
2070/*
2071 * Generate code for the slow path for a store at the end of block
2072 */
2073static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2074{
2075    MemOp opc = get_memop(l->oi);
2076    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2077
2078    /* resolve label address */
2079    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2080    if (label_ptr[1]) {
2081        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2082    }
2083
2084    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2085    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2086
2087    tcg_out_jmp(s, l->raddr);
2088    return true;
2089}
2090
2091#ifdef CONFIG_USER_ONLY
2092static HostAddress x86_guest_base = {
2093    .index = -1
2094};
2095
2096#if defined(__x86_64__) && defined(__linux__)
2097# include <asm/prctl.h>
2098# include <sys/prctl.h>
2099int arch_prctl(int code, unsigned long addr);
2100static inline int setup_guest_base_seg(void)
2101{
2102    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2103        return P_GS;
2104    }
2105    return 0;
2106}
2107#define setup_guest_base_seg  setup_guest_base_seg
2108#elif defined(__x86_64__) && \
2109      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2110# include <machine/sysarch.h>
2111static inline int setup_guest_base_seg(void)
2112{
2113    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2114        return P_GS;
2115    }
2116    return 0;
2117}
2118#define setup_guest_base_seg  setup_guest_base_seg
2119#endif
2120#else
2121# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2122#endif /* CONFIG_USER_ONLY */
2123#ifndef setup_guest_base_seg
2124# define setup_guest_base_seg()  0
2125#endif
2126
2127#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2128
2129/*
2130 * For softmmu, perform the TLB load and compare.
2131 * For useronly, perform any required alignment tests.
2132 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2133 * is required and fill in @h with the host address for the fast path.
2134 */
2135static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2136                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2137{
2138    TCGLabelQemuLdst *ldst = NULL;
2139    MemOp opc = get_memop(oi);
2140    MemOp s_bits = opc & MO_SIZE;
2141    unsigned a_mask;
2142
2143    if (tcg_use_softmmu) {
2144        h->index = TCG_REG_L0;
2145        h->ofs = 0;
2146        h->seg = 0;
2147    } else {
2148        *h = x86_guest_base;
2149    }
2150    h->base = addr;
2151    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2152    a_mask = (1 << h->aa.align) - 1;
2153
2154    if (tcg_use_softmmu) {
2155        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2156                            : offsetof(CPUTLBEntry, addr_write);
2157        TCGType ttype = TCG_TYPE_I32;
2158        TCGType tlbtype = TCG_TYPE_I32;
2159        int trexw = 0, hrexw = 0, tlbrexw = 0;
2160        unsigned mem_index = get_mmuidx(oi);
2161        unsigned s_mask = (1 << s_bits) - 1;
2162        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2163        int tlb_mask;
2164
2165        ldst = new_ldst_label(s);
2166        ldst->is_ld = is_ld;
2167        ldst->oi = oi;
2168        ldst->addr_reg = addr;
2169
2170        if (TCG_TARGET_REG_BITS == 64) {
2171            ttype = s->addr_type;
2172            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2173            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2174                hrexw = P_REXW;
2175                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2176                    tlbtype = TCG_TYPE_I64;
2177                    tlbrexw = P_REXW;
2178                }
2179            }
2180        }
2181
2182        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2183        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2184                       s->page_bits - CPU_TLB_ENTRY_BITS);
2185
2186        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2187                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2188
2189        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2190                             fast_ofs + offsetof(CPUTLBDescFast, table));
2191
2192        /*
2193         * If the required alignment is at least as large as the access,
2194         * simply copy the address and mask.  For lesser alignments,
2195         * check that we don't cross pages for the complete access.
2196         */
2197        if (a_mask >= s_mask) {
2198            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2199        } else {
2200            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2201                                 addr, s_mask - a_mask);
2202        }
2203        tlb_mask = s->page_mask | a_mask;
2204        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2205
2206        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2207        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2208                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2209
2210        /* jne slow_path */
2211        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2212        ldst->label_ptr[0] = s->code_ptr;
2213        s->code_ptr += 4;
2214
2215        /* TLB Hit.  */
2216        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2217                   offsetof(CPUTLBEntry, addend));
2218    } else if (a_mask) {
2219        int jcc;
2220
2221        ldst = new_ldst_label(s);
2222        ldst->is_ld = is_ld;
2223        ldst->oi = oi;
2224        ldst->addr_reg = addr;
2225
2226        /* jne slow_path */
2227        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2228        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2229        ldst->label_ptr[0] = s->code_ptr;
2230        s->code_ptr += 4;
2231    }
2232
2233    return ldst;
2234}
2235
2236static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2237                                   HostAddress h, TCGType type, MemOp memop)
2238{
2239    bool use_movbe = false;
2240    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2241    int movop = OPC_MOVL_GvEv;
2242
2243    /* Do big-endian loads with movbe.  */
2244    if (memop & MO_BSWAP) {
2245        tcg_debug_assert(have_movbe);
2246        use_movbe = true;
2247        movop = OPC_MOVBE_GyMy;
2248    }
2249
2250    switch (memop & MO_SSIZE) {
2251    case MO_UB:
2252        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2253                                 h.base, h.index, 0, h.ofs);
2254        break;
2255    case MO_SB:
2256        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2257                                 h.base, h.index, 0, h.ofs);
2258        break;
2259    case MO_UW:
2260        if (use_movbe) {
2261            /* There is no extending movbe; only low 16-bits are modified.  */
2262            if (datalo != h.base && datalo != h.index) {
2263                /* XOR breaks dependency chains.  */
2264                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2265                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2266                                         datalo, h.base, h.index, 0, h.ofs);
2267            } else {
2268                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2269                                         datalo, h.base, h.index, 0, h.ofs);
2270                tcg_out_ext16u(s, datalo, datalo);
2271            }
2272        } else {
2273            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2274                                     h.base, h.index, 0, h.ofs);
2275        }
2276        break;
2277    case MO_SW:
2278        if (use_movbe) {
2279            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2280                                     datalo, h.base, h.index, 0, h.ofs);
2281            tcg_out_ext16s(s, type, datalo, datalo);
2282        } else {
2283            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2284                                     datalo, h.base, h.index, 0, h.ofs);
2285        }
2286        break;
2287    case MO_UL:
2288        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2289                                 h.base, h.index, 0, h.ofs);
2290        break;
2291#if TCG_TARGET_REG_BITS == 64
2292    case MO_SL:
2293        if (use_movbe) {
2294            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2295                                     h.base, h.index, 0, h.ofs);
2296            tcg_out_ext32s(s, datalo, datalo);
2297        } else {
2298            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2299                                     h.base, h.index, 0, h.ofs);
2300        }
2301        break;
2302#endif
2303    case MO_UQ:
2304        if (TCG_TARGET_REG_BITS == 64) {
2305            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2306                                     h.base, h.index, 0, h.ofs);
2307            break;
2308        }
2309        if (use_movbe) {
2310            TCGReg t = datalo;
2311            datalo = datahi;
2312            datahi = t;
2313        }
2314        if (h.base == datalo || h.index == datalo) {
2315            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2316                                     h.base, h.index, 0, h.ofs);
2317            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2318            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2319        } else {
2320            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2321                                     h.base, h.index, 0, h.ofs);
2322            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2323                                     h.base, h.index, 0, h.ofs + 4);
2324        }
2325        break;
2326
2327    case MO_128:
2328        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2329
2330        /*
2331         * Without 16-byte atomicity, use integer regs.
2332         * That is where we want the data, and it allows bswaps.
2333         */
2334        if (h.aa.atom < MO_128) {
2335            if (use_movbe) {
2336                TCGReg t = datalo;
2337                datalo = datahi;
2338                datahi = t;
2339            }
2340            if (h.base == datalo || h.index == datalo) {
2341                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2342                                         h.base, h.index, 0, h.ofs);
2343                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2344                                     datalo, datahi, 0);
2345                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2346                                     datahi, datahi, 8);
2347            } else {
2348                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2349                                         h.base, h.index, 0, h.ofs);
2350                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2351                                         h.base, h.index, 0, h.ofs + 8);
2352            }
2353            break;
2354        }
2355
2356        /*
2357         * With 16-byte atomicity, a vector load is required.
2358         * If we already have 16-byte alignment, then VMOVDQA always works.
2359         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2360         * Else use we require a runtime test for alignment for VMOVDQA;
2361         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2362         */
2363        if (h.aa.align >= MO_128) {
2364            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2365                                         TCG_TMP_VEC, 0,
2366                                         h.base, h.index, 0, h.ofs);
2367        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2368            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2369                                         TCG_TMP_VEC, 0,
2370                                         h.base, h.index, 0, h.ofs);
2371        } else {
2372            TCGLabel *l1 = gen_new_label();
2373            TCGLabel *l2 = gen_new_label();
2374            int jcc;
2375
2376            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2377            tcg_out_jxx(s, jcc, l1, true);
2378
2379            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2380                                         TCG_TMP_VEC, 0,
2381                                         h.base, h.index, 0, h.ofs);
2382            tcg_out_jxx(s, JCC_JMP, l2, true);
2383
2384            tcg_out_label(s, l1);
2385            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2386                                         TCG_TMP_VEC, 0,
2387                                         h.base, h.index, 0, h.ofs);
2388            tcg_out_label(s, l2);
2389        }
2390        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2391        break;
2392
2393    default:
2394        g_assert_not_reached();
2395    }
2396}
2397
2398static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2399                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2400{
2401    TCGLabelQemuLdst *ldst;
2402    HostAddress h;
2403
2404    ldst = prepare_host_addr(s, &h, addr, oi, true);
2405    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2406
2407    if (ldst) {
2408        ldst->type = data_type;
2409        ldst->datalo_reg = datalo;
2410        ldst->datahi_reg = datahi;
2411        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2412    }
2413}
2414
2415static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2416                                   HostAddress h, MemOp memop)
2417{
2418    bool use_movbe = false;
2419    int movop = OPC_MOVL_EvGv;
2420
2421    /*
2422     * Do big-endian stores with movbe or system-mode.
2423     * User-only without movbe will have its swapping done generically.
2424     */
2425    if (memop & MO_BSWAP) {
2426        tcg_debug_assert(have_movbe);
2427        use_movbe = true;
2428        movop = OPC_MOVBE_MyGy;
2429    }
2430
2431    switch (memop & MO_SIZE) {
2432    case MO_8:
2433        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2434        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2435        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2436                                 datalo, h.base, h.index, 0, h.ofs);
2437        break;
2438    case MO_16:
2439        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2440                                 h.base, h.index, 0, h.ofs);
2441        break;
2442    case MO_32:
2443        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2444                                 h.base, h.index, 0, h.ofs);
2445        break;
2446    case MO_64:
2447        if (TCG_TARGET_REG_BITS == 64) {
2448            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2449                                     h.base, h.index, 0, h.ofs);
2450        } else {
2451            if (use_movbe) {
2452                TCGReg t = datalo;
2453                datalo = datahi;
2454                datahi = t;
2455            }
2456            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2457                                     h.base, h.index, 0, h.ofs);
2458            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2459                                     h.base, h.index, 0, h.ofs + 4);
2460        }
2461        break;
2462
2463    case MO_128:
2464        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2465
2466        /*
2467         * Without 16-byte atomicity, use integer regs.
2468         * That is where we have the data, and it allows bswaps.
2469         */
2470        if (h.aa.atom < MO_128) {
2471            if (use_movbe) {
2472                TCGReg t = datalo;
2473                datalo = datahi;
2474                datahi = t;
2475            }
2476            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2477                                     h.base, h.index, 0, h.ofs);
2478            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2479                                     h.base, h.index, 0, h.ofs + 8);
2480            break;
2481        }
2482
2483        /*
2484         * With 16-byte atomicity, a vector store is required.
2485         * If we already have 16-byte alignment, then VMOVDQA always works.
2486         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2487         * Else use we require a runtime test for alignment for VMOVDQA;
2488         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2489         */
2490        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2491        if (h.aa.align >= MO_128) {
2492            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2493                                         TCG_TMP_VEC, 0,
2494                                         h.base, h.index, 0, h.ofs);
2495        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2496            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2497                                         TCG_TMP_VEC, 0,
2498                                         h.base, h.index, 0, h.ofs);
2499        } else {
2500            TCGLabel *l1 = gen_new_label();
2501            TCGLabel *l2 = gen_new_label();
2502            int jcc;
2503
2504            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2505            tcg_out_jxx(s, jcc, l1, true);
2506
2507            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2508                                         TCG_TMP_VEC, 0,
2509                                         h.base, h.index, 0, h.ofs);
2510            tcg_out_jxx(s, JCC_JMP, l2, true);
2511
2512            tcg_out_label(s, l1);
2513            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2514                                         TCG_TMP_VEC, 0,
2515                                         h.base, h.index, 0, h.ofs);
2516            tcg_out_label(s, l2);
2517        }
2518        break;
2519
2520    default:
2521        g_assert_not_reached();
2522    }
2523}
2524
2525static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2526                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2527{
2528    TCGLabelQemuLdst *ldst;
2529    HostAddress h;
2530
2531    ldst = prepare_host_addr(s, &h, addr, oi, false);
2532    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2533
2534    if (ldst) {
2535        ldst->type = data_type;
2536        ldst->datalo_reg = datalo;
2537        ldst->datahi_reg = datahi;
2538        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2539    }
2540}
2541
2542static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2543{
2544    /* Reuse the zeroing that exists for goto_ptr.  */
2545    if (a0 == 0) {
2546        tcg_out_jmp(s, tcg_code_gen_epilogue);
2547    } else {
2548        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2549        tcg_out_jmp(s, tb_ret_addr);
2550    }
2551}
2552
2553static void tcg_out_goto_tb(TCGContext *s, int which)
2554{
2555    /*
2556     * Jump displacement must be aligned for atomic patching;
2557     * see if we need to add extra nops before jump
2558     */
2559    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2560    if (gap != 1) {
2561        tcg_out_nopn(s, gap - 1);
2562    }
2563    tcg_out8(s, OPC_JMP_long); /* jmp im */
2564    set_jmp_insn_offset(s, which);
2565    tcg_out32(s, 0);
2566    set_jmp_reset_offset(s, which);
2567}
2568
2569void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2570                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2571{
2572    /* patch the branch destination */
2573    uintptr_t addr = tb->jmp_target_addr[n];
2574    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2575    /* no need to flush icache explicitly */
2576}
2577
2578
2579static void tgen_add(TCGContext *s, TCGType type,
2580                     TCGReg a0, TCGReg a1, TCGReg a2)
2581{
2582    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2583
2584    if (a0 == a1) {
2585        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2586    } else if (a0 == a2) {
2587        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2588    } else {
2589        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2590    }
2591}
2592
2593static void tgen_addi(TCGContext *s, TCGType type,
2594                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2595{
2596    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2597
2598    if (a0 == a1) {
2599        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2600    } else {
2601        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2602    }
2603}
2604
2605static const TCGOutOpBinary outop_add = {
2606    .base.static_constraint = C_O1_I2(r, r, re),
2607    .out_rrr = tgen_add,
2608    .out_rri = tgen_addi,
2609};
2610
2611static void tgen_and(TCGContext *s, TCGType type,
2612                     TCGReg a0, TCGReg a1, TCGReg a2)
2613{
2614    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2615    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2616}
2617
2618static void tgen_andi(TCGContext *s, TCGType type,
2619                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2620{
2621    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2622    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2623}
2624
2625static const TCGOutOpBinary outop_and = {
2626    .base.static_constraint = C_O1_I2(r, 0, reZ),
2627    .out_rrr = tgen_and,
2628    .out_rri = tgen_andi,
2629};
2630
2631static void tgen_andc(TCGContext *s, TCGType type,
2632                      TCGReg a0, TCGReg a1, TCGReg a2)
2633{
2634    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2635    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2636}
2637
2638static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2639{
2640    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2641}
2642
2643static const TCGOutOpBinary outop_andc = {
2644    .base.static_constraint = C_Dynamic,
2645    .base.dynamic_constraint = cset_andc,
2646    .out_rrr = tgen_andc,
2647};
2648
2649static void tgen_clz(TCGContext *s, TCGType type,
2650                     TCGReg a0, TCGReg a1, TCGReg a2)
2651{
2652    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2653    int jcc;
2654
2655    if (have_lzcnt) {
2656        tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2657        jcc = JCC_JB;
2658    } else {
2659        /* Recall that the output of BSR is the index not the count.  */
2660        tcg_out_modrm(s, OPC_BSR + rexw, a0, a1);
2661        tgen_arithi(s, ARITH_XOR + rexw, a0, rexw ? 63 : 31, 0);
2662
2663        /* Since we have destroyed the flags from BSR, we have to re-test.  */
2664        jcc = tcg_out_cmp(s, TCG_COND_EQ, a1, 0, 1, rexw);
2665    }
2666    tcg_out_cmov(s, jcc, rexw, a0, a2);
2667}
2668
2669static void tgen_clzi(TCGContext *s, TCGType type,
2670                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2671{
2672    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2673    tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2674}
2675
2676static TCGConstraintSetIndex cset_clz(TCGType type, unsigned flags)
2677{
2678    return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2679}
2680
2681static const TCGOutOpBinary outop_clz = {
2682    .base.static_constraint = C_Dynamic,
2683    .base.dynamic_constraint = cset_clz,
2684    .out_rrr = tgen_clz,
2685    .out_rri = tgen_clzi,
2686};
2687
2688static void tgen_ctpop(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2689{
2690    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2691    tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2692}
2693
2694static TCGConstraintSetIndex cset_ctpop(TCGType type, unsigned flags)
2695{
2696    return have_popcnt ? C_O1_I1(r, r) : C_NotImplemented;
2697}
2698
2699static const TCGOutOpUnary outop_ctpop = {
2700    .base.static_constraint = C_Dynamic,
2701    .base.dynamic_constraint = cset_ctpop,
2702    .out_rr = tgen_ctpop,
2703};
2704
2705static void tgen_ctz(TCGContext *s, TCGType type,
2706                     TCGReg a0, TCGReg a1, TCGReg a2)
2707{
2708    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2709    int jcc;
2710
2711    if (have_bmi1) {
2712        tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2713        jcc = JCC_JB;
2714    } else {
2715        tcg_out_modrm(s, OPC_BSF + rexw, a0, a1);
2716        jcc = JCC_JE;
2717    }
2718    tcg_out_cmov(s, jcc, rexw, a0, a2);
2719}
2720
2721static void tgen_ctzi(TCGContext *s, TCGType type,
2722                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2723{
2724    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2725    tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2726}
2727
2728static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
2729{
2730    return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2731}
2732
2733static const TCGOutOpBinary outop_ctz = {
2734    .base.static_constraint = C_Dynamic,
2735    .base.dynamic_constraint = cset_ctz,
2736    .out_rrr = tgen_ctz,
2737    .out_rri = tgen_ctzi,
2738};
2739
2740static const TCGOutOpBinary outop_divs = {
2741    .base.static_constraint = C_NotImplemented,
2742};
2743
2744static void tgen_divs2(TCGContext *s, TCGType type,
2745                       TCGReg a0, TCGReg a1, TCGReg a4)
2746{
2747    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2748    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, a4);
2749}
2750
2751static const TCGOutOpDivRem outop_divs2 = {
2752    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2753    .out_rr01r = tgen_divs2,
2754};
2755
2756static const TCGOutOpBinary outop_divu = {
2757    .base.static_constraint = C_NotImplemented,
2758};
2759
2760static void tgen_divu2(TCGContext *s, TCGType type,
2761                       TCGReg a0, TCGReg a1, TCGReg a4)
2762{
2763    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2764    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, a4);
2765}
2766
2767static const TCGOutOpDivRem outop_divu2 = {
2768    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2769    .out_rr01r = tgen_divu2,
2770};
2771
2772static const TCGOutOpBinary outop_eqv = {
2773    .base.static_constraint = C_NotImplemented,
2774};
2775
2776static void tgen_mul(TCGContext *s, TCGType type,
2777                     TCGReg a0, TCGReg a1, TCGReg a2)
2778{
2779    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2780    tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2781}
2782
2783static void tgen_muli(TCGContext *s, TCGType type,
2784                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2785{
2786    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2787
2788    if (a2 == (int8_t)a2) {
2789        tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2790        tcg_out8(s, a2);
2791    } else {
2792        tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2793        tcg_out32(s, a2);
2794    }
2795}
2796
2797static const TCGOutOpBinary outop_mul = {
2798    .base.static_constraint = C_O1_I2(r, 0, re),
2799    .out_rrr = tgen_mul,
2800    .out_rri = tgen_muli,
2801};
2802
2803static void tgen_muls2(TCGContext *s, TCGType type,
2804                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2805{
2806    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2807    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, a3);
2808}
2809
2810static const TCGOutOpMul2 outop_muls2 = {
2811    .base.static_constraint = C_O2_I2(a, d, a, r),
2812    .out_rrrr = tgen_muls2,
2813};
2814
2815static const TCGOutOpBinary outop_mulsh = {
2816    .base.static_constraint = C_NotImplemented,
2817};
2818
2819static const TCGOutOpBinary outop_muluh = {
2820    .base.static_constraint = C_NotImplemented,
2821};
2822
2823static void tgen_mulu2(TCGContext *s, TCGType type,
2824                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2825{
2826    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2827    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, a3);
2828}
2829
2830static const TCGOutOpMul2 outop_mulu2 = {
2831    .base.static_constraint = C_O2_I2(a, d, a, r),
2832    .out_rrrr = tgen_mulu2,
2833};
2834
2835static const TCGOutOpBinary outop_nand = {
2836    .base.static_constraint = C_NotImplemented,
2837};
2838
2839static const TCGOutOpBinary outop_nor = {
2840    .base.static_constraint = C_NotImplemented,
2841};
2842
2843static void tgen_or(TCGContext *s, TCGType type,
2844                    TCGReg a0, TCGReg a1, TCGReg a2)
2845{
2846    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2847    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
2848}
2849
2850static void tgen_ori(TCGContext *s, TCGType type,
2851                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2852{
2853    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2854    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
2855}
2856
2857static const TCGOutOpBinary outop_or = {
2858    .base.static_constraint = C_O1_I2(r, 0, re),
2859    .out_rrr = tgen_or,
2860    .out_rri = tgen_ori,
2861};
2862
2863static const TCGOutOpBinary outop_orc = {
2864    .base.static_constraint = C_NotImplemented,
2865};
2866
2867static const TCGOutOpBinary outop_rems = {
2868    .base.static_constraint = C_NotImplemented,
2869};
2870
2871static const TCGOutOpBinary outop_remu = {
2872    .base.static_constraint = C_NotImplemented,
2873};
2874
2875static void tgen_rotl(TCGContext *s, TCGType type,
2876                     TCGReg a0, TCGReg a1, TCGReg a2)
2877{
2878    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2879    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROL, a0);
2880}
2881
2882static void tgen_rotli(TCGContext *s, TCGType type,
2883                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2884{
2885    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2886    tcg_out_shifti(s, SHIFT_ROL + rexw, a0, a2);
2887}
2888
2889static const TCGOutOpBinary outop_rotl = {
2890    .base.static_constraint = C_O1_I2(r, 0, ci),
2891    .out_rrr = tgen_rotl,
2892    .out_rri = tgen_rotli,
2893};
2894
2895static void tgen_rotr(TCGContext *s, TCGType type,
2896                     TCGReg a0, TCGReg a1, TCGReg a2)
2897{
2898    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2899    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROR, a0);
2900}
2901
2902static void tgen_rotri(TCGContext *s, TCGType type,
2903                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2904{
2905    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2906    tcg_out_shifti(s, SHIFT_ROR + rexw, a0, a2);
2907}
2908
2909static const TCGOutOpBinary outop_rotr = {
2910    .base.static_constraint = C_O1_I2(r, 0, ci),
2911    .out_rrr = tgen_rotr,
2912    .out_rri = tgen_rotri,
2913};
2914
2915static TCGConstraintSetIndex cset_shift(TCGType type, unsigned flags)
2916{
2917    return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
2918}
2919
2920static void tgen_sar(TCGContext *s, TCGType type,
2921                     TCGReg a0, TCGReg a1, TCGReg a2)
2922{
2923    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2924    if (have_bmi2) {
2925        tcg_out_vex_modrm(s, OPC_SARX + rexw, a0, a2, a1);
2926    } else {
2927        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SAR, a0);
2928    }
2929}
2930
2931static void tgen_sari(TCGContext *s, TCGType type,
2932                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2933{
2934    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2935
2936    tcg_out_mov(s, type, a0, a1);
2937    tcg_out_shifti(s, SHIFT_SAR + rexw, a0, a2);
2938}
2939
2940static const TCGOutOpBinary outop_sar = {
2941    .base.static_constraint = C_Dynamic,
2942    .base.dynamic_constraint = cset_shift,
2943    .out_rrr = tgen_sar,
2944    .out_rri = tgen_sari,
2945};
2946
2947static void tgen_shl(TCGContext *s, TCGType type,
2948                     TCGReg a0, TCGReg a1, TCGReg a2)
2949{
2950    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2951    if (have_bmi2) {
2952        tcg_out_vex_modrm(s, OPC_SHLX + rexw, a0, a2, a1);
2953    } else {
2954        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHL, a0);
2955    }
2956}
2957
2958static void tgen_shli(TCGContext *s, TCGType type,
2959                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2960{
2961    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2962
2963    /* For small constant 3-operand shift, use LEA.  */
2964    if (a0 != a1 && a2 >= 1 && a2 <= 3) {
2965        if (a2 == 1) {
2966            /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2967            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2968        } else {
2969            /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2970            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2971        }
2972        return;
2973    }
2974    tcg_out_mov(s, type, a0, a1);
2975    tcg_out_shifti(s, SHIFT_SHL + rexw, a0, a2);
2976}
2977
2978static const TCGOutOpBinary outop_shl = {
2979    .base.static_constraint = C_Dynamic,
2980    .base.dynamic_constraint = cset_shift,
2981    .out_rrr = tgen_shl,
2982    .out_rri = tgen_shli,
2983};
2984
2985static void tgen_shr(TCGContext *s, TCGType type,
2986                     TCGReg a0, TCGReg a1, TCGReg a2)
2987{
2988    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2989    if (have_bmi2) {
2990        tcg_out_vex_modrm(s, OPC_SHRX + rexw, a0, a2, a1);
2991    } else {
2992        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHR, a0);
2993    }
2994}
2995
2996static void tgen_shri(TCGContext *s, TCGType type,
2997                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2998{
2999    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3000
3001    tcg_out_mov(s, type, a0, a1);
3002    tcg_out_shifti(s, SHIFT_SHR + rexw, a0, a2);
3003}
3004
3005static const TCGOutOpBinary outop_shr = {
3006    .base.static_constraint = C_Dynamic,
3007    .base.dynamic_constraint = cset_shift,
3008    .out_rrr = tgen_shr,
3009    .out_rri = tgen_shri,
3010};
3011
3012static void tgen_sub(TCGContext *s, TCGType type,
3013                      TCGReg a0, TCGReg a1, TCGReg a2)
3014{
3015    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3016    tgen_arithr(s, ARITH_SUB + rexw, a0, a2);
3017}
3018
3019static const TCGOutOpSubtract outop_sub = {
3020    .base.static_constraint = C_O1_I2(r, 0, r),
3021    .out_rrr = tgen_sub,
3022};
3023
3024static void tgen_xor(TCGContext *s, TCGType type,
3025                     TCGReg a0, TCGReg a1, TCGReg a2)
3026{
3027    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3028    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
3029}
3030
3031static void tgen_xori(TCGContext *s, TCGType type,
3032                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3033{
3034    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3035    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
3036}
3037
3038static const TCGOutOpBinary outop_xor = {
3039    .base.static_constraint = C_O1_I2(r, 0, re),
3040    .out_rrr = tgen_xor,
3041    .out_rri = tgen_xori,
3042};
3043
3044static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3045{
3046    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3047    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
3048}
3049
3050static const TCGOutOpUnary outop_neg = {
3051    .base.static_constraint = C_O1_I1(r, 0),
3052    .out_rr = tgen_neg,
3053};
3054
3055static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3056{
3057    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3058    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
3059}
3060
3061static const TCGOutOpUnary outop_not = {
3062    .base.static_constraint = C_O1_I1(r, 0),
3063    .out_rr = tgen_not,
3064};
3065
3066
3067static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
3068                       const TCGArg args[TCG_MAX_OP_ARGS],
3069                       const int const_args[TCG_MAX_OP_ARGS])
3070{
3071    TCGArg a0, a1, a2;
3072    int const_a2, rexw;
3073
3074#if TCG_TARGET_REG_BITS == 64
3075# define OP_32_64(x) \
3076        case glue(glue(INDEX_op_, x), _i64): \
3077        case glue(glue(INDEX_op_, x), _i32)
3078#else
3079# define OP_32_64(x) \
3080        case glue(glue(INDEX_op_, x), _i32)
3081#endif
3082
3083    /* Hoist the loads of the most common arguments.  */
3084    a0 = args[0];
3085    a1 = args[1];
3086    a2 = args[2];
3087    const_a2 = const_args[2];
3088    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3089
3090    switch (opc) {
3091    case INDEX_op_goto_ptr:
3092        /* jmp to the given host address (could be epilogue) */
3093        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
3094        break;
3095    case INDEX_op_br:
3096        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
3097        break;
3098    OP_32_64(ld8u):
3099        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
3100        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
3101        break;
3102    OP_32_64(ld8s):
3103        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
3104        break;
3105    OP_32_64(ld16u):
3106        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
3107        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
3108        break;
3109    OP_32_64(ld16s):
3110        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
3111        break;
3112#if TCG_TARGET_REG_BITS == 64
3113    case INDEX_op_ld32u_i64:
3114#endif
3115    case INDEX_op_ld_i32:
3116        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
3117        break;
3118
3119    OP_32_64(st8):
3120        if (const_args[0]) {
3121            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
3122            tcg_out8(s, a0);
3123        } else {
3124            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
3125        }
3126        break;
3127    OP_32_64(st16):
3128        if (const_args[0]) {
3129            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
3130            tcg_out16(s, a0);
3131        } else {
3132            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
3133        }
3134        break;
3135#if TCG_TARGET_REG_BITS == 64
3136    case INDEX_op_st32_i64:
3137#endif
3138    case INDEX_op_st_i32:
3139        if (const_args[0]) {
3140            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
3141            tcg_out32(s, a0);
3142        } else {
3143            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
3144        }
3145        break;
3146
3147    OP_32_64(movcond):
3148        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
3149        break;
3150
3151    OP_32_64(bswap16):
3152        if (a2 & TCG_BSWAP_OS) {
3153            /* Output must be sign-extended. */
3154            if (rexw) {
3155                tcg_out_bswap64(s, a0);
3156                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
3157            } else {
3158                tcg_out_bswap32(s, a0);
3159                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
3160            }
3161        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
3162            /* Output must be zero-extended, but input isn't. */
3163            tcg_out_bswap32(s, a0);
3164            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
3165        } else {
3166            tcg_out_rolw_8(s, a0);
3167        }
3168        break;
3169    OP_32_64(bswap32):
3170        tcg_out_bswap32(s, a0);
3171        if (rexw && (a2 & TCG_BSWAP_OS)) {
3172            tcg_out_ext32s(s, a0, a0);
3173        }
3174        break;
3175
3176    case INDEX_op_qemu_ld_i32:
3177        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
3178        break;
3179    case INDEX_op_qemu_ld_i64:
3180        if (TCG_TARGET_REG_BITS == 64) {
3181            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
3182        } else {
3183            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3184        }
3185        break;
3186    case INDEX_op_qemu_ld_i128:
3187        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3188        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3189        break;
3190
3191    case INDEX_op_qemu_st_i32:
3192    case INDEX_op_qemu_st8_i32:
3193        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
3194        break;
3195    case INDEX_op_qemu_st_i64:
3196        if (TCG_TARGET_REG_BITS == 64) {
3197            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
3198        } else {
3199            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3200        }
3201        break;
3202    case INDEX_op_qemu_st_i128:
3203        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3204        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3205        break;
3206
3207    OP_32_64(add2):
3208        if (const_args[4]) {
3209            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
3210        } else {
3211            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
3212        }
3213        if (const_args[5]) {
3214            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
3215        } else {
3216            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
3217        }
3218        break;
3219    OP_32_64(sub2):
3220        if (const_args[4]) {
3221            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
3222        } else {
3223            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
3224        }
3225        if (const_args[5]) {
3226            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
3227        } else {
3228            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
3229        }
3230        break;
3231
3232#if TCG_TARGET_REG_BITS == 32
3233    case INDEX_op_brcond2_i32:
3234        tcg_out_brcond2(s, args, const_args, 0);
3235        break;
3236    case INDEX_op_setcond2_i32:
3237        tcg_out_setcond2(s, args, const_args);
3238        break;
3239#else /* TCG_TARGET_REG_BITS == 64 */
3240    case INDEX_op_ld32s_i64:
3241        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
3242        break;
3243    case INDEX_op_ld_i64:
3244        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
3245        break;
3246    case INDEX_op_st_i64:
3247        if (const_args[0]) {
3248            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
3249            tcg_out32(s, a0);
3250        } else {
3251            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
3252        }
3253        break;
3254
3255    case INDEX_op_bswap64_i64:
3256        tcg_out_bswap64(s, a0);
3257        break;
3258    case INDEX_op_extrh_i64_i32:
3259        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
3260        break;
3261#endif
3262
3263    OP_32_64(deposit):
3264        if (args[3] == 0 && args[4] == 8) {
3265            /* load bits 0..7 */
3266            if (const_a2) {
3267                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
3268                            0, a0, 0);
3269                tcg_out8(s, a2);
3270            } else {
3271                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3272            }
3273        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
3274            /* load bits 8..15 */
3275            if (const_a2) {
3276                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3277                tcg_out8(s, a2);
3278            } else {
3279                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3280            }
3281        } else if (args[3] == 0 && args[4] == 16) {
3282            /* load bits 0..15 */
3283            if (const_a2) {
3284                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
3285                            0, a0, 0);
3286                tcg_out16(s, a2);
3287            } else {
3288                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3289            }
3290        } else {
3291            g_assert_not_reached();
3292        }
3293        break;
3294
3295    case INDEX_op_extract_i64:
3296        if (a2 + args[3] == 32) {
3297            if (a2 == 0) {
3298                tcg_out_ext32u(s, a0, a1);
3299                break;
3300            }
3301            /* This is a 32-bit zero-extending right shift.  */
3302            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3303            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
3304            break;
3305        }
3306        /* FALLTHRU */
3307    case INDEX_op_extract_i32:
3308        if (a2 == 0 && args[3] == 8) {
3309            tcg_out_ext8u(s, a0, a1);
3310        } else if (a2 == 0 && args[3] == 16) {
3311            tcg_out_ext16u(s, a0, a1);
3312        } else if (a2 == 8 && args[3] == 8) {
3313            /*
3314             * On the off-chance that we can use the high-byte registers.
3315             * Otherwise we emit the same ext16 + shift pattern that we
3316             * would have gotten from the normal tcg-op.c expansion.
3317             */
3318            if (a1 < 4 && a0 < 8) {
3319                tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3320            } else {
3321                tcg_out_ext16u(s, a0, a1);
3322                tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3323            }
3324        } else {
3325            g_assert_not_reached();
3326        }
3327        break;
3328
3329    case INDEX_op_sextract_i64:
3330        if (a2 == 0 && args[3] == 8) {
3331            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3332        } else if (a2 == 0 && args[3] == 16) {
3333            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3334        } else if (a2 == 0 && args[3] == 32) {
3335            tcg_out_ext32s(s, a0, a1);
3336        } else {
3337            g_assert_not_reached();
3338        }
3339        break;
3340
3341    case INDEX_op_sextract_i32:
3342        if (a2 == 0 && args[3] == 8) {
3343            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3344        } else if (a2 == 0 && args[3] == 16) {
3345            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3346        } else if (a2 == 8 && args[3] == 8) {
3347            if (a1 < 4 && a0 < 8) {
3348                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3349            } else {
3350                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3351                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3352            }
3353        } else {
3354            g_assert_not_reached();
3355        }
3356        break;
3357
3358    OP_32_64(extract2):
3359        /* Note that SHRD outputs to the r/m operand.  */
3360        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3361        tcg_out8(s, args[3]);
3362        break;
3363
3364    case INDEX_op_mb:
3365        tcg_out_mb(s, a0);
3366        break;
3367    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3368    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3369    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3370    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
3371    case INDEX_op_extu_i32_i64:
3372    case INDEX_op_extrl_i64_i32:
3373    default:
3374        g_assert_not_reached();
3375    }
3376
3377#undef OP_32_64
3378}
3379
3380static int const umin_insn[4] = {
3381    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3382};
3383
3384static int const umax_insn[4] = {
3385    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3386};
3387
3388static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3389                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3390{
3391    static int const cmpeq_insn[4] = {
3392        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3393    };
3394    static int const cmpgt_insn[4] = {
3395        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3396    };
3397
3398    enum {
3399        NEED_INV  = 1,
3400        NEED_SWAP = 2,
3401        NEED_UMIN = 4,
3402        NEED_UMAX = 8,
3403        INVALID   = 16,
3404    };
3405    static const uint8_t cond_fixup[16] = {
3406        [0 ... 15] = INVALID,
3407        [TCG_COND_EQ] = 0,
3408        [TCG_COND_GT] = 0,
3409        [TCG_COND_NE] = NEED_INV,
3410        [TCG_COND_LE] = NEED_INV,
3411        [TCG_COND_LT] = NEED_SWAP,
3412        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3413        [TCG_COND_LEU] = NEED_UMIN,
3414        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3415        [TCG_COND_GEU] = NEED_UMAX,
3416        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3417    };
3418    int fixup = cond_fixup[cond];
3419
3420    assert(!(fixup & INVALID));
3421
3422    if (fixup & NEED_INV) {
3423        cond = tcg_invert_cond(cond);
3424    }
3425
3426    if (fixup & NEED_SWAP) {
3427        TCGReg swap = v1;
3428        v1 = v2;
3429        v2 = swap;
3430        cond = tcg_swap_cond(cond);
3431    }
3432
3433    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3434        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3435
3436        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3437        assert(vece <= MO_32);
3438
3439        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3440        v2 = TCG_TMP_VEC;
3441        cond = TCG_COND_EQ;
3442    }
3443
3444    switch (cond) {
3445    case TCG_COND_EQ:
3446        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3447        break;
3448    case TCG_COND_GT:
3449        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3450        break;
3451    default:
3452        g_assert_not_reached();
3453    }
3454    return fixup & NEED_INV;
3455}
3456
3457static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3458                               TCGReg v1, TCGReg v2, TCGCond cond)
3459{
3460    static const int cmpm_insn[2][4] = {
3461        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3462        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3463    };
3464    static const int testm_insn[4] = {
3465        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3466    };
3467    static const int testnm_insn[4] = {
3468        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3469    };
3470
3471    static const int cond_ext[16] = {
3472        [TCG_COND_EQ] = 0,
3473        [TCG_COND_NE] = 4,
3474        [TCG_COND_LT] = 1,
3475        [TCG_COND_LTU] = 1,
3476        [TCG_COND_LE] = 2,
3477        [TCG_COND_LEU] = 2,
3478        [TCG_COND_NEVER] = 3,
3479        [TCG_COND_GE] = 5,
3480        [TCG_COND_GEU] = 5,
3481        [TCG_COND_GT] = 6,
3482        [TCG_COND_GTU] = 6,
3483        [TCG_COND_ALWAYS] = 7,
3484    };
3485
3486    switch (cond) {
3487    case TCG_COND_TSTNE:
3488        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3489        break;
3490    case TCG_COND_TSTEQ:
3491        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3492        break;
3493    default:
3494        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3495                               /* k1 */ 1, v1, v2, type);
3496        tcg_out8(s, cond_ext[cond]);
3497        break;
3498    }
3499}
3500
3501static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3502                              unsigned vece, TCGReg dest)
3503{
3504    static const int movm_insn[] = {
3505        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3506    };
3507    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3508}
3509
3510static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3511                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3512{
3513    /*
3514     * With avx512, we have a complete set of comparisons into mask.
3515     * Unless there's a single insn expansion for the comparision,
3516     * expand via a mask in k1.
3517     */
3518    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3519        && cond != TCG_COND_EQ
3520        && cond != TCG_COND_LT
3521        && cond != TCG_COND_GT) {
3522        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3523        tcg_out_k1_to_vec(s, type, vece, v0);
3524        return;
3525    }
3526
3527    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3528        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3529        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3530    }
3531}
3532
3533static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3534                                  TCGReg v0, TCGReg c1, TCGReg c2,
3535                                  TCGReg v3, TCGReg v4, TCGCond cond)
3536{
3537    static const int vpblendm_insn[] = {
3538        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3539    };
3540    bool z = false;
3541
3542    /* Swap to place constant in V4 to take advantage of zero-masking. */
3543    if (!v3) {
3544        z = true;
3545        v3 = v4;
3546        cond = tcg_invert_cond(cond);
3547    }
3548
3549    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3550    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3551                            /* k1 */1, z, type);
3552}
3553
3554static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3555                               TCGReg v0, TCGReg c1, TCGReg c2,
3556                               TCGReg v3, TCGReg v4, TCGCond cond)
3557{
3558    bool inv;
3559
3560    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3561        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3562        return;
3563    }
3564
3565    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3566
3567    /*
3568     * Since XMM0 is 16, the only way we get 0 into V3
3569     * is via the constant zero constraint.
3570     */
3571    if (!v3) {
3572        if (inv) {
3573            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3574        } else {
3575            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3576        }
3577    } else {
3578        if (inv) {
3579            TCGReg swap = v3;
3580            v3 = v4;
3581            v4 = swap;
3582        }
3583        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3584        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3585    }
3586}
3587
3588static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3589                           unsigned vecl, unsigned vece,
3590                           const TCGArg args[TCG_MAX_OP_ARGS],
3591                           const int const_args[TCG_MAX_OP_ARGS])
3592{
3593    static int const add_insn[4] = {
3594        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3595    };
3596    static int const ssadd_insn[4] = {
3597        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3598    };
3599    static int const usadd_insn[4] = {
3600        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3601    };
3602    static int const sub_insn[4] = {
3603        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3604    };
3605    static int const sssub_insn[4] = {
3606        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3607    };
3608    static int const ussub_insn[4] = {
3609        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3610    };
3611    static int const mul_insn[4] = {
3612        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3613    };
3614    static int const shift_imm_insn[4] = {
3615        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3616    };
3617    static int const punpckl_insn[4] = {
3618        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3619    };
3620    static int const punpckh_insn[4] = {
3621        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3622    };
3623    static int const packss_insn[4] = {
3624        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3625    };
3626    static int const packus_insn[4] = {
3627        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3628    };
3629    static int const smin_insn[4] = {
3630        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3631    };
3632    static int const smax_insn[4] = {
3633        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3634    };
3635    static int const rotlv_insn[4] = {
3636        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3637    };
3638    static int const rotrv_insn[4] = {
3639        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3640    };
3641    static int const shlv_insn[4] = {
3642        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3643    };
3644    static int const shrv_insn[4] = {
3645        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3646    };
3647    static int const sarv_insn[4] = {
3648        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3649    };
3650    static int const shls_insn[4] = {
3651        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3652    };
3653    static int const shrs_insn[4] = {
3654        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3655    };
3656    static int const sars_insn[4] = {
3657        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3658    };
3659    static int const vpshldi_insn[4] = {
3660        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3661    };
3662    static int const vpshldv_insn[4] = {
3663        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3664    };
3665    static int const vpshrdv_insn[4] = {
3666        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3667    };
3668    static int const abs_insn[4] = {
3669        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3670    };
3671
3672    TCGType type = vecl + TCG_TYPE_V64;
3673    int insn, sub;
3674    TCGArg a0, a1, a2, a3;
3675
3676    a0 = args[0];
3677    a1 = args[1];
3678    a2 = args[2];
3679
3680    switch (opc) {
3681    case INDEX_op_add_vec:
3682        insn = add_insn[vece];
3683        goto gen_simd;
3684    case INDEX_op_ssadd_vec:
3685        insn = ssadd_insn[vece];
3686        goto gen_simd;
3687    case INDEX_op_usadd_vec:
3688        insn = usadd_insn[vece];
3689        goto gen_simd;
3690    case INDEX_op_sub_vec:
3691        insn = sub_insn[vece];
3692        goto gen_simd;
3693    case INDEX_op_sssub_vec:
3694        insn = sssub_insn[vece];
3695        goto gen_simd;
3696    case INDEX_op_ussub_vec:
3697        insn = ussub_insn[vece];
3698        goto gen_simd;
3699    case INDEX_op_mul_vec:
3700        insn = mul_insn[vece];
3701        goto gen_simd;
3702    case INDEX_op_and_vec:
3703        insn = OPC_PAND;
3704        goto gen_simd;
3705    case INDEX_op_or_vec:
3706        insn = OPC_POR;
3707        goto gen_simd;
3708    case INDEX_op_xor_vec:
3709        insn = OPC_PXOR;
3710        goto gen_simd;
3711    case INDEX_op_smin_vec:
3712        insn = smin_insn[vece];
3713        goto gen_simd;
3714    case INDEX_op_umin_vec:
3715        insn = umin_insn[vece];
3716        goto gen_simd;
3717    case INDEX_op_smax_vec:
3718        insn = smax_insn[vece];
3719        goto gen_simd;
3720    case INDEX_op_umax_vec:
3721        insn = umax_insn[vece];
3722        goto gen_simd;
3723    case INDEX_op_shlv_vec:
3724        insn = shlv_insn[vece];
3725        goto gen_simd;
3726    case INDEX_op_shrv_vec:
3727        insn = shrv_insn[vece];
3728        goto gen_simd;
3729    case INDEX_op_sarv_vec:
3730        insn = sarv_insn[vece];
3731        goto gen_simd;
3732    case INDEX_op_rotlv_vec:
3733        insn = rotlv_insn[vece];
3734        goto gen_simd;
3735    case INDEX_op_rotrv_vec:
3736        insn = rotrv_insn[vece];
3737        goto gen_simd;
3738    case INDEX_op_shls_vec:
3739        insn = shls_insn[vece];
3740        goto gen_simd;
3741    case INDEX_op_shrs_vec:
3742        insn = shrs_insn[vece];
3743        goto gen_simd;
3744    case INDEX_op_sars_vec:
3745        insn = sars_insn[vece];
3746        goto gen_simd;
3747    case INDEX_op_x86_punpckl_vec:
3748        insn = punpckl_insn[vece];
3749        goto gen_simd;
3750    case INDEX_op_x86_punpckh_vec:
3751        insn = punpckh_insn[vece];
3752        goto gen_simd;
3753    case INDEX_op_x86_packss_vec:
3754        insn = packss_insn[vece];
3755        goto gen_simd;
3756    case INDEX_op_x86_packus_vec:
3757        insn = packus_insn[vece];
3758        goto gen_simd;
3759    case INDEX_op_x86_vpshldv_vec:
3760        insn = vpshldv_insn[vece];
3761        a1 = a2;
3762        a2 = args[3];
3763        goto gen_simd;
3764    case INDEX_op_x86_vpshrdv_vec:
3765        insn = vpshrdv_insn[vece];
3766        a1 = a2;
3767        a2 = args[3];
3768        goto gen_simd;
3769#if TCG_TARGET_REG_BITS == 32
3770    case INDEX_op_dup2_vec:
3771        /* First merge the two 32-bit inputs to a single 64-bit element. */
3772        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3773        /* Then replicate the 64-bit elements across the rest of the vector. */
3774        if (type != TCG_TYPE_V64) {
3775            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3776        }
3777        break;
3778#endif
3779    case INDEX_op_abs_vec:
3780        insn = abs_insn[vece];
3781        a2 = a1;
3782        a1 = 0;
3783        goto gen_simd;
3784    gen_simd:
3785        tcg_debug_assert(insn != OPC_UD2);
3786        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3787        break;
3788
3789    case INDEX_op_cmp_vec:
3790        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3791        break;
3792
3793    case INDEX_op_cmpsel_vec:
3794        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3795                           args[3], args[4], args[5]);
3796        break;
3797
3798    case INDEX_op_andc_vec:
3799        insn = OPC_PANDN;
3800        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3801        break;
3802
3803    case INDEX_op_shli_vec:
3804        insn = shift_imm_insn[vece];
3805        sub = 6;
3806        goto gen_shift;
3807    case INDEX_op_shri_vec:
3808        insn = shift_imm_insn[vece];
3809        sub = 2;
3810        goto gen_shift;
3811    case INDEX_op_sari_vec:
3812        if (vece == MO_64) {
3813            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3814        } else {
3815            insn = shift_imm_insn[vece];
3816        }
3817        sub = 4;
3818        goto gen_shift;
3819    case INDEX_op_rotli_vec:
3820        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3821        if (vece == MO_64) {
3822            insn |= P_VEXW;
3823        }
3824        sub = 1;
3825        goto gen_shift;
3826    gen_shift:
3827        tcg_debug_assert(vece != MO_8);
3828        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3829        tcg_out8(s, a2);
3830        break;
3831
3832    case INDEX_op_ld_vec:
3833        tcg_out_ld(s, type, a0, a1, a2);
3834        break;
3835    case INDEX_op_st_vec:
3836        tcg_out_st(s, type, a0, a1, a2);
3837        break;
3838    case INDEX_op_dupm_vec:
3839        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3840        break;
3841
3842    case INDEX_op_x86_shufps_vec:
3843        insn = OPC_SHUFPS;
3844        sub = args[3];
3845        goto gen_simd_imm8;
3846    case INDEX_op_x86_blend_vec:
3847        if (vece == MO_16) {
3848            insn = OPC_PBLENDW;
3849        } else if (vece == MO_32) {
3850            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3851        } else {
3852            g_assert_not_reached();
3853        }
3854        sub = args[3];
3855        goto gen_simd_imm8;
3856    case INDEX_op_x86_vperm2i128_vec:
3857        insn = OPC_VPERM2I128;
3858        sub = args[3];
3859        goto gen_simd_imm8;
3860    case INDEX_op_x86_vpshldi_vec:
3861        insn = vpshldi_insn[vece];
3862        sub = args[3];
3863        goto gen_simd_imm8;
3864
3865    case INDEX_op_not_vec:
3866        insn = OPC_VPTERNLOGQ;
3867        a2 = a1;
3868        sub = 0x33; /* !B */
3869        goto gen_simd_imm8;
3870    case INDEX_op_nor_vec:
3871        insn = OPC_VPTERNLOGQ;
3872        sub = 0x11; /* norCB */
3873        goto gen_simd_imm8;
3874    case INDEX_op_nand_vec:
3875        insn = OPC_VPTERNLOGQ;
3876        sub = 0x77; /* nandCB */
3877        goto gen_simd_imm8;
3878    case INDEX_op_eqv_vec:
3879        insn = OPC_VPTERNLOGQ;
3880        sub = 0x99; /* xnorCB */
3881        goto gen_simd_imm8;
3882    case INDEX_op_orc_vec:
3883        insn = OPC_VPTERNLOGQ;
3884        sub = 0xdd; /* orB!C */
3885        goto gen_simd_imm8;
3886
3887    case INDEX_op_bitsel_vec:
3888        insn = OPC_VPTERNLOGQ;
3889        a3 = args[3];
3890        if (a0 == a1) {
3891            a1 = a2;
3892            a2 = a3;
3893            sub = 0xca; /* A?B:C */
3894        } else if (a0 == a2) {
3895            a2 = a3;
3896            sub = 0xe2; /* B?A:C */
3897        } else {
3898            tcg_out_mov(s, type, a0, a3);
3899            sub = 0xb8; /* B?C:A */
3900        }
3901        goto gen_simd_imm8;
3902
3903    gen_simd_imm8:
3904        tcg_debug_assert(insn != OPC_UD2);
3905        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3906        tcg_out8(s, sub);
3907        break;
3908
3909    case INDEX_op_x86_psrldq_vec:
3910        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3911        tcg_out8(s, a2);
3912        break;
3913
3914    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3915    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3916    default:
3917        g_assert_not_reached();
3918    }
3919}
3920
3921static TCGConstraintSetIndex
3922tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3923{
3924    switch (op) {
3925    case INDEX_op_goto_ptr:
3926        return C_O0_I1(r);
3927
3928    case INDEX_op_ld8u_i32:
3929    case INDEX_op_ld8u_i64:
3930    case INDEX_op_ld8s_i32:
3931    case INDEX_op_ld8s_i64:
3932    case INDEX_op_ld16u_i32:
3933    case INDEX_op_ld16u_i64:
3934    case INDEX_op_ld16s_i32:
3935    case INDEX_op_ld16s_i64:
3936    case INDEX_op_ld_i32:
3937    case INDEX_op_ld32u_i64:
3938    case INDEX_op_ld32s_i64:
3939    case INDEX_op_ld_i64:
3940        return C_O1_I1(r, r);
3941
3942    case INDEX_op_st8_i32:
3943    case INDEX_op_st8_i64:
3944        return C_O0_I2(qi, r);
3945
3946    case INDEX_op_st16_i32:
3947    case INDEX_op_st16_i64:
3948    case INDEX_op_st_i32:
3949    case INDEX_op_st32_i64:
3950        return C_O0_I2(ri, r);
3951
3952    case INDEX_op_st_i64:
3953        return C_O0_I2(re, r);
3954
3955    case INDEX_op_bswap16_i32:
3956    case INDEX_op_bswap16_i64:
3957    case INDEX_op_bswap32_i32:
3958    case INDEX_op_bswap32_i64:
3959    case INDEX_op_bswap64_i64:
3960    case INDEX_op_extrh_i64_i32:
3961        return C_O1_I1(r, 0);
3962
3963    case INDEX_op_ext_i32_i64:
3964    case INDEX_op_extu_i32_i64:
3965    case INDEX_op_extrl_i64_i32:
3966    case INDEX_op_extract_i32:
3967    case INDEX_op_extract_i64:
3968    case INDEX_op_sextract_i32:
3969    case INDEX_op_sextract_i64:
3970        return C_O1_I1(r, r);
3971
3972    case INDEX_op_extract2_i32:
3973    case INDEX_op_extract2_i64:
3974        return C_O1_I2(r, 0, r);
3975
3976    case INDEX_op_deposit_i32:
3977    case INDEX_op_deposit_i64:
3978        return C_O1_I2(q, 0, qi);
3979
3980    case INDEX_op_movcond_i32:
3981    case INDEX_op_movcond_i64:
3982        return C_O1_I4(r, r, reT, r, 0);
3983
3984    case INDEX_op_add2_i32:
3985    case INDEX_op_add2_i64:
3986    case INDEX_op_sub2_i32:
3987    case INDEX_op_sub2_i64:
3988        return C_N1_O1_I4(r, r, 0, 1, re, re);
3989
3990    case INDEX_op_qemu_ld_i32:
3991        return C_O1_I1(r, L);
3992
3993    case INDEX_op_qemu_st_i32:
3994        return C_O0_I2(L, L);
3995    case INDEX_op_qemu_st8_i32:
3996        return C_O0_I2(s, L);
3997
3998    case INDEX_op_qemu_ld_i64:
3999        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
4000
4001    case INDEX_op_qemu_st_i64:
4002        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
4003
4004    case INDEX_op_qemu_ld_i128:
4005        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4006        return C_O2_I1(r, r, L);
4007    case INDEX_op_qemu_st_i128:
4008        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4009        return C_O0_I3(L, L, L);
4010
4011    case INDEX_op_brcond2_i32:
4012        return C_O0_I4(r, r, ri, ri);
4013
4014    case INDEX_op_setcond2_i32:
4015        return C_O1_I4(r, r, r, ri, ri);
4016
4017    case INDEX_op_ld_vec:
4018    case INDEX_op_dupm_vec:
4019        return C_O1_I1(x, r);
4020
4021    case INDEX_op_st_vec:
4022        return C_O0_I2(x, r);
4023
4024    case INDEX_op_add_vec:
4025    case INDEX_op_sub_vec:
4026    case INDEX_op_mul_vec:
4027    case INDEX_op_and_vec:
4028    case INDEX_op_or_vec:
4029    case INDEX_op_xor_vec:
4030    case INDEX_op_andc_vec:
4031    case INDEX_op_orc_vec:
4032    case INDEX_op_nand_vec:
4033    case INDEX_op_nor_vec:
4034    case INDEX_op_eqv_vec:
4035    case INDEX_op_ssadd_vec:
4036    case INDEX_op_usadd_vec:
4037    case INDEX_op_sssub_vec:
4038    case INDEX_op_ussub_vec:
4039    case INDEX_op_smin_vec:
4040    case INDEX_op_umin_vec:
4041    case INDEX_op_smax_vec:
4042    case INDEX_op_umax_vec:
4043    case INDEX_op_shlv_vec:
4044    case INDEX_op_shrv_vec:
4045    case INDEX_op_sarv_vec:
4046    case INDEX_op_rotlv_vec:
4047    case INDEX_op_rotrv_vec:
4048    case INDEX_op_shls_vec:
4049    case INDEX_op_shrs_vec:
4050    case INDEX_op_sars_vec:
4051    case INDEX_op_cmp_vec:
4052    case INDEX_op_x86_shufps_vec:
4053    case INDEX_op_x86_blend_vec:
4054    case INDEX_op_x86_packss_vec:
4055    case INDEX_op_x86_packus_vec:
4056    case INDEX_op_x86_vperm2i128_vec:
4057    case INDEX_op_x86_punpckl_vec:
4058    case INDEX_op_x86_punpckh_vec:
4059    case INDEX_op_x86_vpshldi_vec:
4060#if TCG_TARGET_REG_BITS == 32
4061    case INDEX_op_dup2_vec:
4062#endif
4063        return C_O1_I2(x, x, x);
4064
4065    case INDEX_op_abs_vec:
4066    case INDEX_op_dup_vec:
4067    case INDEX_op_not_vec:
4068    case INDEX_op_shli_vec:
4069    case INDEX_op_shri_vec:
4070    case INDEX_op_sari_vec:
4071    case INDEX_op_rotli_vec:
4072    case INDEX_op_x86_psrldq_vec:
4073        return C_O1_I1(x, x);
4074
4075    case INDEX_op_x86_vpshldv_vec:
4076    case INDEX_op_x86_vpshrdv_vec:
4077        return C_O1_I3(x, 0, x, x);
4078
4079    case INDEX_op_bitsel_vec:
4080        return C_O1_I3(x, x, x, x);
4081    case INDEX_op_cmpsel_vec:
4082        return C_O1_I4(x, x, x, xO, x);
4083
4084    default:
4085        return C_NotImplemented;
4086    }
4087}
4088
4089int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
4090{
4091    switch (opc) {
4092    case INDEX_op_add_vec:
4093    case INDEX_op_sub_vec:
4094    case INDEX_op_and_vec:
4095    case INDEX_op_or_vec:
4096    case INDEX_op_xor_vec:
4097    case INDEX_op_andc_vec:
4098    case INDEX_op_orc_vec:
4099    case INDEX_op_nand_vec:
4100    case INDEX_op_nor_vec:
4101    case INDEX_op_eqv_vec:
4102    case INDEX_op_not_vec:
4103    case INDEX_op_bitsel_vec:
4104        return 1;
4105    case INDEX_op_cmp_vec:
4106    case INDEX_op_cmpsel_vec:
4107        return -1;
4108
4109    case INDEX_op_rotli_vec:
4110        return have_avx512vl && vece >= MO_32 ? 1 : -1;
4111
4112    case INDEX_op_shli_vec:
4113    case INDEX_op_shri_vec:
4114        /* We must expand the operation for MO_8.  */
4115        return vece == MO_8 ? -1 : 1;
4116
4117    case INDEX_op_sari_vec:
4118        switch (vece) {
4119        case MO_8:
4120            return -1;
4121        case MO_16:
4122        case MO_32:
4123            return 1;
4124        case MO_64:
4125            if (have_avx512vl) {
4126                return 1;
4127            }
4128            /*
4129             * We can emulate this for MO_64, but it does not pay off
4130             * unless we're producing at least 4 values.
4131             */
4132            return type >= TCG_TYPE_V256 ? -1 : 0;
4133        }
4134        return 0;
4135
4136    case INDEX_op_shls_vec:
4137    case INDEX_op_shrs_vec:
4138        return vece >= MO_16;
4139    case INDEX_op_sars_vec:
4140        switch (vece) {
4141        case MO_16:
4142        case MO_32:
4143            return 1;
4144        case MO_64:
4145            return have_avx512vl;
4146        }
4147        return 0;
4148    case INDEX_op_rotls_vec:
4149        return vece >= MO_16 ? -1 : 0;
4150
4151    case INDEX_op_shlv_vec:
4152    case INDEX_op_shrv_vec:
4153        switch (vece) {
4154        case MO_16:
4155            return have_avx512bw;
4156        case MO_32:
4157        case MO_64:
4158            return have_avx2;
4159        }
4160        return 0;
4161    case INDEX_op_sarv_vec:
4162        switch (vece) {
4163        case MO_16:
4164            return have_avx512bw;
4165        case MO_32:
4166            return have_avx2;
4167        case MO_64:
4168            return have_avx512vl;
4169        }
4170        return 0;
4171    case INDEX_op_rotlv_vec:
4172    case INDEX_op_rotrv_vec:
4173        switch (vece) {
4174        case MO_16:
4175            return have_avx512vbmi2 ? -1 : 0;
4176        case MO_32:
4177        case MO_64:
4178            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4179        }
4180        return 0;
4181
4182    case INDEX_op_mul_vec:
4183        switch (vece) {
4184        case MO_8:
4185            return -1;
4186        case MO_64:
4187            return have_avx512dq;
4188        }
4189        return 1;
4190
4191    case INDEX_op_ssadd_vec:
4192    case INDEX_op_usadd_vec:
4193    case INDEX_op_sssub_vec:
4194    case INDEX_op_ussub_vec:
4195        return vece <= MO_16;
4196    case INDEX_op_smin_vec:
4197    case INDEX_op_smax_vec:
4198    case INDEX_op_umin_vec:
4199    case INDEX_op_umax_vec:
4200    case INDEX_op_abs_vec:
4201        return vece <= MO_32 || have_avx512vl;
4202
4203    default:
4204        return 0;
4205    }
4206}
4207
4208static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4209                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4210{
4211    uint8_t mask;
4212
4213    tcg_debug_assert(vece == MO_8);
4214    if (right) {
4215        mask = 0xff >> imm;
4216        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4217    } else {
4218        mask = 0xff << imm;
4219        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4220    }
4221    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4222}
4223
4224static void expand_vec_sari(TCGType type, unsigned vece,
4225                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4226{
4227    TCGv_vec t1, t2;
4228
4229    switch (vece) {
4230    case MO_8:
4231        /* Unpack to 16-bit, shift, and repack.  */
4232        t1 = tcg_temp_new_vec(type);
4233        t2 = tcg_temp_new_vec(type);
4234        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4235                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4236        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4237                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4238        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4239        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4240        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4241                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4242        tcg_temp_free_vec(t1);
4243        tcg_temp_free_vec(t2);
4244        break;
4245
4246    case MO_64:
4247        t1 = tcg_temp_new_vec(type);
4248        if (imm <= 32) {
4249            /*
4250             * We can emulate a small sign extend by performing an arithmetic
4251             * 32-bit shift and overwriting the high half of a 64-bit logical
4252             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4253             * does not, so we have to bound the smaller shift -- we get the
4254             * same result in the high half either way.
4255             */
4256            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4257            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4258            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4259                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4260                      tcgv_vec_arg(t1), 0xaa);
4261        } else {
4262            /* Otherwise we will need to use a compare vs 0 to produce
4263             * the sign-extend, shift and merge.
4264             */
4265            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4266                            tcg_constant_vec(type, MO_64, 0), v1);
4267            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4268            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4269            tcg_gen_or_vec(MO_64, v0, v0, t1);
4270        }
4271        tcg_temp_free_vec(t1);
4272        break;
4273
4274    default:
4275        g_assert_not_reached();
4276    }
4277}
4278
4279static void expand_vec_rotli(TCGType type, unsigned vece,
4280                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4281{
4282    TCGv_vec t;
4283
4284    if (vece != MO_8 && have_avx512vbmi2) {
4285        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4286                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4287        return;
4288    }
4289
4290    t = tcg_temp_new_vec(type);
4291    tcg_gen_shli_vec(vece, t, v1, imm);
4292    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4293    tcg_gen_or_vec(vece, v0, v0, t);
4294    tcg_temp_free_vec(t);
4295}
4296
4297static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4298                            TCGv_vec v1, TCGv_vec sh, bool right)
4299{
4300    TCGv_vec t;
4301
4302    if (have_avx512vbmi2) {
4303        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4304                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4305                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4306        return;
4307    }
4308
4309    t = tcg_temp_new_vec(type);
4310    tcg_gen_dupi_vec(vece, t, 8 << vece);
4311    tcg_gen_sub_vec(vece, t, t, sh);
4312    if (right) {
4313        tcg_gen_shlv_vec(vece, t, v1, t);
4314        tcg_gen_shrv_vec(vece, v0, v1, sh);
4315    } else {
4316        tcg_gen_shrv_vec(vece, t, v1, t);
4317        tcg_gen_shlv_vec(vece, v0, v1, sh);
4318    }
4319    tcg_gen_or_vec(vece, v0, v0, t);
4320    tcg_temp_free_vec(t);
4321}
4322
4323static void expand_vec_rotls(TCGType type, unsigned vece,
4324                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4325{
4326    TCGv_vec t = tcg_temp_new_vec(type);
4327
4328    tcg_debug_assert(vece != MO_8);
4329
4330    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4331        tcg_gen_dup_i32_vec(vece, t, lsh);
4332        if (vece >= MO_32) {
4333            tcg_gen_rotlv_vec(vece, v0, v1, t);
4334        } else {
4335            expand_vec_rotv(type, vece, v0, v1, t, false);
4336        }
4337    } else {
4338        TCGv_i32 rsh = tcg_temp_new_i32();
4339
4340        tcg_gen_neg_i32(rsh, lsh);
4341        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4342        tcg_gen_shls_vec(vece, t, v1, lsh);
4343        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4344        tcg_gen_or_vec(vece, v0, v0, t);
4345
4346        tcg_temp_free_i32(rsh);
4347    }
4348
4349    tcg_temp_free_vec(t);
4350}
4351
4352static void expand_vec_mul(TCGType type, unsigned vece,
4353                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4354{
4355    TCGv_vec t1, t2, t3, t4, zero;
4356
4357    tcg_debug_assert(vece == MO_8);
4358
4359    /*
4360     * Unpack v1 bytes to words, 0 | x.
4361     * Unpack v2 bytes to words, y | 0.
4362     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4363     * Shift logical right by 8 bits to clear the high 8 bytes before
4364     * using an unsigned saturated pack.
4365     *
4366     * The difference between the V64, V128 and V256 cases is merely how
4367     * we distribute the expansion between temporaries.
4368     */
4369    switch (type) {
4370    case TCG_TYPE_V64:
4371        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4372        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4373        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4374        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4375                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4376        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4377                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4378        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4379        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4380        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4381                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4382        tcg_temp_free_vec(t1);
4383        tcg_temp_free_vec(t2);
4384        break;
4385
4386    case TCG_TYPE_V128:
4387    case TCG_TYPE_V256:
4388        t1 = tcg_temp_new_vec(type);
4389        t2 = tcg_temp_new_vec(type);
4390        t3 = tcg_temp_new_vec(type);
4391        t4 = tcg_temp_new_vec(type);
4392        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4393        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4394                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4395        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4396                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4397        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4398                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4399        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4400                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4401        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4402        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4403        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4404        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4405        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4406                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4407        tcg_temp_free_vec(t1);
4408        tcg_temp_free_vec(t2);
4409        tcg_temp_free_vec(t3);
4410        tcg_temp_free_vec(t4);
4411        break;
4412
4413    default:
4414        g_assert_not_reached();
4415    }
4416}
4417
4418static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4419                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4420{
4421    /*
4422     * Without AVX512, there are no 64-bit unsigned comparisons.
4423     * We must bias the inputs so that they become signed.
4424     * All other swapping and inversion are handled during code generation.
4425     */
4426    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4427        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4428        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4429        TCGv_vec t1 = tcg_temp_new_vec(type);
4430        TCGv_vec t2 = tcg_temp_new_vec(type);
4431        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4432
4433        tcg_gen_sub_vec(vece, t1, v1, t3);
4434        tcg_gen_sub_vec(vece, t2, v2, t3);
4435        *a1 = tcgv_vec_arg(t1);
4436        *a2 = tcgv_vec_arg(t2);
4437        cond = tcg_signed_cond(cond);
4438    }
4439    return cond;
4440}
4441
4442static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4443                           TCGArg a1, TCGArg a2, TCGCond cond)
4444{
4445    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4446    /* Expand directly; do not recurse.  */
4447    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4448}
4449
4450static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4451                              TCGArg a1, TCGArg a2,
4452                              TCGArg a3, TCGArg a4, TCGCond cond)
4453{
4454    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4455    /* Expand directly; do not recurse.  */
4456    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4457}
4458
4459void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4460                       TCGArg a0, ...)
4461{
4462    va_list va;
4463    TCGArg a1, a2, a3, a4, a5;
4464    TCGv_vec v0, v1, v2;
4465
4466    va_start(va, a0);
4467    a1 = va_arg(va, TCGArg);
4468    a2 = va_arg(va, TCGArg);
4469    v0 = temp_tcgv_vec(arg_temp(a0));
4470    v1 = temp_tcgv_vec(arg_temp(a1));
4471
4472    switch (opc) {
4473    case INDEX_op_shli_vec:
4474        expand_vec_shi(type, vece, false, v0, v1, a2);
4475        break;
4476    case INDEX_op_shri_vec:
4477        expand_vec_shi(type, vece, true, v0, v1, a2);
4478        break;
4479    case INDEX_op_sari_vec:
4480        expand_vec_sari(type, vece, v0, v1, a2);
4481        break;
4482
4483    case INDEX_op_rotli_vec:
4484        expand_vec_rotli(type, vece, v0, v1, a2);
4485        break;
4486
4487    case INDEX_op_rotls_vec:
4488        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4489        break;
4490
4491    case INDEX_op_rotlv_vec:
4492        v2 = temp_tcgv_vec(arg_temp(a2));
4493        expand_vec_rotv(type, vece, v0, v1, v2, false);
4494        break;
4495    case INDEX_op_rotrv_vec:
4496        v2 = temp_tcgv_vec(arg_temp(a2));
4497        expand_vec_rotv(type, vece, v0, v1, v2, true);
4498        break;
4499
4500    case INDEX_op_mul_vec:
4501        v2 = temp_tcgv_vec(arg_temp(a2));
4502        expand_vec_mul(type, vece, v0, v1, v2);
4503        break;
4504
4505    case INDEX_op_cmp_vec:
4506        a3 = va_arg(va, TCGArg);
4507        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4508        break;
4509
4510    case INDEX_op_cmpsel_vec:
4511        a3 = va_arg(va, TCGArg);
4512        a4 = va_arg(va, TCGArg);
4513        a5 = va_arg(va, TCGArg);
4514        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4515        break;
4516
4517    default:
4518        break;
4519    }
4520
4521    va_end(va);
4522}
4523
4524static const int tcg_target_callee_save_regs[] = {
4525#if TCG_TARGET_REG_BITS == 64
4526    TCG_REG_RBP,
4527    TCG_REG_RBX,
4528#if defined(_WIN64)
4529    TCG_REG_RDI,
4530    TCG_REG_RSI,
4531#endif
4532    TCG_REG_R12,
4533    TCG_REG_R13,
4534    TCG_REG_R14, /* Currently used for the global env. */
4535    TCG_REG_R15,
4536#else
4537    TCG_REG_EBP, /* Currently used for the global env. */
4538    TCG_REG_EBX,
4539    TCG_REG_ESI,
4540    TCG_REG_EDI,
4541#endif
4542};
4543
4544/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4545   and tcg_register_jit.  */
4546
4547#define PUSH_SIZE \
4548    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4549     * (TCG_TARGET_REG_BITS / 8))
4550
4551#define FRAME_SIZE \
4552    ((PUSH_SIZE \
4553      + TCG_STATIC_CALL_ARGS_SIZE \
4554      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4555      + TCG_TARGET_STACK_ALIGN - 1) \
4556     & ~(TCG_TARGET_STACK_ALIGN - 1))
4557
4558/* Generate global QEMU prologue and epilogue code */
4559static void tcg_target_qemu_prologue(TCGContext *s)
4560{
4561    int i, stack_addend;
4562
4563    /* TB prologue */
4564
4565    /* Reserve some stack space, also for TCG temps.  */
4566    stack_addend = FRAME_SIZE - PUSH_SIZE;
4567    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4568                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4569
4570    /* Save all callee saved registers.  */
4571    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4572        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4573    }
4574
4575    if (!tcg_use_softmmu && guest_base) {
4576        int seg = setup_guest_base_seg();
4577        if (seg != 0) {
4578            x86_guest_base.seg = seg;
4579        } else if (guest_base == (int32_t)guest_base) {
4580            x86_guest_base.ofs = guest_base;
4581        } else {
4582            assert(TCG_TARGET_REG_BITS == 64);
4583            /* Choose R12 because, as a base, it requires a SIB byte. */
4584            x86_guest_base.index = TCG_REG_R12;
4585            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4586            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4587        }
4588    }
4589
4590    if (TCG_TARGET_REG_BITS == 32) {
4591        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4592                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4593        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4594        /* jmp *tb.  */
4595        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4596                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4597                             + stack_addend);
4598    } else {
4599        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4600        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4601        /* jmp *tb.  */
4602        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4603    }
4604
4605    /*
4606     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4607     * and fall through to the rest of the epilogue.
4608     */
4609    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4610    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4611
4612    /* TB epilogue */
4613    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4614
4615    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4616
4617    if (have_avx2) {
4618        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4619    }
4620    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4621        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4622    }
4623    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4624}
4625
4626static void tcg_out_tb_start(TCGContext *s)
4627{
4628    /* nothing to do */
4629}
4630
4631static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4632{
4633    memset(p, 0x90, count);
4634}
4635
4636static void tcg_target_init(TCGContext *s)
4637{
4638    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4639    if (TCG_TARGET_REG_BITS == 64) {
4640        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4641    }
4642    if (have_avx1) {
4643        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4644        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4645    }
4646    if (have_avx2) {
4647        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4648    }
4649
4650    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4651    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4652    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4653    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4654    if (TCG_TARGET_REG_BITS == 64) {
4655#if !defined(_WIN64)
4656        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4657        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4658#endif
4659        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4660        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4661        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4662        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4663    }
4664
4665    s->reserved_regs = 0;
4666    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4667    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4668#ifdef _WIN64
4669    /* These are call saved, and we don't save them, so don't use them. */
4670    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4671    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4672    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4673    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4674    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4675    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4676    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4677    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4678    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4679    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4680#endif
4681}
4682
4683typedef struct {
4684    DebugFrameHeader h;
4685    uint8_t fde_def_cfa[4];
4686    uint8_t fde_reg_ofs[14];
4687} DebugFrame;
4688
4689/* We're expecting a 2 byte uleb128 encoded value.  */
4690QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4691
4692#if !defined(__ELF__)
4693    /* Host machine without ELF. */
4694#elif TCG_TARGET_REG_BITS == 64
4695#define ELF_HOST_MACHINE EM_X86_64
4696static const DebugFrame debug_frame = {
4697    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4698    .h.cie.id = -1,
4699    .h.cie.version = 1,
4700    .h.cie.code_align = 1,
4701    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4702    .h.cie.return_column = 16,
4703
4704    /* Total FDE size does not include the "len" member.  */
4705    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4706
4707    .fde_def_cfa = {
4708        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4709        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4710        (FRAME_SIZE >> 7)
4711    },
4712    .fde_reg_ofs = {
4713        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4714        /* The following ordering must match tcg_target_callee_save_regs.  */
4715        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4716        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4717        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4718        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4719        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4720        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4721    }
4722};
4723#else
4724#define ELF_HOST_MACHINE EM_386
4725static const DebugFrame debug_frame = {
4726    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4727    .h.cie.id = -1,
4728    .h.cie.version = 1,
4729    .h.cie.code_align = 1,
4730    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4731    .h.cie.return_column = 8,
4732
4733    /* Total FDE size does not include the "len" member.  */
4734    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4735
4736    .fde_def_cfa = {
4737        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4738        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4739        (FRAME_SIZE >> 7)
4740    },
4741    .fde_reg_ofs = {
4742        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4743        /* The following ordering must match tcg_target_callee_save_regs.  */
4744        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4745        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4746        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4747        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4748    }
4749};
4750#endif
4751
4752#if defined(ELF_HOST_MACHINE)
4753void tcg_register_jit(const void *buf, size_t buf_size)
4754{
4755    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4756}
4757#endif
4758