xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 641f1c53862aec64810c0b93b5b1de49d55fda92)
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 tgen_qemu_ld(TCGContext *s, TCGType type, TCGReg data,
2426                         TCGReg addr, MemOpIdx oi)
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, data, -1, h, type, get_memop(oi));
2433
2434    if (ldst) {
2435        ldst->type = type;
2436        ldst->datalo_reg = data;
2437        ldst->datahi_reg = -1;
2438        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2439    }
2440}
2441
2442static const TCGOutOpQemuLdSt outop_qemu_ld = {
2443    .base.static_constraint = C_O1_I1(r, L),
2444    .out = tgen_qemu_ld,
2445};
2446
2447static void tgen_qemu_ld2(TCGContext *s, TCGType type, TCGReg datalo,
2448                          TCGReg datahi, TCGReg addr, MemOpIdx oi)
2449{
2450    TCGLabelQemuLdst *ldst;
2451    HostAddress h;
2452
2453    ldst = prepare_host_addr(s, &h, addr, oi, true);
2454    tcg_out_qemu_ld_direct(s, datalo, datahi, h, type, get_memop(oi));
2455
2456    if (ldst) {
2457        ldst->type = type;
2458        ldst->datalo_reg = datalo;
2459        ldst->datahi_reg = datahi;
2460        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2461    }
2462}
2463
2464static const TCGOutOpQemuLdSt2 outop_qemu_ld2 = {
2465    .base.static_constraint = C_O2_I1(r, r, L),
2466    .out = tgen_qemu_ld2,
2467};
2468
2469static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2470                                   HostAddress h, MemOp memop)
2471{
2472    bool use_movbe = false;
2473    int movop = OPC_MOVL_EvGv;
2474
2475    /*
2476     * Do big-endian stores with movbe or system-mode.
2477     * User-only without movbe will have its swapping done generically.
2478     */
2479    if (memop & MO_BSWAP) {
2480        tcg_debug_assert(have_movbe);
2481        use_movbe = true;
2482        movop = OPC_MOVBE_MyGy;
2483    }
2484
2485    switch (memop & MO_SIZE) {
2486    case MO_8:
2487        /* This is handled with constraints in cset_qemu_st(). */
2488        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2489        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2490                                 datalo, h.base, h.index, 0, h.ofs);
2491        break;
2492    case MO_16:
2493        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2494                                 h.base, h.index, 0, h.ofs);
2495        break;
2496    case MO_32:
2497        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2498                                 h.base, h.index, 0, h.ofs);
2499        break;
2500    case MO_64:
2501        if (TCG_TARGET_REG_BITS == 64) {
2502            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2503                                     h.base, h.index, 0, h.ofs);
2504        } else {
2505            if (use_movbe) {
2506                TCGReg t = datalo;
2507                datalo = datahi;
2508                datahi = t;
2509            }
2510            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2511                                     h.base, h.index, 0, h.ofs);
2512            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2513                                     h.base, h.index, 0, h.ofs + 4);
2514        }
2515        break;
2516
2517    case MO_128:
2518        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2519
2520        /*
2521         * Without 16-byte atomicity, use integer regs.
2522         * That is where we have the data, and it allows bswaps.
2523         */
2524        if (h.aa.atom < MO_128) {
2525            if (use_movbe) {
2526                TCGReg t = datalo;
2527                datalo = datahi;
2528                datahi = t;
2529            }
2530            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2531                                     h.base, h.index, 0, h.ofs);
2532            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2533                                     h.base, h.index, 0, h.ofs + 8);
2534            break;
2535        }
2536
2537        /*
2538         * With 16-byte atomicity, a vector store is required.
2539         * If we already have 16-byte alignment, then VMOVDQA always works.
2540         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2541         * Else use we require a runtime test for alignment for VMOVDQA;
2542         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2543         */
2544        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2545        if (h.aa.align >= MO_128) {
2546            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2547                                         TCG_TMP_VEC, 0,
2548                                         h.base, h.index, 0, h.ofs);
2549        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2550            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2551                                         TCG_TMP_VEC, 0,
2552                                         h.base, h.index, 0, h.ofs);
2553        } else {
2554            TCGLabel *l1 = gen_new_label();
2555            TCGLabel *l2 = gen_new_label();
2556            int jcc;
2557
2558            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2559            tcg_out_jxx(s, jcc, l1, true);
2560
2561            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2562                                         TCG_TMP_VEC, 0,
2563                                         h.base, h.index, 0, h.ofs);
2564            tcg_out_jxx(s, JCC_JMP, l2, true);
2565
2566            tcg_out_label(s, l1);
2567            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2568                                         TCG_TMP_VEC, 0,
2569                                         h.base, h.index, 0, h.ofs);
2570            tcg_out_label(s, l2);
2571        }
2572        break;
2573
2574    default:
2575        g_assert_not_reached();
2576    }
2577}
2578
2579static void tgen_qemu_st(TCGContext *s, TCGType type, TCGReg data,
2580                         TCGReg addr, MemOpIdx oi)
2581{
2582    TCGLabelQemuLdst *ldst;
2583    HostAddress h;
2584
2585    ldst = prepare_host_addr(s, &h, addr, oi, false);
2586    tcg_out_qemu_st_direct(s, data, -1, h, get_memop(oi));
2587
2588    if (ldst) {
2589        ldst->type = type;
2590        ldst->datalo_reg = data;
2591        ldst->datahi_reg = -1;
2592        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2593    }
2594}
2595
2596static TCGConstraintSetIndex cset_qemu_st(TCGType type, unsigned flags)
2597{
2598    return flags == MO_8 ? C_O0_I2(s, L) : C_O0_I2(L, L);
2599}
2600
2601static const TCGOutOpQemuLdSt outop_qemu_st = {
2602    .base.static_constraint =
2603        TCG_TARGET_REG_BITS == 32 ? C_Dynamic : C_O0_I2(L, L),
2604    .base.dynamic_constraint =
2605        TCG_TARGET_REG_BITS == 32 ? cset_qemu_st : NULL,
2606    .out = tgen_qemu_st,
2607};
2608
2609static void tgen_qemu_st2(TCGContext *s, TCGType type, TCGReg datalo,
2610                          TCGReg datahi, TCGReg addr, MemOpIdx oi)
2611{
2612    TCGLabelQemuLdst *ldst;
2613    HostAddress h;
2614
2615    ldst = prepare_host_addr(s, &h, addr, oi, false);
2616    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2617
2618    if (ldst) {
2619        ldst->type = type;
2620        ldst->datalo_reg = datalo;
2621        ldst->datahi_reg = datahi;
2622        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2623    }
2624}
2625
2626static const TCGOutOpQemuLdSt2 outop_qemu_st2 = {
2627    .base.static_constraint = C_O0_I3(L, L, L),
2628    .out = tgen_qemu_st2,
2629};
2630
2631static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2632{
2633    /* Reuse the zeroing that exists for goto_ptr.  */
2634    if (a0 == 0) {
2635        tcg_out_jmp(s, tcg_code_gen_epilogue);
2636    } else {
2637        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2638        tcg_out_jmp(s, tb_ret_addr);
2639    }
2640}
2641
2642static void tcg_out_goto_tb(TCGContext *s, int which)
2643{
2644    /*
2645     * Jump displacement must be aligned for atomic patching;
2646     * see if we need to add extra nops before jump
2647     */
2648    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2649    if (gap != 1) {
2650        tcg_out_nopn(s, gap - 1);
2651    }
2652    tcg_out8(s, OPC_JMP_long); /* jmp im */
2653    set_jmp_insn_offset(s, which);
2654    tcg_out32(s, 0);
2655    set_jmp_reset_offset(s, which);
2656}
2657
2658static void tcg_out_goto_ptr(TCGContext *s, TCGReg a0)
2659{
2660    /* Jump to the given host address (could be epilogue) */
2661    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2662}
2663
2664void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2665                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2666{
2667    /* patch the branch destination */
2668    uintptr_t addr = tb->jmp_target_addr[n];
2669    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2670    /* no need to flush icache explicitly */
2671}
2672
2673
2674static void tgen_add(TCGContext *s, TCGType type,
2675                     TCGReg a0, TCGReg a1, TCGReg a2)
2676{
2677    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2678
2679    if (a0 == a1) {
2680        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2681    } else if (a0 == a2) {
2682        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2683    } else {
2684        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2685    }
2686}
2687
2688static void tgen_addi(TCGContext *s, TCGType type,
2689                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2690{
2691    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2692
2693    if (a0 == a1) {
2694        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2695    } else {
2696        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2697    }
2698}
2699
2700static const TCGOutOpBinary outop_add = {
2701    .base.static_constraint = C_O1_I2(r, r, re),
2702    .out_rrr = tgen_add,
2703    .out_rri = tgen_addi,
2704};
2705
2706static void tgen_addco(TCGContext *s, TCGType type,
2707                       TCGReg a0, TCGReg a1, TCGReg a2)
2708{
2709    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2710    tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2711}
2712
2713static void tgen_addco_imm(TCGContext *s, TCGType type,
2714                           TCGReg a0, TCGReg a1, tcg_target_long a2)
2715{
2716    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2717    tgen_arithi(s, ARITH_ADD + rexw, a0, a2, true);
2718}
2719
2720static const TCGOutOpBinary outop_addco = {
2721    .base.static_constraint = C_O1_I2(r, 0, re),
2722    .out_rrr = tgen_addco,
2723    .out_rri = tgen_addco_imm,
2724};
2725
2726static void tgen_addcio(TCGContext *s, TCGType type,
2727                        TCGReg a0, TCGReg a1, TCGReg a2)
2728{
2729    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2730    tgen_arithr(s, ARITH_ADC + rexw, a0, a2);
2731}
2732
2733static void tgen_addcio_imm(TCGContext *s, TCGType type,
2734                            TCGReg a0, TCGReg a1, tcg_target_long a2)
2735{
2736    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2737    tgen_arithi(s, ARITH_ADC + rexw, a0, a2, true);
2738}
2739
2740static const TCGOutOpBinary outop_addcio = {
2741    .base.static_constraint = C_O1_I2(r, 0, re),
2742    .out_rrr = tgen_addcio,
2743    .out_rri = tgen_addcio_imm,
2744};
2745
2746static void tgen_addci_rrr(TCGContext *s, TCGType type,
2747                           TCGReg a0, TCGReg a1, TCGReg a2)
2748{
2749    /* Because "0O" is not a valid constraint, we must match ourselves. */
2750    if (a0 == a2) {
2751        tgen_addcio(s, type, a0, a0, a1);
2752    } else {
2753        tcg_out_mov(s, type, a0, a1);
2754        tgen_addcio(s, type, a0, a0, a2);
2755    }
2756}
2757
2758static void tgen_addci_rri(TCGContext *s, TCGType type,
2759                           TCGReg a0, TCGReg a1, tcg_target_long a2)
2760{
2761    tcg_out_mov(s, type, a0, a1);
2762    tgen_addcio_imm(s, type, a0, a0, a2);
2763}
2764
2765static void tgen_addci_rir(TCGContext *s, TCGType type,
2766                           TCGReg a0, tcg_target_long a1, TCGReg a2)
2767{
2768    tgen_addci_rri(s, type, a0, a2, a1);
2769}
2770
2771static void tgen_addci_rii(TCGContext *s, TCGType type, TCGReg a0,
2772                           tcg_target_long a1, tcg_target_long a2)
2773{
2774    if (a2 == 0) {
2775        /* Implement 0 + 0 + C with -(x - x - c). */
2776        tgen_arithr(s, ARITH_SBB, a0, a0);
2777        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, a0);
2778    } else {
2779        tcg_out_movi(s, type, a0, a2);
2780        tgen_addcio_imm(s, type, a0, a0, a1);
2781    }
2782}
2783
2784static const TCGOutOpAddSubCarry outop_addci = {
2785    .base.static_constraint = C_O1_I2(r, rO, re),
2786    .out_rrr = tgen_addci_rrr,
2787    .out_rri = tgen_addci_rri,
2788    .out_rir = tgen_addci_rir,
2789    .out_rii = tgen_addci_rii,
2790};
2791
2792static void tcg_out_set_carry(TCGContext *s)
2793{
2794    tcg_out8(s, OPC_STC);
2795}
2796
2797static void tgen_and(TCGContext *s, TCGType type,
2798                     TCGReg a0, TCGReg a1, TCGReg a2)
2799{
2800    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2801    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2802}
2803
2804static void tgen_andi(TCGContext *s, TCGType type,
2805                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2806{
2807    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2808    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2809}
2810
2811static const TCGOutOpBinary outop_and = {
2812    .base.static_constraint = C_O1_I2(r, 0, reZ),
2813    .out_rrr = tgen_and,
2814    .out_rri = tgen_andi,
2815};
2816
2817static void tgen_andc(TCGContext *s, TCGType type,
2818                      TCGReg a0, TCGReg a1, TCGReg a2)
2819{
2820    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2821    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2822}
2823
2824static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2825{
2826    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2827}
2828
2829static const TCGOutOpBinary outop_andc = {
2830    .base.static_constraint = C_Dynamic,
2831    .base.dynamic_constraint = cset_andc,
2832    .out_rrr = tgen_andc,
2833};
2834
2835static void tgen_clz(TCGContext *s, TCGType type,
2836                     TCGReg a0, TCGReg a1, TCGReg a2)
2837{
2838    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2839    int jcc;
2840
2841    if (have_lzcnt) {
2842        tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2843        jcc = JCC_JB;
2844    } else {
2845        /* Recall that the output of BSR is the index not the count.  */
2846        tcg_out_modrm(s, OPC_BSR + rexw, a0, a1);
2847        tgen_arithi(s, ARITH_XOR + rexw, a0, rexw ? 63 : 31, 0);
2848
2849        /* Since we have destroyed the flags from BSR, we have to re-test.  */
2850        jcc = tcg_out_cmp(s, TCG_COND_EQ, a1, 0, 1, rexw);
2851    }
2852    tcg_out_cmov(s, jcc, rexw, a0, a2);
2853}
2854
2855static void tgen_clzi(TCGContext *s, TCGType type,
2856                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2857{
2858    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2859    tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2860}
2861
2862static TCGConstraintSetIndex cset_clz(TCGType type, unsigned flags)
2863{
2864    return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2865}
2866
2867static const TCGOutOpBinary outop_clz = {
2868    .base.static_constraint = C_Dynamic,
2869    .base.dynamic_constraint = cset_clz,
2870    .out_rrr = tgen_clz,
2871    .out_rri = tgen_clzi,
2872};
2873
2874static void tgen_ctpop(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2875{
2876    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2877    tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2878}
2879
2880static TCGConstraintSetIndex cset_ctpop(TCGType type, unsigned flags)
2881{
2882    return have_popcnt ? C_O1_I1(r, r) : C_NotImplemented;
2883}
2884
2885static const TCGOutOpUnary outop_ctpop = {
2886    .base.static_constraint = C_Dynamic,
2887    .base.dynamic_constraint = cset_ctpop,
2888    .out_rr = tgen_ctpop,
2889};
2890
2891static void tgen_ctz(TCGContext *s, TCGType type,
2892                     TCGReg a0, TCGReg a1, TCGReg a2)
2893{
2894    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2895    int jcc;
2896
2897    if (have_bmi1) {
2898        tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2899        jcc = JCC_JB;
2900    } else {
2901        tcg_out_modrm(s, OPC_BSF + rexw, a0, a1);
2902        jcc = JCC_JE;
2903    }
2904    tcg_out_cmov(s, jcc, rexw, a0, a2);
2905}
2906
2907static void tgen_ctzi(TCGContext *s, TCGType type,
2908                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2909{
2910    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2911    tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2912}
2913
2914static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
2915{
2916    return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2917}
2918
2919static const TCGOutOpBinary outop_ctz = {
2920    .base.static_constraint = C_Dynamic,
2921    .base.dynamic_constraint = cset_ctz,
2922    .out_rrr = tgen_ctz,
2923    .out_rri = tgen_ctzi,
2924};
2925
2926static const TCGOutOpBinary outop_divs = {
2927    .base.static_constraint = C_NotImplemented,
2928};
2929
2930static void tgen_divs2(TCGContext *s, TCGType type,
2931                       TCGReg a0, TCGReg a1, TCGReg a4)
2932{
2933    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2934    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, a4);
2935}
2936
2937static const TCGOutOpDivRem outop_divs2 = {
2938    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2939    .out_rr01r = tgen_divs2,
2940};
2941
2942static const TCGOutOpBinary outop_divu = {
2943    .base.static_constraint = C_NotImplemented,
2944};
2945
2946static void tgen_divu2(TCGContext *s, TCGType type,
2947                       TCGReg a0, TCGReg a1, TCGReg a4)
2948{
2949    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2950    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, a4);
2951}
2952
2953static const TCGOutOpDivRem outop_divu2 = {
2954    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2955    .out_rr01r = tgen_divu2,
2956};
2957
2958static const TCGOutOpBinary outop_eqv = {
2959    .base.static_constraint = C_NotImplemented,
2960};
2961
2962#if TCG_TARGET_REG_BITS == 64
2963static void tgen_extrh_i64_i32(TCGContext *s, TCGType t, TCGReg a0, TCGReg a1)
2964{
2965    tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2966}
2967
2968static const TCGOutOpUnary outop_extrh_i64_i32 = {
2969    .base.static_constraint = C_O1_I1(r, 0),
2970    .out_rr = tgen_extrh_i64_i32,
2971};
2972#endif /* TCG_TARGET_REG_BITS == 64 */
2973
2974static void tgen_mul(TCGContext *s, TCGType type,
2975                     TCGReg a0, TCGReg a1, TCGReg a2)
2976{
2977    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2978    tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2979}
2980
2981static void tgen_muli(TCGContext *s, TCGType type,
2982                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2983{
2984    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2985
2986    if (a2 == (int8_t)a2) {
2987        tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2988        tcg_out8(s, a2);
2989    } else {
2990        tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2991        tcg_out32(s, a2);
2992    }
2993}
2994
2995static const TCGOutOpBinary outop_mul = {
2996    .base.static_constraint = C_O1_I2(r, 0, re),
2997    .out_rrr = tgen_mul,
2998    .out_rri = tgen_muli,
2999};
3000
3001static void tgen_muls2(TCGContext *s, TCGType type,
3002                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
3003{
3004    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3005    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, a3);
3006}
3007
3008static const TCGOutOpMul2 outop_muls2 = {
3009    .base.static_constraint = C_O2_I2(a, d, a, r),
3010    .out_rrrr = tgen_muls2,
3011};
3012
3013static const TCGOutOpBinary outop_mulsh = {
3014    .base.static_constraint = C_NotImplemented,
3015};
3016
3017static const TCGOutOpBinary outop_muluh = {
3018    .base.static_constraint = C_NotImplemented,
3019};
3020
3021static void tgen_mulu2(TCGContext *s, TCGType type,
3022                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
3023{
3024    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3025    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, a3);
3026}
3027
3028static const TCGOutOpMul2 outop_mulu2 = {
3029    .base.static_constraint = C_O2_I2(a, d, a, r),
3030    .out_rrrr = tgen_mulu2,
3031};
3032
3033static const TCGOutOpBinary outop_nand = {
3034    .base.static_constraint = C_NotImplemented,
3035};
3036
3037static const TCGOutOpBinary outop_nor = {
3038    .base.static_constraint = C_NotImplemented,
3039};
3040
3041static void tgen_or(TCGContext *s, TCGType type,
3042                    TCGReg a0, TCGReg a1, TCGReg a2)
3043{
3044    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3045    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
3046}
3047
3048static void tgen_ori(TCGContext *s, TCGType type,
3049                     TCGReg a0, TCGReg a1, tcg_target_long a2)
3050{
3051    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3052    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
3053}
3054
3055static const TCGOutOpBinary outop_or = {
3056    .base.static_constraint = C_O1_I2(r, 0, re),
3057    .out_rrr = tgen_or,
3058    .out_rri = tgen_ori,
3059};
3060
3061static const TCGOutOpBinary outop_orc = {
3062    .base.static_constraint = C_NotImplemented,
3063};
3064
3065static const TCGOutOpBinary outop_rems = {
3066    .base.static_constraint = C_NotImplemented,
3067};
3068
3069static const TCGOutOpBinary outop_remu = {
3070    .base.static_constraint = C_NotImplemented,
3071};
3072
3073static void tgen_rotl(TCGContext *s, TCGType type,
3074                     TCGReg a0, TCGReg a1, TCGReg a2)
3075{
3076    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3077    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROL, a0);
3078}
3079
3080static void tgen_rotli(TCGContext *s, TCGType type,
3081                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3082{
3083    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3084    tcg_out_shifti(s, SHIFT_ROL + rexw, a0, a2);
3085}
3086
3087static const TCGOutOpBinary outop_rotl = {
3088    .base.static_constraint = C_O1_I2(r, 0, ci),
3089    .out_rrr = tgen_rotl,
3090    .out_rri = tgen_rotli,
3091};
3092
3093static void tgen_rotr(TCGContext *s, TCGType type,
3094                     TCGReg a0, TCGReg a1, TCGReg a2)
3095{
3096    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3097    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROR, a0);
3098}
3099
3100static void tgen_rotri(TCGContext *s, TCGType type,
3101                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3102{
3103    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3104    tcg_out_shifti(s, SHIFT_ROR + rexw, a0, a2);
3105}
3106
3107static const TCGOutOpBinary outop_rotr = {
3108    .base.static_constraint = C_O1_I2(r, 0, ci),
3109    .out_rrr = tgen_rotr,
3110    .out_rri = tgen_rotri,
3111};
3112
3113static TCGConstraintSetIndex cset_shift(TCGType type, unsigned flags)
3114{
3115    return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3116}
3117
3118static void tgen_sar(TCGContext *s, TCGType type,
3119                     TCGReg a0, TCGReg a1, TCGReg a2)
3120{
3121    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3122    if (have_bmi2) {
3123        tcg_out_vex_modrm(s, OPC_SARX + rexw, a0, a2, a1);
3124    } else {
3125        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SAR, a0);
3126    }
3127}
3128
3129static void tgen_sari(TCGContext *s, TCGType type,
3130                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3131{
3132    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3133
3134    tcg_out_mov(s, type, a0, a1);
3135    tcg_out_shifti(s, SHIFT_SAR + rexw, a0, a2);
3136}
3137
3138static const TCGOutOpBinary outop_sar = {
3139    .base.static_constraint = C_Dynamic,
3140    .base.dynamic_constraint = cset_shift,
3141    .out_rrr = tgen_sar,
3142    .out_rri = tgen_sari,
3143};
3144
3145static void tgen_shl(TCGContext *s, TCGType type,
3146                     TCGReg a0, TCGReg a1, TCGReg a2)
3147{
3148    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3149    if (have_bmi2) {
3150        tcg_out_vex_modrm(s, OPC_SHLX + rexw, a0, a2, a1);
3151    } else {
3152        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHL, a0);
3153    }
3154}
3155
3156static void tgen_shli(TCGContext *s, TCGType type,
3157                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3158{
3159    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3160
3161    /* For small constant 3-operand shift, use LEA.  */
3162    if (a0 != a1 && a2 >= 1 && a2 <= 3) {
3163        if (a2 == 1) {
3164            /* shl $1,a1,a0 -> lea (a1,a1),a0 */
3165            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
3166        } else {
3167            /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
3168            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
3169        }
3170        return;
3171    }
3172    tcg_out_mov(s, type, a0, a1);
3173    tcg_out_shifti(s, SHIFT_SHL + rexw, a0, a2);
3174}
3175
3176static const TCGOutOpBinary outop_shl = {
3177    .base.static_constraint = C_Dynamic,
3178    .base.dynamic_constraint = cset_shift,
3179    .out_rrr = tgen_shl,
3180    .out_rri = tgen_shli,
3181};
3182
3183static void tgen_shr(TCGContext *s, TCGType type,
3184                     TCGReg a0, TCGReg a1, TCGReg a2)
3185{
3186    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3187    if (have_bmi2) {
3188        tcg_out_vex_modrm(s, OPC_SHRX + rexw, a0, a2, a1);
3189    } else {
3190        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHR, a0);
3191    }
3192}
3193
3194static void tgen_shri(TCGContext *s, TCGType type,
3195                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3196{
3197    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3198
3199    tcg_out_mov(s, type, a0, a1);
3200    tcg_out_shifti(s, SHIFT_SHR + rexw, a0, a2);
3201}
3202
3203static const TCGOutOpBinary outop_shr = {
3204    .base.static_constraint = C_Dynamic,
3205    .base.dynamic_constraint = cset_shift,
3206    .out_rrr = tgen_shr,
3207    .out_rri = tgen_shri,
3208};
3209
3210static void tgen_sub(TCGContext *s, TCGType type,
3211                     TCGReg a0, TCGReg a1, TCGReg a2)
3212{
3213    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3214    tgen_arithr(s, ARITH_SUB + rexw, a0, a2);
3215}
3216
3217static const TCGOutOpSubtract outop_sub = {
3218    .base.static_constraint = C_O1_I2(r, 0, r),
3219    .out_rrr = tgen_sub,
3220};
3221
3222static void tgen_subbo_rri(TCGContext *s, TCGType type,
3223                           TCGReg a0, TCGReg a1, tcg_target_long a2)
3224{
3225    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3226    tgen_arithi(s, ARITH_SUB + rexw, a0, a2, 1);
3227}
3228
3229static const TCGOutOpAddSubCarry outop_subbo = {
3230    .base.static_constraint = C_O1_I2(r, 0, re),
3231    .out_rrr = tgen_sub,
3232    .out_rri = tgen_subbo_rri,
3233};
3234
3235static void tgen_subbio_rrr(TCGContext *s, TCGType type,
3236                            TCGReg a0, TCGReg a1, TCGReg a2)
3237{
3238    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3239    tgen_arithr(s, ARITH_SBB + rexw, a0, a2);
3240}
3241
3242static void tgen_subbio_rri(TCGContext *s, TCGType type,
3243                            TCGReg a0, TCGReg a1, tcg_target_long a2)
3244{
3245    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3246    tgen_arithi(s, ARITH_SBB + rexw, a0, a2, 1);
3247}
3248
3249static const TCGOutOpAddSubCarry outop_subbio = {
3250    .base.static_constraint = C_O1_I2(r, 0, re),
3251    .out_rrr = tgen_subbio_rrr,
3252    .out_rri = tgen_subbio_rri,
3253};
3254
3255#define outop_subbi  outop_subbio
3256
3257static void tcg_out_set_borrow(TCGContext *s)
3258{
3259    tcg_out8(s, OPC_STC);
3260}
3261
3262static void tgen_xor(TCGContext *s, TCGType type,
3263                     TCGReg a0, TCGReg a1, TCGReg a2)
3264{
3265    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3266    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
3267}
3268
3269static void tgen_xori(TCGContext *s, TCGType type,
3270                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3271{
3272    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3273    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
3274}
3275
3276static const TCGOutOpBinary outop_xor = {
3277    .base.static_constraint = C_O1_I2(r, 0, re),
3278    .out_rrr = tgen_xor,
3279    .out_rri = tgen_xori,
3280};
3281
3282static void tgen_bswap16(TCGContext *s, TCGType type,
3283                         TCGReg a0, TCGReg a1, unsigned flags)
3284{
3285    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3286
3287    if (flags & TCG_BSWAP_OS) {
3288        /* Output must be sign-extended. */
3289        if (rexw) {
3290            tcg_out_bswap64(s, a0);
3291            tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
3292        } else {
3293            tcg_out_bswap32(s, a0);
3294            tcg_out_shifti(s, SHIFT_SAR, a0, 16);
3295        }
3296    } else if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
3297        /* Output must be zero-extended, but input isn't. */
3298        tcg_out_bswap32(s, a0);
3299        tcg_out_shifti(s, SHIFT_SHR, a0, 16);
3300    } else {
3301        tcg_out_rolw_8(s, a0);
3302    }
3303}
3304
3305static const TCGOutOpBswap outop_bswap16 = {
3306    .base.static_constraint = C_O1_I1(r, 0),
3307    .out_rr = tgen_bswap16,
3308};
3309
3310static void tgen_bswap32(TCGContext *s, TCGType type,
3311                         TCGReg a0, TCGReg a1, unsigned flags)
3312{
3313    tcg_out_bswap32(s, a0);
3314    if (flags & TCG_BSWAP_OS) {
3315        tcg_out_ext32s(s, a0, a0);
3316    }
3317}
3318
3319static const TCGOutOpBswap outop_bswap32 = {
3320    .base.static_constraint = C_O1_I1(r, 0),
3321    .out_rr = tgen_bswap32,
3322};
3323
3324#if TCG_TARGET_REG_BITS == 64
3325static void tgen_bswap64(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3326{
3327    tcg_out_bswap64(s, a0);
3328}
3329
3330static const TCGOutOpUnary outop_bswap64 = {
3331    .base.static_constraint = C_O1_I1(r, 0),
3332    .out_rr = tgen_bswap64,
3333};
3334#endif
3335
3336static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3337{
3338    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3339    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
3340}
3341
3342static const TCGOutOpUnary outop_neg = {
3343    .base.static_constraint = C_O1_I1(r, 0),
3344    .out_rr = tgen_neg,
3345};
3346
3347static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3348{
3349    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3350    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
3351}
3352
3353static const TCGOutOpUnary outop_not = {
3354    .base.static_constraint = C_O1_I1(r, 0),
3355    .out_rr = tgen_not,
3356};
3357
3358static void tgen_deposit(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3359                         TCGReg a2, unsigned ofs, unsigned len)
3360{
3361    if (ofs == 0 && len == 8) {
3362        tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3363    } else if (ofs == 0 && len == 16) {
3364        tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3365    } else if (TCG_TARGET_REG_BITS == 32 && ofs == 8 && len == 8) {
3366        tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3367    } else {
3368        g_assert_not_reached();
3369    }
3370}
3371
3372static void tgen_depositi(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3373                          tcg_target_long a2, unsigned ofs, unsigned len)
3374{
3375    if (ofs == 0 && len == 8) {
3376        tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0), 0, a0, 0);
3377        tcg_out8(s, a2);
3378    } else if (ofs == 0 && len == 16) {
3379        tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0), 0, a0, 0);
3380        tcg_out16(s, a2);
3381    } else if (TCG_TARGET_REG_BITS == 32 && ofs == 8 && len == 8) {
3382        tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3383        tcg_out8(s, a2);
3384    } else {
3385        g_assert_not_reached();
3386    }
3387}
3388
3389static const TCGOutOpDeposit outop_deposit = {
3390    .base.static_constraint = C_O1_I2(q, 0, qi),
3391    .out_rrr = tgen_deposit,
3392    .out_rri = tgen_depositi,
3393};
3394
3395static void tgen_extract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3396                         unsigned ofs, unsigned len)
3397{
3398    if (ofs == 0) {
3399        switch (len) {
3400        case 8:
3401            tcg_out_ext8u(s, a0, a1);
3402            return;
3403        case 16:
3404            tcg_out_ext16u(s, a0, a1);
3405            return;
3406        case 32:
3407            tcg_out_ext32u(s, a0, a1);
3408            return;
3409        }
3410    } else if (TCG_TARGET_REG_BITS == 64 && ofs + len == 32) {
3411        /* This is a 32-bit zero-extending right shift.  */
3412        tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3413        tcg_out_shifti(s, SHIFT_SHR, a0, ofs);
3414        return;
3415    } else if (ofs == 8 && len == 8) {
3416        /*
3417         * On the off-chance that we can use the high-byte registers.
3418         * Otherwise we emit the same ext16 + shift pattern that we
3419         * would have gotten from the normal tcg-op.c expansion.
3420         */
3421        if (a1 < 4 && (TCG_TARGET_REG_BITS == 32 || a0 < 8)) {
3422            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3423        } else {
3424            tcg_out_ext16u(s, a0, a1);
3425            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3426        }
3427        return;
3428    }
3429    g_assert_not_reached();
3430}
3431
3432static const TCGOutOpExtract outop_extract = {
3433    .base.static_constraint = C_O1_I1(r, r),
3434    .out_rr = tgen_extract,
3435};
3436
3437static void tgen_sextract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3438                          unsigned ofs, unsigned len)
3439{
3440    if (ofs == 0) {
3441        switch (len) {
3442        case 8:
3443            tcg_out_ext8s(s, type, a0, a1);
3444            return;
3445        case 16:
3446            tcg_out_ext16s(s, type, a0, a1);
3447            return;
3448        case 32:
3449            tcg_out_ext32s(s, a0, a1);
3450            return;
3451        }
3452    } else if (ofs == 8 && len == 8) {
3453        if (type == TCG_TYPE_I32 && a1 < 4 && a0 < 8) {
3454            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3455        } else {
3456            tcg_out_ext16s(s, type, a0, a1);
3457            tgen_sari(s, type, a0, a0, 8);
3458        }
3459        return;
3460    }
3461    g_assert_not_reached();
3462}
3463
3464static const TCGOutOpExtract outop_sextract = {
3465    .base.static_constraint = C_O1_I1(r, r),
3466    .out_rr = tgen_sextract,
3467};
3468
3469static void tgen_extract2(TCGContext *s, TCGType type, TCGReg a0,
3470                          TCGReg a1, TCGReg a2, unsigned shr)
3471{
3472    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3473
3474    /* Note that SHRD outputs to the r/m operand.  */
3475    tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3476    tcg_out8(s, shr);
3477}
3478
3479static const TCGOutOpExtract2 outop_extract2 = {
3480    .base.static_constraint = C_O1_I2(r, 0, r),
3481    .out_rrr = tgen_extract2,
3482};
3483
3484static void tgen_ld8u(TCGContext *s, TCGType type, TCGReg dest,
3485                      TCGReg base, ptrdiff_t offset)
3486{
3487    tcg_out_modrm_offset(s, OPC_MOVZBL, dest, base, offset);
3488}
3489
3490static const TCGOutOpLoad outop_ld8u = {
3491    .base.static_constraint = C_O1_I1(r, r),
3492    .out = tgen_ld8u,
3493};
3494
3495static void tgen_ld8s(TCGContext *s, TCGType type, TCGReg dest,
3496                      TCGReg base, ptrdiff_t offset)
3497{
3498    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3499    tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, dest, base, offset);
3500}
3501
3502static const TCGOutOpLoad outop_ld8s = {
3503    .base.static_constraint = C_O1_I1(r, r),
3504    .out = tgen_ld8s,
3505};
3506
3507static void tgen_ld16u(TCGContext *s, TCGType type, TCGReg dest,
3508                       TCGReg base, ptrdiff_t offset)
3509{
3510    tcg_out_modrm_offset(s, OPC_MOVZWL, dest, base, offset);
3511}
3512
3513static const TCGOutOpLoad outop_ld16u = {
3514    .base.static_constraint = C_O1_I1(r, r),
3515    .out = tgen_ld16u,
3516};
3517
3518static void tgen_ld16s(TCGContext *s, TCGType type, TCGReg dest,
3519                       TCGReg base, ptrdiff_t offset)
3520{
3521    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3522    tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, dest, base, offset);
3523}
3524
3525static const TCGOutOpLoad outop_ld16s = {
3526    .base.static_constraint = C_O1_I1(r, r),
3527    .out = tgen_ld16s,
3528};
3529
3530#if TCG_TARGET_REG_BITS == 64
3531static void tgen_ld32u(TCGContext *s, TCGType type, TCGReg dest,
3532                       TCGReg base, ptrdiff_t offset)
3533{
3534    tcg_out_modrm_offset(s, OPC_MOVL_GvEv, dest, base, offset);
3535}
3536
3537static const TCGOutOpLoad outop_ld32u = {
3538    .base.static_constraint = C_O1_I1(r, r),
3539    .out = tgen_ld32u,
3540};
3541
3542static void tgen_ld32s(TCGContext *s, TCGType type, TCGReg dest,
3543                       TCGReg base, ptrdiff_t offset)
3544{
3545    tcg_out_modrm_offset(s, OPC_MOVSLQ, dest, base, offset);
3546}
3547
3548static const TCGOutOpLoad outop_ld32s = {
3549    .base.static_constraint = C_O1_I1(r, r),
3550    .out = tgen_ld32s,
3551};
3552#endif
3553
3554static void tgen_st8_r(TCGContext *s, TCGType type, TCGReg data,
3555                       TCGReg base, ptrdiff_t offset)
3556{
3557    tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, data, base, offset);
3558}
3559
3560static void tgen_st8_i(TCGContext *s, TCGType type, tcg_target_long data,
3561                       TCGReg base, ptrdiff_t offset)
3562{
3563    tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, base, offset);
3564    tcg_out8(s, data);
3565}
3566
3567static const TCGOutOpStore outop_st8 = {
3568    .base.static_constraint = C_O0_I2(qi, r),
3569    .out_r = tgen_st8_r,
3570    .out_i = tgen_st8_i,
3571};
3572
3573static void tgen_st16_r(TCGContext *s, TCGType type, TCGReg data,
3574                        TCGReg base, ptrdiff_t offset)
3575{
3576    tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, data, base, offset);
3577}
3578
3579static void tgen_st16_i(TCGContext *s, TCGType type, tcg_target_long data,
3580                        TCGReg base, ptrdiff_t offset)
3581{
3582    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, base, offset);
3583    tcg_out16(s, data);
3584}
3585
3586static const TCGOutOpStore outop_st16 = {
3587    .base.static_constraint = C_O0_I2(ri, r),
3588    .out_r = tgen_st16_r,
3589    .out_i = tgen_st16_i,
3590};
3591
3592static void tgen_st_i(TCGContext *s, TCGType type, tcg_target_long data,
3593                      TCGReg base, ptrdiff_t offset)
3594{
3595    bool ok = tcg_out_sti(s, type, data, base, offset);
3596    tcg_debug_assert(ok);
3597}
3598
3599static const TCGOutOpStore outop_st = {
3600    .base.static_constraint = C_O0_I2(re, r),
3601    .out_r = tcg_out_st,
3602    .out_i = tgen_st_i,
3603};
3604
3605static int const umin_insn[4] = {
3606    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3607};
3608
3609static int const umax_insn[4] = {
3610    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3611};
3612
3613static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3614                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3615{
3616    static int const cmpeq_insn[4] = {
3617        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3618    };
3619    static int const cmpgt_insn[4] = {
3620        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3621    };
3622
3623    enum {
3624        NEED_INV  = 1,
3625        NEED_SWAP = 2,
3626        NEED_UMIN = 4,
3627        NEED_UMAX = 8,
3628        INVALID   = 16,
3629    };
3630    static const uint8_t cond_fixup[16] = {
3631        [0 ... 15] = INVALID,
3632        [TCG_COND_EQ] = 0,
3633        [TCG_COND_GT] = 0,
3634        [TCG_COND_NE] = NEED_INV,
3635        [TCG_COND_LE] = NEED_INV,
3636        [TCG_COND_LT] = NEED_SWAP,
3637        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3638        [TCG_COND_LEU] = NEED_UMIN,
3639        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3640        [TCG_COND_GEU] = NEED_UMAX,
3641        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3642    };
3643    int fixup = cond_fixup[cond];
3644
3645    assert(!(fixup & INVALID));
3646
3647    if (fixup & NEED_INV) {
3648        cond = tcg_invert_cond(cond);
3649    }
3650
3651    if (fixup & NEED_SWAP) {
3652        TCGReg swap = v1;
3653        v1 = v2;
3654        v2 = swap;
3655        cond = tcg_swap_cond(cond);
3656    }
3657
3658    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3659        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3660
3661        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3662        assert(vece <= MO_32);
3663
3664        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3665        v2 = TCG_TMP_VEC;
3666        cond = TCG_COND_EQ;
3667    }
3668
3669    switch (cond) {
3670    case TCG_COND_EQ:
3671        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3672        break;
3673    case TCG_COND_GT:
3674        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3675        break;
3676    default:
3677        g_assert_not_reached();
3678    }
3679    return fixup & NEED_INV;
3680}
3681
3682static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3683                               TCGReg v1, TCGReg v2, TCGCond cond)
3684{
3685    static const int cmpm_insn[2][4] = {
3686        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3687        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3688    };
3689    static const int testm_insn[4] = {
3690        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3691    };
3692    static const int testnm_insn[4] = {
3693        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3694    };
3695
3696    static const int cond_ext[16] = {
3697        [TCG_COND_EQ] = 0,
3698        [TCG_COND_NE] = 4,
3699        [TCG_COND_LT] = 1,
3700        [TCG_COND_LTU] = 1,
3701        [TCG_COND_LE] = 2,
3702        [TCG_COND_LEU] = 2,
3703        [TCG_COND_NEVER] = 3,
3704        [TCG_COND_GE] = 5,
3705        [TCG_COND_GEU] = 5,
3706        [TCG_COND_GT] = 6,
3707        [TCG_COND_GTU] = 6,
3708        [TCG_COND_ALWAYS] = 7,
3709    };
3710
3711    switch (cond) {
3712    case TCG_COND_TSTNE:
3713        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3714        break;
3715    case TCG_COND_TSTEQ:
3716        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3717        break;
3718    default:
3719        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3720                               /* k1 */ 1, v1, v2, type);
3721        tcg_out8(s, cond_ext[cond]);
3722        break;
3723    }
3724}
3725
3726static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3727                              unsigned vece, TCGReg dest)
3728{
3729    static const int movm_insn[] = {
3730        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3731    };
3732    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3733}
3734
3735static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3736                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3737{
3738    /*
3739     * With avx512, we have a complete set of comparisons into mask.
3740     * Unless there's a single insn expansion for the comparision,
3741     * expand via a mask in k1.
3742     */
3743    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3744        && cond != TCG_COND_EQ
3745        && cond != TCG_COND_LT
3746        && cond != TCG_COND_GT) {
3747        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3748        tcg_out_k1_to_vec(s, type, vece, v0);
3749        return;
3750    }
3751
3752    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3753        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3754        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3755    }
3756}
3757
3758static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3759                                  TCGReg v0, TCGReg c1, TCGReg c2,
3760                                  TCGReg v3, TCGReg v4, TCGCond cond)
3761{
3762    static const int vpblendm_insn[] = {
3763        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3764    };
3765    bool z = false;
3766
3767    /* Swap to place constant in V4 to take advantage of zero-masking. */
3768    if (!v3) {
3769        z = true;
3770        v3 = v4;
3771        cond = tcg_invert_cond(cond);
3772    }
3773
3774    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3775    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3776                            /* k1 */1, z, type);
3777}
3778
3779static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3780                               TCGReg v0, TCGReg c1, TCGReg c2,
3781                               TCGReg v3, TCGReg v4, TCGCond cond)
3782{
3783    bool inv;
3784
3785    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3786        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3787        return;
3788    }
3789
3790    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3791
3792    /*
3793     * Since XMM0 is 16, the only way we get 0 into V3
3794     * is via the constant zero constraint.
3795     */
3796    if (!v3) {
3797        if (inv) {
3798            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3799        } else {
3800            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3801        }
3802    } else {
3803        if (inv) {
3804            TCGReg swap = v3;
3805            v3 = v4;
3806            v4 = swap;
3807        }
3808        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3809        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3810    }
3811}
3812
3813static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3814                           unsigned vecl, unsigned vece,
3815                           const TCGArg args[TCG_MAX_OP_ARGS],
3816                           const int const_args[TCG_MAX_OP_ARGS])
3817{
3818    static int const add_insn[4] = {
3819        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3820    };
3821    static int const ssadd_insn[4] = {
3822        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3823    };
3824    static int const usadd_insn[4] = {
3825        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3826    };
3827    static int const sub_insn[4] = {
3828        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3829    };
3830    static int const sssub_insn[4] = {
3831        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3832    };
3833    static int const ussub_insn[4] = {
3834        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3835    };
3836    static int const mul_insn[4] = {
3837        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3838    };
3839    static int const shift_imm_insn[4] = {
3840        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3841    };
3842    static int const punpckl_insn[4] = {
3843        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3844    };
3845    static int const punpckh_insn[4] = {
3846        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3847    };
3848    static int const packss_insn[4] = {
3849        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3850    };
3851    static int const packus_insn[4] = {
3852        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3853    };
3854    static int const smin_insn[4] = {
3855        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3856    };
3857    static int const smax_insn[4] = {
3858        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3859    };
3860    static int const rotlv_insn[4] = {
3861        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3862    };
3863    static int const rotrv_insn[4] = {
3864        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3865    };
3866    static int const shlv_insn[4] = {
3867        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3868    };
3869    static int const shrv_insn[4] = {
3870        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3871    };
3872    static int const sarv_insn[4] = {
3873        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3874    };
3875    static int const shls_insn[4] = {
3876        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3877    };
3878    static int const shrs_insn[4] = {
3879        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3880    };
3881    static int const sars_insn[4] = {
3882        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3883    };
3884    static int const vpshldi_insn[4] = {
3885        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3886    };
3887    static int const vpshldv_insn[4] = {
3888        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3889    };
3890    static int const vpshrdv_insn[4] = {
3891        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3892    };
3893    static int const abs_insn[4] = {
3894        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3895    };
3896
3897    TCGType type = vecl + TCG_TYPE_V64;
3898    int insn, sub;
3899    TCGArg a0, a1, a2, a3;
3900
3901    a0 = args[0];
3902    a1 = args[1];
3903    a2 = args[2];
3904
3905    switch (opc) {
3906    case INDEX_op_add_vec:
3907        insn = add_insn[vece];
3908        goto gen_simd;
3909    case INDEX_op_ssadd_vec:
3910        insn = ssadd_insn[vece];
3911        goto gen_simd;
3912    case INDEX_op_usadd_vec:
3913        insn = usadd_insn[vece];
3914        goto gen_simd;
3915    case INDEX_op_sub_vec:
3916        insn = sub_insn[vece];
3917        goto gen_simd;
3918    case INDEX_op_sssub_vec:
3919        insn = sssub_insn[vece];
3920        goto gen_simd;
3921    case INDEX_op_ussub_vec:
3922        insn = ussub_insn[vece];
3923        goto gen_simd;
3924    case INDEX_op_mul_vec:
3925        insn = mul_insn[vece];
3926        goto gen_simd;
3927    case INDEX_op_and_vec:
3928        insn = OPC_PAND;
3929        goto gen_simd;
3930    case INDEX_op_or_vec:
3931        insn = OPC_POR;
3932        goto gen_simd;
3933    case INDEX_op_xor_vec:
3934        insn = OPC_PXOR;
3935        goto gen_simd;
3936    case INDEX_op_smin_vec:
3937        insn = smin_insn[vece];
3938        goto gen_simd;
3939    case INDEX_op_umin_vec:
3940        insn = umin_insn[vece];
3941        goto gen_simd;
3942    case INDEX_op_smax_vec:
3943        insn = smax_insn[vece];
3944        goto gen_simd;
3945    case INDEX_op_umax_vec:
3946        insn = umax_insn[vece];
3947        goto gen_simd;
3948    case INDEX_op_shlv_vec:
3949        insn = shlv_insn[vece];
3950        goto gen_simd;
3951    case INDEX_op_shrv_vec:
3952        insn = shrv_insn[vece];
3953        goto gen_simd;
3954    case INDEX_op_sarv_vec:
3955        insn = sarv_insn[vece];
3956        goto gen_simd;
3957    case INDEX_op_rotlv_vec:
3958        insn = rotlv_insn[vece];
3959        goto gen_simd;
3960    case INDEX_op_rotrv_vec:
3961        insn = rotrv_insn[vece];
3962        goto gen_simd;
3963    case INDEX_op_shls_vec:
3964        insn = shls_insn[vece];
3965        goto gen_simd;
3966    case INDEX_op_shrs_vec:
3967        insn = shrs_insn[vece];
3968        goto gen_simd;
3969    case INDEX_op_sars_vec:
3970        insn = sars_insn[vece];
3971        goto gen_simd;
3972    case INDEX_op_x86_punpckl_vec:
3973        insn = punpckl_insn[vece];
3974        goto gen_simd;
3975    case INDEX_op_x86_punpckh_vec:
3976        insn = punpckh_insn[vece];
3977        goto gen_simd;
3978    case INDEX_op_x86_packss_vec:
3979        insn = packss_insn[vece];
3980        goto gen_simd;
3981    case INDEX_op_x86_packus_vec:
3982        insn = packus_insn[vece];
3983        goto gen_simd;
3984    case INDEX_op_x86_vpshldv_vec:
3985        insn = vpshldv_insn[vece];
3986        a1 = a2;
3987        a2 = args[3];
3988        goto gen_simd;
3989    case INDEX_op_x86_vpshrdv_vec:
3990        insn = vpshrdv_insn[vece];
3991        a1 = a2;
3992        a2 = args[3];
3993        goto gen_simd;
3994#if TCG_TARGET_REG_BITS == 32
3995    case INDEX_op_dup2_vec:
3996        /* First merge the two 32-bit inputs to a single 64-bit element. */
3997        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3998        /* Then replicate the 64-bit elements across the rest of the vector. */
3999        if (type != TCG_TYPE_V64) {
4000            tcg_out_dup_vec(s, type, MO_64, a0, a0);
4001        }
4002        break;
4003#endif
4004    case INDEX_op_abs_vec:
4005        insn = abs_insn[vece];
4006        a2 = a1;
4007        a1 = 0;
4008        goto gen_simd;
4009    gen_simd:
4010        tcg_debug_assert(insn != OPC_UD2);
4011        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
4012        break;
4013
4014    case INDEX_op_cmp_vec:
4015        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
4016        break;
4017
4018    case INDEX_op_cmpsel_vec:
4019        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
4020                           args[3], args[4], args[5]);
4021        break;
4022
4023    case INDEX_op_andc_vec:
4024        insn = OPC_PANDN;
4025        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
4026        break;
4027
4028    case INDEX_op_shli_vec:
4029        insn = shift_imm_insn[vece];
4030        sub = 6;
4031        goto gen_shift;
4032    case INDEX_op_shri_vec:
4033        insn = shift_imm_insn[vece];
4034        sub = 2;
4035        goto gen_shift;
4036    case INDEX_op_sari_vec:
4037        if (vece == MO_64) {
4038            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
4039        } else {
4040            insn = shift_imm_insn[vece];
4041        }
4042        sub = 4;
4043        goto gen_shift;
4044    case INDEX_op_rotli_vec:
4045        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
4046        if (vece == MO_64) {
4047            insn |= P_VEXW;
4048        }
4049        sub = 1;
4050        goto gen_shift;
4051    gen_shift:
4052        tcg_debug_assert(vece != MO_8);
4053        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
4054        tcg_out8(s, a2);
4055        break;
4056
4057    case INDEX_op_ld_vec:
4058        tcg_out_ld(s, type, a0, a1, a2);
4059        break;
4060    case INDEX_op_st_vec:
4061        tcg_out_st(s, type, a0, a1, a2);
4062        break;
4063    case INDEX_op_dupm_vec:
4064        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
4065        break;
4066
4067    case INDEX_op_x86_shufps_vec:
4068        insn = OPC_SHUFPS;
4069        sub = args[3];
4070        goto gen_simd_imm8;
4071    case INDEX_op_x86_blend_vec:
4072        if (vece == MO_16) {
4073            insn = OPC_PBLENDW;
4074        } else if (vece == MO_32) {
4075            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
4076        } else {
4077            g_assert_not_reached();
4078        }
4079        sub = args[3];
4080        goto gen_simd_imm8;
4081    case INDEX_op_x86_vperm2i128_vec:
4082        insn = OPC_VPERM2I128;
4083        sub = args[3];
4084        goto gen_simd_imm8;
4085    case INDEX_op_x86_vpshldi_vec:
4086        insn = vpshldi_insn[vece];
4087        sub = args[3];
4088        goto gen_simd_imm8;
4089
4090    case INDEX_op_not_vec:
4091        insn = OPC_VPTERNLOGQ;
4092        a2 = a1;
4093        sub = 0x33; /* !B */
4094        goto gen_simd_imm8;
4095    case INDEX_op_nor_vec:
4096        insn = OPC_VPTERNLOGQ;
4097        sub = 0x11; /* norCB */
4098        goto gen_simd_imm8;
4099    case INDEX_op_nand_vec:
4100        insn = OPC_VPTERNLOGQ;
4101        sub = 0x77; /* nandCB */
4102        goto gen_simd_imm8;
4103    case INDEX_op_eqv_vec:
4104        insn = OPC_VPTERNLOGQ;
4105        sub = 0x99; /* xnorCB */
4106        goto gen_simd_imm8;
4107    case INDEX_op_orc_vec:
4108        insn = OPC_VPTERNLOGQ;
4109        sub = 0xdd; /* orB!C */
4110        goto gen_simd_imm8;
4111
4112    case INDEX_op_bitsel_vec:
4113        insn = OPC_VPTERNLOGQ;
4114        a3 = args[3];
4115        if (a0 == a1) {
4116            a1 = a2;
4117            a2 = a3;
4118            sub = 0xca; /* A?B:C */
4119        } else if (a0 == a2) {
4120            a2 = a3;
4121            sub = 0xe2; /* B?A:C */
4122        } else {
4123            tcg_out_mov(s, type, a0, a3);
4124            sub = 0xb8; /* B?C:A */
4125        }
4126        goto gen_simd_imm8;
4127
4128    gen_simd_imm8:
4129        tcg_debug_assert(insn != OPC_UD2);
4130        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
4131        tcg_out8(s, sub);
4132        break;
4133
4134    case INDEX_op_x86_psrldq_vec:
4135        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
4136        tcg_out8(s, a2);
4137        break;
4138
4139    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
4140    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
4141    default:
4142        g_assert_not_reached();
4143    }
4144}
4145
4146static TCGConstraintSetIndex
4147tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
4148{
4149    switch (op) {
4150    case INDEX_op_ld_vec:
4151    case INDEX_op_dupm_vec:
4152        return C_O1_I1(x, r);
4153
4154    case INDEX_op_st_vec:
4155        return C_O0_I2(x, r);
4156
4157    case INDEX_op_add_vec:
4158    case INDEX_op_sub_vec:
4159    case INDEX_op_mul_vec:
4160    case INDEX_op_and_vec:
4161    case INDEX_op_or_vec:
4162    case INDEX_op_xor_vec:
4163    case INDEX_op_andc_vec:
4164    case INDEX_op_orc_vec:
4165    case INDEX_op_nand_vec:
4166    case INDEX_op_nor_vec:
4167    case INDEX_op_eqv_vec:
4168    case INDEX_op_ssadd_vec:
4169    case INDEX_op_usadd_vec:
4170    case INDEX_op_sssub_vec:
4171    case INDEX_op_ussub_vec:
4172    case INDEX_op_smin_vec:
4173    case INDEX_op_umin_vec:
4174    case INDEX_op_smax_vec:
4175    case INDEX_op_umax_vec:
4176    case INDEX_op_shlv_vec:
4177    case INDEX_op_shrv_vec:
4178    case INDEX_op_sarv_vec:
4179    case INDEX_op_rotlv_vec:
4180    case INDEX_op_rotrv_vec:
4181    case INDEX_op_shls_vec:
4182    case INDEX_op_shrs_vec:
4183    case INDEX_op_sars_vec:
4184    case INDEX_op_cmp_vec:
4185    case INDEX_op_x86_shufps_vec:
4186    case INDEX_op_x86_blend_vec:
4187    case INDEX_op_x86_packss_vec:
4188    case INDEX_op_x86_packus_vec:
4189    case INDEX_op_x86_vperm2i128_vec:
4190    case INDEX_op_x86_punpckl_vec:
4191    case INDEX_op_x86_punpckh_vec:
4192    case INDEX_op_x86_vpshldi_vec:
4193#if TCG_TARGET_REG_BITS == 32
4194    case INDEX_op_dup2_vec:
4195#endif
4196        return C_O1_I2(x, x, x);
4197
4198    case INDEX_op_abs_vec:
4199    case INDEX_op_dup_vec:
4200    case INDEX_op_not_vec:
4201    case INDEX_op_shli_vec:
4202    case INDEX_op_shri_vec:
4203    case INDEX_op_sari_vec:
4204    case INDEX_op_rotli_vec:
4205    case INDEX_op_x86_psrldq_vec:
4206        return C_O1_I1(x, x);
4207
4208    case INDEX_op_x86_vpshldv_vec:
4209    case INDEX_op_x86_vpshrdv_vec:
4210        return C_O1_I3(x, 0, x, x);
4211
4212    case INDEX_op_bitsel_vec:
4213        return C_O1_I3(x, x, x, x);
4214    case INDEX_op_cmpsel_vec:
4215        return C_O1_I4(x, x, x, xO, x);
4216
4217    default:
4218        return C_NotImplemented;
4219    }
4220}
4221
4222int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
4223{
4224    switch (opc) {
4225    case INDEX_op_add_vec:
4226    case INDEX_op_sub_vec:
4227    case INDEX_op_and_vec:
4228    case INDEX_op_or_vec:
4229    case INDEX_op_xor_vec:
4230    case INDEX_op_andc_vec:
4231    case INDEX_op_orc_vec:
4232    case INDEX_op_nand_vec:
4233    case INDEX_op_nor_vec:
4234    case INDEX_op_eqv_vec:
4235    case INDEX_op_not_vec:
4236    case INDEX_op_bitsel_vec:
4237        return 1;
4238    case INDEX_op_cmp_vec:
4239    case INDEX_op_cmpsel_vec:
4240        return -1;
4241
4242    case INDEX_op_rotli_vec:
4243        return have_avx512vl && vece >= MO_32 ? 1 : -1;
4244
4245    case INDEX_op_shli_vec:
4246    case INDEX_op_shri_vec:
4247        /* We must expand the operation for MO_8.  */
4248        return vece == MO_8 ? -1 : 1;
4249
4250    case INDEX_op_sari_vec:
4251        switch (vece) {
4252        case MO_8:
4253            return -1;
4254        case MO_16:
4255        case MO_32:
4256            return 1;
4257        case MO_64:
4258            if (have_avx512vl) {
4259                return 1;
4260            }
4261            /*
4262             * We can emulate this for MO_64, but it does not pay off
4263             * unless we're producing at least 4 values.
4264             */
4265            return type >= TCG_TYPE_V256 ? -1 : 0;
4266        }
4267        return 0;
4268
4269    case INDEX_op_shls_vec:
4270    case INDEX_op_shrs_vec:
4271        return vece >= MO_16;
4272    case INDEX_op_sars_vec:
4273        switch (vece) {
4274        case MO_16:
4275        case MO_32:
4276            return 1;
4277        case MO_64:
4278            return have_avx512vl;
4279        }
4280        return 0;
4281    case INDEX_op_rotls_vec:
4282        return vece >= MO_16 ? -1 : 0;
4283
4284    case INDEX_op_shlv_vec:
4285    case INDEX_op_shrv_vec:
4286        switch (vece) {
4287        case MO_16:
4288            return have_avx512bw;
4289        case MO_32:
4290        case MO_64:
4291            return have_avx2;
4292        }
4293        return 0;
4294    case INDEX_op_sarv_vec:
4295        switch (vece) {
4296        case MO_16:
4297            return have_avx512bw;
4298        case MO_32:
4299            return have_avx2;
4300        case MO_64:
4301            return have_avx512vl;
4302        }
4303        return 0;
4304    case INDEX_op_rotlv_vec:
4305    case INDEX_op_rotrv_vec:
4306        switch (vece) {
4307        case MO_16:
4308            return have_avx512vbmi2 ? -1 : 0;
4309        case MO_32:
4310        case MO_64:
4311            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4312        }
4313        return 0;
4314
4315    case INDEX_op_mul_vec:
4316        switch (vece) {
4317        case MO_8:
4318            return -1;
4319        case MO_64:
4320            return have_avx512dq;
4321        }
4322        return 1;
4323
4324    case INDEX_op_ssadd_vec:
4325    case INDEX_op_usadd_vec:
4326    case INDEX_op_sssub_vec:
4327    case INDEX_op_ussub_vec:
4328        return vece <= MO_16;
4329    case INDEX_op_smin_vec:
4330    case INDEX_op_smax_vec:
4331    case INDEX_op_umin_vec:
4332    case INDEX_op_umax_vec:
4333    case INDEX_op_abs_vec:
4334        return vece <= MO_32 || have_avx512vl;
4335
4336    default:
4337        return 0;
4338    }
4339}
4340
4341static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4342                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4343{
4344    uint8_t mask;
4345
4346    tcg_debug_assert(vece == MO_8);
4347    if (right) {
4348        mask = 0xff >> imm;
4349        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4350    } else {
4351        mask = 0xff << imm;
4352        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4353    }
4354    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4355}
4356
4357static void expand_vec_sari(TCGType type, unsigned vece,
4358                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4359{
4360    TCGv_vec t1, t2;
4361
4362    switch (vece) {
4363    case MO_8:
4364        /* Unpack to 16-bit, shift, and repack.  */
4365        t1 = tcg_temp_new_vec(type);
4366        t2 = tcg_temp_new_vec(type);
4367        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4368                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4369        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4370                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4371        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4372        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4373        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4374                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4375        tcg_temp_free_vec(t1);
4376        tcg_temp_free_vec(t2);
4377        break;
4378
4379    case MO_64:
4380        t1 = tcg_temp_new_vec(type);
4381        if (imm <= 32) {
4382            /*
4383             * We can emulate a small sign extend by performing an arithmetic
4384             * 32-bit shift and overwriting the high half of a 64-bit logical
4385             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4386             * does not, so we have to bound the smaller shift -- we get the
4387             * same result in the high half either way.
4388             */
4389            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4390            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4391            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4392                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4393                      tcgv_vec_arg(t1), 0xaa);
4394        } else {
4395            /* Otherwise we will need to use a compare vs 0 to produce
4396             * the sign-extend, shift and merge.
4397             */
4398            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4399                            tcg_constant_vec(type, MO_64, 0), v1);
4400            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4401            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4402            tcg_gen_or_vec(MO_64, v0, v0, t1);
4403        }
4404        tcg_temp_free_vec(t1);
4405        break;
4406
4407    default:
4408        g_assert_not_reached();
4409    }
4410}
4411
4412static void expand_vec_rotli(TCGType type, unsigned vece,
4413                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4414{
4415    TCGv_vec t;
4416
4417    if (vece != MO_8 && have_avx512vbmi2) {
4418        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4419                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4420        return;
4421    }
4422
4423    t = tcg_temp_new_vec(type);
4424    tcg_gen_shli_vec(vece, t, v1, imm);
4425    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4426    tcg_gen_or_vec(vece, v0, v0, t);
4427    tcg_temp_free_vec(t);
4428}
4429
4430static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4431                            TCGv_vec v1, TCGv_vec sh, bool right)
4432{
4433    TCGv_vec t;
4434
4435    if (have_avx512vbmi2) {
4436        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4437                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4438                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4439        return;
4440    }
4441
4442    t = tcg_temp_new_vec(type);
4443    tcg_gen_dupi_vec(vece, t, 8 << vece);
4444    tcg_gen_sub_vec(vece, t, t, sh);
4445    if (right) {
4446        tcg_gen_shlv_vec(vece, t, v1, t);
4447        tcg_gen_shrv_vec(vece, v0, v1, sh);
4448    } else {
4449        tcg_gen_shrv_vec(vece, t, v1, t);
4450        tcg_gen_shlv_vec(vece, v0, v1, sh);
4451    }
4452    tcg_gen_or_vec(vece, v0, v0, t);
4453    tcg_temp_free_vec(t);
4454}
4455
4456static void expand_vec_rotls(TCGType type, unsigned vece,
4457                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4458{
4459    TCGv_vec t = tcg_temp_new_vec(type);
4460
4461    tcg_debug_assert(vece != MO_8);
4462
4463    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4464        tcg_gen_dup_i32_vec(vece, t, lsh);
4465        if (vece >= MO_32) {
4466            tcg_gen_rotlv_vec(vece, v0, v1, t);
4467        } else {
4468            expand_vec_rotv(type, vece, v0, v1, t, false);
4469        }
4470    } else {
4471        TCGv_i32 rsh = tcg_temp_new_i32();
4472
4473        tcg_gen_neg_i32(rsh, lsh);
4474        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4475        tcg_gen_shls_vec(vece, t, v1, lsh);
4476        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4477        tcg_gen_or_vec(vece, v0, v0, t);
4478
4479        tcg_temp_free_i32(rsh);
4480    }
4481
4482    tcg_temp_free_vec(t);
4483}
4484
4485static void expand_vec_mul(TCGType type, unsigned vece,
4486                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4487{
4488    TCGv_vec t1, t2, t3, t4, zero;
4489
4490    tcg_debug_assert(vece == MO_8);
4491
4492    /*
4493     * Unpack v1 bytes to words, 0 | x.
4494     * Unpack v2 bytes to words, y | 0.
4495     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4496     * Shift logical right by 8 bits to clear the high 8 bytes before
4497     * using an unsigned saturated pack.
4498     *
4499     * The difference between the V64, V128 and V256 cases is merely how
4500     * we distribute the expansion between temporaries.
4501     */
4502    switch (type) {
4503    case TCG_TYPE_V64:
4504        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4505        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4506        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4507        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4508                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4509        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4510                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4511        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4512        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4513        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4514                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4515        tcg_temp_free_vec(t1);
4516        tcg_temp_free_vec(t2);
4517        break;
4518
4519    case TCG_TYPE_V128:
4520    case TCG_TYPE_V256:
4521        t1 = tcg_temp_new_vec(type);
4522        t2 = tcg_temp_new_vec(type);
4523        t3 = tcg_temp_new_vec(type);
4524        t4 = tcg_temp_new_vec(type);
4525        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4526        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4527                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4528        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4529                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4530        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4531                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4532        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4533                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4534        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4535        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4536        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4537        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4538        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4539                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4540        tcg_temp_free_vec(t1);
4541        tcg_temp_free_vec(t2);
4542        tcg_temp_free_vec(t3);
4543        tcg_temp_free_vec(t4);
4544        break;
4545
4546    default:
4547        g_assert_not_reached();
4548    }
4549}
4550
4551static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4552                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4553{
4554    /*
4555     * Without AVX512, there are no 64-bit unsigned comparisons.
4556     * We must bias the inputs so that they become signed.
4557     * All other swapping and inversion are handled during code generation.
4558     */
4559    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4560        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4561        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4562        TCGv_vec t1 = tcg_temp_new_vec(type);
4563        TCGv_vec t2 = tcg_temp_new_vec(type);
4564        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4565
4566        tcg_gen_sub_vec(vece, t1, v1, t3);
4567        tcg_gen_sub_vec(vece, t2, v2, t3);
4568        *a1 = tcgv_vec_arg(t1);
4569        *a2 = tcgv_vec_arg(t2);
4570        cond = tcg_signed_cond(cond);
4571    }
4572    return cond;
4573}
4574
4575static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4576                           TCGArg a1, TCGArg a2, TCGCond cond)
4577{
4578    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4579    /* Expand directly; do not recurse.  */
4580    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4581}
4582
4583static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4584                              TCGArg a1, TCGArg a2,
4585                              TCGArg a3, TCGArg a4, TCGCond cond)
4586{
4587    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4588    /* Expand directly; do not recurse.  */
4589    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4590}
4591
4592void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4593                       TCGArg a0, ...)
4594{
4595    va_list va;
4596    TCGArg a1, a2, a3, a4, a5;
4597    TCGv_vec v0, v1, v2;
4598
4599    va_start(va, a0);
4600    a1 = va_arg(va, TCGArg);
4601    a2 = va_arg(va, TCGArg);
4602    v0 = temp_tcgv_vec(arg_temp(a0));
4603    v1 = temp_tcgv_vec(arg_temp(a1));
4604
4605    switch (opc) {
4606    case INDEX_op_shli_vec:
4607        expand_vec_shi(type, vece, false, v0, v1, a2);
4608        break;
4609    case INDEX_op_shri_vec:
4610        expand_vec_shi(type, vece, true, v0, v1, a2);
4611        break;
4612    case INDEX_op_sari_vec:
4613        expand_vec_sari(type, vece, v0, v1, a2);
4614        break;
4615
4616    case INDEX_op_rotli_vec:
4617        expand_vec_rotli(type, vece, v0, v1, a2);
4618        break;
4619
4620    case INDEX_op_rotls_vec:
4621        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4622        break;
4623
4624    case INDEX_op_rotlv_vec:
4625        v2 = temp_tcgv_vec(arg_temp(a2));
4626        expand_vec_rotv(type, vece, v0, v1, v2, false);
4627        break;
4628    case INDEX_op_rotrv_vec:
4629        v2 = temp_tcgv_vec(arg_temp(a2));
4630        expand_vec_rotv(type, vece, v0, v1, v2, true);
4631        break;
4632
4633    case INDEX_op_mul_vec:
4634        v2 = temp_tcgv_vec(arg_temp(a2));
4635        expand_vec_mul(type, vece, v0, v1, v2);
4636        break;
4637
4638    case INDEX_op_cmp_vec:
4639        a3 = va_arg(va, TCGArg);
4640        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4641        break;
4642
4643    case INDEX_op_cmpsel_vec:
4644        a3 = va_arg(va, TCGArg);
4645        a4 = va_arg(va, TCGArg);
4646        a5 = va_arg(va, TCGArg);
4647        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4648        break;
4649
4650    default:
4651        break;
4652    }
4653
4654    va_end(va);
4655}
4656
4657static const int tcg_target_callee_save_regs[] = {
4658#if TCG_TARGET_REG_BITS == 64
4659    TCG_REG_RBP,
4660    TCG_REG_RBX,
4661#if defined(_WIN64)
4662    TCG_REG_RDI,
4663    TCG_REG_RSI,
4664#endif
4665    TCG_REG_R12,
4666    TCG_REG_R13,
4667    TCG_REG_R14, /* Currently used for the global env. */
4668    TCG_REG_R15,
4669#else
4670    TCG_REG_EBP, /* Currently used for the global env. */
4671    TCG_REG_EBX,
4672    TCG_REG_ESI,
4673    TCG_REG_EDI,
4674#endif
4675};
4676
4677/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4678   and tcg_register_jit.  */
4679
4680#define PUSH_SIZE \
4681    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4682     * (TCG_TARGET_REG_BITS / 8))
4683
4684#define FRAME_SIZE \
4685    ((PUSH_SIZE \
4686      + TCG_STATIC_CALL_ARGS_SIZE \
4687      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4688      + TCG_TARGET_STACK_ALIGN - 1) \
4689     & ~(TCG_TARGET_STACK_ALIGN - 1))
4690
4691/* Generate global QEMU prologue and epilogue code */
4692static void tcg_target_qemu_prologue(TCGContext *s)
4693{
4694    int i, stack_addend;
4695
4696    /* TB prologue */
4697
4698    /* Reserve some stack space, also for TCG temps.  */
4699    stack_addend = FRAME_SIZE - PUSH_SIZE;
4700    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4701                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4702
4703    /* Save all callee saved registers.  */
4704    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4705        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4706    }
4707
4708    if (!tcg_use_softmmu && guest_base) {
4709        int seg = setup_guest_base_seg();
4710        if (seg != 0) {
4711            x86_guest_base.seg = seg;
4712        } else if (guest_base == (int32_t)guest_base) {
4713            x86_guest_base.ofs = guest_base;
4714        } else {
4715            assert(TCG_TARGET_REG_BITS == 64);
4716            /* Choose R12 because, as a base, it requires a SIB byte. */
4717            x86_guest_base.index = TCG_REG_R12;
4718            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4719            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4720        }
4721    }
4722
4723    if (TCG_TARGET_REG_BITS == 32) {
4724        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4725                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4726        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4727        /* jmp *tb.  */
4728        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4729                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4730                             + stack_addend);
4731    } else {
4732        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4733        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4734        /* jmp *tb.  */
4735        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4736    }
4737
4738    /*
4739     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4740     * and fall through to the rest of the epilogue.
4741     */
4742    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4743    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4744
4745    /* TB epilogue */
4746    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4747
4748    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4749
4750    if (have_avx2) {
4751        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4752    }
4753    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4754        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4755    }
4756    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4757}
4758
4759static void tcg_out_tb_start(TCGContext *s)
4760{
4761    /* nothing to do */
4762}
4763
4764static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4765{
4766    memset(p, 0x90, count);
4767}
4768
4769static void tcg_target_init(TCGContext *s)
4770{
4771    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4772    if (TCG_TARGET_REG_BITS == 64) {
4773        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4774    }
4775    if (have_avx1) {
4776        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4777        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4778    }
4779    if (have_avx2) {
4780        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4781    }
4782
4783    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4784    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4785    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4786    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4787    if (TCG_TARGET_REG_BITS == 64) {
4788#if !defined(_WIN64)
4789        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4790        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4791#endif
4792        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4793        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4794        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4795        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4796    }
4797
4798    s->reserved_regs = 0;
4799    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4800    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4801#ifdef _WIN64
4802    /* These are call saved, and we don't save them, so don't use them. */
4803    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4804    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4805    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4806    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4807    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4808    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4809    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4810    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4811    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4812    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4813#endif
4814}
4815
4816typedef struct {
4817    DebugFrameHeader h;
4818    uint8_t fde_def_cfa[4];
4819    uint8_t fde_reg_ofs[14];
4820} DebugFrame;
4821
4822/* We're expecting a 2 byte uleb128 encoded value.  */
4823QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4824
4825#if !defined(__ELF__)
4826    /* Host machine without ELF. */
4827#elif TCG_TARGET_REG_BITS == 64
4828#define ELF_HOST_MACHINE EM_X86_64
4829static const DebugFrame debug_frame = {
4830    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4831    .h.cie.id = -1,
4832    .h.cie.version = 1,
4833    .h.cie.code_align = 1,
4834    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4835    .h.cie.return_column = 16,
4836
4837    /* Total FDE size does not include the "len" member.  */
4838    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4839
4840    .fde_def_cfa = {
4841        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4842        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4843        (FRAME_SIZE >> 7)
4844    },
4845    .fde_reg_ofs = {
4846        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4847        /* The following ordering must match tcg_target_callee_save_regs.  */
4848        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4849        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4850        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4851        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4852        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4853        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4854    }
4855};
4856#else
4857#define ELF_HOST_MACHINE EM_386
4858static const DebugFrame debug_frame = {
4859    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4860    .h.cie.id = -1,
4861    .h.cie.version = 1,
4862    .h.cie.code_align = 1,
4863    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4864    .h.cie.return_column = 8,
4865
4866    /* Total FDE size does not include the "len" member.  */
4867    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4868
4869    .fde_def_cfa = {
4870        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4871        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4872        (FRAME_SIZE >> 7)
4873    },
4874    .fde_reg_ofs = {
4875        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4876        /* The following ordering must match tcg_target_callee_save_regs.  */
4877        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4878        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4879        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4880        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4881    }
4882};
4883#endif
4884
4885#if defined(ELF_HOST_MACHINE)
4886void tcg_register_jit(const void *buf, size_t buf_size)
4887{
4888    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4889}
4890#endif
4891