xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 33aba058c8fcc9b1581b03a1fbac45d8d91baac6)
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_STC         (0xf9)
428#define OPC_TESTB	(0x84)
429#define OPC_TESTL	(0x85)
430#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
431#define OPC_UD2         (0x0b | P_EXT)
432#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
433#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
434#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
435#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
436#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
437#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
438#define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
439#define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
440#define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
441#define OPC_VPCMPUW     (0x3e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
442#define OPC_VPCMPD      (0x1f | P_EXT3A | P_DATA16 | P_EVEX)
443#define OPC_VPCMPUD     (0x1e | P_EXT3A | P_DATA16 | P_EVEX)
444#define OPC_VPCMPQ      (0x1f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPCMPUQ     (0x1e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
446#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
447#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
448#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
449#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
450#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
451#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
452#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
453#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
454#define OPC_VPMOVM2B    (0x28 | P_EXT38 | P_SIMDF3 | P_EVEX)
455#define OPC_VPMOVM2W    (0x28 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
456#define OPC_VPMOVM2D    (0x38 | P_EXT38 | P_SIMDF3 | P_EVEX)
457#define OPC_VPMOVM2Q    (0x38 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
458#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
459#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
460#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
461#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
462#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
463#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
464#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
465#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
466#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
467#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
468#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
469#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
470#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
471#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
472#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
473#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
474#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
475#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
476#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
477#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
478#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
479#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
480#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
481#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
482#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
483#define OPC_VPTESTMB    (0x26 | P_EXT38 | P_DATA16 | P_EVEX)
484#define OPC_VPTESTMW    (0x26 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
485#define OPC_VPTESTMD    (0x27 | P_EXT38 | P_DATA16 | P_EVEX)
486#define OPC_VPTESTMQ    (0x27 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
487#define OPC_VPTESTNMB   (0x26 | P_EXT38 | P_SIMDF3 | P_EVEX)
488#define OPC_VPTESTNMW   (0x26 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
489#define OPC_VPTESTNMD   (0x27 | P_EXT38 | P_SIMDF3 | P_EVEX)
490#define OPC_VPTESTNMQ   (0x27 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
491#define OPC_VZEROUPPER  (0x77 | P_EXT)
492#define OPC_XCHG_ax_r32	(0x90)
493#define OPC_XCHG_EvGv   (0x87)
494
495#define OPC_GRP3_Eb     (0xf6)
496#define OPC_GRP3_Ev     (0xf7)
497#define OPC_GRP5        (0xff)
498#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
499#define OPC_GRPBT       (0xba | P_EXT)
500
501#define OPC_GRPBT_BT    4
502#define OPC_GRPBT_BTS   5
503#define OPC_GRPBT_BTR   6
504#define OPC_GRPBT_BTC   7
505
506/* Group 1 opcode extensions for 0x80-0x83.
507   These are also used as modifiers for OPC_ARITH.  */
508#define ARITH_ADD 0
509#define ARITH_OR  1
510#define ARITH_ADC 2
511#define ARITH_SBB 3
512#define ARITH_AND 4
513#define ARITH_SUB 5
514#define ARITH_XOR 6
515#define ARITH_CMP 7
516
517/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
518#define SHIFT_ROL 0
519#define SHIFT_ROR 1
520#define SHIFT_SHL 4
521#define SHIFT_SHR 5
522#define SHIFT_SAR 7
523
524/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
525#define EXT3_TESTi 0
526#define EXT3_NOT   2
527#define EXT3_NEG   3
528#define EXT3_MUL   4
529#define EXT3_IMUL  5
530#define EXT3_DIV   6
531#define EXT3_IDIV  7
532
533/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
534#define EXT5_INC_Ev	0
535#define EXT5_DEC_Ev	1
536#define EXT5_CALLN_Ev	2
537#define EXT5_JMPN_Ev	4
538
539/* Condition codes to be added to OPC_JCC_{long,short}.  */
540#define JCC_JMP (-1)
541#define JCC_JO  0x0
542#define JCC_JNO 0x1
543#define JCC_JB  0x2
544#define JCC_JAE 0x3
545#define JCC_JE  0x4
546#define JCC_JNE 0x5
547#define JCC_JBE 0x6
548#define JCC_JA  0x7
549#define JCC_JS  0x8
550#define JCC_JNS 0x9
551#define JCC_JP  0xa
552#define JCC_JNP 0xb
553#define JCC_JL  0xc
554#define JCC_JGE 0xd
555#define JCC_JLE 0xe
556#define JCC_JG  0xf
557
558static const uint8_t tcg_cond_to_jcc[] = {
559    [TCG_COND_EQ] = JCC_JE,
560    [TCG_COND_NE] = JCC_JNE,
561    [TCG_COND_LT] = JCC_JL,
562    [TCG_COND_GE] = JCC_JGE,
563    [TCG_COND_LE] = JCC_JLE,
564    [TCG_COND_GT] = JCC_JG,
565    [TCG_COND_LTU] = JCC_JB,
566    [TCG_COND_GEU] = JCC_JAE,
567    [TCG_COND_LEU] = JCC_JBE,
568    [TCG_COND_GTU] = JCC_JA,
569    [TCG_COND_TSTEQ] = JCC_JE,
570    [TCG_COND_TSTNE] = JCC_JNE,
571};
572
573#if TCG_TARGET_REG_BITS == 64
574static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
575{
576    int rex;
577
578    if (opc & P_GS) {
579        tcg_out8(s, 0x65);
580    }
581    if (opc & P_DATA16) {
582        /* We should never be asking for both 16 and 64-bit operation.  */
583        tcg_debug_assert((opc & P_REXW) == 0);
584        tcg_out8(s, 0x66);
585    }
586    if (opc & P_SIMDF3) {
587        tcg_out8(s, 0xf3);
588    } else if (opc & P_SIMDF2) {
589        tcg_out8(s, 0xf2);
590    }
591
592    rex = 0;
593    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
594    rex |= (r & 8) >> 1;                /* REX.R */
595    rex |= (x & 8) >> 2;                /* REX.X */
596    rex |= (rm & 8) >> 3;               /* REX.B */
597
598    /* P_REXB_{R,RM} indicates that the given register is the low byte.
599       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
600       as otherwise the encoding indicates %[abcd]h.  Note that the values
601       that are ORed in merely indicate that the REX byte must be present;
602       those bits get discarded in output.  */
603    rex |= opc & (r >= 4 ? P_REXB_R : 0);
604    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
605
606    if (rex) {
607        tcg_out8(s, (uint8_t)(rex | 0x40));
608    }
609
610    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
611        tcg_out8(s, 0x0f);
612        if (opc & P_EXT38) {
613            tcg_out8(s, 0x38);
614        } else if (opc & P_EXT3A) {
615            tcg_out8(s, 0x3a);
616        }
617    }
618
619    tcg_out8(s, opc);
620}
621#else
622static void tcg_out_opc(TCGContext *s, int opc)
623{
624    if (opc & P_DATA16) {
625        tcg_out8(s, 0x66);
626    }
627    if (opc & P_SIMDF3) {
628        tcg_out8(s, 0xf3);
629    } else if (opc & P_SIMDF2) {
630        tcg_out8(s, 0xf2);
631    }
632    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
633        tcg_out8(s, 0x0f);
634        if (opc & P_EXT38) {
635            tcg_out8(s, 0x38);
636        } else if (opc & P_EXT3A) {
637            tcg_out8(s, 0x3a);
638        }
639    }
640    tcg_out8(s, opc);
641}
642/* Discard the register arguments to tcg_out_opc early, so as not to penalize
643   the 32-bit compilation paths.  This method works with all versions of gcc,
644   whereas relying on optimization may not be able to exclude them.  */
645#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
646#endif
647
648static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
649{
650    tcg_out_opc(s, opc, r, rm, 0);
651    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
652}
653
654static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
655                            int rm, int index)
656{
657    int tmp;
658
659    if (opc & P_GS) {
660        tcg_out8(s, 0x65);
661    }
662    /* Use the two byte form if possible, which cannot encode
663       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
664    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
665        && ((rm | index) & 8) == 0) {
666        /* Two byte VEX prefix.  */
667        tcg_out8(s, 0xc5);
668
669        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
670    } else {
671        /* Three byte VEX prefix.  */
672        tcg_out8(s, 0xc4);
673
674        /* VEX.m-mmmm */
675        if (opc & P_EXT3A) {
676            tmp = 3;
677        } else if (opc & P_EXT38) {
678            tmp = 2;
679        } else if (opc & P_EXT) {
680            tmp = 1;
681        } else {
682            g_assert_not_reached();
683        }
684        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
685        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
686        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
687        tcg_out8(s, tmp);
688
689        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
690    }
691
692    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
693    /* VEX.pp */
694    if (opc & P_DATA16) {
695        tmp |= 1;                          /* 0x66 */
696    } else if (opc & P_SIMDF3) {
697        tmp |= 2;                          /* 0xf3 */
698    } else if (opc & P_SIMDF2) {
699        tmp |= 3;                          /* 0xf2 */
700    }
701    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
702    tcg_out8(s, tmp);
703    tcg_out8(s, opc);
704}
705
706static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
707                             int rm, int index, int aaa, bool z)
708{
709    /* The entire 4-byte evex prefix; with R' and V' set. */
710    uint32_t p = 0x08041062;
711    int mm, pp;
712
713    tcg_debug_assert(have_avx512vl);
714
715    /* EVEX.mm */
716    if (opc & P_EXT3A) {
717        mm = 3;
718    } else if (opc & P_EXT38) {
719        mm = 2;
720    } else if (opc & P_EXT) {
721        mm = 1;
722    } else {
723        g_assert_not_reached();
724    }
725
726    /* EVEX.pp */
727    if (opc & P_DATA16) {
728        pp = 1;                          /* 0x66 */
729    } else if (opc & P_SIMDF3) {
730        pp = 2;                          /* 0xf3 */
731    } else if (opc & P_SIMDF2) {
732        pp = 3;                          /* 0xf2 */
733    } else {
734        pp = 0;
735    }
736
737    p = deposit32(p, 8, 2, mm);
738    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
739    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
740    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
741    p = deposit32(p, 16, 2, pp);
742    p = deposit32(p, 19, 4, ~v);
743    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
744    p = deposit32(p, 24, 3, aaa);
745    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
746    p = deposit32(p, 31, 1, z);
747
748    tcg_out32(s, p);
749    tcg_out8(s, opc);
750}
751
752static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
753{
754    if (opc & P_EVEX) {
755        tcg_out_evex_opc(s, opc, r, v, rm, 0, 0, false);
756    } else {
757        tcg_out_vex_opc(s, opc, r, v, rm, 0);
758    }
759    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
760}
761
762static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
763                                   int r, int v, int rm, TCGType type)
764{
765    if (type == TCG_TYPE_V256) {
766        opc |= P_VEXL;
767    }
768    tcg_out_vex_modrm(s, opc, r, v, rm);
769}
770
771static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
772                                    int rm, int aaa, bool z, TCGType type)
773{
774    if (type == TCG_TYPE_V256) {
775        opc |= P_VEXL;
776    }
777    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
778    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
779}
780
781/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
782   We handle either RM and INDEX missing with a negative value.  In 64-bit
783   mode for absolute addresses, ~RM is the size of the immediate operand
784   that will follow the instruction.  */
785
786static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
787                               int shift, intptr_t offset)
788{
789    int mod, len;
790
791    if (index < 0 && rm < 0) {
792        if (TCG_TARGET_REG_BITS == 64) {
793            /* Try for a rip-relative addressing mode.  This has replaced
794               the 32-bit-mode absolute addressing encoding.  */
795            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
796            intptr_t disp = offset - pc;
797            if (disp == (int32_t)disp) {
798                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
799                tcg_out32(s, disp);
800                return;
801            }
802
803            /* Try for an absolute address encoding.  This requires the
804               use of the MODRM+SIB encoding and is therefore larger than
805               rip-relative addressing.  */
806            if (offset == (int32_t)offset) {
807                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
808                tcg_out8(s, (4 << 3) | 5);
809                tcg_out32(s, offset);
810                return;
811            }
812
813            /* ??? The memory isn't directly addressable.  */
814            g_assert_not_reached();
815        } else {
816            /* Absolute address.  */
817            tcg_out8(s, (r << 3) | 5);
818            tcg_out32(s, offset);
819            return;
820        }
821    }
822
823    /* Find the length of the immediate addend.  Note that the encoding
824       that would be used for (%ebp) indicates absolute addressing.  */
825    if (rm < 0) {
826        mod = 0, len = 4, rm = 5;
827    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
828        mod = 0, len = 0;
829    } else if (offset == (int8_t)offset) {
830        mod = 0x40, len = 1;
831    } else {
832        mod = 0x80, len = 4;
833    }
834
835    /* Use a single byte MODRM format if possible.  Note that the encoding
836       that would be used for %esp is the escape to the two byte form.  */
837    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
838        /* Single byte MODRM format.  */
839        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
840    } else {
841        /* Two byte MODRM+SIB format.  */
842
843        /* Note that the encoding that would place %esp into the index
844           field indicates no index register.  In 64-bit mode, the REX.X
845           bit counts, so %r12 can be used as the index.  */
846        if (index < 0) {
847            index = 4;
848        } else {
849            tcg_debug_assert(index != TCG_REG_ESP);
850        }
851
852        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
853        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
854    }
855
856    if (len == 1) {
857        tcg_out8(s, offset);
858    } else if (len == 4) {
859        tcg_out32(s, offset);
860    }
861}
862
863static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
864                                     int index, int shift, intptr_t offset)
865{
866    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
867    tcg_out_sib_offset(s, r, rm, index, shift, offset);
868}
869
870static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
871                                         int rm, int index, int shift,
872                                         intptr_t offset)
873{
874    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
875    tcg_out_sib_offset(s, r, rm, index, shift, offset);
876}
877
878/* A simplification of the above with no index or shift.  */
879static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
880                                        int rm, intptr_t offset)
881{
882    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
883}
884
885static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
886                                            int v, int rm, intptr_t offset)
887{
888    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
889}
890
891/* Output an opcode with an expected reference to the constant pool.  */
892static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
893{
894    tcg_out_opc(s, opc, r, 0, 0);
895    /* Absolute for 32-bit, pc-relative for 64-bit.  */
896    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
897    tcg_out32(s, 0);
898}
899
900/* Output an opcode with an expected reference to the constant pool.  */
901static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
902{
903    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
904    /* Absolute for 32-bit, pc-relative for 64-bit.  */
905    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
906    tcg_out32(s, 0);
907}
908
909/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
910static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
911{
912    /* Propagate an opcode prefix, such as P_REXW.  */
913    int ext = subop & ~0x7;
914    subop &= 0x7;
915
916    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
917}
918
919static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
920{
921    int rexw = 0;
922
923    if (arg == ret) {
924        return true;
925    }
926    switch (type) {
927    case TCG_TYPE_I64:
928        rexw = P_REXW;
929        /* fallthru */
930    case TCG_TYPE_I32:
931        if (ret < 16) {
932            if (arg < 16) {
933                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
934            } else {
935                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
936            }
937        } else {
938            if (arg < 16) {
939                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
940            } else {
941                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
942            }
943        }
944        break;
945
946    case TCG_TYPE_V64:
947        tcg_debug_assert(ret >= 16 && arg >= 16);
948        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
949        break;
950    case TCG_TYPE_V128:
951        tcg_debug_assert(ret >= 16 && arg >= 16);
952        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
953        break;
954    case TCG_TYPE_V256:
955        tcg_debug_assert(ret >= 16 && arg >= 16);
956        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
957        break;
958
959    default:
960        g_assert_not_reached();
961    }
962    return true;
963}
964
965static const int avx2_dup_insn[4] = {
966    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
967    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
968};
969
970static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
971                            TCGReg r, TCGReg a)
972{
973    if (have_avx2) {
974        tcg_out_vex_modrm_type(s, avx2_dup_insn[vece], r, 0, a, type);
975    } else {
976        switch (vece) {
977        case MO_8:
978            /* ??? With zero in a register, use PSHUFB.  */
979            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
980            a = r;
981            /* FALLTHRU */
982        case MO_16:
983            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
984            a = r;
985            /* FALLTHRU */
986        case MO_32:
987            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
988            /* imm8 operand: all output lanes selected from input lane 0.  */
989            tcg_out8(s, 0);
990            break;
991        case MO_64:
992            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
993            break;
994        default:
995            g_assert_not_reached();
996        }
997    }
998    return true;
999}
1000
1001static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1002                             TCGReg r, TCGReg base, intptr_t offset)
1003{
1004    if (have_avx2) {
1005        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1006        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
1007                                 r, 0, base, offset);
1008    } else {
1009        switch (vece) {
1010        case MO_64:
1011            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
1012            break;
1013        case MO_32:
1014            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
1015            break;
1016        case MO_16:
1017            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
1018            tcg_out8(s, 0); /* imm8 */
1019            tcg_out_dup_vec(s, type, vece, r, r);
1020            break;
1021        case MO_8:
1022            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
1023            tcg_out8(s, 0); /* imm8 */
1024            tcg_out_dup_vec(s, type, vece, r, r);
1025            break;
1026        default:
1027            g_assert_not_reached();
1028        }
1029    }
1030    return true;
1031}
1032
1033static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1034                             TCGReg ret, int64_t arg)
1035{
1036    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1037
1038    if (arg == 0) {
1039        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1040        return;
1041    }
1042    if (arg == -1) {
1043        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
1044        return;
1045    }
1046
1047    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
1048        if (have_avx2) {
1049            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
1050        } else {
1051            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
1052        }
1053        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1054    } else {
1055        if (type == TCG_TYPE_V64) {
1056            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
1057        } else if (have_avx2) {
1058            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1059        } else {
1060            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1061        }
1062        if (TCG_TARGET_REG_BITS == 64) {
1063            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1064        } else {
1065            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1066        }
1067    }
1068}
1069
1070static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1071                             TCGReg ret, tcg_target_long arg)
1072{
1073    if (arg == 0) {
1074        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1075        return;
1076    }
1077    if (arg == -1) {
1078        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1079        return;
1080    }
1081
1082    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1083    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1084    if (TCG_TARGET_REG_BITS == 64) {
1085        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1086    } else {
1087        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1088    }
1089}
1090
1091static void tcg_out_movi_int(TCGContext *s, TCGType type,
1092                             TCGReg ret, tcg_target_long arg)
1093{
1094    tcg_target_long diff;
1095
1096    if (arg == 0 && !s->carry_live) {
1097        tgen_arithr(s, ARITH_XOR, ret, ret);
1098        return;
1099    }
1100    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1101        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1102        tcg_out32(s, arg);
1103        return;
1104    }
1105    if (arg == (int32_t)arg) {
1106        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1107        tcg_out32(s, arg);
1108        return;
1109    }
1110
1111    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1112    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1113    if (diff == (int32_t)diff) {
1114        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1115        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1116        tcg_out32(s, diff);
1117        return;
1118    }
1119
1120    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1121    tcg_out64(s, arg);
1122}
1123
1124static void tcg_out_movi(TCGContext *s, TCGType type,
1125                         TCGReg ret, tcg_target_long arg)
1126{
1127    switch (type) {
1128    case TCG_TYPE_I32:
1129#if TCG_TARGET_REG_BITS == 64
1130    case TCG_TYPE_I64:
1131#endif
1132        if (ret < 16) {
1133            tcg_out_movi_int(s, type, ret, arg);
1134        } else {
1135            tcg_out_movi_vec(s, type, ret, arg);
1136        }
1137        break;
1138    default:
1139        g_assert_not_reached();
1140    }
1141}
1142
1143static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1144{
1145    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1146    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1147    return true;
1148}
1149
1150static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1151                             tcg_target_long imm)
1152{
1153    /* This function is only used for passing structs by reference. */
1154    tcg_debug_assert(imm == (int32_t)imm);
1155    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1156}
1157
1158static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1159{
1160    if (val == (int8_t)val) {
1161        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1162        tcg_out8(s, val);
1163    } else if (val == (int32_t)val) {
1164        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1165        tcg_out32(s, val);
1166    } else {
1167        g_assert_not_reached();
1168    }
1169}
1170
1171static void tcg_out_mb(TCGContext *s, unsigned a0)
1172{
1173    /* Given the strength of x86 memory ordering, we only need care for
1174       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1175       faster than "mfence", so don't bother with the sse insn.  */
1176    if (a0 & TCG_MO_ST_LD) {
1177        tcg_out8(s, 0xf0);
1178        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1179        tcg_out8(s, 0);
1180    }
1181}
1182
1183static inline void tcg_out_push(TCGContext *s, int reg)
1184{
1185    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1186}
1187
1188static inline void tcg_out_pop(TCGContext *s, int reg)
1189{
1190    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1191}
1192
1193static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1194                       TCGReg arg1, intptr_t arg2)
1195{
1196    switch (type) {
1197    case TCG_TYPE_I32:
1198        if (ret < 16) {
1199            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1200        } else {
1201            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1202        }
1203        break;
1204    case TCG_TYPE_I64:
1205        if (ret < 16) {
1206            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1207            break;
1208        }
1209        /* FALLTHRU */
1210    case TCG_TYPE_V64:
1211        /* There is no instruction that can validate 8-byte alignment.  */
1212        tcg_debug_assert(ret >= 16);
1213        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1214        break;
1215    case TCG_TYPE_V128:
1216        /*
1217         * The gvec infrastructure is asserts that v128 vector loads
1218         * and stores use a 16-byte aligned offset.  Validate that the
1219         * final pointer is aligned by using an insn that will SIGSEGV.
1220         */
1221        tcg_debug_assert(ret >= 16);
1222        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1223        break;
1224    case TCG_TYPE_V256:
1225        /*
1226         * The gvec infrastructure only requires 16-byte alignment,
1227         * so here we must use an unaligned load.
1228         */
1229        tcg_debug_assert(ret >= 16);
1230        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1231                                 ret, 0, arg1, arg2);
1232        break;
1233    default:
1234        g_assert_not_reached();
1235    }
1236}
1237
1238static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1239                       TCGReg arg1, intptr_t arg2)
1240{
1241    switch (type) {
1242    case TCG_TYPE_I32:
1243        if (arg < 16) {
1244            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1245        } else {
1246            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1247        }
1248        break;
1249    case TCG_TYPE_I64:
1250        if (arg < 16) {
1251            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1252            break;
1253        }
1254        /* FALLTHRU */
1255    case TCG_TYPE_V64:
1256        /* There is no instruction that can validate 8-byte alignment.  */
1257        tcg_debug_assert(arg >= 16);
1258        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1259        break;
1260    case TCG_TYPE_V128:
1261        /*
1262         * The gvec infrastructure is asserts that v128 vector loads
1263         * and stores use a 16-byte aligned offset.  Validate that the
1264         * final pointer is aligned by using an insn that will SIGSEGV.
1265         *
1266         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1267         * for _WIN64, which must have SSE2 but may not have AVX.
1268         */
1269        tcg_debug_assert(arg >= 16);
1270        if (have_avx1) {
1271            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1272        } else {
1273            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1274        }
1275        break;
1276    case TCG_TYPE_V256:
1277        /*
1278         * The gvec infrastructure only requires 16-byte alignment,
1279         * so here we must use an unaligned store.
1280         */
1281        tcg_debug_assert(arg >= 16);
1282        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1283                                 arg, 0, arg1, arg2);
1284        break;
1285    default:
1286        g_assert_not_reached();
1287    }
1288}
1289
1290static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1291                        TCGReg base, intptr_t ofs)
1292{
1293    int rexw = 0;
1294    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1295        if (val != (int32_t)val) {
1296            return false;
1297        }
1298        rexw = P_REXW;
1299    } else if (type != TCG_TYPE_I32) {
1300        return false;
1301    }
1302    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1303    tcg_out32(s, val);
1304    return true;
1305}
1306
1307static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1308{
1309    /* Propagate an opcode prefix, such as P_DATA16.  */
1310    int ext = subopc & ~0x7;
1311    subopc &= 0x7;
1312
1313    if (count == 1) {
1314        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1315    } else {
1316        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1317        tcg_out8(s, count);
1318    }
1319}
1320
1321static inline void tcg_out_bswap32(TCGContext *s, int reg)
1322{
1323    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1324}
1325
1326static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1327{
1328    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1329}
1330
1331static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1332{
1333    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1334        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1335        if (dest >= 4) {
1336            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
1337            tcg_out32(s, 0xff);
1338            return;
1339        }
1340        src = dest;
1341    }
1342    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1343}
1344
1345static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1346{
1347    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1348
1349    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1350        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1351        if (dest >= 4) {
1352            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
1353            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
1354            return;
1355        }
1356        src = dest;
1357    }
1358    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1359}
1360
1361static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1362{
1363    /* movzwl */
1364    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1365}
1366
1367static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1368{
1369    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1370    /* movsw[lq] */
1371    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1372}
1373
1374static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1375{
1376    /* 32-bit mov zero extends.  */
1377    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1378}
1379
1380static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1381{
1382    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1383    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1384}
1385
1386static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1387{
1388    tcg_out_ext32s(s, dest, src);
1389}
1390
1391static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1392{
1393    if (dest != src) {
1394        tcg_out_ext32u(s, dest, src);
1395    }
1396}
1397
1398static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1399{
1400    tcg_out_ext32u(s, dest, src);
1401}
1402
1403static inline void tcg_out_bswap64(TCGContext *s, int reg)
1404{
1405    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1406}
1407
1408static void tgen_arithi(TCGContext *s, int c, int r0,
1409                        tcg_target_long val, int cf)
1410{
1411    int rexw = 0;
1412
1413    if (TCG_TARGET_REG_BITS == 64) {
1414        rexw = c & -8;
1415        c &= 7;
1416    }
1417
1418    switch (c) {
1419    case ARITH_ADD:
1420    case ARITH_SUB:
1421        if (!cf) {
1422            /*
1423             * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1424             * partial flags update stalls on Pentium4 and are not recommended
1425             * by current Intel optimization manuals.
1426             */
1427            if (val == 1 || val == -1) {
1428                int is_inc = (c == ARITH_ADD) ^ (val < 0);
1429                if (TCG_TARGET_REG_BITS == 64) {
1430                    /*
1431                     * The single-byte increment encodings are re-tasked
1432                     * as the REX prefixes.  Use the MODRM encoding.
1433                     */
1434                    tcg_out_modrm(s, OPC_GRP5 + rexw,
1435                                  (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1436                } else {
1437                    tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1438                }
1439                return;
1440            }
1441            if (val == 128) {
1442                /*
1443                 * Facilitate using an 8-bit immediate.  Carry is inverted
1444                 * by this transformation, so do it only if cf == 0.
1445                 */
1446                c ^= ARITH_ADD ^ ARITH_SUB;
1447                val = -128;
1448            }
1449        }
1450        break;
1451
1452    case ARITH_AND:
1453        if (TCG_TARGET_REG_BITS == 64) {
1454            if (val == 0xffffffffu) {
1455                tcg_out_ext32u(s, r0, r0);
1456                return;
1457            }
1458            if (val == (uint32_t)val) {
1459                /* AND with no high bits set can use a 32-bit operation.  */
1460                rexw = 0;
1461            }
1462        }
1463        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1464            tcg_out_ext8u(s, r0, r0);
1465            return;
1466        }
1467        if (val == 0xffffu) {
1468            tcg_out_ext16u(s, r0, r0);
1469            return;
1470        }
1471        break;
1472
1473    case ARITH_OR:
1474    case ARITH_XOR:
1475        if (val >= 0x80 && val <= 0xff
1476            && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1477            tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1478            tcg_out8(s, val);
1479            return;
1480        }
1481        break;
1482    }
1483
1484    if (val == (int8_t)val) {
1485        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1486        tcg_out8(s, val);
1487        return;
1488    }
1489    if (rexw == 0 || val == (int32_t)val) {
1490        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1491        tcg_out32(s, val);
1492        return;
1493    }
1494
1495    g_assert_not_reached();
1496}
1497
1498static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1499{
1500    if (val != 0) {
1501        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1502    }
1503}
1504
1505/* Set SMALL to force a short forward branch.  */
1506static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1507{
1508    int32_t val, val1;
1509
1510    if (l->has_value) {
1511        val = tcg_pcrel_diff(s, l->u.value_ptr);
1512        val1 = val - 2;
1513        if ((int8_t)val1 == val1) {
1514            if (opc == -1) {
1515                tcg_out8(s, OPC_JMP_short);
1516            } else {
1517                tcg_out8(s, OPC_JCC_short + opc);
1518            }
1519            tcg_out8(s, val1);
1520        } else {
1521            tcg_debug_assert(!small);
1522            if (opc == -1) {
1523                tcg_out8(s, OPC_JMP_long);
1524                tcg_out32(s, val - 5);
1525            } else {
1526                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1527                tcg_out32(s, val - 6);
1528            }
1529        }
1530    } else if (small) {
1531        if (opc == -1) {
1532            tcg_out8(s, OPC_JMP_short);
1533        } else {
1534            tcg_out8(s, OPC_JCC_short + opc);
1535        }
1536        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1537        s->code_ptr += 1;
1538    } else {
1539        if (opc == -1) {
1540            tcg_out8(s, OPC_JMP_long);
1541        } else {
1542            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1543        }
1544        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1545        s->code_ptr += 4;
1546    }
1547}
1548
1549static void tcg_out_br(TCGContext *s, TCGLabel *l)
1550{
1551    tcg_out_jxx(s, JCC_JMP, l, 0);
1552}
1553
1554static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1555                       TCGArg arg2, int const_arg2, int rexw)
1556{
1557    int jz, js;
1558
1559    if (!is_tst_cond(cond)) {
1560        if (!const_arg2) {
1561            tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1562        } else if (arg2 == 0) {
1563            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1564        } else {
1565            tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1566            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1567        }
1568        return tcg_cond_to_jcc[cond];
1569    }
1570
1571    jz = tcg_cond_to_jcc[cond];
1572    js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1573
1574    if (!const_arg2) {
1575        tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1576        return jz;
1577    }
1578
1579    if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1580        if (arg2 == 0x80) {
1581            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1582            return js;
1583        }
1584        if (arg2 == 0xff) {
1585            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1586            return jz;
1587        }
1588        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1589        tcg_out8(s, arg2);
1590        return jz;
1591    }
1592
1593    if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1594        if (arg2 == 0x8000) {
1595            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1596            return js;
1597        }
1598        if (arg2 == 0xff00) {
1599            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1600            return jz;
1601        }
1602        tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1603        tcg_out8(s, arg2 >> 8);
1604        return jz;
1605    }
1606
1607    if (arg2 == 0xffff) {
1608        tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1609        return jz;
1610    }
1611    if (arg2 == 0xffffffffu) {
1612        tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1613        return jz;
1614    }
1615
1616    if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1617        int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1618        int sh = ctz64(arg2);
1619
1620        rexw = (sh & 32 ? P_REXW : 0);
1621        if ((sh & 31) == 31) {
1622            tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1623            return js;
1624        } else {
1625            tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1626            tcg_out8(s, sh);
1627            return jc;
1628        }
1629    }
1630
1631    if (rexw) {
1632        if (arg2 == (uint32_t)arg2) {
1633            rexw = 0;
1634        } else {
1635            tcg_debug_assert(arg2 == (int32_t)arg2);
1636        }
1637    }
1638    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1639    tcg_out32(s, arg2);
1640    return jz;
1641}
1642
1643static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1644                           TCGArg arg1, TCGArg arg2, int const_arg2,
1645                           TCGLabel *label, bool small)
1646{
1647    int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1648    tcg_out_jxx(s, jcc, label, small);
1649}
1650
1651static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
1652                        TCGReg arg1, TCGReg arg2, TCGLabel *label)
1653{
1654    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1655    tcg_out_brcond(s, rexw, cond, arg1, arg2, false, label, false);
1656}
1657
1658static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
1659                         TCGReg arg1, tcg_target_long arg2, TCGLabel *label)
1660{
1661    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1662    tcg_out_brcond(s, rexw, cond, arg1, arg2, true, label, false);
1663}
1664
1665static const TCGOutOpBrcond outop_brcond = {
1666    .base.static_constraint = C_O0_I2(r, reT),
1667    .out_rr = tgen_brcond,
1668    .out_ri = tgen_brcondi,
1669};
1670
1671static void tcg_out_brcond2(TCGContext *s, TCGCond cond, TCGReg al,
1672                            TCGReg ah, TCGArg bl, bool blconst,
1673                            TCGArg bh, bool bhconst,
1674                            TCGLabel *label_this, bool small)
1675{
1676    TCGLabel *label_next = gen_new_label();
1677
1678    switch (cond) {
1679    case TCG_COND_EQ:
1680    case TCG_COND_TSTEQ:
1681        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1682                       al, bl, blconst, label_next, true);
1683        tcg_out_brcond(s, 0, cond, ah, bh, bhconst, label_this, small);
1684        break;
1685
1686    case TCG_COND_NE:
1687    case TCG_COND_TSTNE:
1688        tcg_out_brcond(s, 0, cond, al, bl, blconst, label_this, small);
1689        tcg_out_brcond(s, 0, cond, ah, bh, bhconst, label_this, small);
1690        break;
1691
1692    default:
1693        tcg_out_brcond(s, 0, tcg_high_cond(cond),
1694                       ah, bh, bhconst, label_this, small);
1695        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1696        tcg_out_brcond(s, 0, tcg_unsigned_cond(cond),
1697                       al, bl, blconst, label_this, small);
1698        break;
1699    }
1700    tcg_out_label(s, label_next);
1701}
1702
1703static void tgen_brcond2(TCGContext *s, TCGCond cond, TCGReg al,
1704                         TCGReg ah, TCGArg bl, bool blconst,
1705                         TCGArg bh, bool bhconst, TCGLabel *l)
1706{
1707    tcg_out_brcond2(s, cond, al, ah, bl, blconst, bh, bhconst, l, false);
1708}
1709
1710#if TCG_TARGET_REG_BITS != 32
1711__attribute__((unused))
1712#endif
1713static const TCGOutOpBrcond2 outop_brcond2 = {
1714    .base.static_constraint = C_O0_I4(r, r, ri, ri),
1715    .out = tgen_brcond2,
1716};
1717
1718static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond,
1719                            TCGReg dest, TCGReg arg1, TCGArg arg2,
1720                            bool const_arg2, bool neg)
1721{
1722    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1723    int cmp_rexw = rexw;
1724    bool inv = false;
1725    bool cleared;
1726    int jcc;
1727
1728    switch (cond) {
1729    case TCG_COND_NE:
1730        inv = true;
1731        /* fall through */
1732    case TCG_COND_EQ:
1733        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1734        if (const_arg2 && arg2 == 0) {
1735            arg2 = 1;
1736            goto do_ltu;
1737        }
1738        break;
1739
1740    case TCG_COND_TSTNE:
1741        inv = true;
1742        /* fall through */
1743    case TCG_COND_TSTEQ:
1744        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1745        if (const_arg2 && arg2 == 0xffffffffu) {
1746            arg2 = 1;
1747            cmp_rexw = 0;
1748            goto do_ltu;
1749        }
1750        break;
1751
1752    case TCG_COND_LEU:
1753        inv = true;
1754        /* fall through */
1755    case TCG_COND_GTU:
1756        /* If arg2 is a register, swap for LTU/GEU. */
1757        if (!const_arg2) {
1758            TCGReg t = arg1;
1759            arg1 = arg2;
1760            arg2 = t;
1761            goto do_ltu;
1762        }
1763        break;
1764
1765    case TCG_COND_GEU:
1766        inv = true;
1767        /* fall through */
1768    case TCG_COND_LTU:
1769    do_ltu:
1770        /*
1771         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1772         * We can then use NEG or INC to produce the desired result.
1773         * This is always smaller than the SETCC expansion.
1774         */
1775        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1776
1777        /* X - X - C = -C = (C ? -1 : 0) */
1778        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1779        if (inv && neg) {
1780            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1781            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1782        } else if (inv) {
1783            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1784            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1785        } else if (!neg) {
1786            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1787            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1788        }
1789        return;
1790
1791    case TCG_COND_GE:
1792        inv = true;
1793        /* fall through */
1794    case TCG_COND_LT:
1795        /* If arg2 is 0, extract the sign bit. */
1796        if (const_arg2 && arg2 == 0) {
1797            tcg_out_mov(s, type, dest, arg1);
1798            if (inv) {
1799                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1800            }
1801            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1802                           dest, rexw ? 63 : 31);
1803            return;
1804        }
1805        break;
1806
1807    default:
1808        break;
1809    }
1810
1811    /*
1812     * If dest does not overlap the inputs, clearing it first is preferred.
1813     * The XOR breaks any false dependency for the low-byte write to dest,
1814     * and is also one byte smaller than MOVZBL.
1815     */
1816    cleared = false;
1817    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1818        tgen_arithr(s, ARITH_XOR, dest, dest);
1819        cleared = true;
1820    }
1821
1822    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1823    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1824
1825    if (!cleared) {
1826        tcg_out_ext8u(s, dest, dest);
1827    }
1828    if (neg) {
1829        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1830    }
1831}
1832
1833static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
1834                         TCGReg dest, TCGReg arg1, TCGReg arg2)
1835{
1836    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, false);
1837}
1838
1839static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
1840                          TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1841{
1842    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, false);
1843}
1844
1845static const TCGOutOpSetcond outop_setcond = {
1846    .base.static_constraint = C_O1_I2(q, r, reT),
1847    .out_rrr = tgen_setcond,
1848    .out_rri = tgen_setcondi,
1849};
1850
1851static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
1852                            TCGReg dest, TCGReg arg1, TCGReg arg2)
1853{
1854    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, true);
1855}
1856
1857static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
1858                             TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1859{
1860    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, true);
1861}
1862
1863static const TCGOutOpSetcond outop_negsetcond = {
1864    .base.static_constraint = C_O1_I2(q, r, reT),
1865    .out_rrr = tgen_negsetcond,
1866    .out_rri = tgen_negsetcondi,
1867};
1868
1869static void tgen_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
1870                          TCGReg al, TCGReg ah,
1871                          TCGArg bl, bool const_bl,
1872                          TCGArg bh, bool const_bh)
1873{
1874    TCGLabel *label_over = gen_new_label();
1875
1876    if (ret == al || ret == ah
1877        || (!const_bl && ret == bl)
1878        || (!const_bh && ret == bh)) {
1879        /*
1880         * When the destination overlaps with one of the argument
1881         * registers, don't do anything tricky.
1882         */
1883        TCGLabel *label_true = gen_new_label();
1884
1885        tcg_out_brcond2(s, cond, al, ah, bl, const_bl,
1886                        bh, const_bh, label_true, true);
1887
1888        tcg_out_movi(s, TCG_TYPE_I32, ret, 0);
1889        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1890        tcg_out_label(s, label_true);
1891
1892        tcg_out_movi(s, TCG_TYPE_I32, ret, 1);
1893    } else {
1894        /*
1895         * When the destination does not overlap one of the arguments,
1896         * clear the destination first, jump if cond false, and emit an
1897         * increment in the true case.  This results in smaller code.
1898         */
1899        tcg_out_movi(s, TCG_TYPE_I32, ret, 0);
1900
1901        tcg_out_brcond2(s, tcg_invert_cond(cond), al, ah, bl, const_bl,
1902                        bh, const_bh, label_over, true);
1903
1904        tgen_arithi(s, ARITH_ADD, ret, 1, 0);
1905    }
1906    tcg_out_label(s, label_over);
1907}
1908
1909#if TCG_TARGET_REG_BITS != 32
1910__attribute__((unused))
1911#endif
1912static const TCGOutOpSetcond2 outop_setcond2 = {
1913    .base.static_constraint = C_O1_I4(r, r, r, ri, ri),
1914    .out = tgen_setcond2,
1915};
1916
1917static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1918                         TCGReg dest, TCGReg v1)
1919{
1920    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1921}
1922
1923static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
1924                         TCGReg dest, TCGReg c1, TCGArg c2, bool const_c2,
1925                         TCGArg vt, bool const_vt,
1926                         TCGArg vf, bool consf_vf)
1927{
1928    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1929    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1930    tcg_out_cmov(s, jcc, rexw, dest, vt);
1931}
1932
1933static const TCGOutOpMovcond outop_movcond = {
1934    .base.static_constraint = C_O1_I4(r, r, reT, r, 0),
1935    .out = tgen_movcond,
1936};
1937
1938static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1939{
1940    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1941
1942    if (disp == (int32_t)disp) {
1943        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1944        tcg_out32(s, disp);
1945    } else {
1946        /* rip-relative addressing into the constant pool.
1947           This is 6 + 8 = 14 bytes, as compared to using an
1948           immediate load 10 + 6 = 16 bytes, plus we may
1949           be able to re-use the pool constant for more calls.  */
1950        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1951        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1952        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1953        tcg_out32(s, 0);
1954    }
1955}
1956
1957static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1958                         const TCGHelperInfo *info)
1959{
1960    tcg_out_branch(s, 1, dest);
1961
1962#ifndef _WIN32
1963    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1964        /*
1965         * The sysv i386 abi for struct return places a reference as the
1966         * first argument of the stack, and pops that argument with the
1967         * return statement.  Since we want to retain the aligned stack
1968         * pointer for the callee, we do not want to actually push that
1969         * argument before the call but rely on the normal store to the
1970         * stack slot.  But we do need to compensate for the pop in order
1971         * to reset our correct stack pointer value.
1972         * Pushing a garbage value back onto the stack is quickest.
1973         */
1974        tcg_out_push(s, TCG_REG_EAX);
1975    }
1976#endif
1977}
1978
1979static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1980{
1981    tcg_out_branch(s, 0, dest);
1982}
1983
1984static void tcg_out_nopn(TCGContext *s, int n)
1985{
1986    int i;
1987    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1988     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1989     * duplicate prefix, and all of the interesting recent cores can
1990     * decode and discard the duplicates in a single cycle.
1991     */
1992    tcg_debug_assert(n >= 1);
1993    for (i = 1; i < n; ++i) {
1994        tcg_out8(s, 0x66);
1995    }
1996    tcg_out8(s, 0x90);
1997}
1998
1999typedef struct {
2000    TCGReg base;
2001    int index;
2002    int ofs;
2003    int seg;
2004    TCGAtomAlign aa;
2005} HostAddress;
2006
2007bool tcg_target_has_memory_bswap(MemOp memop)
2008{
2009    TCGAtomAlign aa;
2010
2011    if (!have_movbe) {
2012        return false;
2013    }
2014    if ((memop & MO_SIZE) < MO_128) {
2015        return true;
2016    }
2017
2018    /*
2019     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
2020     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
2021     */
2022    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
2023    return aa.atom < MO_128;
2024}
2025
2026/*
2027 * Because i686 has no register parameters and because x86_64 has xchg
2028 * to handle addr/data register overlap, we have placed all input arguments
2029 * before we need might need a scratch reg.
2030 *
2031 * Even then, a scratch is only needed for l->raddr.  Rather than expose
2032 * a general-purpose scratch when we don't actually know it's available,
2033 * use the ra_gen hook to load into RAX if needed.
2034 */
2035#if TCG_TARGET_REG_BITS == 64
2036static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
2037{
2038    if (arg < 0) {
2039        arg = TCG_REG_RAX;
2040    }
2041    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2042    return arg;
2043}
2044static const TCGLdstHelperParam ldst_helper_param = {
2045    .ra_gen = ldst_ra_gen
2046};
2047#else
2048static const TCGLdstHelperParam ldst_helper_param = { };
2049#endif
2050
2051static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2052                                TCGReg l, TCGReg h, TCGReg v)
2053{
2054    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2055
2056    /* vpmov{d,q} %v, %l */
2057    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2058    /* vpextr{d,q} $1, %v, %h */
2059    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2060    tcg_out8(s, 1);
2061}
2062
2063static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2064                                TCGReg v, TCGReg l, TCGReg h)
2065{
2066    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2067
2068    /* vmov{d,q} %l, %v */
2069    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2070    /* vpinsr{d,q} $1, %h, %v, %v */
2071    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2072    tcg_out8(s, 1);
2073}
2074
2075/*
2076 * Generate code for the slow path for a load at the end of block
2077 */
2078static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2079{
2080    MemOp opc = get_memop(l->oi);
2081    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2082
2083    /* resolve label address */
2084    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2085    if (label_ptr[1]) {
2086        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2087    }
2088
2089    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2090    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2091    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2092
2093    tcg_out_jmp(s, l->raddr);
2094    return true;
2095}
2096
2097/*
2098 * Generate code for the slow path for a store at the end of block
2099 */
2100static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2101{
2102    MemOp opc = get_memop(l->oi);
2103    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2104
2105    /* resolve label address */
2106    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2107    if (label_ptr[1]) {
2108        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2109    }
2110
2111    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2112    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2113
2114    tcg_out_jmp(s, l->raddr);
2115    return true;
2116}
2117
2118#ifdef CONFIG_USER_ONLY
2119static HostAddress x86_guest_base = {
2120    .index = -1
2121};
2122
2123#if defined(__x86_64__) && defined(__linux__)
2124# include <asm/prctl.h>
2125# include <sys/prctl.h>
2126int arch_prctl(int code, unsigned long addr);
2127static inline int setup_guest_base_seg(void)
2128{
2129    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2130        return P_GS;
2131    }
2132    return 0;
2133}
2134#define setup_guest_base_seg  setup_guest_base_seg
2135#elif defined(__x86_64__) && \
2136      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2137# include <machine/sysarch.h>
2138static inline int setup_guest_base_seg(void)
2139{
2140    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2141        return P_GS;
2142    }
2143    return 0;
2144}
2145#define setup_guest_base_seg  setup_guest_base_seg
2146#endif
2147#else
2148# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2149#endif /* CONFIG_USER_ONLY */
2150#ifndef setup_guest_base_seg
2151# define setup_guest_base_seg()  0
2152#endif
2153
2154#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2155
2156/*
2157 * For softmmu, perform the TLB load and compare.
2158 * For useronly, perform any required alignment tests.
2159 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2160 * is required and fill in @h with the host address for the fast path.
2161 */
2162static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2163                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2164{
2165    TCGLabelQemuLdst *ldst = NULL;
2166    MemOp opc = get_memop(oi);
2167    MemOp s_bits = opc & MO_SIZE;
2168    unsigned a_mask;
2169
2170    if (tcg_use_softmmu) {
2171        h->index = TCG_REG_L0;
2172        h->ofs = 0;
2173        h->seg = 0;
2174    } else {
2175        *h = x86_guest_base;
2176    }
2177    h->base = addr;
2178    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2179    a_mask = (1 << h->aa.align) - 1;
2180
2181    if (tcg_use_softmmu) {
2182        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2183                            : offsetof(CPUTLBEntry, addr_write);
2184        TCGType ttype = TCG_TYPE_I32;
2185        TCGType tlbtype = TCG_TYPE_I32;
2186        int trexw = 0, hrexw = 0, tlbrexw = 0;
2187        unsigned mem_index = get_mmuidx(oi);
2188        unsigned s_mask = (1 << s_bits) - 1;
2189        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2190        int tlb_mask;
2191
2192        ldst = new_ldst_label(s);
2193        ldst->is_ld = is_ld;
2194        ldst->oi = oi;
2195        ldst->addr_reg = addr;
2196
2197        if (TCG_TARGET_REG_BITS == 64) {
2198            ttype = s->addr_type;
2199            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2200            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2201                hrexw = P_REXW;
2202                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2203                    tlbtype = TCG_TYPE_I64;
2204                    tlbrexw = P_REXW;
2205                }
2206            }
2207        }
2208
2209        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2210        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2211                       s->page_bits - CPU_TLB_ENTRY_BITS);
2212
2213        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2214                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2215
2216        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2217                             fast_ofs + offsetof(CPUTLBDescFast, table));
2218
2219        /*
2220         * If the required alignment is at least as large as the access,
2221         * simply copy the address and mask.  For lesser alignments,
2222         * check that we don't cross pages for the complete access.
2223         */
2224        if (a_mask >= s_mask) {
2225            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2226        } else {
2227            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2228                                 addr, s_mask - a_mask);
2229        }
2230        tlb_mask = s->page_mask | a_mask;
2231        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2232
2233        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2234        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2235                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2236
2237        /* jne slow_path */
2238        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2239        ldst->label_ptr[0] = s->code_ptr;
2240        s->code_ptr += 4;
2241
2242        /* TLB Hit.  */
2243        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2244                   offsetof(CPUTLBEntry, addend));
2245    } else if (a_mask) {
2246        int jcc;
2247
2248        ldst = new_ldst_label(s);
2249        ldst->is_ld = is_ld;
2250        ldst->oi = oi;
2251        ldst->addr_reg = addr;
2252
2253        /* jne slow_path */
2254        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2255        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2256        ldst->label_ptr[0] = s->code_ptr;
2257        s->code_ptr += 4;
2258    }
2259
2260    return ldst;
2261}
2262
2263static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2264                                   HostAddress h, TCGType type, MemOp memop)
2265{
2266    bool use_movbe = false;
2267    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2268    int movop = OPC_MOVL_GvEv;
2269
2270    /* Do big-endian loads with movbe.  */
2271    if (memop & MO_BSWAP) {
2272        tcg_debug_assert(have_movbe);
2273        use_movbe = true;
2274        movop = OPC_MOVBE_GyMy;
2275    }
2276
2277    switch (memop & MO_SSIZE) {
2278    case MO_UB:
2279        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2280                                 h.base, h.index, 0, h.ofs);
2281        break;
2282    case MO_SB:
2283        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2284                                 h.base, h.index, 0, h.ofs);
2285        break;
2286    case MO_UW:
2287        if (use_movbe) {
2288            /* There is no extending movbe; only low 16-bits are modified.  */
2289            if (datalo != h.base && datalo != h.index) {
2290                /* XOR breaks dependency chains.  */
2291                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2292                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2293                                         datalo, h.base, h.index, 0, h.ofs);
2294            } else {
2295                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2296                                         datalo, h.base, h.index, 0, h.ofs);
2297                tcg_out_ext16u(s, datalo, datalo);
2298            }
2299        } else {
2300            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2301                                     h.base, h.index, 0, h.ofs);
2302        }
2303        break;
2304    case MO_SW:
2305        if (use_movbe) {
2306            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2307                                     datalo, h.base, h.index, 0, h.ofs);
2308            tcg_out_ext16s(s, type, datalo, datalo);
2309        } else {
2310            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2311                                     datalo, h.base, h.index, 0, h.ofs);
2312        }
2313        break;
2314    case MO_UL:
2315        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2316                                 h.base, h.index, 0, h.ofs);
2317        break;
2318#if TCG_TARGET_REG_BITS == 64
2319    case MO_SL:
2320        if (use_movbe) {
2321            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2322                                     h.base, h.index, 0, h.ofs);
2323            tcg_out_ext32s(s, datalo, datalo);
2324        } else {
2325            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2326                                     h.base, h.index, 0, h.ofs);
2327        }
2328        break;
2329#endif
2330    case MO_UQ:
2331        if (TCG_TARGET_REG_BITS == 64) {
2332            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2333                                     h.base, h.index, 0, h.ofs);
2334            break;
2335        }
2336        if (use_movbe) {
2337            TCGReg t = datalo;
2338            datalo = datahi;
2339            datahi = t;
2340        }
2341        if (h.base == datalo || h.index == datalo) {
2342            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2343                                     h.base, h.index, 0, h.ofs);
2344            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2345            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2346        } else {
2347            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2348                                     h.base, h.index, 0, h.ofs);
2349            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2350                                     h.base, h.index, 0, h.ofs + 4);
2351        }
2352        break;
2353
2354    case MO_128:
2355        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2356
2357        /*
2358         * Without 16-byte atomicity, use integer regs.
2359         * That is where we want the data, and it allows bswaps.
2360         */
2361        if (h.aa.atom < MO_128) {
2362            if (use_movbe) {
2363                TCGReg t = datalo;
2364                datalo = datahi;
2365                datahi = t;
2366            }
2367            if (h.base == datalo || h.index == datalo) {
2368                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2369                                         h.base, h.index, 0, h.ofs);
2370                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2371                                     datalo, datahi, 0);
2372                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2373                                     datahi, datahi, 8);
2374            } else {
2375                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2376                                         h.base, h.index, 0, h.ofs);
2377                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2378                                         h.base, h.index, 0, h.ofs + 8);
2379            }
2380            break;
2381        }
2382
2383        /*
2384         * With 16-byte atomicity, a vector load is required.
2385         * If we already have 16-byte alignment, then VMOVDQA always works.
2386         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2387         * Else use we require a runtime test for alignment for VMOVDQA;
2388         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2389         */
2390        if (h.aa.align >= MO_128) {
2391            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2392                                         TCG_TMP_VEC, 0,
2393                                         h.base, h.index, 0, h.ofs);
2394        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2395            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2396                                         TCG_TMP_VEC, 0,
2397                                         h.base, h.index, 0, h.ofs);
2398        } else {
2399            TCGLabel *l1 = gen_new_label();
2400            TCGLabel *l2 = gen_new_label();
2401            int jcc;
2402
2403            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2404            tcg_out_jxx(s, jcc, l1, true);
2405
2406            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2407                                         TCG_TMP_VEC, 0,
2408                                         h.base, h.index, 0, h.ofs);
2409            tcg_out_jxx(s, JCC_JMP, l2, true);
2410
2411            tcg_out_label(s, l1);
2412            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2413                                         TCG_TMP_VEC, 0,
2414                                         h.base, h.index, 0, h.ofs);
2415            tcg_out_label(s, l2);
2416        }
2417        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2418        break;
2419
2420    default:
2421        g_assert_not_reached();
2422    }
2423}
2424
2425static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2426                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2427{
2428    TCGLabelQemuLdst *ldst;
2429    HostAddress h;
2430
2431    ldst = prepare_host_addr(s, &h, addr, oi, true);
2432    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2433
2434    if (ldst) {
2435        ldst->type = data_type;
2436        ldst->datalo_reg = datalo;
2437        ldst->datahi_reg = datahi;
2438        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2439    }
2440}
2441
2442static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2443                                   HostAddress h, MemOp memop)
2444{
2445    bool use_movbe = false;
2446    int movop = OPC_MOVL_EvGv;
2447
2448    /*
2449     * Do big-endian stores with movbe or system-mode.
2450     * User-only without movbe will have its swapping done generically.
2451     */
2452    if (memop & MO_BSWAP) {
2453        tcg_debug_assert(have_movbe);
2454        use_movbe = true;
2455        movop = OPC_MOVBE_MyGy;
2456    }
2457
2458    switch (memop & MO_SIZE) {
2459    case MO_8:
2460        /* This is handled with constraints on INDEX_op_qemu_st_i32. */
2461        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2462        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2463                                 datalo, h.base, h.index, 0, h.ofs);
2464        break;
2465    case MO_16:
2466        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2467                                 h.base, h.index, 0, h.ofs);
2468        break;
2469    case MO_32:
2470        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2471                                 h.base, h.index, 0, h.ofs);
2472        break;
2473    case MO_64:
2474        if (TCG_TARGET_REG_BITS == 64) {
2475            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2476                                     h.base, h.index, 0, h.ofs);
2477        } else {
2478            if (use_movbe) {
2479                TCGReg t = datalo;
2480                datalo = datahi;
2481                datahi = t;
2482            }
2483            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2484                                     h.base, h.index, 0, h.ofs);
2485            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2486                                     h.base, h.index, 0, h.ofs + 4);
2487        }
2488        break;
2489
2490    case MO_128:
2491        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2492
2493        /*
2494         * Without 16-byte atomicity, use integer regs.
2495         * That is where we have the data, and it allows bswaps.
2496         */
2497        if (h.aa.atom < MO_128) {
2498            if (use_movbe) {
2499                TCGReg t = datalo;
2500                datalo = datahi;
2501                datahi = t;
2502            }
2503            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2504                                     h.base, h.index, 0, h.ofs);
2505            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2506                                     h.base, h.index, 0, h.ofs + 8);
2507            break;
2508        }
2509
2510        /*
2511         * With 16-byte atomicity, a vector store is required.
2512         * If we already have 16-byte alignment, then VMOVDQA always works.
2513         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2514         * Else use we require a runtime test for alignment for VMOVDQA;
2515         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2516         */
2517        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2518        if (h.aa.align >= MO_128) {
2519            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2520                                         TCG_TMP_VEC, 0,
2521                                         h.base, h.index, 0, h.ofs);
2522        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2523            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2524                                         TCG_TMP_VEC, 0,
2525                                         h.base, h.index, 0, h.ofs);
2526        } else {
2527            TCGLabel *l1 = gen_new_label();
2528            TCGLabel *l2 = gen_new_label();
2529            int jcc;
2530
2531            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2532            tcg_out_jxx(s, jcc, l1, true);
2533
2534            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2535                                         TCG_TMP_VEC, 0,
2536                                         h.base, h.index, 0, h.ofs);
2537            tcg_out_jxx(s, JCC_JMP, l2, true);
2538
2539            tcg_out_label(s, l1);
2540            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2541                                         TCG_TMP_VEC, 0,
2542                                         h.base, h.index, 0, h.ofs);
2543            tcg_out_label(s, l2);
2544        }
2545        break;
2546
2547    default:
2548        g_assert_not_reached();
2549    }
2550}
2551
2552static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2553                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2554{
2555    TCGLabelQemuLdst *ldst;
2556    HostAddress h;
2557
2558    ldst = prepare_host_addr(s, &h, addr, oi, false);
2559    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2560
2561    if (ldst) {
2562        ldst->type = data_type;
2563        ldst->datalo_reg = datalo;
2564        ldst->datahi_reg = datahi;
2565        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2566    }
2567}
2568
2569static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2570{
2571    /* Reuse the zeroing that exists for goto_ptr.  */
2572    if (a0 == 0) {
2573        tcg_out_jmp(s, tcg_code_gen_epilogue);
2574    } else {
2575        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2576        tcg_out_jmp(s, tb_ret_addr);
2577    }
2578}
2579
2580static void tcg_out_goto_tb(TCGContext *s, int which)
2581{
2582    /*
2583     * Jump displacement must be aligned for atomic patching;
2584     * see if we need to add extra nops before jump
2585     */
2586    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2587    if (gap != 1) {
2588        tcg_out_nopn(s, gap - 1);
2589    }
2590    tcg_out8(s, OPC_JMP_long); /* jmp im */
2591    set_jmp_insn_offset(s, which);
2592    tcg_out32(s, 0);
2593    set_jmp_reset_offset(s, which);
2594}
2595
2596static void tcg_out_goto_ptr(TCGContext *s, TCGReg a0)
2597{
2598    /* Jump to the given host address (could be epilogue) */
2599    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2600}
2601
2602void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2603                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2604{
2605    /* patch the branch destination */
2606    uintptr_t addr = tb->jmp_target_addr[n];
2607    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2608    /* no need to flush icache explicitly */
2609}
2610
2611
2612static void tgen_add(TCGContext *s, TCGType type,
2613                     TCGReg a0, TCGReg a1, TCGReg a2)
2614{
2615    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2616
2617    if (a0 == a1) {
2618        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2619    } else if (a0 == a2) {
2620        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2621    } else {
2622        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2623    }
2624}
2625
2626static void tgen_addi(TCGContext *s, TCGType type,
2627                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2628{
2629    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2630
2631    if (a0 == a1) {
2632        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2633    } else {
2634        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2635    }
2636}
2637
2638static const TCGOutOpBinary outop_add = {
2639    .base.static_constraint = C_O1_I2(r, r, re),
2640    .out_rrr = tgen_add,
2641    .out_rri = tgen_addi,
2642};
2643
2644static void tgen_addco(TCGContext *s, TCGType type,
2645                       TCGReg a0, TCGReg a1, TCGReg a2)
2646{
2647    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2648    tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2649}
2650
2651static void tgen_addco_imm(TCGContext *s, TCGType type,
2652                           TCGReg a0, TCGReg a1, tcg_target_long a2)
2653{
2654    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2655    tgen_arithi(s, ARITH_ADD + rexw, a0, a2, true);
2656}
2657
2658static const TCGOutOpBinary outop_addco = {
2659    .base.static_constraint = C_O1_I2(r, 0, re),
2660    .out_rrr = tgen_addco,
2661    .out_rri = tgen_addco_imm,
2662};
2663
2664static void tgen_addcio(TCGContext *s, TCGType type,
2665                        TCGReg a0, TCGReg a1, TCGReg a2)
2666{
2667    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2668    tgen_arithr(s, ARITH_ADC + rexw, a0, a2);
2669}
2670
2671static void tgen_addcio_imm(TCGContext *s, TCGType type,
2672                            TCGReg a0, TCGReg a1, tcg_target_long a2)
2673{
2674    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2675    tgen_arithi(s, ARITH_ADC + rexw, a0, a2, true);
2676}
2677
2678static const TCGOutOpBinary outop_addcio = {
2679    .base.static_constraint = C_O1_I2(r, 0, re),
2680    .out_rrr = tgen_addcio,
2681    .out_rri = tgen_addcio_imm,
2682};
2683
2684static void tgen_addci_rrr(TCGContext *s, TCGType type,
2685                           TCGReg a0, TCGReg a1, TCGReg a2)
2686{
2687    /* Because "0O" is not a valid constraint, we must match ourselves. */
2688    if (a0 == a2) {
2689        tgen_addcio(s, type, a0, a0, a1);
2690    } else {
2691        tcg_out_mov(s, type, a0, a1);
2692        tgen_addcio(s, type, a0, a0, a2);
2693    }
2694}
2695
2696static void tgen_addci_rri(TCGContext *s, TCGType type,
2697                           TCGReg a0, TCGReg a1, tcg_target_long a2)
2698{
2699    tcg_out_mov(s, type, a0, a1);
2700    tgen_addcio_imm(s, type, a0, a0, a2);
2701}
2702
2703static void tgen_addci_rir(TCGContext *s, TCGType type,
2704                           TCGReg a0, tcg_target_long a1, TCGReg a2)
2705{
2706    tgen_addci_rri(s, type, a0, a2, a1);
2707}
2708
2709static void tgen_addci_rii(TCGContext *s, TCGType type, TCGReg a0,
2710                           tcg_target_long a1, tcg_target_long a2)
2711{
2712    if (a2 == 0) {
2713        /* Implement 0 + 0 + C with -(x - x - c). */
2714        tgen_arithr(s, ARITH_SBB, a0, a0);
2715        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, a0);
2716    } else {
2717        tcg_out_movi(s, type, a0, a2);
2718        tgen_addcio_imm(s, type, a0, a0, a1);
2719    }
2720}
2721
2722static const TCGOutOpAddSubCarry outop_addci = {
2723    .base.static_constraint = C_O1_I2(r, rO, re),
2724    .out_rrr = tgen_addci_rrr,
2725    .out_rri = tgen_addci_rri,
2726    .out_rir = tgen_addci_rir,
2727    .out_rii = tgen_addci_rii,
2728};
2729
2730static void tcg_out_set_carry(TCGContext *s)
2731{
2732    tcg_out8(s, OPC_STC);
2733}
2734
2735static void tgen_and(TCGContext *s, TCGType type,
2736                     TCGReg a0, TCGReg a1, TCGReg a2)
2737{
2738    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2739    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2740}
2741
2742static void tgen_andi(TCGContext *s, TCGType type,
2743                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2744{
2745    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2746    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2747}
2748
2749static const TCGOutOpBinary outop_and = {
2750    .base.static_constraint = C_O1_I2(r, 0, reZ),
2751    .out_rrr = tgen_and,
2752    .out_rri = tgen_andi,
2753};
2754
2755static void tgen_andc(TCGContext *s, TCGType type,
2756                      TCGReg a0, TCGReg a1, TCGReg a2)
2757{
2758    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2759    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2760}
2761
2762static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2763{
2764    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2765}
2766
2767static const TCGOutOpBinary outop_andc = {
2768    .base.static_constraint = C_Dynamic,
2769    .base.dynamic_constraint = cset_andc,
2770    .out_rrr = tgen_andc,
2771};
2772
2773static void tgen_clz(TCGContext *s, TCGType type,
2774                     TCGReg a0, TCGReg a1, TCGReg a2)
2775{
2776    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2777    int jcc;
2778
2779    if (have_lzcnt) {
2780        tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2781        jcc = JCC_JB;
2782    } else {
2783        /* Recall that the output of BSR is the index not the count.  */
2784        tcg_out_modrm(s, OPC_BSR + rexw, a0, a1);
2785        tgen_arithi(s, ARITH_XOR + rexw, a0, rexw ? 63 : 31, 0);
2786
2787        /* Since we have destroyed the flags from BSR, we have to re-test.  */
2788        jcc = tcg_out_cmp(s, TCG_COND_EQ, a1, 0, 1, rexw);
2789    }
2790    tcg_out_cmov(s, jcc, rexw, a0, a2);
2791}
2792
2793static void tgen_clzi(TCGContext *s, TCGType type,
2794                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2795{
2796    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2797    tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2798}
2799
2800static TCGConstraintSetIndex cset_clz(TCGType type, unsigned flags)
2801{
2802    return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2803}
2804
2805static const TCGOutOpBinary outop_clz = {
2806    .base.static_constraint = C_Dynamic,
2807    .base.dynamic_constraint = cset_clz,
2808    .out_rrr = tgen_clz,
2809    .out_rri = tgen_clzi,
2810};
2811
2812static void tgen_ctpop(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2813{
2814    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2815    tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2816}
2817
2818static TCGConstraintSetIndex cset_ctpop(TCGType type, unsigned flags)
2819{
2820    return have_popcnt ? C_O1_I1(r, r) : C_NotImplemented;
2821}
2822
2823static const TCGOutOpUnary outop_ctpop = {
2824    .base.static_constraint = C_Dynamic,
2825    .base.dynamic_constraint = cset_ctpop,
2826    .out_rr = tgen_ctpop,
2827};
2828
2829static void tgen_ctz(TCGContext *s, TCGType type,
2830                     TCGReg a0, TCGReg a1, TCGReg a2)
2831{
2832    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2833    int jcc;
2834
2835    if (have_bmi1) {
2836        tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2837        jcc = JCC_JB;
2838    } else {
2839        tcg_out_modrm(s, OPC_BSF + rexw, a0, a1);
2840        jcc = JCC_JE;
2841    }
2842    tcg_out_cmov(s, jcc, rexw, a0, a2);
2843}
2844
2845static void tgen_ctzi(TCGContext *s, TCGType type,
2846                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2847{
2848    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2849    tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2850}
2851
2852static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
2853{
2854    return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2855}
2856
2857static const TCGOutOpBinary outop_ctz = {
2858    .base.static_constraint = C_Dynamic,
2859    .base.dynamic_constraint = cset_ctz,
2860    .out_rrr = tgen_ctz,
2861    .out_rri = tgen_ctzi,
2862};
2863
2864static const TCGOutOpBinary outop_divs = {
2865    .base.static_constraint = C_NotImplemented,
2866};
2867
2868static void tgen_divs2(TCGContext *s, TCGType type,
2869                       TCGReg a0, TCGReg a1, TCGReg a4)
2870{
2871    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2872    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, a4);
2873}
2874
2875static const TCGOutOpDivRem outop_divs2 = {
2876    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2877    .out_rr01r = tgen_divs2,
2878};
2879
2880static const TCGOutOpBinary outop_divu = {
2881    .base.static_constraint = C_NotImplemented,
2882};
2883
2884static void tgen_divu2(TCGContext *s, TCGType type,
2885                       TCGReg a0, TCGReg a1, TCGReg a4)
2886{
2887    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2888    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, a4);
2889}
2890
2891static const TCGOutOpDivRem outop_divu2 = {
2892    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2893    .out_rr01r = tgen_divu2,
2894};
2895
2896static const TCGOutOpBinary outop_eqv = {
2897    .base.static_constraint = C_NotImplemented,
2898};
2899
2900#if TCG_TARGET_REG_BITS == 64
2901static void tgen_extrh_i64_i32(TCGContext *s, TCGType t, TCGReg a0, TCGReg a1)
2902{
2903    tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2904}
2905
2906static const TCGOutOpUnary outop_extrh_i64_i32 = {
2907    .base.static_constraint = C_O1_I1(r, 0),
2908    .out_rr = tgen_extrh_i64_i32,
2909};
2910#endif /* TCG_TARGET_REG_BITS == 64 */
2911
2912static void tgen_mul(TCGContext *s, TCGType type,
2913                     TCGReg a0, TCGReg a1, TCGReg a2)
2914{
2915    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2916    tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2917}
2918
2919static void tgen_muli(TCGContext *s, TCGType type,
2920                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2921{
2922    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2923
2924    if (a2 == (int8_t)a2) {
2925        tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2926        tcg_out8(s, a2);
2927    } else {
2928        tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2929        tcg_out32(s, a2);
2930    }
2931}
2932
2933static const TCGOutOpBinary outop_mul = {
2934    .base.static_constraint = C_O1_I2(r, 0, re),
2935    .out_rrr = tgen_mul,
2936    .out_rri = tgen_muli,
2937};
2938
2939static void tgen_muls2(TCGContext *s, TCGType type,
2940                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2941{
2942    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2943    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, a3);
2944}
2945
2946static const TCGOutOpMul2 outop_muls2 = {
2947    .base.static_constraint = C_O2_I2(a, d, a, r),
2948    .out_rrrr = tgen_muls2,
2949};
2950
2951static const TCGOutOpBinary outop_mulsh = {
2952    .base.static_constraint = C_NotImplemented,
2953};
2954
2955static const TCGOutOpBinary outop_muluh = {
2956    .base.static_constraint = C_NotImplemented,
2957};
2958
2959static void tgen_mulu2(TCGContext *s, TCGType type,
2960                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2961{
2962    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2963    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, a3);
2964}
2965
2966static const TCGOutOpMul2 outop_mulu2 = {
2967    .base.static_constraint = C_O2_I2(a, d, a, r),
2968    .out_rrrr = tgen_mulu2,
2969};
2970
2971static const TCGOutOpBinary outop_nand = {
2972    .base.static_constraint = C_NotImplemented,
2973};
2974
2975static const TCGOutOpBinary outop_nor = {
2976    .base.static_constraint = C_NotImplemented,
2977};
2978
2979static void tgen_or(TCGContext *s, TCGType type,
2980                    TCGReg a0, TCGReg a1, TCGReg a2)
2981{
2982    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2983    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
2984}
2985
2986static void tgen_ori(TCGContext *s, TCGType type,
2987                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2988{
2989    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2990    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
2991}
2992
2993static const TCGOutOpBinary outop_or = {
2994    .base.static_constraint = C_O1_I2(r, 0, re),
2995    .out_rrr = tgen_or,
2996    .out_rri = tgen_ori,
2997};
2998
2999static const TCGOutOpBinary outop_orc = {
3000    .base.static_constraint = C_NotImplemented,
3001};
3002
3003static const TCGOutOpBinary outop_rems = {
3004    .base.static_constraint = C_NotImplemented,
3005};
3006
3007static const TCGOutOpBinary outop_remu = {
3008    .base.static_constraint = C_NotImplemented,
3009};
3010
3011static void tgen_rotl(TCGContext *s, TCGType type,
3012                     TCGReg a0, TCGReg a1, TCGReg a2)
3013{
3014    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3015    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROL, a0);
3016}
3017
3018static void tgen_rotli(TCGContext *s, TCGType type,
3019                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3020{
3021    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3022    tcg_out_shifti(s, SHIFT_ROL + rexw, a0, a2);
3023}
3024
3025static const TCGOutOpBinary outop_rotl = {
3026    .base.static_constraint = C_O1_I2(r, 0, ci),
3027    .out_rrr = tgen_rotl,
3028    .out_rri = tgen_rotli,
3029};
3030
3031static void tgen_rotr(TCGContext *s, TCGType type,
3032                     TCGReg a0, TCGReg a1, TCGReg a2)
3033{
3034    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3035    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROR, a0);
3036}
3037
3038static void tgen_rotri(TCGContext *s, TCGType type,
3039                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3040{
3041    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3042    tcg_out_shifti(s, SHIFT_ROR + rexw, a0, a2);
3043}
3044
3045static const TCGOutOpBinary outop_rotr = {
3046    .base.static_constraint = C_O1_I2(r, 0, ci),
3047    .out_rrr = tgen_rotr,
3048    .out_rri = tgen_rotri,
3049};
3050
3051static TCGConstraintSetIndex cset_shift(TCGType type, unsigned flags)
3052{
3053    return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3054}
3055
3056static void tgen_sar(TCGContext *s, TCGType type,
3057                     TCGReg a0, TCGReg a1, TCGReg a2)
3058{
3059    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3060    if (have_bmi2) {
3061        tcg_out_vex_modrm(s, OPC_SARX + rexw, a0, a2, a1);
3062    } else {
3063        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SAR, a0);
3064    }
3065}
3066
3067static void tgen_sari(TCGContext *s, TCGType type,
3068                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3069{
3070    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3071
3072    tcg_out_mov(s, type, a0, a1);
3073    tcg_out_shifti(s, SHIFT_SAR + rexw, a0, a2);
3074}
3075
3076static const TCGOutOpBinary outop_sar = {
3077    .base.static_constraint = C_Dynamic,
3078    .base.dynamic_constraint = cset_shift,
3079    .out_rrr = tgen_sar,
3080    .out_rri = tgen_sari,
3081};
3082
3083static void tgen_shl(TCGContext *s, TCGType type,
3084                     TCGReg a0, TCGReg a1, TCGReg a2)
3085{
3086    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3087    if (have_bmi2) {
3088        tcg_out_vex_modrm(s, OPC_SHLX + rexw, a0, a2, a1);
3089    } else {
3090        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHL, a0);
3091    }
3092}
3093
3094static void tgen_shli(TCGContext *s, TCGType type,
3095                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3096{
3097    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3098
3099    /* For small constant 3-operand shift, use LEA.  */
3100    if (a0 != a1 && a2 >= 1 && a2 <= 3) {
3101        if (a2 == 1) {
3102            /* shl $1,a1,a0 -> lea (a1,a1),a0 */
3103            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
3104        } else {
3105            /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
3106            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
3107        }
3108        return;
3109    }
3110    tcg_out_mov(s, type, a0, a1);
3111    tcg_out_shifti(s, SHIFT_SHL + rexw, a0, a2);
3112}
3113
3114static const TCGOutOpBinary outop_shl = {
3115    .base.static_constraint = C_Dynamic,
3116    .base.dynamic_constraint = cset_shift,
3117    .out_rrr = tgen_shl,
3118    .out_rri = tgen_shli,
3119};
3120
3121static void tgen_shr(TCGContext *s, TCGType type,
3122                     TCGReg a0, TCGReg a1, TCGReg a2)
3123{
3124    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3125    if (have_bmi2) {
3126        tcg_out_vex_modrm(s, OPC_SHRX + rexw, a0, a2, a1);
3127    } else {
3128        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHR, a0);
3129    }
3130}
3131
3132static void tgen_shri(TCGContext *s, TCGType type,
3133                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3134{
3135    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3136
3137    tcg_out_mov(s, type, a0, a1);
3138    tcg_out_shifti(s, SHIFT_SHR + rexw, a0, a2);
3139}
3140
3141static const TCGOutOpBinary outop_shr = {
3142    .base.static_constraint = C_Dynamic,
3143    .base.dynamic_constraint = cset_shift,
3144    .out_rrr = tgen_shr,
3145    .out_rri = tgen_shri,
3146};
3147
3148static void tgen_sub(TCGContext *s, TCGType type,
3149                     TCGReg a0, TCGReg a1, TCGReg a2)
3150{
3151    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3152    tgen_arithr(s, ARITH_SUB + rexw, a0, a2);
3153}
3154
3155static const TCGOutOpSubtract outop_sub = {
3156    .base.static_constraint = C_O1_I2(r, 0, r),
3157    .out_rrr = tgen_sub,
3158};
3159
3160static void tgen_subbo_rri(TCGContext *s, TCGType type,
3161                           TCGReg a0, TCGReg a1, tcg_target_long a2)
3162{
3163    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3164    tgen_arithi(s, ARITH_SUB + rexw, a0, a2, 1);
3165}
3166
3167static const TCGOutOpAddSubCarry outop_subbo = {
3168    .base.static_constraint = C_O1_I2(r, 0, re),
3169    .out_rrr = tgen_sub,
3170    .out_rri = tgen_subbo_rri,
3171};
3172
3173static void tgen_subbio_rrr(TCGContext *s, TCGType type,
3174                            TCGReg a0, TCGReg a1, TCGReg a2)
3175{
3176    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3177    tgen_arithr(s, ARITH_SBB + rexw, a0, a2);
3178}
3179
3180static void tgen_subbio_rri(TCGContext *s, TCGType type,
3181                            TCGReg a0, TCGReg a1, tcg_target_long a2)
3182{
3183    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3184    tgen_arithi(s, ARITH_SBB + rexw, a0, a2, 1);
3185}
3186
3187static const TCGOutOpAddSubCarry outop_subbio = {
3188    .base.static_constraint = C_O1_I2(r, 0, re),
3189    .out_rrr = tgen_subbio_rrr,
3190    .out_rri = tgen_subbio_rri,
3191};
3192
3193#define outop_subbi  outop_subbio
3194
3195static void tcg_out_set_borrow(TCGContext *s)
3196{
3197    tcg_out8(s, OPC_STC);
3198}
3199
3200static void tgen_xor(TCGContext *s, TCGType type,
3201                     TCGReg a0, TCGReg a1, TCGReg a2)
3202{
3203    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3204    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
3205}
3206
3207static void tgen_xori(TCGContext *s, TCGType type,
3208                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3209{
3210    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3211    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
3212}
3213
3214static const TCGOutOpBinary outop_xor = {
3215    .base.static_constraint = C_O1_I2(r, 0, re),
3216    .out_rrr = tgen_xor,
3217    .out_rri = tgen_xori,
3218};
3219
3220static void tgen_bswap16(TCGContext *s, TCGType type,
3221                         TCGReg a0, TCGReg a1, unsigned flags)
3222{
3223    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3224
3225    if (flags & TCG_BSWAP_OS) {
3226        /* Output must be sign-extended. */
3227        if (rexw) {
3228            tcg_out_bswap64(s, a0);
3229            tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
3230        } else {
3231            tcg_out_bswap32(s, a0);
3232            tcg_out_shifti(s, SHIFT_SAR, a0, 16);
3233        }
3234    } else if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
3235        /* Output must be zero-extended, but input isn't. */
3236        tcg_out_bswap32(s, a0);
3237        tcg_out_shifti(s, SHIFT_SHR, a0, 16);
3238    } else {
3239        tcg_out_rolw_8(s, a0);
3240    }
3241}
3242
3243static const TCGOutOpBswap outop_bswap16 = {
3244    .base.static_constraint = C_O1_I1(r, 0),
3245    .out_rr = tgen_bswap16,
3246};
3247
3248static void tgen_bswap32(TCGContext *s, TCGType type,
3249                         TCGReg a0, TCGReg a1, unsigned flags)
3250{
3251    tcg_out_bswap32(s, a0);
3252    if (flags & TCG_BSWAP_OS) {
3253        tcg_out_ext32s(s, a0, a0);
3254    }
3255}
3256
3257static const TCGOutOpBswap outop_bswap32 = {
3258    .base.static_constraint = C_O1_I1(r, 0),
3259    .out_rr = tgen_bswap32,
3260};
3261
3262#if TCG_TARGET_REG_BITS == 64
3263static void tgen_bswap64(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3264{
3265    tcg_out_bswap64(s, a0);
3266}
3267
3268static const TCGOutOpUnary outop_bswap64 = {
3269    .base.static_constraint = C_O1_I1(r, 0),
3270    .out_rr = tgen_bswap64,
3271};
3272#endif
3273
3274static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3275{
3276    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3277    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
3278}
3279
3280static const TCGOutOpUnary outop_neg = {
3281    .base.static_constraint = C_O1_I1(r, 0),
3282    .out_rr = tgen_neg,
3283};
3284
3285static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3286{
3287    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3288    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
3289}
3290
3291static const TCGOutOpUnary outop_not = {
3292    .base.static_constraint = C_O1_I1(r, 0),
3293    .out_rr = tgen_not,
3294};
3295
3296static void tgen_deposit(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3297                         TCGReg a2, unsigned ofs, unsigned len)
3298{
3299    if (ofs == 0 && len == 8) {
3300        tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3301    } else if (ofs == 0 && len == 16) {
3302        tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3303    } else if (TCG_TARGET_REG_BITS == 32 && ofs == 8 && len == 8) {
3304        tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3305    } else {
3306        g_assert_not_reached();
3307    }
3308}
3309
3310static void tgen_depositi(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3311                          tcg_target_long a2, unsigned ofs, unsigned len)
3312{
3313    if (ofs == 0 && len == 8) {
3314        tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0), 0, a0, 0);
3315        tcg_out8(s, a2);
3316    } else if (ofs == 0 && len == 16) {
3317        tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0), 0, a0, 0);
3318        tcg_out16(s, a2);
3319    } else if (TCG_TARGET_REG_BITS == 32 && ofs == 8 && len == 8) {
3320        tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3321        tcg_out8(s, a2);
3322    } else {
3323        g_assert_not_reached();
3324    }
3325}
3326
3327static const TCGOutOpDeposit outop_deposit = {
3328    .base.static_constraint = C_O1_I2(q, 0, qi),
3329    .out_rrr = tgen_deposit,
3330    .out_rri = tgen_depositi,
3331};
3332
3333static void tgen_extract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3334                         unsigned ofs, unsigned len)
3335{
3336    if (ofs == 0) {
3337        switch (len) {
3338        case 8:
3339            tcg_out_ext8u(s, a0, a1);
3340            return;
3341        case 16:
3342            tcg_out_ext16u(s, a0, a1);
3343            return;
3344        case 32:
3345            tcg_out_ext32u(s, a0, a1);
3346            return;
3347        }
3348    } else if (TCG_TARGET_REG_BITS == 64 && ofs + len == 32) {
3349        /* This is a 32-bit zero-extending right shift.  */
3350        tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3351        tcg_out_shifti(s, SHIFT_SHR, a0, ofs);
3352        return;
3353    } else if (ofs == 8 && len == 8) {
3354        /*
3355         * On the off-chance that we can use the high-byte registers.
3356         * Otherwise we emit the same ext16 + shift pattern that we
3357         * would have gotten from the normal tcg-op.c expansion.
3358         */
3359        if (a1 < 4 && (TCG_TARGET_REG_BITS == 32 || a0 < 8)) {
3360            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3361        } else {
3362            tcg_out_ext16u(s, a0, a1);
3363            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3364        }
3365        return;
3366    }
3367    g_assert_not_reached();
3368}
3369
3370static const TCGOutOpExtract outop_extract = {
3371    .base.static_constraint = C_O1_I1(r, r),
3372    .out_rr = tgen_extract,
3373};
3374
3375static void tgen_sextract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3376                          unsigned ofs, unsigned len)
3377{
3378    if (ofs == 0) {
3379        switch (len) {
3380        case 8:
3381            tcg_out_ext8s(s, type, a0, a1);
3382            return;
3383        case 16:
3384            tcg_out_ext16s(s, type, a0, a1);
3385            return;
3386        case 32:
3387            tcg_out_ext32s(s, a0, a1);
3388            return;
3389        }
3390    } else if (ofs == 8 && len == 8) {
3391        if (type == TCG_TYPE_I32 && a1 < 4 && a0 < 8) {
3392            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3393        } else {
3394            tcg_out_ext16s(s, type, a0, a1);
3395            tgen_sari(s, type, a0, a0, 8);
3396        }
3397        return;
3398    }
3399    g_assert_not_reached();
3400}
3401
3402static const TCGOutOpExtract outop_sextract = {
3403    .base.static_constraint = C_O1_I1(r, r),
3404    .out_rr = tgen_sextract,
3405};
3406
3407static void tgen_extract2(TCGContext *s, TCGType type, TCGReg a0,
3408                          TCGReg a1, TCGReg a2, unsigned shr)
3409{
3410    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3411
3412    /* Note that SHRD outputs to the r/m operand.  */
3413    tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3414    tcg_out8(s, shr);
3415}
3416
3417static const TCGOutOpExtract2 outop_extract2 = {
3418    .base.static_constraint = C_O1_I2(r, 0, r),
3419    .out_rrr = tgen_extract2,
3420};
3421
3422static void tgen_ld8u(TCGContext *s, TCGType type, TCGReg dest,
3423                      TCGReg base, ptrdiff_t offset)
3424{
3425    tcg_out_modrm_offset(s, OPC_MOVZBL, dest, base, offset);
3426}
3427
3428static const TCGOutOpLoad outop_ld8u = {
3429    .base.static_constraint = C_O1_I1(r, r),
3430    .out = tgen_ld8u,
3431};
3432
3433static void tgen_ld8s(TCGContext *s, TCGType type, TCGReg dest,
3434                      TCGReg base, ptrdiff_t offset)
3435{
3436    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3437    tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, dest, base, offset);
3438}
3439
3440static const TCGOutOpLoad outop_ld8s = {
3441    .base.static_constraint = C_O1_I1(r, r),
3442    .out = tgen_ld8s,
3443};
3444
3445static void tgen_ld16u(TCGContext *s, TCGType type, TCGReg dest,
3446                       TCGReg base, ptrdiff_t offset)
3447{
3448    tcg_out_modrm_offset(s, OPC_MOVZWL, dest, base, offset);
3449}
3450
3451static const TCGOutOpLoad outop_ld16u = {
3452    .base.static_constraint = C_O1_I1(r, r),
3453    .out = tgen_ld16u,
3454};
3455
3456static void tgen_ld16s(TCGContext *s, TCGType type, TCGReg dest,
3457                       TCGReg base, ptrdiff_t offset)
3458{
3459    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3460    tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, dest, base, offset);
3461}
3462
3463static const TCGOutOpLoad outop_ld16s = {
3464    .base.static_constraint = C_O1_I1(r, r),
3465    .out = tgen_ld16s,
3466};
3467
3468#if TCG_TARGET_REG_BITS == 64
3469static void tgen_ld32u(TCGContext *s, TCGType type, TCGReg dest,
3470                       TCGReg base, ptrdiff_t offset)
3471{
3472    tcg_out_modrm_offset(s, OPC_MOVL_GvEv, dest, base, offset);
3473}
3474
3475static const TCGOutOpLoad outop_ld32u = {
3476    .base.static_constraint = C_O1_I1(r, r),
3477    .out = tgen_ld32u,
3478};
3479
3480static void tgen_ld32s(TCGContext *s, TCGType type, TCGReg dest,
3481                       TCGReg base, ptrdiff_t offset)
3482{
3483    tcg_out_modrm_offset(s, OPC_MOVSLQ, dest, base, offset);
3484}
3485
3486static const TCGOutOpLoad outop_ld32s = {
3487    .base.static_constraint = C_O1_I1(r, r),
3488    .out = tgen_ld32s,
3489};
3490#endif
3491
3492static void tgen_st8_r(TCGContext *s, TCGType type, TCGReg data,
3493                       TCGReg base, ptrdiff_t offset)
3494{
3495    tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, data, base, offset);
3496}
3497
3498static void tgen_st8_i(TCGContext *s, TCGType type, tcg_target_long data,
3499                       TCGReg base, ptrdiff_t offset)
3500{
3501    tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, base, offset);
3502    tcg_out8(s, data);
3503}
3504
3505static const TCGOutOpStore outop_st8 = {
3506    .base.static_constraint = C_O0_I2(qi, r),
3507    .out_r = tgen_st8_r,
3508    .out_i = tgen_st8_i,
3509};
3510
3511static void tgen_st16_r(TCGContext *s, TCGType type, TCGReg data,
3512                        TCGReg base, ptrdiff_t offset)
3513{
3514    tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, data, base, offset);
3515}
3516
3517static void tgen_st16_i(TCGContext *s, TCGType type, tcg_target_long data,
3518                        TCGReg base, ptrdiff_t offset)
3519{
3520    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, base, offset);
3521    tcg_out16(s, data);
3522}
3523
3524static const TCGOutOpStore outop_st16 = {
3525    .base.static_constraint = C_O0_I2(ri, r),
3526    .out_r = tgen_st16_r,
3527    .out_i = tgen_st16_i,
3528};
3529
3530static void tgen_st_i(TCGContext *s, TCGType type, tcg_target_long data,
3531                      TCGReg base, ptrdiff_t offset)
3532{
3533    bool ok = tcg_out_sti(s, type, data, base, offset);
3534    tcg_debug_assert(ok);
3535}
3536
3537static const TCGOutOpStore outop_st = {
3538    .base.static_constraint = C_O0_I2(re, r),
3539    .out_r = tcg_out_st,
3540    .out_i = tgen_st_i,
3541};
3542
3543static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
3544                       const TCGArg args[TCG_MAX_OP_ARGS],
3545                       const int const_args[TCG_MAX_OP_ARGS])
3546{
3547    TCGArg a0, a1, a2;
3548
3549    /* Hoist the loads of the most common arguments.  */
3550    a0 = args[0];
3551    a1 = args[1];
3552    a2 = args[2];
3553
3554    switch (opc) {
3555    case INDEX_op_qemu_ld_i32:
3556        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
3557        break;
3558    case INDEX_op_qemu_ld_i64:
3559        if (TCG_TARGET_REG_BITS == 64) {
3560            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
3561        } else {
3562            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3563        }
3564        break;
3565    case INDEX_op_qemu_ld_i128:
3566        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3567        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3568        break;
3569
3570    case INDEX_op_qemu_st_i32:
3571        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
3572        break;
3573    case INDEX_op_qemu_st_i64:
3574        if (TCG_TARGET_REG_BITS == 64) {
3575            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
3576        } else {
3577            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3578        }
3579        break;
3580    case INDEX_op_qemu_st_i128:
3581        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3582        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3583        break;
3584
3585    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3586    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3587    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3588    default:
3589        g_assert_not_reached();
3590    }
3591}
3592
3593static int const umin_insn[4] = {
3594    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3595};
3596
3597static int const umax_insn[4] = {
3598    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3599};
3600
3601static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3602                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3603{
3604    static int const cmpeq_insn[4] = {
3605        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3606    };
3607    static int const cmpgt_insn[4] = {
3608        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3609    };
3610
3611    enum {
3612        NEED_INV  = 1,
3613        NEED_SWAP = 2,
3614        NEED_UMIN = 4,
3615        NEED_UMAX = 8,
3616        INVALID   = 16,
3617    };
3618    static const uint8_t cond_fixup[16] = {
3619        [0 ... 15] = INVALID,
3620        [TCG_COND_EQ] = 0,
3621        [TCG_COND_GT] = 0,
3622        [TCG_COND_NE] = NEED_INV,
3623        [TCG_COND_LE] = NEED_INV,
3624        [TCG_COND_LT] = NEED_SWAP,
3625        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3626        [TCG_COND_LEU] = NEED_UMIN,
3627        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3628        [TCG_COND_GEU] = NEED_UMAX,
3629        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3630    };
3631    int fixup = cond_fixup[cond];
3632
3633    assert(!(fixup & INVALID));
3634
3635    if (fixup & NEED_INV) {
3636        cond = tcg_invert_cond(cond);
3637    }
3638
3639    if (fixup & NEED_SWAP) {
3640        TCGReg swap = v1;
3641        v1 = v2;
3642        v2 = swap;
3643        cond = tcg_swap_cond(cond);
3644    }
3645
3646    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3647        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3648
3649        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3650        assert(vece <= MO_32);
3651
3652        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3653        v2 = TCG_TMP_VEC;
3654        cond = TCG_COND_EQ;
3655    }
3656
3657    switch (cond) {
3658    case TCG_COND_EQ:
3659        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3660        break;
3661    case TCG_COND_GT:
3662        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3663        break;
3664    default:
3665        g_assert_not_reached();
3666    }
3667    return fixup & NEED_INV;
3668}
3669
3670static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3671                               TCGReg v1, TCGReg v2, TCGCond cond)
3672{
3673    static const int cmpm_insn[2][4] = {
3674        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3675        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3676    };
3677    static const int testm_insn[4] = {
3678        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3679    };
3680    static const int testnm_insn[4] = {
3681        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3682    };
3683
3684    static const int cond_ext[16] = {
3685        [TCG_COND_EQ] = 0,
3686        [TCG_COND_NE] = 4,
3687        [TCG_COND_LT] = 1,
3688        [TCG_COND_LTU] = 1,
3689        [TCG_COND_LE] = 2,
3690        [TCG_COND_LEU] = 2,
3691        [TCG_COND_NEVER] = 3,
3692        [TCG_COND_GE] = 5,
3693        [TCG_COND_GEU] = 5,
3694        [TCG_COND_GT] = 6,
3695        [TCG_COND_GTU] = 6,
3696        [TCG_COND_ALWAYS] = 7,
3697    };
3698
3699    switch (cond) {
3700    case TCG_COND_TSTNE:
3701        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3702        break;
3703    case TCG_COND_TSTEQ:
3704        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3705        break;
3706    default:
3707        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3708                               /* k1 */ 1, v1, v2, type);
3709        tcg_out8(s, cond_ext[cond]);
3710        break;
3711    }
3712}
3713
3714static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3715                              unsigned vece, TCGReg dest)
3716{
3717    static const int movm_insn[] = {
3718        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3719    };
3720    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3721}
3722
3723static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3724                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3725{
3726    /*
3727     * With avx512, we have a complete set of comparisons into mask.
3728     * Unless there's a single insn expansion for the comparision,
3729     * expand via a mask in k1.
3730     */
3731    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3732        && cond != TCG_COND_EQ
3733        && cond != TCG_COND_LT
3734        && cond != TCG_COND_GT) {
3735        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3736        tcg_out_k1_to_vec(s, type, vece, v0);
3737        return;
3738    }
3739
3740    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3741        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3742        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3743    }
3744}
3745
3746static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3747                                  TCGReg v0, TCGReg c1, TCGReg c2,
3748                                  TCGReg v3, TCGReg v4, TCGCond cond)
3749{
3750    static const int vpblendm_insn[] = {
3751        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3752    };
3753    bool z = false;
3754
3755    /* Swap to place constant in V4 to take advantage of zero-masking. */
3756    if (!v3) {
3757        z = true;
3758        v3 = v4;
3759        cond = tcg_invert_cond(cond);
3760    }
3761
3762    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3763    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3764                            /* k1 */1, z, type);
3765}
3766
3767static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3768                               TCGReg v0, TCGReg c1, TCGReg c2,
3769                               TCGReg v3, TCGReg v4, TCGCond cond)
3770{
3771    bool inv;
3772
3773    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3774        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3775        return;
3776    }
3777
3778    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3779
3780    /*
3781     * Since XMM0 is 16, the only way we get 0 into V3
3782     * is via the constant zero constraint.
3783     */
3784    if (!v3) {
3785        if (inv) {
3786            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3787        } else {
3788            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3789        }
3790    } else {
3791        if (inv) {
3792            TCGReg swap = v3;
3793            v3 = v4;
3794            v4 = swap;
3795        }
3796        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3797        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3798    }
3799}
3800
3801static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3802                           unsigned vecl, unsigned vece,
3803                           const TCGArg args[TCG_MAX_OP_ARGS],
3804                           const int const_args[TCG_MAX_OP_ARGS])
3805{
3806    static int const add_insn[4] = {
3807        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3808    };
3809    static int const ssadd_insn[4] = {
3810        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3811    };
3812    static int const usadd_insn[4] = {
3813        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3814    };
3815    static int const sub_insn[4] = {
3816        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3817    };
3818    static int const sssub_insn[4] = {
3819        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3820    };
3821    static int const ussub_insn[4] = {
3822        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3823    };
3824    static int const mul_insn[4] = {
3825        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3826    };
3827    static int const shift_imm_insn[4] = {
3828        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3829    };
3830    static int const punpckl_insn[4] = {
3831        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3832    };
3833    static int const punpckh_insn[4] = {
3834        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3835    };
3836    static int const packss_insn[4] = {
3837        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3838    };
3839    static int const packus_insn[4] = {
3840        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3841    };
3842    static int const smin_insn[4] = {
3843        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3844    };
3845    static int const smax_insn[4] = {
3846        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3847    };
3848    static int const rotlv_insn[4] = {
3849        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3850    };
3851    static int const rotrv_insn[4] = {
3852        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3853    };
3854    static int const shlv_insn[4] = {
3855        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3856    };
3857    static int const shrv_insn[4] = {
3858        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3859    };
3860    static int const sarv_insn[4] = {
3861        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3862    };
3863    static int const shls_insn[4] = {
3864        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3865    };
3866    static int const shrs_insn[4] = {
3867        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3868    };
3869    static int const sars_insn[4] = {
3870        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3871    };
3872    static int const vpshldi_insn[4] = {
3873        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3874    };
3875    static int const vpshldv_insn[4] = {
3876        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3877    };
3878    static int const vpshrdv_insn[4] = {
3879        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3880    };
3881    static int const abs_insn[4] = {
3882        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3883    };
3884
3885    TCGType type = vecl + TCG_TYPE_V64;
3886    int insn, sub;
3887    TCGArg a0, a1, a2, a3;
3888
3889    a0 = args[0];
3890    a1 = args[1];
3891    a2 = args[2];
3892
3893    switch (opc) {
3894    case INDEX_op_add_vec:
3895        insn = add_insn[vece];
3896        goto gen_simd;
3897    case INDEX_op_ssadd_vec:
3898        insn = ssadd_insn[vece];
3899        goto gen_simd;
3900    case INDEX_op_usadd_vec:
3901        insn = usadd_insn[vece];
3902        goto gen_simd;
3903    case INDEX_op_sub_vec:
3904        insn = sub_insn[vece];
3905        goto gen_simd;
3906    case INDEX_op_sssub_vec:
3907        insn = sssub_insn[vece];
3908        goto gen_simd;
3909    case INDEX_op_ussub_vec:
3910        insn = ussub_insn[vece];
3911        goto gen_simd;
3912    case INDEX_op_mul_vec:
3913        insn = mul_insn[vece];
3914        goto gen_simd;
3915    case INDEX_op_and_vec:
3916        insn = OPC_PAND;
3917        goto gen_simd;
3918    case INDEX_op_or_vec:
3919        insn = OPC_POR;
3920        goto gen_simd;
3921    case INDEX_op_xor_vec:
3922        insn = OPC_PXOR;
3923        goto gen_simd;
3924    case INDEX_op_smin_vec:
3925        insn = smin_insn[vece];
3926        goto gen_simd;
3927    case INDEX_op_umin_vec:
3928        insn = umin_insn[vece];
3929        goto gen_simd;
3930    case INDEX_op_smax_vec:
3931        insn = smax_insn[vece];
3932        goto gen_simd;
3933    case INDEX_op_umax_vec:
3934        insn = umax_insn[vece];
3935        goto gen_simd;
3936    case INDEX_op_shlv_vec:
3937        insn = shlv_insn[vece];
3938        goto gen_simd;
3939    case INDEX_op_shrv_vec:
3940        insn = shrv_insn[vece];
3941        goto gen_simd;
3942    case INDEX_op_sarv_vec:
3943        insn = sarv_insn[vece];
3944        goto gen_simd;
3945    case INDEX_op_rotlv_vec:
3946        insn = rotlv_insn[vece];
3947        goto gen_simd;
3948    case INDEX_op_rotrv_vec:
3949        insn = rotrv_insn[vece];
3950        goto gen_simd;
3951    case INDEX_op_shls_vec:
3952        insn = shls_insn[vece];
3953        goto gen_simd;
3954    case INDEX_op_shrs_vec:
3955        insn = shrs_insn[vece];
3956        goto gen_simd;
3957    case INDEX_op_sars_vec:
3958        insn = sars_insn[vece];
3959        goto gen_simd;
3960    case INDEX_op_x86_punpckl_vec:
3961        insn = punpckl_insn[vece];
3962        goto gen_simd;
3963    case INDEX_op_x86_punpckh_vec:
3964        insn = punpckh_insn[vece];
3965        goto gen_simd;
3966    case INDEX_op_x86_packss_vec:
3967        insn = packss_insn[vece];
3968        goto gen_simd;
3969    case INDEX_op_x86_packus_vec:
3970        insn = packus_insn[vece];
3971        goto gen_simd;
3972    case INDEX_op_x86_vpshldv_vec:
3973        insn = vpshldv_insn[vece];
3974        a1 = a2;
3975        a2 = args[3];
3976        goto gen_simd;
3977    case INDEX_op_x86_vpshrdv_vec:
3978        insn = vpshrdv_insn[vece];
3979        a1 = a2;
3980        a2 = args[3];
3981        goto gen_simd;
3982#if TCG_TARGET_REG_BITS == 32
3983    case INDEX_op_dup2_vec:
3984        /* First merge the two 32-bit inputs to a single 64-bit element. */
3985        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3986        /* Then replicate the 64-bit elements across the rest of the vector. */
3987        if (type != TCG_TYPE_V64) {
3988            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3989        }
3990        break;
3991#endif
3992    case INDEX_op_abs_vec:
3993        insn = abs_insn[vece];
3994        a2 = a1;
3995        a1 = 0;
3996        goto gen_simd;
3997    gen_simd:
3998        tcg_debug_assert(insn != OPC_UD2);
3999        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
4000        break;
4001
4002    case INDEX_op_cmp_vec:
4003        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
4004        break;
4005
4006    case INDEX_op_cmpsel_vec:
4007        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
4008                           args[3], args[4], args[5]);
4009        break;
4010
4011    case INDEX_op_andc_vec:
4012        insn = OPC_PANDN;
4013        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
4014        break;
4015
4016    case INDEX_op_shli_vec:
4017        insn = shift_imm_insn[vece];
4018        sub = 6;
4019        goto gen_shift;
4020    case INDEX_op_shri_vec:
4021        insn = shift_imm_insn[vece];
4022        sub = 2;
4023        goto gen_shift;
4024    case INDEX_op_sari_vec:
4025        if (vece == MO_64) {
4026            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
4027        } else {
4028            insn = shift_imm_insn[vece];
4029        }
4030        sub = 4;
4031        goto gen_shift;
4032    case INDEX_op_rotli_vec:
4033        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
4034        if (vece == MO_64) {
4035            insn |= P_VEXW;
4036        }
4037        sub = 1;
4038        goto gen_shift;
4039    gen_shift:
4040        tcg_debug_assert(vece != MO_8);
4041        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
4042        tcg_out8(s, a2);
4043        break;
4044
4045    case INDEX_op_ld_vec:
4046        tcg_out_ld(s, type, a0, a1, a2);
4047        break;
4048    case INDEX_op_st_vec:
4049        tcg_out_st(s, type, a0, a1, a2);
4050        break;
4051    case INDEX_op_dupm_vec:
4052        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
4053        break;
4054
4055    case INDEX_op_x86_shufps_vec:
4056        insn = OPC_SHUFPS;
4057        sub = args[3];
4058        goto gen_simd_imm8;
4059    case INDEX_op_x86_blend_vec:
4060        if (vece == MO_16) {
4061            insn = OPC_PBLENDW;
4062        } else if (vece == MO_32) {
4063            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
4064        } else {
4065            g_assert_not_reached();
4066        }
4067        sub = args[3];
4068        goto gen_simd_imm8;
4069    case INDEX_op_x86_vperm2i128_vec:
4070        insn = OPC_VPERM2I128;
4071        sub = args[3];
4072        goto gen_simd_imm8;
4073    case INDEX_op_x86_vpshldi_vec:
4074        insn = vpshldi_insn[vece];
4075        sub = args[3];
4076        goto gen_simd_imm8;
4077
4078    case INDEX_op_not_vec:
4079        insn = OPC_VPTERNLOGQ;
4080        a2 = a1;
4081        sub = 0x33; /* !B */
4082        goto gen_simd_imm8;
4083    case INDEX_op_nor_vec:
4084        insn = OPC_VPTERNLOGQ;
4085        sub = 0x11; /* norCB */
4086        goto gen_simd_imm8;
4087    case INDEX_op_nand_vec:
4088        insn = OPC_VPTERNLOGQ;
4089        sub = 0x77; /* nandCB */
4090        goto gen_simd_imm8;
4091    case INDEX_op_eqv_vec:
4092        insn = OPC_VPTERNLOGQ;
4093        sub = 0x99; /* xnorCB */
4094        goto gen_simd_imm8;
4095    case INDEX_op_orc_vec:
4096        insn = OPC_VPTERNLOGQ;
4097        sub = 0xdd; /* orB!C */
4098        goto gen_simd_imm8;
4099
4100    case INDEX_op_bitsel_vec:
4101        insn = OPC_VPTERNLOGQ;
4102        a3 = args[3];
4103        if (a0 == a1) {
4104            a1 = a2;
4105            a2 = a3;
4106            sub = 0xca; /* A?B:C */
4107        } else if (a0 == a2) {
4108            a2 = a3;
4109            sub = 0xe2; /* B?A:C */
4110        } else {
4111            tcg_out_mov(s, type, a0, a3);
4112            sub = 0xb8; /* B?C:A */
4113        }
4114        goto gen_simd_imm8;
4115
4116    gen_simd_imm8:
4117        tcg_debug_assert(insn != OPC_UD2);
4118        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
4119        tcg_out8(s, sub);
4120        break;
4121
4122    case INDEX_op_x86_psrldq_vec:
4123        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
4124        tcg_out8(s, a2);
4125        break;
4126
4127    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
4128    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
4129    default:
4130        g_assert_not_reached();
4131    }
4132}
4133
4134static TCGConstraintSetIndex
4135tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
4136{
4137    switch (op) {
4138    case INDEX_op_qemu_ld_i32:
4139        return C_O1_I1(r, L);
4140
4141    case INDEX_op_qemu_st_i32:
4142        return (TCG_TARGET_REG_BITS == 32 && flags == MO_8
4143                ? C_O0_I2(s, L)
4144                : C_O0_I2(L, L));
4145
4146    case INDEX_op_qemu_ld_i64:
4147        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
4148
4149    case INDEX_op_qemu_st_i64:
4150        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
4151
4152    case INDEX_op_qemu_ld_i128:
4153        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4154        return C_O2_I1(r, r, L);
4155    case INDEX_op_qemu_st_i128:
4156        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4157        return C_O0_I3(L, L, L);
4158
4159    case INDEX_op_ld_vec:
4160    case INDEX_op_dupm_vec:
4161        return C_O1_I1(x, r);
4162
4163    case INDEX_op_st_vec:
4164        return C_O0_I2(x, r);
4165
4166    case INDEX_op_add_vec:
4167    case INDEX_op_sub_vec:
4168    case INDEX_op_mul_vec:
4169    case INDEX_op_and_vec:
4170    case INDEX_op_or_vec:
4171    case INDEX_op_xor_vec:
4172    case INDEX_op_andc_vec:
4173    case INDEX_op_orc_vec:
4174    case INDEX_op_nand_vec:
4175    case INDEX_op_nor_vec:
4176    case INDEX_op_eqv_vec:
4177    case INDEX_op_ssadd_vec:
4178    case INDEX_op_usadd_vec:
4179    case INDEX_op_sssub_vec:
4180    case INDEX_op_ussub_vec:
4181    case INDEX_op_smin_vec:
4182    case INDEX_op_umin_vec:
4183    case INDEX_op_smax_vec:
4184    case INDEX_op_umax_vec:
4185    case INDEX_op_shlv_vec:
4186    case INDEX_op_shrv_vec:
4187    case INDEX_op_sarv_vec:
4188    case INDEX_op_rotlv_vec:
4189    case INDEX_op_rotrv_vec:
4190    case INDEX_op_shls_vec:
4191    case INDEX_op_shrs_vec:
4192    case INDEX_op_sars_vec:
4193    case INDEX_op_cmp_vec:
4194    case INDEX_op_x86_shufps_vec:
4195    case INDEX_op_x86_blend_vec:
4196    case INDEX_op_x86_packss_vec:
4197    case INDEX_op_x86_packus_vec:
4198    case INDEX_op_x86_vperm2i128_vec:
4199    case INDEX_op_x86_punpckl_vec:
4200    case INDEX_op_x86_punpckh_vec:
4201    case INDEX_op_x86_vpshldi_vec:
4202#if TCG_TARGET_REG_BITS == 32
4203    case INDEX_op_dup2_vec:
4204#endif
4205        return C_O1_I2(x, x, x);
4206
4207    case INDEX_op_abs_vec:
4208    case INDEX_op_dup_vec:
4209    case INDEX_op_not_vec:
4210    case INDEX_op_shli_vec:
4211    case INDEX_op_shri_vec:
4212    case INDEX_op_sari_vec:
4213    case INDEX_op_rotli_vec:
4214    case INDEX_op_x86_psrldq_vec:
4215        return C_O1_I1(x, x);
4216
4217    case INDEX_op_x86_vpshldv_vec:
4218    case INDEX_op_x86_vpshrdv_vec:
4219        return C_O1_I3(x, 0, x, x);
4220
4221    case INDEX_op_bitsel_vec:
4222        return C_O1_I3(x, x, x, x);
4223    case INDEX_op_cmpsel_vec:
4224        return C_O1_I4(x, x, x, xO, x);
4225
4226    default:
4227        return C_NotImplemented;
4228    }
4229}
4230
4231int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
4232{
4233    switch (opc) {
4234    case INDEX_op_add_vec:
4235    case INDEX_op_sub_vec:
4236    case INDEX_op_and_vec:
4237    case INDEX_op_or_vec:
4238    case INDEX_op_xor_vec:
4239    case INDEX_op_andc_vec:
4240    case INDEX_op_orc_vec:
4241    case INDEX_op_nand_vec:
4242    case INDEX_op_nor_vec:
4243    case INDEX_op_eqv_vec:
4244    case INDEX_op_not_vec:
4245    case INDEX_op_bitsel_vec:
4246        return 1;
4247    case INDEX_op_cmp_vec:
4248    case INDEX_op_cmpsel_vec:
4249        return -1;
4250
4251    case INDEX_op_rotli_vec:
4252        return have_avx512vl && vece >= MO_32 ? 1 : -1;
4253
4254    case INDEX_op_shli_vec:
4255    case INDEX_op_shri_vec:
4256        /* We must expand the operation for MO_8.  */
4257        return vece == MO_8 ? -1 : 1;
4258
4259    case INDEX_op_sari_vec:
4260        switch (vece) {
4261        case MO_8:
4262            return -1;
4263        case MO_16:
4264        case MO_32:
4265            return 1;
4266        case MO_64:
4267            if (have_avx512vl) {
4268                return 1;
4269            }
4270            /*
4271             * We can emulate this for MO_64, but it does not pay off
4272             * unless we're producing at least 4 values.
4273             */
4274            return type >= TCG_TYPE_V256 ? -1 : 0;
4275        }
4276        return 0;
4277
4278    case INDEX_op_shls_vec:
4279    case INDEX_op_shrs_vec:
4280        return vece >= MO_16;
4281    case INDEX_op_sars_vec:
4282        switch (vece) {
4283        case MO_16:
4284        case MO_32:
4285            return 1;
4286        case MO_64:
4287            return have_avx512vl;
4288        }
4289        return 0;
4290    case INDEX_op_rotls_vec:
4291        return vece >= MO_16 ? -1 : 0;
4292
4293    case INDEX_op_shlv_vec:
4294    case INDEX_op_shrv_vec:
4295        switch (vece) {
4296        case MO_16:
4297            return have_avx512bw;
4298        case MO_32:
4299        case MO_64:
4300            return have_avx2;
4301        }
4302        return 0;
4303    case INDEX_op_sarv_vec:
4304        switch (vece) {
4305        case MO_16:
4306            return have_avx512bw;
4307        case MO_32:
4308            return have_avx2;
4309        case MO_64:
4310            return have_avx512vl;
4311        }
4312        return 0;
4313    case INDEX_op_rotlv_vec:
4314    case INDEX_op_rotrv_vec:
4315        switch (vece) {
4316        case MO_16:
4317            return have_avx512vbmi2 ? -1 : 0;
4318        case MO_32:
4319        case MO_64:
4320            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4321        }
4322        return 0;
4323
4324    case INDEX_op_mul_vec:
4325        switch (vece) {
4326        case MO_8:
4327            return -1;
4328        case MO_64:
4329            return have_avx512dq;
4330        }
4331        return 1;
4332
4333    case INDEX_op_ssadd_vec:
4334    case INDEX_op_usadd_vec:
4335    case INDEX_op_sssub_vec:
4336    case INDEX_op_ussub_vec:
4337        return vece <= MO_16;
4338    case INDEX_op_smin_vec:
4339    case INDEX_op_smax_vec:
4340    case INDEX_op_umin_vec:
4341    case INDEX_op_umax_vec:
4342    case INDEX_op_abs_vec:
4343        return vece <= MO_32 || have_avx512vl;
4344
4345    default:
4346        return 0;
4347    }
4348}
4349
4350static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4351                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4352{
4353    uint8_t mask;
4354
4355    tcg_debug_assert(vece == MO_8);
4356    if (right) {
4357        mask = 0xff >> imm;
4358        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4359    } else {
4360        mask = 0xff << imm;
4361        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4362    }
4363    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4364}
4365
4366static void expand_vec_sari(TCGType type, unsigned vece,
4367                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4368{
4369    TCGv_vec t1, t2;
4370
4371    switch (vece) {
4372    case MO_8:
4373        /* Unpack to 16-bit, shift, and repack.  */
4374        t1 = tcg_temp_new_vec(type);
4375        t2 = tcg_temp_new_vec(type);
4376        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4377                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4378        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4379                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4380        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4381        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4382        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4383                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4384        tcg_temp_free_vec(t1);
4385        tcg_temp_free_vec(t2);
4386        break;
4387
4388    case MO_64:
4389        t1 = tcg_temp_new_vec(type);
4390        if (imm <= 32) {
4391            /*
4392             * We can emulate a small sign extend by performing an arithmetic
4393             * 32-bit shift and overwriting the high half of a 64-bit logical
4394             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4395             * does not, so we have to bound the smaller shift -- we get the
4396             * same result in the high half either way.
4397             */
4398            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4399            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4400            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4401                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4402                      tcgv_vec_arg(t1), 0xaa);
4403        } else {
4404            /* Otherwise we will need to use a compare vs 0 to produce
4405             * the sign-extend, shift and merge.
4406             */
4407            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4408                            tcg_constant_vec(type, MO_64, 0), v1);
4409            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4410            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4411            tcg_gen_or_vec(MO_64, v0, v0, t1);
4412        }
4413        tcg_temp_free_vec(t1);
4414        break;
4415
4416    default:
4417        g_assert_not_reached();
4418    }
4419}
4420
4421static void expand_vec_rotli(TCGType type, unsigned vece,
4422                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4423{
4424    TCGv_vec t;
4425
4426    if (vece != MO_8 && have_avx512vbmi2) {
4427        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4428                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4429        return;
4430    }
4431
4432    t = tcg_temp_new_vec(type);
4433    tcg_gen_shli_vec(vece, t, v1, imm);
4434    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4435    tcg_gen_or_vec(vece, v0, v0, t);
4436    tcg_temp_free_vec(t);
4437}
4438
4439static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4440                            TCGv_vec v1, TCGv_vec sh, bool right)
4441{
4442    TCGv_vec t;
4443
4444    if (have_avx512vbmi2) {
4445        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4446                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4447                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4448        return;
4449    }
4450
4451    t = tcg_temp_new_vec(type);
4452    tcg_gen_dupi_vec(vece, t, 8 << vece);
4453    tcg_gen_sub_vec(vece, t, t, sh);
4454    if (right) {
4455        tcg_gen_shlv_vec(vece, t, v1, t);
4456        tcg_gen_shrv_vec(vece, v0, v1, sh);
4457    } else {
4458        tcg_gen_shrv_vec(vece, t, v1, t);
4459        tcg_gen_shlv_vec(vece, v0, v1, sh);
4460    }
4461    tcg_gen_or_vec(vece, v0, v0, t);
4462    tcg_temp_free_vec(t);
4463}
4464
4465static void expand_vec_rotls(TCGType type, unsigned vece,
4466                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4467{
4468    TCGv_vec t = tcg_temp_new_vec(type);
4469
4470    tcg_debug_assert(vece != MO_8);
4471
4472    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4473        tcg_gen_dup_i32_vec(vece, t, lsh);
4474        if (vece >= MO_32) {
4475            tcg_gen_rotlv_vec(vece, v0, v1, t);
4476        } else {
4477            expand_vec_rotv(type, vece, v0, v1, t, false);
4478        }
4479    } else {
4480        TCGv_i32 rsh = tcg_temp_new_i32();
4481
4482        tcg_gen_neg_i32(rsh, lsh);
4483        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4484        tcg_gen_shls_vec(vece, t, v1, lsh);
4485        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4486        tcg_gen_or_vec(vece, v0, v0, t);
4487
4488        tcg_temp_free_i32(rsh);
4489    }
4490
4491    tcg_temp_free_vec(t);
4492}
4493
4494static void expand_vec_mul(TCGType type, unsigned vece,
4495                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4496{
4497    TCGv_vec t1, t2, t3, t4, zero;
4498
4499    tcg_debug_assert(vece == MO_8);
4500
4501    /*
4502     * Unpack v1 bytes to words, 0 | x.
4503     * Unpack v2 bytes to words, y | 0.
4504     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4505     * Shift logical right by 8 bits to clear the high 8 bytes before
4506     * using an unsigned saturated pack.
4507     *
4508     * The difference between the V64, V128 and V256 cases is merely how
4509     * we distribute the expansion between temporaries.
4510     */
4511    switch (type) {
4512    case TCG_TYPE_V64:
4513        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4514        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4515        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4516        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4517                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4518        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4519                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4520        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4521        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4522        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4523                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4524        tcg_temp_free_vec(t1);
4525        tcg_temp_free_vec(t2);
4526        break;
4527
4528    case TCG_TYPE_V128:
4529    case TCG_TYPE_V256:
4530        t1 = tcg_temp_new_vec(type);
4531        t2 = tcg_temp_new_vec(type);
4532        t3 = tcg_temp_new_vec(type);
4533        t4 = tcg_temp_new_vec(type);
4534        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4535        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4536                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4537        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4538                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4539        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4540                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4541        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4542                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4543        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4544        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4545        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4546        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4547        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4548                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4549        tcg_temp_free_vec(t1);
4550        tcg_temp_free_vec(t2);
4551        tcg_temp_free_vec(t3);
4552        tcg_temp_free_vec(t4);
4553        break;
4554
4555    default:
4556        g_assert_not_reached();
4557    }
4558}
4559
4560static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4561                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4562{
4563    /*
4564     * Without AVX512, there are no 64-bit unsigned comparisons.
4565     * We must bias the inputs so that they become signed.
4566     * All other swapping and inversion are handled during code generation.
4567     */
4568    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4569        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4570        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4571        TCGv_vec t1 = tcg_temp_new_vec(type);
4572        TCGv_vec t2 = tcg_temp_new_vec(type);
4573        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4574
4575        tcg_gen_sub_vec(vece, t1, v1, t3);
4576        tcg_gen_sub_vec(vece, t2, v2, t3);
4577        *a1 = tcgv_vec_arg(t1);
4578        *a2 = tcgv_vec_arg(t2);
4579        cond = tcg_signed_cond(cond);
4580    }
4581    return cond;
4582}
4583
4584static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4585                           TCGArg a1, TCGArg a2, TCGCond cond)
4586{
4587    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4588    /* Expand directly; do not recurse.  */
4589    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4590}
4591
4592static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4593                              TCGArg a1, TCGArg a2,
4594                              TCGArg a3, TCGArg a4, TCGCond cond)
4595{
4596    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4597    /* Expand directly; do not recurse.  */
4598    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4599}
4600
4601void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4602                       TCGArg a0, ...)
4603{
4604    va_list va;
4605    TCGArg a1, a2, a3, a4, a5;
4606    TCGv_vec v0, v1, v2;
4607
4608    va_start(va, a0);
4609    a1 = va_arg(va, TCGArg);
4610    a2 = va_arg(va, TCGArg);
4611    v0 = temp_tcgv_vec(arg_temp(a0));
4612    v1 = temp_tcgv_vec(arg_temp(a1));
4613
4614    switch (opc) {
4615    case INDEX_op_shli_vec:
4616        expand_vec_shi(type, vece, false, v0, v1, a2);
4617        break;
4618    case INDEX_op_shri_vec:
4619        expand_vec_shi(type, vece, true, v0, v1, a2);
4620        break;
4621    case INDEX_op_sari_vec:
4622        expand_vec_sari(type, vece, v0, v1, a2);
4623        break;
4624
4625    case INDEX_op_rotli_vec:
4626        expand_vec_rotli(type, vece, v0, v1, a2);
4627        break;
4628
4629    case INDEX_op_rotls_vec:
4630        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4631        break;
4632
4633    case INDEX_op_rotlv_vec:
4634        v2 = temp_tcgv_vec(arg_temp(a2));
4635        expand_vec_rotv(type, vece, v0, v1, v2, false);
4636        break;
4637    case INDEX_op_rotrv_vec:
4638        v2 = temp_tcgv_vec(arg_temp(a2));
4639        expand_vec_rotv(type, vece, v0, v1, v2, true);
4640        break;
4641
4642    case INDEX_op_mul_vec:
4643        v2 = temp_tcgv_vec(arg_temp(a2));
4644        expand_vec_mul(type, vece, v0, v1, v2);
4645        break;
4646
4647    case INDEX_op_cmp_vec:
4648        a3 = va_arg(va, TCGArg);
4649        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4650        break;
4651
4652    case INDEX_op_cmpsel_vec:
4653        a3 = va_arg(va, TCGArg);
4654        a4 = va_arg(va, TCGArg);
4655        a5 = va_arg(va, TCGArg);
4656        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4657        break;
4658
4659    default:
4660        break;
4661    }
4662
4663    va_end(va);
4664}
4665
4666static const int tcg_target_callee_save_regs[] = {
4667#if TCG_TARGET_REG_BITS == 64
4668    TCG_REG_RBP,
4669    TCG_REG_RBX,
4670#if defined(_WIN64)
4671    TCG_REG_RDI,
4672    TCG_REG_RSI,
4673#endif
4674    TCG_REG_R12,
4675    TCG_REG_R13,
4676    TCG_REG_R14, /* Currently used for the global env. */
4677    TCG_REG_R15,
4678#else
4679    TCG_REG_EBP, /* Currently used for the global env. */
4680    TCG_REG_EBX,
4681    TCG_REG_ESI,
4682    TCG_REG_EDI,
4683#endif
4684};
4685
4686/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4687   and tcg_register_jit.  */
4688
4689#define PUSH_SIZE \
4690    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4691     * (TCG_TARGET_REG_BITS / 8))
4692
4693#define FRAME_SIZE \
4694    ((PUSH_SIZE \
4695      + TCG_STATIC_CALL_ARGS_SIZE \
4696      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4697      + TCG_TARGET_STACK_ALIGN - 1) \
4698     & ~(TCG_TARGET_STACK_ALIGN - 1))
4699
4700/* Generate global QEMU prologue and epilogue code */
4701static void tcg_target_qemu_prologue(TCGContext *s)
4702{
4703    int i, stack_addend;
4704
4705    /* TB prologue */
4706
4707    /* Reserve some stack space, also for TCG temps.  */
4708    stack_addend = FRAME_SIZE - PUSH_SIZE;
4709    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4710                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4711
4712    /* Save all callee saved registers.  */
4713    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4714        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4715    }
4716
4717    if (!tcg_use_softmmu && guest_base) {
4718        int seg = setup_guest_base_seg();
4719        if (seg != 0) {
4720            x86_guest_base.seg = seg;
4721        } else if (guest_base == (int32_t)guest_base) {
4722            x86_guest_base.ofs = guest_base;
4723        } else {
4724            assert(TCG_TARGET_REG_BITS == 64);
4725            /* Choose R12 because, as a base, it requires a SIB byte. */
4726            x86_guest_base.index = TCG_REG_R12;
4727            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4728            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4729        }
4730    }
4731
4732    if (TCG_TARGET_REG_BITS == 32) {
4733        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4734                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4735        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4736        /* jmp *tb.  */
4737        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4738                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4739                             + stack_addend);
4740    } else {
4741        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4742        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4743        /* jmp *tb.  */
4744        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4745    }
4746
4747    /*
4748     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4749     * and fall through to the rest of the epilogue.
4750     */
4751    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4752    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4753
4754    /* TB epilogue */
4755    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4756
4757    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4758
4759    if (have_avx2) {
4760        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4761    }
4762    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4763        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4764    }
4765    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4766}
4767
4768static void tcg_out_tb_start(TCGContext *s)
4769{
4770    /* nothing to do */
4771}
4772
4773static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4774{
4775    memset(p, 0x90, count);
4776}
4777
4778static void tcg_target_init(TCGContext *s)
4779{
4780    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4781    if (TCG_TARGET_REG_BITS == 64) {
4782        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4783    }
4784    if (have_avx1) {
4785        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4786        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4787    }
4788    if (have_avx2) {
4789        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4790    }
4791
4792    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4793    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4794    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4795    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4796    if (TCG_TARGET_REG_BITS == 64) {
4797#if !defined(_WIN64)
4798        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4799        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4800#endif
4801        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4802        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4803        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4804        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4805    }
4806
4807    s->reserved_regs = 0;
4808    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4809    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4810#ifdef _WIN64
4811    /* These are call saved, and we don't save them, so don't use them. */
4812    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4813    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4814    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4815    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4816    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4817    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4818    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4819    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4820    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4821    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4822#endif
4823}
4824
4825typedef struct {
4826    DebugFrameHeader h;
4827    uint8_t fde_def_cfa[4];
4828    uint8_t fde_reg_ofs[14];
4829} DebugFrame;
4830
4831/* We're expecting a 2 byte uleb128 encoded value.  */
4832QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4833
4834#if !defined(__ELF__)
4835    /* Host machine without ELF. */
4836#elif TCG_TARGET_REG_BITS == 64
4837#define ELF_HOST_MACHINE EM_X86_64
4838static const DebugFrame debug_frame = {
4839    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4840    .h.cie.id = -1,
4841    .h.cie.version = 1,
4842    .h.cie.code_align = 1,
4843    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4844    .h.cie.return_column = 16,
4845
4846    /* Total FDE size does not include the "len" member.  */
4847    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4848
4849    .fde_def_cfa = {
4850        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4851        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4852        (FRAME_SIZE >> 7)
4853    },
4854    .fde_reg_ofs = {
4855        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4856        /* The following ordering must match tcg_target_callee_save_regs.  */
4857        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4858        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4859        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4860        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4861        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4862        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4863    }
4864};
4865#else
4866#define ELF_HOST_MACHINE EM_386
4867static const DebugFrame debug_frame = {
4868    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4869    .h.cie.id = -1,
4870    .h.cie.version = 1,
4871    .h.cie.code_align = 1,
4872    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4873    .h.cie.return_column = 8,
4874
4875    /* Total FDE size does not include the "len" member.  */
4876    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4877
4878    .fde_def_cfa = {
4879        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4880        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4881        (FRAME_SIZE >> 7)
4882    },
4883    .fde_reg_ofs = {
4884        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4885        /* The following ordering must match tcg_target_callee_save_regs.  */
4886        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4887        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4888        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4889        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4890    }
4891};
4892#endif
4893
4894#if defined(ELF_HOST_MACHINE)
4895void tcg_register_jit(const void *buf, size_t buf_size)
4896{
4897    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4898}
4899#endif
4900